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 regs = task_pt_regs(task);
801 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs);
802 __perf_counter_sched_out(ctx, cpuctx);
804 cpuctx->task_ctx = NULL;
807 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
809 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
813 group_sched_in(struct perf_counter *group_counter,
814 struct perf_cpu_context *cpuctx,
815 struct perf_counter_context *ctx,
818 struct perf_counter *counter, *partial_group;
821 if (group_counter->state == PERF_COUNTER_STATE_OFF)
824 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
826 return ret < 0 ? ret : 0;
828 group_counter->prev_state = group_counter->state;
829 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
833 * Schedule in siblings as one group (if any):
835 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
836 counter->prev_state = counter->state;
837 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
838 partial_group = counter;
847 * Groups can be scheduled in as one unit only, so undo any
848 * partial group before returning:
850 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
851 if (counter == partial_group)
853 counter_sched_out(counter, cpuctx, ctx);
855 counter_sched_out(group_counter, cpuctx, ctx);
861 __perf_counter_sched_in(struct perf_counter_context *ctx,
862 struct perf_cpu_context *cpuctx, int cpu)
864 struct perf_counter *counter;
868 spin_lock(&ctx->lock);
870 if (likely(!ctx->nr_counters))
873 ctx->timestamp = perf_clock();
875 flags = hw_perf_save_disable();
878 * First go through the list and put on any pinned groups
879 * in order to give them the best chance of going on.
881 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
882 if (counter->state <= PERF_COUNTER_STATE_OFF ||
883 !counter->hw_event.pinned)
885 if (counter->cpu != -1 && counter->cpu != cpu)
888 if (group_can_go_on(counter, cpuctx, 1))
889 group_sched_in(counter, cpuctx, ctx, cpu);
892 * If this pinned group hasn't been scheduled,
893 * put it in error state.
895 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
896 update_group_times(counter);
897 counter->state = PERF_COUNTER_STATE_ERROR;
901 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
903 * Ignore counters in OFF or ERROR state, and
904 * ignore pinned counters since we did them already.
906 if (counter->state <= PERF_COUNTER_STATE_OFF ||
907 counter->hw_event.pinned)
911 * Listen to the 'cpu' scheduling filter constraint
914 if (counter->cpu != -1 && counter->cpu != cpu)
917 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
918 if (group_sched_in(counter, cpuctx, ctx, cpu))
922 hw_perf_restore(flags);
924 spin_unlock(&ctx->lock);
928 * Called from scheduler to add the counters of the current task
929 * with interrupts disabled.
931 * We restore the counter value and then enable it.
933 * This does not protect us against NMI, but enable()
934 * sets the enabled bit in the control field of counter _before_
935 * accessing the counter control register. If a NMI hits, then it will
936 * keep the counter running.
938 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
940 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
941 struct perf_counter_context *ctx = &task->perf_counter_ctx;
943 __perf_counter_sched_in(ctx, cpuctx, cpu);
944 cpuctx->task_ctx = ctx;
947 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
949 struct perf_counter_context *ctx = &cpuctx->ctx;
951 __perf_counter_sched_in(ctx, cpuctx, cpu);
954 int perf_counter_task_disable(void)
956 struct task_struct *curr = current;
957 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
958 struct perf_counter *counter;
963 if (likely(!ctx->nr_counters))
966 local_irq_save(flags);
967 cpu = smp_processor_id();
969 perf_counter_task_sched_out(curr, cpu);
971 spin_lock(&ctx->lock);
974 * Disable all the counters:
976 perf_flags = hw_perf_save_disable();
978 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
979 if (counter->state != PERF_COUNTER_STATE_ERROR) {
980 update_group_times(counter);
981 counter->state = PERF_COUNTER_STATE_OFF;
985 hw_perf_restore(perf_flags);
987 spin_unlock_irqrestore(&ctx->lock, flags);
992 int perf_counter_task_enable(void)
994 struct task_struct *curr = current;
995 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
996 struct perf_counter *counter;
1001 if (likely(!ctx->nr_counters))
1004 local_irq_save(flags);
1005 cpu = smp_processor_id();
1007 perf_counter_task_sched_out(curr, cpu);
1009 spin_lock(&ctx->lock);
1012 * Disable all the counters:
1014 perf_flags = hw_perf_save_disable();
1016 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1017 if (counter->state > PERF_COUNTER_STATE_OFF)
1019 counter->state = PERF_COUNTER_STATE_INACTIVE;
1020 counter->tstamp_enabled =
1021 ctx->time - counter->total_time_enabled;
1022 counter->hw_event.disabled = 0;
1024 hw_perf_restore(perf_flags);
1026 spin_unlock(&ctx->lock);
1028 perf_counter_task_sched_in(curr, cpu);
1030 local_irq_restore(flags);
1036 * Round-robin a context's counters:
1038 static void rotate_ctx(struct perf_counter_context *ctx)
1040 struct perf_counter *counter;
1043 if (!ctx->nr_counters)
1046 spin_lock(&ctx->lock);
1048 * Rotate the first entry last (works just fine for group counters too):
1050 perf_flags = hw_perf_save_disable();
1051 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1052 list_move_tail(&counter->list_entry, &ctx->counter_list);
1055 hw_perf_restore(perf_flags);
1057 spin_unlock(&ctx->lock);
1060 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1062 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1063 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1064 const int rotate_percpu = 0;
1067 perf_counter_cpu_sched_out(cpuctx);
1068 perf_counter_task_sched_out(curr, cpu);
1071 rotate_ctx(&cpuctx->ctx);
1075 perf_counter_cpu_sched_in(cpuctx, cpu);
1076 perf_counter_task_sched_in(curr, cpu);
1080 * Cross CPU call to read the hardware counter
1082 static void __read(void *info)
1084 struct perf_counter *counter = info;
1085 struct perf_counter_context *ctx = counter->ctx;
1086 unsigned long flags;
1088 local_irq_save(flags);
1090 update_context_time(ctx);
1091 counter->hw_ops->read(counter);
1092 update_counter_times(counter);
1093 local_irq_restore(flags);
1096 static u64 perf_counter_read(struct perf_counter *counter)
1099 * If counter is enabled and currently active on a CPU, update the
1100 * value in the counter structure:
1102 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1103 smp_call_function_single(counter->oncpu,
1104 __read, counter, 1);
1105 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1106 update_counter_times(counter);
1109 return atomic64_read(&counter->count);
1112 static void put_context(struct perf_counter_context *ctx)
1115 put_task_struct(ctx->task);
1118 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1120 struct perf_cpu_context *cpuctx;
1121 struct perf_counter_context *ctx;
1122 struct task_struct *task;
1125 * If cpu is not a wildcard then this is a percpu counter:
1128 /* Must be root to operate on a CPU counter: */
1129 if (!capable(CAP_SYS_ADMIN))
1130 return ERR_PTR(-EACCES);
1132 if (cpu < 0 || cpu > num_possible_cpus())
1133 return ERR_PTR(-EINVAL);
1136 * We could be clever and allow to attach a counter to an
1137 * offline CPU and activate it when the CPU comes up, but
1140 if (!cpu_isset(cpu, cpu_online_map))
1141 return ERR_PTR(-ENODEV);
1143 cpuctx = &per_cpu(perf_cpu_context, cpu);
1153 task = find_task_by_vpid(pid);
1155 get_task_struct(task);
1159 return ERR_PTR(-ESRCH);
1161 ctx = &task->perf_counter_ctx;
1164 /* Reuse ptrace permission checks for now. */
1165 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1167 return ERR_PTR(-EACCES);
1173 static void free_counter_rcu(struct rcu_head *head)
1175 struct perf_counter *counter;
1177 counter = container_of(head, struct perf_counter, rcu_head);
1181 static void perf_pending_sync(struct perf_counter *counter);
1183 static void free_counter(struct perf_counter *counter)
1185 perf_pending_sync(counter);
1187 if (counter->destroy)
1188 counter->destroy(counter);
1190 call_rcu(&counter->rcu_head, free_counter_rcu);
1194 * Called when the last reference to the file is gone.
