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 int perf_fasync(int fd
, struct file
*filp
, int on
)
1531 struct perf_counter
*counter
= filp
->private_data
;
1532 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1535 mutex_lock(&inode
->i_mutex
);
1536 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
1537 mutex_unlock(&inode
->i_mutex
);
1545 static const struct file_operations perf_fops
= {
1546 .release
= perf_release
,
1549 .unlocked_ioctl
= perf_ioctl
,
1550 .compat_ioctl
= perf_ioctl
,
1552 .fasync
= perf_fasync
,
1556 * Perf counter wakeup
1558 * If there's data, ensure we set the poll() state and publish everything
1559 * to user-space before waking everybody up.
1562 void perf_counter_wakeup(struct perf_counter
*counter
)
1564 struct perf_mmap_data
*data
;
1567 data
= rcu_dereference(counter
->data
);
1569 atomic_set(&data
->wakeup
, POLL_IN
);
1571 * Ensure all data writes are issued before updating the
1572 * user-space data head information. The matching rmb()
1573 * will be in userspace after reading this value.
1576 data
->user_page
->data_head
= atomic_read(&data
->head
);
1580 wake_up_all(&counter
->waitq
);
1581 kill_fasync(&counter
->fasync
, SIGIO
, POLL_IN
);
1584 static void perf_pending_wakeup(struct perf_pending_entry
*entry
)
1586 struct perf_counter
*counter
= container_of(entry
,
1587 struct perf_counter
, pending
);
1589 perf_counter_wakeup(counter
);
1595 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1597 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1598 * single linked list and use cmpxchg() to add entries lockless.
1601 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1603 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
1607 static void perf_pending_queue(struct perf_pending_entry
*entry
,
1608 void (*func
)(struct perf_pending_entry
*))
1610 struct perf_pending_entry
**head
;
1612 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
1617 head
= &get_cpu_var(perf_pending_head
);
1620 entry
->next
= *head
;
1621 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
1623 set_perf_counter_pending();
1625 put_cpu_var(perf_pending_head
);
1628 static int __perf_pending_run(void)
1630 struct perf_pending_entry
*list
;
1633 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
1634 while (list
!= PENDING_TAIL
) {
1635 void (*func
)(struct perf_pending_entry
*);
1636 struct perf_pending_entry
*entry
= list
;
1643 * Ensure we observe the unqueue before we issue the wakeup,
1644 * so that we won't be waiting forever.
1645 * -- see perf_not_pending().
1656 static inline int perf_not_pending(struct perf_counter
*counter
)
1659 * If we flush on whatever cpu we run, there is a chance we don't
1663 __perf_pending_run();
1667 * Ensure we see the proper queue state before going to sleep
1668 * so that we do not miss the wakeup. -- see perf_pending_handle()
1671 return counter
->pending
.next
== NULL
;
1674 static void perf_pending_sync(struct perf_counter
*counter
)
1676 wait_event(counter
->waitq
, perf_not_pending(counter
));
1679 void perf_counter_do_pending(void)
1681 __perf_pending_run();
1685 * Callchain support -- arch specific
1688 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1697 struct perf_output_handle
{
1698 struct perf_counter
*counter
;
1699 struct perf_mmap_data
*data
;
1700 unsigned int offset
;
1706 static inline void __perf_output_wakeup(struct perf_output_handle
*handle
)
1709 perf_pending_queue(&handle
->counter
->pending
,
1710 perf_pending_wakeup
);
1712 perf_counter_wakeup(handle
->counter
);
1715 static int perf_output_begin(struct perf_output_handle
*handle
,
1716 struct perf_counter
*counter
, unsigned int size
,
1719 struct perf_mmap_data
*data
;
1720 unsigned int offset
, head
;
1723 data
= rcu_dereference(counter
->data
);
1727 handle
->counter
= counter
;
1730 if (!data
->nr_pages
)
1734 offset
= head
= atomic_read(&data
->head
);
1736 } while (atomic_cmpxchg(&data
->head
, offset
, head
) != offset
);
1738 handle
->data
= data
;
1739 handle
->offset
= offset
;
1740 handle
->head
= head
;
1741 handle
->wakeup
= (offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
);
1746 __perf_output_wakeup(handle
);
1753 static void perf_output_copy(struct perf_output_handle
*handle
,
1754 void *buf
, unsigned int len
)
1756 unsigned int pages_mask
;
1757 unsigned int offset
;
1761 offset
= handle
->offset
;
1762 pages_mask
= handle
->data
->nr_pages
- 1;
1763 pages
= handle
->data
->data_pages
