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 static 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 raw_spin_lock_irqsave(&ctx->lock, *flags);
207 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
208 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
212 if (!atomic_inc_not_zero(&ctx->refcount)) {
213 raw_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 raw_spin_unlock_irqrestore(&ctx->lock, flags);
239 static void perf_unpin_context(struct perf_event_context *ctx)
243 raw_spin_lock_irqsave(&ctx->lock, flags);
245 raw_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);
343 * If event was in error state, then keep it
344 * that way, otherwise bogus counts will be
345 * returned on read(). The only way to get out
346 * of error state is by explicit re-enabling
349 if (event->state > PERF_EVENT_STATE_OFF)
350 event->state = PERF_EVENT_STATE_OFF;
353 * If this was a group event with sibling events then
354 * upgrade the siblings to singleton events by adding them
355 * to the context list directly:
357 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
359 list_move_tail(&sibling->group_entry, &ctx->group_list);
360 sibling->group_leader = sibling;
365 event_sched_out(struct perf_event *event,
366 struct perf_cpu_context *cpuctx,
367 struct perf_event_context *ctx)
369 if (event->state != PERF_EVENT_STATE_ACTIVE)
372 event->state = PERF_EVENT_STATE_INACTIVE;
373 if (event->pending_disable) {
374 event->pending_disable = 0;
375 event->state = PERF_EVENT_STATE_OFF;
377 event->tstamp_stopped = ctx->time;
378 event->pmu->disable(event);
381 if (!is_software_event(event))
382 cpuctx->active_oncpu--;
384 if (event->attr.exclusive || !cpuctx->active_oncpu)
385 cpuctx->exclusive = 0;
389 group_sched_out(struct perf_event *group_event,
390 struct perf_cpu_context *cpuctx,
391 struct perf_event_context *ctx)
393 struct perf_event *event;
395 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
398 event_sched_out(group_event, cpuctx, ctx);
401 * Schedule out siblings (if any):
403 list_for_each_entry(event, &group_event->sibling_list, group_entry)
404 event_sched_out(event, cpuctx, ctx);
406 if (group_event->attr.exclusive)
407 cpuctx->exclusive = 0;
411 * Cross CPU call to remove a performance event
413 * We disable the event on the hardware level first. After that we
414 * remove it from the context list.
416 static void __perf_event_remove_from_context(void *info)
418 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
419 struct perf_event *event = info;
420 struct perf_event_context *ctx = event->ctx;
423 * If this is a task context, we need to check whether it is
424 * the current task context of this cpu. If not it has been
425 * scheduled out before the smp call arrived.
427 if (ctx->task && cpuctx->task_ctx != ctx)
430 raw_spin_lock(&ctx->lock);
432 * Protect the list operation against NMI by disabling the
433 * events on a global level.
437 event_sched_out(event, cpuctx, ctx);
439 list_del_event(event, ctx);
443 * Allow more per task events with respect to the
446 cpuctx->max_pertask =
447 min(perf_max_events - ctx->nr_events,
448 perf_max_events - perf_reserved_percpu);
452 raw_spin_unlock(&ctx->lock);
457 * Remove the event from a task's (or a CPU's) list of events.
459 * Must be called with ctx->mutex held.
461 * CPU events are removed with a smp call. For task events we only
462 * call when the task is on a CPU.
464 * If event->ctx is a cloned context, callers must make sure that
465 * every task struct that event->ctx->task could possibly point to
466 * remains valid. This is OK when called from perf_release since
467 * that only calls us on the top-level context, which can't be a clone.
468 * When called from perf_event_exit_task, it's OK because the
469 * context has been detached from its task.
471 static void perf_event_remove_from_context(struct perf_event *event)
473 struct perf_event_context *ctx = event->ctx;
474 struct task_struct *task = ctx->task;
478 * Per cpu events are removed via an smp call and
479 * the removal is always successful.
481 smp_call_function_single(event->cpu,
482 __perf_event_remove_from_context,
488 task_oncpu_function_call(task, __perf_event_remove_from_context,
491 raw_spin_lock_irq(&ctx->lock);
493 * If the context is active we need to retry the smp call.
495 if (ctx->nr_active && !list_empty(&event->group_entry)) {
496 raw_spin_unlock_irq(&ctx->lock);
501 * The lock prevents that this context is scheduled in so we
502 * can remove the event safely, if the call above did not
505 if (!list_empty(&event->group_entry))
506 list_del_event(event, ctx);
507 raw_spin_unlock_irq(&ctx->lock);
511 * Update total_time_enabled and total_time_running for all events in a group.
513 static void update_group_times(struct perf_event *leader)
515 struct perf_event *event;
517 update_event_times(leader);
518 list_for_each_entry(event, &leader->sibling_list, group_entry)
519 update_event_times(event);
523 * Cross CPU call to disable a performance event
525 static void __perf_event_disable(void *info)
527 struct perf_event *event = info;
528 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
529 struct perf_event_context *ctx = event->ctx;
532 * If this is a per-task event, need to check whether this
533 * event's task is the current task on this cpu.
535 if (ctx->task && cpuctx->task_ctx != ctx)
538 raw_spin_lock(&ctx->lock);
541 * If the event is on, turn it off.
542 * If it is in error state, leave it in error state.
544 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
545 update_context_time(ctx);
546 update_group_times(event);
547 if (event == event->group_leader)
548 group_sched_out(event, cpuctx, ctx);
550 event_sched_out(event, cpuctx, ctx);
551 event->state = PERF_EVENT_STATE_OFF;
554 raw_spin_unlock(&ctx->lock);
560 * If event->ctx is a cloned context, callers must make sure that
561 * every task struct that event->ctx->task could possibly point to
562 * remains valid. This condition is satisifed when called through
563 * perf_event_for_each_child or perf_event_for_each because they
564 * hold the top-level event's child_mutex, so any descendant that
565 * goes to exit will block in sync_child_event.
566 * When called from perf_pending_event it's OK because event->ctx
567 * is the current context on this CPU and preemption is disabled,
568 * hence we can't get into perf_event_task_sched_out for this context.
570 void perf_event_disable(struct perf_event *event)
572 struct perf_event_context *ctx = event->ctx;
573 struct task_struct *task = ctx->task;
577 * Disable the event on the cpu that it's on
579 smp_call_function_single(event->cpu, __perf_event_disable,
585 task_oncpu_function_call(task, __perf_event_disable, event);
587 raw_spin_lock_irq(&ctx->lock);
589 * If the event is still active, we need to retry the cross-call.
591 if (event->state == PERF_EVENT_STATE_ACTIVE) {
592 raw_spin_unlock_irq(&ctx->lock);
597 * Since we have the lock this context can't be scheduled
598 * in, so we can change the state safely.
600 if (event->state == PERF_EVENT_STATE_INACTIVE) {
601 update_group_times(event);
602 event->state = PERF_EVENT_STATE_OFF;
605 raw_spin_unlock_irq(&ctx->lock);
609 event_sched_in(struct perf_event *event,
610 struct perf_cpu_context *cpuctx,
611 struct perf_event_context *ctx,
614 if (event->state <= PERF_EVENT_STATE_OFF)
617 event->state = PERF_EVENT_STATE_ACTIVE;
618 event->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
620 * The new state must be visible before we turn it on in the hardware:
624 if (event->pmu->enable(event)) {
625 event->state = PERF_EVENT_STATE_INACTIVE;
630 event->tstamp_running += ctx->time - event->tstamp_stopped;
632 if (!is_software_event(event))
633 cpuctx->active_oncpu++;
636 if (event->attr.exclusive)
637 cpuctx->exclusive = 1;
643 group_sched_in(struct perf_event *group_event,
644 struct perf_cpu_context *cpuctx,
645 struct perf_event_context *ctx,
648 struct perf_event *event, *partial_group;
651 if (group_event->state == PERF_EVENT_STATE_OFF)
654 ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
656 return ret < 0 ? ret : 0;
658 if (event_sched_in(group_event, cpuctx, ctx, cpu))
662 * Schedule in siblings as one group (if any):
664 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
665 if (event_sched_in(event, cpuctx, ctx, cpu)) {
666 partial_group = event;
675 * Groups can be scheduled in as one unit only, so undo any
676 * partial group before returning:
678 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
679 if (event == partial_group)
681 event_sched_out(event, cpuctx, ctx);
683 event_sched_out(group_event, cpuctx, ctx);
689 * Return 1 for a group consisting entirely of software events,
690 * 0 if the group contains any hardware events.
692 static int is_software_only_group(struct perf_event *leader)
694 struct perf_event *event;
696 if (!is_software_event(leader))
699 list_for_each_entry(event, &leader->sibling_list, group_entry)
700 if (!is_software_event(event))
707 * Work out whether we can put this event group on the CPU now.
709 static int group_can_go_on(struct perf_event *event,
710 struct perf_cpu_context *cpuctx,
714 * Groups consisting entirely of software events can always go on.
716 if (is_software_only_group(event))
719 * If an exclusive group is already on, no other hardware
722 if (cpuctx->exclusive)
725 * If this group is exclusive and there are already
726 * events on the CPU, it can't go on.
728 if (event->attr.exclusive && cpuctx->active_oncpu)
731 * Otherwise, try to add it if all previous groups were able
737 static void add_event_to_ctx(struct perf_event *event,
738 struct perf_event_context *ctx)
740 list_add_event(event, ctx);
741 event->tstamp_enabled = ctx->time;
742 event->tstamp_running = ctx->time;
743 event->tstamp_stopped = ctx->time;
747 * Cross CPU call to install and enable a performance event
749 * Must be called with ctx->mutex held
751 static void __perf_install_in_context(void *info)
753 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
754 struct perf_event *event = info;
755 struct perf_event_context *ctx = event->ctx;
756 struct perf_event *leader = event->group_leader;
757 int cpu = smp_processor_id();
761 * If this is a task context, we need to check whether it is
762 * the current task context of this cpu. If not it has been
763 * scheduled out before the smp call arrived.
764 * Or possibly this is the right context but it isn't
765 * on this cpu because it had no events.
767 if (ctx->task && cpuctx->task_ctx != ctx) {
768 if (cpuctx->task_ctx || ctx->task != current)
770 cpuctx->task_ctx = ctx;
773 raw_spin_lock(&ctx->lock);
775 update_context_time(ctx);
778 * Protect the list operation against NMI by disabling the
779 * events on a global level. NOP for non NMI based events.
783 add_event_to_ctx(event, ctx);
785 if (event->cpu != -1 && event->cpu != smp_processor_id())
789 * Don't put the event on if it is disabled or if
790 * it is in a group and the group isn't on.
792 if (event->state != PERF_EVENT_STATE_INACTIVE ||
793 (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
797 * An exclusive event can't go on if there are already active
798 * hardware events, and no hardware event can go on if there
799 * is already an exclusive event on.
801 if (!group_can_go_on(event, cpuctx, 1))
804 err = event_sched_in(event, cpuctx, ctx, cpu);
808 * This event couldn't go on. If it is in a group
809 * then we have to pull the whole group off.
810 * If the event group is pinned then put it in error state.
813 group_sched_out(leader, cpuctx, ctx);
814 if (leader->attr.pinned) {
815 update_group_times(leader);
816 leader->state = PERF_EVENT_STATE_ERROR;
820 if (!err && !ctx->task && cpuctx->max_pertask)
821 cpuctx->max_pertask--;
826 raw_spin_unlock(&ctx->lock);
830 * Attach a performance event to a context
832 * First we add the event to the list with the hardware enable bit
833 * in event->hw_config cleared.
835 * If the event is attached to a task which is on a CPU we use a smp
836 * call to enable it in the task context. The task might have been
837 * scheduled away, but we check this in the smp call again.
839 * Must be called with ctx->mutex held.
842 perf_install_in_context(struct perf_event_context *ctx,
843 struct perf_event *event,
846 struct task_struct *task = ctx->task;
850 * Per cpu events are installed via an smp call and
851 * the install is always successful.
853 smp_call_function_single(cpu, __perf_install_in_context,
859 task_oncpu_function_call(task, __perf_install_in_context,
862 raw_spin_lock_irq(&ctx->lock);
864 * we need to retry the smp call.
866 if (ctx->is_active && list_empty(&event->group_entry)) {
867 raw_spin_unlock_irq(&ctx->lock);
872 * The lock prevents that this context is scheduled in so we
873 * can add the event safely, if it the call above did not
876 if (list_empty(&event->group_entry))
877 add_event_to_ctx(event, ctx);
878 raw_spin_unlock_irq(&ctx->lock);
882 * Put a event into inactive state and update time fields.
883 * Enabling the leader of a group effectively enables all
884 * the group members that aren't explicitly disabled, so we
885 * have to update their ->tstamp_enabled also.
886 * Note: this works for group members as well as group leaders
887 * since the non-leader members' sibling_lists will be empty.
889 static void __perf_event_mark_enabled(struct perf_event *event,
890 struct perf_event_context *ctx)
892 struct perf_event *sub;
894 event->state = PERF_EVENT_STATE_INACTIVE;
895 event->tstamp_enabled = ctx->time - event->total_time_enabled;
896 list_for_each_entry(sub, &event->sibling_list, group_entry)
897 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
898 sub->tstamp_enabled =
899 ctx->time - sub->total_time_enabled;
903 * Cross CPU call to enable a performance event
905 static void __perf_event_enable(void *info)
907 struct perf_event *event = info;
908 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
909 struct perf_event_context *ctx = event->ctx;
910 struct perf_event *leader = event->group_leader;
914 * If this is a per-task event, need to check whether this
915 * event's task is the current task on this cpu.
