2 * Performance counter core code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
8 * For licensing details see kernel-base/COPYING
13 #include <linux/cpu.h>
14 #include <linux/smp.h>
15 #include <linux/file.h>
16 #include <linux/poll.h>
17 #include <linux/sysfs.h>
18 #include <linux/ptrace.h>
19 #include <linux/percpu.h>
20 #include <linux/vmstat.h>
21 #include <linux/hardirq.h>
22 #include <linux/rculist.h>
23 #include <linux/uaccess.h>
24 #include <linux/syscalls.h>
25 #include <linux/anon_inodes.h>
26 #include <linux/kernel_stat.h>
27 #include <linux/perf_counter.h>
29 #include <asm/irq_regs.h>
32 * Each CPU has a list of per CPU counters:
34 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
36 int perf_max_counters __read_mostly = 1;
37 static int perf_reserved_percpu __read_mostly;
38 static int perf_overcommit __read_mostly = 1;
41 * Mutex for (sysadmin-configurable) counter reservations:
43 static DEFINE_MUTEX(perf_resource_mutex);
46 * Architecture provided APIs - weak aliases:
48 extern __weak const struct hw_perf_counter_ops *
49 hw_perf_counter_init(struct perf_counter *counter)
54 u64 __weak hw_perf_save_disable(void) { return 0; }
55 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
56 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
57 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
58 struct perf_cpu_context *cpuctx,
59 struct perf_counter_context *ctx, int cpu)
64 void __weak perf_counter_print_debug(void) { }
67 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
69 struct perf_counter *group_leader = counter->group_leader;
72 * Depending on whether it is a standalone or sibling counter,
73 * add it straight to the context's counter list, or to the group
74 * leader's sibling list:
76 if (counter->group_leader == counter)
77 list_add_tail(&counter->list_entry, &ctx->counter_list);
79 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
80 group_leader->nr_siblings++;
83 list_add_rcu(&counter->event_entry, &ctx->event_list);
87 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
89 struct perf_counter *sibling, *tmp;
91 list_del_init(&counter->list_entry);
92 list_del_rcu(&counter->event_entry);
94 if (counter->group_leader != counter)
95 counter->group_leader->nr_siblings--;
98 * If this was a group counter with sibling counters then
99 * upgrade the siblings to singleton counters by adding them
100 * to the context list directly:
102 list_for_each_entry_safe(sibling, tmp,
103 &counter->sibling_list, list_entry) {
105 list_move_tail(&sibling->list_entry, &ctx->counter_list);
106 sibling->group_leader = sibling;
111 counter_sched_out(struct perf_counter *counter,
112 struct perf_cpu_context *cpuctx,
113 struct perf_counter_context *ctx)
115 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
118 counter->state = PERF_COUNTER_STATE_INACTIVE;
119 counter->tstamp_stopped = ctx->time_now;
120 counter->hw_ops->disable(counter);
123 if (!is_software_counter(counter))
124 cpuctx->active_oncpu--;
126 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
127 cpuctx->exclusive = 0;
131 group_sched_out(struct perf_counter *group_counter,
132 struct perf_cpu_context *cpuctx,
133 struct perf_counter_context *ctx)
135 struct perf_counter *counter;
137 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
140 counter_sched_out(group_counter, cpuctx, ctx);
143 * Schedule out siblings (if any):
145 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
146 counter_sched_out(counter, cpuctx, ctx);
148 if (group_counter->hw_event.exclusive)
149 cpuctx->exclusive = 0;
153 * Cross CPU call to remove a performance counter
155 * We disable the counter on the hardware level first. After that we
156 * remove it from the context list.
158 static void __perf_counter_remove_from_context(void *info)
160 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
161 struct perf_counter *counter = info;
162 struct perf_counter_context *ctx = counter->ctx;
167 * If this is a task context, we need to check whether it is
168 * the current task context of this cpu. If not it has been
169 * scheduled out before the smp call arrived.
171 if (ctx->task && cpuctx->task_ctx != ctx)
174 curr_rq_lock_irq_save(&flags);
175 spin_lock(&ctx->lock);
177 counter_sched_out(counter, cpuctx, ctx);
179 counter->task = NULL;
183 * Protect the list operation against NMI by disabling the
184 * counters on a global level. NOP for non NMI based counters.
186 perf_flags = hw_perf_save_disable();
187 list_del_counter(counter, ctx);
188 hw_perf_restore(perf_flags);
192 * Allow more per task counters with respect to the
195 cpuctx->max_pertask =
196 min(perf_max_counters - ctx->nr_counters,
197 perf_max_counters - perf_reserved_percpu);
200 spin_unlock(&ctx->lock);
201 curr_rq_unlock_irq_restore(&flags);
206 * Remove the counter from a task's (or a CPU's) list of counters.
208 * Must be called with counter->mutex and ctx->mutex held.
210 * CPU counters are removed with a smp call. For task counters we only
211 * call when the task is on a CPU.
213 static void perf_counter_remove_from_context(struct perf_counter *counter)
215 struct perf_counter_context *ctx = counter->ctx;
216 struct task_struct *task = ctx->task;
220 * Per cpu counters are removed via an smp call and
221 * the removal is always sucessful.
223 smp_call_function_single(counter->cpu,
224 __perf_counter_remove_from_context,
230 task_oncpu_function_call(task, __perf_counter_remove_from_context,
233 spin_lock_irq(&ctx->lock);
235 * If the context is active we need to retry the smp call.
237 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
238 spin_unlock_irq(&ctx->lock);
243 * The lock prevents that this context is scheduled in so we
244 * can remove the counter safely, if the call above did not
247 if (!list_empty(&counter->list_entry)) {
249 list_del_counter(counter, ctx);
250 counter->task = NULL;
252 spin_unlock_irq(&ctx->lock);
256 * Get the current time for this context.
257 * If this is a task context, we use the task's task clock,
258 * or for a per-cpu context, we use the cpu clock.
260 static u64 get_context_time(struct perf_counter_context *ctx, int update)
262 struct task_struct *curr = ctx->task;
265 return cpu_clock(smp_processor_id());
267 return __task_delta_exec(curr, update) + curr->se.sum_exec_runtime;
271 * Update the record of the current time in a context.
273 static void update_context_time(struct perf_counter_context *ctx, int update)
275 ctx->time_now = get_context_time(ctx, update) - ctx->time_lost;
279 * Update the total_time_enabled and total_time_running fields for a counter.
281 static void update_counter_times(struct perf_counter *counter)
283 struct perf_counter_context *ctx = counter->ctx;
286 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
287 counter->total_time_enabled = ctx->time_now -
288 counter->tstamp_enabled;
289 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
290 run_end = counter->tstamp_stopped;
292 run_end = ctx->time_now;
293 counter->total_time_running = run_end - counter->tstamp_running;
298 * Update total_time_enabled and total_time_running for all counters in a group.
300 static void update_group_times(struct perf_counter *leader)
302 struct perf_counter *counter;
304 update_counter_times(leader);
305 list_for_each_entry(counter, &leader->sibling_list, list_entry)
306 update_counter_times(counter);
310 * Cross CPU call to disable a performance counter
312 static void __perf_counter_disable(void *info)
314 struct perf_counter *counter = info;
315 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
316 struct perf_counter_context *ctx = counter->ctx;
320 * If this is a per-task counter, need to check whether this
321 * counter's task is the current task on this cpu.
323 if (ctx->task && cpuctx->task_ctx != ctx)
326 curr_rq_lock_irq_save(&flags);
327 spin_lock(&ctx->lock);
330 * If the counter is on, turn it off.
331 * If it is in error state, leave it in error state.
333 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
334 update_context_time(ctx, 1);
335 update_counter_times(counter);
336 if (counter == counter->group_leader)
337 group_sched_out(counter, cpuctx, ctx);
339 counter_sched_out(counter, cpuctx, ctx);
340 counter->state = PERF_COUNTER_STATE_OFF;
343 spin_unlock(&ctx->lock);
344 curr_rq_unlock_irq_restore(&flags);
350 static void perf_counter_disable(struct perf_counter *counter)
352 struct perf_counter_context *ctx = counter->ctx;
353 struct task_struct *task = ctx->task;
357 * Disable the counter on the cpu that it's on
359 smp_call_function_single(counter->cpu, __perf_counter_disable,
365 task_oncpu_function_call(task, __perf_counter_disable, counter);
367 spin_lock_irq(&ctx->lock);
369 * If the counter is still active, we need to retry the cross-call.
