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
;
47 * perf counter paranoia level:
49 * 1 - disallow cpu counters to unpriv
50 * 2 - disallow kernel profiling to unpriv
52 int sysctl_perf_counter_paranoid __read_mostly
;
54 static inline bool perf_paranoid_cpu(void)
56 return sysctl_perf_counter_paranoid
> 0;
59 static inline bool perf_paranoid_kernel(void)
61 return sysctl_perf_counter_paranoid
> 1;
64 int sysctl_perf_counter_mlock __read_mostly
= 512; /* 'free' kb per user */
67 * max perf counter sample rate
69 int sysctl_perf_counter_sample_rate __read_mostly
= 100000;
71 static atomic64_t perf_counter_id
;
74 * Lock for (sysadmin-configurable) counter reservations:
76 static DEFINE_SPINLOCK(perf_resource_lock
);
79 * Architecture provided APIs - weak aliases:
81 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
86 void __weak
hw_perf_disable(void) { barrier(); }
87 void __weak
hw_perf_enable(void) { barrier(); }
89 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
92 hw_perf_group_sched_in(struct perf_counter
*group_leader
,
93 struct perf_cpu_context
*cpuctx
,
94 struct perf_counter_context
*ctx
, int cpu
)
99 void __weak
perf_counter_print_debug(void) { }
101 static DEFINE_PER_CPU(int, disable_count
);
103 void __perf_disable(void)
105 __get_cpu_var(disable_count
)++;
108 bool __perf_enable(void)
110 return !--__get_cpu_var(disable_count
);
113 void perf_disable(void)
119 void perf_enable(void)
125 static void get_ctx(struct perf_counter_context
*ctx
)
127 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
130 static void free_ctx(struct rcu_head
*head
)
132 struct perf_counter_context
*ctx
;
134 ctx
= container_of(head
, struct perf_counter_context
, rcu_head
);
138 static void put_ctx(struct perf_counter_context
*ctx
)
140 if (atomic_dec_and_test(&ctx
->refcount
)) {
142 put_ctx(ctx
->parent_ctx
);
144 put_task_struct(ctx
->task
);
145 call_rcu(&ctx
->rcu_head
, free_ctx
);
150 * Get the perf_counter_context for a task and lock it.
151 * This has to cope with with the fact that until it is locked,
152 * the context could get moved to another task.
154 static struct perf_counter_context
*
155 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
157 struct perf_counter_context
*ctx
;
161 ctx
= rcu_dereference(task
->perf_counter_ctxp
);
164 * If this context is a clone of another, it might
165 * get swapped for another underneath us by
166 * perf_counter_task_sched_out, though the
167 * rcu_read_lock() protects us from any context
168 * getting freed. Lock the context and check if it
169 * got swapped before we could get the lock, and retry
170 * if so. If we locked the right context, then it
171 * can't get swapped on us any more.
173 spin_lock_irqsave(&ctx
->lock
, *flags
);
174 if (ctx
!= rcu_dereference(task
->perf_counter_ctxp
)) {
175 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
179 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
180 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
189 * Get the context for a task and increment its pin_count so it
190 * can't get swapped to another task. This also increments its
191 * reference count so that the context can't get freed.
193 static struct perf_counter_context
*perf_pin_task_context(struct task_struct
*task
)
195 struct perf_counter_context
*ctx
;
198 ctx
= perf_lock_task_context(task
, &flags
);
201 spin_unlock_irqrestore(&ctx
->lock
, flags
);
206 static void perf_unpin_context(struct perf_counter_context
*ctx
)
210 spin_lock_irqsave(&ctx
->lock
, flags
);
212 spin_unlock_irqrestore(&ctx
->lock
, flags
);
217 * Add a counter from the lists for its context.
218 * Must be called with ctx->mutex and ctx->lock held.
221 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
223 struct perf_counter
*group_leader
= counter
->group_leader
;
226 * Depending on whether it is a standalone or sibling counter,
227 * add it straight to the context's counter list, or to the group
228 * leader's sibling list:
230 if (group_leader
== counter
)
231 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
233 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
234 group_leader
->nr_siblings
++;
237 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
242 * Remove a counter from the lists for its context.
243 * Must be called with ctx->mutex and ctx->lock held.
246 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
248 struct perf_counter
*sibling
, *tmp
;
250 if (list_empty(&counter
->list_entry
))
254 list_del_init(&counter
->list_entry
);
255 list_del_rcu(&counter
->event_entry
);
257 if (counter
->group_leader
!= counter
)
258 counter
->group_leader
->nr_siblings
--;
261 * If this was a group counter with sibling counters then
262 * upgrade the siblings to singleton counters by adding them
263 * to the context list directly:
265 list_for_each_entry_safe(sibling
, tmp
,
266 &counter
->sibling_list
, list_entry
) {
268 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
269 sibling
->group_leader
= sibling
;
274 counter_sched_out(struct perf_counter
*counter
,
275 struct perf_cpu_context
*cpuctx
,
276 struct perf_counter_context
*ctx
)
278 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
281 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
282 counter
->tstamp_stopped
= ctx
->time
;
283 counter
->pmu
->disable(counter
);
286 if (!is_software_counter(counter
))
287 cpuctx
->active_oncpu
--;
289 if (counter
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
290 cpuctx
->exclusive
= 0;
294 group_sched_out(struct perf_counter
*group_counter
,
295 struct perf_cpu_context
*cpuctx
,
296 struct perf_counter_context
*ctx
)
298 struct perf_counter
*counter
;
300 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
303 counter_sched_out(group_counter
, cpuctx
, ctx
);
306 * Schedule out siblings (if any):
308 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
309 counter_sched_out(counter
, cpuctx
, ctx
);
311 if (group_counter
->attr
.exclusive
)
312 cpuctx
->exclusive
= 0;
316 * Cross CPU call to remove a performance counter
318 * We disable the counter on the hardware level first. After that we
319 * remove it from the context list.
321 static void __perf_counter_remove_from_context(void *info
)
323 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
324 struct perf_counter
*counter
= info
;
325 struct perf_counter_context
*ctx
= counter
->ctx
;
328 * If this is a task context, we need to check whether it is
329 * the current task context of this cpu. If not it has been
330 * scheduled out before the smp call arrived.
332 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
335 spin_lock(&ctx
->lock
);
337 * Protect the list operation against NMI by disabling the
338 * counters on a global level.
342 counter_sched_out(counter
, cpuctx
, ctx
);
344 list_del_counter(counter
, ctx
);
348 * Allow more per task counters with respect to the
351 cpuctx
->max_pertask
=
352 min(perf_max_counters
- ctx
->nr_counters
,
353 perf_max_counters
- perf_reserved_percpu
);
357 spin_unlock(&ctx
->lock
);
362 * Remove the counter from a task's (or a CPU's) list of counters.
364 * Must be called with ctx->mutex held.
366 * CPU counters are removed with a smp call. For task counters we only
367 * call when the task is on a CPU.
369 * If counter->ctx is a cloned context, callers must make sure that
370 * every task struct that counter->ctx->task could possibly point to
371 * remains valid. This is OK when called from perf_release since
372 * that only calls us on the top-level context, which can't be a clone.
373 * When called from perf_counter_exit_task, it's OK because the
374 * context has been detached from its task.
376 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
378 struct perf_counter_context
*ctx
= counter
->ctx
;
379 struct task_struct
*task
= ctx
->task
;
383 * Per cpu counters are removed via an smp call and
384 * the removal is always sucessful.
386 smp_call_function_single(counter
->cpu
,
387 __perf_counter_remove_from_context
,
393 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
396 spin_lock_irq(&ctx
->lock
);
398 * If the context is active we need to retry the smp call.
400 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
401 spin_unlock_irq(&ctx
->lock
);
406 * The lock prevents that this context is scheduled in so we
407 * can remove the counter safely, if the call above did not
410 if (!list_empty(&counter
->list_entry
)) {
411 list_del_counter(counter
, ctx
);
413 spin_unlock_irq(&ctx
->lock
);
416 static inline u64
perf_clock(void)
418 return cpu_clock(smp_processor_id());
422 * Update the record of the current time in a context.
424 static void update_context_time(struct perf_counter_context
*ctx
)
426 u64 now
= perf_clock();
428 ctx
->time
+= now
- ctx
->timestamp
;
429 ctx
->timestamp
= now
;
433 * Update the total_time_enabled and total_time_running fields for a counter.
435 static void update_counter_times(struct perf_counter
*counter
)
437 struct perf_counter_context
*ctx
= counter
->ctx
;
440 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
443 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
445 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
446 run_end
= counter
->tstamp_stopped
;
450 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
454 * Update total_time_enabled and total_time_running for all counters in a group.
456 static void update_group_times(struct perf_counter
*leader
)
458 struct perf_counter
*counter
;
460 update_counter_times(leader
);
461 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
462 update_counter_times(counter
);
466 * Cross CPU call to disable a performance counter
468 static void __perf_counter_disable(void *info
)
470 struct perf_counter
*counter
= info
;
471 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
472 struct perf_counter_context
*ctx
= counter
->ctx
;
475 * If this is a per-task counter, need to check whether this
476 * counter's task is the current task on this cpu.
478 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
481 spin_lock(&ctx
->lock
);
484 * If the counter is on, turn it off.
485 * If it is in error state, leave it in error state.
487 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
488 update_context_time(ctx
);
489 update_counter_times(counter
);
490 if (counter
== counter
->group_leader
)
491 group_sched_out(counter
, cpuctx
, ctx
);
493 counter_sched_out(counter
, cpuctx
, ctx
);
494 counter
->state
= PERF_COUNTER_STATE_OFF
;
497 spin_unlock(&ctx
->lock
);
503 * If counter->ctx is a cloned context, callers must make sure that
504 * every task struct that counter->ctx->task could possibly point to
505 * remains valid. This condition is satisifed when called through
506 * perf_counter_for_each_child or perf_counter_for_each because they
507 * hold the top-level counter's child_mutex, so any descendant that
508 * goes to exit will block in sync_child_counter.
509 * When called from perf_pending_counter it's OK because counter->ctx
510 * is the current context on this CPU and preemption is disabled,
511 * hence we can't get into perf_counter_task_sched_out for this context.
513 static void perf_counter_disable(struct perf_counter
*counter
)
515 struct perf_counter_context
*ctx
= counter
->ctx
;
516 struct task_struct
*task
= ctx
->task
;
520 * Disable the counter on the cpu that it's on
522 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
528 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
530 spin_lock_irq(&ctx
->lock
);
532 * If the counter is still active, we need to retry the cross-call.
534 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
535 spin_unlock_irq(&ctx
->lock
);
540 * Since we have the lock this context can't be scheduled
541 * in, so we can change the state safely.
543 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
544 update_counter_times(counter
);
545 counter
->state
= PERF_COUNTER_STATE_OFF
;
548 spin_unlock_irq(&ctx
->lock
);
552 counter_sched_in(struct perf_counter
*counter
,
553 struct perf_cpu_context
*cpuctx
,
554 struct perf_counter_context
*ctx
,
557 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
560 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
561 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
563 * The new state must be visible before we turn it on in the hardware:
567 if (counter
->pmu
->enable(counter
)) {
568 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
573 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
575 if (!is_software_counter(counter
))
576 cpuctx
->active_oncpu
++;
579 if (counter
->attr
.exclusive
)
580 cpuctx
->exclusive
= 1;
586 group_sched_in(struct perf_counter
*group_counter
,
587 struct perf_cpu_context
*cpuctx
,
588 struct perf_counter_context
*ctx
,
591 struct perf_counter
*counter
, *partial_group
;
594 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
597 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
599 return ret
< 0 ? ret
: 0;
601 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
605 * Schedule in siblings as one group (if any):
607 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
608 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
609 partial_group
= counter
;
618 * Groups can be scheduled in as one unit only, so undo any
619 * partial group before returning:
621 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
622 if (counter
== partial_group
)
624 counter_sched_out(counter
, cpuctx
, ctx
);
626 counter_sched_out(group_counter
, cpuctx
, ctx
);
632 * Return 1 for a group consisting entirely of software counters,
633 * 0 if the group contains any hardware counters.