917 if (ctx->task && cpuctx->task_ctx != ctx) {
918 if (cpuctx->task_ctx || ctx->task != current)
920 cpuctx->task_ctx = ctx;
923 raw_spin_lock(&ctx->lock);
925 update_context_time(ctx);
927 if (event->state >= PERF_EVENT_STATE_INACTIVE)
929 __perf_event_mark_enabled(event, ctx);
931 if (event->cpu != -1 && event->cpu != smp_processor_id())
935 * If the event is in a group and isn't the group leader,
936 * then don't put it on unless the group is on.
938 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
941 if (!group_can_go_on(event, cpuctx, 1)) {
946 err = group_sched_in(event, cpuctx, ctx,
949 err = event_sched_in(event, cpuctx, ctx,
956 * If this event can't go on and it's part of a
957 * group, then the whole group has to come off.
960 group_sched_out(leader, cpuctx, ctx);
961 if (leader->attr.pinned) {
962 update_group_times(leader);
963 leader->state = PERF_EVENT_STATE_ERROR;
968 raw_spin_unlock(&ctx->lock);
974 * If event->ctx is a cloned context, callers must make sure that
975 * every task struct that event->ctx->task could possibly point to
976 * remains valid. This condition is satisfied when called through
977 * perf_event_for_each_child or perf_event_for_each as described
978 * for perf_event_disable.
980 void perf_event_enable(struct perf_event *event)
982 struct perf_event_context *ctx = event->ctx;
983 struct task_struct *task = ctx->task;
987 * Enable the event on the cpu that it's on
989 smp_call_function_single(event->cpu, __perf_event_enable,
994 raw_spin_lock_irq(&ctx->lock);
995 if (event->state >= PERF_EVENT_STATE_INACTIVE)
999 * If the event is in error state, clear that first.
1000 * That way, if we see the event in error state below, we
1001 * know that it has gone back into error state, as distinct
1002 * from the task having been scheduled away before the
1003 * cross-call arrived.
1005 if (event->state == PERF_EVENT_STATE_ERROR)
1006 event->state = PERF_EVENT_STATE_OFF;
1009 raw_spin_unlock_irq(&ctx->lock);
1010 task_oncpu_function_call(task, __perf_event_enable, event);
1012 raw_spin_lock_irq(&ctx->lock);
1015 * If the context is active and the event is still off,
1016 * we need to retry the cross-call.
1018 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1022 * Since we have the lock this context can't be scheduled
1023 * in, so we can change the state safely.
1025 if (event->state == PERF_EVENT_STATE_OFF)
1026 __perf_event_mark_enabled(event, ctx);
1029 raw_spin_unlock_irq(&ctx->lock);
1032 static int perf_event_refresh(struct perf_event *event, int refresh)
1035 * not supported on inherited events
1037 if (event->attr.inherit)
1040 atomic_add(refresh, &event->event_limit);
1041 perf_event_enable(event);
1046 void __perf_event_sched_out(struct perf_event_context *ctx,
1047 struct perf_cpu_context *cpuctx)
1049 struct perf_event *event;
1051 raw_spin_lock(&ctx->lock);
1053 if (likely(!ctx->nr_events))
1055 update_context_time(ctx);
1058 if (ctx->nr_active) {
1059 list_for_each_entry(event, &ctx->group_list, group_entry)
1060 group_sched_out(event, cpuctx, ctx);
1064 raw_spin_unlock(&ctx->lock);
1068 * Test whether two contexts are equivalent, i.e. whether they
1069 * have both been cloned from the same version of the same context
1070 * and they both have the same number of enabled events.
1071 * If the number of enabled events is the same, then the set
1072 * of enabled events should be the same, because these are both
1073 * inherited contexts, therefore we can't access individual events
1074 * in them directly with an fd; we can only enable/disable all
1075 * events via prctl, or enable/disable all events in a family
1076 * via ioctl, which will have the same effect on both contexts.
1078 static int context_equiv(struct perf_event_context *ctx1,
1079 struct perf_event_context *ctx2)
1081 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1082 && ctx1->parent_gen == ctx2->parent_gen
1083 && !ctx1->pin_count && !ctx2->pin_count;
1086 static void __perf_event_sync_stat(struct perf_event *event,
1087 struct perf_event *next_event)
1091 if (!event->attr.inherit_stat)
1095 * Update the event value, we cannot use perf_event_read()
1096 * because we're in the middle of a context switch and have IRQs
1097 * disabled, which upsets smp_call_function_single(), however
1098 * we know the event must be on the current CPU, therefore we
1099 * don't need to use it.
1101 switch (event->state) {
1102 case PERF_EVENT_STATE_ACTIVE:
1103 event->pmu->read(event);
1106 case PERF_EVENT_STATE_INACTIVE:
1107 update_event_times(event);
1115 * In order to keep per-task stats reliable we need to flip the event
1116 * values when we flip the contexts.
1118 value = atomic64_read(&next_event->count);
1119 value = atomic64_xchg(&event->count, value);
1120 atomic64_set(&next_event->count, value);
1122 swap(event->total_time_enabled, next_event->total_time_enabled);
1123 swap(event->total_time_running, next_event->total_time_running);
1126 * Since we swizzled the values, update the user visible data too.
1128 perf_event_update_userpage(event);
1129 perf_event_update_userpage(next_event);
1132 #define list_next_entry(pos, member) \
1133 list_entry(pos->member.next, typeof(*pos), member)
1135 static void perf_event_sync_stat(struct perf_event_context *ctx,
1136 struct perf_event_context *next_ctx)
1138 struct perf_event *event, *next_event;
1143 update_context_time(ctx);
1145 event = list_first_entry(&ctx->event_list,
1146 struct perf_event, event_entry);
1148 next_event = list_first_entry(&next_ctx->event_list,
1149 struct perf_event, event_entry);
1151 while (&event->event_entry != &ctx->event_list &&
1152 &next_event->event_entry != &next_ctx->event_list) {
1154 __perf_event_sync_stat(event, next_event);
1156 event = list_next_entry(event, event_entry);
1157 next_event = list_next_entry(next_event, event_entry);
1162 * Called from scheduler to remove the events of the current task,
1163 * with interrupts disabled.
1165 * We stop each event and update the event value in event->count.
1167 * This does not protect us against NMI, but disable()
1168 * sets the disabled bit in the control field of event _before_
1169 * accessing the event control register. If a NMI hits, then it will
1170 * not restart the event.
1172 void perf_event_task_sched_out(struct task_struct *task,
1173 struct task_struct *next, int cpu)
1175 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1176 struct perf_event_context *ctx = task->perf_event_ctxp;
1177 struct perf_event_context *next_ctx;
1178 struct perf_event_context *parent;
1179 struct pt_regs *regs;
1182 regs = task_pt_regs(task);
1183 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1185 if (likely(!ctx || !cpuctx->task_ctx))
1189 parent = rcu_dereference(ctx->parent_ctx);
1190 next_ctx = next->perf_event_ctxp;
1191 if (parent && next_ctx &&
1192 rcu_dereference(next_ctx->parent_ctx) == parent) {
1194 * Looks like the two contexts are clones, so we might be
1195 * able to optimize the context switch. We lock both
1196 * contexts and check that they are clones under the
1197 * lock (including re-checking that neither has been
1198 * uncloned in the meantime). It doesn't matter which
1199 * order we take the locks because no other cpu could
1200 * be trying to lock both of these tasks.
1202 raw_spin_lock(&ctx->lock);
1203 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1204 if (context_equiv(ctx, next_ctx)) {
1206 * XXX do we need a memory barrier of sorts
1207 * wrt to rcu_dereference() of perf_event_ctxp
1209 task->perf_event_ctxp = next_ctx;
1210 next->perf_event_ctxp = ctx;
1212 next_ctx->task = task;
1215 perf_event_sync_stat(ctx, next_ctx);
1217 raw_spin_unlock(&next_ctx->lock);
1218 raw_spin_unlock(&ctx->lock);
1223 __perf_event_sched_out(ctx, cpuctx);
1224 cpuctx->task_ctx = NULL;
1229 * Called with IRQs disabled
1231 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1233 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1235 if (!cpuctx->task_ctx)
1238 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1241 __perf_event_sched_out(ctx, cpuctx);
1242 cpuctx->task_ctx = NULL;
1246 * Called with IRQs disabled
1248 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1250 __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1254 __perf_event_sched_in(struct perf_event_context *ctx,
1255 struct perf_cpu_context *cpuctx, int cpu)
1257 struct perf_event *event;
1260 raw_spin_lock(&ctx->lock);
1262 if (likely(!ctx->nr_events))
1265 ctx->timestamp = perf_clock();
1270 * First go through the list and put on any pinned groups
1271 * in order to give them the best chance of going on.
1273 list_for_each_entry(event, &ctx->group_list, group_entry) {
1274 if (event->state <= PERF_EVENT_STATE_OFF ||
1275 !event->attr.pinned)
1277 if (event->cpu != -1 && event->cpu != cpu)
1280 if (group_can_go_on(event, cpuctx, 1))
1281 group_sched_in(event, cpuctx, ctx, cpu);
1284 * If this pinned group hasn't been scheduled,
1285 * put it in error state.
1287 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1288 update_group_times(event);
1289 event->state = PERF_EVENT_STATE_ERROR;
1293 list_for_each_entry(event, &ctx->group_list, group_entry) {
1295 * Ignore events in OFF or ERROR state, and
1296 * ignore pinned events since we did them already.
1298 if (event->state <= PERF_EVENT_STATE_OFF ||
1303 * Listen to the 'cpu' scheduling filter constraint
1306 if (event->cpu != -1 && event->cpu != cpu)
1309 if (group_can_go_on(event, cpuctx, can_add_hw))
1310 if (group_sched_in(event, cpuctx, ctx, cpu))
1315 raw_spin_unlock(&ctx->lock);
1319 * Called from scheduler to add the events of the current task
1320 * with interrupts disabled.
1322 * We restore the event value and then enable it.
1324 * This does not protect us against NMI, but enable()
1325 * sets the enabled bit in the control field of event _before_
1326 * accessing the event control register. If a NMI hits, then it will
1327 * keep the event running.
1329 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1331 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1332 struct perf_event_context *ctx = task->perf_event_ctxp;
1336 if (cpuctx->task_ctx == ctx)
1338 __perf_event_sched_in(ctx, cpuctx, cpu);
1339 cpuctx->task_ctx = ctx;
1342 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1344 struct perf_event_context *ctx = &cpuctx->ctx;
1346 __perf_event_sched_in(ctx, cpuctx, cpu);
1349 #define MAX_INTERRUPTS (~0ULL)
1351 static void perf_log_throttle(struct perf_event *event, int enable);
1353 static void perf_adjust_period(struct perf_event *event, u64 events)
1355 struct hw_perf_event *hwc = &event->hw;
1356 u64 period, sample_period;
1359 events *= hwc->sample_period;
1360 period = div64_u64(events, event->attr.sample_freq);
1362 delta = (s64)(period - hwc->sample_period);
1363 delta = (delta + 7) / 8; /* low pass filter */
1365 sample_period = hwc->sample_period + delta;
1370 hwc->sample_period = sample_period;
1373 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1375 struct perf_event *event;
1376 struct hw_perf_event *hwc;
1377 u64 interrupts, freq;
1379 raw_spin_lock(&ctx->lock);
1380 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1381 if (event->state != PERF_EVENT_STATE_ACTIVE)
1384 if (event->cpu != -1 && event->cpu != smp_processor_id())
1389 interrupts = hwc->interrupts;
1390 hwc->interrupts = 0;
1393 * unthrottle events on the tick
1395 if (interrupts == MAX_INTERRUPTS) {
1396 perf_log_throttle(event, 1);
1397 event->pmu->unthrottle(event);
1398 interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1401 if (!event->attr.freq || !event->attr.sample_freq)
1405 * if the specified freq < HZ then we need to skip ticks
1407 if (event->attr.sample_freq < HZ) {
1408 freq = event->attr.sample_freq;
1410 hwc->freq_count += freq;
1411 hwc->freq_interrupts += interrupts;
1413 if (hwc->freq_count < HZ)
1416 interrupts = hwc->freq_interrupts;
1417 hwc->freq_interrupts = 0;
1418 hwc->freq_count -= HZ;
1422 perf_adjust_period(event, freq * interrupts);
1425 * In order to avoid being stalled by an (accidental) huge
1426 * sample period, force reset the sample period if we didn't
1427 * get any events in this freq period.
1431 event->pmu->disable(event);
1432 atomic64_set(&hwc->period_left, 0);
1433 event->pmu->enable(event);
1437 raw_spin_unlock(&ctx->lock);
1441 * Round-robin a context's events:
1443 static void rotate_ctx(struct perf_event_context *ctx)
1445 struct perf_event *event;
1447 if (!ctx->nr_events)
1450 raw_spin_lock(&ctx->lock);
1452 * Rotate the first entry last (works just fine for group events too):
1455 list_for_each_entry(event, &ctx->group_list, group_entry) {
1456 list_move_tail(&event->group_entry, &ctx->group_list);
1461 raw_spin_unlock(&ctx->lock);
1464 void perf_event_task_tick(struct task_struct *curr, int cpu)
1466 struct perf_cpu_context *cpuctx;
1467 struct perf_event_context *ctx;
1469 if (!atomic_read(&nr_events))
1472 cpuctx = &per_cpu(perf_cpu_context, cpu);
1473 ctx = curr->perf_event_ctxp;
1475 perf_ctx_adjust_freq(&cpuctx->ctx);
1477 perf_ctx_adjust_freq(ctx);
1479 perf_event_cpu_sched_out(cpuctx);
1481 __perf_event_task_sched_out(ctx);
1483 rotate_ctx(&cpuctx->ctx);
1487 perf_event_cpu_sched_in(cpuctx, cpu);
1489 perf_event_task_sched_in(curr, cpu);
1493 * Enable all of a task's events that have been marked enable-on-exec.
