2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
34 #include <asm/irq_regs.h>
37 * Each CPU has a list of per CPU events:
39 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly = 1;
59 static inline bool perf_paranoid_tracepoint_raw(void)
61 return sysctl_perf_event_paranoid > -1;
64 static inline bool perf_paranoid_cpu(void)
66 return sysctl_perf_event_paranoid > 0;
69 static inline bool perf_paranoid_kernel(void)
71 return sysctl_perf_event_paranoid > 1;
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
77 * max perf event sample rate
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
81 static atomic64_t perf_event_id;
84 * Lock for (sysadmin-configurable) event reservations:
86 static DEFINE_SPINLOCK(perf_resource_lock);
89 * Architecture provided APIs - weak aliases:
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
96 void __weak hw_perf_disable(void) { barrier(); }
97 void __weak hw_perf_enable(void) { barrier(); }
99 void __weak hw_perf_event_setup(int cpu) { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104 struct perf_cpu_context *cpuctx,
105 struct perf_event_context *ctx, int cpu)
110 void __weak perf_event_print_debug(void) { }
112 static DEFINE_PER_CPU(int, perf_disable_count);
114 void __perf_disable(void)
116 __get_cpu_var(perf_disable_count)++;
119 bool __perf_enable(void)
121 return !--__get_cpu_var(perf_disable_count);
124 void perf_disable(void)
130 void perf_enable(void)
136 static void get_ctx(struct perf_event_context *ctx)
138 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
141 static void free_ctx(struct rcu_head *head)
143 struct perf_event_context *ctx;
145 ctx = container_of(head, struct perf_event_context, rcu_head);
149 static void put_ctx(struct perf_event_context *ctx)
151 if (atomic_dec_and_test(&ctx->refcount)) {
153 put_ctx(ctx->parent_ctx);
155 put_task_struct(ctx->task);
156 call_rcu(&ctx->rcu_head, free_ctx);
160 static void unclone_ctx(struct perf_event_context *ctx)
162 if (ctx->parent_ctx) {
163 put_ctx(ctx->parent_ctx);
164 ctx->parent_ctx = NULL;
169 * If we inherit events we want to return the parent event id
172 static u64 primary_event_id(struct perf_event *event)
177 id = event->parent->id;
183 * Get the perf_event_context for a task and lock it.
184 * This has to cope with with the fact that until it is locked,
185 * the context could get moved to another task.
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
190 struct perf_event_context *ctx;
194 ctx = rcu_dereference(task->perf_event_ctxp);
197 * If this context is a clone of another, it might
198 * get swapped for another underneath us by
199 * perf_event_task_sched_out, though the
200 * rcu_read_lock() protects us from any context
201 * getting freed. Lock the context and check if it
202 * got swapped before we could get the lock, and retry
203 * if so. If we locked the right context, then it
204 * can't get swapped on us any more.
206 spin_lock_irqsave(&ctx->lock, *flags);
207 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208 spin_unlock_irqrestore(&ctx->lock, *flags);
212 if (!atomic_inc_not_zero(&ctx->refcount)) {
213 spin_unlock_irqrestore(&ctx->lock, *flags);
222 * Get the context for a task and increment its pin_count so it
223 * can't get swapped to another task. This also increments its
224 * reference count so that the context can't get freed.
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
228 struct perf_event_context *ctx;
231 ctx = perf_lock_task_context(task, &flags);
234 spin_unlock_irqrestore(&ctx->lock, flags);
239 static void perf_unpin_context(struct perf_event_context *ctx)
243 spin_lock_irqsave(&ctx->lock, flags);
245 spin_unlock_irqrestore(&ctx->lock, flags);
250 * Add a event from the lists for its context.
251 * Must be called with ctx->mutex and ctx->lock held.
254 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
256 struct perf_event *group_leader = event->group_leader;
259 * Depending on whether it is a standalone or sibling event,
260 * add it straight to the context's event list, or to the group
261 * leader's sibling list:
263 if (group_leader == event)
264 list_add_tail(&event->group_entry, &ctx->group_list);
266 list_add_tail(&event->group_entry, &group_leader->sibling_list);
267 group_leader->nr_siblings++;
270 list_add_rcu(&event->event_entry, &ctx->event_list);
272 if (event->attr.inherit_stat)
277 * Remove a event from the lists for its context.
278 * Must be called with ctx->mutex and ctx->lock held.
281 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
283 struct perf_event *sibling, *tmp;
285 if (list_empty(&event->group_entry))
288 if (event->attr.inherit_stat)
291 list_del_init(&event->group_entry);
292 list_del_rcu(&event->event_entry);
294 if (event->group_leader != event)
295 event->group_leader->nr_siblings--;
298 * If this was a group event with sibling events then
299 * upgrade the siblings to singleton events by adding them
300 * to the context list directly:
302 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
304 list_move_tail(&sibling->group_entry, &ctx->group_list);
305 sibling->group_leader = sibling;
310 event_sched_out(struct perf_event *event,
311 struct perf_cpu_context *cpuctx,
312 struct perf_event_context *ctx)
314 if (event->state != PERF_EVENT_STATE_ACTIVE)
317 event->state = PERF_EVENT_STATE_INACTIVE;
318 if (event->pending_disable) {
319 event->pending_disable = 0;
320 event->state = PERF_EVENT_STATE_OFF;
322 event->tstamp_stopped = ctx->time;
323 event->pmu->disable(event);
326 if (!is_software_event(event))
327 cpuctx->active_oncpu--;
329 if (event->attr.exclusive || !cpuctx->active_oncpu)
330 cpuctx->exclusive = 0;
334 group_sched_out(struct perf_event *group_event,
335 struct perf_cpu_context *cpuctx,
336 struct perf_event_context *ctx)
338 struct perf_event *event;
340 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
343 event_sched_out(group_event, cpuctx, ctx);
346 * Schedule out siblings (if any):
348 list_for_each_entry(event, &group_event->sibling_list, group_entry)
349 event_sched_out(event, cpuctx, ctx);
351 if (group_event->attr.exclusive)
352 cpuctx->exclusive = 0;
356 * Cross CPU call to remove a performance event
358 * We disable the event on the hardware level first. After that we
359 * remove it from the context list.
361 static void __perf_event_remove_from_context(void *info)
363 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
364 struct perf_event *event = info;
365 struct perf_event_context *ctx = event->ctx;
368 * If this is a task context, we need to check whether it is
369 * the current task context of this cpu. If not it has been
370 * scheduled out before the smp call arrived.
372 if (ctx->task && cpuctx->task_ctx != ctx)
375 spin_lock(&ctx->lock);
377 * Protect the list operation against NMI by disabling the
378 * events on a global level.
382 event_sched_out(event, cpuctx, ctx);
384 list_del_event(event, ctx);
388 * Allow more per task events with respect to the
391 cpuctx->max_pertask =
392 min(perf_max_events - ctx->nr_events,
393 perf_max_events - perf_reserved_percpu);
397 spin_unlock(&ctx->lock);
402 * Remove the event from a task's (or a CPU's) list of events.
404 * Must be called with ctx->mutex held.
406 * CPU events are removed with a smp call. For task events we only
407 * call when the task is on a CPU.
409 * If event->ctx is a cloned context, callers must make sure that
410 * every task struct that event->ctx->task could possibly point to
411 * remains valid. This is OK when called from perf_release since
412 * that only calls us on the top-level context, which can't be a clone.
413 * When called from perf_event_exit_task, it's OK because the
414 * context has been detached from its task.
416 static void perf_event_remove_from_context(struct perf_event *event)
418 struct perf_event_context *ctx = event->ctx;
419 struct task_struct *task = ctx->task;
423 * Per cpu events are removed via an smp call and
424 * the removal is always sucessful.
426 smp_call_function_single(event->cpu,
427 __perf_event_remove_from_context,
433 task_oncpu_function_call(task, __perf_event_remove_from_context,
436 spin_lock_irq(&ctx->lock);
438 * If the context is active we need to retry the smp call.
440 if (ctx->nr_active && !list_empty(&event->group_entry)) {
441 spin_unlock_irq(&ctx->lock);
446 * The lock prevents that this context is scheduled in so we
447 * can remove the event safely, if the call above did not
450 if (!list_empty(&event->group_entry)) {
451 list_del_event(event, ctx);
453 spin_unlock_irq(&ctx->lock);
456 static inline u64 perf_clock(void)
458 return cpu_clock(smp_processor_id());
462 * Update the record of the current time in a context.
464 static void update_context_time(struct perf_event_context *ctx)
466 u64 now = perf_clock();
468 ctx->time += now - ctx->timestamp;
469 ctx->timestamp = now;
473 * Update the total_time_enabled and total_time_running fields for a event.
475 static void update_event_times(struct perf_event *event)
477 struct perf_event_context *ctx = event->ctx;
480 if (event->state < PERF_EVENT_STATE_INACTIVE ||
481 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
484 event->total_time_enabled = ctx->time - event->tstamp_enabled;
486 if (event->state == PERF_EVENT_STATE_INACTIVE)
487 run_end = event->tstamp_stopped;
491 event->total_time_running = run_end - event->tstamp_running;
495 * Update total_time_enabled and total_time_running for all events in a group.
497 static void update_group_times(struct perf_event *leader)
499 struct perf_event *event;
501 update_event_times(leader);
502 list_for_each_entry(event, &leader->sibling_list, group_entry)
503 update_event_times(event);
507 * Cross CPU call to disable a performance event
509 static void __perf_event_disable(void *info)
511 struct perf_event *event = info;
512 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
513 struct perf_event_context *ctx = event->ctx;
516 * If this is a per-task event, need to check whether this
517 * event's task is the current task on this cpu.
519 if (ctx->task && cpuctx->task_ctx != ctx)
522 spin_lock(&ctx->lock);
525 * If the event is on, turn it off.
526 * If it is in error state, leave it in error state.
528 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
529 update_context_time(ctx);
530 update_group_times(event);
531 if (event == event->group_leader)
532 group_sched_out(event, cpuctx, ctx);
534 event_sched_out(event, cpuctx, ctx);
535 event->state = PERF_EVENT_STATE_OFF;
538 spin_unlock(&ctx->lock);
544 * If event->ctx is a cloned context, callers must make sure that
545 * every task struct that event->ctx->task could possibly point to
546 * remains valid. This condition is satisifed when called through
547 * perf_event_for_each_child or perf_event_for_each because they
548 * hold the top-level event's child_mutex, so any descendant that
549 * goes to exit will block in sync_child_event.
550 * When called from perf_pending_event it's OK because event->ctx
551 * is the current context on this CPU and preemption is disabled,
552 * hence we can't get into perf_event_task_sched_out for this context.
554 static void perf_event_disable(struct perf_event *event)
556 struct perf_event_context *ctx = event->ctx;
557 struct task_struct *task = ctx->task;
561 * Disable the event on the cpu that it's on
563 smp_call_function_single(event->cpu, __perf_event_disable,
569 task_oncpu_function_call(task, __perf_event_disable, event);
571 spin_lock_irq(&ctx->lock);
573 * If the event is still active, we need to retry the cross-call.
575 if (event->state == PERF_EVENT_STATE_ACTIVE) {
576 spin_unlock_irq(&ctx->lock);
581 * Since we have the lock this context can't be scheduled
582 * in, so we can change the state safely.
584 if (event->state == PERF_EVENT_STATE_INACTIVE) {
585 update_group_times(event);
586 event->state = PERF_EVENT_STATE_OFF;
589 spin_unlock_irq(&ctx->lock);
593 event_sched_in(struct perf_event *event,
594 struct perf_cpu_context *cpuctx,
595 struct perf_event_context *ctx,
598 if (event->state <= PERF_EVENT_STATE_OFF)
601 event->state = PERF_EVENT_STATE_ACTIVE;
602 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
604 * The new state must be visible before we turn it on in the hardware:
608 if (event->pmu->enable(event)) {
609 event->state = PERF_EVENT_STATE_INACTIVE;
614 event->tstamp_running += ctx->time - event->tstamp_stopped;
616 if (!is_software_event(event))
617 cpuctx->active_oncpu++;
620 if (event->attr.exclusive)
621 cpuctx->exclusive = 1;
627 group_sched_in(struct perf_event *group_event,
628 struct perf_cpu_context *cpuctx,
629 struct perf_event_context *ctx,
632 struct perf_event *event, *partial_group;
635 if (group_event->state == PERF_EVENT_STATE_OFF)
638 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
640 return ret < 0 ? ret : 0;
642 if (event_sched_in(group_event, cpuctx, ctx, cpu))
646 * Schedule in siblings as one group (if any):
648 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
649 if (event_sched_in(event, cpuctx, ctx, cpu)) {
650 partial_group = event;
659 * Groups can be scheduled in as one unit only, so undo any
660 * partial group before returning:
662 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
663 if (event == partial_group)
665 event_sched_out(event, cpuctx, ctx);
667 event_sched_out(group_event, cpuctx, ctx);
673 * Return 1 for a group consisting entirely of software events,
674 * 0 if the group contains any hardware events.
676 static int is_software_only_group(struct perf_event *leader)
678 struct perf_event *event;
680 if (!is_software_event(leader))
683 list_for_each_entry(event, &leader->sibling_list, group_entry)
684 if (!is_software_event(event))
691 * Work out whether we can put this event group on the CPU now.
693 static int group_can_go_on(struct perf_event *event,
694 struct perf_cpu_context *cpuctx,
698 * Groups consisting entirely of software events can always go on.
700 if (is_software_only_group(event))
703 * If an exclusive group is already on, no other hardware
706 if (cpuctx->exclusive)
709 * If this group is exclusive and there are already
710 * events on the CPU, it can't go on.
