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;
714 if (likely(!cpuctx->task_ctx))
717 __perf_counter_sched_out(ctx, cpuctx);
719 cpuctx->task_ctx = NULL;
722 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
724 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
728 group_sched_in(struct perf_counter *group_counter,
729 struct perf_cpu_context *cpuctx,
730 struct perf_counter_context *ctx,
733 struct perf_counter *counter, *partial_group;
736 if (group_counter->state == PERF_COUNTER_STATE_OFF)
739 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
741 return ret < 0 ? ret : 0;
743 group_counter->prev_state = group_counter->state;
744 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
748 * Schedule in siblings as one group (if any):
750 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
751 counter->prev_state = counter->state;
752 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
753 partial_group = counter;
762 * Groups can be scheduled in as one unit only, so undo any
763 * partial group before returning:
765 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
766 if (counter == partial_group)
768 counter_sched_out(counter, cpuctx, ctx);
770 counter_sched_out(group_counter, cpuctx, ctx);
776 __perf_counter_sched_in(struct perf_counter_context *ctx,
777 struct perf_cpu_context *cpuctx, int cpu)
779 struct perf_counter *counter;
783 spin_lock(&ctx->lock);
785 if (likely(!ctx->nr_counters))
788 flags = hw_perf_save_disable();
791 * First go through the list and put on any pinned groups
792 * in order to give them the best chance of going on.
794 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
795 if (counter->state <= PERF_COUNTER_STATE_OFF ||
796 !counter->hw_event.pinned)
798 if (counter->cpu != -1 && counter->cpu != cpu)
801 if (group_can_go_on(counter, cpuctx, 1))
802 group_sched_in(counter, cpuctx, ctx, cpu);
805 * If this pinned group hasn't been scheduled,
806 * put it in error state.
808 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
809 counter->state = PERF_COUNTER_STATE_ERROR;
812 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
814 * Ignore counters in OFF or ERROR state, and
815 * ignore pinned counters since we did them already.
817 if (counter->state <= PERF_COUNTER_STATE_OFF ||
818 counter->hw_event.pinned)
822 * Listen to the 'cpu' scheduling filter constraint
825 if (counter->cpu != -1 && counter->cpu != cpu)
828 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
829 if (group_sched_in(counter, cpuctx, ctx, cpu))
833 hw_perf_restore(flags);
835 spin_unlock(&ctx->lock);
839 * Called from scheduler to add the counters of the current task
840 * with interrupts disabled.
842 * We restore the counter value and then enable it.
844 * This does not protect us against NMI, but enable()
845 * sets the enabled bit in the control field of counter _before_
846 * accessing the counter control register. If a NMI hits, then it will
847 * keep the counter running.
849 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
851 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
852 struct perf_counter_context *ctx = &task->perf_counter_ctx;
854 __perf_counter_sched_in(ctx, cpuctx, cpu);
855 cpuctx->task_ctx = ctx;
858 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
860 struct perf_counter_context *ctx = &cpuctx->ctx;
862 __perf_counter_sched_in(ctx, cpuctx, cpu);
865 int perf_counter_task_disable(void)
867 struct task_struct *curr = current;
868 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
869 struct perf_counter *counter;
874 if (likely(!ctx->nr_counters))
877 curr_rq_lock_irq_save(&flags);
878 cpu = smp_processor_id();
880 /* force the update of the task clock: */
881 __task_delta_exec(curr, 1);
883 perf_counter_task_sched_out(curr, cpu);
885 spin_lock(&ctx->lock);
888 * Disable all the counters:
890 perf_flags = hw_perf_save_disable();
892 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
893 if (counter->state != PERF_COUNTER_STATE_ERROR)
894 counter->state = PERF_COUNTER_STATE_OFF;
897 hw_perf_restore(perf_flags);
899 spin_unlock(&ctx->lock);
901 curr_rq_unlock_irq_restore(&flags);
906 int perf_counter_task_enable(void)
908 struct task_struct *curr = current;
909 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
910 struct perf_counter *counter;
915 if (likely(!ctx->nr_counters))
918 curr_rq_lock_irq_save(&flags);
919 cpu = smp_processor_id();
921 /* force the update of the task clock: */
922 __task_delta_exec(curr, 1);
924 perf_counter_task_sched_out(curr, cpu);
926 spin_lock(&ctx->lock);
929 * Disable all the counters:
931 perf_flags = hw_perf_save_disable();
933 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
934 if (counter->state > PERF_COUNTER_STATE_OFF)
936 counter->state = PERF_COUNTER_STATE_INACTIVE;
937 counter->hw_event.disabled = 0;
939 hw_perf_restore(perf_flags);
941 spin_unlock(&ctx->lock);
943 perf_counter_task_sched_in(curr, cpu);
945 curr_rq_unlock_irq_restore(&flags);
951 * Round-robin a context's counters:
953 static void rotate_ctx(struct perf_counter_context *ctx)
955 struct perf_counter *counter;
958 if (!ctx->nr_counters)
961 spin_lock(&ctx->lock);
963 * Rotate the first entry last (works just fine for group counters too):
965 perf_flags = hw_perf_save_disable();
966 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
967 list_move_tail(&counter->list_entry, &ctx->counter_list);
970 hw_perf_restore(perf_flags);
972 spin_unlock(&ctx->lock);
975 void perf_counter_task_tick(struct task_struct *curr, int cpu)
977 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
978 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
979 const int rotate_percpu = 0;
982 perf_counter_cpu_sched_out(cpuctx);
983 perf_counter_task_sched_out(curr, cpu);
986 rotate_ctx(&cpuctx->ctx);
990 perf_counter_cpu_sched_in(cpuctx, cpu);
991 perf_counter_task_sched_in(curr, cpu);
995 * Cross CPU call to read the hardware counter
997 static void __read(void *info)
999 struct perf_counter *counter = info;
1000 unsigned long flags;
1002 curr_rq_lock_irq_save(&flags);
1003 counter->hw_ops->read(counter);
1004 curr_rq_unlock_irq_restore(&flags);
1007 static u64 perf_counter_read(struct perf_counter *counter)
1010 * If counter is enabled and currently active on a CPU, update the
1011 * value in the counter structure:
1013 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1014 smp_call_function_single(counter->oncpu,
1015 __read, counter, 1);
1018 return atomic64_read(&counter->count);
1022 * Cross CPU call to switch performance data pointers
1024 static void __perf_switch_irq_data(void *info)
1026 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1027 struct perf_counter *counter = info;
1028 struct perf_counter_context *ctx = counter->ctx;
1029 struct perf_data *oldirqdata = counter->irqdata;
1032 * If this is a task context, we need to check whether it is
1033 * the current task context of this cpu. If not it has been
1034 * scheduled out before the smp call arrived.
1037 if (cpuctx->task_ctx != ctx)
1039 spin_lock(&ctx->lock);
1042 /* Change the pointer NMI safe */
1043 atomic_long_set((atomic_long_t *)&counter->irqdata,
1044 (unsigned long) counter->usrdata);
1045 counter->usrdata = oldirqdata;
1048 spin_unlock(&ctx->lock);
1051 static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
1053 struct perf_counter_context *ctx = counter->ctx;
1054 struct perf_data *oldirqdata = counter->irqdata;
1055 struct task_struct *task = ctx->task;
1058 smp_call_function_single(counter->cpu,
1059 __perf_switch_irq_data,
1061 return counter->usrdata;
1065 spin_lock_irq(&ctx->lock);
1066 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
1067 counter->irqdata = counter->usrdata;
1068 counter->usrdata = oldirqdata;
1069 spin_unlock_irq(&ctx->lock);
1072 spin_unlock_irq(&ctx->lock);
1073 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
1074 /* Might have failed, because task was scheduled out */
1075 if (counter->irqdata == oldirqdata)
1078 return counter->usrdata;
1081 static void put_context(struct perf_counter_context *ctx)
1084 put_task_struct(ctx->task);
1087 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1089 struct perf_cpu_context *cpuctx;
1090 struct perf_counter_context *ctx;
1091 struct task_struct *task;
1094 * If cpu is not a wildcard then this is a percpu counter:
1097 /* Must be root to operate on a CPU counter: */
1098 if (!capable(CAP_SYS_ADMIN))
1099 return ERR_PTR(-EACCES);
1101 if (cpu < 0 || cpu > num_possible_cpus())
1102 return ERR_PTR(-EINVAL);
1105 * We could be clever and allow to attach a counter to an
1106 * offline CPU and activate it when the CPU comes up, but
1109 if (!cpu_isset(cpu, cpu_online_map))
1110 return ERR_PTR(-ENODEV);
1112 cpuctx = &per_cpu(perf_cpu_context, cpu);
1122 task = find_task_by_vpid(pid);
1124 get_task_struct(task);
1128 return ERR_PTR(-ESRCH);
1130 ctx = &task->perf_counter_ctx;
1133 /* Reuse ptrace permission checks for now. */
1134 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1136 return ERR_PTR(-EACCES);
1142 static void free_counter_rcu(struct rcu_head *head)
1144 struct perf_counter *counter;
1146 counter = container_of(head, struct perf_counter, rcu_head);
1151 * Called when the last reference to the file is gone.
1153 static int perf_release(struct inode *inode, struct file *file)
1155 struct perf_counter *counter = file->private_data;
1156 struct perf_counter_context *ctx = counter->ctx;
1158 file->private_data = NULL;
1160 mutex_lock(&ctx->mutex);
1161 mutex_lock(&counter->mutex);
1163 perf_counter_remove_from_context(counter);
1165 mutex_unlock(&counter->mutex);
1166 mutex_unlock(&ctx->mutex);
1168 call_rcu(&counter->rcu_head, free_counter_rcu);
1175 * Read the performance counter - simple non blocking version for now
1178 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1182 if (count != sizeof(cntval))
1186 * Return end-of-file for a read on a counter that is in
1187 * error state (i.e. because it was pinned but it couldn't be
1188 * scheduled on to the CPU at some point).