1494 * This expects task == current.
1496 static void perf_event_enable_on_exec(struct task_struct *task)
1498 struct perf_event_context *ctx;
1499 struct perf_event *event;
1500 unsigned long flags;
1503 local_irq_save(flags);
1504 ctx = task->perf_event_ctxp;
1505 if (!ctx || !ctx->nr_events)
1508 __perf_event_task_sched_out(ctx);
1510 raw_spin_lock(&ctx->lock);
1512 list_for_each_entry(event, &ctx->group_list, group_entry) {
1513 if (!event->attr.enable_on_exec)
1515 event->attr.enable_on_exec = 0;
1516 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1518 __perf_event_mark_enabled(event, ctx);
1523 * Unclone this context if we enabled any event.
1528 raw_spin_unlock(&ctx->lock);
1530 perf_event_task_sched_in(task, smp_processor_id());
1532 local_irq_restore(flags);
1536 * Cross CPU call to read the hardware event
1538 static void __perf_event_read(void *info)
1540 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1541 struct perf_event *event = info;
1542 struct perf_event_context *ctx = event->ctx;
1545 * If this is a task context, we need to check whether it is
1546 * the current task context of this cpu. If not it has been
1547 * scheduled out before the smp call arrived. In that case
1548 * event->count would have been updated to a recent sample
1549 * when the event was scheduled out.
1551 if (ctx->task && cpuctx->task_ctx != ctx)
1554 raw_spin_lock(&ctx->lock);
1555 update_context_time(ctx);
1556 update_event_times(event);
1557 raw_spin_unlock(&ctx->lock);
1559 event->pmu->read(event);
1562 static u64 perf_event_read(struct perf_event *event)
1565 * If event is enabled and currently active on a CPU, update the
1566 * value in the event structure:
1568 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1569 smp_call_function_single(event->oncpu,
1570 __perf_event_read, event, 1);
1571 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1572 struct perf_event_context *ctx = event->ctx;
1573 unsigned long flags;
1575 raw_spin_lock_irqsave(&ctx->lock, flags);
1576 update_context_time(ctx);
1577 update_event_times(event);
1578 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1581 return atomic64_read(&event->count);
1585 * Initialize the perf_event context in a task_struct:
1588 __perf_event_init_context(struct perf_event_context *ctx,
1589 struct task_struct *task)
1591 raw_spin_lock_init(&ctx->lock);
1592 mutex_init(&ctx->mutex);
1593 INIT_LIST_HEAD(&ctx->group_list);
1594 INIT_LIST_HEAD(&ctx->event_list);
1595 atomic_set(&ctx->refcount, 1);
1599 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1601 struct perf_event_context *ctx;
1602 struct perf_cpu_context *cpuctx;
1603 struct task_struct *task;
1604 unsigned long flags;
1607 if (pid == -1 && cpu != -1) {
1608 /* Must be root to operate on a CPU event: */
1609 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1610 return ERR_PTR(-EACCES);
1612 if (cpu < 0 || cpu >= nr_cpumask_bits)
1613 return ERR_PTR(-EINVAL);
1616 * We could be clever and allow to attach a event to an
1617 * offline CPU and activate it when the CPU comes up, but
1620 if (!cpu_online(cpu))
1621 return ERR_PTR(-ENODEV);
1623 cpuctx = &per_cpu(perf_cpu_context, cpu);
1634 task = find_task_by_vpid(pid);
1636 get_task_struct(task);
1640 return ERR_PTR(-ESRCH);
1643 * Can't attach events to a dying task.
1646 if (task->flags & PF_EXITING)
1649 /* Reuse ptrace permission checks for now. */
1651 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1655 ctx = perf_lock_task_context(task, &flags);
1658 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1662 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1666 __perf_event_init_context(ctx, task);
1668 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1670 * We raced with some other task; use
1671 * the context they set.
1676 get_task_struct(task);
1679 put_task_struct(task);
1683 put_task_struct(task);
1684 return ERR_PTR(err);
1687 static void perf_event_free_filter(struct perf_event *event);
1689 static void free_event_rcu(struct rcu_head *head)
1691 struct perf_event *event;
1693 event = container_of(head, struct perf_event, rcu_head);
1695 put_pid_ns(event->ns);
1696 perf_event_free_filter(event);
1700 static void perf_pending_sync(struct perf_event *event);
1702 static void free_event(struct perf_event *event)
1704 perf_pending_sync(event);
1706 if (!event->parent) {
1707 atomic_dec(&nr_events);
1708 if (event->attr.mmap)
1709 atomic_dec(&nr_mmap_events);
1710 if (event->attr.comm)
1711 atomic_dec(&nr_comm_events);
1712 if (event->attr.task)
1713 atomic_dec(&nr_task_events);
1716 if (event->output) {
1717 fput(event->output->filp);
1718 event->output = NULL;
1722 event->destroy(event);
1724 put_ctx(event->ctx);
1725 call_rcu(&event->rcu_head, free_event_rcu);
1728 int perf_event_release_kernel(struct perf_event *event)
1730 struct perf_event_context *ctx = event->ctx;
1732 WARN_ON_ONCE(ctx->parent_ctx);
1733 mutex_lock(&ctx->mutex);
1734 perf_event_remove_from_context(event);
1735 mutex_unlock(&ctx->mutex);
1737 mutex_lock(&event->owner->perf_event_mutex);
1738 list_del_init(&event->owner_entry);
1739 mutex_unlock(&event->owner->perf_event_mutex);
1740 put_task_struct(event->owner);
1746 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1749 * Called when the last reference to the file is gone.
1751 static int perf_release(struct inode *inode, struct file *file)
1753 struct perf_event *event = file->private_data;
1755 file->private_data = NULL;
1757 return perf_event_release_kernel(event);
1760 static int perf_event_read_size(struct perf_event *event)
1762 int entry = sizeof(u64); /* value */
1766 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1767 size += sizeof(u64);
1769 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1770 size += sizeof(u64);
1772 if (event->attr.read_format & PERF_FORMAT_ID)
1773 entry += sizeof(u64);
1775 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1776 nr += event->group_leader->nr_siblings;
1777 size += sizeof(u64);
1785 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1787 struct perf_event *child;
1793 mutex_lock(&event->child_mutex);
1794 total += perf_event_read(event);
1795 *enabled += event->total_time_enabled +
1796 atomic64_read(&event->child_total_time_enabled);
1797 *running += event->total_time_running +
1798 atomic64_read(&event->child_total_time_running);
1800 list_for_each_entry(child, &event->child_list, child_list) {
1801 total += perf_event_read(child);
1802 *enabled += child->total_time_enabled;
1803 *running += child->total_time_running;
1805 mutex_unlock(&event->child_mutex);
1809 EXPORT_SYMBOL_GPL(perf_event_read_value);
1811 static int perf_event_read_group(struct perf_event *event,
1812 u64 read_format, char __user *buf)
1814 struct perf_event *leader = event->group_leader, *sub;
1815 int n = 0, size = 0, ret = -EFAULT;
1816 struct perf_event_context *ctx = leader->ctx;
1818 u64 count, enabled, running;
1820 mutex_lock(&ctx->mutex);
1821 count = perf_event_read_value(leader, &enabled, &running);
1823 values[n++] = 1 + leader->nr_siblings;
1824 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1825 values[n++] = enabled;
1826 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1827 values[n++] = running;
1828 values[n++] = count;
1829 if (read_format & PERF_FORMAT_ID)
1830 values[n++] = primary_event_id(leader);
1832 size = n * sizeof(u64);
1834 if (copy_to_user(buf, values, size))
1839 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1842 values[n++] = perf_event_read_value(sub, &enabled, &running);
1843 if (read_format & PERF_FORMAT_ID)
1844 values[n++] = primary_event_id(sub);
1846 size = n * sizeof(u64);
1848 if (copy_to_user(buf + ret, values, size)) {
1856 mutex_unlock(&ctx->mutex);
1861 static int perf_event_read_one(struct perf_event *event,
1862 u64 read_format, char __user *buf)
1864 u64 enabled, running;
1868 values[n++] = perf_event_read_value(event, &enabled, &running);
1869 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1870 values[n++] = enabled;
1871 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1872 values[n++] = running;
1873 if (read_format & PERF_FORMAT_ID)
1874 values[n++] = primary_event_id(event);
1876 if (copy_to_user(buf, values, n * sizeof(u64)))
1879 return n * sizeof(u64);
1883 * Read the performance event - simple non blocking version for now
1886 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1888 u64 read_format = event->attr.read_format;
1892 * Return end-of-file for a read on a event that is in
1893 * error state (i.e. because it was pinned but it couldn't be
1894 * scheduled on to the CPU at some point).
1896 if (event->state == PERF_EVENT_STATE_ERROR)
1899 if (count < perf_event_read_size(event))
1902 WARN_ON_ONCE(event->ctx->parent_ctx);
1903 if (read_format & PERF_FORMAT_GROUP)
1904 ret = perf_event_read_group(event, read_format, buf);
1906 ret = perf_event_read_one(event, read_format, buf);
1912 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1914 struct perf_event *event = file->private_data;
1916 return perf_read_hw(event, buf, count);
1919 static unsigned int perf_poll(struct file *file, poll_table *wait)
1921 struct perf_event *event = file->private_data;
1922 struct perf_mmap_data *data;
1923 unsigned int events = POLL_HUP;
1926 data = rcu_dereference(event->data);
1928 events = atomic_xchg(&data->poll, 0);
1931 poll_wait(file, &event->waitq, wait);
1936 static void perf_event_reset(struct perf_event *event)
1938 (void)perf_event_read(event);
1939 atomic64_set(&event->count, 0);
1940 perf_event_update_userpage(event);
1944 * Holding the top-level event's child_mutex means that any
1945 * descendant process that has inherited this event will block
1946 * in sync_child_event if it goes to exit, thus satisfying the
1947 * task existence requirements of perf_event_enable/disable.
1949 static void perf_event_for_each_child(struct perf_event *event,
1950 void (*func)(struct perf_event *))
1952 struct perf_event *child;
1954 WARN_ON_ONCE(event->ctx->parent_ctx);
1955 mutex_lock(&event->child_mutex);
1957 list_for_each_entry(child, &event->child_list, child_list)
1959 mutex_unlock(&event->child_mutex);
1962 static void perf_event_for_each(struct perf_event *event,
1963 void (*func)(struct perf_event *))
1965 struct perf_event_context *ctx = event->ctx;
1966 struct perf_event *sibling;
1968 WARN_ON_ONCE(ctx->parent_ctx);
1969 mutex_lock(&ctx->mutex);
1970 event = event->group_leader;
1972 perf_event_for_each_child(event, func);
1974 list_for_each_entry(sibling, &event->sibling_list, group_entry)
1975 perf_event_for_each_child(event, func);
1976 mutex_unlock(&ctx->mutex);
1979 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1981 struct perf_event_context *ctx = event->ctx;
1986 if (!event->attr.sample_period)
1989 size = copy_from_user(&value, arg, sizeof(value));
1990 if (size != sizeof(value))
1996 raw_spin_lock_irq(&ctx->lock);
1997 if (event->attr.freq) {
1998 if (value > sysctl_perf_event_sample_rate) {
2003 event->attr.sample_freq = value;
2005 event->attr.sample_period = value;
2006 event->hw.sample_period = value;
2009 raw_spin_unlock_irq(&ctx->lock);
2014 static int perf_event_set_output(struct perf_event *event, int output_fd);
2015 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2017 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2019 struct perf_event *event = file->private_data;
2020 void (*func)(struct perf_event *);
2024 case PERF_EVENT_IOC_ENABLE:
2025 func = perf_event_enable;
2027 case PERF_EVENT_IOC_DISABLE:
2028 func = perf_event_disable;
2030 case PERF_EVENT_IOC_RESET:
2031 func = perf_event_reset;
2034 case PERF_EVENT_IOC_REFRESH:
2035 return perf_event_refresh(event, arg);
2037 case PERF_EVENT_IOC_PERIOD:
2038 return perf_event_period(event, (u64 __user *)arg);
2040 case PERF_EVENT_IOC_SET_OUTPUT:
2041 return perf_event_set_output(event, arg);
2043 case PERF_EVENT_IOC_SET_FILTER:
2044 return perf_event_set_filter(event, (void __user *)arg);
2050 if (flags & PERF_IOC_FLAG_GROUP)
2051 perf_event_for_each(event, func);
2053 perf_event_for_each_child(event, func);
2058 int perf_event_task_enable(void)
2060 struct perf_event *event;
2062 mutex_lock(¤t->perf_event_mutex);
2063 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2064 perf_event_for_each_child(event, perf_event_enable);
2065 mutex_unlock(¤t->perf_event_mutex);
2070 int perf_event_task_disable(void)
2072 struct perf_event *event;
2074 mutex_lock(¤t->perf_event_mutex);
2075 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
2076 perf_event_for_each_child(event, perf_event_disable);
2077 mutex_unlock(¤t->perf_event_mutex);
2082 #ifndef PERF_EVENT_INDEX_OFFSET
2083 # define PERF_EVENT_INDEX_OFFSET 0
2086 static int perf_event_index(struct perf_event *event)
2088 if (event->state != PERF_EVENT_STATE_ACTIVE)
2091 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2095 * Callers need to ensure there can be no nesting of this function, otherwise
2096 * the seqlock logic goes bad. We can not serialize this because the arch
2097 * code calls this from NMI context.
