2 * Performance counter core code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
8 * For licensing details see kernel-base/COPYING
13 #include <linux/cpu.h>
14 #include <linux/smp.h>
15 #include <linux/file.h>
16 #include <linux/poll.h>
17 #include <linux/sysfs.h>
18 #include <linux/ptrace.h>
19 #include <linux/percpu.h>
20 #include <linux/vmstat.h>
21 #include <linux/hardirq.h>
22 #include <linux/rculist.h>
23 #include <linux/uaccess.h>
24 #include <linux/syscalls.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/kernel_stat.h>
27 #include <linux/perf_counter.h>
28 #include <linux/dcache.h>
30 #include <asm/irq_regs.h>
33 * Each CPU has a list of per CPU counters:
35 DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
37 int perf_max_counters __read_mostly
= 1;
38 static int perf_reserved_percpu __read_mostly
;
39 static int perf_overcommit __read_mostly
= 1;
42 * Mutex for (sysadmin-configurable) counter reservations:
44 static DEFINE_MUTEX(perf_resource_mutex
);
47 * Architecture provided APIs - weak aliases:
49 extern __weak
const struct hw_perf_counter_ops
*
50 hw_perf_counter_init(struct perf_counter
*counter
)
55 u64 __weak
hw_perf_save_disable(void) { return 0; }
56 void __weak
hw_perf_restore(u64 ctrl
) { barrier(); }
57 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
58 int __weak
hw_perf_group_sched_in(struct perf_counter
*group_leader
,
59 struct perf_cpu_context
*cpuctx
,
60 struct perf_counter_context
*ctx
, int cpu
)
65 void __weak
perf_counter_print_debug(void) { }
68 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
70 struct perf_counter
*group_leader
= counter
->group_leader
;
73 * Depending on whether it is a standalone or sibling counter,
74 * add it straight to the context's counter list, or to the group
75 * leader's sibling list:
77 if (counter
->group_leader
== counter
)
78 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
80 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
81 group_leader
->nr_siblings
++;
84 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
88 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
90 struct perf_counter
*sibling
, *tmp
;
92 list_del_init(&counter
->list_entry
);
93 list_del_rcu(&counter
->event_entry
);
95 if (counter
->group_leader
!= counter
)
96 counter
->group_leader
->nr_siblings
--;
99 * If this was a group counter with sibling counters then
100 * upgrade the siblings to singleton counters by adding them
101 * to the context list directly:
103 list_for_each_entry_safe(sibling
, tmp
,
104 &counter
->sibling_list
, list_entry
) {
106 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
107 sibling
->group_leader
= sibling
;
112 counter_sched_out(struct perf_counter
*counter
,
113 struct perf_cpu_context
*cpuctx
,
114 struct perf_counter_context
*ctx
)
116 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
119 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
120 counter
->tstamp_stopped
= ctx
->time_now
;
121 counter
->hw_ops
->disable(counter
);
124 if (!is_software_counter(counter
))
125 cpuctx
->active_oncpu
--;
127 if (counter
->hw_event
.exclusive
|| !cpuctx
->active_oncpu
)
128 cpuctx
->exclusive
= 0;
132 group_sched_out(struct perf_counter
*group_counter
,
133 struct perf_cpu_context
*cpuctx
,
134 struct perf_counter_context
*ctx
)
136 struct perf_counter
*counter
;
138 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
141 counter_sched_out(group_counter
, cpuctx
, ctx
);
144 * Schedule out siblings (if any):
146 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
147 counter_sched_out(counter
, cpuctx
, ctx
);
149 if (group_counter
->hw_event
.exclusive
)
150 cpuctx
->exclusive
= 0;
154 * Cross CPU call to remove a performance counter
156 * We disable the counter on the hardware level first. After that we
157 * remove it from the context list.
159 static void __perf_counter_remove_from_context(void *info
)
161 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
162 struct perf_counter
*counter
= info
;
163 struct perf_counter_context
*ctx
= counter
->ctx
;
168 * If this is a task context, we need to check whether it is
169 * the current task context of this cpu. If not it has been
170 * scheduled out before the smp call arrived.
172 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
175 curr_rq_lock_irq_save(&flags
);
176 spin_lock(&ctx
->lock
);
178 counter_sched_out(counter
, cpuctx
, ctx
);
180 counter
->task
= NULL
;
184 * Protect the list operation against NMI by disabling the
185 * counters on a global level. NOP for non NMI based counters.
187 perf_flags
= hw_perf_save_disable();
188 list_del_counter(counter
, ctx
);
189 hw_perf_restore(perf_flags
);
193 * Allow more per task counters with respect to the
196 cpuctx
->max_pertask
=
197 min(perf_max_counters
- ctx
->nr_counters
,
198 perf_max_counters
- perf_reserved_percpu
);
201 spin_unlock(&ctx
->lock
);
202 curr_rq_unlock_irq_restore(&flags
);
207 * Remove the counter from a task's (or a CPU's) list of counters.
209 * Must be called with counter->mutex and ctx->mutex held.
211 * CPU counters are removed with a smp call. For task counters we only
212 * call when the task is on a CPU.
214 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
216 struct perf_counter_context
*ctx
= counter
->ctx
;
217 struct task_struct
*task
= ctx
->task
;
221 * Per cpu counters are removed via an smp call and
222 * the removal is always sucessful.
224 smp_call_function_single(counter
->cpu
,
225 __perf_counter_remove_from_context
,
231 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
234 spin_lock_irq(&ctx
->lock
);
236 * If the context is active we need to retry the smp call.
238 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
239 spin_unlock_irq(&ctx
->lock
);
244 * The lock prevents that this context is scheduled in so we
245 * can remove the counter safely, if the call above did not
248 if (!list_empty(&counter
->list_entry
)) {
250 list_del_counter(counter
, ctx
);
251 counter
->task
= NULL
;
253 spin_unlock_irq(&ctx
->lock
);
257 * Get the current time for this context.
258 * If this is a task context, we use the task's task clock,
259 * or for a per-cpu context, we use the cpu clock.
261 static u64
get_context_time(struct perf_counter_context
*ctx
, int update
)
263 struct task_struct
*curr
= ctx
->task
;
266 return cpu_clock(smp_processor_id());
268 return __task_delta_exec(curr
, update
) + curr
->se
.sum_exec_runtime
;
272 * Update the record of the current time in a context.
274 static void update_context_time(struct perf_counter_context
*ctx
, int update
)
276 ctx
->time_now
= get_context_time(ctx
, update
) - ctx
->time_lost
;
280 * Update the total_time_enabled and total_time_running fields for a counter.
282 static void update_counter_times(struct perf_counter
*counter
)
284 struct perf_counter_context
*ctx
= counter
->ctx
;
287 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
288 counter
->total_time_enabled
= ctx
->time_now
-
289 counter
->tstamp_enabled
;
290 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
291 run_end
= counter
->tstamp_stopped
;
293 run_end
= ctx
->time_now
;
294 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
299 * Update total_time_enabled and total_time_running for all counters in a group.
301 static void update_group_times(struct perf_counter
*leader
)
303 struct perf_counter
*counter
;
305 update_counter_times(leader
);
306 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
307 update_counter_times(counter
);
311 * Cross CPU call to disable a performance counter
313 static void __perf_counter_disable(void *info
)
315 struct perf_counter
*counter
= info
;
316 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
317 struct perf_counter_context
*ctx
= counter
->ctx
;
321 * If this is a per-task counter, need to check whether this
322 * counter's task is the current task on this cpu.
324 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
327 curr_rq_lock_irq_save(&flags
);
328 spin_lock(&ctx
->lock
);
331 * If the counter is on, turn it off.
332 * If it is in error state, leave it in error state.
334 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
335 update_context_time(ctx
, 1);
336 update_counter_times(counter
);
337 if (counter
== counter
->group_leader
)
338 group_sched_out(counter
, cpuctx
, ctx
);
340 counter_sched_out(counter
, cpuctx
, ctx
);
341 counter
->state
= PERF_COUNTER_STATE_OFF
;
344 spin_unlock(&ctx
->lock
);
345 curr_rq_unlock_irq_restore(&flags
);
351 static void perf_counter_disable(struct perf_counter
*counter
)
353 struct perf_counter_context
*ctx
= counter
->ctx
;
354 struct task_struct
*task
= ctx
->task
;
358 * Disable the counter on the cpu that it's on
360 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
366 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
368 spin_lock_irq(&ctx
->lock
);
370 * If the counter is still active, we need to retry the cross-call.
372 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
373 spin_unlock_irq(&ctx
->lock
);
378 * Since we have the lock this context can't be scheduled
379 * in, so we can change the state safely.
381 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
382 update_counter_times(counter
);
383 counter
->state
= PERF_COUNTER_STATE_OFF
;
386 spin_unlock_irq(&ctx
->lock
);
390 * Disable a counter and all its children.
