2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_counters __read_mostly;
44 static atomic_t nr_comm_counters __read_mostly;
45 static atomic_t nr_task_counters __read_mostly;
48 * perf counter paranoia level:
50 * 1 - disallow cpu counters to unpriv
51 * 2 - disallow kernel profiling to unpriv
53 int sysctl_perf_counter_paranoid __read_mostly;
55 static inline bool perf_paranoid_cpu(void)
57 return sysctl_perf_counter_paranoid > 0;
60 static inline bool perf_paranoid_kernel(void)
62 return sysctl_perf_counter_paranoid > 1;
65 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
68 * max perf counter sample rate
70 int sysctl_perf_counter_sample_rate __read_mostly = 100000;
72 static atomic64_t perf_counter_id;
75 * Lock for (sysadmin-configurable) counter reservations:
77 static DEFINE_SPINLOCK(perf_resource_lock);
80 * Architecture provided APIs - weak aliases:
82 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
87 void __weak hw_perf_disable(void) { barrier(); }
88 void __weak hw_perf_enable(void) { barrier(); }
90 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
91 void __weak hw_perf_counter_setup_online(int cpu) { barrier(); }
94 hw_perf_group_sched_in(struct perf_counter *group_leader,
95 struct perf_cpu_context *cpuctx,
96 struct perf_counter_context *ctx, int cpu)
101 void __weak perf_counter_print_debug(void) { }
103 static DEFINE_PER_CPU(int, perf_disable_count);
105 void __perf_disable(void)
107 __get_cpu_var(perf_disable_count)++;
110 bool __perf_enable(void)
112 return !--__get_cpu_var(perf_disable_count);
115 void perf_disable(void)
121 void perf_enable(void)
127 static void get_ctx(struct perf_counter_context *ctx)
129 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
132 static void free_ctx(struct rcu_head *head)
134 struct perf_counter_context *ctx;
136 ctx = container_of(head, struct perf_counter_context, rcu_head);
140 static void put_ctx(struct perf_counter_context *ctx)
142 if (atomic_dec_and_test(&ctx->refcount)) {
144 put_ctx(ctx->parent_ctx);
146 put_task_struct(ctx->task);
147 call_rcu(&ctx->rcu_head, free_ctx);
151 static void unclone_ctx(struct perf_counter_context *ctx)
153 if (ctx->parent_ctx) {
154 put_ctx(ctx->parent_ctx);
155 ctx->parent_ctx = NULL;
160 * If we inherit counters we want to return the parent counter id
163 static u64 primary_counter_id(struct perf_counter *counter)
165 u64 id = counter->id;
168 id = counter->parent->id;
174 * Get the perf_counter_context for a task and lock it.
175 * This has to cope with with the fact that until it is locked,
176 * the context could get moved to another task.
178 static struct perf_counter_context *
179 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
181 struct perf_counter_context *ctx;
185 ctx = rcu_dereference(task->perf_counter_ctxp);
188 * If this context is a clone of another, it might
189 * get swapped for another underneath us by
190 * perf_counter_task_sched_out, though the
191 * rcu_read_lock() protects us from any context
192 * getting freed. Lock the context and check if it
193 * got swapped before we could get the lock, and retry
194 * if so. If we locked the right context, then it
195 * can't get swapped on us any more.
197 spin_lock_irqsave(&ctx->lock, *flags);
198 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
199 spin_unlock_irqrestore(&ctx->lock, *flags);
203 if (!atomic_inc_not_zero(&ctx->refcount)) {
204 spin_unlock_irqrestore(&ctx->lock, *flags);
213 * Get the context for a task and increment its pin_count so it
214 * can't get swapped to another task. This also increments its
215 * reference count so that the context can't get freed.
217 static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
219 struct perf_counter_context *ctx;
222 ctx = perf_lock_task_context(task, &flags);
225 spin_unlock_irqrestore(&ctx->lock, flags);
230 static void perf_unpin_context(struct perf_counter_context *ctx)
234 spin_lock_irqsave(&ctx->lock, flags);
236 spin_unlock_irqrestore(&ctx->lock, flags);
241 * Add a counter from the lists for its context.
242 * Must be called with ctx->mutex and ctx->lock held.
245 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
247 struct perf_counter *group_leader = counter->group_leader;
250 * Depending on whether it is a standalone or sibling counter,
251 * add it straight to the context's counter list, or to the group
252 * leader's sibling list:
254 if (group_leader == counter)
255 list_add_tail(&counter->list_entry, &ctx->counter_list);
257 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
258 group_leader->nr_siblings++;
261 list_add_rcu(&counter->event_entry, &ctx->event_list);
263 if (counter->attr.inherit_stat)
268 * Remove a counter from the lists for its context.
269 * Must be called with ctx->mutex and ctx->lock held.
272 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
274 struct perf_counter *sibling, *tmp;
276 if (list_empty(&counter->list_entry))
279 if (counter->attr.inherit_stat)
282 list_del_init(&counter->list_entry);
283 list_del_rcu(&counter->event_entry);
285 if (counter->group_leader != counter)
286 counter->group_leader->nr_siblings--;
289 * If this was a group counter with sibling counters then
290 * upgrade the siblings to singleton counters by adding them
291 * to the context list directly:
293 list_for_each_entry_safe(sibling, tmp,
294 &counter->sibling_list, list_entry) {
296 list_move_tail(&sibling->list_entry, &ctx->counter_list);
297 sibling->group_leader = sibling;
302 counter_sched_out(struct perf_counter *counter,
303 struct perf_cpu_context *cpuctx,
304 struct perf_counter_context *ctx)
306 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
309 counter->state = PERF_COUNTER_STATE_INACTIVE;
310 if (counter->pending_disable) {
311 counter->pending_disable = 0;
312 counter->state = PERF_COUNTER_STATE_OFF;
314 counter->tstamp_stopped = ctx->time;
315 counter->pmu->disable(counter);
318 if (!is_software_counter(counter))
319 cpuctx->active_oncpu--;
321 if (counter->attr.exclusive || !cpuctx->active_oncpu)
322 cpuctx->exclusive = 0;
326 group_sched_out(struct perf_counter *group_counter,
327 struct perf_cpu_context *cpuctx,
328 struct perf_counter_context *ctx)
330 struct perf_counter *counter;
332 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
335 counter_sched_out(group_counter, cpuctx, ctx);
338 * Schedule out siblings (if any):
340 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
341 counter_sched_out(counter, cpuctx, ctx);
343 if (group_counter->attr.exclusive)
344 cpuctx->exclusive = 0;
348 * Cross CPU call to remove a performance counter
350 * We disable the counter on the hardware level first. After that we
351 * remove it from the context list.
353 static void __perf_counter_remove_from_context(void *info)
355 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
356 struct perf_counter *counter = info;
357 struct perf_counter_context *ctx = counter->ctx;
360 * If this is a task context, we need to check whether it is
361 * the current task context of this cpu. If not it has been
362 * scheduled out before the smp call arrived.
364 if (ctx->task && cpuctx->task_ctx != ctx)
367 spin_lock(&ctx->lock);
369 * Protect the list operation against NMI by disabling the
370 * counters on a global level.
374 counter_sched_out(counter, cpuctx, ctx);
376 list_del_counter(counter, ctx);
380 * Allow more per task counters with respect to the
383 cpuctx->max_pertask =
384 min(perf_max_counters - ctx->nr_counters,
385 perf_max_counters - perf_reserved_percpu);
389 spin_unlock(&ctx->lock);
394 * Remove the counter from a task's (or a CPU's) list of counters.
396 * Must be called with ctx->mutex held.
398 * CPU counters are removed with a smp call. For task counters we only
399 * call when the task is on a CPU.
401 * If counter->ctx is a cloned context, callers must make sure that
402 * every task struct that counter->ctx->task could possibly point to
403 * remains valid. This is OK when called from perf_release since
404 * that only calls us on the top-level context, which can't be a clone.
405 * When called from perf_counter_exit_task, it's OK because the
406 * context has been detached from its task.
408 static void perf_counter_remove_from_context(struct perf_counter *counter)
410 struct perf_counter_context *ctx = counter->ctx;
411 struct task_struct *task = ctx->task;
415 * Per cpu counters are removed via an smp call and
416 * the removal is always sucessful.
418 smp_call_function_single(counter->cpu,
419 __perf_counter_remove_from_context,
425 task_oncpu_function_call(task, __perf_counter_remove_from_context,
428 spin_lock_irq(&ctx->lock);
430 * If the context is active we need to retry the smp call.
432 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
433 spin_unlock_irq(&ctx->lock);
438 * The lock prevents that this context is scheduled in so we
439 * can remove the counter safely, if the call above did not
442 if (!list_empty(&counter->list_entry)) {
443 list_del_counter(counter, ctx);
445 spin_unlock_irq(&ctx->lock);
448 static inline u64 perf_clock(void)
450 return cpu_clock(smp_processor_id());
454 * Update the record of the current time in a context.
456 static void update_context_time(struct perf_counter_context *ctx)
458 u64 now = perf_clock();
460 ctx->time += now - ctx->timestamp;
461 ctx->timestamp = now;
465 * Update the total_time_enabled and total_time_running fields for a counter.
467 static void update_counter_times(struct perf_counter *counter)
469 struct perf_counter_context *ctx = counter->ctx;
472 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
475 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
477 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
478 run_end = counter->tstamp_stopped;
482 counter->total_time_running = run_end - counter->tstamp_running;
486 * Update total_time_enabled and total_time_running for all counters in a group.
488 static void update_group_times(struct perf_counter *leader)
490 struct perf_counter *counter;
492 update_counter_times(leader);
493 list_for_each_entry(counter, &leader->sibling_list, list_entry)
494 update_counter_times(counter);
498 * Cross CPU call to disable a performance counter
500 static void __perf_counter_disable(void *info)
502 struct perf_counter *counter = info;
503 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
504 struct perf_counter_context *ctx = counter->ctx;
507 * If this is a per-task counter, need to check whether this
508 * counter's task is the current task on this cpu.
510 if (ctx->task && cpuctx->task_ctx != ctx)
513 spin_lock(&ctx->lock);
516 * If the counter is on, turn it off.
517 * If it is in error state, leave it in error state.
519 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
520 update_context_time(ctx);
521 update_counter_times(counter);
522 if (counter == counter->group_leader)
523 group_sched_out(counter, cpuctx, ctx);
525 counter_sched_out(counter, cpuctx, ctx);
526 counter->state = PERF_COUNTER_STATE_OFF;
529 spin_unlock(&ctx->lock);
535 * If counter->ctx is a cloned context, callers must make sure that
536 * every task struct that counter->ctx->task could possibly point to
537 * remains valid. This condition is satisifed when called through
538 * perf_counter_for_each_child or perf_counter_for_each because they
539 * hold the top-level counter's child_mutex, so any descendant that
540 * goes to exit will block in sync_child_counter.
541 * When called from perf_pending_counter it's OK because counter->ctx
542 * is the current context on this CPU and preemption is disabled,
543 * hence we can't get into perf_counter_task_sched_out for this context.
545 static void perf_counter_disable(struct perf_counter *counter)
547 struct perf_counter_context *ctx = counter->ctx;
548 struct task_struct *task = ctx->task;
552 * Disable the counter on the cpu that it's on
554 smp_call_function_single(counter->cpu, __perf_counter_disable,
560 task_oncpu_function_call(task, __perf_counter_disable, counter);
562 spin_lock_irq(&ctx->lock);
564 * If the counter is still active, we need to retry the cross-call.
566 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
567 spin_unlock_irq(&ctx->lock);
572 * Since we have the lock this context can't be scheduled
573 * in, so we can change the state safely.
575 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
576 update_counter_times(counter);
577 counter->state = PERF_COUNTER_STATE_OFF;
580 spin_unlock_irq(&ctx->lock);
584 counter_sched_in(struct perf_counter *counter,
585 struct perf_cpu_context *cpuctx,
586 struct perf_counter_context *ctx,
589 if (counter->state <= PERF_COUNTER_STATE_OFF)
592 counter->state = PERF_COUNTER_STATE_ACTIVE;
593 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
595 * The new state must be visible before we turn it on in the hardware:
599 if (counter->pmu->enable(counter)) {
600 counter->state = PERF_COUNTER_STATE_INACTIVE;
605 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
607 if (!is_software_counter(counter))
608 cpuctx->active_oncpu++;
611 if (counter->attr.exclusive)
612 cpuctx->exclusive = 1;
618 group_sched_in(struct perf_counter *group_counter,
619 struct perf_cpu_context *cpuctx,
620 struct perf_counter_context *ctx,
623 struct perf_counter *counter, *partial_group;
626 if (group_counter->state == PERF_COUNTER_STATE_OFF)
629 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
631 return ret < 0 ? ret : 0;
633 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
637 * Schedule in siblings as one group (if any):
639 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
640 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
641 partial_group = counter;
650 * Groups can be scheduled in as one unit only, so undo any
651 * partial group before returning:
653 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
654 if (counter == partial_group)
656 counter_sched_out(counter, cpuctx, ctx);
658 counter_sched_out(group_counter, cpuctx, ctx);
664 * Return 1 for a group consisting entirely of software counters,
665 * 0 if the group contains any hardware counters.
