2 * Performance counter 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/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
38 int perf_max_counters __read_mostly
= 1;
39 static int perf_reserved_percpu __read_mostly
;
40 static int perf_overcommit __read_mostly
= 1;
42 static atomic_t nr_counters __read_mostly
;
43 static atomic_t nr_mmap_counters __read_mostly
;
44 static atomic_t nr_comm_counters __read_mostly
;
45 static atomic_t nr_task_counters __read_mostly
;
48 * perf counter paranoia level:
50 * 1 - disallow cpu counters to unpriv
51 * 2 - disallow kernel profiling to unpriv
53 int sysctl_perf_counter_paranoid __read_mostly
;
55 static inline bool perf_paranoid_cpu(void)
57 return sysctl_perf_counter_paranoid
> 0;
60 static inline bool perf_paranoid_kernel(void)
62 return sysctl_perf_counter_paranoid
> 1;
65 int sysctl_perf_counter_mlock __read_mostly
= 512; /* 'free' kb per user */
68 * max perf counter sample rate
70 int sysctl_perf_counter_sample_rate __read_mostly
= 100000;
72 static atomic64_t perf_counter_id
;
75 * Lock for (sysadmin-configurable) counter reservations:
77 static DEFINE_SPINLOCK(perf_resource_lock
);
80 * Architecture provided APIs - weak aliases:
82 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
87 void __weak
hw_perf_disable(void) { barrier(); }
88 void __weak
hw_perf_enable(void) { barrier(); }
90 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
91 void __weak
hw_perf_counter_setup_online(int cpu
) { barrier(); }
94 hw_perf_group_sched_in(struct perf_counter
*group_leader
,
95 struct perf_cpu_context
*cpuctx
,
96 struct perf_counter_context
*ctx
, int cpu
)
101 void __weak
perf_counter_print_debug(void) { }
103 static DEFINE_PER_CPU(int, disable_count
);
105 void __perf_disable(void)
107 __get_cpu_var(disable_count
)++;
110 bool __perf_enable(void)
112 return !--__get_cpu_var(disable_count
);
115 void perf_disable(void)
121 void perf_enable(void)
127 static void get_ctx(struct perf_counter_context
*ctx
)
129 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
132 static void free_ctx(struct rcu_head
*head
)
134 struct perf_counter_context
*ctx
;
136 ctx
= container_of(head
, struct perf_counter_context
, rcu_head
);
140 static void put_ctx(struct perf_counter_context
*ctx
)
142 if (atomic_dec_and_test(&ctx
->refcount
)) {
144 put_ctx(ctx
->parent_ctx
);
146 put_task_struct(ctx
->task
);
147 call_rcu(&ctx
->rcu_head
, free_ctx
);
151 static void unclone_ctx(struct perf_counter_context
*ctx
)
153 if (ctx
->parent_ctx
) {
154 put_ctx(ctx
->parent_ctx
);
155 ctx
->parent_ctx
= NULL
;
160 * If we inherit counters we want to return the parent counter id
163 static u64
primary_counter_id(struct perf_counter
*counter
)
165 u64 id
= counter
->id
;
168 id
= counter
->parent
->id
;
174 * Get the perf_counter_context for a task and lock it.
175 * This has to cope with with the fact that until it is locked,
176 * the context could get moved to another task.
178 static struct perf_counter_context
*
179 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
181 struct perf_counter_context
*ctx
;
185 ctx
= rcu_dereference(task
->perf_counter_ctxp
);
188 * If this context is a clone of another, it might
189 * get swapped for another underneath us by
190 * perf_counter_task_sched_out, though the
191 * rcu_read_lock() protects us from any context
192 * getting freed. Lock the context and check if it
193 * got swapped before we could get the lock, and retry
194 * if so. If we locked the right context, then it
195 * can't get swapped on us any more.
197 spin_lock_irqsave(&ctx
->lock
, *flags
);
198 if (ctx
!= rcu_dereference(task
->perf_counter_ctxp
)) {
199 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
203 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
204 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
213 * Get the context for a task and increment its pin_count so it
214 * can't get swapped to another task. This also increments its
215 * reference count so that the context can't get freed.
217 static struct perf_counter_context
*perf_pin_task_context(struct task_struct
*task
)
219 struct perf_counter_context
*ctx
;
222 ctx
= perf_lock_task_context(task
, &flags
);
225 spin_unlock_irqrestore(&ctx
->lock
, flags
);
230 static void perf_unpin_context(struct perf_counter_context
*ctx
)
234 spin_lock_irqsave(&ctx
->lock
, flags
);
236 spin_unlock_irqrestore(&ctx
->lock
, flags
);
241 * Add a counter from the lists for its context.
242 * Must be called with ctx->mutex and ctx->lock held.
245 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
247 struct perf_counter
*group_leader
= counter
->group_leader
;
250 * Depending on whether it is a standalone or sibling counter,
251 * add it straight to the context's counter list, or to the group
252 * leader's sibling list:
254 if (group_leader
== counter
)
255 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
257 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
258 group_leader
->nr_siblings
++;
261 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
263 if (counter
->attr
.inherit_stat
)
268 * Remove a counter from the lists for its context.
269 * Must be called with ctx->mutex and ctx->lock held.
272 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
274 struct perf_counter
*sibling
, *tmp
;
276 if (list_empty(&counter
->list_entry
))
279 if (counter
->attr
.inherit_stat
)
282 list_del_init(&counter
->list_entry
);
283 list_del_rcu(&counter
->event_entry
);
285 if (counter
->group_leader
!= counter
)
286 counter
->group_leader
->nr_siblings
--;
289 * If this was a group counter with sibling counters then
290 * upgrade the siblings to singleton counters by adding them
291 * to the context list directly:
293 list_for_each_entry_safe(sibling
, tmp
,
294 &counter
->sibling_list
, list_entry
) {
296 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
297 sibling
->group_leader
= sibling
;
302 counter_sched_out(struct perf_counter
*counter
,
303 struct perf_cpu_context
*cpuctx
,
304 struct perf_counter_context
*ctx
)
306 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
309 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
310 if (counter
->pending_disable
) {
311 counter
->pending_disable
= 0;
312 counter
->state
= PERF_COUNTER_STATE_OFF
;
314 counter
->tstamp_stopped
= ctx
->time
;
315 counter
->pmu
->disable(counter
);
318 if (!is_software_counter(counter
))
319 cpuctx
->active_oncpu
--;
321 if (counter
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
322 cpuctx
->exclusive
= 0;
326 group_sched_out(struct perf_counter
*group_counter
,
327 struct perf_cpu_context
*cpuctx
,
328 struct perf_counter_context
*ctx
)
330 struct perf_counter
*counter
;
332 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
335 counter_sched_out(group_counter
, cpuctx
, ctx
);
338 * Schedule out siblings (if any):
340 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
341 counter_sched_out(counter
, cpuctx
, ctx
);
343 if (group_counter
->attr
.exclusive
)
344 cpuctx
->exclusive
= 0;
348 * Cross CPU call to remove a performance counter
350 * We disable the counter on the hardware level first. After that we
351 * remove it from the context list.
353 static void __perf_counter_remove_from_context(void *info
)
355 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
356 struct perf_counter
*counter
= info
;
357 struct perf_counter_context
*ctx
= counter
->ctx
;
360 * If this is a task context, we need to check whether it is
361 * the current task context of this cpu. If not it has been
362 * scheduled out before the smp call arrived.
364 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
367 spin_lock(&ctx
->lock
);
369 * Protect the list operation against NMI by disabling the
370 * counters on a global level.
374 counter_sched_out(counter
, cpuctx
, ctx
);
376 list_del_counter(counter
, ctx
);
380 * Allow more per task counters with respect to the
383 cpuctx
->max_pertask
=
384 min(perf_max_counters
- ctx
->nr_counters
,
385 perf_max_counters
- perf_reserved_percpu
);
389 spin_unlock(&ctx
->lock
);
394 * Remove the counter from a task's (or a CPU's) list of counters.
396 * Must be called with ctx->mutex held.
398 * CPU counters are removed with a smp call. For task counters we only
399 * call when the task is on a CPU.
401 * If counter->ctx is a cloned context, callers must make sure that
402 * every task struct that counter->ctx->task could possibly point to
403 * remains valid. This is OK when called from perf_release since
404 * that only calls us on the top-level context, which can't be a clone.
405 * When called from perf_counter_exit_task, it's OK because the
406 * context has been detached from its task.
408 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
410 struct perf_counter_context
*ctx
= counter
->ctx
;
411 struct task_struct
*task
= ctx
->task
;
415 * Per cpu counters are removed via an smp call and
416 * the removal is always sucessful.
418 smp_call_function_single(counter
->cpu
,
419 __perf_counter_remove_from_context
,
425 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
428 spin_lock_irq(&ctx
->lock
);
430 * If the context is active we need to retry the smp call.
432 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
433 spin_unlock_irq(&ctx
->lock
);
438 * The lock prevents that this context is scheduled in so we
439 * can remove the counter safely, if the call above did not
442 if (!list_empty(&counter
->list_entry
)) {
443 list_del_counter(counter
, ctx
);
445 spin_unlock_irq(&ctx
->lock
);
448 static inline u64
perf_clock(void)
450 return cpu_clock(smp_processor_id());
454 * Update the record of the current time in a context.
456 static void update_context_time(struct perf_counter_context
*ctx
)
458 u64 now
= perf_clock();
460 ctx
->time
+= now
- ctx
->timestamp
;
461 ctx
->timestamp
= now
;
465 * Update the total_time_enabled and total_time_running fields for a counter.
467 static void update_counter_times(struct perf_counter
*counter
)
469 struct perf_counter_context
*ctx
= counter
->ctx
;
472 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
475 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
477 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
478 run_end
= counter
->tstamp_stopped
;
482 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
486 * Update total_time_enabled and total_time_running for all counters in a group.
488 static void update_group_times(struct perf_counter
*leader
)
490 struct perf_counter
*counter
;
492 update_counter_times(leader
);
493 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
494 update_counter_times(counter
);
498 * Cross CPU call to disable a performance counter
500 static void __perf_counter_disable(void *info
)
502 struct perf_counter
*counter
= info
;
503 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
504 struct perf_counter_context
*ctx
= counter
->ctx
;
507 * If this is a per-task counter, need to check whether this
508 * counter's task is the current task on this cpu.
510 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
513 spin_lock(&ctx
->lock
);
516 * If the counter is on, turn it off.
517 * If it is in error state, leave it in error state.
519 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
520 update_context_time(ctx
);
521 update_counter_times(counter
);
522 if (counter
== counter
->group_leader
)
523 group_sched_out(counter
, cpuctx
, ctx
);
525 counter_sched_out(counter
, cpuctx
, ctx
);
526 counter
->state
= PERF_COUNTER_STATE_OFF
;
529 spin_unlock(&ctx
->lock
);
535 * If counter->ctx is a cloned context, callers must make sure that
536 * every task struct that counter->ctx->task could possibly point to
537 * remains valid. This condition is satisifed when called through
538 * perf_counter_for_each_child or perf_counter_for_each because they
539 * hold the top-level counter's child_mutex, so any descendant that
540 * goes to exit will block in sync_child_counter.
541 * When called from perf_pending_counter it's OK because counter->ctx
542 * is the current context on this CPU and preemption is disabled,
543 * hence we can't get into perf_counter_task_sched_out for this context.
545 static void perf_counter_disable(struct perf_counter
*counter
)
547 struct perf_counter_context
*ctx
= counter
->ctx
;
548 struct task_struct
*task
= ctx
->task
;
552 * Disable the counter on the cpu that it's on
554 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
560 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
562 spin_lock_irq(&ctx
->lock
);
564 * If the counter is still active, we need to retry the cross-call.
566 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
567 spin_unlock_irq(&ctx
->lock
);
572 * Since we have the lock this context can't be scheduled
573 * in, so we can change the state safely.
575 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
576 update_counter_times(counter
);
577 counter
->state
= PERF_COUNTER_STATE_OFF
;
580 spin_unlock_irq(&ctx
->lock
);
584 counter_sched_in(struct perf_counter
*counter
,
585 struct perf_cpu_context
*cpuctx
,
586 struct perf_counter_context
*ctx
,
589 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
592 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
593 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
595 * The new state must be visible before we turn it on in the hardware:
599 if (counter
->pmu
->enable(counter
)) {
600 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
605 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
607 if (!is_software_counter(counter
))
608 cpuctx
->active_oncpu
++;
611 if (counter
->attr
.exclusive
)
612 cpuctx
->exclusive
= 1;
618 group_sched_in(struct perf_counter
*group_counter
,
619 struct perf_cpu_context
*cpuctx
,
620 struct perf_counter_context
*ctx
,
623 struct perf_counter
*counter
, *partial_group
;
626 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
629 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
631 return ret
< 0 ? ret
: 0;
633 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
637 * Schedule in siblings as one group (if any):
639 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
640 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
641 partial_group
= counter
;
650 * Groups can be scheduled in as one unit only, so undo any
651 * partial group before returning:
653 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
654 if (counter
== partial_group
)
656 counter_sched_out(counter
, cpuctx
, ctx
);
658 counter_sched_out(group_counter
, cpuctx
, ctx
);
664 * Return 1 for a group consisting entirely of software counters,
665 * 0 if the group contains any hardware counters.