392 static void perf_counter_disable_family(struct perf_counter
*counter
)
394 struct perf_counter
*child
;
396 perf_counter_disable(counter
);
399 * Lock the mutex to protect the list of children
401 mutex_lock(&counter
->mutex
);
402 list_for_each_entry(child
, &counter
->child_list
, child_list
)
403 perf_counter_disable(child
);
404 mutex_unlock(&counter
->mutex
);
408 counter_sched_in(struct perf_counter
*counter
,
409 struct perf_cpu_context
*cpuctx
,
410 struct perf_counter_context
*ctx
,
413 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
416 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
417 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
419 * The new state must be visible before we turn it on in the hardware:
423 if (counter
->hw_ops
->enable(counter
)) {
424 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
429 counter
->tstamp_running
+= ctx
->time_now
- counter
->tstamp_stopped
;
431 if (!is_software_counter(counter
))
432 cpuctx
->active_oncpu
++;
435 if (counter
->hw_event
.exclusive
)
436 cpuctx
->exclusive
= 1;
442 * Return 1 for a group consisting entirely of software counters,
443 * 0 if the group contains any hardware counters.
445 static int is_software_only_group(struct perf_counter
*leader
)
447 struct perf_counter
*counter
;
449 if (!is_software_counter(leader
))
452 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
453 if (!is_software_counter(counter
))
460 * Work out whether we can put this counter group on the CPU now.
462 static int group_can_go_on(struct perf_counter
*counter
,
463 struct perf_cpu_context
*cpuctx
,
467 * Groups consisting entirely of software counters can always go on.
469 if (is_software_only_group(counter
))
472 * If an exclusive group is already on, no other hardware
473 * counters can go on.
475 if (cpuctx
->exclusive
)
478 * If this group is exclusive and there are already
479 * counters on the CPU, it can't go on.
481 if (counter
->hw_event
.exclusive
&& cpuctx
->active_oncpu
)
484 * Otherwise, try to add it if all previous groups were able
490 static void add_counter_to_ctx(struct perf_counter
*counter
,
491 struct perf_counter_context
*ctx
)
493 list_add_counter(counter
, ctx
);
495 counter
->prev_state
= PERF_COUNTER_STATE_OFF
;
496 counter
->tstamp_enabled
= ctx
->time_now
;
497 counter
->tstamp_running
= ctx
->time_now
;
498 counter
->tstamp_stopped
= ctx
->time_now
;
502 * Cross CPU call to install and enable a performance counter
504 static void __perf_install_in_context(void *info
)
506 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
507 struct perf_counter
*counter
= info
;
508 struct perf_counter_context
*ctx
= counter
->ctx
;
509 struct perf_counter
*leader
= counter
->group_leader
;
510 int cpu
= smp_processor_id();
516 * If this is a task context, we need to check whether it is
517 * the current task context of this cpu. If not it has been
518 * scheduled out before the smp call arrived.
520 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
523 curr_rq_lock_irq_save(&flags
);
524 spin_lock(&ctx
->lock
);
525 update_context_time(ctx
, 1);
528 * Protect the list operation against NMI by disabling the
529 * counters on a global level. NOP for non NMI based counters.
531 perf_flags
= hw_perf_save_disable();
533 add_counter_to_ctx(counter
, ctx
);
536 * Don't put the counter on if it is disabled or if
537 * it is in a group and the group isn't on.
539 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
540 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
544 * An exclusive counter can't go on if there are already active
545 * hardware counters, and no hardware counter can go on if there
546 * is already an exclusive counter on.
548 if (!group_can_go_on(counter
, cpuctx
, 1))
551 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
555 * This counter couldn't go on. If it is in a group
556 * then we have to pull the whole group off.
557 * If the counter group is pinned then put it in error state.
559 if (leader
!= counter
)
560 group_sched_out(leader
, cpuctx
, ctx
);
561 if (leader
->hw_event
.pinned
) {
562 update_group_times(leader
);
563 leader
->state
= PERF_COUNTER_STATE_ERROR
;
567 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
568 cpuctx
->max_pertask
--;
571 hw_perf_restore(perf_flags
);
573 spin_unlock(&ctx
->lock
);
574 curr_rq_unlock_irq_restore(&flags
);
578 * Attach a performance counter to a context
580 * First we add the counter to the list with the hardware enable bit
581 * in counter->hw_config cleared.
583 * If the counter is attached to a task which is on a CPU we use a smp
584 * call to enable it in the task context. The task might have been
585 * scheduled away, but we check this in the smp call again.
587 * Must be called with ctx->mutex held.
590 perf_install_in_context(struct perf_counter_context
*ctx
,
591 struct perf_counter
*counter
,
594 struct task_struct
*task
= ctx
->task
;
598 * Per cpu counters are installed via an smp call and
599 * the install is always sucessful.
601 smp_call_function_single(cpu
, __perf_install_in_context
,
606 counter
->task
= task
;
608 task_oncpu_function_call(task
, __perf_install_in_context
,
611 spin_lock_irq(&ctx
->lock
);
613 * we need to retry the smp call.
615 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
616 spin_unlock_irq(&ctx
->lock
);
621 * The lock prevents that this context is scheduled in so we
622 * can add the counter safely, if it the call above did not
625 if (list_empty(&counter
->list_entry
))
626 add_counter_to_ctx(counter
, ctx
);
627 spin_unlock_irq(&ctx
->lock
);
631 * Cross CPU call to enable a performance counter
633 static void __perf_counter_enable(void *info
)
635 struct perf_counter
*counter
= info
;
636 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
637 struct perf_counter_context
*ctx
= counter
->ctx
;
638 struct perf_counter
*leader
= counter
->group_leader
;
643 * If this is a per-task counter, need to check whether this
644 * counter's task is the current task on this cpu.
646 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
649 curr_rq_lock_irq_save(&flags
);
650 spin_lock(&ctx
->lock
);
651 update_context_time(ctx
, 1);
653 counter
->prev_state
= counter
->state
;
654 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
656 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
657 counter
->tstamp_enabled
= ctx
->time_now
- counter
->total_time_enabled
;
660 * If the counter is in a group and isn't the group leader,
661 * then don't put it on unless the group is on.
663 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
666 if (!group_can_go_on(counter
, cpuctx
, 1))
669 err
= counter_sched_in(counter
, cpuctx
, ctx
,
674 * If this counter can't go on and it's part of a
675 * group, then the whole group has to come off.
677 if (leader
!= counter
)
678 group_sched_out(leader
, cpuctx
, ctx
);
679 if (leader
->hw_event
.pinned
) {
680 update_group_times(leader
);
681 leader
->state
= PERF_COUNTER_STATE_ERROR
;
686 spin_unlock(&ctx
->lock
);
687 curr_rq_unlock_irq_restore(&flags
);
693 static void perf_counter_enable(struct perf_counter
*counter
)
695 struct perf_counter_context
*ctx
= counter
->ctx
;
696 struct task_struct
*task
= ctx
->task
;
700 * Enable the counter on the cpu that it's on
702 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
707 spin_lock_irq(&ctx
->lock
);
708 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
712 * If the counter is in error state, clear that first.
713 * That way, if we see the counter in error state below, we
714 * know that it has gone back into error state, as distinct
715 * from the task having been scheduled away before the
716 * cross-call arrived.
718 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
719 counter
->state
= PERF_COUNTER_STATE_OFF
;
722 spin_unlock_irq(&ctx
->lock
);
723 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
725 spin_lock_irq(&ctx
->lock
);
728 * If the context is active and the counter is still off,
729 * we need to retry the cross-call.
731 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
735 * Since we have the lock this context can't be scheduled
736 * in, so we can change the state safely.
738 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
739 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
740 counter
->tstamp_enabled
= ctx
->time_now
-
741 counter
->total_time_enabled
;
744 spin_unlock_irq(&ctx
->lock
);
748 * Enable a counter and all its children.
750 static void perf_counter_enable_family(struct perf_counter
*counter
)
752 struct perf_counter
*child
;
754 perf_counter_enable(counter
);
757 * Lock the mutex to protect the list of children
759 mutex_lock(&counter
->mutex
);
760 list_for_each_entry(child
, &counter
->child_list
, child_list
)
761 perf_counter_enable(child
);
762 mutex_unlock(&counter
->mutex
);
765 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
766 struct perf_cpu_context
*cpuctx
)
768 struct perf_counter
*counter
;
771 spin_lock(&ctx
->lock
);
773 if (likely(!ctx
->nr_counters
))
775 update_context_time(ctx
, 0);
777 flags
= hw_perf_save_disable();
778 if (ctx
->nr_active
) {
779 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
)
780 group_sched_out(counter
, cpuctx
, ctx
);
782 hw_perf_restore(flags
);
784 spin_unlock(&ctx
->lock
);
788 * Called from scheduler to remove the counters of the current task,
789 * with interrupts disabled.
