2 * Performance counter core code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
8 * For licensing details see kernel-base/COPYING
13 #include <linux/cpu.h>
14 #include <linux/smp.h>
15 #include <linux/file.h>
16 #include <linux/poll.h>
17 #include <linux/sysfs.h>
18 #include <linux/ptrace.h>
19 #include <linux/percpu.h>
20 #include <linux/vmstat.h>
21 #include <linux/hardirq.h>
22 #include <linux/rculist.h>
23 #include <linux/uaccess.h>
24 #include <linux/syscalls.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/kernel_stat.h>
27 #include <linux/perf_counter.h>
28 #include <linux/dcache.h>
30 #include <asm/irq_regs.h>
33 * Each CPU has a list of per CPU counters:
35 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
37 int perf_max_counters __read_mostly = 1;
38 static int perf_reserved_percpu __read_mostly;
39 static int perf_overcommit __read_mostly = 1;
42 * Mutex for (sysadmin-configurable) counter reservations:
44 static DEFINE_MUTEX(perf_resource_mutex);
47 * Architecture provided APIs - weak aliases:
49 extern __weak const struct hw_perf_counter_ops *
50 hw_perf_counter_init(struct perf_counter *counter)
55 u64 __weak hw_perf_save_disable(void) { return 0; }
56 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
57 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
58 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
59 struct perf_cpu_context *cpuctx,
60 struct perf_counter_context *ctx, int cpu)
65 void __weak perf_counter_print_debug(void) { }
68 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
70 struct perf_counter *group_leader = counter->group_leader;
73 * Depending on whether it is a standalone or sibling counter,
74 * add it straight to the context's counter list, or to the group
75 * leader's sibling list:
77 if (counter->group_leader == counter)
78 list_add_tail(&counter->list_entry, &ctx->counter_list);
80 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
81 group_leader->nr_siblings++;
84 list_add_rcu(&counter->event_entry, &ctx->event_list);
88 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
90 struct perf_counter *sibling, *tmp;
92 list_del_init(&counter->list_entry);
93 list_del_rcu(&counter->event_entry);
95 if (counter->group_leader != counter)
96 counter->group_leader->nr_siblings--;
99 * If this was a group counter with sibling counters then
100 * upgrade the siblings to singleton counters by adding them
101 * to the context list directly:
103 list_for_each_entry_safe(sibling, tmp,
104 &counter->sibling_list, list_entry) {
106 list_move_tail(&sibling->list_entry, &ctx->counter_list);
107 sibling->group_leader = sibling;
112 counter_sched_out(struct perf_counter *counter,
113 struct perf_cpu_context *cpuctx,
114 struct perf_counter_context *ctx)
116 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
119 counter->state = PERF_COUNTER_STATE_INACTIVE;
120 counter->tstamp_stopped = ctx->time;
121 counter->hw_ops->disable(counter);
124 if (!is_software_counter(counter))
125 cpuctx->active_oncpu--;
127 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
128 cpuctx->exclusive = 0;
132 group_sched_out(struct perf_counter *group_counter,
133 struct perf_cpu_context *cpuctx,
134 struct perf_counter_context *ctx)
136 struct perf_counter *counter;
138 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
141 counter_sched_out(group_counter, cpuctx, ctx);
144 * Schedule out siblings (if any):
146 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
147 counter_sched_out(counter, cpuctx, ctx);
149 if (group_counter->hw_event.exclusive)
150 cpuctx->exclusive = 0;
154 * Cross CPU call to remove a performance counter
156 * We disable the counter on the hardware level first. After that we
157 * remove it from the context list.
159 static void __perf_counter_remove_from_context(void *info)
161 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
162 struct perf_counter *counter = info;
163 struct perf_counter_context *ctx = counter->ctx;
168 * If this is a task context, we need to check whether it is
169 * the current task context of this cpu. If not it has been
170 * scheduled out before the smp call arrived.
172 if (ctx->task && cpuctx->task_ctx != ctx)
175 spin_lock_irqsave(&ctx->lock, flags);
177 counter_sched_out(counter, cpuctx, ctx);
179 counter->task = NULL;
183 * Protect the list operation against NMI by disabling the
184 * counters on a global level. NOP for non NMI based counters.
186 perf_flags = hw_perf_save_disable();
187 list_del_counter(counter, ctx);
188 hw_perf_restore(perf_flags);
192 * Allow more per task counters with respect to the
195 cpuctx->max_pertask =
196 min(perf_max_counters - ctx->nr_counters,
197 perf_max_counters - perf_reserved_percpu);
200 spin_unlock_irqrestore(&ctx->lock, flags);
205 * Remove the counter from a task's (or a CPU's) list of counters.
207 * Must be called with counter->mutex and ctx->mutex held.
209 * CPU counters are removed with a smp call. For task counters we only
210 * call when the task is on a CPU.
212 static void perf_counter_remove_from_context(struct perf_counter *counter)
214 struct perf_counter_context *ctx = counter->ctx;
215 struct task_struct *task = ctx->task;
219 * Per cpu counters are removed via an smp call and
220 * the removal is always sucessful.
222 smp_call_function_single(counter->cpu,
223 __perf_counter_remove_from_context,
229 task_oncpu_function_call(task, __perf_counter_remove_from_context,
232 spin_lock_irq(&ctx->lock);
234 * If the context is active we need to retry the smp call.
236 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
237 spin_unlock_irq(&ctx->lock);
242 * The lock prevents that this context is scheduled in so we
243 * can remove the counter safely, if the call above did not
246 if (!list_empty(&counter->list_entry)) {
248 list_del_counter(counter, ctx);
249 counter->task = NULL;
251 spin_unlock_irq(&ctx->lock);
254 static inline u64 perf_clock(void)
256 return cpu_clock(smp_processor_id());
260 * Update the record of the current time in a context.
262 static void update_context_time(struct perf_counter_context *ctx)
264 u64 now = perf_clock();
266 ctx->time += now - ctx->timestamp;
267 ctx->timestamp = now;
271 * Update the total_time_enabled and total_time_running fields for a counter.
273 static void update_counter_times(struct perf_counter *counter)
275 struct perf_counter_context *ctx = counter->ctx;
278 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
281 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
283 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
284 run_end = counter->tstamp_stopped;
288 counter->total_time_running = run_end - counter->tstamp_running;
292 * Update total_time_enabled and total_time_running for all counters in a group.
294 static void update_group_times(struct perf_counter *leader)
296 struct perf_counter *counter;
298 update_counter_times(leader);
299 list_for_each_entry(counter, &leader->sibling_list, list_entry)
300 update_counter_times(counter);
304 * Cross CPU call to disable a performance counter
306 static void __perf_counter_disable(void *info)
308 struct perf_counter *counter = info;
309 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
310 struct perf_counter_context *ctx = counter->ctx;
314 * If this is a per-task counter, need to check whether this
315 * counter's task is the current task on this cpu.
317 if (ctx->task && cpuctx->task_ctx != ctx)
320 spin_lock_irqsave(&ctx->lock, flags);
323 * If the counter is on, turn it off.
324 * If it is in error state, leave it in error state.
326 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
327 update_context_time(ctx);
328 update_counter_times(counter);
329 if (counter == counter->group_leader)
330 group_sched_out(counter, cpuctx, ctx);
332 counter_sched_out(counter, cpuctx, ctx);
333 counter->state = PERF_COUNTER_STATE_OFF;
336 spin_unlock_irqrestore(&ctx->lock, flags);
342 static void perf_counter_disable(struct perf_counter *counter)
344 struct perf_counter_context *ctx = counter->ctx;
345 struct task_struct *task = ctx->task;
349 * Disable the counter on the cpu that it's on
351 smp_call_function_single(counter->cpu, __perf_counter_disable,
357 task_oncpu_function_call(task, __perf_counter_disable, counter);
359 spin_lock_irq(&ctx->lock);
361 * If the counter is still active, we need to retry the cross-call.
363 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
364 spin_unlock_irq(&ctx->lock);
369 * Since we have the lock this context can't be scheduled
370 * in, so we can change the state safely.
372 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
373 update_counter_times(counter);
374 counter->state = PERF_COUNTER_STATE_OFF;
377 spin_unlock_irq(&ctx->lock);
381 * Disable a counter and all its children.
383 static void perf_counter_disable_family(struct perf_counter *counter)
385 struct perf_counter *child;
387 perf_counter_disable(counter);
390 * Lock the mutex to protect the list of children
392 mutex_lock(&counter->mutex);
393 list_for_each_entry(child, &counter->child_list, child_list)
394 perf_counter_disable(child);
395 mutex_unlock(&counter->mutex);
399 counter_sched_in(struct perf_counter *counter,
400 struct perf_cpu_context *cpuctx,
401 struct perf_counter_context *ctx,
404 if (counter->state <= PERF_COUNTER_STATE_OFF)
407 counter->state = PERF_COUNTER_STATE_ACTIVE;
408 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
410 * The new state must be visible before we turn it on in the hardware:
414 if (counter->hw_ops->enable(counter)) {
415 counter->state = PERF_COUNTER_STATE_INACTIVE;
420 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
422 if (!is_software_counter(counter))
423 cpuctx->active_oncpu++;
426 if (counter->hw_event.exclusive)
427 cpuctx->exclusive = 1;
433 * Return 1 for a group consisting entirely of software counters,
434 * 0 if the group contains any hardware counters.