667 static int is_software_only_group(struct perf_counter
*leader
)
669 struct perf_counter
*counter
;
671 if (!is_software_counter(leader
))
674 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
675 if (!is_software_counter(counter
))
682 * Work out whether we can put this counter group on the CPU now.
684 static int group_can_go_on(struct perf_counter
*counter
,
685 struct perf_cpu_context
*cpuctx
,
689 * Groups consisting entirely of software counters can always go on.
691 if (is_software_only_group(counter
))
694 * If an exclusive group is already on, no other hardware
695 * counters can go on.
697 if (cpuctx
->exclusive
)
700 * If this group is exclusive and there are already
701 * counters on the CPU, it can't go on.
703 if (counter
->attr
.exclusive
&& cpuctx
->active_oncpu
)
706 * Otherwise, try to add it if all previous groups were able
712 static void add_counter_to_ctx(struct perf_counter
*counter
,
713 struct perf_counter_context
*ctx
)
715 list_add_counter(counter
, ctx
);
716 counter
->tstamp_enabled
= ctx
->time
;
717 counter
->tstamp_running
= ctx
->time
;
718 counter
->tstamp_stopped
= ctx
->time
;
722 * Cross CPU call to install and enable a performance counter
724 * Must be called with ctx->mutex held
726 static void __perf_install_in_context(void *info
)
728 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
729 struct perf_counter
*counter
= info
;
730 struct perf_counter_context
*ctx
= counter
->ctx
;
731 struct perf_counter
*leader
= counter
->group_leader
;
732 int cpu
= smp_processor_id();
736 * If this is a task context, we need to check whether it is
737 * the current task context of this cpu. If not it has been
738 * scheduled out before the smp call arrived.
739 * Or possibly this is the right context but it isn't
740 * on this cpu because it had no counters.
742 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
743 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
745 cpuctx
->task_ctx
= ctx
;
748 spin_lock(&ctx
->lock
);
750 update_context_time(ctx
);
753 * Protect the list operation against NMI by disabling the
754 * counters on a global level. NOP for non NMI based counters.
758 add_counter_to_ctx(counter
, ctx
);
761 * Don't put the counter on if it is disabled or if
762 * it is in a group and the group isn't on.
764 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
765 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
769 * An exclusive counter can't go on if there are already active
770 * hardware counters, and no hardware counter can go on if there
771 * is already an exclusive counter on.
773 if (!group_can_go_on(counter
, cpuctx
, 1))
776 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
780 * This counter couldn't go on. If it is in a group
781 * then we have to pull the whole group off.
782 * If the counter group is pinned then put it in error state.
784 if (leader
!= counter
)
785 group_sched_out(leader
, cpuctx
, ctx
);
786 if (leader
->attr
.pinned
) {
787 update_group_times(leader
);
788 leader
->state
= PERF_COUNTER_STATE_ERROR
;
792 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
793 cpuctx
->max_pertask
--;
798 spin_unlock(&ctx
->lock
);
802 * Attach a performance counter to a context
804 * First we add the counter to the list with the hardware enable bit
805 * in counter->hw_config cleared.
807 * If the counter is attached to a task which is on a CPU we use a smp
808 * call to enable it in the task context. The task might have been
809 * scheduled away, but we check this in the smp call again.
811 * Must be called with ctx->mutex held.
814 perf_install_in_context(struct perf_counter_context
*ctx
,
815 struct perf_counter
*counter
,
818 struct task_struct
*task
= ctx
->task
;
822 * Per cpu counters are installed via an smp call and
823 * the install is always sucessful.
825 smp_call_function_single(cpu
, __perf_install_in_context
,
831 task_oncpu_function_call(task
, __perf_install_in_context
,
834 spin_lock_irq(&ctx
->lock
);
836 * we need to retry the smp call.
838 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
839 spin_unlock_irq(&ctx
->lock
);
844 * The lock prevents that this context is scheduled in so we
845 * can add the counter safely, if it the call above did not
848 if (list_empty(&counter
->list_entry
))
849 add_counter_to_ctx(counter
, ctx
);
850 spin_unlock_irq(&ctx
->lock
);
854 * Cross CPU call to enable a performance counter
856 static void __perf_counter_enable(void *info
)
858 struct perf_counter
*counter
= info
;
859 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
860 struct perf_counter_context
*ctx
= counter
->ctx
;
861 struct perf_counter
*leader
= counter
->group_leader
;
865 * If this is a per-task counter, need to check whether this
866 * counter's task is the current task on this cpu.
868 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
869 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
871 cpuctx
->task_ctx
= ctx
;
874 spin_lock(&ctx
->lock
);
876 update_context_time(ctx
);
878 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
880 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
881 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
884 * If the counter is in a group and isn't the group leader,
885 * then don't put it on unless the group is on.
887 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
890 if (!group_can_go_on(counter
, cpuctx
, 1)) {
894 if (counter
== leader
)
895 err
= group_sched_in(counter
, cpuctx
, ctx
,
898 err
= counter_sched_in(counter
, cpuctx
, ctx
,
905 * If this counter can't go on and it's part of a
906 * group, then the whole group has to come off.
908 if (leader
!= counter
)
909 group_sched_out(leader
, cpuctx
, ctx
);
910 if (leader
->attr
.pinned
) {
911 update_group_times(leader
);
912 leader
->state
= PERF_COUNTER_STATE_ERROR
;
917 spin_unlock(&ctx
->lock
);
923 * If counter->ctx is a cloned context, callers must make sure that
924 * every task struct that counter->ctx->task could possibly point to
925 * remains valid. This condition is satisfied when called through
926 * perf_counter_for_each_child or perf_counter_for_each as described
927 * for perf_counter_disable.
929 static void perf_counter_enable(struct perf_counter
*counter
)
931 struct perf_counter_context
*ctx
= counter
->ctx
;
932 struct task_struct
*task
= ctx
->task
;
936 * Enable the counter on the cpu that it's on
938 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
943 spin_lock_irq(&ctx
->lock
);
944 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
948 * If the counter is in error state, clear that first.
949 * That way, if we see the counter in error state below, we
950 * know that it has gone back into error state, as distinct
951 * from the task having been scheduled away before the
952 * cross-call arrived.
954 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
955 counter
->state
= PERF_COUNTER_STATE_OFF
;
958 spin_unlock_irq(&ctx
->lock
);
959 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
961 spin_lock_irq(&ctx
->lock
);
964 * If the context is active and the counter is still off,
965 * we need to retry the cross-call.
967 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
971 * Since we have the lock this context can't be scheduled
972 * in, so we can change the state safely.
974 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
975 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
976 counter
->tstamp_enabled
=
977 ctx
->time
- counter
->total_time_enabled
;
980 spin_unlock_irq(&ctx
->lock
);
983 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
986 * not supported on inherited counters
988 if (counter
->attr
.inherit
)
991 atomic_add(refresh
, &counter
->event_limit
);
992 perf_counter_enable(counter
);
997 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
998 struct perf_cpu_context
*cpuctx
)
1000 struct perf_counter
*counter
;
1002 spin_lock(&ctx
->lock
);
1004 if (likely(!ctx
->nr_counters
))
1006 update_context_time(ctx
);
1009 if (ctx
->nr_active
) {
1010 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1011 if (counter
!= counter
->group_leader
)
1012 counter_sched_out(counter
, cpuctx
, ctx
);
1014 group_sched_out(counter
, cpuctx
, ctx
);
1019 spin_unlock(&ctx
->lock
);
1023 * Test whether two contexts are equivalent, i.e. whether they
1024 * have both been cloned from the same version of the same context
1025 * and they both have the same number of enabled counters.
1026 * If the number of enabled counters is the same, then the set
1027 * of enabled counters should be the same, because these are both
1028 * inherited contexts, therefore we can't access individual counters
1029 * in them directly with an fd; we can only enable/disable all
1030 * counters via prctl, or enable/disable all counters in a family
1031 * via ioctl, which will have the same effect on both contexts.
1033 static int context_equiv(struct perf_counter_context
*ctx1
,
1034 struct perf_counter_context
*ctx2
)
1036 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1037 && ctx1
->parent_gen
== ctx2
->parent_gen
1038 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1041 static void __perf_counter_read(void *counter
);
1043 static void __perf_counter_sync_stat(struct perf_counter
*counter
,
1044 struct perf_counter
*next_counter
)
1048 if (!counter
->attr
.inherit_stat
)
1052 * Update the counter value, we cannot use perf_counter_read()
1053 * because we're in the middle of a context switch and have IRQs
1054 * disabled, which upsets smp_call_function_single(), however
1055 * we know the counter must be on the current CPU, therefore we
1056 * don't need to use it.
1058 switch (counter
->state
) {
1059 case PERF_COUNTER_STATE_ACTIVE
:
1060 __perf_counter_read(counter
);
1063 case PERF_COUNTER_STATE_INACTIVE
:
1064 update_counter_times(counter
);
1072 * In order to keep per-task stats reliable we need to flip the counter
1073 * values when we flip the contexts.
1075 value
= atomic64_read(&next_counter
->count
);
1076 value
= atomic64_xchg(&counter
->count
, value
);
1077 atomic64_set(&next_counter
->count
, value
);
1079 swap(counter
->total_time_enabled
, next_counter
->total_time_enabled
);
1080 swap(counter
->total_time_running
, next_counter
->total_time_running
);
1083 * Since we swizzled the values, update the user visible data too.
1085 perf_counter_update_userpage(counter
);
1086 perf_counter_update_userpage(next_counter
);
1089 #define list_next_entry(pos, member) \
1090 list_entry(pos->member.next, typeof(*pos), member)
1092 static void perf_counter_sync_stat(struct perf_counter_context
*ctx
,
1093 struct perf_counter_context
*next_ctx
)
1095 struct perf_counter
*counter
, *next_counter
;
1100 counter
= list_first_entry(&ctx
->event_list
,
1101 struct perf_counter
, event_entry
);
1103 next_counter
= list_first_entry(&next_ctx
->event_list
,
1104 struct perf_counter
, event_entry
);
1106 while (&counter
->event_entry
!= &ctx
->event_list
&&
1107 &next_counter
->event_entry
!= &next_ctx
->event_list
) {
1109 __perf_counter_sync_stat(counter
, next_counter
);
1111 counter
= list_next_entry(counter
, event_entry
);
1112 next_counter
= list_next_entry(next_counter
, event_entry
);
1117 * Called from scheduler to remove the counters of the current task,
1118 * with interrupts disabled.
1120 * We stop each counter and update the counter value in counter->count.
1122 * This does not protect us against NMI, but disable()
1123 * sets the disabled bit in the control field of counter _before_
1124 * accessing the counter control register. If a NMI hits, then it will
1125 * not restart the counter.
1127 void perf_counter_task_sched_out(struct task_struct
*task
,
1128 struct task_struct
*next
, int cpu
)
1130 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1131 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1132 struct perf_counter_context
*next_ctx
;
1133 struct perf_counter_context
*parent
;
1134 struct pt_regs
*regs
;
1137 regs
= task_pt_regs(task
);
1138 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1140 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1143 update_context_time(ctx
);
1146 parent
= rcu_dereference(ctx
->parent_ctx
);
1147 next_ctx
= next
->perf_counter_ctxp
;
1148 if (parent
&& next_ctx
&&
1149 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1151 * Looks like the two contexts are clones, so we might be
1152 * able to optimize the context switch. We lock both
1153 * contexts and check that they are clones under the
1154 * lock (including re-checking that neither has been
1155 * uncloned in the meantime). It doesn't matter which
1156 * order we take the locks because no other cpu could
1157 * be trying to lock both of these tasks.
1159 spin_lock(&ctx
->lock
);
1160 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1161 if (context_equiv(ctx
, next_ctx
)) {
1163 * XXX do we need a memory barrier of sorts
1164 * wrt to rcu_dereference() of perf_counter_ctxp
1166 task
->perf_counter_ctxp
= next_ctx
;
1167 next
->perf_counter_ctxp
= ctx
;
1169 next_ctx
->task
= task
;
1172 perf_counter_sync_stat(ctx
, next_ctx
);
1174 spin_unlock(&next_ctx
->lock
);
1175 spin_unlock(&ctx
->lock
);
1180 __perf_counter_sched_out(ctx
, cpuctx
);
1181 cpuctx
->task_ctx
= NULL
;
1186 * Called with IRQs disabled
1188 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
1190 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1192 if (!cpuctx
->task_ctx
)
1195 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1198 __perf_counter_sched_out(ctx
, cpuctx
);
1199 cpuctx
->task_ctx
= NULL
;
1203 * Called with IRQs disabled
1205 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1207 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
1211 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
1212 struct perf_cpu_context
*cpuctx
, int cpu
)
1214 struct perf_counter
*counter
;
1217 spin_lock(&ctx
->lock
);
1219 if (likely(!ctx
->nr_counters
))
1222 ctx
->timestamp
= perf_clock();
1227 * First go through the list and put on any pinned groups
1228 * in order to give them the best chance of going on.
1230 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1231 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1232 !counter
->attr
.pinned
)
1234 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1237 if (counter
!= counter
->group_leader
)
1238 counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
1240 if (group_can_go_on(counter
, cpuctx
, 1))
1241 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
1245 * If this pinned group hasn't been scheduled,
1246 * put it in error state.