1196 static int perf_release(struct inode *inode, struct file *file)
1198 struct perf_counter *counter = file->private_data;
1199 struct perf_counter_context *ctx = counter->ctx;
1201 file->private_data = NULL;
1203 mutex_lock(&ctx->mutex);
1204 mutex_lock(&counter->mutex);
1206 perf_counter_remove_from_context(counter);
1208 mutex_unlock(&counter->mutex);
1209 mutex_unlock(&ctx->mutex);
1211 free_counter(counter);
1218 * Read the performance counter - simple non blocking version for now
1221 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1227 * Return end-of-file for a read on a counter that is in
1228 * error state (i.e. because it was pinned but it couldn't be
1229 * scheduled on to the CPU at some point).
1231 if (counter->state == PERF_COUNTER_STATE_ERROR)
1234 mutex_lock(&counter->mutex);
1235 values[0] = perf_counter_read(counter);
1237 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1238 values[n++] = counter->total_time_enabled +
1239 atomic64_read(&counter->child_total_time_enabled);
1240 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1241 values[n++] = counter->total_time_running +
1242 atomic64_read(&counter->child_total_time_running);
1243 mutex_unlock(&counter->mutex);
1245 if (count < n * sizeof(u64))
1247 count = n * sizeof(u64);
1249 if (copy_to_user(buf, values, count))
1256 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1258 struct perf_counter *counter = file->private_data;
1260 return perf_read_hw(counter, buf, count);
1263 static unsigned int perf_poll(struct file *file, poll_table *wait)
1265 struct perf_counter *counter = file->private_data;
1266 struct perf_mmap_data *data;
1267 unsigned int events;
1270 data = rcu_dereference(counter->data);
1272 events = atomic_xchg(&data->wakeup, 0);
1277 poll_wait(file, &counter->waitq, wait);
1282 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1284 struct perf_counter *counter = file->private_data;
1288 case PERF_COUNTER_IOC_ENABLE:
1289 perf_counter_enable_family(counter);
1291 case PERF_COUNTER_IOC_DISABLE:
1292 perf_counter_disable_family(counter);
1294 case PERF_COUNTER_IOC_REFRESH:
1295 perf_counter_refresh(counter, arg);
1304 * Callers need to ensure there can be no nesting of this function, otherwise
1305 * the seqlock logic goes bad. We can not serialize this because the arch
1306 * code calls this from NMI context.
1308 void perf_counter_update_userpage(struct perf_counter *counter)
1310 struct perf_mmap_data *data;
1311 struct perf_counter_mmap_page *userpg;
1314 data = rcu_dereference(counter->data);
1318 userpg = data->user_page;
1321 * Disable preemption so as to not let the corresponding user-space
1322 * spin too long if we get preempted.
1327 userpg->index = counter->hw.idx;
1328 userpg->offset = atomic64_read(&counter->count);
1329 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1330 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1339 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1341 struct perf_counter *counter = vma->vm_file->private_data;
1342 struct perf_mmap_data *data;
1343 int ret = VM_FAULT_SIGBUS;
1346 data = rcu_dereference(counter->data);
1350 if (vmf->pgoff == 0) {
1351 vmf->page = virt_to_page(data->user_page);
1353 int nr = vmf->pgoff - 1;
1355 if ((unsigned)nr > data->nr_pages)
1358 vmf->page = virt_to_page(data->data_pages[nr]);
1360 get_page(vmf->page);
1368 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1370 struct perf_mmap_data *data;
1374 WARN_ON(atomic_read(&counter->mmap_count));
1376 size = sizeof(struct perf_mmap_data);
1377 size += nr_pages * sizeof(void *);
1379 data = kzalloc(size, GFP_KERNEL);
1383 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1384 if (!data->user_page)
1385 goto fail_user_page;
1387 for (i = 0; i < nr_pages; i++) {
1388 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1389 if (!data->data_pages[i])
1390 goto fail_data_pages;
1393 data->nr_pages = nr_pages;
1395 rcu_assign_pointer(counter->data, data);
1400 for (i--; i >= 0; i--)
1401 free_page((unsigned long)data->data_pages[i]);
1403 free_page((unsigned long)data->user_page);
1412 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1414 struct perf_mmap_data *data = container_of(rcu_head,
1415 struct perf_mmap_data, rcu_head);
1418 free_page((unsigned long)data->user_page);
1419 for (i = 0; i < data->nr_pages; i++)
1420 free_page((unsigned long)data->data_pages[i]);
1424 static void perf_mmap_data_free(struct perf_counter *counter)
1426 struct perf_mmap_data *data = counter->data;
1428 WARN_ON(atomic_read(&counter->mmap_count));
1430 rcu_assign_pointer(counter->data, NULL);
1431 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1434 static void perf_mmap_open(struct vm_area_struct *vma)
1436 struct perf_counter *counter = vma->vm_file->private_data;
1438 atomic_inc(&counter->mmap_count);
1441 static void perf_mmap_close(struct vm_area_struct *vma)
1443 struct perf_counter *counter = vma->vm_file->private_data;
1445 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1446 &counter->mmap_mutex)) {
1447 vma->vm_mm->locked_vm -= counter->data->nr_pages + 1;
1448 perf_mmap_data_free(counter);
1449 mutex_unlock(&counter->mmap_mutex);
1453 static struct vm_operations_struct perf_mmap_vmops = {
1454 .open = perf_mmap_open,
1455 .close = perf_mmap_close,
1456 .fault = perf_mmap_fault,
1459 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1461 struct perf_counter *counter = file->private_data;
1462 unsigned long vma_size;
1463 unsigned long nr_pages;
1464 unsigned long locked, lock_limit;
1467 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1470 vma_size = vma->vm_end - vma->vm_start;
1471 nr_pages = (vma_size / PAGE_SIZE) - 1;
1474 * If we have data pages ensure they're a power-of-two number, so we
1475 * can do bitmasks instead of modulo.
