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_counters __read_mostly;
43 static atomic_t nr_mmap_tracking __read_mostly;
44 static atomic_t nr_munmap_tracking __read_mostly;
45 static atomic_t nr_comm_tracking __read_mostly;
47 int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
48 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
51 * Lock for (sysadmin-configurable) counter reservations:
53 static DEFINE_SPINLOCK(perf_resource_lock);
56 * Architecture provided APIs - weak aliases:
58 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
63 void __weak hw_perf_disable(void) { barrier(); }
64 void __weak hw_perf_enable(void) { barrier(); }
66 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
67 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
68 struct perf_cpu_context *cpuctx,
69 struct perf_counter_context *ctx, int cpu)
74 void __weak perf_counter_print_debug(void) { }
76 static DEFINE_PER_CPU(int, disable_count);
78 void __perf_disable(void)
80 __get_cpu_var(disable_count)++;
83 bool __perf_enable(void)
85 return !--__get_cpu_var(disable_count);
88 void perf_disable(void)
94 void perf_enable(void)
101 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
103 struct perf_counter *group_leader = counter->group_leader;
106 * Depending on whether it is a standalone or sibling counter,
107 * add it straight to the context's counter list, or to the group
108 * leader's sibling list:
110 if (group_leader == counter)
111 list_add_tail(&counter->list_entry, &ctx->counter_list);
113 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
114 group_leader->nr_siblings++;
117 list_add_rcu(&counter->event_entry, &ctx->event_list);
121 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
123 struct perf_counter *sibling, *tmp;
125 list_del_init(&counter->list_entry);
126 list_del_rcu(&counter->event_entry);
128 if (counter->group_leader != counter)
129 counter->group_leader->nr_siblings--;
132 * If this was a group counter with sibling counters then
133 * upgrade the siblings to singleton counters by adding them
134 * to the context list directly:
136 list_for_each_entry_safe(sibling, tmp,
137 &counter->sibling_list, list_entry) {
139 list_move_tail(&sibling->list_entry, &ctx->counter_list);
140 sibling->group_leader = sibling;
145 counter_sched_out(struct perf_counter *counter,
146 struct perf_cpu_context *cpuctx,
147 struct perf_counter_context *ctx)
149 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
152 counter->state = PERF_COUNTER_STATE_INACTIVE;
153 counter->tstamp_stopped = ctx->time;
154 counter->pmu->disable(counter);
157 if (!is_software_counter(counter))
158 cpuctx->active_oncpu--;
160 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
161 cpuctx->exclusive = 0;
165 group_sched_out(struct perf_counter *group_counter,
166 struct perf_cpu_context *cpuctx,
167 struct perf_counter_context *ctx)
169 struct perf_counter *counter;
171 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
174 counter_sched_out(group_counter, cpuctx, ctx);
177 * Schedule out siblings (if any):
179 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
180 counter_sched_out(counter, cpuctx, ctx);
182 if (group_counter->hw_event.exclusive)
183 cpuctx->exclusive = 0;
187 * Cross CPU call to remove a performance counter
189 * We disable the counter on the hardware level first. After that we
190 * remove it from the context list.
192 static void __perf_counter_remove_from_context(void *info)
194 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
195 struct perf_counter *counter = info;
196 struct perf_counter_context *ctx = counter->ctx;
200 * If this is a task context, we need to check whether it is
201 * the current task context of this cpu. If not it has been
202 * scheduled out before the smp call arrived.
204 if (ctx->task && cpuctx->task_ctx != ctx)
207 spin_lock_irqsave(&ctx->lock, flags);
209 counter_sched_out(counter, cpuctx, ctx);
211 counter->task = NULL;
215 * Protect the list operation against NMI by disabling the
216 * counters on a global level. NOP for non NMI based counters.
219 list_del_counter(counter, ctx);
224 * Allow more per task counters with respect to the
227 cpuctx->max_pertask =
228 min(perf_max_counters - ctx->nr_counters,
229 perf_max_counters - perf_reserved_percpu);
232 spin_unlock_irqrestore(&ctx->lock, flags);
237 * Remove the counter from a task's (or a CPU's) list of counters.
239 * Must be called with counter->mutex and ctx->mutex held.
241 * CPU counters are removed with a smp call. For task counters we only
242 * call when the task is on a CPU.
244 static void perf_counter_remove_from_context(struct perf_counter *counter)
246 struct perf_counter_context *ctx = counter->ctx;
247 struct task_struct *task = ctx->task;
251 * Per cpu counters are removed via an smp call and
252 * the removal is always sucessful.
254 smp_call_function_single(counter->cpu,
255 __perf_counter_remove_from_context,
261 task_oncpu_function_call(task, __perf_counter_remove_from_context,
264 spin_lock_irq(&ctx->lock);
266 * If the context is active we need to retry the smp call.
268 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
269 spin_unlock_irq(&ctx->lock);
274 * The lock prevents that this context is scheduled in so we
275 * can remove the counter safely, if the call above did not
278 if (!list_empty(&counter->list_entry)) {
280 list_del_counter(counter, ctx);
281 counter->task = NULL;
283 spin_unlock_irq(&ctx->lock);
286 static inline u64 perf_clock(void)
288 return cpu_clock(smp_processor_id());
292 * Update the record of the current time in a context.
294 static void update_context_time(struct perf_counter_context *ctx)
296 u64 now = perf_clock();
298 ctx->time += now - ctx->timestamp;
299 ctx->timestamp = now;
303 * Update the total_time_enabled and total_time_running fields for a counter.
305 static void update_counter_times(struct perf_counter *counter)
307 struct perf_counter_context *ctx = counter->ctx;
310 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
313 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
315 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
316 run_end = counter->tstamp_stopped;
320 counter->total_time_running = run_end - counter->tstamp_running;
324 * Update total_time_enabled and total_time_running for all counters in a group.
