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>
29 #include <asm/irq_regs.h>
32 * Each CPU has a list of per CPU counters:
34 DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
36 int perf_max_counters __read_mostly
= 1;
37 static int perf_reserved_percpu __read_mostly
;
38 static int perf_overcommit __read_mostly
= 1;
41 * Mutex for (sysadmin-configurable) counter reservations:
43 static DEFINE_MUTEX(perf_resource_mutex
);
46 * Architecture provided APIs - weak aliases:
48 extern __weak
const struct hw_perf_counter_ops
*
49 hw_perf_counter_init(struct perf_counter
*counter
)
54 u64 __weak
hw_perf_save_disable(void) { return 0; }
55 void __weak
hw_perf_restore(u64 ctrl
) { barrier(); }
56 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
57 int __weak
hw_perf_group_sched_in(struct perf_counter
*group_leader
,
58 struct perf_cpu_context
*cpuctx
,
59 struct perf_counter_context
*ctx
, int cpu
)
64 void __weak
perf_counter_print_debug(void) { }
67 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
69 struct perf_counter
*group_leader
= counter
->group_leader
;
72 * Depending on whether it is a standalone or sibling counter,
73 * add it straight to the context's counter list, or to the group
74 * leader's sibling list:
76 if (counter
->group_leader
== counter
)
77 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
79 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
80 group_leader
->nr_siblings
++;
83 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
87 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
89 struct perf_counter
*sibling
, *tmp
;
91 list_del_init(&counter
->list_entry
);
92 list_del_rcu(&counter
->event_entry
);
94 if (counter
->group_leader
!= counter
)
95 counter
->group_leader
->nr_siblings
--;
98 * If this was a group counter with sibling counters then
99 * upgrade the siblings to singleton counters by adding them
100 * to the context list directly:
102 list_for_each_entry_safe(sibling
, tmp
,
103 &counter
->sibling_list
, list_entry
) {
105 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
106 sibling
->group_leader
= sibling
;
111 counter_sched_out(struct perf_counter
*counter
,
112 struct perf_cpu_context
*cpuctx
,
113 struct perf_counter_context
*ctx
)
115 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
118 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
119 counter
->tstamp_stopped
= ctx
->time_now
;
120 counter
->hw_ops
->disable(counter
);
123 if (!is_software_counter(counter
))
124 cpuctx
->active_oncpu
--;
126 if (counter
->hw_event
.exclusive
|| !cpuctx
->active_oncpu
)
127 cpuctx
->exclusive
= 0;
131 group_sched_out(struct perf_counter
*group_counter
,
132 struct perf_cpu_context
*cpuctx
,
133 struct perf_counter_context
*ctx
)
135 struct perf_counter
*counter
;
137 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
140 counter_sched_out(group_counter
, cpuctx
, ctx
);
143 * Schedule out siblings (if any):
145 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
146 counter_sched_out(counter
, cpuctx
, ctx
);
148 if (group_counter
->hw_event
.exclusive
)
149 cpuctx
->exclusive
= 0;
153 * Cross CPU call to remove a performance counter
155 * We disable the counter on the hardware level first. After that we
156 * remove it from the context list.
158 static void __perf_counter_remove_from_context(void *info
)
160 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
161 struct perf_counter
*counter
= info
;
162 struct perf_counter_context
*ctx
= counter
->ctx
;
167 * If this is a task context, we need to check whether it is
168 * the current task context of this cpu. If not it has been
169 * scheduled out before the smp call arrived.
171 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
174 curr_rq_lock_irq_save(&flags
);
175 spin_lock(&ctx
->lock
);
177 counter_sched_out(counter
, cpuctx
, ctx
);
179 counter
->task
= NULL
;
183 * Protect the list operation against NMI by disabling the
184 * counters on a global level. NOP for non NMI based counters.
186 perf_flags
= hw_perf_save_disable();
187 list_del_counter(counter
, ctx
);
188 hw_perf_restore(perf_flags
);
192 * Allow more per task counters with respect to the
195 cpuctx
->max_pertask
=
196 min(perf_max_counters
- ctx
->nr_counters
,
197 perf_max_counters
- perf_reserved_percpu
);
200 spin_unlock(&ctx
->lock
);
201 curr_rq_unlock_irq_restore(&flags
);
206 * Remove the counter from a task's (or a CPU's) list of counters.
208 * Must be called with counter->mutex and ctx->mutex held.
210 * CPU counters are removed with a smp call. For task counters we only
211 * call when the task is on a CPU.
213 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
215 struct perf_counter_context
*ctx
= counter
->ctx
;
216 struct task_struct
*task
= ctx
->task
;
220 * Per cpu counters are removed via an smp call and
221 * the removal is always sucessful.
223 smp_call_function_single(counter
->cpu
,
224 __perf_counter_remove_from_context
,
230 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
233 spin_lock_irq(&ctx
->lock
);
235 * If the context is active we need to retry the smp call.
237 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
238 spin_unlock_irq(&ctx
->lock
);
243 * The lock prevents that this context is scheduled in so we
244 * can remove the counter safely, if the call above did not
247 if (!list_empty(&counter
->list_entry
)) {
249 list_del_counter(counter
, ctx
);
250 counter
->task
= NULL
;
252 spin_unlock_irq(&ctx
->lock
);
256 * Get the current time for this context.
257 * If this is a task context, we use the task's task clock,
258 * or for a per-cpu context, we use the cpu clock.
260 static u64
get_context_time(struct perf_counter_context
*ctx
, int update
)
262 struct task_struct
*curr
= ctx
->task
;
265 return cpu_clock(smp_processor_id());
267 return __task_delta_exec(curr
, update
) + curr
->se
.sum_exec_runtime
;
271 * Update the record of the current time in a context.
273 static void update_context_time(struct perf_counter_context
*ctx
, int update
)
275 ctx
->time_now
= get_context_time(ctx
, update
) - ctx
->time_lost
;
279 * Update the total_time_enabled and total_time_running fields for a counter.
281 static void update_counter_times(struct perf_counter
*counter
)
283 struct perf_counter_context
*ctx
= counter
->ctx
;
286 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
287 counter
->total_time_enabled
= ctx
->time_now
-
288 counter
->tstamp_enabled
;
289 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
290 run_end
= counter
->tstamp_stopped
;
292 run_end
= ctx
->time_now
;
293 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
298 * Update total_time_enabled and total_time_running for all counters in a group.
300 static void update_group_times(struct perf_counter
*leader
)
302 struct perf_counter
*counter
;
304 update_counter_times(leader
);
305 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
306 update_counter_times(counter
);
310 * Cross CPU call to disable a performance counter
312 static void __perf_counter_disable(void *info
)
314 struct perf_counter
*counter
= info
;
315 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
316 struct perf_counter_context
*ctx
= counter
->ctx
;
320 * If this is a per-task counter, need to check whether this
321 * counter's task is the current task on this cpu.
323 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
326 curr_rq_lock_irq_save(&flags
);
327 spin_lock(&ctx
->lock
);
330 * If the counter is on, turn it off.
331 * If it is in error state, leave it in error state.
333 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
334 update_context_time(ctx
, 1);
335 update_counter_times(counter
);
336 if (counter
== counter
->group_leader
)
337 group_sched_out(counter
, cpuctx
, ctx
);
339 counter_sched_out(counter
, cpuctx
, ctx
);
340 counter
->state
= PERF_COUNTER_STATE_OFF
;
343 spin_unlock(&ctx
->lock
);
344 curr_rq_unlock_irq_restore(&flags
);
350 static void perf_counter_disable(struct perf_counter
*counter
)
352 struct perf_counter_context
*ctx
= counter
->ctx
;
353 struct task_struct
*task
= ctx
->task
;
357 * Disable the counter on the cpu that it's on
359 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
365 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
367 spin_lock_irq(&ctx
->lock
);
369 * If the counter is still active, we need to retry the cross-call.
371 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
372 spin_unlock_irq(&ctx
->lock
);
377 * Since we have the lock this context can't be scheduled
378 * in, so we can change the state safely.
380 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
381 update_counter_times(counter
);
382 counter
->state
= PERF_COUNTER_STATE_OFF
;
385 spin_unlock_irq(&ctx
->lock
);
389 * Disable a counter and all its children.
