2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
38 int perf_max_counters __read_mostly
= 1;
39 static int perf_reserved_percpu __read_mostly
;
40 static int perf_overcommit __read_mostly
= 1;
42 static atomic_t nr_counters __read_mostly
;
43 static atomic_t nr_mmap_counters __read_mostly
;
44 static atomic_t nr_comm_counters __read_mostly
;
45 static atomic_t nr_task_counters __read_mostly
;
48 * perf counter paranoia level:
49 * -1 - not paranoid at all
50 * 0 - disallow raw tracepoint access for unpriv
51 * 1 - disallow cpu counters for unpriv
52 * 2 - disallow kernel profiling for unpriv
54 int sysctl_perf_counter_paranoid __read_mostly
= 1;
56 static inline bool perf_paranoid_tracepoint_raw(void)
58 return sysctl_perf_counter_paranoid
> -1;
61 static inline bool perf_paranoid_cpu(void)
63 return sysctl_perf_counter_paranoid
> 0;
66 static inline bool perf_paranoid_kernel(void)
68 return sysctl_perf_counter_paranoid
> 1;
71 int sysctl_perf_counter_mlock __read_mostly
= 512; /* 'free' kb per user */
74 * max perf counter sample rate
76 int sysctl_perf_counter_sample_rate __read_mostly
= 100000;
78 static atomic64_t perf_counter_id
;
81 * Lock for (sysadmin-configurable) counter reservations:
83 static DEFINE_SPINLOCK(perf_resource_lock
);
86 * Architecture provided APIs - weak aliases:
88 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
93 void __weak
hw_perf_disable(void) { barrier(); }
94 void __weak
hw_perf_enable(void) { barrier(); }
96 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
97 void __weak
hw_perf_counter_setup_online(int cpu
) { barrier(); }
100 hw_perf_group_sched_in(struct perf_counter
*group_leader
,
101 struct perf_cpu_context
*cpuctx
,
102 struct perf_counter_context
*ctx
, int cpu
)
107 void __weak
perf_counter_print_debug(void) { }
109 static DEFINE_PER_CPU(int, perf_disable_count
);
111 void __perf_disable(void)
113 __get_cpu_var(perf_disable_count
)++;
116 bool __perf_enable(void)
118 return !--__get_cpu_var(perf_disable_count
);
121 void perf_disable(void)
127 void perf_enable(void)
133 static void get_ctx(struct perf_counter_context
*ctx
)
135 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
138 static void free_ctx(struct rcu_head
*head
)
140 struct perf_counter_context
*ctx
;
142 ctx
= container_of(head
, struct perf_counter_context
, rcu_head
);
146 static void put_ctx(struct perf_counter_context
*ctx
)
148 if (atomic_dec_and_test(&ctx
->refcount
)) {
150 put_ctx(ctx
->parent_ctx
);
152 put_task_struct(ctx
->task
);
153 call_rcu(&ctx
->rcu_head
, free_ctx
);
157 static void unclone_ctx(struct perf_counter_context
*ctx
)
159 if (ctx
->parent_ctx
) {
160 put_ctx(ctx
->parent_ctx
);
161 ctx
->parent_ctx
= NULL
;
166 * If we inherit counters we want to return the parent counter id
169 static u64
primary_counter_id(struct perf_counter
*counter
)
171 u64 id
= counter
->id
;
174 id
= counter
->parent
->id
;
180 * Get the perf_counter_context for a task and lock it.
181 * This has to cope with with the fact that until it is locked,
182 * the context could get moved to another task.
184 static struct perf_counter_context
*
185 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
187 struct perf_counter_context
*ctx
;
191 ctx
= rcu_dereference(task
->perf_counter_ctxp
);
194 * If this context is a clone of another, it might
195 * get swapped for another underneath us by
196 * perf_counter_task_sched_out, though the
197 * rcu_read_lock() protects us from any context
198 * getting freed. Lock the context and check if it
199 * got swapped before we could get the lock, and retry
200 * if so. If we locked the right context, then it
201 * can't get swapped on us any more.
203 spin_lock_irqsave(&ctx
->lock
, *flags
);
204 if (ctx
!= rcu_dereference(task
->perf_counter_ctxp
)) {
205 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
209 if (!atomic_inc_not_zero(&ctx
->refcount
)) {
210 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
219 * Get the context for a task and increment its pin_count so it
220 * can't get swapped to another task. This also increments its
221 * reference count so that the context can't get freed.
223 static struct perf_counter_context
*perf_pin_task_context(struct task_struct
*task
)
225 struct perf_counter_context
*ctx
;
228 ctx
= perf_lock_task_context(task
, &flags
);
231 spin_unlock_irqrestore(&ctx
->lock
, flags
);
236 static void perf_unpin_context(struct perf_counter_context
*ctx
)
240 spin_lock_irqsave(&ctx
->lock
, flags
);
242 spin_unlock_irqrestore(&ctx
->lock
, flags
);
247 * Add a counter from the lists for its context.
248 * Must be called with ctx->mutex and ctx->lock held.
251 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
253 struct perf_counter
*group_leader
= counter
->group_leader
;
256 * Depending on whether it is a standalone or sibling counter,
257 * add it straight to the context's counter list, or to the group
258 * leader's sibling list:
260 if (group_leader
== counter
)
261 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
263 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
264 group_leader
->nr_siblings
++;
267 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
269 if (counter
->attr
.inherit_stat
)
274 * Remove a counter from the lists for its context.
275 * Must be called with ctx->mutex and ctx->lock held.
278 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
280 struct perf_counter
*sibling
, *tmp
;
282 if (list_empty(&counter
->list_entry
))
285 if (counter
->attr
.inherit_stat
)
288 list_del_init(&counter
->list_entry
);
289 list_del_rcu(&counter
->event_entry
);
291 if (counter
->group_leader
!= counter
)
292 counter
->group_leader
->nr_siblings
--;
295 * If this was a group counter with sibling counters then
296 * upgrade the siblings to singleton counters by adding them
297 * to the context list directly:
299 list_for_each_entry_safe(sibling
, tmp
,
300 &counter
->sibling_list
, list_entry
) {
302 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
303 sibling
->group_leader
= sibling
;
308 counter_sched_out(struct perf_counter
*counter
,
309 struct perf_cpu_context
*cpuctx
,
310 struct perf_counter_context
*ctx
)
312 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
315 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
316 if (counter
->pending_disable
) {
317 counter
->pending_disable
= 0;
318 counter
->state
= PERF_COUNTER_STATE_OFF
;
320 counter
->tstamp_stopped
= ctx
->time
;
321 counter
->pmu
->disable(counter
);
324 if (!is_software_counter(counter
))
325 cpuctx
->active_oncpu
--;
327 if (counter
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
328 cpuctx
->exclusive
= 0;
332 group_sched_out(struct perf_counter
*group_counter
,
333 struct perf_cpu_context
*cpuctx
,
334 struct perf_counter_context
*ctx
)
336 struct perf_counter
*counter
;
338 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
341 counter_sched_out(group_counter
, cpuctx
, ctx
);
344 * Schedule out siblings (if any):
346 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
347 counter_sched_out(counter
, cpuctx
, ctx
);
349 if (group_counter
->attr
.exclusive
)
350 cpuctx
->exclusive
= 0;
354 * Cross CPU call to remove a performance counter
356 * We disable the counter on the hardware level first. After that we
357 * remove it from the context list.
359 static void __perf_counter_remove_from_context(void *info
)
361 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
362 struct perf_counter
*counter
= info
;
363 struct perf_counter_context
*ctx
= counter
->ctx
;
366 * If this is a task context, we need to check whether it is
367 * the current task context of this cpu. If not it has been
368 * scheduled out before the smp call arrived.
370 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
373 spin_lock(&ctx
->lock
);
375 * Protect the list operation against NMI by disabling the
376 * counters on a global level.
380 counter_sched_out(counter
, cpuctx
, ctx
);
382 list_del_counter(counter
, ctx
);
386 * Allow more per task counters with respect to the
389 cpuctx
->max_pertask
=
390 min(perf_max_counters
- ctx
->nr_counters
,
391 perf_max_counters
- perf_reserved_percpu
);
395 spin_unlock(&ctx
->lock
);
400 * Remove the counter from a task's (or a CPU's) list of counters.
402 * Must be called with ctx->mutex held.
404 * CPU counters are removed with a smp call. For task counters we only
405 * call when the task is on a CPU.
407 * If counter->ctx is a cloned context, callers must make sure that
408 * every task struct that counter->ctx->task could possibly point to
409 * remains valid. This is OK when called from perf_release since
410 * that only calls us on the top-level context, which can't be a clone.
411 * When called from perf_counter_exit_task, it's OK because the
412 * context has been detached from its task.
414 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
416 struct perf_counter_context
*ctx
= counter
->ctx
;
417 struct task_struct
*task
= ctx
->task
;
421 * Per cpu counters are removed via an smp call and
422 * the removal is always sucessful.
424 smp_call_function_single(counter
->cpu
,
425 __perf_counter_remove_from_context
,
431 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
434 spin_lock_irq(&ctx
->lock
);
436 * If the context is active we need to retry the smp call.
438 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
439 spin_unlock_irq(&ctx
->lock
);
444 * The lock prevents that this context is scheduled in so we
445 * can remove the counter safely, if the call above did not
448 if (!list_empty(&counter
->list_entry
)) {
449 list_del_counter(counter
, ctx
);
451 spin_unlock_irq(&ctx
->lock
);
454 static inline u64
perf_clock(void)
456 return cpu_clock(smp_processor_id());
460 * Update the record of the current time in a context.
462 static void update_context_time(struct perf_counter_context
*ctx
)
464 u64 now
= perf_clock();
466 ctx
->time
+= now
- ctx
->timestamp
;
467 ctx
->timestamp
= now
;
471 * Update the total_time_enabled and total_time_running fields for a counter.
473 static void update_counter_times(struct perf_counter
*counter
)
475 struct perf_counter_context
*ctx
= counter
->ctx
;
478 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
||
479 counter
->group_leader
->state
< PERF_COUNTER_STATE_INACTIVE
)
482 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
484 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
485 run_end
= counter
->tstamp_stopped
;
489 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
493 * Update total_time_enabled and total_time_running for all counters in a group.
495 static void update_group_times(struct perf_counter
*leader
)
497 struct perf_counter
*counter
;
499 update_counter_times(leader
);
500 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
501 update_counter_times(counter
);
505 * Cross CPU call to disable a performance counter
507 static void __perf_counter_disable(void *info
)
509 struct perf_counter
*counter
= info
;
510 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
511 struct perf_counter_context
*ctx
= counter
->ctx
;
514 * If this is a per-task counter, need to check whether this
515 * counter's task is the current task on this cpu.
517 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
520 spin_lock(&ctx
->lock
);
523 * If the counter is on, turn it off.
524 * If it is in error state, leave it in error state.
526 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
527 update_context_time(ctx
);
528 update_group_times(counter
);
529 if (counter
== counter
->group_leader
)
530 group_sched_out(counter
, cpuctx
, ctx
);
532 counter_sched_out(counter
, cpuctx
, ctx
);
533 counter
->state
= PERF_COUNTER_STATE_OFF
;
536 spin_unlock(&ctx
->lock
);
542 * If counter->ctx is a cloned context, callers must make sure that
543 * every task struct that counter->ctx->task could possibly point to
544 * remains valid. This condition is satisifed when called through
545 * perf_counter_for_each_child or perf_counter_for_each because they
546 * hold the top-level counter's child_mutex, so any descendant that
547 * goes to exit will block in sync_child_counter.
548 * When called from perf_pending_counter it's OK because counter->ctx
549 * is the current context on this CPU and preemption is disabled,
550 * hence we can't get into perf_counter_task_sched_out for this context.
552 static void perf_counter_disable(struct perf_counter
*counter
)
554 struct perf_counter_context
*ctx
= counter
->ctx
;
555 struct task_struct
*task
= ctx
->task
;
559 * Disable the counter on the cpu that it's on
561 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
567 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
569 spin_lock_irq(&ctx
->lock
);
571 * If the counter is still active, we need to retry the cross-call.
573 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
574 spin_unlock_irq(&ctx
->lock
);
579 * Since we have the lock this context can't be scheduled
580 * in, so we can change the state safely.
582 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
583 update_group_times(counter
);
584 counter
->state
= PERF_COUNTER_STATE_OFF
;
587 spin_unlock_irq(&ctx
->lock
);
591 counter_sched_in(struct perf_counter
*counter
,
592 struct perf_cpu_context
*cpuctx
,
593 struct perf_counter_context
*ctx
,
596 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
599 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
600 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
602 * The new state must be visible before we turn it on in the hardware:
606 if (counter
->pmu
->enable(counter
)) {
607 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
612 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
614 if (!is_software_counter(counter
))
615 cpuctx
->active_oncpu
++;
618 if (counter
->attr
.exclusive
)
619 cpuctx
->exclusive
= 1;
625 group_sched_in(struct perf_counter
*group_counter
,
626 struct perf_cpu_context
*cpuctx
,
627 struct perf_counter_context
*ctx
,
630 struct perf_counter
*counter
, *partial_group
;
633 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
636 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
638 return ret
< 0 ? ret
: 0;
640 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
644 * Schedule in siblings as one group (if any):
646 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
647 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
648 partial_group
= counter
;
657 * Groups can be scheduled in as one unit only, so undo any
658 * partial group before returning:
660 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
661 if (counter
== partial_group
)
663 counter_sched_out(counter
, cpuctx
, ctx
);
665 counter_sched_out(group_counter
, cpuctx
, ctx
);
671 * Return 1 for a group consisting entirely of software counters,
672 * 0 if the group contains any hardware counters.
674 static int is_software_only_group(struct perf_counter
*leader
)
676 struct perf_counter
*counter
;
678 if (!is_software_counter(leader
))
681 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
682 if (!is_software_counter(counter
))
689 * Work out whether we can put this counter group on the CPU now.
691 static int group_can_go_on(struct perf_counter
*counter
,
692 struct perf_cpu_context
*cpuctx
,
696 * Groups consisting entirely of software counters can always go on.