667 static int is_software_only_group(struct perf_counter *leader)
669 struct perf_counter *counter;
671 if (!is_software_counter(leader))
674 list_for_each_entry(counter, &leader->sibling_list, list_entry)
675 if (!is_software_counter(counter))
682 * Work out whether we can put this counter group on the CPU now.
684 static int group_can_go_on(struct perf_counter *counter,
685 struct perf_cpu_context *cpuctx,
689 * Groups consisting entirely of software counters can always go on.
691 if (is_software_only_group(counter))
694 * If an exclusive group is already on, no other hardware
695 * counters can go on.
697 if (cpuctx->exclusive)
700 * If this group is exclusive and there are already
701 * counters on the CPU, it can't go on.
703 if (counter->attr.exclusive && cpuctx->active_oncpu)
706 * Otherwise, try to add it if all previous groups were able
712 static void add_counter_to_ctx(struct perf_counter *counter,
713 struct perf_counter_context *ctx)
715 list_add_counter(counter, ctx);
716 counter->tstamp_enabled = ctx->time;
717 counter->tstamp_running = ctx->time;
718 counter->tstamp_stopped = ctx->time;
722 * Cross CPU call to install and enable a performance counter
724 * Must be called with ctx->mutex held
726 static void __perf_install_in_context(void *info)
728 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
729 struct perf_counter *counter = info;
730 struct perf_counter_context *ctx = counter->ctx;
731 struct perf_counter *leader = counter->group_leader;
732 int cpu = smp_processor_id();
736 * If this is a task context, we need to check whether it is
737 * the current task context of this cpu. If not it has been
738 * scheduled out before the smp call arrived.
739 * Or possibly this is the right context but it isn't
740 * on this cpu because it had no counters.
742 if (ctx->task && cpuctx->task_ctx != ctx) {
743 if (cpuctx->task_ctx || ctx->task != current)
745 cpuctx->task_ctx = ctx;
748 spin_lock(&ctx->lock);
750 update_context_time(ctx);
753 * Protect the list operation against NMI by disabling the
754 * counters on a global level. NOP for non NMI based counters.
758 add_counter_to_ctx(counter, ctx);
761 * Don't put the counter on if it is disabled or if
762 * it is in a group and the group isn't on.
764 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
765 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
769 * An exclusive counter can't go on if there are already active
770 * hardware counters, and no hardware counter can go on if there
771 * is already an exclusive counter on.
773 if (!group_can_go_on(counter, cpuctx, 1))
776 err = counter_sched_in(counter, cpuctx, ctx, cpu);
780 * This counter couldn't go on. If it is in a group
781 * then we have to pull the whole group off.
782 * If the counter group is pinned then put it in error state.
784 if (leader != counter)
785 group_sched_out(leader, cpuctx, ctx);
786 if (leader->attr.pinned) {
787 update_group_times(leader);
788 leader->state = PERF_COUNTER_STATE_ERROR;
792 if (!err && !ctx->task && cpuctx->max_pertask)
793 cpuctx->max_pertask--;
798 spin_unlock(&ctx->lock);
802 * Attach a performance counter to a context
804 * First we add the counter to the list with the hardware enable bit
805 * in counter->hw_config cleared.
807 * If the counter is attached to a task which is on a CPU we use a smp
808 * call to enable it in the task context. The task might have been
809 * scheduled away, but we check this in the smp call again.
811 * Must be called with ctx->mutex held.
814 perf_install_in_context(struct perf_counter_context *ctx,
815 struct perf_counter *counter,
818 struct task_struct *task = ctx->task;
822 * Per cpu counters are installed via an smp call and
823 * the install is always sucessful.
825 smp_call_function_single(cpu, __perf_install_in_context,
831 task_oncpu_function_call(task, __perf_install_in_context,
834 spin_lock_irq(&ctx->lock);
836 * we need to retry the smp call.
838 if (ctx->is_active && list_empty(&counter->list_entry)) {
839 spin_unlock_irq(&ctx->lock);
844 * The lock prevents that this context is scheduled in so we
845 * can add the counter safely, if it the call above did not
848 if (list_empty(&counter->list_entry))
849 add_counter_to_ctx(counter, ctx);
850 spin_unlock_irq(&ctx->lock);
854 * Cross CPU call to enable a performance counter
856 static void __perf_counter_enable(void *info)
858 struct perf_counter *counter = info;
859 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
860 struct perf_counter_context *ctx = counter->ctx;
861 struct perf_counter *leader = counter->group_leader;
865 * If this is a per-task counter, need to check whether this
866 * counter's task is the current task on this cpu.
868 if (ctx->task && cpuctx->task_ctx != ctx) {
869 if (cpuctx->task_ctx || ctx->task != current)
871 cpuctx->task_ctx = ctx;
874 spin_lock(&ctx->lock);
876 update_context_time(ctx);
878 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
880 counter->state = PERF_COUNTER_STATE_INACTIVE;
881 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
884 * If the counter is in a group and isn't the group leader,
885 * then don't put it on unless the group is on.
887 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
890 if (!group_can_go_on(counter, cpuctx, 1)) {
894 if (counter == leader)
895 err = group_sched_in(counter, cpuctx, ctx,
898 err = counter_sched_in(counter, cpuctx, ctx,
905 * If this counter can't go on and it's part of a
906 * group, then the whole group has to come off.
908 if (leader != counter)
909 group_sched_out(leader, cpuctx, ctx);
910 if (leader->attr.pinned) {
911 update_group_times(leader);
912 leader->state = PERF_COUNTER_STATE_ERROR;
917 spin_unlock(&ctx->lock);
923 * If counter->ctx is a cloned context, callers must make sure that
924 * every task struct that counter->ctx->task could possibly point to
925 * remains valid. This condition is satisfied when called through
926 * perf_counter_for_each_child or perf_counter_for_each as described
927 * for perf_counter_disable.
929 static void perf_counter_enable(struct perf_counter *counter)
931 struct perf_counter_context *ctx = counter->ctx;
932 struct task_struct *task = ctx->task;
936 * Enable the counter on the cpu that it's on
938 smp_call_function_single(counter->cpu, __perf_counter_enable,
943 spin_lock_irq(&ctx->lock);
944 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
948 * If the counter is in error state, clear that first.
949 * That way, if we see the counter in error state below, we
950 * know that it has gone back into error state, as distinct
951 * from the task having been scheduled away before the
952 * cross-call arrived.
954 if (counter->state == PERF_COUNTER_STATE_ERROR)
955 counter->state = PERF_COUNTER_STATE_OFF;
958 spin_unlock_irq(&ctx->lock);
959 task_oncpu_function_call(task, __perf_counter_enable, counter);
961 spin_lock_irq(&ctx->lock);
964 * If the context is active and the counter is still off,
965 * we need to retry the cross-call.
967 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
971 * Since we have the lock this context can't be scheduled
972 * in, so we can change the state safely.
974 if (counter->state == PERF_COUNTER_STATE_OFF) {
975 counter->state = PERF_COUNTER_STATE_INACTIVE;
976 counter->tstamp_enabled =
977 ctx->time - counter->total_time_enabled;
980 spin_unlock_irq(&ctx->lock);
983 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
986 * not supported on inherited counters
988 if (counter->attr.inherit)
991 atomic_add(refresh, &counter->event_limit);
992 perf_counter_enable(counter);
997 void __perf_counter_sched_out(struct perf_counter_context *ctx,
998 struct perf_cpu_context *cpuctx)
1000 struct perf_counter *counter;
1002 spin_lock(&ctx->lock);
1004 if (likely(!ctx->nr_counters))
1006 update_context_time(ctx);
1009 if (ctx->nr_active) {
1010 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1011 if (counter != counter->group_leader)
1012 counter_sched_out(counter, cpuctx, ctx);
1014 group_sched_out(counter, cpuctx, ctx);
1019 spin_unlock(&ctx->lock);
1023 * Test whether two contexts are equivalent, i.e. whether they
1024 * have both been cloned from the same version of the same context
1025 * and they both have the same number of enabled counters.
1026 * If the number of enabled counters is the same, then the set
1027 * of enabled counters should be the same, because these are both
1028 * inherited contexts, therefore we can't access individual counters
1029 * in them directly with an fd; we can only enable/disable all
1030 * counters via prctl, or enable/disable all counters in a family
1031 * via ioctl, which will have the same effect on both contexts.
1033 static int context_equiv(struct perf_counter_context *ctx1,
1034 struct perf_counter_context *ctx2)
1036 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1037 && ctx1->parent_gen == ctx2->parent_gen
1038 && !ctx1->pin_count && !ctx2->pin_count;
1041 static void __perf_counter_read(void *counter);
1043 static void __perf_counter_sync_stat(struct perf_counter *counter,
1044 struct perf_counter *next_counter)
1048 if (!counter->attr.inherit_stat)
1052 * Update the counter value, we cannot use perf_counter_read()
1053 * because we're in the middle of a context switch and have IRQs
1054 * disabled, which upsets smp_call_function_single(), however
1055 * we know the counter must be on the current CPU, therefore we
1056 * don't need to use it.
1058 switch (counter->state) {
1059 case PERF_COUNTER_STATE_ACTIVE:
1060 __perf_counter_read(counter);
1063 case PERF_COUNTER_STATE_INACTIVE:
1064 update_counter_times(counter);
1072 * In order to keep per-task stats reliable we need to flip the counter
1073 * values when we flip the contexts.
1075 value = atomic64_read(&next_counter->count);
1076 value = atomic64_xchg(&counter->count, value);
1077 atomic64_set(&next_counter->count, value);
1079 swap(counter->total_time_enabled, next_counter->total_time_enabled);
1080 swap(counter->total_time_running, next_counter->total_time_running);
1083 * Since we swizzled the values, update the user visible data too.
1085 perf_counter_update_userpage(counter);
1086 perf_counter_update_userpage(next_counter);
1089 #define list_next_entry(pos, member) \
1090 list_entry(pos->member.next, typeof(*pos), member)
1092 static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1093 struct perf_counter_context *next_ctx)
1095 struct perf_counter *counter, *next_counter;
1100 counter = list_first_entry(&ctx->event_list,
1101 struct perf_counter, event_entry);
1103 next_counter = list_first_entry(&next_ctx->event_list,
1104 struct perf_counter, event_entry);
1106 while (&counter->event_entry != &ctx->event_list &&
1107 &next_counter->event_entry != &next_ctx->event_list) {
1109 __perf_counter_sync_stat(counter, next_counter);
1111 counter = list_next_entry(counter, event_entry);
1112 next_counter = list_next_entry(next_counter, event_entry);
1117 * Called from scheduler to remove the counters of the current task,
1118 * with interrupts disabled.
1120 * We stop each counter and update the counter value in counter->count.
1122 * This does not protect us against NMI, but disable()
1123 * sets the disabled bit in the control field of counter _before_
1124 * accessing the counter control register. If a NMI hits, then it will
1125 * not restart the counter.
1127 void perf_counter_task_sched_out(struct task_struct *task,
1128 struct task_struct *next, int cpu)
1130 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1131 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1132 struct perf_counter_context *next_ctx;
1133 struct perf_counter_context *parent;
1134 struct pt_regs *regs;
1137 regs = task_pt_regs(task);
1138 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1140 if (likely(!ctx || !cpuctx->task_ctx))
1143 update_context_time(ctx);
1146 parent = rcu_dereference(ctx->parent_ctx);
1147 next_ctx = next->perf_counter_ctxp;
1148 if (parent && next_ctx &&
1149 rcu_dereference(next_ctx->parent_ctx) == parent) {
1151 * Looks like the two contexts are clones, so we might be
1152 * able to optimize the context switch. We lock both
1153 * contexts and check that they are clones under the
1154 * lock (including re-checking that neither has been
1155 * uncloned in the meantime). It doesn't matter which
1156 * order we take the locks because no other cpu could
1157 * be trying to lock both of these tasks.
1159 spin_lock(&ctx->lock);
1160 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1161 if (context_equiv(ctx, next_ctx)) {
1163 * XXX do we need a memory barrier of sorts
1164 * wrt to rcu_dereference() of perf_counter_ctxp
1166 task->perf_counter_ctxp = next_ctx;
1167 next->perf_counter_ctxp = ctx;
1169 next_ctx->task = task;
1172 perf_counter_sync_stat(ctx, next_ctx);
1174 spin_unlock(&next_ctx->lock);
1175 spin_unlock(&ctx->lock);
1180 __perf_counter_sched_out(ctx, cpuctx);
1181 cpuctx->task_ctx = NULL;
1186 * Called with IRQs disabled
1188 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1190 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1192 if (!cpuctx->task_ctx)
1195 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1198 __perf_counter_sched_out(ctx, cpuctx);
1199 cpuctx->task_ctx = NULL;
1203 * Called with IRQs disabled
1205 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1207 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1211 __perf_counter_sched_in(struct perf_counter_context *ctx,
1212 struct perf_cpu_context *cpuctx, int cpu)
1214 struct perf_counter *counter;
1217 spin_lock(&ctx->lock);
1219 if (likely(!ctx->nr_counters))
1222 ctx->timestamp = perf_clock();
1227 * First go through the list and put on any pinned groups
1228 * in order to give them the best chance of going on.