1477 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1480 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1483 if (vma->vm_pgoff != 0)
1486 mutex_lock(&counter->mmap_mutex);
1487 if (atomic_inc_not_zero(&counter->mmap_count)) {
1488 if (nr_pages != counter->data->nr_pages)
1493 locked = vma->vm_mm->locked_vm;
1494 locked += nr_pages + 1;
1496 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1497 lock_limit >>= PAGE_SHIFT;
1499 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1504 WARN_ON(counter->data);
1505 ret = perf_mmap_data_alloc(counter, nr_pages);
1509 atomic_set(&counter->mmap_count, 1);
1510 vma->vm_mm->locked_vm += nr_pages + 1;
1512 mutex_unlock(&counter->mmap_mutex);
1514 vma->vm_flags &= ~VM_MAYWRITE;
1515 vma->vm_flags |= VM_RESERVED;
1516 vma->vm_ops = &perf_mmap_vmops;
1521 static int perf_fasync(int fd, struct file *filp, int on)
1523 struct perf_counter *counter = filp->private_data;
1524 struct inode *inode = filp->f_path.dentry->d_inode;
1527 mutex_lock(&inode->i_mutex);
1528 retval = fasync_helper(fd, filp, on, &counter->fasync);
1529 mutex_unlock(&inode->i_mutex);
1537 static const struct file_operations perf_fops = {
1538 .release = perf_release,
1541 .unlocked_ioctl = perf_ioctl,
1542 .compat_ioctl = perf_ioctl,
1544 .fasync = perf_fasync,
1548 * Perf counter wakeup
1550 * If there's data, ensure we set the poll() state and publish everything
1551 * to user-space before waking everybody up.
1554 void perf_counter_wakeup(struct perf_counter *counter)
1556 struct perf_mmap_data *data;
1559 data = rcu_dereference(counter->data);
1561 atomic_set(&data->wakeup, POLL_IN);
1563 * Ensure all data writes are issued before updating the
1564 * user-space data head information. The matching rmb()
1565 * will be in userspace after reading this value.
1568 data->user_page->data_head = atomic_read(&data->head);
1572 wake_up_all(&counter->waitq);
1574 if (counter->pending_kill) {
1575 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1576 counter->pending_kill = 0;
1583 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1585 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1586 * single linked list and use cmpxchg() to add entries lockless.
1589 static void perf_pending_counter(struct perf_pending_entry *entry)
1591 struct perf_counter *counter = container_of(entry,
1592 struct perf_counter, pending);
1594 if (counter->pending_disable) {
1595 counter->pending_disable = 0;
1596 perf_counter_disable(counter);
1599 if (counter->pending_wakeup) {
1600 counter->pending_wakeup = 0;
1601 perf_counter_wakeup(counter);
1605 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1607 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1611 static void perf_pending_queue(struct perf_pending_entry *entry,
1612 void (*func)(struct perf_pending_entry *))
1614 struct perf_pending_entry **head;
1616 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1621 head = &get_cpu_var(perf_pending_head);
1624 entry->next = *head;
1625 } while (cmpxchg(head, entry->next, entry) != entry->next);
1627 set_perf_counter_pending();
1629 put_cpu_var(perf_pending_head);
1632 static int __perf_pending_run(void)
1634 struct perf_pending_entry *list;
1637 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1638 while (list != PENDING_TAIL) {
1639 void (*func)(struct perf_pending_entry *);
1640 struct perf_pending_entry *entry = list;
1647 * Ensure we observe the unqueue before we issue the wakeup,
1648 * so that we won't be waiting forever.
1649 * -- see perf_not_pending().
