2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/ptrace.h>
20 #include <linux/percpu.h>
21 #include <linux/vmstat.h>
22 #include <linux/hardirq.h>
23 #include <linux/rculist.h>
24 #include <linux/uaccess.h>
25 #include <linux/syscalls.h>
26 #include <linux/anon_inodes.h>
27 #include <linux/kernel_stat.h>
28 #include <linux/perf_counter.h>
29 #include <linux/dcache.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
38 int perf_max_counters __read_mostly
= 1;
39 static int perf_reserved_percpu __read_mostly
;
40 static int perf_overcommit __read_mostly
= 1;
42 static atomic_t nr_mmap_tracking __read_mostly
;
43 static atomic_t nr_munmap_tracking __read_mostly
;
44 static atomic_t nr_comm_tracking __read_mostly
;
46 int sysctl_perf_counter_priv __read_mostly
; /* do we need to be privileged */
49 * Mutex for (sysadmin-configurable) counter reservations:
51 static DEFINE_MUTEX(perf_resource_mutex
);
54 * Architecture provided APIs - weak aliases:
56 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
61 u64 __weak
hw_perf_save_disable(void) { return 0; }
62 void __weak
hw_perf_restore(u64 ctrl
) { barrier(); }
63 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
64 int __weak
hw_perf_group_sched_in(struct perf_counter
*group_leader
,
65 struct perf_cpu_context
*cpuctx
,
66 struct perf_counter_context
*ctx
, int cpu
)
71 void __weak
perf_counter_print_debug(void) { }
74 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
76 struct perf_counter
*group_leader
= counter
->group_leader
;
79 * Depending on whether it is a standalone or sibling counter,
80 * add it straight to the context's counter list, or to the group
81 * leader's sibling list:
83 if (counter
->group_leader
== counter
)
84 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
86 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
87 group_leader
->nr_siblings
++;
90 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
94 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
96 struct perf_counter
*sibling
, *tmp
;
98 list_del_init(&counter
->list_entry
);
99 list_del_rcu(&counter
->event_entry
);
101 if (counter
->group_leader
!= counter
)
102 counter
->group_leader
->nr_siblings
--;
105 * If this was a group counter with sibling counters then
106 * upgrade the siblings to singleton counters by adding them
107 * to the context list directly:
109 list_for_each_entry_safe(sibling
, tmp
,
110 &counter
->sibling_list
, list_entry
) {
112 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
113 sibling
->group_leader
= sibling
;
118 counter_sched_out(struct perf_counter
*counter
,
119 struct perf_cpu_context
*cpuctx
,
120 struct perf_counter_context
*ctx
)
122 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
125 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
126 counter
->tstamp_stopped
= ctx
->time
;
127 counter
->pmu
->disable(counter
);
130 if (!is_software_counter(counter
))
131 cpuctx
->active_oncpu
--;
133 if (counter
->hw_event
.exclusive
|| !cpuctx
->active_oncpu
)
134 cpuctx
->exclusive
= 0;
138 group_sched_out(struct perf_counter
*group_counter
,
139 struct perf_cpu_context
*cpuctx
,
140 struct perf_counter_context
*ctx
)
142 struct perf_counter
*counter
;
144 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
147 counter_sched_out(group_counter
, cpuctx
, ctx
);
150 * Schedule out siblings (if any):
152 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
153 counter_sched_out(counter
, cpuctx
, ctx
);
155 if (group_counter
->hw_event
.exclusive
)
156 cpuctx
->exclusive
= 0;
160 * Cross CPU call to remove a performance counter
162 * We disable the counter on the hardware level first. After that we
163 * remove it from the context list.
165 static void __perf_counter_remove_from_context(void *info
)
167 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
168 struct perf_counter
*counter
= info
;
169 struct perf_counter_context
*ctx
= counter
->ctx
;
174 * If this is a task context, we need to check whether it is
175 * the current task context of this cpu. If not it has been
176 * scheduled out before the smp call arrived.
178 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
181 spin_lock_irqsave(&ctx
->lock
, flags
);
183 counter_sched_out(counter
, cpuctx
, ctx
);
185 counter
->task
= NULL
;
189 * Protect the list operation against NMI by disabling the
190 * counters on a global level. NOP for non NMI based counters.
192 perf_flags
= hw_perf_save_disable();
193 list_del_counter(counter
, ctx
);
194 hw_perf_restore(perf_flags
);
198 * Allow more per task counters with respect to the
201 cpuctx
->max_pertask
=
202 min(perf_max_counters
- ctx
->nr_counters
,
203 perf_max_counters
- perf_reserved_percpu
);
206 spin_unlock_irqrestore(&ctx
->lock
, flags
);
211 * Remove the counter from a task's (or a CPU's) list of counters.
213 * Must be called with counter->mutex and ctx->mutex held.
215 * CPU counters are removed with a smp call. For task counters we only
216 * call when the task is on a CPU.
218 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
220 struct perf_counter_context
*ctx
= counter
->ctx
;
221 struct task_struct
*task
= ctx
->task
;
225 * Per cpu counters are removed via an smp call and
226 * the removal is always sucessful.
228 smp_call_function_single(counter
->cpu
,
229 __perf_counter_remove_from_context
,
235 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
238 spin_lock_irq(&ctx
->lock
);
240 * If the context is active we need to retry the smp call.
242 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
243 spin_unlock_irq(&ctx
->lock
);
248 * The lock prevents that this context is scheduled in so we
249 * can remove the counter safely, if the call above did not
252 if (!list_empty(&counter
->list_entry
)) {
254 list_del_counter(counter
, ctx
);
255 counter
->task
= NULL
;
257 spin_unlock_irq(&ctx
->lock
);
260 static inline u64
perf_clock(void)
262 return cpu_clock(smp_processor_id());
266 * Update the record of the current time in a context.
268 static void update_context_time(struct perf_counter_context
*ctx
)
270 u64 now
= perf_clock();
272 ctx
->time
+= now
- ctx
->timestamp
;
273 ctx
->timestamp
= now
;
277 * Update the total_time_enabled and total_time_running fields for a counter.
279 static void update_counter_times(struct perf_counter
*counter
)
281 struct perf_counter_context
*ctx
= counter
->ctx
;
284 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
287 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
289 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
290 run_end
= counter
->tstamp_stopped
;
294 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 spin_lock_irqsave(&ctx
->lock
, flags
);
329 * If the counter is on, turn it off.
330 * If it is in error state, leave it in error state.
332 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
333 update_context_time(ctx
);
334 update_counter_times(counter
);
335 if (counter
== counter
->group_leader
)
336 group_sched_out(counter
, cpuctx
, ctx
);
338 counter_sched_out(counter
, cpuctx
, ctx
);
339 counter
->state
= PERF_COUNTER_STATE_OFF
;
342 spin_unlock_irqrestore(&ctx
->lock
, flags
);
348 static void perf_counter_disable(struct perf_counter
*counter
)
350 struct perf_counter_context
*ctx
= counter
->ctx
;
351 struct task_struct
*task
= ctx
->task
;
355 * Disable the counter on the cpu that it's on
357 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
363 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
365 spin_lock_irq(&ctx
->lock
);
367 * If the counter is still active, we need to retry the cross-call.
369 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
370 spin_unlock_irq(&ctx
->lock
);
375 * Since we have the lock this context can't be scheduled
376 * in, so we can change the state safely.
378 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
379 update_counter_times(counter
);
380 counter
->state
= PERF_COUNTER_STATE_OFF
;
383 spin_unlock_irq(&ctx
->lock
);
387 * Disable a counter and all its children.
389 static void perf_counter_disable_family(struct perf_counter
*counter
)
391 struct perf_counter
*child
;
393 perf_counter_disable(counter
);
396 * Lock the mutex to protect the list of children
398 mutex_lock(&counter
->mutex
);
399 list_for_each_entry(child
, &counter
->child_list
, child_list
)
400 perf_counter_disable(child
);
401 mutex_unlock(&counter
->mutex
);
405 counter_sched_in(struct perf_counter
*counter
,
406 struct perf_cpu_context
*cpuctx
,
407 struct perf_counter_context
*ctx
,
410 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
413 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
414 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
416 * The new state must be visible before we turn it on in the hardware:
420 if (counter
->pmu
->enable(counter
)) {
421 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
426 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
428 if (!is_software_counter(counter
))
429 cpuctx
->active_oncpu
++;
432 if (counter
->hw_event
.exclusive
)
433 cpuctx
->exclusive
= 1;
439 * Return 1 for a group consisting entirely of software counters,
440 * 0 if the group contains any hardware counters.
442 static int is_software_only_group(struct perf_counter
*leader
)
444 struct perf_counter
*counter
;
446 if (!is_software_counter(leader
))
449 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
450 if (!is_software_counter(counter
))
457 * Work out whether we can put this counter group on the CPU now.
459 static int group_can_go_on(struct perf_counter
*counter
,
460 struct perf_cpu_context
*cpuctx
,
464 * Groups consisting entirely of software counters can always go on.