698 if (is_software_only_group(counter
))
701 * If an exclusive group is already on, no other hardware
702 * counters can go on.
704 if (cpuctx
->exclusive
)
707 * If this group is exclusive and there are already
708 * counters on the CPU, it can't go on.
710 if (counter
->attr
.exclusive
&& cpuctx
->active_oncpu
)
713 * Otherwise, try to add it if all previous groups were able
719 static void add_counter_to_ctx(struct perf_counter
*counter
,
720 struct perf_counter_context
*ctx
)
722 list_add_counter(counter
, ctx
);
723 counter
->tstamp_enabled
= ctx
->time
;
724 counter
->tstamp_running
= ctx
->time
;
725 counter
->tstamp_stopped
= ctx
->time
;
729 * Cross CPU call to install and enable a performance counter
731 * Must be called with ctx->mutex held
733 static void __perf_install_in_context(void *info
)
735 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
736 struct perf_counter
*counter
= info
;
737 struct perf_counter_context
*ctx
= counter
->ctx
;
738 struct perf_counter
*leader
= counter
->group_leader
;
739 int cpu
= smp_processor_id();
743 * If this is a task context, we need to check whether it is
744 * the current task context of this cpu. If not it has been
745 * scheduled out before the smp call arrived.
746 * Or possibly this is the right context but it isn't
747 * on this cpu because it had no counters.
749 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
750 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
752 cpuctx
->task_ctx
= ctx
;
755 spin_lock(&ctx
->lock
);
757 update_context_time(ctx
);
760 * Protect the list operation against NMI by disabling the
761 * counters on a global level. NOP for non NMI based counters.
765 add_counter_to_ctx(counter
, ctx
);
768 * Don't put the counter on if it is disabled or if
769 * it is in a group and the group isn't on.
771 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
772 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
776 * An exclusive counter can't go on if there are already active
777 * hardware counters, and no hardware counter can go on if there
778 * is already an exclusive counter on.
780 if (!group_can_go_on(counter
, cpuctx
, 1))
783 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
787 * This counter couldn't go on. If it is in a group
788 * then we have to pull the whole group off.
789 * If the counter group is pinned then put it in error state.
791 if (leader
!= counter
)
792 group_sched_out(leader
, cpuctx
, ctx
);
793 if (leader
->attr
.pinned
) {
794 update_group_times(leader
);
795 leader
->state
= PERF_COUNTER_STATE_ERROR
;
799 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
800 cpuctx
->max_pertask
--;
805 spin_unlock(&ctx
->lock
);
809 * Attach a performance counter to a context
811 * First we add the counter to the list with the hardware enable bit
812 * in counter->hw_config cleared.
814 * If the counter is attached to a task which is on a CPU we use a smp
815 * call to enable it in the task context. The task might have been
816 * scheduled away, but we check this in the smp call again.
818 * Must be called with ctx->mutex held.
821 perf_install_in_context(struct perf_counter_context
*ctx
,
822 struct perf_counter
*counter
,
825 struct task_struct
*task
= ctx
->task
;
829 * Per cpu counters are installed via an smp call and
830 * the install is always sucessful.
832 smp_call_function_single(cpu
, __perf_install_in_context
,
838 task_oncpu_function_call(task
, __perf_install_in_context
,
841 spin_lock_irq(&ctx
->lock
);
843 * we need to retry the smp call.
845 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
846 spin_unlock_irq(&ctx
->lock
);
851 * The lock prevents that this context is scheduled in so we
852 * can add the counter safely, if it the call above did not
855 if (list_empty(&counter
->list_entry
))
856 add_counter_to_ctx(counter
, ctx
);
857 spin_unlock_irq(&ctx
->lock
);
861 * Put a counter into inactive state and update time fields.
862 * Enabling the leader of a group effectively enables all
863 * the group members that aren't explicitly disabled, so we
864 * have to update their ->tstamp_enabled also.
865 * Note: this works for group members as well as group leaders
866 * since the non-leader members' sibling_lists will be empty.
868 static void __perf_counter_mark_enabled(struct perf_counter
*counter
,
869 struct perf_counter_context
*ctx
)
871 struct perf_counter
*sub
;
873 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
874 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
875 list_for_each_entry(sub
, &counter
->sibling_list
, list_entry
)
876 if (sub
->state
>= PERF_COUNTER_STATE_INACTIVE
)
877 sub
->tstamp_enabled
=
878 ctx
->time
- sub
->total_time_enabled
;
882 * Cross CPU call to enable a performance counter
884 static void __perf_counter_enable(void *info
)
886 struct perf_counter
*counter
= info
;
887 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
888 struct perf_counter_context
*ctx
= counter
->ctx
;
889 struct perf_counter
*leader
= counter
->group_leader
;
893 * If this is a per-task counter, need to check whether this
894 * counter's task is the current task on this cpu.
896 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
897 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
899 cpuctx
->task_ctx
= ctx
;
902 spin_lock(&ctx
->lock
);
904 update_context_time(ctx
);
906 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
908 __perf_counter_mark_enabled(counter
, ctx
);
911 * If the counter is in a group and isn't the group leader,
912 * then don't put it on unless the group is on.
914 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
917 if (!group_can_go_on(counter
, cpuctx
, 1)) {
921 if (counter
== leader
)
922 err
= group_sched_in(counter
, cpuctx
, ctx
,
925 err
= counter_sched_in(counter
, cpuctx
, ctx
,
932 * If this counter can't go on and it's part of a
933 * group, then the whole group has to come off.
935 if (leader
!= counter
)
936 group_sched_out(leader
, cpuctx
, ctx
);
937 if (leader
->attr
.pinned
) {
938 update_group_times(leader
);
939 leader
->state
= PERF_COUNTER_STATE_ERROR
;
944 spin_unlock(&ctx
->lock
);
950 * If counter->ctx is a cloned context, callers must make sure that
951 * every task struct that counter->ctx->task could possibly point to
952 * remains valid. This condition is satisfied when called through
953 * perf_counter_for_each_child or perf_counter_for_each as described
954 * for perf_counter_disable.
956 static void perf_counter_enable(struct perf_counter
*counter
)
958 struct perf_counter_context
*ctx
= counter
->ctx
;
959 struct task_struct
*task
= ctx
->task
;
963 * Enable the counter on the cpu that it's on
965 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
970 spin_lock_irq(&ctx
->lock
);
971 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
975 * If the counter is in error state, clear that first.
976 * That way, if we see the counter in error state below, we
977 * know that it has gone back into error state, as distinct
978 * from the task having been scheduled away before the
979 * cross-call arrived.
981 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
982 counter
->state
= PERF_COUNTER_STATE_OFF
;
985 spin_unlock_irq(&ctx
->lock
);
986 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
988 spin_lock_irq(&ctx
->lock
);
991 * If the context is active and the counter is still off,
992 * we need to retry the cross-call.
994 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
998 * Since we have the lock this context can't be scheduled
999 * in, so we can change the state safely.
1001 if (counter
->state
== PERF_COUNTER_STATE_OFF
)
1002 __perf_counter_mark_enabled(counter
, ctx
);
1005 spin_unlock_irq(&ctx
->lock
);
1008 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
1011 * not supported on inherited counters
1013 if (counter
->attr
.inherit
)
1016 atomic_add(refresh
, &counter
->event_limit
);
1017 perf_counter_enable(counter
);
1022 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
1023 struct perf_cpu_context
*cpuctx
)
1025 struct perf_counter
*counter
;
1027 spin_lock(&ctx
->lock
);
1029 if (likely(!ctx
->nr_counters
))
1031 update_context_time(ctx
);
1034 if (ctx
->nr_active
) {
1035 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1036 if (counter
!= counter
->group_leader
)
1037 counter_sched_out(counter
, cpuctx
, ctx
);
1039 group_sched_out(counter
, cpuctx
, ctx
);
1044 spin_unlock(&ctx
->lock
);
1048 * Test whether two contexts are equivalent, i.e. whether they
1049 * have both been cloned from the same version of the same context
1050 * and they both have the same number of enabled counters.
1051 * If the number of enabled counters is the same, then the set
1052 * of enabled counters should be the same, because these are both
1053 * inherited contexts, therefore we can't access individual counters
1054 * in them directly with an fd; we can only enable/disable all
1055 * counters via prctl, or enable/disable all counters in a family
1056 * via ioctl, which will have the same effect on both contexts.
1058 static int context_equiv(struct perf_counter_context
*ctx1
,
1059 struct perf_counter_context
*ctx2
)
1061 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1062 && ctx1
->parent_gen
== ctx2
->parent_gen
1063 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1066 static void __perf_counter_read(void *counter
);
1068 static void __perf_counter_sync_stat(struct perf_counter
*counter
,
1069 struct perf_counter
*next_counter
)
1073 if (!counter
->attr
.inherit_stat
)
1077 * Update the counter value, we cannot use perf_counter_read()
1078 * because we're in the middle of a context switch and have IRQs
1079 * disabled, which upsets smp_call_function_single(), however
1080 * we know the counter must be on the current CPU, therefore we
1081 * don't need to use it.
1083 switch (counter
->state
) {
1084 case PERF_COUNTER_STATE_ACTIVE
:
1085 __perf_counter_read(counter
);
1088 case PERF_COUNTER_STATE_INACTIVE
:
1089 update_counter_times(counter
);
1097 * In order to keep per-task stats reliable we need to flip the counter
1098 * values when we flip the contexts.
1100 value
= atomic64_read(&next_counter
->count
);
1101 value
= atomic64_xchg(&counter
->count
, value
);
1102 atomic64_set(&next_counter
->count
, value
);
1104 swap(counter
->total_time_enabled
, next_counter
->total_time_enabled
);
1105 swap(counter
->total_time_running
, next_counter
->total_time_running
);
1108 * Since we swizzled the values, update the user visible data too.
1110 perf_counter_update_userpage(counter
);
1111 perf_counter_update_userpage(next_counter
);
1114 #define list_next_entry(pos, member) \
1115 list_entry(pos->member.next, typeof(*pos), member)
1117 static void perf_counter_sync_stat(struct perf_counter_context
*ctx
,
1118 struct perf_counter_context
*next_ctx
)
1120 struct perf_counter
*counter
, *next_counter
;
1125 counter
= list_first_entry(&ctx
->event_list
,
1126 struct perf_counter
, event_entry
);
1128 next_counter
= list_first_entry(&next_ctx
->event_list
,
1129 struct perf_counter
, event_entry
);
1131 while (&counter
->event_entry
!= &ctx
->event_list
&&
1132 &next_counter
->event_entry
!= &next_ctx
->event_list
) {
1134 __perf_counter_sync_stat(counter
, next_counter
);
1136 counter
= list_next_entry(counter
, event_entry
);
1137 next_counter
= list_next_entry(next_counter
, event_entry
);
1142 * Called from scheduler to remove the counters of the current task,
1143 * with interrupts disabled.
1145 * We stop each counter and update the counter value in counter->count.
1147 * This does not protect us against NMI, but disable()
1148 * sets the disabled bit in the control field of counter _before_
1149 * accessing the counter control register. If a NMI hits, then it will
1150 * not restart the counter.
1152 void perf_counter_task_sched_out(struct task_struct
*task
,
1153 struct task_struct
*next
, int cpu
)
1155 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1156 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1157 struct perf_counter_context
*next_ctx
;
1158 struct perf_counter_context
*parent
;
1159 struct pt_regs
*regs
;
1162 regs
= task_pt_regs(task
);
1163 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1165 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1168 update_context_time(ctx
);
1171 parent
= rcu_dereference(ctx
->parent_ctx
);
1172 next_ctx
= next
->perf_counter_ctxp
;
1173 if (parent
&& next_ctx
&&
1174 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1176 * Looks like the two contexts are clones, so we might be
1177 * able to optimize the context switch. We lock both
1178 * contexts and check that they are clones under the
1179 * lock (including re-checking that neither has been
1180 * uncloned in the meantime). It doesn't matter which
1181 * order we take the locks because no other cpu could
1182 * be trying to lock both of these tasks.
1184 spin_lock(&ctx
->lock
);
1185 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1186 if (context_equiv(ctx
, next_ctx
)) {
1188 * XXX do we need a memory barrier of sorts
1189 * wrt to rcu_dereference() of perf_counter_ctxp
1191 task
->perf_counter_ctxp
= next_ctx
;
1192 next
->perf_counter_ctxp
= ctx
;
1194 next_ctx
->task
= task
;
1197 perf_counter_sync_stat(ctx
, next_ctx
);
1199 spin_unlock(&next_ctx
->lock
);
1200 spin_unlock(&ctx
->lock
);
1205 __perf_counter_sched_out(ctx
, cpuctx
);
1206 cpuctx
->task_ctx
= NULL
;
1211 * Called with IRQs disabled
1213 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
1215 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1217 if (!cpuctx
->task_ctx
)
1220 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1223 __perf_counter_sched_out(ctx
, cpuctx
);
1224 cpuctx
->task_ctx
= NULL
;
1228 * Called with IRQs disabled
1230 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1232 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
1236 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
1237 struct perf_cpu_context
*cpuctx
, int cpu
)
1239 struct perf_counter
*counter
;
1242 spin_lock(&ctx
->lock
);
1244 if (likely(!ctx
->nr_counters
))
1247 ctx
->timestamp
= perf_clock();
1252 * First go through the list and put on any pinned groups
1253 * in order to give them the best chance of going on.
1255 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1256 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1257 !counter
->attr
.pinned
)
1259 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1262 if (counter
!= counter
->group_leader
)
1263 counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
1265 if (group_can_go_on(counter
, cpuctx
, 1))
1266 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
1270 * If this pinned group hasn't been scheduled,
1271 * put it in error state.
1273 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1274 update_group_times(counter
);
1275 counter
->state
= PERF_COUNTER_STATE_ERROR
;
1279 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1281 * Ignore counters in OFF or ERROR state, and
1282 * ignore pinned counters since we did them already.