1660 static inline int perf_not_pending(struct perf_counter *counter)
1663 * If we flush on whatever cpu we run, there is a chance we don't
1667 __perf_pending_run();
1671 * Ensure we see the proper queue state before going to sleep
1672 * so that we do not miss the wakeup. -- see perf_pending_handle()
1675 return counter->pending.next == NULL;
1678 static void perf_pending_sync(struct perf_counter *counter)
1680 wait_event(counter->waitq, perf_not_pending(counter));
1683 void perf_counter_do_pending(void)
1685 __perf_pending_run();
1689 * Callchain support -- arch specific
1692 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1701 struct perf_output_handle {
1702 struct perf_counter *counter;
1703 struct perf_mmap_data *data;
1704 unsigned int offset;
1711 static inline void __perf_output_wakeup(struct perf_output_handle *handle)
1714 handle->counter->pending_wakeup = 1;
1715 perf_pending_queue(&handle->counter->pending,
1716 perf_pending_counter);
1718 perf_counter_wakeup(handle->counter);
1721 static int perf_output_begin(struct perf_output_handle *handle,
1722 struct perf_counter *counter, unsigned int size,
1723 int nmi, int overflow)
1725 struct perf_mmap_data *data;
1726 unsigned int offset, head;
1729 data = rcu_dereference(counter->data);
1733 handle->counter = counter;
1735 handle->overflow = overflow;
1737 if (!data->nr_pages)
1741 offset = head = atomic_read(&data->head);
1743 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1745 handle->data = data;
1746 handle->offset = offset;
1747 handle->head = head;
1748 handle->wakeup = (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT);
1753 __perf_output_wakeup(handle);
1760 static void perf_output_copy(struct perf_output_handle *handle,
1761 void *buf, unsigned int len)
1763 unsigned int pages_mask;
1764 unsigned int offset;
1768 offset = handle->offset;
1769 pages_mask = handle->data->nr_pages - 1;
1770 pages = handle->data->data_pages;
1773 unsigned int page_offset;
1776 nr = (offset >> PAGE_SHIFT) & pages_mask;
1777 page_offset = offset & (PAGE_SIZE - 1);
1778 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1780 memcpy(pages[nr] + page_offset, buf, size);
1787 handle->offset = offset;
1789 WARN_ON_ONCE(handle->offset > handle->head);
1792 #define perf_output_put(handle, x) \
1793 perf_output_copy((handle), &(x), sizeof(x))
1795 static void perf_output_end(struct perf_output_handle *handle)
1797 int wakeup_events = handle->counter->hw_event.wakeup_events;
1799 if (handle->overflow && wakeup_events) {
1800 int events = atomic_inc_return(&handle->data->events);
1801 if (events >= wakeup_events) {
1802 atomic_sub(wakeup_events, &handle->data->events);
1803 __perf_output_wakeup(handle);
1805 } else if (handle->wakeup)
1806 __perf_output_wakeup(handle);
1810 static void perf_counter_output(struct perf_counter *counter,
1811 int nmi, struct pt_regs *regs)
1814 u64 record_type = counter->hw_event.record_type;
1815 struct perf_output_handle handle;
1816 struct perf_event_header header;
1825 struct perf_callchain_entry *callchain = NULL;
1826 int callchain_size = 0;
1829 header.type = PERF_EVENT_COUNTER_OVERFLOW;
1830 header.size = sizeof(header);
1832 if (record_type & PERF_RECORD_IP) {
1833 ip = instruction_pointer(regs);
1834 header.type |= __PERF_EVENT_IP;
1835 header.size += sizeof(ip);
1838 if (record_type & PERF_RECORD_TID) {
1839 /* namespace issues */
1840 tid_entry.pid = current->group_leader->pid;
1841 tid_entry.tid = current->pid;
1843 header.type |= __PERF_EVENT_TID;
1844 header.size += sizeof(tid_entry);
1847 if (record_type & PERF_RECORD_GROUP) {
1848 header.type |= __PERF_EVENT_GROUP;
1849 header.size += sizeof(u64) +
1850 counter->nr_siblings * sizeof(group_entry);
1853 if (record_type & PERF_RECORD_CALLCHAIN) {
1854 callchain = perf_callchain(regs);
1857 callchain_size = (1 + callchain->nr) * sizeof(u64);
1859 header.type |= __PERF_EVENT_CALLCHAIN;
1860 header.size += callchain_size;
1864 if (record_type & PERF_RECORD_TIME) {
1866 * Maybe do better on x86 and provide cpu_clock_nmi()
1868 time = sched_clock();
1870 header.type |= __PERF_EVENT_TIME;
1871 header.size += sizeof(u64);
1874 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
1878 perf_output_put(&handle, header);
1880 if (record_type & PERF_RECORD_IP)
1881 perf_output_put(&handle, ip);
1883 if (record_type & PERF_RECORD_TID)
1884 perf_output_put(&handle, tid_entry);
1886 if (record_type & PERF_RECORD_GROUP) {
1887 struct perf_counter *leader, *sub;
1888 u64 nr = counter->nr_siblings;
1890 perf_output_put(&handle, nr);
1892 leader = counter->group_leader;
1893 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1895 sub->hw_ops->read(sub);
1897 group_entry.event = sub->hw_event.config;
1898 group_entry.counter = atomic64_read(&sub->count);
1900 perf_output_put(&handle, group_entry);
1905 perf_output_copy(&handle, callchain, callchain_size);
1907 if (record_type & PERF_RECORD_TIME)
1908 perf_output_put(&handle, time);
1910 perf_output_end(&handle);
1917 struct perf_mmap_event {
1923 struct perf_event_header header;
1933 static void perf_counter_mmap_output(struct perf_counter *counter,
1934 struct perf_mmap_event *mmap_event)
1936 struct perf_output_handle handle;
1937 int size = mmap_event->event.header.size;
1938 int ret = perf_output_begin(&handle, counter, size, 0, 0);
1943 perf_output_put(&handle, mmap_event->event);
1944 perf_output_copy(&handle, mmap_event->file_name,
1945 mmap_event->file_size);
1946 perf_output_end(&handle);
1949 static int perf_counter_mmap_match(struct perf_counter *counter,
1950 struct perf_mmap_event *mmap_event)
1952 if (counter->hw_event.mmap &&
1953 mmap_event->event.header.type == PERF_EVENT_MMAP)
1956 if (counter->hw_event.munmap &&
1957 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
1963 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
1964 struct perf_mmap_event *mmap_event)
1966 struct perf_counter *counter;
1968 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
1972 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
1973 if (perf_counter_mmap_match(counter, mmap_event))
1974 perf_counter_mmap_output(counter, mmap_event);
1979 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
1981 struct perf_cpu_context *cpuctx;
1982 struct file *file = mmap_event->file;
1989 buf = kzalloc(PATH_MAX, GFP_KERNEL);
1991 name = strncpy(tmp, "//enomem", sizeof(tmp));
1994 name = dentry_path(file->f_dentry, buf, PATH_MAX);
1996 name = strncpy(tmp, "//toolong", sizeof(tmp));
2000 name = strncpy(tmp, "//anon", sizeof(tmp));
2005 size = ALIGN(strlen(name), sizeof(u64));
2007 mmap_event->file_name = name;
2008 mmap_event->file_size = size;
2010 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2012 cpuctx = &get_cpu_var(perf_cpu_context);
2013 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2014 put_cpu_var(perf_cpu_context);
2016 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2021 void perf_counter_mmap(unsigned long addr, unsigned long len,
2022 unsigned long pgoff, struct file *file)
2024 struct perf_mmap_event mmap_event = {
2027 .header = { .type = PERF_EVENT_MMAP, },
2028 .pid = current->group_leader->pid,
2029 .tid = current->pid,
2036 perf_counter_mmap_event(&mmap_event);
2039 void perf_counter_munmap(unsigned long addr, unsigned long len,
2040 unsigned long pgoff, struct file *file)
2042 struct perf_mmap_event mmap_event = {
2045 .header = { .type = PERF_EVENT_MUNMAP, },
2046 .pid = current->group_leader->pid,
2047 .tid = current->pid,
2054 perf_counter_mmap_event(&mmap_event);
2058 * Generic counter overflow handling.
