2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_counters __read_mostly;
44 static atomic_t nr_comm_counters __read_mostly;
47 * perf counter paranoia level:
49 * 1 - disallow cpu counters to unpriv
50 * 2 - disallow kernel profiling to unpriv
52 int sysctl_perf_counter_paranoid __read_mostly;
54 static inline bool perf_paranoid_cpu(void)
56 return sysctl_perf_counter_paranoid > 0;
59 static inline bool perf_paranoid_kernel(void)
61 return sysctl_perf_counter_paranoid > 1;
64 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
67 * max perf counter sample rate
69 int sysctl_perf_counter_sample_rate __read_mostly = 100000;
71 static atomic64_t perf_counter_id;
74 * Lock for (sysadmin-configurable) counter reservations:
76 static DEFINE_SPINLOCK(perf_resource_lock);
79 * Architecture provided APIs - weak aliases:
81 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
86 void __weak hw_perf_disable(void) { barrier(); }
87 void __weak hw_perf_enable(void) { barrier(); }
89 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
92 hw_perf_group_sched_in(struct perf_counter *group_leader,
93 struct perf_cpu_context *cpuctx,
94 struct perf_counter_context *ctx, int cpu)
99 void __weak perf_counter_print_debug(void) { }
101 static DEFINE_PER_CPU(int, disable_count);
103 void __perf_disable(void)
105 __get_cpu_var(disable_count)++;
108 bool __perf_enable(void)
110 return !--__get_cpu_var(disable_count);
113 void perf_disable(void)
119 void perf_enable(void)
125 static void get_ctx(struct perf_counter_context *ctx)
127 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
130 static void free_ctx(struct rcu_head *head)
132 struct perf_counter_context *ctx;
134 ctx = container_of(head, struct perf_counter_context, rcu_head);
138 static void put_ctx(struct perf_counter_context *ctx)
140 if (atomic_dec_and_test(&ctx->refcount)) {
142 put_ctx(ctx->parent_ctx);
144 put_task_struct(ctx->task);
145 call_rcu(&ctx->rcu_head, free_ctx);
150 * Get the perf_counter_context for a task and lock it.
151 * This has to cope with with the fact that until it is locked,
152 * the context could get moved to another task.
154 static struct perf_counter_context *
155 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
157 struct perf_counter_context *ctx;
161 ctx = rcu_dereference(task->perf_counter_ctxp);
164 * If this context is a clone of another, it might
165 * get swapped for another underneath us by
166 * perf_counter_task_sched_out, though the
167 * rcu_read_lock() protects us from any context
168 * getting freed. Lock the context and check if it
169 * got swapped before we could get the lock, and retry
170 * if so. If we locked the right context, then it
171 * can't get swapped on us any more.
173 spin_lock_irqsave(&ctx->lock, *flags);
174 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
175 spin_unlock_irqrestore(&ctx->lock, *flags);
179 if (!atomic_inc_not_zero(&ctx->refcount)) {
180 spin_unlock_irqrestore(&ctx->lock, *flags);
189 * Get the context for a task and increment its pin_count so it
190 * can't get swapped to another task. This also increments its
191 * reference count so that the context can't get freed.
193 static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
195 struct perf_counter_context *ctx;
198 ctx = perf_lock_task_context(task, &flags);
201 spin_unlock_irqrestore(&ctx->lock, flags);
206 static void perf_unpin_context(struct perf_counter_context *ctx)
210 spin_lock_irqsave(&ctx->lock, flags);
212 spin_unlock_irqrestore(&ctx->lock, flags);
217 * Add a counter from the lists for its context.
218 * Must be called with ctx->mutex and ctx->lock held.
221 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
223 struct perf_counter *group_leader = counter->group_leader;
226 * Depending on whether it is a standalone or sibling counter,
227 * add it straight to the context's counter list, or to the group
228 * leader's sibling list:
230 if (group_leader == counter)
231 list_add_tail(&counter->list_entry, &ctx->counter_list);
233 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
234 group_leader->nr_siblings++;
237 list_add_rcu(&counter->event_entry, &ctx->event_list);
239 if (counter->attr.inherit_stat)
244 * Remove a counter from the lists for its context.
245 * Must be called with ctx->mutex and ctx->lock held.
248 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
250 struct perf_counter *sibling, *tmp;
252 if (list_empty(&counter->list_entry))
255 if (counter->attr.inherit_stat)
258 list_del_init(&counter->list_entry);
259 list_del_rcu(&counter->event_entry);
261 if (counter->group_leader != counter)
262 counter->group_leader->nr_siblings--;
265 * If this was a group counter with sibling counters then
266 * upgrade the siblings to singleton counters by adding them
267 * to the context list directly:
269 list_for_each_entry_safe(sibling, tmp,
270 &counter->sibling_list, list_entry) {
272 list_move_tail(&sibling->list_entry, &ctx->counter_list);
273 sibling->group_leader = sibling;
278 counter_sched_out(struct perf_counter *counter,
279 struct perf_cpu_context *cpuctx,
280 struct perf_counter_context *ctx)
282 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
285 counter->state = PERF_COUNTER_STATE_INACTIVE;
286 counter->tstamp_stopped = ctx->time;
287 counter->pmu->disable(counter);
290 if (!is_software_counter(counter))
291 cpuctx->active_oncpu--;
293 if (counter->attr.exclusive || !cpuctx->active_oncpu)
294 cpuctx->exclusive = 0;
298 group_sched_out(struct perf_counter *group_counter,
299 struct perf_cpu_context *cpuctx,
300 struct perf_counter_context *ctx)
302 struct perf_counter *counter;
304 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
307 counter_sched_out(group_counter, cpuctx, ctx);
310 * Schedule out siblings (if any):
312 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
313 counter_sched_out(counter, cpuctx, ctx);
315 if (group_counter->attr.exclusive)
316 cpuctx->exclusive = 0;
320 * Cross CPU call to remove a performance counter
322 * We disable the counter on the hardware level first. After that we
323 * remove it from the context list.
325 static void __perf_counter_remove_from_context(void *info)
327 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
328 struct perf_counter *counter = info;
329 struct perf_counter_context *ctx = counter->ctx;
332 * If this is a task context, we need to check whether it is
333 * the current task context of this cpu. If not it has been
334 * scheduled out before the smp call arrived.
336 if (ctx->task && cpuctx->task_ctx != ctx)
339 spin_lock(&ctx->lock);
341 * Protect the list operation against NMI by disabling the
342 * counters on a global level.
346 counter_sched_out(counter, cpuctx, ctx);
348 list_del_counter(counter, ctx);
352 * Allow more per task counters with respect to the
355 cpuctx->max_pertask =
356 min(perf_max_counters - ctx->nr_counters,
357 perf_max_counters - perf_reserved_percpu);
361 spin_unlock(&ctx->lock);
366 * Remove the counter from a task's (or a CPU's) list of counters.
368 * Must be called with ctx->mutex held.
370 * CPU counters are removed with a smp call. For task counters we only
371 * call when the task is on a CPU.
373 * If counter->ctx is a cloned context, callers must make sure that
374 * every task struct that counter->ctx->task could possibly point to
375 * remains valid. This is OK when called from perf_release since
376 * that only calls us on the top-level context, which can't be a clone.
377 * When called from perf_counter_exit_task, it's OK because the
378 * context has been detached from its task.
380 static void perf_counter_remove_from_context(struct perf_counter *counter)
382 struct perf_counter_context *ctx = counter->ctx;
383 struct task_struct *task = ctx->task;
387 * Per cpu counters are removed via an smp call and
388 * the removal is always sucessful.
390 smp_call_function_single(counter->cpu,
391 __perf_counter_remove_from_context,
397 task_oncpu_function_call(task, __perf_counter_remove_from_context,
400 spin_lock_irq(&ctx->lock);
402 * If the context is active we need to retry the smp call.
404 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
405 spin_unlock_irq(&ctx->lock);
410 * The lock prevents that this context is scheduled in so we
411 * can remove the counter safely, if the call above did not
414 if (!list_empty(&counter->list_entry)) {
415 list_del_counter(counter, ctx);
417 spin_unlock_irq(&ctx->lock);
420 static inline u64 perf_clock(void)
422 return cpu_clock(smp_processor_id());
426 * Update the record of the current time in a context.
428 static void update_context_time(struct perf_counter_context *ctx)
430 u64 now = perf_clock();
432 ctx->time += now - ctx->timestamp;
433 ctx->timestamp = now;
437 * Update the total_time_enabled and total_time_running fields for a counter.
439 static void update_counter_times(struct perf_counter *counter)
441 struct perf_counter_context *ctx = counter->ctx;
444 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
447 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
449 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
450 run_end = counter->tstamp_stopped;
454 counter->total_time_running = run_end - counter->tstamp_running;
458 * Update total_time_enabled and total_time_running for all counters in a group.
460 static void update_group_times(struct perf_counter *leader)
462 struct perf_counter *counter;
464 update_counter_times(leader);
465 list_for_each_entry(counter, &leader->sibling_list, list_entry)
466 update_counter_times(counter);
470 * Cross CPU call to disable a performance counter
472 static void __perf_counter_disable(void *info)
474 struct perf_counter *counter = info;
475 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
476 struct perf_counter_context *ctx = counter->ctx;
479 * If this is a per-task counter, need to check whether this
480 * counter's task is the current task on this cpu.
482 if (ctx->task && cpuctx->task_ctx != ctx)
485 spin_lock(&ctx->lock);
488 * If the counter is on, turn it off.
489 * If it is in error state, leave it in error state.
491 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
492 update_context_time(ctx);
493 update_counter_times(counter);
494 if (counter == counter->group_leader)
495 group_sched_out(counter, cpuctx, ctx);
497 counter_sched_out(counter, cpuctx, ctx);
498 counter->state = PERF_COUNTER_STATE_OFF;
501 spin_unlock(&ctx->lock);
507 * If counter->ctx is a cloned context, callers must make sure that
508 * every task struct that counter->ctx->task could possibly point to
509 * remains valid. This condition is satisifed when called through
510 * perf_counter_for_each_child or perf_counter_for_each because they
511 * hold the top-level counter's child_mutex, so any descendant that
512 * goes to exit will block in sync_child_counter.
513 * When called from perf_pending_counter it's OK because counter->ctx
514 * is the current context on this CPU and preemption is disabled,
515 * hence we can't get into perf_counter_task_sched_out for this context.
517 static void perf_counter_disable(struct perf_counter *counter)
519 struct perf_counter_context *ctx = counter->ctx;
520 struct task_struct *task = ctx->task;
524 * Disable the counter on the cpu that it's on
526 smp_call_function_single(counter->cpu, __perf_counter_disable,
532 task_oncpu_function_call(task, __perf_counter_disable, counter);
534 spin_lock_irq(&ctx->lock);
536 * If the counter is still active, we need to retry the cross-call.
538 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
539 spin_unlock_irq(&ctx->lock);
544 * Since we have the lock this context can't be scheduled
545 * in, so we can change the state safely.
547 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
548 update_counter_times(counter);
549 counter->state = PERF_COUNTER_STATE_OFF;
552 spin_unlock_irq(&ctx->lock);
556 counter_sched_in(struct perf_counter *counter,
557 struct perf_cpu_context *cpuctx,
558 struct perf_counter_context *ctx,
561 if (counter->state <= PERF_COUNTER_STATE_OFF)
564 counter->state = PERF_COUNTER_STATE_ACTIVE;
565 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
567 * The new state must be visible before we turn it on in the hardware:
571 if (counter->pmu->enable(counter)) {
572 counter->state = PERF_COUNTER_STATE_INACTIVE;
577 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
579 if (!is_software_counter(counter))
580 cpuctx->active_oncpu++;
583 if (counter->attr.exclusive)
584 cpuctx->exclusive = 1;
590 group_sched_in(struct perf_counter *group_counter,
591 struct perf_cpu_context *cpuctx,
592 struct perf_counter_context *ctx,
595 struct perf_counter *counter, *partial_group;
598 if (group_counter->state == PERF_COUNTER_STATE_OFF)
601 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
603 return ret < 0 ? ret : 0;
605 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
609 * Schedule in siblings as one group (if any):
611 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
612 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
613 partial_group = counter;
622 * Groups can be scheduled in as one unit only, so undo any
623 * partial group before returning:
625 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
626 if (counter == partial_group)
628 counter_sched_out(counter, cpuctx, ctx);
630 counter_sched_out(group_counter, cpuctx, ctx);
636 * Return 1 for a group consisting entirely of software counters,
637 * 0 if the group contains any hardware counters.
