2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_counters __read_mostly;
44 static atomic_t nr_comm_counters __read_mostly;
47 * perf counter paranoia level:
49 * 1 - disallow cpu counters to unpriv
50 * 2 - disallow kernel profiling to unpriv
52 int sysctl_perf_counter_paranoid __read_mostly;
54 static inline bool perf_paranoid_cpu(void)
56 return sysctl_perf_counter_paranoid > 0;
59 static inline bool perf_paranoid_kernel(void)
61 return sysctl_perf_counter_paranoid > 1;
64 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
67 * max perf counter sample rate
69 int sysctl_perf_counter_sample_rate __read_mostly = 100000;
71 static atomic64_t perf_counter_id;
74 * Lock for (sysadmin-configurable) counter reservations:
76 static DEFINE_SPINLOCK(perf_resource_lock);
79 * Architecture provided APIs - weak aliases:
81 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
86 void __weak hw_perf_disable(void) { barrier(); }
87 void __weak hw_perf_enable(void) { barrier(); }
89 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
92 hw_perf_group_sched_in(struct perf_counter *group_leader,
93 struct perf_cpu_context *cpuctx,
94 struct perf_counter_context *ctx, int cpu)
99 void __weak perf_counter_print_debug(void) { }
101 static DEFINE_PER_CPU(int, disable_count);
103 void __perf_disable(void)
105 __get_cpu_var(disable_count)++;
108 bool __perf_enable(void)
110 return !--__get_cpu_var(disable_count);
113 void perf_disable(void)
119 void perf_enable(void)
125 static void get_ctx(struct perf_counter_context *ctx)
127 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
130 static void free_ctx(struct rcu_head *head)
132 struct perf_counter_context *ctx;
134 ctx = container_of(head, struct perf_counter_context, rcu_head);
138 static void put_ctx(struct perf_counter_context *ctx)
140 if (atomic_dec_and_test(&ctx->refcount)) {
142 put_ctx(ctx->parent_ctx);
144 put_task_struct(ctx->task);
145 call_rcu(&ctx->rcu_head, free_ctx);
149 static void unclone_ctx(struct perf_counter_context *ctx)
151 if (ctx->parent_ctx) {
152 put_ctx(ctx->parent_ctx);
153 ctx->parent_ctx = NULL;
158 * If we inherit counters we want to return the parent counter id
161 static u64 primary_counter_id(struct perf_counter *counter)
163 u64 id = counter->id;
166 id = counter->parent->id;
172 * Get the perf_counter_context for a task and lock it.
173 * This has to cope with with the fact that until it is locked,
174 * the context could get moved to another task.
176 static struct perf_counter_context *
177 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
179 struct perf_counter_context *ctx;
183 ctx = rcu_dereference(task->perf_counter_ctxp);
186 * If this context is a clone of another, it might
187 * get swapped for another underneath us by
188 * perf_counter_task_sched_out, though the
189 * rcu_read_lock() protects us from any context
190 * getting freed. Lock the context and check if it
191 * got swapped before we could get the lock, and retry
192 * if so. If we locked the right context, then it
193 * can't get swapped on us any more.
195 spin_lock_irqsave(&ctx->lock, *flags);
196 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
197 spin_unlock_irqrestore(&ctx->lock, *flags);
201 if (!atomic_inc_not_zero(&ctx->refcount)) {
202 spin_unlock_irqrestore(&ctx->lock, *flags);
211 * Get the context for a task and increment its pin_count so it
212 * can't get swapped to another task. This also increments its
213 * reference count so that the context can't get freed.
215 static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
217 struct perf_counter_context *ctx;
220 ctx = perf_lock_task_context(task, &flags);
223 spin_unlock_irqrestore(&ctx->lock, flags);
228 static void perf_unpin_context(struct perf_counter_context *ctx)
232 spin_lock_irqsave(&ctx->lock, flags);
234 spin_unlock_irqrestore(&ctx->lock, flags);
239 * Add a counter from the lists for its context.
240 * Must be called with ctx->mutex and ctx->lock held.
243 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
245 struct perf_counter *group_leader = counter->group_leader;
248 * Depending on whether it is a standalone or sibling counter,
249 * add it straight to the context's counter list, or to the group
250 * leader's sibling list:
252 if (group_leader == counter)
253 list_add_tail(&counter->list_entry, &ctx->counter_list);
255 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
256 group_leader->nr_siblings++;
259 list_add_rcu(&counter->event_entry, &ctx->event_list);
261 if (counter->attr.inherit_stat)
266 * Remove a counter from the lists for its context.
267 * Must be called with ctx->mutex and ctx->lock held.
270 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
272 struct perf_counter *sibling, *tmp;
274 if (list_empty(&counter->list_entry))
277 if (counter->attr.inherit_stat)
280 list_del_init(&counter->list_entry);
281 list_del_rcu(&counter->event_entry);
283 if (counter->group_leader != counter)
284 counter->group_leader->nr_siblings--;
287 * If this was a group counter with sibling counters then
288 * upgrade the siblings to singleton counters by adding them
289 * to the context list directly:
291 list_for_each_entry_safe(sibling, tmp,
292 &counter->sibling_list, list_entry) {
294 list_move_tail(&sibling->list_entry, &ctx->counter_list);
295 sibling->group_leader = sibling;
300 counter_sched_out(struct perf_counter *counter,
301 struct perf_cpu_context *cpuctx,
302 struct perf_counter_context *ctx)
304 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
307 counter->state = PERF_COUNTER_STATE_INACTIVE;
308 counter->tstamp_stopped = ctx->time;
309 counter->pmu->disable(counter);
312 if (!is_software_counter(counter))
313 cpuctx->active_oncpu--;
315 if (counter->attr.exclusive || !cpuctx->active_oncpu)
316 cpuctx->exclusive = 0;
320 group_sched_out(struct perf_counter *group_counter,
321 struct perf_cpu_context *cpuctx,
322 struct perf_counter_context *ctx)
324 struct perf_counter *counter;
326 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
329 counter_sched_out(group_counter, cpuctx, ctx);
332 * Schedule out siblings (if any):
334 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
335 counter_sched_out(counter, cpuctx, ctx);
337 if (group_counter->attr.exclusive)
338 cpuctx->exclusive = 0;
342 * Cross CPU call to remove a performance counter
344 * We disable the counter on the hardware level first. After that we
345 * remove it from the context list.
347 static void __perf_counter_remove_from_context(void *info)
349 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
350 struct perf_counter *counter = info;
351 struct perf_counter_context *ctx = counter->ctx;
354 * If this is a task context, we need to check whether it is
355 * the current task context of this cpu. If not it has been
356 * scheduled out before the smp call arrived.
358 if (ctx->task && cpuctx->task_ctx != ctx)
361 spin_lock(&ctx->lock);
363 * Protect the list operation against NMI by disabling the
364 * counters on a global level.
368 counter_sched_out(counter, cpuctx, ctx);
370 list_del_counter(counter, ctx);
374 * Allow more per task counters with respect to the
377 cpuctx->max_pertask =
378 min(perf_max_counters - ctx->nr_counters,
379 perf_max_counters - perf_reserved_percpu);
383 spin_unlock(&ctx->lock);
388 * Remove the counter from a task's (or a CPU's) list of counters.
390 * Must be called with ctx->mutex held.
392 * CPU counters are removed with a smp call. For task counters we only
393 * call when the task is on a CPU.
395 * If counter->ctx is a cloned context, callers must make sure that
396 * every task struct that counter->ctx->task could possibly point to
397 * remains valid. This is OK when called from perf_release since
398 * that only calls us on the top-level context, which can't be a clone.
399 * When called from perf_counter_exit_task, it's OK because the
400 * context has been detached from its task.
402 static void perf_counter_remove_from_context(struct perf_counter *counter)
404 struct perf_counter_context *ctx = counter->ctx;
405 struct task_struct *task = ctx->task;
409 * Per cpu counters are removed via an smp call and
410 * the removal is always sucessful.
412 smp_call_function_single(counter->cpu,
413 __perf_counter_remove_from_context,
419 task_oncpu_function_call(task, __perf_counter_remove_from_context,
422 spin_lock_irq(&ctx->lock);
424 * If the context is active we need to retry the smp call.
426 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
427 spin_unlock_irq(&ctx->lock);
432 * The lock prevents that this context is scheduled in so we
433 * can remove the counter safely, if the call above did not
436 if (!list_empty(&counter->list_entry)) {
437 list_del_counter(counter, ctx);
439 spin_unlock_irq(&ctx->lock);
442 static inline u64 perf_clock(void)
444 return cpu_clock(smp_processor_id());
448 * Update the record of the current time in a context.
450 static void update_context_time(struct perf_counter_context *ctx)
452 u64 now = perf_clock();
454 ctx->time += now - ctx->timestamp;
455 ctx->timestamp = now;
459 * Update the total_time_enabled and total_time_running fields for a counter.
461 static void update_counter_times(struct perf_counter *counter)
463 struct perf_counter_context *ctx = counter->ctx;
466 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
469 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
471 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
472 run_end = counter->tstamp_stopped;
476 counter->total_time_running = run_end - counter->tstamp_running;
480 * Update total_time_enabled and total_time_running for all counters in a group.
482 static void update_group_times(struct perf_counter *leader)
484 struct perf_counter *counter;
486 update_counter_times(leader);
487 list_for_each_entry(counter, &leader->sibling_list, list_entry)
488 update_counter_times(counter);
492 * Cross CPU call to disable a performance counter
494 static void __perf_counter_disable(void *info)
496 struct perf_counter *counter = info;
497 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
498 struct perf_counter_context *ctx = counter->ctx;
501 * If this is a per-task counter, need to check whether this
502 * counter's task is the current task on this cpu.
504 if (ctx->task && cpuctx->task_ctx != ctx)
507 spin_lock(&ctx->lock);
510 * If the counter is on, turn it off.
511 * If it is in error state, leave it in error state.
513 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
514 update_context_time(ctx);
515 update_counter_times(counter);
516 if (counter == counter->group_leader)
517 group_sched_out(counter, cpuctx, ctx);
519 counter_sched_out(counter, cpuctx, ctx);
520 counter->state = PERF_COUNTER_STATE_OFF;
523 spin_unlock(&ctx->lock);
529 * If counter->ctx is a cloned context, callers must make sure that
530 * every task struct that counter->ctx->task could possibly point to
531 * remains valid. This condition is satisifed when called through
532 * perf_counter_for_each_child or perf_counter_for_each because they
533 * hold the top-level counter's child_mutex, so any descendant that
534 * goes to exit will block in sync_child_counter.
535 * When called from perf_pending_counter it's OK because counter->ctx
536 * is the current context on this CPU and preemption is disabled,
537 * hence we can't get into perf_counter_task_sched_out for this context.
539 static void perf_counter_disable(struct perf_counter *counter)
541 struct perf_counter_context *ctx = counter->ctx;
542 struct task_struct *task = ctx->task;
546 * Disable the counter on the cpu that it's on
548 smp_call_function_single(counter->cpu, __perf_counter_disable,
554 task_oncpu_function_call(task, __perf_counter_disable, counter);
556 spin_lock_irq(&ctx->lock);
558 * If the counter is still active, we need to retry the cross-call.
560 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
561 spin_unlock_irq(&ctx->lock);
566 * Since we have the lock this context can't be scheduled
567 * in, so we can change the state safely.
569 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
570 update_counter_times(counter);
571 counter->state = PERF_COUNTER_STATE_OFF;
574 spin_unlock_irq(&ctx->lock);
578 counter_sched_in(struct perf_counter *counter,
579 struct perf_cpu_context *cpuctx,
580 struct perf_counter_context *ctx,
583 if (counter->state <= PERF_COUNTER_STATE_OFF)
586 counter->state = PERF_COUNTER_STATE_ACTIVE;
587 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
589 * The new state must be visible before we turn it on in the hardware:
593 if (counter->pmu->enable(counter)) {
594 counter->state = PERF_COUNTER_STATE_INACTIVE;
599 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
601 if (!is_software_counter(counter))
602 cpuctx->active_oncpu++;
605 if (counter->attr.exclusive)
606 cpuctx->exclusive = 1;
612 group_sched_in(struct perf_counter *group_counter,
613 struct perf_cpu_context *cpuctx,
614 struct perf_counter_context *ctx,
617 struct perf_counter *counter, *partial_group;
620 if (group_counter->state == PERF_COUNTER_STATE_OFF)
623 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
625 return ret < 0 ? ret : 0;
627 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
631 * Schedule in siblings as one group (if any):
633 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
634 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
635 partial_group = counter;
644 * Groups can be scheduled in as one unit only, so undo any
645 * partial group before returning:
647 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
648 if (counter == partial_group)
650 counter_sched_out(counter, cpuctx, ctx);
652 counter_sched_out(group_counter, cpuctx, ctx);
658 * Return 1 for a group consisting entirely of software counters,
659 * 0 if the group contains any hardware counters.
