2 * Performance counter core code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
7 * For licencing details see kernel-base/COPYING
11 #include <linux/cpu.h>
12 #include <linux/smp.h>
13 #include <linux/file.h>
14 #include <linux/poll.h>
15 #include <linux/sysfs.h>
16 #include <linux/ptrace.h>
17 #include <linux/percpu.h>
18 #include <linux/uaccess.h>
19 #include <linux/syscalls.h>
20 #include <linux/anon_inodes.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/perf_counter.h>
24 #include <linux/vmstat.h>
27 * Each CPU has a list of per CPU counters:
29 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
31 int perf_max_counters __read_mostly = 1;
32 static int perf_reserved_percpu __read_mostly;
33 static int perf_overcommit __read_mostly = 1;
36 * Mutex for (sysadmin-configurable) counter reservations:
38 static DEFINE_MUTEX(perf_resource_mutex);
41 * Architecture provided APIs - weak aliases:
43 extern __weak const struct hw_perf_counter_ops *
44 hw_perf_counter_init(struct perf_counter *counter)
49 u64 __weak hw_perf_save_disable(void) { return 0; }
50 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
51 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
52 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
53 struct perf_cpu_context *cpuctx,
54 struct perf_counter_context *ctx, int cpu)
59 void __weak perf_counter_print_debug(void) { }
62 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
64 struct perf_counter *group_leader = counter->group_leader;
67 * Depending on whether it is a standalone or sibling counter,
68 * add it straight to the context's counter list, or to the group
69 * leader's sibling list:
71 if (counter->group_leader == counter)
72 list_add_tail(&counter->list_entry, &ctx->counter_list);
74 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
78 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
80 struct perf_counter *sibling, *tmp;
82 list_del_init(&counter->list_entry);
85 * If this was a group counter with sibling counters then
86 * upgrade the siblings to singleton counters by adding them
87 * to the context list directly:
89 list_for_each_entry_safe(sibling, tmp,
90 &counter->sibling_list, list_entry) {
92 list_del_init(&sibling->list_entry);
93 list_add_tail(&sibling->list_entry, &ctx->counter_list);
94 sibling->group_leader = sibling;
99 counter_sched_out(struct perf_counter *counter,
100 struct perf_cpu_context *cpuctx,
101 struct perf_counter_context *ctx)
103 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
106 counter->state = PERF_COUNTER_STATE_INACTIVE;
107 counter->hw_ops->disable(counter);
110 if (!is_software_counter(counter))
111 cpuctx->active_oncpu--;
113 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
114 cpuctx->exclusive = 0;
118 group_sched_out(struct perf_counter *group_counter,
119 struct perf_cpu_context *cpuctx,
120 struct perf_counter_context *ctx)
122 struct perf_counter *counter;
124 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
127 counter_sched_out(group_counter, cpuctx, ctx);
130 * Schedule out siblings (if any):
132 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
133 counter_sched_out(counter, cpuctx, ctx);
135 if (group_counter->hw_event.exclusive)
136 cpuctx->exclusive = 0;
140 * Cross CPU call to remove a performance counter
142 * We disable the counter on the hardware level first. After that we
143 * remove it from the context list.
145 static void __perf_counter_remove_from_context(void *info)
147 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
148 struct perf_counter *counter = info;
149 struct perf_counter_context *ctx = counter->ctx;
154 * If this is a task context, we need to check whether it is
155 * the current task context of this cpu. If not it has been
156 * scheduled out before the smp call arrived.
158 if (ctx->task && cpuctx->task_ctx != ctx)
161 curr_rq_lock_irq_save(&flags);
162 spin_lock(&ctx->lock);
164 counter_sched_out(counter, cpuctx, ctx);
166 counter->task = NULL;
170 * Protect the list operation against NMI by disabling the
171 * counters on a global level. NOP for non NMI based counters.
173 perf_flags = hw_perf_save_disable();
174 list_del_counter(counter, ctx);
175 hw_perf_restore(perf_flags);
179 * Allow more per task counters with respect to the
182 cpuctx->max_pertask =
183 min(perf_max_counters - ctx->nr_counters,
184 perf_max_counters - perf_reserved_percpu);
187 spin_unlock(&ctx->lock);
188 curr_rq_unlock_irq_restore(&flags);
193 * Remove the counter from a task's (or a CPU's) list of counters.
195 * Must be called with counter->mutex and ctx->mutex held.
197 * CPU counters are removed with a smp call. For task counters we only
198 * call when the task is on a CPU.
200 static void perf_counter_remove_from_context(struct perf_counter *counter)
202 struct perf_counter_context *ctx = counter->ctx;
203 struct task_struct *task = ctx->task;
207 * Per cpu counters are removed via an smp call and
208 * the removal is always sucessful.
210 smp_call_function_single(counter->cpu,
211 __perf_counter_remove_from_context,
217 task_oncpu_function_call(task, __perf_counter_remove_from_context,
220 spin_lock_irq(&ctx->lock);
222 * If the context is active we need to retry the smp call.
224 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
225 spin_unlock_irq(&ctx->lock);
230 * The lock prevents that this context is scheduled in so we
231 * can remove the counter safely, if the call above did not
234 if (!list_empty(&counter->list_entry)) {
236 list_del_counter(counter, ctx);
237 counter->task = NULL;
239 spin_unlock_irq(&ctx->lock);
243 * Cross CPU call to disable a performance counter
245 static void __perf_counter_disable(void *info)
247 struct perf_counter *counter = info;
248 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
249 struct perf_counter_context *ctx = counter->ctx;
253 * If this is a per-task counter, need to check whether this
254 * counter's task is the current task on this cpu.
256 if (ctx->task && cpuctx->task_ctx != ctx)
259 curr_rq_lock_irq_save(&flags);
260 spin_lock(&ctx->lock);
263 * If the counter is on, turn it off.
264 * If it is in error state, leave it in error state.