635 static int is_software_only_group(struct perf_counter
*leader
)
637 struct perf_counter
*counter
;
639 if (!is_software_counter(leader
))
642 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
643 if (!is_software_counter(counter
))
650 * Work out whether we can put this counter group on the CPU now.
652 static int group_can_go_on(struct perf_counter
*counter
,
653 struct perf_cpu_context
*cpuctx
,
657 * Groups consisting entirely of software counters can always go on.
659 if (is_software_only_group(counter
))
662 * If an exclusive group is already on, no other hardware
663 * counters can go on.
665 if (cpuctx
->exclusive
)
668 * If this group is exclusive and there are already
669 * counters on the CPU, it can't go on.
671 if (counter
->attr
.exclusive
&& cpuctx
->active_oncpu
)
674 * Otherwise, try to add it if all previous groups were able
680 static void add_counter_to_ctx(struct perf_counter
*counter
,
681 struct perf_counter_context
*ctx
)
683 list_add_counter(counter
, ctx
);
684 counter
->tstamp_enabled
= ctx
->time
;
685 counter
->tstamp_running
= ctx
->time
;
686 counter
->tstamp_stopped
= ctx
->time
;
690 * Cross CPU call to install and enable a performance counter
692 * Must be called with ctx->mutex held
694 static void __perf_install_in_context(void *info
)
696 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
697 struct perf_counter
*counter
= info
;
698 struct perf_counter_context
*ctx
= counter
->ctx
;
699 struct perf_counter
*leader
= counter
->group_leader
;
700 int cpu
= smp_processor_id();
704 * If this is a task context, we need to check whether it is
705 * the current task context of this cpu. If not it has been
706 * scheduled out before the smp call arrived.
707 * Or possibly this is the right context but it isn't
708 * on this cpu because it had no counters.
710 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
711 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
713 cpuctx
->task_ctx
= ctx
;
716 spin_lock(&ctx
->lock
);
718 update_context_time(ctx
);
721 * Protect the list operation against NMI by disabling the
722 * counters on a global level. NOP for non NMI based counters.
726 add_counter_to_ctx(counter
, ctx
);
729 * Don't put the counter on if it is disabled or if
730 * it is in a group and the group isn't on.
732 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
733 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
737 * An exclusive counter can't go on if there are already active
738 * hardware counters, and no hardware counter can go on if there
739 * is already an exclusive counter on.
741 if (!group_can_go_on(counter
, cpuctx
, 1))
744 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
748 * This counter couldn't go on. If it is in a group
749 * then we have to pull the whole group off.
750 * If the counter group is pinned then put it in error state.
752 if (leader
!= counter
)
753 group_sched_out(leader
, cpuctx
, ctx
);
754 if (leader
->attr
.pinned
) {
755 update_group_times(leader
);
756 leader
->state
= PERF_COUNTER_STATE_ERROR
;
760 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
761 cpuctx
->max_pertask
--;
766 spin_unlock(&ctx
->lock
);
770 * Attach a performance counter to a context
772 * First we add the counter to the list with the hardware enable bit
773 * in counter->hw_config cleared.
775 * If the counter is attached to a task which is on a CPU we use a smp
776 * call to enable it in the task context. The task might have been
777 * scheduled away, but we check this in the smp call again.
779 * Must be called with ctx->mutex held.
782 perf_install_in_context(struct perf_counter_context
*ctx
,
783 struct perf_counter
*counter
,
786 struct task_struct
*task
= ctx
->task
;
790 * Per cpu counters are installed via an smp call and
791 * the install is always sucessful.
793 smp_call_function_single(cpu
, __perf_install_in_context
,
799 task_oncpu_function_call(task
, __perf_install_in_context
,
802 spin_lock_irq(&ctx
->lock
);
804 * we need to retry the smp call.
806 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
807 spin_unlock_irq(&ctx
->lock
);
812 * The lock prevents that this context is scheduled in so we
813 * can add the counter safely, if it the call above did not
816 if (list_empty(&counter
->list_entry
))
817 add_counter_to_ctx(counter
, ctx
);
818 spin_unlock_irq(&ctx
->lock
);
822 * Cross CPU call to enable a performance counter
824 static void __perf_counter_enable(void *info
)
826 struct perf_counter
*counter
= info
;
827 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
828 struct perf_counter_context
*ctx
= counter
->ctx
;
829 struct perf_counter
*leader
= counter
->group_leader
;
833 * If this is a per-task counter, need to check whether this
834 * counter's task is the current task on this cpu.
836 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
837 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
839 cpuctx
->task_ctx
= ctx
;
842 spin_lock(&ctx
->lock
);
844 update_context_time(ctx
);
846 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
848 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
849 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
852 * If the counter is in a group and isn't the group leader,
853 * then don't put it on unless the group is on.
855 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
858 if (!group_can_go_on(counter
, cpuctx
, 1)) {
862 if (counter
== leader
)
863 err
= group_sched_in(counter
, cpuctx
, ctx
,
866 err
= counter_sched_in(counter
, cpuctx
, ctx
,
873 * If this counter can't go on and it's part of a
874 * group, then the whole group has to come off.
876 if (leader
!= counter
)
877 group_sched_out(leader
, cpuctx
, ctx
);
878 if (leader
->attr
.pinned
) {
879 update_group_times(leader
);
880 leader
->state
= PERF_COUNTER_STATE_ERROR
;
885 spin_unlock(&ctx
->lock
);
891 * If counter->ctx is a cloned context, callers must make sure that
892 * every task struct that counter->ctx->task could possibly point to
893 * remains valid. This condition is satisfied when called through
894 * perf_counter_for_each_child or perf_counter_for_each as described
895 * for perf_counter_disable.
897 static void perf_counter_enable(struct perf_counter
*counter
)
899 struct perf_counter_context
*ctx
= counter
->ctx
;
900 struct task_struct
*task
= ctx
->task
;
904 * Enable the counter on the cpu that it's on
906 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
911 spin_lock_irq(&ctx
->lock
);
912 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
916 * If the counter is in error state, clear that first.
917 * That way, if we see the counter in error state below, we
918 * know that it has gone back into error state, as distinct
919 * from the task having been scheduled away before the
920 * cross-call arrived.
922 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
923 counter
->state
= PERF_COUNTER_STATE_OFF
;
926 spin_unlock_irq(&ctx
->lock
);
927 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
929 spin_lock_irq(&ctx
->lock
);
932 * If the context is active and the counter is still off,
933 * we need to retry the cross-call.
935 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
939 * Since we have the lock this context can't be scheduled
940 * in, so we can change the state safely.
942 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
943 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
944 counter
->tstamp_enabled
=
945 ctx
->time
- counter
->total_time_enabled
;
948 spin_unlock_irq(&ctx
->lock
);
951 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
954 * not supported on inherited counters
956 if (counter
->attr
.inherit
)
959 atomic_add(refresh
, &counter
->event_limit
);
960 perf_counter_enable(counter
);
965 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
966 struct perf_cpu_context
*cpuctx
)
968 struct perf_counter
*counter
;
970 spin_lock(&ctx
->lock
);
972 if (likely(!ctx
->nr_counters
))
974 update_context_time(ctx
);
977 if (ctx
->nr_active
) {
978 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
979 if (counter
!= counter
->group_leader
)
980 counter_sched_out(counter
, cpuctx
, ctx
);
982 group_sched_out(counter
, cpuctx
, ctx
);
987 spin_unlock(&ctx
->lock
);
991 * Test whether two contexts are equivalent, i.e. whether they
992 * have both been cloned from the same version of the same context
993 * and they both have the same number of enabled counters.
994 * If the number of enabled counters is the same, then the set
995 * of enabled counters should be the same, because these are both
996 * inherited contexts, therefore we can't access individual counters
997 * in them directly with an fd; we can only enable/disable all
998 * counters via prctl, or enable/disable all counters in a family
999 * via ioctl, which will have the same effect on both contexts.
1001 static int context_equiv(struct perf_counter_context
*ctx1
,
1002 struct perf_counter_context
*ctx2
)
1004 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1005 && ctx1
->parent_gen
== ctx2
->parent_gen
1006 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1010 * Called from scheduler to remove the counters of the current task,
1011 * with interrupts disabled.
1013 * We stop each counter and update the counter value in counter->count.
1015 * This does not protect us against NMI, but disable()
1016 * sets the disabled bit in the control field of counter _before_
1017 * accessing the counter control register. If a NMI hits, then it will
1018 * not restart the counter.
1020 void perf_counter_task_sched_out(struct task_struct
*task
,
1021 struct task_struct
*next
, int cpu
)
1023 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1024 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1025 struct perf_counter_context
*next_ctx
;
1026 struct perf_counter_context
*parent
;
1027 struct pt_regs
*regs
;
1030 regs
= task_pt_regs(task
);
1031 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1033 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1036 update_context_time(ctx
);
1039 parent
= rcu_dereference(ctx
->parent_ctx
);
1040 next_ctx
= next
->perf_counter_ctxp
;
1041 if (parent
&& next_ctx
&&
1042 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1044 * Looks like the two contexts are clones, so we might be
1045 * able to optimize the context switch. We lock both
1046 * contexts and check that they are clones under the
1047 * lock (including re-checking that neither has been
1048 * uncloned in the meantime). It doesn't matter which
1049 * order we take the locks because no other cpu could
1050 * be trying to lock both of these tasks.
1052 spin_lock(&ctx
->lock
);
1053 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1054 if (context_equiv(ctx
, next_ctx
)) {
1056 * XXX do we need a memory barrier of sorts
1057 * wrt to rcu_dereference() of perf_counter_ctxp
1059 task
->perf_counter_ctxp
= next_ctx
;
1060 next
->perf_counter_ctxp
= ctx
;
1062 next_ctx
->task
= task
;
1065 spin_unlock(&next_ctx
->lock
);
1066 spin_unlock(&ctx
->lock
);
1071 __perf_counter_sched_out(ctx
, cpuctx
);
1072 cpuctx
->task_ctx
= NULL
;
1077 * Called with IRQs disabled
1079 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
1081 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1083 if (!cpuctx
->task_ctx
)
1086 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1089 __perf_counter_sched_out(ctx
, cpuctx
);
1090 cpuctx
->task_ctx
= NULL
;
1094 * Called with IRQs disabled
1096 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1098 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
1102 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
1103 struct perf_cpu_context
*cpuctx
, int cpu
)
1105 struct perf_counter
*counter
;
1108 spin_lock(&ctx
->lock
);
1110 if (likely(!ctx
->nr_counters
))
1113 ctx
->timestamp
= perf_clock();
1118 * First go through the list and put on any pinned groups
1119 * in order to give them the best chance of going on.
1121 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1122 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1123 !counter
->attr
.pinned
)
1125 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1128 if (counter
!= counter
->group_leader
)
1129 counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
1131 if (group_can_go_on(counter
, cpuctx
, 1))
1132 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
1136 * If this pinned group hasn't been scheduled,
1137 * put it in error state.
1139 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1140 update_group_times(counter
);
1141 counter
->state
= PERF_COUNTER_STATE_ERROR
;
1145 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1147 * Ignore counters in OFF or ERROR state, and
1148 * ignore pinned counters since we did them already.
1150 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1151 counter
->attr
.pinned
)
1155 * Listen to the 'cpu' scheduling filter constraint
1158 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1161 if (counter
!= counter
->group_leader
) {
1162 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
))
1165 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
1166 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
1173 spin_unlock(&ctx
->lock
);
1177 * Called from scheduler to add the counters of the current task
1178 * with interrupts disabled.
1180 * We restore the counter value and then enable it.
1182 * This does not protect us against NMI, but enable()
1183 * sets the enabled bit in the control field of counter _before_
1184 * accessing the counter control register. If a NMI hits, then it will
1185 * keep the counter running.
