2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/vmalloc.h>
24 #include <linux/hardirq.h>
25 #include <linux/rculist.h>
26 #include <linux/uaccess.h>
27 #include <linux/syscalls.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/perf_event.h>
31 #include <linux/ftrace_event.h>
32 #include <linux/hw_breakpoint.h>
34 #include <asm/irq_regs.h>
37 * Each CPU has a list of per CPU events:
39 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
41 int perf_max_events __read_mostly = 1;
42 static int perf_reserved_percpu __read_mostly;
43 static int perf_overcommit __read_mostly = 1;
45 static atomic_t nr_events __read_mostly;
46 static atomic_t nr_mmap_events __read_mostly;
47 static atomic_t nr_comm_events __read_mostly;
48 static atomic_t nr_task_events __read_mostly;
51 * perf event paranoia level:
52 * -1 - not paranoid at all
53 * 0 - disallow raw tracepoint access for unpriv
54 * 1 - disallow cpu events for unpriv
55 * 2 - disallow kernel profiling for unpriv
57 int sysctl_perf_event_paranoid __read_mostly = 1;
59 static inline bool perf_paranoid_tracepoint_raw(void)
61 return sysctl_perf_event_paranoid > -1;
64 static inline bool perf_paranoid_cpu(void)
66 return sysctl_perf_event_paranoid > 0;
69 static inline bool perf_paranoid_kernel(void)
71 return sysctl_perf_event_paranoid > 1;
74 int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
77 * max perf event sample rate
79 int sysctl_perf_event_sample_rate __read_mostly = 100000;
81 static atomic64_t perf_event_id;
84 * Lock for (sysadmin-configurable) event reservations:
86 static DEFINE_SPINLOCK(perf_resource_lock);
89 * Architecture provided APIs - weak aliases:
91 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
96 void __weak hw_perf_disable(void) { barrier(); }
97 void __weak hw_perf_enable(void) { barrier(); }
99 void __weak hw_perf_event_setup(int cpu) { barrier(); }
100 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
103 hw_perf_group_sched_in(struct perf_event *group_leader,
104 struct perf_cpu_context *cpuctx,
105 struct perf_event_context *ctx, int cpu)
110 void __weak perf_event_print_debug(void) { }
112 static DEFINE_PER_CPU(int, perf_disable_count);
114 void __perf_disable(void)
116 __get_cpu_var(perf_disable_count)++;
119 bool __perf_enable(void)
121 return !--__get_cpu_var(perf_disable_count);
124 void perf_disable(void)
130 void perf_enable(void)
136 static void get_ctx(struct perf_event_context *ctx)
138 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
141 static void free_ctx(struct rcu_head *head)
143 struct perf_event_context *ctx;
145 ctx = container_of(head, struct perf_event_context, rcu_head);
149 static void put_ctx(struct perf_event_context *ctx)
151 if (atomic_dec_and_test(&ctx->refcount)) {
153 put_ctx(ctx->parent_ctx);
155 put_task_struct(ctx->task);
156 call_rcu(&ctx->rcu_head, free_ctx);
160 static void unclone_ctx(struct perf_event_context *ctx)
162 if (ctx->parent_ctx) {
163 put_ctx(ctx->parent_ctx);
164 ctx->parent_ctx = NULL;
169 * If we inherit events we want to return the parent event id
172 static u64 primary_event_id(struct perf_event *event)
177 id = event->parent->id;
183 * Get the perf_event_context for a task and lock it.
184 * This has to cope with with the fact that until it is locked,
185 * the context could get moved to another task.
187 static struct perf_event_context *
188 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
190 struct perf_event_context *ctx;
194 ctx = rcu_dereference(task->perf_event_ctxp);
197 * If this context is a clone of another, it might
198 * get swapped for another underneath us by
199 * perf_event_task_sched_out, though the
200 * rcu_read_lock() protects us from any context
201 * getting freed. Lock the context and check if it
202 * got swapped before we could get the lock, and retry
203 * if so. If we locked the right context, then it
204 * can't get swapped on us any more.
206 spin_lock_irqsave(&ctx->lock, *flags);
207 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208 spin_unlock_irqrestore(&ctx->lock, *flags);
212 if (!atomic_inc_not_zero(&ctx->refcount)) {
213 spin_unlock_irqrestore(&ctx->lock, *flags);
222 * Get the context for a task and increment its pin_count so it
223 * can't get swapped to another task. This also increments its
224 * reference count so that the context can't get freed.
226 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
228 struct perf_event_context *ctx;
231 ctx = perf_lock_task_context(task, &flags);
234 spin_unlock_irqrestore(&ctx->lock, flags);
239 static void perf_unpin_context(struct perf_event_context *ctx)
243 spin_lock_irqsave(&ctx->lock, flags);
245 spin_unlock_irqrestore(&ctx->lock, flags);
250 * Add a event from the lists for its context.
251 * Must be called with ctx->mutex and ctx->lock held.
254 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
256 struct perf_event *group_leader = event->group_leader;
259 * Depending on whether it is a standalone or sibling event,
260 * add it straight to the context's event list, or to the group
261 * leader's sibling list:
263 if (group_leader == event)
264 list_add_tail(&event->group_entry, &ctx->group_list);
266 list_add_tail(&event->group_entry, &group_leader->sibling_list);
267 group_leader->nr_siblings++;
270 list_add_rcu(&event->event_entry, &ctx->event_list);
272 if (event->attr.inherit_stat)
277 * Remove a event from the lists for its context.
278 * Must be called with ctx->mutex and ctx->lock held.
281 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
283 struct perf_event *sibling, *tmp;
285 if (list_empty(&event->group_entry))
288 if (event->attr.inherit_stat)
291 list_del_init(&event->group_entry);
292 list_del_rcu(&event->event_entry);
294 if (event->group_leader != event)
295 event->group_leader->nr_siblings--;
298 * If this was a group event with sibling events then
299 * upgrade the siblings to singleton events by adding them
300 * to the context list directly:
302 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
304 list_move_tail(&sibling->group_entry, &ctx->group_list);
305 sibling->group_leader = sibling;
310 event_sched_out(struct perf_event *event,
311 struct perf_cpu_context *cpuctx,
312 struct perf_event_context *ctx)
314 if (event->state != PERF_EVENT_STATE_ACTIVE)
317 event->state = PERF_EVENT_STATE_INACTIVE;
318 if (event->pending_disable) {
319 event->pending_disable = 0;
320 event->state = PERF_EVENT_STATE_OFF;
322 event->tstamp_stopped = ctx->time;
323 event->pmu->disable(event);
326 if (!is_software_event(event))
327 cpuctx->active_oncpu--;
329 if (event->attr.exclusive || !cpuctx->active_oncpu)
330 cpuctx->exclusive = 0;
334 group_sched_out(struct perf_event *group_event,
335 struct perf_cpu_context *cpuctx,
336 struct perf_event_context *ctx)
338 struct perf_event *event;
340 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
343 event_sched_out(group_event, cpuctx, ctx);
346 * Schedule out siblings (if any):
348 list_for_each_entry(event, &group_event->sibling_list, group_entry)
349 event_sched_out(event, cpuctx, ctx);
351 if (group_event->attr.exclusive)
352 cpuctx->exclusive = 0;
356 * Cross CPU call to remove a performance event
358 * We disable the event on the hardware level first. After that we
359 * remove it from the context list.
361 static void __perf_event_remove_from_context(void *info)
363 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
364 struct perf_event *event = info;
365 struct perf_event_context *ctx = event->ctx;
368 * If this is a task context, we need to check whether it is
369 * the current task context of this cpu. If not it has been
370 * scheduled out before the smp call arrived.
372 if (ctx->task && cpuctx->task_ctx != ctx)
375 spin_lock(&ctx->lock);
377 * Protect the list operation against NMI by disabling the
378 * events on a global level.
382 event_sched_out(event, cpuctx, ctx);
384 list_del_event(event, ctx);
388 * Allow more per task events with respect to the
391 cpuctx->max_pertask =
392 min(perf_max_events - ctx->nr_events,
393 perf_max_events - perf_reserved_percpu);
397 spin_unlock(&ctx->lock);
402 * Remove the event from a task's (or a CPU's) list of events.
404 * Must be called with ctx->mutex held.
406 * CPU events are removed with a smp call. For task events we only
407 * call when the task is on a CPU.
409 * If event->ctx is a cloned context, callers must make sure that
410 * every task struct that event->ctx->task could possibly point to
411 * remains valid. This is OK when called from perf_release since
412 * that only calls us on the top-level context, which can't be a clone.
413 * When called from perf_event_exit_task, it's OK because the
414 * context has been detached from its task.
416 static void perf_event_remove_from_context(struct perf_event *event)
418 struct perf_event_context *ctx = event->ctx;
419 struct task_struct *task = ctx->task;
423 * Per cpu events are removed via an smp call and
424 * the removal is always sucessful.
426 smp_call_function_single(event->cpu,
427 __perf_event_remove_from_context,
433 task_oncpu_function_call(task, __perf_event_remove_from_context,
436 spin_lock_irq(&ctx->lock);
438 * If the context is active we need to retry the smp call.
440 if (ctx->nr_active && !list_empty(&event->group_entry)) {
441 spin_unlock_irq(&ctx->lock);
446 * The lock prevents that this context is scheduled in so we
447 * can remove the event safely, if the call above did not
450 if (!list_empty(&event->group_entry)) {
451 list_del_event(event, ctx);
453 spin_unlock_irq(&ctx->lock);
456 static inline u64 perf_clock(void)
458 return cpu_clock(smp_processor_id());
462 * Update the record of the current time in a context.
464 static void update_context_time(struct perf_event_context *ctx)
466 u64 now = perf_clock();
468 ctx->time += now - ctx->timestamp;
469 ctx->timestamp = now;
473 * Update the total_time_enabled and total_time_running fields for a event.
475 static void update_event_times(struct perf_event *event)
477 struct perf_event_context *ctx = event->ctx;
480 if (event->state < PERF_EVENT_STATE_INACTIVE ||
481 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
484 event->total_time_enabled = ctx->time - event->tstamp_enabled;
486 if (event->state == PERF_EVENT_STATE_INACTIVE)
487 run_end = event->tstamp_stopped;
491 event->total_time_running = run_end - event->tstamp_running;
495 * Update total_time_enabled and total_time_running for all events in a group.
497 static void update_group_times(struct perf_event *leader)
499 struct perf_event *event;
501 update_event_times(leader);
502 list_for_each_entry(event, &leader->sibling_list, group_entry)
503 update_event_times(event);
507 * Cross CPU call to disable a performance event
509 static void __perf_event_disable(void *info)
511 struct perf_event *event = info;
512 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
513 struct perf_event_context *ctx = event->ctx;
516 * If this is a per-task event, need to check whether this
517 * event's task is the current task on this cpu.
519 if (ctx->task && cpuctx->task_ctx != ctx)
522 spin_lock(&ctx->lock);
525 * If the event is on, turn it off.
526 * If it is in error state, leave it in error state.
528 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
529 update_context_time(ctx);
530 update_group_times(event);
531 if (event == event->group_leader)
532 group_sched_out(event, cpuctx, ctx);
534 event_sched_out(event, cpuctx, ctx);
535 event->state = PERF_EVENT_STATE_OFF;
538 spin_unlock(&ctx->lock);
544 * If event->ctx is a cloned context, callers must make sure that
545 * every task struct that event->ctx->task could possibly point to
546 * remains valid. This condition is satisifed when called through
547 * perf_event_for_each_child or perf_event_for_each because they
548 * hold the top-level event's child_mutex, so any descendant that
549 * goes to exit will block in sync_child_event.
550 * When called from perf_pending_event it's OK because event->ctx
551 * is the current context on this CPU and preemption is disabled,
552 * hence we can't get into perf_event_task_sched_out for this context.
554 static void perf_event_disable(struct perf_event *event)
556 struct perf_event_context *ctx = event->ctx;
557 struct task_struct *task = ctx->task;
561 * Disable the event on the cpu that it's on
563 smp_call_function_single(event->cpu, __perf_event_disable,
569 task_oncpu_function_call(task, __perf_event_disable, event);
571 spin_lock_irq(&ctx->lock);
573 * If the event is still active, we need to retry the cross-call.
575 if (event->state == PERF_EVENT_STATE_ACTIVE) {
576 spin_unlock_irq(&ctx->lock);
581 * Since we have the lock this context can't be scheduled
582 * in, so we can change the state safely.
584 if (event->state == PERF_EVENT_STATE_INACTIVE) {
585 update_group_times(event);
586 event->state = PERF_EVENT_STATE_OFF;
589 spin_unlock_irq(&ctx->lock);
593 event_sched_in(struct perf_event *event,
594 struct perf_cpu_context *cpuctx,
595 struct perf_event_context *ctx,
598 if (event->state <= PERF_EVENT_STATE_OFF)
601 event->state = PERF_EVENT_STATE_ACTIVE;
602 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
604 * The new state must be visible before we turn it on in the hardware:
608 if (event->pmu->enable(event)) {
609 event->state = PERF_EVENT_STATE_INACTIVE;
614 event->tstamp_running += ctx->time - event->tstamp_stopped;
616 if (!is_software_event(event))
617 cpuctx->active_oncpu++;
620 if (event->attr.exclusive)
621 cpuctx->exclusive = 1;
627 group_sched_in(struct perf_event *group_event,
628 struct perf_cpu_context *cpuctx,
629 struct perf_event_context *ctx,
632 struct perf_event *event, *partial_group;
635 if (group_event->state == PERF_EVENT_STATE_OFF)
638 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
640 return ret < 0 ? ret : 0;
642 if (event_sched_in(group_event, cpuctx, ctx, cpu))
646 * Schedule in siblings as one group (if any):
648 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
649 if (event_sched_in(event, cpuctx, ctx, cpu)) {
650 partial_group = event;
659 * Groups can be scheduled in as one unit only, so undo any
660 * partial group before returning:
662 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
663 if (event == partial_group)
665 event_sched_out(event, cpuctx, ctx);
667 event_sched_out(group_event, cpuctx, ctx);
673 * Return 1 for a group consisting entirely of software events,
674 * 0 if the group contains any hardware events.
