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
= 512; /* 'free' kb per user */
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
);
1049 void perf_adjust_freq(struct perf_counter_context
*ctx
)
1051 struct perf_counter
*counter
;
1056 spin_lock(&ctx
->lock
);
1057 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1058 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1061 if (!counter
->hw_event
.freq
|| !counter
->hw_event
.irq_freq
)
1064 events
= HZ
* counter
->hw
.interrupts
* counter
->hw
.irq_period
;
1065 period
= div64_u64(events
, counter
->hw_event
.irq_freq
);
1067 delta
= (s64
)(1 + period
- counter
->hw
.irq_period
);
1070 irq_period
= counter
->hw
.irq_period
+ delta
;
1075 counter
->hw
.irq_period
= irq_period
;
1076 counter
->hw
.interrupts
= 0;
1078 spin_unlock(&ctx
->lock
);
1082 * Round-robin a context's counters:
1084 static void rotate_ctx(struct perf_counter_context
*ctx
)
1086 struct perf_counter
*counter
;
1088 if (!ctx
->nr_counters
)
1091 spin_lock(&ctx
->lock
);
1093 * Rotate the first entry last (works just fine for group counters too):
1096 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1097 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1102 spin_unlock(&ctx
->lock
);
1105 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1107 struct perf_cpu_context
*cpuctx
;
1108 struct perf_counter_context
*ctx
;
1110 if (!atomic_read(&nr_counters
))
1113 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1114 ctx
= &curr
->perf_counter_ctx
;
1116 perf_adjust_freq(&cpuctx
->ctx
);
1117 perf_adjust_freq(ctx
);
1119 perf_counter_cpu_sched_out(cpuctx
);
1120 __perf_counter_task_sched_out(ctx
);
1122 rotate_ctx(&cpuctx
->ctx
);
1125 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1126 perf_counter_task_sched_in(curr
, cpu
);
1130 * Cross CPU call to read the hardware counter
1132 static void __read(void *info
)
1134 struct perf_counter
*counter
= info
;
1135 struct perf_counter_context
*ctx
= counter
->ctx
;
1136 unsigned long flags
;
1138 local_irq_save(flags
);
1140 update_context_time(ctx
);
1141 counter
->pmu
->read(counter
);
1142 update_counter_times(counter
);
1143 local_irq_restore(flags
);
1146 static u64
perf_counter_read(struct perf_counter
*counter
)
1149 * If counter is enabled and currently active on a CPU, update the
1150 * value in the counter structure:
1152 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1153 smp_call_function_single(counter
->oncpu
,
1154 __read
, counter
, 1);
1155 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1156 update_counter_times(counter
);
1159 return atomic64_read(&counter
->count
);
1162 static void put_context(struct perf_counter_context
*ctx
)
1165 put_task_struct(ctx
->task
);
1168 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1170 struct perf_cpu_context
*cpuctx
;
1171 struct perf_counter_context
*ctx
;
1172 struct task_struct
*task
;
1175 * If cpu is not a wildcard then this is a percpu counter:
1178 /* Must be root to operate on a CPU counter: */
1179 if (sysctl_perf_counter_priv
&& !capable(CAP_SYS_ADMIN
))
1180 return ERR_PTR(-EACCES
);
1182 if (cpu
< 0 || cpu
> num_possible_cpus())
1183 return ERR_PTR(-EINVAL
);
1186 * We could be clever and allow to attach a counter to an
1187 * offline CPU and activate it when the CPU comes up, but
1190 if (!cpu_isset(cpu
, cpu_online_map
))
1191 return ERR_PTR(-ENODEV
);
1193 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1203 task
= find_task_by_vpid(pid
);
1205 get_task_struct(task
);
1209 return ERR_PTR(-ESRCH
);
1211 ctx
= &task
->perf_counter_ctx
;
1214 /* Reuse ptrace permission checks for now. */
1215 if (!ptrace_may_access(task
, PTRACE_MODE_READ
)) {
1217 return ERR_PTR(-EACCES
);
1223 static void free_counter_rcu(struct rcu_head
*head
)
1225 struct perf_counter
*counter
;
1227 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1231 static void perf_pending_sync(struct perf_counter
*counter
);
1233 static void free_counter(struct perf_counter
*counter
)
1235 perf_pending_sync(counter
);
1237 atomic_dec(&nr_counters
);
1238 if (counter
->hw_event
.mmap
)
1239 atomic_dec(&nr_mmap_tracking
);
1240 if (counter
->hw_event
.munmap
)
1241 atomic_dec(&nr_munmap_tracking
);
1242 if (counter
->hw_event
.comm
)
1243 atomic_dec(&nr_comm_tracking
);
1245 if (counter
->destroy
)
1246 counter
->destroy(counter
);
1248 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1252 * Called when the last reference to the file is gone.
1254 static int perf_release(struct inode
*inode
, struct file
*file
)
1256 struct perf_counter
*counter
= file
->private_data
;
1257 struct perf_counter_context
*ctx
= counter
->ctx
;
1259 file
->private_data
= NULL
;
1261 mutex_lock(&ctx
->mutex
);
1262 mutex_lock(&counter
->mutex
);
1264 perf_counter_remove_from_context(counter
);
1266 mutex_unlock(&counter
->mutex
);
1267 mutex_unlock(&ctx
->mutex
);
1269 free_counter(counter
);
1276 * Read the performance counter - simple non blocking version for now
1279 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1285 * Return end-of-file for a read on a counter that is in
1286 * error state (i.e. because it was pinned but it couldn't be
1287 * scheduled on to the CPU at some point).