1187 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
1189 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1190 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1194 if (cpuctx
->task_ctx
== ctx
)
1196 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1197 cpuctx
->task_ctx
= ctx
;
1200 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1202 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
1204 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1207 #define MAX_INTERRUPTS (~0ULL)
1209 static void perf_log_throttle(struct perf_counter
*counter
, int enable
);
1210 static void perf_log_period(struct perf_counter
*counter
, u64 period
);
1212 static void perf_adjust_period(struct perf_counter
*counter
, u64 events
)
1214 struct hw_perf_counter
*hwc
= &counter
->hw
;
1215 u64 period
, sample_period
;
1218 events
*= hwc
->sample_period
;
1219 period
= div64_u64(events
, counter
->attr
.sample_freq
);
1221 delta
= (s64
)(period
- hwc
->sample_period
);
1222 delta
= (delta
+ 7) / 8; /* low pass filter */
1224 sample_period
= hwc
->sample_period
+ delta
;
1229 perf_log_period(counter
, sample_period
);
1231 hwc
->sample_period
= sample_period
;
1234 static void perf_ctx_adjust_freq(struct perf_counter_context
*ctx
)
1236 struct perf_counter
*counter
;
1237 struct hw_perf_counter
*hwc
;
1238 u64 interrupts
, freq
;
1240 spin_lock(&ctx
->lock
);
1241 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1242 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1247 interrupts
= hwc
->interrupts
;
1248 hwc
->interrupts
= 0;
1251 * unthrottle counters on the tick
1253 if (interrupts
== MAX_INTERRUPTS
) {
1254 perf_log_throttle(counter
, 1);
1255 counter
->pmu
->unthrottle(counter
);
1256 interrupts
= 2*sysctl_perf_counter_sample_rate
/HZ
;
1259 if (!counter
->attr
.freq
|| !counter
->attr
.sample_freq
)
1263 * if the specified freq < HZ then we need to skip ticks
1265 if (counter
->attr
.sample_freq
< HZ
) {
1266 freq
= counter
->attr
.sample_freq
;
1268 hwc
->freq_count
+= freq
;
1269 hwc
->freq_interrupts
+= interrupts
;
1271 if (hwc
->freq_count
< HZ
)
1274 interrupts
= hwc
->freq_interrupts
;
1275 hwc
->freq_interrupts
= 0;
1276 hwc
->freq_count
-= HZ
;
1280 perf_adjust_period(counter
, freq
* interrupts
);
1283 * In order to avoid being stalled by an (accidental) huge
1284 * sample period, force reset the sample period if we didn't
1285 * get any events in this freq period.
1289 counter
->pmu
->disable(counter
);
1290 atomic64_set(&hwc
->period_left
, 0);
1291 counter
->pmu
->enable(counter
);
1295 spin_unlock(&ctx
->lock
);
1299 * Round-robin a context's counters:
1301 static void rotate_ctx(struct perf_counter_context
*ctx
)
1303 struct perf_counter
*counter
;
1305 if (!ctx
->nr_counters
)
1308 spin_lock(&ctx
->lock
);
1310 * Rotate the first entry last (works just fine for group counters too):
1313 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1314 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1319 spin_unlock(&ctx
->lock
);
1322 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1324 struct perf_cpu_context
*cpuctx
;
1325 struct perf_counter_context
*ctx
;
1327 if (!atomic_read(&nr_counters
))
1330 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1331 ctx
= curr
->perf_counter_ctxp
;
1333 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1335 perf_ctx_adjust_freq(ctx
);
1337 perf_counter_cpu_sched_out(cpuctx
);
1339 __perf_counter_task_sched_out(ctx
);
1341 rotate_ctx(&cpuctx
->ctx
);
1345 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1347 perf_counter_task_sched_in(curr
, cpu
);
1351 * Cross CPU call to read the hardware counter
1353 static void __read(void *info
)
1355 struct perf_counter
*counter
= info
;
1356 struct perf_counter_context
*ctx
= counter
->ctx
;
1357 unsigned long flags
;
1359 local_irq_save(flags
);
1361 update_context_time(ctx
);
1362 counter
->pmu
->read(counter
);
1363 update_counter_times(counter
);
1364 local_irq_restore(flags
);
1367 static u64
perf_counter_read(struct perf_counter
*counter
)
1370 * If counter is enabled and currently active on a CPU, update the
1371 * value in the counter structure:
1373 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1374 smp_call_function_single(counter
->oncpu
,
1375 __read
, counter
, 1);
1376 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1377 update_counter_times(counter
);
1380 return atomic64_read(&counter
->count
);
1384 * Initialize the perf_counter context in a task_struct:
1387 __perf_counter_init_context(struct perf_counter_context
*ctx
,
1388 struct task_struct
*task
)
1390 memset(ctx
, 0, sizeof(*ctx
));
1391 spin_lock_init(&ctx
->lock
);
1392 mutex_init(&ctx
->mutex
);
1393 INIT_LIST_HEAD(&ctx
->counter_list
);
1394 INIT_LIST_HEAD(&ctx
->event_list
);
1395 atomic_set(&ctx
->refcount
, 1);
1399 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1401 struct perf_counter_context
*parent_ctx
;
1402 struct perf_counter_context
*ctx
;
1403 struct perf_cpu_context
*cpuctx
;
1404 struct task_struct
*task
;
1405 unsigned long flags
;
1409 * If cpu is not a wildcard then this is a percpu counter:
1412 /* Must be root to operate on a CPU counter: */
1413 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1414 return ERR_PTR(-EACCES
);
1416 if (cpu
< 0 || cpu
> num_possible_cpus())
1417 return ERR_PTR(-EINVAL
);
1420 * We could be clever and allow to attach a counter to an
1421 * offline CPU and activate it when the CPU comes up, but
1424 if (!cpu_isset(cpu
, cpu_online_map
))
1425 return ERR_PTR(-ENODEV
);
1427 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1438 task
= find_task_by_vpid(pid
);
1440 get_task_struct(task
);
1444 return ERR_PTR(-ESRCH
);
1447 * Can't attach counters to a dying task.
1450 if (task
->flags
& PF_EXITING
)
1453 /* Reuse ptrace permission checks for now. */
1455 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1459 ctx
= perf_lock_task_context(task
, &flags
);
1461 parent_ctx
= ctx
->parent_ctx
;
1463 put_ctx(parent_ctx
);
1464 ctx
->parent_ctx
= NULL
; /* no longer a clone */
1466 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1470 ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
1474 __perf_counter_init_context(ctx
, task
);
1476 if (cmpxchg(&task
->perf_counter_ctxp
, NULL
, ctx
)) {
1478 * We raced with some other task; use
1479 * the context they set.
1484 get_task_struct(task
);
1487 put_task_struct(task
);
1491 put_task_struct(task
);
1492 return ERR_PTR(err
);
1495 static void free_counter_rcu(struct rcu_head
*head
)
1497 struct perf_counter
*counter
;
1499 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1501 put_pid_ns(counter
->ns
);
1505 static void perf_pending_sync(struct perf_counter
*counter
);
1507 static void free_counter(struct perf_counter
*counter
)
1509 perf_pending_sync(counter
);
1511 if (!counter
->parent
) {
1512 atomic_dec(&nr_counters
);
1513 if (counter
->attr
.mmap
)
1514 atomic_dec(&nr_mmap_counters
);
1515 if (counter
->attr
.comm
)
1516 atomic_dec(&nr_comm_counters
);
1519 if (counter
->destroy
)
1520 counter
->destroy(counter
);
1522 put_ctx(counter
->ctx
);
1523 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1527 * Called when the last reference to the file is gone.
1529 static int perf_release(struct inode
*inode
, struct file
*file
)
1531 struct perf_counter
*counter
= file
->private_data
;
1532 struct perf_counter_context
*ctx
= counter
->ctx
;
1534 file
->private_data
= NULL
;
1536 WARN_ON_ONCE(ctx
->parent_ctx
);
1537 mutex_lock(&ctx
->mutex
);
1538 perf_counter_remove_from_context(counter
);
1539 mutex_unlock(&ctx
->mutex
);
1541 mutex_lock(&counter
->owner
->perf_counter_mutex
);
1542 list_del_init(&counter
->owner_entry
);
1543 mutex_unlock(&counter
->owner
->perf_counter_mutex
);
1544 put_task_struct(counter
->owner
);
1546 free_counter(counter
);
1552 * Read the performance counter - simple non blocking version for now
1555 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1561 * Return end-of-file for a read on a counter that is in
1562 * error state (i.e. because it was pinned but it couldn't be
1563 * scheduled on to the CPU at some point).
1565 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1568 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1569 mutex_lock(&counter
->child_mutex
);
1570 values
[0] = perf_counter_read(counter
);
1572 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1573 values
[n
++] = counter
->total_time_enabled
+
1574 atomic64_read(&counter
->child_total_time_enabled
);
1575 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1576 values
[n
++] = counter
->total_time_running
+
1577 atomic64_read(&counter
->child_total_time_running
);
1578 if (counter
->attr
.read_format
& PERF_FORMAT_ID
)
1579 values
[n
++] = counter
->id
;
1580 mutex_unlock(&counter
->child_mutex
);
1582 if (count
< n
* sizeof(u64
))
1584 count
= n
* sizeof(u64
);
1586 if (copy_to_user(buf
, values
, count
))
1593 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1595 struct perf_counter
*counter
= file
->private_data
;
1597 return perf_read_hw(counter
, buf
, count
);
1600 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1602 struct perf_counter
*counter
= file
->private_data
;
1603 struct perf_mmap_data
*data
;
1604 unsigned int events
= POLL_HUP
;
1607 data
= rcu_dereference(counter
->data
);
1609 events
= atomic_xchg(&data
->poll
, 0);
1612 poll_wait(file
, &counter
->waitq
, wait
);
1617 static void perf_counter_reset(struct perf_counter
*counter
)
1619 (void)perf_counter_read(counter
);
1620 atomic64_set(&counter
->count
, 0);
1621 perf_counter_update_userpage(counter
);
1625 * Holding the top-level counter's child_mutex means that any
1626 * descendant process that has inherited this counter will block
1627 * in sync_child_counter if it goes to exit, thus satisfying the
1628 * task existence requirements of perf_counter_enable/disable.
1630 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1631 void (*func
)(struct perf_counter
*))
1633 struct perf_counter
*child
;
1635 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1636 mutex_lock(&counter
->child_mutex
);
1638 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1640 mutex_unlock(&counter
->child_mutex
);
1643 static void perf_counter_for_each(struct perf_counter
*counter
,
1644 void (*func
)(struct perf_counter
*))
1646 struct perf_counter_context
*ctx
= counter
->ctx
;
1647 struct perf_counter
*sibling
;
1649 WARN_ON_ONCE(ctx
->parent_ctx
);
1650 mutex_lock(&ctx
->mutex
);
1651 counter
= counter
->group_leader
;
1653 perf_counter_for_each_child(counter
, func
);
1655 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1656 perf_counter_for_each_child(counter
, func
);
1657 mutex_unlock(&ctx
->mutex
);
1660 static int perf_counter_period(struct perf_counter
*counter
, u64 __user
*arg
)
1662 struct perf_counter_context
*ctx
= counter
->ctx
;
1667 if (!counter
->attr
.sample_period
)
1670 size
= copy_from_user(&value
, arg
, sizeof(value
));
1671 if (size
!= sizeof(value
))
1677 spin_lock_irq(&ctx
->lock
);
1678 if (counter
->attr
.freq
) {
1679 if (value
> sysctl_perf_counter_sample_rate
) {
1684 counter
->attr
.sample_freq
= value
;
1686 perf_log_period(counter
, value
);
1688 counter
->attr
.sample_period
= value
;
1689 counter
->hw
.sample_period
= value
;
1692 spin_unlock_irq(&ctx
->lock
);
1697 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1699 struct perf_counter
*counter
= file
->private_data
;
1700 void (*func
)(struct perf_counter
*);
1704 case PERF_COUNTER_IOC_ENABLE
:
1705 func
= perf_counter_enable
;
1707 case PERF_COUNTER_IOC_DISABLE
:
1708 func
= perf_counter_disable
;
1710 case PERF_COUNTER_IOC_RESET
:
1711 func
= perf_counter_reset
;
1714 case PERF_COUNTER_IOC_REFRESH
:
1715 return perf_counter_refresh(counter
, arg
);
1717 case PERF_COUNTER_IOC_PERIOD
:
1718 return perf_counter_period(counter
, (u64 __user
*)arg
);
1724 if (flags
& PERF_IOC_FLAG_GROUP
)
1725 perf_counter_for_each(counter
, func
);
1727 perf_counter_for_each_child(counter
, func
);
1732 int perf_counter_task_enable(void)
1734 struct perf_counter
*counter
;
1736 mutex_lock(¤t
->perf_counter_mutex
);
1737 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1738 perf_counter_for_each_child(counter
, perf_counter_enable
);
1739 mutex_unlock(¤t
->perf_counter_mutex
);
1744 int perf_counter_task_disable(void)
1746 struct perf_counter
*counter
;
1748 mutex_lock(¤t
->perf_counter_mutex
);
1749 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1750 perf_counter_for_each_child(counter
, perf_counter_disable
);
1751 mutex_unlock(¤t
->perf_counter_mutex
);
1757 * Callers need to ensure there can be no nesting of this function, otherwise
1758 * the seqlock logic goes bad. We can not serialize this because the arch
1759 * code calls this from NMI context.