676 static int is_software_only_group(struct perf_event *leader)
678 struct perf_event *event;
680 if (!is_software_event(leader))
683 list_for_each_entry(event, &leader->sibling_list, group_entry)
684 if (!is_software_event(event))
691 * Work out whether we can put this event group on the CPU now.
693 static int group_can_go_on(struct perf_event *event,
694 struct perf_cpu_context *cpuctx,
698 * Groups consisting entirely of software events can always go on.
700 if (is_software_only_group(event))
703 * If an exclusive group is already on, no other hardware
706 if (cpuctx->exclusive)
709 * If this group is exclusive and there are already
710 * events on the CPU, it can't go on.
712 if (event->attr.exclusive && cpuctx->active_oncpu)
715 * Otherwise, try to add it if all previous groups were able
721 static void add_event_to_ctx(struct perf_event *event,
722 struct perf_event_context *ctx)
724 list_add_event(event, ctx);
725 event->tstamp_enabled = ctx->time;
726 event->tstamp_running = ctx->time;
727 event->tstamp_stopped = ctx->time;
731 * Cross CPU call to install and enable a performance event
733 * Must be called with ctx->mutex held
735 static void __perf_install_in_context(void *info)
737 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
738 struct perf_event *event = info;
739 struct perf_event_context *ctx = event->ctx;
740 struct perf_event *leader = event->group_leader;
741 int cpu = smp_processor_id();
745 * If this is a task context, we need to check whether it is
746 * the current task context of this cpu. If not it has been
747 * scheduled out before the smp call arrived.
748 * Or possibly this is the right context but it isn't
749 * on this cpu because it had no events.
751 if (ctx->task && cpuctx->task_ctx != ctx) {
752 if (cpuctx->task_ctx || ctx->task != current)
754 cpuctx->task_ctx = ctx;
757 spin_lock(&ctx->lock);
759 update_context_time(ctx);
762 * Protect the list operation against NMI by disabling the
763 * events on a global level. NOP for non NMI based events.
767 add_event_to_ctx(event, ctx);
770 * Don't put the event on if it is disabled or if
771 * it is in a group and the group isn't on.
773 if (event->state != PERF_EVENT_STATE_INACTIVE ||
774 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
778 * An exclusive event can't go on if there are already active
779 * hardware events, and no hardware event can go on if there
780 * is already an exclusive event on.
782 if (!group_can_go_on(event, cpuctx, 1))
785 err = event_sched_in(event, cpuctx, ctx, cpu);
789 * This event couldn't go on. If it is in a group
790 * then we have to pull the whole group off.
791 * If the event group is pinned then put it in error state.
794 group_sched_out(leader, cpuctx, ctx);
795 if (leader->attr.pinned) {
796 update_group_times(leader);
797 leader->state = PERF_EVENT_STATE_ERROR;
801 if (!err && !ctx->task && cpuctx->max_pertask)
802 cpuctx->max_pertask--;
807 spin_unlock(&ctx->lock);
811 * Attach a performance event to a context
813 * First we add the event to the list with the hardware enable bit
814 * in event->hw_config cleared.
816 * If the event is attached to a task which is on a CPU we use a smp
817 * call to enable it in the task context. The task might have been
818 * scheduled away, but we check this in the smp call again.
820 * Must be called with ctx->mutex held.
823 perf_install_in_context(struct perf_event_context *ctx,
824 struct perf_event *event,
827 struct task_struct *task = ctx->task;
831 * Per cpu events are installed via an smp call and
832 * the install is always sucessful.
834 smp_call_function_single(cpu, __perf_install_in_context,
840 task_oncpu_function_call(task, __perf_install_in_context,
843 spin_lock_irq(&ctx->lock);
845 * we need to retry the smp call.
847 if (ctx->is_active && list_empty(&event->group_entry)) {
848 spin_unlock_irq(&ctx->lock);
853 * The lock prevents that this context is scheduled in so we
854 * can add the event safely, if it the call above did not
857 if (list_empty(&event->group_entry))
858 add_event_to_ctx(event, ctx);
859 spin_unlock_irq(&ctx->lock);
863 * Put a event into inactive state and update time fields.
864 * Enabling the leader of a group effectively enables all
865 * the group members that aren't explicitly disabled, so we
866 * have to update their ->tstamp_enabled also.
867 * Note: this works for group members as well as group leaders
868 * since the non-leader members' sibling_lists will be empty.
870 static void __perf_event_mark_enabled(struct perf_event *event,
871 struct perf_event_context *ctx)
873 struct perf_event *sub;
875 event->state = PERF_EVENT_STATE_INACTIVE;
876 event->tstamp_enabled = ctx->time - event->total_time_enabled;
877 list_for_each_entry(sub, &event->sibling_list, group_entry)
878 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
879 sub->tstamp_enabled =
880 ctx->time - sub->total_time_enabled;
884 * Cross CPU call to enable a performance event
886 static void __perf_event_enable(void *info)
888 struct perf_event *event = info;
889 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
890 struct perf_event_context *ctx = event->ctx;
891 struct perf_event *leader = event->group_leader;
895 * If this is a per-task event, need to check whether this
896 * event's task is the current task on this cpu.
898 if (ctx->task && cpuctx->task_ctx != ctx) {
899 if (cpuctx->task_ctx || ctx->task != current)
901 cpuctx->task_ctx = ctx;
904 spin_lock(&ctx->lock);
906 update_context_time(ctx);
908 if (event->state >= PERF_EVENT_STATE_INACTIVE)
910 __perf_event_mark_enabled(event, ctx);
913 * If the event is in a group and isn't the group leader,
914 * then don't put it on unless the group is on.
916 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
919 if (!group_can_go_on(event, cpuctx, 1)) {
924 err = group_sched_in(event, cpuctx, ctx,
927 err = event_sched_in(event, cpuctx, ctx,
934 * If this event can't go on and it's part of a
935 * group, then the whole group has to come off.
938 group_sched_out(leader, cpuctx, ctx);
939 if (leader->attr.pinned) {
940 update_group_times(leader);
941 leader->state = PERF_EVENT_STATE_ERROR;
946 spin_unlock(&ctx->lock);
952 * If event->ctx is a cloned context, callers must make sure that
953 * every task struct that event->ctx->task could possibly point to
954 * remains valid. This condition is satisfied when called through
955 * perf_event_for_each_child or perf_event_for_each as described
956 * for perf_event_disable.
958 static void perf_event_enable(struct perf_event *event)
960 struct perf_event_context *ctx = event->ctx;
961 struct task_struct *task = ctx->task;
965 * Enable the event on the cpu that it's on
967 smp_call_function_single(event->cpu, __perf_event_enable,
972 spin_lock_irq(&ctx->lock);
973 if (event->state >= PERF_EVENT_STATE_INACTIVE)
977 * If the event is in error state, clear that first.
978 * That way, if we see the event in error state below, we
979 * know that it has gone back into error state, as distinct
980 * from the task having been scheduled away before the
981 * cross-call arrived.
983 if (event->state == PERF_EVENT_STATE_ERROR)
984 event->state = PERF_EVENT_STATE_OFF;
987 spin_unlock_irq(&ctx->lock);
988 task_oncpu_function_call(task, __perf_event_enable, event);
990 spin_lock_irq(&ctx->lock);
993 * If the context is active and the event is still off,
994 * we need to retry the cross-call.
996 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1000 * Since we have the lock this context can't be scheduled
1001 * in, so we can change the state safely.
1003 if (event->state == PERF_EVENT_STATE_OFF)
1004 __perf_event_mark_enabled(event, ctx);
1007 spin_unlock_irq(&ctx->lock);
1010 static int perf_event_refresh(struct perf_event *event, int refresh)
1013 * not supported on inherited events
1015 if (event->attr.inherit)
1018 atomic_add(refresh, &event->event_limit);
1019 perf_event_enable(event);
1024 void __perf_event_sched_out(struct perf_event_context *ctx,
1025 struct perf_cpu_context *cpuctx)
1027 struct perf_event *event;
1029 spin_lock(&ctx->lock);
1031 if (likely(!ctx->nr_events))
1033 update_context_time(ctx);
1037 list_for_each_entry(event, &ctx->group_list, group_entry)
1038 group_sched_out(event, cpuctx, ctx);
1042 spin_unlock(&ctx->lock);
1046 * Test whether two contexts are equivalent, i.e. whether they
1047 * have both been cloned from the same version of the same context
1048 * and they both have the same number of enabled events.
1049 * If the number of enabled events is the same, then the set
1050 * of enabled events should be the same, because these are both
1051 * inherited contexts, therefore we can't access individual events
1052 * in them directly with an fd; we can only enable/disable all
1053 * events via prctl, or enable/disable all events in a family
1054 * via ioctl, which will have the same effect on both contexts.
1056 static int context_equiv(struct perf_event_context *ctx1,
1057 struct perf_event_context *ctx2)
1059 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1060 && ctx1->parent_gen == ctx2->parent_gen
1061 && !ctx1->pin_count && !ctx2->pin_count;
1064 static void __perf_event_sync_stat(struct perf_event *event,
1065 struct perf_event *next_event)
1069 if (!event->attr.inherit_stat)
1073 * Update the event value, we cannot use perf_event_read()
1074 * because we're in the middle of a context switch and have IRQs
1075 * disabled, which upsets smp_call_function_single(), however
1076 * we know the event must be on the current CPU, therefore we
1077 * don't need to use it.
1079 switch (event->state) {
1080 case PERF_EVENT_STATE_ACTIVE:
1081 event->pmu->read(event);
1084 case PERF_EVENT_STATE_INACTIVE:
1085 update_event_times(event);
1093 * In order to keep per-task stats reliable we need to flip the event
1094 * values when we flip the contexts.
1096 value = atomic64_read(&next_event->count);
1097 value = atomic64_xchg(&event->count, value);
1098 atomic64_set(&next_event->count, value);
1100 swap(event->total_time_enabled, next_event->total_time_enabled);
1101 swap(event->total_time_running, next_event->total_time_running);
1104 * Since we swizzled the values, update the user visible data too.
1106 perf_event_update_userpage(event);
1107 perf_event_update_userpage(next_event);
1110 #define list_next_entry(pos, member) \
1111 list_entry(pos->member.next, typeof(*pos), member)
1113 static void perf_event_sync_stat(struct perf_event_context *ctx,
1114 struct perf_event_context *next_ctx)
1116 struct perf_event *event, *next_event;
1121 update_context_time(ctx);
1123 event = list_first_entry(&ctx->event_list,
1124 struct perf_event, event_entry);
1126 next_event = list_first_entry(&next_ctx->event_list,
1127 struct perf_event, event_entry);
1129 while (&event->event_entry != &ctx->event_list &&
1130 &next_event->event_entry != &next_ctx->event_list) {
1132 __perf_event_sync_stat(event, next_event);
1134 event = list_next_entry(event, event_entry);
1135 next_event = list_next_entry(next_event, event_entry);
1140 * Called from scheduler to remove the events of the current task,
1141 * with interrupts disabled.
1143 * We stop each event and update the event value in event->count.
1145 * This does not protect us against NMI, but disable()
1146 * sets the disabled bit in the control field of event _before_
1147 * accessing the event control register. If a NMI hits, then it will
1148 * not restart the event.
1150 void perf_event_task_sched_out(struct task_struct *task,
1151 struct task_struct *next, int cpu)
1153 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1154 struct perf_event_context *ctx = task->perf_event_ctxp;
1155 struct perf_event_context *next_ctx;
1156 struct perf_event_context *parent;
1157 struct pt_regs *regs;
1160 regs = task_pt_regs(task);
1161 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1163 if (likely(!ctx || !cpuctx->task_ctx))
1167 parent = rcu_dereference(ctx->parent_ctx);
1168 next_ctx = next->perf_event_ctxp;
1169 if (parent && next_ctx &&
1170 rcu_dereference(next_ctx->parent_ctx) == parent) {
1172 * Looks like the two contexts are clones, so we might be
1173 * able to optimize the context switch. We lock both
1174 * contexts and check that they are clones under the
1175 * lock (including re-checking that neither has been
1176 * uncloned in the meantime). It doesn't matter which
1177 * order we take the locks because no other cpu could
1178 * be trying to lock both of these tasks.
1180 spin_lock(&ctx->lock);
1181 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1182 if (context_equiv(ctx, next_ctx)) {
1184 * XXX do we need a memory barrier of sorts
1185 * wrt to rcu_dereference() of perf_event_ctxp
1187 task->perf_event_ctxp = next_ctx;
1188 next->perf_event_ctxp = ctx;
1190 next_ctx->task = task;
1193 perf_event_sync_stat(ctx, next_ctx);
1195 spin_unlock(&next_ctx->lock);
1196 spin_unlock(&ctx->lock);
1201 __perf_event_sched_out(ctx, cpuctx);
1202 cpuctx->task_ctx = NULL;
1207 * Called with IRQs disabled
1209 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1211 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1213 if (!cpuctx->task_ctx)
1216 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1219 __perf_event_sched_out(ctx, cpuctx);
1220 cpuctx->task_ctx = NULL;
1224 * Called with IRQs disabled
1226 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1228 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1232 __perf_event_sched_in(struct perf_event_context *ctx,
1233 struct perf_cpu_context *cpuctx, int cpu)
1235 struct perf_event *event;
1238 spin_lock(&ctx->lock);
1240 if (likely(!ctx->nr_events))
1243 ctx->timestamp = perf_clock();
1248 * First go through the list and put on any pinned groups
1249 * in order to give them the best chance of going on.
1251 list_for_each_entry(event, &ctx->group_list, group_entry) {
1252 if (event->state <= PERF_EVENT_STATE_OFF ||
1253 !event->attr.pinned)
1255 if (event->cpu != -1 && event->cpu != cpu)
1258 if (group_can_go_on(event, cpuctx, 1))
1259 group_sched_in(event, cpuctx, ctx, cpu);
1262 * If this pinned group hasn't been scheduled,
1263 * put it in error state.
