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/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
34 #include <asm/irq_regs.h>
37 * Each CPU has a list of per CPU events:
39 static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly = 1;
59 static inline bool perf_paranoid_tracepoint_raw(void)
61 return sysctl_perf_event_paranoid > -1;
64 static inline bool perf_paranoid_cpu(void)
66 return sysctl_perf_event_paranoid > 0;
69 static inline bool perf_paranoid_kernel(void)
71 return sysctl_perf_event_paranoid > 1;
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
77 * max perf event sample rate
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
81 static atomic64_t perf_event_id;
84 * Lock for (sysadmin-configurable) event reservations:
86 static DEFINE_SPINLOCK(perf_resource_lock);
89 * Architecture provided APIs - weak aliases:
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
96 void __weak hw_perf_disable(void) { barrier(); }
97 void __weak hw_perf_enable(void) { barrier(); }
99 void __weak hw_perf_event_setup(int cpu) { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104 struct perf_cpu_context *cpuctx,
105 struct perf_event_context *ctx, int cpu)
110 void __weak perf_event_print_debug(void) { }
112 static DEFINE_PER_CPU(int, perf_disable_count);
114 void __perf_disable(void)
116 __get_cpu_var(perf_disable_count)++;
119 bool __perf_enable(void)
121 return !--__get_cpu_var(perf_disable_count);
124 void perf_disable(void)
130 void perf_enable(void)
136 static void get_ctx(struct perf_event_context *ctx)
138 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
141 static void free_ctx(struct rcu_head *head)
143 struct perf_event_context *ctx;
145 ctx = container_of(head, struct perf_event_context, rcu_head);
149 static void put_ctx(struct perf_event_context *ctx)
151 if (atomic_dec_and_test(&ctx->refcount)) {
153 put_ctx(ctx->parent_ctx);
155 put_task_struct(ctx->task);
156 call_rcu(&ctx->rcu_head, free_ctx);
160 static void unclone_ctx(struct perf_event_context *ctx)
162 if (ctx->parent_ctx) {
163 put_ctx(ctx->parent_ctx);
164 ctx->parent_ctx = NULL;
169 * If we inherit events we want to return the parent event id
172 static u64 primary_event_id(struct perf_event *event)
177 id = event->parent->id;
183 * Get the perf_event_context for a task and lock it.
184 * This has to cope with with the fact that until it is locked,
185 * the context could get moved to another task.
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
190 struct perf_event_context *ctx;
194 ctx = rcu_dereference(task->perf_event_ctxp);
197 * If this context is a clone of another, it might
198 * get swapped for another underneath us by
199 * perf_event_task_sched_out, though the
200 * rcu_read_lock() protects us from any context
201 * getting freed. Lock the context and check if it
202 * got swapped before we could get the lock, and retry
203 * if so. If we locked the right context, then it
204 * can't get swapped on us any more.
206 raw_spin_lock_irqsave(&ctx->lock, *flags);
207 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
212 if (!atomic_inc_not_zero(&ctx->refcount)) {
213 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
222 * Get the context for a task and increment its pin_count so it
223 * can't get swapped to another task. This also increments its
224 * reference count so that the context can't get freed.
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
228 struct perf_event_context *ctx;
231 ctx = perf_lock_task_context(task, &flags);
234 raw_spin_unlock_irqrestore(&ctx->lock, flags);
239 static void perf_unpin_context(struct perf_event_context *ctx)
243 raw_spin_lock_irqsave(&ctx->lock, flags);
245 raw_spin_unlock_irqrestore(&ctx->lock, flags);
249 static inline u64 perf_clock(void)
251 return cpu_clock(smp_processor_id());
255 * Update the record of the current time in a context.
257 static void update_context_time(struct perf_event_context *ctx)
259 u64 now = perf_clock();
261 ctx->time += now - ctx->timestamp;
262 ctx->timestamp = now;
266 * Update the total_time_enabled and total_time_running fields for a event.
268 static void update_event_times(struct perf_event *event)
270 struct perf_event_context *ctx = event->ctx;
273 if (event->state < PERF_EVENT_STATE_INACTIVE ||
274 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
280 run_end = event->tstamp_stopped;
282 event->total_time_enabled = run_end - event->tstamp_enabled;
284 if (event->state == PERF_EVENT_STATE_INACTIVE)
285 run_end = event->tstamp_stopped;
289 event->total_time_running = run_end - event->tstamp_running;
292 static struct list_head *
293 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
295 if (event->attr.pinned)
296 return &ctx->pinned_groups;
298 return &ctx->flexible_groups;
302 * Add a event from the lists for its context.
303 * Must be called with ctx->mutex and ctx->lock held.
306 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
308 struct perf_event *group_leader = event->group_leader;
311 * Depending on whether it is a standalone or sibling event,
312 * add it straight to the context's event list, or to the group
313 * leader's sibling list:
315 if (group_leader == event) {
316 struct list_head *list;
318 if (is_software_event(event))
319 event->group_flags |= PERF_GROUP_SOFTWARE;
321 list = ctx_group_list(event, ctx);
322 list_add_tail(&event->group_entry, list);
324 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
325 !is_software_event(event))
326 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
328 list_add_tail(&event->group_entry, &group_leader->sibling_list);
329 group_leader->nr_siblings++;
332 list_add_rcu(&event->event_entry, &ctx->event_list);
334 if (event->attr.inherit_stat)
339 * Remove a event from the lists for its context.
340 * Must be called with ctx->mutex and ctx->lock held.
343 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
345 struct perf_event *sibling, *tmp;
347 if (list_empty(&event->group_entry))
350 if (event->attr.inherit_stat)
353 list_del_init(&event->group_entry);
354 list_del_rcu(&event->event_entry);
356 if (event->group_leader != event)
357 event->group_leader->nr_siblings--;
359 update_event_times(event);
362 * If event was in error state, then keep it
363 * that way, otherwise bogus counts will be
364 * returned on read(). The only way to get out
365 * of error state is by explicit re-enabling
368 if (event->state > PERF_EVENT_STATE_OFF)
369 event->state = PERF_EVENT_STATE_OFF;
372 * If this was a group event with sibling events then
373 * upgrade the siblings to singleton events by adding them
374 * to the context list directly:
376 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
377 struct list_head *list;
379 list = ctx_group_list(event, ctx);
380 list_move_tail(&sibling->group_entry, list);
381 sibling->group_leader = sibling;
383 /* Inherit group flags from the previous leader */
384 sibling->group_flags = event->group_flags;
389 event_sched_out(struct perf_event *event,
390 struct perf_cpu_context *cpuctx,
391 struct perf_event_context *ctx)
393 if (event->state != PERF_EVENT_STATE_ACTIVE)
396 event->state = PERF_EVENT_STATE_INACTIVE;
397 if (event->pending_disable) {
398 event->pending_disable = 0;
399 event->state = PERF_EVENT_STATE_OFF;
401 event->tstamp_stopped = ctx->time;
402 event->pmu->disable(event);
405 if (!is_software_event(event))
406 cpuctx->active_oncpu--;
408 if (event->attr.exclusive || !cpuctx->active_oncpu)
409 cpuctx->exclusive = 0;
413 group_sched_out(struct perf_event *group_event,
414 struct perf_cpu_context *cpuctx,
415 struct perf_event_context *ctx)
417 struct perf_event *event;
419 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
422 event_sched_out(group_event, cpuctx, ctx);
425 * Schedule out siblings (if any):
427 list_for_each_entry(event, &group_event->sibling_list, group_entry)
428 event_sched_out(event, cpuctx, ctx);
430 if (group_event->attr.exclusive)
431 cpuctx->exclusive = 0;
435 * Cross CPU call to remove a performance event
437 * We disable the event on the hardware level first. After that we
438 * remove it from the context list.
440 static void __perf_event_remove_from_context(void *info)
442 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
443 struct perf_event *event = info;
444 struct perf_event_context *ctx = event->ctx;
447 * If this is a task context, we need to check whether it is
448 * the current task context of this cpu. If not it has been
449 * scheduled out before the smp call arrived.
451 if (ctx->task && cpuctx->task_ctx != ctx)
454 raw_spin_lock(&ctx->lock);
456 * Protect the list operation against NMI by disabling the
457 * events on a global level.
461 event_sched_out(event, cpuctx, ctx);
463 list_del_event(event, ctx);
467 * Allow more per task events with respect to the
470 cpuctx->max_pertask =
471 min(perf_max_events - ctx->nr_events,
472 perf_max_events - perf_reserved_percpu);
476 raw_spin_unlock(&ctx->lock);
481 * Remove the event from a task's (or a CPU's) list of events.
483 * Must be called with ctx->mutex held.
485 * CPU events are removed with a smp call. For task events we only
486 * call when the task is on a CPU.
488 * If event->ctx is a cloned context, callers must make sure that
489 * every task struct that event->ctx->task could possibly point to
490 * remains valid. This is OK when called from perf_release since
491 * that only calls us on the top-level context, which can't be a clone.
492 * When called from perf_event_exit_task, it's OK because the
493 * context has been detached from its task.
495 static void perf_event_remove_from_context(struct perf_event *event)
497 struct perf_event_context *ctx = event->ctx;
498 struct task_struct *task = ctx->task;
502 * Per cpu events are removed via an smp call and
503 * the removal is always successful.
505 smp_call_function_single(event->cpu,
506 __perf_event_remove_from_context,
512 task_oncpu_function_call(task, __perf_event_remove_from_context,
515 raw_spin_lock_irq(&ctx->lock);
517 * If the context is active we need to retry the smp call.
519 if (ctx->nr_active && !list_empty(&event->group_entry)) {
520 raw_spin_unlock_irq(&ctx->lock);
525 * The lock prevents that this context is scheduled in so we
526 * can remove the event safely, if the call above did not
529 if (!list_empty(&event->group_entry))
530 list_del_event(event, ctx);
531 raw_spin_unlock_irq(&ctx->lock);
535 * Update total_time_enabled and total_time_running for all events in a group.
537 static void update_group_times(struct perf_event *leader)
539 struct perf_event *event;
541 update_event_times(leader);
542 list_for_each_entry(event, &leader->sibling_list, group_entry)
543 update_event_times(event);
547 * Cross CPU call to disable a performance event
549 static void __perf_event_disable(void *info)
551 struct perf_event *event = info;
552 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
553 struct perf_event_context *ctx = event->ctx;
556 * If this is a per-task event, need to check whether this
557 * event's task is the current task on this cpu.
559 if (ctx->task && cpuctx->task_ctx != ctx)
562 raw_spin_lock(&ctx->lock);
565 * If the event is on, turn it off.
566 * If it is in error state, leave it in error state.
568 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
569 update_context_time(ctx);
570 update_group_times(event);
571 if (event == event->group_leader)
572 group_sched_out(event, cpuctx, ctx);
574 event_sched_out(event, cpuctx, ctx);
575 event->state = PERF_EVENT_STATE_OFF;
578 raw_spin_unlock(&ctx->lock);
584 * If event->ctx is a cloned context, callers must make sure that
585 * every task struct that event->ctx->task could possibly point to
586 * remains valid. This condition is satisifed when called through
587 * perf_event_for_each_child or perf_event_for_each because they
588 * hold the top-level event's child_mutex, so any descendant that
589 * goes to exit will block in sync_child_event.
590 * When called from perf_pending_event it's OK because event->ctx
591 * is the current context on this CPU and preemption is disabled,
592 * hence we can't get into perf_event_task_sched_out for this context.
594 void perf_event_disable(struct perf_event *event)
596 struct perf_event_context *ctx = event->ctx;
597 struct task_struct *task = ctx->task;
601 * Disable the event on the cpu that it's on
603 smp_call_function_single(event->cpu, __perf_event_disable,
609 task_oncpu_function_call(task, __perf_event_disable, event);
611 raw_spin_lock_irq(&ctx->lock);
613 * If the event is still active, we need to retry the cross-call.
615 if (event->state == PERF_EVENT_STATE_ACTIVE) {
616 raw_spin_unlock_irq(&ctx->lock);
621 * Since we have the lock this context can't be scheduled
622 * in, so we can change the state safely.
624 if (event->state == PERF_EVENT_STATE_INACTIVE) {
625 update_group_times(event);
626 event->state = PERF_EVENT_STATE_OFF;
629 raw_spin_unlock_irq(&ctx->lock);
633 event_sched_in(struct perf_event *event,
634 struct perf_cpu_context *cpuctx,
635 struct perf_event_context *ctx,
638 if (event->state <= PERF_EVENT_STATE_OFF)
641 event->state = PERF_EVENT_STATE_ACTIVE;
642 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
644 * The new state must be visible before we turn it on in the hardware:
648 if (event->pmu->enable(event)) {
649 event->state = PERF_EVENT_STATE_INACTIVE;
654 event->tstamp_running += ctx->time - event->tstamp_stopped;
656 if (!is_software_event(event))
657 cpuctx->active_oncpu++;
660 if (event->attr.exclusive)
661 cpuctx->exclusive = 1;
667 group_sched_in(struct perf_event *group_event,
668 struct perf_cpu_context *cpuctx,
669 struct perf_event_context *ctx,
672 struct perf_event *event, *partial_group;
675 if (group_event->state == PERF_EVENT_STATE_OFF)
678 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
680 return ret < 0 ? ret : 0;
682 if (event_sched_in(group_event, cpuctx, ctx, cpu))
686 * Schedule in siblings as one group (if any):
688 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
689 if (event_sched_in(event, cpuctx, ctx, cpu)) {
690 partial_group = event;
699 * Groups can be scheduled in as one unit only, so undo any
700 * partial group before returning:
702 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
703 if (event == partial_group)
705 event_sched_out(event, cpuctx, ctx);
707 event_sched_out(group_event, cpuctx, ctx);
713 * Work out whether we can put this event group on the CPU now.
715 static int group_can_go_on(struct perf_event *event,
716 struct perf_cpu_context *cpuctx,
720 * Groups consisting entirely of software events can always go on.
722 if (event->group_flags & PERF_GROUP_SOFTWARE)
725 * If an exclusive group is already on, no other hardware
728 if (cpuctx->exclusive)
731 * If this group is exclusive and there are already
732 * events on the CPU, it can't go on.
