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>
35 #include <asm/irq_regs.h>
37 atomic_t perf_task_events __read_mostly;
38 static atomic_t nr_mmap_events __read_mostly;
39 static atomic_t nr_comm_events __read_mostly;
40 static atomic_t nr_task_events __read_mostly;
42 static LIST_HEAD(pmus);
43 static DEFINE_MUTEX(pmus_lock);
44 static struct srcu_struct pmus_srcu;
47 * perf event paranoia level:
48 * -1 - not paranoid at all
49 * 0 - disallow raw tracepoint access for unpriv
50 * 1 - disallow cpu events for unpriv
51 * 2 - disallow kernel profiling for unpriv
53 int sysctl_perf_event_paranoid __read_mostly = 1;
55 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
58 * max perf event sample rate
60 int sysctl_perf_event_sample_rate __read_mostly = 100000;
62 static atomic64_t perf_event_id;
64 void __weak perf_event_print_debug(void) { }
66 extern __weak const char *perf_pmu_name(void)
71 void perf_pmu_disable(struct pmu *pmu)
73 int *count = this_cpu_ptr(pmu->pmu_disable_count);
75 pmu->pmu_disable(pmu);
78 void perf_pmu_enable(struct pmu *pmu)
80 int *count = this_cpu_ptr(pmu->pmu_disable_count);
85 static DEFINE_PER_CPU(struct list_head, rotation_list);
88 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
89 * because they're strictly cpu affine and rotate_start is called with IRQs
90 * disabled, while rotate_context is called from IRQ context.
92 static void perf_pmu_rotate_start(struct pmu *pmu)
94 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
95 struct list_head *head = &__get_cpu_var(rotation_list);
97 WARN_ON(!irqs_disabled());
99 if (list_empty(&cpuctx->rotation_list))
100 list_add(&cpuctx->rotation_list, head);
103 static void get_ctx(struct perf_event_context *ctx)
105 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
108 static void free_ctx(struct rcu_head *head)
110 struct perf_event_context *ctx;
112 ctx = container_of(head, struct perf_event_context, rcu_head);
116 static void put_ctx(struct perf_event_context *ctx)
118 if (atomic_dec_and_test(&ctx->refcount)) {
120 put_ctx(ctx->parent_ctx);
122 put_task_struct(ctx->task);
123 call_rcu(&ctx->rcu_head, free_ctx);
127 static void unclone_ctx(struct perf_event_context *ctx)
129 if (ctx->parent_ctx) {
130 put_ctx(ctx->parent_ctx);
131 ctx->parent_ctx = NULL;
136 * If we inherit events we want to return the parent event id
139 static u64 primary_event_id(struct perf_event *event)
144 id = event->parent->id;
150 * Get the perf_event_context for a task and lock it.
151 * This has to cope with with the fact that until it is locked,
152 * the context could get moved to another task.
154 static struct perf_event_context *
155 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
157 struct perf_event_context *ctx;
161 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
164 * If this context is a clone of another, it might
165 * get swapped for another underneath us by
166 * perf_event_task_sched_out, though the
167 * rcu_read_lock() protects us from any context
168 * getting freed. Lock the context and check if it
169 * got swapped before we could get the lock, and retry
170 * if so. If we locked the right context, then it
171 * can't get swapped on us any more.
173 raw_spin_lock_irqsave(&ctx->lock, *flags);
174 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
175 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
179 if (!atomic_inc_not_zero(&ctx->refcount)) {
180 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
189 * Get the context for a task and increment its pin_count so it
190 * can't get swapped to another task. This also increments its
191 * reference count so that the context can't get freed.
193 static struct perf_event_context *
194 perf_pin_task_context(struct task_struct *task, int ctxn)
196 struct perf_event_context *ctx;
199 ctx = perf_lock_task_context(task, ctxn, &flags);
202 raw_spin_unlock_irqrestore(&ctx->lock, flags);
207 static void perf_unpin_context(struct perf_event_context *ctx)
211 raw_spin_lock_irqsave(&ctx->lock, flags);
213 raw_spin_unlock_irqrestore(&ctx->lock, flags);
217 static inline u64 perf_clock(void)
219 return local_clock();
223 * Update the record of the current time in a context.
225 static void update_context_time(struct perf_event_context *ctx)
227 u64 now = perf_clock();
229 ctx->time += now - ctx->timestamp;
230 ctx->timestamp = now;
234 * Update the total_time_enabled and total_time_running fields for a event.
236 static void update_event_times(struct perf_event *event)
238 struct perf_event_context *ctx = event->ctx;
241 if (event->state < PERF_EVENT_STATE_INACTIVE ||
242 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
248 run_end = event->tstamp_stopped;
250 event->total_time_enabled = run_end - event->tstamp_enabled;
252 if (event->state == PERF_EVENT_STATE_INACTIVE)
253 run_end = event->tstamp_stopped;
257 event->total_time_running = run_end - event->tstamp_running;
261 * Update total_time_enabled and total_time_running for all events in a group.
263 static void update_group_times(struct perf_event *leader)
265 struct perf_event *event;
267 update_event_times(leader);
268 list_for_each_entry(event, &leader->sibling_list, group_entry)
269 update_event_times(event);
272 static struct list_head *
273 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
275 if (event->attr.pinned)
276 return &ctx->pinned_groups;
278 return &ctx->flexible_groups;
282 * Add a event from the lists for its context.
283 * Must be called with ctx->mutex and ctx->lock held.
286 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
288 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
289 event->attach_state |= PERF_ATTACH_CONTEXT;
292 * If we're a stand alone event or group leader, we go to the context
293 * list, group events are kept attached to the group so that
294 * perf_group_detach can, at all times, locate all siblings.
296 if (event->group_leader == event) {
297 struct list_head *list;
299 if (is_software_event(event))
300 event->group_flags |= PERF_GROUP_SOFTWARE;
302 list = ctx_group_list(event, ctx);
303 list_add_tail(&event->group_entry, list);
306 list_add_rcu(&event->event_entry, &ctx->event_list);
308 perf_pmu_rotate_start(ctx->pmu);
310 if (event->attr.inherit_stat)
314 static void perf_group_attach(struct perf_event *event)
316 struct perf_event *group_leader = event->group_leader;
319 * We can have double attach due to group movement in perf_event_open.
321 if (event->attach_state & PERF_ATTACH_GROUP)
324 event->attach_state |= PERF_ATTACH_GROUP;
326 if (group_leader == event)
329 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
330 !is_software_event(event))
331 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
333 list_add_tail(&event->group_entry, &group_leader->sibling_list);
334 group_leader->nr_siblings++;
338 * Remove a event from the lists for its context.
339 * Must be called with ctx->mutex and ctx->lock held.
342 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
345 * We can have double detach due to exit/hot-unplug + close.
347 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
350 event->attach_state &= ~PERF_ATTACH_CONTEXT;
353 if (event->attr.inherit_stat)
356 list_del_rcu(&event->event_entry);
358 if (event->group_leader == event)
359 list_del_init(&event->group_entry);
361 update_group_times(event);
364 * If event was in error state, then keep it
365 * that way, otherwise bogus counts will be
366 * returned on read(). The only way to get out
367 * of error state is by explicit re-enabling
370 if (event->state > PERF_EVENT_STATE_OFF)
371 event->state = PERF_EVENT_STATE_OFF;
374 static void perf_group_detach(struct perf_event *event)
376 struct perf_event *sibling, *tmp;
377 struct list_head *list = NULL;
380 * We can have double detach due to exit/hot-unplug + close.
382 if (!(event->attach_state & PERF_ATTACH_GROUP))
385 event->attach_state &= ~PERF_ATTACH_GROUP;
388 * If this is a sibling, remove it from its group.
390 if (event->group_leader != event) {
391 list_del_init(&event->group_entry);
392 event->group_leader->nr_siblings--;
396 if (!list_empty(&event->group_entry))
397 list = &event->group_entry;
400 * If this was a group event with sibling events then
401 * upgrade the siblings to singleton events by adding them
402 * to whatever list we are on.
404 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
406 list_move_tail(&sibling->group_entry, list);
407 sibling->group_leader = sibling;
409 /* Inherit group flags from the previous leader */
410 sibling->group_flags = event->group_flags;
415 event_filter_match(struct perf_event *event)
417 return event->cpu == -1 || event->cpu == smp_processor_id();
421 __event_sched_out(struct perf_event *event,
422 struct perf_cpu_context *cpuctx,
423 struct perf_event_context *ctx)
427 * An event which could not be activated because of
428 * filter mismatch still needs to have its timings
429 * maintained, otherwise bogus information is return
430 * via read() for time_enabled, time_running:
432 if (event->state == PERF_EVENT_STATE_INACTIVE
433 && !event_filter_match(event)) {
434 delta = ctx->time - event->tstamp_stopped;
435 event->tstamp_running += delta;
436 event->tstamp_stopped = ctx->time;
439 if (event->state != PERF_EVENT_STATE_ACTIVE)
442 event->state = PERF_EVENT_STATE_INACTIVE;
443 if (event->pending_disable) {
444 event->pending_disable = 0;
445 event->state = PERF_EVENT_STATE_OFF;
447 event->pmu->del(event, 0);
450 if (!is_software_event(event))
451 cpuctx->active_oncpu--;
453 if (event->attr.exclusive || !cpuctx->active_oncpu)
454 cpuctx->exclusive = 0;
459 event_sched_out(struct perf_event *event,
460 struct perf_cpu_context *cpuctx,
461 struct perf_event_context *ctx)
465 ret = __event_sched_out(event, cpuctx, ctx);
467 event->tstamp_stopped = ctx->time;
471 group_sched_out(struct perf_event *group_event,
472 struct perf_cpu_context *cpuctx,
473 struct perf_event_context *ctx)
475 struct perf_event *event;
476 int state = group_event->state;
478 event_sched_out(group_event, cpuctx, ctx);
481 * Schedule out siblings (if any):
483 list_for_each_entry(event, &group_event->sibling_list, group_entry)
484 event_sched_out(event, cpuctx, ctx);
486 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
487 cpuctx->exclusive = 0;
490 static inline struct perf_cpu_context *
491 __get_cpu_context(struct perf_event_context *ctx)
493 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
497 * Cross CPU call to remove a performance event
499 * We disable the event on the hardware level first. After that we
500 * remove it from the context list.
502 static void __perf_event_remove_from_context(void *info)
504 struct perf_event *event = info;
505 struct perf_event_context *ctx = event->ctx;
506 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
509 * If this is a task context, we need to check whether it is
510 * the current task context of this cpu. If not it has been
511 * scheduled out before the smp call arrived.
513 if (ctx->task && cpuctx->task_ctx != ctx)
516 raw_spin_lock(&ctx->lock);
518 event_sched_out(event, cpuctx, ctx);
520 list_del_event(event, ctx);
522 raw_spin_unlock(&ctx->lock);
527 * Remove the event from a task's (or a CPU's) list of events.
529 * Must be called with ctx->mutex held.
531 * CPU events are removed with a smp call. For task events we only
532 * call when the task is on a CPU.
534 * If event->ctx is a cloned context, callers must make sure that
535 * every task struct that event->ctx->task could possibly point to
536 * remains valid. This is OK when called from perf_release since
537 * that only calls us on the top-level context, which can't be a clone.
538 * When called from perf_event_exit_task, it's OK because the
539 * context has been detached from its task.
541 static void perf_event_remove_from_context(struct perf_event *event)
543 struct perf_event_context *ctx = event->ctx;
544 struct task_struct *task = ctx->task;
548 * Per cpu events are removed via an smp call and
549 * the removal is always successful.
551 smp_call_function_single(event->cpu,
552 __perf_event_remove_from_context,
558 task_oncpu_function_call(task, __perf_event_remove_from_context,
561 raw_spin_lock_irq(&ctx->lock);
563 * If the context is active we need to retry the smp call.
565 if (ctx->nr_active && !list_empty(&event->group_entry)) {
566 raw_spin_unlock_irq(&ctx->lock);
571 * The lock prevents that this context is scheduled in so we
572 * can remove the event safely, if the call above did not
575 if (!list_empty(&event->group_entry))
576 list_del_event(event, ctx);
577 raw_spin_unlock_irq(&ctx->lock);
581 * Cross CPU call to disable a performance event
583 static void __perf_event_disable(void *info)
585 struct perf_event *event = info;
586 struct perf_event_context *ctx = event->ctx;
587 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
590 * If this is a per-task event, need to check whether this
591 * event's task is the current task on this cpu.
593 if (ctx->task && cpuctx->task_ctx != ctx)
596 raw_spin_lock(&ctx->lock);
599 * If the event is on, turn it off.
600 * If it is in error state, leave it in error state.
602 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
603 update_context_time(ctx);
604 update_group_times(event);
605 if (event == event->group_leader)
606 group_sched_out(event, cpuctx, ctx);
608 event_sched_out(event, cpuctx, ctx);
609 event->state = PERF_EVENT_STATE_OFF;
612 raw_spin_unlock(&ctx->lock);
618 * If event->ctx is a cloned context, callers must make sure that
619 * every task struct that event->ctx->task could possibly point to
620 * remains valid. This condition is satisifed when called through
621 * perf_event_for_each_child or perf_event_for_each because they
622 * hold the top-level event's child_mutex, so any descendant that
623 * goes to exit will block in sync_child_event.
624 * When called from perf_pending_event it's OK because event->ctx
625 * is the current context on this CPU and preemption is disabled,
626 * hence we can't get into perf_event_task_sched_out for this context.
628 void perf_event_disable(struct perf_event *event)
630 struct perf_event_context *ctx = event->ctx;
631 struct task_struct *task = ctx->task;
635 * Disable the event on the cpu that it's on
637 smp_call_function_single(event->cpu, __perf_event_disable,
643 task_oncpu_function_call(task, __perf_event_disable, event);
645 raw_spin_lock_irq(&ctx->lock);
647 * If the event is still active, we need to retry the cross-call.
649 if (event->state == PERF_EVENT_STATE_ACTIVE) {
650 raw_spin_unlock_irq(&ctx->lock);
655 * Since we have the lock this context can't be scheduled
656 * in, so we can change the state safely.
658 if (event->state == PERF_EVENT_STATE_INACTIVE) {
659 update_group_times(event);
660 event->state = PERF_EVENT_STATE_OFF;
663 raw_spin_unlock_irq(&ctx->lock);
667 __event_sched_in(struct perf_event *event,
668 struct perf_cpu_context *cpuctx,
669 struct perf_event_context *ctx)
671 if (event->state <= PERF_EVENT_STATE_OFF)
674 event->state = PERF_EVENT_STATE_ACTIVE;
675 event->oncpu = smp_processor_id();
677 * The new state must be visible before we turn it on in the hardware:
681 if (event->pmu->add(event, PERF_EF_START)) {
682 event->state = PERF_EVENT_STATE_INACTIVE;
687 if (!is_software_event(event))
688 cpuctx->active_oncpu++;
691 if (event->attr.exclusive)
692 cpuctx->exclusive = 1;
698 event_sched_in(struct perf_event *event,
699 struct perf_cpu_context *cpuctx,
700 struct perf_event_context *ctx)
702 int ret = __event_sched_in(event, cpuctx, ctx);
705 event->tstamp_running += ctx->time - event->tstamp_stopped;
710 group_commit_event_sched_in(struct perf_event *group_event,
711 struct perf_cpu_context *cpuctx,
712 struct perf_event_context *ctx)
714 struct perf_event *event;
717 group_event->tstamp_running += now - group_event->tstamp_stopped;
719 * Schedule in siblings as one group (if any):
721 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
722 event->tstamp_running += now - event->tstamp_stopped;
727 group_sched_in(struct perf_event *group_event,
728 struct perf_cpu_context *cpuctx,
729 struct perf_event_context *ctx)
731 struct perf_event *event, *partial_group = NULL;
732 struct pmu *pmu = group_event->pmu;
734 if (group_event->state == PERF_EVENT_STATE_OFF)
740 * use __event_sched_in() to delay updating tstamp_running
741 * until the transaction is committed. In case of failure
742 * we will keep an unmodified tstamp_running which is a
743 * requirement to get correct timing information
745 if (__event_sched_in(group_event, cpuctx, ctx)) {
746 pmu->cancel_txn(pmu);
751 * Schedule in siblings as one group (if any):
753 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
754 if (__event_sched_in(event, cpuctx, ctx)) {
755 partial_group = event;
760 if (!pmu->commit_txn(pmu)) {
761 /* commit tstamp_running */
762 group_commit_event_sched_in(group_event, cpuctx, ctx);
767 * Groups can be scheduled in as one unit only, so undo any
768 * partial group before returning:
770 * use __event_sched_out() to avoid updating tstamp_stopped
771 * because the event never actually ran
773 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
774 if (event == partial_group)
776 __event_sched_out(event, cpuctx, ctx);
778 __event_sched_out(group_event, cpuctx, ctx);
780 pmu->cancel_txn(pmu);
786 * Work out whether we can put this event group on the CPU now.
