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>
33 #include <asm/irq_regs.h>
36 * Each CPU has a list of per CPU events:
38 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
40 int perf_max_events __read_mostly = 1;
41 static int perf_reserved_percpu __read_mostly;
42 static int perf_overcommit __read_mostly = 1;
44 static atomic_t nr_events __read_mostly;
45 static atomic_t nr_mmap_events __read_mostly;
46 static atomic_t nr_comm_events __read_mostly;
47 static atomic_t nr_task_events __read_mostly;
50 * perf event paranoia level:
51 * -1 - not paranoid at all
52 * 0 - disallow raw tracepoint access for unpriv
53 * 1 - disallow cpu events for unpriv
54 * 2 - disallow kernel profiling for unpriv
56 int sysctl_perf_event_paranoid __read_mostly = 1;
58 static inline bool perf_paranoid_tracepoint_raw(void)
60 return sysctl_perf_event_paranoid > -1;
63 static inline bool perf_paranoid_cpu(void)
65 return sysctl_perf_event_paranoid > 0;
68 static inline bool perf_paranoid_kernel(void)
70 return sysctl_perf_event_paranoid > 1;
73 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
76 * max perf event sample rate
78 int sysctl_perf_event_sample_rate __read_mostly = 100000;
80 static atomic64_t perf_event_id;
83 * Lock for (sysadmin-configurable) event reservations:
85 static DEFINE_SPINLOCK(perf_resource_lock);
88 * Architecture provided APIs - weak aliases:
90 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
95 void __weak hw_perf_disable(void) { barrier(); }
96 void __weak hw_perf_enable(void) { barrier(); }
98 void __weak hw_perf_event_setup(int cpu) { barrier(); }
99 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
102 hw_perf_group_sched_in(struct perf_event *group_leader,
103 struct perf_cpu_context *cpuctx,
104 struct perf_event_context *ctx, int cpu)
109 void __weak perf_event_print_debug(void) { }
111 static DEFINE_PER_CPU(int, perf_disable_count);
113 void __perf_disable(void)
115 __get_cpu_var(perf_disable_count)++;
118 bool __perf_enable(void)
120 return !--__get_cpu_var(perf_disable_count);
123 void perf_disable(void)
129 void perf_enable(void)
135 static void get_ctx(struct perf_event_context *ctx)
137 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
140 static void free_ctx(struct rcu_head *head)
142 struct perf_event_context *ctx;
144 ctx = container_of(head, struct perf_event_context, rcu_head);
148 static void put_ctx(struct perf_event_context *ctx)
150 if (atomic_dec_and_test(&ctx->refcount)) {
152 put_ctx(ctx->parent_ctx);
154 put_task_struct(ctx->task);
155 call_rcu(&ctx->rcu_head, free_ctx);
159 static void unclone_ctx(struct perf_event_context *ctx)
161 if (ctx->parent_ctx) {
162 put_ctx(ctx->parent_ctx);
163 ctx->parent_ctx = NULL;
168 * If we inherit events we want to return the parent event id
171 static u64 primary_event_id(struct perf_event *event)
176 id = event->parent->id;
182 * Get the perf_event_context for a task and lock it.
183 * This has to cope with with the fact that until it is locked,
184 * the context could get moved to another task.
186 static struct perf_event_context *
187 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 struct perf_event_context *ctx;
193 ctx = rcu_dereference(task->perf_event_ctxp);
196 * If this context is a clone of another, it might
197 * get swapped for another underneath us by
198 * perf_event_task_sched_out, though the
199 * rcu_read_lock() protects us from any context
200 * getting freed. Lock the context and check if it
201 * got swapped before we could get the lock, and retry
202 * if so. If we locked the right context, then it
203 * can't get swapped on us any more.
205 spin_lock_irqsave(&ctx->lock, *flags);
206 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
207 spin_unlock_irqrestore(&ctx->lock, *flags);
211 if (!atomic_inc_not_zero(&ctx->refcount)) {
212 spin_unlock_irqrestore(&ctx->lock, *flags);
221 * Get the context for a task and increment its pin_count so it
222 * can't get swapped to another task. This also increments its
223 * reference count so that the context can't get freed.
225 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
227 struct perf_event_context *ctx;
230 ctx = perf_lock_task_context(task, &flags);
233 spin_unlock_irqrestore(&ctx->lock, flags);
238 static void perf_unpin_context(struct perf_event_context *ctx)
242 spin_lock_irqsave(&ctx->lock, flags);
244 spin_unlock_irqrestore(&ctx->lock, flags);
249 * Add a event from the lists for its context.
250 * Must be called with ctx->mutex and ctx->lock held.
253 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
255 struct perf_event *group_leader = event->group_leader;
258 * Depending on whether it is a standalone or sibling event,
259 * add it straight to the context's event list, or to the group
260 * leader's sibling list:
262 if (group_leader == event)
263 list_add_tail(&event->group_entry, &ctx->group_list);
265 list_add_tail(&event->group_entry, &group_leader->sibling_list);
266 group_leader->nr_siblings++;
269 list_add_rcu(&event->event_entry, &ctx->event_list);
271 if (event->attr.inherit_stat)
276 * Remove a event from the lists for its context.
277 * Must be called with ctx->mutex and ctx->lock held.
280 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
282 struct perf_event *sibling, *tmp;
284 if (list_empty(&event->group_entry))
287 if (event->attr.inherit_stat)
290 list_del_init(&event->group_entry);
291 list_del_rcu(&event->event_entry);
293 if (event->group_leader != event)
294 event->group_leader->nr_siblings--;
297 * If this was a group event with sibling events then
298 * upgrade the siblings to singleton events by adding them
299 * to the context list directly:
301 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
303 list_move_tail(&sibling->group_entry, &ctx->group_list);
304 sibling->group_leader = sibling;
309 event_sched_out(struct perf_event *event,
310 struct perf_cpu_context *cpuctx,
311 struct perf_event_context *ctx)
313 if (event->state != PERF_EVENT_STATE_ACTIVE)
316 event->state = PERF_EVENT_STATE_INACTIVE;
317 if (event->pending_disable) {
318 event->pending_disable = 0;
319 event->state = PERF_EVENT_STATE_OFF;
321 event->tstamp_stopped = ctx->time;
322 event->pmu->disable(event);
325 if (!is_software_event(event))
326 cpuctx->active_oncpu--;
328 if (event->attr.exclusive || !cpuctx->active_oncpu)
329 cpuctx->exclusive = 0;
333 group_sched_out(struct perf_event *group_event,
334 struct perf_cpu_context *cpuctx,
335 struct perf_event_context *ctx)
337 struct perf_event *event;
339 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
342 event_sched_out(group_event, cpuctx, ctx);
345 * Schedule out siblings (if any):
347 list_for_each_entry(event, &group_event->sibling_list, group_entry)
348 event_sched_out(event, cpuctx, ctx);
350 if (group_event->attr.exclusive)
351 cpuctx->exclusive = 0;
355 * Cross CPU call to remove a performance event
357 * We disable the event on the hardware level first. After that we
358 * remove it from the context list.
360 static void __perf_event_remove_from_context(void *info)
362 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
363 struct perf_event *event = info;
364 struct perf_event_context *ctx = event->ctx;
367 * If this is a task context, we need to check whether it is
368 * the current task context of this cpu. If not it has been
369 * scheduled out before the smp call arrived.
371 if (ctx->task && cpuctx->task_ctx != ctx)
374 spin_lock(&ctx->lock);
376 * Protect the list operation against NMI by disabling the
377 * events on a global level.
381 event_sched_out(event, cpuctx, ctx);
383 list_del_event(event, ctx);
387 * Allow more per task events with respect to the
390 cpuctx->max_pertask =
391 min(perf_max_events - ctx->nr_events,
392 perf_max_events - perf_reserved_percpu);
396 spin_unlock(&ctx->lock);
401 * Remove the event from a task's (or a CPU's) list of events.
403 * Must be called with ctx->mutex held.
405 * CPU events are removed with a smp call. For task events we only
406 * call when the task is on a CPU.
408 * If event->ctx is a cloned context, callers must make sure that
409 * every task struct that event->ctx->task could possibly point to
410 * remains valid. This is OK when called from perf_release since
411 * that only calls us on the top-level context, which can't be a clone.
412 * When called from perf_event_exit_task, it's OK because the
413 * context has been detached from its task.
415 static void perf_event_remove_from_context(struct perf_event *event)
417 struct perf_event_context *ctx = event->ctx;
418 struct task_struct *task = ctx->task;
422 * Per cpu events are removed via an smp call and
423 * the removal is always sucessful.
425 smp_call_function_single(event->cpu,
426 __perf_event_remove_from_context,
432 task_oncpu_function_call(task, __perf_event_remove_from_context,
435 spin_lock_irq(&ctx->lock);
437 * If the context is active we need to retry the smp call.
439 if (ctx->nr_active && !list_empty(&event->group_entry)) {
440 spin_unlock_irq(&ctx->lock);
445 * The lock prevents that this context is scheduled in so we
446 * can remove the event safely, if the call above did not
449 if (!list_empty(&event->group_entry)) {
450 list_del_event(event, ctx);
452 spin_unlock_irq(&ctx->lock);
455 static inline u64 perf_clock(void)
457 return cpu_clock(smp_processor_id());
461 * Update the record of the current time in a context.
463 static void update_context_time(struct perf_event_context *ctx)
465 u64 now = perf_clock();
467 ctx->time += now - ctx->timestamp;
468 ctx->timestamp = now;
472 * Update the total_time_enabled and total_time_running fields for a event.
474 static void update_event_times(struct perf_event *event)
476 struct perf_event_context *ctx = event->ctx;
479 if (event->state < PERF_EVENT_STATE_INACTIVE ||
480 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
483 event->total_time_enabled = ctx->time - event->tstamp_enabled;
485 if (event->state == PERF_EVENT_STATE_INACTIVE)
486 run_end = event->tstamp_stopped;
490 event->total_time_running = run_end - event->tstamp_running;
494 * Update total_time_enabled and total_time_running for all events in a group.
496 static void update_group_times(struct perf_event *leader)
498 struct perf_event *event;
500 update_event_times(leader);
501 list_for_each_entry(event, &leader->sibling_list, group_entry)
502 update_event_times(event);
506 * Cross CPU call to disable a performance event
508 static void __perf_event_disable(void *info)
510 struct perf_event *event = info;
511 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
512 struct perf_event_context *ctx = event->ctx;
515 * If this is a per-task event, need to check whether this
516 * event's task is the current task on this cpu.
518 if (ctx->task && cpuctx->task_ctx != ctx)
521 spin_lock(&ctx->lock);
524 * If the event is on, turn it off.
525 * If it is in error state, leave it in error state.
527 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
528 update_context_time(ctx);
529 update_group_times(event);
530 if (event == event->group_leader)
531 group_sched_out(event, cpuctx, ctx);
533 event_sched_out(event, cpuctx, ctx);
534 event->state = PERF_EVENT_STATE_OFF;
537 spin_unlock(&ctx->lock);
543 * If event->ctx is a cloned context, callers must make sure that
544 * every task struct that event->ctx->task could possibly point to
545 * remains valid. This condition is satisifed when called through
546 * perf_event_for_each_child or perf_event_for_each because they
547 * hold the top-level event's child_mutex, so any descendant that
548 * goes to exit will block in sync_child_event.
549 * When called from perf_pending_event it's OK because event->ctx
550 * is the current context on this CPU and preemption is disabled,
551 * hence we can't get into perf_event_task_sched_out for this context.
553 static void perf_event_disable(struct perf_event *event)
555 struct perf_event_context *ctx = event->ctx;
556 struct task_struct *task = ctx->task;
560 * Disable the event on the cpu that it's on
562 smp_call_function_single(event->cpu, __perf_event_disable,
568 task_oncpu_function_call(task, __perf_event_disable, event);
570 spin_lock_irq(&ctx->lock);
572 * If the event is still active, we need to retry the cross-call.
574 if (event->state == PERF_EVENT_STATE_ACTIVE) {
575 spin_unlock_irq(&ctx->lock);
580 * Since we have the lock this context can't be scheduled
581 * in, so we can change the state safely.
583 if (event->state == PERF_EVENT_STATE_INACTIVE) {
584 update_group_times(event);
585 event->state = PERF_EVENT_STATE_OFF;
588 spin_unlock_irq(&ctx->lock);
592 event_sched_in(struct perf_event *event,
593 struct perf_cpu_context *cpuctx,
594 struct perf_event_context *ctx,
597 if (event->state <= PERF_EVENT_STATE_OFF)
600 event->state = PERF_EVENT_STATE_ACTIVE;
601 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
603 * The new state must be visible before we turn it on in the hardware:
607 if (event->pmu->enable(event)) {
608 event->state = PERF_EVENT_STATE_INACTIVE;
613 event->tstamp_running += ctx->time - event->tstamp_stopped;
615 if (!is_software_event(event))
616 cpuctx->active_oncpu++;
619 if (event->attr.exclusive)
620 cpuctx->exclusive = 1;
626 group_sched_in(struct perf_event *group_event,
627 struct perf_cpu_context *cpuctx,
628 struct perf_event_context *ctx,
631 struct perf_event *event, *partial_group;
634 if (group_event->state == PERF_EVENT_STATE_OFF)
637 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
639 return ret < 0 ? ret : 0;
641 if (event_sched_in(group_event, cpuctx, ctx, cpu))
645 * Schedule in siblings as one group (if any):
647 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
648 if (event_sched_in(event, cpuctx, ctx, cpu)) {
649 partial_group = event;
658 * Groups can be scheduled in as one unit only, so undo any
659 * partial group before returning:
661 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
662 if (event == partial_group)
664 event_sched_out(event, cpuctx, ctx);
666 event_sched_out(group_event, cpuctx, ctx);
672 * Return 1 for a group consisting entirely of software events,
673 * 0 if the group contains any hardware events.