436 static int is_software_only_group(struct perf_counter *leader)
438 struct perf_counter *counter;
440 if (!is_software_counter(leader))
443 list_for_each_entry(counter, &leader->sibling_list, list_entry)
444 if (!is_software_counter(counter))
451 * Work out whether we can put this counter group on the CPU now.
453 static int group_can_go_on(struct perf_counter *counter,
454 struct perf_cpu_context *cpuctx,
458 * Groups consisting entirely of software counters can always go on.
460 if (is_software_only_group(counter))
463 * If an exclusive group is already on, no other hardware
464 * counters can go on.
466 if (cpuctx->exclusive)
469 * If this group is exclusive and there are already
470 * counters on the CPU, it can't go on.
472 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
475 * Otherwise, try to add it if all previous groups were able
481 static void add_counter_to_ctx(struct perf_counter *counter,
482 struct perf_counter_context *ctx)
484 list_add_counter(counter, ctx);
486 counter->prev_state = PERF_COUNTER_STATE_OFF;
487 counter->tstamp_enabled = ctx->time;
488 counter->tstamp_running = ctx->time;
489 counter->tstamp_stopped = ctx->time;
493 * Cross CPU call to install and enable a performance counter
495 static void __perf_install_in_context(void *info)
497 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
498 struct perf_counter *counter = info;
499 struct perf_counter_context *ctx = counter->ctx;
500 struct perf_counter *leader = counter->group_leader;
501 int cpu = smp_processor_id();
507 * If this is a task context, we need to check whether it is
508 * the current task context of this cpu. If not it has been
509 * scheduled out before the smp call arrived.
511 if (ctx->task && cpuctx->task_ctx != ctx)
514 spin_lock_irqsave(&ctx->lock, flags);
515 update_context_time(ctx);
518 * Protect the list operation against NMI by disabling the
519 * counters on a global level. NOP for non NMI based counters.
521 perf_flags = hw_perf_save_disable();
523 add_counter_to_ctx(counter, ctx);
526 * Don't put the counter on if it is disabled or if
527 * it is in a group and the group isn't on.
529 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
530 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
534 * An exclusive counter can't go on if there are already active
535 * hardware counters, and no hardware counter can go on if there
536 * is already an exclusive counter on.
538 if (!group_can_go_on(counter, cpuctx, 1))
541 err = counter_sched_in(counter, cpuctx, ctx, cpu);
545 * This counter couldn't go on. If it is in a group
546 * then we have to pull the whole group off.
547 * If the counter group is pinned then put it in error state.
549 if (leader != counter)
550 group_sched_out(leader, cpuctx, ctx);
551 if (leader->hw_event.pinned) {
552 update_group_times(leader);
553 leader->state = PERF_COUNTER_STATE_ERROR;
557 if (!err && !ctx->task && cpuctx->max_pertask)
558 cpuctx->max_pertask--;
561 hw_perf_restore(perf_flags);
563 spin_unlock_irqrestore(&ctx->lock, flags);
567 * Attach a performance counter to a context
569 * First we add the counter to the list with the hardware enable bit
570 * in counter->hw_config cleared.
572 * If the counter is attached to a task which is on a CPU we use a smp
573 * call to enable it in the task context. The task might have been
574 * scheduled away, but we check this in the smp call again.
576 * Must be called with ctx->mutex held.
579 perf_install_in_context(struct perf_counter_context *ctx,
580 struct perf_counter *counter,
583 struct task_struct *task = ctx->task;
587 * Per cpu counters are installed via an smp call and
588 * the install is always sucessful.
590 smp_call_function_single(cpu, __perf_install_in_context,
595 counter->task = task;
597 task_oncpu_function_call(task, __perf_install_in_context,
600 spin_lock_irq(&ctx->lock);
602 * we need to retry the smp call.
604 if (ctx->is_active && list_empty(&counter->list_entry)) {
605 spin_unlock_irq(&ctx->lock);
610 * The lock prevents that this context is scheduled in so we
611 * can add the counter safely, if it the call above did not
614 if (list_empty(&counter->list_entry))
615 add_counter_to_ctx(counter, ctx);
616 spin_unlock_irq(&ctx->lock);
620 * Cross CPU call to enable a performance counter
622 static void __perf_counter_enable(void *info)
624 struct perf_counter *counter = info;
625 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
626 struct perf_counter_context *ctx = counter->ctx;
627 struct perf_counter *leader = counter->group_leader;
632 * If this is a per-task counter, need to check whether this
633 * counter's task is the current task on this cpu.
635 if (ctx->task && cpuctx->task_ctx != ctx)
638 spin_lock_irqsave(&ctx->lock, flags);
639 update_context_time(ctx);
641 counter->prev_state = counter->state;
642 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
644 counter->state = PERF_COUNTER_STATE_INACTIVE;
645 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
648 * If the counter is in a group and isn't the group leader,
649 * then don't put it on unless the group is on.
651 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
654 if (!group_can_go_on(counter, cpuctx, 1))
657 err = counter_sched_in(counter, cpuctx, ctx,
662 * If this counter can't go on and it's part of a
663 * group, then the whole group has to come off.
665 if (leader != counter)
666 group_sched_out(leader, cpuctx, ctx);
667 if (leader->hw_event.pinned) {
668 update_group_times(leader);
669 leader->state = PERF_COUNTER_STATE_ERROR;
674 spin_unlock_irqrestore(&ctx->lock, flags);
680 static void perf_counter_enable(struct perf_counter *counter)
682 struct perf_counter_context *ctx = counter->ctx;
683 struct task_struct *task = ctx->task;
687 * Enable the counter on the cpu that it's on
689 smp_call_function_single(counter->cpu, __perf_counter_enable,
694 spin_lock_irq(&ctx->lock);
695 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
699 * If the counter is in error state, clear that first.
700 * That way, if we see the counter in error state below, we
701 * know that it has gone back into error state, as distinct
702 * from the task having been scheduled away before the
703 * cross-call arrived.
705 if (counter->state == PERF_COUNTER_STATE_ERROR)
706 counter->state = PERF_COUNTER_STATE_OFF;
709 spin_unlock_irq(&ctx->lock);
710 task_oncpu_function_call(task, __perf_counter_enable, counter);
712 spin_lock_irq(&ctx->lock);
715 * If the context is active and the counter is still off,
716 * we need to retry the cross-call.
718 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
722 * Since we have the lock this context can't be scheduled
723 * in, so we can change the state safely.
725 if (counter->state == PERF_COUNTER_STATE_OFF) {
726 counter->state = PERF_COUNTER_STATE_INACTIVE;
727 counter->tstamp_enabled =
728 ctx->time - counter->total_time_enabled;
731 spin_unlock_irq(&ctx->lock);
734 static void perf_counter_refresh(struct perf_counter *counter, int refresh)
736 atomic_add(refresh, &counter->event_limit);
737 perf_counter_enable(counter);
741 * Enable a counter and all its children.
743 static void perf_counter_enable_family(struct perf_counter *counter)
745 struct perf_counter *child;
747 perf_counter_enable(counter);
750 * Lock the mutex to protect the list of children
752 mutex_lock(&counter->mutex);
753 list_for_each_entry(child, &counter->child_list, child_list)
754 perf_counter_enable(child);
755 mutex_unlock(&counter->mutex);
758 void __perf_counter_sched_out(struct perf_counter_context *ctx,
759 struct perf_cpu_context *cpuctx)
761 struct perf_counter *counter;
764 spin_lock(&ctx->lock);
766 if (likely(!ctx->nr_counters))
768 update_context_time(ctx);
770 flags = hw_perf_save_disable();
771 if (ctx->nr_active) {
772 list_for_each_entry(counter, &ctx->counter_list, list_entry)
773 group_sched_out(counter, cpuctx, ctx);
775 hw_perf_restore(flags);
777 spin_unlock(&ctx->lock);
781 * Called from scheduler to remove the counters of the current task,
782 * with interrupts disabled.
784 * We stop each counter and update the counter value in counter->count.
786 * This does not protect us against NMI, but disable()
787 * sets the disabled bit in the control field of counter _before_
788 * accessing the counter control register. If a NMI hits, then it will
789 * not restart the counter.