1761 void perf_counter_update_userpage(struct perf_counter
*counter
)
1763 struct perf_counter_mmap_page
*userpg
;
1764 struct perf_mmap_data
*data
;
1767 data
= rcu_dereference(counter
->data
);
1771 userpg
= data
->user_page
;
1774 * Disable preemption so as to not let the corresponding user-space
1775 * spin too long if we get preempted.
1780 userpg
->index
= counter
->hw
.idx
;
1781 userpg
->offset
= atomic64_read(&counter
->count
);
1782 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1783 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1785 userpg
->time_enabled
= counter
->total_time_enabled
+
1786 atomic64_read(&counter
->child_total_time_enabled
);
1788 userpg
->time_running
= counter
->total_time_running
+
1789 atomic64_read(&counter
->child_total_time_running
);
1798 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1800 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1801 struct perf_mmap_data
*data
;
1802 int ret
= VM_FAULT_SIGBUS
;
1804 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
1805 if (vmf
->pgoff
== 0)
1811 data
= rcu_dereference(counter
->data
);
1815 if (vmf
->pgoff
== 0) {
1816 vmf
->page
= virt_to_page(data
->user_page
);
1818 int nr
= vmf
->pgoff
- 1;
1820 if ((unsigned)nr
> data
->nr_pages
)
1823 if (vmf
->flags
& FAULT_FLAG_WRITE
)
1826 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1829 get_page(vmf
->page
);
1830 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
1831 vmf
->page
->index
= vmf
->pgoff
;
1840 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1842 struct perf_mmap_data
*data
;
1846 WARN_ON(atomic_read(&counter
->mmap_count
));
1848 size
= sizeof(struct perf_mmap_data
);
1849 size
+= nr_pages
* sizeof(void *);
1851 data
= kzalloc(size
, GFP_KERNEL
);
1855 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1856 if (!data
->user_page
)
1857 goto fail_user_page
;
1859 for (i
= 0; i
< nr_pages
; i
++) {
1860 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1861 if (!data
->data_pages
[i
])
1862 goto fail_data_pages
;
1865 data
->nr_pages
= nr_pages
;
1866 atomic_set(&data
->lock
, -1);
1868 rcu_assign_pointer(counter
->data
, data
);
1873 for (i
--; i
>= 0; i
--)
1874 free_page((unsigned long)data
->data_pages
[i
]);
1876 free_page((unsigned long)data
->user_page
);
1885 static void perf_mmap_free_page(unsigned long addr
)
1887 struct page
*page
= virt_to_page(addr
);
1889 page
->mapping
= NULL
;
1893 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1895 struct perf_mmap_data
*data
;
1898 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
1900 perf_mmap_free_page((unsigned long)data
->user_page
);
1901 for (i
= 0; i
< data
->nr_pages
; i
++)
1902 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
1907 static void perf_mmap_data_free(struct perf_counter
*counter
)
1909 struct perf_mmap_data
*data
= counter
->data
;
1911 WARN_ON(atomic_read(&counter
->mmap_count
));
1913 rcu_assign_pointer(counter
->data
, NULL
);
1914 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1917 static void perf_mmap_open(struct vm_area_struct
*vma
)
1919 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1921 atomic_inc(&counter
->mmap_count
);
1924 static void perf_mmap_close(struct vm_area_struct
*vma
)
1926 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1928 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1929 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
, &counter
->mmap_mutex
)) {
1930 struct user_struct
*user
= current_user();
1932 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
1933 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
1934 perf_mmap_data_free(counter
);
1935 mutex_unlock(&counter
->mmap_mutex
);
1939 static struct vm_operations_struct perf_mmap_vmops
= {
1940 .open
= perf_mmap_open
,
1941 .close
= perf_mmap_close
,
1942 .fault
= perf_mmap_fault
,
1943 .page_mkwrite
= perf_mmap_fault
,
1946 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1948 struct perf_counter
*counter
= file
->private_data
;
1949 unsigned long user_locked
, user_lock_limit
;
1950 struct user_struct
*user
= current_user();
1951 unsigned long locked
, lock_limit
;
1952 unsigned long vma_size
;
1953 unsigned long nr_pages
;
1954 long user_extra
, extra
;
1957 if (!(vma
->vm_flags
& VM_SHARED
))
1960 vma_size
= vma
->vm_end
- vma
->vm_start
;
1961 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1964 * If we have data pages ensure they're a power-of-two number, so we
1965 * can do bitmasks instead of modulo.
1967 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1970 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1973 if (vma
->vm_pgoff
!= 0)
1976 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1977 mutex_lock(&counter
->mmap_mutex
);
1978 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1979 if (nr_pages
!= counter
->data
->nr_pages
)
1984 user_extra
= nr_pages
+ 1;
1985 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
1988 * Increase the limit linearly with more CPUs:
1990 user_lock_limit
*= num_online_cpus();
1992 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
1995 if (user_locked
> user_lock_limit
)
1996 extra
= user_locked
- user_lock_limit
;
1998 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1999 lock_limit
>>= PAGE_SHIFT
;
2000 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2002 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
2007 WARN_ON(counter
->data
);
2008 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
2012 atomic_set(&counter
->mmap_count
, 1);
2013 atomic_long_add(user_extra
, &user
->locked_vm
);
2014 vma
->vm_mm
->locked_vm
+= extra
;
2015 counter
->data
->nr_locked
= extra
;
2016 if (vma
->vm_flags
& VM_WRITE
)
2017 counter
->data
->writable
= 1;
2020 mutex_unlock(&counter
->mmap_mutex
);
2022 vma
->vm_flags
|= VM_RESERVED
;
2023 vma
->vm_ops
= &perf_mmap_vmops
;
2028 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2030 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2031 struct perf_counter
*counter
= filp
->private_data
;
2034 mutex_lock(&inode
->i_mutex
);
2035 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
2036 mutex_unlock(&inode
->i_mutex
);
2044 static const struct file_operations perf_fops
= {
2045 .release
= perf_release
,
2048 .unlocked_ioctl
= perf_ioctl
,
2049 .compat_ioctl
= perf_ioctl
,
2051 .fasync
= perf_fasync
,
2055 * Perf counter wakeup
2057 * If there's data, ensure we set the poll() state and publish everything
2058 * to user-space before waking everybody up.
2061 void perf_counter_wakeup(struct perf_counter
*counter
)
2063 wake_up_all(&counter
->waitq
);
2065 if (counter
->pending_kill
) {
2066 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
2067 counter
->pending_kill
= 0;
2074 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2076 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2077 * single linked list and use cmpxchg() to add entries lockless.
2080 static void perf_pending_counter(struct perf_pending_entry
*entry
)
2082 struct perf_counter
*counter
= container_of(entry
,
2083 struct perf_counter
, pending
);
2085 if (counter
->pending_disable
) {
2086 counter
->pending_disable
= 0;
2087 perf_counter_disable(counter
);
2090 if (counter
->pending_wakeup
) {
2091 counter
->pending_wakeup
= 0;
2092 perf_counter_wakeup(counter
);
2096 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2098 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2102 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2103 void (*func
)(struct perf_pending_entry
*))
2105 struct perf_pending_entry
**head
;
2107 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2112 head
= &get_cpu_var(perf_pending_head
);
2115 entry
->next
= *head
;
2116 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2118 set_perf_counter_pending();
2120 put_cpu_var(perf_pending_head
);
2123 static int __perf_pending_run(void)
2125 struct perf_pending_entry
*list
;
2128 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2129 while (list
!= PENDING_TAIL
) {
2130 void (*func
)(struct perf_pending_entry
*);
2131 struct perf_pending_entry
*entry
= list
;
2138 * Ensure we observe the unqueue before we issue the wakeup,
2139 * so that we won't be waiting forever.
2140 * -- see perf_not_pending().
2151 static inline int perf_not_pending(struct perf_counter
*counter
)
2154 * If we flush on whatever cpu we run, there is a chance we don't
2158 __perf_pending_run();
2162 * Ensure we see the proper queue state before going to sleep
2163 * so that we do not miss the wakeup. -- see perf_pending_handle()
2166 return counter
->pending
.next
== NULL
;
2169 static void perf_pending_sync(struct perf_counter
*counter
)
2171 wait_event(counter
->waitq
, perf_not_pending(counter
));
2174 void perf_counter_do_pending(void)
2176 __perf_pending_run();
2180 * Callchain support -- arch specific
2183 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2192 struct perf_output_handle
{
2193 struct perf_counter
*counter
;
2194 struct perf_mmap_data
*data
;
2196 unsigned long offset
;
2200 unsigned long flags
;
2203 static bool perf_output_space(struct perf_mmap_data
*data
,
2204 unsigned int offset
, unsigned int head
)
2209 if (!data
->writable
)
2212 mask
= (data
->nr_pages
<< PAGE_SHIFT
) - 1;
2214 * Userspace could choose to issue a mb() before updating the tail
2215 * pointer. So that all reads will be completed before the write is
2218 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2221 offset
= (offset
- tail
) & mask
;
2222 head
= (head
- tail
) & mask
;
2224 if ((int)(head
- offset
) < 0)
2230 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2232 atomic_set(&handle
->data
->poll
, POLL_IN
);
2235 handle
->counter
->pending_wakeup
= 1;
2236 perf_pending_queue(&handle
->counter
->pending
,
2237 perf_pending_counter
);
2239 perf_counter_wakeup(handle
->counter
);
2243 * Curious locking construct.
2245 * We need to ensure a later event doesn't publish a head when a former
2246 * event isn't done writing. However since we need to deal with NMIs we
2247 * cannot fully serialize things.
2249 * What we do is serialize between CPUs so we only have to deal with NMI
2250 * nesting on a single CPU.
2252 * We only publish the head (and generate a wakeup) when the outer-most
2255 static void perf_output_lock(struct perf_output_handle
*handle
)
2257 struct perf_mmap_data
*data
= handle
->data
;
2262 local_irq_save(handle
->flags
);
2263 cpu
= smp_processor_id();
2265 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
2268 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2274 static void perf_output_unlock(struct perf_output_handle
*handle
)
2276 struct perf_mmap_data
*data
= handle
->data
;
2280 data
->done_head
= data
->head
;
2282 if (!handle
->locked
)
2287 * The xchg implies a full barrier that ensures all writes are done
2288 * before we publish the new head, matched by a rmb() in userspace when
2289 * reading this position.
2291 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2292 data
->user_page
->data_head
= head
;
2295 * NMI can happen here, which means we can miss a done_head update.
2298 cpu
= atomic_xchg(&data
->lock
, -1);
2299 WARN_ON_ONCE(cpu
!= smp_processor_id());
2302 * Therefore we have to validate we did not indeed do so.