788 static int group_can_go_on(struct perf_event *event,
789 struct perf_cpu_context *cpuctx,
793 * Groups consisting entirely of software events can always go on.
795 if (event->group_flags & PERF_GROUP_SOFTWARE)
798 * If an exclusive group is already on, no other hardware
801 if (cpuctx->exclusive)
804 * If this group is exclusive and there are already
805 * events on the CPU, it can't go on.
807 if (event->attr.exclusive && cpuctx->active_oncpu)
810 * Otherwise, try to add it if all previous groups were able
816 static void add_event_to_ctx(struct perf_event *event,
817 struct perf_event_context *ctx)
819 list_add_event(event, ctx);
820 perf_group_attach(event);
821 event->tstamp_enabled = ctx->time;
822 event->tstamp_running = ctx->time;
823 event->tstamp_stopped = ctx->time;
827 * Cross CPU call to install and enable a performance event
829 * Must be called with ctx->mutex held
831 static void __perf_install_in_context(void *info)
833 struct perf_event *event = info;
834 struct perf_event_context *ctx = event->ctx;
835 struct perf_event *leader = event->group_leader;
836 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
840 * If this is a task context, we need to check whether it is
841 * the current task context of this cpu. If not it has been
842 * scheduled out before the smp call arrived.
843 * Or possibly this is the right context but it isn't
844 * on this cpu because it had no events.
846 if (ctx->task && cpuctx->task_ctx != ctx) {
847 if (cpuctx->task_ctx || ctx->task != current)
849 cpuctx->task_ctx = ctx;
852 raw_spin_lock(&ctx->lock);
854 update_context_time(ctx);
856 add_event_to_ctx(event, ctx);
858 if (event->cpu != -1 && event->cpu != smp_processor_id())
862 * Don't put the event on if it is disabled or if
863 * it is in a group and the group isn't on.
865 if (event->state != PERF_EVENT_STATE_INACTIVE ||
866 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
870 * An exclusive event can't go on if there are already active
871 * hardware events, and no hardware event can go on if there
872 * is already an exclusive event on.
874 if (!group_can_go_on(event, cpuctx, 1))
877 err = event_sched_in(event, cpuctx, ctx);
881 * This event couldn't go on. If it is in a group
882 * then we have to pull the whole group off.
883 * If the event group is pinned then put it in error state.
886 group_sched_out(leader, cpuctx, ctx);
887 if (leader->attr.pinned) {
888 update_group_times(leader);
889 leader->state = PERF_EVENT_STATE_ERROR;
894 raw_spin_unlock(&ctx->lock);
898 * Attach a performance event to a context
900 * First we add the event to the list with the hardware enable bit
901 * in event->hw_config cleared.
903 * If the event is attached to a task which is on a CPU we use a smp
904 * call to enable it in the task context. The task might have been
905 * scheduled away, but we check this in the smp call again.
907 * Must be called with ctx->mutex held.
910 perf_install_in_context(struct perf_event_context *ctx,
911 struct perf_event *event,
914 struct task_struct *task = ctx->task;
920 * Per cpu events are installed via an smp call and
921 * the install is always successful.
923 smp_call_function_single(cpu, __perf_install_in_context,
929 task_oncpu_function_call(task, __perf_install_in_context,
932 raw_spin_lock_irq(&ctx->lock);
934 * we need to retry the smp call.
936 if (ctx->is_active && list_empty(&event->group_entry)) {
937 raw_spin_unlock_irq(&ctx->lock);
942 * The lock prevents that this context is scheduled in so we
943 * can add the event safely, if it the call above did not
946 if (list_empty(&event->group_entry))
947 add_event_to_ctx(event, ctx);
948 raw_spin_unlock_irq(&ctx->lock);
952 * Put a event into inactive state and update time fields.
953 * Enabling the leader of a group effectively enables all
954 * the group members that aren't explicitly disabled, so we
955 * have to update their ->tstamp_enabled also.
956 * Note: this works for group members as well as group leaders
957 * since the non-leader members' sibling_lists will be empty.
959 static void __perf_event_mark_enabled(struct perf_event *event,
960 struct perf_event_context *ctx)
962 struct perf_event *sub;
964 event->state = PERF_EVENT_STATE_INACTIVE;
965 event->tstamp_enabled = ctx->time - event->total_time_enabled;
966 list_for_each_entry(sub, &event->sibling_list, group_entry) {
967 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
968 sub->tstamp_enabled =
969 ctx->time - sub->total_time_enabled;
975 * Cross CPU call to enable a performance event
977 static void __perf_event_enable(void *info)
979 struct perf_event *event = info;
980 struct perf_event_context *ctx = event->ctx;
981 struct perf_event *leader = event->group_leader;
982 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
986 * If this is a per-task event, need to check whether this
987 * event's task is the current task on this cpu.
989 if (ctx->task && cpuctx->task_ctx != ctx) {
990 if (cpuctx->task_ctx || ctx->task != current)
992 cpuctx->task_ctx = ctx;
995 raw_spin_lock(&ctx->lock);
997 update_context_time(ctx);
999 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1001 __perf_event_mark_enabled(event, ctx);
1003 if (event->cpu != -1 && event->cpu != smp_processor_id())
1007 * If the event is in a group and isn't the group leader,
1008 * then don't put it on unless the group is on.
1010 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1013 if (!group_can_go_on(event, cpuctx, 1)) {
1016 if (event == leader)
1017 err = group_sched_in(event, cpuctx, ctx);
1019 err = event_sched_in(event, cpuctx, ctx);
1024 * If this event can't go on and it's part of a
1025 * group, then the whole group has to come off.
1027 if (leader != event)
1028 group_sched_out(leader, cpuctx, ctx);
1029 if (leader->attr.pinned) {
1030 update_group_times(leader);
1031 leader->state = PERF_EVENT_STATE_ERROR;
1036 raw_spin_unlock(&ctx->lock);
1042 * If event->ctx is a cloned context, callers must make sure that
1043 * every task struct that event->ctx->task could possibly point to
1044 * remains valid. This condition is satisfied when called through
1045 * perf_event_for_each_child or perf_event_for_each as described
1046 * for perf_event_disable.
1048 void perf_event_enable(struct perf_event *event)
1050 struct perf_event_context *ctx = event->ctx;
1051 struct task_struct *task = ctx->task;
1055 * Enable the event on the cpu that it's on
1057 smp_call_function_single(event->cpu, __perf_event_enable,
1062 raw_spin_lock_irq(&ctx->lock);
1063 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1067 * If the event is in error state, clear that first.
1068 * That way, if we see the event in error state below, we
1069 * know that it has gone back into error state, as distinct
1070 * from the task having been scheduled away before the
1071 * cross-call arrived.
1073 if (event->state == PERF_EVENT_STATE_ERROR)
1074 event->state = PERF_EVENT_STATE_OFF;
1077 raw_spin_unlock_irq(&ctx->lock);
1078 task_oncpu_function_call(task, __perf_event_enable, event);
1080 raw_spin_lock_irq(&ctx->lock);
1083 * If the context is active and the event is still off,
1084 * we need to retry the cross-call.
1086 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1090 * Since we have the lock this context can't be scheduled
1091 * in, so we can change the state safely.
1093 if (event->state == PERF_EVENT_STATE_OFF)
1094 __perf_event_mark_enabled(event, ctx);
1097 raw_spin_unlock_irq(&ctx->lock);
1100 static int perf_event_refresh(struct perf_event *event, int refresh)
1103 * not supported on inherited events
1105 if (event->attr.inherit)
1108 atomic_add(refresh, &event->event_limit);
1109 perf_event_enable(event);
1115 EVENT_FLEXIBLE = 0x1,
1117 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1120 static void ctx_sched_out(struct perf_event_context *ctx,
1121 struct perf_cpu_context *cpuctx,
1122 enum event_type_t event_type)
1124 struct perf_event *event;
1126 raw_spin_lock(&ctx->lock);
1127 perf_pmu_disable(ctx->pmu);
1129 if (likely(!ctx->nr_events))
1131 update_context_time(ctx);
1133 if (!ctx->nr_active)
1136 if (event_type & EVENT_PINNED) {
1137 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1138 group_sched_out(event, cpuctx, ctx);
1141 if (event_type & EVENT_FLEXIBLE) {
1142 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1143 group_sched_out(event, cpuctx, ctx);
1146 perf_pmu_enable(ctx->pmu);
1147 raw_spin_unlock(&ctx->lock);
1151 * Test whether two contexts are equivalent, i.e. whether they
1152 * have both been cloned from the same version of the same context
1153 * and they both have the same number of enabled events.
1154 * If the number of enabled events is the same, then the set
1155 * of enabled events should be the same, because these are both
1156 * inherited contexts, therefore we can't access individual events
1157 * in them directly with an fd; we can only enable/disable all
1158 * events via prctl, or enable/disable all events in a family
1159 * via ioctl, which will have the same effect on both contexts.
1161 static int context_equiv(struct perf_event_context *ctx1,
1162 struct perf_event_context *ctx2)
1164 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1165 && ctx1->parent_gen == ctx2->parent_gen
1166 && !ctx1->pin_count && !ctx2->pin_count;
1169 static void __perf_event_sync_stat(struct perf_event *event,
1170 struct perf_event *next_event)
1174 if (!event->attr.inherit_stat)
1178 * Update the event value, we cannot use perf_event_read()
1179 * because we're in the middle of a context switch and have IRQs
1180 * disabled, which upsets smp_call_function_single(), however
1181 * we know the event must be on the current CPU, therefore we
1182 * don't need to use it.
1184 switch (event->state) {
1185 case PERF_EVENT_STATE_ACTIVE:
1186 event->pmu->read(event);
1189 case PERF_EVENT_STATE_INACTIVE:
1190 update_event_times(event);
1198 * In order to keep per-task stats reliable we need to flip the event
1199 * values when we flip the contexts.
1201 value = local64_read(&next_event->count);
1202 value = local64_xchg(&event->count, value);
1203 local64_set(&next_event->count, value);
1205 swap(event->total_time_enabled, next_event->total_time_enabled);
1206 swap(event->total_time_running, next_event->total_time_running);
1209 * Since we swizzled the values, update the user visible data too.
1211 perf_event_update_userpage(event);
1212 perf_event_update_userpage(next_event);
1215 #define list_next_entry(pos, member) \
1216 list_entry(pos->member.next, typeof(*pos), member)
1218 static void perf_event_sync_stat(struct perf_event_context *ctx,
1219 struct perf_event_context *next_ctx)
1221 struct perf_event *event, *next_event;
1226 update_context_time(ctx);
1228 event = list_first_entry(&ctx->event_list,
1229 struct perf_event, event_entry);
1231 next_event = list_first_entry(&next_ctx->event_list,
1232 struct perf_event, event_entry);
1234 while (&event->event_entry != &ctx->event_list &&
1235 &next_event->event_entry != &next_ctx->event_list) {
1237 __perf_event_sync_stat(event, next_event);
1239 event = list_next_entry(event, event_entry);
1240 next_event = list_next_entry(next_event, event_entry);
1244 void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1245 struct task_struct *next)
1247 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1248 struct perf_event_context *next_ctx;
1249 struct perf_event_context *parent;
1250 struct perf_cpu_context *cpuctx;
1256 cpuctx = __get_cpu_context(ctx);
1257 if (!cpuctx->task_ctx)
1261 parent = rcu_dereference(ctx->parent_ctx);
1262 next_ctx = next->perf_event_ctxp[ctxn];
1263 if (parent && next_ctx &&
1264 rcu_dereference(next_ctx->parent_ctx) == parent) {
1266 * Looks like the two contexts are clones, so we might be
1267 * able to optimize the context switch. We lock both
1268 * contexts and check that they are clones under the
1269 * lock (including re-checking that neither has been
1270 * uncloned in the meantime). It doesn't matter which
1271 * order we take the locks because no other cpu could
1272 * be trying to lock both of these tasks.
1274 raw_spin_lock(&ctx->lock);
1275 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1276 if (context_equiv(ctx, next_ctx)) {
1278 * XXX do we need a memory barrier of sorts
1279 * wrt to rcu_dereference() of perf_event_ctxp
1281 task->perf_event_ctxp[ctxn] = next_ctx;
1282 next->perf_event_ctxp[ctxn] = ctx;
1284 next_ctx->task = task;
1287 perf_event_sync_stat(ctx, next_ctx);
1289 raw_spin_unlock(&next_ctx->lock);
1290 raw_spin_unlock(&ctx->lock);
1295 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1296 cpuctx->task_ctx = NULL;
1300 #define for_each_task_context_nr(ctxn) \
1301 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1304 * Called from scheduler to remove the events of the current task,
1305 * with interrupts disabled.
1307 * We stop each event and update the event value in event->count.
1309 * This does not protect us against NMI, but disable()
1310 * sets the disabled bit in the control field of event _before_
1311 * accessing the event control register. If a NMI hits, then it will
1312 * not restart the event.
1314 void __perf_event_task_sched_out(struct task_struct *task,
1315 struct task_struct *next)
1319 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1321 for_each_task_context_nr(ctxn)
1322 perf_event_context_sched_out(task, ctxn, next);
1325 static void task_ctx_sched_out(struct perf_event_context *ctx,
1326 enum event_type_t event_type)
1328 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1330 if (!cpuctx->task_ctx)
1333 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1336 ctx_sched_out(ctx, cpuctx, event_type);
1337 cpuctx->task_ctx = NULL;
1341 * Called with IRQs disabled
1343 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1344 enum event_type_t event_type)
1346 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1350 ctx_pinned_sched_in(struct perf_event_context *ctx,
1351 struct perf_cpu_context *cpuctx)
1353 struct perf_event *event;
1355 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1356 if (event->state <= PERF_EVENT_STATE_OFF)
1358 if (event->cpu != -1 && event->cpu != smp_processor_id())
1361 if (group_can_go_on(event, cpuctx, 1))
1362 group_sched_in(event, cpuctx, ctx);
1365 * If this pinned group hasn't been scheduled,
1366 * put it in error state.
1368 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1369 update_group_times(event);
1370 event->state = PERF_EVENT_STATE_ERROR;
1376 ctx_flexible_sched_in(struct perf_event_context *ctx,
1377 struct perf_cpu_context *cpuctx)
1379 struct perf_event *event;
1382 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1383 /* Ignore events in OFF or ERROR state */
1384 if (event->state <= PERF_EVENT_STATE_OFF)
1387 * Listen to the 'cpu' scheduling filter constraint
1390 if (event->cpu != -1 && event->cpu != smp_processor_id())
1393 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1394 if (group_sched_in(event, cpuctx, ctx))
1401 ctx_sched_in(struct perf_event_context *ctx,
1402 struct perf_cpu_context *cpuctx,
1403 enum event_type_t event_type)
1405 raw_spin_lock(&ctx->lock);
1407 if (likely(!ctx->nr_events))
1410 ctx->timestamp = perf_clock();
1413 * First go through the list and put on any pinned groups
1414 * in order to give them the best chance of going on.
1416 if (event_type & EVENT_PINNED)
1417 ctx_pinned_sched_in(ctx, cpuctx);
1419 /* Then walk through the lower prio flexible groups */
1420 if (event_type & EVENT_FLEXIBLE)
1421 ctx_flexible_sched_in(ctx, cpuctx);
1424 raw_spin_unlock(&ctx->lock);
1427 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1428 enum event_type_t event_type)
1430 struct perf_event_context *ctx = &cpuctx->ctx;
1432 ctx_sched_in(ctx, cpuctx, event_type);
1435 static void task_ctx_sched_in(struct perf_event_context *ctx,
1436 enum event_type_t event_type)
1438 struct perf_cpu_context *cpuctx;
1440 cpuctx = __get_cpu_context(ctx);
1441 if (cpuctx->task_ctx == ctx)
1444 ctx_sched_in(ctx, cpuctx, event_type);
1445 cpuctx->task_ctx = ctx;
1448 void perf_event_context_sched_in(struct perf_event_context *ctx)
1450 struct perf_cpu_context *cpuctx;
1452 cpuctx = __get_cpu_context(ctx);
1453 if (cpuctx->task_ctx == ctx)
1456 perf_pmu_disable(ctx->pmu);
1458 * We want to keep the following priority order:
1459 * cpu pinned (that don't need to move), task pinned,
1460 * cpu flexible, task flexible.
1462 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1464 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1465 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1466 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1468 cpuctx->task_ctx = ctx;
1471 * Since these rotations are per-cpu, we need to ensure the
1472 * cpu-context we got scheduled on is actually rotating.