791 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
793 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
794 struct perf_counter_context *ctx = &task->perf_counter_ctx;
795 struct pt_regs *regs;
797 if (likely(!cpuctx->task_ctx))
800 update_context_time(ctx);
802 regs = task_pt_regs(task);
803 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs);
804 __perf_counter_sched_out(ctx, cpuctx);
806 cpuctx->task_ctx = NULL;
809 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
811 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
815 group_sched_in(struct perf_counter *group_counter,
816 struct perf_cpu_context *cpuctx,
817 struct perf_counter_context *ctx,
820 struct perf_counter *counter, *partial_group;
823 if (group_counter->state == PERF_COUNTER_STATE_OFF)
826 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
828 return ret < 0 ? ret : 0;
830 group_counter->prev_state = group_counter->state;
831 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
835 * Schedule in siblings as one group (if any):
837 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
838 counter->prev_state = counter->state;
839 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
840 partial_group = counter;
849 * Groups can be scheduled in as one unit only, so undo any
850 * partial group before returning:
852 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
853 if (counter == partial_group)
855 counter_sched_out(counter, cpuctx, ctx);
857 counter_sched_out(group_counter, cpuctx, ctx);
863 __perf_counter_sched_in(struct perf_counter_context *ctx,
864 struct perf_cpu_context *cpuctx, int cpu)
866 struct perf_counter *counter;
870 spin_lock(&ctx->lock);
872 if (likely(!ctx->nr_counters))
875 ctx->timestamp = perf_clock();
877 flags = hw_perf_save_disable();
880 * First go through the list and put on any pinned groups
881 * in order to give them the best chance of going on.
883 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
884 if (counter->state <= PERF_COUNTER_STATE_OFF ||
885 !counter->hw_event.pinned)
887 if (counter->cpu != -1 && counter->cpu != cpu)
890 if (group_can_go_on(counter, cpuctx, 1))
891 group_sched_in(counter, cpuctx, ctx, cpu);
894 * If this pinned group hasn't been scheduled,
895 * put it in error state.
897 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
898 update_group_times(counter);
899 counter->state = PERF_COUNTER_STATE_ERROR;
903 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
905 * Ignore counters in OFF or ERROR state, and
906 * ignore pinned counters since we did them already.
908 if (counter->state <= PERF_COUNTER_STATE_OFF ||
909 counter->hw_event.pinned)
913 * Listen to the 'cpu' scheduling filter constraint
916 if (counter->cpu != -1 && counter->cpu != cpu)
919 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
920 if (group_sched_in(counter, cpuctx, ctx, cpu))
924 hw_perf_restore(flags);
926 spin_unlock(&ctx->lock);
930 * Called from scheduler to add the counters of the current task
931 * with interrupts disabled.
933 * We restore the counter value and then enable it.
935 * This does not protect us against NMI, but enable()
936 * sets the enabled bit in the control field of counter _before_
937 * accessing the counter control register. If a NMI hits, then it will
938 * keep the counter running.
940 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
942 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
943 struct perf_counter_context *ctx = &task->perf_counter_ctx;
945 __perf_counter_sched_in(ctx, cpuctx, cpu);
946 cpuctx->task_ctx = ctx;
949 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
951 struct perf_counter_context *ctx = &cpuctx->ctx;
953 __perf_counter_sched_in(ctx, cpuctx, cpu);
956 int perf_counter_task_disable(void)
958 struct task_struct *curr = current;
959 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
960 struct perf_counter *counter;
965 if (likely(!ctx->nr_counters))
968 local_irq_save(flags);
969 cpu = smp_processor_id();
971 perf_counter_task_sched_out(curr, cpu);
973 spin_lock(&ctx->lock);
976 * Disable all the counters:
978 perf_flags = hw_perf_save_disable();
980 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
981 if (counter->state != PERF_COUNTER_STATE_ERROR) {
982 update_group_times(counter);
983 counter->state = PERF_COUNTER_STATE_OFF;
987 hw_perf_restore(perf_flags);
989 spin_unlock_irqrestore(&ctx->lock, flags);
994 int perf_counter_task_enable(void)
996 struct task_struct *curr = current;
997 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
998 struct perf_counter *counter;
1003 if (likely(!ctx->nr_counters))
1006 local_irq_save(flags);
1007 cpu = smp_processor_id();
1009 perf_counter_task_sched_out(curr, cpu);
1011 spin_lock(&ctx->lock);
1014 * Disable all the counters:
1016 perf_flags = hw_perf_save_disable();
1018 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1019 if (counter->state > PERF_COUNTER_STATE_OFF)
1021 counter->state = PERF_COUNTER_STATE_INACTIVE;
1022 counter->tstamp_enabled =
1023 ctx->time - counter->total_time_enabled;
1024 counter->hw_event.disabled = 0;
1026 hw_perf_restore(perf_flags);
1028 spin_unlock(&ctx->lock);
1030 perf_counter_task_sched_in(curr, cpu);
1032 local_irq_restore(flags);
1038 * Round-robin a context's counters:
1040 static void rotate_ctx(struct perf_counter_context *ctx)
1042 struct perf_counter *counter;
1045 if (!ctx->nr_counters)
1048 spin_lock(&ctx->lock);
1050 * Rotate the first entry last (works just fine for group counters too):
1052 perf_flags = hw_perf_save_disable();
1053 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1054 list_move_tail(&counter->list_entry, &ctx->counter_list);
1057 hw_perf_restore(perf_flags);
1059 spin_unlock(&ctx->lock);
1062 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1064 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1065 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1066 const int rotate_percpu = 0;
1069 perf_counter_cpu_sched_out(cpuctx);
1070 perf_counter_task_sched_out(curr, cpu);
1073 rotate_ctx(&cpuctx->ctx);
1077 perf_counter_cpu_sched_in(cpuctx, cpu);
1078 perf_counter_task_sched_in(curr, cpu);
1082 * Cross CPU call to read the hardware counter
1084 static void __read(void *info)
1086 struct perf_counter *counter = info;
1087 struct perf_counter_context *ctx = counter->ctx;
1088 unsigned long flags;
1090 local_irq_save(flags);
1092 update_context_time(ctx);
1093 counter->hw_ops->read(counter);
1094 update_counter_times(counter);
1095 local_irq_restore(flags);
1098 static u64 perf_counter_read(struct perf_counter *counter)
1101 * If counter is enabled and currently active on a CPU, update the
1102 * value in the counter structure:
1104 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1105 smp_call_function_single(counter->oncpu,
1106 __read, counter, 1);
1107 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1108 update_counter_times(counter);
1111 return atomic64_read(&counter->count);
1114 static void put_context(struct perf_counter_context *ctx)
1117 put_task_struct(ctx->task);
1120 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1122 struct perf_cpu_context *cpuctx;
1123 struct perf_counter_context *ctx;
1124 struct task_struct *task;
1127 * If cpu is not a wildcard then this is a percpu counter:
1130 /* Must be root to operate on a CPU counter: */
1131 if (!capable(CAP_SYS_ADMIN))
1132 return ERR_PTR(-EACCES);
1134 if (cpu < 0 || cpu > num_possible_cpus())
1135 return ERR_PTR(-EINVAL);
1138 * We could be clever and allow to attach a counter to an
1139 * offline CPU and activate it when the CPU comes up, but
1142 if (!cpu_isset(cpu, cpu_online_map))
1143 return ERR_PTR(-ENODEV);
1145 cpuctx = &per_cpu(perf_cpu_context, cpu);
1155 task = find_task_by_vpid(pid);
1157 get_task_struct(task);
1161 return ERR_PTR(-ESRCH);
1163 ctx = &task->perf_counter_ctx;
1166 /* Reuse ptrace permission checks for now. */
1167 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1169 return ERR_PTR(-EACCES);
1175 static void free_counter_rcu(struct rcu_head *head)
1177 struct perf_counter *counter;
1179 counter = container_of(head, struct perf_counter, rcu_head);
1183 static void perf_pending_sync(struct perf_counter *counter);
1185 static void free_counter(struct perf_counter *counter)
1187 perf_pending_sync(counter);
1189 if (counter->destroy)
1190 counter->destroy(counter);
1192 call_rcu(&counter->rcu_head, free_counter_rcu);
1196 * Called when the last reference to the file is gone.
1198 static int perf_release(struct inode *inode, struct file *file)
1200 struct perf_counter *counter = file->private_data;
1201 struct perf_counter_context *ctx = counter->ctx;
1203 file->private_data = NULL;
1205 mutex_lock(&ctx->mutex);
1206 mutex_lock(&counter->mutex);
1208 perf_counter_remove_from_context(counter);
1210 mutex_unlock(&counter->mutex);
1211 mutex_unlock(&ctx->mutex);
1213 free_counter(counter);
1220 * Read the performance counter - simple non blocking version for now
1223 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1229 * Return end-of-file for a read on a counter that is in
1230 * error state (i.e. because it was pinned but it couldn't be
1231 * scheduled on to the CPU at some point).
