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);
1051 swap(counter->total_time_enabled, next_counter->total_time_enabled);
1052 swap(counter->total_time_running, next_counter->total_time_running);
1055 * Since we swizzled the values, update the user visible data too.
1057 perf_counter_update_userpage(counter);
1058 perf_counter_update_userpage(next_counter);
1061 #define list_next_entry(pos, member) \
1062 list_entry(pos->member.next, typeof(*pos), member)
1064 static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1065 struct perf_counter_context *next_ctx)
1067 struct perf_counter *counter, *next_counter;
1072 counter = list_first_entry(&ctx->event_list,
1073 struct perf_counter, event_entry);
1075 next_counter = list_first_entry(&next_ctx->event_list,
1076 struct perf_counter, event_entry);
1078 while (&counter->event_entry != &ctx->event_list &&
1079 &next_counter->event_entry != &next_ctx->event_list) {
1081 __perf_counter_sync_stat(counter, next_counter);
1083 counter = list_next_entry(counter, event_entry);
1084 next_counter = list_next_entry(counter, event_entry);
1089 * Called from scheduler to remove the counters of the current task,
1090 * with interrupts disabled.
1092 * We stop each counter and update the counter value in counter->count.
1094 * This does not protect us against NMI, but disable()
1095 * sets the disabled bit in the control field of counter _before_
1096 * accessing the counter control register. If a NMI hits, then it will
1097 * not restart the counter.
1099 void perf_counter_task_sched_out(struct task_struct *task,
1100 struct task_struct *next, int cpu)
1102 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1103 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1104 struct perf_counter_context *next_ctx;
1105 struct perf_counter_context *parent;
1106 struct pt_regs *regs;
1109 regs = task_pt_regs(task);
1110 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1112 if (likely(!ctx || !cpuctx->task_ctx))
1115 update_context_time(ctx);
1118 parent = rcu_dereference(ctx->parent_ctx);
1119 next_ctx = next->perf_counter_ctxp;
1120 if (parent && next_ctx &&
1121 rcu_dereference(next_ctx->parent_ctx) == parent) {
1123 * Looks like the two contexts are clones, so we might be
1124 * able to optimize the context switch. We lock both
1125 * contexts and check that they are clones under the
1126 * lock (including re-checking that neither has been
1127 * uncloned in the meantime). It doesn't matter which
1128 * order we take the locks because no other cpu could
1129 * be trying to lock both of these tasks.
1131 spin_lock(&ctx->lock);
1132 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1133 if (context_equiv(ctx, next_ctx)) {
1135 * XXX do we need a memory barrier of sorts
1136 * wrt to rcu_dereference() of perf_counter_ctxp
1138 task->perf_counter_ctxp = next_ctx;
1139 next->perf_counter_ctxp = ctx;
1141 next_ctx->task = task;
1144 perf_counter_sync_stat(ctx, next_ctx);
1146 spin_unlock(&next_ctx->lock);
1147 spin_unlock(&ctx->lock);
1152 __perf_counter_sched_out(ctx, cpuctx);
1153 cpuctx->task_ctx = NULL;
1158 * Called with IRQs disabled
1160 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1162 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1164 if (!cpuctx->task_ctx)
1167 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1170 __perf_counter_sched_out(ctx, cpuctx);
1171 cpuctx->task_ctx = NULL;
1175 * Called with IRQs disabled
1177 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1179 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1183 __perf_counter_sched_in(struct perf_counter_context *ctx,
1184 struct perf_cpu_context *cpuctx, int cpu)
1186 struct perf_counter *counter;
1189 spin_lock(&ctx->lock);
1191 if (likely(!ctx->nr_counters))
1194 ctx->timestamp = perf_clock();
1199 * First go through the list and put on any pinned groups
1200 * in order to give them the best chance of going on.
1202 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1203 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1204 !counter->attr.pinned)
1206 if (counter->cpu != -1 && counter->cpu != cpu)
1209 if (counter != counter->group_leader)
1210 counter_sched_in(counter, cpuctx, ctx, cpu);
1212 if (group_can_go_on(counter, cpuctx, 1))
1213 group_sched_in(counter, cpuctx, ctx, cpu);
1217 * If this pinned group hasn't been scheduled,
1218 * put it in error state.
1220 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1221 update_group_times(counter);
1222 counter->state = PERF_COUNTER_STATE_ERROR;
1226 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1228 * Ignore counters in OFF or ERROR state, and
1229 * ignore pinned counters since we did them already.
1231 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1232 counter->attr.pinned)
1236 * Listen to the 'cpu' scheduling filter constraint
1239 if (counter->cpu != -1 && counter->cpu != cpu)
1242 if (counter != counter->group_leader) {
1243 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1246 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1247 if (group_sched_in(counter, cpuctx, ctx, cpu))
1254 spin_unlock(&ctx->lock);
1258 * Called from scheduler to add the counters of the current task
1259 * with interrupts disabled.
1261 * We restore the counter value and then enable it.
1263 * This does not protect us against NMI, but enable()
1264 * sets the enabled bit in the control field of counter _before_
1265 * accessing the counter control register. If a NMI hits, then it will
1266 * keep the counter running.
1268 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1270 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1271 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1275 if (cpuctx->task_ctx == ctx)
1277 __perf_counter_sched_in(ctx, cpuctx, cpu);
1278 cpuctx->task_ctx = ctx;
1281 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1283 struct perf_counter_context *ctx = &cpuctx->ctx;
1285 __perf_counter_sched_in(ctx, cpuctx, cpu);
1288 #define MAX_INTERRUPTS (~0ULL)
1290 static void perf_log_throttle(struct perf_counter *counter, int enable);
1291 static void perf_log_period(struct perf_counter *counter, u64 period);
1293 static void perf_adjust_period(struct perf_counter *counter, u64 events)
1295 struct hw_perf_counter *hwc = &counter->hw;
1296 u64 period, sample_period;
1299 events *= hwc->sample_period;
1300 period = div64_u64(events, counter->attr.sample_freq);
1302 delta = (s64)(period - hwc->sample_period);
1303 delta = (delta + 7) / 8; /* low pass filter */
1305 sample_period = hwc->sample_period + delta;
1310 perf_log_period(counter, sample_period);
1312 hwc->sample_period = sample_period;
1315 static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1317 struct perf_counter *counter;
1318 struct hw_perf_counter *hwc;
1319 u64 interrupts, freq;
1321 spin_lock(&ctx->lock);
1322 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1323 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1328 interrupts = hwc->interrupts;
1329 hwc->interrupts = 0;
1332 * unthrottle counters on the tick
1334 if (interrupts == MAX_INTERRUPTS) {
1335 perf_log_throttle(counter, 1);
1336 counter->pmu->unthrottle(counter);
1337 interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1340 if (!counter->attr.freq || !counter->attr.sample_freq)
1344 * if the specified freq < HZ then we need to skip ticks
1346 if (counter->attr.sample_freq < HZ) {
1347 freq = counter->attr.sample_freq;
1349 hwc->freq_count += freq;
1350 hwc->freq_interrupts += interrupts;
1352 if (hwc->freq_count < HZ)
1355 interrupts = hwc->freq_interrupts;
1356 hwc->freq_interrupts = 0;
1357 hwc->freq_count -= HZ;
1361 perf_adjust_period(counter, freq * interrupts);
1364 * In order to avoid being stalled by an (accidental) huge
1365 * sample period, force reset the sample period if we didn't
1366 * get any events in this freq period.
1370 counter->pmu->disable(counter);
1371 atomic64_set(&hwc->period_left, 0);
1372 counter->pmu->enable(counter);
1376 spin_unlock(&ctx->lock);
1380 * Round-robin a context's counters:
1382 static void rotate_ctx(struct perf_counter_context *ctx)
1384 struct perf_counter *counter;
1386 if (!ctx->nr_counters)
1389 spin_lock(&ctx->lock);
1391 * Rotate the first entry last (works just fine for group counters too):
1394 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1395 list_move_tail(&counter->list_entry, &ctx->counter_list);
1400 spin_unlock(&ctx->lock);
1403 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1405 struct perf_cpu_context *cpuctx;
1406 struct perf_counter_context *ctx;
1408 if (!atomic_read(&nr_counters))
1411 cpuctx = &per_cpu(perf_cpu_context, cpu);
1412 ctx = curr->perf_counter_ctxp;
1414 perf_ctx_adjust_freq(&cpuctx->ctx);
1416 perf_ctx_adjust_freq(ctx);
1418 perf_counter_cpu_sched_out(cpuctx);
1420 __perf_counter_task_sched_out(ctx);
1422 rotate_ctx(&cpuctx->ctx);
1426 perf_counter_cpu_sched_in(cpuctx, cpu);
1428 perf_counter_task_sched_in(curr, cpu);
1432 * Enable all of a task's counters that have been marked enable-on-exec.
1433 * This expects task == current.
1435 static void perf_counter_enable_on_exec(struct task_struct *task)
1437 struct perf_counter_context *ctx;
1438 struct perf_counter *counter;
1439 unsigned long flags;
1442 local_irq_save(flags);
1443 ctx = task->perf_counter_ctxp;
1444 if (!ctx || !ctx->nr_counters)
1447 __perf_counter_task_sched_out(ctx);
1449 spin_lock(&ctx->lock);
1451 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1452 if (!counter->attr.enable_on_exec)
1454 counter->attr.enable_on_exec = 0;
1455 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
1457 counter->state = PERF_COUNTER_STATE_INACTIVE;
1458 counter->tstamp_enabled =
1459 ctx->time - counter->total_time_enabled;
1464 * Unclone this context if we enabled any counter.