1289 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1292 mutex_lock(&counter
->mutex
);
1293 values
[0] = perf_counter_read(counter
);
1295 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1296 values
[n
++] = counter
->total_time_enabled
+
1297 atomic64_read(&counter
->child_total_time_enabled
);
1298 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1299 values
[n
++] = counter
->total_time_running
+
1300 atomic64_read(&counter
->child_total_time_running
);
1301 mutex_unlock(&counter
->mutex
);
1303 if (count
< n
* sizeof(u64
))
1305 count
= n
* sizeof(u64
);
1307 if (copy_to_user(buf
, values
, count
))
1314 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1316 struct perf_counter
*counter
= file
->private_data
;
1318 return perf_read_hw(counter
, buf
, count
);
1321 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1323 struct perf_counter
*counter
= file
->private_data
;
1324 struct perf_mmap_data
*data
;
1325 unsigned int events
= POLL_HUP
;
1328 data
= rcu_dereference(counter
->data
);
1330 events
= atomic_xchg(&data
->poll
, 0);
1333 poll_wait(file
, &counter
->waitq
, wait
);
1338 static void perf_counter_reset(struct perf_counter
*counter
)
1340 (void)perf_counter_read(counter
);
1341 atomic64_set(&counter
->count
, 0);
1342 perf_counter_update_userpage(counter
);
1345 static void perf_counter_for_each_sibling(struct perf_counter
*counter
,
1346 void (*func
)(struct perf_counter
*))
1348 struct perf_counter_context
*ctx
= counter
->ctx
;
1349 struct perf_counter
*sibling
;
1351 spin_lock_irq(&ctx
->lock
);
1352 counter
= counter
->group_leader
;
1355 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1357 spin_unlock_irq(&ctx
->lock
);
1360 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1361 void (*func
)(struct perf_counter
*))
1363 struct perf_counter
*child
;
1365 mutex_lock(&counter
->mutex
);
1367 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1369 mutex_unlock(&counter
->mutex
);
1372 static void perf_counter_for_each(struct perf_counter
*counter
,
1373 void (*func
)(struct perf_counter
*))
1375 struct perf_counter
*child
;
1377 mutex_lock(&counter
->mutex
);
1378 perf_counter_for_each_sibling(counter
, func
);
1379 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1380 perf_counter_for_each_sibling(child
, func
);
1381 mutex_unlock(&counter
->mutex
);
1384 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1386 struct perf_counter
*counter
= file
->private_data
;
1387 void (*func
)(struct perf_counter
*);
1391 case PERF_COUNTER_IOC_ENABLE
:
1392 func
= perf_counter_enable
;
1394 case PERF_COUNTER_IOC_DISABLE
:
1395 func
= perf_counter_disable
;
1397 case PERF_COUNTER_IOC_RESET
:
1398 func
= perf_counter_reset
;
1401 case PERF_COUNTER_IOC_REFRESH
:
1402 return perf_counter_refresh(counter
, arg
);
1407 if (flags
& PERF_IOC_FLAG_GROUP
)
1408 perf_counter_for_each(counter
, func
);
1410 perf_counter_for_each_child(counter
, func
);
1416 * Callers need to ensure there can be no nesting of this function, otherwise
1417 * the seqlock logic goes bad. We can not serialize this because the arch
1418 * code calls this from NMI context.
1420 void perf_counter_update_userpage(struct perf_counter
*counter
)
1422 struct perf_mmap_data
*data
;
1423 struct perf_counter_mmap_page
*userpg
;
1426 data
= rcu_dereference(counter
->data
);
1430 userpg
= data
->user_page
;
1433 * Disable preemption so as to not let the corresponding user-space
1434 * spin too long if we get preempted.
1439 userpg
->index
= counter
->hw
.idx
;
1440 userpg
->offset
= atomic64_read(&counter
->count
);
1441 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1442 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1451 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1453 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1454 struct perf_mmap_data
*data
;
1455 int ret
= VM_FAULT_SIGBUS
;
1458 data
= rcu_dereference(counter
->data
);
1462 if (vmf
->pgoff
== 0) {
1463 vmf
->page
= virt_to_page(data
->user_page
);
1465 int nr
= vmf
->pgoff
- 1;
1467 if ((unsigned)nr
> data
->nr_pages
)
1470 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1472 get_page(vmf
->page
);
1480 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1482 struct perf_mmap_data
*data
;
1486 WARN_ON(atomic_read(&counter
->mmap_count
));
1488 size
= sizeof(struct perf_mmap_data
);
1489 size
+= nr_pages
* sizeof(void *);
1491 data
= kzalloc(size
, GFP_KERNEL
);
1495 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1496 if (!data
->user_page
)
1497 goto fail_user_page
;
1499 for (i
= 0; i
< nr_pages
; i
++) {
1500 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1501 if (!data
->data_pages
[i
])
1502 goto fail_data_pages
;
1505 data
->nr_pages
= nr_pages
;
1506 atomic_set(&data
->lock
, -1);
1508 rcu_assign_pointer(counter
->data
, data
);
1513 for (i
--; i
>= 0; i
--)
1514 free_page((unsigned long)data
->data_pages
[i
]);
1516 free_page((unsigned long)data
->user_page
);
1525 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1527 struct perf_mmap_data
*data
= container_of(rcu_head
,
1528 struct perf_mmap_data
, rcu_head
);
1531 free_page((unsigned long)data
->user_page
);
1532 for (i
= 0; i
< data
->nr_pages
; i
++)
1533 free_page((unsigned long)data
->data_pages
[i
]);
1537 static void perf_mmap_data_free(struct perf_counter
*counter
)
1539 struct perf_mmap_data
*data
= counter
->data
;
1541 WARN_ON(atomic_read(&counter
->mmap_count
));
1543 rcu_assign_pointer(counter
->data
, NULL
);
1544 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1547 static void perf_mmap_open(struct vm_area_struct
*vma
)
1549 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1551 atomic_inc(&counter
->mmap_count
);
1554 static void perf_mmap_close(struct vm_area_struct
*vma
)
1556 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1558 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
,
1559 &counter
->mmap_mutex
)) {
1560 struct user_struct
*user
= current_user();
1562 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
1563 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
1564 perf_mmap_data_free(counter
);
1565 mutex_unlock(&counter
->mmap_mutex
);
1569 static struct vm_operations_struct perf_mmap_vmops
= {
1570 .open
= perf_mmap_open
,
1571 .close
= perf_mmap_close
,
1572 .fault
= perf_mmap_fault
,
1575 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1577 struct perf_counter
*counter
= file
->private_data
;
1578 struct user_struct
*user
= current_user();
1579 unsigned long vma_size
;
1580 unsigned long nr_pages
;
1581 unsigned long user_locked
, user_lock_limit
;
1582 unsigned long locked
, lock_limit
;
1583 long user_extra
, extra
;
1586 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1589 vma_size
= vma
->vm_end
- vma
->vm_start
;
1590 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1593 * If we have data pages ensure they're a power-of-two number, so we
1594 * can do bitmasks instead of modulo.
1596 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1599 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1602 if (vma
->vm_pgoff
!= 0)
1605 mutex_lock(&counter
->mmap_mutex
);
1606 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1607 if (nr_pages
!= counter
->data
->nr_pages
)
1612 user_extra
= nr_pages
+ 1;
1613 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
1614 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
1617 if (user_locked
> user_lock_limit
)
1618 extra
= user_locked
- user_lock_limit
;
1620 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1621 lock_limit
>>= PAGE_SHIFT
;
1622 locked
= vma
->vm_mm
->locked_vm
+ extra
;
1624 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
1629 WARN_ON(counter
->data
);
1630 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1634 atomic_set(&counter
->mmap_count
, 1);
1635 atomic_long_add(user_extra
, &user
->locked_vm
);
1636 vma
->vm_mm
->locked_vm
+= extra
;
1637 counter
->data
->nr_locked
= extra
;
1639 mutex_unlock(&counter
->mmap_mutex
);
1641 vma
->vm_flags
&= ~VM_MAYWRITE
;
1642 vma
->vm_flags
|= VM_RESERVED
;
1643 vma
->vm_ops
= &perf_mmap_vmops
;
1648 static int perf_fasync(int fd
, struct file
*filp
, int on
)
1650 struct perf_counter
*counter
= filp
->private_data
;
1651 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1654 mutex_lock(&inode
->i_mutex
);
1655 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
1656 mutex_unlock(&inode
->i_mutex
);
1664 static const struct file_operations perf_fops
= {
1665 .release
= perf_release
,
1668 .unlocked_ioctl
= perf_ioctl
,
1669 .compat_ioctl
= perf_ioctl
,
1671 .fasync
= perf_fasync
,
1675 * Perf counter wakeup
1677 * If there's data, ensure we set the poll() state and publish everything
1678 * to user-space before waking everybody up.