1474 perf_pmu_rotate_start(ctx->pmu);
1475 perf_pmu_enable(ctx->pmu);
1479 * Called from scheduler to add the events of the current task
1480 * with interrupts disabled.
1482 * We restore the event value and then enable it.
1484 * This does not protect us against NMI, but enable()
1485 * sets the enabled bit in the control field of event _before_
1486 * accessing the event control register. If a NMI hits, then it will
1487 * keep the event running.
1489 void __perf_event_task_sched_in(struct task_struct *task)
1491 struct perf_event_context *ctx;
1494 for_each_task_context_nr(ctxn) {
1495 ctx = task->perf_event_ctxp[ctxn];
1499 perf_event_context_sched_in(ctx);
1503 #define MAX_INTERRUPTS (~0ULL)
1505 static void perf_log_throttle(struct perf_event *event, int enable);
1507 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1509 u64 frequency = event->attr.sample_freq;
1510 u64 sec = NSEC_PER_SEC;
1511 u64 divisor, dividend;
1513 int count_fls, nsec_fls, frequency_fls, sec_fls;
1515 count_fls = fls64(count);
1516 nsec_fls = fls64(nsec);
1517 frequency_fls = fls64(frequency);
1521 * We got @count in @nsec, with a target of sample_freq HZ
1522 * the target period becomes:
1525 * period = -------------------
1526 * @nsec * sample_freq
1531 * Reduce accuracy by one bit such that @a and @b converge
1532 * to a similar magnitude.
1534 #define REDUCE_FLS(a, b) \
1536 if (a##_fls > b##_fls) { \
1546 * Reduce accuracy until either term fits in a u64, then proceed with
1547 * the other, so that finally we can do a u64/u64 division.
1549 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1550 REDUCE_FLS(nsec, frequency);
1551 REDUCE_FLS(sec, count);
1554 if (count_fls + sec_fls > 64) {
1555 divisor = nsec * frequency;
1557 while (count_fls + sec_fls > 64) {
1558 REDUCE_FLS(count, sec);
1562 dividend = count * sec;
1564 dividend = count * sec;
1566 while (nsec_fls + frequency_fls > 64) {
1567 REDUCE_FLS(nsec, frequency);
1571 divisor = nsec * frequency;
1577 return div64_u64(dividend, divisor);
1580 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1582 struct hw_perf_event *hwc = &event->hw;
1583 s64 period, sample_period;
1586 period = perf_calculate_period(event, nsec, count);
1588 delta = (s64)(period - hwc->sample_period);
1589 delta = (delta + 7) / 8; /* low pass filter */
1591 sample_period = hwc->sample_period + delta;
1596 hwc->sample_period = sample_period;
1598 if (local64_read(&hwc->period_left) > 8*sample_period) {
1599 event->pmu->stop(event, PERF_EF_UPDATE);
1600 local64_set(&hwc->period_left, 0);
1601 event->pmu->start(event, PERF_EF_RELOAD);
1605 static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1607 struct perf_event *event;
1608 struct hw_perf_event *hwc;
1609 u64 interrupts, now;
1612 raw_spin_lock(&ctx->lock);
1613 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1614 if (event->state != PERF_EVENT_STATE_ACTIVE)
1617 if (event->cpu != -1 && event->cpu != smp_processor_id())
1622 interrupts = hwc->interrupts;
1623 hwc->interrupts = 0;
1626 * unthrottle events on the tick
1628 if (interrupts == MAX_INTERRUPTS) {
1629 perf_log_throttle(event, 1);
1630 event->pmu->start(event, 0);
1633 if (!event->attr.freq || !event->attr.sample_freq)
1636 event->pmu->read(event);
1637 now = local64_read(&event->count);
1638 delta = now - hwc->freq_count_stamp;
1639 hwc->freq_count_stamp = now;
1642 perf_adjust_period(event, period, delta);
1644 raw_spin_unlock(&ctx->lock);
1648 * Round-robin a context's events:
1650 static void rotate_ctx(struct perf_event_context *ctx)
1652 raw_spin_lock(&ctx->lock);
1654 /* Rotate the first entry last of non-pinned groups */
1655 list_rotate_left(&ctx->flexible_groups);
1657 raw_spin_unlock(&ctx->lock);
1661 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1662 * because they're strictly cpu affine and rotate_start is called with IRQs
1663 * disabled, while rotate_context is called from IRQ context.
1665 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1667 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1668 struct perf_event_context *ctx = NULL;
1669 int rotate = 0, remove = 1;
1671 if (cpuctx->ctx.nr_events) {
1673 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1677 ctx = cpuctx->task_ctx;
1678 if (ctx && ctx->nr_events) {
1680 if (ctx->nr_events != ctx->nr_active)
1684 perf_pmu_disable(cpuctx->ctx.pmu);
1685 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1687 perf_ctx_adjust_freq(ctx, interval);
1692 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1694 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1696 rotate_ctx(&cpuctx->ctx);
1700 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1702 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1706 list_del_init(&cpuctx->rotation_list);
1708 perf_pmu_enable(cpuctx->ctx.pmu);
1711 void perf_event_task_tick(void)
1713 struct list_head *head = &__get_cpu_var(rotation_list);
1714 struct perf_cpu_context *cpuctx, *tmp;
1716 WARN_ON(!irqs_disabled());
1718 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1719 if (cpuctx->jiffies_interval == 1 ||
1720 !(jiffies % cpuctx->jiffies_interval))
1721 perf_rotate_context(cpuctx);
1725 static int event_enable_on_exec(struct perf_event *event,
1726 struct perf_event_context *ctx)
1728 if (!event->attr.enable_on_exec)
1731 event->attr.enable_on_exec = 0;
1732 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1735 __perf_event_mark_enabled(event, ctx);
1741 * Enable all of a task's events that have been marked enable-on-exec.
1742 * This expects task == current.
1744 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1746 struct perf_event *event;
1747 unsigned long flags;
1751 local_irq_save(flags);
1752 if (!ctx || !ctx->nr_events)
1755 task_ctx_sched_out(ctx, EVENT_ALL);
1757 raw_spin_lock(&ctx->lock);
1759 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1760 ret = event_enable_on_exec(event, ctx);
1765 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1766 ret = event_enable_on_exec(event, ctx);
1772 * Unclone this context if we enabled any event.
1777 raw_spin_unlock(&ctx->lock);
1779 perf_event_context_sched_in(ctx);
1781 local_irq_restore(flags);
1785 * Cross CPU call to read the hardware event
1787 static void __perf_event_read(void *info)
1789 struct perf_event *event = info;
1790 struct perf_event_context *ctx = event->ctx;
1791 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1794 * If this is a task context, we need to check whether it is
1795 * the current task context of this cpu. If not it has been
1796 * scheduled out before the smp call arrived. In that case
1797 * event->count would have been updated to a recent sample
1798 * when the event was scheduled out.
1800 if (ctx->task && cpuctx->task_ctx != ctx)
1803 raw_spin_lock(&ctx->lock);
1804 update_context_time(ctx);
1805 update_event_times(event);
1806 raw_spin_unlock(&ctx->lock);
1808 event->pmu->read(event);
1811 static inline u64 perf_event_count(struct perf_event *event)
1813 return local64_read(&event->count) + atomic64_read(&event->child_count);
1816 static u64 perf_event_read(struct perf_event *event)
1819 * If event is enabled and currently active on a CPU, update the
1820 * value in the event structure:
1822 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1823 smp_call_function_single(event->oncpu,
1824 __perf_event_read, event, 1);
1825 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1826 struct perf_event_context *ctx = event->ctx;
1827 unsigned long flags;
1829 raw_spin_lock_irqsave(&ctx->lock, flags);
1831 * may read while context is not active
1832 * (e.g., thread is blocked), in that case
1833 * we cannot update context time
1836 update_context_time(ctx);
1837 update_event_times(event);
1838 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1841 return perf_event_count(event);
1848 struct callchain_cpus_entries {
1849 struct rcu_head rcu_head;
1850 struct perf_callchain_entry *cpu_entries[0];
1853 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1854 static atomic_t nr_callchain_events;
1855 static DEFINE_MUTEX(callchain_mutex);
1856 struct callchain_cpus_entries *callchain_cpus_entries;
1859 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1860 struct pt_regs *regs)
1864 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
1865 struct pt_regs *regs)
1869 static void release_callchain_buffers_rcu(struct rcu_head *head)
1871 struct callchain_cpus_entries *entries;
1874 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1876 for_each_possible_cpu(cpu)
1877 kfree(entries->cpu_entries[cpu]);
1882 static void release_callchain_buffers(void)
1884 struct callchain_cpus_entries *entries;
1886 entries = callchain_cpus_entries;
1887 rcu_assign_pointer(callchain_cpus_entries, NULL);
1888 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1891 static int alloc_callchain_buffers(void)
1895 struct callchain_cpus_entries *entries;
1898 * We can't use the percpu allocation API for data that can be
1899 * accessed from NMI. Use a temporary manual per cpu allocation
1900 * until that gets sorted out.
1902 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1903 num_possible_cpus();
1905 entries = kzalloc(size, GFP_KERNEL);
1909 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1911 for_each_possible_cpu(cpu) {
1912 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1914 if (!entries->cpu_entries[cpu])
1918 rcu_assign_pointer(callchain_cpus_entries, entries);
1923 for_each_possible_cpu(cpu)
1924 kfree(entries->cpu_entries[cpu]);
1930 static int get_callchain_buffers(void)
1935 mutex_lock(&callchain_mutex);
1937 count = atomic_inc_return(&nr_callchain_events);
1938 if (WARN_ON_ONCE(count < 1)) {
1944 /* If the allocation failed, give up */
1945 if (!callchain_cpus_entries)
1950 err = alloc_callchain_buffers();
1952 release_callchain_buffers();
1954 mutex_unlock(&callchain_mutex);
1959 static void put_callchain_buffers(void)
1961 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1962 release_callchain_buffers();
1963 mutex_unlock(&callchain_mutex);
1967 static int get_recursion_context(int *recursion)
1975 else if (in_softirq())
1980 if (recursion[rctx])
1989 static inline void put_recursion_context(int *recursion, int rctx)
1995 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1998 struct callchain_cpus_entries *entries;
2000 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2004 entries = rcu_dereference(callchain_cpus_entries);
2008 cpu = smp_processor_id();
2010 return &entries->cpu_entries[cpu][*rctx];
2014 put_callchain_entry(int rctx)
2016 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2019 static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2022 struct perf_callchain_entry *entry;
2025 entry = get_callchain_entry(&rctx);
2034 if (!user_mode(regs)) {
2035 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2036 perf_callchain_kernel(entry, regs);
2038 regs = task_pt_regs(current);
2044 perf_callchain_store(entry, PERF_CONTEXT_USER);
2045 perf_callchain_user(entry, regs);
2049 put_callchain_entry(rctx);
2055 * Initialize the perf_event context in a task_struct:
2057 static void __perf_event_init_context(struct perf_event_context *ctx)
2059 raw_spin_lock_init(&ctx->lock);
2060 mutex_init(&ctx->mutex);
2061 INIT_LIST_HEAD(&ctx->pinned_groups);
2062 INIT_LIST_HEAD(&ctx->flexible_groups);
2063 INIT_LIST_HEAD(&ctx->event_list);
2064 atomic_set(&ctx->refcount, 1);
2067 static struct perf_event_context *
2068 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2070 struct perf_event_context *ctx;
2072 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2076 __perf_event_init_context(ctx);
2079 get_task_struct(task);
2086 static struct task_struct *
2087 find_lively_task_by_vpid(pid_t vpid)
2089 struct task_struct *task;
2096 task = find_task_by_vpid(vpid);
2098 get_task_struct(task);
2102 return ERR_PTR(-ESRCH);
2105 * Can't attach events to a dying task.
2108 if (task->flags & PF_EXITING)
2111 /* Reuse ptrace permission checks for now. */
2113 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2118 put_task_struct(task);
2119 return ERR_PTR(err);
2123 static struct perf_event_context *
2124 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2126 struct perf_event_context *ctx;
2127 struct perf_cpu_context *cpuctx;
2128 unsigned long flags;
2131 if (!task && cpu != -1) {
2132 /* Must be root to operate on a CPU event: */
2133 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2134 return ERR_PTR(-EACCES);
2136 if (cpu < 0 || cpu >= nr_cpumask_bits)
2137 return ERR_PTR(-EINVAL);
2140 * We could be clever and allow to attach a event to an
2141 * offline CPU and activate it when the CPU comes up, but
2144 if (!cpu_online(cpu))
2145 return ERR_PTR(-ENODEV);
2147 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2155 ctxn = pmu->task_ctx_nr;
2160 ctx = perf_lock_task_context(task, ctxn, &flags);
2163 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2167 ctx = alloc_perf_context(pmu, task);
2174 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
2176 * We raced with some other task; use
2177 * the context they set.
2179 put_task_struct(task);
2188 return ERR_PTR(err);
2191 static void perf_event_free_filter(struct perf_event *event);
2193 static void free_event_rcu(struct rcu_head *head)
2195 struct perf_event *event;
2197 event = container_of(head, struct perf_event, rcu_head);
2199 put_pid_ns(event->ns);
2200 perf_event_free_filter(event);
2204 static void perf_buffer_put(struct perf_buffer *buffer);
2206 static void free_event(struct perf_event *event)
2208 irq_work_sync(&event->pending);
2210 if (!event->parent) {
2211 if (event->attach_state & PERF_ATTACH_TASK)
2212 jump_label_dec(&perf_task_events);
2213 if (event->attr.mmap || event->attr.mmap_data)
2214 atomic_dec(&nr_mmap_events);
2215 if (event->attr.comm)
2216 atomic_dec(&nr_comm_events);
2217 if (event->attr.task)
2218 atomic_dec(&nr_task_events);
2219 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2220 put_callchain_buffers();
2223 if (event->buffer) {
2224 perf_buffer_put(event->buffer);
2225 event->buffer = NULL;
2229 event->destroy(event);
2232 put_ctx(event->ctx);
2234 call_rcu(&event->rcu_head, free_event_rcu);
2237 int perf_event_release_kernel(struct perf_event *event)
2239 struct perf_event_context *ctx = event->ctx;
2242 * Remove from the PMU, can't get re-enabled since we got
2243 * here because the last ref went.
2245 perf_event_disable(event);
2247 WARN_ON_ONCE(ctx->parent_ctx);
2249 * There are two ways this annotation is useful:
2251 * 1) there is a lock recursion from perf_event_exit_task
2252 * see the comment there.
2254 * 2) there is a lock-inversion with mmap_sem through
2255 * perf_event_read_group(), which takes faults while
2256 * holding ctx->mutex, however this is called after
2257 * the last filedesc died, so there is no possibility
2258 * to trigger the AB-BA case.
2260 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2261 raw_spin_lock_irq(&ctx->lock);
2262 perf_group_detach(event);
2263 list_del_event(event, ctx);
2264 raw_spin_unlock_irq(&ctx->lock);
2265 mutex_unlock(&ctx->mutex);
2267 mutex_lock(&event->owner->perf_event_mutex);
2268 list_del_init(&event->owner_entry);
2269 mutex_unlock(&event->owner->perf_event_mutex);
2270 put_task_struct(event->owner);
2276 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2279 * Called when the last reference to the file is gone.