1466 if (enabled && ctx->parent_ctx) {
1467 put_ctx(ctx->parent_ctx);
1468 ctx->parent_ctx = NULL;
1471 spin_unlock(&ctx->lock);
1473 perf_counter_task_sched_in(task, smp_processor_id());
1475 local_irq_restore(flags);
1479 * Cross CPU call to read the hardware counter
1481 static void __perf_counter_read(void *info)
1483 struct perf_counter *counter = info;
1484 struct perf_counter_context *ctx = counter->ctx;
1485 unsigned long flags;
1487 local_irq_save(flags);
1489 update_context_time(ctx);
1490 counter->pmu->read(counter);
1491 update_counter_times(counter);
1492 local_irq_restore(flags);
1495 static u64 perf_counter_read(struct perf_counter *counter)
1498 * If counter is enabled and currently active on a CPU, update the
1499 * value in the counter structure:
1501 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1502 smp_call_function_single(counter->oncpu,
1503 __perf_counter_read, counter, 1);
1504 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1505 update_counter_times(counter);
1508 return atomic64_read(&counter->count);
1512 * Initialize the perf_counter context in a task_struct:
1515 __perf_counter_init_context(struct perf_counter_context *ctx,
1516 struct task_struct *task)
1518 memset(ctx, 0, sizeof(*ctx));
1519 spin_lock_init(&ctx->lock);
1520 mutex_init(&ctx->mutex);
1521 INIT_LIST_HEAD(&ctx->counter_list);
1522 INIT_LIST_HEAD(&ctx->event_list);
1523 atomic_set(&ctx->refcount, 1);
1527 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1529 struct perf_counter_context *parent_ctx;
1530 struct perf_counter_context *ctx;
1531 struct perf_cpu_context *cpuctx;
1532 struct task_struct *task;
1533 unsigned long flags;
1537 * If cpu is not a wildcard then this is a percpu counter:
1540 /* Must be root to operate on a CPU counter: */
1541 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1542 return ERR_PTR(-EACCES);
1544 if (cpu < 0 || cpu > num_possible_cpus())
1545 return ERR_PTR(-EINVAL);
1548 * We could be clever and allow to attach a counter to an
1549 * offline CPU and activate it when the CPU comes up, but
1552 if (!cpu_isset(cpu, cpu_online_map))
1553 return ERR_PTR(-ENODEV);
1555 cpuctx = &per_cpu(perf_cpu_context, cpu);
1566 task = find_task_by_vpid(pid);
1568 get_task_struct(task);
1572 return ERR_PTR(-ESRCH);
1575 * Can't attach counters to a dying task.
1578 if (task->flags & PF_EXITING)
1581 /* Reuse ptrace permission checks for now. */
1583 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1587 ctx = perf_lock_task_context(task, &flags);
1589 parent_ctx = ctx->parent_ctx;
1591 put_ctx(parent_ctx);
1592 ctx->parent_ctx = NULL; /* no longer a clone */
1594 spin_unlock_irqrestore(&ctx->lock, flags);
1598 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1602 __perf_counter_init_context(ctx, task);
1604 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1606 * We raced with some other task; use
1607 * the context they set.
1612 get_task_struct(task);
1615 put_task_struct(task);
1619 put_task_struct(task);
1620 return ERR_PTR(err);
1623 static void free_counter_rcu(struct rcu_head *head)
1625 struct perf_counter *counter;
1627 counter = container_of(head, struct perf_counter, rcu_head);
1629 put_pid_ns(counter->ns);
1633 static void perf_pending_sync(struct perf_counter *counter);
1635 static void free_counter(struct perf_counter *counter)
1637 perf_pending_sync(counter);
1639 if (!counter->parent) {
1640 atomic_dec(&nr_counters);
1641 if (counter->attr.mmap)
1642 atomic_dec(&nr_mmap_counters);
1643 if (counter->attr.comm)
1644 atomic_dec(&nr_comm_counters);
1647 if (counter->destroy)
1648 counter->destroy(counter);
1650 put_ctx(counter->ctx);
1651 call_rcu(&counter->rcu_head, free_counter_rcu);
1655 * Called when the last reference to the file is gone.
1657 static int perf_release(struct inode *inode, struct file *file)
1659 struct perf_counter *counter = file->private_data;
1660 struct perf_counter_context *ctx = counter->ctx;
1662 file->private_data = NULL;
1664 WARN_ON_ONCE(ctx->parent_ctx);
1665 mutex_lock(&ctx->mutex);
1666 perf_counter_remove_from_context(counter);
1667 mutex_unlock(&ctx->mutex);
1669 mutex_lock(&counter->owner->perf_counter_mutex);
1670 list_del_init(&counter->owner_entry);
1671 mutex_unlock(&counter->owner->perf_counter_mutex);
1672 put_task_struct(counter->owner);
1674 free_counter(counter);
1680 * Read the performance counter - simple non blocking version for now
1683 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1689 * Return end-of-file for a read on a counter that is in
1690 * error state (i.e. because it was pinned but it couldn't be
1691 * scheduled on to the CPU at some point).
1693 if (counter->state == PERF_COUNTER_STATE_ERROR)
1696 WARN_ON_ONCE(counter->ctx->parent_ctx);
1697 mutex_lock(&counter->child_mutex);
1698 values[0] = perf_counter_read(counter);
1700 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1701 values[n++] = counter->total_time_enabled +
1702 atomic64_read(&counter->child_total_time_enabled);
1703 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1704 values[n++] = counter->total_time_running +
1705 atomic64_read(&counter->child_total_time_running);
1706 if (counter->attr.read_format & PERF_FORMAT_ID)
1707 values[n++] = counter->id;
1708 mutex_unlock(&counter->child_mutex);
1710 if (count < n * sizeof(u64))
1712 count = n * sizeof(u64);
1714 if (copy_to_user(buf, values, count))
1721 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1723 struct perf_counter *counter = file->private_data;
1725 return perf_read_hw(counter, buf, count);
1728 static unsigned int perf_poll(struct file *file, poll_table *wait)
1730 struct perf_counter *counter = file->private_data;
1731 struct perf_mmap_data *data;
1732 unsigned int events = POLL_HUP;
1735 data = rcu_dereference(counter->data);
1737 events = atomic_xchg(&data->poll, 0);
1740 poll_wait(file, &counter->waitq, wait);
1745 static void perf_counter_reset(struct perf_counter *counter)
1747 (void)perf_counter_read(counter);
1748 atomic64_set(&counter->count, 0);
1749 perf_counter_update_userpage(counter);
1753 * Holding the top-level counter's child_mutex means that any
1754 * descendant process that has inherited this counter will block
1755 * in sync_child_counter if it goes to exit, thus satisfying the
1756 * task existence requirements of perf_counter_enable/disable.
1758 static void perf_counter_for_each_child(struct perf_counter *counter,
1759 void (*func)(struct perf_counter *))
1761 struct perf_counter *child;
1763 WARN_ON_ONCE(counter->ctx->parent_ctx);
1764 mutex_lock(&counter->child_mutex);
1766 list_for_each_entry(child, &counter->child_list, child_list)
1768 mutex_unlock(&counter->child_mutex);
1771 static void perf_counter_for_each(struct perf_counter *counter,
1772 void (*func)(struct perf_counter *))
1774 struct perf_counter_context *ctx = counter->ctx;
1775 struct perf_counter *sibling;
1777 WARN_ON_ONCE(ctx->parent_ctx);
1778 mutex_lock(&ctx->mutex);
1779 counter = counter->group_leader;
1781 perf_counter_for_each_child(counter, func);
1783 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1784 perf_counter_for_each_child(counter, func);
1785 mutex_unlock(&ctx->mutex);
1788 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1790 struct perf_counter_context *ctx = counter->ctx;
1795 if (!counter->attr.sample_period)
1798 size = copy_from_user(&value, arg, sizeof(value));
1799 if (size != sizeof(value))
1805 spin_lock_irq(&ctx->lock);
1806 if (counter->attr.freq) {
1807 if (value > sysctl_perf_counter_sample_rate) {
1812 counter->attr.sample_freq = value;
1814 perf_log_period(counter, value);
1816 counter->attr.sample_period = value;
1817 counter->hw.sample_period = value;
1820 spin_unlock_irq(&ctx->lock);
1825 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1827 struct perf_counter *counter = file->private_data;
1828 void (*func)(struct perf_counter *);
1832 case PERF_COUNTER_IOC_ENABLE:
1833 func = perf_counter_enable;
1835 case PERF_COUNTER_IOC_DISABLE:
1836 func = perf_counter_disable;
1838 case PERF_COUNTER_IOC_RESET:
1839 func = perf_counter_reset;
1842 case PERF_COUNTER_IOC_REFRESH:
1843 return perf_counter_refresh(counter, arg);
1845 case PERF_COUNTER_IOC_PERIOD:
1846 return perf_counter_period(counter, (u64 __user *)arg);
1852 if (flags & PERF_IOC_FLAG_GROUP)
1853 perf_counter_for_each(counter, func);
1855 perf_counter_for_each_child(counter, func);
1860 int perf_counter_task_enable(void)
1862 struct perf_counter *counter;
1864 mutex_lock(¤t->perf_counter_mutex);
1865 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1866 perf_counter_for_each_child(counter, perf_counter_enable);
1867 mutex_unlock(¤t->perf_counter_mutex);
1872 int perf_counter_task_disable(void)
1874 struct perf_counter *counter;
1876 mutex_lock(¤t->perf_counter_mutex);
1877 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1878 perf_counter_for_each_child(counter, perf_counter_disable);
1879 mutex_unlock(¤t->perf_counter_mutex);
1884 static int perf_counter_index(struct perf_counter *counter)
1886 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1889 return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
1893 * Callers need to ensure there can be no nesting of this function, otherwise
1894 * the seqlock logic goes bad. We can not serialize this because the arch
1895 * code calls this from NMI context.