712 if (event->attr.exclusive && cpuctx->active_oncpu)
715 * Otherwise, try to add it if all previous groups were able
721 static void add_event_to_ctx(struct perf_event *event,
722 struct perf_event_context *ctx)
724 list_add_event(event, ctx);
725 event->tstamp_enabled = ctx->time;
726 event->tstamp_running = ctx->time;
727 event->tstamp_stopped = ctx->time;
731 * Cross CPU call to install and enable a performance event
733 * Must be called with ctx->mutex held
735 static void __perf_install_in_context(void *info)
737 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
738 struct perf_event *event = info;
739 struct perf_event_context *ctx = event->ctx;
740 struct perf_event *leader = event->group_leader;
741 int cpu = smp_processor_id();
745 * If this is a task context, we need to check whether it is
746 * the current task context of this cpu. If not it has been
747 * scheduled out before the smp call arrived.
748 * Or possibly this is the right context but it isn't
749 * on this cpu because it had no events.
751 if (ctx->task && cpuctx->task_ctx != ctx) {
752 if (cpuctx->task_ctx || ctx->task != current)
754 cpuctx->task_ctx = ctx;
757 spin_lock(&ctx->lock);
759 update_context_time(ctx);
762 * Protect the list operation against NMI by disabling the
763 * events on a global level. NOP for non NMI based events.
767 add_event_to_ctx(event, ctx);
770 * Don't put the event on if it is disabled or if
771 * it is in a group and the group isn't on.
773 if (event->state != PERF_EVENT_STATE_INACTIVE ||
774 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
778 * An exclusive event can't go on if there are already active
779 * hardware events, and no hardware event can go on if there
780 * is already an exclusive event on.
782 if (!group_can_go_on(event, cpuctx, 1))
785 err = event_sched_in(event, cpuctx, ctx, cpu);
789 * This event couldn't go on. If it is in a group
790 * then we have to pull the whole group off.
791 * If the event group is pinned then put it in error state.
794 group_sched_out(leader, cpuctx, ctx);
795 if (leader->attr.pinned) {
796 update_group_times(leader);
797 leader->state = PERF_EVENT_STATE_ERROR;
801 if (!err && !ctx->task && cpuctx->max_pertask)
802 cpuctx->max_pertask--;
807 spin_unlock(&ctx->lock);
811 * Attach a performance event to a context
813 * First we add the event to the list with the hardware enable bit
814 * in event->hw_config cleared.
816 * If the event is attached to a task which is on a CPU we use a smp
817 * call to enable it in the task context. The task might have been
818 * scheduled away, but we check this in the smp call again.
820 * Must be called with ctx->mutex held.
823 perf_install_in_context(struct perf_event_context *ctx,
824 struct perf_event *event,
827 struct task_struct *task = ctx->task;
831 * Per cpu events are installed via an smp call and
832 * the install is always sucessful.
834 smp_call_function_single(cpu, __perf_install_in_context,
840 task_oncpu_function_call(task, __perf_install_in_context,
843 spin_lock_irq(&ctx->lock);
845 * we need to retry the smp call.
847 if (ctx->is_active && list_empty(&event->group_entry)) {
848 spin_unlock_irq(&ctx->lock);
853 * The lock prevents that this context is scheduled in so we
854 * can add the event safely, if it the call above did not
857 if (list_empty(&event->group_entry))
858 add_event_to_ctx(event, ctx);
859 spin_unlock_irq(&ctx->lock);
863 * Put a event into inactive state and update time fields.
864 * Enabling the leader of a group effectively enables all
865 * the group members that aren't explicitly disabled, so we
866 * have to update their ->tstamp_enabled also.
867 * Note: this works for group members as well as group leaders
868 * since the non-leader members' sibling_lists will be empty.
870 static void __perf_event_mark_enabled(struct perf_event *event,
871 struct perf_event_context *ctx)
873 struct perf_event *sub;
875 event->state = PERF_EVENT_STATE_INACTIVE;
876 event->tstamp_enabled = ctx->time - event->total_time_enabled;
877 list_for_each_entry(sub, &event->sibling_list, group_entry)
878 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
879 sub->tstamp_enabled =
880 ctx->time - sub->total_time_enabled;
884 * Cross CPU call to enable a performance event
886 static void __perf_event_enable(void *info)
888 struct perf_event *event = info;
889 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
890 struct perf_event_context *ctx = event->ctx;
891 struct perf_event *leader = event->group_leader;
895 * If this is a per-task event, need to check whether this
896 * event's task is the current task on this cpu.
898 if (ctx->task && cpuctx->task_ctx != ctx) {
899 if (cpuctx->task_ctx || ctx->task != current)
901 cpuctx->task_ctx = ctx;
904 spin_lock(&ctx->lock);
906 update_context_time(ctx);
908 if (event->state >= PERF_EVENT_STATE_INACTIVE)
910 __perf_event_mark_enabled(event, ctx);
913 * If the event is in a group and isn't the group leader,
914 * then don't put it on unless the group is on.
916 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
919 if (!group_can_go_on(event, cpuctx, 1)) {
924 err = group_sched_in(event, cpuctx, ctx,
927 err = event_sched_in(event, cpuctx, ctx,
934 * If this event can't go on and it's part of a
935 * group, then the whole group has to come off.
938 group_sched_out(leader, cpuctx, ctx);
939 if (leader->attr.pinned) {
940 update_group_times(leader);
941 leader->state = PERF_EVENT_STATE_ERROR;
946 spin_unlock(&ctx->lock);
952 * If event->ctx is a cloned context, callers must make sure that
953 * every task struct that event->ctx->task could possibly point to
954 * remains valid. This condition is satisfied when called through
955 * perf_event_for_each_child or perf_event_for_each as described
956 * for perf_event_disable.
958 static void perf_event_enable(struct perf_event *event)
960 struct perf_event_context *ctx = event->ctx;
961 struct task_struct *task = ctx->task;
965 * Enable the event on the cpu that it's on
967 smp_call_function_single(event->cpu, __perf_event_enable,
972 spin_lock_irq(&ctx->lock);
973 if (event->state >= PERF_EVENT_STATE_INACTIVE)
977 * If the event is in error state, clear that first.
978 * That way, if we see the event in error state below, we
979 * know that it has gone back into error state, as distinct
980 * from the task having been scheduled away before the
981 * cross-call arrived.
983 if (event->state == PERF_EVENT_STATE_ERROR)
984 event->state = PERF_EVENT_STATE_OFF;
987 spin_unlock_irq(&ctx->lock);
988 task_oncpu_function_call(task, __perf_event_enable, event);
990 spin_lock_irq(&ctx->lock);
993 * If the context is active and the event is still off,
994 * we need to retry the cross-call.
996 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1000 * Since we have the lock this context can't be scheduled
1001 * in, so we can change the state safely.
1003 if (event->state == PERF_EVENT_STATE_OFF)
1004 __perf_event_mark_enabled(event, ctx);
1007 spin_unlock_irq(&ctx->lock);
1010 static int perf_event_refresh(struct perf_event *event, int refresh)
1013 * not supported on inherited events
1015 if (event->attr.inherit)
1018 atomic_add(refresh, &event->event_limit);
1019 perf_event_enable(event);
1024 void __perf_event_sched_out(struct perf_event_context *ctx,
1025 struct perf_cpu_context *cpuctx)
1027 struct perf_event *event;
1029 spin_lock(&ctx->lock);
1031 if (likely(!ctx->nr_events))
1033 update_context_time(ctx);
1037 list_for_each_entry(event, &ctx->group_list, group_entry)
1038 group_sched_out(event, cpuctx, ctx);
1042 spin_unlock(&ctx->lock);
1046 * Test whether two contexts are equivalent, i.e. whether they
1047 * have both been cloned from the same version of the same context
1048 * and they both have the same number of enabled events.
1049 * If the number of enabled events is the same, then the set
1050 * of enabled events should be the same, because these are both
1051 * inherited contexts, therefore we can't access individual events
1052 * in them directly with an fd; we can only enable/disable all
1053 * events via prctl, or enable/disable all events in a family
1054 * via ioctl, which will have the same effect on both contexts.
1056 static int context_equiv(struct perf_event_context *ctx1,
1057 struct perf_event_context *ctx2)
1059 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1060 && ctx1->parent_gen == ctx2->parent_gen
1061 && !ctx1->pin_count && !ctx2->pin_count;
1064 static void __perf_event_sync_stat(struct perf_event *event,
1065 struct perf_event *next_event)
1069 if (!event->attr.inherit_stat)
1073 * Update the event value, we cannot use perf_event_read()
1074 * because we're in the middle of a context switch and have IRQs
1075 * disabled, which upsets smp_call_function_single(), however
1076 * we know the event must be on the current CPU, therefore we
1077 * don't need to use it.
1079 switch (event->state) {
1080 case PERF_EVENT_STATE_ACTIVE:
1081 event->pmu->read(event);
1084 case PERF_EVENT_STATE_INACTIVE:
1085 update_event_times(event);
1093 * In order to keep per-task stats reliable we need to flip the event
1094 * values when we flip the contexts.
1096 value = atomic64_read(&next_event->count);
1097 value = atomic64_xchg(&event->count, value);
1098 atomic64_set(&next_event->count, value);
1100 swap(event->total_time_enabled, next_event->total_time_enabled);
1101 swap(event->total_time_running, next_event->total_time_running);
1104 * Since we swizzled the values, update the user visible data too.
1106 perf_event_update_userpage(event);
1107 perf_event_update_userpage(next_event);
1110 #define list_next_entry(pos, member) \
1111 list_entry(pos->member.next, typeof(*pos), member)
1113 static void perf_event_sync_stat(struct perf_event_context *ctx,
1114 struct perf_event_context *next_ctx)
1116 struct perf_event *event, *next_event;
1121 update_context_time(ctx);
1123 event = list_first_entry(&ctx->event_list,
1124 struct perf_event, event_entry);
1126 next_event = list_first_entry(&next_ctx->event_list,
1127 struct perf_event, event_entry);
1129 while (&event->event_entry != &ctx->event_list &&
1130 &next_event->event_entry != &next_ctx->event_list) {
1132 __perf_event_sync_stat(event, next_event);
1134 event = list_next_entry(event, event_entry);
1135 next_event = list_next_entry(next_event, event_entry);
1140 * Called from scheduler to remove the events of the current task,
1141 * with interrupts disabled.
1143 * We stop each event and update the event value in event->count.
1145 * This does not protect us against NMI, but disable()
1146 * sets the disabled bit in the control field of event _before_
1147 * accessing the event control register. If a NMI hits, then it will
1148 * not restart the event.
1150 void perf_event_task_sched_out(struct task_struct *task,
1151 struct task_struct *next, int cpu)
1153 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1154 struct perf_event_context *ctx = task->perf_event_ctxp;
1155 struct perf_event_context *next_ctx;
1156 struct perf_event_context *parent;
1157 struct pt_regs *regs;
1160 regs = task_pt_regs(task);
1161 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1163 if (likely(!ctx || !cpuctx->task_ctx))
1167 parent = rcu_dereference(ctx->parent_ctx);
1168 next_ctx = next->perf_event_ctxp;
1169 if (parent && next_ctx &&
1170 rcu_dereference(next_ctx->parent_ctx) == parent) {
1172 * Looks like the two contexts are clones, so we might be
1173 * able to optimize the context switch. We lock both
1174 * contexts and check that they are clones under the
1175 * lock (including re-checking that neither has been
1176 * uncloned in the meantime). It doesn't matter which
1177 * order we take the locks because no other cpu could
1178 * be trying to lock both of these tasks.
1180 spin_lock(&ctx->lock);
1181 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1182 if (context_equiv(ctx, next_ctx)) {
1184 * XXX do we need a memory barrier of sorts
1185 * wrt to rcu_dereference() of perf_event_ctxp
1187 task->perf_event_ctxp = next_ctx;
1188 next->perf_event_ctxp = ctx;
1190 next_ctx->task = task;
1193 perf_event_sync_stat(ctx, next_ctx);
1195 spin_unlock(&next_ctx->lock);
1196 spin_unlock(&ctx->lock);
1201 __perf_event_sched_out(ctx, cpuctx);
1202 cpuctx->task_ctx = NULL;
1207 * Called with IRQs disabled
1209 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1211 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1213 if (!cpuctx->task_ctx)
1216 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1219 __perf_event_sched_out(ctx, cpuctx);
1220 cpuctx->task_ctx = NULL;
1224 * Called with IRQs disabled
1226 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1228 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1232 __perf_event_sched_in(struct perf_event_context *ctx,
1233 struct perf_cpu_context *cpuctx, int cpu)
1235 struct perf_event *event;
1238 spin_lock(&ctx->lock);
1240 if (likely(!ctx->nr_events))
1243 ctx->timestamp = perf_clock();
1248 * First go through the list and put on any pinned groups
1249 * in order to give them the best chance of going on.
1251 list_for_each_entry(event, &ctx->group_list, group_entry) {
1252 if (event->state <= PERF_EVENT_STATE_OFF ||
1253 !event->attr.pinned)
1255 if (event->cpu != -1 && event->cpu != cpu)
1258 if (group_can_go_on(event, cpuctx, 1))
1259 group_sched_in(event, cpuctx, ctx, cpu);
1262 * If this pinned group hasn't been scheduled,
1263 * put it in error state.
1265 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1266 update_group_times(event);
1267 event->state = PERF_EVENT_STATE_ERROR;
1271 list_for_each_entry(event, &ctx->group_list, group_entry) {
1273 * Ignore events in OFF or ERROR state, and
1274 * ignore pinned events since we did them already.
1276 if (event->state <= PERF_EVENT_STATE_OFF ||
1281 * Listen to the 'cpu' scheduling filter constraint
1284 if (event->cpu != -1 && event->cpu != cpu)
1287 if (group_can_go_on(event, cpuctx, can_add_hw))
1288 if (group_sched_in(event, cpuctx, ctx, cpu))
1293 spin_unlock(&ctx->lock);
1297 * Called from scheduler to add the events of the current task
1298 * with interrupts disabled.
1300 * We restore the event value and then enable it.