2061 int perf_counter_overflow(struct perf_counter *counter,
2062 int nmi, struct pt_regs *regs)
2064 int events = atomic_read(&counter->event_limit);
2067 counter->pending_kill = POLL_IN;
2068 if (events && atomic_dec_and_test(&counter->event_limit)) {
2070 counter->pending_kill = POLL_HUP;
2072 counter->pending_disable = 1;
2073 perf_pending_queue(&counter->pending,
2074 perf_pending_counter);
2076 perf_counter_disable(counter);
2079 perf_counter_output(counter, nmi, regs);
2084 * Generic software counter infrastructure
2087 static void perf_swcounter_update(struct perf_counter *counter)
2089 struct hw_perf_counter *hwc = &counter->hw;
2094 prev = atomic64_read(&hwc->prev_count);
2095 now = atomic64_read(&hwc->count);
2096 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2101 atomic64_add(delta, &counter->count);
2102 atomic64_sub(delta, &hwc->period_left);
2105 static void perf_swcounter_set_period(struct perf_counter *counter)
2107 struct hw_perf_counter *hwc = &counter->hw;
2108 s64 left = atomic64_read(&hwc->period_left);
2109 s64 period = hwc->irq_period;
2111 if (unlikely(left <= -period)) {
2113 atomic64_set(&hwc->period_left, left);
2116 if (unlikely(left <= 0)) {
2118 atomic64_add(period, &hwc->period_left);
2121 atomic64_set(&hwc->prev_count, -left);
2122 atomic64_set(&hwc->count, -left);
2125 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2127 enum hrtimer_restart ret = HRTIMER_RESTART;
2128 struct perf_counter *counter;
2129 struct pt_regs *regs;
2131 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2132 counter->hw_ops->read(counter);
2134 regs = get_irq_regs();
2136 * In case we exclude kernel IPs or are somehow not in interrupt
2137 * context, provide the next best thing, the user IP.
2139 if ((counter->hw_event.exclude_kernel || !regs) &&
2140 !counter->hw_event.exclude_user)
2141 regs = task_pt_regs(current);
2144 if (perf_counter_overflow(counter, 0, regs))
2145 ret = HRTIMER_NORESTART;
2148 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2153 static void perf_swcounter_overflow(struct perf_counter *counter,
2154 int nmi, struct pt_regs *regs)
2156 perf_swcounter_update(counter);
2157 perf_swcounter_set_period(counter);
2158 if (perf_counter_overflow(counter, nmi, regs))
2159 /* soft-disable the counter */
2164 static int perf_swcounter_match(struct perf_counter *counter,
2165 enum perf_event_types type,
2166 u32 event, struct pt_regs *regs)
2168 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2171 if (perf_event_raw(&counter->hw_event))
2174 if (perf_event_type(&counter->hw_event) != type)
2177 if (perf_event_id(&counter->hw_event) != event)
2180 if (counter->hw_event.exclude_user && user_mode(regs))
2183 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2189 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2190 int nmi, struct pt_regs *regs)
2192 int neg = atomic64_add_negative(nr, &counter->hw.count);
2193 if (counter->hw.irq_period && !neg)
2194 perf_swcounter_overflow(counter, nmi, regs);
2197 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2198 enum perf_event_types type, u32 event,
2199 u64 nr, int nmi, struct pt_regs *regs)
2201 struct perf_counter *counter;
2203 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2207 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2208 if (perf_swcounter_match(counter, type, event, regs))
2209 perf_swcounter_add(counter, nr, nmi, regs);
2214 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2217 return &cpuctx->recursion[3];
2220 return &cpuctx->recursion[2];
2223 return &cpuctx->recursion[1];
2225 return &cpuctx->recursion[0];
2228 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2229 u64 nr, int nmi, struct pt_regs *regs)
2231 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2232 int *recursion = perf_swcounter_recursion_context(cpuctx);
2240 perf_swcounter_ctx_event(&cpuctx->ctx, type, event, nr, nmi, regs);
2241 if (cpuctx->task_ctx) {
2242 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2250 put_cpu_var(perf_cpu_context);
2253 void perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)
2255 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs);
2258 static void perf_swcounter_read(struct perf_counter *counter)
2260 perf_swcounter_update(counter);
2263 static int perf_swcounter_enable(struct perf_counter *counter)
2265 perf_swcounter_set_period(counter);
2269 static void perf_swcounter_disable(struct perf_counter *counter)
2271 perf_swcounter_update(counter);
2274 static const struct hw_perf_counter_ops perf_ops_generic = {
2275 .enable = perf_swcounter_enable,
2276 .disable = perf_swcounter_disable,
2277 .read = perf_swcounter_read,
2281 * Software counter: cpu wall time clock
2284 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2286 int cpu = raw_smp_processor_id();
2290 now = cpu_clock(cpu);
2291 prev = atomic64_read(&counter->hw.prev_count);
2292 atomic64_set(&counter->hw.prev_count, now);
2293 atomic64_add(now - prev, &counter->count);
2296 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2298 struct hw_perf_counter *hwc = &counter->hw;
2299 int cpu = raw_smp_processor_id();
2301 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2302 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2303 hwc->hrtimer.function = perf_swcounter_hrtimer;
2304 if (hwc->irq_period) {
2305 __hrtimer_start_range_ns(&hwc->hrtimer,
2306 ns_to_ktime(hwc->irq_period), 0,
2307 HRTIMER_MODE_REL, 0);
2313 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2315 hrtimer_cancel(&counter->hw.hrtimer);
2316 cpu_clock_perf_counter_update(counter);
2319 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2321 cpu_clock_perf_counter_update(counter);
2324 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
2325 .enable = cpu_clock_perf_counter_enable,
2326 .disable = cpu_clock_perf_counter_disable,
2327 .read = cpu_clock_perf_counter_read,
2331 * Software counter: task time clock
2334 static void task_clock_perf_counter_update(struct perf_counter *counter)
2339 update_context_time(counter->ctx);
2340 now = counter->ctx->time;
2342 prev = atomic64_xchg(&counter->hw.prev_count, now);
2344 atomic64_add(delta, &counter->count);
2347 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2349 struct hw_perf_counter *hwc = &counter->hw;