639 static int is_software_only_group(struct perf_counter *leader)
641 struct perf_counter *counter;
643 if (!is_software_counter(leader))
646 list_for_each_entry(counter, &leader->sibling_list, list_entry)
647 if (!is_software_counter(counter))
654 * Work out whether we can put this counter group on the CPU now.
656 static int group_can_go_on(struct perf_counter *counter,
657 struct perf_cpu_context *cpuctx,
661 * Groups consisting entirely of software counters can always go on.
663 if (is_software_only_group(counter))
666 * If an exclusive group is already on, no other hardware
667 * counters can go on.
669 if (cpuctx->exclusive)
672 * If this group is exclusive and there are already
673 * counters on the CPU, it can't go on.
675 if (counter->attr.exclusive && cpuctx->active_oncpu)
678 * Otherwise, try to add it if all previous groups were able
684 static void add_counter_to_ctx(struct perf_counter *counter,
685 struct perf_counter_context *ctx)
687 list_add_counter(counter, ctx);
688 counter->tstamp_enabled = ctx->time;
689 counter->tstamp_running = ctx->time;
690 counter->tstamp_stopped = ctx->time;
694 * Cross CPU call to install and enable a performance counter
696 * Must be called with ctx->mutex held
698 static void __perf_install_in_context(void *info)
700 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
701 struct perf_counter *counter = info;
702 struct perf_counter_context *ctx = counter->ctx;
703 struct perf_counter *leader = counter->group_leader;
704 int cpu = smp_processor_id();
708 * If this is a task context, we need to check whether it is
709 * the current task context of this cpu. If not it has been
710 * scheduled out before the smp call arrived.
711 * Or possibly this is the right context but it isn't
712 * on this cpu because it had no counters.
714 if (ctx->task && cpuctx->task_ctx != ctx) {
715 if (cpuctx->task_ctx || ctx->task != current)
717 cpuctx->task_ctx = ctx;
720 spin_lock(&ctx->lock);
722 update_context_time(ctx);
725 * Protect the list operation against NMI by disabling the
726 * counters on a global level. NOP for non NMI based counters.
730 add_counter_to_ctx(counter, ctx);
733 * Don't put the counter on if it is disabled or if
734 * it is in a group and the group isn't on.
736 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
737 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
741 * An exclusive counter can't go on if there are already active
742 * hardware counters, and no hardware counter can go on if there
743 * is already an exclusive counter on.
745 if (!group_can_go_on(counter, cpuctx, 1))
748 err = counter_sched_in(counter, cpuctx, ctx, cpu);
752 * This counter couldn't go on. If it is in a group
753 * then we have to pull the whole group off.
754 * If the counter group is pinned then put it in error state.
756 if (leader != counter)
757 group_sched_out(leader, cpuctx, ctx);
758 if (leader->attr.pinned) {
759 update_group_times(leader);
760 leader->state = PERF_COUNTER_STATE_ERROR;
764 if (!err && !ctx->task && cpuctx->max_pertask)
765 cpuctx->max_pertask--;
770 spin_unlock(&ctx->lock);
774 * Attach a performance counter to a context
776 * First we add the counter to the list with the hardware enable bit
777 * in counter->hw_config cleared.
779 * If the counter is attached to a task which is on a CPU we use a smp
780 * call to enable it in the task context. The task might have been
781 * scheduled away, but we check this in the smp call again.
783 * Must be called with ctx->mutex held.
786 perf_install_in_context(struct perf_counter_context *ctx,
787 struct perf_counter *counter,
790 struct task_struct *task = ctx->task;
794 * Per cpu counters are installed via an smp call and
795 * the install is always sucessful.
797 smp_call_function_single(cpu, __perf_install_in_context,
803 task_oncpu_function_call(task, __perf_install_in_context,
806 spin_lock_irq(&ctx->lock);
808 * we need to retry the smp call.
810 if (ctx->is_active && list_empty(&counter->list_entry)) {
811 spin_unlock_irq(&ctx->lock);
816 * The lock prevents that this context is scheduled in so we
817 * can add the counter safely, if it the call above did not
820 if (list_empty(&counter->list_entry))
821 add_counter_to_ctx(counter, ctx);
822 spin_unlock_irq(&ctx->lock);
826 * Cross CPU call to enable a performance counter
828 static void __perf_counter_enable(void *info)
830 struct perf_counter *counter = info;
831 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
832 struct perf_counter_context *ctx = counter->ctx;
833 struct perf_counter *leader = counter->group_leader;
837 * If this is a per-task counter, need to check whether this
838 * counter's task is the current task on this cpu.
840 if (ctx->task && cpuctx->task_ctx != ctx) {
841 if (cpuctx->task_ctx || ctx->task != current)
843 cpuctx->task_ctx = ctx;
846 spin_lock(&ctx->lock);
848 update_context_time(ctx);
850 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
852 counter->state = PERF_COUNTER_STATE_INACTIVE;
853 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
856 * If the counter is in a group and isn't the group leader,
857 * then don't put it on unless the group is on.
859 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
862 if (!group_can_go_on(counter, cpuctx, 1)) {
866 if (counter == leader)
867 err = group_sched_in(counter, cpuctx, ctx,
870 err = counter_sched_in(counter, cpuctx, ctx,
877 * If this counter can't go on and it's part of a
878 * group, then the whole group has to come off.
880 if (leader != counter)
881 group_sched_out(leader, cpuctx, ctx);
882 if (leader->attr.pinned) {
883 update_group_times(leader);
884 leader->state = PERF_COUNTER_STATE_ERROR;
889 spin_unlock(&ctx->lock);
895 * If counter->ctx is a cloned context, callers must make sure that
896 * every task struct that counter->ctx->task could possibly point to
897 * remains valid. This condition is satisfied when called through
898 * perf_counter_for_each_child or perf_counter_for_each as described
899 * for perf_counter_disable.
901 static void perf_counter_enable(struct perf_counter *counter)
903 struct perf_counter_context *ctx = counter->ctx;
904 struct task_struct *task = ctx->task;
908 * Enable the counter on the cpu that it's on
910 smp_call_function_single(counter->cpu, __perf_counter_enable,
915 spin_lock_irq(&ctx->lock);
916 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
920 * If the counter is in error state, clear that first.
921 * That way, if we see the counter in error state below, we
922 * know that it has gone back into error state, as distinct
923 * from the task having been scheduled away before the
924 * cross-call arrived.
926 if (counter->state == PERF_COUNTER_STATE_ERROR)
927 counter->state = PERF_COUNTER_STATE_OFF;
930 spin_unlock_irq(&ctx->lock);
931 task_oncpu_function_call(task, __perf_counter_enable, counter);
933 spin_lock_irq(&ctx->lock);
936 * If the context is active and the counter is still off,
937 * we need to retry the cross-call.
939 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
943 * Since we have the lock this context can't be scheduled
944 * in, so we can change the state safely.
946 if (counter->state == PERF_COUNTER_STATE_OFF) {
947 counter->state = PERF_COUNTER_STATE_INACTIVE;
948 counter->tstamp_enabled =
949 ctx->time - counter->total_time_enabled;
952 spin_unlock_irq(&ctx->lock);
955 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
958 * not supported on inherited counters
960 if (counter->attr.inherit)
963 atomic_add(refresh, &counter->event_limit);
964 perf_counter_enable(counter);
969 void __perf_counter_sched_out(struct perf_counter_context *ctx,
970 struct perf_cpu_context *cpuctx)
972 struct perf_counter *counter;
974 spin_lock(&ctx->lock);
976 if (likely(!ctx->nr_counters))
978 update_context_time(ctx);
981 if (ctx->nr_active) {
982 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
983 if (counter != counter->group_leader)
984 counter_sched_out(counter, cpuctx, ctx);
986 group_sched_out(counter, cpuctx, ctx);
991 spin_unlock(&ctx->lock);
995 * Test whether two contexts are equivalent, i.e. whether they
996 * have both been cloned from the same version of the same context
997 * and they both have the same number of enabled counters.
998 * If the number of enabled counters is the same, then the set
999 * of enabled counters should be the same, because these are both
1000 * inherited contexts, therefore we can't access individual counters
1001 * in them directly with an fd; we can only enable/disable all
1002 * counters via prctl, or enable/disable all counters in a family
1003 * via ioctl, which will have the same effect on both contexts.
1005 static int context_equiv(struct perf_counter_context *ctx1,
1006 struct perf_counter_context *ctx2)
1008 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1009 && ctx1->parent_gen == ctx2->parent_gen
1010 && !ctx1->pin_count && !ctx2->pin_count;
1013 static void __perf_counter_read(void *counter);
1015 static void __perf_counter_sync_stat(struct perf_counter *counter,
1016 struct perf_counter *next_counter)
1020 if (!counter->attr.inherit_stat)
1024 * Update the counter value, we cannot use perf_counter_read()
1025 * because we're in the middle of a context switch and have IRQs
1026 * disabled, which upsets smp_call_function_single(), however
1027 * we know the counter must be on the current CPU, therefore we
1028 * don't need to use it.
1030 switch (counter->state) {
1031 case PERF_COUNTER_STATE_ACTIVE:
1032 __perf_counter_read(counter);
1035 case PERF_COUNTER_STATE_INACTIVE:
1036 update_counter_times(counter);
1044 * In order to keep per-task stats reliable we need to flip the counter
1045 * values when we flip the contexts.
1047 value = atomic64_read(&next_counter->count);
1048 value = atomic64_xchg(&counter->count, value);
1049 atomic64_set(&next_counter->count, value);
1052 * XXX also sync time_enabled and time_running ?
1056 #define list_next_entry(pos, member) \
1057 list_entry(pos->member.next, typeof(*pos), member)
1059 static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1060 struct perf_counter_context *next_ctx)
1062 struct perf_counter *counter, *next_counter;
1067 counter = list_first_entry(&ctx->event_list,
1068 struct perf_counter, event_entry);
1070 next_counter = list_first_entry(&next_ctx->event_list,
1071 struct perf_counter, event_entry);
1073 while (&counter->event_entry != &ctx->event_list &&
1074 &next_counter->event_entry != &next_ctx->event_list) {
1076 __perf_counter_sync_stat(counter, next_counter);
1078 counter = list_next_entry(counter, event_entry);
1079 next_counter = list_next_entry(counter, event_entry);
1084 * Called from scheduler to remove the counters of the current task,
1085 * with interrupts disabled.
1087 * We stop each counter and update the counter value in counter->count.
1089 * This does not protect us against NMI, but disable()
1090 * sets the disabled bit in the control field of counter _before_
1091 * accessing the counter control register. If a NMI hits, then it will
1092 * not restart the counter.
1094 void perf_counter_task_sched_out(struct task_struct *task,
1095 struct task_struct *next, int cpu)
1097 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1098 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1099 struct perf_counter_context *next_ctx;
1100 struct perf_counter_context *parent;
1101 struct pt_regs *regs;
1104 regs = task_pt_regs(task);
1105 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1107 if (likely(!ctx || !cpuctx->task_ctx))
1110 update_context_time(ctx);
1113 parent = rcu_dereference(ctx->parent_ctx);
1114 next_ctx = next->perf_counter_ctxp;
1115 if (parent && next_ctx &&
1116 rcu_dereference(next_ctx->parent_ctx) == parent) {
1118 * Looks like the two contexts are clones, so we might be
1119 * able to optimize the context switch. We lock both
1120 * contexts and check that they are clones under the
1121 * lock (including re-checking that neither has been
1122 * uncloned in the meantime). It doesn't matter which
1123 * order we take the locks because no other cpu could
1124 * be trying to lock both of these tasks.