734 if (event->attr.exclusive && cpuctx->active_oncpu)
737 * Otherwise, try to add it if all previous groups were able
743 static void add_event_to_ctx(struct perf_event *event,
744 struct perf_event_context *ctx)
746 list_add_event(event, ctx);
747 event->tstamp_enabled = ctx->time;
748 event->tstamp_running = ctx->time;
749 event->tstamp_stopped = ctx->time;
753 * Cross CPU call to install and enable a performance event
755 * Must be called with ctx->mutex held
757 static void __perf_install_in_context(void *info)
759 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
760 struct perf_event *event = info;
761 struct perf_event_context *ctx = event->ctx;
762 struct perf_event *leader = event->group_leader;
763 int cpu = smp_processor_id();
767 * If this is a task context, we need to check whether it is
768 * the current task context of this cpu. If not it has been
769 * scheduled out before the smp call arrived.
770 * Or possibly this is the right context but it isn't
771 * on this cpu because it had no events.
773 if (ctx->task && cpuctx->task_ctx != ctx) {
774 if (cpuctx->task_ctx || ctx->task != current)
776 cpuctx->task_ctx = ctx;
779 raw_spin_lock(&ctx->lock);
781 update_context_time(ctx);
784 * Protect the list operation against NMI by disabling the
785 * events on a global level. NOP for non NMI based events.
789 add_event_to_ctx(event, ctx);
791 if (event->cpu != -1 && event->cpu != smp_processor_id())
795 * Don't put the event on if it is disabled or if
796 * it is in a group and the group isn't on.
798 if (event->state != PERF_EVENT_STATE_INACTIVE ||
799 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
803 * An exclusive event can't go on if there are already active
804 * hardware events, and no hardware event can go on if there
805 * is already an exclusive event on.
807 if (!group_can_go_on(event, cpuctx, 1))
810 err = event_sched_in(event, cpuctx, ctx, cpu);
814 * This event couldn't go on. If it is in a group
815 * then we have to pull the whole group off.
816 * If the event group is pinned then put it in error state.
819 group_sched_out(leader, cpuctx, ctx);
820 if (leader->attr.pinned) {
821 update_group_times(leader);
822 leader->state = PERF_EVENT_STATE_ERROR;
826 if (!err && !ctx->task && cpuctx->max_pertask)
827 cpuctx->max_pertask--;
832 raw_spin_unlock(&ctx->lock);
836 * Attach a performance event to a context
838 * First we add the event to the list with the hardware enable bit
839 * in event->hw_config cleared.
841 * If the event is attached to a task which is on a CPU we use a smp
842 * call to enable it in the task context. The task might have been
843 * scheduled away, but we check this in the smp call again.
845 * Must be called with ctx->mutex held.
848 perf_install_in_context(struct perf_event_context *ctx,
849 struct perf_event *event,
852 struct task_struct *task = ctx->task;
856 * Per cpu events are installed via an smp call and
857 * the install is always successful.
859 smp_call_function_single(cpu, __perf_install_in_context,
865 task_oncpu_function_call(task, __perf_install_in_context,
868 raw_spin_lock_irq(&ctx->lock);
870 * we need to retry the smp call.
872 if (ctx->is_active && list_empty(&event->group_entry)) {
873 raw_spin_unlock_irq(&ctx->lock);
878 * The lock prevents that this context is scheduled in so we
879 * can add the event safely, if it the call above did not
882 if (list_empty(&event->group_entry))
883 add_event_to_ctx(event, ctx);
884 raw_spin_unlock_irq(&ctx->lock);
888 * Put a event into inactive state and update time fields.
889 * Enabling the leader of a group effectively enables all
890 * the group members that aren't explicitly disabled, so we
891 * have to update their ->tstamp_enabled also.
892 * Note: this works for group members as well as group leaders
893 * since the non-leader members' sibling_lists will be empty.
895 static void __perf_event_mark_enabled(struct perf_event *event,
896 struct perf_event_context *ctx)
898 struct perf_event *sub;
900 event->state = PERF_EVENT_STATE_INACTIVE;
901 event->tstamp_enabled = ctx->time - event->total_time_enabled;
902 list_for_each_entry(sub, &event->sibling_list, group_entry)
903 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
904 sub->tstamp_enabled =
905 ctx->time - sub->total_time_enabled;
909 * Cross CPU call to enable a performance event
911 static void __perf_event_enable(void *info)
913 struct perf_event *event = info;
914 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
915 struct perf_event_context *ctx = event->ctx;
916 struct perf_event *leader = event->group_leader;
920 * If this is a per-task event, need to check whether this
921 * event's task is the current task on this cpu.
923 if (ctx->task && cpuctx->task_ctx != ctx) {
924 if (cpuctx->task_ctx || ctx->task != current)
926 cpuctx->task_ctx = ctx;
929 raw_spin_lock(&ctx->lock);
931 update_context_time(ctx);
933 if (event->state >= PERF_EVENT_STATE_INACTIVE)
935 __perf_event_mark_enabled(event, ctx);
937 if (event->cpu != -1 && event->cpu != smp_processor_id())
941 * If the event is in a group and isn't the group leader,
942 * then don't put it on unless the group is on.
944 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
947 if (!group_can_go_on(event, cpuctx, 1)) {
952 err = group_sched_in(event, cpuctx, ctx,
955 err = event_sched_in(event, cpuctx, ctx,
962 * If this event can't go on and it's part of a
963 * group, then the whole group has to come off.
966 group_sched_out(leader, cpuctx, ctx);
967 if (leader->attr.pinned) {
968 update_group_times(leader);
969 leader->state = PERF_EVENT_STATE_ERROR;
974 raw_spin_unlock(&ctx->lock);
980 * If event->ctx is a cloned context, callers must make sure that
981 * every task struct that event->ctx->task could possibly point to
982 * remains valid. This condition is satisfied when called through
983 * perf_event_for_each_child or perf_event_for_each as described
984 * for perf_event_disable.
986 void perf_event_enable(struct perf_event *event)
988 struct perf_event_context *ctx = event->ctx;
989 struct task_struct *task = ctx->task;
993 * Enable the event on the cpu that it's on
995 smp_call_function_single(event->cpu, __perf_event_enable,
1000 raw_spin_lock_irq(&ctx->lock);
1001 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1005 * If the event is in error state, clear that first.
1006 * That way, if we see the event in error state below, we
1007 * know that it has gone back into error state, as distinct
1008 * from the task having been scheduled away before the
1009 * cross-call arrived.
1011 if (event->state == PERF_EVENT_STATE_ERROR)
1012 event->state = PERF_EVENT_STATE_OFF;
1015 raw_spin_unlock_irq(&ctx->lock);
1016 task_oncpu_function_call(task, __perf_event_enable, event);
1018 raw_spin_lock_irq(&ctx->lock);
1021 * If the context is active and the event is still off,
1022 * we need to retry the cross-call.
1024 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1028 * Since we have the lock this context can't be scheduled
1029 * in, so we can change the state safely.
1031 if (event->state == PERF_EVENT_STATE_OFF)
1032 __perf_event_mark_enabled(event, ctx);
1035 raw_spin_unlock_irq(&ctx->lock);
1038 static int perf_event_refresh(struct perf_event *event, int refresh)
1041 * not supported on inherited events
1043 if (event->attr.inherit)
1046 atomic_add(refresh, &event->event_limit);
1047 perf_event_enable(event);
1053 EVENT_FLEXIBLE = 0x1,
1055 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
1058 static void ctx_sched_out(struct perf_event_context *ctx,
1059 struct perf_cpu_context *cpuctx,
1060 enum event_type_t event_type)
1062 struct perf_event *event;
1064 raw_spin_lock(&ctx->lock);
1066 if (likely(!ctx->nr_events))
1068 update_context_time(ctx);
1071 if (!ctx->nr_active)
1074 if (event_type & EVENT_PINNED)
1075 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1076 group_sched_out(event, cpuctx, ctx);
1078 if (event_type & EVENT_FLEXIBLE)
1079 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1080 group_sched_out(event, cpuctx, ctx);
1085 raw_spin_unlock(&ctx->lock);
1089 * Test whether two contexts are equivalent, i.e. whether they
1090 * have both been cloned from the same version of the same context
1091 * and they both have the same number of enabled events.
1092 * If the number of enabled events is the same, then the set
1093 * of enabled events should be the same, because these are both
1094 * inherited contexts, therefore we can't access individual events
1095 * in them directly with an fd; we can only enable/disable all
1096 * events via prctl, or enable/disable all events in a family
1097 * via ioctl, which will have the same effect on both contexts.
1099 static int context_equiv(struct perf_event_context *ctx1,
1100 struct perf_event_context *ctx2)
1102 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1103 && ctx1->parent_gen == ctx2->parent_gen
1104 && !ctx1->pin_count && !ctx2->pin_count;
1107 static void __perf_event_sync_stat(struct perf_event *event,
1108 struct perf_event *next_event)
1112 if (!event->attr.inherit_stat)
1116 * Update the event value, we cannot use perf_event_read()
1117 * because we're in the middle of a context switch and have IRQs
1118 * disabled, which upsets smp_call_function_single(), however
1119 * we know the event must be on the current CPU, therefore we
1120 * don't need to use it.
1122 switch (event->state) {
1123 case PERF_EVENT_STATE_ACTIVE:
1124 event->pmu->read(event);
1127 case PERF_EVENT_STATE_INACTIVE:
1128 update_event_times(event);
1136 * In order to keep per-task stats reliable we need to flip the event
1137 * values when we flip the contexts.
1139 value = atomic64_read(&next_event->count);
1140 value = atomic64_xchg(&event->count, value);
1141 atomic64_set(&next_event->count, value);
1143 swap(event->total_time_enabled, next_event->total_time_enabled);
1144 swap(event->total_time_running, next_event->total_time_running);
1147 * Since we swizzled the values, update the user visible data too.
1149 perf_event_update_userpage(event);
1150 perf_event_update_userpage(next_event);
1153 #define list_next_entry(pos, member) \
1154 list_entry(pos->member.next, typeof(*pos), member)
1156 static void perf_event_sync_stat(struct perf_event_context *ctx,
1157 struct perf_event_context *next_ctx)
1159 struct perf_event *event, *next_event;
1164 update_context_time(ctx);
1166 event = list_first_entry(&ctx->event_list,
1167 struct perf_event, event_entry);
1169 next_event = list_first_entry(&next_ctx->event_list,
1170 struct perf_event, event_entry);
1172 while (&event->event_entry != &ctx->event_list &&
1173 &next_event->event_entry != &next_ctx->event_list) {
1175 __perf_event_sync_stat(event, next_event);
1177 event = list_next_entry(event, event_entry);
1178 next_event = list_next_entry(next_event, event_entry);
1183 * Called from scheduler to remove the events of the current task,
1184 * with interrupts disabled.
1186 * We stop each event and update the event value in event->count.
1188 * This does not protect us against NMI, but disable()
1189 * sets the disabled bit in the control field of event _before_
1190 * accessing the event control register. If a NMI hits, then it will
1191 * not restart the event.
1193 void perf_event_task_sched_out(struct task_struct *task,
1194 struct task_struct *next)
1196 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1197 struct perf_event_context *ctx = task->perf_event_ctxp;
1198 struct perf_event_context *next_ctx;
1199 struct perf_event_context *parent;
1200 struct pt_regs *regs;
1203 regs = task_pt_regs(task);
1204 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1206 if (likely(!ctx || !cpuctx->task_ctx))
1210 parent = rcu_dereference(ctx->parent_ctx);
1211 next_ctx = next->perf_event_ctxp;
1212 if (parent && next_ctx &&
1213 rcu_dereference(next_ctx->parent_ctx) == parent) {
1215 * Looks like the two contexts are clones, so we might be
1216 * able to optimize the context switch. We lock both
1217 * contexts and check that they are clones under the
1218 * lock (including re-checking that neither has been
1219 * uncloned in the meantime). It doesn't matter which
1220 * order we take the locks because no other cpu could
1221 * be trying to lock both of these tasks.
1223 raw_spin_lock(&ctx->lock);
1224 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1225 if (context_equiv(ctx, next_ctx)) {
1227 * XXX do we need a memory barrier of sorts
1228 * wrt to rcu_dereference() of perf_event_ctxp
1230 task->perf_event_ctxp = next_ctx;
1231 next->perf_event_ctxp = ctx;
1233 next_ctx->task = task;
1236 perf_event_sync_stat(ctx, next_ctx);
1238 raw_spin_unlock(&next_ctx->lock);
1239 raw_spin_unlock(&ctx->lock);
1244 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1245 cpuctx->task_ctx = NULL;
1249 static void task_ctx_sched_out(struct perf_event_context *ctx,
1250 enum event_type_t event_type)
1252 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1254 if (!cpuctx->task_ctx)
1257 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1260 ctx_sched_out(ctx, cpuctx, event_type);
1261 cpuctx->task_ctx = NULL;
1265 * Called with IRQs disabled
1267 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1269 task_ctx_sched_out(ctx, EVENT_ALL);
1273 * Called with IRQs disabled
1275 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1276 enum event_type_t event_type)
1278 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1282 ctx_pinned_sched_in(struct perf_event_context *ctx,
1283 struct perf_cpu_context *cpuctx,
1286 struct perf_event *event;
1288 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1289 if (event->state <= PERF_EVENT_STATE_OFF)
1291 if (event->cpu != -1 && event->cpu != cpu)
1294 if (group_can_go_on(event, cpuctx, 1))
1295 group_sched_in(event, cpuctx, ctx, cpu);
1298 * If this pinned group hasn't been scheduled,
1299 * put it in error state.
1301 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1302 update_group_times(event);
1303 event->state = PERF_EVENT_STATE_ERROR;
1309 ctx_flexible_sched_in(struct perf_event_context *ctx,
1310 struct perf_cpu_context *cpuctx,
1313 struct perf_event *event;
1316 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1317 /* Ignore events in OFF or ERROR state */
1318 if (event->state <= PERF_EVENT_STATE_OFF)
1321 * Listen to the 'cpu' scheduling filter constraint
1324 if (event->cpu != -1 && event->cpu != cpu)
1327 if (group_can_go_on(event, cpuctx, can_add_hw))
1328 if (group_sched_in(event, cpuctx, ctx, cpu))
1334 ctx_sched_in(struct perf_event_context *ctx,
1335 struct perf_cpu_context *cpuctx,
1336 enum event_type_t event_type)
1338 int cpu = smp_processor_id();
1340 raw_spin_lock(&ctx->lock);
1342 if (likely(!ctx->nr_events))
1345 ctx->timestamp = perf_clock();
1350 * First go through the list and put on any pinned groups
1351 * in order to give them the best chance of going on.
