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(); }
93 hw_perf_group_sched_in(struct perf_counter *group_leader,
94 struct perf_cpu_context *cpuctx,
95 struct perf_counter_context *ctx, int cpu)
100 void __weak perf_counter_print_debug(void) { }
102 static DEFINE_PER_CPU(int, disable_count);
104 void __perf_disable(void)
106 __get_cpu_var(disable_count)++;
109 bool __perf_enable(void)
111 return !--__get_cpu_var(disable_count);
114 void perf_disable(void)
120 void perf_enable(void)
126 static void get_ctx(struct perf_counter_context *ctx)
128 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
131 static void free_ctx(struct rcu_head *head)
133 struct perf_counter_context *ctx;
135 ctx = container_of(head, struct perf_counter_context, rcu_head);
139 static void put_ctx(struct perf_counter_context *ctx)
141 if (atomic_dec_and_test(&ctx->refcount)) {
143 put_ctx(ctx->parent_ctx);
145 put_task_struct(ctx->task);
146 call_rcu(&ctx->rcu_head, free_ctx);
150 static void unclone_ctx(struct perf_counter_context *ctx)
152 if (ctx->parent_ctx) {
153 put_ctx(ctx->parent_ctx);
154 ctx->parent_ctx = NULL;
159 * If we inherit counters we want to return the parent counter id
162 static u64 primary_counter_id(struct perf_counter *counter)
164 u64 id = counter->id;
167 id = counter->parent->id;
173 * Get the perf_counter_context for a task and lock it.
174 * This has to cope with with the fact that until it is locked,
175 * the context could get moved to another task.
177 static struct perf_counter_context *
178 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
180 struct perf_counter_context *ctx;
184 ctx = rcu_dereference(task->perf_counter_ctxp);
187 * If this context is a clone of another, it might
188 * get swapped for another underneath us by
189 * perf_counter_task_sched_out, though the
190 * rcu_read_lock() protects us from any context
191 * getting freed. Lock the context and check if it
192 * got swapped before we could get the lock, and retry
193 * if so. If we locked the right context, then it
194 * can't get swapped on us any more.
196 spin_lock_irqsave(&ctx->lock, *flags);
197 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
198 spin_unlock_irqrestore(&ctx->lock, *flags);
202 if (!atomic_inc_not_zero(&ctx->refcount)) {
203 spin_unlock_irqrestore(&ctx->lock, *flags);
212 * Get the context for a task and increment its pin_count so it
213 * can't get swapped to another task. This also increments its
214 * reference count so that the context can't get freed.
216 static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
218 struct perf_counter_context *ctx;
221 ctx = perf_lock_task_context(task, &flags);
224 spin_unlock_irqrestore(&ctx->lock, flags);
229 static void perf_unpin_context(struct perf_counter_context *ctx)
233 spin_lock_irqsave(&ctx->lock, flags);
235 spin_unlock_irqrestore(&ctx->lock, flags);
240 * Add a counter from the lists for its context.
241 * Must be called with ctx->mutex and ctx->lock held.
244 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
246 struct perf_counter *group_leader = counter->group_leader;
249 * Depending on whether it is a standalone or sibling counter,
250 * add it straight to the context's counter list, or to the group
251 * leader's sibling list:
253 if (group_leader == counter)
254 list_add_tail(&counter->list_entry, &ctx->counter_list);
256 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
257 group_leader->nr_siblings++;
260 list_add_rcu(&counter->event_entry, &ctx->event_list);
262 if (counter->attr.inherit_stat)
267 * Remove a counter from the lists for its context.
268 * Must be called with ctx->mutex and ctx->lock held.
271 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
273 struct perf_counter *sibling, *tmp;
275 if (list_empty(&counter->list_entry))
278 if (counter->attr.inherit_stat)
281 list_del_init(&counter->list_entry);
282 list_del_rcu(&counter->event_entry);
284 if (counter->group_leader != counter)
285 counter->group_leader->nr_siblings--;
288 * If this was a group counter with sibling counters then
289 * upgrade the siblings to singleton counters by adding them
290 * to the context list directly:
292 list_for_each_entry_safe(sibling, tmp,
293 &counter->sibling_list, list_entry) {
295 list_move_tail(&sibling->list_entry, &ctx->counter_list);
296 sibling->group_leader = sibling;
301 counter_sched_out(struct perf_counter *counter,
302 struct perf_cpu_context *cpuctx,
303 struct perf_counter_context *ctx)
305 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
308 counter->state = PERF_COUNTER_STATE_INACTIVE;
309 counter->tstamp_stopped = ctx->time;
310 counter->pmu->disable(counter);
313 if (!is_software_counter(counter))
314 cpuctx->active_oncpu--;
316 if (counter->attr.exclusive || !cpuctx->active_oncpu)
317 cpuctx->exclusive = 0;
321 group_sched_out(struct perf_counter *group_counter,
322 struct perf_cpu_context *cpuctx,
323 struct perf_counter_context *ctx)
325 struct perf_counter *counter;
327 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
330 counter_sched_out(group_counter, cpuctx, ctx);
333 * Schedule out siblings (if any):
335 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
336 counter_sched_out(counter, cpuctx, ctx);
338 if (group_counter->attr.exclusive)
339 cpuctx->exclusive = 0;
343 * Cross CPU call to remove a performance counter
345 * We disable the counter on the hardware level first. After that we
346 * remove it from the context list.
348 static void __perf_counter_remove_from_context(void *info)
350 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
351 struct perf_counter *counter = info;
352 struct perf_counter_context *ctx = counter->ctx;
355 * If this is a task context, we need to check whether it is
356 * the current task context of this cpu. If not it has been
357 * scheduled out before the smp call arrived.
359 if (ctx->task && cpuctx->task_ctx != ctx)
362 spin_lock(&ctx->lock);
364 * Protect the list operation against NMI by disabling the
365 * counters on a global level.
369 counter_sched_out(counter, cpuctx, ctx);
371 list_del_counter(counter, ctx);
375 * Allow more per task counters with respect to the
378 cpuctx->max_pertask =
379 min(perf_max_counters - ctx->nr_counters,
380 perf_max_counters - perf_reserved_percpu);
384 spin_unlock(&ctx->lock);
389 * Remove the counter from a task's (or a CPU's) list of counters.
391 * Must be called with ctx->mutex held.
393 * CPU counters are removed with a smp call. For task counters we only
394 * call when the task is on a CPU.
396 * If counter->ctx is a cloned context, callers must make sure that
397 * every task struct that counter->ctx->task could possibly point to
398 * remains valid. This is OK when called from perf_release since
399 * that only calls us on the top-level context, which can't be a clone.
400 * When called from perf_counter_exit_task, it's OK because the
401 * context has been detached from its task.
403 static void perf_counter_remove_from_context(struct perf_counter *counter)
405 struct perf_counter_context *ctx = counter->ctx;
406 struct task_struct *task = ctx->task;
410 * Per cpu counters are removed via an smp call and
411 * the removal is always sucessful.
413 smp_call_function_single(counter->cpu,
414 __perf_counter_remove_from_context,
420 task_oncpu_function_call(task, __perf_counter_remove_from_context,
423 spin_lock_irq(&ctx->lock);
425 * If the context is active we need to retry the smp call.
427 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
428 spin_unlock_irq(&ctx->lock);
433 * The lock prevents that this context is scheduled in so we
434 * can remove the counter safely, if the call above did not
437 if (!list_empty(&counter->list_entry)) {
438 list_del_counter(counter, ctx);
440 spin_unlock_irq(&ctx->lock);
443 static inline u64 perf_clock(void)
445 return cpu_clock(smp_processor_id());
449 * Update the record of the current time in a context.
451 static void update_context_time(struct perf_counter_context *ctx)
453 u64 now = perf_clock();
455 ctx->time += now - ctx->timestamp;
456 ctx->timestamp = now;
460 * Update the total_time_enabled and total_time_running fields for a counter.
462 static void update_counter_times(struct perf_counter *counter)
464 struct perf_counter_context *ctx = counter->ctx;
467 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
470 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
472 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
473 run_end = counter->tstamp_stopped;
477 counter->total_time_running = run_end - counter->tstamp_running;
481 * Update total_time_enabled and total_time_running for all counters in a group.
483 static void update_group_times(struct perf_counter *leader)
485 struct perf_counter *counter;
487 update_counter_times(leader);
488 list_for_each_entry(counter, &leader->sibling_list, list_entry)
489 update_counter_times(counter);
493 * Cross CPU call to disable a performance counter
495 static void __perf_counter_disable(void *info)
497 struct perf_counter *counter = info;
498 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
499 struct perf_counter_context *ctx = counter->ctx;
502 * If this is a per-task counter, need to check whether this
503 * counter's task is the current task on this cpu.
505 if (ctx->task && cpuctx->task_ctx != ctx)
508 spin_lock(&ctx->lock);
511 * If the counter is on, turn it off.
512 * If it is in error state, leave it in error state.
514 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
515 update_context_time(ctx);
516 update_counter_times(counter);
517 if (counter == counter->group_leader)
518 group_sched_out(counter, cpuctx, ctx);
520 counter_sched_out(counter, cpuctx, ctx);
521 counter->state = PERF_COUNTER_STATE_OFF;
524 spin_unlock(&ctx->lock);
530 * If counter->ctx is a cloned context, callers must make sure that
531 * every task struct that counter->ctx->task could possibly point to
532 * remains valid. This condition is satisifed when called through
533 * perf_counter_for_each_child or perf_counter_for_each because they
534 * hold the top-level counter's child_mutex, so any descendant that
535 * goes to exit will block in sync_child_counter.
536 * When called from perf_pending_counter it's OK because counter->ctx
537 * is the current context on this CPU and preemption is disabled,
538 * hence we can't get into perf_counter_task_sched_out for this context.
540 static void perf_counter_disable(struct perf_counter *counter)
542 struct perf_counter_context *ctx = counter->ctx;
543 struct task_struct *task = ctx->task;
547 * Disable the counter on the cpu that it's on
549 smp_call_function_single(counter->cpu, __perf_counter_disable,
555 task_oncpu_function_call(task, __perf_counter_disable, counter);
557 spin_lock_irq(&ctx->lock);
559 * If the counter is still active, we need to retry the cross-call.
561 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
562 spin_unlock_irq(&ctx->lock);
567 * Since we have the lock this context can't be scheduled
568 * in, so we can change the state safely.
570 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
571 update_counter_times(counter);
572 counter->state = PERF_COUNTER_STATE_OFF;
575 spin_unlock_irq(&ctx->lock);
579 counter_sched_in(struct perf_counter *counter,
580 struct perf_cpu_context *cpuctx,
581 struct perf_counter_context *ctx,
584 if (counter->state <= PERF_COUNTER_STATE_OFF)
587 counter->state = PERF_COUNTER_STATE_ACTIVE;
588 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
590 * The new state must be visible before we turn it on in the hardware:
594 if (counter->pmu->enable(counter)) {
595 counter->state = PERF_COUNTER_STATE_INACTIVE;
600 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
602 if (!is_software_counter(counter))
603 cpuctx->active_oncpu++;
606 if (counter->attr.exclusive)
607 cpuctx->exclusive = 1;
613 group_sched_in(struct perf_counter *group_counter,
614 struct perf_cpu_context *cpuctx,
615 struct perf_counter_context *ctx,
618 struct perf_counter *counter, *partial_group;
621 if (group_counter->state == PERF_COUNTER_STATE_OFF)
624 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
626 return ret < 0 ? ret : 0;
628 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
632 * Schedule in siblings as one group (if any):
634 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
635 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
636 partial_group = counter;
645 * Groups can be scheduled in as one unit only, so undo any
646 * partial group before returning:
648 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
649 if (counter == partial_group)
651 counter_sched_out(counter, cpuctx, ctx);
653 counter_sched_out(group_counter, cpuctx, ctx);
659 * Return 1 for a group consisting entirely of software counters,
660 * 0 if the group contains any hardware counters.
662 static int is_software_only_group(struct perf_counter *leader)
664 struct perf_counter *counter;
666 if (!is_software_counter(leader))
669 list_for_each_entry(counter, &leader->sibling_list, list_entry)
670 if (!is_software_counter(counter))
677 * Work out whether we can put this counter group on the CPU now.
679 static int group_can_go_on(struct perf_counter *counter,
680 struct perf_cpu_context *cpuctx,
684 * Groups consisting entirely of software counters can always go on.