1265 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1266 update_group_times(event);
1267 event->state = PERF_EVENT_STATE_ERROR;
1271 list_for_each_entry(event, &ctx->group_list, group_entry) {
1273 * Ignore events in OFF or ERROR state, and
1274 * ignore pinned events since we did them already.
1276 if (event->state <= PERF_EVENT_STATE_OFF ||
1281 * Listen to the 'cpu' scheduling filter constraint
1284 if (event->cpu != -1 && event->cpu != cpu)
1287 if (group_can_go_on(event, cpuctx, can_add_hw))
1288 if (group_sched_in(event, cpuctx, ctx, cpu))
1293 spin_unlock(&ctx->lock);
1297 * Called from scheduler to add the events of the current task
1298 * with interrupts disabled.
1300 * We restore the event value and then enable it.
1302 * This does not protect us against NMI, but enable()
1303 * sets the enabled bit in the control field of event _before_
1304 * accessing the event control register. If a NMI hits, then it will
1305 * keep the event running.
1307 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1309 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1310 struct perf_event_context *ctx = task->perf_event_ctxp;
1314 if (cpuctx->task_ctx == ctx)
1316 __perf_event_sched_in(ctx, cpuctx, cpu);
1317 cpuctx->task_ctx = ctx;
1320 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1322 struct perf_event_context *ctx = &cpuctx->ctx;
1324 __perf_event_sched_in(ctx, cpuctx, cpu);
1327 #define MAX_INTERRUPTS (~0ULL)
1329 static void perf_log_throttle(struct perf_event *event, int enable);
1331 static void perf_adjust_period(struct perf_event *event, u64 events)
1333 struct hw_perf_event *hwc = &event->hw;
1334 u64 period, sample_period;
1337 events *= hwc->sample_period;
1338 period = div64_u64(events, event->attr.sample_freq);
1340 delta = (s64)(period - hwc->sample_period);
1341 delta = (delta + 7) / 8; /* low pass filter */
1343 sample_period = hwc->sample_period + delta;
1348 hwc->sample_period = sample_period;
1351 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1353 struct perf_event *event;
1354 struct hw_perf_event *hwc;
1355 u64 interrupts, freq;
1357 spin_lock(&ctx->lock);
1358 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1359 if (event->state != PERF_EVENT_STATE_ACTIVE)
1364 interrupts = hwc->interrupts;
1365 hwc->interrupts = 0;
1368 * unthrottle events on the tick
1370 if (interrupts == MAX_INTERRUPTS) {
1371 perf_log_throttle(event, 1);
1372 event->pmu->unthrottle(event);
1373 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1376 if (!event->attr.freq || !event->attr.sample_freq)
1380 * if the specified freq < HZ then we need to skip ticks
1382 if (event->attr.sample_freq < HZ) {
1383 freq = event->attr.sample_freq;
1385 hwc->freq_count += freq;
1386 hwc->freq_interrupts += interrupts;
1388 if (hwc->freq_count < HZ)
1391 interrupts = hwc->freq_interrupts;
1392 hwc->freq_interrupts = 0;
1393 hwc->freq_count -= HZ;
1397 perf_adjust_period(event, freq * interrupts);
1400 * In order to avoid being stalled by an (accidental) huge
1401 * sample period, force reset the sample period if we didn't
1402 * get any events in this freq period.
1406 event->pmu->disable(event);
1407 atomic64_set(&hwc->period_left, 0);
1408 event->pmu->enable(event);
1412 spin_unlock(&ctx->lock);
1416 * Round-robin a context's events:
1418 static void rotate_ctx(struct perf_event_context *ctx)
1420 struct perf_event *event;
1422 if (!ctx->nr_events)
1425 spin_lock(&ctx->lock);
1427 * Rotate the first entry last (works just fine for group events too):
1430 list_for_each_entry(event, &ctx->group_list, group_entry) {
1431 list_move_tail(&event->group_entry, &ctx->group_list);
1436 spin_unlock(&ctx->lock);
1439 void perf_event_task_tick(struct task_struct *curr, int cpu)
1441 struct perf_cpu_context *cpuctx;
1442 struct perf_event_context *ctx;
1444 if (!atomic_read(&nr_events))
1447 cpuctx = &per_cpu(perf_cpu_context, cpu);
1448 ctx = curr->perf_event_ctxp;
1450 perf_ctx_adjust_freq(&cpuctx->ctx);
1452 perf_ctx_adjust_freq(ctx);
1454 perf_event_cpu_sched_out(cpuctx);
1456 __perf_event_task_sched_out(ctx);
1458 rotate_ctx(&cpuctx->ctx);
1462 perf_event_cpu_sched_in(cpuctx, cpu);
1464 perf_event_task_sched_in(curr, cpu);
1468 * Enable all of a task's events that have been marked enable-on-exec.
1469 * This expects task == current.
1471 static void perf_event_enable_on_exec(struct task_struct *task)
1473 struct perf_event_context *ctx;
1474 struct perf_event *event;
1475 unsigned long flags;
1478 local_irq_save(flags);
1479 ctx = task->perf_event_ctxp;
1480 if (!ctx || !ctx->nr_events)
1483 __perf_event_task_sched_out(ctx);
1485 spin_lock(&ctx->lock);
1487 list_for_each_entry(event, &ctx->group_list, group_entry) {
1488 if (!event->attr.enable_on_exec)
1490 event->attr.enable_on_exec = 0;
1491 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1493 __perf_event_mark_enabled(event, ctx);
1498 * Unclone this context if we enabled any event.
1503 spin_unlock(&ctx->lock);
1505 perf_event_task_sched_in(task, smp_processor_id());
1507 local_irq_restore(flags);
1511 * Cross CPU call to read the hardware event
1513 static void __perf_event_read(void *info)
1515 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1516 struct perf_event *event = info;
1517 struct perf_event_context *ctx = event->ctx;
1520 * If this is a task context, we need to check whether it is
1521 * the current task context of this cpu. If not it has been
1522 * scheduled out before the smp call arrived. In that case
1523 * event->count would have been updated to a recent sample
1524 * when the event was scheduled out.
1526 if (ctx->task && cpuctx->task_ctx != ctx)
1529 update_context_time(ctx);
1530 update_event_times(event);
1531 event->pmu->read(event);
1534 static u64 perf_event_read(struct perf_event *event)
1537 * If event is enabled and currently active on a CPU, update the
1538 * value in the event structure:
1540 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1541 smp_call_function_single(event->oncpu,
1542 __perf_event_read, event, 1);
1543 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1544 update_event_times(event);
1547 return atomic64_read(&event->count);
1551 * Initialize the perf_event context in a task_struct:
1554 __perf_event_init_context(struct perf_event_context *ctx,
1555 struct task_struct *task)
1557 memset(ctx, 0, sizeof(*ctx));
1558 spin_lock_init(&ctx->lock);
1559 mutex_init(&ctx->mutex);
1560 INIT_LIST_HEAD(&ctx->group_list);
1561 INIT_LIST_HEAD(&ctx->event_list);
1562 atomic_set(&ctx->refcount, 1);
1566 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1568 struct perf_event_context *ctx;
1569 struct perf_cpu_context *cpuctx;
1570 struct task_struct *task;
1571 unsigned long flags;
1575 * If cpu is not a wildcard then this is a percpu event:
1578 /* Must be root to operate on a CPU event: */
1579 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1580 return ERR_PTR(-EACCES);
1582 if (cpu < 0 || cpu > num_possible_cpus())
1583 return ERR_PTR(-EINVAL);
1586 * We could be clever and allow to attach a event to an
1587 * offline CPU and activate it when the CPU comes up, but
1590 if (!cpu_isset(cpu, cpu_online_map))
1591 return ERR_PTR(-ENODEV);
1593 cpuctx = &per_cpu(perf_cpu_context, cpu);
1604 task = find_task_by_vpid(pid);
1606 get_task_struct(task);
1610 return ERR_PTR(-ESRCH);
1613 * Can't attach events to a dying task.
1616 if (task->flags & PF_EXITING)
1619 /* Reuse ptrace permission checks for now. */
1621 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1625 ctx = perf_lock_task_context(task, &flags);
1628 spin_unlock_irqrestore(&ctx->lock, flags);
1632 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1636 __perf_event_init_context(ctx, task);
1638 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1640 * We raced with some other task; use
1641 * the context they set.
1646 get_task_struct(task);
1649 put_task_struct(task);
1653 put_task_struct(task);
1654 return ERR_PTR(err);
1657 static void perf_event_free_filter(struct perf_event *event);
1659 static void free_event_rcu(struct rcu_head *head)
1661 struct perf_event *event;
1663 event = container_of(head, struct perf_event, rcu_head);
1665 put_pid_ns(event->ns);
1666 perf_event_free_filter(event);
1670 static void perf_pending_sync(struct perf_event *event);
1672 static void free_event(struct perf_event *event)
1674 perf_pending_sync(event);
1676 if (!event->parent) {
1677 atomic_dec(&nr_events);
1678 if (event->attr.mmap)
1679 atomic_dec(&nr_mmap_events);
1680 if (event->attr.comm)
1681 atomic_dec(&nr_comm_events);
1682 if (event->attr.task)
1683 atomic_dec(&nr_task_events);
1686 if (event->output) {
1687 fput(event->output->filp);
1688 event->output = NULL;
1692 event->destroy(event);
1694 put_ctx(event->ctx);
1695 call_rcu(&event->rcu_head, free_event_rcu);
1699 * Called when the last reference to the file is gone.
1701 static int perf_release(struct inode *inode, struct file *file)
1703 struct perf_event *event = file->private_data;
1704 struct perf_event_context *ctx = event->ctx;
1706 file->private_data = NULL;
1708 WARN_ON_ONCE(ctx->parent_ctx);
1709 mutex_lock(&ctx->mutex);
1710 perf_event_remove_from_context(event);
1711 mutex_unlock(&ctx->mutex);
1713 mutex_lock(&event->owner->perf_event_mutex);
1714 list_del_init(&event->owner_entry);
1715 mutex_unlock(&event->owner->perf_event_mutex);
1716 put_task_struct(event->owner);
1723 int perf_event_release_kernel(struct perf_event *event)
1725 struct perf_event_context *ctx = event->ctx;
1727 WARN_ON_ONCE(ctx->parent_ctx);
1728 mutex_lock(&ctx->mutex);
1729 perf_event_remove_from_context(event);
1730 mutex_unlock(&ctx->mutex);
1732 mutex_lock(&event->owner->perf_event_mutex);
1733 list_del_init(&event->owner_entry);
1734 mutex_unlock(&event->owner->perf_event_mutex);
1735 put_task_struct(event->owner);
1741 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1743 static int perf_event_read_size(struct perf_event *event)
1745 int entry = sizeof(u64); /* value */
1749 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1750 size += sizeof(u64);
1752 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1753 size += sizeof(u64);
1755 if (event->attr.read_format & PERF_FORMAT_ID)
1756 entry += sizeof(u64);
1758 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1759 nr += event->group_leader->nr_siblings;
1760 size += sizeof(u64);
1768 u64 perf_event_read_value(struct perf_event *event)
1770 struct perf_event *child;
1773 total += perf_event_read(event);
1774 list_for_each_entry(child, &event->child_list, child_list)
1775 total += perf_event_read(child);
1779 EXPORT_SYMBOL_GPL(perf_event_read_value);
1781 static int perf_event_read_group(struct perf_event *event,
1782 u64 read_format, char __user *buf)
1784 struct perf_event *leader = event->group_leader, *sub;
1785 int n = 0, size = 0, ret = 0;
1789 count = perf_event_read_value(leader);
1791 values[n++] = 1 + leader->nr_siblings;
1792 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1793 values[n++] = leader->total_time_enabled +
1794 atomic64_read(&leader->child_total_time_enabled);
1796 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1797 values[n++] = leader->total_time_running +
1798 atomic64_read(&leader->child_total_time_running);
1800 values[n++] = count;
1801 if (read_format & PERF_FORMAT_ID)
1802 values[n++] = primary_event_id(leader);
1804 size = n * sizeof(u64);
1806 if (copy_to_user(buf, values, size))
1811 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1814 values[n++] = perf_event_read_value(sub);
1815 if (read_format & PERF_FORMAT_ID)
1816 values[n++] = primary_event_id(sub);
1818 size = n * sizeof(u64);
1820 if (copy_to_user(buf + size, values, size))
1829 static int perf_event_read_one(struct perf_event *event,
1830 u64 read_format, char __user *buf)
1835 values[n++] = perf_event_read_value(event);
1836 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1837 values[n++] = event->total_time_enabled +
1838 atomic64_read(&event->child_total_time_enabled);
1840 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1841 values[n++] = event->total_time_running +
1842 atomic64_read(&event->child_total_time_running);
1844 if (read_format & PERF_FORMAT_ID)
1845 values[n++] = primary_event_id(event);
1847 if (copy_to_user(buf, values, n * sizeof(u64)))
1850 return n * sizeof(u64);
1854 * Read the performance event - simple non blocking version for now
1857 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1859 u64 read_format = event->attr.read_format;
1863 * Return end-of-file for a read on a event that is in
1864 * error state (i.e. because it was pinned but it couldn't be
1865 * scheduled on to the CPU at some point).
