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/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context
, perf_cpu_context
);
38 int perf_max_counters __read_mostly
= 1;
39 static int perf_reserved_percpu __read_mostly
;
40 static int perf_overcommit __read_mostly
= 1;
42 static atomic_t nr_counters __read_mostly
;
43 static atomic_t nr_mmap_counters __read_mostly
;
44 static atomic_t nr_comm_counters __read_mostly
;
47 * perf counter paranoia level:
49 * 1 - disallow cpu counters to unpriv
50 * 2 - disallow kernel profiling to unpriv
52 int sysctl_perf_counter_paranoid __read_mostly
;
54 static inline bool perf_paranoid_cpu(void)
56 return sysctl_perf_counter_paranoid
> 0;
59 static inline bool perf_paranoid_kernel(void)
61 return sysctl_perf_counter_paranoid
> 1;
64 int sysctl_perf_counter_mlock __read_mostly
= 512; /* 'free' kb per user */
67 * max perf counter sample rate
69 int sysctl_perf_counter_sample_rate __read_mostly
= 100000;
71 static atomic64_t perf_counter_id
;
74 * Lock for (sysadmin-configurable) counter reservations:
76 static DEFINE_SPINLOCK(perf_resource_lock
);
79 * Architecture provided APIs - weak aliases:
81 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
86 void __weak
hw_perf_disable(void) { barrier(); }
87 void __weak
hw_perf_enable(void) { barrier(); }
89 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
92 hw_perf_group_sched_in(struct perf_counter
*group_leader
,
93 struct perf_cpu_context
*cpuctx
,
94 struct perf_counter_context
*ctx
, int cpu
)
99 void __weak
perf_counter_print_debug(void) { }
101 static DEFINE_PER_CPU(int, disable_count
);
103 void __perf_disable(void)
105 __get_cpu_var(disable_count
)++;
108 bool __perf_enable(void)
110 return !--__get_cpu_var(disable_count
);
113 void perf_disable(void)
119 void perf_enable(void)
125 static void get_ctx(struct perf_counter_context
*ctx
)
127 WARN_ON(!atomic_inc_not_zero(&ctx
->refcount
));
130 static void free_ctx(struct rcu_head
*head
)
132 struct perf_counter_context
*ctx
;
134 ctx
= container_of(head
, struct perf_counter_context
, rcu_head
);
138 static void put_ctx(struct perf_counter_context
*ctx
)
140 if (atomic_dec_and_test(&ctx
->refcount
)) {
142 put_ctx(ctx
->parent_ctx
);
144 put_task_struct(ctx
->task
);
145 call_rcu(&ctx
->rcu_head
, free_ctx
);
150 * Get the perf_counter_context for a task and lock it.
151 * This has to cope with with the fact that until it is locked,
152 * the context could get moved to another task.
154 static struct perf_counter_context
*
155 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
157 struct perf_counter_context
*ctx
;
161 ctx
= rcu_dereference(task
->perf_counter_ctxp
);
164 * If this context is a clone of another, it might
165 * get swapped for another underneath us by
166 * perf_counter_task_sched_out, though the
167 * rcu_read_lock() protects us from any context
168 * getting freed. Lock the context and check if it
169 * got swapped before we could get the lock, and retry
170 * if so. If we locked the right context, then it
171 * can't get swapped on us any more.
173 spin_lock_irqsave(&ctx
->lock
, *flags
);
174 if (ctx
!= rcu_dereference(task
->perf_counter_ctxp
)) {
175 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
184 * Get the context for a task and increment its pin_count so it
185 * can't get swapped to another task. This also increments its
186 * reference count so that the context can't get freed.
188 static struct perf_counter_context
*perf_pin_task_context(struct task_struct
*task
)
190 struct perf_counter_context
*ctx
;
193 ctx
= perf_lock_task_context(task
, &flags
);
197 spin_unlock_irqrestore(&ctx
->lock
, flags
);
202 static void perf_unpin_context(struct perf_counter_context
*ctx
)
206 spin_lock_irqsave(&ctx
->lock
, flags
);
208 spin_unlock_irqrestore(&ctx
->lock
, flags
);
213 * Add a counter from the lists for its context.
214 * Must be called with ctx->mutex and ctx->lock held.
217 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
219 struct perf_counter
*group_leader
= counter
->group_leader
;
222 * Depending on whether it is a standalone or sibling counter,
223 * add it straight to the context's counter list, or to the group
224 * leader's sibling list:
226 if (group_leader
== counter
)
227 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
229 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
230 group_leader
->nr_siblings
++;
233 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
238 * Remove a counter from the lists for its context.
239 * Must be called with ctx->mutex and ctx->lock held.
242 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
244 struct perf_counter
*sibling
, *tmp
;
246 if (list_empty(&counter
->list_entry
))
250 list_del_init(&counter
->list_entry
);
251 list_del_rcu(&counter
->event_entry
);
253 if (counter
->group_leader
!= counter
)
254 counter
->group_leader
->nr_siblings
--;
257 * If this was a group counter with sibling counters then
258 * upgrade the siblings to singleton counters by adding them
259 * to the context list directly:
261 list_for_each_entry_safe(sibling
, tmp
,
262 &counter
->sibling_list
, list_entry
) {
264 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
265 sibling
->group_leader
= sibling
;
270 counter_sched_out(struct perf_counter
*counter
,
271 struct perf_cpu_context
*cpuctx
,
272 struct perf_counter_context
*ctx
)
274 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
277 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
278 counter
->tstamp_stopped
= ctx
->time
;
279 counter
->pmu
->disable(counter
);
282 if (!is_software_counter(counter
))
283 cpuctx
->active_oncpu
--;
285 if (counter
->attr
.exclusive
|| !cpuctx
->active_oncpu
)
286 cpuctx
->exclusive
= 0;
290 group_sched_out(struct perf_counter
*group_counter
,
291 struct perf_cpu_context
*cpuctx
,
292 struct perf_counter_context
*ctx
)
294 struct perf_counter
*counter
;
296 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
299 counter_sched_out(group_counter
, cpuctx
, ctx
);
302 * Schedule out siblings (if any):
304 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
305 counter_sched_out(counter
, cpuctx
, ctx
);
307 if (group_counter
->attr
.exclusive
)
308 cpuctx
->exclusive
= 0;
312 * Cross CPU call to remove a performance counter
314 * We disable the counter on the hardware level first. After that we
315 * remove it from the context list.
317 static void __perf_counter_remove_from_context(void *info
)
319 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
320 struct perf_counter
*counter
= info
;
321 struct perf_counter_context
*ctx
= counter
->ctx
;
324 * If this is a task context, we need to check whether it is
325 * the current task context of this cpu. If not it has been
326 * scheduled out before the smp call arrived.
328 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
331 spin_lock(&ctx
->lock
);
333 * Protect the list operation against NMI by disabling the
334 * counters on a global level.
338 counter_sched_out(counter
, cpuctx
, ctx
);
340 list_del_counter(counter
, ctx
);
344 * Allow more per task counters with respect to the
347 cpuctx
->max_pertask
=
348 min(perf_max_counters
- ctx
->nr_counters
,
349 perf_max_counters
- perf_reserved_percpu
);
353 spin_unlock(&ctx
->lock
);
358 * Remove the counter from a task's (or a CPU's) list of counters.
360 * Must be called with ctx->mutex held.
362 * CPU counters are removed with a smp call. For task counters we only
363 * call when the task is on a CPU.
365 * If counter->ctx is a cloned context, callers must make sure that
366 * every task struct that counter->ctx->task could possibly point to
367 * remains valid. This is OK when called from perf_release since
368 * that only calls us on the top-level context, which can't be a clone.
369 * When called from perf_counter_exit_task, it's OK because the
370 * context has been detached from its task.
372 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
374 struct perf_counter_context
*ctx
= counter
->ctx
;
375 struct task_struct
*task
= ctx
->task
;
379 * Per cpu counters are removed via an smp call and
380 * the removal is always sucessful.
382 smp_call_function_single(counter
->cpu
,
383 __perf_counter_remove_from_context
,
389 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
392 spin_lock_irq(&ctx
->lock
);
394 * If the context is active we need to retry the smp call.
396 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
397 spin_unlock_irq(&ctx
->lock
);
402 * The lock prevents that this context is scheduled in so we
403 * can remove the counter safely, if the call above did not
406 if (!list_empty(&counter
->list_entry
)) {
407 list_del_counter(counter
, ctx
);
409 spin_unlock_irq(&ctx
->lock
);
412 static inline u64
perf_clock(void)
414 return cpu_clock(smp_processor_id());
418 * Update the record of the current time in a context.
420 static void update_context_time(struct perf_counter_context
*ctx
)
422 u64 now
= perf_clock();
424 ctx
->time
+= now
- ctx
->timestamp
;
425 ctx
->timestamp
= now
;
429 * Update the total_time_enabled and total_time_running fields for a counter.
431 static void update_counter_times(struct perf_counter
*counter
)
433 struct perf_counter_context
*ctx
= counter
->ctx
;
436 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
439 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
441 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
442 run_end
= counter
->tstamp_stopped
;
446 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
450 * Update total_time_enabled and total_time_running for all counters in a group.
452 static void update_group_times(struct perf_counter
*leader
)
454 struct perf_counter
*counter
;
456 update_counter_times(leader
);
457 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
458 update_counter_times(counter
);
462 * Cross CPU call to disable a performance counter
464 static void __perf_counter_disable(void *info
)
466 struct perf_counter
*counter
= info
;
467 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
468 struct perf_counter_context
*ctx
= counter
->ctx
;
471 * If this is a per-task counter, need to check whether this
472 * counter's task is the current task on this cpu.
474 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
477 spin_lock(&ctx
->lock
);
480 * If the counter is on, turn it off.
481 * If it is in error state, leave it in error state.
483 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
484 update_context_time(ctx
);
485 update_counter_times(counter
);
486 if (counter
== counter
->group_leader
)
487 group_sched_out(counter
, cpuctx
, ctx
);
489 counter_sched_out(counter
, cpuctx
, ctx
);
490 counter
->state
= PERF_COUNTER_STATE_OFF
;
493 spin_unlock(&ctx
->lock
);
499 * If counter->ctx is a cloned context, callers must make sure that
500 * every task struct that counter->ctx->task could possibly point to
501 * remains valid. This condition is satisifed when called through
502 * perf_counter_for_each_child or perf_counter_for_each because they
503 * hold the top-level counter's child_mutex, so any descendant that
504 * goes to exit will block in sync_child_counter.
505 * When called from perf_pending_counter it's OK because counter->ctx
506 * is the current context on this CPU and preemption is disabled,
507 * hence we can't get into perf_counter_task_sched_out for this context.
509 static void perf_counter_disable(struct perf_counter
*counter
)
511 struct perf_counter_context
*ctx
= counter
->ctx
;
512 struct task_struct
*task
= ctx
->task
;
516 * Disable the counter on the cpu that it's on
518 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
524 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
526 spin_lock_irq(&ctx
->lock
);
528 * If the counter is still active, we need to retry the cross-call.
530 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
531 spin_unlock_irq(&ctx
->lock
);
536 * Since we have the lock this context can't be scheduled
537 * in, so we can change the state safely.
539 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
540 update_counter_times(counter
);
541 counter
->state
= PERF_COUNTER_STATE_OFF
;
544 spin_unlock_irq(&ctx
->lock
);
548 counter_sched_in(struct perf_counter
*counter
,
549 struct perf_cpu_context
*cpuctx
,
550 struct perf_counter_context
*ctx
,
553 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
556 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
557 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
559 * The new state must be visible before we turn it on in the hardware:
563 if (counter
->pmu
->enable(counter
)) {
564 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
569 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
571 if (!is_software_counter(counter
))
572 cpuctx
->active_oncpu
++;
575 if (counter
->attr
.exclusive
)
576 cpuctx
->exclusive
= 1;
582 group_sched_in(struct perf_counter
*group_counter
,
583 struct perf_cpu_context
*cpuctx
,
584 struct perf_counter_context
*ctx
,
587 struct perf_counter
*counter
, *partial_group
;
590 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
593 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
595 return ret
< 0 ? ret
: 0;
597 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
601 * Schedule in siblings as one group (if any):
603 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
604 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
605 partial_group
= counter
;
614 * Groups can be scheduled in as one unit only, so undo any
615 * partial group before returning:
617 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
618 if (counter
== partial_group
)
620 counter_sched_out(counter
, cpuctx
, ctx
);
622 counter_sched_out(group_counter
, cpuctx
, ctx
);
628 * Return 1 for a group consisting entirely of software counters,
629 * 0 if the group contains any hardware counters.