1681 void perf_counter_wakeup(struct perf_counter
*counter
)
1683 wake_up_all(&counter
->waitq
);
1685 if (counter
->pending_kill
) {
1686 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
1687 counter
->pending_kill
= 0;
1694 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1696 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1697 * single linked list and use cmpxchg() to add entries lockless.
1700 static void perf_pending_counter(struct perf_pending_entry
*entry
)
1702 struct perf_counter
*counter
= container_of(entry
,
1703 struct perf_counter
, pending
);
1705 if (counter
->pending_disable
) {
1706 counter
->pending_disable
= 0;
1707 perf_counter_disable(counter
);
1710 if (counter
->pending_wakeup
) {
1711 counter
->pending_wakeup
= 0;
1712 perf_counter_wakeup(counter
);
1716 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1718 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
1722 static void perf_pending_queue(struct perf_pending_entry
*entry
,
1723 void (*func
)(struct perf_pending_entry
*))
1725 struct perf_pending_entry
**head
;
1727 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
1732 head
= &get_cpu_var(perf_pending_head
);
1735 entry
->next
= *head
;
1736 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
1738 set_perf_counter_pending();
1740 put_cpu_var(perf_pending_head
);
1743 static int __perf_pending_run(void)
1745 struct perf_pending_entry
*list
;
1748 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
1749 while (list
!= PENDING_TAIL
) {
1750 void (*func
)(struct perf_pending_entry
*);
1751 struct perf_pending_entry
*entry
= list
;
1758 * Ensure we observe the unqueue before we issue the wakeup,
1759 * so that we won't be waiting forever.
1760 * -- see perf_not_pending().
1771 static inline int perf_not_pending(struct perf_counter
*counter
)
1774 * If we flush on whatever cpu we run, there is a chance we don't
1778 __perf_pending_run();
1782 * Ensure we see the proper queue state before going to sleep
1783 * so that we do not miss the wakeup. -- see perf_pending_handle()
1786 return counter
->pending
.next
== NULL
;
1789 static void perf_pending_sync(struct perf_counter
*counter
)
1791 wait_event(counter
->waitq
, perf_not_pending(counter
));
1794 void perf_counter_do_pending(void)
1796 __perf_pending_run();
1800 * Callchain support -- arch specific
1803 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1812 struct perf_output_handle
{
1813 struct perf_counter
*counter
;
1814 struct perf_mmap_data
*data
;
1815 unsigned int offset
;
1820 unsigned long flags
;
1823 static void perf_output_wakeup(struct perf_output_handle
*handle
)
1825 atomic_set(&handle
->data
->poll
, POLL_IN
);
1828 handle
->counter
->pending_wakeup
= 1;
1829 perf_pending_queue(&handle
->counter
->pending
,
1830 perf_pending_counter
);
1832 perf_counter_wakeup(handle
->counter
);
1836 * Curious locking construct.
1838 * We need to ensure a later event doesn't publish a head when a former
1839 * event isn't done writing. However since we need to deal with NMIs we
1840 * cannot fully serialize things.
1842 * What we do is serialize between CPUs so we only have to deal with NMI
1843 * nesting on a single CPU.
1845 * We only publish the head (and generate a wakeup) when the outer-most
1848 static void perf_output_lock(struct perf_output_handle
*handle
)
1850 struct perf_mmap_data
*data
= handle
->data
;
1855 local_irq_save(handle
->flags
);
1856 cpu
= smp_processor_id();
1858 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
1861 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
1867 static void perf_output_unlock(struct perf_output_handle
*handle
)
1869 struct perf_mmap_data
*data
= handle
->data
;
1872 data
->done_head
= data
->head
;
1874 if (!handle
->locked
)
1879 * The xchg implies a full barrier that ensures all writes are done
1880 * before we publish the new head, matched by a rmb() in userspace when
1881 * reading this position.
1883 while ((head
= atomic_xchg(&data
->done_head
, 0)))
1884 data
->user_page
->data_head
= head
;
1887 * NMI can happen here, which means we can miss a done_head update.
1890 cpu
= atomic_xchg(&data
->lock
, -1);
1891 WARN_ON_ONCE(cpu
!= smp_processor_id());
1894 * Therefore we have to validate we did not indeed do so.
1896 if (unlikely(atomic_read(&data
->done_head
))) {
1898 * Since we had it locked, we can lock it again.
1900 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
1906 if (atomic_xchg(&data
->wakeup
, 0))
1907 perf_output_wakeup(handle
);
1909 local_irq_restore(handle
->flags
);
1912 static int perf_output_begin(struct perf_output_handle
*handle
,
1913 struct perf_counter
*counter
, unsigned int size
,
1914 int nmi
, int overflow
)
1916 struct perf_mmap_data
*data
;
1917 unsigned int offset
, head
;
1920 * For inherited counters we send all the output towards the parent.
1922 if (counter
->parent
)
1923 counter
= counter
->parent
;
1926 data
= rcu_dereference(counter
->data
);
1930 handle
->data
= data
;
1931 handle
->counter
= counter
;
1933 handle
->overflow
= overflow
;
1935 if (!data
->nr_pages
)
1938 perf_output_lock(handle
);
1941 offset
= head
= atomic_read(&data
->head
);
1943 } while (atomic_cmpxchg(&data
->head
, offset
, head
) != offset
);
1945 handle
->offset
= offset
;
1946 handle
->head
= head
;
1948 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
1949 atomic_set(&data
->wakeup
, 1);
1954 perf_output_wakeup(handle
);
1961 static void perf_output_copy(struct perf_output_handle
*handle
,
1962 void *buf
, unsigned int len
)
1964 unsigned int pages_mask
;
1965 unsigned int offset
;
1969 offset
= handle
->offset
;
1970 pages_mask
= handle
->data
->nr_pages
- 1;
1971 pages
= handle
->data
->data_pages
;
1974 unsigned int page_offset
;
1977 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
1978 page_offset
= offset
& (PAGE_SIZE
- 1);
1979 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
1981 memcpy(pages
[nr
] + page_offset
, buf
, size
);
1988 handle
->offset
= offset
;
1991 * Check we didn't copy past our reservation window, taking the
1992 * possible unsigned int wrap into account.
1994 WARN_ON_ONCE(((int)(handle
->head
- handle
->offset
)) < 0);
1997 #define perf_output_put(handle, x) \
1998 perf_output_copy((handle), &(x), sizeof(x))
2000 static void perf_output_end(struct perf_output_handle
*handle
)
2002 struct perf_counter
*counter
= handle
->counter
;
2003 struct perf_mmap_data
*data
= handle
->data
;
2005 int wakeup_events
= counter
->hw_event
.wakeup_events
;
2007 if (handle
->overflow
&& wakeup_events
) {
2008 int events
= atomic_inc_return(&data
->events
);
2009 if (events
>= wakeup_events
) {
2010 atomic_sub(wakeup_events
, &data
->events
);
2011 atomic_set(&data
->wakeup
, 1);
2015 perf_output_unlock(handle
);
2019 static void perf_counter_output(struct perf_counter
*counter
,
2020 int nmi
, struct pt_regs
*regs
, u64 addr
)
2023 u64 record_type
= counter
->hw_event
.record_type
;
2024 struct perf_output_handle handle
;
2025 struct perf_event_header header
;
2034 struct perf_callchain_entry
*callchain
= NULL
;
2035 int callchain_size
= 0;
2042 header
.size
= sizeof(header
);
2044 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
2045 header
.misc
|= user_mode(regs
) ?