1867 if (event->state == PERF_EVENT_STATE_ERROR)
1870 if (count < perf_event_read_size(event))
1873 WARN_ON_ONCE(event->ctx->parent_ctx);
1874 mutex_lock(&event->child_mutex);
1875 if (read_format & PERF_FORMAT_GROUP)
1876 ret = perf_event_read_group(event, read_format, buf);
1878 ret = perf_event_read_one(event, read_format, buf);
1879 mutex_unlock(&event->child_mutex);
1885 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1887 struct perf_event *event = file->private_data;
1889 return perf_read_hw(event, buf, count);
1892 static unsigned int perf_poll(struct file *file, poll_table *wait)
1894 struct perf_event *event = file->private_data;
1895 struct perf_mmap_data *data;
1896 unsigned int events = POLL_HUP;
1899 data = rcu_dereference(event->data);
1901 events = atomic_xchg(&data->poll, 0);
1904 poll_wait(file, &event->waitq, wait);
1909 static void perf_event_reset(struct perf_event *event)
1911 (void)perf_event_read(event);
1912 atomic64_set(&event->count, 0);
1913 perf_event_update_userpage(event);
1917 * Holding the top-level event's child_mutex means that any
1918 * descendant process that has inherited this event will block
1919 * in sync_child_event if it goes to exit, thus satisfying the
1920 * task existence requirements of perf_event_enable/disable.
1922 static void perf_event_for_each_child(struct perf_event *event,
1923 void (*func)(struct perf_event *))
1925 struct perf_event *child;
1927 WARN_ON_ONCE(event->ctx->parent_ctx);
1928 mutex_lock(&event->child_mutex);
1930 list_for_each_entry(child, &event->child_list, child_list)
1932 mutex_unlock(&event->child_mutex);
1935 static void perf_event_for_each(struct perf_event *event,
1936 void (*func)(struct perf_event *))
1938 struct perf_event_context *ctx = event->ctx;
1939 struct perf_event *sibling;
1941 WARN_ON_ONCE(ctx->parent_ctx);
1942 mutex_lock(&ctx->mutex);
1943 event = event->group_leader;
1945 perf_event_for_each_child(event, func);
1947 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1948 perf_event_for_each_child(event, func);
1949 mutex_unlock(&ctx->mutex);
1952 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1954 struct perf_event_context *ctx = event->ctx;
1959 if (!event->attr.sample_period)
1962 size = copy_from_user(&value, arg, sizeof(value));
1963 if (size != sizeof(value))
1969 spin_lock_irq(&ctx->lock);
1970 if (event->attr.freq) {
1971 if (value > sysctl_perf_event_sample_rate) {
1976 event->attr.sample_freq = value;
1978 event->attr.sample_period = value;
1979 event->hw.sample_period = value;
1982 spin_unlock_irq(&ctx->lock);
1987 static int perf_event_set_output(struct perf_event *event, int output_fd);
1988 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
1990 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1992 struct perf_event *event = file->private_data;
1993 void (*func)(struct perf_event *);
1997 case PERF_EVENT_IOC_ENABLE:
1998 func = perf_event_enable;
2000 case PERF_EVENT_IOC_DISABLE:
2001 func = perf_event_disable;
2003 case PERF_EVENT_IOC_RESET:
2004 func = perf_event_reset;
2007 case PERF_EVENT_IOC_REFRESH:
2008 return perf_event_refresh(event, arg);
2010 case PERF_EVENT_IOC_PERIOD:
2011 return perf_event_period(event, (u64 __user *)arg);
2013 case PERF_EVENT_IOC_SET_OUTPUT:
2014 return perf_event_set_output(event, arg);
2016 case PERF_EVENT_IOC_SET_FILTER:
2017 return perf_event_set_filter(event, (void __user *)arg);
2023 if (flags & PERF_IOC_FLAG_GROUP)
2024 perf_event_for_each(event, func);
2026 perf_event_for_each_child(event, func);
2031 int perf_event_task_enable(void)
2033 struct perf_event *event;
2035 mutex_lock(¤t->perf_event_mutex);
2036 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2037 perf_event_for_each_child(event, perf_event_enable);
2038 mutex_unlock(¤t->perf_event_mutex);
2043 int perf_event_task_disable(void)
2045 struct perf_event *event;
2047 mutex_lock(¤t->perf_event_mutex);
2048 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2049 perf_event_for_each_child(event, perf_event_disable);
2050 mutex_unlock(¤t->perf_event_mutex);
2055 #ifndef PERF_EVENT_INDEX_OFFSET
2056 # define PERF_EVENT_INDEX_OFFSET 0
2059 static int perf_event_index(struct perf_event *event)
2061 if (event->state != PERF_EVENT_STATE_ACTIVE)
2064 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2068 * Callers need to ensure there can be no nesting of this function, otherwise
2069 * the seqlock logic goes bad. We can not serialize this because the arch
2070 * code calls this from NMI context.
2072 void perf_event_update_userpage(struct perf_event *event)
2074 struct perf_event_mmap_page *userpg;
2075 struct perf_mmap_data *data;
2078 data = rcu_dereference(event->data);
2082 userpg = data->user_page;
2085 * Disable preemption so as to not let the corresponding user-space
2086 * spin too long if we get preempted.
2091 userpg->index = perf_event_index(event);
2092 userpg->offset = atomic64_read(&event->count);
2093 if (event->state == PERF_EVENT_STATE_ACTIVE)
2094 userpg->offset -= atomic64_read(&event->hw.prev_count);
2096 userpg->time_enabled = event->total_time_enabled +
2097 atomic64_read(&event->child_total_time_enabled);
2099 userpg->time_running = event->total_time_running +
2100 atomic64_read(&event->child_total_time_running);
2109 static unsigned long perf_data_size(struct perf_mmap_data *data)
2111 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2114 #ifndef CONFIG_PERF_USE_VMALLOC
2117 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2120 static struct page *
2121 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2123 if (pgoff > data->nr_pages)
2127 return virt_to_page(data->user_page);
2129 return virt_to_page(data->data_pages[pgoff - 1]);
2132 static struct perf_mmap_data *
2133 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2135 struct perf_mmap_data *data;
2139 WARN_ON(atomic_read(&event->mmap_count));
2141 size = sizeof(struct perf_mmap_data);
2142 size += nr_pages * sizeof(void *);
2144 data = kzalloc(size, GFP_KERNEL);
2148 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2149 if (!data->user_page)
2150 goto fail_user_page;
2152 for (i = 0; i < nr_pages; i++) {
2153 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2154 if (!data->data_pages[i])
2155 goto fail_data_pages;
2158 data->data_order = 0;
2159 data->nr_pages = nr_pages;
2164 for (i--; i >= 0; i--)
2165 free_page((unsigned long)data->data_pages[i]);
2167 free_page((unsigned long)data->user_page);
2176 static void perf_mmap_free_page(unsigned long addr)
2178 struct page *page = virt_to_page((void *)addr);
2180 page->mapping = NULL;
2184 static void perf_mmap_data_free(struct perf_mmap_data *data)
2188 perf_mmap_free_page((unsigned long)data->user_page);
2189 for (i = 0; i < data->nr_pages; i++)
2190 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2196 * Back perf_mmap() with vmalloc memory.
2198 * Required for architectures that have d-cache aliasing issues.
2201 static struct page *
2202 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2204 if (pgoff > (1UL << data->data_order))
2207 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2210 static void perf_mmap_unmark_page(void *addr)
2212 struct page *page = vmalloc_to_page(addr);
2214 page->mapping = NULL;
2217 static void perf_mmap_data_free_work(struct work_struct *work)
2219 struct perf_mmap_data *data;
2223 data = container_of(work, struct perf_mmap_data, work);
2224 nr = 1 << data->data_order;
2226 base = data->user_page;
2227 for (i = 0; i < nr + 1; i++)
2228 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2233 static void perf_mmap_data_free(struct perf_mmap_data *data)
2235 schedule_work(&data->work);
2238 static struct perf_mmap_data *
2239 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2241 struct perf_mmap_data *data;
2245 WARN_ON(atomic_read(&event->mmap_count));
2247 size = sizeof(struct perf_mmap_data);
2248 size += sizeof(void *);
2250 data = kzalloc(size, GFP_KERNEL);
2254 INIT_WORK(&data->work, perf_mmap_data_free_work);
2256 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2260 data->user_page = all_buf;
2261 data->data_pages[0] = all_buf + PAGE_SIZE;
2262 data->data_order = ilog2(nr_pages);
2276 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2278 struct perf_event *event = vma->vm_file->private_data;
2279 struct perf_mmap_data *data;
2280 int ret = VM_FAULT_SIGBUS;
2282 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2283 if (vmf->pgoff == 0)
2289 data = rcu_dereference(event->data);
2293 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2296 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2300 get_page(vmf->page);
2301 vmf->page->mapping = vma->vm_file->f_mapping;
2302 vmf->page->index = vmf->pgoff;
2312 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2314 long max_size = perf_data_size(data);
2316 atomic_set(&data->lock, -1);
2318 if (event->attr.watermark) {
2319 data->watermark = min_t(long, max_size,
2320 event->attr.wakeup_watermark);
2323 if (!data->watermark)
2324 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2327 rcu_assign_pointer(event->data, data);
2330 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2332 struct perf_mmap_data *data;
2334 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2335 perf_mmap_data_free(data);
2339 static void perf_mmap_data_release(struct perf_event *event)
2341 struct perf_mmap_data *data = event->data;
2343 WARN_ON(atomic_read(&event->mmap_count));
2345 rcu_assign_pointer(event->data, NULL);
2346 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2349 static void perf_mmap_open(struct vm_area_struct *vma)
2351 struct perf_event *event = vma->vm_file->private_data;
2353 atomic_inc(&event->mmap_count);
2356 static void perf_mmap_close(struct vm_area_struct *vma)
2358 struct perf_event *event = vma->vm_file->private_data;
2360 WARN_ON_ONCE(event->ctx->parent_ctx);
2361 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2362 unsigned long size = perf_data_size(event->data);
2363 struct user_struct *user = current_user();
2365 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2366 vma->vm_mm->locked_vm -= event->data->nr_locked;
2367 perf_mmap_data_release(event);
2368 mutex_unlock(&event->mmap_mutex);
2372 static const struct vm_operations_struct perf_mmap_vmops = {
2373 .open = perf_mmap_open,
2374 .close = perf_mmap_close,
2375 .fault = perf_mmap_fault,
2376 .page_mkwrite = perf_mmap_fault,
2379 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2381 struct perf_event *event = file->private_data;
2382 unsigned long user_locked, user_lock_limit;
2383 struct user_struct *user = current_user();
2384 unsigned long locked, lock_limit;
2385 struct perf_mmap_data *data;
2386 unsigned long vma_size;
2387 unsigned long nr_pages;
2388 long user_extra, extra;
2391 if (!(vma->vm_flags & VM_SHARED))
2394 vma_size = vma->vm_end - vma->vm_start;
2395 nr_pages = (vma_size / PAGE_SIZE) - 1;
2398 * If we have data pages ensure they're a power-of-two number, so we
2399 * can do bitmasks instead of modulo.
2401 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2404 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2407 if (vma->vm_pgoff != 0)
2410 WARN_ON_ONCE(event->ctx->parent_ctx);
2411 mutex_lock(&event->mmap_mutex);
2412 if (event->output) {
2417 if (atomic_inc_not_zero(&event->mmap_count)) {
2418 if (nr_pages != event->data->nr_pages)
2423 user_extra = nr_pages + 1;
2424 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2427 * Increase the limit linearly with more CPUs:
2429 user_lock_limit *= num_online_cpus();
2431 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2434 if (user_locked > user_lock_limit)
2435 extra = user_locked - user_lock_limit;
2437 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2438 lock_limit >>= PAGE_SHIFT;
2439 locked = vma->vm_mm->locked_vm + extra;
2441 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2442 !capable(CAP_IPC_LOCK)) {
2447 WARN_ON(event->data);
2449 data = perf_mmap_data_alloc(event, nr_pages);
2455 perf_mmap_data_init(event, data);
2457 atomic_set(&event->mmap_count, 1);
2458 atomic_long_add(user_extra, &user->locked_vm);
2459 vma->vm_mm->locked_vm += extra;
2460 event->data->nr_locked = extra;
2461 if (vma->vm_flags & VM_WRITE)
2462 event->data->writable = 1;
2465 mutex_unlock(&event->mmap_mutex);
2467 vma->vm_flags |= VM_RESERVED;
2468 vma->vm_ops = &perf_mmap_vmops;
2473 static int perf_fasync(int fd, struct file *filp, int on)
2475 struct inode *inode = filp->f_path.dentry->d_inode;
2476 struct perf_event *event = filp->private_data;
2479 mutex_lock(&inode->i_mutex);
2480 retval = fasync_helper(fd, filp, on, &event->fasync);
2481 mutex_unlock(&inode->i_mutex);
2489 static const struct file_operations perf_fops = {
2490 .release = perf_release,
2493 .unlocked_ioctl = perf_ioctl,
2494 .compat_ioctl = perf_ioctl,
2496 .fasync = perf_fasync,
2502 * If there's data, ensure we set the poll() state and publish everything
2503 * to user-space before waking everybody up.
2506 void perf_event_wakeup(struct perf_event *event)
2508 wake_up_all(&event->waitq);
2510 if (event->pending_kill) {
2511 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2512 event->pending_kill = 0;
2519 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2521 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2522 * single linked list and use cmpxchg() to add entries lockless.
2525 static void perf_pending_event(struct perf_pending_entry *entry)
2527 struct perf_event *event = container_of(entry,
2528 struct perf_event, pending);
2530 if (event->pending_disable) {
2531 event->pending_disable = 0;
2532 __perf_event_disable(event);
2535 if (event->pending_wakeup) {
2536 event->pending_wakeup = 0;
2537 perf_event_wakeup(event);
2541 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2543 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2547 static void perf_pending_queue(struct perf_pending_entry *entry,
2548 void (*func)(struct perf_pending_entry *))
2550 struct perf_pending_entry **head;
2552 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2557 head = &get_cpu_var(perf_pending_head);
2560 entry->next = *head;
2561 } while (cmpxchg(head, entry->next, entry) != entry->next);
2563 set_perf_event_pending();
2565 put_cpu_var(perf_pending_head);
2568 static int __perf_pending_run(void)
2570 struct perf_pending_entry *list;
2573 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2574 while (list != PENDING_TAIL) {
2575 void (*func)(struct perf_pending_entry *);
2576 struct perf_pending_entry *entry = list;
2583 * Ensure we observe the unqueue before we issue the wakeup,
2584 * so that we won't be waiting forever.