791 * We stop each counter and update the counter value in counter->count.
793 * This does not protect us against NMI, but disable()
794 * sets the disabled bit in the control field of counter _before_
795 * accessing the counter control register. If a NMI hits, then it will
796 * not restart the counter.
798 void perf_counter_task_sched_out(struct task_struct
*task
, int cpu
)
800 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
801 struct perf_counter_context
*ctx
= &task
->perf_counter_ctx
;
802 struct pt_regs
*regs
;
804 if (likely(!cpuctx
->task_ctx
))
807 regs
= task_pt_regs(task
);
808 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES
, 1, 1, regs
);
809 __perf_counter_sched_out(ctx
, cpuctx
);
811 cpuctx
->task_ctx
= NULL
;
814 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
816 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
820 group_sched_in(struct perf_counter
*group_counter
,
821 struct perf_cpu_context
*cpuctx
,
822 struct perf_counter_context
*ctx
,
825 struct perf_counter
*counter
, *partial_group
;
828 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
831 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
833 return ret
< 0 ? ret
: 0;
835 group_counter
->prev_state
= group_counter
->state
;
836 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
840 * Schedule in siblings as one group (if any):
842 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
843 counter
->prev_state
= counter
->state
;
844 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
845 partial_group
= counter
;
854 * Groups can be scheduled in as one unit only, so undo any
855 * partial group before returning:
857 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
858 if (counter
== partial_group
)
860 counter_sched_out(counter
, cpuctx
, ctx
);
862 counter_sched_out(group_counter
, cpuctx
, ctx
);
868 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
869 struct perf_cpu_context
*cpuctx
, int cpu
)
871 struct perf_counter
*counter
;
875 spin_lock(&ctx
->lock
);
877 if (likely(!ctx
->nr_counters
))
881 * Add any time since the last sched_out to the lost time
882 * so it doesn't get included in the total_time_enabled and
883 * total_time_running measures for counters in the context.
885 ctx
->time_lost
= get_context_time(ctx
, 0) - ctx
->time_now
;
887 flags
= hw_perf_save_disable();
890 * First go through the list and put on any pinned groups
891 * in order to give them the best chance of going on.
893 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
894 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
895 !counter
->hw_event
.pinned
)
897 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
900 if (group_can_go_on(counter
, cpuctx
, 1))
901 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
904 * If this pinned group hasn't been scheduled,
905 * put it in error state.
907 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
908 update_group_times(counter
);
909 counter
->state
= PERF_COUNTER_STATE_ERROR
;
913 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
915 * Ignore counters in OFF or ERROR state, and
916 * ignore pinned counters since we did them already.
918 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
919 counter
->hw_event
.pinned
)
923 * Listen to the 'cpu' scheduling filter constraint
926 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
929 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
930 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
934 hw_perf_restore(flags
);
936 spin_unlock(&ctx
->lock
);
940 * Called from scheduler to add the counters of the current task
941 * with interrupts disabled.
943 * We restore the counter value and then enable it.
945 * This does not protect us against NMI, but enable()
946 * sets the enabled bit in the control field of counter _before_
947 * accessing the counter control register. If a NMI hits, then it will
948 * keep the counter running.
950 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
952 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
953 struct perf_counter_context
*ctx
= &task
->perf_counter_ctx
;
955 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
956 cpuctx
->task_ctx
= ctx
;
959 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
961 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
963 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
966 int perf_counter_task_disable(void)
968 struct task_struct
*curr
= current
;
969 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
970 struct perf_counter
*counter
;
975 if (likely(!ctx
->nr_counters
))
978 curr_rq_lock_irq_save(&flags
);
979 cpu
= smp_processor_id();
981 /* force the update of the task clock: */
982 __task_delta_exec(curr
, 1);
984 perf_counter_task_sched_out(curr
, cpu
);
986 spin_lock(&ctx
->lock
);
989 * Disable all the counters:
991 perf_flags
= hw_perf_save_disable();
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
;
1000 hw_perf_restore(perf_flags
);
1002 spin_unlock(&ctx
->lock
);
1004 curr_rq_unlock_irq_restore(&flags
);
1009 int perf_counter_task_enable(void)
1011 struct task_struct
*curr
= current
;
1012 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
1013 struct perf_counter
*counter
;
1014 unsigned long flags
;
1018 if (likely(!ctx
->nr_counters
))
1021 curr_rq_lock_irq_save(&flags
);
1022 cpu
= smp_processor_id();
1024 /* force the update of the task clock: */
1025 __task_delta_exec(curr
, 1);
1027 perf_counter_task_sched_out(curr
, cpu
);
1029 spin_lock(&ctx
->lock
);
1032 * Disable all the counters:
1034 perf_flags
= hw_perf_save_disable();
1036 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1037 if (counter
->state
> PERF_COUNTER_STATE_OFF
)
1039 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
1040 counter
->tstamp_enabled
= ctx
->time_now
-
1041 counter
->total_time_enabled
;
1042 counter
->hw_event
.disabled
= 0;
1044 hw_perf_restore(perf_flags
);
1046 spin_unlock(&ctx
->lock
);
1048 perf_counter_task_sched_in(curr
, cpu
);
1050 curr_rq_unlock_irq_restore(&flags
);
1056 * Round-robin a context's counters:
1058 static void rotate_ctx(struct perf_counter_context
*ctx
)
1060 struct perf_counter
*counter
;
1063 if (!ctx
->nr_counters
)
1066 spin_lock(&ctx
->lock
);
1068 * Rotate the first entry last (works just fine for group counters too):
1070 perf_flags
= hw_perf_save_disable();
1071 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1072 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1075 hw_perf_restore(perf_flags
);
1077 spin_unlock(&ctx
->lock
);
1080 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1082 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1083 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
1084 const int rotate_percpu
= 0;
1087 perf_counter_cpu_sched_out(cpuctx
);
1088 perf_counter_task_sched_out(curr
, cpu
);
1091 rotate_ctx(&cpuctx
->ctx
);
1095 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1096 perf_counter_task_sched_in(curr
, cpu
);
1100 * Cross CPU call to read the hardware counter
1102 static void __read(void *info
)
1104 struct perf_counter
*counter
= info
;
1105 struct perf_counter_context
*ctx
= counter
->ctx
;
1106 unsigned long flags
;
1108 curr_rq_lock_irq_save(&flags
);
1110 update_context_time(ctx
, 1);
1111 counter
->hw_ops
->read(counter
);
1112 update_counter_times(counter
);
1113 curr_rq_unlock_irq_restore(&flags
);
1116 static u64
perf_counter_read(struct perf_counter
*counter
)
1119 * If counter is enabled and currently active on a CPU, update the
1120 * value in the counter structure:
1122 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1123 smp_call_function_single(counter
->oncpu
,
1124 __read
, counter
, 1);
1125 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1126 update_counter_times(counter
);
1129 return atomic64_read(&counter
->count
);
1132 static void put_context(struct perf_counter_context
*ctx
)
1135 put_task_struct(ctx
->task
);
1138 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1140 struct perf_cpu_context
*cpuctx
;
1141 struct perf_counter_context
*ctx
;
1142 struct task_struct
*task
;
1145 * If cpu is not a wildcard then this is a percpu counter:
1148 /* Must be root to operate on a CPU counter: */
1149 if (!capable(CAP_SYS_ADMIN
))
1150 return ERR_PTR(-EACCES
);
1152 if (cpu
< 0 || cpu
> num_possible_cpus())
1153 return ERR_PTR(-EINVAL
);
1156 * We could be clever and allow to attach a counter to an
1157 * offline CPU and activate it when the CPU comes up, but
1160 if (!cpu_isset(cpu
, cpu_online_map
))
1161 return ERR_PTR(-ENODEV
);
1163 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1173 task
= find_task_by_vpid(pid
);
1175 get_task_struct(task
);
1179 return ERR_PTR(-ESRCH
);
1181 ctx
= &task
->perf_counter_ctx
;
1184 /* Reuse ptrace permission checks for now. */
1185 if (!ptrace_may_access(task
, PTRACE_MODE_READ
)) {
1187 return ERR_PTR(-EACCES
);
1193 static void free_counter_rcu(struct rcu_head
*head
)
1195 struct perf_counter
*counter
;
1197 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1201 static void perf_pending_sync(struct perf_counter
*counter
);
1203 static void free_counter(struct perf_counter
*counter
)
1205 perf_pending_sync(counter
);
1207 if (counter
->destroy
)
1208 counter
->destroy(counter
);
1210 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1214 * Called when the last reference to the file is gone.