2046 PERF_EVENT_MISC_USER
: PERF_EVENT_MISC_KERNEL
;
2048 if (record_type
& PERF_RECORD_IP
) {
2049 ip
= instruction_pointer(regs
);
2050 header
.type
|= PERF_RECORD_IP
;
2051 header
.size
+= sizeof(ip
);
2054 if (record_type
& PERF_RECORD_TID
) {
2055 /* namespace issues */
2056 tid_entry
.pid
= current
->group_leader
->pid
;
2057 tid_entry
.tid
= current
->pid
;
2059 header
.type
|= PERF_RECORD_TID
;
2060 header
.size
+= sizeof(tid_entry
);
2063 if (record_type
& PERF_RECORD_TIME
) {
2065 * Maybe do better on x86 and provide cpu_clock_nmi()
2067 time
= sched_clock();
2069 header
.type
|= PERF_RECORD_TIME
;
2070 header
.size
+= sizeof(u64
);
2073 if (record_type
& PERF_RECORD_ADDR
) {
2074 header
.type
|= PERF_RECORD_ADDR
;
2075 header
.size
+= sizeof(u64
);
2078 if (record_type
& PERF_RECORD_CONFIG
) {
2079 header
.type
|= PERF_RECORD_CONFIG
;
2080 header
.size
+= sizeof(u64
);
2083 if (record_type
& PERF_RECORD_CPU
) {
2084 header
.type
|= PERF_RECORD_CPU
;
2085 header
.size
+= sizeof(cpu_entry
);
2087 cpu_entry
.cpu
= raw_smp_processor_id();
2090 if (record_type
& PERF_RECORD_GROUP
) {
2091 header
.type
|= PERF_RECORD_GROUP
;
2092 header
.size
+= sizeof(u64
) +
2093 counter
->nr_siblings
* sizeof(group_entry
);
2096 if (record_type
& PERF_RECORD_CALLCHAIN
) {
2097 callchain
= perf_callchain(regs
);
2100 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2102 header
.type
|= PERF_RECORD_CALLCHAIN
;
2103 header
.size
+= callchain_size
;
2107 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2111 perf_output_put(&handle
, header
);
2113 if (record_type
& PERF_RECORD_IP
)
2114 perf_output_put(&handle
, ip
);
2116 if (record_type
& PERF_RECORD_TID
)
2117 perf_output_put(&handle
, tid_entry
);
2119 if (record_type
& PERF_RECORD_TIME
)
2120 perf_output_put(&handle
, time
);
2122 if (record_type
& PERF_RECORD_ADDR
)
2123 perf_output_put(&handle
, addr
);
2125 if (record_type
& PERF_RECORD_CONFIG
)
2126 perf_output_put(&handle
, counter
->hw_event
.config
);
2128 if (record_type
& PERF_RECORD_CPU
)
2129 perf_output_put(&handle
, cpu_entry
);
2132 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2134 if (record_type
& PERF_RECORD_GROUP
) {
2135 struct perf_counter
*leader
, *sub
;
2136 u64 nr
= counter
->nr_siblings
;
2138 perf_output_put(&handle
, nr
);
2140 leader
= counter
->group_leader
;
2141 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2143 sub
->pmu
->read(sub
);
2145 group_entry
.event
= sub
->hw_event
.config
;
2146 group_entry
.counter
= atomic64_read(&sub
->count
);
2148 perf_output_put(&handle
, group_entry
);
2153 perf_output_copy(&handle
, callchain
, callchain_size
);
2155 perf_output_end(&handle
);
2162 struct perf_comm_event
{
2163 struct task_struct
*task
;
2168 struct perf_event_header header
;
2175 static void perf_counter_comm_output(struct perf_counter
*counter
,
2176 struct perf_comm_event
*comm_event
)
2178 struct perf_output_handle handle
;
2179 int size
= comm_event
->event
.header
.size
;
2180 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2185 perf_output_put(&handle
, comm_event
->event
);
2186 perf_output_copy(&handle
, comm_event
->comm
,
2187 comm_event
->comm_size
);
2188 perf_output_end(&handle
);
2191 static int perf_counter_comm_match(struct perf_counter
*counter
,
2192 struct perf_comm_event
*comm_event
)
2194 if (counter
->hw_event
.comm
&&
2195 comm_event
->event
.header
.type
== PERF_EVENT_COMM
)
2201 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2202 struct perf_comm_event
*comm_event
)
2204 struct perf_counter
*counter
;
2206 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2210 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2211 if (perf_counter_comm_match(counter
, comm_event
))
2212 perf_counter_comm_output(counter
, comm_event
);
2217 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2219 struct perf_cpu_context
*cpuctx
;
2221 char *comm
= comm_event
->task
->comm
;
2223 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2225 comm_event
->comm
= comm
;
2226 comm_event
->comm_size
= size
;
2228 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2230 cpuctx
= &get_cpu_var(perf_cpu_context
);
2231 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2232 put_cpu_var(perf_cpu_context
);
2234 perf_counter_comm_ctx(¤t
->perf_counter_ctx
, comm_event
);
2237 void perf_counter_comm(struct task_struct
*task
)
2239 struct perf_comm_event comm_event
;
2241 if (!atomic_read(&nr_comm_tracking
))
2244 comm_event
= (struct perf_comm_event
){
2247 .header
= { .type
= PERF_EVENT_COMM
, },
2248 .pid
= task
->group_leader
->pid
,
2253 perf_counter_comm_event(&comm_event
);
2260 struct perf_mmap_event
{
2266 struct perf_event_header header
;
2276 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2277 struct perf_mmap_event
*mmap_event
)
2279 struct perf_output_handle handle
;
2280 int size
= mmap_event
->event
.header
.size
;
2281 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2286 perf_output_put(&handle
, mmap_event
->event
);
2287 perf_output_copy(&handle
, mmap_event
->file_name
,
2288 mmap_event
->file_size
);
2289 perf_output_end(&handle
);
2292 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2293 struct perf_mmap_event
*mmap_event
)
2295 if (counter
->hw_event
.mmap
&&
2296 mmap_event
->event
.header
.type
== PERF_EVENT_MMAP
)
2299 if (counter
->hw_event
.munmap
&&
2300 mmap_event
->event
.header
.type
== PERF_EVENT_MUNMAP
)
2306 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2307 struct perf_mmap_event
*mmap_event
)
2309 struct perf_counter
*counter
;
2311 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2315 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2316 if (perf_counter_mmap_match(counter
, mmap_event
))
2317 perf_counter_mmap_output(counter
, mmap_event
);
2322 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2324 struct perf_cpu_context
*cpuctx
;
2325 struct file
*file
= mmap_event
->file
;
2332 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2334 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2337 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2339 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2343 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2348 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2350 mmap_event
->file_name
= name
;
2351 mmap_event
->file_size
= size
;
2353 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2355 cpuctx
= &get_cpu_var(perf_cpu_context
);
2356 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2357 put_cpu_var(perf_cpu_context
);
2359 perf_counter_mmap_ctx(¤t
->perf_counter_ctx
, mmap_event
);
2364 void perf_counter_mmap(unsigned long addr
, unsigned long len
,
2365 unsigned long pgoff
, struct file
*file
)
2367 struct perf_mmap_event mmap_event
;
2369 if (!atomic_read(&nr_mmap_tracking
))
2372 mmap_event
= (struct perf_mmap_event
){
2375 .header
= { .type
= PERF_EVENT_MMAP
, },
2376 .pid
= current
->group_leader
->pid
,
2377 .tid
= current
->pid
,
2384 perf_counter_mmap_event(&mmap_event
);
2387 void perf_counter_munmap(unsigned long addr
, unsigned long len
,
2388 unsigned long pgoff
, struct file
*file
)
2390 struct perf_mmap_event mmap_event
;
2392 if (!atomic_read(&nr_munmap_tracking
))
2395 mmap_event
= (struct perf_mmap_event
){
2398 .header
= { .type
= PERF_EVENT_MUNMAP
, },
2399 .pid
= current
->group_leader
->pid
,
2400 .tid
= current
->pid
,
2407 perf_counter_mmap_event(&mmap_event
);
2411 * Generic counter overflow handling.