1190 if (counter->state == PERF_COUNTER_STATE_ERROR)
1193 mutex_lock(&counter->mutex);
1194 cntval = perf_counter_read(counter);
1195 mutex_unlock(&counter->mutex);
1197 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
1201 perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
1206 count = min(count, (size_t)usrdata->len);
1207 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
1210 /* Adjust the counters */
1211 usrdata->len -= count;
1213 usrdata->rd_idx = 0;
1215 usrdata->rd_idx += count;
1221 perf_read_irq_data(struct perf_counter *counter,
1226 struct perf_data *irqdata, *usrdata;
1227 DECLARE_WAITQUEUE(wait, current);
1230 irqdata = counter->irqdata;
1231 usrdata = counter->usrdata;
1233 if (usrdata->len + irqdata->len >= count)
1239 spin_lock_irq(&counter->waitq.lock);
1240 __add_wait_queue(&counter->waitq, &wait);
1242 set_current_state(TASK_INTERRUPTIBLE);
1243 if (usrdata->len + irqdata->len >= count)
1246 if (signal_pending(current))
1249 if (counter->state == PERF_COUNTER_STATE_ERROR)
1252 spin_unlock_irq(&counter->waitq.lock);
1254 spin_lock_irq(&counter->waitq.lock);
1256 __remove_wait_queue(&counter->waitq, &wait);
1257 __set_current_state(TASK_RUNNING);
1258 spin_unlock_irq(&counter->waitq.lock);
1260 if (usrdata->len + irqdata->len < count &&
1261 counter->state != PERF_COUNTER_STATE_ERROR)
1262 return -ERESTARTSYS;
1264 mutex_lock(&counter->mutex);
1266 /* Drain pending data first: */
1267 res = perf_copy_usrdata(usrdata, buf, count);
1268 if (res < 0 || res == count)
1271 /* Switch irq buffer: */
1272 usrdata = perf_switch_irq_data(counter);
1273 res2 = perf_copy_usrdata(usrdata, buf + res, count - res);
1281 mutex_unlock(&counter->mutex);
1287 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1289 struct perf_counter *counter = file->private_data;
1291 switch (counter->hw_event.record_type) {
1292 case PERF_RECORD_SIMPLE:
1293 return perf_read_hw(counter, buf, count);
1295 case PERF_RECORD_IRQ:
1296 case PERF_RECORD_GROUP:
1297 return perf_read_irq_data(counter, buf, count,
1298 file->f_flags & O_NONBLOCK);
1303 static unsigned int perf_poll(struct file *file, poll_table *wait)
1305 struct perf_counter *counter = file->private_data;
1306 unsigned int events = 0;
1307 unsigned long flags;
1309 poll_wait(file, &counter->waitq, wait);
1311 spin_lock_irqsave(&counter->waitq.lock, flags);
1312 if (counter->usrdata->len || counter->irqdata->len)
1314 spin_unlock_irqrestore(&counter->waitq.lock, flags);
1319 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1321 struct perf_counter *counter = file->private_data;
1325 case PERF_COUNTER_IOC_ENABLE:
1326 perf_counter_enable_family(counter);
1328 case PERF_COUNTER_IOC_DISABLE:
1329 perf_counter_disable_family(counter);
1337 static const struct file_operations perf_fops = {
1338 .release = perf_release,
1341 .unlocked_ioctl = perf_ioctl,
1342 .compat_ioctl = perf_ioctl,
1346 * Generic software counter infrastructure
1349 static void perf_swcounter_update(struct perf_counter *counter)
1351 struct hw_perf_counter *hwc = &counter->hw;
1356 prev = atomic64_read(&hwc->prev_count);
1357 now = atomic64_read(&hwc->count);
1358 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
1363 atomic64_add(delta, &counter->count);
1364 atomic64_sub(delta, &hwc->period_left);
1367 static void perf_swcounter_set_period(struct perf_counter *counter)
1369 struct hw_perf_counter *hwc = &counter->hw;
1370 s64 left = atomic64_read(&hwc->period_left);
1371 s64 period = hwc->irq_period;
1373 if (unlikely(left <= -period)) {
1375 atomic64_set(&hwc->period_left, left);
1378 if (unlikely(left <= 0)) {
1380 atomic64_add(period, &hwc->period_left);
1383 atomic64_set(&hwc->prev_count, -left);
1384 atomic64_set(&hwc->count, -left);
1387 static void perf_swcounter_save_and_restart(struct perf_counter *counter)
1389 perf_swcounter_update(counter);
1390 perf_swcounter_set_period(counter);
1393 static void perf_swcounter_store_irq(struct perf_counter *counter, u64 data)
1395 struct perf_data *irqdata = counter->irqdata;
1397 if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
1400 u64 *p = (u64 *) &irqdata->data[irqdata->len];
1403 irqdata->len += sizeof(u64);
1407 static void perf_swcounter_handle_group(struct perf_counter *sibling)
1409 struct perf_counter *counter, *group_leader = sibling->group_leader;
1411 list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
1412 counter->hw_ops->read(counter);
1413 perf_swcounter_store_irq(sibling, counter->hw_event.type);
1414 perf_swcounter_store_irq(sibling, atomic64_read(&counter->count));
1418 static void perf_swcounter_interrupt(struct perf_counter *counter,
1419 int nmi, struct pt_regs *regs)
1421 switch (counter->hw_event.record_type) {
1422 case PERF_RECORD_SIMPLE:
1425 case PERF_RECORD_IRQ:
1426 perf_swcounter_store_irq(counter, instruction_pointer(regs));
1429 case PERF_RECORD_GROUP:
1430 perf_swcounter_handle_group(counter);
1435 counter->wakeup_pending = 1;
1436 set_perf_counter_pending();
1438 wake_up(&counter->waitq);
1441 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
1443 struct perf_counter *counter;
1444 struct pt_regs *regs;
1446 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
1447 counter->hw_ops->read(counter);
1449 regs = get_irq_regs();
1451 * In case we exclude kernel IPs or are somehow not in interrupt
1452 * context, provide the next best thing, the user IP.
