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_tracking __read_mostly
;
44 static atomic_t nr_munmap_tracking __read_mostly
;
45 static atomic_t nr_comm_tracking __read_mostly
;
47 int sysctl_perf_counter_priv __read_mostly
; /* do we need to be privileged */
48 int sysctl_perf_counter_mlock __read_mostly
= 512; /* 'free' kb per user */
49 int sysctl_perf_counter_limit __read_mostly
= 100000; /* max NMIs per second */
52 * Lock for (sysadmin-configurable) counter reservations:
54 static DEFINE_SPINLOCK(perf_resource_lock
);
57 * Architecture provided APIs - weak aliases:
59 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
64 void __weak
hw_perf_disable(void) { barrier(); }
65 void __weak
hw_perf_enable(void) { barrier(); }
67 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
70 hw_perf_group_sched_in(struct perf_counter
*group_leader
,
71 struct perf_cpu_context
*cpuctx
,
72 struct perf_counter_context
*ctx
, int cpu
)
77 void __weak
perf_counter_print_debug(void) { }
79 static DEFINE_PER_CPU(int, disable_count
);
81 void __perf_disable(void)
83 __get_cpu_var(disable_count
)++;
86 bool __perf_enable(void)
88 return !--__get_cpu_var(disable_count
);
91 void perf_disable(void)
97 void perf_enable(void)
103 static void get_ctx(struct perf_counter_context
*ctx
)
105 atomic_inc(&ctx
->refcount
);
108 static void free_ctx(struct rcu_head
*head
)
110 struct perf_counter_context
*ctx
;
112 ctx
= container_of(head
, struct perf_counter_context
, rcu_head
);
116 static void put_ctx(struct perf_counter_context
*ctx
)
118 if (atomic_dec_and_test(&ctx
->refcount
)) {
120 put_ctx(ctx
->parent_ctx
);
122 put_task_struct(ctx
->task
);
123 call_rcu(&ctx
->rcu_head
, free_ctx
);
128 * Get the perf_counter_context for a task and lock it.
129 * This has to cope with with the fact that until it is locked,
130 * the context could get moved to another task.
132 static struct perf_counter_context
*
133 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
135 struct perf_counter_context
*ctx
;
139 ctx
= rcu_dereference(task
->perf_counter_ctxp
);
142 * If this context is a clone of another, it might
143 * get swapped for another underneath us by
144 * perf_counter_task_sched_out, though the
145 * rcu_read_lock() protects us from any context
146 * getting freed. Lock the context and check if it
147 * got swapped before we could get the lock, and retry
148 * if so. If we locked the right context, then it
149 * can't get swapped on us any more.
151 spin_lock_irqsave(&ctx
->lock
, *flags
);
152 if (ctx
!= rcu_dereference(task
->perf_counter_ctxp
)) {
153 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
162 * Get the context for a task and increment its pin_count so it
163 * can't get swapped to another task. This also increments its
164 * reference count so that the context can't get freed.
166 static struct perf_counter_context
*perf_pin_task_context(struct task_struct
*task
)
168 struct perf_counter_context
*ctx
;
171 ctx
= perf_lock_task_context(task
, &flags
);
175 spin_unlock_irqrestore(&ctx
->lock
, flags
);
180 static void perf_unpin_context(struct perf_counter_context
*ctx
)
184 spin_lock_irqsave(&ctx
->lock
, flags
);
186 spin_unlock_irqrestore(&ctx
->lock
, flags
);
191 * Add a counter from the lists for its context.
192 * Must be called with ctx->mutex and ctx->lock held.
195 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
197 struct perf_counter
*group_leader
= counter
->group_leader
;
200 * Depending on whether it is a standalone or sibling counter,
201 * add it straight to the context's counter list, or to the group
202 * leader's sibling list:
204 if (group_leader
== counter
)
205 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
207 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
208 group_leader
->nr_siblings
++;
211 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
216 * Remove a counter from the lists for its context.
217 * Must be called with ctx->mutex and ctx->lock held.
220 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
222 struct perf_counter
*sibling
, *tmp
;
224 if (list_empty(&counter
->list_entry
))
228 list_del_init(&counter
->list_entry
);
229 list_del_rcu(&counter
->event_entry
);
231 if (counter
->group_leader
!= counter
)
232 counter
->group_leader
->nr_siblings
--;
235 * If this was a group counter with sibling counters then
236 * upgrade the siblings to singleton counters by adding them
237 * to the context list directly:
239 list_for_each_entry_safe(sibling
, tmp
,
240 &counter
->sibling_list
, list_entry
) {
242 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
243 sibling
->group_leader
= sibling
;
248 counter_sched_out(struct perf_counter
*counter
,
249 struct perf_cpu_context
*cpuctx
,
250 struct perf_counter_context
*ctx
)
252 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
255 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
256 counter
->tstamp_stopped
= ctx
->time
;
257 counter
->pmu
->disable(counter
);
260 if (!is_software_counter(counter
))
261 cpuctx
->active_oncpu
--;
263 if (counter
->hw_event
.exclusive
|| !cpuctx
->active_oncpu
)
264 cpuctx
->exclusive
= 0;
268 group_sched_out(struct perf_counter
*group_counter
,
269 struct perf_cpu_context
*cpuctx
,
270 struct perf_counter_context
*ctx
)
272 struct perf_counter
*counter
;
274 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
277 counter_sched_out(group_counter
, cpuctx
, ctx
);
280 * Schedule out siblings (if any):
282 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
283 counter_sched_out(counter
, cpuctx
, ctx
);
285 if (group_counter
->hw_event
.exclusive
)
286 cpuctx
->exclusive
= 0;
290 * Cross CPU call to remove a performance counter
292 * We disable the counter on the hardware level first. After that we
293 * remove it from the context list.
295 static void __perf_counter_remove_from_context(void *info
)
297 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
298 struct perf_counter
*counter
= info
;
299 struct perf_counter_context
*ctx
= counter
->ctx
;
302 * If this is a task context, we need to check whether it is
303 * the current task context of this cpu. If not it has been
304 * scheduled out before the smp call arrived.
306 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
309 spin_lock(&ctx
->lock
);
311 * Protect the list operation against NMI by disabling the
312 * counters on a global level.
316 counter_sched_out(counter
, cpuctx
, ctx
);
318 list_del_counter(counter
, ctx
);
322 * Allow more per task counters with respect to the
325 cpuctx
->max_pertask
=
326 min(perf_max_counters
- ctx
->nr_counters
,
327 perf_max_counters
- perf_reserved_percpu
);
331 spin_unlock(&ctx
->lock
);
336 * Remove the counter from a task's (or a CPU's) list of counters.
338 * Must be called with ctx->mutex held.
340 * CPU counters are removed with a smp call. For task counters we only
341 * call when the task is on a CPU.
343 * If counter->ctx is a cloned context, callers must make sure that
344 * every task struct that counter->ctx->task could possibly point to
345 * remains valid. This is OK when called from perf_release since
346 * that only calls us on the top-level context, which can't be a clone.
347 * When called from perf_counter_exit_task, it's OK because the
348 * context has been detached from its task.
350 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
352 struct perf_counter_context
*ctx
= counter
->ctx
;
353 struct task_struct
*task
= ctx
->task
;
357 * Per cpu counters are removed via an smp call and
358 * the removal is always sucessful.
360 smp_call_function_single(counter
->cpu
,
361 __perf_counter_remove_from_context
,
367 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
370 spin_lock_irq(&ctx
->lock
);
372 * If the context is active we need to retry the smp call.
374 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
375 spin_unlock_irq(&ctx
->lock
);
380 * The lock prevents that this context is scheduled in so we
381 * can remove the counter safely, if the call above did not
384 if (!list_empty(&counter
->list_entry
)) {
385 list_del_counter(counter
, ctx
);
387 spin_unlock_irq(&ctx
->lock
);
390 static inline u64
perf_clock(void)
392 return cpu_clock(smp_processor_id());
396 * Update the record of the current time in a context.
398 static void update_context_time(struct perf_counter_context
*ctx
)
400 u64 now
= perf_clock();
402 ctx
->time
+= now
- ctx
->timestamp
;
403 ctx
->timestamp
= now
;
407 * Update the total_time_enabled and total_time_running fields for a counter.
409 static void update_counter_times(struct perf_counter
*counter
)
411 struct perf_counter_context
*ctx
= counter
->ctx
;
414 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
417 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
419 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
420 run_end
= counter
->tstamp_stopped
;
424 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
428 * Update total_time_enabled and total_time_running for all counters in a group.
430 static void update_group_times(struct perf_counter
*leader
)
432 struct perf_counter
*counter
;
434 update_counter_times(leader
);
435 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
436 update_counter_times(counter
);
440 * Cross CPU call to disable a performance counter
442 static void __perf_counter_disable(void *info
)
444 struct perf_counter
*counter
= info
;
445 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
446 struct perf_counter_context
*ctx
= counter
->ctx
;
449 * If this is a per-task counter, need to check whether this
450 * counter's task is the current task on this cpu.
452 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
455 spin_lock(&ctx
->lock
);
458 * If the counter is on, turn it off.
459 * If it is in error state, leave it in error state.
461 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
462 update_context_time(ctx
);
463 update_counter_times(counter
);
464 if (counter
== counter
->group_leader
)
465 group_sched_out(counter
, cpuctx
, ctx
);
467 counter_sched_out(counter
, cpuctx
, ctx
);
468 counter
->state
= PERF_COUNTER_STATE_OFF
;
471 spin_unlock(&ctx
->lock
);
477 * If counter->ctx is a cloned context, callers must make sure that
478 * every task struct that counter->ctx->task could possibly point to
479 * remains valid. This condition is satisifed when called through
480 * perf_counter_for_each_child or perf_counter_for_each because they
481 * hold the top-level counter's child_mutex, so any descendant that
482 * goes to exit will block in sync_child_counter.
483 * When called from perf_pending_counter it's OK because counter->ctx
484 * is the current context on this CPU and preemption is disabled,
485 * hence we can't get into perf_counter_task_sched_out for this context.
487 static void perf_counter_disable(struct perf_counter
*counter
)
489 struct perf_counter_context
*ctx
= counter
->ctx
;
490 struct task_struct
*task
= ctx
->task
;
494 * Disable the counter on the cpu that it's on
496 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
502 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
504 spin_lock_irq(&ctx
->lock
);
506 * If the counter is still active, we need to retry the cross-call.
508 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
509 spin_unlock_irq(&ctx
->lock
);
514 * Since we have the lock this context can't be scheduled
515 * in, so we can change the state safely.