2414 int perf_counter_overflow(struct perf_counter
*counter
,
2415 int nmi
, struct pt_regs
*regs
, u64 addr
)
2417 int events
= atomic_read(&counter
->event_limit
);
2420 counter
->hw
.interrupts
++;
2423 * XXX event_limit might not quite work as expected on inherited
2427 counter
->pending_kill
= POLL_IN
;
2428 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
2430 counter
->pending_kill
= POLL_HUP
;
2432 counter
->pending_disable
= 1;
2433 perf_pending_queue(&counter
->pending
,
2434 perf_pending_counter
);
2436 perf_counter_disable(counter
);
2439 perf_counter_output(counter
, nmi
, regs
, addr
);
2444 * Generic software counter infrastructure
2447 static void perf_swcounter_update(struct perf_counter
*counter
)
2449 struct hw_perf_counter
*hwc
= &counter
->hw
;
2454 prev
= atomic64_read(&hwc
->prev_count
);
2455 now
= atomic64_read(&hwc
->count
);
2456 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2461 atomic64_add(delta
, &counter
->count
);
2462 atomic64_sub(delta
, &hwc
->period_left
);
2465 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2467 struct hw_perf_counter
*hwc
= &counter
->hw
;
2468 s64 left
= atomic64_read(&hwc
->period_left
);
2469 s64 period
= hwc
->irq_period
;
2471 if (unlikely(left
<= -period
)) {
2473 atomic64_set(&hwc
->period_left
, left
);
2476 if (unlikely(left
<= 0)) {
2478 atomic64_add(period
, &hwc
->period_left
);
2481 atomic64_set(&hwc
->prev_count
, -left
);
2482 atomic64_set(&hwc
->count
, -left
);
2485 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2487 enum hrtimer_restart ret
= HRTIMER_RESTART
;
2488 struct perf_counter
*counter
;
2489 struct pt_regs
*regs
;
2492 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
2493 counter
->pmu
->read(counter
);
2495 regs
= get_irq_regs();
2497 * In case we exclude kernel IPs or are somehow not in interrupt
2498 * context, provide the next best thing, the user IP.
2500 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
2501 !counter
->hw_event
.exclude_user
)
2502 regs
= task_pt_regs(current
);
2505 if (perf_counter_overflow(counter
, 0, regs
, 0))
2506 ret
= HRTIMER_NORESTART
;
2509 period
= max_t(u64
, 10000, counter
->hw
.irq_period
);
2510 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
2515 static void perf_swcounter_overflow(struct perf_counter
*counter
,
2516 int nmi
, struct pt_regs
*regs
, u64 addr
)
2518 perf_swcounter_update(counter
);
2519 perf_swcounter_set_period(counter
);
2520 if (perf_counter_overflow(counter
, nmi
, regs
, addr
))
2521 /* soft-disable the counter */
2526 static int perf_swcounter_match(struct perf_counter
*counter
,
2527 enum perf_event_types type
,
2528 u32 event
, struct pt_regs
*regs
)
2530 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
2533 if (perf_event_raw(&counter
->hw_event
))
2536 if (perf_event_type(&counter
->hw_event
) != type
)
2539 if (perf_event_id(&counter
->hw_event
) != event
)
2542 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
2545 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
2551 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
2552 int nmi
, struct pt_regs
*regs
, u64 addr
)
2554 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
2555 if (counter
->hw
.irq_period
&& !neg
)
2556 perf_swcounter_overflow(counter
, nmi
, regs
, addr
);
2559 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
2560 enum perf_event_types type
, u32 event
,
2561 u64 nr
, int nmi
, struct pt_regs
*regs
,
2564 struct perf_counter
*counter
;
2566 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2570 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2571 if (perf_swcounter_match(counter
, type
, event
, regs
))
2572 perf_swcounter_add(counter
, nr
, nmi
, regs
, addr
);
2577 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
2580 return &cpuctx
->recursion
[3];
2583 return &cpuctx
->recursion
[2];
2586 return &cpuctx
->recursion
[1];
2588 return &cpuctx
->recursion
[0];
2591 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
2592 u64 nr
, int nmi
, struct pt_regs
*regs
,
2595 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
2596 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
2604 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
2605 nr
, nmi
, regs
, addr
);
2606 if (cpuctx
->task_ctx
) {
2607 perf_swcounter_ctx_event(cpuctx
->task_ctx
, type
, event
,
2608 nr
, nmi
, regs
, addr
);
2615 put_cpu_var(perf_cpu_context
);
2619 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
2621 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
, addr
);
2624 static void perf_swcounter_read(struct perf_counter
*counter
)
2626 perf_swcounter_update(counter
);
2629 static int perf_swcounter_enable(struct perf_counter
*counter
)
2631 perf_swcounter_set_period(counter
);
2635 static void perf_swcounter_disable(struct perf_counter
*counter
)
2637 perf_swcounter_update(counter
);
2640 static const struct pmu perf_ops_generic
= {
2641 .enable
= perf_swcounter_enable
,
2642 .disable
= perf_swcounter_disable
,
2643 .read
= perf_swcounter_read
,
2647 * Software counter: cpu wall time clock
2650 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
2652 int cpu
= raw_smp_processor_id();
2656 now
= cpu_clock(cpu
);
2657 prev
= atomic64_read(&counter
->hw
.prev_count
);
2658 atomic64_set(&counter
->hw
.prev_count
, now
);
2659 atomic64_add(now
- prev
, &counter
->count
);
2662 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
2664 struct hw_perf_counter
*hwc
= &counter
->hw
;
2665 int cpu
= raw_smp_processor_id();
2667 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
2668 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2669 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2670 if (hwc
->irq_period
) {
2671 u64 period
= max_t(u64
, 10000, hwc
->irq_period
);
2672 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2673 ns_to_ktime(period
), 0,
2674 HRTIMER_MODE_REL
, 0);
2680 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
2682 hrtimer_cancel(&counter
->hw
.hrtimer
);
2683 cpu_clock_perf_counter_update(counter
);
2686 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
2688 cpu_clock_perf_counter_update(counter
);
2691 static const struct pmu perf_ops_cpu_clock
= {
2692 .enable
= cpu_clock_perf_counter_enable
,
2693 .disable
= cpu_clock_perf_counter_disable
,
2694 .read
= cpu_clock_perf_counter_read
,
2698 * Software counter: task time clock
2701 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
2706 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
2708 atomic64_add(delta
, &counter
->count
);
2711 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