2099 void perf_event_update_userpage(struct perf_event *event)
2101 struct perf_event_mmap_page *userpg;
2102 struct perf_mmap_data *data;
2105 data = rcu_dereference(event->data);
2109 userpg = data->user_page;
2112 * Disable preemption so as to not let the corresponding user-space
2113 * spin too long if we get preempted.
2118 userpg->index = perf_event_index(event);
2119 userpg->offset = atomic64_read(&event->count);
2120 if (event->state == PERF_EVENT_STATE_ACTIVE)
2121 userpg->offset -= atomic64_read(&event->hw.prev_count);
2123 userpg->time_enabled = event->total_time_enabled +
2124 atomic64_read(&event->child_total_time_enabled);
2126 userpg->time_running = event->total_time_running +
2127 atomic64_read(&event->child_total_time_running);
2136 static unsigned long perf_data_size(struct perf_mmap_data *data)
2138 return data->nr_pages << (PAGE_SHIFT + data->data_order);
2141 #ifndef CONFIG_PERF_USE_VMALLOC
2144 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2147 static struct page *
2148 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2150 if (pgoff > data->nr_pages)
2154 return virt_to_page(data->user_page);
2156 return virt_to_page(data->data_pages[pgoff - 1]);
2159 static struct perf_mmap_data *
2160 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2162 struct perf_mmap_data *data;
2166 WARN_ON(atomic_read(&event->mmap_count));
2168 size = sizeof(struct perf_mmap_data);
2169 size += nr_pages * sizeof(void *);
2171 data = kzalloc(size, GFP_KERNEL);
2175 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2176 if (!data->user_page)
2177 goto fail_user_page;
2179 for (i = 0; i < nr_pages; i++) {
2180 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2181 if (!data->data_pages[i])
2182 goto fail_data_pages;
2185 data->data_order = 0;
2186 data->nr_pages = nr_pages;
2191 for (i--; i >= 0; i--)
2192 free_page((unsigned long)data->data_pages[i]);
2194 free_page((unsigned long)data->user_page);
2203 static void perf_mmap_free_page(unsigned long addr)
2205 struct page *page = virt_to_page((void *)addr);
2207 page->mapping = NULL;
2211 static void perf_mmap_data_free(struct perf_mmap_data *data)
2215 perf_mmap_free_page((unsigned long)data->user_page);
2216 for (i = 0; i < data->nr_pages; i++)
2217 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2224 * Back perf_mmap() with vmalloc memory.
2226 * Required for architectures that have d-cache aliasing issues.
2229 static struct page *
2230 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2232 if (pgoff > (1UL << data->data_order))
2235 return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2238 static void perf_mmap_unmark_page(void *addr)
2240 struct page *page = vmalloc_to_page(addr);
2242 page->mapping = NULL;
2245 static void perf_mmap_data_free_work(struct work_struct *work)
2247 struct perf_mmap_data *data;
2251 data = container_of(work, struct perf_mmap_data, work);
2252 nr = 1 << data->data_order;
2254 base = data->user_page;
2255 for (i = 0; i < nr + 1; i++)
2256 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2262 static void perf_mmap_data_free(struct perf_mmap_data *data)
2264 schedule_work(&data->work);
2267 static struct perf_mmap_data *
2268 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2270 struct perf_mmap_data *data;
2274 WARN_ON(atomic_read(&event->mmap_count));
2276 size = sizeof(struct perf_mmap_data);
2277 size += sizeof(void *);
2279 data = kzalloc(size, GFP_KERNEL);
2283 INIT_WORK(&data->work, perf_mmap_data_free_work);
2285 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2289 data->user_page = all_buf;
2290 data->data_pages[0] = all_buf + PAGE_SIZE;
2291 data->data_order = ilog2(nr_pages);
2305 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2307 struct perf_event *event = vma->vm_file->private_data;
2308 struct perf_mmap_data *data;
2309 int ret = VM_FAULT_SIGBUS;
2311 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2312 if (vmf->pgoff == 0)
2318 data = rcu_dereference(event->data);
2322 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2325 vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2329 get_page(vmf->page);
2330 vmf->page->mapping = vma->vm_file->f_mapping;
2331 vmf->page->index = vmf->pgoff;
2341 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2343 long max_size = perf_data_size(data);
2345 atomic_set(&data->lock, -1);
2347 if (event->attr.watermark) {
2348 data->watermark = min_t(long, max_size,
2349 event->attr.wakeup_watermark);
2352 if (!data->watermark)
2353 data->watermark = max_size / 2;
2356 rcu_assign_pointer(event->data, data);
2359 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2361 struct perf_mmap_data *data;
2363 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2364 perf_mmap_data_free(data);
2367 static void perf_mmap_data_release(struct perf_event *event)
2369 struct perf_mmap_data *data = event->data;
2371 WARN_ON(atomic_read(&event->mmap_count));
2373 rcu_assign_pointer(event->data, NULL);
2374 call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2377 static void perf_mmap_open(struct vm_area_struct *vma)
2379 struct perf_event *event = vma->vm_file->private_data;
2381 atomic_inc(&event->mmap_count);
2384 static void perf_mmap_close(struct vm_area_struct *vma)
2386 struct perf_event *event = vma->vm_file->private_data;
2388 WARN_ON_ONCE(event->ctx->parent_ctx);
2389 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2390 unsigned long size = perf_data_size(event->data);
2391 struct user_struct *user = current_user();
2393 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2394 vma->vm_mm->locked_vm -= event->data->nr_locked;
2395 perf_mmap_data_release(event);
2396 mutex_unlock(&event->mmap_mutex);
2400 static const struct vm_operations_struct perf_mmap_vmops = {
2401 .open = perf_mmap_open,
2402 .close = perf_mmap_close,
2403 .fault = perf_mmap_fault,
2404 .page_mkwrite = perf_mmap_fault,
2407 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2409 struct perf_event *event = file->private_data;
2410 unsigned long user_locked, user_lock_limit;
2411 struct user_struct *user = current_user();
2412 unsigned long locked, lock_limit;
2413 struct perf_mmap_data *data;
2414 unsigned long vma_size;
2415 unsigned long nr_pages;
2416 long user_extra, extra;
2419 if (!(vma->vm_flags & VM_SHARED))
2422 vma_size = vma->vm_end - vma->vm_start;
2423 nr_pages = (vma_size / PAGE_SIZE) - 1;
2426 * If we have data pages ensure they're a power-of-two number, so we
2427 * can do bitmasks instead of modulo.
2429 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2432 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2435 if (vma->vm_pgoff != 0)
2438 WARN_ON_ONCE(event->ctx->parent_ctx);
2439 mutex_lock(&event->mmap_mutex);
2440 if (event->output) {
2445 if (atomic_inc_not_zero(&event->mmap_count)) {
2446 if (nr_pages != event->data->nr_pages)
2451 user_extra = nr_pages + 1;
2452 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2455 * Increase the limit linearly with more CPUs:
2457 user_lock_limit *= num_online_cpus();
2459 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2462 if (user_locked > user_lock_limit)
2463 extra = user_locked - user_lock_limit;
2465 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2466 lock_limit >>= PAGE_SHIFT;
2467 locked = vma->vm_mm->locked_vm + extra;
2469 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2470 !capable(CAP_IPC_LOCK)) {
2475 WARN_ON(event->data);
2477 data = perf_mmap_data_alloc(event, nr_pages);
2483 perf_mmap_data_init(event, data);
2485 atomic_set(&event->mmap_count, 1);
2486 atomic_long_add(user_extra, &user->locked_vm);
2487 vma->vm_mm->locked_vm += extra;
2488 event->data->nr_locked = extra;
2489 if (vma->vm_flags & VM_WRITE)
2490 event->data->writable = 1;
2493 mutex_unlock(&event->mmap_mutex);
2495 vma->vm_flags |= VM_RESERVED;
2496 vma->vm_ops = &perf_mmap_vmops;
2501 static int perf_fasync(int fd, struct file *filp, int on)
2503 struct inode *inode = filp->f_path.dentry->d_inode;
2504 struct perf_event *event = filp->private_data;
2507 mutex_lock(&inode->i_mutex);
2508 retval = fasync_helper(fd, filp, on, &event->fasync);
2509 mutex_unlock(&inode->i_mutex);
2517 static const struct file_operations perf_fops = {
2518 .release = perf_release,
2521 .unlocked_ioctl = perf_ioctl,
2522 .compat_ioctl = perf_ioctl,
2524 .fasync = perf_fasync,
2530 * If there's data, ensure we set the poll() state and publish everything
2531 * to user-space before waking everybody up.
2534 void perf_event_wakeup(struct perf_event *event)
2536 wake_up_all(&event->waitq);
2538 if (event->pending_kill) {
2539 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2540 event->pending_kill = 0;
2547 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2549 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2550 * single linked list and use cmpxchg() to add entries lockless.
2553 static void perf_pending_event(struct perf_pending_entry *entry)
2555 struct perf_event *event = container_of(entry,
2556 struct perf_event, pending);
2558 if (event->pending_disable) {
2559 event->pending_disable = 0;
2560 __perf_event_disable(event);
2563 if (event->pending_wakeup) {
2564 event->pending_wakeup = 0;
2565 perf_event_wakeup(event);
2569 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2571 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2575 static void perf_pending_queue(struct perf_pending_entry *entry,
2576 void (*func)(struct perf_pending_entry *))
2578 struct perf_pending_entry **head;
2580 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2585 head = &get_cpu_var(perf_pending_head);
2588 entry->next = *head;
2589 } while (cmpxchg(head, entry->next, entry) != entry->next);
2591 set_perf_event_pending();
2593 put_cpu_var(perf_pending_head);
2596 static int __perf_pending_run(void)
2598 struct perf_pending_entry *list;
2601 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2602 while (list != PENDING_TAIL) {
2603 void (*func)(struct perf_pending_entry *);
2604 struct perf_pending_entry *entry = list;
2611 * Ensure we observe the unqueue before we issue the wakeup,
2612 * so that we won't be waiting forever.
2613 * -- see perf_not_pending().
2624 static inline int perf_not_pending(struct perf_event *event)
2627 * If we flush on whatever cpu we run, there is a chance we don't
2631 __perf_pending_run();
2635 * Ensure we see the proper queue state before going to sleep
2636 * so that we do not miss the wakeup. -- see perf_pending_handle()
2639 return event->pending.next == NULL;
2642 static void perf_pending_sync(struct perf_event *event)
2644 wait_event(event->waitq, perf_not_pending(event));
2647 void perf_event_do_pending(void)
2649 __perf_pending_run();
2653 * Callchain support -- arch specific
2656 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2664 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2665 unsigned long offset, unsigned long head)
2669 if (!data->writable)
2672 mask = perf_data_size(data) - 1;
2674 offset = (offset - tail) & mask;
2675 head = (head - tail) & mask;
2677 if ((int)(head - offset) < 0)
2683 static void perf_output_wakeup(struct perf_output_handle *handle)
2685 atomic_set(&handle->data->poll, POLL_IN);
2688 handle->event->pending_wakeup = 1;
2689 perf_pending_queue(&handle->event->pending,
2690 perf_pending_event);
2692 perf_event_wakeup(handle->event);
2696 * Curious locking construct.
2698 * We need to ensure a later event_id doesn't publish a head when a former
2699 * event_id isn't done writing. However since we need to deal with NMIs we
2700 * cannot fully serialize things.
2702 * What we do is serialize between CPUs so we only have to deal with NMI
2703 * nesting on a single CPU.
2705 * We only publish the head (and generate a wakeup) when the outer-most
2706 * event_id completes.
2708 static void perf_output_lock(struct perf_output_handle *handle)
2710 struct perf_mmap_data *data = handle->data;
2711 int cur, cpu = get_cpu();
2716 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2728 static void perf_output_unlock(struct perf_output_handle *handle)
2730 struct perf_mmap_data *data = handle->data;
2734 data->done_head = data->head;
2736 if (!handle->locked)
2741 * The xchg implies a full barrier that ensures all writes are done
2742 * before we publish the new head, matched by a rmb() in userspace when
2743 * reading this position.
2745 while ((head = atomic_long_xchg(&data->done_head, 0)))
2746 data->user_page->data_head = head;
2749 * NMI can happen here, which means we can miss a done_head update.
2752 cpu = atomic_xchg(&data->lock, -1);
2753 WARN_ON_ONCE(cpu != smp_processor_id());
2756 * Therefore we have to validate we did not indeed do so.
2758 if (unlikely(atomic_long_read(&data->done_head))) {
2760 * Since we had it locked, we can lock it again.
2762 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2768 if (atomic_xchg(&data->wakeup, 0))
2769 perf_output_wakeup(handle);
2774 void perf_output_copy(struct perf_output_handle *handle,
2775 const void *buf, unsigned int len)
2777 unsigned int pages_mask;
2778 unsigned long offset;
2782 offset = handle->offset;
2783 pages_mask = handle->data->nr_pages - 1;
2784 pages = handle->data->data_pages;
2787 unsigned long page_offset;
2788 unsigned long page_size;
2791 nr = (offset >> PAGE_SHIFT) & pages_mask;
2792 page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2793 page_offset = offset & (page_size - 1);
2794 size = min_t(unsigned int, page_size - page_offset, len);
2796 memcpy(pages[nr] + page_offset, buf, size);
2803 handle->offset = offset;
2806 * Check we didn't copy past our reservation window, taking the
2807 * possible unsigned int wrap into account.