391 static void perf_counter_disable_family(struct perf_counter
*counter
)
393 struct perf_counter
*child
;
395 perf_counter_disable(counter
);
398 * Lock the mutex to protect the list of children
400 mutex_lock(&counter
->mutex
);
401 list_for_each_entry(child
, &counter
->child_list
, child_list
)
402 perf_counter_disable(child
);
403 mutex_unlock(&counter
->mutex
);
407 counter_sched_in(struct perf_counter
*counter
,
408 struct perf_cpu_context
*cpuctx
,
409 struct perf_counter_context
*ctx
,
412 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
415 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
416 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
418 * The new state must be visible before we turn it on in the hardware:
422 if (counter
->hw_ops
->enable(counter
)) {
423 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
428 counter
->tstamp_running
+= ctx
->time_now
- counter
->tstamp_stopped
;
430 if (!is_software_counter(counter
))
431 cpuctx
->active_oncpu
++;
434 if (counter
->hw_event
.exclusive
)
435 cpuctx
->exclusive
= 1;
441 * Return 1 for a group consisting entirely of software counters,
442 * 0 if the group contains any hardware counters.
444 static int is_software_only_group(struct perf_counter
*leader
)
446 struct perf_counter
*counter
;
448 if (!is_software_counter(leader
))
451 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
452 if (!is_software_counter(counter
))
459 * Work out whether we can put this counter group on the CPU now.
461 static int group_can_go_on(struct perf_counter
*counter
,
462 struct perf_cpu_context
*cpuctx
,
466 * Groups consisting entirely of software counters can always go on.
468 if (is_software_only_group(counter
))
471 * If an exclusive group is already on, no other hardware
472 * counters can go on.
474 if (cpuctx
->exclusive
)
477 * If this group is exclusive and there are already
478 * counters on the CPU, it can't go on.
480 if (counter
->hw_event
.exclusive
&& cpuctx
->active_oncpu
)
483 * Otherwise, try to add it if all previous groups were able
489 static void add_counter_to_ctx(struct perf_counter
*counter
,
490 struct perf_counter_context
*ctx
)
492 list_add_counter(counter
, ctx
);
494 counter
->prev_state
= PERF_COUNTER_STATE_OFF
;
495 counter
->tstamp_enabled
= ctx
->time_now
;
496 counter
->tstamp_running
= ctx
->time_now
;
497 counter
->tstamp_stopped
= ctx
->time_now
;
501 * Cross CPU call to install and enable a performance counter
503 static void __perf_install_in_context(void *info
)
505 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
506 struct perf_counter
*counter
= info
;
507 struct perf_counter_context
*ctx
= counter
->ctx
;
508 struct perf_counter
*leader
= counter
->group_leader
;
509 int cpu
= smp_processor_id();
515 * If this is a task context, we need to check whether it is
516 * the current task context of this cpu. If not it has been
517 * scheduled out before the smp call arrived.
519 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
522 curr_rq_lock_irq_save(&flags
);
523 spin_lock(&ctx
->lock
);
524 update_context_time(ctx
, 1);
527 * Protect the list operation against NMI by disabling the
528 * counters on a global level. NOP for non NMI based counters.
530 perf_flags
= hw_perf_save_disable();
532 add_counter_to_ctx(counter
, ctx
);
535 * Don't put the counter on if it is disabled or if
536 * it is in a group and the group isn't on.
538 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
539 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
543 * An exclusive counter can't go on if there are already active
544 * hardware counters, and no hardware counter can go on if there
545 * is already an exclusive counter on.
547 if (!group_can_go_on(counter
, cpuctx
, 1))
550 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
554 * This counter couldn't go on. If it is in a group
555 * then we have to pull the whole group off.
556 * If the counter group is pinned then put it in error state.
558 if (leader
!= counter
)
559 group_sched_out(leader
, cpuctx
, ctx
);
560 if (leader
->hw_event
.pinned
) {
561 update_group_times(leader
);
562 leader
->state
= PERF_COUNTER_STATE_ERROR
;
566 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
567 cpuctx
->max_pertask
--;
570 hw_perf_restore(perf_flags
);
572 spin_unlock(&ctx
->lock
);
573 curr_rq_unlock_irq_restore(&flags
);
577 * Attach a performance counter to a context
579 * First we add the counter to the list with the hardware enable bit
580 * in counter->hw_config cleared.
582 * If the counter is attached to a task which is on a CPU we use a smp
583 * call to enable it in the task context. The task might have been
584 * scheduled away, but we check this in the smp call again.
586 * Must be called with ctx->mutex held.
589 perf_install_in_context(struct perf_counter_context
*ctx
,
590 struct perf_counter
*counter
,
593 struct task_struct
*task
= ctx
->task
;
597 * Per cpu counters are installed via an smp call and
598 * the install is always sucessful.
600 smp_call_function_single(cpu
, __perf_install_in_context
,
605 counter
->task
= task
;
607 task_oncpu_function_call(task
, __perf_install_in_context
,
610 spin_lock_irq(&ctx
->lock
);
612 * we need to retry the smp call.
614 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
615 spin_unlock_irq(&ctx
->lock
);
620 * The lock prevents that this context is scheduled in so we
621 * can add the counter safely, if it the call above did not
624 if (list_empty(&counter
->list_entry
))
625 add_counter_to_ctx(counter
, ctx
);
626 spin_unlock_irq(&ctx
->lock
);
630 * Cross CPU call to enable a performance counter
632 static void __perf_counter_enable(void *info
)
634 struct perf_counter
*counter
= info
;
635 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
636 struct perf_counter_context
*ctx
= counter
->ctx
;
637 struct perf_counter
*leader
= counter
->group_leader
;
642 * If this is a per-task counter, need to check whether this
643 * counter's task is the current task on this cpu.
645 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
648 curr_rq_lock_irq_save(&flags
);
649 spin_lock(&ctx
->lock
);
650 update_context_time(ctx
, 1);
652 counter
->prev_state
= counter
->state
;
653 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
655 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
656 counter
->tstamp_enabled
= ctx
->time_now
- counter
->total_time_enabled
;
659 * If the counter is in a group and isn't the group leader,
660 * then don't put it on unless the group is on.
662 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
665 if (!group_can_go_on(counter
, cpuctx
, 1))
668 err
= counter_sched_in(counter
, cpuctx
, ctx
,
673 * If this counter can't go on and it's part of a
674 * group, then the whole group has to come off.
676 if (leader
!= counter
)
677 group_sched_out(leader
, cpuctx
, ctx
);
678 if (leader
->hw_event
.pinned
) {
679 update_group_times(leader
);
680 leader
->state
= PERF_COUNTER_STATE_ERROR
;
685 spin_unlock(&ctx
->lock
);
686 curr_rq_unlock_irq_restore(&flags
);
692 static void perf_counter_enable(struct perf_counter
*counter
)
694 struct perf_counter_context
*ctx
= counter
->ctx
;
695 struct task_struct
*task
= ctx
->task
;
699 * Enable the counter on the cpu that it's on
701 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
706 spin_lock_irq(&ctx
->lock
);
707 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
711 * If the counter is in error state, clear that first.
712 * That way, if we see the counter in error state below, we
713 * know that it has gone back into error state, as distinct
714 * from the task having been scheduled away before the
715 * cross-call arrived.
717 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
718 counter
->state
= PERF_COUNTER_STATE_OFF
;
721 spin_unlock_irq(&ctx
->lock
);
722 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
724 spin_lock_irq(&ctx
->lock
);
727 * If the context is active and the counter is still off,
728 * we need to retry the cross-call.
730 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
734 * Since we have the lock this context can't be scheduled
735 * in, so we can change the state safely.
737 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
738 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
739 counter
->tstamp_enabled
= ctx
->time_now
-
740 counter
->total_time_enabled
;
743 spin_unlock_irq(&ctx
->lock
);
747 * Enable a counter and all its children.