1233 if (counter->state == PERF_COUNTER_STATE_ERROR)
1236 mutex_lock(&counter->mutex);
1237 values[0] = perf_counter_read(counter);
1239 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1240 values[n++] = counter->total_time_enabled +
1241 atomic64_read(&counter->child_total_time_enabled);
1242 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1243 values[n++] = counter->total_time_running +
1244 atomic64_read(&counter->child_total_time_running);
1245 mutex_unlock(&counter->mutex);
1247 if (count < n * sizeof(u64))
1249 count = n * sizeof(u64);
1251 if (copy_to_user(buf, values, count))
1258 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1260 struct perf_counter *counter = file->private_data;
1262 return perf_read_hw(counter, buf, count);
1265 static unsigned int perf_poll(struct file *file, poll_table *wait)
1267 struct perf_counter *counter = file->private_data;
1268 struct perf_mmap_data *data;
1269 unsigned int events;
1272 data = rcu_dereference(counter->data);
1274 events = atomic_xchg(&data->wakeup, 0);
1279 poll_wait(file, &counter->waitq, wait);
1284 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1286 struct perf_counter *counter = file->private_data;
1290 case PERF_COUNTER_IOC_ENABLE:
1291 perf_counter_enable_family(counter);
1293 case PERF_COUNTER_IOC_DISABLE:
1294 perf_counter_disable_family(counter);
1296 case PERF_COUNTER_IOC_REFRESH:
1297 perf_counter_refresh(counter, arg);
1306 * Callers need to ensure there can be no nesting of this function, otherwise
1307 * the seqlock logic goes bad. We can not serialize this because the arch
1308 * code calls this from NMI context.
1310 void perf_counter_update_userpage(struct perf_counter *counter)
1312 struct perf_mmap_data *data;
1313 struct perf_counter_mmap_page *userpg;
1316 data = rcu_dereference(counter->data);
1320 userpg = data->user_page;
1323 * Disable preemption so as to not let the corresponding user-space
1324 * spin too long if we get preempted.
1329 userpg->index = counter->hw.idx;
1330 userpg->offset = atomic64_read(&counter->count);
1331 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1332 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1341 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1343 struct perf_counter *counter = vma->vm_file->private_data;
1344 struct perf_mmap_data *data;
1345 int ret = VM_FAULT_SIGBUS;
1348 data = rcu_dereference(counter->data);
1352 if (vmf->pgoff == 0) {
1353 vmf->page = virt_to_page(data->user_page);
1355 int nr = vmf->pgoff - 1;
1357 if ((unsigned)nr > data->nr_pages)
1360 vmf->page = virt_to_page(data->data_pages[nr]);
1362 get_page(vmf->page);
1370 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1372 struct perf_mmap_data *data;
1376 WARN_ON(atomic_read(&counter->mmap_count));
1378 size = sizeof(struct perf_mmap_data);
1379 size += nr_pages * sizeof(void *);
1381 data = kzalloc(size, GFP_KERNEL);
1385 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1386 if (!data->user_page)
1387 goto fail_user_page;
1389 for (i = 0; i < nr_pages; i++) {
1390 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1391 if (!data->data_pages[i])
1392 goto fail_data_pages;
1395 data->nr_pages = nr_pages;
1397 rcu_assign_pointer(counter->data, data);
1402 for (i--; i >= 0; i--)
1403 free_page((unsigned long)data->data_pages[i]);
1405 free_page((unsigned long)data->user_page);
1414 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1416 struct perf_mmap_data *data = container_of(rcu_head,
1417 struct perf_mmap_data, rcu_head);
1420 free_page((unsigned long)data->user_page);
1421 for (i = 0; i < data->nr_pages; i++)
1422 free_page((unsigned long)data->data_pages[i]);
1426 static void perf_mmap_data_free(struct perf_counter *counter)
1428 struct perf_mmap_data *data = counter->data;
1430 WARN_ON(atomic_read(&counter->mmap_count));
1432 rcu_assign_pointer(counter->data, NULL);
1433 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1436 static void perf_mmap_open(struct vm_area_struct *vma)
1438 struct perf_counter *counter = vma->vm_file->private_data;
1440 atomic_inc(&counter->mmap_count);
1443 static void perf_mmap_close(struct vm_area_struct *vma)
1445 struct perf_counter *counter = vma->vm_file->private_data;
1447 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1448 &counter->mmap_mutex)) {
1449 vma->vm_mm->locked_vm -= counter->data->nr_pages + 1;
1450 perf_mmap_data_free(counter);
1451 mutex_unlock(&counter->mmap_mutex);
1455 static struct vm_operations_struct perf_mmap_vmops = {
1456 .open = perf_mmap_open,
1457 .close = perf_mmap_close,
1458 .fault = perf_mmap_fault,
1461 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1463 struct perf_counter *counter = file->private_data;
1464 unsigned long vma_size;
1465 unsigned long nr_pages;
1466 unsigned long locked, lock_limit;
1469 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1472 vma_size = vma->vm_end - vma->vm_start;
1473 nr_pages = (vma_size / PAGE_SIZE) - 1;
1476 * If we have data pages ensure they're a power-of-two number, so we
1477 * can do bitmasks instead of modulo.
1479 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1482 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1485 if (vma->vm_pgoff != 0)
1488 mutex_lock(&counter->mmap_mutex);
1489 if (atomic_inc_not_zero(&counter->mmap_count)) {
1490 if (nr_pages != counter->data->nr_pages)
1495 locked = vma->vm_mm->locked_vm;
1496 locked += nr_pages + 1;
1498 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1499 lock_limit >>= PAGE_SHIFT;
1501 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1506 WARN_ON(counter->data);
1507 ret = perf_mmap_data_alloc(counter, nr_pages);
1511 atomic_set(&counter->mmap_count, 1);
1512 vma->vm_mm->locked_vm += nr_pages + 1;
1514 mutex_unlock(&counter->mmap_mutex);
1516 vma->vm_flags &= ~VM_MAYWRITE;
1517 vma->vm_flags |= VM_RESERVED;
1518 vma->vm_ops = &perf_mmap_vmops;
1523 static int perf_fasync(int fd, struct file *filp, int on)
1525 struct perf_counter *counter = filp->private_data;
1526 struct inode *inode = filp->f_path.dentry->d_inode;
1529 mutex_lock(&inode->i_mutex);
1530 retval = fasync_helper(fd, filp, on, &counter->fasync);
1531 mutex_unlock(&inode->i_mutex);
1539 static const struct file_operations perf_fops = {
1540 .release = perf_release,
1543 .unlocked_ioctl = perf_ioctl,
1544 .compat_ioctl = perf_ioctl,
1546 .fasync = perf_fasync,
1550 * Perf counter wakeup
1552 * If there's data, ensure we set the poll() state and publish everything
1553 * to user-space before waking everybody up.
1556 void perf_counter_wakeup(struct perf_counter *counter)
1558 struct perf_mmap_data *data;
1561 data = rcu_dereference(counter->data);
1563 atomic_set(&data->wakeup, POLL_IN);
1565 * Ensure all data writes are issued before updating the
1566 * user-space data head information. The matching rmb()
1567 * will be in userspace after reading this value.
1570 data->user_page->data_head = atomic_read(&data->head);
1574 wake_up_all(&counter->waitq);
1576 if (counter->pending_kill) {
1577 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1578 counter->pending_kill = 0;
1585 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1587 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1588 * single linked list and use cmpxchg() to add entries lockless.
1591 static void perf_pending_counter(struct perf_pending_entry *entry)
1593 struct perf_counter *counter = container_of(entry,
1594 struct perf_counter, pending);
1596 if (counter->pending_disable) {
1597 counter->pending_disable = 0;
1598 perf_counter_disable(counter);
1601 if (counter->pending_wakeup) {
1602 counter->pending_wakeup = 0;
1603 perf_counter_wakeup(counter);
1607 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1609 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1613 static void perf_pending_queue(struct perf_pending_entry *entry,
1614 void (*func)(struct perf_pending_entry *))
1616 struct perf_pending_entry **head;
1618 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1623 head = &get_cpu_var(perf_pending_head);
1626 entry->next = *head;
1627 } while (cmpxchg(head, entry->next, entry) != entry->next);
1629 set_perf_counter_pending();
1631 put_cpu_var(perf_pending_head);
1634 static int __perf_pending_run(void)
1636 struct perf_pending_entry *list;
1639 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1640 while (list != PENDING_TAIL) {
1641 void (*func)(struct perf_pending_entry *);
1642 struct perf_pending_entry *entry = list;
1649 * Ensure we observe the unqueue before we issue the wakeup,
1650 * so that we won't be waiting forever.
1651 * -- see perf_not_pending().
