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)
1530 update_context_time(ctx);
1531 event->pmu->read(event);
1532 update_event_times(event);
1535 static u64 perf_event_read(struct perf_event *event)
1538 * If event is enabled and currently active on a CPU, update the
1539 * value in the event structure:
1541 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1542 smp_call_function_single(event->oncpu,
1543 __perf_event_read, event, 1);
1544 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1545 update_event_times(event);
1548 return atomic64_read(&event->count);
1552 * Initialize the perf_event context in a task_struct:
1555 __perf_event_init_context(struct perf_event_context *ctx,
1556 struct task_struct *task)
1558 memset(ctx, 0, sizeof(*ctx));
1559 spin_lock_init(&ctx->lock);
1560 mutex_init(&ctx->mutex);
1561 INIT_LIST_HEAD(&ctx->group_list);
1562 INIT_LIST_HEAD(&ctx->event_list);
1563 atomic_set(&ctx->refcount, 1);
1567 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1569 struct perf_event_context *ctx;
1570 struct perf_cpu_context *cpuctx;
1571 struct task_struct *task;
1572 unsigned long flags;
1576 * If cpu is not a wildcard then this is a percpu event:
1579 /* Must be root to operate on a CPU event: */
1580 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1581 return ERR_PTR(-EACCES);
1583 if (cpu < 0 || cpu > num_possible_cpus())
1584 return ERR_PTR(-EINVAL);
1587 * We could be clever and allow to attach a event to an
1588 * offline CPU and activate it when the CPU comes up, but
1591 if (!cpu_isset(cpu, cpu_online_map))
1592 return ERR_PTR(-ENODEV);
1594 cpuctx = &per_cpu(perf_cpu_context, cpu);
1605 task = find_task_by_vpid(pid);
1607 get_task_struct(task);
1611 return ERR_PTR(-ESRCH);
1614 * Can't attach events to a dying task.
1617 if (task->flags & PF_EXITING)
1620 /* Reuse ptrace permission checks for now. */
1622 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1626 ctx = perf_lock_task_context(task, &flags);
1629 spin_unlock_irqrestore(&ctx->lock, flags);
1633 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1637 __perf_event_init_context(ctx, task);
1639 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1641 * We raced with some other task; use
1642 * the context they set.
1647 get_task_struct(task);
1650 put_task_struct(task);
1654 put_task_struct(task);
1655 return ERR_PTR(err);
1658 static void perf_event_free_filter(struct perf_event *event);
1660 static void free_event_rcu(struct rcu_head *head)
1662 struct perf_event *event;
1664 event = container_of(head, struct perf_event, rcu_head);
1666 put_pid_ns(event->ns);
1667 perf_event_free_filter(event);
1671 static void perf_pending_sync(struct perf_event *event);
1673 static void free_event(struct perf_event *event)
1675 perf_pending_sync(event);
1677 if (!event->parent) {
1678 atomic_dec(&nr_events);
1679 if (event->attr.mmap)
1680 atomic_dec(&nr_mmap_events);
1681 if (event->attr.comm)
1682 atomic_dec(&nr_comm_events);
1683 if (event->attr.task)
1684 atomic_dec(&nr_task_events);
1687 if (event->output) {
1688 fput(event->output->filp);
1689 event->output = NULL;
1693 event->destroy(event);
1695 put_ctx(event->ctx);
1696 call_rcu(&event->rcu_head, free_event_rcu);
1700 * Called when the last reference to the file is gone.
1702 static int perf_release(struct inode *inode, struct file *file)
1704 struct perf_event *event = file->private_data;
1705 struct perf_event_context *ctx = event->ctx;
1707 file->private_data = NULL;
1709 WARN_ON_ONCE(ctx->parent_ctx);
1710 mutex_lock(&ctx->mutex);
1711 perf_event_remove_from_context(event);
1712 mutex_unlock(&ctx->mutex);
1714 mutex_lock(&event->owner->perf_event_mutex);
1715 list_del_init(&event->owner_entry);
1716 mutex_unlock(&event->owner->perf_event_mutex);
1717 put_task_struct(event->owner);
1724 int perf_event_release_kernel(struct perf_event *event)
1726 struct perf_event_context *ctx = event->ctx;
1728 WARN_ON_ONCE(ctx->parent_ctx);
1729 mutex_lock(&ctx->mutex);
1730 perf_event_remove_from_context(event);
1731 mutex_unlock(&ctx->mutex);
1733 mutex_lock(&event->owner->perf_event_mutex);
1734 list_del_init(&event->owner_entry);
1735 mutex_unlock(&event->owner->perf_event_mutex);
1736 put_task_struct(event->owner);
1742 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1744 static int perf_event_read_size(struct perf_event *event)
1746 int entry = sizeof(u64); /* value */
1750 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1751 size += sizeof(u64);
1753 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1754 size += sizeof(u64);
1756 if (event->attr.read_format & PERF_FORMAT_ID)
1757 entry += sizeof(u64);
1759 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1760 nr += event->group_leader->nr_siblings;
1761 size += sizeof(u64);
1769 u64 perf_event_read_value(struct perf_event *event)
1771 struct perf_event *child;
1774 total += perf_event_read(event);
1775 list_for_each_entry(child, &event->child_list, child_list)
1776 total += perf_event_read(child);
1780 EXPORT_SYMBOL_GPL(perf_event_read_value);
1782 static int perf_event_read_group(struct perf_event *event,
1783 u64 read_format, char __user *buf)
1785 struct perf_event *leader = event->group_leader, *sub;
1786 int n = 0, size = 0, ret = 0;
1790 count = perf_event_read_value(leader);
1792 values[n++] = 1 + leader->nr_siblings;
1793 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1794 values[n++] = leader->total_time_enabled +
1795 atomic64_read(&leader->child_total_time_enabled);
1797 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1798 values[n++] = leader->total_time_running +
1799 atomic64_read(&leader->child_total_time_running);
1801 values[n++] = count;
1802 if (read_format & PERF_FORMAT_ID)
1803 values[n++] = primary_event_id(leader);
1805 size = n * sizeof(u64);
1807 if (copy_to_user(buf, values, size))
1812 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1815 values[n++] = perf_event_read_value(sub);
1816 if (read_format & PERF_FORMAT_ID)
1817 values[n++] = primary_event_id(sub);
1819 size = n * sizeof(u64);
1821 if (copy_to_user(buf + size, values, size))
1830 static int perf_event_read_one(struct perf_event *event,
1831 u64 read_format, char __user *buf)
1836 values[n++] = perf_event_read_value(event);
1837 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1838 values[n++] = event->total_time_enabled +
1839 atomic64_read(&event->child_total_time_enabled);
1841 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1842 values[n++] = event->total_time_running +
1843 atomic64_read(&event->child_total_time_running);
1845 if (read_format & PERF_FORMAT_ID)
1846 values[n++] = primary_event_id(event);
1848 if (copy_to_user(buf, values, n * sizeof(u64)))
1851 return n * sizeof(u64);
1855 * Read the performance event - simple non blocking version for now
1858 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1860 u64 read_format = event->attr.read_format;
1864 * Return end-of-file for a read on a event that is in
1865 * error state (i.e. because it was pinned but it couldn't be
1866 * scheduled on to the CPU at some point).
1868 if (event->state == PERF_EVENT_STATE_ERROR)
1871 if (count < perf_event_read_size(event))
1874 WARN_ON_ONCE(event->ctx->parent_ctx);
1875 mutex_lock(&event->child_mutex);
1876 if (read_format & PERF_FORMAT_GROUP)
1877 ret = perf_event_read_group(event, read_format, buf);
1879 ret = perf_event_read_one(event, read_format, buf);
1880 mutex_unlock(&event->child_mutex);
1886 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1888 struct perf_event *event = file->private_data;
1890 return perf_read_hw(event, buf, count);
1893 static unsigned int perf_poll(struct file *file, poll_table *wait)
1895 struct perf_event *event = file->private_data;
1896 struct perf_mmap_data *data;
1897 unsigned int events = POLL_HUP;
1900 data = rcu_dereference(event->data);
1902 events = atomic_xchg(&data->poll, 0);
1905 poll_wait(file, &event->waitq, wait);
1910 static void perf_event_reset(struct perf_event *event)
1912 (void)perf_event_read(event);
1913 atomic64_set(&event->count, 0);
1914 perf_event_update_userpage(event);
1918 * Holding the top-level event's child_mutex means that any
1919 * descendant process that has inherited this event will block
1920 * in sync_child_event if it goes to exit, thus satisfying the
1921 * task existence requirements of perf_event_enable/disable.