675 static int is_software_only_group(struct perf_event *leader)
677 struct perf_event *event;
679 if (!is_software_event(leader))
682 list_for_each_entry(event, &leader->sibling_list, group_entry)
683 if (!is_software_event(event))
690 * Work out whether we can put this event group on the CPU now.
692 static int group_can_go_on(struct perf_event *event,
693 struct perf_cpu_context *cpuctx,
697 * Groups consisting entirely of software events can always go on.
699 if (is_software_only_group(event))
702 * If an exclusive group is already on, no other hardware
705 if (cpuctx->exclusive)
708 * If this group is exclusive and there are already
709 * events on the CPU, it can't go on.
711 if (event->attr.exclusive && cpuctx->active_oncpu)
714 * Otherwise, try to add it if all previous groups were able
720 static void add_event_to_ctx(struct perf_event *event,
721 struct perf_event_context *ctx)
723 list_add_event(event, ctx);
724 event->tstamp_enabled = ctx->time;
725 event->tstamp_running = ctx->time;
726 event->tstamp_stopped = ctx->time;
730 * Cross CPU call to install and enable a performance event
732 * Must be called with ctx->mutex held
734 static void __perf_install_in_context(void *info)
736 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
737 struct perf_event *event = info;
738 struct perf_event_context *ctx = event->ctx;
739 struct perf_event *leader = event->group_leader;
740 int cpu = smp_processor_id();
744 * If this is a task context, we need to check whether it is
745 * the current task context of this cpu. If not it has been
746 * scheduled out before the smp call arrived.
747 * Or possibly this is the right context but it isn't
748 * on this cpu because it had no events.
750 if (ctx->task && cpuctx->task_ctx != ctx) {
751 if (cpuctx->task_ctx || ctx->task != current)
753 cpuctx->task_ctx = ctx;
756 spin_lock(&ctx->lock);
758 update_context_time(ctx);
761 * Protect the list operation against NMI by disabling the
762 * events on a global level. NOP for non NMI based events.
766 add_event_to_ctx(event, ctx);
769 * Don't put the event on if it is disabled or if
770 * it is in a group and the group isn't on.
772 if (event->state != PERF_EVENT_STATE_INACTIVE ||
773 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
777 * An exclusive event can't go on if there are already active
778 * hardware events, and no hardware event can go on if there
779 * is already an exclusive event on.
781 if (!group_can_go_on(event, cpuctx, 1))
784 err = event_sched_in(event, cpuctx, ctx, cpu);
788 * This event couldn't go on. If it is in a group
789 * then we have to pull the whole group off.
790 * If the event group is pinned then put it in error state.
793 group_sched_out(leader, cpuctx, ctx);
794 if (leader->attr.pinned) {
795 update_group_times(leader);
796 leader->state = PERF_EVENT_STATE_ERROR;
800 if (!err && !ctx->task && cpuctx->max_pertask)
801 cpuctx->max_pertask--;
806 spin_unlock(&ctx->lock);
810 * Attach a performance event to a context
812 * First we add the event to the list with the hardware enable bit
813 * in event->hw_config cleared.
815 * If the event is attached to a task which is on a CPU we use a smp
816 * call to enable it in the task context. The task might have been
817 * scheduled away, but we check this in the smp call again.
819 * Must be called with ctx->mutex held.
822 perf_install_in_context(struct perf_event_context *ctx,
823 struct perf_event *event,
826 struct task_struct *task = ctx->task;
830 * Per cpu events are installed via an smp call and
831 * the install is always sucessful.
833 smp_call_function_single(cpu, __perf_install_in_context,
839 task_oncpu_function_call(task, __perf_install_in_context,
842 spin_lock_irq(&ctx->lock);
844 * we need to retry the smp call.
846 if (ctx->is_active && list_empty(&event->group_entry)) {
847 spin_unlock_irq(&ctx->lock);
852 * The lock prevents that this context is scheduled in so we
853 * can add the event safely, if it the call above did not
856 if (list_empty(&event->group_entry))
857 add_event_to_ctx(event, ctx);
858 spin_unlock_irq(&ctx->lock);
862 * Put a event into inactive state and update time fields.
863 * Enabling the leader of a group effectively enables all
864 * the group members that aren't explicitly disabled, so we
865 * have to update their ->tstamp_enabled also.
866 * Note: this works for group members as well as group leaders
867 * since the non-leader members' sibling_lists will be empty.
869 static void __perf_event_mark_enabled(struct perf_event *event,
870 struct perf_event_context *ctx)
872 struct perf_event *sub;
874 event->state = PERF_EVENT_STATE_INACTIVE;
875 event->tstamp_enabled = ctx->time - event->total_time_enabled;
876 list_for_each_entry(sub, &event->sibling_list, group_entry)
877 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
878 sub->tstamp_enabled =
879 ctx->time - sub->total_time_enabled;
883 * Cross CPU call to enable a performance event
885 static void __perf_event_enable(void *info)
887 struct perf_event *event = info;
888 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
889 struct perf_event_context *ctx = event->ctx;
890 struct perf_event *leader = event->group_leader;
894 * If this is a per-task event, need to check whether this
895 * event's task is the current task on this cpu.
897 if (ctx->task && cpuctx->task_ctx != ctx) {
898 if (cpuctx->task_ctx || ctx->task != current)
900 cpuctx->task_ctx = ctx;
903 spin_lock(&ctx->lock);
905 update_context_time(ctx);
907 if (event->state >= PERF_EVENT_STATE_INACTIVE)
909 __perf_event_mark_enabled(event, ctx);
912 * If the event is in a group and isn't the group leader,
913 * then don't put it on unless the group is on.
915 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
918 if (!group_can_go_on(event, cpuctx, 1)) {
923 err = group_sched_in(event, cpuctx, ctx,
926 err = event_sched_in(event, cpuctx, ctx,
933 * If this event can't go on and it's part of a
934 * group, then the whole group has to come off.
937 group_sched_out(leader, cpuctx, ctx);
938 if (leader->attr.pinned) {
939 update_group_times(leader);
940 leader->state = PERF_EVENT_STATE_ERROR;
945 spin_unlock(&ctx->lock);
951 * If event->ctx is a cloned context, callers must make sure that
952 * every task struct that event->ctx->task could possibly point to
953 * remains valid. This condition is satisfied when called through
954 * perf_event_for_each_child or perf_event_for_each as described
955 * for perf_event_disable.
957 static void perf_event_enable(struct perf_event *event)
959 struct perf_event_context *ctx = event->ctx;
960 struct task_struct *task = ctx->task;
964 * Enable the event on the cpu that it's on
966 smp_call_function_single(event->cpu, __perf_event_enable,
971 spin_lock_irq(&ctx->lock);
972 if (event->state >= PERF_EVENT_STATE_INACTIVE)
976 * If the event is in error state, clear that first.
977 * That way, if we see the event in error state below, we
978 * know that it has gone back into error state, as distinct
979 * from the task having been scheduled away before the
980 * cross-call arrived.
982 if (event->state == PERF_EVENT_STATE_ERROR)
983 event->state = PERF_EVENT_STATE_OFF;
986 spin_unlock_irq(&ctx->lock);
987 task_oncpu_function_call(task, __perf_event_enable, event);
989 spin_lock_irq(&ctx->lock);
992 * If the context is active and the event is still off,
993 * we need to retry the cross-call.
995 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
999 * Since we have the lock this context can't be scheduled
1000 * in, so we can change the state safely.
1002 if (event->state == PERF_EVENT_STATE_OFF)
1003 __perf_event_mark_enabled(event, ctx);
1006 spin_unlock_irq(&ctx->lock);
1009 static int perf_event_refresh(struct perf_event *event, int refresh)
1012 * not supported on inherited events
1014 if (event->attr.inherit)
1017 atomic_add(refresh, &event->event_limit);
1018 perf_event_enable(event);
1023 void __perf_event_sched_out(struct perf_event_context *ctx,
1024 struct perf_cpu_context *cpuctx)
1026 struct perf_event *event;
1028 spin_lock(&ctx->lock);
1030 if (likely(!ctx->nr_events))
1032 update_context_time(ctx);
1036 list_for_each_entry(event, &ctx->group_list, group_entry)
1037 group_sched_out(event, cpuctx, ctx);
1041 spin_unlock(&ctx->lock);
1045 * Test whether two contexts are equivalent, i.e. whether they
1046 * have both been cloned from the same version of the same context
1047 * and they both have the same number of enabled events.
1048 * If the number of enabled events is the same, then the set
1049 * of enabled events should be the same, because these are both
1050 * inherited contexts, therefore we can't access individual events
1051 * in them directly with an fd; we can only enable/disable all
1052 * events via prctl, or enable/disable all events in a family
1053 * via ioctl, which will have the same effect on both contexts.
1055 static int context_equiv(struct perf_event_context *ctx1,
1056 struct perf_event_context *ctx2)
1058 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1059 && ctx1->parent_gen == ctx2->parent_gen
1060 && !ctx1->pin_count && !ctx2->pin_count;
1063 static void __perf_event_read(void *event);
1065 static void __perf_event_sync_stat(struct perf_event *event,
1066 struct perf_event *next_event)
1070 if (!event->attr.inherit_stat)
1074 * Update the event value, we cannot use perf_event_read()
1075 * because we're in the middle of a context switch and have IRQs
1076 * disabled, which upsets smp_call_function_single(), however
1077 * we know the event must be on the current CPU, therefore we
1078 * don't need to use it.
1080 switch (event->state) {
1081 case PERF_EVENT_STATE_ACTIVE:
1082 __perf_event_read(event);
1085 case PERF_EVENT_STATE_INACTIVE:
1086 update_event_times(event);
1094 * In order to keep per-task stats reliable we need to flip the event
1095 * values when we flip the contexts.
1097 value = atomic64_read(&next_event->count);
1098 value = atomic64_xchg(&event->count, value);
1099 atomic64_set(&next_event->count, value);
1101 swap(event->total_time_enabled, next_event->total_time_enabled);
1102 swap(event->total_time_running, next_event->total_time_running);
1105 * Since we swizzled the values, update the user visible data too.
1107 perf_event_update_userpage(event);
1108 perf_event_update_userpage(next_event);
1111 #define list_next_entry(pos, member) \
1112 list_entry(pos->member.next, typeof(*pos), member)
1114 static void perf_event_sync_stat(struct perf_event_context *ctx,
1115 struct perf_event_context *next_ctx)
1117 struct perf_event *event, *next_event;
1122 event = list_first_entry(&ctx->event_list,
1123 struct perf_event, event_entry);
1125 next_event = list_first_entry(&next_ctx->event_list,
1126 struct perf_event, event_entry);
1128 while (&event->event_entry != &ctx->event_list &&
1129 &next_event->event_entry != &next_ctx->event_list) {
1131 __perf_event_sync_stat(event, next_event);
1133 event = list_next_entry(event, event_entry);
1134 next_event = list_next_entry(next_event, event_entry);
1139 * Called from scheduler to remove the events of the current task,
1140 * with interrupts disabled.
1142 * We stop each event and update the event value in event->count.
1144 * This does not protect us against NMI, but disable()
1145 * sets the disabled bit in the control field of event _before_
1146 * accessing the event control register. If a NMI hits, then it will
1147 * not restart the event.
1149 void perf_event_task_sched_out(struct task_struct *task,
1150 struct task_struct *next, int cpu)
1152 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1153 struct perf_event_context *ctx = task->perf_event_ctxp;
1154 struct perf_event_context *next_ctx;
1155 struct perf_event_context *parent;
1156 struct pt_regs *regs;
1159 regs = task_pt_regs(task);
1160 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1162 if (likely(!ctx || !cpuctx->task_ctx))
1165 update_context_time(ctx);
1168 parent = rcu_dereference(ctx->parent_ctx);
1169 next_ctx = next->perf_event_ctxp;
1170 if (parent && next_ctx &&
1171 rcu_dereference(next_ctx->parent_ctx) == parent) {
1173 * Looks like the two contexts are clones, so we might be
1174 * able to optimize the context switch. We lock both
1175 * contexts and check that they are clones under the
1176 * lock (including re-checking that neither has been
1177 * uncloned in the meantime). It doesn't matter which
1178 * order we take the locks because no other cpu could
1179 * be trying to lock both of these tasks.
1181 spin_lock(&ctx->lock);
1182 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1183 if (context_equiv(ctx, next_ctx)) {
1185 * XXX do we need a memory barrier of sorts
1186 * wrt to rcu_dereference() of perf_event_ctxp
1188 task->perf_event_ctxp = next_ctx;
1189 next->perf_event_ctxp = ctx;
1191 next_ctx->task = task;
1194 perf_event_sync_stat(ctx, next_ctx);
1196 spin_unlock(&next_ctx->lock);
1197 spin_unlock(&ctx->lock);
1202 __perf_event_sched_out(ctx, cpuctx);
1203 cpuctx->task_ctx = NULL;
1208 * Called with IRQs disabled
1210 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1212 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1214 if (!cpuctx->task_ctx)
1217 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1220 __perf_event_sched_out(ctx, cpuctx);
1221 cpuctx->task_ctx = NULL;
1225 * Called with IRQs disabled
1227 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1229 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1233 __perf_event_sched_in(struct perf_event_context *ctx,
1234 struct perf_cpu_context *cpuctx, int cpu)
1236 struct perf_event *event;
1239 spin_lock(&ctx->lock);
1241 if (likely(!ctx->nr_events))
1244 ctx->timestamp = perf_clock();
1249 * First go through the list and put on any pinned groups
1250 * in order to give them the best chance of going on.