1897 void perf_counter_update_userpage(struct perf_counter *counter)
1899 struct perf_counter_mmap_page *userpg;
1900 struct perf_mmap_data *data;
1903 data = rcu_dereference(counter->data);
1907 userpg = data->user_page;
1910 * Disable preemption so as to not let the corresponding user-space
1911 * spin too long if we get preempted.
1916 userpg->index = perf_counter_index(counter);
1917 userpg->offset = atomic64_read(&counter->count);
1918 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1919 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1921 userpg->time_enabled = counter->total_time_enabled +
1922 atomic64_read(&counter->child_total_time_enabled);
1924 userpg->time_running = counter->total_time_running +
1925 atomic64_read(&counter->child_total_time_running);
1934 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1936 struct perf_counter *counter = vma->vm_file->private_data;
1937 struct perf_mmap_data *data;
1938 int ret = VM_FAULT_SIGBUS;
1940 if (vmf->flags & FAULT_FLAG_MKWRITE) {
1941 if (vmf->pgoff == 0)
1947 data = rcu_dereference(counter->data);
1951 if (vmf->pgoff == 0) {
1952 vmf->page = virt_to_page(data->user_page);
1954 int nr = vmf->pgoff - 1;
1956 if ((unsigned)nr > data->nr_pages)
1959 if (vmf->flags & FAULT_FLAG_WRITE)
1962 vmf->page = virt_to_page(data->data_pages[nr]);
1965 get_page(vmf->page);
1966 vmf->page->mapping = vma->vm_file->f_mapping;
1967 vmf->page->index = vmf->pgoff;
1976 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1978 struct perf_mmap_data *data;
1982 WARN_ON(atomic_read(&counter->mmap_count));
1984 size = sizeof(struct perf_mmap_data);
1985 size += nr_pages * sizeof(void *);
1987 data = kzalloc(size, GFP_KERNEL);
1991 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1992 if (!data->user_page)
1993 goto fail_user_page;
1995 for (i = 0; i < nr_pages; i++) {
1996 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1997 if (!data->data_pages[i])
1998 goto fail_data_pages;
2001 data->nr_pages = nr_pages;
2002 atomic_set(&data->lock, -1);
2004 rcu_assign_pointer(counter->data, data);
2009 for (i--; i >= 0; i--)
2010 free_page((unsigned long)data->data_pages[i]);
2012 free_page((unsigned long)data->user_page);
2021 static void perf_mmap_free_page(unsigned long addr)
2023 struct page *page = virt_to_page(addr);
2025 page->mapping = NULL;
2029 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2031 struct perf_mmap_data *data;
2034 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2036 perf_mmap_free_page((unsigned long)data->user_page);
2037 for (i = 0; i < data->nr_pages; i++)
2038 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2043 static void perf_mmap_data_free(struct perf_counter *counter)
2045 struct perf_mmap_data *data = counter->data;
2047 WARN_ON(atomic_read(&counter->mmap_count));
2049 rcu_assign_pointer(counter->data, NULL);
2050 call_rcu(&data->rcu_head, __perf_mmap_data_free);
2053 static void perf_mmap_open(struct vm_area_struct *vma)
2055 struct perf_counter *counter = vma->vm_file->private_data;
2057 atomic_inc(&counter->mmap_count);
2060 static void perf_mmap_close(struct vm_area_struct *vma)
2062 struct perf_counter *counter = vma->vm_file->private_data;
2064 WARN_ON_ONCE(counter->ctx->parent_ctx);
2065 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2066 struct user_struct *user = current_user();
2068 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2069 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2070 perf_mmap_data_free(counter);
2071 mutex_unlock(&counter->mmap_mutex);
2075 static struct vm_operations_struct perf_mmap_vmops = {
2076 .open = perf_mmap_open,
2077 .close = perf_mmap_close,
2078 .fault = perf_mmap_fault,
2079 .page_mkwrite = perf_mmap_fault,
2082 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2084 struct perf_counter *counter = file->private_data;
2085 unsigned long user_locked, user_lock_limit;
2086 struct user_struct *user = current_user();
2087 unsigned long locked, lock_limit;
2088 unsigned long vma_size;
2089 unsigned long nr_pages;
2090 long user_extra, extra;
2093 if (!(vma->vm_flags & VM_SHARED))
2096 vma_size = vma->vm_end - vma->vm_start;
2097 nr_pages = (vma_size / PAGE_SIZE) - 1;
2100 * If we have data pages ensure they're a power-of-two number, so we
2101 * can do bitmasks instead of modulo.
2103 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2106 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2109 if (vma->vm_pgoff != 0)
2112 WARN_ON_ONCE(counter->ctx->parent_ctx);
2113 mutex_lock(&counter->mmap_mutex);
2114 if (atomic_inc_not_zero(&counter->mmap_count)) {
2115 if (nr_pages != counter->data->nr_pages)
2120 user_extra = nr_pages + 1;
2121 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2124 * Increase the limit linearly with more CPUs:
2126 user_lock_limit *= num_online_cpus();
2128 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2131 if (user_locked > user_lock_limit)
2132 extra = user_locked - user_lock_limit;
2134 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2135 lock_limit >>= PAGE_SHIFT;
2136 locked = vma->vm_mm->locked_vm + extra;
2138 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2143 WARN_ON(counter->data);
2144 ret = perf_mmap_data_alloc(counter, nr_pages);
2148 atomic_set(&counter->mmap_count, 1);
2149 atomic_long_add(user_extra, &user->locked_vm);
2150 vma->vm_mm->locked_vm += extra;
2151 counter->data->nr_locked = extra;
2152 if (vma->vm_flags & VM_WRITE)
2153 counter->data->writable = 1;
2156 mutex_unlock(&counter->mmap_mutex);
2158 vma->vm_flags |= VM_RESERVED;
2159 vma->vm_ops = &perf_mmap_vmops;
2164 static int perf_fasync(int fd, struct file *filp, int on)
2166 struct inode *inode = filp->f_path.dentry->d_inode;
2167 struct perf_counter *counter = filp->private_data;
2170 mutex_lock(&inode->i_mutex);
2171 retval = fasync_helper(fd, filp, on, &counter->fasync);
2172 mutex_unlock(&inode->i_mutex);
2180 static const struct file_operations perf_fops = {
2181 .release = perf_release,
2184 .unlocked_ioctl = perf_ioctl,
2185 .compat_ioctl = perf_ioctl,
2187 .fasync = perf_fasync,
2191 * Perf counter wakeup
2193 * If there's data, ensure we set the poll() state and publish everything
2194 * to user-space before waking everybody up.
2197 void perf_counter_wakeup(struct perf_counter *counter)
2199 wake_up_all(&counter->waitq);
2201 if (counter->pending_kill) {
2202 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2203 counter->pending_kill = 0;
2210 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2212 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2213 * single linked list and use cmpxchg() to add entries lockless.
2216 static void perf_pending_counter(struct perf_pending_entry *entry)
2218 struct perf_counter *counter = container_of(entry,
2219 struct perf_counter, pending);
2221 if (counter->pending_disable) {
2222 counter->pending_disable = 0;
2223 perf_counter_disable(counter);
2226 if (counter->pending_wakeup) {
2227 counter->pending_wakeup = 0;
2228 perf_counter_wakeup(counter);
2232 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2234 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2238 static void perf_pending_queue(struct perf_pending_entry *entry,
2239 void (*func)(struct perf_pending_entry *))
2241 struct perf_pending_entry **head;
2243 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2248 head = &get_cpu_var(perf_pending_head);
2251 entry->next = *head;
2252 } while (cmpxchg(head, entry->next, entry) != entry->next);
2254 set_perf_counter_pending();
2256 put_cpu_var(perf_pending_head);
2259 static int __perf_pending_run(void)
2261 struct perf_pending_entry *list;
2264 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2265 while (list != PENDING_TAIL) {
2266 void (*func)(struct perf_pending_entry *);
2267 struct perf_pending_entry *entry = list;
2274 * Ensure we observe the unqueue before we issue the wakeup,
2275 * so that we won't be waiting forever.
2276 * -- see perf_not_pending().
2287 static inline int perf_not_pending(struct perf_counter *counter)
2290 * If we flush on whatever cpu we run, there is a chance we don't
2294 __perf_pending_run();
2298 * Ensure we see the proper queue state before going to sleep
2299 * so that we do not miss the wakeup. -- see perf_pending_handle()
2302 return counter->pending.next == NULL;
2305 static void perf_pending_sync(struct perf_counter *counter)
2307 wait_event(counter->waitq, perf_not_pending(counter));
2310 void perf_counter_do_pending(void)
2312 __perf_pending_run();
2316 * Callchain support -- arch specific
2319 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2328 struct perf_output_handle {
2329 struct perf_counter *counter;
2330 struct perf_mmap_data *data;
2332 unsigned long offset;
2336 unsigned long flags;
2339 static bool perf_output_space(struct perf_mmap_data *data,
2340 unsigned int offset, unsigned int head)
2345 if (!data->writable)
2348 mask = (data->nr_pages << PAGE_SHIFT) - 1;
2350 * Userspace could choose to issue a mb() before updating the tail
2351 * pointer. So that all reads will be completed before the write is
2354 tail = ACCESS_ONCE(data->user_page->data_tail);
2357 offset = (offset - tail) & mask;
2358 head = (head - tail) & mask;
2360 if ((int)(head - offset) < 0)
2366 static void perf_output_wakeup(struct perf_output_handle *handle)
2368 atomic_set(&handle->data->poll, POLL_IN);
2371 handle->counter->pending_wakeup = 1;
2372 perf_pending_queue(&handle->counter->pending,
2373 perf_pending_counter);
2375 perf_counter_wakeup(handle->counter);
2379 * Curious locking construct.