2304 if (unlikely(atomic_long_read(&data
->done_head
))) {
2306 * Since we had it locked, we can lock it again.
2308 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2314 if (atomic_xchg(&data
->wakeup
, 0))
2315 perf_output_wakeup(handle
);
2317 local_irq_restore(handle
->flags
);
2320 static void perf_output_copy(struct perf_output_handle
*handle
,
2321 const void *buf
, unsigned int len
)
2323 unsigned int pages_mask
;
2324 unsigned int offset
;
2328 offset
= handle
->offset
;
2329 pages_mask
= handle
->data
->nr_pages
- 1;
2330 pages
= handle
->data
->data_pages
;
2333 unsigned int page_offset
;
2336 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2337 page_offset
= offset
& (PAGE_SIZE
- 1);
2338 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
2340 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2347 handle
->offset
= offset
;
2350 * Check we didn't copy past our reservation window, taking the
2351 * possible unsigned int wrap into account.
2353 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2356 #define perf_output_put(handle, x) \
2357 perf_output_copy((handle), &(x), sizeof(x))
2359 static int perf_output_begin(struct perf_output_handle
*handle
,
2360 struct perf_counter
*counter
, unsigned int size
,
2361 int nmi
, int sample
)
2363 struct perf_mmap_data
*data
;
2364 unsigned int offset
, head
;
2367 struct perf_event_header header
;
2373 * For inherited counters we send all the output towards the parent.
2375 if (counter
->parent
)
2376 counter
= counter
->parent
;
2379 data
= rcu_dereference(counter
->data
);
2383 handle
->data
= data
;
2384 handle
->counter
= counter
;
2386 handle
->sample
= sample
;
2388 if (!data
->nr_pages
)
2391 have_lost
= atomic_read(&data
->lost
);
2393 size
+= sizeof(lost_event
);
2395 perf_output_lock(handle
);
2398 offset
= head
= atomic_long_read(&data
->head
);
2400 if (unlikely(!perf_output_space(data
, offset
, head
)))
2402 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2404 handle
->offset
= offset
;
2405 handle
->head
= head
;
2407 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
2408 atomic_set(&data
->wakeup
, 1);
2411 lost_event
.header
.type
= PERF_EVENT_LOST
;
2412 lost_event
.header
.misc
= 0;
2413 lost_event
.header
.size
= sizeof(lost_event
);
2414 lost_event
.id
= counter
->id
;
2415 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2417 perf_output_put(handle
, lost_event
);
2423 atomic_inc(&data
->lost
);
2424 perf_output_unlock(handle
);
2431 static void perf_output_end(struct perf_output_handle
*handle
)
2433 struct perf_counter
*counter
= handle
->counter
;
2434 struct perf_mmap_data
*data
= handle
->data
;
2436 int wakeup_events
= counter
->attr
.wakeup_events
;
2438 if (handle
->sample
&& wakeup_events
) {
2439 int events
= atomic_inc_return(&data
->events
);
2440 if (events
>= wakeup_events
) {
2441 atomic_sub(wakeup_events
, &data
->events
);
2442 atomic_set(&data
->wakeup
, 1);
2446 perf_output_unlock(handle
);
2450 static u32
perf_counter_pid(struct perf_counter
*counter
, struct task_struct
*p
)
2453 * only top level counters have the pid namespace they were created in
2455 if (counter
->parent
)
2456 counter
= counter
->parent
;
2458 return task_tgid_nr_ns(p
, counter
->ns
);
2461 static u32
perf_counter_tid(struct perf_counter
*counter
, struct task_struct
*p
)
2464 * only top level counters have the pid namespace they were created in
2466 if (counter
->parent
)
2467 counter
= counter
->parent
;
2469 return task_pid_nr_ns(p
, counter
->ns
);
2472 static void perf_counter_output(struct perf_counter
*counter
, int nmi
,
2473 struct perf_sample_data
*data
)
2476 u64 sample_type
= counter
->attr
.sample_type
;
2477 struct perf_output_handle handle
;
2478 struct perf_event_header header
;
2487 struct perf_callchain_entry
*callchain
= NULL
;
2488 int callchain_size
= 0;
2495 header
.size
= sizeof(header
);
2497 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
2498 header
.misc
|= perf_misc_flags(data
->regs
);
2500 if (sample_type
& PERF_SAMPLE_IP
) {
2501 ip
= perf_instruction_pointer(data
->regs
);
2502 header
.type
|= PERF_SAMPLE_IP
;
2503 header
.size
+= sizeof(ip
);
2506 if (sample_type
& PERF_SAMPLE_TID
) {
2507 /* namespace issues */
2508 tid_entry
.pid
= perf_counter_pid(counter
, current
);
2509 tid_entry
.tid
= perf_counter_tid(counter
, current
);
2511 header
.type
|= PERF_SAMPLE_TID
;
2512 header
.size
+= sizeof(tid_entry
);
2515 if (sample_type
& PERF_SAMPLE_TIME
) {
2517 * Maybe do better on x86 and provide cpu_clock_nmi()
2519 time
= sched_clock();
2521 header
.type
|= PERF_SAMPLE_TIME
;
2522 header
.size
+= sizeof(u64
);
2525 if (sample_type
& PERF_SAMPLE_ADDR
) {
2526 header
.type
|= PERF_SAMPLE_ADDR
;
2527 header
.size
+= sizeof(u64
);
2530 if (sample_type
& PERF_SAMPLE_ID
) {
2531 header
.type
|= PERF_SAMPLE_ID
;
2532 header
.size
+= sizeof(u64
);
2535 if (sample_type
& PERF_SAMPLE_CPU
) {
2536 header
.type
|= PERF_SAMPLE_CPU
;
2537 header
.size
+= sizeof(cpu_entry
);
2539 cpu_entry
.cpu
= raw_smp_processor_id();
2542 if (sample_type
& PERF_SAMPLE_PERIOD
) {
2543 header
.type
|= PERF_SAMPLE_PERIOD
;
2544 header
.size
+= sizeof(u64
);
2547 if (sample_type
& PERF_SAMPLE_GROUP
) {
2548 header
.type
|= PERF_SAMPLE_GROUP
;
2549 header
.size
+= sizeof(u64
) +
2550 counter
->nr_siblings
* sizeof(group_entry
);
2553 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2554 callchain
= perf_callchain(data
->regs
);
2557 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2559 header
.type
|= PERF_SAMPLE_CALLCHAIN
;
2560 header
.size
+= callchain_size
;
2564 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2568 perf_output_put(&handle
, header
);
2570 if (sample_type
& PERF_SAMPLE_IP
)
2571 perf_output_put(&handle
, ip
);
2573 if (sample_type
& PERF_SAMPLE_TID
)
2574 perf_output_put(&handle
, tid_entry
);
2576 if (sample_type
& PERF_SAMPLE_TIME
)
2577 perf_output_put(&handle
, time
);
2579 if (sample_type
& PERF_SAMPLE_ADDR
)
2580 perf_output_put(&handle
, data
->addr
);
2582 if (sample_type
& PERF_SAMPLE_ID
)
2583 perf_output_put(&handle
, counter
->id
);
2585 if (sample_type
& PERF_SAMPLE_CPU
)
2586 perf_output_put(&handle
, cpu_entry
);
2588 if (sample_type
& PERF_SAMPLE_PERIOD
)
2589 perf_output_put(&handle
, data
->period
);
2592 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2594 if (sample_type
& PERF_SAMPLE_GROUP
) {
2595 struct perf_counter
*leader
, *sub
;
2596 u64 nr
= counter
->nr_siblings
;
2598 perf_output_put(&handle
, nr
);
2600 leader
= counter
->group_leader
;
2601 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2603 sub
->pmu
->read(sub
);
2605 group_entry
.id
= sub
->id
;
2606 group_entry
.counter
= atomic64_read(&sub
->count
);
2608 perf_output_put(&handle
, group_entry
);
2613 perf_output_copy(&handle
, callchain
, callchain_size
);
2615 perf_output_end(&handle
);
2622 struct perf_fork_event
{
2623 struct task_struct
*task
;
2626 struct perf_event_header header
;
2633 static void perf_counter_fork_output(struct perf_counter
*counter
,
2634 struct perf_fork_event
*fork_event
)
2636 struct perf_output_handle handle
;
2637 int size
= fork_event
->event
.header
.size
;
2638 struct task_struct
*task
= fork_event
->task
;
2639 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2644 fork_event
->event
.pid
= perf_counter_pid(counter
, task
);
2645 fork_event
->event
.ppid
= perf_counter_pid(counter
, task
->real_parent
);
2647 perf_output_put(&handle
, fork_event
->event
);
2648 perf_output_end(&handle
);
2651 static int perf_counter_fork_match(struct perf_counter
*counter
)
2653 if (counter
->attr
.comm
|| counter
->attr
.mmap
)
2659 static void perf_counter_fork_ctx(struct perf_counter_context
*ctx
,
2660 struct perf_fork_event
*fork_event
)
2662 struct perf_counter
*counter
;
2664 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2668 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2669 if (perf_counter_fork_match(counter
))
2670 perf_counter_fork_output(counter
, fork_event
);
2675 static void perf_counter_fork_event(struct perf_fork_event
*fork_event
)
2677 struct perf_cpu_context
*cpuctx
;
2678 struct perf_counter_context
*ctx
;
2680 cpuctx
= &get_cpu_var(perf_cpu_context
);
2681 perf_counter_fork_ctx(&cpuctx
->ctx
, fork_event
);
2682 put_cpu_var(perf_cpu_context
);
2686 * doesn't really matter which of the child contexts the
2687 * events ends up in.
2689 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2691 perf_counter_fork_ctx(ctx
, fork_event
);
2695 void perf_counter_fork(struct task_struct
*task
)
2697 struct perf_fork_event fork_event
;
2699 if (!atomic_read(&nr_comm_counters
) &&
2700 !atomic_read(&nr_mmap_counters
))
2703 fork_event
= (struct perf_fork_event
){
2707 .type
= PERF_EVENT_FORK
,
2708 .size
= sizeof(fork_event
.event
),
2713 perf_counter_fork_event(&fork_event
);
2720 struct perf_comm_event
{
2721 struct task_struct
*task
;
2726 struct perf_event_header header
;
2733 static void perf_counter_comm_output(struct perf_counter
*counter
,
2734 struct perf_comm_event
*comm_event
)
2736 struct perf_output_handle handle
;
2737 int size
= comm_event
->event
.header
.size
;
2738 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2743 comm_event
->event
.pid
= perf_counter_pid(counter
, comm_event
->task
);
2744 comm_event
->event
.tid
= perf_counter_tid(counter
, comm_event
->task
);
2746 perf_output_put(&handle
, comm_event
->event
);
2747 perf_output_copy(&handle
, comm_event
->comm
,
2748 comm_event
->comm_size
);
2749 perf_output_end(&handle
);
2752 static int perf_counter_comm_match(struct perf_counter
*counter
)
2754 if (counter
->attr
.comm
)
2760 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2761 struct perf_comm_event
*comm_event
)
2763 struct perf_counter
*counter
;
2765 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2769 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2770 if (perf_counter_comm_match(counter
))
2771 perf_counter_comm_output(counter
, comm_event
);
2776 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2778 struct perf_cpu_context
*cpuctx
;
2779 struct perf_counter_context
*ctx
;
2781 char *comm
= comm_event
->task
->comm
;
2783 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2785 comm_event
->comm
= comm
;
2786 comm_event
->comm_size
= size
;
2788 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2790 cpuctx
= &get_cpu_var(perf_cpu_context
);
2791 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2792 put_cpu_var(perf_cpu_context
);
2796 * doesn't really matter which of the child contexts the
2797 * events ends up in.
