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
;
46 int sysctl_perf_counter_priv __read_mostly
; /* do we need to be privileged */
47 int sysctl_perf_counter_mlock __read_mostly
= 512; /* 'free' kb per user */
48 int sysctl_perf_counter_limit __read_mostly
= 100000; /* max NMIs per second */
50 static atomic64_t perf_counter_id
;
53 * Lock for (sysadmin-configurable) counter reservations:
55 static DEFINE_SPINLOCK(perf_resource_lock
);
58 * Architecture provided APIs - weak aliases:
60 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
65 void __weak
hw_perf_disable(void) { barrier(); }
66 void __weak
hw_perf_enable(void) { barrier(); }
68 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
71 hw_perf_group_sched_in(struct perf_counter
*group_leader
,
72 struct perf_cpu_context
*cpuctx
,
73 struct perf_counter_context
*ctx
, int cpu
)
78 void __weak
perf_counter_print_debug(void) { }
80 static DEFINE_PER_CPU(int, disable_count
);
82 void __perf_disable(void)
84 __get_cpu_var(disable_count
)++;
87 bool __perf_enable(void)
89 return !--__get_cpu_var(disable_count
);
92 void perf_disable(void)
98 void perf_enable(void)
104 static void get_ctx(struct perf_counter_context
*ctx
)
106 atomic_inc(&ctx
->refcount
);
109 static void free_ctx(struct rcu_head
*head
)
111 struct perf_counter_context
*ctx
;
113 ctx
= container_of(head
, struct perf_counter_context
, rcu_head
);
117 static void put_ctx(struct perf_counter_context
*ctx
)
119 if (atomic_dec_and_test(&ctx
->refcount
)) {
121 put_ctx(ctx
->parent_ctx
);
123 put_task_struct(ctx
->task
);
124 call_rcu(&ctx
->rcu_head
, free_ctx
);
129 * Get the perf_counter_context for a task and lock it.
130 * This has to cope with with the fact that until it is locked,
131 * the context could get moved to another task.
133 static struct perf_counter_context
*
134 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
136 struct perf_counter_context
*ctx
;
140 ctx
= rcu_dereference(task
->perf_counter_ctxp
);
143 * If this context is a clone of another, it might
144 * get swapped for another underneath us by
145 * perf_counter_task_sched_out, though the
146 * rcu_read_lock() protects us from any context
147 * getting freed. Lock the context and check if it
148 * got swapped before we could get the lock, and retry
149 * if so. If we locked the right context, then it
150 * can't get swapped on us any more.
152 spin_lock_irqsave(&ctx
->lock
, *flags
);
153 if (ctx
!= rcu_dereference(task
->perf_counter_ctxp
)) {
154 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
163 * Get the context for a task and increment its pin_count so it
164 * can't get swapped to another task. This also increments its
165 * reference count so that the context can't get freed.
167 static struct perf_counter_context
*perf_pin_task_context(struct task_struct
*task
)
169 struct perf_counter_context
*ctx
;
172 ctx
= perf_lock_task_context(task
, &flags
);
176 spin_unlock_irqrestore(&ctx
->lock
, flags
);
181 static void perf_unpin_context(struct perf_counter_context
*ctx
)
185 spin_lock_irqsave(&ctx
->lock
, flags
);
187 spin_unlock_irqrestore(&ctx
->lock
, flags
);
192 * Add a counter from the lists for its context.
193 * Must be called with ctx->mutex and ctx->lock held.
196 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
198 struct perf_counter
*group_leader
= counter
->group_leader
;
201 * Depending on whether it is a standalone or sibling counter,
202 * add it straight to the context's counter list, or to the group
203 * leader's sibling list:
205 if (group_leader
== counter
)
206 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
208 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
209 group_leader
->nr_siblings
++;
212 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
217 * Remove a counter from the lists for its context.
218 * Must be called with ctx->mutex and ctx->lock held.
221 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
223 struct perf_counter
*sibling
, *tmp
;
225 if (list_empty(&counter
->list_entry
))
229 list_del_init(&counter
->list_entry
);
230 list_del_rcu(&counter
->event_entry
);
232 if (counter
->group_leader
!= counter
)
233 counter
->group_leader
->nr_siblings
--;
236 * If this was a group counter with sibling counters then
237 * upgrade the siblings to singleton counters by adding them
238 * to the context list directly:
240 list_for_each_entry_safe(sibling
, tmp
,
241 &counter
->sibling_list
, list_entry
) {
243 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
244 sibling
->group_leader
= sibling
;
249 counter_sched_out(struct perf_counter
*counter
,
250 struct perf_cpu_context
*cpuctx
,
251 struct perf_counter_context
*ctx
)
253 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
256 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
257 counter
->tstamp_stopped
= ctx
->time
;
258 counter
->pmu
->disable(counter
);
261 if (!is_software_counter(counter
))
262 cpuctx
->active_oncpu
--;
264 if (counter
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
265 cpuctx
->exclusive
= 0;
269 group_sched_out(struct perf_counter
*group_counter
,
270 struct perf_cpu_context
*cpuctx
,
271 struct perf_counter_context
*ctx
)
273 struct perf_counter
*counter
;
275 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
278 counter_sched_out(group_counter
, cpuctx
, ctx
);
281 * Schedule out siblings (if any):
283 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
284 counter_sched_out(counter
, cpuctx
, ctx
);
286 if (group_counter
->attr
.exclusive
)
287 cpuctx
->exclusive
= 0;
291 * Cross CPU call to remove a performance counter
293 * We disable the counter on the hardware level first. After that we
294 * remove it from the context list.
296 static void __perf_counter_remove_from_context(void *info
)
298 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
299 struct perf_counter
*counter
= info
;
300 struct perf_counter_context
*ctx
= counter
->ctx
;
303 * If this is a task context, we need to check whether it is
304 * the current task context of this cpu. If not it has been
305 * scheduled out before the smp call arrived.
307 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
310 spin_lock(&ctx
->lock
);
312 * Protect the list operation against NMI by disabling the
313 * counters on a global level.
317 counter_sched_out(counter
, cpuctx
, ctx
);
319 list_del_counter(counter
, ctx
);
323 * Allow more per task counters with respect to the
326 cpuctx
->max_pertask
=
327 min(perf_max_counters
- ctx
->nr_counters
,
328 perf_max_counters
- perf_reserved_percpu
);
332 spin_unlock(&ctx
->lock
);
337 * Remove the counter from a task's (or a CPU's) list of counters.
339 * Must be called with ctx->mutex held.
341 * CPU counters are removed with a smp call. For task counters we only
342 * call when the task is on a CPU.
344 * If counter->ctx is a cloned context, callers must make sure that
345 * every task struct that counter->ctx->task could possibly point to
346 * remains valid. This is OK when called from perf_release since
347 * that only calls us on the top-level context, which can't be a clone.
348 * When called from perf_counter_exit_task, it's OK because the
349 * context has been detached from its task.
351 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
353 struct perf_counter_context
*ctx
= counter
->ctx
;
354 struct task_struct
*task
= ctx
->task
;
358 * Per cpu counters are removed via an smp call and
359 * the removal is always sucessful.
361 smp_call_function_single(counter
->cpu
,
362 __perf_counter_remove_from_context
,
368 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
371 spin_lock_irq(&ctx
->lock
);
373 * If the context is active we need to retry the smp call.
375 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
376 spin_unlock_irq(&ctx
->lock
);
381 * The lock prevents that this context is scheduled in so we
382 * can remove the counter safely, if the call above did not
385 if (!list_empty(&counter
->list_entry
)) {
386 list_del_counter(counter
, ctx
);
388 spin_unlock_irq(&ctx
->lock
);
391 static inline u64
perf_clock(void)
393 return cpu_clock(smp_processor_id());
397 * Update the record of the current time in a context.
399 static void update_context_time(struct perf_counter_context
*ctx
)
401 u64 now
= perf_clock();
403 ctx
->time
+= now
- ctx
->timestamp
;
404 ctx
->timestamp
= now
;
408 * Update the total_time_enabled and total_time_running fields for a counter.
410 static void update_counter_times(struct perf_counter
*counter
)
412 struct perf_counter_context
*ctx
= counter
->ctx
;
415 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
418 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
420 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
421 run_end
= counter
->tstamp_stopped
;
425 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
429 * Update total_time_enabled and total_time_running for all counters in a group.
431 static void update_group_times(struct perf_counter
*leader
)
433 struct perf_counter
*counter
;
435 update_counter_times(leader
);
436 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
437 update_counter_times(counter
);
441 * Cross CPU call to disable a performance counter
443 static void __perf_counter_disable(void *info
)
445 struct perf_counter
*counter
= info
;
446 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
447 struct perf_counter_context
*ctx
= counter
->ctx
;
450 * If this is a per-task counter, need to check whether this
451 * counter's task is the current task on this cpu.
453 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
456 spin_lock(&ctx
->lock
);
459 * If the counter is on, turn it off.
460 * If it is in error state, leave it in error state.
462 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
463 update_context_time(ctx
);
464 update_counter_times(counter
);
465 if (counter
== counter
->group_leader
)
466 group_sched_out(counter
, cpuctx
, ctx
);
468 counter_sched_out(counter
, cpuctx
, ctx
);
469 counter
->state
= PERF_COUNTER_STATE_OFF
;
472 spin_unlock(&ctx
->lock
);
478 * If counter->ctx is a cloned context, callers must make sure that
479 * every task struct that counter->ctx->task could possibly point to
480 * remains valid. This condition is satisifed when called through
481 * perf_counter_for_each_child or perf_counter_for_each because they
482 * hold the top-level counter's child_mutex, so any descendant that
483 * goes to exit will block in sync_child_counter.
484 * When called from perf_pending_counter it's OK because counter->ctx
485 * is the current context on this CPU and preemption is disabled,
486 * hence we can't get into perf_counter_task_sched_out for this context.
488 static void perf_counter_disable(struct perf_counter
*counter
)
490 struct perf_counter_context
*ctx
= counter
->ctx
;
491 struct task_struct
*task
= ctx
->task
;
495 * Disable the counter on the cpu that it's on
497 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
503 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
505 spin_lock_irq(&ctx
->lock
);
507 * If the counter is still active, we need to retry the cross-call.
509 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
510 spin_unlock_irq(&ctx
->lock
);
515 * Since we have the lock this context can't be scheduled
516 * in, so we can change the state safely.
518 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
519 update_counter_times(counter
);
520 counter
->state
= PERF_COUNTER_STATE_OFF
;
523 spin_unlock_irq(&ctx
->lock
);
527 counter_sched_in(struct perf_counter
*counter
,
528 struct perf_cpu_context
*cpuctx
,
529 struct perf_counter_context
*ctx
,
532 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
535 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
536 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
538 * The new state must be visible before we turn it on in the hardware:
542 if (counter
->pmu
->enable(counter
)) {
543 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
548 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
550 if (!is_software_counter(counter
))
551 cpuctx
->active_oncpu
++;
554 if (counter
->attr
.exclusive
)
555 cpuctx
->exclusive
= 1;
561 group_sched_in(struct perf_counter
*group_counter
,
562 struct perf_cpu_context
*cpuctx
,
563 struct perf_counter_context
*ctx
,
566 struct perf_counter
*counter
, *partial_group
;
569 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
572 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
574 return ret
< 0 ? ret
: 0;
576 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
580 * Schedule in siblings as one group (if any):
582 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
583 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
584 partial_group
= counter
;
593 * Groups can be scheduled in as one unit only, so undo any
594 * partial group before returning:
596 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
597 if (counter
== partial_group
)
599 counter_sched_out(counter
, cpuctx
, ctx
);
601 counter_sched_out(group_counter
, cpuctx
, ctx
);
607 * Return 1 for a group consisting entirely of software counters,
608 * 0 if the group contains any hardware counters.