2381 * We need to ensure a later event doesn't publish a head when a former
2382 * event isn't done writing. However since we need to deal with NMIs we
2383 * cannot fully serialize things.
2385 * What we do is serialize between CPUs so we only have to deal with NMI
2386 * nesting on a single CPU.
2388 * We only publish the head (and generate a wakeup) when the outer-most
2391 static void perf_output_lock(struct perf_output_handle *handle)
2393 struct perf_mmap_data *data = handle->data;
2398 local_irq_save(handle->flags);
2399 cpu = smp_processor_id();
2401 if (in_nmi() && atomic_read(&data->lock) == cpu)
2404 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2410 static void perf_output_unlock(struct perf_output_handle *handle)
2412 struct perf_mmap_data *data = handle->data;
2416 data->done_head = data->head;
2418 if (!handle->locked)
2423 * The xchg implies a full barrier that ensures all writes are done
2424 * before we publish the new head, matched by a rmb() in userspace when
2425 * reading this position.
2427 while ((head = atomic_long_xchg(&data->done_head, 0)))
2428 data->user_page->data_head = head;
2431 * NMI can happen here, which means we can miss a done_head update.
2434 cpu = atomic_xchg(&data->lock, -1);
2435 WARN_ON_ONCE(cpu != smp_processor_id());
2438 * Therefore we have to validate we did not indeed do so.
2440 if (unlikely(atomic_long_read(&data->done_head))) {
2442 * Since we had it locked, we can lock it again.
2444 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2450 if (atomic_xchg(&data->wakeup, 0))
2451 perf_output_wakeup(handle);
2453 local_irq_restore(handle->flags);
2456 static void perf_output_copy(struct perf_output_handle *handle,
2457 const void *buf, unsigned int len)
2459 unsigned int pages_mask;
2460 unsigned int offset;
2464 offset = handle->offset;
2465 pages_mask = handle->data->nr_pages - 1;
2466 pages = handle->data->data_pages;
2469 unsigned int page_offset;
2472 nr = (offset >> PAGE_SHIFT) & pages_mask;
2473 page_offset = offset & (PAGE_SIZE - 1);
2474 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2476 memcpy(pages[nr] + page_offset, buf, size);
2483 handle->offset = offset;
2486 * Check we didn't copy past our reservation window, taking the
2487 * possible unsigned int wrap into account.
2489 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2492 #define perf_output_put(handle, x) \
2493 perf_output_copy((handle), &(x), sizeof(x))
2495 static int perf_output_begin(struct perf_output_handle *handle,
2496 struct perf_counter *counter, unsigned int size,
2497 int nmi, int sample)
2499 struct perf_mmap_data *data;
2500 unsigned int offset, head;
2503 struct perf_event_header header;
2509 * For inherited counters we send all the output towards the parent.
2511 if (counter->parent)
2512 counter = counter->parent;
2515 data = rcu_dereference(counter->data);
2519 handle->data = data;
2520 handle->counter = counter;
2522 handle->sample = sample;
2524 if (!data->nr_pages)
2527 have_lost = atomic_read(&data->lost);
2529 size += sizeof(lost_event);
2531 perf_output_lock(handle);
2534 offset = head = atomic_long_read(&data->head);
2536 if (unlikely(!perf_output_space(data, offset, head)))
2538 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2540 handle->offset = offset;
2541 handle->head = head;
2543 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2544 atomic_set(&data->wakeup, 1);
2547 lost_event.header.type = PERF_EVENT_LOST;
2548 lost_event.header.misc = 0;
2549 lost_event.header.size = sizeof(lost_event);
2550 lost_event.id = counter->id;
2551 lost_event.lost = atomic_xchg(&data->lost, 0);
2553 perf_output_put(handle, lost_event);
2559 atomic_inc(&data->lost);
2560 perf_output_unlock(handle);
2567 static void perf_output_end(struct perf_output_handle *handle)
2569 struct perf_counter *counter = handle->counter;
2570 struct perf_mmap_data *data = handle->data;
2572 int wakeup_events = counter->attr.wakeup_events;
2574 if (handle->sample && wakeup_events) {
2575 int events = atomic_inc_return(&data->events);
2576 if (events >= wakeup_events) {
2577 atomic_sub(wakeup_events, &data->events);
2578 atomic_set(&data->wakeup, 1);
2582 perf_output_unlock(handle);
2586 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2589 * only top level counters have the pid namespace they were created in
2591 if (counter->parent)
2592 counter = counter->parent;
2594 return task_tgid_nr_ns(p, counter->ns);
2597 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2600 * only top level counters have the pid namespace they were created in
2602 if (counter->parent)
2603 counter = counter->parent;
2605 return task_pid_nr_ns(p, counter->ns);
2608 static void perf_counter_output(struct perf_counter *counter, int nmi,
2609 struct perf_sample_data *data)
2612 u64 sample_type = counter->attr.sample_type;
2613 struct perf_output_handle handle;
2614 struct perf_event_header header;
2623 struct perf_callchain_entry *callchain = NULL;
2624 int callchain_size = 0;
2630 header.type = PERF_EVENT_SAMPLE;
2631 header.size = sizeof(header);
2634 header.misc |= perf_misc_flags(data->regs);
2636 if (sample_type & PERF_SAMPLE_IP) {
2637 ip = perf_instruction_pointer(data->regs);
2638 header.size += sizeof(ip);
2641 if (sample_type & PERF_SAMPLE_TID) {
2642 /* namespace issues */
2643 tid_entry.pid = perf_counter_pid(counter, current);
2644 tid_entry.tid = perf_counter_tid(counter, current);
2646 header.size += sizeof(tid_entry);
2649 if (sample_type & PERF_SAMPLE_TIME) {
2651 * Maybe do better on x86 and provide cpu_clock_nmi()
2653 time = sched_clock();
2655 header.size += sizeof(u64);
2658 if (sample_type & PERF_SAMPLE_ADDR)
2659 header.size += sizeof(u64);
2661 if (sample_type & PERF_SAMPLE_ID)
2662 header.size += sizeof(u64);
2664 if (sample_type & PERF_SAMPLE_CPU) {
2665 header.size += sizeof(cpu_entry);
2667 cpu_entry.cpu = raw_smp_processor_id();
2670 if (sample_type & PERF_SAMPLE_PERIOD)
2671 header.size += sizeof(u64);
2673 if (sample_type & PERF_SAMPLE_GROUP) {
2674 header.size += sizeof(u64) +
2675 counter->nr_siblings * sizeof(group_entry);
2678 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2679 callchain = perf_callchain(data->regs);
2682 callchain_size = (1 + callchain->nr) * sizeof(u64);
2683 header.size += callchain_size;
2685 header.size += sizeof(u64);
2688 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2692 perf_output_put(&handle, header);
2694 if (sample_type & PERF_SAMPLE_IP)
2695 perf_output_put(&handle, ip);
2697 if (sample_type & PERF_SAMPLE_TID)
2698 perf_output_put(&handle, tid_entry);
2700 if (sample_type & PERF_SAMPLE_TIME)
2701 perf_output_put(&handle, time);
2703 if (sample_type & PERF_SAMPLE_ADDR)
2704 perf_output_put(&handle, data->addr);
2706 if (sample_type & PERF_SAMPLE_ID)
2707 perf_output_put(&handle, counter->id);
2709 if (sample_type & PERF_SAMPLE_CPU)
2710 perf_output_put(&handle, cpu_entry);
2712 if (sample_type & PERF_SAMPLE_PERIOD)
2713 perf_output_put(&handle, data->period);
2716 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2718 if (sample_type & PERF_SAMPLE_GROUP) {
2719 struct perf_counter *leader, *sub;
2720 u64 nr = counter->nr_siblings;
2722 perf_output_put(&handle, nr);
2724 leader = counter->group_leader;
2725 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2727 sub->pmu->read(sub);
2729 group_entry.id = sub->id;
2730 group_entry.counter = atomic64_read(&sub->count);
2732 perf_output_put(&handle, group_entry);
2736 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2738 perf_output_copy(&handle, callchain, callchain_size);
2741 perf_output_put(&handle, nr);
2745 perf_output_end(&handle);
2752 struct perf_read_event {
2753 struct perf_event_header header;
2762 perf_counter_read_event(struct perf_counter *counter,
2763 struct task_struct *task)
2765 struct perf_output_handle handle;
2766 struct perf_read_event event = {
2768 .type = PERF_EVENT_READ,
2770 .size = sizeof(event) - sizeof(event.format),
2772 .pid = perf_counter_pid(counter, task),
2773 .tid = perf_counter_tid(counter, task),
2774 .value = atomic64_read(&counter->count),
2778 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2779 event.header.size += sizeof(u64);
2780 event.format[i++] = counter->total_time_enabled;
2783 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2784 event.header.size += sizeof(u64);
2785 event.format[i++] = counter->total_time_running;
2788 if (counter->attr.read_format & PERF_FORMAT_ID) {
2791 event.header.size += sizeof(u64);
2792 if (counter->parent)
2793 id = counter->parent->id;
2797 event.format[i++] = id;
2800 ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
2804 perf_output_copy(&handle, &event, event.header.size);
2805 perf_output_end(&handle);
2812 struct perf_fork_event {
2813 struct task_struct *task;
2816 struct perf_event_header header;
2823 static void perf_counter_fork_output(struct perf_counter *counter,
2824 struct perf_fork_event *fork_event)
2826 struct perf_output_handle handle;
2827 int size = fork_event->event.header.size;
2828 struct task_struct *task = fork_event->task;
2829 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2834 fork_event->event.pid = perf_counter_pid(counter, task);
2835 fork_event->event.ppid = perf_counter_pid(counter, task->real_parent);
2837 perf_output_put(&handle, fork_event->event);
2838 perf_output_end(&handle);
2841 static int perf_counter_fork_match(struct perf_counter *counter)
2843 if (counter->attr.comm || counter->attr.mmap)
2849 static void perf_counter_fork_ctx(struct perf_counter_context *ctx,
2850 struct perf_fork_event *fork_event)
2852 struct perf_counter *counter;
2854 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2858 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2859 if (perf_counter_fork_match(counter))
2860 perf_counter_fork_output(counter, fork_event);
2865 static void perf_counter_fork_event(struct perf_fork_event *fork_event)
2867 struct perf_cpu_context *cpuctx;
2868 struct perf_counter_context *ctx;
2870 cpuctx = &get_cpu_var(perf_cpu_context);
2871 perf_counter_fork_ctx(&cpuctx->ctx, fork_event);
2872 put_cpu_var(perf_cpu_context);
2876 * doesn't really matter which of the child contexts the
2877 * events ends up in.