326 static void update_group_times(struct perf_counter *leader)
328 struct perf_counter *counter;
330 update_counter_times(leader);
331 list_for_each_entry(counter, &leader->sibling_list, list_entry)
332 update_counter_times(counter);
336 * Cross CPU call to disable a performance counter
338 static void __perf_counter_disable(void *info)
340 struct perf_counter *counter = info;
341 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
342 struct perf_counter_context *ctx = counter->ctx;
346 * If this is a per-task counter, need to check whether this
347 * counter's task is the current task on this cpu.
349 if (ctx->task && cpuctx->task_ctx != ctx)
352 spin_lock_irqsave(&ctx->lock, flags);
355 * If the counter is on, turn it off.
356 * If it is in error state, leave it in error state.
358 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
359 update_context_time(ctx);
360 update_counter_times(counter);
361 if (counter == counter->group_leader)
362 group_sched_out(counter, cpuctx, ctx);
364 counter_sched_out(counter, cpuctx, ctx);
365 counter->state = PERF_COUNTER_STATE_OFF;
368 spin_unlock_irqrestore(&ctx->lock, flags);
374 static void perf_counter_disable(struct perf_counter *counter)
376 struct perf_counter_context *ctx = counter->ctx;
377 struct task_struct *task = ctx->task;
381 * Disable the counter on the cpu that it's on
383 smp_call_function_single(counter->cpu, __perf_counter_disable,
389 task_oncpu_function_call(task, __perf_counter_disable, counter);
391 spin_lock_irq(&ctx->lock);
393 * If the counter is still active, we need to retry the cross-call.
395 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
396 spin_unlock_irq(&ctx->lock);
401 * Since we have the lock this context can't be scheduled
402 * in, so we can change the state safely.
404 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
405 update_counter_times(counter);
406 counter->state = PERF_COUNTER_STATE_OFF;
409 spin_unlock_irq(&ctx->lock);
413 counter_sched_in(struct perf_counter *counter,
414 struct perf_cpu_context *cpuctx,
415 struct perf_counter_context *ctx,
418 if (counter->state <= PERF_COUNTER_STATE_OFF)
421 counter->state = PERF_COUNTER_STATE_ACTIVE;
422 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
424 * The new state must be visible before we turn it on in the hardware:
428 if (counter->pmu->enable(counter)) {
429 counter->state = PERF_COUNTER_STATE_INACTIVE;
434 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
436 if (!is_software_counter(counter))
437 cpuctx->active_oncpu++;
440 if (counter->hw_event.exclusive)
441 cpuctx->exclusive = 1;
447 group_sched_in(struct perf_counter *group_counter,
448 struct perf_cpu_context *cpuctx,
449 struct perf_counter_context *ctx,
452 struct perf_counter *counter, *partial_group;
455 if (group_counter->state == PERF_COUNTER_STATE_OFF)
458 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
460 return ret < 0 ? ret : 0;
462 group_counter->prev_state = group_counter->state;
463 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
467 * Schedule in siblings as one group (if any):
469 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
470 counter->prev_state = counter->state;
471 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
472 partial_group = counter;
481 * Groups can be scheduled in as one unit only, so undo any
482 * partial group before returning:
484 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
485 if (counter == partial_group)
487 counter_sched_out(counter, cpuctx, ctx);
489 counter_sched_out(group_counter, cpuctx, ctx);
495 * Return 1 for a group consisting entirely of software counters,
496 * 0 if the group contains any hardware counters.
498 static int is_software_only_group(struct perf_counter *leader)
500 struct perf_counter *counter;
502 if (!is_software_counter(leader))
505 list_for_each_entry(counter, &leader->sibling_list, list_entry)
506 if (!is_software_counter(counter))
513 * Work out whether we can put this counter group on the CPU now.
515 static int group_can_go_on(struct perf_counter *counter,
516 struct perf_cpu_context *cpuctx,
520 * Groups consisting entirely of software counters can always go on.
522 if (is_software_only_group(counter))
525 * If an exclusive group is already on, no other hardware
526 * counters can go on.
528 if (cpuctx->exclusive)
531 * If this group is exclusive and there are already
532 * counters on the CPU, it can't go on.
534 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
537 * Otherwise, try to add it if all previous groups were able
543 static void add_counter_to_ctx(struct perf_counter *counter,
544 struct perf_counter_context *ctx)
546 list_add_counter(counter, ctx);
548 counter->prev_state = PERF_COUNTER_STATE_OFF;
549 counter->tstamp_enabled = ctx->time;
550 counter->tstamp_running = ctx->time;
551 counter->tstamp_stopped = ctx->time;
555 * Cross CPU call to install and enable a performance counter
557 static void __perf_install_in_context(void *info)
559 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
560 struct perf_counter *counter = info;
561 struct perf_counter_context *ctx = counter->ctx;
562 struct perf_counter *leader = counter->group_leader;
563 int cpu = smp_processor_id();
568 * If this is a task context, we need to check whether it is
569 * the current task context of this cpu. If not it has been
570 * scheduled out before the smp call arrived.
572 if (ctx->task && cpuctx->task_ctx != ctx)
575 spin_lock_irqsave(&ctx->lock, flags);
576 update_context_time(ctx);
579 * Protect the list operation against NMI by disabling the
580 * counters on a global level. NOP for non NMI based counters.
584 add_counter_to_ctx(counter, ctx);
587 * Don't put the counter on if it is disabled or if
588 * it is in a group and the group isn't on.
590 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
591 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
595 * An exclusive counter can't go on if there are already active
596 * hardware counters, and no hardware counter can go on if there
597 * is already an exclusive counter on.
599 if (!group_can_go_on(counter, cpuctx, 1))
602 err = counter_sched_in(counter, cpuctx, ctx, cpu);
606 * This counter couldn't go on. If it is in a group
607 * then we have to pull the whole group off.
608 * If the counter group is pinned then put it in error state.
610 if (leader != counter)
611 group_sched_out(leader, cpuctx, ctx);
612 if (leader->hw_event.pinned) {
613 update_group_times(leader);
614 leader->state = PERF_COUNTER_STATE_ERROR;
618 if (!err && !ctx->task && cpuctx->max_pertask)
619 cpuctx->max_pertask--;
624 spin_unlock_irqrestore(&ctx->lock, flags);
628 * Attach a performance counter to a context
630 * First we add the counter to the list with the hardware enable bit
631 * in counter->hw_config cleared.