1454 if ((counter->hw_event.exclude_kernel || !regs) &&
1455 !counter->hw_event.exclude_user)
1456 regs = task_pt_regs(current);
1459 perf_swcounter_interrupt(counter, 0, regs);
1461 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
1463 return HRTIMER_RESTART;
1466 static void perf_swcounter_overflow(struct perf_counter *counter,
1467 int nmi, struct pt_regs *regs)
1469 perf_swcounter_save_and_restart(counter);
1470 perf_swcounter_interrupt(counter, nmi, regs);
1473 static int perf_swcounter_match(struct perf_counter *counter,
1474 enum hw_event_types event,
1475 struct pt_regs *regs)
1477 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1480 if (counter->hw_event.raw)
1483 if (counter->hw_event.type != event)
1486 if (counter->hw_event.exclude_user && user_mode(regs))
1489 if (counter->hw_event.exclude_kernel && !user_mode(regs))
1495 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
1496 int nmi, struct pt_regs *regs)
1498 int neg = atomic64_add_negative(nr, &counter->hw.count);
1499 if (counter->hw.irq_period && !neg)
1500 perf_swcounter_overflow(counter, nmi, regs);
1503 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
1504 enum hw_event_types event, u64 nr,
1505 int nmi, struct pt_regs *regs)
1507 struct perf_counter *counter;
1509 if (list_empty(&ctx->event_list))
1513 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
1514 if (perf_swcounter_match(counter, event, regs))
1515 perf_swcounter_add(counter, nr, nmi, regs);
1520 void perf_swcounter_event(enum hw_event_types event, u64 nr,
1521 int nmi, struct pt_regs *regs)
1523 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
1525 perf_swcounter_ctx_event(&cpuctx->ctx, event, nr, nmi, regs);
1526 if (cpuctx->task_ctx)
1527 perf_swcounter_ctx_event(cpuctx->task_ctx, event, nr, nmi, regs);
1529 put_cpu_var(perf_cpu_context);
1532 static void perf_swcounter_read(struct perf_counter *counter)
1534 perf_swcounter_update(counter);
1537 static int perf_swcounter_enable(struct perf_counter *counter)
1539 perf_swcounter_set_period(counter);
1543 static void perf_swcounter_disable(struct perf_counter *counter)
1545 perf_swcounter_update(counter);
1548 static const struct hw_perf_counter_ops perf_ops_generic = {
1549 .enable = perf_swcounter_enable,
1550 .disable = perf_swcounter_disable,
1551 .read = perf_swcounter_read,
1555 * Software counter: cpu wall time clock
1558 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
1560 int cpu = raw_smp_processor_id();
1564 now = cpu_clock(cpu);
1565 prev = atomic64_read(&counter->hw.prev_count);
1566 atomic64_set(&counter->hw.prev_count, now);
1567 atomic64_add(now - prev, &counter->count);
1570 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
1572 struct hw_perf_counter *hwc = &counter->hw;
1573 int cpu = raw_smp_processor_id();
1575 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
1576 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1577 hwc->hrtimer.function = perf_swcounter_hrtimer;
1578 if (hwc->irq_period) {
1579 __hrtimer_start_range_ns(&hwc->hrtimer,
1580 ns_to_ktime(hwc->irq_period), 0,
1581 HRTIMER_MODE_REL, 0);
1587 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
1589 hrtimer_cancel(&counter->hw.hrtimer);
1590 cpu_clock_perf_counter_update(counter);
1593 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
1595 cpu_clock_perf_counter_update(counter);
1598 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
1599 .enable = cpu_clock_perf_counter_enable,
1600 .disable = cpu_clock_perf_counter_disable,
1601 .read = cpu_clock_perf_counter_read,
1605 * Software counter: task time clock
1609 * Called from within the scheduler:
1611 static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
1613 struct task_struct *curr = counter->task;
1616 delta = __task_delta_exec(curr, update);
1618 return curr->se.sum_exec_runtime + delta;
1621 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
1626 prev = atomic64_read(&counter->hw.prev_count);
1628 atomic64_set(&counter->hw.prev_count, now);
1632 atomic64_add(delta, &counter->count);
1635 static int task_clock_perf_counter_enable(struct perf_counter *counter)
1637 struct hw_perf_counter *hwc = &counter->hw;