610 static int is_software_only_group(struct perf_counter
*leader
)
612 struct perf_counter
*counter
;
614 if (!is_software_counter(leader
))
617 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
618 if (!is_software_counter(counter
))
625 * Work out whether we can put this counter group on the CPU now.
627 static int group_can_go_on(struct perf_counter
*counter
,
628 struct perf_cpu_context
*cpuctx
,
632 * Groups consisting entirely of software counters can always go on.
634 if (is_software_only_group(counter
))
637 * If an exclusive group is already on, no other hardware
638 * counters can go on.
640 if (cpuctx
->exclusive
)
643 * If this group is exclusive and there are already
644 * counters on the CPU, it can't go on.
646 if (counter
->attr
.exclusive
&& cpuctx
->active_oncpu
)
649 * Otherwise, try to add it if all previous groups were able
655 static void add_counter_to_ctx(struct perf_counter
*counter
,
656 struct perf_counter_context
*ctx
)
658 list_add_counter(counter
, ctx
);
659 counter
->tstamp_enabled
= ctx
->time
;
660 counter
->tstamp_running
= ctx
->time
;
661 counter
->tstamp_stopped
= ctx
->time
;
665 * Cross CPU call to install and enable a performance counter
667 * Must be called with ctx->mutex held
669 static void __perf_install_in_context(void *info
)
671 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
672 struct perf_counter
*counter
= info
;
673 struct perf_counter_context
*ctx
= counter
->ctx
;
674 struct perf_counter
*leader
= counter
->group_leader
;
675 int cpu
= smp_processor_id();
679 * If this is a task context, we need to check whether it is
680 * the current task context of this cpu. If not it has been
681 * scheduled out before the smp call arrived.
682 * Or possibly this is the right context but it isn't
683 * on this cpu because it had no counters.
685 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
686 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
688 cpuctx
->task_ctx
= ctx
;
691 spin_lock(&ctx
->lock
);
693 update_context_time(ctx
);
696 * Protect the list operation against NMI by disabling the
697 * counters on a global level. NOP for non NMI based counters.
701 add_counter_to_ctx(counter
, ctx
);
704 * Don't put the counter on if it is disabled or if
705 * it is in a group and the group isn't on.
707 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
708 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
712 * An exclusive counter can't go on if there are already active
713 * hardware counters, and no hardware counter can go on if there
714 * is already an exclusive counter on.
716 if (!group_can_go_on(counter
, cpuctx
, 1))
719 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
723 * This counter couldn't go on. If it is in a group
724 * then we have to pull the whole group off.
725 * If the counter group is pinned then put it in error state.
727 if (leader
!= counter
)
728 group_sched_out(leader
, cpuctx
, ctx
);
729 if (leader
->attr
.pinned
) {
730 update_group_times(leader
);
731 leader
->state
= PERF_COUNTER_STATE_ERROR
;
735 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
736 cpuctx
->max_pertask
--;
741 spin_unlock(&ctx
->lock
);
745 * Attach a performance counter to a context
747 * First we add the counter to the list with the hardware enable bit
748 * in counter->hw_config cleared.
750 * If the counter is attached to a task which is on a CPU we use a smp
751 * call to enable it in the task context. The task might have been
752 * scheduled away, but we check this in the smp call again.
754 * Must be called with ctx->mutex held.
757 perf_install_in_context(struct perf_counter_context
*ctx
,
758 struct perf_counter
*counter
,
761 struct task_struct
*task
= ctx
->task
;
765 * Per cpu counters are installed via an smp call and
766 * the install is always sucessful.
768 smp_call_function_single(cpu
, __perf_install_in_context
,
774 task_oncpu_function_call(task
, __perf_install_in_context
,
777 spin_lock_irq(&ctx
->lock
);
779 * we need to retry the smp call.
781 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
782 spin_unlock_irq(&ctx
->lock
);
787 * The lock prevents that this context is scheduled in so we
788 * can add the counter safely, if it the call above did not
791 if (list_empty(&counter
->list_entry
))
792 add_counter_to_ctx(counter
, ctx
);
793 spin_unlock_irq(&ctx
->lock
);
797 * Cross CPU call to enable a performance counter
799 static void __perf_counter_enable(void *info
)
801 struct perf_counter
*counter
= info
;
802 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
803 struct perf_counter_context
*ctx
= counter
->ctx
;
804 struct perf_counter
*leader
= counter
->group_leader
;
808 * If this is a per-task counter, need to check whether this
809 * counter's task is the current task on this cpu.
811 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
812 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
814 cpuctx
->task_ctx
= ctx
;
817 spin_lock(&ctx
->lock
);
819 update_context_time(ctx
);
821 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
823 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
824 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
827 * If the counter is in a group and isn't the group leader,
828 * then don't put it on unless the group is on.
830 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
833 if (!group_can_go_on(counter
, cpuctx
, 1)) {
837 if (counter
== leader
)
838 err
= group_sched_in(counter
, cpuctx
, ctx
,
841 err
= counter_sched_in(counter
, cpuctx
, ctx
,
848 * If this counter can't go on and it's part of a
849 * group, then the whole group has to come off.
851 if (leader
!= counter
)
852 group_sched_out(leader
, cpuctx
, ctx
);
853 if (leader
->attr
.pinned
) {
854 update_group_times(leader
);
855 leader
->state
= PERF_COUNTER_STATE_ERROR
;
860 spin_unlock(&ctx
->lock
);
866 * If counter->ctx is a cloned context, callers must make sure that
867 * every task struct that counter->ctx->task could possibly point to
868 * remains valid. This condition is satisfied when called through
869 * perf_counter_for_each_child or perf_counter_for_each as described
870 * for perf_counter_disable.
872 static void perf_counter_enable(struct perf_counter
*counter
)
874 struct perf_counter_context
*ctx
= counter
->ctx
;
875 struct task_struct
*task
= ctx
->task
;
879 * Enable the counter on the cpu that it's on
881 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
886 spin_lock_irq(&ctx
->lock
);
887 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
891 * If the counter is in error state, clear that first.
892 * That way, if we see the counter in error state below, we
893 * know that it has gone back into error state, as distinct
894 * from the task having been scheduled away before the
895 * cross-call arrived.
897 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
898 counter
->state
= PERF_COUNTER_STATE_OFF
;
901 spin_unlock_irq(&ctx
->lock
);
902 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
904 spin_lock_irq(&ctx
->lock
);
907 * If the context is active and the counter is still off,
908 * we need to retry the cross-call.
910 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
914 * Since we have the lock this context can't be scheduled
915 * in, so we can change the state safely.
917 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
918 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
919 counter
->tstamp_enabled
=
920 ctx
->time
- counter
->total_time_enabled
;
923 spin_unlock_irq(&ctx
->lock
);
926 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
929 * not supported on inherited counters
931 if (counter
->attr
.inherit
)
934 atomic_add(refresh
, &counter
->event_limit
);
935 perf_counter_enable(counter
);
940 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
941 struct perf_cpu_context
*cpuctx
)
943 struct perf_counter
*counter
;
945 spin_lock(&ctx
->lock
);
947 if (likely(!ctx
->nr_counters
))
949 update_context_time(ctx
);
952 if (ctx
->nr_active
) {
953 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
954 if (counter
!= counter
->group_leader
)
955 counter_sched_out(counter
, cpuctx
, ctx
);
957 group_sched_out(counter
, cpuctx
, ctx
);
962 spin_unlock(&ctx
->lock
);
966 * Test whether two contexts are equivalent, i.e. whether they
967 * have both been cloned from the same version of the same context
968 * and they both have the same number of enabled counters.
969 * If the number of enabled counters is the same, then the set
970 * of enabled counters should be the same, because these are both
971 * inherited contexts, therefore we can't access individual counters
972 * in them directly with an fd; we can only enable/disable all
973 * counters via prctl, or enable/disable all counters in a family
974 * via ioctl, which will have the same effect on both contexts.
976 static int context_equiv(struct perf_counter_context
*ctx1
,
977 struct perf_counter_context
*ctx2
)
979 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
980 && ctx1
->parent_gen
== ctx2
->parent_gen
981 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
985 * Called from scheduler to remove the counters of the current task,
986 * with interrupts disabled.
988 * We stop each counter and update the counter value in counter->count.
990 * This does not protect us against NMI, but disable()
991 * sets the disabled bit in the control field of counter _before_
992 * accessing the counter control register. If a NMI hits, then it will
993 * not restart the counter.
995 void perf_counter_task_sched_out(struct task_struct
*task
,
996 struct task_struct
*next
, int cpu
)
998 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
999 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1000 struct perf_counter_context
*next_ctx
;
1001 struct perf_counter_context
*parent
;
1002 struct pt_regs
*regs
;
1005 regs
= task_pt_regs(task
);
1006 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1008 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1011 update_context_time(ctx
);
1014 parent
= rcu_dereference(ctx
->parent_ctx
);
1015 next_ctx
= next
->perf_counter_ctxp
;
1016 if (parent
&& next_ctx
&&
1017 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1019 * Looks like the two contexts are clones, so we might be
1020 * able to optimize the context switch. We lock both
1021 * contexts and check that they are clones under the
1022 * lock (including re-checking that neither has been
1023 * uncloned in the meantime). It doesn't matter which
1024 * order we take the locks because no other cpu could
1025 * be trying to lock both of these tasks.
1027 spin_lock(&ctx
->lock
);
1028 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1029 if (context_equiv(ctx
, next_ctx
)) {
1031 * XXX do we need a memory barrier of sorts
1032 * wrt to rcu_dereference() of perf_counter_ctxp
1034 task
->perf_counter_ctxp
= next_ctx
;
1035 next
->perf_counter_ctxp
= ctx
;
1037 next_ctx
->task
= task
;
1040 spin_unlock(&next_ctx
->lock
);
1041 spin_unlock(&ctx
->lock
);
1046 __perf_counter_sched_out(ctx
, cpuctx
);
1047 cpuctx
->task_ctx
= NULL
;
1052 * Called with IRQs disabled
1054 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
1056 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1058 if (!cpuctx
->task_ctx
)
1061 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1064 __perf_counter_sched_out(ctx
, cpuctx
);
1065 cpuctx
->task_ctx
= NULL
;
1069 * Called with IRQs disabled
1071 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1073 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
1077 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
1078 struct perf_cpu_context
*cpuctx
, int cpu
)
1080 struct perf_counter
*counter
;
1083 spin_lock(&ctx
->lock
);
1085 if (likely(!ctx
->nr_counters
))
1088 ctx
->timestamp
= perf_clock();
1093 * First go through the list and put on any pinned groups
1094 * in order to give them the best chance of going on.
1096 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1097 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1098 !counter
->attr
.pinned
)
1100 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1103 if (counter
!= counter
->group_leader
)
1104 counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
1106 if (group_can_go_on(counter
, cpuctx
, 1))
1107 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
1111 * If this pinned group hasn't been scheduled,
1112 * put it in error state.