2281 static int perf_release(struct inode *inode, struct file *file)
2283 struct perf_event *event = file->private_data;
2285 file->private_data = NULL;
2287 return perf_event_release_kernel(event);
2290 static int perf_event_read_size(struct perf_event *event)
2292 int entry = sizeof(u64); /* value */
2296 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2297 size += sizeof(u64);
2299 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2300 size += sizeof(u64);
2302 if (event->attr.read_format & PERF_FORMAT_ID)
2303 entry += sizeof(u64);
2305 if (event->attr.read_format & PERF_FORMAT_GROUP) {
2306 nr += event->group_leader->nr_siblings;
2307 size += sizeof(u64);
2315 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2317 struct perf_event *child;
2323 mutex_lock(&event->child_mutex);
2324 total += perf_event_read(event);
2325 *enabled += event->total_time_enabled +
2326 atomic64_read(&event->child_total_time_enabled);
2327 *running += event->total_time_running +
2328 atomic64_read(&event->child_total_time_running);
2330 list_for_each_entry(child, &event->child_list, child_list) {
2331 total += perf_event_read(child);
2332 *enabled += child->total_time_enabled;
2333 *running += child->total_time_running;
2335 mutex_unlock(&event->child_mutex);
2339 EXPORT_SYMBOL_GPL(perf_event_read_value);
2341 static int perf_event_read_group(struct perf_event *event,
2342 u64 read_format, char __user *buf)
2344 struct perf_event *leader = event->group_leader, *sub;
2345 int n = 0, size = 0, ret = -EFAULT;
2346 struct perf_event_context *ctx = leader->ctx;
2348 u64 count, enabled, running;
2350 mutex_lock(&ctx->mutex);
2351 count = perf_event_read_value(leader, &enabled, &running);
2353 values[n++] = 1 + leader->nr_siblings;
2354 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2355 values[n++] = enabled;
2356 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2357 values[n++] = running;
2358 values[n++] = count;
2359 if (read_format & PERF_FORMAT_ID)
2360 values[n++] = primary_event_id(leader);
2362 size = n * sizeof(u64);
2364 if (copy_to_user(buf, values, size))
2369 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2372 values[n++] = perf_event_read_value(sub, &enabled, &running);
2373 if (read_format & PERF_FORMAT_ID)
2374 values[n++] = primary_event_id(sub);
2376 size = n * sizeof(u64);
2378 if (copy_to_user(buf + ret, values, size)) {
2386 mutex_unlock(&ctx->mutex);
2391 static int perf_event_read_one(struct perf_event *event,
2392 u64 read_format, char __user *buf)
2394 u64 enabled, running;
2398 values[n++] = perf_event_read_value(event, &enabled, &running);
2399 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2400 values[n++] = enabled;
2401 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2402 values[n++] = running;
2403 if (read_format & PERF_FORMAT_ID)
2404 values[n++] = primary_event_id(event);
2406 if (copy_to_user(buf, values, n * sizeof(u64)))
2409 return n * sizeof(u64);
2413 * Read the performance event - simple non blocking version for now
2416 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2418 u64 read_format = event->attr.read_format;
2422 * Return end-of-file for a read on a event that is in
2423 * error state (i.e. because it was pinned but it couldn't be
2424 * scheduled on to the CPU at some point).
2426 if (event->state == PERF_EVENT_STATE_ERROR)
2429 if (count < perf_event_read_size(event))
2432 WARN_ON_ONCE(event->ctx->parent_ctx);
2433 if (read_format & PERF_FORMAT_GROUP)
2434 ret = perf_event_read_group(event, read_format, buf);
2436 ret = perf_event_read_one(event, read_format, buf);
2442 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2444 struct perf_event *event = file->private_data;
2446 return perf_read_hw(event, buf, count);
2449 static unsigned int perf_poll(struct file *file, poll_table *wait)
2451 struct perf_event *event = file->private_data;
2452 struct perf_buffer *buffer;
2453 unsigned int events = POLL_HUP;
2456 buffer = rcu_dereference(event->buffer);
2458 events = atomic_xchg(&buffer->poll, 0);
2461 poll_wait(file, &event->waitq, wait);
2466 static void perf_event_reset(struct perf_event *event)
2468 (void)perf_event_read(event);
2469 local64_set(&event->count, 0);
2470 perf_event_update_userpage(event);
2474 * Holding the top-level event's child_mutex means that any
2475 * descendant process that has inherited this event will block
2476 * in sync_child_event if it goes to exit, thus satisfying the
2477 * task existence requirements of perf_event_enable/disable.
2479 static void perf_event_for_each_child(struct perf_event *event,
2480 void (*func)(struct perf_event *))
2482 struct perf_event *child;
2484 WARN_ON_ONCE(event->ctx->parent_ctx);
2485 mutex_lock(&event->child_mutex);
2487 list_for_each_entry(child, &event->child_list, child_list)
2489 mutex_unlock(&event->child_mutex);
2492 static void perf_event_for_each(struct perf_event *event,
2493 void (*func)(struct perf_event *))
2495 struct perf_event_context *ctx = event->ctx;
2496 struct perf_event *sibling;
2498 WARN_ON_ONCE(ctx->parent_ctx);
2499 mutex_lock(&ctx->mutex);
2500 event = event->group_leader;
2502 perf_event_for_each_child(event, func);
2504 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2505 perf_event_for_each_child(event, func);
2506 mutex_unlock(&ctx->mutex);
2509 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2511 struct perf_event_context *ctx = event->ctx;
2516 if (!event->attr.sample_period)
2519 size = copy_from_user(&value, arg, sizeof(value));
2520 if (size != sizeof(value))
2526 raw_spin_lock_irq(&ctx->lock);
2527 if (event->attr.freq) {
2528 if (value > sysctl_perf_event_sample_rate) {
2533 event->attr.sample_freq = value;
2535 event->attr.sample_period = value;
2536 event->hw.sample_period = value;
2539 raw_spin_unlock_irq(&ctx->lock);
2544 static const struct file_operations perf_fops;
2546 static struct perf_event *perf_fget_light(int fd, int *fput_needed)
2550 file = fget_light(fd, fput_needed);
2552 return ERR_PTR(-EBADF);
2554 if (file->f_op != &perf_fops) {
2555 fput_light(file, *fput_needed);
2557 return ERR_PTR(-EBADF);
2560 return file->private_data;
2563 static int perf_event_set_output(struct perf_event *event,
2564 struct perf_event *output_event);
2565 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2567 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2569 struct perf_event *event = file->private_data;
2570 void (*func)(struct perf_event *);
2574 case PERF_EVENT_IOC_ENABLE:
2575 func = perf_event_enable;
2577 case PERF_EVENT_IOC_DISABLE:
2578 func = perf_event_disable;
2580 case PERF_EVENT_IOC_RESET:
2581 func = perf_event_reset;
2584 case PERF_EVENT_IOC_REFRESH:
2585 return perf_event_refresh(event, arg);
2587 case PERF_EVENT_IOC_PERIOD:
2588 return perf_event_period(event, (u64 __user *)arg);
2590 case PERF_EVENT_IOC_SET_OUTPUT:
2592 struct perf_event *output_event = NULL;
2593 int fput_needed = 0;
2597 output_event = perf_fget_light(arg, &fput_needed);
2598 if (IS_ERR(output_event))
2599 return PTR_ERR(output_event);
2602 ret = perf_event_set_output(event, output_event);
2604 fput_light(output_event->filp, fput_needed);
2609 case PERF_EVENT_IOC_SET_FILTER:
2610 return perf_event_set_filter(event, (void __user *)arg);
2616 if (flags & PERF_IOC_FLAG_GROUP)
2617 perf_event_for_each(event, func);
2619 perf_event_for_each_child(event, func);
2624 int perf_event_task_enable(void)
2626 struct perf_event *event;
2628 mutex_lock(¤t->perf_event_mutex);
2629 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2630 perf_event_for_each_child(event, perf_event_enable);
2631 mutex_unlock(¤t->perf_event_mutex);
2636 int perf_event_task_disable(void)
2638 struct perf_event *event;
2640 mutex_lock(¤t->perf_event_mutex);
2641 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2642 perf_event_for_each_child(event, perf_event_disable);
2643 mutex_unlock(¤t->perf_event_mutex);
2648 #ifndef PERF_EVENT_INDEX_OFFSET
2649 # define PERF_EVENT_INDEX_OFFSET 0
2652 static int perf_event_index(struct perf_event *event)
2654 if (event->hw.state & PERF_HES_STOPPED)
2657 if (event->state != PERF_EVENT_STATE_ACTIVE)
2660 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2664 * Callers need to ensure there can be no nesting of this function, otherwise
2665 * the seqlock logic goes bad. We can not serialize this because the arch
2666 * code calls this from NMI context.
2668 void perf_event_update_userpage(struct perf_event *event)
2670 struct perf_event_mmap_page *userpg;
2671 struct perf_buffer *buffer;
2674 buffer = rcu_dereference(event->buffer);
2678 userpg = buffer->user_page;
2681 * Disable preemption so as to not let the corresponding user-space
2682 * spin too long if we get preempted.
2687 userpg->index = perf_event_index(event);
2688 userpg->offset = perf_event_count(event);
2689 if (event->state == PERF_EVENT_STATE_ACTIVE)
2690 userpg->offset -= local64_read(&event->hw.prev_count);
2692 userpg->time_enabled = event->total_time_enabled +
2693 atomic64_read(&event->child_total_time_enabled);
2695 userpg->time_running = event->total_time_running +
2696 atomic64_read(&event->child_total_time_running);
2705 static unsigned long perf_data_size(struct perf_buffer *buffer);
2708 perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
2710 long max_size = perf_data_size(buffer);
2713 buffer->watermark = min(max_size, watermark);
2715 if (!buffer->watermark)
2716 buffer->watermark = max_size / 2;
2718 if (flags & PERF_BUFFER_WRITABLE)
2719 buffer->writable = 1;
2721 atomic_set(&buffer->refcount, 1);
2724 #ifndef CONFIG_PERF_USE_VMALLOC
2727 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2730 static struct page *
2731 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2733 if (pgoff > buffer->nr_pages)
2737 return virt_to_page(buffer->user_page);
2739 return virt_to_page(buffer->data_pages[pgoff - 1]);
2742 static void *perf_mmap_alloc_page(int cpu)
2747 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
2748 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
2752 return page_address(page);
2755 static struct perf_buffer *
2756 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2758 struct perf_buffer *buffer;
2762 size = sizeof(struct perf_buffer);
2763 size += nr_pages * sizeof(void *);
2765 buffer = kzalloc(size, GFP_KERNEL);
2769 buffer->user_page = perf_mmap_alloc_page(cpu);
2770 if (!buffer->user_page)
2771 goto fail_user_page;
2773 for (i = 0; i < nr_pages; i++) {
2774 buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
2775 if (!buffer->data_pages[i])
2776 goto fail_data_pages;
2779 buffer->nr_pages = nr_pages;
2781 perf_buffer_init(buffer, watermark, flags);
2786 for (i--; i >= 0; i--)
2787 free_page((unsigned long)buffer->data_pages[i]);
2789 free_page((unsigned long)buffer->user_page);
2798 static void perf_mmap_free_page(unsigned long addr)
2800 struct page *page = virt_to_page((void *)addr);
2802 page->mapping = NULL;
2806 static void perf_buffer_free(struct perf_buffer *buffer)
2810 perf_mmap_free_page((unsigned long)buffer->user_page);
2811 for (i = 0; i < buffer->nr_pages; i++)
2812 perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
2816 static inline int page_order(struct perf_buffer *buffer)
2824 * Back perf_mmap() with vmalloc memory.
2826 * Required for architectures that have d-cache aliasing issues.
2829 static inline int page_order(struct perf_buffer *buffer)
2831 return buffer->page_order;
2834 static struct page *
2835 perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
2837 if (pgoff > (1UL << page_order(buffer)))
2840 return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
2843 static void perf_mmap_unmark_page(void *addr)
2845 struct page *page = vmalloc_to_page(addr);
2847 page->mapping = NULL;
2850 static void perf_buffer_free_work(struct work_struct *work)
2852 struct perf_buffer *buffer;
2856 buffer = container_of(work, struct perf_buffer, work);
2857 nr = 1 << page_order(buffer);
2859 base = buffer->user_page;
2860 for (i = 0; i < nr + 1; i++)
2861 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2867 static void perf_buffer_free(struct perf_buffer *buffer)
2869 schedule_work(&buffer->work);
2872 static struct perf_buffer *
2873 perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
2875 struct perf_buffer *buffer;
2879 size = sizeof(struct perf_buffer);
2880 size += sizeof(void *);
2882 buffer = kzalloc(size, GFP_KERNEL);
2886 INIT_WORK(&buffer->work, perf_buffer_free_work);
2888 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2892 buffer->user_page = all_buf;
2893 buffer->data_pages[0] = all_buf + PAGE_SIZE;
2894 buffer->page_order = ilog2(nr_pages);
2895 buffer->nr_pages = 1;
2897 perf_buffer_init(buffer, watermark, flags);
2910 static unsigned long perf_data_size(struct perf_buffer *buffer)
2912 return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
2915 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2917 struct perf_event *event = vma->vm_file->private_data;
2918 struct perf_buffer *buffer;
2919 int ret = VM_FAULT_SIGBUS;
2921 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2922 if (vmf->pgoff == 0)
2928 buffer = rcu_dereference(event->buffer);
2932 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2935 vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
2939 get_page(vmf->page);
2940 vmf->page->mapping = vma->vm_file->f_mapping;
2941 vmf->page->index = vmf->pgoff;
2950 static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
2952 struct perf_buffer *buffer;
2954 buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
2955 perf_buffer_free(buffer);
2958 static struct perf_buffer *perf_buffer_get(struct perf_event *event)
2960 struct perf_buffer *buffer;
2963 buffer = rcu_dereference(event->buffer);
2965 if (!atomic_inc_not_zero(&buffer->refcount))
2973 static void perf_buffer_put(struct perf_buffer *buffer)
2975 if (!atomic_dec_and_test(&buffer->refcount))
2978 call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
2981 static void perf_mmap_open(struct vm_area_struct *vma)
2983 struct perf_event *event = vma->vm_file->private_data;
2985 atomic_inc(&event->mmap_count);
2988 static void perf_mmap_close(struct vm_area_struct *vma)
2990 struct perf_event *event = vma->vm_file->private_data;
2992 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2993 unsigned long size = perf_data_size(event->buffer);
2994 struct user_struct *user = event->mmap_user;
2995 struct perf_buffer *buffer = event->buffer;
2997 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2998 vma->vm_mm->locked_vm -= event->mmap_locked;
2999 rcu_assign_pointer(event->buffer, NULL);
3000 mutex_unlock(&event->mmap_mutex);
3002 perf_buffer_put(buffer);
3007 static const struct vm_operations_struct perf_mmap_vmops = {
3008 .open = perf_mmap_open,
3009 .close = perf_mmap_close,
3010 .fault = perf_mmap_fault,
3011 .page_mkwrite = perf_mmap_fault,
3014 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3016 struct perf_event *event = file->private_data;
3017 unsigned long user_locked, user_lock_limit;
3018 struct user_struct *user = current_user();
3019 unsigned long locked, lock_limit;
3020 struct perf_buffer *buffer;
3021 unsigned long vma_size;
3022 unsigned long nr_pages;
3023 long user_extra, extra;
3024 int ret = 0, flags = 0;
3027 * Don't allow mmap() of inherited per-task counters. This would
3028 * create a performance issue due to all children writing to the
3031 if (event->cpu == -1 && event->attr.inherit)
3034 if (!(vma->vm_flags & VM_SHARED))
3037 vma_size = vma->vm_end - vma->vm_start;
3038 nr_pages = (vma_size / PAGE_SIZE) - 1;
3041 * If we have buffer pages ensure they're a power-of-two number, so we
3042 * can do bitmasks instead of modulo.
3044 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3047 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3050 if (vma->vm_pgoff != 0)
3053 WARN_ON_ONCE(event->ctx->parent_ctx);
3054 mutex_lock(&event->mmap_mutex);
3055 if (event->buffer) {
3056 if (event->buffer->nr_pages == nr_pages)
3057 atomic_inc(&event->buffer->refcount);
3063 user_extra = nr_pages + 1;
3064 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3067 * Increase the limit linearly with more CPUs:
3069 user_lock_limit *= num_online_cpus();
3071 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3074 if (user_locked > user_lock_limit)
3075 extra = user_locked - user_lock_limit;
3077 lock_limit = rlimit(RLIMIT_MEMLOCK);
3078 lock_limit >>= PAGE_SHIFT;
3079 locked = vma->vm_mm->locked_vm + extra;
3081 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3082 !capable(CAP_IPC_LOCK)) {
3087 WARN_ON(event->buffer);
3089 if (vma->vm_flags & VM_WRITE)
3090 flags |= PERF_BUFFER_WRITABLE;
3092 buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
3098 rcu_assign_pointer(event->buffer, buffer);
3100 atomic_long_add(user_extra, &user->locked_vm);
3101 event->mmap_locked = extra;
3102 event->mmap_user = get_current_user();
3103 vma->vm_mm->locked_vm += event->mmap_locked;
3107 atomic_inc(&event->mmap_count);
3108 mutex_unlock(&event->mmap_mutex);
3110 vma->vm_flags |= VM_RESERVED;
3111 vma->vm_ops = &perf_mmap_vmops;
3116 static int perf_fasync(int fd, struct file *filp, int on)
3118 struct inode *inode = filp->f_path.dentry->d_inode;
3119 struct perf_event *event = filp->private_data;
3122 mutex_lock(&inode->i_mutex);
3123 retval = fasync_helper(fd, filp, on, &event->fasync);
3124 mutex_unlock(&inode->i_mutex);
3132 static const struct file_operations perf_fops = {
3133 .llseek = no_llseek,
3134 .release = perf_release,
3137 .unlocked_ioctl = perf_ioctl,
3138 .compat_ioctl = perf_ioctl,
3140 .fasync = perf_fasync,
3146 * If there's data, ensure we set the poll() state and publish everything
3147 * to user-space before waking everybody up.