2713 struct hw_perf_counter
*hwc
= &counter
->hw
;
2716 now
= counter
->ctx
->time
;
2718 atomic64_set(&hwc
->prev_count
, now
);
2719 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2720 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2721 if (hwc
->irq_period
) {
2722 u64 period
= max_t(u64
, 10000, hwc
->irq_period
);
2723 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2724 ns_to_ktime(period
), 0,
2725 HRTIMER_MODE_REL
, 0);
2731 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
2733 hrtimer_cancel(&counter
->hw
.hrtimer
);
2734 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
2738 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
2743 update_context_time(counter
->ctx
);
2744 time
= counter
->ctx
->time
;
2746 u64 now
= perf_clock();
2747 u64 delta
= now
- counter
->ctx
->timestamp
;
2748 time
= counter
->ctx
->time
+ delta
;
2751 task_clock_perf_counter_update(counter
, time
);
2754 static const struct pmu perf_ops_task_clock
= {
2755 .enable
= task_clock_perf_counter_enable
,
2756 .disable
= task_clock_perf_counter_disable
,
2757 .read
= task_clock_perf_counter_read
,
2761 * Software counter: cpu migrations
2764 static inline u64
get_cpu_migrations(struct perf_counter
*counter
)
2766 struct task_struct
*curr
= counter
->ctx
->task
;
2769 return curr
->se
.nr_migrations
;
2770 return cpu_nr_migrations(smp_processor_id());
2773 static void cpu_migrations_perf_counter_update(struct perf_counter
*counter
)
2778 prev
= atomic64_read(&counter
->hw
.prev_count
);
2779 now
= get_cpu_migrations(counter
);
2781 atomic64_set(&counter
->hw
.prev_count
, now
);
2785 atomic64_add(delta
, &counter
->count
);
2788 static void cpu_migrations_perf_counter_read(struct perf_counter
*counter
)
2790 cpu_migrations_perf_counter_update(counter
);
2793 static int cpu_migrations_perf_counter_enable(struct perf_counter
*counter
)
2795 if (counter
->prev_state
<= PERF_COUNTER_STATE_OFF
)
2796 atomic64_set(&counter
->hw
.prev_count
,
2797 get_cpu_migrations(counter
));
2801 static void cpu_migrations_perf_counter_disable(struct perf_counter
*counter
)
2803 cpu_migrations_perf_counter_update(counter
);
2806 static const struct pmu perf_ops_cpu_migrations
= {
2807 .enable
= cpu_migrations_perf_counter_enable
,
2808 .disable
= cpu_migrations_perf_counter_disable
,
2809 .read
= cpu_migrations_perf_counter_read
,
2812 #ifdef CONFIG_EVENT_PROFILE
2813 void perf_tpcounter_event(int event_id
)
2815 struct pt_regs
*regs
= get_irq_regs();
2818 regs
= task_pt_regs(current
);
2820 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
, 0);
2822 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
2824 extern int ftrace_profile_enable(int);
2825 extern void ftrace_profile_disable(int);
2827 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
2829 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
2832 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
2834 int event_id
= perf_event_id(&counter
->hw_event
);
2837 ret
= ftrace_profile_enable(event_id
);
2841 counter
->destroy
= tp_perf_counter_destroy
;
2842 counter
->hw
.irq_period
= counter
->hw_event
.irq_period
;
2844 return &perf_ops_generic
;
2847 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
2853 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
2855 const struct pmu
*pmu
= NULL
;
2858 * Software counters (currently) can't in general distinguish
2859 * between user, kernel and hypervisor events.
2860 * However, context switches and cpu migrations are considered
2861 * to be kernel events, and page faults are never hypervisor
2864 switch (perf_event_id(&counter
->hw_event
)) {
2865 case PERF_COUNT_CPU_CLOCK
:
2866 pmu
= &perf_ops_cpu_clock
;
2869 case PERF_COUNT_TASK_CLOCK
:
2871 * If the user instantiates this as a per-cpu counter,
2872 * use the cpu_clock counter instead.
2874 if (counter
->ctx
->task
)
2875 pmu
= &perf_ops_task_clock
;
2877 pmu
= &perf_ops_cpu_clock
;
2880 case PERF_COUNT_PAGE_FAULTS
:
2881 case PERF_COUNT_PAGE_FAULTS_MIN
:
2882 case PERF_COUNT_PAGE_FAULTS_MAJ
:
2883 case PERF_COUNT_CONTEXT_SWITCHES
:
2884 pmu
= &perf_ops_generic
;
2886 case PERF_COUNT_CPU_MIGRATIONS
:
2887 if (!counter
->hw_event
.exclude_kernel
)
2888 pmu
= &perf_ops_cpu_migrations
;
2896 * Allocate and initialize a counter structure
2898 static struct perf_counter
*
2899 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
2901 struct perf_counter_context
*ctx
,
2902 struct perf_counter
*group_leader
,
2905 const struct pmu
*pmu
;
2906 struct perf_counter
*counter
;
2907 struct hw_perf_counter
*hwc
;
2910 counter
= kzalloc(sizeof(*counter
), gfpflags
);
2912 return ERR_PTR(-ENOMEM
);
2915 * Single counters are their own group leaders, with an
2916 * empty sibling list:
2919 group_leader
= counter
;
2921 mutex_init(&counter
->mutex
);
2922 INIT_LIST_HEAD(&counter
->list_entry
);
2923 INIT_LIST_HEAD(&counter
->event_entry
);
2924 INIT_LIST_HEAD(&counter
->sibling_list
);
2925 init_waitqueue_head(&counter
->waitq
);
2927 mutex_init(&counter
->mmap_mutex
);
2929 INIT_LIST_HEAD(&counter
->child_list
);
2932 counter
->hw_event
= *hw_event
;
2933 counter
->group_leader
= group_leader
;
2934 counter
->pmu
= NULL
;
2937 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2938 if (hw_event
->disabled
)
2939 counter
->state
= PERF_COUNTER_STATE_OFF
;
2944 if (hw_event
->freq
&& hw_event
->irq_freq
)
2945 hwc
->irq_period
= div64_u64(TICK_NSEC
, hw_event
->irq_freq
);
2947 hwc
->irq_period
= hw_event
->irq_period
;
2950 * we currently do not support PERF_RECORD_GROUP on inherited counters
2952 if (hw_event
->inherit
&& (hw_event
->record_type
& PERF_RECORD_GROUP
))
2955 if (perf_event_raw(hw_event
)) {
2956 pmu
= hw_perf_counter_init(counter
);
2960 switch (perf_event_type(hw_event
)) {
2961 case PERF_TYPE_HARDWARE
:
2962 pmu
= hw_perf_counter_init(counter
);
2965 case PERF_TYPE_SOFTWARE
:
2966 pmu
= sw_perf_counter_init(counter
);
2969 case PERF_TYPE_TRACEPOINT
:
2970 pmu
= tp_perf_counter_init(counter
);
2977 else if (IS_ERR(pmu
))
2982 return ERR_PTR(err
);
2987 atomic_inc(&nr_counters
);
2988 if (counter
->hw_event
.mmap
)
2989 atomic_inc(&nr_mmap_tracking
);
2990 if (counter
->hw_event
.munmap
)
2991 atomic_inc(&nr_munmap_tracking
);
2992 if (counter
->hw_event
.comm
)
2993 atomic_inc(&nr_comm_tracking
);
2999 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3001 * @hw_event_uptr: event type attributes for monitoring/sampling
3004 * @group_fd: group leader counter fd
3006 SYSCALL_DEFINE5(perf_counter_open
,
3007 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
3008 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