;
1766 unsigned int page_offset
;
1769 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
1770 page_offset
= offset
& (PAGE_SIZE
- 1);
1771 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
1773 memcpy(pages
[nr
] + page_offset
, buf
, size
);
1780 handle
->offset
= offset
;
1782 WARN_ON_ONCE(handle
->offset
> handle
->head
);
1785 #define perf_output_put(handle, x) \
1786 perf_output_copy((handle), &(x), sizeof(x))
1788 static void perf_output_end(struct perf_output_handle
*handle
)
1790 int wakeup_events
= handle
->counter
->hw_event
.wakeup_events
;
1792 if (wakeup_events
) {
1793 int events
= atomic_inc_return(&handle
->data
->events
);
1794 if (events
>= wakeup_events
) {
1795 atomic_sub(wakeup_events
, &handle
->data
->events
);
1796 __perf_output_wakeup(handle
);
1798 } else if (handle
->wakeup
)
1799 __perf_output_wakeup(handle
);
1803 static void perf_counter_output(struct perf_counter
*counter
,
1804 int nmi
, struct pt_regs
*regs
)
1807 u64 record_type
= counter
->hw_event
.record_type
;
1808 struct perf_output_handle handle
;
1809 struct perf_event_header header
;
1818 struct perf_callchain_entry
*callchain
= NULL
;
1819 int callchain_size
= 0;
1821 header
.type
= PERF_EVENT_COUNTER_OVERFLOW
;
1822 header
.size
= sizeof(header
);
1824 if (record_type
& PERF_RECORD_IP
) {
1825 ip
= instruction_pointer(regs
);
1826 header
.type
|= __PERF_EVENT_IP
;
1827 header
.size
+= sizeof(ip
);
1830 if (record_type
& PERF_RECORD_TID
) {
1831 /* namespace issues */
1832 tid_entry
.pid
= current
->group_leader
->pid
;
1833 tid_entry
.tid
= current
->pid
;
1835 header
.type
|= __PERF_EVENT_TID
;
1836 header
.size
+= sizeof(tid_entry
);
1839 if (record_type
& PERF_RECORD_GROUP
) {
1840 header
.type
|= __PERF_EVENT_GROUP
;
1841 header
.size
+= sizeof(u64
) +
1842 counter
->nr_siblings
* sizeof(group_entry
);
1845 if (record_type
& PERF_RECORD_CALLCHAIN
) {
1846 callchain
= perf_callchain(regs
);
1849 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
1851 header
.type
|= __PERF_EVENT_CALLCHAIN
;
1852 header
.size
+= callchain_size
;
1856 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
);
1860 perf_output_put(&handle
, header
);
1862 if (record_type
& PERF_RECORD_IP
)
1863 perf_output_put(&handle
, ip
);
1865 if (record_type
& PERF_RECORD_TID
)
1866 perf_output_put(&handle
, tid_entry
);
1868 if (record_type
& PERF_RECORD_GROUP
) {
1869 struct perf_counter
*leader
, *sub
;
1870 u64 nr
= counter
->nr_siblings
;
1872 perf_output_put(&handle
, nr
);
1874 leader
= counter
->group_leader
;
1875 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
1877 sub
->hw_ops
->read(sub
);
1879 group_entry
.event
= sub
->hw_event
.config
;
1880 group_entry
.counter
= atomic64_read(&sub
->count
);
1882 perf_output_put(&handle
, group_entry
);
1887 perf_output_copy(&handle
, callchain
, callchain_size
);
1889 perf_output_end(&handle
);
1896 struct perf_mmap_event
{
1902 struct perf_event_header header
;
1912 static void perf_counter_mmap_output(struct perf_counter
*counter
,
1913 struct perf_mmap_event
*mmap_event
)
1915 struct perf_output_handle handle
;
1916 int size
= mmap_event
->event
.header
.size
;
1917 int ret
= perf_output_begin(&handle
, counter
, size
, 0);
1922 perf_output_put(&handle
, mmap_event
->event
);
1923 perf_output_copy(&handle
, mmap_event
->file_name
,
1924 mmap_event
->file_size
);
1925 perf_output_end(&handle
);
1928 static int perf_counter_mmap_match(struct perf_counter
*counter
,
1929 struct perf_mmap_event
*mmap_event
)
1931 if (counter
->hw_event
.mmap
&&
1932 mmap_event
->event
.header
.type
== PERF_EVENT_MMAP
)
1935 if (counter
->hw_event
.munmap
&&
1936 mmap_event
->event
.header
.type
== PERF_EVENT_MUNMAP
)
1942 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
1943 struct perf_mmap_event
*mmap_event
)
1945 struct perf_counter
*counter
;
1947 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
1951 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
1952 if (perf_counter_mmap_match(counter
, mmap_event
))
1953 perf_counter_mmap_output(counter
, mmap_event
);
1958 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
1960 struct perf_cpu_context
*cpuctx
;
1961 struct file
*file
= mmap_event
->file
;
1968 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
1970 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
1973 name
= dentry_path(file
->f_dentry
, buf
, PATH_MAX
);
1975 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