2352 update_context_time(counter->ctx);
2353 now = counter->ctx->time;
2355 atomic64_set(&hwc->prev_count, now);
2356 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2357 hwc->hrtimer.function = perf_swcounter_hrtimer;
2358 if (hwc->irq_period) {
2359 __hrtimer_start_range_ns(&hwc->hrtimer,
2360 ns_to_ktime(hwc->irq_period), 0,
2361 HRTIMER_MODE_REL, 0);
2367 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2369 hrtimer_cancel(&counter->hw.hrtimer);
2370 task_clock_perf_counter_update(counter);
2373 static void task_clock_perf_counter_read(struct perf_counter *counter)
2375 task_clock_perf_counter_update(counter);
2378 static const struct hw_perf_counter_ops perf_ops_task_clock = {
2379 .enable = task_clock_perf_counter_enable,
2380 .disable = task_clock_perf_counter_disable,
2381 .read = task_clock_perf_counter_read,
2385 * Software counter: cpu migrations
2388 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2390 struct task_struct *curr = counter->ctx->task;
2393 return curr->se.nr_migrations;
2394 return cpu_nr_migrations(smp_processor_id());
2397 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2402 prev = atomic64_read(&counter->hw.prev_count);
2403 now = get_cpu_migrations(counter);
2405 atomic64_set(&counter->hw.prev_count, now);
2409 atomic64_add(delta, &counter->count);
2412 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2414 cpu_migrations_perf_counter_update(counter);
2417 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2419 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2420 atomic64_set(&counter->hw.prev_count,
2421 get_cpu_migrations(counter));
2425 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2427 cpu_migrations_perf_counter_update(counter);
2430 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
2431 .enable = cpu_migrations_perf_counter_enable,
2432 .disable = cpu_migrations_perf_counter_disable,
2433 .read = cpu_migrations_perf_counter_read,
2436 #ifdef CONFIG_EVENT_PROFILE
2437 void perf_tpcounter_event(int event_id)
2439 struct pt_regs *regs = get_irq_regs();
2442 regs = task_pt_regs(current);
2444 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs);
2447 extern int ftrace_profile_enable(int);
2448 extern void ftrace_profile_disable(int);
2450 static void tp_perf_counter_destroy(struct perf_counter *counter)
2452 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2455 static const struct hw_perf_counter_ops *
2456 tp_perf_counter_init(struct perf_counter *counter)
2458 int event_id = perf_event_id(&counter->hw_event);
2461 ret = ftrace_profile_enable(event_id);
2465 counter->destroy = tp_perf_counter_destroy;
2466 counter->hw.irq_period = counter->hw_event.irq_period;
2468 return &perf_ops_generic;
2471 static const struct hw_perf_counter_ops *
2472 tp_perf_counter_init(struct perf_counter *counter)
2478 static const struct hw_perf_counter_ops *
2479 sw_perf_counter_init(struct perf_counter *counter)
2481 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2482 const struct hw_perf_counter_ops *hw_ops = NULL;
2483 struct hw_perf_counter *hwc = &counter->hw;
2486 * Software counters (currently) can't in general distinguish
2487 * between user, kernel and hypervisor events.
2488 * However, context switches and cpu migrations are considered
2489 * to be kernel events, and page faults are never hypervisor
2492 switch (perf_event_id(&counter->hw_event)) {
2493 case PERF_COUNT_CPU_CLOCK:
2494 hw_ops = &perf_ops_cpu_clock;
2496 if (hw_event->irq_period && hw_event->irq_period < 10000)
2497 hw_event->irq_period = 10000;
2499 case PERF_COUNT_TASK_CLOCK:
2501 * If the user instantiates this as a per-cpu counter,
2502 * use the cpu_clock counter instead.
2504 if (counter->ctx->task)
2505 hw_ops = &perf_ops_task_clock;
2507 hw_ops = &perf_ops_cpu_clock;
2509 if (hw_event->irq_period && hw_event->irq_period < 10000)
2510 hw_event->irq_period = 10000;
2512 case PERF_COUNT_PAGE_FAULTS:
2513 case PERF_COUNT_PAGE_FAULTS_MIN:
2514 case PERF_COUNT_PAGE_FAULTS_MAJ:
2515 case PERF_COUNT_CONTEXT_SWITCHES:
2516 hw_ops = &perf_ops_generic;
2518 case PERF_COUNT_CPU_MIGRATIONS:
2519 if (!counter->hw_event.exclude_kernel)
2520 hw_ops = &perf_ops_cpu_migrations;
2525 hwc->irq_period = hw_event->irq_period;
2531 * Allocate and initialize a counter structure
2533 static struct perf_counter *
2534 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2536 struct perf_counter_context *ctx,
2537 struct perf_counter *group_leader,
2540 const struct hw_perf_counter_ops *hw_ops;
2541 struct perf_counter *counter;
2544 counter = kzalloc(sizeof(*counter), gfpflags);
2546 return ERR_PTR(-ENOMEM);
2549 * Single counters are their own group leaders, with an
2550 * empty sibling list:
2553 group_leader = counter;
2555 mutex_init(&counter->mutex);
2556 INIT_LIST_HEAD(&counter->list_entry);
2557 INIT_LIST_HEAD(&counter->event_entry);
2558 INIT_LIST_HEAD(&counter->sibling_list);
2559 init_waitqueue_head(&counter->waitq);
2561 mutex_init(&counter->mmap_mutex);
2563 INIT_LIST_HEAD(&counter->child_list);
2566 counter->hw_event = *hw_event;
2567 counter->group_leader = group_leader;
2568 counter->hw_ops = NULL;
2571 counter->state = PERF_COUNTER_STATE_INACTIVE;
2572 if (hw_event->disabled)
2573 counter->state = PERF_COUNTER_STATE_OFF;
2577 if (perf_event_raw(hw_event)) {
2578 hw_ops = hw_perf_counter_init(counter);
2582 switch (perf_event_type(hw_event)) {
2583 case PERF_TYPE_HARDWARE:
2584 hw_ops = hw_perf_counter_init(counter);
2587 case PERF_TYPE_SOFTWARE:
2588 hw_ops = sw_perf_counter_init(counter);
2591 case PERF_TYPE_TRACEPOINT:
2592 hw_ops = tp_perf_counter_init(counter);
2599 else if (IS_ERR(hw_ops))
2600 err = PTR_ERR(hw_ops);
2604 return ERR_PTR(err);
2607 counter->hw_ops = hw_ops;
2613 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2615 * @hw_event_uptr: event type attributes for monitoring/sampling
2618 * @group_fd: group leader counter fd
2620 SYSCALL_DEFINE5(perf_counter_open,
2621 const struct perf_counter_hw_event __user *, hw_event_uptr,
2622 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2624 struct perf_counter *counter, *group_leader;
2625 struct perf_counter_hw_event hw_event;
2626 struct perf_counter_context *ctx;
2627 struct file *counter_file = NULL;
