2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_counters __read_mostly;
44 static atomic_t nr_comm_counters __read_mostly;
45 static atomic_t nr_task_counters __read_mostly;
48 * perf counter paranoia level:
50 * 1 - disallow cpu counters to unpriv
51 * 2 - disallow kernel profiling to unpriv
53 int sysctl_perf_counter_paranoid __read_mostly = 1;
55 static inline bool perf_paranoid_cpu(void)
57 return sysctl_perf_counter_paranoid > 0;
60 static inline bool perf_paranoid_kernel(void)
62 return sysctl_perf_counter_paranoid > 1;
65 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
68 * max perf counter sample rate
70 int sysctl_perf_counter_sample_rate __read_mostly = 100000;
72 static atomic64_t perf_counter_id;
75 * Lock for (sysadmin-configurable) counter reservations:
77 static DEFINE_SPINLOCK(perf_resource_lock);
80 * Architecture provided APIs - weak aliases:
82 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
87 void __weak hw_perf_disable(void) { barrier(); }
88 void __weak hw_perf_enable(void) { barrier(); }
90 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
91 void __weak hw_perf_counter_setup_online(int cpu) { barrier(); }
94 hw_perf_group_sched_in(struct perf_counter *group_leader,
95 struct perf_cpu_context *cpuctx,
96 struct perf_counter_context *ctx, int cpu)
101 void __weak perf_counter_print_debug(void) { }
103 static DEFINE_PER_CPU(int, disable_count);
105 void __perf_disable(void)
107 __get_cpu_var(disable_count)++;
110 bool __perf_enable(void)
112 return !--__get_cpu_var(disable_count);
115 void perf_disable(void)
121 void perf_enable(void)
127 static void get_ctx(struct perf_counter_context *ctx)
129 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
132 static void free_ctx(struct rcu_head *head)
134 struct perf_counter_context *ctx;
136 ctx = container_of(head, struct perf_counter_context, rcu_head);
140 static void put_ctx(struct perf_counter_context *ctx)
142 if (atomic_dec_and_test(&ctx->refcount)) {
144 put_ctx(ctx->parent_ctx);
146 put_task_struct(ctx->task);
147 call_rcu(&ctx->rcu_head, free_ctx);
151 static void unclone_ctx(struct perf_counter_context *ctx)
153 if (ctx->parent_ctx) {
154 put_ctx(ctx->parent_ctx);
155 ctx->parent_ctx = NULL;
160 * If we inherit counters we want to return the parent counter id
163 static u64 primary_counter_id(struct perf_counter *counter)
165 u64 id = counter->id;
168 id = counter->parent->id;
174 * Get the perf_counter_context for a task and lock it.
175 * This has to cope with with the fact that until it is locked,
176 * the context could get moved to another task.
178 static struct perf_counter_context *
179 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
181 struct perf_counter_context *ctx;
185 ctx = rcu_dereference(task->perf_counter_ctxp);
188 * If this context is a clone of another, it might
189 * get swapped for another underneath us by
190 * perf_counter_task_sched_out, though the
191 * rcu_read_lock() protects us from any context
192 * getting freed. Lock the context and check if it
193 * got swapped before we could get the lock, and retry
194 * if so. If we locked the right context, then it
195 * can't get swapped on us any more.
197 spin_lock_irqsave(&ctx->lock, *flags);
198 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
199 spin_unlock_irqrestore(&ctx->lock, *flags);
203 if (!atomic_inc_not_zero(&ctx->refcount)) {
204 spin_unlock_irqrestore(&ctx->lock, *flags);
213 * Get the context for a task and increment its pin_count so it
214 * can't get swapped to another task. This also increments its
215 * reference count so that the context can't get freed.
217 static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
219 struct perf_counter_context *ctx;
222 ctx = perf_lock_task_context(task, &flags);
225 spin_unlock_irqrestore(&ctx->lock, flags);
230 static void perf_unpin_context(struct perf_counter_context *ctx)
234 spin_lock_irqsave(&ctx->lock, flags);
236 spin_unlock_irqrestore(&ctx->lock, flags);
241 * Add a counter from the lists for its context.
242 * Must be called with ctx->mutex and ctx->lock held.
245 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
247 struct perf_counter *group_leader = counter->group_leader;
250 * Depending on whether it is a standalone or sibling counter,
251 * add it straight to the context's counter list, or to the group
252 * leader's sibling list:
254 if (group_leader == counter)
255 list_add_tail(&counter->list_entry, &ctx->counter_list);
257 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
258 group_leader->nr_siblings++;
261 list_add_rcu(&counter->event_entry, &ctx->event_list);
263 if (counter->attr.inherit_stat)
268 * Remove a counter from the lists for its context.
269 * Must be called with ctx->mutex and ctx->lock held.
272 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
274 struct perf_counter *sibling, *tmp;
276 if (list_empty(&counter->list_entry))
279 if (counter->attr.inherit_stat)
282 list_del_init(&counter->list_entry);
283 list_del_rcu(&counter->event_entry);
285 if (counter->group_leader != counter)
286 counter->group_leader->nr_siblings--;
289 * If this was a group counter with sibling counters then
290 * upgrade the siblings to singleton counters by adding them
291 * to the context list directly:
293 list_for_each_entry_safe(sibling, tmp,
294 &counter->sibling_list, list_entry) {
296 list_move_tail(&sibling->list_entry, &ctx->counter_list);
297 sibling->group_leader = sibling;
302 counter_sched_out(struct perf_counter *counter,
303 struct perf_cpu_context *cpuctx,
304 struct perf_counter_context *ctx)
306 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
309 counter->state = PERF_COUNTER_STATE_INACTIVE;
310 if (counter->pending_disable) {
311 counter->pending_disable = 0;
312 counter->state = PERF_COUNTER_STATE_OFF;
314 counter->tstamp_stopped = ctx->time;
315 counter->pmu->disable(counter);
318 if (!is_software_counter(counter))
319 cpuctx->active_oncpu--;
321 if (counter->attr.exclusive || !cpuctx->active_oncpu)
322 cpuctx->exclusive = 0;
326 group_sched_out(struct perf_counter *group_counter,
327 struct perf_cpu_context *cpuctx,
328 struct perf_counter_context *ctx)
330 struct perf_counter *counter;
332 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
335 counter_sched_out(group_counter, cpuctx, ctx);
338 * Schedule out siblings (if any):
340 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
341 counter_sched_out(counter, cpuctx, ctx);
343 if (group_counter->attr.exclusive)
344 cpuctx->exclusive = 0;
348 * Cross CPU call to remove a performance counter
350 * We disable the counter on the hardware level first. After that we
351 * remove it from the context list.
353 static void __perf_counter_remove_from_context(void *info)
355 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
356 struct perf_counter *counter = info;
357 struct perf_counter_context *ctx = counter->ctx;
360 * If this is a task context, we need to check whether it is
361 * the current task context of this cpu. If not it has been
362 * scheduled out before the smp call arrived.
364 if (ctx->task && cpuctx->task_ctx != ctx)
367 spin_lock(&ctx->lock);
369 * Protect the list operation against NMI by disabling the
370 * counters on a global level.
374 counter_sched_out(counter, cpuctx, ctx);
376 list_del_counter(counter, ctx);
380 * Allow more per task counters with respect to the
383 cpuctx->max_pertask =
384 min(perf_max_counters - ctx->nr_counters,
385 perf_max_counters - perf_reserved_percpu);
389 spin_unlock(&ctx->lock);
394 * Remove the counter from a task's (or a CPU's) list of counters.
396 * Must be called with ctx->mutex held.
398 * CPU counters are removed with a smp call. For task counters we only
399 * call when the task is on a CPU.
401 * If counter->ctx is a cloned context, callers must make sure that
402 * every task struct that counter->ctx->task could possibly point to
403 * remains valid. This is OK when called from perf_release since
404 * that only calls us on the top-level context, which can't be a clone.
405 * When called from perf_counter_exit_task, it's OK because the
406 * context has been detached from its task.
408 static void perf_counter_remove_from_context(struct perf_counter *counter)
410 struct perf_counter_context *ctx = counter->ctx;
411 struct task_struct *task = ctx->task;
415 * Per cpu counters are removed via an smp call and
416 * the removal is always sucessful.
418 smp_call_function_single(counter->cpu,
419 __perf_counter_remove_from_context,
425 task_oncpu_function_call(task, __perf_counter_remove_from_context,
428 spin_lock_irq(&ctx->lock);
430 * If the context is active we need to retry the smp call.
432 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
433 spin_unlock_irq(&ctx->lock);
438 * The lock prevents that this context is scheduled in so we
439 * can remove the counter safely, if the call above did not
442 if (!list_empty(&counter->list_entry)) {
443 list_del_counter(counter, ctx);
445 spin_unlock_irq(&ctx->lock);
448 static inline u64 perf_clock(void)
450 return cpu_clock(smp_processor_id());
454 * Update the record of the current time in a context.
456 static void update_context_time(struct perf_counter_context *ctx)
458 u64 now = perf_clock();
460 ctx->time += now - ctx->timestamp;
461 ctx->timestamp = now;
465 * Update the total_time_enabled and total_time_running fields for a counter.
467 static void update_counter_times(struct perf_counter *counter)
469 struct perf_counter_context *ctx = counter->ctx;
472 if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
473 counter->group_leader->state < PERF_COUNTER_STATE_INACTIVE)
476 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
478 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
479 run_end = counter->tstamp_stopped;
483 counter->total_time_running = run_end - counter->tstamp_running;
487 * Update total_time_enabled and total_time_running for all counters in a group.
489 static void update_group_times(struct perf_counter *leader)
491 struct perf_counter *counter;
493 update_counter_times(leader);
494 list_for_each_entry(counter, &leader->sibling_list, list_entry)
495 update_counter_times(counter);
499 * Cross CPU call to disable a performance counter
501 static void __perf_counter_disable(void *info)
503 struct perf_counter *counter = info;
504 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
505 struct perf_counter_context *ctx = counter->ctx;
508 * If this is a per-task counter, need to check whether this
509 * counter's task is the current task on this cpu.
511 if (ctx->task && cpuctx->task_ctx != ctx)
514 spin_lock(&ctx->lock);
517 * If the counter is on, turn it off.
518 * If it is in error state, leave it in error state.
520 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
521 update_context_time(ctx);
522 update_group_times(counter);
523 if (counter == counter->group_leader)
524 group_sched_out(counter, cpuctx, ctx);
526 counter_sched_out(counter, cpuctx, ctx);
527 counter->state = PERF_COUNTER_STATE_OFF;
530 spin_unlock(&ctx->lock);
536 * If counter->ctx is a cloned context, callers must make sure that
537 * every task struct that counter->ctx->task could possibly point to
538 * remains valid. This condition is satisifed when called through
539 * perf_counter_for_each_child or perf_counter_for_each because they
540 * hold the top-level counter's child_mutex, so any descendant that
541 * goes to exit will block in sync_child_counter.
542 * When called from perf_pending_counter it's OK because counter->ctx
543 * is the current context on this CPU and preemption is disabled,
544 * hence we can't get into perf_counter_task_sched_out for this context.
546 static void perf_counter_disable(struct perf_counter *counter)
548 struct perf_counter_context *ctx = counter->ctx;
549 struct task_struct *task = ctx->task;
553 * Disable the counter on the cpu that it's on
555 smp_call_function_single(counter->cpu, __perf_counter_disable,
561 task_oncpu_function_call(task, __perf_counter_disable, counter);
563 spin_lock_irq(&ctx->lock);
565 * If the counter is still active, we need to retry the cross-call.
567 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
568 spin_unlock_irq(&ctx->lock);
573 * Since we have the lock this context can't be scheduled
574 * in, so we can change the state safely.
576 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
577 update_group_times(counter);
578 counter->state = PERF_COUNTER_STATE_OFF;
581 spin_unlock_irq(&ctx->lock);
585 counter_sched_in(struct perf_counter *counter,
586 struct perf_cpu_context *cpuctx,
587 struct perf_counter_context *ctx,
590 if (counter->state <= PERF_COUNTER_STATE_OFF)
593 counter->state = PERF_COUNTER_STATE_ACTIVE;
594 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
596 * The new state must be visible before we turn it on in the hardware:
600 if (counter->pmu->enable(counter)) {
601 counter->state = PERF_COUNTER_STATE_INACTIVE;
606 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
608 if (!is_software_counter(counter))
609 cpuctx->active_oncpu++;
612 if (counter->attr.exclusive)
613 cpuctx->exclusive = 1;
619 group_sched_in(struct perf_counter *group_counter,
620 struct perf_cpu_context *cpuctx,
621 struct perf_counter_context *ctx,
624 struct perf_counter *counter, *partial_group;
627 if (group_counter->state == PERF_COUNTER_STATE_OFF)
630 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
632 return ret < 0 ? ret : 0;
634 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
638 * Schedule in siblings as one group (if any):
640 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
641 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
642 partial_group = counter;
651 * Groups can be scheduled in as one unit only, so undo any
652 * partial group before returning:
654 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
655 if (counter == partial_group)
657 counter_sched_out(counter, cpuctx, ctx);
659 counter_sched_out(group_counter, cpuctx, ctx);
665 * Return 1 for a group consisting entirely of software counters,
666 * 0 if the group contains any hardware counters.