1979 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
1984 size
= ALIGN(strlen(name
), sizeof(u64
));
1986 mmap_event
->file_name
= name
;
1987 mmap_event
->file_size
= size
;
1989 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
1991 cpuctx
= &get_cpu_var(perf_cpu_context
);
1992 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
1993 put_cpu_var(perf_cpu_context
);
1995 perf_counter_mmap_ctx(¤t
->perf_counter_ctx
, mmap_event
);
2000 void perf_counter_mmap(unsigned long addr
, unsigned long len
,
2001 unsigned long pgoff
, struct file
*file
)
2003 struct perf_mmap_event mmap_event
= {
2006 .header
= { .type
= PERF_EVENT_MMAP
, },
2007 .pid
= current
->group_leader
->pid
,
2008 .tid
= current
->pid
,
2015 perf_counter_mmap_event(&mmap_event
);
2018 void perf_counter_munmap(unsigned long addr
, unsigned long len
,
2019 unsigned long pgoff
, struct file
*file
)
2021 struct perf_mmap_event mmap_event
= {
2024 .header
= { .type
= PERF_EVENT_MUNMAP
, },
2025 .pid
= current
->group_leader
->pid
,
2026 .tid
= current
->pid
,
2033 perf_counter_mmap_event(&mmap_event
);
2037 * Generic counter overflow handling.
2040 int perf_counter_overflow(struct perf_counter
*counter
,
2041 int nmi
, struct pt_regs
*regs
)
2043 perf_counter_output(counter
, nmi
, regs
);
2048 * Generic software counter infrastructure
2051 static void perf_swcounter_update(struct perf_counter
*counter
)
2053 struct hw_perf_counter
*hwc
= &counter
->hw
;
2058 prev
= atomic64_read(&hwc
->prev_count
);
2059 now
= atomic64_read(&hwc
->count
);
2060 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2065 atomic64_add(delta
, &counter
->count
);
2066 atomic64_sub(delta
, &hwc
->period_left
);
2069 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2071 struct hw_perf_counter
*hwc
= &counter
->hw
;
2072 s64 left
= atomic64_read(&hwc
->period_left
);
2073 s64 period
= hwc
->irq_period
;
2075 if (unlikely(left
<= -period
)) {
2077 atomic64_set(&hwc
->period_left
, left
);
2080 if (unlikely(left
<= 0)) {
2082 atomic64_add(period
, &hwc
->period_left
);
2085 atomic64_set(&hwc
->prev_count
, -left
);
2086 atomic64_set(&hwc
->count
, -left
);
2089 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2091 enum hrtimer_restart ret
= HRTIMER_RESTART
;
2092 struct perf_counter
*counter
;
2093 struct pt_regs
*regs
;
2095 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
2096 counter
->hw_ops
->read(counter
);
2098 regs
= get_irq_regs();
2100 * In case we exclude kernel IPs or are somehow not in interrupt
2101 * context, provide the next best thing, the user IP.
2103 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
2104 !counter
->hw_event
.exclude_user
)
2105 regs
= task_pt_regs(current
);
2108 if (perf_counter_overflow(counter
, 0, regs
))
2109 ret
= HRTIMER_NORESTART
;
2112 hrtimer_forward_now(hrtimer
, ns_to_ktime(counter
->hw
.irq_period
));
2117 static void perf_swcounter_overflow(struct perf_counter
*counter
,
2118 int nmi
, struct pt_regs
*regs
)
2120 perf_swcounter_update(counter
);
2121 perf_swcounter_set_period(counter
);
2122 if (perf_counter_overflow(counter
, nmi
, regs
))
2123 /* soft-disable the counter */
2128 static int perf_swcounter_match(struct perf_counter
*counter
,
2129 enum perf_event_types type
,
2130 u32 event
, struct pt_regs
*regs
)
2132 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
2135 if (perf_event_raw(&counter
->hw_event
))
2138 if (perf_event_type(&counter
->hw_event
) != type
)
2141 if (perf_event_id(&counter
->hw_event
) != event
)
2144 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
2147 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
2153 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
2154 int nmi
, struct pt_regs
*regs
)
2156 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
2157 if (counter
->hw
.irq_period
&& !neg
)
2158 perf_swcounter_overflow(counter
, nmi
, regs
);
2161 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
2162 enum perf_event_types type
, u32 event
,
2163 u64 nr
, int nmi
, struct pt_regs
*regs
)
2165 struct perf_counter
*counter
;
2167 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2171 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2172 if (perf_swcounter_match(counter
, type
, event
, regs
))
2173 perf_swcounter_add(counter
, nr
, nmi
, regs
);
2178 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
2181 return &cpuctx
->recursion
[3];
2184 return &cpuctx
->recursion
[2];
2187 return &cpuctx
->recursion