1353 if (event_type & EVENT_PINNED)
1354 ctx_pinned_sched_in(ctx, cpuctx, cpu);
1356 /* Then walk through the lower prio flexible groups */
1357 if (event_type & EVENT_FLEXIBLE)
1358 ctx_flexible_sched_in(ctx, cpuctx, cpu);
1362 raw_spin_unlock(&ctx->lock);
1365 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1366 enum event_type_t event_type)
1368 struct perf_event_context *ctx = &cpuctx->ctx;
1370 ctx_sched_in(ctx, cpuctx, event_type);
1373 static void task_ctx_sched_in(struct task_struct *task,
1374 enum event_type_t event_type)
1376 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1377 struct perf_event_context *ctx = task->perf_event_ctxp;
1381 if (cpuctx->task_ctx == ctx)
1383 ctx_sched_in(ctx, cpuctx, event_type);
1384 cpuctx->task_ctx = ctx;
1387 * Called from scheduler to add the events of the current task
1388 * with interrupts disabled.
1390 * We restore the event value and then enable it.
1392 * This does not protect us against NMI, but enable()
1393 * sets the enabled bit in the control field of event _before_
1394 * accessing the event control register. If a NMI hits, then it will
1395 * keep the event running.
1397 void perf_event_task_sched_in(struct task_struct *task)
1399 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1400 struct perf_event_context *ctx = task->perf_event_ctxp;
1405 if (cpuctx->task_ctx == ctx)
1409 * We want to keep the following priority order:
1410 * cpu pinned (that don't need to move), task pinned,
1411 * cpu flexible, task flexible.
1413 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1415 ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1416 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1417 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1419 cpuctx->task_ctx = ctx;
1422 #define MAX_INTERRUPTS (~0ULL)
1424 static void perf_log_throttle(struct perf_event *event, int enable);
1426 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1428 u64 frequency = event->attr.sample_freq;
1429 u64 sec = NSEC_PER_SEC;
1430 u64 divisor, dividend;
1432 int count_fls, nsec_fls, frequency_fls, sec_fls;
1434 count_fls = fls64(count);
1435 nsec_fls = fls64(nsec);
1436 frequency_fls = fls64(frequency);
1440 * We got @count in @nsec, with a target of sample_freq HZ
1441 * the target period becomes:
1444 * period = -------------------
1445 * @nsec * sample_freq
1450 * Reduce accuracy by one bit such that @a and @b converge
1451 * to a similar magnitude.
1453 #define REDUCE_FLS(a, b) \
1455 if (a##_fls > b##_fls) { \
1465 * Reduce accuracy until either term fits in a u64, then proceed with
1466 * the other, so that finally we can do a u64/u64 division.
1468 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1469 REDUCE_FLS(nsec, frequency);
1470 REDUCE_FLS(sec, count);
1473 if (count_fls + sec_fls > 64) {
1474 divisor = nsec * frequency;
1476 while (count_fls + sec_fls > 64) {
1477 REDUCE_FLS(count, sec);
1481 dividend = count * sec;
1483 dividend = count * sec;
1485 while (nsec_fls + frequency_fls > 64) {
1486 REDUCE_FLS(nsec, frequency);
1490 divisor = nsec * frequency;
1493 return div64_u64(dividend, divisor);
1496 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1498 struct hw_perf_event *hwc = &event->hw;
1499 u64 period, sample_period;
1502 period = perf_calculate_period(event, nsec, count);
1504 delta = (s64)(period - hwc->sample_period);
1505 delta = (delta + 7) / 8; /* low pass filter */
1507 sample_period = hwc->sample_period + delta;
1512 hwc->sample_period = sample_period;
1514 if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1516 event->pmu->disable(event);
1517 atomic64_set(&hwc->period_left, 0);
1518 event->pmu->enable(event);
1523 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1525 struct perf_event *event;
1526 struct hw_perf_event *hwc;
1527 u64 interrupts, now;
1530 raw_spin_lock(&ctx->lock);
1531 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1532 if (event->state != PERF_EVENT_STATE_ACTIVE)
1535 if (event->cpu != -1 && event->cpu != smp_processor_id())
1540 interrupts = hwc->interrupts;
1541 hwc->interrupts = 0;
1544 * unthrottle events on the tick
1546 if (interrupts == MAX_INTERRUPTS) {
1547 perf_log_throttle(event, 1);
1548 event->pmu->unthrottle(event);
1551 if (!event->attr.freq || !event->attr.sample_freq)
1554 event->pmu->read(event);
1555 now = atomic64_read(&event->count);
1556 delta = now - hwc->freq_count_stamp;
1557 hwc->freq_count_stamp = now;
1560 perf_adjust_period(event, TICK_NSEC, delta);
1562 raw_spin_unlock(&ctx->lock);
1566 * Round-robin a context's events:
1568 static void rotate_ctx(struct perf_event_context *ctx)
1570 if (!ctx->nr_events)
1573 raw_spin_lock(&ctx->lock);
1575 /* Rotate the first entry last of non-pinned groups */
1576 list_rotate_left(&ctx->flexible_groups);
1578 raw_spin_unlock(&ctx->lock);
1581 void perf_event_task_tick(struct task_struct *curr)
1583 struct perf_cpu_context *cpuctx;
1584 struct perf_event_context *ctx;
1586 if (!atomic_read(&nr_events))
1589 cpuctx = &__get_cpu_var(perf_cpu_context);
1590 ctx = curr->perf_event_ctxp;
1594 perf_ctx_adjust_freq(&cpuctx->ctx);
1596 perf_ctx_adjust_freq(ctx);
1598 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1600 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1602 rotate_ctx(&cpuctx->ctx);
1606 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1608 task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1613 static int event_enable_on_exec(struct perf_event *event,
1614 struct perf_event_context *ctx)
1616 if (!event->attr.enable_on_exec)
1619 event->attr.enable_on_exec = 0;
1620 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1623 __perf_event_mark_enabled(event, ctx);
1629 * Enable all of a task's events that have been marked enable-on-exec.
1630 * This expects task == current.
1632 static void perf_event_enable_on_exec(struct task_struct *task)
1634 struct perf_event_context *ctx;
1635 struct perf_event *event;
1636 unsigned long flags;
1640 local_irq_save(flags);
1641 ctx = task->perf_event_ctxp;
1642 if (!ctx || !ctx->nr_events)
1645 __perf_event_task_sched_out(ctx);
1647 raw_spin_lock(&ctx->lock);
1649 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1650 ret = event_enable_on_exec(event, ctx);
1655 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1656 ret = event_enable_on_exec(event, ctx);
1662 * Unclone this context if we enabled any event.
1667 raw_spin_unlock(&ctx->lock);
1669 perf_event_task_sched_in(task);
1671 local_irq_restore(flags);
1675 * Cross CPU call to read the hardware event
1677 static void __perf_event_read(void *info)
1679 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1680 struct perf_event *event = info;
1681 struct perf_event_context *ctx = event->ctx;
1684 * If this is a task context, we need to check whether it is
1685 * the current task context of this cpu. If not it has been
1686 * scheduled out before the smp call arrived. In that case
1687 * event->count would have been updated to a recent sample
1688 * when the event was scheduled out.
1690 if (ctx->task && cpuctx->task_ctx != ctx)
1693 raw_spin_lock(&ctx->lock);
1694 update_context_time(ctx);
1695 update_event_times(event);
1696 raw_spin_unlock(&ctx->lock);
1698 event->pmu->read(event);
1701 static u64 perf_event_read(struct perf_event *event)
1704 * If event is enabled and currently active on a CPU, update the
1705 * value in the event structure:
1707 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1708 smp_call_function_single(event->oncpu,
1709 __perf_event_read, event, 1);
1710 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1711 struct perf_event_context *ctx = event->ctx;
1712 unsigned long flags;
1714 raw_spin_lock_irqsave(&ctx->lock, flags);
1715 update_context_time(ctx);
1716 update_event_times(event);
1717 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1720 return atomic64_read(&event->count);
1724 * Initialize the perf_event context in a task_struct:
1727 __perf_event_init_context(struct perf_event_context *ctx,
1728 struct task_struct *task)
1730 raw_spin_lock_init(&ctx->lock);
1731 mutex_init(&ctx->mutex);
1732 INIT_LIST_HEAD(&ctx->pinned_groups);
1733 INIT_LIST_HEAD(&ctx->flexible_groups);
1734 INIT_LIST_HEAD(&ctx->event_list);
1735 atomic_set(&ctx->refcount, 1);
1739 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1741 struct perf_event_context *ctx;
1742 struct perf_cpu_context *cpuctx;
1743 struct task_struct *task;
1744 unsigned long flags;
1747 if (pid == -1 && cpu != -1) {
1748 /* Must be root to operate on a CPU event: */
1749 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1750 return ERR_PTR(-EACCES);
1752 if (cpu < 0 || cpu >= nr_cpumask_bits)
1753 return ERR_PTR(-EINVAL);
1756 * We could be clever and allow to attach a event to an
1757 * offline CPU and activate it when the CPU comes up, but
1760 if (!cpu_online(cpu))
1761 return ERR_PTR(-ENODEV);
1763 cpuctx = &per_cpu(perf_cpu_context, cpu);
1774 task = find_task_by_vpid(pid);
1776 get_task_struct(task);
1780 return ERR_PTR(-ESRCH);
1783 * Can't attach events to a dying task.
1786 if (task->flags & PF_EXITING)
1789 /* Reuse ptrace permission checks for now. */
1791 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1795 ctx = perf_lock_task_context(task, &flags);
1798 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1802 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1806 __perf_event_init_context(ctx, task);
1808 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1810 * We raced with some other task; use
1811 * the context they set.
1816 get_task_struct(task);
1819 put_task_struct(task);
1823 put_task_struct(task);
1824 return ERR_PTR(err);
1827 static void perf_event_free_filter(struct perf_event *event);
1829 static void free_event_rcu(struct rcu_head *head)
1831 struct perf_event *event;
1833 event = container_of(head, struct perf_event, rcu_head);
1835 put_pid_ns(event->ns);
1836 perf_event_free_filter(event);
1840 static void perf_pending_sync(struct perf_event *event);
1842 static void free_event(struct perf_event *event)
1844 perf_pending_sync(event);
1846 if (!event->parent) {
1847 atomic_dec(&nr_events);
1848 if (event->attr.mmap)
1849 atomic_dec(&nr_mmap_events);
1850 if (event->attr.comm)
1851 atomic_dec(&nr_comm_events);
1852 if (event->attr.task)
1853 atomic_dec(&nr_task_events);
1856 if (event->output) {
1857 fput(event->output->filp);
1858 event->output = NULL;
1862 event->destroy(event);
1864 put_ctx(event->ctx);
1865 call_rcu(&event->rcu_head, free_event_rcu);
1868 int perf_event_release_kernel(struct perf_event *event)
1870 struct perf_event_context *ctx = event->ctx;
1872 WARN_ON_ONCE(ctx->parent_ctx);
1873 mutex_lock(&ctx->mutex);
1874 perf_event_remove_from_context(event);
1875 mutex_unlock(&ctx->mutex);
1877 mutex_lock(&event->owner->perf_event_mutex);
1878 list_del_init(&event->owner_entry);
1879 mutex_unlock(&event->owner->perf_event_mutex);
1880 put_task_struct(event->owner);
1886 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1889 * Called when the last reference to the file is gone.
1891 static int perf_release(struct inode *inode, struct file *file)
1893 struct perf_event *event = file->private_data;
1895 file->private_data = NULL;
1897 return perf_event_release_kernel(event);
1900 static int perf_event_read_size(struct perf_event *event)
1902 int entry = sizeof(u64); /* value */
1906 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1907 size += sizeof(u64);
1909 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1910 size += sizeof(u64);
1912 if (event->attr.read_format & PERF_FORMAT_ID)
1913 entry += sizeof(u64);
1915 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1916 nr += event->group_leader->nr_siblings;
1917 size += sizeof(u64);
1925 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1927 struct perf_event *child;
1933 mutex_lock(&event->child_mutex);
1934 total += perf_event_read(event);
1935 *enabled += event->total_time_enabled +
1936 atomic64_read(&event->child_total_time_enabled);
1937 *running += event->total_time_running +
1938 atomic64_read(&event->child_total_time_running);
1940 list_for_each_entry(child, &event->child_list, child_list) {
1941 total += perf_event_read(child);
1942 *enabled += child->total_time_enabled;
1943 *running += child->total_time_running;
1945 mutex_unlock(&event->child_mutex);
1949 EXPORT_SYMBOL_GPL(perf_event_read_value);
1951 static int perf_event_read_group(struct perf_event *event,
1952 u64 read_format, char __user *buf)
1954 struct perf_event *leader = event->group_leader, *sub;
1955 int n = 0, size = 0, ret = -EFAULT;
1956 struct perf_event_context *ctx = leader->ctx;
1958 u64 count, enabled, running;
1960 mutex_lock(&ctx->mutex);
1961 count = perf_event_read_value(leader, &enabled, &running);
1963 values[n++] = 1 + leader->nr_siblings;
1964 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1965 values[n++] = enabled;
1966 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1967 values[n++] = running;
1968 values[n++] = count;
1969 if (read_format & PERF_FORMAT_ID)
1970 values[n++] = primary_event_id(leader);
1972 size = n * sizeof(u64);
1974 if (copy_to_user(buf, values, size))
1979 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1982 values[n++] = perf_event_read_value(sub, &enabled, &running);
1983 if (read_format & PERF_FORMAT_ID)
1984 values[n++] = primary_event_id(sub);
1986 size = n * sizeof(u64);
1988 if (copy_to_user(buf + ret, values, size)) {
1996 mutex_unlock(&ctx->mutex);
2001 static int perf_event_read_one(struct perf_event *event,
2002 u64 read_format, char __user *buf)
2004 u64 enabled, running;
2008 values[n++] = perf_event_read_value(event, &enabled, &running);
2009 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2010 values[n++] = enabled;
2011 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2012 values[n++] = running;
2013 if (read_format & PERF_FORMAT_ID)
2014 values[n++] = primary_event_id(event);
2016 if (copy_to_user(buf, values, n * sizeof(u64)))
2019 return n * sizeof(u64);
2023 * Read the performance event - simple non blocking version for now
2026 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
2028 u64 read_format = event->attr.read_format;
2032 * Return end-of-file for a read on a event that is in
2033 * error state (i.e. because it was pinned but it couldn't be
2034 * scheduled on to the CPU at some point).
