2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/ptrace.h>
20 #include <linux/percpu.h>
21 #include <linux/vmstat.h>
22 #include <linux/hardirq.h>
23 #include <linux/rculist.h>
24 #include <linux/uaccess.h>
25 #include <linux/syscalls.h>
26 #include <linux/anon_inodes.h>
27 #include <linux/kernel_stat.h>
28 #include <linux/perf_counter.h>
29 #include <linux/dcache.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_mmap_tracking __read_mostly;
43 static atomic_t nr_munmap_tracking __read_mostly;
44 static atomic_t nr_comm_tracking __read_mostly;
46 int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
49 * Mutex for (sysadmin-configurable) counter reservations:
51 static DEFINE_MUTEX(perf_resource_mutex);
54 * Architecture provided APIs - weak aliases:
56 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
61 u64 __weak hw_perf_save_disable(void) { return 0; }
62 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
63 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
64 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
65 struct perf_cpu_context *cpuctx,
66 struct perf_counter_context *ctx, int cpu)
71 void __weak perf_counter_print_debug(void) { }
74 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
76 struct perf_counter *group_leader = counter->group_leader;
79 * Depending on whether it is a standalone or sibling counter,
80 * add it straight to the context's counter list, or to the group
81 * leader's sibling list:
83 if (counter->group_leader == counter)
84 list_add_tail(&counter->list_entry, &ctx->counter_list);
86 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
87 group_leader->nr_siblings++;
90 list_add_rcu(&counter->event_entry, &ctx->event_list);
94 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
96 struct perf_counter *sibling, *tmp;
98 list_del_init(&counter->list_entry);
99 list_del_rcu(&counter->event_entry);
101 if (counter->group_leader != counter)
102 counter->group_leader->nr_siblings--;
105 * If this was a group counter with sibling counters then
106 * upgrade the siblings to singleton counters by adding them
107 * to the context list directly:
109 list_for_each_entry_safe(sibling, tmp,
110 &counter->sibling_list, list_entry) {
112 list_move_tail(&sibling->list_entry, &ctx->counter_list);
113 sibling->group_leader = sibling;
118 counter_sched_out(struct perf_counter *counter,
119 struct perf_cpu_context *cpuctx,
120 struct perf_counter_context *ctx)
122 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
125 counter->state = PERF_COUNTER_STATE_INACTIVE;
126 counter->tstamp_stopped = ctx->time;
127 counter->pmu->disable(counter);
130 if (!is_software_counter(counter))
131 cpuctx->active_oncpu--;
133 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
134 cpuctx->exclusive = 0;
138 group_sched_out(struct perf_counter *group_counter,
139 struct perf_cpu_context *cpuctx,
140 struct perf_counter_context *ctx)
142 struct perf_counter *counter;
144 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
147 counter_sched_out(group_counter, cpuctx, ctx);
150 * Schedule out siblings (if any):
152 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
153 counter_sched_out(counter, cpuctx, ctx);
155 if (group_counter->hw_event.exclusive)
156 cpuctx->exclusive = 0;
160 * Cross CPU call to remove a performance counter
162 * We disable the counter on the hardware level first. After that we
163 * remove it from the context list.
165 static void __perf_counter_remove_from_context(void *info)
167 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
168 struct perf_counter *counter = info;
169 struct perf_counter_context *ctx = counter->ctx;
174 * If this is a task context, we need to check whether it is
175 * the current task context of this cpu. If not it has been
176 * scheduled out before the smp call arrived.
178 if (ctx->task && cpuctx->task_ctx != ctx)
181 spin_lock_irqsave(&ctx->lock, flags);
183 counter_sched_out(counter, cpuctx, ctx);
185 counter->task = NULL;
189 * Protect the list operation against NMI by disabling the
190 * counters on a global level. NOP for non NMI based counters.
192 perf_flags = hw_perf_save_disable();
193 list_del_counter(counter, ctx);
194 hw_perf_restore(perf_flags);
198 * Allow more per task counters with respect to the
201 cpuctx->max_pertask =
202 min(perf_max_counters - ctx->nr_counters,
203 perf_max_counters - perf_reserved_percpu);
206 spin_unlock_irqrestore(&ctx->lock, flags);
211 * Remove the counter from a task's (or a CPU's) list of counters.
213 * Must be called with counter->mutex and ctx->mutex held.
215 * CPU counters are removed with a smp call. For task counters we only
216 * call when the task is on a CPU.
218 static void perf_counter_remove_from_context(struct perf_counter *counter)
220 struct perf_counter_context *ctx = counter->ctx;
221 struct task_struct *task = ctx->task;
225 * Per cpu counters are removed via an smp call and
226 * the removal is always sucessful.
228 smp_call_function_single(counter->cpu,
229 __perf_counter_remove_from_context,
235 task_oncpu_function_call(task, __perf_counter_remove_from_context,
238 spin_lock_irq(&ctx->lock);
240 * If the context is active we need to retry the smp call.
242 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
243 spin_unlock_irq(&ctx->lock);
248 * The lock prevents that this context is scheduled in so we
249 * can remove the counter safely, if the call above did not
252 if (!list_empty(&counter->list_entry)) {
254 list_del_counter(counter, ctx);
255 counter->task = NULL;
257 spin_unlock_irq(&ctx->lock);
260 static inline u64 perf_clock(void)
262 return cpu_clock(smp_processor_id());
266 * Update the record of the current time in a context.
268 static void update_context_time(struct perf_counter_context *ctx)
270 u64 now = perf_clock();
272 ctx->time += now - ctx->timestamp;
273 ctx->timestamp = now;
277 * Update the total_time_enabled and total_time_running fields for a counter.
279 static void update_counter_times(struct perf_counter *counter)
281 struct perf_counter_context *ctx = counter->ctx;
284 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
287 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
289 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
290 run_end = counter->tstamp_stopped;
294 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 spin_lock_irqsave(&ctx->lock, flags);
329 * If the counter is on, turn it off.
330 * If it is in error state, leave it in error state.
332 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
333 update_context_time(ctx);
334 update_counter_times(counter);
335 if (counter == counter->group_leader)
336 group_sched_out(counter, cpuctx, ctx);
338 counter_sched_out(counter, cpuctx, ctx);
339 counter->state = PERF_COUNTER_STATE_OFF;
342 spin_unlock_irqrestore(&ctx->lock, flags);
348 static void perf_counter_disable(struct perf_counter *counter)
350 struct perf_counter_context *ctx = counter->ctx;
351 struct task_struct *task = ctx->task;
355 * Disable the counter on the cpu that it's on
357 smp_call_function_single(counter->cpu, __perf_counter_disable,
363 task_oncpu_function_call(task, __perf_counter_disable, counter);
365 spin_lock_irq(&ctx->lock);
367 * If the counter is still active, we need to retry the cross-call.
369 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
370 spin_unlock_irq(&ctx->lock);
375 * Since we have the lock this context can't be scheduled
376 * in, so we can change the state safely.
378 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
379 update_counter_times(counter);
380 counter->state = PERF_COUNTER_STATE_OFF;
383 spin_unlock_irq(&ctx->lock);
387 * Disable a counter and all its children.
389 static void perf_counter_disable_family(struct perf_counter *counter)
391 struct perf_counter *child;
393 perf_counter_disable(counter);
396 * Lock the mutex to protect the list of children
398 mutex_lock(&counter->mutex);
399 list_for_each_entry(child, &counter->child_list, child_list)
400 perf_counter_disable(child);
401 mutex_unlock(&counter->mutex);
405 counter_sched_in(struct perf_counter *counter,
406 struct perf_cpu_context *cpuctx,
407 struct perf_counter_context *ctx,
410 if (counter->state <= PERF_COUNTER_STATE_OFF)
413 counter->state = PERF_COUNTER_STATE_ACTIVE;
414 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
416 * The new state must be visible before we turn it on in the hardware:
420 if (counter->pmu->enable(counter)) {
421 counter->state = PERF_COUNTER_STATE_INACTIVE;
426 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
428 if (!is_software_counter(counter))
429 cpuctx->active_oncpu++;
432 if (counter->hw_event.exclusive)
433 cpuctx->exclusive = 1;
439 * Return 1 for a group consisting entirely of software counters,
440 * 0 if the group contains any hardware counters.
442 static int is_software_only_group(struct perf_counter *leader)
444 struct perf_counter *counter;
446 if (!is_software_counter(leader))
449 list_for_each_entry(counter, &leader->sibling_list, list_entry)
450 if (!is_software_counter(counter))
457 * Work out whether we can put this counter group on the CPU now.
