2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/slab.h>
19 #include <linux/hash.h>
20 #include <linux/sysfs.h>
21 #include <linux/dcache.h>
22 #include <linux/percpu.h>
23 #include <linux/ptrace.h>
24 #include <linux/vmstat.h>
25 #include <linux/vmalloc.h>
26 #include <linux/hardirq.h>
27 #include <linux/rculist.h>
28 #include <linux/uaccess.h>
29 #include <linux/syscalls.h>
30 #include <linux/anon_inodes.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/perf_event.h>
33 #include <linux/ftrace_event.h>
34 #include <linux/hw_breakpoint.h>
36 #include <asm/irq_regs.h>
38 atomic_t perf_task_events __read_mostly;
39 static atomic_t nr_mmap_events __read_mostly;
40 static atomic_t nr_comm_events __read_mostly;
41 static atomic_t nr_task_events __read_mostly;
43 static LIST_HEAD(pmus);
44 static DEFINE_MUTEX(pmus_lock);
45 static struct srcu_struct pmus_srcu;
48 * perf event paranoia level:
49 * -1 - not paranoid at all
50 * 0 - disallow raw tracepoint access for unpriv
51 * 1 - disallow cpu events for unpriv
52 * 2 - disallow kernel profiling for unpriv
54 int sysctl_perf_event_paranoid __read_mostly = 1;
56 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
59 * max perf event sample rate
61 int sysctl_perf_event_sample_rate __read_mostly = 100000;
63 static atomic64_t perf_event_id;
65 void __weak perf_event_print_debug(void) { }
67 extern __weak const char *perf_pmu_name(void)
72 void perf_pmu_disable(struct pmu *pmu)
74 int *count = this_cpu_ptr(pmu->pmu_disable_count);
76 pmu->pmu_disable(pmu);
79 void perf_pmu_enable(struct pmu *pmu)
81 int *count = this_cpu_ptr(pmu->pmu_disable_count);
86 static DEFINE_PER_CPU(struct list_head, rotation_list);
89 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
90 * because they're strictly cpu affine and rotate_start is called with IRQs
91 * disabled, while rotate_context is called from IRQ context.
93 static void perf_pmu_rotate_start(struct pmu *pmu)
95 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
96 struct list_head *head = &__get_cpu_var(rotation_list);
98 WARN_ON(!irqs_disabled());
100 if (list_empty(&cpuctx->rotation_list))
101 list_add(&cpuctx->rotation_list, head);
104 static void get_ctx(struct perf_event_context *ctx)
106 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
109 static void free_ctx(struct rcu_head *head)
111 struct perf_event_context *ctx;
113 ctx = container_of(head, struct perf_event_context, rcu_head);
117 static void put_ctx(struct perf_event_context *ctx)
119 if (atomic_dec_and_test(&ctx->refcount)) {
121 put_ctx(ctx->parent_ctx);
123 put_task_struct(ctx->task);
124 call_rcu(&ctx->rcu_head, free_ctx);
128 static void unclone_ctx(struct perf_event_context *ctx)
130 if (ctx->parent_ctx) {
131 put_ctx(ctx->parent_ctx);
132 ctx->parent_ctx = NULL;
137 * If we inherit events we want to return the parent event id
140 static u64 primary_event_id(struct perf_event *event)
145 id = event->parent->id;
151 * Get the perf_event_context for a task and lock it.
152 * This has to cope with with the fact that until it is locked,
153 * the context could get moved to another task.
155 static struct perf_event_context *
156 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
158 struct perf_event_context *ctx;
162 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
165 * If this context is a clone of another, it might
166 * get swapped for another underneath us by
167 * perf_event_task_sched_out, though the
168 * rcu_read_lock() protects us from any context
169 * getting freed. Lock the context and check if it
170 * got swapped before we could get the lock, and retry
171 * if so. If we locked the right context, then it
172 * can't get swapped on us any more.
174 raw_spin_lock_irqsave(&ctx->lock, *flags);
175 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
176 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
180 if (!atomic_inc_not_zero(&ctx->refcount)) {
181 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
190 * Get the context for a task and increment its pin_count so it
191 * can't get swapped to another task. This also increments its
192 * reference count so that the context can't get freed.
194 static struct perf_event_context *
195 perf_pin_task_context(struct task_struct *task, int ctxn)
197 struct perf_event_context *ctx;
200 ctx = perf_lock_task_context(task, ctxn, &flags);
203 raw_spin_unlock_irqrestore(&ctx->lock, flags);
208 static void perf_unpin_context(struct perf_event_context *ctx)
212 raw_spin_lock_irqsave(&ctx->lock, flags);
214 raw_spin_unlock_irqrestore(&ctx->lock, flags);
218 static inline u64 perf_clock(void)
220 return local_clock();
224 * Update the record of the current time in a context.
226 static void update_context_time(struct perf_event_context *ctx)
228 u64 now = perf_clock();
230 ctx->time += now - ctx->timestamp;
231 ctx->timestamp = now;
235 * Update the total_time_enabled and total_time_running fields for a event.
237 static void update_event_times(struct perf_event *event)
239 struct perf_event_context *ctx = event->ctx;
242 if (event->state < PERF_EVENT_STATE_INACTIVE ||
243 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
249 run_end = event->tstamp_stopped;
251 event->total_time_enabled = run_end - event->tstamp_enabled;
253 if (event->state == PERF_EVENT_STATE_INACTIVE)
254 run_end = event->tstamp_stopped;
258 event->total_time_running = run_end - event->tstamp_running;
262 * Update total_time_enabled and total_time_running for all events in a group.
264 static void update_group_times(struct perf_event *leader)
266 struct perf_event *event;
268 update_event_times(leader);
269 list_for_each_entry(event, &leader->sibling_list, group_entry)
270 update_event_times(event);
273 static struct list_head *
274 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
276 if (event->attr.pinned)
277 return &ctx->pinned_groups;
279 return &ctx->flexible_groups;
283 * Add a event from the lists for its context.
284 * Must be called with ctx->mutex and ctx->lock held.
287 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
289 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
290 event->attach_state |= PERF_ATTACH_CONTEXT;
293 * If we're a stand alone event or group leader, we go to the context
294 * list, group events are kept attached to the group so that
295 * perf_group_detach can, at all times, locate all siblings.
297 if (event->group_leader == event) {
298 struct list_head *list;
300 if (is_software_event(event))
301 event->group_flags |= PERF_GROUP_SOFTWARE;
303 list = ctx_group_list(event, ctx);
304 list_add_tail(&event->group_entry, list);
307 list_add_rcu(&event->event_entry, &ctx->event_list);
309 perf_pmu_rotate_start(ctx->pmu);
311 if (event->attr.inherit_stat)
316 * Called at perf_event creation and when events are attached/detached from a
319 static void perf_event__read_size(struct perf_event *event)
321 int entry = sizeof(u64); /* value */
325 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
328 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
331 if (event->attr.read_format & PERF_FORMAT_ID)
332 entry += sizeof(u64);
334 if (event->attr.read_format & PERF_FORMAT_GROUP) {
335 nr += event->group_leader->nr_siblings;
340 event->read_size = size;
343 static void perf_event__header_size(struct perf_event *event)
345 struct perf_sample_data *data;
346 u64 sample_type = event->attr.sample_type;
349 perf_event__read_size(event);
351 if (sample_type & PERF_SAMPLE_IP)
352 size += sizeof(data->ip);
354 if (sample_type & PERF_SAMPLE_TID)
355 size += sizeof(data->tid_entry);
357 if (sample_type & PERF_SAMPLE_TIME)
358 size += sizeof(data->time);
360 if (sample_type & PERF_SAMPLE_ADDR)
361 size += sizeof(data->addr);
363 if (sample_type & PERF_SAMPLE_ID)
364 size += sizeof(data->id);
366 if (sample_type & PERF_SAMPLE_STREAM_ID)
367 size += sizeof(data->stream_id);
369 if (sample_type & PERF_SAMPLE_CPU)
370 size += sizeof(data->cpu_entry);
372 if (sample_type & PERF_SAMPLE_PERIOD)
373 size += sizeof(data->period);
375 if (sample_type & PERF_SAMPLE_READ)
376 size += event->read_size;
378 event->header_size = size;
381 static void perf_group_attach(struct perf_event *event)
383 struct perf_event *group_leader = event->group_leader, *pos;
386 * We can have double attach due to group movement in perf_event_open.
388 if (event->attach_state & PERF_ATTACH_GROUP)
391 event->attach_state |= PERF_ATTACH_GROUP;
393 if (group_leader == event)
396 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
397 !is_software_event(event))
398 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
400 list_add_tail(&event->group_entry, &group_leader->sibling_list);
401 group_leader->nr_siblings++;
403 perf_event__header_size(group_leader);
405 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
406 perf_event__header_size(pos);
410 * Remove a event from the lists for its context.
411 * Must be called with ctx->mutex and ctx->lock held.
414 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
417 * We can have double detach due to exit/hot-unplug + close.
419 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
422 event->attach_state &= ~PERF_ATTACH_CONTEXT;
425 if (event->attr.inherit_stat)
428 list_del_rcu(&event->event_entry);
430 if (event->group_leader == event)
431 list_del_init(&event->group_entry);
433 update_group_times(event);
436 * If event was in error state, then keep it
437 * that way, otherwise bogus counts will be
438 * returned on read(). The only way to get out
439 * of error state is by explicit re-enabling
442 if (event->state > PERF_EVENT_STATE_OFF)
443 event->state = PERF_EVENT_STATE_OFF;
446 static void perf_group_detach(struct perf_event *event)
448 struct perf_event *sibling, *tmp;
449 struct list_head *list = NULL;
452 * We can have double detach due to exit/hot-unplug + close.
454 if (!(event->attach_state & PERF_ATTACH_GROUP))
457 event->attach_state &= ~PERF_ATTACH_GROUP;
460 * If this is a sibling, remove it from its group.
462 if (event->group_leader != event) {
463 list_del_init(&event->group_entry);
464 event->group_leader->nr_siblings--;
468 if (!list_empty(&event->group_entry))
469 list = &event->group_entry;
472 * If this was a group event with sibling events then
473 * upgrade the siblings to singleton events by adding them
474 * to whatever list we are on.
476 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
478 list_move_tail(&sibling->group_entry, list);
479 sibling->group_leader = sibling;
481 /* Inherit group flags from the previous leader */
482 sibling->group_flags = event->group_flags;
486 perf_event__header_size(event->group_leader);
488 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
489 perf_event__header_size(tmp);
493 event_filter_match(struct perf_event *event)
495 return event->cpu == -1 || event->cpu == smp_processor_id();
499 event_sched_out(struct perf_event *event,
500 struct perf_cpu_context *cpuctx,
501 struct perf_event_context *ctx)
505 * An event which could not be activated because of
506 * filter mismatch still needs to have its timings
507 * maintained, otherwise bogus information is return
508 * via read() for time_enabled, time_running:
510 if (event->state == PERF_EVENT_STATE_INACTIVE
511 && !event_filter_match(event)) {
512 delta = ctx->time - event->tstamp_stopped;
513 event->tstamp_running += delta;
514 event->tstamp_stopped = ctx->time;
517 if (event->state != PERF_EVENT_STATE_ACTIVE)
520 event->state = PERF_EVENT_STATE_INACTIVE;
521 if (event->pending_disable) {
522 event->pending_disable = 0;
523 event->state = PERF_EVENT_STATE_OFF;
525 event->tstamp_stopped = ctx->time;
526 event->pmu->del(event, 0);
529 if (!is_software_event(event))
530 cpuctx->active_oncpu--;
532 if (event->attr.exclusive || !cpuctx->active_oncpu)
533 cpuctx->exclusive = 0;
537 group_sched_out(struct perf_event *group_event,
538 struct perf_cpu_context *cpuctx,
539 struct perf_event_context *ctx)
541 struct perf_event *event;
542 int state = group_event->state;
544 event_sched_out(group_event, cpuctx, ctx);
547 * Schedule out siblings (if any):
549 list_for_each_entry(event, &group_event->sibling_list, group_entry)
550 event_sched_out(event, cpuctx, ctx);
552 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
553 cpuctx->exclusive = 0;
556 static inline struct perf_cpu_context *
557 __get_cpu_context(struct perf_event_context *ctx)
559 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
563 * Cross CPU call to remove a performance event
565 * We disable the event on the hardware level first. After that we
566 * remove it from the context list.
568 static void __perf_event_remove_from_context(void *info)
570 struct perf_event *event = info;
571 struct perf_event_context *ctx = event->ctx;
572 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
575 * If this is a task context, we need to check whether it is
576 * the current task context of this cpu. If not it has been
577 * scheduled out before the smp call arrived.
579 if (ctx->task && cpuctx->task_ctx != ctx)
582 raw_spin_lock(&ctx->lock);
584 event_sched_out(event, cpuctx, ctx);
586 list_del_event(event, ctx);
588 raw_spin_unlock(&ctx->lock);
593 * Remove the event from a task's (or a CPU's) list of events.
595 * Must be called with ctx->mutex held.
597 * CPU events are removed with a smp call. For task events we only
598 * call when the task is on a CPU.
600 * If event->ctx is a cloned context, callers must make sure that
601 * every task struct that event->ctx->task could possibly point to
602 * remains valid. This is OK when called from perf_release since
603 * that only calls us on the top-level context, which can't be a clone.
604 * When called from perf_event_exit_task, it's OK because the
605 * context has been detached from its task.
607 static void perf_event_remove_from_context(struct perf_event *event)
609 struct perf_event_context *ctx = event->ctx;
610 struct task_struct *task = ctx->task;
614 * Per cpu events are removed via an smp call and
615 * the removal is always successful.
617 smp_call_function_single(event->cpu,
618 __perf_event_remove_from_context,
624 task_oncpu_function_call(task, __perf_event_remove_from_context,
627 raw_spin_lock_irq(&ctx->lock);
629 * If the context is active we need to retry the smp call.
631 if (ctx->nr_active && !list_empty(&event->group_entry)) {
632 raw_spin_unlock_irq(&ctx->lock);
637 * The lock prevents that this context is scheduled in so we
638 * can remove the event safely, if the call above did not
641 if (!list_empty(&event->group_entry))
642 list_del_event(event, ctx);
643 raw_spin_unlock_irq(&ctx->lock);
647 * Cross CPU call to disable a performance event
649 static void __perf_event_disable(void *info)
651 struct perf_event *event = info;
652 struct perf_event_context *ctx = event->ctx;
653 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
656 * If this is a per-task event, need to check whether this
657 * event's task is the current task on this cpu.
659 if (ctx->task && cpuctx->task_ctx != ctx)
662 raw_spin_lock(&ctx->lock);
665 * If the event is on, turn it off.
666 * If it is in error state, leave it in error state.
668 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
669 update_context_time(ctx);
670 update_group_times(event);
671 if (event == event->group_leader)
672 group_sched_out(event, cpuctx, ctx);
674 event_sched_out(event, cpuctx, ctx);
675 event->state = PERF_EVENT_STATE_OFF;
678 raw_spin_unlock(&ctx->lock);
684 * If event->ctx is a cloned context, callers must make sure that
685 * every task struct that event->ctx->task could possibly point to
686 * remains valid. This condition is satisifed when called through
687 * perf_event_for_each_child or perf_event_for_each because they
688 * hold the top-level event's child_mutex, so any descendant that
689 * goes to exit will block in sync_child_event.
690 * When called from perf_pending_event it's OK because event->ctx
691 * is the current context on this CPU and preemption is disabled,
692 * hence we can't get into perf_event_task_sched_out for this context.
694 void perf_event_disable(struct perf_event *event)
696 struct perf_event_context *ctx = event->ctx;
697 struct task_struct *task = ctx->task;
701 * Disable the event on the cpu that it's on
703 smp_call_function_single(event->cpu, __perf_event_disable,
709 task_oncpu_function_call(task, __perf_event_disable, event);
711 raw_spin_lock_irq(&ctx->lock);
713 * If the event is still active, we need to retry the cross-call.
715 if (event->state == PERF_EVENT_STATE_ACTIVE) {
716 raw_spin_unlock_irq(&ctx->lock);
721 * Since we have the lock this context can't be scheduled
722 * in, so we can change the state safely.
