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);
249 static inline u64 perf_clock(void)
251 return cpu_clock(smp_processor_id());
255 * Update the record of the current time in a context.
257 static void update_context_time(struct perf_event_context *ctx)
259 u64 now = perf_clock();
261 ctx->time += now - ctx->timestamp;
262 ctx->timestamp = now;
266 * Update the total_time_enabled and total_time_running fields for a event.
268 static void update_event_times(struct perf_event *event)
270 struct perf_event_context *ctx = event->ctx;
273 if (event->state < PERF_EVENT_STATE_INACTIVE ||
274 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
280 run_end = event->tstamp_stopped;
282 event->total_time_enabled = run_end - event->tstamp_enabled;
284 if (event->state == PERF_EVENT_STATE_INACTIVE)
285 run_end = event->tstamp_stopped;
289 event->total_time_running = run_end - event->tstamp_running;
293 * Add a event from the lists for its context.
294 * Must be called with ctx->mutex and ctx->lock held.
297 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
299 struct perf_event *group_leader = event->group_leader;
302 * Depending on whether it is a standalone or sibling event,
303 * add it straight to the context's event list, or to the group
304 * leader's sibling list:
306 if (group_leader == event)
307 list_add_tail(&event->group_entry, &ctx->group_list);
309 list_add_tail(&event->group_entry, &group_leader->sibling_list);
310 group_leader->nr_siblings++;
313 list_add_rcu(&event->event_entry, &ctx->event_list);
315 if (event->attr.inherit_stat)
320 * Remove a event from the lists for its context.
321 * Must be called with ctx->mutex and ctx->lock held.
324 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
326 struct perf_event *sibling, *tmp;
328 if (list_empty(&event->group_entry))
331 if (event->attr.inherit_stat)
334 list_del_init(&event->group_entry);
335 list_del_rcu(&event->event_entry);
337 if (event->group_leader != event)
338 event->group_leader->nr_siblings--;
340 update_event_times(event);
341 event->state = PERF_EVENT_STATE_OFF;
344 * If this was a group event with sibling events then
345 * upgrade the siblings to singleton events by adding them
346 * to the context list directly:
348 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
350 list_move_tail(&sibling->group_entry, &ctx->group_list);
351 sibling->group_leader = sibling;
356 event_sched_out(struct perf_event *event,
357 struct perf_cpu_context *cpuctx,
358 struct perf_event_context *ctx)
360 if (event->state != PERF_EVENT_STATE_ACTIVE)
363 event->state = PERF_EVENT_STATE_INACTIVE;
364 if (event->pending_disable) {
365 event->pending_disable = 0;
366 event->state = PERF_EVENT_STATE_OFF;
368 event->tstamp_stopped = ctx->time;
369 event->pmu->disable(event);
372 if (!is_software_event(event))
373 cpuctx->active_oncpu--;
375 if (event->attr.exclusive || !cpuctx->active_oncpu)
376 cpuctx->exclusive = 0;
380 group_sched_out(struct perf_event *group_event,
381 struct perf_cpu_context *cpuctx,
382 struct perf_event_context *ctx)
384 struct perf_event *event;
386 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
389 event_sched_out(group_event, cpuctx, ctx);
392 * Schedule out siblings (if any):
394 list_for_each_entry(event, &group_event->sibling_list, group_entry)
395 event_sched_out(event, cpuctx, ctx);
397 if (group_event->attr.exclusive)
398 cpuctx->exclusive = 0;
402 * Cross CPU call to remove a performance event
404 * We disable the event on the hardware level first. After that we
405 * remove it from the context list.
407 static void __perf_event_remove_from_context(void *info)
409 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
410 struct perf_event *event = info;
411 struct perf_event_context *ctx = event->ctx;
414 * If this is a task context, we need to check whether it is
415 * the current task context of this cpu. If not it has been
416 * scheduled out before the smp call arrived.
418 if (ctx->task && cpuctx->task_ctx != ctx)
421 spin_lock(&ctx->lock);
423 * Protect the list operation against NMI by disabling the
424 * events on a global level.
428 event_sched_out(event, cpuctx, ctx);
430 list_del_event(event, ctx);
434 * Allow more per task events with respect to the
437 cpuctx->max_pertask =
438 min(perf_max_events - ctx->nr_events,
439 perf_max_events - perf_reserved_percpu);
443 spin_unlock(&ctx->lock);
448 * Remove the event from a task's (or a CPU's) list of events.
450 * Must be called with ctx->mutex held.
452 * CPU events are removed with a smp call. For task events we only
453 * call when the task is on a CPU.
455 * If event->ctx is a cloned context, callers must make sure that
456 * every task struct that event->ctx->task could possibly point to
457 * remains valid. This is OK when called from perf_release since
458 * that only calls us on the top-level context, which can't be a clone.
459 * When called from perf_event_exit_task, it's OK because the
460 * context has been detached from its task.
462 static void perf_event_remove_from_context(struct perf_event *event)
464 struct perf_event_context *ctx = event->ctx;
465 struct task_struct *task = ctx->task;
469 * Per cpu events are removed via an smp call and
470 * the removal is always sucessful.
472 smp_call_function_single(event->cpu,
473 __perf_event_remove_from_context,
479 task_oncpu_function_call(task, __perf_event_remove_from_context,
482 spin_lock_irq(&ctx->lock);
484 * If the context is active we need to retry the smp call.
486 if (ctx->nr_active && !list_empty(&event->group_entry)) {
487 spin_unlock_irq(&ctx->lock);
492 * The lock prevents that this context is scheduled in so we
493 * can remove the event safely, if the call above did not
496 if (!list_empty(&event->group_entry))
497 list_del_event(event, ctx);
498 spin_unlock_irq(&ctx->lock);
502 * Update total_time_enabled and total_time_running for all events in a group.
504 static void update_group_times(struct perf_event *leader)
506 struct perf_event *event;
508 update_event_times(leader);
509 list_for_each_entry(event, &leader->sibling_list, group_entry)
510 update_event_times(event);
514 * Cross CPU call to disable a performance event
516 static void __perf_event_disable(void *info)
518 struct perf_event *event = info;
519 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
520 struct perf_event_context *ctx = event->ctx;
523 * If this is a per-task event, need to check whether this
524 * event's task is the current task on this cpu.
526 if (ctx->task && cpuctx->task_ctx != ctx)
529 spin_lock(&ctx->lock);
532 * If the event is on, turn it off.
533 * If it is in error state, leave it in error state.
535 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
536 update_context_time(ctx);
537 update_group_times(event);
538 if (event == event->group_leader)
539 group_sched_out(event, cpuctx, ctx);
541 event_sched_out(event, cpuctx, ctx);
542 event->state = PERF_EVENT_STATE_OFF;
545 spin_unlock(&ctx->lock);
551 * If event->ctx is a cloned context, callers must make sure that
552 * every task struct that event->ctx->task could possibly point to
553 * remains valid. This condition is satisifed when called through
554 * perf_event_for_each_child or perf_event_for_each because they
555 * hold the top-level event's child_mutex, so any descendant that
556 * goes to exit will block in sync_child_event.
557 * When called from perf_pending_event it's OK because event->ctx
558 * is the current context on this CPU and preemption is disabled,
559 * hence we can't get into perf_event_task_sched_out for this context.
561 static void perf_event_disable(struct perf_event *event)
563 struct perf_event_context *ctx = event->ctx;
564 struct task_struct *task = ctx->task;
568 * Disable the event on the cpu that it's on
570 smp_call_function_single(event->cpu, __perf_event_disable,
576 task_oncpu_function_call(task, __perf_event_disable, event);
578 spin_lock_irq(&ctx->lock);
580 * If the event is still active, we need to retry the cross-call.
582 if (event->state == PERF_EVENT_STATE_ACTIVE) {
583 spin_unlock_irq(&ctx->lock);
588 * Since we have the lock this context can't be scheduled
589 * in, so we can change the state safely.
591 if (event->state == PERF_EVENT_STATE_INACTIVE) {
592 update_group_times(event);
593 event->state = PERF_EVENT_STATE_OFF;
596 spin_unlock_irq(&ctx->lock);
600 event_sched_in(struct perf_event *event,
601 struct perf_cpu_context *cpuctx,
602 struct perf_event_context *ctx,
605 if (event->state <= PERF_EVENT_STATE_OFF)
608 event->state = PERF_EVENT_STATE_ACTIVE;
609 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
611 * The new state must be visible before we turn it on in the hardware:
615 if (event->pmu->enable(event)) {
616 event->state = PERF_EVENT_STATE_INACTIVE;
621 event->tstamp_running += ctx->time - event->tstamp_stopped;
623 if (!is_software_event(event))
624 cpuctx->active_oncpu++;
627 if (event->attr.exclusive)
628 cpuctx->exclusive = 1;
634 group_sched_in(struct perf_event *group_event,
635 struct perf_cpu_context *cpuctx,
636 struct perf_event_context *ctx,
639 struct perf_event *event, *partial_group;
642 if (group_event->state == PERF_EVENT_STATE_OFF)
645 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
647 return ret < 0 ? ret : 0;
649 if (event_sched_in(group_event, cpuctx, ctx, cpu))
653 * Schedule in siblings as one group (if any):
655 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
656 if (event_sched_in(event, cpuctx, ctx, cpu)) {
657 partial_group = event;
666 * Groups can be scheduled in as one unit only, so undo any
667 * partial group before returning:
669 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
670 if (event == partial_group)
672 event_sched_out(event, cpuctx, ctx);
674 event_sched_out(group_event, cpuctx, ctx);
680 * Return 1 for a group consisting entirely of software events,
681 * 0 if the group contains any hardware events.
683 static int is_software_only_group(struct perf_event *leader)
685 struct perf_event *event;
687 if (!is_software_event(leader))
690 list_for_each_entry(event, &leader->sibling_list, group_entry)
691 if (!is_software_event(event))
698 * Work out whether we can put this event group on the CPU now.
700 static int group_can_go_on(struct perf_event *event,
701 struct perf_cpu_context *cpuctx,
705 * Groups consisting entirely of software events can always go on.
707 if (is_software_only_group(event))
710 * If an exclusive group is already on, no other hardware
713 if (cpuctx->exclusive)
716 * If this group is exclusive and there are already
717 * events on the CPU, it can't go on.
719 if (event->attr.exclusive && cpuctx->active_oncpu)
722 * Otherwise, try to add it if all previous groups were able
728 static void add_event_to_ctx(struct perf_event *event,
729 struct perf_event_context *ctx)
731 list_add_event(event, ctx);
732 event->tstamp_enabled = ctx->time;
733 event->tstamp_running = ctx->time;
734 event->tstamp_stopped = ctx->time;
738 * Cross CPU call to install and enable a performance event
740 * Must be called with ctx->mutex held
742 static void __perf_install_in_context(void *info)
744 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
745 struct perf_event *event = info;
746 struct perf_event_context *ctx = event->ctx;
747 struct perf_event *leader = event->group_leader;
748 int cpu = smp_processor_id();
752 * If this is a task context, we need to check whether it is
753 * the current task context of this cpu. If not it has been
754 * scheduled out before the smp call arrived.
755 * Or possibly this is the right context but it isn't
756 * on this cpu because it had no events.
758 if (ctx->task && cpuctx->task_ctx != ctx) {
759 if (cpuctx->task_ctx || ctx->task != current)
761 cpuctx->task_ctx = ctx;
764 spin_lock(&ctx->lock);
766 update_context_time(ctx);
769 * Protect the list operation against NMI by disabling the
770 * events on a global level. NOP for non NMI based events.
774 add_event_to_ctx(event, ctx);
777 * Don't put the event on if it is disabled or if
778 * it is in a group and the group isn't on.
780 if (event->state != PERF_EVENT_STATE_INACTIVE ||
781 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
785 * An exclusive event can't go on if there are already active
786 * hardware events, and no hardware event can go on if there
787 * is already an exclusive event on.
789 if (!group_can_go_on(event, cpuctx, 1))
792 err = event_sched_in(event, cpuctx, ctx, cpu);
796 * This event couldn't go on. If it is in a group
797 * then we have to pull the whole group off.
798 * If the event group is pinned then put it in error state.
801 group_sched_out(leader, cpuctx, ctx);
802 if (leader->attr.pinned) {
803 update_group_times(leader);
804 leader->state = PERF_EVENT_STATE_ERROR;
808 if (!err && !ctx->task && cpuctx->max_pertask)
809 cpuctx->max_pertask--;
814 spin_unlock(&ctx->lock);
818 * Attach a performance event to a context
820 * First we add the event to the list with the hardware enable bit
821 * in event->hw_config cleared.
823 * If the event is attached to a task which is on a CPU we use a smp
824 * call to enable it in the task context. The task might have been
825 * scheduled away, but we check this in the smp call again.
827 * Must be called with ctx->mutex held.
830 perf_install_in_context(struct perf_event_context *ctx,
831 struct perf_event *event,
834 struct task_struct *task = ctx->task;
838 * Per cpu events are installed via an smp call and
839 * the install is always sucessful.
841 smp_call_function_single(cpu, __perf_install_in_context,
847 task_oncpu_function_call(task, __perf_install_in_context,
850 spin_lock_irq(&ctx->lock);
852 * we need to retry the smp call.
854 if (ctx->is_active && list_empty(&event->group_entry)) {
855 spin_unlock_irq(&ctx->lock);
860 * The lock prevents that this context is scheduled in so we
861 * can add the event safely, if it the call above did not
864 if (list_empty(&event->group_entry))
865 add_event_to_ctx(event, ctx);
866 spin_unlock_irq(&ctx->lock);
870 * Put a event into inactive state and update time fields.