1216 static int perf_release(struct inode
*inode
, struct file
*file
)
1218 struct perf_counter
*counter
= file
->private_data
;
1219 struct perf_counter_context
*ctx
= counter
->ctx
;
1221 file
->private_data
= NULL
;
1223 mutex_lock(&ctx
->mutex
);
1224 mutex_lock(&counter
->mutex
);
1226 perf_counter_remove_from_context(counter
);
1228 mutex_unlock(&counter
->mutex
);
1229 mutex_unlock(&ctx
->mutex
);
1231 free_counter(counter
);
1238 * Read the performance counter - simple non blocking version for now
1241 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1247 * Return end-of-file for a read on a counter that is in
1248 * error state (i.e. because it was pinned but it couldn't be
1249 * scheduled on to the CPU at some point).
1251 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1254 mutex_lock(&counter
->mutex
);
1255 values
[0] = perf_counter_read(counter
);
1257 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1258 values
[n
++] = counter
->total_time_enabled
+
1259 atomic64_read(&counter
->child_total_time_enabled
);
1260 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1261 values
[n
++] = counter
->total_time_running
+
1262 atomic64_read(&counter
->child_total_time_running
);
1263 mutex_unlock(&counter
->mutex
);
1265 if (count
< n
* sizeof(u64
))
1267 count
= n
* sizeof(u64
);
1269 if (copy_to_user(buf
, values
, count
))
1276 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1278 struct perf_counter
*counter
= file
->private_data
;
1280 return perf_read_hw(counter
, buf
, count
);
1283 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1285 struct perf_counter
*counter
= file
->private_data
;
1286 struct perf_mmap_data
*data
;
1287 unsigned int events
;
1290 data
= rcu_dereference(counter
->data
);
1292 events
= atomic_xchg(&data
->wakeup
, 0);
1297 poll_wait(file
, &counter
->waitq
, wait
);
1302 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1304 struct perf_counter
*counter
= file
->private_data
;
1308 case PERF_COUNTER_IOC_ENABLE
:
1309 perf_counter_enable_family(counter
);
1311 case PERF_COUNTER_IOC_DISABLE
:
1312 perf_counter_disable_family(counter
);
1321 * Callers need to ensure there can be no nesting of this function, otherwise
1322 * the seqlock logic goes bad. We can not serialize this because the arch
1323 * code calls this from NMI context.
1325 void perf_counter_update_userpage(struct perf_counter
*counter
)
1327 struct perf_mmap_data
*data
;
1328 struct perf_counter_mmap_page
*userpg
;
1331 data
= rcu_dereference(counter
->data
);
1335 userpg
= data
->user_page
;
1338 * Disable preemption so as to not let the corresponding user-space
1339 * spin too long if we get preempted.
1344 userpg
->index
= counter
->hw
.idx
;
1345 userpg
->offset
= atomic64_read(&counter
->count
);
1346 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1347 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1356 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1358 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1359 struct perf_mmap_data
*data
;
1360 int ret
= VM_FAULT_SIGBUS
;
1363 data
= rcu_dereference(counter
->data
);
1367 if (vmf
->pgoff
== 0) {
1368 vmf
->page
= virt_to_page(data
->user_page
);
1370 int nr
= vmf
->pgoff
- 1;
1372 if ((unsigned)nr
> data
->nr_pages
)
1375 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1377 get_page(vmf
->page
);
1385 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1387 struct perf_mmap_data
*data
;
1391 WARN_ON(atomic_read(&counter
->mmap_count
));
1393 size
= sizeof(struct perf_mmap_data
);
1394 size
+= nr_pages
* sizeof(void *);
1396 data
= kzalloc(size
, GFP_KERNEL
);
1400 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1401 if (!data
->user_page
)
1402 goto fail_user_page
;
1404 for (i
= 0; i
< nr_pages
; i
++) {
1405 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1406 if (!data
->data_pages
[i
])
1407 goto fail_data_pages
;
1410 data
->nr_pages
= nr_pages
;
1412 rcu_assign_pointer(counter
->data
, data
);
1417 for (i
--; i
>= 0; i
--)
1418 free_page((unsigned long)data
->data_pages
[i
]);
1420 free_page((unsigned long)data
->user_page
);
1429 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1431 struct perf_mmap_data
*data
= container_of(rcu_head
,
1432 struct perf_mmap_data
, rcu_head
);
1435 free_page((unsigned long)data
->user_page
);
1436 for (i
= 0; i
< data
->nr_pages
; i
++)
1437 free_page((unsigned long)data
->data_pages
[i
]);
1441 static void perf_mmap_data_free(struct perf_counter
*counter
)
1443 struct perf_mmap_data
*data
= counter
->data
;
1445 WARN_ON(atomic_read(&counter
->mmap_count
));
1447 rcu_assign_pointer(counter
->data
, NULL
);
1448 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1451 static void perf_mmap_open(struct vm_area_struct
*vma
)
1453 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1455 atomic_inc(&counter
->mmap_count
);
1458 static void perf_mmap_close(struct vm_area_struct
*vma
)
1460 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1462 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
,
1463 &counter
->mmap_mutex
)) {
1464 perf_mmap_data_free(counter
);
1465 mutex_unlock(&counter
->mmap_mutex
);
1469 static struct vm_operations_struct perf_mmap_vmops
= {
1470 .open
= perf_mmap_open
,
1471 .close
= perf_mmap_close
,
1472 .fault
= perf_mmap_fault
,
1475 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1477 struct perf_counter
*counter
= file
->private_data
;
1478 unsigned long vma_size
;
1479 unsigned long nr_pages
;
1480 unsigned long locked
, lock_limit
;
1483 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1486 vma_size
= vma
->vm_end
- vma
->vm_start
;
1487 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1490 * If we have data pages ensure they're a power-of-two number, so we
1491 * can do bitmasks instead of modulo.
1493 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1496 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1499 if (vma
->vm_pgoff
!= 0)
1502 locked
= vma_size
>> PAGE_SHIFT
;
1503 locked
+= vma
->vm_mm
->locked_vm
;
1505 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1506 lock_limit
>>= PAGE_SHIFT
;
1508 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
))
1511 mutex_lock(&counter
->mmap_mutex
);
1512 if (atomic_inc_not_zero(&counter
->mmap_count
))
1515 WARN_ON(counter
->data
);
1516 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1518 atomic_set(&counter
->mmap_count
, 1);
1520 mutex_unlock(&counter
->mmap_mutex
);
1522 vma
->vm_flags
&= ~VM_MAYWRITE
;
1523 vma
->vm_flags
|= VM_RESERVED
;
1524 vma
->vm_ops
= &perf_mmap_vmops
;
1529 static const struct file_operations perf_fops
= {
1530 .release
= perf_release
,
1533 .unlocked_ioctl
= perf_ioctl
,
1534 .compat_ioctl
= perf_ioctl
,
1539 * Perf counter wakeup
1541 * If there's data, ensure we set the poll() state and publish everything
1542 * to user-space before waking everybody up.
1545 void perf_counter_wakeup(struct perf_counter
*counter
)
1547 struct perf_mmap_data
*data
;
1550 data
= rcu_dereference(counter
->data
);
1552 (void)atomic_xchg(&data
->wakeup
, POLL_IN
);
1554 * Ensure all data writes are issued before updating the
1555 * user-space data head information. The matching rmb()
1556 * will be in userspace after reading this value.
1559 data
->user_page
->data_head
= atomic_read(&data
->head
);
1563 wake_up_all(&counter
->waitq
);
1569 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1571 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1572 * single linked list and use cmpxchg() to add entries lockless.
1575 #define PENDING_TAIL ((struct perf_wakeup_entry *)-1UL)
1577 static DEFINE_PER_CPU(struct perf_wakeup_entry
*, perf_wakeup_head
) = {
1581 static void perf_pending_queue(struct perf_counter
*counter
)
1583 struct perf_wakeup_entry
**head
;
1584 struct perf_wakeup_entry
*prev
, *next
;
1586 if (cmpxchg(&counter
->wakeup
.next
, NULL
, PENDING_TAIL
) != NULL
)
1589 head
= &get_cpu_var(perf_wakeup_head
);
1592 prev
= counter
->wakeup
.next
= *head
;
1593 next
= &counter
->wakeup
;
1594 } while (cmpxchg(head
, prev
, next
) != prev
);
1596 set_perf_counter_pending();
1598 put_cpu_var(perf_wakeup_head
);
1601 static int __perf_pending_run(void)
1603 struct perf_wakeup_entry
*list
;
1606 list
= xchg(&__get_cpu_var(perf_wakeup_head
), PENDING_TAIL
);
1607 while (list
!= PENDING_TAIL
) {
1608 struct perf_counter
*counter
= container_of(list
,
1609 struct perf_counter
, wakeup
);
1613 counter
->wakeup
.next
= NULL
;
1615 * Ensure we observe the unqueue before we issue the wakeup,
1616 * so that we won't be waiting forever.