2879 ctx = rcu_dereference(current->perf_counter_ctxp);
2881 perf_counter_fork_ctx(ctx, fork_event);
2885 void perf_counter_fork(struct task_struct *task)
2887 struct perf_fork_event fork_event;
2889 if (!atomic_read(&nr_comm_counters) &&
2890 !atomic_read(&nr_mmap_counters))
2893 fork_event = (struct perf_fork_event){
2897 .type = PERF_EVENT_FORK,
2898 .size = sizeof(fork_event.event),
2903 perf_counter_fork_event(&fork_event);
2910 struct perf_comm_event {
2911 struct task_struct *task;
2916 struct perf_event_header header;
2923 static void perf_counter_comm_output(struct perf_counter *counter,
2924 struct perf_comm_event *comm_event)
2926 struct perf_output_handle handle;
2927 int size = comm_event->event.header.size;
2928 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2933 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
2934 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
2936 perf_output_put(&handle, comm_event->event);
2937 perf_output_copy(&handle, comm_event->comm,
2938 comm_event->comm_size);
2939 perf_output_end(&handle);
2942 static int perf_counter_comm_match(struct perf_counter *counter)
2944 if (counter->attr.comm)
2950 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2951 struct perf_comm_event *comm_event)
2953 struct perf_counter *counter;
2955 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2959 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2960 if (perf_counter_comm_match(counter))
2961 perf_counter_comm_output(counter, comm_event);
2966 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2968 struct perf_cpu_context *cpuctx;
2969 struct perf_counter_context *ctx;
2971 char comm[TASK_COMM_LEN];
2973 memset(comm, 0, sizeof(comm));
2974 strncpy(comm, comm_event->task->comm, sizeof(comm));
2975 size = ALIGN(strlen(comm)+1, sizeof(u64));
2977 comm_event->comm = comm;
2978 comm_event->comm_size = size;
2980 comm_event->event.header.size = sizeof(comm_event->event) + size;
2982 cpuctx = &get_cpu_var(perf_cpu_context);
2983 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2984 put_cpu_var(perf_cpu_context);
2988 * doesn't really matter which of the child contexts the
2989 * events ends up in.
2991 ctx = rcu_dereference(current->perf_counter_ctxp);
2993 perf_counter_comm_ctx(ctx, comm_event);
2997 void perf_counter_comm(struct task_struct *task)
2999 struct perf_comm_event comm_event;
3001 if (task->perf_counter_ctxp)
3002 perf_counter_enable_on_exec(task);
3004 if (!atomic_read(&nr_comm_counters))
3007 comm_event = (struct perf_comm_event){
3010 .header = { .type = PERF_EVENT_COMM, },
3014 perf_counter_comm_event(&comm_event);
3021 struct perf_mmap_event {
3022 struct vm_area_struct *vma;
3024 const char *file_name;
3028 struct perf_event_header header;
3038 static void perf_counter_mmap_output(struct perf_counter *counter,
3039 struct perf_mmap_event *mmap_event)
3041 struct perf_output_handle handle;
3042 int size = mmap_event->event.header.size;
3043 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3048 mmap_event->event.pid = perf_counter_pid(counter, current);
3049 mmap_event->event.tid = perf_counter_tid(counter, current);
3051 perf_output_put(&handle, mmap_event->event);
3052 perf_output_copy(&handle, mmap_event->file_name,
3053 mmap_event->file_size);
3054 perf_output_end(&handle);
3057 static int perf_counter_mmap_match(struct perf_counter *counter,
3058 struct perf_mmap_event *mmap_event)
3060 if (counter->attr.mmap)
3066 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3067 struct perf_mmap_event *mmap_event)
3069 struct perf_counter *counter;
3071 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3075 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3076 if (perf_counter_mmap_match(counter, mmap_event))
3077 perf_counter_mmap_output(counter, mmap_event);
3082 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3084 struct perf_cpu_context *cpuctx;
3085 struct perf_counter_context *ctx;
3086 struct vm_area_struct *vma = mmap_event->vma;
3087 struct file *file = vma->vm_file;
3093 memset(tmp, 0, sizeof(tmp));
3097 * d_path works from the end of the buffer backwards, so we
3098 * need to add enough zero bytes after the string to handle
3099 * the 64bit alignment we do later.
3101 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3103 name = strncpy(tmp, "//enomem", sizeof(tmp));
3106 name = d_path(&file->f_path, buf, PATH_MAX);
3108 name = strncpy(tmp, "//toolong", sizeof(tmp));
3112 if (arch_vma_name(mmap_event->vma)) {
3113 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3119 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3123 name = strncpy(tmp, "//anon", sizeof(tmp));
3128 size = ALIGN(strlen(name)+1, sizeof(u64));
3130 mmap_event->file_name = name;
3131 mmap_event->file_size = size;
3133 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3135 cpuctx = &get_cpu_var(perf_cpu_context);
3136 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3137 put_cpu_var(perf_cpu_context);
3141 * doesn't really matter which of the child contexts the
3142 * events ends up in.
3144 ctx = rcu_dereference(current->perf_counter_ctxp);
3146 perf_counter_mmap_ctx(ctx, mmap_event);
3152 void __perf_counter_mmap(struct vm_area_struct *vma)
3154 struct perf_mmap_event mmap_event;
3156 if (!atomic_read(&nr_mmap_counters))
3159 mmap_event = (struct perf_mmap_event){
3162 .header = { .type = PERF_EVENT_MMAP, },
3163 .start = vma->vm_start,
3164 .len = vma->vm_end - vma->vm_start,
3165 .pgoff = vma->vm_pgoff,
3169 perf_counter_mmap_event(&mmap_event);
3173 * Log sample_period changes so that analyzing tools can re-normalize the
3178 struct perf_event_header header;
3184 static void perf_log_period(struct perf_counter *counter, u64 period)
3186 struct perf_output_handle handle;
3187 struct freq_event event;
3190 if (counter->hw.sample_period == period)
3193 if (counter->attr.sample_type & PERF_SAMPLE_PERIOD)
3196 event = (struct freq_event) {
3198 .type = PERF_EVENT_PERIOD,
3200 .size = sizeof(event),
3202 .time = sched_clock(),
3207 ret = perf_output_begin(&handle, counter, sizeof(event), 1, 0);
3211 perf_output_put(&handle, event);
3212 perf_output_end(&handle);
3216 * IRQ throttle logging
3219 static void perf_log_throttle(struct perf_counter *counter, int enable)
3221 struct perf_output_handle handle;
3225 struct perf_event_header header;
3228 } throttle_event = {
3230 .type = PERF_EVENT_THROTTLE + 1,
3232 .size = sizeof(throttle_event),
3234 .time = sched_clock(),
3238 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3242 perf_output_put(&handle, throttle_event);
3243 perf_output_end(&handle);
3247 * Generic counter overflow handling, sampling.
3250 int perf_counter_overflow(struct perf_counter *counter, int nmi,
3251 struct perf_sample_data *data)
3253 int events = atomic_read(&counter->event_limit);
3254 int throttle = counter->pmu->unthrottle != NULL;
3255 struct hw_perf_counter *hwc = &counter->hw;
3261 if (hwc->interrupts != MAX_INTERRUPTS) {
3263 if (HZ * hwc->interrupts >
3264 (u64)sysctl_perf_counter_sample_rate) {
3265 hwc->interrupts = MAX_INTERRUPTS;
3266 perf_log_throttle(counter, 0);
3271 * Keep re-disabling counters even though on the previous
3272 * pass we disabled it - just in case we raced with a
3273 * sched-in and the counter got enabled again:
3279 if (counter->attr.freq) {
3280 u64 now = sched_clock();
3281 s64 delta = now - hwc->freq_stamp;
3283 hwc->freq_stamp = now;
3285 if (delta > 0 && delta < TICK_NSEC)
3286 perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3290 * XXX event_limit might not quite work as expected on inherited
3294 counter->pending_kill = POLL_IN;
3295 if (events && atomic_dec_and_test(&counter->event_limit)) {
3297 counter->pending_kill = POLL_HUP;
3299 counter->pending_disable = 1;
3300 perf_pending_queue(&counter->pending,
3301 perf_pending_counter);
3303 perf_counter_disable(counter);
3306 perf_counter_output(counter, nmi, data);
3311 * Generic software counter infrastructure
3314 static void perf_swcounter_update(struct perf_counter *counter)
3316 struct hw_perf_counter *hwc = &counter->hw;
3321 prev = atomic64_read(&hwc->prev_count);
3322 now = atomic64_read(&hwc->count);
3323 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
3328 atomic64_add(delta, &counter->count);
3329 atomic64_sub(delta, &hwc->period_left);
3332 static void perf_swcounter_set_period(struct perf_counter *counter)
3334 struct hw_perf_counter *hwc = &counter->hw;
3335 s64 left = atomic64_read(&hwc->period_left);
3336 s64 period = hwc->sample_period;
3338 if (unlikely(left <= -period)) {
3340 atomic64_set(&hwc->period_left, left);
3341 hwc->last_period = period;
3344 if (unlikely(left <= 0)) {
3346 atomic64_add(period, &hwc->period_left);
3347 hwc->last_period = period;
3350 atomic64_set(&hwc->prev_count, -left);
3351 atomic64_set(&hwc->count, -left);
3354 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3356 enum hrtimer_restart ret = HRTIMER_RESTART;
3357 struct perf_sample_data data;
3358 struct perf_counter *counter;
3361 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3362 counter->pmu->read(counter);
3365 data.regs = get_irq_regs();
3367 * In case we exclude kernel IPs or are somehow not in interrupt
3368 * context, provide the next best thing, the user IP.