631 static int is_software_only_group(struct perf_counter
*leader
)
633 struct perf_counter
*counter
;
635 if (!is_software_counter(leader
))
638 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
639 if (!is_software_counter(counter
))
646 * Work out whether we can put this counter group on the CPU now.
648 static int group_can_go_on(struct perf_counter
*counter
,
649 struct perf_cpu_context
*cpuctx
,
653 * Groups consisting entirely of software counters can always go on.
655 if (is_software_only_group(counter
))
658 * If an exclusive group is already on, no other hardware
659 * counters can go on.
661 if (cpuctx
->exclusive
)
664 * If this group is exclusive and there are already
665 * counters on the CPU, it can't go on.
667 if (counter
->attr
.exclusive
&& cpuctx
->active_oncpu
)
670 * Otherwise, try to add it if all previous groups were able
676 static void add_counter_to_ctx(struct perf_counter
*counter
,
677 struct perf_counter_context
*ctx
)
679 list_add_counter(counter
, ctx
);
680 counter
->tstamp_enabled
= ctx
->time
;
681 counter
->tstamp_running
= ctx
->time
;
682 counter
->tstamp_stopped
= ctx
->time
;
686 * Cross CPU call to install and enable a performance counter
688 * Must be called with ctx->mutex held
690 static void __perf_install_in_context(void *info
)
692 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
693 struct perf_counter
*counter
= info
;
694 struct perf_counter_context
*ctx
= counter
->ctx
;
695 struct perf_counter
*leader
= counter
->group_leader
;
696 int cpu
= smp_processor_id();
700 * If this is a task context, we need to check whether it is
701 * the current task context of this cpu. If not it has been
702 * scheduled out before the smp call arrived.
703 * Or possibly this is the right context but it isn't
704 * on this cpu because it had no counters.
706 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
707 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
709 cpuctx
->task_ctx
= ctx
;
712 spin_lock(&ctx
->lock
);
714 update_context_time(ctx
);
717 * Protect the list operation against NMI by disabling the
718 * counters on a global level. NOP for non NMI based counters.
722 add_counter_to_ctx(counter
, ctx
);
725 * Don't put the counter on if it is disabled or if
726 * it is in a group and the group isn't on.
728 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
729 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
733 * An exclusive counter can't go on if there are already active
734 * hardware counters, and no hardware counter can go on if there
735 * is already an exclusive counter on.
737 if (!group_can_go_on(counter
, cpuctx
, 1))
740 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
744 * This counter couldn't go on. If it is in a group
745 * then we have to pull the whole group off.
746 * If the counter group is pinned then put it in error state.
748 if (leader
!= counter
)
749 group_sched_out(leader
, cpuctx
, ctx
);
750 if (leader
->attr
.pinned
) {
751 update_group_times(leader
);
752 leader
->state
= PERF_COUNTER_STATE_ERROR
;
756 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
757 cpuctx
->max_pertask
--;
762 spin_unlock(&ctx
->lock
);
766 * Attach a performance counter to a context
768 * First we add the counter to the list with the hardware enable bit
769 * in counter->hw_config cleared.
771 * If the counter is attached to a task which is on a CPU we use a smp
772 * call to enable it in the task context. The task might have been
773 * scheduled away, but we check this in the smp call again.
775 * Must be called with ctx->mutex held.
778 perf_install_in_context(struct perf_counter_context
*ctx
,
779 struct perf_counter
*counter
,
782 struct task_struct
*task
= ctx
->task
;
786 * Per cpu counters are installed via an smp call and
787 * the install is always sucessful.
789 smp_call_function_single(cpu
, __perf_install_in_context
,
795 task_oncpu_function_call(task
, __perf_install_in_context
,
798 spin_lock_irq(&ctx
->lock
);
800 * we need to retry the smp call.
802 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
803 spin_unlock_irq(&ctx
->lock
);
808 * The lock prevents that this context is scheduled in so we
809 * can add the counter safely, if it the call above did not
812 if (list_empty(&counter
->list_entry
))
813 add_counter_to_ctx(counter
, ctx
);
814 spin_unlock_irq(&ctx
->lock
);
818 * Cross CPU call to enable a performance counter
820 static void __perf_counter_enable(void *info
)
822 struct perf_counter
*counter
= info
;
823 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
824 struct perf_counter_context
*ctx
= counter
->ctx
;
825 struct perf_counter
*leader
= counter
->group_leader
;
829 * If this is a per-task counter, need to check whether this
830 * counter's task is the current task on this cpu.
832 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
833 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
835 cpuctx
->task_ctx
= ctx
;
838 spin_lock(&ctx
->lock
);
840 update_context_time(ctx
);
842 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
844 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
845 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
848 * If the counter is in a group and isn't the group leader,
849 * then don't put it on unless the group is on.
851 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
854 if (!group_can_go_on(counter
, cpuctx
, 1)) {
858 if (counter
== leader
)
859 err
= group_sched_in(counter
, cpuctx
, ctx
,
862 err
= counter_sched_in(counter
, cpuctx
, ctx
,
869 * If this counter can't go on and it's part of a
870 * group, then the whole group has to come off.
872 if (leader
!= counter
)
873 group_sched_out(leader
, cpuctx
, ctx
);
874 if (leader
->attr
.pinned
) {
875 update_group_times(leader
);
876 leader
->state
= PERF_COUNTER_STATE_ERROR
;
881 spin_unlock(&ctx
->lock
);
887 * If counter->ctx is a cloned context, callers must make sure that
888 * every task struct that counter->ctx->task could possibly point to
889 * remains valid. This condition is satisfied when called through
890 * perf_counter_for_each_child or perf_counter_for_each as described
891 * for perf_counter_disable.
893 static void perf_counter_enable(struct perf_counter
*counter
)
895 struct perf_counter_context
*ctx
= counter
->ctx
;
896 struct task_struct
*task
= ctx
->task
;
900 * Enable the counter on the cpu that it's on
902 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
907 spin_lock_irq(&ctx
->lock
);
908 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
912 * If the counter is in error state, clear that first.
913 * That way, if we see the counter in error state below, we
914 * know that it has gone back into error state, as distinct
915 * from the task having been scheduled away before the
916 * cross-call arrived.
918 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
919 counter
->state
= PERF_COUNTER_STATE_OFF
;
922 spin_unlock_irq(&ctx
->lock
);
923 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
925 spin_lock_irq(&ctx
->lock
);
928 * If the context is active and the counter is still off,
929 * we need to retry the cross-call.
931 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
935 * Since we have the lock this context can't be scheduled
936 * in, so we can change the state safely.
938 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
939 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
940 counter
->tstamp_enabled
=
941 ctx
->time
- counter
->total_time_enabled
;
944 spin_unlock_irq(&ctx
->lock
);
947 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
950 * not supported on inherited counters
952 if (counter
->attr
.inherit
)
955 atomic_add(refresh
, &counter
->event_limit
);
956 perf_counter_enable(counter
);
961 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
962 struct perf_cpu_context
*cpuctx
)
964 struct perf_counter
*counter
;
966 spin_lock(&ctx
->lock
);
968 if (likely(!ctx
->nr_counters
))
970 update_context_time(ctx
);
973 if (ctx
->nr_active
) {
974 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
975 if (counter
!= counter
->group_leader
)
976 counter_sched_out(counter
, cpuctx
, ctx
);
978 group_sched_out(counter
, cpuctx
, ctx
);
983 spin_unlock(&ctx
->lock
);
987 * Test whether two contexts are equivalent, i.e. whether they
988 * have both been cloned from the same version of the same context
989 * and they both have the same number of enabled counters.
990 * If the number of enabled counters is the same, then the set
991 * of enabled counters should be the same, because these are both
992 * inherited contexts, therefore we can't access individual counters
993 * in them directly with an fd; we can only enable/disable all
994 * counters via prctl, or enable/disable all counters in a family
995 * via ioctl, which will have the same effect on both contexts.
997 static int context_equiv(struct perf_counter_context
*ctx1
,
998 struct perf_counter_context
*ctx2
)
1000 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
1001 && ctx1
->parent_gen
== ctx2
->parent_gen
1002 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
1006 * Called from scheduler to remove the counters of the current task,
1007 * with interrupts disabled.
1009 * We stop each counter and update the counter value in counter->count.
1011 * This does not protect us against NMI, but disable()
1012 * sets the disabled bit in the control field of counter _before_
1013 * accessing the counter control register. If a NMI hits, then it will
1014 * not restart the counter.
1016 void perf_counter_task_sched_out(struct task_struct
*task
,
1017 struct task_struct
*next
, int cpu
)
1019 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1020 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1021 struct perf_counter_context
*next_ctx
;
1022 struct perf_counter_context
*parent
;
1023 struct pt_regs
*regs
;
1026 regs
= task_pt_regs(task
);
1027 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1029 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1032 update_context_time(ctx
);
1035 parent
= rcu_dereference(ctx
->parent_ctx
);
1036 next_ctx
= next
->perf_counter_ctxp
;
1037 if (parent
&& next_ctx
&&
1038 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1040 * Looks like the two contexts are clones, so we might be
1041 * able to optimize the context switch. We lock both
1042 * contexts and check that they are clones under the
1043 * lock (including re-checking that neither has been
1044 * uncloned in the meantime). It doesn't matter which
1045 * order we take the locks because no other cpu could
1046 * be trying to lock both of these tasks.
1048 spin_lock(&ctx
->lock
);
1049 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1050 if (context_equiv(ctx
, next_ctx
)) {
1052 * XXX do we need a memory barrier of sorts
1053 * wrt to rcu_dereference() of perf_counter_ctxp
1055 task
->perf_counter_ctxp
= next_ctx
;
1056 next
->perf_counter_ctxp
= ctx
;
1058 next_ctx
->task
= task
;
1061 spin_unlock(&next_ctx
->lock
);
1062 spin_unlock(&ctx
->lock
);
1067 __perf_counter_sched_out(ctx
, cpuctx
);
1068 cpuctx
->task_ctx
= NULL
;
1073 * Called with IRQs disabled
1075 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
1077 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1079 if (!cpuctx
->task_ctx
)
1082 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1085 __perf_counter_sched_out(ctx
, cpuctx
);
1086 cpuctx
->task_ctx
= NULL
;
1090 * Called with IRQs disabled
1092 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1094 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
1098 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
1099 struct perf_cpu_context
*cpuctx
, int cpu
)
1101 struct perf_counter
*counter
;
1104 spin_lock(&ctx
->lock
);
1106 if (likely(!ctx
->nr_counters
))
1109 ctx
->timestamp
= perf_clock();
1114 * First go through the list and put on any pinned groups
1115 * in order to give them the best chance of going on.
1117 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1118 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1119 !counter
->attr
.pinned
)
1121 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1124 if (counter
!= counter
->group_leader
)
1125 counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
1127 if (group_can_go_on(counter
, cpuctx
, 1))
1128 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
1132 * If this pinned group hasn't been scheduled,
1133 * put it in error state.
1135 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1136 update_group_times(counter
);
1137 counter
->state
= PERF_COUNTER_STATE_ERROR
;
1141 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1143 * Ignore counters in OFF or ERROR state, and
1144 * ignore pinned counters since we did them already.
1146 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1147 counter
->attr
.pinned
)
1151 * Listen to the 'cpu' scheduling filter constraint
1154 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1157 if (counter
!= counter
->group_leader
) {
1158 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
))
1161 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
1162 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
1169 spin_unlock(&ctx
->lock
);
1173 * Called from scheduler to add the counters of the current task
1174 * with interrupts disabled.
1176 * We restore the counter value and then enable it.