1126 spin_lock(&ctx->lock);
1127 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1128 if (context_equiv(ctx, next_ctx)) {
1130 * XXX do we need a memory barrier of sorts
1131 * wrt to rcu_dereference() of perf_counter_ctxp
1133 task->perf_counter_ctxp = next_ctx;
1134 next->perf_counter_ctxp = ctx;
1136 next_ctx->task = task;
1139 perf_counter_sync_stat(ctx, next_ctx);
1141 spin_unlock(&next_ctx->lock);
1142 spin_unlock(&ctx->lock);
1147 __perf_counter_sched_out(ctx, cpuctx);
1148 cpuctx->task_ctx = NULL;
1153 * Called with IRQs disabled
1155 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1157 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1159 if (!cpuctx->task_ctx)
1162 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1165 __perf_counter_sched_out(ctx, cpuctx);
1166 cpuctx->task_ctx = NULL;
1170 * Called with IRQs disabled
1172 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1174 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1178 __perf_counter_sched_in(struct perf_counter_context *ctx,
1179 struct perf_cpu_context *cpuctx, int cpu)
1181 struct perf_counter *counter;
1184 spin_lock(&ctx->lock);
1186 if (likely(!ctx->nr_counters))
1189 ctx->timestamp = perf_clock();
1194 * First go through the list and put on any pinned groups
1195 * in order to give them the best chance of going on.
1197 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1198 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1199 !counter->attr.pinned)
1201 if (counter->cpu != -1 && counter->cpu != cpu)
1204 if (counter != counter->group_leader)
1205 counter_sched_in(counter, cpuctx, ctx, cpu);
1207 if (group_can_go_on(counter, cpuctx, 1))
1208 group_sched_in(counter, cpuctx, ctx, cpu);
1212 * If this pinned group hasn't been scheduled,
1213 * put it in error state.
1215 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1216 update_group_times(counter);
1217 counter->state = PERF_COUNTER_STATE_ERROR;
1221 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1223 * Ignore counters in OFF or ERROR state, and
1224 * ignore pinned counters since we did them already.
1226 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1227 counter->attr.pinned)
1231 * Listen to the 'cpu' scheduling filter constraint
1234 if (counter->cpu != -1 && counter->cpu != cpu)
1237 if (counter != counter->group_leader) {
1238 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1241 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1242 if (group_sched_in(counter, cpuctx, ctx, cpu))
1249 spin_unlock(&ctx->lock);
1253 * Called from scheduler to add the counters of the current task
1254 * with interrupts disabled.
1256 * We restore the counter value and then enable it.
1258 * This does not protect us against NMI, but enable()
1259 * sets the enabled bit in the control field of counter _before_
1260 * accessing the counter control register. If a NMI hits, then it will
1261 * keep the counter running.
1263 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1265 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1266 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1270 if (cpuctx->task_ctx == ctx)
1272 __perf_counter_sched_in(ctx, cpuctx, cpu);
1273 cpuctx->task_ctx = ctx;
1276 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1278 struct perf_counter_context *ctx = &cpuctx->ctx;
1280 __perf_counter_sched_in(ctx, cpuctx, cpu);
1283 #define MAX_INTERRUPTS (~0ULL)
1285 static void perf_log_throttle(struct perf_counter *counter, int enable);
1286 static void perf_log_period(struct perf_counter *counter, u64 period);
1288 static void perf_adjust_period(struct perf_counter *counter, u64 events)
1290 struct hw_perf_counter *hwc = &counter->hw;
1291 u64 period, sample_period;
1294 events *= hwc->sample_period;
1295 period = div64_u64(events, counter->attr.sample_freq);
1297 delta = (s64)(period - hwc->sample_period);
1298 delta = (delta + 7) / 8; /* low pass filter */
1300 sample_period = hwc->sample_period + delta;
1305 perf_log_period(counter, sample_period);
1307 hwc->sample_period = sample_period;
1310 static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1312 struct perf_counter *counter;
1313 struct hw_perf_counter *hwc;
1314 u64 interrupts, freq;
1316 spin_lock(&ctx->lock);
1317 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1318 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1323 interrupts = hwc->interrupts;
1324 hwc->interrupts = 0;
1327 * unthrottle counters on the tick
1329 if (interrupts == MAX_INTERRUPTS) {
1330 perf_log_throttle(counter, 1);
1331 counter->pmu->unthrottle(counter);
1332 interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1335 if (!counter->attr.freq || !counter->attr.sample_freq)
1339 * if the specified freq < HZ then we need to skip ticks
1341 if (counter->attr.sample_freq < HZ) {
1342 freq = counter->attr.sample_freq;
1344 hwc->freq_count += freq;
1345 hwc->freq_interrupts += interrupts;
1347 if (hwc->freq_count < HZ)
1350 interrupts = hwc->freq_interrupts;
1351 hwc->freq_interrupts = 0;
1352 hwc->freq_count -= HZ;
1356 perf_adjust_period(counter, freq * interrupts);
1359 * In order to avoid being stalled by an (accidental) huge
1360 * sample period, force reset the sample period if we didn't
1361 * get any events in this freq period.
1365 counter->pmu->disable(counter);
1366 atomic64_set(&hwc->period_left, 0);
1367 counter->pmu->enable(counter);
1371 spin_unlock(&ctx->lock);
1375 * Round-robin a context's counters:
1377 static void rotate_ctx(struct perf_counter_context *ctx)
1379 struct perf_counter *counter;
1381 if (!ctx->nr_counters)
1384 spin_lock(&ctx->lock);
1386 * Rotate the first entry last (works just fine for group counters too):
1389 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1390 list_move_tail(&counter->list_entry, &ctx->counter_list);
1395 spin_unlock(&ctx->lock);
1398 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1400 struct perf_cpu_context *cpuctx;
1401 struct perf_counter_context *ctx;
1403 if (!atomic_read(&nr_counters))
1406 cpuctx = &per_cpu(perf_cpu_context, cpu);
1407 ctx = curr->perf_counter_ctxp;
1409 perf_ctx_adjust_freq(&cpuctx->ctx);
1411 perf_ctx_adjust_freq(ctx);
1413 perf_counter_cpu_sched_out(cpuctx);
1415 __perf_counter_task_sched_out(ctx);
1417 rotate_ctx(&cpuctx->ctx);
1421 perf_counter_cpu_sched_in(cpuctx, cpu);
1423 perf_counter_task_sched_in(curr, cpu);
1427 * Cross CPU call to read the hardware counter
1429 static void __perf_counter_read(void *info)
1431 struct perf_counter *counter = info;
1432 struct perf_counter_context *ctx = counter->ctx;
1433 unsigned long flags;
1435 local_irq_save(flags);
1437 update_context_time(ctx);
1438 counter->pmu->read(counter);
1439 update_counter_times(counter);
1440 local_irq_restore(flags);
1443 static u64 perf_counter_read(struct perf_counter *counter)
1446 * If counter is enabled and currently active on a CPU, update the
1447 * value in the counter structure:
1449 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1450 smp_call_function_single(counter->oncpu,
1451 __perf_counter_read, counter, 1);
1452 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1453 update_counter_times(counter);
1456 return atomic64_read(&counter->count);
1460 * Initialize the perf_counter context in a task_struct:
1463 __perf_counter_init_context(struct perf_counter_context *ctx,
1464 struct task_struct *task)
1466 memset(ctx, 0, sizeof(*ctx));
1467 spin_lock_init(&ctx->lock);
1468 mutex_init(&ctx->mutex);
1469 INIT_LIST_HEAD(&ctx->counter_list);
1470 INIT_LIST_HEAD(&ctx->event_list);
1471 atomic_set(&ctx->refcount, 1);
1475 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1477 struct perf_counter_context *parent_ctx;
1478 struct perf_counter_context *ctx;
1479 struct perf_cpu_context *cpuctx;
1480 struct task_struct *task;
1481 unsigned long flags;
1485 * If cpu is not a wildcard then this is a percpu counter:
1488 /* Must be root to operate on a CPU counter: */
1489 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1490 return ERR_PTR(-EACCES);
1492 if (cpu < 0 || cpu > num_possible_cpus())
1493 return ERR_PTR(-EINVAL);
1496 * We could be clever and allow to attach a counter to an
1497 * offline CPU and activate it when the CPU comes up, but
1500 if (!cpu_isset(cpu, cpu_online_map))
1501 return ERR_PTR(-ENODEV);
1503 cpuctx = &per_cpu(perf_cpu_context, cpu);
1514 task = find_task_by_vpid(pid);
1516 get_task_struct(task);
1520 return ERR_PTR(-ESRCH);
1523 * Can't attach counters to a dying task.
1526 if (task->flags & PF_EXITING)
1529 /* Reuse ptrace permission checks for now. */
1531 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1535 ctx = perf_lock_task_context(task, &flags);
1537 parent_ctx = ctx->parent_ctx;
1539 put_ctx(parent_ctx);
1540 ctx->parent_ctx = NULL; /* no longer a clone */
1542 spin_unlock_irqrestore(&ctx->lock, flags);
1546 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1550 __perf_counter_init_context(ctx, task);
1552 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1554 * We raced with some other task; use
1555 * the context they set.
1560 get_task_struct(task);
1563 put_task_struct(task);
1567 put_task_struct(task);
1568 return ERR_PTR(err);
1571 static void free_counter_rcu(struct rcu_head *head)
1573 struct perf_counter *counter;
1575 counter = container_of(head, struct perf_counter, rcu_head);
1577 put_pid_ns(counter->ns);
1581 static void perf_pending_sync(struct perf_counter *counter);
1583 static void free_counter(struct perf_counter *counter)
1585 perf_pending_sync(counter);
1587 if (!counter->parent) {
1588 atomic_dec(&nr_counters);
1589 if (counter->attr.mmap)
1590 atomic_dec(&nr_mmap_counters);
1591 if (counter->attr.comm)
1592 atomic_dec(&nr_comm_counters);
1595 if (counter->destroy)
1596 counter->destroy(counter);
1598 put_ctx(counter->ctx);
1599 call_rcu(&counter->rcu_head, free_counter_rcu);
1603 * Called when the last reference to the file is gone.
1605 static int perf_release(struct inode *inode, struct file *file)
1607 struct perf_counter *counter = file->private_data;
1608 struct perf_counter_context *ctx = counter->ctx;
1610 file->private_data = NULL;
1612 WARN_ON_ONCE(ctx->parent_ctx);
1613 mutex_lock(&ctx->mutex);
1614 perf_counter_remove_from_context(counter);
1615 mutex_unlock(&ctx->mutex);
1617 mutex_lock(&counter->owner->perf_counter_mutex);
1618 list_del_init(&counter->owner_entry);
1619 mutex_unlock(&counter->owner->perf_counter_mutex);
1620 put_task_struct(counter->owner);
1622 free_counter(counter);
1628 * Read the performance counter - simple non blocking version for now
1631 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1637 * Return end-of-file for a read on a counter that is in
1638 * error state (i.e. because it was pinned but it couldn't be
1639 * scheduled on to the CPU at some point).
1641 if (counter->state == PERF_COUNTER_STATE_ERROR)
1644 WARN_ON_ONCE(counter->ctx->parent_ctx);
1645 mutex_lock(&counter->child_mutex);
1646 values[0] = perf_counter_read(counter);
1648 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1649 values[n++] = counter->total_time_enabled +
1650 atomic64_read(&counter->child_total_time_enabled);
1651 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1652 values[n++] = counter->total_time_running +
1653 atomic64_read(&counter->child_total_time_running);
1654 if (counter->attr.read_format & PERF_FORMAT_ID)
1655 values[n++] = counter->id;
1656 mutex_unlock(&counter->child_mutex);
1658 if (count < n * sizeof(u64))
1660 count = n * sizeof(u64);
1662 if (copy_to_user(buf, values, count))
1669 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1671 struct perf_counter *counter = file->private_data;
1673 return perf_read_hw(counter, buf, count);
1676 static unsigned int perf_poll(struct file *file, poll_table *wait)
1678 struct perf_counter *counter = file->private_data;
1679 struct perf_mmap_data *data;
1680 unsigned int events = POLL_HUP;
1683 data = rcu_dereference(counter->data);
1685 events = atomic_xchg(&data->poll, 0);
1688 poll_wait(file, &counter->waitq, wait);
1693 static void perf_counter_reset(struct perf_counter *counter)
1695 (void)perf_counter_read(counter);
1696 atomic64_set(&counter->count, 0);
1697 perf_counter_update_userpage(counter);
1701 * Holding the top-level counter's child_mutex means that any
1702 * descendant process that has inherited this counter will block
1703 * in sync_child_counter if it goes to exit, thus satisfying the
1704 * task existence requirements of perf_counter_enable/disable.