668 static int is_software_only_group(struct perf_counter *leader)
670 struct perf_counter *counter;
672 if (!is_software_counter(leader))
675 list_for_each_entry(counter, &leader->sibling_list, list_entry)
676 if (!is_software_counter(counter))
683 * Work out whether we can put this counter group on the CPU now.
685 static int group_can_go_on(struct perf_counter *counter,
686 struct perf_cpu_context *cpuctx,
690 * Groups consisting entirely of software counters can always go on.
692 if (is_software_only_group(counter))
695 * If an exclusive group is already on, no other hardware
696 * counters can go on.
698 if (cpuctx->exclusive)
701 * If this group is exclusive and there are already
702 * counters on the CPU, it can't go on.
704 if (counter->attr.exclusive && cpuctx->active_oncpu)
707 * Otherwise, try to add it if all previous groups were able
713 static void add_counter_to_ctx(struct perf_counter *counter,
714 struct perf_counter_context *ctx)
716 list_add_counter(counter, ctx);
717 counter->tstamp_enabled = ctx->time;
718 counter->tstamp_running = ctx->time;
719 counter->tstamp_stopped = ctx->time;
723 * Cross CPU call to install and enable a performance counter
725 * Must be called with ctx->mutex held
727 static void __perf_install_in_context(void *info)
729 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
730 struct perf_counter *counter = info;
731 struct perf_counter_context *ctx = counter->ctx;
732 struct perf_counter *leader = counter->group_leader;
733 int cpu = smp_processor_id();
737 * If this is a task context, we need to check whether it is
738 * the current task context of this cpu. If not it has been
739 * scheduled out before the smp call arrived.
740 * Or possibly this is the right context but it isn't
741 * on this cpu because it had no counters.
743 if (ctx->task && cpuctx->task_ctx != ctx) {
744 if (cpuctx->task_ctx || ctx->task != current)
746 cpuctx->task_ctx = ctx;
749 spin_lock(&ctx->lock);
751 update_context_time(ctx);
754 * Protect the list operation against NMI by disabling the
755 * counters on a global level. NOP for non NMI based counters.
759 add_counter_to_ctx(counter, ctx);
762 * Don't put the counter on if it is disabled or if
763 * it is in a group and the group isn't on.
765 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
766 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
770 * An exclusive counter can't go on if there are already active
771 * hardware counters, and no hardware counter can go on if there
772 * is already an exclusive counter on.
774 if (!group_can_go_on(counter, cpuctx, 1))
777 err = counter_sched_in(counter, cpuctx, ctx, cpu);
781 * This counter couldn't go on. If it is in a group
782 * then we have to pull the whole group off.
783 * If the counter group is pinned then put it in error state.
785 if (leader != counter)
786 group_sched_out(leader, cpuctx, ctx);
787 if (leader->attr.pinned) {
788 update_group_times(leader);
789 leader->state = PERF_COUNTER_STATE_ERROR;
793 if (!err && !ctx->task && cpuctx->max_pertask)
794 cpuctx->max_pertask--;
799 spin_unlock(&ctx->lock);
803 * Attach a performance counter to a context
805 * First we add the counter to the list with the hardware enable bit
806 * in counter->hw_config cleared.
808 * If the counter is attached to a task which is on a CPU we use a smp
809 * call to enable it in the task context. The task might have been
810 * scheduled away, but we check this in the smp call again.
812 * Must be called with ctx->mutex held.
815 perf_install_in_context(struct perf_counter_context *ctx,
816 struct perf_counter *counter,
819 struct task_struct *task = ctx->task;
823 * Per cpu counters are installed via an smp call and
824 * the install is always sucessful.
826 smp_call_function_single(cpu, __perf_install_in_context,
832 task_oncpu_function_call(task, __perf_install_in_context,
835 spin_lock_irq(&ctx->lock);
837 * we need to retry the smp call.
839 if (ctx->is_active && list_empty(&counter->list_entry)) {
840 spin_unlock_irq(&ctx->lock);
845 * The lock prevents that this context is scheduled in so we
846 * can add the counter safely, if it the call above did not
849 if (list_empty(&counter->list_entry))
850 add_counter_to_ctx(counter, ctx);
851 spin_unlock_irq(&ctx->lock);
855 * Put a counter into inactive state and update time fields.
856 * Enabling the leader of a group effectively enables all
857 * the group members that aren't explicitly disabled, so we
858 * have to update their ->tstamp_enabled also.
859 * Note: this works for group members as well as group leaders
860 * since the non-leader members' sibling_lists will be empty.
862 static void __perf_counter_mark_enabled(struct perf_counter *counter,
863 struct perf_counter_context *ctx)
865 struct perf_counter *sub;
867 counter->state = PERF_COUNTER_STATE_INACTIVE;
868 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
869 list_for_each_entry(sub, &counter->sibling_list, list_entry)
870 if (sub->state >= PERF_COUNTER_STATE_INACTIVE)
871 sub->tstamp_enabled =
872 ctx->time - sub->total_time_enabled;
876 * Cross CPU call to enable a performance counter
878 static void __perf_counter_enable(void *info)
880 struct perf_counter *counter = info;
881 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
882 struct perf_counter_context *ctx = counter->ctx;
883 struct perf_counter *leader = counter->group_leader;
887 * If this is a per-task counter, need to check whether this
888 * counter's task is the current task on this cpu.
890 if (ctx->task && cpuctx->task_ctx != ctx) {
891 if (cpuctx->task_ctx || ctx->task != current)
893 cpuctx->task_ctx = ctx;
896 spin_lock(&ctx->lock);
898 update_context_time(ctx);
900 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
902 __perf_counter_mark_enabled(counter, ctx);
905 * If the counter is in a group and isn't the group leader,
906 * then don't put it on unless the group is on.
908 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
911 if (!group_can_go_on(counter, cpuctx, 1)) {
915 if (counter == leader)
916 err = group_sched_in(counter, cpuctx, ctx,
919 err = counter_sched_in(counter, cpuctx, ctx,
926 * If this counter can't go on and it's part of a
927 * group, then the whole group has to come off.
929 if (leader != counter)
930 group_sched_out(leader, cpuctx, ctx);
931 if (leader->attr.pinned) {
932 update_group_times(leader);
933 leader->state = PERF_COUNTER_STATE_ERROR;
938 spin_unlock(&ctx->lock);
944 * If counter->ctx is a cloned context, callers must make sure that
945 * every task struct that counter->ctx->task could possibly point to
946 * remains valid. This condition is satisfied when called through
947 * perf_counter_for_each_child or perf_counter_for_each as described
948 * for perf_counter_disable.
950 static void perf_counter_enable(struct perf_counter *counter)
952 struct perf_counter_context *ctx = counter->ctx;
953 struct task_struct *task = ctx->task;
957 * Enable the counter on the cpu that it's on
959 smp_call_function_single(counter->cpu, __perf_counter_enable,
964 spin_lock_irq(&ctx->lock);
965 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
969 * If the counter is in error state, clear that first.
970 * That way, if we see the counter in error state below, we
971 * know that it has gone back into error state, as distinct
972 * from the task having been scheduled away before the
973 * cross-call arrived.
975 if (counter->state == PERF_COUNTER_STATE_ERROR)
976 counter->state = PERF_COUNTER_STATE_OFF;
979 spin_unlock_irq(&ctx->lock);
980 task_oncpu_function_call(task, __perf_counter_enable, counter);
982 spin_lock_irq(&ctx->lock);
985 * If the context is active and the counter is still off,
986 * we need to retry the cross-call.
988 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
992 * Since we have the lock this context can't be scheduled
993 * in, so we can change the state safely.
995 if (counter->state == PERF_COUNTER_STATE_OFF)
996 __perf_counter_mark_enabled(counter, ctx);
999 spin_unlock_irq(&ctx->lock);
1002 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
1005 * not supported on inherited counters
1007 if (counter->attr.inherit)
1010 atomic_add(refresh, &counter->event_limit);
1011 perf_counter_enable(counter);
1016 void __perf_counter_sched_out(struct perf_counter_context *ctx,
1017 struct perf_cpu_context *cpuctx)
1019 struct perf_counter *counter;
1021 spin_lock(&ctx->lock);
1023 if (likely(!ctx->nr_counters))
1025 update_context_time(ctx);
1028 if (ctx->nr_active) {
1029 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1030 if (counter != counter->group_leader)
1031 counter_sched_out(counter, cpuctx, ctx);
1033 group_sched_out(counter, cpuctx, ctx);
1038 spin_unlock(&ctx->lock);
1042 * Test whether two contexts are equivalent, i.e. whether they
1043 * have both been cloned from the same version of the same context
1044 * and they both have the same number of enabled counters.
1045 * If the number of enabled counters is the same, then the set
1046 * of enabled counters should be the same, because these are both
1047 * inherited contexts, therefore we can't access individual counters
1048 * in them directly with an fd; we can only enable/disable all
1049 * counters via prctl, or enable/disable all counters in a family
1050 * via ioctl, which will have the same effect on both contexts.
1052 static int context_equiv(struct perf_counter_context *ctx1,
1053 struct perf_counter_context *ctx2)
1055 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1056 && ctx1->parent_gen == ctx2->parent_gen
1057 && !ctx1->pin_count && !ctx2->pin_count;
1060 static void __perf_counter_read(void *counter);
1062 static void __perf_counter_sync_stat(struct perf_counter *counter,
1063 struct perf_counter *next_counter)
1067 if (!counter->attr.inherit_stat)
1071 * Update the counter value, we cannot use perf_counter_read()
1072 * because we're in the middle of a context switch and have IRQs
1073 * disabled, which upsets smp_call_function_single(), however
1074 * we know the counter must be on the current CPU, therefore we
1075 * don't need to use it.
1077 switch (counter->state) {
1078 case PERF_COUNTER_STATE_ACTIVE:
1079 __perf_counter_read(counter);
1082 case PERF_COUNTER_STATE_INACTIVE:
1083 update_counter_times(counter);
1091 * In order to keep per-task stats reliable we need to flip the counter
1092 * values when we flip the contexts.
1094 value = atomic64_read(&next_counter->count);
1095 value = atomic64_xchg(&counter->count, value);
1096 atomic64_set(&next_counter->count, value);
1098 swap(counter->total_time_enabled, next_counter->total_time_enabled);
1099 swap(counter->total_time_running, next_counter->total_time_running);
1102 * Since we swizzled the values, update the user visible data too.
1104 perf_counter_update_userpage(counter);
1105 perf_counter_update_userpage(next_counter);
1108 #define list_next_entry(pos, member) \
1109 list_entry(pos->member.next, typeof(*pos), member)
1111 static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1112 struct perf_counter_context *next_ctx)
1114 struct perf_counter *counter, *next_counter;
1119 counter = list_first_entry(&ctx->event_list,
1120 struct perf_counter, event_entry);
1122 next_counter = list_first_entry(&next_ctx->event_list,
1123 struct perf_counter, event_entry);
1125 while (&counter->event_entry != &ctx->event_list &&
1126 &next_counter->event_entry != &next_ctx->event_list) {
1128 __perf_counter_sync_stat(counter, next_counter);
1130 counter = list_next_entry(counter, event_entry);
1131 next_counter = list_next_entry(next_counter, event_entry);
1136 * Called from scheduler to remove the counters of the current task,
1137 * with interrupts disabled.
1139 * We stop each counter and update the counter value in counter->count.
1141 * This does not protect us against NMI, but disable()
1142 * sets the disabled bit in the control field of counter _before_
1143 * accessing the counter control register. If a NMI hits, then it will
1144 * not restart the counter.
1146 void perf_counter_task_sched_out(struct task_struct *task,
1147 struct task_struct *next, int cpu)
1149 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1150 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1151 struct perf_counter_context *next_ctx;
1152 struct perf_counter_context *parent;
1153 struct pt_regs *regs;
1156 regs = task_pt_regs(task);
1157 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1159 if (likely(!ctx || !cpuctx->task_ctx))
1162 update_context_time(ctx);
1165 parent = rcu_dereference(ctx->parent_ctx);
1166 next_ctx = next->perf_counter_ctxp;
1167 if (parent && next_ctx &&
1168 rcu_dereference(next_ctx->parent_ctx) == parent) {
1170 * Looks like the two contexts are clones, so we might be
1171 * able to optimize the context switch. We lock both
1172 * contexts and check that they are clones under the
1173 * lock (including re-checking that neither has been
1174 * uncloned in the meantime). It doesn't matter which
1175 * order we take the locks because no other cpu could
1176 * be trying to lock both of these tasks.
1178 spin_lock(&ctx->lock);
1179 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1180 if (context_equiv(ctx, next_ctx)) {
1182 * XXX do we need a memory barrier of sorts
1183 * wrt to rcu_dereference() of perf_counter_ctxp
1185 task->perf_counter_ctxp = next_ctx;
1186 next->perf_counter_ctxp = ctx;
1188 next_ctx->task = task;
1191 perf_counter_sync_stat(ctx, next_ctx);
1193 spin_unlock(&next_ctx->lock);
1194 spin_unlock(&ctx->lock);
1199 __perf_counter_sched_out(ctx, cpuctx);
1200 cpuctx->task_ctx = NULL;
1205 * Called with IRQs disabled
1207 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1209 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1211 if (!cpuctx->task_ctx)
1214 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1217 __perf_counter_sched_out(ctx, cpuctx);
1218 cpuctx->task_ctx = NULL;
1222 * Called with IRQs disabled
1224 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1226 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1230 __perf_counter_sched_in(struct perf_counter_context *ctx,
1231 struct perf_cpu_context *cpuctx, int cpu)
1233 struct perf_counter *counter;
1236 spin_lock(&ctx->lock);
1238 if (likely(!ctx->nr_counters))
1241 ctx->timestamp = perf_clock();
1246 * First go through the list and put on any pinned groups
1247 * in order to give them the best chance of going on.