466 if (is_software_only_group(counter
))
469 * If an exclusive group is already on, no other hardware
470 * counters can go on.
472 if (cpuctx
->exclusive
)
475 * If this group is exclusive and there are already
476 * counters on the CPU, it can't go on.
478 if (counter
->hw_event
.exclusive
&& cpuctx
->active_oncpu
)
481 * Otherwise, try to add it if all previous groups were able
487 static void add_counter_to_ctx(struct perf_counter
*counter
,
488 struct perf_counter_context
*ctx
)
490 list_add_counter(counter
, ctx
);
492 counter
->prev_state
= PERF_COUNTER_STATE_OFF
;
493 counter
->tstamp_enabled
= ctx
->time
;
494 counter
->tstamp_running
= ctx
->time
;
495 counter
->tstamp_stopped
= ctx
->time
;
499 * Cross CPU call to install and enable a performance counter
501 static void __perf_install_in_context(void *info
)
503 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
504 struct perf_counter
*counter
= info
;
505 struct perf_counter_context
*ctx
= counter
->ctx
;
506 struct perf_counter
*leader
= counter
->group_leader
;
507 int cpu
= smp_processor_id();
513 * If this is a task context, we need to check whether it is
514 * the current task context of this cpu. If not it has been
515 * scheduled out before the smp call arrived.
517 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
520 spin_lock_irqsave(&ctx
->lock
, flags
);
521 update_context_time(ctx
);
524 * Protect the list operation against NMI by disabling the
525 * counters on a global level. NOP for non NMI based counters.
527 perf_flags
= hw_perf_save_disable();
529 add_counter_to_ctx(counter
, ctx
);
532 * Don't put the counter on if it is disabled or if
533 * it is in a group and the group isn't on.
535 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
536 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
540 * An exclusive counter can't go on if there are already active
541 * hardware counters, and no hardware counter can go on if there
542 * is already an exclusive counter on.
544 if (!group_can_go_on(counter
, cpuctx
, 1))
547 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
551 * This counter couldn't go on. If it is in a group
552 * then we have to pull the whole group off.
553 * If the counter group is pinned then put it in error state.
555 if (leader
!= counter
)
556 group_sched_out(leader
, cpuctx
, ctx
);
557 if (leader
->hw_event
.pinned
) {
558 update_group_times(leader
);
559 leader
->state
= PERF_COUNTER_STATE_ERROR
;
563 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
564 cpuctx
->max_pertask
--;
567 hw_perf_restore(perf_flags
);
569 spin_unlock_irqrestore(&ctx
->lock
, flags
);
573 * Attach a performance counter to a context
575 * First we add the counter to the list with the hardware enable bit
576 * in counter->hw_config cleared.
578 * If the counter is attached to a task which is on a CPU we use a smp
579 * call to enable it in the task context. The task might have been
580 * scheduled away, but we check this in the smp call again.
582 * Must be called with ctx->mutex held.
585 perf_install_in_context(struct perf_counter_context
*ctx
,
586 struct perf_counter
*counter
,
589 struct task_struct
*task
= ctx
->task
;
593 * Per cpu counters are installed via an smp call and
594 * the install is always sucessful.
596 smp_call_function_single(cpu
, __perf_install_in_context
,
601 counter
->task
= task
;
603 task_oncpu_function_call(task
, __perf_install_in_context
,
606 spin_lock_irq(&ctx
->lock
);
608 * we need to retry the smp call.
610 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
611 spin_unlock_irq(&ctx
->lock
);
616 * The lock prevents that this context is scheduled in so we
617 * can add the counter safely, if it the call above did not
620 if (list_empty(&counter
->list_entry
))
621 add_counter_to_ctx(counter
, ctx
);
622 spin_unlock_irq(&ctx
->lock
);
626 * Cross CPU call to enable a performance counter
628 static void __perf_counter_enable(void *info
)
630 struct perf_counter
*counter
= info
;
631 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
632 struct perf_counter_context
*ctx
= counter
->ctx
;
633 struct perf_counter
*leader
= counter
->group_leader
;
638 * If this is a per-task counter, need to check whether this
639 * counter's task is the current task on this cpu.
641 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
644 spin_lock_irqsave(&ctx
->lock
, flags
);
645 update_context_time(ctx
);
647 counter
->prev_state
= counter
->state
;
648 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
650 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
651 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
654 * If the counter is in a group and isn't the group leader,
655 * then don't put it on unless the group is on.
657 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
660 if (!group_can_go_on(counter
, cpuctx
, 1))
663 err
= counter_sched_in(counter
, cpuctx
, ctx
,
668 * If this counter can't go on and it's part of a
669 * group, then the whole group has to come off.
671 if (leader
!= counter
)
672 group_sched_out(leader
, cpuctx
, ctx
);
673 if (leader
->hw_event
.pinned
) {
674 update_group_times(leader
);
675 leader
->state
= PERF_COUNTER_STATE_ERROR
;
680 spin_unlock_irqrestore(&ctx
->lock
, flags
);
686 static void perf_counter_enable(struct perf_counter
*counter
)
688 struct perf_counter_context
*ctx
= counter
->ctx
;
689 struct task_struct
*task
= ctx
->task
;
693 * Enable the counter on the cpu that it's on
695 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
700 spin_lock_irq(&ctx
->lock
);
701 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
705 * If the counter is in error state, clear that first.
706 * That way, if we see the counter in error state below, we
707 * know that it has gone back into error state, as distinct
708 * from the task having been scheduled away before the
709 * cross-call arrived.
711 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
712 counter
->state
= PERF_COUNTER_STATE_OFF
;
715 spin_unlock_irq(&ctx
->lock
);
716 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
718 spin_lock_irq(&ctx
->lock
);
721 * If the context is active and the counter is still off,
722 * we need to retry the cross-call.
724 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
728 * Since we have the lock this context can't be scheduled
729 * in, so we can change the state safely.
731 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
732 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
733 counter
->tstamp_enabled
=
734 ctx
->time
- counter
->total_time_enabled
;
737 spin_unlock_irq(&ctx
->lock
);
740 static void perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
742 atomic_add(refresh
, &counter
->event_limit
);
743 perf_counter_enable(counter
);
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
);
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 update_context_time(ctx
);
808 regs
= task_pt_regs(task
);
809 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
810 __perf_counter_sched_out(ctx
, cpuctx
);
812 cpuctx
->task_ctx
= NULL
;
815 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
817 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
821 group_sched_in(struct perf_counter
*group_counter
,
822 struct perf_cpu_context
*cpuctx
,
823 struct perf_counter_context
*ctx
,
826 struct perf_counter
*counter
, *partial_group
;
829 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
832 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
834 return ret
< 0 ? ret
: 0;
836 group_counter
->prev_state
= group_counter
->state
;
837 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
841 * Schedule in siblings as one group (if any):
843 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
844 counter
->prev_state
= counter
->state
;
845 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
846 partial_group
= counter
;
855 * Groups can be scheduled in as one unit only, so undo any
856 * partial group before returning:
858 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
859 if (counter
== partial_group
)
861 counter_sched_out(counter
, cpuctx
, ctx
);
863 counter_sched_out(group_counter
, cpuctx
, ctx
);
869 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
870 struct perf_cpu_context
*cpuctx
, int cpu
)
872 struct perf_counter
*counter
;
876 spin_lock(&ctx
->lock
);
878 if (likely(!ctx
->nr_counters
))
881 ctx
->timestamp
= perf_clock();
883 flags
= hw_perf_save_disable();
886 * First go through the list and put on any pinned groups
887 * in order to give them the best chance of going on.
889 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
890 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
891 !counter
->hw_event
.pinned
)
893 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
896 if (group_can_go_on(counter
, cpuctx
, 1))
897 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
900 * If this pinned group hasn't been scheduled,
901 * put it in error state.
903 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
904 update_group_times(counter
);
905 counter
->state
= PERF_COUNTER_STATE_ERROR
;
909 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
911 * Ignore counters in OFF or ERROR state, and
912 * ignore pinned counters since we did them already.
914 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
915 counter
->hw_event
.pinned
)
919 * Listen to the 'cpu' scheduling filter constraint
922 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
925 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
926 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
930 hw_perf_restore(flags
);
932 spin_unlock(&ctx
->lock
);
936 * Called from scheduler to add the counters of the current task
937 * with interrupts disabled.
939 * We restore the counter value and then enable it.
941 * This does not protect us against NMI, but enable()
942 * sets the enabled bit in the control field of counter _before_
943 * accessing the counter control register. If a NMI hits, then it will
944 * keep the counter running.