1248 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1249 update_group_times(counter
);
1250 counter
->state
= PERF_COUNTER_STATE_ERROR
;
1254 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1256 * Ignore counters in OFF or ERROR state, and
1257 * ignore pinned counters since we did them already.
1259 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1260 counter
->attr
.pinned
)
1264 * Listen to the 'cpu' scheduling filter constraint
1267 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1270 if (counter
!= counter
->group_leader
) {
1271 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
))
1274 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
1275 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
1282 spin_unlock(&ctx
->lock
);
1286 * Called from scheduler to add the counters of the current task
1287 * with interrupts disabled.
1289 * We restore the counter value and then enable it.
1291 * This does not protect us against NMI, but enable()
1292 * sets the enabled bit in the control field of counter _before_
1293 * accessing the counter control register. If a NMI hits, then it will
1294 * keep the counter running.
1296 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
1298 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1299 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1303 if (cpuctx
->task_ctx
== ctx
)
1305 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1306 cpuctx
->task_ctx
= ctx
;
1309 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1311 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
1313 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1316 #define MAX_INTERRUPTS (~0ULL)
1318 static void perf_log_throttle(struct perf_counter
*counter
, int enable
);
1320 static void perf_adjust_period(struct perf_counter
*counter
, u64 events
)
1322 struct hw_perf_counter
*hwc
= &counter
->hw
;
1323 u64 period
, sample_period
;
1326 events
*= hwc
->sample_period
;
1327 period
= div64_u64(events
, counter
->attr
.sample_freq
);
1329 delta
= (s64
)(period
- hwc
->sample_period
);
1330 delta
= (delta
+ 7) / 8; /* low pass filter */
1332 sample_period
= hwc
->sample_period
+ delta
;
1337 hwc
->sample_period
= sample_period
;
1340 static void perf_ctx_adjust_freq(struct perf_counter_context
*ctx
)
1342 struct perf_counter
*counter
;
1343 struct hw_perf_counter
*hwc
;
1344 u64 interrupts
, freq
;
1346 spin_lock(&ctx
->lock
);
1347 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1348 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1353 interrupts
= hwc
->interrupts
;
1354 hwc
->interrupts
= 0;
1357 * unthrottle counters on the tick
1359 if (interrupts
== MAX_INTERRUPTS
) {
1360 perf_log_throttle(counter
, 1);
1361 counter
->pmu
->unthrottle(counter
);
1362 interrupts
= 2*sysctl_perf_counter_sample_rate
/HZ
;
1365 if (!counter
->attr
.freq
|| !counter
->attr
.sample_freq
)
1369 * if the specified freq < HZ then we need to skip ticks
1371 if (counter
->attr
.sample_freq
< HZ
) {
1372 freq
= counter
->attr
.sample_freq
;
1374 hwc
->freq_count
+= freq
;
1375 hwc
->freq_interrupts
+= interrupts
;
1377 if (hwc
->freq_count
< HZ
)
1380 interrupts
= hwc
->freq_interrupts
;
1381 hwc
->freq_interrupts
= 0;
1382 hwc
->freq_count
-= HZ
;
1386 perf_adjust_period(counter
, freq
* interrupts
);
1389 * In order to avoid being stalled by an (accidental) huge
1390 * sample period, force reset the sample period if we didn't
1391 * get any events in this freq period.
1395 counter
->pmu
->disable(counter
);
1396 atomic64_set(&hwc
->period_left
, 0);
1397 counter
->pmu
->enable(counter
);
1401 spin_unlock(&ctx
->lock
);
1405 * Round-robin a context's counters:
1407 static void rotate_ctx(struct perf_counter_context
*ctx
)
1409 struct perf_counter
*counter
;
1411 if (!ctx
->nr_counters
)
1414 spin_lock(&ctx
->lock
);
1416 * Rotate the first entry last (works just fine for group counters too):
1419 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1420 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1425 spin_unlock(&ctx
->lock
);
1428 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1430 struct perf_cpu_context
*cpuctx
;
1431 struct perf_counter_context
*ctx
;
1433 if (!atomic_read(&nr_counters
))
1436 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1437 ctx
= curr
->perf_counter_ctxp
;
1439 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1441 perf_ctx_adjust_freq(ctx
);
1443 perf_counter_cpu_sched_out(cpuctx
);
1445 __perf_counter_task_sched_out(ctx
);
1447 rotate_ctx(&cpuctx
->ctx
);
1451 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1453 perf_counter_task_sched_in(curr
, cpu
);
1457 * Enable all of a task's counters that have been marked enable-on-exec.
1458 * This expects task == current.
1460 static void perf_counter_enable_on_exec(struct task_struct
*task
)
1462 struct perf_counter_context
*ctx
;
1463 struct perf_counter
*counter
;
1464 unsigned long flags
;
1467 local_irq_save(flags
);
1468 ctx
= task
->perf_counter_ctxp
;
1469 if (!ctx
|| !ctx
->nr_counters
)
1472 __perf_counter_task_sched_out(ctx
);
1474 spin_lock(&ctx
->lock
);
1476 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1477 if (!counter
->attr
.enable_on_exec
)
1479 counter
->attr
.enable_on_exec
= 0;
1480 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
1482 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
1483 counter
->tstamp_enabled
=
1484 ctx
->time
- counter
->total_time_enabled
;
1489 * Unclone this context if we enabled any counter.
1494 spin_unlock(&ctx
->lock
);
1496 perf_counter_task_sched_in(task
, smp_processor_id());
1498 local_irq_restore(flags
);
1502 * Cross CPU call to read the hardware counter
1504 static void __perf_counter_read(void *info
)
1506 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1507 struct perf_counter
*counter
= info
;
1508 struct perf_counter_context
*ctx
= counter
->ctx
;
1509 unsigned long flags
;
1512 * If this is a task context, we need to check whether it is
1513 * the current task context of this cpu. If not it has been
1514 * scheduled out before the smp call arrived. In that case
1515 * counter->count would have been updated to a recent sample
1516 * when the counter was scheduled out.
1518 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1521 local_irq_save(flags
);
1523 update_context_time(ctx
);
1524 counter
->pmu
->read(counter
);
1525 update_counter_times(counter
);
1526 local_irq_restore(flags
);
1529 static u64
perf_counter_read(struct perf_counter
*counter
)
1532 * If counter is enabled and currently active on a CPU, update the
1533 * value in the counter structure:
1535 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1536 smp_call_function_single(counter
->oncpu
,
1537 __perf_counter_read
, counter
, 1);
1538 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1539 update_counter_times(counter
);
1542 return atomic64_read(&counter
->count
);
1546 * Initialize the perf_counter context in a task_struct:
1549 __perf_counter_init_context(struct perf_counter_context
*ctx
,
1550 struct task_struct
*task
)
1552 memset(ctx
, 0, sizeof(*ctx
));
1553 spin_lock_init(&ctx
->lock
);
1554 mutex_init(&ctx
->mutex
);
1555 INIT_LIST_HEAD(&ctx
->counter_list
);
1556 INIT_LIST_HEAD(&ctx
->event_list
);
1557 atomic_set(&ctx
->refcount
, 1);
1561 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1563 struct perf_counter_context
*ctx
;
1564 struct perf_cpu_context
*cpuctx
;
1565 struct task_struct
*task
;
1566 unsigned long flags
;
1570 * If cpu is not a wildcard then this is a percpu counter:
1573 /* Must be root to operate on a CPU counter: */
1574 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1575 return ERR_PTR(-EACCES
);
1577 if (cpu
< 0 || cpu
> num_possible_cpus())
1578 return ERR_PTR(-EINVAL
);
1581 * We could be clever and allow to attach a counter to an
1582 * offline CPU and activate it when the CPU comes up, but
1585 if (!cpu_isset(cpu
, cpu_online_map
))
1586 return ERR_PTR(-ENODEV
);
1588 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1599 task
= find_task_by_vpid(pid
);
1601 get_task_struct(task
);
1605 return ERR_PTR(-ESRCH
);
1608 * Can't attach counters to a dying task.
1611 if (task
->flags
& PF_EXITING
)
1614 /* Reuse ptrace permission checks for now. */
1616 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1620 ctx
= perf_lock_task_context(task
, &flags
);
1623 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1627 ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
1631 __perf_counter_init_context(ctx
, task
);
1633 if (cmpxchg(&task
->perf_counter_ctxp
, NULL
, ctx
)) {
1635 * We raced with some other task; use
1636 * the context they set.
1641 get_task_struct(task
);
1644 put_task_struct(task
);
1648 put_task_struct(task
);
1649 return ERR_PTR(err
);
1652 static void free_counter_rcu(struct rcu_head
*head
)
1654 struct perf_counter
*counter
;
1656 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1658 put_pid_ns(counter
->ns
);
1662 static void perf_pending_sync(struct perf_counter
*counter
);
1664 static void free_counter(struct perf_counter
*counter
)
1666 perf_pending_sync(counter
);
1668 if (!counter
->parent
) {
1669 atomic_dec(&nr_counters
);
1670 if (counter
->attr
.mmap
)
1671 atomic_dec(&nr_mmap_counters
);
1672 if (counter
->attr
.comm
)
1673 atomic_dec(&nr_comm_counters
);
1674 if (counter
->attr
.task
)
1675 atomic_dec(&nr_task_counters
);
1678 if (counter
->destroy
)
1679 counter
->destroy(counter
);
1681 put_ctx(counter
->ctx
);
1682 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1686 * Called when the last reference to the file is gone.
1688 static int perf_release(struct inode
*inode
, struct file
*file
)
1690 struct perf_counter
*counter
= file
->private_data
;
1691 struct perf_counter_context
*ctx
= counter
->ctx
;
1693 file
->private_data
= NULL
;
1695 WARN_ON_ONCE(ctx
->parent_ctx
);
1696 mutex_lock(&ctx
->mutex
);
1697 perf_counter_remove_from_context(counter
);
1698 mutex_unlock(&ctx
->mutex
);
1700 mutex_lock(&counter
->owner
->perf_counter_mutex
);
1701 list_del_init(&counter
->owner_entry
);
1702 mutex_unlock(&counter
->owner
->perf_counter_mutex
);
1703 put_task_struct(counter
->owner
);
1705 free_counter(counter
);
1710 static int perf_counter_read_size(struct perf_counter
*counter
)
1712 int entry
= sizeof(u64
); /* value */
1716 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1717 size
+= sizeof(u64
);
1719 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1720 size
+= sizeof(u64
);
1722 if (counter
->attr
.read_format
& PERF_FORMAT_ID
)
1723 entry
+= sizeof(u64
);
1725 if (counter
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1726 nr
+= counter
->group_leader
->nr_siblings
;
1727 size
+= sizeof(u64
);
1735 static u64
perf_counter_read_value(struct perf_counter
*counter
)
1737 struct perf_counter
*child
;
1740 total
+= perf_counter_read(counter
);
1741 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1742 total
+= perf_counter_read(child
);
1747 static int perf_counter_read_entry(struct perf_counter
*counter
,
1748 u64 read_format
, char __user
*buf
)
1750 int n
= 0, count
= 0;
1753 values
[n
++] = perf_counter_read_value(counter
);
1754 if (read_format
& PERF_FORMAT_ID
)
1755 values
[n
++] = primary_counter_id(counter
);
1757 count
= n
* sizeof(u64
);
1759 if (copy_to_user(buf
, values
, count
))
1765 static int perf_counter_read_group(struct perf_counter
*counter
,
1766 u64 read_format
, char __user
*buf
)
1768 struct perf_counter
*leader
= counter
->group_leader
, *sub
;
1769 int n
= 0, size
= 0, err
= -EFAULT
;
1772 values
[n
++] = 1 + leader
->nr_siblings
;
1773 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1774 values
[n
++] = leader
->total_time_enabled
+
1775 atomic64_read(&leader
->child_total_time_enabled
);
1777 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1778 values
[n
++] = leader
->total_time_running
+
1779 atomic64_read(&leader
->child_total_time_running
);
1782 size
= n
* sizeof(u64
);
1784 if (copy_to_user(buf
, values
, size
))
1787 err
= perf_counter_read_entry(leader
, read_format
, buf
+ size
);
1793 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
1794 err
= perf_counter_read_entry(counter
, read_format
,
1805 static int perf_counter_read_one(struct perf_counter
*counter
,
1806 u64 read_format
, char __user
*buf
)
1811 values
[n
++] = perf_counter_read_value(counter
);
1812 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1813 values
[n
++] = counter
->total_time_enabled
+
1814 atomic64_read(&counter
->child_total_time_enabled
);
1816 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1817 values
[n
++] = counter
->total_time_running
+
1818 atomic64_read(&counter
->child_total_time_running
);
1820 if (read_format
& PERF_FORMAT_ID
)
1821 values
[n
++] = primary_counter_id(counter
);
1823 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
1826 return n
* sizeof(u64
);
1830 * Read the performance counter - simple non blocking version for now
1833 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1835 u64 read_format
= counter
->attr
.read_format
;
1839 * Return end-of-file for a read on a counter that is in
1840 * error state (i.e. because it was pinned but it couldn't be
1841 * scheduled on to the CPU at some point).