517 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
518 update_counter_times(counter
);
519 counter
->state
= PERF_COUNTER_STATE_OFF
;
522 spin_unlock_irq(&ctx
->lock
);
526 counter_sched_in(struct perf_counter
*counter
,
527 struct perf_cpu_context
*cpuctx
,
528 struct perf_counter_context
*ctx
,
531 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
534 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
535 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
537 * The new state must be visible before we turn it on in the hardware:
541 if (counter
->pmu
->enable(counter
)) {
542 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
547 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
549 if (!is_software_counter(counter
))
550 cpuctx
->active_oncpu
++;
553 if (counter
->hw_event
.exclusive
)
554 cpuctx
->exclusive
= 1;
560 group_sched_in(struct perf_counter
*group_counter
,
561 struct perf_cpu_context
*cpuctx
,
562 struct perf_counter_context
*ctx
,
565 struct perf_counter
*counter
, *partial_group
;
568 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
571 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
573 return ret
< 0 ? ret
: 0;
575 group_counter
->prev_state
= group_counter
->state
;
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 counter
->prev_state
= counter
->state
;
584 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
585 partial_group
= counter
;
594 * Groups can be scheduled in as one unit only, so undo any
595 * partial group before returning:
597 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
598 if (counter
== partial_group
)
600 counter_sched_out(counter
, cpuctx
, ctx
);
602 counter_sched_out(group_counter
, cpuctx
, ctx
);
608 * Return 1 for a group consisting entirely of software counters,
609 * 0 if the group contains any hardware counters.
611 static int is_software_only_group(struct perf_counter
*leader
)
613 struct perf_counter
*counter
;
615 if (!is_software_counter(leader
))
618 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
619 if (!is_software_counter(counter
))
626 * Work out whether we can put this counter group on the CPU now.
628 static int group_can_go_on(struct perf_counter
*counter
,
629 struct perf_cpu_context
*cpuctx
,
633 * Groups consisting entirely of software counters can always go on.
635 if (is_software_only_group(counter
))
638 * If an exclusive group is already on, no other hardware
639 * counters can go on.
641 if (cpuctx
->exclusive
)
644 * If this group is exclusive and there are already
645 * counters on the CPU, it can't go on.
647 if (counter
->hw_event
.exclusive
&& cpuctx
->active_oncpu
)
650 * Otherwise, try to add it if all previous groups were able
656 static void add_counter_to_ctx(struct perf_counter
*counter
,
657 struct perf_counter_context
*ctx
)
659 list_add_counter(counter
, ctx
);
660 counter
->prev_state
= PERF_COUNTER_STATE_OFF
;
661 counter
->tstamp_enabled
= ctx
->time
;
662 counter
->tstamp_running
= ctx
->time
;
663 counter
->tstamp_stopped
= ctx
->time
;
667 * Cross CPU call to install and enable a performance counter
669 * Must be called with ctx->mutex held
671 static void __perf_install_in_context(void *info
)
673 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
674 struct perf_counter
*counter
= info
;
675 struct perf_counter_context
*ctx
= counter
->ctx
;
676 struct perf_counter
*leader
= counter
->group_leader
;
677 int cpu
= smp_processor_id();
681 * If this is a task context, we need to check whether it is
682 * the current task context of this cpu. If not it has been
683 * scheduled out before the smp call arrived.
684 * Or possibly this is the right context but it isn't
685 * on this cpu because it had no counters.
687 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
688 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
690 cpuctx
->task_ctx
= ctx
;
693 spin_lock(&ctx
->lock
);
695 update_context_time(ctx
);
698 * Protect the list operation against NMI by disabling the
699 * counters on a global level. NOP for non NMI based counters.
703 add_counter_to_ctx(counter
, ctx
);
706 * Don't put the counter on if it is disabled or if
707 * it is in a group and the group isn't on.
709 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
710 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
714 * An exclusive counter can't go on if there are already active
715 * hardware counters, and no hardware counter can go on if there
716 * is already an exclusive counter on.
718 if (!group_can_go_on(counter
, cpuctx
, 1))
721 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
725 * This counter couldn't go on. If it is in a group
726 * then we have to pull the whole group off.
727 * If the counter group is pinned then put it in error state.
729 if (leader
!= counter
)
730 group_sched_out(leader
, cpuctx
, ctx
);
731 if (leader
->hw_event
.pinned
) {
732 update_group_times(leader
);
733 leader
->state
= PERF_COUNTER_STATE_ERROR
;
737 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
738 cpuctx
->max_pertask
--;
743 spin_unlock(&ctx
->lock
);
747 * Attach a performance counter to a context
749 * First we add the counter to the list with the hardware enable bit
750 * in counter->hw_config cleared.
752 * If the counter is attached to a task which is on a CPU we use a smp
753 * call to enable it in the task context. The task might have been
754 * scheduled away, but we check this in the smp call again.
756 * Must be called with ctx->mutex held.
759 perf_install_in_context(struct perf_counter_context
*ctx
,
760 struct perf_counter
*counter
,
763 struct task_struct
*task
= ctx
->task
;
767 * Per cpu counters are installed via an smp call and
768 * the install is always sucessful.
770 smp_call_function_single(cpu
, __perf_install_in_context
,
776 task_oncpu_function_call(task
, __perf_install_in_context
,
779 spin_lock_irq(&ctx
->lock
);
781 * we need to retry the smp call.
783 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
784 spin_unlock_irq(&ctx
->lock
);
789 * The lock prevents that this context is scheduled in so we
790 * can add the counter safely, if it the call above did not
793 if (list_empty(&counter
->list_entry
))
794 add_counter_to_ctx(counter
, ctx
);
795 spin_unlock_irq(&ctx
->lock
);
799 * Cross CPU call to enable a performance counter
801 static void __perf_counter_enable(void *info
)
803 struct perf_counter
*counter
= info
;
804 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
805 struct perf_counter_context
*ctx
= counter
->ctx
;
806 struct perf_counter
*leader
= counter
->group_leader
;
810 * If this is a per-task counter, need to check whether this
811 * counter's task is the current task on this cpu.
813 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
814 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
816 cpuctx
->task_ctx
= ctx
;
819 spin_lock(&ctx
->lock
);
821 update_context_time(ctx
);
823 counter
->prev_state
= counter
->state
;
824 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
826 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
827 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
830 * If the counter is in a group and isn't the group leader,
831 * then don't put it on unless the group is on.
833 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
836 if (!group_can_go_on(counter
, cpuctx
, 1)) {
840 if (counter
== leader
)
841 err
= group_sched_in(counter
, cpuctx
, ctx
,
844 err
= counter_sched_in(counter
, cpuctx
, ctx
,
851 * If this counter can't go on and it's part of a
852 * group, then the whole group has to come off.
854 if (leader
!= counter
)
855 group_sched_out(leader
, cpuctx
, ctx
);
856 if (leader
->hw_event
.pinned
) {
857 update_group_times(leader
);
858 leader
->state
= PERF_COUNTER_STATE_ERROR
;
863 spin_unlock(&ctx
->lock
);
869 * If counter->ctx is a cloned context, callers must make sure that
870 * every task struct that counter->ctx->task could possibly point to
871 * remains valid. This condition is satisfied when called through
872 * perf_counter_for_each_child or perf_counter_for_each as described
873 * for perf_counter_disable.
875 static void perf_counter_enable(struct perf_counter
*counter
)
877 struct perf_counter_context
*ctx
= counter
->ctx
;
878 struct task_struct
*task
= ctx
->task
;
882 * Enable the counter on the cpu that it's on
884 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
889 spin_lock_irq(&ctx
->lock
);
890 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
894 * If the counter is in error state, clear that first.
895 * That way, if we see the counter in error state below, we
896 * know that it has gone back into error state, as distinct
897 * from the task having been scheduled away before the
898 * cross-call arrived.
900 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
901 counter
->state
= PERF_COUNTER_STATE_OFF
;
904 spin_unlock_irq(&ctx
->lock
);
905 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
907 spin_lock_irq(&ctx
->lock
);
910 * If the context is active and the counter is still off,
911 * we need to retry the cross-call.
913 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
917 * Since we have the lock this context can't be scheduled
918 * in, so we can change the state safely.
920 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
921 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
922 counter
->tstamp_enabled
=
923 ctx
->time
- counter
->total_time_enabled
;
926 spin_unlock_irq(&ctx
->lock
);
929 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
932 * not supported on inherited counters
934 if (counter
->hw_event
.inherit
)
937 atomic_add(refresh
, &counter
->event_limit
);
938 perf_counter_enable(counter
);
943 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
944 struct perf_cpu_context
*cpuctx
)
946 struct perf_counter
*counter
;
948 spin_lock(&ctx
->lock
);
950 if (likely(!ctx
->nr_counters
))
952 update_context_time(ctx
);
955 if (ctx
->nr_active
) {
956 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
957 if (counter
!= counter
->group_leader
)
958 counter_sched_out(counter
, cpuctx
, ctx
);
960 group_sched_out(counter
, cpuctx
, ctx
);
965 spin_unlock(&ctx
->lock
);
969 * Test whether two contexts are equivalent, i.e. whether they
970 * have both been cloned from the same version of the same context
971 * and they both have the same number of enabled counters.
972 * If the number of enabled counters is the same, then the set
973 * of enabled counters should be the same, because these are both
974 * inherited contexts, therefore we can't access individual counters
975 * in them directly with an fd; we can only enable/disable all
976 * counters via prctl, or enable/disable all counters in a family
977 * via ioctl, which will have the same effect on both contexts.
979 static int context_equiv(struct perf_counter_context
*ctx1
,
980 struct perf_counter_context
*ctx2
)
982 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
983 && ctx1
->parent_gen
== ctx2
->parent_gen
984 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
988 * Called from scheduler to remove the counters of the current task,
989 * with interrupts disabled.
991 * We stop each counter and update the counter value in counter->count.
993 * This does not protect us against NMI, but disable()
994 * sets the disabled bit in the control field of counter _before_
995 * accessing the counter control register. If a NMI hits, then it will
996 * not restart the counter.
998 void perf_counter_task_sched_out(struct task_struct
*task
,
999 struct task_struct
*next
, int cpu
)
1001 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1002 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1003 struct perf_counter_context
*next_ctx
;
1004 struct perf_counter_context
*parent
;
1005 struct pt_regs
*regs
;
1008 regs
= task_pt_regs(task
);
1009 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1011 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1014 update_context_time(ctx
);
1017 parent
= rcu_dereference(ctx
->parent_ctx
);
1018 next_ctx
= next
->perf_counter_ctxp
;
1019 if (parent
&& next_ctx
&&
1020 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1022 * Looks like the two contexts are clones, so we might be
1023 * able to optimize the context switch. We lock both
1024 * contexts and check that they are clones under the
1025 * lock (including re-checking that neither has been
1026 * uncloned in the meantime). It doesn't matter which
1027 * order we take the locks because no other cpu could
1028 * be trying to lock both of these tasks.
1030 spin_lock(&ctx
->lock
);
1031 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1032 if (context_equiv(ctx
, next_ctx
)) {
1034 * XXX do we need a memory barrier of sorts
1035 * wrt to rcu_dereference() of perf_counter_ctxp
1037 task
->perf_counter_ctxp
= next_ctx
;
1038 next
->perf_counter_ctxp
= ctx
;
1040 next_ctx
->task
= task
;
1043 spin_unlock(&next_ctx
->lock
);
1044 spin_unlock(&ctx
->lock
);
1049 __perf_counter_sched_out(ctx
, cpuctx
);
1050 cpuctx
->task_ctx
= NULL
;
1055 * Called with IRQs disabled
1057 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
1059 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1061 if (!cpuctx
->task_ctx
)
1064 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1067 __perf_counter_sched_out(ctx
, cpuctx
);
1068 cpuctx
->task_ctx
= NULL
;
1072 * Called with IRQs disabled
1074 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1076 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
1080 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
1081 struct perf_cpu_context
*cpuctx
, int cpu
)
1083 struct perf_counter
*counter
;
1086 spin_lock(&ctx
->lock
);
1088 if (likely(!ctx
->nr_counters
))
1091 ctx
->timestamp
= perf_clock();
1096 * First go through the list and put on any pinned groups
1097 * in order to give them the best chance of going on.