1706 static void perf_counter_for_each_child(struct perf_counter *counter,
1707 void (*func)(struct perf_counter *))
1709 struct perf_counter *child;
1711 WARN_ON_ONCE(counter->ctx->parent_ctx);
1712 mutex_lock(&counter->child_mutex);
1714 list_for_each_entry(child, &counter->child_list, child_list)
1716 mutex_unlock(&counter->child_mutex);
1719 static void perf_counter_for_each(struct perf_counter *counter,
1720 void (*func)(struct perf_counter *))
1722 struct perf_counter_context *ctx = counter->ctx;
1723 struct perf_counter *sibling;
1725 WARN_ON_ONCE(ctx->parent_ctx);
1726 mutex_lock(&ctx->mutex);
1727 counter = counter->group_leader;
1729 perf_counter_for_each_child(counter, func);
1731 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1732 perf_counter_for_each_child(counter, func);
1733 mutex_unlock(&ctx->mutex);
1736 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1738 struct perf_counter_context *ctx = counter->ctx;
1743 if (!counter->attr.sample_period)
1746 size = copy_from_user(&value, arg, sizeof(value));
1747 if (size != sizeof(value))
1753 spin_lock_irq(&ctx->lock);
1754 if (counter->attr.freq) {
1755 if (value > sysctl_perf_counter_sample_rate) {
1760 counter->attr.sample_freq = value;
1762 perf_log_period(counter, value);
1764 counter->attr.sample_period = value;
1765 counter->hw.sample_period = value;
1768 spin_unlock_irq(&ctx->lock);
1773 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1775 struct perf_counter *counter = file->private_data;
1776 void (*func)(struct perf_counter *);
1780 case PERF_COUNTER_IOC_ENABLE:
1781 func = perf_counter_enable;
1783 case PERF_COUNTER_IOC_DISABLE:
1784 func = perf_counter_disable;
1786 case PERF_COUNTER_IOC_RESET:
1787 func = perf_counter_reset;
1790 case PERF_COUNTER_IOC_REFRESH:
1791 return perf_counter_refresh(counter, arg);
1793 case PERF_COUNTER_IOC_PERIOD:
1794 return perf_counter_period(counter, (u64 __user *)arg);
1800 if (flags & PERF_IOC_FLAG_GROUP)
1801 perf_counter_for_each(counter, func);
1803 perf_counter_for_each_child(counter, func);
1808 int perf_counter_task_enable(void)
1810 struct perf_counter *counter;
1812 mutex_lock(¤t->perf_counter_mutex);
1813 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1814 perf_counter_for_each_child(counter, perf_counter_enable);
1815 mutex_unlock(¤t->perf_counter_mutex);
1820 int perf_counter_task_disable(void)
1822 struct perf_counter *counter;
1824 mutex_lock(¤t->perf_counter_mutex);
1825 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1826 perf_counter_for_each_child(counter, perf_counter_disable);
1827 mutex_unlock(¤t->perf_counter_mutex);
1832 static int perf_counter_index(struct perf_counter *counter)
1834 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1837 return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
1841 * Callers need to ensure there can be no nesting of this function, otherwise
1842 * the seqlock logic goes bad. We can not serialize this because the arch
1843 * code calls this from NMI context.
1845 void perf_counter_update_userpage(struct perf_counter *counter)
1847 struct perf_counter_mmap_page *userpg;
1848 struct perf_mmap_data *data;
1851 data = rcu_dereference(counter->data);
1855 userpg = data->user_page;
1858 * Disable preemption so as to not let the corresponding user-space
1859 * spin too long if we get preempted.
1864 userpg->index = perf_counter_index(counter);
1865 userpg->offset = atomic64_read(&counter->count);
1866 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1867 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1869 userpg->time_enabled = counter->total_time_enabled +
1870 atomic64_read(&counter->child_total_time_enabled);
1872 userpg->time_running = counter->total_time_running +
1873 atomic64_read(&counter->child_total_time_running);
1882 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1884 struct perf_counter *counter = vma->vm_file->private_data;
1885 struct perf_mmap_data *data;
1886 int ret = VM_FAULT_SIGBUS;
1888 if (vmf->flags & FAULT_FLAG_MKWRITE) {
1889 if (vmf->pgoff == 0)
1895 data = rcu_dereference(counter->data);
1899 if (vmf->pgoff == 0) {
1900 vmf->page = virt_to_page(data->user_page);
1902 int nr = vmf->pgoff - 1;
1904 if ((unsigned)nr > data->nr_pages)
1907 if (vmf->flags & FAULT_FLAG_WRITE)
1910 vmf->page = virt_to_page(data->data_pages[nr]);
1913 get_page(vmf->page);
1914 vmf->page->mapping = vma->vm_file->f_mapping;
1915 vmf->page->index = vmf->pgoff;
1924 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1926 struct perf_mmap_data *data;
1930 WARN_ON(atomic_read(&counter->mmap_count));
1932 size = sizeof(struct perf_mmap_data);
1933 size += nr_pages * sizeof(void *);
1935 data = kzalloc(size, GFP_KERNEL);
1939 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1940 if (!data->user_page)
1941 goto fail_user_page;
1943 for (i = 0; i < nr_pages; i++) {
1944 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1945 if (!data->data_pages[i])
1946 goto fail_data_pages;
1949 data->nr_pages = nr_pages;
1950 atomic_set(&data->lock, -1);
1952 rcu_assign_pointer(counter->data, data);
1957 for (i--; i >= 0; i--)
1958 free_page((unsigned long)data->data_pages[i]);
1960 free_page((unsigned long)data->user_page);
1969 static void perf_mmap_free_page(unsigned long addr)
1971 struct page *page = virt_to_page(addr);
1973 page->mapping = NULL;
1977 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1979 struct perf_mmap_data *data;
1982 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
1984 perf_mmap_free_page((unsigned long)data->user_page);
1985 for (i = 0; i < data->nr_pages; i++)
1986 perf_mmap_free_page((unsigned long)data->data_pages[i]);
1991 static void perf_mmap_data_free(struct perf_counter *counter)
1993 struct perf_mmap_data *data = counter->data;
1995 WARN_ON(atomic_read(&counter->mmap_count));
1997 rcu_assign_pointer(counter->data, NULL);
1998 call_rcu(&data->rcu_head, __perf_mmap_data_free);
2001 static void perf_mmap_open(struct vm_area_struct *vma)
2003 struct perf_counter *counter = vma->vm_file->private_data;
2005 atomic_inc(&counter->mmap_count);
2008 static void perf_mmap_close(struct vm_area_struct *vma)
2010 struct perf_counter *counter = vma->vm_file->private_data;
2012 WARN_ON_ONCE(counter->ctx->parent_ctx);
2013 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2014 struct user_struct *user = current_user();
2016 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2017 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2018 perf_mmap_data_free(counter);
2019 mutex_unlock(&counter->mmap_mutex);
2023 static struct vm_operations_struct perf_mmap_vmops = {
2024 .open = perf_mmap_open,
2025 .close = perf_mmap_close,
2026 .fault = perf_mmap_fault,
2027 .page_mkwrite = perf_mmap_fault,
2030 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2032 struct perf_counter *counter = file->private_data;
2033 unsigned long user_locked, user_lock_limit;
2034 struct user_struct *user = current_user();
2035 unsigned long locked, lock_limit;
2036 unsigned long vma_size;
2037 unsigned long nr_pages;
2038 long user_extra, extra;
2041 if (!(vma->vm_flags & VM_SHARED))
2044 vma_size = vma->vm_end - vma->vm_start;
2045 nr_pages = (vma_size / PAGE_SIZE) - 1;
2048 * If we have data pages ensure they're a power-of-two number, so we
2049 * can do bitmasks instead of modulo.
2051 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2054 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2057 if (vma->vm_pgoff != 0)
2060 WARN_ON_ONCE(counter->ctx->parent_ctx);
2061 mutex_lock(&counter->mmap_mutex);
2062 if (atomic_inc_not_zero(&counter->mmap_count)) {
2063 if (nr_pages != counter->data->nr_pages)
2068 user_extra = nr_pages + 1;
2069 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2072 * Increase the limit linearly with more CPUs:
2074 user_lock_limit *= num_online_cpus();
2076 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2079 if (user_locked > user_lock_limit)
2080 extra = user_locked - user_lock_limit;
2082 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2083 lock_limit >>= PAGE_SHIFT;
2084 locked = vma->vm_mm->locked_vm + extra;
2086 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2091 WARN_ON(counter->data);
2092 ret = perf_mmap_data_alloc(counter, nr_pages);
2096 atomic_set(&counter->mmap_count, 1);
2097 atomic_long_add(user_extra, &user->locked_vm);
2098 vma->vm_mm->locked_vm += extra;
2099 counter->data->nr_locked = extra;
2100 if (vma->vm_flags & VM_WRITE)
2101 counter->data->writable = 1;
2104 mutex_unlock(&counter->mmap_mutex);
2106 vma->vm_flags |= VM_RESERVED;
2107 vma->vm_ops = &perf_mmap_vmops;
2112 static int perf_fasync(int fd, struct file *filp, int on)
2114 struct inode *inode = filp->f_path.dentry->d_inode;
2115 struct perf_counter *counter = filp->private_data;
2118 mutex_lock(&inode->i_mutex);
2119 retval = fasync_helper(fd, filp, on, &counter->fasync);
2120 mutex_unlock(&inode->i_mutex);
2128 static const struct file_operations perf_fops = {
2129 .release = perf_release,
2132 .unlocked_ioctl = perf_ioctl,
2133 .compat_ioctl = perf_ioctl,
2135 .fasync = perf_fasync,
2139 * Perf counter wakeup
2141 * If there's data, ensure we set the poll() state and publish everything
2142 * to user-space before waking everybody up.
2145 void perf_counter_wakeup(struct perf_counter *counter)
2147 wake_up_all(&counter->waitq);
2149 if (counter->pending_kill) {
2150 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2151 counter->pending_kill = 0;
2158 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2160 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2161 * single linked list and use cmpxchg() to add entries lockless.
2164 static void perf_pending_counter(struct perf_pending_entry *entry)
2166 struct perf_counter *counter = container_of(entry,
2167 struct perf_counter, pending);
2169 if (counter->pending_disable) {
2170 counter->pending_disable = 0;
2171 perf_counter_disable(counter);
2174 if (counter->pending_wakeup) {
2175 counter->pending_wakeup = 0;
2176 perf_counter_wakeup(counter);
2180 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2182 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2186 static void perf_pending_queue(struct perf_pending_entry *entry,
2187 void (*func)(struct perf_pending_entry *))
2189 struct perf_pending_entry **head;
2191 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2196 head = &get_cpu_var(perf_pending_head);
2199 entry->next = *head;
2200 } while (cmpxchg(head, entry->next, entry) != entry->next);
2202 set_perf_counter_pending();
2204 put_cpu_var(perf_pending_head);
2207 static int __perf_pending_run(void)
2209 struct perf_pending_entry *list;
2212 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2213 while (list != PENDING_TAIL) {
2214 void (*func)(struct perf_pending_entry *);
2215 struct perf_pending_entry *entry = list;
2222 * Ensure we observe the unqueue before we issue the wakeup,
2223 * so that we won't be waiting forever.
2224 * -- see perf_not_pending().