3150 void perf_event_wakeup(struct perf_event *event)
3152 wake_up_all(&event->waitq);
3154 if (event->pending_kill) {
3155 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3156 event->pending_kill = 0;
3160 static void perf_pending_event(struct irq_work *entry)
3162 struct perf_event *event = container_of(entry,
3163 struct perf_event, pending);
3165 if (event->pending_disable) {
3166 event->pending_disable = 0;
3167 __perf_event_disable(event);
3170 if (event->pending_wakeup) {
3171 event->pending_wakeup = 0;
3172 perf_event_wakeup(event);
3177 * We assume there is only KVM supporting the callbacks.
3178 * Later on, we might change it to a list if there is
3179 * another virtualization implementation supporting the callbacks.
3181 struct perf_guest_info_callbacks *perf_guest_cbs;
3183 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3185 perf_guest_cbs = cbs;
3188 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3190 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3192 perf_guest_cbs = NULL;
3195 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3200 static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
3201 unsigned long offset, unsigned long head)
3205 if (!buffer->writable)
3208 mask = perf_data_size(buffer) - 1;
3210 offset = (offset - tail) & mask;
3211 head = (head - tail) & mask;
3213 if ((int)(head - offset) < 0)
3219 static void perf_output_wakeup(struct perf_output_handle *handle)
3221 atomic_set(&handle->buffer->poll, POLL_IN);
3224 handle->event->pending_wakeup = 1;
3225 irq_work_queue(&handle->event->pending);
3227 perf_event_wakeup(handle->event);
3231 * We need to ensure a later event_id doesn't publish a head when a former
3232 * event isn't done writing. However since we need to deal with NMIs we
3233 * cannot fully serialize things.
3235 * We only publish the head (and generate a wakeup) when the outer-most
3238 static void perf_output_get_handle(struct perf_output_handle *handle)
3240 struct perf_buffer *buffer = handle->buffer;
3243 local_inc(&buffer->nest);
3244 handle->wakeup = local_read(&buffer->wakeup);
3247 static void perf_output_put_handle(struct perf_output_handle *handle)
3249 struct perf_buffer *buffer = handle->buffer;
3253 head = local_read(&buffer->head);
3256 * IRQ/NMI can happen here, which means we can miss a head update.
3259 if (!local_dec_and_test(&buffer->nest))
3263 * Publish the known good head. Rely on the full barrier implied
3264 * by atomic_dec_and_test() order the buffer->head read and this
3267 buffer->user_page->data_head = head;
3270 * Now check if we missed an update, rely on the (compiler)
3271 * barrier in atomic_dec_and_test() to re-read buffer->head.
3273 if (unlikely(head != local_read(&buffer->head))) {
3274 local_inc(&buffer->nest);
3278 if (handle->wakeup != local_read(&buffer->wakeup))
3279 perf_output_wakeup(handle);
3285 __always_inline void perf_output_copy(struct perf_output_handle *handle,
3286 const void *buf, unsigned int len)
3289 unsigned long size = min_t(unsigned long, handle->size, len);
3291 memcpy(handle->addr, buf, size);
3294 handle->addr += size;
3296 handle->size -= size;
3297 if (!handle->size) {
3298 struct perf_buffer *buffer = handle->buffer;
3301 handle->page &= buffer->nr_pages - 1;
3302 handle->addr = buffer->data_pages[handle->page];
3303 handle->size = PAGE_SIZE << page_order(buffer);
3308 int perf_output_begin(struct perf_output_handle *handle,
3309 struct perf_event *event, unsigned int size,
3310 int nmi, int sample)
3312 struct perf_buffer *buffer;
3313 unsigned long tail, offset, head;
3316 struct perf_event_header header;
3323 * For inherited events we send all the output towards the parent.
3326 event = event->parent;
3328 buffer = rcu_dereference(event->buffer);
3332 handle->buffer = buffer;
3333 handle->event = event;
3335 handle->sample = sample;
3337 if (!buffer->nr_pages)
3340 have_lost = local_read(&buffer->lost);
3342 size += sizeof(lost_event);
3344 perf_output_get_handle(handle);
3348 * Userspace could choose to issue a mb() before updating the
3349 * tail pointer. So that all reads will be completed before the
3352 tail = ACCESS_ONCE(buffer->user_page->data_tail);
3354 offset = head = local_read(&buffer->head);
3356 if (unlikely(!perf_output_space(buffer, tail, offset, head)))
3358 } while (local_cmpxchg(&buffer->head, offset, head) != offset);
3360 if (head - local_read(&buffer->wakeup) > buffer->watermark)
3361 local_add(buffer->watermark, &buffer->wakeup);
3363 handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
3364 handle->page &= buffer->nr_pages - 1;
3365 handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
3366 handle->addr = buffer->data_pages[handle->page];
3367 handle->addr += handle->size;
3368 handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
3371 lost_event.header.type = PERF_RECORD_LOST;
3372 lost_event.header.misc = 0;
3373 lost_event.header.size = sizeof(lost_event);
3374 lost_event.id = event->id;
3375 lost_event.lost = local_xchg(&buffer->lost, 0);
3377 perf_output_put(handle, lost_event);
3383 local_inc(&buffer->lost);
3384 perf_output_put_handle(handle);
3391 void perf_output_end(struct perf_output_handle *handle)
3393 struct perf_event *event = handle->event;
3394 struct perf_buffer *buffer = handle->buffer;
3396 int wakeup_events = event->attr.wakeup_events;
3398 if (handle->sample && wakeup_events) {
3399 int events = local_inc_return(&buffer->events);
3400 if (events >= wakeup_events) {
3401 local_sub(wakeup_events, &buffer->events);
3402 local_inc(&buffer->wakeup);
3406 perf_output_put_handle(handle);
3410 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3413 * only top level events have the pid namespace they were created in
3416 event = event->parent;
3418 return task_tgid_nr_ns(p, event->ns);
3421 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3424 * only top level events have the pid namespace they were created in
3427 event = event->parent;
3429 return task_pid_nr_ns(p, event->ns);
3432 static void perf_output_read_one(struct perf_output_handle *handle,
3433 struct perf_event *event)
3435 u64 read_format = event->attr.read_format;
3439 values[n++] = perf_event_count(event);
3440 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3441 values[n++] = event->total_time_enabled +
3442 atomic64_read(&event->child_total_time_enabled);
3444 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3445 values[n++] = event->total_time_running +
3446 atomic64_read(&event->child_total_time_running);
3448 if (read_format & PERF_FORMAT_ID)
3449 values[n++] = primary_event_id(event);
3451 perf_output_copy(handle, values, n * sizeof(u64));
3455 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3457 static void perf_output_read_group(struct perf_output_handle *handle,
3458 struct perf_event *event)
3460 struct perf_event *leader = event->group_leader, *sub;
3461 u64 read_format = event->attr.read_format;
3465 values[n++] = 1 + leader->nr_siblings;
3467 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3468 values[n++] = leader->total_time_enabled;
3470 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3471 values[n++] = leader->total_time_running;
3473 if (leader != event)
3474 leader->pmu->read(leader);
3476 values[n++] = perf_event_count(leader);
3477 if (read_format & PERF_FORMAT_ID)
3478 values[n++] = primary_event_id(leader);
3480 perf_output_copy(handle, values, n * sizeof(u64));
3482 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3486 sub->pmu->read(sub);
3488 values[n++] = perf_event_count(sub);
3489 if (read_format & PERF_FORMAT_ID)
3490 values[n++] = primary_event_id(sub);
3492 perf_output_copy(handle, values, n * sizeof(u64));
3496 static void perf_output_read(struct perf_output_handle *handle,
3497 struct perf_event *event)
3499 if (event->attr.read_format & PERF_FORMAT_GROUP)
3500 perf_output_read_group(handle, event);
3502 perf_output_read_one(handle, event);
3505 void perf_output_sample(struct perf_output_handle *handle,
3506 struct perf_event_header *header,
3507 struct perf_sample_data *data,
3508 struct perf_event *event)
3510 u64 sample_type = data->type;
3512 perf_output_put(handle, *header);
3514 if (sample_type & PERF_SAMPLE_IP)
3515 perf_output_put(handle, data->ip);
3517 if (sample_type & PERF_SAMPLE_TID)
3518 perf_output_put(handle, data->tid_entry);
3520 if (sample_type & PERF_SAMPLE_TIME)
3521 perf_output_put(handle, data->time);
3523 if (sample_type & PERF_SAMPLE_ADDR)
3524 perf_output_put(handle, data->addr);
3526 if (sample_type & PERF_SAMPLE_ID)
3527 perf_output_put(handle, data->id);
3529 if (sample_type & PERF_SAMPLE_STREAM_ID)
3530 perf_output_put(handle, data->stream_id);
3532 if (sample_type & PERF_SAMPLE_CPU)
3533 perf_output_put(handle, data->cpu_entry);
3535 if (sample_type & PERF_SAMPLE_PERIOD)
3536 perf_output_put(handle, data->period);
3538 if (sample_type & PERF_SAMPLE_READ)
3539 perf_output_read(handle, event);
3541 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3542 if (data->callchain) {
3545 if (data->callchain)
3546 size += data->callchain->nr;
3548 size *= sizeof(u64);
3550 perf_output_copy(handle, data->callchain, size);
3553 perf_output_put(handle, nr);
3557 if (sample_type & PERF_SAMPLE_RAW) {
3559 perf_output_put(handle, data->raw->size);
3560 perf_output_copy(handle, data->raw->data,
3567 .size = sizeof(u32),
3570 perf_output_put(handle, raw);
3575 void perf_prepare_sample(struct perf_event_header *header,
3576 struct perf_sample_data *data,
3577 struct perf_event *event,
3578 struct pt_regs *regs)
3580 u64 sample_type = event->attr.sample_type;
3582 data->type = sample_type;
3584 header->type = PERF_RECORD_SAMPLE;
3585 header->size = sizeof(*header);
3588 header->misc |= perf_misc_flags(regs);
3590 if (sample_type & PERF_SAMPLE_IP) {
3591 data->ip = perf_instruction_pointer(regs);
3593 header->size += sizeof(data->ip);
3596 if (sample_type & PERF_SAMPLE_TID) {
3597 /* namespace issues */
3598 data->tid_entry.pid = perf_event_pid(event, current);
3599 data->tid_entry.tid = perf_event_tid(event, current);
3601 header->size += sizeof(data->tid_entry);
3604 if (sample_type & PERF_SAMPLE_TIME) {
3605 data->time = perf_clock();
3607 header->size += sizeof(data->time);
3610 if (sample_type & PERF_SAMPLE_ADDR)
3611 header->size += sizeof(data->addr);
3613 if (sample_type & PERF_SAMPLE_ID) {
3614 data->id = primary_event_id(event);
3616 header->size += sizeof(data->id);
3619 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3620 data->stream_id = event->id;
3622 header->size += sizeof(data->stream_id);
3625 if (sample_type & PERF_SAMPLE_CPU) {
3626 data->cpu_entry.cpu = raw_smp_processor_id();
3627 data->cpu_entry.reserved = 0;
3629 header->size += sizeof(data->cpu_entry);
3632 if (sample_type & PERF_SAMPLE_PERIOD)
3633 header->size += sizeof(data->period);
3635 if (sample_type & PERF_SAMPLE_READ)
3636 header->size += perf_event_read_size(event);
3638 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3641 data->callchain = perf_callchain(regs);
3643 if (data->callchain)
3644 size += data->callchain->nr;
3646 header->size += size * sizeof(u64);
3649 if (sample_type & PERF_SAMPLE_RAW) {
3650 int size = sizeof(u32);
3653 size += data->raw->size;
3655 size += sizeof(u32);
3657 WARN_ON_ONCE(size & (sizeof(u64)-1));
3658 header->size += size;
3662 static void perf_event_output(struct perf_event *event, int nmi,
3663 struct perf_sample_data *data,
3664 struct pt_regs *regs)
3666 struct perf_output_handle handle;
3667 struct perf_event_header header;
3669 /* protect the callchain buffers */
3672 perf_prepare_sample(&header, data, event, regs);
3674 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3677 perf_output_sample(&handle, &header, data, event);
3679 perf_output_end(&handle);
3689 struct perf_read_event {
3690 struct perf_event_header header;
3697 perf_event_read_event(struct perf_event *event,
3698 struct task_struct *task)
3700 struct perf_output_handle handle;
3701 struct perf_read_event read_event = {
3703 .type = PERF_RECORD_READ,
3705 .size = sizeof(read_event) + perf_event_read_size(event),
3707 .pid = perf_event_pid(event, task),
3708 .tid = perf_event_tid(event, task),
3712 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3716 perf_output_put(&handle, read_event);
3717 perf_output_read(&handle, event);
3719 perf_output_end(&handle);
3723 * task tracking -- fork/exit
3725 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
3728 struct perf_task_event {
3729 struct task_struct *task;
3730 struct perf_event_context *task_ctx;
3733 struct perf_event_header header;
3743 static void perf_event_task_output(struct perf_event *event,
3744 struct perf_task_event *task_event)
3746 struct perf_output_handle handle;
3747 struct task_struct *task = task_event->task;
3750 size = task_event->event_id.header.size;
3751 ret = perf_output_begin(&handle, event, size, 0, 0);
3756 task_event->event_id.pid = perf_event_pid(event, task);
3757 task_event->event_id.ppid = perf_event_pid(event, current);
3759 task_event->event_id.tid = perf_event_tid(event, task);
3760 task_event->event_id.ptid = perf_event_tid(event, current);
3762 perf_output_put(&handle, task_event->event_id);
3764 perf_output_end(&handle);
3767 static int perf_event_task_match(struct perf_event *event)
3769 if (event->state < PERF_EVENT_STATE_INACTIVE)
3772 if (event->cpu != -1 && event->cpu != smp_processor_id())
3775 if (event->attr.comm || event->attr.mmap ||
3776 event->attr.mmap_data || event->attr.task)
3782 static void perf_event_task_ctx(struct perf_event_context *ctx,
3783 struct perf_task_event *task_event)
3785 struct perf_event *event;
3787 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3788 if (perf_event_task_match(event))
3789 perf_event_task_output(event, task_event);
3793 static void perf_event_task_event(struct perf_task_event *task_event)
3795 struct perf_cpu_context *cpuctx;
3796 struct perf_event_context *ctx;
3801 list_for_each_entry_rcu(pmu, &pmus, entry) {
3802 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3803 perf_event_task_ctx(&cpuctx->ctx, task_event);
3805 ctx = task_event->task_ctx;
3807 ctxn = pmu->task_ctx_nr;
3810 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3813 perf_event_task_ctx(ctx, task_event);
3815 put_cpu_ptr(pmu->pmu_cpu_context);
3820 static void perf_event_task(struct task_struct *task,
3821 struct perf_event_context *task_ctx,
3824 struct perf_task_event task_event;
3826 if (!atomic_read(&nr_comm_events) &&
3827 !atomic_read(&nr_mmap_events) &&
3828 !atomic_read(&nr_task_events))
3831 task_event = (struct perf_task_event){
3833 .task_ctx = task_ctx,
3836 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3838 .size = sizeof(task_event.event_id),
3844 .time = perf_clock(),
3848 perf_event_task_event(&task_event);
3851 void perf_event_fork(struct task_struct *task)
3853 perf_event_task(task, NULL, 1);
3860 struct perf_comm_event {
3861 struct task_struct *task;
3866 struct perf_event_header header;
3873 static void perf_event_comm_output(struct perf_event *event,
3874 struct perf_comm_event *comm_event)
3876 struct perf_output_handle handle;
3877 int size = comm_event->event_id.header.size;
3878 int ret = perf_output_begin(&handle, event, size, 0, 0);
3883 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3884 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3886 perf_output_put(&handle, comm_event->event_id);
3887 perf_output_copy(&handle, comm_event->comm,
3888 comm_event->comm_size);
3889 perf_output_end(&handle);
3892 static int perf_event_comm_match(struct perf_event *event)
3894 if (event->state < PERF_EVENT_STATE_INACTIVE)
3897 if (event->cpu != -1 && event->cpu != smp_processor_id())
3900 if (event->attr.comm)
3906 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3907 struct perf_comm_event *comm_event)
3909 struct perf_event *event;
3911 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3912 if (perf_event_comm_match(event))
3913 perf_event_comm_output(event, comm_event);
3917 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3919 struct perf_cpu_context *cpuctx;
3920 struct perf_event_context *ctx;
3921 char comm[TASK_COMM_LEN];
3926 memset(comm, 0, sizeof(comm));
3927 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3928 size = ALIGN(strlen(comm)+1, sizeof(u64));
3930 comm_event->comm = comm;
3931 comm_event->comm_size = size;
3933 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3936 list_for_each_entry_rcu(pmu, &pmus, entry) {
3937 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3938 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3940 ctxn = pmu->task_ctx_nr;
3944 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3946 perf_event_comm_ctx(ctx, comm_event);
3948 put_cpu_ptr(pmu->pmu_cpu_context);
3953 void perf_event_comm(struct task_struct *task)
3955 struct perf_comm_event comm_event;
3956 struct perf_event_context *ctx;
3959 for_each_task_context_nr(ctxn) {
3960 ctx = task->perf_event_ctxp[ctxn];
3964 perf_event_enable_on_exec(ctx);
3967 if (!atomic_read(&nr_comm_events))
3970 comm_event = (struct perf_comm_event){
3976 .type = PERF_RECORD_COMM,
3985 perf_event_comm_event(&comm_event);
3992 struct perf_mmap_event {
3993 struct vm_area_struct *vma;
3995 const char *file_name;
3999 struct perf_event_header header;
4009 static void perf_event_mmap_output(struct perf_event *event,
4010 struct perf_mmap_event *mmap_event)
4012 struct perf_output_handle handle;
4013 int size = mmap_event->event_id.header.size;
4014 int ret = perf_output_begin(&handle, event, size, 0, 0);
4019 mmap_event->event_id.pid = perf_event_pid(event, current);
4020 mmap_event->event_id.tid = perf_event_tid(event, current);
4022 perf_output_put(&handle, mmap_event->event_id);
4023 perf_output_copy(&handle, mmap_event->file_name,
4024 mmap_event->file_size);
4025 perf_output_end(&handle);
4028 static int perf_event_mmap_match(struct perf_event *event,
4029 struct perf_mmap_event *mmap_event,
4032 if (event->state < PERF_EVENT_STATE_INACTIVE)
4035 if (event->cpu != -1 && event->cpu != smp_processor_id())
4038 if ((!executable && event->attr.mmap_data) ||
4039 (executable && event->attr.mmap))
4045 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4046 struct perf_mmap_event *mmap_event,
4049 struct perf_event *event;
4051 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4052 if (perf_event_mmap_match(event, mmap_event, executable))
4053 perf_event_mmap_output(event, mmap_event);
4057 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4059 struct perf_cpu_context *cpuctx;
4060 struct perf_event_context *ctx;
4061 struct vm_area_struct *vma = mmap_event->vma;
4062 struct file *file = vma->vm_file;
4070 memset(tmp, 0, sizeof(tmp));
4074 * d_path works from the end of the buffer backwards, so we
4075 * need to add enough zero bytes after the string to handle
4076 * the 64bit alignment we do later.