1923 static void perf_event_for_each_child(struct perf_event *event,
1924 void (*func)(struct perf_event *))
1926 struct perf_event *child;
1928 WARN_ON_ONCE(event->ctx->parent_ctx);
1929 mutex_lock(&event->child_mutex);
1931 list_for_each_entry(child, &event->child_list, child_list)
1933 mutex_unlock(&event->child_mutex);
1936 static void perf_event_for_each(struct perf_event *event,
1937 void (*func)(struct perf_event *))
1939 struct perf_event_context *ctx = event->ctx;
1940 struct perf_event *sibling;
1942 WARN_ON_ONCE(ctx->parent_ctx);
1943 mutex_lock(&ctx->mutex);
1944 event = event->group_leader;
1946 perf_event_for_each_child(event, func);
1948 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1949 perf_event_for_each_child(event, func);
1950 mutex_unlock(&ctx->mutex);
1953 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1955 struct perf_event_context *ctx = event->ctx;
1960 if (!event->attr.sample_period)
1963 size = copy_from_user(&value, arg, sizeof(value));
1964 if (size != sizeof(value))
1970 spin_lock_irq(&ctx->lock);
1971 if (event->attr.freq) {
1972 if (value > sysctl_perf_event_sample_rate) {
1977 event->attr.sample_freq = value;
1979 event->attr.sample_period = value;
1980 event->hw.sample_period = value;
1983 spin_unlock_irq(&ctx->lock);
1988 static int perf_event_set_output(struct perf_event *event, int output_fd);
1989 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
1991 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1993 struct perf_event *event = file->private_data;
1994 void (*func)(struct perf_event *);
1998 case PERF_EVENT_IOC_ENABLE:
1999 func = perf_event_enable;
2001 case PERF_EVENT_IOC_DISABLE:
2002 func = perf_event_disable;
2004 case PERF_EVENT_IOC_RESET:
2005 func = perf_event_reset;
2008 case PERF_EVENT_IOC_REFRESH:
2009 return perf_event_refresh(event, arg);
2011 case PERF_EVENT_IOC_PERIOD:
2012 return perf_event_period(event, (u64 __user *)arg);
2014 case PERF_EVENT_IOC_SET_OUTPUT:
2015 return perf_event_set_output(event, arg);
2017 case PERF_EVENT_IOC_SET_FILTER:
2018 return perf_event_set_filter(event, (void __user *)arg);
2024 if (flags & PERF_IOC_FLAG_GROUP)
2025 perf_event_for_each(event, func);
2027 perf_event_for_each_child(event, func);
2032 int perf_event_task_enable(void)
2034 struct perf_event *event;
2036 mutex_lock(¤t->perf_event_mutex);
2037 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2038 perf_event_for_each_child(event, perf_event_enable);
2039 mutex_unlock(¤t->perf_event_mutex);
2044 int perf_event_task_disable(void)
2046 struct perf_event *event;
2048 mutex_lock(¤t->perf_event_mutex);
2049 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2050 perf_event_for_each_child(event, perf_event_disable);
2051 mutex_unlock(¤t->perf_event_mutex);
2056 #ifndef PERF_EVENT_INDEX_OFFSET
2057 # define PERF_EVENT_INDEX_OFFSET 0
2060 static int perf_event_index(struct perf_event *event)
2062 if (event->state != PERF_EVENT_STATE_ACTIVE)
2065 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2069 * Callers need to ensure there can be no nesting of this function, otherwise
2070 * the seqlock logic goes bad. We can not serialize this because the arch
2071 * code calls this from NMI context.
2073 void perf_event_update_userpage(struct perf_event *event)
2075 struct perf_event_mmap_page *userpg;
2076 struct perf_mmap_data *data;
2079 data = rcu_dereference(event->data);
2083 userpg = data->user_page;
2086 * Disable preemption so as to not let the corresponding user-space
2087 * spin too long if we get preempted.
2092 userpg->index = perf_event_index(event);
2093 userpg->offset = atomic64_read(&event->count);
2094 if (event->state == PERF_EVENT_STATE_ACTIVE)
2095 userpg->offset -= atomic64_read(&event->hw.prev_count);
2097 userpg->time_enabled = event->total_time_enabled +
2098 atomic64_read(&event->child_total_time_enabled);
2100 userpg->time_running = event->total_time_running +
2101 atomic64_read(&event->child_total_time_running);
2110 static unsigned long perf_data_size(struct perf_mmap_data *data)
2112 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2115 #ifndef CONFIG_PERF_USE_VMALLOC
2118 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2121 static struct page *
2122 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2124 if (pgoff > data->nr_pages)
2128 return virt_to_page(data->user_page);
2130 return virt_to_page(data->data_pages[pgoff - 1]);
2133 static struct perf_mmap_data *
2134 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2136 struct perf_mmap_data *data;
2140 WARN_ON(atomic_read(&event->mmap_count));
2142 size = sizeof(struct perf_mmap_data);
2143 size += nr_pages * sizeof(void *);
2145 data = kzalloc(size, GFP_KERNEL);
2149 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2150 if (!data->user_page)
2151 goto fail_user_page;
2153 for (i = 0; i < nr_pages; i++) {
2154 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2155 if (!data->data_pages[i])
2156 goto fail_data_pages;
2159 data->data_order = 0;
2160 data->nr_pages = nr_pages;
2165 for (i--; i >= 0; i--)
2166 free_page((unsigned long)data->data_pages[i]);
2168 free_page((unsigned long)data->user_page);
2177 static void perf_mmap_free_page(unsigned long addr)
2179 struct page *page = virt_to_page((void *)addr);
2181 page->mapping = NULL;
2185 static void perf_mmap_data_free(struct perf_mmap_data *data)
2189 perf_mmap_free_page((unsigned long)data->user_page);
2190 for (i = 0; i < data->nr_pages; i++)
2191 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2197 * Back perf_mmap() with vmalloc memory.
2199 * Required for architectures that have d-cache aliasing issues.
2202 static struct page *
2203 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2205 if (pgoff > (1UL << data->data_order))
2208 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2211 static void perf_mmap_unmark_page(void *addr)
2213 struct page *page = vmalloc_to_page(addr);
2215 page->mapping = NULL;
2218 static void perf_mmap_data_free_work(struct work_struct *work)
2220 struct perf_mmap_data *data;
2224 data = container_of(work, struct perf_mmap_data, work);
2225 nr = 1 << data->data_order;
2227 base = data->user_page;
2228 for (i = 0; i < nr + 1; i++)
2229 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2234 static void perf_mmap_data_free(struct perf_mmap_data *data)
2236 schedule_work(&data->work);
2239 static struct perf_mmap_data *
2240 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2242 struct perf_mmap_data *data;
2246 WARN_ON(atomic_read(&event->mmap_count));
2248 size = sizeof(struct perf_mmap_data);
2249 size += sizeof(void *);
2251 data = kzalloc(size, GFP_KERNEL);
2255 INIT_WORK(&data->work, perf_mmap_data_free_work);
2257 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2261 data->user_page = all_buf;
2262 data->data_pages[0] = all_buf + PAGE_SIZE;
2263 data->data_order = ilog2(nr_pages);
2277 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2279 struct perf_event *event = vma->vm_file->private_data;
2280 struct perf_mmap_data *data;
2281 int ret = VM_FAULT_SIGBUS;
2283 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2284 if (vmf->pgoff == 0)
2290 data = rcu_dereference(event->data);
2294 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2297 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2301 get_page(vmf->page);
2302 vmf->page->mapping = vma->vm_file->f_mapping;
2303 vmf->page->index = vmf->pgoff;
2313 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2315 long max_size = perf_data_size(data);
2317 atomic_set(&data->lock, -1);
2319 if (event->attr.watermark) {
2320 data->watermark = min_t(long, max_size,
2321 event->attr.wakeup_watermark);
2324 if (!data->watermark)
2325 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2328 rcu_assign_pointer(event->data, data);
2331 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2333 struct perf_mmap_data *data;
2335 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2336 perf_mmap_data_free(data);
2340 static void perf_mmap_data_release(struct perf_event *event)
2342 struct perf_mmap_data *data = event->data;
2344 WARN_ON(atomic_read(&event->mmap_count));
2346 rcu_assign_pointer(event->data, NULL);
2347 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2350 static void perf_mmap_open(struct vm_area_struct *vma)
2352 struct perf_event *event = vma->vm_file->private_data;
2354 atomic_inc(&event->mmap_count);
2357 static void perf_mmap_close(struct vm_area_struct *vma)
2359 struct perf_event *event = vma->vm_file->private_data;
2361 WARN_ON_ONCE(event->ctx->parent_ctx);
2362 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2363 unsigned long size = perf_data_size(event->data);
2364 struct user_struct *user = current_user();
2366 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2367 vma->vm_mm->locked_vm -= event->data->nr_locked;
2368 perf_mmap_data_release(event);
2369 mutex_unlock(&event->mmap_mutex);
2373 static const struct vm_operations_struct perf_mmap_vmops = {
2374 .open = perf_mmap_open,
2375 .close = perf_mmap_close,
2376 .fault = perf_mmap_fault,
2377 .page_mkwrite = perf_mmap_fault,
2380 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2382 struct perf_event *event = file->private_data;
2383 unsigned long user_locked, user_lock_limit;
2384 struct user_struct *user = current_user();
2385 unsigned long locked, lock_limit;
2386 struct perf_mmap_data *data;
2387 unsigned long vma_size;
2388 unsigned long nr_pages;
2389 long user_extra, extra;
2392 if (!(vma->vm_flags & VM_SHARED))
2395 vma_size = vma->vm_end - vma->vm_start;
2396 nr_pages = (vma_size / PAGE_SIZE) - 1;
2399 * If we have data pages ensure they're a power-of-two number, so we
2400 * can do bitmasks instead of modulo.
2402 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2405 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2408 if (vma->vm_pgoff != 0)
2411 WARN_ON_ONCE(event->ctx->parent_ctx);
2412 mutex_lock(&event->mmap_mutex);
2413 if (event->output) {
2418 if (atomic_inc_not_zero(&event->mmap_count)) {
2419 if (nr_pages != event->data->nr_pages)
2424 user_extra = nr_pages + 1;
2425 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2428 * Increase the limit linearly with more CPUs:
2430 user_lock_limit *= num_online_cpus();
2432 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2435 if (user_locked > user_lock_limit)
2436 extra = user_locked - user_lock_limit;
2438 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2439 lock_limit >>= PAGE_SHIFT;
2440 locked = vma->vm_mm->locked_vm + extra;
2442 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2443 !capable(CAP_IPC_LOCK)) {
2448 WARN_ON(event->data);
2450 data = perf_mmap_data_alloc(event, nr_pages);
2456 perf_mmap_data_init(event, data);
2458 atomic_set(&event->mmap_count, 1);
2459 atomic_long_add(user_extra, &user->locked_vm);
2460 vma->vm_mm->locked_vm += extra;
2461 event->data->nr_locked = extra;
2462 if (vma->vm_flags & VM_WRITE)
2463 event->data->writable = 1;
2466 mutex_unlock(&event->mmap_mutex);
2468 vma->vm_flags |= VM_RESERVED;
2469 vma->vm_ops = &perf_mmap_vmops;
2474 static int perf_fasync(int fd, struct file *filp, int on)
2476 struct inode *inode = filp->f_path.dentry->d_inode;
2477 struct perf_event *event = filp->private_data;
2480 mutex_lock(&inode->i_mutex);
2481 retval = fasync_helper(fd, filp, on, &event->fasync);
2482 mutex_unlock(&inode->i_mutex);
2490 static const struct file_operations perf_fops = {
2491 .release = perf_release,
2494 .unlocked_ioctl = perf_ioctl,
2495 .compat_ioctl = perf_ioctl,
2497 .fasync = perf_fasync,
2503 * If there's data, ensure we set the poll() state and publish everything
2504 * to user-space before waking everybody up.