1617 * -- see perf_not_pending().
1621 perf_counter_wakeup(counter
);
1628 static inline int perf_not_pending(struct perf_counter
*counter
)
1631 * If we flush on whatever cpu we run, there is a chance we don't
1635 __perf_pending_run();
1639 * Ensure we see the proper queue state before going to sleep
1640 * so that we do not miss the wakeup. -- see perf_pending_handle()
1643 return counter
->wakeup
.next
== NULL
;
1646 static void perf_pending_sync(struct perf_counter
*counter
)
1648 wait_event(counter
->waitq
, perf_not_pending(counter
));
1651 void perf_counter_do_pending(void)
1653 __perf_pending_run();
1657 * Callchain support -- arch specific
1660 struct perf_callchain_entry
*
1661 __attribute__((weak
))
1662 perf_callchain(struct pt_regs
*regs
)
1671 struct perf_output_handle
{
1672 struct perf_counter
*counter
;
1673 struct perf_mmap_data
*data
;
1674 unsigned int offset
;
1680 static inline void __perf_output_wakeup(struct perf_output_handle
*handle
)
1683 perf_pending_queue(handle
->counter
);
1685 perf_counter_wakeup(handle
->counter
);
1688 static int perf_output_begin(struct perf_output_handle
*handle
,
1689 struct perf_counter
*counter
, unsigned int size
,
1692 struct perf_mmap_data
*data
;
1693 unsigned int offset
, head
;
1696 data
= rcu_dereference(counter
->data
);
1700 handle
->counter
= counter
;
1703 if (!data
->nr_pages
)
1707 offset
= head
= atomic_read(&data
->head
);
1709 } while (atomic_cmpxchg(&data
->head
, offset
, head
) != offset
);
1711 handle
->data
= data
;
1712 handle
->offset
= offset
;
1713 handle
->head
= head
;
1714 handle
->wakeup
= (offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
);
1719 __perf_output_wakeup(handle
);
1726 static void perf_output_copy(struct perf_output_handle
*handle
,
1727 void *buf
, unsigned int len
)
1729 unsigned int pages_mask
;
1730 unsigned int offset
;
1734 offset
= handle
->offset
;
1735 pages_mask
= handle
->data
->nr_pages
- 1;
1736 pages
= handle
->data
->data_pages
;
1739 unsigned int page_offset
;
1742 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
1743 page_offset
= offset
& (PAGE_SIZE
- 1);
1744 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
1746 memcpy(pages
[nr
] + page_offset
, buf
, size
);
1753 handle
->offset
= offset
;
1755 WARN_ON_ONCE(handle
->offset
> handle
->head
);
1758 #define perf_output_put(handle, x) \
1759 perf_output_copy((handle), &(x), sizeof(x))
1761 static void perf_output_end(struct perf_output_handle
*handle
)
1763 int wakeup_events
= handle
->counter
->hw_event
.wakeup_events
;
1765 if (wakeup_events
) {
1766 int events
= atomic_inc_return(&handle
->data
->events
);
1767 if (events
>= wakeup_events
) {
1768 atomic_sub(wakeup_events
, &handle
->data
->events
);
1769 __perf_output_wakeup(handle
);
1771 } else if (handle
->wakeup
)
1772 __perf_output_wakeup(handle
);
1776 void perf_counter_output(struct perf_counter
*counter
,
1777 int nmi
, struct pt_regs
*regs
)
1780 u64 record_type
= counter
->hw_event
.record_type
;
1781 struct perf_output_handle handle
;
1782 struct perf_event_header header
;
1791 struct perf_callchain_entry
*callchain
= NULL
;
1792 int callchain_size
= 0;
1794 header
.type
= PERF_EVENT_COUNTER_OVERFLOW
;
1795 header
.size
= sizeof(header
);
1797 if (record_type
& PERF_RECORD_IP
) {
1798 ip
= instruction_pointer(regs
);
1799 header
.type
|= __PERF_EVENT_IP
;
1800 header
.size
+= sizeof(ip
);
1803 if (record_type
& PERF_RECORD_TID
) {
1804 /* namespace issues */
1805 tid_entry
.pid
= current
->group_leader
->pid
;
1806 tid_entry
.tid
= current
->pid
;
1808 header
.type
|= __PERF_EVENT_TID
;
1809 header
.size
+= sizeof(tid_entry
);
1812 if (record_type
& PERF_RECORD_GROUP
) {
1813 header
.type
|= __PERF_EVENT_GROUP
;
1814 header
.size
+= sizeof(u64
) +
1815 counter
->nr_siblings
* sizeof(group_entry
);
1818 if (record_type
& PERF_RECORD_CALLCHAIN
) {
1819 callchain
= perf_callchain(regs
);
1822 callchain_size
= (2 + callchain
->nr
) * sizeof(u64
);
1824 header
.type
|= __PERF_EVENT_CALLCHAIN
;
1825 header
.size
+= callchain_size
;
1829 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
);
1833 perf_output_put(&handle
, header
);
1835 if (record_type
& PERF_RECORD_IP
)
1836 perf_output_put(&handle
, ip
);
1838 if (record_type
& PERF_RECORD_TID
)
1839 perf_output_put(&handle
, tid_entry
);
1841 if (record_type
& PERF_RECORD_GROUP
) {
1842 struct perf_counter
*leader
, *sub
;
1843 u64 nr
= counter
->nr_siblings
;
1845 perf_output_put(&handle
, nr
);
1847 leader
= counter
->group_leader
;
1848 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
1850 sub
->hw_ops
->read(sub
);
1852 group_entry
.event
= sub
->hw_event
.config
;
1853 group_entry
.counter
= atomic64_read(&sub
->count
);
1855 perf_output_put(&handle
, group_entry
);
1860 perf_output_copy(&handle
, callchain
, callchain_size
);
1862 perf_output_end(&handle
);
1869 struct perf_mmap_event
{
1875 struct perf_event_header header
;
1885 static void perf_counter_mmap_output(struct perf_counter
*counter
,
1886 struct perf_mmap_event
*mmap_event
)
1888 struct perf_output_handle handle
;
1889 int size
= mmap_event
->event
.header
.size
;
1890 int ret
= perf_output_begin(&handle
, counter
, size
, 0);
1895 perf_output_put(&handle
, mmap_event
->event
);
1896 perf_output_copy(&handle
, mmap_event
->file_name
,
1897 mmap_event
->file_size
);
1898 perf_output_end(&handle
);
1901 static int perf_counter_mmap_match(struct perf_counter
*counter
,
1902 struct perf_mmap_event
*mmap_event
)
1904 if (counter
->hw_event
.mmap
&&
1905 mmap_event
->event
.header
.type
== PERF_EVENT_MMAP
)
1908 if (counter
->hw_event
.munmap
&&
1909 mmap_event
->event
.header
.type
== PERF_EVENT_MUNMAP
)
1915 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
1916 struct perf_mmap_event
*mmap_event
)
1918 struct perf_counter
*counter
;
1920 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
1924 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
1925 if (perf_counter_mmap_match(counter
, mmap_event
))
1926 perf_counter_mmap_output(counter
, mmap_event
);
1931 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
1933 struct perf_cpu_context
*cpuctx
;
1934 struct file
*file
= mmap_event
->file
;
1941 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
1943 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
1946 name
= dentry_path(file
->f_dentry
, buf
, PATH_MAX
);
1948 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
1952 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
1957 size
= ALIGN(strlen(name
), sizeof(u64
));
1959 mmap_event
->file_name
= name
;
1960 mmap_event
->file_size
= size
;
1962 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
1964 cpuctx
= &get_cpu_var(perf_cpu_context
);
1965 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
1966 put_cpu_var(perf_cpu_context
);
1968 perf_counter_mmap_ctx(¤t
->perf_counter_ctx
, mmap_event
);
1973 void perf_counter_mmap(unsigned long addr
, unsigned long len
,
1974 unsigned long pgoff
, struct file
*file
)
1976 struct perf_mmap_event mmap_event
= {
1979 .header
= { .type
= PERF_EVENT_MMAP
, },
1980 .pid
= current
->group_leader
->pid
,
1981 .tid
= current
->pid
,
1988 perf_counter_mmap_event(&mmap_event
);
1991 void perf_counter_munmap(unsigned long addr
, unsigned long len
,
1992 unsigned long pgoff
, struct file
*file
)
1994 struct perf_mmap_event mmap_event
= {
1997 .header
= { .type
= PERF_EVENT_MUNMAP
, },
1998 .pid
= current
->group_leader
->pid
,
1999 .tid
= current
->pid
,
2006 perf_counter_mmap_event(&mmap_event
);
2010 * Generic software counter infrastructure
2013 static void perf_swcounter_update(struct perf_counter
*counter
)
2015 struct hw_perf_counter
*hwc
= &counter
->hw
;
2020 prev
= atomic64_read(&hwc
->prev_count
);
2021 now
= atomic64_read(&hwc
->count
);
2022 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2027 atomic64_add(delta
, &counter
->count
);
2028 atomic64_sub(delta
, &hwc
->period_left
);
2031 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2033 struct hw_perf_counter
*hwc
= &counter
->hw
;
2034 s64 left
= atomic64_read(&hwc
->period_left
);
2035 s64 period
= hwc
->irq_period
;
2037 if (unlikely(left
<= -period
)) {
2039 atomic64_set(&hwc
->period_left
, left
);
2042 if (unlikely(left
<= 0)) {
2044 atomic64_add(period
, &hwc
->period_left
);
2047 atomic64_set(&hwc
->prev_count
, -left
);
2048 atomic64_set(&hwc
->count
, -left
);
2051 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2053 struct perf_counter
*counter
;
2054 struct pt_regs
*regs
;
2056 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
2057 counter
->hw_ops
->read(counter
);
2059 regs
= get_irq_regs();
2061 * In case we exclude kernel IPs or are somehow not in interrupt
2062 * context, provide the next best thing, the user IP.