661 static int is_software_only_group(struct perf_counter *leader)
663 struct perf_counter *counter;
665 if (!is_software_counter(leader))
668 list_for_each_entry(counter, &leader->sibling_list, list_entry)
669 if (!is_software_counter(counter))
676 * Work out whether we can put this counter group on the CPU now.
678 static int group_can_go_on(struct perf_counter *counter,
679 struct perf_cpu_context *cpuctx,
683 * Groups consisting entirely of software counters can always go on.
685 if (is_software_only_group(counter))
688 * If an exclusive group is already on, no other hardware
689 * counters can go on.
691 if (cpuctx->exclusive)
694 * If this group is exclusive and there are already
695 * counters on the CPU, it can't go on.
697 if (counter->attr.exclusive && cpuctx->active_oncpu)
700 * Otherwise, try to add it if all previous groups were able
706 static void add_counter_to_ctx(struct perf_counter *counter,
707 struct perf_counter_context *ctx)
709 list_add_counter(counter, ctx);
710 counter->tstamp_enabled = ctx->time;
711 counter->tstamp_running = ctx->time;
712 counter->tstamp_stopped = ctx->time;
716 * Cross CPU call to install and enable a performance counter
718 * Must be called with ctx->mutex held
720 static void __perf_install_in_context(void *info)
722 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
723 struct perf_counter *counter = info;
724 struct perf_counter_context *ctx = counter->ctx;
725 struct perf_counter *leader = counter->group_leader;
726 int cpu = smp_processor_id();
730 * If this is a task context, we need to check whether it is
731 * the current task context of this cpu. If not it has been
732 * scheduled out before the smp call arrived.
733 * Or possibly this is the right context but it isn't
734 * on this cpu because it had no counters.
736 if (ctx->task && cpuctx->task_ctx != ctx) {
737 if (cpuctx->task_ctx || ctx->task != current)
739 cpuctx->task_ctx = ctx;
742 spin_lock(&ctx->lock);
744 update_context_time(ctx);
747 * Protect the list operation against NMI by disabling the
748 * counters on a global level. NOP for non NMI based counters.
752 add_counter_to_ctx(counter, ctx);
755 * Don't put the counter on if it is disabled or if
756 * it is in a group and the group isn't on.
758 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
759 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
763 * An exclusive counter can't go on if there are already active
764 * hardware counters, and no hardware counter can go on if there
765 * is already an exclusive counter on.
767 if (!group_can_go_on(counter, cpuctx, 1))
770 err = counter_sched_in(counter, cpuctx, ctx, cpu);
774 * This counter couldn't go on. If it is in a group
775 * then we have to pull the whole group off.
776 * If the counter group is pinned then put it in error state.
778 if (leader != counter)
779 group_sched_out(leader, cpuctx, ctx);
780 if (leader->attr.pinned) {
781 update_group_times(leader);
782 leader->state = PERF_COUNTER_STATE_ERROR;
786 if (!err && !ctx->task && cpuctx->max_pertask)
787 cpuctx->max_pertask--;
792 spin_unlock(&ctx->lock);
796 * Attach a performance counter to a context
798 * First we add the counter to the list with the hardware enable bit
799 * in counter->hw_config cleared.
801 * If the counter is attached to a task which is on a CPU we use a smp
802 * call to enable it in the task context. The task might have been
803 * scheduled away, but we check this in the smp call again.
805 * Must be called with ctx->mutex held.
808 perf_install_in_context(struct perf_counter_context *ctx,
809 struct perf_counter *counter,
812 struct task_struct *task = ctx->task;
816 * Per cpu counters are installed via an smp call and
817 * the install is always sucessful.
819 smp_call_function_single(cpu, __perf_install_in_context,
825 task_oncpu_function_call(task, __perf_install_in_context,
828 spin_lock_irq(&ctx->lock);
830 * we need to retry the smp call.
832 if (ctx->is_active && list_empty(&counter->list_entry)) {
833 spin_unlock_irq(&ctx->lock);
838 * The lock prevents that this context is scheduled in so we
839 * can add the counter safely, if it the call above did not
842 if (list_empty(&counter->list_entry))
843 add_counter_to_ctx(counter, ctx);
844 spin_unlock_irq(&ctx->lock);
848 * Cross CPU call to enable a performance counter
850 static void __perf_counter_enable(void *info)
852 struct perf_counter *counter = info;
853 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
854 struct perf_counter_context *ctx = counter->ctx;
855 struct perf_counter *leader = counter->group_leader;
859 * If this is a per-task counter, need to check whether this
860 * counter's task is the current task on this cpu.
862 if (ctx->task && cpuctx->task_ctx != ctx) {
863 if (cpuctx->task_ctx || ctx->task != current)
865 cpuctx->task_ctx = ctx;
868 spin_lock(&ctx->lock);
870 update_context_time(ctx);
872 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
874 counter->state = PERF_COUNTER_STATE_INACTIVE;
875 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
878 * If the counter is in a group and isn't the group leader,
879 * then don't put it on unless the group is on.
881 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
884 if (!group_can_go_on(counter, cpuctx, 1)) {
888 if (counter == leader)
889 err = group_sched_in(counter, cpuctx, ctx,
892 err = counter_sched_in(counter, cpuctx, ctx,
899 * If this counter can't go on and it's part of a
900 * group, then the whole group has to come off.
902 if (leader != counter)
903 group_sched_out(leader, cpuctx, ctx);
904 if (leader->attr.pinned) {
905 update_group_times(leader);
906 leader->state = PERF_COUNTER_STATE_ERROR;
911 spin_unlock(&ctx->lock);
917 * If counter->ctx is a cloned context, callers must make sure that
918 * every task struct that counter->ctx->task could possibly point to
919 * remains valid. This condition is satisfied when called through
920 * perf_counter_for_each_child or perf_counter_for_each as described
921 * for perf_counter_disable.
923 static void perf_counter_enable(struct perf_counter *counter)
925 struct perf_counter_context *ctx = counter->ctx;
926 struct task_struct *task = ctx->task;
930 * Enable the counter on the cpu that it's on
932 smp_call_function_single(counter->cpu, __perf_counter_enable,
937 spin_lock_irq(&ctx->lock);
938 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
942 * If the counter is in error state, clear that first.
943 * That way, if we see the counter in error state below, we
944 * know that it has gone back into error state, as distinct
945 * from the task having been scheduled away before the
946 * cross-call arrived.
948 if (counter->state == PERF_COUNTER_STATE_ERROR)
949 counter->state = PERF_COUNTER_STATE_OFF;
952 spin_unlock_irq(&ctx->lock);
953 task_oncpu_function_call(task, __perf_counter_enable, counter);
955 spin_lock_irq(&ctx->lock);
958 * If the context is active and the counter is still off,
959 * we need to retry the cross-call.
961 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
965 * Since we have the lock this context can't be scheduled
966 * in, so we can change the state safely.
968 if (counter->state == PERF_COUNTER_STATE_OFF) {
969 counter->state = PERF_COUNTER_STATE_INACTIVE;
970 counter->tstamp_enabled =
971 ctx->time - counter->total_time_enabled;
974 spin_unlock_irq(&ctx->lock);
977 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
980 * not supported on inherited counters
982 if (counter->attr.inherit)
985 atomic_add(refresh, &counter->event_limit);
986 perf_counter_enable(counter);
991 void __perf_counter_sched_out(struct perf_counter_context *ctx,
992 struct perf_cpu_context *cpuctx)
994 struct perf_counter *counter;
996 spin_lock(&ctx->lock);
998 if (likely(!ctx->nr_counters))
1000 update_context_time(ctx);
1003 if (ctx->nr_active) {
1004 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1005 if (counter != counter->group_leader)
1006 counter_sched_out(counter, cpuctx, ctx);
1008 group_sched_out(counter, cpuctx, ctx);
1013 spin_unlock(&ctx->lock);
1017 * Test whether two contexts are equivalent, i.e. whether they
1018 * have both been cloned from the same version of the same context
1019 * and they both have the same number of enabled counters.
1020 * If the number of enabled counters is the same, then the set
1021 * of enabled counters should be the same, because these are both
1022 * inherited contexts, therefore we can't access individual counters
1023 * in them directly with an fd; we can only enable/disable all
1024 * counters via prctl, or enable/disable all counters in a family
1025 * via ioctl, which will have the same effect on both contexts.
1027 static int context_equiv(struct perf_counter_context *ctx1,
1028 struct perf_counter_context *ctx2)
1030 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1031 && ctx1->parent_gen == ctx2->parent_gen
1032 && !ctx1->pin_count && !ctx2->pin_count;
1035 static void __perf_counter_read(void *counter);
1037 static void __perf_counter_sync_stat(struct perf_counter *counter,
1038 struct perf_counter *next_counter)
1042 if (!counter->attr.inherit_stat)
1046 * Update the counter value, we cannot use perf_counter_read()
1047 * because we're in the middle of a context switch and have IRQs
1048 * disabled, which upsets smp_call_function_single(), however
1049 * we know the counter must be on the current CPU, therefore we
1050 * don't need to use it.
1052 switch (counter->state) {
1053 case PERF_COUNTER_STATE_ACTIVE:
1054 __perf_counter_read(counter);
1057 case PERF_COUNTER_STATE_INACTIVE:
1058 update_counter_times(counter);
1066 * In order to keep per-task stats reliable we need to flip the counter
1067 * values when we flip the contexts.
1069 value = atomic64_read(&next_counter->count);
1070 value = atomic64_xchg(&counter->count, value);
1071 atomic64_set(&next_counter->count, value);
1073 swap(counter->total_time_enabled, next_counter->total_time_enabled);
1074 swap(counter->total_time_running, next_counter->total_time_running);
1077 * Since we swizzled the values, update the user visible data too.
1079 perf_counter_update_userpage(counter);
1080 perf_counter_update_userpage(next_counter);
1083 #define list_next_entry(pos, member) \
1084 list_entry(pos->member.next, typeof(*pos), member)
1086 static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1087 struct perf_counter_context *next_ctx)
1089 struct perf_counter *counter, *next_counter;
1094 counter = list_first_entry(&ctx->event_list,
1095 struct perf_counter, event_entry);
1097 next_counter = list_first_entry(&next_ctx->event_list,
1098 struct perf_counter, event_entry);
1100 while (&counter->event_entry != &ctx->event_list &&
1101 &next_counter->event_entry != &next_ctx->event_list) {
1103 __perf_counter_sync_stat(counter, next_counter);
1105 counter = list_next_entry(counter, event_entry);
1106 next_counter = list_next_entry(counter, event_entry);
1111 * Called from scheduler to remove the counters of the current task,
1112 * with interrupts disabled.
1114 * We stop each counter and update the counter value in counter->count.
1116 * This does not protect us against NMI, but disable()
1117 * sets the disabled bit in the control field of counter _before_
1118 * accessing the counter control register. If a NMI hits, then it will
1119 * not restart the counter.
1121 void perf_counter_task_sched_out(struct task_struct *task,
1122 struct task_struct *next, int cpu)
1124 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1125 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1126 struct perf_counter_context *next_ctx;
1127 struct perf_counter_context *parent;
1128 struct pt_regs *regs;
1131 regs = task_pt_regs(task);
1132 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1134 if (likely(!ctx || !cpuctx->task_ctx))
1137 update_context_time(ctx);
1140 parent = rcu_dereference(ctx->parent_ctx);
1141 next_ctx = next->perf_counter_ctxp;
1142 if (parent && next_ctx &&
1143 rcu_dereference(next_ctx->parent_ctx) == parent) {
1145 * Looks like the two contexts are clones, so we might be
1146 * able to optimize the context switch. We lock both
1147 * contexts and check that they are clones under the
1148 * lock (including re-checking that neither has been
1149 * uncloned in the meantime). It doesn't matter which
1150 * order we take the locks because no other cpu could
1151 * be trying to lock both of these tasks.