686 if (is_software_only_group(counter))
689 * If an exclusive group is already on, no other hardware
690 * counters can go on.
692 if (cpuctx->exclusive)
695 * If this group is exclusive and there are already
696 * counters on the CPU, it can't go on.
698 if (counter->attr.exclusive && cpuctx->active_oncpu)
701 * Otherwise, try to add it if all previous groups were able
707 static void add_counter_to_ctx(struct perf_counter *counter,
708 struct perf_counter_context *ctx)
710 list_add_counter(counter, ctx);
711 counter->tstamp_enabled = ctx->time;
712 counter->tstamp_running = ctx->time;
713 counter->tstamp_stopped = ctx->time;
717 * Cross CPU call to install and enable a performance counter
719 * Must be called with ctx->mutex held
721 static void __perf_install_in_context(void *info)
723 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
724 struct perf_counter *counter = info;
725 struct perf_counter_context *ctx = counter->ctx;
726 struct perf_counter *leader = counter->group_leader;
727 int cpu = smp_processor_id();
731 * If this is a task context, we need to check whether it is
732 * the current task context of this cpu. If not it has been
733 * scheduled out before the smp call arrived.
734 * Or possibly this is the right context but it isn't
735 * on this cpu because it had no counters.
737 if (ctx->task && cpuctx->task_ctx != ctx) {
738 if (cpuctx->task_ctx || ctx->task != current)
740 cpuctx->task_ctx = ctx;
743 spin_lock(&ctx->lock);
745 update_context_time(ctx);
748 * Protect the list operation against NMI by disabling the
749 * counters on a global level. NOP for non NMI based counters.
753 add_counter_to_ctx(counter, ctx);
756 * Don't put the counter on if it is disabled or if
757 * it is in a group and the group isn't on.
759 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
760 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
764 * An exclusive counter can't go on if there are already active
765 * hardware counters, and no hardware counter can go on if there
766 * is already an exclusive counter on.
768 if (!group_can_go_on(counter, cpuctx, 1))
771 err = counter_sched_in(counter, cpuctx, ctx, cpu);
775 * This counter couldn't go on. If it is in a group
776 * then we have to pull the whole group off.
777 * If the counter group is pinned then put it in error state.
779 if (leader != counter)
780 group_sched_out(leader, cpuctx, ctx);
781 if (leader->attr.pinned) {
782 update_group_times(leader);
783 leader->state = PERF_COUNTER_STATE_ERROR;
787 if (!err && !ctx->task && cpuctx->max_pertask)
788 cpuctx->max_pertask--;
793 spin_unlock(&ctx->lock);
797 * Attach a performance counter to a context
799 * First we add the counter to the list with the hardware enable bit
800 * in counter->hw_config cleared.
802 * If the counter is attached to a task which is on a CPU we use a smp
803 * call to enable it in the task context. The task might have been
804 * scheduled away, but we check this in the smp call again.
806 * Must be called with ctx->mutex held.
809 perf_install_in_context(struct perf_counter_context *ctx,
810 struct perf_counter *counter,
813 struct task_struct *task = ctx->task;
817 * Per cpu counters are installed via an smp call and
818 * the install is always sucessful.
820 smp_call_function_single(cpu, __perf_install_in_context,
826 task_oncpu_function_call(task, __perf_install_in_context,
829 spin_lock_irq(&ctx->lock);
831 * we need to retry the smp call.
833 if (ctx->is_active && list_empty(&counter->list_entry)) {
834 spin_unlock_irq(&ctx->lock);
839 * The lock prevents that this context is scheduled in so we
840 * can add the counter safely, if it the call above did not
843 if (list_empty(&counter->list_entry))
844 add_counter_to_ctx(counter, ctx);
845 spin_unlock_irq(&ctx->lock);
849 * Cross CPU call to enable a performance counter
851 static void __perf_counter_enable(void *info)
853 struct perf_counter *counter = info;
854 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
855 struct perf_counter_context *ctx = counter->ctx;
856 struct perf_counter *leader = counter->group_leader;
860 * If this is a per-task counter, need to check whether this
861 * counter's task is the current task on this cpu.
863 if (ctx->task && cpuctx->task_ctx != ctx) {
864 if (cpuctx->task_ctx || ctx->task != current)
866 cpuctx->task_ctx = ctx;
869 spin_lock(&ctx->lock);
871 update_context_time(ctx);
873 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
875 counter->state = PERF_COUNTER_STATE_INACTIVE;
876 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
879 * If the counter is in a group and isn't the group leader,
880 * then don't put it on unless the group is on.
882 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
885 if (!group_can_go_on(counter, cpuctx, 1)) {
889 if (counter == leader)
890 err = group_sched_in(counter, cpuctx, ctx,
893 err = counter_sched_in(counter, cpuctx, ctx,
900 * If this counter can't go on and it's part of a
901 * group, then the whole group has to come off.
903 if (leader != counter)
904 group_sched_out(leader, cpuctx, ctx);
905 if (leader->attr.pinned) {
906 update_group_times(leader);
907 leader->state = PERF_COUNTER_STATE_ERROR;
912 spin_unlock(&ctx->lock);
918 * If counter->ctx is a cloned context, callers must make sure that
919 * every task struct that counter->ctx->task could possibly point to
920 * remains valid. This condition is satisfied when called through
921 * perf_counter_for_each_child or perf_counter_for_each as described
922 * for perf_counter_disable.
924 static void perf_counter_enable(struct perf_counter *counter)
926 struct perf_counter_context *ctx = counter->ctx;
927 struct task_struct *task = ctx->task;
931 * Enable the counter on the cpu that it's on
933 smp_call_function_single(counter->cpu, __perf_counter_enable,
938 spin_lock_irq(&ctx->lock);
939 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
943 * If the counter is in error state, clear that first.
944 * That way, if we see the counter in error state below, we
945 * know that it has gone back into error state, as distinct
946 * from the task having been scheduled away before the
947 * cross-call arrived.
949 if (counter->state == PERF_COUNTER_STATE_ERROR)
950 counter->state = PERF_COUNTER_STATE_OFF;
953 spin_unlock_irq(&ctx->lock);
954 task_oncpu_function_call(task, __perf_counter_enable, counter);
956 spin_lock_irq(&ctx->lock);
959 * If the context is active and the counter is still off,
960 * we need to retry the cross-call.
962 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
966 * Since we have the lock this context can't be scheduled
967 * in, so we can change the state safely.
969 if (counter->state == PERF_COUNTER_STATE_OFF) {
970 counter->state = PERF_COUNTER_STATE_INACTIVE;
971 counter->tstamp_enabled =
972 ctx->time - counter->total_time_enabled;
975 spin_unlock_irq(&ctx->lock);
978 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
981 * not supported on inherited counters
983 if (counter->attr.inherit)
986 atomic_add(refresh, &counter->event_limit);
987 perf_counter_enable(counter);
992 void __perf_counter_sched_out(struct perf_counter_context *ctx,
993 struct perf_cpu_context *cpuctx)
995 struct perf_counter *counter;
997 spin_lock(&ctx->lock);
999 if (likely(!ctx->nr_counters))
1001 update_context_time(ctx);
1004 if (ctx->nr_active) {
1005 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1006 if (counter != counter->group_leader)
1007 counter_sched_out(counter, cpuctx, ctx);
1009 group_sched_out(counter, cpuctx, ctx);
1014 spin_unlock(&ctx->lock);
1018 * Test whether two contexts are equivalent, i.e. whether they
1019 * have both been cloned from the same version of the same context
1020 * and they both have the same number of enabled counters.
1021 * If the number of enabled counters is the same, then the set
1022 * of enabled counters should be the same, because these are both
1023 * inherited contexts, therefore we can't access individual counters
1024 * in them directly with an fd; we can only enable/disable all
1025 * counters via prctl, or enable/disable all counters in a family
1026 * via ioctl, which will have the same effect on both contexts.
1028 static int context_equiv(struct perf_counter_context *ctx1,
1029 struct perf_counter_context *ctx2)
1031 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1032 && ctx1->parent_gen == ctx2->parent_gen
1033 && !ctx1->pin_count && !ctx2->pin_count;
1036 static void __perf_counter_read(void *counter);
1038 static void __perf_counter_sync_stat(struct perf_counter *counter,
1039 struct perf_counter *next_counter)
1043 if (!counter->attr.inherit_stat)
1047 * Update the counter value, we cannot use perf_counter_read()
1048 * because we're in the middle of a context switch and have IRQs
1049 * disabled, which upsets smp_call_function_single(), however
1050 * we know the counter must be on the current CPU, therefore we
1051 * don't need to use it.
1053 switch (counter->state) {
1054 case PERF_COUNTER_STATE_ACTIVE:
1055 __perf_counter_read(counter);
1058 case PERF_COUNTER_STATE_INACTIVE:
1059 update_counter_times(counter);
1067 * In order to keep per-task stats reliable we need to flip the counter
1068 * values when we flip the contexts.
1070 value = atomic64_read(&next_counter->count);
1071 value = atomic64_xchg(&counter->count, value);
1072 atomic64_set(&next_counter->count, value);
1074 swap(counter->total_time_enabled, next_counter->total_time_enabled);
1075 swap(counter->total_time_running, next_counter->total_time_running);
1078 * Since we swizzled the values, update the user visible data too.
1080 perf_counter_update_userpage(counter);
1081 perf_counter_update_userpage(next_counter);
1084 #define list_next_entry(pos, member) \
1085 list_entry(pos->member.next, typeof(*pos), member)
1087 static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1088 struct perf_counter_context *next_ctx)
1090 struct perf_counter *counter, *next_counter;
1095 counter = list_first_entry(&ctx->event_list,
1096 struct perf_counter, event_entry);
1098 next_counter = list_first_entry(&next_ctx->event_list,
1099 struct perf_counter, event_entry);
1101 while (&counter->event_entry != &ctx->event_list &&
1102 &next_counter->event_entry != &next_ctx->event_list) {
1104 __perf_counter_sync_stat(counter, next_counter);
1106 counter = list_next_entry(counter, event_entry);
1107 next_counter = list_next_entry(next_counter, event_entry);
1112 * Called from scheduler to remove the counters of the current task,
1113 * with interrupts disabled.
1115 * We stop each counter and update the counter value in counter->count.
1117 * This does not protect us against NMI, but disable()
1118 * sets the disabled bit in the control field of counter _before_
1119 * accessing the counter control register. If a NMI hits, then it will
1120 * not restart the counter.
1122 void perf_counter_task_sched_out(struct task_struct *task,
1123 struct task_struct *next, int cpu)
1125 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1126 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1127 struct perf_counter_context *next_ctx;
1128 struct perf_counter_context *parent;
1129 struct pt_regs *regs;
1132 regs = task_pt_regs(task);
1133 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1135 if (likely(!ctx || !cpuctx->task_ctx))
1138 update_context_time(ctx);
1141 parent = rcu_dereference(ctx->parent_ctx);
1142 next_ctx = next->perf_counter_ctxp;
1143 if (parent && next_ctx &&
1144 rcu_dereference(next_ctx->parent_ctx) == parent) {
1146 * Looks like the two contexts are clones, so we might be
1147 * able to optimize the context switch. We lock both
1148 * contexts and check that they are clones under the
1149 * lock (including re-checking that neither has been
1150 * uncloned in the meantime). It doesn't matter which
1151 * order we take the locks because no other cpu could
1152 * be trying to lock both of these tasks.
1154 spin_lock(&ctx->lock);
1155 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1156 if (context_equiv(ctx, next_ctx)) {
1158 * XXX do we need a memory barrier of sorts
1159 * wrt to rcu_dereference() of perf_counter_ctxp
1161 task->perf_counter_ctxp = next_ctx;
1162 next->perf_counter_ctxp = ctx;
1164 next_ctx->task = task;
1167 perf_counter_sync_stat(ctx, next_ctx);
1169 spin_unlock(&next_ctx->lock);
1170 spin_unlock(&ctx->lock);
1175 __perf_counter_sched_out(ctx, cpuctx);
1176 cpuctx->task_ctx = NULL;
1181 * Called with IRQs disabled
1183 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1185 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1187 if (!cpuctx->task_ctx)
1190 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1193 __perf_counter_sched_out(ctx, cpuctx);
1194 cpuctx->task_ctx = NULL;
1198 * Called with IRQs disabled
1200 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1202 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1206 __perf_counter_sched_in(struct perf_counter_context *ctx,
1207 struct perf_cpu_context *cpuctx, int cpu)
1209 struct perf_counter *counter;
1212 spin_lock(&ctx->lock);
1214 if (likely(!ctx->nr_counters))
1217 ctx->timestamp = perf_clock();
1222 * First go through the list and put on any pinned groups
1223 * in order to give them the best chance of going on.
