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/ptrace.h>
20 #include <linux/percpu.h>
21 #include <linux/vmstat.h>
22 #include <linux/hardirq.h>
23 #include <linux/rculist.h>
24 #include <linux/uaccess.h>
25 #include <linux/syscalls.h>
26 #include <linux/anon_inodes.h>
27 #include <linux/kernel_stat.h>
28 #include <linux/perf_counter.h>
29 #include <linux/dcache.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
= 128; /* 'free' kb per counter */
51 * Lock for (sysadmin-configurable) counter reservations:
53 static DEFINE_SPINLOCK(perf_resource_lock
);
56 * Architecture provided APIs - weak aliases:
58 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
63 void __weak
hw_perf_disable(void) { barrier(); }
64 void __weak
hw_perf_enable(void) { barrier(); }
66 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
67 int __weak
hw_perf_group_sched_in(struct perf_counter
*group_leader
,
68 struct perf_cpu_context
*cpuctx
,
69 struct perf_counter_context
*ctx
, int cpu
)
74 void __weak
perf_counter_print_debug(void) { }
76 static DEFINE_PER_CPU(int, disable_count
);
78 void __perf_disable(void)
80 __get_cpu_var(disable_count
)++;
83 bool __perf_enable(void)
85 return !--__get_cpu_var(disable_count
);
88 void perf_disable(void)
94 void perf_enable(void)
101 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
103 struct perf_counter
*group_leader
= counter
->group_leader
;
106 * Depending on whether it is a standalone or sibling counter,
107 * add it straight to the context's counter list, or to the group
108 * leader's sibling list:
110 if (group_leader
== counter
)
111 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
113 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
114 group_leader
->nr_siblings
++;
117 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
121 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
123 struct perf_counter
*sibling
, *tmp
;
125 list_del_init(&counter
->list_entry
);
126 list_del_rcu(&counter
->event_entry
);
128 if (counter
->group_leader
!= counter
)
129 counter
->group_leader
->nr_siblings
--;
132 * If this was a group counter with sibling counters then
133 * upgrade the siblings to singleton counters by adding them
134 * to the context list directly:
136 list_for_each_entry_safe(sibling
, tmp
,
137 &counter
->sibling_list
, list_entry
) {
139 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
140 sibling
->group_leader
= sibling
;
145 counter_sched_out(struct perf_counter
*counter
,
146 struct perf_cpu_context
*cpuctx
,
147 struct perf_counter_context
*ctx
)
149 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
152 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
153 counter
->tstamp_stopped
= ctx
->time
;
154 counter
->pmu
->disable(counter
);
157 if (!is_software_counter(counter
))
158 cpuctx
->active_oncpu
--;
160 if (counter
->hw_event
.exclusive
|| !cpuctx
->active_oncpu
)
161 cpuctx
->exclusive
= 0;
165 group_sched_out(struct perf_counter
*group_counter
,
166 struct perf_cpu_context
*cpuctx
,
167 struct perf_counter_context
*ctx
)
169 struct perf_counter
*counter
;
171 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
174 counter_sched_out(group_counter
, cpuctx
, ctx
);
177 * Schedule out siblings (if any):
179 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
180 counter_sched_out(counter
, cpuctx
, ctx
);
182 if (group_counter
->hw_event
.exclusive
)
183 cpuctx
->exclusive
= 0;
187 * Cross CPU call to remove a performance counter
189 * We disable the counter on the hardware level first. After that we
190 * remove it from the context list.
192 static void __perf_counter_remove_from_context(void *info
)
194 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
195 struct perf_counter
*counter
= info
;
196 struct perf_counter_context
*ctx
= counter
->ctx
;
200 * If this is a task context, we need to check whether it is
201 * the current task context of this cpu. If not it has been
202 * scheduled out before the smp call arrived.
204 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
207 spin_lock_irqsave(&ctx
->lock
, flags
);
209 counter_sched_out(counter
, cpuctx
, ctx
);
211 counter
->task
= NULL
;
215 * Protect the list operation against NMI by disabling the
216 * counters on a global level. NOP for non NMI based counters.
219 list_del_counter(counter
, ctx
);
224 * Allow more per task counters with respect to the
227 cpuctx
->max_pertask
=
228 min(perf_max_counters
- ctx
->nr_counters
,
229 perf_max_counters
- perf_reserved_percpu
);
232 spin_unlock_irqrestore(&ctx
->lock
, flags
);
237 * Remove the counter from a task's (or a CPU's) list of counters.
239 * Must be called with counter->mutex and ctx->mutex held.
241 * CPU counters are removed with a smp call. For task counters we only
242 * call when the task is on a CPU.
244 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
246 struct perf_counter_context
*ctx
= counter
->ctx
;
247 struct task_struct
*task
= ctx
->task
;
251 * Per cpu counters are removed via an smp call and
252 * the removal is always sucessful.
254 smp_call_function_single(counter
->cpu
,
255 __perf_counter_remove_from_context
,
261 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
264 spin_lock_irq(&ctx
->lock
);
266 * If the context is active we need to retry the smp call.
268 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
269 spin_unlock_irq(&ctx
->lock
);
274 * The lock prevents that this context is scheduled in so we
275 * can remove the counter safely, if the call above did not
278 if (!list_empty(&counter
->list_entry
)) {
280 list_del_counter(counter
, ctx
);
281 counter
->task
= NULL
;
283 spin_unlock_irq(&ctx
->lock
);
286 static inline u64
perf_clock(void)
288 return cpu_clock(smp_processor_id());
292 * Update the record of the current time in a context.
294 static void update_context_time(struct perf_counter_context
*ctx
)
296 u64 now
= perf_clock();
298 ctx
->time
+= now
- ctx
->timestamp
;
299 ctx
->timestamp
= now
;
303 * Update the total_time_enabled and total_time_running fields for a counter.
305 static void update_counter_times(struct perf_counter
*counter
)
307 struct perf_counter_context
*ctx
= counter
->ctx
;
310 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
313 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
315 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
316 run_end
= counter
->tstamp_stopped
;
320 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
324 * Update total_time_enabled and total_time_running for all counters in a group.
326 static void update_group_times(struct perf_counter
*leader
)
328 struct perf_counter
*counter
;
330 update_counter_times(leader
);
331 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
332 update_counter_times(counter
);
336 * Cross CPU call to disable a performance counter
338 static void __perf_counter_disable(void *info
)
340 struct perf_counter
*counter
= info
;
341 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
342 struct perf_counter_context
*ctx
= counter
->ctx
;
346 * If this is a per-task counter, need to check whether this
347 * counter's task is the current task on this cpu.
349 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
352 spin_lock_irqsave(&ctx
->lock
, flags
);
355 * If the counter is on, turn it off.
356 * If it is in error state, leave it in error state.
358 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
359 update_context_time(ctx
);
360 update_counter_times(counter
);
361 if (counter
== counter
->group_leader
)
362 group_sched_out(counter
, cpuctx
, ctx
);
364 counter_sched_out(counter
, cpuctx
, ctx
);
365 counter
->state
= PERF_COUNTER_STATE_OFF
;
368 spin_unlock_irqrestore(&ctx
->lock
, flags
);
374 static void perf_counter_disable(struct perf_counter
*counter
)
376 struct perf_counter_context
*ctx
= counter
->ctx
;
377 struct task_struct
*task
= ctx
->task
;
381 * Disable the counter on the cpu that it's on
383 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
389 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
391 spin_lock_irq(&ctx
->lock
);
393 * If the counter is still active, we need to retry the cross-call.
395 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
396 spin_unlock_irq(&ctx
->lock
);
401 * Since we have the lock this context can't be scheduled
402 * in, so we can change the state safely.
404 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
405 update_counter_times(counter
);
406 counter
->state
= PERF_COUNTER_STATE_OFF
;
409 spin_unlock_irq(&ctx
->lock
);
413 counter_sched_in(struct perf_counter
*counter
,
414 struct perf_cpu_context
*cpuctx
,
415 struct perf_counter_context
*ctx
,
418 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
421 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
422 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
424 * The new state must be visible before we turn it on in the hardware:
428 if (counter
->pmu
->enable(counter
)) {
429 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
434 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
436 if (!is_software_counter(counter
))
437 cpuctx
->active_oncpu
++;
440 if (counter
->hw_event
.exclusive
)
441 cpuctx
->exclusive
= 1;
447 group_sched_in(struct perf_counter
*group_counter
,
448 struct perf_cpu_context
*cpuctx
,
449 struct perf_counter_context
*ctx
,
452 struct perf_counter
*counter
, *partial_group
;
455 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
458 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
460 return ret
< 0 ? ret
: 0;
462 group_counter
->prev_state
= group_counter
->state
;
463 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
467 * Schedule in siblings as one group (if any):
469 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
470 counter
->prev_state
= counter
->state
;
471 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
472 partial_group
= counter
;
481 * Groups can be scheduled in as one unit only, so undo any
482 * partial group before returning:
484 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
485 if (counter
== partial_group
)
487 counter_sched_out(counter
, cpuctx
, ctx
);
489 counter_sched_out(group_counter
, cpuctx
, ctx
);
495 * Return 1 for a group consisting entirely of software counters,
496 * 0 if the group contains any hardware counters.
498 static int is_software_only_group(struct perf_counter
*leader
)
500 struct perf_counter
*counter
;
502 if (!is_software_counter(leader
))
505 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
506 if (!is_software_counter(counter
))
513 * Work out whether we can put this counter group on the CPU now.