1249 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1250 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1251 !counter->attr.pinned)
1253 if (counter->cpu != -1 && counter->cpu != cpu)
1256 if (counter != counter->group_leader)
1257 counter_sched_in(counter, cpuctx, ctx, cpu);
1259 if (group_can_go_on(counter, cpuctx, 1))
1260 group_sched_in(counter, cpuctx, ctx, cpu);
1264 * If this pinned group hasn't been scheduled,
1265 * put it in error state.
1267 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1268 update_group_times(counter);
1269 counter->state = PERF_COUNTER_STATE_ERROR;
1273 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1275 * Ignore counters in OFF or ERROR state, and
1276 * ignore pinned counters since we did them already.
1278 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1279 counter->attr.pinned)
1283 * Listen to the 'cpu' scheduling filter constraint
1286 if (counter->cpu != -1 && counter->cpu != cpu)
1289 if (counter != counter->group_leader) {
1290 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1293 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1294 if (group_sched_in(counter, cpuctx, ctx, cpu))
1301 spin_unlock(&ctx->lock);
1305 * Called from scheduler to add the counters of the current task
1306 * with interrupts disabled.
1308 * We restore the counter value and then enable it.
1310 * This does not protect us against NMI, but enable()
1311 * sets the enabled bit in the control field of counter _before_
1312 * accessing the counter control register. If a NMI hits, then it will
1313 * keep the counter running.
1315 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1317 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1318 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1322 if (cpuctx->task_ctx == ctx)
1324 __perf_counter_sched_in(ctx, cpuctx, cpu);
1325 cpuctx->task_ctx = ctx;
1328 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1330 struct perf_counter_context *ctx = &cpuctx->ctx;
1332 __perf_counter_sched_in(ctx, cpuctx, cpu);
1335 #define MAX_INTERRUPTS (~0ULL)
1337 static void perf_log_throttle(struct perf_counter *counter, int enable);
1339 static void perf_adjust_period(struct perf_counter *counter, u64 events)
1341 struct hw_perf_counter *hwc = &counter->hw;
1342 u64 period, sample_period;
1345 events *= hwc->sample_period;
1346 period = div64_u64(events, counter->attr.sample_freq);
1348 delta = (s64)(period - hwc->sample_period);
1349 delta = (delta + 7) / 8; /* low pass filter */
1351 sample_period = hwc->sample_period + delta;
1356 hwc->sample_period = sample_period;
1359 static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1361 struct perf_counter *counter;
1362 struct hw_perf_counter *hwc;
1363 u64 interrupts, freq;
1365 spin_lock(&ctx->lock);
1366 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1367 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1372 interrupts = hwc->interrupts;
1373 hwc->interrupts = 0;
1376 * unthrottle counters on the tick
1378 if (interrupts == MAX_INTERRUPTS) {
1379 perf_log_throttle(counter, 1);
1380 counter->pmu->unthrottle(counter);
1381 interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1384 if (!counter->attr.freq || !counter->attr.sample_freq)
1388 * if the specified freq < HZ then we need to skip ticks
1390 if (counter->attr.sample_freq < HZ) {
1391 freq = counter->attr.sample_freq;
1393 hwc->freq_count += freq;
1394 hwc->freq_interrupts += interrupts;
1396 if (hwc->freq_count < HZ)
1399 interrupts = hwc->freq_interrupts;
1400 hwc->freq_interrupts = 0;
1401 hwc->freq_count -= HZ;
1405 perf_adjust_period(counter, freq * interrupts);
1408 * In order to avoid being stalled by an (accidental) huge
1409 * sample period, force reset the sample period if we didn't
1410 * get any events in this freq period.
1414 counter->pmu->disable(counter);
1415 atomic64_set(&hwc->period_left, 0);
1416 counter->pmu->enable(counter);
1420 spin_unlock(&ctx->lock);
1424 * Round-robin a context's counters:
1426 static void rotate_ctx(struct perf_counter_context *ctx)
1428 struct perf_counter *counter;
1430 if (!ctx->nr_counters)
1433 spin_lock(&ctx->lock);
1435 * Rotate the first entry last (works just fine for group counters too):
1438 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1439 list_move_tail(&counter->list_entry, &ctx->counter_list);
1444 spin_unlock(&ctx->lock);
1447 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1449 struct perf_cpu_context *cpuctx;
1450 struct perf_counter_context *ctx;
1452 if (!atomic_read(&nr_counters))
1455 cpuctx = &per_cpu(perf_cpu_context, cpu);
1456 ctx = curr->perf_counter_ctxp;
1458 perf_ctx_adjust_freq(&cpuctx->ctx);
1460 perf_ctx_adjust_freq(ctx);
1462 perf_counter_cpu_sched_out(cpuctx);
1464 __perf_counter_task_sched_out(ctx);
1466 rotate_ctx(&cpuctx->ctx);
1470 perf_counter_cpu_sched_in(cpuctx, cpu);
1472 perf_counter_task_sched_in(curr, cpu);
1476 * Enable all of a task's counters that have been marked enable-on-exec.
1477 * This expects task == current.
1479 static void perf_counter_enable_on_exec(struct task_struct *task)
1481 struct perf_counter_context *ctx;
1482 struct perf_counter *counter;
1483 unsigned long flags;
1486 local_irq_save(flags);
1487 ctx = task->perf_counter_ctxp;
1488 if (!ctx || !ctx->nr_counters)
1491 __perf_counter_task_sched_out(ctx);
1493 spin_lock(&ctx->lock);
1495 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1496 if (!counter->attr.enable_on_exec)
1498 counter->attr.enable_on_exec = 0;
1499 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
1501 __perf_counter_mark_enabled(counter, ctx);
1506 * Unclone this context if we enabled any counter.
1511 spin_unlock(&ctx->lock);
1513 perf_counter_task_sched_in(task, smp_processor_id());
1515 local_irq_restore(flags);
1519 * Cross CPU call to read the hardware counter
1521 static void __perf_counter_read(void *info)
1523 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1524 struct perf_counter *counter = info;
1525 struct perf_counter_context *ctx = counter->ctx;
1526 unsigned long flags;
1529 * If this is a task context, we need to check whether it is
1530 * the current task context of this cpu. If not it has been
1531 * scheduled out before the smp call arrived. In that case
1532 * counter->count would have been updated to a recent sample
1533 * when the counter was scheduled out.
1535 if (ctx->task && cpuctx->task_ctx != ctx)
1538 local_irq_save(flags);
1540 update_context_time(ctx);
1541 counter->pmu->read(counter);
1542 update_counter_times(counter);
1543 local_irq_restore(flags);
1546 static u64 perf_counter_read(struct perf_counter *counter)
1549 * If counter is enabled and currently active on a CPU, update the
1550 * value in the counter structure:
1552 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1553 smp_call_function_single(counter->oncpu,
1554 __perf_counter_read, counter, 1);
1555 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1556 update_counter_times(counter);
1559 return atomic64_read(&counter->count);
1563 * Initialize the perf_counter context in a task_struct:
1566 __perf_counter_init_context(struct perf_counter_context *ctx,
1567 struct task_struct *task)
1569 memset(ctx, 0, sizeof(*ctx));
1570 spin_lock_init(&ctx->lock);
1571 mutex_init(&ctx->mutex);
1572 INIT_LIST_HEAD(&ctx->counter_list);
1573 INIT_LIST_HEAD(&ctx->event_list);
1574 atomic_set(&ctx->refcount, 1);
1578 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1580 struct perf_counter_context *ctx;
1581 struct perf_cpu_context *cpuctx;
1582 struct task_struct *task;
1583 unsigned long flags;
1587 * If cpu is not a wildcard then this is a percpu counter:
1590 /* Must be root to operate on a CPU counter: */
1591 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1592 return ERR_PTR(-EACCES);
1594 if (cpu < 0 || cpu > num_possible_cpus())
1595 return ERR_PTR(-EINVAL);
1598 * We could be clever and allow to attach a counter to an
1599 * offline CPU and activate it when the CPU comes up, but
1602 if (!cpu_isset(cpu, cpu_online_map))
1603 return ERR_PTR(-ENODEV);
1605 cpuctx = &per_cpu(perf_cpu_context, cpu);
1616 task = find_task_by_vpid(pid);
1618 get_task_struct(task);
1622 return ERR_PTR(-ESRCH);
1625 * Can't attach counters to a dying task.
1628 if (task->flags & PF_EXITING)
1631 /* Reuse ptrace permission checks for now. */
1633 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1637 ctx = perf_lock_task_context(task, &flags);
1640 spin_unlock_irqrestore(&ctx->lock, flags);
1644 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1648 __perf_counter_init_context(ctx, task);
1650 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1652 * We raced with some other task; use
1653 * the context they set.
1658 get_task_struct(task);
1661 put_task_struct(task);
1665 put_task_struct(task);
1666 return ERR_PTR(err);
1669 static void free_counter_rcu(struct rcu_head *head)
1671 struct perf_counter *counter;
1673 counter = container_of(head, struct perf_counter, rcu_head);
1675 put_pid_ns(counter->ns);
1679 static void perf_pending_sync(struct perf_counter *counter);
1681 static void free_counter(struct perf_counter *counter)
1683 perf_pending_sync(counter);
1685 if (!counter->parent) {
1686 atomic_dec(&nr_counters);
1687 if (counter->attr.mmap)
1688 atomic_dec(&nr_mmap_counters);
1689 if (counter->attr.comm)
1690 atomic_dec(&nr_comm_counters);
1691 if (counter->attr.task)
1692 atomic_dec(&nr_task_counters);
1695 if (counter->output) {
1696 fput(counter->output->filp);
1697 counter->output = NULL;
1700 if (counter->destroy)
1701 counter->destroy(counter);
1703 put_ctx(counter->ctx);
1704 call_rcu(&counter->rcu_head, free_counter_rcu);
1708 * Called when the last reference to the file is gone.
1710 static int perf_release(struct inode *inode, struct file *file)
1712 struct perf_counter *counter = file->private_data;
1713 struct perf_counter_context *ctx = counter->ctx;
1715 file->private_data = NULL;
1717 WARN_ON_ONCE(ctx->parent_ctx);
1718 mutex_lock(&ctx->mutex);
1719 perf_counter_remove_from_context(counter);
1720 mutex_unlock(&ctx->mutex);
1722 mutex_lock(&counter->owner->perf_counter_mutex);
1723 list_del_init(&counter->owner_entry);
1724 mutex_unlock(&counter->owner->perf_counter_mutex);
1725 put_task_struct(counter->owner);
1727 free_counter(counter);
1732 static int perf_counter_read_size(struct perf_counter *counter)
1734 int entry = sizeof(u64); /* value */
1738 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1739 size += sizeof(u64);
1741 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1742 size += sizeof(u64);
1744 if (counter->attr.read_format & PERF_FORMAT_ID)
1745 entry += sizeof(u64);
1747 if (counter->attr.read_format & PERF_FORMAT_GROUP) {
1748 nr += counter->group_leader->nr_siblings;
1749 size += sizeof(u64);
1757 static u64 perf_counter_read_value(struct perf_counter *counter)
1759 struct perf_counter *child;
1762 total += perf_counter_read(counter);
1763 list_for_each_entry(child, &counter->child_list, child_list)
1764 total += perf_counter_read(child);
1769 static int perf_counter_read_entry(struct perf_counter *counter,
1770 u64 read_format, char __user *buf)
1772 int n = 0, count = 0;
1775 values[n++] = perf_counter_read_value(counter);
1776 if (read_format & PERF_FORMAT_ID)
1777 values[n++] = primary_counter_id(counter);
1779 count = n * sizeof(u64);
1781 if (copy_to_user(buf, values, count))
1787 static int perf_counter_read_group(struct perf_counter *counter,
1788 u64 read_format, char __user *buf)
1790 struct perf_counter *leader = counter->group_leader, *sub;
1791 int n = 0, size = 0, err = -EFAULT;
1794 values[n++] = 1 + leader->nr_siblings;
1795 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1796 values[n++] = leader->total_time_enabled +
1797 atomic64_read(&leader->child_total_time_enabled);
1799 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1800 values[n++] = leader->total_time_running +
1801 atomic64_read(&leader->child_total_time_running);
1804 size = n * sizeof(u64);
1806 if (copy_to_user(buf, values, size))
1809 err = perf_counter_read_entry(leader, read_format, buf + size);
1815 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1816 err = perf_counter_read_entry(sub, read_format,
1827 static int perf_counter_read_one(struct perf_counter *counter,
1828 u64 read_format, char __user *buf)
1833 values[n++] = perf_counter_read_value(counter);
1834 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1835 values[n++] = counter->total_time_enabled +
1836 atomic64_read(&counter->child_total_time_enabled);
1838 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1839 values[n++] = counter->total_time_running +
1840 atomic64_read(&counter->child_total_time_running);
1842 if (read_format & PERF_FORMAT_ID)
1843 values[n++] = primary_counter_id(counter);
1845 if (copy_to_user(buf, values, n * sizeof(u64)))
1848 return n * sizeof(u64);
1852 * Read the performance counter - simple non blocking version for now
1855 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1857 u64 read_format = counter->attr.read_format;
1861 * Return end-of-file for a read on a counter that is in
1862 * error state (i.e. because it was pinned but it couldn't be
1863 * scheduled on to the CPU at some point).