946 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
948 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
949 struct perf_counter_context
*ctx
= &task
->perf_counter_ctx
;
951 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
952 cpuctx
->task_ctx
= ctx
;
955 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
957 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
959 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
962 int perf_counter_task_disable(void)
964 struct task_struct
*curr
= current
;
965 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
966 struct perf_counter
*counter
;
971 if (likely(!ctx
->nr_counters
))
974 local_irq_save(flags
);
975 cpu
= smp_processor_id();
977 perf_counter_task_sched_out(curr
, cpu
);
979 spin_lock(&ctx
->lock
);
982 * Disable all the counters:
984 perf_flags
= hw_perf_save_disable();
986 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
987 if (counter
->state
!= PERF_COUNTER_STATE_ERROR
) {
988 update_group_times(counter
);
989 counter
->state
= PERF_COUNTER_STATE_OFF
;
993 hw_perf_restore(perf_flags
);
995 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1000 int perf_counter_task_enable(void)
1002 struct task_struct
*curr
= current
;
1003 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
1004 struct perf_counter
*counter
;
1005 unsigned long flags
;
1009 if (likely(!ctx
->nr_counters
))
1012 local_irq_save(flags
);
1013 cpu
= smp_processor_id();
1015 perf_counter_task_sched_out(curr
, cpu
);
1017 spin_lock(&ctx
->lock
);
1020 * Disable all the counters:
1022 perf_flags
= hw_perf_save_disable();
1024 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1025 if (counter
->state
> PERF_COUNTER_STATE_OFF
)
1027 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
1028 counter
->tstamp_enabled
=
1029 ctx
->time
- counter
->total_time_enabled
;
1030 counter
->hw_event
.disabled
= 0;
1032 hw_perf_restore(perf_flags
);
1034 spin_unlock(&ctx
->lock
);
1036 perf_counter_task_sched_in(curr
, cpu
);
1038 local_irq_restore(flags
);
1044 * Round-robin a context's counters:
1046 static void rotate_ctx(struct perf_counter_context
*ctx
)
1048 struct perf_counter
*counter
;
1051 if (!ctx
->nr_counters
)
1054 spin_lock(&ctx
->lock
);
1056 * Rotate the first entry last (works just fine for group counters too):
1058 perf_flags
= hw_perf_save_disable();
1059 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1060 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1063 hw_perf_restore(perf_flags
);
1065 spin_unlock(&ctx
->lock
);
1068 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1070 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1071 struct perf_counter_context
*ctx
= &curr
->perf_counter_ctx
;
1073 perf_counter_cpu_sched_out(cpuctx
);
1074 perf_counter_task_sched_out(curr
, cpu
);
1076 rotate_ctx(&cpuctx
->ctx
);
1079 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1080 perf_counter_task_sched_in(curr
, cpu
);
1084 * Cross CPU call to read the hardware counter
1086 static void __read(void *info
)
1088 struct perf_counter
*counter
= info
;
1089 struct perf_counter_context
*ctx
= counter
->ctx
;
1090 unsigned long flags
;
1092 local_irq_save(flags
);
1094 update_context_time(ctx
);
1095 counter
->pmu
->read(counter
);
1096 update_counter_times(counter
);
1097 local_irq_restore(flags
);
1100 static u64
perf_counter_read(struct perf_counter
*counter
)
1103 * If counter is enabled and currently active on a CPU, update the
1104 * value in the counter structure:
1106 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1107 smp_call_function_single(counter
->oncpu
,
1108 __read
, counter
, 1);
1109 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1110 update_counter_times(counter
);
1113 return atomic64_read(&counter
->count
);
1116 static void put_context(struct perf_counter_context
*ctx
)
1119 put_task_struct(ctx
->task
);
1122 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1124 struct perf_cpu_context
*cpuctx
;
1125 struct perf_counter_context
*ctx
;
1126 struct task_struct
*task
;
1129 * If cpu is not a wildcard then this is a percpu counter:
1132 /* Must be root to operate on a CPU counter: */
1133 if (sysctl_perf_counter_priv
&& !capable(CAP_SYS_ADMIN
))
1134 return ERR_PTR(-EACCES
);
1136 if (cpu
< 0 || cpu
> num_possible_cpus())
1137 return ERR_PTR(-EINVAL
);
1140 * We could be clever and allow to attach a counter to an
1141 * offline CPU and activate it when the CPU comes up, but
1144 if (!cpu_isset(cpu
, cpu_online_map
))
1145 return ERR_PTR(-ENODEV
);
1147 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1157 task
= find_task_by_vpid(pid
);
1159 get_task_struct(task
);
1163 return ERR_PTR(-ESRCH
);
1165 ctx
= &task
->perf_counter_ctx
;
1168 /* Reuse ptrace permission checks for now. */
1169 if (!ptrace_may_access(task
, PTRACE_MODE_READ
)) {
1171 return ERR_PTR(-EACCES
);
1177 static void free_counter_rcu(struct rcu_head
*head
)
1179 struct perf_counter
*counter
;
1181 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1185 static void perf_pending_sync(struct perf_counter
*counter
);
1187 static void free_counter(struct perf_counter
*counter
)
1189 perf_pending_sync(counter
);
1191 if (counter
->hw_event
.mmap
)
1192 atomic_dec(&nr_mmap_tracking
);
1193 if (counter
->hw_event
.munmap
)
1194 atomic_dec(&nr_munmap_tracking
);
1195 if (counter
->hw_event
.comm
)
1196 atomic_dec(&nr_comm_tracking
);
1198 if (counter
->destroy
)
1199 counter
->destroy(counter
);
1201 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1205 * Called when the last reference to the file is gone.
1207 static int perf_release(struct inode
*inode
, struct file
*file
)
1209 struct perf_counter
*counter
= file
->private_data
;
1210 struct perf_counter_context
*ctx
= counter
->ctx
;
1212 file
->private_data
= NULL
;
1214 mutex_lock(&ctx
->mutex
);
1215 mutex_lock(&counter
->mutex
);
1217 perf_counter_remove_from_context(counter
);
1219 mutex_unlock(&counter
->mutex
);
1220 mutex_unlock(&ctx
->mutex
);
1222 free_counter(counter
);
1229 * Read the performance counter - simple non blocking version for now
1232 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1238 * Return end-of-file for a read on a counter that is in
1239 * error state (i.e. because it was pinned but it couldn't be
1240 * scheduled on to the CPU at some point).
1242 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1245 mutex_lock(&counter
->mutex
);
1246 values
[0] = perf_counter_read(counter
);
1248 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1249 values
[n
++] = counter
->total_time_enabled
+
1250 atomic64_read(&counter
->child_total_time_enabled
);
1251 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1252 values
[n
++] = counter
->total_time_running
+
1253 atomic64_read(&counter
->child_total_time_running
);
1254 mutex_unlock(&counter
->mutex
);
1256 if (count
< n
* sizeof(u64
))
1258 count
= n
* sizeof(u64
);
1260 if (copy_to_user(buf
, values
, count
))
1267 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1269 struct perf_counter
*counter
= file
->private_data
;
1271 return perf_read_hw(counter
, buf
, count
);
1274 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1276 struct perf_counter
*counter
= file
->private_data
;
1277 struct perf_mmap_data
*data
;
1278 unsigned int events
= POLL_HUP
;
1281 data
= rcu_dereference(counter
->data
);
1283 events
= atomic_xchg(&data
->poll
, 0);
1286 poll_wait(file
, &counter
->waitq
, wait
);
1291 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1293 struct perf_counter
*counter
= file
->private_data
;
1297 case PERF_COUNTER_IOC_ENABLE
:
1298 perf_counter_enable_family(counter
);
1300 case PERF_COUNTER_IOC_DISABLE
:
1301 perf_counter_disable_family(counter
);
1303 case PERF_COUNTER_IOC_REFRESH
:
1304 perf_counter_refresh(counter
, arg
);
1313 * Callers need to ensure there can be no nesting of this function, otherwise
1314 * the seqlock logic goes bad. We can not serialize this because the arch
1315 * code calls this from NMI context.
1317 void perf_counter_update_userpage(struct perf_counter
*counter
)
1319 struct perf_mmap_data
*data
;
1320 struct perf_counter_mmap_page
*userpg
;
1323 data
= rcu_dereference(counter
->data
);
1327 userpg
= data
->user_page
;
1330 * Disable preemption so as to not let the corresponding user-space
1331 * spin too long if we get preempted.