1225 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1226 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1227 !counter->attr.pinned)
1229 if (counter->cpu != -1 && counter->cpu != cpu)
1232 if (counter != counter->group_leader)
1233 counter_sched_in(counter, cpuctx, ctx, cpu);
1235 if (group_can_go_on(counter, cpuctx, 1))
1236 group_sched_in(counter, cpuctx, ctx, cpu);
1240 * If this pinned group hasn't been scheduled,
1241 * put it in error state.
1243 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1244 update_group_times(counter);
1245 counter->state = PERF_COUNTER_STATE_ERROR;
1249 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1251 * Ignore counters in OFF or ERROR state, and
1252 * ignore pinned counters since we did them already.
1254 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1255 counter->attr.pinned)
1259 * Listen to the 'cpu' scheduling filter constraint
1262 if (counter->cpu != -1 && counter->cpu != cpu)
1265 if (counter != counter->group_leader) {
1266 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1269 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1270 if (group_sched_in(counter, cpuctx, ctx, cpu))
1277 spin_unlock(&ctx->lock);
1281 * Called from scheduler to add the counters of the current task
1282 * with interrupts disabled.
1284 * We restore the counter value and then enable it.
1286 * This does not protect us against NMI, but enable()
1287 * sets the enabled bit in the control field of counter _before_
1288 * accessing the counter control register. If a NMI hits, then it will
1289 * keep the counter running.
1291 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1293 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1294 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1298 if (cpuctx->task_ctx == ctx)
1300 __perf_counter_sched_in(ctx, cpuctx, cpu);
1301 cpuctx->task_ctx = ctx;
1304 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1306 struct perf_counter_context *ctx = &cpuctx->ctx;
1308 __perf_counter_sched_in(ctx, cpuctx, cpu);
1311 #define MAX_INTERRUPTS (~0ULL)
1313 static void perf_log_throttle(struct perf_counter *counter, int enable);
1315 static void perf_adjust_period(struct perf_counter *counter, u64 events)
1317 struct hw_perf_counter *hwc = &counter->hw;
1318 u64 period, sample_period;
1321 events *= hwc->sample_period;
1322 period = div64_u64(events, counter->attr.sample_freq);
1324 delta = (s64)(period - hwc->sample_period);
1325 delta = (delta + 7) / 8; /* low pass filter */
1327 sample_period = hwc->sample_period + delta;
1332 hwc->sample_period = sample_period;
1335 static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1337 struct perf_counter *counter;
1338 struct hw_perf_counter *hwc;
1339 u64 interrupts, freq;
1341 spin_lock(&ctx->lock);
1342 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1343 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1348 interrupts = hwc->interrupts;
1349 hwc->interrupts = 0;
1352 * unthrottle counters on the tick
1354 if (interrupts == MAX_INTERRUPTS) {
1355 perf_log_throttle(counter, 1);
1356 counter->pmu->unthrottle(counter);
1357 interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1360 if (!counter->attr.freq || !counter->attr.sample_freq)
1364 * if the specified freq < HZ then we need to skip ticks
1366 if (counter->attr.sample_freq < HZ) {
1367 freq = counter->attr.sample_freq;
1369 hwc->freq_count += freq;
1370 hwc->freq_interrupts += interrupts;
1372 if (hwc->freq_count < HZ)
1375 interrupts = hwc->freq_interrupts;
1376 hwc->freq_interrupts = 0;
1377 hwc->freq_count -= HZ;
1381 perf_adjust_period(counter, freq * interrupts);
1384 * In order to avoid being stalled by an (accidental) huge
1385 * sample period, force reset the sample period if we didn't
1386 * get any events in this freq period.
1390 counter->pmu->disable(counter);
1391 atomic64_set(&hwc->period_left, 0);
1392 counter->pmu->enable(counter);
1396 spin_unlock(&ctx->lock);
1400 * Round-robin a context's counters:
1402 static void rotate_ctx(struct perf_counter_context *ctx)
1404 struct perf_counter *counter;
1406 if (!ctx->nr_counters)
1409 spin_lock(&ctx->lock);
1411 * Rotate the first entry last (works just fine for group counters too):
1414 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1415 list_move_tail(&counter->list_entry, &ctx->counter_list);
1420 spin_unlock(&ctx->lock);
1423 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1425 struct perf_cpu_context *cpuctx;
1426 struct perf_counter_context *ctx;
1428 if (!atomic_read(&nr_counters))
1431 cpuctx = &per_cpu(perf_cpu_context, cpu);
1432 ctx = curr->perf_counter_ctxp;
1434 perf_ctx_adjust_freq(&cpuctx->ctx);
1436 perf_ctx_adjust_freq(ctx);
1438 perf_counter_cpu_sched_out(cpuctx);
1440 __perf_counter_task_sched_out(ctx);
1442 rotate_ctx(&cpuctx->ctx);
1446 perf_counter_cpu_sched_in(cpuctx, cpu);
1448 perf_counter_task_sched_in(curr, cpu);
1452 * Enable all of a task's counters that have been marked enable-on-exec.
1453 * This expects task == current.
1455 static void perf_counter_enable_on_exec(struct task_struct *task)
1457 struct perf_counter_context *ctx;
1458 struct perf_counter *counter;
1459 unsigned long flags;
1462 local_irq_save(flags);
1463 ctx = task->perf_counter_ctxp;
1464 if (!ctx || !ctx->nr_counters)
1467 __perf_counter_task_sched_out(ctx);
1469 spin_lock(&ctx->lock);
1471 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1472 if (!counter->attr.enable_on_exec)
1474 counter->attr.enable_on_exec = 0;
1475 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
1477 counter->state = PERF_COUNTER_STATE_INACTIVE;
1478 counter->tstamp_enabled =
1479 ctx->time - counter->total_time_enabled;
1484 * Unclone this context if we enabled any counter.
1489 spin_unlock(&ctx->lock);
1491 perf_counter_task_sched_in(task, smp_processor_id());
1493 local_irq_restore(flags);
1497 * Cross CPU call to read the hardware counter
1499 static void __perf_counter_read(void *info)
1501 struct perf_counter *counter = info;
1502 struct perf_counter_context *ctx = counter->ctx;
1503 unsigned long flags;
1505 local_irq_save(flags);
1507 update_context_time(ctx);
1508 counter->pmu->read(counter);
1509 update_counter_times(counter);
1510 local_irq_restore(flags);
1513 static u64 perf_counter_read(struct perf_counter *counter)
1516 * If counter is enabled and currently active on a CPU, update the
1517 * value in the counter structure:
1519 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1520 smp_call_function_single(counter->oncpu,
1521 __perf_counter_read, counter, 1);
1522 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1523 update_counter_times(counter);
1526 return atomic64_read(&counter->count);
1530 * Initialize the perf_counter context in a task_struct:
1533 __perf_counter_init_context(struct perf_counter_context *ctx,
1534 struct task_struct *task)
1536 memset(ctx, 0, sizeof(*ctx));
1537 spin_lock_init(&ctx->lock);
1538 mutex_init(&ctx->mutex);
1539 INIT_LIST_HEAD(&ctx->counter_list);
1540 INIT_LIST_HEAD(&ctx->event_list);
1541 atomic_set(&ctx->refcount, 1);
1545 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1547 struct perf_counter_context *ctx;
1548 struct perf_cpu_context *cpuctx;
1549 struct task_struct *task;
1550 unsigned long flags;
1554 * If cpu is not a wildcard then this is a percpu counter:
1557 /* Must be root to operate on a CPU counter: */
1558 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1559 return ERR_PTR(-EACCES);
1561 if (cpu < 0 || cpu > num_possible_cpus())
1562 return ERR_PTR(-EINVAL);
1565 * We could be clever and allow to attach a counter to an
1566 * offline CPU and activate it when the CPU comes up, but
1569 if (!cpu_isset(cpu, cpu_online_map))
1570 return ERR_PTR(-ENODEV);
1572 cpuctx = &per_cpu(perf_cpu_context, cpu);
1583 task = find_task_by_vpid(pid);
1585 get_task_struct(task);
1589 return ERR_PTR(-ESRCH);
1592 * Can't attach counters to a dying task.
1595 if (task->flags & PF_EXITING)
1598 /* Reuse ptrace permission checks for now. */
1600 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1604 ctx = perf_lock_task_context(task, &flags);
1607 spin_unlock_irqrestore(&ctx->lock, flags);
1611 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1615 __perf_counter_init_context(ctx, task);
1617 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1619 * We raced with some other task; use
1620 * the context they set.
1625 get_task_struct(task);
1628 put_task_struct(task);
1632 put_task_struct(task);
1633 return ERR_PTR(err);
1636 static void free_counter_rcu(struct rcu_head *head)
1638 struct perf_counter *counter;
1640 counter = container_of(head, struct perf_counter, rcu_head);
1642 put_pid_ns(counter->ns);
1646 static void perf_pending_sync(struct perf_counter *counter);
1648 static void free_counter(struct perf_counter *counter)
1650 perf_pending_sync(counter);
1652 if (!counter->parent) {
1653 atomic_dec(&nr_counters);
1654 if (counter->attr.mmap)
1655 atomic_dec(&nr_mmap_counters);
1656 if (counter->attr.comm)
1657 atomic_dec(&nr_comm_counters);
1658 if (counter->attr.task)
1659 atomic_dec(&nr_task_counters);
1662 if (counter->destroy)
1663 counter->destroy(counter);
1665 put_ctx(counter->ctx);
1666 call_rcu(&counter->rcu_head, free_counter_rcu);
1670 * Called when the last reference to the file is gone.
1672 static int perf_release(struct inode *inode, struct file *file)
1674 struct perf_counter *counter = file->private_data;
1675 struct perf_counter_context *ctx = counter->ctx;
1677 file->private_data = NULL;
1679 WARN_ON_ONCE(ctx->parent_ctx);
1680 mutex_lock(&ctx->mutex);
1681 perf_counter_remove_from_context(counter);
1682 mutex_unlock(&ctx->mutex);
1684 mutex_lock(&counter->owner->perf_counter_mutex);
1685 list_del_init(&counter->owner_entry);
1686 mutex_unlock(&counter->owner->perf_counter_mutex);
1687 put_task_struct(counter->owner);
1689 free_counter(counter);
1694 static u64 perf_counter_read_tree(struct perf_counter *counter)
1696 struct perf_counter *child;
1699 total += perf_counter_read(counter);
1700 list_for_each_entry(child, &counter->child_list, child_list)
1701 total += perf_counter_read(child);
1707 * Read the performance counter - simple non blocking version for now
1710 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1716 * Return end-of-file for a read on a counter that is in
1717 * error state (i.e. because it was pinned but it couldn't be
1718 * scheduled on to the CPU at some point).
1720 if (counter->state == PERF_COUNTER_STATE_ERROR)
1723 WARN_ON_ONCE(counter->ctx->parent_ctx);
1724 mutex_lock(&counter->child_mutex);
1725 values[0] = perf_counter_read_tree(counter);
1727 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1728 values[n++] = counter->total_time_enabled +
1729 atomic64_read(&counter->child_total_time_enabled);
1730 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1731 values[n++] = counter->total_time_running +
1732 atomic64_read(&counter->child_total_time_running);
1733 if (counter->attr.read_format & PERF_FORMAT_ID)
1734 values[n++] = primary_counter_id(counter);
1735 mutex_unlock(&counter->child_mutex);
1737 if (count < n * sizeof(u64))
1739 count = n * sizeof(u64);
1741 if (copy_to_user(buf, values, count))
1748 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1750 struct perf_counter *counter = file->private_data;
1752 return perf_read_hw(counter, buf, count);
1755 static unsigned int perf_poll(struct file *file, poll_table *wait)
1757 struct perf_counter *counter = file->private_data;
1758 struct perf_mmap_data *data;
1759 unsigned int events = POLL_HUP;
1762 data = rcu_dereference(counter->data);
1764 events = atomic_xchg(&data->poll, 0);
1767 poll_wait(file, &counter->waitq, wait);
1772 static void perf_counter_reset(struct perf_counter *counter)
1774 (void)perf_counter_read(counter);
1775 atomic64_set(&counter->count, 0);
1776 perf_counter_update_userpage(counter);
1780 * Holding the top-level counter's child_mutex means that any
1781 * descendant process that has inherited this counter will block
1782 * in sync_child_counter if it goes to exit, thus satisfying the
1783 * task existence requirements of perf_counter_enable/disable.