1639 atomic64_set(&hwc->prev_count, task_clock_perf_counter_val(counter, 0));
1640 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1641 hwc->hrtimer.function = perf_swcounter_hrtimer;
1642 if (hwc->irq_period) {
1643 __hrtimer_start_range_ns(&hwc->hrtimer,
1644 ns_to_ktime(hwc->irq_period), 0,
1645 HRTIMER_MODE_REL, 0);
1651 static void task_clock_perf_counter_disable(struct perf_counter *counter)
1653 hrtimer_cancel(&counter->hw.hrtimer);
1654 task_clock_perf_counter_update(counter,
1655 task_clock_perf_counter_val(counter, 0));
1658 static void task_clock_perf_counter_read(struct perf_counter *counter)
1660 task_clock_perf_counter_update(counter,
1661 task_clock_perf_counter_val(counter, 1));
1664 static const struct hw_perf_counter_ops perf_ops_task_clock = {
1665 .enable = task_clock_perf_counter_enable,
1666 .disable = task_clock_perf_counter_disable,
1667 .read = task_clock_perf_counter_read,
1671 * Software counter: context switches
1674 static u64 get_context_switches(struct perf_counter *counter)
1676 struct task_struct *curr = counter->ctx->task;
1679 return curr->nvcsw + curr->nivcsw;
1680 return cpu_nr_switches(smp_processor_id());
1683 static void context_switches_perf_counter_update(struct perf_counter *counter)
1688 prev = atomic64_read(&counter->hw.prev_count);
1689 now = get_context_switches(counter);
1691 atomic64_set(&counter->hw.prev_count, now);
1695 atomic64_add(delta, &counter->count);
1698 static void context_switches_perf_counter_read(struct perf_counter *counter)
1700 context_switches_perf_counter_update(counter);
1703 static int context_switches_perf_counter_enable(struct perf_counter *counter)
1705 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1706 atomic64_set(&counter->hw.prev_count,
1707 get_context_switches(counter));
1711 static void context_switches_perf_counter_disable(struct perf_counter *counter)
1713 context_switches_perf_counter_update(counter);
1716 static const struct hw_perf_counter_ops perf_ops_context_switches = {
1717 .enable = context_switches_perf_counter_enable,
1718 .disable = context_switches_perf_counter_disable,
1719 .read = context_switches_perf_counter_read,
1723 * Software counter: cpu migrations
1726 static inline u64 get_cpu_migrations(struct perf_counter *counter)
1728 struct task_struct *curr = counter->ctx->task;
1731 return curr->se.nr_migrations;
1732 return cpu_nr_migrations(smp_processor_id());
1735 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1740 prev = atomic64_read(&counter->hw.prev_count);
1741 now = get_cpu_migrations(counter);
1743 atomic64_set(&counter->hw.prev_count, now);
1747 atomic64_add(delta, &counter->count);
1750 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1752 cpu_migrations_perf_counter_update(counter);
1755 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1757 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1758 atomic64_set(&counter->hw.prev_count,
1759 get_cpu_migrations(counter));
1763 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1765 cpu_migrations_perf_counter_update(counter);
1768 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1769 .enable = cpu_migrations_perf_counter_enable,
1770 .disable = cpu_migrations_perf_counter_disable,
1771 .read = cpu_migrations_perf_counter_read,
1774 static const struct hw_perf_counter_ops *
1775 sw_perf_counter_init(struct perf_counter *counter)
1777 struct perf_counter_hw_event *hw_event = &counter->hw_event;
1778 const struct hw_perf_counter_ops *hw_ops = NULL;
1779 struct hw_perf_counter *hwc = &counter->hw;
1782 * Software counters (currently) can't in general distinguish
1783 * between user, kernel and hypervisor events.
1784 * However, context switches and cpu migrations are considered
1785 * to be kernel events, and page faults are never hypervisor
1788 switch (counter->hw_event.type) {
1789 case PERF_COUNT_CPU_CLOCK:
1790 hw_ops = &perf_ops_cpu_clock;
1792 if (hw_event->irq_period && hw_event->irq_period < 10000)
1793 hw_event->irq_period = 10000;
1795 case PERF_COUNT_TASK_CLOCK:
1797 * If the user instantiates this as a per-cpu counter,
1798 * use the cpu_clock counter instead.