3370 if ((counter->attr.exclude_kernel || !data.regs) &&
3371 !counter->attr.exclude_user)
3372 data.regs = task_pt_regs(current);
3375 if (perf_counter_overflow(counter, 0, &data))
3376 ret = HRTIMER_NORESTART;
3379 period = max_t(u64, 10000, counter->hw.sample_period);
3380 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3385 static void perf_swcounter_overflow(struct perf_counter *counter,
3386 int nmi, struct perf_sample_data *data)
3388 data->period = counter->hw.last_period;
3390 perf_swcounter_update(counter);
3391 perf_swcounter_set_period(counter);
3392 if (perf_counter_overflow(counter, nmi, data))
3393 /* soft-disable the counter */
3397 static int perf_swcounter_is_counting(struct perf_counter *counter)
3399 struct perf_counter_context *ctx;
3400 unsigned long flags;
3403 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3406 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3410 * If the counter is inactive, it could be just because
3411 * its task is scheduled out, or because it's in a group
3412 * which could not go on the PMU. We want to count in
3413 * the first case but not the second. If the context is
3414 * currently active then an inactive software counter must
3415 * be the second case. If it's not currently active then
3416 * we need to know whether the counter was active when the
3417 * context was last active, which we can determine by
3418 * comparing counter->tstamp_stopped with ctx->time.
3420 * We are within an RCU read-side critical section,
3421 * which protects the existence of *ctx.
3424 spin_lock_irqsave(&ctx->lock, flags);
3426 /* Re-check state now we have the lock */
3427 if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
3428 counter->ctx->is_active ||
3429 counter->tstamp_stopped < ctx->time)
3431 spin_unlock_irqrestore(&ctx->lock, flags);
3435 static int perf_swcounter_match(struct perf_counter *counter,
3436 enum perf_type_id type,
3437 u32 event, struct pt_regs *regs)
3439 if (!perf_swcounter_is_counting(counter))
3442 if (counter->attr.type != type)
3444 if (counter->attr.config != event)
3448 if (counter->attr.exclude_user && user_mode(regs))
3451 if (counter->attr.exclude_kernel && !user_mode(regs))
3458 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3459 int nmi, struct perf_sample_data *data)
3461 int neg = atomic64_add_negative(nr, &counter->hw.count);
3463 if (counter->hw.sample_period && !neg && data->regs)
3464 perf_swcounter_overflow(counter, nmi, data);
3467 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3468 enum perf_type_id type,
3469 u32 event, u64 nr, int nmi,
3470 struct perf_sample_data *data)
3472 struct perf_counter *counter;
3474 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3478 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3479 if (perf_swcounter_match(counter, type, event, data->regs))
3480 perf_swcounter_add(counter, nr, nmi, data);
3485 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3488 return &cpuctx->recursion[3];
3491 return &cpuctx->recursion[2];
3494 return &cpuctx->recursion[1];
3496 return &cpuctx->recursion[0];
3499 static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3501 struct perf_sample_data *data)
3503 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3504 int *recursion = perf_swcounter_recursion_context(cpuctx);
3505 struct perf_counter_context *ctx;
3513 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3517 * doesn't really matter which of the child contexts the
3518 * events ends up in.
3520 ctx = rcu_dereference(current->perf_counter_ctxp);
3522 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3529 put_cpu_var(perf_cpu_context);
3532 void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3533 struct pt_regs *regs, u64 addr)
3535 struct perf_sample_data data = {
3540 do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3543 static void perf_swcounter_read(struct perf_counter *counter)
3545 perf_swcounter_update(counter);
3548 static int perf_swcounter_enable(struct perf_counter *counter)
3550 perf_swcounter_set_period(counter);
3554 static void perf_swcounter_disable(struct perf_counter *counter)
3556 perf_swcounter_update(counter);
3559 static const struct pmu perf_ops_generic = {
3560 .enable = perf_swcounter_enable,
3561 .disable = perf_swcounter_disable,
3562 .read = perf_swcounter_read,
3566 * Software counter: cpu wall time clock
3569 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3571 int cpu = raw_smp_processor_id();
3575 now = cpu_clock(cpu);
3576 prev = atomic64_read(&counter->hw.prev_count);
3577 atomic64_set(&counter->hw.prev_count, now);
3578 atomic64_add(now - prev, &counter->count);
3581 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3583 struct hw_perf_counter *hwc = &counter->hw;
3584 int cpu = raw_smp_processor_id();
3586 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3587 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3588 hwc->hrtimer.function = perf_swcounter_hrtimer;
3589 if (hwc->sample_period) {
3590 u64 period = max_t(u64, 10000, hwc->sample_period);
3591 __hrtimer_start_range_ns(&hwc->hrtimer,
3592 ns_to_ktime(period), 0,
3593 HRTIMER_MODE_REL, 0);
3599 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3601 if (counter->hw.sample_period)
3602 hrtimer_cancel(&counter->hw.hrtimer);
3603 cpu_clock_perf_counter_update(counter);
3606 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3608 cpu_clock_perf_counter_update(counter);
3611 static const struct pmu perf_ops_cpu_clock = {
3612 .enable = cpu_clock_perf_counter_enable,
3613 .disable = cpu_clock_perf_counter_disable,
3614 .read = cpu_clock_perf_counter_read,
3618 * Software counter: task time clock
3621 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3626 prev = atomic64_xchg(&counter->hw.prev_count, now);
3628 atomic64_add(delta, &counter->count);
3631 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3633 struct hw_perf_counter *hwc = &counter->hw;
3636 now = counter->ctx->time;
3638 atomic64_set(&hwc->prev_count, now);
3639 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3640 hwc->hrtimer.function = perf_swcounter_hrtimer;
3641 if (hwc->sample_period) {
3642 u64 period = max_t(u64, 10000, hwc->sample_period);
3643 __hrtimer_start_range_ns(&hwc->hrtimer,
3644 ns_to_ktime(period), 0,
3645 HRTIMER_MODE_REL, 0);
3651 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3653 if (counter->hw.sample_period)
3654 hrtimer_cancel(&counter->hw.hrtimer);
3655 task_clock_perf_counter_update(counter, counter->ctx->time);
3659 static void task_clock_perf_counter_read(struct perf_counter *counter)
3664 update_context_time(counter->ctx);
3665 time = counter->ctx->time;
3667 u64 now = perf_clock();
3668 u64 delta = now - counter->ctx->timestamp;
3669 time = counter->ctx->time + delta;
3672 task_clock_perf_counter_update(counter, time);
3675 static const struct pmu perf_ops_task_clock = {
3676 .enable = task_clock_perf_counter_enable,
3677 .disable = task_clock_perf_counter_disable,
3678 .read = task_clock_perf_counter_read,
3681 #ifdef CONFIG_EVENT_PROFILE
3682 void perf_tpcounter_event(int event_id)
3684 struct perf_sample_data data = {
3685 .regs = get_irq_regs(),
3690 data.regs = task_pt_regs(current);
3692 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, &data);
3694 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3696 extern int ftrace_profile_enable(int);
3697 extern void ftrace_profile_disable(int);
3699 static void tp_perf_counter_destroy(struct perf_counter *counter)
3701 ftrace_profile_disable(counter->attr.config);
3704 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3706 if (ftrace_profile_enable(counter->attr.config))
3709 counter->destroy = tp_perf_counter_destroy;
3711 return &perf_ops_generic;
3714 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3720 atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3722 static void sw_perf_counter_destroy(struct perf_counter *counter)
3724 u64 event = counter->attr.config;
3726 WARN_ON(counter->parent);
3728 atomic_dec(&perf_swcounter_enabled[event]);
3731 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3733 const struct pmu *pmu = NULL;
3734 u64 event = counter->attr.config;
3737 * Software counters (currently) can't in general distinguish
3738 * between user, kernel and hypervisor events.
3739 * However, context switches and cpu migrations are considered
3740 * to be kernel events, and page faults are never hypervisor
3744 case PERF_COUNT_SW_CPU_CLOCK:
3745 pmu = &perf_ops_cpu_clock;
3748 case PERF_COUNT_SW_TASK_CLOCK:
3750 * If the user instantiates this as a per-cpu counter,
3751 * use the cpu_clock counter instead.
