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_tracking __read_mostly
;
44 static atomic_t nr_munmap_tracking __read_mostly
;
45 static atomic_t nr_comm_tracking __read_mostly
;
47 int sysctl_perf_counter_priv __read_mostly
; /* do we need to be privileged */
48 int sysctl_perf_counter_mlock __read_mostly
= 512; /* 'free' kb per user */
49 int sysctl_perf_counter_limit __read_mostly
= 100000; /* max NMIs per second */
52 * Lock for (sysadmin-configurable) counter reservations:
54 static DEFINE_SPINLOCK(perf_resource_lock
);
57 * Architecture provided APIs - weak aliases:
59 extern __weak
const struct pmu
*hw_perf_counter_init(struct perf_counter
*counter
)
64 void __weak
hw_perf_disable(void) { barrier(); }
65 void __weak
hw_perf_enable(void) { barrier(); }
67 void __weak
hw_perf_counter_setup(int cpu
) { barrier(); }
70 hw_perf_group_sched_in(struct perf_counter
*group_leader
,
71 struct perf_cpu_context
*cpuctx
,
72 struct perf_counter_context
*ctx
, int cpu
)
77 void __weak
perf_counter_print_debug(void) { }
79 static DEFINE_PER_CPU(int, disable_count
);
81 void __perf_disable(void)
83 __get_cpu_var(disable_count
)++;
86 bool __perf_enable(void)
88 return !--__get_cpu_var(disable_count
);
91 void perf_disable(void)
97 void perf_enable(void)
103 static void get_ctx(struct perf_counter_context
*ctx
)
105 atomic_inc(&ctx
->refcount
);
108 static void free_ctx(struct rcu_head
*head
)
110 struct perf_counter_context
*ctx
;
112 ctx
= container_of(head
, struct perf_counter_context
, rcu_head
);
116 static void put_ctx(struct perf_counter_context
*ctx
)
118 if (atomic_dec_and_test(&ctx
->refcount
)) {
120 put_ctx(ctx
->parent_ctx
);
122 put_task_struct(ctx
->task
);
123 call_rcu(&ctx
->rcu_head
, free_ctx
);
128 * Get the perf_counter_context for a task and lock it.
129 * This has to cope with with the fact that until it is locked,
130 * the context could get moved to another task.
132 static struct perf_counter_context
*
133 perf_lock_task_context(struct task_struct
*task
, unsigned long *flags
)
135 struct perf_counter_context
*ctx
;
139 ctx
= rcu_dereference(task
->perf_counter_ctxp
);
142 * If this context is a clone of another, it might
143 * get swapped for another underneath us by
144 * perf_counter_task_sched_out, though the
145 * rcu_read_lock() protects us from any context
146 * getting freed. Lock the context and check if it
147 * got swapped before we could get the lock, and retry
148 * if so. If we locked the right context, then it
149 * can't get swapped on us any more.
151 spin_lock_irqsave(&ctx
->lock
, *flags
);
152 if (ctx
!= rcu_dereference(task
->perf_counter_ctxp
)) {
153 spin_unlock_irqrestore(&ctx
->lock
, *flags
);
162 * Get the context for a task and increment its pin_count so it
163 * can't get swapped to another task. This also increments its
164 * reference count so that the context can't get freed.
166 static struct perf_counter_context
*perf_pin_task_context(struct task_struct
*task
)
168 struct perf_counter_context
*ctx
;
171 ctx
= perf_lock_task_context(task
, &flags
);
175 spin_unlock_irqrestore(&ctx
->lock
, flags
);
180 static void perf_unpin_context(struct perf_counter_context
*ctx
)
184 spin_lock_irqsave(&ctx
->lock
, flags
);
186 spin_unlock_irqrestore(&ctx
->lock
, flags
);
191 * Add a counter from the lists for its context.
192 * Must be called with ctx->mutex and ctx->lock held.
195 list_add_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
197 struct perf_counter
*group_leader
= counter
->group_leader
;
200 * Depending on whether it is a standalone or sibling counter,
201 * add it straight to the context's counter list, or to the group
202 * leader's sibling list:
204 if (group_leader
== counter
)
205 list_add_tail(&counter
->list_entry
, &ctx
->counter_list
);
207 list_add_tail(&counter
->list_entry
, &group_leader
->sibling_list
);
208 group_leader
->nr_siblings
++;
211 list_add_rcu(&counter
->event_entry
, &ctx
->event_list
);
216 * Remove a counter from the lists for its context.
217 * Must be called with ctx->mutex and ctx->lock held.
220 list_del_counter(struct perf_counter
*counter
, struct perf_counter_context
*ctx
)
222 struct perf_counter
*sibling
, *tmp
;
224 if (list_empty(&counter
->list_entry
))
228 list_del_init(&counter
->list_entry
);
229 list_del_rcu(&counter
->event_entry
);
231 if (counter
->group_leader
!= counter
)
232 counter
->group_leader
->nr_siblings
--;
235 * If this was a group counter with sibling counters then
236 * upgrade the siblings to singleton counters by adding them
237 * to the context list directly:
239 list_for_each_entry_safe(sibling
, tmp
,
240 &counter
->sibling_list
, list_entry
) {
242 list_move_tail(&sibling
->list_entry
, &ctx
->counter_list
);
243 sibling
->group_leader
= sibling
;
248 counter_sched_out(struct perf_counter
*counter
,
249 struct perf_cpu_context
*cpuctx
,
250 struct perf_counter_context
*ctx
)
252 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
255 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
256 counter
->tstamp_stopped
= ctx
->time
;
257 counter
->pmu
->disable(counter
);
260 if (!is_software_counter(counter
))
261 cpuctx
->active_oncpu
--;
263 if (counter
->hw_event
.exclusive
|| !cpuctx
->active_oncpu
)
264 cpuctx
->exclusive
= 0;
268 group_sched_out(struct perf_counter
*group_counter
,
269 struct perf_cpu_context
*cpuctx
,
270 struct perf_counter_context
*ctx
)
272 struct perf_counter
*counter
;
274 if (group_counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
277 counter_sched_out(group_counter
, cpuctx
, ctx
);
280 * Schedule out siblings (if any):
282 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
)
283 counter_sched_out(counter
, cpuctx
, ctx
);
285 if (group_counter
->hw_event
.exclusive
)
286 cpuctx
->exclusive
= 0;
290 * Cross CPU call to remove a performance counter
292 * We disable the counter on the hardware level first. After that we
293 * remove it from the context list.
295 static void __perf_counter_remove_from_context(void *info
)
297 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
298 struct perf_counter
*counter
= info
;
299 struct perf_counter_context
*ctx
= counter
->ctx
;
302 * If this is a task context, we need to check whether it is
303 * the current task context of this cpu. If not it has been
304 * scheduled out before the smp call arrived.
306 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
309 spin_lock(&ctx
->lock
);
311 * Protect the list operation against NMI by disabling the
312 * counters on a global level.
316 counter_sched_out(counter
, cpuctx
, ctx
);
318 list_del_counter(counter
, ctx
);
322 * Allow more per task counters with respect to the
325 cpuctx
->max_pertask
=
326 min(perf_max_counters
- ctx
->nr_counters
,
327 perf_max_counters
- perf_reserved_percpu
);
331 spin_unlock(&ctx
->lock
);
336 * Remove the counter from a task's (or a CPU's) list of counters.
338 * Must be called with ctx->mutex held.
340 * CPU counters are removed with a smp call. For task counters we only
341 * call when the task is on a CPU.
343 * If counter->ctx is a cloned context, callers must make sure that
344 * every task struct that counter->ctx->task could possibly point to
345 * remains valid. This is OK when called from perf_release since
346 * that only calls us on the top-level context, which can't be a clone.
347 * When called from perf_counter_exit_task, it's OK because the
348 * context has been detached from its task.
350 static void perf_counter_remove_from_context(struct perf_counter
*counter
)
352 struct perf_counter_context
*ctx
= counter
->ctx
;
353 struct task_struct
*task
= ctx
->task
;
357 * Per cpu counters are removed via an smp call and
358 * the removal is always sucessful.
360 smp_call_function_single(counter
->cpu
,
361 __perf_counter_remove_from_context
,
367 task_oncpu_function_call(task
, __perf_counter_remove_from_context
,
370 spin_lock_irq(&ctx
->lock
);
372 * If the context is active we need to retry the smp call.
374 if (ctx
->nr_active
&& !list_empty(&counter
->list_entry
)) {
375 spin_unlock_irq(&ctx
->lock
);
380 * The lock prevents that this context is scheduled in so we
381 * can remove the counter safely, if the call above did not
384 if (!list_empty(&counter
->list_entry
)) {
385 list_del_counter(counter
, ctx
);
387 spin_unlock_irq(&ctx
->lock
);
390 static inline u64
perf_clock(void)
392 return cpu_clock(smp_processor_id());
396 * Update the record of the current time in a context.
398 static void update_context_time(struct perf_counter_context
*ctx
)
400 u64 now
= perf_clock();
402 ctx
->time
+= now
- ctx
->timestamp
;
403 ctx
->timestamp
= now
;
407 * Update the total_time_enabled and total_time_running fields for a counter.
409 static void update_counter_times(struct perf_counter
*counter
)
411 struct perf_counter_context
*ctx
= counter
->ctx
;
414 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
)
417 counter
->total_time_enabled
= ctx
->time
- counter
->tstamp_enabled
;
419 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
)
420 run_end
= counter
->tstamp_stopped
;
424 counter
->total_time_running
= run_end
- counter
->tstamp_running
;
428 * Update total_time_enabled and total_time_running for all counters in a group.
430 static void update_group_times(struct perf_counter
*leader
)
432 struct perf_counter
*counter
;
434 update_counter_times(leader
);
435 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
436 update_counter_times(counter
);
440 * Cross CPU call to disable a performance counter
442 static void __perf_counter_disable(void *info
)
444 struct perf_counter
*counter
= info
;
445 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
446 struct perf_counter_context
*ctx
= counter
->ctx
;
449 * If this is a per-task counter, need to check whether this
450 * counter's task is the current task on this cpu.
452 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
)
455 spin_lock(&ctx
->lock
);
458 * If the counter is on, turn it off.
459 * If it is in error state, leave it in error state.
461 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
) {
462 update_context_time(ctx
);
463 update_counter_times(counter
);
464 if (counter
== counter
->group_leader
)
465 group_sched_out(counter
, cpuctx
, ctx
);
467 counter_sched_out(counter
, cpuctx
, ctx
);
468 counter
->state
= PERF_COUNTER_STATE_OFF
;
471 spin_unlock(&ctx
->lock
);
477 * If counter->ctx is a cloned context, callers must make sure that
478 * every task struct that counter->ctx->task could possibly point to
479 * remains valid. This condition is satisifed when called through
480 * perf_counter_for_each_child or perf_counter_for_each because they
481 * hold the top-level counter's child_mutex, so any descendant that
482 * goes to exit will block in sync_child_counter.
483 * When called from perf_pending_counter it's OK because counter->ctx
484 * is the current context on this CPU and preemption is disabled,
485 * hence we can't get into perf_counter_task_sched_out for this context.
487 static void perf_counter_disable(struct perf_counter
*counter
)
489 struct perf_counter_context
*ctx
= counter
->ctx
;
490 struct task_struct
*task
= ctx
->task
;
494 * Disable the counter on the cpu that it's on
496 smp_call_function_single(counter
->cpu
, __perf_counter_disable
,
502 task_oncpu_function_call(task
, __perf_counter_disable
, counter
);
504 spin_lock_irq(&ctx
->lock
);
506 * If the counter is still active, we need to retry the cross-call.
508 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
509 spin_unlock_irq(&ctx
->lock
);
514 * Since we have the lock this context can't be scheduled
515 * in, so we can change the state safely.
517 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
518 update_counter_times(counter
);
519 counter
->state
= PERF_COUNTER_STATE_OFF
;
522 spin_unlock_irq(&ctx
->lock
);
526 counter_sched_in(struct perf_counter
*counter
,
527 struct perf_cpu_context
*cpuctx
,
528 struct perf_counter_context
*ctx
,
531 if (counter
->state
<= PERF_COUNTER_STATE_OFF
)
534 counter
->state
= PERF_COUNTER_STATE_ACTIVE
;
535 counter
->oncpu
= cpu
; /* TODO: put 'cpu' into cpuctx->cpu */
537 * The new state must be visible before we turn it on in the hardware:
541 if (counter
->pmu
->enable(counter
)) {
542 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
547 counter
->tstamp_running
+= ctx
->time
- counter
->tstamp_stopped
;
549 if (!is_software_counter(counter
))
550 cpuctx
->active_oncpu
++;
553 if (counter
->hw_event
.exclusive
)
554 cpuctx
->exclusive
= 1;
560 group_sched_in(struct perf_counter
*group_counter
,
561 struct perf_cpu_context
*cpuctx
,
562 struct perf_counter_context
*ctx
,
565 struct perf_counter
*counter
, *partial_group
;
568 if (group_counter
->state
== PERF_COUNTER_STATE_OFF
)
571 ret
= hw_perf_group_sched_in(group_counter
, cpuctx
, ctx
, cpu
);
573 return ret
< 0 ? ret
: 0;
575 if (counter_sched_in(group_counter
, cpuctx
, ctx
, cpu
))
579 * Schedule in siblings as one group (if any):
581 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
582 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
)) {
583 partial_group
= counter
;
592 * Groups can be scheduled in as one unit only, so undo any
593 * partial group before returning:
595 list_for_each_entry(counter
, &group_counter
->sibling_list
, list_entry
) {
596 if (counter
== partial_group
)
598 counter_sched_out(counter
, cpuctx
, ctx
);
600 counter_sched_out(group_counter
, cpuctx
, ctx
);
606 * Return 1 for a group consisting entirely of software counters,
607 * 0 if the group contains any hardware counters.
