2 * Performance counter core code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
7 * For licencing details see kernel-base/COPYING
11 #include <linux/cpu.h>
12 #include <linux/smp.h>
13 #include <linux/file.h>
14 #include <linux/poll.h>
15 #include <linux/sysfs.h>
16 #include <linux/ptrace.h>
17 #include <linux/percpu.h>
18 #include <linux/uaccess.h>
19 #include <linux/syscalls.h>
20 #include <linux/anon_inodes.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/perf_counter.h>
24 #include <linux/vmstat.h>
27 * Each CPU has a list of per CPU counters:
29 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
31 int perf_max_counters __read_mostly = 1;
32 static int perf_reserved_percpu __read_mostly;
33 static int perf_overcommit __read_mostly = 1;
36 * Mutex for (sysadmin-configurable) counter reservations:
38 static DEFINE_MUTEX(perf_resource_mutex);
41 * Architecture provided APIs - weak aliases:
43 extern __weak const struct hw_perf_counter_ops *
44 hw_perf_counter_init(struct perf_counter *counter)
49 u64 __weak hw_perf_save_disable(void) { return 0; }
50 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
51 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
52 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
53 struct perf_cpu_context *cpuctx,
54 struct perf_counter_context *ctx, int cpu)
59 void __weak perf_counter_print_debug(void) { }
62 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
64 struct perf_counter *group_leader = counter->group_leader;
67 * Depending on whether it is a standalone or sibling counter,
68 * add it straight to the context's counter list, or to the group
69 * leader's sibling list:
71 if (counter->group_leader == counter)
72 list_add_tail(&counter->list_entry, &ctx->counter_list);
74 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
78 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
80 struct perf_counter *sibling, *tmp;
82 list_del_init(&counter->list_entry);
85 * If this was a group counter with sibling counters then
86 * upgrade the siblings to singleton counters by adding them
87 * to the context list directly:
89 list_for_each_entry_safe(sibling, tmp,
90 &counter->sibling_list, list_entry) {
92 list_move_tail(&sibling->list_entry, &ctx->counter_list);
93 sibling->group_leader = sibling;
98 counter_sched_out(struct perf_counter *counter,
99 struct perf_cpu_context *cpuctx,
100 struct perf_counter_context *ctx)
102 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
105 counter->state = PERF_COUNTER_STATE_INACTIVE;
106 counter->hw_ops->disable(counter);
109 if (!is_software_counter(counter))
110 cpuctx->active_oncpu--;
112 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
113 cpuctx->exclusive = 0;
117 group_sched_out(struct perf_counter *group_counter,
118 struct perf_cpu_context *cpuctx,
119 struct perf_counter_context *ctx)
121 struct perf_counter *counter;
123 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
126 counter_sched_out(group_counter, cpuctx, ctx);
129 * Schedule out siblings (if any):
131 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
132 counter_sched_out(counter, cpuctx, ctx);
134 if (group_counter->hw_event.exclusive)
135 cpuctx->exclusive = 0;
139 * Cross CPU call to remove a performance counter
141 * We disable the counter on the hardware level first. After that we
142 * remove it from the context list.
144 static void __perf_counter_remove_from_context(void *info)
146 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
147 struct perf_counter *counter = info;
148 struct perf_counter_context *ctx = counter->ctx;
153 * If this is a task context, we need to check whether it is
154 * the current task context of this cpu. If not it has been
155 * scheduled out before the smp call arrived.
157 if (ctx->task && cpuctx->task_ctx != ctx)
160 curr_rq_lock_irq_save(&flags);
161 spin_lock(&ctx->lock);
163 counter_sched_out(counter, cpuctx, ctx);
165 counter->task = NULL;
169 * Protect the list operation against NMI by disabling the
170 * counters on a global level. NOP for non NMI based counters.
172 perf_flags = hw_perf_save_disable();
173 list_del_counter(counter, ctx);
174 hw_perf_restore(perf_flags);
178 * Allow more per task counters with respect to the
181 cpuctx->max_pertask =
182 min(perf_max_counters - ctx->nr_counters,
183 perf_max_counters - perf_reserved_percpu);
186 spin_unlock(&ctx->lock);
187 curr_rq_unlock_irq_restore(&flags);
192 * Remove the counter from a task's (or a CPU's) list of counters.
194 * Must be called with counter->mutex and ctx->mutex held.
196 * CPU counters are removed with a smp call. For task counters we only
197 * call when the task is on a CPU.
199 static void perf_counter_remove_from_context(struct perf_counter *counter)
201 struct perf_counter_context *ctx = counter->ctx;
202 struct task_struct *task = ctx->task;
206 * Per cpu counters are removed via an smp call and
207 * the removal is always sucessful.
209 smp_call_function_single(counter->cpu,
210 __perf_counter_remove_from_context,
216 task_oncpu_function_call(task, __perf_counter_remove_from_context,
219 spin_lock_irq(&ctx->lock);
221 * If the context is active we need to retry the smp call.
223 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
224 spin_unlock_irq(&ctx->lock);
229 * The lock prevents that this context is scheduled in so we
230 * can remove the counter safely, if the call above did not
233 if (!list_empty(&counter->list_entry)) {
235 list_del_counter(counter, ctx);
236 counter->task = NULL;
238 spin_unlock_irq(&ctx->lock);
242 * Cross CPU call to disable a performance counter
244 static void __perf_counter_disable(void *info)
246 struct perf_counter *counter = info;
247 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
248 struct perf_counter_context *ctx = counter->ctx;
252 * If this is a per-task counter, need to check whether this
253 * counter's task is the current task on this cpu.
255 if (ctx->task && cpuctx->task_ctx != ctx)
258 curr_rq_lock_irq_save(&flags);
259 spin_lock(&ctx->lock);
262 * If the counter is on, turn it off.
263 * If it is in error state, leave it in error state.
265 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
266 if (counter == counter->group_leader)
267 group_sched_out(counter, cpuctx, ctx);
269 counter_sched_out(counter, cpuctx, ctx);
270 counter->state = PERF_COUNTER_STATE_OFF;
273 spin_unlock(&ctx->lock);
274 curr_rq_unlock_irq_restore(&flags);
280 static void perf_counter_disable(struct perf_counter *counter)
282 struct perf_counter_context *ctx = counter->ctx;
283 struct task_struct *task = ctx->task;
287 * Disable the counter on the cpu that it's on
289 smp_call_function_single(counter->cpu, __perf_counter_disable,
295 task_oncpu_function_call(task, __perf_counter_disable, counter);
297 spin_lock_irq(&ctx->lock);
299 * If the counter is still active, we need to retry the cross-call.
301 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
302 spin_unlock_irq(&ctx->lock);
307 * Since we have the lock this context can't be scheduled
308 * in, so we can change the state safely.