371 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
372 spin_unlock_irq(&ctx->lock);
377 * Since we have the lock this context can't be scheduled
378 * in, so we can change the state safely.
380 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
381 update_counter_times(counter);
382 counter->state = PERF_COUNTER_STATE_OFF;
385 spin_unlock_irq(&ctx->lock);
389 * Disable a counter and all its children.
391 static void perf_counter_disable_family(struct perf_counter *counter)
393 struct perf_counter *child;
395 perf_counter_disable(counter);
398 * Lock the mutex to protect the list of children
400 mutex_lock(&counter->mutex);
401 list_for_each_entry(child, &counter->child_list, child_list)
402 perf_counter_disable(child);
403 mutex_unlock(&counter->mutex);
407 counter_sched_in(struct perf_counter *counter,
408 struct perf_cpu_context *cpuctx,
409 struct perf_counter_context *ctx,
412 if (counter->state <= PERF_COUNTER_STATE_OFF)
415 counter->state = PERF_COUNTER_STATE_ACTIVE;
416 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
418 * The new state must be visible before we turn it on in the hardware:
422 if (counter->hw_ops->enable(counter)) {
423 counter->state = PERF_COUNTER_STATE_INACTIVE;
428 counter->tstamp_running += ctx->time_now - counter->tstamp_stopped;
430 if (!is_software_counter(counter))
431 cpuctx->active_oncpu++;
434 if (counter->hw_event.exclusive)
435 cpuctx->exclusive = 1;
441 * Return 1 for a group consisting entirely of software counters,
442 * 0 if the group contains any hardware counters.
444 static int is_software_only_group(struct perf_counter *leader)
446 struct perf_counter *counter;
448 if (!is_software_counter(leader))
451 list_for_each_entry(counter, &leader->sibling_list, list_entry)
452 if (!is_software_counter(counter))
459 * Work out whether we can put this counter group on the CPU now.
461 static int group_can_go_on(struct perf_counter *counter,
462 struct perf_cpu_context *cpuctx,
466 * Groups consisting entirely of software counters can always go on.
468 if (is_software_only_group(counter))
471 * If an exclusive group is already on, no other hardware
472 * counters can go on.
474 if (cpuctx->exclusive)
477 * If this group is exclusive and there are already
478 * counters on the CPU, it can't go on.
480 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
483 * Otherwise, try to add it if all previous groups were able
489 static void add_counter_to_ctx(struct perf_counter *counter,
490 struct perf_counter_context *ctx)
492 list_add_counter(counter, ctx);
494 counter->prev_state = PERF_COUNTER_STATE_OFF;
495 counter->tstamp_enabled = ctx->time_now;
496 counter->tstamp_running = ctx->time_now;
497 counter->tstamp_stopped = ctx->time_now;
501 * Cross CPU call to install and enable a performance counter
503 static void __perf_install_in_context(void *info)
505 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
506 struct perf_counter *counter = info;
507 struct perf_counter_context *ctx = counter->ctx;
508 struct perf_counter *leader = counter->group_leader;
509 int cpu = smp_processor_id();
515 * If this is a task context, we need to check whether it is
516 * the current task context of this cpu. If not it has been
517 * scheduled out before the smp call arrived.
519 if (ctx->task && cpuctx->task_ctx != ctx)
522 curr_rq_lock_irq_save(&flags);
523 spin_lock(&ctx->lock);
524 update_context_time(ctx, 1);
527 * Protect the list operation against NMI by disabling the
528 * counters on a global level. NOP for non NMI based counters.
530 perf_flags = hw_perf_save_disable();
532 add_counter_to_ctx(counter, ctx);
535 * Don't put the counter on if it is disabled or if
536 * it is in a group and the group isn't on.
538 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
539 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
543 * An exclusive counter can't go on if there are already active
544 * hardware counters, and no hardware counter can go on if there
545 * is already an exclusive counter on.
547 if (!group_can_go_on(counter, cpuctx, 1))
550 err = counter_sched_in(counter, cpuctx, ctx, cpu);
554 * This counter couldn't go on. If it is in a group
555 * then we have to pull the whole group off.
556 * If the counter group is pinned then put it in error state.
558 if (leader != counter)
559 group_sched_out(leader, cpuctx, ctx);
560 if (leader->hw_event.pinned) {
561 update_group_times(leader);
562 leader->state = PERF_COUNTER_STATE_ERROR;
566 if (!err && !ctx->task && cpuctx->max_pertask)
567 cpuctx->max_pertask--;
570 hw_perf_restore(perf_flags);
572 spin_unlock(&ctx->lock);
573 curr_rq_unlock_irq_restore(&flags);
577 * Attach a performance counter to a context
579 * First we add the counter to the list with the hardware enable bit
580 * in counter->hw_config cleared.
582 * If the counter is attached to a task which is on a CPU we use a smp
583 * call to enable it in the task context. The task might have been
584 * scheduled away, but we check this in the smp call again.
586 * Must be called with ctx->mutex held.
589 perf_install_in_context(struct perf_counter_context *ctx,
590 struct perf_counter *counter,
593 struct task_struct *task = ctx->task;
597 * Per cpu counters are installed via an smp call and
598 * the install is always sucessful.
600 smp_call_function_single(cpu, __perf_install_in_context,
605 counter->task = task;
607 task_oncpu_function_call(task, __perf_install_in_context,
610 spin_lock_irq(&ctx->lock);
612 * we need to retry the smp call.
614 if (ctx->is_active && list_empty(&counter->list_entry)) {
615 spin_unlock_irq(&ctx->lock);
620 * The lock prevents that this context is scheduled in so we
621 * can add the counter safely, if it the call above did not
624 if (list_empty(&counter->list_entry))
625 add_counter_to_ctx(counter, ctx);
626 spin_unlock_irq(&ctx->lock);
630 * Cross CPU call to enable a performance counter
632 static void __perf_counter_enable(void *info)
634 struct perf_counter *counter = info;
635 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
636 struct perf_counter_context *ctx = counter->ctx;
637 struct perf_counter *leader = counter->group_leader;
642 * If this is a per-task counter, need to check whether this
643 * counter's task is the current task on this cpu.
645 if (ctx->task && cpuctx->task_ctx != ctx)
648 curr_rq_lock_irq_save(&flags);
649 spin_lock(&ctx->lock);
650 update_context_time(ctx, 1);
652 counter->prev_state = counter->state;
653 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
655 counter->state = PERF_COUNTER_STATE_INACTIVE;
656 counter->tstamp_enabled = ctx->time_now - counter->total_time_enabled;
659 * If the counter is in a group and isn't the group leader,
660 * then don't put it on unless the group is on.
662 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
665 if (!group_can_go_on(counter, cpuctx, 1))
668 err = counter_sched_in(counter, cpuctx, ctx,
673 * If this counter can't go on and it's part of a
674 * group, then the whole group has to come off.
676 if (leader != counter)
677 group_sched_out(leader, cpuctx, ctx);
678 if (leader->hw_event.pinned) {
679 update_group_times(leader);
680 leader->state = PERF_COUNTER_STATE_ERROR;
685 spin_unlock(&ctx->lock);
686 curr_rq_unlock_irq_restore(&flags);
692 static void perf_counter_enable(struct perf_counter *counter)
694 struct perf_counter_context *ctx = counter->ctx;
695 struct task_struct *task = ctx->task;
699 * Enable the counter on the cpu that it's on
701 smp_call_function_single(counter->cpu, __perf_counter_enable,
706 spin_lock_irq(&ctx->lock);
707 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
711 * If the counter is in error state, clear that first.
712 * That way, if we see the counter in error state below, we
713 * know that it has gone back into error state, as distinct
714 * from the task having been scheduled away before the
715 * cross-call arrived.