1336 userpg
->index
= counter
->hw
.idx
;
1337 userpg
->offset
= atomic64_read(&counter
->count
);
1338 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1339 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1348 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1350 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1351 struct perf_mmap_data
*data
;
1352 int ret
= VM_FAULT_SIGBUS
;
1355 data
= rcu_dereference(counter
->data
);
1359 if (vmf
->pgoff
== 0) {
1360 vmf
->page
= virt_to_page(data
->user_page
);
1362 int nr
= vmf
->pgoff
- 1;
1364 if ((unsigned)nr
> data
->nr_pages
)
1367 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1369 get_page(vmf
->page
);
1377 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1379 struct perf_mmap_data
*data
;
1383 WARN_ON(atomic_read(&counter
->mmap_count
));
1385 size
= sizeof(struct perf_mmap_data
);
1386 size
+= nr_pages
* sizeof(void *);
1388 data
= kzalloc(size
, GFP_KERNEL
);
1392 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1393 if (!data
->user_page
)
1394 goto fail_user_page
;
1396 for (i
= 0; i
< nr_pages
; i
++) {
1397 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1398 if (!data
->data_pages
[i
])
1399 goto fail_data_pages
;
1402 data
->nr_pages
= nr_pages
;
1404 rcu_assign_pointer(counter
->data
, data
);
1409 for (i
--; i
>= 0; i
--)
1410 free_page((unsigned long)data
->data_pages
[i
]);
1412 free_page((unsigned long)data
->user_page
);
1421 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1423 struct perf_mmap_data
*data
= container_of(rcu_head
,
1424 struct perf_mmap_data
, rcu_head
);
1427 free_page((unsigned long)data
->user_page
);
1428 for (i
= 0; i
< data
->nr_pages
; i
++)
1429 free_page((unsigned long)data
->data_pages
[i
]);
1433 static void perf_mmap_data_free(struct perf_counter
*counter
)
1435 struct perf_mmap_data
*data
= counter
->data
;
1437 WARN_ON(atomic_read(&counter
->mmap_count
));
1439 rcu_assign_pointer(counter
->data
, NULL
);
1440 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1443 static void perf_mmap_open(struct vm_area_struct
*vma
)
1445 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1447 atomic_inc(&counter
->mmap_count
);
1450 static void perf_mmap_close(struct vm_area_struct
*vma
)
1452 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1454 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
,
1455 &counter
->mmap_mutex
)) {
1456 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_pages
+ 1;
1457 perf_mmap_data_free(counter
);
1458 mutex_unlock(&counter
->mmap_mutex
);
1462 static struct vm_operations_struct perf_mmap_vmops
= {
1463 .open
= perf_mmap_open
,
1464 .close
= perf_mmap_close
,
1465 .fault
= perf_mmap_fault
,
1468 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1470 struct perf_counter
*counter
= file
->private_data
;
1471 unsigned long vma_size
;
1472 unsigned long nr_pages
;
1473 unsigned long locked
, lock_limit
;
1476 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1479 vma_size
= vma
->vm_end
- vma
->vm_start
;
1480 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1483 * If we have data pages ensure they're a power-of-two number, so we
1484 * can do bitmasks instead of modulo.
1486 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1489 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1492 if (vma
->vm_pgoff
!= 0)
1495 mutex_lock(&counter
->mmap_mutex
);
1496 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1497 if (nr_pages
!= counter
->data
->nr_pages
)
1502 locked
= vma
->vm_mm
->locked_vm
;
1503 locked
+= nr_pages
+ 1;
1505 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1506 lock_limit
>>= PAGE_SHIFT
;
1508 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
1513 WARN_ON(counter
->data
);
1514 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1518 atomic_set(&counter
->mmap_count
, 1);
1519 vma
->vm_mm
->locked_vm
+= nr_pages
+ 1;
1521 mutex_unlock(&counter
->mmap_mutex
);
1523 vma
->vm_flags
&= ~VM_MAYWRITE
;
1524 vma
->vm_flags
|= VM_RESERVED
;
1525 vma
->vm_ops
= &perf_mmap_vmops
;
1530 static int perf_fasync(int fd
, struct file
*filp
, int on
)
1532 struct perf_counter
*counter
= filp
->private_data
;
1533 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1536 mutex_lock(&inode
->i_mutex
);
1537 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
1538 mutex_unlock(&inode
->i_mutex
);
1546 static const struct file_operations perf_fops
= {
1547 .release
= perf_release
,
1550 .unlocked_ioctl
= perf_ioctl
,
1551 .compat_ioctl
= perf_ioctl
,
1553 .fasync
= perf_fasync
,
1557 * Perf counter wakeup
1559 * If there's data, ensure we set the poll() state and publish everything
1560 * to user-space before waking everybody up.
1563 void perf_counter_wakeup(struct perf_counter
*counter
)
1565 wake_up_all(&counter
->waitq
);
1567 if (counter
->pending_kill
) {
1568 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
1569 counter
->pending_kill
= 0;
1576 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1578 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1579 * single linked list and use cmpxchg() to add entries lockless.
1582 static void perf_pending_counter(struct perf_pending_entry
*entry
)
1584 struct perf_counter
*counter
= container_of(entry
,
1585 struct perf_counter
, pending
);
1587 if (counter
->pending_disable
) {
1588 counter
->pending_disable
= 0;
1589 perf_counter_disable(counter
);
1592 if (counter
->pending_wakeup
) {
1593 counter
->pending_wakeup
= 0;
1594 perf_counter_wakeup(counter
);
1598 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1600 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
1604 static void perf_pending_queue(struct perf_pending_entry
*entry
,
1605 void (*func
)(struct perf_pending_entry
*))
1607 struct perf_pending_entry
**head
;
1609 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
1614 head
= &get_cpu_var(perf_pending_head
);
1617 entry
->next
= *head
;
1618 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
1620 set_perf_counter_pending();
1622 put_cpu_var(perf_pending_head
);
1625 static int __perf_pending_run(void)
1627 struct perf_pending_entry
*list
;
1630 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
1631 while (list
!= PENDING_TAIL
) {
1632 void (*func
)(struct perf_pending_entry
*);
1633 struct perf_pending_entry
*entry
= list
;
1640 * Ensure we observe the unqueue before we issue the wakeup,
1641 * so that we won't be waiting forever.
1642 * -- see perf_not_pending().
1653 static inline int perf_not_pending(struct perf_counter
*counter
)
1656 * If we flush on whatever cpu we run, there is a chance we don't
1660 __perf_pending_run();
1664 * Ensure we see the proper queue state before going to sleep
1665 * so that we do not miss the wakeup. -- see perf_pending_handle()
1668 return counter
->pending
.next
== NULL
;
1671 static void perf_pending_sync(struct perf_counter
*counter
)
1673 wait_event(counter
->waitq
, perf_not_pending(counter
));
1676 void perf_counter_do_pending(void)
1678 __perf_pending_run();
1682 * Callchain support -- arch specific
1685 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
1694 struct perf_output_handle
{
1695 struct perf_counter
*counter
;
1696 struct perf_mmap_data
*data
;
1697 unsigned int offset
;
1703 unsigned long flags
;
1706 static void perf_output_wakeup(struct perf_output_handle
*handle
)
1708 atomic_set(&handle
->data
->poll
, POLL_IN
);
1711 handle
->counter
->pending_wakeup
= 1;
1712 perf_pending_queue(&handle
->counter
->pending
,
1713 perf_pending_counter
);
1715 perf_counter_wakeup(handle
->counter
);
1719 * Curious locking construct.
1721 * We need to ensure a later event doesn't publish a head when a former
1722 * event isn't done writing. However since we need to deal with NMIs we
1723 * cannot fully serialize things.
1725 * What we do is serialize between CPUs so we only have to deal with NMI
1726 * nesting on a single CPU.
1728 * We only publish the head (and generate a wakeup) when the outer-most
1731 static void perf_output_lock(struct perf_output_handle
*handle
)
1733 struct perf_mmap_data
*data
= handle
->data
;
1738 local_irq_save(handle
->flags
);
1739 cpu
= smp_processor_id();
1741 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
1744 while (atomic_cmpxchg(&data
->lock
, 0, cpu
) != 0)
1750 static void perf_output_unlock(struct perf_output_handle
*handle
)
1752 struct perf_mmap_data
*data
= handle
->data
;
1756 data
->wakeup_head
= data
->head
;
1758 if (!handle
->locked
)
1763 * The xchg implies a full barrier that ensures all writes are done
1764 * before we publish the new head, matched by a rmb() in userspace when
1765 * reading this position.
1767 while ((head
= atomic_xchg(&data
->wakeup_head
, 0))) {
1768 data
->user_page
->data_head
= head
;
1773 * NMI can happen here, which means we can miss a wakeup_head update.
1776 cpu
= atomic_xchg(&data
->lock
, 0);
1777 WARN_ON_ONCE(cpu
!= smp_processor_id());
1780 * Therefore we have to validate we did not indeed do so.
1782 if (unlikely(atomic_read(&data
->wakeup_head
))) {
1784 * Since we had it locked, we can lock it again.