[1];
2189 return &cpuctx
->recursion
[0];
2192 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
2193 u64 nr
, int nmi
, struct pt_regs
*regs
)
2195 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
2196 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
2204 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
, nr
, nmi
, regs
);
2205 if (cpuctx
->task_ctx
) {
2206 perf_swcounter_ctx_event(cpuctx
->task_ctx
, type
, event
,
2214 put_cpu_var(perf_cpu_context
);
2217 void perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
)
2219 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
);
2222 static void perf_swcounter_read(struct perf_counter
*counter
)
2224 perf_swcounter_update(counter
);
2227 static int perf_swcounter_enable(struct perf_counter
*counter
)
2229 perf_swcounter_set_period(counter
);
2233 static void perf_swcounter_disable(struct perf_counter
*counter
)
2235 perf_swcounter_update(counter
);
2238 static const struct hw_perf_counter_ops perf_ops_generic
= {
2239 .enable
= perf_swcounter_enable
,
2240 .disable
= perf_swcounter_disable
,
2241 .read
= perf_swcounter_read
,
2245 * Software counter: cpu wall time clock
2248 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
2250 int cpu
= raw_smp_processor_id();
2254 now
= cpu_clock(cpu
);
2255 prev
= atomic64_read(&counter
->hw
.prev_count
);
2256 atomic64_set(&counter
->hw
.prev_count
, now
);
2257 atomic64_add(now
- prev
, &counter
->count
);
2260 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
2262 struct hw_perf_counter
*hwc
= &counter
->hw
;
2263 int cpu
= raw_smp_processor_id();
2265 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
2266 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2267 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2268 if (hwc
->irq_period
) {
2269 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2270 ns_to_ktime(hwc
->irq_period
), 0,
2271 HRTIMER_MODE_REL
, 0);
2277 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
2279 hrtimer_cancel(&counter
->hw
.hrtimer
);
2280 cpu_clock_perf_counter_update(counter
);
2283 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
2285 cpu_clock_perf_counter_update(counter
);
2288 static const struct hw_perf_counter_ops perf_ops_cpu_clock
= {
2289 .enable
= cpu_clock_perf_counter_enable
,
2290 .disable
= cpu_clock_perf_counter_disable
,
2291 .read
= cpu_clock_perf_counter_read
,
2295 * Software counter: task time clock
2299 * Called from within the scheduler:
2301 static u64
task_clock_perf_counter_val(struct perf_counter
*counter
, int update
)
2303 struct task_struct
*curr
= counter
->task
;
2306 delta
= __task_delta_exec(curr
, update
);
2308 return curr
->se
.sum_exec_runtime
+ delta
;
2311 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
2316 prev
= atomic64_read(&counter
->hw
.prev_count
);
2318 atomic64_set(&counter
->hw
.prev_count
, now
);
2322 atomic64_add(delta
, &counter
->count
);
2325 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
2327 struct hw_perf_counter
*hwc
= &counter
->hw
;
2329 atomic64_set(&hwc
->prev_count
, task_clock_perf_counter_val(counter
, 0));
2330 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2331 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2332 if (hwc
->irq_period
) {
2333 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2334 ns_to_ktime(hwc
->irq_period
), 0,
2335 HRTIMER_MODE_REL
, 0);
2341 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
2343 hrtimer_cancel(&counter
->hw
.hrtimer
);
2344 task_clock_perf_counter_update(counter
,
2345 task_clock_perf_counter_val(counter
, 0));
2348 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
2350 task_clock_perf_counter_update(counter
,
2351 task_clock_perf_counter_val(counter
, 1));
2354 static const struct hw_perf_counter_ops perf_ops_task_clock
= {
2355 .enable
= task_clock_perf_counter_enable
,
2356 .disable
= task_clock_perf_counter_disable
,
2357 .read
= task_clock_perf_counter_read
,
2361 * Software counter: cpu migrations
2364 static inline u64
get_cpu_migrations(struct perf_counter
*counter
)
2366 struct task_struct
*curr
= counter
->ctx
->task
;
2369 return curr
->se
.nr_migrations
;
2370 return cpu_nr_migrations(smp_processor_id());
2373 static void cpu_migrations_perf_counter_update(struct perf_counter
*counter
)
2378 prev
= atomic64_read(&counter
->hw
.prev_count
);
2379 now
= get_cpu_migrations(counter
);
2381 atomic64_set(&counter
->hw
.prev_count
, now
);
2385 atomic64_add(delta