1662 static inline int perf_not_pending(struct perf_counter *counter)
1665 * If we flush on whatever cpu we run, there is a chance we don't
1669 __perf_pending_run();
1673 * Ensure we see the proper queue state before going to sleep
1674 * so that we do not miss the wakeup. -- see perf_pending_handle()
1677 return counter->pending.next == NULL;
1680 static void perf_pending_sync(struct perf_counter *counter)
1682 wait_event(counter->waitq, perf_not_pending(counter));
1685 void perf_counter_do_pending(void)
1687 __perf_pending_run();
1691 * Callchain support -- arch specific
1694 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1703 struct perf_output_handle {
1704 struct perf_counter *counter;
1705 struct perf_mmap_data *data;
1706 unsigned int offset;
1713 static inline void __perf_output_wakeup(struct perf_output_handle *handle)
1716 handle->counter->pending_wakeup = 1;
1717 perf_pending_queue(&handle->counter->pending,
1718 perf_pending_counter);
1720 perf_counter_wakeup(handle->counter);
1723 static int perf_output_begin(struct perf_output_handle *handle,
1724 struct perf_counter *counter, unsigned int size,
1725 int nmi, int overflow)
1727 struct perf_mmap_data *data;
1728 unsigned int offset, head;
1731 data = rcu_dereference(counter->data);
1735 handle->counter = counter;
1737 handle->overflow = overflow;
1739 if (!data->nr_pages)
1743 offset = head = atomic_read(&data->head);
1745 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1747 handle->data = data;
1748 handle->offset = offset;
1749 handle->head = head;
1750 handle->wakeup = (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT);
1755 __perf_output_wakeup(handle);
1762 static void perf_output_copy(struct perf_output_handle *handle,
1763 void *buf, unsigned int len)
1765 unsigned int pages_mask;
1766 unsigned int offset;
1770 offset = handle->offset;
1771 pages_mask = handle->data->nr_pages - 1;
1772 pages = handle->data->data_pages;
1775 unsigned int page_offset;
1778 nr = (offset >> PAGE_SHIFT) & pages_mask;
1779 page_offset = offset & (PAGE_SIZE - 1);
1780 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1782 memcpy(pages[nr] + page_offset, buf, size);
1789 handle->offset = offset;
1791 WARN_ON_ONCE(handle->offset > handle->head);
1794 #define perf_output_put(handle, x) \
1795 perf_output_copy((handle), &(x), sizeof(x))
1797 static void perf_output_end(struct perf_output_handle *handle)
1799 int wakeup_events = handle->counter->hw_event.wakeup_events;
1801 if (handle->overflow && wakeup_events) {
1802 int events = atomic_inc_return(&handle->data->events);
1803 if (events >= wakeup_events) {
1804 atomic_sub(wakeup_events, &handle->data->events);
1805 __perf_output_wakeup(handle);
1807 } else if (handle->wakeup)
1808 __perf_output_wakeup(handle);
1812 static void perf_counter_output(struct perf_counter *counter,
1813 int nmi, struct pt_regs *regs)
1816 u64 record_type = counter->hw_event.record_type;
1817 struct perf_output_handle handle;
1818 struct perf_event_header header;
1827 struct perf_callchain_entry *callchain = NULL;
1828 int callchain_size = 0;
1832 header.size = sizeof(header);
1834 header.misc = PERF_EVENT_MISC_OVERFLOW;
1835 header.misc |= user_mode(regs) ?
1836 PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
1838 if (record_type & PERF_RECORD_IP) {
1839 ip = instruction_pointer(regs);
1840 header.type |= PERF_RECORD_IP;
1841 header.size += sizeof(ip);
1844 if (record_type & PERF_RECORD_TID) {
1845 /* namespace issues */
1846 tid_entry.pid = current->group_leader->pid;
1847 tid_entry.tid = current->pid;
1849 header.type |= PERF_RECORD_TID;
1850 header.size += sizeof(tid_entry);
1853 if (record_type & PERF_RECORD_GROUP) {
1854 header.type |= PERF_RECORD_GROUP;
1855 header.size += sizeof(u64) +
1856 counter->nr_siblings * sizeof(group_entry);
1859 if (record_type & PERF_RECORD_CALLCHAIN) {
1860 callchain = perf_callchain(regs);
1863 callchain_size = (1 + callchain->nr) * sizeof(u64);
1865 header.type |= PERF_RECORD_CALLCHAIN;
1866 header.size += callchain_size;
1870 if (record_type & PERF_RECORD_TIME) {
1872 * Maybe do better on x86 and provide cpu_clock_nmi()
1874 time = sched_clock();
1876 header.type |= PERF_RECORD_TIME;
1877 header.size += sizeof(u64);
1880 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
1884 perf_output_put(&handle, header);
1886 if (record_type & PERF_RECORD_IP)
1887 perf_output_put(&handle, ip);
1889 if (record_type & PERF_RECORD_TID)
1890 perf_output_put(&handle, tid_entry);
1892 if (record_type & PERF_RECORD_GROUP) {
1893 struct perf_counter *leader, *sub;
1894 u64 nr = counter->nr_siblings;
1896 perf_output_put(&handle, nr);
1898 leader = counter->group_leader;
1899 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1901 sub->hw_ops->read(sub);
1903 group_entry.event = sub->hw_event.config;
1904 group_entry.counter = atomic64_read(&sub->count);
1906 perf_output_put(&handle, group_entry);
1911 perf_output_copy(&handle, callchain, callchain_size);
1913 if (record_type & PERF_RECORD_TIME)
1914 perf_output_put(&handle, time);
1916 perf_output_end(&handle);
1923 struct perf_comm_event {
1924 struct task_struct *task;
1929 struct perf_event_header header;
1936 static void perf_counter_comm_output(struct perf_counter *counter,
1937 struct perf_comm_event *comm_event)
1939 struct perf_output_handle handle;
1940 int size = comm_event->event.header.size;
1941 int ret = perf_output_begin(&handle, counter, size, 0, 0);
1946 perf_output_put(&handle, comm_event->event);
1947 perf_output_copy(&handle, comm_event->comm,
1948 comm_event->comm_size);
1949 perf_output_end(&handle);
1952 static int perf_counter_comm_match(struct perf_counter *counter,
1953 struct perf_comm_event *comm_event)
1955 if (counter->hw_event.comm &&
1956 comm_event->event.header.type == PERF_EVENT_COMM)
1962 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
1963 struct perf_comm_event *comm_event)
1965 struct perf_counter *counter;
1967 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
1971 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
1972 if (perf_counter_comm_match(counter, comm_event))
1973 perf_counter_comm_output(counter, comm_event);
1978 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
1980 struct perf_cpu_context *cpuctx;
1982 char *comm = comm_event->task->comm;
1984 size = ALIGN(strlen(comm), sizeof(u64));
1986 comm_event->comm = comm;
1987 comm_event->comm_size = size;
1989 comm_event->event.header.size = sizeof(comm_event->event) + size;
1991 cpuctx = &get_cpu_var(perf_cpu_context);
1992 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
1993 put_cpu_var(perf_cpu_context);
1995 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
1998 void perf_counter_comm(struct task_struct *task)
2000 struct perf_comm_event comm_event = {
2003 .header = { .type = PERF_EVENT_COMM, },
2004 .pid = task->group_leader->pid,
2009 perf_counter_comm_event(&comm_event);
2016 struct perf_mmap_event {
2022 struct perf_event_header header;
2032 static void perf_counter_mmap_output(struct perf_counter *counter,
2033 struct perf_mmap_event *mmap_event)
2035 struct perf_output_handle handle;
2036 int size = mmap_event->event.header.size;
2037 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2042 perf_output_put(&handle, mmap_event->event);
2043 perf_output_copy(&handle, mmap_event->file_name,
2044 mmap_event->file_size);
2045 perf_output_end(&handle);
2048 static int perf_counter_mmap_match(struct perf_counter *counter,
2049 struct perf_mmap_event *mmap_event)
2051 if (counter->hw_event.mmap &&
2052 mmap_event->event.header.type == PERF_EVENT_MMAP)
2055 if (counter->hw_event.munmap &&
2056 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2062 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2063 struct perf_mmap_event *mmap_event)
2065 struct perf_counter *counter;
2067 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2071 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2072 if (perf_counter_mmap_match(counter, mmap_event))
2073 perf_counter_mmap_output(counter, mmap_event);
2078 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2080 struct perf_cpu_context *cpuctx;
2081 struct file *file = mmap_event->file;
2088 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2090 name = strncpy(tmp, "//enomem", sizeof(tmp));
2093 name = dentry_path(file->f_dentry, buf, PATH_MAX);
2095 name = strncpy(tmp, "//toolong", sizeof(tmp));
2099 name = strncpy(tmp, "//anon", sizeof(tmp));
2104 size = ALIGN(strlen(name), sizeof(u64));
2106 mmap_event->file_name = name;
2107 mmap_event->file_size = size;
2109 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2111 cpuctx = &get_cpu_var(perf_cpu_context);
2112 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2113 put_cpu_var(perf_cpu_context);
2115 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2120 void perf_counter_mmap(unsigned long addr, unsigned long len,
2121 unsigned long pgoff, struct file *file)
2123 struct perf_mmap_event mmap_event = {
2126 .header = { .type = PERF_EVENT_MMAP, },
2127 .pid = current->group_leader->pid,
2128 .tid = current->pid,
2135 perf_counter_mmap_event(&mmap_event);
2138 void perf_counter_munmap(unsigned long addr, unsigned long len,
2139 unsigned long pgoff, struct file *file)
2141 struct perf_mmap_event mmap_event = {
2144 .header = { .type = PERF_EVENT_MUNMAP, },
2145 .pid = current->group_leader->pid,
2146 .tid = current->pid,
2153 perf_counter_mmap_event(&mmap_event);
2157 * Generic counter overflow handling.