1099 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1100 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1101 !counter
->hw_event
.pinned
)
1103 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1106 if (counter
!= counter
->group_leader
)
1107 counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
1109 if (group_can_go_on(counter
, cpuctx
, 1))
1110 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
1114 * If this pinned group hasn't been scheduled,
1115 * put it in error state.
1117 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1118 update_group_times(counter
);
1119 counter
->state
= PERF_COUNTER_STATE_ERROR
;
1123 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1125 * Ignore counters in OFF or ERROR state, and
1126 * ignore pinned counters since we did them already.
1128 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1129 counter
->hw_event
.pinned
)
1133 * Listen to the 'cpu' scheduling filter constraint
1136 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1139 if (counter
!= counter
->group_leader
) {
1140 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
))
1143 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
1144 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
1151 spin_unlock(&ctx
->lock
);
1155 * Called from scheduler to add the counters of the current task
1156 * with interrupts disabled.
1158 * We restore the counter value and then enable it.
1160 * This does not protect us against NMI, but enable()
1161 * sets the enabled bit in the control field of counter _before_
1162 * accessing the counter control register. If a NMI hits, then it will
1163 * keep the counter running.
1165 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
1167 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1168 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1172 if (cpuctx
->task_ctx
== ctx
)
1174 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1175 cpuctx
->task_ctx
= ctx
;
1178 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1180 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
1182 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1185 #define MAX_INTERRUPTS (~0ULL)
1187 static void perf_log_throttle(struct perf_counter
*counter
, int enable
);
1188 static void perf_log_period(struct perf_counter
*counter
, u64 period
);
1190 static void perf_adjust_freq(struct perf_counter_context
*ctx
)
1192 struct perf_counter
*counter
;
1193 u64 interrupts
, irq_period
;
1197 spin_lock(&ctx
->lock
);
1198 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1199 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1202 interrupts
= counter
->hw
.interrupts
;
1203 counter
->hw
.interrupts
= 0;
1205 if (interrupts
== MAX_INTERRUPTS
) {
1206 perf_log_throttle(counter
, 1);
1207 counter
->pmu
->unthrottle(counter
);
1208 interrupts
= 2*sysctl_perf_counter_limit
/HZ
;
1211 if (!counter
->hw_event
.freq
|| !counter
->hw_event
.irq_freq
)
1214 events
= HZ
* interrupts
* counter
->hw
.irq_period
;
1215 period
= div64_u64(events
, counter
->hw_event
.irq_freq
);
1217 delta
= (s64
)(1 + period
- counter
->hw
.irq_period
);
1220 irq_period
= counter
->hw
.irq_period
+ delta
;
1225 perf_log_period(counter
, irq_period
);
1227 counter
->hw
.irq_period
= irq_period
;
1229 spin_unlock(&ctx
->lock
);
1233 * Round-robin a context's counters:
1235 static void rotate_ctx(struct perf_counter_context
*ctx
)
1237 struct perf_counter
*counter
;
1239 if (!ctx
->nr_counters
)
1242 spin_lock(&ctx
->lock
);
1244 * Rotate the first entry last (works just fine for group counters too):
1247 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1248 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1253 spin_unlock(&ctx
->lock
);
1256 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1258 struct perf_cpu_context
*cpuctx
;
1259 struct perf_counter_context
*ctx
;
1261 if (!atomic_read(&nr_counters
))
1264 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1265 ctx
= curr
->perf_counter_ctxp
;
1267 perf_adjust_freq(&cpuctx
->ctx
);
1269 perf_adjust_freq(ctx
);
1271 perf_counter_cpu_sched_out(cpuctx
);
1273 __perf_counter_task_sched_out(ctx
);
1275 rotate_ctx(&cpuctx
->ctx
);
1279 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1281 perf_counter_task_sched_in(curr
, cpu
);
1285 * Cross CPU call to read the hardware counter
1287 static void __read(void *info
)
1289 struct perf_counter
*counter
= info
;
1290 struct perf_counter_context
*ctx
= counter
->ctx
;
1291 unsigned long flags
;
1293 local_irq_save(flags
);
1295 update_context_time(ctx
);
1296 counter
->pmu
->read(counter
);
1297 update_counter_times(counter
);
1298 local_irq_restore(flags
);
1301 static u64
perf_counter_read(struct perf_counter
*counter
)
1304 * If counter is enabled and currently active on a CPU, update the
1305 * value in the counter structure:
1307 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1308 smp_call_function_single(counter
->oncpu
,
1309 __read
, counter
, 1);
1310 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1311 update_counter_times(counter
);
1314 return atomic64_read(&counter
->count
);
1318 * Initialize the perf_counter context in a task_struct:
1321 __perf_counter_init_context(struct perf_counter_context
*ctx
,
1322 struct task_struct
*task
)
1324 memset(ctx
, 0, sizeof(*ctx
));
1325 spin_lock_init(&ctx
->lock
);
1326 mutex_init(&ctx
->mutex
);
1327 INIT_LIST_HEAD(&ctx
->counter_list
);
1328 INIT_LIST_HEAD(&ctx
->event_list
);
1329 atomic_set(&ctx
->refcount
, 1);
1333 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1335 struct perf_counter_context
*parent_ctx
;
1336 struct perf_counter_context
*ctx
;
1337 struct perf_cpu_context
*cpuctx
;
1338 struct task_struct
*task
;
1339 unsigned long flags
;
1343 * If cpu is not a wildcard then this is a percpu counter:
1346 /* Must be root to operate on a CPU counter: */
1347 if (sysctl_perf_counter_priv
&& !capable(CAP_SYS_ADMIN
))
1348 return ERR_PTR(-EACCES
);
1350 if (cpu
< 0 || cpu
> num_possible_cpus())
1351 return ERR_PTR(-EINVAL
);
1354 * We could be clever and allow to attach a counter to an
1355 * offline CPU and activate it when the CPU comes up, but
1358 if (!cpu_isset(cpu
, cpu_online_map
))
1359 return ERR_PTR(-ENODEV
);
1361 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1372 task
= find_task_by_vpid(pid
);
1374 get_task_struct(task
);
1378 return ERR_PTR(-ESRCH
);
1381 * Can't attach counters to a dying task.
1384 if (task
->flags
& PF_EXITING
)
1387 /* Reuse ptrace permission checks for now. */
1389 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1393 ctx
= perf_lock_task_context(task
, &flags
);
1395 parent_ctx
= ctx
->parent_ctx
;
1397 put_ctx(parent_ctx
);
1398 ctx
->parent_ctx
= NULL
; /* no longer a clone */
1401 * Get an extra reference before dropping the lock so that
1402 * this context won't get freed if the task exits.
1405 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1409 ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
1413 __perf_counter_init_context(ctx
, task
);
1415 if (cmpxchg(&task
->perf_counter_ctxp
, NULL
, ctx
)) {
1417 * We raced with some other task; use
1418 * the context they set.
1423 get_task_struct(task
);
1426 put_task_struct(task
);
1430 put_task_struct(task
);
1431 return ERR_PTR(err
);
1434 static void free_counter_rcu(struct rcu_head
*head
)
1436 struct perf_counter
*counter
;
1438 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1442 static void perf_pending_sync(struct perf_counter
*counter
);
1444 static void free_counter(struct perf_counter
*counter
)
1446 perf_pending_sync(counter
);
1448 atomic_dec(&nr_counters
);
1449 if (counter
->hw_event
.mmap
)
1450 atomic_dec(&nr_mmap_tracking
);
1451 if (counter
->hw_event
.munmap
)
1452 atomic_dec(&nr_munmap_tracking
);
1453 if (counter
->hw_event
.comm
)
1454 atomic_dec(&nr_comm_tracking
);
1456 if (counter
->destroy
)
1457 counter
->destroy(counter
);
1459 put_ctx(counter
->ctx
);
1460 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1464 * Called when the last reference to the file is gone.
1466 static int perf_release(struct inode
*inode
, struct file
*file
)
1468 struct perf_counter
*counter
= file
->private_data
;
1469 struct perf_counter_context
*ctx
= counter
->ctx
;
1471 file
->private_data
= NULL
;
1473 WARN_ON_ONCE(ctx
->parent_ctx
);
1474 mutex_lock(&ctx
->mutex
);
1475 perf_counter_remove_from_context(counter
);
1476 mutex_unlock(&ctx
->mutex
);
1478 mutex_lock(&counter
->owner
->perf_counter_mutex
);
1479 list_del_init(&counter
->owner_entry
);
1480 mutex_unlock(&counter
->owner
->perf_counter_mutex
);
1481 put_task_struct(counter
->owner
);
1483 free_counter(counter
);
1489 * Read the performance counter - simple non blocking version for now
1492 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1498 * Return end-of-file for a read on a counter that is in
1499 * error state (i.e. because it was pinned but it couldn't be
1500 * scheduled on to the CPU at some point).
1502 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1505 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1506 mutex_lock(&counter
->child_mutex
);
1507 values
[0] = perf_counter_read(counter
);
1509 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1510 values
[n
++] = counter
->total_time_enabled
+
1511 atomic64_read(&counter
->child_total_time_enabled
);
1512 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1513 values
[n
++] = counter
->total_time_running
+
1514 atomic64_read(&counter
->child_total_time_running
);
1515 mutex_unlock(&counter
->child_mutex
);
1517 if (count
< n
* sizeof(u64
))
1519 count
= n
* sizeof(u64
);
1521 if (copy_to_user(buf
, values
, count
))
1528 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1530 struct perf_counter
*counter
= file
->private_data
;
1532 return perf_read_hw(counter
, buf
, count
);
1535 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1537 struct perf_counter
*counter
= file
->private_data
;
1538 struct perf_mmap_data
*data
;
1539 unsigned int events
= POLL_HUP
;
1542 data
= rcu_dereference(counter
->data
);
1544 events
= atomic_xchg(&data
->poll
, 0);
1547 poll_wait(file
, &counter
->waitq
, wait
);
1552 static void perf_counter_reset(struct perf_counter
*counter
)
1554 (void)perf_counter_read(counter
);
1555 atomic64_set(&counter
->count
, 0);
1556 perf_counter_update_userpage(counter
);
1559 static void perf_counter_for_each_sibling(struct perf_counter
*counter
,
1560 void (*func
)(struct perf_counter
*))
1562 struct perf_counter_context
*ctx
= counter
->ctx
;
1563 struct perf_counter
*sibling
;
1565 WARN_ON_ONCE(ctx
->parent_ctx
);
1566 mutex_lock(&ctx
->mutex
);
1567 counter
= counter
->group_leader
;
1570 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1572 mutex_unlock(&ctx
->mutex
);
1576 * Holding the top-level counter's child_mutex means that any
1577 * descendant process that has inherited this counter will block
1578 * in sync_child_counter if it goes to exit, thus satisfying the
1579 * task existence requirements of perf_counter_enable/disable.