871 * Enabling the leader of a group effectively enables all
872 * the group members that aren't explicitly disabled, so we
873 * have to update their ->tstamp_enabled also.
874 * Note: this works for group members as well as group leaders
875 * since the non-leader members' sibling_lists will be empty.
877 static void __perf_event_mark_enabled(struct perf_event *event,
878 struct perf_event_context *ctx)
880 struct perf_event *sub;
882 event->state = PERF_EVENT_STATE_INACTIVE;
883 event->tstamp_enabled = ctx->time - event->total_time_enabled;
884 list_for_each_entry(sub, &event->sibling_list, group_entry)
885 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
886 sub->tstamp_enabled =
887 ctx->time - sub->total_time_enabled;
891 * Cross CPU call to enable a performance event
893 static void __perf_event_enable(void *info)
895 struct perf_event *event = info;
896 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
897 struct perf_event_context *ctx = event->ctx;
898 struct perf_event *leader = event->group_leader;
902 * If this is a per-task event, need to check whether this
903 * event's task is the current task on this cpu.
905 if (ctx->task && cpuctx->task_ctx != ctx) {
906 if (cpuctx->task_ctx || ctx->task != current)
908 cpuctx->task_ctx = ctx;
911 spin_lock(&ctx->lock);
913 update_context_time(ctx);
915 if (event->state >= PERF_EVENT_STATE_INACTIVE)
917 __perf_event_mark_enabled(event, ctx);
920 * If the event is in a group and isn't the group leader,
921 * then don't put it on unless the group is on.
923 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
926 if (!group_can_go_on(event, cpuctx, 1)) {
931 err = group_sched_in(event, cpuctx, ctx,
934 err = event_sched_in(event, cpuctx, ctx,
941 * If this event can't go on and it's part of a
942 * group, then the whole group has to come off.
945 group_sched_out(leader, cpuctx, ctx);
946 if (leader->attr.pinned) {
947 update_group_times(leader);
948 leader->state = PERF_EVENT_STATE_ERROR;
953 spin_unlock(&ctx->lock);
959 * If event->ctx is a cloned context, callers must make sure that
960 * every task struct that event->ctx->task could possibly point to
961 * remains valid. This condition is satisfied when called through
962 * perf_event_for_each_child or perf_event_for_each as described
963 * for perf_event_disable.
965 static void perf_event_enable(struct perf_event *event)
967 struct perf_event_context *ctx = event->ctx;
968 struct task_struct *task = ctx->task;
972 * Enable the event on the cpu that it's on
974 smp_call_function_single(event->cpu, __perf_event_enable,
979 spin_lock_irq(&ctx->lock);
980 if (event->state >= PERF_EVENT_STATE_INACTIVE)
984 * If the event is in error state, clear that first.
985 * That way, if we see the event in error state below, we
986 * know that it has gone back into error state, as distinct
987 * from the task having been scheduled away before the
988 * cross-call arrived.
990 if (event->state == PERF_EVENT_STATE_ERROR)
991 event->state = PERF_EVENT_STATE_OFF;
994 spin_unlock_irq(&ctx->lock);
995 task_oncpu_function_call(task, __perf_event_enable, event);
997 spin_lock_irq(&ctx->lock);
1000 * If the context is active and the event is still off,
1001 * we need to retry the cross-call.
1003 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1007 * Since we have the lock this context can't be scheduled
1008 * in, so we can change the state safely.
1010 if (event->state == PERF_EVENT_STATE_OFF)
1011 __perf_event_mark_enabled(event, ctx);
1014 spin_unlock_irq(&ctx->lock);
1017 static int perf_event_refresh(struct perf_event *event, int refresh)
1020 * not supported on inherited events
1022 if (event->attr.inherit)
1025 atomic_add(refresh, &event->event_limit);
1026 perf_event_enable(event);
1031 void __perf_event_sched_out(struct perf_event_context *ctx,
1032 struct perf_cpu_context *cpuctx)
1034 struct perf_event *event;
1036 spin_lock(&ctx->lock);
1038 if (likely(!ctx->nr_events))
1040 update_context_time(ctx);
1043 if (ctx->nr_active) {
1044 list_for_each_entry(event, &ctx->group_list, group_entry)
1045 group_sched_out(event, cpuctx, ctx);
1049 spin_unlock(&ctx->lock);
1053 * Test whether two contexts are equivalent, i.e. whether they
1054 * have both been cloned from the same version of the same context
1055 * and they both have the same number of enabled events.
1056 * If the number of enabled events is the same, then the set
1057 * of enabled events should be the same, because these are both
1058 * inherited contexts, therefore we can't access individual events
1059 * in them directly with an fd; we can only enable/disable all
1060 * events via prctl, or enable/disable all events in a family
1061 * via ioctl, which will have the same effect on both contexts.
1063 static int context_equiv(struct perf_event_context *ctx1,
1064 struct perf_event_context *ctx2)
1066 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1067 && ctx1->parent_gen == ctx2->parent_gen
1068 && !ctx1->pin_count && !ctx2->pin_count;
1071 static void __perf_event_sync_stat(struct perf_event *event,
1072 struct perf_event *next_event)
1076 if (!event->attr.inherit_stat)
1080 * Update the event value, we cannot use perf_event_read()
1081 * because we're in the middle of a context switch and have IRQs
1082 * disabled, which upsets smp_call_function_single(), however
1083 * we know the event must be on the current CPU, therefore we
1084 * don't need to use it.
1086 switch (event->state) {
1087 case PERF_EVENT_STATE_ACTIVE:
1088 event->pmu->read(event);
1091 case PERF_EVENT_STATE_INACTIVE:
1092 update_event_times(event);
1100 * In order to keep per-task stats reliable we need to flip the event
1101 * values when we flip the contexts.
1103 value = atomic64_read(&next_event->count);
1104 value = atomic64_xchg(&event->count, value);
1105 atomic64_set(&next_event->count, value);
1107 swap(event->total_time_enabled, next_event->total_time_enabled);
1108 swap(event->total_time_running, next_event->total_time_running);
1111 * Since we swizzled the values, update the user visible data too.
1113 perf_event_update_userpage(event);
1114 perf_event_update_userpage(next_event);
1117 #define list_next_entry(pos, member) \
1118 list_entry(pos->member.next, typeof(*pos), member)
1120 static void perf_event_sync_stat(struct perf_event_context *ctx,
1121 struct perf_event_context *next_ctx)
1123 struct perf_event *event, *next_event;
1128 update_context_time(ctx);
1130 event = list_first_entry(&ctx->event_list,
1131 struct perf_event, event_entry);
1133 next_event = list_first_entry(&next_ctx->event_list,
1134 struct perf_event, event_entry);
1136 while (&event->event_entry != &ctx->event_list &&
1137 &next_event->event_entry != &next_ctx->event_list) {
1139 __perf_event_sync_stat(event, next_event);
1141 event = list_next_entry(event, event_entry);
1142 next_event = list_next_entry(next_event, event_entry);
1147 * Called from scheduler to remove the events of the current task,
1148 * with interrupts disabled.
1150 * We stop each event and update the event value in event->count.
1152 * This does not protect us against NMI, but disable()
1153 * sets the disabled bit in the control field of event _before_
1154 * accessing the event control register. If a NMI hits, then it will
1155 * not restart the event.
1157 void perf_event_task_sched_out(struct task_struct *task,
1158 struct task_struct *next, int cpu)
1160 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1161 struct perf_event_context *ctx = task->perf_event_ctxp;
1162 struct perf_event_context *next_ctx;
1163 struct perf_event_context *parent;
1164 struct pt_regs *regs;
1167 regs = task_pt_regs(task);
1168 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1170 if (likely(!ctx || !cpuctx->task_ctx))
1174 parent = rcu_dereference(ctx->parent_ctx);
1175 next_ctx = next->perf_event_ctxp;
1176 if (parent && next_ctx &&
1177 rcu_dereference(next_ctx->parent_ctx) == parent) {
1179 * Looks like the two contexts are clones, so we might be
1180 * able to optimize the context switch. We lock both
1181 * contexts and check that they are clones under the
1182 * lock (including re-checking that neither has been
1183 * uncloned in the meantime). It doesn't matter which
1184 * order we take the locks because no other cpu could
1185 * be trying to lock both of these tasks.
1187 spin_lock(&ctx->lock);
1188 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1189 if (context_equiv(ctx, next_ctx)) {
1191 * XXX do we need a memory barrier of sorts
1192 * wrt to rcu_dereference() of perf_event_ctxp
1194 task->perf_event_ctxp = next_ctx;
1195 next->perf_event_ctxp = ctx;
1197 next_ctx->task = task;
1200 perf_event_sync_stat(ctx, next_ctx);
1202 spin_unlock(&next_ctx->lock);
1203 spin_unlock(&ctx->lock);
1208 __perf_event_sched_out(ctx, cpuctx);
1209 cpuctx->task_ctx = NULL;
1214 * Called with IRQs disabled
1216 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1218 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1220 if (!cpuctx->task_ctx)
1223 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1226 __perf_event_sched_out(ctx, cpuctx);
1227 cpuctx->task_ctx = NULL;
1231 * Called with IRQs disabled
1233 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1235 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1239 __perf_event_sched_in(struct perf_event_context *ctx,
1240 struct perf_cpu_context *cpuctx, int cpu)
1242 struct perf_event *event;
1245 spin_lock(&ctx->lock);
1247 if (likely(!ctx->nr_events))
1250 ctx->timestamp = perf_clock();
1255 * First go through the list and put on any pinned groups
1256 * in order to give them the best chance of going on.
1258 list_for_each_entry(event, &ctx->group_list, group_entry) {
1259 if (event->state <= PERF_EVENT_STATE_OFF ||
1260 !event->attr.pinned)
1262 if (event->cpu != -1 && event->cpu != cpu)
1265 if (group_can_go_on(event, cpuctx, 1))
1266 group_sched_in(event, cpuctx, ctx, cpu);
1269 * If this pinned group hasn't been scheduled,
1270 * put it in error state.
1272 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1273 update_group_times(event);
1274 event->state = PERF_EVENT_STATE_ERROR;
1278 list_for_each_entry(event, &ctx->group_list, group_entry) {
1280 * Ignore events in OFF or ERROR state, and
1281 * ignore pinned events since we did them already.
1283 if (event->state <= PERF_EVENT_STATE_OFF ||
1288 * Listen to the 'cpu' scheduling filter constraint
1291 if (event->cpu != -1 && event->cpu != cpu)
1294 if (group_can_go_on(event, cpuctx, can_add_hw))
1295 if (group_sched_in(event, cpuctx, ctx, cpu))
1300 spin_unlock(&ctx->lock);
1304 * Called from scheduler to add the events of the current task
1305 * with interrupts disabled.
1307 * We restore the event value and then enable it.
1309 * This does not protect us against NMI, but enable()
1310 * sets the enabled bit in the control field of event _before_
1311 * accessing the event control register. If a NMI hits, then it will
1312 * keep the event running.
1314 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1316 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1317 struct perf_event_context *ctx = task->perf_event_ctxp;
1321 if (cpuctx->task_ctx == ctx)
1323 __perf_event_sched_in(ctx, cpuctx, cpu);
1324 cpuctx->task_ctx = ctx;
1327 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1329 struct perf_event_context *ctx = &cpuctx->ctx;
1331 __perf_event_sched_in(ctx, cpuctx, cpu);
1334 #define MAX_INTERRUPTS (~0ULL)
1336 static void perf_log_throttle(struct perf_event *event, int enable);
1338 static void perf_adjust_period(struct perf_event *event, u64 events)
1340 struct hw_perf_event *hwc = &event->hw;
1341 u64 period, sample_period;
1344 events *= hwc->sample_period;
1345 period = div64_u64(events, event->attr.sample_freq);
1347 delta = (s64)(period - hwc->sample_period);
1348 delta = (delta + 7) / 8; /* low pass filter */
1350 sample_period = hwc->sample_period + delta;
1355 hwc->sample_period = sample_period;
1358 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1360 struct perf_event *event;
1361 struct hw_perf_event *hwc;
1362 u64 interrupts, freq;
1364 spin_lock(&ctx->lock);
1365 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1366 if (event->state != PERF_EVENT_STATE_ACTIVE)
1371 interrupts = hwc->interrupts;
1372 hwc->interrupts = 0;
1375 * unthrottle events on the tick
1377 if (interrupts == MAX_INTERRUPTS) {
1378 perf_log_throttle(event, 1);
1379 event->pmu->unthrottle(event);
1380 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1383 if (!event->attr.freq || !event->attr.sample_freq)
1387 * if the specified freq < HZ then we need to skip ticks
1389 if (event->attr.sample_freq < HZ) {
1390 freq = event->attr.sample_freq;
1392 hwc->freq_count += freq;
1393 hwc->freq_interrupts += interrupts;
1395 if (hwc->freq_count < HZ)
1398 interrupts = hwc->freq_interrupts;
1399 hwc->freq_interrupts = 0;
1400 hwc->freq_count -= HZ;
1404 perf_adjust_period(event, freq * interrupts);
1407 * In order to avoid being stalled by an (accidental) huge
1408 * sample period, force reset the sample period if we didn't
1409 * get any events in this freq period.