1252 list_for_each_entry(event, &ctx->group_list, group_entry) {
1253 if (event->state <= PERF_EVENT_STATE_OFF ||
1254 !event->attr.pinned)
1256 if (event->cpu != -1 && event->cpu != cpu)
1259 if (group_can_go_on(event, cpuctx, 1))
1260 group_sched_in(event, cpuctx, ctx, cpu);
1263 * If this pinned group hasn't been scheduled,
1264 * put it in error state.
1266 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1267 update_group_times(event);
1268 event->state = PERF_EVENT_STATE_ERROR;
1272 list_for_each_entry(event, &ctx->group_list, group_entry) {
1274 * Ignore events in OFF or ERROR state, and
1275 * ignore pinned events since we did them already.
1277 if (event->state <= PERF_EVENT_STATE_OFF ||
1282 * Listen to the 'cpu' scheduling filter constraint
1285 if (event->cpu != -1 && event->cpu != cpu)
1288 if (group_can_go_on(event, cpuctx, can_add_hw))
1289 if (group_sched_in(event, cpuctx, ctx, cpu))
1294 spin_unlock(&ctx->lock);
1298 * Called from scheduler to add the events of the current task
1299 * with interrupts disabled.
1301 * We restore the event value and then enable it.
1303 * This does not protect us against NMI, but enable()
1304 * sets the enabled bit in the control field of event _before_
1305 * accessing the event control register. If a NMI hits, then it will
1306 * keep the event running.
1308 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1310 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1311 struct perf_event_context *ctx = task->perf_event_ctxp;
1315 if (cpuctx->task_ctx == ctx)
1317 __perf_event_sched_in(ctx, cpuctx, cpu);
1318 cpuctx->task_ctx = ctx;
1321 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1323 struct perf_event_context *ctx = &cpuctx->ctx;
1325 __perf_event_sched_in(ctx, cpuctx, cpu);
1328 #define MAX_INTERRUPTS (~0ULL)
1330 static void perf_log_throttle(struct perf_event *event, int enable);
1332 static void perf_adjust_period(struct perf_event *event, u64 events)
1334 struct hw_perf_event *hwc = &event->hw;
1335 u64 period, sample_period;
1338 events *= hwc->sample_period;
1339 period = div64_u64(events, event->attr.sample_freq);
1341 delta = (s64)(period - hwc->sample_period);
1342 delta = (delta + 7) / 8; /* low pass filter */
1344 sample_period = hwc->sample_period + delta;
1349 hwc->sample_period = sample_period;
1352 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1354 struct perf_event *event;
1355 struct hw_perf_event *hwc;
1356 u64 interrupts, freq;
1358 spin_lock(&ctx->lock);
1359 list_for_each_entry(event, &ctx->group_list, group_entry) {
1360 if (event->state != PERF_EVENT_STATE_ACTIVE)
1365 interrupts = hwc->interrupts;
1366 hwc->interrupts = 0;
1369 * unthrottle events on the tick
1371 if (interrupts == MAX_INTERRUPTS) {
1372 perf_log_throttle(event, 1);
1373 event->pmu->unthrottle(event);
1374 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1377 if (!event->attr.freq || !event->attr.sample_freq)
1381 * if the specified freq < HZ then we need to skip ticks
1383 if (event->attr.sample_freq < HZ) {
1384 freq = event->attr.sample_freq;
1386 hwc->freq_count += freq;
1387 hwc->freq_interrupts += interrupts;
1389 if (hwc->freq_count < HZ)
1392 interrupts = hwc->freq_interrupts;
1393 hwc->freq_interrupts = 0;
1394 hwc->freq_count -= HZ;
1398 perf_adjust_period(event, freq * interrupts);
1401 * In order to avoid being stalled by an (accidental) huge
1402 * sample period, force reset the sample period if we didn't
1403 * get any events in this freq period.
1407 event->pmu->disable(event);
1408 atomic64_set(&hwc->period_left, 0);
1409 event->pmu->enable(event);
1413 spin_unlock(&ctx->lock);
1417 * Round-robin a context's events:
1419 static void rotate_ctx(struct perf_event_context *ctx)
1421 struct perf_event *event;
1423 if (!ctx->nr_events)
1426 spin_lock(&ctx->lock);
1428 * Rotate the first entry last (works just fine for group events too):
1431 list_for_each_entry(event, &ctx->group_list, group_entry) {
1432 list_move_tail(&event->group_entry, &ctx->group_list);
1437 spin_unlock(&ctx->lock);
1440 void perf_event_task_tick(struct task_struct *curr, int cpu)
1442 struct perf_cpu_context *cpuctx;
1443 struct perf_event_context *ctx;
1445 if (!atomic_read(&nr_events))
1448 cpuctx = &per_cpu(perf_cpu_context, cpu);
1449 ctx = curr->perf_event_ctxp;
1451 perf_ctx_adjust_freq(&cpuctx->ctx);
1453 perf_ctx_adjust_freq(ctx);
1455 perf_event_cpu_sched_out(cpuctx);
1457 __perf_event_task_sched_out(ctx);
1459 rotate_ctx(&cpuctx->ctx);
1463 perf_event_cpu_sched_in(cpuctx, cpu);
1465 perf_event_task_sched_in(curr, cpu);
1469 * Enable all of a task's events that have been marked enable-on-exec.
1470 * This expects task == current.
1472 static void perf_event_enable_on_exec(struct task_struct *task)
1474 struct perf_event_context *ctx;
1475 struct perf_event *event;
1476 unsigned long flags;
1479 local_irq_save(flags);
1480 ctx = task->perf_event_ctxp;
1481 if (!ctx || !ctx->nr_events)
1484 __perf_event_task_sched_out(ctx);
1486 spin_lock(&ctx->lock);
1488 list_for_each_entry(event, &ctx->group_list, group_entry) {
1489 if (!event->attr.enable_on_exec)
1491 event->attr.enable_on_exec = 0;
1492 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1494 __perf_event_mark_enabled(event, ctx);
1499 * Unclone this context if we enabled any event.
1504 spin_unlock(&ctx->lock);
1506 perf_event_task_sched_in(task, smp_processor_id());
1508 local_irq_restore(flags);
1512 * Cross CPU call to read the hardware event
1514 static void __perf_event_read(void *info)
1516 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1517 struct perf_event *event = info;
1518 struct perf_event_context *ctx = event->ctx;
1519 unsigned long flags;
1522 * If this is a task context, we need to check whether it is
1523 * the current task context of this cpu. If not it has been
1524 * scheduled out before the smp call arrived. In that case
1525 * event->count would have been updated to a recent sample
1526 * when the event was scheduled out.
1528 if (ctx->task && cpuctx->task_ctx != ctx)
1531 local_irq_save(flags);
1533 update_context_time(ctx);
1534 event->pmu->read(event);
1535 update_event_times(event);
1536 local_irq_restore(flags);
1539 static u64 perf_event_read(struct perf_event *event)
1542 * If event is enabled and currently active on a CPU, update the
1543 * value in the event structure:
1545 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1546 smp_call_function_single(event->oncpu,
1547 __perf_event_read, event, 1);
1548 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1549 update_event_times(event);
1552 return atomic64_read(&event->count);
1556 * Initialize the perf_event context in a task_struct:
1559 __perf_event_init_context(struct perf_event_context *ctx,
1560 struct task_struct *task)
1562 memset(ctx, 0, sizeof(*ctx));
1563 spin_lock_init(&ctx->lock);
1564 mutex_init(&ctx->mutex);
1565 INIT_LIST_HEAD(&ctx->group_list);
1566 INIT_LIST_HEAD(&ctx->event_list);
1567 atomic_set(&ctx->refcount, 1);
1571 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1573 struct perf_event_context *ctx;
1574 struct perf_cpu_context *cpuctx;
1575 struct task_struct *task;
1576 unsigned long flags;
1580 * If cpu is not a wildcard then this is a percpu event:
1583 /* Must be root to operate on a CPU event: */
1584 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1585 return ERR_PTR(-EACCES);
1587 if (cpu < 0 || cpu > num_possible_cpus())
1588 return ERR_PTR(-EINVAL);
1591 * We could be clever and allow to attach a event to an
1592 * offline CPU and activate it when the CPU comes up, but
1595 if (!cpu_isset(cpu, cpu_online_map))
1596 return ERR_PTR(-ENODEV);
1598 cpuctx = &per_cpu(perf_cpu_context, cpu);
1609 task = find_task_by_vpid(pid);
1611 get_task_struct(task);
1615 return ERR_PTR(-ESRCH);
1618 * Can't attach events to a dying task.
1621 if (task->flags & PF_EXITING)
1624 /* Reuse ptrace permission checks for now. */
1626 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1630 ctx = perf_lock_task_context(task, &flags);
1633 spin_unlock_irqrestore(&ctx->lock, flags);
1637 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1641 __perf_event_init_context(ctx, task);
1643 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1645 * We raced with some other task; use
1646 * the context they set.
1651 get_task_struct(task);
1654 put_task_struct(task);
1658 put_task_struct(task);
1659 return ERR_PTR(err);
1662 static void perf_event_free_filter(struct perf_event *event);
1664 static void free_event_rcu(struct rcu_head *head)
1666 struct perf_event *event;
1668 event = container_of(head, struct perf_event, rcu_head);
1670 put_pid_ns(event->ns);
1671 perf_event_free_filter(event);
1675 static void perf_pending_sync(struct perf_event *event);
1677 static void free_event(struct perf_event *event)
1679 perf_pending_sync(event);
1681 if (!event->parent) {
1682 atomic_dec(&nr_events);
1683 if (event->attr.mmap)
1684 atomic_dec(&nr_mmap_events);
1685 if (event->attr.comm)
1686 atomic_dec(&nr_comm_events);
1687 if (event->attr.task)
1688 atomic_dec(&nr_task_events);
1691 if (event->output) {
1692 fput(event->output->filp);
1693 event->output = NULL;
1697 event->destroy(event);
1699 put_ctx(event->ctx);
1700 call_rcu(&event->rcu_head, free_event_rcu);
1704 * Called when the last reference to the file is gone.
1706 static int perf_release(struct inode *inode, struct file *file)
1708 struct perf_event *event = file->private_data;
1709 struct perf_event_context *ctx = event->ctx;
1711 file->private_data = NULL;
1713 WARN_ON_ONCE(ctx->parent_ctx);
1714 mutex_lock(&ctx->mutex);
1715 perf_event_remove_from_context(event);
1716 mutex_unlock(&ctx->mutex);
1718 mutex_lock(&event->owner->perf_event_mutex);
1719 list_del_init(&event->owner_entry);
1720 mutex_unlock(&event->owner->perf_event_mutex);
1721 put_task_struct(event->owner);
1728 int perf_event_release_kernel(struct perf_event *event)
1730 struct perf_event_context *ctx = event->ctx;
1732 WARN_ON_ONCE(ctx->parent_ctx);
1733 mutex_lock(&ctx->mutex);
1734 perf_event_remove_from_context(event);
1735 mutex_unlock(&ctx->mutex);
1737 mutex_lock(&event->owner->perf_event_mutex);
1738 list_del_init(&event->owner_entry);
1739 mutex_unlock(&event->owner->perf_event_mutex);
1740 put_task_struct(event->owner);
1746 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1748 static int perf_event_read_size(struct perf_event *event)
1750 int entry = sizeof(u64); /* value */
1754 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1755 size += sizeof(u64);
1757 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1758 size += sizeof(u64);
1760 if (event->attr.read_format & PERF_FORMAT_ID)
1761 entry += sizeof(u64);
1763 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1764 nr += event->group_leader->nr_siblings;
1765 size += sizeof(u64);
1773 u64 perf_event_read_value(struct perf_event *event)
1775 struct perf_event *child;
1778 total += perf_event_read(event);
1779 list_for_each_entry(child, &event->child_list, child_list)
1780 total += perf_event_read(child);
1784 EXPORT_SYMBOL_GPL(perf_event_read_value);
1786 static int perf_event_read_entry(struct perf_event *event,
1787 u64 read_format, char __user *buf)
1789 int n = 0, count = 0;
1792 values[n++] = perf_event_read_value(event);
1793 if (read_format & PERF_FORMAT_ID)
1794 values[n++] = primary_event_id(event);
1796 count = n * sizeof(u64);
1798 if (copy_to_user(buf, values, count))
1804 static int perf_event_read_group(struct perf_event *event,
1805 u64 read_format, char __user *buf)
1807 struct perf_event *leader = event->group_leader, *sub;
1808 int n = 0, size = 0, err = -EFAULT;
1811 values[n++] = 1 + leader->nr_siblings;
1812 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1813 values[n++] = leader->total_time_enabled +
1814 atomic64_read(&leader->child_total_time_enabled);
1816 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1817 values[n++] = leader->total_time_running +
1818 atomic64_read(&leader->child_total_time_running);
1821 size = n * sizeof(u64);
1823 if (copy_to_user(buf, values, size))
1826 err = perf_event_read_entry(leader, read_format, buf + size);
1832 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1833 err = perf_event_read_entry(sub, read_format,
1844 static int perf_event_read_one(struct perf_event *event,
1845 u64 read_format, char __user *buf)
1850 values[n++] = perf_event_read_value(event);
1851 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1852 values[n++] = event->total_time_enabled +
1853 atomic64_read(&event->child_total_time_enabled);
1855 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1856 values[n++] = event->total_time_running +
1857 atomic64_read(&event->child_total_time_running);
1859 if (read_format & PERF_FORMAT_ID)
1860 values[n++] = primary_event_id(event);
1862 if (copy_to_user(buf, values, n * sizeof(u64)))
1865 return n * sizeof(u64);
1869 * Read the performance event - simple non blocking version for now
1872 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1874 u64 read_format = event->attr.read_format;
1878 * Return end-of-file for a read on a event that is in
1879 * error state (i.e. because it was pinned but it couldn't be
1880 * scheduled on to the CPU at some point).