633 * If the counter is attached to a task which is on a CPU we use a smp
634 * call to enable it in the task context. The task might have been
635 * scheduled away, but we check this in the smp call again.
637 * Must be called with ctx->mutex held.
640 perf_install_in_context(struct perf_counter_context *ctx,
641 struct perf_counter *counter,
644 struct task_struct *task = ctx->task;
648 * Per cpu counters are installed via an smp call and
649 * the install is always sucessful.
651 smp_call_function_single(cpu, __perf_install_in_context,
656 counter->task = task;
658 task_oncpu_function_call(task, __perf_install_in_context,
661 spin_lock_irq(&ctx->lock);
663 * we need to retry the smp call.
665 if (ctx->is_active && list_empty(&counter->list_entry)) {
666 spin_unlock_irq(&ctx->lock);
671 * The lock prevents that this context is scheduled in so we
672 * can add the counter safely, if it the call above did not
675 if (list_empty(&counter->list_entry))
676 add_counter_to_ctx(counter, ctx);
677 spin_unlock_irq(&ctx->lock);
681 * Cross CPU call to enable a performance counter
683 static void __perf_counter_enable(void *info)
685 struct perf_counter *counter = info;
686 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
687 struct perf_counter_context *ctx = counter->ctx;
688 struct perf_counter *leader = counter->group_leader;
693 * If this is a per-task counter, need to check whether this
694 * counter's task is the current task on this cpu.
696 if (ctx->task && cpuctx->task_ctx != ctx)
699 spin_lock_irqsave(&ctx->lock, flags);
700 update_context_time(ctx);
702 counter->prev_state = counter->state;
703 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
705 counter->state = PERF_COUNTER_STATE_INACTIVE;
706 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
709 * If the counter is in a group and isn't the group leader,
710 * then don't put it on unless the group is on.
712 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
715 if (!group_can_go_on(counter, cpuctx, 1)) {
719 if (counter == leader)
720 err = group_sched_in(counter, cpuctx, ctx,
723 err = counter_sched_in(counter, cpuctx, ctx,
730 * If this counter can't go on and it's part of a
731 * group, then the whole group has to come off.
733 if (leader != counter)
734 group_sched_out(leader, cpuctx, ctx);
735 if (leader->hw_event.pinned) {
736 update_group_times(leader);
737 leader->state = PERF_COUNTER_STATE_ERROR;
742 spin_unlock_irqrestore(&ctx->lock, flags);
748 static void perf_counter_enable(struct perf_counter *counter)
750 struct perf_counter_context *ctx = counter->ctx;
751 struct task_struct *task = ctx->task;
755 * Enable the counter on the cpu that it's on
757 smp_call_function_single(counter->cpu, __perf_counter_enable,
762 spin_lock_irq(&ctx->lock);
763 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
767 * If the counter is in error state, clear that first.
768 * That way, if we see the counter in error state below, we
769 * know that it has gone back into error state, as distinct
770 * from the task having been scheduled away before the
771 * cross-call arrived.
773 if (counter->state == PERF_COUNTER_STATE_ERROR)
774 counter->state = PERF_COUNTER_STATE_OFF;
777 spin_unlock_irq(&ctx->lock);
778 task_oncpu_function_call(task, __perf_counter_enable, counter);
780 spin_lock_irq(&ctx->lock);
783 * If the context is active and the counter is still off,
784 * we need to retry the cross-call.
786 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
790 * Since we have the lock this context can't be scheduled
791 * in, so we can change the state safely.
793 if (counter->state == PERF_COUNTER_STATE_OFF) {
794 counter->state = PERF_COUNTER_STATE_INACTIVE;
795 counter->tstamp_enabled =
796 ctx->time - counter->total_time_enabled;
799 spin_unlock_irq(&ctx->lock);
802 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
805 * not supported on inherited counters
807 if (counter->hw_event.inherit)
810 atomic_add(refresh, &counter->event_limit);
811 perf_counter_enable(counter);
816 void __perf_counter_sched_out(struct perf_counter_context *ctx,
817 struct perf_cpu_context *cpuctx)
819 struct perf_counter *counter;
821 spin_lock(&ctx->lock);
823 if (likely(!ctx->nr_counters))
825 update_context_time(ctx);
828 if (ctx->nr_active) {
829 list_for_each_entry(counter, &ctx->counter_list, list_entry)
830 group_sched_out(counter, cpuctx, ctx);
834 spin_unlock(&ctx->lock);
838 * Called from scheduler to remove the counters of the current task,
839 * with interrupts disabled.
841 * We stop each counter and update the counter value in counter->count.
843 * This does not protect us against NMI, but disable()
844 * sets the disabled bit in the control field of counter _before_
845 * accessing the counter control register. If a NMI hits, then it will
846 * not restart the counter.
848 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
850 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
851 struct perf_counter_context *ctx = &task->perf_counter_ctx;
852 struct pt_regs *regs;
854 if (likely(!cpuctx->task_ctx))
857 update_context_time(ctx);
859 regs = task_pt_regs(task);
860 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
861 __perf_counter_sched_out(ctx, cpuctx);
863 cpuctx->task_ctx = NULL;
866 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
868 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
870 __perf_counter_sched_out(ctx, cpuctx);
871 cpuctx->task_ctx = NULL;
874 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
876 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
880 __perf_counter_sched_in(struct perf_counter_context *ctx,
881 struct perf_cpu_context *cpuctx, int cpu)
883 struct perf_counter *counter;
886 spin_lock(&ctx->lock);
888 if (likely(!ctx->nr_counters))
891 ctx->timestamp = perf_clock();
896 * First go through the list and put on any pinned groups
897 * in order to give them the best chance of going on.
899 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
900 if (counter->state <= PERF_COUNTER_STATE_OFF ||
901 !counter->hw_event.pinned)
903 if (counter->cpu != -1 && counter->cpu != cpu)
906 if (group_can_go_on(counter, cpuctx, 1))
907 group_sched_in(counter, cpuctx, ctx, cpu);
910 * If this pinned group hasn't been scheduled,
911 * put it in error state.