1284 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1285 counter
->attr
.pinned
)
1289 * Listen to the 'cpu' scheduling filter constraint
1292 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1295 if (counter
!= counter
->group_leader
) {
1296 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
))
1299 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
1300 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
1307 spin_unlock(&ctx
->lock
);
1311 * Called from scheduler to add the counters of the current task
1312 * with interrupts disabled.
1314 * We restore the counter value and then enable it.
1316 * This does not protect us against NMI, but enable()
1317 * sets the enabled bit in the control field of counter _before_
1318 * accessing the counter control register. If a NMI hits, then it will
1319 * keep the counter running.
1321 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
1323 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1324 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1328 if (cpuctx
->task_ctx
== ctx
)
1330 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1331 cpuctx
->task_ctx
= ctx
;
1334 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1336 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
1338 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1341 #define MAX_INTERRUPTS (~0ULL)
1343 static void perf_log_throttle(struct perf_counter
*counter
, int enable
);
1345 static void perf_adjust_period(struct perf_counter
*counter
, u64 events
)
1347 struct hw_perf_counter
*hwc
= &counter
->hw
;
1348 u64 period
, sample_period
;
1351 events
*= hwc
->sample_period
;
1352 period
= div64_u64(events
, counter
->attr
.sample_freq
);
1354 delta
= (s64
)(period
- hwc
->sample_period
);
1355 delta
= (delta
+ 7) / 8; /* low pass filter */
1357 sample_period
= hwc
->sample_period
+ delta
;
1362 hwc
->sample_period
= sample_period
;
1365 static void perf_ctx_adjust_freq(struct perf_counter_context
*ctx
)
1367 struct perf_counter
*counter
;
1368 struct hw_perf_counter
*hwc
;
1369 u64 interrupts
, freq
;
1371 spin_lock(&ctx
->lock
);
1372 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1373 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1378 interrupts
= hwc
->interrupts
;
1379 hwc
->interrupts
= 0;
1382 * unthrottle counters on the tick
1384 if (interrupts
== MAX_INTERRUPTS
) {
1385 perf_log_throttle(counter
, 1);
1386 counter
->pmu
->unthrottle(counter
);
1387 interrupts
= 2*sysctl_perf_counter_sample_rate
/HZ
;
1390 if (!counter
->attr
.freq
|| !counter
->attr
.sample_freq
)
1394 * if the specified freq < HZ then we need to skip ticks
1396 if (counter
->attr
.sample_freq
< HZ
) {
1397 freq
= counter
->attr
.sample_freq
;
1399 hwc
->freq_count
+= freq
;
1400 hwc
->freq_interrupts
+= interrupts
;
1402 if (hwc
->freq_count
< HZ
)
1405 interrupts
= hwc
->freq_interrupts
;
1406 hwc
->freq_interrupts
= 0;
1407 hwc
->freq_count
-= HZ
;
1411 perf_adjust_period(counter
, freq
* interrupts
);
1414 * In order to avoid being stalled by an (accidental) huge
1415 * sample period, force reset the sample period if we didn't
1416 * get any events in this freq period.
1420 counter
->pmu
->disable(counter
);
1421 atomic64_set(&hwc
->period_left
, 0);
1422 counter
->pmu
->enable(counter
);
1426 spin_unlock(&ctx
->lock
);
1430 * Round-robin a context's counters:
1432 static void rotate_ctx(struct perf_counter_context
*ctx
)
1434 struct perf_counter
*counter
;
1436 if (!ctx
->nr_counters
)
1439 spin_lock(&ctx
->lock
);
1441 * Rotate the first entry last (works just fine for group counters too):
1444 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1445 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1450 spin_unlock(&ctx
->lock
);
1453 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1455 struct perf_cpu_context
*cpuctx
;
1456 struct perf_counter_context
*ctx
;
1458 if (!atomic_read(&nr_counters
))
1461 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1462 ctx
= curr
->perf_counter_ctxp
;
1464 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1466 perf_ctx_adjust_freq(ctx
);
1468 perf_counter_cpu_sched_out(cpuctx
);
1470 __perf_counter_task_sched_out(ctx
);
1472 rotate_ctx(&cpuctx
->ctx
);
1476 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1478 perf_counter_task_sched_in(curr
, cpu
);
1482 * Enable all of a task's counters that have been marked enable-on-exec.
1483 * This expects task == current.
1485 static void perf_counter_enable_on_exec(struct task_struct
*task
)
1487 struct perf_counter_context
*ctx
;
1488 struct perf_counter
*counter
;
1489 unsigned long flags
;
1492 local_irq_save(flags
);
1493 ctx
= task
->perf_counter_ctxp
;
1494 if (!ctx
|| !ctx
->nr_counters
)
1497 __perf_counter_task_sched_out(ctx
);
1499 spin_lock(&ctx
->lock
);
1501 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1502 if (!counter
->attr
.enable_on_exec
)
1504 counter
->attr
.enable_on_exec
= 0;
1505 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
1507 __perf_counter_mark_enabled(counter
, ctx
);
1512 * Unclone this context if we enabled any counter.
1517 spin_unlock(&ctx
->lock
);
1519 perf_counter_task_sched_in(task
, smp_processor_id());
1521 local_irq_restore(flags
);
1525 * Cross CPU call to read the hardware counter
1527 static void __perf_counter_read(void *info
)
1529 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1530 struct perf_counter
*counter
= info
;
1531 struct perf_counter_context
*ctx
= counter
->ctx
;
1532 unsigned long flags
;
1535 * If this is a task context, we need to check whether it is
1536 * the current task context of this cpu. If not it has been
1537 * scheduled out before the smp call arrived. In that case
1538 * counter->count would have been updated to a recent sample
1539 * when the counter was scheduled out.
1541 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
1544 local_irq_save(flags
);
1546 update_context_time(ctx
);
1547 counter
->pmu
->read(counter
);
1548 update_counter_times(counter
);
1549 local_irq_restore(flags
);
1552 static u64
perf_counter_read(struct perf_counter
*counter
)
1555 * If counter is enabled and currently active on a CPU, update the
1556 * value in the counter structure:
1558 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1559 smp_call_function_single(counter
->oncpu
,
1560 __perf_counter_read
, counter
, 1);
1561 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1562 update_counter_times(counter
);
1565 return atomic64_read(&counter
->count
);
1569 * Initialize the perf_counter context in a task_struct:
1572 __perf_counter_init_context(struct perf_counter_context
*ctx
,
1573 struct task_struct
*task
)
1575 memset(ctx
, 0, sizeof(*ctx
));
1576 spin_lock_init(&ctx
->lock
);
1577 mutex_init(&ctx
->mutex
);
1578 INIT_LIST_HEAD(&ctx
->counter_list
);
1579 INIT_LIST_HEAD(&ctx
->event_list
);
1580 atomic_set(&ctx
->refcount
, 1);
1584 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1586 struct perf_counter_context
*ctx
;
1587 struct perf_cpu_context
*cpuctx
;
1588 struct task_struct
*task
;
1589 unsigned long flags
;
1593 * If cpu is not a wildcard then this is a percpu counter:
1596 /* Must be root to operate on a CPU counter: */
1597 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1598 return ERR_PTR(-EACCES
);
1600 if (cpu
< 0 || cpu
> num_possible_cpus())
1601 return ERR_PTR(-EINVAL
);
1604 * We could be clever and allow to attach a counter to an
1605 * offline CPU and activate it when the CPU comes up, but
1608 if (!cpu_isset(cpu
, cpu_online_map
))
1609 return ERR_PTR(-ENODEV
);
1611 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1622 task
= find_task_by_vpid(pid
);
1624 get_task_struct(task
);
1628 return ERR_PTR(-ESRCH
);
1631 * Can't attach counters to a dying task.
1634 if (task
->flags
& PF_EXITING
)
1637 /* Reuse ptrace permission checks for now. */
1639 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1643 ctx
= perf_lock_task_context(task
, &flags
);
1646 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1650 ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
1654 __perf_counter_init_context(ctx
, task
);
1656 if (cmpxchg(&task
->perf_counter_ctxp
, NULL
, ctx
)) {
1658 * We raced with some other task; use
1659 * the context they set.
1664 get_task_struct(task
);
1667 put_task_struct(task
);
1671 put_task_struct(task
);
1672 return ERR_PTR(err
);
1675 static void free_counter_rcu(struct rcu_head
*head
)
1677 struct perf_counter
*counter
;
1679 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1681 put_pid_ns(counter
->ns
);
1685 static void perf_pending_sync(struct perf_counter
*counter
);
1687 static void free_counter(struct perf_counter
*counter
)
1689 perf_pending_sync(counter
);
1691 if (!counter
->parent
) {
1692 atomic_dec(&nr_counters
);
1693 if (counter
->attr
.mmap
)
1694 atomic_dec(&nr_mmap_counters
);
1695 if (counter
->attr
.comm
)
1696 atomic_dec(&nr_comm_counters
);
1697 if (counter
->attr
.task
)
1698 atomic_dec(&nr_task_counters
);
1701 if (counter
->output
) {
1702 fput(counter
->output
->filp
);
1703 counter
->output
= NULL
;
1706 if (counter
->destroy
)
1707 counter
->destroy(counter
);
1709 put_ctx(counter
->ctx
);
1710 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1714 * Called when the last reference to the file is gone.
1716 static int perf_release(struct inode
*inode
, struct file
*file
)
1718 struct perf_counter
*counter
= file
->private_data
;
1719 struct perf_counter_context
*ctx
= counter
->ctx
;
1721 file
->private_data
= NULL
;
1723 WARN_ON_ONCE(ctx
->parent_ctx
);
1724 mutex_lock(&ctx
->mutex
);
1725 perf_counter_remove_from_context(counter
);
1726 mutex_unlock(&ctx
->mutex
);
1728 mutex_lock(&counter
->owner
->perf_counter_mutex
);
1729 list_del_init(&counter
->owner_entry
);
1730 mutex_unlock(&counter
->owner
->perf_counter_mutex
);
1731 put_task_struct(counter
->owner
);
1733 free_counter(counter
);
1738 static int perf_counter_read_size(struct perf_counter
*counter
)
1740 int entry
= sizeof(u64
); /* value */
1744 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1745 size
+= sizeof(u64
);
1747 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1748 size
+= sizeof(u64
);
1750 if (counter
->attr
.read_format
& PERF_FORMAT_ID
)
1751 entry
+= sizeof(u64
);
1753 if (counter
->attr
.read_format
& PERF_FORMAT_GROUP
) {
1754 nr
+= counter
->group_leader
->nr_siblings
;
1755 size
+= sizeof(u64
);
1763 static u64
perf_counter_read_value(struct perf_counter
*counter
)
1765 struct perf_counter
*child
;
1768 total
+= perf_counter_read(counter
);
1769 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1770 total
+= perf_counter_read(child
);
1775 static int perf_counter_read_entry(struct perf_counter
*counter
,
1776 u64 read_format
, char __user
*buf
)
1778 int n
= 0, count
= 0;
1781 values
[n
++] = perf_counter_read_value(counter
);
1782 if (read_format
& PERF_FORMAT_ID
)
1783 values
[n
++] = primary_counter_id(counter
);
1785 count
= n
* sizeof(u64
);
1787 if (copy_to_user(buf
, values
, count
))
1793 static int perf_counter_read_group(struct perf_counter
*counter
,
1794 u64 read_format
, char __user
*buf
)
1796 struct perf_counter
*leader
= counter
->group_leader
, *sub
;
1797 int n
= 0, size
= 0, err
= -EFAULT
;
1800 values
[n
++] = 1 + leader
->nr_siblings
;
1801 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1802 values
[n
++] = leader
->total_time_enabled
+
1803 atomic64_read(&leader
->child_total_time_enabled
);
1805 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1806 values
[n
++] = leader
->total_time_running
+
1807 atomic64_read(&leader
->child_total_time_running
);
1810 size
= n
* sizeof(u64
);
1812 if (copy_to_user(buf
, values
, size
))
1815 err
= perf_counter_read_entry(leader
, read_format
, buf
+ size
);
1821 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
1822 err
= perf_counter_read_entry(sub
, read_format
,
1833 static int perf_counter_read_one(struct perf_counter
*counter
,
1834 u64 read_format
, char __user
*buf
)
1839 values
[n
++] = perf_counter_read_value(counter
);
1840 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1841 values
[n
++] = counter
->total_time_enabled
+
1842 atomic64_read(&counter
->child_total_time_enabled
);
1844 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1845 values
[n
++] = counter
->total_time_running
+
1846 atomic64_read(&counter
->child_total_time_running
);
1848 if (read_format
& PERF_FORMAT_ID
)
1849 values
[n
++] = primary_counter_id(counter
);
1851 if (copy_to_user(buf
, values
, n
* sizeof(u64
)))
1854 return n
* sizeof(u64
);
1858 * Read the performance counter - simple non blocking version for now
1861 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1863 u64 read_format
= counter
->attr
.read_format
;
1867 * Return end-of-file for a read on a counter that is in
1868 * error state (i.e. because it was pinned but it couldn't be
1869 * scheduled on to the CPU at some point).
1871 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1874 if (count
< perf_counter_read_size(counter
))
1877 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1878 mutex_lock(&counter
->child_mutex
);
1879 if (read_format
& PERF_FORMAT_GROUP
)
1880 ret
= perf_counter_read_group(counter
, read_format
, buf
);
1882 ret
= perf_counter_read_one(counter
, read_format
, buf
);
1883 mutex_unlock(&counter
->child_mutex
);
1889 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1891 struct perf_counter
*counter
= file
->private_data
;
1893 return perf_read_hw(counter
, buf
, count
);
1896 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1898 struct perf_counter
*counter
= file
->private_data
;
1899 struct perf_mmap_data
*data
;
1900 unsigned int events
= POLL_HUP
;
1903 data
= rcu_dereference(counter
->data
);
1905 events
= atomic_xchg(&data
->poll
, 0);
1908 poll_wait(file
, &counter
->waitq
, wait
);
1913 static void perf_counter_reset(struct perf_counter
*counter
)
1915 (void)perf_counter_read(counter
);
1916 atomic64_set(&counter
->count
, 0);
1917 perf_counter_update_userpage(counter
);
1921 * Holding the top-level counter's child_mutex means that any
1922 * descendant process that has inherited this counter will block
1923 * in sync_child_counter if it goes to exit, thus satisfying the
1924 * task existence requirements of perf_counter_enable/disable.