1581 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1582 void (*func
)(struct perf_counter
*))
1584 struct perf_counter
*child
;
1586 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1587 mutex_lock(&counter
->child_mutex
);
1589 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1591 mutex_unlock(&counter
->child_mutex
);
1594 static void perf_counter_for_each(struct perf_counter
*counter
,
1595 void (*func
)(struct perf_counter
*))
1597 struct perf_counter
*child
;
1599 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1600 mutex_lock(&counter
->child_mutex
);
1601 perf_counter_for_each_sibling(counter
, func
);
1602 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1603 perf_counter_for_each_sibling(child
, func
);
1604 mutex_unlock(&counter
->child_mutex
);
1607 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1609 struct perf_counter
*counter
= file
->private_data
;
1610 void (*func
)(struct perf_counter
*);
1614 case PERF_COUNTER_IOC_ENABLE
:
1615 func
= perf_counter_enable
;
1617 case PERF_COUNTER_IOC_DISABLE
:
1618 func
= perf_counter_disable
;
1620 case PERF_COUNTER_IOC_RESET
:
1621 func
= perf_counter_reset
;
1624 case PERF_COUNTER_IOC_REFRESH
:
1625 return perf_counter_refresh(counter
, arg
);
1630 if (flags
& PERF_IOC_FLAG_GROUP
)
1631 perf_counter_for_each(counter
, func
);
1633 perf_counter_for_each_child(counter
, func
);
1638 int perf_counter_task_enable(void)
1640 struct perf_counter
*counter
;
1642 mutex_lock(¤t
->perf_counter_mutex
);
1643 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1644 perf_counter_for_each_child(counter
, perf_counter_enable
);
1645 mutex_unlock(¤t
->perf_counter_mutex
);
1650 int perf_counter_task_disable(void)
1652 struct perf_counter
*counter
;
1654 mutex_lock(¤t
->perf_counter_mutex
);
1655 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1656 perf_counter_for_each_child(counter
, perf_counter_disable
);
1657 mutex_unlock(¤t
->perf_counter_mutex
);
1663 * Callers need to ensure there can be no nesting of this function, otherwise
1664 * the seqlock logic goes bad. We can not serialize this because the arch
1665 * code calls this from NMI context.
1667 void perf_counter_update_userpage(struct perf_counter
*counter
)
1669 struct perf_counter_mmap_page
*userpg
;
1670 struct perf_mmap_data
*data
;
1673 data
= rcu_dereference(counter
->data
);
1677 userpg
= data
->user_page
;
1680 * Disable preemption so as to not let the corresponding user-space
1681 * spin too long if we get preempted.
1686 userpg
->index
= counter
->hw
.idx
;
1687 userpg
->offset
= atomic64_read(&counter
->count
);
1688 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1689 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1698 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1700 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1701 struct perf_mmap_data
*data
;
1702 int ret
= VM_FAULT_SIGBUS
;
1705 data
= rcu_dereference(counter
->data
);
1709 if (vmf
->pgoff
== 0) {
1710 vmf
->page
= virt_to_page(data
->user_page
);
1712 int nr
= vmf
->pgoff
- 1;
1714 if ((unsigned)nr
> data
->nr_pages
)
1717 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1719 get_page(vmf
->page
);
1727 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1729 struct perf_mmap_data
*data
;
1733 WARN_ON(atomic_read(&counter
->mmap_count
));
1735 size
= sizeof(struct perf_mmap_data
);
1736 size
+= nr_pages
* sizeof(void *);
1738 data
= kzalloc(size
, GFP_KERNEL
);
1742 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1743 if (!data
->user_page
)
1744 goto fail_user_page
;
1746 for (i
= 0; i
< nr_pages
; i
++) {
1747 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1748 if (!data
->data_pages
[i
])
1749 goto fail_data_pages
;
1752 data
->nr_pages
= nr_pages
;
1753 atomic_set(&data
->lock
, -1);
1755 rcu_assign_pointer(counter
->data
, data
);
1760 for (i
--; i
>= 0; i
--)
1761 free_page((unsigned long)data
->data_pages
[i
]);
1763 free_page((unsigned long)data
->user_page
);
1772 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1774 struct perf_mmap_data
*data
;
1777 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
1779 free_page((unsigned long)data
->user_page
);
1780 for (i
= 0; i
< data
->nr_pages
; i
++)
1781 free_page((unsigned long)data
->data_pages
[i
]);
1785 static void perf_mmap_data_free(struct perf_counter
*counter
)
1787 struct perf_mmap_data
*data
= counter
->data
;
1789 WARN_ON(atomic_read(&counter
->mmap_count
));
1791 rcu_assign_pointer(counter
->data
, NULL
);
1792 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1795 static void perf_mmap_open(struct vm_area_struct
*vma
)
1797 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1799 atomic_inc(&counter
->mmap_count
);
1802 static void perf_mmap_close(struct vm_area_struct
*vma
)
1804 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1806 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1807 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
, &counter
->mmap_mutex
)) {
1808 struct user_struct
*user
= current_user();
1810 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
1811 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
1812 perf_mmap_data_free(counter
);
1813 mutex_unlock(&counter
->mmap_mutex
);
1817 static struct vm_operations_struct perf_mmap_vmops
= {
1818 .open
= perf_mmap_open
,
1819 .close
= perf_mmap_close
,
1820 .fault
= perf_mmap_fault
,
1823 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1825 struct perf_counter
*counter
= file
->private_data
;
1826 unsigned long user_locked
, user_lock_limit
;
1827 struct user_struct
*user
= current_user();
1828 unsigned long locked
, lock_limit
;
1829 unsigned long vma_size
;
1830 unsigned long nr_pages
;
1831 long user_extra
, extra
;
1834 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1837 vma_size
= vma
->vm_end
- vma
->vm_start
;
1838 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1841 * If we have data pages ensure they're a power-of-two number, so we
1842 * can do bitmasks instead of modulo.
1844 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1847 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1850 if (vma
->vm_pgoff
!= 0)
1853 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1854 mutex_lock(&counter
->mmap_mutex
);
1855 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1856 if (nr_pages
!= counter
->data
->nr_pages
)
1861 user_extra
= nr_pages
+ 1;
1862 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
1865 * Increase the limit linearly with more CPUs:
1867 user_lock_limit
*= num_online_cpus();
1869 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
1872 if (user_locked
> user_lock_limit
)
1873 extra
= user_locked
- user_lock_limit
;
1875 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1876 lock_limit
>>= PAGE_SHIFT
;
1877 locked
= vma
->vm_mm
->locked_vm
+ extra
;
1879 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
1884 WARN_ON(counter
->data
);
1885 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1889 atomic_set(&counter
->mmap_count
, 1);
1890 atomic_long_add(user_extra
, &user
->locked_vm
);
1891 vma
->vm_mm
->locked_vm
+= extra
;
1892 counter
->data
->nr_locked
= extra
;
1894 mutex_unlock(&counter
->mmap_mutex
);
1896 vma
->vm_flags
&= ~VM_MAYWRITE
;
1897 vma
->vm_flags
|= VM_RESERVED
;
1898 vma
->vm_ops
= &perf_mmap_vmops
;
1903 static int perf_fasync(int fd
, struct file
*filp
, int on
)
1905 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1906 struct perf_counter
*counter
= filp
->private_data
;
1909 mutex_lock(&inode
->i_mutex
);
1910 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
1911 mutex_unlock(&inode
->i_mutex
);
1919 static const struct file_operations perf_fops
= {
1920 .release
= perf_release
,
1923 .unlocked_ioctl
= perf_ioctl
,
1924 .compat_ioctl
= perf_ioctl
,
1926 .fasync
= perf_fasync
,
1930 * Perf counter wakeup
1932 * If there's data, ensure we set the poll() state and publish everything
1933 * to user-space before waking everybody up.
1936 void perf_counter_wakeup(struct perf_counter
*counter
)
1938 wake_up_all(&counter
->waitq
);
1940 if (counter
->pending_kill
) {
1941 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
1942 counter
->pending_kill
= 0;
1949 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1951 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1952 * single linked list and use cmpxchg() to add entries lockless.
1955 static void perf_pending_counter(struct perf_pending_entry
*entry
)
1957 struct perf_counter
*counter
= container_of(entry
,
1958 struct perf_counter
, pending
);
1960 if (counter
->pending_disable
) {
1961 counter
->pending_disable
= 0;
1962 perf_counter_disable(counter
);
1965 if (counter
->pending_wakeup
) {
1966 counter
->pending_wakeup
= 0;
1967 perf_counter_wakeup(counter
);
1971 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1973 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
1977 static void perf_pending_queue(struct perf_pending_entry
*entry
,
1978 void (*func
)(struct perf_pending_entry
*))
1980 struct perf_pending_entry
**head
;
1982 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
1987 head
= &get_cpu_var(perf_pending_head
);
1990 entry
->next
= *head
;
1991 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
1993 set_perf_counter_pending();
1995 put_cpu_var(perf_pending_head
);
1998 static int __perf_pending_run(void)
2000 struct perf_pending_entry
*list
;
2003 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2004 while (list
!= PENDING_TAIL
) {
2005 void (*func
)(struct perf_pending_entry
*);
2006 struct perf_pending_entry
*entry
= list
;
2013 * Ensure we observe the unqueue before we issue the wakeup,
2014 * so that we won't be waiting forever.
2015 * -- see perf_not_pending().
2026 static inline int perf_not_pending(struct perf_counter
*counter
)
2029 * If we flush on whatever cpu we run, there is a chance we don't
2033 __perf_pending_run();
2037 * Ensure we see the proper queue state before going to sleep
2038 * so that we do not miss the wakeup. -- see perf_pending_handle()
2041 return counter
->pending
.next
== NULL
;
2044 static void perf_pending_sync(struct perf_counter
*counter
)
2046 wait_event(counter
->waitq
, perf_not_pending(counter
));
2049 void perf_counter_do_pending(void)
2051 __perf_pending_run();
2055 * Callchain support -- arch specific
2058 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2067 struct perf_output_handle
{
2068 struct perf_counter
*counter
;
2069 struct perf_mmap_data
*data
;
2070 unsigned int offset
;
2075 unsigned long flags
;
2078 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2080 atomic_set(&handle
->data
->poll
, POLL_IN
);
2083 handle
->counter
->pending_wakeup
= 1;
2084 perf_pending_queue(&handle
->counter
->pending
,
2085 perf_pending_counter
);
2087 perf_counter_wakeup(handle
->counter
);
2091 * Curious locking construct.
2093 * We need to ensure a later event doesn't publish a head when a former
2094 * event isn't done writing. However since we need to deal with NMIs we
2095 * cannot fully serialize things.