749 static void perf_counter_enable_family(struct perf_counter
*counter
)
751 struct perf_counter
*child
;
753 perf_counter_enable(counter
);
756 * Lock the mutex to protect the list of children
758 mutex_lock(&counter
->mutex
);
759 list_for_each_entry(child
, &counter
->child_list
, child_list
)
760 perf_counter_enable(child
);
761 mutex_unlock(&counter
->mutex
);
764 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
765 struct perf_cpu_context
*cpuctx
)
767 struct perf_counter
*counter
;
770 spin_lock(&ctx
->lock
);
772 if (likely(!ctx
->nr_counters
))
774 update_context_time(ctx
, 0);
776 flags
= hw_perf_save_disable();
777 if (ctx
->nr_active
) {
778 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
)
779 group_sched_out(counter
, cpuctx
, ctx
);
781 hw_perf_restore(flags
);
783 spin_unlock(&ctx
->lock
);
787 * Called from scheduler to remove the counters of the current task,
788 * with interrupts disabled.
790 * We stop each counter and update the counter value in counter->count.
792 * This does not protect us against NMI, but disable()
793 * sets the disabled bit in the control field of counter _before_
794 * accessing the counter control register. If a NMI hits, then it will
795 * not restart the counter.
797 void perf_counter_task_sched_out(struct task_struct
*task
, int cpu
)
799 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
800 struct perf_counter_context
*ctx
= &task
->perf_counter_ctx
;
801 struct pt_regs
*regs
;
803 if (likely(!cpuctx
->task_ctx
))
806 regs
= task_pt_regs(task
);
807 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES
, 1, 1, regs
);
808 __perf_counter_sched_out(ctx
, cpuctx
);
810 cpuctx
->task_ctx
= NULL
;
813 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
815 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
819 group_sched_in(struct perf_counter
*group_counter
,
820 struct perf_cpu_context
*cpuctx
,
821 struct perf_counter_context
*ctx
,
824 struct perf_counter
*counter
, *partial_group
;
827 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
830 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
832 return ret
< 0 ? ret
: 0;
834 group_counter
->prev_state
= group_counter
->state
;
835 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
839 * Schedule in siblings as one group (if any):
841 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
842 counter
->prev_state
= counter
->state
;
843 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
844 partial_group
= counter
;
853 * Groups can be scheduled in as one unit only, so undo any
854 * partial group before returning:
856 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
857 if (counter
== partial_group
)
859 counter_sched_out(counter
, cpuctx
, ctx
);
861 counter_sched_out(group_counter
, cpuctx
, ctx
);
867 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
868 struct perf_cpu_context
*cpuctx
, int cpu
)
870 struct perf_counter
*counter
;
874 spin_lock(&ctx
->lock
);
876 if (likely(!ctx
->nr_counters
))
880 * Add any time since the last sched_out to the lost time
881 * so it doesn't get included in the total_time_enabled and
882 * total_time_running measures for counters in the context.
884 ctx
->time_lost
= get_context_time(ctx
, 0) - ctx
->time_now
;
886 flags
= hw_perf_save_disable();
889 * First go through the list and put on any pinned groups
890 * in order to give them the best chance of going on.
892 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
893 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
894 !counter
->hw_event
.pinned
)
896 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
899 if (group_can_go_on(counter
, cpuctx
, 1))
900 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
903 * If this pinned group hasn't been scheduled,
904 * put it in error state.
906 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
907 update_group_times(counter
);
908 counter
->state
= PERF_COUNTER_STATE_ERROR
;
912 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
914 * Ignore counters in OFF or ERROR state, and
915 * ignore pinned counters since we did them already.
917 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
918 counter
->hw_event
.pinned
)
922 * Listen to the 'cpu' scheduling filter constraint
925 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
928 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
929 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
933 hw_perf_restore(flags
);
935 spin_unlock(&ctx
->lock
);
939 * Called from scheduler to add the counters of the current task
940 * with interrupts disabled.
942 * We restore the counter value and then enable it.
944 * This does not protect us against NMI, but enable()
945 * sets the enabled bit in the control field of counter _before_
946 * accessing the counter control register. If a NMI hits, then it will
947 * keep the counter running.
949 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
951 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
952 struct perf_counter_context
*ctx
= &task
->perf_counter_ctx
;
954 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
955 cpuctx
->task_ctx
= ctx
;
958 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
960 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
962 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
965 int perf_counter_task_disable(void)
967 struct task_struct
*curr
= current
;
968 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
969 struct perf_counter
*counter
;
974 if (likely(!ctx
->nr_counters
))
977 curr_rq_lock_irq_save(&flags
);
978 cpu
= smp_processor_id();
980 /* force the update of the task clock: */
981 __task_delta_exec(curr
, 1);
983 perf_counter_task_sched_out(curr
, cpu
);
985 spin_lock(&ctx
->lock
);
988 * Disable all the counters:
990 perf_flags
= hw_perf_save_disable();
992 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
993 if (counter
->state
!= PERF_COUNTER_STATE_ERROR
) {
994 update_group_times(counter
);
995 counter
->state
= PERF_COUNTER_STATE_OFF
;
999 hw_perf_restore(perf_flags
);
1001 spin_unlock(&ctx
->lock
);
1003 curr_rq_unlock_irq_restore(&flags
);
1008 int perf_counter_task_enable(void)
1010 struct task_struct
*curr
= current
;
1011 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
1012 struct perf_counter
*counter
;
1013 unsigned long flags
;
1017 if (likely(!ctx
->nr_counters
))
1020 curr_rq_lock_irq_save(&flags
);
1021 cpu
= smp_processor_id();
1023 /* force the update of the task clock: */
1024 __task_delta_exec(curr
, 1);
1026 perf_counter_task_sched_out(curr
, cpu
);
1028 spin_lock(&ctx
->lock
);
1031 * Disable all the counters:
1033 perf_flags
= hw_perf_save_disable();
1035 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1036 if (counter
->state
> PERF_COUNTER_STATE_OFF
)
1038 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
1039 counter
->tstamp_enabled
= ctx
->time_now
-
1040 counter
->total_time_enabled
;
1041 counter
->hw_event
.disabled
= 0;
1043 hw_perf_restore(perf_flags
);
1045 spin_unlock(&ctx
->lock
);
1047 perf_counter_task_sched_in(curr
, cpu
);
1049 curr_rq_unlock_irq_restore(&flags
);
1055 * Round-robin a context's counters:
1057 static void rotate_ctx(struct perf_counter_context
*ctx
)
1059 struct perf_counter
*counter
;
1062 if (!ctx
->nr_counters
)
1065 spin_lock(&ctx
->lock
);
1067 * Rotate the first entry last (works just fine for group counters too):
1069 perf_flags
= hw_perf_save_disable();
1070 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1071 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1074 hw_perf_restore(perf_flags
);
1076 spin_unlock(&ctx
->lock
);
1079 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1081 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1082 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
1083 const int rotate_percpu
= 0;
1086 perf_counter_cpu_sched_out(cpuctx
);
1087 perf_counter_task_sched_out(curr
, cpu
);
1090 rotate_ctx(&cpuctx
->ctx
);
1094 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1095 perf_counter_task_sched_in(curr
, cpu
);
1099 * Cross CPU call to read the hardware counter
1101 static void __read(void *info
)
1103 struct perf_counter
*counter
= info
;
1104 struct perf_counter_context
*ctx
= counter
->ctx
;
1105 unsigned long flags
;
1107 curr_rq_lock_irq_save(&flags
);
1109 update_context_time(ctx
, 1);
1110 counter
->hw_ops
->read(counter
);
1111 update_counter_times(counter
);
1112 curr_rq_unlock_irq_restore(&flags
);
1115 static u64
perf_counter_read(struct perf_counter
*counter
)
1118 * If counter is enabled and currently active on a CPU, update the
1119 * value in the counter structure:
1121 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1122 smp_call_function_single(counter
->oncpu
,
1123 __read
, counter
, 1);
1124 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1125 update_counter_times(counter
);
1128 return atomic64_read(&counter
->count
);
1131 static void put_context(struct perf_counter_context
*ctx
)
1134 put_task_struct(ctx
->task
);
1137 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1139 struct perf_cpu_context
*cpuctx
;
1140 struct perf_counter_context
*ctx
;
1141 struct task_struct
*task
;
1144 * If cpu is not a wildcard then this is a percpu counter:
1147 /* Must be root to operate on a CPU counter: */
1148 if (!capable(CAP_SYS_ADMIN
))
1149 return ERR_PTR(-EACCES
);
1151 if (cpu
< 0 || cpu
> num_possible_cpus())
1152 return ERR_PTR(-EINVAL
);
1155 * We could be clever and allow to attach a counter to an
1156 * offline CPU and activate it when the CPU comes up, but
1159 if (!cpu_isset(cpu
, cpu_online_map
))
1160 return ERR_PTR(-ENODEV
);
1162 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1172 task
= find_task_by_vpid(pid
);
1174 get_task_struct(task
);
1178 return ERR_PTR(-ESRCH
);
1180 ctx
= &task
->perf_counter_ctx
;
1183 /* Reuse ptrace permission checks for now. */
1184 if (!ptrace_may_access(task
, PTRACE_MODE_READ
)) {
1186 return ERR_PTR(-EACCES
);
1192 static void free_counter_rcu(struct rcu_head
*head
)
1194 struct perf_counter
*counter
;
1196 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1200 static void perf_pending_sync(struct perf_counter
*counter
);
1202 static void free_counter(struct perf_counter
*counter
)
1204 perf_pending_sync(counter
);
1206 if (counter
->destroy
)
1207 counter
->destroy(counter
);
1209 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1213 * Called when the last reference to the file is gone.