2799 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2801 perf_counter_comm_ctx(ctx
, comm_event
);
2805 void perf_counter_comm(struct task_struct
*task
)
2807 struct perf_comm_event comm_event
;
2809 if (!atomic_read(&nr_comm_counters
))
2812 comm_event
= (struct perf_comm_event
){
2815 .header
= { .type
= PERF_EVENT_COMM
, },
2819 perf_counter_comm_event(&comm_event
);
2826 struct perf_mmap_event
{
2827 struct vm_area_struct
*vma
;
2829 const char *file_name
;
2833 struct perf_event_header header
;
2843 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2844 struct perf_mmap_event
*mmap_event
)
2846 struct perf_output_handle handle
;
2847 int size
= mmap_event
->event
.header
.size
;
2848 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2853 mmap_event
->event
.pid
= perf_counter_pid(counter
, current
);
2854 mmap_event
->event
.tid
= perf_counter_tid(counter
, current
);
2856 perf_output_put(&handle
, mmap_event
->event
);
2857 perf_output_copy(&handle
, mmap_event
->file_name
,
2858 mmap_event
->file_size
);
2859 perf_output_end(&handle
);
2862 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2863 struct perf_mmap_event
*mmap_event
)
2865 if (counter
->attr
.mmap
)
2871 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2872 struct perf_mmap_event
*mmap_event
)
2874 struct perf_counter
*counter
;
2876 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2880 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2881 if (perf_counter_mmap_match(counter
, mmap_event
))
2882 perf_counter_mmap_output(counter
, mmap_event
);
2887 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2889 struct perf_cpu_context
*cpuctx
;
2890 struct perf_counter_context
*ctx
;
2891 struct vm_area_struct
*vma
= mmap_event
->vma
;
2892 struct file
*file
= vma
->vm_file
;
2899 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2901 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2904 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2906 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2910 name
= arch_vma_name(mmap_event
->vma
);
2915 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
2919 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2924 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2926 mmap_event
->file_name
= name
;
2927 mmap_event
->file_size
= size
;
2929 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2931 cpuctx
= &get_cpu_var(perf_cpu_context
);
2932 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2933 put_cpu_var(perf_cpu_context
);
2937 * doesn't really matter which of the child contexts the
2938 * events ends up in.
2940 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2942 perf_counter_mmap_ctx(ctx
, mmap_event
);
2948 void __perf_counter_mmap(struct vm_area_struct
*vma
)
2950 struct perf_mmap_event mmap_event
;
2952 if (!atomic_read(&nr_mmap_counters
))
2955 mmap_event
= (struct perf_mmap_event
){
2958 .header
= { .type
= PERF_EVENT_MMAP
, },
2959 .start
= vma
->vm_start
,
2960 .len
= vma
->vm_end
- vma
->vm_start
,
2961 .pgoff
= vma
->vm_pgoff
,
2965 perf_counter_mmap_event(&mmap_event
);
2969 * Log sample_period changes so that analyzing tools can re-normalize the
2974 struct perf_event_header header
;
2980 static void perf_log_period(struct perf_counter
*counter
, u64 period
)
2982 struct perf_output_handle handle
;
2983 struct freq_event event
;
2986 if (counter
->hw
.sample_period
== period
)
2989 if (counter
->attr
.sample_type
& PERF_SAMPLE_PERIOD
)
2992 event
= (struct freq_event
) {
2994 .type
= PERF_EVENT_PERIOD
,
2996 .size
= sizeof(event
),
2998 .time
= sched_clock(),
3003 ret
= perf_output_begin(&handle
, counter
, sizeof(event
), 1, 0);
3007 perf_output_put(&handle
, event
);
3008 perf_output_end(&handle
);
3012 * IRQ throttle logging
3015 static void perf_log_throttle(struct perf_counter
*counter
, int enable
)
3017 struct perf_output_handle handle
;
3021 struct perf_event_header header
;
3024 } throttle_event
= {
3026 .type
= PERF_EVENT_THROTTLE
+ 1,
3028 .size
= sizeof(throttle_event
),
3030 .time
= sched_clock(),
3034 ret
= perf_output_begin(&handle
, counter
, sizeof(throttle_event
), 1, 0);
3038 perf_output_put(&handle
, throttle_event
);
3039 perf_output_end(&handle
);
3043 * Generic counter overflow handling, sampling.
3046 int perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
3047 struct perf_sample_data
*data
)
3049 int events
= atomic_read(&counter
->event_limit
);
3050 int throttle
= counter
->pmu
->unthrottle
!= NULL
;
3051 struct hw_perf_counter
*hwc
= &counter
->hw
;
3057 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3059 if (HZ
* hwc
->interrupts
>
3060 (u64
)sysctl_perf_counter_sample_rate
) {
3061 hwc
->interrupts
= MAX_INTERRUPTS
;
3062 perf_log_throttle(counter
, 0);
3067 * Keep re-disabling counters even though on the previous
3068 * pass we disabled it - just in case we raced with a
3069 * sched-in and the counter got enabled again:
3075 if (counter
->attr
.freq
) {
3076 u64 now
= sched_clock();
3077 s64 delta
= now
- hwc
->freq_stamp
;
3079 hwc
->freq_stamp
= now
;
3081 if (delta
> 0 && delta
< TICK_NSEC
)
3082 perf_adjust_period(counter
, NSEC_PER_SEC
/ (int)delta
);
3086 * XXX event_limit might not quite work as expected on inherited
3090 counter
->pending_kill
= POLL_IN
;
3091 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
3093 counter
->pending_kill
= POLL_HUP
;
3095 counter
->pending_disable
= 1;
3096 perf_pending_queue(&counter
->pending
,
3097 perf_pending_counter
);
3099 perf_counter_disable(counter
);
3102 perf_counter_output(counter
, nmi
, data
);
3107 * Generic software counter infrastructure
3110 static void perf_swcounter_update(struct perf_counter
*counter
)
3112 struct hw_perf_counter
*hwc
= &counter
->hw
;
3117 prev
= atomic64_read(&hwc
->prev_count
);
3118 now
= atomic64_read(&hwc
->count
);
3119 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
3124 atomic64_add(delta
, &counter
->count
);
3125 atomic64_sub(delta
, &hwc
->period_left
);
3128 static void perf_swcounter_set_period(struct perf_counter
*counter
)
3130 struct hw_perf_counter
*hwc
= &counter
->hw
;
3131 s64 left
= atomic64_read(&hwc
->period_left
);
3132 s64 period
= hwc
->sample_period
;
3134 if (unlikely(left
<= -period
)) {
3136 atomic64_set(&hwc
->period_left
, left
);
3137 hwc
->last_period
= period
;
3140 if (unlikely(left
<= 0)) {
3142 atomic64_add(period
, &hwc
->period_left
);
3143 hwc
->last_period
= period
;
3146 atomic64_set(&hwc
->prev_count
, -left
);
3147 atomic64_set(&hwc
->count
, -left
);
3150 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
3152 enum hrtimer_restart ret
= HRTIMER_RESTART
;
3153 struct perf_sample_data data
;
3154 struct perf_counter
*counter
;
3157 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
3158 counter
->pmu
->read(counter
);
3161 data
.regs
= get_irq_regs();
3163 * In case we exclude kernel IPs or are somehow not in interrupt
3164 * context, provide the next best thing, the user IP.
3166 if ((counter
->attr
.exclude_kernel
|| !data
.regs
) &&
3167 !counter
->attr
.exclude_user
)
3168 data
.regs
= task_pt_regs(current
);
3171 if (perf_counter_overflow(counter
, 0, &data
))
3172 ret
= HRTIMER_NORESTART
;
3175 period
= max_t(u64
, 10000, counter
->hw
.sample_period
);
3176 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
3181 static void perf_swcounter_overflow(struct perf_counter
*counter
,
3182 int nmi
, struct perf_sample_data
*data
)
3184 data
->period
= counter
->hw
.last_period
;
3186 perf_swcounter_update(counter
);
3187 perf_swcounter_set_period(counter
);
3188 if (perf_counter_overflow(counter
, nmi
, data
))
3189 /* soft-disable the counter */
3193 static int perf_swcounter_is_counting(struct perf_counter
*counter
)
3195 struct perf_counter_context
*ctx
;
3196 unsigned long flags
;
3199 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
3202 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
)
3206 * If the counter is inactive, it could be just because
3207 * its task is scheduled out, or because it's in a group
3208 * which could not go on the PMU. We want to count in
3209 * the first case but not the second. If the context is
3210 * currently active then an inactive software counter must
3211 * be the second case. If it's not currently active then
3212 * we need to know whether the counter was active when the
3213 * context was last active, which we can determine by
3214 * comparing counter->tstamp_stopped with ctx->time.
3216 * We are within an RCU read-side critical section,
3217 * which protects the existence of *ctx.
3220 spin_lock_irqsave(&ctx
->lock
, flags
);
3222 /* Re-check state now we have the lock */
3223 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
||
3224 counter
->ctx
->is_active
||
3225 counter
->tstamp_stopped
< ctx
->time
)
3227 spin_unlock_irqrestore(&ctx
->lock
, flags
);
3231 static int perf_swcounter_match(struct perf_counter
*counter
,
3232 enum perf_type_id type
,
3233 u32 event
, struct pt_regs
*regs
)
3235 if (!perf_swcounter_is_counting(counter
))
3238 if (counter
->attr
.type
!= type
)
3240 if (counter
->attr
.config
!= event
)
3244 if (counter
->attr
.exclude_user
&& user_mode(regs
))
3247 if (counter
->attr
.exclude_kernel
&& !user_mode(regs
))
3254 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
3255 int nmi
, struct perf_sample_data
*data
)
3257 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
3259 if (counter
->hw
.sample_period
&& !neg
&& data
->regs
)
3260 perf_swcounter_overflow(counter
, nmi
, data
);
3263 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
3264 enum perf_type_id type
,
3265 u32 event
, u64 nr
, int nmi
,
3266 struct perf_sample_data
*data
)
3268 struct perf_counter
*counter
;
3270 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3274 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3275 if (perf_swcounter_match(counter
, type
, event
, data
->regs
))
3276 perf_swcounter_add(counter
, nr
, nmi
, data
);
3281 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
3284 return &cpuctx
->recursion
[3];
3287 return &cpuctx
->recursion
[2];
3290 return &cpuctx
->recursion
[1];
3292 return &cpuctx
->recursion
[0];
3295 static void do_perf_swcounter_event(enum perf_type_id type
, u32 event
,
3297 struct perf_sample_data
*data
)
3299 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3300 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
3301 struct perf_counter_context
*ctx
;
3309 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
3313 * doesn't really matter which of the child contexts the
3314 * events ends up in.