1413 event->pmu->disable(event);
1414 atomic64_set(&hwc->period_left, 0);
1415 event->pmu->enable(event);
1419 spin_unlock(&ctx->lock);
1423 * Round-robin a context's events:
1425 static void rotate_ctx(struct perf_event_context *ctx)
1427 struct perf_event *event;
1429 if (!ctx->nr_events)
1432 spin_lock(&ctx->lock);
1434 * Rotate the first entry last (works just fine for group events too):
1437 list_for_each_entry(event, &ctx->group_list, group_entry) {
1438 list_move_tail(&event->group_entry, &ctx->group_list);
1443 spin_unlock(&ctx->lock);
1446 void perf_event_task_tick(struct task_struct *curr, int cpu)
1448 struct perf_cpu_context *cpuctx;
1449 struct perf_event_context *ctx;
1451 if (!atomic_read(&nr_events))
1454 cpuctx = &per_cpu(perf_cpu_context, cpu);
1455 ctx = curr->perf_event_ctxp;
1457 perf_ctx_adjust_freq(&cpuctx->ctx);
1459 perf_ctx_adjust_freq(ctx);
1461 perf_event_cpu_sched_out(cpuctx);
1463 __perf_event_task_sched_out(ctx);
1465 rotate_ctx(&cpuctx->ctx);
1469 perf_event_cpu_sched_in(cpuctx, cpu);
1471 perf_event_task_sched_in(curr, cpu);
1475 * Enable all of a task's events that have been marked enable-on-exec.
1476 * This expects task == current.
1478 static void perf_event_enable_on_exec(struct task_struct *task)
1480 struct perf_event_context *ctx;
1481 struct perf_event *event;
1482 unsigned long flags;
1485 local_irq_save(flags);
1486 ctx = task->perf_event_ctxp;
1487 if (!ctx || !ctx->nr_events)
1490 __perf_event_task_sched_out(ctx);
1492 spin_lock(&ctx->lock);
1494 list_for_each_entry(event, &ctx->group_list, group_entry) {
1495 if (!event->attr.enable_on_exec)
1497 event->attr.enable_on_exec = 0;
1498 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1500 __perf_event_mark_enabled(event, ctx);
1505 * Unclone this context if we enabled any event.
1510 spin_unlock(&ctx->lock);
1512 perf_event_task_sched_in(task, smp_processor_id());
1514 local_irq_restore(flags);
1518 * Cross CPU call to read the hardware event
1520 static void __perf_event_read(void *info)
1522 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1523 struct perf_event *event = info;
1524 struct perf_event_context *ctx = event->ctx;
1527 * If this is a task context, we need to check whether it is
1528 * the current task context of this cpu. If not it has been
1529 * scheduled out before the smp call arrived. In that case
1530 * event->count would have been updated to a recent sample
1531 * when the event was scheduled out.
1533 if (ctx->task && cpuctx->task_ctx != ctx)
1536 spin_lock(&ctx->lock);
1537 update_context_time(ctx);
1538 update_event_times(event);
1539 spin_unlock(&ctx->lock);
1541 event->pmu->read(event);
1544 static u64 perf_event_read(struct perf_event *event)
1547 * If event is enabled and currently active on a CPU, update the
1548 * value in the event structure:
1550 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1551 smp_call_function_single(event->oncpu,
1552 __perf_event_read, event, 1);
1553 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1554 struct perf_event_context *ctx = event->ctx;
1555 unsigned long flags;
1557 spin_lock_irqsave(&ctx->lock, flags);
1558 update_context_time(ctx);
1559 update_event_times(event);
1560 spin_unlock_irqrestore(&ctx->lock, flags);
1563 return atomic64_read(&event->count);
1567 * Initialize the perf_event context in a task_struct:
1570 __perf_event_init_context(struct perf_event_context *ctx,
1571 struct task_struct *task)
1573 memset(ctx, 0, sizeof(*ctx));
1574 spin_lock_init(&ctx->lock);
1575 mutex_init(&ctx->mutex);
1576 INIT_LIST_HEAD(&ctx->group_list);
1577 INIT_LIST_HEAD(&ctx->event_list);
1578 atomic_set(&ctx->refcount, 1);
1582 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1584 struct perf_event_context *ctx;
1585 struct perf_cpu_context *cpuctx;
1586 struct task_struct *task;
1587 unsigned long flags;
1591 * If cpu is not a wildcard then this is a percpu event:
1594 /* Must be root to operate on a CPU event: */
1595 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1596 return ERR_PTR(-EACCES);
1598 if (cpu < 0 || cpu > num_possible_cpus())
1599 return ERR_PTR(-EINVAL);
1602 * We could be clever and allow to attach a event to an
1603 * offline CPU and activate it when the CPU comes up, but
1606 if (!cpu_isset(cpu, cpu_online_map))
1607 return ERR_PTR(-ENODEV);
1609 cpuctx = &per_cpu(perf_cpu_context, cpu);
1620 task = find_task_by_vpid(pid);
1622 get_task_struct(task);
1626 return ERR_PTR(-ESRCH);
1629 * Can't attach events to a dying task.
1632 if (task->flags & PF_EXITING)
1635 /* Reuse ptrace permission checks for now. */
1637 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1641 ctx = perf_lock_task_context(task, &flags);
1644 spin_unlock_irqrestore(&ctx->lock, flags);
1648 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1652 __perf_event_init_context(ctx, task);
1654 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1656 * We raced with some other task; use
1657 * the context they set.
1662 get_task_struct(task);
1665 put_task_struct(task);
1669 put_task_struct(task);
1670 return ERR_PTR(err);
1673 static void perf_event_free_filter(struct perf_event *event);
1675 static void free_event_rcu(struct rcu_head *head)
1677 struct perf_event *event;
1679 event = container_of(head, struct perf_event, rcu_head);
1681 put_pid_ns(event->ns);
1682 perf_event_free_filter(event);
1686 static void perf_pending_sync(struct perf_event *event);
1688 static void free_event(struct perf_event *event)
1690 perf_pending_sync(event);
1692 if (!event->parent) {
1693 atomic_dec(&nr_events);
1694 if (event->attr.mmap)
1695 atomic_dec(&nr_mmap_events);
1696 if (event->attr.comm)
1697 atomic_dec(&nr_comm_events);
1698 if (event->attr.task)
1699 atomic_dec(&nr_task_events);
1702 if (event->output) {
1703 fput(event->output->filp);
1704 event->output = NULL;
1708 event->destroy(event);
1710 put_ctx(event->ctx);
1711 call_rcu(&event->rcu_head, free_event_rcu);
1714 int perf_event_release_kernel(struct perf_event *event)
1716 struct perf_event_context *ctx = event->ctx;
1718 WARN_ON_ONCE(ctx->parent_ctx);
1719 mutex_lock(&ctx->mutex);
1720 perf_event_remove_from_context(event);
1721 mutex_unlock(&ctx->mutex);
1723 mutex_lock(&event->owner->perf_event_mutex);
1724 list_del_init(&event->owner_entry);
1725 mutex_unlock(&event->owner->perf_event_mutex);
1726 put_task_struct(event->owner);
1732 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1735 * Called when the last reference to the file is gone.
1737 static int perf_release(struct inode *inode, struct file *file)
1739 struct perf_event *event = file->private_data;
1741 file->private_data = NULL;
1743 return perf_event_release_kernel(event);
1746 static int perf_event_read_size(struct perf_event *event)
1748 int entry = sizeof(u64); /* value */
1752 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1753 size += sizeof(u64);
1755 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1756 size += sizeof(u64);
1758 if (event->attr.read_format & PERF_FORMAT_ID)
1759 entry += sizeof(u64);
1761 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1762 nr += event->group_leader->nr_siblings;
1763 size += sizeof(u64);
1771 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1773 struct perf_event *child;
1779 mutex_lock(&event->child_mutex);
1780 total += perf_event_read(event);
1781 *enabled += event->total_time_enabled +
1782 atomic64_read(&event->child_total_time_enabled);
1783 *running += event->total_time_running +
1784 atomic64_read(&event->child_total_time_running);
1786 list_for_each_entry(child, &event->child_list, child_list) {
1787 total += perf_event_read(child);
1788 *enabled += child->total_time_enabled;
1789 *running += child->total_time_running;
1791 mutex_unlock(&event->child_mutex);
1795 EXPORT_SYMBOL_GPL(perf_event_read_value);
1797 static int perf_event_read_group(struct perf_event *event,
1798 u64 read_format, char __user *buf)
1800 struct perf_event *leader = event->group_leader, *sub;
1801 int n = 0, size = 0, ret = -EFAULT;
1802 struct perf_event_context *ctx = leader->ctx;
1804 u64 count, enabled, running;
1806 mutex_lock(&ctx->mutex);
1807 count = perf_event_read_value(leader, &enabled, &running);
1809 values[n++] = 1 + leader->nr_siblings;
1810 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1811 values[n++] = enabled;
1812 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1813 values[n++] = running;
1814 values[n++] = count;
1815 if (read_format & PERF_FORMAT_ID)
1816 values[n++] = primary_event_id(leader);
1818 size = n * sizeof(u64);
1820 if (copy_to_user(buf, values, size))
1825 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1828 values[n++] = perf_event_read_value(sub, &enabled, &running);
1829 if (read_format & PERF_FORMAT_ID)
1830 values[n++] = primary_event_id(sub);
1832 size = n * sizeof(u64);
1834 if (copy_to_user(buf + ret, values, size)) {
1842 mutex_unlock(&ctx->mutex);
1847 static int perf_event_read_one(struct perf_event *event,
1848 u64 read_format, char __user *buf)
1850 u64 enabled, running;
1854 values[n++] = perf_event_read_value(event, &enabled, &running);
1855 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1856 values[n++] = enabled;
1857 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1858 values[n++] = running;
1859 if (read_format & PERF_FORMAT_ID)
1860 values[n++] = primary_event_id(event);
1862 if (copy_to_user(buf, values, n * sizeof(u64)))
1865 return n * sizeof(u64);
1869 * Read the performance event - simple non blocking version for now
1872 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1874 u64 read_format = event->attr.read_format;
1878 * Return end-of-file for a read on a event that is in
1879 * error state (i.e. because it was pinned but it couldn't be
1880 * scheduled on to the CPU at some point).
1882 if (event->state == PERF_EVENT_STATE_ERROR)
1885 if (count < perf_event_read_size(event))
1888 WARN_ON_ONCE(event->ctx->parent_ctx);
1889 if (read_format & PERF_FORMAT_GROUP)
1890 ret = perf_event_read_group(event, read_format, buf);
1892 ret = perf_event_read_one(event, read_format, buf);
1898 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1900 struct perf_event *event = file->private_data;
1902 return perf_read_hw(event, buf, count);
1905 static unsigned int perf_poll(struct file *file, poll_table *wait)
1907 struct perf_event *event = file->private_data;
1908 struct perf_mmap_data *data;
1909 unsigned int events = POLL_HUP;
1912 data = rcu_dereference(event->data);
1914 events = atomic_xchg(&data->poll, 0);
1917 poll_wait(file, &event->waitq, wait);
1922 static void perf_event_reset(struct perf_event *event)
1924 (void)perf_event_read(event);
1925 atomic64_set(&event->count, 0);
1926 perf_event_update_userpage(event);
1930 * Holding the top-level event's child_mutex means that any
1931 * descendant process that has inherited this event will block
1932 * in sync_child_event if it goes to exit, thus satisfying the
1933 * task existence requirements of perf_event_enable/disable.
1935 static void perf_event_for_each_child(struct perf_event *event,
1936 void (*func)(struct perf_event *))
1938 struct perf_event *child;
1940 WARN_ON_ONCE(event->ctx->parent_ctx);
1941 mutex_lock(&event->child_mutex);
1943 list_for_each_entry(child, &event->child_list, child_list)
1945 mutex_unlock(&event->child_mutex);
1948 static void perf_event_for_each(struct perf_event *event,
1949 void (*func)(struct perf_event *))
1951 struct perf_event_context *ctx = event->ctx;
1952 struct perf_event *sibling;
1954 WARN_ON_ONCE(ctx->parent_ctx);
1955 mutex_lock(&ctx->mutex);
1956 event = event->group_leader;
1958 perf_event_for_each_child(event, func);
1960 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1961 perf_event_for_each_child(event, func);
1962 mutex_unlock(&ctx->mutex);
1965 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1967 struct perf_event_context *ctx = event->ctx;
1972 if (!event->attr.sample_period)
1975 size = copy_from_user(&value, arg, sizeof(value));
1976 if (size != sizeof(value))
1982 spin_lock_irq(&ctx->lock);
1983 if (event->attr.freq) {
1984 if (value > sysctl_perf_event_sample_rate) {
1989 event->attr.sample_freq = value;
1991 event->attr.sample_period = value;
1992 event->hw.sample_period = value;
1995 spin_unlock_irq(&ctx->lock);
2000 static int perf_event_set_output(struct perf_event *event, int output_fd);
2001 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2003 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2005 struct perf_event *event = file->private_data;
2006 void (*func)(struct perf_event *);
2010 case PERF_EVENT_IOC_ENABLE:
2011 func = perf_event_enable;
2013 case PERF_EVENT_IOC_DISABLE:
2014 func = perf_event_disable;
2016 case PERF_EVENT_IOC_RESET:
2017 func = perf_event_reset;
2020 case PERF_EVENT_IOC_REFRESH:
2021 return perf_event_refresh(event, arg);
2023 case PERF_EVENT_IOC_PERIOD:
2024 return perf_event_period(event, (u64 __user *)arg);
2026 case PERF_EVENT_IOC_SET_OUTPUT:
2027 return perf_event_set_output(event, arg);
2029 case PERF_EVENT_IOC_SET_FILTER:
2030 return perf_event_set_filter(event, (void __user *)arg);
2036 if (flags & PERF_IOC_FLAG_GROUP)
2037 perf_event_for_each(event, func);
2039 perf_event_for_each_child(event, func);
2044 int perf_event_task_enable(void)
2046 struct perf_event *event;
2048 mutex_lock(¤t->perf_event_mutex);
2049 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2050 perf_event_for_each_child(event, perf_event_enable);
2051 mutex_unlock(¤t->perf_event_mutex);
2056 int perf_event_task_disable(void)
2058 struct perf_event *event;
2060 mutex_lock(¤t->perf_event_mutex);
2061 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2062 perf_event_for_each_child(event, perf_event_disable);
2063 mutex_unlock(¤t->perf_event_mutex);
2068 #ifndef PERF_EVENT_INDEX_OFFSET
2069 # define PERF_EVENT_INDEX_OFFSET 0
2072 static int perf_event_index(struct perf_event *event)
2074 if (event->state != PERF_EVENT_STATE_ACTIVE)
2077 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2081 * Callers need to ensure there can be no nesting of this function, otherwise
2082 * the seqlock logic goes bad. We can not serialize this because the arch
2083 * code calls this from NMI context.