913 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
914 update_group_times(counter);
915 counter->state = PERF_COUNTER_STATE_ERROR;
919 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
921 * Ignore counters in OFF or ERROR state, and
922 * ignore pinned counters since we did them already.
924 if (counter->state <= PERF_COUNTER_STATE_OFF ||
925 counter->hw_event.pinned)
929 * Listen to the 'cpu' scheduling filter constraint
932 if (counter->cpu != -1 && counter->cpu != cpu)
935 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
936 if (group_sched_in(counter, cpuctx, ctx, cpu))
942 spin_unlock(&ctx->lock);
946 * Called from scheduler to add the counters of the current task
947 * with interrupts disabled.
949 * We restore the counter value and then enable it.
951 * This does not protect us against NMI, but enable()
952 * sets the enabled bit in the control field of counter _before_
953 * accessing the counter control register. If a NMI hits, then it will
954 * keep the counter running.
956 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
958 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
959 struct perf_counter_context *ctx = &task->perf_counter_ctx;
961 __perf_counter_sched_in(ctx, cpuctx, cpu);
962 cpuctx->task_ctx = ctx;
965 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
967 struct perf_counter_context *ctx = &cpuctx->ctx;
969 __perf_counter_sched_in(ctx, cpuctx, cpu);
972 int perf_counter_task_disable(void)
974 struct task_struct *curr = current;
975 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
976 struct perf_counter *counter;
979 if (likely(!ctx->nr_counters))
982 local_irq_save(flags);
984 __perf_counter_task_sched_out(ctx);
986 spin_lock(&ctx->lock);
989 * Disable all the counters:
993 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
994 if (counter->state != PERF_COUNTER_STATE_ERROR) {
995 update_group_times(counter);
996 counter->state = PERF_COUNTER_STATE_OFF;
1002 spin_unlock_irqrestore(&ctx->lock, flags);
1007 int perf_counter_task_enable(void)
1009 struct task_struct *curr = current;
1010 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1011 struct perf_counter *counter;
1012 unsigned long flags;
1015 if (likely(!ctx->nr_counters))
1018 local_irq_save(flags);
1019 cpu = smp_processor_id();
1021 __perf_counter_task_sched_out(ctx);
1023 spin_lock(&ctx->lock);
1026 * Disable all the counters:
1030 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1031 if (counter->state > PERF_COUNTER_STATE_OFF)
1033 counter->state = PERF_COUNTER_STATE_INACTIVE;
1034 counter->tstamp_enabled =
1035 ctx->time - counter->total_time_enabled;
1036 counter->hw_event.disabled = 0;
1040 spin_unlock(&ctx->lock);
1042 perf_counter_task_sched_in(curr, cpu);
1044 local_irq_restore(flags);
1050 * Round-robin a context's counters:
1052 static void rotate_ctx(struct perf_counter_context *ctx)
1054 struct perf_counter *counter;
1056 if (!ctx->nr_counters)
1059 spin_lock(&ctx->lock);
1061 * Rotate the first entry last (works just fine for group counters too):
1064 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1065 list_move_tail(&counter->list_entry, &ctx->counter_list);
1070 spin_unlock(&ctx->lock);
1073 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1075 struct perf_cpu_context *cpuctx;
1076 struct perf_counter_context *ctx;
1078 if (!atomic_read(&nr_counters))
1081 cpuctx = &per_cpu(perf_cpu_context, cpu);
1082 ctx = &curr->perf_counter_ctx;
1084 perf_counter_cpu_sched_out(cpuctx);
1085 __perf_counter_task_sched_out(ctx);
1087 rotate_ctx(&cpuctx->ctx);
1090 perf_counter_cpu_sched_in(cpuctx, cpu);
1091 perf_counter_task_sched_in(curr, cpu);
1095 * Cross CPU call to read the hardware counter
1097 static void __read(void *info)
1099 struct perf_counter *counter = info;
1100 struct perf_counter_context *ctx = counter->ctx;
1101 unsigned long flags;
1103 local_irq_save(flags);
1105 update_context_time(ctx);
1106 counter->pmu->read(counter);
1107 update_counter_times(counter);
1108 local_irq_restore(flags);
1111 static u64 perf_counter_read(struct perf_counter *counter)
1114 * If counter is enabled and currently active on a CPU, update the
1115 * value in the counter structure:
1117 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1118 smp_call_function_single(counter->oncpu,
1119 __read, counter, 1);
1120 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1121 update_counter_times(counter);
1124 return atomic64_read(&counter->count);
1127 static void put_context(struct perf_counter_context *ctx)
1130 put_task_struct(ctx->task);
1133 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1135 struct perf_cpu_context *cpuctx;
1136 struct perf_counter_context *ctx;
1137 struct task_struct *task;
1140 * If cpu is not a wildcard then this is a percpu counter:
1143 /* Must be root to operate on a CPU counter: */
1144 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1145 return ERR_PTR(-EACCES);
1147 if (cpu < 0 || cpu > num_possible_cpus())
1148 return ERR_PTR(-EINVAL);
1151 * We could be clever and allow to attach a counter to an
1152 * offline CPU and activate it when the CPU comes up, but
1155 if (!cpu_isset(cpu, cpu_online_map))
1156 return ERR_PTR(-ENODEV);
1158 cpuctx = &per_cpu(perf_cpu_context, cpu);
1168 task = find_task_by_vpid(pid);
1170 get_task_struct(task);
1174 return ERR_PTR(-ESRCH);
1176 ctx = &task->perf_counter_ctx;
1179 /* Reuse ptrace permission checks for now. */
1180 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1182 return ERR_PTR(-EACCES);
1188 static void free_counter_rcu(struct rcu_head *head)
1190 struct perf_counter *counter;
1192 counter = container_of(head, struct perf_counter, rcu_head);
1196 static void perf_pending_sync(struct perf_counter *counter);
1198 static void free_counter(struct perf_counter *counter)
1200 perf_pending_sync(counter);
1202 atomic_dec(&nr_counters);
1203 if (counter->hw_event.mmap)
1204 atomic_dec(&nr_mmap_tracking);
1205 if (counter->hw_event.munmap)
1206 atomic_dec(&nr_munmap_tracking);
1207 if (counter->hw_event.comm)
1208 atomic_dec(&nr_comm_tracking);
1210 if (counter->destroy)
1211 counter->destroy(counter);
1213 call_rcu(&counter->rcu_head, free_counter_rcu);
1217 * Called when the last reference to the file is gone.