1882 if (event->state == PERF_EVENT_STATE_ERROR)
1885 if (count < perf_event_read_size(event))
1888 WARN_ON_ONCE(event->ctx->parent_ctx);
1889 mutex_lock(&event->child_mutex);
1890 if (read_format & PERF_FORMAT_GROUP)
1891 ret = perf_event_read_group(event, read_format, buf);
1893 ret = perf_event_read_one(event, read_format, buf);
1894 mutex_unlock(&event->child_mutex);
1900 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1902 struct perf_event *event = file->private_data;
1904 return perf_read_hw(event, buf, count);
1907 static unsigned int perf_poll(struct file *file, poll_table *wait)
1909 struct perf_event *event = file->private_data;
1910 struct perf_mmap_data *data;
1911 unsigned int events = POLL_HUP;
1914 data = rcu_dereference(event->data);
1916 events = atomic_xchg(&data->poll, 0);
1919 poll_wait(file, &event->waitq, wait);
1924 static void perf_event_reset(struct perf_event *event)
1926 (void)perf_event_read(event);
1927 atomic64_set(&event->count, 0);
1928 perf_event_update_userpage(event);
1932 * Holding the top-level event's child_mutex means that any
1933 * descendant process that has inherited this event will block
1934 * in sync_child_event if it goes to exit, thus satisfying the
1935 * task existence requirements of perf_event_enable/disable.
1937 static void perf_event_for_each_child(struct perf_event *event,
1938 void (*func)(struct perf_event *))
1940 struct perf_event *child;
1942 WARN_ON_ONCE(event->ctx->parent_ctx);
1943 mutex_lock(&event->child_mutex);
1945 list_for_each_entry(child, &event->child_list, child_list)
1947 mutex_unlock(&event->child_mutex);
1950 static void perf_event_for_each(struct perf_event *event,
1951 void (*func)(struct perf_event *))
1953 struct perf_event_context *ctx = event->ctx;
1954 struct perf_event *sibling;
1956 WARN_ON_ONCE(ctx->parent_ctx);
1957 mutex_lock(&ctx->mutex);
1958 event = event->group_leader;
1960 perf_event_for_each_child(event, func);
1962 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1963 perf_event_for_each_child(event, func);
1964 mutex_unlock(&ctx->mutex);
1967 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1969 struct perf_event_context *ctx = event->ctx;
1974 if (!event->attr.sample_period)
1977 size = copy_from_user(&value, arg, sizeof(value));
1978 if (size != sizeof(value))
1984 spin_lock_irq(&ctx->lock);
1985 if (event->attr.freq) {
1986 if (value > sysctl_perf_event_sample_rate) {
1991 event->attr.sample_freq = value;
1993 event->attr.sample_period = value;
1994 event->hw.sample_period = value;
1997 spin_unlock_irq(&ctx->lock);
2002 static int perf_event_set_output(struct perf_event *event, int output_fd);
2003 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2005 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2007 struct perf_event *event = file->private_data;
2008 void (*func)(struct perf_event *);
2012 case PERF_EVENT_IOC_ENABLE:
2013 func = perf_event_enable;
2015 case PERF_EVENT_IOC_DISABLE:
2016 func = perf_event_disable;
2018 case PERF_EVENT_IOC_RESET:
2019 func = perf_event_reset;
2022 case PERF_EVENT_IOC_REFRESH:
2023 return perf_event_refresh(event, arg);
2025 case PERF_EVENT_IOC_PERIOD:
2026 return perf_event_period(event, (u64 __user *)arg);
2028 case PERF_EVENT_IOC_SET_OUTPUT:
2029 return perf_event_set_output(event, arg);
2031 case PERF_EVENT_IOC_SET_FILTER:
2032 return perf_event_set_filter(event, (void __user *)arg);
2038 if (flags & PERF_IOC_FLAG_GROUP)
2039 perf_event_for_each(event, func);
2041 perf_event_for_each_child(event, func);
2046 int perf_event_task_enable(void)
2048 struct perf_event *event;
2050 mutex_lock(¤t->perf_event_mutex);
2051 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2052 perf_event_for_each_child(event, perf_event_enable);
2053 mutex_unlock(¤t->perf_event_mutex);
2058 int perf_event_task_disable(void)
2060 struct perf_event *event;
2062 mutex_lock(¤t->perf_event_mutex);
2063 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2064 perf_event_for_each_child(event, perf_event_disable);
2065 mutex_unlock(¤t->perf_event_mutex);
2070 #ifndef PERF_EVENT_INDEX_OFFSET
2071 # define PERF_EVENT_INDEX_OFFSET 0
2074 static int perf_event_index(struct perf_event *event)
2076 if (event->state != PERF_EVENT_STATE_ACTIVE)
2079 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2083 * Callers need to ensure there can be no nesting of this function, otherwise
2084 * the seqlock logic goes bad. We can not serialize this because the arch
2085 * code calls this from NMI context.
2087 void perf_event_update_userpage(struct perf_event *event)
2089 struct perf_event_mmap_page *userpg;
2090 struct perf_mmap_data *data;
2093 data = rcu_dereference(event->data);
2097 userpg = data->user_page;
2100 * Disable preemption so as to not let the corresponding user-space
2101 * spin too long if we get preempted.
2106 userpg->index = perf_event_index(event);
2107 userpg->offset = atomic64_read(&event->count);
2108 if (event->state == PERF_EVENT_STATE_ACTIVE)
2109 userpg->offset -= atomic64_read(&event->hw.prev_count);
2111 userpg->time_enabled = event->total_time_enabled +
2112 atomic64_read(&event->child_total_time_enabled);
2114 userpg->time_running = event->total_time_running +
2115 atomic64_read(&event->child_total_time_running);
2124 static unsigned long perf_data_size(struct perf_mmap_data *data)
2126 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2129 #ifndef CONFIG_PERF_USE_VMALLOC
2132 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2135 static struct page *
2136 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2138 if (pgoff > data->nr_pages)
2142 return virt_to_page(data->user_page);
2144 return virt_to_page(data->data_pages[pgoff - 1]);
2147 static struct perf_mmap_data *
2148 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2150 struct perf_mmap_data *data;
2154 WARN_ON(atomic_read(&event->mmap_count));
2156 size = sizeof(struct perf_mmap_data);
2157 size += nr_pages * sizeof(void *);
2159 data = kzalloc(size, GFP_KERNEL);
2163 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2164 if (!data->user_page)
2165 goto fail_user_page;
2167 for (i = 0; i < nr_pages; i++) {
2168 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2169 if (!data->data_pages[i])
2170 goto fail_data_pages;
2173 data->data_order = 0;
2174 data->nr_pages = nr_pages;
2179 for (i--; i >= 0; i--)
2180 free_page((unsigned long)data->data_pages[i]);
2182 free_page((unsigned long)data->user_page);
2191 static void perf_mmap_free_page(unsigned long addr)
2193 struct page *page = virt_to_page((void *)addr);
2195 page->mapping = NULL;
2199 static void perf_mmap_data_free(struct perf_mmap_data *data)
2203 perf_mmap_free_page((unsigned long)data->user_page);
2204 for (i = 0; i < data->nr_pages; i++)
2205 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2211 * Back perf_mmap() with vmalloc memory.
2213 * Required for architectures that have d-cache aliasing issues.
2216 static struct page *
2217 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2219 if (pgoff > (1UL << data->data_order))
2222 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2225 static void perf_mmap_unmark_page(void *addr)
2227 struct page *page = vmalloc_to_page(addr);
2229 page->mapping = NULL;
2232 static void perf_mmap_data_free_work(struct work_struct *work)
2234 struct perf_mmap_data *data;
2238 data = container_of(work, struct perf_mmap_data, work);
2239 nr = 1 << data->data_order;
2241 base = data->user_page;
2242 for (i = 0; i < nr + 1; i++)
2243 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2248 static void perf_mmap_data_free(struct perf_mmap_data *data)
2250 schedule_work(&data->work);
2253 static struct perf_mmap_data *
2254 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2256 struct perf_mmap_data *data;
2260 WARN_ON(atomic_read(&event->mmap_count));
2262 size = sizeof(struct perf_mmap_data);
2263 size += sizeof(void *);
2265 data = kzalloc(size, GFP_KERNEL);
2269 INIT_WORK(&data->work, perf_mmap_data_free_work);
2271 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2275 data->user_page = all_buf;
2276 data->data_pages[0] = all_buf + PAGE_SIZE;
2277 data->data_order = ilog2(nr_pages);
2291 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2293 struct perf_event *event = vma->vm_file->private_data;
2294 struct perf_mmap_data *data;
2295 int ret = VM_FAULT_SIGBUS;
2297 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2298 if (vmf->pgoff == 0)
2304 data = rcu_dereference(event->data);
2308 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2311 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2315 get_page(vmf->page);
2316 vmf->page->mapping = vma->vm_file->f_mapping;
2317 vmf->page->index = vmf->pgoff;
2327 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2329 long max_size = perf_data_size(data);
2331 atomic_set(&data->lock, -1);
2333 if (event->attr.watermark) {
2334 data->watermark = min_t(long, max_size,
2335 event->attr.wakeup_watermark);
2338 if (!data->watermark)
2339 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2342 rcu_assign_pointer(event->data, data);
2345 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2347 struct perf_mmap_data *data;
2349 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2350 perf_mmap_data_free(data);
2354 static void perf_mmap_data_release(struct perf_event *event)
2356 struct perf_mmap_data *data = event->data;
2358 WARN_ON(atomic_read(&event->mmap_count));
2360 rcu_assign_pointer(event->data, NULL);
2361 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2364 static void perf_mmap_open(struct vm_area_struct *vma)
2366 struct perf_event *event = vma->vm_file->private_data;
2368 atomic_inc(&event->mmap_count);
2371 static void perf_mmap_close(struct vm_area_struct *vma)
2373 struct perf_event *event = vma->vm_file->private_data;
2375 WARN_ON_ONCE(event->ctx->parent_ctx);
2376 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2377 unsigned long size = perf_data_size(event->data);
2378 struct user_struct *user = current_user();
2380 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2381 vma->vm_mm->locked_vm -= event->data->nr_locked;
2382 perf_mmap_data_release(event);
2383 mutex_unlock(&event->mmap_mutex);
2387 static const struct vm_operations_struct perf_mmap_vmops = {
2388 .open = perf_mmap_open,
2389 .close = perf_mmap_close,
2390 .fault = perf_mmap_fault,
2391 .page_mkwrite = perf_mmap_fault,
2394 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2396 struct perf_event *event = file->private_data;
2397 unsigned long user_locked, user_lock_limit;
2398 struct user_struct *user = current_user();
2399 unsigned long locked, lock_limit;
2400 struct perf_mmap_data *data;
2401 unsigned long vma_size;
2402 unsigned long nr_pages;
2403 long user_extra, extra;
2406 if (!(vma->vm_flags & VM_SHARED))
2409 vma_size = vma->vm_end - vma->vm_start;
2410 nr_pages = (vma_size / PAGE_SIZE) - 1;
2413 * If we have data pages ensure they're a power-of-two number, so we
2414 * can do bitmasks instead of modulo.
2416 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2419 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2422 if (vma->vm_pgoff != 0)
2425 WARN_ON_ONCE(event->ctx->parent_ctx);
2426 mutex_lock(&event->mmap_mutex);
2427 if (event->output) {
2432 if (atomic_inc_not_zero(&event->mmap_count)) {
2433 if (nr_pages != event->data->nr_pages)
2438 user_extra = nr_pages + 1;
2439 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2442 * Increase the limit linearly with more CPUs:
2444 user_lock_limit *= num_online_cpus();
2446 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2449 if (user_locked > user_lock_limit)
2450 extra = user_locked - user_lock_limit;
2452 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2453 lock_limit >>= PAGE_SHIFT;
2454 locked = vma->vm_mm->locked_vm + extra;
2456 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2457 !capable(CAP_IPC_LOCK)) {
2462 WARN_ON(event->data);
2464 data = perf_mmap_data_alloc(event, nr_pages);
2470 perf_mmap_data_init(event, data);
2472 atomic_set(&event->mmap_count, 1);
2473 atomic_long_add(user_extra, &user->locked_vm);
2474 vma->vm_mm->locked_vm += extra;
2475 event->data->nr_locked = extra;
2476 if (vma->vm_flags & VM_WRITE)
2477 event->data->writable = 1;
2480 mutex_unlock(&event->mmap_mutex);
2482 vma->vm_flags |= VM_RESERVED;
2483 vma->vm_ops = &perf_mmap_vmops;
2488 static int perf_fasync(int fd, struct file *filp, int on)
2490 struct inode *inode = filp->f_path.dentry->d_inode;
2491 struct perf_event *event = filp->private_data;
2494 mutex_lock(&inode->i_mutex);
2495 retval = fasync_helper(fd, filp, on, &event->fasync);
2496 mutex_unlock(&inode->i_mutex);
2504 static const struct file_operations perf_fops = {
2505 .release = perf_release,
2508 .unlocked_ioctl = perf_ioctl,
2509 .compat_ioctl = perf_ioctl,
2511 .fasync = perf_fasync,
2517 * If there's data, ensure we set the poll() state and publish everything
2518 * to user-space before waking everybody up.
2521 void perf_event_wakeup(struct perf_event *event)
2523 wake_up_all(&event->waitq);
2525 if (event->pending_kill) {
2526 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2527 event->pending_kill = 0;
2534 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2536 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2537 * single linked list and use cmpxchg() to add entries lockless.