1785 static void perf_counter_for_each_child(struct perf_counter *counter,
1786 void (*func)(struct perf_counter *))
1788 struct perf_counter *child;
1790 WARN_ON_ONCE(counter->ctx->parent_ctx);
1791 mutex_lock(&counter->child_mutex);
1793 list_for_each_entry(child, &counter->child_list, child_list)
1795 mutex_unlock(&counter->child_mutex);
1798 static void perf_counter_for_each(struct perf_counter *counter,
1799 void (*func)(struct perf_counter *))
1801 struct perf_counter_context *ctx = counter->ctx;
1802 struct perf_counter *sibling;
1804 WARN_ON_ONCE(ctx->parent_ctx);
1805 mutex_lock(&ctx->mutex);
1806 counter = counter->group_leader;
1808 perf_counter_for_each_child(counter, func);
1810 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1811 perf_counter_for_each_child(counter, func);
1812 mutex_unlock(&ctx->mutex);
1815 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1817 struct perf_counter_context *ctx = counter->ctx;
1822 if (!counter->attr.sample_period)
1825 size = copy_from_user(&value, arg, sizeof(value));
1826 if (size != sizeof(value))
1832 spin_lock_irq(&ctx->lock);
1833 if (counter->attr.freq) {
1834 if (value > sysctl_perf_counter_sample_rate) {
1839 counter->attr.sample_freq = value;
1841 counter->attr.sample_period = value;
1842 counter->hw.sample_period = value;
1845 spin_unlock_irq(&ctx->lock);
1850 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1852 struct perf_counter *counter = file->private_data;
1853 void (*func)(struct perf_counter *);
1857 case PERF_COUNTER_IOC_ENABLE:
1858 func = perf_counter_enable;
1860 case PERF_COUNTER_IOC_DISABLE:
1861 func = perf_counter_disable;
1863 case PERF_COUNTER_IOC_RESET:
1864 func = perf_counter_reset;
1867 case PERF_COUNTER_IOC_REFRESH:
1868 return perf_counter_refresh(counter, arg);
1870 case PERF_COUNTER_IOC_PERIOD:
1871 return perf_counter_period(counter, (u64 __user *)arg);
1877 if (flags & PERF_IOC_FLAG_GROUP)
1878 perf_counter_for_each(counter, func);
1880 perf_counter_for_each_child(counter, func);
1885 int perf_counter_task_enable(void)
1887 struct perf_counter *counter;
1889 mutex_lock(¤t->perf_counter_mutex);
1890 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1891 perf_counter_for_each_child(counter, perf_counter_enable);
1892 mutex_unlock(¤t->perf_counter_mutex);
1897 int perf_counter_task_disable(void)
1899 struct perf_counter *counter;
1901 mutex_lock(¤t->perf_counter_mutex);
1902 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1903 perf_counter_for_each_child(counter, perf_counter_disable);
1904 mutex_unlock(¤t->perf_counter_mutex);
1909 static int perf_counter_index(struct perf_counter *counter)
1911 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1914 return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
1918 * Callers need to ensure there can be no nesting of this function, otherwise
1919 * the seqlock logic goes bad. We can not serialize this because the arch
1920 * code calls this from NMI context.
1922 void perf_counter_update_userpage(struct perf_counter *counter)
1924 struct perf_counter_mmap_page *userpg;
1925 struct perf_mmap_data *data;
1928 data = rcu_dereference(counter->data);
1932 userpg = data->user_page;
1935 * Disable preemption so as to not let the corresponding user-space
1936 * spin too long if we get preempted.
1941 userpg->index = perf_counter_index(counter);
1942 userpg->offset = atomic64_read(&counter->count);
1943 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1944 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1946 userpg->time_enabled = counter->total_time_enabled +
1947 atomic64_read(&counter->child_total_time_enabled);
1949 userpg->time_running = counter->total_time_running +
1950 atomic64_read(&counter->child_total_time_running);
1959 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1961 struct perf_counter *counter = vma->vm_file->private_data;
1962 struct perf_mmap_data *data;
1963 int ret = VM_FAULT_SIGBUS;
1965 if (vmf->flags & FAULT_FLAG_MKWRITE) {
1966 if (vmf->pgoff == 0)
1972 data = rcu_dereference(counter->data);
1976 if (vmf->pgoff == 0) {
1977 vmf->page = virt_to_page(data->user_page);
1979 int nr = vmf->pgoff - 1;
1981 if ((unsigned)nr > data->nr_pages)
1984 if (vmf->flags & FAULT_FLAG_WRITE)
1987 vmf->page = virt_to_page(data->data_pages[nr]);
1990 get_page(vmf->page);
1991 vmf->page->mapping = vma->vm_file->f_mapping;
1992 vmf->page->index = vmf->pgoff;
2001 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
2003 struct perf_mmap_data *data;
2007 WARN_ON(atomic_read(&counter->mmap_count));
2009 size = sizeof(struct perf_mmap_data);
2010 size += nr_pages * sizeof(void *);
2012 data = kzalloc(size, GFP_KERNEL);
2016 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2017 if (!data->user_page)
2018 goto fail_user_page;
2020 for (i = 0; i < nr_pages; i++) {
2021 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2022 if (!data->data_pages[i])
2023 goto fail_data_pages;
2026 data->nr_pages = nr_pages;
2027 atomic_set(&data->lock, -1);
2029 rcu_assign_pointer(counter->data, data);
2034 for (i--; i >= 0; i--)
2035 free_page((unsigned long)data->data_pages[i]);
2037 free_page((unsigned long)data->user_page);
2046 static void perf_mmap_free_page(unsigned long addr)
2048 struct page *page = virt_to_page((void *)addr);
2050 page->mapping = NULL;
2054 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2056 struct perf_mmap_data *data;
2059 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2061 perf_mmap_free_page((unsigned long)data->user_page);
2062 for (i = 0; i < data->nr_pages; i++)
2063 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2068 static void perf_mmap_data_free(struct perf_counter *counter)
2070 struct perf_mmap_data *data = counter->data;
2072 WARN_ON(atomic_read(&counter->mmap_count));
2074 rcu_assign_pointer(counter->data, NULL);
2075 call_rcu(&data->rcu_head, __perf_mmap_data_free);
2078 static void perf_mmap_open(struct vm_area_struct *vma)
2080 struct perf_counter *counter = vma->vm_file->private_data;
2082 atomic_inc(&counter->mmap_count);
2085 static void perf_mmap_close(struct vm_area_struct *vma)
2087 struct perf_counter *counter = vma->vm_file->private_data;
2089 WARN_ON_ONCE(counter->ctx->parent_ctx);
2090 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2091 struct user_struct *user = current_user();
2093 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2094 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2095 perf_mmap_data_free(counter);
2096 mutex_unlock(&counter->mmap_mutex);
2100 static struct vm_operations_struct perf_mmap_vmops = {
2101 .open = perf_mmap_open,
2102 .close = perf_mmap_close,
2103 .fault = perf_mmap_fault,
2104 .page_mkwrite = perf_mmap_fault,
2107 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2109 struct perf_counter *counter = file->private_data;
2110 unsigned long user_locked, user_lock_limit;
2111 struct user_struct *user = current_user();
2112 unsigned long locked, lock_limit;
2113 unsigned long vma_size;
2114 unsigned long nr_pages;
2115 long user_extra, extra;
2118 if (!(vma->vm_flags & VM_SHARED))
2121 vma_size = vma->vm_end - vma->vm_start;
2122 nr_pages = (vma_size / PAGE_SIZE) - 1;
2125 * If we have data pages ensure they're a power-of-two number, so we
2126 * can do bitmasks instead of modulo.
2128 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2131 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2134 if (vma->vm_pgoff != 0)
2137 WARN_ON_ONCE(counter->ctx->parent_ctx);
2138 mutex_lock(&counter->mmap_mutex);
2139 if (atomic_inc_not_zero(&counter->mmap_count)) {
2140 if (nr_pages != counter->data->nr_pages)
2145 user_extra = nr_pages + 1;
2146 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2149 * Increase the limit linearly with more CPUs:
2151 user_lock_limit *= num_online_cpus();
2153 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2156 if (user_locked > user_lock_limit)
2157 extra = user_locked - user_lock_limit;
2159 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2160 lock_limit >>= PAGE_SHIFT;
2161 locked = vma->vm_mm->locked_vm + extra;
2163 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2168 WARN_ON(counter->data);
2169 ret = perf_mmap_data_alloc(counter, nr_pages);
2173 atomic_set(&counter->mmap_count, 1);
2174 atomic_long_add(user_extra, &user->locked_vm);
2175 vma->vm_mm->locked_vm += extra;
2176 counter->data->nr_locked = extra;
2177 if (vma->vm_flags & VM_WRITE)
2178 counter->data->writable = 1;
2181 mutex_unlock(&counter->mmap_mutex);
2183 vma->vm_flags |= VM_RESERVED;
2184 vma->vm_ops = &perf_mmap_vmops;
2189 static int perf_fasync(int fd, struct file *filp, int on)
2191 struct inode *inode = filp->f_path.dentry->d_inode;
2192 struct perf_counter *counter = filp->private_data;
2195 mutex_lock(&inode->i_mutex);
2196 retval = fasync_helper(fd, filp, on, &counter->fasync);
2197 mutex_unlock(&inode->i_mutex);
2205 static const struct file_operations perf_fops = {
2206 .release = perf_release,
2209 .unlocked_ioctl = perf_ioctl,
2210 .compat_ioctl = perf_ioctl,
2212 .fasync = perf_fasync,
2216 * Perf counter wakeup
2218 * If there's data, ensure we set the poll() state and publish everything
2219 * to user-space before waking everybody up.
2222 void perf_counter_wakeup(struct perf_counter *counter)
2224 wake_up_all(&counter->waitq);
2226 if (counter->pending_kill) {
2227 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2228 counter->pending_kill = 0;
2235 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2237 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2238 * single linked list and use cmpxchg() to add entries lockless.
2241 static void perf_pending_counter(struct perf_pending_entry *entry)
2243 struct perf_counter *counter = container_of(entry,
2244 struct perf_counter, pending);
2246 if (counter->pending_disable) {
2247 counter->pending_disable = 0;
2248 perf_counter_disable(counter);
2251 if (counter->pending_wakeup) {
2252 counter->pending_wakeup = 0;
2253 perf_counter_wakeup(counter);
2257 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2259 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2263 static void perf_pending_queue(struct perf_pending_entry *entry,
2264 void (*func)(struct perf_pending_entry *))
2266 struct perf_pending_entry **head;
2268 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2273 head = &get_cpu_var(perf_pending_head);
2276 entry->next = *head;
2277 } while (cmpxchg(head, entry->next, entry) != entry->next);
2279 set_perf_counter_pending();
2281 put_cpu_var(perf_pending_head);
2284 static int __perf_pending_run(void)
2286 struct perf_pending_entry *list;
2289 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2290 while (list != PENDING_TAIL) {
2291 void (*func)(struct perf_pending_entry *);
2292 struct perf_pending_entry *entry = list;
2299 * Ensure we observe the unqueue before we issue the wakeup,
2300 * so that we won't be waiting forever.
2301 * -- see perf_not_pending().
2312 static inline int perf_not_pending(struct perf_counter *counter)
2315 * If we flush on whatever cpu we run, there is a chance we don't
2319 __perf_pending_run();
2323 * Ensure we see the proper queue state before going to sleep
2324 * so that we do not miss the wakeup. -- see perf_pending_handle()
2327 return counter->pending.next == NULL;
2330 static void perf_pending_sync(struct perf_counter *counter)
2332 wait_event(counter->waitq, perf_not_pending(counter));
2335 void perf_counter_do_pending(void)
2337 __perf_pending_run();
2341 * Callchain support -- arch specific
2344 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2353 struct perf_output_handle {
2354 struct perf_counter *counter;
2355 struct perf_mmap_data *data;
2357 unsigned long offset;
2361 unsigned long flags;
2364 static bool perf_output_space(struct perf_mmap_data *data,
2365 unsigned int offset, unsigned int head)
2370 if (!data->writable)
2373 mask = (data->nr_pages << PAGE_SHIFT) - 1;
2375 * Userspace could choose to issue a mb() before updating the tail
2376 * pointer. So that all reads will be completed before the write is
2379 tail = ACCESS_ONCE(data->user_page->data_tail);
2382 offset = (offset - tail) & mask;
2383 head = (head - tail) & mask;
2385 if ((int)(head - offset) < 0)
2391 static void perf_output_wakeup(struct perf_output_handle *handle)
2393 atomic_set(&handle->data->poll, POLL_IN);
2396 handle->counter->pending_wakeup = 1;
2397 perf_pending_queue(&handle->counter->pending,
2398 perf_pending_counter);
2400 perf_counter_wakeup(handle->counter);
2404 * Curious locking construct.