2507 void perf_event_wakeup(struct perf_event *event)
2509 wake_up_all(&event->waitq);
2511 if (event->pending_kill) {
2512 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2513 event->pending_kill = 0;
2520 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2522 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2523 * single linked list and use cmpxchg() to add entries lockless.
2526 static void perf_pending_event(struct perf_pending_entry *entry)
2528 struct perf_event *event = container_of(entry,
2529 struct perf_event, pending);
2531 if (event->pending_disable) {
2532 event->pending_disable = 0;
2533 __perf_event_disable(event);
2536 if (event->pending_wakeup) {
2537 event->pending_wakeup = 0;
2538 perf_event_wakeup(event);
2542 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2544 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2548 static void perf_pending_queue(struct perf_pending_entry *entry,
2549 void (*func)(struct perf_pending_entry *))
2551 struct perf_pending_entry **head;
2553 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2558 head = &get_cpu_var(perf_pending_head);
2561 entry->next = *head;
2562 } while (cmpxchg(head, entry->next, entry) != entry->next);
2564 set_perf_event_pending();
2566 put_cpu_var(perf_pending_head);
2569 static int __perf_pending_run(void)
2571 struct perf_pending_entry *list;
2574 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2575 while (list != PENDING_TAIL) {
2576 void (*func)(struct perf_pending_entry *);
2577 struct perf_pending_entry *entry = list;
2584 * Ensure we observe the unqueue before we issue the wakeup,
2585 * so that we won't be waiting forever.
2586 * -- see perf_not_pending().
2597 static inline int perf_not_pending(struct perf_event *event)
2600 * If we flush on whatever cpu we run, there is a chance we don't
2604 __perf_pending_run();
2608 * Ensure we see the proper queue state before going to sleep
2609 * so that we do not miss the wakeup. -- see perf_pending_handle()
2612 return event->pending.next == NULL;
2615 static void perf_pending_sync(struct perf_event *event)
2617 wait_event(event->waitq, perf_not_pending(event));
2620 void perf_event_do_pending(void)
2622 __perf_pending_run();
2626 * Callchain support -- arch specific
2629 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2637 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2638 unsigned long offset, unsigned long head)
2642 if (!data->writable)
2645 mask = perf_data_size(data) - 1;
2647 offset = (offset - tail) & mask;
2648 head = (head - tail) & mask;
2650 if ((int)(head - offset) < 0)
2656 static void perf_output_wakeup(struct perf_output_handle *handle)
2658 atomic_set(&handle->data->poll, POLL_IN);
2661 handle->event->pending_wakeup = 1;
2662 perf_pending_queue(&handle->event->pending,
2663 perf_pending_event);
2665 perf_event_wakeup(handle->event);
2669 * Curious locking construct.
2671 * We need to ensure a later event_id doesn't publish a head when a former
2672 * event_id isn't done writing. However since we need to deal with NMIs we
2673 * cannot fully serialize things.
2675 * What we do is serialize between CPUs so we only have to deal with NMI
2676 * nesting on a single CPU.
2678 * We only publish the head (and generate a wakeup) when the outer-most
2679 * event_id completes.
2681 static void perf_output_lock(struct perf_output_handle *handle)
2683 struct perf_mmap_data *data = handle->data;
2684 int cur, cpu = get_cpu();
2689 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2701 static void perf_output_unlock(struct perf_output_handle *handle)
2703 struct perf_mmap_data *data = handle->data;
2707 data->done_head = data->head;
2709 if (!handle->locked)
2714 * The xchg implies a full barrier that ensures all writes are done
2715 * before we publish the new head, matched by a rmb() in userspace when
2716 * reading this position.
2718 while ((head = atomic_long_xchg(&data->done_head, 0)))
2719 data->user_page->data_head = head;
2722 * NMI can happen here, which means we can miss a done_head update.
2725 cpu = atomic_xchg(&data->lock, -1);
2726 WARN_ON_ONCE(cpu != smp_processor_id());
2729 * Therefore we have to validate we did not indeed do so.
2731 if (unlikely(atomic_long_read(&data->done_head))) {
2733 * Since we had it locked, we can lock it again.
2735 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2741 if (atomic_xchg(&data->wakeup, 0))
2742 perf_output_wakeup(handle);
2747 void perf_output_copy(struct perf_output_handle *handle,
2748 const void *buf, unsigned int len)
2750 unsigned int pages_mask;
2751 unsigned long offset;
2755 offset = handle->offset;
2756 pages_mask = handle->data->nr_pages - 1;
2757 pages = handle->data->data_pages;
2760 unsigned long page_offset;
2761 unsigned long page_size;
2764 nr = (offset >> PAGE_SHIFT) & pages_mask;
2765 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2766 page_offset = offset & (page_size - 1);
2767 size = min_t(unsigned int, page_size - page_offset, len);
2769 memcpy(pages[nr] + page_offset, buf, size);
2776 handle->offset = offset;
2779 * Check we didn't copy past our reservation window, taking the
2780 * possible unsigned int wrap into account.
2782 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2785 int perf_output_begin(struct perf_output_handle *handle,
2786 struct perf_event *event, unsigned int size,
2787 int nmi, int sample)
2789 struct perf_event *output_event;
2790 struct perf_mmap_data *data;
2791 unsigned long tail, offset, head;
2794 struct perf_event_header header;
2801 * For inherited events we send all the output towards the parent.
2804 event = event->parent;
2806 output_event = rcu_dereference(event->output);
2808 event = output_event;
2810 data = rcu_dereference(event->data);
2814 handle->data = data;
2815 handle->event = event;
2817 handle->sample = sample;
2819 if (!data->nr_pages)
2822 have_lost = atomic_read(&data->lost);
2824 size += sizeof(lost_event);
2826 perf_output_lock(handle);
2830 * Userspace could choose to issue a mb() before updating the
2831 * tail pointer. So that all reads will be completed before the
2834 tail = ACCESS_ONCE(data->user_page->data_tail);
2836 offset = head = atomic_long_read(&data->head);
2838 if (unlikely(!perf_output_space(data, tail, offset, head)))
2840 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2842 handle->offset = offset;
2843 handle->head = head;
2845 if (head - tail > data->watermark)
2846 atomic_set(&data->wakeup, 1);
2849 lost_event.header.type = PERF_RECORD_LOST;
2850 lost_event.header.misc = 0;
2851 lost_event.header.size = sizeof(lost_event);
2852 lost_event.id = event->id;
2853 lost_event.lost = atomic_xchg(&data->lost, 0);
2855 perf_output_put(handle, lost_event);
2861 atomic_inc(&data->lost);
2862 perf_output_unlock(handle);
2869 void perf_output_end(struct perf_output_handle *handle)
2871 struct perf_event *event = handle->event;
2872 struct perf_mmap_data *data = handle->data;
2874 int wakeup_events = event->attr.wakeup_events;
2876 if (handle->sample && wakeup_events) {
2877 int events = atomic_inc_return(&data->events);
2878 if (events >= wakeup_events) {
2879 atomic_sub(wakeup_events, &data->events);
2880 atomic_set(&data->wakeup, 1);
2884 perf_output_unlock(handle);
2888 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2891 * only top level events have the pid namespace they were created in
2894 event = event->parent;
2896 return task_tgid_nr_ns(p, event->ns);
2899 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2902 * only top level events have the pid namespace they were created in
2905 event = event->parent;
2907 return task_pid_nr_ns(p, event->ns);
2910 static void perf_output_read_one(struct perf_output_handle *handle,
2911 struct perf_event *event)
2913 u64 read_format = event->attr.read_format;
2917 values[n++] = atomic64_read(&event->count);
2918 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2919 values[n++] = event->total_time_enabled +
2920 atomic64_read(&event->child_total_time_enabled);
2922 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2923 values[n++] = event->total_time_running +
2924 atomic64_read(&event->child_total_time_running);
2926 if (read_format & PERF_FORMAT_ID)
2927 values[n++] = primary_event_id(event);
2929 perf_output_copy(handle, values, n * sizeof(u64));
2933 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2935 static void perf_output_read_group(struct perf_output_handle *handle,
2936 struct perf_event *event)
2938 struct perf_event *leader = event->group_leader, *sub;
2939 u64 read_format = event->attr.read_format;
2943 values[n++] = 1 + leader->nr_siblings;
2945 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2946 values[n++] = leader->total_time_enabled;
2948 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2949 values[n++] = leader->total_time_running;
2951 if (leader != event)
2952 leader->pmu->read(leader);
2954 values[n++] = atomic64_read(&leader->count);
2955 if (read_format & PERF_FORMAT_ID)