609 static int is_software_only_group(struct perf_counter
*leader
)
611 struct perf_counter
*counter
;
613 if (!is_software_counter(leader
))
616 list_for_each_entry(counter
, &leader
->sibling_list
, list_entry
)
617 if (!is_software_counter(counter
))
624 * Work out whether we can put this counter group on the CPU now.
626 static int group_can_go_on(struct perf_counter
*counter
,
627 struct perf_cpu_context
*cpuctx
,
631 * Groups consisting entirely of software counters can always go on.
633 if (is_software_only_group(counter
))
636 * If an exclusive group is already on, no other hardware
637 * counters can go on.
639 if (cpuctx
->exclusive
)
642 * If this group is exclusive and there are already
643 * counters on the CPU, it can't go on.
645 if (counter
->hw_event
.exclusive
&& cpuctx
->active_oncpu
)
648 * Otherwise, try to add it if all previous groups were able
654 static void add_counter_to_ctx(struct perf_counter
*counter
,
655 struct perf_counter_context
*ctx
)
657 list_add_counter(counter
, ctx
);
658 counter
->tstamp_enabled
= ctx
->time
;
659 counter
->tstamp_running
= ctx
->time
;
660 counter
->tstamp_stopped
= ctx
->time
;
664 * Cross CPU call to install and enable a performance counter
666 * Must be called with ctx->mutex held
668 static void __perf_install_in_context(void *info
)
670 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
671 struct perf_counter
*counter
= info
;
672 struct perf_counter_context
*ctx
= counter
->ctx
;
673 struct perf_counter
*leader
= counter
->group_leader
;
674 int cpu
= smp_processor_id();
678 * If this is a task context, we need to check whether it is
679 * the current task context of this cpu. If not it has been
680 * scheduled out before the smp call arrived.
681 * Or possibly this is the right context but it isn't
682 * on this cpu because it had no counters.
684 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
685 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
687 cpuctx
->task_ctx
= ctx
;
690 spin_lock(&ctx
->lock
);
692 update_context_time(ctx
);
695 * Protect the list operation against NMI by disabling the
696 * counters on a global level. NOP for non NMI based counters.
700 add_counter_to_ctx(counter
, ctx
);
703 * Don't put the counter on if it is disabled or if
704 * it is in a group and the group isn't on.
706 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
||
707 (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
))
711 * An exclusive counter can't go on if there are already active
712 * hardware counters, and no hardware counter can go on if there
713 * is already an exclusive counter on.
715 if (!group_can_go_on(counter
, cpuctx
, 1))
718 err
= counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
722 * This counter couldn't go on. If it is in a group
723 * then we have to pull the whole group off.
724 * If the counter group is pinned then put it in error state.
726 if (leader
!= counter
)
727 group_sched_out(leader
, cpuctx
, ctx
);
728 if (leader
->hw_event
.pinned
) {
729 update_group_times(leader
);
730 leader
->state
= PERF_COUNTER_STATE_ERROR
;
734 if (!err
&& !ctx
->task
&& cpuctx
->max_pertask
)
735 cpuctx
->max_pertask
--;
740 spin_unlock(&ctx
->lock
);
744 * Attach a performance counter to a context
746 * First we add the counter to the list with the hardware enable bit
747 * in counter->hw_config cleared.
749 * If the counter is attached to a task which is on a CPU we use a smp
750 * call to enable it in the task context. The task might have been
751 * scheduled away, but we check this in the smp call again.
753 * Must be called with ctx->mutex held.
756 perf_install_in_context(struct perf_counter_context
*ctx
,
757 struct perf_counter
*counter
,
760 struct task_struct
*task
= ctx
->task
;
764 * Per cpu counters are installed via an smp call and
765 * the install is always sucessful.
767 smp_call_function_single(cpu
, __perf_install_in_context
,
773 task_oncpu_function_call(task
, __perf_install_in_context
,
776 spin_lock_irq(&ctx
->lock
);
778 * we need to retry the smp call.
780 if (ctx
->is_active
&& list_empty(&counter
->list_entry
)) {
781 spin_unlock_irq(&ctx
->lock
);
786 * The lock prevents that this context is scheduled in so we
787 * can add the counter safely, if it the call above did not
790 if (list_empty(&counter
->list_entry
))
791 add_counter_to_ctx(counter
, ctx
);
792 spin_unlock_irq(&ctx
->lock
);
796 * Cross CPU call to enable a performance counter
798 static void __perf_counter_enable(void *info
)
800 struct perf_counter
*counter
= info
;
801 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
802 struct perf_counter_context
*ctx
= counter
->ctx
;
803 struct perf_counter
*leader
= counter
->group_leader
;
807 * If this is a per-task counter, need to check whether this
808 * counter's task is the current task on this cpu.
810 if (ctx
->task
&& cpuctx
->task_ctx
!= ctx
) {
811 if (cpuctx
->task_ctx
|| ctx
->task
!= current
)
813 cpuctx
->task_ctx
= ctx
;
816 spin_lock(&ctx
->lock
);
818 update_context_time(ctx
);
820 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
822 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
823 counter
->tstamp_enabled
= ctx
->time
- counter
->total_time_enabled
;
826 * If the counter is in a group and isn't the group leader,
827 * then don't put it on unless the group is on.
829 if (leader
!= counter
&& leader
->state
!= PERF_COUNTER_STATE_ACTIVE
)
832 if (!group_can_go_on(counter
, cpuctx
, 1)) {
836 if (counter
== leader
)
837 err
= group_sched_in(counter
, cpuctx
, ctx
,
840 err
= counter_sched_in(counter
, cpuctx
, ctx
,
847 * If this counter can't go on and it's part of a
848 * group, then the whole group has to come off.
850 if (leader
!= counter
)
851 group_sched_out(leader
, cpuctx
, ctx
);
852 if (leader
->hw_event
.pinned
) {
853 update_group_times(leader
);
854 leader
->state
= PERF_COUNTER_STATE_ERROR
;
859 spin_unlock(&ctx
->lock
);
865 * If counter->ctx is a cloned context, callers must make sure that
866 * every task struct that counter->ctx->task could possibly point to
867 * remains valid. This condition is satisfied when called through
868 * perf_counter_for_each_child or perf_counter_for_each as described
869 * for perf_counter_disable.
871 static void perf_counter_enable(struct perf_counter
*counter
)
873 struct perf_counter_context
*ctx
= counter
->ctx
;
874 struct task_struct
*task
= ctx
->task
;
878 * Enable the counter on the cpu that it's on
880 smp_call_function_single(counter
->cpu
, __perf_counter_enable
,
885 spin_lock_irq(&ctx
->lock
);
886 if (counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
890 * If the counter is in error state, clear that first.
891 * That way, if we see the counter in error state below, we
892 * know that it has gone back into error state, as distinct
893 * from the task having been scheduled away before the
894 * cross-call arrived.
896 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
897 counter
->state
= PERF_COUNTER_STATE_OFF
;
900 spin_unlock_irq(&ctx
->lock
);
901 task_oncpu_function_call(task
, __perf_counter_enable
, counter
);
903 spin_lock_irq(&ctx
->lock
);
906 * If the context is active and the counter is still off,
907 * we need to retry the cross-call.
909 if (ctx
->is_active
&& counter
->state
== PERF_COUNTER_STATE_OFF
)
913 * Since we have the lock this context can't be scheduled
914 * in, so we can change the state safely.
916 if (counter
->state
== PERF_COUNTER_STATE_OFF
) {
917 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
918 counter
->tstamp_enabled
=
919 ctx
->time
- counter
->total_time_enabled
;
922 spin_unlock_irq(&ctx
->lock
);
925 static int perf_counter_refresh(struct perf_counter
*counter
, int refresh
)
928 * not supported on inherited counters
930 if (counter
->hw_event
.inherit
)
933 atomic_add(refresh
, &counter
->event_limit
);
934 perf_counter_enable(counter
);
939 void __perf_counter_sched_out(struct perf_counter_context
*ctx
,
940 struct perf_cpu_context
*cpuctx
)
942 struct perf_counter
*counter
;
944 spin_lock(&ctx
->lock
);
946 if (likely(!ctx
->nr_counters
))
948 update_context_time(ctx
);
951 if (ctx
->nr_active
) {
952 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
953 if (counter
!= counter
->group_leader
)
954 counter_sched_out(counter
, cpuctx
, ctx
);
956 group_sched_out(counter
, cpuctx
, ctx
);
961 spin_unlock(&ctx
->lock
);
965 * Test whether two contexts are equivalent, i.e. whether they
966 * have both been cloned from the same version of the same context
967 * and they both have the same number of enabled counters.
968 * If the number of enabled counters is the same, then the set
969 * of enabled counters should be the same, because these are both
970 * inherited contexts, therefore we can't access individual counters
971 * in them directly with an fd; we can only enable/disable all
972 * counters via prctl, or enable/disable all counters in a family
973 * via ioctl, which will have the same effect on both contexts.
975 static int context_equiv(struct perf_counter_context
*ctx1
,
976 struct perf_counter_context
*ctx2
)
978 return ctx1
->parent_ctx
&& ctx1
->parent_ctx
== ctx2
->parent_ctx
979 && ctx1
->parent_gen
== ctx2
->parent_gen
980 && !ctx1
->pin_count
&& !ctx2
->pin_count
;
984 * Called from scheduler to remove the counters of the current task,
985 * with interrupts disabled.
987 * We stop each counter and update the counter value in counter->count.
989 * This does not protect us against NMI, but disable()
990 * sets the disabled bit in the control field of counter _before_
991 * accessing the counter control register. If a NMI hits, then it will
992 * not restart the counter.
994 void perf_counter_task_sched_out(struct task_struct
*task
,
995 struct task_struct
*next
, int cpu
)
997 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
998 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
999 struct perf_counter_context
*next_ctx
;
1000 struct perf_counter_context
*parent
;
1001 struct pt_regs
*regs
;
1004 regs
= task_pt_regs(task
);
1005 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES
, 1, 1, regs
, 0);
1007 if (likely(!ctx
|| !cpuctx
->task_ctx
))
1010 update_context_time(ctx
);
1013 parent
= rcu_dereference(ctx
->parent_ctx
);
1014 next_ctx
= next
->perf_counter_ctxp
;
1015 if (parent
&& next_ctx
&&
1016 rcu_dereference(next_ctx
->parent_ctx
) == parent
) {
1018 * Looks like the two contexts are clones, so we might be
1019 * able to optimize the context switch. We lock both
1020 * contexts and check that they are clones under the
1021 * lock (including re-checking that neither has been
1022 * uncloned in the meantime). It doesn't matter which
1023 * order we take the locks because no other cpu could
1024 * be trying to lock both of these tasks.
1026 spin_lock(&ctx
->lock
);
1027 spin_lock_nested(&next_ctx
->lock
, SINGLE_DEPTH_NESTING
);
1028 if (context_equiv(ctx
, next_ctx
)) {
1030 * XXX do we need a memory barrier of sorts
1031 * wrt to rcu_dereference() of perf_counter_ctxp
1033 task
->perf_counter_ctxp
= next_ctx
;
1034 next
->perf_counter_ctxp
= ctx
;
1036 next_ctx
->task
= task
;
1039 spin_unlock(&next_ctx
->lock
);
1040 spin_unlock(&ctx
->lock
);
1045 __perf_counter_sched_out(ctx
, cpuctx
);
1046 cpuctx
->task_ctx
= NULL
;
1051 * Called with IRQs disabled
1053 static void __perf_counter_task_sched_out(struct perf_counter_context
*ctx
)
1055 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
1057 if (!cpuctx
->task_ctx
)
1060 if (WARN_ON_ONCE(ctx
!= cpuctx
->task_ctx
))
1063 __perf_counter_sched_out(ctx
, cpuctx
);
1064 cpuctx
->task_ctx
= NULL
;
1068 * Called with IRQs disabled
1070 static void perf_counter_cpu_sched_out(struct perf_cpu_context
*cpuctx
)
1072 __perf_counter_sched_out(&cpuctx
->ctx
, cpuctx
);
1076 __perf_counter_sched_in(struct perf_counter_context
*ctx
,
1077 struct perf_cpu_context
*cpuctx
, int cpu
)
1079 struct perf_counter
*counter
;
1082 spin_lock(&ctx
->lock
);
1084 if (likely(!ctx
->nr_counters
))
1087 ctx
->timestamp
= perf_clock();
1092 * First go through the list and put on any pinned groups
1093 * in order to give them the best chance of going on.
1095 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1096 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1097 !counter
->hw_event
.pinned
)
1099 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1102 if (counter
!= counter
->group_leader
)
1103 counter_sched_in(counter
, cpuctx
, ctx
, cpu
);
1105 if (group_can_go_on(counter
, cpuctx
, 1))
1106 group_sched_in(counter
, cpuctx
, ctx
, cpu
);
1110 * If this pinned group hasn't been scheduled,
1111 * put it in error state.