1153 spin_lock(&ctx->lock);
1154 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1155 if (context_equiv(ctx, next_ctx)) {
1157 * XXX do we need a memory barrier of sorts
1158 * wrt to rcu_dereference() of perf_counter_ctxp
1160 task->perf_counter_ctxp = next_ctx;
1161 next->perf_counter_ctxp = ctx;
1163 next_ctx->task = task;
1166 perf_counter_sync_stat(ctx, next_ctx);
1168 spin_unlock(&next_ctx->lock);
1169 spin_unlock(&ctx->lock);
1174 __perf_counter_sched_out(ctx, cpuctx);
1175 cpuctx->task_ctx = NULL;
1180 * Called with IRQs disabled
1182 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1184 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1186 if (!cpuctx->task_ctx)
1189 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1192 __perf_counter_sched_out(ctx, cpuctx);
1193 cpuctx->task_ctx = NULL;
1197 * Called with IRQs disabled
1199 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1201 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1205 __perf_counter_sched_in(struct perf_counter_context *ctx,
1206 struct perf_cpu_context *cpuctx, int cpu)
1208 struct perf_counter *counter;
1211 spin_lock(&ctx->lock);
1213 if (likely(!ctx->nr_counters))
1216 ctx->timestamp = perf_clock();
1221 * First go through the list and put on any pinned groups
1222 * in order to give them the best chance of going on.
1224 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1225 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1226 !counter->attr.pinned)
1228 if (counter->cpu != -1 && counter->cpu != cpu)
1231 if (counter != counter->group_leader)
1232 counter_sched_in(counter, cpuctx, ctx, cpu);
1234 if (group_can_go_on(counter, cpuctx, 1))
1235 group_sched_in(counter, cpuctx, ctx, cpu);
1239 * If this pinned group hasn't been scheduled,
1240 * put it in error state.
1242 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1243 update_group_times(counter);
1244 counter->state = PERF_COUNTER_STATE_ERROR;
1248 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1250 * Ignore counters in OFF or ERROR state, and
1251 * ignore pinned counters since we did them already.
1253 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1254 counter->attr.pinned)
1258 * Listen to the 'cpu' scheduling filter constraint
1261 if (counter->cpu != -1 && counter->cpu != cpu)
1264 if (counter != counter->group_leader) {
1265 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1268 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1269 if (group_sched_in(counter, cpuctx, ctx, cpu))
1276 spin_unlock(&ctx->lock);
1280 * Called from scheduler to add the counters of the current task
1281 * with interrupts disabled.
1283 * We restore the counter value and then enable it.
1285 * This does not protect us against NMI, but enable()
1286 * sets the enabled bit in the control field of counter _before_
1287 * accessing the counter control register. If a NMI hits, then it will
1288 * keep the counter running.
1290 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1292 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1293 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1297 if (cpuctx->task_ctx == ctx)
1299 __perf_counter_sched_in(ctx, cpuctx, cpu);
1300 cpuctx->task_ctx = ctx;
1303 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1305 struct perf_counter_context *ctx = &cpuctx->ctx;
1307 __perf_counter_sched_in(ctx, cpuctx, cpu);
1310 #define MAX_INTERRUPTS (~0ULL)
1312 static void perf_log_throttle(struct perf_counter *counter, int enable);
1314 static void perf_adjust_period(struct perf_counter *counter, u64 events)
1316 struct hw_perf_counter *hwc = &counter->hw;
1317 u64 period, sample_period;
1320 events *= hwc->sample_period;
1321 period = div64_u64(events, counter->attr.sample_freq);
1323 delta = (s64)(period - hwc->sample_period);
1324 delta = (delta + 7) / 8; /* low pass filter */
1326 sample_period = hwc->sample_period + delta;
1331 hwc->sample_period = sample_period;
1334 static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1336 struct perf_counter *counter;
1337 struct hw_perf_counter *hwc;
1338 u64 interrupts, freq;
1340 spin_lock(&ctx->lock);
1341 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1342 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1347 interrupts = hwc->interrupts;
1348 hwc->interrupts = 0;
1351 * unthrottle counters on the tick
1353 if (interrupts == MAX_INTERRUPTS) {
1354 perf_log_throttle(counter, 1);
1355 counter->pmu->unthrottle(counter);
1356 interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1359 if (!counter->attr.freq || !counter->attr.sample_freq)
1363 * if the specified freq < HZ then we need to skip ticks
1365 if (counter->attr.sample_freq < HZ) {
1366 freq = counter->attr.sample_freq;
1368 hwc->freq_count += freq;
1369 hwc->freq_interrupts += interrupts;
1371 if (hwc->freq_count < HZ)
1374 interrupts = hwc->freq_interrupts;
1375 hwc->freq_interrupts = 0;
1376 hwc->freq_count -= HZ;
1380 perf_adjust_period(counter, freq * interrupts);
1383 * In order to avoid being stalled by an (accidental) huge
1384 * sample period, force reset the sample period if we didn't
1385 * get any events in this freq period.
1389 counter->pmu->disable(counter);
1390 atomic64_set(&hwc->period_left, 0);
1391 counter->pmu->enable(counter);
1395 spin_unlock(&ctx->lock);
1399 * Round-robin a context's counters:
1401 static void rotate_ctx(struct perf_counter_context *ctx)
1403 struct perf_counter *counter;
1405 if (!ctx->nr_counters)
1408 spin_lock(&ctx->lock);
1410 * Rotate the first entry last (works just fine for group counters too):
1413 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1414 list_move_tail(&counter->list_entry, &ctx->counter_list);
1419 spin_unlock(&ctx->lock);
1422 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1424 struct perf_cpu_context *cpuctx;
1425 struct perf_counter_context *ctx;
1427 if (!atomic_read(&nr_counters))
1430 cpuctx = &per_cpu(perf_cpu_context, cpu);
1431 ctx = curr->perf_counter_ctxp;
1433 perf_ctx_adjust_freq(&cpuctx->ctx);
1435 perf_ctx_adjust_freq(ctx);
1437 perf_counter_cpu_sched_out(cpuctx);
1439 __perf_counter_task_sched_out(ctx);
1441 rotate_ctx(&cpuctx->ctx);
1445 perf_counter_cpu_sched_in(cpuctx, cpu);
1447 perf_counter_task_sched_in(curr, cpu);
1451 * Enable all of a task's counters that have been marked enable-on-exec.
1452 * This expects task == current.
1454 static void perf_counter_enable_on_exec(struct task_struct *task)
1456 struct perf_counter_context *ctx;
1457 struct perf_counter *counter;
1458 unsigned long flags;
1461 local_irq_save(flags);
1462 ctx = task->perf_counter_ctxp;
1463 if (!ctx || !ctx->nr_counters)
1466 __perf_counter_task_sched_out(ctx);
1468 spin_lock(&ctx->lock);
1470 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1471 if (!counter->attr.enable_on_exec)
1473 counter->attr.enable_on_exec = 0;
1474 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
1476 counter->state = PERF_COUNTER_STATE_INACTIVE;
1477 counter->tstamp_enabled =
1478 ctx->time - counter->total_time_enabled;
1483 * Unclone this context if we enabled any counter.
1488 spin_unlock(&ctx->lock);
1490 perf_counter_task_sched_in(task, smp_processor_id());
1492 local_irq_restore(flags);
1496 * Cross CPU call to read the hardware counter
1498 static void __perf_counter_read(void *info)
1500 struct perf_counter *counter = info;
1501 struct perf_counter_context *ctx = counter->ctx;
1502 unsigned long flags;
1504 local_irq_save(flags);
1506 update_context_time(ctx);
1507 counter->pmu->read(counter);
1508 update_counter_times(counter);
1509 local_irq_restore(flags);
1512 static u64 perf_counter_read(struct perf_counter *counter)
1515 * If counter is enabled and currently active on a CPU, update the
1516 * value in the counter structure:
1518 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1519 smp_call_function_single(counter->oncpu,
1520 __perf_counter_read, counter, 1);
1521 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1522 update_counter_times(counter);
1525 return atomic64_read(&counter->count);
1529 * Initialize the perf_counter context in a task_struct:
1532 __perf_counter_init_context(struct perf_counter_context *ctx,
1533 struct task_struct *task)
1535 memset(ctx, 0, sizeof(*ctx));
1536 spin_lock_init(&ctx->lock);
1537 mutex_init(&ctx->mutex);
1538 INIT_LIST_HEAD(&ctx->counter_list);
1539 INIT_LIST_HEAD(&ctx->event_list);
1540 atomic_set(&ctx->refcount, 1);
1544 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1546 struct perf_counter_context *ctx;
1547 struct perf_cpu_context *cpuctx;
1548 struct task_struct *task;
1549 unsigned long flags;
1553 * If cpu is not a wildcard then this is a percpu counter:
1556 /* Must be root to operate on a CPU counter: */
1557 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1558 return ERR_PTR(-EACCES);
1560 if (cpu < 0 || cpu > num_possible_cpus())
1561 return ERR_PTR(-EINVAL);
1564 * We could be clever and allow to attach a counter to an
1565 * offline CPU and activate it when the CPU comes up, but
1568 if (!cpu_isset(cpu, cpu_online_map))
1569 return ERR_PTR(-ENODEV);
1571 cpuctx = &per_cpu(perf_cpu_context, cpu);
1582 task = find_task_by_vpid(pid);
1584 get_task_struct(task);
1588 return ERR_PTR(-ESRCH);
1591 * Can't attach counters to a dying task.
1594 if (task->flags & PF_EXITING)
1597 /* Reuse ptrace permission checks for now. */
1599 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1603 ctx = perf_lock_task_context(task, &flags);
1606 spin_unlock_irqrestore(&ctx->lock, flags);
1610 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1614 __perf_counter_init_context(ctx, task);
1616 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1618 * We raced with some other task; use
1619 * the context they set.
1624 get_task_struct(task);
1627 put_task_struct(task);
1631 put_task_struct(task);
1632 return ERR_PTR(err);
1635 static void free_counter_rcu(struct rcu_head *head)
1637 struct perf_counter *counter;
1639 counter = container_of(head, struct perf_counter, rcu_head);
1641 put_pid_ns(counter->ns);
1645 static void perf_pending_sync(struct perf_counter *counter);
1647 static void free_counter(struct perf_counter *counter)
1649 perf_pending_sync(counter);
1651 if (!counter->parent) {
1652 atomic_dec(&nr_counters);
1653 if (counter->attr.mmap)
1654 atomic_dec(&nr_mmap_counters);
1655 if (counter->attr.comm)
1656 atomic_dec(&nr_comm_counters);
1659 if (counter->destroy)
1660 counter->destroy(counter);
1662 put_ctx(counter->ctx);
1663 call_rcu(&counter->rcu_head, free_counter_rcu);
1667 * Called when the last reference to the file is gone.
1669 static int perf_release(struct inode *inode, struct file *file)
1671 struct perf_counter *counter = file->private_data;
1672 struct perf_counter_context *ctx = counter->ctx;
1674 file->private_data = NULL;
1676 WARN_ON_ONCE(ctx->parent_ctx);
1677 mutex_lock(&ctx->mutex);
1678 perf_counter_remove_from_context(counter);
1679 mutex_unlock(&ctx->mutex);
1681 mutex_lock(&counter->owner->perf_counter_mutex);
1682 list_del_init(&counter->owner_entry);
1683 mutex_unlock(&counter->owner->perf_counter_mutex);
1684 put_task_struct(counter->owner);
1686 free_counter(counter);
1691 static u64 perf_counter_read_tree(struct perf_counter *counter)
1693 struct perf_counter *child;
1696 total += perf_counter_read(counter);
1697 list_for_each_entry(child, &counter->child_list, child_list)
1698 total += perf_counter_read(child);
1704 * Read the performance counter - simple non blocking version for now
1707 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1713 * Return end-of-file for a read on a counter that is in
1714 * error state (i.e. because it was pinned but it couldn't be
1715 * scheduled on to the CPU at some point).