717 if (counter->state == PERF_COUNTER_STATE_ERROR)
718 counter->state = PERF_COUNTER_STATE_OFF;
721 spin_unlock_irq(&ctx->lock);
722 task_oncpu_function_call(task, __perf_counter_enable, counter);
724 spin_lock_irq(&ctx->lock);
727 * If the context is active and the counter is still off,
728 * we need to retry the cross-call.
730 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
734 * Since we have the lock this context can't be scheduled
735 * in, so we can change the state safely.
737 if (counter->state == PERF_COUNTER_STATE_OFF) {
738 counter->state = PERF_COUNTER_STATE_INACTIVE;
739 counter->tstamp_enabled = ctx->time_now -
740 counter->total_time_enabled;
743 spin_unlock_irq(&ctx->lock);
747 * Enable a counter and all its children.
749 static void perf_counter_enable_family(struct perf_counter *counter)
751 struct perf_counter *child;
753 perf_counter_enable(counter);
756 * Lock the mutex to protect the list of children
758 mutex_lock(&counter->mutex);
759 list_for_each_entry(child, &counter->child_list, child_list)
760 perf_counter_enable(child);
761 mutex_unlock(&counter->mutex);
764 void __perf_counter_sched_out(struct perf_counter_context *ctx,
765 struct perf_cpu_context *cpuctx)
767 struct perf_counter *counter;
770 spin_lock(&ctx->lock);
772 if (likely(!ctx->nr_counters))
774 update_context_time(ctx, 0);
776 flags = hw_perf_save_disable();
777 if (ctx->nr_active) {
778 list_for_each_entry(counter, &ctx->counter_list, list_entry)
779 group_sched_out(counter, cpuctx, ctx);
781 hw_perf_restore(flags);
783 spin_unlock(&ctx->lock);
787 * Called from scheduler to remove the counters of the current task,
788 * with interrupts disabled.
790 * We stop each counter and update the counter value in counter->count.
792 * This does not protect us against NMI, but disable()
793 * sets the disabled bit in the control field of counter _before_
794 * accessing the counter control register. If a NMI hits, then it will
795 * not restart the counter.
797 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
799 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
800 struct perf_counter_context *ctx = &task->perf_counter_ctx;
801 struct pt_regs *regs;
803 if (likely(!cpuctx->task_ctx))
806 regs = task_pt_regs(task);
807 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs);
808 __perf_counter_sched_out(ctx, cpuctx);
810 cpuctx->task_ctx = NULL;
813 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
815 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
819 group_sched_in(struct perf_counter *group_counter,
820 struct perf_cpu_context *cpuctx,
821 struct perf_counter_context *ctx,
824 struct perf_counter *counter, *partial_group;
827 if (group_counter->state == PERF_COUNTER_STATE_OFF)
830 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
832 return ret < 0 ? ret : 0;
834 group_counter->prev_state = group_counter->state;
835 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
839 * Schedule in siblings as one group (if any):
841 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
842 counter->prev_state = counter->state;
843 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
844 partial_group = counter;
853 * Groups can be scheduled in as one unit only, so undo any
854 * partial group before returning:
856 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
857 if (counter == partial_group)
859 counter_sched_out(counter, cpuctx, ctx);
861 counter_sched_out(group_counter, cpuctx, ctx);
867 __perf_counter_sched_in(struct perf_counter_context *ctx,
868 struct perf_cpu_context *cpuctx, int cpu)
870 struct perf_counter *counter;
874 spin_lock(&ctx->lock);
876 if (likely(!ctx->nr_counters))
880 * Add any time since the last sched_out to the lost time
881 * so it doesn't get included in the total_time_enabled and
882 * total_time_running measures for counters in the context.
884 ctx->time_lost = get_context_time(ctx, 0) - ctx->time_now;
886 flags = hw_perf_save_disable();
889 * First go through the list and put on any pinned groups
890 * in order to give them the best chance of going on.
892 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
893 if (counter->state <= PERF_COUNTER_STATE_OFF ||
894 !counter->hw_event.pinned)
896 if (counter->cpu != -1 && counter->cpu != cpu)
899 if (group_can_go_on(counter, cpuctx, 1))
900 group_sched_in(counter, cpuctx, ctx, cpu);
903 * If this pinned group hasn't been scheduled,
904 * put it in error state.
906 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
907 update_group_times(counter);
908 counter->state = PERF_COUNTER_STATE_ERROR;
912 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
914 * Ignore counters in OFF or ERROR state, and
915 * ignore pinned counters since we did them already.
917 if (counter->state <= PERF_COUNTER_STATE_OFF ||
918 counter->hw_event.pinned)
922 * Listen to the 'cpu' scheduling filter constraint
925 if (counter->cpu != -1 && counter->cpu != cpu)
928 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
929 if (group_sched_in(counter, cpuctx, ctx, cpu))
933 hw_perf_restore(flags);
935 spin_unlock(&ctx->lock);
939 * Called from scheduler to add the counters of the current task
940 * with interrupts disabled.
942 * We restore the counter value and then enable it.
944 * This does not protect us against NMI, but enable()
945 * sets the enabled bit in the control field of counter _before_
946 * accessing the counter control register. If a NMI hits, then it will
947 * keep the counter running.
949 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
951 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
952 struct perf_counter_context *ctx = &task->perf_counter_ctx;
954 __perf_counter_sched_in(ctx, cpuctx, cpu);
955 cpuctx->task_ctx = ctx;
958 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
960 struct perf_counter_context *ctx = &cpuctx->ctx;
962 __perf_counter_sched_in(ctx, cpuctx, cpu);
965 int perf_counter_task_disable(void)
967 struct task_struct *curr = current;
968 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
969 struct perf_counter *counter;
974 if (likely(!ctx->nr_counters))
977 curr_rq_lock_irq_save(&flags);
978 cpu = smp_processor_id();
980 /* force the update of the task clock: */
981 __task_delta_exec(curr, 1);
983 perf_counter_task_sched_out(curr, cpu);
985 spin_lock(&ctx->lock);
988 * Disable all the counters:
990 perf_flags = hw_perf_save_disable();
992 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
993 if (counter->state != PERF_COUNTER_STATE_ERROR) {
994 update_group_times(counter);
995 counter->state = PERF_COUNTER_STATE_OFF;
999 hw_perf_restore(perf_flags);
1001 spin_unlock(&ctx->lock);
1003 curr_rq_unlock_irq_restore(&flags);
1008 int perf_counter_task_enable(void)
1010 struct task_struct *curr = current;
1011 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1012 struct perf_counter *counter;
1013 unsigned long flags;
1017 if (likely(!ctx->nr_counters))
1020 curr_rq_lock_irq_save(&flags);
1021 cpu = smp_processor_id();
1023 /* force the update of the task clock: */
1024 __task_delta_exec(curr, 1);
1026 perf_counter_task_sched_out(curr, cpu);
1028 spin_lock(&ctx->lock);
1031 * Disable all the counters:
1033 perf_flags = hw_perf_save_disable();
1035 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1036 if (counter->state > PERF_COUNTER_STATE_OFF)
1038 counter->state = PERF_COUNTER_STATE_INACTIVE;
1039 counter->tstamp_enabled = ctx->time_now -
1040 counter->total_time_enabled;
1041 counter->hw_event.disabled = 0;
1043 hw_perf_restore(perf_flags);
1045 spin_unlock(&ctx->lock);
1047 perf_counter_task_sched_in(curr, cpu);
1049 curr_rq_unlock_irq_restore(&flags);
1055 * Round-robin a context's counters:
1057 static void rotate_ctx(struct perf_counter_context *ctx)
1059 struct perf_counter *counter;
1062 if (!ctx->nr_counters)
1065 spin_lock(&ctx->lock);
1067 * Rotate the first entry last (works just fine for group counters too):
1069 perf_flags = hw_perf_save_disable();
1070 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1071 list_move_tail(&counter->list_entry, &ctx->counter_list);
1074 hw_perf_restore(perf_flags);
1076 spin_unlock(&ctx->lock);
1079 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1081 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1082 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1083 const int rotate_percpu = 0;