2064 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
2065 !counter
->hw_event
.exclude_user
)
2066 regs
= task_pt_regs(current
);
2069 perf_counter_output(counter
, 0, regs
);
2071 hrtimer_forward_now(hrtimer
, ns_to_ktime(counter
->hw
.irq_period
));
2073 return HRTIMER_RESTART
;
2076 static void perf_swcounter_overflow(struct perf_counter
*counter
,
2077 int nmi
, struct pt_regs
*regs
)
2079 perf_swcounter_update(counter
);
2080 perf_swcounter_set_period(counter
);
2081 perf_counter_output(counter
, nmi
, regs
);
2084 static int perf_swcounter_match(struct perf_counter
*counter
,
2085 enum perf_event_types type
,
2086 u32 event
, struct pt_regs
*regs
)
2088 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
2091 if (perf_event_raw(&counter
->hw_event
))
2094 if (perf_event_type(&counter
->hw_event
) != type
)
2097 if (perf_event_id(&counter
->hw_event
) != event
)
2100 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
2103 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
2109 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
2110 int nmi
, struct pt_regs
*regs
)
2112 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
2113 if (counter
->hw
.irq_period
&& !neg
)
2114 perf_swcounter_overflow(counter
, nmi
, regs
);
2117 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
2118 enum perf_event_types type
, u32 event
,
2119 u64 nr
, int nmi
, struct pt_regs
*regs
)
2121 struct perf_counter
*counter
;
2123 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2127 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2128 if (perf_swcounter_match(counter
, type
, event
, regs
))
2129 perf_swcounter_add(counter
, nr
, nmi
, regs
);
2134 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
2137 return &cpuctx
->recursion
[3];
2140 return &cpuctx
->recursion
[2];
2143 return &cpuctx
->recursion
[1];
2145 return &cpuctx
->recursion
[0];
2148 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
2149 u64 nr
, int nmi
, struct pt_regs
*regs
)
2151 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
2152 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
2160 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
, nr
, nmi
, regs
);
2161 if (cpuctx
->task_ctx
) {
2162 perf_swcounter_ctx_event(cpuctx
->task_ctx
, type
, event
,
2170 put_cpu_var(perf_cpu_context
);
2173 void perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
)
2175 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
);
2178 static void perf_swcounter_read(struct perf_counter
*counter
)
2180 perf_swcounter_update(counter
);
2183 static int perf_swcounter_enable(struct perf_counter
*counter
)
2185 perf_swcounter_set_period(counter
);
2189 static void perf_swcounter_disable(struct perf_counter
*counter
)
2191 perf_swcounter_update(counter
);
2194 static const struct hw_perf_counter_ops perf_ops_generic
= {
2195 .enable
= perf_swcounter_enable
,
2196 .disable
= perf_swcounter_disable
,
2197 .read
= perf_swcounter_read
,
2201 * Software counter: cpu wall time clock
2204 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
2206 int cpu
= raw_smp_processor_id();
2210 now
= cpu_clock(cpu
);
2211 prev
= atomic64_read(&counter
->hw
.prev_count
);
2212 atomic64_set(&counter
->hw
.prev_count
, now
);
2213 atomic64_add(now
- prev
, &counter
->count
);
2216 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
2218 struct hw_perf_counter
*hwc
= &counter
->hw
;
2219 int cpu
= raw_smp_processor_id();
2221 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
2222 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2223 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2224 if (hwc
->irq_period
) {
2225 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2226 ns_to_ktime(hwc
->irq_period
), 0,
2227 HRTIMER_MODE_REL
, 0);
2233 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
2235 hrtimer_cancel(&counter
->hw
.hrtimer
);
2236 cpu_clock_perf_counter_update(counter
);
2239 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
2241 cpu_clock_perf_counter_update(counter
);
2244 static const struct hw_perf_counter_ops perf_ops_cpu_clock
= {
2245 .enable
= cpu_clock_perf_counter_enable
,
2246 .disable
= cpu_clock_perf_counter_disable
,
2247 .read
= cpu_clock_perf_counter_read
,
2251 * Software counter: task time clock
2255 * Called from within the scheduler:
2257 static u64
task_clock_perf_counter_val(struct perf_counter
*counter
, int update
)
2259 struct task_struct
*curr
= counter
->task
;
2262 delta
= __task_delta_exec(curr
, update
);
2264 return curr
->se
.sum_exec_runtime
+ delta
;
2267 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
2272 prev
= atomic64_read(&counter
->hw
.prev_count
);
2274 atomic64_set(&counter
->hw
.prev_count
, now
);
2278 atomic64_add(delta
, &counter
->count
);
2281 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
2283 struct hw_perf_counter
*hwc
= &counter
->hw
;
2285 atomic64_set(&hwc
->prev_count
, task_clock_perf_counter_val(counter
, 0));
2286 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2287 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2288 if (hwc
->irq_period
) {
2289 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2290 ns_to_ktime(hwc
->irq_period
), 0,
2291 HRTIMER_MODE_REL
, 0);
2297 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
2299 hrtimer_cancel(&counter
->hw
.hrtimer
);
2300 task_clock_perf_counter_update(counter
,
2301 task_clock_perf_counter_val(counter
, 0));
2304 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
2306 task_clock_perf_counter_update(counter
,
2307 task_clock_perf_counter_val(counter
, 1));
2310 static const struct hw_perf_counter_ops perf_ops_task_clock
= {
2311 .enable
= task_clock_perf_counter_enable
,
2312 .disable
= task_clock_perf_counter_disable
,
2313 .read
= task_clock_perf_counter_read
,
2317 * Software counter: cpu migrations
2320 static inline u64
get_cpu_migrations(struct perf_counter
*counter
)
2322 struct task_struct
*curr
= counter
->ctx
->task
;
2325 return curr
->se
.nr_migrations
;
2326 return cpu_nr_migrations(smp_processor_id());
2329 static void cpu_migrations_perf_counter_update(struct perf_counter
*counter
)
2334 prev
= atomic64_read(&counter
->hw
.prev_count
);
2335 now
= get_cpu_migrations(counter
);
2337 atomic64_set(&counter
->hw
.prev_count
, now
);
2341 atomic64_add(delta
, &counter
->count
);
2344 static void cpu_migrations_perf_counter_read(struct perf_counter
*counter
)
2346 cpu_migrations_perf_counter_update(counter
);
2349 static int cpu_migrations_perf_counter_enable(struct perf_counter
*counter
)
2351 if (counter
->prev_state
<= PERF_COUNTER_STATE_OFF
)
2352 atomic64_set(&counter
->hw
.prev_count
,
2353 get_cpu_migrations(counter
));
2357 static void cpu_migrations_perf_counter_disable(struct perf_counter
*counter
)
2359 cpu_migrations_perf_counter_update(counter
);
2362 static const struct hw_perf_counter_ops perf_ops_cpu_migrations
= {
2363 .enable
= cpu_migrations_perf_counter_enable
,
2364 .disable
= cpu_migrations_perf_counter_disable
,
2365 .read
= cpu_migrations_perf_counter_read
,
2368 #ifdef CONFIG_EVENT_PROFILE
2369 void perf_tpcounter_event(int event_id
)
2371 struct pt_regs
*regs
= get_irq_regs();
2374 regs
= task_pt_regs(current
);
2376 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
);
2379 extern int ftrace_profile_enable(int);
2380 extern void ftrace_profile_disable(int);
2382 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
2384 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
2387 static const struct hw_perf_counter_ops
*
2388 tp_perf_counter_init(struct perf_counter
*counter
)
2390 int event_id
= perf_event_id(&counter
->hw_event
);
2393 ret
= ftrace_profile_enable(event_id
);
2397 counter
->destroy
= tp_perf_counter_destroy
;
2398 counter
->hw
.irq_period
= counter
->hw_event
.irq_period
;
2400 return &perf_ops_generic
;
2403 static const struct hw_perf_counter_ops
*
2404 tp_perf_counter_init(struct perf_counter
*counter
)
2410 static const struct hw_perf_counter_ops
*
2411 sw_perf_counter_init(struct perf_counter
*counter
)
2413 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
2414 const struct hw_perf_counter_ops
*hw_ops
= NULL
;
2415 struct hw_perf_counter
*hwc
= &counter
->hw
;
2418 * Software counters (currently) can't in general distinguish
2419 * between user, kernel and hypervisor events.