4078 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4080 name = strncpy(tmp, "//enomem", sizeof(tmp));
4083 name = d_path(&file->f_path, buf, PATH_MAX);
4085 name = strncpy(tmp, "//toolong", sizeof(tmp));
4089 if (arch_vma_name(mmap_event->vma)) {
4090 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4096 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4098 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4099 vma->vm_end >= vma->vm_mm->brk) {
4100 name = strncpy(tmp, "[heap]", sizeof(tmp));
4102 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4103 vma->vm_end >= vma->vm_mm->start_stack) {
4104 name = strncpy(tmp, "[stack]", sizeof(tmp));
4108 name = strncpy(tmp, "//anon", sizeof(tmp));
4113 size = ALIGN(strlen(name)+1, sizeof(u64));
4115 mmap_event->file_name = name;
4116 mmap_event->file_size = size;
4118 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4121 list_for_each_entry_rcu(pmu, &pmus, entry) {
4122 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4123 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4124 vma->vm_flags & VM_EXEC);
4126 ctxn = pmu->task_ctx_nr;
4130 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4132 perf_event_mmap_ctx(ctx, mmap_event,
4133 vma->vm_flags & VM_EXEC);
4136 put_cpu_ptr(pmu->pmu_cpu_context);
4143 void perf_event_mmap(struct vm_area_struct *vma)
4145 struct perf_mmap_event mmap_event;
4147 if (!atomic_read(&nr_mmap_events))
4150 mmap_event = (struct perf_mmap_event){
4156 .type = PERF_RECORD_MMAP,
4157 .misc = PERF_RECORD_MISC_USER,
4162 .start = vma->vm_start,
4163 .len = vma->vm_end - vma->vm_start,
4164 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4168 perf_event_mmap_event(&mmap_event);
4172 * IRQ throttle logging
4175 static void perf_log_throttle(struct perf_event *event, int enable)
4177 struct perf_output_handle handle;
4181 struct perf_event_header header;
4185 } throttle_event = {
4187 .type = PERF_RECORD_THROTTLE,
4189 .size = sizeof(throttle_event),
4191 .time = perf_clock(),
4192 .id = primary_event_id(event),
4193 .stream_id = event->id,
4197 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4199 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
4203 perf_output_put(&handle, throttle_event);
4204 perf_output_end(&handle);
4208 * Generic event overflow handling, sampling.
4211 static int __perf_event_overflow(struct perf_event *event, int nmi,
4212 int throttle, struct perf_sample_data *data,
4213 struct pt_regs *regs)
4215 int events = atomic_read(&event->event_limit);
4216 struct hw_perf_event *hwc = &event->hw;
4222 if (hwc->interrupts != MAX_INTERRUPTS) {
4224 if (HZ * hwc->interrupts >
4225 (u64)sysctl_perf_event_sample_rate) {
4226 hwc->interrupts = MAX_INTERRUPTS;
4227 perf_log_throttle(event, 0);
4232 * Keep re-disabling events even though on the previous
4233 * pass we disabled it - just in case we raced with a
4234 * sched-in and the event got enabled again:
4240 if (event->attr.freq) {
4241 u64 now = perf_clock();
4242 s64 delta = now - hwc->freq_time_stamp;
4244 hwc->freq_time_stamp = now;
4246 if (delta > 0 && delta < 2*TICK_NSEC)
4247 perf_adjust_period(event, delta, hwc->last_period);
4251 * XXX event_limit might not quite work as expected on inherited
4255 event->pending_kill = POLL_IN;
4256 if (events && atomic_dec_and_test(&event->event_limit)) {
4258 event->pending_kill = POLL_HUP;
4260 event->pending_disable = 1;
4261 irq_work_queue(&event->pending);
4263 perf_event_disable(event);
4266 if (event->overflow_handler)
4267 event->overflow_handler(event, nmi, data, regs);
4269 perf_event_output(event, nmi, data, regs);
4274 int perf_event_overflow(struct perf_event *event, int nmi,
4275 struct perf_sample_data *data,
4276 struct pt_regs *regs)
4278 return __perf_event_overflow(event, nmi, 1, data, regs);
4282 * Generic software event infrastructure
4285 struct swevent_htable {
4286 struct swevent_hlist *swevent_hlist;
4287 struct mutex hlist_mutex;
4290 /* Recursion avoidance in each contexts */
4291 int recursion[PERF_NR_CONTEXTS];
4294 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4297 * We directly increment event->count and keep a second value in
4298 * event->hw.period_left to count intervals. This period event
4299 * is kept in the range [-sample_period, 0] so that we can use the
4303 static u64 perf_swevent_set_period(struct perf_event *event)
4305 struct hw_perf_event *hwc = &event->hw;
4306 u64 period = hwc->last_period;
4310 hwc->last_period = hwc->sample_period;
4313 old = val = local64_read(&hwc->period_left);
4317 nr = div64_u64(period + val, period);
4318 offset = nr * period;
4320 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4326 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4327 int nmi, struct perf_sample_data *data,
4328 struct pt_regs *regs)
4330 struct hw_perf_event *hwc = &event->hw;
4333 data->period = event->hw.last_period;
4335 overflow = perf_swevent_set_period(event);
4337 if (hwc->interrupts == MAX_INTERRUPTS)
4340 for (; overflow; overflow--) {
4341 if (__perf_event_overflow(event, nmi, throttle,
4344 * We inhibit the overflow from happening when
4345 * hwc->interrupts == MAX_INTERRUPTS.
4353 static void perf_swevent_event(struct perf_event *event, u64 nr,
4354 int nmi, struct perf_sample_data *data,
4355 struct pt_regs *regs)
4357 struct hw_perf_event *hwc = &event->hw;
4359 local64_add(nr, &event->count);
4364 if (!hwc->sample_period)
4367 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4368 return perf_swevent_overflow(event, 1, nmi, data, regs);
4370 if (local64_add_negative(nr, &hwc->period_left))
4373 perf_swevent_overflow(event, 0, nmi, data, regs);
4376 static int perf_exclude_event(struct perf_event *event,
4377 struct pt_regs *regs)
4379 if (event->hw.state & PERF_HES_STOPPED)
4383 if (event->attr.exclude_user && user_mode(regs))
4386 if (event->attr.exclude_kernel && !user_mode(regs))
4393 static int perf_swevent_match(struct perf_event *event,
4394 enum perf_type_id type,
4396 struct perf_sample_data *data,
4397 struct pt_regs *regs)
4399 if (event->attr.type != type)
4402 if (event->attr.config != event_id)
4405 if (perf_exclude_event(event, regs))
4411 static inline u64 swevent_hash(u64 type, u32 event_id)
4413 u64 val = event_id | (type << 32);
4415 return hash_64(val, SWEVENT_HLIST_BITS);
4418 static inline struct hlist_head *
4419 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4421 u64 hash = swevent_hash(type, event_id);
4423 return &hlist->heads[hash];
4426 /* For the read side: events when they trigger */
4427 static inline struct hlist_head *
4428 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4430 struct swevent_hlist *hlist;
4432 hlist = rcu_dereference(swhash->swevent_hlist);
4436 return __find_swevent_head(hlist, type, event_id);
4439 /* For the event head insertion and removal in the hlist */
4440 static inline struct hlist_head *
4441 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4443 struct swevent_hlist *hlist;
4444 u32 event_id = event->attr.config;
4445 u64 type = event->attr.type;
4448 * Event scheduling is always serialized against hlist allocation
4449 * and release. Which makes the protected version suitable here.
4450 * The context lock guarantees that.
4452 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4453 lockdep_is_held(&event->ctx->lock));
4457 return __find_swevent_head(hlist, type, event_id);
4460 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4462 struct perf_sample_data *data,
4463 struct pt_regs *regs)
4465 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4466 struct perf_event *event;
4467 struct hlist_node *node;
4468 struct hlist_head *head;
4471 head = find_swevent_head_rcu(swhash, type, event_id);
4475 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4476 if (perf_swevent_match(event, type, event_id, data, regs))
4477 perf_swevent_event(event, nr, nmi, data, regs);
4483 int perf_swevent_get_recursion_context(void)
4485 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4487 return get_recursion_context(swhash->recursion);
4489 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4491 void inline perf_swevent_put_recursion_context(int rctx)
4493 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4495 put_recursion_context(swhash->recursion, rctx);
4498 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4499 struct pt_regs *regs, u64 addr)
4501 struct perf_sample_data data;
4504 preempt_disable_notrace();
4505 rctx = perf_swevent_get_recursion_context();
4509 perf_sample_data_init(&data, addr);
4511 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4513 perf_swevent_put_recursion_context(rctx);
4514 preempt_enable_notrace();
4517 static void perf_swevent_read(struct perf_event *event)
4521 static int perf_swevent_add(struct perf_event *event, int flags)
4523 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4524 struct hw_perf_event *hwc = &event->hw;
4525 struct hlist_head *head;
4527 if (hwc->sample_period) {
4528 hwc->last_period = hwc->sample_period;
4529 perf_swevent_set_period(event);
4532 hwc->state = !(flags & PERF_EF_START);
4534 head = find_swevent_head(swhash, event);
4535 if (WARN_ON_ONCE(!head))
4538 hlist_add_head_rcu(&event->hlist_entry, head);
4543 static void perf_swevent_del(struct perf_event *event, int flags)
4545 hlist_del_rcu(&event->hlist_entry);
4548 static void perf_swevent_start(struct perf_event *event, int flags)
4550 event->hw.state = 0;
4553 static void perf_swevent_stop(struct perf_event *event, int flags)
4555 event->hw.state = PERF_HES_STOPPED;
4558 /* Deref the hlist from the update side */
4559 static inline struct swevent_hlist *
4560 swevent_hlist_deref(struct swevent_htable *swhash)
4562 return rcu_dereference_protected(swhash->swevent_hlist,
4563 lockdep_is_held(&swhash->hlist_mutex));
4566 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4568 struct swevent_hlist *hlist;
4570 hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
4574 static void swevent_hlist_release(struct swevent_htable *swhash)
4576 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4581 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4582 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4585 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4587 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4589 mutex_lock(&swhash->hlist_mutex);
4591 if (!--swhash->hlist_refcount)
4592 swevent_hlist_release(swhash);
4594 mutex_unlock(&swhash->hlist_mutex);
4597 static void swevent_hlist_put(struct perf_event *event)
4601 if (event->cpu != -1) {
4602 swevent_hlist_put_cpu(event, event->cpu);
4606 for_each_possible_cpu(cpu)
4607 swevent_hlist_put_cpu(event, cpu);
4610 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4612 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4615 mutex_lock(&swhash->hlist_mutex);
4617 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4618 struct swevent_hlist *hlist;
4620 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
4625 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4627 swhash->hlist_refcount++;
4629 mutex_unlock(&swhash->hlist_mutex);
4634 static int swevent_hlist_get(struct perf_event *event)
4637 int cpu, failed_cpu;
4639 if (event->cpu != -1)
4640 return swevent_hlist_get_cpu(event, event->cpu);
4643 for_each_possible_cpu(cpu) {
4644 err = swevent_hlist_get_cpu(event, cpu);
4654 for_each_possible_cpu(cpu) {
4655 if (cpu == failed_cpu)
4657 swevent_hlist_put_cpu(event, cpu);
4664 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4666 static void sw_perf_event_destroy(struct perf_event *event)
4668 u64 event_id = event->attr.config;
4670 WARN_ON(event->parent);
4672 jump_label_dec(&perf_swevent_enabled[event_id]);
4673 swevent_hlist_put(event);
4676 static int perf_swevent_init(struct perf_event *event)
4678 int event_id = event->attr.config;
4680 if (event->attr.type != PERF_TYPE_SOFTWARE)
4684 case PERF_COUNT_SW_CPU_CLOCK:
4685 case PERF_COUNT_SW_TASK_CLOCK:
4692 if (event_id > PERF_COUNT_SW_MAX)
4695 if (!event->parent) {
4698 err = swevent_hlist_get(event);
4702 jump_label_inc(&perf_swevent_enabled[event_id]);
4703 event->destroy = sw_perf_event_destroy;
4709 static struct pmu perf_swevent = {
4710 .task_ctx_nr = perf_sw_context,
4712 .event_init = perf_swevent_init,
4713 .add = perf_swevent_add,
4714 .del = perf_swevent_del,
4715 .start = perf_swevent_start,
4716 .stop = perf_swevent_stop,
4717 .read = perf_swevent_read,
4720 #ifdef CONFIG_EVENT_TRACING
4722 static int perf_tp_filter_match(struct perf_event *event,
4723 struct perf_sample_data *data)
4725 void *record = data->raw->data;
4727 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4732 static int perf_tp_event_match(struct perf_event *event,
4733 struct perf_sample_data *data,
4734 struct pt_regs *regs)
4737 * All tracepoints are from kernel-space.