1086 perf_counter_cpu_sched_out(cpuctx);
1087 perf_counter_task_sched_out(curr, cpu);
1090 rotate_ctx(&cpuctx->ctx);
1094 perf_counter_cpu_sched_in(cpuctx, cpu);
1095 perf_counter_task_sched_in(curr, cpu);
1099 * Cross CPU call to read the hardware counter
1101 static void __read(void *info)
1103 struct perf_counter *counter = info;
1104 struct perf_counter_context *ctx = counter->ctx;
1105 unsigned long flags;
1107 curr_rq_lock_irq_save(&flags);
1109 update_context_time(ctx, 1);
1110 counter->hw_ops->read(counter);
1111 update_counter_times(counter);
1112 curr_rq_unlock_irq_restore(&flags);
1115 static u64 perf_counter_read(struct perf_counter *counter)
1118 * If counter is enabled and currently active on a CPU, update the
1119 * value in the counter structure:
1121 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1122 smp_call_function_single(counter->oncpu,
1123 __read, counter, 1);
1124 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1125 update_counter_times(counter);
1128 return atomic64_read(&counter->count);
1131 static void put_context(struct perf_counter_context *ctx)
1134 put_task_struct(ctx->task);
1137 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1139 struct perf_cpu_context *cpuctx;
1140 struct perf_counter_context *ctx;
1141 struct task_struct *task;
1144 * If cpu is not a wildcard then this is a percpu counter:
1147 /* Must be root to operate on a CPU counter: */
1148 if (!capable(CAP_SYS_ADMIN))
1149 return ERR_PTR(-EACCES);
1151 if (cpu < 0 || cpu > num_possible_cpus())
1152 return ERR_PTR(-EINVAL);
1155 * We could be clever and allow to attach a counter to an
1156 * offline CPU and activate it when the CPU comes up, but
1159 if (!cpu_isset(cpu, cpu_online_map))
1160 return ERR_PTR(-ENODEV);
1162 cpuctx = &per_cpu(perf_cpu_context, cpu);
1172 task = find_task_by_vpid(pid);
1174 get_task_struct(task);
1178 return ERR_PTR(-ESRCH);
1180 ctx = &task->perf_counter_ctx;
1183 /* Reuse ptrace permission checks for now. */
1184 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1186 return ERR_PTR(-EACCES);
1192 static void free_counter_rcu(struct rcu_head *head)
1194 struct perf_counter *counter;
1196 counter = container_of(head, struct perf_counter, rcu_head);
1200 static void perf_pending_sync(struct perf_counter *counter);
1202 static void free_counter(struct perf_counter *counter)
1204 perf_pending_sync(counter);
1206 if (counter->destroy)
1207 counter->destroy(counter);
1209 call_rcu(&counter->rcu_head, free_counter_rcu);
1213 * Called when the last reference to the file is gone.
1215 static int perf_release(struct inode *inode, struct file *file)
1217 struct perf_counter *counter = file->private_data;
1218 struct perf_counter_context *ctx = counter->ctx;
1220 file->private_data = NULL;
1222 mutex_lock(&ctx->mutex);
1223 mutex_lock(&counter->mutex);
1225 perf_counter_remove_from_context(counter);
1227 mutex_unlock(&counter->mutex);
1228 mutex_unlock(&ctx->mutex);
1230 free_counter(counter);
1237 * Read the performance counter - simple non blocking version for now
1240 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1246 * Return end-of-file for a read on a counter that is in
1247 * error state (i.e. because it was pinned but it couldn't be
1248 * scheduled on to the CPU at some point).
1250 if (counter->state == PERF_COUNTER_STATE_ERROR)
1253 mutex_lock(&counter->mutex);
1254 values[0] = perf_counter_read(counter);
1256 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1257 values[n++] = counter->total_time_enabled +
1258 atomic64_read(&counter->child_total_time_enabled);
1259 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1260 values[n++] = counter->total_time_running +
1261 atomic64_read(&counter->child_total_time_running);
1262 mutex_unlock(&counter->mutex);
1264 if (count < n * sizeof(u64))
1266 count = n * sizeof(u64);
1268 if (copy_to_user(buf, values, count))
1275 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1277 struct perf_counter *counter = file->private_data;
1279 return perf_read_hw(counter, buf, count);
1282 static unsigned int perf_poll(struct file *file, poll_table *wait)
1284 struct perf_counter *counter = file->private_data;
1285 struct perf_mmap_data *data;
1286 unsigned int events;
1289 data = rcu_dereference(counter->data);
1291 events = atomic_xchg(&data->wakeup, 0);
1296 poll_wait(file, &counter->waitq, wait);
1301 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1303 struct perf_counter *counter = file->private_data;
1307 case PERF_COUNTER_IOC_ENABLE:
1308 perf_counter_enable_family(counter);
1310 case PERF_COUNTER_IOC_DISABLE:
1311 perf_counter_disable_family(counter);
1320 * Callers need to ensure there can be no nesting of this function, otherwise
1321 * the seqlock logic goes bad. We can not serialize this because the arch
1322 * code calls this from NMI context.
1324 void perf_counter_update_userpage(struct perf_counter *counter)
1326 struct perf_mmap_data *data;
1327 struct perf_counter_mmap_page *userpg;
1330 data = rcu_dereference(counter->data);
1334 userpg = data->user_page;
1337 * Disable preemption so as to not let the corresponding user-space
1338 * spin too long if we get preempted.
1343 userpg->index = counter->hw.idx;
1344 userpg->offset = atomic64_read(&counter->count);
1345 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1346 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1355 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1357 struct perf_counter *counter = vma->vm_file->private_data;
1358 struct perf_mmap_data *data;
1359 int ret = VM_FAULT_SIGBUS;
1362 data = rcu_dereference(counter->data);
1366 if (vmf->pgoff == 0) {
1367 vmf->page = virt_to_page(data->user_page);
1369 int nr = vmf->pgoff - 1;
1371 if ((unsigned)nr > data->nr_pages)
1374 vmf->page = virt_to_page(data->data_pages[nr]);
1376 get_page(vmf->page);
1384 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1386 struct perf_mmap_data *data;
1390 WARN_ON(atomic_read(&counter->mmap_count));
1392 size = sizeof(struct perf_mmap_data);
1393 size += nr_pages * sizeof(void *);
1395 data = kzalloc(size, GFP_KERNEL);
1399 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1400 if (!data->user_page)
1401 goto fail_user_page;
1403 for (i = 0; i < nr_pages; i++) {
1404 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1405 if (!data->data_pages[i])
1406 goto fail_data_pages;
1409 data->nr_pages = nr_pages;
1411 rcu_assign_pointer(counter->data, data);
1416 for (i--; i >= 0; i--)
1417 free_page((unsigned long)data->data_pages[i]);
1419 free_page((unsigned long)data->user_page);
1428 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1430 struct perf_mmap_data *data = container_of(rcu_head,
1431 struct perf_mmap_data, rcu_head);
1434 free_page((unsigned long)data->user_page);
1435 for (i = 0; i < data->nr_pages; i++)
1436 free_page((unsigned long)data->data_pages[i]);
1440 static void perf_mmap_data_free(struct perf_counter *counter)
1442 struct perf_mmap_data *data = counter->data;
1444 WARN_ON(atomic_read(&counter->mmap_count));
1446 rcu_assign_pointer(counter->data, NULL);
1447 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1450 static void perf_mmap_open(struct vm_area_struct *vma)
1452 struct perf_counter *counter = vma->vm_file->private_data;
1454 atomic_inc(&counter->mmap_count);
1457 static void perf_mmap_close(struct vm_area_struct *vma)
1459 struct perf_counter *counter = vma->vm_file->private_data;
1461 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1462 &counter->mmap_mutex)) {
1463 perf_mmap_data_free(counter);
1464 mutex_unlock(&counter->mmap_mutex);
1468 static struct vm_operations_struct perf_mmap_vmops = {
1469 .open = perf_mmap_open,
1470 .close = perf_mmap_close,
1471 .fault = perf_mmap_fault,
1474 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1476 struct perf_counter *counter = file->private_data;
1477 unsigned long vma_size;
1478 unsigned long nr_pages;
1479 unsigned long locked, lock_limit;
1482 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1485 vma_size = vma->vm_end - vma->vm_start;
1486 nr_pages = (vma_size / PAGE_SIZE) - 1;
1489 * If we have data pages ensure they're a power-of-two number, so we
1490 * can do bitmasks instead of modulo.