1302 * This does not protect us against NMI, but enable()
1303 * sets the enabled bit in the control field of event _before_
1304 * accessing the event control register. If a NMI hits, then it will
1305 * keep the event running.
1307 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1309 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1310 struct perf_event_context *ctx = task->perf_event_ctxp;
1314 if (cpuctx->task_ctx == ctx)
1316 __perf_event_sched_in(ctx, cpuctx, cpu);
1317 cpuctx->task_ctx = ctx;
1320 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1322 struct perf_event_context *ctx = &cpuctx->ctx;
1324 __perf_event_sched_in(ctx, cpuctx, cpu);
1327 #define MAX_INTERRUPTS (~0ULL)
1329 static void perf_log_throttle(struct perf_event *event, int enable);
1331 static void perf_adjust_period(struct perf_event *event, u64 events)
1333 struct hw_perf_event *hwc = &event->hw;
1334 u64 period, sample_period;
1337 events *= hwc->sample_period;
1338 period = div64_u64(events, event->attr.sample_freq);
1340 delta = (s64)(period - hwc->sample_period);
1341 delta = (delta + 7) / 8; /* low pass filter */
1343 sample_period = hwc->sample_period + delta;
1348 hwc->sample_period = sample_period;
1351 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1353 struct perf_event *event;
1354 struct hw_perf_event *hwc;
1355 u64 interrupts, freq;
1357 spin_lock(&ctx->lock);
1358 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1359 if (event->state != PERF_EVENT_STATE_ACTIVE)
1364 interrupts = hwc->interrupts;
1365 hwc->interrupts = 0;
1368 * unthrottle events on the tick
1370 if (interrupts == MAX_INTERRUPTS) {
1371 perf_log_throttle(event, 1);
1372 event->pmu->unthrottle(event);
1373 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1376 if (!event->attr.freq || !event->attr.sample_freq)
1380 * if the specified freq < HZ then we need to skip ticks
1382 if (event->attr.sample_freq < HZ) {
1383 freq = event->attr.sample_freq;
1385 hwc->freq_count += freq;
1386 hwc->freq_interrupts += interrupts;
1388 if (hwc->freq_count < HZ)
1391 interrupts = hwc->freq_interrupts;
1392 hwc->freq_interrupts = 0;
1393 hwc->freq_count -= HZ;
1397 perf_adjust_period(event, freq * interrupts);
1400 * In order to avoid being stalled by an (accidental) huge
1401 * sample period, force reset the sample period if we didn't
1402 * get any events in this freq period.
1406 event->pmu->disable(event);
1407 atomic64_set(&hwc->period_left, 0);
1408 event->pmu->enable(event);
1412 spin_unlock(&ctx->lock);
1416 * Round-robin a context's events:
1418 static void rotate_ctx(struct perf_event_context *ctx)
1420 struct perf_event *event;
1422 if (!ctx->nr_events)
1425 spin_lock(&ctx->lock);
1427 * Rotate the first entry last (works just fine for group events too):
1430 list_for_each_entry(event, &ctx->group_list, group_entry) {
1431 list_move_tail(&event->group_entry, &ctx->group_list);
1436 spin_unlock(&ctx->lock);
1439 void perf_event_task_tick(struct task_struct *curr, int cpu)
1441 struct perf_cpu_context *cpuctx;
1442 struct perf_event_context *ctx;
1444 if (!atomic_read(&nr_events))
1447 cpuctx = &per_cpu(perf_cpu_context, cpu);
1448 ctx = curr->perf_event_ctxp;
1450 perf_ctx_adjust_freq(&cpuctx->ctx);
1452 perf_ctx_adjust_freq(ctx);
1454 perf_event_cpu_sched_out(cpuctx);
1456 __perf_event_task_sched_out(ctx);
1458 rotate_ctx(&cpuctx->ctx);
1462 perf_event_cpu_sched_in(cpuctx, cpu);
1464 perf_event_task_sched_in(curr, cpu);
1468 * Enable all of a task's events that have been marked enable-on-exec.
1469 * This expects task == current.
1471 static void perf_event_enable_on_exec(struct task_struct *task)
1473 struct perf_event_context *ctx;
1474 struct perf_event *event;
1475 unsigned long flags;
1478 local_irq_save(flags);
1479 ctx = task->perf_event_ctxp;
1480 if (!ctx || !ctx->nr_events)
1483 __perf_event_task_sched_out(ctx);
1485 spin_lock(&ctx->lock);
1487 list_for_each_entry(event, &ctx->group_list, group_entry) {
1488 if (!event->attr.enable_on_exec)
1490 event->attr.enable_on_exec = 0;
1491 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1493 __perf_event_mark_enabled(event, ctx);
1498 * Unclone this context if we enabled any event.
1503 spin_unlock(&ctx->lock);
1505 perf_event_task_sched_in(task, smp_processor_id());
1507 local_irq_restore(flags);
1511 * Cross CPU call to read the hardware event
1513 static void __perf_event_read(void *info)
1515 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1516 struct perf_event *event = info;
1517 struct perf_event_context *ctx = event->ctx;
1520 * If this is a task context, we need to check whether it is
1521 * the current task context of this cpu. If not it has been
1522 * scheduled out before the smp call arrived. In that case
1523 * event->count would have been updated to a recent sample
1524 * when the event was scheduled out.
1526 if (ctx->task && cpuctx->task_ctx != ctx)
1529 spin_lock(&ctx->lock);
1530 update_context_time(ctx);
1531 update_event_times(event);
1532 spin_unlock(&ctx->lock);
1534 event->pmu->read(event);
1537 static u64 perf_event_read(struct perf_event *event)
1540 * If event is enabled and currently active on a CPU, update the
1541 * value in the event structure:
1543 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1544 smp_call_function_single(event->oncpu,
1545 __perf_event_read, event, 1);
1546 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1547 struct perf_event_context *ctx = event->ctx;
1548 unsigned long flags;
1550 spin_lock_irqsave(&ctx->lock, flags);
1551 update_context_time(ctx);
1552 update_event_times(event);
1553 spin_unlock_irqrestore(&ctx->lock, flags);
1556 return atomic64_read(&event->count);
1560 * Initialize the perf_event context in a task_struct:
1563 __perf_event_init_context(struct perf_event_context *ctx,
1564 struct task_struct *task)
1566 memset(ctx, 0, sizeof(*ctx));
1567 spin_lock_init(&ctx->lock);
1568 mutex_init(&ctx->mutex);
1569 INIT_LIST_HEAD(&ctx->group_list);
1570 INIT_LIST_HEAD(&ctx->event_list);
1571 atomic_set(&ctx->refcount, 1);
1575 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1577 struct perf_event_context *ctx;
1578 struct perf_cpu_context *cpuctx;
1579 struct task_struct *task;
1580 unsigned long flags;
1584 * If cpu is not a wildcard then this is a percpu event:
1587 /* Must be root to operate on a CPU event: */
1588 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1589 return ERR_PTR(-EACCES);
1591 if (cpu < 0 || cpu > num_possible_cpus())
1592 return ERR_PTR(-EINVAL);
1595 * We could be clever and allow to attach a event to an
1596 * offline CPU and activate it when the CPU comes up, but
1599 if (!cpu_isset(cpu, cpu_online_map))
1600 return ERR_PTR(-ENODEV);
1602 cpuctx = &per_cpu(perf_cpu_context, cpu);
1613 task = find_task_by_vpid(pid);
1615 get_task_struct(task);
1619 return ERR_PTR(-ESRCH);
1622 * Can't attach events to a dying task.
1625 if (task->flags & PF_EXITING)
1628 /* Reuse ptrace permission checks for now. */
1630 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1634 ctx = perf_lock_task_context(task, &flags);
1637 spin_unlock_irqrestore(&ctx->lock, flags);
1641 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1645 __perf_event_init_context(ctx, task);
1647 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1649 * We raced with some other task; use
1650 * the context they set.
1655 get_task_struct(task);
1658 put_task_struct(task);
1662 put_task_struct(task);
1663 return ERR_PTR(err);
1666 static void perf_event_free_filter(struct perf_event *event);
1668 static void free_event_rcu(struct rcu_head *head)
1670 struct perf_event *event;
1672 event = container_of(head, struct perf_event, rcu_head);
1674 put_pid_ns(event->ns);
1675 perf_event_free_filter(event);
1679 static void perf_pending_sync(struct perf_event *event);
1681 static void free_event(struct perf_event *event)
1683 perf_pending_sync(event);
1685 if (!event->parent) {
1686 atomic_dec(&nr_events);
1687 if (event->attr.mmap)
1688 atomic_dec(&nr_mmap_events);
1689 if (event->attr.comm)
1690 atomic_dec(&nr_comm_events);
1691 if (event->attr.task)
1692 atomic_dec(&nr_task_events);
1695 if (event->output) {
1696 fput(event->output->filp);
1697 event->output = NULL;
1701 event->destroy(event);
1703 put_ctx(event->ctx);
1704 call_rcu(&event->rcu_head, free_event_rcu);
1707 int perf_event_release_kernel(struct perf_event *event)
1709 struct perf_event_context *ctx = event->ctx;
1711 WARN_ON_ONCE(ctx->parent_ctx);
1712 mutex_lock(&ctx->mutex);
1713 perf_event_remove_from_context(event);
1714 mutex_unlock(&ctx->mutex);
1716 mutex_lock(&event->owner->perf_event_mutex);
1717 list_del_init(&event->owner_entry);
1718 mutex_unlock(&event->owner->perf_event_mutex);
1719 put_task_struct(event->owner);
1725 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1728 * Called when the last reference to the file is gone.
1730 static int perf_release(struct inode *inode, struct file *file)
1732 struct perf_event *event = file->private_data;
1734 file->private_data = NULL;
1736 return perf_event_release_kernel(event);
1739 static int perf_event_read_size(struct perf_event *event)
1741 int entry = sizeof(u64); /* value */
1745 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1746 size += sizeof(u64);
1748 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1749 size += sizeof(u64);
1751 if (event->attr.read_format & PERF_FORMAT_ID)
1752 entry += sizeof(u64);
1754 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1755 nr += event->group_leader->nr_siblings;
1756 size += sizeof(u64);
1764 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1766 struct perf_event *child;
1772 mutex_lock(&event->child_mutex);
1773 total += perf_event_read(event);
1774 *enabled += event->total_time_enabled +
1775 atomic64_read(&event->child_total_time_enabled);
1776 *running += event->total_time_running +
1777 atomic64_read(&event->child_total_time_running);
1779 list_for_each_entry(child, &event->child_list, child_list) {
1780 total += perf_event_read(child);
1781 *enabled += child->total_time_enabled;
1782 *running += child->total_time_running;
1784 mutex_unlock(&event->child_mutex);
1788 EXPORT_SYMBOL_GPL(perf_event_read_value);
1790 static int perf_event_read_group(struct perf_event *event,
1791 u64 read_format, char __user *buf)
1793 struct perf_event *leader = event->group_leader, *sub;
1794 int n = 0, size = 0, ret = -EFAULT;
1795 struct perf_event_context *ctx = leader->ctx;
1797 u64 count, enabled, running;
1799 mutex_lock(&ctx->mutex);
1800 count = perf_event_read_value(leader, &enabled, &running);
1802 values[n++] = 1 + leader->nr_siblings;
1803 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1804 values[n++] = enabled;
1805 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1806 values[n++] = running;
1807 values[n++] = count;
1808 if (read_format & PERF_FORMAT_ID)
1809 values[n++] = primary_event_id(leader);
1811 size = n * sizeof(u64);
1813 if (copy_to_user(buf, values, size))
1818 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1821 values[n++] = perf_event_read_value(sub, &enabled, &running);
1822 if (read_format & PERF_FORMAT_ID)
1823 values[n++] = primary_event_id(sub);
1825 size = n * sizeof(u64);
1827 if (copy_to_user(buf + size, values, size)) {
1835 mutex_unlock(&ctx->mutex);
1840 static int perf_event_read_one(struct perf_event *event,
1841 u64 read_format, char __user *buf)
1843 u64 enabled, running;
1847 values[n++] = perf_event_read_value(event, &enabled, &running);
1848 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1849 values[n++] = enabled;
1850 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1851 values[n++] = running;
1852 if (read_format & PERF_FORMAT_ID)
1853 values[n++] = primary_event_id(event);
1855 if (copy_to_user(buf, values, n * sizeof(u64)))
1858 return n * sizeof(u64);
1862 * Read the performance event - simple non blocking version for now
1865 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1867 u64 read_format = event->attr.read_format;
1871 * Return end-of-file for a read on a event that is in
1872 * error state (i.e. because it was pinned but it couldn't be
1873 * scheduled on to the CPU at some point).
1875 if (event->state == PERF_EVENT_STATE_ERROR)
1878 if (count < perf_event_read_size(event))
1881 WARN_ON_ONCE(event->ctx->parent_ctx);
1882 if (read_format & PERF_FORMAT_GROUP)
1883 ret = perf_event_read_group(event, read_format, buf);
1885 ret = perf_event_read_one(event, read_format, buf);
1891 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1893 struct perf_event *event = file->private_data;
1895 return perf_read_hw(event, buf, count);
1898 static unsigned int perf_poll(struct file *file, poll_table *wait)
1900 struct perf_event *event = file->private_data;
1901 struct perf_mmap_data *data;
1902 unsigned int events = POLL_HUP;
1905 data = rcu_dereference(event->data);
1907 events = atomic_xchg(&data->poll, 0);
1910 poll_wait(file, &event->waitq, wait);
1915 static void perf_event_reset(struct perf_event *event)
1917 (void)perf_event_read(event);
1918 atomic64_set(&event->count, 0);
1919 perf_event_update_userpage(event);
1923 * Holding the top-level event's child_mutex means that any
1924 * descendant process that has inherited this event will block
1925 * in sync_child_event if it goes to exit, thus satisfying the
1926 * task existence requirements of perf_event_enable/disable.