310 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
311 counter->state = PERF_COUNTER_STATE_OFF;
313 spin_unlock_irq(&ctx->lock);
317 * Disable a counter and all its children.
319 static void perf_counter_disable_family(struct perf_counter *counter)
321 struct perf_counter *child;
323 perf_counter_disable(counter);
326 * Lock the mutex to protect the list of children
328 mutex_lock(&counter->mutex);
329 list_for_each_entry(child, &counter->child_list, child_list)
330 perf_counter_disable(child);
331 mutex_unlock(&counter->mutex);
335 counter_sched_in(struct perf_counter *counter,
336 struct perf_cpu_context *cpuctx,
337 struct perf_counter_context *ctx,
340 if (counter->state <= PERF_COUNTER_STATE_OFF)
343 counter->state = PERF_COUNTER_STATE_ACTIVE;
344 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
346 * The new state must be visible before we turn it on in the hardware:
350 if (counter->hw_ops->enable(counter)) {
351 counter->state = PERF_COUNTER_STATE_INACTIVE;
356 if (!is_software_counter(counter))
357 cpuctx->active_oncpu++;
360 if (counter->hw_event.exclusive)
361 cpuctx->exclusive = 1;
367 * Return 1 for a group consisting entirely of software counters,
368 * 0 if the group contains any hardware counters.
370 static int is_software_only_group(struct perf_counter *leader)
372 struct perf_counter *counter;
374 if (!is_software_counter(leader))
376 list_for_each_entry(counter, &leader->sibling_list, list_entry)
377 if (!is_software_counter(counter))
383 * Work out whether we can put this counter group on the CPU now.
385 static int group_can_go_on(struct perf_counter *counter,
386 struct perf_cpu_context *cpuctx,
390 * Groups consisting entirely of software counters can always go on.
392 if (is_software_only_group(counter))
395 * If an exclusive group is already on, no other hardware
396 * counters can go on.
398 if (cpuctx->exclusive)
401 * If this group is exclusive and there are already
402 * counters on the CPU, it can't go on.
404 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
407 * Otherwise, try to add it if all previous groups were able
414 * Cross CPU call to install and enable a performance counter
416 static void __perf_install_in_context(void *info)
418 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
419 struct perf_counter *counter = info;
420 struct perf_counter_context *ctx = counter->ctx;
421 struct perf_counter *leader = counter->group_leader;
422 int cpu = smp_processor_id();
428 * If this is a task context, we need to check whether it is
429 * the current task context of this cpu. If not it has been
430 * scheduled out before the smp call arrived.
432 if (ctx->task && cpuctx->task_ctx != ctx)
435 curr_rq_lock_irq_save(&flags);
436 spin_lock(&ctx->lock);
439 * Protect the list operation against NMI by disabling the
440 * counters on a global level. NOP for non NMI based counters.
442 perf_flags = hw_perf_save_disable();
444 list_add_counter(counter, ctx);
446 counter->prev_state = PERF_COUNTER_STATE_OFF;
449 * Don't put the counter on if it is disabled or if
450 * it is in a group and the group isn't on.
452 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
453 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
457 * An exclusive counter can't go on if there are already active
458 * hardware counters, and no hardware counter can go on if there
459 * is already an exclusive counter on.
461 if (!group_can_go_on(counter, cpuctx, 1))
464 err = counter_sched_in(counter, cpuctx, ctx, cpu);
468 * This counter couldn't go on. If it is in a group
469 * then we have to pull the whole group off.
470 * If the counter group is pinned then put it in error state.
472 if (leader != counter)
473 group_sched_out(leader, cpuctx, ctx);
474 if (leader->hw_event.pinned)
475 leader->state = PERF_COUNTER_STATE_ERROR;
478 if (!err && !ctx->task && cpuctx->max_pertask)
479 cpuctx->max_pertask--;
482 hw_perf_restore(perf_flags);
484 spin_unlock(&ctx->lock);
485 curr_rq_unlock_irq_restore(&flags);
489 * Attach a performance counter to a context
491 * First we add the counter to the list with the hardware enable bit
492 * in counter->hw_config cleared.
494 * If the counter is attached to a task which is on a CPU we use a smp
495 * call to enable it in the task context. The task might have been
496 * scheduled away, but we check this in the smp call again.
498 * Must be called with ctx->mutex held.
501 perf_install_in_context(struct perf_counter_context *ctx,
502 struct perf_counter *counter,
505 struct task_struct *task = ctx->task;
509 * Per cpu counters are installed via an smp call and
510 * the install is always sucessful.
512 smp_call_function_single(cpu, __perf_install_in_context,
517 counter->task = task;
519 task_oncpu_function_call(task, __perf_install_in_context,
522 spin_lock_irq(&ctx->lock);
524 * we need to retry the smp call.
526 if (ctx->is_active && list_empty(&counter->list_entry)) {
527 spin_unlock_irq(&ctx->lock);
532 * The lock prevents that this context is scheduled in so we
533 * can add the counter safely, if it the call above did not
536 if (list_empty(&counter->list_entry)) {
537 list_add_counter(counter, ctx);
540 spin_unlock_irq(&ctx->lock);
544 * Cross CPU call to enable a performance counter
546 static void __perf_counter_enable(void *info)
548 struct perf_counter *counter = info;
549 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
550 struct perf_counter_context *ctx = counter->ctx;
551 struct perf_counter *leader = counter->group_leader;
556 * If this is a per-task counter, need to check whether this
557 * counter's task is the current task on this cpu.
559 if (ctx->task && cpuctx->task_ctx != ctx)
562 curr_rq_lock_irq_save(&flags);
563 spin_lock(&ctx->lock);
565 counter->prev_state = counter->state;
566 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
568 counter->state = PERF_COUNTER_STATE_INACTIVE;
571 * If the counter is in a group and isn't the group leader,
572 * then don't put it on unless the group is on.
574 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
577 if (!group_can_go_on(counter, cpuctx, 1))
580 err = counter_sched_in(counter, cpuctx, ctx,
585 * If this counter can't go on and it's part of a
586 * group, then the whole group has to come off.
588 if (leader != counter)
589 group_sched_out(leader, cpuctx, ctx);
590 if (leader->hw_event.pinned)
591 leader->state = PERF_COUNTER_STATE_ERROR;
595 spin_unlock(&ctx->lock);
596 curr_rq_unlock_irq_restore(&flags);
602 static void perf_counter_enable(struct perf_counter *counter)
604 struct perf_counter_context *ctx = counter->ctx;
605 struct task_struct *task = ctx->task;
609 * Enable the counter on the cpu that it's on
611 smp_call_function_single(counter->cpu, __perf_counter_enable,
616 spin_lock_irq(&ctx->lock);
617 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
621 * If the counter is in error state, clear that first.