1865 if (counter->state == PERF_COUNTER_STATE_ERROR)
1868 if (count < perf_counter_read_size(counter))
1871 WARN_ON_ONCE(counter->ctx->parent_ctx);
1872 mutex_lock(&counter->child_mutex);
1873 if (read_format & PERF_FORMAT_GROUP)
1874 ret = perf_counter_read_group(counter, read_format, buf);
1876 ret = perf_counter_read_one(counter, read_format, buf);
1877 mutex_unlock(&counter->child_mutex);
1883 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1885 struct perf_counter *counter = file->private_data;
1887 return perf_read_hw(counter, buf, count);
1890 static unsigned int perf_poll(struct file *file, poll_table *wait)
1892 struct perf_counter *counter = file->private_data;
1893 struct perf_mmap_data *data;
1894 unsigned int events = POLL_HUP;
1897 data = rcu_dereference(counter->data);
1899 events = atomic_xchg(&data->poll, 0);
1902 poll_wait(file, &counter->waitq, wait);
1907 static void perf_counter_reset(struct perf_counter *counter)
1909 (void)perf_counter_read(counter);
1910 atomic64_set(&counter->count, 0);
1911 perf_counter_update_userpage(counter);
1915 * Holding the top-level counter's child_mutex means that any
1916 * descendant process that has inherited this counter will block
1917 * in sync_child_counter if it goes to exit, thus satisfying the
1918 * task existence requirements of perf_counter_enable/disable.
1920 static void perf_counter_for_each_child(struct perf_counter *counter,
1921 void (*func)(struct perf_counter *))
1923 struct perf_counter *child;
1925 WARN_ON_ONCE(counter->ctx->parent_ctx);
1926 mutex_lock(&counter->child_mutex);
1928 list_for_each_entry(child, &counter->child_list, child_list)
1930 mutex_unlock(&counter->child_mutex);
1933 static void perf_counter_for_each(struct perf_counter *counter,
1934 void (*func)(struct perf_counter *))
1936 struct perf_counter_context *ctx = counter->ctx;
1937 struct perf_counter *sibling;
1939 WARN_ON_ONCE(ctx->parent_ctx);
1940 mutex_lock(&ctx->mutex);
1941 counter = counter->group_leader;
1943 perf_counter_for_each_child(counter, func);
1945 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1946 perf_counter_for_each_child(counter, func);
1947 mutex_unlock(&ctx->mutex);
1950 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1952 struct perf_counter_context *ctx = counter->ctx;
1957 if (!counter->attr.sample_period)
1960 size = copy_from_user(&value, arg, sizeof(value));
1961 if (size != sizeof(value))
1967 spin_lock_irq(&ctx->lock);
1968 if (counter->attr.freq) {
1969 if (value > sysctl_perf_counter_sample_rate) {
1974 counter->attr.sample_freq = value;
1976 counter->attr.sample_period = value;
1977 counter->hw.sample_period = value;
1980 spin_unlock_irq(&ctx->lock);
1985 int perf_counter_set_output(struct perf_counter *counter, int output_fd);
1987 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1989 struct perf_counter *counter = file->private_data;
1990 void (*func)(struct perf_counter *);
1994 case PERF_COUNTER_IOC_ENABLE:
1995 func = perf_counter_enable;
1997 case PERF_COUNTER_IOC_DISABLE:
1998 func = perf_counter_disable;
2000 case PERF_COUNTER_IOC_RESET:
2001 func = perf_counter_reset;
2004 case PERF_COUNTER_IOC_REFRESH:
2005 return perf_counter_refresh(counter, arg);
2007 case PERF_COUNTER_IOC_PERIOD:
2008 return perf_counter_period(counter, (u64 __user *)arg);
2010 case PERF_COUNTER_IOC_SET_OUTPUT:
2011 return perf_counter_set_output(counter, arg);
2017 if (flags & PERF_IOC_FLAG_GROUP)
2018 perf_counter_for_each(counter, func);
2020 perf_counter_for_each_child(counter, func);
2025 int perf_counter_task_enable(void)
2027 struct perf_counter *counter;
2029 mutex_lock(¤t->perf_counter_mutex);
2030 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
2031 perf_counter_for_each_child(counter, perf_counter_enable);
2032 mutex_unlock(¤t->perf_counter_mutex);
2037 int perf_counter_task_disable(void)
2039 struct perf_counter *counter;
2041 mutex_lock(¤t->perf_counter_mutex);
2042 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
2043 perf_counter_for_each_child(counter, perf_counter_disable);
2044 mutex_unlock(¤t->perf_counter_mutex);
2049 #ifndef PERF_COUNTER_INDEX_OFFSET
2050 # define PERF_COUNTER_INDEX_OFFSET 0
2053 static int perf_counter_index(struct perf_counter *counter)
2055 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2058 return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
2062 * Callers need to ensure there can be no nesting of this function, otherwise
2063 * the seqlock logic goes bad. We can not serialize this because the arch
2064 * code calls this from NMI context.
2066 void perf_counter_update_userpage(struct perf_counter *counter)
2068 struct perf_counter_mmap_page *userpg;
2069 struct perf_mmap_data *data;
2072 data = rcu_dereference(counter->data);
2076 userpg = data->user_page;
2079 * Disable preemption so as to not let the corresponding user-space
2080 * spin too long if we get preempted.
2085 userpg->index = perf_counter_index(counter);
2086 userpg->offset = atomic64_read(&counter->count);
2087 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
2088 userpg->offset -= atomic64_read(&counter->hw.prev_count);
2090 userpg->time_enabled = counter->total_time_enabled +
2091 atomic64_read(&counter->child_total_time_enabled);
2093 userpg->time_running = counter->total_time_running +
2094 atomic64_read(&counter->child_total_time_running);
2103 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2105 struct perf_counter *counter = vma->vm_file->private_data;
2106 struct perf_mmap_data *data;
2107 int ret = VM_FAULT_SIGBUS;
2109 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2110 if (vmf->pgoff == 0)
2116 data = rcu_dereference(counter->data);
2120 if (vmf->pgoff == 0) {
2121 vmf->page = virt_to_page(data->user_page);
2123 int nr = vmf->pgoff - 1;
2125 if ((unsigned)nr > data->nr_pages)
2128 if (vmf->flags & FAULT_FLAG_WRITE)
2131 vmf->page = virt_to_page(data->data_pages[nr]);
2134 get_page(vmf->page);
2135 vmf->page->mapping = vma->vm_file->f_mapping;
2136 vmf->page->index = vmf->pgoff;
2145 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
2147 struct perf_mmap_data *data;
2151 WARN_ON(atomic_read(&counter->mmap_count));
2153 size = sizeof(struct perf_mmap_data);
2154 size += nr_pages * sizeof(void *);
2156 data = kzalloc(size, GFP_KERNEL);
2160 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2161 if (!data->user_page)
2162 goto fail_user_page;
2164 for (i = 0; i < nr_pages; i++) {
2165 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2166 if (!data->data_pages[i])
2167 goto fail_data_pages;
2170 data->nr_pages = nr_pages;
2171 atomic_set(&data->lock, -1);
2173 rcu_assign_pointer(counter->data, data);
2178 for (i--; i >= 0; i--)
2179 free_page((unsigned long)data->data_pages[i]);
2181 free_page((unsigned long)data->user_page);
2190 static void perf_mmap_free_page(unsigned long addr)
2192 struct page *page = virt_to_page((void *)addr);
2194 page->mapping = NULL;
2198 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2200 struct perf_mmap_data *data;
2203 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2205 perf_mmap_free_page((unsigned long)data->user_page);
2206 for (i = 0; i < data->nr_pages; i++)
2207 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2212 static void perf_mmap_data_free(struct perf_counter *counter)
2214 struct perf_mmap_data *data = counter->data;
2216 WARN_ON(atomic_read(&counter->mmap_count));
2218 rcu_assign_pointer(counter->data, NULL);
2219 call_rcu(&data->rcu_head, __perf_mmap_data_free);
2222 static void perf_mmap_open(struct vm_area_struct *vma)
2224 struct perf_counter *counter = vma->vm_file->private_data;
2226 atomic_inc(&counter->mmap_count);
2229 static void perf_mmap_close(struct vm_area_struct *vma)
2231 struct perf_counter *counter = vma->vm_file->private_data;
2233 WARN_ON_ONCE(counter->ctx->parent_ctx);
2234 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2235 struct user_struct *user = current_user();
2237 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2238 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2239 perf_mmap_data_free(counter);
2240 mutex_unlock(&counter->mmap_mutex);
2244 static struct vm_operations_struct perf_mmap_vmops = {
2245 .open = perf_mmap_open,
2246 .close = perf_mmap_close,
2247 .fault = perf_mmap_fault,
2248 .page_mkwrite = perf_mmap_fault,
2251 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2253 struct perf_counter *counter = file->private_data;
2254 unsigned long user_locked, user_lock_limit;
2255 struct user_struct *user = current_user();
2256 unsigned long locked, lock_limit;
2257 unsigned long vma_size;
2258 unsigned long nr_pages;
2259 long user_extra, extra;
2262 if (!(vma->vm_flags & VM_SHARED))
2265 vma_size = vma->vm_end - vma->vm_start;
2266 nr_pages = (vma_size / PAGE_SIZE) - 1;
2269 * If we have data pages ensure they're a power-of-two number, so we
2270 * can do bitmasks instead of modulo.
2272 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2275 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2278 if (vma->vm_pgoff != 0)
2281 WARN_ON_ONCE(counter->ctx->parent_ctx);
2282 mutex_lock(&counter->mmap_mutex);
2283 if (counter->output) {
2288 if (atomic_inc_not_zero(&counter->mmap_count)) {
2289 if (nr_pages != counter->data->nr_pages)
2294 user_extra = nr_pages + 1;
2295 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2298 * Increase the limit linearly with more CPUs:
2300 user_lock_limit *= num_online_cpus();
2302 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2305 if (user_locked > user_lock_limit)
2306 extra = user_locked - user_lock_limit;
2308 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2309 lock_limit >>= PAGE_SHIFT;
2310 locked = vma->vm_mm->locked_vm + extra;
2312 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2317 WARN_ON(counter->data);
2318 ret = perf_mmap_data_alloc(counter, nr_pages);
2322 atomic_set(&counter->mmap_count, 1);
2323 atomic_long_add(user_extra, &user->locked_vm);
2324 vma->vm_mm->locked_vm += extra;
2325 counter->data->nr_locked = extra;
2326 if (vma->vm_flags & VM_WRITE)
2327 counter->data->writable = 1;
2330 mutex_unlock(&counter->mmap_mutex);
2332 vma->vm_flags |= VM_RESERVED;
2333 vma->vm_ops = &perf_mmap_vmops;
2338 static int perf_fasync(int fd, struct file *filp, int on)
2340 struct inode *inode = filp->f_path.dentry->d_inode;
2341 struct perf_counter *counter = filp->private_data;
2344 mutex_lock(&inode->i_mutex);
2345 retval = fasync_helper(fd, filp, on, &counter->fasync);
2346 mutex_unlock(&inode->i_mutex);
2354 static const struct file_operations perf_fops = {
2355 .release = perf_release,
2358 .unlocked_ioctl = perf_ioctl,
2359 .compat_ioctl = perf_ioctl,
2361 .fasync = perf_fasync,
2365 * Perf counter wakeup
2367 * If there's data, ensure we set the poll() state and publish everything
2368 * to user-space before waking everybody up.
2371 void perf_counter_wakeup(struct perf_counter *counter)
2373 wake_up_all(&counter->waitq);
2375 if (counter->pending_kill) {
2376 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2377 counter->pending_kill = 0;
2384 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2386 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2387 * single linked list and use cmpxchg() to add entries lockless.
2390 static void perf_pending_counter(struct perf_pending_entry *entry)
2392 struct perf_counter *counter = container_of(entry,
2393 struct perf_counter, pending);
2395 if (counter->pending_disable) {
2396 counter->pending_disable = 0;
2397 __perf_counter_disable(counter);
2400 if (counter->pending_wakeup) {
2401 counter->pending_wakeup = 0;
2402 perf_counter_wakeup(counter);
2406 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2408 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2412 static void perf_pending_queue(struct perf_pending_entry *entry,
2413 void (*func)(struct perf_pending_entry *))
2415 struct perf_pending_entry **head;
2417 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2422 head = &get_cpu_var(perf_pending_head);
2425 entry->next = *head;
2426 } while (cmpxchg(head, entry->next, entry) != entry->next);
2428 set_perf_counter_pending();
2430 put_cpu_var(perf_pending_head);
2433 static int __perf_pending_run(void)
2435 struct perf_pending_entry *list;
2438 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2439 while (list != PENDING_TAIL) {
2440 void (*func)(struct perf_pending_entry *);
2441 struct perf_pending_entry *entry = list;
2448 * Ensure we observe the unqueue before we issue the wakeup,
2449 * so that we won't be waiting forever.