1786 while (atomic_cmpxchg(&data
->lock
, 0, cpu
) != 0)
1793 perf_output_wakeup(handle
);
1795 local_irq_restore(handle
->flags
);
1798 static int perf_output_begin(struct perf_output_handle
*handle
,
1799 struct perf_counter
*counter
, unsigned int size
,
1800 int nmi
, int overflow
)
1802 struct perf_mmap_data
*data
;
1803 unsigned int offset
, head
;
1806 data
= rcu_dereference(counter
->data
);
1810 handle
->data
= data
;
1811 handle
->counter
= counter
;
1813 handle
->overflow
= overflow
;
1815 if (!data
->nr_pages
)
1818 perf_output_lock(handle
);
1821 offset
= head
= atomic_read(&data
->head
);
1823 } while (atomic_cmpxchg(&data
->head
, offset
, head
) != offset
);
1825 handle
->offset
= offset
;
1826 handle
->head
= head
;
1827 handle
->wakeup
= (offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
);
1832 perf_output_wakeup(handle
);
1839 static void perf_output_copy(struct perf_output_handle
*handle
,
1840 void *buf
, unsigned int len
)
1842 unsigned int pages_mask
;
1843 unsigned int offset
;
1847 offset
= handle
->offset
;
1848 pages_mask
= handle
->data
->nr_pages
- 1;
1849 pages
= handle
->data
->data_pages
;
1852 unsigned int page_offset
;
1855 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
1856 page_offset
= offset
& (PAGE_SIZE
- 1);
1857 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
1859 memcpy(pages
[nr
] + page_offset
, buf
, size
);
1866 handle
->offset
= offset
;
1868 WARN_ON_ONCE(handle
->offset
> handle
->head
);
1871 #define perf_output_put(handle, x) \
1872 perf_output_copy((handle), &(x), sizeof(x))
1874 static void perf_output_end(struct perf_output_handle
*handle
)
1876 struct perf_counter
*counter
= handle
->counter
;
1877 struct perf_mmap_data
*data
= handle
->data
;
1879 int wakeup_events
= counter
->hw_event
.wakeup_events
;
1881 if (handle
->overflow
&& wakeup_events
) {
1882 int events
= atomic_inc_return(&data
->events
);
1883 if (events
>= wakeup_events
) {
1884 atomic_sub(wakeup_events
, &data
->events
);
1889 perf_output_unlock(handle
);
1893 static void perf_counter_output(struct perf_counter
*counter
,
1894 int nmi
, struct pt_regs
*regs
, u64 addr
)
1897 u64 record_type
= counter
->hw_event
.record_type
;
1898 struct perf_output_handle handle
;
1899 struct perf_event_header header
;
1908 struct perf_callchain_entry
*callchain
= NULL
;
1909 int callchain_size
= 0;
1913 header
.size
= sizeof(header
);
1915 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
1916 header
.misc
|= user_mode(regs
) ?
1917 PERF_EVENT_MISC_USER
: PERF_EVENT_MISC_KERNEL
;
1919 if (record_type
& PERF_RECORD_IP
) {
1920 ip
= instruction_pointer(regs
);
1921 header
.type
|= PERF_RECORD_IP
;
1922 header
.size
+= sizeof(ip
);
1925 if (record_type
& PERF_RECORD_TID
) {
1926 /* namespace issues */
1927 tid_entry
.pid
= current
->group_leader
->pid
;
1928 tid_entry
.tid
= current
->pid
;
1930 header
.type
|= PERF_RECORD_TID
;
1931 header
.size
+= sizeof(tid_entry
);
1934 if (record_type
& PERF_RECORD_TIME
) {
1936 * Maybe do better on x86 and provide cpu_clock_nmi()
1938 time
= sched_clock();
1940 header
.type
|= PERF_RECORD_TIME
;
1941 header
.size
+= sizeof(u64
);
1944 if (record_type
& PERF_RECORD_ADDR
) {
1945 header
.type
|= PERF_RECORD_ADDR
;
1946 header
.size
+= sizeof(u64
);
1949 if (record_type
& PERF_RECORD_GROUP
) {
1950 header
.type
|= PERF_RECORD_GROUP
;
1951 header
.size
+= sizeof(u64
) +
1952 counter
->nr_siblings
* sizeof(group_entry
);
1955 if (record_type
& PERF_RECORD_CALLCHAIN
) {
1956 callchain
= perf_callchain(regs
);
1959 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
1961 header
.type
|= PERF_RECORD_CALLCHAIN
;
1962 header
.size
+= callchain_size
;
1966 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
1970 perf_output_put(&handle
, header
);
1972 if (record_type
& PERF_RECORD_IP
)
1973 perf_output_put(&handle
, ip
);
1975 if (record_type
& PERF_RECORD_TID
)
1976 perf_output_put(&handle
, tid_entry
);
1978 if (record_type
& PERF_RECORD_TIME
)
1979 perf_output_put(&handle
, time
);
1981 if (record_type
& PERF_RECORD_ADDR
)
1982 perf_output_put(&handle
, addr
);
1984 if (record_type
& PERF_RECORD_GROUP
) {
1985 struct perf_counter
*leader
, *sub
;
1986 u64 nr
= counter
->nr_siblings
;
1988 perf_output_put(&handle
, nr
);
1990 leader
= counter
->group_leader
;
1991 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
1993 sub
->pmu
->read(sub
);
1995 group_entry
.event
= sub
->hw_event
.config
;
1996 group_entry
.counter
= atomic64_read(&sub
->count
);
1998 perf_output_put(&handle
, group_entry
);
2003 perf_output_copy(&handle
, callchain
, callchain_size
);
2005 perf_output_end(&handle
);
2012 struct perf_comm_event
{
2013 struct task_struct
*task
;
2018 struct perf_event_header header
;
2025 static void perf_counter_comm_output(struct perf_counter
*counter
,
2026 struct perf_comm_event
*comm_event
)
2028 struct perf_output_handle handle
;
2029 int size
= comm_event
->event
.header
.size
;
2030 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2035 perf_output_put(&handle
, comm_event
->event
);
2036 perf_output_copy(&handle
, comm_event
->comm
,
2037 comm_event
->comm_size
);
2038 perf_output_end(&handle
);
2041 static int perf_counter_comm_match(struct perf_counter
*counter
,
2042 struct perf_comm_event
*comm_event
)
2044 if (counter
->hw_event
.comm
&&
2045 comm_event
->event
.header
.type
== PERF_EVENT_COMM
)
2051 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2052 struct perf_comm_event
*comm_event
)
2054 struct perf_counter
*counter
;
2056 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2060 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2061 if (perf_counter_comm_match(counter
, comm_event
))
2062 perf_counter_comm_output(counter
, comm_event
);
2067 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2069 struct perf_cpu_context
*cpuctx
;
2071 char *comm
= comm_event
->task
->comm
;
2073 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2075 comm_event
->comm
= comm
;
2076 comm_event
->comm_size
= size
;
2078 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2080 cpuctx
= &get_cpu_var(perf_cpu_context
);
2081 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2082 put_cpu_var(perf_cpu_context
);
2084 perf_counter_comm_ctx(¤t
->perf_counter_ctx
, comm_event
);
2087 void perf_counter_comm(struct task_struct
*task
)
2089 struct perf_comm_event comm_event
;
2091 if (!atomic_read(&nr_comm_tracking
))
2094 comm_event
= (struct perf_comm_event
){
2097 .header
= { .type
= PERF_EVENT_COMM
, },
2098 .pid
= task
->group_leader
->pid
,
2103 perf_counter_comm_event(&comm_event
);
2110 struct perf_mmap_event
{
2116 struct perf_event_header header
;
2126 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2127 struct perf_mmap_event
*mmap_event
)
2129 struct perf_output_handle handle
;
2130 int size
= mmap_event
->event
.header
.size
;
2131 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2136 perf_output_put(&handle
, mmap_event
->event
);
2137 perf_output_copy(&handle
, mmap_event
->file_name
,
2138 mmap_event
->file_size
);
2139 perf_output_end(&handle
);
2142 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2143 struct perf_mmap_event
*mmap_event
)
2145 if (counter
->hw_event
.mmap
&&
2146 mmap_event
->event
.header
.type
== PERF_EVENT_MMAP
)
2149 if (counter
->hw_event
.munmap
&&
2150 mmap_event
->event
.header
.type
== PERF_EVENT_MUNMAP
)
2156 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2157 struct perf_mmap_event
*mmap_event
)
2159 struct perf_counter
*counter
;
2161 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2165 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2166 if (perf_counter_mmap_match(counter
, mmap_event
))
2167 perf_counter_mmap_output(counter
, mmap_event
);
2172 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2174 struct perf_cpu_context
*cpuctx
;
2175 struct file
*file
= mmap_event
->file
;
2182 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2184 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2187 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2189 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2193 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2198 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2200 mmap_event
->file_name
= name
;
2201 mmap_event
->file_size
= size
;
2203 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2205 cpuctx
= &get_cpu_var(perf_cpu_context
);
2206 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2207 put_cpu_var(perf_cpu_context
);
2209 perf_counter_mmap_ctx(¤t
->perf_counter_ctx
, mmap_event
);
2214 void perf_counter_mmap(unsigned long addr
, unsigned long len
,
2215 unsigned long pgoff
, struct file
*file
)
2217 struct perf_mmap_event mmap_event
;
2219 if (!atomic_read(&nr_mmap_tracking
))
2222 mmap_event
= (struct perf_mmap_event
){
2225 .header
= { .type
= PERF_EVENT_MMAP
, },
2226 .pid
= current
->group_leader
->pid
,
2227 .tid
= current
->pid
,
2234 perf_counter_mmap_event(&mmap_event
);
2237 void perf_counter_munmap(unsigned long addr
, unsigned long len
,
2238 unsigned long pgoff
, struct file
*file
)
2240 struct perf_mmap_event mmap_event
;
2242 if (!atomic_read(&nr_munmap_tracking
))
2245 mmap_event
= (struct perf_mmap_event
){
2248 .header
= { .type
= PERF_EVENT_MUNMAP
, },
2249 .pid
= current
->group_leader
->pid
,
2250 .tid
= current
->pid
,
2257 perf_counter_mmap_event(&mmap_event
);
2261 * Generic counter overflow handling.