2420 * However, context switches and cpu migrations are considered
2421 * to be kernel events, and page faults are never hypervisor
2424 switch (perf_event_id(&counter
->hw_event
)) {
2425 case PERF_COUNT_CPU_CLOCK
:
2426 hw_ops
= &perf_ops_cpu_clock
;
2428 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2429 hw_event
->irq_period
= 10000;
2431 case PERF_COUNT_TASK_CLOCK
:
2433 * If the user instantiates this as a per-cpu counter,
2434 * use the cpu_clock counter instead.
2436 if (counter
->ctx
->task
)
2437 hw_ops
= &perf_ops_task_clock
;
2439 hw_ops
= &perf_ops_cpu_clock
;
2441 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2442 hw_event
->irq_period
= 10000;
2444 case PERF_COUNT_PAGE_FAULTS
:
2445 case PERF_COUNT_PAGE_FAULTS_MIN
:
2446 case PERF_COUNT_PAGE_FAULTS_MAJ
:
2447 case PERF_COUNT_CONTEXT_SWITCHES
:
2448 hw_ops
= &perf_ops_generic
;
2450 case PERF_COUNT_CPU_MIGRATIONS
:
2451 if (!counter
->hw_event
.exclude_kernel
)
2452 hw_ops
= &perf_ops_cpu_migrations
;
2457 hwc
->irq_period
= hw_event
->irq_period
;
2463 * Allocate and initialize a counter structure
2465 static struct perf_counter
*
2466 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
2468 struct perf_counter_context
*ctx
,
2469 struct perf_counter
*group_leader
,
2472 const struct hw_perf_counter_ops
*hw_ops
;
2473 struct perf_counter
*counter
;
2476 counter
= kzalloc(sizeof(*counter
), gfpflags
);
2478 return ERR_PTR(-ENOMEM
);
2481 * Single counters are their own group leaders, with an
2482 * empty sibling list:
2485 group_leader
= counter
;
2487 mutex_init(&counter
->mutex
);
2488 INIT_LIST_HEAD(&counter
->list_entry
);
2489 INIT_LIST_HEAD(&counter
->event_entry
);
2490 INIT_LIST_HEAD(&counter
->sibling_list
);
2491 init_waitqueue_head(&counter
->waitq
);
2493 mutex_init(&counter
->mmap_mutex
);
2495 INIT_LIST_HEAD(&counter
->child_list
);
2498 counter
->hw_event
= *hw_event
;
2499 counter
->group_leader
= group_leader
;
2500 counter
->hw_ops
= NULL
;
2503 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2504 if (hw_event
->disabled
)
2505 counter
->state
= PERF_COUNTER_STATE_OFF
;
2509 if (perf_event_raw(hw_event
)) {
2510 hw_ops
= hw_perf_counter_init(counter
);
2514 switch (perf_event_type(hw_event
)) {
2515 case PERF_TYPE_HARDWARE
:
2516 hw_ops
= hw_perf_counter_init(counter
);
2519 case PERF_TYPE_SOFTWARE
:
2520 hw_ops
= sw_perf_counter_init(counter
);
2523 case PERF_TYPE_TRACEPOINT
:
2524 hw_ops
= tp_perf_counter_init(counter
);
2531 else if (IS_ERR(hw_ops
))
2532 err
= PTR_ERR(hw_ops
);
2536 return ERR_PTR(err
);
2539 counter
->hw_ops
= hw_ops
;
2545 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2547 * @hw_event_uptr: event type attributes for monitoring/sampling
2550 * @group_fd: group leader counter fd
2552 SYSCALL_DEFINE5(perf_counter_open
,
2553 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
2554 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
2556 struct perf_counter
*counter
, *group_leader
;
2557 struct perf_counter_hw_event hw_event
;
2558 struct perf_counter_context
*ctx
;
2559 struct file
*counter_file
= NULL
;
2560 struct file
*group_file
= NULL
;
2561 int fput_needed
= 0;
2562 int fput_needed2
= 0;
2565 /* for future expandability... */
2569 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
2573 * Get the target context (task or percpu):
2575 ctx
= find_get_context(pid
, cpu
);
2577 return PTR_ERR(ctx
);
2580 * Look up the group leader (we will attach this counter to it):
2582 group_leader
= NULL
;
2583 if (group_fd
!= -1) {
2585 group_file
= fget_light(group_fd
, &fput_needed
);
2587 goto err_put_context
;
2588 if (group_file
->f_op
!= &perf_fops
)
2589 goto err_put_context
;
2591 group_leader
= group_file
->private_data
;
2593 * Do not allow a recursive hierarchy (this new sibling
2594 * becoming part of another group-sibling):
2596 if (group_leader
->group_leader
!= group_leader
)
2597 goto err_put_context
;
2599 * Do not allow to attach to a group in a different
2600 * task or CPU context:
2602 if (group_leader
->ctx
!= ctx
)
2603 goto err_put_context
;
2605 * Only a group leader can be exclusive or pinned
2607 if (hw_event
.exclusive
|| hw_event
.pinned
)
2608 goto err_put_context
;
2611 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
2613 ret
= PTR_ERR(counter
);
2614 if (IS_ERR(counter
))
2615 goto err_put_context
;
2617 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
2619 goto err_free_put_context
;
2621 counter_file
= fget_light(ret
, &fput_needed2
);
2623 goto err_free_put_context
;
2625 counter
->filp
= counter_file
;
2626 mutex_lock(&ctx
->mutex
);
2627 perf_install_in_context(ctx
, counter
, cpu
);
2628 mutex_unlock(&ctx
->mutex
);
2630 fput_light(counter_file
, fput_needed2
);
2633 fput_light(group_file
, fput_needed
);
2637 err_free_put_context
:
2647 * Initialize the perf_counter context in a task_struct:
2650 __perf_counter_init_context(struct perf_counter_context
*ctx
,
2651 struct task_struct
*task
)
2653 memset(ctx
, 0, sizeof(*ctx
));
2654 spin_lock_init(&ctx
->lock
);
2655 mutex_init(&ctx
->mutex
);
2656 INIT_LIST_HEAD(&ctx
->counter_list
);
2657 INIT_LIST_HEAD(&ctx
->event_list
);
2662 * inherit a counter from parent task to child task:
2664 static struct perf_counter
*
2665 inherit_counter(struct perf_counter
*parent_counter
,
2666 struct task_struct
*parent
,
2667 struct perf_counter_context
*parent_ctx
,
2668 struct task_struct
*child
,
2669 struct perf_counter
*group_leader
,
2670 struct perf_counter_context
*child_ctx
)
2672 struct perf_counter
*child_counter
;
2675 * Instead of creating recursive hierarchies of counters,
2676 * we link inherited counters back to the original parent,
2677 * which has a filp for sure, which we use as the reference
2680 if (parent_counter
->parent
)
2681 parent_counter
= parent_counter
->parent
;
2683 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
2684 parent_counter
->cpu
, child_ctx
,
2685 group_leader
, GFP_KERNEL
);
2686 if (IS_ERR(child_counter
))
2687 return child_counter
;
2690 * Link it up in the child's context:
2692 child_counter
->task
= child
;
2693 add_counter_to_ctx(child_counter
, child_ctx
);
2695 child_counter
->parent
= parent_counter
;
2697 * inherit into child's child as well:
2699 child_counter
->hw_event
.inherit
= 1;
2702 * Get a reference to the parent filp - we will fput it
2703 * when the child counter exits. This is safe to do because
2704 * we are in the parent and we know that the filp still
2705 * exists and has a nonzero count:
2707 atomic_long_inc(&parent_counter
->filp
->f_count
);
2710 * Link this into the parent counter's child list
2712 mutex_lock(&parent_counter
->mutex
);
2713 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
2716 * Make the child state follow the state of the parent counter,
2717 * not its hw_event.disabled bit. We hold the parent's mutex,
2718 * so we won't race with perf_counter_{en,dis}able_family.