1926 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1927 void (*func
)(struct perf_counter
*))
1929 struct perf_counter
*child
;
1931 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1932 mutex_lock(&counter
->child_mutex
);
1934 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1936 mutex_unlock(&counter
->child_mutex
);
1939 static void perf_counter_for_each(struct perf_counter
*counter
,
1940 void (*func
)(struct perf_counter
*))
1942 struct perf_counter_context
*ctx
= counter
->ctx
;
1943 struct perf_counter
*sibling
;
1945 WARN_ON_ONCE(ctx
->parent_ctx
);
1946 mutex_lock(&ctx
->mutex
);
1947 counter
= counter
->group_leader
;
1949 perf_counter_for_each_child(counter
, func
);
1951 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1952 perf_counter_for_each_child(counter
, func
);
1953 mutex_unlock(&ctx
->mutex
);
1956 static int perf_counter_period(struct perf_counter
*counter
, u64 __user
*arg
)
1958 struct perf_counter_context
*ctx
= counter
->ctx
;
1963 if (!counter
->attr
.sample_period
)
1966 size
= copy_from_user(&value
, arg
, sizeof(value
));
1967 if (size
!= sizeof(value
))
1973 spin_lock_irq(&ctx
->lock
);
1974 if (counter
->attr
.freq
) {
1975 if (value
> sysctl_perf_counter_sample_rate
) {
1980 counter
->attr
.sample_freq
= value
;
1982 counter
->attr
.sample_period
= value
;
1983 counter
->hw
.sample_period
= value
;
1986 spin_unlock_irq(&ctx
->lock
);
1991 int perf_counter_set_output(struct perf_counter
*counter
, int output_fd
);
1993 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1995 struct perf_counter
*counter
= file
->private_data
;
1996 void (*func
)(struct perf_counter
*);
2000 case PERF_COUNTER_IOC_ENABLE
:
2001 func
= perf_counter_enable
;
2003 case PERF_COUNTER_IOC_DISABLE
:
2004 func
= perf_counter_disable
;
2006 case PERF_COUNTER_IOC_RESET
:
2007 func
= perf_counter_reset
;
2010 case PERF_COUNTER_IOC_REFRESH
:
2011 return perf_counter_refresh(counter
, arg
);
2013 case PERF_COUNTER_IOC_PERIOD
:
2014 return perf_counter_period(counter
, (u64 __user
*)arg
);
2016 case PERF_COUNTER_IOC_SET_OUTPUT
:
2017 return perf_counter_set_output(counter
, arg
);
2023 if (flags
& PERF_IOC_FLAG_GROUP
)
2024 perf_counter_for_each(counter
, func
);
2026 perf_counter_for_each_child(counter
, func
);
2031 int perf_counter_task_enable(void)
2033 struct perf_counter
*counter
;
2035 mutex_lock(¤t
->perf_counter_mutex
);
2036 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
2037 perf_counter_for_each_child(counter
, perf_counter_enable
);
2038 mutex_unlock(¤t
->perf_counter_mutex
);
2043 int perf_counter_task_disable(void)
2045 struct perf_counter
*counter
;
2047 mutex_lock(¤t
->perf_counter_mutex
);
2048 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
2049 perf_counter_for_each_child(counter
, perf_counter_disable
);
2050 mutex_unlock(¤t
->perf_counter_mutex
);
2055 #ifndef PERF_COUNTER_INDEX_OFFSET
2056 # define PERF_COUNTER_INDEX_OFFSET 0
2059 static int perf_counter_index(struct perf_counter
*counter
)
2061 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
2064 return counter
->hw
.idx
+ 1 - PERF_COUNTER_INDEX_OFFSET
;
2068 * Callers need to ensure there can be no nesting of this function, otherwise
2069 * the seqlock logic goes bad. We can not serialize this because the arch
2070 * code calls this from NMI context.
2072 void perf_counter_update_userpage(struct perf_counter
*counter
)
2074 struct perf_counter_mmap_page
*userpg
;
2075 struct perf_mmap_data
*data
;
2078 data
= rcu_dereference(counter
->data
);
2082 userpg
= data
->user_page
;
2085 * Disable preemption so as to not let the corresponding user-space
2086 * spin too long if we get preempted.
2091 userpg
->index
= perf_counter_index(counter
);
2092 userpg
->offset
= atomic64_read(&counter
->count
);
2093 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
2094 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
2096 userpg
->time_enabled
= counter
->total_time_enabled
+
2097 atomic64_read(&counter
->child_total_time_enabled
);
2099 userpg
->time_running
= counter
->total_time_running
+
2100 atomic64_read(&counter
->child_total_time_running
);
2109 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2111 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
2112 struct perf_mmap_data
*data
;
2113 int ret
= VM_FAULT_SIGBUS
;
2115 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
2116 if (vmf
->pgoff
== 0)
2122 data
= rcu_dereference(counter
->data
);
2126 if (vmf
->pgoff
== 0) {
2127 vmf
->page
= virt_to_page(data
->user_page
);
2129 int nr
= vmf
->pgoff
- 1;
2131 if ((unsigned)nr
> data
->nr_pages
)
2134 if (vmf
->flags
& FAULT_FLAG_WRITE
)
2137 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
2140 get_page(vmf
->page
);
2141 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
2142 vmf
->page
->index
= vmf
->pgoff
;
2151 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
2153 struct perf_mmap_data
*data
;
2157 WARN_ON(atomic_read(&counter
->mmap_count
));
2159 size
= sizeof(struct perf_mmap_data
);
2160 size
+= nr_pages
* sizeof(void *);
2162 data
= kzalloc(size
, GFP_KERNEL
);
2166 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
2167 if (!data
->user_page
)
2168 goto fail_user_page
;
2170 for (i
= 0; i
< nr_pages
; i
++) {
2171 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
2172 if (!data
->data_pages
[i
])
2173 goto fail_data_pages
;
2176 data
->nr_pages
= nr_pages
;
2177 atomic_set(&data
->lock
, -1);
2179 if (counter
->attr
.watermark
) {
2180 data
->watermark
= min_t(long, PAGE_SIZE
* nr_pages
,
2181 counter
->attr
.wakeup_watermark
);
2183 if (!data
->watermark
)
2184 data
->watermark
= max(PAGE_SIZE
, PAGE_SIZE
* nr_pages
/ 4);
2186 rcu_assign_pointer(counter
->data
, data
);
2191 for (i
--; i
>= 0; i
--)
2192 free_page((unsigned long)data
->data_pages
[i
]);
2194 free_page((unsigned long)data
->user_page
);
2203 static void perf_mmap_free_page(unsigned long addr
)
2205 struct page
*page
= virt_to_page((void *)addr
);
2207 page
->mapping
= NULL
;
2211 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
2213 struct perf_mmap_data
*data
;
2216 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
2218 perf_mmap_free_page((unsigned long)data
->user_page
);
2219 for (i
= 0; i
< data
->nr_pages
; i
++)
2220 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
2225 static void perf_mmap_data_free(struct perf_counter
*counter
)
2227 struct perf_mmap_data
*data
= counter
->data
;
2229 WARN_ON(atomic_read(&counter
->mmap_count
));
2231 rcu_assign_pointer(counter
->data
, NULL
);
2232 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
2235 static void perf_mmap_open(struct vm_area_struct
*vma
)
2237 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
2239 atomic_inc(&counter
->mmap_count
);
2242 static void perf_mmap_close(struct vm_area_struct
*vma
)
2244 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
2246 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
2247 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
, &counter
->mmap_mutex
)) {
2248 struct user_struct
*user
= current_user();
2250 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
2251 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
2252 perf_mmap_data_free(counter
);
2253 mutex_unlock(&counter
->mmap_mutex
);
2257 static struct vm_operations_struct perf_mmap_vmops
= {
2258 .open
= perf_mmap_open
,
2259 .close
= perf_mmap_close
,
2260 .fault
= perf_mmap_fault
,
2261 .page_mkwrite
= perf_mmap_fault
,
2264 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2266 struct perf_counter
*counter
= file
->private_data
;
2267 unsigned long user_locked
, user_lock_limit
;
2268 struct user_struct
*user
= current_user();
2269 unsigned long locked
, lock_limit
;
2270 unsigned long vma_size
;
2271 unsigned long nr_pages
;
2272 long user_extra
, extra
;
2275 if (!(vma
->vm_flags
& VM_SHARED
))
2278 vma_size
= vma
->vm_end
- vma
->vm_start
;
2279 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
2282 * If we have data pages ensure they're a power-of-two number, so we
2283 * can do bitmasks instead of modulo.
2285 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
2288 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
2291 if (vma
->vm_pgoff
!= 0)
2294 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
2295 mutex_lock(&counter
->mmap_mutex
);
2296 if (counter
->output
) {
2301 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
2302 if (nr_pages
!= counter
->data
->nr_pages
)
2307 user_extra
= nr_pages
+ 1;
2308 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
2311 * Increase the limit linearly with more CPUs:
2313 user_lock_limit
*= num_online_cpus();
2315 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
2318 if (user_locked
> user_lock_limit
)
2319 extra
= user_locked
- user_lock_limit
;
2321 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
2322 lock_limit
>>= PAGE_SHIFT
;
2323 locked
= vma
->vm_mm
->locked_vm
+ extra
;
2325 if ((locked
> lock_limit
) && perf_paranoid_tracepoint_raw() &&
2326 !capable(CAP_IPC_LOCK
)) {
2331 WARN_ON(counter
->data
);
2332 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
2336 atomic_set(&counter
->mmap_count
, 1);
2337 atomic_long_add(user_extra
, &user
->locked_vm
);
2338 vma
->vm_mm
->locked_vm
+= extra
;
2339 counter
->data
->nr_locked
= extra
;
2340 if (vma
->vm_flags
& VM_WRITE
)
2341 counter
->data
->writable
= 1;
2344 mutex_unlock(&counter
->mmap_mutex
);
2346 vma
->vm_flags
|= VM_RESERVED
;
2347 vma
->vm_ops
= &perf_mmap_vmops
;
2352 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2354 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2355 struct perf_counter
*counter
= filp
->private_data
;
2358 mutex_lock(&inode
->i_mutex
);
2359 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
2360 mutex_unlock(&inode
->i_mutex
);
2368 static const struct file_operations perf_fops
= {
2369 .release
= perf_release
,
2372 .unlocked_ioctl
= perf_ioctl
,
2373 .compat_ioctl
= perf_ioctl
,
2375 .fasync
= perf_fasync
,
2379 * Perf counter wakeup
2381 * If there's data, ensure we set the poll() state and publish everything
2382 * to user-space before waking everybody up.
2385 void perf_counter_wakeup(struct perf_counter
*counter
)
2387 wake_up_all(&counter
->waitq
);
2389 if (counter
->pending_kill
) {
2390 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
2391 counter
->pending_kill
= 0;
2398 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2400 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2401 * single linked list and use cmpxchg() to add entries lockless.
2404 static void perf_pending_counter(struct perf_pending_entry
*entry
)
2406 struct perf_counter
*counter
= container_of(entry
,
2407 struct perf_counter
, pending
);
2409 if (counter
->pending_disable
) {
2410 counter
->pending_disable
= 0;
2411 __perf_counter_disable(counter
);
2414 if (counter
->pending_wakeup
) {
2415 counter
->pending_wakeup
= 0;
2416 perf_counter_wakeup(counter
);
2420 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2422 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2426 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2427 void (*func
)(struct perf_pending_entry
*))
2429 struct perf_pending_entry
**head
;
2431 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2436 head
= &get_cpu_var(perf_pending_head
);
2439 entry
->next
= *head
;
2440 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2442 set_perf_counter_pending();
2444 put_cpu_var(perf_pending_head
);
2447 static int __perf_pending_run(void)
2449 struct perf_pending_entry
*list
;
2452 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2453 while (list
!= PENDING_TAIL
) {
2454 void (*func
)(struct perf_pending_entry
*);
2455 struct perf_pending_entry
*entry
= list
;
2462 * Ensure we observe the unqueue before we issue the wakeup,
2463 * so that we won't be waiting forever.
2464 * -- see perf_not_pending().
2475 static inline int perf_not_pending(struct perf_counter
*counter
)
2478 * If we flush on whatever cpu we run, there is a chance we don't
2482 __perf_pending_run();
2486 * Ensure we see the proper queue state before going to sleep
2487 * so that we do not miss the wakeup. -- see perf_pending_handle()
2490 return counter
->pending
.next
== NULL
;
2493 static void perf_pending_sync(struct perf_counter
*counter
)
2495 wait_event(counter
->waitq
, perf_not_pending(counter
));
2498 void perf_counter_do_pending(void)
2500 __perf_pending_run();
2504 * Callchain support -- arch specific
2507 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2515 static bool perf_output_space(struct perf_mmap_data
*data
, unsigned long tail
,
2516 unsigned long offset
, unsigned long head
)
2520 if (!data
->writable
)
2523 mask
= (data
->nr_pages
<< PAGE_SHIFT
) - 1;
2525 offset
= (offset
- tail
) & mask
;
2526 head
= (head
- tail
) & mask
;
2528 if ((int)(head
- offset
) < 0)
2534 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2536 atomic_set(&handle
->data
->poll
, POLL_IN
);
2539 handle
->counter
->pending_wakeup
= 1;
2540 perf_pending_queue(&handle
->counter
->pending
,
2541 perf_pending_counter
);
2543 perf_counter_wakeup(handle
->counter
);
2547 * Curious locking construct.
2549 * We need to ensure a later event doesn't publish a head when a former
2550 * event isn't done writing. However since we need to deal with NMIs we
2551 * cannot fully serialize things.
2553 * What we do is serialize between CPUs so we only have to deal with NMI
2554 * nesting on a single CPU.