2264 int perf_counter_overflow(struct perf_counter
*counter
,
2265 int nmi
, struct pt_regs
*regs
, u64 addr
)
2267 int events
= atomic_read(&counter
->event_limit
);
2270 counter
->pending_kill
= POLL_IN
;
2271 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
2273 counter
->pending_kill
= POLL_HUP
;
2275 counter
->pending_disable
= 1;
2276 perf_pending_queue(&counter
->pending
,
2277 perf_pending_counter
);
2279 perf_counter_disable(counter
);
2282 perf_counter_output(counter
, nmi
, regs
, addr
);
2287 * Generic software counter infrastructure
2290 static void perf_swcounter_update(struct perf_counter
*counter
)
2292 struct hw_perf_counter
*hwc
= &counter
->hw
;
2297 prev
= atomic64_read(&hwc
->prev_count
);
2298 now
= atomic64_read(&hwc
->count
);
2299 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2304 atomic64_add(delta
, &counter
->count
);
2305 atomic64_sub(delta
, &hwc
->period_left
);
2308 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2310 struct hw_perf_counter
*hwc
= &counter
->hw
;
2311 s64 left
= atomic64_read(&hwc
->period_left
);
2312 s64 period
= hwc
->irq_period
;
2314 if (unlikely(left
<= -period
)) {
2316 atomic64_set(&hwc
->period_left
, left
);
2319 if (unlikely(left
<= 0)) {
2321 atomic64_add(period
, &hwc
->period_left
);
2324 atomic64_set(&hwc
->prev_count
, -left
);
2325 atomic64_set(&hwc
->count
, -left
);
2328 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2330 enum hrtimer_restart ret
= HRTIMER_RESTART
;
2331 struct perf_counter
*counter
;
2332 struct pt_regs
*regs
;
2334 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
2335 counter
->pmu
->read(counter
);
2337 regs
= get_irq_regs();
2339 * In case we exclude kernel IPs or are somehow not in interrupt
2340 * context, provide the next best thing, the user IP.
2342 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
2343 !counter
->hw_event
.exclude_user
)
2344 regs
= task_pt_regs(current
);
2347 if (perf_counter_overflow(counter
, 0, regs
, 0))
2348 ret
= HRTIMER_NORESTART
;
2351 hrtimer_forward_now(hrtimer
, ns_to_ktime(counter
->hw
.irq_period
));
2356 static void perf_swcounter_overflow(struct perf_counter
*counter
,
2357 int nmi
, struct pt_regs
*regs
, u64 addr
)
2359 perf_swcounter_update(counter
);
2360 perf_swcounter_set_period(counter
);
2361 if (perf_counter_overflow(counter
, nmi
, regs
, addr
))
2362 /* soft-disable the counter */
2367 static int perf_swcounter_match(struct perf_counter
*counter
,
2368 enum perf_event_types type
,
2369 u32 event
, struct pt_regs
*regs
)
2371 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
2374 if (perf_event_raw(&counter
->hw_event
))
2377 if (perf_event_type(&counter
->hw_event
) != type
)
2380 if (perf_event_id(&counter
->hw_event
) != event
)
2383 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
2386 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
2392 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
2393 int nmi
, struct pt_regs
*regs
, u64 addr
)
2395 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
2396 if (counter
->hw
.irq_period
&& !neg
)
2397 perf_swcounter_overflow(counter
, nmi
, regs
, addr
);
2400 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
2401 enum perf_event_types type
, u32 event
,
2402 u64 nr
, int nmi
, struct pt_regs
*regs
,
2405 struct perf_counter
*counter
;
2407 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2411 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2412 if (perf_swcounter_match(counter
, type
, event
, regs
))
2413 perf_swcounter_add(counter
, nr
, nmi
, regs
, addr
);
2418 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
2421 return &cpuctx
->recursion
[3];
2424 return &cpuctx
->recursion
[2];
2427 return &cpuctx
->recursion
[1];
2429 return &cpuctx
->recursion
[0];
2432 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
2433 u64 nr
, int nmi
, struct pt_regs
*regs
,
2436 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
2437 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
2445 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
2446 nr
, nmi
, regs
, addr
);
2447 if (cpuctx
->task_ctx
) {
2448 perf_swcounter_ctx_event(cpuctx
->task_ctx
, type
, event
,
2449 nr
, nmi
, regs
, addr
);
2456 put_cpu_var(perf_cpu_context
);
2460 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
2462 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
, addr
);
2465 static void perf_swcounter_read(struct perf_counter
*counter
)
2467 perf_swcounter_update(counter
);
2470 static int perf_swcounter_enable(struct perf_counter
*counter
)
2472 perf_swcounter_set_period(counter
);
2476 static void perf_swcounter_disable(struct perf_counter
*counter
)
2478 perf_swcounter_update(counter
);
2481 static const struct pmu perf_ops_generic
= {
2482 .enable
= perf_swcounter_enable
,
2483 .disable
= perf_swcounter_disable
,
2484 .read
= perf_swcounter_read
,
2488 * Software counter: cpu wall time clock
2491 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
2493 int cpu
= raw_smp_processor_id();
2497 now
= cpu_clock(cpu
);
2498 prev
= atomic64_read(&counter
->hw
.prev_count
);
2499 atomic64_set(&counter
->hw
.prev_count
, now
);
2500 atomic64_add(now
- prev
, &counter
->count
);
2503 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
2505 struct hw_perf_counter
*hwc
= &counter
->hw
;
2506 int cpu
= raw_smp_processor_id();
2508 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
2509 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2510 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2511 if (hwc
->irq_period
) {
2512 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2513 ns_to_ktime(hwc
->irq_period
), 0,
2514 HRTIMER_MODE_REL
, 0);
2520 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
2522 hrtimer_cancel(&counter
->hw
.hrtimer
);
2523 cpu_clock_perf_counter_update(counter
);
2526 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
2528 cpu_clock_perf_counter_update(counter
);
2531 static const struct pmu perf_ops_cpu_clock
= {
2532 .enable
= cpu_clock_perf_counter_enable
,
2533 .disable
= cpu_clock_perf_counter_disable
,
2534 .read
= cpu_clock_perf_counter_read
,
2538 * Software counter: task time clock
2541 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
2546 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
2548 atomic64_add(delta
, &counter
->count
);
2551 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
2553 struct hw_perf_counter
*hwc
= &counter
->hw
;
2556 now
= counter
->ctx
->time
;
2558 atomic64_set(&hwc
->prev_count
, now
);
2559 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2560 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
2561 if (hwc
->irq_period
) {
2562 __hrtimer_start_range_ns(&hwc
->hrtimer
,
2563 ns_to_ktime(hwc
->irq_period
), 0,
2564 HRTIMER_MODE_REL
, 0);
2570 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
2572 hrtimer_cancel(&counter
->hw
.hrtimer
);
2573 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
2577 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
2582 update_context_time(counter
->ctx
);
2583 time
= counter
->ctx
->time
;
2585 u64 now
= perf_clock();
2586 u64 delta
= now
- counter
->ctx
->timestamp
;
2587 time
= counter
->ctx
->time
+ delta
;
2590 task_clock_perf_counter_update(counter
, time
);
2593 static const struct pmu perf_ops_task_clock
= {
2594 .enable
= task_clock_perf_counter_enable
,
2595 .disable
= task_clock_perf_counter_disable
,
2596 .read
= task_clock_perf_counter_read
,
2600 * Software counter: cpu migrations
2603 static inline u64
get_cpu_migrations(struct perf_counter
*counter
)
2605 struct task_struct
*curr
= counter
->ctx
->task
;
2608 return curr
->se
.nr_migrations
;
2609 return cpu_nr_migrations(smp_processor_id());
2612 static void cpu_migrations_perf_counter_update(struct perf_counter
*counter
)
2617 prev
= atomic64_read(&counter
->hw
.prev_count
);
2618 now
= get_cpu_migrations(counter
);
2620 atomic64_set(&counter
->hw
.prev_count
, now
);
2624 atomic64_add(delta
, &counter
->count
);
2627 static void cpu_migrations_perf_counter_read(struct perf_counter
*counter
)
2629 cpu_migrations_perf_counter_update(counter
);
2632 static int cpu_migrations_perf_counter_enable(struct perf_counter
*counter
)
2634 if (counter
->prev_state
<= PERF_COUNTER_STATE_OFF
)
2635 atomic64_set(&counter
->hw
.prev_count
,
2636 get_cpu_migrations(counter
));
2640 static void cpu_migrations_perf_counter_disable(struct perf_counter
*counter
)
2642 cpu_migrations_perf_counter_update(counter
);
2645 static const struct pmu perf_ops_cpu_migrations
= {
2646 .enable
= cpu_migrations_perf_counter_enable
,
2647 .disable
= cpu_migrations_perf_counter_disable
,
2648 .read
= cpu_migrations_perf_counter_read
,
2651 #ifdef CONFIG_EVENT_PROFILE
2652 void perf_tpcounter_event(int event_id
)
2654 struct pt_regs
*regs
= get_irq_regs();
2657 regs
= task_pt_regs(current
);
2659 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
, 0);
2661 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
2663 extern int ftrace_profile_enable(int);
2664 extern void ftrace_profile_disable(int);
2666 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
2668 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
2671 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
2673 int event_id
= perf_event_id(&counter
->hw_event
);
2676 ret
= ftrace_profile_enable(event_id
);
2680 counter
->destroy
= tp_perf_counter_destroy
;
2681 counter
->hw
.irq_period
= counter
->hw_event
.irq_period
;
2683 return &perf_ops_generic
;
2686 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
2692 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
2694 struct perf_counter_hw_event
*hw_event
= &counter
->hw_event
;
2695 const struct pmu
*pmu
= NULL
;
2696 struct hw_perf_counter
*hwc
= &counter
->hw
;
2699 * Software counters (currently) can't in general distinguish
2700 * between user, kernel and hypervisor events.