2809 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2812 int perf_output_begin(struct perf_output_handle *handle,
2813 struct perf_event *event, unsigned int size,
2814 int nmi, int sample)
2816 struct perf_event *output_event;
2817 struct perf_mmap_data *data;
2818 unsigned long tail, offset, head;
2821 struct perf_event_header header;
2828 * For inherited events we send all the output towards the parent.
2831 event = event->parent;
2833 output_event = rcu_dereference(event->output);
2835 event = output_event;
2837 data = rcu_dereference(event->data);
2841 handle->data = data;
2842 handle->event = event;
2844 handle->sample = sample;
2846 if (!data->nr_pages)
2849 have_lost = atomic_read(&data->lost);
2851 size += sizeof(lost_event);
2853 perf_output_lock(handle);
2857 * Userspace could choose to issue a mb() before updating the
2858 * tail pointer. So that all reads will be completed before the
2861 tail = ACCESS_ONCE(data->user_page->data_tail);
2863 offset = head = atomic_long_read(&data->head);
2865 if (unlikely(!perf_output_space(data, tail, offset, head)))
2867 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2869 handle->offset = offset;
2870 handle->head = head;
2872 if (head - tail > data->watermark)
2873 atomic_set(&data->wakeup, 1);
2876 lost_event.header.type = PERF_RECORD_LOST;
2877 lost_event.header.misc = 0;
2878 lost_event.header.size = sizeof(lost_event);
2879 lost_event.id = event->id;
2880 lost_event.lost = atomic_xchg(&data->lost, 0);
2882 perf_output_put(handle, lost_event);
2888 atomic_inc(&data->lost);
2889 perf_output_unlock(handle);
2896 void perf_output_end(struct perf_output_handle *handle)
2898 struct perf_event *event = handle->event;
2899 struct perf_mmap_data *data = handle->data;
2901 int wakeup_events = event->attr.wakeup_events;
2903 if (handle->sample && wakeup_events) {
2904 int events = atomic_inc_return(&data->events);
2905 if (events >= wakeup_events) {
2906 atomic_sub(wakeup_events, &data->events);
2907 atomic_set(&data->wakeup, 1);
2911 perf_output_unlock(handle);
2915 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2918 * only top level events have the pid namespace they were created in
2921 event = event->parent;
2923 return task_tgid_nr_ns(p, event->ns);
2926 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2929 * only top level events have the pid namespace they were created in
2932 event = event->parent;
2934 return task_pid_nr_ns(p, event->ns);
2937 static void perf_output_read_one(struct perf_output_handle *handle,
2938 struct perf_event *event)
2940 u64 read_format = event->attr.read_format;
2944 values[n++] = atomic64_read(&event->count);
2945 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2946 values[n++] = event->total_time_enabled +
2947 atomic64_read(&event->child_total_time_enabled);
2949 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2950 values[n++] = event->total_time_running +
2951 atomic64_read(&event->child_total_time_running);
2953 if (read_format & PERF_FORMAT_ID)
2954 values[n++] = primary_event_id(event);
2956 perf_output_copy(handle, values, n * sizeof(u64));
2960 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2962 static void perf_output_read_group(struct perf_output_handle *handle,
2963 struct perf_event *event)
2965 struct perf_event *leader = event->group_leader, *sub;
2966 u64 read_format = event->attr.read_format;
2970 values[n++] = 1 + leader->nr_siblings;
2972 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2973 values[n++] = leader->total_time_enabled;
2975 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2976 values[n++] = leader->total_time_running;
2978 if (leader != event)
2979 leader->pmu->read(leader);
2981 values[n++] = atomic64_read(&leader->count);
2982 if (read_format & PERF_FORMAT_ID)
2983 values[n++] = primary_event_id(leader);
2985 perf_output_copy(handle, values, n * sizeof(u64));
2987 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2991 sub->pmu->read(sub);
2993 values[n++] = atomic64_read(&sub->count);
2994 if (read_format & PERF_FORMAT_ID)
2995 values[n++] = primary_event_id(sub);
2997 perf_output_copy(handle, values, n * sizeof(u64));
3001 static void perf_output_read(struct perf_output_handle *handle,
3002 struct perf_event *event)
3004 if (event->attr.read_format & PERF_FORMAT_GROUP)
3005 perf_output_read_group(handle, event);
3007 perf_output_read_one(handle, event);
3010 void perf_output_sample(struct perf_output_handle *handle,
3011 struct perf_event_header *header,
3012 struct perf_sample_data *data,
3013 struct perf_event *event)
3015 u64 sample_type = data->type;
3017 perf_output_put(handle, *header);
3019 if (sample_type & PERF_SAMPLE_IP)
3020 perf_output_put(handle, data->ip);
3022 if (sample_type & PERF_SAMPLE_TID)
3023 perf_output_put(handle, data->tid_entry);
3025 if (sample_type & PERF_SAMPLE_TIME)
3026 perf_output_put(handle, data->time);
3028 if (sample_type & PERF_SAMPLE_ADDR)
3029 perf_output_put(handle, data->addr);
3031 if (sample_type & PERF_SAMPLE_ID)
3032 perf_output_put(handle, data->id);
3034 if (sample_type & PERF_SAMPLE_STREAM_ID)
3035 perf_output_put(handle, data->stream_id);
3037 if (sample_type & PERF_SAMPLE_CPU)
3038 perf_output_put(handle, data->cpu_entry);
3040 if (sample_type & PERF_SAMPLE_PERIOD)
3041 perf_output_put(handle, data->period);
3043 if (sample_type & PERF_SAMPLE_READ)
3044 perf_output_read(handle, event);
3046 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3047 if (data->callchain) {
3050 if (data->callchain)
3051 size += data->callchain->nr;
3053 size *= sizeof(u64);
3055 perf_output_copy(handle, data->callchain, size);
3058 perf_output_put(handle, nr);
3062 if (sample_type & PERF_SAMPLE_RAW) {
3064 perf_output_put(handle, data->raw->size);
3065 perf_output_copy(handle, data->raw->data,
3072 .size = sizeof(u32),
3075 perf_output_put(handle, raw);
3080 void perf_prepare_sample(struct perf_event_header *header,
3081 struct perf_sample_data *data,
3082 struct perf_event *event,
3083 struct pt_regs *regs)
3085 u64 sample_type = event->attr.sample_type;
3087 data->type = sample_type;
3089 header->type = PERF_RECORD_SAMPLE;
3090 header->size = sizeof(*header);
3093 header->misc |= perf_misc_flags(regs);
3095 if (sample_type & PERF_SAMPLE_IP) {
3096 data->ip = perf_instruction_pointer(regs);
3098 header->size += sizeof(data->ip);
3101 if (sample_type & PERF_SAMPLE_TID) {
3102 /* namespace issues */
3103 data->tid_entry.pid = perf_event_pid(event, current);
3104 data->tid_entry.tid = perf_event_tid(event, current);
3106 header->size += sizeof(data->tid_entry);
3109 if (sample_type & PERF_SAMPLE_TIME) {
3110 data->time = perf_clock();
3112 header->size += sizeof(data->time);
3115 if (sample_type & PERF_SAMPLE_ADDR)
3116 header->size += sizeof(data->addr);
3118 if (sample_type & PERF_SAMPLE_ID) {
3119 data->id = primary_event_id(event);
3121 header->size += sizeof(data->id);
3124 if (sample_type & PERF_SAMPLE_STREAM_ID) {
3125 data->stream_id = event->id;
3127 header->size += sizeof(data->stream_id);
3130 if (sample_type & PERF_SAMPLE_CPU) {
3131 data->cpu_entry.cpu = raw_smp_processor_id();
3132 data->cpu_entry.reserved = 0;
3134 header->size += sizeof(data->cpu_entry);
3137 if (sample_type & PERF_SAMPLE_PERIOD)
3138 header->size += sizeof(data->period);
3140 if (sample_type & PERF_SAMPLE_READ)
3141 header->size += perf_event_read_size(event);
3143 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3146 data->callchain = perf_callchain(regs);
3148 if (data->callchain)
3149 size += data->callchain->nr;
3151 header->size += size * sizeof(u64);
3154 if (sample_type & PERF_SAMPLE_RAW) {
3155 int size = sizeof(u32);
3158 size += data->raw->size;
3160 size += sizeof(u32);
3162 WARN_ON_ONCE(size & (sizeof(u64)-1));
3163 header->size += size;
3167 static void perf_event_output(struct perf_event *event, int nmi,
3168 struct perf_sample_data *data,
3169 struct pt_regs *regs)
3171 struct perf_output_handle handle;
3172 struct perf_event_header header;
3174 perf_prepare_sample(&header, data, event, regs);
3176 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3179 perf_output_sample(&handle, &header, data, event);
3181 perf_output_end(&handle);
3188 struct perf_read_event {
3189 struct perf_event_header header;
3196 perf_event_read_event(struct perf_event *event,
3197 struct task_struct *task)
3199 struct perf_output_handle handle;
3200 struct perf_read_event read_event = {
3202 .type = PERF_RECORD_READ,
3204 .size = sizeof(read_event) + perf_event_read_size(event),
3206 .pid = perf_event_pid(event, task),
3207 .tid = perf_event_tid(event, task),
3211 ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3215 perf_output_put(&handle, read_event);
3216 perf_output_read(&handle, event);
3218 perf_output_end(&handle);
3222 * task tracking -- fork/exit
3224 * enabled by: attr.comm | attr.mmap | attr.task
3227 struct perf_task_event {
3228 struct task_struct *task;
3229 struct perf_event_context *task_ctx;
3232 struct perf_event_header header;
3242 static void perf_event_task_output(struct perf_event *event,
3243 struct perf_task_event *task_event)
3245 struct perf_output_handle handle;
3247 struct task_struct *task = task_event->task;
3250 size = task_event->event_id.header.size;
3251 ret = perf_output_begin(&handle, event, size, 0, 0);
3256 task_event->event_id.pid = perf_event_pid(event, task);
3257 task_event->event_id.ppid = perf_event_pid(event, current);
3259 task_event->event_id.tid = perf_event_tid(event, task);
3260 task_event->event_id.ptid = perf_event_tid(event, current);
3262 task_event->event_id.time = perf_clock();
3264 perf_output_put(&handle, task_event->event_id);
3266 perf_output_end(&handle);
3269 static int perf_event_task_match(struct perf_event *event)
3271 if (event->state != PERF_EVENT_STATE_ACTIVE)
3274 if (event->cpu != -1 && event->cpu != smp_processor_id())
3277 if (event->attr.comm || event->attr.mmap || event->attr.task)
3283 static void perf_event_task_ctx(struct perf_event_context *ctx,
3284 struct perf_task_event *task_event)
3286 struct perf_event *event;
3288 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3289 if (perf_event_task_match(event))
3290 perf_event_task_output(event, task_event);
3294 static void perf_event_task_event(struct perf_task_event *task_event)
3296 struct perf_cpu_context *cpuctx;
3297 struct perf_event_context *ctx = task_event->task_ctx;
3300 cpuctx = &get_cpu_var(perf_cpu_context);
3301 perf_event_task_ctx(&cpuctx->ctx, task_event);
3303 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3305 perf_event_task_ctx(ctx, task_event);
3306 put_cpu_var(perf_cpu_context);
3310 static void perf_event_task(struct task_struct *task,
3311 struct perf_event_context *task_ctx,
3314 struct perf_task_event task_event;
3316 if (!atomic_read(&nr_comm_events) &&
3317 !atomic_read(&nr_mmap_events) &&
3318 !atomic_read(&nr_task_events))
3321 task_event = (struct perf_task_event){
3323 .task_ctx = task_ctx,
3326 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3328 .size = sizeof(task_event.event_id),
3337 perf_event_task_event(&task_event);
3340 void perf_event_fork(struct task_struct *task)
3342 perf_event_task(task, NULL, 1);
3349 struct perf_comm_event {
3350 struct task_struct *task;
3355 struct perf_event_header header;
3362 static void perf_event_comm_output(struct perf_event *event,
3363 struct perf_comm_event *comm_event)
3365 struct perf_output_handle handle;
3366 int size = comm_event->event_id.header.size;
3367 int ret = perf_output_begin(&handle, event, size, 0, 0);
3372 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3373 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3375 perf_output_put(&handle, comm_event->event_id);
3376 perf_output_copy(&handle, comm_event->comm,
3377 comm_event->comm_size);
3378 perf_output_end(&handle);
3381 static int perf_event_comm_match(struct perf_event *event)
3383 if (event->state != PERF_EVENT_STATE_ACTIVE)
3386 if (event->cpu != -1 && event->cpu != smp_processor_id())
3389 if (event->attr.comm)
3395 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3396 struct perf_comm_event *comm_event)
3398 struct perf_event *event;
3400 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3401 if (perf_event_comm_match(event))
3402 perf_event_comm_output(event, comm_event);
3406 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3408 struct perf_cpu_context *cpuctx;
3409 struct perf_event_context *ctx;
3411 char comm[TASK_COMM_LEN];
3413 memset(comm, 0, sizeof(comm));
3414 strlcpy(comm, comm_event->task->comm, sizeof(comm));
3415 size = ALIGN(strlen(comm)+1, sizeof(u64));
3417 comm_event->comm = comm;
3418 comm_event->comm_size = size;
3420 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3423 cpuctx = &get_cpu_var(perf_cpu_context);
3424 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3425 ctx = rcu_dereference(current->perf_event_ctxp);
3427 perf_event_comm_ctx(ctx, comm_event);
3428 put_cpu_var(perf_cpu_context);
3432 void perf_event_comm(struct task_struct *task)
3434 struct perf_comm_event comm_event;
3436 if (task->perf_event_ctxp)
3437 perf_event_enable_on_exec(task);
3439 if (!atomic_read(&nr_comm_events))
3442 comm_event = (struct perf_comm_event){
3448 .type = PERF_RECORD_COMM,
3457 perf_event_comm_event(&comm_event);
3464 struct perf_mmap_event {
3465 struct vm_area_struct *vma;
3467 const char *file_name;
3471 struct perf_event_header header;
3481 static void perf_event_mmap_output(struct perf_event *event,
3482 struct perf_mmap_event *mmap_event)
3484 struct perf_output_handle handle;
3485 int size = mmap_event->event_id.header.size;
3486 int ret = perf_output_begin(&handle, event, size, 0, 0);
3491 mmap_event->event_id.pid = perf_event_pid(event, current);
3492 mmap_event->event_id.tid = perf_event_tid(event, current);
3494 perf_output_put(&handle, mmap_event->event_id);
3495 perf_output_copy(&handle, mmap_event->file_name,
3496 mmap_event->file_size);
3497 perf_output_end(&handle);
3500 static int perf_event_mmap_match(struct perf_event *event,
3501 struct perf_mmap_event *mmap_event)
3503 if (event->state != PERF_EVENT_STATE_ACTIVE)
3506 if (event->cpu != -1 && event->cpu != smp_processor_id())
3509 if (event->attr.mmap)
3515 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3516 struct perf_mmap_event *mmap_event)
3518 struct perf_event *event;
3520 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3521 if (perf_event_mmap_match(event, mmap_event))
3522 perf_event_mmap_output(event, mmap_event);
3526 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3528 struct perf_cpu_context *cpuctx;
3529 struct perf_event_context *ctx;
3530 struct vm_area_struct *vma = mmap_event->vma;
3531 struct file *file = vma->vm_file;
3537 memset(tmp, 0, sizeof(tmp));
3541 * d_path works from the end of the buffer backwards, so we
3542 * need to add enough zero bytes after the string to handle
3543 * the 64bit alignment we do later.