459 static int group_can_go_on(struct perf_counter *counter,
460 struct perf_cpu_context *cpuctx,
464 * Groups consisting entirely of software counters can always go on.
466 if (is_software_only_group(counter))
469 * If an exclusive group is already on, no other hardware
470 * counters can go on.
472 if (cpuctx->exclusive)
475 * If this group is exclusive and there are already
476 * counters on the CPU, it can't go on.
478 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
481 * Otherwise, try to add it if all previous groups were able
487 static void add_counter_to_ctx(struct perf_counter *counter,
488 struct perf_counter_context *ctx)
490 list_add_counter(counter, ctx);
492 counter->prev_state = PERF_COUNTER_STATE_OFF;
493 counter->tstamp_enabled = ctx->time;
494 counter->tstamp_running = ctx->time;
495 counter->tstamp_stopped = ctx->time;
499 * Cross CPU call to install and enable a performance counter
501 static void __perf_install_in_context(void *info)
503 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
504 struct perf_counter *counter = info;
505 struct perf_counter_context *ctx = counter->ctx;
506 struct perf_counter *leader = counter->group_leader;
507 int cpu = smp_processor_id();
513 * If this is a task context, we need to check whether it is
514 * the current task context of this cpu. If not it has been
515 * scheduled out before the smp call arrived.
517 if (ctx->task && cpuctx->task_ctx != ctx)
520 spin_lock_irqsave(&ctx->lock, flags);
521 update_context_time(ctx);
524 * Protect the list operation against NMI by disabling the
525 * counters on a global level. NOP for non NMI based counters.
527 perf_flags = hw_perf_save_disable();
529 add_counter_to_ctx(counter, ctx);
532 * Don't put the counter on if it is disabled or if
533 * it is in a group and the group isn't on.
535 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
536 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
540 * An exclusive counter can't go on if there are already active
541 * hardware counters, and no hardware counter can go on if there
542 * is already an exclusive counter on.
544 if (!group_can_go_on(counter, cpuctx, 1))
547 err = counter_sched_in(counter, cpuctx, ctx, cpu);
551 * This counter couldn't go on. If it is in a group
552 * then we have to pull the whole group off.
553 * If the counter group is pinned then put it in error state.
555 if (leader != counter)
556 group_sched_out(leader, cpuctx, ctx);
557 if (leader->hw_event.pinned) {
558 update_group_times(leader);
559 leader->state = PERF_COUNTER_STATE_ERROR;
563 if (!err && !ctx->task && cpuctx->max_pertask)
564 cpuctx->max_pertask--;
567 hw_perf_restore(perf_flags);
569 spin_unlock_irqrestore(&ctx->lock, flags);
573 * Attach a performance counter to a context
575 * First we add the counter to the list with the hardware enable bit
576 * in counter->hw_config cleared.
578 * If the counter is attached to a task which is on a CPU we use a smp
579 * call to enable it in the task context. The task might have been
580 * scheduled away, but we check this in the smp call again.
582 * Must be called with ctx->mutex held.
585 perf_install_in_context(struct perf_counter_context *ctx,
586 struct perf_counter *counter,
589 struct task_struct *task = ctx->task;
593 * Per cpu counters are installed via an smp call and
594 * the install is always sucessful.
596 smp_call_function_single(cpu, __perf_install_in_context,
601 counter->task = task;
603 task_oncpu_function_call(task, __perf_install_in_context,
606 spin_lock_irq(&ctx->lock);
608 * we need to retry the smp call.
610 if (ctx->is_active && list_empty(&counter->list_entry)) {
611 spin_unlock_irq(&ctx->lock);
616 * The lock prevents that this context is scheduled in so we
617 * can add the counter safely, if it the call above did not
620 if (list_empty(&counter->list_entry))
621 add_counter_to_ctx(counter, ctx);
622 spin_unlock_irq(&ctx->lock);
626 * Cross CPU call to enable a performance counter
628 static void __perf_counter_enable(void *info)
630 struct perf_counter *counter = info;
631 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
632 struct perf_counter_context *ctx = counter->ctx;
633 struct perf_counter *leader = counter->group_leader;
638 * If this is a per-task counter, need to check whether this
639 * counter's task is the current task on this cpu.
641 if (ctx->task && cpuctx->task_ctx != ctx)
644 spin_lock_irqsave(&ctx->lock, flags);
645 update_context_time(ctx);
647 counter->prev_state = counter->state;
648 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
650 counter->state = PERF_COUNTER_STATE_INACTIVE;
651 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
654 * If the counter is in a group and isn't the group leader,
655 * then don't put it on unless the group is on.
657 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
660 if (!group_can_go_on(counter, cpuctx, 1))
663 err = counter_sched_in(counter, cpuctx, ctx,
668 * If this counter can't go on and it's part of a
669 * group, then the whole group has to come off.
671 if (leader != counter)
672 group_sched_out(leader, cpuctx, ctx);
673 if (leader->hw_event.pinned) {
674 update_group_times(leader);
675 leader->state = PERF_COUNTER_STATE_ERROR;
680 spin_unlock_irqrestore(&ctx->lock, flags);
686 static void perf_counter_enable(struct perf_counter *counter)
688 struct perf_counter_context *ctx = counter->ctx;
689 struct task_struct *task = ctx->task;
693 * Enable the counter on the cpu that it's on
695 smp_call_function_single(counter->cpu, __perf_counter_enable,
700 spin_lock_irq(&ctx->lock);
701 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
705 * If the counter is in error state, clear that first.
706 * That way, if we see the counter in error state below, we
707 * know that it has gone back into error state, as distinct
708 * from the task having been scheduled away before the
709 * cross-call arrived.
711 if (counter->state == PERF_COUNTER_STATE_ERROR)
712 counter->state = PERF_COUNTER_STATE_OFF;
715 spin_unlock_irq(&ctx->lock);
716 task_oncpu_function_call(task, __perf_counter_enable, counter);
718 spin_lock_irq(&ctx->lock);
721 * If the context is active and the counter is still off,
722 * we need to retry the cross-call.
724 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
728 * Since we have the lock this context can't be scheduled
729 * in, so we can change the state safely.
731 if (counter->state == PERF_COUNTER_STATE_OFF) {
732 counter->state = PERF_COUNTER_STATE_INACTIVE;
733 counter->tstamp_enabled =
734 ctx->time - counter->total_time_enabled;
737 spin_unlock_irq(&ctx->lock);
740 static void perf_counter_refresh(struct perf_counter *counter, int refresh)
742 atomic_add(refresh, &counter->event_limit);
743 perf_counter_enable(counter);
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);
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 update_context_time(ctx);
808 regs = task_pt_regs(task);
809 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
810 __perf_counter_sched_out(ctx, cpuctx);
812 cpuctx->task_ctx = NULL;
815 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
817 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
821 group_sched_in(struct perf_counter *group_counter,
822 struct perf_cpu_context *cpuctx,
823 struct perf_counter_context *ctx,
826 struct perf_counter *counter, *partial_group;
829 if (group_counter->state == PERF_COUNTER_STATE_OFF)
832 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
834 return ret < 0 ? ret : 0;
836 group_counter->prev_state = group_counter->state;
837 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
841 * Schedule in siblings as one group (if any):
843 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
844 counter->prev_state = counter->state;
845 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
846 partial_group = counter;
855 * Groups can be scheduled in as one unit only, so undo any
856 * partial group before returning:
858 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
859 if (counter == partial_group)
861 counter_sched_out(counter, cpuctx, ctx);
863 counter_sched_out(group_counter, cpuctx, ctx);
869 __perf_counter_sched_in(struct perf_counter_context *ctx,
870 struct perf_cpu_context *cpuctx, int cpu)
872 struct perf_counter *counter;
876 spin_lock(&ctx->lock);
878 if (likely(!ctx->nr_counters))
881 ctx->timestamp = perf_clock();
883 flags = hw_perf_save_disable();
886 * First go through the list and put on any pinned groups
887 * in order to give them the best chance of going on.
889 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
890 if (counter->state <= PERF_COUNTER_STATE_OFF ||
891 !counter->hw_event.pinned)
893 if (counter->cpu != -1 && counter->cpu != cpu)
896 if (group_can_go_on(counter, cpuctx, 1))
897 group_sched_in(counter, cpuctx, ctx, cpu);
900 * If this pinned group hasn't been scheduled,
901 * put it in error state.