2085 void perf_event_update_userpage(struct perf_event *event)
2087 struct perf_event_mmap_page *userpg;
2088 struct perf_mmap_data *data;
2091 data = rcu_dereference(event->data);
2095 userpg = data->user_page;
2098 * Disable preemption so as to not let the corresponding user-space
2099 * spin too long if we get preempted.
2104 userpg->index = perf_event_index(event);
2105 userpg->offset = atomic64_read(&event->count);
2106 if (event->state == PERF_EVENT_STATE_ACTIVE)
2107 userpg->offset -= atomic64_read(&event->hw.prev_count);
2109 userpg->time_enabled = event->total_time_enabled +
2110 atomic64_read(&event->child_total_time_enabled);
2112 userpg->time_running = event->total_time_running +
2113 atomic64_read(&event->child_total_time_running);
2122 static unsigned long perf_data_size(struct perf_mmap_data *data)
2124 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2127 #ifndef CONFIG_PERF_USE_VMALLOC
2130 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2133 static struct page *
2134 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2136 if (pgoff > data->nr_pages)
2140 return virt_to_page(data->user_page);
2142 return virt_to_page(data->data_pages[pgoff - 1]);
2145 static struct perf_mmap_data *
2146 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2148 struct perf_mmap_data *data;
2152 WARN_ON(atomic_read(&event->mmap_count));
2154 size = sizeof(struct perf_mmap_data);
2155 size += nr_pages * sizeof(void *);
2157 data = kzalloc(size, GFP_KERNEL);
2161 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2162 if (!data->user_page)
2163 goto fail_user_page;
2165 for (i = 0; i < nr_pages; i++) {
2166 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2167 if (!data->data_pages[i])
2168 goto fail_data_pages;
2171 data->data_order = 0;
2172 data->nr_pages = nr_pages;
2177 for (i--; i >= 0; i--)
2178 free_page((unsigned long)data->data_pages[i]);
2180 free_page((unsigned long)data->user_page);
2189 static void perf_mmap_free_page(unsigned long addr)
2191 struct page *page = virt_to_page((void *)addr);
2193 page->mapping = NULL;
2197 static void perf_mmap_data_free(struct perf_mmap_data *data)
2201 perf_mmap_free_page((unsigned long)data->user_page);
2202 for (i = 0; i < data->nr_pages; i++)
2203 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2209 * Back perf_mmap() with vmalloc memory.
2211 * Required for architectures that have d-cache aliasing issues.
2214 static struct page *
2215 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2217 if (pgoff > (1UL << data->data_order))
2220 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2223 static void perf_mmap_unmark_page(void *addr)
2225 struct page *page = vmalloc_to_page(addr);
2227 page->mapping = NULL;
2230 static void perf_mmap_data_free_work(struct work_struct *work)
2232 struct perf_mmap_data *data;
2236 data = container_of(work, struct perf_mmap_data, work);
2237 nr = 1 << data->data_order;
2239 base = data->user_page;
2240 for (i = 0; i < nr + 1; i++)
2241 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2246 static void perf_mmap_data_free(struct perf_mmap_data *data)
2248 schedule_work(&data->work);
2251 static struct perf_mmap_data *
2252 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2254 struct perf_mmap_data *data;
2258 WARN_ON(atomic_read(&event->mmap_count));
2260 size = sizeof(struct perf_mmap_data);
2261 size += sizeof(void *);
2263 data = kzalloc(size, GFP_KERNEL);
2267 INIT_WORK(&data->work, perf_mmap_data_free_work);
2269 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2273 data->user_page = all_buf;
2274 data->data_pages[0] = all_buf + PAGE_SIZE;
2275 data->data_order = ilog2(nr_pages);
2289 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2291 struct perf_event *event = vma->vm_file->private_data;
2292 struct perf_mmap_data *data;
2293 int ret = VM_FAULT_SIGBUS;
2295 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2296 if (vmf->pgoff == 0)
2302 data = rcu_dereference(event->data);
2306 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2309 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2313 get_page(vmf->page);
2314 vmf->page->mapping = vma->vm_file->f_mapping;
2315 vmf->page->index = vmf->pgoff;
2325 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2327 long max_size = perf_data_size(data);
2329 atomic_set(&data->lock, -1);
2331 if (event->attr.watermark) {
2332 data->watermark = min_t(long, max_size,
2333 event->attr.wakeup_watermark);
2336 if (!data->watermark)
2337 data->watermark = max_size / 2;
2340 rcu_assign_pointer(event->data, data);
2343 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2345 struct perf_mmap_data *data;
2347 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2348 perf_mmap_data_free(data);
2352 static void perf_mmap_data_release(struct perf_event *event)
2354 struct perf_mmap_data *data = event->data;
2356 WARN_ON(atomic_read(&event->mmap_count));
2358 rcu_assign_pointer(event->data, NULL);
2359 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2362 static void perf_mmap_open(struct vm_area_struct *vma)
2364 struct perf_event *event = vma->vm_file->private_data;
2366 atomic_inc(&event->mmap_count);
2369 static void perf_mmap_close(struct vm_area_struct *vma)
2371 struct perf_event *event = vma->vm_file->private_data;
2373 WARN_ON_ONCE(event->ctx->parent_ctx);
2374 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2375 unsigned long size = perf_data_size(event->data);
2376 struct user_struct *user = current_user();
2378 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2379 vma->vm_mm->locked_vm -= event->data->nr_locked;
2380 perf_mmap_data_release(event);
2381 mutex_unlock(&event->mmap_mutex);
2385 static const struct vm_operations_struct perf_mmap_vmops = {
2386 .open = perf_mmap_open,
2387 .close = perf_mmap_close,
2388 .fault = perf_mmap_fault,
2389 .page_mkwrite = perf_mmap_fault,
2392 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2394 struct perf_event *event = file->private_data;
2395 unsigned long user_locked, user_lock_limit;
2396 struct user_struct *user = current_user();
2397 unsigned long locked, lock_limit;
2398 struct perf_mmap_data *data;
2399 unsigned long vma_size;
2400 unsigned long nr_pages;
2401 long user_extra, extra;
2404 if (!(vma->vm_flags & VM_SHARED))
2407 vma_size = vma->vm_end - vma->vm_start;
2408 nr_pages = (vma_size / PAGE_SIZE) - 1;
2411 * If we have data pages ensure they're a power-of-two number, so we
2412 * can do bitmasks instead of modulo.
2414 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2417 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2420 if (vma->vm_pgoff != 0)
2423 WARN_ON_ONCE(event->ctx->parent_ctx);
2424 mutex_lock(&event->mmap_mutex);
2425 if (event->output) {
2430 if (atomic_inc_not_zero(&event->mmap_count)) {
2431 if (nr_pages != event->data->nr_pages)
2436 user_extra = nr_pages + 1;
2437 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2440 * Increase the limit linearly with more CPUs:
2442 user_lock_limit *= num_online_cpus();
2444 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2447 if (user_locked > user_lock_limit)
2448 extra = user_locked - user_lock_limit;
2450 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2451 lock_limit >>= PAGE_SHIFT;
2452 locked = vma->vm_mm->locked_vm + extra;
2454 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2455 !capable(CAP_IPC_LOCK)) {
2460 WARN_ON(event->data);
2462 data = perf_mmap_data_alloc(event, nr_pages);
2468 perf_mmap_data_init(event, data);
2470 atomic_set(&event->mmap_count, 1);
2471 atomic_long_add(user_extra, &user->locked_vm);
2472 vma->vm_mm->locked_vm += extra;
2473 event->data->nr_locked = extra;
2474 if (vma->vm_flags & VM_WRITE)
2475 event->data->writable = 1;
2478 mutex_unlock(&event->mmap_mutex);
2480 vma->vm_flags |= VM_RESERVED;
2481 vma->vm_ops = &perf_mmap_vmops;
2486 static int perf_fasync(int fd, struct file *filp, int on)
2488 struct inode *inode = filp->f_path.dentry->d_inode;
2489 struct perf_event *event = filp->private_data;
2492 mutex_lock(&inode->i_mutex);
2493 retval = fasync_helper(fd, filp, on, &event->fasync);
2494 mutex_unlock(&inode->i_mutex);
2502 static const struct file_operations perf_fops = {
2503 .release = perf_release,
2506 .unlocked_ioctl = perf_ioctl,
2507 .compat_ioctl = perf_ioctl,
2509 .fasync = perf_fasync,
2515 * If there's data, ensure we set the poll() state and publish everything
2516 * to user-space before waking everybody up.
2519 void perf_event_wakeup(struct perf_event *event)
2521 wake_up_all(&event->waitq);
2523 if (event->pending_kill) {
2524 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2525 event->pending_kill = 0;
2532 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2534 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2535 * single linked list and use cmpxchg() to add entries lockless.
2538 static void perf_pending_event(struct perf_pending_entry *entry)
2540 struct perf_event *event = container_of(entry,
2541 struct perf_event, pending);
2543 if (event->pending_disable) {
2544 event->pending_disable = 0;
2545 __perf_event_disable(event);
2548 if (event->pending_wakeup) {
2549 event->pending_wakeup = 0;
2550 perf_event_wakeup(event);
2554 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2556 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2560 static void perf_pending_queue(struct perf_pending_entry *entry,
2561 void (*func)(struct perf_pending_entry *))
2563 struct perf_pending_entry **head;
2565 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2570 head = &get_cpu_var(perf_pending_head);
2573 entry->next = *head;
2574 } while (cmpxchg(head, entry->next, entry) != entry->next);
2576 set_perf_event_pending();
2578 put_cpu_var(perf_pending_head);
2581 static int __perf_pending_run(void)
2583 struct perf_pending_entry *list;
2586 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2587 while (list != PENDING_TAIL) {
2588 void (*func)(struct perf_pending_entry *);
2589 struct perf_pending_entry *entry = list;
2596 * Ensure we observe the unqueue before we issue the wakeup,
2597 * so that we won't be waiting forever.
2598 * -- see perf_not_pending().
2609 static inline int perf_not_pending(struct perf_event *event)
2612 * If we flush on whatever cpu we run, there is a chance we don't
2616 __perf_pending_run();
2620 * Ensure we see the proper queue state before going to sleep
2621 * so that we do not miss the wakeup. -- see perf_pending_handle()
2624 return event->pending.next == NULL;
2627 static void perf_pending_sync(struct perf_event *event)
2629 wait_event(event->waitq, perf_not_pending(event));
2632 void perf_event_do_pending(void)
2634 __perf_pending_run();
2638 * Callchain support -- arch specific
2641 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2649 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2650 unsigned long offset, unsigned long head)
2654 if (!data->writable)
2657 mask = perf_data_size(data) - 1;
2659 offset = (offset - tail) & mask;
2660 head = (head - tail) & mask;
2662 if ((int)(head - offset) < 0)
2668 static void perf_output_wakeup(struct perf_output_handle *handle)
2670 atomic_set(&handle->data->poll, POLL_IN);
2673 handle->event->pending_wakeup = 1;
2674 perf_pending_queue(&handle->event->pending,
2675 perf_pending_event);
2677 perf_event_wakeup(handle->event);
2681 * Curious locking construct.
2683 * We need to ensure a later event_id doesn't publish a head when a former
2684 * event_id isn't done writing. However since we need to deal with NMIs we
2685 * cannot fully serialize things.
2687 * What we do is serialize between CPUs so we only have to deal with NMI
2688 * nesting on a single CPU.
2690 * We only publish the head (and generate a wakeup) when the outer-most
2691 * event_id completes.
2693 static void perf_output_lock(struct perf_output_handle *handle)
2695 struct perf_mmap_data *data = handle->data;
2696 int cur, cpu = get_cpu();
2701 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2713 static void perf_output_unlock(struct perf_output_handle *handle)
2715 struct perf_mmap_data *data = handle->data;
2719 data->done_head = data->head;
2721 if (!handle->locked)
2726 * The xchg implies a full barrier that ensures all writes are done
2727 * before we publish the new head, matched by a rmb() in userspace when
2728 * reading this position.
2730 while ((head = atomic_long_xchg(&data->done_head, 0)))
2731 data->user_page->data_head = head;
2734 * NMI can happen here, which means we can miss a done_head update.
2737 cpu = atomic_xchg(&data->lock, -1);
2738 WARN_ON_ONCE(cpu != smp_processor_id());
2741 * Therefore we have to validate we did not indeed do so.