2235 static inline int perf_not_pending(struct perf_counter *counter)
2238 * If we flush on whatever cpu we run, there is a chance we don't
2242 __perf_pending_run();
2246 * Ensure we see the proper queue state before going to sleep
2247 * so that we do not miss the wakeup. -- see perf_pending_handle()
2250 return counter->pending.next == NULL;
2253 static void perf_pending_sync(struct perf_counter *counter)
2255 wait_event(counter->waitq, perf_not_pending(counter));
2258 void perf_counter_do_pending(void)
2260 __perf_pending_run();
2264 * Callchain support -- arch specific
2267 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2276 struct perf_output_handle {
2277 struct perf_counter *counter;
2278 struct perf_mmap_data *data;
2280 unsigned long offset;
2284 unsigned long flags;
2287 static bool perf_output_space(struct perf_mmap_data *data,
2288 unsigned int offset, unsigned int head)
2293 if (!data->writable)
2296 mask = (data->nr_pages << PAGE_SHIFT) - 1;
2298 * Userspace could choose to issue a mb() before updating the tail
2299 * pointer. So that all reads will be completed before the write is
2302 tail = ACCESS_ONCE(data->user_page->data_tail);
2305 offset = (offset - tail) & mask;
2306 head = (head - tail) & mask;
2308 if ((int)(head - offset) < 0)
2314 static void perf_output_wakeup(struct perf_output_handle *handle)
2316 atomic_set(&handle->data->poll, POLL_IN);
2319 handle->counter->pending_wakeup = 1;
2320 perf_pending_queue(&handle->counter->pending,
2321 perf_pending_counter);
2323 perf_counter_wakeup(handle->counter);
2327 * Curious locking construct.
2329 * We need to ensure a later event doesn't publish a head when a former
2330 * event isn't done writing. However since we need to deal with NMIs we
2331 * cannot fully serialize things.
2333 * What we do is serialize between CPUs so we only have to deal with NMI
2334 * nesting on a single CPU.
2336 * We only publish the head (and generate a wakeup) when the outer-most
2339 static void perf_output_lock(struct perf_output_handle *handle)
2341 struct perf_mmap_data *data = handle->data;
2346 local_irq_save(handle->flags);
2347 cpu = smp_processor_id();
2349 if (in_nmi() && atomic_read(&data->lock) == cpu)
2352 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2358 static void perf_output_unlock(struct perf_output_handle *handle)
2360 struct perf_mmap_data *data = handle->data;
2364 data->done_head = data->head;
2366 if (!handle->locked)
2371 * The xchg implies a full barrier that ensures all writes are done
2372 * before we publish the new head, matched by a rmb() in userspace when
2373 * reading this position.
2375 while ((head = atomic_long_xchg(&data->done_head, 0)))
2376 data->user_page->data_head = head;
2379 * NMI can happen here, which means we can miss a done_head update.
2382 cpu = atomic_xchg(&data->lock, -1);
2383 WARN_ON_ONCE(cpu != smp_processor_id());
2386 * Therefore we have to validate we did not indeed do so.
2388 if (unlikely(atomic_long_read(&data->done_head))) {
2390 * Since we had it locked, we can lock it again.
2392 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2398 if (atomic_xchg(&data->wakeup, 0))
2399 perf_output_wakeup(handle);
2401 local_irq_restore(handle->flags);
2404 static void perf_output_copy(struct perf_output_handle *handle,
2405 const void *buf, unsigned int len)
2407 unsigned int pages_mask;
2408 unsigned int offset;
2412 offset = handle->offset;
2413 pages_mask = handle->data->nr_pages - 1;
2414 pages = handle->data->data_pages;
2417 unsigned int page_offset;
2420 nr = (offset >> PAGE_SHIFT) & pages_mask;
2421 page_offset = offset & (PAGE_SIZE - 1);
2422 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2424 memcpy(pages[nr] + page_offset, buf, size);
2431 handle->offset = offset;
2434 * Check we didn't copy past our reservation window, taking the
2435 * possible unsigned int wrap into account.
2437 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2440 #define perf_output_put(handle, x) \
2441 perf_output_copy((handle), &(x), sizeof(x))
2443 static int perf_output_begin(struct perf_output_handle *handle,
2444 struct perf_counter *counter, unsigned int size,
2445 int nmi, int sample)
2447 struct perf_mmap_data *data;
2448 unsigned int offset, head;
2451 struct perf_event_header header;
2457 * For inherited counters we send all the output towards the parent.
2459 if (counter->parent)
2460 counter = counter->parent;
2463 data = rcu_dereference(counter->data);
2467 handle->data = data;
2468 handle->counter = counter;
2470 handle->sample = sample;
2472 if (!data->nr_pages)
2475 have_lost = atomic_read(&data->lost);
2477 size += sizeof(lost_event);
2479 perf_output_lock(handle);
2482 offset = head = atomic_long_read(&data->head);
2484 if (unlikely(!perf_output_space(data, offset, head)))
2486 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2488 handle->offset = offset;
2489 handle->head = head;
2491 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2492 atomic_set(&data->wakeup, 1);
2495 lost_event.header.type = PERF_EVENT_LOST;
2496 lost_event.header.misc = 0;
2497 lost_event.header.size = sizeof(lost_event);
2498 lost_event.id = counter->id;
2499 lost_event.lost = atomic_xchg(&data->lost, 0);
2501 perf_output_put(handle, lost_event);
2507 atomic_inc(&data->lost);
2508 perf_output_unlock(handle);
2515 static void perf_output_end(struct perf_output_handle *handle)
2517 struct perf_counter *counter = handle->counter;
2518 struct perf_mmap_data *data = handle->data;
2520 int wakeup_events = counter->attr.wakeup_events;
2522 if (handle->sample && wakeup_events) {
2523 int events = atomic_inc_return(&data->events);
2524 if (events >= wakeup_events) {
2525 atomic_sub(wakeup_events, &data->events);
2526 atomic_set(&data->wakeup, 1);
2530 perf_output_unlock(handle);
2534 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2537 * only top level counters have the pid namespace they were created in
2539 if (counter->parent)
2540 counter = counter->parent;
2542 return task_tgid_nr_ns(p, counter->ns);
2545 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2548 * only top level counters have the pid namespace they were created in
2550 if (counter->parent)
2551 counter = counter->parent;
2553 return task_pid_nr_ns(p, counter->ns);
2556 static void perf_counter_output(struct perf_counter *counter, int nmi,
2557 struct perf_sample_data *data)
2560 u64 sample_type = counter->attr.sample_type;
2561 struct perf_output_handle handle;
2562 struct perf_event_header header;
2571 struct perf_callchain_entry *callchain = NULL;
2572 int callchain_size = 0;
2579 header.size = sizeof(header);
2581 header.misc = PERF_EVENT_MISC_OVERFLOW;
2582 header.misc |= perf_misc_flags(data->regs);
2584 if (sample_type & PERF_SAMPLE_IP) {
2585 ip = perf_instruction_pointer(data->regs);
2586 header.type |= PERF_SAMPLE_IP;
2587 header.size += sizeof(ip);
2590 if (sample_type & PERF_SAMPLE_TID) {
2591 /* namespace issues */
2592 tid_entry.pid = perf_counter_pid(counter, current);
2593 tid_entry.tid = perf_counter_tid(counter, current);
2595 header.type |= PERF_SAMPLE_TID;
2596 header.size += sizeof(tid_entry);
2599 if (sample_type & PERF_SAMPLE_TIME) {
2601 * Maybe do better on x86 and provide cpu_clock_nmi()
2603 time = sched_clock();
2605 header.type |= PERF_SAMPLE_TIME;
2606 header.size += sizeof(u64);
2609 if (sample_type & PERF_SAMPLE_ADDR) {
2610 header.type |= PERF_SAMPLE_ADDR;
2611 header.size += sizeof(u64);
2614 if (sample_type & PERF_SAMPLE_ID) {
2615 header.type |= PERF_SAMPLE_ID;
2616 header.size += sizeof(u64);
2619 if (sample_type & PERF_SAMPLE_CPU) {
2620 header.type |= PERF_SAMPLE_CPU;
2621 header.size += sizeof(cpu_entry);
2623 cpu_entry.cpu = raw_smp_processor_id();
2626 if (sample_type & PERF_SAMPLE_PERIOD) {
2627 header.type |= PERF_SAMPLE_PERIOD;
2628 header.size += sizeof(u64);
2631 if (sample_type & PERF_SAMPLE_GROUP) {
2632 header.type |= PERF_SAMPLE_GROUP;
2633 header.size += sizeof(u64) +
2634 counter->nr_siblings * sizeof(group_entry);
2637 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2638 callchain = perf_callchain(data->regs);
2641 callchain_size = (1 + callchain->nr) * sizeof(u64);
2643 header.type |= PERF_SAMPLE_CALLCHAIN;
2644 header.size += callchain_size;
2648 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2652 perf_output_put(&handle, header);
2654 if (sample_type & PERF_SAMPLE_IP)
2655 perf_output_put(&handle, ip);
2657 if (sample_type & PERF_SAMPLE_TID)
2658 perf_output_put(&handle, tid_entry);
2660 if (sample_type & PERF_SAMPLE_TIME)
2661 perf_output_put(&handle, time);
2663 if (sample_type & PERF_SAMPLE_ADDR)
2664 perf_output_put(&handle, data->addr);
2666 if (sample_type & PERF_SAMPLE_ID)
2667 perf_output_put(&handle, counter->id);
2669 if (sample_type & PERF_SAMPLE_CPU)
2670 perf_output_put(&handle, cpu_entry);
2672 if (sample_type & PERF_SAMPLE_PERIOD)
2673 perf_output_put(&handle, data->period);
2676 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2678 if (sample_type & PERF_SAMPLE_GROUP) {
2679 struct perf_counter *leader, *sub;
2680 u64 nr = counter->nr_siblings;
2682 perf_output_put(&handle, nr);
2684 leader = counter->group_leader;
2685 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2687 sub->pmu->read(sub);
2689 group_entry.id = sub->id;
2690 group_entry.counter = atomic64_read(&sub->count);
2692 perf_output_put(&handle, group_entry);
2697 perf_output_copy(&handle, callchain, callchain_size);
2699 perf_output_end(&handle);
2706 struct perf_read_event {
2707 struct perf_event_header header;
2716 perf_counter_read_event(struct perf_counter *counter,
2717 struct task_struct *task)
2719 struct perf_output_handle handle;
2720 struct perf_read_event event = {
2722 .type = PERF_EVENT_READ,
2724 .size = sizeof(event) - sizeof(event.format),
2726 .pid = perf_counter_pid(counter, task),
2727 .tid = perf_counter_tid(counter, task),
2728 .value = atomic64_read(&counter->count),
2732 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2733 event.header.size += sizeof(u64);
2734 event.format[i++] = counter->total_time_enabled;
2737 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2738 event.header.size += sizeof(u64);
2739 event.format[i++] = counter->total_time_running;
2742 if (counter->attr.read_format & PERF_FORMAT_ID) {
2745 event.header.size += sizeof(u64);
2746 if (counter->parent)
2747 id = counter->parent->id;
2751 event.format[i++] = id;
2754 ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
2758 perf_output_copy(&handle, &event, event.header.size);
2759 perf_output_end(&handle);
2766 struct perf_fork_event {
2767 struct task_struct *task;
2770 struct perf_event_header header;
2777 static void perf_counter_fork_output(struct perf_counter *counter,
2778 struct perf_fork_event *fork_event)
2780 struct perf_output_handle handle;
2781 int size = fork_event->event.header.size;
2782 struct task_struct *task = fork_event->task;
2783 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2788 fork_event->event.pid = perf_counter_pid(counter, task);
2789 fork_event->event.ppid = perf_counter_pid(counter, task->real_parent);
2791 perf_output_put(&handle, fork_event->event);
2792 perf_output_end(&handle);
2795 static int perf_counter_fork_match(struct perf_counter *counter)
2797 if (counter->attr.comm || counter->attr.mmap)
2803 static void perf_counter_fork_ctx(struct perf_counter_context *ctx,
2804 struct perf_fork_event *fork_event)
2806 struct perf_counter *counter;
2808 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2812 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2813 if (perf_counter_fork_match(counter))
2814 perf_counter_fork_output(counter, fork_event);
2819 static void perf_counter_fork_event(struct perf_fork_event *fork_event)
2821 struct perf_cpu_context *cpuctx;
2822 struct perf_counter_context *ctx;
2824 cpuctx = &get_cpu_var(perf_cpu_context);
2825 perf_counter_fork_ctx(&cpuctx->ctx, fork_event);
2826 put_cpu_var(perf_cpu_context);
2830 * doesn't really matter which of the child contexts the
2831 * events ends up in.