1219 static int perf_release(struct inode *inode, struct file *file)
1221 struct perf_counter *counter = file->private_data;
1222 struct perf_counter_context *ctx = counter->ctx;
1224 file->private_data = NULL;
1226 mutex_lock(&ctx->mutex);
1227 mutex_lock(&counter->mutex);
1229 perf_counter_remove_from_context(counter);
1231 mutex_unlock(&counter->mutex);
1232 mutex_unlock(&ctx->mutex);
1234 free_counter(counter);
1241 * Read the performance counter - simple non blocking version for now
1244 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1250 * Return end-of-file for a read on a counter that is in
1251 * error state (i.e. because it was pinned but it couldn't be
1252 * scheduled on to the CPU at some point).
1254 if (counter->state == PERF_COUNTER_STATE_ERROR)
1257 mutex_lock(&counter->mutex);
1258 values[0] = perf_counter_read(counter);
1260 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1261 values[n++] = counter->total_time_enabled +
1262 atomic64_read(&counter->child_total_time_enabled);
1263 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1264 values[n++] = counter->total_time_running +
1265 atomic64_read(&counter->child_total_time_running);
1266 mutex_unlock(&counter->mutex);
1268 if (count < n * sizeof(u64))
1270 count = n * sizeof(u64);
1272 if (copy_to_user(buf, values, count))
1279 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1281 struct perf_counter *counter = file->private_data;
1283 return perf_read_hw(counter, buf, count);
1286 static unsigned int perf_poll(struct file *file, poll_table *wait)
1288 struct perf_counter *counter = file->private_data;
1289 struct perf_mmap_data *data;
1290 unsigned int events = POLL_HUP;
1293 data = rcu_dereference(counter->data);
1295 events = atomic_xchg(&data->poll, 0);
1298 poll_wait(file, &counter->waitq, wait);
1303 static void perf_counter_reset(struct perf_counter *counter)
1305 (void)perf_counter_read(counter);
1306 atomic64_set(&counter->count, 0);
1307 perf_counter_update_userpage(counter);
1310 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1311 void (*func)(struct perf_counter *))
1313 struct perf_counter_context *ctx = counter->ctx;
1314 struct perf_counter *sibling;
1316 spin_lock_irq(&ctx->lock);
1317 counter = counter->group_leader;
1320 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1322 spin_unlock_irq(&ctx->lock);
1325 static void perf_counter_for_each_child(struct perf_counter *counter,
1326 void (*func)(struct perf_counter *))
1328 struct perf_counter *child;
1330 mutex_lock(&counter->mutex);
1332 list_for_each_entry(child, &counter->child_list, child_list)
1334 mutex_unlock(&counter->mutex);
1337 static void perf_counter_for_each(struct perf_counter *counter,
1338 void (*func)(struct perf_counter *))
1340 struct perf_counter *child;
1342 mutex_lock(&counter->mutex);
1343 perf_counter_for_each_sibling(counter, func);
1344 list_for_each_entry(child, &counter->child_list, child_list)
1345 perf_counter_for_each_sibling(child, func);
1346 mutex_unlock(&counter->mutex);
1349 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1351 struct perf_counter *counter = file->private_data;
1352 void (*func)(struct perf_counter *);
1356 case PERF_COUNTER_IOC_ENABLE:
1357 func = perf_counter_enable;
1359 case PERF_COUNTER_IOC_DISABLE:
1360 func = perf_counter_disable;
1362 case PERF_COUNTER_IOC_RESET:
1363 func = perf_counter_reset;
1366 case PERF_COUNTER_IOC_REFRESH:
1367 return perf_counter_refresh(counter, arg);
1372 if (flags & PERF_IOC_FLAG_GROUP)
1373 perf_counter_for_each(counter, func);
1375 perf_counter_for_each_child(counter, func);
1381 * Callers need to ensure there can be no nesting of this function, otherwise
1382 * the seqlock logic goes bad. We can not serialize this because the arch
1383 * code calls this from NMI context.
1385 void perf_counter_update_userpage(struct perf_counter *counter)
1387 struct perf_mmap_data *data;
1388 struct perf_counter_mmap_page *userpg;
1391 data = rcu_dereference(counter->data);
1395 userpg = data->user_page;
1398 * Disable preemption so as to not let the corresponding user-space
1399 * spin too long if we get preempted.