515 static int group_can_go_on(struct perf_counter
*counter
,
516 struct perf_cpu_context
*cpuctx
,
520 * Groups consisting entirely of software counters can always go on.
522 if (is_software_only_group(counter
))
525 * If an exclusive group is already on, no other hardware
526 * counters can go on.
528 if (cpuctx
->exclusive
)
531 * If this group is exclusive and there are already
532 * counters on the CPU, it can't go on.
534 if (counter
->hw_event
.exclusive
&& cpuctx
->active_oncpu
)
537 * Otherwise, try to add it if all previous groups were able
543 static void add_counter_to_ctx(struct perf_counter
*counter
,
544 struct perf_counter_context
*ctx
)
546 list_add_counter(counter
, ctx
);
548 counter
->prev_state
= PERF_COUNTER_STATE_OFF
;
549 counter
->tstamp_enabled
= ctx
->time
;
550 counter
->tstamp_running
= ctx
->time
;
551 counter
->tstamp_stopped
= ctx
->time
;
555 * Cross CPU call to install and enable a performance counter
557 static void __perf_install_in_context(void *info
)
559 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
560 struct perf_counter
*counter
= info
;
561 struct perf_counter_context
*ctx
= counter
->ctx
;
562 struct perf_counter
*leader
= counter
->group_leader
;
563 int cpu
= smp_processor_id();
568 * If this is a task context, we need to check whether it is
569 * the current task context of this cpu. If not it has been
570 * scheduled out before the smp call arrived.
572 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
575 spin_lock_irqsave(&ctx
->lock
, flags
);
576 update_context_time(ctx
);
579 * Protect the list operation against NMI by disabling the
580 * counters on a global level. NOP for non NMI based counters.
584 add_counter_to_ctx(counter
, ctx
);
587 * Don't put the counter on if it is disabled or if
588 * it is in a group and the group isn't on.
590 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
591 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
595 * An exclusive counter can't go on if there are already active
596 * hardware counters, and no hardware counter can go on if there
597 * is already an exclusive counter on.
599 if (!group_can_go_on(counter
, cpuctx
, 1))
602 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
606 * This counter couldn't go on. If it is in a group
607 * then we have to pull the whole group off.
608 * If the counter group is pinned then put it in error state.
610 if (leader
!= counter
)
611 group_sched_out(leader
, cpuctx
, ctx
);
612 if (leader
->hw_event
.pinned
) {
613 update_group_times(leader
);
614 leader
->state
= PERF_COUNTER_STATE_ERROR
;
618 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
619 cpuctx
->max_pertask
--;
624 spin_unlock_irqrestore(&ctx
->lock
, flags
);
628 * Attach a performance counter to a context
630 * First we add the counter to the list with the hardware enable bit
631 * in counter->hw_config cleared.
633 * If the counter is attached to a task which is on a CPU we use a smp
634 * call to enable it in the task context. The task might have been
635 * scheduled away, but we check this in the smp call again.
637 * Must be called with ctx->mutex held.
640 perf_install_in_context(struct perf_counter_context
*ctx
,
641 struct perf_counter
*counter
,
644 struct task_struct
*task
= ctx
->task
;
648 * Per cpu counters are installed via an smp call and
649 * the install is always sucessful.
651 smp_call_function_single(cpu
, __perf_install_in_context
,
656 counter
->task
= task
;
658 task_oncpu_function_call(task
, __perf_install_in_context
,
661 spin_lock_irq(&ctx
->lock
);
663 * we need to retry the smp call.
665 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
666 spin_unlock_irq(&ctx
->lock
);
671 * The lock prevents that this context is scheduled in so we
672 * can add the counter safely, if it the call above did not
675 if (list_empty(&counter
->list_entry
))
676 add_counter_to_ctx(counter
, ctx
);
677 spin_unlock_irq(&ctx
->lock
);
681 * Cross CPU call to enable a performance counter
683 static void __perf_counter_enable(void *info
)
685 struct perf_counter
*counter
= info
;
686 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
687 struct perf_counter_context
*ctx
= counter
->ctx
;
688 struct perf_counter
*leader
= counter
->group_leader
;
693 * If this is a per-task counter, need to check whether this
694 * counter's task is the current task on this cpu.
696 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
699 spin_lock_irqsave(&ctx
->lock
, flags
);
700 update_context_time(ctx
);
702 counter
->prev_state
= counter
->state
;
703 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
705 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
706 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
709 * If the counter is in a group and isn't the group leader,
710 * then don't put it on unless the group is on.
712 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
715 if (!group_can_go_on(counter
, cpuctx
, 1)) {
719 if (counter
== leader
)
720 err
= group_sched_in(counter
, cpuctx
, ctx
,
723 err
= counter_sched_in(counter
, cpuctx
, ctx
,
730 * If this counter can't go on and it's part of a
731 * group, then the whole group has to come off.
733 if (leader
!= counter
)
734 group_sched_out(leader
, cpuctx
, ctx
);
735 if (leader
->hw_event
.pinned
) {
736 update_group_times(leader
);
737 leader
->state
= PERF_COUNTER_STATE_ERROR
;
742 spin_unlock_irqrestore(&ctx
->lock
, flags
);
748 static void perf_counter_enable(struct perf_counter
*counter
)
750 struct perf_counter_context
*ctx
= counter
->ctx
;
751 struct task_struct
*task
= ctx
->task
;
755 * Enable the counter on the cpu that it's on
757 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
762 spin_lock_irq(&ctx
->lock
);
763 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
767 * If the counter is in error state, clear that first.
768 * That way, if we see the counter in error state below, we
769 * know that it has gone back into error state, as distinct
770 * from the task having been scheduled away before the
771 * cross-call arrived.
773 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
774 counter
->state
= PERF_COUNTER_STATE_OFF
;
777 spin_unlock_irq(&ctx
->lock
);
778 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
780 spin_lock_irq(&ctx
->lock
);
783 * If the context is active and the counter is still off,
784 * we need to retry the cross-call.
786 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
790 * Since we have the lock this context can't be scheduled
791 * in, so we can change the state safely.
793 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
794 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
795 counter
->tstamp_enabled
=
796 ctx
->time
- counter
->total_time_enabled
;
799 spin_unlock_irq(&ctx
->lock
);
802 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
805 * not supported on inherited counters
807 if (counter
->hw_event
.inherit
)
810 atomic_add(refresh
, &counter
->event_limit
);
811 perf_counter_enable(counter
);
816 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
817 struct perf_cpu_context
*cpuctx
)
819 struct perf_counter
*counter
;
821 spin_lock(&ctx
->lock
);
823 if (likely(!ctx
->nr_counters
))
825 update_context_time(ctx
);
828 if (ctx
->nr_active
) {
829 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
)
830 group_sched_out(counter
, cpuctx
, ctx
);
834 spin_unlock(&ctx
->lock
);
838 * Called from scheduler to remove the counters of the current task,
839 * with interrupts disabled.
841 * We stop each counter and update the counter value in counter->count.
843 * This does not protect us against NMI, but disable()
844 * sets the disabled bit in the control field of counter _before_
845 * accessing the counter control register. If a NMI hits, then it will
846 * not restart the counter.
848 void perf_counter_task_sched_out(struct task_struct
*task
, int cpu
)
850 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
851 struct perf_counter_context
*ctx
= &task
->perf_counter_ctx
;
852 struct pt_regs
*regs
;
854 if (likely(!cpuctx
->task_ctx
))
857 update_context_time(ctx
);
859 regs
= task_pt_regs(task
);
860 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
861 __perf_counter_sched_out(ctx
, cpuctx
);
863 cpuctx
->task_ctx
= NULL
;
866 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
868 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
870 __perf_counter_sched_out(ctx
, cpuctx
);
871 cpuctx
->task_ctx
= NULL
;
874 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
876 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
880 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
881 struct perf_cpu_context
*cpuctx
, int cpu
)
883 struct perf_counter
*counter
;
886 spin_lock(&ctx
->lock
);
888 if (likely(!ctx
->nr_counters
))
891 ctx
->timestamp
= perf_clock();
896 * First go through the list and put on any pinned groups
897 * in order to give them the best chance of going on.
899 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
900 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
901 !counter
->hw_event
.pinned
)
903 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
906 if (group_can_go_on(counter
, cpuctx
, 1))
907 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
910 * If this pinned group hasn't been scheduled,
911 * put it in error state.
913 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
914 update_group_times(counter
);
915 counter
->state
= PERF_COUNTER_STATE_ERROR
;
919 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
921 * Ignore counters in OFF or ERROR state, and
922 * ignore pinned counters since we did them already.
924 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
925 counter
->hw_event
.pinned
)
929 * Listen to the 'cpu' scheduling filter constraint
932 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
935 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
936 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
942 spin_unlock(&ctx
->lock
);
946 * Called from scheduler to add the counters of the current task
947 * with interrupts disabled.
949 * We restore the counter value and then enable it.
951 * This does not protect us against NMI, but enable()
952 * sets the enabled bit in the control field of counter _before_
953 * accessing the counter control register. If a NMI hits, then it will
954 * keep the counter running.