2628 struct file *group_file = NULL;
2629 int fput_needed = 0;
2630 int fput_needed2 = 0;
2633 /* for future expandability... */
2637 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2641 * Get the target context (task or percpu):
2643 ctx = find_get_context(pid, cpu);
2645 return PTR_ERR(ctx);
2648 * Look up the group leader (we will attach this counter to it):
2650 group_leader = NULL;
2651 if (group_fd != -1) {
2653 group_file = fget_light(group_fd, &fput_needed);
2655 goto err_put_context;
2656 if (group_file->f_op != &perf_fops)
2657 goto err_put_context;
2659 group_leader = group_file->private_data;
2661 * Do not allow a recursive hierarchy (this new sibling
2662 * becoming part of another group-sibling):
2664 if (group_leader->group_leader != group_leader)
2665 goto err_put_context;
2667 * Do not allow to attach to a group in a different
2668 * task or CPU context:
2670 if (group_leader->ctx != ctx)
2671 goto err_put_context;
2673 * Only a group leader can be exclusive or pinned
2675 if (hw_event.exclusive || hw_event.pinned)
2676 goto err_put_context;
2679 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2681 ret = PTR_ERR(counter);
2682 if (IS_ERR(counter))
2683 goto err_put_context;
2685 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2687 goto err_free_put_context;
2689 counter_file = fget_light(ret, &fput_needed2);
2691 goto err_free_put_context;
2693 counter->filp = counter_file;
2694 mutex_lock(&ctx->mutex);
2695 perf_install_in_context(ctx, counter, cpu);
2696 mutex_unlock(&ctx->mutex);
2698 fput_light(counter_file, fput_needed2);
2701 fput_light(group_file, fput_needed);
2705 err_free_put_context:
2715 * Initialize the perf_counter context in a task_struct:
2718 __perf_counter_init_context(struct perf_counter_context *ctx,
2719 struct task_struct *task)
2721 memset(ctx, 0, sizeof(*ctx));
2722 spin_lock_init(&ctx->lock);
2723 mutex_init(&ctx->mutex);
2724 INIT_LIST_HEAD(&ctx->counter_list);
2725 INIT_LIST_HEAD(&ctx->event_list);
2730 * inherit a counter from parent task to child task:
2732 static struct perf_counter *
2733 inherit_counter(struct perf_counter *parent_counter,
2734 struct task_struct *parent,
2735 struct perf_counter_context *parent_ctx,
2736 struct task_struct *child,
2737 struct perf_counter *group_leader,
2738 struct perf_counter_context *child_ctx)
2740 struct perf_counter *child_counter;
2743 * Instead of creating recursive hierarchies of counters,
2744 * we link inherited counters back to the original parent,
2745 * which has a filp for sure, which we use as the reference
2748 if (parent_counter->parent)
2749 parent_counter = parent_counter->parent;
2751 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2752 parent_counter->cpu, child_ctx,
2753 group_leader, GFP_KERNEL);
2754 if (IS_ERR(child_counter))
2755 return child_counter;
2758 * Link it up in the child's context:
2760 child_counter->task = child;
2761 add_counter_to_ctx(child_counter, child_ctx);
2763 child_counter->parent = parent_counter;
2765 * inherit into child's child as well:
2767 child_counter->hw_event.inherit = 1;
2770 * Get a reference to the parent filp - we will fput it
2771 * when the child counter exits. This is safe to do because
2772 * we are in the parent and we know that the filp still
2773 * exists and has a nonzero count:
2775 atomic_long_inc(&parent_counter->filp->f_count);
2778 * Link this into the parent counter's child list
2780 mutex_lock(&parent_counter->mutex);
2781 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2784 * Make the child state follow the state of the parent counter,
2785 * not its hw_event.disabled bit. We hold the parent's mutex,
2786 * so we won't race with perf_counter_{en,dis}able_family.
2788 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2789 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2791 child_counter->state = PERF_COUNTER_STATE_OFF;
2793 mutex_unlock(&parent_counter->mutex);
2795 return child_counter;
2798 static int inherit_group(struct perf_counter *parent_counter,
2799 struct task_struct *parent,
2800 struct perf_counter_context *parent_ctx,
2801 struct task_struct *child,
2802 struct perf_counter_context *child_ctx)
2804 struct perf_counter *leader;
2805 struct perf_counter *sub;
2806 struct perf_counter *child_ctr;
2808 leader = inherit_counter(parent_counter, parent, parent_ctx,
2809 child, NULL, child_ctx);
2811 return PTR_ERR(leader);
2812 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2813 child_ctr = inherit_counter(sub, parent, parent_ctx,
2814 child, leader, child_ctx);
2815 if (IS_ERR(child_ctr))
2816 return PTR_ERR(child_ctr);
2821 static void sync_child_counter(struct perf_counter *child_counter,
2822 struct perf_counter *parent_counter)
2824 u64 parent_val, child_val;
2826 parent_val = atomic64_read(&parent_counter->count);
2827 child_val = atomic64_read(&child_counter->count);
2830 * Add back the child's count to the parent's count:
2832 atomic64_add(child_val, &parent_counter->count);
2833 atomic64_add(child_counter->total_time_enabled,
2834 &parent_counter->child_total_time_enabled);
2835 atomic64_add(child_counter->total_time_running,
2836 &parent_counter->child_total_time_running);
2839 * Remove this counter from the parent's list
2841 mutex_lock(&parent_counter->mutex);
2842 list_del_init(&child_counter->child_list);
2843 mutex_unlock(&parent_counter->mutex);
2846 * Release the parent counter, if this was the last
2849 fput(parent_counter->filp);
2853 __perf_counter_exit_task(struct task_struct *child,
2854 struct perf_counter *child_counter,
2855 struct perf_counter_context *child_ctx)
2857 struct perf_counter *parent_counter;
2858 struct perf_counter *sub, *tmp;
2861 * If we do not self-reap then we have to wait for the
2862 * child task to unschedule (it will happen for sure),
2863 * so that its counter is at its final count. (This
2864 * condition triggers rarely - child tasks usually get
2865 * off their CPU before the parent has a chance to
2866 * get this far into the reaping action)
2868 if (child != current) {
2869 wait_task_inactive(child, 0);
2870 list_del_init(&child_counter->list_entry);
2871 update_counter_times(child_counter);
2873 struct perf_cpu_context *cpuctx;
2874 unsigned long flags;
2878 * Disable and unlink this counter.