1230 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1231 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1232 !counter->attr.pinned)
1234 if (counter->cpu != -1 && counter->cpu != cpu)
1237 if (counter != counter->group_leader)
1238 counter_sched_in(counter, cpuctx, ctx, cpu);
1240 if (group_can_go_on(counter, cpuctx, 1))
1241 group_sched_in(counter, cpuctx, ctx, cpu);
1245 * If this pinned group hasn't been scheduled,
1246 * put it in error state.
1248 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1249 update_group_times(counter);
1250 counter->state = PERF_COUNTER_STATE_ERROR;
1254 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1256 * Ignore counters in OFF or ERROR state, and
1257 * ignore pinned counters since we did them already.
1259 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1260 counter->attr.pinned)
1264 * Listen to the 'cpu' scheduling filter constraint
1267 if (counter->cpu != -1 && counter->cpu != cpu)
1270 if (counter != counter->group_leader) {
1271 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1274 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1275 if (group_sched_in(counter, cpuctx, ctx, cpu))
1282 spin_unlock(&ctx->lock);
1286 * Called from scheduler to add the counters of the current task
1287 * with interrupts disabled.
1289 * We restore the counter value and then enable it.
1291 * This does not protect us against NMI, but enable()
1292 * sets the enabled bit in the control field of counter _before_
1293 * accessing the counter control register. If a NMI hits, then it will
1294 * keep the counter running.
1296 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1298 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1299 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1303 if (cpuctx->task_ctx == ctx)
1305 __perf_counter_sched_in(ctx, cpuctx, cpu);
1306 cpuctx->task_ctx = ctx;
1309 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1311 struct perf_counter_context *ctx = &cpuctx->ctx;
1313 __perf_counter_sched_in(ctx, cpuctx, cpu);
1316 #define MAX_INTERRUPTS (~0ULL)
1318 static void perf_log_throttle(struct perf_counter *counter, int enable);
1320 static void perf_adjust_period(struct perf_counter *counter, u64 events)
1322 struct hw_perf_counter *hwc = &counter->hw;
1323 u64 period, sample_period;
1326 events *= hwc->sample_period;
1327 period = div64_u64(events, counter->attr.sample_freq);
1329 delta = (s64)(period - hwc->sample_period);
1330 delta = (delta + 7) / 8; /* low pass filter */
1332 sample_period = hwc->sample_period + delta;
1337 hwc->sample_period = sample_period;
1340 static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1342 struct perf_counter *counter;
1343 struct hw_perf_counter *hwc;
1344 u64 interrupts, freq;
1346 spin_lock(&ctx->lock);
1347 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1348 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1353 interrupts = hwc->interrupts;
1354 hwc->interrupts = 0;
1357 * unthrottle counters on the tick
1359 if (interrupts == MAX_INTERRUPTS) {
1360 perf_log_throttle(counter, 1);
1361 counter->pmu->unthrottle(counter);
1362 interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1365 if (!counter->attr.freq || !counter->attr.sample_freq)
1369 * if the specified freq < HZ then we need to skip ticks
1371 if (counter->attr.sample_freq < HZ) {
1372 freq = counter->attr.sample_freq;
1374 hwc->freq_count += freq;
1375 hwc->freq_interrupts += interrupts;
1377 if (hwc->freq_count < HZ)
1380 interrupts = hwc->freq_interrupts;
1381 hwc->freq_interrupts = 0;
1382 hwc->freq_count -= HZ;
1386 perf_adjust_period(counter, freq * interrupts);
1389 * In order to avoid being stalled by an (accidental) huge
1390 * sample period, force reset the sample period if we didn't
1391 * get any events in this freq period.
1395 counter->pmu->disable(counter);
1396 atomic64_set(&hwc->period_left, 0);
1397 counter->pmu->enable(counter);
1401 spin_unlock(&ctx->lock);
1405 * Round-robin a context's counters:
1407 static void rotate_ctx(struct perf_counter_context *ctx)
1409 struct perf_counter *counter;
1411 if (!ctx->nr_counters)
1414 spin_lock(&ctx->lock);
1416 * Rotate the first entry last (works just fine for group counters too):
1419 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1420 list_move_tail(&counter->list_entry, &ctx->counter_list);
1425 spin_unlock(&ctx->lock);
1428 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1430 struct perf_cpu_context *cpuctx;
1431 struct perf_counter_context *ctx;
1433 if (!atomic_read(&nr_counters))
1436 cpuctx = &per_cpu(perf_cpu_context, cpu);
1437 ctx = curr->perf_counter_ctxp;
1439 perf_ctx_adjust_freq(&cpuctx->ctx);
1441 perf_ctx_adjust_freq(ctx);
1443 perf_counter_cpu_sched_out(cpuctx);
1445 __perf_counter_task_sched_out(ctx);
1447 rotate_ctx(&cpuctx->ctx);
1451 perf_counter_cpu_sched_in(cpuctx, cpu);
1453 perf_counter_task_sched_in(curr, cpu);
1457 * Enable all of a task's counters that have been marked enable-on-exec.
1458 * This expects task == current.
1460 static void perf_counter_enable_on_exec(struct task_struct *task)
1462 struct perf_counter_context *ctx;
1463 struct perf_counter *counter;
1464 unsigned long flags;
1467 local_irq_save(flags);
1468 ctx = task->perf_counter_ctxp;
1469 if (!ctx || !ctx->nr_counters)
1472 __perf_counter_task_sched_out(ctx);
1474 spin_lock(&ctx->lock);
1476 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1477 if (!counter->attr.enable_on_exec)
1479 counter->attr.enable_on_exec = 0;
1480 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
1482 counter->state = PERF_COUNTER_STATE_INACTIVE;
1483 counter->tstamp_enabled =
1484 ctx->time - counter->total_time_enabled;
1489 * Unclone this context if we enabled any counter.
1494 spin_unlock(&ctx->lock);
1496 perf_counter_task_sched_in(task, smp_processor_id());
1498 local_irq_restore(flags);
1502 * Cross CPU call to read the hardware counter
1504 static void __perf_counter_read(void *info)
1506 struct perf_counter *counter = info;
1507 struct perf_counter_context *ctx = counter->ctx;
1508 unsigned long flags;
1510 local_irq_save(flags);
1512 update_context_time(ctx);
1513 counter->pmu->read(counter);
1514 update_counter_times(counter);
1515 local_irq_restore(flags);
1518 static u64 perf_counter_read(struct perf_counter *counter)
1521 * If counter is enabled and currently active on a CPU, update the
1522 * value in the counter structure:
1524 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1525 smp_call_function_single(counter->oncpu,
1526 __perf_counter_read, counter, 1);
1527 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1528 update_counter_times(counter);
1531 return atomic64_read(&counter->count);
1535 * Initialize the perf_counter context in a task_struct:
1538 __perf_counter_init_context(struct perf_counter_context *ctx,
1539 struct task_struct *task)
1541 memset(ctx, 0, sizeof(*ctx));
1542 spin_lock_init(&ctx->lock);
1543 mutex_init(&ctx->mutex);
1544 INIT_LIST_HEAD(&ctx->counter_list);
1545 INIT_LIST_HEAD(&ctx->event_list);
1546 atomic_set(&ctx->refcount, 1);
1550 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1552 struct perf_counter_context *ctx;
1553 struct perf_cpu_context *cpuctx;
1554 struct task_struct *task;
1555 unsigned long flags;
1559 * If cpu is not a wildcard then this is a percpu counter:
1562 /* Must be root to operate on a CPU counter: */
1563 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1564 return ERR_PTR(-EACCES);
1566 if (cpu < 0 || cpu > num_possible_cpus())
1567 return ERR_PTR(-EINVAL);
1570 * We could be clever and allow to attach a counter to an
1571 * offline CPU and activate it when the CPU comes up, but
1574 if (!cpu_isset(cpu, cpu_online_map))
1575 return ERR_PTR(-ENODEV);
1577 cpuctx = &per_cpu(perf_cpu_context, cpu);
1588 task = find_task_by_vpid(pid);
1590 get_task_struct(task);
1594 return ERR_PTR(-ESRCH);
1597 * Can't attach counters to a dying task.
1600 if (task->flags & PF_EXITING)
1603 /* Reuse ptrace permission checks for now. */
1605 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1609 ctx = perf_lock_task_context(task, &flags);
1612 spin_unlock_irqrestore(&ctx->lock, flags);
1616 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1620 __perf_counter_init_context(ctx, task);
1622 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1624 * We raced with some other task; use
1625 * the context they set.
1630 get_task_struct(task);
1633 put_task_struct(task);
1637 put_task_struct(task);
1638 return ERR_PTR(err);
1641 static void free_counter_rcu(struct rcu_head *head)
1643 struct perf_counter *counter;
1645 counter = container_of(head, struct perf_counter, rcu_head);
1647 put_pid_ns(counter->ns);
1651 static void perf_pending_sync(struct perf_counter *counter);
1653 static void free_counter(struct perf_counter *counter)
1655 perf_pending_sync(counter);
1657 if (!counter->parent) {
1658 atomic_dec(&nr_counters);
1659 if (counter->attr.mmap)
1660 atomic_dec(&nr_mmap_counters);
1661 if (counter->attr.comm)
1662 atomic_dec(&nr_comm_counters);
1663 if (counter->attr.task)
1664 atomic_dec(&nr_task_counters);
1667 if (counter->destroy)
1668 counter->destroy(counter);
1670 put_ctx(counter->ctx);
1671 call_rcu(&counter->rcu_head, free_counter_rcu);
1675 * Called when the last reference to the file is gone.
1677 static int perf_release(struct inode *inode, struct file *file)
1679 struct perf_counter *counter = file->private_data;
1680 struct perf_counter_context *ctx = counter->ctx;
1682 file->private_data = NULL;
1684 WARN_ON_ONCE(ctx->parent_ctx);
1685 mutex_lock(&ctx->mutex);
1686 perf_counter_remove_from_context(counter);
1687 mutex_unlock(&ctx->mutex);
1689 mutex_lock(&counter->owner->perf_counter_mutex);
1690 list_del_init(&counter->owner_entry);
1691 mutex_unlock(&counter->owner->perf_counter_mutex);
1692 put_task_struct(counter->owner);
1694 free_counter(counter);
1699 static int perf_counter_read_size(struct perf_counter *counter)
1701 int entry = sizeof(u64); /* value */
1705 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1706 size += sizeof(u64);
1708 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1709 size += sizeof(u64);
1711 if (counter->attr.read_format & PERF_FORMAT_ID)
1712 entry += sizeof(u64);
1714 if (counter->attr.read_format & PERF_FORMAT_GROUP) {
1715 nr += counter->group_leader->nr_siblings;
1716 size += sizeof(u64);
1724 static u64 perf_counter_read_value(struct perf_counter *counter)
1726 struct perf_counter *child;
1729 total += perf_counter_read(counter);
1730 list_for_each_entry(child, &counter->child_list, child_list)
1731 total += perf_counter_read(child);
1736 static int perf_counter_read_entry(struct perf_counter *counter,
1737 u64 read_format, char __user *buf)
1739 int n = 0, count = 0;
1742 values[n++] = perf_counter_read_value(counter);
1743 if (read_format & PERF_FORMAT_ID)
1744 values[n++] = primary_counter_id(counter);
1746 count = n * sizeof(u64);
1748 if (copy_to_user(buf, values, count))
1754 static int perf_counter_read_group(struct perf_counter *counter,
1755 u64 read_format, char __user *buf)
1757 struct perf_counter *leader = counter->group_leader, *sub;
1758 int n = 0, size = 0, err = -EFAULT;
1761 values[n++] = 1 + leader->nr_siblings;
1762 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1763 values[n++] = leader->total_time_enabled +
1764 atomic64_read(&leader->child_total_time_enabled);
1766 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1767 values[n++] = leader->total_time_running +
1768 atomic64_read(&leader->child_total_time_running);
1771 size = n * sizeof(u64);
1773 if (copy_to_user(buf, values, size))
1776 err = perf_counter_read_entry(leader, read_format, buf + size);
1782 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1783 err = perf_counter_read_entry(counter, read_format,
1794 static int perf_counter_read_one(struct perf_counter *counter,
1795 u64 read_format, char __user *buf)
1800 values[n++] = perf_counter_read_value(counter);
1801 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1802 values[n++] = counter->total_time_enabled +
1803 atomic64_read(&counter->child_total_time_enabled);
1805 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1806 values[n++] = counter->total_time_running +
1807 atomic64_read(&counter->child_total_time_running);
1809 if (read_format & PERF_FORMAT_ID)
1810 values[n++] = primary_counter_id(counter);
1812 if (copy_to_user(buf, values, n * sizeof(u64)))
1815 return n * sizeof(u64);
1819 * Read the performance counter - simple non blocking version for now
1822 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1824 u64 read_format = counter->attr.read_format;
1828 * Return end-of-file for a read on a counter that is in
1829 * error state (i.e. because it was pinned but it couldn't be
1830 * scheduled on to the CPU at some point).