3753 if (counter->ctx->task)
3754 pmu = &perf_ops_task_clock;
3756 pmu = &perf_ops_cpu_clock;
3759 case PERF_COUNT_SW_PAGE_FAULTS:
3760 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
3761 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
3762 case PERF_COUNT_SW_CONTEXT_SWITCHES:
3763 case PERF_COUNT_SW_CPU_MIGRATIONS:
3764 if (!counter->parent) {
3765 atomic_inc(&perf_swcounter_enabled[event]);
3766 counter->destroy = sw_perf_counter_destroy;
3768 pmu = &perf_ops_generic;
3776 * Allocate and initialize a counter structure
3778 static struct perf_counter *
3779 perf_counter_alloc(struct perf_counter_attr *attr,
3781 struct perf_counter_context *ctx,
3782 struct perf_counter *group_leader,
3783 struct perf_counter *parent_counter,
3786 const struct pmu *pmu;
3787 struct perf_counter *counter;
3788 struct hw_perf_counter *hwc;
3791 counter = kzalloc(sizeof(*counter), gfpflags);
3793 return ERR_PTR(-ENOMEM);
3796 * Single counters are their own group leaders, with an
3797 * empty sibling list:
3800 group_leader = counter;
3802 mutex_init(&counter->child_mutex);
3803 INIT_LIST_HEAD(&counter->child_list);
3805 INIT_LIST_HEAD(&counter->list_entry);
3806 INIT_LIST_HEAD(&counter->event_entry);
3807 INIT_LIST_HEAD(&counter->sibling_list);
3808 init_waitqueue_head(&counter->waitq);
3810 mutex_init(&counter->mmap_mutex);
3813 counter->attr = *attr;
3814 counter->group_leader = group_leader;
3815 counter->pmu = NULL;
3817 counter->oncpu = -1;
3819 counter->parent = parent_counter;
3821 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
3822 counter->id = atomic64_inc_return(&perf_counter_id);
3824 counter->state = PERF_COUNTER_STATE_INACTIVE;
3827 counter->state = PERF_COUNTER_STATE_OFF;
3832 hwc->sample_period = attr->sample_period;
3833 if (attr->freq && attr->sample_freq)
3834 hwc->sample_period = 1;
3836 atomic64_set(&hwc->period_left, hwc->sample_period);
3839 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3841 if (attr->inherit && (attr->sample_type & PERF_SAMPLE_GROUP))
3844 switch (attr->type) {
3846 case PERF_TYPE_HARDWARE:
3847 case PERF_TYPE_HW_CACHE:
3848 pmu = hw_perf_counter_init(counter);
3851 case PERF_TYPE_SOFTWARE:
3852 pmu = sw_perf_counter_init(counter);
3855 case PERF_TYPE_TRACEPOINT:
3856 pmu = tp_perf_counter_init(counter);
3866 else if (IS_ERR(pmu))
3871 put_pid_ns(counter->ns);
3873 return ERR_PTR(err);
3878 if (!counter->parent) {
3879 atomic_inc(&nr_counters);
3880 if (counter->attr.mmap)
3881 atomic_inc(&nr_mmap_counters);
3882 if (counter->attr.comm)
3883 atomic_inc(&nr_comm_counters);
3889 static int perf_copy_attr(struct perf_counter_attr __user *uattr,
3890 struct perf_counter_attr *attr)
3895 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
3899 * zero the full structure, so that a short copy will be nice.
3901 memset(attr, 0, sizeof(*attr));
3903 ret = get_user(size, &uattr->size);
3907 if (size > PAGE_SIZE) /* silly large */
3910 if (!size) /* abi compat */
3911 size = PERF_ATTR_SIZE_VER0;
3913 if (size < PERF_ATTR_SIZE_VER0)
3917 * If we're handed a bigger struct than we know of,
3918 * ensure all the unknown bits are 0.
3920 if (size > sizeof(*attr)) {
3922 unsigned long __user *addr;
3923 unsigned long __user *end;
3925 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
3926 sizeof(unsigned long));
3927 end = PTR_ALIGN((void __user *)uattr + size,
3928 sizeof(unsigned long));
3930 for (; addr < end; addr += sizeof(unsigned long)) {
3931 ret = get_user(val, addr);
3939 ret = copy_from_user(attr, uattr, size);
3944 * If the type exists, the corresponding creation will verify
3947 if (attr->type >= PERF_TYPE_MAX)
3950 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
3953 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
3956 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
3963 put_user(sizeof(*attr), &uattr->size);
3969 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3971 * @attr_uptr: event type attributes for monitoring/sampling
3974 * @group_fd: group leader counter fd
3976 SYSCALL_DEFINE5(perf_counter_open,
3977 struct perf_counter_attr __user *, attr_uptr,
3978 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3980 struct perf_counter *counter, *group_leader;
3981 struct perf_counter_attr attr;
3982 struct perf_counter_context *ctx;
3983 struct file *counter_file = NULL;
3984 struct file *group_file = NULL;
3985 int fput_needed = 0;
3986 int fput_needed2 = 0;
3989 /* for future expandability... */
3993 ret = perf_copy_attr(attr_uptr, &attr);
3997 if (!attr.exclude_kernel) {
3998 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4003 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
4008 * Get the target context (task or percpu):
4010 ctx = find_get_context(pid, cpu);
4012 return PTR_ERR(ctx);
4015 * Look up the group leader (we will attach this counter to it):
4017 group_leader = NULL;
4018 if (group_fd != -1) {
4020 group_file = fget_light(group_fd, &fput_needed);
4022 goto err_put_context;
4023 if (group_file->f_op != &perf_fops)
4024 goto err_put_context;
4026 group_leader = group_file->private_data;
4028 * Do not allow a recursive hierarchy (this new sibling
4029 * becoming part of another group-sibling):
4031 if (group_leader->group_leader != group_leader)
4032 goto err_put_context;
4034 * Do not allow to attach to a group in a different
4035 * task or CPU context:
4037 if (group_leader->ctx != ctx)
4038 goto err_put_context;
4040 * Only a group leader can be exclusive or pinned
4042 if (attr.exclusive || attr.pinned)
4043 goto err_put_context;
4046 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
4048 ret = PTR_ERR(counter);
4049 if (IS_ERR(counter))
4050 goto err_put_context;
4052 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
4054 goto err_free_put_context;
4056 counter_file = fget_light(ret, &fput_needed2);
4058 goto err_free_put_context;
4060 counter->filp = counter_file;
4061 WARN_ON_ONCE(ctx->parent_ctx);
4062 mutex_lock(&ctx->mutex);
4063 perf_install_in_context(ctx, counter, cpu);
4065 mutex_unlock(&ctx->mutex);
4067 counter->owner = current;
4068 get_task_struct(current);
4069 mutex_lock(¤t->perf_counter_mutex);
4070 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
4071 mutex_unlock(¤t->perf_counter_mutex);
4073 fput_light(counter_file, fput_needed2);
4076 fput_light(group_file, fput_needed);
4080 err_free_put_context:
4090 * inherit a counter from parent task to child task:
4092 static struct perf_counter *
4093 inherit_counter(struct perf_counter *parent_counter,
4094 struct task_struct *parent,
4095 struct perf_counter_context *parent_ctx,
4096 struct task_struct *child,
4097 struct perf_counter *group_leader,
4098 struct perf_counter_context *child_ctx)
4100 struct perf_counter *child_counter;
4103 * Instead of creating recursive hierarchies of counters,
4104 * we link inherited counters back to the original parent,
4105 * which has a filp for sure, which we use as the reference
4108 if (parent_counter->parent)
4109 parent_counter = parent_counter->parent;
4111 child_counter = perf_counter_alloc(&parent_counter->attr,
4112 parent_counter->cpu, child_ctx,
4113 group_leader, parent_counter,
4115 if (IS_ERR(child_counter))
4116 return child_counter;
4120 * Make the child state follow the state of the parent counter,
4121 * not its attr.disabled bit. We hold the parent's mutex,
4122 * so we won't race with perf_counter_{en, dis}able_family.
4124 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4125 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4127 child_counter->state = PERF_COUNTER_STATE_OFF;
4129 if (parent_counter->attr.freq)
4130 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4133 * Link it up in the child's context:
4135 add_counter_to_ctx(child_counter, child_ctx);
4138 * Get a reference to the parent filp - we will fput it
4139 * when the child counter exits. This is safe to do because
4140 * we are in the parent and we know that the filp still
4141 * exists and has a nonzero count:
4143 atomic_long_inc(&parent_counter->filp->f_count);
4146 * Link this into the parent counter's child list
4148 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4149 mutex_lock(&parent_counter->child_mutex);
4150 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4151 mutex_unlock(&parent_counter->child_mutex);
4153 return child_counter;
4156 static int inherit_group(struct perf_counter *parent_counter,
4157 struct task_struct *parent,
4158 struct perf_counter_context *parent_ctx,
4159 struct task_struct *child,
4160 struct perf_counter_context *child_ctx)
4162 struct perf_counter *leader;
4163 struct perf_counter *sub;
4164 struct perf_counter *child_ctr;
4166 leader = inherit_counter(parent_counter, parent, parent_ctx,
4167 child, NULL, child_ctx);
4169 return PTR_ERR(leader);
4170 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4171 child_ctr = inherit_counter(sub, parent, parent_ctx,
4172 child, leader, child_ctx);
4173 if (IS_ERR(child_ctr))
4174 return PTR_ERR(child_ctr);
4179 static void sync_child_counter(struct perf_counter *child_counter,
4180 struct task_struct *child)
4182 struct perf_counter *parent_counter = child_counter->parent;
4185 if (child_counter->attr.inherit_stat)
4186 perf_counter_read_event(child_counter, child);
4188 child_val = atomic64_read(&child_counter->count);
4191 * Add back the child's count to the parent's count:
4193 atomic64_add(child_val, &parent_counter->count);
4194 atomic64_add(child_counter->total_time_enabled,
4195 &parent_counter->child_total_time_enabled);
4196 atomic64_add(child_counter->total_time_running,
4197 &parent_counter->child_total_time_running);
4200 * Remove this counter from the parent's list
4202 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4203 mutex_lock(&parent_counter->child_mutex);
4204 list_del_init(&child_counter->child_list);
4205 mutex_unlock(&parent_counter->child_mutex);
4208 * Release the parent counter, if this was the last
4211 fput(parent_counter->filp);
4215 __perf_counter_exit_task(struct perf_counter *child_counter,
4216 struct perf_counter_context *child_ctx,
4217 struct task_struct *child)
4219 struct perf_counter *parent_counter;
4221 update_counter_times(child_counter);
4222 perf_counter_remove_from_context(child_counter);
4224 parent_counter = child_counter->parent;
4226 * It can happen that parent exits first, and has counters
4227 * that are still around due to the child reference. These
4228 * counters need to be zapped - but otherwise linger.