2160 int perf_counter_overflow(struct perf_counter *counter,
2161 int nmi, struct pt_regs *regs)
2163 int events = atomic_read(&counter->event_limit);
2166 counter->pending_kill = POLL_IN;
2167 if (events && atomic_dec_and_test(&counter->event_limit)) {
2169 counter->pending_kill = POLL_HUP;
2171 counter->pending_disable = 1;
2172 perf_pending_queue(&counter->pending,
2173 perf_pending_counter);
2175 perf_counter_disable(counter);
2178 perf_counter_output(counter, nmi, regs);
2183 * Generic software counter infrastructure
2186 static void perf_swcounter_update(struct perf_counter *counter)
2188 struct hw_perf_counter *hwc = &counter->hw;
2193 prev = atomic64_read(&hwc->prev_count);
2194 now = atomic64_read(&hwc->count);
2195 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2200 atomic64_add(delta, &counter->count);
2201 atomic64_sub(delta, &hwc->period_left);
2204 static void perf_swcounter_set_period(struct perf_counter *counter)
2206 struct hw_perf_counter *hwc = &counter->hw;
2207 s64 left = atomic64_read(&hwc->period_left);
2208 s64 period = hwc->irq_period;
2210 if (unlikely(left <= -period)) {
2212 atomic64_set(&hwc->period_left, left);
2215 if (unlikely(left <= 0)) {
2217 atomic64_add(period, &hwc->period_left);
2220 atomic64_set(&hwc->prev_count, -left);
2221 atomic64_set(&hwc->count, -left);
2224 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2226 enum hrtimer_restart ret = HRTIMER_RESTART;
2227 struct perf_counter *counter;
2228 struct pt_regs *regs;
2230 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2231 counter->hw_ops->read(counter);
2233 regs = get_irq_regs();
2235 * In case we exclude kernel IPs or are somehow not in interrupt
2236 * context, provide the next best thing, the user IP.
2238 if ((counter->hw_event.exclude_kernel || !regs) &&
2239 !counter->hw_event.exclude_user)
2240 regs = task_pt_regs(current);
2243 if (perf_counter_overflow(counter, 0, regs))
2244 ret = HRTIMER_NORESTART;
2247 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2252 static void perf_swcounter_overflow(struct perf_counter *counter,
2253 int nmi, struct pt_regs *regs)
2255 perf_swcounter_update(counter);
2256 perf_swcounter_set_period(counter);
2257 if (perf_counter_overflow(counter, nmi, regs))
2258 /* soft-disable the counter */
2263 static int perf_swcounter_match(struct perf_counter *counter,
2264 enum perf_event_types type,
2265 u32 event, struct pt_regs *regs)
2267 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2270 if (perf_event_raw(&counter->hw_event))
2273 if (perf_event_type(&counter->hw_event) != type)
2276 if (perf_event_id(&counter->hw_event) != event)
2279 if (counter->hw_event.exclude_user && user_mode(regs))
2282 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2288 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2289 int nmi, struct pt_regs *regs)
2291 int neg = atomic64_add_negative(nr, &counter->hw.count);
2292 if (counter->hw.irq_period && !neg)
2293 perf_swcounter_overflow(counter, nmi, regs);
2296 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2297 enum perf_event_types type, u32 event,
2298 u64 nr, int nmi, struct pt_regs *regs)
2300 struct perf_counter *counter;
2302 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2306 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2307 if (perf_swcounter_match(counter, type, event, regs))
2308 perf_swcounter_add(counter, nr, nmi, regs);
2313 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2316 return &cpuctx->recursion[3];
2319 return &cpuctx->recursion[2];
2322 return &cpuctx->recursion[1];
2324 return &cpuctx->recursion[0];
2327 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2328 u64 nr, int nmi, struct pt_regs *regs)
2330 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2331 int *recursion = perf_swcounter_recursion_context(cpuctx);
2339 perf_swcounter_ctx_event(&cpuctx->ctx, type, event, nr, nmi, regs);
2340 if (cpuctx->task_ctx) {
2341 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2349 put_cpu_var(perf_cpu_context);
2352 void perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)
2354 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs);
2357 static void perf_swcounter_read(struct perf_counter *counter)
2359 perf_swcounter_update(counter);
2362 static int perf_swcounter_enable(struct perf_counter *counter)
2364 perf_swcounter_set_period(counter);
2368 static void perf_swcounter_disable(struct perf_counter *counter)
2370 perf_swcounter_update(counter);
2373 static const struct hw_perf_counter_ops perf_ops_generic = {
2374 .enable = perf_swcounter_enable,
2375 .disable = perf_swcounter_disable,
2376 .read = perf_swcounter_read,
2380 * Software counter: cpu wall time clock
2383 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2385 int cpu = raw_smp_processor_id();
2389 now = cpu_clock(cpu);
2390 prev = atomic64_read(&counter->hw.prev_count);
2391 atomic64_set(&counter->hw.prev_count, now);
2392 atomic64_add(now - prev, &counter->count);
2395 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2397 struct hw_perf_counter *hwc = &counter->hw;
2398 int cpu = raw_smp_processor_id();
2400 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2401 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2402 hwc->hrtimer.function = perf_swcounter_hrtimer;
2403 if (hwc->irq_period) {
2404 __hrtimer_start_range_ns(&hwc->hrtimer,
2405 ns_to_ktime(hwc->irq_period), 0,
2406 HRTIMER_MODE_REL, 0);
2412 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2414 hrtimer_cancel(&counter->hw.hrtimer);
2415 cpu_clock_perf_counter_update(counter);
2418 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2420 cpu_clock_perf_counter_update(counter);
2423 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
2424 .enable = cpu_clock_perf_counter_enable,
2425 .disable = cpu_clock_perf_counter_disable,
2426 .read = cpu_clock_perf_counter_read,
2430 * Software counter: task time clock
2433 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2438 prev = atomic64_xchg(&counter->hw.prev_count, now);
2440 atomic64_add(delta, &counter->count);
2443 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2445 struct hw_perf_counter *hwc = &counter->hw;
2448 now = counter->ctx->time;
2450 atomic64_set(&hwc->prev_count, now);
2451 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2452 hwc->hrtimer.function = perf_swcounter_hrtimer;
2453 if (hwc->irq_period) {
2454 __hrtimer_start_range_ns(&hwc->hrtimer,
2455 ns_to_ktime(hwc->irq_period), 0,
2456 HRTIMER_MODE_REL, 0);
2462 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2464 hrtimer_cancel(&counter->hw.hrtimer);
2465 task_clock_perf_counter_update(counter, counter->ctx->time);
2469 static void task_clock_perf_counter_read(struct perf_counter *counter)
2474 update_context_time(counter->ctx);
2475 time = counter->ctx->time;
2477 u64 now = perf_clock();
2478 u64 delta = now - counter->ctx->timestamp;
2479 time = counter->ctx->time + delta;
2482 task_clock_perf_counter_update(counter, time);
2485 static const struct hw_perf_counter_ops perf_ops_task_clock = {
2486 .enable = task_clock_perf_counter_enable,
2487 .disable = task_clock_perf_counter_disable,
2488 .read = task_clock_perf_counter_read,
2492 * Software counter: cpu migrations
2495 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2497 struct task_struct *curr = counter->ctx->task;
2500 return curr->se.nr_migrations;
2501 return cpu_nr_migrations(smp_processor_id());
2504 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2509 prev = atomic64_read(&counter->hw.prev_count);
2510 now = get_cpu_migrations(counter);
2512 atomic64_set(&counter->hw.prev_count, now);
2516 atomic64_add(delta, &counter->count);
2519 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2521 cpu_migrations_perf_counter_update(counter);
2524 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2526 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2527 atomic64_set(&counter->hw.prev_count,
2528 get_cpu_migrations(counter));
2532 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2534 cpu_migrations_perf_counter_update(counter);
2537 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
2538 .enable = cpu_migrations_perf_counter_enable,
2539 .disable = cpu_migrations_perf_counter_disable,
2540 .read = cpu_migrations_perf_counter_read,
2543 #ifdef CONFIG_EVENT_PROFILE
2544 void perf_tpcounter_event(int event_id)
2546 struct pt_regs *regs = get_irq_regs();
2549 regs = task_pt_regs(current);
2551 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs);
2554 extern int ftrace_profile_enable(int);
2555 extern void ftrace_profile_disable(int);
2557 static void tp_perf_counter_destroy(struct perf_counter *counter)
2559 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2562 static const struct hw_perf_counter_ops *
2563 tp_perf_counter_init(struct perf_counter *counter)
2565 int event_id = perf_event_id(&counter->hw_event);
2568 ret = ftrace_profile_enable(event_id);
2572 counter->destroy = tp_perf_counter_destroy;
2573 counter->hw.irq_period = counter->hw_event.irq_period;
2575 return &perf_ops_generic;
2578 static const struct hw_perf_counter_ops *
2579 tp_perf_counter_init(struct perf_counter *counter)
2585 static const struct hw_perf_counter_ops *
2586 sw_perf_counter_init(struct perf_counter *counter)
2588 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2589 const struct hw_perf_counter_ops *hw_ops = NULL;
2590 struct hw_perf_counter *hwc = &counter->hw;
2593 * Software counters (currently) can't in general distinguish
2594 * between user, kernel and hypervisor events.
