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>
25 #include <linux/rculist.h>
27 #include <asm/irq_regs.h>
30 * Each CPU has a list of per CPU counters:
32 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
34 int perf_max_counters __read_mostly = 1;
35 static int perf_reserved_percpu __read_mostly;
36 static int perf_overcommit __read_mostly = 1;
39 * Mutex for (sysadmin-configurable) counter reservations:
41 static DEFINE_MUTEX(perf_resource_mutex);
44 * Architecture provided APIs - weak aliases:
46 extern __weak const struct hw_perf_counter_ops *
47 hw_perf_counter_init(struct perf_counter *counter)
52 u64 __weak hw_perf_save_disable(void) { return 0; }
53 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
54 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
55 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
56 struct perf_cpu_context *cpuctx,
57 struct perf_counter_context *ctx, int cpu)
62 void __weak perf_counter_print_debug(void) { }
65 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
67 struct perf_counter *group_leader = counter->group_leader;
70 * Depending on whether it is a standalone or sibling counter,
71 * add it straight to the context's counter list, or to the group
72 * leader's sibling list:
74 if (counter->group_leader == counter)
75 list_add_tail(&counter->list_entry, &ctx->counter_list);
77 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
79 list_add_rcu(&counter->event_entry, &ctx->event_list);
83 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
85 struct perf_counter *sibling, *tmp;
87 list_del_init(&counter->list_entry);
88 list_del_rcu(&counter->event_entry);
91 * If this was a group counter with sibling counters then
92 * upgrade the siblings to singleton counters by adding them
93 * to the context list directly:
95 list_for_each_entry_safe(sibling, tmp,
96 &counter->sibling_list, list_entry) {
98 list_move_tail(&sibling->list_entry, &ctx->counter_list);
99 sibling->group_leader = sibling;
104 counter_sched_out(struct perf_counter *counter,
105 struct perf_cpu_context *cpuctx,
106 struct perf_counter_context *ctx)
108 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
111 counter->state = PERF_COUNTER_STATE_INACTIVE;
112 counter->hw_ops->disable(counter);
115 if (!is_software_counter(counter))
116 cpuctx->active_oncpu--;
118 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
119 cpuctx->exclusive = 0;
123 group_sched_out(struct perf_counter *group_counter,
124 struct perf_cpu_context *cpuctx,
125 struct perf_counter_context *ctx)
127 struct perf_counter *counter;
129 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
132 counter_sched_out(group_counter, cpuctx, ctx);
135 * Schedule out siblings (if any):
137 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
138 counter_sched_out(counter, cpuctx, ctx);
140 if (group_counter->hw_event.exclusive)
141 cpuctx->exclusive = 0;
145 * Cross CPU call to remove a performance counter
147 * We disable the counter on the hardware level first. After that we
148 * remove it from the context list.
150 static void __perf_counter_remove_from_context(void *info)
152 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
153 struct perf_counter *counter = info;
154 struct perf_counter_context *ctx = counter->ctx;
159 * If this is a task context, we need to check whether it is
160 * the current task context of this cpu. If not it has been
161 * scheduled out before the smp call arrived.
163 if (ctx->task && cpuctx->task_ctx != ctx)
166 curr_rq_lock_irq_save(&flags);
167 spin_lock(&ctx->lock);
169 counter_sched_out(counter, cpuctx, ctx);
171 counter->task = NULL;
175 * Protect the list operation against NMI by disabling the
176 * counters on a global level. NOP for non NMI based counters.
178 perf_flags = hw_perf_save_disable();
179 list_del_counter(counter, ctx);
180 hw_perf_restore(perf_flags);
184 * Allow more per task counters with respect to the
187 cpuctx->max_pertask =
188 min(perf_max_counters - ctx->nr_counters,
189 perf_max_counters - perf_reserved_percpu);
192 spin_unlock(&ctx->lock);
193 curr_rq_unlock_irq_restore(&flags);
198 * Remove the counter from a task's (or a CPU's) list of counters.
200 * Must be called with counter->mutex and ctx->mutex held.
202 * CPU counters are removed with a smp call. For task counters we only
203 * call when the task is on a CPU.
205 static void perf_counter_remove_from_context(struct perf_counter *counter)
207 struct perf_counter_context *ctx = counter->ctx;
208 struct task_struct *task = ctx->task;
212 * Per cpu counters are removed via an smp call and
213 * the removal is always sucessful.
215 smp_call_function_single(counter->cpu,
216 __perf_counter_remove_from_context,
222 task_oncpu_function_call(task, __perf_counter_remove_from_context,
225 spin_lock_irq(&ctx->lock);
227 * If the context is active we need to retry the smp call.
229 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
230 spin_unlock_irq(&ctx->lock);
235 * The lock prevents that this context is scheduled in so we
236 * can remove the counter safely, if the call above did not
239 if (!list_empty(&counter->list_entry)) {
241 list_del_counter(counter, ctx);
242 counter->task = NULL;
244 spin_unlock_irq(&ctx->lock);
248 * Cross CPU call to disable a performance counter
250 static void __perf_counter_disable(void *info)
252 struct perf_counter *counter = info;
253 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
254 struct perf_counter_context *ctx = counter->ctx;
258 * If this is a per-task counter, need to check whether this
259 * counter's task is the current task on this cpu.
261 if (ctx->task && cpuctx->task_ctx != ctx)
264 curr_rq_lock_irq_save(&flags);
265 spin_lock(&ctx->lock);
268 * If the counter is on, turn it off.
269 * If it is in error state, leave it in error state.
271 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
272 if (counter == counter->group_leader)
273 group_sched_out(counter, cpuctx, ctx);
275 counter_sched_out(counter, cpuctx, ctx);
276 counter->state = PERF_COUNTER_STATE_OFF;
279 spin_unlock(&ctx->lock);
280 curr_rq_unlock_irq_restore(&flags);
286 static void perf_counter_disable(struct perf_counter *counter)
288 struct perf_counter_context *ctx = counter->ctx;
289 struct task_struct *task = ctx->task;
293 * Disable the counter on the cpu that it's on
295 smp_call_function_single(counter->cpu, __perf_counter_disable,
301 task_oncpu_function_call(task, __perf_counter_disable, counter);
303 spin_lock_irq(&ctx->lock);
305 * If the counter is still active, we need to retry the cross-call.
307 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
308 spin_unlock_irq(&ctx->lock);
313 * Since we have the lock this context can't be scheduled
314 * in, so we can change the state safely.
316 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
317 counter->state = PERF_COUNTER_STATE_OFF;
319 spin_unlock_irq(&ctx->lock);
323 * Disable a counter and all its children.