3545 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3547 name = strncpy(tmp, "//enomem", sizeof(tmp));
3550 name = d_path(&file->f_path, buf, PATH_MAX);
3552 name = strncpy(tmp, "//toolong", sizeof(tmp));
3556 if (arch_vma_name(mmap_event->vma)) {
3557 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3563 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3567 name = strncpy(tmp, "//anon", sizeof(tmp));
3572 size = ALIGN(strlen(name)+1, sizeof(u64));
3574 mmap_event->file_name = name;
3575 mmap_event->file_size = size;
3577 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3580 cpuctx = &get_cpu_var(perf_cpu_context);
3581 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3582 ctx = rcu_dereference(current->perf_event_ctxp);
3584 perf_event_mmap_ctx(ctx, mmap_event);
3585 put_cpu_var(perf_cpu_context);
3591 void __perf_event_mmap(struct vm_area_struct *vma)
3593 struct perf_mmap_event mmap_event;
3595 if (!atomic_read(&nr_mmap_events))
3598 mmap_event = (struct perf_mmap_event){
3604 .type = PERF_RECORD_MMAP,
3610 .start = vma->vm_start,
3611 .len = vma->vm_end - vma->vm_start,
3612 .pgoff = vma->vm_pgoff,
3616 perf_event_mmap_event(&mmap_event);
3620 * IRQ throttle logging
3623 static void perf_log_throttle(struct perf_event *event, int enable)
3625 struct perf_output_handle handle;
3629 struct perf_event_header header;
3633 } throttle_event = {
3635 .type = PERF_RECORD_THROTTLE,
3637 .size = sizeof(throttle_event),
3639 .time = perf_clock(),
3640 .id = primary_event_id(event),
3641 .stream_id = event->id,
3645 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3647 ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3651 perf_output_put(&handle, throttle_event);
3652 perf_output_end(&handle);
3656 * Generic event overflow handling, sampling.
3659 static int __perf_event_overflow(struct perf_event *event, int nmi,
3660 int throttle, struct perf_sample_data *data,
3661 struct pt_regs *regs)
3663 int events = atomic_read(&event->event_limit);
3664 struct hw_perf_event *hwc = &event->hw;
3667 throttle = (throttle && event->pmu->unthrottle != NULL);
3672 if (hwc->interrupts != MAX_INTERRUPTS) {
3674 if (HZ * hwc->interrupts >
3675 (u64)sysctl_perf_event_sample_rate) {
3676 hwc->interrupts = MAX_INTERRUPTS;
3677 perf_log_throttle(event, 0);
3682 * Keep re-disabling events even though on the previous
3683 * pass we disabled it - just in case we raced with a
3684 * sched-in and the event got enabled again:
3690 if (event->attr.freq) {
3691 u64 now = perf_clock();
3692 s64 delta = now - hwc->freq_stamp;
3694 hwc->freq_stamp = now;
3696 if (delta > 0 && delta < TICK_NSEC)
3697 perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3701 * XXX event_limit might not quite work as expected on inherited
3705 event->pending_kill = POLL_IN;
3706 if (events && atomic_dec_and_test(&event->event_limit)) {
3708 event->pending_kill = POLL_HUP;
3710 event->pending_disable = 1;
3711 perf_pending_queue(&event->pending,
3712 perf_pending_event);
3714 perf_event_disable(event);
3717 if (event->overflow_handler)
3718 event->overflow_handler(event, nmi, data, regs);
3720 perf_event_output(event, nmi, data, regs);
3725 int perf_event_overflow(struct perf_event *event, int nmi,
3726 struct perf_sample_data *data,
3727 struct pt_regs *regs)
3729 return __perf_event_overflow(event, nmi, 1, data, regs);
3733 * Generic software event infrastructure
3737 * We directly increment event->count and keep a second value in
3738 * event->hw.period_left to count intervals. This period event
3739 * is kept in the range [-sample_period, 0] so that we can use the
3743 static u64 perf_swevent_set_period(struct perf_event *event)
3745 struct hw_perf_event *hwc = &event->hw;
3746 u64 period = hwc->last_period;
3750 hwc->last_period = hwc->sample_period;
3753 old = val = atomic64_read(&hwc->period_left);
3757 nr = div64_u64(period + val, period);
3758 offset = nr * period;
3760 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3766 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3767 int nmi, struct perf_sample_data *data,
3768 struct pt_regs *regs)
3770 struct hw_perf_event *hwc = &event->hw;
3773 data->period = event->hw.last_period;
3775 overflow = perf_swevent_set_period(event);
3777 if (hwc->interrupts == MAX_INTERRUPTS)
3780 for (; overflow; overflow--) {
3781 if (__perf_event_overflow(event, nmi, throttle,
3784 * We inhibit the overflow from happening when
3785 * hwc->interrupts == MAX_INTERRUPTS.
3793 static void perf_swevent_unthrottle(struct perf_event *event)
3796 * Nothing to do, we already reset hwc->interrupts.
3800 static void perf_swevent_add(struct perf_event *event, u64 nr,
3801 int nmi, struct perf_sample_data *data,
3802 struct pt_regs *regs)
3804 struct hw_perf_event *hwc = &event->hw;
3806 atomic64_add(nr, &event->count);
3811 if (!hwc->sample_period)
3814 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3815 return perf_swevent_overflow(event, 1, nmi, data, regs);
3817 if (atomic64_add_negative(nr, &hwc->period_left))
3820 perf_swevent_overflow(event, 0, nmi, data, regs);
3823 static int perf_swevent_is_counting(struct perf_event *event)
3826 * The event is active, we're good!
3828 if (event->state == PERF_EVENT_STATE_ACTIVE)
3832 * The event is off/error, not counting.
3834 if (event->state != PERF_EVENT_STATE_INACTIVE)
3838 * The event is inactive, if the context is active
3839 * we're part of a group that didn't make it on the 'pmu',
3842 if (event->ctx->is_active)
3846 * We're inactive and the context is too, this means the
3847 * task is scheduled out, we're counting events that happen
3848 * to us, like migration events.
3853 static int perf_tp_event_match(struct perf_event *event,
3854 struct perf_sample_data *data);
3856 static int perf_exclude_event(struct perf_event *event,
3857 struct pt_regs *regs)
3860 if (event->attr.exclude_user && user_mode(regs))
3863 if (event->attr.exclude_kernel && !user_mode(regs))
3870 static int perf_swevent_match(struct perf_event *event,
3871 enum perf_type_id type,
3873 struct perf_sample_data *data,
3874 struct pt_regs *regs)
3876 if (event->cpu != -1 && event->cpu != smp_processor_id())
3879 if (!perf_swevent_is_counting(event))
3882 if (event->attr.type != type)
3885 if (event->attr.config != event_id)
3888 if (perf_exclude_event(event, regs))
3891 if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3892 !perf_tp_event_match(event, data))
3898 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3899 enum perf_type_id type,
3900 u32 event_id, u64 nr, int nmi,
3901 struct perf_sample_data *data,
3902 struct pt_regs *regs)
3904 struct perf_event *event;
3906 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3907 if (perf_swevent_match(event, type, event_id, data, regs))
3908 perf_swevent_add(event, nr, nmi, data, regs);
3912 int perf_swevent_get_recursion_context(void)
3914 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3921 else if (in_softirq())
3926 if (cpuctx->recursion[rctx]) {
3927 put_cpu_var(perf_cpu_context);
3931 cpuctx->recursion[rctx]++;
3936 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3938 void perf_swevent_put_recursion_context(int rctx)
3940 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3942 cpuctx->recursion[rctx]--;
3943 put_cpu_var(perf_cpu_context);
3945 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3947 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3949 struct perf_sample_data *data,
3950 struct pt_regs *regs)
3952 struct perf_cpu_context *cpuctx;
3953 struct perf_event_context *ctx;
3955 cpuctx = &__get_cpu_var(perf_cpu_context);
3957 perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3958 nr, nmi, data, regs);
3960 * doesn't really matter which of the child contexts the
3961 * events ends up in.
3963 ctx = rcu_dereference(current->perf_event_ctxp);
3965 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3969 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3970 struct pt_regs *regs, u64 addr)
3972 struct perf_sample_data data;
3975 rctx = perf_swevent_get_recursion_context();
3982 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
3984 perf_swevent_put_recursion_context(rctx);
3987 static void perf_swevent_read(struct perf_event *event)
3991 static int perf_swevent_enable(struct perf_event *event)
3993 struct hw_perf_event *hwc = &event->hw;
3995 if (hwc->sample_period) {
3996 hwc->last_period = hwc->sample_period;
3997 perf_swevent_set_period(event);
4002 static void perf_swevent_disable(struct perf_event *event)
4006 static const struct pmu perf_ops_generic = {
4007 .enable = perf_swevent_enable,
4008 .disable = perf_swevent_disable,
4009 .read = perf_swevent_read,
4010 .unthrottle = perf_swevent_unthrottle,
4014 * hrtimer based swevent callback
4017 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4019 enum hrtimer_restart ret = HRTIMER_RESTART;
4020 struct perf_sample_data data;
4021 struct pt_regs *regs;
4022 struct perf_event *event;
4025 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4026 event->pmu->read(event);
4030 data.period = event->hw.last_period;
4031 regs = get_irq_regs();
4033 * In case we exclude kernel IPs or are somehow not in interrupt
4034 * context, provide the next best thing, the user IP.