724 if (event->state == PERF_EVENT_STATE_INACTIVE) {
725 update_group_times(event);
726 event->state = PERF_EVENT_STATE_OFF;
729 raw_spin_unlock_irq(&ctx->lock);
733 event_sched_in(struct perf_event *event,
734 struct perf_cpu_context *cpuctx,
735 struct perf_event_context *ctx)
737 if (event->state <= PERF_EVENT_STATE_OFF)
740 event->state = PERF_EVENT_STATE_ACTIVE;
741 event->oncpu = smp_processor_id();
743 * The new state must be visible before we turn it on in the hardware:
747 if (event->pmu->add(event, PERF_EF_START)) {
748 event->state = PERF_EVENT_STATE_INACTIVE;
753 event->tstamp_running += ctx->time - event->tstamp_stopped;
755 event->shadow_ctx_time = ctx->time - ctx->timestamp;
757 if (!is_software_event(event))
758 cpuctx->active_oncpu++;
761 if (event->attr.exclusive)
762 cpuctx->exclusive = 1;
768 group_sched_in(struct perf_event *group_event,
769 struct perf_cpu_context *cpuctx,
770 struct perf_event_context *ctx)
772 struct perf_event *event, *partial_group = NULL;
773 struct pmu *pmu = group_event->pmu;
775 bool simulate = false;
777 if (group_event->state == PERF_EVENT_STATE_OFF)
782 if (event_sched_in(group_event, cpuctx, ctx)) {
783 pmu->cancel_txn(pmu);
788 * Schedule in siblings as one group (if any):
790 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
791 if (event_sched_in(event, cpuctx, ctx)) {
792 partial_group = event;
797 if (!pmu->commit_txn(pmu))
802 * Groups can be scheduled in as one unit only, so undo any
803 * partial group before returning:
804 * The events up to the failed event are scheduled out normally,
805 * tstamp_stopped will be updated.
807 * The failed events and the remaining siblings need to have
808 * their timings updated as if they had gone thru event_sched_in()
809 * and event_sched_out(). This is required to get consistent timings
810 * across the group. This also takes care of the case where the group
811 * could never be scheduled by ensuring tstamp_stopped is set to mark
812 * the time the event was actually stopped, such that time delta
813 * calculation in update_event_times() is correct.
815 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
816 if (event == partial_group)
820 event->tstamp_running += now - event->tstamp_stopped;
821 event->tstamp_stopped = now;
823 event_sched_out(event, cpuctx, ctx);
826 event_sched_out(group_event, cpuctx, ctx);
828 pmu->cancel_txn(pmu);
834 * Work out whether we can put this event group on the CPU now.
836 static int group_can_go_on(struct perf_event *event,
837 struct perf_cpu_context *cpuctx,
841 * Groups consisting entirely of software events can always go on.
843 if (event->group_flags & PERF_GROUP_SOFTWARE)
846 * If an exclusive group is already on, no other hardware
849 if (cpuctx->exclusive)
852 * If this group is exclusive and there are already
853 * events on the CPU, it can't go on.
855 if (event->attr.exclusive && cpuctx->active_oncpu)
858 * Otherwise, try to add it if all previous groups were able
864 static void add_event_to_ctx(struct perf_event *event,
865 struct perf_event_context *ctx)
867 list_add_event(event, ctx);
868 perf_group_attach(event);
869 event->tstamp_enabled = ctx->time;
870 event->tstamp_running = ctx->time;
871 event->tstamp_stopped = ctx->time;
875 * Cross CPU call to install and enable a performance event
877 * Must be called with ctx->mutex held
879 static void __perf_install_in_context(void *info)
881 struct perf_event *event = info;
882 struct perf_event_context *ctx = event->ctx;
883 struct perf_event *leader = event->group_leader;
884 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
888 * If this is a task context, we need to check whether it is
889 * the current task context of this cpu. If not it has been
890 * scheduled out before the smp call arrived.
891 * Or possibly this is the right context but it isn't
892 * on this cpu because it had no events.
894 if (ctx->task && cpuctx->task_ctx != ctx) {
895 if (cpuctx->task_ctx || ctx->task != current)
897 cpuctx->task_ctx = ctx;
900 raw_spin_lock(&ctx->lock);
902 update_context_time(ctx);
904 add_event_to_ctx(event, ctx);
906 if (event->cpu != -1 && event->cpu != smp_processor_id())
910 * Don't put the event on if it is disabled or if
911 * it is in a group and the group isn't on.
913 if (event->state != PERF_EVENT_STATE_INACTIVE ||
914 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
918 * An exclusive event can't go on if there are already active
919 * hardware events, and no hardware event can go on if there
920 * is already an exclusive event on.
922 if (!group_can_go_on(event, cpuctx, 1))
925 err = event_sched_in(event, cpuctx, ctx);
929 * This event couldn't go on. If it is in a group
930 * then we have to pull the whole group off.
931 * If the event group is pinned then put it in error state.
934 group_sched_out(leader, cpuctx, ctx);
935 if (leader->attr.pinned) {
936 update_group_times(leader);
937 leader->state = PERF_EVENT_STATE_ERROR;
942 raw_spin_unlock(&ctx->lock);
946 * Attach a performance event to a context
948 * First we add the event to the list with the hardware enable bit
949 * in event->hw_config cleared.
951 * If the event is attached to a task which is on a CPU we use a smp
952 * call to enable it in the task context. The task might have been
953 * scheduled away, but we check this in the smp call again.
955 * Must be called with ctx->mutex held.
958 perf_install_in_context(struct perf_event_context *ctx,
959 struct perf_event *event,
962 struct task_struct *task = ctx->task;
968 * Per cpu events are installed via an smp call and
969 * the install is always successful.
971 smp_call_function_single(cpu, __perf_install_in_context,
977 task_oncpu_function_call(task, __perf_install_in_context,
980 raw_spin_lock_irq(&ctx->lock);
982 * we need to retry the smp call.
984 if (ctx->is_active && list_empty(&event->group_entry)) {
985 raw_spin_unlock_irq(&ctx->lock);
990 * The lock prevents that this context is scheduled in so we
991 * can add the event safely, if it the call above did not
994 if (list_empty(&event->group_entry))
995 add_event_to_ctx(event, ctx);
996 raw_spin_unlock_irq(&ctx->lock);
1000 * Put a event into inactive state and update time fields.
1001 * Enabling the leader of a group effectively enables all
1002 * the group members that aren't explicitly disabled, so we
1003 * have to update their ->tstamp_enabled also.
1004 * Note: this works for group members as well as group leaders
1005 * since the non-leader members' sibling_lists will be empty.
1007 static void __perf_event_mark_enabled(struct perf_event *event,
1008 struct perf_event_context *ctx)
1010 struct perf_event *sub;
1012 event->state = PERF_EVENT_STATE_INACTIVE;
1013 event->tstamp_enabled = ctx->time - event->total_time_enabled;
1014 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1015 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
1016 sub->tstamp_enabled =
1017 ctx->time - sub->total_time_enabled;
1023 * Cross CPU call to enable a performance event
1025 static void __perf_event_enable(void *info)
1027 struct perf_event *event = info;
1028 struct perf_event_context *ctx = event->ctx;
1029 struct perf_event *leader = event->group_leader;
1030 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1034 * If this is a per-task event, need to check whether this
1035 * event's task is the current task on this cpu.
1037 if (ctx->task && cpuctx->task_ctx != ctx) {
1038 if (cpuctx->task_ctx || ctx->task != current)
1040 cpuctx->task_ctx = ctx;
1043 raw_spin_lock(&ctx->lock);
1045 update_context_time(ctx);
1047 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1049 __perf_event_mark_enabled(event, ctx);
1051 if (event->cpu != -1 && event->cpu != smp_processor_id())
1055 * If the event is in a group and isn't the group leader,
1056 * then don't put it on unless the group is on.
1058 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1061 if (!group_can_go_on(event, cpuctx, 1)) {
1064 if (event == leader)
1065 err = group_sched_in(event, cpuctx, ctx);
1067 err = event_sched_in(event, cpuctx, ctx);
1072 * If this event can't go on and it's part of a
1073 * group, then the whole group has to come off.
1075 if (leader != event)
1076 group_sched_out(leader, cpuctx, ctx);
1077 if (leader->attr.pinned) {
1078 update_group_times(leader);
1079 leader->state = PERF_EVENT_STATE_ERROR;
1084 raw_spin_unlock(&ctx->lock);
1090 * If event->ctx is a cloned context, callers must make sure that
1091 * every task struct that event->ctx->task could possibly point to
1092 * remains valid. This condition is satisfied when called through
1093 * perf_event_for_each_child or perf_event_for_each as described
1094 * for perf_event_disable.
1096 void perf_event_enable(struct perf_event *event)
1098 struct perf_event_context *ctx = event->ctx;
1099 struct task_struct *task = ctx->task;
1103 * Enable the event on the cpu that it's on
1105 smp_call_function_single(event->cpu, __perf_event_enable,
1110 raw_spin_lock_irq(&ctx->lock);
1111 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1115 * If the event is in error state, clear that first.
1116 * That way, if we see the event in error state below, we
1117 * know that it has gone back into error state, as distinct
1118 * from the task having been scheduled away before the
1119 * cross-call arrived.
1121 if (event->state == PERF_EVENT_STATE_ERROR)
1122 event->state = PERF_EVENT_STATE_OFF;
1125 raw_spin_unlock_irq(&ctx->lock);
1126 task_oncpu_function_call(task, __perf_event_enable, event);
1128 raw_spin_lock_irq(&ctx->lock);
1131 * If the context is active and the event is still off,
1132 * we need to retry the cross-call.
1134 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1138 * Since we have the lock this context can't be scheduled
1139 * in, so we can change the state safely.
1141 if (event->state == PERF_EVENT_STATE_OFF)
1142 __perf_event_mark_enabled(event, ctx);
1145 raw_spin_unlock_irq(&ctx->lock);
1148 static int perf_event_refresh(struct perf_event *event, int refresh)
1151 * not supported on inherited events
1153 if (event->attr.inherit || !is_sampling_event(event))
1156 atomic_add(refresh, &event->event_limit);
1157 perf_event_enable(event);
1163 EVENT_FLEXIBLE = 0x1,
1165 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1168 static void ctx_sched_out(struct perf_event_context *ctx,
1169 struct perf_cpu_context *cpuctx,
1170 enum event_type_t event_type)
1172 struct perf_event *event;
1174 raw_spin_lock(&ctx->lock);
1175 perf_pmu_disable(ctx->pmu);
1177 if (likely(!ctx->nr_events))
1179 update_context_time(ctx);
1181 if (!ctx->nr_active)
1184 if (event_type & EVENT_PINNED) {
1185 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1186 group_sched_out(event, cpuctx, ctx);
1189 if (event_type & EVENT_FLEXIBLE) {
1190 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1191 group_sched_out(event, cpuctx, ctx);
1194 perf_pmu_enable(ctx->pmu);
1195 raw_spin_unlock(&ctx->lock);
1199 * Test whether two contexts are equivalent, i.e. whether they
1200 * have both been cloned from the same version of the same context
1201 * and they both have the same number of enabled events.
1202 * If the number of enabled events is the same, then the set
1203 * of enabled events should be the same, because these are both
1204 * inherited contexts, therefore we can't access individual events
1205 * in them directly with an fd; we can only enable/disable all
1206 * events via prctl, or enable/disable all events in a family
1207 * via ioctl, which will have the same effect on both contexts.
1209 static int context_equiv(struct perf_event_context *ctx1,
1210 struct perf_event_context *ctx2)
1212 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1213 && ctx1->parent_gen == ctx2->parent_gen
1214 && !ctx1->pin_count && !ctx2->pin_count;
1217 static void __perf_event_sync_stat(struct perf_event *event,
1218 struct perf_event *next_event)
1222 if (!event->attr.inherit_stat)
1226 * Update the event value, we cannot use perf_event_read()
1227 * because we're in the middle of a context switch and have IRQs
1228 * disabled, which upsets smp_call_function_single(), however
1229 * we know the event must be on the current CPU, therefore we
1230 * don't need to use it.
1232 switch (event->state) {
1233 case PERF_EVENT_STATE_ACTIVE:
1234 event->pmu->read(event);
1237 case PERF_EVENT_STATE_INACTIVE:
1238 update_event_times(event);
1246 * In order to keep per-task stats reliable we need to flip the event
1247 * values when we flip the contexts.
1249 value = local64_read(&next_event->count);
1250 value = local64_xchg(&event->count, value);
1251 local64_set(&next_event->count, value);
1253 swap(event->total_time_enabled, next_event->total_time_enabled);
1254 swap(event->total_time_running, next_event->total_time_running);
1257 * Since we swizzled the values, update the user visible data too.
1259 perf_event_update_userpage(event);
1260 perf_event_update_userpage(next_event);
1263 #define list_next_entry(pos, member) \
1264 list_entry(pos->member.next, typeof(*pos), member)
1266 static void perf_event_sync_stat(struct perf_event_context *ctx,
1267 struct perf_event_context *next_ctx)
1269 struct perf_event *event, *next_event;
1274 update_context_time(ctx);
1276 event = list_first_entry(&ctx->event_list,
1277 struct perf_event, event_entry);
1279 next_event = list_first_entry(&next_ctx->event_list,
1280 struct perf_event, event_entry);
1282 while (&event->event_entry != &ctx->event_list &&
1283 &next_event->event_entry != &next_ctx->event_list) {
1285 __perf_event_sync_stat(event, next_event);
1287 event = list_next_entry(event, event_entry);
1288 next_event = list_next_entry(next_event, event_entry);
1292 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1293 struct task_struct *next)
1295 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1296 struct perf_event_context *next_ctx;
1297 struct perf_event_context *parent;
1298 struct perf_cpu_context *cpuctx;
1304 cpuctx = __get_cpu_context(ctx);
1305 if (!cpuctx->task_ctx)
1309 parent = rcu_dereference(ctx->parent_ctx);
1310 next_ctx = next->perf_event_ctxp[ctxn];
1311 if (parent && next_ctx &&
1312 rcu_dereference(next_ctx->parent_ctx) == parent) {
1314 * Looks like the two contexts are clones, so we might be
1315 * able to optimize the context switch. We lock both
1316 * contexts and check that they are clones under the
1317 * lock (including re-checking that neither has been
1318 * uncloned in the meantime). It doesn't matter which
1319 * order we take the locks because no other cpu could
1320 * be trying to lock both of these tasks.
1322 raw_spin_lock(&ctx->lock);
1323 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1324 if (context_equiv(ctx, next_ctx)) {
1326 * XXX do we need a memory barrier of sorts
1327 * wrt to rcu_dereference() of perf_event_ctxp
1329 task->perf_event_ctxp[ctxn] = next_ctx;
1330 next->perf_event_ctxp[ctxn] = ctx;
1332 next_ctx->task = task;
1335 perf_event_sync_stat(ctx, next_ctx);
1337 raw_spin_unlock(&next_ctx->lock);
1338 raw_spin_unlock(&ctx->lock);
1343 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1344 cpuctx->task_ctx = NULL;
1348 #define for_each_task_context_nr(ctxn) \
1349 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1352 * Called from scheduler to remove the events of the current task,
1353 * with interrupts disabled.
1355 * We stop each event and update the event value in event->count.
1357 * This does not protect us against NMI, but disable()
1358 * sets the disabled bit in the control field of event _before_
1359 * accessing the event control register. If a NMI hits, then it will
1360 * not restart the event.
1362 void __perf_event_task_sched_out(struct task_struct *task,
1363 struct task_struct *next)
1367 for_each_task_context_nr(ctxn)
1368 perf_event_context_sched_out(task, ctxn, next);
1371 static void task_ctx_sched_out(struct perf_event_context *ctx,
1372 enum event_type_t event_type)
1374 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1376 if (!cpuctx->task_ctx)
1379 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1382 ctx_sched_out(ctx, cpuctx, event_type);
1383 cpuctx->task_ctx = NULL;
1387 * Called with IRQs disabled
1389 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1390 enum event_type_t event_type)
1392 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1396 ctx_pinned_sched_in(struct perf_event_context *ctx,
1397 struct perf_cpu_context *cpuctx)
1399 struct perf_event *event;
1401 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1402 if (event->state <= PERF_EVENT_STATE_OFF)
1404 if (event->cpu != -1 && event->cpu != smp_processor_id())
1407 if (group_can_go_on(event, cpuctx, 1))
1408 group_sched_in(event, cpuctx, ctx);
1411 * If this pinned group hasn't been scheduled,
1412 * put it in error state.
1414 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1415 update_group_times(event);
1416 event->state = PERF_EVENT_STATE_ERROR;
1422 ctx_flexible_sched_in(struct perf_event_context *ctx,
1423 struct perf_cpu_context *cpuctx)
1425 struct perf_event *event;
1428 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1429 /* Ignore events in OFF or ERROR state */
1430 if (event->state <= PERF_EVENT_STATE_OFF)
1433 * Listen to the 'cpu' scheduling filter constraint
1436 if (event->cpu != -1 && event->cpu != smp_processor_id())
1439 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1440 if (group_sched_in(event, cpuctx, ctx))
1447 ctx_sched_in(struct perf_event_context *ctx,
1448 struct perf_cpu_context *cpuctx,
1449 enum event_type_t event_type)
1451 raw_spin_lock(&ctx->lock);
1453 if (likely(!ctx->nr_events))
1456 ctx->timestamp = perf_clock();
1459 * First go through the list and put on any pinned groups
1460 * in order to give them the best chance of going on.