266 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
267 if (counter == counter->group_leader)
268 group_sched_out(counter, cpuctx, ctx);
270 counter_sched_out(counter, cpuctx, ctx);
271 counter->state = PERF_COUNTER_STATE_OFF;
274 spin_unlock(&ctx->lock);
275 curr_rq_unlock_irq_restore(&flags);
281 static void perf_counter_disable(struct perf_counter *counter)
283 struct perf_counter_context *ctx = counter->ctx;
284 struct task_struct *task = ctx->task;
288 * Disable the counter on the cpu that it's on
290 smp_call_function_single(counter->cpu, __perf_counter_disable,
296 task_oncpu_function_call(task, __perf_counter_disable, counter);
298 spin_lock_irq(&ctx->lock);
300 * If the counter is still active, we need to retry the cross-call.
302 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
303 spin_unlock_irq(&ctx->lock);
308 * Since we have the lock this context can't be scheduled
309 * in, so we can change the state safely.
311 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
312 counter->state = PERF_COUNTER_STATE_OFF;
314 spin_unlock_irq(&ctx->lock);
318 * Disable a counter and all its children.
320 static void perf_counter_disable_family(struct perf_counter *counter)
322 struct perf_counter *child;
324 perf_counter_disable(counter);
327 * Lock the mutex to protect the list of children
329 mutex_lock(&counter->mutex);
330 list_for_each_entry(child, &counter->child_list, child_list)
331 perf_counter_disable(child);
332 mutex_unlock(&counter->mutex);
336 counter_sched_in(struct perf_counter *counter,
337 struct perf_cpu_context *cpuctx,
338 struct perf_counter_context *ctx,
341 if (counter->state <= PERF_COUNTER_STATE_OFF)
344 counter->state = PERF_COUNTER_STATE_ACTIVE;
345 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
347 * The new state must be visible before we turn it on in the hardware:
351 if (counter->hw_ops->enable(counter)) {
352 counter->state = PERF_COUNTER_STATE_INACTIVE;
357 if (!is_software_counter(counter))
358 cpuctx->active_oncpu++;
361 if (counter->hw_event.exclusive)
362 cpuctx->exclusive = 1;
368 * Return 1 for a group consisting entirely of software counters,
369 * 0 if the group contains any hardware counters.
371 static int is_software_only_group(struct perf_counter *leader)
373 struct perf_counter *counter;
375 if (!is_software_counter(leader))
377 list_for_each_entry(counter, &leader->sibling_list, list_entry)
378 if (!is_software_counter(counter))
384 * Work out whether we can put this counter group on the CPU now.
386 static int group_can_go_on(struct perf_counter *counter,
387 struct perf_cpu_context *cpuctx,
391 * Groups consisting entirely of software counters can always go on.
393 if (is_software_only_group(counter))
396 * If an exclusive group is already on, no other hardware
397 * counters can go on.
399 if (cpuctx->exclusive)
402 * If this group is exclusive and there are already
403 * counters on the CPU, it can't go on.
405 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
408 * Otherwise, try to add it if all previous groups were able
415 * Cross CPU call to install and enable a performance counter
417 static void __perf_install_in_context(void *info)
419 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
420 struct perf_counter *counter = info;
421 struct perf_counter_context *ctx = counter->ctx;
422 struct perf_counter *leader = counter->group_leader;
423 int cpu = smp_processor_id();
429 * If this is a task context, we need to check whether it is
430 * the current task context of this cpu. If not it has been
431 * scheduled out before the smp call arrived.
433 if (ctx->task && cpuctx->task_ctx != ctx)
436 curr_rq_lock_irq_save(&flags);
437 spin_lock(&ctx->lock);
440 * Protect the list operation against NMI by disabling the
441 * counters on a global level. NOP for non NMI based counters.
443 perf_flags = hw_perf_save_disable();
445 list_add_counter(counter, ctx);
449 * Don't put the counter on if it is disabled or if
450 * it is in a group and the group isn't on.
452 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
453 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
457 * An exclusive counter can't go on if there are already active
458 * hardware counters, and no hardware counter can go on if there
459 * is already an exclusive counter on.
461 if (!group_can_go_on(counter, cpuctx, 1))
464 err = counter_sched_in(counter, cpuctx, ctx, cpu);
468 * This counter couldn't go on. If it is in a group
469 * then we have to pull the whole group off.
470 * If the counter group is pinned then put it in error state.
472 if (leader != counter)
473 group_sched_out(leader, cpuctx, ctx);
474 if (leader->hw_event.pinned)
475 leader->state = PERF_COUNTER_STATE_ERROR;
478 if (!err && !ctx->task && cpuctx->max_pertask)
479 cpuctx->max_pertask--;
482 hw_perf_restore(perf_flags);
484 spin_unlock(&ctx->lock);
485 curr_rq_unlock_irq_restore(&flags);
489 * Attach a performance counter to a context
491 * First we add the counter to the list with the hardware enable bit
492 * in counter->hw_config cleared.
494 * If the counter is attached to a task which is on a CPU we use a smp
495 * call to enable it in the task context. The task might have been
496 * scheduled away, but we check this in the smp call again.
498 * Must be called with ctx->mutex held.
501 perf_install_in_context(struct perf_counter_context *ctx,
502 struct perf_counter *counter,
505 struct task_struct *task = ctx->task;
509 * Per cpu counters are installed via an smp call and
510 * the install is always sucessful.
512 smp_call_function_single(cpu, __perf_install_in_context,
517 counter->task = task;
519 task_oncpu_function_call(task, __perf_install_in_context,
522 spin_lock_irq(&ctx->lock);
524 * we need to retry the smp call.
526 if (ctx->is_active && list_empty(&counter->list_entry)) {
527 spin_unlock_irq(&ctx->lock);
532 * The lock prevents that this context is scheduled in so we
533 * can add the counter safely, if it the call above did not
536 if (list_empty(&counter->list_entry)) {
537 list_add_counter(counter, ctx);
540 spin_unlock_irq(&ctx->lock);
544 * Cross CPU call to enable a performance counter
546 static void __perf_counter_enable(void *info)
548 struct perf_counter *counter = info;
549 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
550 struct perf_counter_context *ctx = counter->ctx;
551 struct perf_counter *leader = counter->group_leader;
556 * If this is a per-task counter, need to check whether this
557 * counter's task is the current task on this cpu.
559 if (ctx->task && cpuctx->task_ctx != ctx)
562 curr_rq_lock_irq_save(&flags);
563 spin_lock(&ctx->lock);
565 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
567 counter->state = PERF_COUNTER_STATE_INACTIVE;
570 * If the counter is in a group and isn't the group leader,
571 * then don't put it on unless the group is on.