4036 if ((event->attr.exclude_kernel || !regs) &&
4037 !event->attr.exclude_user)
4038 regs = task_pt_regs(current);
4041 if (!(event->attr.exclude_idle && current->pid == 0))
4042 if (perf_event_overflow(event, 0, &data, regs))
4043 ret = HRTIMER_NORESTART;
4046 period = max_t(u64, 10000, event->hw.sample_period);
4047 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4052 static void perf_swevent_start_hrtimer(struct perf_event *event)
4054 struct hw_perf_event *hwc = &event->hw;
4056 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4057 hwc->hrtimer.function = perf_swevent_hrtimer;
4058 if (hwc->sample_period) {
4061 if (hwc->remaining) {
4062 if (hwc->remaining < 0)
4065 period = hwc->remaining;
4068 period = max_t(u64, 10000, hwc->sample_period);
4070 __hrtimer_start_range_ns(&hwc->hrtimer,
4071 ns_to_ktime(period), 0,
4072 HRTIMER_MODE_REL, 0);
4076 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4078 struct hw_perf_event *hwc = &event->hw;
4080 if (hwc->sample_period) {
4081 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4082 hwc->remaining = ktime_to_ns(remaining);
4084 hrtimer_cancel(&hwc->hrtimer);
4089 * Software event: cpu wall time clock
4092 static void cpu_clock_perf_event_update(struct perf_event *event)
4094 int cpu = raw_smp_processor_id();
4098 now = cpu_clock(cpu);
4099 prev = atomic64_xchg(&event->hw.prev_count, now);
4100 atomic64_add(now - prev, &event->count);
4103 static int cpu_clock_perf_event_enable(struct perf_event *event)
4105 struct hw_perf_event *hwc = &event->hw;
4106 int cpu = raw_smp_processor_id();
4108 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4109 perf_swevent_start_hrtimer(event);
4114 static void cpu_clock_perf_event_disable(struct perf_event *event)
4116 perf_swevent_cancel_hrtimer(event);
4117 cpu_clock_perf_event_update(event);
4120 static void cpu_clock_perf_event_read(struct perf_event *event)
4122 cpu_clock_perf_event_update(event);
4125 static const struct pmu perf_ops_cpu_clock = {
4126 .enable = cpu_clock_perf_event_enable,
4127 .disable = cpu_clock_perf_event_disable,
4128 .read = cpu_clock_perf_event_read,
4132 * Software event: task time clock
4135 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4140 prev = atomic64_xchg(&event->hw.prev_count, now);
4142 atomic64_add(delta, &event->count);
4145 static int task_clock_perf_event_enable(struct perf_event *event)
4147 struct hw_perf_event *hwc = &event->hw;
4150 now = event->ctx->time;
4152 atomic64_set(&hwc->prev_count, now);
4154 perf_swevent_start_hrtimer(event);
4159 static void task_clock_perf_event_disable(struct perf_event *event)
4161 perf_swevent_cancel_hrtimer(event);
4162 task_clock_perf_event_update(event, event->ctx->time);
4166 static void task_clock_perf_event_read(struct perf_event *event)
4171 update_context_time(event->ctx);
4172 time = event->ctx->time;
4174 u64 now = perf_clock();
4175 u64 delta = now - event->ctx->timestamp;
4176 time = event->ctx->time + delta;
4179 task_clock_perf_event_update(event, time);
4182 static const struct pmu perf_ops_task_clock = {
4183 .enable = task_clock_perf_event_enable,
4184 .disable = task_clock_perf_event_disable,
4185 .read = task_clock_perf_event_read,
4188 #ifdef CONFIG_EVENT_PROFILE
4190 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4193 struct perf_raw_record raw = {
4198 struct perf_sample_data data = {
4203 struct pt_regs *regs = get_irq_regs();
4206 regs = task_pt_regs(current);
4208 /* Trace events already protected against recursion */
4209 do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4212 EXPORT_SYMBOL_GPL(perf_tp_event);
4214 static int perf_tp_event_match(struct perf_event *event,
4215 struct perf_sample_data *data)
4217 void *record = data->raw->data;
4219 if (likely(!event->filter) || filter_match_preds(event->filter, record))
4224 static void tp_perf_event_destroy(struct perf_event *event)
4226 ftrace_profile_disable(event->attr.config);
4229 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4232 * Raw tracepoint data is a severe data leak, only allow root to
4235 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4236 perf_paranoid_tracepoint_raw() &&
4237 !capable(CAP_SYS_ADMIN))
4238 return ERR_PTR(-EPERM);
4240 if (ftrace_profile_enable(event->attr.config))
4243 event->destroy = tp_perf_event_destroy;
4245 return &perf_ops_generic;
4248 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4253 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4256 filter_str = strndup_user(arg, PAGE_SIZE);
4257 if (IS_ERR(filter_str))
4258 return PTR_ERR(filter_str);
4260 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4266 static void perf_event_free_filter(struct perf_event *event)
4268 ftrace_profile_free_filter(event);
4273 static int perf_tp_event_match(struct perf_event *event,
4274 struct perf_sample_data *data)
4279 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4284 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4289 static void perf_event_free_filter(struct perf_event *event)
4293 #endif /* CONFIG_EVENT_PROFILE */
4295 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4296 static void bp_perf_event_destroy(struct perf_event *event)
4298 release_bp_slot(event);
4301 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4305 err = register_perf_hw_breakpoint(bp);
4307 return ERR_PTR(err);
4309 bp->destroy = bp_perf_event_destroy;
4311 return &perf_ops_bp;
4314 void perf_bp_event(struct perf_event *bp, void *data)
4316 struct perf_sample_data sample;
4317 struct pt_regs *regs = data;
4320 sample.addr = bp->attr.bp_addr;
4322 if (!perf_exclude_event(bp, regs))
4323 perf_swevent_add(bp, 1, 1, &sample, regs);
4326 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4331 void perf_bp_event(struct perf_event *bp, void *regs)
4336 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4338 static void sw_perf_event_destroy(struct perf_event *event)
4340 u64 event_id = event->attr.config;
4342 WARN_ON(event->parent);
4344 atomic_dec(&perf_swevent_enabled[event_id]);
4347 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4349 const struct pmu *pmu = NULL;
4350 u64 event_id = event->attr.config;
4353 * Software events (currently) can't in general distinguish
4354 * between user, kernel and hypervisor events.
4355 * However, context switches and cpu migrations are considered
4356 * to be kernel events, and page faults are never hypervisor
4360 case PERF_COUNT_SW_CPU_CLOCK:
4361 pmu = &perf_ops_cpu_clock;
4364 case PERF_COUNT_SW_TASK_CLOCK:
4366 * If the user instantiates this as a per-cpu event,
4367 * use the cpu_clock event instead.
4369 if (event->ctx->task)
4370 pmu = &perf_ops_task_clock;
4372 pmu = &perf_ops_cpu_clock;
4375 case PERF_COUNT_SW_PAGE_FAULTS:
4376 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4377 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4378 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4379 case PERF_COUNT_SW_CPU_MIGRATIONS:
4380 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4381 case PERF_COUNT_SW_EMULATION_FAULTS:
4382 if (!event->parent) {
4383 atomic_inc(&perf_swevent_enabled[event_id]);
4384 event->destroy = sw_perf_event_destroy;
4386 pmu = &perf_ops_generic;
4394 * Allocate and initialize a event structure
4396 static struct perf_event *
4397 perf_event_alloc(struct perf_event_attr *attr,
4399 struct perf_event_context *ctx,
4400 struct perf_event *group_leader,
4401 struct perf_event *parent_event,
4402 perf_overflow_handler_t overflow_handler,
4405 const struct pmu *pmu;
4406 struct perf_event *event;
4407 struct hw_perf_event *hwc;
4410 event = kzalloc(sizeof(*event), gfpflags);
4412 return ERR_PTR(-ENOMEM);
4415 * Single events are their own group leaders, with an
4416 * empty sibling list:
4419 group_leader = event;
4421 mutex_init(&event->child_mutex);
4422 INIT_LIST_HEAD(&event->child_list);
4424 INIT_LIST_HEAD(&event->group_entry);
4425 INIT_LIST_HEAD(&event->event_entry);
4426 INIT_LIST_HEAD(&event->sibling_list);
4427 init_waitqueue_head(&event->waitq);
4429 mutex_init(&event->mmap_mutex);
4432 event->attr = *attr;
4433 event->group_leader = group_leader;
4438 event->parent = parent_event;
4440 event->ns = get_pid_ns(current->nsproxy->pid_ns);
4441 event->id = atomic64_inc_return(&perf_event_id);
4443 event->state = PERF_EVENT_STATE_INACTIVE;
4445 if (!overflow_handler && parent_event)
4446 overflow_handler = parent_event->overflow_handler;
4448 event->overflow_handler = overflow_handler;
4451 event->state = PERF_EVENT_STATE_OFF;
4456 hwc->sample_period = attr->sample_period;
4457 if (attr->freq && attr->sample_freq)
4458 hwc->sample_period = 1;
4459 hwc->last_period = hwc->sample_period;
4461 atomic64_set(&hwc->period_left, hwc->sample_period);
4464 * we currently do not support PERF_FORMAT_GROUP on inherited events
4466 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4469 switch (attr->type) {
4471 case PERF_TYPE_HARDWARE:
4472 case PERF_TYPE_HW_CACHE:
4473 pmu = hw_perf_event_init(event);
4476 case PERF_TYPE_SOFTWARE:
4477 pmu = sw_perf_event_init(event);
4480 case PERF_TYPE_TRACEPOINT:
4481 pmu = tp_perf_event_init(event);
4484 case PERF_TYPE_BREAKPOINT:
4485 pmu = bp_perf_event_init(event);
4496 else if (IS_ERR(pmu))
4501 put_pid_ns(event->ns);
4503 return ERR_PTR(err);
4508 if (!event->parent) {
4509 atomic_inc(&nr_events);
4510 if (event->attr.mmap)
4511 atomic_inc(&nr_mmap_events);
4512 if (event->attr.comm)
4513 atomic_inc(&nr_comm_events);
4514 if (event->attr.task)
4515 atomic_inc(&nr_task_events);
4521 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4522 struct perf_event_attr *attr)
4527 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4531 * zero the full structure, so that a short copy will be nice.
4533 memset(attr, 0, sizeof(*attr));
4535 ret = get_user(size, &uattr->size);
4539 if (size > PAGE_SIZE) /* silly large */
4542 if (!size) /* abi compat */
4543 size = PERF_ATTR_SIZE_VER0;
4545 if (size < PERF_ATTR_SIZE_VER0)
4549 * If we're handed a bigger struct than we know of,
4550 * ensure all the unknown bits are 0 - i.e. new
4551 * user-space does not rely on any kernel feature
4552 * extensions we dont know about yet.
4554 if (size > sizeof(*attr)) {
4555 unsigned char __user *addr;
4556 unsigned char __user *end;
4559 addr = (void __user *)uattr + sizeof(*attr);
4560 end = (void __user *)uattr + size;
4562 for (; addr < end; addr++) {
4563 ret = get_user(val, addr);
4569 size = sizeof(*attr);
4572 ret = copy_from_user(attr, uattr, size);
4577 * If the type exists, the corresponding creation will verify
4580 if (attr->type >= PERF_TYPE_MAX)
4583 if (attr->__reserved_1 || attr->__reserved_2)
4586 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4589 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4596 put_user(sizeof(*attr), &uattr->size);
4601 static int perf_event_set_output(struct perf_event *event, int output_fd)
4603 struct perf_event *output_event = NULL;
4604 struct file *output_file = NULL;
4605 struct perf_event *old_output;
4606 int fput_needed = 0;
4612 output_file = fget_light(output_fd, &fput_needed);
4616 if (output_file->f_op != &perf_fops)
4619 output_event = output_file->private_data;
4621 /* Don't chain output fds */
4622 if (output_event->output)
4625 /* Don't set an output fd when we already have an output channel */
4629 atomic_long_inc(&output_file->f_count);
4632 mutex_lock(&event->mmap_mutex);
4633 old_output = event->output;
4634 rcu_assign_pointer(event->output, output_event);
4635 mutex_unlock(&event->mmap_mutex);
4639 * we need to make sure no existing perf_output_*()
4640 * is still referencing this event.
4643 fput(old_output->filp);
4648 fput_light(output_file, fput_needed);
4653 * sys_perf_event_open - open a performance event, associate it to a task/cpu
4655 * @attr_uptr: event_id type attributes for monitoring/sampling
4658 * @group_fd: group leader event fd
4660 SYSCALL_DEFINE5(perf_event_open,
4661 struct perf_event_attr __user *, attr_uptr,
4662 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4664 struct perf_event *event, *group_leader;
4665 struct perf_event_attr attr;
4666 struct perf_event_context *ctx;
4667 struct file *event_file = NULL;
4668 struct file *group_file = NULL;
4669 int fput_needed = 0;
4670 int fput_needed2 = 0;
4673 /* for future expandability... */
4674 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4677 err = perf_copy_attr(attr_uptr, &attr);
4681 if (!attr.exclude_kernel) {
4682 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4687 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4692 * Get the target context (task or percpu):
4694 ctx = find_get_context(pid, cpu);
4696 return PTR_ERR(ctx);
4699 * Look up the group leader (we will attach this event to it):
4701 group_leader = NULL;
4702 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4704 group_file = fget_light(group_fd, &fput_needed);
4706 goto err_put_context;
4707 if (group_file->f_op != &perf_fops)
4708 goto err_put_context;
4710 group_leader = group_file->private_data;
4712 * Do not allow a recursive hierarchy (this new sibling
4713 * becoming part of another group-sibling):
4715 if (group_leader->group_leader != group_leader)
4716 goto err_put_context;
4718 * Do not allow to attach to a group in a different
4719 * task or CPU context:
4721 if (group_leader->ctx != ctx)
4722 goto err_put_context;
4724 * Only a group leader can be exclusive or pinned
4726 if (attr.exclusive || attr.pinned)
4727 goto err_put_context;
4730 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4731 NULL, NULL, GFP_KERNEL);
4732 err = PTR_ERR(event);
4734 goto err_put_context;
4736 err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4738 goto err_free_put_context;
4740 event_file = fget_light(err, &fput_needed2);
4742 goto err_free_put_context;
4744 if (flags & PERF_FLAG_FD_OUTPUT) {
4745 err = perf_event_set_output(event, group_fd);
4747 goto err_fput_free_put_context;
4750 event->filp = event_file;
4751 WARN_ON_ONCE(ctx->parent_ctx);
4752 mutex_lock(&ctx->mutex);
4753 perf_install_in_context(ctx, event, cpu);
4755 mutex_unlock(&ctx->mutex);
4757 event->owner = current;
4758 get_task_struct(current);
4759 mutex_lock(¤t->perf_event_mutex);
4760 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4761 mutex_unlock(¤t->perf_event_mutex);
4763 err_fput_free_put_context:
4764 fput_light(event_file, fput_needed2);
4766 err_free_put_context:
4774 fput_light(group_file, fput_needed);
4780 * perf_event_create_kernel_counter
4782 * @attr: attributes of the counter to create
4783 * @cpu: cpu in which the counter is bound
4784 * @pid: task to profile
4787 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4789 perf_overflow_handler_t overflow_handler)
4791 struct perf_event *event;
4792 struct perf_event_context *ctx;
4796 * Get the target context (task or percpu):
4799 ctx = find_get_context(pid, cpu);
4805 event = perf_event_alloc(attr, cpu, ctx, NULL,
4806 NULL, overflow_handler, GFP_KERNEL);
4807 if (IS_ERR(event)) {
4808 err = PTR_ERR(event);
4809 goto err_put_context;
4813 WARN_ON_ONCE(ctx->parent_ctx);
4814 mutex_lock(&ctx->mutex);
4815 perf_install_in_context(ctx, event, cpu);
4817 mutex_unlock(&ctx->mutex);
4819 event->owner = current;
4820 get_task_struct(current);
4821 mutex_lock(¤t->perf_event_mutex);
4822 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
4823 mutex_unlock(¤t->perf_event_mutex);
4830 return ERR_PTR(err);
4832 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4835 * inherit a event from parent task to child task:
4837 static struct perf_event *
4838 inherit_event(struct perf_event *parent_event,
4839 struct task_struct *parent,
4840 struct perf_event_context *parent_ctx,
4841 struct task_struct *child,
4842 struct perf_event *group_leader,
4843 struct perf_event_context *child_ctx)
4845 struct perf_event *child_event;
4848 * Instead of creating recursive hierarchies of events,
4849 * we link inherited events back to the original parent,
4850 * which has a filp for sure, which we use as the reference
4853 if (parent_event->parent)
4854 parent_event = parent_event->parent;
4856 child_event = perf_event_alloc(&parent_event->attr,
4857 parent_event->cpu, child_ctx,
4858 group_leader, parent_event,
4860 if (IS_ERR(child_event))
4865 * Make the child state follow the state of the parent event,
4866 * not its attr.disabled bit. We hold the parent's mutex,
4867 * so we won't race with perf_event_{en, dis}able_family.