325 static void perf_counter_disable_family(struct perf_counter *counter)
327 struct perf_counter *child;
329 perf_counter_disable(counter);
332 * Lock the mutex to protect the list of children
334 mutex_lock(&counter->mutex);
335 list_for_each_entry(child, &counter->child_list, child_list)
336 perf_counter_disable(child);
337 mutex_unlock(&counter->mutex);
341 counter_sched_in(struct perf_counter *counter,
342 struct perf_cpu_context *cpuctx,
343 struct perf_counter_context *ctx,
346 if (counter->state <= PERF_COUNTER_STATE_OFF)
349 counter->state = PERF_COUNTER_STATE_ACTIVE;
350 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
352 * The new state must be visible before we turn it on in the hardware:
356 if (counter->hw_ops->enable(counter)) {
357 counter->state = PERF_COUNTER_STATE_INACTIVE;
362 if (!is_software_counter(counter))
363 cpuctx->active_oncpu++;
366 if (counter->hw_event.exclusive)
367 cpuctx->exclusive = 1;
373 * Return 1 for a group consisting entirely of software counters,
374 * 0 if the group contains any hardware counters.
376 static int is_software_only_group(struct perf_counter *leader)
378 struct perf_counter *counter;
380 if (!is_software_counter(leader))
382 list_for_each_entry(counter, &leader->sibling_list, list_entry)
383 if (!is_software_counter(counter))
389 * Work out whether we can put this counter group on the CPU now.
391 static int group_can_go_on(struct perf_counter *counter,
392 struct perf_cpu_context *cpuctx,
396 * Groups consisting entirely of software counters can always go on.
398 if (is_software_only_group(counter))
401 * If an exclusive group is already on, no other hardware
402 * counters can go on.
404 if (cpuctx->exclusive)
407 * If this group is exclusive and there are already
408 * counters on the CPU, it can't go on.
410 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
413 * Otherwise, try to add it if all previous groups were able
420 * Cross CPU call to install and enable a performance counter
422 static void __perf_install_in_context(void *info)
424 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
425 struct perf_counter *counter = info;
426 struct perf_counter_context *ctx = counter->ctx;
427 struct perf_counter *leader = counter->group_leader;
428 int cpu = smp_processor_id();
434 * If this is a task context, we need to check whether it is
435 * the current task context of this cpu. If not it has been
436 * scheduled out before the smp call arrived.
438 if (ctx->task && cpuctx->task_ctx != ctx)
441 curr_rq_lock_irq_save(&flags);
442 spin_lock(&ctx->lock);
445 * Protect the list operation against NMI by disabling the
446 * counters on a global level. NOP for non NMI based counters.
448 perf_flags = hw_perf_save_disable();
450 list_add_counter(counter, ctx);
452 counter->prev_state = PERF_COUNTER_STATE_OFF;
455 * Don't put the counter on if it is disabled or if
456 * it is in a group and the group isn't on.
458 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
459 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
463 * An exclusive counter can't go on if there are already active
464 * hardware counters, and no hardware counter can go on if there
465 * is already an exclusive counter on.
467 if (!group_can_go_on(counter, cpuctx, 1))
470 err = counter_sched_in(counter, cpuctx, ctx, cpu);
474 * This counter couldn't go on. If it is in a group
475 * then we have to pull the whole group off.
476 * If the counter group is pinned then put it in error state.
478 if (leader != counter)
479 group_sched_out(leader, cpuctx, ctx);
480 if (leader->hw_event.pinned)
481 leader->state = PERF_COUNTER_STATE_ERROR;
484 if (!err && !ctx->task && cpuctx->max_pertask)
485 cpuctx->max_pertask--;
488 hw_perf_restore(perf_flags);
490 spin_unlock(&ctx->lock);
491 curr_rq_unlock_irq_restore(&flags);
495 * Attach a performance counter to a context
497 * First we add the counter to the list with the hardware enable bit
498 * in counter->hw_config cleared.
500 * If the counter is attached to a task which is on a CPU we use a smp
501 * call to enable it in the task context. The task might have been
502 * scheduled away, but we check this in the smp call again.
504 * Must be called with ctx->mutex held.
507 perf_install_in_context(struct perf_counter_context *ctx,
508 struct perf_counter *counter,
511 struct task_struct *task = ctx->task;
515 * Per cpu counters are installed via an smp call and
516 * the install is always sucessful.
518 smp_call_function_single(cpu, __perf_install_in_context,
523 counter->task = task;
525 task_oncpu_function_call(task, __perf_install_in_context,
528 spin_lock_irq(&ctx->lock);
530 * we need to retry the smp call.
532 if (ctx->is_active && list_empty(&counter->list_entry)) {
533 spin_unlock_irq(&ctx->lock);
538 * The lock prevents that this context is scheduled in so we
539 * can add the counter safely, if it the call above did not
542 if (list_empty(&counter->list_entry)) {
543 list_add_counter(counter, ctx);
546 spin_unlock_irq(&ctx->lock);
550 * Cross CPU call to enable a performance counter
552 static void __perf_counter_enable(void *info)
554 struct perf_counter *counter = info;
555 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
556 struct perf_counter_context *ctx = counter->ctx;
557 struct perf_counter *leader = counter->group_leader;
562 * If this is a per-task counter, need to check whether this
563 * counter's task is the current task on this cpu.
565 if (ctx->task && cpuctx->task_ctx != ctx)
568 curr_rq_lock_irq_save(&flags);
569 spin_lock(&ctx->lock);
571 counter->prev_state = counter->state;
572 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
574 counter->state = PERF_COUNTER_STATE_INACTIVE;
577 * If the counter is in a group and isn't the group leader,
578 * then don't put it on unless the group is on.
580 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
583 if (!group_can_go_on(counter, cpuctx, 1))
586 err = counter_sched_in(counter, cpuctx, ctx,
591 * If this counter can't go on and it's part of a
592 * group, then the whole group has to come off.
594 if (leader != counter)
595 group_sched_out(leader, cpuctx, ctx);
596 if (leader->hw_event.pinned)
597 leader->state = PERF_COUNTER_STATE_ERROR;
601 spin_unlock(&ctx->lock);
602 curr_rq_unlock_irq_restore(&flags);
608 static void perf_counter_enable(struct perf_counter *counter)
610 struct perf_counter_context *ctx = counter->ctx;
611 struct task_struct *task = ctx->task;
615 * Enable the counter on the cpu that it's on
617 smp_call_function_single(counter->cpu, __perf_counter_enable,
622 spin_lock_irq(&ctx->lock);
623 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
627 * If the counter is in error state, clear that first.
628 * That way, if we see the counter in error state below, we
629 * know that it has gone back into error state, as distinct
630 * from the task having been scheduled away before the
631 * cross-call arrived.
633 if (counter->state == PERF_COUNTER_STATE_ERROR)
634 counter->state = PERF_COUNTER_STATE_OFF;
637 spin_unlock_irq(&ctx->lock);
638 task_oncpu_function_call(task, __perf_counter_enable, counter);
640 spin_lock_irq(&ctx->lock);
643 * If the context is active and the counter is still off,
644 * we need to retry the cross-call.