573 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
576 if (!group_can_go_on(counter, cpuctx, 1))
579 err = counter_sched_in(counter, cpuctx, ctx,
584 * If this counter can't go on and it's part of a
585 * group, then the whole group has to come off.
587 if (leader != counter)
588 group_sched_out(leader, cpuctx, ctx);
589 if (leader->hw_event.pinned)
590 leader->state = PERF_COUNTER_STATE_ERROR;
594 spin_unlock(&ctx->lock);
595 curr_rq_unlock_irq_restore(&flags);
601 static void perf_counter_enable(struct perf_counter *counter)
603 struct perf_counter_context *ctx = counter->ctx;
604 struct task_struct *task = ctx->task;
608 * Enable the counter on the cpu that it's on
610 smp_call_function_single(counter->cpu, __perf_counter_enable,
615 spin_lock_irq(&ctx->lock);
616 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
620 * If the counter is in error state, clear that first.
621 * That way, if we see the counter in error state below, we
622 * know that it has gone back into error state, as distinct
623 * from the task having been scheduled away before the
624 * cross-call arrived.
626 if (counter->state == PERF_COUNTER_STATE_ERROR)
627 counter->state = PERF_COUNTER_STATE_OFF;
630 spin_unlock_irq(&ctx->lock);
631 task_oncpu_function_call(task, __perf_counter_enable, counter);
633 spin_lock_irq(&ctx->lock);
636 * If the context is active and the counter is still off,
637 * we need to retry the cross-call.
639 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
643 * Since we have the lock this context can't be scheduled
644 * in, so we can change the state safely.
646 if (counter->state == PERF_COUNTER_STATE_OFF)
647 counter->state = PERF_COUNTER_STATE_INACTIVE;
649 spin_unlock_irq(&ctx->lock);
653 * Enable a counter and all its children.
655 static void perf_counter_enable_family(struct perf_counter *counter)
657 struct perf_counter *child;
659 perf_counter_enable(counter);
662 * Lock the mutex to protect the list of children
664 mutex_lock(&counter->mutex);
665 list_for_each_entry(child, &counter->child_list, child_list)
666 perf_counter_enable(child);
667 mutex_unlock(&counter->mutex);
670 void __perf_counter_sched_out(struct perf_counter_context *ctx,
671 struct perf_cpu_context *cpuctx)
673 struct perf_counter *counter;
676 spin_lock(&ctx->lock);
678 if (likely(!ctx->nr_counters))
681 flags = hw_perf_save_disable();
682 if (ctx->nr_active) {
683 list_for_each_entry(counter, &ctx->counter_list, list_entry)
684 group_sched_out(counter, cpuctx, ctx);
686 hw_perf_restore(flags);
688 spin_unlock(&ctx->lock);
692 * Called from scheduler to remove the counters of the current task,
693 * with interrupts disabled.
695 * We stop each counter and update the counter value in counter->count.
697 * This does not protect us against NMI, but disable()
698 * sets the disabled bit in the control field of counter _before_
699 * accessing the counter control register. If a NMI hits, then it will
700 * not restart the counter.
702 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
704 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
705 struct perf_counter_context *ctx = &task->perf_counter_ctx;
707 if (likely(!cpuctx->task_ctx))
710 __perf_counter_sched_out(ctx, cpuctx);
712 cpuctx->task_ctx = NULL;
715 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
717 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
721 group_sched_in(struct perf_counter *group_counter,
722 struct perf_cpu_context *cpuctx,
723 struct perf_counter_context *ctx,
726 struct perf_counter *counter, *partial_group;
729 if (group_counter->state == PERF_COUNTER_STATE_OFF)
732 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
734 return ret < 0 ? ret : 0;
736 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
740 * Schedule in siblings as one group (if any):
742 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
743 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
744 partial_group = counter;
753 * Groups can be scheduled in as one unit only, so undo any
754 * partial group before returning:
756 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
757 if (counter == partial_group)
759 counter_sched_out(counter, cpuctx, ctx);
761 counter_sched_out(group_counter, cpuctx, ctx);
767 __perf_counter_sched_in(struct perf_counter_context *ctx,
768 struct perf_cpu_context *cpuctx, int cpu)
770 struct perf_counter *counter;
774 spin_lock(&ctx->lock);
776 if (likely(!ctx->nr_counters))
779 flags = hw_perf_save_disable();
782 * First go through the list and put on any pinned groups
783 * in order to give them the best chance of going on.
785 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
786 if (counter->state <= PERF_COUNTER_STATE_OFF ||
787 !counter->hw_event.pinned)
789 if (counter->cpu != -1 && counter->cpu != cpu)
792 if (group_can_go_on(counter, cpuctx, 1))
793 group_sched_in(counter, cpuctx, ctx, cpu);
796 * If this pinned group hasn't been scheduled,
797 * put it in error state.
799 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
800 counter->state = PERF_COUNTER_STATE_ERROR;
803 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
805 * Ignore counters in OFF or ERROR state, and
806 * ignore pinned counters since we did them already.
808 if (counter->state <= PERF_COUNTER_STATE_OFF ||
809 counter->hw_event.pinned)
813 * Listen to the 'cpu' scheduling filter constraint
816 if (counter->cpu != -1 && counter->cpu != cpu)
819 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
820 if (group_sched_in(counter, cpuctx, ctx, cpu))
824 hw_perf_restore(flags);
826 spin_unlock(&ctx->lock);
830 * Called from scheduler to add the counters of the current task
831 * with interrupts disabled.
833 * We restore the counter value and then enable it.
835 * This does not protect us against NMI, but enable()
836 * sets the enabled bit in the control field of counter _before_
837 * accessing the counter control register. If a NMI hits, then it will
838 * keep the counter running.