956 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
958 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
959 struct perf_counter_context
*ctx
= &task
->perf_counter_ctx
;
961 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
962 cpuctx
->task_ctx
= ctx
;
965 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
967 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
969 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
972 int perf_counter_task_disable(void)
974 struct task_struct
*curr
= current
;
975 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
976 struct perf_counter
*counter
;
979 if (likely(!ctx
->nr_counters
))
982 local_irq_save(flags
);
984 __perf_counter_task_sched_out(ctx
);
986 spin_lock(&ctx
->lock
);
989 * Disable all the counters:
993 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
994 if (counter
->state
!= PERF_COUNTER_STATE_ERROR
) {
995 update_group_times(counter
);
996 counter
->state
= PERF_COUNTER_STATE_OFF
;
1002 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1007 int perf_counter_task_enable(void)
1009 struct task_struct
*curr
= current
;
1010 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
1011 struct perf_counter
*counter
;
1012 unsigned long flags
;
1015 if (likely(!ctx
->nr_counters
))
1018 local_irq_save(flags
);
1019 cpu
= smp_processor_id();
1021 __perf_counter_task_sched_out(ctx
);
1023 spin_lock(&ctx
->lock
);
1026 * Disable all the counters:
1030 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1031 if (counter
->state
> PERF_COUNTER_STATE_OFF
)
1033 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
1034 counter
->tstamp_enabled
=
1035 ctx
->time
- counter
->total_time_enabled
;
1036 counter
->hw_event
.disabled
= 0;
1040 spin_unlock(&ctx
->lock
);
1042 perf_counter_task_sched_in(curr
, cpu
);
1044 local_irq_restore(flags
);
1050 * Round-robin a context's counters:
1052 static void rotate_ctx(struct perf_counter_context
*ctx
)
1054 struct perf_counter
*counter
;
1056 if (!ctx
->nr_counters
)
1059 spin_lock(&ctx
->lock
);
1061 * Rotate the first entry last (works just fine for group counters too):
1064 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1065 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1070 spin_unlock(&ctx
->lock
);
1073 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1075 struct perf_cpu_context
*cpuctx
;
1076 struct perf_counter_context
*ctx
;
1078 if (!atomic_read(&nr_counters
))
1081 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1082 ctx
= &curr
->perf_counter_ctx
;
1084 perf_counter_cpu_sched_out(cpuctx
);
1085 __perf_counter_task_sched_out(ctx
);
1087 rotate_ctx(&cpuctx
->ctx
);
1090 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1091 perf_counter_task_sched_in(curr
, cpu
);
1095 * Cross CPU call to read the hardware counter
1097 static void __read(void *info
)
1099 struct perf_counter
*counter
= info
;
1100 struct perf_counter_context
*ctx
= counter
->ctx
;
1101 unsigned long flags
;
1103 local_irq_save(flags
);
1105 update_context_time(ctx
);
1106 counter
->pmu
->read(counter
);
1107 update_counter_times(counter
);
1108 local_irq_restore(flags
);
1111 static u64
perf_counter_read(struct perf_counter
*counter
)
1114 * If counter is enabled and currently active on a CPU, update the
1115 * value in the counter structure:
1117 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1118 smp_call_function_single(counter
->oncpu
,
1119 __read
, counter
, 1);
1120 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1121 update_counter_times(counter
);
1124 return atomic64_read(&counter
->count
);
1127 static void put_context(struct perf_counter_context
*ctx
)
1130 put_task_struct(ctx
->task
);
1133 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1135 struct perf_cpu_context
*cpuctx
;
1136 struct perf_counter_context
*ctx
;
1137 struct task_struct
*task
;
1140 * If cpu is not a wildcard then this is a percpu counter:
1143 /* Must be root to operate on a CPU counter: */
1144 if (sysctl_perf_counter_priv
&& !capable(CAP_SYS_ADMIN
))
1145 return ERR_PTR(-EACCES
);
1147 if (cpu
< 0 || cpu
> num_possible_cpus())
1148 return ERR_PTR(-EINVAL
);
1151 * We could be clever and allow to attach a counter to an
1152 * offline CPU and activate it when the CPU comes up, but
1155 if (!cpu_isset(cpu
, cpu_online_map
))
1156 return ERR_PTR(-ENODEV
);
1158 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1168 task
= find_task_by_vpid(pid
);
1170 get_task_struct(task
);
1174 return ERR_PTR(-ESRCH
);
1176 ctx
= &task
->perf_counter_ctx
;
1179 /* Reuse ptrace permission checks for now. */
1180 if (!ptrace_may_access(task
, PTRACE_MODE_READ
)) {
1182 return ERR_PTR(-EACCES
);
1188 static void free_counter_rcu(struct rcu_head
*head
)
1190 struct perf_counter
*counter
;
1192 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1196 static void perf_pending_sync(struct perf_counter
*counter
);
1198 static void free_counter(struct perf_counter
*counter
)
1200 perf_pending_sync(counter
);
1202 atomic_dec(&nr_counters
);
1203 if (counter
->hw_event
.mmap
)
1204 atomic_dec(&nr_mmap_tracking
);
1205 if (counter
->hw_event
.munmap
)
1206 atomic_dec(&nr_munmap_tracking
);
1207 if (counter
->hw_event
.comm
)
1208 atomic_dec(&nr_comm_tracking
);
1210 if (counter
->destroy
)
1211 counter
->destroy(counter
);
1213 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1217 * Called when the last reference to the file is gone.
1219 static int perf_release(struct inode
*inode
, struct file
*file
)
1221 struct perf_counter
*counter
= file
->private_data
;
1222 struct perf_counter_context
*ctx
= counter
->ctx
;
1224 file
->private_data
= NULL
;
1226 mutex_lock(&ctx
->mutex
);
1227 mutex_lock(&counter
->mutex
);
1229 perf_counter_remove_from_context(counter
);
1231 mutex_unlock(&counter
->mutex
);
1232 mutex_unlock(&ctx
->mutex
);
1234 free_counter(counter
);
1241 * Read the performance counter - simple non blocking version for now
1244 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1250 * Return end-of-file for a read on a counter that is in
1251 * error state (i.e. because it was pinned but it couldn't be
1252 * scheduled on to the CPU at some point).
1254 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1257 mutex_lock(&counter
->mutex
);
1258 values
[0] = perf_counter_read(counter
);
1260 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1261 values
[n
++] = counter
->total_time_enabled
+
1262 atomic64_read(&counter
->child_total_time_enabled
);
1263 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1264 values
[n
++] = counter
->total_time_running
+
1265 atomic64_read(&counter
->child_total_time_running
);
1266 mutex_unlock(&counter
->mutex
);
1268 if (count
< n
* sizeof(u64
))
1270 count
= n
* sizeof(u64
);
1272 if (copy_to_user(buf
, values
, count
))
1279 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1281 struct perf_counter
*counter
= file
->private_data
;
1283 return perf_read_hw(counter
, buf
, count
);
1286 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1288 struct perf_counter
*counter
= file
->private_data
;
1289 struct perf_mmap_data
*data
;
1290 unsigned int events
= POLL_HUP
;
1293 data
= rcu_dereference(counter
->data
);
1295 events
= atomic_xchg(&data
->poll
, 0);
1298 poll_wait(file
, &counter
->waitq
, wait
);
1303 static void perf_counter_reset(struct perf_counter
*counter
)
1305 (void)perf_counter_read(counter
);
1306 atomic64_set(&counter
->count
, 0);
1307 perf_counter_update_userpage(counter
);
1310 static void perf_counter_for_each_sibling(struct perf_counter
*counter
,
1311 void (*func
)(struct perf_counter
*))
1313 struct perf_counter_context
*ctx
= counter
->ctx
;
1314 struct perf_counter
*sibling
;
1316 spin_lock_irq(&ctx
->lock
);
1317 counter
= counter
->group_leader
;
1320 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1322 spin_unlock_irq(&ctx
->lock
);
1325 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1326 void (*func
)(struct perf_counter
*))
1328 struct perf_counter
*child
;
1330 mutex_lock(&counter
->mutex
);
1332 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1334 mutex_unlock(&counter
->mutex
);
1337 static void perf_counter_for_each(struct perf_counter
*counter
,
1338 void (*func
)(struct perf_counter
*))
1340 struct perf_counter
*child
;
1342 mutex_lock(&counter
->mutex
);
1343 perf_counter_for_each_sibling(counter
, func
);
1344 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1345 perf_counter_for_each_sibling(child
, func
);
1346 mutex_unlock(&counter
->mutex
);
1349 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1351 struct perf_counter
*counter
= file
->private_data
;
1352 void (*func
)(struct perf_counter
*);
1356 case PERF_COUNTER_IOC_ENABLE
:
1357 func
= perf_counter_enable
;
1359 case PERF_COUNTER_IOC_DISABLE
:
1360 func
= perf_counter_disable
;
1362 case PERF_COUNTER_IOC_RESET
:
1363 func
= perf_counter_reset
;
1366 case PERF_COUNTER_IOC_REFRESH
:
1367 return perf_counter_refresh(counter
, arg
);
1372 if (flags
& PERF_IOC_FLAG_GROUP
)
1373 perf_counter_for_each(counter
, func
);
1375 perf_counter_for_each_child(counter
, func
);
1381 * Callers need to ensure there can be no nesting of this function, otherwise
1382 * the seqlock logic goes bad. We can not serialize this because the arch
1383 * code calls this from NMI context.
1385 void perf_counter_update_userpage(struct perf_counter
*counter
)
1387 struct perf_mmap_data
*data
;
1388 struct perf_counter_mmap_page
*userpg
;
1391 data
= rcu_dereference(counter
->data
);
1395 userpg
= data
->user_page
;
1398 * Disable preemption so as to not let the corresponding user-space
1399 * spin too long if we get preempted.