2540 static void perf_pending_event(struct perf_pending_entry *entry)
2542 struct perf_event *event = container_of(entry,
2543 struct perf_event, pending);
2545 if (event->pending_disable) {
2546 event->pending_disable = 0;
2547 __perf_event_disable(event);
2550 if (event->pending_wakeup) {
2551 event->pending_wakeup = 0;
2552 perf_event_wakeup(event);
2556 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2558 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2562 static void perf_pending_queue(struct perf_pending_entry *entry,
2563 void (*func)(struct perf_pending_entry *))
2565 struct perf_pending_entry **head;
2567 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2572 head = &get_cpu_var(perf_pending_head);
2575 entry->next = *head;
2576 } while (cmpxchg(head, entry->next, entry) != entry->next);
2578 set_perf_event_pending();
2580 put_cpu_var(perf_pending_head);
2583 static int __perf_pending_run(void)
2585 struct perf_pending_entry *list;
2588 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2589 while (list != PENDING_TAIL) {
2590 void (*func)(struct perf_pending_entry *);
2591 struct perf_pending_entry *entry = list;
2598 * Ensure we observe the unqueue before we issue the wakeup,
2599 * so that we won't be waiting forever.
2600 * -- see perf_not_pending().
2611 static inline int perf_not_pending(struct perf_event *event)
2614 * If we flush on whatever cpu we run, there is a chance we don't
2618 __perf_pending_run();
2622 * Ensure we see the proper queue state before going to sleep
2623 * so that we do not miss the wakeup. -- see perf_pending_handle()
2626 return event->pending.next == NULL;
2629 static void perf_pending_sync(struct perf_event *event)
2631 wait_event(event->waitq, perf_not_pending(event));
2634 void perf_event_do_pending(void)
2636 __perf_pending_run();
2640 * Callchain support -- arch specific
2643 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2651 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2652 unsigned long offset, unsigned long head)
2656 if (!data->writable)
2659 mask = perf_data_size(data) - 1;
2661 offset = (offset - tail) & mask;
2662 head = (head - tail) & mask;
2664 if ((int)(head - offset) < 0)
2670 static void perf_output_wakeup(struct perf_output_handle *handle)
2672 atomic_set(&handle->data->poll, POLL_IN);
2675 handle->event->pending_wakeup = 1;
2676 perf_pending_queue(&handle->event->pending,
2677 perf_pending_event);
2679 perf_event_wakeup(handle->event);
2683 * Curious locking construct.
2685 * We need to ensure a later event_id doesn't publish a head when a former
2686 * event_id isn't done writing. However since we need to deal with NMIs we
2687 * cannot fully serialize things.
2689 * What we do is serialize between CPUs so we only have to deal with NMI
2690 * nesting on a single CPU.
2692 * We only publish the head (and generate a wakeup) when the outer-most
2693 * event_id completes.
2695 static void perf_output_lock(struct perf_output_handle *handle)
2697 struct perf_mmap_data *data = handle->data;
2702 local_irq_save(handle->flags);
2703 cpu = smp_processor_id();
2705 if (in_nmi() && atomic_read(&data->lock) == cpu)
2708 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2714 static void perf_output_unlock(struct perf_output_handle *handle)
2716 struct perf_mmap_data *data = handle->data;
2720 data->done_head = data->head;
2722 if (!handle->locked)
2727 * The xchg implies a full barrier that ensures all writes are done
2728 * before we publish the new head, matched by a rmb() in userspace when
2729 * reading this position.
2731 while ((head = atomic_long_xchg(&data->done_head, 0)))
2732 data->user_page->data_head = head;
2735 * NMI can happen here, which means we can miss a done_head update.
2738 cpu = atomic_xchg(&data->lock, -1);
2739 WARN_ON_ONCE(cpu != smp_processor_id());
2742 * Therefore we have to validate we did not indeed do so.
2744 if (unlikely(atomic_long_read(&data->done_head))) {
2746 * Since we had it locked, we can lock it again.
2748 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2754 if (atomic_xchg(&data->wakeup, 0))
2755 perf_output_wakeup(handle);
2757 local_irq_restore(handle->flags);
2760 void perf_output_copy(struct perf_output_handle *handle,
2761 const void *buf, unsigned int len)
2763 unsigned int pages_mask;
2764 unsigned long offset;
2768 offset = handle->offset;
2769 pages_mask = handle->data->nr_pages - 1;
2770 pages = handle->data->data_pages;
2773 unsigned long page_offset;
2774 unsigned long page_size;
2777 nr = (offset >> PAGE_SHIFT) & pages_mask;
2778 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2779 page_offset = offset & (page_size - 1);
2780 size = min_t(unsigned int, page_size - page_offset, len);
2782 memcpy(pages[nr] + page_offset, buf, size);
2789 handle->offset = offset;
2792 * Check we didn't copy past our reservation window, taking the
2793 * possible unsigned int wrap into account.
2795 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2798 int perf_output_begin(struct perf_output_handle *handle,
2799 struct perf_event *event, unsigned int size,
2800 int nmi, int sample)
2802 struct perf_event *output_event;
2803 struct perf_mmap_data *data;
2804 unsigned long tail, offset, head;
2807 struct perf_event_header header;
2814 * For inherited events we send all the output towards the parent.
2817 event = event->parent;
2819 output_event = rcu_dereference(event->output);
2821 event = output_event;
2823 data = rcu_dereference(event->data);
2827 handle->data = data;
2828 handle->event = event;
2830 handle->sample = sample;
2832 if (!data->nr_pages)
2835 have_lost = atomic_read(&data->lost);
2837 size += sizeof(lost_event);
2839 perf_output_lock(handle);
2843 * Userspace could choose to issue a mb() before updating the
2844 * tail pointer. So that all reads will be completed before the
2847 tail = ACCESS_ONCE(data->user_page->data_tail);
2849 offset = head = atomic_long_read(&data->head);
2851 if (unlikely(!perf_output_space(data, tail, offset, head)))
2853 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2855 handle->offset = offset;
2856 handle->head = head;
2858 if (head - tail > data->watermark)
2859 atomic_set(&data->wakeup, 1);
2862 lost_event.header.type = PERF_RECORD_LOST;
2863 lost_event.header.misc = 0;
2864 lost_event.header.size = sizeof(lost_event);
2865 lost_event.id = event->id;
2866 lost_event.lost = atomic_xchg(&data->lost, 0);
2868 perf_output_put(handle, lost_event);
2874 atomic_inc(&data->lost);
2875 perf_output_unlock(handle);
2882 void perf_output_end(struct perf_output_handle *handle)
2884 struct perf_event *event = handle->event;
2885 struct perf_mmap_data *data = handle->data;
2887 int wakeup_events = event->attr.wakeup_events;
2889 if (handle->sample && wakeup_events) {
2890 int events = atomic_inc_return(&data->events);
2891 if (events >= wakeup_events) {
2892 atomic_sub(wakeup_events, &data->events);
2893 atomic_set(&data->wakeup, 1);
2897 perf_output_unlock(handle);
2901 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2904 * only top level events have the pid namespace they were created in
2907 event = event->parent;
2909 return task_tgid_nr_ns(p, event->ns);
2912 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2915 * only top level events have the pid namespace they were created in
2918 event = event->parent;
2920 return task_pid_nr_ns(p, event->ns);
2923 static void perf_output_read_one(struct perf_output_handle *handle,
2924 struct perf_event *event)
2926 u64 read_format = event->attr.read_format;
2930 values[n++] = atomic64_read(&event->count);
2931 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2932 values[n++] = event->total_time_enabled +
2933 atomic64_read(&event->child_total_time_enabled);
2935 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2936 values[n++] = event->total_time_running +
2937 atomic64_read(&event->child_total_time_running);
2939 if (read_format & PERF_FORMAT_ID)
2940 values[n++] = primary_event_id(event);
2942 perf_output_copy(handle, values, n * sizeof(u64));
2946 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2948 static void perf_output_read_group(struct perf_output_handle *handle,
2949 struct perf_event *event)
2951 struct perf_event *leader = event->group_leader, *sub;
2952 u64 read_format = event->attr.read_format;
2956 values[n++] = 1 + leader->nr_siblings;
2958 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2959 values[n++] = leader->total_time_enabled;
2961 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2962 values[n++] = leader->total_time_running;
2964 if (leader != event)
2965 leader->pmu->read(leader);
2967 values[n++] = atomic64_read(&leader->count);
2968 if (read_format & PERF_FORMAT_ID)
2969 values[n++] = primary_event_id(leader);
2971 perf_output_copy(handle, values, n * sizeof(u64));
2973 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2977 sub->pmu->read(sub);
2979 values[n++] = atomic64_read(&sub->count);
2980 if (read_format & PERF_FORMAT_ID)
2981 values[n++] = primary_event_id(sub);
2983 perf_output_copy(handle, values, n * sizeof(u64));
2987 static void perf_output_read(struct perf_output_handle *handle,
2988 struct perf_event *event)
2990 if (event->attr.read_format & PERF_FORMAT_GROUP)
2991 perf_output_read_group(handle, event);
2993 perf_output_read_one(handle, event);
2996 void perf_output_sample(struct perf_output_handle *handle,
2997 struct perf_event_header *header,
2998 struct perf_sample_data *data,
2999 struct perf_event *event)
3001 u64 sample_type = data->type;
3003 perf_output_put(handle, *header);
3005 if (sample_type & PERF_SAMPLE_IP)
3006 perf_output_put(handle, data->ip);
3008 if (sample_type & PERF_SAMPLE_TID)
3009 perf_output_put(handle, data->tid_entry);
3011 if (sample_type & PERF_SAMPLE_TIME)
3012 perf_output_put(handle, data->time);
3014 if (sample_type & PERF_SAMPLE_ADDR)
3015 perf_output_put(handle, data->addr);
3017 if (sample_type & PERF_SAMPLE_ID)
3018 perf_output_put(handle, data->id);
3020 if (sample_type & PERF_SAMPLE_STREAM_ID)
3021 perf_output_put(handle, data->stream_id);
3023 if (sample_type & PERF_SAMPLE_CPU)
3024 perf_output_put(handle, data->cpu_entry);
3026 if (sample_type & PERF_SAMPLE_PERIOD)
3027 perf_output_put(handle, data->period);
3029 if (sample_type & PERF_SAMPLE_READ)
3030 perf_output_read(handle, event);
3032 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3033 if (data->callchain) {
3036 if (data->callchain)
3037 size += data->callchain->nr;
3039 size *= sizeof(u64);
3041 perf_output_copy(handle, data->callchain, size);
3044 perf_output_put(handle, nr);
3048 if (sample_type & PERF_SAMPLE_RAW) {
3050 perf_output_put(handle, data->raw->size);
3051 perf_output_copy(handle, data->raw->data,
3058 .size = sizeof(u32),
3061 perf_output_put(handle, raw);
3066 void perf_prepare_sample(struct perf_event_header *header,
3067 struct perf_sample_data *data,
3068 struct perf_event *event,
3069 struct pt_regs *regs)
3071 u64 sample_type = event->attr.sample_type;
3073 data->type = sample_type;
3075 header->type = PERF_RECORD_SAMPLE;
3076 header->size = sizeof(*header);
3079 header->misc |= perf_misc_flags(regs);
3081 if (sample_type & PERF_SAMPLE_IP) {
3082 data->ip = perf_instruction_pointer(regs);
3084 header->size += sizeof(data->ip);
3087 if (sample_type & PERF_SAMPLE_TID) {
3088 /* namespace issues */
3089 data->tid_entry.pid = perf_event_pid(event, current);
3090 data->tid_entry.tid = perf_event_tid(event, current);
3092 header->size += sizeof(data->tid_entry);
3095 if (sample_type & PERF_SAMPLE_TIME) {
3096 data->time = perf_clock();
3098 header->size += sizeof(data->time);
3101 if (sample_type & PERF_SAMPLE_ADDR)
3102 header->size += sizeof(data->addr);
3104 if (sample_type & PERF_SAMPLE_ID) {
3105 data->id = primary_event_id(event);
3107 header->size += sizeof(data->id);
3110 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3111 data->stream_id = event->id;
3113 header->size += sizeof(data->stream_id);
3116 if (sample_type & PERF_SAMPLE_CPU) {
3117 data->cpu_entry.cpu = raw_smp_processor_id();
3118 data->cpu_entry.reserved = 0;
3120 header->size += sizeof(data->cpu_entry);
3123 if (sample_type & PERF_SAMPLE_PERIOD)
3124 header->size += sizeof(data->period);
3126 if (sample_type & PERF_SAMPLE_READ)
3127 header->size += perf_event_read_size(event);
3129 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3132 data->callchain = perf_callchain(regs);
3134 if (data->callchain)
3135 size += data->callchain->nr;
3137 header->size += size * sizeof(u64);
3140 if (sample_type & PERF_SAMPLE_RAW) {
3141 int size = sizeof(u32);
3144 size += data->raw->size;
3146 size += sizeof(u32);
3148 WARN_ON_ONCE(size & (sizeof(u64)-1));
3149 header->size += size;
3153 static void perf_event_output(struct perf_event *event, int nmi,
3154 struct perf_sample_data *data,
3155 struct pt_regs *regs)
3157 struct perf_output_handle handle;
3158 struct perf_event_header header;
3160 perf_prepare_sample(&header, data, event, regs);
3162 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3165 perf_output_sample(&handle, &header, data, event);
3167 perf_output_end(&handle);
3174 struct perf_read_event {
3175 struct perf_event_header header;
3182 perf_event_read_event(struct perf_event *event,
3183 struct task_struct *task)
3185 struct perf_output_handle handle;
3186 struct perf_read_event read_event = {
3188 .type = PERF_RECORD_READ,
3190 .size = sizeof(read_event) + perf_event_read_size(event),
3192 .pid = perf_event_pid(event, task),
3193 .tid = perf_event_tid(event, task),
3197 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3201 perf_output_put(&handle, read_event);
3202 perf_output_read(&handle, event);
3204 perf_output_end(&handle);
3208 * task tracking -- fork/exit
3210 * enabled by: attr.comm | attr.mmap | attr.task
3213 struct perf_task_event {
3214 struct task_struct *task;
3215 struct perf_event_context *task_ctx;
3218 struct perf_event_header header;
3228 static void perf_event_task_output(struct perf_event *event,
3229 struct perf_task_event *task_event)
3231 struct perf_output_handle handle;
3233 struct task_struct *task = task_event->task;
3236 size = task_event->event_id.header.size;
3237 ret = perf_output_begin(&handle, event, size, 0, 0);
3242 task_event->event_id.pid = perf_event_pid(event, task);
3243 task_event->event_id.ppid = perf_event_pid(event, current);
3245 task_event->event_id.tid = perf_event_tid(event, task);
3246 task_event->event_id.ptid = perf_event_tid(event, current);
3248 task_event->event_id.time = perf_clock();
3250 perf_output_put(&handle, task_event->event_id);
3252 perf_output_end(&handle);
3255 static int perf_event_task_match(struct perf_event *event)
3257 if (event->attr.comm || event->attr.mmap || event->attr.task)
3263 static void perf_event_task_ctx(struct perf_event_context *ctx,
3264 struct perf_task_event *task_event)
3266 struct perf_event *event;
3268 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3272 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3273 if (perf_event_task_match(event))
3274 perf_event_task_output(event, task_event);
3279 static void perf_event_task_event(struct perf_task_event *task_event)
3281 struct perf_cpu_context *cpuctx;
3282 struct perf_event_context *ctx = task_event->task_ctx;
3284 cpuctx = &get_cpu_var(perf_cpu_context);
3285 perf_event_task_ctx(&cpuctx->ctx, task_event);
3286 put_cpu_var(perf_cpu_context);
3290 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3292 perf_event_task_ctx(ctx, task_event);
3296 static void perf_event_task(struct task_struct *task,
3297 struct perf_event_context *task_ctx,
3300 struct perf_task_event task_event;
3302 if (!atomic_read(&nr_comm_events) &&
3303 !atomic_read(&nr_mmap_events) &&
3304 !atomic_read(&nr_task_events))
3307 task_event = (struct perf_task_event){
3309 .task_ctx = task_ctx,
3312 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3314 .size = sizeof(task_event.event_id),
3323 perf_event_task_event(&task_event);
3326 void perf_event_fork(struct task_struct *task)
3328 perf_event_task(task, NULL, 1);
3335 struct perf_comm_event {
3336 struct task_struct *task;
3341 struct perf_event_header header;
3348 static void perf_event_comm_output(struct perf_event *event,
3349 struct perf_comm_event *comm_event)
3351 struct perf_output_handle handle;
3352 int size = comm_event->event_id.header.size;
3353 int ret = perf_output_begin(&handle, event, size, 0, 0);
3358 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3359 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3361 perf_output_put(&handle, comm_event->event_id);
3362 perf_output_copy(&handle, comm_event->comm,
3363 comm_event->comm_size);
3364 perf_output_end(&handle);
3367 static int perf_event_comm_match(struct perf_event *event)
3369 if (event->attr.comm)
3375 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3376 struct perf_comm_event *comm_event)
3378 struct perf_event *event;
3380 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3384 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3385 if (perf_event_comm_match(event))
3386 perf_event_comm_output(event, comm_event);
3391 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3393 struct perf_cpu_context *cpuctx;
3394 struct perf_event_context *ctx;
3396 char comm[TASK_COMM_LEN];
3398 memset(comm, 0, sizeof(comm));
3399 strncpy(comm, comm_event->task->comm, sizeof(comm));
3400 size = ALIGN(strlen(comm)+1, sizeof(u64));
3402 comm_event->comm = comm;
3403 comm_event->comm_size = size;
3405 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3407 cpuctx = &get_cpu_var(perf_cpu_context);
3408 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3409 put_cpu_var(perf_cpu_context);
3413 * doesn't really matter which of the child contexts the
3414 * events ends up in.