3010 struct perf_counter
*counter
, *group_leader
;
3011 struct perf_counter_hw_event hw_event
;
3012 struct perf_counter_context
*ctx
;
3013 struct file
*counter_file
= NULL
;
3014 struct file
*group_file
= NULL
;
3015 int fput_needed
= 0;
3016 int fput_needed2
= 0;
3019 /* for future expandability... */
3023 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
3027 * Get the target context (task or percpu):
3029 ctx
= find_get_context(pid
, cpu
);
3031 return PTR_ERR(ctx
);
3034 * Look up the group leader (we will attach this counter to it):
3036 group_leader
= NULL
;
3037 if (group_fd
!= -1) {
3039 group_file
= fget_light(group_fd
, &fput_needed
);
3041 goto err_put_context
;
3042 if (group_file
->f_op
!= &perf_fops
)
3043 goto err_put_context
;
3045 group_leader
= group_file
->private_data
;
3047 * Do not allow a recursive hierarchy (this new sibling
3048 * becoming part of another group-sibling):
3050 if (group_leader
->group_leader
!= group_leader
)
3051 goto err_put_context
;
3053 * Do not allow to attach to a group in a different
3054 * task or CPU context:
3056 if (group_leader
->ctx
!= ctx
)
3057 goto err_put_context
;
3059 * Only a group leader can be exclusive or pinned
3061 if (hw_event
.exclusive
|| hw_event
.pinned
)
3062 goto err_put_context
;
3065 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
3067 ret
= PTR_ERR(counter
);
3068 if (IS_ERR(counter
))
3069 goto err_put_context
;
3071 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
3073 goto err_free_put_context
;
3075 counter_file
= fget_light(ret
, &fput_needed2
);
3077 goto err_free_put_context
;
3079 counter
->filp
= counter_file
;
3080 mutex_lock(&ctx
->mutex
);
3081 perf_install_in_context(ctx
, counter
, cpu
);
3082 mutex_unlock(&ctx
->mutex
);
3084 fput_light(counter_file
, fput_needed2
);
3087 fput_light(group_file
, fput_needed
);
3091 err_free_put_context
:
3101 * Initialize the perf_counter context in a task_struct:
3104 __perf_counter_init_context(struct perf_counter_context
*ctx
,
3105 struct task_struct
*task
)
3107 memset(ctx
, 0, sizeof(*ctx
));
3108 spin_lock_init(&ctx
->lock
);
3109 mutex_init(&ctx
->mutex
);
3110 INIT_LIST_HEAD(&ctx
->counter_list
);
3111 INIT_LIST_HEAD(&ctx
->event_list
);
3116 * inherit a counter from parent task to child task:
3118 static struct perf_counter
*
3119 inherit_counter(struct perf_counter
*parent_counter
,
3120 struct task_struct
*parent
,
3121 struct perf_counter_context
*parent_ctx
,
3122 struct task_struct
*child
,
3123 struct perf_counter
*group_leader
,
3124 struct perf_counter_context
*child_ctx
)
3126 struct perf_counter
*child_counter
;
3129 * Instead of creating recursive hierarchies of counters,
3130 * we link inherited counters back to the original parent,
3131 * which has a filp for sure, which we use as the reference
3134 if (parent_counter
->parent
)
3135 parent_counter
= parent_counter
->parent
;
3137 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
3138 parent_counter
->cpu
, child_ctx
,
3139 group_leader
, GFP_KERNEL
);
3140 if (IS_ERR(child_counter
))
3141 return child_counter
;
3144 * Link it up in the child's context:
3146 child_counter
->task
= child
;
3147 add_counter_to_ctx(child_counter
, child_ctx
);
3149 child_counter
->parent
= parent_counter
;
3151 * inherit into child's child as well:
3153 child_counter
->hw_event
.inherit
= 1;
3156 * Get a reference to the parent filp - we will fput it
3157 * when the child counter exits. This is safe to do because
3158 * we are in the parent and we know that the filp still
3159 * exists and has a nonzero count:
3161 atomic_long_inc(&parent_counter
->filp
->f_count
);
3164 * Link this into the parent counter's child list
3166 mutex_lock(&parent_counter
->mutex
);
3167 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3170 * Make the child state follow the state of the parent counter,
3171 * not its hw_event.disabled bit. We hold the parent's mutex,
3172 * so we won't race with perf_counter_{en,dis}able_family.
3174 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3175 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3177 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3179 mutex_unlock(&parent_counter
->mutex
);
3181 return child_counter
;
3184 static int inherit_group(struct perf_counter
*parent_counter
,
3185 struct task_struct
*parent
,
3186 struct perf_counter_context
*parent_ctx
,
3187 struct task_struct
*child
,
3188 struct perf_counter_context
*child_ctx
)
3190 struct perf_counter
*leader
;
3191 struct perf_counter
*sub
;
3192 struct perf_counter
*child_ctr
;
3194 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3195 child
, NULL
, child_ctx
);
3197 return PTR_ERR(leader
);
3198 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3199 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3200 child
, leader
, child_ctx
);
3201 if (IS_ERR(child_ctr
))
3202 return PTR_ERR(child_ctr
);
3207 static void sync_child_counter(struct perf_counter
*child_counter
,
3208 struct perf_counter
*parent_counter
)
3210 u64 parent_val
, child_val
;
3212 parent_val
= atomic64_read(&parent_counter
->count
);
3213 child_val
= atomic64_read(&child_counter
->count
);
3216 * Add back the child's count to the parent's count:
3218 atomic64_add(child_val
, &parent_counter
->count
);
3219 atomic64_add(child_counter
->total_time_enabled
,
3220 &parent_counter
->child_total_time_enabled
);
3221 atomic64_add(child_counter
->total_time_running
,
3222 &parent_counter
->child_total_time_running
);
3225 * Remove this counter from the parent's list
3227 mutex_lock(&parent_counter
->mutex
);
3228 list_del_init(&child_counter
->child_list
);
3229 mutex_unlock(&parent_counter
->mutex
);
3232 * Release the parent counter, if this was the last
3235 fput(parent_counter
->filp
);
3239 __perf_counter_exit_task(struct task_struct
*child
,
3240 struct perf_counter
*child_counter
,
3241 struct perf_counter_context
*child_ctx
)
3243 struct perf_counter
*parent_counter
;
3244 struct perf_counter
*sub
, *tmp
;
3247 * If we do not self-reap then we have to wait for the
3248 * child task to unschedule (it will happen for sure),
3249 * so that its counter is at its final count. (This
3250 * condition triggers rarely - child tasks usually get
3251 * off their CPU before the parent has a chance to
3252 * get this far into the reaping action)
3254 if (child
!= current
) {
3255 wait_task_inactive(child
, 0);
3256 list_del_init(&child_counter
->list_entry
);
3257 update_counter_times(child_counter
);
3259 struct perf_cpu_context
*cpuctx
;
3260 unsigned long flags
;
3263 * Disable and unlink this counter.