, &counter
->count
);
2388 static void cpu_migrations_perf_counter_read(struct perf_counter
*counter
)
2390 cpu_migrations_perf_counter_update(counter
);
2393 static int cpu_migrations_perf_counter_enable(struct perf_counter
*counter
)
2395 if (counter
->prev_state
<= PERF_COUNTER_STATE_OFF
)
2396 atomic64_set(&counter
->hw
.prev_count
,
2397 get_cpu_migrations(counter
));
2401 static void cpu_migrations_perf_counter_disable(struct perf_counter
*counter
)
2403 cpu_migrations_perf_counter_update(counter
);
2406 static const struct hw_perf_counter_ops perf_ops_cpu_migrations
= {
2407 .enable
= cpu_migrations_perf_counter_enable
,
2408 .disable
= cpu_migrations_perf_counter_disable
,
2409 .read
= cpu_migrations_perf_counter_read
,
2412 #ifdef CONFIG_EVENT_PROFILE
2413 void perf_tpcounter_event(int event_id
)
2415 struct pt_regs
*regs
= get_irq_regs();
2418 regs
= task_pt_regs(current
);
2420 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
);
2423 extern int ftrace_profile_enable(int);
2424 extern void ftrace_profile_disable(int);
2426 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
2428 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
2431 static const struct hw_perf_counter_ops
*
2432 tp_perf_counter_init(struct perf_counter
*counter
)
2434 int event_id
= perf_event_id(&counter
->hw_event
);
2437 ret
= ftrace_profile_enable(event_id
);
2441 counter
->destroy
= tp_perf_counter_destroy
;
2442 counter
->hw
.irq_period
= counter
->hw_event
.irq_period
;
2444 return &perf_ops_generic
;
2447 static const struct hw_perf_counter_ops
*
2448 tp_perf_counter_init(struct perf_counter
*counter
)
2454 static const struct hw_perf_counter_ops
*
2455 sw_perf_counter_init(struct perf_counter
*counter
)
2457 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
2458 const struct hw_perf_counter_ops
*hw_ops
= NULL
;
2459 struct hw_perf_counter
*hwc
= &counter
->hw
;
2462 * Software counters (currently) can't in general distinguish
2463 * between user, kernel and hypervisor events.
2464 * However, context switches and cpu migrations are considered
2465 * to be kernel events, and page faults are never hypervisor
2468 switch (perf_event_id(&counter
->hw_event
)) {
2469 case PERF_COUNT_CPU_CLOCK
:
2470 hw_ops
= &perf_ops_cpu_clock
;
2472 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2473 hw_event
->irq_period
= 10000;
2475 case PERF_COUNT_TASK_CLOCK
:
2477 * If the user instantiates this as a per-cpu counter,
2478 * use the cpu_clock counter instead.
2480 if (counter
->ctx
->task
)
2481 hw_ops
= &perf_ops_task_clock
;
2483 hw_ops
= &perf_ops_cpu_clock
;
2485 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2486 hw_event
->irq_period
= 10000;
2488 case PERF_COUNT_PAGE_FAULTS
:
2489 case PERF_COUNT_PAGE_FAULTS_MIN
:
2490 case PERF_COUNT_PAGE_FAULTS_MAJ
:
2491 case PERF_COUNT_CONTEXT_SWITCHES
:
2492 hw_ops
= &perf_ops_generic
;
2494 case PERF_COUNT_CPU_MIGRATIONS
:
2495 if (!counter
->hw_event
.exclude_kernel
)
2496 hw_ops
= &perf_ops_cpu_migrations
;
2501 hwc
->irq_period
= hw_event
->irq_period
;
2507 * Allocate and initialize a counter structure
2509 static struct perf_counter
*
2510 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
2512 struct perf_counter_context
*ctx
,
2513 struct perf_counter
*group_leader
,
2516 const struct hw_perf_counter_ops
*hw_ops
;
2517 struct perf_counter
*counter
;
2520 counter
= kzalloc(sizeof(*counter
), gfpflags
);
2522 return ERR_PTR(-ENOMEM
);
2525 * Single counters are their own group leaders, with an
2526 * empty sibling list:
2529 group_leader
= counter
;
2531 mutex_init(&counter
->mutex
);
2532 INIT_LIST_HEAD(&counter
->list_entry
);
2533 INIT_LIST_HEAD(&counter
->event_entry
);
2534 INIT_LIST_HEAD(&counter
->sibling_list
);
2535 init_waitqueue_head(&counter
->waitq
);
2537 mutex_init(&counter
->mmap_mutex
);
2539 INIT_LIST_HEAD(&counter
->child_list
);
2542 counter
->hw_event
= *hw_event
;
2543 counter
->group_leader
= group_leader
;
2544 counter
->hw_ops
= NULL
;
2547 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2548 if (hw_event
->disabled
)
2549 counter
->state
= PERF_COUNTER_STATE_OFF
;
2553 if (perf_event_raw(hw_event
)) {
2554 hw_ops
= hw_perf_counter_init(counter
);
2558 switch (perf_event_type(hw_event
)) {
2559 case PERF_TYPE_HARDWARE
:
2560 hw_ops
= hw_perf_counter_init(counter
);
2563 case PERF_TYPE_SOFTWARE
:
2564 hw_ops
= sw_perf_counter_init(counter
);
2567 case PERF_TYPE_TRACEPOINT
:
2568 hw_ops