2743 if (unlikely(atomic_long_read(&data->done_head))) {
2745 * Since we had it locked, we can lock it again.
2747 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2753 if (atomic_xchg(&data->wakeup, 0))
2754 perf_output_wakeup(handle);
2759 void perf_output_copy(struct perf_output_handle *handle,
2760 const void *buf, unsigned int len)
2762 unsigned int pages_mask;
2763 unsigned long offset;
2767 offset = handle->offset;
2768 pages_mask = handle->data->nr_pages - 1;
2769 pages = handle->data->data_pages;
2772 unsigned long page_offset;
2773 unsigned long page_size;
2776 nr = (offset >> PAGE_SHIFT) & pages_mask;
2777 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2778 page_offset = offset & (page_size - 1);
2779 size = min_t(unsigned int, page_size - page_offset, len);
2781 memcpy(pages[nr] + page_offset, buf, size);
2788 handle->offset = offset;
2791 * Check we didn't copy past our reservation window, taking the
2792 * possible unsigned int wrap into account.
2794 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2797 int perf_output_begin(struct perf_output_handle *handle,
2798 struct perf_event *event, unsigned int size,
2799 int nmi, int sample)
2801 struct perf_event *output_event;
2802 struct perf_mmap_data *data;
2803 unsigned long tail, offset, head;
2806 struct perf_event_header header;
2813 * For inherited events we send all the output towards the parent.
2816 event = event->parent;
2818 output_event = rcu_dereference(event->output);
2820 event = output_event;
2822 data = rcu_dereference(event->data);
2826 handle->data = data;
2827 handle->event = event;
2829 handle->sample = sample;
2831 if (!data->nr_pages)
2834 have_lost = atomic_read(&data->lost);
2836 size += sizeof(lost_event);
2838 perf_output_lock(handle);
2842 * Userspace could choose to issue a mb() before updating the
2843 * tail pointer. So that all reads will be completed before the
2846 tail = ACCESS_ONCE(data->user_page->data_tail);
2848 offset = head = atomic_long_read(&data->head);
2850 if (unlikely(!perf_output_space(data, tail, offset, head)))
2852 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2854 handle->offset = offset;
2855 handle->head = head;
2857 if (head - tail > data->watermark)
2858 atomic_set(&data->wakeup, 1);
2861 lost_event.header.type = PERF_RECORD_LOST;
2862 lost_event.header.misc = 0;
2863 lost_event.header.size = sizeof(lost_event);
2864 lost_event.id = event->id;
2865 lost_event.lost = atomic_xchg(&data->lost, 0);
2867 perf_output_put(handle, lost_event);
2873 atomic_inc(&data->lost);
2874 perf_output_unlock(handle);
2881 void perf_output_end(struct perf_output_handle *handle)
2883 struct perf_event *event = handle->event;
2884 struct perf_mmap_data *data = handle->data;
2886 int wakeup_events = event->attr.wakeup_events;
2888 if (handle->sample && wakeup_events) {
2889 int events = atomic_inc_return(&data->events);
2890 if (events >= wakeup_events) {
2891 atomic_sub(wakeup_events, &data->events);
2892 atomic_set(&data->wakeup, 1);
2896 perf_output_unlock(handle);
2900 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2903 * only top level events have the pid namespace they were created in
2906 event = event->parent;
2908 return task_tgid_nr_ns(p, event->ns);
2911 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2914 * only top level events have the pid namespace they were created in
2917 event = event->parent;
2919 return task_pid_nr_ns(p, event->ns);
2922 static void perf_output_read_one(struct perf_output_handle *handle,
2923 struct perf_event *event)
2925 u64 read_format = event->attr.read_format;
2929 values[n++] = atomic64_read(&event->count);
2930 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2931 values[n++] = event->total_time_enabled +
2932 atomic64_read(&event->child_total_time_enabled);
2934 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2935 values[n++] = event->total_time_running +
2936 atomic64_read(&event->child_total_time_running);
2938 if (read_format & PERF_FORMAT_ID)
2939 values[n++] = primary_event_id(event);
2941 perf_output_copy(handle, values, n * sizeof(u64));
2945 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2947 static void perf_output_read_group(struct perf_output_handle *handle,
2948 struct perf_event *event)
2950 struct perf_event *leader = event->group_leader, *sub;
2951 u64 read_format = event->attr.read_format;
2955 values[n++] = 1 + leader->nr_siblings;
2957 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2958 values[n++] = leader->total_time_enabled;
2960 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2961 values[n++] = leader->total_time_running;
2963 if (leader != event)
2964 leader->pmu->read(leader);
2966 values[n++] = atomic64_read(&leader->count);
2967 if (read_format & PERF_FORMAT_ID)
2968 values[n++] = primary_event_id(leader);
2970 perf_output_copy(handle, values, n * sizeof(u64));
2972 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2976 sub->pmu->read(sub);
2978 values[n++] = atomic64_read(&sub->count);
2979 if (read_format & PERF_FORMAT_ID)
2980 values[n++] = primary_event_id(sub);
2982 perf_output_copy(handle, values, n * sizeof(u64));
2986 static void perf_output_read(struct perf_output_handle *handle,
2987 struct perf_event *event)
2989 if (event->attr.read_format & PERF_FORMAT_GROUP)
2990 perf_output_read_group(handle, event);
2992 perf_output_read_one(handle, event);
2995 void perf_output_sample(struct perf_output_handle *handle,
2996 struct perf_event_header *header,
2997 struct perf_sample_data *data,
2998 struct perf_event *event)
3000 u64 sample_type = data->type;
3002 perf_output_put(handle, *header);
3004 if (sample_type & PERF_SAMPLE_IP)
3005 perf_output_put(handle, data->ip);
3007 if (sample_type & PERF_SAMPLE_TID)
3008 perf_output_put(handle, data->tid_entry);
3010 if (sample_type & PERF_SAMPLE_TIME)
3011 perf_output_put(handle, data->time);
3013 if (sample_type & PERF_SAMPLE_ADDR)
3014 perf_output_put(handle, data->addr);
3016 if (sample_type & PERF_SAMPLE_ID)
3017 perf_output_put(handle, data->id);
3019 if (sample_type & PERF_SAMPLE_STREAM_ID)
3020 perf_output_put(handle, data->stream_id);
3022 if (sample_type & PERF_SAMPLE_CPU)
3023 perf_output_put(handle, data->cpu_entry);
3025 if (sample_type & PERF_SAMPLE_PERIOD)
3026 perf_output_put(handle, data->period);
3028 if (sample_type & PERF_SAMPLE_READ)
3029 perf_output_read(handle, event);
3031 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3032 if (data->callchain) {
3035 if (data->callchain)
3036 size += data->callchain->nr;
3038 size *= sizeof(u64);
3040 perf_output_copy(handle, data->callchain, size);
3043 perf_output_put(handle, nr);
3047 if (sample_type & PERF_SAMPLE_RAW) {
3049 perf_output_put(handle, data->raw->size);
3050 perf_output_copy(handle, data->raw->data,
3057 .size = sizeof(u32),
3060 perf_output_put(handle, raw);
3065 void perf_prepare_sample(struct perf_event_header *header,
3066 struct perf_sample_data *data,
3067 struct perf_event *event,
3068 struct pt_regs *regs)
3070 u64 sample_type = event->attr.sample_type;
3072 data->type = sample_type;
3074 header->type = PERF_RECORD_SAMPLE;
3075 header->size = sizeof(*header);
3078 header->misc |= perf_misc_flags(regs);
3080 if (sample_type & PERF_SAMPLE_IP) {
3081 data->ip = perf_instruction_pointer(regs);
3083 header->size += sizeof(data->ip);
3086 if (sample_type & PERF_SAMPLE_TID) {
3087 /* namespace issues */
3088 data->tid_entry.pid = perf_event_pid(event, current);
3089 data->tid_entry.tid = perf_event_tid(event, current);
3091 header->size += sizeof(data->tid_entry);
3094 if (sample_type & PERF_SAMPLE_TIME) {
3095 data->time = perf_clock();
3097 header->size += sizeof(data->time);
3100 if (sample_type & PERF_SAMPLE_ADDR)
3101 header->size += sizeof(data->addr);
3103 if (sample_type & PERF_SAMPLE_ID) {
3104 data->id = primary_event_id(event);
3106 header->size += sizeof(data->id);
3109 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3110 data->stream_id = event->id;
3112 header->size += sizeof(data->stream_id);
3115 if (sample_type & PERF_SAMPLE_CPU) {
3116 data->cpu_entry.cpu = raw_smp_processor_id();
3117 data->cpu_entry.reserved = 0;
3119 header->size += sizeof(data->cpu_entry);
3122 if (sample_type & PERF_SAMPLE_PERIOD)
3123 header->size += sizeof(data->period);
3125 if (sample_type & PERF_SAMPLE_READ)
3126 header->size += perf_event_read_size(event);
3128 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3131 data->callchain = perf_callchain(regs);
3133 if (data->callchain)
3134 size += data->callchain->nr;
3136 header->size += size * sizeof(u64);
3139 if (sample_type & PERF_SAMPLE_RAW) {
3140 int size = sizeof(u32);
3143 size += data->raw->size;
3145 size += sizeof(u32);
3147 WARN_ON_ONCE(size & (sizeof(u64)-1));
3148 header->size += size;
3152 static void perf_event_output(struct perf_event *event, int nmi,
3153 struct perf_sample_data *data,
3154 struct pt_regs *regs)
3156 struct perf_output_handle handle;
3157 struct perf_event_header header;
3159 perf_prepare_sample(&header, data, event, regs);
3161 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3164 perf_output_sample(&handle, &header, data, event);
3166 perf_output_end(&handle);
3173 struct perf_read_event {
3174 struct perf_event_header header;
3181 perf_event_read_event(struct perf_event *event,
3182 struct task_struct *task)
3184 struct perf_output_handle handle;
3185 struct perf_read_event read_event = {
3187 .type = PERF_RECORD_READ,
3189 .size = sizeof(read_event) + perf_event_read_size(event),
3191 .pid = perf_event_pid(event, task),
3192 .tid = perf_event_tid(event, task),
3196 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3200 perf_output_put(&handle, read_event);
3201 perf_output_read(&handle, event);
3203 perf_output_end(&handle);
3207 * task tracking -- fork/exit
3209 * enabled by: attr.comm | attr.mmap | attr.task
3212 struct perf_task_event {
3213 struct task_struct *task;
3214 struct perf_event_context *task_ctx;
3217 struct perf_event_header header;
3227 static void perf_event_task_output(struct perf_event *event,
3228 struct perf_task_event *task_event)
3230 struct perf_output_handle handle;
3232 struct task_struct *task = task_event->task;
3235 size = task_event->event_id.header.size;
3236 ret = perf_output_begin(&handle, event, size, 0, 0);
3241 task_event->event_id.pid = perf_event_pid(event, task);
3242 task_event->event_id.ppid = perf_event_pid(event, current);
3244 task_event->event_id.tid = perf_event_tid(event, task);
3245 task_event->event_id.ptid = perf_event_tid(event, current);
3247 task_event->event_id.time = perf_clock();
3249 perf_output_put(&handle, task_event->event_id);
3251 perf_output_end(&handle);
3254 static int perf_event_task_match(struct perf_event *event)
3256 if (event->attr.comm || event->attr.mmap || event->attr.task)
3262 static void perf_event_task_ctx(struct perf_event_context *ctx,
3263 struct perf_task_event *task_event)
3265 struct perf_event *event;
3267 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3268 if (perf_event_task_match(event))
3269 perf_event_task_output(event, task_event);
3273 static void perf_event_task_event(struct perf_task_event *task_event)
3275 struct perf_cpu_context *cpuctx;
3276 struct perf_event_context *ctx = task_event->task_ctx;
3279 cpuctx = &get_cpu_var(perf_cpu_context);
3280 perf_event_task_ctx(&cpuctx->ctx, task_event);
3281 put_cpu_var(perf_cpu_context);
3284 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3286 perf_event_task_ctx(ctx, task_event);
3290 static void perf_event_task(struct task_struct *task,
3291 struct perf_event_context *task_ctx,
3294 struct perf_task_event task_event;
3296 if (!atomic_read(&nr_comm_events) &&
3297 !atomic_read(&nr_mmap_events) &&
3298 !atomic_read(&nr_task_events))
3301 task_event = (struct perf_task_event){
3303 .task_ctx = task_ctx,
3306 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3308 .size = sizeof(task_event.event_id),
3317 perf_event_task_event(&task_event);
3320 void perf_event_fork(struct task_struct *task)
3322 perf_event_task(task, NULL, 1);
3329 struct perf_comm_event {
3330 struct task_struct *task;
3335 struct perf_event_header header;
3342 static void perf_event_comm_output(struct perf_event *event,
3343 struct perf_comm_event *comm_event)
3345 struct perf_output_handle handle;
3346 int size = comm_event->event_id.header.size;
3347 int ret = perf_output_begin(&handle, event, size, 0, 0);
3352 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3353 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3355 perf_output_put(&handle, comm_event->event_id);
3356 perf_output_copy(&handle, comm_event->comm,
3357 comm_event->comm_size);
3358 perf_output_end(&handle);
3361 static int perf_event_comm_match(struct perf_event *event)
3363 if (event->attr.comm)
3369 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3370 struct perf_comm_event *comm_event)
3372 struct perf_event *event;
3374 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3375 if (perf_event_comm_match(event))
3376 perf_event_comm_output(event, comm_event);
3380 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3382 struct perf_cpu_context *cpuctx;
3383 struct perf_event_context *ctx;
3385 char comm[TASK_COMM_LEN];
3387 memset(comm, 0, sizeof(comm));
3388 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3389 size = ALIGN(strlen(comm)+1, sizeof(u64));
3391 comm_event->comm = comm;
3392 comm_event->comm_size = size;
3394 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3397 cpuctx = &get_cpu_var(perf_cpu_context);
3398 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3399 put_cpu_var(perf_cpu_context);
3402 * doesn't really matter which of the child contexts the
3403 * events ends up in.