1928 static void perf_event_for_each_child(struct perf_event *event,
1929 void (*func)(struct perf_event *))
1931 struct perf_event *child;
1933 WARN_ON_ONCE(event->ctx->parent_ctx);
1934 mutex_lock(&event->child_mutex);
1936 list_for_each_entry(child, &event->child_list, child_list)
1938 mutex_unlock(&event->child_mutex);
1941 static void perf_event_for_each(struct perf_event *event,
1942 void (*func)(struct perf_event *))
1944 struct perf_event_context *ctx = event->ctx;
1945 struct perf_event *sibling;
1947 WARN_ON_ONCE(ctx->parent_ctx);
1948 mutex_lock(&ctx->mutex);
1949 event = event->group_leader;
1951 perf_event_for_each_child(event, func);
1953 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1954 perf_event_for_each_child(event, func);
1955 mutex_unlock(&ctx->mutex);
1958 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1960 struct perf_event_context *ctx = event->ctx;
1965 if (!event->attr.sample_period)
1968 size = copy_from_user(&value, arg, sizeof(value));
1969 if (size != sizeof(value))
1975 spin_lock_irq(&ctx->lock);
1976 if (event->attr.freq) {
1977 if (value > sysctl_perf_event_sample_rate) {
1982 event->attr.sample_freq = value;
1984 event->attr.sample_period = value;
1985 event->hw.sample_period = value;
1988 spin_unlock_irq(&ctx->lock);
1993 static int perf_event_set_output(struct perf_event *event, int output_fd);
1994 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
1996 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1998 struct perf_event *event = file->private_data;
1999 void (*func)(struct perf_event *);
2003 case PERF_EVENT_IOC_ENABLE:
2004 func = perf_event_enable;
2006 case PERF_EVENT_IOC_DISABLE:
2007 func = perf_event_disable;
2009 case PERF_EVENT_IOC_RESET:
2010 func = perf_event_reset;
2013 case PERF_EVENT_IOC_REFRESH:
2014 return perf_event_refresh(event, arg);
2016 case PERF_EVENT_IOC_PERIOD:
2017 return perf_event_period(event, (u64 __user *)arg);
2019 case PERF_EVENT_IOC_SET_OUTPUT:
2020 return perf_event_set_output(event, arg);
2022 case PERF_EVENT_IOC_SET_FILTER:
2023 return perf_event_set_filter(event, (void __user *)arg);
2029 if (flags & PERF_IOC_FLAG_GROUP)
2030 perf_event_for_each(event, func);
2032 perf_event_for_each_child(event, func);
2037 int perf_event_task_enable(void)
2039 struct perf_event *event;
2041 mutex_lock(¤t->perf_event_mutex);
2042 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2043 perf_event_for_each_child(event, perf_event_enable);
2044 mutex_unlock(¤t->perf_event_mutex);
2049 int perf_event_task_disable(void)
2051 struct perf_event *event;
2053 mutex_lock(¤t->perf_event_mutex);
2054 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2055 perf_event_for_each_child(event, perf_event_disable);
2056 mutex_unlock(¤t->perf_event_mutex);
2061 #ifndef PERF_EVENT_INDEX_OFFSET
2062 # define PERF_EVENT_INDEX_OFFSET 0
2065 static int perf_event_index(struct perf_event *event)
2067 if (event->state != PERF_EVENT_STATE_ACTIVE)
2070 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2074 * Callers need to ensure there can be no nesting of this function, otherwise
2075 * the seqlock logic goes bad. We can not serialize this because the arch
2076 * code calls this from NMI context.
2078 void perf_event_update_userpage(struct perf_event *event)
2080 struct perf_event_mmap_page *userpg;
2081 struct perf_mmap_data *data;
2084 data = rcu_dereference(event->data);
2088 userpg = data->user_page;
2091 * Disable preemption so as to not let the corresponding user-space
2092 * spin too long if we get preempted.
2097 userpg->index = perf_event_index(event);
2098 userpg->offset = atomic64_read(&event->count);
2099 if (event->state == PERF_EVENT_STATE_ACTIVE)
2100 userpg->offset -= atomic64_read(&event->hw.prev_count);
2102 userpg->time_enabled = event->total_time_enabled +
2103 atomic64_read(&event->child_total_time_enabled);
2105 userpg->time_running = event->total_time_running +
2106 atomic64_read(&event->child_total_time_running);
2115 static unsigned long perf_data_size(struct perf_mmap_data *data)
2117 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2120 #ifndef CONFIG_PERF_USE_VMALLOC
2123 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2126 static struct page *
2127 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2129 if (pgoff > data->nr_pages)
2133 return virt_to_page(data->user_page);
2135 return virt_to_page(data->data_pages[pgoff - 1]);
2138 static struct perf_mmap_data *
2139 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2141 struct perf_mmap_data *data;
2145 WARN_ON(atomic_read(&event->mmap_count));
2147 size = sizeof(struct perf_mmap_data);
2148 size += nr_pages * sizeof(void *);
2150 data = kzalloc(size, GFP_KERNEL);
2154 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2155 if (!data->user_page)
2156 goto fail_user_page;
2158 for (i = 0; i < nr_pages; i++) {
2159 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2160 if (!data->data_pages[i])
2161 goto fail_data_pages;
2164 data->data_order = 0;
2165 data->nr_pages = nr_pages;
2170 for (i--; i >= 0; i--)
2171 free_page((unsigned long)data->data_pages[i]);
2173 free_page((unsigned long)data->user_page);
2182 static void perf_mmap_free_page(unsigned long addr)
2184 struct page *page = virt_to_page((void *)addr);
2186 page->mapping = NULL;
2190 static void perf_mmap_data_free(struct perf_mmap_data *data)
2194 perf_mmap_free_page((unsigned long)data->user_page);
2195 for (i = 0; i < data->nr_pages; i++)
2196 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2202 * Back perf_mmap() with vmalloc memory.
2204 * Required for architectures that have d-cache aliasing issues.
2207 static struct page *
2208 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2210 if (pgoff > (1UL << data->data_order))
2213 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2216 static void perf_mmap_unmark_page(void *addr)
2218 struct page *page = vmalloc_to_page(addr);
2220 page->mapping = NULL;
2223 static void perf_mmap_data_free_work(struct work_struct *work)
2225 struct perf_mmap_data *data;
2229 data = container_of(work, struct perf_mmap_data, work);
2230 nr = 1 << data->data_order;
2232 base = data->user_page;
2233 for (i = 0; i < nr + 1; i++)
2234 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2239 static void perf_mmap_data_free(struct perf_mmap_data *data)
2241 schedule_work(&data->work);
2244 static struct perf_mmap_data *
2245 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2247 struct perf_mmap_data *data;
2251 WARN_ON(atomic_read(&event->mmap_count));
2253 size = sizeof(struct perf_mmap_data);
2254 size += sizeof(void *);
2256 data = kzalloc(size, GFP_KERNEL);
2260 INIT_WORK(&data->work, perf_mmap_data_free_work);
2262 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2266 data->user_page = all_buf;
2267 data->data_pages[0] = all_buf + PAGE_SIZE;
2268 data->data_order = ilog2(nr_pages);
2282 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2284 struct perf_event *event = vma->vm_file->private_data;
2285 struct perf_mmap_data *data;
2286 int ret = VM_FAULT_SIGBUS;
2288 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2289 if (vmf->pgoff == 0)
2295 data = rcu_dereference(event->data);
2299 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2302 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2306 get_page(vmf->page);
2307 vmf->page->mapping = vma->vm_file->f_mapping;
2308 vmf->page->index = vmf->pgoff;
2318 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2320 long max_size = perf_data_size(data);
2322 atomic_set(&data->lock, -1);
2324 if (event->attr.watermark) {
2325 data->watermark = min_t(long, max_size,
2326 event->attr.wakeup_watermark);
2329 if (!data->watermark)
2330 data->watermark = max_size / 2;
2333 rcu_assign_pointer(event->data, data);
2336 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2338 struct perf_mmap_data *data;
2340 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2341 perf_mmap_data_free(data);
2345 static void perf_mmap_data_release(struct perf_event *event)
2347 struct perf_mmap_data *data = event->data;
2349 WARN_ON(atomic_read(&event->mmap_count));
2351 rcu_assign_pointer(event->data, NULL);
2352 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2355 static void perf_mmap_open(struct vm_area_struct *vma)
2357 struct perf_event *event = vma->vm_file->private_data;
2359 atomic_inc(&event->mmap_count);
2362 static void perf_mmap_close(struct vm_area_struct *vma)
2364 struct perf_event *event = vma->vm_file->private_data;
2366 WARN_ON_ONCE(event->ctx->parent_ctx);
2367 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2368 unsigned long size = perf_data_size(event->data);
2369 struct user_struct *user = current_user();
2371 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2372 vma->vm_mm->locked_vm -= event->data->nr_locked;
2373 perf_mmap_data_release(event);
2374 mutex_unlock(&event->mmap_mutex);
2378 static const struct vm_operations_struct perf_mmap_vmops = {
2379 .open = perf_mmap_open,
2380 .close = perf_mmap_close,
2381 .fault = perf_mmap_fault,
2382 .page_mkwrite = perf_mmap_fault,
2385 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2387 struct perf_event *event = file->private_data;
2388 unsigned long user_locked, user_lock_limit;
2389 struct user_struct *user = current_user();
2390 unsigned long locked, lock_limit;
2391 struct perf_mmap_data *data;
2392 unsigned long vma_size;
2393 unsigned long nr_pages;
2394 long user_extra, extra;
2397 if (!(vma->vm_flags & VM_SHARED))
2400 vma_size = vma->vm_end - vma->vm_start;
2401 nr_pages = (vma_size / PAGE_SIZE) - 1;
2404 * If we have data pages ensure they're a power-of-two number, so we
2405 * can do bitmasks instead of modulo.
2407 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2410 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2413 if (vma->vm_pgoff != 0)
2416 WARN_ON_ONCE(event->ctx->parent_ctx);
2417 mutex_lock(&event->mmap_mutex);
2418 if (event->output) {
2423 if (atomic_inc_not_zero(&event->mmap_count)) {
2424 if (nr_pages != event->data->nr_pages)
2429 user_extra = nr_pages + 1;
2430 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2433 * Increase the limit linearly with more CPUs:
2435 user_lock_limit *= num_online_cpus();
2437 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2440 if (user_locked > user_lock_limit)
2441 extra = user_locked - user_lock_limit;
2443 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2444 lock_limit >>= PAGE_SHIFT;
2445 locked = vma->vm_mm->locked_vm + extra;
2447 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2448 !capable(CAP_IPC_LOCK)) {
2453 WARN_ON(event->data);
2455 data = perf_mmap_data_alloc(event, nr_pages);
2461 perf_mmap_data_init(event, data);
2463 atomic_set(&event->mmap_count, 1);
2464 atomic_long_add(user_extra, &user->locked_vm);
2465 vma->vm_mm->locked_vm += extra;
2466 event->data->nr_locked = extra;
2467 if (vma->vm_flags & VM_WRITE)
2468 event->data->writable = 1;
2471 mutex_unlock(&event->mmap_mutex);
2473 vma->vm_flags |= VM_RESERVED;
2474 vma->vm_ops = &perf_mmap_vmops;
2479 static int perf_fasync(int fd, struct file *filp, int on)
2481 struct inode *inode = filp->f_path.dentry->d_inode;
2482 struct perf_event *event = filp->private_data;
2485 mutex_lock(&inode->i_mutex);
2486 retval = fasync_helper(fd, filp, on, &event->fasync);
2487 mutex_unlock(&inode->i_mutex);
2495 static const struct file_operations perf_fops = {
2496 .release = perf_release,
2499 .unlocked_ioctl = perf_ioctl,
2500 .compat_ioctl = perf_ioctl,
2502 .fasync = perf_fasync,
2508 * If there's data, ensure we set the poll() state and publish everything
2509 * to user-space before waking everybody up.
2512 void perf_event_wakeup(struct perf_event *event)
2514 wake_up_all(&event->waitq);
2516 if (event->pending_kill) {
2517 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2518 event->pending_kill = 0;
2525 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2527 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2528 * single linked list and use cmpxchg() to add entries lockless.
2531 static void perf_pending_event(struct perf_pending_entry *entry)
2533 struct perf_event *event = container_of(entry,
2534 struct perf_event, pending);
2536 if (event->pending_disable) {
2537 event->pending_disable = 0;
2538 __perf_event_disable(event);
2541 if (event->pending_wakeup) {
2542 event->pending_wakeup = 0;
2543 perf_event_wakeup(event);
2547 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2549 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2553 static void perf_pending_queue(struct perf_pending_entry *entry,
2554 void (*func)(struct perf_pending_entry *))
2556 struct perf_pending_entry **head;
2558 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2563 head = &get_cpu_var(perf_pending_head);
2566 entry->next = *head;
2567 } while (cmpxchg(head, entry->next, entry) != entry->next);
2569 set_perf_event_pending();
2571 put_cpu_var(perf_pending_head);
2574 static int __perf_pending_run(void)
2576 struct perf_pending_entry *list;
2579 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2580 while (list != PENDING_TAIL) {
2581 void (*func)(struct perf_pending_entry *);
2582 struct perf_pending_entry *entry = list;
2589 * Ensure we observe the unqueue before we issue the wakeup,
2590 * so that we won't be waiting forever.
2591 * -- see perf_not_pending().
2602 static inline int perf_not_pending(struct perf_event *event)
2605 * If we flush on whatever cpu we run, there is a chance we don't
2609 __perf_pending_run();
2613 * Ensure we see the proper queue state before going to sleep
2614 * so that we do not miss the wakeup. -- see perf_pending_handle()
2617 return event->pending.next == NULL;
2620 static void perf_pending_sync(struct perf_event *event)
2622 wait_event(event->waitq, perf_not_pending(event));
2625 void perf_event_do_pending(void)
2627 __perf_pending_run();
2631 * Callchain support -- arch specific
2634 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2642 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2643 unsigned long offset, unsigned long head)
2647 if (!data->writable)
2650 mask = perf_data_size(data) - 1;
2652 offset = (offset - tail) & mask;
2653 head = (head - tail) & mask;
2655 if ((int)(head - offset) < 0)
2661 static void perf_output_wakeup(struct perf_output_handle *handle)
2663 atomic_set(&handle->data->poll, POLL_IN);
2666 handle->event->pending_wakeup = 1;
2667 perf_pending_queue(&handle->event->pending,
2668 perf_pending_event);
2670 perf_event_wakeup(handle->event);
2674 * Curious locking construct.