1492 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1495 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1498 if (vma->vm_pgoff != 0)
1501 locked = vma_size >> PAGE_SHIFT;
1502 locked += vma->vm_mm->locked_vm;
1504 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1505 lock_limit >>= PAGE_SHIFT;
1507 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK))
1510 mutex_lock(&counter->mmap_mutex);
1511 if (atomic_inc_not_zero(&counter->mmap_count))
1514 WARN_ON(counter->data);
1515 ret = perf_mmap_data_alloc(counter, nr_pages);
1517 atomic_set(&counter->mmap_count, 1);
1519 mutex_unlock(&counter->mmap_mutex);
1521 vma->vm_flags &= ~VM_MAYWRITE;
1522 vma->vm_flags |= VM_RESERVED;
1523 vma->vm_ops = &perf_mmap_vmops;
1528 static const struct file_operations perf_fops = {
1529 .release = perf_release,
1532 .unlocked_ioctl = perf_ioctl,
1533 .compat_ioctl = perf_ioctl,
1538 * Perf counter wakeup
1540 * If there's data, ensure we set the poll() state and publish everything
1541 * to user-space before waking everybody up.
1544 void perf_counter_wakeup(struct perf_counter *counter)
1546 struct perf_mmap_data *data;
1549 data = rcu_dereference(counter->data);
1551 (void)atomic_xchg(&data->wakeup, POLL_IN);
1553 * Ensure all data writes are issued before updating the
1554 * user-space data head information. The matching rmb()
1555 * will be in userspace after reading this value.
1558 data->user_page->data_head = atomic_read(&data->head);
1562 wake_up_all(&counter->waitq);
1568 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1570 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1571 * single linked list and use cmpxchg() to add entries lockless.
1574 #define PENDING_TAIL ((struct perf_wakeup_entry *)-1UL)
1576 static DEFINE_PER_CPU(struct perf_wakeup_entry *, perf_wakeup_head) = {
1580 static void perf_pending_queue(struct perf_counter *counter)
1582 struct perf_wakeup_entry **head;
1583 struct perf_wakeup_entry *prev, *next;
1585 if (cmpxchg(&counter->wakeup.next, NULL, PENDING_TAIL) != NULL)
1588 head = &get_cpu_var(perf_wakeup_head);
1591 prev = counter->wakeup.next = *head;
1592 next = &counter->wakeup;
1593 } while (cmpxchg(head, prev, next) != prev);
1595 set_perf_counter_pending();
1597 put_cpu_var(perf_wakeup_head);
1600 static int __perf_pending_run(void)
1602 struct perf_wakeup_entry *list;
1605 list = xchg(&__get_cpu_var(perf_wakeup_head), PENDING_TAIL);
1606 while (list != PENDING_TAIL) {
1607 struct perf_counter *counter = container_of(list,
1608 struct perf_counter, wakeup);
1612 counter->wakeup.next = NULL;
1614 * Ensure we observe the unqueue before we issue the wakeup,
1615 * so that we won't be waiting forever.
1616 * -- see perf_not_pending().
1620 perf_counter_wakeup(counter);
1627 static inline int perf_not_pending(struct perf_counter *counter)
1630 * If we flush on whatever cpu we run, there is a chance we don't
1634 __perf_pending_run();
1638 * Ensure we see the proper queue state before going to sleep
1639 * so that we do not miss the wakeup. -- see perf_pending_handle()
1642 return counter->wakeup.next == NULL;
1645 static void perf_pending_sync(struct perf_counter *counter)
1647 wait_event(counter->waitq, perf_not_pending(counter));
1650 void perf_counter_do_pending(void)
1652 __perf_pending_run();
1659 struct perf_output_handle {
1660 struct perf_counter *counter;
1661 struct perf_mmap_data *data;
1662 unsigned int offset;
1667 static int perf_output_begin(struct perf_output_handle *handle,
1668 struct perf_counter *counter, unsigned int size)
1670 struct perf_mmap_data *data;
1671 unsigned int offset, head;
1674 data = rcu_dereference(counter->data);
1678 if (!data->nr_pages)
1682 offset = head = atomic_read(&data->head);
1684 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1686 handle->counter = counter;
1687 handle->data = data;
1688 handle->offset = offset;
1689 handle->head = head;
1690 handle->wakeup = (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT);
1700 static void perf_output_copy(struct perf_output_handle *handle,
1701 void *buf, unsigned int len)
1703 unsigned int pages_mask;
1704 unsigned int offset;
1708 offset = handle->offset;
1709 pages_mask = handle->data->nr_pages - 1;
1710 pages = handle->data->data_pages;
1713 unsigned int page_offset;
1716 nr = (offset >> PAGE_SHIFT) & pages_mask;
1717 page_offset = offset & (PAGE_SIZE - 1);
1718 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1720 memcpy(pages[nr] + page_offset, buf, size);
1727 handle->offset = offset;
1729 WARN_ON_ONCE(handle->offset > handle->head);
1732 #define perf_output_put(handle, x) \
1733 perf_output_copy((handle), &(x), sizeof(x))
1735 static void perf_output_end(struct perf_output_handle *handle, int nmi)
1737 if (handle->wakeup) {
1739 perf_pending_queue(handle->counter);
1741 perf_counter_wakeup(handle->counter);
1746 static int perf_output_write(struct perf_counter *counter, int nmi,
1747 void *buf, ssize_t size)
1749 struct perf_output_handle handle;
1752 ret = perf_output_begin(&handle, counter, size);
1756 perf_output_copy(&handle, buf, size);
1757 perf_output_end(&handle, nmi);
1763 static void perf_output_simple(struct perf_counter *counter,
1764 int nmi, struct pt_regs *regs)
1768 struct perf_event_header header;
1773 event.header.type = PERF_EVENT_IP;
1774 event.ip = instruction_pointer(regs);
1776 size = sizeof(event);
1778 if (counter->hw_event.include_tid) {
1779 /* namespace issues */
1780 event.pid = current->group_leader->pid;
1781 event.tid = current->pid;
1783 event.header.type |= __PERF_EVENT_TID;
1785 size -= sizeof(u64);
1787 event.header.size = size;
1789 perf_output_write(counter, nmi, &event, size);
1792 static void perf_output_group(struct perf_counter *counter, int nmi)
1794 struct perf_output_handle handle;
1795 struct perf_event_header header;
1796 struct perf_counter *leader, *sub;
1804 size = sizeof(header) + counter->nr_siblings * sizeof(entry);
1806 ret = perf_output_begin(&handle, counter, size);
1810 header.type = PERF_EVENT_GROUP;
1813 perf_output_put(&handle, header);
1815 leader = counter->group_leader;
1816 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1818 sub->hw_ops->read(sub);
1820 entry.event = sub->hw_event.config;
1821 entry.counter = atomic64_read(&sub->count);
1823 perf_output_put(&handle, entry);
1826 perf_output_end(&handle, nmi);
1829 void perf_counter_output(struct perf_counter *counter,
1830 int nmi, struct pt_regs *regs)
1832 switch (counter->hw_event.record_type) {
1833 case PERF_RECORD_SIMPLE:
1836 case PERF_RECORD_IRQ:
1837 perf_output_simple(counter, nmi, regs);
1840 case PERF_RECORD_GROUP:
1841 perf_output_group(counter, nmi);
1847 * Generic software counter infrastructure
1850 static void perf_swcounter_update(struct perf_counter *counter)
1852 struct hw_perf_counter *hwc = &counter->hw;
1857 prev = atomic64_read(&hwc->prev_count);
1858 now = atomic64_read(&hwc->count);
1859 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
1864 atomic64_add(delta, &counter->count);
1865 atomic64_sub(delta, &hwc->period_left);
1868 static void perf_swcounter_set_period(struct perf_counter *counter)
1870 struct hw_perf_counter *hwc = &counter->hw;
1871 s64 left = atomic64_read(&hwc->period_left);
1872 s64 period = hwc->irq_period;
1874 if (unlikely(left <= -period)) {
1876 atomic64_set(&hwc->period_left, left);
1879 if (unlikely(left <= 0)) {
1881 atomic64_add(period, &hwc->period_left);
1884 atomic64_set(&hwc->prev_count, -left);
1885 atomic64_set(&hwc->count, -left);
1888 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
1890 struct perf_counter *counter;
1891 struct pt_regs *regs;
1893 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
1894 counter->hw_ops->read(counter);
1896 regs = get_irq_regs();
1898 * In case we exclude kernel IPs or are somehow not in interrupt
1899 * context, provide the next best thing, the user IP.