646 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
650 * Since we have the lock this context can't be scheduled
651 * in, so we can change the state safely.
653 if (counter->state == PERF_COUNTER_STATE_OFF)
654 counter->state = PERF_COUNTER_STATE_INACTIVE;
656 spin_unlock_irq(&ctx->lock);
660 * Enable a counter and all its children.
662 static void perf_counter_enable_family(struct perf_counter *counter)
664 struct perf_counter *child;
666 perf_counter_enable(counter);
669 * Lock the mutex to protect the list of children
671 mutex_lock(&counter->mutex);
672 list_for_each_entry(child, &counter->child_list, child_list)
673 perf_counter_enable(child);
674 mutex_unlock(&counter->mutex);
677 void __perf_counter_sched_out(struct perf_counter_context *ctx,
678 struct perf_cpu_context *cpuctx)
680 struct perf_counter *counter;
683 spin_lock(&ctx->lock);
685 if (likely(!ctx->nr_counters))
688 flags = hw_perf_save_disable();
689 if (ctx->nr_active) {
690 list_for_each_entry(counter, &ctx->counter_list, list_entry)
691 group_sched_out(counter, cpuctx, ctx);
693 hw_perf_restore(flags);
695 spin_unlock(&ctx->lock);
699 * Called from scheduler to remove the counters of the current task,
700 * with interrupts disabled.
702 * We stop each counter and update the counter value in counter->count.
704 * This does not protect us against NMI, but disable()
705 * sets the disabled bit in the control field of counter _before_
706 * accessing the counter control register. If a NMI hits, then it will
707 * not restart the counter.
709 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
711 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
712 struct perf_counter_context *ctx = &task->perf_counter_ctx;
713 struct pt_regs *regs;
715 if (likely(!cpuctx->task_ctx))
718 regs = task_pt_regs(task);
719 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs);
720 __perf_counter_sched_out(ctx, cpuctx);
722 cpuctx->task_ctx = NULL;
725 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
727 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
731 group_sched_in(struct perf_counter *group_counter,
732 struct perf_cpu_context *cpuctx,
733 struct perf_counter_context *ctx,
736 struct perf_counter *counter, *partial_group;
739 if (group_counter->state == PERF_COUNTER_STATE_OFF)
742 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
744 return ret < 0 ? ret : 0;
746 group_counter->prev_state = group_counter->state;
747 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
751 * Schedule in siblings as one group (if any):
753 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
754 counter->prev_state = counter->state;
755 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
756 partial_group = counter;
765 * Groups can be scheduled in as one unit only, so undo any
766 * partial group before returning:
768 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
769 if (counter == partial_group)
771 counter_sched_out(counter, cpuctx, ctx);
773 counter_sched_out(group_counter, cpuctx, ctx);
779 __perf_counter_sched_in(struct perf_counter_context *ctx,
780 struct perf_cpu_context *cpuctx, int cpu)
782 struct perf_counter *counter;
786 spin_lock(&ctx->lock);
788 if (likely(!ctx->nr_counters))
791 flags = hw_perf_save_disable();
794 * First go through the list and put on any pinned groups
795 * in order to give them the best chance of going on.
797 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
798 if (counter->state <= PERF_COUNTER_STATE_OFF ||
799 !counter->hw_event.pinned)
801 if (counter->cpu != -1 && counter->cpu != cpu)
804 if (group_can_go_on(counter, cpuctx, 1))
805 group_sched_in(counter, cpuctx, ctx, cpu);
808 * If this pinned group hasn't been scheduled,
809 * put it in error state.
811 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
812 counter->state = PERF_COUNTER_STATE_ERROR;
815 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
817 * Ignore counters in OFF or ERROR state, and
818 * ignore pinned counters since we did them already.
820 if (counter->state <= PERF_COUNTER_STATE_OFF ||
821 counter->hw_event.pinned)
825 * Listen to the 'cpu' scheduling filter constraint
828 if (counter->cpu != -1 && counter->cpu != cpu)
831 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
832 if (group_sched_in(counter, cpuctx, ctx, cpu))
836 hw_perf_restore(flags);
838 spin_unlock(&ctx->lock);
842 * Called from scheduler to add the counters of the current task
843 * with interrupts disabled.
845 * We restore the counter value and then enable it.
847 * This does not protect us against NMI, but enable()
848 * sets the enabled bit in the control field of counter _before_
849 * accessing the counter control register. If a NMI hits, then it will
850 * keep the counter running.
852 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
854 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
855 struct perf_counter_context *ctx = &task->perf_counter_ctx;
857 __perf_counter_sched_in(ctx, cpuctx, cpu);
858 cpuctx->task_ctx = ctx;
861 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
863 struct perf_counter_context *ctx = &cpuctx->ctx;
865 __perf_counter_sched_in(ctx, cpuctx, cpu);
868 int perf_counter_task_disable(void)
870 struct task_struct *curr = current;
871 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
872 struct perf_counter *counter;
877 if (likely(!ctx->nr_counters))
880 curr_rq_lock_irq_save(&flags);
881 cpu = smp_processor_id();
883 /* force the update of the task clock: */
884 __task_delta_exec(curr, 1);
886 perf_counter_task_sched_out(curr, cpu);
888 spin_lock(&ctx->lock);
891 * Disable all the counters:
893 perf_flags = hw_perf_save_disable();
895 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
896 if (counter->state != PERF_COUNTER_STATE_ERROR)
897 counter->state = PERF_COUNTER_STATE_OFF;
900 hw_perf_restore(perf_flags);
902 spin_unlock(&ctx->lock);
904 curr_rq_unlock_irq_restore(&flags);
909 int perf_counter_task_enable(void)
911 struct task_struct *curr = current;
912 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
913 struct perf_counter *counter;
918 if (likely(!ctx->nr_counters))
921 curr_rq_lock_irq_save(&flags);
922 cpu = smp_processor_id();
924 /* force the update of the task clock: */
925 __task_delta_exec(curr, 1);
927 perf_counter_task_sched_out(curr, cpu);
929 spin_lock(&ctx->lock);
932 * Disable all the counters:
934 perf_flags = hw_perf_save_disable();
936 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
937 if (counter->state > PERF_COUNTER_STATE_OFF)
939 counter->state = PERF_COUNTER_STATE_INACTIVE;
940 counter->hw_event.disabled = 0;
942 hw_perf_restore(perf_flags);
944 spin_unlock(&ctx->lock);
946 perf_counter_task_sched_in(curr, cpu);
948 curr_rq_unlock_irq_restore(&flags);
954 * Round-robin a context's counters:
956 static void rotate_ctx(struct perf_counter_context *ctx)
958 struct perf_counter *counter;
961 if (!ctx->nr_counters)
964 spin_lock(&ctx->lock);
966 * Rotate the first entry last (works just fine for group counters too):
968 perf_flags = hw_perf_save_disable();
969 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
970 list_move_tail(&counter->list_entry, &ctx->counter_list);
973 hw_perf_restore(perf_flags);
975 spin_unlock(&ctx->lock);
978 void perf_counter_task_tick(struct task_struct *curr, int cpu)
980 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
981 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
982 const int rotate_percpu = 0;
985 perf_counter_cpu_sched_out(cpuctx);
986 perf_counter_task_sched_out(curr, cpu);
989 rotate_ctx(&cpuctx->ctx);
993 perf_counter_cpu_sched_in(cpuctx, cpu);
994 perf_counter_task_sched_in(curr, cpu);
998 * Cross CPU call to read the hardware counter
1000 static void __read(void *info)
1002 struct perf_counter *counter = info;
1003 unsigned long flags;
1005 curr_rq_lock_irq_save(&flags);
1006 counter->hw_ops->read(counter);
1007 curr_rq_unlock_irq_restore(&flags);
1010 static u64 perf_counter_read(struct perf_counter *counter)
1013 * If counter is enabled and currently active on a CPU, update the
1014 * value in the counter structure:
1016 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1017 smp_call_function_single(counter->oncpu,
1018 __read, counter, 1);
1021 return atomic64_read(&counter->count);
1025 * Cross CPU call to switch performance data pointers
1027 static void __perf_switch_irq_data(void *info)
1029 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1030 struct perf_counter *counter = info;
1031 struct perf_counter_context *ctx = counter->ctx;
1032 struct perf_data *oldirqdata = counter->irqdata;
1035 * If this is a task context, we need to check whether it is
1036 * the current task context of this cpu. If not it has been
1037 * scheduled out before the smp call arrived.