622 * That way, if we see the counter in error state below, we
623 * know that it has gone back into error state, as distinct
624 * from the task having been scheduled away before the
625 * cross-call arrived.
627 if (counter->state == PERF_COUNTER_STATE_ERROR)
628 counter->state = PERF_COUNTER_STATE_OFF;
631 spin_unlock_irq(&ctx->lock);
632 task_oncpu_function_call(task, __perf_counter_enable, counter);
634 spin_lock_irq(&ctx->lock);
637 * If the context is active and the counter is still off,
638 * we need to retry the cross-call.
640 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
644 * Since we have the lock this context can't be scheduled
645 * in, so we can change the state safely.
647 if (counter->state == PERF_COUNTER_STATE_OFF)
648 counter->state = PERF_COUNTER_STATE_INACTIVE;
650 spin_unlock_irq(&ctx->lock);
654 * Enable a counter and all its children.
656 static void perf_counter_enable_family(struct perf_counter *counter)
658 struct perf_counter *child;
660 perf_counter_enable(counter);
663 * Lock the mutex to protect the list of children
665 mutex_lock(&counter->mutex);
666 list_for_each_entry(child, &counter->child_list, child_list)
667 perf_counter_enable(child);
668 mutex_unlock(&counter->mutex);
671 void __perf_counter_sched_out(struct perf_counter_context *ctx,
672 struct perf_cpu_context *cpuctx)
674 struct perf_counter *counter;
677 spin_lock(&ctx->lock);
679 if (likely(!ctx->nr_counters))
682 flags = hw_perf_save_disable();
683 if (ctx->nr_active) {
684 list_for_each_entry(counter, &ctx->counter_list, list_entry)
685 group_sched_out(counter, cpuctx, ctx);
687 hw_perf_restore(flags);
689 spin_unlock(&ctx->lock);
693 * Called from scheduler to remove the counters of the current task,
694 * with interrupts disabled.
696 * We stop each counter and update the counter value in counter->count.
698 * This does not protect us against NMI, but disable()
699 * sets the disabled bit in the control field of counter _before_
700 * accessing the counter control register. If a NMI hits, then it will
701 * not restart the counter.
703 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
705 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
706 struct perf_counter_context *ctx = &task->perf_counter_ctx;
708 if (likely(!cpuctx->task_ctx))
711 __perf_counter_sched_out(ctx, cpuctx);
713 cpuctx->task_ctx = NULL;
716 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
718 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
722 group_sched_in(struct perf_counter *group_counter,
723 struct perf_cpu_context *cpuctx,
724 struct perf_counter_context *ctx,
727 struct perf_counter *counter, *partial_group;
730 if (group_counter->state == PERF_COUNTER_STATE_OFF)
733 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
735 return ret < 0 ? ret : 0;
737 group_counter->prev_state = group_counter->state;
738 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
742 * Schedule in siblings as one group (if any):
744 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
745 counter->prev_state = counter->state;
746 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
747 partial_group = counter;
756 * Groups can be scheduled in as one unit only, so undo any
757 * partial group before returning:
759 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
760 if (counter == partial_group)
762 counter_sched_out(counter, cpuctx, ctx);
764 counter_sched_out(group_counter, cpuctx, ctx);
770 __perf_counter_sched_in(struct perf_counter_context *ctx,
771 struct perf_cpu_context *cpuctx, int cpu)
773 struct perf_counter *counter;
777 spin_lock(&ctx->lock);
779 if (likely(!ctx->nr_counters))
782 flags = hw_perf_save_disable();
785 * First go through the list and put on any pinned groups
786 * in order to give them the best chance of going on.
788 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
789 if (counter->state <= PERF_COUNTER_STATE_OFF ||
790 !counter->hw_event.pinned)
792 if (counter->cpu != -1 && counter->cpu != cpu)
795 if (group_can_go_on(counter, cpuctx, 1))
796 group_sched_in(counter, cpuctx, ctx, cpu);
799 * If this pinned group hasn't been scheduled,
800 * put it in error state.
802 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
803 counter->state = PERF_COUNTER_STATE_ERROR;
806 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
808 * Ignore counters in OFF or ERROR state, and
809 * ignore pinned counters since we did them already.
811 if (counter->state <= PERF_COUNTER_STATE_OFF ||
812 counter->hw_event.pinned)
816 * Listen to the 'cpu' scheduling filter constraint
819 if (counter->cpu != -1 && counter->cpu != cpu)
822 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
823 if (group_sched_in(counter, cpuctx, ctx, cpu))
827 hw_perf_restore(flags);
829 spin_unlock(&ctx->lock);
833 * Called from scheduler to add the counters of the current task
834 * with interrupts disabled.
836 * We restore the counter value and then enable it.
838 * This does not protect us against NMI, but enable()
839 * sets the enabled bit in the control field of counter _before_
840 * accessing the counter control register. If a NMI hits, then it will
841 * keep the counter running.