1717 if (counter->state == PERF_COUNTER_STATE_ERROR)
1720 WARN_ON_ONCE(counter->ctx->parent_ctx);
1721 mutex_lock(&counter->child_mutex);
1722 values[0] = perf_counter_read_tree(counter);
1724 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1725 values[n++] = counter->total_time_enabled +
1726 atomic64_read(&counter->child_total_time_enabled);
1727 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1728 values[n++] = counter->total_time_running +
1729 atomic64_read(&counter->child_total_time_running);
1730 if (counter->attr.read_format & PERF_FORMAT_ID)
1731 values[n++] = primary_counter_id(counter);
1732 mutex_unlock(&counter->child_mutex);
1734 if (count < n * sizeof(u64))
1736 count = n * sizeof(u64);
1738 if (copy_to_user(buf, values, count))
1745 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1747 struct perf_counter *counter = file->private_data;
1749 return perf_read_hw(counter, buf, count);
1752 static unsigned int perf_poll(struct file *file, poll_table *wait)
1754 struct perf_counter *counter = file->private_data;
1755 struct perf_mmap_data *data;
1756 unsigned int events = POLL_HUP;
1759 data = rcu_dereference(counter->data);
1761 events = atomic_xchg(&data->poll, 0);
1764 poll_wait(file, &counter->waitq, wait);
1769 static void perf_counter_reset(struct perf_counter *counter)
1771 (void)perf_counter_read(counter);
1772 atomic64_set(&counter->count, 0);
1773 perf_counter_update_userpage(counter);
1777 * Holding the top-level counter's child_mutex means that any
1778 * descendant process that has inherited this counter will block
1779 * in sync_child_counter if it goes to exit, thus satisfying the
1780 * task existence requirements of perf_counter_enable/disable.
1782 static void perf_counter_for_each_child(struct perf_counter *counter,
1783 void (*func)(struct perf_counter *))
1785 struct perf_counter *child;
1787 WARN_ON_ONCE(counter->ctx->parent_ctx);
1788 mutex_lock(&counter->child_mutex);
1790 list_for_each_entry(child, &counter->child_list, child_list)
1792 mutex_unlock(&counter->child_mutex);
1795 static void perf_counter_for_each(struct perf_counter *counter,
1796 void (*func)(struct perf_counter *))
1798 struct perf_counter_context *ctx = counter->ctx;
1799 struct perf_counter *sibling;
1801 WARN_ON_ONCE(ctx->parent_ctx);
1802 mutex_lock(&ctx->mutex);
1803 counter = counter->group_leader;
1805 perf_counter_for_each_child(counter, func);
1807 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1808 perf_counter_for_each_child(counter, func);
1809 mutex_unlock(&ctx->mutex);
1812 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1814 struct perf_counter_context *ctx = counter->ctx;
1819 if (!counter->attr.sample_period)
1822 size = copy_from_user(&value, arg, sizeof(value));
1823 if (size != sizeof(value))
1829 spin_lock_irq(&ctx->lock);
1830 if (counter->attr.freq) {
1831 if (value > sysctl_perf_counter_sample_rate) {
1836 counter->attr.sample_freq = value;
1838 counter->attr.sample_period = value;
1839 counter->hw.sample_period = value;
1842 spin_unlock_irq(&ctx->lock);
1847 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1849 struct perf_counter *counter = file->private_data;
1850 void (*func)(struct perf_counter *);
1854 case PERF_COUNTER_IOC_ENABLE:
1855 func = perf_counter_enable;
1857 case PERF_COUNTER_IOC_DISABLE:
1858 func = perf_counter_disable;
1860 case PERF_COUNTER_IOC_RESET:
1861 func = perf_counter_reset;
1864 case PERF_COUNTER_IOC_REFRESH:
1865 return perf_counter_refresh(counter, arg);
1867 case PERF_COUNTER_IOC_PERIOD:
1868 return perf_counter_period(counter, (u64 __user *)arg);
1874 if (flags & PERF_IOC_FLAG_GROUP)
1875 perf_counter_for_each(counter, func);
1877 perf_counter_for_each_child(counter, func);
1882 int perf_counter_task_enable(void)
1884 struct perf_counter *counter;
1886 mutex_lock(¤t->perf_counter_mutex);
1887 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1888 perf_counter_for_each_child(counter, perf_counter_enable);
1889 mutex_unlock(¤t->perf_counter_mutex);
1894 int perf_counter_task_disable(void)
1896 struct perf_counter *counter;
1898 mutex_lock(¤t->perf_counter_mutex);
1899 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1900 perf_counter_for_each_child(counter, perf_counter_disable);
1901 mutex_unlock(¤t->perf_counter_mutex);
1906 static int perf_counter_index(struct perf_counter *counter)
1908 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1911 return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
1915 * Callers need to ensure there can be no nesting of this function, otherwise
1916 * the seqlock logic goes bad. We can not serialize this because the arch
1917 * code calls this from NMI context.
1919 void perf_counter_update_userpage(struct perf_counter *counter)
1921 struct perf_counter_mmap_page *userpg;
1922 struct perf_mmap_data *data;
1925 data = rcu_dereference(counter->data);
1929 userpg = data->user_page;
1932 * Disable preemption so as to not let the corresponding user-space
1933 * spin too long if we get preempted.
1938 userpg->index = perf_counter_index(counter);
1939 userpg->offset = atomic64_read(&counter->count);
1940 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1941 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1943 userpg->time_enabled = counter->total_time_enabled +
1944 atomic64_read(&counter->child_total_time_enabled);
1946 userpg->time_running = counter->total_time_running +
1947 atomic64_read(&counter->child_total_time_running);
1956 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1958 struct perf_counter *counter = vma->vm_file->private_data;
1959 struct perf_mmap_data *data;
1960 int ret = VM_FAULT_SIGBUS;
1962 if (vmf->flags & FAULT_FLAG_MKWRITE) {
1963 if (vmf->pgoff == 0)
1969 data = rcu_dereference(counter->data);
1973 if (vmf->pgoff == 0) {
1974 vmf->page = virt_to_page(data->user_page);
1976 int nr = vmf->pgoff - 1;
1978 if ((unsigned)nr > data->nr_pages)
1981 if (vmf->flags & FAULT_FLAG_WRITE)
1984 vmf->page = virt_to_page(data->data_pages[nr]);
1987 get_page(vmf->page);
1988 vmf->page->mapping = vma->vm_file->f_mapping;
1989 vmf->page->index = vmf->pgoff;
1998 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
2000 struct perf_mmap_data *data;
2004 WARN_ON(atomic_read(&counter->mmap_count));
2006 size = sizeof(struct perf_mmap_data);
2007 size += nr_pages * sizeof(void *);
2009 data = kzalloc(size, GFP_KERNEL);
2013 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2014 if (!data->user_page)
2015 goto fail_user_page;
2017 for (i = 0; i < nr_pages; i++) {
2018 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2019 if (!data->data_pages[i])
2020 goto fail_data_pages;
2023 data->nr_pages = nr_pages;
2024 atomic_set(&data->lock, -1);
2026 rcu_assign_pointer(counter->data, data);
2031 for (i--; i >= 0; i--)
2032 free_page((unsigned long)data->data_pages[i]);
2034 free_page((unsigned long)data->user_page);
2043 static void perf_mmap_free_page(unsigned long addr)
2045 struct page *page = virt_to_page((void *)addr);
2047 page->mapping = NULL;
2051 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2053 struct perf_mmap_data *data;
2056 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2058 perf_mmap_free_page((unsigned long)data->user_page);
2059 for (i = 0; i < data->nr_pages; i++)
2060 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2065 static void perf_mmap_data_free(struct perf_counter *counter)
2067 struct perf_mmap_data *data = counter->data;
2069 WARN_ON(atomic_read(&counter->mmap_count));
2071 rcu_assign_pointer(counter->data, NULL);
2072 call_rcu(&data->rcu_head, __perf_mmap_data_free);
2075 static void perf_mmap_open(struct vm_area_struct *vma)
2077 struct perf_counter *counter = vma->vm_file->private_data;
2079 atomic_inc(&counter->mmap_count);
2082 static void perf_mmap_close(struct vm_area_struct *vma)
2084 struct perf_counter *counter = vma->vm_file->private_data;
2086 WARN_ON_ONCE(counter->ctx->parent_ctx);
2087 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2088 struct user_struct *user = current_user();
2090 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2091 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2092 perf_mmap_data_free(counter);
2093 mutex_unlock(&counter->mmap_mutex);
2097 static struct vm_operations_struct perf_mmap_vmops = {
2098 .open = perf_mmap_open,
2099 .close = perf_mmap_close,
2100 .fault = perf_mmap_fault,
2101 .page_mkwrite = perf_mmap_fault,
2104 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2106 struct perf_counter *counter = file->private_data;
2107 unsigned long user_locked, user_lock_limit;
2108 struct user_struct *user = current_user();
2109 unsigned long locked, lock_limit;
2110 unsigned long vma_size;
2111 unsigned long nr_pages;
2112 long user_extra, extra;
2115 if (!(vma->vm_flags & VM_SHARED))
2118 vma_size = vma->vm_end - vma->vm_start;
2119 nr_pages = (vma_size / PAGE_SIZE) - 1;
2122 * If we have data pages ensure they're a power-of-two number, so we
2123 * can do bitmasks instead of modulo.
2125 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2128 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2131 if (vma->vm_pgoff != 0)
2134 WARN_ON_ONCE(counter->ctx->parent_ctx);
2135 mutex_lock(&counter->mmap_mutex);
2136 if (atomic_inc_not_zero(&counter->mmap_count)) {
2137 if (nr_pages != counter->data->nr_pages)
2142 user_extra = nr_pages + 1;
2143 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2146 * Increase the limit linearly with more CPUs:
2148 user_lock_limit *= num_online_cpus();
2150 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2153 if (user_locked > user_lock_limit)
2154 extra = user_locked - user_lock_limit;
2156 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2157 lock_limit >>= PAGE_SHIFT;
2158 locked = vma->vm_mm->locked_vm + extra;
2160 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2165 WARN_ON(counter->data);
2166 ret = perf_mmap_data_alloc(counter, nr_pages);
2170 atomic_set(&counter->mmap_count, 1);
2171 atomic_long_add(user_extra, &user->locked_vm);
2172 vma->vm_mm->locked_vm += extra;
2173 counter->data->nr_locked = extra;
2174 if (vma->vm_flags & VM_WRITE)
2175 counter->data->writable = 1;
2178 mutex_unlock(&counter->mmap_mutex);
2180 vma->vm_flags |= VM_RESERVED;
2181 vma->vm_ops = &perf_mmap_vmops;
2186 static int perf_fasync(int fd, struct file *filp, int on)
2188 struct inode *inode = filp->f_path.dentry->d_inode;
2189 struct perf_counter *counter = filp->private_data;
2192 mutex_lock(&inode->i_mutex);
2193 retval = fasync_helper(fd, filp, on, &counter->fasync);
2194 mutex_unlock(&inode->i_mutex);
2202 static const struct file_operations perf_fops = {
2203 .release = perf_release,
2206 .unlocked_ioctl = perf_ioctl,
2207 .compat_ioctl = perf_ioctl,
2209 .fasync = perf_fasync,
2213 * Perf counter wakeup
2215 * If there's data, ensure we set the poll() state and publish everything
2216 * to user-space before waking everybody up.
2219 void perf_counter_wakeup(struct perf_counter *counter)
2221 wake_up_all(&counter->waitq);
2223 if (counter->pending_kill) {
2224 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2225 counter->pending_kill = 0;
2232 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2234 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2235 * single linked list and use cmpxchg() to add entries lockless.
2238 static void perf_pending_counter(struct perf_pending_entry *entry)
2240 struct perf_counter *counter = container_of(entry,
2241 struct perf_counter, pending);
2243 if (counter->pending_disable) {
2244 counter->pending_disable = 0;
2245 perf_counter_disable(counter);
2248 if (counter->pending_wakeup) {
2249 counter->pending_wakeup = 0;
2250 perf_counter_wakeup(counter);
2254 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2256 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2260 static void perf_pending_queue(struct perf_pending_entry *entry,
2261 void (*func)(struct perf_pending_entry *))
2263 struct perf_pending_entry **head;
2265 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2270 head = &get_cpu_var(perf_pending_head);
2273 entry->next = *head;
2274 } while (cmpxchg(head, entry->next, entry) != entry->next);
2276 set_perf_counter_pending();
2278 put_cpu_var(perf_pending_head);
2281 static int __perf_pending_run(void)
2283 struct perf_pending_entry *list;
2286 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2287 while (list != PENDING_TAIL) {
2288 void (*func)(struct perf_pending_entry *);
2289 struct perf_pending_entry *entry = list;
2296 * Ensure we observe the unqueue before we issue the wakeup,
2297 * so that we won't be waiting forever.