4230 if (parent_counter) {
4231 sync_child_counter(child_counter, child);
4232 free_counter(child_counter);
4237 * When a child task exits, feed back counter values to parent counters.
4239 void perf_counter_exit_task(struct task_struct *child)
4241 struct perf_counter *child_counter, *tmp;
4242 struct perf_counter_context *child_ctx;
4243 unsigned long flags;
4245 if (likely(!child->perf_counter_ctxp))
4248 local_irq_save(flags);
4250 * We can't reschedule here because interrupts are disabled,
4251 * and either child is current or it is a task that can't be
4252 * scheduled, so we are now safe from rescheduling changing
4255 child_ctx = child->perf_counter_ctxp;
4256 __perf_counter_task_sched_out(child_ctx);
4259 * Take the context lock here so that if find_get_context is
4260 * reading child->perf_counter_ctxp, we wait until it has
4261 * incremented the context's refcount before we do put_ctx below.
4263 spin_lock(&child_ctx->lock);
4264 child->perf_counter_ctxp = NULL;
4265 if (child_ctx->parent_ctx) {
4267 * This context is a clone; unclone it so it can't get
4268 * swapped to another process while we're removing all
4269 * the counters from it.
4271 put_ctx(child_ctx->parent_ctx);
4272 child_ctx->parent_ctx = NULL;
4274 spin_unlock(&child_ctx->lock);
4275 local_irq_restore(flags);
4278 * We can recurse on the same lock type through:
4280 * __perf_counter_exit_task()
4281 * sync_child_counter()
4282 * fput(parent_counter->filp)
4284 * mutex_lock(&ctx->mutex)
4286 * But since its the parent context it won't be the same instance.
4288 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4291 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4293 __perf_counter_exit_task(child_counter, child_ctx, child);
4296 * If the last counter was a group counter, it will have appended all
4297 * its siblings to the list, but we obtained 'tmp' before that which
4298 * will still point to the list head terminating the iteration.
4300 if (!list_empty(&child_ctx->counter_list))
4303 mutex_unlock(&child_ctx->mutex);
4309 * free an unexposed, unused context as created by inheritance by
4310 * init_task below, used by fork() in case of fail.
4312 void perf_counter_free_task(struct task_struct *task)
4314 struct perf_counter_context *ctx = task->perf_counter_ctxp;
4315 struct perf_counter *counter, *tmp;
4320 mutex_lock(&ctx->mutex);
4322 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4323 struct perf_counter *parent = counter->parent;
4325 if (WARN_ON_ONCE(!parent))
4328 mutex_lock(&parent->child_mutex);
4329 list_del_init(&counter->child_list);
4330 mutex_unlock(&parent->child_mutex);
4334 list_del_counter(counter, ctx);
4335 free_counter(counter);
4338 if (!list_empty(&ctx->counter_list))
4341 mutex_unlock(&ctx->mutex);
4347 * Initialize the perf_counter context in task_struct
4349 int perf_counter_init_task(struct task_struct *child)
4351 struct perf_counter_context *child_ctx, *parent_ctx;
4352 struct perf_counter_context *cloned_ctx;
4353 struct perf_counter *counter;
4354 struct task_struct *parent = current;
4355 int inherited_all = 1;
4358 child->perf_counter_ctxp = NULL;
4360 mutex_init(&child->perf_counter_mutex);
4361 INIT_LIST_HEAD(&child->perf_counter_list);
4363 if (likely(!parent->perf_counter_ctxp))
4367 * This is executed from the parent task context, so inherit
4368 * counters that have been marked for cloning.
4369 * First allocate and initialize a context for the child.
4372 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4376 __perf_counter_init_context(child_ctx, child);
4377 child->perf_counter_ctxp = child_ctx;
4378 get_task_struct(child);
4381 * If the parent's context is a clone, pin it so it won't get
4384 parent_ctx = perf_pin_task_context(parent);
4387 * No need to check if parent_ctx != NULL here; since we saw
4388 * it non-NULL earlier, the only reason for it to become NULL
4389 * is if we exit, and since we're currently in the middle of
4390 * a fork we can't be exiting at the same time.
4394 * Lock the parent list. No need to lock the child - not PID
4395 * hashed yet and not running, so nobody can access it.
4397 mutex_lock(&parent_ctx->mutex);
4400 * We dont have to disable NMIs - we are only looking at
4401 * the list, not manipulating it:
4403 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4404 if (counter != counter->group_leader)
4407 if (!counter->attr.inherit) {
4412 ret = inherit_group(counter, parent, parent_ctx,
4420 if (inherited_all) {
4422 * Mark the child context as a clone of the parent
4423 * context, or of whatever the parent is a clone of.
4424 * Note that if the parent is a clone, it could get
4425 * uncloned at any point, but that doesn't matter
4426 * because the list of counters and the generation
4427 * count can't have changed since we took the mutex.
4429 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4431 child_ctx->parent_ctx = cloned_ctx;
4432 child_ctx->parent_gen = parent_ctx->parent_gen;
4434 child_ctx->parent_ctx = parent_ctx;
4435 child_ctx->parent_gen = parent_ctx->generation;
4437 get_ctx(child_ctx->parent_ctx);
4440 mutex_unlock(&parent_ctx->mutex);
4442 perf_unpin_context(parent_ctx);
4447 static void __cpuinit perf_counter_init_cpu(int cpu)
4449 struct perf_cpu_context *cpuctx;
4451 cpuctx = &per_cpu(perf_cpu_context, cpu);
4452 __perf_counter_init_context(&cpuctx->ctx, NULL);
4454 spin_lock(&perf_resource_lock);
4455 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4456 spin_unlock(&perf_resource_lock);
4458 hw_perf_counter_setup(cpu);
4461 #ifdef CONFIG_HOTPLUG_CPU
4462 static void __perf_counter_exit_cpu(void *info)
4464 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4465 struct perf_counter_context *ctx = &cpuctx->ctx;
4466 struct perf_counter *counter, *tmp;
4468 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4469 __perf_counter_remove_from_context(counter);
4471 static void perf_counter_exit_cpu(int cpu)
4473 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4474 struct perf_counter_context *ctx = &cpuctx->ctx;
4476 mutex_lock(&ctx->mutex);
4477 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4478 mutex_unlock(&ctx->mutex);
4481 static inline void perf_counter_exit_cpu(int cpu) { }
4484 static int __cpuinit
4485 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4487 unsigned int cpu = (long)hcpu;
4491 case CPU_UP_PREPARE:
4492 case CPU_UP_PREPARE_FROZEN:
4493 perf_counter_init_cpu(cpu);
4496 case CPU_DOWN_PREPARE:
4497 case CPU_DOWN_PREPARE_FROZEN:
4498 perf_counter_exit_cpu(cpu);
4509 * This has to have a higher priority than migration_notifier in sched.c.
4511 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4512 .notifier_call = perf_cpu_notify,
4516 void __init perf_counter_init(void)
4518 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4519 (void *)(long)smp_processor_id());
4520 register_cpu_notifier(&perf_cpu_nb);
4523 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4525 return sprintf(buf, "%d\n", perf_reserved_percpu);
4529 perf_set_reserve_percpu(struct sysdev_class *class,
4533 struct perf_cpu_context *cpuctx;
4537 err = strict_strtoul(buf, 10, &val);
4540 if (val > perf_max_counters)
4543 spin_lock(&perf_resource_lock);
4544 perf_reserved_percpu = val;
4545 for_each_online_cpu(cpu) {
4546 cpuctx = &per_cpu(perf_cpu_context, cpu);
4547 spin_lock_irq(&cpuctx->ctx.lock);
4548 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4549 perf_max_counters - perf_reserved_percpu);
4550 cpuctx->max_pertask = mpt;
4551 spin_unlock_irq(&cpuctx->ctx.lock);
4553 spin_unlock(&perf_resource_lock);
4558 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4560 return sprintf(buf, "%d\n", perf_overcommit);
4564 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4569 err = strict_strtoul(buf, 10, &val);
4575 spin_lock(&perf_resource_lock);
4576 perf_overcommit = val;
4577 spin_unlock(&perf_resource_lock);
4582 static SYSDEV_CLASS_ATTR(
4585 perf_show_reserve_percpu,
4586 perf_set_reserve_percpu
4589 static SYSDEV_CLASS_ATTR(
4592 perf_show_overcommit,
4596 static struct attribute *perfclass_attrs[] = {
4597 &attr_reserve_percpu.attr,
4598 &attr_overcommit.attr,
4602 static struct attribute_group perfclass_attr_group = {
4603 .attrs = perfclass_attrs,
4604 .name = "perf_counters",
4607 static int __init perf_counter_sysfs_init(void)
4609 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4610 &perfclass_attr_group);
4612 device_initcall(perf_counter_sysfs_init);