2097 * What we do is serialize between CPUs so we only have to deal with NMI
2098 * nesting on a single CPU.
2100 * We only publish the head (and generate a wakeup) when the outer-most
2103 static void perf_output_lock(struct perf_output_handle
*handle
)
2105 struct perf_mmap_data
*data
= handle
->data
;
2110 local_irq_save(handle
->flags
);
2111 cpu
= smp_processor_id();
2113 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
2116 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2122 static void perf_output_unlock(struct perf_output_handle
*handle
)
2124 struct perf_mmap_data
*data
= handle
->data
;
2127 data
->done_head
= data
->head
;
2129 if (!handle
->locked
)
2134 * The xchg implies a full barrier that ensures all writes are done
2135 * before we publish the new head, matched by a rmb() in userspace when
2136 * reading this position.
2138 while ((head
= atomic_xchg(&data
->done_head
, 0)))
2139 data
->user_page
->data_head
= head
;
2142 * NMI can happen here, which means we can miss a done_head update.
2145 cpu
= atomic_xchg(&data
->lock
, -1);
2146 WARN_ON_ONCE(cpu
!= smp_processor_id());
2149 * Therefore we have to validate we did not indeed do so.
2151 if (unlikely(atomic_read(&data
->done_head
))) {
2153 * Since we had it locked, we can lock it again.
2155 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2161 if (atomic_xchg(&data
->wakeup
, 0))
2162 perf_output_wakeup(handle
);
2164 local_irq_restore(handle
->flags
);
2167 static int perf_output_begin(struct perf_output_handle
*handle
,
2168 struct perf_counter
*counter
, unsigned int size
,
2169 int nmi
, int overflow
)
2171 struct perf_mmap_data
*data
;
2172 unsigned int offset
, head
;
2175 * For inherited counters we send all the output towards the parent.
2177 if (counter
->parent
)
2178 counter
= counter
->parent
;
2181 data
= rcu_dereference(counter
->data
);
2185 handle
->data
= data
;
2186 handle
->counter
= counter
;
2188 handle
->overflow
= overflow
;
2190 if (!data
->nr_pages
)
2193 perf_output_lock(handle
);
2196 offset
= head
= atomic_read(&data
->head
);
2198 } while (atomic_cmpxchg(&data
->head
, offset
, head
) != offset
);
2200 handle
->offset
= offset
;
2201 handle
->head
= head
;
2203 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
2204 atomic_set(&data
->wakeup
, 1);
2209 perf_output_wakeup(handle
);
2216 static void perf_output_copy(struct perf_output_handle
*handle
,
2217 void *buf
, unsigned int len
)
2219 unsigned int pages_mask
;
2220 unsigned int offset
;
2224 offset
= handle
->offset
;
2225 pages_mask
= handle
->data
->nr_pages
- 1;
2226 pages
= handle
->data
->data_pages
;
2229 unsigned int page_offset
;
2232 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2233 page_offset
= offset
& (PAGE_SIZE
- 1);
2234 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
2236 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2243 handle
->offset
= offset
;
2246 * Check we didn't copy past our reservation window, taking the
2247 * possible unsigned int wrap into account.
2249 WARN_ON_ONCE(((int)(handle
->head
- handle
->offset
)) < 0);
2252 #define perf_output_put(handle, x) \
2253 perf_output_copy((handle), &(x), sizeof(x))
2255 static void perf_output_end(struct perf_output_handle
*handle
)
2257 struct perf_counter
*counter
= handle
->counter
;
2258 struct perf_mmap_data
*data
= handle
->data
;
2260 int wakeup_events
= counter
->hw_event
.wakeup_events
;
2262 if (handle
->overflow
&& wakeup_events
) {
2263 int events
= atomic_inc_return(&data
->events
);
2264 if (events
>= wakeup_events
) {
2265 atomic_sub(wakeup_events
, &data
->events
);
2266 atomic_set(&data
->wakeup
, 1);
2270 perf_output_unlock(handle
);
2274 static void perf_counter_output(struct perf_counter
*counter
,
2275 int nmi
, struct pt_regs
*regs
, u64 addr
)
2278 u64 record_type
= counter
->hw_event
.record_type
;
2279 struct perf_output_handle handle
;
2280 struct perf_event_header header
;
2289 struct perf_callchain_entry
*callchain
= NULL
;
2290 int callchain_size
= 0;
2297 header
.size
= sizeof(header
);
2299 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
2300 header
.misc
|= perf_misc_flags(regs
);
2302 if (record_type
& PERF_RECORD_IP
) {
2303 ip
= perf_instruction_pointer(regs
);
2304 header
.type
|= PERF_RECORD_IP
;
2305 header
.size
+= sizeof(ip
);
2308 if (record_type
& PERF_RECORD_TID
) {
2309 /* namespace issues */
2310 tid_entry
.pid
= current
->group_leader
->pid
;
2311 tid_entry
.tid
= current
->pid
;
2313 header
.type
|= PERF_RECORD_TID
;
2314 header
.size
+= sizeof(tid_entry
);
2317 if (record_type
& PERF_RECORD_TIME
) {
2319 * Maybe do better on x86 and provide cpu_clock_nmi()
2321 time
= sched_clock();
2323 header
.type
|= PERF_RECORD_TIME
;
2324 header
.size
+= sizeof(u64
);
2327 if (record_type
& PERF_RECORD_ADDR
) {
2328 header
.type
|= PERF_RECORD_ADDR
;
2329 header
.size
+= sizeof(u64
);
2332 if (record_type
& PERF_RECORD_CONFIG
) {
2333 header
.type
|= PERF_RECORD_CONFIG
;
2334 header
.size
+= sizeof(u64
);
2337 if (record_type
& PERF_RECORD_CPU
) {
2338 header
.type
|= PERF_RECORD_CPU
;
2339 header
.size
+= sizeof(cpu_entry
);
2341 cpu_entry
.cpu
= raw_smp_processor_id();
2344 if (record_type
& PERF_RECORD_GROUP
) {
2345 header
.type
|= PERF_RECORD_GROUP
;
2346 header
.size
+= sizeof(u64
) +
2347 counter
->nr_siblings
* sizeof(group_entry
);
2350 if (record_type
& PERF_RECORD_CALLCHAIN
) {
2351 callchain
= perf_callchain(regs
);
2354 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2356 header
.type
|= PERF_RECORD_CALLCHAIN
;
2357 header
.size
+= callchain_size
;
2361 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2365 perf_output_put(&handle
, header
);
2367 if (record_type
& PERF_RECORD_IP
)
2368 perf_output_put(&handle
, ip
);
2370 if (record_type
& PERF_RECORD_TID
)
2371 perf_output_put(&handle
, tid_entry
);
2373 if (record_type
& PERF_RECORD_TIME
)
2374 perf_output_put(&handle
, time
);
2376 if (record_type
& PERF_RECORD_ADDR
)
2377 perf_output_put(&handle
, addr
);
2379 if (record_type
& PERF_RECORD_CONFIG
)
2380 perf_output_put(&handle
, counter
->hw_event
.config
);
2382 if (record_type
& PERF_RECORD_CPU
)
2383 perf_output_put(&handle
, cpu_entry
);
2386 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2388 if (record_type
& PERF_RECORD_GROUP
) {
2389 struct perf_counter
*leader
, *sub
;
2390 u64 nr
= counter
->nr_siblings
;
2392 perf_output_put(&handle
, nr
);
2394 leader
= counter
->group_leader
;
2395 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2397 sub
->pmu
->read(sub
);
2399 group_entry
.event
= sub
->hw_event
.config
;
2400 group_entry
.counter
= atomic64_read(&sub
->count
);
2402 perf_output_put(&handle
, group_entry
);
2407 perf_output_copy(&handle
, callchain
, callchain_size
);
2409 perf_output_end(&handle
);
2416 struct perf_comm_event
{
2417 struct task_struct
*task
;
2422 struct perf_event_header header
;
2429 static void perf_counter_comm_output(struct perf_counter
*counter
,
2430 struct perf_comm_event
*comm_event
)
2432 struct perf_output_handle handle
;
2433 int size
= comm_event
->event
.header
.size
;
2434 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2439 perf_output_put(&handle
, comm_event
->event
);
2440 perf_output_copy(&handle
, comm_event
->comm
,
2441 comm_event
->comm_size
);
2442 perf_output_end(&handle
);
2445 static int perf_counter_comm_match(struct perf_counter
*counter
,
2446 struct perf_comm_event
*comm_event
)
2448 if (counter
->hw_event
.comm
&&
2449 comm_event
->event
.header
.type
== PERF_EVENT_COMM
)
2455 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2456 struct perf_comm_event
*comm_event
)
2458 struct perf_counter
*counter
;
2460 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2464 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2465 if (perf_counter_comm_match(counter
, comm_event
))
2466 perf_counter_comm_output(counter
, comm_event
);
2471 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2473 struct perf_cpu_context
*cpuctx
;
2474 struct perf_counter_context
*ctx
;
2476 char *comm
= comm_event
->task
->comm
;
2478 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2480 comm_event
->comm
= comm
;
2481 comm_event
->comm_size
= size
;
2483 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2485 cpuctx
= &get_cpu_var(perf_cpu_context
);
2486 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2487 put_cpu_var(perf_cpu_context
);
2491 * doesn't really matter which of the child contexts the
2492 * events ends up in.
2494 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2496 perf_counter_comm_ctx(ctx
, comm_event
);
2500 void perf_counter_comm(struct task_struct
*task
)
2502 struct perf_comm_event comm_event
;
2504 if (!atomic_read(&nr_comm_tracking
))
2507 comm_event
= (struct perf_comm_event
){
2510 .header
= { .type
= PERF_EVENT_COMM
, },
2511 .pid
= task
->group_leader
->pid
,
2516 perf_counter_comm_event(&comm_event
);
2523 struct perf_mmap_event
{
2529 struct perf_event_header header
;
2539 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2540 struct perf_mmap_event
*mmap_event
)
2542 struct perf_output_handle handle
;
2543 int size
= mmap_event
->event
.header
.size
;
2544 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2549 perf_output_put(&handle
, mmap_event
->event
);
2550 perf_output_copy(&handle
, mmap_event
->file_name
,
2551 mmap_event
->file_size
);
2552 perf_output_end(&handle
);
2555 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2556 struct perf_mmap_event
*mmap_event
)
2558 if (counter
->hw_event
.mmap
&&
2559 mmap_event
->event
.header
.type
== PERF_EVENT_MMAP
)
2562 if (counter
->hw_event
.munmap
&&
2563 mmap_event
->event
.header
.type
== PERF_EVENT_MUNMAP
)
2569 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2570 struct perf_mmap_event
*mmap_event
)
2572 struct perf_counter
*counter
;
2574 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2578 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2579 if (perf_counter_mmap_match(counter
, mmap_event
))
2580 perf_counter_mmap_output(counter
, mmap_event
);
2585 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2587 struct perf_cpu_context
*cpuctx
;
2588 struct perf_counter_context
*ctx
;
2589 struct file
*file
= mmap_event
->file
;
2596 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2598 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2601 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2603 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2607 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2612 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2614 mmap_event
->file_name
= name
;
2615 mmap_event
->file_size
= size
;
2617 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2619 cpuctx
= &get_cpu_var(perf_cpu_context
);
2620 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2621 put_cpu_var(perf_cpu_context
);
2625 * doesn't really matter which of the child contexts the
2626 * events ends up in.