3316 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3318 perf_swcounter_ctx_event(ctx
, type
, event
, nr
, nmi
, data
);
3325 put_cpu_var(perf_cpu_context
);
3328 void __perf_swcounter_event(u32 event
, u64 nr
, int nmi
,
3329 struct pt_regs
*regs
, u64 addr
)
3331 struct perf_sample_data data
= {
3336 do_perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, &data
);
3339 static void perf_swcounter_read(struct perf_counter
*counter
)
3341 perf_swcounter_update(counter
);
3344 static int perf_swcounter_enable(struct perf_counter
*counter
)
3346 perf_swcounter_set_period(counter
);
3350 static void perf_swcounter_disable(struct perf_counter
*counter
)
3352 perf_swcounter_update(counter
);
3355 static const struct pmu perf_ops_generic
= {
3356 .enable
= perf_swcounter_enable
,
3357 .disable
= perf_swcounter_disable
,
3358 .read
= perf_swcounter_read
,
3362 * Software counter: cpu wall time clock
3365 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
3367 int cpu
= raw_smp_processor_id();
3371 now
= cpu_clock(cpu
);
3372 prev
= atomic64_read(&counter
->hw
.prev_count
);
3373 atomic64_set(&counter
->hw
.prev_count
, now
);
3374 atomic64_add(now
- prev
, &counter
->count
);
3377 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
3379 struct hw_perf_counter
*hwc
= &counter
->hw
;
3380 int cpu
= raw_smp_processor_id();
3382 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
3383 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3384 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3385 if (hwc
->sample_period
) {
3386 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3387 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3388 ns_to_ktime(period
), 0,
3389 HRTIMER_MODE_REL
, 0);
3395 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
3397 if (counter
->hw
.sample_period
)
3398 hrtimer_cancel(&counter
->hw
.hrtimer
);
3399 cpu_clock_perf_counter_update(counter
);
3402 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
3404 cpu_clock_perf_counter_update(counter
);
3407 static const struct pmu perf_ops_cpu_clock
= {
3408 .enable
= cpu_clock_perf_counter_enable
,
3409 .disable
= cpu_clock_perf_counter_disable
,
3410 .read
= cpu_clock_perf_counter_read
,
3414 * Software counter: task time clock
3417 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
3422 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
3424 atomic64_add(delta
, &counter
->count
);
3427 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
3429 struct hw_perf_counter
*hwc
= &counter
->hw
;
3432 now
= counter
->ctx
->time
;
3434 atomic64_set(&hwc
->prev_count
, now
);
3435 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3436 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3437 if (hwc
->sample_period
) {
3438 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3439 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3440 ns_to_ktime(period
), 0,
3441 HRTIMER_MODE_REL
, 0);
3447 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
3449 if (counter
->hw
.sample_period
)
3450 hrtimer_cancel(&counter
->hw
.hrtimer
);
3451 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
3455 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
3460 update_context_time(counter
->ctx
);
3461 time
= counter
->ctx
->time
;
3463 u64 now
= perf_clock();
3464 u64 delta
= now
- counter
->ctx
->timestamp
;
3465 time
= counter
->ctx
->time
+ delta
;
3468 task_clock_perf_counter_update(counter
, time
);
3471 static const struct pmu perf_ops_task_clock
= {
3472 .enable
= task_clock_perf_counter_enable
,
3473 .disable
= task_clock_perf_counter_disable
,
3474 .read
= task_clock_perf_counter_read
,
3477 #ifdef CONFIG_EVENT_PROFILE
3478 void perf_tpcounter_event(int event_id
)
3480 struct perf_sample_data data
= {
3481 .regs
= get_irq_regs();
3486 data
.regs
= task_pt_regs(current
);
3488 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, &data
);
3490 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
3492 extern int ftrace_profile_enable(int);
3493 extern void ftrace_profile_disable(int);
3495 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
3497 ftrace_profile_disable(perf_event_id(&counter
->attr
));
3500 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3502 int event_id
= perf_event_id(&counter
->attr
);
3505 ret
= ftrace_profile_enable(event_id
);
3509 counter
->destroy
= tp_perf_counter_destroy
;
3511 return &perf_ops_generic
;
3514 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3520 atomic_t perf_swcounter_enabled
[PERF_COUNT_SW_MAX
];
3522 static void sw_perf_counter_destroy(struct perf_counter
*counter
)
3524 u64 event
= counter
->attr
.config
;
3526 WARN_ON(counter
->parent
);
3528 atomic_dec(&perf_swcounter_enabled
[event
]);
3531 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
3533 const struct pmu
*pmu
= NULL
;
3534 u64 event
= counter
->attr
.config
;
3537 * Software counters (currently) can't in general distinguish
3538 * between user, kernel and hypervisor events.
3539 * However, context switches and cpu migrations are considered
3540 * to be kernel events, and page faults are never hypervisor
3544 case PERF_COUNT_SW_CPU_CLOCK
:
3545 pmu
= &perf_ops_cpu_clock
;
3548 case PERF_COUNT_SW_TASK_CLOCK
:
3550 * If the user instantiates this as a per-cpu counter,
3551 * use the cpu_clock counter instead.
3553 if (counter
->ctx
->task
)
3554 pmu
= &perf_ops_task_clock
;
3556 pmu
= &perf_ops_cpu_clock
;
3559 case PERF_COUNT_SW_PAGE_FAULTS
:
3560 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
3561 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
3562 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
3563 case PERF_COUNT_SW_CPU_MIGRATIONS
:
3564 if (!counter
->parent
) {
3565 atomic_inc(&perf_swcounter_enabled
[event
]);
3566 counter
->destroy
= sw_perf_counter_destroy
;
3568 pmu
= &perf_ops_generic
;
3576 * Allocate and initialize a counter structure
3578 static struct perf_counter
*
3579 perf_counter_alloc(struct perf_counter_attr
*attr
,
3581 struct perf_counter_context
*ctx
,
3582 struct perf_counter
*group_leader
,
3583 struct perf_counter
*parent_counter
,
3586 const struct pmu
*pmu
;
3587 struct perf_counter
*counter
;
3588 struct hw_perf_counter
*hwc
;
3591 counter
= kzalloc(sizeof(*counter
), gfpflags
);
3593 return ERR_PTR(-ENOMEM
);
3596 * Single counters are their own group leaders, with an
3597 * empty sibling list:
3600 group_leader
= counter
;
3602 mutex_init(&counter
->child_mutex
);
3603 INIT_LIST_HEAD(&counter
->child_list
);
3605 INIT_LIST_HEAD(&counter
->list_entry
);
3606 INIT_LIST_HEAD(&counter
->event_entry
);
3607 INIT_LIST_HEAD(&counter
->sibling_list
);
3608 init_waitqueue_head(&counter
->waitq
);
3610 mutex_init(&counter
->mmap_mutex
);
3613 counter
->attr
= *attr
;
3614 counter
->group_leader
= group_leader
;
3615 counter
->pmu
= NULL
;
3617 counter
->oncpu
= -1;
3619 counter
->parent
= parent_counter
;
3621 counter
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
3622 counter
->id
= atomic64_inc_return(&perf_counter_id
);
3624 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3627 counter
->state
= PERF_COUNTER_STATE_OFF
;
3632 hwc
->sample_period
= attr
->sample_period
;
3633 if (attr
->freq
&& attr
->sample_freq
)
3634 hwc
->sample_period
= 1;
3636 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
3639 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3641 if (attr
->inherit
&& (attr
->sample_type
& PERF_SAMPLE_GROUP
))
3644 switch (attr
->type
) {
3646 case PERF_TYPE_HARDWARE
:
3647 case PERF_TYPE_HW_CACHE
:
3648 pmu
= hw_perf_counter_init(counter
);
3651 case PERF_TYPE_SOFTWARE
:
3652 pmu
= sw_perf_counter_init(counter
);
3655 case PERF_TYPE_TRACEPOINT
:
3656 pmu
= tp_perf_counter_init(counter
);
3666 else if (IS_ERR(pmu
))
3671 put_pid_ns(counter
->ns
);
3673 return ERR_PTR(err
);
3678 if (!counter
->parent
) {
3679 atomic_inc(&nr_counters
);
3680 if (counter
->attr
.mmap
)
3681 atomic_inc(&nr_mmap_counters
);
3682 if (counter
->attr
.comm
)
3683 atomic_inc(&nr_comm_counters
);
3689 static int perf_copy_attr(struct perf_counter_attr __user
*uattr
,
3690 struct perf_counter_attr
*attr
)
3695 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
3699 * zero the full structure, so that a short copy will be nice.
3701 memset(attr
, 0, sizeof(*attr
));
3703 ret
= get_user(size
, &uattr
->size
);
3707 if (size
> PAGE_SIZE
) /* silly large */
3710 if (!size
) /* abi compat */
3711 size
= PERF_ATTR_SIZE_VER0
;
3713 if (size
< PERF_ATTR_SIZE_VER0
)
3717 * If we're handed a bigger struct than we know of,
3718 * ensure all the unknown bits are 0.
3720 if (size
> sizeof(*attr
)) {
3722 unsigned long __user
*addr
;
3723 unsigned long __user
*end
;
3725 addr
= PTR_ALIGN((void __user
*)uattr
+ sizeof(*attr
),
3726 sizeof(unsigned long));
3727 end
= PTR_ALIGN((void __user
*)uattr
+ size
,
3728 sizeof(unsigned long));
3730 for (; addr
< end
; addr
+= sizeof(unsigned long)) {
3731 ret
= get_user(val
, addr
);
3739 ret
= copy_from_user(attr
, uattr
, size
);
3744 * If the type exists, the corresponding creation will verify
3747 if (attr
->type
>= PERF_TYPE_MAX
)
3750 if (attr
->__reserved_1
|| attr
->__reserved_2
|| attr
->__reserved_3
)
3753 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
3756 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
3763 put_user(sizeof(*attr
), &uattr
->size
);
3769 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3771 * @attr_uptr: event type attributes for monitoring/sampling
3774 * @group_fd: group leader counter fd
3776 SYSCALL_DEFINE5(perf_counter_open
,
3777 struct perf_counter_attr __user
*, attr_uptr
,
3778 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
3780 struct perf_counter
*counter
, *group_leader
;
3781 struct perf_counter_attr attr
;
3782 struct perf_counter_context
*ctx
;
3783 struct file
*counter_file
= NULL
;
3784 struct file
*group_file
= NULL
;
3785 int fput_needed
= 0;
3786 int fput_needed2
= 0;
3789 /* for future expandability... */
3793 ret
= perf_copy_attr(attr_uptr
, &attr
);
3797 if (!attr
.exclude_kernel
) {
3798 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
3803 if (attr
.sample_freq
> sysctl_perf_counter_sample_rate
)
3808 * Get the target context (task or percpu):
3810 ctx
= find_get_context(pid
, cpu
);
3812 return PTR_ERR(ctx
);
3815 * Look up the group leader (we will attach this counter to it):
3817 group_leader
= NULL
;
3818 if (group_fd
!= -1) {
3820 group_file
= fget_light(group_fd
, &fput_needed
);
3822 goto err_put_context
;
3823 if (group_file
->f_op
!= &perf_fops
)
3824 goto err_put_context
;
3826 group_leader
= group_file
->private_data
;
3828 * Do not allow a recursive hierarchy (this new sibling
3829 * becoming part of another group-sibling):
3831 if (group_leader
->group_leader
!= group_leader
)
3832 goto err_put_context
;
3834 * Do not allow to attach to a group in a different
3835 * task or CPU context:
3837 if (group_leader
->ctx
!= ctx
)
3838 goto err_put_context
;
3840 * Only a group leader can be exclusive or pinned
3842 if (attr
.exclusive
|| attr
.pinned
)
3843 goto err_put_context
;
3846 counter
= perf_counter_alloc(&attr
, cpu
, ctx
, group_leader
,
3848 ret
= PTR_ERR(counter
);
3849 if (IS_ERR(counter
))
3850 goto err_put_context
;
3852 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
3854 goto err_free_put_context
;
3856 counter_file
= fget_light(ret
, &fput_needed2
);
3858 goto err_free_put_context
;
3860 counter
->filp
= counter_file
;
3861 WARN_ON_ONCE(ctx
->parent_ctx
);
3862 mutex_lock(&ctx
->mutex
);
3863 perf_install_in_context(ctx
, counter
, cpu
);
3865 mutex_unlock(&ctx
->mutex
);
3867 counter
->owner
= current
;
3868 get_task_struct(current
);
3869 mutex_lock(¤t
->perf_counter_mutex
);
3870 list_add_tail(&counter
->owner_entry
, ¤t
->perf_counter_list
);
3871 mutex_unlock(¤t
->perf_counter_mutex
);
3873 fput_light(counter_file
, fput_needed2
);
3876 fput_light(group_file
, fput_needed
);
3880 err_free_put_context
:
3890 * inherit a counter from parent task to child task:
3892 static struct perf_counter
*
3893 inherit_counter(struct perf_counter
*parent_counter
,
3894 struct task_struct
*parent
,
3895 struct perf_counter_context
*parent_ctx
,
3896 struct task_struct
*child
,
3897 struct perf_counter
*group_leader
,
3898 struct perf_counter_context
*child_ctx
)
3900 struct perf_counter
*child_counter
;
3903 * Instead of creating recursive hierarchies of counters,
3904 * we link inherited counters back to the original parent,
3905 * which has a filp for sure, which we use as the reference
3908 if (parent_counter
->parent
)
3909 parent_counter
= parent_counter
->parent
;
3911 child_counter
= perf_counter_alloc(&parent_counter
->attr
,
3912 parent_counter
->cpu
, child_ctx
,
3913 group_leader
, parent_counter
,
3915 if (IS_ERR(child_counter
))
3916 return child_counter
;
3920 * Make the child state follow the state of the parent counter,
3921 * not its attr.disabled bit. We hold the parent's mutex,
3922 * so we won't race with perf_counter_{en, dis}able_family.