1114 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1115 update_group_times(counter
);
1116 counter
->state
= PERF_COUNTER_STATE_ERROR
;
1120 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1122 * Ignore counters in OFF or ERROR state, and
1123 * ignore pinned counters since we did them already.
1125 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1126 counter
->attr
.pinned
)
1130 * Listen to the 'cpu' scheduling filter constraint
1133 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1136 if (counter
!= counter
->group_leader
) {
1137 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
))
1140 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
1141 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
1148 spin_unlock(&ctx
->lock
);
1152 * Called from scheduler to add the counters of the current task
1153 * with interrupts disabled.
1155 * We restore the counter value and then enable it.
1157 * This does not protect us against NMI, but enable()
1158 * sets the enabled bit in the control field of counter _before_
1159 * accessing the counter control register. If a NMI hits, then it will
1160 * keep the counter running.
1162 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
1164 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1165 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1169 if (cpuctx
->task_ctx
== ctx
)
1171 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1172 cpuctx
->task_ctx
= ctx
;
1175 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1177 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
1179 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1182 #define MAX_INTERRUPTS (~0ULL)
1184 static void perf_log_throttle(struct perf_counter
*counter
, int enable
);
1185 static void perf_log_period(struct perf_counter
*counter
, u64 period
);
1187 static void perf_adjust_freq(struct perf_counter_context
*ctx
)
1189 struct perf_counter
*counter
;
1190 u64 interrupts
, sample_period
;
1194 spin_lock(&ctx
->lock
);
1195 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1196 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1199 interrupts
= counter
->hw
.interrupts
;
1200 counter
->hw
.interrupts
= 0;
1202 if (interrupts
== MAX_INTERRUPTS
) {
1203 perf_log_throttle(counter
, 1);
1204 counter
->pmu
->unthrottle(counter
);
1205 interrupts
= 2*sysctl_perf_counter_limit
/HZ
;
1208 if (!counter
->attr
.freq
|| !counter
->attr
.sample_freq
)
1211 events
= HZ
* interrupts
* counter
->hw
.sample_period
;
1212 period
= div64_u64(events
, counter
->attr
.sample_freq
);
1214 delta
= (s64
)(1 + period
- counter
->hw
.sample_period
);
1217 sample_period
= counter
->hw
.sample_period
+ delta
;
1222 perf_log_period(counter
, sample_period
);
1224 counter
->hw
.sample_period
= sample_period
;
1226 spin_unlock(&ctx
->lock
);
1230 * Round-robin a context's counters:
1232 static void rotate_ctx(struct perf_counter_context
*ctx
)
1234 struct perf_counter
*counter
;
1236 if (!ctx
->nr_counters
)
1239 spin_lock(&ctx
->lock
);
1241 * Rotate the first entry last (works just fine for group counters too):
1244 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1245 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1250 spin_unlock(&ctx
->lock
);
1253 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1255 struct perf_cpu_context
*cpuctx
;
1256 struct perf_counter_context
*ctx
;
1258 if (!atomic_read(&nr_counters
))
1261 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1262 ctx
= curr
->perf_counter_ctxp
;
1264 perf_adjust_freq(&cpuctx
->ctx
);
1266 perf_adjust_freq(ctx
);
1268 perf_counter_cpu_sched_out(cpuctx
);
1270 __perf_counter_task_sched_out(ctx
);
1272 rotate_ctx(&cpuctx
->ctx
);
1276 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1278 perf_counter_task_sched_in(curr
, cpu
);
1282 * Cross CPU call to read the hardware counter
1284 static void __read(void *info
)
1286 struct perf_counter
*counter
= info
;
1287 struct perf_counter_context
*ctx
= counter
->ctx
;
1288 unsigned long flags
;
1290 local_irq_save(flags
);
1292 update_context_time(ctx
);
1293 counter
->pmu
->read(counter
);
1294 update_counter_times(counter
);
1295 local_irq_restore(flags
);
1298 static u64
perf_counter_read(struct perf_counter
*counter
)
1301 * If counter is enabled and currently active on a CPU, update the
1302 * value in the counter structure:
1304 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1305 smp_call_function_single(counter
->oncpu
,
1306 __read
, counter
, 1);
1307 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1308 update_counter_times(counter
);
1311 return atomic64_read(&counter
->count
);
1315 * Initialize the perf_counter context in a task_struct:
1318 __perf_counter_init_context(struct perf_counter_context
*ctx
,
1319 struct task_struct
*task
)
1321 memset(ctx
, 0, sizeof(*ctx
));
1322 spin_lock_init(&ctx
->lock
);
1323 mutex_init(&ctx
->mutex
);
1324 INIT_LIST_HEAD(&ctx
->counter_list
);
1325 INIT_LIST_HEAD(&ctx
->event_list
);
1326 atomic_set(&ctx
->refcount
, 1);
1330 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1332 struct perf_counter_context
*parent_ctx
;
1333 struct perf_counter_context
*ctx
;
1334 struct perf_cpu_context
*cpuctx
;
1335 struct task_struct
*task
;
1336 unsigned long flags
;
1340 * If cpu is not a wildcard then this is a percpu counter:
1343 /* Must be root to operate on a CPU counter: */
1344 if (sysctl_perf_counter_priv
&& !capable(CAP_SYS_ADMIN
))
1345 return ERR_PTR(-EACCES
);
1347 if (cpu
< 0 || cpu
> num_possible_cpus())
1348 return ERR_PTR(-EINVAL
);
1351 * We could be clever and allow to attach a counter to an
1352 * offline CPU and activate it when the CPU comes up, but
1355 if (!cpu_isset(cpu
, cpu_online_map
))
1356 return ERR_PTR(-ENODEV
);
1358 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1369 task
= find_task_by_vpid(pid
);
1371 get_task_struct(task
);
1375 return ERR_PTR(-ESRCH
);
1378 * Can't attach counters to a dying task.
1381 if (task
->flags
& PF_EXITING
)
1384 /* Reuse ptrace permission checks for now. */
1386 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1390 ctx
= perf_lock_task_context(task
, &flags
);
1392 parent_ctx
= ctx
->parent_ctx
;
1394 put_ctx(parent_ctx
);
1395 ctx
->parent_ctx
= NULL
; /* no longer a clone */
1398 * Get an extra reference before dropping the lock so that
1399 * this context won't get freed if the task exits.
1402 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1406 ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
1410 __perf_counter_init_context(ctx
, task
);
1412 if (cmpxchg(&task
->perf_counter_ctxp
, NULL
, ctx
)) {
1414 * We raced with some other task; use
1415 * the context they set.
1420 get_task_struct(task
);
1423 put_task_struct(task
);
1427 put_task_struct(task
);
1428 return ERR_PTR(err
);
1431 static void free_counter_rcu(struct rcu_head
*head
)
1433 struct perf_counter
*counter
;
1435 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1437 put_pid_ns(counter
->ns
);
1441 static void perf_pending_sync(struct perf_counter
*counter
);
1443 static void free_counter(struct perf_counter
*counter
)
1445 perf_pending_sync(counter
);
1447 atomic_dec(&nr_counters
);
1448 if (counter
->attr
.mmap
)
1449 atomic_dec(&nr_mmap_counters
);
1450 if (counter
->attr
.comm
)
1451 atomic_dec(&nr_comm_counters
);
1453 if (counter
->destroy
)
1454 counter
->destroy(counter
);
1456 put_ctx(counter
->ctx
);
1457 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1461 * Called when the last reference to the file is gone.
1463 static int perf_release(struct inode
*inode
, struct file
*file
)
1465 struct perf_counter
*counter
= file
->private_data
;
1466 struct perf_counter_context
*ctx
= counter
->ctx
;
1468 file
->private_data
= NULL
;
1470 WARN_ON_ONCE(ctx
->parent_ctx
);
1471 mutex_lock(&ctx
->mutex
);
1472 perf_counter_remove_from_context(counter
);
1473 mutex_unlock(&ctx
->mutex
);
1475 mutex_lock(&counter
->owner
->perf_counter_mutex
);
1476 list_del_init(&counter
->owner_entry
);
1477 mutex_unlock(&counter
->owner
->perf_counter_mutex
);
1478 put_task_struct(counter
->owner
);
1480 free_counter(counter
);
1486 * Read the performance counter - simple non blocking version for now
1489 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1495 * Return end-of-file for a read on a counter that is in
1496 * error state (i.e. because it was pinned but it couldn't be
1497 * scheduled on to the CPU at some point).
1499 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1502 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1503 mutex_lock(&counter
->child_mutex
);
1504 values
[0] = perf_counter_read(counter
);
1506 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1507 values
[n
++] = counter
->total_time_enabled
+
1508 atomic64_read(&counter
->child_total_time_enabled
);
1509 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1510 values
[n
++] = counter
->total_time_running
+
1511 atomic64_read(&counter
->child_total_time_running
);
1512 if (counter
->attr
.read_format
& PERF_FORMAT_ID
)
1513 values
[n
++] = counter
->id
;
1514 mutex_unlock(&counter
->child_mutex
);
1516 if (count
< n
* sizeof(u64
))
1518 count
= n
* sizeof(u64
);
1520 if (copy_to_user(buf
, values
, count
))
1527 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1529 struct perf_counter
*counter
= file
->private_data
;
1531 return perf_read_hw(counter
, buf
, count
);
1534 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1536 struct perf_counter
*counter
= file
->private_data
;
1537 struct perf_mmap_data
*data
;
1538 unsigned int events
= POLL_HUP
;
1541 data
= rcu_dereference(counter
->data
);
1543 events
= atomic_xchg(&data
->poll
, 0);
1546 poll_wait(file
, &counter
->waitq
, wait
);
1551 static void perf_counter_reset(struct perf_counter
*counter
)
1553 (void)perf_counter_read(counter
);
1554 atomic64_set(&counter
->count
, 0);
1555 perf_counter_update_userpage(counter
);
1558 static void perf_counter_for_each_sibling(struct perf_counter
*counter
,
1559 void (*func
)(struct perf_counter
*))
1561 struct perf_counter_context
*ctx
= counter
->ctx
;
1562 struct perf_counter
*sibling
;
1564 WARN_ON_ONCE(ctx
->parent_ctx
);
1565 mutex_lock(&ctx
->mutex
);
1566 counter
= counter
->group_leader
;
1569 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1571 mutex_unlock(&ctx
->mutex
);
1575 * Holding the top-level counter's child_mutex means that any
1576 * descendant process that has inherited this counter will block
1577 * in sync_child_counter if it goes to exit, thus satisfying the
1578 * task existence requirements of perf_counter_enable/disable.