1901 if ((counter->hw_event.exclude_kernel || !regs) &&
1902 !counter->hw_event.exclude_user)
1903 regs = task_pt_regs(current);
1906 perf_counter_output(counter, 0, regs);
1908 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
1910 return HRTIMER_RESTART;
1913 static void perf_swcounter_overflow(struct perf_counter *counter,
1914 int nmi, struct pt_regs *regs)
1916 perf_swcounter_update(counter);
1917 perf_swcounter_set_period(counter);
1918 perf_counter_output(counter, nmi, regs);
1921 static int perf_swcounter_match(struct perf_counter *counter,
1922 enum perf_event_types type,
1923 u32 event, struct pt_regs *regs)
1925 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1928 if (perf_event_raw(&counter->hw_event))
1931 if (perf_event_type(&counter->hw_event) != type)
1934 if (perf_event_id(&counter->hw_event) != event)
1937 if (counter->hw_event.exclude_user && user_mode(regs))
1940 if (counter->hw_event.exclude_kernel && !user_mode(regs))
1946 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
1947 int nmi, struct pt_regs *regs)
1949 int neg = atomic64_add_negative(nr, &counter->hw.count);
1950 if (counter->hw.irq_period && !neg)
1951 perf_swcounter_overflow(counter, nmi, regs);
1954 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
1955 enum perf_event_types type, u32 event,
1956 u64 nr, int nmi, struct pt_regs *regs)
1958 struct perf_counter *counter;
1960 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
1964 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
1965 if (perf_swcounter_match(counter, type, event, regs))
1966 perf_swcounter_add(counter, nr, nmi, regs);
1971 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
1974 return &cpuctx->recursion[3];
1977 return &cpuctx->recursion[2];
1980 return &cpuctx->recursion[1];
1982 return &cpuctx->recursion[0];
1985 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
1986 u64 nr, int nmi, struct pt_regs *regs)
1988 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
1989 int *recursion = perf_swcounter_recursion_context(cpuctx);
1997 perf_swcounter_ctx_event(&cpuctx->ctx, type, event, nr, nmi, regs);
1998 if (cpuctx->task_ctx) {
1999 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2007 put_cpu_var(perf_cpu_context);
2010 void perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)
2012 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs);
2015 static void perf_swcounter_read(struct perf_counter *counter)
2017 perf_swcounter_update(counter);
2020 static int perf_swcounter_enable(struct perf_counter *counter)
2022 perf_swcounter_set_period(counter);
2026 static void perf_swcounter_disable(struct perf_counter *counter)
2028 perf_swcounter_update(counter);
2031 static const struct hw_perf_counter_ops perf_ops_generic = {
2032 .enable = perf_swcounter_enable,
2033 .disable = perf_swcounter_disable,
2034 .read = perf_swcounter_read,
2038 * Software counter: cpu wall time clock
2041 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2043 int cpu = raw_smp_processor_id();
2047 now = cpu_clock(cpu);
2048 prev = atomic64_read(&counter->hw.prev_count);
2049 atomic64_set(&counter->hw.prev_count, now);
2050 atomic64_add(now - prev, &counter->count);
2053 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2055 struct hw_perf_counter *hwc = &counter->hw;
2056 int cpu = raw_smp_processor_id();
2058 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2059 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2060 hwc->hrtimer.function = perf_swcounter_hrtimer;
2061 if (hwc->irq_period) {
2062 __hrtimer_start_range_ns(&hwc->hrtimer,
2063 ns_to_ktime(hwc->irq_period), 0,
2064 HRTIMER_MODE_REL, 0);
2070 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2072 hrtimer_cancel(&counter->hw.hrtimer);
2073 cpu_clock_perf_counter_update(counter);
2076 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2078 cpu_clock_perf_counter_update(counter);
2081 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
2082 .enable = cpu_clock_perf_counter_enable,
2083 .disable = cpu_clock_perf_counter_disable,
2084 .read = cpu_clock_perf_counter_read,
2088 * Software counter: task time clock
2092 * Called from within the scheduler:
2094 static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
2096 struct task_struct *curr = counter->task;
2099 delta = __task_delta_exec(curr, update);
2101 return curr->se.sum_exec_runtime + delta;
2104 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2109 prev = atomic64_read(&counter->hw.prev_count);
2111 atomic64_set(&counter->hw.prev_count, now);
2115 atomic64_add(delta, &counter->count);
2118 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2120 struct hw_perf_counter *hwc = &counter->hw;
2122 atomic64_set(&hwc->prev_count, task_clock_perf_counter_val(counter, 0));
2123 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2124 hwc->hrtimer.function = perf_swcounter_hrtimer;
2125 if (hwc->irq_period) {
2126 __hrtimer_start_range_ns(&hwc->hrtimer,
2127 ns_to_ktime(hwc->irq_period), 0,
2128 HRTIMER_MODE_REL, 0);
2134 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2136 hrtimer_cancel(&counter->hw.hrtimer);
2137 task_clock_perf_counter_update(counter,
2138 task_clock_perf_counter_val(counter, 0));
2141 static void task_clock_perf_counter_read(struct perf_counter *counter)
2143 task_clock_perf_counter_update(counter,
2144 task_clock_perf_counter_val(counter, 1));
2147 static const struct hw_perf_counter_ops perf_ops_task_clock = {
2148 .enable = task_clock_perf_counter_enable,
2149 .disable = task_clock_perf_counter_disable,
2150 .read = task_clock_perf_counter_read,
2154 * Software counter: cpu migrations
2157 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2159 struct task_struct *curr = counter->ctx->task;
2162 return curr->se.nr_migrations;
2163 return cpu_nr_migrations(smp_processor_id());
2166 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2171 prev = atomic64_read(&counter->hw.prev_count);
2172 now = get_cpu_migrations(counter);
2174 atomic64_set(&counter->hw.prev_count, now);
2178 atomic64_add(delta, &counter->count);
2181 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2183 cpu_migrations_perf_counter_update(counter);
2186 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2188 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2189 atomic64_set(&counter->hw.prev_count,
2190 get_cpu_migrations(counter));
2194 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2196 cpu_migrations_perf_counter_update(counter);
2199 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
2200 .enable = cpu_migrations_perf_counter_enable,
2201 .disable = cpu_migrations_perf_counter_disable,
2202 .read = cpu_migrations_perf_counter_read,
2205 #ifdef CONFIG_EVENT_PROFILE
2206 void perf_tpcounter_event(int event_id)
2208 struct pt_regs *regs = get_irq_regs();
2211 regs = task_pt_regs(current);
2213 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs);
2216 extern int ftrace_profile_enable(int);
2217 extern void ftrace_profile_disable(int);
2219 static void tp_perf_counter_destroy(struct perf_counter *counter)
2221 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2224 static const struct hw_perf_counter_ops *
2225 tp_perf_counter_init(struct perf_counter *counter)
2227 int event_id = perf_event_id(&counter->hw_event);
2230 ret = ftrace_profile_enable(event_id);
2234 counter->destroy = tp_perf_counter_destroy;
2235 counter->hw.irq_period = counter->hw_event.irq_period;
2237 return &perf_ops_generic;
2240 static const struct hw_perf_counter_ops *
2241 tp_perf_counter_init(struct perf_counter *counter)
2247 static const struct hw_perf_counter_ops *
2248 sw_perf_counter_init(struct perf_counter *counter)
2250 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2251 const struct hw_perf_counter_ops *hw_ops = NULL;
2252 struct hw_perf_counter *hwc = &counter->hw;
2255 * Software counters (currently) can't in general distinguish
2256 * between user, kernel and hypervisor events.