3416 ctx = rcu_dereference(current->perf_event_ctxp);
3418 perf_event_comm_ctx(ctx, comm_event);
3422 void perf_event_comm(struct task_struct *task)
3424 struct perf_comm_event comm_event;
3426 if (task->perf_event_ctxp)
3427 perf_event_enable_on_exec(task);
3429 if (!atomic_read(&nr_comm_events))
3432 comm_event = (struct perf_comm_event){
3438 .type = PERF_RECORD_COMM,
3447 perf_event_comm_event(&comm_event);
3454 struct perf_mmap_event {
3455 struct vm_area_struct *vma;
3457 const char *file_name;
3461 struct perf_event_header header;
3471 static void perf_event_mmap_output(struct perf_event *event,
3472 struct perf_mmap_event *mmap_event)
3474 struct perf_output_handle handle;
3475 int size = mmap_event->event_id.header.size;
3476 int ret = perf_output_begin(&handle, event, size, 0, 0);
3481 mmap_event->event_id.pid = perf_event_pid(event, current);
3482 mmap_event->event_id.tid = perf_event_tid(event, current);
3484 perf_output_put(&handle, mmap_event->event_id);
3485 perf_output_copy(&handle, mmap_event->file_name,
3486 mmap_event->file_size);
3487 perf_output_end(&handle);
3490 static int perf_event_mmap_match(struct perf_event *event,
3491 struct perf_mmap_event *mmap_event)
3493 if (event->attr.mmap)
3499 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3500 struct perf_mmap_event *mmap_event)
3502 struct perf_event *event;
3504 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3508 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3509 if (perf_event_mmap_match(event, mmap_event))
3510 perf_event_mmap_output(event, mmap_event);
3515 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3517 struct perf_cpu_context *cpuctx;
3518 struct perf_event_context *ctx;
3519 struct vm_area_struct *vma = mmap_event->vma;
3520 struct file *file = vma->vm_file;
3526 memset(tmp, 0, sizeof(tmp));
3530 * d_path works from the end of the buffer backwards, so we
3531 * need to add enough zero bytes after the string to handle
3532 * the 64bit alignment we do later.
3534 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3536 name = strncpy(tmp, "//enomem", sizeof(tmp));
3539 name = d_path(&file->f_path, buf, PATH_MAX);
3541 name = strncpy(tmp, "//toolong", sizeof(tmp));
3545 if (arch_vma_name(mmap_event->vma)) {
3546 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3552 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3556 name = strncpy(tmp, "//anon", sizeof(tmp));
3561 size = ALIGN(strlen(name)+1, sizeof(u64));
3563 mmap_event->file_name = name;
3564 mmap_event->file_size = size;
3566 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3568 cpuctx = &get_cpu_var(perf_cpu_context);
3569 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3570 put_cpu_var(perf_cpu_context);
3574 * doesn't really matter which of the child contexts the
3575 * events ends up in.
3577 ctx = rcu_dereference(current->perf_event_ctxp);
3579 perf_event_mmap_ctx(ctx, mmap_event);
3585 void __perf_event_mmap(struct vm_area_struct *vma)
3587 struct perf_mmap_event mmap_event;
3589 if (!atomic_read(&nr_mmap_events))
3592 mmap_event = (struct perf_mmap_event){
3598 .type = PERF_RECORD_MMAP,
3604 .start = vma->vm_start,
3605 .len = vma->vm_end - vma->vm_start,
3606 .pgoff = vma->vm_pgoff,
3610 perf_event_mmap_event(&mmap_event);
3614 * IRQ throttle logging
3617 static void perf_log_throttle(struct perf_event *event, int enable)
3619 struct perf_output_handle handle;
3623 struct perf_event_header header;
3627 } throttle_event = {
3629 .type = PERF_RECORD_THROTTLE,
3631 .size = sizeof(throttle_event),
3633 .time = perf_clock(),
3634 .id = primary_event_id(event),
3635 .stream_id = event->id,
3639 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3641 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3645 perf_output_put(&handle, throttle_event);
3646 perf_output_end(&handle);
3650 * Generic event overflow handling, sampling.
3653 static int __perf_event_overflow(struct perf_event *event, int nmi,
3654 int throttle, struct perf_sample_data *data,
3655 struct pt_regs *regs)
3657 int events = atomic_read(&event->event_limit);
3658 struct hw_perf_event *hwc = &event->hw;
3661 throttle = (throttle && event->pmu->unthrottle != NULL);
3666 if (hwc->interrupts != MAX_INTERRUPTS) {
3668 if (HZ * hwc->interrupts >
3669 (u64)sysctl_perf_event_sample_rate) {
3670 hwc->interrupts = MAX_INTERRUPTS;
3671 perf_log_throttle(event, 0);
3676 * Keep re-disabling events even though on the previous
3677 * pass we disabled it - just in case we raced with a
3678 * sched-in and the event got enabled again:
3684 if (event->attr.freq) {
3685 u64 now = perf_clock();
3686 s64 delta = now - hwc->freq_stamp;
3688 hwc->freq_stamp = now;
3690 if (delta > 0 && delta < TICK_NSEC)
3691 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3695 * XXX event_limit might not quite work as expected on inherited
3699 event->pending_kill = POLL_IN;
3700 if (events && atomic_dec_and_test(&event->event_limit)) {
3702 event->pending_kill = POLL_HUP;
3704 event->pending_disable = 1;
3705 perf_pending_queue(&event->pending,
3706 perf_pending_event);
3708 perf_event_disable(event);
3711 perf_event_output(event, nmi, data, regs);
3715 int perf_event_overflow(struct perf_event *event, int nmi,
3716 struct perf_sample_data *data,
3717 struct pt_regs *regs)
3719 return __perf_event_overflow(event, nmi, 1, data, regs);
3723 * Generic software event infrastructure
3727 * We directly increment event->count and keep a second value in
3728 * event->hw.period_left to count intervals. This period event
3729 * is kept in the range [-sample_period, 0] so that we can use the
3733 static u64 perf_swevent_set_period(struct perf_event *event)
3735 struct hw_perf_event *hwc = &event->hw;
3736 u64 period = hwc->last_period;
3740 hwc->last_period = hwc->sample_period;
3743 old = val = atomic64_read(&hwc->period_left);
3747 nr = div64_u64(period + val, period);
3748 offset = nr * period;
3750 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3756 static void perf_swevent_overflow(struct perf_event *event,
3757 int nmi, struct perf_sample_data *data,
3758 struct pt_regs *regs)
3760 struct hw_perf_event *hwc = &event->hw;
3764 data->period = event->hw.last_period;
3765 overflow = perf_swevent_set_period(event);
3767 if (hwc->interrupts == MAX_INTERRUPTS)
3770 for (; overflow; overflow--) {
3771 if (__perf_event_overflow(event, nmi, throttle,
3774 * We inhibit the overflow from happening when
3775 * hwc->interrupts == MAX_INTERRUPTS.
3783 static void perf_swevent_unthrottle(struct perf_event *event)
3786 * Nothing to do, we already reset hwc->interrupts.
3790 static void perf_swevent_add(struct perf_event *event, u64 nr,
3791 int nmi, struct perf_sample_data *data,
3792 struct pt_regs *regs)
3794 struct hw_perf_event *hwc = &event->hw;
3796 atomic64_add(nr, &event->count);
3798 if (!hwc->sample_period)
3804 if (!atomic64_add_negative(nr, &hwc->period_left))
3805 perf_swevent_overflow(event, nmi, data, regs);
3808 static int perf_swevent_is_counting(struct perf_event *event)
3811 * The event is active, we're good!
3813 if (event->state == PERF_EVENT_STATE_ACTIVE)
3817 * The event is off/error, not counting.
3819 if (event->state != PERF_EVENT_STATE_INACTIVE)
3823 * The event is inactive, if the context is active
3824 * we're part of a group that didn't make it on the 'pmu',
3827 if (event->ctx->is_active)
3831 * We're inactive and the context is too, this means the
3832 * task is scheduled out, we're counting events that happen
3833 * to us, like migration events.
3838 static int perf_tp_event_match(struct perf_event *event,
3839 struct perf_sample_data *data);
3841 static int perf_swevent_match(struct perf_event *event,
3842 enum perf_type_id type,
3844 struct perf_sample_data *data,
3845 struct pt_regs *regs)
3847 if (!perf_swevent_is_counting(event))
3850 if (event->attr.type != type)
3852 if (event->attr.config != event_id)
3856 if (event->attr.exclude_user && user_mode(regs))
3859 if (event->attr.exclude_kernel && !user_mode(regs))
3863 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3864 !perf_tp_event_match(event, data))
3870 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3871 enum perf_type_id type,
3872 u32 event_id, u64 nr, int nmi,
3873 struct perf_sample_data *data,
3874 struct pt_regs *regs)
3876 struct perf_event *event;
3878 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3882 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3883 if (perf_swevent_match(event, type, event_id, data, regs))
3884 perf_swevent_add(event, nr, nmi, data, regs);
3889 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3892 return &cpuctx->recursion[3];
3895 return &cpuctx->recursion[2];
3898 return &cpuctx->recursion[1];
3900 return &cpuctx->recursion[0];
3903 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3905 struct perf_sample_data *data,
3906 struct pt_regs *regs)
3908 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3909 int *recursion = perf_swevent_recursion_context(cpuctx);
3910 struct perf_event_context *ctx;
3918 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3919 nr, nmi, data, regs);
3922 * doesn't really matter which of the child contexts the
3923 * events ends up in.