3405 ctx = rcu_dereference(current->perf_event_ctxp);
3407 perf_event_comm_ctx(ctx, comm_event);
3411 void perf_event_comm(struct task_struct *task)
3413 struct perf_comm_event comm_event;
3415 if (task->perf_event_ctxp)
3416 perf_event_enable_on_exec(task);
3418 if (!atomic_read(&nr_comm_events))
3421 comm_event = (struct perf_comm_event){
3427 .type = PERF_RECORD_COMM,
3436 perf_event_comm_event(&comm_event);
3443 struct perf_mmap_event {
3444 struct vm_area_struct *vma;
3446 const char *file_name;
3450 struct perf_event_header header;
3460 static void perf_event_mmap_output(struct perf_event *event,
3461 struct perf_mmap_event *mmap_event)
3463 struct perf_output_handle handle;
3464 int size = mmap_event->event_id.header.size;
3465 int ret = perf_output_begin(&handle, event, size, 0, 0);
3470 mmap_event->event_id.pid = perf_event_pid(event, current);
3471 mmap_event->event_id.tid = perf_event_tid(event, current);
3473 perf_output_put(&handle, mmap_event->event_id);
3474 perf_output_copy(&handle, mmap_event->file_name,
3475 mmap_event->file_size);
3476 perf_output_end(&handle);
3479 static int perf_event_mmap_match(struct perf_event *event,
3480 struct perf_mmap_event *mmap_event)
3482 if (event->attr.mmap)
3488 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3489 struct perf_mmap_event *mmap_event)
3491 struct perf_event *event;
3493 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3494 if (perf_event_mmap_match(event, mmap_event))
3495 perf_event_mmap_output(event, mmap_event);
3499 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3501 struct perf_cpu_context *cpuctx;
3502 struct perf_event_context *ctx;
3503 struct vm_area_struct *vma = mmap_event->vma;
3504 struct file *file = vma->vm_file;
3510 memset(tmp, 0, sizeof(tmp));
3514 * d_path works from the end of the buffer backwards, so we
3515 * need to add enough zero bytes after the string to handle
3516 * the 64bit alignment we do later.
3518 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3520 name = strncpy(tmp, "//enomem", sizeof(tmp));
3523 name = d_path(&file->f_path, buf, PATH_MAX);
3525 name = strncpy(tmp, "//toolong", sizeof(tmp));
3529 if (arch_vma_name(mmap_event->vma)) {
3530 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3536 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3540 name = strncpy(tmp, "//anon", sizeof(tmp));
3545 size = ALIGN(strlen(name)+1, sizeof(u64));
3547 mmap_event->file_name = name;
3548 mmap_event->file_size = size;
3550 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3553 cpuctx = &get_cpu_var(perf_cpu_context);
3554 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3555 put_cpu_var(perf_cpu_context);
3558 * doesn't really matter which of the child contexts the
3559 * events ends up in.
3561 ctx = rcu_dereference(current->perf_event_ctxp);
3563 perf_event_mmap_ctx(ctx, mmap_event);
3569 void __perf_event_mmap(struct vm_area_struct *vma)
3571 struct perf_mmap_event mmap_event;
3573 if (!atomic_read(&nr_mmap_events))
3576 mmap_event = (struct perf_mmap_event){
3582 .type = PERF_RECORD_MMAP,
3588 .start = vma->vm_start,
3589 .len = vma->vm_end - vma->vm_start,
3590 .pgoff = vma->vm_pgoff,
3594 perf_event_mmap_event(&mmap_event);
3598 * IRQ throttle logging
3601 static void perf_log_throttle(struct perf_event *event, int enable)
3603 struct perf_output_handle handle;
3607 struct perf_event_header header;
3611 } throttle_event = {
3613 .type = PERF_RECORD_THROTTLE,
3615 .size = sizeof(throttle_event),
3617 .time = perf_clock(),
3618 .id = primary_event_id(event),
3619 .stream_id = event->id,
3623 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3625 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3629 perf_output_put(&handle, throttle_event);
3630 perf_output_end(&handle);
3634 * Generic event overflow handling, sampling.
3637 static int __perf_event_overflow(struct perf_event *event, int nmi,
3638 int throttle, struct perf_sample_data *data,
3639 struct pt_regs *regs)
3641 int events = atomic_read(&event->event_limit);
3642 struct hw_perf_event *hwc = &event->hw;
3645 throttle = (throttle && event->pmu->unthrottle != NULL);
3650 if (hwc->interrupts != MAX_INTERRUPTS) {
3652 if (HZ * hwc->interrupts >
3653 (u64)sysctl_perf_event_sample_rate) {
3654 hwc->interrupts = MAX_INTERRUPTS;
3655 perf_log_throttle(event, 0);
3660 * Keep re-disabling events even though on the previous
3661 * pass we disabled it - just in case we raced with a
3662 * sched-in and the event got enabled again:
3668 if (event->attr.freq) {
3669 u64 now = perf_clock();
3670 s64 delta = now - hwc->freq_stamp;
3672 hwc->freq_stamp = now;
3674 if (delta > 0 && delta < TICK_NSEC)
3675 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3679 * XXX event_limit might not quite work as expected on inherited
3683 event->pending_kill = POLL_IN;
3684 if (events && atomic_dec_and_test(&event->event_limit)) {
3686 event->pending_kill = POLL_HUP;
3688 event->pending_disable = 1;
3689 perf_pending_queue(&event->pending,
3690 perf_pending_event);
3692 perf_event_disable(event);
3695 if (event->overflow_handler)
3696 event->overflow_handler(event, nmi, data, regs);
3698 perf_event_output(event, nmi, data, regs);
3703 int perf_event_overflow(struct perf_event *event, int nmi,
3704 struct perf_sample_data *data,
3705 struct pt_regs *regs)
3707 return __perf_event_overflow(event, nmi, 1, data, regs);
3711 * Generic software event infrastructure
3715 * We directly increment event->count and keep a second value in
3716 * event->hw.period_left to count intervals. This period event
3717 * is kept in the range [-sample_period, 0] so that we can use the
3721 static u64 perf_swevent_set_period(struct perf_event *event)
3723 struct hw_perf_event *hwc = &event->hw;
3724 u64 period = hwc->last_period;
3728 hwc->last_period = hwc->sample_period;
3731 old = val = atomic64_read(&hwc->period_left);
3735 nr = div64_u64(period + val, period);
3736 offset = nr * period;
3738 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3744 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3745 int nmi, struct perf_sample_data *data,
3746 struct pt_regs *regs)
3748 struct hw_perf_event *hwc = &event->hw;
3751 data->period = event->hw.last_period;
3753 overflow = perf_swevent_set_period(event);
3755 if (hwc->interrupts == MAX_INTERRUPTS)
3758 for (; overflow; overflow--) {
3759 if (__perf_event_overflow(event, nmi, throttle,
3762 * We inhibit the overflow from happening when
3763 * hwc->interrupts == MAX_INTERRUPTS.
3771 static void perf_swevent_unthrottle(struct perf_event *event)
3774 * Nothing to do, we already reset hwc->interrupts.
3778 static void perf_swevent_add(struct perf_event *event, u64 nr,
3779 int nmi, struct perf_sample_data *data,
3780 struct pt_regs *regs)
3782 struct hw_perf_event *hwc = &event->hw;
3784 atomic64_add(nr, &event->count);
3789 if (!hwc->sample_period)
3792 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3793 return perf_swevent_overflow(event, 1, nmi, data, regs);
3795 if (atomic64_add_negative(nr, &hwc->period_left))
3798 perf_swevent_overflow(event, 0, nmi, data, regs);
3801 static int perf_swevent_is_counting(struct perf_event *event)
3804 * The event is active, we're good!
3806 if (event->state == PERF_EVENT_STATE_ACTIVE)
3810 * The event is off/error, not counting.
3812 if (event->state != PERF_EVENT_STATE_INACTIVE)
3816 * The event is inactive, if the context is active
3817 * we're part of a group that didn't make it on the 'pmu',
3820 if (event->ctx->is_active)
3824 * We're inactive and the context is too, this means the
3825 * task is scheduled out, we're counting events that happen
3826 * to us, like migration events.
3831 static int perf_tp_event_match(struct perf_event *event,
3832 struct perf_sample_data *data);
3834 static int perf_exclude_event(struct perf_event *event,
3835 struct pt_regs *regs)
3838 if (event->attr.exclude_user && user_mode(regs))
3841 if (event->attr.exclude_kernel && !user_mode(regs))
3848 static int perf_swevent_match(struct perf_event *event,
3849 enum perf_type_id type,
3851 struct perf_sample_data *data,
3852 struct pt_regs *regs)
3854 if (!perf_swevent_is_counting(event))
3857 if (event->attr.type != type)
3860 if (event->attr.config != event_id)
3863 if (perf_exclude_event(event, regs))
3866 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3867 !perf_tp_event_match(event, data))
3873 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3874 enum perf_type_id type,
3875 u32 event_id, u64 nr, int nmi,
3876 struct perf_sample_data *data,
3877 struct pt_regs *regs)
3879 struct perf_event *event;
3881 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3882 if (perf_swevent_match(event, type, event_id, data, regs))
3883 perf_swevent_add(event, nr, nmi, data, regs);
3887 int perf_swevent_get_recursion_context(void)
3889 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3896 else if (in_softirq())
3901 if (cpuctx->recursion[rctx]) {
3902 put_cpu_var(perf_cpu_context);
3906 cpuctx->recursion[rctx]++;
3911 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3913 void perf_swevent_put_recursion_context(int rctx)
3915 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3917 cpuctx->recursion[rctx]--;
3918 put_cpu_var(perf_cpu_context);
3920 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3922 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3924 struct perf_sample_data *data,
3925 struct pt_regs *regs)
3927 struct perf_cpu_context *cpuctx;
3928 struct perf_event_context *ctx;
3930 cpuctx = &__get_cpu_var(perf_cpu_context);
3932 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3933 nr, nmi, data, regs);
3935 * doesn't really matter which of the child contexts the
3936 * events ends up in.
3938 ctx = rcu_dereference(current->perf_event_ctxp);
3940 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3944 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3945 struct pt_regs *regs, u64 addr)
3947 struct perf_sample_data data;
3950 rctx = perf_swevent_get_recursion_context();
3957 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3959 perf_swevent_put_recursion_context(rctx);
3962 static void perf_swevent_read(struct perf_event *event)
3966 static int perf_swevent_enable(struct perf_event *event)
3968 struct hw_perf_event *hwc = &event->hw;
3970 if (hwc->sample_period) {
3971 hwc->last_period = hwc->sample_period;
3972 perf_swevent_set_period(event);
3977 static void perf_swevent_disable(struct perf_event *event)
3981 static const struct pmu perf_ops_generic = {
3982 .enable = perf_swevent_enable,
3983 .disable = perf_swevent_disable,
3984 .read = perf_swevent_read,
3985 .unthrottle = perf_swevent_unthrottle,
3989 * hrtimer based swevent callback
3992 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3994 enum hrtimer_restart ret = HRTIMER_RESTART;
3995 struct perf_sample_data data;
3996 struct pt_regs *regs;
3997 struct perf_event *event;
4000 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4001 event->pmu->read(event);
4004 regs = get_irq_regs();
4006 * In case we exclude kernel IPs or are somehow not in interrupt
4007 * context, provide the next best thing, the user IP.