2257 * However, context switches and cpu migrations are considered
2258 * to be kernel events, and page faults are never hypervisor
2261 switch (perf_event_id(&counter->hw_event)) {
2262 case PERF_COUNT_CPU_CLOCK:
2263 hw_ops = &perf_ops_cpu_clock;
2265 if (hw_event->irq_period && hw_event->irq_period < 10000)
2266 hw_event->irq_period = 10000;
2268 case PERF_COUNT_TASK_CLOCK:
2270 * If the user instantiates this as a per-cpu counter,
2271 * use the cpu_clock counter instead.
2273 if (counter->ctx->task)
2274 hw_ops = &perf_ops_task_clock;
2276 hw_ops = &perf_ops_cpu_clock;
2278 if (hw_event->irq_period && hw_event->irq_period < 10000)
2279 hw_event->irq_period = 10000;
2281 case PERF_COUNT_PAGE_FAULTS:
2282 case PERF_COUNT_PAGE_FAULTS_MIN:
2283 case PERF_COUNT_PAGE_FAULTS_MAJ:
2284 case PERF_COUNT_CONTEXT_SWITCHES:
2285 hw_ops = &perf_ops_generic;
2287 case PERF_COUNT_CPU_MIGRATIONS:
2288 if (!counter->hw_event.exclude_kernel)
2289 hw_ops = &perf_ops_cpu_migrations;
2294 hwc->irq_period = hw_event->irq_period;
2300 * Allocate and initialize a counter structure
2302 static struct perf_counter *
2303 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2305 struct perf_counter_context *ctx,
2306 struct perf_counter *group_leader,
2309 const struct hw_perf_counter_ops *hw_ops;
2310 struct perf_counter *counter;
2312 counter = kzalloc(sizeof(*counter), gfpflags);
2317 * Single counters are their own group leaders, with an
2318 * empty sibling list:
2321 group_leader = counter;
2323 mutex_init(&counter->mutex);
2324 INIT_LIST_HEAD(&counter->list_entry);
2325 INIT_LIST_HEAD(&counter->event_entry);
2326 INIT_LIST_HEAD(&counter->sibling_list);
2327 init_waitqueue_head(&counter->waitq);
2329 mutex_init(&counter->mmap_mutex);
2331 INIT_LIST_HEAD(&counter->child_list);
2334 counter->hw_event = *hw_event;
2335 counter->group_leader = group_leader;
2336 counter->hw_ops = NULL;
2339 counter->state = PERF_COUNTER_STATE_INACTIVE;
2340 if (hw_event->disabled)
2341 counter->state = PERF_COUNTER_STATE_OFF;
2345 if (perf_event_raw(hw_event)) {
2346 hw_ops = hw_perf_counter_init(counter);
2350 switch (perf_event_type(hw_event)) {
2351 case PERF_TYPE_HARDWARE:
2352 hw_ops = hw_perf_counter_init(counter);
2355 case PERF_TYPE_SOFTWARE:
2356 hw_ops = sw_perf_counter_init(counter);
2359 case PERF_TYPE_TRACEPOINT:
2360 hw_ops = tp_perf_counter_init(counter);
2369 counter->hw_ops = hw_ops;
2375 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2377 * @hw_event_uptr: event type attributes for monitoring/sampling
2380 * @group_fd: group leader counter fd
2382 SYSCALL_DEFINE5(perf_counter_open,
2383 const struct perf_counter_hw_event __user *, hw_event_uptr,
2384 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2386 struct perf_counter *counter, *group_leader;
2387 struct perf_counter_hw_event hw_event;
2388 struct perf_counter_context *ctx;
2389 struct file *counter_file = NULL;
2390 struct file *group_file = NULL;
2391 int fput_needed = 0;
2392 int fput_needed2 = 0;
2395 /* for future expandability... */
2399 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2403 * Get the target context (task or percpu):
2405 ctx = find_get_context(pid, cpu);
2407 return PTR_ERR(ctx);
2410 * Look up the group leader (we will attach this counter to it):
2412 group_leader = NULL;
2413 if (group_fd != -1) {
2415 group_file = fget_light(group_fd, &fput_needed);
2417 goto err_put_context;
2418 if (group_file->f_op != &perf_fops)
2419 goto err_put_context;
2421 group_leader = group_file->private_data;
2423 * Do not allow a recursive hierarchy (this new sibling
2424 * becoming part of another group-sibling):
2426 if (group_leader->group_leader != group_leader)
2427 goto err_put_context;
2429 * Do not allow to attach to a group in a different
2430 * task or CPU context:
2432 if (group_leader->ctx != ctx)
2433 goto err_put_context;
2435 * Only a group leader can be exclusive or pinned
2437 if (hw_event.exclusive || hw_event.pinned)
2438 goto err_put_context;
2442 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2445 goto err_put_context;
2447 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2449 goto err_free_put_context;
2451 counter_file = fget_light(ret, &fput_needed2);
2453 goto err_free_put_context;
2455 counter->filp = counter_file;
2456 mutex_lock(&ctx->mutex);
2457 perf_install_in_context(ctx, counter, cpu);
2458 mutex_unlock(&ctx->mutex);
2460 fput_light(counter_file, fput_needed2);
2463 fput_light(group_file, fput_needed);
2467 err_free_put_context:
2477 * Initialize the perf_counter context in a task_struct:
2480 __perf_counter_init_context(struct perf_counter_context *ctx,
2481 struct task_struct *task)
2483 memset(ctx, 0, sizeof(*ctx));
2484 spin_lock_init(&ctx->lock);
2485 mutex_init(&ctx->mutex);
2486 INIT_LIST_HEAD(&ctx->counter_list);
2487 INIT_LIST_HEAD(&ctx->event_list);
2492 * inherit a counter from parent task to child task:
2494 static struct perf_counter *
2495 inherit_counter(struct perf_counter *parent_counter,
2496 struct task_struct *parent,
2497 struct perf_counter_context *parent_ctx,
2498 struct task_struct *child,
2499 struct perf_counter *group_leader,
2500 struct perf_counter_context *child_ctx)
2502 struct perf_counter *child_counter;
2505 * Instead of creating recursive hierarchies of counters,
2506 * we link inherited counters back to the original parent,
2507 * which has a filp for sure, which we use as the reference
2510 if (parent_counter->parent)
2511 parent_counter = parent_counter->parent;
2513 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2514 parent_counter->cpu, child_ctx,
2515 group_leader, GFP_KERNEL);
2520 * Link it up in the child's context:
2522 child_counter->task = child;
2523 add_counter_to_ctx(child_counter, child_ctx);
2525 child_counter->parent = parent_counter;
2527 * inherit into child's child as well:
2529 child_counter->hw_event.inherit = 1;
2532 * Get a reference to the parent filp - we will fput it
2533 * when the child counter exits. This is safe to do because
2534 * we are in the parent and we know that the filp still
2535 * exists and has a nonzero count:
2537 atomic_long_inc(&parent_counter->filp->f_count);
2540 * Link this into the parent counter's child list
2542 mutex_lock(&parent_counter->mutex);
2543 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2546 * Make the child state follow the state of the parent counter,
2547 * not its hw_event.disabled bit. We hold the parent's mutex,
2548 * so we won't race with perf_counter_{en,dis}able_family.