2556 * We only publish the head (and generate a wakeup) when the outer-most
2559 static void perf_output_lock(struct perf_output_handle
*handle
)
2561 struct perf_mmap_data
*data
= handle
->data
;
2566 local_irq_save(handle
->flags
);
2567 cpu
= smp_processor_id();
2569 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
2572 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2578 static void perf_output_unlock(struct perf_output_handle
*handle
)
2580 struct perf_mmap_data
*data
= handle
->data
;
2584 data
->done_head
= data
->head
;
2586 if (!handle
->locked
)
2591 * The xchg implies a full barrier that ensures all writes are done
2592 * before we publish the new head, matched by a rmb() in userspace when
2593 * reading this position.
2595 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2596 data
->user_page
->data_head
= head
;
2599 * NMI can happen here, which means we can miss a done_head update.
2602 cpu
= atomic_xchg(&data
->lock
, -1);
2603 WARN_ON_ONCE(cpu
!= smp_processor_id());
2606 * Therefore we have to validate we did not indeed do so.
2608 if (unlikely(atomic_long_read(&data
->done_head
))) {
2610 * Since we had it locked, we can lock it again.
2612 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2618 if (atomic_xchg(&data
->wakeup
, 0))
2619 perf_output_wakeup(handle
);
2621 local_irq_restore(handle
->flags
);
2624 void perf_output_copy(struct perf_output_handle
*handle
,
2625 const void *buf
, unsigned int len
)
2627 unsigned int pages_mask
;
2628 unsigned int offset
;
2632 offset
= handle
->offset
;
2633 pages_mask
= handle
->data
->nr_pages
- 1;
2634 pages
= handle
->data
->data_pages
;
2637 unsigned int page_offset
;
2640 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2641 page_offset
= offset
& (PAGE_SIZE
- 1);
2642 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
2644 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2651 handle
->offset
= offset
;
2654 * Check we didn't copy past our reservation window, taking the
2655 * possible unsigned int wrap into account.
2657 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2660 int perf_output_begin(struct perf_output_handle
*handle
,
2661 struct perf_counter
*counter
, unsigned int size
,
2662 int nmi
, int sample
)
2664 struct perf_counter
*output_counter
;
2665 struct perf_mmap_data
*data
;
2666 unsigned long tail
, offset
, head
;
2669 struct perf_event_header header
;
2676 * For inherited counters we send all the output towards the parent.
2678 if (counter
->parent
)
2679 counter
= counter
->parent
;
2681 output_counter
= rcu_dereference(counter
->output
);
2683 counter
= output_counter
;
2685 data
= rcu_dereference(counter
->data
);
2689 handle
->data
= data
;
2690 handle
->counter
= counter
;
2692 handle
->sample
= sample
;
2694 if (!data
->nr_pages
)
2697 have_lost
= atomic_read(&data
->lost
);
2699 size
+= sizeof(lost_event
);
2701 perf_output_lock(handle
);
2705 * Userspace could choose to issue a mb() before updating the
2706 * tail pointer. So that all reads will be completed before the
2709 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2711 offset
= head
= atomic_long_read(&data
->head
);
2713 if (unlikely(!perf_output_space(data
, tail
, offset
, head
)))
2715 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2717 handle
->offset
= offset
;
2718 handle
->head
= head
;
2720 if (head
- tail
> data
->watermark
)
2721 atomic_set(&data
->wakeup
, 1);
2724 lost_event
.header
.type
= PERF_EVENT_LOST
;
2725 lost_event
.header
.misc
= 0;
2726 lost_event
.header
.size
= sizeof(lost_event
);
2727 lost_event
.id
= counter
->id
;
2728 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2730 perf_output_put(handle
, lost_event
);
2736 atomic_inc(&data
->lost
);
2737 perf_output_unlock(handle
);
2744 void perf_output_end(struct perf_output_handle
*handle
)
2746 struct perf_counter
*counter
= handle
->counter
;
2747 struct perf_mmap_data
*data
= handle
->data
;
2749 int wakeup_events
= counter
->attr
.wakeup_events
;
2751 if (handle
->sample
&& wakeup_events
) {
2752 int events
= atomic_inc_return(&data
->events
);
2753 if (events
>= wakeup_events
) {
2754 atomic_sub(wakeup_events
, &data
->events
);
2755 atomic_set(&data
->wakeup
, 1);
2759 perf_output_unlock(handle
);
2763 static u32
perf_counter_pid(struct perf_counter
*counter
, struct task_struct
*p
)
2766 * only top level counters have the pid namespace they were created in
2768 if (counter
->parent
)
2769 counter
= counter
->parent
;
2771 return task_tgid_nr_ns(p
, counter
->ns
);
2774 static u32
perf_counter_tid(struct perf_counter
*counter
, struct task_struct
*p
)
2777 * only top level counters have the pid namespace they were created in
2779 if (counter
->parent
)
2780 counter
= counter
->parent
;
2782 return task_pid_nr_ns(p
, counter
->ns
);
2785 static void perf_output_read_one(struct perf_output_handle
*handle
,
2786 struct perf_counter
*counter
)
2788 u64 read_format
= counter
->attr
.read_format
;
2792 values
[n
++] = atomic64_read(&counter
->count
);
2793 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
2794 values
[n
++] = counter
->total_time_enabled
+
2795 atomic64_read(&counter
->child_total_time_enabled
);
2797 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
2798 values
[n
++] = counter
->total_time_running
+
2799 atomic64_read(&counter
->child_total_time_running
);
2801 if (read_format
& PERF_FORMAT_ID
)
2802 values
[n
++] = primary_counter_id(counter
);
2804 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2808 * XXX PERF_FORMAT_GROUP vs inherited counters seems difficult.
2810 static void perf_output_read_group(struct perf_output_handle
*handle
,
2811 struct perf_counter
*counter
)
2813 struct perf_counter
*leader
= counter
->group_leader
, *sub
;
2814 u64 read_format
= counter
->attr
.read_format
;
2818 values
[n
++] = 1 + leader
->nr_siblings
;
2820 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2821 values
[n
++] = leader
->total_time_enabled
;
2823 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2824 values
[n
++] = leader
->total_time_running
;
2826 if (leader
!= counter
)
2827 leader
->pmu
->read(leader
);
2829 values
[n
++] = atomic64_read(&leader
->count
);
2830 if (read_format
& PERF_FORMAT_ID
)
2831 values
[n
++] = primary_counter_id(leader
);
2833 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2835 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2839 sub
->pmu
->read(sub
);
2841 values
[n
++] = atomic64_read(&sub
->count
);
2842 if (read_format
& PERF_FORMAT_ID
)
2843 values
[n
++] = primary_counter_id(sub
);
2845 perf_output_copy(handle
, values
, n
* sizeof(u64
));
2849 static void perf_output_read(struct perf_output_handle
*handle
,
2850 struct perf_counter
*counter
)
2852 if (counter
->attr
.read_format
& PERF_FORMAT_GROUP
)
2853 perf_output_read_group(handle
, counter
);
2855 perf_output_read_one(handle
, counter
);
2858 void perf_output_sample(struct perf_output_handle
*handle
,
2859 struct perf_event_header
*header
,
2860 struct perf_sample_data
*data
,
2861 struct perf_counter
*counter
)
2863 u64 sample_type
= data
->type
;
2865 perf_output_put(handle
, *header
);
2867 if (sample_type
& PERF_SAMPLE_IP
)
2868 perf_output_put(handle
, data
->ip
);
2870 if (sample_type
& PERF_SAMPLE_TID
)
2871 perf_output_put(handle
, data
->tid_entry
);
2873 if (sample_type
& PERF_SAMPLE_TIME
)
2874 perf_output_put(handle
, data
->time
);
2876 if (sample_type
& PERF_SAMPLE_ADDR
)
2877 perf_output_put(handle
, data
->addr
);
2879 if (sample_type
& PERF_SAMPLE_ID
)
2880 perf_output_put(handle
, data
->id
);
2882 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
2883 perf_output_put(handle
, data
->stream_id
);
2885 if (sample_type
& PERF_SAMPLE_CPU
)
2886 perf_output_put(handle
, data
->cpu_entry
);
2888 if (sample_type
& PERF_SAMPLE_PERIOD
)
2889 perf_output_put(handle
, data
->period
);
2891 if (sample_type
& PERF_SAMPLE_READ
)
2892 perf_output_read(handle
, counter
);
2894 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2895 if (data
->callchain
) {
2898 if (data
->callchain
)
2899 size
+= data
->callchain
->nr
;
2901 size
*= sizeof(u64
);
2903 perf_output_copy(handle
, data
->callchain
, size
);
2906 perf_output_put(handle
, nr
);
2910 if (sample_type
& PERF_SAMPLE_RAW
) {
2912 perf_output_put(handle
, data
->raw
->size
);
2913 perf_output_copy(handle
, data
->raw
->data
,
2920 .size
= sizeof(u32
),
2923 perf_output_put(handle
, raw
);
2928 void perf_prepare_sample(struct perf_event_header
*header
,
2929 struct perf_sample_data
*data
,
2930 struct perf_counter
*counter
,
2931 struct pt_regs
*regs
)
2933 u64 sample_type
= counter
->attr
.sample_type
;
2935 data
->type
= sample_type
;
2937 header
->type
= PERF_EVENT_SAMPLE
;
2938 header
->size
= sizeof(*header
);
2941 header
->misc
|= perf_misc_flags(regs
);
2943 if (sample_type
& PERF_SAMPLE_IP
) {
2944 data
->ip
= perf_instruction_pointer(regs
);
2946 header
->size
+= sizeof(data
->ip
);
2949 if (sample_type
& PERF_SAMPLE_TID
) {
2950 /* namespace issues */
2951 data
->tid_entry
.pid
= perf_counter_pid(counter
, current
);
2952 data
->tid_entry
.tid
= perf_counter_tid(counter
, current
);
2954 header
->size
+= sizeof(data
->tid_entry
);
2957 if (sample_type
& PERF_SAMPLE_TIME
) {
2958 data
->time
= perf_clock();
2960 header
->size
+= sizeof(data
->time
);
2963 if (sample_type
& PERF_SAMPLE_ADDR
)
2964 header
->size
+= sizeof(data
->addr
);
2966 if (sample_type
& PERF_SAMPLE_ID
) {
2967 data
->id
= primary_counter_id(counter
);
2969 header
->size
+= sizeof(data
->id
);
2972 if (sample_type
& PERF_SAMPLE_STREAM_ID
) {
2973 data
->stream_id
= counter
->id
;
2975 header
->size
+= sizeof(data
->stream_id
);
2978 if (sample_type
& PERF_SAMPLE_CPU
) {
2979 data
->cpu_entry
.cpu
= raw_smp_processor_id();
2980 data
->cpu_entry
.reserved
= 0;
2982 header
->size
+= sizeof(data
->cpu_entry
);
2985 if (sample_type
& PERF_SAMPLE_PERIOD
)
2986 header
->size
+= sizeof(data
->period
);
2988 if (sample_type
& PERF_SAMPLE_READ
)
2989 header
->size
+= perf_counter_read_size(counter
);
2991 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2994 data
->callchain
= perf_callchain(regs
);
2996 if (data
->callchain
)
2997 size
+= data
->callchain
->nr
;
2999 header
->size
+= size
* sizeof(u64
);
3002 if (sample_type
& PERF_SAMPLE_RAW
) {
3003 int size
= sizeof(u32
);
3006 size
+= data
->raw
->size
;
3008 size
+= sizeof(u32
);
3010 WARN_ON_ONCE(size
& (sizeof(u64
)-1));
3011 header
->size
+= size
;
3015 static void perf_counter_output(struct perf_counter
*counter
, int nmi
,
3016 struct perf_sample_data
*data
,
3017 struct pt_regs
*regs
)
3019 struct perf_output_handle handle
;
3020 struct perf_event_header header
;
3022 perf_prepare_sample(&header
, data
, counter
, regs
);
3024 if (perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1))
3027 perf_output_sample(&handle
, &header
, data
, counter
);
3029 perf_output_end(&handle
);
3036 struct perf_read_event
{
3037 struct perf_event_header header
;
3044 perf_counter_read_event(struct perf_counter
*counter
,
3045 struct task_struct
*task
)
3047 struct perf_output_handle handle
;
3048 struct perf_read_event event
= {
3050 .type
= PERF_EVENT_READ
,
3052 .size
= sizeof(event
) + perf_counter_read_size(counter
),
3054 .pid
= perf_counter_pid(counter
, task
),
3055 .tid
= perf_counter_tid(counter
, task
),
3059 ret
= perf_output_begin(&handle
, counter
, event
.header
.size
, 0, 0);
3063 perf_output_put(&handle
, event
);
3064 perf_output_read(&handle
, counter
);
3066 perf_output_end(&handle
);
3070 * task tracking -- fork/exit
3072 * enabled by: attr.comm | attr.mmap | attr.task
3075 struct perf_task_event
{
3076 struct task_struct
*task
;
3077 struct perf_counter_context
*task_ctx
;
3080 struct perf_event_header header
;
3090 static void perf_counter_task_output(struct perf_counter
*counter
,
3091 struct perf_task_event
*task_event
)
3093 struct perf_output_handle handle
;
3095 struct task_struct
*task
= task_event
->task
;
3098 size
= task_event
->event
.header
.size
;
3099 ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
3104 task_event
->event
.pid
= perf_counter_pid(counter
, task
);
3105 task_event
->event
.ppid
= perf_counter_pid(counter
, current
);
3107 task_event
->event
.tid
= perf_counter_tid(counter
, task
);
3108 task_event
->event
.ptid
= perf_counter_tid(counter
, current
);
3110 task_event
->event
.time
= perf_clock();
3112 perf_output_put(&handle
, task_event
->event
);
3114 perf_output_end(&handle
);
3117 static int perf_counter_task_match(struct perf_counter
*counter
)
3119 if (counter
->attr
.comm
|| counter
->attr
.mmap
|| counter
->attr
.task
)
3125 static void perf_counter_task_ctx(struct perf_counter_context
*ctx
,
3126 struct perf_task_event
*task_event
)
3128 struct perf_counter
*counter
;
3130 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3134 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3135 if (perf_counter_task_match(counter
))
3136 perf_counter_task_output(counter
, task_event
);
3141 static void perf_counter_task_event(struct perf_task_event
*task_event
)
3143 struct perf_cpu_context
*cpuctx
;
3144 struct perf_counter_context
*ctx
= task_event
->task_ctx
;
3146 cpuctx
= &get_cpu_var(perf_cpu_context
);
3147 perf_counter_task_ctx(&cpuctx
->ctx
, task_event
);
3148 put_cpu_var(perf_cpu_context
);
3152 ctx
= rcu_dereference(task_event
->task
->perf_counter_ctxp
);
3154 perf_counter_task_ctx(ctx
, task_event
);
3158 static void perf_counter_task(struct task_struct
*task
,
3159 struct perf_counter_context
*task_ctx
,
3162 struct perf_task_event task_event
;
3164 if (!atomic_read(&nr_comm_counters
) &&
3165 !atomic_read(&nr_mmap_counters
) &&
3166 !atomic_read(&nr_task_counters
))
3169 task_event
= (struct perf_task_event
){
3171 .task_ctx
= task_ctx
,
3174 .type
= new ? PERF_EVENT_FORK
: PERF_EVENT_EXIT
,
3176 .size
= sizeof(task_event
.event
),
3185 perf_counter_task_event(&task_event
);
3188 void perf_counter_fork(struct task_struct
*task
)
3190 perf_counter_task(task
, NULL
, 1);
3197 struct perf_comm_event
{
3198 struct task_struct
*task
;
3203 struct perf_event_header header
;
3210 static void perf_counter_comm_output(struct perf_counter
*counter
,
3211 struct perf_comm_event
*comm_event
)
3213 struct perf_output_handle handle
;
3214 int size
= comm_event
->event
.header
.size
;
3215 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
3220 comm_event
->event
.pid
= perf_counter_pid(counter
, comm_event
->task
);
3221 comm_event
->event
.tid
= perf_counter_tid(counter
, comm_event
->task
);
3223 perf_output_put(&handle
, comm_event
->event
);
3224 perf_output_copy(&handle
, comm_event
->comm
,
3225 comm_event
->comm_size
);
3226 perf_output_end(&handle
);
3229 static int perf_counter_comm_match(struct perf_counter
*counter
)
3231 if (counter
->attr
.comm
)
3237 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
3238 struct perf_comm_event
*comm_event
)
3240 struct perf_counter
*counter
;
3242 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3246 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3247 if (perf_counter_comm_match(counter
))
3248 perf_counter_comm_output(counter
, comm_event
);
3253 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
3255 struct perf_cpu_context
*cpuctx
;
3256 struct perf_counter_context
*ctx
;
3258 char comm
[TASK_COMM_LEN
];
3260 memset(comm
, 0, sizeof(comm
));
3261 strncpy(comm
, comm_event
->task
->comm
, sizeof(comm
));
3262 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
3264 comm_event
->comm
= comm
;
3265 comm_event
->comm_size
= size
;
3267 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
3269 cpuctx
= &get_cpu_var(perf_cpu_context
);
3270 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
3271 put_cpu_var(perf_cpu_context
);
3275 * doesn't really matter which of the child contexts the
3276 * events ends up in.