2298 * -- see perf_not_pending().
2309 static inline int perf_not_pending(struct perf_counter *counter)
2312 * If we flush on whatever cpu we run, there is a chance we don't
2316 __perf_pending_run();
2320 * Ensure we see the proper queue state before going to sleep
2321 * so that we do not miss the wakeup. -- see perf_pending_handle()
2324 return counter->pending.next == NULL;
2327 static void perf_pending_sync(struct perf_counter *counter)
2329 wait_event(counter->waitq, perf_not_pending(counter));
2332 void perf_counter_do_pending(void)
2334 __perf_pending_run();
2338 * Callchain support -- arch specific
2341 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2350 struct perf_output_handle {
2351 struct perf_counter *counter;
2352 struct perf_mmap_data *data;
2354 unsigned long offset;
2358 unsigned long flags;
2361 static bool perf_output_space(struct perf_mmap_data *data,
2362 unsigned int offset, unsigned int head)
2367 if (!data->writable)
2370 mask = (data->nr_pages << PAGE_SHIFT) - 1;
2372 * Userspace could choose to issue a mb() before updating the tail
2373 * pointer. So that all reads will be completed before the write is
2376 tail = ACCESS_ONCE(data->user_page->data_tail);
2379 offset = (offset - tail) & mask;
2380 head = (head - tail) & mask;
2382 if ((int)(head - offset) < 0)
2388 static void perf_output_wakeup(struct perf_output_handle *handle)
2390 atomic_set(&handle->data->poll, POLL_IN);
2393 handle->counter->pending_wakeup = 1;
2394 perf_pending_queue(&handle->counter->pending,
2395 perf_pending_counter);
2397 perf_counter_wakeup(handle->counter);
2401 * Curious locking construct.
2403 * We need to ensure a later event doesn't publish a head when a former
2404 * event isn't done writing. However since we need to deal with NMIs we
2405 * cannot fully serialize things.
2407 * What we do is serialize between CPUs so we only have to deal with NMI
2408 * nesting on a single CPU.
2410 * We only publish the head (and generate a wakeup) when the outer-most
2413 static void perf_output_lock(struct perf_output_handle *handle)
2415 struct perf_mmap_data *data = handle->data;
2420 local_irq_save(handle->flags);
2421 cpu = smp_processor_id();
2423 if (in_nmi() && atomic_read(&data->lock) == cpu)
2426 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2432 static void perf_output_unlock(struct perf_output_handle *handle)
2434 struct perf_mmap_data *data = handle->data;
2438 data->done_head = data->head;
2440 if (!handle->locked)
2445 * The xchg implies a full barrier that ensures all writes are done
2446 * before we publish the new head, matched by a rmb() in userspace when
2447 * reading this position.
2449 while ((head = atomic_long_xchg(&data->done_head, 0)))
2450 data->user_page->data_head = head;
2453 * NMI can happen here, which means we can miss a done_head update.
2456 cpu = atomic_xchg(&data->lock, -1);
2457 WARN_ON_ONCE(cpu != smp_processor_id());
2460 * Therefore we have to validate we did not indeed do so.
2462 if (unlikely(atomic_long_read(&data->done_head))) {
2464 * Since we had it locked, we can lock it again.
2466 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2472 if (atomic_xchg(&data->wakeup, 0))
2473 perf_output_wakeup(handle);
2475 local_irq_restore(handle->flags);
2478 static void perf_output_copy(struct perf_output_handle *handle,
2479 const void *buf, unsigned int len)
2481 unsigned int pages_mask;
2482 unsigned int offset;
2486 offset = handle->offset;
2487 pages_mask = handle->data->nr_pages - 1;
2488 pages = handle->data->data_pages;
2491 unsigned int page_offset;
2494 nr = (offset >> PAGE_SHIFT) & pages_mask;
2495 page_offset = offset & (PAGE_SIZE - 1);
2496 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2498 memcpy(pages[nr] + page_offset, buf, size);
2505 handle->offset = offset;
2508 * Check we didn't copy past our reservation window, taking the
2509 * possible unsigned int wrap into account.
2511 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2514 #define perf_output_put(handle, x) \
2515 perf_output_copy((handle), &(x), sizeof(x))
2517 static int perf_output_begin(struct perf_output_handle *handle,
2518 struct perf_counter *counter, unsigned int size,
2519 int nmi, int sample)
2521 struct perf_mmap_data *data;
2522 unsigned int offset, head;
2525 struct perf_event_header header;
2531 * For inherited counters we send all the output towards the parent.
2533 if (counter->parent)
2534 counter = counter->parent;
2537 data = rcu_dereference(counter->data);
2541 handle->data = data;
2542 handle->counter = counter;
2544 handle->sample = sample;
2546 if (!data->nr_pages)
2549 have_lost = atomic_read(&data->lost);
2551 size += sizeof(lost_event);
2553 perf_output_lock(handle);
2556 offset = head = atomic_long_read(&data->head);
2558 if (unlikely(!perf_output_space(data, offset, head)))
2560 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2562 handle->offset = offset;
2563 handle->head = head;
2565 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2566 atomic_set(&data->wakeup, 1);
2569 lost_event.header.type = PERF_EVENT_LOST;
2570 lost_event.header.misc = 0;
2571 lost_event.header.size = sizeof(lost_event);
2572 lost_event.id = counter->id;
2573 lost_event.lost = atomic_xchg(&data->lost, 0);
2575 perf_output_put(handle, lost_event);
2581 atomic_inc(&data->lost);
2582 perf_output_unlock(handle);
2589 static void perf_output_end(struct perf_output_handle *handle)
2591 struct perf_counter *counter = handle->counter;
2592 struct perf_mmap_data *data = handle->data;
2594 int wakeup_events = counter->attr.wakeup_events;
2596 if (handle->sample && wakeup_events) {
2597 int events = atomic_inc_return(&data->events);
2598 if (events >= wakeup_events) {
2599 atomic_sub(wakeup_events, &data->events);
2600 atomic_set(&data->wakeup, 1);
2604 perf_output_unlock(handle);
2608 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2611 * only top level counters have the pid namespace they were created in
2613 if (counter->parent)
2614 counter = counter->parent;
2616 return task_tgid_nr_ns(p, counter->ns);
2619 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2622 * only top level counters have the pid namespace they were created in
2624 if (counter->parent)
2625 counter = counter->parent;
2627 return task_pid_nr_ns(p, counter->ns);
2630 static void perf_counter_output(struct perf_counter *counter, int nmi,
2631 struct perf_sample_data *data)
2634 u64 sample_type = counter->attr.sample_type;
2635 struct perf_output_handle handle;
2636 struct perf_event_header header;
2645 struct perf_callchain_entry *callchain = NULL;
2646 int callchain_size = 0;
2652 header.type = PERF_EVENT_SAMPLE;
2653 header.size = sizeof(header);
2656 header.misc |= perf_misc_flags(data->regs);
2658 if (sample_type & PERF_SAMPLE_IP) {
2659 ip = perf_instruction_pointer(data->regs);
2660 header.size += sizeof(ip);
2663 if (sample_type & PERF_SAMPLE_TID) {
2664 /* namespace issues */
2665 tid_entry.pid = perf_counter_pid(counter, current);
2666 tid_entry.tid = perf_counter_tid(counter, current);
2668 header.size += sizeof(tid_entry);
2671 if (sample_type & PERF_SAMPLE_TIME) {
2673 * Maybe do better on x86 and provide cpu_clock_nmi()
2675 time = sched_clock();
2677 header.size += sizeof(u64);
2680 if (sample_type & PERF_SAMPLE_ADDR)
2681 header.size += sizeof(u64);
2683 if (sample_type & PERF_SAMPLE_ID)
2684 header.size += sizeof(u64);
2686 if (sample_type & PERF_SAMPLE_STREAM_ID)
2687 header.size += sizeof(u64);
2689 if (sample_type & PERF_SAMPLE_CPU) {
2690 header.size += sizeof(cpu_entry);
2692 cpu_entry.cpu = raw_smp_processor_id();
2693 cpu_entry.reserved = 0;
2696 if (sample_type & PERF_SAMPLE_PERIOD)
2697 header.size += sizeof(u64);
2699 if (sample_type & PERF_SAMPLE_GROUP) {
2700 header.size += sizeof(u64) +
2701 counter->nr_siblings * sizeof(group_entry);
2704 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2705 callchain = perf_callchain(data->regs);
2708 callchain_size = (1 + callchain->nr) * sizeof(u64);
2709 header.size += callchain_size;
2711 header.size += sizeof(u64);
2714 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2718 perf_output_put(&handle, header);
2720 if (sample_type & PERF_SAMPLE_IP)
2721 perf_output_put(&handle, ip);
2723 if (sample_type & PERF_SAMPLE_TID)
2724 perf_output_put(&handle, tid_entry);
2726 if (sample_type & PERF_SAMPLE_TIME)
2727 perf_output_put(&handle, time);
2729 if (sample_type & PERF_SAMPLE_ADDR)
2730 perf_output_put(&handle, data->addr);
2732 if (sample_type & PERF_SAMPLE_ID) {
2733 u64 id = primary_counter_id(counter);
2735 perf_output_put(&handle, id);
2738 if (sample_type & PERF_SAMPLE_STREAM_ID)
2739 perf_output_put(&handle, counter->id);
2741 if (sample_type & PERF_SAMPLE_CPU)
2742 perf_output_put(&handle, cpu_entry);
2744 if (sample_type & PERF_SAMPLE_PERIOD)
2745 perf_output_put(&handle, data->period);
2748 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2750 if (sample_type & PERF_SAMPLE_GROUP) {
2751 struct perf_counter *leader, *sub;
2752 u64 nr = counter->nr_siblings;
2754 perf_output_put(&handle, nr);
2756 leader = counter->group_leader;
2757 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2759 sub->pmu->read(sub);
2761 group_entry.id = primary_counter_id(sub);
2762 group_entry.counter = atomic64_read(&sub->count);
2764 perf_output_put(&handle, group_entry);
2768 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2770 perf_output_copy(&handle, callchain, callchain_size);
2773 perf_output_put(&handle, nr);
2777 perf_output_end(&handle);
2784 struct perf_read_event {
2785 struct perf_event_header header;
2794 perf_counter_read_event(struct perf_counter *counter,
2795 struct task_struct *task)
2797 struct perf_output_handle handle;
2798 struct perf_read_event event = {
2800 .type = PERF_EVENT_READ,
2802 .size = sizeof(event) - sizeof(event.format),
2804 .pid = perf_counter_pid(counter, task),
2805 .tid = perf_counter_tid(counter, task),
2806 .value = atomic64_read(&counter->count),
2810 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2811 event.header.size += sizeof(u64);
2812 event.format[i++] = counter->total_time_enabled;
2815 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2816 event.header.size += sizeof(u64);
2817 event.format[i++] = counter->total_time_running;
2820 if (counter->attr.read_format & PERF_FORMAT_ID) {
2821 event.header.size += sizeof(u64);
2822 event.format[i++] = primary_counter_id(counter);
2825 ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
2829 perf_output_copy(&handle, &event, event.header.size);
2830 perf_output_end(&handle);
2837 struct perf_fork_event {
2838 struct task_struct *task;
2841 struct perf_event_header header;
2848 static void perf_counter_fork_output(struct perf_counter *counter,
2849 struct perf_fork_event *fork_event)
2851 struct perf_output_handle handle;
2852 int size = fork_event->event.header.size;
2853 struct task_struct *task = fork_event->task;
2854 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2859 fork_event->event.pid = perf_counter_pid(counter, task);
2860 fork_event->event.ppid = perf_counter_pid(counter, task->real_parent);
2862 perf_output_put(&handle, fork_event->event);
2863 perf_output_end(&handle);
2866 static int perf_counter_fork_match(struct perf_counter *counter)
2868 if (counter->attr.comm || counter->attr.mmap)
2874 static void perf_counter_fork_ctx(struct perf_counter_context *ctx,
2875 struct perf_fork_event *fork_event)
2877 struct perf_counter *counter;
2879 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2883 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2884 if (perf_counter_fork_match(counter))
2885 perf_counter_fork_output(counter, fork_event);
2890 static void perf_counter_fork_event(struct perf_fork_event *fork_event)
2892 struct perf_cpu_context *cpuctx;
2893 struct perf_counter_context *ctx;
2895 cpuctx = &get_cpu_var(perf_cpu_context);
2896 perf_counter_fork_ctx(&cpuctx->ctx, fork_event);
2897 put_cpu_var(perf_cpu_context);
2901 * doesn't really matter which of the child contexts the
2902 * events ends up in.