1800 if (counter->ctx->task)
1801 hw_ops = &perf_ops_task_clock;
1803 hw_ops = &perf_ops_cpu_clock;
1805 if (hw_event->irq_period && hw_event->irq_period < 10000)
1806 hw_event->irq_period = 10000;
1808 case PERF_COUNT_PAGE_FAULTS:
1809 case PERF_COUNT_PAGE_FAULTS_MIN:
1810 case PERF_COUNT_PAGE_FAULTS_MAJ:
1811 hw_ops = &perf_ops_generic;
1813 case PERF_COUNT_CONTEXT_SWITCHES:
1814 if (!counter->hw_event.exclude_kernel)
1815 hw_ops = &perf_ops_context_switches;
1817 case PERF_COUNT_CPU_MIGRATIONS:
1818 if (!counter->hw_event.exclude_kernel)
1819 hw_ops = &perf_ops_cpu_migrations;
1826 hwc->irq_period = hw_event->irq_period;
1832 * Allocate and initialize a counter structure
1834 static struct perf_counter *
1835 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1837 struct perf_counter_context *ctx,
1838 struct perf_counter *group_leader,
1841 const struct hw_perf_counter_ops *hw_ops;
1842 struct perf_counter *counter;
1844 counter = kzalloc(sizeof(*counter), gfpflags);
1849 * Single counters are their own group leaders, with an
1850 * empty sibling list:
1853 group_leader = counter;
1855 mutex_init(&counter->mutex);
1856 INIT_LIST_HEAD(&counter->list_entry);
1857 INIT_LIST_HEAD(&counter->event_entry);
1858 INIT_LIST_HEAD(&counter->sibling_list);
1859 init_waitqueue_head(&counter->waitq);
1861 INIT_LIST_HEAD(&counter->child_list);
1863 counter->irqdata = &counter->data[0];
1864 counter->usrdata = &counter->data[1];
1866 counter->hw_event = *hw_event;
1867 counter->wakeup_pending = 0;
1868 counter->group_leader = group_leader;
1869 counter->hw_ops = NULL;
1872 counter->state = PERF_COUNTER_STATE_INACTIVE;
1873 if (hw_event->disabled)
1874 counter->state = PERF_COUNTER_STATE_OFF;
1877 if (!hw_event->raw && hw_event->type < 0)
1878 hw_ops = sw_perf_counter_init(counter);
1880 hw_ops = hw_perf_counter_init(counter);
1886 counter->hw_ops = hw_ops;
1892 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
1894 * @hw_event_uptr: event type attributes for monitoring/sampling
1897 * @group_fd: group leader counter fd
1899 SYSCALL_DEFINE5(perf_counter_open,
1900 const struct perf_counter_hw_event __user *, hw_event_uptr,
1901 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
1903 struct perf_counter *counter, *group_leader;
1904 struct perf_counter_hw_event hw_event;
1905 struct perf_counter_context *ctx;
1906 struct file *counter_file = NULL;
1907 struct file *group_file = NULL;
1908 int fput_needed = 0;
1909 int fput_needed2 = 0;
1912 /* for future expandability... */
1916 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1920 * Get the target context (task or percpu):
1922 ctx = find_get_context(pid, cpu);
1924 return PTR_ERR(ctx);
1927 * Look up the group leader (we will attach this counter to it):
1929 group_leader = NULL;
1930 if (group_fd != -1) {
1932 group_file = fget_light(group_fd, &fput_needed);
1934 goto err_put_context;
1935 if (group_file->f_op != &perf_fops)
1936 goto err_put_context;
1938 group_leader = group_file->private_data;
1940 * Do not allow a recursive hierarchy (this new sibling
1941 * becoming part of another group-sibling):
1943 if (group_leader->group_leader != group_leader)
1944 goto err_put_context;
1946 * Do not allow to attach to a group in a different
1947 * task or CPU context:
1949 if (group_leader->ctx != ctx)
1950 goto err_put_context;
1952 * Only a group leader can be exclusive or pinned
1954 if (hw_event.exclusive || hw_event.pinned)
1955 goto err_put_context;
1959 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
1962 goto err_put_context;
1964 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
1966 goto err_free_put_context;
1968 counter_file = fget_light(ret, &fput_needed2);
1970 goto err_free_put_context;
1972 counter->filp = counter_file;
1973 mutex_lock(&ctx->mutex);
1974 perf_install_in_context(ctx, counter, cpu);
1975 mutex_unlock(&ctx->mutex);
1977 fput_light(counter_file, fput_needed2);
1980 fput_light(group_file, fput_needed);
1984 err_free_put_context:
1994 * Initialize the perf_counter context in a task_struct:
1997 __perf_counter_init_context(struct perf_counter_context *ctx,
1998 struct task_struct *task)
2000 memset(ctx, 0, sizeof(*ctx));
2001 spin_lock_init(&ctx->lock);
2002 mutex_init(&ctx->mutex);
2003 INIT_LIST_HEAD(&ctx->counter_list);
2004 INIT_LIST_HEAD(&ctx->event_list);
2009 * inherit a counter from parent task to child task:
2011 static struct perf_counter *
2012 inherit_counter(struct perf_counter *parent_counter,
2013 struct task_struct *parent,
2014 struct perf_counter_context *parent_ctx,
2015 struct task_struct *child,
2016 struct perf_counter *group_leader,