1178 * This does not protect us against NMI, but enable()
1179 * sets the enabled bit in the control field of counter _before_
1180 * accessing the counter control register. If a NMI hits, then it will
1181 * keep the counter running.
1183 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
1185 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1186 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1190 if (cpuctx
->task_ctx
== ctx
)
1192 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1193 cpuctx
->task_ctx
= ctx
;
1196 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1198 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
1200 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1203 #define MAX_INTERRUPTS (~0ULL)
1205 static void perf_log_throttle(struct perf_counter
*counter
, int enable
);
1206 static void perf_log_period(struct perf_counter
*counter
, u64 period
);
1208 static void perf_adjust_period(struct perf_counter
*counter
, u64 events
)
1210 struct hw_perf_counter
*hwc
= &counter
->hw
;
1211 u64 period
, sample_period
;
1214 events
*= hwc
->sample_period
;
1215 period
= div64_u64(events
, counter
->attr
.sample_freq
);
1217 delta
= (s64
)(period
- hwc
->sample_period
);
1218 delta
= (delta
+ 7) / 8; /* low pass filter */
1220 sample_period
= hwc
->sample_period
+ delta
;
1225 perf_log_period(counter
, sample_period
);
1227 hwc
->sample_period
= sample_period
;
1230 static void perf_ctx_adjust_freq(struct perf_counter_context
*ctx
)
1232 struct perf_counter
*counter
;
1233 struct hw_perf_counter
*hwc
;
1234 u64 interrupts
, freq
;
1236 spin_lock(&ctx
->lock
);
1237 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1238 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1243 interrupts
= hwc
->interrupts
;
1244 hwc
->interrupts
= 0;
1247 * unthrottle counters on the tick
1249 if (interrupts
== MAX_INTERRUPTS
) {
1250 perf_log_throttle(counter
, 1);
1251 counter
->pmu
->unthrottle(counter
);
1252 interrupts
= 2*sysctl_perf_counter_sample_rate
/HZ
;
1255 if (!counter
->attr
.freq
|| !counter
->attr
.sample_freq
)
1259 * if the specified freq < HZ then we need to skip ticks
1261 if (counter
->attr
.sample_freq
< HZ
) {
1262 freq
= counter
->attr
.sample_freq
;
1264 hwc
->freq_count
+= freq
;
1265 hwc
->freq_interrupts
+= interrupts
;
1267 if (hwc
->freq_count
< HZ
)
1270 interrupts
= hwc
->freq_interrupts
;
1271 hwc
->freq_interrupts
= 0;
1272 hwc
->freq_count
-= HZ
;
1276 perf_adjust_period(counter
, freq
* interrupts
);
1279 * In order to avoid being stalled by an (accidental) huge
1280 * sample period, force reset the sample period if we didn't
1281 * get any events in this freq period.
1285 counter
->pmu
->disable(counter
);
1286 atomic64_set(&hwc
->period_left
, 0);
1287 counter
->pmu
->enable(counter
);
1291 spin_unlock(&ctx
->lock
);
1295 * Round-robin a context's counters:
1297 static void rotate_ctx(struct perf_counter_context
*ctx
)
1299 struct perf_counter
*counter
;
1301 if (!ctx
->nr_counters
)
1304 spin_lock(&ctx
->lock
);
1306 * Rotate the first entry last (works just fine for group counters too):
1309 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1310 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1315 spin_unlock(&ctx
->lock
);
1318 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1320 struct perf_cpu_context
*cpuctx
;
1321 struct perf_counter_context
*ctx
;
1323 if (!atomic_read(&nr_counters
))
1326 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1327 ctx
= curr
->perf_counter_ctxp
;
1329 perf_ctx_adjust_freq(&cpuctx
->ctx
);
1331 perf_ctx_adjust_freq(ctx
);
1333 perf_counter_cpu_sched_out(cpuctx
);
1335 __perf_counter_task_sched_out(ctx
);
1337 rotate_ctx(&cpuctx
->ctx
);
1341 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1343 perf_counter_task_sched_in(curr
, cpu
);
1347 * Cross CPU call to read the hardware counter
1349 static void __read(void *info
)
1351 struct perf_counter
*counter
= info
;
1352 struct perf_counter_context
*ctx
= counter
->ctx
;
1353 unsigned long flags
;
1355 local_irq_save(flags
);
1357 update_context_time(ctx
);
1358 counter
->pmu
->read(counter
);
1359 update_counter_times(counter
);
1360 local_irq_restore(flags
);
1363 static u64
perf_counter_read(struct perf_counter
*counter
)
1366 * If counter is enabled and currently active on a CPU, update the
1367 * value in the counter structure:
1369 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1370 smp_call_function_single(counter
->oncpu
,
1371 __read
, counter
, 1);
1372 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1373 update_counter_times(counter
);
1376 return atomic64_read(&counter
->count
);
1380 * Initialize the perf_counter context in a task_struct:
1383 __perf_counter_init_context(struct perf_counter_context
*ctx
,
1384 struct task_struct
*task
)
1386 memset(ctx
, 0, sizeof(*ctx
));
1387 spin_lock_init(&ctx
->lock
);
1388 mutex_init(&ctx
->mutex
);
1389 INIT_LIST_HEAD(&ctx
->counter_list
);
1390 INIT_LIST_HEAD(&ctx
->event_list
);
1391 atomic_set(&ctx
->refcount
, 1);
1395 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1397 struct perf_counter_context
*parent_ctx
;
1398 struct perf_counter_context
*ctx
;
1399 struct perf_cpu_context
*cpuctx
;
1400 struct task_struct
*task
;
1401 unsigned long flags
;
1405 * If cpu is not a wildcard then this is a percpu counter:
1408 /* Must be root to operate on a CPU counter: */
1409 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN
))
1410 return ERR_PTR(-EACCES
);
1412 if (cpu
< 0 || cpu
> num_possible_cpus())
1413 return ERR_PTR(-EINVAL
);
1416 * We could be clever and allow to attach a counter to an
1417 * offline CPU and activate it when the CPU comes up, but
1420 if (!cpu_isset(cpu
, cpu_online_map
))
1421 return ERR_PTR(-ENODEV
);
1423 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1434 task
= find_task_by_vpid(pid
);
1436 get_task_struct(task
);
1440 return ERR_PTR(-ESRCH
);
1443 * Can't attach counters to a dying task.
1446 if (task
->flags
& PF_EXITING
)
1449 /* Reuse ptrace permission checks for now. */
1451 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1455 ctx
= perf_lock_task_context(task
, &flags
);
1457 parent_ctx
= ctx
->parent_ctx
;
1459 put_ctx(parent_ctx
);
1460 ctx
->parent_ctx
= NULL
; /* no longer a clone */
1463 * Get an extra reference before dropping the lock so that
1464 * this context won't get freed if the task exits.
1467 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1471 ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
1475 __perf_counter_init_context(ctx
, task
);
1477 if (cmpxchg(&task
->perf_counter_ctxp
, NULL
, ctx
)) {
1479 * We raced with some other task; use
1480 * the context they set.
1485 get_task_struct(task
);
1488 put_task_struct(task
);
1492 put_task_struct(task
);
1493 return ERR_PTR(err
);
1496 static void free_counter_rcu(struct rcu_head
*head
)
1498 struct perf_counter
*counter
;
1500 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1502 put_pid_ns(counter
->ns
);
1506 static void perf_pending_sync(struct perf_counter
*counter
);
1508 static void free_counter(struct perf_counter
*counter
)
1510 perf_pending_sync(counter
);
1512 atomic_dec(&nr_counters
);
1513 if (counter
->attr
.mmap
)
1514 atomic_dec(&nr_mmap_counters
);
1515 if (counter
->attr
.comm
)
1516 atomic_dec(&nr_comm_counters
);
1518 if (counter
->destroy
)
1519 counter
->destroy(counter
);
1521 put_ctx(counter
->ctx
);
1522 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1526 * Called when the last reference to the file is gone.
1528 static int perf_release(struct inode
*inode
, struct file
*file
)
1530 struct perf_counter
*counter
= file
->private_data
;
1531 struct perf_counter_context
*ctx
= counter
->ctx
;
1533 file
->private_data
= NULL
;
1535 WARN_ON_ONCE(ctx
->parent_ctx
);
1536 mutex_lock(&ctx
->mutex
);
1537 perf_counter_remove_from_context(counter
);
1538 mutex_unlock(&ctx
->mutex
);
1540 mutex_lock(&counter
->owner
->perf_counter_mutex
);
1541 list_del_init(&counter
->owner_entry
);
1542 mutex_unlock(&counter
->owner
->perf_counter_mutex
);
1543 put_task_struct(counter
->owner
);
1545 free_counter(counter
);
1551 * Read the performance counter - simple non blocking version for now
1554 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1560 * Return end-of-file for a read on a counter that is in
1561 * error state (i.e. because it was pinned but it couldn't be
1562 * scheduled on to the CPU at some point).
1564 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1567 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1568 mutex_lock(&counter
->child_mutex
);
1569 values
[0] = perf_counter_read(counter
);
1571 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1572 values
[n
++] = counter
->total_time_enabled
+
1573 atomic64_read(&counter
->child_total_time_enabled
);
1574 if (counter
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1575 values
[n
++] = counter
->total_time_running
+
1576 atomic64_read(&counter
->child_total_time_running
);
1577 if (counter
->attr
.read_format
& PERF_FORMAT_ID
)
1578 values
[n
++] = counter
->id
;
1579 mutex_unlock(&counter
->child_mutex
);
1581 if (count
< n
* sizeof(u64
))
1583 count
= n
* sizeof(u64
);
1585 if (copy_to_user(buf
, values
, count
))
1592 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1594 struct perf_counter
*counter
= file
->private_data
;
1596 return perf_read_hw(counter
, buf
, count
);
1599 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1601 struct perf_counter
*counter
= file
->private_data
;
1602 struct perf_mmap_data
*data
;
1603 unsigned int events
= POLL_HUP
;
1606 data
= rcu_dereference(counter
->data
);
1608 events
= atomic_xchg(&data
->poll
, 0);
1611 poll_wait(file
, &counter
->waitq
, wait
);
1616 static void perf_counter_reset(struct perf_counter
*counter
)
1618 (void)perf_counter_read(counter
);
1619 atomic64_set(&counter
->count
, 0);
1620 perf_counter_update_userpage(counter
);
1624 * Holding the top-level counter's child_mutex means that any
1625 * descendant process that has inherited this counter will block
1626 * in sync_child_counter if it goes to exit, thus satisfying the
1627 * task existence requirements of perf_counter_enable/disable.
1629 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1630 void (*func
)(struct perf_counter
*))
1632 struct perf_counter
*child
;
1634 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1635 mutex_lock(&counter
->child_mutex
);
1637 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1639 mutex_unlock(&counter
->child_mutex
);
1642 static void perf_counter_for_each(struct perf_counter
*counter
,
1643 void (*func
)(struct perf_counter
*))
1645 struct perf_counter_context
*ctx
= counter
->ctx
;
1646 struct perf_counter
*sibling
;
1648 WARN_ON_ONCE(ctx
->parent_ctx
);
1649 mutex_lock(&ctx
->mutex
);
1650 counter
= counter
->group_leader
;
1652 perf_counter_for_each_child(counter
, func
);
1654 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1655 perf_counter_for_each_child(counter
, func
);
1656 mutex_unlock(&ctx
->mutex
);
1659 static int perf_counter_period(struct perf_counter
*counter
, u64 __user
*arg
)
1661 struct perf_counter_context
*ctx
= counter
->ctx
;
1666 if (!counter
->attr
.sample_period
)
1669 size
= copy_from_user(&value
, arg
, sizeof(value
));
1670 if (size
!= sizeof(value
))
1676 spin_lock_irq(&ctx
->lock
);
1677 if (counter
->attr
.freq
) {
1678 if (value
> sysctl_perf_counter_sample_rate
) {
1683 counter
->attr
.sample_freq
= value
;
1685 perf_log_period(counter
, value
);
1687 counter
->attr
.sample_period
= value
;
1688 counter
->hw
.sample_period
= value
;
1691 spin_unlock_irq(&ctx
->lock
);
1696 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1698 struct perf_counter
*counter
= file
->private_data
;
1699 void (*func
)(struct perf_counter
*);
1703 case PERF_COUNTER_IOC_ENABLE
:
1704 func
= perf_counter_enable
;
1706 case PERF_COUNTER_IOC_DISABLE
:
1707 func
= perf_counter_disable
;
1709 case PERF_COUNTER_IOC_RESET
:
1710 func
= perf_counter_reset
;
1713 case PERF_COUNTER_IOC_REFRESH
:
1714 return perf_counter_refresh(counter
, arg
);
1716 case PERF_COUNTER_IOC_PERIOD
:
1717 return perf_counter_period(counter
, (u64 __user
*)arg
);
1723 if (flags
& PERF_IOC_FLAG_GROUP
)
1724 perf_counter_for_each(counter
, func
);
1726 perf_counter_for_each_child(counter
, func
);
1731 int perf_counter_task_enable(void)
1733 struct perf_counter
*counter
;
1735 mutex_lock(¤t
->perf_counter_mutex
);
1736 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1737 perf_counter_for_each_child(counter
, perf_counter_enable
);
1738 mutex_unlock(¤t
->perf_counter_mutex
);
1743 int perf_counter_task_disable(void)
1745 struct perf_counter
*counter
;
1747 mutex_lock(¤t
->perf_counter_mutex
);
1748 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1749 perf_counter_for_each_child(counter
, perf_counter_disable
);
1750 mutex_unlock(¤t
->perf_counter_mutex
);
1756 * Callers need to ensure there can be no nesting of this function, otherwise
1757 * the seqlock logic goes bad. We can not serialize this because the arch
1758 * code calls this from NMI context.