2406 * We need to ensure a later event doesn't publish a head when a former
2407 * event isn't done writing. However since we need to deal with NMIs we
2408 * cannot fully serialize things.
2410 * What we do is serialize between CPUs so we only have to deal with NMI
2411 * nesting on a single CPU.
2413 * We only publish the head (and generate a wakeup) when the outer-most
2416 static void perf_output_lock(struct perf_output_handle *handle)
2418 struct perf_mmap_data *data = handle->data;
2423 local_irq_save(handle->flags);
2424 cpu = smp_processor_id();
2426 if (in_nmi() && atomic_read(&data->lock) == cpu)
2429 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2435 static void perf_output_unlock(struct perf_output_handle *handle)
2437 struct perf_mmap_data *data = handle->data;
2441 data->done_head = data->head;
2443 if (!handle->locked)
2448 * The xchg implies a full barrier that ensures all writes are done
2449 * before we publish the new head, matched by a rmb() in userspace when
2450 * reading this position.
2452 while ((head = atomic_long_xchg(&data->done_head, 0)))
2453 data->user_page->data_head = head;
2456 * NMI can happen here, which means we can miss a done_head update.
2459 cpu = atomic_xchg(&data->lock, -1);
2460 WARN_ON_ONCE(cpu != smp_processor_id());
2463 * Therefore we have to validate we did not indeed do so.
2465 if (unlikely(atomic_long_read(&data->done_head))) {
2467 * Since we had it locked, we can lock it again.
2469 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2475 if (atomic_xchg(&data->wakeup, 0))
2476 perf_output_wakeup(handle);
2478 local_irq_restore(handle->flags);
2481 static void perf_output_copy(struct perf_output_handle *handle,
2482 const void *buf, unsigned int len)
2484 unsigned int pages_mask;
2485 unsigned int offset;
2489 offset = handle->offset;
2490 pages_mask = handle->data->nr_pages - 1;
2491 pages = handle->data->data_pages;
2494 unsigned int page_offset;
2497 nr = (offset >> PAGE_SHIFT) & pages_mask;
2498 page_offset = offset & (PAGE_SIZE - 1);
2499 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2501 memcpy(pages[nr] + page_offset, buf, size);
2508 handle->offset = offset;
2511 * Check we didn't copy past our reservation window, taking the
2512 * possible unsigned int wrap into account.
2514 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2517 #define perf_output_put(handle, x) \
2518 perf_output_copy((handle), &(x), sizeof(x))
2520 static int perf_output_begin(struct perf_output_handle *handle,
2521 struct perf_counter *counter, unsigned int size,
2522 int nmi, int sample)
2524 struct perf_mmap_data *data;
2525 unsigned int offset, head;
2528 struct perf_event_header header;
2534 * For inherited counters we send all the output towards the parent.
2536 if (counter->parent)
2537 counter = counter->parent;
2540 data = rcu_dereference(counter->data);
2544 handle->data = data;
2545 handle->counter = counter;
2547 handle->sample = sample;
2549 if (!data->nr_pages)
2552 have_lost = atomic_read(&data->lost);
2554 size += sizeof(lost_event);
2556 perf_output_lock(handle);
2559 offset = head = atomic_long_read(&data->head);
2561 if (unlikely(!perf_output_space(data, offset, head)))
2563 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2565 handle->offset = offset;
2566 handle->head = head;
2568 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2569 atomic_set(&data->wakeup, 1);
2572 lost_event.header.type = PERF_EVENT_LOST;
2573 lost_event.header.misc = 0;
2574 lost_event.header.size = sizeof(lost_event);
2575 lost_event.id = counter->id;
2576 lost_event.lost = atomic_xchg(&data->lost, 0);
2578 perf_output_put(handle, lost_event);
2584 atomic_inc(&data->lost);
2585 perf_output_unlock(handle);
2592 static void perf_output_end(struct perf_output_handle *handle)
2594 struct perf_counter *counter = handle->counter;
2595 struct perf_mmap_data *data = handle->data;
2597 int wakeup_events = counter->attr.wakeup_events;
2599 if (handle->sample && wakeup_events) {
2600 int events = atomic_inc_return(&data->events);
2601 if (events >= wakeup_events) {
2602 atomic_sub(wakeup_events, &data->events);
2603 atomic_set(&data->wakeup, 1);
2607 perf_output_unlock(handle);
2611 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2614 * only top level counters have the pid namespace they were created in
2616 if (counter->parent)
2617 counter = counter->parent;
2619 return task_tgid_nr_ns(p, counter->ns);
2622 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2625 * only top level counters have the pid namespace they were created in
2627 if (counter->parent)
2628 counter = counter->parent;
2630 return task_pid_nr_ns(p, counter->ns);
2633 static void perf_counter_output(struct perf_counter *counter, int nmi,
2634 struct perf_sample_data *data)
2637 u64 sample_type = counter->attr.sample_type;
2638 struct perf_output_handle handle;
2639 struct perf_event_header header;
2648 struct perf_callchain_entry *callchain = NULL;
2649 struct perf_raw_record *raw = NULL;
2650 int callchain_size = 0;
2656 header.type = PERF_EVENT_SAMPLE;
2657 header.size = sizeof(header);
2660 header.misc |= perf_misc_flags(data->regs);
2662 if (sample_type & PERF_SAMPLE_IP) {
2663 ip = perf_instruction_pointer(data->regs);
2664 header.size += sizeof(ip);
2667 if (sample_type & PERF_SAMPLE_TID) {
2668 /* namespace issues */
2669 tid_entry.pid = perf_counter_pid(counter, current);
2670 tid_entry.tid = perf_counter_tid(counter, current);
2672 header.size += sizeof(tid_entry);
2675 if (sample_type & PERF_SAMPLE_TIME) {
2677 * Maybe do better on x86 and provide cpu_clock_nmi()
2679 time = sched_clock();
2681 header.size += sizeof(u64);
2684 if (sample_type & PERF_SAMPLE_ADDR)
2685 header.size += sizeof(u64);
2687 if (sample_type & PERF_SAMPLE_ID)
2688 header.size += sizeof(u64);
2690 if (sample_type & PERF_SAMPLE_STREAM_ID)
2691 header.size += sizeof(u64);
2693 if (sample_type & PERF_SAMPLE_CPU) {
2694 header.size += sizeof(cpu_entry);
2696 cpu_entry.cpu = raw_smp_processor_id();
2697 cpu_entry.reserved = 0;
2700 if (sample_type & PERF_SAMPLE_PERIOD)
2701 header.size += sizeof(u64);
2703 if (sample_type & PERF_SAMPLE_GROUP) {
2704 header.size += sizeof(u64) +
2705 counter->nr_siblings * sizeof(group_entry);
2708 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2709 callchain = perf_callchain(data->regs);
2712 callchain_size = (1 + callchain->nr) * sizeof(u64);
2713 header.size += callchain_size;
2715 header.size += sizeof(u64);
2718 if (sample_type & PERF_SAMPLE_RAW) {
2721 header.size += raw->size;
2724 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2728 perf_output_put(&handle, header);
2730 if (sample_type & PERF_SAMPLE_IP)
2731 perf_output_put(&handle, ip);
2733 if (sample_type & PERF_SAMPLE_TID)
2734 perf_output_put(&handle, tid_entry);
2736 if (sample_type & PERF_SAMPLE_TIME)
2737 perf_output_put(&handle, time);
2739 if (sample_type & PERF_SAMPLE_ADDR)
2740 perf_output_put(&handle, data->addr);
2742 if (sample_type & PERF_SAMPLE_ID) {
2743 u64 id = primary_counter_id(counter);
2745 perf_output_put(&handle, id);
2748 if (sample_type & PERF_SAMPLE_STREAM_ID)
2749 perf_output_put(&handle, counter->id);
2751 if (sample_type & PERF_SAMPLE_CPU)
2752 perf_output_put(&handle, cpu_entry);
2754 if (sample_type & PERF_SAMPLE_PERIOD)
2755 perf_output_put(&handle, data->period);
2758 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2760 if (sample_type & PERF_SAMPLE_GROUP) {
2761 struct perf_counter *leader, *sub;
2762 u64 nr = counter->nr_siblings;
2764 perf_output_put(&handle, nr);
2766 leader = counter->group_leader;
2767 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2769 sub->pmu->read(sub);
2771 group_entry.id = primary_counter_id(sub);
2772 group_entry.counter = atomic64_read(&sub->count);
2774 perf_output_put(&handle, group_entry);
2778 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2780 perf_output_copy(&handle, callchain, callchain_size);
2783 perf_output_put(&handle, nr);
2787 if ((sample_type & PERF_SAMPLE_RAW) && raw)
2788 perf_output_copy(&handle, raw->data, raw->size);
2790 perf_output_end(&handle);
2797 struct perf_read_event {
2798 struct perf_event_header header;
2807 perf_counter_read_event(struct perf_counter *counter,
2808 struct task_struct *task)
2810 struct perf_output_handle handle;
2811 struct perf_read_event event = {
2813 .type = PERF_EVENT_READ,
2815 .size = sizeof(event) - sizeof(event.format),
2817 .pid = perf_counter_pid(counter, task),
2818 .tid = perf_counter_tid(counter, task),
2819 .value = atomic64_read(&counter->count),
2823 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2824 event.header.size += sizeof(u64);
2825 event.format[i++] = counter->total_time_enabled;
2828 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2829 event.header.size += sizeof(u64);
2830 event.format[i++] = counter->total_time_running;
2833 if (counter->attr.read_format & PERF_FORMAT_ID) {
2834 event.header.size += sizeof(u64);
2835 event.format[i++] = primary_counter_id(counter);
2838 ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
2842 perf_output_copy(&handle, &event, event.header.size);
2843 perf_output_end(&handle);
2847 * task tracking -- fork/exit
2849 * enabled by: attr.comm | attr.mmap | attr.task
2852 struct perf_task_event {
2853 struct task_struct *task;
2856 struct perf_event_header header;
2865 static void perf_counter_task_output(struct perf_counter *counter,
2866 struct perf_task_event *task_event)
2868 struct perf_output_handle handle;
2869 int size = task_event->event.header.size;
2870 struct task_struct *task = task_event->task;
2871 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2876 task_event->event.pid = perf_counter_pid(counter, task);
2877 task_event->event.ppid = perf_counter_pid(counter, task->real_parent);
2879 task_event->event.tid = perf_counter_tid(counter, task);
2880 task_event->event.ptid = perf_counter_tid(counter, task->real_parent);
2882 perf_output_put(&handle, task_event->event);
2883 perf_output_end(&handle);
2886 static int perf_counter_task_match(struct perf_counter *counter)
2888 if (counter->attr.comm || counter->attr.mmap || counter->attr.task)
2894 static void perf_counter_task_ctx(struct perf_counter_context *ctx,
2895 struct perf_task_event *task_event)
2897 struct perf_counter *counter;
2899 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2903 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2904 if (perf_counter_task_match(counter))
2905 perf_counter_task_output(counter, task_event);
2910 static void perf_counter_task_event(struct perf_task_event *task_event)
2912 struct perf_cpu_context *cpuctx;
2913 struct perf_counter_context *ctx;
2915 cpuctx = &get_cpu_var(perf_cpu_context);
2916 perf_counter_task_ctx(&cpuctx->ctx, task_event);
2917 put_cpu_var(perf_cpu_context);
2921 * doesn't really matter which of the child contexts the
2922 * events ends up in.