1404 userpg
->index
= counter
->hw
.idx
;
1405 userpg
->offset
= atomic64_read(&counter
->count
);
1406 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1407 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1416 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1418 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1419 struct perf_mmap_data
*data
;
1420 int ret
= VM_FAULT_SIGBUS
;
1423 data
= rcu_dereference(counter
->data
);
1427 if (vmf
->pgoff
== 0) {
1428 vmf
->page
= virt_to_page(data
->user_page
);
1430 int nr
= vmf
->pgoff
- 1;
1432 if ((unsigned)nr
> data
->nr_pages
)
1435 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1437 get_page(vmf
->page
);
1445 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1447 struct perf_mmap_data
*data
;
1451 WARN_ON(atomic_read(&counter
->mmap_count
));
1453 size
= sizeof(struct perf_mmap_data
);
1454 size
+= nr_pages
* sizeof(void *);
1456 data
= kzalloc(size
, GFP_KERNEL
);
1460 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1461 if (!data
->user_page
)
1462 goto fail_user_page
;
1464 for (i
= 0; i
< nr_pages
; i
++) {
1465 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1466 if (!data
->data_pages
[i
])
1467 goto fail_data_pages
;
1470 data
->nr_pages
= nr_pages
;
1471 atomic_set(&data
->lock
, -1);
1473 rcu_assign_pointer(counter
->data
, data
);
1478 for (i
--; i
>= 0; i
--)
1479 free_page((unsigned long)data
->data_pages
[i
]);
1481 free_page((unsigned long)data
->user_page
);
1490 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1492 struct perf_mmap_data
*data
= container_of(rcu_head
,
1493 struct perf_mmap_data
, rcu_head
);
1496 free_page((unsigned long)data
->user_page
);
1497 for (i
= 0; i
< data
->nr_pages
; i
++)
1498 free_page((unsigned long)data
->data_pages
[i
]);
1502 static void perf_mmap_data_free(struct perf_counter
*counter
)
1504 struct perf_mmap_data
*data
= counter
->data
;
1506 WARN_ON(atomic_read(&counter
->mmap_count
));
1508 rcu_assign_pointer(counter
->data
, NULL
);
1509 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1512 static void perf_mmap_open(struct vm_area_struct
*vma
)
1514 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1516 atomic_inc(&counter
->mmap_count
);
1519 static void perf_mmap_close(struct vm_area_struct
*vma
)
1521 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1523 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
,
1524 &counter
->mmap_mutex
)) {
1525 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
1526 perf_mmap_data_free(counter
);
1527 mutex_unlock(&counter
->mmap_mutex
);
1531 static struct vm_operations_struct perf_mmap_vmops
= {
1532 .open
= perf_mmap_open
,
1533 .close
= perf_mmap_close
,
1534 .fault
= perf_mmap_fault
,
1537 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1539 struct perf_counter
*counter
= file
->private_data
;
1540 unsigned long vma_size
;
1541 unsigned long nr_pages
;
1542 unsigned long locked
, lock_limit
;
1546 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1549 vma_size
= vma
->vm_end
- vma
->vm_start
;
1550 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1553 * If we have data pages ensure they're a power-of-two number, so we
1554 * can do bitmasks instead of modulo.
1556 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1559 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1562 if (vma
->vm_pgoff
!= 0)
1565 mutex_lock(&counter
->mmap_mutex
);
1566 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1567 if (nr_pages
!= counter
->data
->nr_pages
)
1572 extra
= nr_pages
/* + 1 only account the data pages */;
1573 extra
-= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
1577 locked
= vma
->vm_mm
->locked_vm
+ extra
;
1579 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1580 lock_limit
>>= PAGE_SHIFT
;
1582 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
1587 WARN_ON(counter
->data
);
1588 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1592 atomic_set(&counter
->mmap_count
, 1);
1593 vma
->vm_mm
->locked_vm
+= extra
;
1594 counter
->data
->nr_locked
= extra
;
1596 mutex_unlock(&counter
->mmap_mutex
);
1598 vma
->vm_flags
&= ~VM_MAYWRITE
;
1599 vma
->vm_flags
|= VM_RESERVED
;
1600 vma
->vm_ops
= &perf_mmap_vmops
;
1605 static int perf_fasync(int fd
, struct file
*filp
, int on
)
1607 struct perf_counter
*counter
= filp
->private_data
;
1608 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1611 mutex_lock(&inode
->i_mutex
);
1612 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
1613 mutex_unlock(&inode
->i_mutex
);
1621 static const struct file_operations perf_fops
= {
1622 .release
= perf_release
,
1625 .unlocked_ioctl
= perf_ioctl
,
1626 .compat_ioctl
= perf_ioctl
,
1628 .fasync
= perf_fasync
,
1632 * Perf counter wakeup
1634 * If there's data, ensure we set the poll() state and publish everything
1635 * to user-space before waking everybody up.
1638 void perf_counter_wakeup(struct perf_counter
*counter
)
1640 wake_up_all(&counter
->waitq
);
1642 if (counter
->pending_kill
) {
1643 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
1644 counter
->pending_kill
= 0;
1651 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1653 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1654 * single linked list and use cmpxchg() to add entries lockless.
1657 static void perf_pending_counter(struct perf_pending_entry
*entry
)
1659 struct perf_counter
*counter
= container_of(entry
,
1660 struct perf_counter
, pending
);
1662 if (counter
->pending_disable
) {
1663 counter
->pending_disable
= 0;
1664 perf_counter_disable(counter
);
1667 if (counter
->pending_wakeup
) {
1668 counter
->pending_wakeup
= 0;
1669 perf_counter_wakeup(counter
);
1673 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1675 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
1679 static void perf_pending_queue(struct perf_pending_entry
*entry
,
1680 void (*func
)(struct perf_pending_entry
*))
1682 struct perf_pending_entry
**head
;
1684 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
1689 head
= &get_cpu_var(perf_pending_head
);
1692 entry
->next
= *head
;
1693 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
1695 set_perf_counter_pending();
1697 put_cpu_var(perf_pending_head
);
1700 static int __perf_pending_run(void)
1702 struct perf_pending_entry
*list
;
1705 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
1706 while (list
!= PENDING_TAIL
) {
1707 void (*func
)(struct perf_pending_entry
*);
1708 struct perf_pending_entry
*entry
= list
;
1715 * Ensure we observe the unqueue before we issue the wakeup,
1716 * so that we won't be waiting forever.
1717 * -- see perf_not_pending().
1728 static inline int perf_not_pending(struct perf_counter
*counter
)
1731 * If we flush on whatever cpu we run, there is a chance we don't
1735 __perf_pending_run();
1739 * Ensure we see the proper queue state before going to sleep
1740 * so that we do not miss the wakeup. -- see perf_pending_handle()
1743 return counter
->pending
.next
== NULL
;
1746 static void perf_pending_sync(struct perf_counter
*counter
)
1748 wait_event(counter
->waitq
, perf_not_pending(counter
));
1751 void perf_counter_do_pending(void)
1753 __perf_pending_run();
1757 * Callchain support -- arch specific
1760 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1769 struct perf_output_handle
{
1770 struct perf_counter
*counter
;
1771 struct perf_mmap_data
*data
;
1772 unsigned int offset
;
1777 unsigned long flags
;
1780 static void perf_output_wakeup(struct perf_output_handle
*handle
)
1782 atomic_set(&handle
->data
->poll
, POLL_IN
);
1785 handle
->counter
->pending_wakeup
= 1;
1786 perf_pending_queue(&handle
->counter
->pending
,
1787 perf_pending_counter
);
1789 perf_counter_wakeup(handle
->counter
);
1793 * Curious locking construct.
1795 * We need to ensure a later event doesn't publish a head when a former
1796 * event isn't done writing. However since we need to deal with NMIs we
1797 * cannot fully serialize things.
1799 * What we do is serialize between CPUs so we only have to deal with NMI
1800 * nesting on a single CPU.
1802 * We only publish the head (and generate a wakeup) when the outer-most
1805 static void perf_output_lock(struct perf_output_handle
*handle
)
1807 struct perf_mmap_data
*data
= handle
->data
;
1812 local_irq_save(handle
->flags
);
1813 cpu
= smp_processor_id();
1815 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
1818 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
1824 static void perf_output_unlock(struct perf_output_handle
*handle
)
1826 struct perf_mmap_data
*data
= handle
->data
;
1829 data
->done_head
= data
->head
;
1831 if (!handle
->locked
)
1836 * The xchg implies a full barrier that ensures all writes are done
1837 * before we publish the new head, matched by a rmb() in userspace when
1838 * reading this position.
1840 while ((head
= atomic_xchg(&data
->done_head
, 0)))
1841 data
->user_page
->data_head
= head
;
1844 * NMI can happen here, which means we can miss a done_head update.
1847 cpu
= atomic_xchg(&data
->lock
, -1);
1848 WARN_ON_ONCE(cpu
!= smp_processor_id());
1851 * Therefore we have to validate we did not indeed do so.
1853 if (unlikely(atomic_read(&data
->done_head
))) {
1855 * Since we had it locked, we can lock it again.
1857 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
1863 if (atomic_xchg(&data
->wakeup
, 0))
1864 perf_output_wakeup(handle
);
1866 local_irq_restore(handle
->flags
);
1869 static int perf_output_begin(struct perf_output_handle
*handle
,
1870 struct perf_counter
*counter
, unsigned int size
,
1871 int nmi
, int overflow
)
1873 struct perf_mmap_data
*data
;
1874 unsigned int offset
, head
;
1877 * For inherited counters we send all the output towards the parent.