1843 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1846 if (count
< perf_counter_read_size(counter
))
1849 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1850 mutex_lock(&counter
->child_mutex
);
1851 if (read_format
& PERF_FORMAT_GROUP
)
1852 ret
= perf_counter_read_group(counter
, read_format
, buf
);
1854 ret
= perf_counter_read_one(counter
, read_format
, buf
);
1855 mutex_unlock(&counter
->child_mutex
);
1861 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1863 struct perf_counter
*counter
= file
->private_data
;
1865 return perf_read_hw(counter
, buf
, count
);
1868 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1870 struct perf_counter
*counter
= file
->private_data
;
1871 struct perf_mmap_data
*data
;
1872 unsigned int events
= POLL_HUP
;
1875 data
= rcu_dereference(counter
->data
);
1877 events
= atomic_xchg(&data
->poll
, 0);
1880 poll_wait(file
, &counter
->waitq
, wait
);
1885 static void perf_counter_reset(struct perf_counter
*counter
)
1887 (void)perf_counter_read(counter
);
1888 atomic64_set(&counter
->count
, 0);
1889 perf_counter_update_userpage(counter
);
1893 * Holding the top-level counter's child_mutex means that any
1894 * descendant process that has inherited this counter will block
1895 * in sync_child_counter if it goes to exit, thus satisfying the
1896 * task existence requirements of perf_counter_enable/disable.
1898 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1899 void (*func
)(struct perf_counter
*))
1901 struct perf_counter
*child
;
1903 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1904 mutex_lock(&counter
->child_mutex
);
1906 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1908 mutex_unlock(&counter
->child_mutex
);
1911 static void perf_counter_for_each(struct perf_counter
*counter
,
1912 void (*func
)(struct perf_counter
*))
1914 struct perf_counter_context
*ctx
= counter
->ctx
;
1915 struct perf_counter
*sibling
;
1917 WARN_ON_ONCE(ctx
->parent_ctx
);
1918 mutex_lock(&ctx
->mutex
);
1919 counter
= counter
->group_leader
;
1921 perf_counter_for_each_child(counter
, func
);
1923 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1924 perf_counter_for_each_child(counter
, func
);
1925 mutex_unlock(&ctx
->mutex
);
1928 static int perf_counter_period(struct perf_counter
*counter
, u64 __user
*arg
)
1930 struct perf_counter_context
*ctx
= counter
->ctx
;
1935 if (!counter
->attr
.sample_period
)
1938 size
= copy_from_user(&value
, arg
, sizeof(value
));
1939 if (size
!= sizeof(value
))
1945 spin_lock_irq(&ctx
->lock
);
1946 if (counter
->attr
.freq
) {
1947 if (value
> sysctl_perf_counter_sample_rate
) {
1952 counter
->attr
.sample_freq
= value
;
1954 counter
->attr
.sample_period
= value
;
1955 counter
->hw
.sample_period
= value
;
1958 spin_unlock_irq(&ctx
->lock
);
1963 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1965 struct perf_counter
*counter
= file
->private_data
;
1966 void (*func
)(struct perf_counter
*);
1970 case PERF_COUNTER_IOC_ENABLE
:
1971 func
= perf_counter_enable
;
1973 case PERF_COUNTER_IOC_DISABLE
:
1974 func
= perf_counter_disable
;
1976 case PERF_COUNTER_IOC_RESET
:
1977 func
= perf_counter_reset
;
1980 case PERF_COUNTER_IOC_REFRESH
:
1981 return perf_counter_refresh(counter
, arg
);
1983 case PERF_COUNTER_IOC_PERIOD
:
1984 return perf_counter_period(counter
, (u64 __user
*)arg
);
1990 if (flags
& PERF_IOC_FLAG_GROUP
)
1991 perf_counter_for_each(counter
, func
);
1993 perf_counter_for_each_child(counter
, func
);
1998 int perf_counter_task_enable(void)
2000 struct perf_counter
*counter
;
2002 mutex_lock(¤t
->perf_counter_mutex
);
2003 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
2004 perf_counter_for_each_child(counter
, perf_counter_enable
);
2005 mutex_unlock(¤t
->perf_counter_mutex
);
2010 int perf_counter_task_disable(void)
2012 struct perf_counter
*counter
;
2014 mutex_lock(¤t
->perf_counter_mutex
);
2015 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
2016 perf_counter_for_each_child(counter
, perf_counter_disable
);
2017 mutex_unlock(¤t
->perf_counter_mutex
);
2022 static int perf_counter_index(struct perf_counter
*counter
)
2024 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
2027 return counter
->hw
.idx
+ 1 - PERF_COUNTER_INDEX_OFFSET
;
2031 * Callers need to ensure there can be no nesting of this function, otherwise
2032 * the seqlock logic goes bad. We can not serialize this because the arch
2033 * code calls this from NMI context.
2035 void perf_counter_update_userpage(struct perf_counter
*counter
)
2037 struct perf_counter_mmap_page
*userpg
;
2038 struct perf_mmap_data
*data
;
2041 data
= rcu_dereference(counter
->data
);
2045 userpg
= data
->user_page
;
2048 * Disable preemption so as to not let the corresponding user-space
2049 * spin too long if we get preempted.
2054 userpg
->index
= perf_counter_index(counter
);
2055 userpg
->offset
= atomic64_read(&counter
->count
);
2056 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
2057 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
2059 userpg
->time_enabled
= counter
->total_time_enabled
+
2060 atomic64_read(&counter
->child_total_time_enabled
);
2062 userpg
->time_running
= counter
->total_time_running
+
2063 atomic64_read(&counter
->child_total_time_running
);
2072 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2074 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
2075 struct perf_mmap_data
*data
;
2076 int ret
= VM_FAULT_SIGBUS
;
2078 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2079 if (vmf
->pgoff
== 0)
2085 data
= rcu_dereference(counter
->data
);
2089 if (vmf
->pgoff
== 0) {
2090 vmf
->page
= virt_to_page(data
->user_page
);
2092 int nr
= vmf
->pgoff
- 1;
2094 if ((unsigned)nr
> data
->nr_pages
)
2097 if (vmf
->flags
& FAULT_FLAG_WRITE
)
2100 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
2103 get_page(vmf
->page
);
2104 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2105 vmf
->page
->index
= vmf
->pgoff
;
2114 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
2116 struct perf_mmap_data
*data
;
2120 WARN_ON(atomic_read(&counter
->mmap_count
));
2122 size
= sizeof(struct perf_mmap_data
);
2123 size
+= nr_pages
* sizeof(void *);
2125 data
= kzalloc(size
, GFP_KERNEL
);
2129 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2130 if (!data
->user_page
)
2131 goto fail_user_page
;
2133 for (i
= 0; i
< nr_pages
; i
++) {
2134 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2135 if (!data
->data_pages
[i
])
2136 goto fail_data_pages
;
2139 data
->nr_pages
= nr_pages
;
2140 atomic_set(&data
->lock
, -1);
2142 rcu_assign_pointer(counter
->data
, data
);
2147 for (i
--; i
>= 0; i
--)
2148 free_page((unsigned long)data
->data_pages
[i
]);
2150 free_page((unsigned long)data
->user_page
);
2159 static void perf_mmap_free_page(unsigned long addr
)
2161 struct page
*page
= virt_to_page((void *)addr
);
2163 page
->mapping
= NULL
;
2167 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
2169 struct perf_mmap_data
*data
;
2172 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2174 perf_mmap_free_page((unsigned long)data
->user_page
);
2175 for (i
= 0; i
< data
->nr_pages
; i
++)
2176 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2181 static void perf_mmap_data_free(struct perf_counter
*counter
)
2183 struct perf_mmap_data
*data
= counter
->data
;
2185 WARN_ON(atomic_read(&counter
->mmap_count
));
2187 rcu_assign_pointer(counter
->data
, NULL
);
2188 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
2191 static void perf_mmap_open(struct vm_area_struct
*vma
)
2193 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
2195 atomic_inc(&counter
->mmap_count
);
2198 static void perf_mmap_close(struct vm_area_struct
*vma
)
2200 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
2202 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
2203 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
, &counter
->mmap_mutex
)) {
2204 struct user_struct
*user
= current_user();
2206 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
2207 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
2208 perf_mmap_data_free(counter
);
2209 mutex_unlock(&counter
->mmap_mutex
);
2213 static struct vm_operations_struct perf_mmap_vmops
= {
2214 .open
= perf_mmap_open
,
2215 .close
= perf_mmap_close
,
2216 .fault
= perf_mmap_fault
,
2217 .page_mkwrite
= perf_mmap_fault
,
2220 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2222 struct perf_counter
*counter
= file
->private_data
;
2223 unsigned long user_locked
, user_lock_limit
;
2224 struct user_struct
*user
= current_user();
2225 unsigned long locked
, lock_limit
;
2226 unsigned long vma_size
;
2227 unsigned long nr_pages
;
2228 long user_extra
, extra
;
2231 if (!(vma
->vm_flags
& VM_SHARED
))
2234 vma_size
= vma
->vm_end
- vma
->vm_start
;
2235 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2238 * If we have data pages ensure they're a power-of-two number, so we
2239 * can do bitmasks instead of modulo.
2241 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2244 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2247 if (vma
->vm_pgoff
!= 0)
2250 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
2251 mutex_lock(&counter
->mmap_mutex
);
2252 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
2253 if (nr_pages
!= counter
->data
->nr_pages
)
2258 user_extra
= nr_pages
+ 1;
2259 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
2262 * Increase the limit linearly with more CPUs:
2264 user_lock_limit
*= num_online_cpus();
2266 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2269 if (user_locked
> user_lock_limit
)
2270 extra
= user_locked
- user_lock_limit
;
2272 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2273 lock_limit
>>= PAGE_SHIFT
;
2274 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2276 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
2281 WARN_ON(counter
->data
);
2282 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
2286 atomic_set(&counter
->mmap_count
, 1);
2287 atomic_long_add(user_extra
, &user
->locked_vm
);
2288 vma
->vm_mm
->locked_vm
+= extra
;
2289 counter
->data
->nr_locked
= extra
;
2290 if (vma
->vm_flags
& VM_WRITE
)
2291 counter
->data
->writable
= 1;
2294 mutex_unlock(&counter
->mmap_mutex
);
2296 vma
->vm_flags
|= VM_RESERVED
;
2297 vma
->vm_ops
= &perf_mmap_vmops
;
2302 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2304 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2305 struct perf_counter
*counter
= filp
->private_data
;
2308 mutex_lock(&inode
->i_mutex
);
2309 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
2310 mutex_unlock(&inode
->i_mutex
);
2318 static const struct file_operations perf_fops
= {
2319 .release
= perf_release
,
2322 .unlocked_ioctl
= perf_ioctl
,
2323 .compat_ioctl
= perf_ioctl
,
2325 .fasync
= perf_fasync
,
2329 * Perf counter wakeup
2331 * If there's data, ensure we set the poll() state and publish everything
2332 * to user-space before waking everybody up.
2335 void perf_counter_wakeup(struct perf_counter
*counter
)
2337 wake_up_all(&counter
->waitq
);
2339 if (counter
->pending_kill
) {
2340 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
2341 counter
->pending_kill
= 0;
2348 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2350 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2351 * single linked list and use cmpxchg() to add entries lockless.
2354 static void perf_pending_counter(struct perf_pending_entry
*entry
)
2356 struct perf_counter
*counter
= container_of(entry
,
2357 struct perf_counter
, pending
);
2359 if (counter
->pending_disable
) {
2360 counter
->pending_disable
= 0;
2361 __perf_counter_disable(counter
);
2364 if (counter
->pending_wakeup
) {
2365 counter
->pending_wakeup
= 0;
2366 perf_counter_wakeup(counter
);
2370 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2372 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2376 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2377 void (*func
)(struct perf_pending_entry
*))
2379 struct perf_pending_entry
**head
;
2381 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2386 head
= &get_cpu_var(perf_pending_head
);
2389 entry
->next
= *head
;
2390 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2392 set_perf_counter_pending();
2394 put_cpu_var(perf_pending_head
);
2397 static int __perf_pending_run(void)
2399 struct perf_pending_entry
*list
;
2402 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2403 while (list
!= PENDING_TAIL
) {
2404 void (*func
)(struct perf_pending_entry
*);
2405 struct perf_pending_entry
*entry
= list
;
2412 * Ensure we observe the unqueue before we issue the wakeup,
2413 * so that we won't be waiting forever.
2414 * -- see perf_not_pending().
2425 static inline int perf_not_pending(struct perf_counter
*counter
)
2428 * If we flush on whatever cpu we run, there is a chance we don't
2432 __perf_pending_run();
2436 * Ensure we see the proper queue state before going to sleep
2437 * so that we do not miss the wakeup. -- see perf_pending_handle()
2440 return counter
->pending
.next
== NULL
;
2443 static void perf_pending_sync(struct perf_counter
*counter
)
2445 wait_event(counter
->waitq
, perf_not_pending(counter
));
2448 void perf_counter_do_pending(void)
2450 __perf_pending_run();
2454 * Callchain support -- arch specific
2457 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2466 struct perf_output_handle
{
2467 struct perf_counter
*counter
;
2468 struct perf_mmap_data
*data
;
2470 unsigned long offset
;
2474 unsigned long flags
;
2477 static bool perf_output_space(struct perf_mmap_data
*data
,
2478 unsigned int offset
, unsigned int head
)
2483 if (!data
->writable
)
2486 mask
= (data
->nr_pages
<< PAGE_SHIFT
) - 1;
2488 * Userspace could choose to issue a mb() before updating the tail
2489 * pointer. So that all reads will be completed before the write is
2492 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2495 offset
= (offset
- tail
) & mask
;
2496 head
= (head
- tail
) & mask
;
2498 if ((int)(head
- offset
) < 0)
2504 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2506 atomic_set(&handle
->data
->poll
, POLL_IN
);
2509 handle
->counter
->pending_wakeup
= 1;
2510 perf_pending_queue(&handle
->counter
->pending
,
2511 perf_pending_counter
);
2513 perf_counter_wakeup(handle
->counter
);
2517 * Curious locking construct.
