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 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 spin_lock_irqsave(&ctx->lock, *flags);
207 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208 spin_unlock_irqrestore(&ctx->lock, *flags);
212 if (!atomic_inc_not_zero(&ctx->refcount)) {
213 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 spin_unlock_irqrestore(&ctx->lock, flags);
239 static void perf_unpin_context(struct perf_event_context *ctx)
243 spin_lock_irqsave(&ctx->lock, flags);
245 spin_unlock_irqrestore(&ctx->lock, flags);
250 * Add a event from the lists for its context.
251 * Must be called with ctx->mutex and ctx->lock held.
254 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
256 struct perf_event *group_leader = event->group_leader;
259 * Depending on whether it is a standalone or sibling event,
260 * add it straight to the context's event list, or to the group
261 * leader's sibling list:
263 if (group_leader == event)
264 list_add_tail(&event->group_entry, &ctx->group_list);
266 list_add_tail(&event->group_entry, &group_leader->sibling_list);
267 group_leader->nr_siblings++;
270 list_add_rcu(&event->event_entry, &ctx->event_list);
272 if (event->attr.inherit_stat)
277 * Remove a event from the lists for its context.
278 * Must be called with ctx->mutex and ctx->lock held.
281 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
283 struct perf_event *sibling, *tmp;
285 if (list_empty(&event->group_entry))
288 if (event->attr.inherit_stat)
291 list_del_init(&event->group_entry);
292 list_del_rcu(&event->event_entry);
294 if (event->group_leader != event)
295 event->group_leader->nr_siblings--;
298 * If this was a group event with sibling events then
299 * upgrade the siblings to singleton events by adding them
300 * to the context list directly:
302 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
304 list_move_tail(&sibling->group_entry, &ctx->group_list);
305 sibling->group_leader = sibling;
310 event_sched_out(struct perf_event *event,
311 struct perf_cpu_context *cpuctx,
312 struct perf_event_context *ctx)
314 if (event->state != PERF_EVENT_STATE_ACTIVE)
317 event->state = PERF_EVENT_STATE_INACTIVE;
318 if (event->pending_disable) {
319 event->pending_disable = 0;
320 event->state = PERF_EVENT_STATE_OFF;
322 event->tstamp_stopped = ctx->time;
323 event->pmu->disable(event);
326 if (!is_software_event(event))
327 cpuctx->active_oncpu--;
329 if (event->attr.exclusive || !cpuctx->active_oncpu)
330 cpuctx->exclusive = 0;
334 group_sched_out(struct perf_event *group_event,
335 struct perf_cpu_context *cpuctx,
336 struct perf_event_context *ctx)
338 struct perf_event *event;
340 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
343 event_sched_out(group_event, cpuctx, ctx);
346 * Schedule out siblings (if any):
348 list_for_each_entry(event, &group_event->sibling_list, group_entry)
349 event_sched_out(event, cpuctx, ctx);
351 if (group_event->attr.exclusive)
352 cpuctx->exclusive = 0;
356 * Cross CPU call to remove a performance event
358 * We disable the event on the hardware level first. After that we
359 * remove it from the context list.
361 static void __perf_event_remove_from_context(void *info)
363 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
364 struct perf_event *event = info;
365 struct perf_event_context *ctx = event->ctx;
368 * If this is a task context, we need to check whether it is
369 * the current task context of this cpu. If not it has been
370 * scheduled out before the smp call arrived.
372 if (ctx->task && cpuctx->task_ctx != ctx)
375 spin_lock(&ctx->lock);
377 * Protect the list operation against NMI by disabling the
378 * events on a global level.
382 event_sched_out(event, cpuctx, ctx);
384 list_del_event(event, ctx);
388 * Allow more per task events with respect to the
391 cpuctx->max_pertask =
392 min(perf_max_events - ctx->nr_events,
393 perf_max_events - perf_reserved_percpu);
397 spin_unlock(&ctx->lock);
402 * Remove the event from a task's (or a CPU's) list of events.
404 * Must be called with ctx->mutex held.
406 * CPU events are removed with a smp call. For task events we only
407 * call when the task is on a CPU.
409 * If event->ctx is a cloned context, callers must make sure that
410 * every task struct that event->ctx->task could possibly point to
411 * remains valid. This is OK when called from perf_release since
412 * that only calls us on the top-level context, which can't be a clone.
413 * When called from perf_event_exit_task, it's OK because the
414 * context has been detached from its task.
416 static void perf_event_remove_from_context(struct perf_event *event)
418 struct perf_event_context *ctx = event->ctx;
419 struct task_struct *task = ctx->task;
423 * Per cpu events are removed via an smp call and
424 * the removal is always sucessful.
426 smp_call_function_single(event->cpu,
427 __perf_event_remove_from_context,
433 task_oncpu_function_call(task, __perf_event_remove_from_context,
436 spin_lock_irq(&ctx->lock);
438 * If the context is active we need to retry the smp call.
440 if (ctx->nr_active && !list_empty(&event->group_entry)) {
441 spin_unlock_irq(&ctx->lock);
446 * The lock prevents that this context is scheduled in so we
447 * can remove the event safely, if the call above did not
450 if (!list_empty(&event->group_entry)) {
451 list_del_event(event, ctx);
453 spin_unlock_irq(&ctx->lock);
456 static inline u64 perf_clock(void)
458 return cpu_clock(smp_processor_id());
462 * Update the record of the current time in a context.
464 static void update_context_time(struct perf_event_context *ctx)
466 u64 now = perf_clock();
468 ctx->time += now - ctx->timestamp;
469 ctx->timestamp = now;
473 * Update the total_time_enabled and total_time_running fields for a event.
475 static void update_event_times(struct perf_event *event)
477 struct perf_event_context *ctx = event->ctx;
480 if (event->state < PERF_EVENT_STATE_INACTIVE ||
481 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
484 event->total_time_enabled = ctx->time - event->tstamp_enabled;
486 if (event->state == PERF_EVENT_STATE_INACTIVE)
487 run_end = event->tstamp_stopped;
491 event->total_time_running = run_end - event->tstamp_running;
495 * Update total_time_enabled and total_time_running for all events in a group.
497 static void update_group_times(struct perf_event *leader)
499 struct perf_event *event;
501 update_event_times(leader);
502 list_for_each_entry(event, &leader->sibling_list, group_entry)
503 update_event_times(event);
507 * Cross CPU call to disable a performance event
509 static void __perf_event_disable(void *info)
511 struct perf_event *event = info;
512 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
513 struct perf_event_context *ctx = event->ctx;
516 * If this is a per-task event, need to check whether this
517 * event's task is the current task on this cpu.
519 if (ctx->task && cpuctx->task_ctx != ctx)
522 spin_lock(&ctx->lock);
525 * If the event is on, turn it off.
526 * If it is in error state, leave it in error state.
528 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
529 update_context_time(ctx);
530 update_group_times(event);
531 if (event == event->group_leader)
532 group_sched_out(event, cpuctx, ctx);
534 event_sched_out(event, cpuctx, ctx);
535 event->state = PERF_EVENT_STATE_OFF;
538 spin_unlock(&ctx->lock);
544 * If event->ctx is a cloned context, callers must make sure that
545 * every task struct that event->ctx->task could possibly point to
546 * remains valid. This condition is satisifed when called through
547 * perf_event_for_each_child or perf_event_for_each because they
548 * hold the top-level event's child_mutex, so any descendant that
549 * goes to exit will block in sync_child_event.
550 * When called from perf_pending_event it's OK because event->ctx
551 * is the current context on this CPU and preemption is disabled,
552 * hence we can't get into perf_event_task_sched_out for this context.
554 static void perf_event_disable(struct perf_event *event)
556 struct perf_event_context *ctx = event->ctx;
557 struct task_struct *task = ctx->task;
561 * Disable the event on the cpu that it's on
563 smp_call_function_single(event->cpu, __perf_event_disable,
569 task_oncpu_function_call(task, __perf_event_disable, event);
571 spin_lock_irq(&ctx->lock);
573 * If the event is still active, we need to retry the cross-call.
575 if (event->state == PERF_EVENT_STATE_ACTIVE) {
576 spin_unlock_irq(&ctx->lock);
581 * Since we have the lock this context can't be scheduled
582 * in, so we can change the state safely.
584 if (event->state == PERF_EVENT_STATE_INACTIVE) {
585 update_group_times(event);
586 event->state = PERF_EVENT_STATE_OFF;
589 spin_unlock_irq(&ctx->lock);
593 event_sched_in(struct perf_event *event,
594 struct perf_cpu_context *cpuctx,
595 struct perf_event_context *ctx,
598 if (event->state <= PERF_EVENT_STATE_OFF)
601 event->state = PERF_EVENT_STATE_ACTIVE;
602 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
604 * The new state must be visible before we turn it on in the hardware:
608 if (event->pmu->enable(event)) {
609 event->state = PERF_EVENT_STATE_INACTIVE;
614 event->tstamp_running += ctx->time - event->tstamp_stopped;
616 if (!is_software_event(event))
617 cpuctx->active_oncpu++;
620 if (event->attr.exclusive)
621 cpuctx->exclusive = 1;
627 group_sched_in(struct perf_event *group_event,
628 struct perf_cpu_context *cpuctx,
629 struct perf_event_context *ctx,
632 struct perf_event *event, *partial_group;
635 if (group_event->state == PERF_EVENT_STATE_OFF)
638 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
640 return ret < 0 ? ret : 0;
642 if (event_sched_in(group_event, cpuctx, ctx, cpu))
646 * Schedule in siblings as one group (if any):
648 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
649 if (event_sched_in(event, cpuctx, ctx, cpu)) {
650 partial_group = event;
659 * Groups can be scheduled in as one unit only, so undo any
660 * partial group before returning:
662 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
663 if (event == partial_group)
665 event_sched_out(event, cpuctx, ctx);
667 event_sched_out(group_event, cpuctx, ctx);
673 * Return 1 for a group consisting entirely of software events,
674 * 0 if the group contains any hardware events.
676 static int is_software_only_group(struct perf_event *leader)
678 struct perf_event *event;
680 if (!is_software_event(leader))
683 list_for_each_entry(event, &leader->sibling_list, group_entry)
684 if (!is_software_event(event))
691 * Work out whether we can put this event group on the CPU now.
693 static int group_can_go_on(struct perf_event *event,
694 struct perf_cpu_context *cpuctx,
698 * Groups consisting entirely of software events can always go on.
700 if (is_software_only_group(event))
703 * If an exclusive group is already on, no other hardware
706 if (cpuctx->exclusive)
709 * If this group is exclusive and there are already
710 * events on the CPU, it can't go on.
712 if (event->attr.exclusive && cpuctx->active_oncpu)
715 * Otherwise, try to add it if all previous groups were able
721 static void add_event_to_ctx(struct perf_event *event,
722 struct perf_event_context *ctx)
724 list_add_event(event, ctx);
725 event->tstamp_enabled = ctx->time;
726 event->tstamp_running = ctx->time;
727 event->tstamp_stopped = ctx->time;
731 * Cross CPU call to install and enable a performance event
733 * Must be called with ctx->mutex held
735 static void __perf_install_in_context(void *info)
737 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
738 struct perf_event *event = info;
739 struct perf_event_context *ctx = event->ctx;
740 struct perf_event *leader = event->group_leader;
741 int cpu = smp_processor_id();
745 * If this is a task context, we need to check whether it is
746 * the current task context of this cpu. If not it has been
747 * scheduled out before the smp call arrived.
748 * Or possibly this is the right context but it isn't
749 * on this cpu because it had no events.
751 if (ctx->task && cpuctx->task_ctx != ctx) {
752 if (cpuctx->task_ctx || ctx->task != current)
754 cpuctx->task_ctx = ctx;
757 spin_lock(&ctx->lock);
759 update_context_time(ctx);
762 * Protect the list operation against NMI by disabling the
763 * events on a global level. NOP for non NMI based events.
767 add_event_to_ctx(event, ctx);
770 * Don't put the event on if it is disabled or if
771 * it is in a group and the group isn't on.
773 if (event->state != PERF_EVENT_STATE_INACTIVE ||
774 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
778 * An exclusive event can't go on if there are already active
779 * hardware events, and no hardware event can go on if there
780 * is already an exclusive event on.
782 if (!group_can_go_on(event, cpuctx, 1))
785 err = event_sched_in(event, cpuctx, ctx, cpu);
789 * This event couldn't go on. If it is in a group
790 * then we have to pull the whole group off.
791 * If the event group is pinned then put it in error state.
794 group_sched_out(leader, cpuctx, ctx);
795 if (leader->attr.pinned) {
796 update_group_times(leader);
797 leader->state = PERF_EVENT_STATE_ERROR;
801 if (!err && !ctx->task && cpuctx->max_pertask)
802 cpuctx->max_pertask--;
807 spin_unlock(&ctx->lock);
811 * Attach a performance event to a context
813 * First we add the event to the list with the hardware enable bit
814 * in event->hw_config cleared.
816 * If the event is attached to a task which is on a CPU we use a smp
817 * call to enable it in the task context. The task might have been
818 * scheduled away, but we check this in the smp call again.
820 * Must be called with ctx->mutex held.
823 perf_install_in_context(struct perf_event_context *ctx,
824 struct perf_event *event,
827 struct task_struct *task = ctx->task;
831 * Per cpu events are installed via an smp call and
832 * the install is always sucessful.
834 smp_call_function_single(cpu, __perf_install_in_context,
840 task_oncpu_function_call(task, __perf_install_in_context,
843 spin_lock_irq(&ctx->lock);
845 * we need to retry the smp call.