903 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
904 update_group_times(counter);
905 counter->state = PERF_COUNTER_STATE_ERROR;
909 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
911 * Ignore counters in OFF or ERROR state, and
912 * ignore pinned counters since we did them already.
914 if (counter->state <= PERF_COUNTER_STATE_OFF ||
915 counter->hw_event.pinned)
919 * Listen to the 'cpu' scheduling filter constraint
922 if (counter->cpu != -1 && counter->cpu != cpu)
925 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
926 if (group_sched_in(counter, cpuctx, ctx, cpu))
930 hw_perf_restore(flags);
932 spin_unlock(&ctx->lock);
936 * Called from scheduler to add the counters of the current task
937 * with interrupts disabled.
939 * We restore the counter value and then enable it.
941 * This does not protect us against NMI, but enable()
942 * sets the enabled bit in the control field of counter _before_
943 * accessing the counter control register. If a NMI hits, then it will
944 * keep the counter running.
946 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
948 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
949 struct perf_counter_context *ctx = &task->perf_counter_ctx;
951 __perf_counter_sched_in(ctx, cpuctx, cpu);
952 cpuctx->task_ctx = ctx;
955 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
957 struct perf_counter_context *ctx = &cpuctx->ctx;
959 __perf_counter_sched_in(ctx, cpuctx, cpu);
962 int perf_counter_task_disable(void)
964 struct task_struct *curr = current;
965 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
966 struct perf_counter *counter;
971 if (likely(!ctx->nr_counters))
974 local_irq_save(flags);
975 cpu = smp_processor_id();
977 perf_counter_task_sched_out(curr, cpu);
979 spin_lock(&ctx->lock);
982 * Disable all the counters:
984 perf_flags = hw_perf_save_disable();
986 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
987 if (counter->state != PERF_COUNTER_STATE_ERROR) {
988 update_group_times(counter);
989 counter->state = PERF_COUNTER_STATE_OFF;
993 hw_perf_restore(perf_flags);
995 spin_unlock_irqrestore(&ctx->lock, flags);
1000 int perf_counter_task_enable(void)
1002 struct task_struct *curr = current;
1003 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1004 struct perf_counter *counter;
1005 unsigned long flags;
1009 if (likely(!ctx->nr_counters))
1012 local_irq_save(flags);
1013 cpu = smp_processor_id();
1015 perf_counter_task_sched_out(curr, cpu);
1017 spin_lock(&ctx->lock);
1020 * Disable all the counters:
1022 perf_flags = hw_perf_save_disable();
1024 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1025 if (counter->state > PERF_COUNTER_STATE_OFF)
1027 counter->state = PERF_COUNTER_STATE_INACTIVE;
1028 counter->tstamp_enabled =
1029 ctx->time - counter->total_time_enabled;
1030 counter->hw_event.disabled = 0;
1032 hw_perf_restore(perf_flags);
1034 spin_unlock(&ctx->lock);
1036 perf_counter_task_sched_in(curr, cpu);
1038 local_irq_restore(flags);
1044 * Round-robin a context's counters:
1046 static void rotate_ctx(struct perf_counter_context *ctx)
1048 struct perf_counter *counter;
1051 if (!ctx->nr_counters)
1054 spin_lock(&ctx->lock);
1056 * Rotate the first entry last (works just fine for group counters too):
1058 perf_flags = hw_perf_save_disable();
1059 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1060 list_move_tail(&counter->list_entry, &ctx->counter_list);
1063 hw_perf_restore(perf_flags);
1065 spin_unlock(&ctx->lock);
1068 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1070 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1071 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1072 const int rotate_percpu = 0;
1075 perf_counter_cpu_sched_out(cpuctx);
1076 perf_counter_task_sched_out(curr, cpu);
1079 rotate_ctx(&cpuctx->ctx);
1083 perf_counter_cpu_sched_in(cpuctx, cpu);
1084 perf_counter_task_sched_in(curr, cpu);
1088 * Cross CPU call to read the hardware counter
1090 static void __read(void *info)
1092 struct perf_counter *counter = info;
1093 struct perf_counter_context *ctx = counter->ctx;
1094 unsigned long flags;
1096 local_irq_save(flags);
1098 update_context_time(ctx);
1099 counter->pmu->read(counter);
1100 update_counter_times(counter);
1101 local_irq_restore(flags);
1104 static u64 perf_counter_read(struct perf_counter *counter)
1107 * If counter is enabled and currently active on a CPU, update the
1108 * value in the counter structure:
1110 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1111 smp_call_function_single(counter->oncpu,
1112 __read, counter, 1);
1113 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1114 update_counter_times(counter);
1117 return atomic64_read(&counter->count);
1120 static void put_context(struct perf_counter_context *ctx)
1123 put_task_struct(ctx->task);
1126 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1128 struct perf_cpu_context *cpuctx;
1129 struct perf_counter_context *ctx;
1130 struct task_struct *task;
1133 * If cpu is not a wildcard then this is a percpu counter:
1136 /* Must be root to operate on a CPU counter: */
1137 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1138 return ERR_PTR(-EACCES);
1140 if (cpu < 0 || cpu > num_possible_cpus())
1141 return ERR_PTR(-EINVAL);
1144 * We could be clever and allow to attach a counter to an
1145 * offline CPU and activate it when the CPU comes up, but
1148 if (!cpu_isset(cpu, cpu_online_map))
1149 return ERR_PTR(-ENODEV);
1151 cpuctx = &per_cpu(perf_cpu_context, cpu);
1161 task = find_task_by_vpid(pid);
1163 get_task_struct(task);
1167 return ERR_PTR(-ESRCH);
1169 ctx = &task->perf_counter_ctx;
1172 /* Reuse ptrace permission checks for now. */
1173 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1175 return ERR_PTR(-EACCES);
1181 static void free_counter_rcu(struct rcu_head *head)
1183 struct perf_counter *counter;
1185 counter = container_of(head, struct perf_counter, rcu_head);
1189 static void perf_pending_sync(struct perf_counter *counter);
1191 static void free_counter(struct perf_counter *counter)
1193 perf_pending_sync(counter);
1195 if (counter->hw_event.mmap)
1196 atomic_dec(&nr_mmap_tracking);
1197 if (counter->hw_event.munmap)
1198 atomic_dec(&nr_munmap_tracking);
1199 if (counter->hw_event.comm)
1200 atomic_dec(&nr_comm_tracking);
1202 if (counter->destroy)
1203 counter->destroy(counter);
1205 call_rcu(&counter->rcu_head, free_counter_rcu);
1209 * Called when the last reference to the file is gone.
1211 static int perf_release(struct inode *inode, struct file *file)
1213 struct perf_counter *counter = file->private_data;
1214 struct perf_counter_context *ctx = counter->ctx;
1216 file->private_data = NULL;
1218 mutex_lock(&ctx->mutex);
1219 mutex_lock(&counter->mutex);
1221 perf_counter_remove_from_context(counter);
1223 mutex_unlock(&counter->mutex);
1224 mutex_unlock(&ctx->mutex);
1226 free_counter(counter);
1233 * Read the performance counter - simple non blocking version for now
1236 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1242 * Return end-of-file for a read on a counter that is in
1243 * error state (i.e. because it was pinned but it couldn't be
1244 * scheduled on to the CPU at some point).
1246 if (counter->state == PERF_COUNTER_STATE_ERROR)
1249 mutex_lock(&counter->mutex);
1250 values[0] = perf_counter_read(counter);
1252 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1253 values[n++] = counter->total_time_enabled +
1254 atomic64_read(&counter->child_total_time_enabled);
1255 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1256 values[n++] = counter->total_time_running +
1257 atomic64_read(&counter->child_total_time_running);
1258 mutex_unlock(&counter->mutex);
1260 if (count < n * sizeof(u64))
1262 count = n * sizeof(u64);
1264 if (copy_to_user(buf, values, count))
1271 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1273 struct perf_counter *counter = file->private_data;
1275 return perf_read_hw(counter, buf, count);
1278 static unsigned int perf_poll(struct file *file, poll_table *wait)
1280 struct perf_counter *counter = file->private_data;
1281 struct perf_mmap_data *data;
1282 unsigned int events;
1285 data = rcu_dereference(counter->data);
1287 events = atomic_xchg(&data->wakeup, 0);
1292 poll_wait(file, &counter->waitq, wait);
1297 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1299 struct perf_counter *counter = file->private_data;
1303 case PERF_COUNTER_IOC_ENABLE:
1304 perf_counter_enable_family(counter);
1306 case PERF_COUNTER_IOC_DISABLE:
1307 perf_counter_disable_family(counter);
1309 case PERF_COUNTER_IOC_REFRESH:
1310 perf_counter_refresh(counter, arg);
1319 * Callers need to ensure there can be no nesting of this function, otherwise
1320 * the seqlock logic goes bad. We can not serialize this because the arch
1321 * code calls this from NMI context.