1832 if (counter->state == PERF_COUNTER_STATE_ERROR)
1835 if (count < perf_counter_read_size(counter))
1838 WARN_ON_ONCE(counter->ctx->parent_ctx);
1839 mutex_lock(&counter->child_mutex);
1840 if (read_format & PERF_FORMAT_GROUP)
1841 ret = perf_counter_read_group(counter, read_format, buf);
1843 ret = perf_counter_read_one(counter, read_format, buf);
1844 mutex_unlock(&counter->child_mutex);
1850 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1852 struct perf_counter *counter = file->private_data;
1854 return perf_read_hw(counter, buf, count);
1857 static unsigned int perf_poll(struct file *file, poll_table *wait)
1859 struct perf_counter *counter = file->private_data;
1860 struct perf_mmap_data *data;
1861 unsigned int events = POLL_HUP;
1864 data = rcu_dereference(counter->data);
1866 events = atomic_xchg(&data->poll, 0);
1869 poll_wait(file, &counter->waitq, wait);
1874 static void perf_counter_reset(struct perf_counter *counter)
1876 (void)perf_counter_read(counter);
1877 atomic64_set(&counter->count, 0);
1878 perf_counter_update_userpage(counter);
1882 * Holding the top-level counter's child_mutex means that any
1883 * descendant process that has inherited this counter will block
1884 * in sync_child_counter if it goes to exit, thus satisfying the
1885 * task existence requirements of perf_counter_enable/disable.
1887 static void perf_counter_for_each_child(struct perf_counter *counter,
1888 void (*func)(struct perf_counter *))
1890 struct perf_counter *child;
1892 WARN_ON_ONCE(counter->ctx->parent_ctx);
1893 mutex_lock(&counter->child_mutex);
1895 list_for_each_entry(child, &counter->child_list, child_list)
1897 mutex_unlock(&counter->child_mutex);
1900 static void perf_counter_for_each(struct perf_counter *counter,
1901 void (*func)(struct perf_counter *))
1903 struct perf_counter_context *ctx = counter->ctx;
1904 struct perf_counter *sibling;
1906 WARN_ON_ONCE(ctx->parent_ctx);
1907 mutex_lock(&ctx->mutex);
1908 counter = counter->group_leader;
1910 perf_counter_for_each_child(counter, func);
1912 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1913 perf_counter_for_each_child(counter, func);
1914 mutex_unlock(&ctx->mutex);
1917 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1919 struct perf_counter_context *ctx = counter->ctx;
1924 if (!counter->attr.sample_period)
1927 size = copy_from_user(&value, arg, sizeof(value));
1928 if (size != sizeof(value))
1934 spin_lock_irq(&ctx->lock);
1935 if (counter->attr.freq) {
1936 if (value > sysctl_perf_counter_sample_rate) {
1941 counter->attr.sample_freq = value;
1943 counter->attr.sample_period = value;
1944 counter->hw.sample_period = value;
1947 spin_unlock_irq(&ctx->lock);
1952 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1954 struct perf_counter *counter = file->private_data;
1955 void (*func)(struct perf_counter *);
1959 case PERF_COUNTER_IOC_ENABLE:
1960 func = perf_counter_enable;
1962 case PERF_COUNTER_IOC_DISABLE:
1963 func = perf_counter_disable;
1965 case PERF_COUNTER_IOC_RESET:
1966 func = perf_counter_reset;
1969 case PERF_COUNTER_IOC_REFRESH:
1970 return perf_counter_refresh(counter, arg);
1972 case PERF_COUNTER_IOC_PERIOD:
1973 return perf_counter_period(counter, (u64 __user *)arg);
1979 if (flags & PERF_IOC_FLAG_GROUP)
1980 perf_counter_for_each(counter, func);
1982 perf_counter_for_each_child(counter, func);
1987 int perf_counter_task_enable(void)
1989 struct perf_counter *counter;
1991 mutex_lock(¤t->perf_counter_mutex);
1992 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1993 perf_counter_for_each_child(counter, perf_counter_enable);
1994 mutex_unlock(¤t->perf_counter_mutex);
1999 int perf_counter_task_disable(void)
2001 struct perf_counter *counter;
2003 mutex_lock(¤t->perf_counter_mutex);
2004 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
2005 perf_counter_for_each_child(counter, perf_counter_disable);
2006 mutex_unlock(¤t->perf_counter_mutex);
2011 static int perf_counter_index(struct perf_counter *counter)
2013 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2016 return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
2020 * Callers need to ensure there can be no nesting of this function, otherwise
2021 * the seqlock logic goes bad. We can not serialize this because the arch
2022 * code calls this from NMI context.
2024 void perf_counter_update_userpage(struct perf_counter *counter)
2026 struct perf_counter_mmap_page *userpg;
2027 struct perf_mmap_data *data;
2030 data = rcu_dereference(counter->data);
2034 userpg = data->user_page;
2037 * Disable preemption so as to not let the corresponding user-space
2038 * spin too long if we get preempted.
2043 userpg->index = perf_counter_index(counter);
2044 userpg->offset = atomic64_read(&counter->count);
2045 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
2046 userpg->offset -= atomic64_read(&counter->hw.prev_count);
2048 userpg->time_enabled = counter->total_time_enabled +
2049 atomic64_read(&counter->child_total_time_enabled);
2051 userpg->time_running = counter->total_time_running +
2052 atomic64_read(&counter->child_total_time_running);
2061 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2063 struct perf_counter *counter = vma->vm_file->private_data;
2064 struct perf_mmap_data *data;
2065 int ret = VM_FAULT_SIGBUS;
2067 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2068 if (vmf->pgoff == 0)
2074 data = rcu_dereference(counter->data);
2078 if (vmf->pgoff == 0) {
2079 vmf->page = virt_to_page(data->user_page);
2081 int nr = vmf->pgoff - 1;
2083 if ((unsigned)nr > data->nr_pages)
2086 if (vmf->flags & FAULT_FLAG_WRITE)
2089 vmf->page = virt_to_page(data->data_pages[nr]);
2092 get_page(vmf->page);
2093 vmf->page->mapping = vma->vm_file->f_mapping;
2094 vmf->page->index = vmf->pgoff;
2103 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
2105 struct perf_mmap_data *data;
2109 WARN_ON(atomic_read(&counter->mmap_count));
2111 size = sizeof(struct perf_mmap_data);
2112 size += nr_pages * sizeof(void *);
2114 data = kzalloc(size, GFP_KERNEL);
2118 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2119 if (!data->user_page)
2120 goto fail_user_page;
2122 for (i = 0; i < nr_pages; i++) {
2123 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2124 if (!data->data_pages[i])
2125 goto fail_data_pages;
2128 data->nr_pages = nr_pages;
2129 atomic_set(&data->lock, -1);
2131 rcu_assign_pointer(counter->data, data);
2136 for (i--; i >= 0; i--)
2137 free_page((unsigned long)data->data_pages[i]);
2139 free_page((unsigned long)data->user_page);
2148 static void perf_mmap_free_page(unsigned long addr)
2150 struct page *page = virt_to_page((void *)addr);
2152 page->mapping = NULL;
2156 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2158 struct perf_mmap_data *data;
2161 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2163 perf_mmap_free_page((unsigned long)data->user_page);
2164 for (i = 0; i < data->nr_pages; i++)
2165 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2170 static void perf_mmap_data_free(struct perf_counter *counter)
2172 struct perf_mmap_data *data = counter->data;
2174 WARN_ON(atomic_read(&counter->mmap_count));
2176 rcu_assign_pointer(counter->data, NULL);
2177 call_rcu(&data->rcu_head, __perf_mmap_data_free);
2180 static void perf_mmap_open(struct vm_area_struct *vma)
2182 struct perf_counter *counter = vma->vm_file->private_data;
2184 atomic_inc(&counter->mmap_count);
2187 static void perf_mmap_close(struct vm_area_struct *vma)
2189 struct perf_counter *counter = vma->vm_file->private_data;
2191 WARN_ON_ONCE(counter->ctx->parent_ctx);
2192 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2193 struct user_struct *user = current_user();
2195 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2196 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2197 perf_mmap_data_free(counter);
2198 mutex_unlock(&counter->mmap_mutex);
2202 static struct vm_operations_struct perf_mmap_vmops = {
2203 .open = perf_mmap_open,
2204 .close = perf_mmap_close,
2205 .fault = perf_mmap_fault,
2206 .page_mkwrite = perf_mmap_fault,
2209 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2211 struct perf_counter *counter = file->private_data;
2212 unsigned long user_locked, user_lock_limit;
2213 struct user_struct *user = current_user();
2214 unsigned long locked, lock_limit;
2215 unsigned long vma_size;
2216 unsigned long nr_pages;
2217 long user_extra, extra;
2220 if (!(vma->vm_flags & VM_SHARED))
2223 vma_size = vma->vm_end - vma->vm_start;
2224 nr_pages = (vma_size / PAGE_SIZE) - 1;
2227 * If we have data pages ensure they're a power-of-two number, so we
2228 * can do bitmasks instead of modulo.
2230 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2233 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2236 if (vma->vm_pgoff != 0)
2239 WARN_ON_ONCE(counter->ctx->parent_ctx);
2240 mutex_lock(&counter->mmap_mutex);
2241 if (atomic_inc_not_zero(&counter->mmap_count)) {
2242 if (nr_pages != counter->data->nr_pages)
2247 user_extra = nr_pages + 1;
2248 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2251 * Increase the limit linearly with more CPUs:
2253 user_lock_limit *= num_online_cpus();
2255 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2258 if (user_locked > user_lock_limit)
2259 extra = user_locked - user_lock_limit;
2261 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2262 lock_limit >>= PAGE_SHIFT;
2263 locked = vma->vm_mm->locked_vm + extra;
2265 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2270 WARN_ON(counter->data);
2271 ret = perf_mmap_data_alloc(counter, nr_pages);
2275 atomic_set(&counter->mmap_count, 1);
2276 atomic_long_add(user_extra, &user->locked_vm);
2277 vma->vm_mm->locked_vm += extra;
2278 counter->data->nr_locked = extra;
2279 if (vma->vm_flags & VM_WRITE)
2280 counter->data->writable = 1;
2283 mutex_unlock(&counter->mmap_mutex);
2285 vma->vm_flags |= VM_RESERVED;
2286 vma->vm_ops = &perf_mmap_vmops;
2291 static int perf_fasync(int fd, struct file *filp, int on)
2293 struct inode *inode = filp->f_path.dentry->d_inode;
2294 struct perf_counter *counter = filp->private_data;
2297 mutex_lock(&inode->i_mutex);
2298 retval = fasync_helper(fd, filp, on, &counter->fasync);
2299 mutex_unlock(&inode->i_mutex);
2307 static const struct file_operations perf_fops = {
2308 .release = perf_release,
2311 .unlocked_ioctl = perf_ioctl,
2312 .compat_ioctl = perf_ioctl,
2314 .fasync = perf_fasync,
2318 * Perf counter wakeup
2320 * If there's data, ensure we set the poll() state and publish everything
2321 * to user-space before waking everybody up.
2324 void perf_counter_wakeup(struct perf_counter *counter)
2326 wake_up_all(&counter->waitq);
2328 if (counter->pending_kill) {
2329 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2330 counter->pending_kill = 0;
2337 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2339 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2340 * single linked list and use cmpxchg() to add entries lockless.
2343 static void perf_pending_counter(struct perf_pending_entry *entry)
2345 struct perf_counter *counter = container_of(entry,
2346 struct perf_counter, pending);
2348 if (counter->pending_disable) {
2349 counter->pending_disable = 0;
2350 __perf_counter_disable(counter);
2353 if (counter->pending_wakeup) {
2354 counter->pending_wakeup = 0;
2355 perf_counter_wakeup(counter);
2359 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2361 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2365 static void perf_pending_queue(struct perf_pending_entry *entry,
2366 void (*func)(struct perf_pending_entry *))
2368 struct perf_pending_entry **head;
2370 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2375 head = &get_cpu_var(perf_pending_head);
2378 entry->next = *head;
2379 } while (cmpxchg(head, entry->next, entry) != entry->next);
2381 set_perf_counter_pending();
2383 put_cpu_var(perf_pending_head);
2386 static int __perf_pending_run(void)
2388 struct perf_pending_entry *list;
2391 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2392 while (list != PENDING_TAIL) {
2393 void (*func)(struct perf_pending_entry *);
2394 struct perf_pending_entry *entry = list;
2401 * Ensure we observe the unqueue before we issue the wakeup,
2402 * so that we won't be waiting forever.