2450 * -- see perf_not_pending().
2461 static inline int perf_not_pending(struct perf_counter *counter)
2464 * If we flush on whatever cpu we run, there is a chance we don't
2468 __perf_pending_run();
2472 * Ensure we see the proper queue state before going to sleep
2473 * so that we do not miss the wakeup. -- see perf_pending_handle()
2476 return counter->pending.next == NULL;
2479 static void perf_pending_sync(struct perf_counter *counter)
2481 wait_event(counter->waitq, perf_not_pending(counter));
2484 void perf_counter_do_pending(void)
2486 __perf_pending_run();
2490 * Callchain support -- arch specific
2493 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2502 struct perf_output_handle {
2503 struct perf_counter *counter;
2504 struct perf_mmap_data *data;
2506 unsigned long offset;
2510 unsigned long flags;
2513 static bool perf_output_space(struct perf_mmap_data *data,
2514 unsigned int offset, unsigned int head)
2519 if (!data->writable)
2522 mask = (data->nr_pages << PAGE_SHIFT) - 1;
2524 * Userspace could choose to issue a mb() before updating the tail
2525 * pointer. So that all reads will be completed before the write is
2528 tail = ACCESS_ONCE(data->user_page->data_tail);
2531 offset = (offset - tail) & mask;
2532 head = (head - tail) & mask;
2534 if ((int)(head - offset) < 0)
2540 static void perf_output_wakeup(struct perf_output_handle *handle)
2542 atomic_set(&handle->data->poll, POLL_IN);
2545 handle->counter->pending_wakeup = 1;
2546 perf_pending_queue(&handle->counter->pending,
2547 perf_pending_counter);
2549 perf_counter_wakeup(handle->counter);
2553 * Curious locking construct.
2555 * We need to ensure a later event doesn't publish a head when a former
2556 * event isn't done writing. However since we need to deal with NMIs we
2557 * cannot fully serialize things.
2559 * What we do is serialize between CPUs so we only have to deal with NMI
2560 * nesting on a single CPU.
2562 * We only publish the head (and generate a wakeup) when the outer-most
2565 static void perf_output_lock(struct perf_output_handle *handle)
2567 struct perf_mmap_data *data = handle->data;
2572 local_irq_save(handle->flags);
2573 cpu = smp_processor_id();
2575 if (in_nmi() && atomic_read(&data->lock) == cpu)
2578 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2584 static void perf_output_unlock(struct perf_output_handle *handle)
2586 struct perf_mmap_data *data = handle->data;
2590 data->done_head = data->head;
2592 if (!handle->locked)
2597 * The xchg implies a full barrier that ensures all writes are done
2598 * before we publish the new head, matched by a rmb() in userspace when
2599 * reading this position.
2601 while ((head = atomic_long_xchg(&data->done_head, 0)))
2602 data->user_page->data_head = head;
2605 * NMI can happen here, which means we can miss a done_head update.
2608 cpu = atomic_xchg(&data->lock, -1);
2609 WARN_ON_ONCE(cpu != smp_processor_id());
2612 * Therefore we have to validate we did not indeed do so.
2614 if (unlikely(atomic_long_read(&data->done_head))) {
2616 * Since we had it locked, we can lock it again.
2618 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2624 if (atomic_xchg(&data->wakeup, 0))
2625 perf_output_wakeup(handle);
2627 local_irq_restore(handle->flags);
2630 static void perf_output_copy(struct perf_output_handle *handle,
2631 const void *buf, unsigned int len)
2633 unsigned int pages_mask;
2634 unsigned int offset;
2638 offset = handle->offset;
2639 pages_mask = handle->data->nr_pages - 1;
2640 pages = handle->data->data_pages;
2643 unsigned int page_offset;
2646 nr = (offset >> PAGE_SHIFT) & pages_mask;
2647 page_offset = offset & (PAGE_SIZE - 1);
2648 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2650 memcpy(pages[nr] + page_offset, buf, size);
2657 handle->offset = offset;
2660 * Check we didn't copy past our reservation window, taking the
2661 * possible unsigned int wrap into account.
2663 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2666 #define perf_output_put(handle, x) \
2667 perf_output_copy((handle), &(x), sizeof(x))
2669 static int perf_output_begin(struct perf_output_handle *handle,
2670 struct perf_counter *counter, unsigned int size,
2671 int nmi, int sample)
2673 struct perf_counter *output_counter;
2674 struct perf_mmap_data *data;
2675 unsigned int offset, head;
2678 struct perf_event_header header;
2685 * For inherited counters we send all the output towards the parent.
2687 if (counter->parent)
2688 counter = counter->parent;
2690 output_counter = rcu_dereference(counter->output);
2692 counter = output_counter;
2694 data = rcu_dereference(counter->data);
2698 handle->data = data;
2699 handle->counter = counter;
2701 handle->sample = sample;
2703 if (!data->nr_pages)
2706 have_lost = atomic_read(&data->lost);
2708 size += sizeof(lost_event);
2710 perf_output_lock(handle);
2713 offset = head = atomic_long_read(&data->head);
2715 if (unlikely(!perf_output_space(data, offset, head)))
2717 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2719 handle->offset = offset;
2720 handle->head = head;
2722 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2723 atomic_set(&data->wakeup, 1);
2726 lost_event.header.type = PERF_EVENT_LOST;
2727 lost_event.header.misc = 0;
2728 lost_event.header.size = sizeof(lost_event);
2729 lost_event.id = counter->id;
2730 lost_event.lost = atomic_xchg(&data->lost, 0);
2732 perf_output_put(handle, lost_event);
2738 atomic_inc(&data->lost);
2739 perf_output_unlock(handle);
2746 static void perf_output_end(struct perf_output_handle *handle)
2748 struct perf_counter *counter = handle->counter;
2749 struct perf_mmap_data *data = handle->data;
2751 int wakeup_events = counter->attr.wakeup_events;
2753 if (handle->sample && wakeup_events) {
2754 int events = atomic_inc_return(&data->events);
2755 if (events >= wakeup_events) {
2756 atomic_sub(wakeup_events, &data->events);
2757 atomic_set(&data->wakeup, 1);
2761 perf_output_unlock(handle);
2765 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2768 * only top level counters have the pid namespace they were created in
2770 if (counter->parent)
2771 counter = counter->parent;
2773 return task_tgid_nr_ns(p, counter->ns);
2776 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2779 * only top level counters have the pid namespace they were created in
2781 if (counter->parent)
2782 counter = counter->parent;
2784 return task_pid_nr_ns(p, counter->ns);
2787 static void perf_output_read_one(struct perf_output_handle *handle,
2788 struct perf_counter *counter)
2790 u64 read_format = counter->attr.read_format;
2794 values[n++] = atomic64_read(&counter->count);
2795 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2796 values[n++] = counter->total_time_enabled +
2797 atomic64_read(&counter->child_total_time_enabled);
2799 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2800 values[n++] = counter->total_time_running +
2801 atomic64_read(&counter->child_total_time_running);
2803 if (read_format & PERF_FORMAT_ID)
2804 values[n++] = primary_counter_id(counter);
2806 perf_output_copy(handle, values, n * sizeof(u64));
2810 * XXX PERF_FORMAT_GROUP vs inherited counters seems difficult.
2812 static void perf_output_read_group(struct perf_output_handle *handle,
2813 struct perf_counter *counter)
2815 struct perf_counter *leader = counter->group_leader, *sub;
2816 u64 read_format = counter->attr.read_format;
2820 values[n++] = 1 + leader->nr_siblings;
2822 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2823 values[n++] = leader->total_time_enabled;
2825 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2826 values[n++] = leader->total_time_running;
2828 if (leader != counter)
2829 leader->pmu->read(leader);
2831 values[n++] = atomic64_read(&leader->count);
2832 if (read_format & PERF_FORMAT_ID)
2833 values[n++] = primary_counter_id(leader);
2835 perf_output_copy(handle, values, n * sizeof(u64));
2837 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2841 sub->pmu->read(sub);
2843 values[n++] = atomic64_read(&sub->count);
2844 if (read_format & PERF_FORMAT_ID)
2845 values[n++] = primary_counter_id(sub);
2847 perf_output_copy(handle, values, n * sizeof(u64));
2851 static void perf_output_read(struct perf_output_handle *handle,
2852 struct perf_counter *counter)
2854 if (counter->attr.read_format & PERF_FORMAT_GROUP)
2855 perf_output_read_group(handle, counter);
2857 perf_output_read_one(handle, counter);
2860 void perf_counter_output(struct perf_counter *counter, int nmi,
2861 struct perf_sample_data *data)
2864 u64 sample_type = counter->attr.sample_type;
2865 struct perf_output_handle handle;
2866 struct perf_event_header header;
2871 struct perf_callchain_entry *callchain = NULL;
2872 int callchain_size = 0;
2878 header.type = PERF_EVENT_SAMPLE;
2879 header.size = sizeof(header);
2882 header.misc |= perf_misc_flags(data->regs);
2884 if (sample_type & PERF_SAMPLE_IP) {
2885 ip = perf_instruction_pointer(data->regs);
2886 header.size += sizeof(ip);
2889 if (sample_type & PERF_SAMPLE_TID) {
2890 /* namespace issues */
2891 tid_entry.pid = perf_counter_pid(counter, current);
2892 tid_entry.tid = perf_counter_tid(counter, current);
2894 header.size += sizeof(tid_entry);
2897 if (sample_type & PERF_SAMPLE_TIME) {
2899 * Maybe do better on x86 and provide cpu_clock_nmi()
2901 time = sched_clock();
2903 header.size += sizeof(u64);
2906 if (sample_type & PERF_SAMPLE_ADDR)
2907 header.size += sizeof(u64);
2909 if (sample_type & PERF_SAMPLE_ID)
2910 header.size += sizeof(u64);
2912 if (sample_type & PERF_SAMPLE_STREAM_ID)
2913 header.size += sizeof(u64);
2915 if (sample_type & PERF_SAMPLE_CPU) {
2916 header.size += sizeof(cpu_entry);
2918 cpu_entry.cpu = raw_smp_processor_id();
2919 cpu_entry.reserved = 0;
2922 if (sample_type & PERF_SAMPLE_PERIOD)
2923 header.size += sizeof(u64);
2925 if (sample_type & PERF_SAMPLE_READ)
2926 header.size += perf_counter_read_size(counter);
2928 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2929 callchain = perf_callchain(data->regs);
2932 callchain_size = (1 + callchain->nr) * sizeof(u64);
2933 header.size += callchain_size;
2935 header.size += sizeof(u64);
2938 if (sample_type & PERF_SAMPLE_RAW) {
2939 int size = sizeof(u32);
2942 size += data->raw->size;
2944 size += sizeof(u32);
2946 WARN_ON_ONCE(size & (sizeof(u64)-1));
2947 header.size += size;
2950 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2954 perf_output_put(&handle, header);
2956 if (sample_type & PERF_SAMPLE_IP)
2957 perf_output_put(&handle, ip);
2959 if (sample_type & PERF_SAMPLE_TID)
2960 perf_output_put(&handle, tid_entry);
2962 if (sample_type & PERF_SAMPLE_TIME)
2963 perf_output_put(&handle, time);
2965 if (sample_type & PERF_SAMPLE_ADDR)
2966 perf_output_put(&handle, data->addr);
2968 if (sample_type & PERF_SAMPLE_ID) {
2969 u64 id = primary_counter_id(counter);
2971 perf_output_put(&handle, id);
2974 if (sample_type & PERF_SAMPLE_STREAM_ID)
2975 perf_output_put(&handle, counter->id);
2977 if (sample_type & PERF_SAMPLE_CPU)
2978 perf_output_put(&handle, cpu_entry);
2980 if (sample_type & PERF_SAMPLE_PERIOD)
2981 perf_output_put(&handle, data->period);
2983 if (sample_type & PERF_SAMPLE_READ)
2984 perf_output_read(&handle, counter);
2986 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2988 perf_output_copy(&handle, callchain, callchain_size);
2991 perf_output_put(&handle, nr);
2995 if (sample_type & PERF_SAMPLE_RAW) {
2997 perf_output_put(&handle, data->raw->size);
2998 perf_output_copy(&handle, data->raw->data, data->raw->size);
3004 .size = sizeof(u32),
3007 perf_output_put(&handle, raw);
3011 perf_output_end(&handle);
3018 struct perf_read_event {
3019 struct perf_event_header header;
3026 perf_counter_read_event(struct perf_counter *counter,
3027 struct task_struct *task)
3029 struct perf_output_handle handle;
3030 struct perf_read_event event = {
3032 .type = PERF_EVENT_READ,
3034 .size = sizeof(event) + perf_counter_read_size(counter),
3036 .pid = perf_counter_pid(counter, task),
3037 .tid = perf_counter_tid(counter, task),
3041 ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
3045 perf_output_put(&handle, event);
3046 perf_output_read(&handle, counter);
3048 perf_output_end(&handle);
3052 * task tracking -- fork/exit
3054 * enabled by: attr.comm | attr.mmap | attr.task
3057 struct perf_task_event {
3058 struct task_struct *task;
3059 struct perf_counter_context *task_ctx;
3062 struct perf_event_header header;
3071 static void perf_counter_task_output(struct perf_counter *counter,
3072 struct perf_task_event *task_event)
3074 struct perf_output_handle handle;
3075 int size = task_event->event.header.size;
3076 struct task_struct *task = task_event->task;
3077 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3082 task_event->event.pid = perf_counter_pid(counter, task);
3083 task_event->event.ppid = perf_counter_pid(counter, current);
3085 task_event->event.tid = perf_counter_tid(counter, task);
3086 task_event->event.ptid = perf_counter_tid(counter, current);
3088 perf_output_put(&handle, task_event->event);
3089 perf_output_end(&handle);
3092 static int perf_counter_task_match(struct perf_counter *counter)
3094 if (counter->attr.comm || counter->attr.mmap || counter->attr.task)
3100 static void perf_counter_task_ctx(struct perf_counter_context *ctx,
3101 struct perf_task_event *task_event)
3103 struct perf_counter *counter;
3105 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3109 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3110 if (perf_counter_task_match(counter))
3111 perf_counter_task_output(counter, task_event);
3116 static void perf_counter_task_event(struct perf_task_event *task_event)
3118 struct perf_cpu_context *cpuctx;
3119 struct perf_counter_context *ctx = task_event->task_ctx;
3121 cpuctx = &get_cpu_var(perf_cpu_context);
3122 perf_counter_task_ctx(&cpuctx->ctx, task_event);
3123 put_cpu_var(perf_cpu_context);
3127 ctx = rcu_dereference(task_event->task->perf_counter_ctxp);
3129 perf_counter_task_ctx(ctx, task_event);
3133 static void perf_counter_task(struct task_struct *task,
3134 struct perf_counter_context *task_ctx,
3137 struct perf_task_event task_event;
3139 if (!atomic_read(&nr_comm_counters) &&
3140 !atomic_read(&nr_mmap_counters) &&
3141 !atomic_read(&nr_task_counters))
3144 task_event = (struct perf_task_event){
3146 .task_ctx = task_ctx,
3149 .type = new ? PERF_EVENT_FORK : PERF_EVENT_EXIT,
3151 .size = sizeof(task_event.event),
3160 perf_counter_task_event(&task_event);
3163 void perf_counter_fork(struct task_struct *task)
3165 perf_counter_task(task, NULL, 1);
3172 struct perf_comm_event {
3173 struct task_struct *task;
3178 struct perf_event_header header;
3185 static void perf_counter_comm_output(struct perf_counter *counter,
3186 struct perf_comm_event *comm_event)
3188 struct perf_output_handle handle;
3189 int size = comm_event->event.header.size;
3190 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3195 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
3196 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
3198 perf_output_put(&handle, comm_event->event);
3199 perf_output_copy(&handle, comm_event->comm,
3200 comm_event->comm_size);
3201 perf_output_end(&handle);
3204 static int perf_counter_comm_match(struct perf_counter *counter)
3206 if (counter->attr.comm)
3212 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
3213 struct perf_comm_event *comm_event)
3215 struct perf_counter *counter;
3217 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3221 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3222 if (perf_counter_comm_match(counter))
3223 perf_counter_comm_output(counter, comm_event);
3228 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
3230 struct perf_cpu_context *cpuctx;
3231 struct perf_counter_context *ctx;
3233 char comm[TASK_COMM_LEN];
3235 memset(comm, 0, sizeof(comm));
3236 strncpy(comm, comm_event->task->comm, sizeof(comm));
3237 size = ALIGN(strlen(comm)+1, sizeof(u64));
3239 comm_event->comm = comm;
3240 comm_event->comm_size = size;
3242 comm_event->event.header.size = sizeof(comm_event->event) + size;
3244 cpuctx = &get_cpu_var(perf_cpu_context);
3245 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
3246 put_cpu_var(perf_cpu_context);
3250 * doesn't really matter which of the child contexts the
3251 * events ends up in.