3278 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3280 perf_counter_comm_ctx(ctx
, comm_event
);
3284 void perf_counter_comm(struct task_struct
*task
)
3286 struct perf_comm_event comm_event
;
3288 if (task
->perf_counter_ctxp
)
3289 perf_counter_enable_on_exec(task
);
3291 if (!atomic_read(&nr_comm_counters
))
3294 comm_event
= (struct perf_comm_event
){
3300 .type
= PERF_EVENT_COMM
,
3309 perf_counter_comm_event(&comm_event
);
3316 struct perf_mmap_event
{
3317 struct vm_area_struct
*vma
;
3319 const char *file_name
;
3323 struct perf_event_header header
;
3333 static void perf_counter_mmap_output(struct perf_counter
*counter
,
3334 struct perf_mmap_event
*mmap_event
)
3336 struct perf_output_handle handle
;
3337 int size
= mmap_event
->event
.header
.size
;
3338 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
3343 mmap_event
->event
.pid
= perf_counter_pid(counter
, current
);
3344 mmap_event
->event
.tid
= perf_counter_tid(counter
, current
);
3346 perf_output_put(&handle
, mmap_event
->event
);
3347 perf_output_copy(&handle
, mmap_event
->file_name
,
3348 mmap_event
->file_size
);
3349 perf_output_end(&handle
);
3352 static int perf_counter_mmap_match(struct perf_counter
*counter
,
3353 struct perf_mmap_event
*mmap_event
)
3355 if (counter
->attr
.mmap
)
3361 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
3362 struct perf_mmap_event
*mmap_event
)
3364 struct perf_counter
*counter
;
3366 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3370 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3371 if (perf_counter_mmap_match(counter
, mmap_event
))
3372 perf_counter_mmap_output(counter
, mmap_event
);
3377 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
3379 struct perf_cpu_context
*cpuctx
;
3380 struct perf_counter_context
*ctx
;
3381 struct vm_area_struct
*vma
= mmap_event
->vma
;
3382 struct file
*file
= vma
->vm_file
;
3388 memset(tmp
, 0, sizeof(tmp
));
3392 * d_path works from the end of the buffer backwards, so we
3393 * need to add enough zero bytes after the string to handle
3394 * the 64bit alignment we do later.
3396 buf
= kzalloc(PATH_MAX
+ sizeof(u64
), GFP_KERNEL
);
3398 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
3401 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
3403 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
3407 if (arch_vma_name(mmap_event
->vma
)) {
3408 name
= strncpy(tmp
, arch_vma_name(mmap_event
->vma
),
3414 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
3418 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
3423 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
3425 mmap_event
->file_name
= name
;
3426 mmap_event
->file_size
= size
;
3428 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
3430 cpuctx
= &get_cpu_var(perf_cpu_context
);
3431 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
3432 put_cpu_var(perf_cpu_context
);
3436 * doesn't really matter which of the child contexts the
3437 * events ends up in.
3439 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3441 perf_counter_mmap_ctx(ctx
, mmap_event
);
3447 void __perf_counter_mmap(struct vm_area_struct
*vma
)
3449 struct perf_mmap_event mmap_event
;
3451 if (!atomic_read(&nr_mmap_counters
))
3454 mmap_event
= (struct perf_mmap_event
){
3460 .type
= PERF_EVENT_MMAP
,
3466 .start
= vma
->vm_start
,
3467 .len
= vma
->vm_end
- vma
->vm_start
,
3468 .pgoff
= vma
->vm_pgoff
,
3472 perf_counter_mmap_event(&mmap_event
);
3476 * IRQ throttle logging
3479 static void perf_log_throttle(struct perf_counter
*counter
, int enable
)
3481 struct perf_output_handle handle
;
3485 struct perf_event_header header
;
3489 } throttle_event
= {
3491 .type
= PERF_EVENT_THROTTLE
,
3493 .size
= sizeof(throttle_event
),
3495 .time
= perf_clock(),
3496 .id
= primary_counter_id(counter
),
3497 .stream_id
= counter
->id
,
3501 throttle_event
.header
.type
= PERF_EVENT_UNTHROTTLE
;
3503 ret
= perf_output_begin(&handle
, counter
, sizeof(throttle_event
), 1, 0);
3507 perf_output_put(&handle
, throttle_event
);
3508 perf_output_end(&handle
);
3512 * Generic counter overflow handling, sampling.
3515 static int __perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
3516 int throttle
, struct perf_sample_data
*data
,
3517 struct pt_regs
*regs
)
3519 int events
= atomic_read(&counter
->event_limit
);
3520 struct hw_perf_counter
*hwc
= &counter
->hw
;
3523 throttle
= (throttle
&& counter
->pmu
->unthrottle
!= NULL
);
3528 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3530 if (HZ
* hwc
->interrupts
>
3531 (u64
)sysctl_perf_counter_sample_rate
) {
3532 hwc
->interrupts
= MAX_INTERRUPTS
;
3533 perf_log_throttle(counter
, 0);
3538 * Keep re-disabling counters even though on the previous
3539 * pass we disabled it - just in case we raced with a
3540 * sched-in and the counter got enabled again:
3546 if (counter
->attr
.freq
) {
3547 u64 now
= perf_clock();
3548 s64 delta
= now
- hwc
->freq_stamp
;
3550 hwc
->freq_stamp
= now
;
3552 if (delta
> 0 && delta
< TICK_NSEC
)
3553 perf_adjust_period(counter
, NSEC_PER_SEC
/ (int)delta
);
3557 * XXX event_limit might not quite work as expected on inherited
3561 counter
->pending_kill
= POLL_IN
;
3562 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
3564 counter
->pending_kill
= POLL_HUP
;
3566 counter
->pending_disable
= 1;
3567 perf_pending_queue(&counter
->pending
,
3568 perf_pending_counter
);
3570 perf_counter_disable(counter
);
3573 perf_counter_output(counter
, nmi
, data
, regs
);
3577 int perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
3578 struct perf_sample_data
*data
,
3579 struct pt_regs
*regs
)
3581 return __perf_counter_overflow(counter
, nmi
, 1, data
, regs
);
3585 * Generic software counter infrastructure
3589 * We directly increment counter->count and keep a second value in
3590 * counter->hw.period_left to count intervals. This period counter
3591 * is kept in the range [-sample_period, 0] so that we can use the
3595 static u64
perf_swcounter_set_period(struct perf_counter
*counter
)
3597 struct hw_perf_counter
*hwc
= &counter
->hw
;
3598 u64 period
= hwc
->last_period
;
3602 hwc
->last_period
= hwc
->sample_period
;
3605 old
= val
= atomic64_read(&hwc
->period_left
);
3609 nr
= div64_u64(period
+ val
, period
);
3610 offset
= nr
* period
;
3612 if (atomic64_cmpxchg(&hwc
->period_left
, old
, val
) != old
)
3618 static void perf_swcounter_overflow(struct perf_counter
*counter
,
3619 int nmi
, struct perf_sample_data
*data
,
3620 struct pt_regs
*regs
)
3622 struct hw_perf_counter
*hwc
= &counter
->hw
;
3626 data
->period
= counter
->hw
.last_period
;
3627 overflow
= perf_swcounter_set_period(counter
);
3629 if (hwc
->interrupts
== MAX_INTERRUPTS
)
3632 for (; overflow
; overflow
--) {
3633 if (__perf_counter_overflow(counter
, nmi
, throttle
,
3636 * We inhibit the overflow from happening when
3637 * hwc->interrupts == MAX_INTERRUPTS.
3645 static void perf_swcounter_unthrottle(struct perf_counter
*counter
)
3648 * Nothing to do, we already reset hwc->interrupts.
3652 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
3653 int nmi
, struct perf_sample_data
*data
,
3654 struct pt_regs
*regs
)
3656 struct hw_perf_counter
*hwc
= &counter
->hw
;
3658 atomic64_add(nr
, &counter
->count
);
3660 if (!hwc
->sample_period
)
3666 if (!atomic64_add_negative(nr
, &hwc
->period_left
))
3667 perf_swcounter_overflow(counter
, nmi
, data
, regs
);
3670 static int perf_swcounter_is_counting(struct perf_counter
*counter
)
3673 * The counter is active, we're good!
3675 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
3679 * The counter is off/error, not counting.
3681 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
)
3685 * The counter is inactive, if the context is active
3686 * we're part of a group that didn't make it on the 'pmu',
3689 if (counter
->ctx
->is_active
)
3693 * We're inactive and the context is too, this means the
3694 * task is scheduled out, we're counting events that happen
3695 * to us, like migration events.
3700 static int perf_swcounter_match(struct perf_counter
*counter
,
3701 enum perf_type_id type
,
3702 u32 event
, struct pt_regs
*regs
)
3704 if (!perf_swcounter_is_counting(counter
))
3707 if (counter
->attr
.type
!= type
)
3709 if (counter
->attr
.config
!= event
)
3713 if (counter
->attr
.exclude_user
&& user_mode(regs
))
3716 if (counter
->attr
.exclude_kernel
&& !user_mode(regs
))
3723 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
3724 enum perf_type_id type
,
3725 u32 event
, u64 nr
, int nmi
,
3726 struct perf_sample_data
*data
,
3727 struct pt_regs
*regs
)
3729 struct perf_counter
*counter
;
3731 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3735 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3736 if (perf_swcounter_match(counter
, type
, event
, regs
))
3737 perf_swcounter_add(counter
, nr
, nmi
, data
, regs
);
3742 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
3745 return &cpuctx
->recursion
[3];
3748 return &cpuctx
->recursion
[2];
3751 return &cpuctx
->recursion
[1];
3753 return &cpuctx
->recursion
[0];
3756 static void do_perf_swcounter_event(enum perf_type_id type
, u32 event
,
3758 struct perf_sample_data
*data
,
3759 struct pt_regs
*regs
)
3761 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3762 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
3763 struct perf_counter_context
*ctx
;
3771 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
3772 nr
, nmi
, data
, regs
);
3775 * doesn't really matter which of the child contexts the
3776 * events ends up in.
3778 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3780 perf_swcounter_ctx_event(ctx
, type
, event
, nr
, nmi
, data
, regs
);
3787 put_cpu_var(perf_cpu_context
);
3790 void __perf_swcounter_event(u32 event
, u64 nr
, int nmi
,
3791 struct pt_regs
*regs
, u64 addr
)
3793 struct perf_sample_data data
= {
3797 do_perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
,
3801 static void perf_swcounter_read(struct perf_counter
*counter
)
3805 static int perf_swcounter_enable(struct perf_counter
*counter
)
3807 struct hw_perf_counter
*hwc
= &counter
->hw
;
3809 if (hwc
->sample_period
) {
3810 hwc
->last_period
= hwc
->sample_period
;
3811 perf_swcounter_set_period(counter
);
3816 static void perf_swcounter_disable(struct perf_counter
*counter
)
3820 static const struct pmu perf_ops_generic
= {
3821 .enable
= perf_swcounter_enable
,
3822 .disable
= perf_swcounter_disable
,
3823 .read
= perf_swcounter_read
,
3824 .unthrottle
= perf_swcounter_unthrottle
,
3828 * hrtimer based swcounter callback
3831 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
3833 enum hrtimer_restart ret
= HRTIMER_RESTART
;
3834 struct perf_sample_data data
;
3835 struct pt_regs
*regs
;
3836 struct perf_counter
*counter
;
3839 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
3840 counter
->pmu
->read(counter
);
3843 regs
= get_irq_regs();
3845 * In case we exclude kernel IPs or are somehow not in interrupt
3846 * context, provide the next best thing, the user IP.