= tp_perf_counter_init(counter
);
2575 else if (IS_ERR(hw_ops
))
2576 err
= PTR_ERR(hw_ops
);
2580 return ERR_PTR(err
);
2583 counter
->hw_ops
= hw_ops
;
2589 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2591 * @hw_event_uptr: event type attributes for monitoring/sampling
2594 * @group_fd: group leader counter fd
2596 SYSCALL_DEFINE5(perf_counter_open
,
2597 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
2598 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
2600 struct perf_counter
*counter
, *group_leader
;
2601 struct perf_counter_hw_event hw_event
;
2602 struct perf_counter_context
*ctx
;
2603 struct file
*counter_file
= NULL
;
2604 struct file
*group_file
= NULL
;
2605 int fput_needed
= 0;
2606 int fput_needed2
= 0;
2609 /* for future expandability... */
2613 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
2617 * Get the target context (task or percpu):
2619 ctx
= find_get_context(pid
, cpu
);
2621 return PTR_ERR(ctx
);
2624 * Look up the group leader (we will attach this counter to it):
2626 group_leader
= NULL
;
2627 if (group_fd
!= -1) {
2629 group_file
= fget_light(group_fd
, &fput_needed
);
2631 goto err_put_context
;
2632 if (group_file
->f_op
!= &perf_fops
)
2633 goto err_put_context
;
2635 group_leader
= group_file
->private_data
;
2637 * Do not allow a recursive hierarchy (this new sibling
2638 * becoming part of another group-sibling):
2640 if (group_leader
->group_leader
!= group_leader
)
2641 goto err_put_context
;
2643 * Do not allow to attach to a group in a different
2644 * task or CPU context:
2646 if (group_leader
->ctx
!= ctx
)
2647 goto err_put_context
;
2649 * Only a group leader can be exclusive or pinned
2651 if (hw_event
.exclusive
|| hw_event
.pinned
)
2652 goto err_put_context
;
2655 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
2657 ret
= PTR_ERR(counter
);
2658 if (IS_ERR(counter
))
2659 goto err_put_context
;
2661 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
2663 goto err_free_put_context
;
2665 counter_file
= fget_light(ret
, &fput_needed2
);
2667 goto err_free_put_context
;
2669 counter
->filp
= counter_file
;
2670 mutex_lock(&ctx
->mutex
);
2671 perf_install_in_context(ctx
, counter
, cpu
);
2672 mutex_unlock(&ctx
->mutex
);
2674 fput_light(counter_file
, fput_needed2
);
2677 fput_light(group_file
, fput_needed
);
2681 err_free_put_context
:
2691 * Initialize the perf_counter context in a task_struct:
2694 __perf_counter_init_context(struct perf_counter_context
*ctx
,
2695 struct task_struct
*task
)
2697 memset(ctx
, 0, sizeof(*ctx
));
2698 spin_lock_init(&ctx
->lock
);
2699 mutex_init(&ctx
->mutex
);
2700 INIT_LIST_HEAD(&ctx
->counter_list
);
2701 INIT_LIST_HEAD(&ctx
->event_list
);
2706 * inherit a counter from parent task to child task:
2708 static struct perf_counter
*
2709 inherit_counter(struct perf_counter
*parent_counter
,
2710 struct task_struct
*parent
,
2711 struct perf_counter_context
*parent_ctx
,
2712 struct task_struct
*child
,
2713 struct perf_counter
*group_leader
,
2714 struct perf_counter_context
*child_ctx
)
2716 struct perf_counter
*child_counter
;
2719 * Instead of creating recursive hierarchies of counters,
2720 * we link inherited counters back to the original parent,
2721 * which has a filp for sure, which we use as the reference
2724 if (parent_counter
->parent
)
2725 parent_counter
= parent_counter
->parent
;
2727 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
2728 parent_counter
->cpu
, child_ctx
,
2729 group_leader
, GFP_KERNEL
);
2730 if (IS_ERR(child_counter
))
2731 return child_counter
;
2734 * Link it up in the child's context:
2736 child_counter
->task
= child
;
2737 add_counter_to_ctx(child_counter
, child_ctx
);
2739 child_counter
->parent
= parent_counter
;
2741 * inherit into child's child as well:
2743 child_counter
->hw_event
.inherit
= 1;
2746 * Get a reference to the parent filp - we will fput it
2747 * when the child counter exits. This is safe to do because
2748 * we are in the parent and we know that the filp still
2749 * exists and has a nonzero count:
2751 atomic_long_inc(&parent_counter
->filp
->f_count
);
2754 * Link this into the parent counter's child list
2756 mutex_lock(&parent_counter
->mutex
);
2757 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
2760 * Make the child state follow the state of the parent counter,
2761 * not its hw_event.disabled bit. We hold the parent's mutex,
2762 * so we won't race with perf_counter_{en,dis}able_family.