2701 * However, context switches and cpu migrations are considered
2702 * to be kernel events, and page faults are never hypervisor
2705 switch (perf_event_id(&counter
->hw_event
)) {
2706 case PERF_COUNT_CPU_CLOCK
:
2707 pmu
= &perf_ops_cpu_clock
;
2709 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2710 hw_event
->irq_period
= 10000;
2712 case PERF_COUNT_TASK_CLOCK
:
2714 * If the user instantiates this as a per-cpu counter,
2715 * use the cpu_clock counter instead.
2717 if (counter
->ctx
->task
)
2718 pmu
= &perf_ops_task_clock
;
2720 pmu
= &perf_ops_cpu_clock
;
2722 if (hw_event
->irq_period
&& hw_event
->irq_period
< 10000)
2723 hw_event
->irq_period
= 10000;
2725 case PERF_COUNT_PAGE_FAULTS
:
2726 case PERF_COUNT_PAGE_FAULTS_MIN
:
2727 case PERF_COUNT_PAGE_FAULTS_MAJ
:
2728 case PERF_COUNT_CONTEXT_SWITCHES
:
2729 pmu
= &perf_ops_generic
;
2731 case PERF_COUNT_CPU_MIGRATIONS
:
2732 if (!counter
->hw_event
.exclude_kernel
)
2733 pmu
= &perf_ops_cpu_migrations
;
2738 hwc
->irq_period
= hw_event
->irq_period
;
2744 * Allocate and initialize a counter structure
2746 static struct perf_counter
*
2747 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
2749 struct perf_counter_context
*ctx
,
2750 struct perf_counter
*group_leader
,
2753 const struct pmu
*pmu
;
2754 struct perf_counter
*counter
;
2757 counter
= kzalloc(sizeof(*counter
), gfpflags
);
2759 return ERR_PTR(-ENOMEM
);
2762 * Single counters are their own group leaders, with an
2763 * empty sibling list:
2766 group_leader
= counter
;
2768 mutex_init(&counter
->mutex
);
2769 INIT_LIST_HEAD(&counter
->list_entry
);
2770 INIT_LIST_HEAD(&counter
->event_entry
);
2771 INIT_LIST_HEAD(&counter
->sibling_list
);
2772 init_waitqueue_head(&counter
->waitq
);
2774 mutex_init(&counter
->mmap_mutex
);
2776 INIT_LIST_HEAD(&counter
->child_list
);
2779 counter
->hw_event
= *hw_event
;
2780 counter
->group_leader
= group_leader
;
2781 counter
->pmu
= NULL
;
2784 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
2785 if (hw_event
->disabled
)
2786 counter
->state
= PERF_COUNTER_STATE_OFF
;
2790 if (perf_event_raw(hw_event
)) {
2791 pmu
= hw_perf_counter_init(counter
);
2795 switch (perf_event_type(hw_event
)) {
2796 case PERF_TYPE_HARDWARE
:
2797 pmu
= hw_perf_counter_init(counter
);
2800 case PERF_TYPE_SOFTWARE
:
2801 pmu
= sw_perf_counter_init(counter
);
2804 case PERF_TYPE_TRACEPOINT
:
2805 pmu
= tp_perf_counter_init(counter
);
2812 else if (IS_ERR(pmu
))
2817 return ERR_PTR(err
);
2822 if (counter
->hw_event
.mmap
)
2823 atomic_inc(&nr_mmap_tracking
);
2824 if (counter
->hw_event
.munmap
)
2825 atomic_inc(&nr_munmap_tracking
);
2826 if (counter
->hw_event
.comm
)
2827 atomic_inc(&nr_comm_tracking
);
2833 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2835 * @hw_event_uptr: event type attributes for monitoring/sampling
2838 * @group_fd: group leader counter fd
2840 SYSCALL_DEFINE5(perf_counter_open
,
2841 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
2842 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
2844 struct perf_counter
*counter
, *group_leader
;
2845 struct perf_counter_hw_event hw_event
;
2846 struct perf_counter_context
*ctx
;
2847 struct file
*counter_file
= NULL
;
2848 struct file
*group_file
= NULL
;
2849 int fput_needed
= 0;
2850 int fput_needed2
= 0;
2853 /* for future expandability... */
2857 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
2861 * Get the target context (task or percpu):
2863 ctx
= find_get_context(pid
, cpu
);
2865 return PTR_ERR(ctx
);
2868 * Look up the group leader (we will attach this counter to it):
2870 group_leader
= NULL
;
2871 if (group_fd
!= -1) {
2873 group_file
= fget_light(group_fd
, &fput_needed
);
2875 goto err_put_context
;
2876 if (group_file
->f_op
!= &perf_fops
)
2877 goto err_put_context
;
2879 group_leader
= group_file
->private_data
;
2881 * Do not allow a recursive hierarchy (this new sibling
2882 * becoming part of another group-sibling):
2884 if (group_leader
->group_leader
!= group_leader
)
2885 goto err_put_context
;
2887 * Do not allow to attach to a group in a different
2888 * task or CPU context:
2890 if (group_leader
->ctx
!= ctx
)
2891 goto err_put_context
;
2893 * Only a group leader can be exclusive or pinned
2895 if (hw_event
.exclusive
|| hw_event
.pinned
)
2896 goto err_put_context
;
2899 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
2901 ret
= PTR_ERR(counter
);
2902 if (IS_ERR(counter
))
2903 goto err_put_context
;
2905 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
2907 goto err_free_put_context
;
2909 counter_file
= fget_light(ret
, &fput_needed2
);
2911 goto err_free_put_context
;
2913 counter
->filp
= counter_file
;
2914 mutex_lock(&ctx
->mutex
);
2915 perf_install_in_context(ctx
, counter
, cpu
);
2916 mutex_unlock(&ctx
->mutex
);
2918 fput_light(counter_file
, fput_needed2
);
2921 fput_light(group_file
, fput_needed
);
2925 err_free_put_context
:
2935 * Initialize the perf_counter context in a task_struct:
2938 __perf_counter_init_context(struct perf_counter_context
*ctx
,
2939 struct task_struct
*task
)
2941 memset(ctx
, 0, sizeof(*ctx
));
2942 spin_lock_init(&ctx
->lock
);
2943 mutex_init(&ctx
->mutex
);
2944 INIT_LIST_HEAD(&ctx
->counter_list
);
2945 INIT_LIST_HEAD(&ctx
->event_list
);
2950 * inherit a counter from parent task to child task:
2952 static struct perf_counter
*
2953 inherit_counter(struct perf_counter
*parent_counter
,
2954 struct task_struct
*parent
,
2955 struct perf_counter_context
*parent_ctx
,
2956 struct task_struct
*child
,
2957 struct perf_counter
*group_leader
,
2958 struct perf_counter_context
*child_ctx
)
2960 struct perf_counter
*child_counter
;
2963 * Instead of creating recursive hierarchies of counters,
2964 * we link inherited counters back to the original parent,
2965 * which has a filp for sure, which we use as the reference
2968 if (parent_counter
->parent
)
2969 parent_counter
= parent_counter
->parent
;
2971 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
2972 parent_counter
->cpu
, child_ctx
,
2973 group_leader
, GFP_KERNEL
);
2974 if (IS_ERR(child_counter
))
2975 return child_counter
;
2978 * Link it up in the child's context:
2980 child_counter
->task
= child
;
2981 add_counter_to_ctx(child_counter
, child_ctx
);
2983 child_counter
->parent
= parent_counter
;
2985 * inherit into child's child as well:
2987 child_counter
->hw_event
.inherit
= 1;
2990 * Get a reference to the parent filp - we will fput it
2991 * when the child counter exits. This is safe to do because
2992 * we are in the parent and we know that the filp still
2993 * exists and has a nonzero count:
2995 atomic_long_inc(&parent_counter
->filp
->f_count
);
2998 * Link this into the parent counter's child list
3000 mutex_lock(&parent_counter
->mutex
);
3001 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3004 * Make the child state follow the state of the parent counter,
3005 * not its hw_event.disabled bit. We hold the parent's mutex,
3006 * so we won't race with perf_counter_{en,dis}able_family.