2628 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2630 perf_counter_mmap_ctx(ctx
, mmap_event
);
2636 void perf_counter_mmap(unsigned long addr
, unsigned long len
,
2637 unsigned long pgoff
, struct file
*file
)
2639 struct perf_mmap_event mmap_event
;
2641 if (!atomic_read(&nr_mmap_tracking
))
2644 mmap_event
= (struct perf_mmap_event
){
2647 .header
= { .type
= PERF_EVENT_MMAP
, },
2648 .pid
= current
->group_leader
->pid
,
2649 .tid
= current
->pid
,
2656 perf_counter_mmap_event(&mmap_event
);
2659 void perf_counter_munmap(unsigned long addr
, unsigned long len
,
2660 unsigned long pgoff
, struct file
*file
)
2662 struct perf_mmap_event mmap_event
;
2664 if (!atomic_read(&nr_munmap_tracking
))
2667 mmap_event
= (struct perf_mmap_event
){
2670 .header
= { .type
= PERF_EVENT_MUNMAP
, },
2671 .pid
= current
->group_leader
->pid
,
2672 .tid
= current
->pid
,
2679 perf_counter_mmap_event(&mmap_event
);
2683 * Log irq_period changes so that analyzing tools can re-normalize the
2687 static void perf_log_period(struct perf_counter
*counter
, u64 period
)
2689 struct perf_output_handle handle
;
2693 struct perf_event_header header
;
2698 .type
= PERF_EVENT_PERIOD
,
2700 .size
= sizeof(freq_event
),
2702 .time
= sched_clock(),
2706 if (counter
->hw
.irq_period
== period
)
2709 ret
= perf_output_begin(&handle
, counter
, sizeof(freq_event
), 0, 0);
2713 perf_output_put(&handle
, freq_event
);
2714 perf_output_end(&handle
);
2718 * IRQ throttle logging
2721 static void perf_log_throttle(struct perf_counter
*counter
, int enable
)
2723 struct perf_output_handle handle
;
2727 struct perf_event_header header
;
2729 } throttle_event
= {
2731 .type
= PERF_EVENT_THROTTLE
+ 1,
2733 .size
= sizeof(throttle_event
),
2735 .time
= sched_clock(),
2738 ret
= perf_output_begin(&handle
, counter
, sizeof(throttle_event
), 1, 0);
2742 perf_output_put(&handle
, throttle_event
);
2743 perf_output_end(&handle
);
2747 * Generic counter overflow handling.
2750 int perf_counter_overflow(struct perf_counter
*counter
,
2751 int nmi
, struct pt_regs
*regs
, u64 addr
)
2753 int events
= atomic_read(&counter
->event_limit
);
2754 int throttle
= counter
->pmu
->unthrottle
!= NULL
;
2758 counter
->hw
.interrupts
++;
2759 } else if (counter
->hw
.interrupts
!= MAX_INTERRUPTS
) {
2760 counter
->hw
.interrupts
++;
2761 if (HZ
*counter
->hw
.interrupts
> (u64
)sysctl_perf_counter_limit
) {
2762 counter
->hw
.interrupts
= MAX_INTERRUPTS
;
2763 perf_log_throttle(counter
, 0);
2769 * XXX event_limit might not quite work as expected on inherited
2773 counter
->pending_kill
= POLL_IN
;
2774 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
2776 counter
->pending_kill
= POLL_HUP
;
2778 counter
->pending_disable
= 1;
2779 perf_pending_queue(&counter
->pending
,
2780 perf_pending_counter
);
2782 perf_counter_disable(counter
);
2785 perf_counter_output(counter
, nmi
, regs
, addr
);
2790 * Generic software counter infrastructure
2793 static void perf_swcounter_update(struct perf_counter
*counter
)
2795 struct hw_perf_counter
*hwc
= &counter
->hw
;
2800 prev
= atomic64_read(&hwc
->prev_count
);
2801 now
= atomic64_read(&hwc
->count
);
2802 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2807 atomic64_add(delta
, &counter
->count
);
2808 atomic64_sub(delta
, &hwc
->period_left
);
2811 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2813 struct hw_perf_counter
*hwc
= &counter
->hw
;
2814 s64 left
= atomic64_read(&hwc
->period_left
);
2815 s64 period
= hwc
->irq_period
;
2817 if (unlikely(left
<= -period
)) {
2819 atomic64_set(&hwc
->period_left
, left
);
2822 if (unlikely(left
<= 0)) {
2824 atomic64_add(period
, &hwc
->period_left
);
2827 atomic64_set(&hwc
->prev_count
, -left
);
2828 atomic64_set(&hwc
->count
, -left
);
2831 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2833 enum hrtimer_restart ret
= HRTIMER_RESTART
;
2834 struct perf_counter
*counter
;
2835 struct pt_regs
*regs
;
2838 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
2839 counter
->pmu
->read(counter
);
2841 regs
= get_irq_regs();
2843 * In case we exclude kernel IPs or are somehow not in interrupt
2844 * context, provide the next best thing, the user IP.
2846 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
2847 !counter
->hw_event
.exclude_user
)
2848 regs
= task_pt_regs(current
);
2851 if (perf_counter_overflow(counter
, 0, regs
, 0))
2852 ret
= HRTIMER_NORESTART
;
2855 period
= max_t(u64
, 10000, counter
->hw
.irq_period
);
2856 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
2861 static void perf_swcounter_overflow(struct perf_counter
*counter
,
2862 int nmi
, struct pt_regs
*regs
, u64 addr
)
2864 perf_swcounter_update(counter
);
2865 perf_swcounter_set_period(counter
);
2866 if (perf_counter_overflow(counter
, nmi
, regs
, addr
))
2867 /* soft-disable the counter */
2872 static int perf_swcounter_is_counting(struct perf_counter
*counter
)
2874 struct perf_counter_context
*ctx
;
2875 unsigned long flags
;
2878 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
2881 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
)
2885 * If the counter is inactive, it could be just because
2886 * its task is scheduled out, or because it's in a group
2887 * which could not go on the PMU. We want to count in
2888 * the first case but not the second. If the context is
2889 * currently active then an inactive software counter must
2890 * be the second case. If it's not currently active then
2891 * we need to know whether the counter was active when the
2892 * context was last active, which we can determine by
2893 * comparing counter->tstamp_stopped with ctx->time.
2895 * We are within an RCU read-side critical section,
2896 * which protects the existence of *ctx.
2899 spin_lock_irqsave(&ctx
->lock
, flags
);
2901 /* Re-check state now we have the lock */
2902 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
||
2903 counter
->ctx
->is_active
||
2904 counter
->tstamp_stopped
< ctx
->time
)
2906 spin_unlock_irqrestore(&ctx
->lock
, flags
);
2910 static int perf_swcounter_match(struct perf_counter
*counter
,
2911 enum perf_event_types type
,
2912 u32 event
, struct pt_regs
*regs
)
2916 event_config
= ((u64
) type
<< PERF_COUNTER_TYPE_SHIFT
) | event
;
2918 if (!perf_swcounter_is_counting(counter
))
2921 if (counter
->hw_event
.config
!= event_config
)
2924 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
2927 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
2933 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
2934 int nmi
, struct pt_regs
*regs
, u64 addr
)
2936 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
2938 if (counter
->hw
.irq_period
&& !neg
)
2939 perf_swcounter_overflow(counter
, nmi
, regs
, addr
);
2942 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
2943 enum perf_event_types type
, u32 event
,
2944 u64 nr
, int nmi
, struct pt_regs
*regs
,
2947 struct perf_counter
*counter
;
2949 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2953 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2954 if (perf_swcounter_match(counter
, type
, event
, regs
))
2955 perf_swcounter_add(counter
, nr
, nmi
, regs
, addr
);
2960 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
2963 return &cpuctx
->recursion
[3];
2966 return &cpuctx
->recursion
[2];
2969 return &cpuctx
->recursion
[1];
2971 return &cpuctx
->recursion
[0];
2974 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
2975 u64 nr
, int nmi
, struct pt_regs
*regs
,
2978 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
2979 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
2980 struct perf_counter_context
*ctx
;
2988 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
2989 nr
, nmi
, regs
, addr
);
2992 * doesn't really matter which of the child contexts the
2993 * events ends up in.