843 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
845 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
846 struct perf_counter_context *ctx = &task->perf_counter_ctx;
848 __perf_counter_sched_in(ctx, cpuctx, cpu);
849 cpuctx->task_ctx = ctx;
852 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
854 struct perf_counter_context *ctx = &cpuctx->ctx;
856 __perf_counter_sched_in(ctx, cpuctx, cpu);
859 int perf_counter_task_disable(void)
861 struct task_struct *curr = current;
862 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
863 struct perf_counter *counter;
868 if (likely(!ctx->nr_counters))
871 curr_rq_lock_irq_save(&flags);
872 cpu = smp_processor_id();
874 /* force the update of the task clock: */
875 __task_delta_exec(curr, 1);
877 perf_counter_task_sched_out(curr, cpu);
879 spin_lock(&ctx->lock);
882 * Disable all the counters:
884 perf_flags = hw_perf_save_disable();
886 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
887 if (counter->state != PERF_COUNTER_STATE_ERROR)
888 counter->state = PERF_COUNTER_STATE_OFF;
891 hw_perf_restore(perf_flags);
893 spin_unlock(&ctx->lock);
895 curr_rq_unlock_irq_restore(&flags);
900 int perf_counter_task_enable(void)
902 struct task_struct *curr = current;
903 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
904 struct perf_counter *counter;
909 if (likely(!ctx->nr_counters))
912 curr_rq_lock_irq_save(&flags);
913 cpu = smp_processor_id();
915 /* force the update of the task clock: */
916 __task_delta_exec(curr, 1);
918 perf_counter_task_sched_out(curr, cpu);
920 spin_lock(&ctx->lock);
923 * Disable all the counters:
925 perf_flags = hw_perf_save_disable();
927 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
928 if (counter->state > PERF_COUNTER_STATE_OFF)
930 counter->state = PERF_COUNTER_STATE_INACTIVE;
931 counter->hw_event.disabled = 0;
933 hw_perf_restore(perf_flags);
935 spin_unlock(&ctx->lock);
937 perf_counter_task_sched_in(curr, cpu);
939 curr_rq_unlock_irq_restore(&flags);
945 * Round-robin a context's counters:
947 static void rotate_ctx(struct perf_counter_context *ctx)
949 struct perf_counter *counter;
952 if (!ctx->nr_counters)
955 spin_lock(&ctx->lock);
957 * Rotate the first entry last (works just fine for group counters too):
959 perf_flags = hw_perf_save_disable();
960 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
961 list_move_tail(&counter->list_entry, &ctx->counter_list);
964 hw_perf_restore(perf_flags);
966 spin_unlock(&ctx->lock);
969 void perf_counter_task_tick(struct task_struct *curr, int cpu)
971 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
972 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
973 const int rotate_percpu = 0;
976 perf_counter_cpu_sched_out(cpuctx);
977 perf_counter_task_sched_out(curr, cpu);
980 rotate_ctx(&cpuctx->ctx);
984 perf_counter_cpu_sched_in(cpuctx, cpu);
985 perf_counter_task_sched_in(curr, cpu);
989 * Cross CPU call to read the hardware counter
991 static void __read(void *info)
993 struct perf_counter *counter = info;
996 curr_rq_lock_irq_save(&flags);
997 counter->hw_ops->read(counter);
998 curr_rq_unlock_irq_restore(&flags);
1001 static u64 perf_counter_read(struct perf_counter *counter)
1004 * If counter is enabled and currently active on a CPU, update the
1005 * value in the counter structure:
1007 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1008 smp_call_function_single(counter->oncpu,
1009 __read, counter, 1);
1012 return atomic64_read(&counter->count);
1016 * Cross CPU call to switch performance data pointers
1018 static void __perf_switch_irq_data(void *info)
1020 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1021 struct perf_counter *counter = info;
1022 struct perf_counter_context *ctx = counter->ctx;
1023 struct perf_data *oldirqdata = counter->irqdata;
1026 * If this is a task context, we need to check whether it is
1027 * the current task context of this cpu. If not it has been
1028 * scheduled out before the smp call arrived.
1031 if (cpuctx->task_ctx != ctx)
1033 spin_lock(&ctx->lock);
1036 /* Change the pointer NMI safe */
1037 atomic_long_set((atomic_long_t *)&counter->irqdata,
1038 (unsigned long) counter->usrdata);
1039 counter->usrdata = oldirqdata;
1042 spin_unlock(&ctx->lock);
1045 static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
1047 struct perf_counter_context *ctx = counter->ctx;
1048 struct perf_data *oldirqdata = counter->irqdata;
1049 struct task_struct *task = ctx->task;
1052 smp_call_function_single(counter->cpu,
1053 __perf_switch_irq_data,
1055 return counter->usrdata;
1059 spin_lock_irq(&ctx->lock);
1060 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
1061 counter->irqdata = counter->usrdata;
1062 counter->usrdata = oldirqdata;
1063 spin_unlock_irq(&ctx->lock);
1066 spin_unlock_irq(&ctx->lock);
1067 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
1068 /* Might have failed, because task was scheduled out */
1069 if (counter->irqdata == oldirqdata)
1072 return counter->usrdata;
1075 static void put_context(struct perf_counter_context *ctx)
1078 put_task_struct(ctx->task);
1081 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1083 struct perf_cpu_context *cpuctx;
1084 struct perf_counter_context *ctx;
1085 struct task_struct *task;
1088 * If cpu is not a wildcard then this is a percpu counter:
1091 /* Must be root to operate on a CPU counter: */
1092 if (!capable(CAP_SYS_ADMIN))
1093 return ERR_PTR(-EACCES);
1095 if (cpu < 0 || cpu > num_possible_cpus())
1096 return ERR_PTR(-EINVAL);
1099 * We could be clever and allow to attach a counter to an
1100 * offline CPU and activate it when the CPU comes up, but
1103 if (!cpu_isset(cpu, cpu_online_map))
1104 return ERR_PTR(-ENODEV);
1106 cpuctx = &per_cpu(perf_cpu_context, cpu);
1116 task = find_task_by_vpid(pid);
1118 get_task_struct(task);
1122 return ERR_PTR(-ESRCH);
1124 ctx = &task->perf_counter_ctx;
1127 /* Reuse ptrace permission checks for now. */
1128 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1130 return ERR_PTR(-EACCES);
1137 * Called when the last reference to the file is gone.
1139 static int perf_release(struct inode *inode, struct file *file)
1141 struct perf_counter *counter = file->private_data;
1142 struct perf_counter_context *ctx = counter->ctx;
1144 file->private_data = NULL;
1146 mutex_lock(&ctx->mutex);
1147 mutex_lock(&counter->mutex);
1149 perf_counter_remove_from_context(counter);
1151 mutex_unlock(&counter->mutex);
1152 mutex_unlock(&ctx->mutex);
1161 * Read the performance counter - simple non blocking version for now
1164 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1168 if (count != sizeof(cntval))
1172 * Return end-of-file for a read on a counter that is in
1173 * error state (i.e. because it was pinned but it couldn't be
1174 * scheduled on to the CPU at some point).