2924 ctx = rcu_dereference(current->perf_counter_ctxp);
2926 perf_counter_task_ctx(ctx, task_event);
2930 static void perf_counter_task(struct task_struct *task, int new)
2932 struct perf_task_event task_event;
2934 if (!atomic_read(&nr_comm_counters) &&
2935 !atomic_read(&nr_mmap_counters) &&
2936 !atomic_read(&nr_task_counters))
2939 task_event = (struct perf_task_event){
2943 .type = new ? PERF_EVENT_FORK : PERF_EVENT_EXIT,
2945 .size = sizeof(task_event.event),
2954 perf_counter_task_event(&task_event);
2957 void perf_counter_fork(struct task_struct *task)
2959 perf_counter_task(task, 1);
2966 struct perf_comm_event {
2967 struct task_struct *task;
2972 struct perf_event_header header;
2979 static void perf_counter_comm_output(struct perf_counter *counter,
2980 struct perf_comm_event *comm_event)
2982 struct perf_output_handle handle;
2983 int size = comm_event->event.header.size;
2984 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2989 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
2990 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
2992 perf_output_put(&handle, comm_event->event);
2993 perf_output_copy(&handle, comm_event->comm,
2994 comm_event->comm_size);
2995 perf_output_end(&handle);
2998 static int perf_counter_comm_match(struct perf_counter *counter)
3000 if (counter->attr.comm)
3006 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
3007 struct perf_comm_event *comm_event)
3009 struct perf_counter *counter;
3011 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3015 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3016 if (perf_counter_comm_match(counter))
3017 perf_counter_comm_output(counter, comm_event);
3022 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
3024 struct perf_cpu_context *cpuctx;
3025 struct perf_counter_context *ctx;
3027 char comm[TASK_COMM_LEN];
3029 memset(comm, 0, sizeof(comm));
3030 strncpy(comm, comm_event->task->comm, sizeof(comm));
3031 size = ALIGN(strlen(comm)+1, sizeof(u64));
3033 comm_event->comm = comm;
3034 comm_event->comm_size = size;
3036 comm_event->event.header.size = sizeof(comm_event->event) + size;
3038 cpuctx = &get_cpu_var(perf_cpu_context);
3039 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
3040 put_cpu_var(perf_cpu_context);
3044 * doesn't really matter which of the child contexts the
3045 * events ends up in.
3047 ctx = rcu_dereference(current->perf_counter_ctxp);
3049 perf_counter_comm_ctx(ctx, comm_event);
3053 void perf_counter_comm(struct task_struct *task)
3055 struct perf_comm_event comm_event;
3057 if (task->perf_counter_ctxp)
3058 perf_counter_enable_on_exec(task);
3060 if (!atomic_read(&nr_comm_counters))
3063 comm_event = (struct perf_comm_event){
3069 .type = PERF_EVENT_COMM,
3078 perf_counter_comm_event(&comm_event);
3085 struct perf_mmap_event {
3086 struct vm_area_struct *vma;
3088 const char *file_name;
3092 struct perf_event_header header;
3102 static void perf_counter_mmap_output(struct perf_counter *counter,
3103 struct perf_mmap_event *mmap_event)
3105 struct perf_output_handle handle;
3106 int size = mmap_event->event.header.size;
3107 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3112 mmap_event->event.pid = perf_counter_pid(counter, current);
3113 mmap_event->event.tid = perf_counter_tid(counter, current);
3115 perf_output_put(&handle, mmap_event->event);
3116 perf_output_copy(&handle, mmap_event->file_name,
3117 mmap_event->file_size);
3118 perf_output_end(&handle);
3121 static int perf_counter_mmap_match(struct perf_counter *counter,
3122 struct perf_mmap_event *mmap_event)
3124 if (counter->attr.mmap)
3130 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3131 struct perf_mmap_event *mmap_event)
3133 struct perf_counter *counter;
3135 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3139 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3140 if (perf_counter_mmap_match(counter, mmap_event))
3141 perf_counter_mmap_output(counter, mmap_event);
3146 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3148 struct perf_cpu_context *cpuctx;
3149 struct perf_counter_context *ctx;
3150 struct vm_area_struct *vma = mmap_event->vma;
3151 struct file *file = vma->vm_file;
3157 memset(tmp, 0, sizeof(tmp));
3161 * d_path works from the end of the buffer backwards, so we
3162 * need to add enough zero bytes after the string to handle
3163 * the 64bit alignment we do later.
3165 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3167 name = strncpy(tmp, "//enomem", sizeof(tmp));
3170 name = d_path(&file->f_path, buf, PATH_MAX);
3172 name = strncpy(tmp, "//toolong", sizeof(tmp));
3176 if (arch_vma_name(mmap_event->vma)) {
3177 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3183 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3187 name = strncpy(tmp, "//anon", sizeof(tmp));
3192 size = ALIGN(strlen(name)+1, sizeof(u64));
3194 mmap_event->file_name = name;
3195 mmap_event->file_size = size;
3197 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3199 cpuctx = &get_cpu_var(perf_cpu_context);
3200 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3201 put_cpu_var(perf_cpu_context);
3205 * doesn't really matter which of the child contexts the
3206 * events ends up in.
3208 ctx = rcu_dereference(current->perf_counter_ctxp);
3210 perf_counter_mmap_ctx(ctx, mmap_event);
3216 void __perf_counter_mmap(struct vm_area_struct *vma)
3218 struct perf_mmap_event mmap_event;
3220 if (!atomic_read(&nr_mmap_counters))
3223 mmap_event = (struct perf_mmap_event){
3229 .type = PERF_EVENT_MMAP,
3235 .start = vma->vm_start,
3236 .len = vma->vm_end - vma->vm_start,
3237 .pgoff = vma->vm_pgoff,
3241 perf_counter_mmap_event(&mmap_event);
3245 * IRQ throttle logging
3248 static void perf_log_throttle(struct perf_counter *counter, int enable)
3250 struct perf_output_handle handle;
3254 struct perf_event_header header;
3258 } throttle_event = {
3260 .type = PERF_EVENT_THROTTLE,
3262 .size = sizeof(throttle_event),
3264 .time = sched_clock(),
3265 .id = primary_counter_id(counter),
3266 .stream_id = counter->id,
3270 throttle_event.header.type = PERF_EVENT_UNTHROTTLE;
3272 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3276 perf_output_put(&handle, throttle_event);
3277 perf_output_end(&handle);
3281 * Generic counter overflow handling, sampling.
3284 int perf_counter_overflow(struct perf_counter *counter, int nmi,
3285 struct perf_sample_data *data)
3287 int events = atomic_read(&counter->event_limit);
3288 int throttle = counter->pmu->unthrottle != NULL;
3289 struct hw_perf_counter *hwc = &counter->hw;
3295 if (hwc->interrupts != MAX_INTERRUPTS) {
3297 if (HZ * hwc->interrupts >
3298 (u64)sysctl_perf_counter_sample_rate) {
3299 hwc->interrupts = MAX_INTERRUPTS;
3300 perf_log_throttle(counter, 0);
3305 * Keep re-disabling counters even though on the previous
3306 * pass we disabled it - just in case we raced with a
3307 * sched-in and the counter got enabled again:
3313 if (counter->attr.freq) {
3314 u64 now = sched_clock();
3315 s64 delta = now - hwc->freq_stamp;
3317 hwc->freq_stamp = now;
3319 if (delta > 0 && delta < TICK_NSEC)
3320 perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3324 * XXX event_limit might not quite work as expected on inherited
3328 counter->pending_kill = POLL_IN;
3329 if (events && atomic_dec_and_test(&counter->event_limit)) {
3331 counter->pending_kill = POLL_HUP;
3333 counter->pending_disable = 1;
3334 perf_pending_queue(&counter->pending,
3335 perf_pending_counter);
3337 perf_counter_disable(counter);
3340 perf_counter_output(counter, nmi, data);
3345 * Generic software counter infrastructure
3349 * We directly increment counter->count and keep a second value in
3350 * counter->hw.period_left to count intervals. This period counter
3351 * is kept in the range [-sample_period, 0] so that we can use the
3355 static u64 perf_swcounter_set_period(struct perf_counter *counter)
3357 struct hw_perf_counter *hwc = &counter->hw;
3358 u64 period = hwc->last_period;
3362 hwc->last_period = hwc->sample_period;
3365 old = val = atomic64_read(&hwc->period_left);
3369 nr = div64_u64(period + val, period);
3370 offset = nr * period;
3372 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3378 static void perf_swcounter_overflow(struct perf_counter *counter,
3379 int nmi, struct perf_sample_data *data)
3381 struct hw_perf_counter *hwc = &counter->hw;
3384 data->period = counter->hw.last_period;
3385 overflow = perf_swcounter_set_period(counter);
3387 if (hwc->interrupts == MAX_INTERRUPTS)
3390 for (; overflow; overflow--) {
3391 if (perf_counter_overflow(counter, nmi, data)) {
3393 * We inhibit the overflow from happening when
3394 * hwc->interrupts == MAX_INTERRUPTS.
3401 static void perf_swcounter_unthrottle(struct perf_counter *counter)
3404 * Nothing to do, we already reset hwc->interrupts.
3408 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3409 int nmi, struct perf_sample_data *data)
3411 struct hw_perf_counter *hwc = &counter->hw;
3413 atomic64_add(nr, &counter->count);
3415 if (!hwc->sample_period)
3421 if (!atomic64_add_negative(nr, &hwc->period_left))
3422 perf_swcounter_overflow(counter, nmi, data);
3425 static int perf_swcounter_is_counting(struct perf_counter *counter)
3427 struct perf_counter_context *ctx;
3428 unsigned long flags;
3431 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3434 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3438 * If the counter is inactive, it could be just because
3439 * its task is scheduled out, or because it's in a group
3440 * which could not go on the PMU. We want to count in
3441 * the first case but not the second. If the context is
3442 * currently active then an inactive software counter must
3443 * be the second case. If it's not currently active then
3444 * we need to know whether the counter was active when the
3445 * context was last active, which we can determine by
3446 * comparing counter->tstamp_stopped with ctx->time.
3448 * We are within an RCU read-side critical section,
3449 * which protects the existence of *ctx.
3452 spin_lock_irqsave(&ctx->lock, flags);
3454 /* Re-check state now we have the lock */
3455 if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
3456 counter->ctx->is_active ||
3457 counter->tstamp_stopped < ctx->time)
3459 spin_unlock_irqrestore(&ctx->lock, flags);
3463 static int perf_swcounter_match(struct perf_counter *counter,
3464 enum perf_type_id type,
3465 u32 event, struct pt_regs *regs)
3467 if (!perf_swcounter_is_counting(counter))
3470 if (counter->attr.type != type)
3472 if (counter->attr.config != event)
3476 if (counter->attr.exclude_user && user_mode(regs))
3479 if (counter->attr.exclude_kernel && !user_mode(regs))
3486 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3487 enum perf_type_id type,
3488 u32 event, u64 nr, int nmi,
3489 struct perf_sample_data *data)
3491 struct perf_counter *counter;
3493 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3497 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3498 if (perf_swcounter_match(counter, type, event, data->regs))
3499 perf_swcounter_add(counter, nr, nmi, data);
3504 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3507 return &cpuctx->recursion[3];
3510 return &cpuctx->recursion[2];
3513 return &cpuctx->recursion[1];
3515 return &cpuctx->recursion[0];
3518 static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3520 struct perf_sample_data *data)
3522 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3523 int *recursion = perf_swcounter_recursion_context(cpuctx);
3524 struct perf_counter_context *ctx;
3532 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3536 * doesn't really matter which of the child contexts the
3537 * events ends up in.
3539 ctx = rcu_dereference(current->perf_counter_ctxp);
3541 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3548 put_cpu_var(perf_cpu_context);
3551 void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3552 struct pt_regs *regs, u64 addr)
3554 struct perf_sample_data data = {
3559 do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3562 static void perf_swcounter_read(struct perf_counter *counter)
3566 static int perf_swcounter_enable(struct perf_counter *counter)
3568 struct hw_perf_counter *hwc = &counter->hw;
3570 if (hwc->sample_period) {
3571 hwc->last_period = hwc->sample_period;
3572 perf_swcounter_set_period(counter);
3577 static void perf_swcounter_disable(struct perf_counter *counter)
3581 static const struct pmu perf_ops_generic = {
3582 .enable = perf_swcounter_enable,
3583 .disable = perf_swcounter_disable,
3584 .read = perf_swcounter_read,
3585 .unthrottle = perf_swcounter_unthrottle,
3589 * hrtimer based swcounter callback
3592 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3594 enum hrtimer_restart ret = HRTIMER_RESTART;
3595 struct perf_sample_data data;
3596 struct perf_counter *counter;
3599 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3600 counter->pmu->read(counter);
3603 data.regs = get_irq_regs();
3605 * In case we exclude kernel IPs or are somehow not in interrupt
3606 * context, provide the next best thing, the user IP.