1462 if (event_type & EVENT_PINNED)
1463 ctx_pinned_sched_in(ctx, cpuctx);
1465 /* Then walk through the lower prio flexible groups */
1466 if (event_type & EVENT_FLEXIBLE)
1467 ctx_flexible_sched_in(ctx, cpuctx);
1470 raw_spin_unlock(&ctx->lock);
1473 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1474 enum event_type_t event_type)
1476 struct perf_event_context *ctx = &cpuctx->ctx;
1478 ctx_sched_in(ctx, cpuctx, event_type);
1481 static void task_ctx_sched_in(struct perf_event_context *ctx,
1482 enum event_type_t event_type)
1484 struct perf_cpu_context *cpuctx;
1486 cpuctx = __get_cpu_context(ctx);
1487 if (cpuctx->task_ctx == ctx)
1490 ctx_sched_in(ctx, cpuctx, event_type);
1491 cpuctx->task_ctx = ctx;
1494 void perf_event_context_sched_in(struct perf_event_context *ctx)
1496 struct perf_cpu_context *cpuctx;
1498 cpuctx = __get_cpu_context(ctx);
1499 if (cpuctx->task_ctx == ctx)
1502 perf_pmu_disable(ctx->pmu);
1504 * We want to keep the following priority order:
1505 * cpu pinned (that don't need to move), task pinned,
1506 * cpu flexible, task flexible.
1508 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1510 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1511 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1512 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1514 cpuctx->task_ctx = ctx;
1517 * Since these rotations are per-cpu, we need to ensure the
1518 * cpu-context we got scheduled on is actually rotating.
1520 perf_pmu_rotate_start(ctx->pmu);
1521 perf_pmu_enable(ctx->pmu);
1525 * Called from scheduler to add the events of the current task
1526 * with interrupts disabled.
1528 * We restore the event value and then enable it.
1530 * This does not protect us against NMI, but enable()
1531 * sets the enabled bit in the control field of event _before_
1532 * accessing the event control register. If a NMI hits, then it will
1533 * keep the event running.
1535 void __perf_event_task_sched_in(struct task_struct *task)
1537 struct perf_event_context *ctx;
1540 for_each_task_context_nr(ctxn) {
1541 ctx = task->perf_event_ctxp[ctxn];
1545 perf_event_context_sched_in(ctx);
1549 #define MAX_INTERRUPTS (~0ULL)
1551 static void perf_log_throttle(struct perf_event *event, int enable);
1553 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1555 u64 frequency = event->attr.sample_freq;
1556 u64 sec = NSEC_PER_SEC;
1557 u64 divisor, dividend;
1559 int count_fls, nsec_fls, frequency_fls, sec_fls;
1561 count_fls = fls64(count);
1562 nsec_fls = fls64(nsec);
1563 frequency_fls = fls64(frequency);
1567 * We got @count in @nsec, with a target of sample_freq HZ
1568 * the target period becomes:
1571 * period = -------------------
1572 * @nsec * sample_freq
1577 * Reduce accuracy by one bit such that @a and @b converge
1578 * to a similar magnitude.
1580 #define REDUCE_FLS(a, b) \
1582 if (a##_fls > b##_fls) { \
1592 * Reduce accuracy until either term fits in a u64, then proceed with
1593 * the other, so that finally we can do a u64/u64 division.
1595 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1596 REDUCE_FLS(nsec, frequency);
1597 REDUCE_FLS(sec, count);
1600 if (count_fls + sec_fls > 64) {
1601 divisor = nsec * frequency;
1603 while (count_fls + sec_fls > 64) {
1604 REDUCE_FLS(count, sec);
1608 dividend = count * sec;
1610 dividend = count * sec;
1612 while (nsec_fls + frequency_fls > 64) {
1613 REDUCE_FLS(nsec, frequency);
1617 divisor = nsec * frequency;
1623 return div64_u64(dividend, divisor);
1626 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1628 struct hw_perf_event *hwc = &event->hw;
1629 s64 period, sample_period;
1632 period = perf_calculate_period(event, nsec, count);
1634 delta = (s64)(period - hwc->sample_period);
1635 delta = (delta + 7) / 8; /* low pass filter */
1637 sample_period = hwc->sample_period + delta;
1642 hwc->sample_period = sample_period;
1644 if (local64_read(&hwc->period_left) > 8*sample_period) {
1645 event->pmu->stop(event, PERF_EF_UPDATE);
1646 local64_set(&hwc->period_left, 0);
1647 event->pmu->start(event, PERF_EF_RELOAD);
1651 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1653 struct perf_event *event;
1654 struct hw_perf_event *hwc;
1655 u64 interrupts, now;
1658 raw_spin_lock(&ctx->lock);
1659 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1660 if (event->state != PERF_EVENT_STATE_ACTIVE)
1663 if (event->cpu != -1 && event->cpu != smp_processor_id())
1668 interrupts = hwc->interrupts;
1669 hwc->interrupts = 0;
1672 * unthrottle events on the tick
1674 if (interrupts == MAX_INTERRUPTS) {
1675 perf_log_throttle(event, 1);
1676 event->pmu->start(event, 0);
1679 if (!event->attr.freq || !event->attr.sample_freq)
1682 event->pmu->read(event);
1683 now = local64_read(&event->count);
1684 delta = now - hwc->freq_count_stamp;
1685 hwc->freq_count_stamp = now;
1688 perf_adjust_period(event, period, delta);
1690 raw_spin_unlock(&ctx->lock);
1694 * Round-robin a context's events:
1696 static void rotate_ctx(struct perf_event_context *ctx)
1698 raw_spin_lock(&ctx->lock);
1701 * Rotate the first entry last of non-pinned groups. Rotation might be
1702 * disabled by the inheritance code.
1704 if (!ctx->rotate_disable)
1705 list_rotate_left(&ctx->flexible_groups);
1707 raw_spin_unlock(&ctx->lock);
1711 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1712 * because they're strictly cpu affine and rotate_start is called with IRQs
1713 * disabled, while rotate_context is called from IRQ context.
1715 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1717 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1718 struct perf_event_context *ctx = NULL;
1719 int rotate = 0, remove = 1;
1721 if (cpuctx->ctx.nr_events) {
1723 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1727 ctx = cpuctx->task_ctx;
1728 if (ctx && ctx->nr_events) {
1730 if (ctx->nr_events != ctx->nr_active)
1734 perf_pmu_disable(cpuctx->ctx.pmu);
1735 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1737 perf_ctx_adjust_freq(ctx, interval);
1742 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1744 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1746 rotate_ctx(&cpuctx->ctx);
1750 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1752 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1756 list_del_init(&cpuctx->rotation_list);
1758 perf_pmu_enable(cpuctx->ctx.pmu);
1761 void perf_event_task_tick(void)
1763 struct list_head *head = &__get_cpu_var(rotation_list);
1764 struct perf_cpu_context *cpuctx, *tmp;
1766 WARN_ON(!irqs_disabled());
1768 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1769 if (cpuctx->jiffies_interval == 1 ||
1770 !(jiffies % cpuctx->jiffies_interval))
1771 perf_rotate_context(cpuctx);
1775 static int event_enable_on_exec(struct perf_event *event,
1776 struct perf_event_context *ctx)
1778 if (!event->attr.enable_on_exec)
1781 event->attr.enable_on_exec = 0;
1782 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1785 __perf_event_mark_enabled(event, ctx);
1791 * Enable all of a task's events that have been marked enable-on-exec.
1792 * This expects task == current.
1794 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1796 struct perf_event *event;
1797 unsigned long flags;
1801 local_irq_save(flags);
1802 if (!ctx || !ctx->nr_events)
1805 task_ctx_sched_out(ctx, EVENT_ALL);
1807 raw_spin_lock(&ctx->lock);
1809 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1810 ret = event_enable_on_exec(event, ctx);
1815 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1816 ret = event_enable_on_exec(event, ctx);
1822 * Unclone this context if we enabled any event.
1827 raw_spin_unlock(&ctx->lock);
1829 perf_event_context_sched_in(ctx);
1831 local_irq_restore(flags);
1835 * Cross CPU call to read the hardware event
1837 static void __perf_event_read(void *info)
1839 struct perf_event *event = info;
1840 struct perf_event_context *ctx = event->ctx;
1841 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1844 * If this is a task context, we need to check whether it is
1845 * the current task context of this cpu. If not it has been
1846 * scheduled out before the smp call arrived. In that case
1847 * event->count would have been updated to a recent sample
1848 * when the event was scheduled out.
1850 if (ctx->task && cpuctx->task_ctx != ctx)
1853 raw_spin_lock(&ctx->lock);
1854 update_context_time(ctx);
1855 update_event_times(event);
1856 raw_spin_unlock(&ctx->lock);
1858 event->pmu->read(event);
1861 static inline u64 perf_event_count(struct perf_event *event)
1863 return local64_read(&event->count) + atomic64_read(&event->child_count);
1866 static u64 perf_event_read(struct perf_event *event)
1869 * If event is enabled and currently active on a CPU, update the
1870 * value in the event structure:
1872 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1873 smp_call_function_single(event->oncpu,
1874 __perf_event_read, event, 1);
1875 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1876 struct perf_event_context *ctx = event->ctx;
1877 unsigned long flags;
1879 raw_spin_lock_irqsave(&ctx->lock, flags);
1881 * may read while context is not active
1882 * (e.g., thread is blocked), in that case
1883 * we cannot update context time
1886 update_context_time(ctx);
1887 update_event_times(event);
1888 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1891 return perf_event_count(event);
1898 struct callchain_cpus_entries {
1899 struct rcu_head rcu_head;
1900 struct perf_callchain_entry *cpu_entries[0];
1903 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1904 static atomic_t nr_callchain_events;
1905 static DEFINE_MUTEX(callchain_mutex);
1906 struct callchain_cpus_entries *callchain_cpus_entries;
1909 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1910 struct pt_regs *regs)
1914 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1915 struct pt_regs *regs)
1919 static void release_callchain_buffers_rcu(struct rcu_head *head)
1921 struct callchain_cpus_entries *entries;
1924 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1926 for_each_possible_cpu(cpu)
1927 kfree(entries->cpu_entries[cpu]);
1932 static void release_callchain_buffers(void)
1934 struct callchain_cpus_entries *entries;
1936 entries = callchain_cpus_entries;
1937 rcu_assign_pointer(callchain_cpus_entries, NULL);
1938 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1941 static int alloc_callchain_buffers(void)
1945 struct callchain_cpus_entries *entries;
1948 * We can't use the percpu allocation API for data that can be
1949 * accessed from NMI. Use a temporary manual per cpu allocation
1950 * until that gets sorted out.
1952 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1953 num_possible_cpus();
1955 entries = kzalloc(size, GFP_KERNEL);
1959 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1961 for_each_possible_cpu(cpu) {
1962 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1964 if (!entries->cpu_entries[cpu])
1968 rcu_assign_pointer(callchain_cpus_entries, entries);
1973 for_each_possible_cpu(cpu)
1974 kfree(entries->cpu_entries[cpu]);
1980 static int get_callchain_buffers(void)
1985 mutex_lock(&callchain_mutex);
1987 count = atomic_inc_return(&nr_callchain_events);
1988 if (WARN_ON_ONCE(count < 1)) {
1994 /* If the allocation failed, give up */
1995 if (!callchain_cpus_entries)
2000 err = alloc_callchain_buffers();
2002 release_callchain_buffers();
2004 mutex_unlock(&callchain_mutex);
2009 static void put_callchain_buffers(void)
2011 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2012 release_callchain_buffers();
2013 mutex_unlock(&callchain_mutex);
2017 static int get_recursion_context(int *recursion)
2025 else if (in_softirq())
2030 if (recursion[rctx])
2039 static inline void put_recursion_context(int *recursion, int rctx)
2045 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2048 struct callchain_cpus_entries *entries;
2050 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2054 entries = rcu_dereference(callchain_cpus_entries);
2058 cpu = smp_processor_id();
2060 return &entries->cpu_entries[cpu][*rctx];
2064 put_callchain_entry(int rctx)
2066 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2069 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2072 struct perf_callchain_entry *entry;
2075 entry = get_callchain_entry(&rctx);
2084 if (!user_mode(regs)) {
2085 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2086 perf_callchain_kernel(entry, regs);
2088 regs = task_pt_regs(current);
2094 perf_callchain_store(entry, PERF_CONTEXT_USER);
2095 perf_callchain_user(entry, regs);
2099 put_callchain_entry(rctx);
2105 * Initialize the perf_event context in a task_struct:
2107 static void __perf_event_init_context(struct perf_event_context *ctx)
2109 raw_spin_lock_init(&ctx->lock);
2110 mutex_init(&ctx->mutex);
2111 INIT_LIST_HEAD(&ctx->pinned_groups);
2112 INIT_LIST_HEAD(&ctx->flexible_groups);
2113 INIT_LIST_HEAD(&ctx->event_list);
2114 atomic_set(&ctx->refcount, 1);
2117 static struct perf_event_context *
2118 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2120 struct perf_event_context *ctx;
2122 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2126 __perf_event_init_context(ctx);
2129 get_task_struct(task);
2136 static struct task_struct *
2137 find_lively_task_by_vpid(pid_t vpid)
2139 struct task_struct *task;
2146 task = find_task_by_vpid(vpid);
2148 get_task_struct(task);
2152 return ERR_PTR(-ESRCH);
2155 * Can't attach events to a dying task.
2158 if (task->flags & PF_EXITING)
2161 /* Reuse ptrace permission checks for now. */
2163 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2168 put_task_struct(task);
2169 return ERR_PTR(err);
2173 static struct perf_event_context *
2174 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2176 struct perf_event_context *ctx;
2177 struct perf_cpu_context *cpuctx;
2178 unsigned long flags;
2181 if (!task && cpu != -1) {
2182 /* Must be root to operate on a CPU event: */
2183 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2184 return ERR_PTR(-EACCES);
2186 if (cpu < 0 || cpu >= nr_cpumask_bits)
2187 return ERR_PTR(-EINVAL);
2190 * We could be clever and allow to attach a event to an
2191 * offline CPU and activate it when the CPU comes up, but
2194 if (!cpu_online(cpu))
2195 return ERR_PTR(-ENODEV);
2197 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2205 ctxn = pmu->task_ctx_nr;
2210 ctx = perf_lock_task_context(task, ctxn, &flags);
2213 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2217 ctx = alloc_perf_context(pmu, task);
2224 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
2226 * We raced with some other task; use
2227 * the context they set.
2229 put_task_struct(task);
2238 return ERR_PTR(err);
2241 static void perf_event_free_filter(struct perf_event *event);
2243 static void free_event_rcu(struct rcu_head *head)
2245 struct perf_event *event;
2247 event = container_of(head, struct perf_event, rcu_head);
2249 put_pid_ns(event->ns);
2250 perf_event_free_filter(event);
2254 static void perf_buffer_put(struct perf_buffer *buffer);
2256 static void free_event(struct perf_event *event)
2258 irq_work_sync(&event->pending);
2260 if (!event->parent) {
2261 if (event->attach_state & PERF_ATTACH_TASK)
2262 jump_label_dec(&perf_task_events);
2263 if (event->attr.mmap || event->attr.mmap_data)
2264 atomic_dec(&nr_mmap_events);
2265 if (event->attr.comm)
2266 atomic_dec(&nr_comm_events);
2267 if (event->attr.task)
2268 atomic_dec(&nr_task_events);
2269 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2270 put_callchain_buffers();
2273 if (event->buffer) {
2274 perf_buffer_put(event->buffer);
2275 event->buffer = NULL;
2279 event->destroy(event);
2282 put_ctx(event->ctx);
2284 call_rcu(&event->rcu_head, free_event_rcu);
2287 int perf_event_release_kernel(struct perf_event *event)
2289 struct perf_event_context *ctx = event->ctx;
2292 * Remove from the PMU, can't get re-enabled since we got
2293 * here because the last ref went.
2295 perf_event_disable(event);
2297 WARN_ON_ONCE(ctx->parent_ctx);
2299 * There are two ways this annotation is useful:
2301 * 1) there is a lock recursion from perf_event_exit_task
2302 * see the comment there.
2304 * 2) there is a lock-inversion with mmap_sem through
2305 * perf_event_read_group(), which takes faults while
2306 * holding ctx->mutex, however this is called after
2307 * the last filedesc died, so there is no possibility
2308 * to trigger the AB-BA case.
2310 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2311 raw_spin_lock_irq(&ctx->lock);
2312 perf_group_detach(event);
2313 list_del_event(event, ctx);
2314 raw_spin_unlock_irq(&ctx->lock);
2315 mutex_unlock(&ctx->mutex);
2321 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2324 * Called when the last reference to the file is gone.
2326 static int perf_release(struct inode *inode, struct file *file)
2328 struct perf_event *event = file->private_data;
2329 struct task_struct *owner;
2331 file->private_data = NULL;
2334 owner = ACCESS_ONCE(event->owner);
2336 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2337 * !owner it means the list deletion is complete and we can indeed
2338 * free this event, otherwise we need to serialize on
2339 * owner->perf_event_mutex.