3008 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3009 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3011 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3013 mutex_unlock(&parent_counter
->mutex
);
3015 return child_counter
;
3018 static int inherit_group(struct perf_counter
*parent_counter
,
3019 struct task_struct
*parent
,
3020 struct perf_counter_context
*parent_ctx
,
3021 struct task_struct
*child
,
3022 struct perf_counter_context
*child_ctx
)
3024 struct perf_counter
*leader
;
3025 struct perf_counter
*sub
;
3026 struct perf_counter
*child_ctr
;
3028 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3029 child
, NULL
, child_ctx
);
3031 return PTR_ERR(leader
);
3032 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3033 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3034 child
, leader
, child_ctx
);
3035 if (IS_ERR(child_ctr
))
3036 return PTR_ERR(child_ctr
);
3041 static void sync_child_counter(struct perf_counter
*child_counter
,
3042 struct perf_counter
*parent_counter
)
3044 u64 parent_val
, child_val
;
3046 parent_val
= atomic64_read(&parent_counter
->count
);
3047 child_val
= atomic64_read(&child_counter
->count
);
3050 * Add back the child's count to the parent's count:
3052 atomic64_add(child_val
, &parent_counter
->count
);
3053 atomic64_add(child_counter
->total_time_enabled
,
3054 &parent_counter
->child_total_time_enabled
);
3055 atomic64_add(child_counter
->total_time_running
,
3056 &parent_counter
->child_total_time_running
);
3059 * Remove this counter from the parent's list
3061 mutex_lock(&parent_counter
->mutex
);
3062 list_del_init(&child_counter
->child_list
);
3063 mutex_unlock(&parent_counter
->mutex
);
3066 * Release the parent counter, if this was the last
3069 fput(parent_counter
->filp
);
3073 __perf_counter_exit_task(struct task_struct
*child
,
3074 struct perf_counter
*child_counter
,
3075 struct perf_counter_context
*child_ctx
)
3077 struct perf_counter
*parent_counter
;
3078 struct perf_counter
*sub
, *tmp
;
3081 * If we do not self-reap then we have to wait for the
3082 * child task to unschedule (it will happen for sure),
3083 * so that its counter is at its final count. (This
3084 * condition triggers rarely - child tasks usually get
3085 * off their CPU before the parent has a chance to
3086 * get this far into the reaping action)
3088 if (child
!= current
) {
3089 wait_task_inactive(child
, 0);
3090 list_del_init(&child_counter
->list_entry
);
3091 update_counter_times(child_counter
);
3093 struct perf_cpu_context
*cpuctx
;
3094 unsigned long flags
;
3098 * Disable and unlink this counter.
3100 * Be careful about zapping the list - IRQ/NMI context
3101 * could still be processing it:
3103 local_irq_save(flags
);
3104 perf_flags
= hw_perf_save_disable();
3106 cpuctx
= &__get_cpu_var(perf_cpu_context
);
3108 group_sched_out(child_counter
, cpuctx
, child_ctx
);
3109 update_counter_times(child_counter
);
3111 list_del_init(&child_counter
->list_entry
);
3113 child_ctx
->nr_counters
--;
3115 hw_perf_restore(perf_flags
);
3116 local_irq_restore(flags
);
3119 parent_counter
= child_counter
->parent
;
3121 * It can happen that parent exits first, and has counters
3122 * that are still around due to the child reference. These
3123 * counters need to be zapped - but otherwise linger.
3125 if (parent_counter
) {
3126 sync_child_counter(child_counter
, parent_counter
);
3127 list_for_each_entry_safe(sub
, tmp
, &child_counter
->sibling_list
,
3130 sync_child_counter(sub
, sub
->parent
);
3134 free_counter(child_counter
);
3139 * When a child task exits, feed back counter values to parent counters.
3141 * Note: we may be running in child context, but the PID is not hashed
3142 * anymore so new counters will not be added.
3144 void perf_counter_exit_task(struct task_struct
*child
)
3146 struct perf_counter
*child_counter
, *tmp
;
3147 struct perf_counter_context
*child_ctx
;
3149 child_ctx
= &child
->perf_counter_ctx
;
3151 if (likely(!child_ctx
->nr_counters
))
3154 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
3156 __perf_counter_exit_task(child
, child_counter
, child_ctx
);
3160 * Initialize the perf_counter context in task_struct
3162 void perf_counter_init_task(struct task_struct
*child
)
3164 struct perf_counter_context
*child_ctx
, *parent_ctx
;
3165 struct perf_counter
*counter
;
3166 struct task_struct
*parent
= current
;
3168 child_ctx
= &child
->perf_counter_ctx
;
3169 parent_ctx
= &parent
->perf_counter_ctx
;
3171 __perf_counter_init_context(child_ctx
, child
);
3174 * This is executed from the parent task context, so inherit
3175 * counters that have been marked for cloning:
3178 if (likely(!parent_ctx
->nr_counters
))
3182 * Lock the parent list. No need to lock the child - not PID
3183 * hashed yet and not running, so nobody can access it.
3185 mutex_lock(&parent_ctx
->mutex
);
3188 * We dont have to disable NMIs - we are only looking at
3189 * the list, not manipulating it:
3191 list_for_each_entry(counter
, &parent_ctx
->counter_list
, list_entry
) {
3192 if (!counter
->hw_event
.inherit
)
3195 if (inherit_group(counter
, parent
,
3196 parent_ctx
, child
, child_ctx
))
3200 mutex_unlock(&parent_ctx
->mutex
);
3203 static void __cpuinit
perf_counter_init_cpu(int cpu
)
3205 struct perf_cpu_context
*cpuctx
;
3207 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3208 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
3210 mutex_lock(&perf_resource_mutex
);
3211 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
3212 mutex_unlock(&perf_resource_mutex
);
3214 hw_perf_counter_setup(cpu
);
3217 #ifdef CONFIG_HOTPLUG_CPU
3218 static void __perf_counter_exit_cpu(void *info
)
3220 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3221 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3222 struct perf_counter
*counter
, *tmp
;
3224 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
3225 __perf_counter_remove_from_context(counter
);
3227 static void perf_counter_exit_cpu(int cpu
)
3229 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3230 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3232 mutex_lock(&ctx
->mutex
);
3233 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
3234 mutex_unlock(&ctx
->mutex
);
3237 static inline void perf_counter_exit_cpu(int cpu
) { }
3240 static int __cpuinit
3241 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
3243 unsigned int cpu
= (long)hcpu
;
3247 case CPU_UP_PREPARE
:
3248 case CPU_UP_PREPARE_FROZEN
:
3249 perf_counter_init_cpu(cpu
);
3252 case CPU_DOWN_PREPARE
:
3253 case CPU_DOWN_PREPARE_FROZEN
:
3254 perf_counter_exit_cpu(cpu
);
3264 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
3265 .notifier_call
= perf_cpu_notify
,
3268 static int __init
perf_counter_init(void)
3270 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
3271 (void *)(long)smp_processor_id());
3272 register_cpu_notifier(&perf_cpu_nb
);
3276 early_initcall(perf_counter_init
);
3278 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
3280 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
3284 perf_set_reserve_percpu(struct sysdev_class
*class,
3288 struct perf_cpu_context
*cpuctx
;
3292 err
= strict_strtoul(buf
, 10, &val
);
3295 if (val
> perf_max_counters
)
3298 mutex_lock(&perf_resource_mutex
);
3299 perf_reserved_percpu
= val
;
3300 for_each_online_cpu(cpu
) {
3301 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3302 spin_lock_irq(&cpuctx
->ctx
.lock
);
3303 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
3304 perf_max_counters
- perf_reserved_percpu
);
3305 cpuctx
->max_pertask
= mpt
;
3306 spin_unlock_irq(&cpuctx
->ctx
.lock
);
3308 mutex_unlock(&perf_resource_mutex
);
3313 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
3315 return sprintf(buf
, "%d\n", perf_overcommit
);
3319 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
3324 err
= strict_strtoul(buf
, 10, &val
);
3330 mutex_lock(&perf_resource_mutex
);
3331 perf_overcommit
= val
;
3332 mutex_unlock(&perf_resource_mutex
);
3337 static SYSDEV_CLASS_ATTR(
3340 perf_show_reserve_percpu
,
3341 perf_set_reserve_percpu
3344 static SYSDEV_CLASS_ATTR(
3347 perf_show_overcommit
,
3351 static struct attribute
*perfclass_attrs
[] = {
3352 &attr_reserve_percpu
.attr
,
3353 &attr_overcommit
.attr
,
3357 static struct attribute_group perfclass_attr_group
= {
3358 .attrs
= perfclass_attrs
,
3359 .name
= "perf_counters",
3362 static int __init
perf_counter_sysfs_init(void)
3364 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
3365 &perfclass_attr_group
);
3367 device_initcall(perf_counter_sysfs_init
);