2956 values[n++] = primary_event_id(leader);
2958 perf_output_copy(handle, values, n * sizeof(u64));
2960 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2964 sub->pmu->read(sub);
2966 values[n++] = atomic64_read(&sub->count);
2967 if (read_format & PERF_FORMAT_ID)
2968 values[n++] = primary_event_id(sub);
2970 perf_output_copy(handle, values, n * sizeof(u64));
2974 static void perf_output_read(struct perf_output_handle *handle,
2975 struct perf_event *event)
2977 if (event->attr.read_format & PERF_FORMAT_GROUP)
2978 perf_output_read_group(handle, event);
2980 perf_output_read_one(handle, event);
2983 void perf_output_sample(struct perf_output_handle *handle,
2984 struct perf_event_header *header,
2985 struct perf_sample_data *data,
2986 struct perf_event *event)
2988 u64 sample_type = data->type;
2990 perf_output_put(handle, *header);
2992 if (sample_type & PERF_SAMPLE_IP)
2993 perf_output_put(handle, data->ip);
2995 if (sample_type & PERF_SAMPLE_TID)
2996 perf_output_put(handle, data->tid_entry);
2998 if (sample_type & PERF_SAMPLE_TIME)
2999 perf_output_put(handle, data->time);
3001 if (sample_type & PERF_SAMPLE_ADDR)
3002 perf_output_put(handle, data->addr);
3004 if (sample_type & PERF_SAMPLE_ID)
3005 perf_output_put(handle, data->id);
3007 if (sample_type & PERF_SAMPLE_STREAM_ID)
3008 perf_output_put(handle, data->stream_id);
3010 if (sample_type & PERF_SAMPLE_CPU)
3011 perf_output_put(handle, data->cpu_entry);
3013 if (sample_type & PERF_SAMPLE_PERIOD)
3014 perf_output_put(handle, data->period);
3016 if (sample_type & PERF_SAMPLE_READ)
3017 perf_output_read(handle, event);
3019 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3020 if (data->callchain) {
3023 if (data->callchain)
3024 size += data->callchain->nr;
3026 size *= sizeof(u64);
3028 perf_output_copy(handle, data->callchain, size);
3031 perf_output_put(handle, nr);
3035 if (sample_type & PERF_SAMPLE_RAW) {
3037 perf_output_put(handle, data->raw->size);
3038 perf_output_copy(handle, data->raw->data,
3045 .size = sizeof(u32),
3048 perf_output_put(handle, raw);
3053 void perf_prepare_sample(struct perf_event_header *header,
3054 struct perf_sample_data *data,
3055 struct perf_event *event,
3056 struct pt_regs *regs)
3058 u64 sample_type = event->attr.sample_type;
3060 data->type = sample_type;
3062 header->type = PERF_RECORD_SAMPLE;
3063 header->size = sizeof(*header);
3066 header->misc |= perf_misc_flags(regs);
3068 if (sample_type & PERF_SAMPLE_IP) {
3069 data->ip = perf_instruction_pointer(regs);
3071 header->size += sizeof(data->ip);
3074 if (sample_type & PERF_SAMPLE_TID) {
3075 /* namespace issues */
3076 data->tid_entry.pid = perf_event_pid(event, current);
3077 data->tid_entry.tid = perf_event_tid(event, current);
3079 header->size += sizeof(data->tid_entry);
3082 if (sample_type & PERF_SAMPLE_TIME) {
3083 data->time = perf_clock();
3085 header->size += sizeof(data->time);
3088 if (sample_type & PERF_SAMPLE_ADDR)
3089 header->size += sizeof(data->addr);
3091 if (sample_type & PERF_SAMPLE_ID) {
3092 data->id = primary_event_id(event);
3094 header->size += sizeof(data->id);
3097 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3098 data->stream_id = event->id;
3100 header->size += sizeof(data->stream_id);
3103 if (sample_type & PERF_SAMPLE_CPU) {
3104 data->cpu_entry.cpu = raw_smp_processor_id();
3105 data->cpu_entry.reserved = 0;
3107 header->size += sizeof(data->cpu_entry);
3110 if (sample_type & PERF_SAMPLE_PERIOD)
3111 header->size += sizeof(data->period);
3113 if (sample_type & PERF_SAMPLE_READ)
3114 header->size += perf_event_read_size(event);
3116 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3119 data->callchain = perf_callchain(regs);
3121 if (data->callchain)
3122 size += data->callchain->nr;
3124 header->size += size * sizeof(u64);
3127 if (sample_type & PERF_SAMPLE_RAW) {
3128 int size = sizeof(u32);
3131 size += data->raw->size;
3133 size += sizeof(u32);
3135 WARN_ON_ONCE(size & (sizeof(u64)-1));
3136 header->size += size;
3140 static void perf_event_output(struct perf_event *event, int nmi,
3141 struct perf_sample_data *data,
3142 struct pt_regs *regs)
3144 struct perf_output_handle handle;
3145 struct perf_event_header header;
3147 perf_prepare_sample(&header, data, event, regs);
3149 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3152 perf_output_sample(&handle, &header, data, event);
3154 perf_output_end(&handle);
3161 struct perf_read_event {
3162 struct perf_event_header header;
3169 perf_event_read_event(struct perf_event *event,
3170 struct task_struct *task)
3172 struct perf_output_handle handle;
3173 struct perf_read_event read_event = {
3175 .type = PERF_RECORD_READ,
3177 .size = sizeof(read_event) + perf_event_read_size(event),
3179 .pid = perf_event_pid(event, task),
3180 .tid = perf_event_tid(event, task),
3184 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3188 perf_output_put(&handle, read_event);
3189 perf_output_read(&handle, event);
3191 perf_output_end(&handle);
3195 * task tracking -- fork/exit
3197 * enabled by: attr.comm | attr.mmap | attr.task
3200 struct perf_task_event {
3201 struct task_struct *task;
3202 struct perf_event_context *task_ctx;
3205 struct perf_event_header header;
3215 static void perf_event_task_output(struct perf_event *event,
3216 struct perf_task_event *task_event)
3218 struct perf_output_handle handle;
3220 struct task_struct *task = task_event->task;
3223 size = task_event->event_id.header.size;
3224 ret = perf_output_begin(&handle, event, size, 0, 0);
3229 task_event->event_id.pid = perf_event_pid(event, task);
3230 task_event->event_id.ppid = perf_event_pid(event, current);
3232 task_event->event_id.tid = perf_event_tid(event, task);
3233 task_event->event_id.ptid = perf_event_tid(event, current);
3235 task_event->event_id.time = perf_clock();
3237 perf_output_put(&handle, task_event->event_id);
3239 perf_output_end(&handle);
3242 static int perf_event_task_match(struct perf_event *event)
3244 if (event->attr.comm || event->attr.mmap || event->attr.task)
3250 static void perf_event_task_ctx(struct perf_event_context *ctx,
3251 struct perf_task_event *task_event)
3253 struct perf_event *event;
3255 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3256 if (perf_event_task_match(event))
3257 perf_event_task_output(event, task_event);
3261 static void perf_event_task_event(struct perf_task_event *task_event)
3263 struct perf_cpu_context *cpuctx;
3264 struct perf_event_context *ctx = task_event->task_ctx;
3267 cpuctx = &get_cpu_var(perf_cpu_context);
3268 perf_event_task_ctx(&cpuctx->ctx, task_event);
3269 put_cpu_var(perf_cpu_context);
3272 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3274 perf_event_task_ctx(ctx, task_event);
3278 static void perf_event_task(struct task_struct *task,
3279 struct perf_event_context *task_ctx,
3282 struct perf_task_event task_event;
3284 if (!atomic_read(&nr_comm_events) &&
3285 !atomic_read(&nr_mmap_events) &&
3286 !atomic_read(&nr_task_events))
3289 task_event = (struct perf_task_event){
3291 .task_ctx = task_ctx,
3294 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3296 .size = sizeof(task_event.event_id),
3305 perf_event_task_event(&task_event);
3308 void perf_event_fork(struct task_struct *task)
3310 perf_event_task(task, NULL, 1);
3317 struct perf_comm_event {
3318 struct task_struct *task;
3323 struct perf_event_header header;
3330 static void perf_event_comm_output(struct perf_event *event,
3331 struct perf_comm_event *comm_event)
3333 struct perf_output_handle handle;
3334 int size = comm_event->event_id.header.size;
3335 int ret = perf_output_begin(&handle, event, size, 0, 0);
3340 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3341 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3343 perf_output_put(&handle, comm_event->event_id);
3344 perf_output_copy(&handle, comm_event->comm,
3345 comm_event->comm_size);
3346 perf_output_end(&handle);
3349 static int perf_event_comm_match(struct perf_event *event)
3351 if (event->attr.comm)
3357 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3358 struct perf_comm_event *comm_event)
3360 struct perf_event *event;
3362 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3363 if (perf_event_comm_match(event))
3364 perf_event_comm_output(event, comm_event);
3368 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3370 struct perf_cpu_context *cpuctx;
3371 struct perf_event_context *ctx;
3373 char comm[TASK_COMM_LEN];
3375 memset(comm, 0, sizeof(comm));
3376 strncpy(comm, comm_event->task->comm, sizeof(comm));
3377 size = ALIGN(strlen(comm)+1, sizeof(u64));
3379 comm_event->comm = comm;
3380 comm_event->comm_size = size;
3382 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3385 cpuctx = &get_cpu_var(perf_cpu_context);
3386 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3387 put_cpu_var(perf_cpu_context);
3390 * doesn't really matter which of the child contexts the
3391 * events ends up in.