1113 if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1114 update_group_times(counter
);
1115 counter
->state
= PERF_COUNTER_STATE_ERROR
;
1119 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1121 * Ignore counters in OFF or ERROR state, and
1122 * ignore pinned counters since we did them already.
1124 if (counter
->state
<= PERF_COUNTER_STATE_OFF
||
1125 counter
->hw_event
.pinned
)
1129 * Listen to the 'cpu' scheduling filter constraint
1132 if (counter
->cpu
!= -1 && counter
->cpu
!= cpu
)
1135 if (counter
!= counter
->group_leader
) {
1136 if (counter_sched_in(counter
, cpuctx
, ctx
, cpu
))
1139 if (group_can_go_on(counter
, cpuctx
, can_add_hw
)) {
1140 if (group_sched_in(counter
, cpuctx
, ctx
, cpu
))
1147 spin_unlock(&ctx
->lock
);
1151 * Called from scheduler to add the counters of the current task
1152 * with interrupts disabled.
1154 * We restore the counter value and then enable it.
1156 * This does not protect us against NMI, but enable()
1157 * sets the enabled bit in the control field of counter _before_
1158 * accessing the counter control register. If a NMI hits, then it will
1159 * keep the counter running.
1161 void perf_counter_task_sched_in(struct task_struct
*task
, int cpu
)
1163 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1164 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
1168 if (cpuctx
->task_ctx
== ctx
)
1170 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1171 cpuctx
->task_ctx
= ctx
;
1174 static void perf_counter_cpu_sched_in(struct perf_cpu_context
*cpuctx
, int cpu
)
1176 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
1178 __perf_counter_sched_in(ctx
, cpuctx
, cpu
);
1181 #define MAX_INTERRUPTS (~0ULL)
1183 static void perf_log_throttle(struct perf_counter
*counter
, int enable
);
1184 static void perf_log_period(struct perf_counter
*counter
, u64 period
);
1186 static void perf_adjust_freq(struct perf_counter_context
*ctx
)
1188 struct perf_counter
*counter
;
1189 u64 interrupts
, sample_period
;
1193 spin_lock(&ctx
->lock
);
1194 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1195 if (counter
->state
!= PERF_COUNTER_STATE_ACTIVE
)
1198 interrupts
= counter
->hw
.interrupts
;
1199 counter
->hw
.interrupts
= 0;
1201 if (interrupts
== MAX_INTERRUPTS
) {
1202 perf_log_throttle(counter
, 1);
1203 counter
->pmu
->unthrottle(counter
);
1204 interrupts
= 2*sysctl_perf_counter_limit
/HZ
;
1207 if (!counter
->hw_event
.freq
|| !counter
->hw_event
.sample_freq
)
1210 events
= HZ
* interrupts
* counter
->hw
.sample_period
;
1211 period
= div64_u64(events
, counter
->hw_event
.sample_freq
);
1213 delta
= (s64
)(1 + period
- counter
->hw
.sample_period
);
1216 sample_period
= counter
->hw
.sample_period
+ delta
;
1221 perf_log_period(counter
, sample_period
);
1223 counter
->hw
.sample_period
= sample_period
;
1225 spin_unlock(&ctx
->lock
);
1229 * Round-robin a context's counters:
1231 static void rotate_ctx(struct perf_counter_context
*ctx
)
1233 struct perf_counter
*counter
;
1235 if (!ctx
->nr_counters
)
1238 spin_lock(&ctx
->lock
);
1240 * Rotate the first entry last (works just fine for group counters too):
1243 list_for_each_entry(counter
, &ctx
->counter_list
, list_entry
) {
1244 list_move_tail(&counter
->list_entry
, &ctx
->counter_list
);
1249 spin_unlock(&ctx
->lock
);
1252 void perf_counter_task_tick(struct task_struct
*curr
, int cpu
)
1254 struct perf_cpu_context
*cpuctx
;
1255 struct perf_counter_context
*ctx
;
1257 if (!atomic_read(&nr_counters
))
1260 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1261 ctx
= curr
->perf_counter_ctxp
;
1263 perf_adjust_freq(&cpuctx
->ctx
);
1265 perf_adjust_freq(ctx
);
1267 perf_counter_cpu_sched_out(cpuctx
);
1269 __perf_counter_task_sched_out(ctx
);
1271 rotate_ctx(&cpuctx
->ctx
);
1275 perf_counter_cpu_sched_in(cpuctx
, cpu
);
1277 perf_counter_task_sched_in(curr
, cpu
);
1281 * Cross CPU call to read the hardware counter
1283 static void __read(void *info
)
1285 struct perf_counter
*counter
= info
;
1286 struct perf_counter_context
*ctx
= counter
->ctx
;
1287 unsigned long flags
;
1289 local_irq_save(flags
);
1291 update_context_time(ctx
);
1292 counter
->pmu
->read(counter
);
1293 update_counter_times(counter
);
1294 local_irq_restore(flags
);
1297 static u64
perf_counter_read(struct perf_counter
*counter
)
1300 * If counter is enabled and currently active on a CPU, update the
1301 * value in the counter structure:
1303 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
) {
1304 smp_call_function_single(counter
->oncpu
,
1305 __read
, counter
, 1);
1306 } else if (counter
->state
== PERF_COUNTER_STATE_INACTIVE
) {
1307 update_counter_times(counter
);
1310 return atomic64_read(&counter
->count
);
1314 * Initialize the perf_counter context in a task_struct:
1317 __perf_counter_init_context(struct perf_counter_context
*ctx
,
1318 struct task_struct
*task
)
1320 memset(ctx
, 0, sizeof(*ctx
));
1321 spin_lock_init(&ctx
->lock
);
1322 mutex_init(&ctx
->mutex
);
1323 INIT_LIST_HEAD(&ctx
->counter_list
);
1324 INIT_LIST_HEAD(&ctx
->event_list
);
1325 atomic_set(&ctx
->refcount
, 1);
1329 static struct perf_counter_context
*find_get_context(pid_t pid
, int cpu
)
1331 struct perf_counter_context
*parent_ctx
;
1332 struct perf_counter_context
*ctx
;
1333 struct perf_cpu_context
*cpuctx
;
1334 struct task_struct
*task
;
1335 unsigned long flags
;
1339 * If cpu is not a wildcard then this is a percpu counter:
1342 /* Must be root to operate on a CPU counter: */
1343 if (sysctl_perf_counter_priv
&& !capable(CAP_SYS_ADMIN
))
1344 return ERR_PTR(-EACCES
);
1346 if (cpu
< 0 || cpu
> num_possible_cpus())
1347 return ERR_PTR(-EINVAL
);
1350 * We could be clever and allow to attach a counter to an
1351 * offline CPU and activate it when the CPU comes up, but
1354 if (!cpu_isset(cpu
, cpu_online_map
))
1355 return ERR_PTR(-ENODEV
);
1357 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
1368 task
= find_task_by_vpid(pid
);
1370 get_task_struct(task
);
1374 return ERR_PTR(-ESRCH
);
1377 * Can't attach counters to a dying task.
1380 if (task
->flags
& PF_EXITING
)
1383 /* Reuse ptrace permission checks for now. */
1385 if (!ptrace_may_access(task
, PTRACE_MODE_READ
))
1389 ctx
= perf_lock_task_context(task
, &flags
);
1391 parent_ctx
= ctx
->parent_ctx
;
1393 put_ctx(parent_ctx
);
1394 ctx
->parent_ctx
= NULL
; /* no longer a clone */
1397 * Get an extra reference before dropping the lock so that
1398 * this context won't get freed if the task exits.
1401 spin_unlock_irqrestore(&ctx
->lock
, flags
);
1405 ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
1409 __perf_counter_init_context(ctx
, task
);
1411 if (cmpxchg(&task
->perf_counter_ctxp
, NULL
, ctx
)) {
1413 * We raced with some other task; use
1414 * the context they set.
1419 get_task_struct(task
);
1422 put_task_struct(task
);
1426 put_task_struct(task
);
1427 return ERR_PTR(err
);
1430 static void free_counter_rcu(struct rcu_head
*head
)
1432 struct perf_counter
*counter
;
1434 counter
= container_of(head
, struct perf_counter
, rcu_head
);
1436 put_pid_ns(counter
->ns
);
1440 static void perf_pending_sync(struct perf_counter
*counter
);
1442 static void free_counter(struct perf_counter
*counter
)
1444 perf_pending_sync(counter
);
1446 atomic_dec(&nr_counters
);
1447 if (counter
->hw_event
.mmap
)
1448 atomic_dec(&nr_mmap_tracking
);
1449 if (counter
->hw_event
.munmap
)
1450 atomic_dec(&nr_munmap_tracking
);
1451 if (counter
->hw_event
.comm
)
1452 atomic_dec(&nr_comm_tracking
);
1454 if (counter
->destroy
)
1455 counter
->destroy(counter
);
1457 put_ctx(counter
->ctx
);
1458 call_rcu(&counter
->rcu_head
, free_counter_rcu
);
1462 * Called when the last reference to the file is gone.
1464 static int perf_release(struct inode
*inode
, struct file
*file
)
1466 struct perf_counter
*counter
= file
->private_data
;
1467 struct perf_counter_context
*ctx
= counter
->ctx
;
1469 file
->private_data
= NULL
;
1471 WARN_ON_ONCE(ctx
->parent_ctx
);
1472 mutex_lock(&ctx
->mutex
);
1473 perf_counter_remove_from_context(counter
);
1474 mutex_unlock(&ctx
->mutex
);
1476 mutex_lock(&counter
->owner
->perf_counter_mutex
);
1477 list_del_init(&counter
->owner_entry
);
1478 mutex_unlock(&counter
->owner
->perf_counter_mutex
);
1479 put_task_struct(counter
->owner
);
1481 free_counter(counter
);
1487 * Read the performance counter - simple non blocking version for now
1490 perf_read_hw(struct perf_counter
*counter
, char __user
*buf
, size_t count
)
1496 * Return end-of-file for a read on a counter that is in
1497 * error state (i.e. because it was pinned but it couldn't be
1498 * scheduled on to the CPU at some point).
1500 if (counter
->state
== PERF_COUNTER_STATE_ERROR
)
1503 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1504 mutex_lock(&counter
->child_mutex
);
1505 values
[0] = perf_counter_read(counter
);
1507 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1508 values
[n
++] = counter
->total_time_enabled
+
1509 atomic64_read(&counter
->child_total_time_enabled
);
1510 if (counter
->hw_event
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1511 values
[n
++] = counter
->total_time_running
+
1512 atomic64_read(&counter
->child_total_time_running
);
1513 if (counter
->hw_event
.read_format
& PERF_FORMAT_ID
)
1514 values
[n
++] = counter
->id
;
1515 mutex_unlock(&counter
->child_mutex
);
1517 if (count
< n
* sizeof(u64
))
1519 count
= n
* sizeof(u64
);
1521 if (copy_to_user(buf
, values
, count
))
1528 perf_read(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
1530 struct perf_counter
*counter
= file
->private_data
;
1532 return perf_read_hw(counter
, buf
, count
);
1535 static unsigned int perf_poll(struct file
*file
, poll_table
*wait
)
1537 struct perf_counter
*counter
= file
->private_data
;
1538 struct perf_mmap_data
*data
;
1539 unsigned int events
= POLL_HUP
;
1542 data
= rcu_dereference(counter
->data
);
1544 events
= atomic_xchg(&data
->poll
, 0);
1547 poll_wait(file
, &counter
->waitq
, wait
);
1552 static void perf_counter_reset(struct perf_counter
*counter
)
1554 (void)perf_counter_read(counter
);
1555 atomic64_set(&counter
->count
, 0);
1556 perf_counter_update_userpage(counter
);
1559 static void perf_counter_for_each_sibling(struct perf_counter
*counter
,
1560 void (*func
)(struct perf_counter
*))
1562 struct perf_counter_context
*ctx
= counter
->ctx
;
1563 struct perf_counter
*sibling
;
1565 WARN_ON_ONCE(ctx
->parent_ctx
);
1566 mutex_lock(&ctx
->mutex
);
1567 counter
= counter
->group_leader
;
1570 list_for_each_entry(sibling
, &counter
->sibling_list
, list_entry
)
1572 mutex_unlock(&ctx
->mutex
);
1576 * Holding the top-level counter's child_mutex means that any
1577 * descendant process that has inherited this counter will block
1578 * in sync_child_counter if it goes to exit, thus satisfying the
1579 * task existence requirements of perf_counter_enable/disable.