4739 if (event->attr.exclude_kernel)
4742 if (!perf_tp_filter_match(event, data))
4748 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4749 struct pt_regs *regs, struct hlist_head *head, int rctx)
4751 struct perf_sample_data data;
4752 struct perf_event *event;
4753 struct hlist_node *node;
4755 struct perf_raw_record raw = {
4760 perf_sample_data_init(&data, addr);
4763 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4764 if (perf_tp_event_match(event, &data, regs))
4765 perf_swevent_event(event, count, 1, &data, regs);
4768 perf_swevent_put_recursion_context(rctx);
4770 EXPORT_SYMBOL_GPL(perf_tp_event);
4772 static void tp_perf_event_destroy(struct perf_event *event)
4774 perf_trace_destroy(event);
4777 static int perf_tp_event_init(struct perf_event *event)
4781 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4785 * Raw tracepoint data is a severe data leak, only allow root to
4788 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4789 perf_paranoid_tracepoint_raw() &&
4790 !capable(CAP_SYS_ADMIN))
4793 err = perf_trace_init(event);
4797 event->destroy = tp_perf_event_destroy;
4802 static struct pmu perf_tracepoint = {
4803 .task_ctx_nr = perf_sw_context,
4805 .event_init = perf_tp_event_init,
4806 .add = perf_trace_add,
4807 .del = perf_trace_del,
4808 .start = perf_swevent_start,
4809 .stop = perf_swevent_stop,
4810 .read = perf_swevent_read,
4813 static inline void perf_tp_register(void)
4815 perf_pmu_register(&perf_tracepoint);
4818 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4823 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4826 filter_str = strndup_user(arg, PAGE_SIZE);
4827 if (IS_ERR(filter_str))
4828 return PTR_ERR(filter_str);
4830 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4836 static void perf_event_free_filter(struct perf_event *event)
4838 ftrace_profile_free_filter(event);
4843 static inline void perf_tp_register(void)
4847 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4852 static void perf_event_free_filter(struct perf_event *event)
4856 #endif /* CONFIG_EVENT_TRACING */
4858 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4859 void perf_bp_event(struct perf_event *bp, void *data)
4861 struct perf_sample_data sample;
4862 struct pt_regs *regs = data;
4864 perf_sample_data_init(&sample, bp->attr.bp_addr);
4866 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4867 perf_swevent_event(bp, 1, 1, &sample, regs);
4872 * hrtimer based swevent callback
4875 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4877 enum hrtimer_restart ret = HRTIMER_RESTART;
4878 struct perf_sample_data data;
4879 struct pt_regs *regs;
4880 struct perf_event *event;
4883 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4884 event->pmu->read(event);
4886 perf_sample_data_init(&data, 0);
4887 data.period = event->hw.last_period;
4888 regs = get_irq_regs();
4890 if (regs && !perf_exclude_event(event, regs)) {
4891 if (!(event->attr.exclude_idle && current->pid == 0))
4892 if (perf_event_overflow(event, 0, &data, regs))
4893 ret = HRTIMER_NORESTART;
4896 period = max_t(u64, 10000, event->hw.sample_period);
4897 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4902 static void perf_swevent_start_hrtimer(struct perf_event *event)
4904 struct hw_perf_event *hwc = &event->hw;
4906 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4907 hwc->hrtimer.function = perf_swevent_hrtimer;
4908 if (hwc->sample_period) {
4909 s64 period = local64_read(&hwc->period_left);
4915 local64_set(&hwc->period_left, 0);
4917 period = max_t(u64, 10000, hwc->sample_period);
4919 __hrtimer_start_range_ns(&hwc->hrtimer,
4920 ns_to_ktime(period), 0,
4921 HRTIMER_MODE_REL_PINNED, 0);
4925 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4927 struct hw_perf_event *hwc = &event->hw;
4929 if (hwc->sample_period) {
4930 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4931 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4933 hrtimer_cancel(&hwc->hrtimer);
4938 * Software event: cpu wall time clock
4941 static void cpu_clock_event_update(struct perf_event *event)
4946 now = local_clock();
4947 prev = local64_xchg(&event->hw.prev_count, now);
4948 local64_add(now - prev, &event->count);
4951 static void cpu_clock_event_start(struct perf_event *event, int flags)
4953 local64_set(&event->hw.prev_count, local_clock());
4954 perf_swevent_start_hrtimer(event);
4957 static void cpu_clock_event_stop(struct perf_event *event, int flags)
4959 perf_swevent_cancel_hrtimer(event);
4960 cpu_clock_event_update(event);
4963 static int cpu_clock_event_add(struct perf_event *event, int flags)
4965 if (flags & PERF_EF_START)
4966 cpu_clock_event_start(event, flags);
4971 static void cpu_clock_event_del(struct perf_event *event, int flags)
4973 cpu_clock_event_stop(event, flags);
4976 static void cpu_clock_event_read(struct perf_event *event)
4978 cpu_clock_event_update(event);
4981 static int cpu_clock_event_init(struct perf_event *event)
4983 if (event->attr.type != PERF_TYPE_SOFTWARE)
4986 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
4992 static struct pmu perf_cpu_clock = {
4993 .task_ctx_nr = perf_sw_context,
4995 .event_init = cpu_clock_event_init,
4996 .add = cpu_clock_event_add,
4997 .del = cpu_clock_event_del,
4998 .start = cpu_clock_event_start,
4999 .stop = cpu_clock_event_stop,
5000 .read = cpu_clock_event_read,
5004 * Software event: task time clock
5007 static void task_clock_event_update(struct perf_event *event, u64 now)
5012 prev = local64_xchg(&event->hw.prev_count, now);
5014 local64_add(delta, &event->count);
5017 static void task_clock_event_start(struct perf_event *event, int flags)
5019 local64_set(&event->hw.prev_count, event->ctx->time);
5020 perf_swevent_start_hrtimer(event);
5023 static void task_clock_event_stop(struct perf_event *event, int flags)
5025 perf_swevent_cancel_hrtimer(event);
5026 task_clock_event_update(event, event->ctx->time);
5029 static int task_clock_event_add(struct perf_event *event, int flags)
5031 if (flags & PERF_EF_START)
5032 task_clock_event_start(event, flags);
5037 static void task_clock_event_del(struct perf_event *event, int flags)
5039 task_clock_event_stop(event, PERF_EF_UPDATE);
5042 static void task_clock_event_read(struct perf_event *event)
5047 update_context_time(event->ctx);
5048 time = event->ctx->time;
5050 u64 now = perf_clock();
5051 u64 delta = now - event->ctx->timestamp;
5052 time = event->ctx->time + delta;
5055 task_clock_event_update(event, time);
5058 static int task_clock_event_init(struct perf_event *event)
5060 if (event->attr.type != PERF_TYPE_SOFTWARE)
5063 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5069 static struct pmu perf_task_clock = {
5070 .task_ctx_nr = perf_sw_context,
5072 .event_init = task_clock_event_init,
5073 .add = task_clock_event_add,
5074 .del = task_clock_event_del,
5075 .start = task_clock_event_start,
5076 .stop = task_clock_event_stop,
5077 .read = task_clock_event_read,
5080 static void perf_pmu_nop_void(struct pmu *pmu)
5084 static int perf_pmu_nop_int(struct pmu *pmu)
5089 static void perf_pmu_start_txn(struct pmu *pmu)
5091 perf_pmu_disable(pmu);
5094 static int perf_pmu_commit_txn(struct pmu *pmu)
5096 perf_pmu_enable(pmu);
5100 static void perf_pmu_cancel_txn(struct pmu *pmu)
5102 perf_pmu_enable(pmu);
5106 * Ensures all contexts with the same task_ctx_nr have the same
5107 * pmu_cpu_context too.
5109 static void *find_pmu_context(int ctxn)
5116 list_for_each_entry(pmu, &pmus, entry) {
5117 if (pmu->task_ctx_nr == ctxn)
5118 return pmu->pmu_cpu_context;
5124 static void free_pmu_context(void * __percpu cpu_context)
5128 mutex_lock(&pmus_lock);
5130 * Like a real lame refcount.
5132 list_for_each_entry(pmu, &pmus, entry) {
5133 if (pmu->pmu_cpu_context == cpu_context)
5137 free_percpu(cpu_context);
5139 mutex_unlock(&pmus_lock);
5142 int perf_pmu_register(struct pmu *pmu)
5146 mutex_lock(&pmus_lock);
5148 pmu->pmu_disable_count = alloc_percpu(int);
5149 if (!pmu->pmu_disable_count)
5152 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5153 if (pmu->pmu_cpu_context)
5154 goto got_cpu_context;
5156 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5157 if (!pmu->pmu_cpu_context)
5160 for_each_possible_cpu(cpu) {
5161 struct perf_cpu_context *cpuctx;
5163 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5164 __perf_event_init_context(&cpuctx->ctx);
5165 cpuctx->ctx.type = cpu_context;
5166 cpuctx->ctx.pmu = pmu;
5167 cpuctx->jiffies_interval = 1;
5168 INIT_LIST_HEAD(&cpuctx->rotation_list);
5172 if (!pmu->start_txn) {
5173 if (pmu->pmu_enable) {
5175 * If we have pmu_enable/pmu_disable calls, install
5176 * transaction stubs that use that to try and batch
5177 * hardware accesses.
5179 pmu->start_txn = perf_pmu_start_txn;
5180 pmu->commit_txn = perf_pmu_commit_txn;
5181 pmu->cancel_txn = perf_pmu_cancel_txn;
5183 pmu->start_txn = perf_pmu_nop_void;
5184 pmu->commit_txn = perf_pmu_nop_int;
5185 pmu->cancel_txn = perf_pmu_nop_void;
5189 if (!pmu->pmu_enable) {
5190 pmu->pmu_enable = perf_pmu_nop_void;
5191 pmu->pmu_disable = perf_pmu_nop_void;
5194 list_add_rcu(&pmu->entry, &pmus);
5197 mutex_unlock(&pmus_lock);
5202 free_percpu(pmu->pmu_disable_count);
5206 void perf_pmu_unregister(struct pmu *pmu)
5208 mutex_lock(&pmus_lock);
5209 list_del_rcu(&pmu->entry);
5210 mutex_unlock(&pmus_lock);
5213 * We dereference the pmu list under both SRCU and regular RCU, so
5214 * synchronize against both of those.
5216 synchronize_srcu(&pmus_srcu);
5219 free_percpu(pmu->pmu_disable_count);
5220 free_pmu_context(pmu->pmu_cpu_context);
5223 struct pmu *perf_init_event(struct perf_event *event)
5225 struct pmu *pmu = NULL;
5228 idx = srcu_read_lock(&pmus_srcu);
5229 list_for_each_entry_rcu(pmu, &pmus, entry) {
5230 int ret = pmu->event_init(event);
5234 if (ret != -ENOENT) {
5239 pmu = ERR_PTR(-ENOENT);
5241 srcu_read_unlock(&pmus_srcu, idx);
5247 * Allocate and initialize a event structure
5249 static struct perf_event *
5250 perf_event_alloc(struct perf_event_attr *attr, int cpu,
5251 struct task_struct *task,
5252 struct perf_event *group_leader,
5253 struct perf_event *parent_event,
5254 perf_overflow_handler_t overflow_handler)
5257 struct perf_event *event;
5258 struct hw_perf_event *hwc;
5261 event = kzalloc(sizeof(*event), GFP_KERNEL);
5263 return ERR_PTR(-ENOMEM);
5266 * Single events are their own group leaders, with an
5267 * empty sibling list:
5270 group_leader = event;
5272 mutex_init(&event->child_mutex);
5273 INIT_LIST_HEAD(&event->child_list);
5275 INIT_LIST_HEAD(&event->group_entry);
5276 INIT_LIST_HEAD(&event->event_entry);
5277 INIT_LIST_HEAD(&event->sibling_list);
5278 init_waitqueue_head(&event->waitq);
5279 init_irq_work(&event->pending, perf_pending_event);
5281 mutex_init(&event->mmap_mutex);
5284 event->attr = *attr;
5285 event->group_leader = group_leader;
5289 event->parent = parent_event;
5291 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5292 event->id = atomic64_inc_return(&perf_event_id);
5294 event->state = PERF_EVENT_STATE_INACTIVE;
5297 event->attach_state = PERF_ATTACH_TASK;
5298 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5300 * hw_breakpoint is a bit difficult here..
5302 if (attr->type == PERF_TYPE_BREAKPOINT)
5303 event->hw.bp_target = task;
5307 if (!overflow_handler && parent_event)
5308 overflow_handler = parent_event->overflow_handler;
5310 event->overflow_handler = overflow_handler;
5313 event->state = PERF_EVENT_STATE_OFF;
5318 hwc->sample_period = attr->sample_period;
5319 if (attr->freq && attr->sample_freq)
5320 hwc->sample_period = 1;
5321 hwc->last_period = hwc->sample_period;
5323 local64_set(&hwc->period_left, hwc->sample_period);
5326 * we currently do not support PERF_FORMAT_GROUP on inherited events
5328 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5331 pmu = perf_init_event(event);
5337 else if (IS_ERR(pmu))
5342 put_pid_ns(event->ns);
5344 return ERR_PTR(err);
5349 if (!event->parent) {
5350 if (event->attach_state & PERF_ATTACH_TASK)
5351 jump_label_inc(&perf_task_events);
5352 if (event->attr.mmap || event->attr.mmap_data)
5353 atomic_inc(&nr_mmap_events);
5354 if (event->attr.comm)
5355 atomic_inc(&nr_comm_events);
5356 if (event->attr.task)
5357 atomic_inc(&nr_task_events);
5358 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5359 err = get_callchain_buffers();
5362 return ERR_PTR(err);
5370 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5371 struct perf_event_attr *attr)
5376 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5380 * zero the full structure, so that a short copy will be nice.
5382 memset(attr, 0, sizeof(*attr));
5384 ret = get_user(size, &uattr->size);
5388 if (size > PAGE_SIZE) /* silly large */
5391 if (!size) /* abi compat */
5392 size = PERF_ATTR_SIZE_VER0;
5394 if (size < PERF_ATTR_SIZE_VER0)
5398 * If we're handed a bigger struct than we know of,
5399 * ensure all the unknown bits are 0 - i.e. new
5400 * user-space does not rely on any kernel feature
5401 * extensions we dont know about yet.
5403 if (size > sizeof(*attr)) {
5404 unsigned char __user *addr;
5405 unsigned char __user *end;
5408 addr = (void __user *)uattr + sizeof(*attr);
5409 end = (void __user *)uattr + size;
5411 for (; addr < end; addr++) {
5412 ret = get_user(val, addr);
5418 size = sizeof(*attr);
5421 ret = copy_from_user(attr, uattr, size);
5426 * If the type exists, the corresponding creation will verify
5429 if (attr->type >= PERF_TYPE_MAX)
5432 if (attr->__reserved_1)
5435 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5438 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5445 put_user(sizeof(*attr), &uattr->size);
5451 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5453 struct perf_buffer *buffer = NULL, *old_buffer = NULL;
5459 /* don't allow circular references */
5460 if (event == output_event)
5464 * Don't allow cross-cpu buffers
5466 if (output_event->cpu != event->cpu)
5470 * If its not a per-cpu buffer, it must be the same task.
5472 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
5476 mutex_lock(&event->mmap_mutex);
5477 /* Can't redirect output if we've got an active mmap() */
5478 if (atomic_read(&event->mmap_count))
5482 /* get the buffer we want to redirect to */
5483 buffer = perf_buffer_get(output_event);
5488 old_buffer = event->buffer;
5489 rcu_assign_pointer(event->buffer, buffer);
5492 mutex_unlock(&event->mmap_mutex);
5495 perf_buffer_put(old_buffer);
5501 * sys_perf_event_open - open a performance event, associate it to a task/cpu
5503 * @attr_uptr: event_id type attributes for monitoring/sampling
5506 * @group_fd: group leader event fd
5508 SYSCALL_DEFINE5(perf_event_open,
5509 struct perf_event_attr __user *, attr_uptr,
5510 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5512 struct perf_event *group_leader = NULL, *output_event = NULL;
5513 struct perf_event *event, *sibling;
5514 struct perf_event_attr attr;
5515 struct perf_event_context *ctx;
5516 struct file *event_file = NULL;
5517 struct file *group_file = NULL;
5518 struct task_struct *task = NULL;
5522 int fput_needed = 0;
5525 /* for future expandability... */
5526 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
5529 err = perf_copy_attr(attr_uptr, &attr);
5533 if (!attr.exclude_kernel) {
5534 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
5539 if (attr.sample_freq > sysctl_perf_event_sample_rate)
5543 event_fd = get_unused_fd_flags(O_RDWR);
5547 if (group_fd != -1) {
5548 group_leader = perf_fget_light(group_fd, &fput_needed);
5549 if (IS_ERR(group_leader)) {
5550 err = PTR_ERR(group_leader);
5553 group_file = group_leader->filp;
5554 if (flags & PERF_FLAG_FD_OUTPUT)
5555 output_event = group_leader;
5556 if (flags & PERF_FLAG_FD_NO_GROUP)
5557 group_leader = NULL;
5561 task = find_lively_task_by_vpid(pid);
5563 err = PTR_ERR(task);
5568 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5569 if (IS_ERR(event)) {
5570 err = PTR_ERR(event);
5575 * Special case software events and allow them to be part of
5576 * any hardware group.
5581 (is_software_event(event) != is_software_event(group_leader))) {
5582 if (is_software_event(event)) {
5584 * If event and group_leader are not both a software
5585 * event, and event is, then group leader is not.
5587 * Allow the addition of software events to !software
5588 * groups, this is safe because software events never
5591 pmu = group_leader->pmu;
5592 } else if (is_software_event(group_leader) &&
5593 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5595 * In case the group is a pure software group, and we
5596 * try to add a hardware event, move the whole group to
5597 * the hardware context.