1215 static int perf_release(struct inode
*inode
, struct file
*file
)
1217 struct perf_counter
*counter
= file
->private_data
;
1218 struct perf_counter_context
*ctx
= counter
->ctx
;
1220 file
->private_data
= NULL
;
1222 mutex_lock(&ctx
->mutex
);
1223 mutex_lock(&counter
->mutex
);
1225 perf_counter_remove_from_context(counter
);
1227 mutex_unlock(&counter
->mutex
);
1228 mutex_unlock(&ctx
->mutex
);
1230 free_counter(counter
);
1237 * Read the performance counter - simple non blocking version for now
1240 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1246 * Return end-of-file for a read on a counter that is in
1247 * error state (i.e. because it was pinned but it couldn't be
1248 * scheduled on to the CPU at some point).
1250 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1253 mutex_lock(&counter
->mutex
);
1254 values
[0] = perf_counter_read(counter
);
1256 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1257 values
[n
++] = counter
->total_time_enabled
+
1258 atomic64_read(&counter
->child_total_time_enabled
);
1259 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1260 values
[n
++] = counter
->total_time_running
+
1261 atomic64_read(&counter
->child_total_time_running
);
1262 mutex_unlock(&counter
->mutex
);
1264 if (count
< n
* sizeof(u64
))
1266 count
= n
* sizeof(u64
);
1268 if (copy_to_user(buf
, values
, count
))
1275 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1277 struct perf_counter
*counter
= file
->private_data
;
1279 return perf_read_hw(counter
, buf
, count
);
1282 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1284 struct perf_counter
*counter
= file
->private_data
;
1285 struct perf_mmap_data
*data
;
1286 unsigned int events
;
1289 data
= rcu_dereference(counter
->data
);
1291 events
= atomic_xchg(&data
->wakeup
, 0);
1296 poll_wait(file
, &counter
->waitq
, wait
);
1301 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1303 struct perf_counter
*counter
= file
->private_data
;
1307 case PERF_COUNTER_IOC_ENABLE
:
1308 perf_counter_enable_family(counter
);
1310 case PERF_COUNTER_IOC_DISABLE
:
1311 perf_counter_disable_family(counter
);
1319 static void __perf_counter_update_userpage(struct perf_counter
*counter
,
1320 struct perf_mmap_data
*data
)
1322 struct perf_counter_mmap_page
*userpg
= data
->user_page
;
1325 * Disable preemption so as to not let the corresponding user-space
1326 * spin too long if we get preempted.
1331 userpg
->index
= counter
->hw
.idx
;
1332 userpg
->offset
= atomic64_read(&counter
->count
);
1333 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1334 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1336 userpg
->data_head
= atomic_read(&data
->head
);
1342 void perf_counter_update_userpage(struct perf_counter
*counter
)
1344 struct perf_mmap_data
*data
;
1347 data
= rcu_dereference(counter
->data
);
1349 __perf_counter_update_userpage(counter
, data
);
1353 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1355 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1356 struct perf_mmap_data
*data
;
1357 int ret
= VM_FAULT_SIGBUS
;
1360 data
= rcu_dereference(counter
->data
);
1364 if (vmf
->pgoff
== 0) {
1365 vmf
->page
= virt_to_page(data
->user_page
);
1367 int nr
= vmf
->pgoff
- 1;
1369 if ((unsigned)nr
> data
->nr_pages
)
1372 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1374 get_page(vmf
->page
);
1382 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1384 struct perf_mmap_data
*data
;
1388 WARN_ON(atomic_read(&counter
->mmap_count
));
1390 size
= sizeof(struct perf_mmap_data
);
1391 size
+= nr_pages
* sizeof(void *);
1393 data
= kzalloc(size
, GFP_KERNEL
);
1397 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1398 if (!data
->user_page
)
1399 goto fail_user_page
;
1401 for (i
= 0; i
< nr_pages
; i
++) {
1402 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1403 if (!data
->data_pages
[i
])
1404 goto fail_data_pages
;
1407 data
->nr_pages
= nr_pages
;
1409 rcu_assign_pointer(counter
->data
, data
);
1414 for (i
--; i
>= 0; i
--)
1415 free_page((unsigned long)data
->data_pages
[i
]);
1417 free_page((unsigned long)data
->user_page
);
1426 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1428 struct perf_mmap_data
*data
= container_of(rcu_head
,
1429 struct perf_mmap_data
, rcu_head
);
1432 free_page((unsigned long)data
->user_page
);
1433 for (i
= 0; i
< data
->nr_pages
; i
++)
1434 free_page((unsigned long)data
->data_pages
[i
]);
1438 static void perf_mmap_data_free(struct perf_counter
*counter
)
1440 struct perf_mmap_data
*data
= counter
->data
;
1442 WARN_ON(atomic_read(&counter
->mmap_count
));
1444 rcu_assign_pointer(counter
->data
, NULL
);
1445 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1448 static void perf_mmap_open(struct vm_area_struct
*vma
)
1450 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1452 atomic_inc(&counter
->mmap_count
);
1455 static void perf_mmap_close(struct vm_area_struct
*vma
)
1457 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1459 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
,
1460 &counter
->mmap_mutex
)) {
1461 perf_mmap_data_free(counter
);
1462 mutex_unlock(&counter
->mmap_mutex
);
1466 static struct vm_operations_struct perf_mmap_vmops
= {
1467 .open
= perf_mmap_open
,
1468 .close
= perf_mmap_close
,
1469 .fault
= perf_mmap_fault
,
1472 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1474 struct perf_counter
*counter
= file
->private_data
;
1475 unsigned long vma_size
;
1476 unsigned long nr_pages
;
1477 unsigned long locked
, lock_limit
;
1480 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1483 vma_size
= vma
->vm_end
- vma
->vm_start
;
1484 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1487 * If we have data pages ensure they're a power-of-two number, so we
1488 * can do bitmasks instead of modulo.
1490 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1493 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1496 if (vma
->vm_pgoff
!= 0)
1499 locked
= vma_size
>> PAGE_SHIFT
;
1500 locked
+= vma
->vm_mm
->locked_vm
;
1502 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1503 lock_limit
>>= PAGE_SHIFT
;
1505 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
))
1508 mutex_lock(&counter
->mmap_mutex
);
1509 if (atomic_inc_not_zero(&counter
->mmap_count
))
1512 WARN_ON(counter
->data
);
1513 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1515 atomic_set(&counter
->mmap_count
, 1);
1517 mutex_unlock(&counter
->mmap_mutex
);
1519 vma
->vm_flags
&= ~VM_MAYWRITE
;
1520 vma
->vm_flags
|= VM_RESERVED
;
1521 vma
->vm_ops
= &perf_mmap_vmops
;
1526 static const struct file_operations perf_fops
= {
1527 .release
= perf_release
,
1530 .unlocked_ioctl
= perf_ioctl
,
1531 .compat_ioctl
= perf_ioctl
,
1536 * Perf counter wakeup
1538 * If there's data, ensure we set the poll() state and publish everything
1539 * to user-space before waking everybody up.