1404 userpg->index = counter->hw.idx;
1405 userpg->offset = atomic64_read(&counter->count);
1406 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1407 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1416 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1418 struct perf_counter *counter = vma->vm_file->private_data;
1419 struct perf_mmap_data *data;
1420 int ret = VM_FAULT_SIGBUS;
1423 data = rcu_dereference(counter->data);
1427 if (vmf->pgoff == 0) {
1428 vmf->page = virt_to_page(data->user_page);
1430 int nr = vmf->pgoff - 1;
1432 if ((unsigned)nr > data->nr_pages)
1435 vmf->page = virt_to_page(data->data_pages[nr]);
1437 get_page(vmf->page);
1445 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1447 struct perf_mmap_data *data;
1451 WARN_ON(atomic_read(&counter->mmap_count));
1453 size = sizeof(struct perf_mmap_data);
1454 size += nr_pages * sizeof(void *);
1456 data = kzalloc(size, GFP_KERNEL);
1460 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1461 if (!data->user_page)
1462 goto fail_user_page;
1464 for (i = 0; i < nr_pages; i++) {
1465 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1466 if (!data->data_pages[i])
1467 goto fail_data_pages;
1470 data->nr_pages = nr_pages;
1471 atomic_set(&data->lock, -1);
1473 rcu_assign_pointer(counter->data, data);
1478 for (i--; i >= 0; i--)
1479 free_page((unsigned long)data->data_pages[i]);
1481 free_page((unsigned long)data->user_page);
1490 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1492 struct perf_mmap_data *data = container_of(rcu_head,
1493 struct perf_mmap_data, rcu_head);
1496 free_page((unsigned long)data->user_page);
1497 for (i = 0; i < data->nr_pages; i++)
1498 free_page((unsigned long)data->data_pages[i]);
1502 static void perf_mmap_data_free(struct perf_counter *counter)
1504 struct perf_mmap_data *data = counter->data;
1506 WARN_ON(atomic_read(&counter->mmap_count));
1508 rcu_assign_pointer(counter->data, NULL);
1509 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1512 static void perf_mmap_open(struct vm_area_struct *vma)
1514 struct perf_counter *counter = vma->vm_file->private_data;
1516 atomic_inc(&counter->mmap_count);
1519 static void perf_mmap_close(struct vm_area_struct *vma)
1521 struct perf_counter *counter = vma->vm_file->private_data;
1523 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1524 &counter->mmap_mutex)) {
1525 struct user_struct *user = current_user();
1527 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
1528 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1529 perf_mmap_data_free(counter);
1530 mutex_unlock(&counter->mmap_mutex);
1534 static struct vm_operations_struct perf_mmap_vmops = {
1535 .open = perf_mmap_open,
1536 .close = perf_mmap_close,
1537 .fault = perf_mmap_fault,
1540 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1542 struct perf_counter *counter = file->private_data;
1543 struct user_struct *user = current_user();
1544 unsigned long vma_size;
1545 unsigned long nr_pages;
1546 unsigned long user_locked, user_lock_limit;
1547 unsigned long locked, lock_limit;
1548 long user_extra, extra;
1551 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1554 vma_size = vma->vm_end - vma->vm_start;
1555 nr_pages = (vma_size / PAGE_SIZE) - 1;
1558 * If we have data pages ensure they're a power-of-two number, so we
1559 * can do bitmasks instead of modulo.
1561 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1564 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1567 if (vma->vm_pgoff != 0)
1570 mutex_lock(&counter->mmap_mutex);
1571 if (atomic_inc_not_zero(&counter->mmap_count)) {
1572 if (nr_pages != counter->data->nr_pages)
1577 user_extra = nr_pages + 1;
1578 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1579 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
1582 if (user_locked > user_lock_limit)
1583 extra = user_locked - user_lock_limit;
1585 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1586 lock_limit >>= PAGE_SHIFT;
1587 locked = vma->vm_mm->locked_vm + extra;
1589 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1594 WARN_ON(counter->data);
1595 ret = perf_mmap_data_alloc(counter, nr_pages);
1599 atomic_set(&counter->mmap_count, 1);
1600 atomic_long_add(user_extra, &user->locked_vm);
1601 vma->vm_mm->locked_vm += extra;
1602 counter->data->nr_locked = extra;
1604 mutex_unlock(&counter->mmap_mutex);
1606 vma->vm_flags &= ~VM_MAYWRITE;
1607 vma->vm_flags |= VM_RESERVED;
1608 vma->vm_ops = &perf_mmap_vmops;
1613 static int perf_fasync(int fd, struct file *filp, int on)
1615 struct perf_counter *counter = filp->private_data;
1616 struct inode *inode = filp->f_path.dentry->d_inode;
1619 mutex_lock(&inode->i_mutex);
1620 retval = fasync_helper(fd, filp, on, &counter->fasync);
1621 mutex_unlock(&inode->i_mutex);
1629 static const struct file_operations perf_fops = {
1630 .release = perf_release,
1633 .unlocked_ioctl = perf_ioctl,
1634 .compat_ioctl = perf_ioctl,
1636 .fasync = perf_fasync,
1640 * Perf counter wakeup
1642 * If there's data, ensure we set the poll() state and publish everything
1643 * to user-space before waking everybody up.
1646 void perf_counter_wakeup(struct perf_counter *counter)
1648 wake_up_all(&counter->waitq);
1650 if (counter->pending_kill) {
1651 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1652 counter->pending_kill = 0;
1659 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1661 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1662 * single linked list and use cmpxchg() to add entries lockless.
1665 static void perf_pending_counter(struct perf_pending_entry *entry)
1667 struct perf_counter *counter = container_of(entry,
1668 struct perf_counter, pending);
1670 if (counter->pending_disable) {
1671 counter->pending_disable = 0;
1672 perf_counter_disable(counter);
1675 if (counter->pending_wakeup) {
1676 counter->pending_wakeup = 0;
1677 perf_counter_wakeup(counter);
1681 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1683 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1687 static void perf_pending_queue(struct perf_pending_entry *entry,
1688 void (*func)(struct perf_pending_entry *))
1690 struct perf_pending_entry **head;
1692 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1697 head = &get_cpu_var(perf_pending_head);
1700 entry->next = *head;
1701 } while (cmpxchg(head, entry->next, entry) != entry->next);
1703 set_perf_counter_pending();
1705 put_cpu_var(perf_pending_head);
1708 static int __perf_pending_run(void)
1710 struct perf_pending_entry *list;
1713 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1714 while (list != PENDING_TAIL) {
1715 void (*func)(struct perf_pending_entry *);
1716 struct perf_pending_entry *entry = list;
1723 * Ensure we observe the unqueue before we issue the wakeup,
1724 * so that we won't be waiting forever.
1725 * -- see perf_not_pending().