840 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
842 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
843 struct perf_counter_context *ctx = &task->perf_counter_ctx;
845 __perf_counter_sched_in(ctx, cpuctx, cpu);
846 cpuctx->task_ctx = ctx;
849 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
851 struct perf_counter_context *ctx = &cpuctx->ctx;
853 __perf_counter_sched_in(ctx, cpuctx, cpu);
856 int perf_counter_task_disable(void)
858 struct task_struct *curr = current;
859 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
860 struct perf_counter *counter;
865 if (likely(!ctx->nr_counters))
868 curr_rq_lock_irq_save(&flags);
869 cpu = smp_processor_id();
871 /* force the update of the task clock: */
872 __task_delta_exec(curr, 1);
874 perf_counter_task_sched_out(curr, cpu);
876 spin_lock(&ctx->lock);
879 * Disable all the counters:
881 perf_flags = hw_perf_save_disable();
883 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
884 if (counter->state != PERF_COUNTER_STATE_ERROR)
885 counter->state = PERF_COUNTER_STATE_OFF;
888 hw_perf_restore(perf_flags);
890 spin_unlock(&ctx->lock);
892 curr_rq_unlock_irq_restore(&flags);
897 int perf_counter_task_enable(void)
899 struct task_struct *curr = current;
900 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
901 struct perf_counter *counter;
906 if (likely(!ctx->nr_counters))
909 curr_rq_lock_irq_save(&flags);
910 cpu = smp_processor_id();
912 /* force the update of the task clock: */
913 __task_delta_exec(curr, 1);
915 perf_counter_task_sched_out(curr, cpu);
917 spin_lock(&ctx->lock);
920 * Disable all the counters:
922 perf_flags = hw_perf_save_disable();
924 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
925 if (counter->state > PERF_COUNTER_STATE_OFF)
927 counter->state = PERF_COUNTER_STATE_INACTIVE;
928 counter->hw_event.disabled = 0;
930 hw_perf_restore(perf_flags);
932 spin_unlock(&ctx->lock);
934 perf_counter_task_sched_in(curr, cpu);
936 curr_rq_unlock_irq_restore(&flags);
942 * Round-robin a context's counters:
944 static void rotate_ctx(struct perf_counter_context *ctx)
946 struct perf_counter *counter;
949 if (!ctx->nr_counters)
952 spin_lock(&ctx->lock);
954 * Rotate the first entry last (works just fine for group counters too):
956 perf_flags = hw_perf_save_disable();
957 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
958 list_del(&counter->list_entry);
959 list_add_tail(&counter->list_entry, &ctx->counter_list);
962 hw_perf_restore(perf_flags);
964 spin_unlock(&ctx->lock);
967 void perf_counter_task_tick(struct task_struct *curr, int cpu)
969 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
970 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
971 const int rotate_percpu = 0;
974 perf_counter_cpu_sched_out(cpuctx);
975 perf_counter_task_sched_out(curr, cpu);
978 rotate_ctx(&cpuctx->ctx);
982 perf_counter_cpu_sched_in(cpuctx, cpu);
983 perf_counter_task_sched_in(curr, cpu);
987 * Cross CPU call to read the hardware counter
989 static void __read(void *info)
991 struct perf_counter *counter = info;
994 curr_rq_lock_irq_save(&flags);
995 counter->hw_ops->read(counter);
996 curr_rq_unlock_irq_restore(&flags);
999 static u64 perf_counter_read(struct perf_counter *counter)
1002 * If counter is enabled and currently active on a CPU, update the
1003 * value in the counter structure:
1005 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1006 smp_call_function_single(counter->oncpu,
1007 __read, counter, 1);
1010 return atomic64_read(&counter->count);
1014 * Cross CPU call to switch performance data pointers
1016 static void __perf_switch_irq_data(void *info)
1018 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1019 struct perf_counter *counter = info;
1020 struct perf_counter_context *ctx = counter->ctx;
1021 struct perf_data *oldirqdata = counter->irqdata;
1024 * If this is a task context, we need to check whether it is
1025 * the current task context of this cpu. If not it has been
1026 * scheduled out before the smp call arrived.
1029 if (cpuctx->task_ctx != ctx)
1031 spin_lock(&ctx->lock);
1034 /* Change the pointer NMI safe */
1035 atomic_long_set((atomic_long_t *)&counter->irqdata,
1036 (unsigned long) counter->usrdata);
1037 counter->usrdata = oldirqdata;
1040 spin_unlock(&ctx->lock);
1043 static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
1045 struct perf_counter_context *ctx = counter->ctx;
1046 struct perf_data *oldirqdata = counter->irqdata;
1047 struct task_struct *task = ctx->task;
1050 smp_call_function_single(counter->cpu,
1051 __perf_switch_irq_data,
1053 return counter->usrdata;
1057 spin_lock_irq(&ctx->lock);
1058 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
1059 counter->irqdata = counter->usrdata;
1060 counter->usrdata = oldirqdata;
1061 spin_unlock_irq(&ctx->lock);
1064 spin_unlock_irq(&ctx->lock);
1065 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
1066 /* Might have failed, because task was scheduled out */
1067 if (counter->irqdata == oldirqdata)
1070 return counter->usrdata;
1073 static void put_context(struct perf_counter_context *ctx)
1076 put_task_struct(ctx->task);
1079 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1081 struct perf_cpu_context *cpuctx;
1082 struct perf_counter_context *ctx;
1083 struct task_struct *task;
1086 * If cpu is not a wildcard then this is a percpu counter:
1089 /* Must be root to operate on a CPU counter: */
1090 if (!capable(CAP_SYS_ADMIN))
1091 return ERR_PTR(-EACCES);
1093 if (cpu < 0 || cpu > num_possible_cpus())
1094 return ERR_PTR(-EINVAL);
1097 * We could be clever and allow to attach a counter to an
1098 * offline CPU and activate it when the CPU comes up, but
1101 if (!cpu_isset(cpu, cpu_online_map))
1102 return ERR_PTR(-ENODEV);
1104 cpuctx = &per_cpu(perf_cpu_context, cpu);
1114 task = find_task_by_vpid(pid);
1116 get_task_struct(task);
1120 return ERR_PTR(-ESRCH);
1122 ctx = &task->perf_counter_ctx;
1125 /* Reuse ptrace permission checks for now. */
1126 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1128 return ERR_PTR(-EACCES);
1135 * Called when the last reference to the file is gone.