1581 static void perf_counter_for_each_child(struct perf_counter
*counter
,
1582 void (*func
)(struct perf_counter
*))
1584 struct perf_counter
*child
;
1586 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1587 mutex_lock(&counter
->child_mutex
);
1589 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1591 mutex_unlock(&counter
->child_mutex
);
1594 static void perf_counter_for_each(struct perf_counter
*counter
,
1595 void (*func
)(struct perf_counter
*))
1597 struct perf_counter
*child
;
1599 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1600 mutex_lock(&counter
->child_mutex
);
1601 perf_counter_for_each_sibling(counter
, func
);
1602 list_for_each_entry(child
, &counter
->child_list
, child_list
)
1603 perf_counter_for_each_sibling(child
, func
);
1604 mutex_unlock(&counter
->child_mutex
);
1607 static int perf_counter_period(struct perf_counter
*counter
, u64 __user
*arg
)
1609 struct perf_counter_context
*ctx
= counter
->ctx
;
1614 if (!counter
->hw_event
.sample_period
)
1617 size
= copy_from_user(&value
, arg
, sizeof(value
));
1618 if (size
!= sizeof(value
))
1624 spin_lock_irq(&ctx
->lock
);
1625 if (counter
->hw_event
.freq
) {
1626 if (value
> sysctl_perf_counter_limit
) {
1631 counter
->hw_event
.sample_freq
= value
;
1633 counter
->hw_event
.sample_period
= value
;
1634 counter
->hw
.sample_period
= value
;
1636 perf_log_period(counter
, value
);
1639 spin_unlock_irq(&ctx
->lock
);
1644 static long perf_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1646 struct perf_counter
*counter
= file
->private_data
;
1647 void (*func
)(struct perf_counter
*);
1651 case PERF_COUNTER_IOC_ENABLE
:
1652 func
= perf_counter_enable
;
1654 case PERF_COUNTER_IOC_DISABLE
:
1655 func
= perf_counter_disable
;
1657 case PERF_COUNTER_IOC_RESET
:
1658 func
= perf_counter_reset
;
1661 case PERF_COUNTER_IOC_REFRESH
:
1662 return perf_counter_refresh(counter
, arg
);
1664 case PERF_COUNTER_IOC_PERIOD
:
1665 return perf_counter_period(counter
, (u64 __user
*)arg
);
1671 if (flags
& PERF_IOC_FLAG_GROUP
)
1672 perf_counter_for_each(counter
, func
);
1674 perf_counter_for_each_child(counter
, func
);
1679 int perf_counter_task_enable(void)
1681 struct perf_counter
*counter
;
1683 mutex_lock(¤t
->perf_counter_mutex
);
1684 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1685 perf_counter_for_each_child(counter
, perf_counter_enable
);
1686 mutex_unlock(¤t
->perf_counter_mutex
);
1691 int perf_counter_task_disable(void)
1693 struct perf_counter
*counter
;
1695 mutex_lock(¤t
->perf_counter_mutex
);
1696 list_for_each_entry(counter
, ¤t
->perf_counter_list
, owner_entry
)
1697 perf_counter_for_each_child(counter
, perf_counter_disable
);
1698 mutex_unlock(¤t
->perf_counter_mutex
);
1704 * Callers need to ensure there can be no nesting of this function, otherwise
1705 * the seqlock logic goes bad. We can not serialize this because the arch
1706 * code calls this from NMI context.
1708 void perf_counter_update_userpage(struct perf_counter
*counter
)
1710 struct perf_counter_mmap_page
*userpg
;
1711 struct perf_mmap_data
*data
;
1714 data
= rcu_dereference(counter
->data
);
1718 userpg
= data
->user_page
;
1721 * Disable preemption so as to not let the corresponding user-space
1722 * spin too long if we get preempted.
1727 userpg
->index
= counter
->hw
.idx
;
1728 userpg
->offset
= atomic64_read(&counter
->count
);
1729 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
1730 userpg
->offset
-= atomic64_read(&counter
->hw
.prev_count
);
1739 static int perf_mmap_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1741 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1742 struct perf_mmap_data
*data
;
1743 int ret
= VM_FAULT_SIGBUS
;
1746 data
= rcu_dereference(counter
->data
);
1750 if (vmf
->pgoff
== 0) {
1751 vmf
->page
= virt_to_page(data
->user_page
);
1753 int nr
= vmf
->pgoff
- 1;
1755 if ((unsigned)nr
> data
->nr_pages
)
1758 vmf
->page
= virt_to_page(data
->data_pages
[nr
]);
1760 get_page(vmf
->page
);
1768 static int perf_mmap_data_alloc(struct perf_counter
*counter
, int nr_pages
)
1770 struct perf_mmap_data
*data
;
1774 WARN_ON(atomic_read(&counter
->mmap_count
));
1776 size
= sizeof(struct perf_mmap_data
);
1777 size
+= nr_pages
* sizeof(void *);
1779 data
= kzalloc(size
, GFP_KERNEL
);
1783 data
->user_page
= (void *)get_zeroed_page(GFP_KERNEL
);
1784 if (!data
->user_page
)
1785 goto fail_user_page
;
1787 for (i
= 0; i
< nr_pages
; i
++) {
1788 data
->data_pages
[i
] = (void *)get_zeroed_page(GFP_KERNEL
);
1789 if (!data
->data_pages
[i
])
1790 goto fail_data_pages
;
1793 data
->nr_pages
= nr_pages
;
1794 atomic_set(&data
->lock
, -1);
1796 rcu_assign_pointer(counter
->data
, data
);
1801 for (i
--; i
>= 0; i
--)
1802 free_page((unsigned long)data
->data_pages
[i
]);
1804 free_page((unsigned long)data
->user_page
);
1813 static void __perf_mmap_data_free(struct rcu_head
*rcu_head
)
1815 struct perf_mmap_data
*data
;
1818 data
= container_of(rcu_head
, struct perf_mmap_data
, rcu_head
);
1820 free_page((unsigned long)data
->user_page
);
1821 for (i
= 0; i
< data
->nr_pages
; i
++)
1822 free_page((unsigned long)data
->data_pages
[i
]);
1826 static void perf_mmap_data_free(struct perf_counter
*counter
)
1828 struct perf_mmap_data
*data
= counter
->data
;
1830 WARN_ON(atomic_read(&counter
->mmap_count
));
1832 rcu_assign_pointer(counter
->data
, NULL
);
1833 call_rcu(&data
->rcu_head
, __perf_mmap_data_free
);
1836 static void perf_mmap_open(struct vm_area_struct
*vma
)
1838 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1840 atomic_inc(&counter
->mmap_count
);
1843 static void perf_mmap_close(struct vm_area_struct
*vma
)
1845 struct perf_counter
*counter
= vma
->vm_file
->private_data
;
1847 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1848 if (atomic_dec_and_mutex_lock(&counter
->mmap_count
, &counter
->mmap_mutex
)) {
1849 struct user_struct
*user
= current_user();
1851 atomic_long_sub(counter
->data
->nr_pages
+ 1, &user
->locked_vm
);
1852 vma
->vm_mm
->locked_vm
-= counter
->data
->nr_locked
;
1853 perf_mmap_data_free(counter
);
1854 mutex_unlock(&counter
->mmap_mutex
);
1858 static struct vm_operations_struct perf_mmap_vmops
= {
1859 .open
= perf_mmap_open
,
1860 .close
= perf_mmap_close
,
1861 .fault
= perf_mmap_fault
,
1864 static int perf_mmap(struct file
*file
, struct vm_area_struct
*vma
)
1866 struct perf_counter
*counter
= file
->private_data
;
1867 unsigned long user_locked
, user_lock_limit
;
1868 struct user_struct
*user
= current_user();
1869 unsigned long locked
, lock_limit
;
1870 unsigned long vma_size
;
1871 unsigned long nr_pages
;
1872 long user_extra
, extra
;
1875 if (!(vma
->vm_flags
& VM_SHARED
) || (vma
->vm_flags
& VM_WRITE
))
1878 vma_size
= vma
->vm_end
- vma
->vm_start
;
1879 nr_pages
= (vma_size
/ PAGE_SIZE
) - 1;
1882 * If we have data pages ensure they're a power-of-two number, so we
1883 * can do bitmasks instead of modulo.
1885 if (nr_pages
!= 0 && !is_power_of_2(nr_pages
))
1888 if (vma_size
!= PAGE_SIZE
* (1 + nr_pages
))
1891 if (vma
->vm_pgoff
!= 0)
1894 WARN_ON_ONCE(counter
->ctx
->parent_ctx
);
1895 mutex_lock(&counter
->mmap_mutex
);
1896 if (atomic_inc_not_zero(&counter
->mmap_count
)) {
1897 if (nr_pages
!= counter
->data
->nr_pages
)
1902 user_extra
= nr_pages
+ 1;
1903 user_lock_limit
= sysctl_perf_counter_mlock
>> (PAGE_SHIFT
- 10);
1906 * Increase the limit linearly with more CPUs:
1908 user_lock_limit
*= num_online_cpus();
1910 user_locked
= atomic_long_read(&user
->locked_vm
) + user_extra
;
1913 if (user_locked
> user_lock_limit
)
1914 extra
= user_locked
- user_lock_limit
;
1916 lock_limit
= current
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
;
1917 lock_limit
>>= PAGE_SHIFT
;
1918 locked
= vma
->vm_mm
->locked_vm
+ extra
;
1920 if ((locked
> lock_limit
) && !capable(CAP_IPC_LOCK
)) {
1925 WARN_ON(counter
->data
);
1926 ret
= perf_mmap_data_alloc(counter
, nr_pages
);
1930 atomic_set(&counter
->mmap_count
, 1);
1931 atomic_long_add(user_extra
, &user
->locked_vm
);
1932 vma
->vm_mm
->locked_vm
+= extra
;
1933 counter
->data
->nr_locked
= extra
;
1935 mutex_unlock(&counter
->mmap_mutex
);
1937 vma
->vm_flags
&= ~VM_MAYWRITE
;
1938 vma
->vm_flags
|= VM_RESERVED
;
1939 vma
->vm_ops
= &perf_mmap_vmops
;
1944 static int perf_fasync(int fd
, struct file
*filp
, int on
)
1946 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1947 struct perf_counter
*counter
= filp
->private_data
;
1950 mutex_lock(&inode
->i_mutex
);
1951 retval
= fasync_helper(fd
, filp
, on
, &counter
->fasync
);
1952 mutex_unlock(&inode
->i_mutex
);
1960 static const struct file_operations perf_fops
= {
1961 .release
= perf_release
,
1964 .unlocked_ioctl
= perf_ioctl
,
1965 .compat_ioctl
= perf_ioctl
,
1967 .fasync
= perf_fasync
,
1971 * Perf counter wakeup
1973 * If there's data, ensure we set the poll() state and publish everything
1974 * to user-space before waking everybody up.
1977 void perf_counter_wakeup(struct perf_counter
*counter
)
1979 wake_up_all(&counter
->waitq
);
1981 if (counter
->pending_kill
) {
1982 kill_fasync(&counter
->fasync
, SIGIO
, counter
->pending_kill
);
1983 counter
->pending_kill
= 0;
1990 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1992 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1993 * single linked list and use cmpxchg() to add entries lockless.
1996 static void perf_pending_counter(struct perf_pending_entry
*entry
)
1998 struct perf_counter
*counter
= container_of(entry
,
1999 struct perf_counter
, pending
);
2001 if (counter
->pending_disable
) {
2002 counter
->pending_disable
= 0;
2003 perf_counter_disable(counter
);
2006 if (counter
->pending_wakeup
) {
2007 counter
->pending_wakeup
= 0;
2008 perf_counter_wakeup(counter
);
2012 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2014 static DEFINE_PER_CPU(struct perf_pending_entry
*, perf_pending_head
) = {
2018 static void perf_pending_queue(struct perf_pending_entry
*entry
,
2019 void (*func
)(struct perf_pending_entry
*))
2021 struct perf_pending_entry
**head
;
2023 if (cmpxchg(&entry
->next
, NULL
, PENDING_TAIL
) != NULL
)
2028 head
= &get_cpu_var(perf_pending_head
);
2031 entry
->next
= *head
;
2032 } while (cmpxchg(head
, entry
->next
, entry
) != entry
->next
);
2034 set_perf_counter_pending();
2036 put_cpu_var(perf_pending_head
);
2039 static int __perf_pending_run(void)
2041 struct perf_pending_entry
*list
;
2044 list
= xchg(&__get_cpu_var(perf_pending_head
), PENDING_TAIL
);
2045 while (list
!= PENDING_TAIL
) {
2046 void (*func
)(struct perf_pending_entry
*);
2047 struct perf_pending_entry
*entry
= list
;
2054 * Ensure we observe the unqueue before we issue the wakeup,
2055 * so that we won't be waiting forever.
2056 * -- see perf_not_pending().
2067 static inline int perf_not_pending(struct perf_counter
*counter
)
2070 * If we flush on whatever cpu we run, there is a chance we don't
2074 __perf_pending_run();
2078 * Ensure we see the proper queue state before going to sleep
2079 * so that we do not miss the wakeup. -- see perf_pending_handle()
2082 return counter
->pending
.next
== NULL
;
2085 static void perf_pending_sync(struct perf_counter
*counter
)
2087 wait_event(counter
->waitq
, perf_not_pending(counter
));
2090 void perf_counter_do_pending(void)
2092 __perf_pending_run();
2096 * Callchain support -- arch specific
2099 __weak
struct perf_callchain_entry
*perf_callchain(struct pt_regs
*regs
)
2108 struct perf_output_handle
{
2109 struct perf_counter
*counter
;
2110 struct perf_mmap_data
*data
;
2112 unsigned long offset
;
2116 unsigned long flags
;
2119 static void perf_output_wakeup(struct perf_output_handle
*handle
)
2121 atomic_set(&handle
->data
->poll
, POLL_IN
);
2124 handle
->counter
->pending_wakeup
= 1;
2125 perf_pending_queue(&handle
->counter
->pending
,
2126 perf_pending_counter
);
2128 perf_counter_wakeup(handle
->counter
);
2132 * Curious locking construct.