2676 * We need to ensure a later event_id doesn't publish a head when a former
2677 * event_id isn't done writing. However since we need to deal with NMIs we
2678 * cannot fully serialize things.
2680 * What we do is serialize between CPUs so we only have to deal with NMI
2681 * nesting on a single CPU.
2683 * We only publish the head (and generate a wakeup) when the outer-most
2684 * event_id completes.
2686 static void perf_output_lock(struct perf_output_handle *handle)
2688 struct perf_mmap_data *data = handle->data;
2689 int cur, cpu = get_cpu();
2694 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2706 static void perf_output_unlock(struct perf_output_handle *handle)
2708 struct perf_mmap_data *data = handle->data;
2712 data->done_head = data->head;
2714 if (!handle->locked)
2719 * The xchg implies a full barrier that ensures all writes are done
2720 * before we publish the new head, matched by a rmb() in userspace when
2721 * reading this position.
2723 while ((head = atomic_long_xchg(&data->done_head, 0)))
2724 data->user_page->data_head = head;
2727 * NMI can happen here, which means we can miss a done_head update.
2730 cpu = atomic_xchg(&data->lock, -1);
2731 WARN_ON_ONCE(cpu != smp_processor_id());
2734 * Therefore we have to validate we did not indeed do so.
2736 if (unlikely(atomic_long_read(&data->done_head))) {
2738 * Since we had it locked, we can lock it again.
2740 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2746 if (atomic_xchg(&data->wakeup, 0))
2747 perf_output_wakeup(handle);
2752 void perf_output_copy(struct perf_output_handle *handle,
2753 const void *buf, unsigned int len)
2755 unsigned int pages_mask;
2756 unsigned long offset;
2760 offset = handle->offset;
2761 pages_mask = handle->data->nr_pages - 1;
2762 pages = handle->data->data_pages;
2765 unsigned long page_offset;
2766 unsigned long page_size;
2769 nr = (offset >> PAGE_SHIFT) & pages_mask;
2770 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2771 page_offset = offset & (page_size - 1);
2772 size = min_t(unsigned int, page_size - page_offset, len);
2774 memcpy(pages[nr] + page_offset, buf, size);
2781 handle->offset = offset;
2784 * Check we didn't copy past our reservation window, taking the
2785 * possible unsigned int wrap into account.
2787 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2790 int perf_output_begin(struct perf_output_handle *handle,
2791 struct perf_event *event, unsigned int size,
2792 int nmi, int sample)
2794 struct perf_event *output_event;
2795 struct perf_mmap_data *data;
2796 unsigned long tail, offset, head;
2799 struct perf_event_header header;
2806 * For inherited events we send all the output towards the parent.
2809 event = event->parent;
2811 output_event = rcu_dereference(event->output);
2813 event = output_event;
2815 data = rcu_dereference(event->data);
2819 handle->data = data;
2820 handle->event = event;
2822 handle->sample = sample;
2824 if (!data->nr_pages)
2827 have_lost = atomic_read(&data->lost);
2829 size += sizeof(lost_event);
2831 perf_output_lock(handle);
2835 * Userspace could choose to issue a mb() before updating the
2836 * tail pointer. So that all reads will be completed before the
2839 tail = ACCESS_ONCE(data->user_page->data_tail);
2841 offset = head = atomic_long_read(&data->head);
2843 if (unlikely(!perf_output_space(data, tail, offset, head)))
2845 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2847 handle->offset = offset;
2848 handle->head = head;
2850 if (head - tail > data->watermark)
2851 atomic_set(&data->wakeup, 1);
2854 lost_event.header.type = PERF_RECORD_LOST;
2855 lost_event.header.misc = 0;
2856 lost_event.header.size = sizeof(lost_event);
2857 lost_event.id = event->id;
2858 lost_event.lost = atomic_xchg(&data->lost, 0);
2860 perf_output_put(handle, lost_event);
2866 atomic_inc(&data->lost);
2867 perf_output_unlock(handle);
2874 void perf_output_end(struct perf_output_handle *handle)
2876 struct perf_event *event = handle->event;
2877 struct perf_mmap_data *data = handle->data;
2879 int wakeup_events = event->attr.wakeup_events;
2881 if (handle->sample && wakeup_events) {
2882 int events = atomic_inc_return(&data->events);
2883 if (events >= wakeup_events) {
2884 atomic_sub(wakeup_events, &data->events);
2885 atomic_set(&data->wakeup, 1);
2889 perf_output_unlock(handle);
2893 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2896 * only top level events have the pid namespace they were created in
2899 event = event->parent;
2901 return task_tgid_nr_ns(p, event->ns);
2904 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2907 * only top level events have the pid namespace they were created in
2910 event = event->parent;
2912 return task_pid_nr_ns(p, event->ns);
2915 static void perf_output_read_one(struct perf_output_handle *handle,
2916 struct perf_event *event)
2918 u64 read_format = event->attr.read_format;
2922 values[n++] = atomic64_read(&event->count);
2923 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2924 values[n++] = event->total_time_enabled +
2925 atomic64_read(&event->child_total_time_enabled);
2927 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2928 values[n++] = event->total_time_running +
2929 atomic64_read(&event->child_total_time_running);
2931 if (read_format & PERF_FORMAT_ID)
2932 values[n++] = primary_event_id(event);
2934 perf_output_copy(handle, values, n * sizeof(u64));
2938 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2940 static void perf_output_read_group(struct perf_output_handle *handle,
2941 struct perf_event *event)
2943 struct perf_event *leader = event->group_leader, *sub;
2944 u64 read_format = event->attr.read_format;
2948 values[n++] = 1 + leader->nr_siblings;
2950 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2951 values[n++] = leader->total_time_enabled;
2953 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2954 values[n++] = leader->total_time_running;
2956 if (leader != event)
2957 leader->pmu->read(leader);
2959 values[n++] = atomic64_read(&leader->count);
2960 if (read_format & PERF_FORMAT_ID)
2961 values[n++] = primary_event_id(leader);
2963 perf_output_copy(handle, values, n * sizeof(u64));
2965 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2969 sub->pmu->read(sub);
2971 values[n++] = atomic64_read(&sub->count);
2972 if (read_format & PERF_FORMAT_ID)
2973 values[n++] = primary_event_id(sub);
2975 perf_output_copy(handle, values, n * sizeof(u64));
2979 static void perf_output_read(struct perf_output_handle *handle,
2980 struct perf_event *event)
2982 if (event->attr.read_format & PERF_FORMAT_GROUP)
2983 perf_output_read_group(handle, event);
2985 perf_output_read_one(handle, event);
2988 void perf_output_sample(struct perf_output_handle *handle,
2989 struct perf_event_header *header,
2990 struct perf_sample_data *data,
2991 struct perf_event *event)
2993 u64 sample_type = data->type;
2995 perf_output_put(handle, *header);
2997 if (sample_type & PERF_SAMPLE_IP)
2998 perf_output_put(handle, data->ip);
3000 if (sample_type & PERF_SAMPLE_TID)
3001 perf_output_put(handle, data->tid_entry);
3003 if (sample_type & PERF_SAMPLE_TIME)
3004 perf_output_put(handle, data->time);
3006 if (sample_type & PERF_SAMPLE_ADDR)
3007 perf_output_put(handle, data->addr);
3009 if (sample_type & PERF_SAMPLE_ID)
3010 perf_output_put(handle, data->id);
3012 if (sample_type & PERF_SAMPLE_STREAM_ID)
3013 perf_output_put(handle, data->stream_id);
3015 if (sample_type & PERF_SAMPLE_CPU)
3016 perf_output_put(handle, data->cpu_entry);
3018 if (sample_type & PERF_SAMPLE_PERIOD)
3019 perf_output_put(handle, data->period);
3021 if (sample_type & PERF_SAMPLE_READ)
3022 perf_output_read(handle, event);
3024 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3025 if (data->callchain) {
3028 if (data->callchain)
3029 size += data->callchain->nr;
3031 size *= sizeof(u64);
3033 perf_output_copy(handle, data->callchain, size);
3036 perf_output_put(handle, nr);
3040 if (sample_type & PERF_SAMPLE_RAW) {
3042 perf_output_put(handle, data->raw->size);
3043 perf_output_copy(handle, data->raw->data,
3050 .size = sizeof(u32),
3053 perf_output_put(handle, raw);
3058 void perf_prepare_sample(struct perf_event_header *header,
3059 struct perf_sample_data *data,
3060 struct perf_event *event,
3061 struct pt_regs *regs)
3063 u64 sample_type = event->attr.sample_type;
3065 data->type = sample_type;
3067 header->type = PERF_RECORD_SAMPLE;
3068 header->size = sizeof(*header);
3071 header->misc |= perf_misc_flags(regs);
3073 if (sample_type & PERF_SAMPLE_IP) {
3074 data->ip = perf_instruction_pointer(regs);
3076 header->size += sizeof(data->ip);
3079 if (sample_type & PERF_SAMPLE_TID) {
3080 /* namespace issues */
3081 data->tid_entry.pid = perf_event_pid(event, current);
3082 data->tid_entry.tid = perf_event_tid(event, current);
3084 header->size += sizeof(data->tid_entry);
3087 if (sample_type & PERF_SAMPLE_TIME) {
3088 data->time = perf_clock();
3090 header->size += sizeof(data->time);
3093 if (sample_type & PERF_SAMPLE_ADDR)
3094 header->size += sizeof(data->addr);
3096 if (sample_type & PERF_SAMPLE_ID) {
3097 data->id = primary_event_id(event);
3099 header->size += sizeof(data->id);
3102 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3103 data->stream_id = event->id;
3105 header->size += sizeof(data->stream_id);
3108 if (sample_type & PERF_SAMPLE_CPU) {
3109 data->cpu_entry.cpu = raw_smp_processor_id();
3110 data->cpu_entry.reserved = 0;
3112 header->size += sizeof(data->cpu_entry);
3115 if (sample_type & PERF_SAMPLE_PERIOD)
3116 header->size += sizeof(data->period);
3118 if (sample_type & PERF_SAMPLE_READ)
3119 header->size += perf_event_read_size(event);
3121 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3124 data->callchain = perf_callchain(regs);
3126 if (data->callchain)
3127 size += data->callchain->nr;
3129 header->size += size * sizeof(u64);
3132 if (sample_type & PERF_SAMPLE_RAW) {
3133 int size = sizeof(u32);
3136 size += data->raw->size;
3138 size += sizeof(u32);
3140 WARN_ON_ONCE(size & (sizeof(u64)-1));
3141 header->size += size;
3145 static void perf_event_output(struct perf_event *event, int nmi,
3146 struct perf_sample_data *data,
3147 struct pt_regs *regs)
3149 struct perf_output_handle handle;
3150 struct perf_event_header header;
3152 perf_prepare_sample(&header, data, event, regs);
3154 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3157 perf_output_sample(&handle, &header, data, event);
3159 perf_output_end(&handle);
3166 struct perf_read_event {
3167 struct perf_event_header header;
3174 perf_event_read_event(struct perf_event *event,
3175 struct task_struct *task)
3177 struct perf_output_handle handle;
3178 struct perf_read_event read_event = {
3180 .type = PERF_RECORD_READ,
3182 .size = sizeof(read_event) + perf_event_read_size(event),
3184 .pid = perf_event_pid(event, task),
3185 .tid = perf_event_tid(event, task),
3189 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3193 perf_output_put(&handle, read_event);
3194 perf_output_read(&handle, event);
3196 perf_output_end(&handle);
3200 * task tracking -- fork/exit
3202 * enabled by: attr.comm | attr.mmap | attr.task
3205 struct perf_task_event {
3206 struct task_struct *task;
3207 struct perf_event_context *task_ctx;
3210 struct perf_event_header header;
3220 static void perf_event_task_output(struct perf_event *event,
3221 struct perf_task_event *task_event)
3223 struct perf_output_handle handle;
3225 struct task_struct *task = task_event->task;
3228 size = task_event->event_id.header.size;
3229 ret = perf_output_begin(&handle, event, size, 0, 0);
3234 task_event->event_id.pid = perf_event_pid(event, task);
3235 task_event->event_id.ppid = perf_event_pid(event, current);
3237 task_event->event_id.tid = perf_event_tid(event, task);
3238 task_event->event_id.ptid = perf_event_tid(event, current);
3240 task_event->event_id.time = perf_clock();
3242 perf_output_put(&handle, task_event->event_id);
3244 perf_output_end(&handle);
3247 static int perf_event_task_match(struct perf_event *event)
3249 if (event->attr.comm || event->attr.mmap || event->attr.task)
3255 static void perf_event_task_ctx(struct perf_event_context *ctx,
3256 struct perf_task_event *task_event)
3258 struct perf_event *event;
3260 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3261 if (perf_event_task_match(event))
3262 perf_event_task_output(event, task_event);
3266 static void perf_event_task_event(struct perf_task_event *task_event)
3268 struct perf_cpu_context *cpuctx;
3269 struct perf_event_context *ctx = task_event->task_ctx;
3272 cpuctx = &get_cpu_var(perf_cpu_context);
3273 perf_event_task_ctx(&cpuctx->ctx, task_event);
3274 put_cpu_var(perf_cpu_context);
3277 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3279 perf_event_task_ctx(ctx, task_event);
3283 static void perf_event_task(struct task_struct *task,
3284 struct perf_event_context *task_ctx,
3287 struct perf_task_event task_event;
3289 if (!atomic_read(&nr_comm_events) &&
3290 !atomic_read(&nr_mmap_events) &&
3291 !atomic_read(&nr_task_events))
3294 task_event = (struct perf_task_event){
3296 .task_ctx = task_ctx,
3299 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3301 .size = sizeof(task_event.event_id),
3310 perf_event_task_event(&task_event);
3313 void perf_event_fork(struct task_struct *task)
3315 perf_event_task(task, NULL, 1);
3322 struct perf_comm_event {
3323 struct task_struct *task;
3328 struct perf_event_header header;
3335 static void perf_event_comm_output(struct perf_event *event,
3336 struct perf_comm_event *comm_event)
3338 struct perf_output_handle handle;
3339 int size = comm_event->event_id.header.size;
3340 int ret = perf_output_begin(&handle, event, size, 0, 0);
3345 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3346 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3348 perf_output_put(&handle, comm_event->event_id);
3349 perf_output_copy(&handle, comm_event->comm,
3350 comm_event->comm_size);
3351 perf_output_end(&handle);
3354 static int perf_event_comm_match(struct perf_event *event)
3356 if (event->attr.comm)
3362 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3363 struct perf_comm_event *comm_event)
3365 struct perf_event *event;
3367 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3368 if (perf_event_comm_match(event))
3369 perf_event_comm_output(event, comm_event);
3373 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3375 struct perf_cpu_context *cpuctx;
3376 struct perf_event_context *ctx;
3378 char comm[TASK_COMM_LEN];
3380 memset(comm, 0, sizeof(comm));
3381 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3382 size = ALIGN(strlen(comm)+1, sizeof(u64));
3384 comm_event->comm = comm;
3385 comm_event->comm_size = size;
3387 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3390 cpuctx = &get_cpu_var(perf_cpu_context);
3391 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3392 put_cpu_var(perf_cpu_context);
3395 * doesn't really matter which of the child contexts the
3396 * events ends up in.