2995 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2997 perf_swcounter_ctx_event(ctx
, type
, event
, nr
, nmi
, regs
, addr
);
3004 put_cpu_var(perf_cpu_context
);
3008 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
3010 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
, addr
);
3013 static void perf_swcounter_read(struct perf_counter
*counter
)
3015 perf_swcounter_update(counter
);
3018 static int perf_swcounter_enable(struct perf_counter
*counter
)
3020 perf_swcounter_set_period(counter
);
3024 static void perf_swcounter_disable(struct perf_counter
*counter
)
3026 perf_swcounter_update(counter
);
3029 static const struct pmu perf_ops_generic
= {
3030 .enable
= perf_swcounter_enable
,
3031 .disable
= perf_swcounter_disable
,
3032 .read
= perf_swcounter_read
,
3036 * Software counter: cpu wall time clock
3039 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
3041 int cpu
= raw_smp_processor_id();
3045 now
= cpu_clock(cpu
);
3046 prev
= atomic64_read(&counter
->hw
.prev_count
);
3047 atomic64_set(&counter
->hw
.prev_count
, now
);
3048 atomic64_add(now
- prev
, &counter
->count
);
3051 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
3053 struct hw_perf_counter
*hwc
= &counter
->hw
;
3054 int cpu
= raw_smp_processor_id();
3056 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
3057 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3058 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3059 if (hwc
->irq_period
) {
3060 u64 period
= max_t(u64
, 10000, hwc
->irq_period
);
3061 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3062 ns_to_ktime(period
), 0,
3063 HRTIMER_MODE_REL
, 0);
3069 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
3071 if (counter
->hw
.irq_period
)
3072 hrtimer_cancel(&counter
->hw
.hrtimer
);
3073 cpu_clock_perf_counter_update(counter
);
3076 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
3078 cpu_clock_perf_counter_update(counter
);
3081 static const struct pmu perf_ops_cpu_clock
= {
3082 .enable
= cpu_clock_perf_counter_enable
,
3083 .disable
= cpu_clock_perf_counter_disable
,
3084 .read
= cpu_clock_perf_counter_read
,
3088 * Software counter: task time clock
3091 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
3096 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
3098 atomic64_add(delta
, &counter
->count
);
3101 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
3103 struct hw_perf_counter
*hwc
= &counter
->hw
;
3106 now
= counter
->ctx
->time
;
3108 atomic64_set(&hwc
->prev_count
, now
);
3109 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3110 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3111 if (hwc
->irq_period
) {
3112 u64 period
= max_t(u64
, 10000, hwc
->irq_period
);
3113 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3114 ns_to_ktime(period
), 0,
3115 HRTIMER_MODE_REL
, 0);
3121 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
3123 if (counter
->hw
.irq_period
)
3124 hrtimer_cancel(&counter
->hw
.hrtimer
);
3125 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
3129 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
3134 update_context_time(counter
->ctx
);
3135 time
= counter
->ctx
->time
;
3137 u64 now
= perf_clock();
3138 u64 delta
= now
- counter
->ctx
->timestamp
;
3139 time
= counter
->ctx
->time
+ delta
;
3142 task_clock_perf_counter_update(counter
, time
);
3145 static const struct pmu perf_ops_task_clock
= {
3146 .enable
= task_clock_perf_counter_enable
,
3147 .disable
= task_clock_perf_counter_disable
,
3148 .read
= task_clock_perf_counter_read
,
3152 * Software counter: cpu migrations
3155 static inline u64
get_cpu_migrations(struct perf_counter
*counter
)
3157 struct task_struct
*curr
= counter
->ctx
->task
;
3160 return curr
->se
.nr_migrations
;
3161 return cpu_nr_migrations(smp_processor_id());
3164 static void cpu_migrations_perf_counter_update(struct perf_counter
*counter
)
3169 prev
= atomic64_read(&counter
->hw
.prev_count
);
3170 now
= get_cpu_migrations(counter
);
3172 atomic64_set(&counter
->hw
.prev_count
, now
);
3176 atomic64_add(delta
, &counter
->count
);
3179 static void cpu_migrations_perf_counter_read(struct perf_counter
*counter
)
3181 cpu_migrations_perf_counter_update(counter
);
3184 static int cpu_migrations_perf_counter_enable(struct perf_counter
*counter
)
3186 if (counter
->prev_state
<= PERF_COUNTER_STATE_OFF
)
3187 atomic64_set(&counter
->hw
.prev_count
,
3188 get_cpu_migrations(counter
));
3192 static void cpu_migrations_perf_counter_disable(struct perf_counter
*counter
)
3194 cpu_migrations_perf_counter_update(counter
);
3197 static const struct pmu perf_ops_cpu_migrations
= {
3198 .enable
= cpu_migrations_perf_counter_enable
,
3199 .disable
= cpu_migrations_perf_counter_disable
,
3200 .read
= cpu_migrations_perf_counter_read
,
3203 #ifdef CONFIG_EVENT_PROFILE
3204 void perf_tpcounter_event(int event_id
)
3206 struct pt_regs
*regs
= get_irq_regs();
3209 regs
= task_pt_regs(current
);
3211 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
, 0);
3213 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
3215 extern int ftrace_profile_enable(int);
3216 extern void ftrace_profile_disable(int);
3218 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
3220 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
3223 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3225 int event_id
= perf_event_id(&counter
->hw_event
);
3228 ret
= ftrace_profile_enable(event_id
);
3232 counter
->destroy
= tp_perf_counter_destroy
;
3233 counter
->hw
.irq_period
= counter
->hw_event
.irq_period
;
3235 return &perf_ops_generic
;
3238 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3244 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
3246 const struct pmu
*pmu
= NULL
;
3249 * Software counters (currently) can't in general distinguish
3250 * between user, kernel and hypervisor events.
3251 * However, context switches and cpu migrations are considered
3252 * to be kernel events, and page faults are never hypervisor
3255 switch (perf_event_id(&counter
->hw_event
)) {
3256 case PERF_COUNT_CPU_CLOCK
:
3257 pmu
= &perf_ops_cpu_clock
;
3260 case PERF_COUNT_TASK_CLOCK
:
3262 * If the user instantiates this as a per-cpu counter,
3263 * use the cpu_clock counter instead.
3265 if (counter
->ctx
->task
)
3266 pmu
= &perf_ops_task_clock
;
3268 pmu
= &perf_ops_cpu_clock
;
3271 case PERF_COUNT_PAGE_FAULTS
:
3272 case PERF_COUNT_PAGE_FAULTS_MIN
:
3273 case PERF_COUNT_PAGE_FAULTS_MAJ
:
3274 case PERF_COUNT_CONTEXT_SWITCHES
:
3275 pmu
= &perf_ops_generic
;
3277 case PERF_COUNT_CPU_MIGRATIONS
:
3278 if (!counter
->hw_event
.exclude_kernel
)
3279 pmu
= &perf_ops_cpu_migrations
;
3287 * Allocate and initialize a counter structure
3289 static struct perf_counter
*
3290 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
3292 struct perf_counter_context
*ctx
,
3293 struct perf_counter
*group_leader
,
3296 const struct pmu
*pmu
;
3297 struct perf_counter
*counter
;
3298 struct hw_perf_counter
*hwc
;
3301 counter
= kzalloc(sizeof(*counter
), gfpflags
);
3303 return ERR_PTR(-ENOMEM
);
3306 * Single counters are their own group leaders, with an
3307 * empty sibling list:
3310 group_leader
= counter
;
3312 mutex_init(&counter
->child_mutex
);
3313 INIT_LIST_HEAD(&counter
->child_list
);
3315 INIT_LIST_HEAD(&counter
->list_entry
);
3316 INIT_LIST_HEAD(&counter
->event_entry
);
3317 INIT_LIST_HEAD(&counter
->sibling_list
);
3318 init_waitqueue_head(&counter
->waitq
);
3320 mutex_init(&counter
->mmap_mutex
);
3323 counter
->hw_event
= *hw_event
;
3324 counter
->group_leader
= group_leader
;
3325 counter
->pmu
= NULL
;
3327 counter
->oncpu
= -1;
3329 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3330 if (hw_event
->disabled
)
3331 counter
->state
= PERF_COUNTER_STATE_OFF
;
3336 if (hw_event
->freq
&& hw_event
->irq_freq
)
3337 hwc
->irq_period
= div64_u64(TICK_NSEC
, hw_event
->irq_freq
);
3339 hwc
->irq_period
= hw_event
->irq_period
;
3342 * we currently do not support PERF_RECORD_GROUP on inherited counters
3344 if (hw_event
->inherit
&& (hw_event
->record_type
& PERF_RECORD_GROUP
))
3347 if (perf_event_raw(hw_event
)) {
3348 pmu
= hw_perf_counter_init(counter
);
3352 switch (perf_event_type(hw_event
)) {
3353 case PERF_TYPE_HARDWARE
:
3354 pmu
= hw_perf_counter_init(counter
);
3357 case PERF_TYPE_SOFTWARE
:
3358 pmu
= sw_perf_counter_init(counter
);
3361 case PERF_TYPE_TRACEPOINT
:
3362 pmu
= tp_perf_counter_init(counter
);
3369 else if (IS_ERR(pmu
))
3374 return ERR_PTR(err
);
3379 atomic_inc(&nr_counters
);
3380 if (counter
->hw_event
.mmap
)
3381 atomic_inc(&nr_mmap_tracking
);
3382 if (counter
->hw_event
.munmap
)
3383 atomic_inc(&nr_munmap_tracking
);
3384 if (counter
->hw_event
.comm
)
3385 atomic_inc(&nr_comm_tracking
);
3391 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3393 * @hw_event_uptr: event type attributes for monitoring/sampling
3396 * @group_fd: group leader counter fd
3398 SYSCALL_DEFINE5(perf_counter_open
,
3399 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
3400 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
3402 struct perf_counter
*counter
, *group_leader
;
3403 struct perf_counter_hw_event hw_event
;
3404 struct perf_counter_context
*ctx
;
3405 struct file
*counter_file
= NULL
;
3406 struct file
*group_file
= NULL
;
3407 int fput_needed
= 0;
3408 int fput_needed2
= 0;
3411 /* for future expandability... */
3415 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
3419 * Get the target context (task or percpu):
3421 ctx
= find_get_context(pid
, cpu
);
3423 return PTR_ERR(ctx
);
3426 * Look up the group leader (we will attach this counter to it):
3428 group_leader
= NULL
;
3429 if (group_fd
!= -1) {
3431 group_file
= fget_light(group_fd
, &fput_needed
);
3433 goto err_put_context
;
3434 if (group_file
->f_op
!= &perf_fops
)
3435 goto err_put_context
;
3437 group_leader
= group_file
->private_data
;
3439 * Do not allow a recursive hierarchy (this new sibling
3440 * becoming part of another group-sibling):
3442 if (group_leader
->group_leader
!= group_leader
)
3443 goto err_put_context
;
3445 * Do not allow to attach to a group in a different
3446 * task or CPU context:
3448 if (group_leader
->ctx
!= ctx
)
3449 goto err_put_context
;
3451 * Only a group leader can be exclusive or pinned
3453 if (hw_event
.exclusive
|| hw_event
.pinned
)
3454 goto err_put_context
;
3457 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
3459 ret
= PTR_ERR(counter
);
3460 if (IS_ERR(counter
))
3461 goto err_put_context
;
3463 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
3465 goto err_free_put_context
;
3467 counter_file
= fget_light(ret
, &fput_needed2
);
3469 goto err_free_put_context
;
3471 counter
->filp
= counter_file
;
3472 WARN_ON_ONCE(ctx
->parent_ctx
);
3473 mutex_lock(&ctx
->mutex
);
3474 perf_install_in_context(ctx
, counter
, cpu
);
3476 mutex_unlock(&ctx
->mutex
);
3478 counter
->owner
= current
;
3479 get_task_struct(current
);
3480 mutex_lock(¤t
->perf_counter_mutex
);
3481 list_add_tail(&counter
->owner_entry
, ¤t
->perf_counter_list
);
3482 mutex_unlock(¤t
->perf_counter_mutex
);
3484 fput_light(counter_file
, fput_needed2
);
3487 fput_light(group_file
, fput_needed
);
3491 err_free_put_context
:
3501 * inherit a counter from parent task to child task:
3503 static struct perf_counter
*
3504 inherit_counter(struct perf_counter
*parent_counter
,
3505 struct task_struct
*parent
,
3506 struct perf_counter_context
*parent_ctx
,
3507 struct task_struct
*child
,
3508 struct perf_counter
*group_leader
,
3509 struct perf_counter_context
*child_ctx
)
3511 struct perf_counter
*child_counter
;
3514 * Instead of creating recursive hierarchies of counters,
3515 * we link inherited counters back to the original parent,
3516 * which has a filp for sure, which we use as the reference
3519 if (parent_counter
->parent
)
3520 parent_counter
= parent_counter
->parent
;
3522 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
3523 parent_counter
->cpu
, child_ctx
,
3524 group_leader
, GFP_KERNEL
);
3525 if (IS_ERR(child_counter
))
3526 return child_counter
;
3530 * Make the child state follow the state of the parent counter,
3531 * not its hw_event.disabled bit. We hold the parent's mutex,
3532 * so we won't race with perf_counter_{en, dis}able_family.