3393 ctx = rcu_dereference(current->perf_event_ctxp);
3395 perf_event_comm_ctx(ctx, comm_event);
3399 void perf_event_comm(struct task_struct *task)
3401 struct perf_comm_event comm_event;
3403 if (task->perf_event_ctxp)
3404 perf_event_enable_on_exec(task);
3406 if (!atomic_read(&nr_comm_events))
3409 comm_event = (struct perf_comm_event){
3415 .type = PERF_RECORD_COMM,
3424 perf_event_comm_event(&comm_event);
3431 struct perf_mmap_event {
3432 struct vm_area_struct *vma;
3434 const char *file_name;
3438 struct perf_event_header header;
3448 static void perf_event_mmap_output(struct perf_event *event,
3449 struct perf_mmap_event *mmap_event)
3451 struct perf_output_handle handle;
3452 int size = mmap_event->event_id.header.size;
3453 int ret = perf_output_begin(&handle, event, size, 0, 0);
3458 mmap_event->event_id.pid = perf_event_pid(event, current);
3459 mmap_event->event_id.tid = perf_event_tid(event, current);
3461 perf_output_put(&handle, mmap_event->event_id);
3462 perf_output_copy(&handle, mmap_event->file_name,
3463 mmap_event->file_size);
3464 perf_output_end(&handle);
3467 static int perf_event_mmap_match(struct perf_event *event,
3468 struct perf_mmap_event *mmap_event)
3470 if (event->attr.mmap)
3476 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3477 struct perf_mmap_event *mmap_event)
3479 struct perf_event *event;
3481 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3482 if (perf_event_mmap_match(event, mmap_event))
3483 perf_event_mmap_output(event, mmap_event);
3487 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3489 struct perf_cpu_context *cpuctx;
3490 struct perf_event_context *ctx;
3491 struct vm_area_struct *vma = mmap_event->vma;
3492 struct file *file = vma->vm_file;
3498 memset(tmp, 0, sizeof(tmp));
3502 * d_path works from the end of the buffer backwards, so we
3503 * need to add enough zero bytes after the string to handle
3504 * the 64bit alignment we do later.
3506 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3508 name = strncpy(tmp, "//enomem", sizeof(tmp));
3511 name = d_path(&file->f_path, buf, PATH_MAX);
3513 name = strncpy(tmp, "//toolong", sizeof(tmp));
3517 if (arch_vma_name(mmap_event->vma)) {
3518 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3524 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3528 name = strncpy(tmp, "//anon", sizeof(tmp));
3533 size = ALIGN(strlen(name)+1, sizeof(u64));
3535 mmap_event->file_name = name;
3536 mmap_event->file_size = size;
3538 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3541 cpuctx = &get_cpu_var(perf_cpu_context);
3542 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3543 put_cpu_var(perf_cpu_context);
3546 * doesn't really matter which of the child contexts the
3547 * events ends up in.
3549 ctx = rcu_dereference(current->perf_event_ctxp);
3551 perf_event_mmap_ctx(ctx, mmap_event);
3557 void __perf_event_mmap(struct vm_area_struct *vma)
3559 struct perf_mmap_event mmap_event;
3561 if (!atomic_read(&nr_mmap_events))
3564 mmap_event = (struct perf_mmap_event){
3570 .type = PERF_RECORD_MMAP,
3576 .start = vma->vm_start,
3577 .len = vma->vm_end - vma->vm_start,
3578 .pgoff = vma->vm_pgoff,
3582 perf_event_mmap_event(&mmap_event);
3586 * IRQ throttle logging
3589 static void perf_log_throttle(struct perf_event *event, int enable)
3591 struct perf_output_handle handle;
3595 struct perf_event_header header;
3599 } throttle_event = {
3601 .type = PERF_RECORD_THROTTLE,
3603 .size = sizeof(throttle_event),
3605 .time = perf_clock(),
3606 .id = primary_event_id(event),
3607 .stream_id = event->id,
3611 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3613 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3617 perf_output_put(&handle, throttle_event);
3618 perf_output_end(&handle);
3622 * Generic event overflow handling, sampling.
3625 static int __perf_event_overflow(struct perf_event *event, int nmi,
3626 int throttle, struct perf_sample_data *data,
3627 struct pt_regs *regs)
3629 int events = atomic_read(&event->event_limit);
3630 struct hw_perf_event *hwc = &event->hw;
3633 throttle = (throttle && event->pmu->unthrottle != NULL);
3638 if (hwc->interrupts != MAX_INTERRUPTS) {
3640 if (HZ * hwc->interrupts >
3641 (u64)sysctl_perf_event_sample_rate) {
3642 hwc->interrupts = MAX_INTERRUPTS;
3643 perf_log_throttle(event, 0);
3648 * Keep re-disabling events even though on the previous
3649 * pass we disabled it - just in case we raced with a
3650 * sched-in and the event got enabled again:
3656 if (event->attr.freq) {
3657 u64 now = perf_clock();
3658 s64 delta = now - hwc->freq_stamp;
3660 hwc->freq_stamp = now;
3662 if (delta > 0 && delta < TICK_NSEC)
3663 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3667 * XXX event_limit might not quite work as expected on inherited
3671 event->pending_kill = POLL_IN;
3672 if (events && atomic_dec_and_test(&event->event_limit)) {
3674 event->pending_kill = POLL_HUP;
3676 event->pending_disable = 1;
3677 perf_pending_queue(&event->pending,
3678 perf_pending_event);
3680 perf_event_disable(event);
3683 if (event->overflow_handler)
3684 event->overflow_handler(event, nmi, data, regs);
3686 perf_event_output(event, nmi, data, regs);
3691 int perf_event_overflow(struct perf_event *event, int nmi,
3692 struct perf_sample_data *data,
3693 struct pt_regs *regs)
3695 return __perf_event_overflow(event, nmi, 1, data, regs);
3699 * Generic software event infrastructure
3703 * We directly increment event->count and keep a second value in
3704 * event->hw.period_left to count intervals. This period event
3705 * is kept in the range [-sample_period, 0] so that we can use the
3709 static u64 perf_swevent_set_period(struct perf_event *event)
3711 struct hw_perf_event *hwc = &event->hw;
3712 u64 period = hwc->last_period;
3716 hwc->last_period = hwc->sample_period;
3719 old = val = atomic64_read(&hwc->period_left);
3723 nr = div64_u64(period + val, period);
3724 offset = nr * period;
3726 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3732 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3733 int nmi, struct perf_sample_data *data,
3734 struct pt_regs *regs)
3736 struct hw_perf_event *hwc = &event->hw;
3739 data->period = event->hw.last_period;
3741 overflow = perf_swevent_set_period(event);
3743 if (hwc->interrupts == MAX_INTERRUPTS)
3746 for (; overflow; overflow--) {
3747 if (__perf_event_overflow(event, nmi, throttle,
3750 * We inhibit the overflow from happening when
3751 * hwc->interrupts == MAX_INTERRUPTS.
3759 static void perf_swevent_unthrottle(struct perf_event *event)
3762 * Nothing to do, we already reset hwc->interrupts.
3766 static void perf_swevent_add(struct perf_event *event, u64 nr,
3767 int nmi, struct perf_sample_data *data,
3768 struct pt_regs *regs)
3770 struct hw_perf_event *hwc = &event->hw;
3772 atomic64_add(nr, &event->count);
3777 if (!hwc->sample_period)
3780 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3781 return perf_swevent_overflow(event, 1, nmi, data, regs);
3783 if (atomic64_add_negative(nr, &hwc->period_left))
3786 perf_swevent_overflow(event, 0, nmi, data, regs);
3789 static int perf_swevent_is_counting(struct perf_event *event)
3792 * The event is active, we're good!
3794 if (event->state == PERF_EVENT_STATE_ACTIVE)
3798 * The event is off/error, not counting.
3800 if (event->state != PERF_EVENT_STATE_INACTIVE)
3804 * The event is inactive, if the context is active
3805 * we're part of a group that didn't make it on the 'pmu',
3808 if (event->ctx->is_active)
3812 * We're inactive and the context is too, this means the
3813 * task is scheduled out, we're counting events that happen
3814 * to us, like migration events.
3819 static int perf_tp_event_match(struct perf_event *event,
3820 struct perf_sample_data *data);
3822 static int perf_swevent_match(struct perf_event *event,
3823 enum perf_type_id type,
3825 struct perf_sample_data *data,
3826 struct pt_regs *regs)
3828 if (!perf_swevent_is_counting(event))
3831 if (event->attr.type != type)
3833 if (event->attr.config != event_id)
3837 if (event->attr.exclude_user && user_mode(regs))
3840 if (event->attr.exclude_kernel && !user_mode(regs))
3844 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3845 !perf_tp_event_match(event, data))
3851 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3852 enum perf_type_id type,
3853 u32 event_id, u64 nr, int nmi,
3854 struct perf_sample_data *data,
3855 struct pt_regs *regs)
3857 struct perf_event *event;
3859 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3860 if (perf_swevent_match(event, type, event_id, data, regs))
3861 perf_swevent_add(event, nr, nmi, data, regs);
3865 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3868 return &cpuctx->recursion[3];
3871 return &cpuctx->recursion[2];
3874 return &cpuctx->recursion[1];
3876 return &cpuctx->recursion[0];
3879 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3881 struct perf_sample_data *data,
3882 struct pt_regs *regs)
3884 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3885 int *recursion = perf_swevent_recursion_context(cpuctx);
3886 struct perf_event_context *ctx;
3895 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3896 nr, nmi, data, regs);
3898 * doesn't really matter which of the child contexts the
3899 * events ends up in.