1580 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1581 void (*func
)(struct perf_counter
*))
1583 struct perf_counter
*child
;
1585 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1586 mutex_lock(&counter
->child_mutex
);
1588 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1590 mutex_unlock(&counter
->child_mutex
);
1593 static void perf_counter_for_each(struct perf_counter
*counter
,
1594 void (*func
)(struct perf_counter
*))
1596 struct perf_counter
*child
;
1598 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1599 mutex_lock(&counter
->child_mutex
);
1600 perf_counter_for_each_sibling(counter
, func
);
1601 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1602 perf_counter_for_each_sibling(child
, func
);
1603 mutex_unlock(&counter
->child_mutex
);
1606 static int perf_counter_period(struct perf_counter
*counter
, u64 __user
*arg
)
1608 struct perf_counter_context
*ctx
= counter
->ctx
;
1613 if (!counter
->attr
.sample_period
)
1616 size
= copy_from_user(&value
, arg
, sizeof(value
));
1617 if (size
!= sizeof(value
))
1623 spin_lock_irq(&ctx
->lock
);
1624 if (counter
->attr
.freq
) {
1625 if (value
> sysctl_perf_counter_limit
) {
1630 counter
->attr
.sample_freq
= value
;
1632 counter
->attr
.sample_period
= value
;
1633 counter
->hw
.sample_period
= value
;
1635 perf_log_period(counter
, value
);
1638 spin_unlock_irq(&ctx
->lock
);
1643 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1645 struct perf_counter
*counter
= file
->private_data
;
1646 void (*func
)(struct perf_counter
*);
1650 case PERF_COUNTER_IOC_ENABLE
:
1651 func
= perf_counter_enable
;
1653 case PERF_COUNTER_IOC_DISABLE
:
1654 func
= perf_counter_disable
;
1656 case PERF_COUNTER_IOC_RESET
:
1657 func
= perf_counter_reset
;
1660 case PERF_COUNTER_IOC_REFRESH
:
1661 return perf_counter_refresh(counter
, arg
);
1663 case PERF_COUNTER_IOC_PERIOD
:
1664 return perf_counter_period(counter
, (u64 __user
*)arg
);
1670 if (flags
& PERF_IOC_FLAG_GROUP
)
1671 perf_counter_for_each(counter
, func
);
1673 perf_counter_for_each_child(counter
, func
);
1678 int perf_counter_task_enable(void)
1680 struct perf_counter
*counter
;
1682 mutex_lock(¤t
->perf_counter_mutex
);
1683 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1684 perf_counter_for_each_child(counter
, perf_counter_enable
);
1685 mutex_unlock(¤t
->perf_counter_mutex
);
1690 int perf_counter_task_disable(void)
1692 struct perf_counter
*counter
;
1694 mutex_lock(¤t
->perf_counter_mutex
);
1695 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1696 perf_counter_for_each_child(counter
, perf_counter_disable
);
1697 mutex_unlock(¤t
->perf_counter_mutex
);
1703 * Callers need to ensure there can be no nesting of this function, otherwise
1704 * the seqlock logic goes bad. We can not serialize this because the arch
1705 * code calls this from NMI context.
1707 void perf_counter_update_userpage(struct perf_counter
*counter
)
1709 struct perf_counter_mmap_page
*userpg
;
1710 struct perf_mmap_data
*data
;
1713 data
= rcu_dereference(counter
->data
);
1717 userpg
= data
->user_page
;
1720 * Disable preemption so as to not let the corresponding user-space
1721 * spin too long if we get preempted.
1726 userpg
->index
= counter
->hw
.idx
;
1727 userpg
->offset
= atomic64_read(&counter
->count
);
1728 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1729 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1738 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1740 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1741 struct perf_mmap_data
*data
;
1742 int ret
= VM_FAULT_SIGBUS
;
1745 data
= rcu_dereference(counter
->data
);
1749 if (vmf
->pgoff
== 0) {
1750 vmf
->page
= virt_to_page(data
->user_page
);
1752 int nr
= vmf
->pgoff
- 1;
1754 if ((unsigned)nr
> data
->nr_pages
)
1757 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1759 get_page(vmf
->page
);
1767 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1769 struct perf_mmap_data
*data
;
1773 WARN_ON(atomic_read(&counter
->mmap_count
));
1775 size
= sizeof(struct perf_mmap_data
);
1776 size
+= nr_pages
* sizeof(void *);
1778 data
= kzalloc(size
, GFP_KERNEL
);
1782 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1783 if (!data
->user_page
)
1784 goto fail_user_page
;
1786 for (i
= 0; i
< nr_pages
; i
++) {
1787 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1788 if (!data
->data_pages
[i
])
1789 goto fail_data_pages
;
1792 data
->nr_pages
= nr_pages
;
1793 atomic_set(&data
->lock
, -1);
1795 rcu_assign_pointer(counter
->data
, data
);
1800 for (i
--; i
>= 0; i
--)
1801 free_page((unsigned long)data
->data_pages
[i
]);
1803 free_page((unsigned long)data
->user_page
);
1812 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1814 struct perf_mmap_data
*data
;
1817 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
1819 free_page((unsigned long)data
->user_page
);
1820 for (i
= 0; i
< data
->nr_pages
; i
++)
1821 free_page((unsigned long)data
->data_pages
[i
]);
1825 static void perf_mmap_data_free(struct perf_counter
*counter
)
1827 struct perf_mmap_data
*data
= counter
->data
;
1829 WARN_ON(atomic_read(&counter
->mmap_count
));
1831 rcu_assign_pointer(counter
->data
, NULL
);
1832 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1835 static void perf_mmap_open(struct vm_area_struct
*vma
)
1837 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1839 atomic_inc(&counter
->mmap_count
);
1842 static void perf_mmap_close(struct vm_area_struct
*vma
)
1844 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1846 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1847 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
, &counter
->mmap_mutex
)) {
1848 struct user_struct
*user
= current_user();
1850 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
1851 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
1852 perf_mmap_data_free(counter
);
1853 mutex_unlock(&counter
->mmap_mutex
);
1857 static struct vm_operations_struct perf_mmap_vmops
= {
1858 .open
= perf_mmap_open
,
1859 .close
= perf_mmap_close
,
1860 .fault
= perf_mmap_fault
,
1863 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1865 struct perf_counter
*counter
= file
->private_data
;
1866 unsigned long user_locked
, user_lock_limit
;
1867 struct user_struct
*user
= current_user();
1868 unsigned long locked
, lock_limit
;
1869 unsigned long vma_size
;
1870 unsigned long nr_pages
;
1871 long user_extra
, extra
;
1874 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1877 vma_size
= vma
->vm_end
- vma
->vm_start
;
1878 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1881 * If we have data pages ensure they're a power-of-two number, so we
1882 * can do bitmasks instead of modulo.
1884 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1887 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1890 if (vma
->vm_pgoff
!= 0)
1893 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1894 mutex_lock(&counter
->mmap_mutex
);
1895 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1896 if (nr_pages
!= counter
->data
->nr_pages
)
1901 user_extra
= nr_pages
+ 1;
1902 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
1905 * Increase the limit linearly with more CPUs:
1907 user_lock_limit
*= num_online_cpus();
1909 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
1912 if (user_locked
> user_lock_limit
)
1913 extra
= user_locked
- user_lock_limit
;
1915 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1916 lock_limit
>>= PAGE_SHIFT
;
1917 locked
= vma
->vm_mm
->locked_vm
+ extra
;
1919 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
1924 WARN_ON(counter
->data
);
1925 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1929 atomic_set(&counter
->mmap_count
, 1);
1930 atomic_long_add(user_extra
, &user
->locked_vm
);
1931 vma
->vm_mm
->locked_vm
+= extra
;
1932 counter
->data
->nr_locked
= extra
;
1934 mutex_unlock(&counter
->mmap_mutex
);
1936 vma
->vm_flags
&= ~VM_MAYWRITE
;
1937 vma
->vm_flags
|= VM_RESERVED
;
1938 vma
->vm_ops
= &perf_mmap_vmops
;
1943 static int perf_fasync(int fd
, struct file
*filp
, int on
)
1945 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1946 struct perf_counter
*counter
= filp
->private_data
;
1949 mutex_lock(&inode
->i_mutex
);
1950 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
1951 mutex_unlock(&inode
->i_mutex
);
1959 static const struct file_operations perf_fops
= {
1960 .release
= perf_release
,
1963 .unlocked_ioctl
= perf_ioctl
,
1964 .compat_ioctl
= perf_ioctl
,
1966 .fasync
= perf_fasync
,
1970 * Perf counter wakeup
1972 * If there's data, ensure we set the poll() state and publish everything
1973 * to user-space before waking everybody up.
1976 void perf_counter_wakeup(struct perf_counter
*counter
)
1978 wake_up_all(&counter
->waitq
);
1980 if (counter
->pending_kill
) {
1981 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
1982 counter
->pending_kill
= 0;
1989 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1991 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1992 * single linked list and use cmpxchg() to add entries lockless.
1995 static void perf_pending_counter(struct perf_pending_entry
*entry
)
1997 struct perf_counter
*counter
= container_of(entry
,
1998 struct perf_counter
, pending
);
2000 if (counter
->pending_disable
) {
2001 counter
->pending_disable
= 0;
2002 perf_counter_disable(counter
);
2005 if (counter
->pending_wakeup
) {
2006 counter
->pending_wakeup
= 0;
2007 perf_counter_wakeup(counter
);
2011 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2013 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2017 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2018 void (*func
)(struct perf_pending_entry
*))
2020 struct perf_pending_entry
**head
;
2022 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2027 head
= &get_cpu_var(perf_pending_head
);
2030 entry
->next
= *head
;
2031 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2033 set_perf_counter_pending();
2035 put_cpu_var(perf_pending_head
);
2038 static int __perf_pending_run(void)
2040 struct perf_pending_entry
*list
;
2043 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2044 while (list
!= PENDING_TAIL
) {
2045 void (*func
)(struct perf_pending_entry
*);
2046 struct perf_pending_entry
*entry
= list
;
2053 * Ensure we observe the unqueue before we issue the wakeup,
2054 * so that we won't be waiting forever.
2055 * -- see perf_not_pending().
2066 static inline int perf_not_pending(struct perf_counter
*counter
)
2069 * If we flush on whatever cpu we run, there is a chance we don't
2073 __perf_pending_run();
2077 * Ensure we see the proper queue state before going to sleep
2078 * so that we do not miss the wakeup. -- see perf_pending_handle()
2081 return counter
->pending
.next
== NULL
;
2084 static void perf_pending_sync(struct perf_counter
*counter
)
2086 wait_event(counter
->waitq
, perf_not_pending(counter
));
2089 void perf_counter_do_pending(void)
2091 __perf_pending_run();
2095 * Callchain support -- arch specific
2098 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2107 struct perf_output_handle
{
2108 struct perf_counter
*counter
;
2109 struct perf_mmap_data
*data
;
2111 unsigned long offset
;
2115 unsigned long flags
;
2118 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2120 atomic_set(&handle
->data
->poll
, POLL_IN
);
2123 handle
->counter
->pending_wakeup
= 1;
2124 perf_pending_queue(&handle
->counter
->pending
,
2125 perf_pending_counter
);
2127 perf_counter_wakeup(handle
->counter
);
2131 * Curious locking construct.
2133 * We need to ensure a later event doesn't publish a head when a former
2134 * event isn't done writing. However since we need to deal with NMIs we
2135 * cannot fully serialize things.
2137 * What we do is serialize between CPUs so we only have to deal with NMI
2138 * nesting on a single CPU.
2140 * We only publish the head (and generate a wakeup) when the outer-most
2143 static void perf_output_lock(struct perf_output_handle
*handle
)
2145 struct perf_mmap_data
*data
= handle
->data
;
2150 local_irq_save(handle
->flags
);
2151 cpu
= smp_processor_id();
2153 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
2156 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2162 static void perf_output_unlock(struct perf_output_handle
*handle
)
2164 struct perf_mmap_data
*data
= handle
->data
;
2168 data
->done_head
= data
->head
;
2170 if (!handle
->locked
)
2175 * The xchg implies a full barrier that ensures all writes are done
2176 * before we publish the new head, matched by a rmb() in userspace when
2177 * reading this position.