2403 * -- see perf_not_pending().
2414 static inline int perf_not_pending(struct perf_counter *counter)
2417 * If we flush on whatever cpu we run, there is a chance we don't
2421 __perf_pending_run();
2425 * Ensure we see the proper queue state before going to sleep
2426 * so that we do not miss the wakeup. -- see perf_pending_handle()
2429 return counter->pending.next == NULL;
2432 static void perf_pending_sync(struct perf_counter *counter)
2434 wait_event(counter->waitq, perf_not_pending(counter));
2437 void perf_counter_do_pending(void)
2439 __perf_pending_run();
2443 * Callchain support -- arch specific
2446 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2455 struct perf_output_handle {
2456 struct perf_counter *counter;
2457 struct perf_mmap_data *data;
2459 unsigned long offset;
2463 unsigned long flags;
2466 static bool perf_output_space(struct perf_mmap_data *data,
2467 unsigned int offset, unsigned int head)
2472 if (!data->writable)
2475 mask = (data->nr_pages << PAGE_SHIFT) - 1;
2477 * Userspace could choose to issue a mb() before updating the tail
2478 * pointer. So that all reads will be completed before the write is
2481 tail = ACCESS_ONCE(data->user_page->data_tail);
2484 offset = (offset - tail) & mask;
2485 head = (head - tail) & mask;
2487 if ((int)(head - offset) < 0)
2493 static void perf_output_wakeup(struct perf_output_handle *handle)
2495 atomic_set(&handle->data->poll, POLL_IN);
2498 handle->counter->pending_wakeup = 1;
2499 perf_pending_queue(&handle->counter->pending,
2500 perf_pending_counter);
2502 perf_counter_wakeup(handle->counter);
2506 * Curious locking construct.
2508 * We need to ensure a later event doesn't publish a head when a former
2509 * event isn't done writing. However since we need to deal with NMIs we
2510 * cannot fully serialize things.
2512 * What we do is serialize between CPUs so we only have to deal with NMI
2513 * nesting on a single CPU.
2515 * We only publish the head (and generate a wakeup) when the outer-most
2518 static void perf_output_lock(struct perf_output_handle *handle)
2520 struct perf_mmap_data *data = handle->data;
2525 local_irq_save(handle->flags);
2526 cpu = smp_processor_id();
2528 if (in_nmi() && atomic_read(&data->lock) == cpu)
2531 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2537 static void perf_output_unlock(struct perf_output_handle *handle)
2539 struct perf_mmap_data *data = handle->data;
2543 data->done_head = data->head;
2545 if (!handle->locked)
2550 * The xchg implies a full barrier that ensures all writes are done
2551 * before we publish the new head, matched by a rmb() in userspace when
2552 * reading this position.
2554 while ((head = atomic_long_xchg(&data->done_head, 0)))
2555 data->user_page->data_head = head;
2558 * NMI can happen here, which means we can miss a done_head update.
2561 cpu = atomic_xchg(&data->lock, -1);
2562 WARN_ON_ONCE(cpu != smp_processor_id());
2565 * Therefore we have to validate we did not indeed do so.
2567 if (unlikely(atomic_long_read(&data->done_head))) {
2569 * Since we had it locked, we can lock it again.
2571 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2577 if (atomic_xchg(&data->wakeup, 0))
2578 perf_output_wakeup(handle);
2580 local_irq_restore(handle->flags);
2583 static void perf_output_copy(struct perf_output_handle *handle,
2584 const void *buf, unsigned int len)
2586 unsigned int pages_mask;
2587 unsigned int offset;
2591 offset = handle->offset;
2592 pages_mask = handle->data->nr_pages - 1;
2593 pages = handle->data->data_pages;
2596 unsigned int page_offset;
2599 nr = (offset >> PAGE_SHIFT) & pages_mask;
2600 page_offset = offset & (PAGE_SIZE - 1);
2601 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2603 memcpy(pages[nr] + page_offset, buf, size);
2610 handle->offset = offset;
2613 * Check we didn't copy past our reservation window, taking the
2614 * possible unsigned int wrap into account.
2616 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2619 #define perf_output_put(handle, x) \
2620 perf_output_copy((handle), &(x), sizeof(x))
2622 static int perf_output_begin(struct perf_output_handle *handle,
2623 struct perf_counter *counter, unsigned int size,
2624 int nmi, int sample)
2626 struct perf_mmap_data *data;
2627 unsigned int offset, head;
2630 struct perf_event_header header;
2636 * For inherited counters we send all the output towards the parent.
2638 if (counter->parent)
2639 counter = counter->parent;
2642 data = rcu_dereference(counter->data);
2646 handle->data = data;
2647 handle->counter = counter;
2649 handle->sample = sample;
2651 if (!data->nr_pages)
2654 have_lost = atomic_read(&data->lost);
2656 size += sizeof(lost_event);
2658 perf_output_lock(handle);
2661 offset = head = atomic_long_read(&data->head);
2663 if (unlikely(!perf_output_space(data, offset, head)))
2665 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2667 handle->offset = offset;
2668 handle->head = head;
2670 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2671 atomic_set(&data->wakeup, 1);
2674 lost_event.header.type = PERF_EVENT_LOST;
2675 lost_event.header.misc = 0;
2676 lost_event.header.size = sizeof(lost_event);
2677 lost_event.id = counter->id;
2678 lost_event.lost = atomic_xchg(&data->lost, 0);
2680 perf_output_put(handle, lost_event);
2686 atomic_inc(&data->lost);
2687 perf_output_unlock(handle);
2694 static void perf_output_end(struct perf_output_handle *handle)
2696 struct perf_counter *counter = handle->counter;
2697 struct perf_mmap_data *data = handle->data;
2699 int wakeup_events = counter->attr.wakeup_events;
2701 if (handle->sample && wakeup_events) {
2702 int events = atomic_inc_return(&data->events);
2703 if (events >= wakeup_events) {
2704 atomic_sub(wakeup_events, &data->events);
2705 atomic_set(&data->wakeup, 1);
2709 perf_output_unlock(handle);
2713 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2716 * only top level counters have the pid namespace they were created in
2718 if (counter->parent)
2719 counter = counter->parent;
2721 return task_tgid_nr_ns(p, counter->ns);
2724 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2727 * only top level counters have the pid namespace they were created in
2729 if (counter->parent)
2730 counter = counter->parent;
2732 return task_pid_nr_ns(p, counter->ns);
2735 static void perf_output_read_one(struct perf_output_handle *handle,
2736 struct perf_counter *counter)
2738 u64 read_format = counter->attr.read_format;
2742 values[n++] = atomic64_read(&counter->count);
2743 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2744 values[n++] = counter->total_time_enabled +
2745 atomic64_read(&counter->child_total_time_enabled);
2747 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2748 values[n++] = counter->total_time_running +
2749 atomic64_read(&counter->child_total_time_running);
2751 if (read_format & PERF_FORMAT_ID)
2752 values[n++] = primary_counter_id(counter);
2754 perf_output_copy(handle, values, n * sizeof(u64));
2758 * XXX PERF_FORMAT_GROUP vs inherited counters seems difficult.
2760 static void perf_output_read_group(struct perf_output_handle *handle,
2761 struct perf_counter *counter)
2763 struct perf_counter *leader = counter->group_leader, *sub;
2764 u64 read_format = counter->attr.read_format;
2768 values[n++] = 1 + leader->nr_siblings;
2770 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2771 values[n++] = leader->total_time_enabled;
2773 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2774 values[n++] = leader->total_time_running;
2776 if (leader != counter)
2777 leader->pmu->read(leader);
2779 values[n++] = atomic64_read(&leader->count);
2780 if (read_format & PERF_FORMAT_ID)
2781 values[n++] = primary_counter_id(leader);
2783 perf_output_copy(handle, values, n * sizeof(u64));
2785 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2789 sub->pmu->read(sub);
2791 values[n++] = atomic64_read(&sub->count);
2792 if (read_format & PERF_FORMAT_ID)
2793 values[n++] = primary_counter_id(sub);
2795 perf_output_copy(handle, values, n * sizeof(u64));
2799 static void perf_output_read(struct perf_output_handle *handle,
2800 struct perf_counter *counter)
2802 if (counter->attr.read_format & PERF_FORMAT_GROUP)
2803 perf_output_read_group(handle, counter);
2805 perf_output_read_one(handle, counter);
2808 void perf_counter_output(struct perf_counter *counter, int nmi,
2809 struct perf_sample_data *data)
2812 u64 sample_type = counter->attr.sample_type;
2813 struct perf_output_handle handle;
2814 struct perf_event_header header;
2819 struct perf_callchain_entry *callchain = NULL;
2820 int callchain_size = 0;
2826 header.type = PERF_EVENT_SAMPLE;
2827 header.size = sizeof(header);
2830 header.misc |= perf_misc_flags(data->regs);
2832 if (sample_type & PERF_SAMPLE_IP) {
2833 ip = perf_instruction_pointer(data->regs);
2834 header.size += sizeof(ip);
2837 if (sample_type & PERF_SAMPLE_TID) {
2838 /* namespace issues */
2839 tid_entry.pid = perf_counter_pid(counter, current);
2840 tid_entry.tid = perf_counter_tid(counter, current);
2842 header.size += sizeof(tid_entry);
2845 if (sample_type & PERF_SAMPLE_TIME) {
2847 * Maybe do better on x86 and provide cpu_clock_nmi()
2849 time = sched_clock();
2851 header.size += sizeof(u64);
2854 if (sample_type & PERF_SAMPLE_ADDR)
2855 header.size += sizeof(u64);
2857 if (sample_type & PERF_SAMPLE_ID)
2858 header.size += sizeof(u64);
2860 if (sample_type & PERF_SAMPLE_STREAM_ID)
2861 header.size += sizeof(u64);
2863 if (sample_type & PERF_SAMPLE_CPU) {
2864 header.size += sizeof(cpu_entry);
2866 cpu_entry.cpu = raw_smp_processor_id();
2867 cpu_entry.reserved = 0;
2870 if (sample_type & PERF_SAMPLE_PERIOD)
2871 header.size += sizeof(u64);
2873 if (sample_type & PERF_SAMPLE_READ)
2874 header.size += perf_counter_read_size(counter);
2876 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2877 callchain = perf_callchain(data->regs);
2880 callchain_size = (1 + callchain->nr) * sizeof(u64);
2881 header.size += callchain_size;
2883 header.size += sizeof(u64);
2886 if (sample_type & PERF_SAMPLE_RAW) {
2887 int size = sizeof(u32);
2890 size += data->raw->size;
2892 size += sizeof(u32);
2894 WARN_ON_ONCE(size & (sizeof(u64)-1));
2895 header.size += size;
2898 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2902 perf_output_put(&handle, header);
2904 if (sample_type & PERF_SAMPLE_IP)
2905 perf_output_put(&handle, ip);
2907 if (sample_type & PERF_SAMPLE_TID)
2908 perf_output_put(&handle, tid_entry);
2910 if (sample_type & PERF_SAMPLE_TIME)
2911 perf_output_put(&handle, time);
2913 if (sample_type & PERF_SAMPLE_ADDR)
2914 perf_output_put(&handle, data->addr);
2916 if (sample_type & PERF_SAMPLE_ID) {
2917 u64 id = primary_counter_id(counter);
2919 perf_output_put(&handle, id);
2922 if (sample_type & PERF_SAMPLE_STREAM_ID)
2923 perf_output_put(&handle, counter->id);
2925 if (sample_type & PERF_SAMPLE_CPU)
2926 perf_output_put(&handle, cpu_entry);
2928 if (sample_type & PERF_SAMPLE_PERIOD)
2929 perf_output_put(&handle, data->period);
2931 if (sample_type & PERF_SAMPLE_READ)
2932 perf_output_read(&handle, counter);
2934 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2936 perf_output_copy(&handle, callchain, callchain_size);
2939 perf_output_put(&handle, nr);
2943 if (sample_type & PERF_SAMPLE_RAW) {
2945 perf_output_put(&handle, data->raw->size);
2946 perf_output_copy(&handle, data->raw->data, data->raw->size);
2952 .size = sizeof(u32),
2955 perf_output_put(&handle, raw);
2959 perf_output_end(&handle);
2966 struct perf_read_event {
2967 struct perf_event_header header;
2974 perf_counter_read_event(struct perf_counter *counter,
2975 struct task_struct *task)
2977 struct perf_output_handle handle;
2978 struct perf_read_event event = {
2980 .type = PERF_EVENT_READ,
2982 .size = sizeof(event) + perf_counter_read_size(counter),
2984 .pid = perf_counter_pid(counter, task),
2985 .tid = perf_counter_tid(counter, task),
2989 ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
2993 perf_output_put(&handle, event);
2994 perf_output_read(&handle, counter);
2996 perf_output_end(&handle);
3000 * task tracking -- fork/exit
3002 * enabled by: attr.comm | attr.mmap | attr.task
3005 struct perf_task_event {
3006 struct task_struct *task;
3007 struct perf_counter_context *task_ctx;
3010 struct perf_event_header header;
3019 static void perf_counter_task_output(struct perf_counter *counter,
3020 struct perf_task_event *task_event)
3022 struct perf_output_handle handle;
3023 int size = task_event->event.header.size;