1176 if (counter->state == PERF_COUNTER_STATE_ERROR)
1179 mutex_lock(&counter->mutex);
1180 cntval = perf_counter_read(counter);
1181 mutex_unlock(&counter->mutex);
1183 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
1187 perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
1192 count = min(count, (size_t)usrdata->len);
1193 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
1196 /* Adjust the counters */
1197 usrdata->len -= count;
1199 usrdata->rd_idx = 0;
1201 usrdata->rd_idx += count;
1207 perf_read_irq_data(struct perf_counter *counter,
1212 struct perf_data *irqdata, *usrdata;
1213 DECLARE_WAITQUEUE(wait, current);
1216 irqdata = counter->irqdata;
1217 usrdata = counter->usrdata;
1219 if (usrdata->len + irqdata->len >= count)
1225 spin_lock_irq(&counter->waitq.lock);
1226 __add_wait_queue(&counter->waitq, &wait);
1228 set_current_state(TASK_INTERRUPTIBLE);
1229 if (usrdata->len + irqdata->len >= count)
1232 if (signal_pending(current))
1235 if (counter->state == PERF_COUNTER_STATE_ERROR)
1238 spin_unlock_irq(&counter->waitq.lock);
1240 spin_lock_irq(&counter->waitq.lock);
1242 __remove_wait_queue(&counter->waitq, &wait);
1243 __set_current_state(TASK_RUNNING);
1244 spin_unlock_irq(&counter->waitq.lock);
1246 if (usrdata->len + irqdata->len < count &&
1247 counter->state != PERF_COUNTER_STATE_ERROR)
1248 return -ERESTARTSYS;
1250 mutex_lock(&counter->mutex);
1252 /* Drain pending data first: */
1253 res = perf_copy_usrdata(usrdata, buf, count);
1254 if (res < 0 || res == count)
1257 /* Switch irq buffer: */
1258 usrdata = perf_switch_irq_data(counter);
1259 res2 = perf_copy_usrdata(usrdata, buf + res, count - res);
1267 mutex_unlock(&counter->mutex);
1273 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1275 struct perf_counter *counter = file->private_data;
1277 switch (counter->hw_event.record_type) {
1278 case PERF_RECORD_SIMPLE:
1279 return perf_read_hw(counter, buf, count);
1281 case PERF_RECORD_IRQ:
1282 case PERF_RECORD_GROUP:
1283 return perf_read_irq_data(counter, buf, count,
1284 file->f_flags & O_NONBLOCK);
1289 static unsigned int perf_poll(struct file *file, poll_table *wait)
1291 struct perf_counter *counter = file->private_data;
1292 unsigned int events = 0;
1293 unsigned long flags;
1295 poll_wait(file, &counter->waitq, wait);
1297 spin_lock_irqsave(&counter->waitq.lock, flags);
1298 if (counter->usrdata->len || counter->irqdata->len)
1300 spin_unlock_irqrestore(&counter->waitq.lock, flags);
1305 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1307 struct perf_counter *counter = file->private_data;
1311 case PERF_COUNTER_IOC_ENABLE:
1312 perf_counter_enable_family(counter);
1314 case PERF_COUNTER_IOC_DISABLE:
1315 perf_counter_disable_family(counter);
1323 static const struct file_operations perf_fops = {
1324 .release = perf_release,
1327 .unlocked_ioctl = perf_ioctl,
1328 .compat_ioctl = perf_ioctl,
1332 * Generic software counter infrastructure
1335 static void perf_swcounter_update(struct perf_counter *counter)
1337 struct hw_perf_counter *hwc = &counter->hw;
1342 prev = atomic64_read(&hwc->prev_count);
1343 now = atomic64_read(&hwc->count);
1344 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
1349 atomic64_add(delta, &counter->count);
1350 atomic64_sub(delta, &hwc->period_left);
1353 static void perf_swcounter_set_period(struct perf_counter *counter)
1355 struct hw_perf_counter *hwc = &counter->hw;
1356 s64 left = atomic64_read(&hwc->period_left);
1357 s64 period = hwc->irq_period;
1359 if (unlikely(left <= -period)) {
1361 atomic64_set(&hwc->period_left, left);
1364 if (unlikely(left <= 0)) {
1366 atomic64_add(period, &hwc->period_left);
1369 atomic64_set(&hwc->prev_count, -left);
1370 atomic64_set(&hwc->count, -left);
1373 static void perf_swcounter_save_and_restart(struct perf_counter *counter)
1375 perf_swcounter_update(counter);
1376 perf_swcounter_set_period(counter);
1379 static void perf_swcounter_store_irq(struct perf_counter *counter, u64 data)
1381 struct perf_data *irqdata = counter->irqdata;
1383 if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
1386 u64 *p = (u64 *) &irqdata->data[irqdata->len];
1389 irqdata->len += sizeof(u64);
1393 static void perf_swcounter_handle_group(struct perf_counter *sibling)
1395 struct perf_counter *counter, *group_leader = sibling->group_leader;
1397 list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
1398 perf_swcounter_update(counter);
1399 perf_swcounter_store_irq(sibling, counter->hw_event.type);
1400 perf_swcounter_store_irq(sibling, atomic64_read(&counter->count));
1404 static void perf_swcounter_interrupt(struct perf_counter *counter,
1405 int nmi, struct pt_regs *regs)
1407 perf_swcounter_save_and_restart(counter);
1409 switch (counter->hw_event.record_type) {
1410 case PERF_RECORD_SIMPLE:
1413 case PERF_RECORD_IRQ:
1414 perf_swcounter_store_irq(counter, instruction_pointer(regs));
1417 case PERF_RECORD_GROUP:
1418 perf_swcounter_handle_group(counter);
1423 counter->wakeup_pending = 1;
1424 set_tsk_thread_flag(current, TIF_PERF_COUNTERS);
1426 wake_up(&counter->waitq);
1429 static int perf_swcounter_match(struct perf_counter *counter,
1430 enum hw_event_types event,
1431 struct pt_regs *regs)
1433 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1436 if (counter->hw_event.raw)
1439 if (counter->hw_event.type != event)
1442 if (counter->hw_event.exclude_user && user_mode(regs))
1445 if (counter->hw_event.exclude_kernel && !user_mode(regs))
1451 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
1452 enum hw_event_types event, u64 nr,
1453 int nmi, struct pt_regs *regs)
1455 struct perf_counter *counter;
1456 unsigned long flags;
1459 if (list_empty(&ctx->counter_list))
1462 spin_lock_irqsave(&ctx->lock, flags);
1465 * XXX: make counter_list RCU safe
1467 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1468 if (perf_swcounter_match(counter, event, regs)) {
1469 neg = atomic64_add_negative(nr, &counter->hw.count);
1470 if (counter->hw.irq_period && !neg)
1471 perf_swcounter_interrupt(counter, nmi, regs);
1475 spin_unlock_irqrestore(&ctx->lock, flags);
1478 void perf_swcounter_event(enum hw_event_types event, u64 nr,
1479 int nmi, struct pt_regs *regs)
1481 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