2036 if (event->state == PERF_EVENT_STATE_ERROR)
2039 if (count < perf_event_read_size(event))
2042 WARN_ON_ONCE(event->ctx->parent_ctx);
2043 if (read_format & PERF_FORMAT_GROUP)
2044 ret = perf_event_read_group(event, read_format, buf);
2046 ret = perf_event_read_one(event, read_format, buf);
2052 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
2054 struct perf_event *event = file->private_data;
2056 return perf_read_hw(event, buf, count);
2059 static unsigned int perf_poll(struct file *file, poll_table *wait)
2061 struct perf_event *event = file->private_data;
2062 struct perf_mmap_data *data;
2063 unsigned int events = POLL_HUP;
2066 data = rcu_dereference(event->data);
2068 events = atomic_xchg(&data->poll, 0);
2071 poll_wait(file, &event->waitq, wait);
2076 static void perf_event_reset(struct perf_event *event)
2078 (void)perf_event_read(event);
2079 atomic64_set(&event->count, 0);
2080 perf_event_update_userpage(event);
2084 * Holding the top-level event's child_mutex means that any
2085 * descendant process that has inherited this event will block
2086 * in sync_child_event if it goes to exit, thus satisfying the
2087 * task existence requirements of perf_event_enable/disable.
2089 static void perf_event_for_each_child(struct perf_event *event,
2090 void (*func)(struct perf_event *))
2092 struct perf_event *child;
2094 WARN_ON_ONCE(event->ctx->parent_ctx);
2095 mutex_lock(&event->child_mutex);
2097 list_for_each_entry(child, &event->child_list, child_list)
2099 mutex_unlock(&event->child_mutex);
2102 static void perf_event_for_each(struct perf_event *event,
2103 void (*func)(struct perf_event *))
2105 struct perf_event_context *ctx = event->ctx;
2106 struct perf_event *sibling;
2108 WARN_ON_ONCE(ctx->parent_ctx);
2109 mutex_lock(&ctx->mutex);
2110 event = event->group_leader;
2112 perf_event_for_each_child(event, func);
2114 list_for_each_entry(sibling, &event->sibling_list, group_entry)
2115 perf_event_for_each_child(event, func);
2116 mutex_unlock(&ctx->mutex);
2119 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2121 struct perf_event_context *ctx = event->ctx;
2126 if (!event->attr.sample_period)
2129 size = copy_from_user(&value, arg, sizeof(value));
2130 if (size != sizeof(value))
2136 raw_spin_lock_irq(&ctx->lock);
2137 if (event->attr.freq) {
2138 if (value > sysctl_perf_event_sample_rate) {
2143 event->attr.sample_freq = value;
2145 event->attr.sample_period = value;
2146 event->hw.sample_period = value;
2149 raw_spin_unlock_irq(&ctx->lock);
2154 static int perf_event_set_output(struct perf_event *event, int output_fd);
2155 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2157 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2159 struct perf_event *event = file->private_data;
2160 void (*func)(struct perf_event *);
2164 case PERF_EVENT_IOC_ENABLE:
2165 func = perf_event_enable;
2167 case PERF_EVENT_IOC_DISABLE:
2168 func = perf_event_disable;
2170 case PERF_EVENT_IOC_RESET:
2171 func = perf_event_reset;
2174 case PERF_EVENT_IOC_REFRESH:
2175 return perf_event_refresh(event, arg);
2177 case PERF_EVENT_IOC_PERIOD:
2178 return perf_event_period(event, (u64 __user *)arg);
2180 case PERF_EVENT_IOC_SET_OUTPUT:
2181 return perf_event_set_output(event, arg);
2183 case PERF_EVENT_IOC_SET_FILTER:
2184 return perf_event_set_filter(event, (void __user *)arg);
2190 if (flags & PERF_IOC_FLAG_GROUP)
2191 perf_event_for_each(event, func);
2193 perf_event_for_each_child(event, func);
2198 int perf_event_task_enable(void)
2200 struct perf_event *event;
2202 mutex_lock(¤t->perf_event_mutex);
2203 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2204 perf_event_for_each_child(event, perf_event_enable);
2205 mutex_unlock(¤t->perf_event_mutex);
2210 int perf_event_task_disable(void)
2212 struct perf_event *event;
2214 mutex_lock(¤t->perf_event_mutex);
2215 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2216 perf_event_for_each_child(event, perf_event_disable);
2217 mutex_unlock(¤t->perf_event_mutex);
2222 #ifndef PERF_EVENT_INDEX_OFFSET
2223 # define PERF_EVENT_INDEX_OFFSET 0
2226 static int perf_event_index(struct perf_event *event)
2228 if (event->state != PERF_EVENT_STATE_ACTIVE)
2231 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2235 * Callers need to ensure there can be no nesting of this function, otherwise
2236 * the seqlock logic goes bad. We can not serialize this because the arch
2237 * code calls this from NMI context.
2239 void perf_event_update_userpage(struct perf_event *event)
2241 struct perf_event_mmap_page *userpg;
2242 struct perf_mmap_data *data;
2245 data = rcu_dereference(event->data);
2249 userpg = data->user_page;
2252 * Disable preemption so as to not let the corresponding user-space
2253 * spin too long if we get preempted.
2258 userpg->index = perf_event_index(event);
2259 userpg->offset = atomic64_read(&event->count);
2260 if (event->state == PERF_EVENT_STATE_ACTIVE)
2261 userpg->offset -= atomic64_read(&event->hw.prev_count);
2263 userpg->time_enabled = event->total_time_enabled +
2264 atomic64_read(&event->child_total_time_enabled);
2266 userpg->time_running = event->total_time_running +
2267 atomic64_read(&event->child_total_time_running);
2276 static unsigned long perf_data_size(struct perf_mmap_data *data)
2278 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2281 #ifndef CONFIG_PERF_USE_VMALLOC
2284 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2287 static struct page *
2288 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2290 if (pgoff > data->nr_pages)
2294 return virt_to_page(data->user_page);
2296 return virt_to_page(data->data_pages[pgoff - 1]);
2299 static struct perf_mmap_data *
2300 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2302 struct perf_mmap_data *data;
2306 WARN_ON(atomic_read(&event->mmap_count));
2308 size = sizeof(struct perf_mmap_data);
2309 size += nr_pages * sizeof(void *);
2311 data = kzalloc(size, GFP_KERNEL);
2315 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2316 if (!data->user_page)
2317 goto fail_user_page;
2319 for (i = 0; i < nr_pages; i++) {
2320 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2321 if (!data->data_pages[i])
2322 goto fail_data_pages;
2325 data->data_order = 0;
2326 data->nr_pages = nr_pages;
2331 for (i--; i >= 0; i--)
2332 free_page((unsigned long)data->data_pages[i]);
2334 free_page((unsigned long)data->user_page);
2343 static void perf_mmap_free_page(unsigned long addr)
2345 struct page *page = virt_to_page((void *)addr);
2347 page->mapping = NULL;
2351 static void perf_mmap_data_free(struct perf_mmap_data *data)
2355 perf_mmap_free_page((unsigned long)data->user_page);
2356 for (i = 0; i < data->nr_pages; i++)
2357 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2364 * Back perf_mmap() with vmalloc memory.
2366 * Required for architectures that have d-cache aliasing issues.
2369 static struct page *
2370 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2372 if (pgoff > (1UL << data->data_order))
2375 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2378 static void perf_mmap_unmark_page(void *addr)
2380 struct page *page = vmalloc_to_page(addr);
2382 page->mapping = NULL;
2385 static void perf_mmap_data_free_work(struct work_struct *work)
2387 struct perf_mmap_data *data;
2391 data = container_of(work, struct perf_mmap_data, work);
2392 nr = 1 << data->data_order;
2394 base = data->user_page;
2395 for (i = 0; i < nr + 1; i++)
2396 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2402 static void perf_mmap_data_free(struct perf_mmap_data *data)
2404 schedule_work(&data->work);
2407 static struct perf_mmap_data *
2408 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2410 struct perf_mmap_data *data;
2414 WARN_ON(atomic_read(&event->mmap_count));
2416 size = sizeof(struct perf_mmap_data);
2417 size += sizeof(void *);
2419 data = kzalloc(size, GFP_KERNEL);
2423 INIT_WORK(&data->work, perf_mmap_data_free_work);
2425 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2429 data->user_page = all_buf;
2430 data->data_pages[0] = all_buf + PAGE_SIZE;
2431 data->data_order = ilog2(nr_pages);
2445 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2447 struct perf_event *event = vma->vm_file->private_data;
2448 struct perf_mmap_data *data;
2449 int ret = VM_FAULT_SIGBUS;
2451 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2452 if (vmf->pgoff == 0)
2458 data = rcu_dereference(event->data);
2462 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2465 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2469 get_page(vmf->page);
2470 vmf->page->mapping = vma->vm_file->f_mapping;
2471 vmf->page->index = vmf->pgoff;
2481 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2483 long max_size = perf_data_size(data);
2485 atomic_set(&data->lock, -1);
2487 if (event->attr.watermark) {
2488 data->watermark = min_t(long, max_size,
2489 event->attr.wakeup_watermark);
2492 if (!data->watermark)
2493 data->watermark = max_size / 2;
2496 rcu_assign_pointer(event->data, data);
2499 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2501 struct perf_mmap_data *data;
2503 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2504 perf_mmap_data_free(data);
2507 static void perf_mmap_data_release(struct perf_event *event)
2509 struct perf_mmap_data *data = event->data;
2511 WARN_ON(atomic_read(&event->mmap_count));
2513 rcu_assign_pointer(event->data, NULL);
2514 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2517 static void perf_mmap_open(struct vm_area_struct *vma)
2519 struct perf_event *event = vma->vm_file->private_data;
2521 atomic_inc(&event->mmap_count);
2524 static void perf_mmap_close(struct vm_area_struct *vma)
2526 struct perf_event *event = vma->vm_file->private_data;
2528 WARN_ON_ONCE(event->ctx->parent_ctx);
2529 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2530 unsigned long size = perf_data_size(event->data);
2531 struct user_struct *user = current_user();
2533 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2534 vma->vm_mm->locked_vm -= event->data->nr_locked;
2535 perf_mmap_data_release(event);
2536 mutex_unlock(&event->mmap_mutex);
2540 static const struct vm_operations_struct perf_mmap_vmops = {
2541 .open = perf_mmap_open,
2542 .close = perf_mmap_close,
2543 .fault = perf_mmap_fault,
2544 .page_mkwrite = perf_mmap_fault,
2547 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2549 struct perf_event *event = file->private_data;
2550 unsigned long user_locked, user_lock_limit;
2551 struct user_struct *user = current_user();
2552 unsigned long locked, lock_limit;
2553 struct perf_mmap_data *data;
2554 unsigned long vma_size;
2555 unsigned long nr_pages;
2556 long user_extra, extra;
2559 if (!(vma->vm_flags & VM_SHARED))
2562 vma_size = vma->vm_end - vma->vm_start;
2563 nr_pages = (vma_size / PAGE_SIZE) - 1;
2566 * If we have data pages ensure they're a power-of-two number, so we
2567 * can do bitmasks instead of modulo.
2569 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2572 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2575 if (vma->vm_pgoff != 0)
2578 WARN_ON_ONCE(event->ctx->parent_ctx);
2579 mutex_lock(&event->mmap_mutex);
2580 if (event->output) {
2585 if (atomic_inc_not_zero(&event->mmap_count)) {
2586 if (nr_pages != event->data->nr_pages)
2591 user_extra = nr_pages + 1;
2592 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2595 * Increase the limit linearly with more CPUs:
2597 user_lock_limit *= num_online_cpus();
2599 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2602 if (user_locked > user_lock_limit)
2603 extra = user_locked - user_lock_limit;
2605 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2606 lock_limit >>= PAGE_SHIFT;
2607 locked = vma->vm_mm->locked_vm + extra;
2609 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2610 !capable(CAP_IPC_LOCK)) {
2615 WARN_ON(event->data);
2617 data = perf_mmap_data_alloc(event, nr_pages);
2623 perf_mmap_data_init(event, data);
2625 atomic_set(&event->mmap_count, 1);
2626 atomic_long_add(user_extra, &user->locked_vm);
2627 vma->vm_mm->locked_vm += extra;
2628 event->data->nr_locked = extra;
2629 if (vma->vm_flags & VM_WRITE)
2630 event->data->writable = 1;
2633 mutex_unlock(&event->mmap_mutex);
2635 vma->vm_flags |= VM_RESERVED;
2636 vma->vm_ops = &perf_mmap_vmops;
2641 static int perf_fasync(int fd, struct file *filp, int on)
2643 struct inode *inode = filp->f_path.dentry->d_inode;
2644 struct perf_event *event = filp->private_data;
2647 mutex_lock(&inode->i_mutex);
2648 retval = fasync_helper(fd, filp, on, &event->fasync);
2649 mutex_unlock(&inode->i_mutex);
2657 static const struct file_operations perf_fops = {
2658 .release = perf_release,
2661 .unlocked_ioctl = perf_ioctl,
2662 .compat_ioctl = perf_ioctl,
2664 .fasync = perf_fasync,
2670 * If there's data, ensure we set the poll() state and publish everything
2671 * to user-space before waking everybody up.
2674 void perf_event_wakeup(struct perf_event *event)
2676 wake_up_all(&event->waitq);
2678 if (event->pending_kill) {
2679 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2680 event->pending_kill = 0;
2687 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2689 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2690 * single linked list and use cmpxchg() to add entries lockless.
2693 static void perf_pending_event(struct perf_pending_entry *entry)
2695 struct perf_event *event = container_of(entry,
2696 struct perf_event, pending);
2698 if (event->pending_disable) {
2699 event->pending_disable = 0;
2700 __perf_event_disable(event);
2703 if (event->pending_wakeup) {
2704 event->pending_wakeup = 0;
2705 perf_event_wakeup(event);
2709 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2711 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2715 static void perf_pending_queue(struct perf_pending_entry *entry,
2716 void (*func)(struct perf_pending_entry *))
2718 struct perf_pending_entry **head;
2720 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2725 head = &get_cpu_var(perf_pending_head);
2728 entry->next = *head;
2729 } while (cmpxchg(head, entry->next, entry) != entry->next);
2731 set_perf_event_pending();
2733 put_cpu_var(perf_pending_head);
2736 static int __perf_pending_run(void)
2738 struct perf_pending_entry *list;
2741 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2742 while (list != PENDING_TAIL) {
2743 void (*func)(struct perf_pending_entry *);
2744 struct perf_pending_entry *entry = list;
2751 * Ensure we observe the unqueue before we issue the wakeup,
2752 * so that we won't be waiting forever.