3024 struct task_struct *task = task_event->task;
3025 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3030 task_event->event.pid = perf_counter_pid(counter, task);
3031 task_event->event.ppid = perf_counter_pid(counter, current);
3033 task_event->event.tid = perf_counter_tid(counter, task);
3034 task_event->event.ptid = perf_counter_tid(counter, current);
3036 perf_output_put(&handle, task_event->event);
3037 perf_output_end(&handle);
3040 static int perf_counter_task_match(struct perf_counter *counter)
3042 if (counter->attr.comm || counter->attr.mmap || counter->attr.task)
3048 static void perf_counter_task_ctx(struct perf_counter_context *ctx,
3049 struct perf_task_event *task_event)
3051 struct perf_counter *counter;
3053 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3057 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3058 if (perf_counter_task_match(counter))
3059 perf_counter_task_output(counter, task_event);
3064 static void perf_counter_task_event(struct perf_task_event *task_event)
3066 struct perf_cpu_context *cpuctx;
3067 struct perf_counter_context *ctx = task_event->task_ctx;
3069 cpuctx = &get_cpu_var(perf_cpu_context);
3070 perf_counter_task_ctx(&cpuctx->ctx, task_event);
3071 put_cpu_var(perf_cpu_context);
3075 ctx = rcu_dereference(task_event->task->perf_counter_ctxp);
3077 perf_counter_task_ctx(ctx, task_event);
3081 static void perf_counter_task(struct task_struct *task,
3082 struct perf_counter_context *task_ctx,
3085 struct perf_task_event task_event;
3087 if (!atomic_read(&nr_comm_counters) &&
3088 !atomic_read(&nr_mmap_counters) &&
3089 !atomic_read(&nr_task_counters))
3092 task_event = (struct perf_task_event){
3094 .task_ctx = task_ctx,
3097 .type = new ? PERF_EVENT_FORK : PERF_EVENT_EXIT,
3099 .size = sizeof(task_event.event),
3108 perf_counter_task_event(&task_event);
3111 void perf_counter_fork(struct task_struct *task)
3113 perf_counter_task(task, NULL, 1);
3120 struct perf_comm_event {
3121 struct task_struct *task;
3126 struct perf_event_header header;
3133 static void perf_counter_comm_output(struct perf_counter *counter,
3134 struct perf_comm_event *comm_event)
3136 struct perf_output_handle handle;
3137 int size = comm_event->event.header.size;
3138 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3143 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
3144 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
3146 perf_output_put(&handle, comm_event->event);
3147 perf_output_copy(&handle, comm_event->comm,
3148 comm_event->comm_size);
3149 perf_output_end(&handle);
3152 static int perf_counter_comm_match(struct perf_counter *counter)
3154 if (counter->attr.comm)
3160 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
3161 struct perf_comm_event *comm_event)
3163 struct perf_counter *counter;
3165 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3169 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3170 if (perf_counter_comm_match(counter))
3171 perf_counter_comm_output(counter, comm_event);
3176 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
3178 struct perf_cpu_context *cpuctx;
3179 struct perf_counter_context *ctx;
3181 char comm[TASK_COMM_LEN];
3183 memset(comm, 0, sizeof(comm));
3184 strncpy(comm, comm_event->task->comm, sizeof(comm));
3185 size = ALIGN(strlen(comm)+1, sizeof(u64));
3187 comm_event->comm = comm;
3188 comm_event->comm_size = size;
3190 comm_event->event.header.size = sizeof(comm_event->event) + size;
3192 cpuctx = &get_cpu_var(perf_cpu_context);
3193 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
3194 put_cpu_var(perf_cpu_context);
3198 * doesn't really matter which of the child contexts the
3199 * events ends up in.
3201 ctx = rcu_dereference(current->perf_counter_ctxp);
3203 perf_counter_comm_ctx(ctx, comm_event);
3207 void perf_counter_comm(struct task_struct *task)
3209 struct perf_comm_event comm_event;
3211 if (task->perf_counter_ctxp)
3212 perf_counter_enable_on_exec(task);
3214 if (!atomic_read(&nr_comm_counters))
3217 comm_event = (struct perf_comm_event){
3223 .type = PERF_EVENT_COMM,
3232 perf_counter_comm_event(&comm_event);
3239 struct perf_mmap_event {
3240 struct vm_area_struct *vma;
3242 const char *file_name;
3246 struct perf_event_header header;
3256 static void perf_counter_mmap_output(struct perf_counter *counter,
3257 struct perf_mmap_event *mmap_event)
3259 struct perf_output_handle handle;
3260 int size = mmap_event->event.header.size;
3261 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3266 mmap_event->event.pid = perf_counter_pid(counter, current);
3267 mmap_event->event.tid = perf_counter_tid(counter, current);
3269 perf_output_put(&handle, mmap_event->event);
3270 perf_output_copy(&handle, mmap_event->file_name,
3271 mmap_event->file_size);
3272 perf_output_end(&handle);
3275 static int perf_counter_mmap_match(struct perf_counter *counter,
3276 struct perf_mmap_event *mmap_event)
3278 if (counter->attr.mmap)
3284 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3285 struct perf_mmap_event *mmap_event)
3287 struct perf_counter *counter;
3289 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3293 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3294 if (perf_counter_mmap_match(counter, mmap_event))
3295 perf_counter_mmap_output(counter, mmap_event);
3300 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3302 struct perf_cpu_context *cpuctx;
3303 struct perf_counter_context *ctx;
3304 struct vm_area_struct *vma = mmap_event->vma;
3305 struct file *file = vma->vm_file;
3311 memset(tmp, 0, sizeof(tmp));
3315 * d_path works from the end of the buffer backwards, so we
3316 * need to add enough zero bytes after the string to handle
3317 * the 64bit alignment we do later.
3319 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3321 name = strncpy(tmp, "//enomem", sizeof(tmp));
3324 name = d_path(&file->f_path, buf, PATH_MAX);
3326 name = strncpy(tmp, "//toolong", sizeof(tmp));
3330 if (arch_vma_name(mmap_event->vma)) {
3331 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3337 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3341 name = strncpy(tmp, "//anon", sizeof(tmp));
3346 size = ALIGN(strlen(name)+1, sizeof(u64));
3348 mmap_event->file_name = name;
3349 mmap_event->file_size = size;
3351 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3353 cpuctx = &get_cpu_var(perf_cpu_context);
3354 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3355 put_cpu_var(perf_cpu_context);
3359 * doesn't really matter which of the child contexts the
3360 * events ends up in.
3362 ctx = rcu_dereference(current->perf_counter_ctxp);
3364 perf_counter_mmap_ctx(ctx, mmap_event);
3370 void __perf_counter_mmap(struct vm_area_struct *vma)
3372 struct perf_mmap_event mmap_event;
3374 if (!atomic_read(&nr_mmap_counters))
3377 mmap_event = (struct perf_mmap_event){
3383 .type = PERF_EVENT_MMAP,
3389 .start = vma->vm_start,
3390 .len = vma->vm_end - vma->vm_start,
3391 .pgoff = vma->vm_pgoff,
3395 perf_counter_mmap_event(&mmap_event);
3399 * IRQ throttle logging
3402 static void perf_log_throttle(struct perf_counter *counter, int enable)
3404 struct perf_output_handle handle;
3408 struct perf_event_header header;
3412 } throttle_event = {
3414 .type = PERF_EVENT_THROTTLE,
3416 .size = sizeof(throttle_event),
3418 .time = sched_clock(),
3419 .id = primary_counter_id(counter),
3420 .stream_id = counter->id,
3424 throttle_event.header.type = PERF_EVENT_UNTHROTTLE;
3426 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3430 perf_output_put(&handle, throttle_event);
3431 perf_output_end(&handle);
3435 * Generic counter overflow handling, sampling.
3438 int perf_counter_overflow(struct perf_counter *counter, int nmi,
3439 struct perf_sample_data *data)
3441 int events = atomic_read(&counter->event_limit);
3442 int throttle = counter->pmu->unthrottle != NULL;
3443 struct hw_perf_counter *hwc = &counter->hw;
3449 if (hwc->interrupts != MAX_INTERRUPTS) {
3451 if (HZ * hwc->interrupts >
3452 (u64)sysctl_perf_counter_sample_rate) {
3453 hwc->interrupts = MAX_INTERRUPTS;
3454 perf_log_throttle(counter, 0);
3459 * Keep re-disabling counters even though on the previous
3460 * pass we disabled it - just in case we raced with a
3461 * sched-in and the counter got enabled again:
3467 if (counter->attr.freq) {
3468 u64 now = sched_clock();
3469 s64 delta = now - hwc->freq_stamp;
3471 hwc->freq_stamp = now;
3473 if (delta > 0 && delta < TICK_NSEC)
3474 perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3478 * XXX event_limit might not quite work as expected on inherited
3482 counter->pending_kill = POLL_IN;
3483 if (events && atomic_dec_and_test(&counter->event_limit)) {
3485 counter->pending_kill = POLL_HUP;
3487 counter->pending_disable = 1;
3488 perf_pending_queue(&counter->pending,
3489 perf_pending_counter);
3491 perf_counter_disable(counter);
3494 perf_counter_output(counter, nmi, data);
3499 * Generic software counter infrastructure
3503 * We directly increment counter->count and keep a second value in
3504 * counter->hw.period_left to count intervals. This period counter
3505 * is kept in the range [-sample_period, 0] so that we can use the
3509 static u64 perf_swcounter_set_period(struct perf_counter *counter)
3511 struct hw_perf_counter *hwc = &counter->hw;
3512 u64 period = hwc->last_period;
3516 hwc->last_period = hwc->sample_period;
3519 old = val = atomic64_read(&hwc->period_left);
3523 nr = div64_u64(period + val, period);
3524 offset = nr * period;
3526 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3532 static void perf_swcounter_overflow(struct perf_counter *counter,
3533 int nmi, struct perf_sample_data *data)
3535 struct hw_perf_counter *hwc = &counter->hw;
3538 data->period = counter->hw.last_period;
3539 overflow = perf_swcounter_set_period(counter);
3541 if (hwc->interrupts == MAX_INTERRUPTS)
3544 for (; overflow; overflow--) {
3545 if (perf_counter_overflow(counter, nmi, data)) {
3547 * We inhibit the overflow from happening when
3548 * hwc->interrupts == MAX_INTERRUPTS.
3555 static void perf_swcounter_unthrottle(struct perf_counter *counter)
3558 * Nothing to do, we already reset hwc->interrupts.
3562 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3563 int nmi, struct perf_sample_data *data)
3565 struct hw_perf_counter *hwc = &counter->hw;
3567 atomic64_add(nr, &counter->count);
3569 if (!hwc->sample_period)
3575 if (!atomic64_add_negative(nr, &hwc->period_left))
3576 perf_swcounter_overflow(counter, nmi, data);
3579 static int perf_swcounter_is_counting(struct perf_counter *counter)
3582 * The counter is active, we're good!
3584 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3588 * The counter is off/error, not counting.
3590 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3594 * The counter is inactive, if the context is active
3595 * we're part of a group that didn't make it on the 'pmu',
3598 if (counter->ctx->is_active)
3602 * We're inactive and the context is too, this means the
3603 * task is scheduled out, we're counting events that happen
3604 * to us, like migration events.
3609 static int perf_swcounter_match(struct perf_counter *counter,
3610 enum perf_type_id type,
3611 u32 event, struct pt_regs *regs)
3613 if (!perf_swcounter_is_counting(counter))
3616 if (counter->attr.type != type)
3618 if (counter->attr.config != event)
3622 if (counter->attr.exclude_user && user_mode(regs))
3625 if (counter->attr.exclude_kernel && !user_mode(regs))
3632 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3633 enum perf_type_id type,
3634 u32 event, u64 nr, int nmi,
3635 struct perf_sample_data *data)
3637 struct perf_counter *counter;
3639 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3643 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3644 if (perf_swcounter_match(counter, type, event, data->regs))
3645 perf_swcounter_add(counter, nr, nmi, data);
3650 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3653 return &cpuctx->recursion[3];
3656 return &cpuctx->recursion[2];
3659 return &cpuctx->recursion[1];
3661 return &cpuctx->recursion[0];
3664 static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3666 struct perf_sample_data *data)
3668 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3669 int *recursion = perf_swcounter_recursion_context(cpuctx);
3670 struct perf_counter_context *ctx;
3678 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3682 * doesn't really matter which of the child contexts the
3683 * events ends up in.