1879 if (counter
->parent
)
1880 counter
= counter
->parent
;
1883 data
= rcu_dereference(counter
->data
);
1887 handle
->data
= data
;
1888 handle
->counter
= counter
;
1890 handle
->overflow
= overflow
;
1892 if (!data
->nr_pages
)
1895 perf_output_lock(handle
);
1898 offset
= head
= atomic_read(&data
->head
);
1900 } while (atomic_cmpxchg(&data
->head
, offset
, head
) != offset
);
1902 handle
->offset
= offset
;
1903 handle
->head
= head
;
1905 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
1906 atomic_set(&data
->wakeup
, 1);
1911 perf_output_wakeup(handle
);
1918 static void perf_output_copy(struct perf_output_handle
*handle
,
1919 void *buf
, unsigned int len
)
1921 unsigned int pages_mask
;
1922 unsigned int offset
;
1926 offset
= handle
->offset
;
1927 pages_mask
= handle
->data
->nr_pages
- 1;
1928 pages
= handle
->data
->data_pages
;
1931 unsigned int page_offset
;
1934 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
1935 page_offset
= offset
& (PAGE_SIZE
- 1);
1936 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
1938 memcpy(pages
[nr
] + page_offset
, buf
, size
);
1945 handle
->offset
= offset
;
1948 * Check we didn't copy past our reservation window, taking the
1949 * possible unsigned int wrap into account.
1951 WARN_ON_ONCE(((int)(handle
->head
- handle
->offset
)) < 0);
1954 #define perf_output_put(handle, x) \
1955 perf_output_copy((handle), &(x), sizeof(x))
1957 static void perf_output_end(struct perf_output_handle
*handle
)
1959 struct perf_counter
*counter
= handle
->counter
;
1960 struct perf_mmap_data
*data
= handle
->data
;
1962 int wakeup_events
= counter
->hw_event
.wakeup_events
;
1964 if (handle
->overflow
&& wakeup_events
) {
1965 int events
= atomic_inc_return(&data
->events
);
1966 if (events
>= wakeup_events
) {
1967 atomic_sub(wakeup_events
, &data
->events
);
1968 atomic_set(&data
->wakeup
, 1);
1972 perf_output_unlock(handle
);
1976 static void perf_counter_output(struct perf_counter
*counter
,
1977 int nmi
, struct pt_regs
*regs
, u64 addr
)
1980 u64 record_type
= counter
->hw_event
.record_type
;
1981 struct perf_output_handle handle
;
1982 struct perf_event_header header
;
1991 struct perf_callchain_entry
*callchain
= NULL
;
1992 int callchain_size
= 0;
1999 header
.size
= sizeof(header
);
2001 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
2002 header
.misc
|= user_mode(regs
) ?
2003 PERF_EVENT_MISC_USER
: PERF_EVENT_MISC_KERNEL
;
2005 if (record_type
& PERF_RECORD_IP
) {
2006 ip
= instruction_pointer(regs
);
2007 header
.type
|= PERF_RECORD_IP
;
2008 header
.size
+= sizeof(ip
);
2011 if (record_type
& PERF_RECORD_TID
) {
2012 /* namespace issues */
2013 tid_entry
.pid
= current
->group_leader
->pid
;
2014 tid_entry
.tid
= current
->pid
;
2016 header
.type
|= PERF_RECORD_TID
;
2017 header
.size
+= sizeof(tid_entry
);
2020 if (record_type
& PERF_RECORD_TIME
) {
2022 * Maybe do better on x86 and provide cpu_clock_nmi()
2024 time
= sched_clock();
2026 header
.type
|= PERF_RECORD_TIME
;
2027 header
.size
+= sizeof(u64
);
2030 if (record_type
& PERF_RECORD_ADDR
) {
2031 header
.type
|= PERF_RECORD_ADDR
;
2032 header
.size
+= sizeof(u64
);
2035 if (record_type
& PERF_RECORD_CONFIG
) {
2036 header
.type
|= PERF_RECORD_CONFIG
;
2037 header
.size
+= sizeof(u64
);
2040 if (record_type
& PERF_RECORD_CPU
) {
2041 header
.type
|= PERF_RECORD_CPU
;
2042 header
.size
+= sizeof(cpu_entry
);
2044 cpu_entry
.cpu
= raw_smp_processor_id();
2047 if (record_type
& PERF_RECORD_GROUP
) {
2048 header
.type
|= PERF_RECORD_GROUP
;
2049 header
.size
+= sizeof(u64
) +
2050 counter
->nr_siblings
* sizeof(group_entry
);
2053 if (record_type
& PERF_RECORD_CALLCHAIN
) {
2054 callchain
= perf_callchain(regs
);
2057 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2059 header
.type
|= PERF_RECORD_CALLCHAIN
;
2060 header
.size
+= callchain_size
;
2064 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2068 perf_output_put(&handle
, header
);
2070 if (record_type
& PERF_RECORD_IP
)
2071 perf_output_put(&handle
, ip
);
2073 if (record_type
& PERF_RECORD_TID
)
2074 perf_output_put(&handle
, tid_entry
);
2076 if (record_type
& PERF_RECORD_TIME
)
2077 perf_output_put(&handle
, time
);
2079 if (record_type
& PERF_RECORD_ADDR
)
2080 perf_output_put(&handle
, addr
);
2082 if (record_type
& PERF_RECORD_CONFIG
)
2083 perf_output_put(&handle
, counter
->hw_event
.config
);
2085 if (record_type
& PERF_RECORD_CPU
)
2086 perf_output_put(&handle
, cpu_entry
);
2089 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2091 if (record_type
& PERF_RECORD_GROUP
) {
2092 struct perf_counter
*leader
, *sub
;
2093 u64 nr
= counter
->nr_siblings
;
2095 perf_output_put(&handle
, nr
);
2097 leader
= counter
->group_leader
;
2098 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2100 sub
->pmu
->read(sub
);
2102 group_entry
.event
= sub
->hw_event
.config
;
2103 group_entry
.counter
= atomic64_read(&sub
->count
);
2105 perf_output_put(&handle
, group_entry
);
2110 perf_output_copy(&handle
, callchain
, callchain_size
);
2112 perf_output_end(&handle
);
2119 struct perf_comm_event
{
2120 struct task_struct
*task
;
2125 struct perf_event_header header
;
2132 static void perf_counter_comm_output(struct perf_counter
*counter
,
2133 struct perf_comm_event
*comm_event
)
2135 struct perf_output_handle handle
;
2136 int size
= comm_event
->event
.header
.size
;
2137 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2142 perf_output_put(&handle
, comm_event
->event
);
2143 perf_output_copy(&handle
, comm_event
->comm
,
2144 comm_event
->comm_size
);
2145 perf_output_end(&handle
);
2148 static int perf_counter_comm_match(struct perf_counter
*counter
,
2149 struct perf_comm_event
*comm_event
)
2151 if (counter
->hw_event
.comm
&&
2152 comm_event
->event
.header
.type
== PERF_EVENT_COMM
)
2158 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2159 struct perf_comm_event
*comm_event
)
2161 struct perf_counter
*counter
;
2163 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2167 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2168 if (perf_counter_comm_match(counter
, comm_event
))
2169 perf_counter_comm_output(counter
, comm_event
);
2174 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2176 struct perf_cpu_context
*cpuctx
;
2178 char *comm
= comm_event
->task
->comm
;
2180 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2182 comm_event
->comm
= comm
;
2183 comm_event
->comm_size
= size
;
2185 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2187 cpuctx
= &get_cpu_var(perf_cpu_context
);
2188 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2189 put_cpu_var(perf_cpu_context
);
2191 perf_counter_comm_ctx(¤t
->perf_counter_ctx
, comm_event
);
2194 void perf_counter_comm(struct task_struct
*task
)
2196 struct perf_comm_event comm_event
;
2198 if (!atomic_read(&nr_comm_tracking
))
2201 comm_event
= (struct perf_comm_event
){
2204 .header
= { .type
= PERF_EVENT_COMM
, },
2205 .pid
= task
->group_leader
->pid
,
2210 perf_counter_comm_event(&comm_event
);
2217 struct perf_mmap_event
{
2223 struct perf_event_header header
;
2233 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2234 struct perf_mmap_event
*mmap_event
)
2236 struct perf_output_handle handle
;
2237 int size
= mmap_event
->event
.header
.size
;
2238 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2243 perf_output_put(&handle
, mmap_event
->event
);
2244 perf_output_copy(&handle
, mmap_event
->file_name
,
2245 mmap_event
->file_size
);
2246 perf_output_end(&handle
);
2249 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2250 struct perf_mmap_event
*mmap_event
)
2252 if (counter
->hw_event
.mmap
&&
2253 mmap_event
->event
.header
.type
== PERF_EVENT_MMAP
)
2256 if (counter
->hw_event
.munmap
&&
2257 mmap_event
->event
.header
.type
== PERF_EVENT_MUNMAP
)
2263 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2264 struct perf_mmap_event
*mmap_event
)
2266 struct perf_counter
*counter
;
2268 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2272 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2273 if (perf_counter_mmap_match(counter
, mmap_event
))
2274 perf_counter_mmap_output(counter
, mmap_event
);
2279 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2281 struct perf_cpu_context
*cpuctx
;
2282 struct file
*file
= mmap_event
->file
;
2289 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2291 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2294 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2296 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2300 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2305 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2307 mmap_event
->file_name
= name
;
2308 mmap_event
->file_size
= size
;
2310 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2312 cpuctx
= &get_cpu_var(perf_cpu_context
);
2313 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2314 put_cpu_var(perf_cpu_context
);
2316 perf_counter_mmap_ctx(¤t
->perf_counter_ctx
, mmap_event
);
2321 void perf_counter_mmap(unsigned long addr
, unsigned long len
,
2322 unsigned long pgoff
, struct file
*file
)
2324 struct perf_mmap_event mmap_event
;
2326 if (!atomic_read(&nr_mmap_tracking
))
2329 mmap_event
= (struct perf_mmap_event
){
2332 .header
= { .type
= PERF_EVENT_MMAP
, },
2333 .pid
= current
->group_leader
->pid
,
2334 .tid
= current
->pid
,
2341 perf_counter_mmap_event(&mmap_event
);
2344 void perf_counter_munmap(unsigned long addr
, unsigned long len
,
2345 unsigned long pgoff
, struct file
*file
)
2347 struct perf_mmap_event mmap_event
;
2349 if (!atomic_read(&nr_munmap_tracking
))
2352 mmap_event
= (struct perf_mmap_event
){
2355 .header
= { .type
= PERF_EVENT_MUNMAP
, },
2356 .pid
= current
->group_leader
->pid
,
2357 .tid
= current
->pid
,
2364 perf_counter_mmap_event(&mmap_event
);
2368 * Generic counter overflow handling.