3901 ctx = rcu_dereference(current->perf_event_ctxp);
3903 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3910 put_cpu_var(perf_cpu_context);
3913 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3914 struct pt_regs *regs, u64 addr)
3916 struct perf_sample_data data = {
3920 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3924 static void perf_swevent_read(struct perf_event *event)
3928 static int perf_swevent_enable(struct perf_event *event)
3930 struct hw_perf_event *hwc = &event->hw;
3932 if (hwc->sample_period) {
3933 hwc->last_period = hwc->sample_period;
3934 perf_swevent_set_period(event);
3939 static void perf_swevent_disable(struct perf_event *event)
3943 static const struct pmu perf_ops_generic = {
3944 .enable = perf_swevent_enable,
3945 .disable = perf_swevent_disable,
3946 .read = perf_swevent_read,
3947 .unthrottle = perf_swevent_unthrottle,
3951 * hrtimer based swevent callback
3954 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3956 enum hrtimer_restart ret = HRTIMER_RESTART;
3957 struct perf_sample_data data;
3958 struct pt_regs *regs;
3959 struct perf_event *event;
3962 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
3963 event->pmu->read(event);
3966 regs = get_irq_regs();
3968 * In case we exclude kernel IPs or are somehow not in interrupt
3969 * context, provide the next best thing, the user IP.
3971 if ((event->attr.exclude_kernel || !regs) &&
3972 !event->attr.exclude_user)
3973 regs = task_pt_regs(current);
3976 if (!(event->attr.exclude_idle && current->pid == 0))
3977 if (perf_event_overflow(event, 0, &data, regs))
3978 ret = HRTIMER_NORESTART;
3981 period = max_t(u64, 10000, event->hw.sample_period);
3982 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3987 static void perf_swevent_start_hrtimer(struct perf_event *event)
3989 struct hw_perf_event *hwc = &event->hw;
3991 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3992 hwc->hrtimer.function = perf_swevent_hrtimer;
3993 if (hwc->sample_period) {
3996 if (hwc->remaining) {
3997 if (hwc->remaining < 0)
4000 period = hwc->remaining;
4003 period = max_t(u64, 10000, hwc->sample_period);
4005 __hrtimer_start_range_ns(&hwc->hrtimer,
4006 ns_to_ktime(period), 0,
4007 HRTIMER_MODE_REL, 0);
4011 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4013 struct hw_perf_event *hwc = &event->hw;
4015 if (hwc->sample_period) {
4016 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4017 hwc->remaining = ktime_to_ns(remaining);
4019 hrtimer_cancel(&hwc->hrtimer);
4024 * Software event: cpu wall time clock
4027 static void cpu_clock_perf_event_update(struct perf_event *event)
4029 int cpu = raw_smp_processor_id();
4033 now = cpu_clock(cpu);
4034 prev = atomic64_read(&event->hw.prev_count);
4035 atomic64_set(&event->hw.prev_count, now);
4036 atomic64_add(now - prev, &event->count);
4039 static int cpu_clock_perf_event_enable(struct perf_event *event)
4041 struct hw_perf_event *hwc = &event->hw;
4042 int cpu = raw_smp_processor_id();
4044 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4045 perf_swevent_start_hrtimer(event);
4050 static void cpu_clock_perf_event_disable(struct perf_event *event)
4052 perf_swevent_cancel_hrtimer(event);
4053 cpu_clock_perf_event_update(event);
4056 static void cpu_clock_perf_event_read(struct perf_event *event)
4058 cpu_clock_perf_event_update(event);
4061 static const struct pmu perf_ops_cpu_clock = {
4062 .enable = cpu_clock_perf_event_enable,
4063 .disable = cpu_clock_perf_event_disable,
4064 .read = cpu_clock_perf_event_read,
4068 * Software event: task time clock
4071 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4076 prev = atomic64_xchg(&event->hw.prev_count, now);
4078 atomic64_add(delta, &event->count);
4081 static int task_clock_perf_event_enable(struct perf_event *event)
4083 struct hw_perf_event *hwc = &event->hw;
4086 now = event->ctx->time;
4088 atomic64_set(&hwc->prev_count, now);
4090 perf_swevent_start_hrtimer(event);
4095 static void task_clock_perf_event_disable(struct perf_event *event)
4097 perf_swevent_cancel_hrtimer(event);
4098 task_clock_perf_event_update(event, event->ctx->time);
4102 static void task_clock_perf_event_read(struct perf_event *event)
4107 update_context_time(event->ctx);
4108 time = event->ctx->time;
4110 u64 now = perf_clock();
4111 u64 delta = now - event->ctx->timestamp;
4112 time = event->ctx->time + delta;
4115 task_clock_perf_event_update(event, time);
4118 static const struct pmu perf_ops_task_clock = {
4119 .enable = task_clock_perf_event_enable,
4120 .disable = task_clock_perf_event_disable,
4121 .read = task_clock_perf_event_read,
4124 #ifdef CONFIG_EVENT_PROFILE
4126 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4129 struct perf_raw_record raw = {
4134 struct perf_sample_data data = {
4139 struct pt_regs *regs = get_irq_regs();
4142 regs = task_pt_regs(current);
4144 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4147 EXPORT_SYMBOL_GPL(perf_tp_event);
4149 static int perf_tp_event_match(struct perf_event *event,
4150 struct perf_sample_data *data)
4152 void *record = data->raw->data;
4154 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4159 static void tp_perf_event_destroy(struct perf_event *event)
4161 ftrace_profile_disable(event->attr.config);
4164 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4167 * Raw tracepoint data is a severe data leak, only allow root to
4170 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4171 perf_paranoid_tracepoint_raw() &&
4172 !capable(CAP_SYS_ADMIN))
4173 return ERR_PTR(-EPERM);
4175 if (ftrace_profile_enable(event->attr.config))
4178 event->destroy = tp_perf_event_destroy;
4180 return &perf_ops_generic;
4183 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4188 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4191 filter_str = strndup_user(arg, PAGE_SIZE);
4192 if (IS_ERR(filter_str))
4193 return PTR_ERR(filter_str);
4195 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4201 static void perf_event_free_filter(struct perf_event *event)
4203 ftrace_profile_free_filter(event);
4208 static int perf_tp_event_match(struct perf_event *event,
4209 struct perf_sample_data *data)
4214 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4219 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4224 static void perf_event_free_filter(struct perf_event *event)
4228 #endif /* CONFIG_EVENT_PROFILE */
4230 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4231 static void bp_perf_event_destroy(struct perf_event *event)
4233 release_bp_slot(event);
4236 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4240 * The breakpoint is already filled if we haven't created the counter
4241 * through perf syscall
4242 * FIXME: manage to get trigerred to NULL if it comes from syscalls
4245 err = register_perf_hw_breakpoint(bp);
4247 err = __register_perf_hw_breakpoint(bp);
4249 return ERR_PTR(err);
4251 bp->destroy = bp_perf_event_destroy;
4253 return &perf_ops_bp;
4256 void perf_bp_event(struct perf_event *bp, void *regs)
4261 static void bp_perf_event_destroy(struct perf_event *event)
4265 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4270 void perf_bp_event(struct perf_event *bp, void *regs)
4275 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4277 static void sw_perf_event_destroy(struct perf_event *event)
4279 u64 event_id = event->attr.config;
4281 WARN_ON(event->parent);
4283 atomic_dec(&perf_swevent_enabled[event_id]);
4286 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4288 const struct pmu *pmu = NULL;
4289 u64 event_id = event->attr.config;
4292 * Software events (currently) can't in general distinguish
4293 * between user, kernel and hypervisor events.
4294 * However, context switches and cpu migrations are considered
4295 * to be kernel events, and page faults are never hypervisor
4299 case PERF_COUNT_SW_CPU_CLOCK:
4300 pmu = &perf_ops_cpu_clock;
4303 case PERF_COUNT_SW_TASK_CLOCK:
4305 * If the user instantiates this as a per-cpu event,
4306 * use the cpu_clock event instead.
4308 if (event->ctx->task)
4309 pmu = &perf_ops_task_clock;
4311 pmu = &perf_ops_cpu_clock;
4314 case PERF_COUNT_SW_PAGE_FAULTS:
4315 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4316 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4317 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4318 case PERF_COUNT_SW_CPU_MIGRATIONS:
4319 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4320 case PERF_COUNT_SW_EMULATION_FAULTS:
4321 if (!event->parent) {
4322 atomic_inc(&perf_swevent_enabled[event_id]);
4323 event->destroy = sw_perf_event_destroy;
4325 pmu = &perf_ops_generic;
4333 * Allocate and initialize a event structure
4335 static struct perf_event *
4336 perf_event_alloc(struct perf_event_attr *attr,
4338 struct perf_event_context *ctx,
4339 struct perf_event *group_leader,
4340 struct perf_event *parent_event,
4341 perf_callback_t callback,
4344 const struct pmu *pmu;
4345 struct perf_event *event;
4346 struct hw_perf_event *hwc;
4349 event = kzalloc(sizeof(*event), gfpflags);
4351 return ERR_PTR(-ENOMEM);
4354 * Single events are their own group leaders, with an
4355 * empty sibling list:
4358 group_leader = event;
4360 mutex_init(&event->child_mutex);
4361 INIT_LIST_HEAD(&event->child_list);
4363 INIT_LIST_HEAD(&event->group_entry);
4364 INIT_LIST_HEAD(&event->event_entry);
4365 INIT_LIST_HEAD(&event->sibling_list);
4366 init_waitqueue_head(&event->waitq);
4368 mutex_init(&event->mmap_mutex);
4371 event->attr = *attr;
4372 event->group_leader = group_leader;
4377 event->parent = parent_event;
4379 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4380 event->id = atomic64_inc_return(&perf_event_id);
4382 event->state = PERF_EVENT_STATE_INACTIVE;
4384 if (!callback && parent_event)
4385 callback = parent_event->callback;
4387 event->callback = callback;
4390 event->state = PERF_EVENT_STATE_OFF;
4395 hwc->sample_period = attr->sample_period;
4396 if (attr->freq && attr->sample_freq)
4397 hwc->sample_period = 1;
4398 hwc->last_period = hwc->sample_period;
4400 atomic64_set(&hwc->period_left, hwc->sample_period);
4403 * we currently do not support PERF_FORMAT_GROUP on inherited events
4405 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4408 switch (attr->type) {
4410 case PERF_TYPE_HARDWARE:
4411 case PERF_TYPE_HW_CACHE:
4412 pmu = hw_perf_event_init(event);
4415 case PERF_TYPE_SOFTWARE:
4416 pmu = sw_perf_event_init(event);
4419 case PERF_TYPE_TRACEPOINT:
4420 pmu = tp_perf_event_init(event);
4423 case PERF_TYPE_BREAKPOINT:
4424 pmu = bp_perf_event_init(event);
4435 else if (IS_ERR(pmu))
4440 put_pid_ns(event->ns);
4442 return ERR_PTR(err);
4447 if (!event->parent) {
4448 atomic_inc(&nr_events);
4449 if (event->attr.mmap)
4450 atomic_inc(&nr_mmap_events);
4451 if (event->attr.comm)
4452 atomic_inc(&nr_comm_events);
4453 if (event->attr.task)
4454 atomic_inc(&nr_task_events);
4460 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4461 struct perf_event_attr *attr)
4466 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4470 * zero the full structure, so that a short copy will be nice.