2753 * -- see perf_not_pending().
2764 static inline int perf_not_pending(struct perf_event *event)
2767 * If we flush on whatever cpu we run, there is a chance we don't
2771 __perf_pending_run();
2775 * Ensure we see the proper queue state before going to sleep
2776 * so that we do not miss the wakeup. -- see perf_pending_handle()
2779 return event->pending.next == NULL;
2782 static void perf_pending_sync(struct perf_event *event)
2784 wait_event(event->waitq, perf_not_pending(event));
2787 void perf_event_do_pending(void)
2789 __perf_pending_run();
2793 * Callchain support -- arch specific
2796 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2804 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2805 unsigned long offset, unsigned long head)
2809 if (!data->writable)
2812 mask = perf_data_size(data) - 1;
2814 offset = (offset - tail) & mask;
2815 head = (head - tail) & mask;
2817 if ((int)(head - offset) < 0)
2823 static void perf_output_wakeup(struct perf_output_handle *handle)
2825 atomic_set(&handle->data->poll, POLL_IN);
2828 handle->event->pending_wakeup = 1;
2829 perf_pending_queue(&handle->event->pending,
2830 perf_pending_event);
2832 perf_event_wakeup(handle->event);
2836 * Curious locking construct.
2838 * We need to ensure a later event_id doesn't publish a head when a former
2839 * event_id isn't done writing. However since we need to deal with NMIs we
2840 * cannot fully serialize things.
2842 * What we do is serialize between CPUs so we only have to deal with NMI
2843 * nesting on a single CPU.
2845 * We only publish the head (and generate a wakeup) when the outer-most
2846 * event_id completes.
2848 static void perf_output_lock(struct perf_output_handle *handle)
2850 struct perf_mmap_data *data = handle->data;
2851 int cur, cpu = get_cpu();
2856 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2868 static void perf_output_unlock(struct perf_output_handle *handle)
2870 struct perf_mmap_data *data = handle->data;
2874 data->done_head = data->head;
2876 if (!handle->locked)
2881 * The xchg implies a full barrier that ensures all writes are done
2882 * before we publish the new head, matched by a rmb() in userspace when
2883 * reading this position.
2885 while ((head = atomic_long_xchg(&data->done_head, 0)))
2886 data->user_page->data_head = head;
2889 * NMI can happen here, which means we can miss a done_head update.
2892 cpu = atomic_xchg(&data->lock, -1);
2893 WARN_ON_ONCE(cpu != smp_processor_id());
2896 * Therefore we have to validate we did not indeed do so.
2898 if (unlikely(atomic_long_read(&data->done_head))) {
2900 * Since we had it locked, we can lock it again.
2902 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2908 if (atomic_xchg(&data->wakeup, 0))
2909 perf_output_wakeup(handle);
2914 void perf_output_copy(struct perf_output_handle *handle,
2915 const void *buf, unsigned int len)
2917 unsigned int pages_mask;
2918 unsigned long offset;
2922 offset = handle->offset;
2923 pages_mask = handle->data->nr_pages - 1;
2924 pages = handle->data->data_pages;
2927 unsigned long page_offset;
2928 unsigned long page_size;
2931 nr = (offset >> PAGE_SHIFT) & pages_mask;
2932 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2933 page_offset = offset & (page_size - 1);
2934 size = min_t(unsigned int, page_size - page_offset, len);
2936 memcpy(pages[nr] + page_offset, buf, size);
2943 handle->offset = offset;
2946 * Check we didn't copy past our reservation window, taking the
2947 * possible unsigned int wrap into account.
2949 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2952 int perf_output_begin(struct perf_output_handle *handle,
2953 struct perf_event *event, unsigned int size,
2954 int nmi, int sample)
2956 struct perf_event *output_event;
2957 struct perf_mmap_data *data;
2958 unsigned long tail, offset, head;
2961 struct perf_event_header header;
2968 * For inherited events we send all the output towards the parent.
2971 event = event->parent;
2973 output_event = rcu_dereference(event->output);
2975 event = output_event;
2977 data = rcu_dereference(event->data);
2981 handle->data = data;
2982 handle->event = event;
2984 handle->sample = sample;
2986 if (!data->nr_pages)
2989 have_lost = atomic_read(&data->lost);
2991 size += sizeof(lost_event);
2993 perf_output_lock(handle);
2997 * Userspace could choose to issue a mb() before updating the
2998 * tail pointer. So that all reads will be completed before the
3001 tail = ACCESS_ONCE(data->user_page->data_tail);
3003 offset = head = atomic_long_read(&data->head);
3005 if (unlikely(!perf_output_space(data, tail, offset, head)))
3007 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3009 handle->offset = offset;
3010 handle->head = head;
3012 if (head - tail > data->watermark)
3013 atomic_set(&data->wakeup, 1);
3016 lost_event.header.type = PERF_RECORD_LOST;
3017 lost_event.header.misc = 0;
3018 lost_event.header.size = sizeof(lost_event);
3019 lost_event.id = event->id;
3020 lost_event.lost = atomic_xchg(&data->lost, 0);
3022 perf_output_put(handle, lost_event);
3028 atomic_inc(&data->lost);
3029 perf_output_unlock(handle);
3036 void perf_output_end(struct perf_output_handle *handle)
3038 struct perf_event *event = handle->event;
3039 struct perf_mmap_data *data = handle->data;
3041 int wakeup_events = event->attr.wakeup_events;
3043 if (handle->sample && wakeup_events) {
3044 int events = atomic_inc_return(&data->events);
3045 if (events >= wakeup_events) {
3046 atomic_sub(wakeup_events, &data->events);
3047 atomic_set(&data->wakeup, 1);
3051 perf_output_unlock(handle);
3055 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
3058 * only top level events have the pid namespace they were created in
3061 event = event->parent;
3063 return task_tgid_nr_ns(p, event->ns);
3066 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3069 * only top level events have the pid namespace they were created in
3072 event = event->parent;
3074 return task_pid_nr_ns(p, event->ns);
3077 static void perf_output_read_one(struct perf_output_handle *handle,
3078 struct perf_event *event)
3080 u64 read_format = event->attr.read_format;
3084 values[n++] = atomic64_read(&event->count);
3085 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3086 values[n++] = event->total_time_enabled +
3087 atomic64_read(&event->child_total_time_enabled);
3089 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3090 values[n++] = event->total_time_running +
3091 atomic64_read(&event->child_total_time_running);
3093 if (read_format & PERF_FORMAT_ID)
3094 values[n++] = primary_event_id(event);
3096 perf_output_copy(handle, values, n * sizeof(u64));
3100 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3102 static void perf_output_read_group(struct perf_output_handle *handle,
3103 struct perf_event *event)
3105 struct perf_event *leader = event->group_leader, *sub;
3106 u64 read_format = event->attr.read_format;
3110 values[n++] = 1 + leader->nr_siblings;
3112 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3113 values[n++] = leader->total_time_enabled;
3115 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3116 values[n++] = leader->total_time_running;
3118 if (leader != event)
3119 leader->pmu->read(leader);
3121 values[n++] = atomic64_read(&leader->count);
3122 if (read_format & PERF_FORMAT_ID)
3123 values[n++] = primary_event_id(leader);
3125 perf_output_copy(handle, values, n * sizeof(u64));
3127 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3131 sub->pmu->read(sub);
3133 values[n++] = atomic64_read(&sub->count);
3134 if (read_format & PERF_FORMAT_ID)
3135 values[n++] = primary_event_id(sub);
3137 perf_output_copy(handle, values, n * sizeof(u64));
3141 static void perf_output_read(struct perf_output_handle *handle,
3142 struct perf_event *event)
3144 if (event->attr.read_format & PERF_FORMAT_GROUP)
3145 perf_output_read_group(handle, event);
3147 perf_output_read_one(handle, event);
3150 void perf_output_sample(struct perf_output_handle *handle,
3151 struct perf_event_header *header,
3152 struct perf_sample_data *data,
3153 struct perf_event *event)
3155 u64 sample_type = data->type;
3157 perf_output_put(handle, *header);
3159 if (sample_type & PERF_SAMPLE_IP)
3160 perf_output_put(handle, data->ip);
3162 if (sample_type & PERF_SAMPLE_TID)
3163 perf_output_put(handle, data->tid_entry);
3165 if (sample_type & PERF_SAMPLE_TIME)
3166 perf_output_put(handle, data->time);
3168 if (sample_type & PERF_SAMPLE_ADDR)
3169 perf_output_put(handle, data->addr);
3171 if (sample_type & PERF_SAMPLE_ID)
3172 perf_output_put(handle, data->id);
3174 if (sample_type & PERF_SAMPLE_STREAM_ID)
3175 perf_output_put(handle, data->stream_id);
3177 if (sample_type & PERF_SAMPLE_CPU)
3178 perf_output_put(handle, data->cpu_entry);
3180 if (sample_type & PERF_SAMPLE_PERIOD)
3181 perf_output_put(handle, data->period);
3183 if (sample_type & PERF_SAMPLE_READ)
3184 perf_output_read(handle, event);
3186 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3187 if (data->callchain) {
3190 if (data->callchain)
3191 size += data->callchain->nr;
3193 size *= sizeof(u64);
3195 perf_output_copy(handle, data->callchain, size);
3198 perf_output_put(handle, nr);
3202 if (sample_type & PERF_SAMPLE_RAW) {
3204 perf_output_put(handle, data->raw->size);
3205 perf_output_copy(handle, data->raw->data,
3212 .size = sizeof(u32),
3215 perf_output_put(handle, raw);
3220 void perf_prepare_sample(struct perf_event_header *header,
3221 struct perf_sample_data *data,
3222 struct perf_event *event,
3223 struct pt_regs *regs)
3225 u64 sample_type = event->attr.sample_type;
3227 data->type = sample_type;
3229 header->type = PERF_RECORD_SAMPLE;
3230 header->size = sizeof(*header);
3233 header->misc |= perf_misc_flags(regs);
3235 if (sample_type & PERF_SAMPLE_IP) {
3236 data->ip = perf_instruction_pointer(regs);
3238 header->size += sizeof(data->ip);
3241 if (sample_type & PERF_SAMPLE_TID) {
3242 /* namespace issues */
3243 data->tid_entry.pid = perf_event_pid(event, current);
3244 data->tid_entry.tid = perf_event_tid(event, current);
3246 header->size += sizeof(data->tid_entry);
3249 if (sample_type & PERF_SAMPLE_TIME) {
3250 data->time = perf_clock();
3252 header->size += sizeof(data->time);
3255 if (sample_type & PERF_SAMPLE_ADDR)
3256 header->size += sizeof(data->addr);
3258 if (sample_type & PERF_SAMPLE_ID) {
3259 data->id = primary_event_id(event);
3261 header->size += sizeof(data->id);
3264 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3265 data->stream_id = event->id;
3267 header->size += sizeof(data->stream_id);
3270 if (sample_type & PERF_SAMPLE_CPU) {
3271 data->cpu_entry.cpu = raw_smp_processor_id();
3272 data->cpu_entry.reserved = 0;
3274 header->size += sizeof(data->cpu_entry);
3277 if (sample_type & PERF_SAMPLE_PERIOD)
3278 header->size += sizeof(data->period);
3280 if (sample_type & PERF_SAMPLE_READ)
3281 header->size += perf_event_read_size(event);
3283 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3286 data->callchain = perf_callchain(regs);
3288 if (data->callchain)
3289 size += data->callchain->nr;
3291 header->size += size * sizeof(u64);
3294 if (sample_type & PERF_SAMPLE_RAW) {
3295 int size = sizeof(u32);
3298 size += data->raw->size;
3300 size += sizeof(u32);
3302 WARN_ON_ONCE(size & (sizeof(u64)-1));
3303 header->size += size;
3307 static void perf_event_output(struct perf_event *event, int nmi,
3308 struct perf_sample_data *data,
3309 struct pt_regs *regs)
3311 struct perf_output_handle handle;
3312 struct perf_event_header header;
3314 perf_prepare_sample(&header, data, event, regs);
3316 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3319 perf_output_sample(&handle, &header, data, event);
3321 perf_output_end(&handle);
3328 struct perf_read_event {
3329 struct perf_event_header header;
3336 perf_event_read_event(struct perf_event *event,
3337 struct task_struct *task)
3339 struct perf_output_handle handle;
3340 struct perf_read_event read_event = {
3342 .type = PERF_RECORD_READ,
3344 .size = sizeof(read_event) + perf_event_read_size(event),
3346 .pid = perf_event_pid(event, task),
3347 .tid = perf_event_tid(event, task),
3351 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3355 perf_output_put(&handle, read_event);
3356 perf_output_read(&handle, event);
3358 perf_output_end(&handle);
3362 * task tracking -- fork/exit
3364 * enabled by: attr.comm | attr.mmap | attr.task
3367 struct perf_task_event {
3368 struct task_struct *task;
3369 struct perf_event_context *task_ctx;
3372 struct perf_event_header header;
3382 static void perf_event_task_output(struct perf_event *event,
3383 struct perf_task_event *task_event)
3385 struct perf_output_handle handle;
3387 struct task_struct *task = task_event->task;
3390 size = task_event->event_id.header.size;
3391 ret = perf_output_begin(&handle, event, size, 0, 0);
3396 task_event->event_id.pid = perf_event_pid(event, task);
3397 task_event->event_id.ppid = perf_event_pid(event, current);
3399 task_event->event_id.tid = perf_event_tid(event, task);
3400 task_event->event_id.ptid = perf_event_tid(event, current);
3402 task_event->event_id.time = perf_clock();
3404 perf_output_put(&handle, task_event->event_id);
3406 perf_output_end(&handle);
3409 static int perf_event_task_match(struct perf_event *event)
3411 if (event->state != PERF_EVENT_STATE_ACTIVE)
3414 if (event->cpu != -1 && event->cpu != smp_processor_id())
3417 if (event->attr.comm || event->attr.mmap || event->attr.task)
3423 static void perf_event_task_ctx(struct perf_event_context *ctx,
3424 struct perf_task_event *task_event)
3426 struct perf_event *event;
3428 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3429 if (perf_event_task_match(event))
3430 perf_event_task_output(event, task_event);
3434 static void perf_event_task_event(struct perf_task_event *task_event)
3436 struct perf_cpu_context *cpuctx;