1323 void perf_counter_update_userpage(struct perf_counter *counter)
1325 struct perf_mmap_data *data;
1326 struct perf_counter_mmap_page *userpg;
1329 data = rcu_dereference(counter->data);
1333 userpg = data->user_page;
1336 * Disable preemption so as to not let the corresponding user-space
1337 * spin too long if we get preempted.
1342 userpg->index = counter->hw.idx;
1343 userpg->offset = atomic64_read(&counter->count);
1344 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1345 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1354 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1356 struct perf_counter *counter = vma->vm_file->private_data;
1357 struct perf_mmap_data *data;
1358 int ret = VM_FAULT_SIGBUS;
1361 data = rcu_dereference(counter->data);
1365 if (vmf->pgoff == 0) {
1366 vmf->page = virt_to_page(data->user_page);
1368 int nr = vmf->pgoff - 1;
1370 if ((unsigned)nr > data->nr_pages)
1373 vmf->page = virt_to_page(data->data_pages[nr]);
1375 get_page(vmf->page);
1383 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1385 struct perf_mmap_data *data;
1389 WARN_ON(atomic_read(&counter->mmap_count));
1391 size = sizeof(struct perf_mmap_data);
1392 size += nr_pages * sizeof(void *);
1394 data = kzalloc(size, GFP_KERNEL);
1398 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1399 if (!data->user_page)
1400 goto fail_user_page;
1402 for (i = 0; i < nr_pages; i++) {
1403 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1404 if (!data->data_pages[i])
1405 goto fail_data_pages;
1408 data->nr_pages = nr_pages;
1410 rcu_assign_pointer(counter->data, data);
1415 for (i--; i >= 0; i--)
1416 free_page((unsigned long)data->data_pages[i]);
1418 free_page((unsigned long)data->user_page);
1427 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1429 struct perf_mmap_data *data = container_of(rcu_head,
1430 struct perf_mmap_data, rcu_head);
1433 free_page((unsigned long)data->user_page);
1434 for (i = 0; i < data->nr_pages; i++)
1435 free_page((unsigned long)data->data_pages[i]);
1439 static void perf_mmap_data_free(struct perf_counter *counter)
1441 struct perf_mmap_data *data = counter->data;
1443 WARN_ON(atomic_read(&counter->mmap_count));
1445 rcu_assign_pointer(counter->data, NULL);
1446 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1449 static void perf_mmap_open(struct vm_area_struct *vma)
1451 struct perf_counter *counter = vma->vm_file->private_data;
1453 atomic_inc(&counter->mmap_count);
1456 static void perf_mmap_close(struct vm_area_struct *vma)
1458 struct perf_counter *counter = vma->vm_file->private_data;
1460 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1461 &counter->mmap_mutex)) {
1462 vma->vm_mm->locked_vm -= counter->data->nr_pages + 1;
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 mutex_lock(&counter->mmap_mutex);
1502 if (atomic_inc_not_zero(&counter->mmap_count)) {
1503 if (nr_pages != counter->data->nr_pages)
1508 locked = vma->vm_mm->locked_vm;
1509 locked += nr_pages + 1;
1511 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1512 lock_limit >>= PAGE_SHIFT;
1514 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1519 WARN_ON(counter->data);
1520 ret = perf_mmap_data_alloc(counter, nr_pages);
1524 atomic_set(&counter->mmap_count, 1);
1525 vma->vm_mm->locked_vm += nr_pages + 1;
1527 mutex_unlock(&counter->mmap_mutex);
1529 vma->vm_flags &= ~VM_MAYWRITE;
1530 vma->vm_flags |= VM_RESERVED;
1531 vma->vm_ops = &perf_mmap_vmops;
1536 static int perf_fasync(int fd, struct file *filp, int on)
1538 struct perf_counter *counter = filp->private_data;
1539 struct inode *inode = filp->f_path.dentry->d_inode;
1542 mutex_lock(&inode->i_mutex);
1543 retval = fasync_helper(fd, filp, on, &counter->fasync);
1544 mutex_unlock(&inode->i_mutex);
1552 static const struct file_operations perf_fops = {
1553 .release = perf_release,
1556 .unlocked_ioctl = perf_ioctl,
1557 .compat_ioctl = perf_ioctl,
1559 .fasync = perf_fasync,
1563 * Perf counter wakeup
1565 * If there's data, ensure we set the poll() state and publish everything
1566 * to user-space before waking everybody up.
1569 void perf_counter_wakeup(struct perf_counter *counter)
1571 struct perf_mmap_data *data;
1574 data = rcu_dereference(counter->data);
1576 atomic_set(&data->wakeup, POLL_IN);
1578 * Ensure all data writes are issued before updating the
1579 * user-space data head information. The matching rmb()
1580 * will be in userspace after reading this value.
1583 data->user_page->data_head = atomic_read(&data->head);
1587 wake_up_all(&counter->waitq);
1589 if (counter->pending_kill) {
1590 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1591 counter->pending_kill = 0;
1598 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1600 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1601 * single linked list and use cmpxchg() to add entries lockless.
1604 static void perf_pending_counter(struct perf_pending_entry *entry)
1606 struct perf_counter *counter = container_of(entry,
1607 struct perf_counter, pending);
1609 if (counter->pending_disable) {
1610 counter->pending_disable = 0;
1611 perf_counter_disable(counter);
1614 if (counter->pending_wakeup) {
1615 counter->pending_wakeup = 0;
1616 perf_counter_wakeup(counter);
1620 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1622 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1626 static void perf_pending_queue(struct perf_pending_entry *entry,
1627 void (*func)(struct perf_pending_entry *))
1629 struct perf_pending_entry **head;
1631 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1636 head = &get_cpu_var(perf_pending_head);
1639 entry->next = *head;
1640 } while (cmpxchg(head, entry->next, entry) != entry->next);
1642 set_perf_counter_pending();
1644 put_cpu_var(perf_pending_head);
1647 static int __perf_pending_run(void)
1649 struct perf_pending_entry *list;
1652 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1653 while (list != PENDING_TAIL) {
1654 void (*func)(struct perf_pending_entry *);
1655 struct perf_pending_entry *entry = list;
1662 * Ensure we observe the unqueue before we issue the wakeup,
1663 * so that we won't be waiting forever.
1664 * -- see perf_not_pending().
1675 static inline int perf_not_pending(struct perf_counter *counter)
1678 * If we flush on whatever cpu we run, there is a chance we don't
1682 __perf_pending_run();
1686 * Ensure we see the proper queue state before going to sleep
1687 * so that we do not miss the wakeup. -- see perf_pending_handle()
1690 return counter->pending.next == NULL;
1693 static void perf_pending_sync(struct perf_counter *counter)
1695 wait_event(counter->waitq, perf_not_pending(counter));
1698 void perf_counter_do_pending(void)
1700 __perf_pending_run();
1704 * Callchain support -- arch specific
1707 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1716 struct perf_output_handle {
1717 struct perf_counter *counter;
1718 struct perf_mmap_data *data;
1719 unsigned int offset;
1726 static inline void __perf_output_wakeup(struct perf_output_handle *handle)
1729 handle->counter->pending_wakeup = 1;
1730 perf_pending_queue(&handle->counter->pending,
1731 perf_pending_counter);
1733 perf_counter_wakeup(handle->counter);
1736 static int perf_output_begin(struct perf_output_handle *handle,
1737 struct perf_counter *counter, unsigned int size,
1738 int nmi, int overflow)
1740 struct perf_mmap_data *data;
1741 unsigned int offset, head;
1744 data = rcu_dereference(counter->data);
1748 handle->counter = counter;
1750 handle->overflow = overflow;
1752 if (!data->nr_pages)
1756 offset = head = atomic_read(&data->head);
1758 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1760 handle->data = data;
1761 handle->offset = offset;
1762 handle->head = head;
1763 handle->wakeup = (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT);
1768 __perf_output_wakeup(handle);
1775 static void perf_output_copy(struct perf_output_handle *handle,
1776 void *buf, unsigned int len)
1778 unsigned int pages_mask;
1779 unsigned int offset;
1783 offset = handle->offset;
1784 pages_mask = handle->data->nr_pages - 1;
1785 pages = handle->data->data_pages;
1788 unsigned int page_offset;
1791 nr = (offset >> PAGE_SHIFT) & pages_mask;
1792 page_offset = offset & (PAGE_SIZE - 1);
1793 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1795 memcpy(pages[nr] + page_offset, buf, size);
1802 handle->offset = offset;
1804 WARN_ON_ONCE(handle->offset > handle->head);
1807 #define perf_output_put(handle, x) \
1808 perf_output_copy((handle), &(x), sizeof(x))
1810 static void perf_output_end(struct perf_output_handle *handle)
1812 int wakeup_events = handle->counter->hw_event.wakeup_events;
1814 if (handle->overflow && wakeup_events) {
1815 int events = atomic_inc_return(&handle->data->events);
1816 if (events >= wakeup_events) {
1817 atomic_sub(wakeup_events, &handle->data->events);
1818 __perf_output_wakeup(handle);
1820 } else if (handle->wakeup)
1821 __perf_output_wakeup(handle);
1825 static void perf_counter_output(struct perf_counter *counter,
1826 int nmi, struct pt_regs *regs, u64 addr)
1829 u64 record_type = counter->hw_event.record_type;
1830 struct perf_output_handle handle;
1831 struct perf_event_header header;
1840 struct perf_callchain_entry *callchain = NULL;
1841 int callchain_size = 0;
1845 header.size = sizeof(header);
1847 header.misc = PERF_EVENT_MISC_OVERFLOW;
1848 header.misc |= user_mode(regs) ?