2833 ctx = rcu_dereference(current->perf_counter_ctxp);
2835 perf_counter_fork_ctx(ctx, fork_event);
2839 void perf_counter_fork(struct task_struct *task)
2841 struct perf_fork_event fork_event;
2843 if (!atomic_read(&nr_comm_counters) &&
2844 !atomic_read(&nr_mmap_counters))
2847 fork_event = (struct perf_fork_event){
2851 .type = PERF_EVENT_FORK,
2852 .size = sizeof(fork_event.event),
2857 perf_counter_fork_event(&fork_event);
2864 struct perf_comm_event {
2865 struct task_struct *task;
2870 struct perf_event_header header;
2877 static void perf_counter_comm_output(struct perf_counter *counter,
2878 struct perf_comm_event *comm_event)
2880 struct perf_output_handle handle;
2881 int size = comm_event->event.header.size;
2882 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2887 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
2888 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
2890 perf_output_put(&handle, comm_event->event);
2891 perf_output_copy(&handle, comm_event->comm,
2892 comm_event->comm_size);
2893 perf_output_end(&handle);
2896 static int perf_counter_comm_match(struct perf_counter *counter)
2898 if (counter->attr.comm)
2904 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2905 struct perf_comm_event *comm_event)
2907 struct perf_counter *counter;
2909 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2913 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2914 if (perf_counter_comm_match(counter))
2915 perf_counter_comm_output(counter, comm_event);
2920 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2922 struct perf_cpu_context *cpuctx;
2923 struct perf_counter_context *ctx;
2925 char *comm = comm_event->task->comm;
2927 size = ALIGN(strlen(comm)+1, sizeof(u64));
2929 comm_event->comm = comm;
2930 comm_event->comm_size = size;
2932 comm_event->event.header.size = sizeof(comm_event->event) + size;
2934 cpuctx = &get_cpu_var(perf_cpu_context);
2935 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2936 put_cpu_var(perf_cpu_context);
2940 * doesn't really matter which of the child contexts the
2941 * events ends up in.
2943 ctx = rcu_dereference(current->perf_counter_ctxp);
2945 perf_counter_comm_ctx(ctx, comm_event);
2949 void perf_counter_comm(struct task_struct *task)
2951 struct perf_comm_event comm_event;
2953 if (!atomic_read(&nr_comm_counters))
2956 comm_event = (struct perf_comm_event){
2959 .header = { .type = PERF_EVENT_COMM, },
2963 perf_counter_comm_event(&comm_event);
2970 struct perf_mmap_event {
2971 struct vm_area_struct *vma;
2973 const char *file_name;
2977 struct perf_event_header header;
2987 static void perf_counter_mmap_output(struct perf_counter *counter,
2988 struct perf_mmap_event *mmap_event)
2990 struct perf_output_handle handle;
2991 int size = mmap_event->event.header.size;
2992 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2997 mmap_event->event.pid = perf_counter_pid(counter, current);
2998 mmap_event->event.tid = perf_counter_tid(counter, current);
3000 perf_output_put(&handle, mmap_event->event);
3001 perf_output_copy(&handle, mmap_event->file_name,
3002 mmap_event->file_size);
3003 perf_output_end(&handle);
3006 static int perf_counter_mmap_match(struct perf_counter *counter,
3007 struct perf_mmap_event *mmap_event)
3009 if (counter->attr.mmap)
3015 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3016 struct perf_mmap_event *mmap_event)
3018 struct perf_counter *counter;
3020 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3024 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3025 if (perf_counter_mmap_match(counter, mmap_event))
3026 perf_counter_mmap_output(counter, mmap_event);
3031 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3033 struct perf_cpu_context *cpuctx;
3034 struct perf_counter_context *ctx;
3035 struct vm_area_struct *vma = mmap_event->vma;
3036 struct file *file = vma->vm_file;
3043 buf = kzalloc(PATH_MAX, GFP_KERNEL);
3045 name = strncpy(tmp, "//enomem", sizeof(tmp));
3048 name = d_path(&file->f_path, buf, PATH_MAX);
3050 name = strncpy(tmp, "//toolong", sizeof(tmp));
3054 name = arch_vma_name(mmap_event->vma);
3059 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3063 name = strncpy(tmp, "//anon", sizeof(tmp));
3068 size = ALIGN(strlen(name)+1, sizeof(u64));
3070 mmap_event->file_name = name;
3071 mmap_event->file_size = size;
3073 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3075 cpuctx = &get_cpu_var(perf_cpu_context);
3076 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3077 put_cpu_var(perf_cpu_context);
3081 * doesn't really matter which of the child contexts the
3082 * events ends up in.
3084 ctx = rcu_dereference(current->perf_counter_ctxp);
3086 perf_counter_mmap_ctx(ctx, mmap_event);
3092 void __perf_counter_mmap(struct vm_area_struct *vma)
3094 struct perf_mmap_event mmap_event;
3096 if (!atomic_read(&nr_mmap_counters))
3099 mmap_event = (struct perf_mmap_event){
3102 .header = { .type = PERF_EVENT_MMAP, },
3103 .start = vma->vm_start,
3104 .len = vma->vm_end - vma->vm_start,
3105 .pgoff = vma->vm_pgoff,
3109 perf_counter_mmap_event(&mmap_event);
3113 * Log sample_period changes so that analyzing tools can re-normalize the
3118 struct perf_event_header header;
3124 static void perf_log_period(struct perf_counter *counter, u64 period)
3126 struct perf_output_handle handle;
3127 struct freq_event event;
3130 if (counter->hw.sample_period == period)
3133 if (counter->attr.sample_type & PERF_SAMPLE_PERIOD)
3136 event = (struct freq_event) {
3138 .type = PERF_EVENT_PERIOD,
3140 .size = sizeof(event),
3142 .time = sched_clock(),
3147 ret = perf_output_begin(&handle, counter, sizeof(event), 1, 0);
3151 perf_output_put(&handle, event);
3152 perf_output_end(&handle);
3156 * IRQ throttle logging
3159 static void perf_log_throttle(struct perf_counter *counter, int enable)
3161 struct perf_output_handle handle;
3165 struct perf_event_header header;
3168 } throttle_event = {
3170 .type = PERF_EVENT_THROTTLE + 1,
3172 .size = sizeof(throttle_event),
3174 .time = sched_clock(),
3178 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3182 perf_output_put(&handle, throttle_event);
3183 perf_output_end(&handle);
3187 * Generic counter overflow handling, sampling.
3190 int perf_counter_overflow(struct perf_counter *counter, int nmi,
3191 struct perf_sample_data *data)
3193 int events = atomic_read(&counter->event_limit);
3194 int throttle = counter->pmu->unthrottle != NULL;
3195 struct hw_perf_counter *hwc = &counter->hw;
3201 if (hwc->interrupts != MAX_INTERRUPTS) {
3203 if (HZ * hwc->interrupts >
3204 (u64)sysctl_perf_counter_sample_rate) {
3205 hwc->interrupts = MAX_INTERRUPTS;
3206 perf_log_throttle(counter, 0);
3211 * Keep re-disabling counters even though on the previous
3212 * pass we disabled it - just in case we raced with a
3213 * sched-in and the counter got enabled again:
3219 if (counter->attr.freq) {
3220 u64 now = sched_clock();
3221 s64 delta = now - hwc->freq_stamp;
3223 hwc->freq_stamp = now;
3225 if (delta > 0 && delta < TICK_NSEC)
3226 perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3230 * XXX event_limit might not quite work as expected on inherited
3234 counter->pending_kill = POLL_IN;
3235 if (events && atomic_dec_and_test(&counter->event_limit)) {
3237 counter->pending_kill = POLL_HUP;
3239 counter->pending_disable = 1;
3240 perf_pending_queue(&counter->pending,
3241 perf_pending_counter);
3243 perf_counter_disable(counter);
3246 perf_counter_output(counter, nmi, data);
3251 * Generic software counter infrastructure
3254 static void perf_swcounter_update(struct perf_counter *counter)
3256 struct hw_perf_counter *hwc = &counter->hw;
3261 prev = atomic64_read(&hwc->prev_count);
3262 now = atomic64_read(&hwc->count);
3263 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
3268 atomic64_add(delta, &counter->count);
3269 atomic64_sub(delta, &hwc->period_left);
3272 static void perf_swcounter_set_period(struct perf_counter *counter)
3274 struct hw_perf_counter *hwc = &counter->hw;
3275 s64 left = atomic64_read(&hwc->period_left);
3276 s64 period = hwc->sample_period;
3278 if (unlikely(left <= -period)) {
3280 atomic64_set(&hwc->period_left, left);
3281 hwc->last_period = period;
3284 if (unlikely(left <= 0)) {
3286 atomic64_add(period, &hwc->period_left);
3287 hwc->last_period = period;
3290 atomic64_set(&hwc->prev_count, -left);
3291 atomic64_set(&hwc->count, -left);
3294 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3296 enum hrtimer_restart ret = HRTIMER_RESTART;
3297 struct perf_sample_data data;
3298 struct perf_counter *counter;
3301 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3302 counter->pmu->read(counter);
3305 data.regs = get_irq_regs();
3307 * In case we exclude kernel IPs or are somehow not in interrupt
3308 * context, provide the next best thing, the user IP.
3310 if ((counter->attr.exclude_kernel || !data.regs) &&
3311 !counter->attr.exclude_user)
3312 data.regs = task_pt_regs(current);
3315 if (perf_counter_overflow(counter, 0, &data))
3316 ret = HRTIMER_NORESTART;
3319 period = max_t(u64, 10000, counter->hw.sample_period);
3320 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3325 static void perf_swcounter_overflow(struct perf_counter *counter,
3326 int nmi, struct perf_sample_data *data)
3328 data->period = counter->hw.last_period;
3330 perf_swcounter_update(counter);
3331 perf_swcounter_set_period(counter);
3332 if (perf_counter_overflow(counter, nmi, data))
3333 /* soft-disable the counter */
3337 static int perf_swcounter_is_counting(struct perf_counter *counter)
3339 struct perf_counter_context *ctx;
3340 unsigned long flags;
3343 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3346 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3350 * If the counter is inactive, it could be just because
3351 * its task is scheduled out, or because it's in a group
3352 * which could not go on the PMU. We want to count in
3353 * the first case but not the second. If the context is
3354 * currently active then an inactive software counter must
3355 * be the second case. If it's not currently active then
3356 * we need to know whether the counter was active when the
3357 * context was last active, which we can determine by
3358 * comparing counter->tstamp_stopped with ctx->time.
3360 * We are within an RCU read-side critical section,
3361 * which protects the existence of *ctx.
3364 spin_lock_irqsave(&ctx->lock, flags);
3366 /* Re-check state now we have the lock */
3367 if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
3368 counter->ctx->is_active ||
3369 counter->tstamp_stopped < ctx->time)
3371 spin_unlock_irqrestore(&ctx->lock, flags);
3375 static int perf_swcounter_match(struct perf_counter *counter,
3376 enum perf_type_id type,
3377 u32 event, struct pt_regs *regs)
3379 if (!perf_swcounter_is_counting(counter))
3382 if (counter->attr.type != type)
3384 if (counter->attr.config != event)
3388 if (counter->attr.exclude_user && user_mode(regs))
3391 if (counter->attr.exclude_kernel && !user_mode(regs))
3398 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3399 int nmi, struct perf_sample_data *data)
3401 int neg = atomic64_add_negative(nr, &counter->hw.count);
3403 if (counter->hw.sample_period && !neg && data->regs)
3404 perf_swcounter_overflow(counter, nmi, data);
3407 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3408 enum perf_type_id type,
3409 u32 event, u64 nr, int nmi,
3410 struct perf_sample_data *data)
3412 struct perf_counter *counter;
3414 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3418 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3419 if (perf_swcounter_match(counter, type, event, data->regs))
3420 perf_swcounter_add(counter, nr, nmi, data);
3425 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3428 return &cpuctx->recursion[3];
3431 return &cpuctx->recursion[2];
3434 return &cpuctx->recursion[1];
3436 return &cpuctx->recursion[0];
3439 static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3441 struct perf_sample_data *data)
3443 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3444 int *recursion = perf_swcounter_recursion_context(cpuctx);
3445 struct perf_counter_context *ctx;
3453 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3457 * doesn't really matter which of the child contexts the
3458 * events ends up in.