3437 struct perf_event_context *ctx = task_event->task_ctx;
3440 cpuctx = &get_cpu_var(perf_cpu_context);
3441 perf_event_task_ctx(&cpuctx->ctx, task_event);
3443 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3445 perf_event_task_ctx(ctx, task_event);
3446 put_cpu_var(perf_cpu_context);
3450 static void perf_event_task(struct task_struct *task,
3451 struct perf_event_context *task_ctx,
3454 struct perf_task_event task_event;
3456 if (!atomic_read(&nr_comm_events) &&
3457 !atomic_read(&nr_mmap_events) &&
3458 !atomic_read(&nr_task_events))
3461 task_event = (struct perf_task_event){
3463 .task_ctx = task_ctx,
3466 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3468 .size = sizeof(task_event.event_id),
3477 perf_event_task_event(&task_event);
3480 void perf_event_fork(struct task_struct *task)
3482 perf_event_task(task, NULL, 1);
3489 struct perf_comm_event {
3490 struct task_struct *task;
3495 struct perf_event_header header;
3502 static void perf_event_comm_output(struct perf_event *event,
3503 struct perf_comm_event *comm_event)
3505 struct perf_output_handle handle;
3506 int size = comm_event->event_id.header.size;
3507 int ret = perf_output_begin(&handle, event, size, 0, 0);
3512 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3513 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3515 perf_output_put(&handle, comm_event->event_id);
3516 perf_output_copy(&handle, comm_event->comm,
3517 comm_event->comm_size);
3518 perf_output_end(&handle);
3521 static int perf_event_comm_match(struct perf_event *event)
3523 if (event->state != PERF_EVENT_STATE_ACTIVE)
3526 if (event->cpu != -1 && event->cpu != smp_processor_id())
3529 if (event->attr.comm)
3535 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3536 struct perf_comm_event *comm_event)
3538 struct perf_event *event;
3540 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3541 if (perf_event_comm_match(event))
3542 perf_event_comm_output(event, comm_event);
3546 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3548 struct perf_cpu_context *cpuctx;
3549 struct perf_event_context *ctx;
3551 char comm[TASK_COMM_LEN];
3553 memset(comm, 0, sizeof(comm));
3554 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3555 size = ALIGN(strlen(comm)+1, sizeof(u64));
3557 comm_event->comm = comm;
3558 comm_event->comm_size = size;
3560 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3563 cpuctx = &get_cpu_var(perf_cpu_context);
3564 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3565 ctx = rcu_dereference(current->perf_event_ctxp);
3567 perf_event_comm_ctx(ctx, comm_event);
3568 put_cpu_var(perf_cpu_context);
3572 void perf_event_comm(struct task_struct *task)
3574 struct perf_comm_event comm_event;
3576 if (task->perf_event_ctxp)
3577 perf_event_enable_on_exec(task);
3579 if (!atomic_read(&nr_comm_events))
3582 comm_event = (struct perf_comm_event){
3588 .type = PERF_RECORD_COMM,
3597 perf_event_comm_event(&comm_event);
3604 struct perf_mmap_event {
3605 struct vm_area_struct *vma;
3607 const char *file_name;
3611 struct perf_event_header header;
3621 static void perf_event_mmap_output(struct perf_event *event,
3622 struct perf_mmap_event *mmap_event)
3624 struct perf_output_handle handle;
3625 int size = mmap_event->event_id.header.size;
3626 int ret = perf_output_begin(&handle, event, size, 0, 0);
3631 mmap_event->event_id.pid = perf_event_pid(event, current);
3632 mmap_event->event_id.tid = perf_event_tid(event, current);
3634 perf_output_put(&handle, mmap_event->event_id);
3635 perf_output_copy(&handle, mmap_event->file_name,
3636 mmap_event->file_size);
3637 perf_output_end(&handle);
3640 static int perf_event_mmap_match(struct perf_event *event,
3641 struct perf_mmap_event *mmap_event)
3643 if (event->state != PERF_EVENT_STATE_ACTIVE)
3646 if (event->cpu != -1 && event->cpu != smp_processor_id())
3649 if (event->attr.mmap)
3655 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3656 struct perf_mmap_event *mmap_event)
3658 struct perf_event *event;
3660 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3661 if (perf_event_mmap_match(event, mmap_event))
3662 perf_event_mmap_output(event, mmap_event);
3666 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3668 struct perf_cpu_context *cpuctx;
3669 struct perf_event_context *ctx;
3670 struct vm_area_struct *vma = mmap_event->vma;
3671 struct file *file = vma->vm_file;
3677 memset(tmp, 0, sizeof(tmp));
3681 * d_path works from the end of the buffer backwards, so we
3682 * need to add enough zero bytes after the string to handle
3683 * the 64bit alignment we do later.
3685 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3687 name = strncpy(tmp, "//enomem", sizeof(tmp));
3690 name = d_path(&file->f_path, buf, PATH_MAX);
3692 name = strncpy(tmp, "//toolong", sizeof(tmp));
3696 if (arch_vma_name(mmap_event->vma)) {
3697 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3703 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3707 name = strncpy(tmp, "//anon", sizeof(tmp));
3712 size = ALIGN(strlen(name)+1, sizeof(u64));
3714 mmap_event->file_name = name;
3715 mmap_event->file_size = size;
3717 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3720 cpuctx = &get_cpu_var(perf_cpu_context);
3721 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3722 ctx = rcu_dereference(current->perf_event_ctxp);
3724 perf_event_mmap_ctx(ctx, mmap_event);
3725 put_cpu_var(perf_cpu_context);
3731 void __perf_event_mmap(struct vm_area_struct *vma)
3733 struct perf_mmap_event mmap_event;
3735 if (!atomic_read(&nr_mmap_events))
3738 mmap_event = (struct perf_mmap_event){
3744 .type = PERF_RECORD_MMAP,
3750 .start = vma->vm_start,
3751 .len = vma->vm_end - vma->vm_start,
3752 .pgoff = vma->vm_pgoff,
3756 perf_event_mmap_event(&mmap_event);
3760 * IRQ throttle logging
3763 static void perf_log_throttle(struct perf_event *event, int enable)
3765 struct perf_output_handle handle;
3769 struct perf_event_header header;
3773 } throttle_event = {
3775 .type = PERF_RECORD_THROTTLE,
3777 .size = sizeof(throttle_event),
3779 .time = perf_clock(),
3780 .id = primary_event_id(event),
3781 .stream_id = event->id,
3785 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3787 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3791 perf_output_put(&handle, throttle_event);
3792 perf_output_end(&handle);
3796 * Generic event overflow handling, sampling.
3799 static int __perf_event_overflow(struct perf_event *event, int nmi,
3800 int throttle, struct perf_sample_data *data,
3801 struct pt_regs *regs)
3803 int events = atomic_read(&event->event_limit);
3804 struct hw_perf_event *hwc = &event->hw;
3807 throttle = (throttle && event->pmu->unthrottle != NULL);
3812 if (hwc->interrupts != MAX_INTERRUPTS) {
3814 if (HZ * hwc->interrupts >
3815 (u64)sysctl_perf_event_sample_rate) {
3816 hwc->interrupts = MAX_INTERRUPTS;
3817 perf_log_throttle(event, 0);
3822 * Keep re-disabling events even though on the previous
3823 * pass we disabled it - just in case we raced with a
3824 * sched-in and the event got enabled again:
3830 if (event->attr.freq) {
3831 u64 now = perf_clock();
3832 s64 delta = now - hwc->freq_time_stamp;
3834 hwc->freq_time_stamp = now;
3836 if (delta > 0 && delta < 2*TICK_NSEC)
3837 perf_adjust_period(event, delta, hwc->last_period);
3841 * XXX event_limit might not quite work as expected on inherited
3845 event->pending_kill = POLL_IN;
3846 if (events && atomic_dec_and_test(&event->event_limit)) {
3848 event->pending_kill = POLL_HUP;
3850 event->pending_disable = 1;
3851 perf_pending_queue(&event->pending,
3852 perf_pending_event);
3854 perf_event_disable(event);
3857 if (event->overflow_handler)
3858 event->overflow_handler(event, nmi, data, regs);
3860 perf_event_output(event, nmi, data, regs);
3865 int perf_event_overflow(struct perf_event *event, int nmi,
3866 struct perf_sample_data *data,
3867 struct pt_regs *regs)
3869 return __perf_event_overflow(event, nmi, 1, data, regs);
3873 * Generic software event infrastructure
3877 * We directly increment event->count and keep a second value in
3878 * event->hw.period_left to count intervals. This period event
3879 * is kept in the range [-sample_period, 0] so that we can use the
3883 static u64 perf_swevent_set_period(struct perf_event *event)
3885 struct hw_perf_event *hwc = &event->hw;
3886 u64 period = hwc->last_period;
3890 hwc->last_period = hwc->sample_period;
3893 old = val = atomic64_read(&hwc->period_left);
3897 nr = div64_u64(period + val, period);
3898 offset = nr * period;
3900 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3906 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3907 int nmi, struct perf_sample_data *data,
3908 struct pt_regs *regs)
3910 struct hw_perf_event *hwc = &event->hw;
3913 data->period = event->hw.last_period;
3915 overflow = perf_swevent_set_period(event);
3917 if (hwc->interrupts == MAX_INTERRUPTS)
3920 for (; overflow; overflow--) {
3921 if (__perf_event_overflow(event, nmi, throttle,
3924 * We inhibit the overflow from happening when
3925 * hwc->interrupts == MAX_INTERRUPTS.
3933 static void perf_swevent_unthrottle(struct perf_event *event)
3936 * Nothing to do, we already reset hwc->interrupts.
3940 static void perf_swevent_add(struct perf_event *event, u64 nr,
3941 int nmi, struct perf_sample_data *data,
3942 struct pt_regs *regs)
3944 struct hw_perf_event *hwc = &event->hw;
3946 atomic64_add(nr, &event->count);
3951 if (!hwc->sample_period)
3954 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3955 return perf_swevent_overflow(event, 1, nmi, data, regs);
3957 if (atomic64_add_negative(nr, &hwc->period_left))
3960 perf_swevent_overflow(event, 0, nmi, data, regs);
3963 static int perf_swevent_is_counting(struct perf_event *event)
3966 * The event is active, we're good!
3968 if (event->state == PERF_EVENT_STATE_ACTIVE)
3972 * The event is off/error, not counting.
3974 if (event->state != PERF_EVENT_STATE_INACTIVE)
3978 * The event is inactive, if the context is active
3979 * we're part of a group that didn't make it on the 'pmu',
3982 if (event->ctx->is_active)
3986 * We're inactive and the context is too, this means the
3987 * task is scheduled out, we're counting events that happen
3988 * to us, like migration events.
3993 static int perf_tp_event_match(struct perf_event *event,
3994 struct perf_sample_data *data);
3996 static int perf_exclude_event(struct perf_event *event,
3997 struct pt_regs *regs)
4000 if (event->attr.exclude_user && user_mode(regs))
4003 if (event->attr.exclude_kernel && !user_mode(regs))
4010 static int perf_swevent_match(struct perf_event *event,
4011 enum perf_type_id type,
4013 struct perf_sample_data *data,
4014 struct pt_regs *regs)
4016 if (event->cpu != -1 && event->cpu != smp_processor_id())
4019 if (!perf_swevent_is_counting(event))
4022 if (event->attr.type != type)
4025 if (event->attr.config != event_id)
4028 if (perf_exclude_event(event, regs))
4031 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4032 !perf_tp_event_match(event, data))
4038 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4039 enum perf_type_id type,
4040 u32 event_id, u64 nr, int nmi,
4041 struct perf_sample_data *data,
4042 struct pt_regs *regs)
4044 struct perf_event *event;
4046 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4047 if (perf_swevent_match(event, type, event_id, data, regs))
4048 perf_swevent_add(event, nr, nmi, data, regs);
4052 int perf_swevent_get_recursion_context(void)
4054 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4061 else if (in_softirq())
4066 if (cpuctx->recursion[rctx]) {
4067 put_cpu_var(perf_cpu_context);
4071 cpuctx->recursion[rctx]++;
4076 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4078 void perf_swevent_put_recursion_context(int rctx)
4080 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4082 cpuctx->recursion[rctx]--;
4083 put_cpu_var(perf_cpu_context);
4085 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4087 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4089 struct perf_sample_data *data,
4090 struct pt_regs *regs)
4092 struct perf_cpu_context *cpuctx;
4093 struct perf_event_context *ctx;
4095 cpuctx = &__get_cpu_var(perf_cpu_context);
4097 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4098 nr, nmi, data, regs);
4100 * doesn't really matter which of the child contexts the
4101 * events ends up in.
4103 ctx = rcu_dereference(current->perf_event_ctxp);
4105 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4109 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4110 struct pt_regs *regs, u64 addr)
4112 struct perf_sample_data data;
4115 rctx = perf_swevent_get_recursion_context();
4122 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4124 perf_swevent_put_recursion_context(rctx);
4127 static void perf_swevent_read(struct perf_event *event)
4131 static int perf_swevent_enable(struct perf_event *event)
4133 struct hw_perf_event *hwc = &event->hw;
4135 if (hwc->sample_period) {
4136 hwc->last_period = hwc->sample_period;
4137 perf_swevent_set_period(event);
4142 static void perf_swevent_disable(struct perf_event *event)
4146 static const struct pmu perf_ops_generic = {
4147 .enable = perf_swevent_enable,
4148 .disable = perf_swevent_disable,
4149 .read = perf_swevent_read,
4150 .unthrottle = perf_swevent_unthrottle,
4154 * hrtimer based swevent callback
4157 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4159 enum hrtimer_restart ret = HRTIMER_RESTART;
4160 struct perf_sample_data data;
4161 struct pt_regs *regs;
4162 struct perf_event *event;
4165 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4166 event->pmu->read(event);
4170 data.period = event->hw.last_period;
4171 regs = get_irq_regs();
4173 * In case we exclude kernel IPs or are somehow not in interrupt
4174 * context, provide the next best thing, the user IP.