1542 void perf_counter_wakeup(struct perf_counter
*counter
)
1544 struct perf_mmap_data
*data
;
1547 data
= rcu_dereference(counter
->data
);
1549 (void)atomic_xchg(&data
->wakeup
, POLL_IN
);
1550 __perf_counter_update_userpage(counter
, data
);
1554 wake_up_all(&counter
->waitq
);
1560 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1562 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1563 * single linked list and use cmpxchg() to add entries lockless.
1566 #define PENDING_TAIL ((struct perf_wakeup_entry *)-1UL)
1568 static DEFINE_PER_CPU(struct perf_wakeup_entry
*, perf_wakeup_head
) = {
1572 static void perf_pending_queue(struct perf_counter
*counter
)
1574 struct perf_wakeup_entry
**head
;
1575 struct perf_wakeup_entry
*prev
, *next
;
1577 if (cmpxchg(&counter
->wakeup
.next
, NULL
, PENDING_TAIL
) != NULL
)
1580 head
= &get_cpu_var(perf_wakeup_head
);
1583 prev
= counter
->wakeup
.next
= *head
;
1584 next
= &counter
->wakeup
;
1585 } while (cmpxchg(head
, prev
, next
) != prev
);
1587 set_perf_counter_pending();
1589 put_cpu_var(perf_wakeup_head
);
1592 static int __perf_pending_run(void)
1594 struct perf_wakeup_entry
*list
;
1597 list
= xchg(&__get_cpu_var(perf_wakeup_head
), PENDING_TAIL
);
1598 while (list
!= PENDING_TAIL
) {
1599 struct perf_counter
*counter
= container_of(list
,
1600 struct perf_counter
, wakeup
);
1604 counter
->wakeup
.next
= NULL
;
1606 * Ensure we observe the unqueue before we issue the wakeup,
1607 * so that we won't be waiting forever.
1608 * -- see perf_not_pending().
1612 perf_counter_wakeup(counter
);
1619 static inline int perf_not_pending(struct perf_counter
*counter
)
1622 * If we flush on whatever cpu we run, there is a chance we don't
1626 __perf_pending_run();
1630 * Ensure we see the proper queue state before going to sleep
1631 * so that we do not miss the wakeup. -- see perf_pending_handle()
1634 return counter
->wakeup
.next
== NULL
;
1637 static void perf_pending_sync(struct perf_counter
*counter
)
1639 wait_event(counter
->waitq
, perf_not_pending(counter
));
1642 void perf_counter_do_pending(void)
1644 __perf_pending_run();
1651 struct perf_output_handle
{
1652 struct perf_counter
*counter
;
1653 struct perf_mmap_data
*data
;
1654 unsigned int offset
;
1659 static int perf_output_begin(struct perf_output_handle
*handle
,
1660 struct perf_counter
*counter
, unsigned int size
)
1662 struct perf_mmap_data
*data
;
1663 unsigned int offset
, head
;
1666 data
= rcu_dereference(counter
->data
);
1670 if (!data
->nr_pages
)
1674 offset
= head
= atomic_read(&data
->head
);
1676 } while (atomic_cmpxchg(&data
->head
, offset
, head
) != offset
);
1678 handle
->counter
= counter
;
1679 handle
->data
= data
;
1680 handle
->offset
= offset
;
1681 handle
->head
= head
;
1682 handle
->wakeup
= (offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
);
1692 static void perf_output_copy(struct perf_output_handle
*handle
,
1693 void *buf
, unsigned int len
)
1695 unsigned int pages_mask
;
1696 unsigned int offset
;
1700 offset
= handle
->offset
;
1701 pages_mask
= handle
->data
->nr_pages
- 1;
1702 pages
= handle
->data
->data_pages
;
1705 unsigned int page_offset
;
1708 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
1709 page_offset
= offset
& (PAGE_SIZE
- 1);
1710 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
1712 memcpy(pages
[nr
] + page_offset
, buf
, size
);
1719 handle
->offset
= offset
;
1721 WARN_ON_ONCE(handle
->offset
> handle
->head
);
1724 #define perf_output_put(handle, x) \
1725 perf_output_copy((handle), &(x), sizeof(x))
1727 static void perf_output_end(struct perf_output_handle
*handle
, int nmi
)
1729 if (handle
->wakeup
) {
1731 perf_pending_queue(handle
->counter
);
1733 perf_counter_wakeup(handle
->counter
);
1738 static int perf_output_write(struct perf_counter
*counter
, int nmi
,
1739 void *buf
, ssize_t size
)
1741 struct perf_output_handle handle
;
1744 ret
= perf_output_begin(&handle
, counter
, size
);
1748 perf_output_copy(&handle
, buf
, size
);
1749 perf_output_end(&handle
, nmi
);
1755 static void perf_output_simple(struct perf_counter
*counter
,
1756 int nmi
, struct pt_regs
*regs
)
1760 struct perf_event_header header
;
1765 event
.header
.type
= PERF_EVENT_IP
;
1766 event
.ip
= instruction_pointer(regs
);
1768 size
= sizeof(event
);
1770 if (counter
->hw_event
.include_tid
) {
1771 /* namespace issues */
1772 event
.pid
= current
->group_leader
->pid
;
1773 event
.tid
= current
->pid
;
1775 event
.header
.type
|= __PERF_EVENT_TID
;
1777 size
-= sizeof(u64
);
1779 event
.header
.size
= size
;
1781 perf_output_write(counter
, nmi
, &event
, size
);
1784 static void perf_output_group(struct perf_counter
*counter
, int nmi
)
1786 struct perf_output_handle handle
;
1787 struct perf_event_header header
;
1788 struct perf_counter
*leader
, *sub
;
1796 size
= sizeof(header
) + counter
->nr_siblings
* sizeof(entry
);
1798 ret
= perf_output_begin(&handle
, counter
, size
);
1802 header
.type
= PERF_EVENT_GROUP
;
1805 perf_output_put(&handle
, header
);
1807 leader
= counter
->group_leader
;
1808 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
1810 sub
->hw_ops
->read(sub
);
1812 entry
.event
= sub
->hw_event
.config
;
1813 entry
.counter
= atomic64_read(&sub
->count
);
1815 perf_output_put(&handle
, entry
);
1818 perf_output_end(&handle
, nmi
);
1821 void perf_counter_output(struct perf_counter
*counter
,
1822 int nmi
, struct pt_regs
*regs
)
1824 switch (counter
->hw_event
.record_type
) {
1825 case PERF_RECORD_SIMPLE
:
1828 case PERF_RECORD_IRQ
:
1829 perf_output_simple(counter
, nmi
, regs
);
1832 case PERF_RECORD_GROUP
:
1833 perf_output_group(counter
, nmi
);
1839 * Generic software counter infrastructure
1842 static void perf_swcounter_update(struct perf_counter
*counter
)
1844 struct hw_perf_counter
*hwc
= &counter
->hw
;
1849 prev
= atomic64_read(&hwc
->prev_count
);
1850 now
= atomic64_read(&hwc
->count
);
1851 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
1856 atomic64_add(delta
, &counter
->count
);
1857 atomic64_sub(delta
, &hwc
->period_left
);
1860 static void perf_swcounter_set_period(struct perf_counter
*counter
)
1862 struct hw_perf_counter
*hwc
= &counter
->hw
;
1863 s64 left
= atomic64_read(&hwc
->period_left
);
1864 s64 period
= hwc
->irq_period
;
1866 if (unlikely(left
<= -period
)) {
1868 atomic64_set(&hwc
->period_left
, left
);
1871 if (unlikely(left
<= 0)) {
1873 atomic64_add(period
, &hwc
->period_left
);
1876 atomic64_set(&hwc
->prev_count
, -left
);
1877 atomic64_set(&hwc
->count
, -left
);
1880 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
1882 struct perf_counter
*counter
;
1883 struct pt_regs
*regs
;
1885 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
1886 counter
->hw_ops
->read(counter
);
1888 regs
= get_irq_regs();
1890 * In case we exclude kernel IPs or are somehow not in interrupt
1891 * context, provide the next best thing, the user IP.