2341 smp_read_barrier_depends();
2344 * Since delayed_put_task_struct() also drops the last
2345 * task reference we can safely take a new reference
2346 * while holding the rcu_read_lock().
2348 get_task_struct(owner);
2353 mutex_lock(&owner->perf_event_mutex);
2355 * We have to re-check the event->owner field, if it is cleared
2356 * we raced with perf_event_exit_task(), acquiring the mutex
2357 * ensured they're done, and we can proceed with freeing the
2361 list_del_init(&event->owner_entry);
2362 mutex_unlock(&owner->perf_event_mutex);
2363 put_task_struct(owner);
2366 return perf_event_release_kernel(event);
2369 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2371 struct perf_event *child;
2377 mutex_lock(&event->child_mutex);
2378 total += perf_event_read(event);
2379 *enabled += event->total_time_enabled +
2380 atomic64_read(&event->child_total_time_enabled);
2381 *running += event->total_time_running +
2382 atomic64_read(&event->child_total_time_running);
2384 list_for_each_entry(child, &event->child_list, child_list) {
2385 total += perf_event_read(child);
2386 *enabled += child->total_time_enabled;
2387 *running += child->total_time_running;
2389 mutex_unlock(&event->child_mutex);
2393 EXPORT_SYMBOL_GPL(perf_event_read_value);
2395 static int perf_event_read_group(struct perf_event *event,
2396 u64 read_format, char __user *buf)
2398 struct perf_event *leader = event->group_leader, *sub;
2399 int n = 0, size = 0, ret = -EFAULT;
2400 struct perf_event_context *ctx = leader->ctx;
2402 u64 count, enabled, running;
2404 mutex_lock(&ctx->mutex);
2405 count = perf_event_read_value(leader, &enabled, &running);
2407 values[n++] = 1 + leader->nr_siblings;
2408 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2409 values[n++] = enabled;
2410 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2411 values[n++] = running;
2412 values[n++] = count;
2413 if (read_format & PERF_FORMAT_ID)
2414 values[n++] = primary_event_id(leader);
2416 size = n * sizeof(u64);
2418 if (copy_to_user(buf, values, size))
2423 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2426 values[n++] = perf_event_read_value(sub, &enabled, &running);
2427 if (read_format & PERF_FORMAT_ID)
2428 values[n++] = primary_event_id(sub);
2430 size = n * sizeof(u64);
2432 if (copy_to_user(buf + ret, values, size)) {
2440 mutex_unlock(&ctx->mutex);
2445 static int perf_event_read_one(struct perf_event *event,
2446 u64 read_format, char __user *buf)
2448 u64 enabled, running;
2452 values[n++] = perf_event_read_value(event, &enabled, &running);
2453 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2454 values[n++] = enabled;
2455 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2456 values[n++] = running;
2457 if (read_format & PERF_FORMAT_ID)
2458 values[n++] = primary_event_id(event);
2460 if (copy_to_user(buf, values, n * sizeof(u64)))
2463 return n * sizeof(u64);
2467 * Read the performance event - simple non blocking version for now
2470 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2472 u64 read_format = event->attr.read_format;
2476 * Return end-of-file for a read on a event that is in
2477 * error state (i.e. because it was pinned but it couldn't be
2478 * scheduled on to the CPU at some point).
2480 if (event->state == PERF_EVENT_STATE_ERROR)
2483 if (count < event->read_size)
2486 WARN_ON_ONCE(event->ctx->parent_ctx);
2487 if (read_format & PERF_FORMAT_GROUP)
2488 ret = perf_event_read_group(event, read_format, buf);
2490 ret = perf_event_read_one(event, read_format, buf);
2496 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2498 struct perf_event *event = file->private_data;
2500 return perf_read_hw(event, buf, count);
2503 static unsigned int perf_poll(struct file *file, poll_table *wait)
2505 struct perf_event *event = file->private_data;
2506 struct perf_buffer *buffer;
2507 unsigned int events = POLL_HUP;
2510 buffer = rcu_dereference(event->buffer);
2512 events = atomic_xchg(&buffer->poll, 0);
2515 poll_wait(file, &event->waitq, wait);
2520 static void perf_event_reset(struct perf_event *event)
2522 (void)perf_event_read(event);
2523 local64_set(&event->count, 0);
2524 perf_event_update_userpage(event);
2528 * Holding the top-level event's child_mutex means that any
2529 * descendant process that has inherited this event will block
2530 * in sync_child_event if it goes to exit, thus satisfying the
2531 * task existence requirements of perf_event_enable/disable.
2533 static void perf_event_for_each_child(struct perf_event *event,
2534 void (*func)(struct perf_event *))
2536 struct perf_event *child;
2538 WARN_ON_ONCE(event->ctx->parent_ctx);
2539 mutex_lock(&event->child_mutex);
2541 list_for_each_entry(child, &event->child_list, child_list)
2543 mutex_unlock(&event->child_mutex);
2546 static void perf_event_for_each(struct perf_event *event,
2547 void (*func)(struct perf_event *))
2549 struct perf_event_context *ctx = event->ctx;
2550 struct perf_event *sibling;
2552 WARN_ON_ONCE(ctx->parent_ctx);
2553 mutex_lock(&ctx->mutex);
2554 event = event->group_leader;
2556 perf_event_for_each_child(event, func);
2558 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2559 perf_event_for_each_child(event, func);
2560 mutex_unlock(&ctx->mutex);
2563 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2565 struct perf_event_context *ctx = event->ctx;
2569 if (!is_sampling_event(event))
2572 if (copy_from_user(&value, arg, sizeof(value)))
2578 raw_spin_lock_irq(&ctx->lock);
2579 if (event->attr.freq) {
2580 if (value > sysctl_perf_event_sample_rate) {
2585 event->attr.sample_freq = value;
2587 event->attr.sample_period = value;
2588 event->hw.sample_period = value;
2591 raw_spin_unlock_irq(&ctx->lock);
2596 static const struct file_operations perf_fops;
2598 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2602 file = fget_light(fd, fput_needed);
2604 return ERR_PTR(-EBADF);
2606 if (file->f_op != &perf_fops) {
2607 fput_light(file, *fput_needed);
2609 return ERR_PTR(-EBADF);
2612 return file->private_data;
2615 static int perf_event_set_output(struct perf_event *event,
2616 struct perf_event *output_event);
2617 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2619 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2621 struct perf_event *event = file->private_data;
2622 void (*func)(struct perf_event *);
2626 case PERF_EVENT_IOC_ENABLE:
2627 func = perf_event_enable;
2629 case PERF_EVENT_IOC_DISABLE:
2630 func = perf_event_disable;
2632 case PERF_EVENT_IOC_RESET:
2633 func = perf_event_reset;
2636 case PERF_EVENT_IOC_REFRESH:
2637 return perf_event_refresh(event, arg);
2639 case PERF_EVENT_IOC_PERIOD:
2640 return perf_event_period(event, (u64 __user *)arg);
2642 case PERF_EVENT_IOC_SET_OUTPUT:
2644 struct perf_event *output_event = NULL;
2645 int fput_needed = 0;
2649 output_event = perf_fget_light(arg, &fput_needed);
2650 if (IS_ERR(output_event))
2651 return PTR_ERR(output_event);
2654 ret = perf_event_set_output(event, output_event);
2656 fput_light(output_event->filp, fput_needed);
2661 case PERF_EVENT_IOC_SET_FILTER:
2662 return perf_event_set_filter(event, (void __user *)arg);
2668 if (flags & PERF_IOC_FLAG_GROUP)
2669 perf_event_for_each(event, func);
2671 perf_event_for_each_child(event, func);
2676 int perf_event_task_enable(void)
2678 struct perf_event *event;
2680 mutex_lock(¤t->perf_event_mutex);
2681 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2682 perf_event_for_each_child(event, perf_event_enable);
2683 mutex_unlock(¤t->perf_event_mutex);
2688 int perf_event_task_disable(void)
2690 struct perf_event *event;
2692 mutex_lock(¤t->perf_event_mutex);
2693 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2694 perf_event_for_each_child(event, perf_event_disable);
2695 mutex_unlock(¤t->perf_event_mutex);
2700 #ifndef PERF_EVENT_INDEX_OFFSET
2701 # define PERF_EVENT_INDEX_OFFSET 0
2704 static int perf_event_index(struct perf_event *event)
2706 if (event->hw.state & PERF_HES_STOPPED)
2709 if (event->state != PERF_EVENT_STATE_ACTIVE)
2712 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2716 * Callers need to ensure there can be no nesting of this function, otherwise
2717 * the seqlock logic goes bad. We can not serialize this because the arch
2718 * code calls this from NMI context.
2720 void perf_event_update_userpage(struct perf_event *event)
2722 struct perf_event_mmap_page *userpg;
2723 struct perf_buffer *buffer;
2726 buffer = rcu_dereference(event->buffer);
2730 userpg = buffer->user_page;
2733 * Disable preemption so as to not let the corresponding user-space
2734 * spin too long if we get preempted.
2739 userpg->index = perf_event_index(event);
2740 userpg->offset = perf_event_count(event);
2741 if (event->state == PERF_EVENT_STATE_ACTIVE)
2742 userpg->offset -= local64_read(&event->hw.prev_count);
2744 userpg->time_enabled = event->total_time_enabled +
2745 atomic64_read(&event->child_total_time_enabled);
2747 userpg->time_running = event->total_time_running +
2748 atomic64_read(&event->child_total_time_running);
2757 static unsigned long perf_data_size(struct perf_buffer *buffer);
2760 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2762 long max_size = perf_data_size(buffer);
2765 buffer->watermark = min(max_size, watermark);
2767 if (!buffer->watermark)
2768 buffer->watermark = max_size / 2;
2770 if (flags & PERF_BUFFER_WRITABLE)
2771 buffer->writable = 1;
2773 atomic_set(&buffer->refcount, 1);
2776 #ifndef CONFIG_PERF_USE_VMALLOC
2779 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2782 static struct page *
2783 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2785 if (pgoff > buffer->nr_pages)
2789 return virt_to_page(buffer->user_page);
2791 return virt_to_page(buffer->data_pages[pgoff - 1]);
2794 static void *perf_mmap_alloc_page(int cpu)
2799 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2800 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2804 return page_address(page);
2807 static struct perf_buffer *
2808 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2810 struct perf_buffer *buffer;
2814 size = sizeof(struct perf_buffer);
2815 size += nr_pages * sizeof(void *);
2817 buffer = kzalloc(size, GFP_KERNEL);
2821 buffer->user_page = perf_mmap_alloc_page(cpu);
2822 if (!buffer->user_page)
2823 goto fail_user_page;
2825 for (i = 0; i < nr_pages; i++) {
2826 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2827 if (!buffer->data_pages[i])
2828 goto fail_data_pages;
2831 buffer->nr_pages = nr_pages;
2833 perf_buffer_init(buffer, watermark, flags);
2838 for (i--; i >= 0; i--)
2839 free_page((unsigned long)buffer->data_pages[i]);
2841 free_page((unsigned long)buffer->user_page);
2850 static void perf_mmap_free_page(unsigned long addr)
2852 struct page *page = virt_to_page((void *)addr);
2854 page->mapping = NULL;
2858 static void perf_buffer_free(struct perf_buffer *buffer)
2862 perf_mmap_free_page((unsigned long)buffer->user_page);
2863 for (i = 0; i < buffer->nr_pages; i++)
2864 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2868 static inline int page_order(struct perf_buffer *buffer)
2876 * Back perf_mmap() with vmalloc memory.
2878 * Required for architectures that have d-cache aliasing issues.
2881 static inline int page_order(struct perf_buffer *buffer)
2883 return buffer->page_order;
2886 static struct page *
2887 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2889 if (pgoff > (1UL << page_order(buffer)))
2892 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2895 static void perf_mmap_unmark_page(void *addr)
2897 struct page *page = vmalloc_to_page(addr);
2899 page->mapping = NULL;
2902 static void perf_buffer_free_work(struct work_struct *work)
2904 struct perf_buffer *buffer;
2908 buffer = container_of(work, struct perf_buffer, work);
2909 nr = 1 << page_order(buffer);
2911 base = buffer->user_page;
2912 for (i = 0; i < nr + 1; i++)
2913 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2919 static void perf_buffer_free(struct perf_buffer *buffer)
2921 schedule_work(&buffer->work);
2924 static struct perf_buffer *
2925 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2927 struct perf_buffer *buffer;
2931 size = sizeof(struct perf_buffer);
2932 size += sizeof(void *);
2934 buffer = kzalloc(size, GFP_KERNEL);
2938 INIT_WORK(&buffer->work, perf_buffer_free_work);
2940 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2944 buffer->user_page = all_buf;
2945 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2946 buffer->page_order = ilog2(nr_pages);
2947 buffer->nr_pages = 1;
2949 perf_buffer_init(buffer, watermark, flags);
2962 static unsigned long perf_data_size(struct perf_buffer *buffer)
2964 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2967 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2969 struct perf_event *event = vma->vm_file->private_data;
2970 struct perf_buffer *buffer;
2971 int ret = VM_FAULT_SIGBUS;
2973 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2974 if (vmf->pgoff == 0)
2980 buffer = rcu_dereference(event->buffer);
2984 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2987 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2991 get_page(vmf->page);
2992 vmf->page->mapping = vma->vm_file->f_mapping;
2993 vmf->page->index = vmf->pgoff;
3002 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
3004 struct perf_buffer *buffer;
3006 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
3007 perf_buffer_free(buffer);
3010 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
3012 struct perf_buffer *buffer;
3015 buffer = rcu_dereference(event->buffer);
3017 if (!atomic_inc_not_zero(&buffer->refcount))
3025 static void perf_buffer_put(struct perf_buffer *buffer)
3027 if (!atomic_dec_and_test(&buffer->refcount))
3030 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
3033 static void perf_mmap_open(struct vm_area_struct *vma)
3035 struct perf_event *event = vma->vm_file->private_data;
3037 atomic_inc(&event->mmap_count);
3040 static void perf_mmap_close(struct vm_area_struct *vma)
3042 struct perf_event *event = vma->vm_file->private_data;
3044 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3045 unsigned long size = perf_data_size(event->buffer);
3046 struct user_struct *user = event->mmap_user;
3047 struct perf_buffer *buffer = event->buffer;
3049 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3050 vma->vm_mm->locked_vm -= event->mmap_locked;
3051 rcu_assign_pointer(event->buffer, NULL);
3052 mutex_unlock(&event->mmap_mutex);
3054 perf_buffer_put(buffer);
3059 static const struct vm_operations_struct perf_mmap_vmops = {
3060 .open = perf_mmap_open,
3061 .close = perf_mmap_close,
3062 .fault = perf_mmap_fault,
3063 .page_mkwrite = perf_mmap_fault,
3066 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3068 struct perf_event *event = file->private_data;
3069 unsigned long user_locked, user_lock_limit;
3070 struct user_struct *user = current_user();
3071 unsigned long locked, lock_limit;
3072 struct perf_buffer *buffer;
3073 unsigned long vma_size;
3074 unsigned long nr_pages;
3075 long user_extra, extra;
3076 int ret = 0, flags = 0;
3079 * Don't allow mmap() of inherited per-task counters. This would
3080 * create a performance issue due to all children writing to the
3083 if (event->cpu == -1 && event->attr.inherit)
3086 if (!(vma->vm_flags & VM_SHARED))
3089 vma_size = vma->vm_end - vma->vm_start;
3090 nr_pages = (vma_size / PAGE_SIZE) - 1;
3093 * If we have buffer pages ensure they're a power-of-two number, so we
3094 * can do bitmasks instead of modulo.
3096 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3099 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3102 if (vma->vm_pgoff != 0)
3105 WARN_ON_ONCE(event->ctx->parent_ctx);
3106 mutex_lock(&event->mmap_mutex);
3107 if (event->buffer) {
3108 if (event->buffer->nr_pages == nr_pages)
3109 atomic_inc(&event->buffer->refcount);
3115 user_extra = nr_pages + 1;
3116 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3119 * Increase the limit linearly with more CPUs:
3121 user_lock_limit *= num_online_cpus();
3123 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3126 if (user_locked > user_lock_limit)
3127 extra = user_locked - user_lock_limit;
3129 lock_limit = rlimit(RLIMIT_MEMLOCK);
3130 lock_limit >>= PAGE_SHIFT;
3131 locked = vma->vm_mm->locked_vm + extra;
3133 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3134 !capable(CAP_IPC_LOCK)) {
3139 WARN_ON(event->buffer);
3141 if (vma->vm_flags & VM_WRITE)
3142 flags |= PERF_BUFFER_WRITABLE;
3144 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3150 rcu_assign_pointer(event->buffer, buffer);
3152 atomic_long_add(user_extra, &user->locked_vm);
3153 event->mmap_locked = extra;
3154 event->mmap_user = get_current_user();
3155 vma->vm_mm->locked_vm += event->mmap_locked;
3159 atomic_inc(&event->mmap_count);
3160 mutex_unlock(&event->mmap_mutex);
3162 vma->vm_flags |= VM_RESERVED;
3163 vma->vm_ops = &perf_mmap_vmops;
3168 static int perf_fasync(int fd, struct file *filp, int on)
3170 struct inode *inode = filp->f_path.dentry->d_inode;
3171 struct perf_event *event = filp->private_data;
3174 mutex_lock(&inode->i_mutex);
3175 retval = fasync_helper(fd, filp, on, &event->fasync);
3176 mutex_unlock(&inode->i_mutex);
3184 static const struct file_operations perf_fops = {
3185 .llseek = no_llseek,
3186 .release = perf_release,
3189 .unlocked_ioctl = perf_ioctl,
3190 .compat_ioctl = perf_ioctl,
3192 .fasync = perf_fasync,
3198 * If there's data, ensure we set the poll() state and publish everything
3199 * to user-space before waking everybody up.