2904 ctx = rcu_dereference(current->perf_counter_ctxp);
2906 perf_counter_fork_ctx(ctx, fork_event);
2910 void perf_counter_fork(struct task_struct *task)
2912 struct perf_fork_event fork_event;
2914 if (!atomic_read(&nr_comm_counters) &&
2915 !atomic_read(&nr_mmap_counters))
2918 fork_event = (struct perf_fork_event){
2922 .type = PERF_EVENT_FORK,
2924 .size = sizeof(fork_event.event),
2931 perf_counter_fork_event(&fork_event);
2938 struct perf_comm_event {
2939 struct task_struct *task;
2944 struct perf_event_header header;
2951 static void perf_counter_comm_output(struct perf_counter *counter,
2952 struct perf_comm_event *comm_event)
2954 struct perf_output_handle handle;
2955 int size = comm_event->event.header.size;
2956 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2961 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
2962 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
2964 perf_output_put(&handle, comm_event->event);
2965 perf_output_copy(&handle, comm_event->comm,
2966 comm_event->comm_size);
2967 perf_output_end(&handle);
2970 static int perf_counter_comm_match(struct perf_counter *counter)
2972 if (counter->attr.comm)
2978 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2979 struct perf_comm_event *comm_event)
2981 struct perf_counter *counter;
2983 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2987 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2988 if (perf_counter_comm_match(counter))
2989 perf_counter_comm_output(counter, comm_event);
2994 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2996 struct perf_cpu_context *cpuctx;
2997 struct perf_counter_context *ctx;
2999 char comm[TASK_COMM_LEN];
3001 memset(comm, 0, sizeof(comm));
3002 strncpy(comm, comm_event->task->comm, sizeof(comm));
3003 size = ALIGN(strlen(comm)+1, sizeof(u64));
3005 comm_event->comm = comm;
3006 comm_event->comm_size = size;
3008 comm_event->event.header.size = sizeof(comm_event->event) + size;
3010 cpuctx = &get_cpu_var(perf_cpu_context);
3011 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
3012 put_cpu_var(perf_cpu_context);
3016 * doesn't really matter which of the child contexts the
3017 * events ends up in.
3019 ctx = rcu_dereference(current->perf_counter_ctxp);
3021 perf_counter_comm_ctx(ctx, comm_event);
3025 void perf_counter_comm(struct task_struct *task)
3027 struct perf_comm_event comm_event;
3029 if (task->perf_counter_ctxp)
3030 perf_counter_enable_on_exec(task);
3032 if (!atomic_read(&nr_comm_counters))
3035 comm_event = (struct perf_comm_event){
3041 .type = PERF_EVENT_COMM,
3050 perf_counter_comm_event(&comm_event);
3057 struct perf_mmap_event {
3058 struct vm_area_struct *vma;
3060 const char *file_name;
3064 struct perf_event_header header;
3074 static void perf_counter_mmap_output(struct perf_counter *counter,
3075 struct perf_mmap_event *mmap_event)
3077 struct perf_output_handle handle;
3078 int size = mmap_event->event.header.size;
3079 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3084 mmap_event->event.pid = perf_counter_pid(counter, current);
3085 mmap_event->event.tid = perf_counter_tid(counter, current);
3087 perf_output_put(&handle, mmap_event->event);
3088 perf_output_copy(&handle, mmap_event->file_name,
3089 mmap_event->file_size);
3090 perf_output_end(&handle);
3093 static int perf_counter_mmap_match(struct perf_counter *counter,
3094 struct perf_mmap_event *mmap_event)
3096 if (counter->attr.mmap)
3102 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3103 struct perf_mmap_event *mmap_event)
3105 struct perf_counter *counter;
3107 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3111 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3112 if (perf_counter_mmap_match(counter, mmap_event))
3113 perf_counter_mmap_output(counter, mmap_event);
3118 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3120 struct perf_cpu_context *cpuctx;
3121 struct perf_counter_context *ctx;
3122 struct vm_area_struct *vma = mmap_event->vma;
3123 struct file *file = vma->vm_file;
3129 memset(tmp, 0, sizeof(tmp));
3133 * d_path works from the end of the buffer backwards, so we
3134 * need to add enough zero bytes after the string to handle
3135 * the 64bit alignment we do later.
3137 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3139 name = strncpy(tmp, "//enomem", sizeof(tmp));
3142 name = d_path(&file->f_path, buf, PATH_MAX);
3144 name = strncpy(tmp, "//toolong", sizeof(tmp));
3148 if (arch_vma_name(mmap_event->vma)) {
3149 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3155 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3159 name = strncpy(tmp, "//anon", sizeof(tmp));
3164 size = ALIGN(strlen(name)+1, sizeof(u64));
3166 mmap_event->file_name = name;
3167 mmap_event->file_size = size;
3169 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3171 cpuctx = &get_cpu_var(perf_cpu_context);
3172 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3173 put_cpu_var(perf_cpu_context);
3177 * doesn't really matter which of the child contexts the
3178 * events ends up in.
3180 ctx = rcu_dereference(current->perf_counter_ctxp);
3182 perf_counter_mmap_ctx(ctx, mmap_event);
3188 void __perf_counter_mmap(struct vm_area_struct *vma)
3190 struct perf_mmap_event mmap_event;
3192 if (!atomic_read(&nr_mmap_counters))
3195 mmap_event = (struct perf_mmap_event){
3201 .type = PERF_EVENT_MMAP,
3207 .start = vma->vm_start,
3208 .len = vma->vm_end - vma->vm_start,
3209 .pgoff = vma->vm_pgoff,
3213 perf_counter_mmap_event(&mmap_event);
3217 * IRQ throttle logging
3220 static void perf_log_throttle(struct perf_counter *counter, int enable)
3222 struct perf_output_handle handle;
3226 struct perf_event_header header;
3230 } throttle_event = {
3232 .type = PERF_EVENT_THROTTLE,
3234 .size = sizeof(throttle_event),
3236 .time = sched_clock(),
3237 .id = primary_counter_id(counter),
3238 .stream_id = counter->id,
3242 throttle_event.header.type = PERF_EVENT_UNTHROTTLE;
3244 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3248 perf_output_put(&handle, throttle_event);
3249 perf_output_end(&handle);
3253 * Generic counter overflow handling, sampling.
3256 int perf_counter_overflow(struct perf_counter *counter, int nmi,
3257 struct perf_sample_data *data)
3259 int events = atomic_read(&counter->event_limit);
3260 int throttle = counter->pmu->unthrottle != NULL;
3261 struct hw_perf_counter *hwc = &counter->hw;
3267 if (hwc->interrupts != MAX_INTERRUPTS) {
3269 if (HZ * hwc->interrupts >
3270 (u64)sysctl_perf_counter_sample_rate) {
3271 hwc->interrupts = MAX_INTERRUPTS;
3272 perf_log_throttle(counter, 0);
3277 * Keep re-disabling counters even though on the previous
3278 * pass we disabled it - just in case we raced with a
3279 * sched-in and the counter got enabled again:
3285 if (counter->attr.freq) {
3286 u64 now = sched_clock();
3287 s64 delta = now - hwc->freq_stamp;
3289 hwc->freq_stamp = now;
3291 if (delta > 0 && delta < TICK_NSEC)
3292 perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3296 * XXX event_limit might not quite work as expected on inherited
3300 counter->pending_kill = POLL_IN;
3301 if (events && atomic_dec_and_test(&counter->event_limit)) {
3303 counter->pending_kill = POLL_HUP;
3305 counter->pending_disable = 1;
3306 perf_pending_queue(&counter->pending,
3307 perf_pending_counter);
3309 perf_counter_disable(counter);
3312 perf_counter_output(counter, nmi, data);
3317 * Generic software counter infrastructure
3320 static void perf_swcounter_update(struct perf_counter *counter)
3322 struct hw_perf_counter *hwc = &counter->hw;
3327 prev = atomic64_read(&hwc->prev_count);
3328 now = atomic64_read(&hwc->count);
3329 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
3334 atomic64_add(delta, &counter->count);
3335 atomic64_sub(delta, &hwc->period_left);
3338 static void perf_swcounter_set_period(struct perf_counter *counter)
3340 struct hw_perf_counter *hwc = &counter->hw;
3341 s64 left = atomic64_read(&hwc->period_left);
3342 s64 period = hwc->sample_period;
3344 if (unlikely(left <= -period)) {
3346 atomic64_set(&hwc->period_left, left);
3347 hwc->last_period = period;
3350 if (unlikely(left <= 0)) {
3352 atomic64_add(period, &hwc->period_left);
3353 hwc->last_period = period;
3356 atomic64_set(&hwc->prev_count, -left);
3357 atomic64_set(&hwc->count, -left);
3360 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3362 enum hrtimer_restart ret = HRTIMER_RESTART;
3363 struct perf_sample_data data;
3364 struct perf_counter *counter;
3367 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3368 counter->pmu->read(counter);
3371 data.regs = get_irq_regs();
3373 * In case we exclude kernel IPs or are somehow not in interrupt
3374 * context, provide the next best thing, the user IP.
3376 if ((counter->attr.exclude_kernel || !data.regs) &&
3377 !counter->attr.exclude_user)
3378 data.regs = task_pt_regs(current);
3381 if (perf_counter_overflow(counter, 0, &data))
3382 ret = HRTIMER_NORESTART;
3385 period = max_t(u64, 10000, counter->hw.sample_period);
3386 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3391 static void perf_swcounter_overflow(struct perf_counter *counter,
3392 int nmi, struct perf_sample_data *data)
3394 data->period = counter->hw.last_period;
3396 perf_swcounter_update(counter);
3397 perf_swcounter_set_period(counter);
3398 if (perf_counter_overflow(counter, nmi, data))
3399 /* soft-disable the counter */
3403 static int perf_swcounter_is_counting(struct perf_counter *counter)
3405 struct perf_counter_context *ctx;
3406 unsigned long flags;
3409 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3412 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3416 * If the counter is inactive, it could be just because
3417 * its task is scheduled out, or because it's in a group
3418 * which could not go on the PMU. We want to count in
3419 * the first case but not the second. If the context is
3420 * currently active then an inactive software counter must
3421 * be the second case. If it's not currently active then
3422 * we need to know whether the counter was active when the
3423 * context was last active, which we can determine by
3424 * comparing counter->tstamp_stopped with ctx->time.
3426 * We are within an RCU read-side critical section,
3427 * which protects the existence of *ctx.
3430 spin_lock_irqsave(&ctx->lock, flags);
3432 /* Re-check state now we have the lock */
3433 if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
3434 counter->ctx->is_active ||
3435 counter->tstamp_stopped < ctx->time)
3437 spin_unlock_irqrestore(&ctx->lock, flags);
3441 static int perf_swcounter_match(struct perf_counter *counter,
3442 enum perf_type_id type,
3443 u32 event, struct pt_regs *regs)
3445 if (!perf_swcounter_is_counting(counter))
3448 if (counter->attr.type != type)
3450 if (counter->attr.config != event)
3454 if (counter->attr.exclude_user && user_mode(regs))
3457 if (counter->attr.exclude_kernel && !user_mode(regs))
3464 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3465 int nmi, struct perf_sample_data *data)
3467 int neg = atomic64_add_negative(nr, &counter->hw.count);
3469 if (counter->hw.sample_period && !neg && data->regs)
3470 perf_swcounter_overflow(counter, nmi, data);
3473 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3474 enum perf_type_id type,
3475 u32 event, u64 nr, int nmi,
3476 struct perf_sample_data *data)
3478 struct perf_counter *counter;
3480 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3484 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3485 if (perf_swcounter_match(counter, type, event, data->regs))
3486 perf_swcounter_add(counter, nr, nmi, data);
3491 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3494 return &cpuctx->recursion[3];
3497 return &cpuctx->recursion[2];
3500 return &cpuctx->recursion[1];
3502 return &cpuctx->recursion[0];
3505 static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3507 struct perf_sample_data *data)
3509 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3510 int *recursion = perf_swcounter_recursion_context(cpuctx);
3511 struct perf_counter_context *ctx;
3519 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3523 * doesn't really matter which of the child contexts the
3524 * events ends up in.