2550 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2551 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2553 child_counter->state = PERF_COUNTER_STATE_OFF;
2555 mutex_unlock(&parent_counter->mutex);
2557 return child_counter;
2560 static int inherit_group(struct perf_counter *parent_counter,
2561 struct task_struct *parent,
2562 struct perf_counter_context *parent_ctx,
2563 struct task_struct *child,
2564 struct perf_counter_context *child_ctx)
2566 struct perf_counter *leader;
2567 struct perf_counter *sub;
2569 leader = inherit_counter(parent_counter, parent, parent_ctx,
2570 child, NULL, child_ctx);
2573 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2574 if (!inherit_counter(sub, parent, parent_ctx,
2575 child, leader, child_ctx))
2581 static void sync_child_counter(struct perf_counter *child_counter,
2582 struct perf_counter *parent_counter)
2584 u64 parent_val, child_val;
2586 parent_val = atomic64_read(&parent_counter->count);
2587 child_val = atomic64_read(&child_counter->count);
2590 * Add back the child's count to the parent's count:
2592 atomic64_add(child_val, &parent_counter->count);
2593 atomic64_add(child_counter->total_time_enabled,
2594 &parent_counter->child_total_time_enabled);
2595 atomic64_add(child_counter->total_time_running,
2596 &parent_counter->child_total_time_running);
2599 * Remove this counter from the parent's list
2601 mutex_lock(&parent_counter->mutex);
2602 list_del_init(&child_counter->child_list);
2603 mutex_unlock(&parent_counter->mutex);
2606 * Release the parent counter, if this was the last
2609 fput(parent_counter->filp);
2613 __perf_counter_exit_task(struct task_struct *child,
2614 struct perf_counter *child_counter,
2615 struct perf_counter_context *child_ctx)
2617 struct perf_counter *parent_counter;
2618 struct perf_counter *sub, *tmp;
2621 * If we do not self-reap then we have to wait for the
2622 * child task to unschedule (it will happen for sure),
2623 * so that its counter is at its final count. (This
2624 * condition triggers rarely - child tasks usually get
2625 * off their CPU before the parent has a chance to
2626 * get this far into the reaping action)
2628 if (child != current) {
2629 wait_task_inactive(child, 0);
2630 list_del_init(&child_counter->list_entry);
2631 update_counter_times(child_counter);
2633 struct perf_cpu_context *cpuctx;
2634 unsigned long flags;
2638 * Disable and unlink this counter.
2640 * Be careful about zapping the list - IRQ/NMI context
2641 * could still be processing it:
2643 curr_rq_lock_irq_save(&flags);
2644 perf_flags = hw_perf_save_disable();
2646 cpuctx = &__get_cpu_var(perf_cpu_context);
2648 group_sched_out(child_counter, cpuctx, child_ctx);
2649 update_counter_times(child_counter);
2651 list_del_init(&child_counter->list_entry);
2653 child_ctx->nr_counters--;
2655 hw_perf_restore(perf_flags);
2656 curr_rq_unlock_irq_restore(&flags);
2659 parent_counter = child_counter->parent;
2661 * It can happen that parent exits first, and has counters
2662 * that are still around due to the child reference. These
2663 * counters need to be zapped - but otherwise linger.
2665 if (parent_counter) {
2666 sync_child_counter(child_counter, parent_counter);
2667 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
2670 sync_child_counter(sub, sub->parent);
2674 free_counter(child_counter);
2679 * When a child task exits, feed back counter values to parent counters.
2681 * Note: we may be running in child context, but the PID is not hashed
2682 * anymore so new counters will not be added.
2684 void perf_counter_exit_task(struct task_struct *child)
2686 struct perf_counter *child_counter, *tmp;
2687 struct perf_counter_context *child_ctx;
2689 child_ctx = &child->perf_counter_ctx;
2691 if (likely(!child_ctx->nr_counters))
2694 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
2696 __perf_counter_exit_task(child, child_counter, child_ctx);
2700 * Initialize the perf_counter context in task_struct
2702 void perf_counter_init_task(struct task_struct *child)
2704 struct perf_counter_context *child_ctx, *parent_ctx;
2705 struct perf_counter *counter;
2706 struct task_struct *parent = current;
2708 child_ctx = &child->perf_counter_ctx;
2709 parent_ctx = &parent->perf_counter_ctx;
2711 __perf_counter_init_context(child_ctx, child);
2714 * This is executed from the parent task context, so inherit
2715 * counters that have been marked for cloning:
2718 if (likely(!parent_ctx->nr_counters))
2722 * Lock the parent list. No need to lock the child - not PID
2723 * hashed yet and not running, so nobody can access it.
2725 mutex_lock(&parent_ctx->mutex);
2728 * We dont have to disable NMIs - we are only looking at
2729 * the list, not manipulating it:
2731 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2732 if (!counter->hw_event.inherit)
2735 if (inherit_group(counter, parent,
2736 parent_ctx, child, child_ctx))
2740 mutex_unlock(&parent_ctx->mutex);
2743 static void __cpuinit perf_counter_init_cpu(int cpu)
2745 struct perf_cpu_context *cpuctx;
2747 cpuctx = &per_cpu(perf_cpu_context, cpu);
2748 __perf_counter_init_context(&cpuctx->ctx, NULL);
2750 mutex_lock(&perf_resource_mutex);
2751 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2752 mutex_unlock(&perf_resource_mutex);
2754 hw_perf_counter_setup(cpu);
2757 #ifdef CONFIG_HOTPLUG_CPU
2758 static void __perf_counter_exit_cpu(void *info)
2760 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2761 struct perf_counter_context *ctx = &cpuctx->ctx;
2762 struct perf_counter *counter, *tmp;
2764 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2765 __perf_counter_remove_from_context(counter);
2767 static void perf_counter_exit_cpu(int cpu)
2769 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2770 struct perf_counter_context *ctx = &cpuctx->ctx;
2772 mutex_lock(&ctx->mutex);
2773 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2774 mutex_unlock(&ctx->mutex);
2777 static inline void perf_counter_exit_cpu(int cpu) { }
2780 static int __cpuinit
2781 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2783 unsigned int cpu = (long)hcpu;
2787 case CPU_UP_PREPARE:
2788 case CPU_UP_PREPARE_FROZEN:
2789 perf_counter_init_cpu(cpu);
2792 case CPU_DOWN_PREPARE:
2793 case CPU_DOWN_PREPARE_FROZEN:
2794 perf_counter_exit_cpu(cpu);
2804 static struct notifier_block __cpuinitdata perf_cpu_nb = {
2805 .notifier_call = perf_cpu_notify,
2808 static int __init perf_counter_init(void)
2810 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2811 (void *)(long)smp_processor_id());
2812 register_cpu_notifier(&perf_cpu_nb);
2816 early_initcall(perf_counter_init);
2818 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2820 return sprintf(buf, "%d\n", perf_reserved_percpu);
2824 perf_set_reserve_percpu(struct sysdev_class *class,
2828 struct perf_cpu_context *cpuctx;
2832 err = strict_strtoul(buf, 10, &val);
2835 if (val > perf_max_counters)
2838 mutex_lock(&perf_resource_mutex);
2839 perf_reserved_percpu = val;
2840 for_each_online_cpu(cpu) {
2841 cpuctx = &per_cpu(perf_cpu_context, cpu);
2842 spin_lock_irq(&cpuctx->ctx.lock);
2843 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2844 perf_max_counters - perf_reserved_percpu);
2845 cpuctx->max_pertask = mpt;
2846 spin_unlock_irq(&cpuctx->ctx.lock);
2848 mutex_unlock(&perf_resource_mutex);
2853 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2855 return sprintf(buf, "%d\n", perf_overcommit);
2859 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2864 err = strict_strtoul(buf, 10, &val);
2870 mutex_lock(&perf_resource_mutex);
2871 perf_overcommit = val;
2872 mutex_unlock(&perf_resource_mutex);
2877 static SYSDEV_CLASS_ATTR(
2880 perf_show_reserve_percpu,
2881 perf_set_reserve_percpu
2884 static SYSDEV_CLASS_ATTR(
2887 perf_show_overcommit,
2891 static struct attribute *perfclass_attrs[] = {
2892 &attr_reserve_percpu.attr,
2893 &attr_overcommit.attr,
2897 static struct attribute_group perfclass_attr_group = {
2898 .attrs = perfclass_attrs,
2899 .name = "perf_counters",
2902 static int __init perf_counter_sysfs_init(void)
2904 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2905 &perfclass_attr_group);
2907 device_initcall(perf_counter_sysfs_init);