1137 static int perf_release(struct inode *inode, struct file *file)
1139 struct perf_counter *counter = file->private_data;
1140 struct perf_counter_context *ctx = counter->ctx;
1142 file->private_data = NULL;
1144 mutex_lock(&ctx->mutex);
1145 mutex_lock(&counter->mutex);
1147 perf_counter_remove_from_context(counter);
1149 mutex_unlock(&counter->mutex);
1150 mutex_unlock(&ctx->mutex);
1159 * Read the performance counter - simple non blocking version for now
1162 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1166 if (count != sizeof(cntval))
1170 * Return end-of-file for a read on a counter that is in
1171 * error state (i.e. because it was pinned but it couldn't be
1172 * scheduled on to the CPU at some point).
1174 if (counter->state == PERF_COUNTER_STATE_ERROR)
1177 mutex_lock(&counter->mutex);
1178 cntval = perf_counter_read(counter);
1179 mutex_unlock(&counter->mutex);
1181 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
1185 perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
1190 count = min(count, (size_t)usrdata->len);
1191 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
1194 /* Adjust the counters */
1195 usrdata->len -= count;
1197 usrdata->rd_idx = 0;
1199 usrdata->rd_idx += count;
1205 perf_read_irq_data(struct perf_counter *counter,
1210 struct perf_data *irqdata, *usrdata;
1211 DECLARE_WAITQUEUE(wait, current);
1214 irqdata = counter->irqdata;
1215 usrdata = counter->usrdata;
1217 if (usrdata->len + irqdata->len >= count)
1223 spin_lock_irq(&counter->waitq.lock);
1224 __add_wait_queue(&counter->waitq, &wait);
1226 set_current_state(TASK_INTERRUPTIBLE);
1227 if (usrdata->len + irqdata->len >= count)
1230 if (signal_pending(current))
1233 if (counter->state == PERF_COUNTER_STATE_ERROR)
1236 spin_unlock_irq(&counter->waitq.lock);
1238 spin_lock_irq(&counter->waitq.lock);
1240 __remove_wait_queue(&counter->waitq, &wait);
1241 __set_current_state(TASK_RUNNING);
1242 spin_unlock_irq(&counter->waitq.lock);
1244 if (usrdata->len + irqdata->len < count &&
1245 counter->state != PERF_COUNTER_STATE_ERROR)
1246 return -ERESTARTSYS;
1248 mutex_lock(&counter->mutex);
1250 /* Drain pending data first: */
1251 res = perf_copy_usrdata(usrdata, buf, count);
1252 if (res < 0 || res == count)
1255 /* Switch irq buffer: */
1256 usrdata = perf_switch_irq_data(counter);
1257 res2 = perf_copy_usrdata(usrdata, buf + res, count - res);
1265 mutex_unlock(&counter->mutex);
1271 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1273 struct perf_counter *counter = file->private_data;
1275 switch (counter->hw_event.record_type) {
1276 case PERF_RECORD_SIMPLE:
1277 return perf_read_hw(counter, buf, count);
1279 case PERF_RECORD_IRQ:
1280 case PERF_RECORD_GROUP:
1281 return perf_read_irq_data(counter, buf, count,
1282 file->f_flags & O_NONBLOCK);
1287 static unsigned int perf_poll(struct file *file, poll_table *wait)
1289 struct perf_counter *counter = file->private_data;
1290 unsigned int events = 0;
1291 unsigned long flags;
1293 poll_wait(file, &counter->waitq, wait);
1295 spin_lock_irqsave(&counter->waitq.lock, flags);
1296 if (counter->usrdata->len || counter->irqdata->len)
1298 spin_unlock_irqrestore(&counter->waitq.lock, flags);
1303 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1305 struct perf_counter *counter = file->private_data;
1309 case PERF_COUNTER_IOC_ENABLE:
1310 perf_counter_enable_family(counter);
1312 case PERF_COUNTER_IOC_DISABLE:
1313 perf_counter_disable_family(counter);
1321 static const struct file_operations perf_fops = {
1322 .release = perf_release,
1325 .unlocked_ioctl = perf_ioctl,
1326 .compat_ioctl = perf_ioctl,
1329 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
1331 int cpu = raw_smp_processor_id();
1333 atomic64_set(&counter->hw.prev_count, cpu_clock(cpu));
1337 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
1339 int cpu = raw_smp_processor_id();
1343 now = cpu_clock(cpu);
1344 prev = atomic64_read(&counter->hw.prev_count);
1345 atomic64_set(&counter->hw.prev_count, now);
1346 atomic64_add(now - prev, &counter->count);
1349 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
1351 cpu_clock_perf_counter_update(counter);
1354 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
1356 cpu_clock_perf_counter_update(counter);
1359 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
1360 .enable = cpu_clock_perf_counter_enable,
1361 .disable = cpu_clock_perf_counter_disable,
1362 .read = cpu_clock_perf_counter_read,
1366 * Called from within the scheduler:
1368 static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
1370 struct task_struct *curr = counter->task;
1373 delta = __task_delta_exec(curr, update);
1375 return curr->se.sum_exec_runtime + delta;
1378 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
1383 prev = atomic64_read(&counter->hw.prev_count);
1385 atomic64_set(&counter->hw.prev_count, now);
1389 atomic64_add(delta, &counter->count);
1392 static void task_clock_perf_counter_read(struct perf_counter *counter)
1394 u64 now = task_clock_perf_counter_val(counter, 1);
1396 task_clock_perf_counter_update(counter, now);
1399 static int task_clock_perf_counter_enable(struct perf_counter *counter)
1401 u64 now = task_clock_perf_counter_val(counter, 0);
1403 atomic64_set(&counter->hw.prev_count, now);
1408 static void task_clock_perf_counter_disable(struct perf_counter *counter)
1410 u64 now = task_clock_perf_counter_val(counter, 0);
1412 task_clock_perf_counter_update(counter, now);
1415 static const struct hw_perf_counter_ops perf_ops_task_clock = {
1416 .enable = task_clock_perf_counter_enable,
1417 .disable = task_clock_perf_counter_disable,
1418 .read = task_clock_perf_counter_read,
1421 #ifdef CONFIG_VM_EVENT_COUNTERS
1422 #define cpu_page_faults() __get_cpu_var(vm_event_states).event[PGFAULT]
1424 #define cpu_page_faults() 0
1427 static u64 get_page_faults(struct perf_counter *counter)
1429 struct task_struct *curr = counter->ctx->task;