3608 if ((counter->attr.exclude_kernel || !data.regs) &&
3609 !counter->attr.exclude_user)
3610 data.regs = task_pt_regs(current);
3613 if (perf_counter_overflow(counter, 0, &data))
3614 ret = HRTIMER_NORESTART;
3617 period = max_t(u64, 10000, counter->hw.sample_period);
3618 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3624 * Software counter: cpu wall time clock
3627 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3629 int cpu = raw_smp_processor_id();
3633 now = cpu_clock(cpu);
3634 prev = atomic64_read(&counter->hw.prev_count);
3635 atomic64_set(&counter->hw.prev_count, now);
3636 atomic64_add(now - prev, &counter->count);
3639 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3641 struct hw_perf_counter *hwc = &counter->hw;
3642 int cpu = raw_smp_processor_id();
3644 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3645 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3646 hwc->hrtimer.function = perf_swcounter_hrtimer;
3647 if (hwc->sample_period) {
3648 u64 period = max_t(u64, 10000, hwc->sample_period);
3649 __hrtimer_start_range_ns(&hwc->hrtimer,
3650 ns_to_ktime(period), 0,
3651 HRTIMER_MODE_REL, 0);
3657 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3659 if (counter->hw.sample_period)
3660 hrtimer_cancel(&counter->hw.hrtimer);
3661 cpu_clock_perf_counter_update(counter);
3664 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3666 cpu_clock_perf_counter_update(counter);
3669 static const struct pmu perf_ops_cpu_clock = {
3670 .enable = cpu_clock_perf_counter_enable,
3671 .disable = cpu_clock_perf_counter_disable,
3672 .read = cpu_clock_perf_counter_read,
3676 * Software counter: task time clock
3679 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3684 prev = atomic64_xchg(&counter->hw.prev_count, now);
3686 atomic64_add(delta, &counter->count);
3689 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3691 struct hw_perf_counter *hwc = &counter->hw;
3694 now = counter->ctx->time;
3696 atomic64_set(&hwc->prev_count, now);
3697 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3698 hwc->hrtimer.function = perf_swcounter_hrtimer;
3699 if (hwc->sample_period) {
3700 u64 period = max_t(u64, 10000, hwc->sample_period);
3701 __hrtimer_start_range_ns(&hwc->hrtimer,
3702 ns_to_ktime(period), 0,
3703 HRTIMER_MODE_REL, 0);
3709 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3711 if (counter->hw.sample_period)
3712 hrtimer_cancel(&counter->hw.hrtimer);
3713 task_clock_perf_counter_update(counter, counter->ctx->time);
3717 static void task_clock_perf_counter_read(struct perf_counter *counter)
3722 update_context_time(counter->ctx);
3723 time = counter->ctx->time;
3725 u64 now = perf_clock();
3726 u64 delta = now - counter->ctx->timestamp;
3727 time = counter->ctx->time + delta;
3730 task_clock_perf_counter_update(counter, time);
3733 static const struct pmu perf_ops_task_clock = {
3734 .enable = task_clock_perf_counter_enable,
3735 .disable = task_clock_perf_counter_disable,
3736 .read = task_clock_perf_counter_read,
3739 #ifdef CONFIG_EVENT_PROFILE
3740 void perf_tpcounter_event(int event_id, u64 addr, u64 count, void *record,
3743 struct perf_raw_record raw = {
3748 struct perf_sample_data data = {
3749 .regs = get_irq_regs(),
3755 data.regs = task_pt_regs(current);
3757 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, count, 1, &data);
3759 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3761 extern int ftrace_profile_enable(int);
3762 extern void ftrace_profile_disable(int);
3764 static void tp_perf_counter_destroy(struct perf_counter *counter)
3766 ftrace_profile_disable(counter->attr.config);
3769 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3771 if (ftrace_profile_enable(counter->attr.config))
3774 counter->destroy = tp_perf_counter_destroy;
3776 return &perf_ops_generic;
3779 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3785 atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3787 static void sw_perf_counter_destroy(struct perf_counter *counter)
3789 u64 event = counter->attr.config;
3791 WARN_ON(counter->parent);
3793 atomic_dec(&perf_swcounter_enabled[event]);
3796 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3798 const struct pmu *pmu = NULL;
3799 u64 event = counter->attr.config;
3802 * Software counters (currently) can't in general distinguish
3803 * between user, kernel and hypervisor events.
3804 * However, context switches and cpu migrations are considered
3805 * to be kernel events, and page faults are never hypervisor
3809 case PERF_COUNT_SW_CPU_CLOCK:
3810 pmu = &perf_ops_cpu_clock;
3813 case PERF_COUNT_SW_TASK_CLOCK:
3815 * If the user instantiates this as a per-cpu counter,
3816 * use the cpu_clock counter instead.
3818 if (counter->ctx->task)
3819 pmu = &perf_ops_task_clock;
3821 pmu = &perf_ops_cpu_clock;
3824 case PERF_COUNT_SW_PAGE_FAULTS:
3825 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
3826 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
3827 case PERF_COUNT_SW_CONTEXT_SWITCHES:
3828 case PERF_COUNT_SW_CPU_MIGRATIONS:
3829 if (!counter->parent) {
3830 atomic_inc(&perf_swcounter_enabled[event]);
3831 counter->destroy = sw_perf_counter_destroy;
3833 pmu = &perf_ops_generic;
3841 * Allocate and initialize a counter structure
3843 static struct perf_counter *
3844 perf_counter_alloc(struct perf_counter_attr *attr,
3846 struct perf_counter_context *ctx,
3847 struct perf_counter *group_leader,
3848 struct perf_counter *parent_counter,
3851 const struct pmu *pmu;
3852 struct perf_counter *counter;
3853 struct hw_perf_counter *hwc;
3856 counter = kzalloc(sizeof(*counter), gfpflags);
3858 return ERR_PTR(-ENOMEM);
3861 * Single counters are their own group leaders, with an
3862 * empty sibling list:
3865 group_leader = counter;
3867 mutex_init(&counter->child_mutex);
3868 INIT_LIST_HEAD(&counter->child_list);
3870 INIT_LIST_HEAD(&counter->list_entry);
3871 INIT_LIST_HEAD(&counter->event_entry);
3872 INIT_LIST_HEAD(&counter->sibling_list);
3873 init_waitqueue_head(&counter->waitq);
3875 mutex_init(&counter->mmap_mutex);
3878 counter->attr = *attr;
3879 counter->group_leader = group_leader;
3880 counter->pmu = NULL;
3882 counter->oncpu = -1;
3884 counter->parent = parent_counter;
3886 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
3887 counter->id = atomic64_inc_return(&perf_counter_id);
3889 counter->state = PERF_COUNTER_STATE_INACTIVE;
3892 counter->state = PERF_COUNTER_STATE_OFF;
3897 hwc->sample_period = attr->sample_period;
3898 if (attr->freq && attr->sample_freq)
3899 hwc->sample_period = 1;
3901 atomic64_set(&hwc->period_left, hwc->sample_period);
3904 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3906 if (attr->inherit && (attr->sample_type & PERF_SAMPLE_GROUP))
3909 switch (attr->type) {
3911 case PERF_TYPE_HARDWARE:
3912 case PERF_TYPE_HW_CACHE:
3913 pmu = hw_perf_counter_init(counter);
3916 case PERF_TYPE_SOFTWARE:
3917 pmu = sw_perf_counter_init(counter);
3920 case PERF_TYPE_TRACEPOINT:
3921 pmu = tp_perf_counter_init(counter);
3931 else if (IS_ERR(pmu))
3936 put_pid_ns(counter->ns);
3938 return ERR_PTR(err);
3943 if (!counter->parent) {
3944 atomic_inc(&nr_counters);
3945 if (counter->attr.mmap)
3946 atomic_inc(&nr_mmap_counters);
3947 if (counter->attr.comm)
3948 atomic_inc(&nr_comm_counters);
3949 if (counter->attr.task)
3950 atomic_inc(&nr_task_counters);
3956 static int perf_copy_attr(struct perf_counter_attr __user *uattr,
3957 struct perf_counter_attr *attr)
3962 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
3966 * zero the full structure, so that a short copy will be nice.
3968 memset(attr, 0, sizeof(*attr));
3970 ret = get_user(size, &uattr->size);
3974 if (size > PAGE_SIZE) /* silly large */
3977 if (!size) /* abi compat */
3978 size = PERF_ATTR_SIZE_VER0;
3980 if (size < PERF_ATTR_SIZE_VER0)
3984 * If we're handed a bigger struct than we know of,
3985 * ensure all the unknown bits are 0.
3987 if (size > sizeof(*attr)) {
3989 unsigned long __user *addr;
3990 unsigned long __user *end;
3992 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
3993 sizeof(unsigned long));
3994 end = PTR_ALIGN((void __user *)uattr + size,
3995 sizeof(unsigned long));
3997 for (; addr < end; addr += sizeof(unsigned long)) {
3998 ret = get_user(val, addr);
4006 ret = copy_from_user(attr, uattr, size);
4011 * If the type exists, the corresponding creation will verify
4014 if (attr->type >= PERF_TYPE_MAX)
4017 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4020 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4023 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4030 put_user(sizeof(*attr), &uattr->size);
4036 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
4038 * @attr_uptr: event type attributes for monitoring/sampling
4041 * @group_fd: group leader counter fd
4043 SYSCALL_DEFINE5(perf_counter_open,
4044 struct perf_counter_attr __user *, attr_uptr,
4045 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4047 struct perf_counter *counter, *group_leader;
4048 struct perf_counter_attr attr;
4049 struct perf_counter_context *ctx;
4050 struct file *counter_file = NULL;
4051 struct file *group_file = NULL;
4052 int fput_needed = 0;
4053 int fput_needed2 = 0;
4056 /* for future expandability... */
4060 ret = perf_copy_attr(attr_uptr, &attr);
4064 if (!attr.exclude_kernel) {
4065 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4070 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
4075 * Get the target context (task or percpu):
4077 ctx = find_get_context(pid, cpu);
4079 return PTR_ERR(ctx);
4082 * Look up the group leader (we will attach this counter to it):
4084 group_leader = NULL;
4085 if (group_fd != -1) {
4087 group_file = fget_light(group_fd, &fput_needed);
4089 goto err_put_context;
4090 if (group_file->f_op != &perf_fops)
4091 goto err_put_context;
4093 group_leader = group_file->private_data;
4095 * Do not allow a recursive hierarchy (this new sibling
4096 * becoming part of another group-sibling):
4098 if (group_leader->group_leader != group_leader)
4099 goto err_put_context;
4101 * Do not allow to attach to a group in a different
4102 * task or CPU context:
4104 if (group_leader->ctx != ctx)
4105 goto err_put_context;
4107 * Only a group leader can be exclusive or pinned
4109 if (attr.exclusive || attr.pinned)
4110 goto err_put_context;
4113 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
4115 ret = PTR_ERR(counter);
4116 if (IS_ERR(counter))
4117 goto err_put_context;
4119 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
4121 goto err_free_put_context;
4123 counter_file = fget_light(ret, &fput_needed2);
4125 goto err_free_put_context;
4127 counter->filp = counter_file;
4128 WARN_ON_ONCE(ctx->parent_ctx);
4129 mutex_lock(&ctx->mutex);
4130 perf_install_in_context(ctx, counter, cpu);
4132 mutex_unlock(&ctx->mutex);
4134 counter->owner = current;
4135 get_task_struct(current);
4136 mutex_lock(¤t->perf_counter_mutex);
4137 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
4138 mutex_unlock(¤t->perf_counter_mutex);
4140 fput_light(counter_file, fput_needed2);
4143 fput_light(group_file, fput_needed);
4147 err_free_put_context:
4157 * inherit a counter from parent task to child task:
4159 static struct perf_counter *
4160 inherit_counter(struct perf_counter *parent_counter,
4161 struct task_struct *parent,
4162 struct perf_counter_context *parent_ctx,
4163 struct task_struct *child,
4164 struct perf_counter *group_leader,
4165 struct perf_counter_context *child_ctx)
4167 struct perf_counter *child_counter;
4170 * Instead of creating recursive hierarchies of counters,
4171 * we link inherited counters back to the original parent,
4172 * which has a filp for sure, which we use as the reference
4175 if (parent_counter->parent)
4176 parent_counter = parent_counter->parent;
4178 child_counter = perf_counter_alloc(&parent_counter->attr,
4179 parent_counter->cpu, child_ctx,
4180 group_leader, parent_counter,
4182 if (IS_ERR(child_counter))
4183 return child_counter;
4187 * Make the child state follow the state of the parent counter,
4188 * not its attr.disabled bit. We hold the parent's mutex,
4189 * so we won't race with perf_counter_{en, dis}able_family.