1760 void perf_counter_update_userpage(struct perf_counter
*counter
)
1762 struct perf_counter_mmap_page
*userpg
;
1763 struct perf_mmap_data
*data
;
1766 data
= rcu_dereference(counter
->data
);
1770 userpg
= data
->user_page
;
1773 * Disable preemption so as to not let the corresponding user-space
1774 * spin too long if we get preempted.
1779 userpg
->index
= counter
->hw
.idx
;
1780 userpg
->offset
= atomic64_read(&counter
->count
);
1781 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1782 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1791 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1793 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1794 struct perf_mmap_data
*data
;
1795 int ret
= VM_FAULT_SIGBUS
;
1797 if (vmf
->flags
& FAULT_FLAG_MKWRITE
) {
1798 if (vmf
->pgoff
== 0)
1804 data
= rcu_dereference(counter
->data
);
1808 if (vmf
->pgoff
== 0) {
1809 vmf
->page
= virt_to_page(data
->user_page
);
1811 int nr
= vmf
->pgoff
- 1;
1813 if ((unsigned)nr
> data
->nr_pages
)
1816 if (vmf
->flags
& FAULT_FLAG_WRITE
)
1819 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1822 get_page(vmf
->page
);
1823 vmf
->page
->mapping
= vma
->vm_file
->f_mapping
;
1824 vmf
->page
->index
= vmf
->pgoff
;
1833 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1835 struct perf_mmap_data
*data
;
1839 WARN_ON(atomic_read(&counter
->mmap_count
));
1841 size
= sizeof(struct perf_mmap_data
);
1842 size
+= nr_pages
* sizeof(void *);
1844 data
= kzalloc(size
, GFP_KERNEL
);
1848 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1849 if (!data
->user_page
)
1850 goto fail_user_page
;
1852 for (i
= 0; i
< nr_pages
; i
++) {
1853 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1854 if (!data
->data_pages
[i
])
1855 goto fail_data_pages
;
1858 data
->nr_pages
= nr_pages
;
1859 atomic_set(&data
->lock
, -1);
1861 rcu_assign_pointer(counter
->data
, data
);
1866 for (i
--; i
>= 0; i
--)
1867 free_page((unsigned long)data
->data_pages
[i
]);
1869 free_page((unsigned long)data
->user_page
);
1878 static void perf_mmap_free_page(unsigned long addr
)
1880 struct page
*page
= virt_to_page(addr
);
1882 page
->mapping
= NULL
;
1886 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1888 struct perf_mmap_data
*data
;
1891 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
1893 perf_mmap_free_page((unsigned long)data
->user_page
);
1894 for (i
= 0; i
< data
->nr_pages
; i
++)
1895 perf_mmap_free_page((unsigned long)data
->data_pages
[i
]);
1900 static void perf_mmap_data_free(struct perf_counter
*counter
)
1902 struct perf_mmap_data
*data
= counter
->data
;
1904 WARN_ON(atomic_read(&counter
->mmap_count
));
1906 rcu_assign_pointer(counter
->data
, NULL
);
1907 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1910 static void perf_mmap_open(struct vm_area_struct
*vma
)
1912 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1914 atomic_inc(&counter
->mmap_count
);
1917 static void perf_mmap_close(struct vm_area_struct
*vma
)
1919 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1921 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1922 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
, &counter
->mmap_mutex
)) {
1923 struct user_struct
*user
= current_user();
1925 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
1926 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
1927 perf_mmap_data_free(counter
);
1928 mutex_unlock(&counter
->mmap_mutex
);
1932 static struct vm_operations_struct perf_mmap_vmops
= {
1933 .open
= perf_mmap_open
,
1934 .close
= perf_mmap_close
,
1935 .fault
= perf_mmap_fault
,
1936 .page_mkwrite
= perf_mmap_fault
,
1939 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1941 struct perf_counter
*counter
= file
->private_data
;
1942 unsigned long user_locked
, user_lock_limit
;
1943 struct user_struct
*user
= current_user();
1944 unsigned long locked
, lock_limit
;
1945 unsigned long vma_size
;
1946 unsigned long nr_pages
;
1947 long user_extra
, extra
;
1950 if (!(vma
->vm_flags
& VM_SHARED
))
1953 vma_size
= vma
->vm_end
- vma
->vm_start
;
1954 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1957 * If we have data pages ensure they're a power-of-two number, so we
1958 * can do bitmasks instead of modulo.
1960 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1963 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1966 if (vma
->vm_pgoff
!= 0)
1969 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1970 mutex_lock(&counter
->mmap_mutex
);
1971 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1972 if (nr_pages
!= counter
->data
->nr_pages
)
1977 user_extra
= nr_pages
+ 1;
1978 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
1981 * Increase the limit linearly with more CPUs:
1983 user_lock_limit
*= num_online_cpus();
1985 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
1988 if (user_locked
> user_lock_limit
)
1989 extra
= user_locked
- user_lock_limit
;
1991 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1992 lock_limit
>>= PAGE_SHIFT
;
1993 locked
= vma
->vm_mm
->locked_vm
+ extra
;
1995 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
2000 WARN_ON(counter
->data
);
2001 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
2005 atomic_set(&counter
->mmap_count
, 1);
2006 atomic_long_add(user_extra
, &user
->locked_vm
);
2007 vma
->vm_mm
->locked_vm
+= extra
;
2008 counter
->data
->nr_locked
= extra
;
2009 if (vma
->vm_flags
& VM_WRITE
)
2010 counter
->data
->writable
= 1;
2013 mutex_unlock(&counter
->mmap_mutex
);
2015 vma
->vm_flags
|= VM_RESERVED
;
2016 vma
->vm_ops
= &perf_mmap_vmops
;
2021 static int perf_fasync(int fd
, struct file
*filp
, int on
)
2023 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
2024 struct perf_counter
*counter
= filp
->private_data
;
2027 mutex_lock(&inode
->i_mutex
);
2028 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
2029 mutex_unlock(&inode
->i_mutex
);
2037 static const struct file_operations perf_fops
= {
2038 .release
= perf_release
,
2041 .unlocked_ioctl
= perf_ioctl
,
2042 .compat_ioctl
= perf_ioctl
,
2044 .fasync
= perf_fasync
,
2048 * Perf counter wakeup
2050 * If there's data, ensure we set the poll() state and publish everything
2051 * to user-space before waking everybody up.
2054 void perf_counter_wakeup(struct perf_counter
*counter
)
2056 wake_up_all(&counter
->waitq
);
2058 if (counter
->pending_kill
) {
2059 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
2060 counter
->pending_kill
= 0;
2067 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2069 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2070 * single linked list and use cmpxchg() to add entries lockless.
2073 static void perf_pending_counter(struct perf_pending_entry
*entry
)
2075 struct perf_counter
*counter
= container_of(entry
,
2076 struct perf_counter
, pending
);
2078 if (counter
->pending_disable
) {
2079 counter
->pending_disable
= 0;
2080 perf_counter_disable(counter
);
2083 if (counter
->pending_wakeup
) {
2084 counter
->pending_wakeup
= 0;
2085 perf_counter_wakeup(counter
);
2089 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2091 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2095 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2096 void (*func
)(struct perf_pending_entry
*))
2098 struct perf_pending_entry
**head
;
2100 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2105 head
= &get_cpu_var(perf_pending_head
);
2108 entry
->next
= *head
;
2109 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2111 set_perf_counter_pending();
2113 put_cpu_var(perf_pending_head
);
2116 static int __perf_pending_run(void)
2118 struct perf_pending_entry
*list
;
2121 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2122 while (list
!= PENDING_TAIL
) {
2123 void (*func
)(struct perf_pending_entry
*);
2124 struct perf_pending_entry
*entry
= list
;
2131 * Ensure we observe the unqueue before we issue the wakeup,
2132 * so that we won't be waiting forever.
2133 * -- see perf_not_pending().
2144 static inline int perf_not_pending(struct perf_counter
*counter
)
2147 * If we flush on whatever cpu we run, there is a chance we don't
2151 __perf_pending_run();
2155 * Ensure we see the proper queue state before going to sleep
2156 * so that we do not miss the wakeup. -- see perf_pending_handle()
2159 return counter
->pending
.next
== NULL
;
2162 static void perf_pending_sync(struct perf_counter
*counter
)
2164 wait_event(counter
->waitq
, perf_not_pending(counter
));
2167 void perf_counter_do_pending(void)
2169 __perf_pending_run();
2173 * Callchain support -- arch specific
2176 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2185 struct perf_output_handle
{
2186 struct perf_counter
*counter
;
2187 struct perf_mmap_data
*data
;
2189 unsigned long offset
;
2193 unsigned long flags
;
2196 static bool perf_output_space(struct perf_mmap_data
*data
,
2197 unsigned int offset
, unsigned int head
)
2202 if (!data
->writable
)
2205 mask
= (data
->nr_pages
<< PAGE_SHIFT
) - 1;
2207 * Userspace could choose to issue a mb() before updating the tail
2208 * pointer. So that all reads will be completed before the write is
2211 tail
= ACCESS_ONCE(data
->user_page
->data_tail
);
2214 offset
= (offset
- tail
) & mask
;
2215 head
= (head
- tail
) & mask
;
2217 if ((int)(head
- offset
) < 0)
2223 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2225 atomic_set(&handle
->data
->poll
, POLL_IN
);
2228 handle
->counter
->pending_wakeup
= 1;
2229 perf_pending_queue(&handle
->counter
->pending
,
2230 perf_pending_counter
);
2232 perf_counter_wakeup(handle
->counter
);
2236 * Curious locking construct.
2238 * We need to ensure a later event doesn't publish a head when a former
2239 * event isn't done writing. However since we need to deal with NMIs we
2240 * cannot fully serialize things.
2242 * What we do is serialize between CPUs so we only have to deal with NMI
2243 * nesting on a single CPU.
2245 * We only publish the head (and generate a wakeup) when the outer-most
2248 static void perf_output_lock(struct perf_output_handle
*handle
)
2250 struct perf_mmap_data
*data
= handle
->data
;
2255 local_irq_save(handle
->flags
);
2256 cpu
= smp_processor_id();
2258 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
2261 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2267 static void perf_output_unlock(struct perf_output_handle
*handle
)
2269 struct perf_mmap_data
*data
= handle
->data
;
2273 data
->done_head
= data
->head
;
2275 if (!handle
->locked
)
2280 * The xchg implies a full barrier that ensures all writes are done
2281 * before we publish the new head, matched by a rmb() in userspace when
2282 * reading this position.