4009 if ((event->attr.exclude_kernel || !regs) &&
4010 !event->attr.exclude_user)
4011 regs = task_pt_regs(current);
4014 if (!(event->attr.exclude_idle && current->pid == 0))
4015 if (perf_event_overflow(event, 0, &data, regs))
4016 ret = HRTIMER_NORESTART;
4019 period = max_t(u64, 10000, event->hw.sample_period);
4020 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4025 static void perf_swevent_start_hrtimer(struct perf_event *event)
4027 struct hw_perf_event *hwc = &event->hw;
4029 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4030 hwc->hrtimer.function = perf_swevent_hrtimer;
4031 if (hwc->sample_period) {
4034 if (hwc->remaining) {
4035 if (hwc->remaining < 0)
4038 period = hwc->remaining;
4041 period = max_t(u64, 10000, hwc->sample_period);
4043 __hrtimer_start_range_ns(&hwc->hrtimer,
4044 ns_to_ktime(period), 0,
4045 HRTIMER_MODE_REL, 0);
4049 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4051 struct hw_perf_event *hwc = &event->hw;
4053 if (hwc->sample_period) {
4054 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4055 hwc->remaining = ktime_to_ns(remaining);
4057 hrtimer_cancel(&hwc->hrtimer);
4062 * Software event: cpu wall time clock
4065 static void cpu_clock_perf_event_update(struct perf_event *event)
4067 int cpu = raw_smp_processor_id();
4071 now = cpu_clock(cpu);
4072 prev = atomic64_read(&event->hw.prev_count);
4073 atomic64_set(&event->hw.prev_count, now);
4074 atomic64_add(now - prev, &event->count);
4077 static int cpu_clock_perf_event_enable(struct perf_event *event)
4079 struct hw_perf_event *hwc = &event->hw;
4080 int cpu = raw_smp_processor_id();
4082 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4083 perf_swevent_start_hrtimer(event);
4088 static void cpu_clock_perf_event_disable(struct perf_event *event)
4090 perf_swevent_cancel_hrtimer(event);
4091 cpu_clock_perf_event_update(event);
4094 static void cpu_clock_perf_event_read(struct perf_event *event)
4096 cpu_clock_perf_event_update(event);
4099 static const struct pmu perf_ops_cpu_clock = {
4100 .enable = cpu_clock_perf_event_enable,
4101 .disable = cpu_clock_perf_event_disable,
4102 .read = cpu_clock_perf_event_read,
4106 * Software event: task time clock
4109 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4114 prev = atomic64_xchg(&event->hw.prev_count, now);
4116 atomic64_add(delta, &event->count);
4119 static int task_clock_perf_event_enable(struct perf_event *event)
4121 struct hw_perf_event *hwc = &event->hw;
4124 now = event->ctx->time;
4126 atomic64_set(&hwc->prev_count, now);
4128 perf_swevent_start_hrtimer(event);
4133 static void task_clock_perf_event_disable(struct perf_event *event)
4135 perf_swevent_cancel_hrtimer(event);
4136 task_clock_perf_event_update(event, event->ctx->time);
4140 static void task_clock_perf_event_read(struct perf_event *event)
4145 update_context_time(event->ctx);
4146 time = event->ctx->time;
4148 u64 now = perf_clock();
4149 u64 delta = now - event->ctx->timestamp;
4150 time = event->ctx->time + delta;
4153 task_clock_perf_event_update(event, time);
4156 static const struct pmu perf_ops_task_clock = {
4157 .enable = task_clock_perf_event_enable,
4158 .disable = task_clock_perf_event_disable,
4159 .read = task_clock_perf_event_read,
4162 #ifdef CONFIG_EVENT_PROFILE
4164 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4167 struct perf_raw_record raw = {
4172 struct perf_sample_data data = {
4177 struct pt_regs *regs = get_irq_regs();
4180 regs = task_pt_regs(current);
4182 /* Trace events already protected against recursion */
4183 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4186 EXPORT_SYMBOL_GPL(perf_tp_event);
4188 static int perf_tp_event_match(struct perf_event *event,
4189 struct perf_sample_data *data)
4191 void *record = data->raw->data;
4193 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4198 static void tp_perf_event_destroy(struct perf_event *event)
4200 ftrace_profile_disable(event->attr.config);
4203 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4206 * Raw tracepoint data is a severe data leak, only allow root to
4209 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4210 perf_paranoid_tracepoint_raw() &&
4211 !capable(CAP_SYS_ADMIN))
4212 return ERR_PTR(-EPERM);
4214 if (ftrace_profile_enable(event->attr.config))
4217 event->destroy = tp_perf_event_destroy;
4219 return &perf_ops_generic;
4222 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4227 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4230 filter_str = strndup_user(arg, PAGE_SIZE);
4231 if (IS_ERR(filter_str))
4232 return PTR_ERR(filter_str);
4234 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4240 static void perf_event_free_filter(struct perf_event *event)
4242 ftrace_profile_free_filter(event);
4247 static int perf_tp_event_match(struct perf_event *event,
4248 struct perf_sample_data *data)
4253 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4258 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4263 static void perf_event_free_filter(struct perf_event *event)
4267 #endif /* CONFIG_EVENT_PROFILE */
4269 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4270 static void bp_perf_event_destroy(struct perf_event *event)
4272 release_bp_slot(event);
4275 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4279 * The breakpoint is already filled if we haven't created the counter
4280 * through perf syscall
4281 * FIXME: manage to get trigerred to NULL if it comes from syscalls
4284 err = register_perf_hw_breakpoint(bp);
4286 err = __register_perf_hw_breakpoint(bp);
4288 return ERR_PTR(err);
4290 bp->destroy = bp_perf_event_destroy;
4292 return &perf_ops_bp;
4295 void perf_bp_event(struct perf_event *bp, void *data)
4297 struct perf_sample_data sample;
4298 struct pt_regs *regs = data;
4300 sample.addr = bp->attr.bp_addr;
4302 if (!perf_exclude_event(bp, regs))
4303 perf_swevent_add(bp, 1, 1, &sample, regs);
4306 static void bp_perf_event_destroy(struct perf_event *event)
4310 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4315 void perf_bp_event(struct perf_event *bp, void *regs)
4320 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4322 static void sw_perf_event_destroy(struct perf_event *event)
4324 u64 event_id = event->attr.config;
4326 WARN_ON(event->parent);
4328 atomic_dec(&perf_swevent_enabled[event_id]);
4331 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4333 const struct pmu *pmu = NULL;
4334 u64 event_id = event->attr.config;
4337 * Software events (currently) can't in general distinguish
4338 * between user, kernel and hypervisor events.
4339 * However, context switches and cpu migrations are considered
4340 * to be kernel events, and page faults are never hypervisor
4344 case PERF_COUNT_SW_CPU_CLOCK:
4345 pmu = &perf_ops_cpu_clock;
4348 case PERF_COUNT_SW_TASK_CLOCK:
4350 * If the user instantiates this as a per-cpu event,
4351 * use the cpu_clock event instead.
4353 if (event->ctx->task)
4354 pmu = &perf_ops_task_clock;
4356 pmu = &perf_ops_cpu_clock;
4359 case PERF_COUNT_SW_PAGE_FAULTS:
4360 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4361 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4362 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4363 case PERF_COUNT_SW_CPU_MIGRATIONS:
4364 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4365 case PERF_COUNT_SW_EMULATION_FAULTS:
4366 if (!event->parent) {
4367 atomic_inc(&perf_swevent_enabled[event_id]);
4368 event->destroy = sw_perf_event_destroy;
4370 pmu = &perf_ops_generic;
4378 * Allocate and initialize a event structure
4380 static struct perf_event *
4381 perf_event_alloc(struct perf_event_attr *attr,
4383 struct perf_event_context *ctx,
4384 struct perf_event *group_leader,
4385 struct perf_event *parent_event,
4386 perf_callback_t callback,
4389 const struct pmu *pmu;
4390 struct perf_event *event;
4391 struct hw_perf_event *hwc;
4394 event = kzalloc(sizeof(*event), gfpflags);
4396 return ERR_PTR(-ENOMEM);
4399 * Single events are their own group leaders, with an
4400 * empty sibling list:
4403 group_leader = event;
4405 mutex_init(&event->child_mutex);
4406 INIT_LIST_HEAD(&event->child_list);
4408 INIT_LIST_HEAD(&event->group_entry);
4409 INIT_LIST_HEAD(&event->event_entry);
4410 INIT_LIST_HEAD(&event->sibling_list);
4411 init_waitqueue_head(&event->waitq);
4413 mutex_init(&event->mmap_mutex);
4416 event->attr = *attr;
4417 event->group_leader = group_leader;
4422 event->parent = parent_event;
4424 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4425 event->id = atomic64_inc_return(&perf_event_id);
4427 event->state = PERF_EVENT_STATE_INACTIVE;
4429 if (!callback && parent_event)
4430 callback = parent_event->callback;
4432 event->callback = callback;
4435 event->state = PERF_EVENT_STATE_OFF;
4440 hwc->sample_period = attr->sample_period;
4441 if (attr->freq && attr->sample_freq)
4442 hwc->sample_period = 1;
4443 hwc->last_period = hwc->sample_period;
4445 atomic64_set(&hwc->period_left, hwc->sample_period);
4448 * we currently do not support PERF_FORMAT_GROUP on inherited events
4450 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4453 switch (attr->type) {
4455 case PERF_TYPE_HARDWARE:
4456 case PERF_TYPE_HW_CACHE:
4457 pmu = hw_perf_event_init(event);
4460 case PERF_TYPE_SOFTWARE:
4461 pmu = sw_perf_event_init(event);
4464 case PERF_TYPE_TRACEPOINT:
4465 pmu = tp_perf_event_init(event);
4468 case PERF_TYPE_BREAKPOINT:
4469 pmu = bp_perf_event_init(event);
4480 else if (IS_ERR(pmu))
4485 put_pid_ns(event->ns);
4487 return ERR_PTR(err);
4492 if (!event->parent) {
4493 atomic_inc(&nr_events);
4494 if (event->attr.mmap)
4495 atomic_inc(&nr_mmap_events);
4496 if (event->attr.comm)
4497 atomic_inc(&nr_comm_events);
4498 if (event->attr.task)
4499 atomic_inc(&nr_task_events);
4505 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4506 struct perf_event_attr *attr)
4511 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4515 * zero the full structure, so that a short copy will be nice.
4517 memset(attr, 0, sizeof(*attr));
4519 ret = get_user(size, &uattr->size);
4523 if (size > PAGE_SIZE) /* silly large */
4526 if (!size) /* abi compat */
4527 size = PERF_ATTR_SIZE_VER0;
4529 if (size < PERF_ATTR_SIZE_VER0)
4533 * If we're handed a bigger struct than we know of,
4534 * ensure all the unknown bits are 0 - i.e. new
4535 * user-space does not rely on any kernel feature
4536 * extensions we dont know about yet.
4538 if (size > sizeof(*attr)) {
4539 unsigned char __user *addr;
4540 unsigned char __user *end;
4543 addr = (void __user *)uattr + sizeof(*attr);
4544 end = (void __user *)uattr + size;
4546 for (; addr < end; addr++) {
4547 ret = get_user(val, addr);
4553 size = sizeof(*attr);
4556 ret = copy_from_user(attr, uattr, size);
4561 * If the type exists, the corresponding creation will verify
4564 if (attr->type >= PERF_TYPE_MAX)
4567 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4570 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4573 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4580 put_user(sizeof(*attr), &uattr->size);
4585 static int perf_event_set_output(struct perf_event *event, int output_fd)
4587 struct perf_event *output_event = NULL;
4588 struct file *output_file = NULL;
4589 struct perf_event *old_output;
4590 int fput_needed = 0;
4596 output_file = fget_light(output_fd, &fput_needed);
4600 if (output_file->f_op != &perf_fops)
4603 output_event = output_file->private_data;
4605 /* Don't chain output fds */
4606 if (output_event->output)
4609 /* Don't set an output fd when we already have an output channel */
4613 atomic_long_inc(&output_file->f_count);
4616 mutex_lock(&event->mmap_mutex);
4617 old_output = event->output;
4618 rcu_assign_pointer(event->output, output_event);
4619 mutex_unlock(&event->mmap_mutex);
4623 * we need to make sure no existing perf_output_*()
4624 * is still referencing this event.
4627 fput(old_output->filp);
4632 fput_light(output_file, fput_needed);
4637 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4639 * @attr_uptr: event_id type attributes for monitoring/sampling
4642 * @group_fd: group leader event fd
4644 SYSCALL_DEFINE5(perf_event_open,
4645 struct perf_event_attr __user *, attr_uptr,
4646 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4648 struct perf_event *event, *group_leader;
4649 struct perf_event_attr attr;
4650 struct perf_event_context *ctx;
4651 struct file *event_file = NULL;
4652 struct file *group_file = NULL;
4653 int fput_needed = 0;
4654 int fput_needed2 = 0;
4657 /* for future expandability... */
4658 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4661 err = perf_copy_attr(attr_uptr, &attr);
4665 if (!attr.exclude_kernel) {
4666 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4671 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4676 * Get the target context (task or percpu):
4678 ctx = find_get_context(pid, cpu);
4680 return PTR_ERR(ctx);
4683 * Look up the group leader (we will attach this event to it):
4685 group_leader = NULL;
4686 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4688 group_file = fget_light(group_fd, &fput_needed);
4690 goto err_put_context;
4691 if (group_file->f_op != &perf_fops)
4692 goto err_put_context;
4694 group_leader = group_file->private_data;
4696 * Do not allow a recursive hierarchy (this new sibling
4697 * becoming part of another group-sibling):
4699 if (group_leader->group_leader != group_leader)
4700 goto err_put_context;
4702 * Do not allow to attach to a group in a different
4703 * task or CPU context:
4705 if (group_leader->ctx != ctx)
4706 goto err_put_context;
4708 * Only a group leader can be exclusive or pinned
4710 if (attr.exclusive || attr.pinned)
4711 goto err_put_context;
4714 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4715 NULL, NULL, GFP_KERNEL);
4716 err = PTR_ERR(event);
4718 goto err_put_context;
4720 err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
4722 goto err_free_put_context;
4724 event_file = fget_light(err, &fput_needed2);
4726 goto err_free_put_context;
4728 if (flags & PERF_FLAG_FD_OUTPUT) {
4729 err = perf_event_set_output(event, group_fd);
4731 goto err_fput_free_put_context;
4734 event->filp = event_file;
4735 WARN_ON_ONCE(ctx->parent_ctx);
4736 mutex_lock(&ctx->mutex);
4737 perf_install_in_context(ctx, event, cpu);
4739 mutex_unlock(&ctx->mutex);
4741 event->owner = current;
4742 get_task_struct(current);
4743 mutex_lock(¤t->perf_event_mutex);
4744 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4745 mutex_unlock(¤t->perf_event_mutex);
4747 err_fput_free_put_context:
4748 fput_light(event_file, fput_needed2);
4750 err_free_put_context:
4758 fput_light(group_file, fput_needed);
4764 * perf_event_create_kernel_counter
4766 * @attr: attributes of the counter to create
4767 * @cpu: cpu in which the counter is bound
4768 * @pid: task to profile
4771 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4772 pid_t pid, perf_callback_t callback)
4774 struct perf_event *event;
4775 struct perf_event_context *ctx;
4779 * Get the target context (task or percpu):
4782 ctx = find_get_context(pid, cpu);
4786 event = perf_event_alloc(attr, cpu, ctx, NULL,
4787 NULL, callback, GFP_KERNEL);
4788 err = PTR_ERR(event);
4790 goto err_put_context;
4793 WARN_ON_ONCE(ctx->parent_ctx);
4794 mutex_lock(&ctx->mutex);
4795 perf_install_in_context(ctx, event, cpu);
4797 mutex_unlock(&ctx->mutex);
4799 event->owner = current;
4800 get_task_struct(current);
4801 mutex_lock(¤t->perf_event_mutex);
4802 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4803 mutex_unlock(¤t->perf_event_mutex);
4813 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4816 * inherit a event from parent task to child task:
4818 static struct perf_event *
4819 inherit_event(struct perf_event *parent_event,
4820 struct task_struct *parent,
4821 struct perf_event_context *parent_ctx,
4822 struct task_struct *child,
4823 struct perf_event *group_leader,
4824 struct perf_event_context *child_ctx)
4826 struct perf_event *child_event;
4829 * Instead of creating recursive hierarchies of events,
4830 * we link inherited events back to the original parent,
4831 * which has a filp for sure, which we use as the reference
4834 if (parent_event->parent)
4835 parent_event = parent_event->parent;
4837 child_event = perf_event_alloc(&parent_event->attr,
4838 parent_event->cpu, child_ctx,
4839 group_leader, parent_event,
4841 if (IS_ERR(child_event))
4846 * Make the child state follow the state of the parent event,
4847 * not its attr.disabled bit. We hold the parent's mutex,
4848 * so we won't race with perf_event_{en, dis}able_family.