3398 ctx = rcu_dereference(current->perf_event_ctxp);
3400 perf_event_comm_ctx(ctx, comm_event);
3404 void perf_event_comm(struct task_struct *task)
3406 struct perf_comm_event comm_event;
3408 if (task->perf_event_ctxp)
3409 perf_event_enable_on_exec(task);
3411 if (!atomic_read(&nr_comm_events))
3414 comm_event = (struct perf_comm_event){
3420 .type = PERF_RECORD_COMM,
3429 perf_event_comm_event(&comm_event);
3436 struct perf_mmap_event {
3437 struct vm_area_struct *vma;
3439 const char *file_name;
3443 struct perf_event_header header;
3453 static void perf_event_mmap_output(struct perf_event *event,
3454 struct perf_mmap_event *mmap_event)
3456 struct perf_output_handle handle;
3457 int size = mmap_event->event_id.header.size;
3458 int ret = perf_output_begin(&handle, event, size, 0, 0);
3463 mmap_event->event_id.pid = perf_event_pid(event, current);
3464 mmap_event->event_id.tid = perf_event_tid(event, current);
3466 perf_output_put(&handle, mmap_event->event_id);
3467 perf_output_copy(&handle, mmap_event->file_name,
3468 mmap_event->file_size);
3469 perf_output_end(&handle);
3472 static int perf_event_mmap_match(struct perf_event *event,
3473 struct perf_mmap_event *mmap_event)
3475 if (event->attr.mmap)
3481 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3482 struct perf_mmap_event *mmap_event)
3484 struct perf_event *event;
3486 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3487 if (perf_event_mmap_match(event, mmap_event))
3488 perf_event_mmap_output(event, mmap_event);
3492 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3494 struct perf_cpu_context *cpuctx;
3495 struct perf_event_context *ctx;
3496 struct vm_area_struct *vma = mmap_event->vma;
3497 struct file *file = vma->vm_file;
3503 memset(tmp, 0, sizeof(tmp));
3507 * d_path works from the end of the buffer backwards, so we
3508 * need to add enough zero bytes after the string to handle
3509 * the 64bit alignment we do later.
3511 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3513 name = strncpy(tmp, "//enomem", sizeof(tmp));
3516 name = d_path(&file->f_path, buf, PATH_MAX);
3518 name = strncpy(tmp, "//toolong", sizeof(tmp));
3522 if (arch_vma_name(mmap_event->vma)) {
3523 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3529 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3533 name = strncpy(tmp, "//anon", sizeof(tmp));
3538 size = ALIGN(strlen(name)+1, sizeof(u64));
3540 mmap_event->file_name = name;
3541 mmap_event->file_size = size;
3543 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3546 cpuctx = &get_cpu_var(perf_cpu_context);
3547 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3548 put_cpu_var(perf_cpu_context);
3551 * doesn't really matter which of the child contexts the
3552 * events ends up in.
3554 ctx = rcu_dereference(current->perf_event_ctxp);
3556 perf_event_mmap_ctx(ctx, mmap_event);
3562 void __perf_event_mmap(struct vm_area_struct *vma)
3564 struct perf_mmap_event mmap_event;
3566 if (!atomic_read(&nr_mmap_events))
3569 mmap_event = (struct perf_mmap_event){
3575 .type = PERF_RECORD_MMAP,
3581 .start = vma->vm_start,
3582 .len = vma->vm_end - vma->vm_start,
3583 .pgoff = vma->vm_pgoff,
3587 perf_event_mmap_event(&mmap_event);
3591 * IRQ throttle logging
3594 static void perf_log_throttle(struct perf_event *event, int enable)
3596 struct perf_output_handle handle;
3600 struct perf_event_header header;
3604 } throttle_event = {
3606 .type = PERF_RECORD_THROTTLE,
3608 .size = sizeof(throttle_event),
3610 .time = perf_clock(),
3611 .id = primary_event_id(event),
3612 .stream_id = event->id,
3616 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3618 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3622 perf_output_put(&handle, throttle_event);
3623 perf_output_end(&handle);
3627 * Generic event overflow handling, sampling.
3630 static int __perf_event_overflow(struct perf_event *event, int nmi,
3631 int throttle, struct perf_sample_data *data,
3632 struct pt_regs *regs)
3634 int events = atomic_read(&event->event_limit);
3635 struct hw_perf_event *hwc = &event->hw;
3638 throttle = (throttle && event->pmu->unthrottle != NULL);
3643 if (hwc->interrupts != MAX_INTERRUPTS) {
3645 if (HZ * hwc->interrupts >
3646 (u64)sysctl_perf_event_sample_rate) {
3647 hwc->interrupts = MAX_INTERRUPTS;
3648 perf_log_throttle(event, 0);
3653 * Keep re-disabling events even though on the previous
3654 * pass we disabled it - just in case we raced with a
3655 * sched-in and the event got enabled again:
3661 if (event->attr.freq) {
3662 u64 now = perf_clock();
3663 s64 delta = now - hwc->freq_stamp;
3665 hwc->freq_stamp = now;
3667 if (delta > 0 && delta < TICK_NSEC)
3668 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3672 * XXX event_limit might not quite work as expected on inherited
3676 event->pending_kill = POLL_IN;
3677 if (events && atomic_dec_and_test(&event->event_limit)) {
3679 event->pending_kill = POLL_HUP;
3681 event->pending_disable = 1;
3682 perf_pending_queue(&event->pending,
3683 perf_pending_event);
3685 perf_event_disable(event);
3688 if (event->overflow_handler)
3689 event->overflow_handler(event, nmi, data, regs);
3691 perf_event_output(event, nmi, data, regs);
3696 int perf_event_overflow(struct perf_event *event, int nmi,
3697 struct perf_sample_data *data,
3698 struct pt_regs *regs)
3700 return __perf_event_overflow(event, nmi, 1, data, regs);
3704 * Generic software event infrastructure
3708 * We directly increment event->count and keep a second value in
3709 * event->hw.period_left to count intervals. This period event
3710 * is kept in the range [-sample_period, 0] so that we can use the
3714 static u64 perf_swevent_set_period(struct perf_event *event)
3716 struct hw_perf_event *hwc = &event->hw;
3717 u64 period = hwc->last_period;
3721 hwc->last_period = hwc->sample_period;
3724 old = val = atomic64_read(&hwc->period_left);
3728 nr = div64_u64(period + val, period);
3729 offset = nr * period;
3731 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3737 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3738 int nmi, struct perf_sample_data *data,
3739 struct pt_regs *regs)
3741 struct hw_perf_event *hwc = &event->hw;
3744 data->period = event->hw.last_period;
3746 overflow = perf_swevent_set_period(event);
3748 if (hwc->interrupts == MAX_INTERRUPTS)
3751 for (; overflow; overflow--) {
3752 if (__perf_event_overflow(event, nmi, throttle,
3755 * We inhibit the overflow from happening when
3756 * hwc->interrupts == MAX_INTERRUPTS.
3764 static void perf_swevent_unthrottle(struct perf_event *event)
3767 * Nothing to do, we already reset hwc->interrupts.
3771 static void perf_swevent_add(struct perf_event *event, u64 nr,
3772 int nmi, struct perf_sample_data *data,
3773 struct pt_regs *regs)
3775 struct hw_perf_event *hwc = &event->hw;
3777 atomic64_add(nr, &event->count);
3782 if (!hwc->sample_period)
3785 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3786 return perf_swevent_overflow(event, 1, nmi, data, regs);
3788 if (atomic64_add_negative(nr, &hwc->period_left))
3791 perf_swevent_overflow(event, 0, nmi, data, regs);
3794 static int perf_swevent_is_counting(struct perf_event *event)
3797 * The event is active, we're good!
3799 if (event->state == PERF_EVENT_STATE_ACTIVE)
3803 * The event is off/error, not counting.
3805 if (event->state != PERF_EVENT_STATE_INACTIVE)
3809 * The event is inactive, if the context is active
3810 * we're part of a group that didn't make it on the 'pmu',
3813 if (event->ctx->is_active)
3817 * We're inactive and the context is too, this means the
3818 * task is scheduled out, we're counting events that happen
3819 * to us, like migration events.
3824 static int perf_tp_event_match(struct perf_event *event,
3825 struct perf_sample_data *data);
3827 static int perf_swevent_match(struct perf_event *event,
3828 enum perf_type_id type,
3830 struct perf_sample_data *data,
3831 struct pt_regs *regs)
3833 if (!perf_swevent_is_counting(event))
3836 if (event->attr.type != type)
3838 if (event->attr.config != event_id)
3842 if (event->attr.exclude_user && user_mode(regs))
3845 if (event->attr.exclude_kernel && !user_mode(regs))
3849 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3850 !perf_tp_event_match(event, data))
3856 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3857 enum perf_type_id type,
3858 u32 event_id, u64 nr, int nmi,
3859 struct perf_sample_data *data,
3860 struct pt_regs *regs)
3862 struct perf_event *event;
3864 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3865 if (perf_swevent_match(event, type, event_id, data, regs))
3866 perf_swevent_add(event, nr, nmi, data, regs);
3871 * Must be called with preemption disabled
3873 int perf_swevent_get_recursion_context(int **recursion)
3875 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3878 *recursion = &cpuctx->recursion[3];
3880 *recursion = &cpuctx->recursion[2];
3881 else if (in_softirq())
3882 *recursion = &cpuctx->recursion[1];
3884 *recursion = &cpuctx->recursion[0];
3893 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3895 void perf_swevent_put_recursion_context(int *recursion)
3899 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3901 static void __do_perf_sw_event(enum perf_type_id type, u32 event_id,
3903 struct perf_sample_data *data,
3904 struct pt_regs *regs)
3906 struct perf_event_context *ctx;
3907 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3910 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3911 nr, nmi, data, regs);
3913 * doesn't really matter which of the child contexts the
3914 * events ends up in.
3916 ctx = rcu_dereference(current->perf_event_ctxp);
3918 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3922 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3924 struct perf_sample_data *data,
3925 struct pt_regs *regs)
3931 if (perf_swevent_get_recursion_context(&recursion))
3934 __do_perf_sw_event(type, event_id, nr, nmi, data, regs);
3936 perf_swevent_put_recursion_context(recursion);
3941 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3942 struct pt_regs *regs, u64 addr)
3944 struct perf_sample_data data;
3949 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3952 static void perf_swevent_read(struct perf_event *event)
3956 static int perf_swevent_enable(struct perf_event *event)
3958 struct hw_perf_event *hwc = &event->hw;
3960 if (hwc->sample_period) {
3961 hwc->last_period = hwc->sample_period;
3962 perf_swevent_set_period(event);
3967 static void perf_swevent_disable(struct perf_event *event)
3971 static const struct pmu perf_ops_generic = {
3972 .enable = perf_swevent_enable,
3973 .disable = perf_swevent_disable,
3974 .read = perf_swevent_read,
3975 .unthrottle = perf_swevent_unthrottle,
3979 * hrtimer based swevent callback
3982 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3984 enum hrtimer_restart ret = HRTIMER_RESTART;
3985 struct perf_sample_data data;
3986 struct pt_regs *regs;
3987 struct perf_event *event;
3990 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
3991 event->pmu->read(event);
3994 regs = get_irq_regs();
3996 * In case we exclude kernel IPs or are somehow not in interrupt
3997 * context, provide the next best thing, the user IP.