1483 perf_swcounter_ctx_event(&cpuctx->ctx, event, nr, nmi, regs);
1484 if (cpuctx->task_ctx)
1485 perf_swcounter_ctx_event(cpuctx->task_ctx, event, nr, nmi, regs);
1487 put_cpu_var(perf_cpu_context);
1490 static void perf_swcounter_read(struct perf_counter *counter)
1492 perf_swcounter_update(counter);
1495 static int perf_swcounter_enable(struct perf_counter *counter)
1497 perf_swcounter_set_period(counter);
1501 static void perf_swcounter_disable(struct perf_counter *counter)
1503 perf_swcounter_update(counter);
1506 static const struct hw_perf_counter_ops perf_ops_generic = {
1507 .enable = perf_swcounter_enable,
1508 .disable = perf_swcounter_disable,
1509 .read = perf_swcounter_read,
1513 * Software counter: cpu wall time clock
1516 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
1518 int cpu = raw_smp_processor_id();
1520 atomic64_set(&counter->hw.prev_count, cpu_clock(cpu));
1524 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
1526 int cpu = raw_smp_processor_id();
1530 now = cpu_clock(cpu);
1531 prev = atomic64_read(&counter->hw.prev_count);
1532 atomic64_set(&counter->hw.prev_count, now);
1533 atomic64_add(now - prev, &counter->count);
1536 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
1538 cpu_clock_perf_counter_update(counter);
1541 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
1543 cpu_clock_perf_counter_update(counter);
1546 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
1547 .enable = cpu_clock_perf_counter_enable,
1548 .disable = cpu_clock_perf_counter_disable,
1549 .read = cpu_clock_perf_counter_read,
1553 * Software counter: task time clock
1557 * Called from within the scheduler:
1559 static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
1561 struct task_struct *curr = counter->task;
1564 delta = __task_delta_exec(curr, update);
1566 return curr->se.sum_exec_runtime + delta;
1569 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
1574 prev = atomic64_read(&counter->hw.prev_count);
1576 atomic64_set(&counter->hw.prev_count, now);
1580 atomic64_add(delta, &counter->count);
1583 static void task_clock_perf_counter_read(struct perf_counter *counter)
1585 u64 now = task_clock_perf_counter_val(counter, 1);
1587 task_clock_perf_counter_update(counter, now);
1590 static int task_clock_perf_counter_enable(struct perf_counter *counter)
1592 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1593 atomic64_set(&counter->hw.prev_count,
1594 task_clock_perf_counter_val(counter, 0));
1599 static void task_clock_perf_counter_disable(struct perf_counter *counter)
1601 u64 now = task_clock_perf_counter_val(counter, 0);
1603 task_clock_perf_counter_update(counter, now);
1606 static const struct hw_perf_counter_ops perf_ops_task_clock = {
1607 .enable = task_clock_perf_counter_enable,
1608 .disable = task_clock_perf_counter_disable,
1609 .read = task_clock_perf_counter_read,
1613 * Software counter: context switches
1616 static u64 get_context_switches(struct perf_counter *counter)
1618 struct task_struct *curr = counter->ctx->task;
1621 return curr->nvcsw + curr->nivcsw;
1622 return cpu_nr_switches(smp_processor_id());
1625 static void context_switches_perf_counter_update(struct perf_counter *counter)
1630 prev = atomic64_read(&counter->hw.prev_count);
1631 now = get_context_switches(counter);
1633 atomic64_set(&counter->hw.prev_count, now);
1637 atomic64_add(delta, &counter->count);
1640 static void context_switches_perf_counter_read(struct perf_counter *counter)
1642 context_switches_perf_counter_update(counter);
1645 static int context_switches_perf_counter_enable(struct perf_counter *counter)
1647 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1648 atomic64_set(&counter->hw.prev_count,
1649 get_context_switches(counter));
1653 static void context_switches_perf_counter_disable(struct perf_counter *counter)
1655 context_switches_perf_counter_update(counter);
1658 static const struct hw_perf_counter_ops perf_ops_context_switches = {
1659 .enable = context_switches_perf_counter_enable,
1660 .disable = context_switches_perf_counter_disable,
1661 .read = context_switches_perf_counter_read,
1665 * Software counter: cpu migrations
1668 static inline u64 get_cpu_migrations(struct perf_counter *counter)
1670 struct task_struct *curr = counter->ctx->task;
1673 return curr->se.nr_migrations;
1674 return cpu_nr_migrations(smp_processor_id());
1677 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1682 prev = atomic64_read(&counter->hw.prev_count);
1683 now = get_cpu_migrations(counter);
1685 atomic64_set(&counter->hw.prev_count, now);
1689 atomic64_add(delta, &counter->count);
1692 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1694 cpu_migrations_perf_counter_update(counter);
1697 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1699 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1700 atomic64_set(&counter->hw.prev_count,
1701 get_cpu_migrations(counter));
1705 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1707 cpu_migrations_perf_counter_update(counter);
1710 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1711 .enable = cpu_migrations_perf_counter_enable,
1712 .disable = cpu_migrations_perf_counter_disable,
1713 .read = cpu_migrations_perf_counter_read,
1716 static const struct hw_perf_counter_ops *
1717 sw_perf_counter_init(struct perf_counter *counter)
1719 struct perf_counter_hw_event *hw_event = &counter->hw_event;
1720 const struct hw_perf_counter_ops *hw_ops = NULL;
1721 struct hw_perf_counter *hwc = &counter->hw;
1724 * Software counters (currently) can't in general distinguish
1725 * between user, kernel and hypervisor events.
1726 * However, context switches and cpu migrations are considered
1727 * to be kernel events, and page faults are never hypervisor
1730 switch (counter->hw_event.type) {
1731 case PERF_COUNT_CPU_CLOCK:
1732 if (!(counter->hw_event.exclude_user ||
1733 counter->hw_event.exclude_kernel ||
1734 counter->hw_event.exclude_hv))
1735 hw_ops = &perf_ops_cpu_clock;
1737 case PERF_COUNT_TASK_CLOCK:
1738 if (counter->hw_event.exclude_user ||
1739 counter->hw_event.exclude_kernel ||
1740 counter->hw_event.exclude_hv)
1743 * If the user instantiates this as a per-cpu counter,
1744 * use the cpu_clock counter instead.