2371 int perf_counter_overflow(struct perf_counter
*counter
,
2372 int nmi
, struct pt_regs
*regs
, u64 addr
)
2374 int events
= atomic_read(&counter
->event_limit
);
2378 * XXX event_limit might not quite work as expected on inherited
2382 counter
->pending_kill
= POLL_IN
;
2383 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
2385 counter
->pending_kill
= POLL_HUP
;
2387 counter
->pending_disable
= 1;
2388 perf_pending_queue(&counter
->pending
,
2389 perf_pending_counter
);
2391 perf_counter_disable(counter
);
2394 perf_counter_output(counter
, nmi
, regs
, addr
);
2399 * Generic software counter infrastructure
2402 static void perf_swcounter_update(struct perf_counter
*counter
)
2404 struct hw_perf_counter
*hwc
= &counter
->hw
;
2409 prev
= atomic64_read(&hwc
->prev_count
);
2410 now
= atomic64_read(&hwc
->count
);
2411 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2416 atomic64_add(delta
, &counter
->count
);
2417 atomic64_sub(delta
, &hwc
->period_left
);
2420 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2422 struct hw_perf_counter
*hwc
= &counter
->hw
;
2423 s64 left
= atomic64_read(&hwc
->period_left
);
2424 s64 period
= hwc
->irq_period
;
2426 if (unlikely(left
<= -period
)) {
2428 atomic64_set(&hwc
->period_left
, left
);
2431 if (unlikely(left
<= 0)) {
2433 atomic64_add(period
, &hwc
->period_left
);
2436 atomic64_set(&hwc
->prev_count
, -left
);
2437 atomic64_set(&hwc
->count
, -left
);
2440 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2442 enum hrtimer_restart ret
= HRTIMER_RESTART
;
2443 struct perf_counter
*counter
;
2444 struct pt_regs
*regs
;
2446 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
2447 counter
->pmu
->read(counter
);
2449 regs
= get_irq_regs();
2451 * In case we exclude kernel IPs or are somehow not in interrupt
2452 * context, provide the next best thing, the user IP.
2454 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
2455 !counter
->hw_event
.exclude_user
)
2456 regs
= task_pt_regs(current
);
2459 if (perf_counter_overflow(counter
, 0, regs
, 0))
2460 ret
= HRTIMER_NORESTART
;
2463 hrtimer_forward_now(hrtimer
, ns_to_ktime(counter
->hw
.irq_period
));
2468 static void perf_swcounter_overflow(struct perf_counter
*counter
,
2469 int nmi
, struct pt_regs
*regs
, u64 addr
)
2471 perf_swcounter_update(counter
);
2472 perf_swcounter_set_period(counter
);
2473 if (perf_counter_overflow(counter
, nmi
, regs
, addr
))
2474 /* soft-disable the counter */
2479 static int perf_swcounter_match(struct perf_counter
*counter
,
2480 enum perf_event_types type
,
2481 u32 event
, struct pt_regs
*regs
)
2483 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
2486 if (perf_event_raw(&counter
->hw_event
))
2489 if (perf_event_type(&counter
->hw_event
) != type
)
2492 if (perf_event_id(&counter
->hw_event
) != event
)
2495 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
2498 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
2504 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
2505 int nmi
, struct pt_regs
*regs
, u64 addr
)
2507 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
2508 if (counter
->hw
.irq_period
&& !neg
)
2509 perf_swcounter_overflow(counter
, nmi
, regs
, addr
);
2512 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
2513 enum perf_event_types type
, u32 event
,
2514 u64 nr
, int nmi
, struct pt_regs
*regs
,
2517 struct perf_counter
*counter
;
2519 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2523 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2524 if (perf_swcounter_match(counter
, type
, event
, regs
))
2525 perf_swcounter_add(counter
, nr
, nmi
, regs
, addr
);
2530 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
2533 return &cpuctx
->recursion
[3];
2536 return &cpuctx
->recursion
[2];
2539 return &cpuctx
->recursion
[1];
2541 return &cpuctx
->recursion
[0];
2544 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
2545 u64 nr
, int nmi
, struct pt_regs
*regs
,
2548 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
2549 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
2557 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
2558 nr
, nmi
, regs
, addr
);
2559 if (cpuctx
->task_ctx
) {
2560 perf_swcounter_ctx_event(cpuctx
->task_ctx
, type
, event
,
2561 nr
, nmi
, regs
, addr
);
2568 put_cpu_var(perf_cpu_context
);
2572 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
2574 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
, addr
);
2577 static void perf_swcounter_read(struct perf_counter
*counter
)
2579 perf_swcounter_update(counter
);
2582 static int perf_swcounter_enable(struct perf_counter
*counter
)
2584 perf_swcounter_set_period(counter
);
2588 static void perf_swcounter_disable(struct perf_counter
*counter
)
2590 perf_swcounter_update(counter
);
2593 static const struct pmu perf_ops_generic
= {
2594 .enable
= perf_swcounter_enable
,
2595 .disable
= perf_swcounter_disable
,
2596 .read
= perf_swcounter_read
,
2600 * Software counter: cpu wall time clock
2603 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
2605 int cpu
= raw_smp_processor_id();
2609 now
= cpu_clock(cpu
);
2610 prev
= atomic64_read(&counter
->hw
.prev_count
);
2611 atomic64_set(&counter
->hw
.prev_count
, now
);
2612 atomic64_add(now
- prev
, &counter
->count
);
2615 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
2617 struct hw_perf_counter
*hwc
= &counter
->hw
;
2618 int cpu
= raw_smp_processor_id();
2620 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
2621 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2622 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2623 if (hwc
->irq_period
) {
2624 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2625 ns_to_ktime(hwc
->irq_period
), 0,
2626 HRTIMER_MODE_REL
, 0);
2632 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
2634 hrtimer_cancel(&counter
->hw
.hrtimer
);
2635 cpu_clock_perf_counter_update(counter
);
2638 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
2640 cpu_clock_perf_counter_update(counter
);
2643 static const struct pmu perf_ops_cpu_clock
= {
2644 .enable
= cpu_clock_perf_counter_enable
,
2645 .disable
= cpu_clock_perf_counter_disable
,
2646 .read
= cpu_clock_perf_counter_read
,
2650 * Software counter: task time clock
2653 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
2658 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
2660 atomic64_add(delta
, &counter
->count
);
2663 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
2665 struct hw_perf_counter
*hwc
= &counter
->hw
;
2668 now
= counter
->ctx
->time
;
2670 atomic64_set(&hwc
->prev_count
, now
);
2671 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2672 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2673 if (hwc
->irq_period
) {
2674 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2675 ns_to_ktime(hwc
->irq_period
), 0,
2676 HRTIMER_MODE_REL
, 0);
2682 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
2684 hrtimer_cancel(&counter
->hw
.hrtimer
);
2685 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
2689 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
2694 update_context_time(counter
->ctx
);
2695 time
= counter
->ctx
->time
;
2697 u64 now
= perf_clock();
2698 u64 delta
= now
- counter
->ctx
->timestamp
;
2699 time
= counter
->ctx
->time
+ delta
;
2702 task_clock_perf_counter_update(counter
, time
);
2705 static const struct pmu perf_ops_task_clock
= {
2706 .enable
= task_clock_perf_counter_enable
,
2707 .disable
= task_clock_perf_counter_disable
,
2708 .read
= task_clock_perf_counter_read
,
2712 * Software counter: cpu migrations
2715 static inline u64
get_cpu_migrations(struct perf_counter
*counter
)
2717 struct task_struct
*curr
= counter
->ctx
->task
;
2720 return curr
->se
.nr_migrations
;
2721 return cpu_nr_migrations(smp_processor_id());
2724 static void cpu_migrations_perf_counter_update(struct perf_counter
*counter
)
2729 prev
= atomic64_read(&counter
->hw
.prev_count
);
2730 now
= get_cpu_migrations(counter
);
2732 atomic64_set(&counter
->hw
.prev_count
, now
);
2736 atomic64_add(delta
, &counter
->count
);
2739 static void cpu_migrations_perf_counter_read(struct perf_counter
*counter
)
2741 cpu_migrations_perf_counter_update(counter
);
2744 static int cpu_migrations_perf_counter_enable(struct perf_counter
*counter
)
2746 if (counter
->prev_state
<= PERF_COUNTER_STATE_OFF
)
2747 atomic64_set(&counter
->hw
.prev_count
,
2748 get_cpu_migrations(counter
));
2752 static void cpu_migrations_perf_counter_disable(struct perf_counter
*counter
)
2754 cpu_migrations_perf_counter_update(counter
);
2757 static const struct pmu perf_ops_cpu_migrations
= {
2758 .enable
= cpu_migrations_perf_counter_enable
,
2759 .disable
= cpu_migrations_perf_counter_disable
,
2760 .read
= cpu_migrations_perf_counter_read
,
2763 #ifdef CONFIG_EVENT_PROFILE
2764 void perf_tpcounter_event(int event_id
)
2766 struct pt_regs
*regs
= get_irq_regs();
2769 regs
= task_pt_regs(current
);
2771 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
, 0);
2773 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
2775 extern int ftrace_profile_enable(int);
2776 extern void ftrace_profile_disable(int);
2778 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
2780 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
2783 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
2785 int event_id
= perf_event_id(&counter
->hw_event
);
2788 ret
= ftrace_profile_enable(event_id
);
2792 counter
->destroy
= tp_perf_counter_destroy
;
2793 counter
->hw
.irq_period
= counter
->hw_event
.irq_period
;
2795 return &perf_ops_generic
;
2798 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
2804 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
2806 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
2807 const struct pmu
*pmu
= NULL
;
2808 struct hw_perf_counter
*hwc
= &counter
->hw
;
2811 * Software counters (currently) can't in general distinguish
2812 * between user, kernel and hypervisor events.