3925 ctx = rcu_dereference(current->perf_event_ctxp);
3927 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3934 put_cpu_var(perf_cpu_context);
3937 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3938 struct pt_regs *regs, u64 addr)
3940 struct perf_sample_data data = {
3944 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3948 static void perf_swevent_read(struct perf_event *event)
3952 static int perf_swevent_enable(struct perf_event *event)
3954 struct hw_perf_event *hwc = &event->hw;
3956 if (hwc->sample_period) {
3957 hwc->last_period = hwc->sample_period;
3958 perf_swevent_set_period(event);
3963 static void perf_swevent_disable(struct perf_event *event)
3967 static const struct pmu perf_ops_generic = {
3968 .enable = perf_swevent_enable,
3969 .disable = perf_swevent_disable,
3970 .read = perf_swevent_read,
3971 .unthrottle = perf_swevent_unthrottle,
3975 * hrtimer based swevent callback
3978 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3980 enum hrtimer_restart ret = HRTIMER_RESTART;
3981 struct perf_sample_data data;
3982 struct pt_regs *regs;
3983 struct perf_event *event;
3986 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
3987 event->pmu->read(event);
3990 regs = get_irq_regs();
3992 * In case we exclude kernel IPs or are somehow not in interrupt
3993 * context, provide the next best thing, the user IP.
3995 if ((event->attr.exclude_kernel || !regs) &&
3996 !event->attr.exclude_user)
3997 regs = task_pt_regs(current);
4000 if (perf_event_overflow(event, 0, &data, regs))
4001 ret = HRTIMER_NORESTART;
4004 period = max_t(u64, 10000, event->hw.sample_period);
4005 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4011 * Software event: cpu wall time clock
4014 static void cpu_clock_perf_event_update(struct perf_event *event)
4016 int cpu = raw_smp_processor_id();
4020 now = cpu_clock(cpu);
4021 prev = atomic64_read(&event->hw.prev_count);
4022 atomic64_set(&event->hw.prev_count, now);
4023 atomic64_add(now - prev, &event->count);
4026 static int cpu_clock_perf_event_enable(struct perf_event *event)
4028 struct hw_perf_event *hwc = &event->hw;
4029 int cpu = raw_smp_processor_id();
4031 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4032 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4033 hwc->hrtimer.function = perf_swevent_hrtimer;
4034 if (hwc->sample_period) {
4035 u64 period = max_t(u64, 10000, hwc->sample_period);
4036 __hrtimer_start_range_ns(&hwc->hrtimer,
4037 ns_to_ktime(period), 0,
4038 HRTIMER_MODE_REL, 0);
4044 static void cpu_clock_perf_event_disable(struct perf_event *event)
4046 if (event->hw.sample_period)
4047 hrtimer_cancel(&event->hw.hrtimer);
4048 cpu_clock_perf_event_update(event);
4051 static void cpu_clock_perf_event_read(struct perf_event *event)
4053 cpu_clock_perf_event_update(event);
4056 static const struct pmu perf_ops_cpu_clock = {
4057 .enable = cpu_clock_perf_event_enable,
4058 .disable = cpu_clock_perf_event_disable,
4059 .read = cpu_clock_perf_event_read,
4063 * Software event: task time clock
4066 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4071 prev = atomic64_xchg(&event->hw.prev_count, now);
4073 atomic64_add(delta, &event->count);
4076 static int task_clock_perf_event_enable(struct perf_event *event)
4078 struct hw_perf_event *hwc = &event->hw;
4081 now = event->ctx->time;
4083 atomic64_set(&hwc->prev_count, now);
4084 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4085 hwc->hrtimer.function = perf_swevent_hrtimer;
4086 if (hwc->sample_period) {
4087 u64 period = max_t(u64, 10000, hwc->sample_period);
4088 __hrtimer_start_range_ns(&hwc->hrtimer,
4089 ns_to_ktime(period), 0,
4090 HRTIMER_MODE_REL, 0);
4096 static void task_clock_perf_event_disable(struct perf_event *event)
4098 if (event->hw.sample_period)
4099 hrtimer_cancel(&event->hw.hrtimer);
4100 task_clock_perf_event_update(event, event->ctx->time);
4104 static void task_clock_perf_event_read(struct perf_event *event)
4109 update_context_time(event->ctx);
4110 time = event->ctx->time;
4112 u64 now = perf_clock();
4113 u64 delta = now - event->ctx->timestamp;
4114 time = event->ctx->time + delta;
4117 task_clock_perf_event_update(event, time);
4120 static const struct pmu perf_ops_task_clock = {
4121 .enable = task_clock_perf_event_enable,
4122 .disable = task_clock_perf_event_disable,
4123 .read = task_clock_perf_event_read,
4126 #ifdef CONFIG_EVENT_PROFILE
4128 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4131 struct perf_raw_record raw = {
4136 struct perf_sample_data data = {
4141 struct pt_regs *regs = get_irq_regs();
4144 regs = task_pt_regs(current);
4146 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4149 EXPORT_SYMBOL_GPL(perf_tp_event);
4151 static int perf_tp_event_match(struct perf_event *event,
4152 struct perf_sample_data *data)
4154 void *record = data->raw->data;
4156 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4161 static void tp_perf_event_destroy(struct perf_event *event)
4163 ftrace_profile_disable(event->attr.config);
4166 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4169 * Raw tracepoint data is a severe data leak, only allow root to
4172 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4173 perf_paranoid_tracepoint_raw() &&
4174 !capable(CAP_SYS_ADMIN))
4175 return ERR_PTR(-EPERM);
4177 if (ftrace_profile_enable(event->attr.config))
4180 event->destroy = tp_perf_event_destroy;
4182 return &perf_ops_generic;
4185 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4190 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4193 filter_str = strndup_user(arg, PAGE_SIZE);
4194 if (IS_ERR(filter_str))
4195 return PTR_ERR(filter_str);
4197 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4203 static void perf_event_free_filter(struct perf_event *event)
4205 ftrace_profile_free_filter(event);
4210 static int perf_tp_event_match(struct perf_event *event,
4211 struct perf_sample_data *data)
4216 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4221 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4226 static void perf_event_free_filter(struct perf_event *event)
4230 #endif /* CONFIG_EVENT_PROFILE */
4232 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4234 static void sw_perf_event_destroy(struct perf_event *event)
4236 u64 event_id = event->attr.config;
4238 WARN_ON(event->parent);
4240 atomic_dec(&perf_swevent_enabled[event_id]);
4243 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4245 const struct pmu *pmu = NULL;
4246 u64 event_id = event->attr.config;
4249 * Software events (currently) can't in general distinguish
4250 * between user, kernel and hypervisor events.
4251 * However, context switches and cpu migrations are considered
4252 * to be kernel events, and page faults are never hypervisor
4256 case PERF_COUNT_SW_CPU_CLOCK:
4257 pmu = &perf_ops_cpu_clock;
4260 case PERF_COUNT_SW_TASK_CLOCK:
4262 * If the user instantiates this as a per-cpu event,
4263 * use the cpu_clock event instead.
4265 if (event->ctx->task)
4266 pmu = &perf_ops_task_clock;
4268 pmu = &perf_ops_cpu_clock;
4271 case PERF_COUNT_SW_PAGE_FAULTS:
4272 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4273 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4274 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4275 case PERF_COUNT_SW_CPU_MIGRATIONS:
4276 if (!event->parent) {
4277 atomic_inc(&perf_swevent_enabled[event_id]);
4278 event->destroy = sw_perf_event_destroy;
4280 pmu = &perf_ops_generic;
4288 * Allocate and initialize a event structure
4290 static struct perf_event *
4291 perf_event_alloc(struct perf_event_attr *attr,
4293 struct perf_event_context *ctx,
4294 struct perf_event *group_leader,
4295 struct perf_event *parent_event,
4298 const struct pmu *pmu;
4299 struct perf_event *event;
4300 struct hw_perf_event *hwc;
4303 event = kzalloc(sizeof(*event), gfpflags);
4305 return ERR_PTR(-ENOMEM);
4308 * Single events are their own group leaders, with an
4309 * empty sibling list:
4312 group_leader = event;
4314 mutex_init(&event->child_mutex);
4315 INIT_LIST_HEAD(&event->child_list);
4317 INIT_LIST_HEAD(&event->group_entry);
4318 INIT_LIST_HEAD(&event->event_entry);
4319 INIT_LIST_HEAD(&event->sibling_list);
4320 init_waitqueue_head(&event->waitq);
4322 mutex_init(&event->mmap_mutex);
4325 event->attr = *attr;
4326 event->group_leader = group_leader;
4331 event->parent = parent_event;
4333 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4334 event->id = atomic64_inc_return(&perf_event_id);
4336 event->state = PERF_EVENT_STATE_INACTIVE;
4339 event->state = PERF_EVENT_STATE_OFF;
4344 hwc->sample_period = attr->sample_period;
4345 if (attr->freq && attr->sample_freq)
4346 hwc->sample_period = 1;
4347 hwc->last_period = hwc->sample_period;
4349 atomic64_set(&hwc->period_left, hwc->sample_period);
4352 * we currently do not support PERF_FORMAT_GROUP on inherited events
4354 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4357 switch (attr->type) {
4359 case PERF_TYPE_HARDWARE:
4360 case PERF_TYPE_HW_CACHE:
4361 pmu = hw_perf_event_init(event);
4364 case PERF_TYPE_SOFTWARE:
4365 pmu = sw_perf_event_init(event);
4368 case PERF_TYPE_TRACEPOINT:
4369 pmu = tp_perf_event_init(event);
4379 else if (IS_ERR(pmu))
4384 put_pid_ns(event->ns);
4386 return ERR_PTR(err);
4391 if (!event->parent) {
4392 atomic_inc(&nr_events);
4393 if (event->attr.mmap)
4394 atomic_inc(&nr_mmap_events);
4395 if (event->attr.comm)
4396 atomic_inc(&nr_comm_events);
4397 if (event->attr.task)
4398 atomic_inc(&nr_task_events);
4404 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4405 struct perf_event_attr *attr)
4410 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4414 * zero the full structure, so that a short copy will be nice.
4416 memset(attr, 0, sizeof(*attr));
4418 ret = get_user(size, &uattr->size);
4422 if (size > PAGE_SIZE) /* silly large */
4425 if (!size) /* abi compat */
4426 size = PERF_ATTR_SIZE_VER0;
4428 if (size < PERF_ATTR_SIZE_VER0)
4432 * If we're handed a bigger struct than we know of,
4433 * ensure all the unknown bits are 0 - i.e. new
4434 * user-space does not rely on any kernel feature
4435 * extensions we dont know about yet.
4437 if (size > sizeof(*attr)) {
4438 unsigned char __user *addr;
4439 unsigned char __user *end;
4442 addr = (void __user *)uattr + sizeof(*attr);
4443 end = (void __user *)uattr + size;
4445 for (; addr < end; addr++) {
4446 ret = get_user(val, addr);
4452 size = sizeof(*attr);
4455 ret = copy_from_user(attr, uattr, size);
4460 * If the type exists, the corresponding creation will verify
4463 if (attr->type >= PERF_TYPE_MAX)
4466 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4469 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4472 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4479 put_user(sizeof(*attr), &uattr->size);
4484 static int perf_event_set_output(struct perf_event *event, int output_fd)
4486 struct perf_event *output_event = NULL;
4487 struct file *output_file = NULL;
4488 struct perf_event *old_output;
4489 int fput_needed = 0;
4495 output_file = fget_light(output_fd, &fput_needed);
4499 if (output_file->f_op != &perf_fops)
4502 output_event = output_file->private_data;
4504 /* Don't chain output fds */
4505 if (output_event->output)
4508 /* Don't set an output fd when we already have an output channel */
4512 atomic_long_inc(&output_file->f_count);
4515 mutex_lock(&event->mmap_mutex);
4516 old_output = event->output;
4517 rcu_assign_pointer(event->output, output_event);
4518 mutex_unlock(&event->mmap_mutex);
4522 * we need to make sure no existing perf_output_*()
4523 * is still referencing this event.
4526 fput(old_output->filp);
4531 fput_light(output_file, fput_needed);
4536 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4538 * @attr_uptr: event_id type attributes for monitoring/sampling
4541 * @group_fd: group leader event fd
4543 SYSCALL_DEFINE5(perf_event_open,
4544 struct perf_event_attr __user *, attr_uptr,
4545 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4547 struct perf_event *event, *group_leader;
4548 struct perf_event_attr attr;
4549 struct perf_event_context *ctx;
4550 struct file *event_file = NULL;
4551 struct file *group_file = NULL;
4552 int fput_needed = 0;
4553 int fput_needed2 = 0;
4556 /* for future expandability... */
4557 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4560 err = perf_copy_attr(attr_uptr, &attr);
4564 if (!attr.exclude_kernel) {
4565 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4570 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4575 * Get the target context (task or percpu):
4577 ctx = find_get_context(pid, cpu);
4579 return PTR_ERR(ctx);
4582 * Look up the group leader (we will attach this event to it):
4584 group_leader = NULL;
4585 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4587 group_file = fget_light(group_fd, &fput_needed);
4589 goto err_put_context;
4590 if (group_file->f_op != &perf_fops)
4591 goto err_put_context;
4593 group_leader = group_file->private_data;
4595 * Do not allow a recursive hierarchy (this new sibling
4596 * becoming part of another group-sibling):
4598 if (group_leader->group_leader != group_leader)
4599 goto err_put_context;
4601 * Do not allow to attach to a group in a different
4602 * task or CPU context:
4604 if (group_leader->ctx != ctx)
4605 goto err_put_context;
4607 * Only a group leader can be exclusive or pinned
4609 if (attr.exclusive || attr.pinned)
4610 goto err_put_context;
4613 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4615 err = PTR_ERR(event);
4617 goto err_put_context;
4619 err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4621 goto err_free_put_context;
4623 event_file = fget_light(err, &fput_needed2);
4625 goto err_free_put_context;
4627 if (flags & PERF_FLAG_FD_OUTPUT) {
4628 err = perf_event_set_output(event, group_fd);
4630 goto err_fput_free_put_context;
4633 event->filp = event_file;
4634 WARN_ON_ONCE(ctx->parent_ctx);
4635 mutex_lock(&ctx->mutex);
4636 perf_install_in_context(ctx, event, cpu);
4638 mutex_unlock(&ctx->mutex);
4640 event->owner = current;
4641 get_task_struct(current);
4642 mutex_lock(¤t->perf_event_mutex);
4643 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4644 mutex_unlock(¤t->perf_event_mutex);
4646 err_fput_free_put_context:
4647 fput_light(event_file, fput_needed2);
4649 err_free_put_context:
4657 fput_light(group_file, fput_needed);
4663 * perf_event_create_kernel_counter
4665 * @attr: attributes of the counter to create
4666 * @cpu: cpu in which the counter is bound
4667 * @pid: task to profile
4670 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4673 struct perf_event *event;
4674 struct perf_event_context *ctx;
4678 * Get the target context (task or percpu):
4681 ctx = find_get_context(pid, cpu);
4685 event = perf_event_alloc(attr, cpu, ctx, NULL,
4687 err = PTR_ERR(event);
4689 goto err_put_context;
4692 WARN_ON_ONCE(ctx->parent_ctx);
4693 mutex_lock(&ctx->mutex);
4694 perf_install_in_context(ctx, event, cpu);
4696 mutex_unlock(&ctx->mutex);
4698 event->owner = current;
4699 get_task_struct(current);
4700 mutex_lock(¤t->perf_event_mutex);
4701 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4702 mutex_unlock(¤t->perf_event_mutex);
4712 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4715 * inherit a event from parent task to child task:
4717 static struct perf_event *
4718 inherit_event(struct perf_event *parent_event,
4719 struct task_struct *parent,
4720 struct perf_event_context *parent_ctx,
4721 struct task_struct *child,
4722 struct perf_event *group_leader,
4723 struct perf_event_context *child_ctx)
4725 struct perf_event *child_event;
4728 * Instead of creating recursive hierarchies of events,
4729 * we link inherited events back to the original parent,
4730 * which has a filp for sure, which we use as the reference
4733 if (parent_event->parent)
4734 parent_event = parent_event->parent;
4736 child_event = perf_event_alloc(&parent_event->attr,
4737 parent_event->cpu, child_ctx,
4738 group_leader, parent_event,
4740 if (IS_ERR(child_event))
4745 * Make the child state follow the state of the parent event,
4746 * not its attr.disabled bit. We hold the parent's mutex,
4747 * so we won't race with perf_event_{en, dis}able_family.