1432 return curr->maj_flt + curr->min_flt;
1433 return cpu_page_faults();
1436 static void page_faults_perf_counter_update(struct perf_counter *counter)
1441 prev = atomic64_read(&counter->hw.prev_count);
1442 now = get_page_faults(counter);
1444 atomic64_set(&counter->hw.prev_count, now);
1448 atomic64_add(delta, &counter->count);
1451 static void page_faults_perf_counter_read(struct perf_counter *counter)
1453 page_faults_perf_counter_update(counter);
1456 static int page_faults_perf_counter_enable(struct perf_counter *counter)
1458 atomic64_set(&counter->hw.prev_count, get_page_faults(counter));
1462 static void page_faults_perf_counter_disable(struct perf_counter *counter)
1464 page_faults_perf_counter_update(counter);
1467 static const struct hw_perf_counter_ops perf_ops_page_faults = {
1468 .enable = page_faults_perf_counter_enable,
1469 .disable = page_faults_perf_counter_disable,
1470 .read = page_faults_perf_counter_read,
1473 static u64 get_context_switches(struct perf_counter *counter)
1475 struct task_struct *curr = counter->ctx->task;
1478 return curr->nvcsw + curr->nivcsw;
1479 return cpu_nr_switches(smp_processor_id());
1482 static void context_switches_perf_counter_update(struct perf_counter *counter)
1487 prev = atomic64_read(&counter->hw.prev_count);
1488 now = get_context_switches(counter);
1490 atomic64_set(&counter->hw.prev_count, now);
1494 atomic64_add(delta, &counter->count);
1497 static void context_switches_perf_counter_read(struct perf_counter *counter)
1499 context_switches_perf_counter_update(counter);
1502 static int context_switches_perf_counter_enable(struct perf_counter *counter)
1504 atomic64_set(&counter->hw.prev_count, get_context_switches(counter));
1508 static void context_switches_perf_counter_disable(struct perf_counter *counter)
1510 context_switches_perf_counter_update(counter);
1513 static const struct hw_perf_counter_ops perf_ops_context_switches = {
1514 .enable = context_switches_perf_counter_enable,
1515 .disable = context_switches_perf_counter_disable,
1516 .read = context_switches_perf_counter_read,
1519 static inline u64 get_cpu_migrations(struct perf_counter *counter)
1521 struct task_struct *curr = counter->ctx->task;
1524 return curr->se.nr_migrations;
1525 return cpu_nr_migrations(smp_processor_id());
1528 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1533 prev = atomic64_read(&counter->hw.prev_count);
1534 now = get_cpu_migrations(counter);
1536 atomic64_set(&counter->hw.prev_count, now);
1540 atomic64_add(delta, &counter->count);
1543 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1545 cpu_migrations_perf_counter_update(counter);
1548 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1550 atomic64_set(&counter->hw.prev_count, get_cpu_migrations(counter));
1554 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1556 cpu_migrations_perf_counter_update(counter);
1559 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1560 .enable = cpu_migrations_perf_counter_enable,
1561 .disable = cpu_migrations_perf_counter_disable,
1562 .read = cpu_migrations_perf_counter_read,
1565 static const struct hw_perf_counter_ops *
1566 sw_perf_counter_init(struct perf_counter *counter)
1568 const struct hw_perf_counter_ops *hw_ops = NULL;
1571 * Software counters (currently) can't in general distinguish
1572 * between user, kernel and hypervisor events.
1573 * However, context switches and cpu migrations are considered
1574 * to be kernel events, and page faults are never hypervisor
1577 switch (counter->hw_event.type) {
1578 case PERF_COUNT_CPU_CLOCK:
1579 if (!(counter->hw_event.exclude_user ||
1580 counter->hw_event.exclude_kernel ||
1581 counter->hw_event.exclude_hv))
1582 hw_ops = &perf_ops_cpu_clock;
1584 case PERF_COUNT_TASK_CLOCK:
1585 if (counter->hw_event.exclude_user ||
1586 counter->hw_event.exclude_kernel ||
1587 counter->hw_event.exclude_hv)
1590 * If the user instantiates this as a per-cpu counter,
1591 * use the cpu_clock counter instead.
1593 if (counter->ctx->task)
1594 hw_ops = &perf_ops_task_clock;
1596 hw_ops = &perf_ops_cpu_clock;
1598 case PERF_COUNT_PAGE_FAULTS:
1599 if (!(counter->hw_event.exclude_user ||
1600 counter->hw_event.exclude_kernel))
1601 hw_ops = &perf_ops_page_faults;
1603 case PERF_COUNT_CONTEXT_SWITCHES:
1604 if (!counter->hw_event.exclude_kernel)
1605 hw_ops = &perf_ops_context_switches;
1607 case PERF_COUNT_CPU_MIGRATIONS:
1608 if (!counter->hw_event.exclude_kernel)
1609 hw_ops = &perf_ops_cpu_migrations;
1618 * Allocate and initialize a counter structure
1620 static struct perf_counter *
1621 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1623 struct perf_counter_context *ctx,
1624 struct perf_counter *group_leader,
1627 const struct hw_perf_counter_ops *hw_ops;
1628 struct perf_counter *counter;
1630 counter = kzalloc(sizeof(*counter), gfpflags);
1635 * Single counters are their own group leaders, with an
1636 * empty sibling list:
1639 group_leader = counter;
1641 mutex_init(&counter->mutex);
1642 INIT_LIST_HEAD(&counter->list_entry);
1643 INIT_LIST_HEAD(&counter->sibling_list);
1644 init_waitqueue_head(&counter->waitq);
1646 INIT_LIST_HEAD(&counter->child_list);
1648 counter->irqdata = &counter->data[0];
1649 counter->usrdata = &counter->data[1];
1651 counter->hw_event = *hw_event;
1652 counter->wakeup_pending = 0;
1653 counter->group_leader = group_leader;
1654 counter->hw_ops = NULL;
1657 counter->state = PERF_COUNTER_STATE_INACTIVE;
1658 if (hw_event->disabled)
1659 counter->state = PERF_COUNTER_STATE_OFF;
1662 if (!hw_event->raw && hw_event->type < 0)
1663 hw_ops = sw_perf_counter_init(counter);
1665 hw_ops = hw_perf_counter_init(counter);
1671 counter->hw_ops = hw_ops;
1677 * sys_perf_task_open - open a performance counter, associate it to a task/cpu
1679 * @hw_event_uptr: event type attributes for monitoring/sampling
1682 * @group_fd: group leader counter fd
1685 sys_perf_counter_open(struct perf_counter_hw_event *hw_event_uptr __user,