2813 * However, context switches and cpu migrations are considered
2814 * to be kernel events, and page faults are never hypervisor
2817 switch (perf_event_id(&counter
->hw_event
)) {
2818 case PERF_COUNT_CPU_CLOCK
:
2819 pmu
= &perf_ops_cpu_clock
;
2821 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2822 hw_event
->irq_period
= 10000;
2824 case PERF_COUNT_TASK_CLOCK
:
2826 * If the user instantiates this as a per-cpu counter,
2827 * use the cpu_clock counter instead.
2829 if (counter
->ctx
->task
)
2830 pmu
= &perf_ops_task_clock
;
2832 pmu
= &perf_ops_cpu_clock
;
2834 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2835 hw_event
->irq_period
= 10000;
2837 case PERF_COUNT_PAGE_FAULTS
:
2838 case PERF_COUNT_PAGE_FAULTS_MIN
:
2839 case PERF_COUNT_PAGE_FAULTS_MAJ
:
2840 case PERF_COUNT_CONTEXT_SWITCHES
:
2841 pmu
= &perf_ops_generic
;
2843 case PERF_COUNT_CPU_MIGRATIONS
:
2844 if (!counter
->hw_event
.exclude_kernel
)
2845 pmu
= &perf_ops_cpu_migrations
;
2850 hwc
->irq_period
= hw_event
->irq_period
;
2856 * Allocate and initialize a counter structure
2858 static struct perf_counter
*
2859 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
2861 struct perf_counter_context
*ctx
,
2862 struct perf_counter
*group_leader
,
2865 const struct pmu
*pmu
;
2866 struct perf_counter
*counter
;
2869 counter
= kzalloc(sizeof(*counter
), gfpflags
);
2871 return ERR_PTR(-ENOMEM
);
2874 * Single counters are their own group leaders, with an
2875 * empty sibling list:
2878 group_leader
= counter
;
2880 mutex_init(&counter
->mutex
);
2881 INIT_LIST_HEAD(&counter
->list_entry
);
2882 INIT_LIST_HEAD(&counter
->event_entry
);
2883 INIT_LIST_HEAD(&counter
->sibling_list
);
2884 init_waitqueue_head(&counter
->waitq
);
2886 mutex_init(&counter
->mmap_mutex
);
2888 INIT_LIST_HEAD(&counter
->child_list
);
2891 counter
->hw_event
= *hw_event
;
2892 counter
->group_leader
= group_leader
;
2893 counter
->pmu
= NULL
;
2896 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2897 if (hw_event
->disabled
)
2898 counter
->state
= PERF_COUNTER_STATE_OFF
;
2903 * we currently do not support PERF_RECORD_GROUP on inherited counters
2905 if (hw_event
->inherit
&& (hw_event
->record_type
& PERF_RECORD_GROUP
))
2908 if (perf_event_raw(hw_event
)) {
2909 pmu
= hw_perf_counter_init(counter
);
2913 switch (perf_event_type(hw_event
)) {
2914 case PERF_TYPE_HARDWARE
:
2915 pmu
= hw_perf_counter_init(counter
);
2918 case PERF_TYPE_SOFTWARE
:
2919 pmu
= sw_perf_counter_init(counter
);
2922 case PERF_TYPE_TRACEPOINT
:
2923 pmu
= tp_perf_counter_init(counter
);
2930 else if (IS_ERR(pmu
))
2935 return ERR_PTR(err
);
2940 atomic_inc(&nr_counters
);
2941 if (counter
->hw_event
.mmap
)
2942 atomic_inc(&nr_mmap_tracking
);
2943 if (counter
->hw_event
.munmap
)
2944 atomic_inc(&nr_munmap_tracking
);
2945 if (counter
->hw_event
.comm
)
2946 atomic_inc(&nr_comm_tracking
);
2952 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2954 * @hw_event_uptr: event type attributes for monitoring/sampling
2957 * @group_fd: group leader counter fd
2959 SYSCALL_DEFINE5(perf_counter_open
,
2960 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
2961 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
2963 struct perf_counter
*counter
, *group_leader
;
2964 struct perf_counter_hw_event hw_event
;
2965 struct perf_counter_context
*ctx
;
2966 struct file
*counter_file
= NULL
;
2967 struct file
*group_file
= NULL
;
2968 int fput_needed
= 0;
2969 int fput_needed2
= 0;
2972 /* for future expandability... */
2976 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
2980 * Get the target context (task or percpu):
2982 ctx
= find_get_context(pid
, cpu
);
2984 return PTR_ERR(ctx
);
2987 * Look up the group leader (we will attach this counter to it):
2989 group_leader
= NULL
;
2990 if (group_fd
!= -1) {
2992 group_file
= fget_light(group_fd
, &fput_needed
);
2994 goto err_put_context
;
2995 if (group_file
->f_op
!= &perf_fops
)
2996 goto err_put_context
;
2998 group_leader
= group_file
->private_data
;
3000 * Do not allow a recursive hierarchy (this new sibling
3001 * becoming part of another group-sibling):
3003 if (group_leader
->group_leader
!= group_leader
)
3004 goto err_put_context
;
3006 * Do not allow to attach to a group in a different
3007 * task or CPU context:
3009 if (group_leader
->ctx
!= ctx
)
3010 goto err_put_context
;
3012 * Only a group leader can be exclusive or pinned
3014 if (hw_event
.exclusive
|| hw_event
.pinned
)
3015 goto err_put_context
;
3018 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
3020 ret
= PTR_ERR(counter
);
3021 if (IS_ERR(counter
))
3022 goto err_put_context
;
3024 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
3026 goto err_free_put_context
;
3028 counter_file
= fget_light(ret
, &fput_needed2
);
3030 goto err_free_put_context
;
3032 counter
->filp
= counter_file
;
3033 mutex_lock(&ctx
->mutex
);
3034 perf_install_in_context(ctx
, counter
, cpu
);
3035 mutex_unlock(&ctx
->mutex
);
3037 fput_light(counter_file
, fput_needed2
);
3040 fput_light(group_file
, fput_needed
);
3044 err_free_put_context
:
3054 * Initialize the perf_counter context in a task_struct:
3057 __perf_counter_init_context(struct perf_counter_context
*ctx
,
3058 struct task_struct
*task
)
3060 memset(ctx
, 0, sizeof(*ctx
));
3061 spin_lock_init(&ctx
->lock
);
3062 mutex_init(&ctx
->mutex
);
3063 INIT_LIST_HEAD(&ctx
->counter_list
);
3064 INIT_LIST_HEAD(&ctx
->event_list
);
3069 * inherit a counter from parent task to child task:
3071 static struct perf_counter
*
3072 inherit_counter(struct perf_counter
*parent_counter
,
3073 struct task_struct
*parent
,
3074 struct perf_counter_context
*parent_ctx
,
3075 struct task_struct
*child
,
3076 struct perf_counter
*group_leader
,
3077 struct perf_counter_context
*child_ctx
)
3079 struct perf_counter
*child_counter
;
3082 * Instead of creating recursive hierarchies of counters,
3083 * we link inherited counters back to the original parent,
3084 * which has a filp for sure, which we use as the reference
3087 if (parent_counter
->parent
)
3088 parent_counter
= parent_counter
->parent
;
3090 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
3091 parent_counter
->cpu
, child_ctx
,
3092 group_leader
, GFP_KERNEL
);
3093 if (IS_ERR(child_counter
))
3094 return child_counter
;
3097 * Link it up in the child's context:
3099 child_counter
->task
= child
;
3100 add_counter_to_ctx(child_counter
, child_ctx
);
3102 child_counter
->parent
= parent_counter
;
3104 * inherit into child's child as well:
3106 child_counter
->hw_event
.inherit
= 1;
3109 * Get a reference to the parent filp - we will fput it
3110 * when the child counter exits. This is safe to do because
3111 * we are in the parent and we know that the filp still
3112 * exists and has a nonzero count:
3114 atomic_long_inc(&parent_counter
->filp
->f_count
);
3117 * Link this into the parent counter's child list
3119 mutex_lock(&parent_counter
->mutex
);
3120 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3123 * Make the child state follow the state of the parent counter,
3124 * not its hw_event.disabled bit. We hold the parent's mutex,
3125 * so we won't race with perf_counter_{en,dis}able_family.