3460 ctx = rcu_dereference(current->perf_counter_ctxp);
3462 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3469 put_cpu_var(perf_cpu_context);
3472 void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3473 struct pt_regs *regs, u64 addr)
3475 struct perf_sample_data data = {
3480 do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3483 static void perf_swcounter_read(struct perf_counter *counter)
3485 perf_swcounter_update(counter);
3488 static int perf_swcounter_enable(struct perf_counter *counter)
3490 perf_swcounter_set_period(counter);
3494 static void perf_swcounter_disable(struct perf_counter *counter)
3496 perf_swcounter_update(counter);
3499 static const struct pmu perf_ops_generic = {
3500 .enable = perf_swcounter_enable,
3501 .disable = perf_swcounter_disable,
3502 .read = perf_swcounter_read,
3506 * Software counter: cpu wall time clock
3509 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3511 int cpu = raw_smp_processor_id();
3515 now = cpu_clock(cpu);
3516 prev = atomic64_read(&counter->hw.prev_count);
3517 atomic64_set(&counter->hw.prev_count, now);
3518 atomic64_add(now - prev, &counter->count);
3521 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3523 struct hw_perf_counter *hwc = &counter->hw;
3524 int cpu = raw_smp_processor_id();
3526 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3527 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3528 hwc->hrtimer.function = perf_swcounter_hrtimer;
3529 if (hwc->sample_period) {
3530 u64 period = max_t(u64, 10000, hwc->sample_period);
3531 __hrtimer_start_range_ns(&hwc->hrtimer,
3532 ns_to_ktime(period), 0,
3533 HRTIMER_MODE_REL, 0);
3539 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3541 if (counter->hw.sample_period)
3542 hrtimer_cancel(&counter->hw.hrtimer);
3543 cpu_clock_perf_counter_update(counter);
3546 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3548 cpu_clock_perf_counter_update(counter);
3551 static const struct pmu perf_ops_cpu_clock = {
3552 .enable = cpu_clock_perf_counter_enable,
3553 .disable = cpu_clock_perf_counter_disable,
3554 .read = cpu_clock_perf_counter_read,
3558 * Software counter: task time clock
3561 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3566 prev = atomic64_xchg(&counter->hw.prev_count, now);
3568 atomic64_add(delta, &counter->count);
3571 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3573 struct hw_perf_counter *hwc = &counter->hw;
3576 now = counter->ctx->time;
3578 atomic64_set(&hwc->prev_count, now);
3579 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3580 hwc->hrtimer.function = perf_swcounter_hrtimer;
3581 if (hwc->sample_period) {
3582 u64 period = max_t(u64, 10000, hwc->sample_period);
3583 __hrtimer_start_range_ns(&hwc->hrtimer,
3584 ns_to_ktime(period), 0,
3585 HRTIMER_MODE_REL, 0);
3591 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3593 if (counter->hw.sample_period)
3594 hrtimer_cancel(&counter->hw.hrtimer);
3595 task_clock_perf_counter_update(counter, counter->ctx->time);
3599 static void task_clock_perf_counter_read(struct perf_counter *counter)
3604 update_context_time(counter->ctx);
3605 time = counter->ctx->time;
3607 u64 now = perf_clock();
3608 u64 delta = now - counter->ctx->timestamp;
3609 time = counter->ctx->time + delta;
3612 task_clock_perf_counter_update(counter, time);
3615 static const struct pmu perf_ops_task_clock = {
3616 .enable = task_clock_perf_counter_enable,
3617 .disable = task_clock_perf_counter_disable,
3618 .read = task_clock_perf_counter_read,
3621 #ifdef CONFIG_EVENT_PROFILE
3622 void perf_tpcounter_event(int event_id)
3624 struct perf_sample_data data = {
3625 .regs = get_irq_regs();
3630 data.regs = task_pt_regs(current);
3632 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, &data);
3634 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3636 extern int ftrace_profile_enable(int);
3637 extern void ftrace_profile_disable(int);
3639 static void tp_perf_counter_destroy(struct perf_counter *counter)
3641 ftrace_profile_disable(perf_event_id(&counter->attr));
3644 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3646 int event_id = perf_event_id(&counter->attr);
3649 ret = ftrace_profile_enable(event_id);
3653 counter->destroy = tp_perf_counter_destroy;
3655 return &perf_ops_generic;
3658 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3664 atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3666 static void sw_perf_counter_destroy(struct perf_counter *counter)
3668 u64 event = counter->attr.config;
3670 WARN_ON(counter->parent);
3672 atomic_dec(&perf_swcounter_enabled[event]);
3675 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3677 const struct pmu *pmu = NULL;
3678 u64 event = counter->attr.config;
3681 * Software counters (currently) can't in general distinguish
3682 * between user, kernel and hypervisor events.
3683 * However, context switches and cpu migrations are considered
3684 * to be kernel events, and page faults are never hypervisor
3688 case PERF_COUNT_SW_CPU_CLOCK:
3689 pmu = &perf_ops_cpu_clock;
3692 case PERF_COUNT_SW_TASK_CLOCK:
3694 * If the user instantiates this as a per-cpu counter,
3695 * use the cpu_clock counter instead.
3697 if (counter->ctx->task)
3698 pmu = &perf_ops_task_clock;
3700 pmu = &perf_ops_cpu_clock;
3703 case PERF_COUNT_SW_PAGE_FAULTS:
3704 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
3705 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
3706 case PERF_COUNT_SW_CONTEXT_SWITCHES:
3707 case PERF_COUNT_SW_CPU_MIGRATIONS:
3708 if (!counter->parent) {
3709 atomic_inc(&perf_swcounter_enabled[event]);
3710 counter->destroy = sw_perf_counter_destroy;
3712 pmu = &perf_ops_generic;
3720 * Allocate and initialize a counter structure
3722 static struct perf_counter *
3723 perf_counter_alloc(struct perf_counter_attr *attr,
3725 struct perf_counter_context *ctx,
3726 struct perf_counter *group_leader,
3727 struct perf_counter *parent_counter,
3730 const struct pmu *pmu;
3731 struct perf_counter *counter;
3732 struct hw_perf_counter *hwc;
3735 counter = kzalloc(sizeof(*counter), gfpflags);
3737 return ERR_PTR(-ENOMEM);
3740 * Single counters are their own group leaders, with an
3741 * empty sibling list:
3744 group_leader = counter;
3746 mutex_init(&counter->child_mutex);
3747 INIT_LIST_HEAD(&counter->child_list);
3749 INIT_LIST_HEAD(&counter->list_entry);
3750 INIT_LIST_HEAD(&counter->event_entry);
3751 INIT_LIST_HEAD(&counter->sibling_list);
3752 init_waitqueue_head(&counter->waitq);
3754 mutex_init(&counter->mmap_mutex);
3757 counter->attr = *attr;
3758 counter->group_leader = group_leader;
3759 counter->pmu = NULL;
3761 counter->oncpu = -1;
3763 counter->parent = parent_counter;
3765 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
3766 counter->id = atomic64_inc_return(&perf_counter_id);
3768 counter->state = PERF_COUNTER_STATE_INACTIVE;
3771 counter->state = PERF_COUNTER_STATE_OFF;
3776 hwc->sample_period = attr->sample_period;
3777 if (attr->freq && attr->sample_freq)
3778 hwc->sample_period = 1;
3780 atomic64_set(&hwc->period_left, hwc->sample_period);
3783 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3785 if (attr->inherit && (attr->sample_type & PERF_SAMPLE_GROUP))
3788 switch (attr->type) {
3790 case PERF_TYPE_HARDWARE:
3791 case PERF_TYPE_HW_CACHE:
3792 pmu = hw_perf_counter_init(counter);
3795 case PERF_TYPE_SOFTWARE:
3796 pmu = sw_perf_counter_init(counter);
3799 case PERF_TYPE_TRACEPOINT:
3800 pmu = tp_perf_counter_init(counter);
3810 else if (IS_ERR(pmu))
3815 put_pid_ns(counter->ns);
3817 return ERR_PTR(err);
3822 if (!counter->parent) {
3823 atomic_inc(&nr_counters);
3824 if (counter->attr.mmap)
3825 atomic_inc(&nr_mmap_counters);
3826 if (counter->attr.comm)
3827 atomic_inc(&nr_comm_counters);
3833 static int perf_copy_attr(struct perf_counter_attr __user *uattr,
3834 struct perf_counter_attr *attr)
3839 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
3843 * zero the full structure, so that a short copy will be nice.
3845 memset(attr, 0, sizeof(*attr));
3847 ret = get_user(size, &uattr->size);
3851 if (size > PAGE_SIZE) /* silly large */
3854 if (!size) /* abi compat */
3855 size = PERF_ATTR_SIZE_VER0;
3857 if (size < PERF_ATTR_SIZE_VER0)
3861 * If we're handed a bigger struct than we know of,
3862 * ensure all the unknown bits are 0.
3864 if (size > sizeof(*attr)) {
3866 unsigned long __user *addr;
3867 unsigned long __user *end;
3869 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
3870 sizeof(unsigned long));
3871 end = PTR_ALIGN((void __user *)uattr + size,
3872 sizeof(unsigned long));
3874 for (; addr < end; addr += sizeof(unsigned long)) {
3875 ret = get_user(val, addr);
3883 ret = copy_from_user(attr, uattr, size);
3888 * If the type exists, the corresponding creation will verify
3891 if (attr->type >= PERF_TYPE_MAX)
3894 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
3897 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
3900 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
3907 put_user(sizeof(*attr), &uattr->size);
3913 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3915 * @attr_uptr: event type attributes for monitoring/sampling
3918 * @group_fd: group leader counter fd
3920 SYSCALL_DEFINE5(perf_counter_open,
3921 struct perf_counter_attr __user *, attr_uptr,
3922 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3924 struct perf_counter *counter, *group_leader;
3925 struct perf_counter_attr attr;
3926 struct perf_counter_context *ctx;
3927 struct file *counter_file = NULL;
3928 struct file *group_file = NULL;
3929 int fput_needed = 0;
3930 int fput_needed2 = 0;
3933 /* for future expandability... */
3937 ret = perf_copy_attr(attr_uptr, &attr);
3941 if (!attr.exclude_kernel) {
3942 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
3947 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
3952 * Get the target context (task or percpu):
3954 ctx = find_get_context(pid, cpu);
3956 return PTR_ERR(ctx);
3959 * Look up the group leader (we will attach this counter to it):
3961 group_leader = NULL;
3962 if (group_fd != -1) {
3964 group_file = fget_light(group_fd, &fput_needed);
3966 goto err_put_context;
3967 if (group_file->f_op != &perf_fops)
3968 goto err_put_context;
3970 group_leader = group_file->private_data;
3972 * Do not allow a recursive hierarchy (this new sibling
3973 * becoming part of another group-sibling):
3975 if (group_leader->group_leader != group_leader)
3976 goto err_put_context;
3978 * Do not allow to attach to a group in a different
3979 * task or CPU context:
3981 if (group_leader->ctx != ctx)
3982 goto err_put_context;
3984 * Only a group leader can be exclusive or pinned
3986 if (attr.exclusive || attr.pinned)
3987 goto err_put_context;
3990 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
3992 ret = PTR_ERR(counter);
3993 if (IS_ERR(counter))
3994 goto err_put_context;
3996 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3998 goto err_free_put_context;
4000 counter_file = fget_light(ret, &fput_needed2);
4002 goto err_free_put_context;
4004 counter->filp = counter_file;
4005 WARN_ON_ONCE(ctx->parent_ctx);
4006 mutex_lock(&ctx->mutex);
4007 perf_install_in_context(ctx, counter, cpu);
4009 mutex_unlock(&ctx->mutex);
4011 counter->owner = current;
4012 get_task_struct(current);
4013 mutex_lock(¤t->perf_counter_mutex);
4014 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
4015 mutex_unlock(¤t->perf_counter_mutex);
4017 fput_light(counter_file, fput_needed2);
4020 fput_light(group_file, fput_needed);
4024 err_free_put_context:
4034 * inherit a counter from parent task to child task:
4036 static struct perf_counter *
4037 inherit_counter(struct perf_counter *parent_counter,
4038 struct task_struct *parent,
4039 struct perf_counter_context *parent_ctx,
4040 struct task_struct *child,
4041 struct perf_counter *group_leader,
4042 struct perf_counter_context *child_ctx)
4044 struct perf_counter *child_counter;
4047 * Instead of creating recursive hierarchies of counters,
4048 * we link inherited counters back to the original parent,
4049 * which has a filp for sure, which we use as the reference
4052 if (parent_counter->parent)
4053 parent_counter = parent_counter->parent;
4055 child_counter = perf_counter_alloc(&parent_counter->attr,
4056 parent_counter->cpu, child_ctx,
4057 group_leader, parent_counter,
4059 if (IS_ERR(child_counter))
4060 return child_counter;
4064 * Make the child state follow the state of the parent counter,
4065 * not its attr.disabled bit. We hold the parent's mutex,
4066 * so we won't race with perf_counter_{en, dis}able_family.