2179 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2180 data
->user_page
->data_head
= head
;
2183 * NMI can happen here, which means we can miss a done_head update.
2186 cpu
= atomic_xchg(&data
->lock
, -1);
2187 WARN_ON_ONCE(cpu
!= smp_processor_id());
2190 * Therefore we have to validate we did not indeed do so.
2192 if (unlikely(atomic_long_read(&data
->done_head
))) {
2194 * Since we had it locked, we can lock it again.
2196 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2202 if (atomic_xchg(&data
->wakeup
, 0))
2203 perf_output_wakeup(handle
);
2205 local_irq_restore(handle
->flags
);
2208 static int perf_output_begin(struct perf_output_handle
*handle
,
2209 struct perf_counter
*counter
, unsigned int size
,
2210 int nmi
, int overflow
)
2212 struct perf_mmap_data
*data
;
2213 unsigned int offset
, head
;
2216 * For inherited counters we send all the output towards the parent.
2218 if (counter
->parent
)
2219 counter
= counter
->parent
;
2222 data
= rcu_dereference(counter
->data
);
2226 handle
->data
= data
;
2227 handle
->counter
= counter
;
2229 handle
->overflow
= overflow
;
2231 if (!data
->nr_pages
)
2234 perf_output_lock(handle
);
2237 offset
= head
= atomic_long_read(&data
->head
);
2239 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2241 handle
->offset
= offset
;
2242 handle
->head
= head
;
2244 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
2245 atomic_set(&data
->wakeup
, 1);
2250 perf_output_wakeup(handle
);
2257 static void perf_output_copy(struct perf_output_handle
*handle
,
2258 const void *buf
, unsigned int len
)
2260 unsigned int pages_mask
;
2261 unsigned int offset
;
2265 offset
= handle
->offset
;
2266 pages_mask
= handle
->data
->nr_pages
- 1;
2267 pages
= handle
->data
->data_pages
;
2270 unsigned int page_offset
;
2273 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2274 page_offset
= offset
& (PAGE_SIZE
- 1);
2275 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
2277 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2284 handle
->offset
= offset
;
2287 * Check we didn't copy past our reservation window, taking the
2288 * possible unsigned int wrap into account.
2290 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2293 #define perf_output_put(handle, x) \
2294 perf_output_copy((handle), &(x), sizeof(x))
2296 static void perf_output_end(struct perf_output_handle
*handle
)
2298 struct perf_counter
*counter
= handle
->counter
;
2299 struct perf_mmap_data
*data
= handle
->data
;
2301 int wakeup_events
= counter
->attr
.wakeup_events
;
2303 if (handle
->overflow
&& wakeup_events
) {
2304 int events
= atomic_inc_return(&data
->events
);
2305 if (events
>= wakeup_events
) {
2306 atomic_sub(wakeup_events
, &data
->events
);
2307 atomic_set(&data
->wakeup
, 1);
2311 perf_output_unlock(handle
);
2315 static u32
perf_counter_pid(struct perf_counter
*counter
, struct task_struct
*p
)
2318 * only top level counters have the pid namespace they were created in
2320 if (counter
->parent
)
2321 counter
= counter
->parent
;
2323 return task_tgid_nr_ns(p
, counter
->ns
);
2326 static u32
perf_counter_tid(struct perf_counter
*counter
, struct task_struct
*p
)
2329 * only top level counters have the pid namespace they were created in
2331 if (counter
->parent
)
2332 counter
= counter
->parent
;
2334 return task_pid_nr_ns(p
, counter
->ns
);
2337 static void perf_counter_output(struct perf_counter
*counter
,
2338 int nmi
, struct pt_regs
*regs
, u64 addr
)
2341 u64 sample_type
= counter
->attr
.sample_type
;
2342 struct perf_output_handle handle
;
2343 struct perf_event_header header
;
2352 struct perf_callchain_entry
*callchain
= NULL
;
2353 int callchain_size
= 0;
2360 header
.size
= sizeof(header
);
2362 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
2363 header
.misc
|= perf_misc_flags(regs
);
2365 if (sample_type
& PERF_SAMPLE_IP
) {
2366 ip
= perf_instruction_pointer(regs
);
2367 header
.type
|= PERF_SAMPLE_IP
;
2368 header
.size
+= sizeof(ip
);
2371 if (sample_type
& PERF_SAMPLE_TID
) {
2372 /* namespace issues */
2373 tid_entry
.pid
= perf_counter_pid(counter
, current
);
2374 tid_entry
.tid
= perf_counter_tid(counter
, current
);
2376 header
.type
|= PERF_SAMPLE_TID
;
2377 header
.size
+= sizeof(tid_entry
);
2380 if (sample_type
& PERF_SAMPLE_TIME
) {
2382 * Maybe do better on x86 and provide cpu_clock_nmi()
2384 time
= sched_clock();
2386 header
.type
|= PERF_SAMPLE_TIME
;
2387 header
.size
+= sizeof(u64
);
2390 if (sample_type
& PERF_SAMPLE_ADDR
) {
2391 header
.type
|= PERF_SAMPLE_ADDR
;
2392 header
.size
+= sizeof(u64
);
2395 if (sample_type
& PERF_SAMPLE_ID
) {
2396 header
.type
|= PERF_SAMPLE_ID
;
2397 header
.size
+= sizeof(u64
);
2400 if (sample_type
& PERF_SAMPLE_CPU
) {
2401 header
.type
|= PERF_SAMPLE_CPU
;
2402 header
.size
+= sizeof(cpu_entry
);
2404 cpu_entry
.cpu
= raw_smp_processor_id();
2407 if (sample_type
& PERF_SAMPLE_PERIOD
) {
2408 header
.type
|= PERF_SAMPLE_PERIOD
;
2409 header
.size
+= sizeof(u64
);
2412 if (sample_type
& PERF_SAMPLE_GROUP
) {
2413 header
.type
|= PERF_SAMPLE_GROUP
;
2414 header
.size
+= sizeof(u64
) +
2415 counter
->nr_siblings
* sizeof(group_entry
);
2418 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2419 callchain
= perf_callchain(regs
);
2422 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2424 header
.type
|= PERF_SAMPLE_CALLCHAIN
;
2425 header
.size
+= callchain_size
;
2429 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2433 perf_output_put(&handle
, header
);
2435 if (sample_type
& PERF_SAMPLE_IP
)
2436 perf_output_put(&handle
, ip
);
2438 if (sample_type
& PERF_SAMPLE_TID
)
2439 perf_output_put(&handle
, tid_entry
);
2441 if (sample_type
& PERF_SAMPLE_TIME
)
2442 perf_output_put(&handle
, time
);
2444 if (sample_type
& PERF_SAMPLE_ADDR
)
2445 perf_output_put(&handle
, addr
);
2447 if (sample_type
& PERF_SAMPLE_ID
)
2448 perf_output_put(&handle
, counter
->id
);
2450 if (sample_type
& PERF_SAMPLE_CPU
)
2451 perf_output_put(&handle
, cpu_entry
);
2453 if (sample_type
& PERF_SAMPLE_PERIOD
)
2454 perf_output_put(&handle
, counter
->hw
.sample_period
);
2457 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2459 if (sample_type
& PERF_SAMPLE_GROUP
) {
2460 struct perf_counter
*leader
, *sub
;
2461 u64 nr
= counter
->nr_siblings
;
2463 perf_output_put(&handle
, nr
);
2465 leader
= counter
->group_leader
;
2466 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2468 sub
->pmu
->read(sub
);
2470 group_entry
.id
= sub
->id
;
2471 group_entry
.counter
= atomic64_read(&sub
->count
);
2473 perf_output_put(&handle
, group_entry
);
2478 perf_output_copy(&handle
, callchain
, callchain_size
);
2480 perf_output_end(&handle
);
2487 struct perf_fork_event
{
2488 struct task_struct
*task
;
2491 struct perf_event_header header
;
2498 static void perf_counter_fork_output(struct perf_counter
*counter
,
2499 struct perf_fork_event
*fork_event
)
2501 struct perf_output_handle handle
;
2502 int size
= fork_event
->event
.header
.size
;
2503 struct task_struct
*task
= fork_event
->task
;
2504 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2509 fork_event
->event
.pid
= perf_counter_pid(counter
, task
);
2510 fork_event
->event
.ppid
= perf_counter_pid(counter
, task
->real_parent
);
2512 perf_output_put(&handle
, fork_event
->event
);
2513 perf_output_end(&handle
);
2516 static int perf_counter_fork_match(struct perf_counter
*counter
)
2518 if (counter
->attr
.comm
|| counter
->attr
.mmap
)
2524 static void perf_counter_fork_ctx(struct perf_counter_context
*ctx
,
2525 struct perf_fork_event
*fork_event
)
2527 struct perf_counter
*counter
;
2529 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2533 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2534 if (perf_counter_fork_match(counter
))
2535 perf_counter_fork_output(counter
, fork_event
);
2540 static void perf_counter_fork_event(struct perf_fork_event
*fork_event
)
2542 struct perf_cpu_context
*cpuctx
;
2543 struct perf_counter_context
*ctx
;
2545 cpuctx
= &get_cpu_var(perf_cpu_context
);
2546 perf_counter_fork_ctx(&cpuctx
->ctx
, fork_event
);
2547 put_cpu_var(perf_cpu_context
);
2551 * doesn't really matter which of the child contexts the
2552 * events ends up in.
2554 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2556 perf_counter_fork_ctx(ctx
, fork_event
);
2560 void perf_counter_fork(struct task_struct
*task
)
2562 struct perf_fork_event fork_event
;
2564 if (!atomic_read(&nr_comm_counters
) &&
2565 !atomic_read(&nr_mmap_counters
))
2568 fork_event
= (struct perf_fork_event
){
2572 .type
= PERF_EVENT_FORK
,
2573 .size
= sizeof(fork_event
.event
),
2578 perf_counter_fork_event(&fork_event
);
2585 struct perf_comm_event
{
2586 struct task_struct
*task
;
2591 struct perf_event_header header
;
2598 static void perf_counter_comm_output(struct perf_counter
*counter
,
2599 struct perf_comm_event
*comm_event
)
2601 struct perf_output_handle handle
;
2602 int size
= comm_event
->event
.header
.size
;
2603 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2608 comm_event
->event
.pid
= perf_counter_pid(counter
, comm_event
->task
);
2609 comm_event
->event
.tid
= perf_counter_tid(counter
, comm_event
->task
);
2611 perf_output_put(&handle
, comm_event
->event
);
2612 perf_output_copy(&handle
, comm_event
->comm
,
2613 comm_event
->comm_size
);
2614 perf_output_end(&handle
);
2617 static int perf_counter_comm_match(struct perf_counter
*counter
)
2619 if (counter
->attr
.comm
)
2625 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2626 struct perf_comm_event
*comm_event
)
2628 struct perf_counter
*counter
;
2630 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2634 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2635 if (perf_counter_comm_match(counter
))
2636 perf_counter_comm_output(counter
, comm_event
);
2641 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2643 struct perf_cpu_context
*cpuctx
;
2644 struct perf_counter_context
*ctx
;
2646 char *comm
= comm_event
->task
->comm
;
2648 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2650 comm_event
->comm
= comm
;
2651 comm_event
->comm_size
= size
;
2653 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2655 cpuctx
= &get_cpu_var(perf_cpu_context
);
2656 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2657 put_cpu_var(perf_cpu_context
);
2661 * doesn't really matter which of the child contexts the
2662 * events ends up in.