2284 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2285 data
->user_page
->data_head
= head
;
2288 * NMI can happen here, which means we can miss a done_head update.
2291 cpu
= atomic_xchg(&data
->lock
, -1);
2292 WARN_ON_ONCE(cpu
!= smp_processor_id());
2295 * Therefore we have to validate we did not indeed do so.
2297 if (unlikely(atomic_long_read(&data
->done_head
))) {
2299 * Since we had it locked, we can lock it again.
2301 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2307 if (atomic_xchg(&data
->wakeup
, 0))
2308 perf_output_wakeup(handle
);
2310 local_irq_restore(handle
->flags
);
2313 static void perf_output_copy(struct perf_output_handle
*handle
,
2314 const void *buf
, unsigned int len
)
2316 unsigned int pages_mask
;
2317 unsigned int offset
;
2321 offset
= handle
->offset
;
2322 pages_mask
= handle
->data
->nr_pages
- 1;
2323 pages
= handle
->data
->data_pages
;
2326 unsigned int page_offset
;
2329 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2330 page_offset
= offset
& (PAGE_SIZE
- 1);
2331 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
2333 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2340 handle
->offset
= offset
;
2343 * Check we didn't copy past our reservation window, taking the
2344 * possible unsigned int wrap into account.
2346 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2349 #define perf_output_put(handle, x) \
2350 perf_output_copy((handle), &(x), sizeof(x))
2352 static int perf_output_begin(struct perf_output_handle
*handle
,
2353 struct perf_counter
*counter
, unsigned int size
,
2354 int nmi
, int sample
)
2356 struct perf_mmap_data
*data
;
2357 unsigned int offset
, head
;
2360 struct perf_event_header header
;
2366 * For inherited counters we send all the output towards the parent.
2368 if (counter
->parent
)
2369 counter
= counter
->parent
;
2372 data
= rcu_dereference(counter
->data
);
2376 handle
->data
= data
;
2377 handle
->counter
= counter
;
2379 handle
->sample
= sample
;
2381 if (!data
->nr_pages
)
2384 have_lost
= atomic_read(&data
->lost
);
2386 size
+= sizeof(lost_event
);
2388 perf_output_lock(handle
);
2391 offset
= head
= atomic_long_read(&data
->head
);
2393 if (unlikely(!perf_output_space(data
, offset
, head
)))
2395 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2397 handle
->offset
= offset
;
2398 handle
->head
= head
;
2400 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
2401 atomic_set(&data
->wakeup
, 1);
2404 lost_event
.header
.type
= PERF_EVENT_LOST
;
2405 lost_event
.header
.misc
= 0;
2406 lost_event
.header
.size
= sizeof(lost_event
);
2407 lost_event
.id
= counter
->id
;
2408 lost_event
.lost
= atomic_xchg(&data
->lost
, 0);
2410 perf_output_put(handle
, lost_event
);
2416 atomic_inc(&data
->lost
);
2417 perf_output_unlock(handle
);
2424 static void perf_output_end(struct perf_output_handle
*handle
)
2426 struct perf_counter
*counter
= handle
->counter
;
2427 struct perf_mmap_data
*data
= handle
->data
;
2429 int wakeup_events
= counter
->attr
.wakeup_events
;
2431 if (handle
->sample
&& wakeup_events
) {
2432 int events
= atomic_inc_return(&data
->events
);
2433 if (events
>= wakeup_events
) {
2434 atomic_sub(wakeup_events
, &data
->events
);
2435 atomic_set(&data
->wakeup
, 1);
2439 perf_output_unlock(handle
);
2443 static u32
perf_counter_pid(struct perf_counter
*counter
, struct task_struct
*p
)
2446 * only top level counters have the pid namespace they were created in
2448 if (counter
->parent
)
2449 counter
= counter
->parent
;
2451 return task_tgid_nr_ns(p
, counter
->ns
);
2454 static u32
perf_counter_tid(struct perf_counter
*counter
, struct task_struct
*p
)
2457 * only top level counters have the pid namespace they were created in
2459 if (counter
->parent
)
2460 counter
= counter
->parent
;
2462 return task_pid_nr_ns(p
, counter
->ns
);
2465 static void perf_counter_output(struct perf_counter
*counter
, int nmi
,
2466 struct perf_sample_data
*data
)
2469 u64 sample_type
= counter
->attr
.sample_type
;
2470 struct perf_output_handle handle
;
2471 struct perf_event_header header
;
2480 struct perf_callchain_entry
*callchain
= NULL
;
2481 int callchain_size
= 0;
2488 header
.size
= sizeof(header
);
2490 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
2491 header
.misc
|= perf_misc_flags(data
->regs
);
2493 if (sample_type
& PERF_SAMPLE_IP
) {
2494 ip
= perf_instruction_pointer(data
->regs
);
2495 header
.type
|= PERF_SAMPLE_IP
;
2496 header
.size
+= sizeof(ip
);
2499 if (sample_type
& PERF_SAMPLE_TID
) {
2500 /* namespace issues */
2501 tid_entry
.pid
= perf_counter_pid(counter
, current
);
2502 tid_entry
.tid
= perf_counter_tid(counter
, current
);
2504 header
.type
|= PERF_SAMPLE_TID
;
2505 header
.size
+= sizeof(tid_entry
);
2508 if (sample_type
& PERF_SAMPLE_TIME
) {
2510 * Maybe do better on x86 and provide cpu_clock_nmi()
2512 time
= sched_clock();
2514 header
.type
|= PERF_SAMPLE_TIME
;
2515 header
.size
+= sizeof(u64
);
2518 if (sample_type
& PERF_SAMPLE_ADDR
) {
2519 header
.type
|= PERF_SAMPLE_ADDR
;
2520 header
.size
+= sizeof(u64
);
2523 if (sample_type
& PERF_SAMPLE_ID
) {
2524 header
.type
|= PERF_SAMPLE_ID
;
2525 header
.size
+= sizeof(u64
);
2528 if (sample_type
& PERF_SAMPLE_CPU
) {
2529 header
.type
|= PERF_SAMPLE_CPU
;
2530 header
.size
+= sizeof(cpu_entry
);
2532 cpu_entry
.cpu
= raw_smp_processor_id();
2535 if (sample_type
& PERF_SAMPLE_PERIOD
) {
2536 header
.type
|= PERF_SAMPLE_PERIOD
;
2537 header
.size
+= sizeof(u64
);
2540 if (sample_type
& PERF_SAMPLE_GROUP
) {
2541 header
.type
|= PERF_SAMPLE_GROUP
;
2542 header
.size
+= sizeof(u64
) +
2543 counter
->nr_siblings
* sizeof(group_entry
);
2546 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2547 callchain
= perf_callchain(data
->regs
);
2550 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2552 header
.type
|= PERF_SAMPLE_CALLCHAIN
;
2553 header
.size
+= callchain_size
;
2557 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2561 perf_output_put(&handle
, header
);
2563 if (sample_type
& PERF_SAMPLE_IP
)
2564 perf_output_put(&handle
, ip
);
2566 if (sample_type
& PERF_SAMPLE_TID
)
2567 perf_output_put(&handle
, tid_entry
);
2569 if (sample_type
& PERF_SAMPLE_TIME
)
2570 perf_output_put(&handle
, time
);
2572 if (sample_type
& PERF_SAMPLE_ADDR
)
2573 perf_output_put(&handle
, data
->addr
);
2575 if (sample_type
& PERF_SAMPLE_ID
)
2576 perf_output_put(&handle
, counter
->id
);
2578 if (sample_type
& PERF_SAMPLE_CPU
)
2579 perf_output_put(&handle
, cpu_entry
);
2581 if (sample_type
& PERF_SAMPLE_PERIOD
)
2582 perf_output_put(&handle
, data
->period
);
2585 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2587 if (sample_type
& PERF_SAMPLE_GROUP
) {
2588 struct perf_counter
*leader
, *sub
;
2589 u64 nr
= counter
->nr_siblings
;
2591 perf_output_put(&handle
, nr
);
2593 leader
= counter
->group_leader
;
2594 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2596 sub
->pmu
->read(sub
);
2598 group_entry
.id
= sub
->id
;
2599 group_entry
.counter
= atomic64_read(&sub
->count
);
2601 perf_output_put(&handle
, group_entry
);
2606 perf_output_copy(&handle
, callchain
, callchain_size
);
2608 perf_output_end(&handle
);
2615 struct perf_fork_event
{
2616 struct task_struct
*task
;
2619 struct perf_event_header header
;
2626 static void perf_counter_fork_output(struct perf_counter
*counter
,
2627 struct perf_fork_event
*fork_event
)
2629 struct perf_output_handle handle
;
2630 int size
= fork_event
->event
.header
.size
;
2631 struct task_struct
*task
= fork_event
->task
;
2632 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2637 fork_event
->event
.pid
= perf_counter_pid(counter
, task
);
2638 fork_event
->event
.ppid
= perf_counter_pid(counter
, task
->real_parent
);
2640 perf_output_put(&handle
, fork_event
->event
);
2641 perf_output_end(&handle
);
2644 static int perf_counter_fork_match(struct perf_counter
*counter
)
2646 if (counter
->attr
.comm
|| counter
->attr
.mmap
)
2652 static void perf_counter_fork_ctx(struct perf_counter_context
*ctx
,
2653 struct perf_fork_event
*fork_event
)
2655 struct perf_counter
*counter
;
2657 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2661 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2662 if (perf_counter_fork_match(counter
))
2663 perf_counter_fork_output(counter
, fork_event
);
2668 static void perf_counter_fork_event(struct perf_fork_event
*fork_event
)
2670 struct perf_cpu_context
*cpuctx
;
2671 struct perf_counter_context
*ctx
;
2673 cpuctx
= &get_cpu_var(perf_cpu_context
);
2674 perf_counter_fork_ctx(&cpuctx
->ctx
, fork_event
);
2675 put_cpu_var(perf_cpu_context
);
2679 * doesn't really matter which of the child contexts the
2680 * events ends up in.
2682 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2684 perf_counter_fork_ctx(ctx
, fork_event
);
2688 void perf_counter_fork(struct task_struct
*task
)
2690 struct perf_fork_event fork_event
;
2692 if (!atomic_read(&nr_comm_counters
) &&
2693 !atomic_read(&nr_mmap_counters
))
2696 fork_event
= (struct perf_fork_event
){
2700 .type
= PERF_EVENT_FORK
,
2701 .size
= sizeof(fork_event
.event
),
2706 perf_counter_fork_event(&fork_event
);
2713 struct perf_comm_event
{
2714 struct task_struct
*task
;
2719 struct perf_event_header header
;
2726 static void perf_counter_comm_output(struct perf_counter
*counter
,
2727 struct perf_comm_event
*comm_event
)
2729 struct perf_output_handle handle
;
2730 int size
= comm_event
->event
.header
.size
;
2731 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2736 comm_event
->event
.pid
= perf_counter_pid(counter
, comm_event
->task
);
2737 comm_event
->event
.tid
= perf_counter_tid(counter
, comm_event
->task
);
2739 perf_output_put(&handle
, comm_event
->event
);
2740 perf_output_copy(&handle
, comm_event
->comm
,
2741 comm_event
->comm_size
);
2742 perf_output_end(&handle
);
2745 static int perf_counter_comm_match(struct perf_counter
*counter
)
2747 if (counter
->attr
.comm
)
2753 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2754 struct perf_comm_event
*comm_event
)
2756 struct perf_counter
*counter
;
2758 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2762 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2763 if (perf_counter_comm_match(counter
))
2764 perf_counter_comm_output(counter
, comm_event
);
2769 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2771 struct perf_cpu_context
*cpuctx
;
2772 struct perf_counter_context
*ctx
;
2774 char *comm
= comm_event
->task
->comm
;
2776 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2778 comm_event
->comm
= comm
;
2779 comm_event
->comm_size
= size
;
2781 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2783 cpuctx
= &get_cpu_var(perf_cpu_context
);
2784 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2785 put_cpu_var(perf_cpu_context
);
2789 * doesn't really matter which of the child contexts the
2790 * events ends up in.