2134 * We need to ensure a later event doesn't publish a head when a former
2135 * event isn't done writing. However since we need to deal with NMIs we
2136 * cannot fully serialize things.
2138 * What we do is serialize between CPUs so we only have to deal with NMI
2139 * nesting on a single CPU.
2141 * We only publish the head (and generate a wakeup) when the outer-most
2144 static void perf_output_lock(struct perf_output_handle
*handle
)
2146 struct perf_mmap_data
*data
= handle
->data
;
2151 local_irq_save(handle
->flags
);
2152 cpu
= smp_processor_id();
2154 if (in_nmi() && atomic_read(&data
->lock
) == cpu
)
2157 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2163 static void perf_output_unlock(struct perf_output_handle
*handle
)
2165 struct perf_mmap_data
*data
= handle
->data
;
2169 data
->done_head
= data
->head
;
2171 if (!handle
->locked
)
2176 * The xchg implies a full barrier that ensures all writes are done
2177 * before we publish the new head, matched by a rmb() in userspace when
2178 * reading this position.
2180 while ((head
= atomic_long_xchg(&data
->done_head
, 0)))
2181 data
->user_page
->data_head
= head
;
2184 * NMI can happen here, which means we can miss a done_head update.
2187 cpu
= atomic_xchg(&data
->lock
, -1);
2188 WARN_ON_ONCE(cpu
!= smp_processor_id());
2191 * Therefore we have to validate we did not indeed do so.
2193 if (unlikely(atomic_long_read(&data
->done_head
))) {
2195 * Since we had it locked, we can lock it again.
2197 while (atomic_cmpxchg(&data
->lock
, -1, cpu
) != -1)
2203 if (atomic_xchg(&data
->wakeup
, 0))
2204 perf_output_wakeup(handle
);
2206 local_irq_restore(handle
->flags
);
2209 static int perf_output_begin(struct perf_output_handle
*handle
,
2210 struct perf_counter
*counter
, unsigned int size
,
2211 int nmi
, int overflow
)
2213 struct perf_mmap_data
*data
;
2214 unsigned int offset
, head
;
2217 * For inherited counters we send all the output towards the parent.
2219 if (counter
->parent
)
2220 counter
= counter
->parent
;
2223 data
= rcu_dereference(counter
->data
);
2227 handle
->data
= data
;
2228 handle
->counter
= counter
;
2230 handle
->overflow
= overflow
;
2232 if (!data
->nr_pages
)
2235 perf_output_lock(handle
);
2238 offset
= head
= atomic_read(&data
->head
);
2240 } while (atomic_long_cmpxchg(&data
->head
, offset
, head
) != offset
);
2242 handle
->offset
= offset
;
2243 handle
->head
= head
;
2245 if ((offset
>> PAGE_SHIFT
) != (head
>> PAGE_SHIFT
))
2246 atomic_set(&data
->wakeup
, 1);
2251 perf_output_wakeup(handle
);
2258 static void perf_output_copy(struct perf_output_handle
*handle
,
2259 void *buf
, unsigned int len
)
2261 unsigned int pages_mask
;
2262 unsigned int offset
;
2266 offset
= handle
->offset
;
2267 pages_mask
= handle
->data
->nr_pages
- 1;
2268 pages
= handle
->data
->data_pages
;
2271 unsigned int page_offset
;
2274 nr
= (offset
>> PAGE_SHIFT
) & pages_mask
;
2275 page_offset
= offset
& (PAGE_SIZE
- 1);
2276 size
= min_t(unsigned int, PAGE_SIZE
- page_offset
, len
);
2278 memcpy(pages
[nr
] + page_offset
, buf
, size
);
2285 handle
->offset
= offset
;
2288 * Check we didn't copy past our reservation window, taking the
2289 * possible unsigned int wrap into account.
2291 WARN_ON_ONCE(((long)(handle
->head
- handle
->offset
)) < 0);
2294 #define perf_output_put(handle, x) \
2295 perf_output_copy((handle), &(x), sizeof(x))
2297 static void perf_output_end(struct perf_output_handle
*handle
)
2299 struct perf_counter
*counter
= handle
->counter
;
2300 struct perf_mmap_data
*data
= handle
->data
;
2302 int wakeup_events
= counter
->hw_event
.wakeup_events
;
2304 if (handle
->overflow
&& wakeup_events
) {
2305 int events
= atomic_inc_return(&data
->events
);
2306 if (events
>= wakeup_events
) {
2307 atomic_sub(wakeup_events
, &data
->events
);
2308 atomic_set(&data
->wakeup
, 1);
2312 perf_output_unlock(handle
);
2316 static u32
perf_counter_pid(struct perf_counter
*counter
, struct task_struct
*p
)
2319 * only top level counters have the pid namespace they were created in
2321 if (counter
->parent
)
2322 counter
= counter
->parent
;
2324 return task_tgid_nr_ns(p
, counter
->ns
);
2327 static u32
perf_counter_tid(struct perf_counter
*counter
, struct task_struct
*p
)
2330 * only top level counters have the pid namespace they were created in
2332 if (counter
->parent
)
2333 counter
= counter
->parent
;
2335 return task_pid_nr_ns(p
, counter
->ns
);
2338 static void perf_counter_output(struct perf_counter
*counter
,
2339 int nmi
, struct pt_regs
*regs
, u64 addr
)
2342 u64 sample_type
= counter
->hw_event
.sample_type
;
2343 struct perf_output_handle handle
;
2344 struct perf_event_header header
;
2353 struct perf_callchain_entry
*callchain
= NULL
;
2354 int callchain_size
= 0;
2361 header
.size
= sizeof(header
);
2363 header
.misc
= PERF_EVENT_MISC_OVERFLOW
;
2364 header
.misc
|= perf_misc_flags(regs
);
2366 if (sample_type
& PERF_SAMPLE_IP
) {
2367 ip
= perf_instruction_pointer(regs
);
2368 header
.type
|= PERF_SAMPLE_IP
;
2369 header
.size
+= sizeof(ip
);
2372 if (sample_type
& PERF_SAMPLE_TID
) {
2373 /* namespace issues */
2374 tid_entry
.pid
= perf_counter_pid(counter
, current
);
2375 tid_entry
.tid
= perf_counter_tid(counter
, current
);
2377 header
.type
|= PERF_SAMPLE_TID
;
2378 header
.size
+= sizeof(tid_entry
);
2381 if (sample_type
& PERF_SAMPLE_TIME
) {
2383 * Maybe do better on x86 and provide cpu_clock_nmi()
2385 time
= sched_clock();
2387 header
.type
|= PERF_SAMPLE_TIME
;
2388 header
.size
+= sizeof(u64
);
2391 if (sample_type
& PERF_SAMPLE_ADDR
) {
2392 header
.type
|= PERF_SAMPLE_ADDR
;
2393 header
.size
+= sizeof(u64
);
2396 if (sample_type
& PERF_SAMPLE_CONFIG
) {
2397 header
.type
|= PERF_SAMPLE_CONFIG
;
2398 header
.size
+= sizeof(u64
);
2401 if (sample_type
& PERF_SAMPLE_CPU
) {
2402 header
.type
|= PERF_SAMPLE_CPU
;
2403 header
.size
+= sizeof(cpu_entry
);
2405 cpu_entry
.cpu
= raw_smp_processor_id();
2408 if (sample_type
& PERF_SAMPLE_GROUP
) {
2409 header
.type
|= PERF_SAMPLE_GROUP
;
2410 header
.size
+= sizeof(u64
) +
2411 counter
->nr_siblings
* sizeof(group_entry
);
2414 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
2415 callchain
= perf_callchain(regs
);
2418 callchain_size
= (1 + callchain
->nr
) * sizeof(u64
);
2420 header
.type
|= PERF_SAMPLE_CALLCHAIN
;
2421 header
.size
+= callchain_size
;
2425 ret
= perf_output_begin(&handle
, counter
, header
.size
, nmi
, 1);
2429 perf_output_put(&handle
, header
);
2431 if (sample_type
& PERF_SAMPLE_IP
)
2432 perf_output_put(&handle
, ip
);
2434 if (sample_type
& PERF_SAMPLE_TID
)
2435 perf_output_put(&handle
, tid_entry
);
2437 if (sample_type
& PERF_SAMPLE_TIME
)
2438 perf_output_put(&handle
, time
);
2440 if (sample_type
& PERF_SAMPLE_ADDR
)
2441 perf_output_put(&handle
, addr
);
2443 if (sample_type
& PERF_SAMPLE_CONFIG
)
2444 perf_output_put(&handle
, counter
->hw_event
.config
);
2446 if (sample_type
& PERF_SAMPLE_CPU
)
2447 perf_output_put(&handle
, cpu_entry
);
2450 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2452 if (sample_type
& PERF_SAMPLE_GROUP
) {
2453 struct perf_counter
*leader
, *sub
;
2454 u64 nr
= counter
->nr_siblings
;
2456 perf_output_put(&handle
, nr
);
2458 leader
= counter
->group_leader
;
2459 list_for_each_entry(sub
, &leader
->sibling_list
, list_entry
) {
2461 sub
->pmu
->read(sub
);
2463 group_entry
.id
= sub
->id
;
2464 group_entry
.counter
= atomic64_read(&sub
->count
);
2466 perf_output_put(&handle
, group_entry
);
2471 perf_output_copy(&handle
, callchain
, callchain_size
);
2473 perf_output_end(&handle
);
2480 struct perf_comm_event
{
2481 struct task_struct
*task
;
2486 struct perf_event_header header
;
2493 static void perf_counter_comm_output(struct perf_counter
*counter
,
2494 struct perf_comm_event
*comm_event
)
2496 struct perf_output_handle handle
;
2497 int size
= comm_event
->event
.header
.size
;
2498 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2503 comm_event
->event
.pid
= perf_counter_pid(counter
, comm_event
->task
);
2504 comm_event
->event
.tid
= perf_counter_tid(counter
, comm_event
->task
);
2506 perf_output_put(&handle
, comm_event
->event
);
2507 perf_output_copy(&handle
, comm_event
->comm
,
2508 comm_event
->comm_size
);
2509 perf_output_end(&handle
);
2512 static int perf_counter_comm_match(struct perf_counter
*counter
,
2513 struct perf_comm_event
*comm_event
)
2515 if (counter
->hw_event
.comm
&&
2516 comm_event
->event
.header
.type
== PERF_EVENT_COMM
)
2522 static void perf_counter_comm_ctx(struct perf_counter_context
*ctx
,
2523 struct perf_comm_event
*comm_event
)
2525 struct perf_counter
*counter
;
2527 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2531 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2532 if (perf_counter_comm_match(counter
, comm_event
))
2533 perf_counter_comm_output(counter
, comm_event
);
2538 static void perf_counter_comm_event(struct perf_comm_event
*comm_event
)
2540 struct perf_cpu_context
*cpuctx
;
2541 struct perf_counter_context
*ctx
;
2543 char *comm
= comm_event
->task
->comm
;
2545 size
= ALIGN(strlen(comm
)+1, sizeof(u64
));
2547 comm_event
->comm
= comm
;
2548 comm_event
->comm_size
= size
;
2550 comm_event
->event
.header
.size
= sizeof(comm_event
->event
) + size
;
2552 cpuctx
= &get_cpu_var(perf_cpu_context
);
2553 perf_counter_comm_ctx(&cpuctx
->ctx
, comm_event
);
2554 put_cpu_var(perf_cpu_context
);
2558 * doesn't really matter which of the child contexts the
2559 * events ends up in.