2720 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
2721 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2723 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
2725 mutex_unlock(&parent_counter
->mutex
);
2727 return child_counter
;
2730 static int inherit_group(struct perf_counter
*parent_counter
,
2731 struct task_struct
*parent
,
2732 struct perf_counter_context
*parent_ctx
,
2733 struct task_struct
*child
,
2734 struct perf_counter_context
*child_ctx
)
2736 struct perf_counter
*leader
;
2737 struct perf_counter
*sub
;
2738 struct perf_counter
*child_ctr
;
2740 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
2741 child
, NULL
, child_ctx
);
2743 return PTR_ERR(leader
);
2744 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
2745 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
2746 child
, leader
, child_ctx
);
2747 if (IS_ERR(child_ctr
))
2748 return PTR_ERR(child_ctr
);
2753 static void sync_child_counter(struct perf_counter
*child_counter
,
2754 struct perf_counter
*parent_counter
)
2756 u64 parent_val
, child_val
;
2758 parent_val
= atomic64_read(&parent_counter
->count
);
2759 child_val
= atomic64_read(&child_counter
->count
);
2762 * Add back the child's count to the parent's count:
2764 atomic64_add(child_val
, &parent_counter
->count
);
2765 atomic64_add(child_counter
->total_time_enabled
,
2766 &parent_counter
->child_total_time_enabled
);
2767 atomic64_add(child_counter
->total_time_running
,
2768 &parent_counter
->child_total_time_running
);
2771 * Remove this counter from the parent's list
2773 mutex_lock(&parent_counter
->mutex
);
2774 list_del_init(&child_counter
->child_list
);
2775 mutex_unlock(&parent_counter
->mutex
);
2778 * Release the parent counter, if this was the last
2781 fput(parent_counter
->filp
);
2785 __perf_counter_exit_task(struct task_struct
*child
,
2786 struct perf_counter
*child_counter
,
2787 struct perf_counter_context
*child_ctx
)
2789 struct perf_counter
*parent_counter
;
2790 struct perf_counter
*sub
, *tmp
;
2793 * If we do not self-reap then we have to wait for the
2794 * child task to unschedule (it will happen for sure),
2795 * so that its counter is at its final count. (This
2796 * condition triggers rarely - child tasks usually get
2797 * off their CPU before the parent has a chance to
2798 * get this far into the reaping action)
2800 if (child
!= current
) {
2801 wait_task_inactive(child
, 0);
2802 list_del_init(&child_counter
->list_entry
);
2803 update_counter_times(child_counter
);
2805 struct perf_cpu_context
*cpuctx
;
2806 unsigned long flags
;
2810 * Disable and unlink this counter.
2812 * Be careful about zapping the list - IRQ/NMI context
2813 * could still be processing it:
2815 curr_rq_lock_irq_save(&flags
);
2816 perf_flags
= hw_perf_save_disable();
2818 cpuctx
= &__get_cpu_var(perf_cpu_context
);
2820 group_sched_out(child_counter
, cpuctx
, child_ctx
);
2821 update_counter_times(child_counter
);
2823 list_del_init(&child_counter
->list_entry
);
2825 child_ctx
->nr_counters
--;
2827 hw_perf_restore(perf_flags
);
2828 curr_rq_unlock_irq_restore(&flags
);
2831 parent_counter
= child_counter
->parent
;
2833 * It can happen that parent exits first, and has counters
2834 * that are still around due to the child reference. These
2835 * counters need to be zapped - but otherwise linger.
2837 if (parent_counter
) {
2838 sync_child_counter(child_counter
, parent_counter
);
2839 list_for_each_entry_safe(sub
, tmp
, &child_counter
->sibling_list
,
2842 sync_child_counter(sub
, sub
->parent
);
2846 free_counter(child_counter
);
2851 * When a child task exits, feed back counter values to parent counters.
2853 * Note: we may be running in child context, but the PID is not hashed
2854 * anymore so new counters will not be added.
2856 void perf_counter_exit_task(struct task_struct
*child
)
2858 struct perf_counter
*child_counter
, *tmp
;
2859 struct perf_counter_context
*child_ctx
;
2861 child_ctx
= &child
->perf_counter_ctx
;
2863 if (likely(!child_ctx
->nr_counters
))
2866 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
2868 __perf_counter_exit_task(child
, child_counter
, child_ctx
);
2872 * Initialize the perf_counter context in task_struct
2874 void perf_counter_init_task(struct task_struct
*child
)
2876 struct perf_counter_context
*child_ctx
, *parent_ctx
;
2877 struct perf_counter
*counter
;
2878 struct task_struct
*parent
= current
;
2880 child_ctx
= &child
->perf_counter_ctx
;
2881 parent_ctx
= &parent
->perf_counter_ctx
;
2883 __perf_counter_init_context(child_ctx
, child
);
2886 * This is executed from the parent task context, so inherit
2887 * counters that have been marked for cloning:
2890 if (likely(!parent_ctx
->nr_counters
))
2894 * Lock the parent list. No need to lock the child - not PID
2895 * hashed yet and not running, so nobody can access it.
2897 mutex_lock(&parent_ctx
->mutex
);
2900 * We dont have to disable NMIs - we are only looking at
2901 * the list, not manipulating it:
2903 list_for_each_entry(counter
, &parent_ctx
->counter_list
, list_entry
) {
2904 if (!counter
->hw_event
.inherit
)
2907 if (inherit_group(counter
, parent
,
2908 parent_ctx
, child
, child_ctx
))
2912 mutex_unlock(&parent_ctx
->mutex
);
2915 static void __cpuinit
perf_counter_init_cpu(int cpu
)
2917 struct perf_cpu_context
*cpuctx
;
2919 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2920 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
2922 mutex_lock(&perf_resource_mutex
);
2923 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
2924 mutex_unlock(&perf_resource_mutex
);
2926 hw_perf_counter_setup(cpu
);
2929 #ifdef CONFIG_HOTPLUG_CPU
2930 static void __perf_counter_exit_cpu(void *info
)
2932 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
2933 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
2934 struct perf_counter
*counter
, *tmp
;
2936 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
2937 __perf_counter_remove_from_context(counter
);
2939 static void perf_counter_exit_cpu(int cpu
)
2941 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2942 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
2944 mutex_lock(&ctx
->mutex
);
2945 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
2946 mutex_unlock(&ctx
->mutex
);
2949 static inline void perf_counter_exit_cpu(int cpu
) { }
2952 static int __cpuinit
2953 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
2955 unsigned int cpu
= (long)hcpu
;
2959 case CPU_UP_PREPARE
:
2960 case CPU_UP_PREPARE_FROZEN
:
2961 perf_counter_init_cpu(cpu
);
2964 case CPU_DOWN_PREPARE
:
2965 case CPU_DOWN_PREPARE_FROZEN
:
2966 perf_counter_exit_cpu(cpu
);
2976 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
2977 .notifier_call
= perf_cpu_notify
,
2980 static int __init
perf_counter_init(void)
2982 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
2983 (void *)(long)smp_processor_id());
2984 register_cpu_notifier(&perf_cpu_nb
);
2988 early_initcall(perf_counter_init
);
2990 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
2992 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
2996 perf_set_reserve_percpu(struct sysdev_class
*class,
3000 struct perf_cpu_context
*cpuctx
;
3004 err
= strict_strtoul(buf
, 10, &val
);
3007 if (val
> perf_max_counters
)
3010 mutex_lock(&perf_resource_mutex
);
3011 perf_reserved_percpu
= val
;
3012 for_each_online_cpu(cpu
) {
3013 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3014 spin_lock_irq(&cpuctx
->ctx
.lock
);
3015 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
3016 perf_max_counters
- perf_reserved_percpu
);
3017 cpuctx
->max_pertask
= mpt
;
3018 spin_unlock_irq(&cpuctx
->ctx
.lock
);
3020 mutex_unlock(&perf_resource_mutex
);
3025 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
3027 return sprintf(buf
, "%d\n", perf_overcommit
);
3031 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
3036 err
= strict_strtoul(buf
, 10, &val
);
3042 mutex_lock(&perf_resource_mutex
);
3043 perf_overcommit
= val
;
3044 mutex_unlock(&perf_resource_mutex
);
3049 static SYSDEV_CLASS_ATTR(
3052 perf_show_reserve_percpu
,
3053 perf_set_reserve_percpu
3056 static SYSDEV_CLASS_ATTR(
3059 perf_show_overcommit
,
3063 static struct attribute
*perfclass_attrs
[] = {
3064 &attr_reserve_percpu
.attr
,
3065 &attr_overcommit
.attr
,
3069 static struct attribute_group perfclass_attr_group
= {
3070 .attrs
= perfclass_attrs
,
3071 .name
= "perf_counters",
3074 static int __init
perf_counter_sysfs_init(void)
3076 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
3077 &perfclass_attr_group
);
3079 device_initcall(perf_counter_sysfs_init
);