2595 * However, context switches and cpu migrations are considered
2596 * to be kernel events, and page faults are never hypervisor
2599 switch (perf_event_id(&counter->hw_event)) {
2600 case PERF_COUNT_CPU_CLOCK:
2601 hw_ops = &perf_ops_cpu_clock;
2603 if (hw_event->irq_period && hw_event->irq_period < 10000)
2604 hw_event->irq_period = 10000;
2606 case PERF_COUNT_TASK_CLOCK:
2608 * If the user instantiates this as a per-cpu counter,
2609 * use the cpu_clock counter instead.
2611 if (counter->ctx->task)
2612 hw_ops = &perf_ops_task_clock;
2614 hw_ops = &perf_ops_cpu_clock;
2616 if (hw_event->irq_period && hw_event->irq_period < 10000)
2617 hw_event->irq_period = 10000;
2619 case PERF_COUNT_PAGE_FAULTS:
2620 case PERF_COUNT_PAGE_FAULTS_MIN:
2621 case PERF_COUNT_PAGE_FAULTS_MAJ:
2622 case PERF_COUNT_CONTEXT_SWITCHES:
2623 hw_ops = &perf_ops_generic;
2625 case PERF_COUNT_CPU_MIGRATIONS:
2626 if (!counter->hw_event.exclude_kernel)
2627 hw_ops = &perf_ops_cpu_migrations;
2632 hwc->irq_period = hw_event->irq_period;
2638 * Allocate and initialize a counter structure
2640 static struct perf_counter *
2641 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2643 struct perf_counter_context *ctx,
2644 struct perf_counter *group_leader,
2647 const struct hw_perf_counter_ops *hw_ops;
2648 struct perf_counter *counter;
2651 counter = kzalloc(sizeof(*counter), gfpflags);
2653 return ERR_PTR(-ENOMEM);
2656 * Single counters are their own group leaders, with an
2657 * empty sibling list:
2660 group_leader = counter;
2662 mutex_init(&counter->mutex);
2663 INIT_LIST_HEAD(&counter->list_entry);
2664 INIT_LIST_HEAD(&counter->event_entry);
2665 INIT_LIST_HEAD(&counter->sibling_list);
2666 init_waitqueue_head(&counter->waitq);
2668 mutex_init(&counter->mmap_mutex);
2670 INIT_LIST_HEAD(&counter->child_list);
2673 counter->hw_event = *hw_event;
2674 counter->group_leader = group_leader;
2675 counter->hw_ops = NULL;
2678 counter->state = PERF_COUNTER_STATE_INACTIVE;
2679 if (hw_event->disabled)
2680 counter->state = PERF_COUNTER_STATE_OFF;
2684 if (perf_event_raw(hw_event)) {
2685 hw_ops = hw_perf_counter_init(counter);
2689 switch (perf_event_type(hw_event)) {
2690 case PERF_TYPE_HARDWARE:
2691 hw_ops = hw_perf_counter_init(counter);
2694 case PERF_TYPE_SOFTWARE:
2695 hw_ops = sw_perf_counter_init(counter);
2698 case PERF_TYPE_TRACEPOINT:
2699 hw_ops = tp_perf_counter_init(counter);
2706 else if (IS_ERR(hw_ops))
2707 err = PTR_ERR(hw_ops);
2711 return ERR_PTR(err);
2714 counter->hw_ops = hw_ops;
2720 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2722 * @hw_event_uptr: event type attributes for monitoring/sampling
2725 * @group_fd: group leader counter fd
2727 SYSCALL_DEFINE5(perf_counter_open,
2728 const struct perf_counter_hw_event __user *, hw_event_uptr,
2729 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2731 struct perf_counter *counter, *group_leader;
2732 struct perf_counter_hw_event hw_event;
2733 struct perf_counter_context *ctx;
2734 struct file *counter_file = NULL;
2735 struct file *group_file = NULL;
2736 int fput_needed = 0;
2737 int fput_needed2 = 0;
2740 /* for future expandability... */
2744 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2748 * Get the target context (task or percpu):
2750 ctx = find_get_context(pid, cpu);
2752 return PTR_ERR(ctx);
2755 * Look up the group leader (we will attach this counter to it):
2757 group_leader = NULL;
2758 if (group_fd != -1) {
2760 group_file = fget_light(group_fd, &fput_needed);
2762 goto err_put_context;
2763 if (group_file->f_op != &perf_fops)
2764 goto err_put_context;
2766 group_leader = group_file->private_data;
2768 * Do not allow a recursive hierarchy (this new sibling
2769 * becoming part of another group-sibling):
2771 if (group_leader->group_leader != group_leader)
2772 goto err_put_context;
2774 * Do not allow to attach to a group in a different
2775 * task or CPU context:
2777 if (group_leader->ctx != ctx)
2778 goto err_put_context;
2780 * Only a group leader can be exclusive or pinned
2782 if (hw_event.exclusive || hw_event.pinned)
2783 goto err_put_context;
2786 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2788 ret = PTR_ERR(counter);
2789 if (IS_ERR(counter))
2790 goto err_put_context;
2792 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2794 goto err_free_put_context;
2796 counter_file = fget_light(ret, &fput_needed2);
2798 goto err_free_put_context;
2800 counter->filp = counter_file;
2801 mutex_lock(&ctx->mutex);
2802 perf_install_in_context(ctx, counter, cpu);
2803 mutex_unlock(&ctx->mutex);
2805 fput_light(counter_file, fput_needed2);
2808 fput_light(group_file, fput_needed);
2812 err_free_put_context:
2822 * Initialize the perf_counter context in a task_struct:
2825 __perf_counter_init_context(struct perf_counter_context *ctx,
2826 struct task_struct *task)
2828 memset(ctx, 0, sizeof(*ctx));
2829 spin_lock_init(&ctx->lock);
2830 mutex_init(&ctx->mutex);
2831 INIT_LIST_HEAD(&ctx->counter_list);
2832 INIT_LIST_HEAD(&ctx->event_list);
2837 * inherit a counter from parent task to child task:
2839 static struct perf_counter *
2840 inherit_counter(struct perf_counter *parent_counter,
2841 struct task_struct *parent,
2842 struct perf_counter_context *parent_ctx,
2843 struct task_struct *child,
2844 struct perf_counter *group_leader,
2845 struct perf_counter_context *child_ctx)
2847 struct perf_counter *child_counter;
2850 * Instead of creating recursive hierarchies of counters,
2851 * we link inherited counters back to the original parent,
2852 * which has a filp for sure, which we use as the reference
2855 if (parent_counter->parent)
2856 parent_counter = parent_counter->parent;
2858 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2859 parent_counter->cpu, child_ctx,
2860 group_leader, GFP_KERNEL);
2861 if (IS_ERR(child_counter))
2862 return child_counter;
2865 * Link it up in the child's context:
2867 child_counter->task = child;
2868 add_counter_to_ctx(child_counter, child_ctx);
2870 child_counter->parent = parent_counter;
2872 * inherit into child's child as well:
2874 child_counter->hw_event.inherit = 1;
2877 * Get a reference to the parent filp - we will fput it
2878 * when the child counter exits. This is safe to do because
2879 * we are in the parent and we know that the filp still
2880 * exists and has a nonzero count:
2882 atomic_long_inc(&parent_counter->filp->f_count);
2885 * Link this into the parent counter's child list
2887 mutex_lock(&parent_counter->mutex);
2888 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2891 * Make the child state follow the state of the parent counter,
2892 * not its hw_event.disabled bit. We hold the parent's mutex,
2893 * so we won't race with perf_counter_{en,dis}able_family.