2585 * -- see perf_not_pending().
2596 static inline int perf_not_pending(struct perf_event *event)
2599 * If we flush on whatever cpu we run, there is a chance we don't
2603 __perf_pending_run();
2607 * Ensure we see the proper queue state before going to sleep
2608 * so that we do not miss the wakeup. -- see perf_pending_handle()
2611 return event->pending.next == NULL;
2614 static void perf_pending_sync(struct perf_event *event)
2616 wait_event(event->waitq, perf_not_pending(event));
2619 void perf_event_do_pending(void)
2621 __perf_pending_run();
2625 * Callchain support -- arch specific
2628 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2636 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2637 unsigned long offset, unsigned long head)
2641 if (!data->writable)
2644 mask = perf_data_size(data) - 1;
2646 offset = (offset - tail) & mask;
2647 head = (head - tail) & mask;
2649 if ((int)(head - offset) < 0)
2655 static void perf_output_wakeup(struct perf_output_handle *handle)
2657 atomic_set(&handle->data->poll, POLL_IN);
2660 handle->event->pending_wakeup = 1;
2661 perf_pending_queue(&handle->event->pending,
2662 perf_pending_event);
2664 perf_event_wakeup(handle->event);
2668 * Curious locking construct.
2670 * We need to ensure a later event_id doesn't publish a head when a former
2671 * event_id isn't done writing. However since we need to deal with NMIs we
2672 * cannot fully serialize things.
2674 * What we do is serialize between CPUs so we only have to deal with NMI
2675 * nesting on a single CPU.
2677 * We only publish the head (and generate a wakeup) when the outer-most
2678 * event_id completes.
2680 static void perf_output_lock(struct perf_output_handle *handle)
2682 struct perf_mmap_data *data = handle->data;
2683 int cur, cpu = get_cpu();
2688 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2700 static void perf_output_unlock(struct perf_output_handle *handle)
2702 struct perf_mmap_data *data = handle->data;
2706 data->done_head = data->head;
2708 if (!handle->locked)
2713 * The xchg implies a full barrier that ensures all writes are done
2714 * before we publish the new head, matched by a rmb() in userspace when
2715 * reading this position.
2717 while ((head = atomic_long_xchg(&data->done_head, 0)))
2718 data->user_page->data_head = head;
2721 * NMI can happen here, which means we can miss a done_head update.
2724 cpu = atomic_xchg(&data->lock, -1);
2725 WARN_ON_ONCE(cpu != smp_processor_id());
2728 * Therefore we have to validate we did not indeed do so.
2730 if (unlikely(atomic_long_read(&data->done_head))) {
2732 * Since we had it locked, we can lock it again.
2734 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2740 if (atomic_xchg(&data->wakeup, 0))
2741 perf_output_wakeup(handle);
2746 void perf_output_copy(struct perf_output_handle *handle,
2747 const void *buf, unsigned int len)
2749 unsigned int pages_mask;
2750 unsigned long offset;
2754 offset = handle->offset;
2755 pages_mask = handle->data->nr_pages - 1;
2756 pages = handle->data->data_pages;
2759 unsigned long page_offset;
2760 unsigned long page_size;
2763 nr = (offset >> PAGE_SHIFT) & pages_mask;
2764 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2765 page_offset = offset & (page_size - 1);
2766 size = min_t(unsigned int, page_size - page_offset, len);
2768 memcpy(pages[nr] + page_offset, buf, size);
2775 handle->offset = offset;
2778 * Check we didn't copy past our reservation window, taking the
2779 * possible unsigned int wrap into account.
2781 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2784 int perf_output_begin(struct perf_output_handle *handle,
2785 struct perf_event *event, unsigned int size,
2786 int nmi, int sample)
2788 struct perf_event *output_event;
2789 struct perf_mmap_data *data;
2790 unsigned long tail, offset, head;
2793 struct perf_event_header header;
2800 * For inherited events we send all the output towards the parent.
2803 event = event->parent;
2805 output_event = rcu_dereference(event->output);
2807 event = output_event;
2809 data = rcu_dereference(event->data);
2813 handle->data = data;
2814 handle->event = event;
2816 handle->sample = sample;
2818 if (!data->nr_pages)
2821 have_lost = atomic_read(&data->lost);
2823 size += sizeof(lost_event);
2825 perf_output_lock(handle);
2829 * Userspace could choose to issue a mb() before updating the
2830 * tail pointer. So that all reads will be completed before the
2833 tail = ACCESS_ONCE(data->user_page->data_tail);
2835 offset = head = atomic_long_read(&data->head);
2837 if (unlikely(!perf_output_space(data, tail, offset, head)))
2839 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2841 handle->offset = offset;
2842 handle->head = head;
2844 if (head - tail > data->watermark)
2845 atomic_set(&data->wakeup, 1);
2848 lost_event.header.type = PERF_RECORD_LOST;
2849 lost_event.header.misc = 0;
2850 lost_event.header.size = sizeof(lost_event);
2851 lost_event.id = event->id;
2852 lost_event.lost = atomic_xchg(&data->lost, 0);
2854 perf_output_put(handle, lost_event);
2860 atomic_inc(&data->lost);
2861 perf_output_unlock(handle);
2868 void perf_output_end(struct perf_output_handle *handle)
2870 struct perf_event *event = handle->event;
2871 struct perf_mmap_data *data = handle->data;
2873 int wakeup_events = event->attr.wakeup_events;
2875 if (handle->sample && wakeup_events) {
2876 int events = atomic_inc_return(&data->events);
2877 if (events >= wakeup_events) {
2878 atomic_sub(wakeup_events, &data->events);
2879 atomic_set(&data->wakeup, 1);
2883 perf_output_unlock(handle);
2887 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2890 * only top level events have the pid namespace they were created in
2893 event = event->parent;
2895 return task_tgid_nr_ns(p, event->ns);
2898 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2901 * only top level events have the pid namespace they were created in
2904 event = event->parent;
2906 return task_pid_nr_ns(p, event->ns);
2909 static void perf_output_read_one(struct perf_output_handle *handle,
2910 struct perf_event *event)
2912 u64 read_format = event->attr.read_format;
2916 values[n++] = atomic64_read(&event->count);
2917 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2918 values[n++] = event->total_time_enabled +
2919 atomic64_read(&event->child_total_time_enabled);
2921 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2922 values[n++] = event->total_time_running +
2923 atomic64_read(&event->child_total_time_running);
2925 if (read_format & PERF_FORMAT_ID)
2926 values[n++] = primary_event_id(event);
2928 perf_output_copy(handle, values, n * sizeof(u64));
2932 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2934 static void perf_output_read_group(struct perf_output_handle *handle,
2935 struct perf_event *event)
2937 struct perf_event *leader = event->group_leader, *sub;
2938 u64 read_format = event->attr.read_format;
2942 values[n++] = 1 + leader->nr_siblings;
2944 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2945 values[n++] = leader->total_time_enabled;
2947 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2948 values[n++] = leader->total_time_running;
2950 if (leader != event)
2951 leader->pmu->read(leader);
2953 values[n++] = atomic64_read(&leader->count);
2954 if (read_format & PERF_FORMAT_ID)
2955 values[n++] = primary_event_id(leader);
2957 perf_output_copy(handle, values, n * sizeof(u64));
2959 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2963 sub->pmu->read(sub);
2965 values[n++] = atomic64_read(&sub->count);
2966 if (read_format & PERF_FORMAT_ID)
2967 values[n++] = primary_event_id(sub);
2969 perf_output_copy(handle, values, n * sizeof(u64));
2973 static void perf_output_read(struct perf_output_handle *handle,
2974 struct perf_event *event)
2976 if (event->attr.read_format & PERF_FORMAT_GROUP)
2977 perf_output_read_group(handle, event);
2979 perf_output_read_one(handle, event);
2982 void perf_output_sample(struct perf_output_handle *handle,
2983 struct perf_event_header *header,
2984 struct perf_sample_data *data,
2985 struct perf_event *event)
2987 u64 sample_type = data->type;
2989 perf_output_put(handle, *header);
2991 if (sample_type & PERF_SAMPLE_IP)
2992 perf_output_put(handle, data->ip);
2994 if (sample_type & PERF_SAMPLE_TID)
2995 perf_output_put(handle, data->tid_entry);
2997 if (sample_type & PERF_SAMPLE_TIME)
2998 perf_output_put(handle, data->time);
3000 if (sample_type & PERF_SAMPLE_ADDR)
3001 perf_output_put(handle, data->addr);
3003 if (sample_type & PERF_SAMPLE_ID)
3004 perf_output_put(handle, data->id);
3006 if (sample_type & PERF_SAMPLE_STREAM_ID)
3007 perf_output_put(handle, data->stream_id);
3009 if (sample_type & PERF_SAMPLE_CPU)
3010 perf_output_put(handle, data->cpu_entry);
3012 if (sample_type & PERF_SAMPLE_PERIOD)
3013 perf_output_put(handle, data->period);
3015 if (sample_type & PERF_SAMPLE_READ)
3016 perf_output_read(handle, event);
3018 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3019 if (data->callchain) {
3022 if (data->callchain)
3023 size += data->callchain->nr;
3025 size *= sizeof(u64);
3027 perf_output_copy(handle, data->callchain, size);
3030 perf_output_put(handle, nr);
3034 if (sample_type & PERF_SAMPLE_RAW) {
3036 perf_output_put(handle, data->raw->size);
3037 perf_output_copy(handle, data->raw->data,
3044 .size = sizeof(u32),
3047 perf_output_put(handle, raw);
3052 void perf_prepare_sample(struct perf_event_header *header,
3053 struct perf_sample_data *data,
3054 struct perf_event *event,
3055 struct pt_regs *regs)
3057 u64 sample_type = event->attr.sample_type;
3059 data->type = sample_type;
3061 header->type = PERF_RECORD_SAMPLE;
3062 header->size = sizeof(*header);
3065 header->misc |= perf_misc_flags(regs);
3067 if (sample_type & PERF_SAMPLE_IP) {
3068 data->ip = perf_instruction_pointer(regs);
3070 header->size += sizeof(data->ip);
3073 if (sample_type & PERF_SAMPLE_TID) {
3074 /* namespace issues */
3075 data->tid_entry.pid = perf_event_pid(event, current);
3076 data->tid_entry.tid = perf_event_tid(event, current);
3078 header->size += sizeof(data->tid_entry);
3081 if (sample_type & PERF_SAMPLE_TIME) {
3082 data->time = perf_clock();
3084 header->size += sizeof(data->time);
3087 if (sample_type & PERF_SAMPLE_ADDR)
3088 header->size += sizeof(data->addr);
3090 if (sample_type & PERF_SAMPLE_ID) {
3091 data->id = primary_event_id(event);
3093 header->size += sizeof(data->id);
3096 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3097 data->stream_id = event->id;
3099 header->size += sizeof(data->stream_id);
3102 if (sample_type & PERF_SAMPLE_CPU) {
3103 data->cpu_entry.cpu = raw_smp_processor_id();
3104 data->cpu_entry.reserved = 0;
3106 header->size += sizeof(data->cpu_entry);
3109 if (sample_type & PERF_SAMPLE_PERIOD)
3110 header->size += sizeof(data->period);
3112 if (sample_type & PERF_SAMPLE_READ)
3113 header->size += perf_event_read_size(event);
3115 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3118 data->callchain = perf_callchain(regs);
3120 if (data->callchain)
3121 size += data->callchain->nr;
3123 header->size += size * sizeof(u64);
3126 if (sample_type & PERF_SAMPLE_RAW) {
3127 int size = sizeof(u32);
3130 size += data->raw->size;
3132 size += sizeof(u32);
3134 WARN_ON_ONCE(size & (sizeof(u64)-1));
3135 header->size += size;
3139 static void perf_event_output(struct perf_event *event, int nmi,
3140 struct perf_sample_data *data,
3141 struct pt_regs *regs)
3143 struct perf_output_handle handle;
3144 struct perf_event_header header;
3146 perf_prepare_sample(&header, data, event, regs);
3148 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3151 perf_output_sample(&handle, &header, data, event);
3153 perf_output_end(&handle);
3160 struct perf_read_event {
3161 struct perf_event_header header;
3168 perf_event_read_event(struct perf_event *event,
3169 struct task_struct *task)
3171 struct perf_output_handle handle;
3172 struct perf_read_event read_event = {
3174 .type = PERF_RECORD_READ,
3176 .size = sizeof(read_event) + perf_event_read_size(event),
3178 .pid = perf_event_pid(event, task),
3179 .tid = perf_event_tid(event, task),
3183 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3187 perf_output_put(&handle, read_event);
3188 perf_output_read(&handle, event);
3190 perf_output_end(&handle);
3194 * task tracking -- fork/exit
3196 * enabled by: attr.comm | attr.mmap | attr.task
3199 struct perf_task_event {
3200 struct task_struct *task;
3201 struct perf_event_context *task_ctx;
3204 struct perf_event_header header;
3214 static void perf_event_task_output(struct perf_event *event,
3215 struct perf_task_event *task_event)
3217 struct perf_output_handle handle;
3219 struct task_struct *task = task_event->task;
3222 size = task_event->event_id.header.size;
3223 ret = perf_output_begin(&handle, event, size, 0, 0);
3228 task_event->event_id.pid = perf_event_pid(event, task);
3229 task_event->event_id.ppid = perf_event_pid(event, current);
3231 task_event->event_id.tid = perf_event_tid(event, task);
3232 task_event->event_id.ptid = perf_event_tid(event, current);
3234 task_event->event_id.time = perf_clock();
3236 perf_output_put(&handle, task_event->event_id);
3238 perf_output_end(&handle);
3241 static int perf_event_task_match(struct perf_event *event)
3243 if (event->attr.comm || event->attr.mmap || event->attr.task)
3249 static void perf_event_task_ctx(struct perf_event_context *ctx,
3250 struct perf_task_event *task_event)
3252 struct perf_event *event;
3254 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3255 if (perf_event_task_match(event))
3256 perf_event_task_output(event, task_event);
3260 static void perf_event_task_event(struct perf_task_event *task_event)
3262 struct perf_cpu_context *cpuctx;
3263 struct perf_event_context *ctx = task_event->task_ctx;
3266 cpuctx = &get_cpu_var(perf_cpu_context);
3267 perf_event_task_ctx(&cpuctx->ctx, task_event);
3268 put_cpu_var(perf_cpu_context);
3271 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3273 perf_event_task_ctx(ctx, task_event);
3277 static void perf_event_task(struct task_struct *task,
3278 struct perf_event_context *task_ctx,
3281 struct perf_task_event task_event;
3283 if (!atomic_read(&nr_comm_events) &&
3284 !atomic_read(&nr_mmap_events) &&
3285 !atomic_read(&nr_task_events))
3288 task_event = (struct perf_task_event){
3290 .task_ctx = task_ctx,
3293 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3295 .size = sizeof(task_event.event_id),
3304 perf_event_task_event(&task_event);
3307 void perf_event_fork(struct task_struct *task)
3309 perf_event_task(task, NULL, 1);
3316 struct perf_comm_event {
3317 struct task_struct *task;
3322 struct perf_event_header header;
3329 static void perf_event_comm_output(struct perf_event *event,
3330 struct perf_comm_event *comm_event)
3332 struct perf_output_handle handle;
3333 int size = comm_event->event_id.header.size;
3334 int ret = perf_output_begin(&handle, event, size, 0, 0);
3339 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3340 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3342 perf_output_put(&handle, comm_event->event_id);
3343 perf_output_copy(&handle, comm_event->comm,
3344 comm_event->comm_size);
3345 perf_output_end(&handle);
3348 static int perf_event_comm_match(struct perf_event *event)
3350 if (event->attr.comm)
3356 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3357 struct perf_comm_event *comm_event)
3359 struct perf_event *event;
3361 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3362 if (perf_event_comm_match(event))
3363 perf_event_comm_output(event, comm_event);
3367 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3369 struct perf_cpu_context *cpuctx;
3370 struct perf_event_context *ctx;
3372 char comm[TASK_COMM_LEN];
3374 memset(comm, 0, sizeof(comm));
3375 strncpy(comm, comm_event->task->comm, sizeof(comm));
3376 size = ALIGN(strlen(comm)+1, sizeof(u64));
3378 comm_event->comm = comm;
3379 comm_event->comm_size = size;
3381 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3384 cpuctx = &get_cpu_var(perf_cpu_context);
3385 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3386 put_cpu_var(perf_cpu_context);
3389 * doesn't really matter which of the child contexts the
3390 * events ends up in.