847 if (ctx->is_active && list_empty(&event->group_entry)) {
848 spin_unlock_irq(&ctx->lock);
853 * The lock prevents that this context is scheduled in so we
854 * can add the event safely, if it the call above did not
857 if (list_empty(&event->group_entry))
858 add_event_to_ctx(event, ctx);
859 spin_unlock_irq(&ctx->lock);
863 * Put a event into inactive state and update time fields.
864 * Enabling the leader of a group effectively enables all
865 * the group members that aren't explicitly disabled, so we
866 * have to update their ->tstamp_enabled also.
867 * Note: this works for group members as well as group leaders
868 * since the non-leader members' sibling_lists will be empty.
870 static void __perf_event_mark_enabled(struct perf_event *event,
871 struct perf_event_context *ctx)
873 struct perf_event *sub;
875 event->state = PERF_EVENT_STATE_INACTIVE;
876 event->tstamp_enabled = ctx->time - event->total_time_enabled;
877 list_for_each_entry(sub, &event->sibling_list, group_entry)
878 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
879 sub->tstamp_enabled =
880 ctx->time - sub->total_time_enabled;
884 * Cross CPU call to enable a performance event
886 static void __perf_event_enable(void *info)
888 struct perf_event *event = info;
889 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
890 struct perf_event_context *ctx = event->ctx;
891 struct perf_event *leader = event->group_leader;
895 * If this is a per-task event, need to check whether this
896 * event's task is the current task on this cpu.
898 if (ctx->task && cpuctx->task_ctx != ctx) {
899 if (cpuctx->task_ctx || ctx->task != current)
901 cpuctx->task_ctx = ctx;
904 spin_lock(&ctx->lock);
906 update_context_time(ctx);
908 if (event->state >= PERF_EVENT_STATE_INACTIVE)
910 __perf_event_mark_enabled(event, ctx);
913 * If the event is in a group and isn't the group leader,
914 * then don't put it on unless the group is on.
916 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
919 if (!group_can_go_on(event, cpuctx, 1)) {
924 err = group_sched_in(event, cpuctx, ctx,
927 err = event_sched_in(event, cpuctx, ctx,
934 * If this event can't go on and it's part of a
935 * group, then the whole group has to come off.
938 group_sched_out(leader, cpuctx, ctx);
939 if (leader->attr.pinned) {
940 update_group_times(leader);
941 leader->state = PERF_EVENT_STATE_ERROR;
946 spin_unlock(&ctx->lock);
952 * If event->ctx is a cloned context, callers must make sure that
953 * every task struct that event->ctx->task could possibly point to
954 * remains valid. This condition is satisfied when called through
955 * perf_event_for_each_child or perf_event_for_each as described
956 * for perf_event_disable.
958 static void perf_event_enable(struct perf_event *event)
960 struct perf_event_context *ctx = event->ctx;
961 struct task_struct *task = ctx->task;
965 * Enable the event on the cpu that it's on
967 smp_call_function_single(event->cpu, __perf_event_enable,
972 spin_lock_irq(&ctx->lock);
973 if (event->state >= PERF_EVENT_STATE_INACTIVE)
977 * If the event is in error state, clear that first.
978 * That way, if we see the event in error state below, we
979 * know that it has gone back into error state, as distinct
980 * from the task having been scheduled away before the
981 * cross-call arrived.
983 if (event->state == PERF_EVENT_STATE_ERROR)
984 event->state = PERF_EVENT_STATE_OFF;
987 spin_unlock_irq(&ctx->lock);
988 task_oncpu_function_call(task, __perf_event_enable, event);
990 spin_lock_irq(&ctx->lock);
993 * If the context is active and the event is still off,
994 * we need to retry the cross-call.
996 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1000 * Since we have the lock this context can't be scheduled
1001 * in, so we can change the state safely.
1003 if (event->state == PERF_EVENT_STATE_OFF)
1004 __perf_event_mark_enabled(event, ctx);
1007 spin_unlock_irq(&ctx->lock);
1010 static int perf_event_refresh(struct perf_event *event, int refresh)
1013 * not supported on inherited events
1015 if (event->attr.inherit)
1018 atomic_add(refresh, &event->event_limit);
1019 perf_event_enable(event);
1024 void __perf_event_sched_out(struct perf_event_context *ctx,
1025 struct perf_cpu_context *cpuctx)
1027 struct perf_event *event;
1029 spin_lock(&ctx->lock);
1031 if (likely(!ctx->nr_events))
1033 update_context_time(ctx);
1037 list_for_each_entry(event, &ctx->group_list, group_entry)
1038 group_sched_out(event, cpuctx, ctx);
1042 spin_unlock(&ctx->lock);
1046 * Test whether two contexts are equivalent, i.e. whether they
1047 * have both been cloned from the same version of the same context
1048 * and they both have the same number of enabled events.
1049 * If the number of enabled events is the same, then the set
1050 * of enabled events should be the same, because these are both
1051 * inherited contexts, therefore we can't access individual events
1052 * in them directly with an fd; we can only enable/disable all
1053 * events via prctl, or enable/disable all events in a family
1054 * via ioctl, which will have the same effect on both contexts.
1056 static int context_equiv(struct perf_event_context *ctx1,
1057 struct perf_event_context *ctx2)
1059 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1060 && ctx1->parent_gen == ctx2->parent_gen
1061 && !ctx1->pin_count && !ctx2->pin_count;
1064 static void __perf_event_read(void *event);
1066 static void __perf_event_sync_stat(struct perf_event *event,
1067 struct perf_event *next_event)
1071 if (!event->attr.inherit_stat)
1075 * Update the event value, we cannot use perf_event_read()
1076 * because we're in the middle of a context switch and have IRQs
1077 * disabled, which upsets smp_call_function_single(), however
1078 * we know the event must be on the current CPU, therefore we
1079 * don't need to use it.
1081 switch (event->state) {
1082 case PERF_EVENT_STATE_ACTIVE:
1083 __perf_event_read(event);
1086 case PERF_EVENT_STATE_INACTIVE:
1087 update_event_times(event);
1095 * In order to keep per-task stats reliable we need to flip the event
1096 * values when we flip the contexts.
1098 value = atomic64_read(&next_event->count);
1099 value = atomic64_xchg(&event->count, value);
1100 atomic64_set(&next_event->count, value);
1102 swap(event->total_time_enabled, next_event->total_time_enabled);
1103 swap(event->total_time_running, next_event->total_time_running);
1106 * Since we swizzled the values, update the user visible data too.
1108 perf_event_update_userpage(event);
1109 perf_event_update_userpage(next_event);
1112 #define list_next_entry(pos, member) \
1113 list_entry(pos->member.next, typeof(*pos), member)
1115 static void perf_event_sync_stat(struct perf_event_context *ctx,
1116 struct perf_event_context *next_ctx)
1118 struct perf_event *event, *next_event;
1123 event = list_first_entry(&ctx->event_list,
1124 struct perf_event, event_entry);
1126 next_event = list_first_entry(&next_ctx->event_list,
1127 struct perf_event, event_entry);
1129 while (&event->event_entry != &ctx->event_list &&
1130 &next_event->event_entry != &next_ctx->event_list) {
1132 __perf_event_sync_stat(event, next_event);
1134 event = list_next_entry(event, event_entry);
1135 next_event = list_next_entry(next_event, event_entry);
1140 * Called from scheduler to remove the events of the current task,
1141 * with interrupts disabled.
1143 * We stop each event and update the event value in event->count.
1145 * This does not protect us against NMI, but disable()
1146 * sets the disabled bit in the control field of event _before_
1147 * accessing the event control register. If a NMI hits, then it will
1148 * not restart the event.
1150 void perf_event_task_sched_out(struct task_struct *task,
1151 struct task_struct *next, int cpu)
1153 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1154 struct perf_event_context *ctx = task->perf_event_ctxp;
1155 struct perf_event_context *next_ctx;
1156 struct perf_event_context *parent;
1157 struct pt_regs *regs;
1160 regs = task_pt_regs(task);
1161 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1163 if (likely(!ctx || !cpuctx->task_ctx))
1166 update_context_time(ctx);
1169 parent = rcu_dereference(ctx->parent_ctx);
1170 next_ctx = next->perf_event_ctxp;
1171 if (parent && next_ctx &&
1172 rcu_dereference(next_ctx->parent_ctx) == parent) {
1174 * Looks like the two contexts are clones, so we might be
1175 * able to optimize the context switch. We lock both
1176 * contexts and check that they are clones under the
1177 * lock (including re-checking that neither has been
1178 * uncloned in the meantime). It doesn't matter which
1179 * order we take the locks because no other cpu could
1180 * be trying to lock both of these tasks.
1182 spin_lock(&ctx->lock);
1183 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1184 if (context_equiv(ctx, next_ctx)) {
1186 * XXX do we need a memory barrier of sorts
1187 * wrt to rcu_dereference() of perf_event_ctxp
1189 task->perf_event_ctxp = next_ctx;
1190 next->perf_event_ctxp = ctx;
1192 next_ctx->task = task;
1195 perf_event_sync_stat(ctx, next_ctx);
1197 spin_unlock(&next_ctx->lock);
1198 spin_unlock(&ctx->lock);
1203 __perf_event_sched_out(ctx, cpuctx);
1204 cpuctx->task_ctx = NULL;
1209 * Called with IRQs disabled
1211 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1213 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1215 if (!cpuctx->task_ctx)
1218 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1221 __perf_event_sched_out(ctx, cpuctx);
1222 cpuctx->task_ctx = NULL;
1226 * Called with IRQs disabled
1228 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1230 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1234 __perf_event_sched_in(struct perf_event_context *ctx,
1235 struct perf_cpu_context *cpuctx, int cpu)
1237 struct perf_event *event;
1240 spin_lock(&ctx->lock);
1242 if (likely(!ctx->nr_events))
1245 ctx->timestamp = perf_clock();
1250 * First go through the list and put on any pinned groups
1251 * in order to give them the best chance of going on.
1253 list_for_each_entry(event, &ctx->group_list, group_entry) {
1254 if (event->state <= PERF_EVENT_STATE_OFF ||
1255 !event->attr.pinned)
1257 if (event->cpu != -1 && event->cpu != cpu)
1260 if (group_can_go_on(event, cpuctx, 1))
1261 group_sched_in(event, cpuctx, ctx, cpu);
1264 * If this pinned group hasn't been scheduled,
1265 * put it in error state.
1267 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1268 update_group_times(event);
1269 event->state = PERF_EVENT_STATE_ERROR;
1273 list_for_each_entry(event, &ctx->group_list, group_entry) {
1275 * Ignore events in OFF or ERROR state, and
1276 * ignore pinned events since we did them already.
1278 if (event->state <= PERF_EVENT_STATE_OFF ||
1283 * Listen to the 'cpu' scheduling filter constraint
1286 if (event->cpu != -1 && event->cpu != cpu)
1289 if (group_can_go_on(event, cpuctx, can_add_hw))
1290 if (group_sched_in(event, cpuctx, ctx, cpu))
1295 spin_unlock(&ctx->lock);
1299 * Called from scheduler to add the events of the current task
1300 * with interrupts disabled.
1302 * We restore the event value and then enable it.
1304 * This does not protect us against NMI, but enable()
1305 * sets the enabled bit in the control field of event _before_
1306 * accessing the event control register. If a NMI hits, then it will
1307 * keep the event running.
1309 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1311 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1312 struct perf_event_context *ctx = task->perf_event_ctxp;
1316 if (cpuctx->task_ctx == ctx)
1318 __perf_event_sched_in(ctx, cpuctx, cpu);
1319 cpuctx->task_ctx = ctx;
1322 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1324 struct perf_event_context *ctx = &cpuctx->ctx;
1326 __perf_event_sched_in(ctx, cpuctx, cpu);
1329 #define MAX_INTERRUPTS (~0ULL)
1331 static void perf_log_throttle(struct perf_event *event, int enable);
1333 static void perf_adjust_period(struct perf_event *event, u64 events)
1335 struct hw_perf_event *hwc = &event->hw;
1336 u64 period, sample_period;
1339 events *= hwc->sample_period;
1340 period = div64_u64(events, event->attr.sample_freq);
1342 delta = (s64)(period - hwc->sample_period);
1343 delta = (delta + 7) / 8; /* low pass filter */
1345 sample_period = hwc->sample_period + delta;
1350 hwc->sample_period = sample_period;
1353 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1355 struct perf_event *event;
1356 struct hw_perf_event *hwc;
1357 u64 interrupts, freq;
1359 spin_lock(&ctx->lock);
1360 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1361 if (event->state != PERF_EVENT_STATE_ACTIVE)
1366 interrupts = hwc->interrupts;
1367 hwc->interrupts = 0;
1370 * unthrottle events on the tick
1372 if (interrupts == MAX_INTERRUPTS) {
1373 perf_log_throttle(event, 1);
1374 event->pmu->unthrottle(event);
1375 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1378 if (!event->attr.freq || !event->attr.sample_freq)
1382 * if the specified freq < HZ then we need to skip ticks
1384 if (event->attr.sample_freq < HZ) {
1385 freq = event->attr.sample_freq;
1387 hwc->freq_count += freq;
1388 hwc->freq_interrupts += interrupts;
1390 if (hwc->freq_count < HZ)
1393 interrupts = hwc->freq_interrupts;
1394 hwc->freq_interrupts = 0;
1395 hwc->freq_count -= HZ;
1399 perf_adjust_period(event, freq * interrupts);
1402 * In order to avoid being stalled by an (accidental) huge
1403 * sample period, force reset the sample period if we didn't
1404 * get any events in this freq period.