1040 if (cpuctx->task_ctx != ctx)
1042 spin_lock(&ctx->lock);
1045 /* Change the pointer NMI safe */
1046 atomic_long_set((atomic_long_t *)&counter->irqdata,
1047 (unsigned long) counter->usrdata);
1048 counter->usrdata = oldirqdata;
1051 spin_unlock(&ctx->lock);
1054 static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
1056 struct perf_counter_context *ctx = counter->ctx;
1057 struct perf_data *oldirqdata = counter->irqdata;
1058 struct task_struct *task = ctx->task;
1061 smp_call_function_single(counter->cpu,
1062 __perf_switch_irq_data,
1064 return counter->usrdata;
1068 spin_lock_irq(&ctx->lock);
1069 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
1070 counter->irqdata = counter->usrdata;
1071 counter->usrdata = oldirqdata;
1072 spin_unlock_irq(&ctx->lock);
1075 spin_unlock_irq(&ctx->lock);
1076 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
1077 /* Might have failed, because task was scheduled out */
1078 if (counter->irqdata == oldirqdata)
1081 return counter->usrdata;
1084 static void put_context(struct perf_counter_context *ctx)
1087 put_task_struct(ctx->task);
1090 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1092 struct perf_cpu_context *cpuctx;
1093 struct perf_counter_context *ctx;
1094 struct task_struct *task;
1097 * If cpu is not a wildcard then this is a percpu counter:
1100 /* Must be root to operate on a CPU counter: */
1101 if (!capable(CAP_SYS_ADMIN))
1102 return ERR_PTR(-EACCES);
1104 if (cpu < 0 || cpu > num_possible_cpus())
1105 return ERR_PTR(-EINVAL);
1108 * We could be clever and allow to attach a counter to an
1109 * offline CPU and activate it when the CPU comes up, but
1112 if (!cpu_isset(cpu, cpu_online_map))
1113 return ERR_PTR(-ENODEV);
1115 cpuctx = &per_cpu(perf_cpu_context, cpu);
1125 task = find_task_by_vpid(pid);
1127 get_task_struct(task);
1131 return ERR_PTR(-ESRCH);
1133 ctx = &task->perf_counter_ctx;
1136 /* Reuse ptrace permission checks for now. */
1137 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1139 return ERR_PTR(-EACCES);
1145 static void free_counter_rcu(struct rcu_head *head)
1147 struct perf_counter *counter;
1149 counter = container_of(head, struct perf_counter, rcu_head);
1153 static void free_counter(struct perf_counter *counter)
1155 if (counter->destroy)
1156 counter->destroy(counter);
1158 call_rcu(&counter->rcu_head, free_counter_rcu);
1162 * Called when the last reference to the file is gone.
1164 static int perf_release(struct inode *inode, struct file *file)
1166 struct perf_counter *counter = file->private_data;
1167 struct perf_counter_context *ctx = counter->ctx;
1169 file->private_data = NULL;
1171 mutex_lock(&ctx->mutex);
1172 mutex_lock(&counter->mutex);
1174 perf_counter_remove_from_context(counter);
1176 mutex_unlock(&counter->mutex);
1177 mutex_unlock(&ctx->mutex);
1179 free_counter(counter);
1186 * Read the performance counter - simple non blocking version for now
1189 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1193 if (count != sizeof(cntval))
1197 * Return end-of-file for a read on a counter that is in
1198 * error state (i.e. because it was pinned but it couldn't be
1199 * scheduled on to the CPU at some point).