3202 void perf_event_wakeup(struct perf_event *event)
3204 wake_up_all(&event->waitq);
3206 if (event->pending_kill) {
3207 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3208 event->pending_kill = 0;
3212 static void perf_pending_event(struct irq_work *entry)
3214 struct perf_event *event = container_of(entry,
3215 struct perf_event, pending);
3217 if (event->pending_disable) {
3218 event->pending_disable = 0;
3219 __perf_event_disable(event);
3222 if (event->pending_wakeup) {
3223 event->pending_wakeup = 0;
3224 perf_event_wakeup(event);
3229 * We assume there is only KVM supporting the callbacks.
3230 * Later on, we might change it to a list if there is
3231 * another virtualization implementation supporting the callbacks.
3233 struct perf_guest_info_callbacks *perf_guest_cbs;
3235 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3237 perf_guest_cbs = cbs;
3240 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3242 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3244 perf_guest_cbs = NULL;
3247 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3252 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3253 unsigned long offset, unsigned long head)
3257 if (!buffer->writable)
3260 mask = perf_data_size(buffer) - 1;
3262 offset = (offset - tail) & mask;
3263 head = (head - tail) & mask;
3265 if ((int)(head - offset) < 0)
3271 static void perf_output_wakeup(struct perf_output_handle *handle)
3273 atomic_set(&handle->buffer->poll, POLL_IN);
3276 handle->event->pending_wakeup = 1;
3277 irq_work_queue(&handle->event->pending);
3279 perf_event_wakeup(handle->event);
3283 * We need to ensure a later event_id doesn't publish a head when a former
3284 * event isn't done writing. However since we need to deal with NMIs we
3285 * cannot fully serialize things.
3287 * We only publish the head (and generate a wakeup) when the outer-most
3290 static void perf_output_get_handle(struct perf_output_handle *handle)
3292 struct perf_buffer *buffer = handle->buffer;
3295 local_inc(&buffer->nest);
3296 handle->wakeup = local_read(&buffer->wakeup);
3299 static void perf_output_put_handle(struct perf_output_handle *handle)
3301 struct perf_buffer *buffer = handle->buffer;
3305 head = local_read(&buffer->head);
3308 * IRQ/NMI can happen here, which means we can miss a head update.
3311 if (!local_dec_and_test(&buffer->nest))
3315 * Publish the known good head. Rely on the full barrier implied
3316 * by atomic_dec_and_test() order the buffer->head read and this
3319 buffer->user_page->data_head = head;
3322 * Now check if we missed an update, rely on the (compiler)
3323 * barrier in atomic_dec_and_test() to re-read buffer->head.
3325 if (unlikely(head != local_read(&buffer->head))) {
3326 local_inc(&buffer->nest);
3330 if (handle->wakeup != local_read(&buffer->wakeup))
3331 perf_output_wakeup(handle);
3337 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3338 const void *buf, unsigned int len)
3341 unsigned long size = min_t(unsigned long, handle->size, len);
3343 memcpy(handle->addr, buf, size);
3346 handle->addr += size;
3348 handle->size -= size;
3349 if (!handle->size) {
3350 struct perf_buffer *buffer = handle->buffer;
3353 handle->page &= buffer->nr_pages - 1;
3354 handle->addr = buffer->data_pages[handle->page];
3355 handle->size = PAGE_SIZE << page_order(buffer);
3360 int perf_output_begin(struct perf_output_handle *handle,
3361 struct perf_event *event, unsigned int size,
3362 int nmi, int sample)
3364 struct perf_buffer *buffer;
3365 unsigned long tail, offset, head;
3368 struct perf_event_header header;
3375 * For inherited events we send all the output towards the parent.
3378 event = event->parent;
3380 buffer = rcu_dereference(event->buffer);
3384 handle->buffer = buffer;
3385 handle->event = event;
3387 handle->sample = sample;
3389 if (!buffer->nr_pages)
3392 have_lost = local_read(&buffer->lost);
3394 size += sizeof(lost_event);
3396 perf_output_get_handle(handle);
3400 * Userspace could choose to issue a mb() before updating the
3401 * tail pointer. So that all reads will be completed before the
3404 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3406 offset = head = local_read(&buffer->head);
3408 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3410 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3412 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3413 local_add(buffer->watermark, &buffer->wakeup);
3415 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3416 handle->page &= buffer->nr_pages - 1;
3417 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3418 handle->addr = buffer->data_pages[handle->page];
3419 handle->addr += handle->size;
3420 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3423 lost_event.header.type = PERF_RECORD_LOST;
3424 lost_event.header.misc = 0;
3425 lost_event.header.size = sizeof(lost_event);
3426 lost_event.id = event->id;
3427 lost_event.lost = local_xchg(&buffer->lost, 0);
3429 perf_output_put(handle, lost_event);
3435 local_inc(&buffer->lost);
3436 perf_output_put_handle(handle);
3443 void perf_output_end(struct perf_output_handle *handle)
3445 struct perf_event *event = handle->event;
3446 struct perf_buffer *buffer = handle->buffer;
3448 int wakeup_events = event->attr.wakeup_events;
3450 if (handle->sample && wakeup_events) {
3451 int events = local_inc_return(&buffer->events);
3452 if (events >= wakeup_events) {
3453 local_sub(wakeup_events, &buffer->events);
3454 local_inc(&buffer->wakeup);
3458 perf_output_put_handle(handle);
3462 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3465 * only top level events have the pid namespace they were created in
3468 event = event->parent;
3470 return task_tgid_nr_ns(p, event->ns);
3473 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3476 * only top level events have the pid namespace they were created in
3479 event = event->parent;
3481 return task_pid_nr_ns(p, event->ns);
3484 static void perf_output_read_one(struct perf_output_handle *handle,
3485 struct perf_event *event,
3486 u64 enabled, u64 running)
3488 u64 read_format = event->attr.read_format;
3492 values[n++] = perf_event_count(event);
3493 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3494 values[n++] = enabled +
3495 atomic64_read(&event->child_total_time_enabled);
3497 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3498 values[n++] = running +
3499 atomic64_read(&event->child_total_time_running);
3501 if (read_format & PERF_FORMAT_ID)
3502 values[n++] = primary_event_id(event);
3504 perf_output_copy(handle, values, n * sizeof(u64));
3508 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3510 static void perf_output_read_group(struct perf_output_handle *handle,
3511 struct perf_event *event,
3512 u64 enabled, u64 running)
3514 struct perf_event *leader = event->group_leader, *sub;
3515 u64 read_format = event->attr.read_format;
3519 values[n++] = 1 + leader->nr_siblings;
3521 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3522 values[n++] = enabled;
3524 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3525 values[n++] = running;
3527 if (leader != event)
3528 leader->pmu->read(leader);
3530 values[n++] = perf_event_count(leader);
3531 if (read_format & PERF_FORMAT_ID)
3532 values[n++] = primary_event_id(leader);
3534 perf_output_copy(handle, values, n * sizeof(u64));
3536 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3540 sub->pmu->read(sub);
3542 values[n++] = perf_event_count(sub);
3543 if (read_format & PERF_FORMAT_ID)
3544 values[n++] = primary_event_id(sub);
3546 perf_output_copy(handle, values, n * sizeof(u64));
3550 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3551 PERF_FORMAT_TOTAL_TIME_RUNNING)
3553 static void perf_output_read(struct perf_output_handle *handle,
3554 struct perf_event *event)
3556 u64 enabled = 0, running = 0, now, ctx_time;
3557 u64 read_format = event->attr.read_format;
3560 * compute total_time_enabled, total_time_running
3561 * based on snapshot values taken when the event
3562 * was last scheduled in.
3564 * we cannot simply called update_context_time()
3565 * because of locking issue as we are called in
3568 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3570 ctx_time = event->shadow_ctx_time + now;
3571 enabled = ctx_time - event->tstamp_enabled;
3572 running = ctx_time - event->tstamp_running;
3575 if (event->attr.read_format & PERF_FORMAT_GROUP)
3576 perf_output_read_group(handle, event, enabled, running);
3578 perf_output_read_one(handle, event, enabled, running);
3581 void perf_output_sample(struct perf_output_handle *handle,
3582 struct perf_event_header *header,
3583 struct perf_sample_data *data,
3584 struct perf_event *event)
3586 u64 sample_type = data->type;
3588 perf_output_put(handle, *header);
3590 if (sample_type & PERF_SAMPLE_IP)
3591 perf_output_put(handle, data->ip);
3593 if (sample_type & PERF_SAMPLE_TID)
3594 perf_output_put(handle, data->tid_entry);
3596 if (sample_type & PERF_SAMPLE_TIME)
3597 perf_output_put(handle, data->time);
3599 if (sample_type & PERF_SAMPLE_ADDR)
3600 perf_output_put(handle, data->addr);
3602 if (sample_type & PERF_SAMPLE_ID)
3603 perf_output_put(handle, data->id);
3605 if (sample_type & PERF_SAMPLE_STREAM_ID)
3606 perf_output_put(handle, data->stream_id);
3608 if (sample_type & PERF_SAMPLE_CPU)
3609 perf_output_put(handle, data->cpu_entry);
3611 if (sample_type & PERF_SAMPLE_PERIOD)
3612 perf_output_put(handle, data->period);
3614 if (sample_type & PERF_SAMPLE_READ)
3615 perf_output_read(handle, event);
3617 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3618 if (data->callchain) {
3621 if (data->callchain)
3622 size += data->callchain->nr;
3624 size *= sizeof(u64);
3626 perf_output_copy(handle, data->callchain, size);
3629 perf_output_put(handle, nr);
3633 if (sample_type & PERF_SAMPLE_RAW) {
3635 perf_output_put(handle, data->raw->size);
3636 perf_output_copy(handle, data->raw->data,
3643 .size = sizeof(u32),
3646 perf_output_put(handle, raw);
3651 void perf_prepare_sample(struct perf_event_header *header,
3652 struct perf_sample_data *data,
3653 struct perf_event *event,
3654 struct pt_regs *regs)
3656 u64 sample_type = event->attr.sample_type;
3658 data->type = sample_type;
3660 header->type = PERF_RECORD_SAMPLE;
3661 header->size = sizeof(*header) + event->header_size;
3664 header->misc |= perf_misc_flags(regs);
3666 if (sample_type & PERF_SAMPLE_IP)
3667 data->ip = perf_instruction_pointer(regs);
3669 if (sample_type & PERF_SAMPLE_TID) {
3670 /* namespace issues */
3671 data->tid_entry.pid = perf_event_pid(event, current);
3672 data->tid_entry.tid = perf_event_tid(event, current);
3675 if (sample_type & PERF_SAMPLE_TIME)
3676 data->time = perf_clock();
3678 if (sample_type & PERF_SAMPLE_ID)
3679 data->id = primary_event_id(event);
3681 if (sample_type & PERF_SAMPLE_STREAM_ID)
3682 data->stream_id = event->id;
3684 if (sample_type & PERF_SAMPLE_CPU) {
3685 data->cpu_entry.cpu = raw_smp_processor_id();
3686 data->cpu_entry.reserved = 0;
3689 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3692 data->callchain = perf_callchain(regs);
3694 if (data->callchain)
3695 size += data->callchain->nr;
3697 header->size += size * sizeof(u64);
3700 if (sample_type & PERF_SAMPLE_RAW) {
3701 int size = sizeof(u32);
3704 size += data->raw->size;
3706 size += sizeof(u32);
3708 WARN_ON_ONCE(size & (sizeof(u64)-1));
3709 header->size += size;
3713 static void perf_event_output(struct perf_event *event, int nmi,
3714 struct perf_sample_data *data,
3715 struct pt_regs *regs)
3717 struct perf_output_handle handle;
3718 struct perf_event_header header;
3720 /* protect the callchain buffers */
3723 perf_prepare_sample(&header, data, event, regs);
3725 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3728 perf_output_sample(&handle, &header, data, event);
3730 perf_output_end(&handle);
3740 struct perf_read_event {
3741 struct perf_event_header header;
3748 perf_event_read_event(struct perf_event *event,
3749 struct task_struct *task)
3751 struct perf_output_handle handle;
3752 struct perf_read_event read_event = {
3754 .type = PERF_RECORD_READ,
3756 .size = sizeof(read_event) + event->read_size,
3758 .pid = perf_event_pid(event, task),
3759 .tid = perf_event_tid(event, task),
3763 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3767 perf_output_put(&handle, read_event);
3768 perf_output_read(&handle, event);
3770 perf_output_end(&handle);
3774 * task tracking -- fork/exit
3776 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3779 struct perf_task_event {
3780 struct task_struct *task;
3781 struct perf_event_context *task_ctx;
3784 struct perf_event_header header;
3794 static void perf_event_task_output(struct perf_event *event,
3795 struct perf_task_event *task_event)
3797 struct perf_output_handle handle;
3798 struct task_struct *task = task_event->task;
3801 size = task_event->event_id.header.size;
3802 ret = perf_output_begin(&handle, event, size, 0, 0);
3807 task_event->event_id.pid = perf_event_pid(event, task);
3808 task_event->event_id.ppid = perf_event_pid(event, current);
3810 task_event->event_id.tid = perf_event_tid(event, task);
3811 task_event->event_id.ptid = perf_event_tid(event, current);
3813 perf_output_put(&handle, task_event->event_id);
3815 perf_output_end(&handle);
3818 static int perf_event_task_match(struct perf_event *event)
3820 if (event->state < PERF_EVENT_STATE_INACTIVE)
3823 if (event->cpu != -1 && event->cpu != smp_processor_id())
3826 if (event->attr.comm || event->attr.mmap ||
3827 event->attr.mmap_data || event->attr.task)
3833 static void perf_event_task_ctx(struct perf_event_context *ctx,
3834 struct perf_task_event *task_event)
3836 struct perf_event *event;
3838 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3839 if (perf_event_task_match(event))
3840 perf_event_task_output(event, task_event);
3844 static void perf_event_task_event(struct perf_task_event *task_event)
3846 struct perf_cpu_context *cpuctx;
3847 struct perf_event_context *ctx;
3852 list_for_each_entry_rcu(pmu, &pmus, entry) {
3853 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3854 perf_event_task_ctx(&cpuctx->ctx, task_event);
3856 ctx = task_event->task_ctx;
3858 ctxn = pmu->task_ctx_nr;
3861 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3864 perf_event_task_ctx(ctx, task_event);
3866 put_cpu_ptr(pmu->pmu_cpu_context);
3871 static void perf_event_task(struct task_struct *task,
3872 struct perf_event_context *task_ctx,
3875 struct perf_task_event task_event;
3877 if (!atomic_read(&nr_comm_events) &&
3878 !atomic_read(&nr_mmap_events) &&
3879 !atomic_read(&nr_task_events))
3882 task_event = (struct perf_task_event){
3884 .task_ctx = task_ctx,
3887 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3889 .size = sizeof(task_event.event_id),
3895 .time = perf_clock(),
3899 perf_event_task_event(&task_event);
3902 void perf_event_fork(struct task_struct *task)
3904 perf_event_task(task, NULL, 1);
3911 struct perf_comm_event {
3912 struct task_struct *task;
3917 struct perf_event_header header;
3924 static void perf_event_comm_output(struct perf_event *event,
3925 struct perf_comm_event *comm_event)
3927 struct perf_output_handle handle;
3928 int size = comm_event->event_id.header.size;
3929 int ret = perf_output_begin(&handle, event, size, 0, 0);
3934 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3935 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3937 perf_output_put(&handle, comm_event->event_id);
3938 perf_output_copy(&handle, comm_event->comm,
3939 comm_event->comm_size);
3940 perf_output_end(&handle);
3943 static int perf_event_comm_match(struct perf_event *event)
3945 if (event->state < PERF_EVENT_STATE_INACTIVE)
3948 if (event->cpu != -1 && event->cpu != smp_processor_id())
3951 if (event->attr.comm)
3957 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3958 struct perf_comm_event *comm_event)
3960 struct perf_event *event;
3962 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3963 if (perf_event_comm_match(event))
3964 perf_event_comm_output(event, comm_event);
3968 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3970 struct perf_cpu_context *cpuctx;
3971 struct perf_event_context *ctx;
3972 char comm[TASK_COMM_LEN];
3977 memset(comm, 0, sizeof(comm));
3978 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3979 size = ALIGN(strlen(comm)+1, sizeof(u64));
3981 comm_event->comm = comm;
3982 comm_event->comm_size = size;
3984 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3987 list_for_each_entry_rcu(pmu, &pmus, entry) {
3988 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3989 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3991 ctxn = pmu->task_ctx_nr;
3995 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3997 perf_event_comm_ctx(ctx, comm_event);
3999 put_cpu_ptr(pmu->pmu_cpu_context);
4004 void perf_event_comm(struct task_struct *task)
4006 struct perf_comm_event comm_event;
4007 struct perf_event_context *ctx;
4010 for_each_task_context_nr(ctxn) {
4011 ctx = task->perf_event_ctxp[ctxn];
4015 perf_event_enable_on_exec(ctx);
4018 if (!atomic_read(&nr_comm_events))
4021 comm_event = (struct perf_comm_event){
4027 .type = PERF_RECORD_COMM,
4036 perf_event_comm_event(&comm_event);
4043 struct perf_mmap_event {
4044 struct vm_area_struct *vma;
4046 const char *file_name;
4050 struct perf_event_header header;
4060 static void perf_event_mmap_output(struct perf_event *event,
4061 struct perf_mmap_event *mmap_event)
4063 struct perf_output_handle handle;
4064 int size = mmap_event->event_id.header.size;
4065 int ret = perf_output_begin(&handle, event, size, 0, 0);
4070 mmap_event->event_id.pid = perf_event_pid(event, current);
4071 mmap_event->event_id.tid = perf_event_tid(event, current);
4073 perf_output_put(&handle, mmap_event->event_id);
4074 perf_output_copy(&handle, mmap_event->file_name,
4075 mmap_event->file_size);
4076 perf_output_end(&handle);
4079 static int perf_event_mmap_match(struct perf_event *event,
4080 struct perf_mmap_event *mmap_event,
4083 if (event->state < PERF_EVENT_STATE_INACTIVE)
4086 if (event->cpu != -1 && event->cpu != smp_processor_id())
4089 if ((!executable && event->attr.mmap_data) ||
4090 (executable && event->attr.mmap))
4096 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4097 struct perf_mmap_event *mmap_event,
4100 struct perf_event *event;
4102 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4103 if (perf_event_mmap_match(event, mmap_event, executable))
4104 perf_event_mmap_output(event, mmap_event);
4108 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4110 struct perf_cpu_context *cpuctx;
4111 struct perf_event_context *ctx;
4112 struct vm_area_struct *vma = mmap_event->vma;
4113 struct file *file = vma->vm_file;
4121 memset(tmp, 0, sizeof(tmp));
4125 * d_path works from the end of the buffer backwards, so we
4126 * need to add enough zero bytes after the string to handle
4127 * the 64bit alignment we do later.