2880 * Be careful about zapping the list - IRQ/NMI context
2881 * could still be processing it:
2883 local_irq_save(flags);
2884 perf_flags = hw_perf_save_disable();
2886 cpuctx = &__get_cpu_var(perf_cpu_context);
2888 group_sched_out(child_counter, cpuctx, child_ctx);
2889 update_counter_times(child_counter);
2891 list_del_init(&child_counter->list_entry);
2893 child_ctx->nr_counters--;
2895 hw_perf_restore(perf_flags);
2896 local_irq_restore(flags);
2899 parent_counter = child_counter->parent;
2901 * It can happen that parent exits first, and has counters
2902 * that are still around due to the child reference. These
2903 * counters need to be zapped - but otherwise linger.
2905 if (parent_counter) {
2906 sync_child_counter(child_counter, parent_counter);
2907 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
2910 sync_child_counter(sub, sub->parent);
2914 free_counter(child_counter);
2919 * When a child task exits, feed back counter values to parent counters.
2921 * Note: we may be running in child context, but the PID is not hashed
2922 * anymore so new counters will not be added.
2924 void perf_counter_exit_task(struct task_struct *child)
2926 struct perf_counter *child_counter, *tmp;
2927 struct perf_counter_context *child_ctx;
2929 child_ctx = &child->perf_counter_ctx;
2931 if (likely(!child_ctx->nr_counters))
2934 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
2936 __perf_counter_exit_task(child, child_counter, child_ctx);
2940 * Initialize the perf_counter context in task_struct
2942 void perf_counter_init_task(struct task_struct *child)
2944 struct perf_counter_context *child_ctx, *parent_ctx;
2945 struct perf_counter *counter;
2946 struct task_struct *parent = current;
2948 child_ctx = &child->perf_counter_ctx;
2949 parent_ctx = &parent->perf_counter_ctx;
2951 __perf_counter_init_context(child_ctx, child);
2954 * This is executed from the parent task context, so inherit
2955 * counters that have been marked for cloning:
2958 if (likely(!parent_ctx->nr_counters))
2962 * Lock the parent list. No need to lock the child - not PID
2963 * hashed yet and not running, so nobody can access it.
2965 mutex_lock(&parent_ctx->mutex);
2968 * We dont have to disable NMIs - we are only looking at
2969 * the list, not manipulating it:
2971 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2972 if (!counter->hw_event.inherit)
2975 if (inherit_group(counter, parent,
2976 parent_ctx, child, child_ctx))
2980 mutex_unlock(&parent_ctx->mutex);
2983 static void __cpuinit perf_counter_init_cpu(int cpu)
2985 struct perf_cpu_context *cpuctx;
2987 cpuctx = &per_cpu(perf_cpu_context, cpu);
2988 __perf_counter_init_context(&cpuctx->ctx, NULL);
2990 mutex_lock(&perf_resource_mutex);
2991 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2992 mutex_unlock(&perf_resource_mutex);
2994 hw_perf_counter_setup(cpu);
2997 #ifdef CONFIG_HOTPLUG_CPU
2998 static void __perf_counter_exit_cpu(void *info)
3000 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3001 struct perf_counter_context *ctx = &cpuctx->ctx;
3002 struct perf_counter *counter, *tmp;
3004 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3005 __perf_counter_remove_from_context(counter);
3007 static void perf_counter_exit_cpu(int cpu)
3009 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3010 struct perf_counter_context *ctx = &cpuctx->ctx;
3012 mutex_lock(&ctx->mutex);
3013 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3014 mutex_unlock(&ctx->mutex);
3017 static inline void perf_counter_exit_cpu(int cpu) { }
3020 static int __cpuinit
3021 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3023 unsigned int cpu = (long)hcpu;
3027 case CPU_UP_PREPARE:
3028 case CPU_UP_PREPARE_FROZEN:
3029 perf_counter_init_cpu(cpu);
3032 case CPU_DOWN_PREPARE:
3033 case CPU_DOWN_PREPARE_FROZEN:
3034 perf_counter_exit_cpu(cpu);
3044 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3045 .notifier_call = perf_cpu_notify,
3048 static int __init perf_counter_init(void)
3050 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3051 (void *)(long)smp_processor_id());
3052 register_cpu_notifier(&perf_cpu_nb);
3056 early_initcall(perf_counter_init);
3058 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3060 return sprintf(buf, "%d\n", perf_reserved_percpu);
3064 perf_set_reserve_percpu(struct sysdev_class *class,
3068 struct perf_cpu_context *cpuctx;
3072 err = strict_strtoul(buf, 10, &val);
3075 if (val > perf_max_counters)
3078 mutex_lock(&perf_resource_mutex);
3079 perf_reserved_percpu = val;
3080 for_each_online_cpu(cpu) {
3081 cpuctx = &per_cpu(perf_cpu_context, cpu);
3082 spin_lock_irq(&cpuctx->ctx.lock);
3083 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3084 perf_max_counters - perf_reserved_percpu);
3085 cpuctx->max_pertask = mpt;
3086 spin_unlock_irq(&cpuctx->ctx.lock);
3088 mutex_unlock(&perf_resource_mutex);
3093 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3095 return sprintf(buf, "%d\n", perf_overcommit);
3099 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3104 err = strict_strtoul(buf, 10, &val);
3110 mutex_lock(&perf_resource_mutex);
3111 perf_overcommit = val;
3112 mutex_unlock(&perf_resource_mutex);
3117 static SYSDEV_CLASS_ATTR(
3120 perf_show_reserve_percpu,
3121 perf_set_reserve_percpu
3124 static SYSDEV_CLASS_ATTR(
3127 perf_show_overcommit,
3131 static struct attribute *perfclass_attrs[] = {
3132 &attr_reserve_percpu.attr,
3133 &attr_overcommit.attr,
3137 static struct attribute_group perfclass_attr_group = {
3138 .attrs = perfclass_attrs,
3139 .name = "perf_counters",
3142 static int __init perf_counter_sysfs_init(void)
3144 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3145 &perfclass_attr_group);
3147 device_initcall(perf_counter_sysfs_init);