1893 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
1894 !counter
->hw_event
.exclude_user
)
1895 regs
= task_pt_regs(current
);
1898 perf_counter_output(counter
, 0, regs
);
1900 hrtimer_forward_now(hrtimer
, ns_to_ktime(counter
->hw
.irq_period
));
1902 return HRTIMER_RESTART
;
1905 static void perf_swcounter_overflow(struct perf_counter
*counter
,
1906 int nmi
, struct pt_regs
*regs
)
1908 perf_swcounter_update(counter
);
1909 perf_swcounter_set_period(counter
);
1910 perf_counter_output(counter
, nmi
, regs
);
1913 static int perf_swcounter_match(struct perf_counter
*counter
,
1914 enum perf_event_types type
,
1915 u32 event
, struct pt_regs
*regs
)
1917 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1920 if (perf_event_raw(&counter
->hw_event
))
1923 if (perf_event_type(&counter
->hw_event
) != type
)
1926 if (perf_event_id(&counter
->hw_event
) != event
)
1929 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
1932 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
1938 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
1939 int nmi
, struct pt_regs
*regs
)
1941 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
1942 if (counter
->hw
.irq_period
&& !neg
)
1943 perf_swcounter_overflow(counter
, nmi
, regs
);
1946 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
1947 enum perf_event_types type
, u32 event
,
1948 u64 nr
, int nmi
, struct pt_regs
*regs
)
1950 struct perf_counter
*counter
;
1952 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
1956 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
1957 if (perf_swcounter_match(counter
, type
, event
, regs
))
1958 perf_swcounter_add(counter
, nr
, nmi
, regs
);
1963 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
1966 return &cpuctx
->recursion
[3];
1969 return &cpuctx
->recursion
[2];
1972 return &cpuctx
->recursion
[1];
1974 return &cpuctx
->recursion
[0];
1977 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
1978 u64 nr
, int nmi
, struct pt_regs
*regs
)
1980 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
1981 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
1989 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
, nr
, nmi
, regs
);
1990 if (cpuctx
->task_ctx
) {
1991 perf_swcounter_ctx_event(cpuctx
->task_ctx
, type
, event
,
1999 put_cpu_var(perf_cpu_context
);
2002 void perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
)
2004 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
);
2007 static void perf_swcounter_read(struct perf_counter
*counter
)
2009 perf_swcounter_update(counter
);
2012 static int perf_swcounter_enable(struct perf_counter
*counter
)
2014 perf_swcounter_set_period(counter
);
2018 static void perf_swcounter_disable(struct perf_counter
*counter
)
2020 perf_swcounter_update(counter
);
2023 static const struct hw_perf_counter_ops perf_ops_generic
= {
2024 .enable
= perf_swcounter_enable
,
2025 .disable
= perf_swcounter_disable
,
2026 .read
= perf_swcounter_read
,
2030 * Software counter: cpu wall time clock
2033 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
2035 int cpu
= raw_smp_processor_id();
2039 now
= cpu_clock(cpu
);
2040 prev
= atomic64_read(&counter
->hw
.prev_count
);
2041 atomic64_set(&counter
->hw
.prev_count
, now
);
2042 atomic64_add(now
- prev
, &counter
->count
);
2045 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
2047 struct hw_perf_counter
*hwc
= &counter
->hw
;
2048 int cpu
= raw_smp_processor_id();
2050 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
2051 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2052 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2053 if (hwc
->irq_period
) {
2054 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2055 ns_to_ktime(hwc
->irq_period
), 0,
2056 HRTIMER_MODE_REL
, 0);
2062 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
2064 hrtimer_cancel(&counter
->hw
.hrtimer
);
2065 cpu_clock_perf_counter_update(counter
);
2068 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
2070 cpu_clock_perf_counter_update(counter
);
2073 static const struct hw_perf_counter_ops perf_ops_cpu_clock
= {
2074 .enable
= cpu_clock_perf_counter_enable
,
2075 .disable
= cpu_clock_perf_counter_disable
,
2076 .read
= cpu_clock_perf_counter_read
,
2080 * Software counter: task time clock
2084 * Called from within the scheduler:
2086 static u64
task_clock_perf_counter_val(struct perf_counter
*counter
, int update
)
2088 struct task_struct
*curr
= counter
->task
;
2091 delta
= __task_delta_exec(curr
, update
);
2093 return curr
->se
.sum_exec_runtime
+ delta
;
2096 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
2101 prev
= atomic64_read(&counter
->hw
.prev_count
);
2103 atomic64_set(&counter
->hw
.prev_count
, now
);
2107 atomic64_add(delta
, &counter
->count
);
2110 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
2112 struct hw_perf_counter
*hwc
= &counter
->hw
;
2114 atomic64_set(&hwc
->prev_count
, task_clock_perf_counter_val(counter
, 0));
2115 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2116 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2117 if (hwc
->irq_period
) {
2118 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2119 ns_to_ktime(hwc
->irq_period
), 0,
2120 HRTIMER_MODE_REL
, 0);
2126 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
2128 hrtimer_cancel(&counter
->hw
.hrtimer
);
2129 task_clock_perf_counter_update(counter
,
2130 task_clock_perf_counter_val(counter
, 0));
2133 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
2135 task_clock_perf_counter_update(counter
,
2136 task_clock_perf_counter_val(counter
, 1));
2139 static const struct hw_perf_counter_ops perf_ops_task_clock
= {
2140 .enable
= task_clock_perf_counter_enable
,
2141 .disable
= task_clock_perf_counter_disable
,
2142 .read
= task_clock_perf_counter_read
,
2146 * Software counter: cpu migrations
2149 static inline u64
get_cpu_migrations(struct perf_counter
*counter
)
2151 struct task_struct
*curr
= counter
->ctx
->task
;
2154 return curr
->se
.nr_migrations
;
2155 return cpu_nr_migrations(smp_processor_id());
2158 static void cpu_migrations_perf_counter_update(struct perf_counter
*counter
)
2163 prev
= atomic64_read(&counter
->hw
.prev_count
);
2164 now
= get_cpu_migrations(counter
);
2166 atomic64_set(&counter
->hw
.prev_count
, now
);
2170 atomic64_add(delta
, &counter
->count
);
2173 static void cpu_migrations_perf_counter_read(struct perf_counter
*counter
)
2175 cpu_migrations_perf_counter_update(counter
);
2178 static int cpu_migrations_perf_counter_enable(struct perf_counter
*counter
)
2180 if (counter
->prev_state
<= PERF_COUNTER_STATE_OFF
)
2181 atomic64_set(&counter
->hw
.prev_count
,
2182 get_cpu_migrations(counter
));
2186 static void cpu_migrations_perf_counter_disable(struct perf_counter
*counter
)
2188 cpu_migrations_perf_counter_update(counter
);
2191 static const struct hw_perf_counter_ops perf_ops_cpu_migrations
= {
2192 .enable
= cpu_migrations_perf_counter_enable
,
2193 .disable
= cpu_migrations_perf_counter_disable
,
2194 .read
= cpu_migrations_perf_counter_read
,
2197 #ifdef CONFIG_EVENT_PROFILE
2198 void perf_tpcounter_event(int event_id
)
2200 struct pt_regs
*regs
= get_irq_regs();
2203 regs
= task_pt_regs(current
);
2205 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
);
2208 extern int ftrace_profile_enable(int);
2209 extern void ftrace_profile_disable(int);
2211 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
2213 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
2216 static const struct hw_perf_counter_ops
*
2217 tp_perf_counter_init(struct perf_counter
*counter
)
2219 int event_id
= perf_event_id(&counter
->hw_event
);
2222 ret
= ftrace_profile_enable(event_id
);
2226 counter
->destroy
= tp_perf_counter_destroy
;
2227 counter
->hw
.irq_period
= counter
->hw_event
.irq_period
;
2229 return &perf_ops_generic
;
2232 static const struct hw_perf_counter_ops
*
2233 tp_perf_counter_init(struct perf_counter
*counter
)
2239 static const struct hw_perf_counter_ops
*
2240 sw_perf_counter_init(struct perf_counter
*counter
)
2242 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
2243 const struct hw_perf_counter_ops
*hw_ops
= NULL
;
2244 struct hw_perf_counter
*hwc
= &counter
->hw
;
2247 * Software counters (currently) can't in general distinguish
2248 * between user, kernel and hypervisor events.