2561 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2563 perf_counter_comm_ctx(ctx
, comm_event
);
2567 void perf_counter_comm(struct task_struct
*task
)
2569 struct perf_comm_event comm_event
;
2571 if (!atomic_read(&nr_comm_tracking
))
2574 comm_event
= (struct perf_comm_event
){
2577 .header
= { .type
= PERF_EVENT_COMM
, },
2581 perf_counter_comm_event(&comm_event
);
2588 struct perf_mmap_event
{
2594 struct perf_event_header header
;
2604 static void perf_counter_mmap_output(struct perf_counter
*counter
,
2605 struct perf_mmap_event
*mmap_event
)
2607 struct perf_output_handle handle
;
2608 int size
= mmap_event
->event
.header
.size
;
2609 int ret
= perf_output_begin(&handle
, counter
, size
, 0, 0);
2614 mmap_event
->event
.pid
= perf_counter_pid(counter
, current
);
2615 mmap_event
->event
.tid
= perf_counter_tid(counter
, current
);
2617 perf_output_put(&handle
, mmap_event
->event
);
2618 perf_output_copy(&handle
, mmap_event
->file_name
,
2619 mmap_event
->file_size
);
2620 perf_output_end(&handle
);
2623 static int perf_counter_mmap_match(struct perf_counter
*counter
,
2624 struct perf_mmap_event
*mmap_event
)
2626 if (counter
->hw_event
.mmap
&&
2627 mmap_event
->event
.header
.type
== PERF_EVENT_MMAP
)
2630 if (counter
->hw_event
.munmap
&&
2631 mmap_event
->event
.header
.type
== PERF_EVENT_MUNMAP
)
2637 static void perf_counter_mmap_ctx(struct perf_counter_context
*ctx
,
2638 struct perf_mmap_event
*mmap_event
)
2640 struct perf_counter
*counter
;
2642 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
2646 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
2647 if (perf_counter_mmap_match(counter
, mmap_event
))
2648 perf_counter_mmap_output(counter
, mmap_event
);
2653 static void perf_counter_mmap_event(struct perf_mmap_event
*mmap_event
)
2655 struct perf_cpu_context
*cpuctx
;
2656 struct perf_counter_context
*ctx
;
2657 struct file
*file
= mmap_event
->file
;
2664 buf
= kzalloc(PATH_MAX
, GFP_KERNEL
);
2666 name
= strncpy(tmp
, "//enomem", sizeof(tmp
));
2669 name
= d_path(&file
->f_path
, buf
, PATH_MAX
);
2671 name
= strncpy(tmp
, "//toolong", sizeof(tmp
));
2675 name
= strncpy(tmp
, "//anon", sizeof(tmp
));
2680 size
= ALIGN(strlen(name
)+1, sizeof(u64
));
2682 mmap_event
->file_name
= name
;
2683 mmap_event
->file_size
= size
;
2685 mmap_event
->event
.header
.size
= sizeof(mmap_event
->event
) + size
;
2687 cpuctx
= &get_cpu_var(perf_cpu_context
);
2688 perf_counter_mmap_ctx(&cpuctx
->ctx
, mmap_event
);
2689 put_cpu_var(perf_cpu_context
);
2693 * doesn't really matter which of the child contexts the
2694 * events ends up in.
2696 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
2698 perf_counter_mmap_ctx(ctx
, mmap_event
);
2704 void perf_counter_mmap(unsigned long addr
, unsigned long len
,
2705 unsigned long pgoff
, struct file
*file
)
2707 struct perf_mmap_event mmap_event
;
2709 if (!atomic_read(&nr_mmap_tracking
))
2712 mmap_event
= (struct perf_mmap_event
){
2715 .header
= { .type
= PERF_EVENT_MMAP
, },
2722 perf_counter_mmap_event(&mmap_event
);
2725 void perf_counter_munmap(unsigned long addr
, unsigned long len
,
2726 unsigned long pgoff
, struct file
*file
)
2728 struct perf_mmap_event mmap_event
;
2730 if (!atomic_read(&nr_munmap_tracking
))
2733 mmap_event
= (struct perf_mmap_event
){
2736 .header
= { .type
= PERF_EVENT_MUNMAP
, },
2743 perf_counter_mmap_event(&mmap_event
);
2747 * Log sample_period changes so that analyzing tools can re-normalize the
2751 static void perf_log_period(struct perf_counter
*counter
, u64 period
)
2753 struct perf_output_handle handle
;
2757 struct perf_event_header header
;
2762 .type
= PERF_EVENT_PERIOD
,
2764 .size
= sizeof(freq_event
),
2766 .time
= sched_clock(),
2770 if (counter
->hw
.sample_period
== period
)
2773 ret
= perf_output_begin(&handle
, counter
, sizeof(freq_event
), 0, 0);
2777 perf_output_put(&handle
, freq_event
);
2778 perf_output_end(&handle
);
2782 * IRQ throttle logging
2785 static void perf_log_throttle(struct perf_counter
*counter
, int enable
)
2787 struct perf_output_handle handle
;
2791 struct perf_event_header header
;
2793 } throttle_event
= {
2795 .type
= PERF_EVENT_THROTTLE
+ 1,
2797 .size
= sizeof(throttle_event
),
2799 .time
= sched_clock(),
2802 ret
= perf_output_begin(&handle
, counter
, sizeof(throttle_event
), 1, 0);
2806 perf_output_put(&handle
, throttle_event
);
2807 perf_output_end(&handle
);
2811 * Generic counter overflow handling.
2814 int perf_counter_overflow(struct perf_counter
*counter
,
2815 int nmi
, struct pt_regs
*regs
, u64 addr
)
2817 int events
= atomic_read(&counter
->event_limit
);
2818 int throttle
= counter
->pmu
->unthrottle
!= NULL
;
2822 counter
->hw
.interrupts
++;
2823 } else if (counter
->hw
.interrupts
!= MAX_INTERRUPTS
) {
2824 counter
->hw
.interrupts
++;
2825 if (HZ
*counter
->hw
.interrupts
> (u64
)sysctl_perf_counter_limit
) {
2826 counter
->hw
.interrupts
= MAX_INTERRUPTS
;
2827 perf_log_throttle(counter
, 0);
2833 * XXX event_limit might not quite work as expected on inherited
2837 counter
->pending_kill
= POLL_IN
;
2838 if (events
&& atomic_dec_and_test(&counter
->event_limit
)) {
2840 counter
->pending_kill
= POLL_HUP
;
2842 counter
->pending_disable
= 1;
2843 perf_pending_queue(&counter
->pending
,
2844 perf_pending_counter
);
2846 perf_counter_disable(counter
);
2849 perf_counter_output(counter
, nmi
, regs
, addr
);
2854 * Generic software counter infrastructure
2857 static void perf_swcounter_update(struct perf_counter
*counter
)
2859 struct hw_perf_counter
*hwc
= &counter
->hw
;
2864 prev
= atomic64_read(&hwc
->prev_count
);
2865 now
= atomic64_read(&hwc
->count
);
2866 if (atomic64_cmpxchg(&hwc
->prev_count
, prev
, now
) != prev
)
2871 atomic64_add(delta
, &counter
->count
);
2872 atomic64_sub(delta
, &hwc
->period_left
);
2875 static void perf_swcounter_set_period(struct perf_counter
*counter
)
2877 struct hw_perf_counter
*hwc
= &counter
->hw
;
2878 s64 left
= atomic64_read(&hwc
->period_left
);
2879 s64 period
= hwc
->sample_period
;
2881 if (unlikely(left
<= -period
)) {
2883 atomic64_set(&hwc
->period_left
, left
);
2886 if (unlikely(left
<= 0)) {
2888 atomic64_add(period
, &hwc
->period_left
);
2891 atomic64_set(&hwc
->prev_count
, -left
);
2892 atomic64_set(&hwc
->count
, -left
);
2895 static enum hrtimer_restart
perf_swcounter_hrtimer(struct hrtimer
*hrtimer
)
2897 enum hrtimer_restart ret
= HRTIMER_RESTART
;
2898 struct perf_counter
*counter
;
2899 struct pt_regs
*regs
;
2902 counter
= container_of(hrtimer
, struct perf_counter
, hw
.hrtimer
);
2903 counter
->pmu
->read(counter
);
2905 regs
= get_irq_regs();
2907 * In case we exclude kernel IPs or are somehow not in interrupt
2908 * context, provide the next best thing, the user IP.
2910 if ((counter
->hw_event
.exclude_kernel
|| !regs
) &&
2911 !counter
->hw_event
.exclude_user
)
2912 regs
= task_pt_regs(current
);
2915 if (perf_counter_overflow(counter
, 0, regs
, 0))
2916 ret
= HRTIMER_NORESTART
;
2919 period
= max_t(u64
, 10000, counter
->hw
.sample_period
);
2920 hrtimer_forward_now(hrtimer
, ns_to_ktime(period
));
2925 static void perf_swcounter_overflow(struct perf_counter
*counter
,
2926 int nmi
, struct pt_regs
*regs
, u64 addr
)
2928 perf_swcounter_update(counter
);
2929 perf_swcounter_set_period(counter
);
2930 if (perf_counter_overflow(counter
, nmi
, regs
, addr
))
2931 /* soft-disable the counter */
2936 static int perf_swcounter_is_counting(struct perf_counter
*counter
)
2938 struct perf_counter_context
*ctx
;
2939 unsigned long flags
;
2942 if (counter
->state
== PERF_COUNTER_STATE_ACTIVE
)
2945 if (counter
->state
!= PERF_COUNTER_STATE_INACTIVE
)
2949 * If the counter is inactive, it could be just because
2950 * its task is scheduled out, or because it's in a group
2951 * which could not go on the PMU. We want to count in
2952 * the first case but not the second. If the context is
2953 * currently active then an inactive software counter must
2954 * be the second case. If it's not currently active then
2955 * we need to know whether the counter was active when the
2956 * context was last active, which we can determine by
2957 * comparing counter->tstamp_stopped with ctx->time.
2959 * We are within an RCU read-side critical section,
2960 * which protects the existence of *ctx.
2963 spin_lock_irqsave(&ctx
->lock
, flags
);
2965 /* Re-check state now we have the lock */
2966 if (counter
->state
< PERF_COUNTER_STATE_INACTIVE
||
2967 counter
->ctx
->is_active
||
2968 counter
->tstamp_stopped
< ctx
->time
)
2970 spin_unlock_irqrestore(&ctx
->lock
, flags
);
2974 static int perf_swcounter_match(struct perf_counter
*counter
,
2975 enum perf_event_types type
,
2976 u32 event
, struct pt_regs
*regs
)
2980 event_config
= ((u64
) type
<< PERF_COUNTER_TYPE_SHIFT
) | event
;
2982 if (!perf_swcounter_is_counting(counter
))
2985 if (counter
->hw_event
.config
!= event_config
)
2989 if (counter
->hw_event
.exclude_user
&& user_mode(regs
))
2992 if (counter
->hw_event
.exclude_kernel
&& !user_mode(regs
))
2999 static void perf_swcounter_add(struct perf_counter
*counter
, u64 nr
,
3000 int nmi
, struct pt_regs
*regs
, u64 addr
)
3002 int neg
= atomic64_add_negative(nr
, &counter
->hw
.count
);
3004 if (counter
->hw
.sample_period
&& !neg
&& regs
)
3005 perf_swcounter_overflow(counter
, nmi
, regs
, addr
);
3008 static void perf_swcounter_ctx_event(struct perf_counter_context
*ctx
,
3009 enum perf_event_types type
, u32 event
,
3010 u64 nr
, int nmi
, struct pt_regs
*regs
,
3013 struct perf_counter
*counter
;
3015 if (system_state
!= SYSTEM_RUNNING
|| list_empty(&ctx
->event_list
))
3019 list_for_each_entry_rcu(counter
, &ctx
->event_list
, event_entry
) {
3020 if (perf_swcounter_match(counter
, type
, event
, regs
))
3021 perf_swcounter_add(counter
, nr
, nmi
, regs
, addr
);
3026 static int *perf_swcounter_recursion_context(struct perf_cpu_context
*cpuctx
)
3029 return &cpuctx
->recursion
[3];
3032 return &cpuctx
->recursion
[2];
3035 return &cpuctx
->recursion
[1];
3037 return &cpuctx
->recursion
[0];
3040 static void __perf_swcounter_event(enum perf_event_types type
, u32 event
,
3041 u64 nr
, int nmi
, struct pt_regs
*regs
,
3044 struct perf_cpu_context
*cpuctx
= &get_cpu_var(perf_cpu_context
);
3045 int *recursion
= perf_swcounter_recursion_context(cpuctx
);
3046 struct perf_counter_context
*ctx
;
3054 perf_swcounter_ctx_event(&cpuctx
->ctx
, type
, event
,
3055 nr
, nmi
, regs
, addr
);
3058 * doesn't really matter which of the child contexts the
3059 * events ends up in.