1408 event->pmu->disable(event);
1409 atomic64_set(&hwc->period_left, 0);
1410 event->pmu->enable(event);
1414 spin_unlock(&ctx->lock);
1418 * Round-robin a context's events:
1420 static void rotate_ctx(struct perf_event_context *ctx)
1422 struct perf_event *event;
1424 if (!ctx->nr_events)
1427 spin_lock(&ctx->lock);
1429 * Rotate the first entry last (works just fine for group events too):
1432 list_for_each_entry(event, &ctx->group_list, group_entry) {
1433 list_move_tail(&event->group_entry, &ctx->group_list);
1438 spin_unlock(&ctx->lock);
1441 void perf_event_task_tick(struct task_struct *curr, int cpu)
1443 struct perf_cpu_context *cpuctx;
1444 struct perf_event_context *ctx;
1446 if (!atomic_read(&nr_events))
1449 cpuctx = &per_cpu(perf_cpu_context, cpu);
1450 ctx = curr->perf_event_ctxp;
1452 perf_ctx_adjust_freq(&cpuctx->ctx);
1454 perf_ctx_adjust_freq(ctx);
1456 perf_event_cpu_sched_out(cpuctx);
1458 __perf_event_task_sched_out(ctx);
1460 rotate_ctx(&cpuctx->ctx);
1464 perf_event_cpu_sched_in(cpuctx, cpu);
1466 perf_event_task_sched_in(curr, cpu);
1470 * Enable all of a task's events that have been marked enable-on-exec.
1471 * This expects task == current.
1473 static void perf_event_enable_on_exec(struct task_struct *task)
1475 struct perf_event_context *ctx;
1476 struct perf_event *event;
1477 unsigned long flags;
1480 local_irq_save(flags);
1481 ctx = task->perf_event_ctxp;
1482 if (!ctx || !ctx->nr_events)
1485 __perf_event_task_sched_out(ctx);
1487 spin_lock(&ctx->lock);
1489 list_for_each_entry(event, &ctx->group_list, group_entry) {
1490 if (!event->attr.enable_on_exec)
1492 event->attr.enable_on_exec = 0;
1493 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1495 __perf_event_mark_enabled(event, ctx);
1500 * Unclone this context if we enabled any event.
1505 spin_unlock(&ctx->lock);
1507 perf_event_task_sched_in(task, smp_processor_id());
1509 local_irq_restore(flags);
1513 * Cross CPU call to read the hardware event
1515 static void __perf_event_read(void *info)
1517 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1518 struct perf_event *event = info;
1519 struct perf_event_context *ctx = event->ctx;
1520 unsigned long flags;
1523 * If this is a task context, we need to check whether it is
1524 * the current task context of this cpu. If not it has been
1525 * scheduled out before the smp call arrived. In that case
1526 * event->count would have been updated to a recent sample
1527 * when the event was scheduled out.
1529 if (ctx->task && cpuctx->task_ctx != ctx)
1532 local_irq_save(flags);
1534 update_context_time(ctx);
1535 event->pmu->read(event);
1536 update_event_times(event);
1537 local_irq_restore(flags);
1540 static u64 perf_event_read(struct perf_event *event)
1543 * If event is enabled and currently active on a CPU, update the
1544 * value in the event structure:
1546 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1547 smp_call_function_single(event->oncpu,
1548 __perf_event_read, event, 1);
1549 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1550 update_event_times(event);
1553 return atomic64_read(&event->count);
1557 * Initialize the perf_event context in a task_struct:
1560 __perf_event_init_context(struct perf_event_context *ctx,
1561 struct task_struct *task)
1563 memset(ctx, 0, sizeof(*ctx));
1564 spin_lock_init(&ctx->lock);
1565 mutex_init(&ctx->mutex);
1566 INIT_LIST_HEAD(&ctx->group_list);
1567 INIT_LIST_HEAD(&ctx->event_list);
1568 atomic_set(&ctx->refcount, 1);
1572 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1574 struct perf_event_context *ctx;
1575 struct perf_cpu_context *cpuctx;
1576 struct task_struct *task;
1577 unsigned long flags;
1581 * If cpu is not a wildcard then this is a percpu event:
1584 /* Must be root to operate on a CPU event: */
1585 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1586 return ERR_PTR(-EACCES);
1588 if (cpu < 0 || cpu > num_possible_cpus())
1589 return ERR_PTR(-EINVAL);
1592 * We could be clever and allow to attach a event to an
1593 * offline CPU and activate it when the CPU comes up, but
1596 if (!cpu_isset(cpu, cpu_online_map))
1597 return ERR_PTR(-ENODEV);
1599 cpuctx = &per_cpu(perf_cpu_context, cpu);
1610 task = find_task_by_vpid(pid);
1612 get_task_struct(task);
1616 return ERR_PTR(-ESRCH);
1619 * Can't attach events to a dying task.
1622 if (task->flags & PF_EXITING)
1625 /* Reuse ptrace permission checks for now. */
1627 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1631 ctx = perf_lock_task_context(task, &flags);
1634 spin_unlock_irqrestore(&ctx->lock, flags);
1638 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1642 __perf_event_init_context(ctx, task);
1644 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1646 * We raced with some other task; use
1647 * the context they set.
1652 get_task_struct(task);
1655 put_task_struct(task);
1659 put_task_struct(task);
1660 return ERR_PTR(err);
1663 static void perf_event_free_filter(struct perf_event *event);
1665 static void free_event_rcu(struct rcu_head *head)
1667 struct perf_event *event;
1669 event = container_of(head, struct perf_event, rcu_head);
1671 put_pid_ns(event->ns);
1672 perf_event_free_filter(event);
1676 static void perf_pending_sync(struct perf_event *event);
1678 static void free_event(struct perf_event *event)
1680 perf_pending_sync(event);
1682 if (!event->parent) {
1683 atomic_dec(&nr_events);
1684 if (event->attr.mmap)
1685 atomic_dec(&nr_mmap_events);
1686 if (event->attr.comm)
1687 atomic_dec(&nr_comm_events);
1688 if (event->attr.task)
1689 atomic_dec(&nr_task_events);
1692 if (event->output) {
1693 fput(event->output->filp);
1694 event->output = NULL;
1698 event->destroy(event);
1700 put_ctx(event->ctx);
1701 call_rcu(&event->rcu_head, free_event_rcu);
1705 * Called when the last reference to the file is gone.
1707 static int perf_release(struct inode *inode, struct file *file)
1709 struct perf_event *event = file->private_data;
1710 struct perf_event_context *ctx = event->ctx;
1712 file->private_data = NULL;
1714 WARN_ON_ONCE(ctx->parent_ctx);
1715 mutex_lock(&ctx->mutex);
1716 perf_event_remove_from_context(event);
1717 mutex_unlock(&ctx->mutex);
1719 mutex_lock(&event->owner->perf_event_mutex);
1720 list_del_init(&event->owner_entry);
1721 mutex_unlock(&event->owner->perf_event_mutex);
1722 put_task_struct(event->owner);
1729 int perf_event_release_kernel(struct perf_event *event)
1731 struct perf_event_context *ctx = event->ctx;
1733 WARN_ON_ONCE(ctx->parent_ctx);
1734 mutex_lock(&ctx->mutex);
1735 perf_event_remove_from_context(event);
1736 mutex_unlock(&ctx->mutex);
1738 mutex_lock(&event->owner->perf_event_mutex);
1739 list_del_init(&event->owner_entry);
1740 mutex_unlock(&event->owner->perf_event_mutex);
1741 put_task_struct(event->owner);
1747 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1749 static int perf_event_read_size(struct perf_event *event)
1751 int entry = sizeof(u64); /* value */
1755 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1756 size += sizeof(u64);
1758 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1759 size += sizeof(u64);
1761 if (event->attr.read_format & PERF_FORMAT_ID)
1762 entry += sizeof(u64);
1764 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1765 nr += event->group_leader->nr_siblings;
1766 size += sizeof(u64);
1774 u64 perf_event_read_value(struct perf_event *event)
1776 struct perf_event *child;
1779 total += perf_event_read(event);
1780 list_for_each_entry(child, &event->child_list, child_list)
1781 total += perf_event_read(child);
1785 EXPORT_SYMBOL_GPL(perf_event_read_value);
1787 static int perf_event_read_entry(struct perf_event *event,
1788 u64 read_format, char __user *buf)
1790 int n = 0, count = 0;
1793 values[n++] = perf_event_read_value(event);
1794 if (read_format & PERF_FORMAT_ID)
1795 values[n++] = primary_event_id(event);
1797 count = n * sizeof(u64);
1799 if (copy_to_user(buf, values, count))
1805 static int perf_event_read_group(struct perf_event *event,
1806 u64 read_format, char __user *buf)
1808 struct perf_event *leader = event->group_leader, *sub;
1809 int n = 0, size = 0, err = -EFAULT;
1812 values[n++] = 1 + leader->nr_siblings;
1813 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1814 values[n++] = leader->total_time_enabled +
1815 atomic64_read(&leader->child_total_time_enabled);
1817 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1818 values[n++] = leader->total_time_running +
1819 atomic64_read(&leader->child_total_time_running);
1822 size = n * sizeof(u64);
1824 if (copy_to_user(buf, values, size))
1827 err = perf_event_read_entry(leader, read_format, buf + size);
1833 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1834 err = perf_event_read_entry(sub, read_format,
1845 static int perf_event_read_one(struct perf_event *event,
1846 u64 read_format, char __user *buf)
1851 values[n++] = perf_event_read_value(event);
1852 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1853 values[n++] = event->total_time_enabled +
1854 atomic64_read(&event->child_total_time_enabled);
1856 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1857 values[n++] = event->total_time_running +
1858 atomic64_read(&event->child_total_time_running);
1860 if (read_format & PERF_FORMAT_ID)
1861 values[n++] = primary_event_id(event);
1863 if (copy_to_user(buf, values, n * sizeof(u64)))
1866 return n * sizeof(u64);
1870 * Read the performance event - simple non blocking version for now
1873 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1875 u64 read_format = event->attr.read_format;
1879 * Return end-of-file for a read on a event that is in
1880 * error state (i.e. because it was pinned but it couldn't be
1881 * scheduled on to the CPU at some point).
1883 if (event->state == PERF_EVENT_STATE_ERROR)
1886 if (count < perf_event_read_size(event))
1889 WARN_ON_ONCE(event->ctx->parent_ctx);
1890 mutex_lock(&event->child_mutex);
1891 if (read_format & PERF_FORMAT_GROUP)
1892 ret = perf_event_read_group(event, read_format, buf);
1894 ret = perf_event_read_one(event, read_format, buf);
1895 mutex_unlock(&event->child_mutex);
1901 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1903 struct perf_event *event = file->private_data;
1905 return perf_read_hw(event, buf, count);
1908 static unsigned int perf_poll(struct file *file, poll_table *wait)
1910 struct perf_event *event = file->private_data;
1911 struct perf_mmap_data *data;
1912 unsigned int events = POLL_HUP;
1915 data = rcu_dereference(event->data);
1917 events = atomic_xchg(&data->poll, 0);
1920 poll_wait(file, &event->waitq, wait);
1925 static void perf_event_reset(struct perf_event *event)
1927 (void)perf_event_read(event);
1928 atomic64_set(&event->count, 0);
1929 perf_event_update_userpage(event);
1933 * Holding the top-level event's child_mutex means that any
1934 * descendant process that has inherited this event will block
1935 * in sync_child_event if it goes to exit, thus satisfying the
1936 * task existence requirements of perf_event_enable/disable.
1938 static void perf_event_for_each_child(struct perf_event *event,
1939 void (*func)(struct perf_event *))
1941 struct perf_event *child;
1943 WARN_ON_ONCE(event->ctx->parent_ctx);
1944 mutex_lock(&event->child_mutex);
1946 list_for_each_entry(child, &event->child_list, child_list)
1948 mutex_unlock(&event->child_mutex);
1951 static void perf_event_for_each(struct perf_event *event,
1952 void (*func)(struct perf_event *))
1954 struct perf_event_context *ctx = event->ctx;
1955 struct perf_event *sibling;
1957 WARN_ON_ONCE(ctx->parent_ctx);
1958 mutex_lock(&ctx->mutex);
1959 event = event->group_leader;
1961 perf_event_for_each_child(event, func);
1963 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1964 perf_event_for_each_child(event, func);
1965 mutex_unlock(&ctx->mutex);
1968 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1970 struct perf_event_context *ctx = event->ctx;
1975 if (!event->attr.sample_period)
1978 size = copy_from_user(&value, arg, sizeof(value));
1979 if (size != sizeof(value))
1985 spin_lock_irq(&ctx->lock);
1986 if (event->attr.freq) {
1987 if (value > sysctl_perf_event_sample_rate) {
1992 event->attr.sample_freq = value;
1994 event->attr.sample_period = value;
1995 event->hw.sample_period = value;
1998 spin_unlock_irq(&ctx->lock);
2003 static int perf_event_set_output(struct perf_event *event, int output_fd);
2004 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2006 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2008 struct perf_event *event = file->private_data;
2009 void (*func)(struct perf_event *);
2013 case PERF_EVENT_IOC_ENABLE:
2014 func = perf_event_enable;
2016 case PERF_EVENT_IOC_DISABLE:
2017 func = perf_event_disable;
2019 case PERF_EVENT_IOC_RESET:
2020 func = perf_event_reset;
2023 case PERF_EVENT_IOC_REFRESH:
2024 return perf_event_refresh(event, arg);
2026 case PERF_EVENT_IOC_PERIOD:
2027 return perf_event_period(event, (u64 __user *)arg);
2029 case PERF_EVENT_IOC_SET_OUTPUT:
2030 return perf_event_set_output(event, arg);
2032 case PERF_EVENT_IOC_SET_FILTER:
2033 return perf_event_set_filter(event, (void __user *)arg);
2039 if (flags & PERF_IOC_FLAG_GROUP)
2040 perf_event_for_each(event, func);
2042 perf_event_for_each_child(event, func);
2047 int perf_event_task_enable(void)
2049 struct perf_event *event;
2051 mutex_lock(¤t->perf_event_mutex);
2052 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2053 perf_event_for_each_child(event, perf_event_enable);
2054 mutex_unlock(¤t->perf_event_mutex);
2059 int perf_event_task_disable(void)
2061 struct perf_event *event;
2063 mutex_lock(¤t->perf_event_mutex);
2064 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2065 perf_event_for_each_child(event, perf_event_disable);
2066 mutex_unlock(¤t->perf_event_mutex);
2071 #ifndef PERF_EVENT_INDEX_OFFSET
2072 # define PERF_EVENT_INDEX_OFFSET 0
2075 static int perf_event_index(struct perf_event *event)
2077 if (event->state != PERF_EVENT_STATE_ACTIVE)
2080 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2084 * Callers need to ensure there can be no nesting of this function, otherwise
2085 * the seqlock logic goes bad. We can not serialize this because the arch
2086 * code calls this from NMI context.