1746 if (counter->ctx->task)
1747 hw_ops = &perf_ops_task_clock;
1749 hw_ops = &perf_ops_cpu_clock;
1751 case PERF_COUNT_PAGE_FAULTS:
1752 case PERF_COUNT_PAGE_FAULTS_MIN:
1753 case PERF_COUNT_PAGE_FAULTS_MAJ:
1754 hw_ops = &perf_ops_generic;
1756 case PERF_COUNT_CONTEXT_SWITCHES:
1757 if (!counter->hw_event.exclude_kernel)
1758 hw_ops = &perf_ops_context_switches;
1760 case PERF_COUNT_CPU_MIGRATIONS:
1761 if (!counter->hw_event.exclude_kernel)
1762 hw_ops = &perf_ops_cpu_migrations;
1769 hwc->irq_period = hw_event->irq_period;
1775 * Allocate and initialize a counter structure
1777 static struct perf_counter *
1778 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1780 struct perf_counter_context *ctx,
1781 struct perf_counter *group_leader,
1784 const struct hw_perf_counter_ops *hw_ops;
1785 struct perf_counter *counter;
1787 counter = kzalloc(sizeof(*counter), gfpflags);
1792 * Single counters are their own group leaders, with an
1793 * empty sibling list:
1796 group_leader = counter;
1798 mutex_init(&counter->mutex);
1799 INIT_LIST_HEAD(&counter->list_entry);
1800 INIT_LIST_HEAD(&counter->sibling_list);
1801 init_waitqueue_head(&counter->waitq);
1803 INIT_LIST_HEAD(&counter->child_list);
1805 counter->irqdata = &counter->data[0];
1806 counter->usrdata = &counter->data[1];
1808 counter->hw_event = *hw_event;
1809 counter->wakeup_pending = 0;
1810 counter->group_leader = group_leader;
1811 counter->hw_ops = NULL;
1814 counter->state = PERF_COUNTER_STATE_INACTIVE;
1815 if (hw_event->disabled)
1816 counter->state = PERF_COUNTER_STATE_OFF;
1819 if (!hw_event->raw && hw_event->type < 0)
1820 hw_ops = sw_perf_counter_init(counter);
1822 hw_ops = hw_perf_counter_init(counter);
1828 counter->hw_ops = hw_ops;
1834 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
1836 * @hw_event_uptr: event type attributes for monitoring/sampling
1839 * @group_fd: group leader counter fd
1841 SYSCALL_DEFINE5(perf_counter_open,
1842 const struct perf_counter_hw_event __user *, hw_event_uptr,
1843 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
1845 struct perf_counter *counter, *group_leader;
1846 struct perf_counter_hw_event hw_event;
1847 struct perf_counter_context *ctx;
1848 struct file *counter_file = NULL;
1849 struct file *group_file = NULL;
1850 int fput_needed = 0;
1851 int fput_needed2 = 0;
1854 /* for future expandability... */
1858 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1862 * Get the target context (task or percpu):
1864 ctx = find_get_context(pid, cpu);
1866 return PTR_ERR(ctx);
1869 * Look up the group leader (we will attach this counter to it):
1871 group_leader = NULL;
1872 if (group_fd != -1) {
1874 group_file = fget_light(group_fd, &fput_needed);
1876 goto err_put_context;
1877 if (group_file->f_op != &perf_fops)
1878 goto err_put_context;
1880 group_leader = group_file->private_data;
1882 * Do not allow a recursive hierarchy (this new sibling
1883 * becoming part of another group-sibling):
1885 if (group_leader->group_leader != group_leader)
1886 goto err_put_context;
1888 * Do not allow to attach to a group in a different
1889 * task or CPU context:
1891 if (group_leader->ctx != ctx)
1892 goto err_put_context;
1894 * Only a group leader can be exclusive or pinned
1896 if (hw_event.exclusive || hw_event.pinned)
1897 goto err_put_context;
1901 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
1904 goto err_put_context;
1906 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
1908 goto err_free_put_context;
1910 counter_file = fget_light(ret, &fput_needed2);
1912 goto err_free_put_context;
1914 counter->filp = counter_file;
1915 mutex_lock(&ctx->mutex);
1916 perf_install_in_context(ctx, counter, cpu);
1917 mutex_unlock(&ctx->mutex);
1919 fput_light(counter_file, fput_needed2);
1922 fput_light(group_file, fput_needed);
1926 err_free_put_context:
1936 * Initialize the perf_counter context in a task_struct:
1939 __perf_counter_init_context(struct perf_counter_context *ctx,
1940 struct task_struct *task)
1942 memset(ctx, 0, sizeof(*ctx));
1943 spin_lock_init(&ctx->lock);
1944 mutex_init(&ctx->mutex);
1945 INIT_LIST_HEAD(&ctx->counter_list);
1950 * inherit a counter from parent task to child task:
1952 static struct perf_counter *
1953 inherit_counter(struct perf_counter *parent_counter,
1954 struct task_struct *parent,
1955 struct perf_counter_context *parent_ctx,
1956 struct task_struct *child,
1957 struct perf_counter *group_leader,
1958 struct perf_counter_context *child_ctx)
1960 struct perf_counter *child_counter;
1963 * Instead of creating recursive hierarchies of counters,
1964 * we link inherited counters back to the original parent,
1965 * which has a filp for sure, which we use as the reference
1968 if (parent_counter->parent)
1969 parent_counter = parent_counter->parent;
1971 child_counter = perf_counter_alloc(&parent_counter->hw_event,
1972 parent_counter->cpu, child_ctx,
1973 group_leader, GFP_KERNEL);
1978 * Link it up in the child's context:
1980 child_counter->task = child;
1981 list_add_counter(child_counter, child_ctx);
1982 child_ctx->nr_counters++;
1984 child_counter->parent = parent_counter;
1986 * inherit into child's child as well:
1988 child_counter->hw_event.inherit = 1;
1991 * Get a reference to the parent filp - we will fput it
1992 * when the child counter exits. This is safe to do because
1993 * we are in the parent and we know that the filp still
1994 * exists and has a nonzero count:
1996 atomic_long_inc(&parent_counter->filp->f_count);
1999 * Link this into the parent counter's child list
2001 mutex_lock(&parent_counter->mutex);
2002 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2005 * Make the child state follow the state of the parent counter,
2006 * not its hw_event.disabled bit. We hold the parent's mutex,
2007 * so we won't race with perf_counter_{en,dis}able_family.
2009 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2010 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2012 child_counter->state = PERF_COUNTER_STATE_OFF;
2014 mutex_unlock(&parent_counter->mutex);
2016 return child_counter;
2019 static int inherit_group(struct perf_counter *parent_counter,
2020 struct task_struct *parent,
2021 struct perf_counter_context *parent_ctx,
2022 struct task_struct *child,
2023 struct perf_counter_context *child_ctx)
2025 struct perf_counter *leader;
2026 struct perf_counter *sub;
2028 leader = inherit_counter(parent_counter, parent, parent_ctx,
2029 child, NULL, child_ctx);
2032 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2033 if (!inherit_counter(sub, parent, parent_ctx,
2034 child, leader, child_ctx))
2040 static void sync_child_counter(struct perf_counter *child_counter,
2041 struct perf_counter *parent_counter)
2043 u64 parent_val, child_val;
2045 parent_val = atomic64_read(&parent_counter->count);
2046 child_val = atomic64_read(&child_counter->count);
2049 * Add back the child's count to the parent's count:
2051 atomic64_add(child_val, &parent_counter->count);
2054 * Remove this counter from the parent's list
2056 mutex_lock(&parent_counter->mutex);
2057 list_del_init(&child_counter->child_list);
2058 mutex_unlock(&parent_counter->mutex);
2061 * Release the parent counter, if this was the last
2064 fput(parent_counter->filp);
2068 __perf_counter_exit_task(struct task_struct *child,
2069 struct perf_counter *child_counter,
2070 struct perf_counter_context *child_ctx)
2072 struct perf_counter *parent_counter;
2073 struct perf_counter *sub, *tmp;
2076 * If we do not self-reap then we have to wait for the
2077 * child task to unschedule (it will happen for sure),
2078 * so that its counter is at its final count. (This
2079 * condition triggers rarely - child tasks usually get
2080 * off their CPU before the parent has a chance to
2081 * get this far into the reaping action)
2083 if (child != current) {
2084 wait_task_inactive(child, 0);
2085 list_del_init(&child_counter->list_entry);
2087 struct perf_cpu_context *cpuctx;
2088 unsigned long flags;
2092 * Disable and unlink this counter.