1201 if (counter->state == PERF_COUNTER_STATE_ERROR)
1204 mutex_lock(&counter->mutex);
1205 cntval = perf_counter_read(counter);
1206 mutex_unlock(&counter->mutex);
1208 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
1212 perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
1217 count = min(count, (size_t)usrdata->len);
1218 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
1221 /* Adjust the counters */
1222 usrdata->len -= count;
1224 usrdata->rd_idx = 0;
1226 usrdata->rd_idx += count;
1232 perf_read_irq_data(struct perf_counter *counter,
1237 struct perf_data *irqdata, *usrdata;
1238 DECLARE_WAITQUEUE(wait, current);
1241 irqdata = counter->irqdata;
1242 usrdata = counter->usrdata;
1244 if (usrdata->len + irqdata->len >= count)
1250 spin_lock_irq(&counter->waitq.lock);
1251 __add_wait_queue(&counter->waitq, &wait);
1253 set_current_state(TASK_INTERRUPTIBLE);
1254 if (usrdata->len + irqdata->len >= count)
1257 if (signal_pending(current))
1260 if (counter->state == PERF_COUNTER_STATE_ERROR)
1263 spin_unlock_irq(&counter->waitq.lock);
1265 spin_lock_irq(&counter->waitq.lock);
1267 __remove_wait_queue(&counter->waitq, &wait);
1268 __set_current_state(TASK_RUNNING);
1269 spin_unlock_irq(&counter->waitq.lock);
1271 if (usrdata->len + irqdata->len < count &&
1272 counter->state != PERF_COUNTER_STATE_ERROR)
1273 return -ERESTARTSYS;
1275 mutex_lock(&counter->mutex);
1277 /* Drain pending data first: */
1278 res = perf_copy_usrdata(usrdata, buf, count);
1279 if (res < 0 || res == count)
1282 /* Switch irq buffer: */
1283 usrdata = perf_switch_irq_data(counter);
1284 res2 = perf_copy_usrdata(usrdata, buf + res, count - res);
1292 mutex_unlock(&counter->mutex);
1298 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1300 struct perf_counter *counter = file->private_data;
1302 switch (counter->hw_event.record_type) {
1303 case PERF_RECORD_SIMPLE:
1304 return perf_read_hw(counter, buf, count);
1306 case PERF_RECORD_IRQ:
1307 case PERF_RECORD_GROUP:
1308 return perf_read_irq_data(counter, buf, count,
1309 file->f_flags & O_NONBLOCK);
1314 static unsigned int perf_poll(struct file *file, poll_table *wait)
1316 struct perf_counter *counter = file->private_data;
1317 unsigned int events = 0;
1318 unsigned long flags;
1320 poll_wait(file, &counter->waitq, wait);
1322 spin_lock_irqsave(&counter->waitq.lock, flags);
1323 if (counter->usrdata->len || counter->irqdata->len)
1325 spin_unlock_irqrestore(&counter->waitq.lock, flags);
1330 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1332 struct perf_counter *counter = file->private_data;
1336 case PERF_COUNTER_IOC_ENABLE:
1337 perf_counter_enable_family(counter);
1339 case PERF_COUNTER_IOC_DISABLE:
1340 perf_counter_disable_family(counter);
1348 static const struct file_operations perf_fops = {
1349 .release = perf_release,
1352 .unlocked_ioctl = perf_ioctl,
1353 .compat_ioctl = perf_ioctl,
1357 * Generic software counter infrastructure
1360 static void perf_swcounter_update(struct perf_counter *counter)
1362 struct hw_perf_counter *hwc = &counter->hw;
1367 prev = atomic64_read(&hwc->prev_count);
1368 now = atomic64_read(&hwc->count);
1369 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
1374 atomic64_add(delta, &counter->count);
1375 atomic64_sub(delta, &hwc->period_left);
1378 static void perf_swcounter_set_period(struct perf_counter *counter)
1380 struct hw_perf_counter *hwc = &counter->hw;
1381 s64 left = atomic64_read(&hwc->period_left);
1382 s64 period = hwc->irq_period;
1384 if (unlikely(left <= -period)) {
1386 atomic64_set(&hwc->period_left, left);
1389 if (unlikely(left <= 0)) {
1391 atomic64_add(period, &hwc->period_left);
1394 atomic64_set(&hwc->prev_count, -left);
1395 atomic64_set(&hwc->count, -left);
1398 static void perf_swcounter_store_irq(struct perf_counter *counter, u64 data)
1400 struct perf_data *irqdata = counter->irqdata;
1402 if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
1405 u64 *p = (u64 *) &irqdata->data[irqdata->len];
1408 irqdata->len += sizeof(u64);
1412 static void perf_swcounter_handle_group(struct perf_counter *sibling)
1414 struct perf_counter *counter, *group_leader = sibling->group_leader;
1416 list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
1417 counter->hw_ops->read(counter);
1418 perf_swcounter_store_irq(sibling, counter->hw_event.event_config);
1419 perf_swcounter_store_irq(sibling, atomic64_read(&counter->count));
1423 static void perf_swcounter_interrupt(struct perf_counter *counter,
1424 int nmi, struct pt_regs *regs)
1426 switch (counter->hw_event.record_type) {
1427 case PERF_RECORD_SIMPLE:
1430 case PERF_RECORD_IRQ:
1431 perf_swcounter_store_irq(counter, instruction_pointer(regs));
1434 case PERF_RECORD_GROUP:
1435 perf_swcounter_handle_group(counter);
1440 counter->wakeup_pending = 1;
1441 set_perf_counter_pending();
1443 wake_up(&counter->waitq);
1446 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
1448 struct perf_counter *counter;
1449 struct pt_regs *regs;
1451 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
1452 counter->hw_ops->read(counter);
1454 regs = get_irq_regs();
1456 * In case we exclude kernel IPs or are somehow not in interrupt
1457 * context, provide the next best thing, the user IP.
1459 if ((counter->hw_event.exclude_kernel || !regs) &&
1460 !counter->hw_event.exclude_user)
1461 regs = task_pt_regs(current);
1464 perf_swcounter_interrupt(counter, 0, regs);
1466 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
1468 return HRTIMER_RESTART;
1471 static void perf_swcounter_overflow(struct perf_counter *counter,
1472 int nmi, struct pt_regs *regs)
1474 perf_swcounter_update(counter);
1475 perf_swcounter_set_period(counter);
1476 perf_swcounter_interrupt(counter, nmi, regs);
1479 static int perf_swcounter_match(struct perf_counter *counter,
1480 enum perf_event_types type,
1481 u32 event, struct pt_regs *regs)
1483 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1486 if (counter->hw_event.raw_type)
1489 if (counter->hw_event.type != type)
1492 if (counter->hw_event.event_id != event)
1495 if (counter->hw_event.exclude_user && user_mode(regs))
1498 if (counter->hw_event.exclude_kernel && !user_mode(regs))
1504 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
1505 int nmi, struct pt_regs *regs)
1507 int neg = atomic64_add_negative(nr, &counter->hw.count);
1508 if (counter->hw.irq_period && !neg)
1509 perf_swcounter_overflow(counter, nmi, regs);
1512 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
1513 enum perf_event_types type, u32 event,
1514 u64 nr, int nmi, struct pt_regs *regs)
1516 struct perf_counter *counter;
1518 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
1522 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