2249 * However, context switches and cpu migrations are considered
2250 * to be kernel events, and page faults are never hypervisor
2253 switch (perf_event_id(&counter
->hw_event
)) {
2254 case PERF_COUNT_CPU_CLOCK
:
2255 hw_ops
= &perf_ops_cpu_clock
;
2257 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2258 hw_event
->irq_period
= 10000;
2260 case PERF_COUNT_TASK_CLOCK
:
2262 * If the user instantiates this as a per-cpu counter,
2263 * use the cpu_clock counter instead.
2265 if (counter
->ctx
->task
)
2266 hw_ops
= &perf_ops_task_clock
;
2268 hw_ops
= &perf_ops_cpu_clock
;
2270 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2271 hw_event
->irq_period
= 10000;
2273 case PERF_COUNT_PAGE_FAULTS
:
2274 case PERF_COUNT_PAGE_FAULTS_MIN
:
2275 case PERF_COUNT_PAGE_FAULTS_MAJ
:
2276 case PERF_COUNT_CONTEXT_SWITCHES
:
2277 hw_ops
= &perf_ops_generic
;
2279 case PERF_COUNT_CPU_MIGRATIONS
:
2280 if (!counter
->hw_event
.exclude_kernel
)
2281 hw_ops
= &perf_ops_cpu_migrations
;
2286 hwc
->irq_period
= hw_event
->irq_period
;
2292 * Allocate and initialize a counter structure
2294 static struct perf_counter
*
2295 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
2297 struct perf_counter_context
*ctx
,
2298 struct perf_counter
*group_leader
,
2301 const struct hw_perf_counter_ops
*hw_ops
;
2302 struct perf_counter
*counter
;
2304 counter
= kzalloc(sizeof(*counter
), gfpflags
);
2309 * Single counters are their own group leaders, with an
2310 * empty sibling list:
2313 group_leader
= counter
;
2315 mutex_init(&counter
->mutex
);
2316 INIT_LIST_HEAD(&counter
->list_entry
);
2317 INIT_LIST_HEAD(&counter
->event_entry
);
2318 INIT_LIST_HEAD(&counter
->sibling_list
);
2319 init_waitqueue_head(&counter
->waitq
);
2321 mutex_init(&counter
->mmap_mutex
);
2323 INIT_LIST_HEAD(&counter
->child_list
);
2326 counter
->hw_event
= *hw_event
;
2327 counter
->group_leader
= group_leader
;
2328 counter
->hw_ops
= NULL
;
2331 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2332 if (hw_event
->disabled
)
2333 counter
->state
= PERF_COUNTER_STATE_OFF
;
2337 if (perf_event_raw(hw_event
)) {
2338 hw_ops
= hw_perf_counter_init(counter
);
2342 switch (perf_event_type(hw_event
)) {
2343 case PERF_TYPE_HARDWARE
:
2344 hw_ops
= hw_perf_counter_init(counter
);
2347 case PERF_TYPE_SOFTWARE
:
2348 hw_ops
= sw_perf_counter_init(counter
);
2351 case PERF_TYPE_TRACEPOINT
:
2352 hw_ops
= tp_perf_counter_init(counter
);
2361 counter
->hw_ops
= hw_ops
;
2367 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2369 * @hw_event_uptr: event type attributes for monitoring/sampling
2372 * @group_fd: group leader counter fd
2374 SYSCALL_DEFINE5(perf_counter_open
,
2375 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
2376 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
2378 struct perf_counter
*counter
, *group_leader
;
2379 struct perf_counter_hw_event hw_event
;
2380 struct perf_counter_context
*ctx
;
2381 struct file
*counter_file
= NULL
;
2382 struct file
*group_file
= NULL
;
2383 int fput_needed
= 0;
2384 int fput_needed2
= 0;
2387 /* for future expandability... */
2391 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
2395 * Get the target context (task or percpu):
2397 ctx
= find_get_context(pid
, cpu
);
2399 return PTR_ERR(ctx
);
2402 * Look up the group leader (we will attach this counter to it):
2404 group_leader
= NULL
;
2405 if (group_fd
!= -1) {
2407 group_file
= fget_light(group_fd
, &fput_needed
);
2409 goto err_put_context
;
2410 if (group_file
->f_op
!= &perf_fops
)
2411 goto err_put_context
;
2413 group_leader
= group_file
->private_data
;
2415 * Do not allow a recursive hierarchy (this new sibling
2416 * becoming part of another group-sibling):
2418 if (group_leader
->group_leader
!= group_leader
)
2419 goto err_put_context
;
2421 * Do not allow to attach to a group in a different
2422 * task or CPU context:
2424 if (group_leader
->ctx
!= ctx
)
2425 goto err_put_context
;
2427 * Only a group leader can be exclusive or pinned
2429 if (hw_event
.exclusive
|| hw_event
.pinned
)
2430 goto err_put_context
;
2434 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
2437 goto err_put_context
;
2439 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
2441 goto err_free_put_context
;
2443 counter_file
= fget_light(ret
, &fput_needed2
);
2445 goto err_free_put_context
;
2447 counter
->filp
= counter_file
;
2448 mutex_lock(&ctx
->mutex
);
2449 perf_install_in_context(ctx
, counter
, cpu
);
2450 mutex_unlock(&ctx
->mutex
);
2452 fput_light(counter_file
, fput_needed2
);
2455 fput_light(group_file
, fput_needed
);
2459 err_free_put_context
:
2469 * Initialize the perf_counter context in a task_struct:
2472 __perf_counter_init_context(struct perf_counter_context
*ctx
,
2473 struct task_struct
*task
)
2475 memset(ctx
, 0, sizeof(*ctx
));
2476 spin_lock_init(&ctx
->lock
);
2477 mutex_init(&ctx
->mutex
);
2478 INIT_LIST_HEAD(&ctx
->counter_list
);
2479 INIT_LIST_HEAD(&ctx
->event_list
);
2484 * inherit a counter from parent task to child task:
2486 static struct perf_counter
*
2487 inherit_counter(struct perf_counter
*parent_counter
,
2488 struct task_struct
*parent
,
2489 struct perf_counter_context
*parent_ctx
,
2490 struct task_struct
*child
,
2491 struct perf_counter
*group_leader
,
2492 struct perf_counter_context
*child_ctx
)
2494 struct perf_counter
*child_counter
;
2497 * Instead of creating recursive hierarchies of counters,
2498 * we link inherited counters back to the original parent,
2499 * which has a filp for sure, which we use as the reference
2502 if (parent_counter
->parent
)
2503 parent_counter
= parent_counter
->parent
;
2505 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
2506 parent_counter
->cpu
, child_ctx
,
2507 group_leader
, GFP_KERNEL
);
2512 * Link it up in the child's context:
2514 child_counter
->task
= child
;
2515 add_counter_to_ctx(child_counter
, child_ctx
);
2517 child_counter
->parent
= parent_counter
;
2519 * inherit into child's child as well:
2521 child_counter
->hw_event
.inherit
= 1;
2524 * Get a reference to the parent filp - we will fput it
2525 * when the child counter exits. This is safe to do because
2526 * we are in the parent and we know that the filp still
2527 * exists and has a nonzero count:
2529 atomic_long_inc(&parent_counter
->filp
->f_count
);
2532 * Link this into the parent counter's child list
2534 mutex_lock(&parent_counter
->mutex
);
2535 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
2538 * Make the child state follow the state of the parent counter,
2539 * not its hw_event.disabled bit. We hold the parent's mutex,
2540 * so we won't race with perf_counter_{en,dis}able_family.