1849 PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
1851 if (record_type & PERF_RECORD_IP) {
1852 ip = instruction_pointer(regs);
1853 header.type |= PERF_RECORD_IP;
1854 header.size += sizeof(ip);
1857 if (record_type & PERF_RECORD_TID) {
1858 /* namespace issues */
1859 tid_entry.pid = current->group_leader->pid;
1860 tid_entry.tid = current->pid;
1862 header.type |= PERF_RECORD_TID;
1863 header.size += sizeof(tid_entry);
1866 if (record_type & PERF_RECORD_TIME) {
1868 * Maybe do better on x86 and provide cpu_clock_nmi()
1870 time = sched_clock();
1872 header.type |= PERF_RECORD_TIME;
1873 header.size += sizeof(u64);
1876 if (record_type & PERF_RECORD_ADDR) {
1877 header.type |= PERF_RECORD_ADDR;
1878 header.size += sizeof(u64);
1881 if (record_type & PERF_RECORD_GROUP) {
1882 header.type |= PERF_RECORD_GROUP;
1883 header.size += sizeof(u64) +
1884 counter->nr_siblings * sizeof(group_entry);
1887 if (record_type & PERF_RECORD_CALLCHAIN) {
1888 callchain = perf_callchain(regs);
1891 callchain_size = (1 + callchain->nr) * sizeof(u64);
1893 header.type |= PERF_RECORD_CALLCHAIN;
1894 header.size += callchain_size;
1898 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
1902 perf_output_put(&handle, header);
1904 if (record_type & PERF_RECORD_IP)
1905 perf_output_put(&handle, ip);
1907 if (record_type & PERF_RECORD_TID)
1908 perf_output_put(&handle, tid_entry);
1910 if (record_type & PERF_RECORD_TIME)
1911 perf_output_put(&handle, time);
1913 if (record_type & PERF_RECORD_ADDR)
1914 perf_output_put(&handle, addr);
1916 if (record_type & PERF_RECORD_GROUP) {
1917 struct perf_counter *leader, *sub;
1918 u64 nr = counter->nr_siblings;
1920 perf_output_put(&handle, nr);
1922 leader = counter->group_leader;
1923 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1925 sub->pmu->read(sub);
1927 group_entry.event = sub->hw_event.config;
1928 group_entry.counter = atomic64_read(&sub->count);
1930 perf_output_put(&handle, group_entry);
1935 perf_output_copy(&handle, callchain, callchain_size);
1937 perf_output_end(&handle);
1944 struct perf_comm_event {
1945 struct task_struct *task;
1950 struct perf_event_header header;
1957 static void perf_counter_comm_output(struct perf_counter *counter,
1958 struct perf_comm_event *comm_event)
1960 struct perf_output_handle handle;
1961 int size = comm_event->event.header.size;
1962 int ret = perf_output_begin(&handle, counter, size, 0, 0);
1967 perf_output_put(&handle, comm_event->event);
1968 perf_output_copy(&handle, comm_event->comm,
1969 comm_event->comm_size);
1970 perf_output_end(&handle);
1973 static int perf_counter_comm_match(struct perf_counter *counter,
1974 struct perf_comm_event *comm_event)
1976 if (counter->hw_event.comm &&
1977 comm_event->event.header.type == PERF_EVENT_COMM)
1983 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
1984 struct perf_comm_event *comm_event)
1986 struct perf_counter *counter;
1988 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
1992 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
1993 if (perf_counter_comm_match(counter, comm_event))
1994 perf_counter_comm_output(counter, comm_event);
1999 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2001 struct perf_cpu_context *cpuctx;
2003 char *comm = comm_event->task->comm;
2005 size = ALIGN(strlen(comm)+1, sizeof(u64));
2007 comm_event->comm = comm;
2008 comm_event->comm_size = size;
2010 comm_event->event.header.size = sizeof(comm_event->event) + size;
2012 cpuctx = &get_cpu_var(perf_cpu_context);
2013 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2014 put_cpu_var(perf_cpu_context);
2016 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
2019 void perf_counter_comm(struct task_struct *task)
2021 struct perf_comm_event comm_event;
2023 if (!atomic_read(&nr_comm_tracking))
2026 comm_event = (struct perf_comm_event){
2029 .header = { .type = PERF_EVENT_COMM, },
2030 .pid = task->group_leader->pid,
2035 perf_counter_comm_event(&comm_event);
2042 struct perf_mmap_event {
2048 struct perf_event_header header;
2058 static void perf_counter_mmap_output(struct perf_counter *counter,
2059 struct perf_mmap_event *mmap_event)
2061 struct perf_output_handle handle;
2062 int size = mmap_event->event.header.size;
2063 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2068 perf_output_put(&handle, mmap_event->event);
2069 perf_output_copy(&handle, mmap_event->file_name,
2070 mmap_event->file_size);
2071 perf_output_end(&handle);
2074 static int perf_counter_mmap_match(struct perf_counter *counter,
2075 struct perf_mmap_event *mmap_event)
2077 if (counter->hw_event.mmap &&
2078 mmap_event->event.header.type == PERF_EVENT_MMAP)
2081 if (counter->hw_event.munmap &&
2082 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2088 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2089 struct perf_mmap_event *mmap_event)
2091 struct perf_counter *counter;
2093 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2097 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2098 if (perf_counter_mmap_match(counter, mmap_event))
2099 perf_counter_mmap_output(counter, mmap_event);
2104 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2106 struct perf_cpu_context *cpuctx;
2107 struct file *file = mmap_event->file;
2114 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2116 name = strncpy(tmp, "//enomem", sizeof(tmp));
2119 name = d_path(&file->f_path, buf, PATH_MAX);
2121 name = strncpy(tmp, "//toolong", sizeof(tmp));
2125 name = strncpy(tmp, "//anon", sizeof(tmp));
2130 size = ALIGN(strlen(name)+1, sizeof(u64));
2132 mmap_event->file_name = name;
2133 mmap_event->file_size = size;
2135 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2137 cpuctx = &get_cpu_var(perf_cpu_context);
2138 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2139 put_cpu_var(perf_cpu_context);
2141 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2146 void perf_counter_mmap(unsigned long addr, unsigned long len,
2147 unsigned long pgoff, struct file *file)
2149 struct perf_mmap_event mmap_event;
2151 if (!atomic_read(&nr_mmap_tracking))
2154 mmap_event = (struct perf_mmap_event){
2157 .header = { .type = PERF_EVENT_MMAP, },
2158 .pid = current->group_leader->pid,
2159 .tid = current->pid,
2166 perf_counter_mmap_event(&mmap_event);
2169 void perf_counter_munmap(unsigned long addr, unsigned long len,
2170 unsigned long pgoff, struct file *file)
2172 struct perf_mmap_event mmap_event;
2174 if (!atomic_read(&nr_munmap_tracking))
2177 mmap_event = (struct perf_mmap_event){
2180 .header = { .type = PERF_EVENT_MUNMAP, },
2181 .pid = current->group_leader->pid,
2182 .tid = current->pid,
2189 perf_counter_mmap_event(&mmap_event);
2193 * Generic counter overflow handling.