2664 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2666 perf_counter_comm_ctx(ctx
, comm_event
);
2670 void perf_counter_comm(struct task_struct
*task
)
2672 struct perf_comm_event comm_event
;
2674 if (!atomic_read(&nr_comm_counters
))
2677 comm_event
= (struct perf_comm_event
){
2680 .header
= { .type
= PERF_EVENT_COMM
, },
2684 perf_counter_comm_event(&comm_event
);
2691 struct perf_mmap_event
{
2692 struct vm_area_struct
*vma
;
2694 const char *file_name
;
2698 struct perf_event_header header
;
2708 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2709 struct perf_mmap_event
*mmap_event
)
2711 struct perf_output_handle handle
;
2712 int size
= mmap_event
->event
.header
.size
;
2713 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2718 mmap_event
->event
.pid
= perf_counter_pid(counter
, current
);
2719 mmap_event
->event
.tid
= perf_counter_tid(counter
, current
);
2721 perf_output_put(&handle
, mmap_event
->event
);
2722 perf_output_copy(&handle
, mmap_event
->file_name
,
2723 mmap_event
->file_size
);
2724 perf_output_end(&handle
);
2727 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2728 struct perf_mmap_event
*mmap_event
)
2730 if (counter
->attr
.mmap
)
2736 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2737 struct perf_mmap_event
*mmap_event
)
2739 struct perf_counter
*counter
;
2741 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2745 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2746 if (perf_counter_mmap_match(counter
, mmap_event
))
2747 perf_counter_mmap_output(counter
, mmap_event
);
2752 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2754 struct perf_cpu_context
*cpuctx
;
2755 struct perf_counter_context
*ctx
;
2756 struct vm_area_struct
*vma
= mmap_event
->vma
;
2757 struct file
*file
= vma
->vm_file
;
2764 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2766 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2769 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2771 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2775 name
= arch_vma_name(mmap_event
->vma
);
2780 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
2784 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2789 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2791 mmap_event
->file_name
= name
;
2792 mmap_event
->file_size
= size
;
2794 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2796 cpuctx
= &get_cpu_var(perf_cpu_context
);
2797 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2798 put_cpu_var(perf_cpu_context
);
2802 * doesn't really matter which of the child contexts the
2803 * events ends up in.
2805 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2807 perf_counter_mmap_ctx(ctx
, mmap_event
);
2813 void __perf_counter_mmap(struct vm_area_struct
*vma
)
2815 struct perf_mmap_event mmap_event
;
2817 if (!atomic_read(&nr_mmap_counters
))
2820 mmap_event
= (struct perf_mmap_event
){
2823 .header
= { .type
= PERF_EVENT_MMAP
, },
2824 .start
= vma
->vm_start
,
2825 .len
= vma
->vm_end
- vma
->vm_start
,
2826 .pgoff
= vma
->vm_pgoff
,
2830 perf_counter_mmap_event(&mmap_event
);
2834 * Log sample_period changes so that analyzing tools can re-normalize the
2838 static void perf_log_period(struct perf_counter
*counter
, u64 period
)
2840 struct perf_output_handle handle
;
2844 struct perf_event_header header
;
2850 .type
= PERF_EVENT_PERIOD
,
2852 .size
= sizeof(freq_event
),
2854 .time
= sched_clock(),
2859 if (counter
->hw
.sample_period
== period
)
2862 ret
= perf_output_begin(&handle
, counter
, sizeof(freq_event
), 0, 0);
2866 perf_output_put(&handle
, freq_event
);
2867 perf_output_end(&handle
);
2871 * IRQ throttle logging
2874 static void perf_log_throttle(struct perf_counter
*counter
, int enable
)
2876 struct perf_output_handle handle
;
2880 struct perf_event_header header
;
2882 } throttle_event
= {
2884 .type
= PERF_EVENT_THROTTLE
+ 1,
2886 .size
= sizeof(throttle_event
),
2888 .time
= sched_clock(),
2891 ret
= perf_output_begin(&handle
, counter
, sizeof(throttle_event
), 1, 0);
2895 perf_output_put(&handle
, throttle_event
);
2896 perf_output_end(&handle
);
2900 * Generic counter overflow handling.
2903 int perf_counter_overflow(struct perf_counter
*counter
,
2904 int nmi
, struct pt_regs
*regs
, u64 addr
)
2906 int events
= atomic_read(&counter
->event_limit
);
2907 int throttle
= counter
->pmu
->unthrottle
!= NULL
;
2911 counter
->hw
.interrupts
++;
2913 if (counter
->hw
.interrupts
!= MAX_INTERRUPTS
) {
2914 counter
->hw
.interrupts
++;
2915 if (HZ
*counter
->hw
.interrupts
> (u64
)sysctl_perf_counter_limit
) {
2916 counter
->hw
.interrupts
= MAX_INTERRUPTS
;
2917 perf_log_throttle(counter
, 0);
2922 * Keep re-disabling counters even though on the previous
2923 * pass we disabled it - just in case we raced with a
2924 * sched-in and the counter got enabled again:
2931 * XXX event_limit might not quite work as expected on inherited
2935 counter
->pending_kill
= POLL_IN
;
2936 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
2938 counter
->pending_kill
= POLL_HUP
;
2940 counter
->pending_disable
= 1;
2941 perf_pending_queue(&counter
->pending
,
2942 perf_pending_counter
);
2944 perf_counter_disable(counter
);
2947 perf_counter_output(counter
, nmi
, regs
, addr
);
2952 * Generic software counter infrastructure
2955 static void perf_swcounter_update(struct perf_counter
*counter
)
2957 struct hw_perf_counter
*hwc
= &counter
->hw
;
2962 prev
= atomic64_read(&hwc
->prev_count
);
2963 now
= atomic64_read(&hwc
->count
);
2964 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2969 atomic64_add(delta
, &counter
->count
);
2970 atomic64_sub(delta
, &hwc
->period_left
);
2973 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2975 struct hw_perf_counter
*hwc
= &counter
->hw
;
2976 s64 left
= atomic64_read(&hwc
->period_left
);
2977 s64 period
= hwc
->sample_period
;
2979 if (unlikely(left
<= -period
)) {
2981 atomic64_set(&hwc
->period_left
, left
);
2984 if (unlikely(left
<= 0)) {
2986 atomic64_add(period
, &hwc
->period_left
);
2989 atomic64_set(&hwc
->prev_count
, -left
);
2990 atomic64_set(&hwc
->count
, -left
);
2993 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2995 enum hrtimer_restart ret
= HRTIMER_RESTART
;
2996 struct perf_counter
*counter
;
2997 struct pt_regs
*regs
;
3000 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
3001 counter
->pmu
->read(counter
);
3003 regs
= get_irq_regs();
3005 * In case we exclude kernel IPs or are somehow not in interrupt
3006 * context, provide the next best thing, the user IP.
3008 if ((counter
->attr
.exclude_kernel
|| !regs
) &&
3009 !counter
->attr
.exclude_user
)
3010 regs
= task_pt_regs(current
);
3013 if (perf_counter_overflow(counter
, 0, regs
, 0))
3014 ret
= HRTIMER_NORESTART
;
3017 period
= max_t(u64
, 10000, counter
->hw
.sample_period
);
3018 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
3023 static void perf_swcounter_overflow(struct perf_counter
*counter
,
3024 int nmi
, struct pt_regs
*regs
, u64 addr
)
3026 perf_swcounter_update(counter
);
3027 perf_swcounter_set_period(counter
);
3028 if (perf_counter_overflow(counter
, nmi
, regs
, addr
))
3029 /* soft-disable the counter */
3034 static int perf_swcounter_is_counting(struct perf_counter
*counter
)
3036 struct perf_counter_context
*ctx
;
3037 unsigned long flags
;
3040 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
3043 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
)
3047 * If the counter is inactive, it could be just because
3048 * its task is scheduled out, or because it's in a group
3049 * which could not go on the PMU. We want to count in
3050 * the first case but not the second. If the context is
3051 * currently active then an inactive software counter must
3052 * be the second case. If it's not currently active then
3053 * we need to know whether the counter was active when the
3054 * context was last active, which we can determine by
3055 * comparing counter->tstamp_stopped with ctx->time.
3057 * We are within an RCU read-side critical section,
3058 * which protects the existence of *ctx.
3061 spin_lock_irqsave(&ctx
->lock
, flags
);
3063 /* Re-check state now we have the lock */
3064 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
||
3065 counter
->ctx
->is_active
||
3066 counter
->tstamp_stopped
< ctx
->time
)
3068 spin_unlock_irqrestore(&ctx
->lock
, flags
);
3072 static int perf_swcounter_match(struct perf_counter
*counter
,
3073 enum perf_event_types type
,
3074 u32 event
, struct pt_regs
*regs
)
3078 event_config
= ((u64
) type
<< PERF_COUNTER_TYPE_SHIFT
) | event
;
3080 if (!perf_swcounter_is_counting(counter
))
3083 if (counter
->attr
.config
!= event_config
)
3087 if (counter
->attr
.exclude_user
&& user_mode(regs
))
3090 if (counter
->attr
.exclude_kernel
&& !user_mode(regs
))
3097 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
3098 int nmi
, struct pt_regs
*regs
, u64 addr
)
3100 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
3102 if (counter
->hw
.sample_period
&& !neg
&& regs
)
3103 perf_swcounter_overflow(counter
, nmi
, regs
, addr
);
3106 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
3107 enum perf_event_types type
, u32 event
,
3108 u64 nr
, int nmi
, struct pt_regs
*regs
,
3111 struct perf_counter
*counter
;
3113 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3117 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3118 if (perf_swcounter_match(counter
, type
, event
, regs
))
3119 perf_swcounter_add(counter
, nr
, nmi
, regs
, addr
);
3124 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
3127 return &cpuctx
->recursion
[3];
3130 return &cpuctx
->recursion
[2];
3133 return &cpuctx
->recursion
[1];
3135 return &cpuctx
->recursion
[0];
3138 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
3139 u64 nr
, int nmi
, struct pt_regs
*regs
,
3142 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3143 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
3144 struct perf_counter_context
*ctx
;
3152 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
3153 nr
, nmi
, regs
, addr
);
3156 * doesn't really matter which of the child contexts the
3157 * events ends up in.