4068 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4069 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4071 child_counter->state = PERF_COUNTER_STATE_OFF;
4073 if (parent_counter->attr.freq)
4074 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4077 * Link it up in the child's context:
4079 add_counter_to_ctx(child_counter, child_ctx);
4082 * Get a reference to the parent filp - we will fput it
4083 * when the child counter exits. This is safe to do because
4084 * we are in the parent and we know that the filp still
4085 * exists and has a nonzero count:
4087 atomic_long_inc(&parent_counter->filp->f_count);
4090 * Link this into the parent counter's child list
4092 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4093 mutex_lock(&parent_counter->child_mutex);
4094 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4095 mutex_unlock(&parent_counter->child_mutex);
4097 return child_counter;
4100 static int inherit_group(struct perf_counter *parent_counter,
4101 struct task_struct *parent,
4102 struct perf_counter_context *parent_ctx,
4103 struct task_struct *child,
4104 struct perf_counter_context *child_ctx)
4106 struct perf_counter *leader;
4107 struct perf_counter *sub;
4108 struct perf_counter *child_ctr;
4110 leader = inherit_counter(parent_counter, parent, parent_ctx,
4111 child, NULL, child_ctx);
4113 return PTR_ERR(leader);
4114 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4115 child_ctr = inherit_counter(sub, parent, parent_ctx,
4116 child, leader, child_ctx);
4117 if (IS_ERR(child_ctr))
4118 return PTR_ERR(child_ctr);
4123 static void sync_child_counter(struct perf_counter *child_counter,
4124 struct task_struct *child)
4126 struct perf_counter *parent_counter = child_counter->parent;
4129 if (child_counter->attr.inherit_stat)
4130 perf_counter_read_event(child_counter, child);
4132 child_val = atomic64_read(&child_counter->count);
4135 * Add back the child's count to the parent's count:
4137 atomic64_add(child_val, &parent_counter->count);
4138 atomic64_add(child_counter->total_time_enabled,
4139 &parent_counter->child_total_time_enabled);
4140 atomic64_add(child_counter->total_time_running,
4141 &parent_counter->child_total_time_running);
4144 * Remove this counter from the parent's list
4146 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4147 mutex_lock(&parent_counter->child_mutex);
4148 list_del_init(&child_counter->child_list);
4149 mutex_unlock(&parent_counter->child_mutex);
4152 * Release the parent counter, if this was the last
4155 fput(parent_counter->filp);
4159 __perf_counter_exit_task(struct perf_counter *child_counter,
4160 struct perf_counter_context *child_ctx,
4161 struct task_struct *child)
4163 struct perf_counter *parent_counter;
4165 update_counter_times(child_counter);
4166 perf_counter_remove_from_context(child_counter);
4168 parent_counter = child_counter->parent;
4170 * It can happen that parent exits first, and has counters
4171 * that are still around due to the child reference. These
4172 * counters need to be zapped - but otherwise linger.
4174 if (parent_counter) {
4175 sync_child_counter(child_counter, child);
4176 free_counter(child_counter);
4181 * When a child task exits, feed back counter values to parent counters.
4183 void perf_counter_exit_task(struct task_struct *child)
4185 struct perf_counter *child_counter, *tmp;
4186 struct perf_counter_context *child_ctx;
4187 unsigned long flags;
4189 if (likely(!child->perf_counter_ctxp))
4192 local_irq_save(flags);
4194 * We can't reschedule here because interrupts are disabled,
4195 * and either child is current or it is a task that can't be
4196 * scheduled, so we are now safe from rescheduling changing
4199 child_ctx = child->perf_counter_ctxp;
4200 __perf_counter_task_sched_out(child_ctx);
4203 * Take the context lock here so that if find_get_context is
4204 * reading child->perf_counter_ctxp, we wait until it has
4205 * incremented the context's refcount before we do put_ctx below.
4207 spin_lock(&child_ctx->lock);
4208 child->perf_counter_ctxp = NULL;
4209 if (child_ctx->parent_ctx) {
4211 * This context is a clone; unclone it so it can't get
4212 * swapped to another process while we're removing all
4213 * the counters from it.
4215 put_ctx(child_ctx->parent_ctx);
4216 child_ctx->parent_ctx = NULL;
4218 spin_unlock(&child_ctx->lock);
4219 local_irq_restore(flags);
4222 * We can recurse on the same lock type through:
4224 * __perf_counter_exit_task()
4225 * sync_child_counter()
4226 * fput(parent_counter->filp)
4228 * mutex_lock(&ctx->mutex)
4230 * But since its the parent context it won't be the same instance.
4232 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4235 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4237 __perf_counter_exit_task(child_counter, child_ctx, child);
4240 * If the last counter was a group counter, it will have appended all
4241 * its siblings to the list, but we obtained 'tmp' before that which
4242 * will still point to the list head terminating the iteration.
4244 if (!list_empty(&child_ctx->counter_list))
4247 mutex_unlock(&child_ctx->mutex);
4253 * free an unexposed, unused context as created by inheritance by
4254 * init_task below, used by fork() in case of fail.
4256 void perf_counter_free_task(struct task_struct *task)
4258 struct perf_counter_context *ctx = task->perf_counter_ctxp;
4259 struct perf_counter *counter, *tmp;
4264 mutex_lock(&ctx->mutex);
4266 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4267 struct perf_counter *parent = counter->parent;
4269 if (WARN_ON_ONCE(!parent))
4272 mutex_lock(&parent->child_mutex);
4273 list_del_init(&counter->child_list);
4274 mutex_unlock(&parent->child_mutex);
4278 list_del_counter(counter, ctx);
4279 free_counter(counter);
4282 if (!list_empty(&ctx->counter_list))
4285 mutex_unlock(&ctx->mutex);
4291 * Initialize the perf_counter context in task_struct
4293 int perf_counter_init_task(struct task_struct *child)
4295 struct perf_counter_context *child_ctx, *parent_ctx;
4296 struct perf_counter_context *cloned_ctx;
4297 struct perf_counter *counter;
4298 struct task_struct *parent = current;
4299 int inherited_all = 1;
4302 child->perf_counter_ctxp = NULL;
4304 mutex_init(&child->perf_counter_mutex);
4305 INIT_LIST_HEAD(&child->perf_counter_list);
4307 if (likely(!parent->perf_counter_ctxp))
4311 * This is executed from the parent task context, so inherit
4312 * counters that have been marked for cloning.
4313 * First allocate and initialize a context for the child.
4316 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4320 __perf_counter_init_context(child_ctx, child);
4321 child->perf_counter_ctxp = child_ctx;
4322 get_task_struct(child);
4325 * If the parent's context is a clone, pin it so it won't get
4328 parent_ctx = perf_pin_task_context(parent);
4331 * No need to check if parent_ctx != NULL here; since we saw
4332 * it non-NULL earlier, the only reason for it to become NULL
4333 * is if we exit, and since we're currently in the middle of
4334 * a fork we can't be exiting at the same time.
4338 * Lock the parent list. No need to lock the child - not PID
4339 * hashed yet and not running, so nobody can access it.
4341 mutex_lock(&parent_ctx->mutex);
4344 * We dont have to disable NMIs - we are only looking at
4345 * the list, not manipulating it:
4347 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4348 if (counter != counter->group_leader)
4351 if (!counter->attr.inherit) {
4356 ret = inherit_group(counter, parent, parent_ctx,
4364 if (inherited_all) {
4366 * Mark the child context as a clone of the parent
4367 * context, or of whatever the parent is a clone of.
4368 * Note that if the parent is a clone, it could get
4369 * uncloned at any point, but that doesn't matter
4370 * because the list of counters and the generation
4371 * count can't have changed since we took the mutex.
4373 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4375 child_ctx->parent_ctx = cloned_ctx;
4376 child_ctx->parent_gen = parent_ctx->parent_gen;
4378 child_ctx->parent_ctx = parent_ctx;
4379 child_ctx->parent_gen = parent_ctx->generation;
4381 get_ctx(child_ctx->parent_ctx);
4384 mutex_unlock(&parent_ctx->mutex);
4386 perf_unpin_context(parent_ctx);
4391 static void __cpuinit perf_counter_init_cpu(int cpu)
4393 struct perf_cpu_context *cpuctx;
4395 cpuctx = &per_cpu(perf_cpu_context, cpu);
4396 __perf_counter_init_context(&cpuctx->ctx, NULL);
4398 spin_lock(&perf_resource_lock);
4399 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4400 spin_unlock(&perf_resource_lock);
4402 hw_perf_counter_setup(cpu);
4405 #ifdef CONFIG_HOTPLUG_CPU
4406 static void __perf_counter_exit_cpu(void *info)
4408 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4409 struct perf_counter_context *ctx = &cpuctx->ctx;
4410 struct perf_counter *counter, *tmp;
4412 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4413 __perf_counter_remove_from_context(counter);
4415 static void perf_counter_exit_cpu(int cpu)
4417 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4418 struct perf_counter_context *ctx = &cpuctx->ctx;
4420 mutex_lock(&ctx->mutex);
4421 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4422 mutex_unlock(&ctx->mutex);
4425 static inline void perf_counter_exit_cpu(int cpu) { }
4428 static int __cpuinit
4429 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4431 unsigned int cpu = (long)hcpu;
4435 case CPU_UP_PREPARE:
4436 case CPU_UP_PREPARE_FROZEN:
4437 perf_counter_init_cpu(cpu);
4440 case CPU_DOWN_PREPARE:
4441 case CPU_DOWN_PREPARE_FROZEN:
4442 perf_counter_exit_cpu(cpu);
4453 * This has to have a higher priority than migration_notifier in sched.c.
4455 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4456 .notifier_call = perf_cpu_notify,
4460 void __init perf_counter_init(void)
4462 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4463 (void *)(long)smp_processor_id());
4464 register_cpu_notifier(&perf_cpu_nb);
4467 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4469 return sprintf(buf, "%d\n", perf_reserved_percpu);
4473 perf_set_reserve_percpu(struct sysdev_class *class,
4477 struct perf_cpu_context *cpuctx;
4481 err = strict_strtoul(buf, 10, &val);
4484 if (val > perf_max_counters)
4487 spin_lock(&perf_resource_lock);
4488 perf_reserved_percpu = val;
4489 for_each_online_cpu(cpu) {
4490 cpuctx = &per_cpu(perf_cpu_context, cpu);
4491 spin_lock_irq(&cpuctx->ctx.lock);
4492 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4493 perf_max_counters - perf_reserved_percpu);
4494 cpuctx->max_pertask = mpt;
4495 spin_unlock_irq(&cpuctx->ctx.lock);
4497 spin_unlock(&perf_resource_lock);
4502 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4504 return sprintf(buf, "%d\n", perf_overcommit);
4508 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4513 err = strict_strtoul(buf, 10, &val);
4519 spin_lock(&perf_resource_lock);
4520 perf_overcommit = val;
4521 spin_unlock(&perf_resource_lock);
4526 static SYSDEV_CLASS_ATTR(
4529 perf_show_reserve_percpu,
4530 perf_set_reserve_percpu
4533 static SYSDEV_CLASS_ATTR(
4536 perf_show_overcommit,
4540 static struct attribute *perfclass_attrs[] = {
4541 &attr_reserve_percpu.attr,
4542 &attr_overcommit.attr,
4546 static struct attribute_group perfclass_attr_group = {
4547 .attrs = perfclass_attrs,
4548 .name = "perf_counters",
4551 static int __init perf_counter_sysfs_init(void)
4553 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4554 &perfclass_attr_group);
4556 device_initcall(perf_counter_sysfs_init);