2094 * Be careful about zapping the list - IRQ/NMI context
2095 * could still be processing it:
2097 curr_rq_lock_irq_save(&flags);
2098 perf_flags = hw_perf_save_disable();
2100 cpuctx = &__get_cpu_var(perf_cpu_context);
2102 group_sched_out(child_counter, cpuctx, child_ctx);
2104 list_del_init(&child_counter->list_entry);
2106 child_ctx->nr_counters--;
2108 hw_perf_restore(perf_flags);
2109 curr_rq_unlock_irq_restore(&flags);
2112 parent_counter = child_counter->parent;
2114 * It can happen that parent exits first, and has counters
2115 * that are still around due to the child reference. These
2116 * counters need to be zapped - but otherwise linger.
2118 if (parent_counter) {
2119 sync_child_counter(child_counter, parent_counter);
2120 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
2123 sync_child_counter(sub, sub->parent);
2127 kfree(child_counter);
2132 * When a child task exits, feed back counter values to parent counters.
2134 * Note: we may be running in child context, but the PID is not hashed
2135 * anymore so new counters will not be added.
2137 void perf_counter_exit_task(struct task_struct *child)
2139 struct perf_counter *child_counter, *tmp;
2140 struct perf_counter_context *child_ctx;
2142 child_ctx = &child->perf_counter_ctx;
2144 if (likely(!child_ctx->nr_counters))
2147 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
2149 __perf_counter_exit_task(child, child_counter, child_ctx);
2153 * Initialize the perf_counter context in task_struct
2155 void perf_counter_init_task(struct task_struct *child)
2157 struct perf_counter_context *child_ctx, *parent_ctx;
2158 struct perf_counter *counter;
2159 struct task_struct *parent = current;
2161 child_ctx = &child->perf_counter_ctx;
2162 parent_ctx = &parent->perf_counter_ctx;
2164 __perf_counter_init_context(child_ctx, child);
2167 * This is executed from the parent task context, so inherit
2168 * counters that have been marked for cloning:
2171 if (likely(!parent_ctx->nr_counters))
2175 * Lock the parent list. No need to lock the child - not PID
2176 * hashed yet and not running, so nobody can access it.
2178 mutex_lock(&parent_ctx->mutex);
2181 * We dont have to disable NMIs - we are only looking at
2182 * the list, not manipulating it:
2184 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2185 if (!counter->hw_event.inherit)
2188 if (inherit_group(counter, parent,
2189 parent_ctx, child, child_ctx))
2193 mutex_unlock(&parent_ctx->mutex);
2196 static void __cpuinit perf_counter_init_cpu(int cpu)
2198 struct perf_cpu_context *cpuctx;
2200 cpuctx = &per_cpu(perf_cpu_context, cpu);
2201 __perf_counter_init_context(&cpuctx->ctx, NULL);
2203 mutex_lock(&perf_resource_mutex);
2204 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2205 mutex_unlock(&perf_resource_mutex);
2207 hw_perf_counter_setup(cpu);
2210 #ifdef CONFIG_HOTPLUG_CPU
2211 static void __perf_counter_exit_cpu(void *info)
2213 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2214 struct perf_counter_context *ctx = &cpuctx->ctx;
2215 struct perf_counter *counter, *tmp;
2217 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2218 __perf_counter_remove_from_context(counter);
2220 static void perf_counter_exit_cpu(int cpu)
2222 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2223 struct perf_counter_context *ctx = &cpuctx->ctx;
2225 mutex_lock(&ctx->mutex);
2226 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2227 mutex_unlock(&ctx->mutex);
2230 static inline void perf_counter_exit_cpu(int cpu) { }
2233 static int __cpuinit
2234 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2236 unsigned int cpu = (long)hcpu;
2240 case CPU_UP_PREPARE:
2241 case CPU_UP_PREPARE_FROZEN:
2242 perf_counter_init_cpu(cpu);
2245 case CPU_DOWN_PREPARE:
2246 case CPU_DOWN_PREPARE_FROZEN:
2247 perf_counter_exit_cpu(cpu);
2257 static struct notifier_block __cpuinitdata perf_cpu_nb = {
2258 .notifier_call = perf_cpu_notify,
2261 static int __init perf_counter_init(void)
2263 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2264 (void *)(long)smp_processor_id());
2265 register_cpu_notifier(&perf_cpu_nb);
2269 early_initcall(perf_counter_init);
2271 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2273 return sprintf(buf, "%d\n", perf_reserved_percpu);
2277 perf_set_reserve_percpu(struct sysdev_class *class,
2281 struct perf_cpu_context *cpuctx;
2285 err = strict_strtoul(buf, 10, &val);
2288 if (val > perf_max_counters)
2291 mutex_lock(&perf_resource_mutex);
2292 perf_reserved_percpu = val;
2293 for_each_online_cpu(cpu) {
2294 cpuctx = &per_cpu(perf_cpu_context, cpu);
2295 spin_lock_irq(&cpuctx->ctx.lock);
2296 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2297 perf_max_counters - perf_reserved_percpu);
2298 cpuctx->max_pertask = mpt;
2299 spin_unlock_irq(&cpuctx->ctx.lock);
2301 mutex_unlock(&perf_resource_mutex);
2306 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2308 return sprintf(buf, "%d\n", perf_overcommit);
2312 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2317 err = strict_strtoul(buf, 10, &val);
2323 mutex_lock(&perf_resource_mutex);
2324 perf_overcommit = val;
2325 mutex_unlock(&perf_resource_mutex);
2330 static SYSDEV_CLASS_ATTR(
2333 perf_show_reserve_percpu,
2334 perf_set_reserve_percpu
2337 static SYSDEV_CLASS_ATTR(
2340 perf_show_overcommit,
2344 static struct attribute *perfclass_attrs[] = {
2345 &attr_reserve_percpu.attr,
2346 &attr_overcommit.attr,
2350 static struct attribute_group perfclass_attr_group = {
2351 .attrs = perfclass_attrs,
2352 .name = "perf_counters",
2355 static int __init perf_counter_sysfs_init(void)
2357 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2358 &perfclass_attr_group);
2360 device_initcall(perf_counter_sysfs_init);