3127 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3128 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3130 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3132 mutex_unlock(&parent_counter
->mutex
);
3134 return child_counter
;
3137 static int inherit_group(struct perf_counter
*parent_counter
,
3138 struct task_struct
*parent
,
3139 struct perf_counter_context
*parent_ctx
,
3140 struct task_struct
*child
,
3141 struct perf_counter_context
*child_ctx
)
3143 struct perf_counter
*leader
;
3144 struct perf_counter
*sub
;
3145 struct perf_counter
*child_ctr
;
3147 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3148 child
, NULL
, child_ctx
);
3150 return PTR_ERR(leader
);
3151 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3152 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3153 child
, leader
, child_ctx
);
3154 if (IS_ERR(child_ctr
))
3155 return PTR_ERR(child_ctr
);
3160 static void sync_child_counter(struct perf_counter
*child_counter
,
3161 struct perf_counter
*parent_counter
)
3163 u64 parent_val
, child_val
;
3165 parent_val
= atomic64_read(&parent_counter
->count
);
3166 child_val
= atomic64_read(&child_counter
->count
);
3169 * Add back the child's count to the parent's count:
3171 atomic64_add(child_val
, &parent_counter
->count
);
3172 atomic64_add(child_counter
->total_time_enabled
,
3173 &parent_counter
->child_total_time_enabled
);
3174 atomic64_add(child_counter
->total_time_running
,
3175 &parent_counter
->child_total_time_running
);
3178 * Remove this counter from the parent's list
3180 mutex_lock(&parent_counter
->mutex
);
3181 list_del_init(&child_counter
->child_list
);
3182 mutex_unlock(&parent_counter
->mutex
);
3185 * Release the parent counter, if this was the last
3188 fput(parent_counter
->filp
);
3192 __perf_counter_exit_task(struct task_struct
*child
,
3193 struct perf_counter
*child_counter
,
3194 struct perf_counter_context
*child_ctx
)
3196 struct perf_counter
*parent_counter
;
3197 struct perf_counter
*sub
, *tmp
;
3200 * If we do not self-reap then we have to wait for the
3201 * child task to unschedule (it will happen for sure),
3202 * so that its counter is at its final count. (This
3203 * condition triggers rarely - child tasks usually get
3204 * off their CPU before the parent has a chance to
3205 * get this far into the reaping action)
3207 if (child
!= current
) {
3208 wait_task_inactive(child
, 0);
3209 list_del_init(&child_counter
->list_entry
);
3210 update_counter_times(child_counter
);
3212 struct perf_cpu_context
*cpuctx
;
3213 unsigned long flags
;
3216 * Disable and unlink this counter.
3218 * Be careful about zapping the list - IRQ/NMI context
3219 * could still be processing it:
3221 local_irq_save(flags
);
3224 cpuctx
= &__get_cpu_var(perf_cpu_context
);
3226 group_sched_out(child_counter
, cpuctx
, child_ctx
);
3227 update_counter_times(child_counter
);
3229 list_del_init(&child_counter
->list_entry
);
3231 child_ctx
->nr_counters
--;
3234 local_irq_restore(flags
);
3237 parent_counter
= child_counter
->parent
;
3239 * It can happen that parent exits first, and has counters
3240 * that are still around due to the child reference. These
3241 * counters need to be zapped - but otherwise linger.
3243 if (parent_counter
) {
3244 sync_child_counter(child_counter
, parent_counter
);
3245 list_for_each_entry_safe(sub
, tmp
, &child_counter
->sibling_list
,
3248 sync_child_counter(sub
, sub
->parent
);
3252 free_counter(child_counter
);
3257 * When a child task exits, feed back counter values to parent counters.
3259 * Note: we may be running in child context, but the PID is not hashed
3260 * anymore so new counters will not be added.
3262 void perf_counter_exit_task(struct task_struct
*child
)
3264 struct perf_counter
*child_counter
, *tmp
;
3265 struct perf_counter_context
*child_ctx
;
3267 child_ctx
= &child
->perf_counter_ctx
;
3269 if (likely(!child_ctx
->nr_counters
))
3272 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
3274 __perf_counter_exit_task(child
, child_counter
, child_ctx
);
3278 * Initialize the perf_counter context in task_struct
3280 void perf_counter_init_task(struct task_struct
*child
)
3282 struct perf_counter_context
*child_ctx
, *parent_ctx
;
3283 struct perf_counter
*counter
;
3284 struct task_struct
*parent
= current
;
3286 child_ctx
= &child
->perf_counter_ctx
;
3287 parent_ctx
= &parent
->perf_counter_ctx
;
3289 __perf_counter_init_context(child_ctx
, child
);
3292 * This is executed from the parent task context, so inherit
3293 * counters that have been marked for cloning:
3296 if (likely(!parent_ctx
->nr_counters
))
3300 * Lock the parent list. No need to lock the child - not PID
3301 * hashed yet and not running, so nobody can access it.
3303 mutex_lock(&parent_ctx
->mutex
);
3306 * We dont have to disable NMIs - we are only looking at
3307 * the list, not manipulating it:
3309 list_for_each_entry(counter
, &parent_ctx
->counter_list
, list_entry
) {
3310 if (!counter
->hw_event
.inherit
)
3313 if (inherit_group(counter
, parent
,
3314 parent_ctx
, child
, child_ctx
))
3318 mutex_unlock(&parent_ctx
->mutex
);
3321 static void __cpuinit
perf_counter_init_cpu(int cpu
)
3323 struct perf_cpu_context
*cpuctx
;
3325 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3326 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
3328 spin_lock(&perf_resource_lock
);
3329 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
3330 spin_unlock(&perf_resource_lock
);
3332 hw_perf_counter_setup(cpu
);
3335 #ifdef CONFIG_HOTPLUG_CPU
3336 static void __perf_counter_exit_cpu(void *info
)
3338 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3339 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3340 struct perf_counter
*counter
, *tmp
;
3342 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
3343 __perf_counter_remove_from_context(counter
);
3345 static void perf_counter_exit_cpu(int cpu
)
3347 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3348 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3350 mutex_lock(&ctx
->mutex
);
3351 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
3352 mutex_unlock(&ctx
->mutex
);
3355 static inline void perf_counter_exit_cpu(int cpu
) { }
3358 static int __cpuinit
3359 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
3361 unsigned int cpu
= (long)hcpu
;
3365 case CPU_UP_PREPARE
:
3366 case CPU_UP_PREPARE_FROZEN
:
3367 perf_counter_init_cpu(cpu
);
3370 case CPU_DOWN_PREPARE
:
3371 case CPU_DOWN_PREPARE_FROZEN
:
3372 perf_counter_exit_cpu(cpu
);
3382 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
3383 .notifier_call
= perf_cpu_notify
,
3386 void __init
perf_counter_init(void)
3388 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
3389 (void *)(long)smp_processor_id());
3390 register_cpu_notifier(&perf_cpu_nb
);
3393 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
3395 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
3399 perf_set_reserve_percpu(struct sysdev_class
*class,
3403 struct perf_cpu_context
*cpuctx
;
3407 err
= strict_strtoul(buf
, 10, &val
);
3410 if (val
> perf_max_counters
)
3413 spin_lock(&perf_resource_lock
);
3414 perf_reserved_percpu
= val
;
3415 for_each_online_cpu(cpu
) {
3416 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3417 spin_lock_irq(&cpuctx
->ctx
.lock
);
3418 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
3419 perf_max_counters
- perf_reserved_percpu
);
3420 cpuctx
->max_pertask
= mpt
;
3421 spin_unlock_irq(&cpuctx
->ctx
.lock
);
3423 spin_unlock(&perf_resource_lock
);
3428 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
3430 return sprintf(buf
, "%d\n", perf_overcommit
);
3434 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
3439 err
= strict_strtoul(buf
, 10, &val
);
3445 spin_lock(&perf_resource_lock
);
3446 perf_overcommit
= val
;
3447 spin_unlock(&perf_resource_lock
);
3452 static SYSDEV_CLASS_ATTR(
3455 perf_show_reserve_percpu
,
3456 perf_set_reserve_percpu
3459 static SYSDEV_CLASS_ATTR(
3462 perf_show_overcommit
,
3466 static struct attribute
*perfclass_attrs
[] = {
3467 &attr_reserve_percpu
.attr
,
3468 &attr_overcommit
.attr
,
3472 static struct attribute_group perfclass_attr_group
= {
3473 .attrs
= perfclass_attrs
,
3474 .name
= "perf_counters",
3477 static int __init
perf_counter_sysfs_init(void)
3479 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
3480 &perfclass_attr_group
);
3482 device_initcall(perf_counter_sysfs_init
);