1523 if (perf_swcounter_match(counter, type, event, regs))
1524 perf_swcounter_add(counter, nr, nmi, regs);
1529 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
1530 u64 nr, int nmi, struct pt_regs *regs)
1532 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
1534 perf_swcounter_ctx_event(&cpuctx->ctx, type, event, nr, nmi, regs);
1535 if (cpuctx->task_ctx) {
1536 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
1540 put_cpu_var(perf_cpu_context);
1543 void perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)
1545 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs);
1548 static void perf_swcounter_read(struct perf_counter *counter)
1550 perf_swcounter_update(counter);
1553 static int perf_swcounter_enable(struct perf_counter *counter)
1555 perf_swcounter_set_period(counter);
1559 static void perf_swcounter_disable(struct perf_counter *counter)
1561 perf_swcounter_update(counter);
1564 static const struct hw_perf_counter_ops perf_ops_generic = {
1565 .enable = perf_swcounter_enable,
1566 .disable = perf_swcounter_disable,
1567 .read = perf_swcounter_read,
1571 * Software counter: cpu wall time clock
1574 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
1576 int cpu = raw_smp_processor_id();
1580 now = cpu_clock(cpu);
1581 prev = atomic64_read(&counter->hw.prev_count);
1582 atomic64_set(&counter->hw.prev_count, now);
1583 atomic64_add(now - prev, &counter->count);
1586 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
1588 struct hw_perf_counter *hwc = &counter->hw;
1589 int cpu = raw_smp_processor_id();
1591 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
1592 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1593 hwc->hrtimer.function = perf_swcounter_hrtimer;
1594 if (hwc->irq_period) {
1595 __hrtimer_start_range_ns(&hwc->hrtimer,
1596 ns_to_ktime(hwc->irq_period), 0,
1597 HRTIMER_MODE_REL, 0);
1603 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
1605 hrtimer_cancel(&counter->hw.hrtimer);
1606 cpu_clock_perf_counter_update(counter);
1609 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
1611 cpu_clock_perf_counter_update(counter);
1614 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
1615 .enable = cpu_clock_perf_counter_enable,
1616 .disable = cpu_clock_perf_counter_disable,
1617 .read = cpu_clock_perf_counter_read,
1621 * Software counter: task time clock
1625 * Called from within the scheduler:
1627 static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
1629 struct task_struct *curr = counter->task;
1632 delta = __task_delta_exec(curr, update);
1634 return curr->se.sum_exec_runtime + delta;
1637 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
1642 prev = atomic64_read(&counter->hw.prev_count);
1644 atomic64_set(&counter->hw.prev_count, now);
1648 atomic64_add(delta, &counter->count);
1651 static int task_clock_perf_counter_enable(struct perf_counter *counter)
1653 struct hw_perf_counter *hwc = &counter->hw;
1655 atomic64_set(&hwc->prev_count, task_clock_perf_counter_val(counter, 0));
1656 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1657 hwc->hrtimer.function = perf_swcounter_hrtimer;
1658 if (hwc->irq_period) {
1659 __hrtimer_start_range_ns(&hwc->hrtimer,
1660 ns_to_ktime(hwc->irq_period), 0,
1661 HRTIMER_MODE_REL, 0);
1667 static void task_clock_perf_counter_disable(struct perf_counter *counter)
1669 hrtimer_cancel(&counter->hw.hrtimer);
1670 task_clock_perf_counter_update(counter,
1671 task_clock_perf_counter_val(counter, 0));
1674 static void task_clock_perf_counter_read(struct perf_counter *counter)
1676 task_clock_perf_counter_update(counter,
1677 task_clock_perf_counter_val(counter, 1));
1680 static const struct hw_perf_counter_ops perf_ops_task_clock = {
1681 .enable = task_clock_perf_counter_enable,
1682 .disable = task_clock_perf_counter_disable,
1683 .read = task_clock_perf_counter_read,
1687 * Software counter: cpu migrations
1690 static inline u64 get_cpu_migrations(struct perf_counter *counter)
1692 struct task_struct *curr = counter->ctx->task;
1695 return curr->se.nr_migrations;
1696 return cpu_nr_migrations(smp_processor_id());
1699 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1704 prev = atomic64_read(&counter->hw.prev_count);
1705 now = get_cpu_migrations(counter);
1707 atomic64_set(&counter->hw.prev_count, now);
1711 atomic64_add(delta, &counter->count);
1714 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1716 cpu_migrations_perf_counter_update(counter);
1719 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1721 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1722 atomic64_set(&counter->hw.prev_count,
1723 get_cpu_migrations(counter));
1727 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1729 cpu_migrations_perf_counter_update(counter);
1732 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1733 .enable = cpu_migrations_perf_counter_enable,
1734 .disable = cpu_migrations_perf_counter_disable,
1735 .read = cpu_migrations_perf_counter_read,
1738 #ifdef CONFIG_EVENT_PROFILE
1739 void perf_tpcounter_event(int event_id)
1741 struct pt_regs *regs = get_irq_regs();
1744 regs = task_pt_regs(current);
1746 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs);
1749 extern int ftrace_profile_enable(int);
1750 extern void ftrace_profile_disable(int);
1752 static void tp_perf_counter_destroy(struct perf_counter *counter)
1754 ftrace_profile_disable(counter->hw_event.event_id);
1757 static const struct hw_perf_counter_ops *
1758 tp_perf_counter_init(struct perf_counter *counter)
1760 int event_id = counter->hw_event.event_id;
1763 ret = ftrace_profile_enable(event_id);
1767 counter->destroy = tp_perf_counter_destroy;
1768 counter->hw.irq_period = counter->hw_event.irq_period;
1770 return &perf_ops_generic;
1773 static const struct hw_perf_counter_ops *
1774 tp_perf_counter_init(struct perf_counter *counter)
1780 static const struct hw_perf_counter_ops *
1781 sw_perf_counter_init(struct perf_counter *counter)
1783 struct perf_counter_hw_event *hw_event = &counter->hw_event;
1784 const struct hw_perf_counter_ops *hw_ops = NULL;
1785 struct hw_perf_counter *hwc = &counter->hw;
1788 * Software counters (currently) can't in general distinguish
1789 * between user, kernel and hypervisor events.
1790 * However, context switches and cpu migrations are considered
1791 * to be kernel events, and page faults are never hypervisor
1794 switch (counter->hw_event.event_id) {
1795 case PERF_COUNT_CPU_CLOCK:
1796 hw_ops = &perf_ops_cpu_clock;
1798 if (hw_event->irq_period && hw_event->irq_period < 10000)
1799 hw_event->irq_period = 10000;
1801 case PERF_COUNT_TASK_CLOCK:
1803 * If the user instantiates this as a per-cpu counter,
1804 * use the cpu_clock counter instead.