2542 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
2543 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2545 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
2547 mutex_unlock(&parent_counter
->mutex
);
2549 return child_counter
;
2552 static int inherit_group(struct perf_counter
*parent_counter
,
2553 struct task_struct
*parent
,
2554 struct perf_counter_context
*parent_ctx
,
2555 struct task_struct
*child
,
2556 struct perf_counter_context
*child_ctx
)
2558 struct perf_counter
*leader
;
2559 struct perf_counter
*sub
;
2561 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
2562 child
, NULL
, child_ctx
);
2565 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
2566 if (!inherit_counter(sub
, parent
, parent_ctx
,
2567 child
, leader
, child_ctx
))
2573 static void sync_child_counter(struct perf_counter
*child_counter
,
2574 struct perf_counter
*parent_counter
)
2576 u64 parent_val
, child_val
;
2578 parent_val
= atomic64_read(&parent_counter
->count
);
2579 child_val
= atomic64_read(&child_counter
->count
);
2582 * Add back the child's count to the parent's count:
2584 atomic64_add(child_val
, &parent_counter
->count
);
2585 atomic64_add(child_counter
->total_time_enabled
,
2586 &parent_counter
->child_total_time_enabled
);
2587 atomic64_add(child_counter
->total_time_running
,
2588 &parent_counter
->child_total_time_running
);
2591 * Remove this counter from the parent's list
2593 mutex_lock(&parent_counter
->mutex
);
2594 list_del_init(&child_counter
->child_list
);
2595 mutex_unlock(&parent_counter
->mutex
);
2598 * Release the parent counter, if this was the last
2601 fput(parent_counter
->filp
);
2605 __perf_counter_exit_task(struct task_struct
*child
,
2606 struct perf_counter
*child_counter
,
2607 struct perf_counter_context
*child_ctx
)
2609 struct perf_counter
*parent_counter
;
2610 struct perf_counter
*sub
, *tmp
;
2613 * If we do not self-reap then we have to wait for the
2614 * child task to unschedule (it will happen for sure),
2615 * so that its counter is at its final count. (This
2616 * condition triggers rarely - child tasks usually get
2617 * off their CPU before the parent has a chance to
2618 * get this far into the reaping action)
2620 if (child
!= current
) {
2621 wait_task_inactive(child
, 0);
2622 list_del_init(&child_counter
->list_entry
);
2623 update_counter_times(child_counter
);
2625 struct perf_cpu_context
*cpuctx
;
2626 unsigned long flags
;
2630 * Disable and unlink this counter.
2632 * Be careful about zapping the list - IRQ/NMI context
2633 * could still be processing it:
2635 curr_rq_lock_irq_save(&flags
);
2636 perf_flags
= hw_perf_save_disable();
2638 cpuctx
= &__get_cpu_var(perf_cpu_context
);
2640 group_sched_out(child_counter
, cpuctx
, child_ctx
);
2641 update_counter_times(child_counter
);
2643 list_del_init(&child_counter
->list_entry
);
2645 child_ctx
->nr_counters
--;
2647 hw_perf_restore(perf_flags
);
2648 curr_rq_unlock_irq_restore(&flags
);
2651 parent_counter
= child_counter
->parent
;
2653 * It can happen that parent exits first, and has counters
2654 * that are still around due to the child reference. These
2655 * counters need to be zapped - but otherwise linger.
2657 if (parent_counter
) {
2658 sync_child_counter(child_counter
, parent_counter
);
2659 list_for_each_entry_safe(sub
, tmp
, &child_counter
->sibling_list
,
2662 sync_child_counter(sub
, sub
->parent
);
2666 free_counter(child_counter
);
2671 * When a child task exits, feed back counter values to parent counters.
2673 * Note: we may be running in child context, but the PID is not hashed
2674 * anymore so new counters will not be added.
2676 void perf_counter_exit_task(struct task_struct
*child
)
2678 struct perf_counter
*child_counter
, *tmp
;
2679 struct perf_counter_context
*child_ctx
;
2681 child_ctx
= &child
->perf_counter_ctx
;
2683 if (likely(!child_ctx
->nr_counters
))
2686 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
2688 __perf_counter_exit_task(child
, child_counter
, child_ctx
);
2692 * Initialize the perf_counter context in task_struct
2694 void perf_counter_init_task(struct task_struct
*child
)
2696 struct perf_counter_context
*child_ctx
, *parent_ctx
;
2697 struct perf_counter
*counter
;
2698 struct task_struct
*parent
= current
;
2700 child_ctx
= &child
->perf_counter_ctx
;
2701 parent_ctx
= &parent
->perf_counter_ctx
;
2703 __perf_counter_init_context(child_ctx
, child
);
2706 * This is executed from the parent task context, so inherit
2707 * counters that have been marked for cloning:
2710 if (likely(!parent_ctx
->nr_counters
))
2714 * Lock the parent list. No need to lock the child - not PID
2715 * hashed yet and not running, so nobody can access it.
2717 mutex_lock(&parent_ctx
->mutex
);
2720 * We dont have to disable NMIs - we are only looking at
2721 * the list, not manipulating it:
2723 list_for_each_entry(counter
, &parent_ctx
->counter_list
, list_entry
) {
2724 if (!counter
->hw_event
.inherit
)
2727 if (inherit_group(counter
, parent
,
2728 parent_ctx
, child
, child_ctx
))
2732 mutex_unlock(&parent_ctx
->mutex
);
2735 static void __cpuinit
perf_counter_init_cpu(int cpu
)
2737 struct perf_cpu_context
*cpuctx
;
2739 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2740 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
2742 mutex_lock(&perf_resource_mutex
);
2743 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
2744 mutex_unlock(&perf_resource_mutex
);
2746 hw_perf_counter_setup(cpu
);
2749 #ifdef CONFIG_HOTPLUG_CPU
2750 static void __perf_counter_exit_cpu(void *info
)
2752 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
2753 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
2754 struct perf_counter
*counter
, *tmp
;
2756 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
2757 __perf_counter_remove_from_context(counter
);
2759 static void perf_counter_exit_cpu(int cpu
)
2761 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2762 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
2764 mutex_lock(&ctx
->mutex
);
2765 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
2766 mutex_unlock(&ctx
->mutex
);
2769 static inline void perf_counter_exit_cpu(int cpu
) { }
2772 static int __cpuinit
2773 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
2775 unsigned int cpu
= (long)hcpu
;
2779 case CPU_UP_PREPARE
:
2780 case CPU_UP_PREPARE_FROZEN
:
2781 perf_counter_init_cpu(cpu
);
2784 case CPU_DOWN_PREPARE
:
2785 case CPU_DOWN_PREPARE_FROZEN
:
2786 perf_counter_exit_cpu(cpu
);
2796 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
2797 .notifier_call
= perf_cpu_notify
,
2800 static int __init
perf_counter_init(void)
2802 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
2803 (void *)(long)smp_processor_id());
2804 register_cpu_notifier(&perf_cpu_nb
);
2808 early_initcall(perf_counter_init
);
2810 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
2812 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
2816 perf_set_reserve_percpu(struct sysdev_class
*class,
2820 struct perf_cpu_context
*cpuctx
;
2824 err
= strict_strtoul(buf
, 10, &val
);
2827 if (val
> perf_max_counters
)
2830 mutex_lock(&perf_resource_mutex
);
2831 perf_reserved_percpu
= val
;
2832 for_each_online_cpu(cpu
) {
2833 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
2834 spin_lock_irq(&cpuctx
->ctx
.lock
);
2835 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
2836 perf_max_counters
- perf_reserved_percpu
);
2837 cpuctx
->max_pertask
= mpt
;
2838 spin_unlock_irq(&cpuctx
->ctx
.lock
);
2840 mutex_unlock(&perf_resource_mutex
);
2845 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
2847 return sprintf(buf
, "%d\n", perf_overcommit
);
2851 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
2856 err
= strict_strtoul(buf
, 10, &val
);
2862 mutex_lock(&perf_resource_mutex
);
2863 perf_overcommit
= val
;
2864 mutex_unlock(&perf_resource_mutex
);
2869 static SYSDEV_CLASS_ATTR(
2872 perf_show_reserve_percpu
,
2873 perf_set_reserve_percpu
2876 static SYSDEV_CLASS_ATTR(
2879 perf_show_overcommit
,
2883 static struct attribute
*perfclass_attrs
[] = {
2884 &attr_reserve_percpu
.attr
,
2885 &attr_overcommit
.attr
,
2889 static struct attribute_group perfclass_attr_group
= {
2890 .attrs
= perfclass_attrs
,
2891 .name
= "perf_counters",
2894 static int __init
perf_counter_sysfs_init(void)
2896 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
2897 &perfclass_attr_group
);
2899 device_initcall(perf_counter_sysfs_init
);