2519 * We need to ensure a later event doesn't publish a head when a former
2520 * event isn't done writing. However since we need to deal with NMIs we
2521 * cannot fully serialize things.
2523 * What we do is serialize between CPUs so we only have to deal with NMI
2524 * nesting on a single CPU.
2526 * We only publish the head (and generate a wakeup) when the outer-most
2529 static void perf_output_lock(struct perf_output_handle
*handle
)
2531 struct perf_mmap_data
*data
= handle
->data
;
2536 local_irq_save(handle
->flags
);
2537 cpu
= smp_processor_id();
2539 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
2542 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2548 static void perf_output_unlock(struct perf_output_handle
*handle
)
2550 struct perf_mmap_data
*data
= handle
->data
;
2554 data
->done_head
= data
->head
;
2556 if (!handle
->locked
)
2561 * The xchg implies a full barrier that ensures all writes are done
2562 * before we publish the new head, matched by a rmb() in userspace when
2563 * reading this position.
2565 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2566 data
->user_page
->data_head
= head
;
2569 * NMI can happen here, which means we can miss a done_head update.
2572 cpu
= atomic_xchg(&data
->lock
, -1);
2573 WARN_ON_ONCE(cpu
!= smp_processor_id());
2576 * Therefore we have to validate we did not indeed do so.
2578 if (unlikely(atomic_long_read(&data
->done_head
))) {
2580 * Since we had it locked, we can lock it again.
2582 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2588 if (atomic_xchg(&data
->wakeup
, 0))
2589 perf_output_wakeup(handle
);
2591 local_irq_restore(handle
->flags
);
2594 static void perf_output_copy(struct perf_output_handle
*handle
,
2595 const void *buf
, unsigned int len
)
2597 unsigned int pages_mask
;
2598 unsigned int offset
;
2602 offset
= handle
->offset
;
2603 pages_mask
= handle
->data
->nr_pages
- 1;
2604 pages
= handle
->data
->data_pages
;
2607 unsigned int page_offset
;
2610 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2611 page_offset
= offset
& (PAGE_SIZE
- 1);
2612 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
2614 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2621 handle
->offset
= offset
;
2624 * Check we didn't copy past our reservation window, taking the
2625 * possible unsigned int wrap into account.
2627 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2630 #define perf_output_put(handle, x) \
2631 perf_output_copy((handle), &(x), sizeof(x))
2633 static int perf_output_begin(struct perf_output_handle
*handle
,
2634 struct perf_counter
*counter
, unsigned int size
,
2635 int nmi
, int sample
)
2637 struct perf_mmap_data
*data
;
2638 unsigned int offset
, head
;
2641 struct perf_event_header header
;
2647 * For inherited counters we send all the output towards the parent.
2649 if (counter
->parent
)
2650 counter
= counter
->parent
;
2653 data
= rcu_dereference(counter
->data
);
2657 handle
->data
= data
;
2658 handle
->counter
= counter
;
2660 handle
->sample
= sample
;
2662 if (!data
->nr_pages
)
2665 have_lost
= atomic_read(&data
->lost
);
2667 size
+= sizeof(lost_event
);
2669 perf_output_lock(handle
);
2672 offset
= head
= atomic_long_read(&data
->head
);
2674 if (unlikely(!perf_output_space(data
, offset
, head
)))
2676 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2678 handle
->offset
= offset
;
2679 handle
->head
= head
;
2681 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
2682 atomic_set(&data
->wakeup
, 1);
2685 lost_event
.header
.type
= PERF_EVENT_LOST
;
2686 lost_event
.header
.misc
= 0;
2687 lost_event
.header
.size
= sizeof(lost_event
);
2688 lost_event
.id
= counter
->id
;
2689 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2691 perf_output_put(handle
, lost_event
);
2697 atomic_inc(&data
->lost
);
2698 perf_output_unlock(handle
);
2705 static void perf_output_end(struct perf_output_handle
*handle
)
2707 struct perf_counter
*counter
= handle
->counter
;
2708 struct perf_mmap_data
*data
= handle
->data
;
2710 int wakeup_events
= counter
->attr
.wakeup_events
;
2712 if (handle
->sample
&& wakeup_events
) {
2713 int events
= atomic_inc_return(&data
->events
);
2714 if (events
>= wakeup_events
) {
2715 atomic_sub(wakeup_events
, &data
->events
);
2716 atomic_set(&data
->wakeup
, 1);
2720 perf_output_unlock(handle
);
2724 static u32
perf_counter_pid(struct perf_counter
*counter
, struct task_struct
*p
)
2727 * only top level counters have the pid namespace they were created in
2729 if (counter
->parent
)
2730 counter
= counter
->parent
;
2732 return task_tgid_nr_ns(p
, counter
->ns
);
2735 static u32
perf_counter_tid(struct perf_counter
*counter
, struct task_struct
*p
)
2738 * only top level counters have the pid namespace they were created in
2740 if (counter
->parent
)
2741 counter
= counter
->parent
;
2743 return task_pid_nr_ns(p
, counter
->ns
);
2746 static void perf_output_read_one(struct perf_output_handle
*handle
,
2747 struct perf_counter
*counter
)
2749 u64 read_format
= counter
->attr
.read_format
;
2753 values
[n
++] = atomic64_read(&counter
->count
);
2754 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
2755 values
[n
++] = counter
->total_time_enabled
+
2756 atomic64_read(&counter
->child_total_time_enabled
);
2758 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
2759 values
[n
++] = counter
->total_time_running
+
2760 atomic64_read(&counter
->child_total_time_running
);
2762 if (read_format
& PERF_FORMAT_ID
)
2763 values
[n
++] = primary_counter_id(counter
);
2765 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2769 * XXX PERF_FORMAT_GROUP vs inherited counters seems difficult.
2771 static void perf_output_read_group(struct perf_output_handle
*handle
,
2772 struct perf_counter
*counter
)
2774 struct perf_counter
*leader
= counter
->group_leader
, *sub
;
2775 u64 read_format
= counter
->attr
.read_format
;
2779 values
[n
++] = 1 + leader
->nr_siblings
;
2781 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2782 values
[n
++] = leader
->total_time_enabled
;
2784 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2785 values
[n
++] = leader
->total_time_running
;
2787 if (leader
!= counter
)
2788 leader
->pmu
->read(leader
);
2790 values
[n
++] = atomic64_read(&leader
->count
);
2791 if (read_format
& PERF_FORMAT_ID
)
2792 values
[n
++] = primary_counter_id(leader
);
2794 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2796 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2800 sub
->pmu
->read(sub
);
2802 values
[n
++] = atomic64_read(&sub
->count
);
2803 if (read_format
& PERF_FORMAT_ID
)
2804 values
[n
++] = primary_counter_id(sub
);
2806 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2810 static void perf_output_read(struct perf_output_handle
*handle
,
2811 struct perf_counter
*counter
)
2813 if (counter
->attr
.read_format
& PERF_FORMAT_GROUP
)
2814 perf_output_read_group(handle
, counter
);
2816 perf_output_read_one(handle
, counter
);
2819 void perf_counter_output(struct perf_counter
*counter
, int nmi
,
2820 struct perf_sample_data
*data
)
2823 u64 sample_type
= counter
->attr
.sample_type
;
2824 struct perf_output_handle handle
;
2825 struct perf_event_header header
;
2830 struct perf_callchain_entry
*callchain
= NULL
;
2831 int callchain_size
= 0;
2837 header
.type
= PERF_EVENT_SAMPLE
;
2838 header
.size
= sizeof(header
);
2841 header
.misc
|= perf_misc_flags(data
->regs
);
2843 if (sample_type
& PERF_SAMPLE_IP
) {
2844 ip
= perf_instruction_pointer(data
->regs
);
2845 header
.size
+= sizeof(ip
);
2848 if (sample_type
& PERF_SAMPLE_TID
) {
2849 /* namespace issues */
2850 tid_entry
.pid
= perf_counter_pid(counter
, current
);
2851 tid_entry
.tid
= perf_counter_tid(counter
, current
);
2853 header
.size
+= sizeof(tid_entry
);
2856 if (sample_type
& PERF_SAMPLE_TIME
) {
2858 * Maybe do better on x86 and provide cpu_clock_nmi()
2860 time
= sched_clock();
2862 header
.size
+= sizeof(u64
);
2865 if (sample_type
& PERF_SAMPLE_ADDR
)
2866 header
.size
+= sizeof(u64
);
2868 if (sample_type
& PERF_SAMPLE_ID
)
2869 header
.size
+= sizeof(u64
);
2871 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
2872 header
.size
+= sizeof(u64
);
2874 if (sample_type
& PERF_SAMPLE_CPU
) {
2875 header
.size
+= sizeof(cpu_entry
);
2877 cpu_entry
.cpu
= raw_smp_processor_id();
2878 cpu_entry
.reserved
= 0;
2881 if (sample_type
& PERF_SAMPLE_PERIOD
)
2882 header
.size
+= sizeof(u64
);
2884 if (sample_type
& PERF_SAMPLE_READ
)
2885 header
.size
+= perf_counter_read_size(counter
);
2887 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2888 callchain
= perf_callchain(data
->regs
);
2891 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2892 header
.size
+= callchain_size
;
2894 header
.size
+= sizeof(u64
);
2897 if (sample_type
& PERF_SAMPLE_RAW
) {
2898 int size
= sizeof(u32
);
2901 size
+= data
->raw
->size
;
2903 size
+= sizeof(u32
);
2905 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
2906 header
.size
+= size
;
2909 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2913 perf_output_put(&handle
, header
);
2915 if (sample_type
& PERF_SAMPLE_IP
)
2916 perf_output_put(&handle
, ip
);
2918 if (sample_type
& PERF_SAMPLE_TID
)
2919 perf_output_put(&handle
, tid_entry
);
2921 if (sample_type
& PERF_SAMPLE_TIME
)
2922 perf_output_put(&handle
, time
);
2924 if (sample_type
& PERF_SAMPLE_ADDR
)
2925 perf_output_put(&handle
, data
->addr
);
2927 if (sample_type
& PERF_SAMPLE_ID
) {
2928 u64 id
= primary_counter_id(counter
);
2930 perf_output_put(&handle
, id
);
2933 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
2934 perf_output_put(&handle
, counter
->id
);
2936 if (sample_type
& PERF_SAMPLE_CPU
)
2937 perf_output_put(&handle
, cpu_entry
);
2939 if (sample_type
& PERF_SAMPLE_PERIOD
)
2940 perf_output_put(&handle
, data
->period
);
2942 if (sample_type
& PERF_SAMPLE_READ
)
2943 perf_output_read(&handle
, counter
);
2945 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2947 perf_output_copy(&handle
, callchain
, callchain_size
);
2950 perf_output_put(&handle
, nr
);
2954 if (sample_type
& PERF_SAMPLE_RAW
) {
2956 perf_output_put(&handle
, data
->raw
->size
);
2957 perf_output_copy(&handle
, data
->raw
->data
, data
->raw
->size
);
2963 .size
= sizeof(u32
),
2966 perf_output_put(&handle
, raw
);
2970 perf_output_end(&handle
);
2977 struct perf_read_event
{
2978 struct perf_event_header header
;
2985 perf_counter_read_event(struct perf_counter
*counter
,
2986 struct task_struct
*task
)
2988 struct perf_output_handle handle
;
2989 struct perf_read_event event
= {
2991 .type
= PERF_EVENT_READ
,
2993 .size
= sizeof(event
) + perf_counter_read_size(counter
),
2995 .pid
= perf_counter_pid(counter
, task
),
2996 .tid
= perf_counter_tid(counter
, task
),
3000 ret
= perf_output_begin(&handle
, counter
, event
.header
.size
, 0, 0);
3004 perf_output_put(&handle
, event
);
3005 perf_output_read(&handle
, counter
);
3007 perf_output_end(&handle
);
3011 * task tracking -- fork/exit
3013 * enabled by: attr.comm | attr.mmap | attr.task
3016 struct perf_task_event
{
3017 struct task_struct
*task
;
3018 struct perf_counter_context
*task_ctx
;
3021 struct perf_event_header header
;
3030 static void perf_counter_task_output(struct perf_counter
*counter
,
3031 struct perf_task_event
*task_event
)
3033 struct perf_output_handle handle
;
3034 int size
= task_event
->event
.header
.size
;
3035 struct task_struct
*task
= task_event
->task
;
3036 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
3041 task_event
->event
.pid
= perf_counter_pid(counter
, task
);
3042 task_event
->event
.ppid
= perf_counter_pid(counter
, current
);
3044 task_event
->event
.tid
= perf_counter_tid(counter
, task
);
3045 task_event
->event
.ptid
= perf_counter_tid(counter
, current
);
3047 perf_output_put(&handle
, task_event
->event
);
3048 perf_output_end(&handle
);
3051 static int perf_counter_task_match(struct perf_counter
*counter
)
3053 if (counter
->attr
.comm
|| counter
->attr
.mmap
|| counter
->attr
.task
)
3059 static void perf_counter_task_ctx(struct perf_counter_context
*ctx
,
3060 struct perf_task_event
*task_event
)
3062 struct perf_counter
*counter
;
3064 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3068 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3069 if (perf_counter_task_match(counter
))
3070 perf_counter_task_output(counter
, task_event
);
3075 static void perf_counter_task_event(struct perf_task_event
*task_event
)
3077 struct perf_cpu_context
*cpuctx
;
3078 struct perf_counter_context
*ctx
= task_event
->task_ctx
;
3080 cpuctx
= &get_cpu_var(perf_cpu_context
);
3081 perf_counter_task_ctx(&cpuctx
->ctx
, task_event
);
3082 put_cpu_var(perf_cpu_context
);
3086 ctx
= rcu_dereference(task_event
->task
->perf_counter_ctxp
);
3088 perf_counter_task_ctx(ctx
, task_event
);
3092 static void perf_counter_task(struct task_struct
*task
,
3093 struct perf_counter_context
*task_ctx
,
3096 struct perf_task_event task_event
;
3098 if (!atomic_read(&nr_comm_counters
) &&
3099 !atomic_read(&nr_mmap_counters
) &&
3100 !atomic_read(&nr_task_counters
))
3103 task_event
= (struct perf_task_event
){
3105 .task_ctx
= task_ctx
,
3108 .type
= new ? PERF_EVENT_FORK
: PERF_EVENT_EXIT
,
3110 .size
= sizeof(task_event
.event
),
3119 perf_counter_task_event(&task_event
);
3122 void perf_counter_fork(struct task_struct
*task
)
3124 perf_counter_task(task
, NULL
, 1);
3131 struct perf_comm_event
{
3132 struct task_struct
*task
;
3137 struct perf_event_header header
;
3144 static void perf_counter_comm_output(struct perf_counter
*counter
,
3145 struct perf_comm_event
*comm_event
)
3147 struct perf_output_handle handle
;
3148 int size
= comm_event
->event
.header
.size
;
3149 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
3154 comm_event
->event
.pid
= perf_counter_pid(counter
, comm_event
->task
);
3155 comm_event
->event
.tid
= perf_counter_tid(counter
, comm_event
->task
);
3157 perf_output_put(&handle
, comm_event
->event
);
3158 perf_output_copy(&handle
, comm_event
->comm
,
3159 comm_event
->comm_size
);
3160 perf_output_end(&handle
);
3163 static int perf_counter_comm_match(struct perf_counter
*counter
)
3165 if (counter
->attr
.comm
)
3171 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
3172 struct perf_comm_event
*comm_event
)
3174 struct perf_counter
*counter
;
3176 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3180 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3181 if (perf_counter_comm_match(counter
))
3182 perf_counter_comm_output(counter
, comm_event
);
3187 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
3189 struct perf_cpu_context
*cpuctx
;
3190 struct perf_counter_context
*ctx
;
3192 char comm
[TASK_COMM_LEN
];
3194 memset(comm
, 0, sizeof(comm
));
3195 strncpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3196 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3198 comm_event
->comm
= comm
;
3199 comm_event
->comm_size
= size
;
3201 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
3203 cpuctx
= &get_cpu_var(perf_cpu_context
);
3204 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
3205 put_cpu_var(perf_cpu_context
);
3209 * doesn't really matter which of the child contexts the
3210 * events ends up in.