4129 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4131 name = strncpy(tmp, "//enomem", sizeof(tmp));
4134 name = d_path(&file->f_path, buf, PATH_MAX);
4136 name = strncpy(tmp, "//toolong", sizeof(tmp));
4140 if (arch_vma_name(mmap_event->vma)) {
4141 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4147 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4149 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4150 vma->vm_end >= vma->vm_mm->brk) {
4151 name = strncpy(tmp, "[heap]", sizeof(tmp));
4153 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4154 vma->vm_end >= vma->vm_mm->start_stack) {
4155 name = strncpy(tmp, "[stack]", sizeof(tmp));
4159 name = strncpy(tmp, "//anon", sizeof(tmp));
4164 size = ALIGN(strlen(name)+1, sizeof(u64));
4166 mmap_event->file_name = name;
4167 mmap_event->file_size = size;
4169 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4172 list_for_each_entry_rcu(pmu, &pmus, entry) {
4173 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4174 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4175 vma->vm_flags & VM_EXEC);
4177 ctxn = pmu->task_ctx_nr;
4181 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4183 perf_event_mmap_ctx(ctx, mmap_event,
4184 vma->vm_flags & VM_EXEC);
4187 put_cpu_ptr(pmu->pmu_cpu_context);
4194 void perf_event_mmap(struct vm_area_struct *vma)
4196 struct perf_mmap_event mmap_event;
4198 if (!atomic_read(&nr_mmap_events))
4201 mmap_event = (struct perf_mmap_event){
4207 .type = PERF_RECORD_MMAP,
4208 .misc = PERF_RECORD_MISC_USER,
4213 .start = vma->vm_start,
4214 .len = vma->vm_end - vma->vm_start,
4215 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4219 perf_event_mmap_event(&mmap_event);
4223 * IRQ throttle logging
4226 static void perf_log_throttle(struct perf_event *event, int enable)
4228 struct perf_output_handle handle;
4232 struct perf_event_header header;
4236 } throttle_event = {
4238 .type = PERF_RECORD_THROTTLE,
4240 .size = sizeof(throttle_event),
4242 .time = perf_clock(),
4243 .id = primary_event_id(event),
4244 .stream_id = event->id,
4248 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4250 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4254 perf_output_put(&handle, throttle_event);
4255 perf_output_end(&handle);
4259 * Generic event overflow handling, sampling.
4262 static int __perf_event_overflow(struct perf_event *event, int nmi,
4263 int throttle, struct perf_sample_data *data,
4264 struct pt_regs *regs)
4266 int events = atomic_read(&event->event_limit);
4267 struct hw_perf_event *hwc = &event->hw;
4271 * Non-sampling counters might still use the PMI to fold short
4272 * hardware counters, ignore those.
4274 if (unlikely(!is_sampling_event(event)))
4280 if (hwc->interrupts != MAX_INTERRUPTS) {
4282 if (HZ * hwc->interrupts >
4283 (u64)sysctl_perf_event_sample_rate) {
4284 hwc->interrupts = MAX_INTERRUPTS;
4285 perf_log_throttle(event, 0);
4290 * Keep re-disabling events even though on the previous
4291 * pass we disabled it - just in case we raced with a
4292 * sched-in and the event got enabled again:
4298 if (event->attr.freq) {
4299 u64 now = perf_clock();
4300 s64 delta = now - hwc->freq_time_stamp;
4302 hwc->freq_time_stamp = now;
4304 if (delta > 0 && delta < 2*TICK_NSEC)
4305 perf_adjust_period(event, delta, hwc->last_period);
4309 * XXX event_limit might not quite work as expected on inherited
4313 event->pending_kill = POLL_IN;
4314 if (events && atomic_dec_and_test(&event->event_limit)) {
4316 event->pending_kill = POLL_HUP;
4318 event->pending_disable = 1;
4319 irq_work_queue(&event->pending);
4321 perf_event_disable(event);
4324 if (event->overflow_handler)
4325 event->overflow_handler(event, nmi, data, regs);
4327 perf_event_output(event, nmi, data, regs);
4332 int perf_event_overflow(struct perf_event *event, int nmi,
4333 struct perf_sample_data *data,
4334 struct pt_regs *regs)
4336 return __perf_event_overflow(event, nmi, 1, data, regs);
4340 * Generic software event infrastructure
4343 struct swevent_htable {
4344 struct swevent_hlist *swevent_hlist;
4345 struct mutex hlist_mutex;
4348 /* Recursion avoidance in each contexts */
4349 int recursion[PERF_NR_CONTEXTS];
4352 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4355 * We directly increment event->count and keep a second value in
4356 * event->hw.period_left to count intervals. This period event
4357 * is kept in the range [-sample_period, 0] so that we can use the
4361 static u64 perf_swevent_set_period(struct perf_event *event)
4363 struct hw_perf_event *hwc = &event->hw;
4364 u64 period = hwc->last_period;
4368 hwc->last_period = hwc->sample_period;
4371 old = val = local64_read(&hwc->period_left);
4375 nr = div64_u64(period + val, period);
4376 offset = nr * period;
4378 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4384 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4385 int nmi, struct perf_sample_data *data,
4386 struct pt_regs *regs)
4388 struct hw_perf_event *hwc = &event->hw;
4391 data->period = event->hw.last_period;
4393 overflow = perf_swevent_set_period(event);
4395 if (hwc->interrupts == MAX_INTERRUPTS)
4398 for (; overflow; overflow--) {
4399 if (__perf_event_overflow(event, nmi, throttle,
4402 * We inhibit the overflow from happening when
4403 * hwc->interrupts == MAX_INTERRUPTS.
4411 static void perf_swevent_event(struct perf_event *event, u64 nr,
4412 int nmi, struct perf_sample_data *data,
4413 struct pt_regs *regs)
4415 struct hw_perf_event *hwc = &event->hw;
4417 local64_add(nr, &event->count);
4422 if (!is_sampling_event(event))
4425 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4426 return perf_swevent_overflow(event, 1, nmi, data, regs);
4428 if (local64_add_negative(nr, &hwc->period_left))
4431 perf_swevent_overflow(event, 0, nmi, data, regs);
4434 static int perf_exclude_event(struct perf_event *event,
4435 struct pt_regs *regs)
4437 if (event->hw.state & PERF_HES_STOPPED)
4441 if (event->attr.exclude_user && user_mode(regs))
4444 if (event->attr.exclude_kernel && !user_mode(regs))
4451 static int perf_swevent_match(struct perf_event *event,
4452 enum perf_type_id type,
4454 struct perf_sample_data *data,
4455 struct pt_regs *regs)
4457 if (event->attr.type != type)
4460 if (event->attr.config != event_id)
4463 if (perf_exclude_event(event, regs))
4469 static inline u64 swevent_hash(u64 type, u32 event_id)
4471 u64 val = event_id | (type << 32);
4473 return hash_64(val, SWEVENT_HLIST_BITS);
4476 static inline struct hlist_head *
4477 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4479 u64 hash = swevent_hash(type, event_id);
4481 return &hlist->heads[hash];
4484 /* For the read side: events when they trigger */
4485 static inline struct hlist_head *
4486 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4488 struct swevent_hlist *hlist;
4490 hlist = rcu_dereference(swhash->swevent_hlist);
4494 return __find_swevent_head(hlist, type, event_id);
4497 /* For the event head insertion and removal in the hlist */
4498 static inline struct hlist_head *
4499 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4501 struct swevent_hlist *hlist;
4502 u32 event_id = event->attr.config;
4503 u64 type = event->attr.type;
4506 * Event scheduling is always serialized against hlist allocation
4507 * and release. Which makes the protected version suitable here.
4508 * The context lock guarantees that.
4510 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4511 lockdep_is_held(&event->ctx->lock));
4515 return __find_swevent_head(hlist, type, event_id);
4518 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4520 struct perf_sample_data *data,
4521 struct pt_regs *regs)
4523 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4524 struct perf_event *event;
4525 struct hlist_node *node;
4526 struct hlist_head *head;
4529 head = find_swevent_head_rcu(swhash, type, event_id);
4533 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4534 if (perf_swevent_match(event, type, event_id, data, regs))
4535 perf_swevent_event(event, nr, nmi, data, regs);
4541 int perf_swevent_get_recursion_context(void)
4543 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4545 return get_recursion_context(swhash->recursion);
4547 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4549 void inline perf_swevent_put_recursion_context(int rctx)
4551 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4553 put_recursion_context(swhash->recursion, rctx);
4556 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4557 struct pt_regs *regs, u64 addr)
4559 struct perf_sample_data data;
4562 preempt_disable_notrace();
4563 rctx = perf_swevent_get_recursion_context();
4567 perf_sample_data_init(&data, addr);
4569 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4571 perf_swevent_put_recursion_context(rctx);
4572 preempt_enable_notrace();
4575 static void perf_swevent_read(struct perf_event *event)
4579 static int perf_swevent_add(struct perf_event *event, int flags)
4581 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4582 struct hw_perf_event *hwc = &event->hw;
4583 struct hlist_head *head;
4585 if (is_sampling_event(event)) {
4586 hwc->last_period = hwc->sample_period;
4587 perf_swevent_set_period(event);
4590 hwc->state = !(flags & PERF_EF_START);
4592 head = find_swevent_head(swhash, event);
4593 if (WARN_ON_ONCE(!head))
4596 hlist_add_head_rcu(&event->hlist_entry, head);
4601 static void perf_swevent_del(struct perf_event *event, int flags)
4603 hlist_del_rcu(&event->hlist_entry);
4606 static void perf_swevent_start(struct perf_event *event, int flags)
4608 event->hw.state = 0;
4611 static void perf_swevent_stop(struct perf_event *event, int flags)
4613 event->hw.state = PERF_HES_STOPPED;
4616 /* Deref the hlist from the update side */
4617 static inline struct swevent_hlist *
4618 swevent_hlist_deref(struct swevent_htable *swhash)
4620 return rcu_dereference_protected(swhash->swevent_hlist,
4621 lockdep_is_held(&swhash->hlist_mutex));
4624 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4626 struct swevent_hlist *hlist;
4628 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4632 static void swevent_hlist_release(struct swevent_htable *swhash)
4634 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4639 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4640 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4643 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4645 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4647 mutex_lock(&swhash->hlist_mutex);
4649 if (!--swhash->hlist_refcount)
4650 swevent_hlist_release(swhash);
4652 mutex_unlock(&swhash->hlist_mutex);
4655 static void swevent_hlist_put(struct perf_event *event)
4659 if (event->cpu != -1) {
4660 swevent_hlist_put_cpu(event, event->cpu);
4664 for_each_possible_cpu(cpu)
4665 swevent_hlist_put_cpu(event, cpu);
4668 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4670 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4673 mutex_lock(&swhash->hlist_mutex);
4675 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4676 struct swevent_hlist *hlist;
4678 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4683 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4685 swhash->hlist_refcount++;
4687 mutex_unlock(&swhash->hlist_mutex);
4692 static int swevent_hlist_get(struct perf_event *event)
4695 int cpu, failed_cpu;
4697 if (event->cpu != -1)
4698 return swevent_hlist_get_cpu(event, event->cpu);
4701 for_each_possible_cpu(cpu) {
4702 err = swevent_hlist_get_cpu(event, cpu);
4712 for_each_possible_cpu(cpu) {
4713 if (cpu == failed_cpu)
4715 swevent_hlist_put_cpu(event, cpu);
4722 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4724 static void sw_perf_event_destroy(struct perf_event *event)
4726 u64 event_id = event->attr.config;
4728 WARN_ON(event->parent);
4730 jump_label_dec(&perf_swevent_enabled[event_id]);
4731 swevent_hlist_put(event);
4734 static int perf_swevent_init(struct perf_event *event)
4736 int event_id = event->attr.config;
4738 if (event->attr.type != PERF_TYPE_SOFTWARE)
4742 case PERF_COUNT_SW_CPU_CLOCK:
4743 case PERF_COUNT_SW_TASK_CLOCK:
4750 if (event_id > PERF_COUNT_SW_MAX)
4753 if (!event->parent) {
4756 err = swevent_hlist_get(event);
4760 jump_label_inc(&perf_swevent_enabled[event_id]);
4761 event->destroy = sw_perf_event_destroy;
4767 static struct pmu perf_swevent = {
4768 .task_ctx_nr = perf_sw_context,
4770 .event_init = perf_swevent_init,
4771 .add = perf_swevent_add,
4772 .del = perf_swevent_del,
4773 .start = perf_swevent_start,
4774 .stop = perf_swevent_stop,
4775 .read = perf_swevent_read,
4778 #ifdef CONFIG_EVENT_TRACING
4780 static int perf_tp_filter_match(struct perf_event *event,
4781 struct perf_sample_data *data)
4783 void *record = data->raw->data;
4785 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4790 static int perf_tp_event_match(struct perf_event *event,
4791 struct perf_sample_data *data,
4792 struct pt_regs *regs)
4795 * All tracepoints are from kernel-space.