4850 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4851 child_event->state = PERF_EVENT_STATE_INACTIVE;
4853 child_event->state = PERF_EVENT_STATE_OFF;
4855 if (parent_event->attr.freq)
4856 child_event->hw.sample_period = parent_event->hw.sample_period;
4858 child_event->overflow_handler = parent_event->overflow_handler;
4861 * Link it up in the child's context:
4863 add_event_to_ctx(child_event, child_ctx);
4866 * Get a reference to the parent filp - we will fput it
4867 * when the child event exits. This is safe to do because
4868 * we are in the parent and we know that the filp still
4869 * exists and has a nonzero count:
4871 atomic_long_inc(&parent_event->filp->f_count);
4874 * Link this into the parent event's child list
4876 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4877 mutex_lock(&parent_event->child_mutex);
4878 list_add_tail(&child_event->child_list, &parent_event->child_list);
4879 mutex_unlock(&parent_event->child_mutex);
4884 static int inherit_group(struct perf_event *parent_event,
4885 struct task_struct *parent,
4886 struct perf_event_context *parent_ctx,
4887 struct task_struct *child,
4888 struct perf_event_context *child_ctx)
4890 struct perf_event *leader;
4891 struct perf_event *sub;
4892 struct perf_event *child_ctr;
4894 leader = inherit_event(parent_event, parent, parent_ctx,
4895 child, NULL, child_ctx);
4897 return PTR_ERR(leader);
4898 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4899 child_ctr = inherit_event(sub, parent, parent_ctx,
4900 child, leader, child_ctx);
4901 if (IS_ERR(child_ctr))
4902 return PTR_ERR(child_ctr);
4907 static void sync_child_event(struct perf_event *child_event,
4908 struct task_struct *child)
4910 struct perf_event *parent_event = child_event->parent;
4913 if (child_event->attr.inherit_stat)
4914 perf_event_read_event(child_event, child);
4916 child_val = atomic64_read(&child_event->count);
4919 * Add back the child's count to the parent's count:
4921 atomic64_add(child_val, &parent_event->count);
4922 atomic64_add(child_event->total_time_enabled,
4923 &parent_event->child_total_time_enabled);
4924 atomic64_add(child_event->total_time_running,
4925 &parent_event->child_total_time_running);
4928 * Remove this event from the parent's list
4930 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4931 mutex_lock(&parent_event->child_mutex);
4932 list_del_init(&child_event->child_list);
4933 mutex_unlock(&parent_event->child_mutex);
4936 * Release the parent event, if this was the last
4939 fput(parent_event->filp);
4943 __perf_event_exit_task(struct perf_event *child_event,
4944 struct perf_event_context *child_ctx,
4945 struct task_struct *child)
4947 struct perf_event *parent_event;
4949 perf_event_remove_from_context(child_event);
4951 parent_event = child_event->parent;
4953 * It can happen that parent exits first, and has events
4954 * that are still around due to the child reference. These
4955 * events need to be zapped - but otherwise linger.
4958 sync_child_event(child_event, child);
4959 free_event(child_event);
4964 * When a child task exits, feed back event values to parent events.
4966 void perf_event_exit_task(struct task_struct *child)
4968 struct perf_event *child_event, *tmp;
4969 struct perf_event_context *child_ctx;
4970 unsigned long flags;
4972 if (likely(!child->perf_event_ctxp)) {
4973 perf_event_task(child, NULL, 0);
4977 local_irq_save(flags);
4979 * We can't reschedule here because interrupts are disabled,
4980 * and either child is current or it is a task that can't be
4981 * scheduled, so we are now safe from rescheduling changing
4984 child_ctx = child->perf_event_ctxp;
4985 __perf_event_task_sched_out(child_ctx);
4988 * Take the context lock here so that if find_get_context is
4989 * reading child->perf_event_ctxp, we wait until it has
4990 * incremented the context's refcount before we do put_ctx below.
4992 spin_lock(&child_ctx->lock);
4993 child->perf_event_ctxp = NULL;
4995 * If this context is a clone; unclone it so it can't get
4996 * swapped to another process while we're removing all
4997 * the events from it.
4999 unclone_ctx(child_ctx);
5000 update_context_time(child_ctx);
5001 spin_unlock_irqrestore(&child_ctx->lock, flags);
5004 * Report the task dead after unscheduling the events so that we
5005 * won't get any samples after PERF_RECORD_EXIT. We can however still
5006 * get a few PERF_RECORD_READ events.
5008 perf_event_task(child, child_ctx, 0);
5011 * We can recurse on the same lock type through:
5013 * __perf_event_exit_task()
5014 * sync_child_event()
5015 * fput(parent_event->filp)
5017 * mutex_lock(&ctx->mutex)
5019 * But since its the parent context it won't be the same instance.
5021 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5024 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5026 __perf_event_exit_task(child_event, child_ctx, child);
5029 * If the last event was a group event, it will have appended all
5030 * its siblings to the list, but we obtained 'tmp' before that which
5031 * will still point to the list head terminating the iteration.
5033 if (!list_empty(&child_ctx->group_list))
5036 mutex_unlock(&child_ctx->mutex);
5042 * free an unexposed, unused context as created by inheritance by
5043 * init_task below, used by fork() in case of fail.
5045 void perf_event_free_task(struct task_struct *task)
5047 struct perf_event_context *ctx = task->perf_event_ctxp;
5048 struct perf_event *event, *tmp;
5053 mutex_lock(&ctx->mutex);
5055 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5056 struct perf_event *parent = event->parent;
5058 if (WARN_ON_ONCE(!parent))
5061 mutex_lock(&parent->child_mutex);
5062 list_del_init(&event->child_list);
5063 mutex_unlock(&parent->child_mutex);
5067 list_del_event(event, ctx);
5071 if (!list_empty(&ctx->group_list))
5074 mutex_unlock(&ctx->mutex);
5080 * Initialize the perf_event context in task_struct
5082 int perf_event_init_task(struct task_struct *child)
5084 struct perf_event_context *child_ctx, *parent_ctx;
5085 struct perf_event_context *cloned_ctx;
5086 struct perf_event *event;
5087 struct task_struct *parent = current;
5088 int inherited_all = 1;
5091 child->perf_event_ctxp = NULL;
5093 mutex_init(&child->perf_event_mutex);
5094 INIT_LIST_HEAD(&child->perf_event_list);
5096 if (likely(!parent->perf_event_ctxp))
5100 * This is executed from the parent task context, so inherit
5101 * events that have been marked for cloning.
5102 * First allocate and initialize a context for the child.
5105 child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
5109 __perf_event_init_context(child_ctx, child);
5110 child->perf_event_ctxp = child_ctx;
5111 get_task_struct(child);
5114 * If the parent's context is a clone, pin it so it won't get
5117 parent_ctx = perf_pin_task_context(parent);
5120 * No need to check if parent_ctx != NULL here; since we saw
5121 * it non-NULL earlier, the only reason for it to become NULL
5122 * is if we exit, and since we're currently in the middle of
5123 * a fork we can't be exiting at the same time.
5127 * Lock the parent list. No need to lock the child - not PID
5128 * hashed yet and not running, so nobody can access it.
5130 mutex_lock(&parent_ctx->mutex);
5133 * We dont have to disable NMIs - we are only looking at
5134 * the list, not manipulating it:
5136 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5138 if (!event->attr.inherit) {
5143 ret = inherit_group(event, parent, parent_ctx,
5151 if (inherited_all) {
5153 * Mark the child context as a clone of the parent
5154 * context, or of whatever the parent is a clone of.
5155 * Note that if the parent is a clone, it could get
5156 * uncloned at any point, but that doesn't matter
5157 * because the list of events and the generation
5158 * count can't have changed since we took the mutex.
5160 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5162 child_ctx->parent_ctx = cloned_ctx;
5163 child_ctx->parent_gen = parent_ctx->parent_gen;
5165 child_ctx->parent_ctx = parent_ctx;
5166 child_ctx->parent_gen = parent_ctx->generation;
5168 get_ctx(child_ctx->parent_ctx);
5171 mutex_unlock(&parent_ctx->mutex);
5173 perf_unpin_context(parent_ctx);
5178 static void __cpuinit perf_event_init_cpu(int cpu)
5180 struct perf_cpu_context *cpuctx;
5182 cpuctx = &per_cpu(perf_cpu_context, cpu);
5183 __perf_event_init_context(&cpuctx->ctx, NULL);
5185 spin_lock(&perf_resource_lock);
5186 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5187 spin_unlock(&perf_resource_lock);
5189 hw_perf_event_setup(cpu);
5192 #ifdef CONFIG_HOTPLUG_CPU
5193 static void __perf_event_exit_cpu(void *info)
5195 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5196 struct perf_event_context *ctx = &cpuctx->ctx;
5197 struct perf_event *event, *tmp;
5199 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5200 __perf_event_remove_from_context(event);
5202 static void perf_event_exit_cpu(int cpu)
5204 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5205 struct perf_event_context *ctx = &cpuctx->ctx;
5207 mutex_lock(&ctx->mutex);
5208 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5209 mutex_unlock(&ctx->mutex);
5212 static inline void perf_event_exit_cpu(int cpu) { }
5215 static int __cpuinit
5216 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5218 unsigned int cpu = (long)hcpu;
5222 case CPU_UP_PREPARE:
5223 case CPU_UP_PREPARE_FROZEN:
5224 perf_event_init_cpu(cpu);
5228 case CPU_ONLINE_FROZEN:
5229 hw_perf_event_setup_online(cpu);
5232 case CPU_DOWN_PREPARE:
5233 case CPU_DOWN_PREPARE_FROZEN:
5234 perf_event_exit_cpu(cpu);
5245 * This has to have a higher priority than migration_notifier in sched.c.
5247 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5248 .notifier_call = perf_cpu_notify,
5252 void __init perf_event_init(void)
5254 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5255 (void *)(long)smp_processor_id());
5256 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5257 (void *)(long)smp_processor_id());
5258 register_cpu_notifier(&perf_cpu_nb);
5261 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5263 return sprintf(buf, "%d\n", perf_reserved_percpu);
5267 perf_set_reserve_percpu(struct sysdev_class *class,
5271 struct perf_cpu_context *cpuctx;
5275 err = strict_strtoul(buf, 10, &val);
5278 if (val > perf_max_events)
5281 spin_lock(&perf_resource_lock);
5282 perf_reserved_percpu = val;
5283 for_each_online_cpu(cpu) {
5284 cpuctx = &per_cpu(perf_cpu_context, cpu);
5285 spin_lock_irq(&cpuctx->ctx.lock);
5286 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5287 perf_max_events - perf_reserved_percpu);
5288 cpuctx->max_pertask = mpt;
5289 spin_unlock_irq(&cpuctx->ctx.lock);
5291 spin_unlock(&perf_resource_lock);
5296 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5298 return sprintf(buf, "%d\n", perf_overcommit);
5302 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5307 err = strict_strtoul(buf, 10, &val);
5313 spin_lock(&perf_resource_lock);
5314 perf_overcommit = val;
5315 spin_unlock(&perf_resource_lock);
5320 static SYSDEV_CLASS_ATTR(
5323 perf_show_reserve_percpu,
5324 perf_set_reserve_percpu
5327 static SYSDEV_CLASS_ATTR(
5330 perf_show_overcommit,
5334 static struct attribute *perfclass_attrs[] = {
5335 &attr_reserve_percpu.attr,
5336 &attr_overcommit.attr,
5340 static struct attribute_group perfclass_attr_group = {
5341 .attrs = perfclass_attrs,
5342 .name = "perf_events",
5345 static int __init perf_event_sysfs_init(void)
5347 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5348 &perfclass_attr_group);
5350 device_initcall(perf_event_sysfs_init);