2196 int perf_counter_overflow(struct perf_counter *counter,
2197 int nmi, struct pt_regs *regs, u64 addr)
2199 int events = atomic_read(&counter->event_limit);
2202 counter->pending_kill = POLL_IN;
2203 if (events && atomic_dec_and_test(&counter->event_limit)) {
2205 counter->pending_kill = POLL_HUP;
2207 counter->pending_disable = 1;
2208 perf_pending_queue(&counter->pending,
2209 perf_pending_counter);
2211 perf_counter_disable(counter);
2214 perf_counter_output(counter, nmi, regs, addr);
2219 * Generic software counter infrastructure
2222 static void perf_swcounter_update(struct perf_counter *counter)
2224 struct hw_perf_counter *hwc = &counter->hw;
2229 prev = atomic64_read(&hwc->prev_count);
2230 now = atomic64_read(&hwc->count);
2231 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2236 atomic64_add(delta, &counter->count);
2237 atomic64_sub(delta, &hwc->period_left);
2240 static void perf_swcounter_set_period(struct perf_counter *counter)
2242 struct hw_perf_counter *hwc = &counter->hw;
2243 s64 left = atomic64_read(&hwc->period_left);
2244 s64 period = hwc->irq_period;
2246 if (unlikely(left <= -period)) {
2248 atomic64_set(&hwc->period_left, left);
2251 if (unlikely(left <= 0)) {
2253 atomic64_add(period, &hwc->period_left);
2256 atomic64_set(&hwc->prev_count, -left);
2257 atomic64_set(&hwc->count, -left);
2260 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2262 enum hrtimer_restart ret = HRTIMER_RESTART;
2263 struct perf_counter *counter;
2264 struct pt_regs *regs;
2266 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2267 counter->pmu->read(counter);
2269 regs = get_irq_regs();
2271 * In case we exclude kernel IPs or are somehow not in interrupt
2272 * context, provide the next best thing, the user IP.
2274 if ((counter->hw_event.exclude_kernel || !regs) &&
2275 !counter->hw_event.exclude_user)
2276 regs = task_pt_regs(current);
2279 if (perf_counter_overflow(counter, 0, regs, 0))
2280 ret = HRTIMER_NORESTART;
2283 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2288 static void perf_swcounter_overflow(struct perf_counter *counter,
2289 int nmi, struct pt_regs *regs, u64 addr)
2291 perf_swcounter_update(counter);
2292 perf_swcounter_set_period(counter);
2293 if (perf_counter_overflow(counter, nmi, regs, addr))
2294 /* soft-disable the counter */
2299 static int perf_swcounter_match(struct perf_counter *counter,
2300 enum perf_event_types type,
2301 u32 event, struct pt_regs *regs)
2303 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2306 if (perf_event_raw(&counter->hw_event))
2309 if (perf_event_type(&counter->hw_event) != type)
2312 if (perf_event_id(&counter->hw_event) != event)
2315 if (counter->hw_event.exclude_user && user_mode(regs))
2318 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2324 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2325 int nmi, struct pt_regs *regs, u64 addr)
2327 int neg = atomic64_add_negative(nr, &counter->hw.count);
2328 if (counter->hw.irq_period && !neg)
2329 perf_swcounter_overflow(counter, nmi, regs, addr);
2332 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2333 enum perf_event_types type, u32 event,
2334 u64 nr, int nmi, struct pt_regs *regs,
2337 struct perf_counter *counter;
2339 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2343 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2344 if (perf_swcounter_match(counter, type, event, regs))
2345 perf_swcounter_add(counter, nr, nmi, regs, addr);
2350 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2353 return &cpuctx->recursion[3];
2356 return &cpuctx->recursion[2];
2359 return &cpuctx->recursion[1];
2361 return &cpuctx->recursion[0];
2364 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2365 u64 nr, int nmi, struct pt_regs *regs,
2368 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2369 int *recursion = perf_swcounter_recursion_context(cpuctx);
2377 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2378 nr, nmi, regs, addr);
2379 if (cpuctx->task_ctx) {
2380 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2381 nr, nmi, regs, addr);
2388 put_cpu_var(perf_cpu_context);
2392 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2394 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2397 static void perf_swcounter_read(struct perf_counter *counter)
2399 perf_swcounter_update(counter);
2402 static int perf_swcounter_enable(struct perf_counter *counter)
2404 perf_swcounter_set_period(counter);
2408 static void perf_swcounter_disable(struct perf_counter *counter)
2410 perf_swcounter_update(counter);
2413 static const struct pmu perf_ops_generic = {
2414 .enable = perf_swcounter_enable,
2415 .disable = perf_swcounter_disable,
2416 .read = perf_swcounter_read,
2420 * Software counter: cpu wall time clock
2423 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2425 int cpu = raw_smp_processor_id();
2429 now = cpu_clock(cpu);
2430 prev = atomic64_read(&counter->hw.prev_count);
2431 atomic64_set(&counter->hw.prev_count, now);
2432 atomic64_add(now - prev, &counter->count);
2435 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2437 struct hw_perf_counter *hwc = &counter->hw;
2438 int cpu = raw_smp_processor_id();
2440 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2441 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2442 hwc->hrtimer.function = perf_swcounter_hrtimer;
2443 if (hwc->irq_period) {
2444 __hrtimer_start_range_ns(&hwc->hrtimer,
2445 ns_to_ktime(hwc->irq_period), 0,
2446 HRTIMER_MODE_REL, 0);
2452 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2454 hrtimer_cancel(&counter->hw.hrtimer);
2455 cpu_clock_perf_counter_update(counter);
2458 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2460 cpu_clock_perf_counter_update(counter);
2463 static const struct pmu perf_ops_cpu_clock = {
2464 .enable = cpu_clock_perf_counter_enable,
2465 .disable = cpu_clock_perf_counter_disable,
2466 .read = cpu_clock_perf_counter_read,
2470 * Software counter: task time clock
2473 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2478 prev = atomic64_xchg(&counter->hw.prev_count, now);
2480 atomic64_add(delta, &counter->count);
2483 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2485 struct hw_perf_counter *hwc = &counter->hw;
2488 now = counter->ctx->time;
2490 atomic64_set(&hwc->prev_count, now);
2491 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2492 hwc->hrtimer.function = perf_swcounter_hrtimer;
2493 if (hwc->irq_period) {
2494 __hrtimer_start_range_ns(&hwc->hrtimer,
2495 ns_to_ktime(hwc->irq_period), 0,
2496 HRTIMER_MODE_REL, 0);
2502 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2504 hrtimer_cancel(&counter->hw.hrtimer);
2505 task_clock_perf_counter_update(counter, counter->ctx->time);
2509 static void task_clock_perf_counter_read(struct perf_counter *counter)
2514 update_context_time(counter->ctx);
2515 time = counter->ctx->time;
2517 u64 now = perf_clock();
2518 u64 delta = now - counter->ctx->timestamp;
2519 time = counter->ctx->time + delta;
2522 task_clock_perf_counter_update(counter, time);
2525 static const struct pmu perf_ops_task_clock = {
2526 .enable = task_clock_perf_counter_enable,
2527 .disable = task_clock_perf_counter_disable,
2528 .read = task_clock_perf_counter_read,
2532 * Software counter: cpu migrations
2535 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2537 struct task_struct *curr = counter->ctx->task;
2540 return curr->se.nr_migrations;
2541 return cpu_nr_migrations(smp_processor_id());
2544 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2549 prev = atomic64_read(&counter->hw.prev_count);
2550 now = get_cpu_migrations(counter);
2552 atomic64_set(&counter->hw.prev_count, now);
2556 atomic64_add(delta, &counter->count);
2559 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2561 cpu_migrations_perf_counter_update(counter);
2564 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2566 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2567 atomic64_set(&counter->hw.prev_count,
2568 get_cpu_migrations(counter));
2572 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2574 cpu_migrations_perf_counter_update(counter);
2577 static const struct pmu perf_ops_cpu_migrations = {
2578 .enable = cpu_migrations_perf_counter_enable,
2579 .disable = cpu_migrations_perf_counter_disable,
2580 .read = cpu_migrations_perf_counter_read,
2583 #ifdef CONFIG_EVENT_PROFILE
2584 void perf_tpcounter_event(int event_id)
2586 struct pt_regs *regs = get_irq_regs();
2589 regs = task_pt_regs(current);
2591 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2593 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2595 extern int ftrace_profile_enable(int);
2596 extern void ftrace_profile_disable(int);
2598 static void tp_perf_counter_destroy(struct perf_counter *counter)
2600 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2603 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2605 int event_id = perf_event_id(&counter->hw_event);
2608 ret = ftrace_profile_enable(event_id);
2612 counter->destroy = tp_perf_counter_destroy;
2613 counter->hw.irq_period = counter->hw_event.irq_period;
2615 return &perf_ops_generic;
2618 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2624 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2626 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2627 const struct pmu *pmu = NULL;
2628 struct hw_perf_counter *hwc = &counter->hw;
2631 * Software counters (currently) can't in general distinguish
2632 * between user, kernel and hypervisor events.