4176 if ((event->attr.exclude_kernel || !regs) &&
4177 !event->attr.exclude_user)
4178 regs = task_pt_regs(current);
4181 if (!(event->attr.exclude_idle && current->pid == 0))
4182 if (perf_event_overflow(event, 0, &data, regs))
4183 ret = HRTIMER_NORESTART;
4186 period = max_t(u64, 10000, event->hw.sample_period);
4187 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4192 static void perf_swevent_start_hrtimer(struct perf_event *event)
4194 struct hw_perf_event *hwc = &event->hw;
4196 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4197 hwc->hrtimer.function = perf_swevent_hrtimer;
4198 if (hwc->sample_period) {
4201 if (hwc->remaining) {
4202 if (hwc->remaining < 0)
4205 period = hwc->remaining;
4208 period = max_t(u64, 10000, hwc->sample_period);
4210 __hrtimer_start_range_ns(&hwc->hrtimer,
4211 ns_to_ktime(period), 0,
4212 HRTIMER_MODE_REL, 0);
4216 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4218 struct hw_perf_event *hwc = &event->hw;
4220 if (hwc->sample_period) {
4221 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4222 hwc->remaining = ktime_to_ns(remaining);
4224 hrtimer_cancel(&hwc->hrtimer);
4229 * Software event: cpu wall time clock
4232 static void cpu_clock_perf_event_update(struct perf_event *event)
4234 int cpu = raw_smp_processor_id();
4238 now = cpu_clock(cpu);
4239 prev = atomic64_xchg(&event->hw.prev_count, now);
4240 atomic64_add(now - prev, &event->count);
4243 static int cpu_clock_perf_event_enable(struct perf_event *event)
4245 struct hw_perf_event *hwc = &event->hw;
4246 int cpu = raw_smp_processor_id();
4248 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4249 perf_swevent_start_hrtimer(event);
4254 static void cpu_clock_perf_event_disable(struct perf_event *event)
4256 perf_swevent_cancel_hrtimer(event);
4257 cpu_clock_perf_event_update(event);
4260 static void cpu_clock_perf_event_read(struct perf_event *event)
4262 cpu_clock_perf_event_update(event);
4265 static const struct pmu perf_ops_cpu_clock = {
4266 .enable = cpu_clock_perf_event_enable,
4267 .disable = cpu_clock_perf_event_disable,
4268 .read = cpu_clock_perf_event_read,
4272 * Software event: task time clock
4275 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4280 prev = atomic64_xchg(&event->hw.prev_count, now);
4282 atomic64_add(delta, &event->count);
4285 static int task_clock_perf_event_enable(struct perf_event *event)
4287 struct hw_perf_event *hwc = &event->hw;
4290 now = event->ctx->time;
4292 atomic64_set(&hwc->prev_count, now);
4294 perf_swevent_start_hrtimer(event);
4299 static void task_clock_perf_event_disable(struct perf_event *event)
4301 perf_swevent_cancel_hrtimer(event);
4302 task_clock_perf_event_update(event, event->ctx->time);
4306 static void task_clock_perf_event_read(struct perf_event *event)
4311 update_context_time(event->ctx);
4312 time = event->ctx->time;
4314 u64 now = perf_clock();
4315 u64 delta = now - event->ctx->timestamp;
4316 time = event->ctx->time + delta;
4319 task_clock_perf_event_update(event, time);
4322 static const struct pmu perf_ops_task_clock = {
4323 .enable = task_clock_perf_event_enable,
4324 .disable = task_clock_perf_event_disable,
4325 .read = task_clock_perf_event_read,
4328 #ifdef CONFIG_EVENT_TRACING
4330 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4333 struct perf_raw_record raw = {
4338 struct perf_sample_data data = {
4343 struct pt_regs *regs = get_irq_regs();
4346 regs = task_pt_regs(current);
4348 /* Trace events already protected against recursion */
4349 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4352 EXPORT_SYMBOL_GPL(perf_tp_event);
4354 static int perf_tp_event_match(struct perf_event *event,
4355 struct perf_sample_data *data)
4357 void *record = data->raw->data;
4359 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4364 static void tp_perf_event_destroy(struct perf_event *event)
4366 ftrace_profile_disable(event->attr.config);
4369 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4372 * Raw tracepoint data is a severe data leak, only allow root to
4375 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4376 perf_paranoid_tracepoint_raw() &&
4377 !capable(CAP_SYS_ADMIN))
4378 return ERR_PTR(-EPERM);
4380 if (ftrace_profile_enable(event->attr.config))
4383 event->destroy = tp_perf_event_destroy;
4385 return &perf_ops_generic;
4388 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4393 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4396 filter_str = strndup_user(arg, PAGE_SIZE);
4397 if (IS_ERR(filter_str))
4398 return PTR_ERR(filter_str);
4400 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4406 static void perf_event_free_filter(struct perf_event *event)
4408 ftrace_profile_free_filter(event);
4413 static int perf_tp_event_match(struct perf_event *event,
4414 struct perf_sample_data *data)
4419 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4424 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4429 static void perf_event_free_filter(struct perf_event *event)
4433 #endif /* CONFIG_EVENT_TRACING */
4435 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4436 static void bp_perf_event_destroy(struct perf_event *event)
4438 release_bp_slot(event);
4441 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4445 err = register_perf_hw_breakpoint(bp);
4447 return ERR_PTR(err);
4449 bp->destroy = bp_perf_event_destroy;
4451 return &perf_ops_bp;
4454 void perf_bp_event(struct perf_event *bp, void *data)
4456 struct perf_sample_data sample;
4457 struct pt_regs *regs = data;
4460 sample.addr = bp->attr.bp_addr;
4462 if (!perf_exclude_event(bp, regs))
4463 perf_swevent_add(bp, 1, 1, &sample, regs);
4466 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4471 void perf_bp_event(struct perf_event *bp, void *regs)
4476 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4478 static void sw_perf_event_destroy(struct perf_event *event)
4480 u64 event_id = event->attr.config;
4482 WARN_ON(event->parent);
4484 atomic_dec(&perf_swevent_enabled[event_id]);
4487 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4489 const struct pmu *pmu = NULL;
4490 u64 event_id = event->attr.config;
4493 * Software events (currently) can't in general distinguish
4494 * between user, kernel and hypervisor events.
4495 * However, context switches and cpu migrations are considered
4496 * to be kernel events, and page faults are never hypervisor
4500 case PERF_COUNT_SW_CPU_CLOCK:
4501 pmu = &perf_ops_cpu_clock;
4504 case PERF_COUNT_SW_TASK_CLOCK:
4506 * If the user instantiates this as a per-cpu event,
4507 * use the cpu_clock event instead.
4509 if (event->ctx->task)
4510 pmu = &perf_ops_task_clock;
4512 pmu = &perf_ops_cpu_clock;
4515 case PERF_COUNT_SW_PAGE_FAULTS:
4516 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4517 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4518 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4519 case PERF_COUNT_SW_CPU_MIGRATIONS:
4520 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4521 case PERF_COUNT_SW_EMULATION_FAULTS:
4522 if (!event->parent) {
4523 atomic_inc(&perf_swevent_enabled[event_id]);
4524 event->destroy = sw_perf_event_destroy;
4526 pmu = &perf_ops_generic;
4534 * Allocate and initialize a event structure
4536 static struct perf_event *
4537 perf_event_alloc(struct perf_event_attr *attr,
4539 struct perf_event_context *ctx,
4540 struct perf_event *group_leader,
4541 struct perf_event *parent_event,
4542 perf_overflow_handler_t overflow_handler,
4545 const struct pmu *pmu;
4546 struct perf_event *event;
4547 struct hw_perf_event *hwc;
4550 event = kzalloc(sizeof(*event), gfpflags);
4552 return ERR_PTR(-ENOMEM);
4555 * Single events are their own group leaders, with an
4556 * empty sibling list:
4559 group_leader = event;
4561 mutex_init(&event->child_mutex);
4562 INIT_LIST_HEAD(&event->child_list);
4564 INIT_LIST_HEAD(&event->group_entry);
4565 INIT_LIST_HEAD(&event->event_entry);
4566 INIT_LIST_HEAD(&event->sibling_list);
4567 init_waitqueue_head(&event->waitq);
4569 mutex_init(&event->mmap_mutex);
4572 event->attr = *attr;
4573 event->group_leader = group_leader;
4578 event->parent = parent_event;
4580 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4581 event->id = atomic64_inc_return(&perf_event_id);
4583 event->state = PERF_EVENT_STATE_INACTIVE;
4585 if (!overflow_handler && parent_event)
4586 overflow_handler = parent_event->overflow_handler;
4588 event->overflow_handler = overflow_handler;
4591 event->state = PERF_EVENT_STATE_OFF;
4596 hwc->sample_period = attr->sample_period;
4597 if (attr->freq && attr->sample_freq)
4598 hwc->sample_period = 1;
4599 hwc->last_period = hwc->sample_period;
4601 atomic64_set(&hwc->period_left, hwc->sample_period);
4604 * we currently do not support PERF_FORMAT_GROUP on inherited events
4606 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4609 switch (attr->type) {
4611 case PERF_TYPE_HARDWARE:
4612 case PERF_TYPE_HW_CACHE:
4613 pmu = hw_perf_event_init(event);
4616 case PERF_TYPE_SOFTWARE:
4617 pmu = sw_perf_event_init(event);
4620 case PERF_TYPE_TRACEPOINT:
4621 pmu = tp_perf_event_init(event);
4624 case PERF_TYPE_BREAKPOINT:
4625 pmu = bp_perf_event_init(event);
4636 else if (IS_ERR(pmu))
4641 put_pid_ns(event->ns);
4643 return ERR_PTR(err);
4648 if (!event->parent) {
4649 atomic_inc(&nr_events);
4650 if (event->attr.mmap)
4651 atomic_inc(&nr_mmap_events);
4652 if (event->attr.comm)
4653 atomic_inc(&nr_comm_events);
4654 if (event->attr.task)
4655 atomic_inc(&nr_task_events);
4661 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4662 struct perf_event_attr *attr)
4667 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4671 * zero the full structure, so that a short copy will be nice.
4673 memset(attr, 0, sizeof(*attr));
4675 ret = get_user(size, &uattr->size);
4679 if (size > PAGE_SIZE) /* silly large */
4682 if (!size) /* abi compat */
4683 size = PERF_ATTR_SIZE_VER0;
4685 if (size < PERF_ATTR_SIZE_VER0)
4689 * If we're handed a bigger struct than we know of,
4690 * ensure all the unknown bits are 0 - i.e. new
4691 * user-space does not rely on any kernel feature
4692 * extensions we dont know about yet.
4694 if (size > sizeof(*attr)) {
4695 unsigned char __user *addr;
4696 unsigned char __user *end;
4699 addr = (void __user *)uattr + sizeof(*attr);
4700 end = (void __user *)uattr + size;
4702 for (; addr < end; addr++) {
4703 ret = get_user(val, addr);
4709 size = sizeof(*attr);
4712 ret = copy_from_user(attr, uattr, size);
4717 * If the type exists, the corresponding creation will verify
4720 if (attr->type >= PERF_TYPE_MAX)
4723 if (attr->__reserved_1 || attr->__reserved_2)
4726 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4729 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4736 put_user(sizeof(*attr), &uattr->size);
4741 static int perf_event_set_output(struct perf_event *event, int output_fd)
4743 struct perf_event *output_event = NULL;
4744 struct file *output_file = NULL;
4745 struct perf_event *old_output;
4746 int fput_needed = 0;
4752 output_file = fget_light(output_fd, &fput_needed);
4756 if (output_file->f_op != &perf_fops)
4759 output_event = output_file->private_data;
4761 /* Don't chain output fds */
4762 if (output_event->output)
4765 /* Don't set an output fd when we already have an output channel */
4769 atomic_long_inc(&output_file->f_count);
4772 mutex_lock(&event->mmap_mutex);
4773 old_output = event->output;
4774 rcu_assign_pointer(event->output, output_event);
4775 mutex_unlock(&event->mmap_mutex);
4779 * we need to make sure no existing perf_output_*()
4780 * is still referencing this event.
4783 fput(old_output->filp);
4788 fput_light(output_file, fput_needed);
4793 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4795 * @attr_uptr: event_id type attributes for monitoring/sampling
4798 * @group_fd: group leader event fd
4800 SYSCALL_DEFINE5(perf_event_open,
4801 struct perf_event_attr __user *, attr_uptr,
4802 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4804 struct perf_event *event, *group_leader;
4805 struct perf_event_attr attr;
4806 struct perf_event_context *ctx;
4807 struct file *event_file = NULL;
4808 struct file *group_file = NULL;
4809 int fput_needed = 0;
4810 int fput_needed2 = 0;
4813 /* for future expandability... */
4814 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4817 err = perf_copy_attr(attr_uptr, &attr);
4821 if (!attr.exclude_kernel) {
4822 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4827 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4832 * Get the target context (task or percpu):
4834 ctx = find_get_context(pid, cpu);
4836 return PTR_ERR(ctx);
4839 * Look up the group leader (we will attach this event to it):
4841 group_leader = NULL;
4842 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4844 group_file = fget_light(group_fd, &fput_needed);
4846 goto err_put_context;
4847 if (group_file->f_op != &perf_fops)
4848 goto err_put_context;
4850 group_leader = group_file->private_data;
4852 * Do not allow a recursive hierarchy (this new sibling
4853 * becoming part of another group-sibling):
4855 if (group_leader->group_leader != group_leader)
4856 goto err_put_context;
4858 * Do not allow to attach to a group in a different
4859 * task or CPU context:
4861 if (group_leader->ctx != ctx)
4862 goto err_put_context;
4864 * Only a group leader can be exclusive or pinned
4866 if (attr.exclusive || attr.pinned)
4867 goto err_put_context;
4870 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4871 NULL, NULL, GFP_KERNEL);
4872 err = PTR_ERR(event);
4874 goto err_put_context;
4876 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4878 goto err_free_put_context;
4880 event_file = fget_light(err, &fput_needed2);
4882 goto err_free_put_context;
4884 if (flags & PERF_FLAG_FD_OUTPUT) {
4885 err = perf_event_set_output(event, group_fd);
4887 goto err_fput_free_put_context;
4890 event->filp = event_file;
4891 WARN_ON_ONCE(ctx->parent_ctx);
4892 mutex_lock(&ctx->mutex);
4893 perf_install_in_context(ctx, event, cpu);
4895 mutex_unlock(&ctx->mutex);
4897 event->owner = current;
4898 get_task_struct(current);
4899 mutex_lock(¤t->perf_event_mutex);
4900 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4901 mutex_unlock(¤t->perf_event_mutex);
4903 err_fput_free_put_context:
4904 fput_light(event_file, fput_needed2);
4906 err_free_put_context:
4914 fput_light(group_file, fput_needed);
4920 * perf_event_create_kernel_counter
4922 * @attr: attributes of the counter to create
4923 * @cpu: cpu in which the counter is bound
4924 * @pid: task to profile
4927 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4929 perf_overflow_handler_t overflow_handler)
4931 struct perf_event *event;
4932 struct perf_event_context *ctx;
4936 * Get the target context (task or percpu):
4939 ctx = find_get_context(pid, cpu);
4945 event = perf_event_alloc(attr, cpu, ctx, NULL,
4946 NULL, overflow_handler, GFP_KERNEL);
4947 if (IS_ERR(event)) {
4948 err = PTR_ERR(event);
4949 goto err_put_context;
4953 WARN_ON_ONCE(ctx->parent_ctx);
4954 mutex_lock(&ctx->mutex);
4955 perf_install_in_context(ctx, event, cpu);
4957 mutex_unlock(&ctx->mutex);
4959 event->owner = current;
4960 get_task_struct(current);
4961 mutex_lock(¤t->perf_event_mutex);
4962 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4963 mutex_unlock(¤t->perf_event_mutex);
4970 return ERR_PTR(err);
4972 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4975 * inherit a event from parent task to child task:
4977 static struct perf_event *
4978 inherit_event(struct perf_event *parent_event,
4979 struct task_struct *parent,
4980 struct perf_event_context *parent_ctx,
4981 struct task_struct *child,
4982 struct perf_event *group_leader,
4983 struct perf_event_context *child_ctx)
4985 struct perf_event *child_event;
4988 * Instead of creating recursive hierarchies of events,
4989 * we link inherited events back to the original parent,
4990 * which has a filp for sure, which we use as the reference
4993 if (parent_event->parent)
4994 parent_event = parent_event->parent;
4996 child_event = perf_event_alloc(&parent_event->attr,
4997 parent_event->cpu, child_ctx,
4998 group_leader, parent_event,
5000 if (IS_ERR(child_event))
5005 * Make the child state follow the state of the parent event,
5006 * not its attr.disabled bit. We hold the parent's mutex,
5007 * so we won't race with perf_event_{en, dis}able_family.