1686 pid_t pid, int cpu, int group_fd)
1688 struct perf_counter *counter, *group_leader;
1689 struct perf_counter_hw_event hw_event;
1690 struct perf_counter_context *ctx;
1691 struct file *counter_file = NULL;
1692 struct file *group_file = NULL;
1693 int fput_needed = 0;
1694 int fput_needed2 = 0;
1697 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1701 * Get the target context (task or percpu):
1703 ctx = find_get_context(pid, cpu);
1705 return PTR_ERR(ctx);
1708 * Look up the group leader (we will attach this counter to it):
1710 group_leader = NULL;
1711 if (group_fd != -1) {
1713 group_file = fget_light(group_fd, &fput_needed);
1715 goto err_put_context;
1716 if (group_file->f_op != &perf_fops)
1717 goto err_put_context;
1719 group_leader = group_file->private_data;
1721 * Do not allow a recursive hierarchy (this new sibling
1722 * becoming part of another group-sibling):
1724 if (group_leader->group_leader != group_leader)
1725 goto err_put_context;
1727 * Do not allow to attach to a group in a different
1728 * task or CPU context:
1730 if (group_leader->ctx != ctx)
1731 goto err_put_context;
1733 * Only a group leader can be exclusive or pinned
1735 if (hw_event.exclusive || hw_event.pinned)
1736 goto err_put_context;
1740 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
1743 goto err_put_context;
1745 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
1747 goto err_free_put_context;
1749 counter_file = fget_light(ret, &fput_needed2);
1751 goto err_free_put_context;
1753 counter->filp = counter_file;
1754 mutex_lock(&ctx->mutex);
1755 perf_install_in_context(ctx, counter, cpu);
1756 mutex_unlock(&ctx->mutex);
1758 fput_light(counter_file, fput_needed2);
1761 fput_light(group_file, fput_needed);
1765 err_free_put_context:
1775 * Initialize the perf_counter context in a task_struct:
1778 __perf_counter_init_context(struct perf_counter_context *ctx,
1779 struct task_struct *task)
1781 memset(ctx, 0, sizeof(*ctx));
1782 spin_lock_init(&ctx->lock);
1783 mutex_init(&ctx->mutex);
1784 INIT_LIST_HEAD(&ctx->counter_list);
1789 * inherit a counter from parent task to child task:
1791 static struct perf_counter *
1792 inherit_counter(struct perf_counter *parent_counter,
1793 struct task_struct *parent,
1794 struct perf_counter_context *parent_ctx,
1795 struct task_struct *child,
1796 struct perf_counter *group_leader,
1797 struct perf_counter_context *child_ctx)
1799 struct perf_counter *child_counter;
1802 * Instead of creating recursive hierarchies of counters,
1803 * we link inherited counters back to the original parent,
1804 * which has a filp for sure, which we use as the reference
1807 if (parent_counter->parent)
1808 parent_counter = parent_counter->parent;
1810 child_counter = perf_counter_alloc(&parent_counter->hw_event,
1811 parent_counter->cpu, child_ctx,
1812 group_leader, GFP_KERNEL);
1817 * Link it up in the child's context:
1819 child_counter->task = child;
1820 list_add_counter(child_counter, child_ctx);
1821 child_ctx->nr_counters++;
1823 child_counter->parent = parent_counter;
1825 * inherit into child's child as well:
1827 child_counter->hw_event.inherit = 1;
1830 * Get a reference to the parent filp - we will fput it
1831 * when the child counter exits. This is safe to do because
1832 * we are in the parent and we know that the filp still
1833 * exists and has a nonzero count:
1835 atomic_long_inc(&parent_counter->filp->f_count);
1838 * Link this into the parent counter's child list
1840 mutex_lock(&parent_counter->mutex);
1841 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
1844 * Make the child state follow the state of the parent counter,
1845 * not its hw_event.disabled bit. We hold the parent's mutex,
1846 * so we won't race with perf_counter_{en,dis}able_family.
1848 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
1849 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
1851 child_counter->state = PERF_COUNTER_STATE_OFF;
1853 mutex_unlock(&parent_counter->mutex);
1855 return child_counter;
1858 static int inherit_group(struct perf_counter *parent_counter,
1859 struct task_struct *parent,
1860 struct perf_counter_context *parent_ctx,
1861 struct task_struct *child,
1862 struct perf_counter_context *child_ctx)
1864 struct perf_counter *leader;
1865 struct perf_counter *sub;
1867 leader = inherit_counter(parent_counter, parent, parent_ctx,
1868 child, NULL, child_ctx);
1871 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
1872 if (!inherit_counter(sub, parent, parent_ctx,
1873 child, leader, child_ctx))
1879 static void sync_child_counter(struct perf_counter *child_counter,
1880 struct perf_counter *parent_counter)
1882 u64 parent_val, child_val;
1884 parent_val = atomic64_read(&parent_counter->count);
1885 child_val = atomic64_read(&child_counter->count);
1888 * Add back the child's count to the parent's count:
1890 atomic64_add(child_val, &parent_counter->count);
1893 * Remove this counter from the parent's list
1895 mutex_lock(&parent_counter->mutex);
1896 list_del_init(&child_counter->child_list);
1897 mutex_unlock(&parent_counter->mutex);
1900 * Release the parent counter, if this was the last
1903 fput(parent_counter->filp);
1907 __perf_counter_exit_task(struct task_struct *child,
1908 struct perf_counter *child_counter,
1909 struct perf_counter_context *child_ctx)
1911 struct perf_counter *parent_counter;
1912 struct perf_counter *sub, *tmp;
1915 * If we do not self-reap then we have to wait for the
1916 * child task to unschedule (it will happen for sure),
1917 * so that its counter is at its final count. (This
1918 * condition triggers rarely - child tasks usually get
1919 * off their CPU before the parent has a chance to
1920 * get this far into the reaping action)
1922 if (child != current) {
1923 wait_task_inactive(child, 0);
1924 list_del_init(&child_counter->list_entry);
1926 struct perf_cpu_context *cpuctx;
1927 unsigned long flags;
1931 * Disable and unlink this counter.