2633 * However, context switches and cpu migrations are considered
2634 * to be kernel events, and page faults are never hypervisor
2637 switch (perf_event_id(&counter->hw_event)) {
2638 case PERF_COUNT_CPU_CLOCK:
2639 pmu = &perf_ops_cpu_clock;
2641 if (hw_event->irq_period && hw_event->irq_period < 10000)
2642 hw_event->irq_period = 10000;
2644 case PERF_COUNT_TASK_CLOCK:
2646 * If the user instantiates this as a per-cpu counter,
2647 * use the cpu_clock counter instead.
2649 if (counter->ctx->task)
2650 pmu = &perf_ops_task_clock;
2652 pmu = &perf_ops_cpu_clock;
2654 if (hw_event->irq_period && hw_event->irq_period < 10000)
2655 hw_event->irq_period = 10000;
2657 case PERF_COUNT_PAGE_FAULTS:
2658 case PERF_COUNT_PAGE_FAULTS_MIN:
2659 case PERF_COUNT_PAGE_FAULTS_MAJ:
2660 case PERF_COUNT_CONTEXT_SWITCHES:
2661 pmu = &perf_ops_generic;
2663 case PERF_COUNT_CPU_MIGRATIONS:
2664 if (!counter->hw_event.exclude_kernel)
2665 pmu = &perf_ops_cpu_migrations;
2670 hwc->irq_period = hw_event->irq_period;
2676 * Allocate and initialize a counter structure
2678 static struct perf_counter *
2679 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2681 struct perf_counter_context *ctx,
2682 struct perf_counter *group_leader,
2685 const struct pmu *pmu;
2686 struct perf_counter *counter;
2689 counter = kzalloc(sizeof(*counter), gfpflags);
2691 return ERR_PTR(-ENOMEM);
2694 * Single counters are their own group leaders, with an
2695 * empty sibling list:
2698 group_leader = counter;
2700 mutex_init(&counter->mutex);
2701 INIT_LIST_HEAD(&counter->list_entry);
2702 INIT_LIST_HEAD(&counter->event_entry);
2703 INIT_LIST_HEAD(&counter->sibling_list);
2704 init_waitqueue_head(&counter->waitq);
2706 mutex_init(&counter->mmap_mutex);
2708 INIT_LIST_HEAD(&counter->child_list);
2711 counter->hw_event = *hw_event;
2712 counter->group_leader = group_leader;
2713 counter->pmu = NULL;
2716 counter->state = PERF_COUNTER_STATE_INACTIVE;
2717 if (hw_event->disabled)
2718 counter->state = PERF_COUNTER_STATE_OFF;
2722 if (perf_event_raw(hw_event)) {
2723 pmu = hw_perf_counter_init(counter);
2727 switch (perf_event_type(hw_event)) {
2728 case PERF_TYPE_HARDWARE:
2729 pmu = hw_perf_counter_init(counter);
2732 case PERF_TYPE_SOFTWARE:
2733 pmu = sw_perf_counter_init(counter);
2736 case PERF_TYPE_TRACEPOINT:
2737 pmu = tp_perf_counter_init(counter);
2744 else if (IS_ERR(pmu))
2749 return ERR_PTR(err);
2754 if (counter->hw_event.mmap)
2755 atomic_inc(&nr_mmap_tracking);
2756 if (counter->hw_event.munmap)
2757 atomic_inc(&nr_munmap_tracking);
2758 if (counter->hw_event.comm)
2759 atomic_inc(&nr_comm_tracking);
2765 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2767 * @hw_event_uptr: event type attributes for monitoring/sampling
2770 * @group_fd: group leader counter fd
2772 SYSCALL_DEFINE5(perf_counter_open,
2773 const struct perf_counter_hw_event __user *, hw_event_uptr,
2774 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2776 struct perf_counter *counter, *group_leader;
2777 struct perf_counter_hw_event hw_event;
2778 struct perf_counter_context *ctx;
2779 struct file *counter_file = NULL;
2780 struct file *group_file = NULL;
2781 int fput_needed = 0;
2782 int fput_needed2 = 0;
2785 /* for future expandability... */
2789 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2793 * Get the target context (task or percpu):
2795 ctx = find_get_context(pid, cpu);
2797 return PTR_ERR(ctx);
2800 * Look up the group leader (we will attach this counter to it):
2802 group_leader = NULL;
2803 if (group_fd != -1) {
2805 group_file = fget_light(group_fd, &fput_needed);
2807 goto err_put_context;
2808 if (group_file->f_op != &perf_fops)
2809 goto err_put_context;
2811 group_leader = group_file->private_data;
2813 * Do not allow a recursive hierarchy (this new sibling
2814 * becoming part of another group-sibling):
2816 if (group_leader->group_leader != group_leader)
2817 goto err_put_context;
2819 * Do not allow to attach to a group in a different
2820 * task or CPU context:
2822 if (group_leader->ctx != ctx)
2823 goto err_put_context;
2825 * Only a group leader can be exclusive or pinned
2827 if (hw_event.exclusive || hw_event.pinned)
2828 goto err_put_context;
2831 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2833 ret = PTR_ERR(counter);
2834 if (IS_ERR(counter))
2835 goto err_put_context;
2837 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2839 goto err_free_put_context;
2841 counter_file = fget_light(ret, &fput_needed2);
2843 goto err_free_put_context;
2845 counter->filp = counter_file;
2846 mutex_lock(&ctx->mutex);
2847 perf_install_in_context(ctx, counter, cpu);
2848 mutex_unlock(&ctx->mutex);
2850 fput_light(counter_file, fput_needed2);
2853 fput_light(group_file, fput_needed);
2857 err_free_put_context:
2867 * Initialize the perf_counter context in a task_struct:
2870 __perf_counter_init_context(struct perf_counter_context *ctx,
2871 struct task_struct *task)
2873 memset(ctx, 0, sizeof(*ctx));
2874 spin_lock_init(&ctx->lock);
2875 mutex_init(&ctx->mutex);
2876 INIT_LIST_HEAD(&ctx->counter_list);
2877 INIT_LIST_HEAD(&ctx->event_list);
2882 * inherit a counter from parent task to child task:
2884 static struct perf_counter *
2885 inherit_counter(struct perf_counter *parent_counter,
2886 struct task_struct *parent,
2887 struct perf_counter_context *parent_ctx,
2888 struct task_struct *child,
2889 struct perf_counter *group_leader,
2890 struct perf_counter_context *child_ctx)
2892 struct perf_counter *child_counter;
2895 * Instead of creating recursive hierarchies of counters,
2896 * we link inherited counters back to the original parent,
2897 * which has a filp for sure, which we use as the reference
2900 if (parent_counter->parent)
2901 parent_counter = parent_counter->parent;
2903 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2904 parent_counter->cpu, child_ctx,
2905 group_leader, GFP_KERNEL);
2906 if (IS_ERR(child_counter))
2907 return child_counter;
2910 * Link it up in the child's context:
2912 child_counter->task = child;
2913 add_counter_to_ctx(child_counter, child_ctx);
2915 child_counter->parent = parent_counter;
2917 * inherit into child's child as well:
2919 child_counter->hw_event.inherit = 1;
2922 * Get a reference to the parent filp - we will fput it
2923 * when the child counter exits. This is safe to do because
2924 * we are in the parent and we know that the filp still
2925 * exists and has a nonzero count:
2927 atomic_long_inc(&parent_counter->filp->f_count);
2930 * Link this into the parent counter's child list
2932 mutex_lock(&parent_counter->mutex);
2933 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2936 * Make the child state follow the state of the parent counter,
2937 * not its hw_event.disabled bit. We hold the parent's mutex,
2938 * so we won't race with perf_counter_{en,dis}able_family.