2088 void perf_event_update_userpage(struct perf_event *event)
2090 struct perf_event_mmap_page *userpg;
2091 struct perf_mmap_data *data;
2094 data = rcu_dereference(event->data);
2098 userpg = data->user_page;
2101 * Disable preemption so as to not let the corresponding user-space
2102 * spin too long if we get preempted.
2107 userpg->index = perf_event_index(event);
2108 userpg->offset = atomic64_read(&event->count);
2109 if (event->state == PERF_EVENT_STATE_ACTIVE)
2110 userpg->offset -= atomic64_read(&event->hw.prev_count);
2112 userpg->time_enabled = event->total_time_enabled +
2113 atomic64_read(&event->child_total_time_enabled);
2115 userpg->time_running = event->total_time_running +
2116 atomic64_read(&event->child_total_time_running);
2125 static unsigned long perf_data_size(struct perf_mmap_data *data)
2127 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2130 #ifndef CONFIG_PERF_USE_VMALLOC
2133 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2136 static struct page *
2137 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2139 if (pgoff > data->nr_pages)
2143 return virt_to_page(data->user_page);
2145 return virt_to_page(data->data_pages[pgoff - 1]);
2148 static struct perf_mmap_data *
2149 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2151 struct perf_mmap_data *data;
2155 WARN_ON(atomic_read(&event->mmap_count));
2157 size = sizeof(struct perf_mmap_data);
2158 size += nr_pages * sizeof(void *);
2160 data = kzalloc(size, GFP_KERNEL);
2164 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2165 if (!data->user_page)
2166 goto fail_user_page;
2168 for (i = 0; i < nr_pages; i++) {
2169 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2170 if (!data->data_pages[i])
2171 goto fail_data_pages;
2174 data->data_order = 0;
2175 data->nr_pages = nr_pages;
2180 for (i--; i >= 0; i--)
2181 free_page((unsigned long)data->data_pages[i]);
2183 free_page((unsigned long)data->user_page);
2192 static void perf_mmap_free_page(unsigned long addr)
2194 struct page *page = virt_to_page((void *)addr);
2196 page->mapping = NULL;
2200 static void perf_mmap_data_free(struct perf_mmap_data *data)
2204 perf_mmap_free_page((unsigned long)data->user_page);
2205 for (i = 0; i < data->nr_pages; i++)
2206 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2212 * Back perf_mmap() with vmalloc memory.
2214 * Required for architectures that have d-cache aliasing issues.
2217 static struct page *
2218 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2220 if (pgoff > (1UL << data->data_order))
2223 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2226 static void perf_mmap_unmark_page(void *addr)
2228 struct page *page = vmalloc_to_page(addr);
2230 page->mapping = NULL;
2233 static void perf_mmap_data_free_work(struct work_struct *work)
2235 struct perf_mmap_data *data;
2239 data = container_of(work, struct perf_mmap_data, work);
2240 nr = 1 << data->data_order;
2242 base = data->user_page;
2243 for (i = 0; i < nr + 1; i++)
2244 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2249 static void perf_mmap_data_free(struct perf_mmap_data *data)
2251 schedule_work(&data->work);
2254 static struct perf_mmap_data *
2255 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2257 struct perf_mmap_data *data;
2261 WARN_ON(atomic_read(&event->mmap_count));
2263 size = sizeof(struct perf_mmap_data);
2264 size += sizeof(void *);
2266 data = kzalloc(size, GFP_KERNEL);
2270 INIT_WORK(&data->work, perf_mmap_data_free_work);
2272 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2276 data->user_page = all_buf;
2277 data->data_pages[0] = all_buf + PAGE_SIZE;
2278 data->data_order = ilog2(nr_pages);
2292 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2294 struct perf_event *event = vma->vm_file->private_data;
2295 struct perf_mmap_data *data;
2296 int ret = VM_FAULT_SIGBUS;
2298 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2299 if (vmf->pgoff == 0)
2305 data = rcu_dereference(event->data);
2309 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2312 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2316 get_page(vmf->page);
2317 vmf->page->mapping = vma->vm_file->f_mapping;
2318 vmf->page->index = vmf->pgoff;
2328 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2330 long max_size = perf_data_size(data);
2332 atomic_set(&data->lock, -1);
2334 if (event->attr.watermark) {
2335 data->watermark = min_t(long, max_size,
2336 event->attr.wakeup_watermark);
2339 if (!data->watermark)
2340 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2343 rcu_assign_pointer(event->data, data);
2346 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2348 struct perf_mmap_data *data;
2350 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2351 perf_mmap_data_free(data);
2355 static void perf_mmap_data_release(struct perf_event *event)
2357 struct perf_mmap_data *data = event->data;
2359 WARN_ON(atomic_read(&event->mmap_count));
2361 rcu_assign_pointer(event->data, NULL);
2362 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2365 static void perf_mmap_open(struct vm_area_struct *vma)
2367 struct perf_event *event = vma->vm_file->private_data;
2369 atomic_inc(&event->mmap_count);
2372 static void perf_mmap_close(struct vm_area_struct *vma)
2374 struct perf_event *event = vma->vm_file->private_data;
2376 WARN_ON_ONCE(event->ctx->parent_ctx);
2377 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2378 unsigned long size = perf_data_size(event->data);
2379 struct user_struct *user = current_user();
2381 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2382 vma->vm_mm->locked_vm -= event->data->nr_locked;
2383 perf_mmap_data_release(event);
2384 mutex_unlock(&event->mmap_mutex);
2388 static const struct vm_operations_struct perf_mmap_vmops = {
2389 .open = perf_mmap_open,
2390 .close = perf_mmap_close,
2391 .fault = perf_mmap_fault,
2392 .page_mkwrite = perf_mmap_fault,
2395 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2397 struct perf_event *event = file->private_data;
2398 unsigned long user_locked, user_lock_limit;
2399 struct user_struct *user = current_user();
2400 unsigned long locked, lock_limit;
2401 struct perf_mmap_data *data;
2402 unsigned long vma_size;
2403 unsigned long nr_pages;
2404 long user_extra, extra;
2407 if (!(vma->vm_flags & VM_SHARED))
2410 vma_size = vma->vm_end - vma->vm_start;
2411 nr_pages = (vma_size / PAGE_SIZE) - 1;
2414 * If we have data pages ensure they're a power-of-two number, so we
2415 * can do bitmasks instead of modulo.
2417 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2420 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2423 if (vma->vm_pgoff != 0)
2426 WARN_ON_ONCE(event->ctx->parent_ctx);
2427 mutex_lock(&event->mmap_mutex);
2428 if (event->output) {
2433 if (atomic_inc_not_zero(&event->mmap_count)) {
2434 if (nr_pages != event->data->nr_pages)
2439 user_extra = nr_pages + 1;
2440 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2443 * Increase the limit linearly with more CPUs:
2445 user_lock_limit *= num_online_cpus();
2447 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2450 if (user_locked > user_lock_limit)
2451 extra = user_locked - user_lock_limit;
2453 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2454 lock_limit >>= PAGE_SHIFT;
2455 locked = vma->vm_mm->locked_vm + extra;
2457 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2458 !capable(CAP_IPC_LOCK)) {
2463 WARN_ON(event->data);
2465 data = perf_mmap_data_alloc(event, nr_pages);
2471 perf_mmap_data_init(event, data);
2473 atomic_set(&event->mmap_count, 1);
2474 atomic_long_add(user_extra, &user->locked_vm);
2475 vma->vm_mm->locked_vm += extra;
2476 event->data->nr_locked = extra;
2477 if (vma->vm_flags & VM_WRITE)
2478 event->data->writable = 1;
2481 mutex_unlock(&event->mmap_mutex);
2483 vma->vm_flags |= VM_RESERVED;
2484 vma->vm_ops = &perf_mmap_vmops;
2489 static int perf_fasync(int fd, struct file *filp, int on)
2491 struct inode *inode = filp->f_path.dentry->d_inode;
2492 struct perf_event *event = filp->private_data;
2495 mutex_lock(&inode->i_mutex);
2496 retval = fasync_helper(fd, filp, on, &event->fasync);
2497 mutex_unlock(&inode->i_mutex);
2505 static const struct file_operations perf_fops = {
2506 .release = perf_release,
2509 .unlocked_ioctl = perf_ioctl,
2510 .compat_ioctl = perf_ioctl,
2512 .fasync = perf_fasync,
2518 * If there's data, ensure we set the poll() state and publish everything
2519 * to user-space before waking everybody up.
2522 void perf_event_wakeup(struct perf_event *event)
2524 wake_up_all(&event->waitq);
2526 if (event->pending_kill) {
2527 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2528 event->pending_kill = 0;
2535 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2537 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2538 * single linked list and use cmpxchg() to add entries lockless.
2541 static void perf_pending_event(struct perf_pending_entry *entry)
2543 struct perf_event *event = container_of(entry,
2544 struct perf_event, pending);
2546 if (event->pending_disable) {
2547 event->pending_disable = 0;
2548 __perf_event_disable(event);
2551 if (event->pending_wakeup) {
2552 event->pending_wakeup = 0;
2553 perf_event_wakeup(event);
2557 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2559 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2563 static void perf_pending_queue(struct perf_pending_entry *entry,
2564 void (*func)(struct perf_pending_entry *))
2566 struct perf_pending_entry **head;
2568 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2573 head = &get_cpu_var(perf_pending_head);
2576 entry->next = *head;
2577 } while (cmpxchg(head, entry->next, entry) != entry->next);
2579 set_perf_event_pending();
2581 put_cpu_var(perf_pending_head);
2584 static int __perf_pending_run(void)
2586 struct perf_pending_entry *list;
2589 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2590 while (list != PENDING_TAIL) {
2591 void (*func)(struct perf_pending_entry *);
2592 struct perf_pending_entry *entry = list;
2599 * Ensure we observe the unqueue before we issue the wakeup,
2600 * so that we won't be waiting forever.
2601 * -- see perf_not_pending().
2612 static inline int perf_not_pending(struct perf_event *event)
2615 * If we flush on whatever cpu we run, there is a chance we don't
2619 __perf_pending_run();
2623 * Ensure we see the proper queue state before going to sleep
2624 * so that we do not miss the wakeup. -- see perf_pending_handle()
2627 return event->pending.next == NULL;
2630 static void perf_pending_sync(struct perf_event *event)
2632 wait_event(event->waitq, perf_not_pending(event));
2635 void perf_event_do_pending(void)
2637 __perf_pending_run();
2641 * Callchain support -- arch specific
2644 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2652 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2653 unsigned long offset, unsigned long head)
2657 if (!data->writable)
2660 mask = perf_data_size(data) - 1;
2662 offset = (offset - tail) & mask;
2663 head = (head - tail) & mask;
2665 if ((int)(head - offset) < 0)
2671 static void perf_output_wakeup(struct perf_output_handle *handle)
2673 atomic_set(&handle->data->poll, POLL_IN);
2676 handle->event->pending_wakeup = 1;
2677 perf_pending_queue(&handle->event->pending,
2678 perf_pending_event);
2680 perf_event_wakeup(handle->event);
2684 * Curious locking construct.
2686 * We need to ensure a later event_id doesn't publish a head when a former
2687 * event_id isn't done writing. However since we need to deal with NMIs we
2688 * cannot fully serialize things.
2690 * What we do is serialize between CPUs so we only have to deal with NMI
2691 * nesting on a single CPU.
2693 * We only publish the head (and generate a wakeup) when the outer-most
2694 * event_id completes.
2696 static void perf_output_lock(struct perf_output_handle *handle)
2698 struct perf_mmap_data *data = handle->data;
2699 int cur, cpu = get_cpu();
2704 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2716 static void perf_output_unlock(struct perf_output_handle *handle)
2718 struct perf_mmap_data *data = handle->data;
2722 data->done_head = data->head;
2724 if (!handle->locked)
2729 * The xchg implies a full barrier that ensures all writes are done
2730 * before we publish the new head, matched by a rmb() in userspace when
2731 * reading this position.