1933 * Be careful about zapping the list - IRQ/NMI context
1934 * could still be processing it:
1936 curr_rq_lock_irq_save(&flags);
1937 perf_flags = hw_perf_save_disable();
1939 cpuctx = &__get_cpu_var(perf_cpu_context);
1941 group_sched_out(child_counter, cpuctx, child_ctx);
1943 list_del_init(&child_counter->list_entry);
1945 child_ctx->nr_counters--;
1947 hw_perf_restore(perf_flags);
1948 curr_rq_unlock_irq_restore(&flags);
1951 parent_counter = child_counter->parent;
1953 * It can happen that parent exits first, and has counters
1954 * that are still around due to the child reference. These
1955 * counters need to be zapped - but otherwise linger.
1957 if (parent_counter) {
1958 sync_child_counter(child_counter, parent_counter);
1959 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
1962 sync_child_counter(sub, sub->parent);
1967 if (!child_counter->filp || !atomic_long_read(&child_counter->filp->f_count))
1968 kfree(child_counter);
1972 * When a child task exits, feed back counter values to parent counters.
1974 * Note: we may be running in child context, but the PID is not hashed
1975 * anymore so new counters will not be added.
1977 void perf_counter_exit_task(struct task_struct *child)
1979 struct perf_counter *child_counter, *tmp;
1980 struct perf_counter_context *child_ctx;
1982 child_ctx = &child->perf_counter_ctx;
1984 if (likely(!child_ctx->nr_counters))
1987 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
1989 __perf_counter_exit_task(child, child_counter, child_ctx);
1993 * Initialize the perf_counter context in task_struct
1995 void perf_counter_init_task(struct task_struct *child)
1997 struct perf_counter_context *child_ctx, *parent_ctx;
1998 struct perf_counter *counter;
1999 struct task_struct *parent = current;
2001 child_ctx = &child->perf_counter_ctx;
2002 parent_ctx = &parent->perf_counter_ctx;
2004 __perf_counter_init_context(child_ctx, child);
2007 * This is executed from the parent task context, so inherit
2008 * counters that have been marked for cloning:
2011 if (likely(!parent_ctx->nr_counters))
2015 * Lock the parent list. No need to lock the child - not PID
2016 * hashed yet and not running, so nobody can access it.
2018 mutex_lock(&parent_ctx->mutex);
2021 * We dont have to disable NMIs - we are only looking at
2022 * the list, not manipulating it:
2024 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2025 if (!counter->hw_event.inherit)
2028 if (inherit_group(counter, parent,
2029 parent_ctx, child, child_ctx))
2033 mutex_unlock(&parent_ctx->mutex);
2036 static void __cpuinit perf_counter_init_cpu(int cpu)
2038 struct perf_cpu_context *cpuctx;
2040 cpuctx = &per_cpu(perf_cpu_context, cpu);
2041 __perf_counter_init_context(&cpuctx->ctx, NULL);
2043 mutex_lock(&perf_resource_mutex);
2044 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2045 mutex_unlock(&perf_resource_mutex);
2047 hw_perf_counter_setup(cpu);
2050 #ifdef CONFIG_HOTPLUG_CPU
2051 static void __perf_counter_exit_cpu(void *info)
2053 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2054 struct perf_counter_context *ctx = &cpuctx->ctx;
2055 struct perf_counter *counter, *tmp;
2057 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2058 __perf_counter_remove_from_context(counter);
2060 static void perf_counter_exit_cpu(int cpu)
2062 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2063 struct perf_counter_context *ctx = &cpuctx->ctx;
2065 mutex_lock(&ctx->mutex);
2066 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2067 mutex_unlock(&ctx->mutex);
2070 static inline void perf_counter_exit_cpu(int cpu) { }
2073 static int __cpuinit
2074 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2076 unsigned int cpu = (long)hcpu;
2080 case CPU_UP_PREPARE:
2081 case CPU_UP_PREPARE_FROZEN:
2082 perf_counter_init_cpu(cpu);
2085 case CPU_DOWN_PREPARE:
2086 case CPU_DOWN_PREPARE_FROZEN:
2087 perf_counter_exit_cpu(cpu);
2097 static struct notifier_block __cpuinitdata perf_cpu_nb = {
2098 .notifier_call = perf_cpu_notify,
2101 static int __init perf_counter_init(void)
2103 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2104 (void *)(long)smp_processor_id());
2105 register_cpu_notifier(&perf_cpu_nb);
2109 early_initcall(perf_counter_init);
2111 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2113 return sprintf(buf, "%d\n", perf_reserved_percpu);
2117 perf_set_reserve_percpu(struct sysdev_class *class,
2121 struct perf_cpu_context *cpuctx;
2125 err = strict_strtoul(buf, 10, &val);
2128 if (val > perf_max_counters)
2131 mutex_lock(&perf_resource_mutex);
2132 perf_reserved_percpu = val;
2133 for_each_online_cpu(cpu) {
2134 cpuctx = &per_cpu(perf_cpu_context, cpu);
2135 spin_lock_irq(&cpuctx->ctx.lock);
2136 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2137 perf_max_counters - perf_reserved_percpu);
2138 cpuctx->max_pertask = mpt;
2139 spin_unlock_irq(&cpuctx->ctx.lock);
2141 mutex_unlock(&perf_resource_mutex);
2146 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2148 return sprintf(buf, "%d\n", perf_overcommit);
2152 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2157 err = strict_strtoul(buf, 10, &val);
2163 mutex_lock(&perf_resource_mutex);
2164 perf_overcommit = val;
2165 mutex_unlock(&perf_resource_mutex);
2170 static SYSDEV_CLASS_ATTR(
2173 perf_show_reserve_percpu,
2174 perf_set_reserve_percpu
2177 static SYSDEV_CLASS_ATTR(
2180 perf_show_overcommit,
2184 static struct attribute *perfclass_attrs[] = {
2185 &attr_reserve_percpu.attr,
2186 &attr_overcommit.attr,
2190 static struct attribute_group perfclass_attr_group = {
2191 .attrs = perfclass_attrs,
2192 .name = "perf_counters",
2195 static int __init perf_counter_sysfs_init(void)
2197 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2198 &perfclass_attr_group);
2200 device_initcall(perf_counter_sysfs_init);