3685 ctx = rcu_dereference(current->perf_counter_ctxp);
3687 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3694 put_cpu_var(perf_cpu_context);
3697 void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3698 struct pt_regs *regs, u64 addr)
3700 struct perf_sample_data data = {
3705 do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3708 static void perf_swcounter_read(struct perf_counter *counter)
3712 static int perf_swcounter_enable(struct perf_counter *counter)
3714 struct hw_perf_counter *hwc = &counter->hw;
3716 if (hwc->sample_period) {
3717 hwc->last_period = hwc->sample_period;
3718 perf_swcounter_set_period(counter);
3723 static void perf_swcounter_disable(struct perf_counter *counter)
3727 static const struct pmu perf_ops_generic = {
3728 .enable = perf_swcounter_enable,
3729 .disable = perf_swcounter_disable,
3730 .read = perf_swcounter_read,
3731 .unthrottle = perf_swcounter_unthrottle,
3735 * hrtimer based swcounter callback
3738 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3740 enum hrtimer_restart ret = HRTIMER_RESTART;
3741 struct perf_sample_data data;
3742 struct perf_counter *counter;
3745 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3746 counter->pmu->read(counter);
3749 data.regs = get_irq_regs();
3751 * In case we exclude kernel IPs or are somehow not in interrupt
3752 * context, provide the next best thing, the user IP.
3754 if ((counter->attr.exclude_kernel || !data.regs) &&
3755 !counter->attr.exclude_user)
3756 data.regs = task_pt_regs(current);
3759 if (perf_counter_overflow(counter, 0, &data))
3760 ret = HRTIMER_NORESTART;
3763 period = max_t(u64, 10000, counter->hw.sample_period);
3764 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3770 * Software counter: cpu wall time clock
3773 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3775 int cpu = raw_smp_processor_id();
3779 now = cpu_clock(cpu);
3780 prev = atomic64_read(&counter->hw.prev_count);
3781 atomic64_set(&counter->hw.prev_count, now);
3782 atomic64_add(now - prev, &counter->count);
3785 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3787 struct hw_perf_counter *hwc = &counter->hw;
3788 int cpu = raw_smp_processor_id();
3790 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3791 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3792 hwc->hrtimer.function = perf_swcounter_hrtimer;
3793 if (hwc->sample_period) {
3794 u64 period = max_t(u64, 10000, hwc->sample_period);
3795 __hrtimer_start_range_ns(&hwc->hrtimer,
3796 ns_to_ktime(period), 0,
3797 HRTIMER_MODE_REL, 0);
3803 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3805 if (counter->hw.sample_period)
3806 hrtimer_cancel(&counter->hw.hrtimer);
3807 cpu_clock_perf_counter_update(counter);
3810 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3812 cpu_clock_perf_counter_update(counter);
3815 static const struct pmu perf_ops_cpu_clock = {
3816 .enable = cpu_clock_perf_counter_enable,
3817 .disable = cpu_clock_perf_counter_disable,
3818 .read = cpu_clock_perf_counter_read,
3822 * Software counter: task time clock
3825 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3830 prev = atomic64_xchg(&counter->hw.prev_count, now);
3832 atomic64_add(delta, &counter->count);
3835 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3837 struct hw_perf_counter *hwc = &counter->hw;
3840 now = counter->ctx->time;
3842 atomic64_set(&hwc->prev_count, now);
3843 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3844 hwc->hrtimer.function = perf_swcounter_hrtimer;
3845 if (hwc->sample_period) {
3846 u64 period = max_t(u64, 10000, hwc->sample_period);
3847 __hrtimer_start_range_ns(&hwc->hrtimer,
3848 ns_to_ktime(period), 0,
3849 HRTIMER_MODE_REL, 0);
3855 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3857 if (counter->hw.sample_period)
3858 hrtimer_cancel(&counter->hw.hrtimer);
3859 task_clock_perf_counter_update(counter, counter->ctx->time);
3863 static void task_clock_perf_counter_read(struct perf_counter *counter)
3868 update_context_time(counter->ctx);
3869 time = counter->ctx->time;
3871 u64 now = perf_clock();
3872 u64 delta = now - counter->ctx->timestamp;
3873 time = counter->ctx->time + delta;
3876 task_clock_perf_counter_update(counter, time);
3879 static const struct pmu perf_ops_task_clock = {
3880 .enable = task_clock_perf_counter_enable,
3881 .disable = task_clock_perf_counter_disable,
3882 .read = task_clock_perf_counter_read,
3885 #ifdef CONFIG_EVENT_PROFILE
3886 void perf_tpcounter_event(int event_id, u64 addr, u64 count, void *record,
3889 struct perf_raw_record raw = {
3894 struct perf_sample_data data = {
3895 .regs = get_irq_regs(),
3901 data.regs = task_pt_regs(current);
3903 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, count, 1, &data);
3905 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3907 extern int ftrace_profile_enable(int);
3908 extern void ftrace_profile_disable(int);
3910 static void tp_perf_counter_destroy(struct perf_counter *counter)
3912 ftrace_profile_disable(counter->attr.config);
3915 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3918 * Raw tracepoint data is a severe data leak, only allow root to
3921 if ((counter->attr.sample_type & PERF_SAMPLE_RAW) &&
3922 !capable(CAP_SYS_ADMIN))
3923 return ERR_PTR(-EPERM);
3925 if (ftrace_profile_enable(counter->attr.config))
3928 counter->destroy = tp_perf_counter_destroy;
3930 return &perf_ops_generic;
3933 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3939 atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3941 static void sw_perf_counter_destroy(struct perf_counter *counter)
3943 u64 event = counter->attr.config;
3945 WARN_ON(counter->parent);
3947 atomic_dec(&perf_swcounter_enabled[event]);
3950 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3952 const struct pmu *pmu = NULL;
3953 u64 event = counter->attr.config;
3956 * Software counters (currently) can't in general distinguish
3957 * between user, kernel and hypervisor events.
3958 * However, context switches and cpu migrations are considered
3959 * to be kernel events, and page faults are never hypervisor
3963 case PERF_COUNT_SW_CPU_CLOCK:
3964 pmu = &perf_ops_cpu_clock;
3967 case PERF_COUNT_SW_TASK_CLOCK:
3969 * If the user instantiates this as a per-cpu counter,
3970 * use the cpu_clock counter instead.
3972 if (counter->ctx->task)
3973 pmu = &perf_ops_task_clock;
3975 pmu = &perf_ops_cpu_clock;
3978 case PERF_COUNT_SW_PAGE_FAULTS:
3979 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
3980 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
3981 case PERF_COUNT_SW_CONTEXT_SWITCHES:
3982 case PERF_COUNT_SW_CPU_MIGRATIONS:
3983 if (!counter->parent) {
3984 atomic_inc(&perf_swcounter_enabled[event]);
3985 counter->destroy = sw_perf_counter_destroy;
3987 pmu = &perf_ops_generic;
3995 * Allocate and initialize a counter structure
3997 static struct perf_counter *
3998 perf_counter_alloc(struct perf_counter_attr *attr,
4000 struct perf_counter_context *ctx,
4001 struct perf_counter *group_leader,
4002 struct perf_counter *parent_counter,
4005 const struct pmu *pmu;
4006 struct perf_counter *counter;
4007 struct hw_perf_counter *hwc;
4010 counter = kzalloc(sizeof(*counter), gfpflags);
4012 return ERR_PTR(-ENOMEM);
4015 * Single counters are their own group leaders, with an
4016 * empty sibling list:
4019 group_leader = counter;
4021 mutex_init(&counter->child_mutex);
4022 INIT_LIST_HEAD(&counter->child_list);
4024 INIT_LIST_HEAD(&counter->list_entry);
4025 INIT_LIST_HEAD(&counter->event_entry);
4026 INIT_LIST_HEAD(&counter->sibling_list);
4027 init_waitqueue_head(&counter->waitq);
4029 mutex_init(&counter->mmap_mutex);
4032 counter->attr = *attr;
4033 counter->group_leader = group_leader;
4034 counter->pmu = NULL;
4036 counter->oncpu = -1;
4038 counter->parent = parent_counter;
4040 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
4041 counter->id = atomic64_inc_return(&perf_counter_id);
4043 counter->state = PERF_COUNTER_STATE_INACTIVE;
4046 counter->state = PERF_COUNTER_STATE_OFF;
4051 hwc->sample_period = attr->sample_period;
4052 if (attr->freq && attr->sample_freq)
4053 hwc->sample_period = 1;
4055 atomic64_set(&hwc->period_left, hwc->sample_period);
4058 * we currently do not support PERF_FORMAT_GROUP on inherited counters
4060 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4063 switch (attr->type) {
4065 case PERF_TYPE_HARDWARE:
4066 case PERF_TYPE_HW_CACHE:
4067 pmu = hw_perf_counter_init(counter);
4070 case PERF_TYPE_SOFTWARE:
4071 pmu = sw_perf_counter_init(counter);
4074 case PERF_TYPE_TRACEPOINT:
4075 pmu = tp_perf_counter_init(counter);
4085 else if (IS_ERR(pmu))
4090 put_pid_ns(counter->ns);
4092 return ERR_PTR(err);
4097 if (!counter->parent) {
4098 atomic_inc(&nr_counters);
4099 if (counter->attr.mmap)
4100 atomic_inc(&nr_mmap_counters);
4101 if (counter->attr.comm)
4102 atomic_inc(&nr_comm_counters);
4103 if (counter->attr.task)
4104 atomic_inc(&nr_task_counters);
4110 static int perf_copy_attr(struct perf_counter_attr __user *uattr,
4111 struct perf_counter_attr *attr)
4116 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4120 * zero the full structure, so that a short copy will be nice.
4122 memset(attr, 0, sizeof(*attr));
4124 ret = get_user(size, &uattr->size);
4128 if (size > PAGE_SIZE) /* silly large */
4131 if (!size) /* abi compat */
4132 size = PERF_ATTR_SIZE_VER0;
4134 if (size < PERF_ATTR_SIZE_VER0)
4138 * If we're handed a bigger struct than we know of,
4139 * ensure all the unknown bits are 0.
4141 if (size > sizeof(*attr)) {
4143 unsigned long __user *addr;
4144 unsigned long __user *end;
4146 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
4147 sizeof(unsigned long));
4148 end = PTR_ALIGN((void __user *)uattr + size,
4149 sizeof(unsigned long));
4151 for (; addr < end; addr += sizeof(unsigned long)) {
4152 ret = get_user(val, addr);
4160 ret = copy_from_user(attr, uattr, size);
4165 * If the type exists, the corresponding creation will verify
4168 if (attr->type >= PERF_TYPE_MAX)
4171 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4174 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4177 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4184 put_user(sizeof(*attr), &uattr->size);
4190 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
4192 * @attr_uptr: event type attributes for monitoring/sampling
4195 * @group_fd: group leader counter fd
4197 SYSCALL_DEFINE5(perf_counter_open,
4198 struct perf_counter_attr __user *, attr_uptr,
4199 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4201 struct perf_counter *counter, *group_leader;
4202 struct perf_counter_attr attr;
4203 struct perf_counter_context *ctx;
4204 struct file *counter_file = NULL;
4205 struct file *group_file = NULL;
4206 int fput_needed = 0;
4207 int fput_needed2 = 0;
4210 /* for future expandability... */
4214 ret = perf_copy_attr(attr_uptr, &attr);
4218 if (!attr.exclude_kernel) {
4219 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4224 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
4229 * Get the target context (task or percpu):
4231 ctx = find_get_context(pid, cpu);
4233 return PTR_ERR(ctx);
4236 * Look up the group leader (we will attach this counter to it):
4238 group_leader = NULL;
4239 if (group_fd != -1) {
4241 group_file = fget_light(group_fd, &fput_needed);
4243 goto err_put_context;
4244 if (group_file->f_op != &perf_fops)
4245 goto err_put_context;
4247 group_leader = group_file->private_data;
4249 * Do not allow a recursive hierarchy (this new sibling
4250 * becoming part of another group-sibling):
4252 if (group_leader->group_leader != group_leader)
4253 goto err_put_context;
4255 * Do not allow to attach to a group in a different
4256 * task or CPU context:
4258 if (group_leader->ctx != ctx)
4259 goto err_put_context;
4261 * Only a group leader can be exclusive or pinned
4263 if (attr.exclusive || attr.pinned)
4264 goto err_put_context;
4267 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
4269 ret = PTR_ERR(counter);
4270 if (IS_ERR(counter))
4271 goto err_put_context;
4273 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
4275 goto err_free_put_context;
4277 counter_file = fget_light(ret, &fput_needed2);
4279 goto err_free_put_context;
4281 counter->filp = counter_file;
4282 WARN_ON_ONCE(ctx->parent_ctx);
4283 mutex_lock(&ctx->mutex);
4284 perf_install_in_context(ctx, counter, cpu);
4286 mutex_unlock(&ctx->mutex);
4288 counter->owner = current;
4289 get_task_struct(current);
4290 mutex_lock(¤t->perf_counter_mutex);
4291 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
4292 mutex_unlock(¤t->perf_counter_mutex);
4294 fput_light(counter_file, fput_needed2);
4297 fput_light(group_file, fput_needed);
4301 err_free_put_context:
4311 * inherit a counter from parent task to child task:
4313 static struct perf_counter *
4314 inherit_counter(struct perf_counter *parent_counter,
4315 struct task_struct *parent,
4316 struct perf_counter_context *parent_ctx,
4317 struct task_struct *child,
4318 struct perf_counter *group_leader,
4319 struct perf_counter_context *child_ctx)
4321 struct perf_counter *child_counter;
4324 * Instead of creating recursive hierarchies of counters,
4325 * we link inherited counters back to the original parent,
4326 * which has a filp for sure, which we use as the reference
4329 if (parent_counter->parent)
4330 parent_counter = parent_counter->parent;
4332 child_counter = perf_counter_alloc(&parent_counter->attr,
4333 parent_counter->cpu, child_ctx,
4334 group_leader, parent_counter,
4336 if (IS_ERR(child_counter))
4337 return child_counter;
4341 * Make the child state follow the state of the parent counter,
4342 * not its attr.disabled bit. We hold the parent's mutex,
4343 * so we won't race with perf_counter_{en, dis}able_family.