2895 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2896 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2898 child_counter->state = PERF_COUNTER_STATE_OFF;
2900 mutex_unlock(&parent_counter->mutex);
2902 return child_counter;
2905 static int inherit_group(struct perf_counter *parent_counter,
2906 struct task_struct *parent,
2907 struct perf_counter_context *parent_ctx,
2908 struct task_struct *child,
2909 struct perf_counter_context *child_ctx)
2911 struct perf_counter *leader;
2912 struct perf_counter *sub;
2913 struct perf_counter *child_ctr;
2915 leader = inherit_counter(parent_counter, parent, parent_ctx,
2916 child, NULL, child_ctx);
2918 return PTR_ERR(leader);
2919 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2920 child_ctr = inherit_counter(sub, parent, parent_ctx,
2921 child, leader, child_ctx);
2922 if (IS_ERR(child_ctr))
2923 return PTR_ERR(child_ctr);
2928 static void sync_child_counter(struct perf_counter *child_counter,
2929 struct perf_counter *parent_counter)
2931 u64 parent_val, child_val;
2933 parent_val = atomic64_read(&parent_counter->count);
2934 child_val = atomic64_read(&child_counter->count);
2937 * Add back the child's count to the parent's count:
2939 atomic64_add(child_val, &parent_counter->count);
2940 atomic64_add(child_counter->total_time_enabled,
2941 &parent_counter->child_total_time_enabled);
2942 atomic64_add(child_counter->total_time_running,
2943 &parent_counter->child_total_time_running);
2946 * Remove this counter from the parent's list
2948 mutex_lock(&parent_counter->mutex);
2949 list_del_init(&child_counter->child_list);
2950 mutex_unlock(&parent_counter->mutex);
2953 * Release the parent counter, if this was the last
2956 fput(parent_counter->filp);
2960 __perf_counter_exit_task(struct task_struct *child,
2961 struct perf_counter *child_counter,
2962 struct perf_counter_context *child_ctx)
2964 struct perf_counter *parent_counter;
2965 struct perf_counter *sub, *tmp;
2968 * If we do not self-reap then we have to wait for the
2969 * child task to unschedule (it will happen for sure),
2970 * so that its counter is at its final count. (This
2971 * condition triggers rarely - child tasks usually get
2972 * off their CPU before the parent has a chance to
2973 * get this far into the reaping action)
2975 if (child != current) {
2976 wait_task_inactive(child, 0);
2977 list_del_init(&child_counter->list_entry);
2978 update_counter_times(child_counter);
2980 struct perf_cpu_context *cpuctx;
2981 unsigned long flags;
2985 * Disable and unlink this counter.
2987 * Be careful about zapping the list - IRQ/NMI context
2988 * could still be processing it:
2990 local_irq_save(flags);
2991 perf_flags = hw_perf_save_disable();
2993 cpuctx = &__get_cpu_var(perf_cpu_context);
2995 group_sched_out(child_counter, cpuctx, child_ctx);
2996 update_counter_times(child_counter);
2998 list_del_init(&child_counter->list_entry);
3000 child_ctx->nr_counters--;
3002 hw_perf_restore(perf_flags);
3003 local_irq_restore(flags);
3006 parent_counter = child_counter->parent;
3008 * It can happen that parent exits first, and has counters
3009 * that are still around due to the child reference. These
3010 * counters need to be zapped - but otherwise linger.
3012 if (parent_counter) {
3013 sync_child_counter(child_counter, parent_counter);
3014 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
3017 sync_child_counter(sub, sub->parent);
3021 free_counter(child_counter);
3026 * When a child task exits, feed back counter values to parent counters.
3028 * Note: we may be running in child context, but the PID is not hashed
3029 * anymore so new counters will not be added.
3031 void perf_counter_exit_task(struct task_struct *child)
3033 struct perf_counter *child_counter, *tmp;
3034 struct perf_counter_context *child_ctx;
3036 child_ctx = &child->perf_counter_ctx;
3038 if (likely(!child_ctx->nr_counters))
3041 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3043 __perf_counter_exit_task(child, child_counter, child_ctx);
3047 * Initialize the perf_counter context in task_struct
3049 void perf_counter_init_task(struct task_struct *child)
3051 struct perf_counter_context *child_ctx, *parent_ctx;
3052 struct perf_counter *counter;
3053 struct task_struct *parent = current;
3055 child_ctx = &child->perf_counter_ctx;
3056 parent_ctx = &parent->perf_counter_ctx;
3058 __perf_counter_init_context(child_ctx, child);
3061 * This is executed from the parent task context, so inherit
3062 * counters that have been marked for cloning:
3065 if (likely(!parent_ctx->nr_counters))
3069 * Lock the parent list. No need to lock the child - not PID
3070 * hashed yet and not running, so nobody can access it.
3072 mutex_lock(&parent_ctx->mutex);
3075 * We dont have to disable NMIs - we are only looking at
3076 * the list, not manipulating it:
3078 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
3079 if (!counter->hw_event.inherit)
3082 if (inherit_group(counter, parent,
3083 parent_ctx, child, child_ctx))
3087 mutex_unlock(&parent_ctx->mutex);
3090 static void __cpuinit perf_counter_init_cpu(int cpu)
3092 struct perf_cpu_context *cpuctx;
3094 cpuctx = &per_cpu(perf_cpu_context, cpu);
3095 __perf_counter_init_context(&cpuctx->ctx, NULL);
3097 mutex_lock(&perf_resource_mutex);
3098 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3099 mutex_unlock(&perf_resource_mutex);
3101 hw_perf_counter_setup(cpu);
3104 #ifdef CONFIG_HOTPLUG_CPU
3105 static void __perf_counter_exit_cpu(void *info)
3107 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3108 struct perf_counter_context *ctx = &cpuctx->ctx;
3109 struct perf_counter *counter, *tmp;
3111 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3112 __perf_counter_remove_from_context(counter);
3114 static void perf_counter_exit_cpu(int cpu)
3116 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3117 struct perf_counter_context *ctx = &cpuctx->ctx;
3119 mutex_lock(&ctx->mutex);
3120 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3121 mutex_unlock(&ctx->mutex);
3124 static inline void perf_counter_exit_cpu(int cpu) { }
3127 static int __cpuinit
3128 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3130 unsigned int cpu = (long)hcpu;
3134 case CPU_UP_PREPARE:
3135 case CPU_UP_PREPARE_FROZEN:
3136 perf_counter_init_cpu(cpu);
3139 case CPU_DOWN_PREPARE:
3140 case CPU_DOWN_PREPARE_FROZEN:
3141 perf_counter_exit_cpu(cpu);
3151 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3152 .notifier_call = perf_cpu_notify,
3155 static int __init perf_counter_init(void)
3157 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3158 (void *)(long)smp_processor_id());
3159 register_cpu_notifier(&perf_cpu_nb);
3163 early_initcall(perf_counter_init);
3165 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3167 return sprintf(buf, "%d\n", perf_reserved_percpu);
3171 perf_set_reserve_percpu(struct sysdev_class *class,
3175 struct perf_cpu_context *cpuctx;
3179 err = strict_strtoul(buf, 10, &val);
3182 if (val > perf_max_counters)
3185 mutex_lock(&perf_resource_mutex);
3186 perf_reserved_percpu = val;
3187 for_each_online_cpu(cpu) {
3188 cpuctx = &per_cpu(perf_cpu_context, cpu);
3189 spin_lock_irq(&cpuctx->ctx.lock);
3190 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3191 perf_max_counters - perf_reserved_percpu);
3192 cpuctx->max_pertask = mpt;
3193 spin_unlock_irq(&cpuctx->ctx.lock);
3195 mutex_unlock(&perf_resource_mutex);
3200 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3202 return sprintf(buf, "%d\n", perf_overcommit);
3206 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3211 err = strict_strtoul(buf, 10, &val);
3217 mutex_lock(&perf_resource_mutex);
3218 perf_overcommit = val;
3219 mutex_unlock(&perf_resource_mutex);
3224 static SYSDEV_CLASS_ATTR(
3227 perf_show_reserve_percpu,
3228 perf_set_reserve_percpu
3231 static SYSDEV_CLASS_ATTR(
3234 perf_show_overcommit,
3238 static struct attribute *perfclass_attrs[] = {
3239 &attr_reserve_percpu.attr,
3240 &attr_overcommit.attr,
3244 static struct attribute_group perfclass_attr_group = {
3245 .attrs = perfclass_attrs,
3246 .name = "perf_counters",
3249 static int __init perf_counter_sysfs_init(void)
3251 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3252 &perfclass_attr_group);
3254 device_initcall(perf_counter_sysfs_init);