1736 static inline int perf_not_pending(struct perf_counter *counter)
1739 * If we flush on whatever cpu we run, there is a chance we don't
1743 __perf_pending_run();
1747 * Ensure we see the proper queue state before going to sleep
1748 * so that we do not miss the wakeup. -- see perf_pending_handle()
1751 return counter->pending.next == NULL;
1754 static void perf_pending_sync(struct perf_counter *counter)
1756 wait_event(counter->waitq, perf_not_pending(counter));
1759 void perf_counter_do_pending(void)
1761 __perf_pending_run();
1765 * Callchain support -- arch specific
1768 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1777 struct perf_output_handle {
1778 struct perf_counter *counter;
1779 struct perf_mmap_data *data;
1780 unsigned int offset;
1785 unsigned long flags;
1788 static void perf_output_wakeup(struct perf_output_handle *handle)
1790 atomic_set(&handle->data->poll, POLL_IN);
1793 handle->counter->pending_wakeup = 1;
1794 perf_pending_queue(&handle->counter->pending,
1795 perf_pending_counter);
1797 perf_counter_wakeup(handle->counter);
1801 * Curious locking construct.
1803 * We need to ensure a later event doesn't publish a head when a former
1804 * event isn't done writing. However since we need to deal with NMIs we
1805 * cannot fully serialize things.
1807 * What we do is serialize between CPUs so we only have to deal with NMI
1808 * nesting on a single CPU.
1810 * We only publish the head (and generate a wakeup) when the outer-most
1813 static void perf_output_lock(struct perf_output_handle *handle)
1815 struct perf_mmap_data *data = handle->data;
1820 local_irq_save(handle->flags);
1821 cpu = smp_processor_id();
1823 if (in_nmi() && atomic_read(&data->lock) == cpu)
1826 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1832 static void perf_output_unlock(struct perf_output_handle *handle)
1834 struct perf_mmap_data *data = handle->data;
1837 data->done_head = data->head;
1839 if (!handle->locked)
1844 * The xchg implies a full barrier that ensures all writes are done
1845 * before we publish the new head, matched by a rmb() in userspace when
1846 * reading this position.
1848 while ((head = atomic_xchg(&data->done_head, 0)))
1849 data->user_page->data_head = head;
1852 * NMI can happen here, which means we can miss a done_head update.
1855 cpu = atomic_xchg(&data->lock, -1);
1856 WARN_ON_ONCE(cpu != smp_processor_id());
1859 * Therefore we have to validate we did not indeed do so.
1861 if (unlikely(atomic_read(&data->done_head))) {
1863 * Since we had it locked, we can lock it again.
1865 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1871 if (atomic_xchg(&data->wakeup, 0))
1872 perf_output_wakeup(handle);
1874 local_irq_restore(handle->flags);
1877 static int perf_output_begin(struct perf_output_handle *handle,
1878 struct perf_counter *counter, unsigned int size,
1879 int nmi, int overflow)
1881 struct perf_mmap_data *data;
1882 unsigned int offset, head;
1885 * For inherited counters we send all the output towards the parent.
1887 if (counter->parent)
1888 counter = counter->parent;
1891 data = rcu_dereference(counter->data);
1895 handle->data = data;
1896 handle->counter = counter;
1898 handle->overflow = overflow;
1900 if (!data->nr_pages)
1903 perf_output_lock(handle);
1906 offset = head = atomic_read(&data->head);
1908 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1910 handle->offset = offset;
1911 handle->head = head;
1913 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
1914 atomic_set(&data->wakeup, 1);
1919 perf_output_wakeup(handle);
1926 static void perf_output_copy(struct perf_output_handle *handle,
1927 void *buf, unsigned int len)
1929 unsigned int pages_mask;
1930 unsigned int offset;
1934 offset = handle->offset;
1935 pages_mask = handle->data->nr_pages - 1;
1936 pages = handle->data->data_pages;
1939 unsigned int page_offset;
1942 nr = (offset >> PAGE_SHIFT) & pages_mask;
1943 page_offset = offset & (PAGE_SIZE - 1);
1944 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1946 memcpy(pages[nr] + page_offset, buf, size);
1953 handle->offset = offset;
1956 * Check we didn't copy past our reservation window, taking the
1957 * possible unsigned int wrap into account.
1959 WARN_ON_ONCE(((int)(handle->head - handle->offset)) < 0);
1962 #define perf_output_put(handle, x) \
1963 perf_output_copy((handle), &(x), sizeof(x))
1965 static void perf_output_end(struct perf_output_handle *handle)
1967 struct perf_counter *counter = handle->counter;
1968 struct perf_mmap_data *data = handle->data;
1970 int wakeup_events = counter->hw_event.wakeup_events;
1972 if (handle->overflow && wakeup_events) {
1973 int events = atomic_inc_return(&data->events);
1974 if (events >= wakeup_events) {
1975 atomic_sub(wakeup_events, &data->events);
1976 atomic_set(&data->wakeup, 1);
1980 perf_output_unlock(handle);
1984 static void perf_counter_output(struct perf_counter *counter,
1985 int nmi, struct pt_regs *regs, u64 addr)
1988 u64 record_type = counter->hw_event.record_type;
1989 struct perf_output_handle handle;
1990 struct perf_event_header header;
1999 struct perf_callchain_entry *callchain = NULL;
2000 int callchain_size = 0;
2007 header.size = sizeof(header);
2009 header.misc = PERF_EVENT_MISC_OVERFLOW;
2010 header.misc |= user_mode(regs) ?