3159 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3161 perf_swcounter_ctx_event(ctx
, type
, event
, nr
, nmi
, regs
, addr
);
3168 put_cpu_var(perf_cpu_context
);
3172 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
3174 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
, addr
);
3177 static void perf_swcounter_read(struct perf_counter
*counter
)
3179 perf_swcounter_update(counter
);
3182 static int perf_swcounter_enable(struct perf_counter
*counter
)
3184 perf_swcounter_set_period(counter
);
3188 static void perf_swcounter_disable(struct perf_counter
*counter
)
3190 perf_swcounter_update(counter
);
3193 static const struct pmu perf_ops_generic
= {
3194 .enable
= perf_swcounter_enable
,
3195 .disable
= perf_swcounter_disable
,
3196 .read
= perf_swcounter_read
,
3200 * Software counter: cpu wall time clock
3203 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
3205 int cpu
= raw_smp_processor_id();
3209 now
= cpu_clock(cpu
);
3210 prev
= atomic64_read(&counter
->hw
.prev_count
);
3211 atomic64_set(&counter
->hw
.prev_count
, now
);
3212 atomic64_add(now
- prev
, &counter
->count
);
3215 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
3217 struct hw_perf_counter
*hwc
= &counter
->hw
;
3218 int cpu
= raw_smp_processor_id();
3220 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
3221 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3222 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3223 if (hwc
->sample_period
) {
3224 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3225 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3226 ns_to_ktime(period
), 0,
3227 HRTIMER_MODE_REL
, 0);
3233 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
3235 if (counter
->hw
.sample_period
)
3236 hrtimer_cancel(&counter
->hw
.hrtimer
);
3237 cpu_clock_perf_counter_update(counter
);
3240 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
3242 cpu_clock_perf_counter_update(counter
);
3245 static const struct pmu perf_ops_cpu_clock
= {
3246 .enable
= cpu_clock_perf_counter_enable
,
3247 .disable
= cpu_clock_perf_counter_disable
,
3248 .read
= cpu_clock_perf_counter_read
,
3252 * Software counter: task time clock
3255 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
3260 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
3262 atomic64_add(delta
, &counter
->count
);
3265 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
3267 struct hw_perf_counter
*hwc
= &counter
->hw
;
3270 now
= counter
->ctx
->time
;
3272 atomic64_set(&hwc
->prev_count
, now
);
3273 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3274 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3275 if (hwc
->sample_period
) {
3276 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3277 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3278 ns_to_ktime(period
), 0,
3279 HRTIMER_MODE_REL
, 0);
3285 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
3287 if (counter
->hw
.sample_period
)
3288 hrtimer_cancel(&counter
->hw
.hrtimer
);
3289 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
3293 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
3298 update_context_time(counter
->ctx
);
3299 time
= counter
->ctx
->time
;
3301 u64 now
= perf_clock();
3302 u64 delta
= now
- counter
->ctx
->timestamp
;
3303 time
= counter
->ctx
->time
+ delta
;
3306 task_clock_perf_counter_update(counter
, time
);
3309 static const struct pmu perf_ops_task_clock
= {
3310 .enable
= task_clock_perf_counter_enable
,
3311 .disable
= task_clock_perf_counter_disable
,
3312 .read
= task_clock_perf_counter_read
,
3316 * Software counter: cpu migrations
3318 void perf_counter_task_migration(struct task_struct
*task
, int cpu
)
3320 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3321 struct perf_counter_context
*ctx
;
3323 perf_swcounter_ctx_event(&cpuctx
->ctx
, PERF_TYPE_SOFTWARE
,
3324 PERF_COUNT_CPU_MIGRATIONS
,
3327 ctx
= perf_pin_task_context(task
);
3329 perf_swcounter_ctx_event(ctx
, PERF_TYPE_SOFTWARE
,
3330 PERF_COUNT_CPU_MIGRATIONS
,
3332 perf_unpin_context(ctx
);
3336 #ifdef CONFIG_EVENT_PROFILE
3337 void perf_tpcounter_event(int event_id
)
3339 struct pt_regs
*regs
= get_irq_regs();
3342 regs
= task_pt_regs(current
);
3344 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
, 0);
3346 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
3348 extern int ftrace_profile_enable(int);
3349 extern void ftrace_profile_disable(int);
3351 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
3353 ftrace_profile_disable(perf_event_id(&counter
->attr
));
3356 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3358 int event_id
= perf_event_id(&counter
->attr
);
3361 ret
= ftrace_profile_enable(event_id
);
3365 counter
->destroy
= tp_perf_counter_destroy
;
3366 counter
->hw
.sample_period
= counter
->attr
.sample_period
;
3368 return &perf_ops_generic
;
3371 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3377 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
3379 const struct pmu
*pmu
= NULL
;
3382 * Software counters (currently) can't in general distinguish
3383 * between user, kernel and hypervisor events.
3384 * However, context switches and cpu migrations are considered
3385 * to be kernel events, and page faults are never hypervisor
3388 switch (perf_event_id(&counter
->attr
)) {
3389 case PERF_COUNT_CPU_CLOCK
:
3390 pmu
= &perf_ops_cpu_clock
;
3393 case PERF_COUNT_TASK_CLOCK
:
3395 * If the user instantiates this as a per-cpu counter,
3396 * use the cpu_clock counter instead.
3398 if (counter
->ctx
->task
)
3399 pmu
= &perf_ops_task_clock
;
3401 pmu
= &perf_ops_cpu_clock
;
3404 case PERF_COUNT_PAGE_FAULTS
:
3405 case PERF_COUNT_PAGE_FAULTS_MIN
:
3406 case PERF_COUNT_PAGE_FAULTS_MAJ
:
3407 case PERF_COUNT_CONTEXT_SWITCHES
:
3408 case PERF_COUNT_CPU_MIGRATIONS
:
3409 pmu
= &perf_ops_generic
;
3417 * Allocate and initialize a counter structure
3419 static struct perf_counter
*
3420 perf_counter_alloc(struct perf_counter_attr
*attr
,
3422 struct perf_counter_context
*ctx
,
3423 struct perf_counter
*group_leader
,
3426 const struct pmu
*pmu
;
3427 struct perf_counter
*counter
;
3428 struct hw_perf_counter
*hwc
;
3431 counter
= kzalloc(sizeof(*counter
), gfpflags
);
3433 return ERR_PTR(-ENOMEM
);
3436 * Single counters are their own group leaders, with an
3437 * empty sibling list:
3440 group_leader
= counter
;
3442 mutex_init(&counter
->child_mutex
);
3443 INIT_LIST_HEAD(&counter
->child_list
);
3445 INIT_LIST_HEAD(&counter
->list_entry
);
3446 INIT_LIST_HEAD(&counter
->event_entry
);
3447 INIT_LIST_HEAD(&counter
->sibling_list
);
3448 init_waitqueue_head(&counter
->waitq
);
3450 mutex_init(&counter
->mmap_mutex
);
3453 counter
->attr
= *attr
;
3454 counter
->group_leader
= group_leader
;
3455 counter
->pmu
= NULL
;
3457 counter
->oncpu
= -1;
3459 counter
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
3460 counter
->id
= atomic64_inc_return(&perf_counter_id
);
3462 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3465 counter
->state
= PERF_COUNTER_STATE_OFF
;
3470 if (attr
->freq
&& attr
->sample_freq
)
3471 hwc
->sample_period
= div64_u64(TICK_NSEC
, attr
->sample_freq
);
3473 hwc
->sample_period
= attr
->sample_period
;
3476 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3478 if (attr
->inherit
&& (attr
->sample_type
& PERF_SAMPLE_GROUP
))
3481 if (perf_event_raw(attr
)) {
3482 pmu
= hw_perf_counter_init(counter
);
3486 switch (perf_event_type(attr
)) {
3487 case PERF_TYPE_HARDWARE
:
3488 pmu
= hw_perf_counter_init(counter
);
3491 case PERF_TYPE_SOFTWARE
:
3492 pmu
= sw_perf_counter_init(counter
);
3495 case PERF_TYPE_TRACEPOINT
:
3496 pmu
= tp_perf_counter_init(counter
);
3503 else if (IS_ERR(pmu
))
3508 put_pid_ns(counter
->ns
);
3510 return ERR_PTR(err
);
3515 atomic_inc(&nr_counters
);
3516 if (counter
->attr
.mmap
)
3517 atomic_inc(&nr_mmap_counters
);
3518 if (counter
->attr
.comm
)
3519 atomic_inc(&nr_comm_counters
);
3525 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3527 * @attr_uptr: event type attributes for monitoring/sampling
3530 * @group_fd: group leader counter fd
3532 SYSCALL_DEFINE5(perf_counter_open
,
3533 const struct perf_counter_attr __user
*, attr_uptr
,
3534 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
3536 struct perf_counter
*counter
, *group_leader
;
3537 struct perf_counter_attr attr
;
3538 struct perf_counter_context
*ctx
;
3539 struct file
*counter_file
= NULL
;
3540 struct file
*group_file
= NULL
;
3541 int fput_needed
= 0;
3542 int fput_needed2
= 0;
3545 /* for future expandability... */
3549 if (copy_from_user(&attr
, attr_uptr
, sizeof(attr
)) != 0)
3553 * Get the target context (task or percpu):
3555 ctx
= find_get_context(pid
, cpu
);
3557 return PTR_ERR(ctx
);
3560 * Look up the group leader (we will attach this counter to it):
3562 group_leader
= NULL
;
3563 if (group_fd
!= -1) {
3565 group_file
= fget_light(group_fd
, &fput_needed
);
3567 goto err_put_context
;
3568 if (group_file
->f_op
!= &perf_fops
)
3569 goto err_put_context
;
3571 group_leader
= group_file
->private_data
;
3573 * Do not allow a recursive hierarchy (this new sibling
3574 * becoming part of another group-sibling):
3576 if (group_leader
->group_leader
!= group_leader
)
3577 goto err_put_context
;
3579 * Do not allow to attach to a group in a different
3580 * task or CPU context:
3582 if (group_leader
->ctx
!= ctx
)
3583 goto err_put_context
;
3585 * Only a group leader can be exclusive or pinned
3587 if (attr
.exclusive
|| attr
.pinned
)
3588 goto err_put_context
;
3591 counter
= perf_counter_alloc(&attr
, cpu
, ctx
, group_leader
,
3593 ret
= PTR_ERR(counter
);
3594 if (IS_ERR(counter
))
3595 goto err_put_context
;
3597 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
3599 goto err_free_put_context
;
3601 counter_file
= fget_light(ret
, &fput_needed2
);
3603 goto err_free_put_context
;
3605 counter
->filp
= counter_file
;
3606 WARN_ON_ONCE(ctx
->parent_ctx
);
3607 mutex_lock(&ctx
->mutex
);
3608 perf_install_in_context(ctx
, counter
, cpu
);
3610 mutex_unlock(&ctx
->mutex
);
3612 counter
->owner
= current
;
3613 get_task_struct(current
);
3614 mutex_lock(¤t
->perf_counter_mutex
);
3615 list_add_tail(&counter
->owner_entry
, ¤t
->perf_counter_list
);
3616 mutex_unlock(¤t
->perf_counter_mutex
);
3618 fput_light(counter_file
, fput_needed2
);
3621 fput_light(group_file
, fput_needed
);
3625 err_free_put_context
:
3635 * inherit a counter from parent task to child task:
3637 static struct perf_counter
*
3638 inherit_counter(struct perf_counter
*parent_counter
,
3639 struct task_struct
*parent
,
3640 struct perf_counter_context
*parent_ctx
,
3641 struct task_struct
*child
,
3642 struct perf_counter
*group_leader
,
3643 struct perf_counter_context
*child_ctx
)
3645 struct perf_counter
*child_counter
;
3648 * Instead of creating recursive hierarchies of counters,
3649 * we link inherited counters back to the original parent,
3650 * which has a filp for sure, which we use as the reference
3653 if (parent_counter
->parent
)
3654 parent_counter
= parent_counter
->parent
;
3656 child_counter
= perf_counter_alloc(&parent_counter
->attr
,
3657 parent_counter
->cpu
, child_ctx
,
3658 group_leader
, GFP_KERNEL
);
3659 if (IS_ERR(child_counter
))
3660 return child_counter
;
3664 * Make the child state follow the state of the parent counter,
3665 * not its attr.disabled bit. We hold the parent's mutex,
3666 * so we won't race with perf_counter_{en, dis}able_family.