2017 struct perf_counter_context *child_ctx)
2019 struct perf_counter *child_counter;
2022 * Instead of creating recursive hierarchies of counters,
2023 * we link inherited counters back to the original parent,
2024 * which has a filp for sure, which we use as the reference
2027 if (parent_counter->parent)
2028 parent_counter = parent_counter->parent;
2030 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2031 parent_counter->cpu, child_ctx,
2032 group_leader, GFP_KERNEL);
2037 * Link it up in the child's context:
2039 child_counter->task = child;
2040 list_add_counter(child_counter, child_ctx);
2041 child_ctx->nr_counters++;
2043 child_counter->parent = parent_counter;
2045 * inherit into child's child as well:
2047 child_counter->hw_event.inherit = 1;
2050 * Get a reference to the parent filp - we will fput it
2051 * when the child counter exits. This is safe to do because
2052 * we are in the parent and we know that the filp still
2053 * exists and has a nonzero count:
2055 atomic_long_inc(&parent_counter->filp->f_count);
2058 * Link this into the parent counter's child list
2060 mutex_lock(&parent_counter->mutex);
2061 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2064 * Make the child state follow the state of the parent counter,
2065 * not its hw_event.disabled bit. We hold the parent's mutex,
2066 * so we won't race with perf_counter_{en,dis}able_family.
2068 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2069 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2071 child_counter->state = PERF_COUNTER_STATE_OFF;
2073 mutex_unlock(&parent_counter->mutex);
2075 return child_counter;
2078 static int inherit_group(struct perf_counter *parent_counter,
2079 struct task_struct *parent,
2080 struct perf_counter_context *parent_ctx,
2081 struct task_struct *child,
2082 struct perf_counter_context *child_ctx)
2084 struct perf_counter *leader;
2085 struct perf_counter *sub;
2087 leader = inherit_counter(parent_counter, parent, parent_ctx,
2088 child, NULL, child_ctx);
2091 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2092 if (!inherit_counter(sub, parent, parent_ctx,
2093 child, leader, child_ctx))
2099 static void sync_child_counter(struct perf_counter *child_counter,
2100 struct perf_counter *parent_counter)
2102 u64 parent_val, child_val;
2104 parent_val = atomic64_read(&parent_counter->count);
2105 child_val = atomic64_read(&child_counter->count);
2108 * Add back the child's count to the parent's count:
2110 atomic64_add(child_val, &parent_counter->count);
2113 * Remove this counter from the parent's list
2115 mutex_lock(&parent_counter->mutex);
2116 list_del_init(&child_counter->child_list);
2117 mutex_unlock(&parent_counter->mutex);
2120 * Release the parent counter, if this was the last
2123 fput(parent_counter->filp);
2127 __perf_counter_exit_task(struct task_struct *child,
2128 struct perf_counter *child_counter,
2129 struct perf_counter_context *child_ctx)
2131 struct perf_counter *parent_counter;
2132 struct perf_counter *sub, *tmp;
2135 * If we do not self-reap then we have to wait for the
2136 * child task to unschedule (it will happen for sure),
2137 * so that its counter is at its final count. (This
2138 * condition triggers rarely - child tasks usually get
2139 * off their CPU before the parent has a chance to
2140 * get this far into the reaping action)
2142 if (child != current) {
2143 wait_task_inactive(child, 0);
2144 list_del_init(&child_counter->list_entry);
2146 struct perf_cpu_context *cpuctx;
2147 unsigned long flags;
2151 * Disable and unlink this counter.
2153 * Be careful about zapping the list - IRQ/NMI context
2154 * could still be processing it:
2156 curr_rq_lock_irq_save(&flags);
2157 perf_flags = hw_perf_save_disable();
2159 cpuctx = &__get_cpu_var(perf_cpu_context);
2161 group_sched_out(child_counter, cpuctx, child_ctx);
2163 list_del_init(&child_counter->list_entry);
2165 child_ctx->nr_counters--;
2167 hw_perf_restore(perf_flags);
2168 curr_rq_unlock_irq_restore(&flags);
2171 parent_counter = child_counter->parent;
2173 * It can happen that parent exits first, and has counters
2174 * that are still around due to the child reference. These
2175 * counters need to be zapped - but otherwise linger.
2177 if (parent_counter) {
2178 sync_child_counter(child_counter, parent_counter);
2179 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
2182 sync_child_counter(sub, sub->parent);
2186 kfree(child_counter);
2191 * When a child task exits, feed back counter values to parent counters.
2193 * Note: we may be running in child context, but the PID is not hashed
2194 * anymore so new counters will not be added.