4797 if (event->attr.exclude_kernel)
4800 if (!perf_tp_filter_match(event, data))
4806 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4807 struct pt_regs *regs, struct hlist_head *head, int rctx)
4809 struct perf_sample_data data;
4810 struct perf_event *event;
4811 struct hlist_node *node;
4813 struct perf_raw_record raw = {
4818 perf_sample_data_init(&data, addr);
4821 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4822 if (perf_tp_event_match(event, &data, regs))
4823 perf_swevent_event(event, count, 1, &data, regs);
4826 perf_swevent_put_recursion_context(rctx);
4828 EXPORT_SYMBOL_GPL(perf_tp_event);
4830 static void tp_perf_event_destroy(struct perf_event *event)
4832 perf_trace_destroy(event);
4835 static int perf_tp_event_init(struct perf_event *event)
4839 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4842 err = perf_trace_init(event);
4846 event->destroy = tp_perf_event_destroy;
4851 static struct pmu perf_tracepoint = {
4852 .task_ctx_nr = perf_sw_context,
4854 .event_init = perf_tp_event_init,
4855 .add = perf_trace_add,
4856 .del = perf_trace_del,
4857 .start = perf_swevent_start,
4858 .stop = perf_swevent_stop,
4859 .read = perf_swevent_read,
4862 static inline void perf_tp_register(void)
4864 perf_pmu_register(&perf_tracepoint);
4867 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4872 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4875 filter_str = strndup_user(arg, PAGE_SIZE);
4876 if (IS_ERR(filter_str))
4877 return PTR_ERR(filter_str);
4879 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4885 static void perf_event_free_filter(struct perf_event *event)
4887 ftrace_profile_free_filter(event);
4892 static inline void perf_tp_register(void)
4896 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4901 static void perf_event_free_filter(struct perf_event *event)
4905 #endif /* CONFIG_EVENT_TRACING */
4907 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4908 void perf_bp_event(struct perf_event *bp, void *data)
4910 struct perf_sample_data sample;
4911 struct pt_regs *regs = data;
4913 perf_sample_data_init(&sample, bp->attr.bp_addr);
4915 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4916 perf_swevent_event(bp, 1, 1, &sample, regs);
4921 * hrtimer based swevent callback
4924 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4926 enum hrtimer_restart ret = HRTIMER_RESTART;
4927 struct perf_sample_data data;
4928 struct pt_regs *regs;
4929 struct perf_event *event;
4932 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4933 event->pmu->read(event);
4935 perf_sample_data_init(&data, 0);
4936 data.period = event->hw.last_period;
4937 regs = get_irq_regs();
4939 if (regs && !perf_exclude_event(event, regs)) {
4940 if (!(event->attr.exclude_idle && current->pid == 0))
4941 if (perf_event_overflow(event, 0, &data, regs))
4942 ret = HRTIMER_NORESTART;
4945 period = max_t(u64, 10000, event->hw.sample_period);
4946 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4951 static void perf_swevent_start_hrtimer(struct perf_event *event)
4953 struct hw_perf_event *hwc = &event->hw;
4956 if (!is_sampling_event(event))
4959 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4960 hwc->hrtimer.function = perf_swevent_hrtimer;
4962 period = local64_read(&hwc->period_left);
4967 local64_set(&hwc->period_left, 0);
4969 period = max_t(u64, 10000, hwc->sample_period);
4971 __hrtimer_start_range_ns(&hwc->hrtimer,
4972 ns_to_ktime(period), 0,
4973 HRTIMER_MODE_REL_PINNED, 0);
4976 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4978 struct hw_perf_event *hwc = &event->hw;
4980 if (is_sampling_event(event)) {
4981 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4982 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4984 hrtimer_cancel(&hwc->hrtimer);
4989 * Software event: cpu wall time clock
4992 static void cpu_clock_event_update(struct perf_event *event)
4997 now = local_clock();
4998 prev = local64_xchg(&event->hw.prev_count, now);
4999 local64_add(now - prev, &event->count);
5002 static void cpu_clock_event_start(struct perf_event *event, int flags)
5004 local64_set(&event->hw.prev_count, local_clock());
5005 perf_swevent_start_hrtimer(event);
5008 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5010 perf_swevent_cancel_hrtimer(event);
5011 cpu_clock_event_update(event);
5014 static int cpu_clock_event_add(struct perf_event *event, int flags)
5016 if (flags & PERF_EF_START)
5017 cpu_clock_event_start(event, flags);
5022 static void cpu_clock_event_del(struct perf_event *event, int flags)
5024 cpu_clock_event_stop(event, flags);
5027 static void cpu_clock_event_read(struct perf_event *event)
5029 cpu_clock_event_update(event);
5032 static int cpu_clock_event_init(struct perf_event *event)
5034 if (event->attr.type != PERF_TYPE_SOFTWARE)
5037 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5043 static struct pmu perf_cpu_clock = {
5044 .task_ctx_nr = perf_sw_context,
5046 .event_init = cpu_clock_event_init,
5047 .add = cpu_clock_event_add,
5048 .del = cpu_clock_event_del,
5049 .start = cpu_clock_event_start,
5050 .stop = cpu_clock_event_stop,
5051 .read = cpu_clock_event_read,
5055 * Software event: task time clock
5058 static void task_clock_event_update(struct perf_event *event, u64 now)
5063 prev = local64_xchg(&event->hw.prev_count, now);
5065 local64_add(delta, &event->count);
5068 static void task_clock_event_start(struct perf_event *event, int flags)
5070 local64_set(&event->hw.prev_count, event->ctx->time);
5071 perf_swevent_start_hrtimer(event);
5074 static void task_clock_event_stop(struct perf_event *event, int flags)
5076 perf_swevent_cancel_hrtimer(event);
5077 task_clock_event_update(event, event->ctx->time);
5080 static int task_clock_event_add(struct perf_event *event, int flags)
5082 if (flags & PERF_EF_START)
5083 task_clock_event_start(event, flags);
5088 static void task_clock_event_del(struct perf_event *event, int flags)
5090 task_clock_event_stop(event, PERF_EF_UPDATE);
5093 static void task_clock_event_read(struct perf_event *event)
5098 update_context_time(event->ctx);
5099 time = event->ctx->time;
5101 u64 now = perf_clock();
5102 u64 delta = now - event->ctx->timestamp;
5103 time = event->ctx->time + delta;
5106 task_clock_event_update(event, time);
5109 static int task_clock_event_init(struct perf_event *event)
5111 if (event->attr.type != PERF_TYPE_SOFTWARE)
5114 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5120 static struct pmu perf_task_clock = {
5121 .task_ctx_nr = perf_sw_context,
5123 .event_init = task_clock_event_init,
5124 .add = task_clock_event_add,
5125 .del = task_clock_event_del,
5126 .start = task_clock_event_start,
5127 .stop = task_clock_event_stop,
5128 .read = task_clock_event_read,
5131 static void perf_pmu_nop_void(struct pmu *pmu)
5135 static int perf_pmu_nop_int(struct pmu *pmu)
5140 static void perf_pmu_start_txn(struct pmu *pmu)
5142 perf_pmu_disable(pmu);
5145 static int perf_pmu_commit_txn(struct pmu *pmu)
5147 perf_pmu_enable(pmu);
5151 static void perf_pmu_cancel_txn(struct pmu *pmu)
5153 perf_pmu_enable(pmu);
5157 * Ensures all contexts with the same task_ctx_nr have the same
5158 * pmu_cpu_context too.
5160 static void *find_pmu_context(int ctxn)
5167 list_for_each_entry(pmu, &pmus, entry) {
5168 if (pmu->task_ctx_nr == ctxn)
5169 return pmu->pmu_cpu_context;
5175 static void free_pmu_context(void * __percpu cpu_context)
5179 mutex_lock(&pmus_lock);
5181 * Like a real lame refcount.
5183 list_for_each_entry(pmu, &pmus, entry) {
5184 if (pmu->pmu_cpu_context == cpu_context)
5188 free_percpu(cpu_context);
5190 mutex_unlock(&pmus_lock);
5193 int perf_pmu_register(struct pmu *pmu)
5197 mutex_lock(&pmus_lock);
5199 pmu->pmu_disable_count = alloc_percpu(int);
5200 if (!pmu->pmu_disable_count)
5203 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5204 if (pmu->pmu_cpu_context)
5205 goto got_cpu_context;
5207 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5208 if (!pmu->pmu_cpu_context)
5211 for_each_possible_cpu(cpu) {
5212 struct perf_cpu_context *cpuctx;
5214 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5215 __perf_event_init_context(&cpuctx->ctx);
5216 cpuctx->ctx.type = cpu_context;
5217 cpuctx->ctx.pmu = pmu;
5218 cpuctx->jiffies_interval = 1;
5219 INIT_LIST_HEAD(&cpuctx->rotation_list);
5223 if (!pmu->start_txn) {
5224 if (pmu->pmu_enable) {
5226 * If we have pmu_enable/pmu_disable calls, install
5227 * transaction stubs that use that to try and batch
5228 * hardware accesses.
5230 pmu->start_txn = perf_pmu_start_txn;
5231 pmu->commit_txn = perf_pmu_commit_txn;
5232 pmu->cancel_txn = perf_pmu_cancel_txn;
5234 pmu->start_txn = perf_pmu_nop_void;
5235 pmu->commit_txn = perf_pmu_nop_int;
5236 pmu->cancel_txn = perf_pmu_nop_void;
5240 if (!pmu->pmu_enable) {
5241 pmu->pmu_enable = perf_pmu_nop_void;
5242 pmu->pmu_disable = perf_pmu_nop_void;
5245 list_add_rcu(&pmu->entry, &pmus);
5248 mutex_unlock(&pmus_lock);
5253 free_percpu(pmu->pmu_disable_count);
5257 void perf_pmu_unregister(struct pmu *pmu)
5259 mutex_lock(&pmus_lock);
5260 list_del_rcu(&pmu->entry);
5261 mutex_unlock(&pmus_lock);
5264 * We dereference the pmu list under both SRCU and regular RCU, so
5265 * synchronize against both of those.
5267 synchronize_srcu(&pmus_srcu);
5270 free_percpu(pmu->pmu_disable_count);
5271 free_pmu_context(pmu->pmu_cpu_context);
5274 struct pmu *perf_init_event(struct perf_event *event)
5276 struct pmu *pmu = NULL;
5279 idx = srcu_read_lock(&pmus_srcu);
5280 list_for_each_entry_rcu(pmu, &pmus, entry) {
5281 int ret = pmu->event_init(event);
5285 if (ret != -ENOENT) {
5290 pmu = ERR_PTR(-ENOENT);
5292 srcu_read_unlock(&pmus_srcu, idx);
5298 * Allocate and initialize a event structure
5300 static struct perf_event *
5301 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5302 struct task_struct *task,
5303 struct perf_event *group_leader,
5304 struct perf_event *parent_event,
5305 perf_overflow_handler_t overflow_handler)
5308 struct perf_event *event;
5309 struct hw_perf_event *hwc;
5312 event = kzalloc(sizeof(*event), GFP_KERNEL);
5314 return ERR_PTR(-ENOMEM);
5317 * Single events are their own group leaders, with an
5318 * empty sibling list:
5321 group_leader = event;
5323 mutex_init(&event->child_mutex);
5324 INIT_LIST_HEAD(&event->child_list);
5326 INIT_LIST_HEAD(&event->group_entry);
5327 INIT_LIST_HEAD(&event->event_entry);
5328 INIT_LIST_HEAD(&event->sibling_list);
5329 init_waitqueue_head(&event->waitq);
5330 init_irq_work(&event->pending, perf_pending_event);
5332 mutex_init(&event->mmap_mutex);
5335 event->attr = *attr;
5336 event->group_leader = group_leader;
5340 event->parent = parent_event;
5342 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5343 event->id = atomic64_inc_return(&perf_event_id);
5345 event->state = PERF_EVENT_STATE_INACTIVE;
5348 event->attach_state = PERF_ATTACH_TASK;
5349 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5351 * hw_breakpoint is a bit difficult here..
5353 if (attr->type == PERF_TYPE_BREAKPOINT)
5354 event->hw.bp_target = task;
5358 if (!overflow_handler && parent_event)
5359 overflow_handler = parent_event->overflow_handler;
5361 event->overflow_handler = overflow_handler;
5364 event->state = PERF_EVENT_STATE_OFF;
5369 hwc->sample_period = attr->sample_period;
5370 if (attr->freq && attr->sample_freq)
5371 hwc->sample_period = 1;
5372 hwc->last_period = hwc->sample_period;
5374 local64_set(&hwc->period_left, hwc->sample_period);
5377 * we currently do not support PERF_FORMAT_GROUP on inherited events
5379 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5382 pmu = perf_init_event(event);
5388 else if (IS_ERR(pmu))
5393 put_pid_ns(event->ns);
5395 return ERR_PTR(err);
5400 if (!event->parent) {
5401 if (event->attach_state & PERF_ATTACH_TASK)
5402 jump_label_inc(&perf_task_events);
5403 if (event->attr.mmap || event->attr.mmap_data)
5404 atomic_inc(&nr_mmap_events);
5405 if (event->attr.comm)
5406 atomic_inc(&nr_comm_events);
5407 if (event->attr.task)
5408 atomic_inc(&nr_task_events);
5409 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5410 err = get_callchain_buffers();
5413 return ERR_PTR(err);
5421 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5422 struct perf_event_attr *attr)
5427 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5431 * zero the full structure, so that a short copy will be nice.
5433 memset(attr, 0, sizeof(*attr));
5435 ret = get_user(size, &uattr->size);
5439 if (size > PAGE_SIZE) /* silly large */
5442 if (!size) /* abi compat */
5443 size = PERF_ATTR_SIZE_VER0;
5445 if (size < PERF_ATTR_SIZE_VER0)
5449 * If we're handed a bigger struct than we know of,
5450 * ensure all the unknown bits are 0 - i.e. new
5451 * user-space does not rely on any kernel feature
5452 * extensions we dont know about yet.
5454 if (size > sizeof(*attr)) {
5455 unsigned char __user *addr;
5456 unsigned char __user *end;
5459 addr = (void __user *)uattr + sizeof(*attr);
5460 end = (void __user *)uattr + size;
5462 for (; addr < end; addr++) {
5463 ret = get_user(val, addr);
5469 size = sizeof(*attr);
5472 ret = copy_from_user(attr, uattr, size);
5477 * If the type exists, the corresponding creation will verify
5480 if (attr->type >= PERF_TYPE_MAX)
5483 if (attr->__reserved_1)
5486 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5489 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5496 put_user(sizeof(*attr), &uattr->size);
5502 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5504 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5510 /* don't allow circular references */
5511 if (event == output_event)
5515 * Don't allow cross-cpu buffers
5517 if (output_event->cpu != event->cpu)
5521 * If its not a per-cpu buffer, it must be the same task.
5523 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5527 mutex_lock(&event->mmap_mutex);
5528 /* Can't redirect output if we've got an active mmap() */
5529 if (atomic_read(&event->mmap_count))
5533 /* get the buffer we want to redirect to */
5534 buffer = perf_buffer_get(output_event);
5539 old_buffer = event->buffer;
5540 rcu_assign_pointer(event->buffer, buffer);
5543 mutex_unlock(&event->mmap_mutex);
5546 perf_buffer_put(old_buffer);
5552 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5554 * @attr_uptr: event_id type attributes for monitoring/sampling
5557 * @group_fd: group leader event fd
5559 SYSCALL_DEFINE5(perf_event_open,
5560 struct perf_event_attr __user *, attr_uptr,
5561 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5563 struct perf_event *group_leader = NULL, *output_event = NULL;
5564 struct perf_event *event, *sibling;
5565 struct perf_event_attr attr;
5566 struct perf_event_context *ctx;
5567 struct file *event_file = NULL;
5568 struct file *group_file = NULL;
5569 struct task_struct *task = NULL;
5573 int fput_needed = 0;
5576 /* for future expandability... */
5577 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5580 err = perf_copy_attr(attr_uptr, &attr);
5584 if (!attr.exclude_kernel) {
5585 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5590 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5594 event_fd = get_unused_fd_flags(O_RDWR);
5598 if (group_fd != -1) {
5599 group_leader = perf_fget_light(group_fd, &fput_needed);
5600 if (IS_ERR(group_leader)) {
5601 err = PTR_ERR(group_leader);
5604 group_file = group_leader->filp;
5605 if (flags & PERF_FLAG_FD_OUTPUT)
5606 output_event = group_leader;
5607 if (flags & PERF_FLAG_FD_NO_GROUP)
5608 group_leader = NULL;
5612 task = find_lively_task_by_vpid(pid);
5614 err = PTR_ERR(task);
5619 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5620 if (IS_ERR(event)) {
5621 err = PTR_ERR(event);
5626 * Special case software events and allow them to be part of
5627 * any hardware group.
5632 (is_software_event(event) != is_software_event(group_leader))) {
5633 if (is_software_event(event)) {
5635 * If event and group_leader are not both a software
5636 * event, and event is, then group leader is not.
5638 * Allow the addition of software events to !software
5639 * groups, this is safe because software events never
5642 pmu = group_leader->pmu;
5643 } else if (is_software_event(group_leader) &&
5644 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5646 * In case the group is a pure software group, and we
5647 * try to add a hardware event, move the whole group to
5648 * the hardware context.