3848 if ((counter
->attr
.exclude_kernel
|| !regs
) &&
3849 !counter
->attr
.exclude_user
)
3850 regs
= task_pt_regs(current
);
3853 if (perf_counter_overflow(counter
, 0, &data
, regs
))
3854 ret
= HRTIMER_NORESTART
;
3857 period
= max_t(u64
, 10000, counter
->hw
.sample_period
);
3858 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
3864 * Software counter: cpu wall time clock
3867 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
3869 int cpu
= raw_smp_processor_id();
3873 now
= cpu_clock(cpu
);
3874 prev
= atomic64_read(&counter
->hw
.prev_count
);
3875 atomic64_set(&counter
->hw
.prev_count
, now
);
3876 atomic64_add(now
- prev
, &counter
->count
);
3879 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
3881 struct hw_perf_counter
*hwc
= &counter
->hw
;
3882 int cpu
= raw_smp_processor_id();
3884 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
3885 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3886 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3887 if (hwc
->sample_period
) {
3888 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3889 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3890 ns_to_ktime(period
), 0,
3891 HRTIMER_MODE_REL
, 0);
3897 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
3899 if (counter
->hw
.sample_period
)
3900 hrtimer_cancel(&counter
->hw
.hrtimer
);
3901 cpu_clock_perf_counter_update(counter
);
3904 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
3906 cpu_clock_perf_counter_update(counter
);
3909 static const struct pmu perf_ops_cpu_clock
= {
3910 .enable
= cpu_clock_perf_counter_enable
,
3911 .disable
= cpu_clock_perf_counter_disable
,
3912 .read
= cpu_clock_perf_counter_read
,
3916 * Software counter: task time clock
3919 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
3924 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
3926 atomic64_add(delta
, &counter
->count
);
3929 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
3931 struct hw_perf_counter
*hwc
= &counter
->hw
;
3934 now
= counter
->ctx
->time
;
3936 atomic64_set(&hwc
->prev_count
, now
);
3937 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3938 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3939 if (hwc
->sample_period
) {
3940 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3941 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3942 ns_to_ktime(period
), 0,
3943 HRTIMER_MODE_REL
, 0);
3949 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
3951 if (counter
->hw
.sample_period
)
3952 hrtimer_cancel(&counter
->hw
.hrtimer
);
3953 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
3957 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
3962 update_context_time(counter
->ctx
);
3963 time
= counter
->ctx
->time
;
3965 u64 now
= perf_clock();
3966 u64 delta
= now
- counter
->ctx
->timestamp
;
3967 time
= counter
->ctx
->time
+ delta
;
3970 task_clock_perf_counter_update(counter
, time
);
3973 static const struct pmu perf_ops_task_clock
= {
3974 .enable
= task_clock_perf_counter_enable
,
3975 .disable
= task_clock_perf_counter_disable
,
3976 .read
= task_clock_perf_counter_read
,
3979 #ifdef CONFIG_EVENT_PROFILE
3980 void perf_tpcounter_event(int event_id
, u64 addr
, u64 count
, void *record
,
3983 struct perf_raw_record raw
= {
3988 struct perf_sample_data data
= {
3993 struct pt_regs
*regs
= get_irq_regs();
3996 regs
= task_pt_regs(current
);
3998 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, count
, 1,
4001 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
4003 extern int ftrace_profile_enable(int);
4004 extern void ftrace_profile_disable(int);
4006 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
4008 ftrace_profile_disable(counter
->attr
.config
);
4011 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
4014 * Raw tracepoint data is a severe data leak, only allow root to
4017 if ((counter
->attr
.sample_type
& PERF_SAMPLE_RAW
) &&
4018 perf_paranoid_tracepoint_raw() &&
4019 !capable(CAP_SYS_ADMIN
))
4020 return ERR_PTR(-EPERM
);
4022 if (ftrace_profile_enable(counter
->attr
.config
))
4025 counter
->destroy
= tp_perf_counter_destroy
;
4027 return &perf_ops_generic
;
4030 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
4036 atomic_t perf_swcounter_enabled
[PERF_COUNT_SW_MAX
];
4038 static void sw_perf_counter_destroy(struct perf_counter
*counter
)
4040 u64 event
= counter
->attr
.config
;
4042 WARN_ON(counter
->parent
);
4044 atomic_dec(&perf_swcounter_enabled
[event
]);
4047 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
4049 const struct pmu
*pmu
= NULL
;
4050 u64 event
= counter
->attr
.config
;
4053 * Software counters (currently) can't in general distinguish
4054 * between user, kernel and hypervisor events.
4055 * However, context switches and cpu migrations are considered
4056 * to be kernel events, and page faults are never hypervisor
4060 case PERF_COUNT_SW_CPU_CLOCK
:
4061 pmu
= &perf_ops_cpu_clock
;
4064 case PERF_COUNT_SW_TASK_CLOCK
:
4066 * If the user instantiates this as a per-cpu counter,
4067 * use the cpu_clock counter instead.
4069 if (counter
->ctx
->task
)
4070 pmu
= &perf_ops_task_clock
;
4072 pmu
= &perf_ops_cpu_clock
;
4075 case PERF_COUNT_SW_PAGE_FAULTS
:
4076 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
4077 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
4078 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
4079 case PERF_COUNT_SW_CPU_MIGRATIONS
:
4080 if (!counter
->parent
) {
4081 atomic_inc(&perf_swcounter_enabled
[event
]);
4082 counter
->destroy
= sw_perf_counter_destroy
;
4084 pmu
= &perf_ops_generic
;
4092 * Allocate and initialize a counter structure
4094 static struct perf_counter
*
4095 perf_counter_alloc(struct perf_counter_attr
*attr
,
4097 struct perf_counter_context
*ctx
,
4098 struct perf_counter
*group_leader
,
4099 struct perf_counter
*parent_counter
,
4102 const struct pmu
*pmu
;
4103 struct perf_counter
*counter
;
4104 struct hw_perf_counter
*hwc
;
4107 counter
= kzalloc(sizeof(*counter
), gfpflags
);
4109 return ERR_PTR(-ENOMEM
);
4112 * Single counters are their own group leaders, with an
4113 * empty sibling list:
4116 group_leader
= counter
;
4118 mutex_init(&counter
->child_mutex
);
4119 INIT_LIST_HEAD(&counter
->child_list
);
4121 INIT_LIST_HEAD(&counter
->list_entry
);
4122 INIT_LIST_HEAD(&counter
->event_entry
);
4123 INIT_LIST_HEAD(&counter
->sibling_list
);
4124 init_waitqueue_head(&counter
->waitq
);
4126 mutex_init(&counter
->mmap_mutex
);
4129 counter
->attr
= *attr
;
4130 counter
->group_leader
= group_leader
;
4131 counter
->pmu
= NULL
;
4133 counter
->oncpu
= -1;
4135 counter
->parent
= parent_counter
;
4137 counter
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
4138 counter
->id
= atomic64_inc_return(&perf_counter_id
);
4140 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
4143 counter
->state
= PERF_COUNTER_STATE_OFF
;
4148 hwc
->sample_period
= attr
->sample_period
;
4149 if (attr
->freq
&& attr
->sample_freq
)
4150 hwc
->sample_period
= 1;
4151 hwc
->last_period
= hwc
->sample_period
;
4153 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
4156 * we currently do not support PERF_FORMAT_GROUP on inherited counters
4158 if (attr
->inherit
&& (attr
->read_format
& PERF_FORMAT_GROUP
))
4161 switch (attr
->type
) {
4163 case PERF_TYPE_HARDWARE
:
4164 case PERF_TYPE_HW_CACHE
:
4165 pmu
= hw_perf_counter_init(counter
);
4168 case PERF_TYPE_SOFTWARE
:
4169 pmu
= sw_perf_counter_init(counter
);
4172 case PERF_TYPE_TRACEPOINT
:
4173 pmu
= tp_perf_counter_init(counter
);
4183 else if (IS_ERR(pmu
))
4188 put_pid_ns(counter
->ns
);
4190 return ERR_PTR(err
);
4195 if (!counter
->parent
) {
4196 atomic_inc(&nr_counters
);
4197 if (counter
->attr
.mmap
)
4198 atomic_inc(&nr_mmap_counters
);
4199 if (counter
->attr
.comm
)
4200 atomic_inc(&nr_comm_counters
);
4201 if (counter
->attr
.task
)
4202 atomic_inc(&nr_task_counters
);
4208 static int perf_copy_attr(struct perf_counter_attr __user
*uattr
,
4209 struct perf_counter_attr
*attr
)
4214 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
4218 * zero the full structure, so that a short copy will be nice.
4220 memset(attr
, 0, sizeof(*attr
));
4222 ret
= get_user(size
, &uattr
->size
);
4226 if (size
> PAGE_SIZE
) /* silly large */
4229 if (!size
) /* abi compat */
4230 size
= PERF_ATTR_SIZE_VER0
;
4232 if (size
< PERF_ATTR_SIZE_VER0
)
4236 * If we're handed a bigger struct than we know of,
4237 * ensure all the unknown bits are 0 - i.e. new
4238 * user-space does not rely on any kernel feature
4239 * extensions we dont know about yet.
4241 if (size
> sizeof(*attr
)) {
4242 unsigned char __user
*addr
;
4243 unsigned char __user
*end
;
4246 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4247 end
= (void __user
*)uattr
+ size
;
4249 for (; addr
< end
; addr
++) {
4250 ret
= get_user(val
, addr
);
4256 size
= sizeof(*attr
);
4259 ret
= copy_from_user(attr
, uattr
, size
);
4264 * If the type exists, the corresponding creation will verify
4267 if (attr
->type
>= PERF_TYPE_MAX
)
4270 if (attr
->__reserved_1
|| attr
->__reserved_2
|| attr
->__reserved_3
)
4273 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
4276 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
4283 put_user(sizeof(*attr
), &uattr
->size
);
4288 int perf_counter_set_output(struct perf_counter
*counter
, int output_fd
)
4290 struct perf_counter
*output_counter
= NULL
;
4291 struct file
*output_file
= NULL
;
4292 struct perf_counter
*old_output
;
4293 int fput_needed
= 0;
4299 output_file
= fget_light(output_fd
, &fput_needed
);
4303 if (output_file
->f_op
!= &perf_fops
)
4306 output_counter
= output_file
->private_data
;
4308 /* Don't chain output fds */
4309 if (output_counter
->output
)
4312 /* Don't set an output fd when we already have an output channel */
4316 atomic_long_inc(&output_file
->f_count
);
4319 mutex_lock(&counter
->mmap_mutex
);
4320 old_output
= counter
->output
;
4321 rcu_assign_pointer(counter
->output
, output_counter
);
4322 mutex_unlock(&counter
->mmap_mutex
);
4326 * we need to make sure no existing perf_output_*()
4327 * is still referencing this counter.
4330 fput(old_output
->filp
);
4335 fput_light(output_file
, fput_needed
);
4340 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
4342 * @attr_uptr: event type attributes for monitoring/sampling
4345 * @group_fd: group leader counter fd
4347 SYSCALL_DEFINE5(perf_counter_open
,
4348 struct perf_counter_attr __user
*, attr_uptr
,
4349 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
4351 struct perf_counter
*counter
, *group_leader
;
4352 struct perf_counter_attr attr
;
4353 struct perf_counter_context
*ctx
;
4354 struct file
*counter_file
= NULL
;
4355 struct file
*group_file
= NULL
;
4356 int fput_needed
= 0;
4357 int fput_needed2
= 0;
4360 /* for future expandability... */
4361 if (flags
& ~(PERF_FLAG_FD_NO_GROUP
| PERF_FLAG_FD_OUTPUT
))
4364 err
= perf_copy_attr(attr_uptr
, &attr
);
4368 if (!attr
.exclude_kernel
) {
4369 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
4374 if (attr
.sample_freq
> sysctl_perf_counter_sample_rate
)
4379 * Get the target context (task or percpu):
4381 ctx
= find_get_context(pid
, cpu
);
4383 return PTR_ERR(ctx
);
4386 * Look up the group leader (we will attach this counter to it):
4388 group_leader
= NULL
;
4389 if (group_fd
!= -1 && !(flags
& PERF_FLAG_FD_NO_GROUP
)) {
4391 group_file
= fget_light(group_fd
, &fput_needed
);
4393 goto err_put_context
;
4394 if (group_file
->f_op
!= &perf_fops
)
4395 goto err_put_context
;
4397 group_leader
= group_file
->private_data
;
4399 * Do not allow a recursive hierarchy (this new sibling
4400 * becoming part of another group-sibling):
4402 if (group_leader
->group_leader
!= group_leader
)
4403 goto err_put_context
;
4405 * Do not allow to attach to a group in a different
4406 * task or CPU context:
4408 if (group_leader
->ctx
!= ctx
)
4409 goto err_put_context
;
4411 * Only a group leader can be exclusive or pinned
4413 if (attr
.exclusive
|| attr
.pinned
)
4414 goto err_put_context
;
4417 counter
= perf_counter_alloc(&attr
, cpu
, ctx
, group_leader
,
4419 err
= PTR_ERR(counter
);
4420 if (IS_ERR(counter
))
4421 goto err_put_context
;
4423 err
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
4425 goto err_free_put_context
;
4427 counter_file
= fget_light(err
, &fput_needed2
);
4429 goto err_free_put_context
;
4431 if (flags
& PERF_FLAG_FD_OUTPUT
) {
4432 err
= perf_counter_set_output(counter
, group_fd
);
4434 goto err_fput_free_put_context
;
4437 counter
->filp
= counter_file
;
4438 WARN_ON_ONCE(ctx
->parent_ctx
);
4439 mutex_lock(&ctx
->mutex
);
4440 perf_install_in_context(ctx
, counter
, cpu
);
4442 mutex_unlock(&ctx
->mutex
);
4444 counter
->owner
= current
;
4445 get_task_struct(current
);
4446 mutex_lock(¤t
->perf_counter_mutex
);
4447 list_add_tail(&counter
->owner_entry
, ¤t
->perf_counter_list
);
4448 mutex_unlock(¤t
->perf_counter_mutex
);
4450 err_fput_free_put_context
:
4451 fput_light(counter_file
, fput_needed2
);
4453 err_free_put_context
:
4461 fput_light(group_file
, fput_needed
);
4467 * inherit a counter from parent task to child task:
4469 static struct perf_counter
*
4470 inherit_counter(struct perf_counter
*parent_counter
,
4471 struct task_struct
*parent
,
4472 struct perf_counter_context
*parent_ctx
,
4473 struct task_struct
*child
,
4474 struct perf_counter
*group_leader
,
4475 struct perf_counter_context
*child_ctx
)
4477 struct perf_counter
*child_counter
;
4480 * Instead of creating recursive hierarchies of counters,
4481 * we link inherited counters back to the original parent,
4482 * which has a filp for sure, which we use as the reference
4485 if (parent_counter
->parent
)
4486 parent_counter
= parent_counter
->parent
;
4488 child_counter
= perf_counter_alloc(&parent_counter
->attr
,
4489 parent_counter
->cpu
, child_ctx
,
4490 group_leader
, parent_counter
,
4492 if (IS_ERR(child_counter
))
4493 return child_counter
;
4497 * Make the child state follow the state of the parent counter,
4498 * not its attr.disabled bit. We hold the parent's mutex,
4499 * so we won't race with perf_counter_{en, dis}able_family.