2764 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
2765 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2767 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
2769 mutex_unlock(&parent_counter
->mutex
);
2771 return child_counter
;
2774 static int inherit_group(struct perf_counter
*parent_counter
,
2775 struct task_struct
*parent
,
2776 struct perf_counter_context
*parent_ctx
,
2777 struct task_struct
*child
,
2778 struct perf_counter_context
*child_ctx
)
2780 struct perf_counter
*leader
;
2781 struct perf_counter
*sub
;
2782 struct perf_counter
*child_ctr
;
2784 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
2785 child
, NULL
, child_ctx
);
2787 return PTR_ERR(leader
);
2788 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
2789 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
2790 child
, leader
, child_ctx
);
2791 if (IS_ERR(child_ctr
))
2792 return PTR_ERR(child_ctr
);
2797 static void sync_child_counter(struct perf_counter
*child_counter
,
2798 struct perf_counter
*parent_counter
)
2800 u64 parent_val
, child_val
;
2802 parent_val
= atomic64_read(&parent_counter
->count
);
2803 child_val
= atomic64_read(&child_counter
->count
);
2806 * Add back the child's count to the parent's count:
2808 atomic64_add(child_val
, &parent_counter
->count
);
2809 atomic64_add(child_counter
->total_time_enabled
,
2810 &parent_counter
->child_total_time_enabled
);
2811 atomic64_add(child_counter
->total_time_running
,
2812 &parent_counter
->child_total_time_running
);
2815 * Remove this counter from the parent's list
2817 mutex_lock(&parent_counter
->mutex
);
2818 list_del_init(&child_counter
->child_list
);
2819 mutex_unlock(&parent_counter
->mutex
);
2822 * Release the parent counter, if this was the last
2825 fput(parent_counter
->filp
);
2829 __perf_counter_exit_task(struct task_struct
*child
,
2830 struct perf_counter
*child_counter
,
2831 struct perf_counter_context
*child_ctx
)
2833 struct perf_counter
*parent_counter
;
2834 struct perf_counter
*sub
, *tmp
;
2837 * If we do not self-reap then we have to wait for the
2838 * child task to unschedule (it will happen for sure),
2839 * so that its counter is at its final count. (This
2840 * condition triggers rarely - child tasks usually get
2841 * off their CPU before the parent has a chance to
2842 * get this far into the reaping action)
2844 if (child
!= current
) {
2845 wait_task_inactive(child
, 0);
2846 list_del_init(&child_counter
->list_entry
);
2847 update_counter_times(child_counter
);
2849 struct perf_cpu_context
*cpuctx
;
2850 unsigned long flags
;
2854 * Disable and unlink this counter.
2856 * Be careful about zapping the list - IRQ/NMI context
2857 * could still be processing it:
2859 curr_rq_lock_irq_save(&flags
);
2860 perf_flags
= hw_perf_save_disable();
2862 cpuctx
= &__get_cpu_var(perf_cpu_context
);
2864 group_sched_out(child_counter
, cpuctx
, child_ctx
);
2865 update_counter_times(child_counter
);
2867 list_del_init(&child_counter
->list_entry
);
2869 child_ctx
->nr_counters
--;
2871 hw_perf_restore(perf_flags
);
2872 curr_rq_unlock_irq_restore(&flags
);
2875 parent_counter
= child_counter
->parent
;
2877 * It can happen that parent exits first, and has counters
2878 * that are still around due to the child reference. These
2879 * counters need to be zapped - but otherwise linger.
2881 if (parent_counter
) {
2882 sync_child_counter(child_counter
, parent_counter
);
2883 list_for_each_entry_safe(sub
, tmp
, &child_counter
->sibling_list
,
2886 sync_child_counter(sub
, sub
->parent
);
2890 free_counter(child_counter
);
2895 * When a child task exits, feed back counter values to parent counters.
2897 * Note: we may be running in child context, but the PID is not hashed
2898 * anymore so new counters will not be added.