4749 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4750 child_event->state = PERF_EVENT_STATE_INACTIVE;
4752 child_event->state = PERF_EVENT_STATE_OFF;
4754 if (parent_event->attr.freq)
4755 child_event->hw.sample_period = parent_event->hw.sample_period;
4758 * Link it up in the child's context:
4760 add_event_to_ctx(child_event, child_ctx);
4763 * Get a reference to the parent filp - we will fput it
4764 * when the child event exits. This is safe to do because
4765 * we are in the parent and we know that the filp still
4766 * exists and has a nonzero count:
4768 atomic_long_inc(&parent_event->filp->f_count);
4771 * Link this into the parent event's child list
4773 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4774 mutex_lock(&parent_event->child_mutex);
4775 list_add_tail(&child_event->child_list, &parent_event->child_list);
4776 mutex_unlock(&parent_event->child_mutex);
4781 static int inherit_group(struct perf_event *parent_event,
4782 struct task_struct *parent,
4783 struct perf_event_context *parent_ctx,
4784 struct task_struct *child,
4785 struct perf_event_context *child_ctx)
4787 struct perf_event *leader;
4788 struct perf_event *sub;
4789 struct perf_event *child_ctr;
4791 leader = inherit_event(parent_event, parent, parent_ctx,
4792 child, NULL, child_ctx);
4794 return PTR_ERR(leader);
4795 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4796 child_ctr = inherit_event(sub, parent, parent_ctx,
4797 child, leader, child_ctx);
4798 if (IS_ERR(child_ctr))
4799 return PTR_ERR(child_ctr);
4804 static void sync_child_event(struct perf_event *child_event,
4805 struct task_struct *child)
4807 struct perf_event *parent_event = child_event->parent;
4810 if (child_event->attr.inherit_stat)
4811 perf_event_read_event(child_event, child);
4813 child_val = atomic64_read(&child_event->count);
4816 * Add back the child's count to the parent's count:
4818 atomic64_add(child_val, &parent_event->count);
4819 atomic64_add(child_event->total_time_enabled,
4820 &parent_event->child_total_time_enabled);
4821 atomic64_add(child_event->total_time_running,
4822 &parent_event->child_total_time_running);
4825 * Remove this event from the parent's list
4827 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4828 mutex_lock(&parent_event->child_mutex);
4829 list_del_init(&child_event->child_list);
4830 mutex_unlock(&parent_event->child_mutex);
4833 * Release the parent event, if this was the last
4836 fput(parent_event->filp);
4840 __perf_event_exit_task(struct perf_event *child_event,
4841 struct perf_event_context *child_ctx,
4842 struct task_struct *child)
4844 struct perf_event *parent_event;
4846 update_event_times(child_event);
4847 perf_event_remove_from_context(child_event);
4849 parent_event = child_event->parent;
4851 * It can happen that parent exits first, and has events
4852 * that are still around due to the child reference. These
4853 * events need to be zapped - but otherwise linger.
4856 sync_child_event(child_event, child);
4857 free_event(child_event);
4862 * When a child task exits, feed back event values to parent events.
4864 void perf_event_exit_task(struct task_struct *child)
4866 struct perf_event *child_event, *tmp;
4867 struct perf_event_context *child_ctx;
4868 unsigned long flags;
4870 if (likely(!child->perf_event_ctxp)) {
4871 perf_event_task(child, NULL, 0);
4875 local_irq_save(flags);
4877 * We can't reschedule here because interrupts are disabled,
4878 * and either child is current or it is a task that can't be
4879 * scheduled, so we are now safe from rescheduling changing
4882 child_ctx = child->perf_event_ctxp;
4883 __perf_event_task_sched_out(child_ctx);
4886 * Take the context lock here so that if find_get_context is
4887 * reading child->perf_event_ctxp, we wait until it has
4888 * incremented the context's refcount before we do put_ctx below.
4890 spin_lock(&child_ctx->lock);
4891 child->perf_event_ctxp = NULL;
4893 * If this context is a clone; unclone it so it can't get
4894 * swapped to another process while we're removing all
4895 * the events from it.
4897 unclone_ctx(child_ctx);
4898 spin_unlock_irqrestore(&child_ctx->lock, flags);
4901 * Report the task dead after unscheduling the events so that we
4902 * won't get any samples after PERF_RECORD_EXIT. We can however still
4903 * get a few PERF_RECORD_READ events.
4905 perf_event_task(child, child_ctx, 0);
4908 * We can recurse on the same lock type through:
4910 * __perf_event_exit_task()
4911 * sync_child_event()
4912 * fput(parent_event->filp)
4914 * mutex_lock(&ctx->mutex)
4916 * But since its the parent context it won't be the same instance.
4918 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4921 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4923 __perf_event_exit_task(child_event, child_ctx, child);
4926 * If the last event was a group event, it will have appended all
4927 * its siblings to the list, but we obtained 'tmp' before that which
4928 * will still point to the list head terminating the iteration.
4930 if (!list_empty(&child_ctx->group_list))
4933 mutex_unlock(&child_ctx->mutex);
4939 * free an unexposed, unused context as created by inheritance by
4940 * init_task below, used by fork() in case of fail.
4942 void perf_event_free_task(struct task_struct *task)
4944 struct perf_event_context *ctx = task->perf_event_ctxp;
4945 struct perf_event *event, *tmp;
4950 mutex_lock(&ctx->mutex);
4952 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
4953 struct perf_event *parent = event->parent;
4955 if (WARN_ON_ONCE(!parent))
4958 mutex_lock(&parent->child_mutex);
4959 list_del_init(&event->child_list);
4960 mutex_unlock(&parent->child_mutex);
4964 list_del_event(event, ctx);
4968 if (!list_empty(&ctx->group_list))
4971 mutex_unlock(&ctx->mutex);
4977 * Initialize the perf_event context in task_struct
4979 int perf_event_init_task(struct task_struct *child)
4981 struct perf_event_context *child_ctx, *parent_ctx;
4982 struct perf_event_context *cloned_ctx;
4983 struct perf_event *event;
4984 struct task_struct *parent = current;
4985 int inherited_all = 1;
4988 child->perf_event_ctxp = NULL;
4990 mutex_init(&child->perf_event_mutex);
4991 INIT_LIST_HEAD(&child->perf_event_list);
4993 if (likely(!parent->perf_event_ctxp))
4997 * This is executed from the parent task context, so inherit
4998 * events that have been marked for cloning.
4999 * First allocate and initialize a context for the child.
5002 child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5006 __perf_event_init_context(child_ctx, child);
5007 child->perf_event_ctxp = child_ctx;
5008 get_task_struct(child);
5011 * If the parent's context is a clone, pin it so it won't get
5014 parent_ctx = perf_pin_task_context(parent);
5017 * No need to check if parent_ctx != NULL here; since we saw
5018 * it non-NULL earlier, the only reason for it to become NULL
5019 * is if we exit, and since we're currently in the middle of
5020 * a fork we can't be exiting at the same time.
5024 * Lock the parent list. No need to lock the child - not PID
5025 * hashed yet and not running, so nobody can access it.
5027 mutex_lock(&parent_ctx->mutex);
5030 * We dont have to disable NMIs - we are only looking at
5031 * the list, not manipulating it:
5033 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5035 if (!event->attr.inherit) {
5040 ret = inherit_group(event, parent, parent_ctx,
5048 if (inherited_all) {
5050 * Mark the child context as a clone of the parent
5051 * context, or of whatever the parent is a clone of.
5052 * Note that if the parent is a clone, it could get
5053 * uncloned at any point, but that doesn't matter
5054 * because the list of events and the generation
5055 * count can't have changed since we took the mutex.
5057 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5059 child_ctx->parent_ctx = cloned_ctx;
5060 child_ctx->parent_gen = parent_ctx->parent_gen;
5062 child_ctx->parent_ctx = parent_ctx;
5063 child_ctx->parent_gen = parent_ctx->generation;
5065 get_ctx(child_ctx->parent_ctx);
5068 mutex_unlock(&parent_ctx->mutex);
5070 perf_unpin_context(parent_ctx);
5075 static void __cpuinit perf_event_init_cpu(int cpu)
5077 struct perf_cpu_context *cpuctx;
5079 cpuctx = &per_cpu(perf_cpu_context, cpu);
5080 __perf_event_init_context(&cpuctx->ctx, NULL);
5082 spin_lock(&perf_resource_lock);
5083 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5084 spin_unlock(&perf_resource_lock);
5086 hw_perf_event_setup(cpu);
5089 #ifdef CONFIG_HOTPLUG_CPU
5090 static void __perf_event_exit_cpu(void *info)
5092 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5093 struct perf_event_context *ctx = &cpuctx->ctx;
5094 struct perf_event *event, *tmp;
5096 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5097 __perf_event_remove_from_context(event);
5099 static void perf_event_exit_cpu(int cpu)
5101 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5102 struct perf_event_context *ctx = &cpuctx->ctx;
5104 mutex_lock(&ctx->mutex);
5105 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5106 mutex_unlock(&ctx->mutex);
5109 static inline void perf_event_exit_cpu(int cpu) { }
5112 static int __cpuinit
5113 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5115 unsigned int cpu = (long)hcpu;
5119 case CPU_UP_PREPARE:
5120 case CPU_UP_PREPARE_FROZEN:
5121 perf_event_init_cpu(cpu);
5125 case CPU_ONLINE_FROZEN:
5126 hw_perf_event_setup_online(cpu);
5129 case CPU_DOWN_PREPARE:
5130 case CPU_DOWN_PREPARE_FROZEN:
5131 perf_event_exit_cpu(cpu);
5142 * This has to have a higher priority than migration_notifier in sched.c.
5144 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5145 .notifier_call = perf_cpu_notify,
5149 void __init perf_event_init(void)
5151 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5152 (void *)(long)smp_processor_id());
5153 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5154 (void *)(long)smp_processor_id());
5155 register_cpu_notifier(&perf_cpu_nb);
5158 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5160 return sprintf(buf, "%d\n", perf_reserved_percpu);
5164 perf_set_reserve_percpu(struct sysdev_class *class,
5168 struct perf_cpu_context *cpuctx;
5172 err = strict_strtoul(buf, 10, &val);
5175 if (val > perf_max_events)
5178 spin_lock(&perf_resource_lock);
5179 perf_reserved_percpu = val;
5180 for_each_online_cpu(cpu) {
5181 cpuctx = &per_cpu(perf_cpu_context, cpu);
5182 spin_lock_irq(&cpuctx->ctx.lock);
5183 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5184 perf_max_events - perf_reserved_percpu);
5185 cpuctx->max_pertask = mpt;
5186 spin_unlock_irq(&cpuctx->ctx.lock);
5188 spin_unlock(&perf_resource_lock);
5193 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5195 return sprintf(buf, "%d\n", perf_overcommit);
5199 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5204 err = strict_strtoul(buf, 10, &val);
5210 spin_lock(&perf_resource_lock);
5211 perf_overcommit = val;
5212 spin_unlock(&perf_resource_lock);
5217 static SYSDEV_CLASS_ATTR(
5220 perf_show_reserve_percpu,
5221 perf_set_reserve_percpu
5224 static SYSDEV_CLASS_ATTR(
5227 perf_show_overcommit,
5231 static struct attribute *perfclass_attrs[] = {
5232 &attr_reserve_percpu.attr,
5233 &attr_overcommit.attr,
5237 static struct attribute_group perfclass_attr_group = {
5238 .attrs = perfclass_attrs,
5239 .name = "perf_events",
5242 static int __init perf_event_sysfs_init(void)
5244 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5245 &perfclass_attr_group);
5247 device_initcall(perf_event_sysfs_init);