3253 ctx = rcu_dereference(current->perf_counter_ctxp);
3255 perf_counter_comm_ctx(ctx, comm_event);
3259 void perf_counter_comm(struct task_struct *task)
3261 struct perf_comm_event comm_event;
3263 if (task->perf_counter_ctxp)
3264 perf_counter_enable_on_exec(task);
3266 if (!atomic_read(&nr_comm_counters))
3269 comm_event = (struct perf_comm_event){
3275 .type = PERF_EVENT_COMM,
3284 perf_counter_comm_event(&comm_event);
3291 struct perf_mmap_event {
3292 struct vm_area_struct *vma;
3294 const char *file_name;
3298 struct perf_event_header header;
3308 static void perf_counter_mmap_output(struct perf_counter *counter,
3309 struct perf_mmap_event *mmap_event)
3311 struct perf_output_handle handle;
3312 int size = mmap_event->event.header.size;
3313 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3318 mmap_event->event.pid = perf_counter_pid(counter, current);
3319 mmap_event->event.tid = perf_counter_tid(counter, current);
3321 perf_output_put(&handle, mmap_event->event);
3322 perf_output_copy(&handle, mmap_event->file_name,
3323 mmap_event->file_size);
3324 perf_output_end(&handle);
3327 static int perf_counter_mmap_match(struct perf_counter *counter,
3328 struct perf_mmap_event *mmap_event)
3330 if (counter->attr.mmap)
3336 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3337 struct perf_mmap_event *mmap_event)
3339 struct perf_counter *counter;
3341 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3345 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3346 if (perf_counter_mmap_match(counter, mmap_event))
3347 perf_counter_mmap_output(counter, mmap_event);
3352 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3354 struct perf_cpu_context *cpuctx;
3355 struct perf_counter_context *ctx;
3356 struct vm_area_struct *vma = mmap_event->vma;
3357 struct file *file = vma->vm_file;
3363 memset(tmp, 0, sizeof(tmp));
3367 * d_path works from the end of the buffer backwards, so we
3368 * need to add enough zero bytes after the string to handle
3369 * the 64bit alignment we do later.
3371 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3373 name = strncpy(tmp, "//enomem", sizeof(tmp));
3376 name = d_path(&file->f_path, buf, PATH_MAX);
3378 name = strncpy(tmp, "//toolong", sizeof(tmp));
3382 if (arch_vma_name(mmap_event->vma)) {
3383 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3389 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3393 name = strncpy(tmp, "//anon", sizeof(tmp));
3398 size = ALIGN(strlen(name)+1, sizeof(u64));
3400 mmap_event->file_name = name;
3401 mmap_event->file_size = size;
3403 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3405 cpuctx = &get_cpu_var(perf_cpu_context);
3406 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3407 put_cpu_var(perf_cpu_context);
3411 * doesn't really matter which of the child contexts the
3412 * events ends up in.
3414 ctx = rcu_dereference(current->perf_counter_ctxp);
3416 perf_counter_mmap_ctx(ctx, mmap_event);
3422 void __perf_counter_mmap(struct vm_area_struct *vma)
3424 struct perf_mmap_event mmap_event;
3426 if (!atomic_read(&nr_mmap_counters))
3429 mmap_event = (struct perf_mmap_event){
3435 .type = PERF_EVENT_MMAP,
3441 .start = vma->vm_start,
3442 .len = vma->vm_end - vma->vm_start,
3443 .pgoff = vma->vm_pgoff,
3447 perf_counter_mmap_event(&mmap_event);
3451 * IRQ throttle logging
3454 static void perf_log_throttle(struct perf_counter *counter, int enable)
3456 struct perf_output_handle handle;
3460 struct perf_event_header header;
3464 } throttle_event = {
3466 .type = PERF_EVENT_THROTTLE,
3468 .size = sizeof(throttle_event),
3470 .time = sched_clock(),
3471 .id = primary_counter_id(counter),
3472 .stream_id = counter->id,
3476 throttle_event.header.type = PERF_EVENT_UNTHROTTLE;
3478 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3482 perf_output_put(&handle, throttle_event);
3483 perf_output_end(&handle);
3487 * Generic counter overflow handling, sampling.
3490 int perf_counter_overflow(struct perf_counter *counter, int nmi,
3491 struct perf_sample_data *data)
3493 int events = atomic_read(&counter->event_limit);
3494 int throttle = counter->pmu->unthrottle != NULL;
3495 struct hw_perf_counter *hwc = &counter->hw;
3501 if (hwc->interrupts != MAX_INTERRUPTS) {
3503 if (HZ * hwc->interrupts >
3504 (u64)sysctl_perf_counter_sample_rate) {
3505 hwc->interrupts = MAX_INTERRUPTS;
3506 perf_log_throttle(counter, 0);
3511 * Keep re-disabling counters even though on the previous
3512 * pass we disabled it - just in case we raced with a
3513 * sched-in and the counter got enabled again:
3519 if (counter->attr.freq) {
3520 u64 now = sched_clock();
3521 s64 delta = now - hwc->freq_stamp;
3523 hwc->freq_stamp = now;
3525 if (delta > 0 && delta < TICK_NSEC)
3526 perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3530 * XXX event_limit might not quite work as expected on inherited
3534 counter->pending_kill = POLL_IN;
3535 if (events && atomic_dec_and_test(&counter->event_limit)) {
3537 counter->pending_kill = POLL_HUP;
3539 counter->pending_disable = 1;
3540 perf_pending_queue(&counter->pending,
3541 perf_pending_counter);
3543 perf_counter_disable(counter);
3546 perf_counter_output(counter, nmi, data);
3551 * Generic software counter infrastructure
3555 * We directly increment counter->count and keep a second value in
3556 * counter->hw.period_left to count intervals. This period counter
3557 * is kept in the range [-sample_period, 0] so that we can use the
3561 static u64 perf_swcounter_set_period(struct perf_counter *counter)
3563 struct hw_perf_counter *hwc = &counter->hw;
3564 u64 period = hwc->last_period;
3568 hwc->last_period = hwc->sample_period;
3571 old = val = atomic64_read(&hwc->period_left);
3575 nr = div64_u64(period + val, period);
3576 offset = nr * period;
3578 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3584 static void perf_swcounter_overflow(struct perf_counter *counter,
3585 int nmi, struct perf_sample_data *data)
3587 struct hw_perf_counter *hwc = &counter->hw;
3590 data->period = counter->hw.last_period;
3591 overflow = perf_swcounter_set_period(counter);
3593 if (hwc->interrupts == MAX_INTERRUPTS)
3596 for (; overflow; overflow--) {
3597 if (perf_counter_overflow(counter, nmi, data)) {
3599 * We inhibit the overflow from happening when
3600 * hwc->interrupts == MAX_INTERRUPTS.
3607 static void perf_swcounter_unthrottle(struct perf_counter *counter)
3610 * Nothing to do, we already reset hwc->interrupts.
3614 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3615 int nmi, struct perf_sample_data *data)
3617 struct hw_perf_counter *hwc = &counter->hw;
3619 atomic64_add(nr, &counter->count);
3621 if (!hwc->sample_period)
3627 if (!atomic64_add_negative(nr, &hwc->period_left))
3628 perf_swcounter_overflow(counter, nmi, data);
3631 static int perf_swcounter_is_counting(struct perf_counter *counter)
3634 * The counter is active, we're good!
3636 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3640 * The counter is off/error, not counting.
3642 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3646 * The counter is inactive, if the context is active
3647 * we're part of a group that didn't make it on the 'pmu',
3650 if (counter->ctx->is_active)
3654 * We're inactive and the context is too, this means the
3655 * task is scheduled out, we're counting events that happen
3656 * to us, like migration events.
3661 static int perf_swcounter_match(struct perf_counter *counter,
3662 enum perf_type_id type,
3663 u32 event, struct pt_regs *regs)
3665 if (!perf_swcounter_is_counting(counter))
3668 if (counter->attr.type != type)
3670 if (counter->attr.config != event)
3674 if (counter->attr.exclude_user && user_mode(regs))
3677 if (counter->attr.exclude_kernel && !user_mode(regs))
3684 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3685 enum perf_type_id type,
3686 u32 event, u64 nr, int nmi,
3687 struct perf_sample_data *data)
3689 struct perf_counter *counter;
3691 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3695 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3696 if (perf_swcounter_match(counter, type, event, data->regs))
3697 perf_swcounter_add(counter, nr, nmi, data);
3702 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3705 return &cpuctx->recursion[3];
3708 return &cpuctx->recursion[2];
3711 return &cpuctx->recursion[1];
3713 return &cpuctx->recursion[0];
3716 static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3718 struct perf_sample_data *data)
3720 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3721 int *recursion = perf_swcounter_recursion_context(cpuctx);
3722 struct perf_counter_context *ctx;
3730 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3734 * doesn't really matter which of the child contexts the
3735 * events ends up in.