1806 if (counter->ctx->task)
1807 hw_ops = &perf_ops_task_clock;
1809 hw_ops = &perf_ops_cpu_clock;
1811 if (hw_event->irq_period && hw_event->irq_period < 10000)
1812 hw_event->irq_period = 10000;
1814 case PERF_COUNT_PAGE_FAULTS:
1815 case PERF_COUNT_PAGE_FAULTS_MIN:
1816 case PERF_COUNT_PAGE_FAULTS_MAJ:
1817 case PERF_COUNT_CONTEXT_SWITCHES:
1818 hw_ops = &perf_ops_generic;
1820 case PERF_COUNT_CPU_MIGRATIONS:
1821 if (!counter->hw_event.exclude_kernel)
1822 hw_ops = &perf_ops_cpu_migrations;
1827 hwc->irq_period = hw_event->irq_period;
1833 * Allocate and initialize a counter structure
1835 static struct perf_counter *
1836 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1838 struct perf_counter_context *ctx,
1839 struct perf_counter *group_leader,
1842 const struct hw_perf_counter_ops *hw_ops;
1843 struct perf_counter *counter;
1845 counter = kzalloc(sizeof(*counter), gfpflags);
1850 * Single counters are their own group leaders, with an
1851 * empty sibling list:
1854 group_leader = counter;
1856 mutex_init(&counter->mutex);
1857 INIT_LIST_HEAD(&counter->list_entry);
1858 INIT_LIST_HEAD(&counter->event_entry);
1859 INIT_LIST_HEAD(&counter->sibling_list);
1860 init_waitqueue_head(&counter->waitq);
1862 INIT_LIST_HEAD(&counter->child_list);
1864 counter->irqdata = &counter->data[0];
1865 counter->usrdata = &counter->data[1];
1867 counter->hw_event = *hw_event;
1868 counter->wakeup_pending = 0;
1869 counter->group_leader = group_leader;
1870 counter->hw_ops = NULL;
1873 counter->state = PERF_COUNTER_STATE_INACTIVE;
1874 if (hw_event->disabled)
1875 counter->state = PERF_COUNTER_STATE_OFF;
1879 if (hw_event->raw_type)
1880 hw_ops = hw_perf_counter_init(counter);
1881 else switch (hw_event->type) {
1882 case PERF_TYPE_HARDWARE:
1883 hw_ops = hw_perf_counter_init(counter);
1886 case PERF_TYPE_SOFTWARE:
1887 hw_ops = sw_perf_counter_init(counter);
1890 case PERF_TYPE_TRACEPOINT:
1891 hw_ops = tp_perf_counter_init(counter);
1899 counter->hw_ops = hw_ops;
1905 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
1907 * @hw_event_uptr: event type attributes for monitoring/sampling
1910 * @group_fd: group leader counter fd
1912 SYSCALL_DEFINE5(perf_counter_open,
1913 const struct perf_counter_hw_event __user *, hw_event_uptr,
1914 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
1916 struct perf_counter *counter, *group_leader;
1917 struct perf_counter_hw_event hw_event;
1918 struct perf_counter_context *ctx;
1919 struct file *counter_file = NULL;
1920 struct file *group_file = NULL;
1921 int fput_needed = 0;
1922 int fput_needed2 = 0;
1925 /* for future expandability... */
1929 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1933 * Get the target context (task or percpu):
1935 ctx = find_get_context(pid, cpu);
1937 return PTR_ERR(ctx);
1940 * Look up the group leader (we will attach this counter to it):
1942 group_leader = NULL;
1943 if (group_fd != -1) {
1945 group_file = fget_light(group_fd, &fput_needed);
1947 goto err_put_context;
1948 if (group_file->f_op != &perf_fops)
1949 goto err_put_context;
1951 group_leader = group_file->private_data;
1953 * Do not allow a recursive hierarchy (this new sibling
1954 * becoming part of another group-sibling):
1956 if (group_leader->group_leader != group_leader)
1957 goto err_put_context;
1959 * Do not allow to attach to a group in a different
1960 * task or CPU context:
1962 if (group_leader->ctx != ctx)
1963 goto err_put_context;
1965 * Only a group leader can be exclusive or pinned
1967 if (hw_event.exclusive || hw_event.pinned)
1968 goto err_put_context;
1972 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
1975 goto err_put_context;
1977 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
1979 goto err_free_put_context;
1981 counter_file = fget_light(ret, &fput_needed2);
1983 goto err_free_put_context;
1985 counter->filp = counter_file;
1986 mutex_lock(&ctx->mutex);
1987 perf_install_in_context(ctx, counter, cpu);
1988 mutex_unlock(&ctx->mutex);
1990 fput_light(counter_file, fput_needed2);
1993 fput_light(group_file, fput_needed);
1997 err_free_put_context:
2007 * Initialize the perf_counter context in a task_struct:
2010 __perf_counter_init_context(struct perf_counter_context *ctx,
2011 struct task_struct *task)
2013 memset(ctx, 0, sizeof(*ctx));
2014 spin_lock_init(&ctx->lock);
2015 mutex_init(&ctx->mutex);
2016 INIT_LIST_HEAD(&ctx->counter_list);
2017 INIT_LIST_HEAD(&ctx->event_list);
2022 * inherit a counter from parent task to child task:
2024 static struct perf_counter *
2025 inherit_counter(struct perf_counter *parent_counter,
2026 struct task_struct *parent,
2027 struct perf_counter_context *parent_ctx,
2028 struct task_struct *child,
2029 struct perf_counter *group_leader,
2030 struct perf_counter_context *child_ctx)
2032 struct perf_counter *child_counter;
2035 * Instead of creating recursive hierarchies of counters,
2036 * we link inherited counters back to the original parent,
2037 * which has a filp for sure, which we use as the reference
2040 if (parent_counter->parent)
2041 parent_counter = parent_counter->parent;
2043 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2044 parent_counter->cpu, child_ctx,
2045 group_leader, GFP_KERNEL);
2050 * Link it up in the child's context:
2052 child_counter->task = child;
2053 list_add_counter(child_counter, child_ctx);
2054 child_ctx->nr_counters++;
2056 child_counter->parent = parent_counter;
2058 * inherit into child's child as well:
2060 child_counter->hw_event.inherit = 1;
2063 * Get a reference to the parent filp - we will fput it
2064 * when the child counter exits. This is safe to do because
2065 * we are in the parent and we know that the filp still
2066 * exists and has a nonzero count:
2068 atomic_long_inc(&parent_counter->filp->f_count);
2071 * Link this into the parent counter's child list
2073 mutex_lock(&parent_counter->mutex);
2074 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2077 * Make the child state follow the state of the parent counter,
2078 * not its hw_event.disabled bit. We hold the parent's mutex,
2079 * so we won't race with perf_counter_{en,dis}able_family.
2081 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2082 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2084 child_counter->state = PERF_COUNTER_STATE_OFF;
2086 mutex_unlock(&parent_counter->mutex);
2088 return child_counter;
2091 static int inherit_group(struct perf_counter *parent_counter,
2092 struct task_struct *parent,
2093 struct perf_counter_context *parent_ctx,
2094 struct task_struct *child,
2095 struct perf_counter_context *child_ctx)
2097 struct perf_counter *leader;
2098 struct perf_counter *sub;
2100 leader = inherit_counter(parent_counter, parent, parent_ctx,
2101 child, NULL, child_ctx);
2104 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2105 if (!inherit_counter(sub, parent, parent_ctx,
2106 child, leader, child_ctx))
2112 static void sync_child_counter(struct perf_counter *child_counter,
2113 struct perf_counter *parent_counter)
2115 u64 parent_val, child_val;
2117 parent_val = atomic64_read(&parent_counter->count);
2118 child_val = atomic64_read(&child_counter->count);
2121 * Add back the child's count to the parent's count:
2123 atomic64_add(child_val, &parent_counter->count);
2126 * Remove this counter from the parent's list
2128 mutex_lock(&parent_counter->mutex);
2129 list_del_init(&child_counter->child_list);
2130 mutex_unlock(&parent_counter->mutex);
2133 * Release the parent counter, if this was the last
2136 fput(parent_counter->filp);
2140 __perf_counter_exit_task(struct task_struct *child,
2141 struct perf_counter *child_counter,
2142 struct perf_counter_context *child_ctx)
2144 struct perf_counter *parent_counter;
2145 struct perf_counter *sub, *tmp;
2148 * If we do not self-reap then we have to wait for the
2149 * child task to unschedule (it will happen for sure),
2150 * so that its counter is at its final count. (This
2151 * condition triggers rarely - child tasks usually get
2152 * off their CPU before the parent has a chance to
2153 * get this far into the reaping action)
2155 if (child != current) {
2156 wait_task_inactive(child, 0);
2157 list_del_init(&child_counter->list_entry);
2159 struct perf_cpu_context *cpuctx;
2160 unsigned long flags;
2164 * Disable and unlink this counter.