2196 void perf_counter_exit_task(struct task_struct *child)
2198 struct perf_counter *child_counter, *tmp;
2199 struct perf_counter_context *child_ctx;
2201 child_ctx = &child->perf_counter_ctx;
2203 if (likely(!child_ctx->nr_counters))
2206 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
2208 __perf_counter_exit_task(child, child_counter, child_ctx);
2212 * Initialize the perf_counter context in task_struct
2214 void perf_counter_init_task(struct task_struct *child)
2216 struct perf_counter_context *child_ctx, *parent_ctx;
2217 struct perf_counter *counter;
2218 struct task_struct *parent = current;
2220 child_ctx = &child->perf_counter_ctx;
2221 parent_ctx = &parent->perf_counter_ctx;
2223 __perf_counter_init_context(child_ctx, child);
2226 * This is executed from the parent task context, so inherit
2227 * counters that have been marked for cloning:
2230 if (likely(!parent_ctx->nr_counters))
2234 * Lock the parent list. No need to lock the child - not PID
2235 * hashed yet and not running, so nobody can access it.
2237 mutex_lock(&parent_ctx->mutex);
2240 * We dont have to disable NMIs - we are only looking at
2241 * the list, not manipulating it:
2243 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2244 if (!counter->hw_event.inherit)
2247 if (inherit_group(counter, parent,
2248 parent_ctx, child, child_ctx))
2252 mutex_unlock(&parent_ctx->mutex);
2255 static void __cpuinit perf_counter_init_cpu(int cpu)
2257 struct perf_cpu_context *cpuctx;
2259 cpuctx = &per_cpu(perf_cpu_context, cpu);
2260 __perf_counter_init_context(&cpuctx->ctx, NULL);
2262 mutex_lock(&perf_resource_mutex);
2263 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2264 mutex_unlock(&perf_resource_mutex);
2266 hw_perf_counter_setup(cpu);
2269 #ifdef CONFIG_HOTPLUG_CPU
2270 static void __perf_counter_exit_cpu(void *info)
2272 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2273 struct perf_counter_context *ctx = &cpuctx->ctx;
2274 struct perf_counter *counter, *tmp;
2276 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2277 __perf_counter_remove_from_context(counter);
2279 static void perf_counter_exit_cpu(int cpu)
2281 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2282 struct perf_counter_context *ctx = &cpuctx->ctx;
2284 mutex_lock(&ctx->mutex);
2285 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2286 mutex_unlock(&ctx->mutex);
2289 static inline void perf_counter_exit_cpu(int cpu) { }
2292 static int __cpuinit
2293 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2295 unsigned int cpu = (long)hcpu;
2299 case CPU_UP_PREPARE:
2300 case CPU_UP_PREPARE_FROZEN:
2301 perf_counter_init_cpu(cpu);
2304 case CPU_DOWN_PREPARE:
2305 case CPU_DOWN_PREPARE_FROZEN:
2306 perf_counter_exit_cpu(cpu);
2316 static struct notifier_block __cpuinitdata perf_cpu_nb = {
2317 .notifier_call = perf_cpu_notify,
2320 static int __init perf_counter_init(void)
2322 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2323 (void *)(long)smp_processor_id());
2324 register_cpu_notifier(&perf_cpu_nb);
2328 early_initcall(perf_counter_init);
2330 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2332 return sprintf(buf, "%d\n", perf_reserved_percpu);
2336 perf_set_reserve_percpu(struct sysdev_class *class,
2340 struct perf_cpu_context *cpuctx;
2344 err = strict_strtoul(buf, 10, &val);
2347 if (val > perf_max_counters)
2350 mutex_lock(&perf_resource_mutex);
2351 perf_reserved_percpu = val;
2352 for_each_online_cpu(cpu) {
2353 cpuctx = &per_cpu(perf_cpu_context, cpu);
2354 spin_lock_irq(&cpuctx->ctx.lock);
2355 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2356 perf_max_counters - perf_reserved_percpu);
2357 cpuctx->max_pertask = mpt;
2358 spin_unlock_irq(&cpuctx->ctx.lock);
2360 mutex_unlock(&perf_resource_mutex);
2365 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2367 return sprintf(buf, "%d\n", perf_overcommit);
2371 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2376 err = strict_strtoul(buf, 10, &val);
2382 mutex_lock(&perf_resource_mutex);
2383 perf_overcommit = val;
2384 mutex_unlock(&perf_resource_mutex);
2389 static SYSDEV_CLASS_ATTR(
2392 perf_show_reserve_percpu,
2393 perf_set_reserve_percpu
2396 static SYSDEV_CLASS_ATTR(
2399 perf_show_overcommit,
2403 static struct attribute *perfclass_attrs[] = {
2404 &attr_reserve_percpu.attr,
2405 &attr_overcommit.attr,
2409 static struct attribute_group perfclass_attr_group = {
2410 .attrs = perfclass_attrs,
2411 .name = "perf_counters",
2414 static int __init perf_counter_sysfs_init(void)
2416 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2417 &perfclass_attr_group);
2419 device_initcall(perf_counter_sysfs_init);