3737 ctx = rcu_dereference(current->perf_counter_ctxp);
3739 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3746 put_cpu_var(perf_cpu_context);
3749 void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3750 struct pt_regs *regs, u64 addr)
3752 struct perf_sample_data data = {
3757 do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3760 static void perf_swcounter_read(struct perf_counter *counter)
3764 static int perf_swcounter_enable(struct perf_counter *counter)
3766 struct hw_perf_counter *hwc = &counter->hw;
3768 if (hwc->sample_period) {
3769 hwc->last_period = hwc->sample_period;
3770 perf_swcounter_set_period(counter);
3775 static void perf_swcounter_disable(struct perf_counter *counter)
3779 static const struct pmu perf_ops_generic = {
3780 .enable = perf_swcounter_enable,
3781 .disable = perf_swcounter_disable,
3782 .read = perf_swcounter_read,
3783 .unthrottle = perf_swcounter_unthrottle,
3787 * hrtimer based swcounter callback
3790 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3792 enum hrtimer_restart ret = HRTIMER_RESTART;
3793 struct perf_sample_data data;
3794 struct perf_counter *counter;
3797 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3798 counter->pmu->read(counter);
3801 data.regs = get_irq_regs();
3803 * In case we exclude kernel IPs or are somehow not in interrupt
3804 * context, provide the next best thing, the user IP.
3806 if ((counter->attr.exclude_kernel || !data.regs) &&
3807 !counter->attr.exclude_user)
3808 data.regs = task_pt_regs(current);
3811 if (perf_counter_overflow(counter, 0, &data))
3812 ret = HRTIMER_NORESTART;
3815 period = max_t(u64, 10000, counter->hw.sample_period);
3816 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3822 * Software counter: cpu wall time clock
3825 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3827 int cpu = raw_smp_processor_id();
3831 now = cpu_clock(cpu);
3832 prev = atomic64_read(&counter->hw.prev_count);
3833 atomic64_set(&counter->hw.prev_count, now);
3834 atomic64_add(now - prev, &counter->count);
3837 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3839 struct hw_perf_counter *hwc = &counter->hw;
3840 int cpu = raw_smp_processor_id();
3842 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3843 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3844 hwc->hrtimer.function = perf_swcounter_hrtimer;
3845 if (hwc->sample_period) {
3846 u64 period = max_t(u64, 10000, hwc->sample_period);
3847 __hrtimer_start_range_ns(&hwc->hrtimer,
3848 ns_to_ktime(period), 0,
3849 HRTIMER_MODE_REL, 0);
3855 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3857 if (counter->hw.sample_period)
3858 hrtimer_cancel(&counter->hw.hrtimer);
3859 cpu_clock_perf_counter_update(counter);
3862 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3864 cpu_clock_perf_counter_update(counter);
3867 static const struct pmu perf_ops_cpu_clock = {
3868 .enable = cpu_clock_perf_counter_enable,
3869 .disable = cpu_clock_perf_counter_disable,
3870 .read = cpu_clock_perf_counter_read,
3874 * Software counter: task time clock
3877 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3882 prev = atomic64_xchg(&counter->hw.prev_count, now);
3884 atomic64_add(delta, &counter->count);
3887 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3889 struct hw_perf_counter *hwc = &counter->hw;
3892 now = counter->ctx->time;
3894 atomic64_set(&hwc->prev_count, now);
3895 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3896 hwc->hrtimer.function = perf_swcounter_hrtimer;
3897 if (hwc->sample_period) {
3898 u64 period = max_t(u64, 10000, hwc->sample_period);
3899 __hrtimer_start_range_ns(&hwc->hrtimer,
3900 ns_to_ktime(period), 0,
3901 HRTIMER_MODE_REL, 0);
3907 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3909 if (counter->hw.sample_period)
3910 hrtimer_cancel(&counter->hw.hrtimer);
3911 task_clock_perf_counter_update(counter, counter->ctx->time);
3915 static void task_clock_perf_counter_read(struct perf_counter *counter)
3920 update_context_time(counter->ctx);
3921 time = counter->ctx->time;
3923 u64 now = perf_clock();
3924 u64 delta = now - counter->ctx->timestamp;
3925 time = counter->ctx->time + delta;
3928 task_clock_perf_counter_update(counter, time);
3931 static const struct pmu perf_ops_task_clock = {
3932 .enable = task_clock_perf_counter_enable,
3933 .disable = task_clock_perf_counter_disable,
3934 .read = task_clock_perf_counter_read,
3937 #ifdef CONFIG_EVENT_PROFILE
3938 void perf_tpcounter_event(int event_id, u64 addr, u64 count, void *record,
3941 struct perf_raw_record raw = {
3946 struct perf_sample_data data = {
3947 .regs = get_irq_regs(),
3953 data.regs = task_pt_regs(current);
3955 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, count, 1, &data);
3957 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3959 extern int ftrace_profile_enable(int);
3960 extern void ftrace_profile_disable(int);
3962 static void tp_perf_counter_destroy(struct perf_counter *counter)
3964 ftrace_profile_disable(counter->attr.config);
3967 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3970 * Raw tracepoint data is a severe data leak, only allow root to
3973 if ((counter->attr.sample_type & PERF_SAMPLE_RAW) &&
3974 !capable(CAP_SYS_ADMIN))
3975 return ERR_PTR(-EPERM);
3977 if (ftrace_profile_enable(counter->attr.config))
3980 counter->destroy = tp_perf_counter_destroy;
3982 return &perf_ops_generic;
3985 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3991 atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3993 static void sw_perf_counter_destroy(struct perf_counter *counter)
3995 u64 event = counter->attr.config;
3997 WARN_ON(counter->parent);
3999 atomic_dec(&perf_swcounter_enabled[event]);
4002 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
4004 const struct pmu *pmu = NULL;
4005 u64 event = counter->attr.config;
4008 * Software counters (currently) can't in general distinguish
4009 * between user, kernel and hypervisor events.
4010 * However, context switches and cpu migrations are considered
4011 * to be kernel events, and page faults are never hypervisor
4015 case PERF_COUNT_SW_CPU_CLOCK:
4016 pmu = &perf_ops_cpu_clock;
4019 case PERF_COUNT_SW_TASK_CLOCK:
4021 * If the user instantiates this as a per-cpu counter,
4022 * use the cpu_clock counter instead.
4024 if (counter->ctx->task)
4025 pmu = &perf_ops_task_clock;
4027 pmu = &perf_ops_cpu_clock;
4030 case PERF_COUNT_SW_PAGE_FAULTS:
4031 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4032 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4033 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4034 case PERF_COUNT_SW_CPU_MIGRATIONS:
4035 if (!counter->parent) {
4036 atomic_inc(&perf_swcounter_enabled[event]);
4037 counter->destroy = sw_perf_counter_destroy;
4039 pmu = &perf_ops_generic;
4047 * Allocate and initialize a counter structure
4049 static struct perf_counter *
4050 perf_counter_alloc(struct perf_counter_attr *attr,
4052 struct perf_counter_context *ctx,
4053 struct perf_counter *group_leader,
4054 struct perf_counter *parent_counter,
4057 const struct pmu *pmu;
4058 struct perf_counter *counter;
4059 struct hw_perf_counter *hwc;
4062 counter = kzalloc(sizeof(*counter), gfpflags);
4064 return ERR_PTR(-ENOMEM);
4067 * Single counters are their own group leaders, with an
4068 * empty sibling list:
4071 group_leader = counter;
4073 mutex_init(&counter->child_mutex);
4074 INIT_LIST_HEAD(&counter->child_list);
4076 INIT_LIST_HEAD(&counter->list_entry);
4077 INIT_LIST_HEAD(&counter->event_entry);
4078 INIT_LIST_HEAD(&counter->sibling_list);
4079 init_waitqueue_head(&counter->waitq);
4081 mutex_init(&counter->mmap_mutex);
4084 counter->attr = *attr;
4085 counter->group_leader = group_leader;
4086 counter->pmu = NULL;
4088 counter->oncpu = -1;
4090 counter->parent = parent_counter;
4092 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
4093 counter->id = atomic64_inc_return(&perf_counter_id);
4095 counter->state = PERF_COUNTER_STATE_INACTIVE;
4098 counter->state = PERF_COUNTER_STATE_OFF;
4103 hwc->sample_period = attr->sample_period;
4104 if (attr->freq && attr->sample_freq)
4105 hwc->sample_period = 1;
4106 hwc->last_period = hwc->sample_period;
4108 atomic64_set(&hwc->period_left, hwc->sample_period);
4111 * we currently do not support PERF_FORMAT_GROUP on inherited counters
4113 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4116 switch (attr->type) {
4118 case PERF_TYPE_HARDWARE:
4119 case PERF_TYPE_HW_CACHE:
4120 pmu = hw_perf_counter_init(counter);
4123 case PERF_TYPE_SOFTWARE:
4124 pmu = sw_perf_counter_init(counter);
4127 case PERF_TYPE_TRACEPOINT:
4128 pmu = tp_perf_counter_init(counter);
4138 else if (IS_ERR(pmu))
4143 put_pid_ns(counter->ns);
4145 return ERR_PTR(err);
4150 if (!counter->parent) {
4151 atomic_inc(&nr_counters);
4152 if (counter->attr.mmap)
4153 atomic_inc(&nr_mmap_counters);
4154 if (counter->attr.comm)
4155 atomic_inc(&nr_comm_counters);
4156 if (counter->attr.task)
4157 atomic_inc(&nr_task_counters);
4163 static int perf_copy_attr(struct perf_counter_attr __user *uattr,
4164 struct perf_counter_attr *attr)
4169 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4173 * zero the full structure, so that a short copy will be nice.
4175 memset(attr, 0, sizeof(*attr));
4177 ret = get_user(size, &uattr->size);
4181 if (size > PAGE_SIZE) /* silly large */
4184 if (!size) /* abi compat */
4185 size = PERF_ATTR_SIZE_VER0;
4187 if (size < PERF_ATTR_SIZE_VER0)
4191 * If we're handed a bigger struct than we know of,
4192 * ensure all the unknown bits are 0.
4194 if (size > sizeof(*attr)) {
4196 unsigned long __user *addr;
4197 unsigned long __user *end;
4199 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
4200 sizeof(unsigned long));
4201 end = PTR_ALIGN((void __user *)uattr + size,
4202 sizeof(unsigned long));
4204 for (; addr < end; addr += sizeof(unsigned long)) {
4205 ret = get_user(val, addr);
4213 ret = copy_from_user(attr, uattr, size);
4218 * If the type exists, the corresponding creation will verify
4221 if (attr->type >= PERF_TYPE_MAX)
4224 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4227 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4230 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4237 put_user(sizeof(*attr), &uattr->size);
4242 int perf_counter_set_output(struct perf_counter *counter, int output_fd)
4244 struct perf_counter *output_counter = NULL;
4245 struct file *output_file = NULL;
4246 struct perf_counter *old_output;
4247 int fput_needed = 0;
4253 output_file = fget_light(output_fd, &fput_needed);
4257 if (output_file->f_op != &perf_fops)
4260 output_counter = output_file->private_data;
4262 /* Don't chain output fds */
4263 if (output_counter->output)
4266 /* Don't set an output fd when we already have an output channel */
4270 atomic_long_inc(&output_file->f_count);
4273 mutex_lock(&counter->mmap_mutex);
4274 old_output = counter->output;
4275 rcu_assign_pointer(counter->output, output_counter);
4276 mutex_unlock(&counter->mmap_mutex);
4280 * we need to make sure no existing perf_output_*()
4281 * is still referencing this counter.
4284 fput(old_output->filp);
4289 fput_light(output_file, fput_needed);
4294 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
4296 * @attr_uptr: event type attributes for monitoring/sampling
4299 * @group_fd: group leader counter fd
4301 SYSCALL_DEFINE5(perf_counter_open,
4302 struct perf_counter_attr __user *, attr_uptr,
4303 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4305 struct perf_counter *counter, *group_leader;
4306 struct perf_counter_attr attr;
4307 struct perf_counter_context *ctx;
4308 struct file *counter_file = NULL;
4309 struct file *group_file = NULL;
4310 int fput_needed = 0;
4311 int fput_needed2 = 0;
4314 /* for future expandability... */
4315 if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4318 ret = perf_copy_attr(attr_uptr, &attr);
4322 if (!attr.exclude_kernel) {
4323 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4328 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
4333 * Get the target context (task or percpu):
4335 ctx = find_get_context(pid, cpu);
4337 return PTR_ERR(ctx);
4340 * Look up the group leader (we will attach this counter to it):
4342 group_leader = NULL;
4343 if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4345 group_file = fget_light(group_fd, &fput_needed);
4347 goto err_put_context;
4348 if (group_file->f_op != &perf_fops)
4349 goto err_put_context;
4351 group_leader = group_file->private_data;
4353 * Do not allow a recursive hierarchy (this new sibling
4354 * becoming part of another group-sibling):
4356 if (group_leader->group_leader != group_leader)
4357 goto err_put_context;
4359 * Do not allow to attach to a group in a different
4360 * task or CPU context:
4362 if (group_leader->ctx != ctx)
4363 goto err_put_context;
4365 * Only a group leader can be exclusive or pinned
4367 if (attr.exclusive || attr.pinned)
4368 goto err_put_context;
4371 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
4373 ret = PTR_ERR(counter);
4374 if (IS_ERR(counter))
4375 goto err_put_context;
4377 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
4379 goto err_free_put_context;
4381 counter_file = fget_light(ret, &fput_needed2);
4383 goto err_free_put_context;
4385 if (flags & PERF_FLAG_FD_OUTPUT) {
4386 ret = perf_counter_set_output(counter, group_fd);
4388 goto err_free_put_context;
4391 counter->filp = counter_file;
4392 WARN_ON_ONCE(ctx->parent_ctx);
4393 mutex_lock(&ctx->mutex);
4394 perf_install_in_context(ctx, counter, cpu);
4396 mutex_unlock(&ctx->mutex);
4398 counter->owner = current;
4399 get_task_struct(current);
4400 mutex_lock(¤t->perf_counter_mutex);
4401 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
4402 mutex_unlock(¤t->perf_counter_mutex);
4404 fput_light(counter_file, fput_needed2);
4407 fput_light(group_file, fput_needed);
4411 err_free_put_context:
4421 * inherit a counter from parent task to child task:
4423 static struct perf_counter *
4424 inherit_counter(struct perf_counter *parent_counter,
4425 struct task_struct *parent,
4426 struct perf_counter_context *parent_ctx,
4427 struct task_struct *child,
4428 struct perf_counter *group_leader,
4429 struct perf_counter_context *child_ctx)
4431 struct perf_counter *child_counter;
4434 * Instead of creating recursive hierarchies of counters,
4435 * we link inherited counters back to the original parent,
4436 * which has a filp for sure, which we use as the reference
4439 if (parent_counter->parent)
4440 parent_counter = parent_counter->parent;
4442 child_counter = perf_counter_alloc(&parent_counter->attr,
4443 parent_counter->cpu, child_ctx,
4444 group_leader, parent_counter,
4446 if (IS_ERR(child_counter))
4447 return child_counter;
4451 * Make the child state follow the state of the parent counter,
4452 * not its attr.disabled bit. We hold the parent's mutex,
4453 * so we won't race with perf_counter_{en, dis}able_family.