5655 * Get the target context (task or percpu):
5657 ctx = find_get_context(pmu, task, cpu);
5664 * Look up the group leader (we will attach this event to it):
5670 * Do not allow a recursive hierarchy (this new sibling
5671 * becoming part of another group-sibling):
5673 if (group_leader->group_leader != group_leader)
5676 * Do not allow to attach to a group in a different
5677 * task or CPU context:
5680 if (group_leader->ctx->type != ctx->type)
5683 if (group_leader->ctx != ctx)
5688 * Only a group leader can be exclusive or pinned
5690 if (attr.exclusive || attr.pinned)
5695 err = perf_event_set_output(event, output_event);
5700 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5701 if (IS_ERR(event_file)) {
5702 err = PTR_ERR(event_file);
5707 struct perf_event_context *gctx = group_leader->ctx;
5709 mutex_lock(&gctx->mutex);
5710 perf_event_remove_from_context(group_leader);
5711 list_for_each_entry(sibling, &group_leader->sibling_list,
5713 perf_event_remove_from_context(sibling);
5716 mutex_unlock(&gctx->mutex);
5720 event->filp = event_file;
5721 WARN_ON_ONCE(ctx->parent_ctx);
5722 mutex_lock(&ctx->mutex);
5725 perf_install_in_context(ctx, group_leader, cpu);
5727 list_for_each_entry(sibling, &group_leader->sibling_list,
5729 perf_install_in_context(ctx, sibling, cpu);
5734 perf_install_in_context(ctx, event, cpu);
5736 mutex_unlock(&ctx->mutex);
5738 event->owner = current;
5740 mutex_lock(¤t->perf_event_mutex);
5741 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5742 mutex_unlock(¤t->perf_event_mutex);
5745 * Precalculate sample_data sizes
5747 perf_event__header_size(event);
5750 * Drop the reference on the group_event after placing the
5751 * new event on the sibling_list. This ensures destruction
5752 * of the group leader will find the pointer to itself in
5753 * perf_group_detach().
5755 fput_light(group_file, fput_needed);
5756 fd_install(event_fd, event_file);
5765 put_task_struct(task);
5767 fput_light(group_file, fput_needed);
5769 put_unused_fd(event_fd);
5774 * perf_event_create_kernel_counter
5776 * @attr: attributes of the counter to create
5777 * @cpu: cpu in which the counter is bound
5778 * @task: task to profile (NULL for percpu)
5781 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5782 struct task_struct *task,
5783 perf_overflow_handler_t overflow_handler)
5785 struct perf_event_context *ctx;
5786 struct perf_event *event;
5790 * Get the target context (task or percpu):
5793 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
5794 if (IS_ERR(event)) {
5795 err = PTR_ERR(event);
5799 ctx = find_get_context(event->pmu, task, cpu);
5806 WARN_ON_ONCE(ctx->parent_ctx);
5807 mutex_lock(&ctx->mutex);
5808 perf_install_in_context(ctx, event, cpu);
5810 mutex_unlock(&ctx->mutex);
5817 return ERR_PTR(err);
5819 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5821 static void sync_child_event(struct perf_event *child_event,
5822 struct task_struct *child)
5824 struct perf_event *parent_event = child_event->parent;
5827 if (child_event->attr.inherit_stat)
5828 perf_event_read_event(child_event, child);
5830 child_val = perf_event_count(child_event);
5833 * Add back the child's count to the parent's count:
5835 atomic64_add(child_val, &parent_event->child_count);
5836 atomic64_add(child_event->total_time_enabled,
5837 &parent_event->child_total_time_enabled);
5838 atomic64_add(child_event->total_time_running,
5839 &parent_event->child_total_time_running);
5842 * Remove this event from the parent's list
5844 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5845 mutex_lock(&parent_event->child_mutex);
5846 list_del_init(&child_event->child_list);
5847 mutex_unlock(&parent_event->child_mutex);
5850 * Release the parent event, if this was the last
5853 fput(parent_event->filp);
5857 __perf_event_exit_task(struct perf_event *child_event,
5858 struct perf_event_context *child_ctx,
5859 struct task_struct *child)
5861 struct perf_event *parent_event;
5863 perf_event_remove_from_context(child_event);
5865 parent_event = child_event->parent;
5867 * It can happen that parent exits first, and has events
5868 * that are still around due to the child reference. These
5869 * events need to be zapped - but otherwise linger.
5872 sync_child_event(child_event, child);
5873 free_event(child_event);
5877 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
5879 struct perf_event *child_event, *tmp;
5880 struct perf_event_context *child_ctx;
5881 unsigned long flags;
5883 if (likely(!child->perf_event_ctxp[ctxn])) {
5884 perf_event_task(child, NULL, 0);
5888 local_irq_save(flags);
5890 * We can't reschedule here because interrupts are disabled,
5891 * and either child is current or it is a task that can't be
5892 * scheduled, so we are now safe from rescheduling changing
5895 child_ctx = child->perf_event_ctxp[ctxn];
5896 task_ctx_sched_out(child_ctx, EVENT_ALL);
5899 * Take the context lock here so that if find_get_context is
5900 * reading child->perf_event_ctxp, we wait until it has
5901 * incremented the context's refcount before we do put_ctx below.
5903 raw_spin_lock(&child_ctx->lock);
5904 child->perf_event_ctxp[ctxn] = NULL;
5906 * If this context is a clone; unclone it so it can't get
5907 * swapped to another process while we're removing all
5908 * the events from it.
5910 unclone_ctx(child_ctx);
5911 update_context_time(child_ctx);
5912 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5915 * Report the task dead after unscheduling the events so that we
5916 * won't get any samples after PERF_RECORD_EXIT. We can however still
5917 * get a few PERF_RECORD_READ events.
5919 perf_event_task(child, child_ctx, 0);
5922 * We can recurse on the same lock type through:
5924 * __perf_event_exit_task()
5925 * sync_child_event()
5926 * fput(parent_event->filp)
5928 * mutex_lock(&ctx->mutex)
5930 * But since its the parent context it won't be the same instance.
5932 mutex_lock(&child_ctx->mutex);
5935 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5937 __perf_event_exit_task(child_event, child_ctx, child);
5939 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5941 __perf_event_exit_task(child_event, child_ctx, child);
5944 * If the last event was a group event, it will have appended all
5945 * its siblings to the list, but we obtained 'tmp' before that which
5946 * will still point to the list head terminating the iteration.
5948 if (!list_empty(&child_ctx->pinned_groups) ||
5949 !list_empty(&child_ctx->flexible_groups))
5952 mutex_unlock(&child_ctx->mutex);
5958 * When a child task exits, feed back event values to parent events.
5960 void perf_event_exit_task(struct task_struct *child)
5962 struct perf_event *event, *tmp;
5965 mutex_lock(&child->perf_event_mutex);
5966 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
5968 list_del_init(&event->owner_entry);
5971 * Ensure the list deletion is visible before we clear
5972 * the owner, closes a race against perf_release() where
5973 * we need to serialize on the owner->perf_event_mutex.
5976 event->owner = NULL;
5978 mutex_unlock(&child->perf_event_mutex);
5980 for_each_task_context_nr(ctxn)
5981 perf_event_exit_task_context(child, ctxn);
5984 static void perf_free_event(struct perf_event *event,
5985 struct perf_event_context *ctx)
5987 struct perf_event *parent = event->parent;
5989 if (WARN_ON_ONCE(!parent))
5992 mutex_lock(&parent->child_mutex);
5993 list_del_init(&event->child_list);
5994 mutex_unlock(&parent->child_mutex);
5998 perf_group_detach(event);
5999 list_del_event(event, ctx);
6004 * free an unexposed, unused context as created by inheritance by
6005 * perf_event_init_task below, used by fork() in case of fail.
6007 void perf_event_free_task(struct task_struct *task)
6009 struct perf_event_context *ctx;
6010 struct perf_event *event, *tmp;
6013 for_each_task_context_nr(ctxn) {
6014 ctx = task->perf_event_ctxp[ctxn];
6018 mutex_lock(&ctx->mutex);
6020 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6022 perf_free_event(event, ctx);
6024 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6026 perf_free_event(event, ctx);
6028 if (!list_empty(&ctx->pinned_groups) ||
6029 !list_empty(&ctx->flexible_groups))
6032 mutex_unlock(&ctx->mutex);
6038 void perf_event_delayed_put(struct task_struct *task)
6042 for_each_task_context_nr(ctxn)
6043 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6047 * inherit a event from parent task to child task:
6049 static struct perf_event *
6050 inherit_event(struct perf_event *parent_event,
6051 struct task_struct *parent,
6052 struct perf_event_context *parent_ctx,
6053 struct task_struct *child,
6054 struct perf_event *group_leader,
6055 struct perf_event_context *child_ctx)
6057 struct perf_event *child_event;
6058 unsigned long flags;
6061 * Instead of creating recursive hierarchies of events,
6062 * we link inherited events back to the original parent,
6063 * which has a filp for sure, which we use as the reference
6066 if (parent_event->parent)
6067 parent_event = parent_event->parent;
6069 child_event = perf_event_alloc(&parent_event->attr,
6072 group_leader, parent_event,
6074 if (IS_ERR(child_event))
6079 * Make the child state follow the state of the parent event,
6080 * not its attr.disabled bit. We hold the parent's mutex,
6081 * so we won't race with perf_event_{en, dis}able_family.
6083 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6084 child_event->state = PERF_EVENT_STATE_INACTIVE;
6086 child_event->state = PERF_EVENT_STATE_OFF;
6088 if (parent_event->attr.freq) {
6089 u64 sample_period = parent_event->hw.sample_period;
6090 struct hw_perf_event *hwc = &child_event->hw;
6092 hwc->sample_period = sample_period;
6093 hwc->last_period = sample_period;
6095 local64_set(&hwc->period_left, sample_period);
6098 child_event->ctx = child_ctx;
6099 child_event->overflow_handler = parent_event->overflow_handler;
6102 * Link it up in the child's context:
6104 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6105 add_event_to_ctx(child_event, child_ctx);
6106 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6109 * Get a reference to the parent filp - we will fput it
6110 * when the child event exits. This is safe to do because
6111 * we are in the parent and we know that the filp still
6112 * exists and has a nonzero count:
6114 atomic_long_inc(&parent_event->filp->f_count);
6117 * Link this into the parent event's child list
6119 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6120 mutex_lock(&parent_event->child_mutex);
6121 list_add_tail(&child_event->child_list, &parent_event->child_list);
6122 mutex_unlock(&parent_event->child_mutex);
6127 static int inherit_group(struct perf_event *parent_event,
6128 struct task_struct *parent,
6129 struct perf_event_context *parent_ctx,
6130 struct task_struct *child,
6131 struct perf_event_context *child_ctx)
6133 struct perf_event *leader;
6134 struct perf_event *sub;
6135 struct perf_event *child_ctr;
6137 leader = inherit_event(parent_event, parent, parent_ctx,
6138 child, NULL, child_ctx);
6140 return PTR_ERR(leader);
6141 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6142 child_ctr = inherit_event(sub, parent, parent_ctx,
6143 child, leader, child_ctx);
6144 if (IS_ERR(child_ctr))
6145 return PTR_ERR(child_ctr);
6151 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6152 struct perf_event_context *parent_ctx,
6153 struct task_struct *child, int ctxn,
6157 struct perf_event_context *child_ctx;
6159 if (!event->attr.inherit) {
6164 child_ctx = child->perf_event_ctxp[ctxn];
6167 * This is executed from the parent task context, so
6168 * inherit events that have been marked for cloning.
6169 * First allocate and initialize a context for the
6173 child_ctx = alloc_perf_context(event->pmu, child);
6177 child->perf_event_ctxp[ctxn] = child_ctx;
6180 ret = inherit_group(event, parent, parent_ctx,
6190 * Initialize the perf_event context in task_struct
6192 int perf_event_init_context(struct task_struct *child, int ctxn)
6194 struct perf_event_context *child_ctx, *parent_ctx;
6195 struct perf_event_context *cloned_ctx;
6196 struct perf_event *event;
6197 struct task_struct *parent = current;
6198 int inherited_all = 1;
6199 unsigned long flags;
6202 child->perf_event_ctxp[ctxn] = NULL;
6204 mutex_init(&child->perf_event_mutex);
6205 INIT_LIST_HEAD(&child->perf_event_list);
6207 if (likely(!parent->perf_event_ctxp[ctxn]))
6211 * If the parent's context is a clone, pin it so it won't get
6214 parent_ctx = perf_pin_task_context(parent, ctxn);
6217 * No need to check if parent_ctx != NULL here; since we saw
6218 * it non-NULL earlier, the only reason for it to become NULL
6219 * is if we exit, and since we're currently in the middle of
6220 * a fork we can't be exiting at the same time.
6224 * Lock the parent list. No need to lock the child - not PID
6225 * hashed yet and not running, so nobody can access it.
6227 mutex_lock(&parent_ctx->mutex);
6230 * We dont have to disable NMIs - we are only looking at
6231 * the list, not manipulating it:
6233 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6234 ret = inherit_task_group(event, parent, parent_ctx,
6235 child, ctxn, &inherited_all);
6241 * We can't hold ctx->lock when iterating the ->flexible_group list due
6242 * to allocations, but we need to prevent rotation because
6243 * rotate_ctx() will change the list from interrupt context.
6245 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6246 parent_ctx->rotate_disable = 1;
6247 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6249 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6250 ret = inherit_task_group(event, parent, parent_ctx,
6251 child, ctxn, &inherited_all);
6256 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6257 parent_ctx->rotate_disable = 0;
6258 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6260 child_ctx = child->perf_event_ctxp[ctxn];
6262 if (child_ctx && inherited_all) {
6264 * Mark the child context as a clone of the parent
6265 * context, or of whatever the parent is a clone of.
6266 * Note that if the parent is a clone, it could get
6267 * uncloned at any point, but that doesn't matter
6268 * because the list of events and the generation
6269 * count can't have changed since we took the mutex.
6271 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
6273 child_ctx->parent_ctx = cloned_ctx;
6274 child_ctx->parent_gen = parent_ctx->parent_gen;
6276 child_ctx->parent_ctx = parent_ctx;
6277 child_ctx->parent_gen = parent_ctx->generation;
6279 get_ctx(child_ctx->parent_ctx);
6282 mutex_unlock(&parent_ctx->mutex);
6284 perf_unpin_context(parent_ctx);
6290 * Initialize the perf_event context in task_struct
6292 int perf_event_init_task(struct task_struct *child)
6296 for_each_task_context_nr(ctxn) {
6297 ret = perf_event_init_context(child, ctxn);
6305 static void __init perf_event_init_all_cpus(void)
6307 struct swevent_htable *swhash;
6310 for_each_possible_cpu(cpu) {
6311 swhash = &per_cpu(swevent_htable, cpu);
6312 mutex_init(&swhash->hlist_mutex);
6313 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6317 static void __cpuinit perf_event_init_cpu(int cpu)
6319 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6321 mutex_lock(&swhash->hlist_mutex);
6322 if (swhash->hlist_refcount > 0) {
6323 struct swevent_hlist *hlist;
6325 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6327 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6329 mutex_unlock(&swhash->hlist_mutex);
6332 #ifdef CONFIG_HOTPLUG_CPU
6333 static void perf_pmu_rotate_stop(struct pmu *pmu)
6335 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6337 WARN_ON(!irqs_disabled());
6339 list_del_init(&cpuctx->rotation_list);
6342 static void __perf_event_exit_context(void *__info)
6344 struct perf_event_context *ctx = __info;
6345 struct perf_event *event, *tmp;
6347 perf_pmu_rotate_stop(ctx->pmu);
6349 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6350 __perf_event_remove_from_context(event);
6351 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6352 __perf_event_remove_from_context(event);
6355 static void perf_event_exit_cpu_context(int cpu)
6357 struct perf_event_context *ctx;
6361 idx = srcu_read_lock(&pmus_srcu);
6362 list_for_each_entry_rcu(pmu, &pmus, entry) {
6363 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6365 mutex_lock(&ctx->mutex);
6366 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6367 mutex_unlock(&ctx->mutex);
6369 srcu_read_unlock(&pmus_srcu, idx);
6372 static void perf_event_exit_cpu(int cpu)
6374 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6376 mutex_lock(&swhash->hlist_mutex);
6377 swevent_hlist_release(swhash);
6378 mutex_unlock(&swhash->hlist_mutex);
6380 perf_event_exit_cpu_context(cpu);
6383 static inline void perf_event_exit_cpu(int cpu) { }
6386 static int __cpuinit
6387 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6389 unsigned int cpu = (long)hcpu;
6391 switch (action & ~CPU_TASKS_FROZEN) {
6393 case CPU_UP_PREPARE:
6394 case CPU_DOWN_FAILED:
6395 perf_event_init_cpu(cpu);
6398 case CPU_UP_CANCELED:
6399 case CPU_DOWN_PREPARE:
6400 perf_event_exit_cpu(cpu);
6410 void __init perf_event_init(void)
6414 perf_event_init_all_cpus();
6415 init_srcu_struct(&pmus_srcu);
6416 perf_pmu_register(&perf_swevent);
6417 perf_pmu_register(&perf_cpu_clock);
6418 perf_pmu_register(&perf_task_clock);
6420 perf_cpu_notifier(perf_cpu_notify);
6422 ret = init_hw_breakpoint();
6423 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);