2792 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2794 perf_counter_comm_ctx(ctx
, comm_event
);
2798 void perf_counter_comm(struct task_struct
*task
)
2800 struct perf_comm_event comm_event
;
2802 if (!atomic_read(&nr_comm_counters
))
2805 comm_event
= (struct perf_comm_event
){
2808 .header
= { .type
= PERF_EVENT_COMM
, },
2812 perf_counter_comm_event(&comm_event
);
2819 struct perf_mmap_event
{
2820 struct vm_area_struct
*vma
;
2822 const char *file_name
;
2826 struct perf_event_header header
;
2836 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2837 struct perf_mmap_event
*mmap_event
)
2839 struct perf_output_handle handle
;
2840 int size
= mmap_event
->event
.header
.size
;
2841 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2846 mmap_event
->event
.pid
= perf_counter_pid(counter
, current
);
2847 mmap_event
->event
.tid
= perf_counter_tid(counter
, current
);
2849 perf_output_put(&handle
, mmap_event
->event
);
2850 perf_output_copy(&handle
, mmap_event
->file_name
,
2851 mmap_event
->file_size
);
2852 perf_output_end(&handle
);
2855 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2856 struct perf_mmap_event
*mmap_event
)
2858 if (counter
->attr
.mmap
)
2864 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2865 struct perf_mmap_event
*mmap_event
)
2867 struct perf_counter
*counter
;
2869 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2873 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2874 if (perf_counter_mmap_match(counter
, mmap_event
))
2875 perf_counter_mmap_output(counter
, mmap_event
);
2880 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2882 struct perf_cpu_context
*cpuctx
;
2883 struct perf_counter_context
*ctx
;
2884 struct vm_area_struct
*vma
= mmap_event
->vma
;
2885 struct file
*file
= vma
->vm_file
;
2892 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2894 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2897 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2899 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2903 name
= arch_vma_name(mmap_event
->vma
);
2908 name
= strncpy(tmp
, "[vdso]", sizeof(tmp
));
2912 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2917 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2919 mmap_event
->file_name
= name
;
2920 mmap_event
->file_size
= size
;
2922 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2924 cpuctx
= &get_cpu_var(perf_cpu_context
);
2925 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2926 put_cpu_var(perf_cpu_context
);
2930 * doesn't really matter which of the child contexts the
2931 * events ends up in.
2933 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2935 perf_counter_mmap_ctx(ctx
, mmap_event
);
2941 void __perf_counter_mmap(struct vm_area_struct
*vma
)
2943 struct perf_mmap_event mmap_event
;
2945 if (!atomic_read(&nr_mmap_counters
))
2948 mmap_event
= (struct perf_mmap_event
){
2951 .header
= { .type
= PERF_EVENT_MMAP
, },
2952 .start
= vma
->vm_start
,
2953 .len
= vma
->vm_end
- vma
->vm_start
,
2954 .pgoff
= vma
->vm_pgoff
,
2958 perf_counter_mmap_event(&mmap_event
);
2962 * Log sample_period changes so that analyzing tools can re-normalize the
2967 struct perf_event_header header
;
2973 static void perf_log_period(struct perf_counter
*counter
, u64 period
)
2975 struct perf_output_handle handle
;
2976 struct freq_event event
;
2979 if (counter
->hw
.sample_period
== period
)
2982 if (counter
->attr
.sample_type
& PERF_SAMPLE_PERIOD
)
2985 event
= (struct freq_event
) {
2987 .type
= PERF_EVENT_PERIOD
,
2989 .size
= sizeof(event
),
2991 .time
= sched_clock(),
2996 ret
= perf_output_begin(&handle
, counter
, sizeof(event
), 1, 0);
3000 perf_output_put(&handle
, event
);
3001 perf_output_end(&handle
);
3005 * IRQ throttle logging
3008 static void perf_log_throttle(struct perf_counter
*counter
, int enable
)
3010 struct perf_output_handle handle
;
3014 struct perf_event_header header
;
3017 } throttle_event
= {
3019 .type
= PERF_EVENT_THROTTLE
+ 1,
3021 .size
= sizeof(throttle_event
),
3023 .time
= sched_clock(),
3027 ret
= perf_output_begin(&handle
, counter
, sizeof(throttle_event
), 1, 0);
3031 perf_output_put(&handle
, throttle_event
);
3032 perf_output_end(&handle
);
3036 * Generic counter overflow handling, sampling.
3039 int perf_counter_overflow(struct perf_counter
*counter
, int nmi
,
3040 struct perf_sample_data
*data
)
3042 int events
= atomic_read(&counter
->event_limit
);
3043 int throttle
= counter
->pmu
->unthrottle
!= NULL
;
3044 struct hw_perf_counter
*hwc
= &counter
->hw
;
3050 if (hwc
->interrupts
!= MAX_INTERRUPTS
) {
3052 if (HZ
* hwc
->interrupts
>
3053 (u64
)sysctl_perf_counter_sample_rate
) {
3054 hwc
->interrupts
= MAX_INTERRUPTS
;
3055 perf_log_throttle(counter
, 0);
3060 * Keep re-disabling counters even though on the previous
3061 * pass we disabled it - just in case we raced with a
3062 * sched-in and the counter got enabled again:
3068 if (counter
->attr
.freq
) {
3069 u64 now
= sched_clock();
3070 s64 delta
= now
- hwc
->freq_stamp
;
3072 hwc
->freq_stamp
= now
;
3074 if (delta
> 0 && delta
< TICK_NSEC
)
3075 perf_adjust_period(counter
, NSEC_PER_SEC
/ (int)delta
);
3079 * XXX event_limit might not quite work as expected on inherited
3083 counter
->pending_kill
= POLL_IN
;
3084 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
3086 counter
->pending_kill
= POLL_HUP
;
3088 counter
->pending_disable
= 1;
3089 perf_pending_queue(&counter
->pending
,
3090 perf_pending_counter
);
3092 perf_counter_disable(counter
);
3095 perf_counter_output(counter
, nmi
, data
);
3100 * Generic software counter infrastructure
3103 static void perf_swcounter_update(struct perf_counter
*counter
)
3105 struct hw_perf_counter
*hwc
= &counter
->hw
;
3110 prev
= atomic64_read(&hwc
->prev_count
);
3111 now
= atomic64_read(&hwc
->count
);
3112 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
3117 atomic64_add(delta
, &counter
->count
);
3118 atomic64_sub(delta
, &hwc
->period_left
);
3121 static void perf_swcounter_set_period(struct perf_counter
*counter
)
3123 struct hw_perf_counter
*hwc
= &counter
->hw
;
3124 s64 left
= atomic64_read(&hwc
->period_left
);
3125 s64 period
= hwc
->sample_period
;
3127 if (unlikely(left
<= -period
)) {
3129 atomic64_set(&hwc
->period_left
, left
);
3130 hwc
->last_period
= period
;
3133 if (unlikely(left
<= 0)) {
3135 atomic64_add(period
, &hwc
->period_left
);
3136 hwc
->last_period
= period
;
3139 atomic64_set(&hwc
->prev_count
, -left
);
3140 atomic64_set(&hwc
->count
, -left
);
3143 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
3145 enum hrtimer_restart ret
= HRTIMER_RESTART
;
3146 struct perf_sample_data data
;
3147 struct perf_counter
*counter
;
3150 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
3151 counter
->pmu
->read(counter
);
3154 data
.regs
= get_irq_regs();
3156 * In case we exclude kernel IPs or are somehow not in interrupt
3157 * context, provide the next best thing, the user IP.
3159 if ((counter
->attr
.exclude_kernel
|| !data
.regs
) &&
3160 !counter
->attr
.exclude_user
)
3161 data
.regs
= task_pt_regs(current
);
3164 if (perf_counter_overflow(counter
, 0, &data
))
3165 ret
= HRTIMER_NORESTART
;
3168 period
= max_t(u64
, 10000, counter
->hw
.sample_period
);
3169 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
3174 static void perf_swcounter_overflow(struct perf_counter
*counter
,
3175 int nmi
, struct pt_regs
*regs
, u64 addr
)
3177 struct perf_sample_data data
= {
3180 .period
= counter
->hw
.last_period
,
3183 perf_swcounter_update(counter
);
3184 perf_swcounter_set_period(counter
);
3185 if (perf_counter_overflow(counter
, nmi
, &data
))
3186 /* soft-disable the counter */
3191 static int perf_swcounter_is_counting(struct perf_counter
*counter
)
3193 struct perf_counter_context
*ctx
;
3194 unsigned long flags
;
3197 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
3200 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
)
3204 * If the counter is inactive, it could be just because
3205 * its task is scheduled out, or because it's in a group
3206 * which could not go on the PMU. We want to count in
3207 * the first case but not the second. If the context is
3208 * currently active then an inactive software counter must
3209 * be the second case. If it's not currently active then
3210 * we need to know whether the counter was active when the
3211 * context was last active, which we can determine by
3212 * comparing counter->tstamp_stopped with ctx->time.
3214 * We are within an RCU read-side critical section,
3215 * which protects the existence of *ctx.
3218 spin_lock_irqsave(&ctx
->lock
, flags
);
3220 /* Re-check state now we have the lock */
3221 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
||
3222 counter
->ctx
->is_active
||
3223 counter
->tstamp_stopped
< ctx
->time
)
3225 spin_unlock_irqrestore(&ctx
->lock
, flags
);
3229 static int perf_swcounter_match(struct perf_counter
*counter
,
3230 enum perf_type_id type
,
3231 u32 event
, struct pt_regs
*regs
)
3233 if (!perf_swcounter_is_counting(counter
))
3236 if (counter
->attr
.type
!= type
)
3238 if (counter
->attr
.config
!= event
)
3242 if (counter
->attr
.exclude_user
&& user_mode(regs
))
3245 if (counter
->attr
.exclude_kernel
&& !user_mode(regs
))
3252 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
3253 int nmi
, struct pt_regs
*regs
, u64 addr
)
3255 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
3257 if (counter
->hw
.sample_period
&& !neg
&& regs
)
3258 perf_swcounter_overflow(counter
, nmi
, regs
, addr
);
3261 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
3262 enum perf_type_id type
, u32 event
,
3263 u64 nr
, int nmi
, struct pt_regs
*regs
,
3266 struct perf_counter
*counter
;
3268 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3272 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3273 if (perf_swcounter_match(counter
, type
, event
, regs
))
3274 perf_swcounter_add(counter
, nr
, nmi
, regs
, addr
);
3279 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
3282 return &cpuctx
->recursion
[3];
3285 return &cpuctx
->recursion
[2];
3288 return &cpuctx
->recursion
[1];
3290 return &cpuctx
->recursion
[0];
3293 static void __perf_swcounter_event(enum perf_type_id type
, u32 event
,
3294 u64 nr
, int nmi
, struct pt_regs
*regs
,
3297 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3298 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
3299 struct perf_counter_context
*ctx
;
3307 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
3308 nr
, nmi
, regs
, addr
);
3311 * doesn't really matter which of the child contexts the
3312 * events ends up in.