4455 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4456 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4458 child_counter->state = PERF_COUNTER_STATE_OFF;
4460 if (parent_counter->attr.freq)
4461 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4464 * Link it up in the child's context:
4466 add_counter_to_ctx(child_counter, child_ctx);
4469 * Get a reference to the parent filp - we will fput it
4470 * when the child counter exits. This is safe to do because
4471 * we are in the parent and we know that the filp still
4472 * exists and has a nonzero count:
4474 atomic_long_inc(&parent_counter->filp->f_count);
4477 * Link this into the parent counter's child list
4479 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4480 mutex_lock(&parent_counter->child_mutex);
4481 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4482 mutex_unlock(&parent_counter->child_mutex);
4484 return child_counter;
4487 static int inherit_group(struct perf_counter *parent_counter,
4488 struct task_struct *parent,
4489 struct perf_counter_context *parent_ctx,
4490 struct task_struct *child,
4491 struct perf_counter_context *child_ctx)
4493 struct perf_counter *leader;
4494 struct perf_counter *sub;
4495 struct perf_counter *child_ctr;
4497 leader = inherit_counter(parent_counter, parent, parent_ctx,
4498 child, NULL, child_ctx);
4500 return PTR_ERR(leader);
4501 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4502 child_ctr = inherit_counter(sub, parent, parent_ctx,
4503 child, leader, child_ctx);
4504 if (IS_ERR(child_ctr))
4505 return PTR_ERR(child_ctr);
4510 static void sync_child_counter(struct perf_counter *child_counter,
4511 struct task_struct *child)
4513 struct perf_counter *parent_counter = child_counter->parent;
4516 if (child_counter->attr.inherit_stat)
4517 perf_counter_read_event(child_counter, child);
4519 child_val = atomic64_read(&child_counter->count);
4522 * Add back the child's count to the parent's count:
4524 atomic64_add(child_val, &parent_counter->count);
4525 atomic64_add(child_counter->total_time_enabled,
4526 &parent_counter->child_total_time_enabled);
4527 atomic64_add(child_counter->total_time_running,
4528 &parent_counter->child_total_time_running);
4531 * Remove this counter from the parent's list
4533 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4534 mutex_lock(&parent_counter->child_mutex);
4535 list_del_init(&child_counter->child_list);
4536 mutex_unlock(&parent_counter->child_mutex);
4539 * Release the parent counter, if this was the last
4542 fput(parent_counter->filp);
4546 __perf_counter_exit_task(struct perf_counter *child_counter,
4547 struct perf_counter_context *child_ctx,
4548 struct task_struct *child)
4550 struct perf_counter *parent_counter;
4552 update_counter_times(child_counter);
4553 perf_counter_remove_from_context(child_counter);
4555 parent_counter = child_counter->parent;
4557 * It can happen that parent exits first, and has counters
4558 * that are still around due to the child reference. These
4559 * counters need to be zapped - but otherwise linger.
4561 if (parent_counter) {
4562 sync_child_counter(child_counter, child);
4563 free_counter(child_counter);
4568 * When a child task exits, feed back counter values to parent counters.
4570 void perf_counter_exit_task(struct task_struct *child)
4572 struct perf_counter *child_counter, *tmp;
4573 struct perf_counter_context *child_ctx;
4574 unsigned long flags;
4576 if (likely(!child->perf_counter_ctxp)) {
4577 perf_counter_task(child, NULL, 0);
4581 local_irq_save(flags);
4583 * We can't reschedule here because interrupts are disabled,
4584 * and either child is current or it is a task that can't be
4585 * scheduled, so we are now safe from rescheduling changing
4588 child_ctx = child->perf_counter_ctxp;
4589 __perf_counter_task_sched_out(child_ctx);
4592 * Take the context lock here so that if find_get_context is
4593 * reading child->perf_counter_ctxp, we wait until it has
4594 * incremented the context's refcount before we do put_ctx below.
4596 spin_lock(&child_ctx->lock);
4597 child->perf_counter_ctxp = NULL;
4599 * If this context is a clone; unclone it so it can't get
4600 * swapped to another process while we're removing all
4601 * the counters from it.
4603 unclone_ctx(child_ctx);
4604 spin_unlock_irqrestore(&child_ctx->lock, flags);
4607 * Report the task dead after unscheduling the counters so that we
4608 * won't get any samples after PERF_EVENT_EXIT. We can however still
4609 * get a few PERF_EVENT_READ events.
4611 perf_counter_task(child, child_ctx, 0);
4614 * We can recurse on the same lock type through:
4616 * __perf_counter_exit_task()
4617 * sync_child_counter()
4618 * fput(parent_counter->filp)
4620 * mutex_lock(&ctx->mutex)
4622 * But since its the parent context it won't be the same instance.
4624 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4627 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4629 __perf_counter_exit_task(child_counter, child_ctx, child);
4632 * If the last counter was a group counter, it will have appended all
4633 * its siblings to the list, but we obtained 'tmp' before that which
4634 * will still point to the list head terminating the iteration.
4636 if (!list_empty(&child_ctx->counter_list))
4639 mutex_unlock(&child_ctx->mutex);
4645 * free an unexposed, unused context as created by inheritance by
4646 * init_task below, used by fork() in case of fail.
4648 void perf_counter_free_task(struct task_struct *task)
4650 struct perf_counter_context *ctx = task->perf_counter_ctxp;
4651 struct perf_counter *counter, *tmp;
4656 mutex_lock(&ctx->mutex);
4658 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4659 struct perf_counter *parent = counter->parent;
4661 if (WARN_ON_ONCE(!parent))
4664 mutex_lock(&parent->child_mutex);
4665 list_del_init(&counter->child_list);
4666 mutex_unlock(&parent->child_mutex);
4670 list_del_counter(counter, ctx);
4671 free_counter(counter);
4674 if (!list_empty(&ctx->counter_list))
4677 mutex_unlock(&ctx->mutex);
4683 * Initialize the perf_counter context in task_struct
4685 int perf_counter_init_task(struct task_struct *child)
4687 struct perf_counter_context *child_ctx, *parent_ctx;
4688 struct perf_counter_context *cloned_ctx;
4689 struct perf_counter *counter;
4690 struct task_struct *parent = current;
4691 int inherited_all = 1;
4694 child->perf_counter_ctxp = NULL;
4696 mutex_init(&child->perf_counter_mutex);
4697 INIT_LIST_HEAD(&child->perf_counter_list);
4699 if (likely(!parent->perf_counter_ctxp))
4703 * This is executed from the parent task context, so inherit
4704 * counters that have been marked for cloning.
4705 * First allocate and initialize a context for the child.
4708 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4712 __perf_counter_init_context(child_ctx, child);
4713 child->perf_counter_ctxp = child_ctx;
4714 get_task_struct(child);
4717 * If the parent's context is a clone, pin it so it won't get
4720 parent_ctx = perf_pin_task_context(parent);
4723 * No need to check if parent_ctx != NULL here; since we saw
4724 * it non-NULL earlier, the only reason for it to become NULL
4725 * is if we exit, and since we're currently in the middle of
4726 * a fork we can't be exiting at the same time.
4730 * Lock the parent list. No need to lock the child - not PID
4731 * hashed yet and not running, so nobody can access it.
4733 mutex_lock(&parent_ctx->mutex);
4736 * We dont have to disable NMIs - we are only looking at
4737 * the list, not manipulating it:
4739 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4740 if (counter != counter->group_leader)
4743 if (!counter->attr.inherit) {
4748 ret = inherit_group(counter, parent, parent_ctx,
4756 if (inherited_all) {
4758 * Mark the child context as a clone of the parent
4759 * context, or of whatever the parent is a clone of.
4760 * Note that if the parent is a clone, it could get
4761 * uncloned at any point, but that doesn't matter
4762 * because the list of counters and the generation
4763 * count can't have changed since we took the mutex.
4765 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4767 child_ctx->parent_ctx = cloned_ctx;
4768 child_ctx->parent_gen = parent_ctx->parent_gen;
4770 child_ctx->parent_ctx = parent_ctx;
4771 child_ctx->parent_gen = parent_ctx->generation;
4773 get_ctx(child_ctx->parent_ctx);
4776 mutex_unlock(&parent_ctx->mutex);
4778 perf_unpin_context(parent_ctx);
4783 static void __cpuinit perf_counter_init_cpu(int cpu)
4785 struct perf_cpu_context *cpuctx;
4787 cpuctx = &per_cpu(perf_cpu_context, cpu);
4788 __perf_counter_init_context(&cpuctx->ctx, NULL);
4790 spin_lock(&perf_resource_lock);
4791 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4792 spin_unlock(&perf_resource_lock);
4794 hw_perf_counter_setup(cpu);
4797 #ifdef CONFIG_HOTPLUG_CPU
4798 static void __perf_counter_exit_cpu(void *info)
4800 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4801 struct perf_counter_context *ctx = &cpuctx->ctx;
4802 struct perf_counter *counter, *tmp;
4804 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4805 __perf_counter_remove_from_context(counter);
4807 static void perf_counter_exit_cpu(int cpu)
4809 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4810 struct perf_counter_context *ctx = &cpuctx->ctx;
4812 mutex_lock(&ctx->mutex);
4813 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4814 mutex_unlock(&ctx->mutex);
4817 static inline void perf_counter_exit_cpu(int cpu) { }
4820 static int __cpuinit
4821 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4823 unsigned int cpu = (long)hcpu;
4827 case CPU_UP_PREPARE:
4828 case CPU_UP_PREPARE_FROZEN:
4829 perf_counter_init_cpu(cpu);
4833 case CPU_ONLINE_FROZEN:
4834 hw_perf_counter_setup_online(cpu);
4837 case CPU_DOWN_PREPARE:
4838 case CPU_DOWN_PREPARE_FROZEN:
4839 perf_counter_exit_cpu(cpu);
4850 * This has to have a higher priority than migration_notifier in sched.c.
4852 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4853 .notifier_call = perf_cpu_notify,
4857 void __init perf_counter_init(void)
4859 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4860 (void *)(long)smp_processor_id());
4861 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
4862 (void *)(long)smp_processor_id());
4863 register_cpu_notifier(&perf_cpu_nb);
4866 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4868 return sprintf(buf, "%d\n", perf_reserved_percpu);
4872 perf_set_reserve_percpu(struct sysdev_class *class,
4876 struct perf_cpu_context *cpuctx;
4880 err = strict_strtoul(buf, 10, &val);
4883 if (val > perf_max_counters)
4886 spin_lock(&perf_resource_lock);
4887 perf_reserved_percpu = val;
4888 for_each_online_cpu(cpu) {
4889 cpuctx = &per_cpu(perf_cpu_context, cpu);
4890 spin_lock_irq(&cpuctx->ctx.lock);
4891 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4892 perf_max_counters - perf_reserved_percpu);
4893 cpuctx->max_pertask = mpt;
4894 spin_unlock_irq(&cpuctx->ctx.lock);
4896 spin_unlock(&perf_resource_lock);
4901 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4903 return sprintf(buf, "%d\n", perf_overcommit);
4907 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4912 err = strict_strtoul(buf, 10, &val);
4918 spin_lock(&perf_resource_lock);
4919 perf_overcommit = val;
4920 spin_unlock(&perf_resource_lock);
4925 static SYSDEV_CLASS_ATTR(
4928 perf_show_reserve_percpu,
4929 perf_set_reserve_percpu
4932 static SYSDEV_CLASS_ATTR(
4935 perf_show_overcommit,
4939 static struct attribute *perfclass_attrs[] = {
4940 &attr_reserve_percpu.attr,
4941 &attr_overcommit.attr,
4945 static struct attribute_group perfclass_attr_group = {
4946 .attrs = perfclass_attrs,
4947 .name = "perf_counters",
4950 static int __init perf_counter_sysfs_init(void)
4952 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4953 &perfclass_attr_group);
4955 device_initcall(perf_counter_sysfs_init);