3392 ctx = rcu_dereference(current->perf_event_ctxp);
3394 perf_event_comm_ctx(ctx, comm_event);
3398 void perf_event_comm(struct task_struct *task)
3400 struct perf_comm_event comm_event;
3402 if (task->perf_event_ctxp)
3403 perf_event_enable_on_exec(task);
3405 if (!atomic_read(&nr_comm_events))
3408 comm_event = (struct perf_comm_event){
3414 .type = PERF_RECORD_COMM,
3423 perf_event_comm_event(&comm_event);
3430 struct perf_mmap_event {
3431 struct vm_area_struct *vma;
3433 const char *file_name;
3437 struct perf_event_header header;
3447 static void perf_event_mmap_output(struct perf_event *event,
3448 struct perf_mmap_event *mmap_event)
3450 struct perf_output_handle handle;
3451 int size = mmap_event->event_id.header.size;
3452 int ret = perf_output_begin(&handle, event, size, 0, 0);
3457 mmap_event->event_id.pid = perf_event_pid(event, current);
3458 mmap_event->event_id.tid = perf_event_tid(event, current);
3460 perf_output_put(&handle, mmap_event->event_id);
3461 perf_output_copy(&handle, mmap_event->file_name,
3462 mmap_event->file_size);
3463 perf_output_end(&handle);
3466 static int perf_event_mmap_match(struct perf_event *event,
3467 struct perf_mmap_event *mmap_event)
3469 if (event->attr.mmap)
3475 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3476 struct perf_mmap_event *mmap_event)
3478 struct perf_event *event;
3480 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3481 if (perf_event_mmap_match(event, mmap_event))
3482 perf_event_mmap_output(event, mmap_event);
3486 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3488 struct perf_cpu_context *cpuctx;
3489 struct perf_event_context *ctx;
3490 struct vm_area_struct *vma = mmap_event->vma;
3491 struct file *file = vma->vm_file;
3497 memset(tmp, 0, sizeof(tmp));
3501 * d_path works from the end of the buffer backwards, so we
3502 * need to add enough zero bytes after the string to handle
3503 * the 64bit alignment we do later.
3505 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3507 name = strncpy(tmp, "//enomem", sizeof(tmp));
3510 name = d_path(&file->f_path, buf, PATH_MAX);
3512 name = strncpy(tmp, "//toolong", sizeof(tmp));
3516 if (arch_vma_name(mmap_event->vma)) {
3517 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3523 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3527 name = strncpy(tmp, "//anon", sizeof(tmp));
3532 size = ALIGN(strlen(name)+1, sizeof(u64));
3534 mmap_event->file_name = name;
3535 mmap_event->file_size = size;
3537 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3540 cpuctx = &get_cpu_var(perf_cpu_context);
3541 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3542 put_cpu_var(perf_cpu_context);
3545 * doesn't really matter which of the child contexts the
3546 * events ends up in.
3548 ctx = rcu_dereference(current->perf_event_ctxp);
3550 perf_event_mmap_ctx(ctx, mmap_event);
3556 void __perf_event_mmap(struct vm_area_struct *vma)
3558 struct perf_mmap_event mmap_event;
3560 if (!atomic_read(&nr_mmap_events))
3563 mmap_event = (struct perf_mmap_event){
3569 .type = PERF_RECORD_MMAP,
3575 .start = vma->vm_start,
3576 .len = vma->vm_end - vma->vm_start,
3577 .pgoff = vma->vm_pgoff,
3581 perf_event_mmap_event(&mmap_event);
3585 * IRQ throttle logging
3588 static void perf_log_throttle(struct perf_event *event, int enable)
3590 struct perf_output_handle handle;
3594 struct perf_event_header header;
3598 } throttle_event = {
3600 .type = PERF_RECORD_THROTTLE,
3602 .size = sizeof(throttle_event),
3604 .time = perf_clock(),
3605 .id = primary_event_id(event),
3606 .stream_id = event->id,
3610 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3612 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3616 perf_output_put(&handle, throttle_event);
3617 perf_output_end(&handle);
3621 * Generic event overflow handling, sampling.
3624 static int __perf_event_overflow(struct perf_event *event, int nmi,
3625 int throttle, struct perf_sample_data *data,
3626 struct pt_regs *regs)
3628 int events = atomic_read(&event->event_limit);
3629 struct hw_perf_event *hwc = &event->hw;
3632 throttle = (throttle && event->pmu->unthrottle != NULL);
3637 if (hwc->interrupts != MAX_INTERRUPTS) {
3639 if (HZ * hwc->interrupts >
3640 (u64)sysctl_perf_event_sample_rate) {
3641 hwc->interrupts = MAX_INTERRUPTS;
3642 perf_log_throttle(event, 0);
3647 * Keep re-disabling events even though on the previous
3648 * pass we disabled it - just in case we raced with a
3649 * sched-in and the event got enabled again:
3655 if (event->attr.freq) {
3656 u64 now = perf_clock();
3657 s64 delta = now - hwc->freq_stamp;
3659 hwc->freq_stamp = now;
3661 if (delta > 0 && delta < TICK_NSEC)
3662 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3666 * XXX event_limit might not quite work as expected on inherited
3670 event->pending_kill = POLL_IN;
3671 if (events && atomic_dec_and_test(&event->event_limit)) {
3673 event->pending_kill = POLL_HUP;
3675 event->pending_disable = 1;
3676 perf_pending_queue(&event->pending,
3677 perf_pending_event);
3679 perf_event_disable(event);
3682 if (event->overflow_handler)
3683 event->overflow_handler(event, nmi, data, regs);
3685 perf_event_output(event, nmi, data, regs);
3690 int perf_event_overflow(struct perf_event *event, int nmi,
3691 struct perf_sample_data *data,
3692 struct pt_regs *regs)
3694 return __perf_event_overflow(event, nmi, 1, data, regs);
3698 * Generic software event infrastructure
3702 * We directly increment event->count and keep a second value in
3703 * event->hw.period_left to count intervals. This period event
3704 * is kept in the range [-sample_period, 0] so that we can use the
3708 static u64 perf_swevent_set_period(struct perf_event *event)
3710 struct hw_perf_event *hwc = &event->hw;
3711 u64 period = hwc->last_period;
3715 hwc->last_period = hwc->sample_period;
3718 old = val = atomic64_read(&hwc->period_left);
3722 nr = div64_u64(period + val, period);
3723 offset = nr * period;
3725 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3731 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3732 int nmi, struct perf_sample_data *data,
3733 struct pt_regs *regs)
3735 struct hw_perf_event *hwc = &event->hw;
3738 data->period = event->hw.last_period;
3740 overflow = perf_swevent_set_period(event);
3742 if (hwc->interrupts == MAX_INTERRUPTS)
3745 for (; overflow; overflow--) {
3746 if (__perf_event_overflow(event, nmi, throttle,
3749 * We inhibit the overflow from happening when
3750 * hwc->interrupts == MAX_INTERRUPTS.
3758 static void perf_swevent_unthrottle(struct perf_event *event)
3761 * Nothing to do, we already reset hwc->interrupts.
3765 static void perf_swevent_add(struct perf_event *event, u64 nr,
3766 int nmi, struct perf_sample_data *data,
3767 struct pt_regs *regs)
3769 struct hw_perf_event *hwc = &event->hw;
3771 atomic64_add(nr, &event->count);
3776 if (!hwc->sample_period)
3779 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3780 return perf_swevent_overflow(event, 1, nmi, data, regs);
3782 if (atomic64_add_negative(nr, &hwc->period_left))
3785 perf_swevent_overflow(event, 0, nmi, data, regs);
3788 static int perf_swevent_is_counting(struct perf_event *event)
3791 * The event is active, we're good!
3793 if (event->state == PERF_EVENT_STATE_ACTIVE)
3797 * The event is off/error, not counting.
3799 if (event->state != PERF_EVENT_STATE_INACTIVE)
3803 * The event is inactive, if the context is active
3804 * we're part of a group that didn't make it on the 'pmu',
3807 if (event->ctx->is_active)
3811 * We're inactive and the context is too, this means the
3812 * task is scheduled out, we're counting events that happen
3813 * to us, like migration events.
3818 static int perf_tp_event_match(struct perf_event *event,
3819 struct perf_sample_data *data);
3821 static int perf_swevent_match(struct perf_event *event,
3822 enum perf_type_id type,
3824 struct perf_sample_data *data,
3825 struct pt_regs *regs)
3827 if (!perf_swevent_is_counting(event))
3830 if (event->attr.type != type)
3832 if (event->attr.config != event_id)
3836 if (event->attr.exclude_user && user_mode(regs))
3839 if (event->attr.exclude_kernel && !user_mode(regs))
3843 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3844 !perf_tp_event_match(event, data))
3850 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3851 enum perf_type_id type,
3852 u32 event_id, u64 nr, int nmi,
3853 struct perf_sample_data *data,
3854 struct pt_regs *regs)
3856 struct perf_event *event;
3858 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3859 if (perf_swevent_match(event, type, event_id, data, regs))
3860 perf_swevent_add(event, nr, nmi, data, regs);
3864 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3867 return &cpuctx->recursion[3];
3870 return &cpuctx->recursion[2];
3873 return &cpuctx->recursion[1];
3875 return &cpuctx->recursion[0];
3878 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3880 struct perf_sample_data *data,
3881 struct pt_regs *regs)
3883 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3884 int *recursion = perf_swevent_recursion_context(cpuctx);
3885 struct perf_event_context *ctx;
3894 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3895 nr, nmi, data, regs);
3897 * doesn't really matter which of the child contexts the
3898 * events ends up in.
3900 ctx = rcu_dereference(current->perf_event_ctxp);
3902 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3909 put_cpu_var(perf_cpu_context);
3912 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3913 struct pt_regs *regs, u64 addr)
3915 struct perf_sample_data data = {
3919 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3923 static void perf_swevent_read(struct perf_event *event)
3927 static int perf_swevent_enable(struct perf_event *event)
3929 struct hw_perf_event *hwc = &event->hw;
3931 if (hwc->sample_period) {
3932 hwc->last_period = hwc->sample_period;
3933 perf_swevent_set_period(event);
3938 static void perf_swevent_disable(struct perf_event *event)
3942 static const struct pmu perf_ops_generic = {
3943 .enable = perf_swevent_enable,
3944 .disable = perf_swevent_disable,
3945 .read = perf_swevent_read,
3946 .unthrottle = perf_swevent_unthrottle,
3950 * hrtimer based swevent callback
3953 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3955 enum hrtimer_restart ret = HRTIMER_RESTART;
3956 struct perf_sample_data data;
3957 struct pt_regs *regs;
3958 struct perf_event *event;
3961 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
3962 event->pmu->read(event);
3965 regs = get_irq_regs();
3967 * In case we exclude kernel IPs or are somehow not in interrupt
3968 * context, provide the next best thing, the user IP.