2733 while ((head = atomic_long_xchg(&data->done_head, 0)))
2734 data->user_page->data_head = head;
2737 * NMI can happen here, which means we can miss a done_head update.
2740 cpu = atomic_xchg(&data->lock, -1);
2741 WARN_ON_ONCE(cpu != smp_processor_id());
2744 * Therefore we have to validate we did not indeed do so.
2746 if (unlikely(atomic_long_read(&data->done_head))) {
2748 * Since we had it locked, we can lock it again.
2750 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2756 if (atomic_xchg(&data->wakeup, 0))
2757 perf_output_wakeup(handle);
2762 void perf_output_copy(struct perf_output_handle *handle,
2763 const void *buf, unsigned int len)
2765 unsigned int pages_mask;
2766 unsigned long offset;
2770 offset = handle->offset;
2771 pages_mask = handle->data->nr_pages - 1;
2772 pages = handle->data->data_pages;
2775 unsigned long page_offset;
2776 unsigned long page_size;
2779 nr = (offset >> PAGE_SHIFT) & pages_mask;
2780 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2781 page_offset = offset & (page_size - 1);
2782 size = min_t(unsigned int, page_size - page_offset, len);
2784 memcpy(pages[nr] + page_offset, buf, size);
2791 handle->offset = offset;
2794 * Check we didn't copy past our reservation window, taking the
2795 * possible unsigned int wrap into account.
2797 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2800 int perf_output_begin(struct perf_output_handle *handle,
2801 struct perf_event *event, unsigned int size,
2802 int nmi, int sample)
2804 struct perf_event *output_event;
2805 struct perf_mmap_data *data;
2806 unsigned long tail, offset, head;
2809 struct perf_event_header header;
2816 * For inherited events we send all the output towards the parent.
2819 event = event->parent;
2821 output_event = rcu_dereference(event->output);
2823 event = output_event;
2825 data = rcu_dereference(event->data);
2829 handle->data = data;
2830 handle->event = event;
2832 handle->sample = sample;
2834 if (!data->nr_pages)
2837 have_lost = atomic_read(&data->lost);
2839 size += sizeof(lost_event);
2841 perf_output_lock(handle);
2845 * Userspace could choose to issue a mb() before updating the
2846 * tail pointer. So that all reads will be completed before the
2849 tail = ACCESS_ONCE(data->user_page->data_tail);
2851 offset = head = atomic_long_read(&data->head);
2853 if (unlikely(!perf_output_space(data, tail, offset, head)))
2855 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2857 handle->offset = offset;
2858 handle->head = head;
2860 if (head - tail > data->watermark)
2861 atomic_set(&data->wakeup, 1);
2864 lost_event.header.type = PERF_RECORD_LOST;
2865 lost_event.header.misc = 0;
2866 lost_event.header.size = sizeof(lost_event);
2867 lost_event.id = event->id;
2868 lost_event.lost = atomic_xchg(&data->lost, 0);
2870 perf_output_put(handle, lost_event);
2876 atomic_inc(&data->lost);
2877 perf_output_unlock(handle);
2884 void perf_output_end(struct perf_output_handle *handle)
2886 struct perf_event *event = handle->event;
2887 struct perf_mmap_data *data = handle->data;
2889 int wakeup_events = event->attr.wakeup_events;
2891 if (handle->sample && wakeup_events) {
2892 int events = atomic_inc_return(&data->events);
2893 if (events >= wakeup_events) {
2894 atomic_sub(wakeup_events, &data->events);
2895 atomic_set(&data->wakeup, 1);
2899 perf_output_unlock(handle);
2903 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2906 * only top level events have the pid namespace they were created in
2909 event = event->parent;
2911 return task_tgid_nr_ns(p, event->ns);
2914 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2917 * only top level events have the pid namespace they were created in
2920 event = event->parent;
2922 return task_pid_nr_ns(p, event->ns);
2925 static void perf_output_read_one(struct perf_output_handle *handle,
2926 struct perf_event *event)
2928 u64 read_format = event->attr.read_format;
2932 values[n++] = atomic64_read(&event->count);
2933 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2934 values[n++] = event->total_time_enabled +
2935 atomic64_read(&event->child_total_time_enabled);
2937 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2938 values[n++] = event->total_time_running +
2939 atomic64_read(&event->child_total_time_running);
2941 if (read_format & PERF_FORMAT_ID)
2942 values[n++] = primary_event_id(event);
2944 perf_output_copy(handle, values, n * sizeof(u64));
2948 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2950 static void perf_output_read_group(struct perf_output_handle *handle,
2951 struct perf_event *event)
2953 struct perf_event *leader = event->group_leader, *sub;
2954 u64 read_format = event->attr.read_format;
2958 values[n++] = 1 + leader->nr_siblings;
2960 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2961 values[n++] = leader->total_time_enabled;
2963 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2964 values[n++] = leader->total_time_running;
2966 if (leader != event)
2967 leader->pmu->read(leader);
2969 values[n++] = atomic64_read(&leader->count);
2970 if (read_format & PERF_FORMAT_ID)
2971 values[n++] = primary_event_id(leader);
2973 perf_output_copy(handle, values, n * sizeof(u64));
2975 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2979 sub->pmu->read(sub);
2981 values[n++] = atomic64_read(&sub->count);
2982 if (read_format & PERF_FORMAT_ID)
2983 values[n++] = primary_event_id(sub);
2985 perf_output_copy(handle, values, n * sizeof(u64));
2989 static void perf_output_read(struct perf_output_handle *handle,
2990 struct perf_event *event)
2992 if (event->attr.read_format & PERF_FORMAT_GROUP)
2993 perf_output_read_group(handle, event);
2995 perf_output_read_one(handle, event);
2998 void perf_output_sample(struct perf_output_handle *handle,
2999 struct perf_event_header *header,
3000 struct perf_sample_data *data,
3001 struct perf_event *event)
3003 u64 sample_type = data->type;
3005 perf_output_put(handle, *header);
3007 if (sample_type & PERF_SAMPLE_IP)
3008 perf_output_put(handle, data->ip);
3010 if (sample_type & PERF_SAMPLE_TID)
3011 perf_output_put(handle, data->tid_entry);
3013 if (sample_type & PERF_SAMPLE_TIME)
3014 perf_output_put(handle, data->time);
3016 if (sample_type & PERF_SAMPLE_ADDR)
3017 perf_output_put(handle, data->addr);
3019 if (sample_type & PERF_SAMPLE_ID)
3020 perf_output_put(handle, data->id);
3022 if (sample_type & PERF_SAMPLE_STREAM_ID)
3023 perf_output_put(handle, data->stream_id);
3025 if (sample_type & PERF_SAMPLE_CPU)
3026 perf_output_put(handle, data->cpu_entry);
3028 if (sample_type & PERF_SAMPLE_PERIOD)
3029 perf_output_put(handle, data->period);
3031 if (sample_type & PERF_SAMPLE_READ)
3032 perf_output_read(handle, event);
3034 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3035 if (data->callchain) {
3038 if (data->callchain)
3039 size += data->callchain->nr;
3041 size *= sizeof(u64);
3043 perf_output_copy(handle, data->callchain, size);
3046 perf_output_put(handle, nr);
3050 if (sample_type & PERF_SAMPLE_RAW) {
3052 perf_output_put(handle, data->raw->size);
3053 perf_output_copy(handle, data->raw->data,
3060 .size = sizeof(u32),
3063 perf_output_put(handle, raw);
3068 void perf_prepare_sample(struct perf_event_header *header,
3069 struct perf_sample_data *data,
3070 struct perf_event *event,
3071 struct pt_regs *regs)
3073 u64 sample_type = event->attr.sample_type;
3075 data->type = sample_type;
3077 header->type = PERF_RECORD_SAMPLE;
3078 header->size = sizeof(*header);
3081 header->misc |= perf_misc_flags(regs);
3083 if (sample_type & PERF_SAMPLE_IP) {
3084 data->ip = perf_instruction_pointer(regs);
3086 header->size += sizeof(data->ip);
3089 if (sample_type & PERF_SAMPLE_TID) {
3090 /* namespace issues */
3091 data->tid_entry.pid = perf_event_pid(event, current);
3092 data->tid_entry.tid = perf_event_tid(event, current);
3094 header->size += sizeof(data->tid_entry);
3097 if (sample_type & PERF_SAMPLE_TIME) {
3098 data->time = perf_clock();
3100 header->size += sizeof(data->time);
3103 if (sample_type & PERF_SAMPLE_ADDR)
3104 header->size += sizeof(data->addr);
3106 if (sample_type & PERF_SAMPLE_ID) {
3107 data->id = primary_event_id(event);
3109 header->size += sizeof(data->id);
3112 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3113 data->stream_id = event->id;
3115 header->size += sizeof(data->stream_id);
3118 if (sample_type & PERF_SAMPLE_CPU) {
3119 data->cpu_entry.cpu = raw_smp_processor_id();
3120 data->cpu_entry.reserved = 0;
3122 header->size += sizeof(data->cpu_entry);
3125 if (sample_type & PERF_SAMPLE_PERIOD)
3126 header->size += sizeof(data->period);
3128 if (sample_type & PERF_SAMPLE_READ)
3129 header->size += perf_event_read_size(event);
3131 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3134 data->callchain = perf_callchain(regs);
3136 if (data->callchain)
3137 size += data->callchain->nr;
3139 header->size += size * sizeof(u64);
3142 if (sample_type & PERF_SAMPLE_RAW) {
3143 int size = sizeof(u32);
3146 size += data->raw->size;
3148 size += sizeof(u32);
3150 WARN_ON_ONCE(size & (sizeof(u64)-1));
3151 header->size += size;
3155 static void perf_event_output(struct perf_event *event, int nmi,
3156 struct perf_sample_data *data,
3157 struct pt_regs *regs)
3159 struct perf_output_handle handle;
3160 struct perf_event_header header;
3162 perf_prepare_sample(&header, data, event, regs);
3164 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3167 perf_output_sample(&handle, &header, data, event);
3169 perf_output_end(&handle);
3176 struct perf_read_event {
3177 struct perf_event_header header;
3184 perf_event_read_event(struct perf_event *event,
3185 struct task_struct *task)
3187 struct perf_output_handle handle;
3188 struct perf_read_event read_event = {
3190 .type = PERF_RECORD_READ,
3192 .size = sizeof(read_event) + perf_event_read_size(event),
3194 .pid = perf_event_pid(event, task),
3195 .tid = perf_event_tid(event, task),
3199 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3203 perf_output_put(&handle, read_event);
3204 perf_output_read(&handle, event);
3206 perf_output_end(&handle);
3210 * task tracking -- fork/exit
3212 * enabled by: attr.comm | attr.mmap | attr.task
3215 struct perf_task_event {
3216 struct task_struct *task;
3217 struct perf_event_context *task_ctx;
3220 struct perf_event_header header;
3230 static void perf_event_task_output(struct perf_event *event,
3231 struct perf_task_event *task_event)
3233 struct perf_output_handle handle;
3235 struct task_struct *task = task_event->task;
3238 size = task_event->event_id.header.size;
3239 ret = perf_output_begin(&handle, event, size, 0, 0);
3244 task_event->event_id.pid = perf_event_pid(event, task);
3245 task_event->event_id.ppid = perf_event_pid(event, current);
3247 task_event->event_id.tid = perf_event_tid(event, task);
3248 task_event->event_id.ptid = perf_event_tid(event, current);
3250 task_event->event_id.time = perf_clock();
3252 perf_output_put(&handle, task_event->event_id);
3254 perf_output_end(&handle);
3257 static int perf_event_task_match(struct perf_event *event)
3259 if (event->attr.comm || event->attr.mmap || event->attr.task)
3265 static void perf_event_task_ctx(struct perf_event_context *ctx,
3266 struct perf_task_event *task_event)
3268 struct perf_event *event;
3270 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3271 if (perf_event_task_match(event))
3272 perf_event_task_output(event, task_event);
3276 static void perf_event_task_event(struct perf_task_event *task_event)
3278 struct perf_cpu_context *cpuctx;
3279 struct perf_event_context *ctx = task_event->task_ctx;
3282 cpuctx = &get_cpu_var(perf_cpu_context);
3283 perf_event_task_ctx(&cpuctx->ctx, task_event);
3284 put_cpu_var(perf_cpu_context);
3287 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3289 perf_event_task_ctx(ctx, task_event);
3293 static void perf_event_task(struct task_struct *task,
3294 struct perf_event_context *task_ctx,
3297 struct perf_task_event task_event;
3299 if (!atomic_read(&nr_comm_events) &&
3300 !atomic_read(&nr_mmap_events) &&
3301 !atomic_read(&nr_task_events))
3304 task_event = (struct perf_task_event){
3306 .task_ctx = task_ctx,
3309 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3311 .size = sizeof(task_event.event_id),
3320 perf_event_task_event(&task_event);
3323 void perf_event_fork(struct task_struct *task)
3325 perf_event_task(task, NULL, 1);
3332 struct perf_comm_event {
3333 struct task_struct *task;
3338 struct perf_event_header header;
3345 static void perf_event_comm_output(struct perf_event *event,
3346 struct perf_comm_event *comm_event)
3348 struct perf_output_handle handle;
3349 int size = comm_event->event_id.header.size;
3350 int ret = perf_output_begin(&handle, event, size, 0, 0);
3355 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3356 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3358 perf_output_put(&handle, comm_event->event_id);
3359 perf_output_copy(&handle, comm_event->comm,
3360 comm_event->comm_size);
3361 perf_output_end(&handle);
3364 static int perf_event_comm_match(struct perf_event *event)
3366 if (event->attr.comm)
3372 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3373 struct perf_comm_event *comm_event)
3375 struct perf_event *event;
3377 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3381 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3382 if (perf_event_comm_match(event))
3383 perf_event_comm_output(event, comm_event);
3388 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3390 struct perf_cpu_context *cpuctx;
3391 struct perf_event_context *ctx;
3393 char comm[TASK_COMM_LEN];
3395 memset(comm, 0, sizeof(comm));
3396 strncpy(comm, comm_event->task->comm, sizeof(comm));
3397 size = ALIGN(strlen(comm)+1, sizeof(u64));
3399 comm_event->comm = comm;
3400 comm_event->comm_size = size;
3402 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3404 cpuctx = &get_cpu_var(perf_cpu_context);
3405 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3406 put_cpu_var(perf_cpu_context);
3410 * doesn't really matter which of the child contexts the
3411 * events ends up in.