4869 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4870 child_event->state = PERF_EVENT_STATE_INACTIVE;
4872 child_event->state = PERF_EVENT_STATE_OFF;
4874 if (parent_event->attr.freq)
4875 child_event->hw.sample_period = parent_event->hw.sample_period;
4877 child_event->overflow_handler = parent_event->overflow_handler;
4880 * Link it up in the child's context:
4882 add_event_to_ctx(child_event, child_ctx);
4885 * Get a reference to the parent filp - we will fput it
4886 * when the child event exits. This is safe to do because
4887 * we are in the parent and we know that the filp still
4888 * exists and has a nonzero count:
4890 atomic_long_inc(&parent_event->filp->f_count);
4893 * Link this into the parent event's child list
4895 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4896 mutex_lock(&parent_event->child_mutex);
4897 list_add_tail(&child_event->child_list, &parent_event->child_list);
4898 mutex_unlock(&parent_event->child_mutex);
4903 static int inherit_group(struct perf_event *parent_event,
4904 struct task_struct *parent,
4905 struct perf_event_context *parent_ctx,
4906 struct task_struct *child,
4907 struct perf_event_context *child_ctx)
4909 struct perf_event *leader;
4910 struct perf_event *sub;
4911 struct perf_event *child_ctr;
4913 leader = inherit_event(parent_event, parent, parent_ctx,
4914 child, NULL, child_ctx);
4916 return PTR_ERR(leader);
4917 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4918 child_ctr = inherit_event(sub, parent, parent_ctx,
4919 child, leader, child_ctx);
4920 if (IS_ERR(child_ctr))
4921 return PTR_ERR(child_ctr);
4926 static void sync_child_event(struct perf_event *child_event,
4927 struct task_struct *child)
4929 struct perf_event *parent_event = child_event->parent;
4932 if (child_event->attr.inherit_stat)
4933 perf_event_read_event(child_event, child);
4935 child_val = atomic64_read(&child_event->count);
4938 * Add back the child's count to the parent's count:
4940 atomic64_add(child_val, &parent_event->count);
4941 atomic64_add(child_event->total_time_enabled,
4942 &parent_event->child_total_time_enabled);
4943 atomic64_add(child_event->total_time_running,
4944 &parent_event->child_total_time_running);
4947 * Remove this event from the parent's list
4949 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4950 mutex_lock(&parent_event->child_mutex);
4951 list_del_init(&child_event->child_list);
4952 mutex_unlock(&parent_event->child_mutex);
4955 * Release the parent event, if this was the last
4958 fput(parent_event->filp);
4962 __perf_event_exit_task(struct perf_event *child_event,
4963 struct perf_event_context *child_ctx,
4964 struct task_struct *child)
4966 struct perf_event *parent_event;
4968 perf_event_remove_from_context(child_event);
4970 parent_event = child_event->parent;
4972 * It can happen that parent exits first, and has events
4973 * that are still around due to the child reference. These
4974 * events need to be zapped - but otherwise linger.
4977 sync_child_event(child_event, child);
4978 free_event(child_event);
4983 * When a child task exits, feed back event values to parent events.
4985 void perf_event_exit_task(struct task_struct *child)
4987 struct perf_event *child_event, *tmp;
4988 struct perf_event_context *child_ctx;
4989 unsigned long flags;
4991 if (likely(!child->perf_event_ctxp)) {
4992 perf_event_task(child, NULL, 0);
4996 local_irq_save(flags);
4998 * We can't reschedule here because interrupts are disabled,
4999 * and either child is current or it is a task that can't be
5000 * scheduled, so we are now safe from rescheduling changing
5003 child_ctx = child->perf_event_ctxp;
5004 __perf_event_task_sched_out(child_ctx);
5007 * Take the context lock here so that if find_get_context is
5008 * reading child->perf_event_ctxp, we wait until it has
5009 * incremented the context's refcount before we do put_ctx below.
5011 raw_spin_lock(&child_ctx->lock);
5012 child->perf_event_ctxp = NULL;
5014 * If this context is a clone; unclone it so it can't get
5015 * swapped to another process while we're removing all
5016 * the events from it.
5018 unclone_ctx(child_ctx);
5019 update_context_time(child_ctx);
5020 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5023 * Report the task dead after unscheduling the events so that we
5024 * won't get any samples after PERF_RECORD_EXIT. We can however still
5025 * get a few PERF_RECORD_READ events.
5027 perf_event_task(child, child_ctx, 0);
5030 * We can recurse on the same lock type through:
5032 * __perf_event_exit_task()
5033 * sync_child_event()
5034 * fput(parent_event->filp)
5036 * mutex_lock(&ctx->mutex)
5038 * But since its the parent context it won't be the same instance.
5040 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5043 list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
5045 __perf_event_exit_task(child_event, child_ctx, child);
5048 * If the last event was a group event, it will have appended all
5049 * its siblings to the list, but we obtained 'tmp' before that which
5050 * will still point to the list head terminating the iteration.
5052 if (!list_empty(&child_ctx->group_list))
5055 mutex_unlock(&child_ctx->mutex);
5061 * free an unexposed, unused context as created by inheritance by
5062 * init_task below, used by fork() in case of fail.
5064 void perf_event_free_task(struct task_struct *task)
5066 struct perf_event_context *ctx = task->perf_event_ctxp;
5067 struct perf_event *event, *tmp;
5072 mutex_lock(&ctx->mutex);
5074 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
5075 struct perf_event *parent = event->parent;
5077 if (WARN_ON_ONCE(!parent))
5080 mutex_lock(&parent->child_mutex);
5081 list_del_init(&event->child_list);
5082 mutex_unlock(&parent->child_mutex);
5086 list_del_event(event, ctx);
5090 if (!list_empty(&ctx->group_list))
5093 mutex_unlock(&ctx->mutex);
5099 * Initialize the perf_event context in task_struct
5101 int perf_event_init_task(struct task_struct *child)
5103 struct perf_event_context *child_ctx = NULL, *parent_ctx;
5104 struct perf_event_context *cloned_ctx;
5105 struct perf_event *event;
5106 struct task_struct *parent = current;
5107 int inherited_all = 1;
5110 child->perf_event_ctxp = NULL;
5112 mutex_init(&child->perf_event_mutex);
5113 INIT_LIST_HEAD(&child->perf_event_list);
5115 if (likely(!parent->perf_event_ctxp))
5119 * If the parent's context is a clone, pin it so it won't get
5122 parent_ctx = perf_pin_task_context(parent);
5125 * No need to check if parent_ctx != NULL here; since we saw
5126 * it non-NULL earlier, the only reason for it to become NULL
5127 * is if we exit, and since we're currently in the middle of
5128 * a fork we can't be exiting at the same time.
5132 * Lock the parent list. No need to lock the child - not PID
5133 * hashed yet and not running, so nobody can access it.
5135 mutex_lock(&parent_ctx->mutex);
5138 * We dont have to disable NMIs - we are only looking at
5139 * the list, not manipulating it:
5141 list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
5143 if (!event->attr.inherit) {
5148 if (!child->perf_event_ctxp) {
5150 * This is executed from the parent task context, so
5151 * inherit events that have been marked for cloning.
5152 * First allocate and initialize a context for the
5156 child_ctx = kzalloc(sizeof(struct perf_event_context),
5163 __perf_event_init_context(child_ctx, child);
5164 child->perf_event_ctxp = child_ctx;
5165 get_task_struct(child);
5168 ret = inherit_group(event, parent, parent_ctx,
5176 if (child_ctx && inherited_all) {
5178 * Mark the child context as a clone of the parent
5179 * context, or of whatever the parent is a clone of.
5180 * Note that if the parent is a clone, it could get
5181 * uncloned at any point, but that doesn't matter
5182 * because the list of events and the generation
5183 * count can't have changed since we took the mutex.
5185 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5187 child_ctx->parent_ctx = cloned_ctx;
5188 child_ctx->parent_gen = parent_ctx->parent_gen;
5190 child_ctx->parent_ctx = parent_ctx;
5191 child_ctx->parent_gen = parent_ctx->generation;
5193 get_ctx(child_ctx->parent_ctx);
5196 mutex_unlock(&parent_ctx->mutex);
5198 perf_unpin_context(parent_ctx);
5203 static void __cpuinit perf_event_init_cpu(int cpu)
5205 struct perf_cpu_context *cpuctx;
5207 cpuctx = &per_cpu(perf_cpu_context, cpu);
5208 __perf_event_init_context(&cpuctx->ctx, NULL);
5210 spin_lock(&perf_resource_lock);
5211 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5212 spin_unlock(&perf_resource_lock);
5214 hw_perf_event_setup(cpu);
5217 #ifdef CONFIG_HOTPLUG_CPU
5218 static void __perf_event_exit_cpu(void *info)
5220 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5221 struct perf_event_context *ctx = &cpuctx->ctx;
5222 struct perf_event *event, *tmp;
5224 list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5225 __perf_event_remove_from_context(event);
5227 static void perf_event_exit_cpu(int cpu)
5229 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5230 struct perf_event_context *ctx = &cpuctx->ctx;
5232 mutex_lock(&ctx->mutex);
5233 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5234 mutex_unlock(&ctx->mutex);
5237 static inline void perf_event_exit_cpu(int cpu) { }
5240 static int __cpuinit
5241 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5243 unsigned int cpu = (long)hcpu;
5247 case CPU_UP_PREPARE:
5248 case CPU_UP_PREPARE_FROZEN:
5249 perf_event_init_cpu(cpu);
5253 case CPU_ONLINE_FROZEN:
5254 hw_perf_event_setup_online(cpu);
5257 case CPU_DOWN_PREPARE:
5258 case CPU_DOWN_PREPARE_FROZEN:
5259 perf_event_exit_cpu(cpu);
5270 * This has to have a higher priority than migration_notifier in sched.c.
5272 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5273 .notifier_call = perf_cpu_notify,
5277 void __init perf_event_init(void)
5279 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5280 (void *)(long)smp_processor_id());
5281 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5282 (void *)(long)smp_processor_id());
5283 register_cpu_notifier(&perf_cpu_nb);
5286 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5288 return sprintf(buf, "%d\n", perf_reserved_percpu);
5292 perf_set_reserve_percpu(struct sysdev_class *class,
5296 struct perf_cpu_context *cpuctx;
5300 err = strict_strtoul(buf, 10, &val);
5303 if (val > perf_max_events)
5306 spin_lock(&perf_resource_lock);
5307 perf_reserved_percpu = val;
5308 for_each_online_cpu(cpu) {
5309 cpuctx = &per_cpu(perf_cpu_context, cpu);
5310 raw_spin_lock_irq(&cpuctx->ctx.lock);
5311 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5312 perf_max_events - perf_reserved_percpu);
5313 cpuctx->max_pertask = mpt;
5314 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5316 spin_unlock(&perf_resource_lock);
5321 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5323 return sprintf(buf, "%d\n", perf_overcommit);
5327 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5332 err = strict_strtoul(buf, 10, &val);
5338 spin_lock(&perf_resource_lock);
5339 perf_overcommit = val;
5340 spin_unlock(&perf_resource_lock);
5345 static SYSDEV_CLASS_ATTR(
5348 perf_show_reserve_percpu,
5349 perf_set_reserve_percpu
5352 static SYSDEV_CLASS_ATTR(
5355 perf_show_overcommit,
5359 static struct attribute *perfclass_attrs[] = {
5360 &attr_reserve_percpu.attr,
5361 &attr_overcommit.attr,
5365 static struct attribute_group perfclass_attr_group = {
5366 .attrs = perfclass_attrs,
5367 .name = "perf_events",
5370 static int __init perf_event_sysfs_init(void)
5372 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5373 &perfclass_attr_group);
5375 device_initcall(perf_event_sysfs_init);