3061 ctx
= rcu_dereference(current
->perf_counter_ctxp
);
3063 perf_swcounter_ctx_event(ctx
, type
, event
, nr
, nmi
, regs
, addr
);
3070 put_cpu_var(perf_cpu_context
);
3074 perf_swcounter_event(u32 event
, u64 nr
, int nmi
, struct pt_regs
*regs
, u64 addr
)
3076 __perf_swcounter_event(PERF_TYPE_SOFTWARE
, event
, nr
, nmi
, regs
, addr
);
3079 static void perf_swcounter_read(struct perf_counter
*counter
)
3081 perf_swcounter_update(counter
);
3084 static int perf_swcounter_enable(struct perf_counter
*counter
)
3086 perf_swcounter_set_period(counter
);
3090 static void perf_swcounter_disable(struct perf_counter
*counter
)
3092 perf_swcounter_update(counter
);
3095 static const struct pmu perf_ops_generic
= {
3096 .enable
= perf_swcounter_enable
,
3097 .disable
= perf_swcounter_disable
,
3098 .read
= perf_swcounter_read
,
3102 * Software counter: cpu wall time clock
3105 static void cpu_clock_perf_counter_update(struct perf_counter
*counter
)
3107 int cpu
= raw_smp_processor_id();
3111 now
= cpu_clock(cpu
);
3112 prev
= atomic64_read(&counter
->hw
.prev_count
);
3113 atomic64_set(&counter
->hw
.prev_count
, now
);
3114 atomic64_add(now
- prev
, &counter
->count
);
3117 static int cpu_clock_perf_counter_enable(struct perf_counter
*counter
)
3119 struct hw_perf_counter
*hwc
= &counter
->hw
;
3120 int cpu
= raw_smp_processor_id();
3122 atomic64_set(&hwc
->prev_count
, cpu_clock(cpu
));
3123 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3124 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3125 if (hwc
->sample_period
) {
3126 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3127 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3128 ns_to_ktime(period
), 0,
3129 HRTIMER_MODE_REL
, 0);
3135 static void cpu_clock_perf_counter_disable(struct perf_counter
*counter
)
3137 if (counter
->hw
.sample_period
)
3138 hrtimer_cancel(&counter
->hw
.hrtimer
);
3139 cpu_clock_perf_counter_update(counter
);
3142 static void cpu_clock_perf_counter_read(struct perf_counter
*counter
)
3144 cpu_clock_perf_counter_update(counter
);
3147 static const struct pmu perf_ops_cpu_clock
= {
3148 .enable
= cpu_clock_perf_counter_enable
,
3149 .disable
= cpu_clock_perf_counter_disable
,
3150 .read
= cpu_clock_perf_counter_read
,
3154 * Software counter: task time clock
3157 static void task_clock_perf_counter_update(struct perf_counter
*counter
, u64 now
)
3162 prev
= atomic64_xchg(&counter
->hw
.prev_count
, now
);
3164 atomic64_add(delta
, &counter
->count
);
3167 static int task_clock_perf_counter_enable(struct perf_counter
*counter
)
3169 struct hw_perf_counter
*hwc
= &counter
->hw
;
3172 now
= counter
->ctx
->time
;
3174 atomic64_set(&hwc
->prev_count
, now
);
3175 hrtimer_init(&hwc
->hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3176 hwc
->hrtimer
.function
= perf_swcounter_hrtimer
;
3177 if (hwc
->sample_period
) {
3178 u64 period
= max_t(u64
, 10000, hwc
->sample_period
);
3179 __hrtimer_start_range_ns(&hwc
->hrtimer
,
3180 ns_to_ktime(period
), 0,
3181 HRTIMER_MODE_REL
, 0);
3187 static void task_clock_perf_counter_disable(struct perf_counter
*counter
)
3189 if (counter
->hw
.sample_period
)
3190 hrtimer_cancel(&counter
->hw
.hrtimer
);
3191 task_clock_perf_counter_update(counter
, counter
->ctx
->time
);
3195 static void task_clock_perf_counter_read(struct perf_counter
*counter
)
3200 update_context_time(counter
->ctx
);
3201 time
= counter
->ctx
->time
;
3203 u64 now
= perf_clock();
3204 u64 delta
= now
- counter
->ctx
->timestamp
;
3205 time
= counter
->ctx
->time
+ delta
;
3208 task_clock_perf_counter_update(counter
, time
);
3211 static const struct pmu perf_ops_task_clock
= {
3212 .enable
= task_clock_perf_counter_enable
,
3213 .disable
= task_clock_perf_counter_disable
,
3214 .read
= task_clock_perf_counter_read
,
3218 * Software counter: cpu migrations
3220 void perf_counter_task_migration(struct task_struct
*task
, int cpu
)
3222 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3223 struct perf_counter_context
*ctx
;
3225 perf_swcounter_ctx_event(&cpuctx
->ctx
, PERF_TYPE_SOFTWARE
,
3226 PERF_COUNT_CPU_MIGRATIONS
,
3229 ctx
= perf_pin_task_context(task
);
3231 perf_swcounter_ctx_event(ctx
, PERF_TYPE_SOFTWARE
,
3232 PERF_COUNT_CPU_MIGRATIONS
,
3234 perf_unpin_context(ctx
);
3238 #ifdef CONFIG_EVENT_PROFILE
3239 void perf_tpcounter_event(int event_id
)
3241 struct pt_regs
*regs
= get_irq_regs();
3244 regs
= task_pt_regs(current
);
3246 __perf_swcounter_event(PERF_TYPE_TRACEPOINT
, event_id
, 1, 1, regs
, 0);
3248 EXPORT_SYMBOL_GPL(perf_tpcounter_event
);
3250 extern int ftrace_profile_enable(int);
3251 extern void ftrace_profile_disable(int);
3253 static void tp_perf_counter_destroy(struct perf_counter
*counter
)
3255 ftrace_profile_disable(perf_event_id(&counter
->hw_event
));
3258 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3260 int event_id
= perf_event_id(&counter
->hw_event
);
3263 ret
= ftrace_profile_enable(event_id
);
3267 counter
->destroy
= tp_perf_counter_destroy
;
3268 counter
->hw
.sample_period
= counter
->hw_event
.sample_period
;
3270 return &perf_ops_generic
;
3273 static const struct pmu
*tp_perf_counter_init(struct perf_counter
*counter
)
3279 static const struct pmu
*sw_perf_counter_init(struct perf_counter
*counter
)
3281 const struct pmu
*pmu
= NULL
;
3284 * Software counters (currently) can't in general distinguish
3285 * between user, kernel and hypervisor events.
3286 * However, context switches and cpu migrations are considered
3287 * to be kernel events, and page faults are never hypervisor
3290 switch (perf_event_id(&counter
->hw_event
)) {
3291 case PERF_COUNT_CPU_CLOCK
:
3292 pmu
= &perf_ops_cpu_clock
;
3295 case PERF_COUNT_TASK_CLOCK
:
3297 * If the user instantiates this as a per-cpu counter,
3298 * use the cpu_clock counter instead.
3300 if (counter
->ctx
->task
)
3301 pmu
= &perf_ops_task_clock
;
3303 pmu
= &perf_ops_cpu_clock
;
3306 case PERF_COUNT_PAGE_FAULTS
:
3307 case PERF_COUNT_PAGE_FAULTS_MIN
:
3308 case PERF_COUNT_PAGE_FAULTS_MAJ
:
3309 case PERF_COUNT_CONTEXT_SWITCHES
:
3310 case PERF_COUNT_CPU_MIGRATIONS
:
3311 pmu
= &perf_ops_generic
;
3319 * Allocate and initialize a counter structure
3321 static struct perf_counter
*
3322 perf_counter_alloc(struct perf_counter_hw_event
*hw_event
,
3324 struct perf_counter_context
*ctx
,
3325 struct perf_counter
*group_leader
,
3328 const struct pmu
*pmu
;
3329 struct perf_counter
*counter
;
3330 struct hw_perf_counter
*hwc
;
3333 counter
= kzalloc(sizeof(*counter
), gfpflags
);
3335 return ERR_PTR(-ENOMEM
);
3338 * Single counters are their own group leaders, with an
3339 * empty sibling list:
3342 group_leader
= counter
;
3344 mutex_init(&counter
->child_mutex
);
3345 INIT_LIST_HEAD(&counter
->child_list
);
3347 INIT_LIST_HEAD(&counter
->list_entry
);
3348 INIT_LIST_HEAD(&counter
->event_entry
);
3349 INIT_LIST_HEAD(&counter
->sibling_list
);
3350 init_waitqueue_head(&counter
->waitq
);
3352 mutex_init(&counter
->mmap_mutex
);
3355 counter
->hw_event
= *hw_event
;
3356 counter
->group_leader
= group_leader
;
3357 counter
->pmu
= NULL
;
3359 counter
->oncpu
= -1;
3361 counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3362 if (hw_event
->disabled
)
3363 counter
->state
= PERF_COUNTER_STATE_OFF
;
3368 if (hw_event
->freq
&& hw_event
->sample_freq
)
3369 hwc
->sample_period
= div64_u64(TICK_NSEC
, hw_event
->sample_freq
);
3371 hwc
->sample_period
= hw_event
->sample_period
;
3374 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3376 if (hw_event
->inherit
&& (hw_event
->sample_type
& PERF_SAMPLE_GROUP
))
3379 if (perf_event_raw(hw_event
)) {
3380 pmu
= hw_perf_counter_init(counter
);
3384 switch (perf_event_type(hw_event
)) {
3385 case PERF_TYPE_HARDWARE
:
3386 pmu
= hw_perf_counter_init(counter
);
3389 case PERF_TYPE_SOFTWARE
:
3390 pmu
= sw_perf_counter_init(counter
);
3393 case PERF_TYPE_TRACEPOINT
:
3394 pmu
= tp_perf_counter_init(counter
);
3401 else if (IS_ERR(pmu
))
3406 return ERR_PTR(err
);
3411 atomic_inc(&nr_counters
);
3412 if (counter
->hw_event
.mmap
)
3413 atomic_inc(&nr_mmap_tracking
);
3414 if (counter
->hw_event
.munmap
)
3415 atomic_inc(&nr_munmap_tracking
);
3416 if (counter
->hw_event
.comm
)
3417 atomic_inc(&nr_comm_tracking
);
3422 static atomic64_t perf_counter_id
;
3425 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3427 * @hw_event_uptr: event type attributes for monitoring/sampling
3430 * @group_fd: group leader counter fd
3432 SYSCALL_DEFINE5(perf_counter_open
,
3433 const struct perf_counter_hw_event __user
*, hw_event_uptr
,
3434 pid_t
, pid
, int, cpu
, int, group_fd
, unsigned long, flags
)
3436 struct perf_counter
*counter
, *group_leader
;
3437 struct perf_counter_hw_event hw_event
;
3438 struct perf_counter_context
*ctx
;
3439 struct file
*counter_file
= NULL
;
3440 struct file
*group_file
= NULL
;
3441 int fput_needed
= 0;
3442 int fput_needed2
= 0;
3445 /* for future expandability... */
3449 if (copy_from_user(&hw_event
, hw_event_uptr
, sizeof(hw_event
)) != 0)
3453 * Get the target context (task or percpu):
3455 ctx
= find_get_context(pid
, cpu
);
3457 return PTR_ERR(ctx
);
3460 * Look up the group leader (we will attach this counter to it):
3462 group_leader
= NULL
;
3463 if (group_fd
!= -1) {
3465 group_file
= fget_light(group_fd
, &fput_needed
);
3467 goto err_put_context
;
3468 if (group_file
->f_op
!= &perf_fops
)
3469 goto err_put_context
;
3471 group_leader
= group_file
->private_data
;
3473 * Do not allow a recursive hierarchy (this new sibling
3474 * becoming part of another group-sibling):
3476 if (group_leader
->group_leader
!= group_leader
)
3477 goto err_put_context
;
3479 * Do not allow to attach to a group in a different
3480 * task or CPU context:
3482 if (group_leader
->ctx
!= ctx
)
3483 goto err_put_context
;
3485 * Only a group leader can be exclusive or pinned
3487 if (hw_event
.exclusive
|| hw_event
.pinned
)
3488 goto err_put_context
;
3491 counter
= perf_counter_alloc(&hw_event
, cpu
, ctx
, group_leader
,
3493 ret
= PTR_ERR(counter
);
3494 if (IS_ERR(counter
))
3495 goto err_put_context
;
3497 ret
= anon_inode_getfd("[perf_counter]", &perf_fops
, counter
, 0);
3499 goto err_free_put_context
;
3501 counter_file
= fget_light(ret
, &fput_needed2
);
3503 goto err_free_put_context
;
3505 counter
->filp
= counter_file
;
3506 WARN_ON_ONCE(ctx
->parent_ctx
);
3507 mutex_lock(&ctx
->mutex
);
3508 perf_install_in_context(ctx
, counter
, cpu
);
3510 mutex_unlock(&ctx
->mutex
);
3512 counter
->owner
= current
;
3513 get_task_struct(current
);
3514 mutex_lock(¤t
->perf_counter_mutex
);
3515 list_add_tail(&counter
->owner_entry
, ¤t
->perf_counter_list
);
3516 mutex_unlock(¤t
->perf_counter_mutex
);
3518 counter
->ns
= get_pid_ns(current
->nsproxy
->pid_ns
);
3519 counter
->id
= atomic64_inc_return(&perf_counter_id
);
3521 fput_light(counter_file
, fput_needed2
);
3524 fput_light(group_file
, fput_needed
);
3528 err_free_put_context
:
3538 * inherit a counter from parent task to child task:
3540 static struct perf_counter
*
3541 inherit_counter(struct perf_counter
*parent_counter
,
3542 struct task_struct
*parent
,
3543 struct perf_counter_context
*parent_ctx
,
3544 struct task_struct
*child
,
3545 struct perf_counter
*group_leader
,
3546 struct perf_counter_context
*child_ctx
)
3548 struct perf_counter
*child_counter
;
3551 * Instead of creating recursive hierarchies of counters,
3552 * we link inherited counters back to the original parent,
3553 * which has a filp for sure, which we use as the reference
3556 if (parent_counter
->parent
)
3557 parent_counter
= parent_counter
->parent
;
3559 child_counter
= perf_counter_alloc(&parent_counter
->hw_event
,
3560 parent_counter
->cpu
, child_ctx
,
3561 group_leader
, GFP_KERNEL
);
3562 if (IS_ERR(child_counter
))
3563 return child_counter
;
3567 * Make the child state follow the state of the parent counter,
3568 * not its hw_event.disabled bit. We hold the parent's mutex,
3569 * so we won't race with perf_counter_{en, dis}able_family.