4472 memset(attr, 0, sizeof(*attr));
4474 ret = get_user(size, &uattr->size);
4478 if (size > PAGE_SIZE) /* silly large */
4481 if (!size) /* abi compat */
4482 size = PERF_ATTR_SIZE_VER0;
4484 if (size < PERF_ATTR_SIZE_VER0)
4488 * If we're handed a bigger struct than we know of,
4489 * ensure all the unknown bits are 0 - i.e. new
4490 * user-space does not rely on any kernel feature
4491 * extensions we dont know about yet.
4493 if (size > sizeof(*attr)) {
4494 unsigned char __user *addr;
4495 unsigned char __user *end;
4498 addr = (void __user *)uattr + sizeof(*attr);
4499 end = (void __user *)uattr + size;
4501 for (; addr < end; addr++) {
4502 ret = get_user(val, addr);
4508 size = sizeof(*attr);
4511 ret = copy_from_user(attr, uattr, size);
4516 * If the type exists, the corresponding creation will verify
4519 if (attr->type >= PERF_TYPE_MAX)
4522 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4525 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4528 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4535 put_user(sizeof(*attr), &uattr->size);
4540 static int perf_event_set_output(struct perf_event *event, int output_fd)
4542 struct perf_event *output_event = NULL;
4543 struct file *output_file = NULL;
4544 struct perf_event *old_output;
4545 int fput_needed = 0;
4551 output_file = fget_light(output_fd, &fput_needed);
4555 if (output_file->f_op != &perf_fops)
4558 output_event = output_file->private_data;
4560 /* Don't chain output fds */
4561 if (output_event->output)
4564 /* Don't set an output fd when we already have an output channel */
4568 atomic_long_inc(&output_file->f_count);
4571 mutex_lock(&event->mmap_mutex);
4572 old_output = event->output;
4573 rcu_assign_pointer(event->output, output_event);
4574 mutex_unlock(&event->mmap_mutex);
4578 * we need to make sure no existing perf_output_*()
4579 * is still referencing this event.
4582 fput(old_output->filp);
4587 fput_light(output_file, fput_needed);
4592 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4594 * @attr_uptr: event_id type attributes for monitoring/sampling
4597 * @group_fd: group leader event fd
4599 SYSCALL_DEFINE5(perf_event_open,
4600 struct perf_event_attr __user *, attr_uptr,
4601 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4603 struct perf_event *event, *group_leader;
4604 struct perf_event_attr attr;
4605 struct perf_event_context *ctx;
4606 struct file *event_file = NULL;
4607 struct file *group_file = NULL;
4608 int fput_needed = 0;
4609 int fput_needed2 = 0;
4612 /* for future expandability... */
4613 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4616 err = perf_copy_attr(attr_uptr, &attr);
4620 if (!attr.exclude_kernel) {
4621 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4626 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4631 * Get the target context (task or percpu):
4633 ctx = find_get_context(pid, cpu);
4635 return PTR_ERR(ctx);
4638 * Look up the group leader (we will attach this event to it):
4640 group_leader = NULL;
4641 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4643 group_file = fget_light(group_fd, &fput_needed);
4645 goto err_put_context;
4646 if (group_file->f_op != &perf_fops)
4647 goto err_put_context;
4649 group_leader = group_file->private_data;
4651 * Do not allow a recursive hierarchy (this new sibling
4652 * becoming part of another group-sibling):
4654 if (group_leader->group_leader != group_leader)
4655 goto err_put_context;
4657 * Do not allow to attach to a group in a different
4658 * task or CPU context:
4660 if (group_leader->ctx != ctx)
4661 goto err_put_context;
4663 * Only a group leader can be exclusive or pinned
4665 if (attr.exclusive || attr.pinned)
4666 goto err_put_context;
4669 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4670 NULL, NULL, GFP_KERNEL);
4671 err = PTR_ERR(event);
4673 goto err_put_context;
4675 err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4677 goto err_free_put_context;
4679 event_file = fget_light(err, &fput_needed2);
4681 goto err_free_put_context;
4683 if (flags & PERF_FLAG_FD_OUTPUT) {
4684 err = perf_event_set_output(event, group_fd);
4686 goto err_fput_free_put_context;
4689 event->filp = event_file;
4690 WARN_ON_ONCE(ctx->parent_ctx);
4691 mutex_lock(&ctx->mutex);
4692 perf_install_in_context(ctx, event, cpu);
4694 mutex_unlock(&ctx->mutex);
4696 event->owner = current;
4697 get_task_struct(current);
4698 mutex_lock(¤t->perf_event_mutex);
4699 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4700 mutex_unlock(¤t->perf_event_mutex);
4702 err_fput_free_put_context:
4703 fput_light(event_file, fput_needed2);
4705 err_free_put_context:
4713 fput_light(group_file, fput_needed);
4719 * perf_event_create_kernel_counter
4721 * @attr: attributes of the counter to create
4722 * @cpu: cpu in which the counter is bound
4723 * @pid: task to profile
4726 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4727 pid_t pid, perf_callback_t callback)
4729 struct perf_event *event;
4730 struct perf_event_context *ctx;
4734 * Get the target context (task or percpu):
4737 ctx = find_get_context(pid, cpu);
4741 event = perf_event_alloc(attr, cpu, ctx, NULL,
4742 NULL, callback, GFP_KERNEL);
4743 err = PTR_ERR(event);
4745 goto err_put_context;
4748 WARN_ON_ONCE(ctx->parent_ctx);
4749 mutex_lock(&ctx->mutex);
4750 perf_install_in_context(ctx, event, cpu);
4752 mutex_unlock(&ctx->mutex);
4754 event->owner = current;
4755 get_task_struct(current);
4756 mutex_lock(¤t->perf_event_mutex);
4757 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4758 mutex_unlock(¤t->perf_event_mutex);
4768 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4771 * inherit a event from parent task to child task:
4773 static struct perf_event *
4774 inherit_event(struct perf_event *parent_event,
4775 struct task_struct *parent,
4776 struct perf_event_context *parent_ctx,
4777 struct task_struct *child,
4778 struct perf_event *group_leader,
4779 struct perf_event_context *child_ctx)
4781 struct perf_event *child_event;
4784 * Instead of creating recursive hierarchies of events,
4785 * we link inherited events back to the original parent,
4786 * which has a filp for sure, which we use as the reference
4789 if (parent_event->parent)
4790 parent_event = parent_event->parent;
4792 child_event = perf_event_alloc(&parent_event->attr,
4793 parent_event->cpu, child_ctx,
4794 group_leader, parent_event,
4796 if (IS_ERR(child_event))
4801 * Make the child state follow the state of the parent event,
4802 * not its attr.disabled bit. We hold the parent's mutex,
4803 * so we won't race with perf_event_{en, dis}able_family.
4805 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4806 child_event->state = PERF_EVENT_STATE_INACTIVE;
4808 child_event->state = PERF_EVENT_STATE_OFF;
4810 if (parent_event->attr.freq)
4811 child_event->hw.sample_period = parent_event->hw.sample_period;
4813 child_event->overflow_handler = parent_event->overflow_handler;
4816 * Link it up in the child's context:
4818 add_event_to_ctx(child_event, child_ctx);
4821 * Get a reference to the parent filp - we will fput it
4822 * when the child event exits. This is safe to do because
4823 * we are in the parent and we know that the filp still
4824 * exists and has a nonzero count:
4826 atomic_long_inc(&parent_event->filp->f_count);
4829 * Link this into the parent event's child list
4831 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4832 mutex_lock(&parent_event->child_mutex);
4833 list_add_tail(&child_event->child_list, &parent_event->child_list);
4834 mutex_unlock(&parent_event->child_mutex);
4839 static int inherit_group(struct perf_event *parent_event,
4840 struct task_struct *parent,
4841 struct perf_event_context *parent_ctx,
4842 struct task_struct *child,
4843 struct perf_event_context *child_ctx)
4845 struct perf_event *leader;
4846 struct perf_event *sub;
4847 struct perf_event *child_ctr;
4849 leader = inherit_event(parent_event, parent, parent_ctx,
4850 child, NULL, child_ctx);
4852 return PTR_ERR(leader);
4853 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4854 child_ctr = inherit_event(sub, parent, parent_ctx,
4855 child, leader, child_ctx);
4856 if (IS_ERR(child_ctr))
4857 return PTR_ERR(child_ctr);
4862 static void sync_child_event(struct perf_event *child_event,
4863 struct task_struct *child)
4865 struct perf_event *parent_event = child_event->parent;
4868 if (child_event->attr.inherit_stat)
4869 perf_event_read_event(child_event, child);
4871 child_val = atomic64_read(&child_event->count);
4874 * Add back the child's count to the parent's count:
4876 atomic64_add(child_val, &parent_event->count);
4877 atomic64_add(child_event->total_time_enabled,
4878 &parent_event->child_total_time_enabled);
4879 atomic64_add(child_event->total_time_running,
4880 &parent_event->child_total_time_running);
4883 * Remove this event from the parent's list
4885 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4886 mutex_lock(&parent_event->child_mutex);
4887 list_del_init(&child_event->child_list);
4888 mutex_unlock(&parent_event->child_mutex);
4891 * Release the parent event, if this was the last
4894 fput(parent_event->filp);
4898 __perf_event_exit_task(struct perf_event *child_event,
4899 struct perf_event_context *child_ctx,
4900 struct task_struct *child)
4902 struct perf_event *parent_event;
4904 update_event_times(child_event);
4905 perf_event_remove_from_context(child_event);
4907 parent_event = child_event->parent;
4909 * It can happen that parent exits first, and has events
4910 * that are still around due to the child reference. These
4911 * events need to be zapped - but otherwise linger.