3668 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3669 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3671 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3674 * Link it up in the child's context:
3676 add_counter_to_ctx(child_counter
, child_ctx
);
3678 child_counter
->parent
= parent_counter
;
3680 * inherit into child's child as well:
3682 child_counter
->attr
.inherit
= 1;
3685 * Get a reference to the parent filp - we will fput it
3686 * when the child counter exits. This is safe to do because
3687 * we are in the parent and we know that the filp still
3688 * exists and has a nonzero count:
3690 atomic_long_inc(&parent_counter
->filp
->f_count
);
3693 * Link this into the parent counter's child list
3695 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3696 mutex_lock(&parent_counter
->child_mutex
);
3697 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3698 mutex_unlock(&parent_counter
->child_mutex
);
3700 return child_counter
;
3703 static int inherit_group(struct perf_counter
*parent_counter
,
3704 struct task_struct
*parent
,
3705 struct perf_counter_context
*parent_ctx
,
3706 struct task_struct
*child
,
3707 struct perf_counter_context
*child_ctx
)
3709 struct perf_counter
*leader
;
3710 struct perf_counter
*sub
;
3711 struct perf_counter
*child_ctr
;
3713 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3714 child
, NULL
, child_ctx
);
3716 return PTR_ERR(leader
);
3717 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3718 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3719 child
, leader
, child_ctx
);
3720 if (IS_ERR(child_ctr
))
3721 return PTR_ERR(child_ctr
);
3726 static void sync_child_counter(struct perf_counter
*child_counter
,
3727 struct perf_counter
*parent_counter
)
3731 child_val
= atomic64_read(&child_counter
->count
);
3734 * Add back the child's count to the parent's count:
3736 atomic64_add(child_val
, &parent_counter
->count
);
3737 atomic64_add(child_counter
->total_time_enabled
,
3738 &parent_counter
->child_total_time_enabled
);
3739 atomic64_add(child_counter
->total_time_running
,
3740 &parent_counter
->child_total_time_running
);
3743 * Remove this counter from the parent's list
3745 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3746 mutex_lock(&parent_counter
->child_mutex
);
3747 list_del_init(&child_counter
->child_list
);
3748 mutex_unlock(&parent_counter
->child_mutex
);
3751 * Release the parent counter, if this was the last
3754 fput(parent_counter
->filp
);
3758 __perf_counter_exit_task(struct perf_counter
*child_counter
,
3759 struct perf_counter_context
*child_ctx
)
3761 struct perf_counter
*parent_counter
;
3763 update_counter_times(child_counter
);
3764 perf_counter_remove_from_context(child_counter
);
3766 parent_counter
= child_counter
->parent
;
3768 * It can happen that parent exits first, and has counters
3769 * that are still around due to the child reference. These
3770 * counters need to be zapped - but otherwise linger.
3772 if (parent_counter
) {
3773 sync_child_counter(child_counter
, parent_counter
);
3774 free_counter(child_counter
);
3779 * When a child task exits, feed back counter values to parent counters.
3781 void perf_counter_exit_task(struct task_struct
*child
)
3783 struct perf_counter
*child_counter
, *tmp
;
3784 struct perf_counter_context
*child_ctx
;
3785 unsigned long flags
;
3787 if (likely(!child
->perf_counter_ctxp
))
3790 local_irq_save(flags
);
3792 * We can't reschedule here because interrupts are disabled,
3793 * and either child is current or it is a task that can't be
3794 * scheduled, so we are now safe from rescheduling changing
3797 child_ctx
= child
->perf_counter_ctxp
;
3798 __perf_counter_task_sched_out(child_ctx
);
3801 * Take the context lock here so that if find_get_context is
3802 * reading child->perf_counter_ctxp, we wait until it has
3803 * incremented the context's refcount before we do put_ctx below.
3805 spin_lock(&child_ctx
->lock
);
3806 child
->perf_counter_ctxp
= NULL
;
3807 if (child_ctx
->parent_ctx
) {
3809 * This context is a clone; unclone it so it can't get
3810 * swapped to another process while we're removing all
3811 * the counters from it.
3813 put_ctx(child_ctx
->parent_ctx
);
3814 child_ctx
->parent_ctx
= NULL
;
3816 spin_unlock(&child_ctx
->lock
);
3817 local_irq_restore(flags
);
3819 mutex_lock(&child_ctx
->mutex
);
3822 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
3824 __perf_counter_exit_task(child_counter
, child_ctx
);
3827 * If the last counter was a group counter, it will have appended all
3828 * its siblings to the list, but we obtained 'tmp' before that which
3829 * will still point to the list head terminating the iteration.
3831 if (!list_empty(&child_ctx
->counter_list
))
3834 mutex_unlock(&child_ctx
->mutex
);
3840 * free an unexposed, unused context as created by inheritance by
3841 * init_task below, used by fork() in case of fail.
3843 void perf_counter_free_task(struct task_struct
*task
)
3845 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
3846 struct perf_counter
*counter
, *tmp
;
3851 mutex_lock(&ctx
->mutex
);
3853 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
) {
3854 struct perf_counter
*parent
= counter
->parent
;
3856 if (WARN_ON_ONCE(!parent
))
3859 mutex_lock(&parent
->child_mutex
);
3860 list_del_init(&counter
->child_list
);
3861 mutex_unlock(&parent
->child_mutex
);
3865 list_del_counter(counter
, ctx
);
3866 free_counter(counter
);
3869 if (!list_empty(&ctx
->counter_list
))
3872 mutex_unlock(&ctx
->mutex
);
3878 * Initialize the perf_counter context in task_struct
3880 int perf_counter_init_task(struct task_struct
*child
)
3882 struct perf_counter_context
*child_ctx
, *parent_ctx
;
3883 struct perf_counter_context
*cloned_ctx
;
3884 struct perf_counter
*counter
;
3885 struct task_struct
*parent
= current
;
3886 int inherited_all
= 1;
3889 child
->perf_counter_ctxp
= NULL
;
3891 mutex_init(&child
->perf_counter_mutex
);
3892 INIT_LIST_HEAD(&child
->perf_counter_list
);
3894 if (likely(!parent
->perf_counter_ctxp
))
3898 * This is executed from the parent task context, so inherit
3899 * counters that have been marked for cloning.
3900 * First allocate and initialize a context for the child.
3903 child_ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
3907 __perf_counter_init_context(child_ctx
, child
);
3908 child
->perf_counter_ctxp
= child_ctx
;
3909 get_task_struct(child
);
3912 * If the parent's context is a clone, pin it so it won't get
3915 parent_ctx
= perf_pin_task_context(parent
);
3918 * No need to check if parent_ctx != NULL here; since we saw
3919 * it non-NULL earlier, the only reason for it to become NULL
3920 * is if we exit, and since we're currently in the middle of
3921 * a fork we can't be exiting at the same time.
3925 * Lock the parent list. No need to lock the child - not PID
3926 * hashed yet and not running, so nobody can access it.
3928 mutex_lock(&parent_ctx
->mutex
);
3931 * We dont have to disable NMIs - we are only looking at
3932 * the list, not manipulating it:
3934 list_for_each_entry_rcu(counter
, &parent_ctx
->event_list
, event_entry
) {
3935 if (counter
!= counter
->group_leader
)
3938 if (!counter
->attr
.inherit
) {
3943 ret
= inherit_group(counter
, parent
, parent_ctx
,
3951 if (inherited_all
) {
3953 * Mark the child context as a clone of the parent
3954 * context, or of whatever the parent is a clone of.
3955 * Note that if the parent is a clone, it could get
3956 * uncloned at any point, but that doesn't matter
3957 * because the list of counters and the generation
3958 * count can't have changed since we took the mutex.
3960 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
3962 child_ctx
->parent_ctx
= cloned_ctx
;
3963 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
3965 child_ctx
->parent_ctx
= parent_ctx
;
3966 child_ctx
->parent_gen
= parent_ctx
->generation
;
3968 get_ctx(child_ctx
->parent_ctx
);
3971 mutex_unlock(&parent_ctx
->mutex
);
3973 perf_unpin_context(parent_ctx
);
3978 static void __cpuinit
perf_counter_init_cpu(int cpu
)
3980 struct perf_cpu_context
*cpuctx
;
3982 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3983 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
3985 spin_lock(&perf_resource_lock
);
3986 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
3987 spin_unlock(&perf_resource_lock
);
3989 hw_perf_counter_setup(cpu
);
3992 #ifdef CONFIG_HOTPLUG_CPU
3993 static void __perf_counter_exit_cpu(void *info
)
3995 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3996 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3997 struct perf_counter
*counter
, *tmp
;
3999 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
4000 __perf_counter_remove_from_context(counter
);
4002 static void perf_counter_exit_cpu(int cpu
)
4004 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4005 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4007 mutex_lock(&ctx
->mutex
);
4008 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
4009 mutex_unlock(&ctx
->mutex
);
4012 static inline void perf_counter_exit_cpu(int cpu
) { }
4015 static int __cpuinit
4016 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
4018 unsigned int cpu
= (long)hcpu
;
4022 case CPU_UP_PREPARE
:
4023 case CPU_UP_PREPARE_FROZEN
:
4024 perf_counter_init_cpu(cpu
);
4027 case CPU_DOWN_PREPARE
:
4028 case CPU_DOWN_PREPARE_FROZEN
:
4029 perf_counter_exit_cpu(cpu
);
4040 * This has to have a higher priority than migration_notifier in sched.c.
4042 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
4043 .notifier_call
= perf_cpu_notify
,
4047 void __init
perf_counter_init(void)
4049 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
4050 (void *)(long)smp_processor_id());
4051 register_cpu_notifier(&perf_cpu_nb
);
4054 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
4056 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
4060 perf_set_reserve_percpu(struct sysdev_class
*class,
4064 struct perf_cpu_context
*cpuctx
;
4068 err
= strict_strtoul(buf
, 10, &val
);
4071 if (val
> perf_max_counters
)
4074 spin_lock(&perf_resource_lock
);
4075 perf_reserved_percpu
= val
;
4076 for_each_online_cpu(cpu
) {
4077 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4078 spin_lock_irq(&cpuctx
->ctx
.lock
);
4079 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
4080 perf_max_counters
- perf_reserved_percpu
);
4081 cpuctx
->max_pertask
= mpt
;
4082 spin_unlock_irq(&cpuctx
->ctx
.lock
);
4084 spin_unlock(&perf_resource_lock
);
4089 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
4091 return sprintf(buf
, "%d\n", perf_overcommit
);
4095 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
4100 err
= strict_strtoul(buf
, 10, &val
);
4106 spin_lock(&perf_resource_lock
);
4107 perf_overcommit
= val
;
4108 spin_unlock(&perf_resource_lock
);
4113 static SYSDEV_CLASS_ATTR(
4116 perf_show_reserve_percpu
,
4117 perf_set_reserve_percpu
4120 static SYSDEV_CLASS_ATTR(
4123 perf_show_overcommit
,
4127 static struct attribute
*perfclass_attrs
[] = {
4128 &attr_reserve_percpu
.attr
,
4129 &attr_overcommit
.attr
,
4133 static struct attribute_group perfclass_attr_group
= {
4134 .attrs
= perfclass_attrs
,
4135 .name
= "perf_counters",
4138 static int __init
perf_counter_sysfs_init(void)
4140 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
4141 &perfclass_attr_group
);
4143 device_initcall(perf_counter_sysfs_init
);