2166 * Be careful about zapping the list - IRQ/NMI context
2167 * could still be processing it:
2169 curr_rq_lock_irq_save(&flags);
2170 perf_flags = hw_perf_save_disable();
2172 cpuctx = &__get_cpu_var(perf_cpu_context);
2174 group_sched_out(child_counter, cpuctx, child_ctx);
2176 list_del_init(&child_counter->list_entry);
2178 child_ctx->nr_counters--;
2180 hw_perf_restore(perf_flags);
2181 curr_rq_unlock_irq_restore(&flags);
2184 parent_counter = child_counter->parent;
2186 * It can happen that parent exits first, and has counters
2187 * that are still around due to the child reference. These
2188 * counters need to be zapped - but otherwise linger.
2190 if (parent_counter) {
2191 sync_child_counter(child_counter, parent_counter);
2192 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
2195 sync_child_counter(sub, sub->parent);
2199 free_counter(child_counter);
2204 * When a child task exits, feed back counter values to parent counters.
2206 * Note: we may be running in child context, but the PID is not hashed
2207 * anymore so new counters will not be added.
2209 void perf_counter_exit_task(struct task_struct *child)
2211 struct perf_counter *child_counter, *tmp;
2212 struct perf_counter_context *child_ctx;
2214 child_ctx = &child->perf_counter_ctx;
2216 if (likely(!child_ctx->nr_counters))
2219 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
2221 __perf_counter_exit_task(child, child_counter, child_ctx);
2225 * Initialize the perf_counter context in task_struct
2227 void perf_counter_init_task(struct task_struct *child)
2229 struct perf_counter_context *child_ctx, *parent_ctx;
2230 struct perf_counter *counter;
2231 struct task_struct *parent = current;
2233 child_ctx = &child->perf_counter_ctx;
2234 parent_ctx = &parent->perf_counter_ctx;
2236 __perf_counter_init_context(child_ctx, child);
2239 * This is executed from the parent task context, so inherit
2240 * counters that have been marked for cloning:
2243 if (likely(!parent_ctx->nr_counters))
2247 * Lock the parent list. No need to lock the child - not PID
2248 * hashed yet and not running, so nobody can access it.
2250 mutex_lock(&parent_ctx->mutex);
2253 * We dont have to disable NMIs - we are only looking at
2254 * the list, not manipulating it:
2256 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2257 if (!counter->hw_event.inherit)
2260 if (inherit_group(counter, parent,
2261 parent_ctx, child, child_ctx))
2265 mutex_unlock(&parent_ctx->mutex);
2268 static void __cpuinit perf_counter_init_cpu(int cpu)
2270 struct perf_cpu_context *cpuctx;
2272 cpuctx = &per_cpu(perf_cpu_context, cpu);
2273 __perf_counter_init_context(&cpuctx->ctx, NULL);
2275 mutex_lock(&perf_resource_mutex);
2276 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2277 mutex_unlock(&perf_resource_mutex);
2279 hw_perf_counter_setup(cpu);
2282 #ifdef CONFIG_HOTPLUG_CPU
2283 static void __perf_counter_exit_cpu(void *info)
2285 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2286 struct perf_counter_context *ctx = &cpuctx->ctx;
2287 struct perf_counter *counter, *tmp;
2289 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2290 __perf_counter_remove_from_context(counter);
2292 static void perf_counter_exit_cpu(int cpu)
2294 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2295 struct perf_counter_context *ctx = &cpuctx->ctx;
2297 mutex_lock(&ctx->mutex);
2298 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2299 mutex_unlock(&ctx->mutex);
2302 static inline void perf_counter_exit_cpu(int cpu) { }
2305 static int __cpuinit
2306 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2308 unsigned int cpu = (long)hcpu;
2312 case CPU_UP_PREPARE:
2313 case CPU_UP_PREPARE_FROZEN:
2314 perf_counter_init_cpu(cpu);
2317 case CPU_DOWN_PREPARE:
2318 case CPU_DOWN_PREPARE_FROZEN:
2319 perf_counter_exit_cpu(cpu);
2329 static struct notifier_block __cpuinitdata perf_cpu_nb = {
2330 .notifier_call = perf_cpu_notify,
2333 static int __init perf_counter_init(void)
2335 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2336 (void *)(long)smp_processor_id());
2337 register_cpu_notifier(&perf_cpu_nb);
2341 early_initcall(perf_counter_init);
2343 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2345 return sprintf(buf, "%d\n", perf_reserved_percpu);
2349 perf_set_reserve_percpu(struct sysdev_class *class,
2353 struct perf_cpu_context *cpuctx;
2357 err = strict_strtoul(buf, 10, &val);
2360 if (val > perf_max_counters)
2363 mutex_lock(&perf_resource_mutex);
2364 perf_reserved_percpu = val;
2365 for_each_online_cpu(cpu) {
2366 cpuctx = &per_cpu(perf_cpu_context, cpu);
2367 spin_lock_irq(&cpuctx->ctx.lock);
2368 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2369 perf_max_counters - perf_reserved_percpu);
2370 cpuctx->max_pertask = mpt;
2371 spin_unlock_irq(&cpuctx->ctx.lock);
2373 mutex_unlock(&perf_resource_mutex);
2378 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2380 return sprintf(buf, "%d\n", perf_overcommit);
2384 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2389 err = strict_strtoul(buf, 10, &val);
2395 mutex_lock(&perf_resource_mutex);
2396 perf_overcommit = val;
2397 mutex_unlock(&perf_resource_mutex);
2402 static SYSDEV_CLASS_ATTR(
2405 perf_show_reserve_percpu,
2406 perf_set_reserve_percpu
2409 static SYSDEV_CLASS_ATTR(
2412 perf_show_overcommit,
2416 static struct attribute *perfclass_attrs[] = {
2417 &attr_reserve_percpu.attr,
2418 &attr_overcommit.attr,
2422 static struct attribute_group perfclass_attr_group = {
2423 .attrs = perfclass_attrs,
2424 .name = "perf_counters",
2427 static int __init perf_counter_sysfs_init(void)
2429 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2430 &perfclass_attr_group);
2432 device_initcall(perf_counter_sysfs_init);