3970 if ((event->attr.exclude_kernel || !regs) &&
3971 !event->attr.exclude_user)
3972 regs = task_pt_regs(current);
3975 if (!(event->attr.exclude_idle && current->pid == 0))
3976 if (perf_event_overflow(event, 0, &data, regs))
3977 ret = HRTIMER_NORESTART;
3980 period = max_t(u64, 10000, event->hw.sample_period);
3981 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3986 static void perf_swevent_start_hrtimer(struct perf_event *event)
3988 struct hw_perf_event *hwc = &event->hw;
3990 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3991 hwc->hrtimer.function = perf_swevent_hrtimer;
3992 if (hwc->sample_period) {
3995 if (hwc->remaining) {
3996 if (hwc->remaining < 0)
3999 period = hwc->remaining;
4002 period = max_t(u64, 10000, hwc->sample_period);
4004 __hrtimer_start_range_ns(&hwc->hrtimer,
4005 ns_to_ktime(period), 0,
4006 HRTIMER_MODE_REL, 0);
4010 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4012 struct hw_perf_event *hwc = &event->hw;
4014 if (hwc->sample_period) {
4015 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4016 hwc->remaining = ktime_to_ns(remaining);
4018 hrtimer_cancel(&hwc->hrtimer);
4023 * Software event: cpu wall time clock
4026 static void cpu_clock_perf_event_update(struct perf_event *event)
4028 int cpu = raw_smp_processor_id();
4032 now = cpu_clock(cpu);
4033 prev = atomic64_read(&event->hw.prev_count);
4034 atomic64_set(&event->hw.prev_count, now);
4035 atomic64_add(now - prev, &event->count);
4038 static int cpu_clock_perf_event_enable(struct perf_event *event)
4040 struct hw_perf_event *hwc = &event->hw;
4041 int cpu = raw_smp_processor_id();
4043 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4044 perf_swevent_start_hrtimer(event);
4049 static void cpu_clock_perf_event_disable(struct perf_event *event)
4051 perf_swevent_cancel_hrtimer(event);
4052 cpu_clock_perf_event_update(event);
4055 static void cpu_clock_perf_event_read(struct perf_event *event)
4057 cpu_clock_perf_event_update(event);
4060 static const struct pmu perf_ops_cpu_clock = {
4061 .enable = cpu_clock_perf_event_enable,
4062 .disable = cpu_clock_perf_event_disable,
4063 .read = cpu_clock_perf_event_read,
4067 * Software event: task time clock
4070 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4075 prev = atomic64_xchg(&event->hw.prev_count, now);
4077 atomic64_add(delta, &event->count);
4080 static int task_clock_perf_event_enable(struct perf_event *event)
4082 struct hw_perf_event *hwc = &event->hw;
4085 now = event->ctx->time;
4087 atomic64_set(&hwc->prev_count, now);
4089 perf_swevent_start_hrtimer(event);
4094 static void task_clock_perf_event_disable(struct perf_event *event)
4096 perf_swevent_cancel_hrtimer(event);
4097 task_clock_perf_event_update(event, event->ctx->time);
4101 static void task_clock_perf_event_read(struct perf_event *event)
4106 update_context_time(event->ctx);
4107 time = event->ctx->time;
4109 u64 now = perf_clock();
4110 u64 delta = now - event->ctx->timestamp;
4111 time = event->ctx->time + delta;
4114 task_clock_perf_event_update(event, time);
4117 static const struct pmu perf_ops_task_clock = {
4118 .enable = task_clock_perf_event_enable,
4119 .disable = task_clock_perf_event_disable,
4120 .read = task_clock_perf_event_read,
4123 #ifdef CONFIG_EVENT_PROFILE
4125 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4128 struct perf_raw_record raw = {
4133 struct perf_sample_data data = {
4138 struct pt_regs *regs = get_irq_regs();
4141 regs = task_pt_regs(current);
4143 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4146 EXPORT_SYMBOL_GPL(perf_tp_event);
4148 static int perf_tp_event_match(struct perf_event *event,
4149 struct perf_sample_data *data)
4151 void *record = data->raw->data;
4153 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4158 static void tp_perf_event_destroy(struct perf_event *event)
4160 ftrace_profile_disable(event->attr.config);
4163 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4166 * Raw tracepoint data is a severe data leak, only allow root to
4169 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4170 perf_paranoid_tracepoint_raw() &&
4171 !capable(CAP_SYS_ADMIN))
4172 return ERR_PTR(-EPERM);
4174 if (ftrace_profile_enable(event->attr.config))
4177 event->destroy = tp_perf_event_destroy;
4179 return &perf_ops_generic;
4182 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4187 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4190 filter_str = strndup_user(arg, PAGE_SIZE);
4191 if (IS_ERR(filter_str))
4192 return PTR_ERR(filter_str);
4194 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4200 static void perf_event_free_filter(struct perf_event *event)
4202 ftrace_profile_free_filter(event);
4207 static int perf_tp_event_match(struct perf_event *event,
4208 struct perf_sample_data *data)
4213 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4218 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4223 static void perf_event_free_filter(struct perf_event *event)
4227 #endif /* CONFIG_EVENT_PROFILE */
4229 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4230 static void bp_perf_event_destroy(struct perf_event *event)
4232 release_bp_slot(event);
4235 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4239 * The breakpoint is already filled if we haven't created the counter
4240 * through perf syscall
4241 * FIXME: manage to get trigerred to NULL if it comes from syscalls
4244 err = register_perf_hw_breakpoint(bp);
4246 err = __register_perf_hw_breakpoint(bp);
4248 return ERR_PTR(err);
4250 bp->destroy = bp_perf_event_destroy;
4252 return &perf_ops_bp;
4255 void perf_bp_event(struct perf_event *bp, void *regs)
4260 static void bp_perf_event_destroy(struct perf_event *event)
4264 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4269 void perf_bp_event(struct perf_event *bp, void *regs)
4274 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4276 static void sw_perf_event_destroy(struct perf_event *event)
4278 u64 event_id = event->attr.config;
4280 WARN_ON(event->parent);
4282 atomic_dec(&perf_swevent_enabled[event_id]);
4285 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4287 const struct pmu *pmu = NULL;
4288 u64 event_id = event->attr.config;
4291 * Software events (currently) can't in general distinguish
4292 * between user, kernel and hypervisor events.
4293 * However, context switches and cpu migrations are considered
4294 * to be kernel events, and page faults are never hypervisor
4298 case PERF_COUNT_SW_CPU_CLOCK:
4299 pmu = &perf_ops_cpu_clock;
4302 case PERF_COUNT_SW_TASK_CLOCK:
4304 * If the user instantiates this as a per-cpu event,
4305 * use the cpu_clock event instead.
4307 if (event->ctx->task)
4308 pmu = &perf_ops_task_clock;
4310 pmu = &perf_ops_cpu_clock;
4313 case PERF_COUNT_SW_PAGE_FAULTS:
4314 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4315 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4316 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4317 case PERF_COUNT_SW_CPU_MIGRATIONS:
4318 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4319 case PERF_COUNT_SW_EMULATION_FAULTS:
4320 if (!event->parent) {
4321 atomic_inc(&perf_swevent_enabled[event_id]);
4322 event->destroy = sw_perf_event_destroy;
4324 pmu = &perf_ops_generic;
4332 * Allocate and initialize a event structure
4334 static struct perf_event *
4335 perf_event_alloc(struct perf_event_attr *attr,
4337 struct perf_event_context *ctx,
4338 struct perf_event *group_leader,
4339 struct perf_event *parent_event,
4340 perf_callback_t callback,
4343 const struct pmu *pmu;
4344 struct perf_event *event;
4345 struct hw_perf_event *hwc;
4348 event = kzalloc(sizeof(*event), gfpflags);
4350 return ERR_PTR(-ENOMEM);
4353 * Single events are their own group leaders, with an
4354 * empty sibling list:
4357 group_leader = event;
4359 mutex_init(&event->child_mutex);
4360 INIT_LIST_HEAD(&event->child_list);
4362 INIT_LIST_HEAD(&event->group_entry);
4363 INIT_LIST_HEAD(&event->event_entry);
4364 INIT_LIST_HEAD(&event->sibling_list);
4365 init_waitqueue_head(&event->waitq);
4367 mutex_init(&event->mmap_mutex);
4370 event->attr = *attr;
4371 event->group_leader = group_leader;
4376 event->parent = parent_event;
4378 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4379 event->id = atomic64_inc_return(&perf_event_id);
4381 event->state = PERF_EVENT_STATE_INACTIVE;
4383 if (!callback && parent_event)
4384 callback = parent_event->callback;
4386 event->callback = callback;
4389 event->state = PERF_EVENT_STATE_OFF;
4394 hwc->sample_period = attr->sample_period;
4395 if (attr->freq && attr->sample_freq)
4396 hwc->sample_period = 1;
4397 hwc->last_period = hwc->sample_period;
4399 atomic64_set(&hwc->period_left, hwc->sample_period);
4402 * we currently do not support PERF_FORMAT_GROUP on inherited events
4404 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4407 switch (attr->type) {
4409 case PERF_TYPE_HARDWARE:
4410 case PERF_TYPE_HW_CACHE:
4411 pmu = hw_perf_event_init(event);
4414 case PERF_TYPE_SOFTWARE:
4415 pmu = sw_perf_event_init(event);
4418 case PERF_TYPE_TRACEPOINT:
4419 pmu = tp_perf_event_init(event);
4422 case PERF_TYPE_BREAKPOINT:
4423 pmu = bp_perf_event_init(event);
4434 else if (IS_ERR(pmu))
4439 put_pid_ns(event->ns);
4441 return ERR_PTR(err);
4446 if (!event->parent) {
4447 atomic_inc(&nr_events);
4448 if (event->attr.mmap)
4449 atomic_inc(&nr_mmap_events);
4450 if (event->attr.comm)
4451 atomic_inc(&nr_comm_events);
4452 if (event->attr.task)
4453 atomic_inc(&nr_task_events);
4459 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4460 struct perf_event_attr *attr)
4465 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4469 * zero the full structure, so that a short copy will be nice.
4471 memset(attr, 0, sizeof(*attr));
4473 ret = get_user(size, &uattr->size);
4477 if (size > PAGE_SIZE) /* silly large */
4480 if (!size) /* abi compat */
4481 size = PERF_ATTR_SIZE_VER0;
4483 if (size < PERF_ATTR_SIZE_VER0)
4487 * If we're handed a bigger struct than we know of,
4488 * ensure all the unknown bits are 0 - i.e. new
4489 * user-space does not rely on any kernel feature
4490 * extensions we dont know about yet.
4492 if (size > sizeof(*attr)) {
4493 unsigned char __user *addr;
4494 unsigned char __user *end;
4497 addr = (void __user *)uattr + sizeof(*attr);
4498 end = (void __user *)uattr + size;
4500 for (; addr < end; addr++) {
4501 ret = get_user(val, addr);
4507 size = sizeof(*attr);
4510 ret = copy_from_user(attr, uattr, size);
4515 * If the type exists, the corresponding creation will verify
4518 if (attr->type >= PERF_TYPE_MAX)
4521 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4524 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4527 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4534 put_user(sizeof(*attr), &uattr->size);
4539 static int perf_event_set_output(struct perf_event *event, int output_fd)
4541 struct perf_event *output_event = NULL;
4542 struct file *output_file = NULL;
4543 struct perf_event *old_output;
4544 int fput_needed = 0;
4550 output_file = fget_light(output_fd, &fput_needed);
4554 if (output_file->f_op != &perf_fops)
4557 output_event = output_file->private_data;
4559 /* Don't chain output fds */
4560 if (output_event->output)
4563 /* Don't set an output fd when we already have an output channel */
4567 atomic_long_inc(&output_file->f_count);
4570 mutex_lock(&event->mmap_mutex);
4571 old_output = event->output;
4572 rcu_assign_pointer(event->output, output_event);
4573 mutex_unlock(&event->mmap_mutex);
4577 * we need to make sure no existing perf_output_*()
4578 * is still referencing this event.
4581 fput(old_output->filp);
4586 fput_light(output_file, fput_needed);
4591 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4593 * @attr_uptr: event_id type attributes for monitoring/sampling
4596 * @group_fd: group leader event fd
4598 SYSCALL_DEFINE5(perf_event_open,
4599 struct perf_event_attr __user *, attr_uptr,
4600 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4602 struct perf_event *event, *group_leader;
4603 struct perf_event_attr attr;
4604 struct perf_event_context *ctx;
4605 struct file *event_file = NULL;
4606 struct file *group_file = NULL;
4607 int fput_needed = 0;
4608 int fput_needed2 = 0;
4611 /* for future expandability... */
4612 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4615 err = perf_copy_attr(attr_uptr, &attr);
4619 if (!attr.exclude_kernel) {
4620 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4625 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4630 * Get the target context (task or percpu):
4632 ctx = find_get_context(pid, cpu);
4634 return PTR_ERR(ctx);
4637 * Look up the group leader (we will attach this event to it):
4639 group_leader = NULL;
4640 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4642 group_file = fget_light(group_fd, &fput_needed);
4644 goto err_put_context;
4645 if (group_file->f_op != &perf_fops)
4646 goto err_put_context;
4648 group_leader = group_file->private_data;
4650 * Do not allow a recursive hierarchy (this new sibling
4651 * becoming part of another group-sibling):
4653 if (group_leader->group_leader != group_leader)
4654 goto err_put_context;
4656 * Do not allow to attach to a group in a different
4657 * task or CPU context:
4659 if (group_leader->ctx != ctx)
4660 goto err_put_context;
4662 * Only a group leader can be exclusive or pinned
4664 if (attr.exclusive || attr.pinned)
4665 goto err_put_context;
4668 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4669 NULL, NULL, GFP_KERNEL);
4670 err = PTR_ERR(event);
4672 goto err_put_context;
4674 err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4676 goto err_free_put_context;
4678 event_file = fget_light(err, &fput_needed2);
4680 goto err_free_put_context;
4682 if (flags & PERF_FLAG_FD_OUTPUT) {
4683 err = perf_event_set_output(event, group_fd);
4685 goto err_fput_free_put_context;
4688 event->filp = event_file;
4689 WARN_ON_ONCE(ctx->parent_ctx);
4690 mutex_lock(&ctx->mutex);
4691 perf_install_in_context(ctx, event, cpu);
4693 mutex_unlock(&ctx->mutex);
4695 event->owner = current;
4696 get_task_struct(current);
4697 mutex_lock(¤t->perf_event_mutex);
4698 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4699 mutex_unlock(¤t->perf_event_mutex);
4701 err_fput_free_put_context:
4702 fput_light(event_file, fput_needed2);
4704 err_free_put_context:
4712 fput_light(group_file, fput_needed);
4718 * perf_event_create_kernel_counter
4720 * @attr: attributes of the counter to create
4721 * @cpu: cpu in which the counter is bound
4722 * @pid: task to profile
4725 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4726 pid_t pid, perf_callback_t callback)
4728 struct perf_event *event;
4729 struct perf_event_context *ctx;
4733 * Get the target context (task or percpu):
4736 ctx = find_get_context(pid, cpu);
4740 event = perf_event_alloc(attr, cpu, ctx, NULL,
4741 NULL, callback, GFP_KERNEL);
4742 err = PTR_ERR(event);
4744 goto err_put_context;
4747 WARN_ON_ONCE(ctx->parent_ctx);
4748 mutex_lock(&ctx->mutex);
4749 perf_install_in_context(ctx, event, cpu);
4751 mutex_unlock(&ctx->mutex);
4753 event->owner = current;
4754 get_task_struct(current);
4755 mutex_lock(¤t->perf_event_mutex);
4756 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4757 mutex_unlock(¤t->perf_event_mutex);
4767 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4770 * inherit a event from parent task to child task:
4772 static struct perf_event *
4773 inherit_event(struct perf_event *parent_event,
4774 struct task_struct *parent,
4775 struct perf_event_context *parent_ctx,
4776 struct task_struct *child,
4777 struct perf_event *group_leader,
4778 struct perf_event_context *child_ctx)
4780 struct perf_event *child_event;
4783 * Instead of creating recursive hierarchies of events,
4784 * we link inherited events back to the original parent,
4785 * which has a filp for sure, which we use as the reference
4788 if (parent_event->parent)
4789 parent_event = parent_event->parent;
4791 child_event = perf_event_alloc(&parent_event->attr,
4792 parent_event->cpu, child_ctx,
4793 group_leader, parent_event,
4795 if (IS_ERR(child_event))
4800 * Make the child state follow the state of the parent event,
4801 * not its attr.disabled bit. We hold the parent's mutex,
4802 * so we won't race with perf_event_{en, dis}able_family.