3314 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3316 perf_swcounter_ctx_event(ctx
, type
, event
, nr
, nmi
, regs
, addr
);
3323 put_cpu_var(perf_cpu_context
);
3327 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
3329 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
, addr
);
3332 static void perf_swcounter_read(struct perf_counter
*counter
)
3334 perf_swcounter_update(counter
);
3337 static int perf_swcounter_enable(struct perf_counter
*counter
)
3339 perf_swcounter_set_period(counter
);
3343 static void perf_swcounter_disable(struct perf_counter
*counter
)
3345 perf_swcounter_update(counter
);
3348 static const struct pmu perf_ops_generic
= {
3349 .enable
= perf_swcounter_enable
,
3350 .disable
= perf_swcounter_disable
,
3351 .read
= perf_swcounter_read
,
3355 * Software counter: cpu wall time clock
3358 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
3360 int cpu
= raw_smp_processor_id();
3364 now
= cpu_clock(cpu
);
3365 prev
= atomic64_read(&counter
->hw
.prev_count
);
3366 atomic64_set(&counter
->hw
.prev_count
, now
);
3367 atomic64_add(now
- prev
, &counter
->count
);
3370 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
3372 struct hw_perf_counter
*hwc
= &counter
->hw
;
3373 int cpu
= raw_smp_processor_id();
3375 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
3376 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3377 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3378 if (hwc
->sample_period
) {
3379 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3380 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3381 ns_to_ktime(period
), 0,
3382 HRTIMER_MODE_REL
, 0);
3388 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
3390 if (counter
->hw
.sample_period
)
3391 hrtimer_cancel(&counter
->hw
.hrtimer
);
3392 cpu_clock_perf_counter_update(counter
);
3395 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
3397 cpu_clock_perf_counter_update(counter
);
3400 static const struct pmu perf_ops_cpu_clock
= {
3401 .enable
= cpu_clock_perf_counter_enable
,
3402 .disable
= cpu_clock_perf_counter_disable
,
3403 .read
= cpu_clock_perf_counter_read
,
3407 * Software counter: task time clock
3410 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
3415 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
3417 atomic64_add(delta
, &counter
->count
);
3420 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
3422 struct hw_perf_counter
*hwc
= &counter
->hw
;
3425 now
= counter
->ctx
->time
;
3427 atomic64_set(&hwc
->prev_count
, now
);
3428 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3429 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3430 if (hwc
->sample_period
) {
3431 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3432 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3433 ns_to_ktime(period
), 0,
3434 HRTIMER_MODE_REL
, 0);
3440 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
3442 if (counter
->hw
.sample_period
)
3443 hrtimer_cancel(&counter
->hw
.hrtimer
);
3444 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
3448 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
3453 update_context_time(counter
->ctx
);
3454 time
= counter
->ctx
->time
;
3456 u64 now
= perf_clock();
3457 u64 delta
= now
- counter
->ctx
->timestamp
;
3458 time
= counter
->ctx
->time
+ delta
;
3461 task_clock_perf_counter_update(counter
, time
);
3464 static const struct pmu perf_ops_task_clock
= {
3465 .enable
= task_clock_perf_counter_enable
,
3466 .disable
= task_clock_perf_counter_disable
,
3467 .read
= task_clock_perf_counter_read
,
3470 #ifdef CONFIG_EVENT_PROFILE
3471 void perf_tpcounter_event(int event_id
)
3473 struct pt_regs
*regs
= get_irq_regs();
3476 regs
= task_pt_regs(current
);
3478 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
, 0);
3480 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
3482 extern int ftrace_profile_enable(int);
3483 extern void ftrace_profile_disable(int);
3485 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
3487 ftrace_profile_disable(perf_event_id(&counter
->attr
));
3490 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3492 int event_id
= perf_event_id(&counter
->attr
);
3495 ret
= ftrace_profile_enable(event_id
);
3499 counter
->destroy
= tp_perf_counter_destroy
;
3501 return &perf_ops_generic
;
3504 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3510 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
3512 const struct pmu
*pmu
= NULL
;
3515 * Software counters (currently) can't in general distinguish
3516 * between user, kernel and hypervisor events.
3517 * However, context switches and cpu migrations are considered
3518 * to be kernel events, and page faults are never hypervisor
3521 switch (counter
->attr
.config
) {
3522 case PERF_COUNT_SW_CPU_CLOCK
:
3523 pmu
= &perf_ops_cpu_clock
;
3526 case PERF_COUNT_SW_TASK_CLOCK
:
3528 * If the user instantiates this as a per-cpu counter,
3529 * use the cpu_clock counter instead.
3531 if (counter
->ctx
->task
)
3532 pmu
= &perf_ops_task_clock
;
3534 pmu
= &perf_ops_cpu_clock
;
3537 case PERF_COUNT_SW_PAGE_FAULTS
:
3538 case PERF_COUNT_SW_PAGE_FAULTS_MIN
:
3539 case PERF_COUNT_SW_PAGE_FAULTS_MAJ
:
3540 case PERF_COUNT_SW_CONTEXT_SWITCHES
:
3541 case PERF_COUNT_SW_CPU_MIGRATIONS
:
3542 pmu
= &perf_ops_generic
;
3550 * Allocate and initialize a counter structure
3552 static struct perf_counter
*
3553 perf_counter_alloc(struct perf_counter_attr
*attr
,
3555 struct perf_counter_context
*ctx
,
3556 struct perf_counter
*group_leader
,
3559 const struct pmu
*pmu
;
3560 struct perf_counter
*counter
;
3561 struct hw_perf_counter
*hwc
;
3564 counter
= kzalloc(sizeof(*counter
), gfpflags
);
3566 return ERR_PTR(-ENOMEM
);
3569 * Single counters are their own group leaders, with an
3570 * empty sibling list:
3573 group_leader
= counter
;
3575 mutex_init(&counter
->child_mutex
);
3576 INIT_LIST_HEAD(&counter
->child_list
);
3578 INIT_LIST_HEAD(&counter
->list_entry
);
3579 INIT_LIST_HEAD(&counter
->event_entry
);
3580 INIT_LIST_HEAD(&counter
->sibling_list
);
3581 init_waitqueue_head(&counter
->waitq
);
3583 mutex_init(&counter
->mmap_mutex
);
3586 counter
->attr
= *attr
;
3587 counter
->group_leader
= group_leader
;
3588 counter
->pmu
= NULL
;
3590 counter
->oncpu
= -1;
3592 counter
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
3593 counter
->id
= atomic64_inc_return(&perf_counter_id
);
3595 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3598 counter
->state
= PERF_COUNTER_STATE_OFF
;
3603 hwc
->sample_period
= attr
->sample_period
;
3604 if (attr
->freq
&& attr
->sample_freq
)
3605 hwc
->sample_period
= 1;
3607 atomic64_set(&hwc
->period_left
, hwc
->sample_period
);
3610 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3612 if (attr
->inherit
&& (attr
->sample_type
& PERF_SAMPLE_GROUP
))
3615 switch (attr
->type
) {
3617 case PERF_TYPE_HARDWARE
:
3618 case PERF_TYPE_HW_CACHE
:
3619 pmu
= hw_perf_counter_init(counter
);
3622 case PERF_TYPE_SOFTWARE
:
3623 pmu
= sw_perf_counter_init(counter
);
3626 case PERF_TYPE_TRACEPOINT
:
3627 pmu
= tp_perf_counter_init(counter
);
3637 else if (IS_ERR(pmu
))
3642 put_pid_ns(counter
->ns
);
3644 return ERR_PTR(err
);
3649 atomic_inc(&nr_counters
);
3650 if (counter
->attr
.mmap
)
3651 atomic_inc(&nr_mmap_counters
);
3652 if (counter
->attr
.comm
)
3653 atomic_inc(&nr_comm_counters
);
3658 static int perf_copy_attr(struct perf_counter_attr __user
*uattr
,
3659 struct perf_counter_attr
*attr
)
3664 if (!access_ok(VERIFY_WRITE
, uattr
, PERF_ATTR_SIZE_VER0
))
3668 * zero the full structure, so that a short copy will be nice.
3670 memset(attr
, 0, sizeof(*attr
));
3672 ret
= get_user(size
, &uattr
->size
);
3676 if (size
> PAGE_SIZE
) /* silly large */
3679 if (!size
) /* abi compat */
3680 size
= PERF_ATTR_SIZE_VER0
;
3682 if (size
< PERF_ATTR_SIZE_VER0
)
3686 * If we're handed a bigger struct than we know of,
3687 * ensure all the unknown bits are 0.
3689 if (size
> sizeof(*attr
)) {
3691 unsigned long __user
*addr
;
3692 unsigned long __user
*end
;
3694 addr
= PTR_ALIGN((void __user
*)uattr
+ sizeof(*attr
),
3695 sizeof(unsigned long));
3696 end
= PTR_ALIGN((void __user
*)uattr
+ size
,
3697 sizeof(unsigned long));
3699 for (; addr
< end
; addr
+= sizeof(unsigned long)) {
3700 ret
= get_user(val
, addr
);
3708 ret
= copy_from_user(attr
, uattr
, size
);
3713 * If the type exists, the corresponding creation will verify
3716 if (attr
->type
>= PERF_TYPE_MAX
)
3719 if (attr
->__reserved_1
|| attr
->__reserved_2
|| attr
->__reserved_3
)
3722 if (attr
->sample_type
& ~(PERF_SAMPLE_MAX
-1))
3725 if (attr
->read_format
& ~(PERF_FORMAT_MAX
-1))
3732 put_user(sizeof(*attr
), &uattr
->size
);
3738 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3740 * @attr_uptr: event type attributes for monitoring/sampling
3743 * @group_fd: group leader counter fd
3745 SYSCALL_DEFINE5(perf_counter_open
,
3746 struct perf_counter_attr __user
*, attr_uptr
,
3747 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
3749 struct perf_counter
*counter
, *group_leader
;
3750 struct perf_counter_attr attr
;
3751 struct perf_counter_context
*ctx
;
3752 struct file
*counter_file
= NULL
;
3753 struct file
*group_file
= NULL
;
3754 int fput_needed
= 0;
3755 int fput_needed2
= 0;
3758 /* for future expandability... */
3762 ret
= perf_copy_attr(attr_uptr
, &attr
);
3766 if (!attr
.exclude_kernel
) {
3767 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN
))
3772 if (attr
.sample_freq
> sysctl_perf_counter_sample_rate
)
3777 * Get the target context (task or percpu):
3779 ctx
= find_get_context(pid
, cpu
);
3781 return PTR_ERR(ctx
);
3784 * Look up the group leader (we will attach this counter to it):
3786 group_leader
= NULL
;
3787 if (group_fd
!= -1) {
3789 group_file
= fget_light(group_fd
, &fput_needed
);
3791 goto err_put_context
;
3792 if (group_file
->f_op
!= &perf_fops
)
3793 goto err_put_context
;
3795 group_leader
= group_file
->private_data
;
3797 * Do not allow a recursive hierarchy (this new sibling
3798 * becoming part of another group-sibling):
3800 if (group_leader
->group_leader
!= group_leader
)
3801 goto err_put_context
;
3803 * Do not allow to attach to a group in a different
3804 * task or CPU context:
3806 if (group_leader
->ctx
!= ctx
)
3807 goto err_put_context
;
3809 * Only a group leader can be exclusive or pinned
3811 if (attr
.exclusive
|| attr
.pinned
)
3812 goto err_put_context
;
3815 counter
= perf_counter_alloc(&attr
, cpu
, ctx
, group_leader
,
3817 ret
= PTR_ERR(counter
);
3818 if (IS_ERR(counter
))
3819 goto err_put_context
;
3821 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
3823 goto err_free_put_context
;
3825 counter_file
= fget_light(ret
, &fput_needed2
);
3827 goto err_free_put_context
;
3829 counter
->filp
= counter_file
;
3830 WARN_ON_ONCE(ctx
->parent_ctx
);
3831 mutex_lock(&ctx
->mutex
);
3832 perf_install_in_context(ctx
, counter
, cpu
);
3834 mutex_unlock(&ctx
->mutex
);
3836 counter
->owner
= current
;
3837 get_task_struct(current
);
3838 mutex_lock(¤t
->perf_counter_mutex
);
3839 list_add_tail(&counter
->owner_entry
, ¤t
->perf_counter_list
);
3840 mutex_unlock(¤t
->perf_counter_mutex
);
3842 fput_light(counter_file
, fput_needed2
);
3845 fput_light(group_file
, fput_needed
);
3849 err_free_put_context
:
3859 * inherit a counter from parent task to child task:
3861 static struct perf_counter
*
3862 inherit_counter(struct perf_counter
*parent_counter
,
3863 struct task_struct
*parent
,
3864 struct perf_counter_context
*parent_ctx
,
3865 struct task_struct
*child
,
3866 struct perf_counter
*group_leader
,
3867 struct perf_counter_context
*child_ctx
)
3869 struct perf_counter
*child_counter
;
3872 * Instead of creating recursive hierarchies of counters,
3873 * we link inherited counters back to the original parent,
3874 * which has a filp for sure, which we use as the reference
3877 if (parent_counter
->parent
)
3878 parent_counter
= parent_counter
->parent
;
3880 child_counter
= perf_counter_alloc(&parent_counter
->attr
,
3881 parent_counter
->cpu
, child_ctx
,
3882 group_leader
, GFP_KERNEL
);
3883 if (IS_ERR(child_counter
))
3884 return child_counter
;
3888 * Make the child state follow the state of the parent counter,
3889 * not its attr.disabled bit. We hold the parent's mutex,
3890 * so we won't race with perf_counter_{en, dis}able_family.