3212 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3214 perf_counter_comm_ctx(ctx
, comm_event
);
3218 void perf_counter_comm(struct task_struct
*task
)
3220 struct perf_comm_event comm_event
;
3222 if (task
->perf_counter_ctxp
)
3223 perf_counter_enable_on_exec(task
);
3225 if (!atomic_read(&nr_comm_counters
))
3228 comm_event
= (struct perf_comm_event
){
3234 .type
= PERF_EVENT_COMM
,
3243 perf_counter_comm_event(&comm_event
);
3250 struct perf_mmap_event
{
3251 struct vm_area_struct
*vma
;
3253 const char *file_name
;
3257 struct perf_event_header header
;
3267 static void perf_counter_mmap_output(struct perf_counter
*counter
,
3268 struct perf_mmap_event
*mmap_event
)
3270 struct perf_output_handle handle
;
3271 int size
= mmap_event
->event
.header
.size
;
3272 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
3277 mmap_event
->event
.pid
= perf_counter_pid(counter
, current
);
3278 mmap_event
->event
.tid
= perf_counter_tid(counter
, current
);
3280 perf_output_put(&handle
, mmap_event
->event
);
3281 perf_output_copy(&handle
, mmap_event
->file_name
,
3282 mmap_event
->file_size
);
3283 perf_output_end(&handle
);
3286 static int perf_counter_mmap_match(struct perf_counter
*counter
,
3287 struct perf_mmap_event
*mmap_event
)
3289 if (counter
->attr
.mmap
)
3295 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
3296 struct perf_mmap_event
*mmap_event
)
3298 struct perf_counter
*counter
;
3300 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3304 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3305 if (perf_counter_mmap_match(counter
, mmap_event
))
3306 perf_counter_mmap_output(counter
, mmap_event
);
3311 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
3313 struct perf_cpu_context
*cpuctx
;
3314 struct perf_counter_context
*ctx
;
3315 struct vm_area_struct
*vma
= mmap_event
->vma
;
3316 struct file
*file
= vma
->vm_file
;
3322 memset(tmp
, 0, sizeof(tmp
));
3326 * d_path works from the end of the buffer backwards, so we
3327 * need to add enough zero bytes after the string to handle
3328 * the 64bit alignment we do later.
3330 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3332 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3335 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3337 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3341 if (arch_vma_name(mmap_event
->vma
)) {
3342 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3348 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3352 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3357 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3359 mmap_event
->file_name
= name
;
3360 mmap_event
->file_size
= size
;
3362 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
3364 cpuctx
= &get_cpu_var(perf_cpu_context
);
3365 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3366 put_cpu_var(perf_cpu_context
);
3370 * doesn't really matter which of the child contexts the
3371 * events ends up in.
3373 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3375 perf_counter_mmap_ctx(ctx
, mmap_event
);
3381 void __perf_counter_mmap(struct vm_area_struct
*vma
)
3383 struct perf_mmap_event mmap_event
;
3385 if (!atomic_read(&nr_mmap_counters
))
3388 mmap_event
= (struct perf_mmap_event
){
3394 .type
= PERF_EVENT_MMAP
,
3400 .start
= vma
->vm_start
,
3401 .len
= vma
->vm_end
- vma
->vm_start
,
3402 .pgoff
= vma
->vm_pgoff
,
3406 perf_counter_mmap_event(&mmap_event
);
3410 * IRQ throttle logging
3413 static void perf_log_throttle(struct perf_counter
*counter
, int enable
)
3415 struct perf_output_handle handle
;
3419 struct perf_event_header header
;
3423 } throttle_event
= {
3425 .type
= PERF_EVENT_THROTTLE
,
3427 .size
= sizeof(throttle_event
),
3429 .time
= sched_clock(),
3430 .id
= primary_counter_id(counter
),
3431 .stream_id
= counter
->id
,
3435 throttle_event
.header
.type
= PERF_EVENT_UNTHROTTLE
;
3437 ret
= perf_output_begin(&handle
, counter
, sizeof(throttle_event
), 1, 0);
3441 perf_output_put(&handle
, throttle_event
);
3442 perf_output_end(&handle
);
3446 * Generic counter overflow handling, sampling.
3449 int perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
3450 struct perf_sample_data
*data
)
3452 int events
= atomic_read(&counter
->event_limit
);
3453 int throttle
= counter
->pmu
->unthrottle
!= NULL
;
3454 struct hw_perf_counter
*hwc
= &counter
->hw
;
3460 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3462 if (HZ
* hwc
->interrupts
>
3463 (u64
)sysctl_perf_counter_sample_rate
) {
3464 hwc
->interrupts
= MAX_INTERRUPTS
;
3465 perf_log_throttle(counter
, 0);
3470 * Keep re-disabling counters even though on the previous
3471 * pass we disabled it - just in case we raced with a
3472 * sched-in and the counter got enabled again:
3478 if (counter
->attr
.freq
) {
3479 u64 now
= sched_clock();
3480 s64 delta
= now
- hwc
->freq_stamp
;
3482 hwc
->freq_stamp
= now
;
3484 if (delta
> 0 && delta
< TICK_NSEC
)
3485 perf_adjust_period(counter
, NSEC_PER_SEC
/ (int)delta
);
3489 * XXX event_limit might not quite work as expected on inherited
3493 counter
->pending_kill
= POLL_IN
;
3494 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
3496 counter
->pending_kill
= POLL_HUP
;
3498 counter
->pending_disable
= 1;
3499 perf_pending_queue(&counter
->pending
,
3500 perf_pending_counter
);
3502 perf_counter_disable(counter
);
3505 perf_counter_output(counter
, nmi
, data
);
3510 * Generic software counter infrastructure
3514 * We directly increment counter->count and keep a second value in
3515 * counter->hw.period_left to count intervals. This period counter
3516 * is kept in the range [-sample_period, 0] so that we can use the
3520 static u64
perf_swcounter_set_period(struct perf_counter
*counter
)
3522 struct hw_perf_counter
*hwc
= &counter
->hw
;
3523 u64 period
= hwc
->last_period
;
3527 hwc
->last_period
= hwc
->sample_period
;
3530 old
= val
= atomic64_read(&hwc
->period_left
);
3534 nr
= div64_u64(period
+ val
, period
);
3535 offset
= nr
* period
;
3537 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3543 static void perf_swcounter_overflow(struct perf_counter
*counter
,
3544 int nmi
, struct perf_sample_data
*data
)
3546 struct hw_perf_counter
*hwc
= &counter
->hw
;
3549 data
->period
= counter
->hw
.last_period
;
3550 overflow
= perf_swcounter_set_period(counter
);
3552 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3555 for (; overflow
; overflow
--) {
3556 if (perf_counter_overflow(counter
, nmi
, data
)) {
3558 * We inhibit the overflow from happening when
3559 * hwc->interrupts == MAX_INTERRUPTS.
3566 static void perf_swcounter_unthrottle(struct perf_counter
*counter
)
3569 * Nothing to do, we already reset hwc->interrupts.
3573 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
3574 int nmi
, struct perf_sample_data
*data
)
3576 struct hw_perf_counter
*hwc
= &counter
->hw
;
3578 atomic64_add(nr
, &counter
->count
);
3580 if (!hwc
->sample_period
)
3586 if (!atomic64_add_negative(nr
, &hwc
->period_left
))
3587 perf_swcounter_overflow(counter
, nmi
, data
);
3590 static int perf_swcounter_is_counting(struct perf_counter
*counter
)
3593 * The counter is active, we're good!
3595 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
3599 * The counter is off/error, not counting.
3601 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
)
3605 * The counter is inactive, if the context is active
3606 * we're part of a group that didn't make it on the 'pmu',
3609 if (counter
->ctx
->is_active
)
3613 * We're inactive and the context is too, this means the
3614 * task is scheduled out, we're counting events that happen
3615 * to us, like migration events.
3620 static int perf_swcounter_match(struct perf_counter
*counter
,
3621 enum perf_type_id type
,
3622 u32 event
, struct pt_regs
*regs
)
3624 if (!perf_swcounter_is_counting(counter
))
3627 if (counter
->attr
.type
!= type
)
3629 if (counter
->attr
.config
!= event
)
3633 if (counter
->attr
.exclude_user
&& user_mode(regs
))
3636 if (counter
->attr
.exclude_kernel
&& !user_mode(regs
))
3643 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
3644 enum perf_type_id type
,
3645 u32 event
, u64 nr
, int nmi
,
3646 struct perf_sample_data
*data
)
3648 struct perf_counter
*counter
;
3650 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3654 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3655 if (perf_swcounter_match(counter
, type
, event
, data
->regs
))
3656 perf_swcounter_add(counter
, nr
, nmi
, data
);
3661 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
3664 return &cpuctx
->recursion
[3];
3667 return &cpuctx
->recursion
[2];
3670 return &cpuctx
->recursion
[1];
3672 return &cpuctx
->recursion
[0];
3675 static void do_perf_swcounter_event(enum perf_type_id type
, u32 event
,
3677 struct perf_sample_data
*data
)
3679 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3680 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
3681 struct perf_counter_context
*ctx
;
3689 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
3693 * doesn't really matter which of the child contexts the
3694 * events ends up in.