2940 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2941 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2943 child_counter->state = PERF_COUNTER_STATE_OFF;
2945 mutex_unlock(&parent_counter->mutex);
2947 return child_counter;
2950 static int inherit_group(struct perf_counter *parent_counter,
2951 struct task_struct *parent,
2952 struct perf_counter_context *parent_ctx,
2953 struct task_struct *child,
2954 struct perf_counter_context *child_ctx)
2956 struct perf_counter *leader;
2957 struct perf_counter *sub;
2958 struct perf_counter *child_ctr;
2960 leader = inherit_counter(parent_counter, parent, parent_ctx,
2961 child, NULL, child_ctx);
2963 return PTR_ERR(leader);
2964 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2965 child_ctr = inherit_counter(sub, parent, parent_ctx,
2966 child, leader, child_ctx);
2967 if (IS_ERR(child_ctr))
2968 return PTR_ERR(child_ctr);
2973 static void sync_child_counter(struct perf_counter *child_counter,
2974 struct perf_counter *parent_counter)
2976 u64 parent_val, child_val;
2978 parent_val = atomic64_read(&parent_counter->count);
2979 child_val = atomic64_read(&child_counter->count);
2982 * Add back the child's count to the parent's count:
2984 atomic64_add(child_val, &parent_counter->count);
2985 atomic64_add(child_counter->total_time_enabled,
2986 &parent_counter->child_total_time_enabled);
2987 atomic64_add(child_counter->total_time_running,
2988 &parent_counter->child_total_time_running);
2991 * Remove this counter from the parent's list
2993 mutex_lock(&parent_counter->mutex);
2994 list_del_init(&child_counter->child_list);
2995 mutex_unlock(&parent_counter->mutex);
2998 * Release the parent counter, if this was the last
3001 fput(parent_counter->filp);
3005 __perf_counter_exit_task(struct task_struct *child,
3006 struct perf_counter *child_counter,
3007 struct perf_counter_context *child_ctx)
3009 struct perf_counter *parent_counter;
3010 struct perf_counter *sub, *tmp;
3013 * If we do not self-reap then we have to wait for the
3014 * child task to unschedule (it will happen for sure),
3015 * so that its counter is at its final count. (This
3016 * condition triggers rarely - child tasks usually get
3017 * off their CPU before the parent has a chance to
3018 * get this far into the reaping action)
3020 if (child != current) {
3021 wait_task_inactive(child, 0);
3022 list_del_init(&child_counter->list_entry);
3023 update_counter_times(child_counter);
3025 struct perf_cpu_context *cpuctx;
3026 unsigned long flags;
3030 * Disable and unlink this counter.
3032 * Be careful about zapping the list - IRQ/NMI context
3033 * could still be processing it:
3035 local_irq_save(flags);
3036 perf_flags = hw_perf_save_disable();
3038 cpuctx = &__get_cpu_var(perf_cpu_context);
3040 group_sched_out(child_counter, cpuctx, child_ctx);
3041 update_counter_times(child_counter);
3043 list_del_init(&child_counter->list_entry);
3045 child_ctx->nr_counters--;
3047 hw_perf_restore(perf_flags);
3048 local_irq_restore(flags);
3051 parent_counter = child_counter->parent;
3053 * It can happen that parent exits first, and has counters
3054 * that are still around due to the child reference. These
3055 * counters need to be zapped - but otherwise linger.
3057 if (parent_counter) {
3058 sync_child_counter(child_counter, parent_counter);
3059 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
3062 sync_child_counter(sub, sub->parent);
3066 free_counter(child_counter);
3071 * When a child task exits, feed back counter values to parent counters.
3073 * Note: we may be running in child context, but the PID is not hashed
3074 * anymore so new counters will not be added.
3076 void perf_counter_exit_task(struct task_struct *child)
3078 struct perf_counter *child_counter, *tmp;
3079 struct perf_counter_context *child_ctx;
3081 child_ctx = &child->perf_counter_ctx;
3083 if (likely(!child_ctx->nr_counters))
3086 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3088 __perf_counter_exit_task(child, child_counter, child_ctx);
3092 * Initialize the perf_counter context in task_struct
3094 void perf_counter_init_task(struct task_struct *child)
3096 struct perf_counter_context *child_ctx, *parent_ctx;
3097 struct perf_counter *counter;
3098 struct task_struct *parent = current;
3100 child_ctx = &child->perf_counter_ctx;
3101 parent_ctx = &parent->perf_counter_ctx;
3103 __perf_counter_init_context(child_ctx, child);
3106 * This is executed from the parent task context, so inherit
3107 * counters that have been marked for cloning:
3110 if (likely(!parent_ctx->nr_counters))
3114 * Lock the parent list. No need to lock the child - not PID
3115 * hashed yet and not running, so nobody can access it.
3117 mutex_lock(&parent_ctx->mutex);
3120 * We dont have to disable NMIs - we are only looking at
3121 * the list, not manipulating it:
3123 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
3124 if (!counter->hw_event.inherit)
3127 if (inherit_group(counter, parent,
3128 parent_ctx, child, child_ctx))
3132 mutex_unlock(&parent_ctx->mutex);
3135 static void __cpuinit perf_counter_init_cpu(int cpu)
3137 struct perf_cpu_context *cpuctx;
3139 cpuctx = &per_cpu(perf_cpu_context, cpu);
3140 __perf_counter_init_context(&cpuctx->ctx, NULL);
3142 mutex_lock(&perf_resource_mutex);
3143 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3144 mutex_unlock(&perf_resource_mutex);
3146 hw_perf_counter_setup(cpu);
3149 #ifdef CONFIG_HOTPLUG_CPU
3150 static void __perf_counter_exit_cpu(void *info)
3152 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3153 struct perf_counter_context *ctx = &cpuctx->ctx;
3154 struct perf_counter *counter, *tmp;
3156 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3157 __perf_counter_remove_from_context(counter);
3159 static void perf_counter_exit_cpu(int cpu)
3161 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3162 struct perf_counter_context *ctx = &cpuctx->ctx;
3164 mutex_lock(&ctx->mutex);
3165 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3166 mutex_unlock(&ctx->mutex);
3169 static inline void perf_counter_exit_cpu(int cpu) { }
3172 static int __cpuinit
3173 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3175 unsigned int cpu = (long)hcpu;
3179 case CPU_UP_PREPARE:
3180 case CPU_UP_PREPARE_FROZEN:
3181 perf_counter_init_cpu(cpu);
3184 case CPU_DOWN_PREPARE:
3185 case CPU_DOWN_PREPARE_FROZEN:
3186 perf_counter_exit_cpu(cpu);
3196 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3197 .notifier_call = perf_cpu_notify,
3200 static int __init perf_counter_init(void)
3202 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3203 (void *)(long)smp_processor_id());
3204 register_cpu_notifier(&perf_cpu_nb);
3208 early_initcall(perf_counter_init);
3210 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3212 return sprintf(buf, "%d\n", perf_reserved_percpu);
3216 perf_set_reserve_percpu(struct sysdev_class *class,
3220 struct perf_cpu_context *cpuctx;
3224 err = strict_strtoul(buf, 10, &val);
3227 if (val > perf_max_counters)
3230 mutex_lock(&perf_resource_mutex);
3231 perf_reserved_percpu = val;
3232 for_each_online_cpu(cpu) {
3233 cpuctx = &per_cpu(perf_cpu_context, cpu);
3234 spin_lock_irq(&cpuctx->ctx.lock);
3235 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3236 perf_max_counters - perf_reserved_percpu);
3237 cpuctx->max_pertask = mpt;
3238 spin_unlock_irq(&cpuctx->ctx.lock);
3240 mutex_unlock(&perf_resource_mutex);
3245 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3247 return sprintf(buf, "%d\n", perf_overcommit);
3251 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3256 err = strict_strtoul(buf, 10, &val);
3262 mutex_lock(&perf_resource_mutex);
3263 perf_overcommit = val;
3264 mutex_unlock(&perf_resource_mutex);
3269 static SYSDEV_CLASS_ATTR(
3272 perf_show_reserve_percpu,
3273 perf_set_reserve_percpu
3276 static SYSDEV_CLASS_ATTR(
3279 perf_show_overcommit,
3283 static struct attribute *perfclass_attrs[] = {
3284 &attr_reserve_percpu.attr,
3285 &attr_overcommit.attr,
3289 static struct attribute_group perfclass_attr_group = {
3290 .attrs = perfclass_attrs,
3291 .name = "perf_counters",
3294 static int __init perf_counter_sysfs_init(void)
3296 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3297 &perfclass_attr_group);
3299 device_initcall(perf_counter_sysfs_init);