3571 if (parent_counter
->state
>= PERF_COUNTER_STATE_INACTIVE
)
3572 child_counter
->state
= PERF_COUNTER_STATE_INACTIVE
;
3574 child_counter
->state
= PERF_COUNTER_STATE_OFF
;
3577 * Link it up in the child's context:
3579 add_counter_to_ctx(child_counter
, child_ctx
);
3581 child_counter
->parent
= parent_counter
;
3583 * inherit into child's child as well:
3585 child_counter
->hw_event
.inherit
= 1;
3588 * Get a reference to the parent filp - we will fput it
3589 * when the child counter exits. This is safe to do because
3590 * we are in the parent and we know that the filp still
3591 * exists and has a nonzero count:
3593 atomic_long_inc(&parent_counter
->filp
->f_count
);
3596 * Link this into the parent counter's child list
3598 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3599 mutex_lock(&parent_counter
->child_mutex
);
3600 list_add_tail(&child_counter
->child_list
, &parent_counter
->child_list
);
3601 mutex_unlock(&parent_counter
->child_mutex
);
3603 return child_counter
;
3606 static int inherit_group(struct perf_counter
*parent_counter
,
3607 struct task_struct
*parent
,
3608 struct perf_counter_context
*parent_ctx
,
3609 struct task_struct
*child
,
3610 struct perf_counter_context
*child_ctx
)
3612 struct perf_counter
*leader
;
3613 struct perf_counter
*sub
;
3614 struct perf_counter
*child_ctr
;
3616 leader
= inherit_counter(parent_counter
, parent
, parent_ctx
,
3617 child
, NULL
, child_ctx
);
3619 return PTR_ERR(leader
);
3620 list_for_each_entry(sub
, &parent_counter
->sibling_list
, list_entry
) {
3621 child_ctr
= inherit_counter(sub
, parent
, parent_ctx
,
3622 child
, leader
, child_ctx
);
3623 if (IS_ERR(child_ctr
))
3624 return PTR_ERR(child_ctr
);
3629 static void sync_child_counter(struct perf_counter
*child_counter
,
3630 struct perf_counter
*parent_counter
)
3634 child_val
= atomic64_read(&child_counter
->count
);
3637 * Add back the child's count to the parent's count:
3639 atomic64_add(child_val
, &parent_counter
->count
);
3640 atomic64_add(child_counter
->total_time_enabled
,
3641 &parent_counter
->child_total_time_enabled
);
3642 atomic64_add(child_counter
->total_time_running
,
3643 &parent_counter
->child_total_time_running
);
3646 * Remove this counter from the parent's list
3648 WARN_ON_ONCE(parent_counter
->ctx
->parent_ctx
);
3649 mutex_lock(&parent_counter
->child_mutex
);
3650 list_del_init(&child_counter
->child_list
);
3651 mutex_unlock(&parent_counter
->child_mutex
);
3654 * Release the parent counter, if this was the last
3657 fput(parent_counter
->filp
);
3661 __perf_counter_exit_task(struct perf_counter
*child_counter
,
3662 struct perf_counter_context
*child_ctx
)
3664 struct perf_counter
*parent_counter
;
3666 update_counter_times(child_counter
);
3667 perf_counter_remove_from_context(child_counter
);
3669 parent_counter
= child_counter
->parent
;
3671 * It can happen that parent exits first, and has counters
3672 * that are still around due to the child reference. These
3673 * counters need to be zapped - but otherwise linger.
3675 if (parent_counter
) {
3676 sync_child_counter(child_counter
, parent_counter
);
3677 free_counter(child_counter
);
3682 * When a child task exits, feed back counter values to parent counters.
3684 void perf_counter_exit_task(struct task_struct
*child
)
3686 struct perf_counter
*child_counter
, *tmp
;
3687 struct perf_counter_context
*child_ctx
;
3688 unsigned long flags
;
3690 if (likely(!child
->perf_counter_ctxp
))
3693 local_irq_save(flags
);
3695 * We can't reschedule here because interrupts are disabled,
3696 * and either child is current or it is a task that can't be
3697 * scheduled, so we are now safe from rescheduling changing
3700 child_ctx
= child
->perf_counter_ctxp
;
3701 __perf_counter_task_sched_out(child_ctx
);
3704 * Take the context lock here so that if find_get_context is
3705 * reading child->perf_counter_ctxp, we wait until it has
3706 * incremented the context's refcount before we do put_ctx below.
3708 spin_lock(&child_ctx
->lock
);
3709 child
->perf_counter_ctxp
= NULL
;
3710 if (child_ctx
->parent_ctx
) {
3712 * This context is a clone; unclone it so it can't get
3713 * swapped to another process while we're removing all
3714 * the counters from it.
3716 put_ctx(child_ctx
->parent_ctx
);
3717 child_ctx
->parent_ctx
= NULL
;
3719 spin_unlock(&child_ctx
->lock
);
3720 local_irq_restore(flags
);
3722 mutex_lock(&child_ctx
->mutex
);
3725 list_for_each_entry_safe(child_counter
, tmp
, &child_ctx
->counter_list
,
3727 __perf_counter_exit_task(child_counter
, child_ctx
);
3730 * If the last counter was a group counter, it will have appended all
3731 * its siblings to the list, but we obtained 'tmp' before that which
3732 * will still point to the list head terminating the iteration.
3734 if (!list_empty(&child_ctx
->counter_list
))
3737 mutex_unlock(&child_ctx
->mutex
);
3743 * free an unexposed, unused context as created by inheritance by
3744 * init_task below, used by fork() in case of fail.
3746 void perf_counter_free_task(struct task_struct
*task
)
3748 struct perf_counter_context
*ctx
= task
->perf_counter_ctxp
;
3749 struct perf_counter
*counter
, *tmp
;
3754 mutex_lock(&ctx
->mutex
);
3756 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
) {
3757 struct perf_counter
*parent
= counter
->parent
;
3759 if (WARN_ON_ONCE(!parent
))
3762 mutex_lock(&parent
->child_mutex
);
3763 list_del_init(&counter
->child_list
);
3764 mutex_unlock(&parent
->child_mutex
);
3768 list_del_counter(counter
, ctx
);
3769 free_counter(counter
);
3772 if (!list_empty(&ctx
->counter_list
))
3775 mutex_unlock(&ctx
->mutex
);
3781 * Initialize the perf_counter context in task_struct
3783 int perf_counter_init_task(struct task_struct
*child
)
3785 struct perf_counter_context
*child_ctx
, *parent_ctx
;
3786 struct perf_counter_context
*cloned_ctx
;
3787 struct perf_counter
*counter
;
3788 struct task_struct
*parent
= current
;
3789 int inherited_all
= 1;
3792 child
->perf_counter_ctxp
= NULL
;
3794 mutex_init(&child
->perf_counter_mutex
);
3795 INIT_LIST_HEAD(&child
->perf_counter_list
);
3797 if (likely(!parent
->perf_counter_ctxp
))
3801 * This is executed from the parent task context, so inherit
3802 * counters that have been marked for cloning.
3803 * First allocate and initialize a context for the child.
3806 child_ctx
= kmalloc(sizeof(struct perf_counter_context
), GFP_KERNEL
);
3810 __perf_counter_init_context(child_ctx
, child
);
3811 child
->perf_counter_ctxp
= child_ctx
;
3812 get_task_struct(child
);
3815 * If the parent's context is a clone, pin it so it won't get
3818 parent_ctx
= perf_pin_task_context(parent
);
3821 * No need to check if parent_ctx != NULL here; since we saw
3822 * it non-NULL earlier, the only reason for it to become NULL
3823 * is if we exit, and since we're currently in the middle of
3824 * a fork we can't be exiting at the same time.
3828 * Lock the parent list. No need to lock the child - not PID
3829 * hashed yet and not running, so nobody can access it.
3831 mutex_lock(&parent_ctx
->mutex
);
3834 * We dont have to disable NMIs - we are only looking at
3835 * the list, not manipulating it:
3837 list_for_each_entry_rcu(counter
, &parent_ctx
->event_list
, event_entry
) {
3838 if (counter
!= counter
->group_leader
)
3841 if (!counter
->hw_event
.inherit
) {
3846 ret
= inherit_group(counter
, parent
, parent_ctx
,
3854 if (inherited_all
) {
3856 * Mark the child context as a clone of the parent
3857 * context, or of whatever the parent is a clone of.
3858 * Note that if the parent is a clone, it could get
3859 * uncloned at any point, but that doesn't matter
3860 * because the list of counters and the generation
3861 * count can't have changed since we took the mutex.
3863 cloned_ctx
= rcu_dereference(parent_ctx
->parent_ctx
);
3865 child_ctx
->parent_ctx
= cloned_ctx
;
3866 child_ctx
->parent_gen
= parent_ctx
->parent_gen
;
3868 child_ctx
->parent_ctx
= parent_ctx
;
3869 child_ctx
->parent_gen
= parent_ctx
->generation
;
3871 get_ctx(child_ctx
->parent_ctx
);
3874 mutex_unlock(&parent_ctx
->mutex
);
3876 perf_unpin_context(parent_ctx
);
3881 static void __cpuinit
perf_counter_init_cpu(int cpu
)
3883 struct perf_cpu_context
*cpuctx
;
3885 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3886 __perf_counter_init_context(&cpuctx
->ctx
, NULL
);
3888 spin_lock(&perf_resource_lock
);
3889 cpuctx
->max_pertask
= perf_max_counters
- perf_reserved_percpu
;
3890 spin_unlock(&perf_resource_lock
);
3892 hw_perf_counter_setup(cpu
);
3895 #ifdef CONFIG_HOTPLUG_CPU
3896 static void __perf_counter_exit_cpu(void *info
)
3898 struct perf_cpu_context
*cpuctx
= &__get_cpu_var(perf_cpu_context
);
3899 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3900 struct perf_counter
*counter
, *tmp
;
3902 list_for_each_entry_safe(counter
, tmp
, &ctx
->counter_list
, list_entry
)
3903 __perf_counter_remove_from_context(counter
);
3905 static void perf_counter_exit_cpu(int cpu
)
3907 struct perf_cpu_context
*cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3908 struct perf_counter_context
*ctx
= &cpuctx
->ctx
;
3910 mutex_lock(&ctx
->mutex
);
3911 smp_call_function_single(cpu
, __perf_counter_exit_cpu
, NULL
, 1);
3912 mutex_unlock(&ctx
->mutex
);
3915 static inline void perf_counter_exit_cpu(int cpu
) { }
3918 static int __cpuinit
3919 perf_cpu_notify(struct notifier_block
*self
, unsigned long action
, void *hcpu
)
3921 unsigned int cpu
= (long)hcpu
;
3925 case CPU_UP_PREPARE
:
3926 case CPU_UP_PREPARE_FROZEN
:
3927 perf_counter_init_cpu(cpu
);
3930 case CPU_DOWN_PREPARE
:
3931 case CPU_DOWN_PREPARE_FROZEN
:
3932 perf_counter_exit_cpu(cpu
);
3943 * This has to have a higher priority than migration_notifier in sched.c.
3945 static struct notifier_block __cpuinitdata perf_cpu_nb
= {
3946 .notifier_call
= perf_cpu_notify
,
3950 void __init
perf_counter_init(void)
3952 perf_cpu_notify(&perf_cpu_nb
, (unsigned long)CPU_UP_PREPARE
,
3953 (void *)(long)smp_processor_id());
3954 register_cpu_notifier(&perf_cpu_nb
);
3957 static ssize_t
perf_show_reserve_percpu(struct sysdev_class
*class, char *buf
)
3959 return sprintf(buf
, "%d\n", perf_reserved_percpu
);
3963 perf_set_reserve_percpu(struct sysdev_class
*class,
3967 struct perf_cpu_context
*cpuctx
;
3971 err
= strict_strtoul(buf
, 10, &val
);
3974 if (val
> perf_max_counters
)
3977 spin_lock(&perf_resource_lock
);
3978 perf_reserved_percpu
= val
;
3979 for_each_online_cpu(cpu
) {
3980 cpuctx
= &per_cpu(perf_cpu_context
, cpu
);
3981 spin_lock_irq(&cpuctx
->ctx
.lock
);
3982 mpt
= min(perf_max_counters
- cpuctx
->ctx
.nr_counters
,
3983 perf_max_counters
- perf_reserved_percpu
);
3984 cpuctx
->max_pertask
= mpt
;
3985 spin_unlock_irq(&cpuctx
->ctx
.lock
);
3987 spin_unlock(&perf_resource_lock
);
3992 static ssize_t
perf_show_overcommit(struct sysdev_class
*class, char *buf
)
3994 return sprintf(buf
, "%d\n", perf_overcommit
);
3998 perf_set_overcommit(struct sysdev_class
*class, const char *buf
, size_t count
)
4003 err
= strict_strtoul(buf
, 10, &val
);
4009 spin_lock(&perf_resource_lock
);
4010 perf_overcommit
= val
;
4011 spin_unlock(&perf_resource_lock
);
4016 static SYSDEV_CLASS_ATTR(
4019 perf_show_reserve_percpu
,
4020 perf_set_reserve_percpu
4023 static SYSDEV_CLASS_ATTR(
4026 perf_show_overcommit
,
4030 static struct attribute
*perfclass_attrs
[] = {
4031 &attr_reserve_percpu
.attr
,
4032 &attr_overcommit
.attr
,
4036 static struct attribute_group perfclass_attr_group
= {
4037 .attrs
= perfclass_attrs
,
4038 .name
= "perf_counters",
4041 static int __init
perf_counter_sysfs_init(void)
4043 return sysfs_create_group(&cpu_sysdev_class
.kset
.kobj
,
4044 &perfclass_attr_group
);
4046 device_initcall(perf_counter_sysfs_init
);