3696 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3698 perf_swcounter_ctx_event(ctx
, type
, event
, nr
, nmi
, data
);
3705 put_cpu_var(perf_cpu_context
);
3708 void __perf_swcounter_event(u32 event
, u64 nr
, int nmi
,
3709 struct pt_regs
*regs
, u64 addr
)
3711 struct perf_sample_data data
= {
3716 do_perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, &data
);
3719 static void perf_swcounter_read(struct perf_counter
*counter
)
3723 static int perf_swcounter_enable(struct perf_counter
*counter
)
3725 struct hw_perf_counter
*hwc
= &counter
->hw
;
3727 if (hwc
->sample_period
) {
3728 hwc
->last_period
= hwc
->sample_period
;
3729 perf_swcounter_set_period(counter
);
3734 static void perf_swcounter_disable(struct perf_counter
*counter
)
3738 static const struct pmu perf_ops_generic
= {
3739 .enable
= perf_swcounter_enable
,
3740 .disable
= perf_swcounter_disable
,
3741 .read
= perf_swcounter_read
,
3742 .unthrottle
= perf_swcounter_unthrottle
,
3746 * hrtimer based swcounter callback
3749 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
3751 enum hrtimer_restart ret
= HRTIMER_RESTART
;
3752 struct perf_sample_data data
;
3753 struct perf_counter
*counter
;
3756 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
3757 counter
->pmu
->read(counter
);
3760 data
.regs
= get_irq_regs();
3762 * In case we exclude kernel IPs or are somehow not in interrupt
3763 * context, provide the next best thing, the user IP.
3765 if ((counter
->attr
.exclude_kernel
|| !data
.regs
) &&
3766 !counter
->attr
.exclude_user
)
3767 data
.regs
= task_pt_regs(current
);
3770 if (perf_counter_overflow(counter
, 0, &data
))
3771 ret
= HRTIMER_NORESTART
;
3774 period
= max_t(u64
, 10000, counter
->hw
.sample_period
);
3775 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
3781 * Software counter: cpu wall time clock
3784 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
3786 int cpu
= raw_smp_processor_id();
3790 now
= cpu_clock(cpu
);
3791 prev
= atomic64_read(&counter
->hw
.prev_count
);
3792 atomic64_set(&counter
->hw
.prev_count
, now
);
3793 atomic64_add(now
- prev
, &counter
->count
);
3796 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
3798 struct hw_perf_counter
*hwc
= &counter
->hw
;
3799 int cpu
= raw_smp_processor_id();
3801 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
3802 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3803 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3804 if (hwc
->sample_period
) {
3805 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3806 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3807 ns_to_ktime(period
), 0,
3808 HRTIMER_MODE_REL
, 0);
3814 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
3816 if (counter
->hw
.sample_period
)
3817 hrtimer_cancel(&counter
->hw
.hrtimer
);
3818 cpu_clock_perf_counter_update(counter
);
3821 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
3823 cpu_clock_perf_counter_update(counter
);
3826 static const struct pmu perf_ops_cpu_clock
= {
3827 .enable
= cpu_clock_perf_counter_enable
,
3828 .disable
= cpu_clock_perf_counter_disable
,
3829 .read
= cpu_clock_perf_counter_read
,
3833 * Software counter: task time clock
3836 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
3841 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
3843 atomic64_add(delta
, &counter
->count
);
3846 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
3848 struct hw_perf_counter
*hwc
= &counter
->hw
;
3851 now
= counter
->ctx
->time
;
3853 atomic64_set(&hwc
->prev_count
, now
);
3854 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3855 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3856 if (hwc
->sample_period
) {
3857 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3858 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3859 ns_to_ktime(period
), 0,
3860 HRTIMER_MODE_REL
, 0);
3866 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
3868 if (counter
->hw
.sample_period
)
3869 hrtimer_cancel(&counter
->hw
.hrtimer
);
3870 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
3874 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
3879 update_context_time(counter
->ctx
);
3880 time
= counter
->ctx
->time
;
3882 u64 now
= perf_clock();
3883 u64 delta
= now
- counter
->ctx
->timestamp
;
3884 time
= counter
->ctx
->time
+ delta
;
3887 task_clock_perf_counter_update(counter
, time
);
3890 static const struct pmu perf_ops_task_clock
= {
3891 .enable
= task_clock_perf_counter_enable
,
3892 .disable
= task_clock_perf_counter_disable
,
3893 .read
= task_clock_perf_counter_read
,
3896 #ifdef CONFIG_EVENT_PROFILE
3897 void perf_tpcounter_event(int event_id
, u64 addr
, u64 count
, void *record
,
3900 struct perf_raw_record raw
= {
3905 struct perf_sample_data data
= {
3906 .regs
= get_irq_regs(),
3912 data
.regs
= task_pt_regs(current
);
3914 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1, &data
);
3916 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
3918 extern int ftrace_profile_enable(int);
3919 extern void ftrace_profile_disable(int);
3921 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
3923 ftrace_profile_disable(counter
->attr
.config
);
3926 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3929 * Raw tracepoint data is a severe data leak, only allow root to
3932 if ((counter
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
3933 !capable(CAP_SYS_ADMIN
))
3934 return ERR_PTR(-EPERM
);
3936 if (ftrace_profile_enable(counter
->attr
.config
))
3939 counter
->destroy
= tp_perf_counter_destroy
;
3941 return &perf_ops_generic
;
3944 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3950 atomic_t perf_swcounter_enabled
[PERF_COUNT_SW_MAX
];
3952 static void sw_perf_counter_destroy(struct perf_counter
*counter
)
3954 u64 event
= counter
->attr
.config
;
3956 WARN_ON(counter
->parent
);
3958 atomic_dec(&perf_swcounter_enabled
[event
]);
3961 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
3963 const struct pmu
*pmu
= NULL
;
3964 u64 event
= counter
->attr
.config
;
3967 * Software counters (currently) can't in general distinguish
3968 * between user, kernel and hypervisor events.
3969 * However, context switches and cpu migrations are considered
3970 * to be kernel events, and page faults are never hypervisor
3974 case PERF_COUNT_SW_CPU_CLOCK
:
3975 pmu
= &perf_ops_cpu_clock
;
3978 case PERF_COUNT_SW_TASK_CLOCK
:
3980 * If the user instantiates this as a per-cpu counter,
3981 * use the cpu_clock counter instead.
3983 if (counter
->ctx
->task
)
3984 pmu
= &perf_ops_task_clock
;
3986 pmu
= &perf_ops_cpu_clock
;
3989 case PERF_COUNT_SW_PAGE_FAULTS
:
3990 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
3991 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
3992 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
3993 case PERF_COUNT_SW_CPU_MIGRATIONS
:
3994 if (!counter
->parent
) {
3995 atomic_inc(&perf_swcounter_enabled
[event
]);
3996 counter
->destroy
= sw_perf_counter_destroy
;
3998 pmu
= &perf_ops_generic
;
4006 * Allocate and initialize a counter structure
4008 static struct perf_counter
*
4009 perf_counter_alloc(struct perf_counter_attr
*attr
,
4011 struct perf_counter_context
*ctx
,
4012 struct perf_counter
*group_leader
,
4013 struct perf_counter
*parent_counter
,
4016 const struct pmu
*pmu
;
4017 struct perf_counter
*counter
;
4018 struct hw_perf_counter
*hwc
;
4021 counter
= kzalloc(sizeof(*counter
), gfpflags
);
4023 return ERR_PTR(-ENOMEM
);
4026 * Single counters are their own group leaders, with an
4027 * empty sibling list:
4030 group_leader
= counter
;
4032 mutex_init(&counter
->child_mutex
);
4033 INIT_LIST_HEAD(&counter
->child_list
);
4035 INIT_LIST_HEAD(&counter
->list_entry
);
4036 INIT_LIST_HEAD(&counter
->event_entry
);
4037 INIT_LIST_HEAD(&counter
->sibling_list
);
4038 init_waitqueue_head(&counter
->waitq
);
4040 mutex_init(&counter
->mmap_mutex
);
4043 counter
->attr
= *attr
;
4044 counter
->group_leader
= group_leader
;
4045 counter
->pmu
= NULL
;
4047 counter
->oncpu
= -1;
4049 counter
->parent
= parent_counter
;
4051 counter
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4052 counter
->id
= atomic64_inc_return(&perf_counter_id
);
4054 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
4057 counter
->state
= PERF_COUNTER_STATE_OFF
;
4062 hwc
->sample_period
= attr
->sample_period
;
4063 if (attr
->freq
&& attr
->sample_freq
)
4064 hwc
->sample_period
= 1;
4066 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4069 * we currently do not support PERF_FORMAT_GROUP on inherited counters
4071 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4074 switch (attr
->type
) {
4076 case PERF_TYPE_HARDWARE
:
4077 case PERF_TYPE_HW_CACHE
:
4078 pmu
= hw_perf_counter_init(counter
);
4081 case PERF_TYPE_SOFTWARE
:
4082 pmu
= sw_perf_counter_init(counter
);
4085 case PERF_TYPE_TRACEPOINT
:
4086 pmu
= tp_perf_counter_init(counter
);
4096 else if (IS_ERR(pmu
))
4101 put_pid_ns(counter
->ns
);
4103 return ERR_PTR(err
);
4108 if (!counter
->parent
) {
4109 atomic_inc(&nr_counters
);
4110 if (counter
->attr
.mmap
)
4111 atomic_inc(&nr_mmap_counters
);
4112 if (counter
->attr
.comm
)
4113 atomic_inc(&nr_comm_counters
);
4114 if (counter
->attr
.task
)
4115 atomic_inc(&nr_task_counters
);
4121 static int perf_copy_attr(struct perf_counter_attr __user
*uattr
,
4122 struct perf_counter_attr
*attr
)
4127 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4131 * zero the full structure, so that a short copy will be nice.
4133 memset(attr
, 0, sizeof(*attr
));
4135 ret
= get_user(size
, &uattr
->size
);
4139 if (size
> PAGE_SIZE
) /* silly large */
4142 if (!size
) /* abi compat */
4143 size
= PERF_ATTR_SIZE_VER0
;
4145 if (size
< PERF_ATTR_SIZE_VER0
)
4149 * If we're handed a bigger struct than we know of,
4150 * ensure all the unknown bits are 0.
4152 if (size
> sizeof(*attr
)) {
4154 unsigned long __user
*addr
;
4155 unsigned long __user
*end
;
4157 addr
= PTR_ALIGN((void __user
*)uattr
+ sizeof(*attr
),
4158 sizeof(unsigned long));
4159 end
= PTR_ALIGN((void __user
*)uattr
+ size
,
4160 sizeof(unsigned long));
4162 for (; addr
< end
; addr
+= sizeof(unsigned long)) {
4163 ret
= get_user(val
, addr
);
4171 ret
= copy_from_user(attr
, uattr
, size
);
4176 * If the type exists, the corresponding creation will verify
4179 if (attr
->type
>= PERF_TYPE_MAX
)
4182 if (attr
->__reserved_1
|| attr
->__reserved_2
|| attr
->__reserved_3
)
4185 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4188 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4195 put_user(sizeof(*attr
), &uattr
->size
);
4201 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
4203 * @attr_uptr: event type attributes for monitoring/sampling
4206 * @group_fd: group leader counter fd
4208 SYSCALL_DEFINE5(perf_counter_open
,
4209 struct perf_counter_attr __user
*, attr_uptr
,
4210 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4212 struct perf_counter
*counter
, *group_leader
;
4213 struct perf_counter_attr attr
;
4214 struct perf_counter_context
*ctx
;
4215 struct file
*counter_file
= NULL
;
4216 struct file
*group_file
= NULL
;
4217 int fput_needed
= 0;
4218 int fput_needed2
= 0;
4221 /* for future expandability... */
4225 ret
= perf_copy_attr(attr_uptr
, &attr
);
4229 if (!attr
.exclude_kernel
) {
4230 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4235 if (attr
.sample_freq
> sysctl_perf_counter_sample_rate
)
4240 * Get the target context (task or percpu):
4242 ctx
= find_get_context(pid
, cpu
);
4244 return PTR_ERR(ctx
);
4247 * Look up the group leader (we will attach this counter to it):
4249 group_leader
= NULL
;
4250 if (group_fd
!= -1) {
4252 group_file
= fget_light(group_fd
, &fput_needed
);
4254 goto err_put_context
;
4255 if (group_file
->f_op
!= &perf_fops
)
4256 goto err_put_context
;
4258 group_leader
= group_file
->private_data
;
4260 * Do not allow a recursive hierarchy (this new sibling
4261 * becoming part of another group-sibling):
4263 if (group_leader
->group_leader
!= group_leader
)
4264 goto err_put_context
;
4266 * Do not allow to attach to a group in a different
4267 * task or CPU context:
4269 if (group_leader
->ctx
!= ctx
)
4270 goto err_put_context
;
4272 * Only a group leader can be exclusive or pinned
4274 if (attr
.exclusive
|| attr
.pinned
)
4275 goto err_put_context
;
4278 counter
= perf_counter_alloc(&attr
, cpu
, ctx
, group_leader
,
4280 ret
= PTR_ERR(counter
);
4281 if (IS_ERR(counter
))
4282 goto err_put_context
;
4284 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
4286 goto err_free_put_context
;
4288 counter_file
= fget_light(ret
, &fput_needed2
);
4290 goto err_free_put_context
;
4292 counter
->filp
= counter_file
;
4293 WARN_ON_ONCE(ctx
->parent_ctx
);
4294 mutex_lock(&ctx
->mutex
);
4295 perf_install_in_context(ctx
, counter
, cpu
);
4297 mutex_unlock(&ctx
->mutex
);
4299 counter
->owner
= current
;
4300 get_task_struct(current
);
4301 mutex_lock(¤t
->perf_counter_mutex
);
4302 list_add_tail(&counter
->owner_entry
, ¤t
->perf_counter_list
);
4303 mutex_unlock(¤t
->perf_counter_mutex
);
4305 fput_light(counter_file
, fput_needed2
);
4308 fput_light(group_file
, fput_needed
);
4312 err_free_put_context
:
4322 * inherit a counter from parent task to child task:
4324 static struct perf_counter
*
4325 inherit_counter(struct perf_counter
*parent_counter
,
4326 struct task_struct
*parent
,
4327 struct perf_counter_context
*parent_ctx
,
4328 struct task_struct
*child
,
4329 struct perf_counter
*group_leader
,
4330 struct perf_counter_context
*child_ctx
)
4332 struct perf_counter
*child_counter
;
4335 * Instead of creating recursive hierarchies of counters,
4336 * we link inherited counters back to the original parent,
4337 * which has a filp for sure, which we use as the reference
4340 if (parent_counter
->parent
)
4341 parent_counter
= parent_counter
->parent
;
4343 child_counter
= perf_counter_alloc(&parent_counter
->attr
,
4344 parent_counter
->cpu
, child_ctx
,
4345 group_leader
, parent_counter
,
4347 if (IS_ERR(child_counter
))
4348 return child_counter
;
4352 * Make the child state follow the state of the parent counter,
4353 * not its attr.disabled bit. We hold the parent's mutex,
4354 * so we won't race with perf_counter_{en, dis}able_family.