3413 ctx = rcu_dereference(current->perf_event_ctxp);
3415 perf_event_comm_ctx(ctx, comm_event);
3419 void perf_event_comm(struct task_struct *task)
3421 struct perf_comm_event comm_event;
3423 if (task->perf_event_ctxp)
3424 perf_event_enable_on_exec(task);
3426 if (!atomic_read(&nr_comm_events))
3429 comm_event = (struct perf_comm_event){
3435 .type = PERF_RECORD_COMM,
3444 perf_event_comm_event(&comm_event);
3451 struct perf_mmap_event {
3452 struct vm_area_struct *vma;
3454 const char *file_name;
3458 struct perf_event_header header;
3468 static void perf_event_mmap_output(struct perf_event *event,
3469 struct perf_mmap_event *mmap_event)
3471 struct perf_output_handle handle;
3472 int size = mmap_event->event_id.header.size;
3473 int ret = perf_output_begin(&handle, event, size, 0, 0);
3478 mmap_event->event_id.pid = perf_event_pid(event, current);
3479 mmap_event->event_id.tid = perf_event_tid(event, current);
3481 perf_output_put(&handle, mmap_event->event_id);
3482 perf_output_copy(&handle, mmap_event->file_name,
3483 mmap_event->file_size);
3484 perf_output_end(&handle);
3487 static int perf_event_mmap_match(struct perf_event *event,
3488 struct perf_mmap_event *mmap_event)
3490 if (event->attr.mmap)
3496 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3497 struct perf_mmap_event *mmap_event)
3499 struct perf_event *event;
3501 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3505 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3506 if (perf_event_mmap_match(event, mmap_event))
3507 perf_event_mmap_output(event, mmap_event);
3512 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3514 struct perf_cpu_context *cpuctx;
3515 struct perf_event_context *ctx;
3516 struct vm_area_struct *vma = mmap_event->vma;
3517 struct file *file = vma->vm_file;
3523 memset(tmp, 0, sizeof(tmp));
3527 * d_path works from the end of the buffer backwards, so we
3528 * need to add enough zero bytes after the string to handle
3529 * the 64bit alignment we do later.
3531 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3533 name = strncpy(tmp, "//enomem", sizeof(tmp));
3536 name = d_path(&file->f_path, buf, PATH_MAX);
3538 name = strncpy(tmp, "//toolong", sizeof(tmp));
3542 if (arch_vma_name(mmap_event->vma)) {
3543 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3549 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3553 name = strncpy(tmp, "//anon", sizeof(tmp));
3558 size = ALIGN(strlen(name)+1, sizeof(u64));
3560 mmap_event->file_name = name;
3561 mmap_event->file_size = size;
3563 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3565 cpuctx = &get_cpu_var(perf_cpu_context);
3566 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3567 put_cpu_var(perf_cpu_context);
3571 * doesn't really matter which of the child contexts the
3572 * events ends up in.
3574 ctx = rcu_dereference(current->perf_event_ctxp);
3576 perf_event_mmap_ctx(ctx, mmap_event);
3582 void __perf_event_mmap(struct vm_area_struct *vma)
3584 struct perf_mmap_event mmap_event;
3586 if (!atomic_read(&nr_mmap_events))
3589 mmap_event = (struct perf_mmap_event){
3595 .type = PERF_RECORD_MMAP,
3601 .start = vma->vm_start,
3602 .len = vma->vm_end - vma->vm_start,
3603 .pgoff = vma->vm_pgoff,
3607 perf_event_mmap_event(&mmap_event);
3611 * IRQ throttle logging
3614 static void perf_log_throttle(struct perf_event *event, int enable)
3616 struct perf_output_handle handle;
3620 struct perf_event_header header;
3624 } throttle_event = {
3626 .type = PERF_RECORD_THROTTLE,
3628 .size = sizeof(throttle_event),
3630 .time = perf_clock(),
3631 .id = primary_event_id(event),
3632 .stream_id = event->id,
3636 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3638 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3642 perf_output_put(&handle, throttle_event);
3643 perf_output_end(&handle);
3647 * Generic event overflow handling, sampling.
3650 static int __perf_event_overflow(struct perf_event *event, int nmi,
3651 int throttle, struct perf_sample_data *data,
3652 struct pt_regs *regs)
3654 int events = atomic_read(&event->event_limit);
3655 struct hw_perf_event *hwc = &event->hw;
3658 throttle = (throttle && event->pmu->unthrottle != NULL);
3663 if (hwc->interrupts != MAX_INTERRUPTS) {
3665 if (HZ * hwc->interrupts >
3666 (u64)sysctl_perf_event_sample_rate) {
3667 hwc->interrupts = MAX_INTERRUPTS;
3668 perf_log_throttle(event, 0);
3673 * Keep re-disabling events even though on the previous
3674 * pass we disabled it - just in case we raced with a
3675 * sched-in and the event got enabled again:
3681 if (event->attr.freq) {
3682 u64 now = perf_clock();
3683 s64 delta = now - hwc->freq_stamp;
3685 hwc->freq_stamp = now;
3687 if (delta > 0 && delta < TICK_NSEC)
3688 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3692 * XXX event_limit might not quite work as expected on inherited
3696 event->pending_kill = POLL_IN;
3697 if (events && atomic_dec_and_test(&event->event_limit)) {
3699 event->pending_kill = POLL_HUP;
3701 event->pending_disable = 1;
3702 perf_pending_queue(&event->pending,
3703 perf_pending_event);
3705 perf_event_disable(event);
3708 if (event->overflow_handler)
3709 event->overflow_handler(event, nmi, data, regs);
3711 perf_event_output(event, nmi, data, regs);
3716 int perf_event_overflow(struct perf_event *event, int nmi,
3717 struct perf_sample_data *data,
3718 struct pt_regs *regs)
3720 return __perf_event_overflow(event, nmi, 1, data, regs);
3724 * Generic software event infrastructure
3728 * We directly increment event->count and keep a second value in
3729 * event->hw.period_left to count intervals. This period event
3730 * is kept in the range [-sample_period, 0] so that we can use the
3734 static u64 perf_swevent_set_period(struct perf_event *event)
3736 struct hw_perf_event *hwc = &event->hw;
3737 u64 period = hwc->last_period;
3741 hwc->last_period = hwc->sample_period;
3744 old = val = atomic64_read(&hwc->period_left);
3748 nr = div64_u64(period + val, period);
3749 offset = nr * period;
3751 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3757 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3758 int nmi, struct perf_sample_data *data,
3759 struct pt_regs *regs)
3761 struct hw_perf_event *hwc = &event->hw;
3764 data->period = event->hw.last_period;
3766 overflow = perf_swevent_set_period(event);
3768 if (hwc->interrupts == MAX_INTERRUPTS)
3771 for (; overflow; overflow--) {
3772 if (__perf_event_overflow(event, nmi, throttle,
3775 * We inhibit the overflow from happening when
3776 * hwc->interrupts == MAX_INTERRUPTS.
3784 static void perf_swevent_unthrottle(struct perf_event *event)
3787 * Nothing to do, we already reset hwc->interrupts.
3791 static void perf_swevent_add(struct perf_event *event, u64 nr,
3792 int nmi, struct perf_sample_data *data,
3793 struct pt_regs *regs)
3795 struct hw_perf_event *hwc = &event->hw;
3797 atomic64_add(nr, &event->count);
3802 if (!hwc->sample_period)
3805 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3806 return perf_swevent_overflow(event, 1, nmi, data, regs);
3808 if (atomic64_add_negative(nr, &hwc->period_left))
3811 perf_swevent_overflow(event, 0, nmi, data, regs);
3814 static int perf_swevent_is_counting(struct perf_event *event)
3817 * The event is active, we're good!
3819 if (event->state == PERF_EVENT_STATE_ACTIVE)
3823 * The event is off/error, not counting.
3825 if (event->state != PERF_EVENT_STATE_INACTIVE)
3829 * The event is inactive, if the context is active
3830 * we're part of a group that didn't make it on the 'pmu',
3833 if (event->ctx->is_active)
3837 * We're inactive and the context is too, this means the
3838 * task is scheduled out, we're counting events that happen
3839 * to us, like migration events.
3844 static int perf_tp_event_match(struct perf_event *event,
3845 struct perf_sample_data *data);
3847 static int perf_swevent_match(struct perf_event *event,
3848 enum perf_type_id type,
3850 struct perf_sample_data *data,
3851 struct pt_regs *regs)
3853 if (!perf_swevent_is_counting(event))
3856 if (event->attr.type != type)
3858 if (event->attr.config != event_id)
3862 if (event->attr.exclude_user && user_mode(regs))
3865 if (event->attr.exclude_kernel && !user_mode(regs))
3869 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3870 !perf_tp_event_match(event, data))
3876 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3877 enum perf_type_id type,
3878 u32 event_id, u64 nr, int nmi,
3879 struct perf_sample_data *data,
3880 struct pt_regs *regs)
3882 struct perf_event *event;
3884 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3885 if (perf_swevent_match(event, type, event_id, data, regs))
3886 perf_swevent_add(event, nr, nmi, data, regs);
3890 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3893 return &cpuctx->recursion[3];
3896 return &cpuctx->recursion[2];
3899 return &cpuctx->recursion[1];
3901 return &cpuctx->recursion[0];
3904 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3906 struct perf_sample_data *data,
3907 struct pt_regs *regs)
3909 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3910 int *recursion = perf_swevent_recursion_context(cpuctx);
3911 struct perf_event_context *ctx;
3920 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3921 nr, nmi, data, regs);
3923 * doesn't really matter which of the child contexts the
3924 * events ends up in.
3926 ctx = rcu_dereference(current->perf_event_ctxp);
3928 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3935 put_cpu_var(perf_cpu_context);
3938 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3939 struct pt_regs *regs, u64 addr)
3941 struct perf_sample_data data = {
3945 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3949 static void perf_swevent_read(struct perf_event *event)
3953 static int perf_swevent_enable(struct perf_event *event)
3955 struct hw_perf_event *hwc = &event->hw;
3957 if (hwc->sample_period) {
3958 hwc->last_period = hwc->sample_period;
3959 perf_swevent_set_period(event);
3964 static void perf_swevent_disable(struct perf_event *event)
3968 static const struct pmu perf_ops_generic = {
3969 .enable = perf_swevent_enable,
3970 .disable = perf_swevent_disable,
3971 .read = perf_swevent_read,
3972 .unthrottle = perf_swevent_unthrottle,
3976 * hrtimer based swevent callback
3979 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3981 enum hrtimer_restart ret = HRTIMER_RESTART;
3982 struct perf_sample_data data;
3983 struct pt_regs *regs;
3984 struct perf_event *event;
3987 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
3988 event->pmu->read(event);
3991 regs = get_irq_regs();
3993 * In case we exclude kernel IPs or are somehow not in interrupt
3994 * context, provide the next best thing, the user IP.