4345 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4346 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4348 child_counter->state = PERF_COUNTER_STATE_OFF;
4350 if (parent_counter->attr.freq)
4351 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4354 * Link it up in the child's context:
4356 add_counter_to_ctx(child_counter, child_ctx);
4359 * Get a reference to the parent filp - we will fput it
4360 * when the child counter exits. This is safe to do because
4361 * we are in the parent and we know that the filp still
4362 * exists and has a nonzero count:
4364 atomic_long_inc(&parent_counter->filp->f_count);
4367 * Link this into the parent counter's child list
4369 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4370 mutex_lock(&parent_counter->child_mutex);
4371 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4372 mutex_unlock(&parent_counter->child_mutex);
4374 return child_counter;
4377 static int inherit_group(struct perf_counter *parent_counter,
4378 struct task_struct *parent,
4379 struct perf_counter_context *parent_ctx,
4380 struct task_struct *child,
4381 struct perf_counter_context *child_ctx)
4383 struct perf_counter *leader;
4384 struct perf_counter *sub;
4385 struct perf_counter *child_ctr;
4387 leader = inherit_counter(parent_counter, parent, parent_ctx,
4388 child, NULL, child_ctx);
4390 return PTR_ERR(leader);
4391 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4392 child_ctr = inherit_counter(sub, parent, parent_ctx,
4393 child, leader, child_ctx);
4394 if (IS_ERR(child_ctr))
4395 return PTR_ERR(child_ctr);
4400 static void sync_child_counter(struct perf_counter *child_counter,
4401 struct task_struct *child)
4403 struct perf_counter *parent_counter = child_counter->parent;
4406 if (child_counter->attr.inherit_stat)
4407 perf_counter_read_event(child_counter, child);
4409 child_val = atomic64_read(&child_counter->count);
4412 * Add back the child's count to the parent's count:
4414 atomic64_add(child_val, &parent_counter->count);
4415 atomic64_add(child_counter->total_time_enabled,
4416 &parent_counter->child_total_time_enabled);
4417 atomic64_add(child_counter->total_time_running,
4418 &parent_counter->child_total_time_running);
4421 * Remove this counter from the parent's list
4423 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4424 mutex_lock(&parent_counter->child_mutex);
4425 list_del_init(&child_counter->child_list);
4426 mutex_unlock(&parent_counter->child_mutex);
4429 * Release the parent counter, if this was the last
4432 fput(parent_counter->filp);
4436 __perf_counter_exit_task(struct perf_counter *child_counter,
4437 struct perf_counter_context *child_ctx,
4438 struct task_struct *child)
4440 struct perf_counter *parent_counter;
4442 update_counter_times(child_counter);
4443 perf_counter_remove_from_context(child_counter);
4445 parent_counter = child_counter->parent;
4447 * It can happen that parent exits first, and has counters
4448 * that are still around due to the child reference. These
4449 * counters need to be zapped - but otherwise linger.
4451 if (parent_counter) {
4452 sync_child_counter(child_counter, child);
4453 free_counter(child_counter);
4458 * When a child task exits, feed back counter values to parent counters.
4460 void perf_counter_exit_task(struct task_struct *child)
4462 struct perf_counter *child_counter, *tmp;
4463 struct perf_counter_context *child_ctx;
4464 unsigned long flags;
4466 if (likely(!child->perf_counter_ctxp)) {
4467 perf_counter_task(child, NULL, 0);
4471 local_irq_save(flags);
4473 * We can't reschedule here because interrupts are disabled,
4474 * and either child is current or it is a task that can't be
4475 * scheduled, so we are now safe from rescheduling changing
4478 child_ctx = child->perf_counter_ctxp;
4479 __perf_counter_task_sched_out(child_ctx);
4482 * Take the context lock here so that if find_get_context is
4483 * reading child->perf_counter_ctxp, we wait until it has
4484 * incremented the context's refcount before we do put_ctx below.
4486 spin_lock(&child_ctx->lock);
4487 child->perf_counter_ctxp = NULL;
4489 * If this context is a clone; unclone it so it can't get
4490 * swapped to another process while we're removing all
4491 * the counters from it.
4493 unclone_ctx(child_ctx);
4494 spin_unlock_irqrestore(&child_ctx->lock, flags);
4497 * Report the task dead after unscheduling the counters so that we
4498 * won't get any samples after PERF_EVENT_EXIT. We can however still
4499 * get a few PERF_EVENT_READ events.
4501 perf_counter_task(child, child_ctx, 0);
4504 * We can recurse on the same lock type through:
4506 * __perf_counter_exit_task()
4507 * sync_child_counter()
4508 * fput(parent_counter->filp)
4510 * mutex_lock(&ctx->mutex)
4512 * But since its the parent context it won't be the same instance.
4514 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4517 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4519 __perf_counter_exit_task(child_counter, child_ctx, child);
4522 * If the last counter was a group counter, it will have appended all
4523 * its siblings to the list, but we obtained 'tmp' before that which
4524 * will still point to the list head terminating the iteration.
4526 if (!list_empty(&child_ctx->counter_list))
4529 mutex_unlock(&child_ctx->mutex);
4535 * free an unexposed, unused context as created by inheritance by
4536 * init_task below, used by fork() in case of fail.
4538 void perf_counter_free_task(struct task_struct *task)
4540 struct perf_counter_context *ctx = task->perf_counter_ctxp;
4541 struct perf_counter *counter, *tmp;
4546 mutex_lock(&ctx->mutex);
4548 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4549 struct perf_counter *parent = counter->parent;
4551 if (WARN_ON_ONCE(!parent))
4554 mutex_lock(&parent->child_mutex);
4555 list_del_init(&counter->child_list);
4556 mutex_unlock(&parent->child_mutex);
4560 list_del_counter(counter, ctx);
4561 free_counter(counter);
4564 if (!list_empty(&ctx->counter_list))
4567 mutex_unlock(&ctx->mutex);
4573 * Initialize the perf_counter context in task_struct
4575 int perf_counter_init_task(struct task_struct *child)
4577 struct perf_counter_context *child_ctx, *parent_ctx;
4578 struct perf_counter_context *cloned_ctx;
4579 struct perf_counter *counter;
4580 struct task_struct *parent = current;
4581 int inherited_all = 1;
4584 child->perf_counter_ctxp = NULL;
4586 mutex_init(&child->perf_counter_mutex);
4587 INIT_LIST_HEAD(&child->perf_counter_list);
4589 if (likely(!parent->perf_counter_ctxp))
4593 * This is executed from the parent task context, so inherit
4594 * counters that have been marked for cloning.
4595 * First allocate and initialize a context for the child.
4598 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4602 __perf_counter_init_context(child_ctx, child);
4603 child->perf_counter_ctxp = child_ctx;
4604 get_task_struct(child);
4607 * If the parent's context is a clone, pin it so it won't get
4610 parent_ctx = perf_pin_task_context(parent);
4613 * No need to check if parent_ctx != NULL here; since we saw
4614 * it non-NULL earlier, the only reason for it to become NULL
4615 * is if we exit, and since we're currently in the middle of
4616 * a fork we can't be exiting at the same time.
4620 * Lock the parent list. No need to lock the child - not PID
4621 * hashed yet and not running, so nobody can access it.
4623 mutex_lock(&parent_ctx->mutex);
4626 * We dont have to disable NMIs - we are only looking at
4627 * the list, not manipulating it:
4629 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4630 if (counter != counter->group_leader)
4633 if (!counter->attr.inherit) {
4638 ret = inherit_group(counter, parent, parent_ctx,
4646 if (inherited_all) {
4648 * Mark the child context as a clone of the parent
4649 * context, or of whatever the parent is a clone of.
4650 * Note that if the parent is a clone, it could get
4651 * uncloned at any point, but that doesn't matter
4652 * because the list of counters and the generation
4653 * count can't have changed since we took the mutex.
4655 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4657 child_ctx->parent_ctx = cloned_ctx;
4658 child_ctx->parent_gen = parent_ctx->parent_gen;
4660 child_ctx->parent_ctx = parent_ctx;
4661 child_ctx->parent_gen = parent_ctx->generation;
4663 get_ctx(child_ctx->parent_ctx);
4666 mutex_unlock(&parent_ctx->mutex);
4668 perf_unpin_context(parent_ctx);
4673 static void __cpuinit perf_counter_init_cpu(int cpu)
4675 struct perf_cpu_context *cpuctx;
4677 cpuctx = &per_cpu(perf_cpu_context, cpu);
4678 __perf_counter_init_context(&cpuctx->ctx, NULL);
4680 spin_lock(&perf_resource_lock);
4681 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4682 spin_unlock(&perf_resource_lock);
4684 hw_perf_counter_setup(cpu);
4687 #ifdef CONFIG_HOTPLUG_CPU
4688 static void __perf_counter_exit_cpu(void *info)
4690 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4691 struct perf_counter_context *ctx = &cpuctx->ctx;
4692 struct perf_counter *counter, *tmp;
4694 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4695 __perf_counter_remove_from_context(counter);
4697 static void perf_counter_exit_cpu(int cpu)
4699 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4700 struct perf_counter_context *ctx = &cpuctx->ctx;
4702 mutex_lock(&ctx->mutex);
4703 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4704 mutex_unlock(&ctx->mutex);
4707 static inline void perf_counter_exit_cpu(int cpu) { }
4710 static int __cpuinit
4711 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4713 unsigned int cpu = (long)hcpu;
4717 case CPU_UP_PREPARE:
4718 case CPU_UP_PREPARE_FROZEN:
4719 perf_counter_init_cpu(cpu);
4723 case CPU_ONLINE_FROZEN:
4724 hw_perf_counter_setup_online(cpu);
4727 case CPU_DOWN_PREPARE:
4728 case CPU_DOWN_PREPARE_FROZEN:
4729 perf_counter_exit_cpu(cpu);
4740 * This has to have a higher priority than migration_notifier in sched.c.
4742 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4743 .notifier_call = perf_cpu_notify,
4747 void __init perf_counter_init(void)
4749 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4750 (void *)(long)smp_processor_id());
4751 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
4752 (void *)(long)smp_processor_id());
4753 register_cpu_notifier(&perf_cpu_nb);
4756 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4758 return sprintf(buf, "%d\n", perf_reserved_percpu);
4762 perf_set_reserve_percpu(struct sysdev_class *class,
4766 struct perf_cpu_context *cpuctx;
4770 err = strict_strtoul(buf, 10, &val);
4773 if (val > perf_max_counters)
4776 spin_lock(&perf_resource_lock);
4777 perf_reserved_percpu = val;
4778 for_each_online_cpu(cpu) {
4779 cpuctx = &per_cpu(perf_cpu_context, cpu);
4780 spin_lock_irq(&cpuctx->ctx.lock);
4781 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4782 perf_max_counters - perf_reserved_percpu);
4783 cpuctx->max_pertask = mpt;
4784 spin_unlock_irq(&cpuctx->ctx.lock);
4786 spin_unlock(&perf_resource_lock);
4791 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4793 return sprintf(buf, "%d\n", perf_overcommit);
4797 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4802 err = strict_strtoul(buf, 10, &val);
4808 spin_lock(&perf_resource_lock);
4809 perf_overcommit = val;
4810 spin_unlock(&perf_resource_lock);
4815 static SYSDEV_CLASS_ATTR(
4818 perf_show_reserve_percpu,
4819 perf_set_reserve_percpu
4822 static SYSDEV_CLASS_ATTR(
4825 perf_show_overcommit,
4829 static struct attribute *perfclass_attrs[] = {
4830 &attr_reserve_percpu.attr,
4831 &attr_overcommit.attr,
4835 static struct attribute_group perfclass_attr_group = {
4836 .attrs = perfclass_attrs,
4837 .name = "perf_counters",
4840 static int __init perf_counter_sysfs_init(void)
4842 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4843 &perfclass_attr_group);
4845 device_initcall(perf_counter_sysfs_init);