3892 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3893 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3895 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3897 if (parent_counter
->attr
.freq
)
3898 child_counter
->hw
.sample_period
= parent_counter
->hw
.sample_period
;
3901 * Link it up in the child's context:
3903 add_counter_to_ctx(child_counter
, child_ctx
);
3905 child_counter
->parent
= parent_counter
;
3907 * inherit into child's child as well:
3909 child_counter
->attr
.inherit
= 1;
3912 * Get a reference to the parent filp - we will fput it
3913 * when the child counter exits. This is safe to do because
3914 * we are in the parent and we know that the filp still
3915 * exists and has a nonzero count:
3917 atomic_long_inc(&parent_counter
->filp
->f_count
);
3920 * Link this into the parent counter's child list
3922 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3923 mutex_lock(&parent_counter
->child_mutex
);
3924 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3925 mutex_unlock(&parent_counter
->child_mutex
);
3927 return child_counter
;
3930 static int inherit_group(struct perf_counter
*parent_counter
,
3931 struct task_struct
*parent
,
3932 struct perf_counter_context
*parent_ctx
,
3933 struct task_struct
*child
,
3934 struct perf_counter_context
*child_ctx
)
3936 struct perf_counter
*leader
;
3937 struct perf_counter
*sub
;
3938 struct perf_counter
*child_ctr
;
3940 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3941 child
, NULL
, child_ctx
);
3943 return PTR_ERR(leader
);
3944 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3945 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3946 child
, leader
, child_ctx
);
3947 if (IS_ERR(child_ctr
))
3948 return PTR_ERR(child_ctr
);
3953 static void sync_child_counter(struct perf_counter
*child_counter
,
3954 struct perf_counter
*parent_counter
)
3958 child_val
= atomic64_read(&child_counter
->count
);
3961 * Add back the child's count to the parent's count:
3963 atomic64_add(child_val
, &parent_counter
->count
);
3964 atomic64_add(child_counter
->total_time_enabled
,
3965 &parent_counter
->child_total_time_enabled
);
3966 atomic64_add(child_counter
->total_time_running
,
3967 &parent_counter
->child_total_time_running
);
3970 * Remove this counter from the parent's list
3972 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3973 mutex_lock(&parent_counter
->child_mutex
);
3974 list_del_init(&child_counter
->child_list
);
3975 mutex_unlock(&parent_counter
->child_mutex
);
3978 * Release the parent counter, if this was the last
3981 fput(parent_counter
->filp
);
3985 __perf_counter_exit_task(struct perf_counter
*child_counter
,
3986 struct perf_counter_context
*child_ctx
)
3988 struct perf_counter
*parent_counter
;
3990 update_counter_times(child_counter
);
3991 perf_counter_remove_from_context(child_counter
);
3993 parent_counter
= child_counter
->parent
;
3995 * It can happen that parent exits first, and has counters
3996 * that are still around due to the child reference. These
3997 * counters need to be zapped - but otherwise linger.
3999 if (parent_counter
) {
4000 sync_child_counter(child_counter
, parent_counter
);
4001 free_counter(child_counter
);
4006 * When a child task exits, feed back counter values to parent counters.
4008 void perf_counter_exit_task(struct task_struct
*child
)
4010 struct perf_counter
*child_counter
, *tmp
;
4011 struct perf_counter_context
*child_ctx
;
4012 unsigned long flags
;
4014 if (likely(!child
->perf_counter_ctxp
))
4017 local_irq_save(flags
);
4019 * We can't reschedule here because interrupts are disabled,
4020 * and either child is current or it is a task that can't be
4021 * scheduled, so we are now safe from rescheduling changing
4024 child_ctx
= child
->perf_counter_ctxp
;
4025 __perf_counter_task_sched_out(child_ctx
);
4028 * Take the context lock here so that if find_get_context is
4029 * reading child->perf_counter_ctxp, we wait until it has
4030 * incremented the context's refcount before we do put_ctx below.
4032 spin_lock(&child_ctx
->lock
);
4033 child
->perf_counter_ctxp
= NULL
;
4034 if (child_ctx
->parent_ctx
) {
4036 * This context is a clone; unclone it so it can't get
4037 * swapped to another process while we're removing all
4038 * the counters from it.
4040 put_ctx(child_ctx
->parent_ctx
);
4041 child_ctx
->parent_ctx
= NULL
;
4043 spin_unlock(&child_ctx
->lock
);
4044 local_irq_restore(flags
);
4047 * We can recurse on the same lock type through:
4049 * __perf_counter_exit_task()
4050 * sync_child_counter()
4051 * fput(parent_counter->filp)
4053 * mutex_lock(&ctx->mutex)
4055 * But since its the parent context it won't be the same instance.
4057 mutex_lock_nested(&child_ctx
->mutex
, SINGLE_DEPTH_NESTING
);
4060 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
4062 __perf_counter_exit_task(child_counter
, child_ctx
);
4065 * If the last counter was a group counter, it will have appended all
4066 * its siblings to the list, but we obtained 'tmp' before that which
4067 * will still point to the list head terminating the iteration.
4069 if (!list_empty(&child_ctx
->counter_list
))
4072 mutex_unlock(&child_ctx
->mutex
);
4078 * free an unexposed, unused context as created by inheritance by
4079 * init_task below, used by fork() in case of fail.
4081 void perf_counter_free_task(struct task_struct
*task
)
4083 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
4084 struct perf_counter
*counter
, *tmp
;
4089 mutex_lock(&ctx
->mutex
);
4091 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
) {
4092 struct perf_counter
*parent
= counter
->parent
;
4094 if (WARN_ON_ONCE(!parent
))
4097 mutex_lock(&parent
->child_mutex
);
4098 list_del_init(&counter
->child_list
);
4099 mutex_unlock(&parent
->child_mutex
);
4103 list_del_counter(counter
, ctx
);
4104 free_counter(counter
);
4107 if (!list_empty(&ctx
->counter_list
))
4110 mutex_unlock(&ctx
->mutex
);
4116 * Initialize the perf_counter context in task_struct
4118 int perf_counter_init_task(struct task_struct
*child
)
4120 struct perf_counter_context
*child_ctx
, *parent_ctx
;
4121 struct perf_counter_context
*cloned_ctx
;
4122 struct perf_counter
*counter
;
4123 struct task_struct
*parent
= current
;
4124 int inherited_all
= 1;
4127 child
->perf_counter_ctxp
= NULL
;
4129 mutex_init(&child
->perf_counter_mutex
);
4130 INIT_LIST_HEAD(&child
->perf_counter_list
);
4132 if (likely(!parent
->perf_counter_ctxp
))
4136 * This is executed from the parent task context, so inherit
4137 * counters that have been marked for cloning.
4138 * First allocate and initialize a context for the child.
4141 child_ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
4145 __perf_counter_init_context(child_ctx
, child
);
4146 child
->perf_counter_ctxp
= child_ctx
;
4147 get_task_struct(child
);
4150 * If the parent's context is a clone, pin it so it won't get
4153 parent_ctx
= perf_pin_task_context(parent
);
4156 * No need to check if parent_ctx != NULL here; since we saw
4157 * it non-NULL earlier, the only reason for it to become NULL
4158 * is if we exit, and since we're currently in the middle of
4159 * a fork we can't be exiting at the same time.
4163 * Lock the parent list. No need to lock the child - not PID
4164 * hashed yet and not running, so nobody can access it.
4166 mutex_lock(&parent_ctx
->mutex
);
4169 * We dont have to disable NMIs - we are only looking at
4170 * the list, not manipulating it:
4172 list_for_each_entry_rcu(counter
, &parent_ctx
->event_list
, event_entry
) {
4173 if (counter
!= counter
->group_leader
)
4176 if (!counter
->attr
.inherit
) {
4181 ret
= inherit_group(counter
, parent
, parent_ctx
,
4189 if (inherited_all
) {
4191 * Mark the child context as a clone of the parent
4192 * context, or of whatever the parent is a clone of.
4193 * Note that if the parent is a clone, it could get
4194 * uncloned at any point, but that doesn't matter
4195 * because the list of counters and the generation
4196 * count can't have changed since we took the mutex.
4198 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
4200 child_ctx
->parent_ctx
= cloned_ctx
;
4201 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
4203 child_ctx
->parent_ctx
= parent_ctx
;
4204 child_ctx
->parent_gen
= parent_ctx
->generation
;
4206 get_ctx(child_ctx
->parent_ctx
);
4209 mutex_unlock(&parent_ctx
->mutex
);
4211 perf_unpin_context(parent_ctx
);
4216 static void __cpuinit
perf_counter_init_cpu(int cpu
)
4218 struct perf_cpu_context
*cpuctx
;
4220 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4221 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
4223 spin_lock(&perf_resource_lock
);
4224 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
4225 spin_unlock(&perf_resource_lock
);
4227 hw_perf_counter_setup(cpu
);
4230 #ifdef CONFIG_HOTPLUG_CPU
4231 static void __perf_counter_exit_cpu(void *info
)
4233 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
4234 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4235 struct perf_counter
*counter
, *tmp
;
4237 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
4238 __perf_counter_remove_from_context(counter
);
4240 static void perf_counter_exit_cpu(int cpu
)
4242 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4243 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
4245 mutex_lock(&ctx
->mutex
);
4246 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
4247 mutex_unlock(&ctx
->mutex
);
4250 static inline void perf_counter_exit_cpu(int cpu
) { }
4253 static int __cpuinit
4254 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
4256 unsigned int cpu
= (long)hcpu
;
4260 case CPU_UP_PREPARE
:
4261 case CPU_UP_PREPARE_FROZEN
:
4262 perf_counter_init_cpu(cpu
);
4265 case CPU_DOWN_PREPARE
:
4266 case CPU_DOWN_PREPARE_FROZEN
:
4267 perf_counter_exit_cpu(cpu
);
4278 * This has to have a higher priority than migration_notifier in sched.c.
4280 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
4281 .notifier_call
= perf_cpu_notify
,
4285 void __init
perf_counter_init(void)
4287 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
4288 (void *)(long)smp_processor_id());
4289 register_cpu_notifier(&perf_cpu_nb
);
4292 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
4294 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
4298 perf_set_reserve_percpu(struct sysdev_class
*class,
4302 struct perf_cpu_context
*cpuctx
;
4306 err
= strict_strtoul(buf
, 10, &val
);
4309 if (val
> perf_max_counters
)
4312 spin_lock(&perf_resource_lock
);
4313 perf_reserved_percpu
= val
;
4314 for_each_online_cpu(cpu
) {
4315 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
4316 spin_lock_irq(&cpuctx
->ctx
.lock
);
4317 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
4318 perf_max_counters
- perf_reserved_percpu
);
4319 cpuctx
->max_pertask
= mpt
;
4320 spin_unlock_irq(&cpuctx
->ctx
.lock
);
4322 spin_unlock(&perf_resource_lock
);
4327 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
4329 return sprintf(buf
, "%d\n", perf_overcommit
);
4333 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
4338 err
= strict_strtoul(buf
, 10, &val
);
4344 spin_lock(&perf_resource_lock
);
4345 perf_overcommit
= val
;
4346 spin_unlock(&perf_resource_lock
);
4351 static SYSDEV_CLASS_ATTR(
4354 perf_show_reserve_percpu
,
4355 perf_set_reserve_percpu
4358 static SYSDEV_CLASS_ATTR(
4361 perf_show_overcommit
,
4365 static struct attribute
*perfclass_attrs
[] = {
4366 &attr_reserve_percpu
.attr
,
4367 &attr_overcommit
.attr
,
4371 static struct attribute_group perfclass_attr_group
= {
4372 .attrs
= perfclass_attrs
,
4373 .name
= "perf_counters",
4376 static int __init
perf_counter_sysfs_init(void)
4378 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
4379 &perfclass_attr_group
);
4381 device_initcall(perf_counter_sysfs_init
);