3534 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3535 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3537 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3540 * Link it up in the child's context:
3542 add_counter_to_ctx(child_counter
, child_ctx
);
3544 child_counter
->parent
= parent_counter
;
3546 * inherit into child's child as well:
3548 child_counter
->hw_event
.inherit
= 1;
3551 * Get a reference to the parent filp - we will fput it
3552 * when the child counter exits. This is safe to do because
3553 * we are in the parent and we know that the filp still
3554 * exists and has a nonzero count:
3556 atomic_long_inc(&parent_counter
->filp
->f_count
);
3559 * Link this into the parent counter's child list
3561 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3562 mutex_lock(&parent_counter
->child_mutex
);
3563 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3564 mutex_unlock(&parent_counter
->child_mutex
);
3566 return child_counter
;
3569 static int inherit_group(struct perf_counter
*parent_counter
,
3570 struct task_struct
*parent
,
3571 struct perf_counter_context
*parent_ctx
,
3572 struct task_struct
*child
,
3573 struct perf_counter_context
*child_ctx
)
3575 struct perf_counter
*leader
;
3576 struct perf_counter
*sub
;
3577 struct perf_counter
*child_ctr
;
3579 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3580 child
, NULL
, child_ctx
);
3582 return PTR_ERR(leader
);
3583 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3584 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3585 child
, leader
, child_ctx
);
3586 if (IS_ERR(child_ctr
))
3587 return PTR_ERR(child_ctr
);
3592 static void sync_child_counter(struct perf_counter
*child_counter
,
3593 struct perf_counter
*parent_counter
)
3597 child_val
= atomic64_read(&child_counter
->count
);
3600 * Add back the child's count to the parent's count:
3602 atomic64_add(child_val
, &parent_counter
->count
);
3603 atomic64_add(child_counter
->total_time_enabled
,
3604 &parent_counter
->child_total_time_enabled
);
3605 atomic64_add(child_counter
->total_time_running
,
3606 &parent_counter
->child_total_time_running
);
3609 * Remove this counter from the parent's list
3611 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3612 mutex_lock(&parent_counter
->child_mutex
);
3613 list_del_init(&child_counter
->child_list
);
3614 mutex_unlock(&parent_counter
->child_mutex
);
3617 * Release the parent counter, if this was the last
3620 fput(parent_counter
->filp
);
3624 __perf_counter_exit_task(struct perf_counter
*child_counter
,
3625 struct perf_counter_context
*child_ctx
)
3627 struct perf_counter
*parent_counter
;
3629 update_counter_times(child_counter
);
3630 perf_counter_remove_from_context(child_counter
);
3632 parent_counter
= child_counter
->parent
;
3634 * It can happen that parent exits first, and has counters
3635 * that are still around due to the child reference. These
3636 * counters need to be zapped - but otherwise linger.
3638 if (parent_counter
) {
3639 sync_child_counter(child_counter
, parent_counter
);
3640 free_counter(child_counter
);
3645 * When a child task exits, feed back counter values to parent counters.
3647 void perf_counter_exit_task(struct task_struct
*child
)
3649 struct perf_counter
*child_counter
, *tmp
;
3650 struct perf_counter_context
*child_ctx
;
3651 unsigned long flags
;
3653 if (likely(!child
->perf_counter_ctxp
))
3656 local_irq_save(flags
);
3658 * We can't reschedule here because interrupts are disabled,
3659 * and either child is current or it is a task that can't be
3660 * scheduled, so we are now safe from rescheduling changing
3663 child_ctx
= child
->perf_counter_ctxp
;
3664 __perf_counter_task_sched_out(child_ctx
);
3667 * Take the context lock here so that if find_get_context is
3668 * reading child->perf_counter_ctxp, we wait until it has
3669 * incremented the context's refcount before we do put_ctx below.
3671 spin_lock(&child_ctx
->lock
);
3672 child
->perf_counter_ctxp
= NULL
;
3673 if (child_ctx
->parent_ctx
) {
3675 * This context is a clone; unclone it so it can't get
3676 * swapped to another process while we're removing all
3677 * the counters from it.
3679 put_ctx(child_ctx
->parent_ctx
);
3680 child_ctx
->parent_ctx
= NULL
;
3682 spin_unlock(&child_ctx
->lock
);
3683 local_irq_restore(flags
);
3685 mutex_lock(&child_ctx
->mutex
);
3688 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
3690 __perf_counter_exit_task(child_counter
, child_ctx
);
3693 * If the last counter was a group counter, it will have appended all
3694 * its siblings to the list, but we obtained 'tmp' before that which
3695 * will still point to the list head terminating the iteration.
3697 if (!list_empty(&child_ctx
->counter_list
))
3700 mutex_unlock(&child_ctx
->mutex
);
3706 * free an unexposed, unused context as created by inheritance by
3707 * init_task below, used by fork() in case of fail.
3709 void perf_counter_free_task(struct task_struct
*task
)
3711 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
3712 struct perf_counter
*counter
, *tmp
;
3717 mutex_lock(&ctx
->mutex
);
3719 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
) {
3720 struct perf_counter
*parent
= counter
->parent
;
3722 if (WARN_ON_ONCE(!parent
))
3725 mutex_lock(&parent
->child_mutex
);
3726 list_del_init(&counter
->child_list
);
3727 mutex_unlock(&parent
->child_mutex
);
3731 list_del_counter(counter
, ctx
);
3732 free_counter(counter
);
3735 if (!list_empty(&ctx
->counter_list
))
3738 mutex_unlock(&ctx
->mutex
);
3744 * Initialize the perf_counter context in task_struct
3746 int perf_counter_init_task(struct task_struct
*child
)
3748 struct perf_counter_context
*child_ctx
, *parent_ctx
;
3749 struct perf_counter_context
*cloned_ctx
;
3750 struct perf_counter
*counter
;
3751 struct task_struct
*parent
= current
;
3752 int inherited_all
= 1;
3755 child
->perf_counter_ctxp
= NULL
;
3757 mutex_init(&child
->perf_counter_mutex
);
3758 INIT_LIST_HEAD(&child
->perf_counter_list
);
3760 if (likely(!parent
->perf_counter_ctxp
))
3764 * This is executed from the parent task context, so inherit
3765 * counters that have been marked for cloning.
3766 * First allocate and initialize a context for the child.
3769 child_ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
3773 __perf_counter_init_context(child_ctx
, child
);
3774 child
->perf_counter_ctxp
= child_ctx
;
3775 get_task_struct(child
);
3778 * If the parent's context is a clone, pin it so it won't get
3781 parent_ctx
= perf_pin_task_context(parent
);
3784 * No need to check if parent_ctx != NULL here; since we saw
3785 * it non-NULL earlier, the only reason for it to become NULL
3786 * is if we exit, and since we're currently in the middle of
3787 * a fork we can't be exiting at the same time.
3791 * Lock the parent list. No need to lock the child - not PID
3792 * hashed yet and not running, so nobody can access it.
3794 mutex_lock(&parent_ctx
->mutex
);
3797 * We dont have to disable NMIs - we are only looking at
3798 * the list, not manipulating it:
3800 list_for_each_entry_rcu(counter
, &parent_ctx
->event_list
, event_entry
) {
3801 if (counter
!= counter
->group_leader
)
3804 if (!counter
->hw_event
.inherit
) {
3809 ret
= inherit_group(counter
, parent
, parent_ctx
,
3817 if (inherited_all
) {
3819 * Mark the child context as a clone of the parent
3820 * context, or of whatever the parent is a clone of.
3821 * Note that if the parent is a clone, it could get
3822 * uncloned at any point, but that doesn't matter
3823 * because the list of counters and the generation
3824 * count can't have changed since we took the mutex.
3826 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
3828 child_ctx
->parent_ctx
= cloned_ctx
;
3829 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
3831 child_ctx
->parent_ctx
= parent_ctx
;
3832 child_ctx
->parent_gen
= parent_ctx
->generation
;
3834 get_ctx(child_ctx
->parent_ctx
);
3837 mutex_unlock(&parent_ctx
->mutex
);
3839 perf_unpin_context(parent_ctx
);
3844 static void __cpuinit
perf_counter_init_cpu(int cpu
)
3846 struct perf_cpu_context
*cpuctx
;
3848 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3849 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
3851 spin_lock(&perf_resource_lock
);
3852 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
3853 spin_unlock(&perf_resource_lock
);
3855 hw_perf_counter_setup(cpu
);
3858 #ifdef CONFIG_HOTPLUG_CPU
3859 static void __perf_counter_exit_cpu(void *info
)
3861 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3862 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3863 struct perf_counter
*counter
, *tmp
;
3865 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
3866 __perf_counter_remove_from_context(counter
);
3868 static void perf_counter_exit_cpu(int cpu
)
3870 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3871 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3873 mutex_lock(&ctx
->mutex
);
3874 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
3875 mutex_unlock(&ctx
->mutex
);
3878 static inline void perf_counter_exit_cpu(int cpu
) { }
3881 static int __cpuinit
3882 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
3884 unsigned int cpu
= (long)hcpu
;
3888 case CPU_UP_PREPARE
:
3889 case CPU_UP_PREPARE_FROZEN
:
3890 perf_counter_init_cpu(cpu
);
3893 case CPU_DOWN_PREPARE
:
3894 case CPU_DOWN_PREPARE_FROZEN
:
3895 perf_counter_exit_cpu(cpu
);
3905 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
3906 .notifier_call
= perf_cpu_notify
,
3909 void __init
perf_counter_init(void)
3911 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
3912 (void *)(long)smp_processor_id());
3913 register_cpu_notifier(&perf_cpu_nb
);
3916 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
3918 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
3922 perf_set_reserve_percpu(struct sysdev_class
*class,
3926 struct perf_cpu_context
*cpuctx
;
3930 err
= strict_strtoul(buf
, 10, &val
);
3933 if (val
> perf_max_counters
)
3936 spin_lock(&perf_resource_lock
);
3937 perf_reserved_percpu
= val
;
3938 for_each_online_cpu(cpu
) {
3939 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3940 spin_lock_irq(&cpuctx
->ctx
.lock
);
3941 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
3942 perf_max_counters
- perf_reserved_percpu
);
3943 cpuctx
->max_pertask
= mpt
;
3944 spin_unlock_irq(&cpuctx
->ctx
.lock
);
3946 spin_unlock(&perf_resource_lock
);
3951 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
3953 return sprintf(buf
, "%d\n", perf_overcommit
);
3957 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
3962 err
= strict_strtoul(buf
, 10, &val
);
3968 spin_lock(&perf_resource_lock
);
3969 perf_overcommit
= val
;
3970 spin_unlock(&perf_resource_lock
);
3975 static SYSDEV_CLASS_ATTR(
3978 perf_show_reserve_percpu
,
3979 perf_set_reserve_percpu
3982 static SYSDEV_CLASS_ATTR(
3985 perf_show_overcommit
,
3989 static struct attribute
*perfclass_attrs
[] = {
3990 &attr_reserve_percpu
.attr
,
3991 &attr_overcommit
.attr
,
3995 static struct attribute_group perfclass_attr_group
= {
3996 .attrs
= perfclass_attrs
,
3997 .name
= "perf_counters",
4000 static int __init
perf_counter_sysfs_init(void)
4002 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
4003 &perfclass_attr_group
);
4005 device_initcall(perf_counter_sysfs_init
);