5009 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5010 child_event->state = PERF_EVENT_STATE_INACTIVE;
5012 child_event->state = PERF_EVENT_STATE_OFF;
5014 if (parent_event->attr.freq) {
5015 u64 sample_period = parent_event->hw.sample_period;
5016 struct hw_perf_event *hwc = &child_event->hw;
5018 hwc->sample_period = sample_period;
5019 hwc->last_period = sample_period;
5021 atomic64_set(&hwc->period_left, sample_period);
5024 child_event->overflow_handler = parent_event->overflow_handler;
5027 * Link it up in the child's context:
5029 add_event_to_ctx(child_event, child_ctx);
5032 * Get a reference to the parent filp - we will fput it
5033 * when the child event exits. This is safe to do because
5034 * we are in the parent and we know that the filp still
5035 * exists and has a nonzero count:
5037 atomic_long_inc(&parent_event->filp->f_count);
5040 * Link this into the parent event's child list
5042 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5043 mutex_lock(&parent_event->child_mutex);
5044 list_add_tail(&child_event->child_list, &parent_event->child_list);
5045 mutex_unlock(&parent_event->child_mutex);
5050 static int inherit_group(struct perf_event *parent_event,
5051 struct task_struct *parent,
5052 struct perf_event_context *parent_ctx,
5053 struct task_struct *child,
5054 struct perf_event_context *child_ctx)
5056 struct perf_event *leader;
5057 struct perf_event *sub;
5058 struct perf_event *child_ctr;
5060 leader = inherit_event(parent_event, parent, parent_ctx,
5061 child, NULL, child_ctx);
5063 return PTR_ERR(leader);
5064 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5065 child_ctr = inherit_event(sub, parent, parent_ctx,
5066 child, leader, child_ctx);
5067 if (IS_ERR(child_ctr))
5068 return PTR_ERR(child_ctr);
5073 static void sync_child_event(struct perf_event *child_event,
5074 struct task_struct *child)
5076 struct perf_event *parent_event = child_event->parent;
5079 if (child_event->attr.inherit_stat)
5080 perf_event_read_event(child_event, child);
5082 child_val = atomic64_read(&child_event->count);
5085 * Add back the child's count to the parent's count:
5087 atomic64_add(child_val, &parent_event->count);
5088 atomic64_add(child_event->total_time_enabled,
5089 &parent_event->child_total_time_enabled);
5090 atomic64_add(child_event->total_time_running,
5091 &parent_event->child_total_time_running);
5094 * Remove this event from the parent's list
5096 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5097 mutex_lock(&parent_event->child_mutex);
5098 list_del_init(&child_event->child_list);
5099 mutex_unlock(&parent_event->child_mutex);
5102 * Release the parent event, if this was the last
5105 fput(parent_event->filp);
5109 __perf_event_exit_task(struct perf_event *child_event,
5110 struct perf_event_context *child_ctx,
5111 struct task_struct *child)
5113 struct perf_event *parent_event;
5115 perf_event_remove_from_context(child_event);
5117 parent_event = child_event->parent;
5119 * It can happen that parent exits first, and has events
5120 * that are still around due to the child reference. These
5121 * events need to be zapped - but otherwise linger.
5124 sync_child_event(child_event, child);
5125 free_event(child_event);
5130 * When a child task exits, feed back event values to parent events.
5132 void perf_event_exit_task(struct task_struct *child)
5134 struct perf_event *child_event, *tmp;
5135 struct perf_event_context *child_ctx;
5136 unsigned long flags;
5138 if (likely(!child->perf_event_ctxp)) {
5139 perf_event_task(child, NULL, 0);
5143 local_irq_save(flags);
5145 * We can't reschedule here because interrupts are disabled,
5146 * and either child is current or it is a task that can't be
5147 * scheduled, so we are now safe from rescheduling changing
5150 child_ctx = child->perf_event_ctxp;
5151 __perf_event_task_sched_out(child_ctx);
5154 * Take the context lock here so that if find_get_context is
5155 * reading child->perf_event_ctxp, we wait until it has
5156 * incremented the context's refcount before we do put_ctx below.
5158 raw_spin_lock(&child_ctx->lock);
5159 child->perf_event_ctxp = NULL;
5161 * If this context is a clone; unclone it so it can't get
5162 * swapped to another process while we're removing all
5163 * the events from it.
5165 unclone_ctx(child_ctx);
5166 update_context_time(child_ctx);
5167 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5170 * Report the task dead after unscheduling the events so that we
5171 * won't get any samples after PERF_RECORD_EXIT. We can however still
5172 * get a few PERF_RECORD_READ events.
5174 perf_event_task(child, child_ctx, 0);
5177 * We can recurse on the same lock type through:
5179 * __perf_event_exit_task()
5180 * sync_child_event()
5181 * fput(parent_event->filp)
5183 * mutex_lock(&ctx->mutex)
5185 * But since its the parent context it won't be the same instance.
5187 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5190 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5192 __perf_event_exit_task(child_event, child_ctx, child);
5194 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5196 __perf_event_exit_task(child_event, child_ctx, child);
5199 * If the last event was a group event, it will have appended all
5200 * its siblings to the list, but we obtained 'tmp' before that which
5201 * will still point to the list head terminating the iteration.
5203 if (!list_empty(&child_ctx->pinned_groups) ||
5204 !list_empty(&child_ctx->flexible_groups))
5207 mutex_unlock(&child_ctx->mutex);
5212 static void perf_free_event(struct perf_event *event,
5213 struct perf_event_context *ctx)
5215 struct perf_event *parent = event->parent;
5217 if (WARN_ON_ONCE(!parent))
5220 mutex_lock(&parent->child_mutex);
5221 list_del_init(&event->child_list);
5222 mutex_unlock(&parent->child_mutex);
5226 list_del_event(event, ctx);
5231 * free an unexposed, unused context as created by inheritance by
5232 * init_task below, used by fork() in case of fail.
5234 void perf_event_free_task(struct task_struct *task)
5236 struct perf_event_context *ctx = task->perf_event_ctxp;
5237 struct perf_event *event, *tmp;
5242 mutex_lock(&ctx->mutex);
5244 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5245 perf_free_event(event, ctx);
5247 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5249 perf_free_event(event, ctx);
5251 if (!list_empty(&ctx->pinned_groups) ||
5252 !list_empty(&ctx->flexible_groups))
5255 mutex_unlock(&ctx->mutex);
5261 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5262 struct perf_event_context *parent_ctx,
5263 struct task_struct *child,
5267 struct perf_event_context *child_ctx = child->perf_event_ctxp;
5269 if (!event->attr.inherit) {
5276 * This is executed from the parent task context, so
5277 * inherit events that have been marked for cloning.
5278 * First allocate and initialize a context for the
5282 child_ctx = kzalloc(sizeof(struct perf_event_context),
5287 __perf_event_init_context(child_ctx, child);
5288 child->perf_event_ctxp = child_ctx;
5289 get_task_struct(child);
5292 ret = inherit_group(event, parent, parent_ctx,
5303 * Initialize the perf_event context in task_struct
5305 int perf_event_init_task(struct task_struct *child)
5307 struct perf_event_context *child_ctx, *parent_ctx;
5308 struct perf_event_context *cloned_ctx;
5309 struct perf_event *event;
5310 struct task_struct *parent = current;
5311 int inherited_all = 1;
5314 child->perf_event_ctxp = NULL;
5316 mutex_init(&child->perf_event_mutex);
5317 INIT_LIST_HEAD(&child->perf_event_list);
5319 if (likely(!parent->perf_event_ctxp))
5323 * If the parent's context is a clone, pin it so it won't get
5326 parent_ctx = perf_pin_task_context(parent);
5329 * No need to check if parent_ctx != NULL here; since we saw
5330 * it non-NULL earlier, the only reason for it to become NULL
5331 * is if we exit, and since we're currently in the middle of
5332 * a fork we can't be exiting at the same time.
5336 * Lock the parent list. No need to lock the child - not PID
5337 * hashed yet and not running, so nobody can access it.
5339 mutex_lock(&parent_ctx->mutex);
5342 * We dont have to disable NMIs - we are only looking at
5343 * the list, not manipulating it:
5345 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5346 ret = inherit_task_group(event, parent, parent_ctx, child,
5352 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5353 ret = inherit_task_group(event, parent, parent_ctx, child,
5359 child_ctx = child->perf_event_ctxp;
5361 if (child_ctx && inherited_all) {
5363 * Mark the child context as a clone of the parent
5364 * context, or of whatever the parent is a clone of.
5365 * Note that if the parent is a clone, it could get
5366 * uncloned at any point, but that doesn't matter
5367 * because the list of events and the generation
5368 * count can't have changed since we took the mutex.
5370 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5372 child_ctx->parent_ctx = cloned_ctx;
5373 child_ctx->parent_gen = parent_ctx->parent_gen;
5375 child_ctx->parent_ctx = parent_ctx;
5376 child_ctx->parent_gen = parent_ctx->generation;
5378 get_ctx(child_ctx->parent_ctx);
5381 mutex_unlock(&parent_ctx->mutex);
5383 perf_unpin_context(parent_ctx);
5388 static void __cpuinit perf_event_init_cpu(int cpu)
5390 struct perf_cpu_context *cpuctx;
5392 cpuctx = &per_cpu(perf_cpu_context, cpu);
5393 __perf_event_init_context(&cpuctx->ctx, NULL);
5395 spin_lock(&perf_resource_lock);
5396 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5397 spin_unlock(&perf_resource_lock);
5399 hw_perf_event_setup(cpu);
5402 #ifdef CONFIG_HOTPLUG_CPU
5403 static void __perf_event_exit_cpu(void *info)
5405 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5406 struct perf_event_context *ctx = &cpuctx->ctx;
5407 struct perf_event *event, *tmp;
5409 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5410 __perf_event_remove_from_context(event);
5411 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5412 __perf_event_remove_from_context(event);
5414 static void perf_event_exit_cpu(int cpu)
5416 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5417 struct perf_event_context *ctx = &cpuctx->ctx;
5419 mutex_lock(&ctx->mutex);
5420 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5421 mutex_unlock(&ctx->mutex);
5424 static inline void perf_event_exit_cpu(int cpu) { }
5427 static int __cpuinit
5428 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5430 unsigned int cpu = (long)hcpu;
5434 case CPU_UP_PREPARE:
5435 case CPU_UP_PREPARE_FROZEN:
5436 perf_event_init_cpu(cpu);
5440 case CPU_ONLINE_FROZEN:
5441 hw_perf_event_setup_online(cpu);
5444 case CPU_DOWN_PREPARE:
5445 case CPU_DOWN_PREPARE_FROZEN:
5446 perf_event_exit_cpu(cpu);
5457 * This has to have a higher priority than migration_notifier in sched.c.
5459 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5460 .notifier_call = perf_cpu_notify,
5464 void __init perf_event_init(void)
5466 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5467 (void *)(long)smp_processor_id());
5468 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5469 (void *)(long)smp_processor_id());
5470 register_cpu_notifier(&perf_cpu_nb);
5473 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5475 return sprintf(buf, "%d\n", perf_reserved_percpu);
5479 perf_set_reserve_percpu(struct sysdev_class *class,
5483 struct perf_cpu_context *cpuctx;
5487 err = strict_strtoul(buf, 10, &val);
5490 if (val > perf_max_events)
5493 spin_lock(&perf_resource_lock);
5494 perf_reserved_percpu = val;
5495 for_each_online_cpu(cpu) {
5496 cpuctx = &per_cpu(perf_cpu_context, cpu);
5497 raw_spin_lock_irq(&cpuctx->ctx.lock);
5498 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5499 perf_max_events - perf_reserved_percpu);
5500 cpuctx->max_pertask = mpt;
5501 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5503 spin_unlock(&perf_resource_lock);
5508 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5510 return sprintf(buf, "%d\n", perf_overcommit);
5514 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5519 err = strict_strtoul(buf, 10, &val);
5525 spin_lock(&perf_resource_lock);
5526 perf_overcommit = val;
5527 spin_unlock(&perf_resource_lock);
5532 static SYSDEV_CLASS_ATTR(
5535 perf_show_reserve_percpu,
5536 perf_set_reserve_percpu
5539 static SYSDEV_CLASS_ATTR(
5542 perf_show_overcommit,
5546 static struct attribute *perfclass_attrs[] = {
5547 &attr_reserve_percpu.attr,
5548 &attr_overcommit.attr,
5552 static struct attribute_group perfclass_attr_group = {
5553 .attrs = perfclass_attrs,
5554 .name = "perf_events",
5557 static int __init perf_event_sysfs_init(void)
5559 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5560 &perfclass_attr_group);
5562 device_initcall(perf_event_sysfs_init);