4501 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
4502 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
4504 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
4506 if (parent_counter
->attr
.freq
)
4507 child_counter
->hw
.sample_period
= parent_counter
->hw
.sample_period
;
4510 * Link it up in the child's context:
4512 add_counter_to_ctx(child_counter
, child_ctx
);
4515 * Get a reference to the parent filp - we will fput it
4516 * when the child counter exits. This is safe to do because
4517 * we are in the parent and we know that the filp still
4518 * exists and has a nonzero count:
4520 atomic_long_inc(&parent_counter
->filp
->f_count
);
4523 * Link this into the parent counter's child list
4525 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
4526 mutex_lock(&parent_counter
->child_mutex
);
4527 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
4528 mutex_unlock(&parent_counter
->child_mutex
);
4530 return child_counter
;
4533 static int inherit_group(struct perf_counter
*parent_counter
,
4534 struct task_struct
*parent
,
4535 struct perf_counter_context
*parent_ctx
,
4536 struct task_struct
*child
,
4537 struct perf_counter_context
*child_ctx
)
4539 struct perf_counter
*leader
;
4540 struct perf_counter
*sub
;
4541 struct perf_counter
*child_ctr
;
4543 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
4544 child
, NULL
, child_ctx
);
4546 return PTR_ERR(leader
);
4547 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
4548 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
4549 child
, leader
, child_ctx
);
4550 if (IS_ERR(child_ctr
))
4551 return PTR_ERR(child_ctr
);
4556 static void sync_child_counter(struct perf_counter
*child_counter
,
4557 struct task_struct
*child
)
4559 struct perf_counter
*parent_counter
= child_counter
->parent
;
4562 if (child_counter
->attr
.inherit_stat
)
4563 perf_counter_read_event(child_counter
, child
);
4565 child_val
= atomic64_read(&child_counter
->count
);
4568 * Add back the child's count to the parent's count:
4570 atomic64_add(child_val
, &parent_counter
->count
);
4571 atomic64_add(child_counter
->total_time_enabled
,
4572 &parent_counter
->child_total_time_enabled
);
4573 atomic64_add(child_counter
->total_time_running
,
4574 &parent_counter
->child_total_time_running
);
4577 * Remove this counter from the parent's list
4579 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
4580 mutex_lock(&parent_counter
->child_mutex
);
4581 list_del_init(&child_counter
->child_list
);
4582 mutex_unlock(&parent_counter
->child_mutex
);
4585 * Release the parent counter, if this was the last
4588 fput(parent_counter
->filp
);
4592 __perf_counter_exit_task(struct perf_counter
*child_counter
,
4593 struct perf_counter_context
*child_ctx
,
4594 struct task_struct
*child
)
4596 struct perf_counter
*parent_counter
;
4598 update_counter_times(child_counter
);
4599 perf_counter_remove_from_context(child_counter
);
4601 parent_counter
= child_counter
->parent
;
4603 * It can happen that parent exits first, and has counters
4604 * that are still around due to the child reference. These
4605 * counters need to be zapped - but otherwise linger.
4607 if (parent_counter
) {
4608 sync_child_counter(child_counter
, child
);
4609 free_counter(child_counter
);
4614 * When a child task exits, feed back counter values to parent counters.
4616 void perf_counter_exit_task(struct task_struct
*child
)
4618 struct perf_counter
*child_counter
, *tmp
;
4619 struct perf_counter_context
*child_ctx
;
4620 unsigned long flags
;
4622 if (likely(!child
->perf_counter_ctxp
)) {
4623 perf_counter_task(child
, NULL
, 0);
4627 local_irq_save(flags
);
4629 * We can't reschedule here because interrupts are disabled,
4630 * and either child is current or it is a task that can't be
4631 * scheduled, so we are now safe from rescheduling changing
4634 child_ctx
= child
->perf_counter_ctxp
;
4635 __perf_counter_task_sched_out(child_ctx
);
4638 * Take the context lock here so that if find_get_context is
4639 * reading child->perf_counter_ctxp, we wait until it has
4640 * incremented the context's refcount before we do put_ctx below.
4642 spin_lock(&child_ctx
->lock
);
4643 child
->perf_counter_ctxp
= NULL
;
4645 * If this context is a clone; unclone it so it can't get
4646 * swapped to another process while we're removing all
4647 * the counters from it.
4649 unclone_ctx(child_ctx
);
4650 spin_unlock_irqrestore(&child_ctx
->lock
, flags
);
4653 * Report the task dead after unscheduling the counters so that we
4654 * won't get any samples after PERF_EVENT_EXIT. We can however still
4655 * get a few PERF_EVENT_READ events.
4657 perf_counter_task(child
, child_ctx
, 0);
4660 * We can recurse on the same lock type through:
4662 * __perf_counter_exit_task()
4663 * sync_child_counter()
4664 * fput(parent_counter->filp)
4666 * mutex_lock(&ctx->mutex)
4668 * But since its the parent context it won't be the same instance.
4670 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
4673 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
4675 __perf_counter_exit_task(child_counter
, child_ctx
, child
);
4678 * If the last counter was a group counter, it will have appended all
4679 * its siblings to the list, but we obtained 'tmp' before that which
4680 * will still point to the list head terminating the iteration.
4682 if (!list_empty(&child_ctx
->counter_list
))
4685 mutex_unlock(&child_ctx
->mutex
);
4691 * free an unexposed, unused context as created by inheritance by
4692 * init_task below, used by fork() in case of fail.
4694 void perf_counter_free_task(struct task_struct
*task
)
4696 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
4697 struct perf_counter
*counter
, *tmp
;
4702 mutex_lock(&ctx
->mutex
);
4704 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
) {
4705 struct perf_counter
*parent
= counter
->parent
;
4707 if (WARN_ON_ONCE(!parent
))
4710 mutex_lock(&parent
->child_mutex
);
4711 list_del_init(&counter
->child_list
);
4712 mutex_unlock(&parent
->child_mutex
);
4716 list_del_counter(counter
, ctx
);
4717 free_counter(counter
);
4720 if (!list_empty(&ctx
->counter_list
))
4723 mutex_unlock(&ctx
->mutex
);
4729 * Initialize the perf_counter context in task_struct
4731 int perf_counter_init_task(struct task_struct
*child
)
4733 struct perf_counter_context
*child_ctx
, *parent_ctx
;
4734 struct perf_counter_context
*cloned_ctx
;
4735 struct perf_counter
*counter
;
4736 struct task_struct
*parent
= current
;
4737 int inherited_all
= 1;
4740 child
->perf_counter_ctxp
= NULL
;
4742 mutex_init(&child
->perf_counter_mutex
);
4743 INIT_LIST_HEAD(&child
->perf_counter_list
);
4745 if (likely(!parent
->perf_counter_ctxp
))
4749 * This is executed from the parent task context, so inherit
4750 * counters that have been marked for cloning.
4751 * First allocate and initialize a context for the child.
4754 child_ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
4758 __perf_counter_init_context(child_ctx
, child
);
4759 child
->perf_counter_ctxp
= child_ctx
;
4760 get_task_struct(child
);
4763 * If the parent's context is a clone, pin it so it won't get
4766 parent_ctx
= perf_pin_task_context(parent
);
4769 * No need to check if parent_ctx != NULL here; since we saw
4770 * it non-NULL earlier, the only reason for it to become NULL
4771 * is if we exit, and since we're currently in the middle of
4772 * a fork we can't be exiting at the same time.
4776 * Lock the parent list. No need to lock the child - not PID
4777 * hashed yet and not running, so nobody can access it.
4779 mutex_lock(&parent_ctx
->mutex
);
4782 * We dont have to disable NMIs - we are only looking at
4783 * the list, not manipulating it:
4785 list_for_each_entry_rcu(counter
, &parent_ctx
->event_list
, event_entry
) {
4786 if (counter
!= counter
->group_leader
)
4789 if (!counter
->attr
.inherit
) {
4794 ret
= inherit_group(counter
, parent
, parent_ctx
,
4802 if (inherited_all
) {
4804 * Mark the child context as a clone of the parent
4805 * context, or of whatever the parent is a clone of.
4806 * Note that if the parent is a clone, it could get
4807 * uncloned at any point, but that doesn't matter
4808 * because the list of counters and the generation
4809 * count can't have changed since we took the mutex.
4811 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
4813 child_ctx
->parent_ctx
= cloned_ctx
;
4814 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
4816 child_ctx
->parent_ctx
= parent_ctx
;
4817 child_ctx
->parent_gen
= parent_ctx
->generation
;
4819 get_ctx(child_ctx
->parent_ctx
);
4822 mutex_unlock(&parent_ctx
->mutex
);
4824 perf_unpin_context(parent_ctx
);
4829 static void __cpuinit
perf_counter_init_cpu(int cpu
)
4831 struct perf_cpu_context
*cpuctx
;
4833 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4834 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
4836 spin_lock(&perf_resource_lock
);
4837 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
4838 spin_unlock(&perf_resource_lock
);
4840 hw_perf_counter_setup(cpu
);
4843 #ifdef CONFIG_HOTPLUG_CPU
4844 static void __perf_counter_exit_cpu(void *info
)
4846 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4847 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4848 struct perf_counter
*counter
, *tmp
;
4850 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
4851 __perf_counter_remove_from_context(counter
);
4853 static void perf_counter_exit_cpu(int cpu
)
4855 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4856 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4858 mutex_lock(&ctx
->mutex
);
4859 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
4860 mutex_unlock(&ctx
->mutex
);
4863 static inline void perf_counter_exit_cpu(int cpu
) { }
4866 static int __cpuinit
4867 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
4869 unsigned int cpu
= (long)hcpu
;
4873 case CPU_UP_PREPARE
:
4874 case CPU_UP_PREPARE_FROZEN
:
4875 perf_counter_init_cpu(cpu
);
4879 case CPU_ONLINE_FROZEN
:
4880 hw_perf_counter_setup_online(cpu
);
4883 case CPU_DOWN_PREPARE
:
4884 case CPU_DOWN_PREPARE_FROZEN
:
4885 perf_counter_exit_cpu(cpu
);
4896 * This has to have a higher priority than migration_notifier in sched.c.
4898 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
4899 .notifier_call
= perf_cpu_notify
,
4903 void __init
perf_counter_init(void)
4905 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
4906 (void *)(long)smp_processor_id());
4907 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_ONLINE
,
4908 (void *)(long)smp_processor_id());
4909 register_cpu_notifier(&perf_cpu_nb
);
4912 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
4914 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
4918 perf_set_reserve_percpu(struct sysdev_class
*class,
4922 struct perf_cpu_context
*cpuctx
;
4926 err
= strict_strtoul(buf
, 10, &val
);
4929 if (val
> perf_max_counters
)
4932 spin_lock(&perf_resource_lock
);
4933 perf_reserved_percpu
= val
;
4934 for_each_online_cpu(cpu
) {
4935 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4936 spin_lock_irq(&cpuctx
->ctx
.lock
);
4937 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
4938 perf_max_counters
- perf_reserved_percpu
);
4939 cpuctx
->max_pertask
= mpt
;
4940 spin_unlock_irq(&cpuctx
->ctx
.lock
);
4942 spin_unlock(&perf_resource_lock
);
4947 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
4949 return sprintf(buf
, "%d\n", perf_overcommit
);
4953 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
4958 err
= strict_strtoul(buf
, 10, &val
);
4964 spin_lock(&perf_resource_lock
);
4965 perf_overcommit
= val
;
4966 spin_unlock(&perf_resource_lock
);
4971 static SYSDEV_CLASS_ATTR(
4974 perf_show_reserve_percpu
,
4975 perf_set_reserve_percpu
4978 static SYSDEV_CLASS_ATTR(
4981 perf_show_overcommit
,
4985 static struct attribute
*perfclass_attrs
[] = {
4986 &attr_reserve_percpu
.attr
,
4987 &attr_overcommit
.attr
,
4991 static struct attribute_group perfclass_attr_group
= {
4992 .attrs
= perfclass_attrs
,
4993 .name
= "perf_counters",
4996 static int __init
perf_counter_sysfs_init(void)
4998 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
4999 &perfclass_attr_group
);
5001 device_initcall(perf_counter_sysfs_init
);