2900 void perf_counter_exit_task(struct task_struct
*child
)
2902 struct perf_counter
*child_counter
, *tmp
;
2903 struct perf_counter_context
*child_ctx
;
2905 child_ctx
= &child
->perf_counter_ctx
;
2907 if (likely(!child_ctx
->nr_counters
))
2910 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
2912 __perf_counter_exit_task(child
, child_counter
, child_ctx
);
2916 * Initialize the perf_counter context in task_struct
2918 void perf_counter_init_task(struct task_struct
*child
)
2920 struct perf_counter_context
*child_ctx
, *parent_ctx
;
2921 struct perf_counter
*counter
;
2922 struct task_struct
*parent
= current
;
2924 child_ctx
= &child
->perf_counter_ctx
;
2925 parent_ctx
= &parent
->perf_counter_ctx
;
2927 __perf_counter_init_context(child_ctx
, child
);
2930 * This is executed from the parent task context, so inherit
2931 * counters that have been marked for cloning:
2934 if (likely(!parent_ctx
->nr_counters
))
2938 * Lock the parent list. No need to lock the child - not PID
2939 * hashed yet and not running, so nobody can access it.
2941 mutex_lock(&parent_ctx
->mutex
);
2944 * We dont have to disable NMIs - we are only looking at
2945 * the list, not manipulating it:
2947 list_for_each_entry(counter
, &parent_ctx
->counter_list
, list_entry
) {
2948 if (!counter
->hw_event
.inherit
)
2951 if (inherit_group(counter
, parent
,
2952 parent_ctx
, child
, child_ctx
))
2956 mutex_unlock(&parent_ctx
->mutex
);
2959 static void __cpuinit
perf_counter_init_cpu(int cpu
)
2961 struct perf_cpu_context
*cpuctx
;
2963 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2964 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
2966 mutex_lock(&perf_resource_mutex
);
2967 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
2968 mutex_unlock(&perf_resource_mutex
);
2970 hw_perf_counter_setup(cpu
);
2973 #ifdef CONFIG_HOTPLUG_CPU
2974 static void __perf_counter_exit_cpu(void *info
)
2976 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
2977 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
2978 struct perf_counter
*counter
, *tmp
;
2980 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
2981 __perf_counter_remove_from_context(counter
);
2983 static void perf_counter_exit_cpu(int cpu
)
2985 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2986 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
2988 mutex_lock(&ctx
->mutex
);
2989 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
2990 mutex_unlock(&ctx
->mutex
);
2993 static inline void perf_counter_exit_cpu(int cpu
) { }
2996 static int __cpuinit
2997 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
2999 unsigned int cpu
= (long)hcpu
;
3003 case CPU_UP_PREPARE
:
3004 case CPU_UP_PREPARE_FROZEN
:
3005 perf_counter_init_cpu(cpu
);
3008 case CPU_DOWN_PREPARE
:
3009 case CPU_DOWN_PREPARE_FROZEN
:
3010 perf_counter_exit_cpu(cpu
);
3020 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
3021 .notifier_call
= perf_cpu_notify
,
3024 static int __init
perf_counter_init(void)
3026 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
3027 (void *)(long)smp_processor_id());
3028 register_cpu_notifier(&perf_cpu_nb
);
3032 early_initcall(perf_counter_init
);
3034 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
3036 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
3040 perf_set_reserve_percpu(struct sysdev_class
*class,
3044 struct perf_cpu_context
*cpuctx
;
3048 err
= strict_strtoul(buf
, 10, &val
);
3051 if (val
> perf_max_counters
)
3054 mutex_lock(&perf_resource_mutex
);
3055 perf_reserved_percpu
= val
;
3056 for_each_online_cpu(cpu
) {
3057 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3058 spin_lock_irq(&cpuctx
->ctx
.lock
);
3059 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
3060 perf_max_counters
- perf_reserved_percpu
);
3061 cpuctx
->max_pertask
= mpt
;
3062 spin_unlock_irq(&cpuctx
->ctx
.lock
);
3064 mutex_unlock(&perf_resource_mutex
);
3069 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
3071 return sprintf(buf
, "%d\n", perf_overcommit
);
3075 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
3080 err
= strict_strtoul(buf
, 10, &val
);
3086 mutex_lock(&perf_resource_mutex
);
3087 perf_overcommit
= val
;
3088 mutex_unlock(&perf_resource_mutex
);
3093 static SYSDEV_CLASS_ATTR(
3096 perf_show_reserve_percpu
,
3097 perf_set_reserve_percpu
3100 static SYSDEV_CLASS_ATTR(
3103 perf_show_overcommit
,
3107 static struct attribute
*perfclass_attrs
[] = {
3108 &attr_reserve_percpu
.attr
,
3109 &attr_overcommit
.attr
,
3113 static struct attribute_group perfclass_attr_group
= {
3114 .attrs
= perfclass_attrs
,
3115 .name
= "perf_counters",
3118 static int __init
perf_counter_sysfs_init(void)
3120 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
3121 &perfclass_attr_group
);
3123 device_initcall(perf_counter_sysfs_init
);