4804 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4805 child_event->state = PERF_EVENT_STATE_INACTIVE;
4807 child_event->state = PERF_EVENT_STATE_OFF;
4809 if (parent_event->attr.freq)
4810 child_event->hw.sample_period = parent_event->hw.sample_period;
4812 child_event->overflow_handler = parent_event->overflow_handler;
4815 * Link it up in the child's context:
4817 add_event_to_ctx(child_event, child_ctx);
4820 * Get a reference to the parent filp - we will fput it
4821 * when the child event exits. This is safe to do because
4822 * we are in the parent and we know that the filp still
4823 * exists and has a nonzero count:
4825 atomic_long_inc(&parent_event->filp->f_count);
4828 * Link this into the parent event's child list
4830 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4831 mutex_lock(&parent_event->child_mutex);
4832 list_add_tail(&child_event->child_list, &parent_event->child_list);
4833 mutex_unlock(&parent_event->child_mutex);
4838 static int inherit_group(struct perf_event *parent_event,
4839 struct task_struct *parent,
4840 struct perf_event_context *parent_ctx,
4841 struct task_struct *child,
4842 struct perf_event_context *child_ctx)
4844 struct perf_event *leader;
4845 struct perf_event *sub;
4846 struct perf_event *child_ctr;
4848 leader = inherit_event(parent_event, parent, parent_ctx,
4849 child, NULL, child_ctx);
4851 return PTR_ERR(leader);
4852 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4853 child_ctr = inherit_event(sub, parent, parent_ctx,
4854 child, leader, child_ctx);
4855 if (IS_ERR(child_ctr))
4856 return PTR_ERR(child_ctr);
4861 static void sync_child_event(struct perf_event *child_event,
4862 struct task_struct *child)
4864 struct perf_event *parent_event = child_event->parent;
4867 if (child_event->attr.inherit_stat)
4868 perf_event_read_event(child_event, child);
4870 child_val = atomic64_read(&child_event->count);
4873 * Add back the child's count to the parent's count:
4875 atomic64_add(child_val, &parent_event->count);
4876 atomic64_add(child_event->total_time_enabled,
4877 &parent_event->child_total_time_enabled);
4878 atomic64_add(child_event->total_time_running,
4879 &parent_event->child_total_time_running);
4882 * Remove this event from the parent's list
4884 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4885 mutex_lock(&parent_event->child_mutex);
4886 list_del_init(&child_event->child_list);
4887 mutex_unlock(&parent_event->child_mutex);
4890 * Release the parent event, if this was the last
4893 fput(parent_event->filp);
4897 __perf_event_exit_task(struct perf_event *child_event,
4898 struct perf_event_context *child_ctx,
4899 struct task_struct *child)
4901 struct perf_event *parent_event;
4903 update_event_times(child_event);
4904 perf_event_remove_from_context(child_event);
4906 parent_event = child_event->parent;
4908 * It can happen that parent exits first, and has events
4909 * that are still around due to the child reference. These
4910 * events need to be zapped - but otherwise linger.
4913 sync_child_event(child_event, child);
4914 free_event(child_event);
4919 * When a child task exits, feed back event values to parent events.
4921 void perf_event_exit_task(struct task_struct *child)
4923 struct perf_event *child_event, *tmp;
4924 struct perf_event_context *child_ctx;
4925 unsigned long flags;
4927 if (likely(!child->perf_event_ctxp)) {
4928 perf_event_task(child, NULL, 0);
4932 local_irq_save(flags);
4934 * We can't reschedule here because interrupts are disabled,
4935 * and either child is current or it is a task that can't be
4936 * scheduled, so we are now safe from rescheduling changing
4939 child_ctx = child->perf_event_ctxp;
4940 __perf_event_task_sched_out(child_ctx);
4943 * Take the context lock here so that if find_get_context is
4944 * reading child->perf_event_ctxp, we wait until it has
4945 * incremented the context's refcount before we do put_ctx below.
4947 spin_lock(&child_ctx->lock);
4948 child->perf_event_ctxp = NULL;
4950 * If this context is a clone; unclone it so it can't get
4951 * swapped to another process while we're removing all
4952 * the events from it.
4954 unclone_ctx(child_ctx);
4955 spin_unlock_irqrestore(&child_ctx->lock, flags);
4958 * Report the task dead after unscheduling the events so that we
4959 * won't get any samples after PERF_RECORD_EXIT. We can however still
4960 * get a few PERF_RECORD_READ events.
4962 perf_event_task(child, child_ctx, 0);
4965 * We can recurse on the same lock type through:
4967 * __perf_event_exit_task()
4968 * sync_child_event()
4969 * fput(parent_event->filp)
4971 * mutex_lock(&ctx->mutex)
4973 * But since its the parent context it won't be the same instance.
4975 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4978 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4980 __perf_event_exit_task(child_event, child_ctx, child);
4983 * If the last event was a group event, it will have appended all
4984 * its siblings to the list, but we obtained 'tmp' before that which
4985 * will still point to the list head terminating the iteration.
4987 if (!list_empty(&child_ctx->group_list))
4990 mutex_unlock(&child_ctx->mutex);
4996 * free an unexposed, unused context as created by inheritance by
4997 * init_task below, used by fork() in case of fail.
4999 void perf_event_free_task(struct task_struct *task)
5001 struct perf_event_context *ctx = task->perf_event_ctxp;
5002 struct perf_event *event, *tmp;
5007 mutex_lock(&ctx->mutex);
5009 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5010 struct perf_event *parent = event->parent;
5012 if (WARN_ON_ONCE(!parent))
5015 mutex_lock(&parent->child_mutex);
5016 list_del_init(&event->child_list);
5017 mutex_unlock(&parent->child_mutex);
5021 list_del_event(event, ctx);
5025 if (!list_empty(&ctx->group_list))
5028 mutex_unlock(&ctx->mutex);
5034 * Initialize the perf_event context in task_struct
5036 int perf_event_init_task(struct task_struct *child)
5038 struct perf_event_context *child_ctx, *parent_ctx;
5039 struct perf_event_context *cloned_ctx;
5040 struct perf_event *event;
5041 struct task_struct *parent = current;
5042 int inherited_all = 1;
5045 child->perf_event_ctxp = NULL;
5047 mutex_init(&child->perf_event_mutex);
5048 INIT_LIST_HEAD(&child->perf_event_list);
5050 if (likely(!parent->perf_event_ctxp))
5054 * This is executed from the parent task context, so inherit
5055 * events that have been marked for cloning.
5056 * First allocate and initialize a context for the child.
5059 child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5063 __perf_event_init_context(child_ctx, child);
5064 child->perf_event_ctxp = child_ctx;
5065 get_task_struct(child);
5068 * If the parent's context is a clone, pin it so it won't get
5071 parent_ctx = perf_pin_task_context(parent);
5074 * No need to check if parent_ctx != NULL here; since we saw
5075 * it non-NULL earlier, the only reason for it to become NULL
5076 * is if we exit, and since we're currently in the middle of
5077 * a fork we can't be exiting at the same time.
5081 * Lock the parent list. No need to lock the child - not PID
5082 * hashed yet and not running, so nobody can access it.
5084 mutex_lock(&parent_ctx->mutex);
5087 * We dont have to disable NMIs - we are only looking at
5088 * the list, not manipulating it:
5090 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5092 if (!event->attr.inherit) {
5097 ret = inherit_group(event, parent, parent_ctx,
5105 if (inherited_all) {
5107 * Mark the child context as a clone of the parent
5108 * context, or of whatever the parent is a clone of.
5109 * Note that if the parent is a clone, it could get
5110 * uncloned at any point, but that doesn't matter
5111 * because the list of events and the generation
5112 * count can't have changed since we took the mutex.
5114 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5116 child_ctx->parent_ctx = cloned_ctx;
5117 child_ctx->parent_gen = parent_ctx->parent_gen;
5119 child_ctx->parent_ctx = parent_ctx;
5120 child_ctx->parent_gen = parent_ctx->generation;
5122 get_ctx(child_ctx->parent_ctx);
5125 mutex_unlock(&parent_ctx->mutex);
5127 perf_unpin_context(parent_ctx);
5132 static void __cpuinit perf_event_init_cpu(int cpu)
5134 struct perf_cpu_context *cpuctx;
5136 cpuctx = &per_cpu(perf_cpu_context, cpu);
5137 __perf_event_init_context(&cpuctx->ctx, NULL);
5139 spin_lock(&perf_resource_lock);
5140 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5141 spin_unlock(&perf_resource_lock);
5143 hw_perf_event_setup(cpu);
5146 #ifdef CONFIG_HOTPLUG_CPU
5147 static void __perf_event_exit_cpu(void *info)
5149 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5150 struct perf_event_context *ctx = &cpuctx->ctx;
5151 struct perf_event *event, *tmp;
5153 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5154 __perf_event_remove_from_context(event);
5156 static void perf_event_exit_cpu(int cpu)
5158 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5159 struct perf_event_context *ctx = &cpuctx->ctx;
5161 mutex_lock(&ctx->mutex);
5162 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5163 mutex_unlock(&ctx->mutex);
5166 static inline void perf_event_exit_cpu(int cpu) { }
5169 static int __cpuinit
5170 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5172 unsigned int cpu = (long)hcpu;
5176 case CPU_UP_PREPARE:
5177 case CPU_UP_PREPARE_FROZEN:
5178 perf_event_init_cpu(cpu);
5182 case CPU_ONLINE_FROZEN:
5183 hw_perf_event_setup_online(cpu);
5186 case CPU_DOWN_PREPARE:
5187 case CPU_DOWN_PREPARE_FROZEN:
5188 perf_event_exit_cpu(cpu);
5199 * This has to have a higher priority than migration_notifier in sched.c.
5201 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5202 .notifier_call = perf_cpu_notify,
5206 void __init perf_event_init(void)
5208 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5209 (void *)(long)smp_processor_id());
5210 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5211 (void *)(long)smp_processor_id());
5212 register_cpu_notifier(&perf_cpu_nb);
5215 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5217 return sprintf(buf, "%d\n", perf_reserved_percpu);
5221 perf_set_reserve_percpu(struct sysdev_class *class,
5225 struct perf_cpu_context *cpuctx;
5229 err = strict_strtoul(buf, 10, &val);
5232 if (val > perf_max_events)
5235 spin_lock(&perf_resource_lock);
5236 perf_reserved_percpu = val;
5237 for_each_online_cpu(cpu) {
5238 cpuctx = &per_cpu(perf_cpu_context, cpu);
5239 spin_lock_irq(&cpuctx->ctx.lock);
5240 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5241 perf_max_events - perf_reserved_percpu);
5242 cpuctx->max_pertask = mpt;
5243 spin_unlock_irq(&cpuctx->ctx.lock);
5245 spin_unlock(&perf_resource_lock);
5250 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5252 return sprintf(buf, "%d\n", perf_overcommit);
5256 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5261 err = strict_strtoul(buf, 10, &val);
5267 spin_lock(&perf_resource_lock);
5268 perf_overcommit = val;
5269 spin_unlock(&perf_resource_lock);
5274 static SYSDEV_CLASS_ATTR(
5277 perf_show_reserve_percpu,
5278 perf_set_reserve_percpu
5281 static SYSDEV_CLASS_ATTR(
5284 perf_show_overcommit,
5288 static struct attribute *perfclass_attrs[] = {
5289 &attr_reserve_percpu.attr,
5290 &attr_overcommit.attr,
5294 static struct attribute_group perfclass_attr_group = {
5295 .attrs = perfclass_attrs,
5296 .name = "perf_events",
5299 static int __init perf_event_sysfs_init(void)
5301 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5302 &perfclass_attr_group);
5304 device_initcall(perf_event_sysfs_init);