3996 if ((event->attr.exclude_kernel || !regs) &&
3997 !event->attr.exclude_user)
3998 regs = task_pt_regs(current);
4001 if (!(event->attr.exclude_idle && current->pid == 0))
4002 if (perf_event_overflow(event, 0, &data, regs))
4003 ret = HRTIMER_NORESTART;
4006 period = max_t(u64, 10000, event->hw.sample_period);
4007 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4012 static void perf_swevent_start_hrtimer(struct perf_event *event)
4014 struct hw_perf_event *hwc = &event->hw;
4016 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4017 hwc->hrtimer.function = perf_swevent_hrtimer;
4018 if (hwc->sample_period) {
4021 if (hwc->remaining) {
4022 if (hwc->remaining < 0)
4025 period = hwc->remaining;
4028 period = max_t(u64, 10000, hwc->sample_period);
4030 __hrtimer_start_range_ns(&hwc->hrtimer,
4031 ns_to_ktime(period), 0,
4032 HRTIMER_MODE_REL, 0);
4036 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4038 struct hw_perf_event *hwc = &event->hw;
4040 if (hwc->sample_period) {
4041 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4042 hwc->remaining = ktime_to_ns(remaining);
4044 hrtimer_cancel(&hwc->hrtimer);
4049 * Software event: cpu wall time clock
4052 static void cpu_clock_perf_event_update(struct perf_event *event)
4054 int cpu = raw_smp_processor_id();
4058 now = cpu_clock(cpu);
4059 prev = atomic64_read(&event->hw.prev_count);
4060 atomic64_set(&event->hw.prev_count, now);
4061 atomic64_add(now - prev, &event->count);
4064 static int cpu_clock_perf_event_enable(struct perf_event *event)
4066 struct hw_perf_event *hwc = &event->hw;
4067 int cpu = raw_smp_processor_id();
4069 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4070 perf_swevent_start_hrtimer(event);
4075 static void cpu_clock_perf_event_disable(struct perf_event *event)
4077 perf_swevent_cancel_hrtimer(event);
4078 cpu_clock_perf_event_update(event);
4081 static void cpu_clock_perf_event_read(struct perf_event *event)
4083 cpu_clock_perf_event_update(event);
4086 static const struct pmu perf_ops_cpu_clock = {
4087 .enable = cpu_clock_perf_event_enable,
4088 .disable = cpu_clock_perf_event_disable,
4089 .read = cpu_clock_perf_event_read,
4093 * Software event: task time clock
4096 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4101 prev = atomic64_xchg(&event->hw.prev_count, now);
4103 atomic64_add(delta, &event->count);
4106 static int task_clock_perf_event_enable(struct perf_event *event)
4108 struct hw_perf_event *hwc = &event->hw;
4111 now = event->ctx->time;
4113 atomic64_set(&hwc->prev_count, now);
4115 perf_swevent_start_hrtimer(event);
4120 static void task_clock_perf_event_disable(struct perf_event *event)
4122 perf_swevent_cancel_hrtimer(event);
4123 task_clock_perf_event_update(event, event->ctx->time);
4127 static void task_clock_perf_event_read(struct perf_event *event)
4132 update_context_time(event->ctx);
4133 time = event->ctx->time;
4135 u64 now = perf_clock();
4136 u64 delta = now - event->ctx->timestamp;
4137 time = event->ctx->time + delta;
4140 task_clock_perf_event_update(event, time);
4143 static const struct pmu perf_ops_task_clock = {
4144 .enable = task_clock_perf_event_enable,
4145 .disable = task_clock_perf_event_disable,
4146 .read = task_clock_perf_event_read,
4149 #ifdef CONFIG_EVENT_PROFILE
4151 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4154 struct perf_raw_record raw = {
4159 struct perf_sample_data data = {
4164 struct pt_regs *regs = get_irq_regs();
4167 regs = task_pt_regs(current);
4169 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4172 EXPORT_SYMBOL_GPL(perf_tp_event);
4174 static int perf_tp_event_match(struct perf_event *event,
4175 struct perf_sample_data *data)
4177 void *record = data->raw->data;
4179 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4184 static void tp_perf_event_destroy(struct perf_event *event)
4186 ftrace_profile_disable(event->attr.config);
4189 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4192 * Raw tracepoint data is a severe data leak, only allow root to
4195 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4196 perf_paranoid_tracepoint_raw() &&
4197 !capable(CAP_SYS_ADMIN))
4198 return ERR_PTR(-EPERM);
4200 if (ftrace_profile_enable(event->attr.config))
4203 event->destroy = tp_perf_event_destroy;
4205 return &perf_ops_generic;
4208 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4213 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4216 filter_str = strndup_user(arg, PAGE_SIZE);
4217 if (IS_ERR(filter_str))
4218 return PTR_ERR(filter_str);
4220 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4226 static void perf_event_free_filter(struct perf_event *event)
4228 ftrace_profile_free_filter(event);
4233 static int perf_tp_event_match(struct perf_event *event,
4234 struct perf_sample_data *data)
4239 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4244 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4249 static void perf_event_free_filter(struct perf_event *event)
4253 #endif /* CONFIG_EVENT_PROFILE */
4255 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4256 static void bp_perf_event_destroy(struct perf_event *event)
4258 release_bp_slot(event);
4261 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4265 * The breakpoint is already filled if we haven't created the counter
4266 * through perf syscall
4267 * FIXME: manage to get trigerred to NULL if it comes from syscalls
4270 err = register_perf_hw_breakpoint(bp);
4272 err = __register_perf_hw_breakpoint(bp);
4274 return ERR_PTR(err);
4276 bp->destroy = bp_perf_event_destroy;
4278 return &perf_ops_bp;
4281 void perf_bp_event(struct perf_event *bp, void *regs)
4286 static void bp_perf_event_destroy(struct perf_event *event)
4290 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4295 void perf_bp_event(struct perf_event *bp, void *regs)
4300 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4302 static void sw_perf_event_destroy(struct perf_event *event)
4304 u64 event_id = event->attr.config;
4306 WARN_ON(event->parent);
4308 atomic_dec(&perf_swevent_enabled[event_id]);
4311 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4313 const struct pmu *pmu = NULL;
4314 u64 event_id = event->attr.config;
4317 * Software events (currently) can't in general distinguish
4318 * between user, kernel and hypervisor events.
4319 * However, context switches and cpu migrations are considered
4320 * to be kernel events, and page faults are never hypervisor
4324 case PERF_COUNT_SW_CPU_CLOCK:
4325 pmu = &perf_ops_cpu_clock;
4328 case PERF_COUNT_SW_TASK_CLOCK:
4330 * If the user instantiates this as a per-cpu event,
4331 * use the cpu_clock event instead.
4333 if (event->ctx->task)
4334 pmu = &perf_ops_task_clock;
4336 pmu = &perf_ops_cpu_clock;
4339 case PERF_COUNT_SW_PAGE_FAULTS:
4340 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4341 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4342 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4343 case PERF_COUNT_SW_CPU_MIGRATIONS:
4344 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4345 case PERF_COUNT_SW_EMULATION_FAULTS:
4346 if (!event->parent) {
4347 atomic_inc(&perf_swevent_enabled[event_id]);
4348 event->destroy = sw_perf_event_destroy;
4350 pmu = &perf_ops_generic;
4358 * Allocate and initialize a event structure
4360 static struct perf_event *
4361 perf_event_alloc(struct perf_event_attr *attr,
4363 struct perf_event_context *ctx,
4364 struct perf_event *group_leader,
4365 struct perf_event *parent_event,
4366 perf_callback_t callback,
4369 const struct pmu *pmu;
4370 struct perf_event *event;
4371 struct hw_perf_event *hwc;
4374 event = kzalloc(sizeof(*event), gfpflags);
4376 return ERR_PTR(-ENOMEM);
4379 * Single events are their own group leaders, with an
4380 * empty sibling list:
4383 group_leader = event;
4385 mutex_init(&event->child_mutex);
4386 INIT_LIST_HEAD(&event->child_list);
4388 INIT_LIST_HEAD(&event->group_entry);
4389 INIT_LIST_HEAD(&event->event_entry);
4390 INIT_LIST_HEAD(&event->sibling_list);
4391 init_waitqueue_head(&event->waitq);
4393 mutex_init(&event->mmap_mutex);
4396 event->attr = *attr;
4397 event->group_leader = group_leader;
4402 event->parent = parent_event;
4404 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4405 event->id = atomic64_inc_return(&perf_event_id);
4407 event->state = PERF_EVENT_STATE_INACTIVE;
4409 if (!callback && parent_event)
4410 callback = parent_event->callback;
4412 event->callback = callback;
4415 event->state = PERF_EVENT_STATE_OFF;
4420 hwc->sample_period = attr->sample_period;
4421 if (attr->freq && attr->sample_freq)
4422 hwc->sample_period = 1;
4423 hwc->last_period = hwc->sample_period;
4425 atomic64_set(&hwc->period_left, hwc->sample_period);
4428 * we currently do not support PERF_FORMAT_GROUP on inherited events
4430 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4433 switch (attr->type) {
4435 case PERF_TYPE_HARDWARE:
4436 case PERF_TYPE_HW_CACHE:
4437 pmu = hw_perf_event_init(event);
4440 case PERF_TYPE_SOFTWARE:
4441 pmu = sw_perf_event_init(event);
4444 case PERF_TYPE_TRACEPOINT:
4445 pmu = tp_perf_event_init(event);
4448 case PERF_TYPE_BREAKPOINT:
4449 pmu = bp_perf_event_init(event);
4460 else if (IS_ERR(pmu))
4465 put_pid_ns(event->ns);
4467 return ERR_PTR(err);
4472 if (!event->parent) {
4473 atomic_inc(&nr_events);
4474 if (event->attr.mmap)
4475 atomic_inc(&nr_mmap_events);
4476 if (event->attr.comm)
4477 atomic_inc(&nr_comm_events);
4478 if (event->attr.task)
4479 atomic_inc(&nr_task_events);
4485 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4486 struct perf_event_attr *attr)
4491 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4495 * zero the full structure, so that a short copy will be nice.
4497 memset(attr, 0, sizeof(*attr));
4499 ret = get_user(size, &uattr->size);
4503 if (size > PAGE_SIZE) /* silly large */
4506 if (!size) /* abi compat */
4507 size = PERF_ATTR_SIZE_VER0;
4509 if (size < PERF_ATTR_SIZE_VER0)
4513 * If we're handed a bigger struct than we know of,
4514 * ensure all the unknown bits are 0 - i.e. new
4515 * user-space does not rely on any kernel feature
4516 * extensions we dont know about yet.
4518 if (size > sizeof(*attr)) {
4519 unsigned char __user *addr;
4520 unsigned char __user *end;
4523 addr = (void __user *)uattr + sizeof(*attr);
4524 end = (void __user *)uattr + size;
4526 for (; addr < end; addr++) {
4527 ret = get_user(val, addr);
4533 size = sizeof(*attr);
4536 ret = copy_from_user(attr, uattr, size);
4541 * If the type exists, the corresponding creation will verify
4544 if (attr->type >= PERF_TYPE_MAX)
4547 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4550 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4553 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4560 put_user(sizeof(*attr), &uattr->size);
4565 static int perf_event_set_output(struct perf_event *event, int output_fd)
4567 struct perf_event *output_event = NULL;
4568 struct file *output_file = NULL;
4569 struct perf_event *old_output;
4570 int fput_needed = 0;
4576 output_file = fget_light(output_fd, &fput_needed);
4580 if (output_file->f_op != &perf_fops)
4583 output_event = output_file->private_data;
4585 /* Don't chain output fds */
4586 if (output_event->output)
4589 /* Don't set an output fd when we already have an output channel */
4593 atomic_long_inc(&output_file->f_count);
4596 mutex_lock(&event->mmap_mutex);
4597 old_output = event->output;
4598 rcu_assign_pointer(event->output, output_event);
4599 mutex_unlock(&event->mmap_mutex);
4603 * we need to make sure no existing perf_output_*()
4604 * is still referencing this event.
4607 fput(old_output->filp);
4612 fput_light(output_file, fput_needed);
4617 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4619 * @attr_uptr: event_id type attributes for monitoring/sampling
4622 * @group_fd: group leader event fd
4624 SYSCALL_DEFINE5(perf_event_open,
4625 struct perf_event_attr __user *, attr_uptr,
4626 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4628 struct perf_event *event, *group_leader;
4629 struct perf_event_attr attr;
4630 struct perf_event_context *ctx;
4631 struct file *event_file = NULL;
4632 struct file *group_file = NULL;
4633 int fput_needed = 0;
4634 int fput_needed2 = 0;
4637 /* for future expandability... */
4638 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4641 err = perf_copy_attr(attr_uptr, &attr);
4645 if (!attr.exclude_kernel) {
4646 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4651 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4656 * Get the target context (task or percpu):
4658 ctx = find_get_context(pid, cpu);
4660 return PTR_ERR(ctx);
4663 * Look up the group leader (we will attach this event to it):
4665 group_leader = NULL;
4666 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4668 group_file = fget_light(group_fd, &fput_needed);
4670 goto err_put_context;
4671 if (group_file->f_op != &perf_fops)
4672 goto err_put_context;
4674 group_leader = group_file->private_data;
4676 * Do not allow a recursive hierarchy (this new sibling
4677 * becoming part of another group-sibling):
4679 if (group_leader->group_leader != group_leader)
4680 goto err_put_context;
4682 * Do not allow to attach to a group in a different
4683 * task or CPU context:
4685 if (group_leader->ctx != ctx)
4686 goto err_put_context;
4688 * Only a group leader can be exclusive or pinned
4690 if (attr.exclusive || attr.pinned)
4691 goto err_put_context;
4694 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4695 NULL, NULL, GFP_KERNEL);
4696 err = PTR_ERR(event);
4698 goto err_put_context;
4700 err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4702 goto err_free_put_context;
4704 event_file = fget_light(err, &fput_needed2);
4706 goto err_free_put_context;
4708 if (flags & PERF_FLAG_FD_OUTPUT) {
4709 err = perf_event_set_output(event, group_fd);
4711 goto err_fput_free_put_context;
4714 event->filp = event_file;
4715 WARN_ON_ONCE(ctx->parent_ctx);
4716 mutex_lock(&ctx->mutex);
4717 perf_install_in_context(ctx, event, cpu);
4719 mutex_unlock(&ctx->mutex);
4721 event->owner = current;
4722 get_task_struct(current);
4723 mutex_lock(¤t->perf_event_mutex);
4724 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4725 mutex_unlock(¤t->perf_event_mutex);
4727 err_fput_free_put_context:
4728 fput_light(event_file, fput_needed2);
4730 err_free_put_context:
4738 fput_light(group_file, fput_needed);
4744 * perf_event_create_kernel_counter
4746 * @attr: attributes of the counter to create
4747 * @cpu: cpu in which the counter is bound
4748 * @pid: task to profile
4751 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4752 pid_t pid, perf_callback_t callback)
4754 struct perf_event *event;
4755 struct perf_event_context *ctx;
4759 * Get the target context (task or percpu):
4762 ctx = find_get_context(pid, cpu);
4766 event = perf_event_alloc(attr, cpu, ctx, NULL,
4767 NULL, callback, GFP_KERNEL);
4768 err = PTR_ERR(event);
4770 goto err_put_context;
4773 WARN_ON_ONCE(ctx->parent_ctx);
4774 mutex_lock(&ctx->mutex);
4775 perf_install_in_context(ctx, event, cpu);
4777 mutex_unlock(&ctx->mutex);
4779 event->owner = current;
4780 get_task_struct(current);
4781 mutex_lock(¤t->perf_event_mutex);
4782 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4783 mutex_unlock(¤t->perf_event_mutex);
4793 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4796 * inherit a event from parent task to child task:
4798 static struct perf_event *
4799 inherit_event(struct perf_event *parent_event,
4800 struct task_struct *parent,
4801 struct perf_event_context *parent_ctx,
4802 struct task_struct *child,
4803 struct perf_event *group_leader,
4804 struct perf_event_context *child_ctx)
4806 struct perf_event *child_event;
4809 * Instead of creating recursive hierarchies of events,
4810 * we link inherited events back to the original parent,
4811 * which has a filp for sure, which we use as the reference
4814 if (parent_event->parent)
4815 parent_event = parent_event->parent;
4817 child_event = perf_event_alloc(&parent_event->attr,
4818 parent_event->cpu, child_ctx,
4819 group_leader, parent_event,
4821 if (IS_ERR(child_event))
4826 * Make the child state follow the state of the parent event,
4827 * not its attr.disabled bit. We hold the parent's mutex,
4828 * so we won't race with perf_event_{en, dis}able_family.