4356 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
4357 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
4359 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
4361 if (parent_counter
->attr
.freq
)
4362 child_counter
->hw
.sample_period
= parent_counter
->hw
.sample_period
;
4365 * Link it up in the child's context:
4367 add_counter_to_ctx(child_counter
, child_ctx
);
4370 * Get a reference to the parent filp - we will fput it
4371 * when the child counter exits. This is safe to do because
4372 * we are in the parent and we know that the filp still
4373 * exists and has a nonzero count:
4375 atomic_long_inc(&parent_counter
->filp
->f_count
);
4378 * Link this into the parent counter's child list
4380 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
4381 mutex_lock(&parent_counter
->child_mutex
);
4382 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
4383 mutex_unlock(&parent_counter
->child_mutex
);
4385 return child_counter
;
4388 static int inherit_group(struct perf_counter
*parent_counter
,
4389 struct task_struct
*parent
,
4390 struct perf_counter_context
*parent_ctx
,
4391 struct task_struct
*child
,
4392 struct perf_counter_context
*child_ctx
)
4394 struct perf_counter
*leader
;
4395 struct perf_counter
*sub
;
4396 struct perf_counter
*child_ctr
;
4398 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
4399 child
, NULL
, child_ctx
);
4401 return PTR_ERR(leader
);
4402 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
4403 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
4404 child
, leader
, child_ctx
);
4405 if (IS_ERR(child_ctr
))
4406 return PTR_ERR(child_ctr
);
4411 static void sync_child_counter(struct perf_counter
*child_counter
,
4412 struct task_struct
*child
)
4414 struct perf_counter
*parent_counter
= child_counter
->parent
;
4417 if (child_counter
->attr
.inherit_stat
)
4418 perf_counter_read_event(child_counter
, child
);
4420 child_val
= atomic64_read(&child_counter
->count
);
4423 * Add back the child's count to the parent's count:
4425 atomic64_add(child_val
, &parent_counter
->count
);
4426 atomic64_add(child_counter
->total_time_enabled
,
4427 &parent_counter
->child_total_time_enabled
);
4428 atomic64_add(child_counter
->total_time_running
,
4429 &parent_counter
->child_total_time_running
);
4432 * Remove this counter from the parent's list
4434 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
4435 mutex_lock(&parent_counter
->child_mutex
);
4436 list_del_init(&child_counter
->child_list
);
4437 mutex_unlock(&parent_counter
->child_mutex
);
4440 * Release the parent counter, if this was the last
4443 fput(parent_counter
->filp
);
4447 __perf_counter_exit_task(struct perf_counter
*child_counter
,
4448 struct perf_counter_context
*child_ctx
,
4449 struct task_struct
*child
)
4451 struct perf_counter
*parent_counter
;
4453 update_counter_times(child_counter
);
4454 perf_counter_remove_from_context(child_counter
);
4456 parent_counter
= child_counter
->parent
;
4458 * It can happen that parent exits first, and has counters
4459 * that are still around due to the child reference. These
4460 * counters need to be zapped - but otherwise linger.
4462 if (parent_counter
) {
4463 sync_child_counter(child_counter
, child
);
4464 free_counter(child_counter
);
4469 * When a child task exits, feed back counter values to parent counters.
4471 void perf_counter_exit_task(struct task_struct
*child
)
4473 struct perf_counter
*child_counter
, *tmp
;
4474 struct perf_counter_context
*child_ctx
;
4475 unsigned long flags
;
4477 if (likely(!child
->perf_counter_ctxp
)) {
4478 perf_counter_task(child
, NULL
, 0);
4482 local_irq_save(flags
);
4484 * We can't reschedule here because interrupts are disabled,
4485 * and either child is current or it is a task that can't be
4486 * scheduled, so we are now safe from rescheduling changing
4489 child_ctx
= child
->perf_counter_ctxp
;
4490 __perf_counter_task_sched_out(child_ctx
);
4493 * Take the context lock here so that if find_get_context is
4494 * reading child->perf_counter_ctxp, we wait until it has
4495 * incremented the context's refcount before we do put_ctx below.
4497 spin_lock(&child_ctx
->lock
);
4498 child
->perf_counter_ctxp
= NULL
;
4500 * If this context is a clone; unclone it so it can't get
4501 * swapped to another process while we're removing all
4502 * the counters from it.
4504 unclone_ctx(child_ctx
);
4505 spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
4508 * Report the task dead after unscheduling the counters so that we
4509 * won't get any samples after PERF_EVENT_EXIT. We can however still
4510 * get a few PERF_EVENT_READ events.
4512 perf_counter_task(child
, child_ctx
, 0);
4515 * We can recurse on the same lock type through:
4517 * __perf_counter_exit_task()
4518 * sync_child_counter()
4519 * fput(parent_counter->filp)
4521 * mutex_lock(&ctx->mutex)
4523 * But since its the parent context it won't be the same instance.
4525 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
4528 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
4530 __perf_counter_exit_task(child_counter
, child_ctx
, child
);
4533 * If the last counter was a group counter, it will have appended all
4534 * its siblings to the list, but we obtained 'tmp' before that which
4535 * will still point to the list head terminating the iteration.
4537 if (!list_empty(&child_ctx
->counter_list
))
4540 mutex_unlock(&child_ctx
->mutex
);
4546 * free an unexposed, unused context as created by inheritance by
4547 * init_task below, used by fork() in case of fail.
4549 void perf_counter_free_task(struct task_struct
*task
)
4551 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
4552 struct perf_counter
*counter
, *tmp
;
4557 mutex_lock(&ctx
->mutex
);
4559 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
) {
4560 struct perf_counter
*parent
= counter
->parent
;
4562 if (WARN_ON_ONCE(!parent
))
4565 mutex_lock(&parent
->child_mutex
);
4566 list_del_init(&counter
->child_list
);
4567 mutex_unlock(&parent
->child_mutex
);
4571 list_del_counter(counter
, ctx
);
4572 free_counter(counter
);
4575 if (!list_empty(&ctx
->counter_list
))
4578 mutex_unlock(&ctx
->mutex
);
4584 * Initialize the perf_counter context in task_struct
4586 int perf_counter_init_task(struct task_struct
*child
)
4588 struct perf_counter_context
*child_ctx
, *parent_ctx
;
4589 struct perf_counter_context
*cloned_ctx
;
4590 struct perf_counter
*counter
;
4591 struct task_struct
*parent
= current
;
4592 int inherited_all
= 1;
4595 child
->perf_counter_ctxp
= NULL
;
4597 mutex_init(&child
->perf_counter_mutex
);
4598 INIT_LIST_HEAD(&child
->perf_counter_list
);
4600 if (likely(!parent
->perf_counter_ctxp
))
4604 * This is executed from the parent task context, so inherit
4605 * counters that have been marked for cloning.
4606 * First allocate and initialize a context for the child.
4609 child_ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
4613 __perf_counter_init_context(child_ctx
, child
);
4614 child
->perf_counter_ctxp
= child_ctx
;
4615 get_task_struct(child
);
4618 * If the parent's context is a clone, pin it so it won't get
4621 parent_ctx
= perf_pin_task_context(parent
);
4624 * No need to check if parent_ctx != NULL here; since we saw
4625 * it non-NULL earlier, the only reason for it to become NULL
4626 * is if we exit, and since we're currently in the middle of
4627 * a fork we can't be exiting at the same time.
4631 * Lock the parent list. No need to lock the child - not PID
4632 * hashed yet and not running, so nobody can access it.
4634 mutex_lock(&parent_ctx
->mutex
);
4637 * We dont have to disable NMIs - we are only looking at
4638 * the list, not manipulating it:
4640 list_for_each_entry_rcu(counter
, &parent_ctx
->event_list
, event_entry
) {
4641 if (counter
!= counter
->group_leader
)
4644 if (!counter
->attr
.inherit
) {
4649 ret
= inherit_group(counter
, parent
, parent_ctx
,
4657 if (inherited_all
) {
4659 * Mark the child context as a clone of the parent
4660 * context, or of whatever the parent is a clone of.
4661 * Note that if the parent is a clone, it could get
4662 * uncloned at any point, but that doesn't matter
4663 * because the list of counters and the generation
4664 * count can't have changed since we took the mutex.
4666 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
4668 child_ctx
->parent_ctx
= cloned_ctx
;
4669 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
4671 child_ctx
->parent_ctx
= parent_ctx
;
4672 child_ctx
->parent_gen
= parent_ctx
->generation
;
4674 get_ctx(child_ctx
->parent_ctx
);
4677 mutex_unlock(&parent_ctx
->mutex
);
4679 perf_unpin_context(parent_ctx
);
4684 static void __cpuinit
perf_counter_init_cpu(int cpu
)
4686 struct perf_cpu_context
*cpuctx
;
4688 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4689 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
4691 spin_lock(&perf_resource_lock
);
4692 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
4693 spin_unlock(&perf_resource_lock
);
4695 hw_perf_counter_setup(cpu
);
4698 #ifdef CONFIG_HOTPLUG_CPU
4699 static void __perf_counter_exit_cpu(void *info
)
4701 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4702 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4703 struct perf_counter
*counter
, *tmp
;
4705 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
4706 __perf_counter_remove_from_context(counter
);
4708 static void perf_counter_exit_cpu(int cpu
)
4710 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4711 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4713 mutex_lock(&ctx
->mutex
);
4714 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
4715 mutex_unlock(&ctx
->mutex
);
4718 static inline void perf_counter_exit_cpu(int cpu
) { }
4721 static int __cpuinit
4722 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
4724 unsigned int cpu
= (long)hcpu
;
4728 case CPU_UP_PREPARE
:
4729 case CPU_UP_PREPARE_FROZEN
:
4730 perf_counter_init_cpu(cpu
);
4734 case CPU_ONLINE_FROZEN
:
4735 hw_perf_counter_setup_online(cpu
);
4738 case CPU_DOWN_PREPARE
:
4739 case CPU_DOWN_PREPARE_FROZEN
:
4740 perf_counter_exit_cpu(cpu
);
4751 * This has to have a higher priority than migration_notifier in sched.c.
4753 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
4754 .notifier_call
= perf_cpu_notify
,
4758 void __init
perf_counter_init(void)
4760 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
4761 (void *)(long)smp_processor_id());
4762 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
4763 (void *)(long)smp_processor_id());
4764 register_cpu_notifier(&perf_cpu_nb
);
4767 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
4769 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
4773 perf_set_reserve_percpu(struct sysdev_class
*class,
4777 struct perf_cpu_context
*cpuctx
;
4781 err
= strict_strtoul(buf
, 10, &val
);
4784 if (val
> perf_max_counters
)
4787 spin_lock(&perf_resource_lock
);
4788 perf_reserved_percpu
= val
;
4789 for_each_online_cpu(cpu
) {
4790 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4791 spin_lock_irq(&cpuctx
->ctx
.lock
);
4792 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
4793 perf_max_counters
- perf_reserved_percpu
);
4794 cpuctx
->max_pertask
= mpt
;
4795 spin_unlock_irq(&cpuctx
->ctx
.lock
);
4797 spin_unlock(&perf_resource_lock
);
4802 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
4804 return sprintf(buf
, "%d\n", perf_overcommit
);
4808 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
4813 err
= strict_strtoul(buf
, 10, &val
);
4819 spin_lock(&perf_resource_lock
);
4820 perf_overcommit
= val
;
4821 spin_unlock(&perf_resource_lock
);
4826 static SYSDEV_CLASS_ATTR(
4829 perf_show_reserve_percpu
,
4830 perf_set_reserve_percpu
4833 static SYSDEV_CLASS_ATTR(
4836 perf_show_overcommit
,
4840 static struct attribute
*perfclass_attrs
[] = {
4841 &attr_reserve_percpu
.attr
,
4842 &attr_overcommit
.attr
,
4846 static struct attribute_group perfclass_attr_group
= {
4847 .attrs
= perfclass_attrs
,
4848 .name
= "perf_counters",
4851 static int __init
perf_counter_sysfs_init(void)
4853 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
4854 &perfclass_attr_group
);
4856 device_initcall(perf_counter_sysfs_init
);