3999 if ((event->attr.exclude_kernel || !regs) &&
4000 !event->attr.exclude_user)
4001 regs = task_pt_regs(current);
4004 if (!(event->attr.exclude_idle && current->pid == 0))
4005 if (perf_event_overflow(event, 0, &data, regs))
4006 ret = HRTIMER_NORESTART;
4009 period = max_t(u64, 10000, event->hw.sample_period);
4010 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4015 static void perf_swevent_start_hrtimer(struct perf_event *event)
4017 struct hw_perf_event *hwc = &event->hw;
4019 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4020 hwc->hrtimer.function = perf_swevent_hrtimer;
4021 if (hwc->sample_period) {
4024 if (hwc->remaining) {
4025 if (hwc->remaining < 0)
4028 period = hwc->remaining;
4031 period = max_t(u64, 10000, hwc->sample_period);
4033 __hrtimer_start_range_ns(&hwc->hrtimer,
4034 ns_to_ktime(period), 0,
4035 HRTIMER_MODE_REL, 0);
4039 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4041 struct hw_perf_event *hwc = &event->hw;
4043 if (hwc->sample_period) {
4044 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4045 hwc->remaining = ktime_to_ns(remaining);
4047 hrtimer_cancel(&hwc->hrtimer);
4052 * Software event: cpu wall time clock
4055 static void cpu_clock_perf_event_update(struct perf_event *event)
4057 int cpu = raw_smp_processor_id();
4061 now = cpu_clock(cpu);
4062 prev = atomic64_read(&event->hw.prev_count);
4063 atomic64_set(&event->hw.prev_count, now);
4064 atomic64_add(now - prev, &event->count);
4067 static int cpu_clock_perf_event_enable(struct perf_event *event)
4069 struct hw_perf_event *hwc = &event->hw;
4070 int cpu = raw_smp_processor_id();
4072 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4073 perf_swevent_start_hrtimer(event);
4078 static void cpu_clock_perf_event_disable(struct perf_event *event)
4080 perf_swevent_cancel_hrtimer(event);
4081 cpu_clock_perf_event_update(event);
4084 static void cpu_clock_perf_event_read(struct perf_event *event)
4086 cpu_clock_perf_event_update(event);
4089 static const struct pmu perf_ops_cpu_clock = {
4090 .enable = cpu_clock_perf_event_enable,
4091 .disable = cpu_clock_perf_event_disable,
4092 .read = cpu_clock_perf_event_read,
4096 * Software event: task time clock
4099 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4104 prev = atomic64_xchg(&event->hw.prev_count, now);
4106 atomic64_add(delta, &event->count);
4109 static int task_clock_perf_event_enable(struct perf_event *event)
4111 struct hw_perf_event *hwc = &event->hw;
4114 now = event->ctx->time;
4116 atomic64_set(&hwc->prev_count, now);
4118 perf_swevent_start_hrtimer(event);
4123 static void task_clock_perf_event_disable(struct perf_event *event)
4125 perf_swevent_cancel_hrtimer(event);
4126 task_clock_perf_event_update(event, event->ctx->time);
4130 static void task_clock_perf_event_read(struct perf_event *event)
4135 update_context_time(event->ctx);
4136 time = event->ctx->time;
4138 u64 now = perf_clock();
4139 u64 delta = now - event->ctx->timestamp;
4140 time = event->ctx->time + delta;
4143 task_clock_perf_event_update(event, time);
4146 static const struct pmu perf_ops_task_clock = {
4147 .enable = task_clock_perf_event_enable,
4148 .disable = task_clock_perf_event_disable,
4149 .read = task_clock_perf_event_read,
4152 #ifdef CONFIG_EVENT_PROFILE
4154 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4157 struct perf_raw_record raw = {
4162 struct perf_sample_data data = {
4167 struct pt_regs *regs = get_irq_regs();
4170 regs = task_pt_regs(current);
4172 /* Trace events already protected against recursion */
4173 __do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4176 EXPORT_SYMBOL_GPL(perf_tp_event);
4178 static int perf_tp_event_match(struct perf_event *event,
4179 struct perf_sample_data *data)
4181 void *record = data->raw->data;
4183 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4188 static void tp_perf_event_destroy(struct perf_event *event)
4190 ftrace_profile_disable(event->attr.config);
4193 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4196 * Raw tracepoint data is a severe data leak, only allow root to
4199 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4200 perf_paranoid_tracepoint_raw() &&
4201 !capable(CAP_SYS_ADMIN))
4202 return ERR_PTR(-EPERM);
4204 if (ftrace_profile_enable(event->attr.config))
4207 event->destroy = tp_perf_event_destroy;
4209 return &perf_ops_generic;
4212 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4217 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4220 filter_str = strndup_user(arg, PAGE_SIZE);
4221 if (IS_ERR(filter_str))
4222 return PTR_ERR(filter_str);
4224 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4230 static void perf_event_free_filter(struct perf_event *event)
4232 ftrace_profile_free_filter(event);
4237 static int perf_tp_event_match(struct perf_event *event,
4238 struct perf_sample_data *data)
4243 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4248 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4253 static void perf_event_free_filter(struct perf_event *event)
4257 #endif /* CONFIG_EVENT_PROFILE */
4259 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4260 static void bp_perf_event_destroy(struct perf_event *event)
4262 release_bp_slot(event);
4265 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4269 * The breakpoint is already filled if we haven't created the counter
4270 * through perf syscall
4271 * FIXME: manage to get trigerred to NULL if it comes from syscalls
4274 err = register_perf_hw_breakpoint(bp);
4276 err = __register_perf_hw_breakpoint(bp);
4278 return ERR_PTR(err);
4280 bp->destroy = bp_perf_event_destroy;
4282 return &perf_ops_bp;
4285 void perf_bp_event(struct perf_event *bp, void *regs)
4290 static void bp_perf_event_destroy(struct perf_event *event)
4294 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4299 void perf_bp_event(struct perf_event *bp, void *regs)
4304 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4306 static void sw_perf_event_destroy(struct perf_event *event)
4308 u64 event_id = event->attr.config;
4310 WARN_ON(event->parent);
4312 atomic_dec(&perf_swevent_enabled[event_id]);
4315 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4317 const struct pmu *pmu = NULL;
4318 u64 event_id = event->attr.config;
4321 * Software events (currently) can't in general distinguish
4322 * between user, kernel and hypervisor events.
4323 * However, context switches and cpu migrations are considered
4324 * to be kernel events, and page faults are never hypervisor
4328 case PERF_COUNT_SW_CPU_CLOCK:
4329 pmu = &perf_ops_cpu_clock;
4332 case PERF_COUNT_SW_TASK_CLOCK:
4334 * If the user instantiates this as a per-cpu event,
4335 * use the cpu_clock event instead.
4337 if (event->ctx->task)
4338 pmu = &perf_ops_task_clock;
4340 pmu = &perf_ops_cpu_clock;
4343 case PERF_COUNT_SW_PAGE_FAULTS:
4344 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4345 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4346 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4347 case PERF_COUNT_SW_CPU_MIGRATIONS:
4348 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4349 case PERF_COUNT_SW_EMULATION_FAULTS:
4350 if (!event->parent) {
4351 atomic_inc(&perf_swevent_enabled[event_id]);
4352 event->destroy = sw_perf_event_destroy;
4354 pmu = &perf_ops_generic;
4362 * Allocate and initialize a event structure
4364 static struct perf_event *
4365 perf_event_alloc(struct perf_event_attr *attr,
4367 struct perf_event_context *ctx,
4368 struct perf_event *group_leader,
4369 struct perf_event *parent_event,
4370 perf_callback_t callback,
4373 const struct pmu *pmu;
4374 struct perf_event *event;
4375 struct hw_perf_event *hwc;
4378 event = kzalloc(sizeof(*event), gfpflags);
4380 return ERR_PTR(-ENOMEM);
4383 * Single events are their own group leaders, with an
4384 * empty sibling list:
4387 group_leader = event;
4389 mutex_init(&event->child_mutex);
4390 INIT_LIST_HEAD(&event->child_list);
4392 INIT_LIST_HEAD(&event->group_entry);
4393 INIT_LIST_HEAD(&event->event_entry);
4394 INIT_LIST_HEAD(&event->sibling_list);
4395 init_waitqueue_head(&event->waitq);
4397 mutex_init(&event->mmap_mutex);
4400 event->attr = *attr;
4401 event->group_leader = group_leader;
4406 event->parent = parent_event;
4408 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4409 event->id = atomic64_inc_return(&perf_event_id);
4411 event->state = PERF_EVENT_STATE_INACTIVE;
4413 if (!callback && parent_event)
4414 callback = parent_event->callback;
4416 event->callback = callback;
4419 event->state = PERF_EVENT_STATE_OFF;
4424 hwc->sample_period = attr->sample_period;
4425 if (attr->freq && attr->sample_freq)
4426 hwc->sample_period = 1;
4427 hwc->last_period = hwc->sample_period;
4429 atomic64_set(&hwc->period_left, hwc->sample_period);
4432 * we currently do not support PERF_FORMAT_GROUP on inherited events
4434 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4437 switch (attr->type) {
4439 case PERF_TYPE_HARDWARE:
4440 case PERF_TYPE_HW_CACHE:
4441 pmu = hw_perf_event_init(event);
4444 case PERF_TYPE_SOFTWARE:
4445 pmu = sw_perf_event_init(event);
4448 case PERF_TYPE_TRACEPOINT:
4449 pmu = tp_perf_event_init(event);
4452 case PERF_TYPE_BREAKPOINT:
4453 pmu = bp_perf_event_init(event);
4464 else if (IS_ERR(pmu))
4469 put_pid_ns(event->ns);
4471 return ERR_PTR(err);
4476 if (!event->parent) {
4477 atomic_inc(&nr_events);
4478 if (event->attr.mmap)
4479 atomic_inc(&nr_mmap_events);
4480 if (event->attr.comm)
4481 atomic_inc(&nr_comm_events);
4482 if (event->attr.task)
4483 atomic_inc(&nr_task_events);
4489 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4490 struct perf_event_attr *attr)
4495 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4499 * zero the full structure, so that a short copy will be nice.
4501 memset(attr, 0, sizeof(*attr));
4503 ret = get_user(size, &uattr->size);
4507 if (size > PAGE_SIZE) /* silly large */
4510 if (!size) /* abi compat */
4511 size = PERF_ATTR_SIZE_VER0;
4513 if (size < PERF_ATTR_SIZE_VER0)
4517 * If we're handed a bigger struct than we know of,
4518 * ensure all the unknown bits are 0 - i.e. new
4519 * user-space does not rely on any kernel feature
4520 * extensions we dont know about yet.
4522 if (size > sizeof(*attr)) {
4523 unsigned char __user *addr;
4524 unsigned char __user *end;
4527 addr = (void __user *)uattr + sizeof(*attr);
4528 end = (void __user *)uattr + size;
4530 for (; addr < end; addr++) {
4531 ret = get_user(val, addr);
4537 size = sizeof(*attr);
4540 ret = copy_from_user(attr, uattr, size);
4545 * If the type exists, the corresponding creation will verify
4548 if (attr->type >= PERF_TYPE_MAX)
4551 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4554 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4557 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4564 put_user(sizeof(*attr), &uattr->size);
4569 static int perf_event_set_output(struct perf_event *event, int output_fd)
4571 struct perf_event *output_event = NULL;
4572 struct file *output_file = NULL;
4573 struct perf_event *old_output;
4574 int fput_needed = 0;
4580 output_file = fget_light(output_fd, &fput_needed);
4584 if (output_file->f_op != &perf_fops)
4587 output_event = output_file->private_data;
4589 /* Don't chain output fds */
4590 if (output_event->output)
4593 /* Don't set an output fd when we already have an output channel */
4597 atomic_long_inc(&output_file->f_count);
4600 mutex_lock(&event->mmap_mutex);
4601 old_output = event->output;
4602 rcu_assign_pointer(event->output, output_event);
4603 mutex_unlock(&event->mmap_mutex);
4607 * we need to make sure no existing perf_output_*()
4608 * is still referencing this event.
4611 fput(old_output->filp);
4616 fput_light(output_file, fput_needed);
4621 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4623 * @attr_uptr: event_id type attributes for monitoring/sampling
4626 * @group_fd: group leader event fd
4628 SYSCALL_DEFINE5(perf_event_open,
4629 struct perf_event_attr __user *, attr_uptr,
4630 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4632 struct perf_event *event, *group_leader;
4633 struct perf_event_attr attr;
4634 struct perf_event_context *ctx;
4635 struct file *event_file = NULL;
4636 struct file *group_file = NULL;
4637 int fput_needed = 0;
4638 int fput_needed2 = 0;
4641 /* for future expandability... */
4642 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4645 err = perf_copy_attr(attr_uptr, &attr);
4649 if (!attr.exclude_kernel) {
4650 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4655 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4660 * Get the target context (task or percpu):
4662 ctx = find_get_context(pid, cpu);
4664 return PTR_ERR(ctx);
4667 * Look up the group leader (we will attach this event to it):
4669 group_leader = NULL;
4670 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4672 group_file = fget_light(group_fd, &fput_needed);
4674 goto err_put_context;
4675 if (group_file->f_op != &perf_fops)
4676 goto err_put_context;
4678 group_leader = group_file->private_data;
4680 * Do not allow a recursive hierarchy (this new sibling
4681 * becoming part of another group-sibling):
4683 if (group_leader->group_leader != group_leader)
4684 goto err_put_context;
4686 * Do not allow to attach to a group in a different
4687 * task or CPU context:
4689 if (group_leader->ctx != ctx)
4690 goto err_put_context;
4692 * Only a group leader can be exclusive or pinned
4694 if (attr.exclusive || attr.pinned)
4695 goto err_put_context;
4698 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4699 NULL, NULL, GFP_KERNEL);
4700 err = PTR_ERR(event);
4702 goto err_put_context;
4704 err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4706 goto err_free_put_context;
4708 event_file = fget_light(err, &fput_needed2);
4710 goto err_free_put_context;
4712 if (flags & PERF_FLAG_FD_OUTPUT) {
4713 err = perf_event_set_output(event, group_fd);
4715 goto err_fput_free_put_context;
4718 event->filp = event_file;
4719 WARN_ON_ONCE(ctx->parent_ctx);
4720 mutex_lock(&ctx->mutex);
4721 perf_install_in_context(ctx, event, cpu);
4723 mutex_unlock(&ctx->mutex);
4725 event->owner = current;
4726 get_task_struct(current);
4727 mutex_lock(¤t->perf_event_mutex);
4728 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4729 mutex_unlock(¤t->perf_event_mutex);
4731 err_fput_free_put_context:
4732 fput_light(event_file, fput_needed2);
4734 err_free_put_context:
4742 fput_light(group_file, fput_needed);
4748 * perf_event_create_kernel_counter
4750 * @attr: attributes of the counter to create
4751 * @cpu: cpu in which the counter is bound
4752 * @pid: task to profile
4755 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4756 pid_t pid, perf_callback_t callback)
4758 struct perf_event *event;
4759 struct perf_event_context *ctx;
4763 * Get the target context (task or percpu):
4766 ctx = find_get_context(pid, cpu);
4770 event = perf_event_alloc(attr, cpu, ctx, NULL,
4771 NULL, callback, GFP_KERNEL);
4772 err = PTR_ERR(event);
4774 goto err_put_context;
4777 WARN_ON_ONCE(ctx->parent_ctx);
4778 mutex_lock(&ctx->mutex);
4779 perf_install_in_context(ctx, event, cpu);
4781 mutex_unlock(&ctx->mutex);
4783 event->owner = current;
4784 get_task_struct(current);
4785 mutex_lock(¤t->perf_event_mutex);
4786 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4787 mutex_unlock(¤t->perf_event_mutex);
4797 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4800 * inherit a event from parent task to child task:
4802 static struct perf_event *
4803 inherit_event(struct perf_event *parent_event,
4804 struct task_struct *parent,
4805 struct perf_event_context *parent_ctx,
4806 struct task_struct *child,
4807 struct perf_event *group_leader,
4808 struct perf_event_context *child_ctx)
4810 struct perf_event *child_event;
4813 * Instead of creating recursive hierarchies of events,
4814 * we link inherited events back to the original parent,
4815 * which has a filp for sure, which we use as the reference
4818 if (parent_event->parent)
4819 parent_event = parent_event->parent;
4821 child_event = perf_event_alloc(&parent_event->attr,
4822 parent_event->cpu, child_ctx,
4823 group_leader, parent_event,
4825 if (IS_ERR(child_event))
4830 * Make the child state follow the state of the parent event,
4831 * not its attr.disabled bit. We hold the parent's mutex,
4832 * so we won't race with perf_event_{en, dis}able_family.