5604 * Get the target context (task or percpu):
5606 ctx = find_get_context(pmu, task, cpu);
5613 * Look up the group leader (we will attach this event to it):
5619 * Do not allow a recursive hierarchy (this new sibling
5620 * becoming part of another group-sibling):
5622 if (group_leader->group_leader != group_leader)
5625 * Do not allow to attach to a group in a different
5626 * task or CPU context:
5629 if (group_leader->ctx->type != ctx->type)
5632 if (group_leader->ctx != ctx)
5637 * Only a group leader can be exclusive or pinned
5639 if (attr.exclusive || attr.pinned)
5644 err = perf_event_set_output(event, output_event);
5649 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5650 if (IS_ERR(event_file)) {
5651 err = PTR_ERR(event_file);
5656 struct perf_event_context *gctx = group_leader->ctx;
5658 mutex_lock(&gctx->mutex);
5659 perf_event_remove_from_context(group_leader);
5660 list_for_each_entry(sibling, &group_leader->sibling_list,
5662 perf_event_remove_from_context(sibling);
5665 mutex_unlock(&gctx->mutex);
5669 event->filp = event_file;
5670 WARN_ON_ONCE(ctx->parent_ctx);
5671 mutex_lock(&ctx->mutex);
5674 perf_install_in_context(ctx, group_leader, cpu);
5676 list_for_each_entry(sibling, &group_leader->sibling_list,
5678 perf_install_in_context(ctx, sibling, cpu);
5683 perf_install_in_context(ctx, event, cpu);
5685 mutex_unlock(&ctx->mutex);
5687 event->owner = current;
5688 get_task_struct(current);
5689 mutex_lock(¤t->perf_event_mutex);
5690 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5691 mutex_unlock(¤t->perf_event_mutex);
5694 * Drop the reference on the group_event after placing the
5695 * new event on the sibling_list. This ensures destruction
5696 * of the group leader will find the pointer to itself in
5697 * perf_group_detach().
5699 fput_light(group_file, fput_needed);
5700 fd_install(event_fd, event_file);
5709 put_task_struct(task);
5711 fput_light(group_file, fput_needed);
5713 put_unused_fd(event_fd);
5718 * perf_event_create_kernel_counter
5720 * @attr: attributes of the counter to create
5721 * @cpu: cpu in which the counter is bound
5722 * @task: task to profile (NULL for percpu)
5725 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5726 struct task_struct *task,
5727 perf_overflow_handler_t overflow_handler)
5729 struct perf_event_context *ctx;
5730 struct perf_event *event;
5734 * Get the target context (task or percpu):
5737 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
5738 if (IS_ERR(event)) {
5739 err = PTR_ERR(event);
5743 ctx = find_get_context(event->pmu, task, cpu);
5750 WARN_ON_ONCE(ctx->parent_ctx);
5751 mutex_lock(&ctx->mutex);
5752 perf_install_in_context(ctx, event, cpu);
5754 mutex_unlock(&ctx->mutex);
5756 event->owner = current;
5757 get_task_struct(current);
5758 mutex_lock(¤t->perf_event_mutex);
5759 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
5760 mutex_unlock(¤t->perf_event_mutex);
5767 return ERR_PTR(err);
5769 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5771 static void sync_child_event(struct perf_event *child_event,
5772 struct task_struct *child)
5774 struct perf_event *parent_event = child_event->parent;
5777 if (child_event->attr.inherit_stat)
5778 perf_event_read_event(child_event, child);
5780 child_val = perf_event_count(child_event);
5783 * Add back the child's count to the parent's count:
5785 atomic64_add(child_val, &parent_event->child_count);
5786 atomic64_add(child_event->total_time_enabled,
5787 &parent_event->child_total_time_enabled);
5788 atomic64_add(child_event->total_time_running,
5789 &parent_event->child_total_time_running);
5792 * Remove this event from the parent's list
5794 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5795 mutex_lock(&parent_event->child_mutex);
5796 list_del_init(&child_event->child_list);
5797 mutex_unlock(&parent_event->child_mutex);
5800 * Release the parent event, if this was the last
5803 fput(parent_event->filp);
5807 __perf_event_exit_task(struct perf_event *child_event,
5808 struct perf_event_context *child_ctx,
5809 struct task_struct *child)
5811 struct perf_event *parent_event;
5813 perf_event_remove_from_context(child_event);
5815 parent_event = child_event->parent;
5817 * It can happen that parent exits first, and has events
5818 * that are still around due to the child reference. These
5819 * events need to be zapped - but otherwise linger.
5822 sync_child_event(child_event, child);
5823 free_event(child_event);
5827 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
5829 struct perf_event *child_event, *tmp;
5830 struct perf_event_context *child_ctx;
5831 unsigned long flags;
5833 if (likely(!child->perf_event_ctxp[ctxn])) {
5834 perf_event_task(child, NULL, 0);
5838 local_irq_save(flags);
5840 * We can't reschedule here because interrupts are disabled,
5841 * and either child is current or it is a task that can't be
5842 * scheduled, so we are now safe from rescheduling changing
5845 child_ctx = child->perf_event_ctxp[ctxn];
5846 task_ctx_sched_out(child_ctx, EVENT_ALL);
5849 * Take the context lock here so that if find_get_context is
5850 * reading child->perf_event_ctxp, we wait until it has
5851 * incremented the context's refcount before we do put_ctx below.
5853 raw_spin_lock(&child_ctx->lock);
5854 child->perf_event_ctxp[ctxn] = NULL;
5856 * If this context is a clone; unclone it so it can't get
5857 * swapped to another process while we're removing all
5858 * the events from it.
5860 unclone_ctx(child_ctx);
5861 update_context_time(child_ctx);
5862 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5865 * Report the task dead after unscheduling the events so that we
5866 * won't get any samples after PERF_RECORD_EXIT. We can however still
5867 * get a few PERF_RECORD_READ events.
5869 perf_event_task(child, child_ctx, 0);
5872 * We can recurse on the same lock type through:
5874 * __perf_event_exit_task()
5875 * sync_child_event()
5876 * fput(parent_event->filp)
5878 * mutex_lock(&ctx->mutex)
5880 * But since its the parent context it won't be the same instance.
5882 mutex_lock(&child_ctx->mutex);
5885 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5887 __perf_event_exit_task(child_event, child_ctx, child);
5889 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5891 __perf_event_exit_task(child_event, child_ctx, child);
5894 * If the last event was a group event, it will have appended all
5895 * its siblings to the list, but we obtained 'tmp' before that which
5896 * will still point to the list head terminating the iteration.
5898 if (!list_empty(&child_ctx->pinned_groups) ||
5899 !list_empty(&child_ctx->flexible_groups))
5902 mutex_unlock(&child_ctx->mutex);
5908 * When a child task exits, feed back event values to parent events.
5910 void perf_event_exit_task(struct task_struct *child)
5914 for_each_task_context_nr(ctxn)
5915 perf_event_exit_task_context(child, ctxn);
5918 static void perf_free_event(struct perf_event *event,
5919 struct perf_event_context *ctx)
5921 struct perf_event *parent = event->parent;
5923 if (WARN_ON_ONCE(!parent))
5926 mutex_lock(&parent->child_mutex);
5927 list_del_init(&event->child_list);
5928 mutex_unlock(&parent->child_mutex);
5932 perf_group_detach(event);
5933 list_del_event(event, ctx);
5938 * free an unexposed, unused context as created by inheritance by
5939 * perf_event_init_task below, used by fork() in case of fail.
5941 void perf_event_free_task(struct task_struct *task)
5943 struct perf_event_context *ctx;
5944 struct perf_event *event, *tmp;
5947 for_each_task_context_nr(ctxn) {
5948 ctx = task->perf_event_ctxp[ctxn];
5952 mutex_lock(&ctx->mutex);
5954 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
5956 perf_free_event(event, ctx);
5958 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5960 perf_free_event(event, ctx);
5962 if (!list_empty(&ctx->pinned_groups) ||
5963 !list_empty(&ctx->flexible_groups))
5966 mutex_unlock(&ctx->mutex);
5972 void perf_event_delayed_put(struct task_struct *task)
5976 for_each_task_context_nr(ctxn)
5977 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
5981 * inherit a event from parent task to child task:
5983 static struct perf_event *
5984 inherit_event(struct perf_event *parent_event,
5985 struct task_struct *parent,
5986 struct perf_event_context *parent_ctx,
5987 struct task_struct *child,
5988 struct perf_event *group_leader,
5989 struct perf_event_context *child_ctx)
5991 struct perf_event *child_event;
5992 unsigned long flags;
5995 * Instead of creating recursive hierarchies of events,
5996 * we link inherited events back to the original parent,
5997 * which has a filp for sure, which we use as the reference
6000 if (parent_event->parent)
6001 parent_event = parent_event->parent;
6003 child_event = perf_event_alloc(&parent_event->attr,
6006 group_leader, parent_event,
6008 if (IS_ERR(child_event))
6013 * Make the child state follow the state of the parent event,
6014 * not its attr.disabled bit. We hold the parent's mutex,
6015 * so we won't race with perf_event_{en, dis}able_family.
6017 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6018 child_event->state = PERF_EVENT_STATE_INACTIVE;
6020 child_event->state = PERF_EVENT_STATE_OFF;
6022 if (parent_event->attr.freq) {
6023 u64 sample_period = parent_event->hw.sample_period;
6024 struct hw_perf_event *hwc = &child_event->hw;
6026 hwc->sample_period = sample_period;
6027 hwc->last_period = sample_period;
6029 local64_set(&hwc->period_left, sample_period);
6032 child_event->ctx = child_ctx;
6033 child_event->overflow_handler = parent_event->overflow_handler;
6036 * Link it up in the child's context:
6038 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6039 add_event_to_ctx(child_event, child_ctx);
6040 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6043 * Get a reference to the parent filp - we will fput it
6044 * when the child event exits. This is safe to do because
6045 * we are in the parent and we know that the filp still
6046 * exists and has a nonzero count:
6048 atomic_long_inc(&parent_event->filp->f_count);
6051 * Link this into the parent event's child list
6053 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6054 mutex_lock(&parent_event->child_mutex);
6055 list_add_tail(&child_event->child_list, &parent_event->child_list);
6056 mutex_unlock(&parent_event->child_mutex);
6061 static int inherit_group(struct perf_event *parent_event,
6062 struct task_struct *parent,
6063 struct perf_event_context *parent_ctx,
6064 struct task_struct *child,
6065 struct perf_event_context *child_ctx)
6067 struct perf_event *leader;
6068 struct perf_event *sub;
6069 struct perf_event *child_ctr;
6071 leader = inherit_event(parent_event, parent, parent_ctx,
6072 child, NULL, child_ctx);
6074 return PTR_ERR(leader);
6075 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6076 child_ctr = inherit_event(sub, parent, parent_ctx,
6077 child, leader, child_ctx);
6078 if (IS_ERR(child_ctr))
6079 return PTR_ERR(child_ctr);
6085 inherit_task_group(struct perf_event *event, struct task_struct *parent,
6086 struct perf_event_context *parent_ctx,
6087 struct task_struct *child, int ctxn,
6091 struct perf_event_context *child_ctx;
6093 if (!event->attr.inherit) {
6098 child_ctx = child->perf_event_ctxp[ctxn];
6101 * This is executed from the parent task context, so
6102 * inherit events that have been marked for cloning.
6103 * First allocate and initialize a context for the
6107 child_ctx = alloc_perf_context(event->pmu, child);
6111 child->perf_event_ctxp[ctxn] = child_ctx;
6114 ret = inherit_group(event, parent, parent_ctx,
6124 * Initialize the perf_event context in task_struct
6126 int perf_event_init_context(struct task_struct *child, int ctxn)
6128 struct perf_event_context *child_ctx, *parent_ctx;
6129 struct perf_event_context *cloned_ctx;
6130 struct perf_event *event;
6131 struct task_struct *parent = current;
6132 int inherited_all = 1;
6135 child->perf_event_ctxp[ctxn] = NULL;
6137 mutex_init(&child->perf_event_mutex);
6138 INIT_LIST_HEAD(&child->perf_event_list);
6140 if (likely(!parent->perf_event_ctxp[ctxn]))
6144 * If the parent's context is a clone, pin it so it won't get
6147 parent_ctx = perf_pin_task_context(parent, ctxn);
6150 * No need to check if parent_ctx != NULL here; since we saw
6151 * it non-NULL earlier, the only reason for it to become NULL
6152 * is if we exit, and since we're currently in the middle of
6153 * a fork we can't be exiting at the same time.
6157 * Lock the parent list. No need to lock the child - not PID
6158 * hashed yet and not running, so nobody can access it.
6160 mutex_lock(&parent_ctx->mutex);
6163 * We dont have to disable NMIs - we are only looking at
6164 * the list, not manipulating it:
6166 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6167 ret = inherit_task_group(event, parent, parent_ctx,
6168 child, ctxn, &inherited_all);
6173 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6174 ret = inherit_task_group(event, parent, parent_ctx,
6175 child, ctxn, &inherited_all);
6180 child_ctx = child->perf_event_ctxp[ctxn];
6182 if (child_ctx && inherited_all) {
6184 * Mark the child context as a clone of the parent
6185 * context, or of whatever the parent is a clone of.
6186 * Note that if the parent is a clone, it could get
6187 * uncloned at any point, but that doesn't matter
6188 * because the list of events and the generation
6189 * count can't have changed since we took the mutex.
6191 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
6193 child_ctx->parent_ctx = cloned_ctx;
6194 child_ctx->parent_gen = parent_ctx->parent_gen;
6196 child_ctx->parent_ctx = parent_ctx;
6197 child_ctx->parent_gen = parent_ctx->generation;
6199 get_ctx(child_ctx->parent_ctx);
6202 mutex_unlock(&parent_ctx->mutex);
6204 perf_unpin_context(parent_ctx);
6210 * Initialize the perf_event context in task_struct
6212 int perf_event_init_task(struct task_struct *child)
6216 for_each_task_context_nr(ctxn) {
6217 ret = perf_event_init_context(child, ctxn);
6225 static void __init perf_event_init_all_cpus(void)
6227 struct swevent_htable *swhash;
6230 for_each_possible_cpu(cpu) {
6231 swhash = &per_cpu(swevent_htable, cpu);
6232 mutex_init(&swhash->hlist_mutex);
6233 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6237 static void __cpuinit perf_event_init_cpu(int cpu)
6239 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6241 mutex_lock(&swhash->hlist_mutex);
6242 if (swhash->hlist_refcount > 0) {
6243 struct swevent_hlist *hlist;
6245 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6247 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6249 mutex_unlock(&swhash->hlist_mutex);
6252 #ifdef CONFIG_HOTPLUG_CPU
6253 static void perf_pmu_rotate_stop(struct pmu *pmu)
6255 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6257 WARN_ON(!irqs_disabled());
6259 list_del_init(&cpuctx->rotation_list);
6262 static void __perf_event_exit_context(void *__info)
6264 struct perf_event_context *ctx = __info;
6265 struct perf_event *event, *tmp;
6267 perf_pmu_rotate_stop(ctx->pmu);
6269 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6270 __perf_event_remove_from_context(event);
6271 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6272 __perf_event_remove_from_context(event);
6275 static void perf_event_exit_cpu_context(int cpu)
6277 struct perf_event_context *ctx;
6281 idx = srcu_read_lock(&pmus_srcu);
6282 list_for_each_entry_rcu(pmu, &pmus, entry) {
6283 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6285 mutex_lock(&ctx->mutex);
6286 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6287 mutex_unlock(&ctx->mutex);
6289 srcu_read_unlock(&pmus_srcu, idx);
6292 static void perf_event_exit_cpu(int cpu)
6294 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6296 mutex_lock(&swhash->hlist_mutex);
6297 swevent_hlist_release(swhash);
6298 mutex_unlock(&swhash->hlist_mutex);
6300 perf_event_exit_cpu_context(cpu);
6303 static inline void perf_event_exit_cpu(int cpu) { }
6306 static int __cpuinit
6307 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6309 unsigned int cpu = (long)hcpu;
6311 switch (action & ~CPU_TASKS_FROZEN) {
6313 case CPU_UP_PREPARE:
6314 case CPU_DOWN_FAILED:
6315 perf_event_init_cpu(cpu);
6318 case CPU_UP_CANCELED:
6319 case CPU_DOWN_PREPARE:
6320 perf_event_exit_cpu(cpu);
6330 void __init perf_event_init(void)
6332 perf_event_init_all_cpus();
6333 init_srcu_struct(&pmus_srcu);
6334 perf_pmu_register(&perf_swevent);
6335 perf_pmu_register(&perf_cpu_clock);
6336 perf_pmu_register(&perf_task_clock);
6338 perf_cpu_notifier(perf_cpu_notify);