3526 ctx = rcu_dereference(current->perf_counter_ctxp);
3528 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3535 put_cpu_var(perf_cpu_context);
3538 void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3539 struct pt_regs *regs, u64 addr)
3541 struct perf_sample_data data = {
3546 do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3549 static void perf_swcounter_read(struct perf_counter *counter)
3551 perf_swcounter_update(counter);
3554 static int perf_swcounter_enable(struct perf_counter *counter)
3556 perf_swcounter_set_period(counter);
3560 static void perf_swcounter_disable(struct perf_counter *counter)
3562 perf_swcounter_update(counter);
3565 static const struct pmu perf_ops_generic = {
3566 .enable = perf_swcounter_enable,
3567 .disable = perf_swcounter_disable,
3568 .read = perf_swcounter_read,
3572 * Software counter: cpu wall time clock
3575 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3577 int cpu = raw_smp_processor_id();
3581 now = cpu_clock(cpu);
3582 prev = atomic64_read(&counter->hw.prev_count);
3583 atomic64_set(&counter->hw.prev_count, now);
3584 atomic64_add(now - prev, &counter->count);
3587 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3589 struct hw_perf_counter *hwc = &counter->hw;
3590 int cpu = raw_smp_processor_id();
3592 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3593 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3594 hwc->hrtimer.function = perf_swcounter_hrtimer;
3595 if (hwc->sample_period) {
3596 u64 period = max_t(u64, 10000, hwc->sample_period);
3597 __hrtimer_start_range_ns(&hwc->hrtimer,
3598 ns_to_ktime(period), 0,
3599 HRTIMER_MODE_REL, 0);
3605 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3607 if (counter->hw.sample_period)
3608 hrtimer_cancel(&counter->hw.hrtimer);
3609 cpu_clock_perf_counter_update(counter);
3612 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3614 cpu_clock_perf_counter_update(counter);
3617 static const struct pmu perf_ops_cpu_clock = {
3618 .enable = cpu_clock_perf_counter_enable,
3619 .disable = cpu_clock_perf_counter_disable,
3620 .read = cpu_clock_perf_counter_read,
3624 * Software counter: task time clock
3627 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3632 prev = atomic64_xchg(&counter->hw.prev_count, now);
3634 atomic64_add(delta, &counter->count);
3637 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3639 struct hw_perf_counter *hwc = &counter->hw;
3642 now = counter->ctx->time;
3644 atomic64_set(&hwc->prev_count, now);
3645 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3646 hwc->hrtimer.function = perf_swcounter_hrtimer;
3647 if (hwc->sample_period) {
3648 u64 period = max_t(u64, 10000, hwc->sample_period);
3649 __hrtimer_start_range_ns(&hwc->hrtimer,
3650 ns_to_ktime(period), 0,
3651 HRTIMER_MODE_REL, 0);
3657 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3659 if (counter->hw.sample_period)
3660 hrtimer_cancel(&counter->hw.hrtimer);
3661 task_clock_perf_counter_update(counter, counter->ctx->time);
3665 static void task_clock_perf_counter_read(struct perf_counter *counter)
3670 update_context_time(counter->ctx);
3671 time = counter->ctx->time;
3673 u64 now = perf_clock();
3674 u64 delta = now - counter->ctx->timestamp;
3675 time = counter->ctx->time + delta;
3678 task_clock_perf_counter_update(counter, time);
3681 static const struct pmu perf_ops_task_clock = {
3682 .enable = task_clock_perf_counter_enable,
3683 .disable = task_clock_perf_counter_disable,
3684 .read = task_clock_perf_counter_read,
3687 #ifdef CONFIG_EVENT_PROFILE
3688 void perf_tpcounter_event(int event_id)
3690 struct perf_sample_data data = {
3691 .regs = get_irq_regs(),
3696 data.regs = task_pt_regs(current);
3698 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, &data);
3700 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3702 extern int ftrace_profile_enable(int);
3703 extern void ftrace_profile_disable(int);
3705 static void tp_perf_counter_destroy(struct perf_counter *counter)
3707 ftrace_profile_disable(counter->attr.config);
3710 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3712 if (ftrace_profile_enable(counter->attr.config))
3715 counter->destroy = tp_perf_counter_destroy;
3717 return &perf_ops_generic;
3720 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3726 atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3728 static void sw_perf_counter_destroy(struct perf_counter *counter)
3730 u64 event = counter->attr.config;
3732 WARN_ON(counter->parent);
3734 atomic_dec(&perf_swcounter_enabled[event]);
3737 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3739 const struct pmu *pmu = NULL;
3740 u64 event = counter->attr.config;
3743 * Software counters (currently) can't in general distinguish
3744 * between user, kernel and hypervisor events.
3745 * However, context switches and cpu migrations are considered
3746 * to be kernel events, and page faults are never hypervisor
3750 case PERF_COUNT_SW_CPU_CLOCK:
3751 pmu = &perf_ops_cpu_clock;
3754 case PERF_COUNT_SW_TASK_CLOCK:
3756 * If the user instantiates this as a per-cpu counter,
3757 * use the cpu_clock counter instead.
3759 if (counter->ctx->task)
3760 pmu = &perf_ops_task_clock;
3762 pmu = &perf_ops_cpu_clock;
3765 case PERF_COUNT_SW_PAGE_FAULTS:
3766 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
3767 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
3768 case PERF_COUNT_SW_CONTEXT_SWITCHES:
3769 case PERF_COUNT_SW_CPU_MIGRATIONS:
3770 if (!counter->parent) {
3771 atomic_inc(&perf_swcounter_enabled[event]);
3772 counter->destroy = sw_perf_counter_destroy;
3774 pmu = &perf_ops_generic;
3782 * Allocate and initialize a counter structure
3784 static struct perf_counter *
3785 perf_counter_alloc(struct perf_counter_attr *attr,
3787 struct perf_counter_context *ctx,
3788 struct perf_counter *group_leader,
3789 struct perf_counter *parent_counter,
3792 const struct pmu *pmu;
3793 struct perf_counter *counter;
3794 struct hw_perf_counter *hwc;
3797 counter = kzalloc(sizeof(*counter), gfpflags);
3799 return ERR_PTR(-ENOMEM);
3802 * Single counters are their own group leaders, with an
3803 * empty sibling list:
3806 group_leader = counter;
3808 mutex_init(&counter->child_mutex);
3809 INIT_LIST_HEAD(&counter->child_list);
3811 INIT_LIST_HEAD(&counter->list_entry);
3812 INIT_LIST_HEAD(&counter->event_entry);
3813 INIT_LIST_HEAD(&counter->sibling_list);
3814 init_waitqueue_head(&counter->waitq);
3816 mutex_init(&counter->mmap_mutex);
3819 counter->attr = *attr;
3820 counter->group_leader = group_leader;
3821 counter->pmu = NULL;
3823 counter->oncpu = -1;
3825 counter->parent = parent_counter;
3827 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
3828 counter->id = atomic64_inc_return(&perf_counter_id);
3830 counter->state = PERF_COUNTER_STATE_INACTIVE;
3833 counter->state = PERF_COUNTER_STATE_OFF;
3838 hwc->sample_period = attr->sample_period;
3839 if (attr->freq && attr->sample_freq)
3840 hwc->sample_period = 1;
3842 atomic64_set(&hwc->period_left, hwc->sample_period);
3845 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3847 if (attr->inherit && (attr->sample_type & PERF_SAMPLE_GROUP))
3850 switch (attr->type) {
3852 case PERF_TYPE_HARDWARE:
3853 case PERF_TYPE_HW_CACHE:
3854 pmu = hw_perf_counter_init(counter);
3857 case PERF_TYPE_SOFTWARE:
3858 pmu = sw_perf_counter_init(counter);
3861 case PERF_TYPE_TRACEPOINT:
3862 pmu = tp_perf_counter_init(counter);
3872 else if (IS_ERR(pmu))
3877 put_pid_ns(counter->ns);
3879 return ERR_PTR(err);
3884 if (!counter->parent) {
3885 atomic_inc(&nr_counters);
3886 if (counter->attr.mmap)
3887 atomic_inc(&nr_mmap_counters);
3888 if (counter->attr.comm)
3889 atomic_inc(&nr_comm_counters);
3895 static int perf_copy_attr(struct perf_counter_attr __user *uattr,
3896 struct perf_counter_attr *attr)
3901 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
3905 * zero the full structure, so that a short copy will be nice.
3907 memset(attr, 0, sizeof(*attr));
3909 ret = get_user(size, &uattr->size);
3913 if (size > PAGE_SIZE) /* silly large */
3916 if (!size) /* abi compat */
3917 size = PERF_ATTR_SIZE_VER0;
3919 if (size < PERF_ATTR_SIZE_VER0)
3923 * If we're handed a bigger struct than we know of,
3924 * ensure all the unknown bits are 0.
3926 if (size > sizeof(*attr)) {
3928 unsigned long __user *addr;
3929 unsigned long __user *end;
3931 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
3932 sizeof(unsigned long));
3933 end = PTR_ALIGN((void __user *)uattr + size,
3934 sizeof(unsigned long));
3936 for (; addr < end; addr += sizeof(unsigned long)) {
3937 ret = get_user(val, addr);
3945 ret = copy_from_user(attr, uattr, size);
3950 * If the type exists, the corresponding creation will verify
3953 if (attr->type >= PERF_TYPE_MAX)
3956 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
3959 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
3962 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
3969 put_user(sizeof(*attr), &uattr->size);
3975 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3977 * @attr_uptr: event type attributes for monitoring/sampling
3980 * @group_fd: group leader counter fd
3982 SYSCALL_DEFINE5(perf_counter_open,
3983 struct perf_counter_attr __user *, attr_uptr,
3984 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3986 struct perf_counter *counter, *group_leader;
3987 struct perf_counter_attr attr;
3988 struct perf_counter_context *ctx;
3989 struct file *counter_file = NULL;
3990 struct file *group_file = NULL;
3991 int fput_needed = 0;
3992 int fput_needed2 = 0;
3995 /* for future expandability... */
3999 ret = perf_copy_attr(attr_uptr, &attr);
4003 if (!attr.exclude_kernel) {
4004 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4009 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
4014 * Get the target context (task or percpu):
4016 ctx = find_get_context(pid, cpu);
4018 return PTR_ERR(ctx);
4021 * Look up the group leader (we will attach this counter to it):
4023 group_leader = NULL;
4024 if (group_fd != -1) {
4026 group_file = fget_light(group_fd, &fput_needed);
4028 goto err_put_context;
4029 if (group_file->f_op != &perf_fops)
4030 goto err_put_context;
4032 group_leader = group_file->private_data;
4034 * Do not allow a recursive hierarchy (this new sibling
4035 * becoming part of another group-sibling):
4037 if (group_leader->group_leader != group_leader)
4038 goto err_put_context;
4040 * Do not allow to attach to a group in a different
4041 * task or CPU context:
4043 if (group_leader->ctx != ctx)
4044 goto err_put_context;
4046 * Only a group leader can be exclusive or pinned
4048 if (attr.exclusive || attr.pinned)
4049 goto err_put_context;
4052 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
4054 ret = PTR_ERR(counter);
4055 if (IS_ERR(counter))
4056 goto err_put_context;
4058 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
4060 goto err_free_put_context;
4062 counter_file = fget_light(ret, &fput_needed2);
4064 goto err_free_put_context;
4066 counter->filp = counter_file;
4067 WARN_ON_ONCE(ctx->parent_ctx);
4068 mutex_lock(&ctx->mutex);
4069 perf_install_in_context(ctx, counter, cpu);
4071 mutex_unlock(&ctx->mutex);
4073 counter->owner = current;
4074 get_task_struct(current);
4075 mutex_lock(¤t->perf_counter_mutex);
4076 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
4077 mutex_unlock(¤t->perf_counter_mutex);
4079 fput_light(counter_file, fput_needed2);
4082 fput_light(group_file, fput_needed);
4086 err_free_put_context:
4096 * inherit a counter from parent task to child task:
4098 static struct perf_counter *
4099 inherit_counter(struct perf_counter *parent_counter,
4100 struct task_struct *parent,
4101 struct perf_counter_context *parent_ctx,
4102 struct task_struct *child,
4103 struct perf_counter *group_leader,
4104 struct perf_counter_context *child_ctx)
4106 struct perf_counter *child_counter;
4109 * Instead of creating recursive hierarchies of counters,
4110 * we link inherited counters back to the original parent,
4111 * which has a filp for sure, which we use as the reference
4114 if (parent_counter->parent)
4115 parent_counter = parent_counter->parent;
4117 child_counter = perf_counter_alloc(&parent_counter->attr,
4118 parent_counter->cpu, child_ctx,
4119 group_leader, parent_counter,
4121 if (IS_ERR(child_counter))
4122 return child_counter;
4126 * Make the child state follow the state of the parent counter,
4127 * not its attr.disabled bit. We hold the parent's mutex,
4128 * so we won't race with perf_counter_{en, dis}able_family.