3924 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3925 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3927 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3929 if (parent_counter
->attr
.freq
)
3930 child_counter
->hw
.sample_period
= parent_counter
->hw
.sample_period
;
3933 * Link it up in the child's context:
3935 add_counter_to_ctx(child_counter
, child_ctx
);
3938 * Get a reference to the parent filp - we will fput it
3939 * when the child counter exits. This is safe to do because
3940 * we are in the parent and we know that the filp still
3941 * exists and has a nonzero count:
3943 atomic_long_inc(&parent_counter
->filp
->f_count
);
3946 * Link this into the parent counter's child list
3948 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3949 mutex_lock(&parent_counter
->child_mutex
);
3950 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3951 mutex_unlock(&parent_counter
->child_mutex
);
3953 return child_counter
;
3956 static int inherit_group(struct perf_counter
*parent_counter
,
3957 struct task_struct
*parent
,
3958 struct perf_counter_context
*parent_ctx
,
3959 struct task_struct
*child
,
3960 struct perf_counter_context
*child_ctx
)
3962 struct perf_counter
*leader
;
3963 struct perf_counter
*sub
;
3964 struct perf_counter
*child_ctr
;
3966 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3967 child
, NULL
, child_ctx
);
3969 return PTR_ERR(leader
);
3970 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3971 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3972 child
, leader
, child_ctx
);
3973 if (IS_ERR(child_ctr
))
3974 return PTR_ERR(child_ctr
);
3979 static void sync_child_counter(struct perf_counter
*child_counter
,
3980 struct perf_counter
*parent_counter
)
3984 child_val
= atomic64_read(&child_counter
->count
);
3987 * Add back the child's count to the parent's count:
3989 atomic64_add(child_val
, &parent_counter
->count
);
3990 atomic64_add(child_counter
->total_time_enabled
,
3991 &parent_counter
->child_total_time_enabled
);
3992 atomic64_add(child_counter
->total_time_running
,
3993 &parent_counter
->child_total_time_running
);
3996 * Remove this counter from the parent's list
3998 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3999 mutex_lock(&parent_counter
->child_mutex
);
4000 list_del_init(&child_counter
->child_list
);
4001 mutex_unlock(&parent_counter
->child_mutex
);
4004 * Release the parent counter, if this was the last
4007 fput(parent_counter
->filp
);
4011 __perf_counter_exit_task(struct perf_counter
*child_counter
,
4012 struct perf_counter_context
*child_ctx
)
4014 struct perf_counter
*parent_counter
;
4016 update_counter_times(child_counter
);
4017 perf_counter_remove_from_context(child_counter
);
4019 parent_counter
= child_counter
->parent
;
4021 * It can happen that parent exits first, and has counters
4022 * that are still around due to the child reference. These
4023 * counters need to be zapped - but otherwise linger.
4025 if (parent_counter
) {
4026 sync_child_counter(child_counter
, parent_counter
);
4027 free_counter(child_counter
);
4032 * When a child task exits, feed back counter values to parent counters.
4034 void perf_counter_exit_task(struct task_struct
*child
)
4036 struct perf_counter
*child_counter
, *tmp
;
4037 struct perf_counter_context
*child_ctx
;
4038 unsigned long flags
;
4040 if (likely(!child
->perf_counter_ctxp
))
4043 local_irq_save(flags
);
4045 * We can't reschedule here because interrupts are disabled,
4046 * and either child is current or it is a task that can't be
4047 * scheduled, so we are now safe from rescheduling changing
4050 child_ctx
= child
->perf_counter_ctxp
;
4051 __perf_counter_task_sched_out(child_ctx
);
4054 * Take the context lock here so that if find_get_context is
4055 * reading child->perf_counter_ctxp, we wait until it has
4056 * incremented the context's refcount before we do put_ctx below.
4058 spin_lock(&child_ctx
->lock
);
4059 child
->perf_counter_ctxp
= NULL
;
4060 if (child_ctx
->parent_ctx
) {
4062 * This context is a clone; unclone it so it can't get
4063 * swapped to another process while we're removing all
4064 * the counters from it.
4066 put_ctx(child_ctx
->parent_ctx
);
4067 child_ctx
->parent_ctx
= NULL
;
4069 spin_unlock(&child_ctx
->lock
);
4070 local_irq_restore(flags
);
4073 * We can recurse on the same lock type through:
4075 * __perf_counter_exit_task()
4076 * sync_child_counter()
4077 * fput(parent_counter->filp)
4079 * mutex_lock(&ctx->mutex)
4081 * But since its the parent context it won't be the same instance.
4083 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
4086 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
4088 __perf_counter_exit_task(child_counter
, child_ctx
);
4091 * If the last counter was a group counter, it will have appended all
4092 * its siblings to the list, but we obtained 'tmp' before that which
4093 * will still point to the list head terminating the iteration.
4095 if (!list_empty(&child_ctx
->counter_list
))
4098 mutex_unlock(&child_ctx
->mutex
);
4104 * free an unexposed, unused context as created by inheritance by
4105 * init_task below, used by fork() in case of fail.
4107 void perf_counter_free_task(struct task_struct
*task
)
4109 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
4110 struct perf_counter
*counter
, *tmp
;
4115 mutex_lock(&ctx
->mutex
);
4117 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
) {
4118 struct perf_counter
*parent
= counter
->parent
;
4120 if (WARN_ON_ONCE(!parent
))
4123 mutex_lock(&parent
->child_mutex
);
4124 list_del_init(&counter
->child_list
);
4125 mutex_unlock(&parent
->child_mutex
);
4129 list_del_counter(counter
, ctx
);
4130 free_counter(counter
);
4133 if (!list_empty(&ctx
->counter_list
))
4136 mutex_unlock(&ctx
->mutex
);
4142 * Initialize the perf_counter context in task_struct
4144 int perf_counter_init_task(struct task_struct
*child
)
4146 struct perf_counter_context
*child_ctx
, *parent_ctx
;
4147 struct perf_counter_context
*cloned_ctx
;
4148 struct perf_counter
*counter
;
4149 struct task_struct
*parent
= current
;
4150 int inherited_all
= 1;
4153 child
->perf_counter_ctxp
= NULL
;
4155 mutex_init(&child
->perf_counter_mutex
);
4156 INIT_LIST_HEAD(&child
->perf_counter_list
);
4158 if (likely(!parent
->perf_counter_ctxp
))
4162 * This is executed from the parent task context, so inherit
4163 * counters that have been marked for cloning.
4164 * First allocate and initialize a context for the child.
4167 child_ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
4171 __perf_counter_init_context(child_ctx
, child
);
4172 child
->perf_counter_ctxp
= child_ctx
;
4173 get_task_struct(child
);
4176 * If the parent's context is a clone, pin it so it won't get
4179 parent_ctx
= perf_pin_task_context(parent
);
4182 * No need to check if parent_ctx != NULL here; since we saw
4183 * it non-NULL earlier, the only reason for it to become NULL
4184 * is if we exit, and since we're currently in the middle of
4185 * a fork we can't be exiting at the same time.
4189 * Lock the parent list. No need to lock the child - not PID
4190 * hashed yet and not running, so nobody can access it.
4192 mutex_lock(&parent_ctx
->mutex
);
4195 * We dont have to disable NMIs - we are only looking at
4196 * the list, not manipulating it:
4198 list_for_each_entry_rcu(counter
, &parent_ctx
->event_list
, event_entry
) {
4199 if (counter
!= counter
->group_leader
)
4202 if (!counter
->attr
.inherit
) {
4207 ret
= inherit_group(counter
, parent
, parent_ctx
,
4215 if (inherited_all
) {
4217 * Mark the child context as a clone of the parent
4218 * context, or of whatever the parent is a clone of.
4219 * Note that if the parent is a clone, it could get
4220 * uncloned at any point, but that doesn't matter
4221 * because the list of counters and the generation
4222 * count can't have changed since we took the mutex.
4224 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
4226 child_ctx
->parent_ctx
= cloned_ctx
;
4227 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
4229 child_ctx
->parent_ctx
= parent_ctx
;
4230 child_ctx
->parent_gen
= parent_ctx
->generation
;
4232 get_ctx(child_ctx
->parent_ctx
);
4235 mutex_unlock(&parent_ctx
->mutex
);
4237 perf_unpin_context(parent_ctx
);
4242 static void __cpuinit
perf_counter_init_cpu(int cpu
)
4244 struct perf_cpu_context
*cpuctx
;
4246 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4247 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
4249 spin_lock(&perf_resource_lock
);
4250 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
4251 spin_unlock(&perf_resource_lock
);
4253 hw_perf_counter_setup(cpu
);
4256 #ifdef CONFIG_HOTPLUG_CPU
4257 static void __perf_counter_exit_cpu(void *info
)
4259 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4260 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4261 struct perf_counter
*counter
, *tmp
;
4263 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
4264 __perf_counter_remove_from_context(counter
);
4266 static void perf_counter_exit_cpu(int cpu
)
4268 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4269 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4271 mutex_lock(&ctx
->mutex
);
4272 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
4273 mutex_unlock(&ctx
->mutex
);
4276 static inline void perf_counter_exit_cpu(int cpu
) { }
4279 static int __cpuinit
4280 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
4282 unsigned int cpu
= (long)hcpu
;
4286 case CPU_UP_PREPARE
:
4287 case CPU_UP_PREPARE_FROZEN
:
4288 perf_counter_init_cpu(cpu
);
4291 case CPU_DOWN_PREPARE
:
4292 case CPU_DOWN_PREPARE_FROZEN
:
4293 perf_counter_exit_cpu(cpu
);
4304 * This has to have a higher priority than migration_notifier in sched.c.
4306 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
4307 .notifier_call
= perf_cpu_notify
,
4311 void __init
perf_counter_init(void)
4313 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
4314 (void *)(long)smp_processor_id());
4315 register_cpu_notifier(&perf_cpu_nb
);
4318 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
4320 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
4324 perf_set_reserve_percpu(struct sysdev_class
*class,
4328 struct perf_cpu_context
*cpuctx
;
4332 err
= strict_strtoul(buf
, 10, &val
);
4335 if (val
> perf_max_counters
)
4338 spin_lock(&perf_resource_lock
);
4339 perf_reserved_percpu
= val
;
4340 for_each_online_cpu(cpu
) {
4341 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4342 spin_lock_irq(&cpuctx
->ctx
.lock
);
4343 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
4344 perf_max_counters
- perf_reserved_percpu
);
4345 cpuctx
->max_pertask
= mpt
;
4346 spin_unlock_irq(&cpuctx
->ctx
.lock
);
4348 spin_unlock(&perf_resource_lock
);
4353 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
4355 return sprintf(buf
, "%d\n", perf_overcommit
);
4359 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
4364 err
= strict_strtoul(buf
, 10, &val
);
4370 spin_lock(&perf_resource_lock
);
4371 perf_overcommit
= val
;
4372 spin_unlock(&perf_resource_lock
);
4377 static SYSDEV_CLASS_ATTR(
4380 perf_show_reserve_percpu
,
4381 perf_set_reserve_percpu
4384 static SYSDEV_CLASS_ATTR(
4387 perf_show_overcommit
,
4391 static struct attribute
*perfclass_attrs
[] = {
4392 &attr_reserve_percpu
.attr
,
4393 &attr_overcommit
.attr
,
4397 static struct attribute_group perfclass_attr_group
= {
4398 .attrs
= perfclass_attrs
,
4399 .name
= "perf_counters",
4402 static int __init
perf_counter_sysfs_init(void)
4404 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
4405 &perfclass_attr_group
);
4407 device_initcall(perf_counter_sysfs_init
);