4834 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4835 child_event->state = PERF_EVENT_STATE_INACTIVE;
4837 child_event->state = PERF_EVENT_STATE_OFF;
4839 if (parent_event->attr.freq)
4840 child_event->hw.sample_period = parent_event->hw.sample_period;
4842 child_event->overflow_handler = parent_event->overflow_handler;
4845 * Link it up in the child's context:
4847 add_event_to_ctx(child_event, child_ctx);
4850 * Get a reference to the parent filp - we will fput it
4851 * when the child event exits. This is safe to do because
4852 * we are in the parent and we know that the filp still
4853 * exists and has a nonzero count:
4855 atomic_long_inc(&parent_event->filp->f_count);
4858 * Link this into the parent event's child list
4860 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4861 mutex_lock(&parent_event->child_mutex);
4862 list_add_tail(&child_event->child_list, &parent_event->child_list);
4863 mutex_unlock(&parent_event->child_mutex);
4868 static int inherit_group(struct perf_event *parent_event,
4869 struct task_struct *parent,
4870 struct perf_event_context *parent_ctx,
4871 struct task_struct *child,
4872 struct perf_event_context *child_ctx)
4874 struct perf_event *leader;
4875 struct perf_event *sub;
4876 struct perf_event *child_ctr;
4878 leader = inherit_event(parent_event, parent, parent_ctx,
4879 child, NULL, child_ctx);
4881 return PTR_ERR(leader);
4882 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4883 child_ctr = inherit_event(sub, parent, parent_ctx,
4884 child, leader, child_ctx);
4885 if (IS_ERR(child_ctr))
4886 return PTR_ERR(child_ctr);
4891 static void sync_child_event(struct perf_event *child_event,
4892 struct task_struct *child)
4894 struct perf_event *parent_event = child_event->parent;
4897 if (child_event->attr.inherit_stat)
4898 perf_event_read_event(child_event, child);
4900 child_val = atomic64_read(&child_event->count);
4903 * Add back the child's count to the parent's count:
4905 atomic64_add(child_val, &parent_event->count);
4906 atomic64_add(child_event->total_time_enabled,
4907 &parent_event->child_total_time_enabled);
4908 atomic64_add(child_event->total_time_running,
4909 &parent_event->child_total_time_running);
4912 * Remove this event from the parent's list
4914 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4915 mutex_lock(&parent_event->child_mutex);
4916 list_del_init(&child_event->child_list);
4917 mutex_unlock(&parent_event->child_mutex);
4920 * Release the parent event, if this was the last
4923 fput(parent_event->filp);
4927 __perf_event_exit_task(struct perf_event *child_event,
4928 struct perf_event_context *child_ctx,
4929 struct task_struct *child)
4931 struct perf_event *parent_event;
4933 update_event_times(child_event);
4934 perf_event_remove_from_context(child_event);
4936 parent_event = child_event->parent;
4938 * It can happen that parent exits first, and has events
4939 * that are still around due to the child reference. These
4940 * events need to be zapped - but otherwise linger.
4943 sync_child_event(child_event, child);
4944 free_event(child_event);
4949 * When a child task exits, feed back event values to parent events.
4951 void perf_event_exit_task(struct task_struct *child)
4953 struct perf_event *child_event, *tmp;
4954 struct perf_event_context *child_ctx;
4955 unsigned long flags;
4957 if (likely(!child->perf_event_ctxp)) {
4958 perf_event_task(child, NULL, 0);
4962 local_irq_save(flags);
4964 * We can't reschedule here because interrupts are disabled,
4965 * and either child is current or it is a task that can't be
4966 * scheduled, so we are now safe from rescheduling changing
4969 child_ctx = child->perf_event_ctxp;
4970 __perf_event_task_sched_out(child_ctx);
4973 * Take the context lock here so that if find_get_context is
4974 * reading child->perf_event_ctxp, we wait until it has
4975 * incremented the context's refcount before we do put_ctx below.
4977 spin_lock(&child_ctx->lock);
4978 child->perf_event_ctxp = NULL;
4980 * If this context is a clone; unclone it so it can't get
4981 * swapped to another process while we're removing all
4982 * the events from it.
4984 unclone_ctx(child_ctx);
4985 spin_unlock_irqrestore(&child_ctx->lock, flags);
4988 * Report the task dead after unscheduling the events so that we
4989 * won't get any samples after PERF_RECORD_EXIT. We can however still
4990 * get a few PERF_RECORD_READ events.
4992 perf_event_task(child, child_ctx, 0);
4995 * We can recurse on the same lock type through:
4997 * __perf_event_exit_task()
4998 * sync_child_event()
4999 * fput(parent_event->filp)
5001 * mutex_lock(&ctx->mutex)
5003 * But since its the parent context it won't be the same instance.
5005 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5008 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5010 __perf_event_exit_task(child_event, child_ctx, child);
5013 * If the last event was a group event, it will have appended all
5014 * its siblings to the list, but we obtained 'tmp' before that which
5015 * will still point to the list head terminating the iteration.
5017 if (!list_empty(&child_ctx->group_list))
5020 mutex_unlock(&child_ctx->mutex);
5026 * free an unexposed, unused context as created by inheritance by
5027 * init_task below, used by fork() in case of fail.
5029 void perf_event_free_task(struct task_struct *task)
5031 struct perf_event_context *ctx = task->perf_event_ctxp;
5032 struct perf_event *event, *tmp;
5037 mutex_lock(&ctx->mutex);
5039 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5040 struct perf_event *parent = event->parent;
5042 if (WARN_ON_ONCE(!parent))
5045 mutex_lock(&parent->child_mutex);
5046 list_del_init(&event->child_list);
5047 mutex_unlock(&parent->child_mutex);
5051 list_del_event(event, ctx);
5055 if (!list_empty(&ctx->group_list))
5058 mutex_unlock(&ctx->mutex);
5064 * Initialize the perf_event context in task_struct
5066 int perf_event_init_task(struct task_struct *child)
5068 struct perf_event_context *child_ctx, *parent_ctx;
5069 struct perf_event_context *cloned_ctx;
5070 struct perf_event *event;
5071 struct task_struct *parent = current;
5072 int inherited_all = 1;
5075 child->perf_event_ctxp = NULL;
5077 mutex_init(&child->perf_event_mutex);
5078 INIT_LIST_HEAD(&child->perf_event_list);
5080 if (likely(!parent->perf_event_ctxp))
5084 * This is executed from the parent task context, so inherit
5085 * events that have been marked for cloning.
5086 * First allocate and initialize a context for the child.
5089 child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5093 __perf_event_init_context(child_ctx, child);
5094 child->perf_event_ctxp = child_ctx;
5095 get_task_struct(child);
5098 * If the parent's context is a clone, pin it so it won't get
5101 parent_ctx = perf_pin_task_context(parent);
5104 * No need to check if parent_ctx != NULL here; since we saw
5105 * it non-NULL earlier, the only reason for it to become NULL
5106 * is if we exit, and since we're currently in the middle of
5107 * a fork we can't be exiting at the same time.
5111 * Lock the parent list. No need to lock the child - not PID
5112 * hashed yet and not running, so nobody can access it.
5114 mutex_lock(&parent_ctx->mutex);
5117 * We dont have to disable NMIs - we are only looking at
5118 * the list, not manipulating it:
5120 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5122 if (!event->attr.inherit) {
5127 ret = inherit_group(event, parent, parent_ctx,
5135 if (inherited_all) {
5137 * Mark the child context as a clone of the parent
5138 * context, or of whatever the parent is a clone of.
5139 * Note that if the parent is a clone, it could get
5140 * uncloned at any point, but that doesn't matter
5141 * because the list of events and the generation
5142 * count can't have changed since we took the mutex.
5144 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5146 child_ctx->parent_ctx = cloned_ctx;
5147 child_ctx->parent_gen = parent_ctx->parent_gen;
5149 child_ctx->parent_ctx = parent_ctx;
5150 child_ctx->parent_gen = parent_ctx->generation;
5152 get_ctx(child_ctx->parent_ctx);
5155 mutex_unlock(&parent_ctx->mutex);
5157 perf_unpin_context(parent_ctx);
5162 static void __cpuinit perf_event_init_cpu(int cpu)
5164 struct perf_cpu_context *cpuctx;
5166 cpuctx = &per_cpu(perf_cpu_context, cpu);
5167 __perf_event_init_context(&cpuctx->ctx, NULL);
5169 spin_lock(&perf_resource_lock);
5170 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5171 spin_unlock(&perf_resource_lock);
5173 hw_perf_event_setup(cpu);
5176 #ifdef CONFIG_HOTPLUG_CPU
5177 static void __perf_event_exit_cpu(void *info)
5179 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5180 struct perf_event_context *ctx = &cpuctx->ctx;
5181 struct perf_event *event, *tmp;
5183 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5184 __perf_event_remove_from_context(event);
5186 static void perf_event_exit_cpu(int cpu)
5188 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5189 struct perf_event_context *ctx = &cpuctx->ctx;
5191 mutex_lock(&ctx->mutex);
5192 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5193 mutex_unlock(&ctx->mutex);
5196 static inline void perf_event_exit_cpu(int cpu) { }
5199 static int __cpuinit
5200 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5202 unsigned int cpu = (long)hcpu;
5206 case CPU_UP_PREPARE:
5207 case CPU_UP_PREPARE_FROZEN:
5208 perf_event_init_cpu(cpu);
5212 case CPU_ONLINE_FROZEN:
5213 hw_perf_event_setup_online(cpu);
5216 case CPU_DOWN_PREPARE:
5217 case CPU_DOWN_PREPARE_FROZEN:
5218 perf_event_exit_cpu(cpu);
5229 * This has to have a higher priority than migration_notifier in sched.c.
5231 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5232 .notifier_call = perf_cpu_notify,
5236 void __init perf_event_init(void)
5238 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5239 (void *)(long)smp_processor_id());
5240 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5241 (void *)(long)smp_processor_id());
5242 register_cpu_notifier(&perf_cpu_nb);
5245 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5247 return sprintf(buf, "%d\n", perf_reserved_percpu);
5251 perf_set_reserve_percpu(struct sysdev_class *class,
5255 struct perf_cpu_context *cpuctx;
5259 err = strict_strtoul(buf, 10, &val);
5262 if (val > perf_max_events)
5265 spin_lock(&perf_resource_lock);
5266 perf_reserved_percpu = val;
5267 for_each_online_cpu(cpu) {
5268 cpuctx = &per_cpu(perf_cpu_context, cpu);
5269 spin_lock_irq(&cpuctx->ctx.lock);
5270 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5271 perf_max_events - perf_reserved_percpu);
5272 cpuctx->max_pertask = mpt;
5273 spin_unlock_irq(&cpuctx->ctx.lock);
5275 spin_unlock(&perf_resource_lock);
5280 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5282 return sprintf(buf, "%d\n", perf_overcommit);
5286 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5291 err = strict_strtoul(buf, 10, &val);
5297 spin_lock(&perf_resource_lock);
5298 perf_overcommit = val;
5299 spin_unlock(&perf_resource_lock);
5304 static SYSDEV_CLASS_ATTR(
5307 perf_show_reserve_percpu,
5308 perf_set_reserve_percpu
5311 static SYSDEV_CLASS_ATTR(
5314 perf_show_overcommit,
5318 static struct attribute *perfclass_attrs[] = {
5319 &attr_reserve_percpu.attr,
5320 &attr_overcommit.attr,
5324 static struct attribute_group perfclass_attr_group = {
5325 .attrs = perfclass_attrs,
5326 .name = "perf_events",
5329 static int __init perf_event_sysfs_init(void)
5331 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5332 &perfclass_attr_group);
5334 device_initcall(perf_event_sysfs_init);