4191 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4192 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4194 child_counter->state = PERF_COUNTER_STATE_OFF;
4196 if (parent_counter->attr.freq)
4197 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4200 * Link it up in the child's context:
4202 add_counter_to_ctx(child_counter, child_ctx);
4205 * Get a reference to the parent filp - we will fput it
4206 * when the child counter exits. This is safe to do because
4207 * we are in the parent and we know that the filp still
4208 * exists and has a nonzero count:
4210 atomic_long_inc(&parent_counter->filp->f_count);
4213 * Link this into the parent counter's child list
4215 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4216 mutex_lock(&parent_counter->child_mutex);
4217 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4218 mutex_unlock(&parent_counter->child_mutex);
4220 return child_counter;
4223 static int inherit_group(struct perf_counter *parent_counter,
4224 struct task_struct *parent,
4225 struct perf_counter_context *parent_ctx,
4226 struct task_struct *child,
4227 struct perf_counter_context *child_ctx)
4229 struct perf_counter *leader;
4230 struct perf_counter *sub;
4231 struct perf_counter *child_ctr;
4233 leader = inherit_counter(parent_counter, parent, parent_ctx,
4234 child, NULL, child_ctx);
4236 return PTR_ERR(leader);
4237 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4238 child_ctr = inherit_counter(sub, parent, parent_ctx,
4239 child, leader, child_ctx);
4240 if (IS_ERR(child_ctr))
4241 return PTR_ERR(child_ctr);
4246 static void sync_child_counter(struct perf_counter *child_counter,
4247 struct task_struct *child)
4249 struct perf_counter *parent_counter = child_counter->parent;
4252 if (child_counter->attr.inherit_stat)
4253 perf_counter_read_event(child_counter, child);
4255 child_val = atomic64_read(&child_counter->count);
4258 * Add back the child's count to the parent's count:
4260 atomic64_add(child_val, &parent_counter->count);
4261 atomic64_add(child_counter->total_time_enabled,
4262 &parent_counter->child_total_time_enabled);
4263 atomic64_add(child_counter->total_time_running,
4264 &parent_counter->child_total_time_running);
4267 * Remove this counter from the parent's list
4269 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4270 mutex_lock(&parent_counter->child_mutex);
4271 list_del_init(&child_counter->child_list);
4272 mutex_unlock(&parent_counter->child_mutex);
4275 * Release the parent counter, if this was the last
4278 fput(parent_counter->filp);
4282 __perf_counter_exit_task(struct perf_counter *child_counter,
4283 struct perf_counter_context *child_ctx,
4284 struct task_struct *child)
4286 struct perf_counter *parent_counter;
4288 update_counter_times(child_counter);
4289 perf_counter_remove_from_context(child_counter);
4291 parent_counter = child_counter->parent;
4293 * It can happen that parent exits first, and has counters
4294 * that are still around due to the child reference. These
4295 * counters need to be zapped - but otherwise linger.
4297 if (parent_counter) {
4298 sync_child_counter(child_counter, child);
4299 free_counter(child_counter);
4304 * When a child task exits, feed back counter values to parent counters.
4306 void perf_counter_exit_task(struct task_struct *child)
4308 struct perf_counter *child_counter, *tmp;
4309 struct perf_counter_context *child_ctx;
4310 unsigned long flags;
4312 if (likely(!child->perf_counter_ctxp)) {
4313 perf_counter_task(child, 0);
4317 local_irq_save(flags);
4319 * We can't reschedule here because interrupts are disabled,
4320 * and either child is current or it is a task that can't be
4321 * scheduled, so we are now safe from rescheduling changing
4324 child_ctx = child->perf_counter_ctxp;
4325 __perf_counter_task_sched_out(child_ctx);
4328 * Take the context lock here so that if find_get_context is
4329 * reading child->perf_counter_ctxp, we wait until it has
4330 * incremented the context's refcount before we do put_ctx below.
4332 spin_lock(&child_ctx->lock);
4334 * If this context is a clone; unclone it so it can't get
4335 * swapped to another process while we're removing all
4336 * the counters from it.
4338 unclone_ctx(child_ctx);
4339 spin_unlock_irqrestore(&child_ctx->lock, flags);
4342 * Report the task dead after unscheduling the counters so that we
4343 * won't get any samples after PERF_EVENT_EXIT. We can however still
4344 * get a few PERF_EVENT_READ events.
4346 perf_counter_task(child, 0);
4348 child->perf_counter_ctxp = NULL;
4351 * We can recurse on the same lock type through:
4353 * __perf_counter_exit_task()
4354 * sync_child_counter()
4355 * fput(parent_counter->filp)
4357 * mutex_lock(&ctx->mutex)
4359 * But since its the parent context it won't be the same instance.
4361 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4364 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4366 __perf_counter_exit_task(child_counter, child_ctx, child);
4369 * If the last counter was a group counter, it will have appended all
4370 * its siblings to the list, but we obtained 'tmp' before that which
4371 * will still point to the list head terminating the iteration.
4373 if (!list_empty(&child_ctx->counter_list))
4376 mutex_unlock(&child_ctx->mutex);
4382 * free an unexposed, unused context as created by inheritance by
4383 * init_task below, used by fork() in case of fail.
4385 void perf_counter_free_task(struct task_struct *task)
4387 struct perf_counter_context *ctx = task->perf_counter_ctxp;
4388 struct perf_counter *counter, *tmp;
4393 mutex_lock(&ctx->mutex);
4395 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4396 struct perf_counter *parent = counter->parent;
4398 if (WARN_ON_ONCE(!parent))
4401 mutex_lock(&parent->child_mutex);
4402 list_del_init(&counter->child_list);
4403 mutex_unlock(&parent->child_mutex);
4407 list_del_counter(counter, ctx);
4408 free_counter(counter);
4411 if (!list_empty(&ctx->counter_list))
4414 mutex_unlock(&ctx->mutex);
4420 * Initialize the perf_counter context in task_struct
4422 int perf_counter_init_task(struct task_struct *child)
4424 struct perf_counter_context *child_ctx, *parent_ctx;
4425 struct perf_counter_context *cloned_ctx;
4426 struct perf_counter *counter;
4427 struct task_struct *parent = current;
4428 int inherited_all = 1;
4431 child->perf_counter_ctxp = NULL;
4433 mutex_init(&child->perf_counter_mutex);
4434 INIT_LIST_HEAD(&child->perf_counter_list);
4436 if (likely(!parent->perf_counter_ctxp))
4440 * This is executed from the parent task context, so inherit
4441 * counters that have been marked for cloning.
4442 * First allocate and initialize a context for the child.
4445 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4449 __perf_counter_init_context(child_ctx, child);
4450 child->perf_counter_ctxp = child_ctx;
4451 get_task_struct(child);
4454 * If the parent's context is a clone, pin it so it won't get
4457 parent_ctx = perf_pin_task_context(parent);
4460 * No need to check if parent_ctx != NULL here; since we saw
4461 * it non-NULL earlier, the only reason for it to become NULL
4462 * is if we exit, and since we're currently in the middle of
4463 * a fork we can't be exiting at the same time.
4467 * Lock the parent list. No need to lock the child - not PID
4468 * hashed yet and not running, so nobody can access it.
4470 mutex_lock(&parent_ctx->mutex);
4473 * We dont have to disable NMIs - we are only looking at
4474 * the list, not manipulating it:
4476 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4477 if (counter != counter->group_leader)
4480 if (!counter->attr.inherit) {
4485 ret = inherit_group(counter, parent, parent_ctx,
4493 if (inherited_all) {
4495 * Mark the child context as a clone of the parent
4496 * context, or of whatever the parent is a clone of.
4497 * Note that if the parent is a clone, it could get
4498 * uncloned at any point, but that doesn't matter
4499 * because the list of counters and the generation
4500 * count can't have changed since we took the mutex.
4502 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4504 child_ctx->parent_ctx = cloned_ctx;
4505 child_ctx->parent_gen = parent_ctx->parent_gen;
4507 child_ctx->parent_ctx = parent_ctx;
4508 child_ctx->parent_gen = parent_ctx->generation;
4510 get_ctx(child_ctx->parent_ctx);
4513 mutex_unlock(&parent_ctx->mutex);
4515 perf_unpin_context(parent_ctx);
4520 static void __cpuinit perf_counter_init_cpu(int cpu)
4522 struct perf_cpu_context *cpuctx;
4524 cpuctx = &per_cpu(perf_cpu_context, cpu);
4525 __perf_counter_init_context(&cpuctx->ctx, NULL);
4527 spin_lock(&perf_resource_lock);
4528 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4529 spin_unlock(&perf_resource_lock);
4531 hw_perf_counter_setup(cpu);
4534 #ifdef CONFIG_HOTPLUG_CPU
4535 static void __perf_counter_exit_cpu(void *info)
4537 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4538 struct perf_counter_context *ctx = &cpuctx->ctx;
4539 struct perf_counter *counter, *tmp;
4541 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4542 __perf_counter_remove_from_context(counter);
4544 static void perf_counter_exit_cpu(int cpu)
4546 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4547 struct perf_counter_context *ctx = &cpuctx->ctx;
4549 mutex_lock(&ctx->mutex);
4550 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4551 mutex_unlock(&ctx->mutex);
4554 static inline void perf_counter_exit_cpu(int cpu) { }
4557 static int __cpuinit
4558 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4560 unsigned int cpu = (long)hcpu;
4564 case CPU_UP_PREPARE:
4565 case CPU_UP_PREPARE_FROZEN:
4566 perf_counter_init_cpu(cpu);
4569 case CPU_DOWN_PREPARE:
4570 case CPU_DOWN_PREPARE_FROZEN:
4571 perf_counter_exit_cpu(cpu);
4582 * This has to have a higher priority than migration_notifier in sched.c.
4584 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4585 .notifier_call = perf_cpu_notify,
4589 void __init perf_counter_init(void)
4591 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4592 (void *)(long)smp_processor_id());
4593 register_cpu_notifier(&perf_cpu_nb);
4596 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4598 return sprintf(buf, "%d\n", perf_reserved_percpu);
4602 perf_set_reserve_percpu(struct sysdev_class *class,
4606 struct perf_cpu_context *cpuctx;
4610 err = strict_strtoul(buf, 10, &val);
4613 if (val > perf_max_counters)
4616 spin_lock(&perf_resource_lock);
4617 perf_reserved_percpu = val;
4618 for_each_online_cpu(cpu) {
4619 cpuctx = &per_cpu(perf_cpu_context, cpu);
4620 spin_lock_irq(&cpuctx->ctx.lock);
4621 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4622 perf_max_counters - perf_reserved_percpu);
4623 cpuctx->max_pertask = mpt;
4624 spin_unlock_irq(&cpuctx->ctx.lock);
4626 spin_unlock(&perf_resource_lock);
4631 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4633 return sprintf(buf, "%d\n", perf_overcommit);
4637 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4642 err = strict_strtoul(buf, 10, &val);
4648 spin_lock(&perf_resource_lock);
4649 perf_overcommit = val;
4650 spin_unlock(&perf_resource_lock);
4655 static SYSDEV_CLASS_ATTR(
4658 perf_show_reserve_percpu,
4659 perf_set_reserve_percpu
4662 static SYSDEV_CLASS_ATTR(
4665 perf_show_overcommit,
4669 static struct attribute *perfclass_attrs[] = {
4670 &attr_reserve_percpu.attr,
4671 &attr_overcommit.attr,
4675 static struct attribute_group perfclass_attr_group = {
4676 .attrs = perfclass_attrs,
4677 .name = "perf_counters",
4680 static int __init perf_counter_sysfs_init(void)
4682 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4683 &perfclass_attr_group);
4685 device_initcall(perf_counter_sysfs_init);