4914 sync_child_event(child_event, child);
4915 free_event(child_event);
4920 * When a child task exits, feed back event values to parent events.
4922 void perf_event_exit_task(struct task_struct *child)
4924 struct perf_event *child_event, *tmp;
4925 struct perf_event_context *child_ctx;
4926 unsigned long flags;
4928 if (likely(!child->perf_event_ctxp)) {
4929 perf_event_task(child, NULL, 0);
4933 local_irq_save(flags);
4935 * We can't reschedule here because interrupts are disabled,
4936 * and either child is current or it is a task that can't be
4937 * scheduled, so we are now safe from rescheduling changing
4940 child_ctx = child->perf_event_ctxp;
4941 __perf_event_task_sched_out(child_ctx);
4944 * Take the context lock here so that if find_get_context is
4945 * reading child->perf_event_ctxp, we wait until it has
4946 * incremented the context's refcount before we do put_ctx below.
4948 spin_lock(&child_ctx->lock);
4949 child->perf_event_ctxp = NULL;
4951 * If this context is a clone; unclone it so it can't get
4952 * swapped to another process while we're removing all
4953 * the events from it.
4955 unclone_ctx(child_ctx);
4956 spin_unlock_irqrestore(&child_ctx->lock, flags);
4959 * Report the task dead after unscheduling the events so that we
4960 * won't get any samples after PERF_RECORD_EXIT. We can however still
4961 * get a few PERF_RECORD_READ events.
4963 perf_event_task(child, child_ctx, 0);
4966 * We can recurse on the same lock type through:
4968 * __perf_event_exit_task()
4969 * sync_child_event()
4970 * fput(parent_event->filp)
4972 * mutex_lock(&ctx->mutex)
4974 * But since its the parent context it won't be the same instance.
4976 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4979 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4981 __perf_event_exit_task(child_event, child_ctx, child);
4984 * If the last event was a group event, it will have appended all
4985 * its siblings to the list, but we obtained 'tmp' before that which
4986 * will still point to the list head terminating the iteration.
4988 if (!list_empty(&child_ctx->group_list))
4991 mutex_unlock(&child_ctx->mutex);
4997 * free an unexposed, unused context as created by inheritance by
4998 * init_task below, used by fork() in case of fail.
5000 void perf_event_free_task(struct task_struct *task)
5002 struct perf_event_context *ctx = task->perf_event_ctxp;
5003 struct perf_event *event, *tmp;
5008 mutex_lock(&ctx->mutex);
5010 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5011 struct perf_event *parent = event->parent;
5013 if (WARN_ON_ONCE(!parent))
5016 mutex_lock(&parent->child_mutex);
5017 list_del_init(&event->child_list);
5018 mutex_unlock(&parent->child_mutex);
5022 list_del_event(event, ctx);
5026 if (!list_empty(&ctx->group_list))
5029 mutex_unlock(&ctx->mutex);
5035 * Initialize the perf_event context in task_struct
5037 int perf_event_init_task(struct task_struct *child)
5039 struct perf_event_context *child_ctx, *parent_ctx;
5040 struct perf_event_context *cloned_ctx;
5041 struct perf_event *event;
5042 struct task_struct *parent = current;
5043 int inherited_all = 1;
5046 child->perf_event_ctxp = NULL;
5048 mutex_init(&child->perf_event_mutex);
5049 INIT_LIST_HEAD(&child->perf_event_list);
5051 if (likely(!parent->perf_event_ctxp))
5055 * This is executed from the parent task context, so inherit
5056 * events that have been marked for cloning.
5057 * First allocate and initialize a context for the child.
5060 child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5064 __perf_event_init_context(child_ctx, child);
5065 child->perf_event_ctxp = child_ctx;
5066 get_task_struct(child);
5069 * If the parent's context is a clone, pin it so it won't get
5072 parent_ctx = perf_pin_task_context(parent);
5075 * No need to check if parent_ctx != NULL here; since we saw
5076 * it non-NULL earlier, the only reason for it to become NULL
5077 * is if we exit, and since we're currently in the middle of
5078 * a fork we can't be exiting at the same time.
5082 * Lock the parent list. No need to lock the child - not PID
5083 * hashed yet and not running, so nobody can access it.
5085 mutex_lock(&parent_ctx->mutex);
5088 * We dont have to disable NMIs - we are only looking at
5089 * the list, not manipulating it:
5091 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5093 if (!event->attr.inherit) {
5098 ret = inherit_group(event, parent, parent_ctx,
5106 if (inherited_all) {
5108 * Mark the child context as a clone of the parent
5109 * context, or of whatever the parent is a clone of.
5110 * Note that if the parent is a clone, it could get
5111 * uncloned at any point, but that doesn't matter
5112 * because the list of events and the generation
5113 * count can't have changed since we took the mutex.
5115 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5117 child_ctx->parent_ctx = cloned_ctx;
5118 child_ctx->parent_gen = parent_ctx->parent_gen;
5120 child_ctx->parent_ctx = parent_ctx;
5121 child_ctx->parent_gen = parent_ctx->generation;
5123 get_ctx(child_ctx->parent_ctx);
5126 mutex_unlock(&parent_ctx->mutex);
5128 perf_unpin_context(parent_ctx);
5133 static void __cpuinit perf_event_init_cpu(int cpu)
5135 struct perf_cpu_context *cpuctx;
5137 cpuctx = &per_cpu(perf_cpu_context, cpu);
5138 __perf_event_init_context(&cpuctx->ctx, NULL);
5140 spin_lock(&perf_resource_lock);
5141 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5142 spin_unlock(&perf_resource_lock);
5144 hw_perf_event_setup(cpu);
5147 #ifdef CONFIG_HOTPLUG_CPU
5148 static void __perf_event_exit_cpu(void *info)
5150 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5151 struct perf_event_context *ctx = &cpuctx->ctx;
5152 struct perf_event *event, *tmp;
5154 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5155 __perf_event_remove_from_context(event);
5157 static void perf_event_exit_cpu(int cpu)
5159 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5160 struct perf_event_context *ctx = &cpuctx->ctx;
5162 mutex_lock(&ctx->mutex);
5163 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5164 mutex_unlock(&ctx->mutex);
5167 static inline void perf_event_exit_cpu(int cpu) { }
5170 static int __cpuinit
5171 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5173 unsigned int cpu = (long)hcpu;
5177 case CPU_UP_PREPARE:
5178 case CPU_UP_PREPARE_FROZEN:
5179 perf_event_init_cpu(cpu);
5183 case CPU_ONLINE_FROZEN:
5184 hw_perf_event_setup_online(cpu);
5187 case CPU_DOWN_PREPARE:
5188 case CPU_DOWN_PREPARE_FROZEN:
5189 perf_event_exit_cpu(cpu);
5200 * This has to have a higher priority than migration_notifier in sched.c.
5202 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5203 .notifier_call = perf_cpu_notify,
5207 void __init perf_event_init(void)
5209 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5210 (void *)(long)smp_processor_id());
5211 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5212 (void *)(long)smp_processor_id());
5213 register_cpu_notifier(&perf_cpu_nb);
5216 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5218 return sprintf(buf, "%d\n", perf_reserved_percpu);
5222 perf_set_reserve_percpu(struct sysdev_class *class,
5226 struct perf_cpu_context *cpuctx;
5230 err = strict_strtoul(buf, 10, &val);
5233 if (val > perf_max_events)
5236 spin_lock(&perf_resource_lock);
5237 perf_reserved_percpu = val;
5238 for_each_online_cpu(cpu) {
5239 cpuctx = &per_cpu(perf_cpu_context, cpu);
5240 spin_lock_irq(&cpuctx->ctx.lock);
5241 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5242 perf_max_events - perf_reserved_percpu);
5243 cpuctx->max_pertask = mpt;
5244 spin_unlock_irq(&cpuctx->ctx.lock);
5246 spin_unlock(&perf_resource_lock);
5251 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5253 return sprintf(buf, "%d\n", perf_overcommit);
5257 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5262 err = strict_strtoul(buf, 10, &val);
5268 spin_lock(&perf_resource_lock);
5269 perf_overcommit = val;
5270 spin_unlock(&perf_resource_lock);
5275 static SYSDEV_CLASS_ATTR(
5278 perf_show_reserve_percpu,
5279 perf_set_reserve_percpu
5282 static SYSDEV_CLASS_ATTR(
5285 perf_show_overcommit,
5289 static struct attribute *perfclass_attrs[] = {
5290 &attr_reserve_percpu.attr,
5291 &attr_overcommit.attr,
5295 static struct attribute_group perfclass_attr_group = {
5296 .attrs = perfclass_attrs,
5297 .name = "perf_events",
5300 static int __init perf_event_sysfs_init(void)
5302 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5303 &perfclass_attr_group);
5305 device_initcall(perf_event_sysfs_init);