3265 * Be careful about zapping the list - IRQ/NMI context
3266 * could still be processing it:
3268 local_irq_save(flags
);
3271 cpuctx
= &__get_cpu_var(perf_cpu_context
);
3273 group_sched_out(child_counter
, cpuctx
, child_ctx
);
3274 update_counter_times(child_counter
);
3276 list_del_init(&child_counter
->list_entry
);
3278 child_ctx
->nr_counters
--;
3281 local_irq_restore(flags
);
3284 parent_counter
= child_counter
->parent
;
3286 * It can happen that parent exits first, and has counters
3287 * that are still around due to the child reference. These
3288 * counters need to be zapped - but otherwise linger.
3290 if (parent_counter
) {
3291 sync_child_counter(child_counter
, parent_counter
);
3292 list_for_each_entry_safe(sub
, tmp
, &child_counter
->sibling_list
,
3295 sync_child_counter(sub
, sub
->parent
);
3299 free_counter(child_counter
);
3304 * When a child task exits, feed back counter values to parent counters.
3306 * Note: we may be running in child context, but the PID is not hashed
3307 * anymore so new counters will not be added.
3309 void perf_counter_exit_task(struct task_struct
*child
)
3311 struct perf_counter
*child_counter
, *tmp
;
3312 struct perf_counter_context
*child_ctx
;
3314 child_ctx
= &child
->perf_counter_ctx
;
3316 if (likely(!child_ctx
->nr_counters
))
3319 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
3321 __perf_counter_exit_task(child
, child_counter
, child_ctx
);
3325 * Initialize the perf_counter context in task_struct
3327 void perf_counter_init_task(struct task_struct
*child
)
3329 struct perf_counter_context
*child_ctx
, *parent_ctx
;
3330 struct perf_counter
*counter
;
3331 struct task_struct
*parent
= current
;
3333 child_ctx
= &child
->perf_counter_ctx
;
3334 parent_ctx
= &parent
->perf_counter_ctx
;
3336 __perf_counter_init_context(child_ctx
, child
);
3339 * This is executed from the parent task context, so inherit
3340 * counters that have been marked for cloning:
3343 if (likely(!parent_ctx
->nr_counters
))
3347 * Lock the parent list. No need to lock the child - not PID
3348 * hashed yet and not running, so nobody can access it.
3350 mutex_lock(&parent_ctx
->mutex
);
3353 * We dont have to disable NMIs - we are only looking at
3354 * the list, not manipulating it:
3356 list_for_each_entry(counter
, &parent_ctx
->counter_list
, list_entry
) {
3357 if (!counter
->hw_event
.inherit
)
3360 if (inherit_group(counter
, parent
,
3361 parent_ctx
, child
, child_ctx
))
3365 mutex_unlock(&parent_ctx
->mutex
);
3368 static void __cpuinit
perf_counter_init_cpu(int cpu
)
3370 struct perf_cpu_context
*cpuctx
;
3372 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3373 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
3375 spin_lock(&perf_resource_lock
);
3376 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
3377 spin_unlock(&perf_resource_lock
);
3379 hw_perf_counter_setup(cpu
);
3382 #ifdef CONFIG_HOTPLUG_CPU
3383 static void __perf_counter_exit_cpu(void *info
)
3385 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3386 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3387 struct perf_counter
*counter
, *tmp
;
3389 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
3390 __perf_counter_remove_from_context(counter
);
3392 static void perf_counter_exit_cpu(int cpu
)
3394 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3395 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3397 mutex_lock(&ctx
->mutex
);
3398 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
3399 mutex_unlock(&ctx
->mutex
);
3402 static inline void perf_counter_exit_cpu(int cpu
) { }
3405 static int __cpuinit
3406 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
3408 unsigned int cpu
= (long)hcpu
;
3412 case CPU_UP_PREPARE
:
3413 case CPU_UP_PREPARE_FROZEN
:
3414 perf_counter_init_cpu(cpu
);
3417 case CPU_DOWN_PREPARE
:
3418 case CPU_DOWN_PREPARE_FROZEN
:
3419 perf_counter_exit_cpu(cpu
);
3429 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
3430 .notifier_call
= perf_cpu_notify
,
3433 void __init
perf_counter_init(void)
3435 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
3436 (void *)(long)smp_processor_id());
3437 register_cpu_notifier(&perf_cpu_nb
);
3440 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
3442 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
3446 perf_set_reserve_percpu(struct sysdev_class
*class,
3450 struct perf_cpu_context
*cpuctx
;
3454 err
= strict_strtoul(buf
, 10, &val
);
3457 if (val
> perf_max_counters
)
3460 spin_lock(&perf_resource_lock
);
3461 perf_reserved_percpu
= val
;
3462 for_each_online_cpu(cpu
) {
3463 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3464 spin_lock_irq(&cpuctx
->ctx
.lock
);
3465 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
3466 perf_max_counters
- perf_reserved_percpu
);
3467 cpuctx
->max_pertask
= mpt
;
3468 spin_unlock_irq(&cpuctx
->ctx
.lock
);
3470 spin_unlock(&perf_resource_lock
);
3475 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
3477 return sprintf(buf
, "%d\n", perf_overcommit
);
3481 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
3486 err
= strict_strtoul(buf
, 10, &val
);
3492 spin_lock(&perf_resource_lock
);
3493 perf_overcommit
= val
;
3494 spin_unlock(&perf_resource_lock
);
3499 static SYSDEV_CLASS_ATTR(
3502 perf_show_reserve_percpu
,
3503 perf_set_reserve_percpu
3506 static SYSDEV_CLASS_ATTR(
3509 perf_show_overcommit
,
3513 static struct attribute
*perfclass_attrs
[] = {
3514 &attr_reserve_percpu
.attr
,
3515 &attr_overcommit
.attr
,
3519 static struct attribute_group perfclass_attr_group
= {
3520 .attrs
= perfclass_attrs
,
3521 .name
= "perf_counters",
3524 static int __init
perf_counter_sysfs_init(void)
3526 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
3527 &perfclass_attr_group
);
3529 device_initcall(perf_counter_sysfs_init
);