4130 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4131 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4133 child_counter->state = PERF_COUNTER_STATE_OFF;
4135 if (parent_counter->attr.freq)
4136 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4139 * Link it up in the child's context:
4141 add_counter_to_ctx(child_counter, child_ctx);
4144 * Get a reference to the parent filp - we will fput it
4145 * when the child counter exits. This is safe to do because
4146 * we are in the parent and we know that the filp still
4147 * exists and has a nonzero count:
4149 atomic_long_inc(&parent_counter->filp->f_count);
4152 * Link this into the parent counter's child list
4154 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4155 mutex_lock(&parent_counter->child_mutex);
4156 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4157 mutex_unlock(&parent_counter->child_mutex);
4159 return child_counter;
4162 static int inherit_group(struct perf_counter *parent_counter,
4163 struct task_struct *parent,
4164 struct perf_counter_context *parent_ctx,
4165 struct task_struct *child,
4166 struct perf_counter_context *child_ctx)
4168 struct perf_counter *leader;
4169 struct perf_counter *sub;
4170 struct perf_counter *child_ctr;
4172 leader = inherit_counter(parent_counter, parent, parent_ctx,
4173 child, NULL, child_ctx);
4175 return PTR_ERR(leader);
4176 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4177 child_ctr = inherit_counter(sub, parent, parent_ctx,
4178 child, leader, child_ctx);
4179 if (IS_ERR(child_ctr))
4180 return PTR_ERR(child_ctr);
4185 static void sync_child_counter(struct perf_counter *child_counter,
4186 struct task_struct *child)
4188 struct perf_counter *parent_counter = child_counter->parent;
4191 if (child_counter->attr.inherit_stat)
4192 perf_counter_read_event(child_counter, child);
4194 child_val = atomic64_read(&child_counter->count);
4197 * Add back the child's count to the parent's count:
4199 atomic64_add(child_val, &parent_counter->count);
4200 atomic64_add(child_counter->total_time_enabled,
4201 &parent_counter->child_total_time_enabled);
4202 atomic64_add(child_counter->total_time_running,
4203 &parent_counter->child_total_time_running);
4206 * Remove this counter from the parent's list
4208 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4209 mutex_lock(&parent_counter->child_mutex);
4210 list_del_init(&child_counter->child_list);
4211 mutex_unlock(&parent_counter->child_mutex);
4214 * Release the parent counter, if this was the last
4217 fput(parent_counter->filp);
4221 __perf_counter_exit_task(struct perf_counter *child_counter,
4222 struct perf_counter_context *child_ctx,
4223 struct task_struct *child)
4225 struct perf_counter *parent_counter;
4227 update_counter_times(child_counter);
4228 perf_counter_remove_from_context(child_counter);
4230 parent_counter = child_counter->parent;
4232 * It can happen that parent exits first, and has counters
4233 * that are still around due to the child reference. These
4234 * counters need to be zapped - but otherwise linger.
4236 if (parent_counter) {
4237 sync_child_counter(child_counter, child);
4238 free_counter(child_counter);
4243 * When a child task exits, feed back counter values to parent counters.
4245 void perf_counter_exit_task(struct task_struct *child)
4247 struct perf_counter *child_counter, *tmp;
4248 struct perf_counter_context *child_ctx;
4249 unsigned long flags;
4251 if (likely(!child->perf_counter_ctxp))
4254 local_irq_save(flags);
4256 * We can't reschedule here because interrupts are disabled,
4257 * and either child is current or it is a task that can't be
4258 * scheduled, so we are now safe from rescheduling changing
4261 child_ctx = child->perf_counter_ctxp;
4262 __perf_counter_task_sched_out(child_ctx);
4265 * Take the context lock here so that if find_get_context is
4266 * reading child->perf_counter_ctxp, we wait until it has
4267 * incremented the context's refcount before we do put_ctx below.
4269 spin_lock(&child_ctx->lock);
4270 child->perf_counter_ctxp = NULL;
4272 * If this context is a clone; unclone it so it can't get
4273 * swapped to another process while we're removing all
4274 * the counters from it.
4276 unclone_ctx(child_ctx);
4277 spin_unlock(&child_ctx->lock);
4278 local_irq_restore(flags);
4281 * We can recurse on the same lock type through:
4283 * __perf_counter_exit_task()
4284 * sync_child_counter()
4285 * fput(parent_counter->filp)
4287 * mutex_lock(&ctx->mutex)
4289 * But since its the parent context it won't be the same instance.
4291 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4294 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4296 __perf_counter_exit_task(child_counter, child_ctx, child);
4299 * If the last counter was a group counter, it will have appended all
4300 * its siblings to the list, but we obtained 'tmp' before that which
4301 * will still point to the list head terminating the iteration.
4303 if (!list_empty(&child_ctx->counter_list))
4306 mutex_unlock(&child_ctx->mutex);
4312 * free an unexposed, unused context as created by inheritance by
4313 * init_task below, used by fork() in case of fail.
4315 void perf_counter_free_task(struct task_struct *task)
4317 struct perf_counter_context *ctx = task->perf_counter_ctxp;
4318 struct perf_counter *counter, *tmp;
4323 mutex_lock(&ctx->mutex);
4325 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4326 struct perf_counter *parent = counter->parent;
4328 if (WARN_ON_ONCE(!parent))
4331 mutex_lock(&parent->child_mutex);
4332 list_del_init(&counter->child_list);
4333 mutex_unlock(&parent->child_mutex);
4337 list_del_counter(counter, ctx);
4338 free_counter(counter);
4341 if (!list_empty(&ctx->counter_list))
4344 mutex_unlock(&ctx->mutex);
4350 * Initialize the perf_counter context in task_struct
4352 int perf_counter_init_task(struct task_struct *child)
4354 struct perf_counter_context *child_ctx, *parent_ctx;
4355 struct perf_counter_context *cloned_ctx;
4356 struct perf_counter *counter;
4357 struct task_struct *parent = current;
4358 int inherited_all = 1;
4361 child->perf_counter_ctxp = NULL;
4363 mutex_init(&child->perf_counter_mutex);
4364 INIT_LIST_HEAD(&child->perf_counter_list);
4366 if (likely(!parent->perf_counter_ctxp))
4370 * This is executed from the parent task context, so inherit
4371 * counters that have been marked for cloning.
4372 * First allocate and initialize a context for the child.
4375 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4379 __perf_counter_init_context(child_ctx, child);
4380 child->perf_counter_ctxp = child_ctx;
4381 get_task_struct(child);
4384 * If the parent's context is a clone, pin it so it won't get
4387 parent_ctx = perf_pin_task_context(parent);
4390 * No need to check if parent_ctx != NULL here; since we saw
4391 * it non-NULL earlier, the only reason for it to become NULL
4392 * is if we exit, and since we're currently in the middle of
4393 * a fork we can't be exiting at the same time.
4397 * Lock the parent list. No need to lock the child - not PID
4398 * hashed yet and not running, so nobody can access it.
4400 mutex_lock(&parent_ctx->mutex);
4403 * We dont have to disable NMIs - we are only looking at
4404 * the list, not manipulating it:
4406 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4407 if (counter != counter->group_leader)
4410 if (!counter->attr.inherit) {
4415 ret = inherit_group(counter, parent, parent_ctx,
4423 if (inherited_all) {
4425 * Mark the child context as a clone of the parent
4426 * context, or of whatever the parent is a clone of.
4427 * Note that if the parent is a clone, it could get
4428 * uncloned at any point, but that doesn't matter
4429 * because the list of counters and the generation
4430 * count can't have changed since we took the mutex.
4432 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4434 child_ctx->parent_ctx = cloned_ctx;
4435 child_ctx->parent_gen = parent_ctx->parent_gen;
4437 child_ctx->parent_ctx = parent_ctx;
4438 child_ctx->parent_gen = parent_ctx->generation;
4440 get_ctx(child_ctx->parent_ctx);
4443 mutex_unlock(&parent_ctx->mutex);
4445 perf_unpin_context(parent_ctx);
4450 static void __cpuinit perf_counter_init_cpu(int cpu)
4452 struct perf_cpu_context *cpuctx;
4454 cpuctx = &per_cpu(perf_cpu_context, cpu);
4455 __perf_counter_init_context(&cpuctx->ctx, NULL);
4457 spin_lock(&perf_resource_lock);
4458 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4459 spin_unlock(&perf_resource_lock);
4461 hw_perf_counter_setup(cpu);
4464 #ifdef CONFIG_HOTPLUG_CPU
4465 static void __perf_counter_exit_cpu(void *info)
4467 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4468 struct perf_counter_context *ctx = &cpuctx->ctx;
4469 struct perf_counter *counter, *tmp;
4471 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4472 __perf_counter_remove_from_context(counter);
4474 static void perf_counter_exit_cpu(int cpu)
4476 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4477 struct perf_counter_context *ctx = &cpuctx->ctx;
4479 mutex_lock(&ctx->mutex);
4480 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4481 mutex_unlock(&ctx->mutex);
4484 static inline void perf_counter_exit_cpu(int cpu) { }
4487 static int __cpuinit
4488 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4490 unsigned int cpu = (long)hcpu;
4494 case CPU_UP_PREPARE:
4495 case CPU_UP_PREPARE_FROZEN:
4496 perf_counter_init_cpu(cpu);
4499 case CPU_DOWN_PREPARE:
4500 case CPU_DOWN_PREPARE_FROZEN:
4501 perf_counter_exit_cpu(cpu);
4512 * This has to have a higher priority than migration_notifier in sched.c.
4514 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4515 .notifier_call = perf_cpu_notify,
4519 void __init perf_counter_init(void)
4521 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4522 (void *)(long)smp_processor_id());
4523 register_cpu_notifier(&perf_cpu_nb);
4526 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4528 return sprintf(buf, "%d\n", perf_reserved_percpu);
4532 perf_set_reserve_percpu(struct sysdev_class *class,
4536 struct perf_cpu_context *cpuctx;
4540 err = strict_strtoul(buf, 10, &val);
4543 if (val > perf_max_counters)
4546 spin_lock(&perf_resource_lock);
4547 perf_reserved_percpu = val;
4548 for_each_online_cpu(cpu) {
4549 cpuctx = &per_cpu(perf_cpu_context, cpu);
4550 spin_lock_irq(&cpuctx->ctx.lock);
4551 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4552 perf_max_counters - perf_reserved_percpu);
4553 cpuctx->max_pertask = mpt;
4554 spin_unlock_irq(&cpuctx->ctx.lock);
4556 spin_unlock(&perf_resource_lock);
4561 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4563 return sprintf(buf, "%d\n", perf_overcommit);
4567 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4572 err = strict_strtoul(buf, 10, &val);
4578 spin_lock(&perf_resource_lock);
4579 perf_overcommit = val;
4580 spin_unlock(&perf_resource_lock);
4585 static SYSDEV_CLASS_ATTR(
4588 perf_show_reserve_percpu,
4589 perf_set_reserve_percpu
4592 static SYSDEV_CLASS_ATTR(
4595 perf_show_overcommit,
4599 static struct attribute *perfclass_attrs[] = {
4600 &attr_reserve_percpu.attr,
4601 &attr_overcommit.attr,
4605 static struct attribute_group perfclass_attr_group = {
4606 .attrs = perfclass_attrs,
4607 .name = "perf_counters",
4610 static int __init perf_counter_sysfs_init(void)
4612 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4613 &perfclass_attr_group);
4615 device_initcall(perf_counter_sysfs_init);