4830 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4831 child_event->state = PERF_EVENT_STATE_INACTIVE;
4833 child_event->state = PERF_EVENT_STATE_OFF;
4835 if (parent_event->attr.freq)
4836 child_event->hw.sample_period = parent_event->hw.sample_period;
4838 child_event->overflow_handler = parent_event->overflow_handler;
4841 * Link it up in the child's context:
4843 add_event_to_ctx(child_event, child_ctx);
4846 * Get a reference to the parent filp - we will fput it
4847 * when the child event exits. This is safe to do because
4848 * we are in the parent and we know that the filp still
4849 * exists and has a nonzero count:
4851 atomic_long_inc(&parent_event->filp->f_count);
4854 * Link this into the parent event's child list
4856 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4857 mutex_lock(&parent_event->child_mutex);
4858 list_add_tail(&child_event->child_list, &parent_event->child_list);
4859 mutex_unlock(&parent_event->child_mutex);
4864 static int inherit_group(struct perf_event *parent_event,
4865 struct task_struct *parent,
4866 struct perf_event_context *parent_ctx,
4867 struct task_struct *child,
4868 struct perf_event_context *child_ctx)
4870 struct perf_event *leader;
4871 struct perf_event *sub;
4872 struct perf_event *child_ctr;
4874 leader = inherit_event(parent_event, parent, parent_ctx,
4875 child, NULL, child_ctx);
4877 return PTR_ERR(leader);
4878 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4879 child_ctr = inherit_event(sub, parent, parent_ctx,
4880 child, leader, child_ctx);
4881 if (IS_ERR(child_ctr))
4882 return PTR_ERR(child_ctr);
4887 static void sync_child_event(struct perf_event *child_event,
4888 struct task_struct *child)
4890 struct perf_event *parent_event = child_event->parent;
4893 if (child_event->attr.inherit_stat)
4894 perf_event_read_event(child_event, child);
4896 child_val = atomic64_read(&child_event->count);
4899 * Add back the child's count to the parent's count:
4901 atomic64_add(child_val, &parent_event->count);
4902 atomic64_add(child_event->total_time_enabled,
4903 &parent_event->child_total_time_enabled);
4904 atomic64_add(child_event->total_time_running,
4905 &parent_event->child_total_time_running);
4908 * Remove this event from the parent's list
4910 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4911 mutex_lock(&parent_event->child_mutex);
4912 list_del_init(&child_event->child_list);
4913 mutex_unlock(&parent_event->child_mutex);
4916 * Release the parent event, if this was the last
4919 fput(parent_event->filp);
4923 __perf_event_exit_task(struct perf_event *child_event,
4924 struct perf_event_context *child_ctx,
4925 struct task_struct *child)
4927 struct perf_event *parent_event;
4929 update_event_times(child_event);
4930 perf_event_remove_from_context(child_event);
4932 parent_event = child_event->parent;
4934 * It can happen that parent exits first, and has events
4935 * that are still around due to the child reference. These
4936 * events need to be zapped - but otherwise linger.
4939 sync_child_event(child_event, child);
4940 free_event(child_event);
4945 * When a child task exits, feed back event values to parent events.
4947 void perf_event_exit_task(struct task_struct *child)
4949 struct perf_event *child_event, *tmp;
4950 struct perf_event_context *child_ctx;
4951 unsigned long flags;
4953 if (likely(!child->perf_event_ctxp)) {
4954 perf_event_task(child, NULL, 0);
4958 local_irq_save(flags);
4960 * We can't reschedule here because interrupts are disabled,
4961 * and either child is current or it is a task that can't be
4962 * scheduled, so we are now safe from rescheduling changing
4965 child_ctx = child->perf_event_ctxp;
4966 __perf_event_task_sched_out(child_ctx);
4969 * Take the context lock here so that if find_get_context is
4970 * reading child->perf_event_ctxp, we wait until it has
4971 * incremented the context's refcount before we do put_ctx below.
4973 spin_lock(&child_ctx->lock);
4974 child->perf_event_ctxp = NULL;
4976 * If this context is a clone; unclone it so it can't get
4977 * swapped to another process while we're removing all
4978 * the events from it.
4980 unclone_ctx(child_ctx);
4981 spin_unlock_irqrestore(&child_ctx->lock, flags);
4984 * Report the task dead after unscheduling the events so that we
4985 * won't get any samples after PERF_RECORD_EXIT. We can however still
4986 * get a few PERF_RECORD_READ events.
4988 perf_event_task(child, child_ctx, 0);
4991 * We can recurse on the same lock type through:
4993 * __perf_event_exit_task()
4994 * sync_child_event()
4995 * fput(parent_event->filp)
4997 * mutex_lock(&ctx->mutex)
4999 * But since its the parent context it won't be the same instance.
5001 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5004 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5006 __perf_event_exit_task(child_event, child_ctx, child);
5009 * If the last event was a group event, it will have appended all
5010 * its siblings to the list, but we obtained 'tmp' before that which
5011 * will still point to the list head terminating the iteration.
5013 if (!list_empty(&child_ctx->group_list))
5016 mutex_unlock(&child_ctx->mutex);
5022 * free an unexposed, unused context as created by inheritance by
5023 * init_task below, used by fork() in case of fail.
5025 void perf_event_free_task(struct task_struct *task)
5027 struct perf_event_context *ctx = task->perf_event_ctxp;
5028 struct perf_event *event, *tmp;
5033 mutex_lock(&ctx->mutex);
5035 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5036 struct perf_event *parent = event->parent;
5038 if (WARN_ON_ONCE(!parent))
5041 mutex_lock(&parent->child_mutex);
5042 list_del_init(&event->child_list);
5043 mutex_unlock(&parent->child_mutex);
5047 list_del_event(event, ctx);
5051 if (!list_empty(&ctx->group_list))
5054 mutex_unlock(&ctx->mutex);
5060 * Initialize the perf_event context in task_struct
5062 int perf_event_init_task(struct task_struct *child)
5064 struct perf_event_context *child_ctx, *parent_ctx;
5065 struct perf_event_context *cloned_ctx;
5066 struct perf_event *event;
5067 struct task_struct *parent = current;
5068 int inherited_all = 1;
5071 child->perf_event_ctxp = NULL;
5073 mutex_init(&child->perf_event_mutex);
5074 INIT_LIST_HEAD(&child->perf_event_list);
5076 if (likely(!parent->perf_event_ctxp))
5080 * This is executed from the parent task context, so inherit
5081 * events that have been marked for cloning.
5082 * First allocate and initialize a context for the child.
5085 child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5089 __perf_event_init_context(child_ctx, child);
5090 child->perf_event_ctxp = child_ctx;
5091 get_task_struct(child);
5094 * If the parent's context is a clone, pin it so it won't get
5097 parent_ctx = perf_pin_task_context(parent);
5100 * No need to check if parent_ctx != NULL here; since we saw
5101 * it non-NULL earlier, the only reason for it to become NULL
5102 * is if we exit, and since we're currently in the middle of
5103 * a fork we can't be exiting at the same time.
5107 * Lock the parent list. No need to lock the child - not PID
5108 * hashed yet and not running, so nobody can access it.
5110 mutex_lock(&parent_ctx->mutex);
5113 * We dont have to disable NMIs - we are only looking at
5114 * the list, not manipulating it:
5116 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5118 if (!event->attr.inherit) {
5123 ret = inherit_group(event, parent, parent_ctx,
5131 if (inherited_all) {
5133 * Mark the child context as a clone of the parent
5134 * context, or of whatever the parent is a clone of.
5135 * Note that if the parent is a clone, it could get
5136 * uncloned at any point, but that doesn't matter
5137 * because the list of events and the generation
5138 * count can't have changed since we took the mutex.
5140 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5142 child_ctx->parent_ctx = cloned_ctx;
5143 child_ctx->parent_gen = parent_ctx->parent_gen;
5145 child_ctx->parent_ctx = parent_ctx;
5146 child_ctx->parent_gen = parent_ctx->generation;
5148 get_ctx(child_ctx->parent_ctx);
5151 mutex_unlock(&parent_ctx->mutex);
5153 perf_unpin_context(parent_ctx);
5158 static void __cpuinit perf_event_init_cpu(int cpu)
5160 struct perf_cpu_context *cpuctx;
5162 cpuctx = &per_cpu(perf_cpu_context, cpu);
5163 __perf_event_init_context(&cpuctx->ctx, NULL);
5165 spin_lock(&perf_resource_lock);
5166 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5167 spin_unlock(&perf_resource_lock);
5169 hw_perf_event_setup(cpu);
5172 #ifdef CONFIG_HOTPLUG_CPU
5173 static void __perf_event_exit_cpu(void *info)
5175 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5176 struct perf_event_context *ctx = &cpuctx->ctx;
5177 struct perf_event *event, *tmp;
5179 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5180 __perf_event_remove_from_context(event);
5182 static void perf_event_exit_cpu(int cpu)
5184 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5185 struct perf_event_context *ctx = &cpuctx->ctx;
5187 mutex_lock(&ctx->mutex);
5188 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5189 mutex_unlock(&ctx->mutex);
5192 static inline void perf_event_exit_cpu(int cpu) { }
5195 static int __cpuinit
5196 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5198 unsigned int cpu = (long)hcpu;
5202 case CPU_UP_PREPARE:
5203 case CPU_UP_PREPARE_FROZEN:
5204 perf_event_init_cpu(cpu);
5208 case CPU_ONLINE_FROZEN:
5209 hw_perf_event_setup_online(cpu);
5212 case CPU_DOWN_PREPARE:
5213 case CPU_DOWN_PREPARE_FROZEN:
5214 perf_event_exit_cpu(cpu);
5225 * This has to have a higher priority than migration_notifier in sched.c.
5227 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5228 .notifier_call = perf_cpu_notify,
5232 void __init perf_event_init(void)
5234 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5235 (void *)(long)smp_processor_id());
5236 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5237 (void *)(long)smp_processor_id());
5238 register_cpu_notifier(&perf_cpu_nb);
5241 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5243 return sprintf(buf, "%d\n", perf_reserved_percpu);
5247 perf_set_reserve_percpu(struct sysdev_class *class,
5251 struct perf_cpu_context *cpuctx;
5255 err = strict_strtoul(buf, 10, &val);
5258 if (val > perf_max_events)
5261 spin_lock(&perf_resource_lock);
5262 perf_reserved_percpu = val;
5263 for_each_online_cpu(cpu) {
5264 cpuctx = &per_cpu(perf_cpu_context, cpu);
5265 spin_lock_irq(&cpuctx->ctx.lock);
5266 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5267 perf_max_events - perf_reserved_percpu);
5268 cpuctx->max_pertask = mpt;
5269 spin_unlock_irq(&cpuctx->ctx.lock);
5271 spin_unlock(&perf_resource_lock);
5276 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5278 return sprintf(buf, "%d\n", perf_overcommit);
5282 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5287 err = strict_strtoul(buf, 10, &val);
5293 spin_lock(&perf_resource_lock);
5294 perf_overcommit = val;
5295 spin_unlock(&perf_resource_lock);
5300 static SYSDEV_CLASS_ATTR(
5303 perf_show_reserve_percpu,
5304 perf_set_reserve_percpu
5307 static SYSDEV_CLASS_ATTR(
5310 perf_show_overcommit,
5314 static struct attribute *perfclass_attrs[] = {
5315 &attr_reserve_percpu.attr,
5316 &attr_overcommit.attr,
5320 static struct attribute_group perfclass_attr_group = {
5321 .attrs = perfclass_attrs,
5322 .name = "perf_events",
5325 static int __init perf_event_sysfs_init(void)
5327 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5328 &perfclass_attr_group);
5330 device_initcall(perf_event_sysfs_init);