2011 PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
2013 if (record_type & PERF_RECORD_IP) {
2014 ip = instruction_pointer(regs);
2015 header.type |= PERF_RECORD_IP;
2016 header.size += sizeof(ip);
2019 if (record_type & PERF_RECORD_TID) {
2020 /* namespace issues */
2021 tid_entry.pid = current->group_leader->pid;
2022 tid_entry.tid = current->pid;
2024 header.type |= PERF_RECORD_TID;
2025 header.size += sizeof(tid_entry);
2028 if (record_type & PERF_RECORD_TIME) {
2030 * Maybe do better on x86 and provide cpu_clock_nmi()
2032 time = sched_clock();
2034 header.type |= PERF_RECORD_TIME;
2035 header.size += sizeof(u64);
2038 if (record_type & PERF_RECORD_ADDR) {
2039 header.type |= PERF_RECORD_ADDR;
2040 header.size += sizeof(u64);
2043 if (record_type & PERF_RECORD_CONFIG) {
2044 header.type |= PERF_RECORD_CONFIG;
2045 header.size += sizeof(u64);
2048 if (record_type & PERF_RECORD_CPU) {
2049 header.type |= PERF_RECORD_CPU;
2050 header.size += sizeof(cpu_entry);
2052 cpu_entry.cpu = raw_smp_processor_id();
2055 if (record_type & PERF_RECORD_GROUP) {
2056 header.type |= PERF_RECORD_GROUP;
2057 header.size += sizeof(u64) +
2058 counter->nr_siblings * sizeof(group_entry);
2061 if (record_type & PERF_RECORD_CALLCHAIN) {
2062 callchain = perf_callchain(regs);
2065 callchain_size = (1 + callchain->nr) * sizeof(u64);
2067 header.type |= PERF_RECORD_CALLCHAIN;
2068 header.size += callchain_size;
2072 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2076 perf_output_put(&handle, header);
2078 if (record_type & PERF_RECORD_IP)
2079 perf_output_put(&handle, ip);
2081 if (record_type & PERF_RECORD_TID)
2082 perf_output_put(&handle, tid_entry);
2084 if (record_type & PERF_RECORD_TIME)
2085 perf_output_put(&handle, time);
2087 if (record_type & PERF_RECORD_ADDR)
2088 perf_output_put(&handle, addr);
2090 if (record_type & PERF_RECORD_CONFIG)
2091 perf_output_put(&handle, counter->hw_event.config);
2093 if (record_type & PERF_RECORD_CPU)
2094 perf_output_put(&handle, cpu_entry);
2097 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2099 if (record_type & PERF_RECORD_GROUP) {
2100 struct perf_counter *leader, *sub;
2101 u64 nr = counter->nr_siblings;
2103 perf_output_put(&handle, nr);
2105 leader = counter->group_leader;
2106 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2108 sub->pmu->read(sub);
2110 group_entry.event = sub->hw_event.config;
2111 group_entry.counter = atomic64_read(&sub->count);
2113 perf_output_put(&handle, group_entry);
2118 perf_output_copy(&handle, callchain, callchain_size);
2120 perf_output_end(&handle);
2127 struct perf_comm_event {
2128 struct task_struct *task;
2133 struct perf_event_header header;
2140 static void perf_counter_comm_output(struct perf_counter *counter,
2141 struct perf_comm_event *comm_event)
2143 struct perf_output_handle handle;
2144 int size = comm_event->event.header.size;
2145 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2150 perf_output_put(&handle, comm_event->event);
2151 perf_output_copy(&handle, comm_event->comm,
2152 comm_event->comm_size);
2153 perf_output_end(&handle);
2156 static int perf_counter_comm_match(struct perf_counter *counter,
2157 struct perf_comm_event *comm_event)
2159 if (counter->hw_event.comm &&
2160 comm_event->event.header.type == PERF_EVENT_COMM)
2166 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2167 struct perf_comm_event *comm_event)
2169 struct perf_counter *counter;
2171 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2175 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2176 if (perf_counter_comm_match(counter, comm_event))
2177 perf_counter_comm_output(counter, comm_event);
2182 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2184 struct perf_cpu_context *cpuctx;
2186 char *comm = comm_event->task->comm;
2188 size = ALIGN(strlen(comm)+1, sizeof(u64));
2190 comm_event->comm = comm;
2191 comm_event->comm_size = size;
2193 comm_event->event.header.size = sizeof(comm_event->event) + size;
2195 cpuctx = &get_cpu_var(perf_cpu_context);
2196 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2197 put_cpu_var(perf_cpu_context);
2199 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
2202 void perf_counter_comm(struct task_struct *task)
2204 struct perf_comm_event comm_event;
2206 if (!atomic_read(&nr_comm_tracking))
2209 comm_event = (struct perf_comm_event){
2212 .header = { .type = PERF_EVENT_COMM, },
2213 .pid = task->group_leader->pid,
2218 perf_counter_comm_event(&comm_event);
2225 struct perf_mmap_event {
2231 struct perf_event_header header;
2241 static void perf_counter_mmap_output(struct perf_counter *counter,
2242 struct perf_mmap_event *mmap_event)
2244 struct perf_output_handle handle;
2245 int size = mmap_event->event.header.size;
2246 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2251 perf_output_put(&handle, mmap_event->event);
2252 perf_output_copy(&handle, mmap_event->file_name,
2253 mmap_event->file_size);
2254 perf_output_end(&handle);
2257 static int perf_counter_mmap_match(struct perf_counter *counter,
2258 struct perf_mmap_event *mmap_event)
2260 if (counter->hw_event.mmap &&
2261 mmap_event->event.header.type == PERF_EVENT_MMAP)
2264 if (counter->hw_event.munmap &&
2265 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2271 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2272 struct perf_mmap_event *mmap_event)
2274 struct perf_counter *counter;
2276 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2280 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2281 if (perf_counter_mmap_match(counter, mmap_event))
2282 perf_counter_mmap_output(counter, mmap_event);
2287 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2289 struct perf_cpu_context *cpuctx;
2290 struct file *file = mmap_event->file;
2297 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2299 name = strncpy(tmp, "//enomem", sizeof(tmp));
2302 name = d_path(&file->f_path, buf, PATH_MAX);
2304 name = strncpy(tmp, "//toolong", sizeof(tmp));
2308 name = strncpy(tmp, "//anon", sizeof(tmp));
2313 size = ALIGN(strlen(name)+1, sizeof(u64));
2315 mmap_event->file_name = name;
2316 mmap_event->file_size = size;
2318 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2320 cpuctx = &get_cpu_var(perf_cpu_context);
2321 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2322 put_cpu_var(perf_cpu_context);
2324 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2329 void perf_counter_mmap(unsigned long addr, unsigned long len,
2330 unsigned long pgoff, struct file *file)
2332 struct perf_mmap_event mmap_event;
2334 if (!atomic_read(&nr_mmap_tracking))
2337 mmap_event = (struct perf_mmap_event){
2340 .header = { .type = PERF_EVENT_MMAP, },
2341 .pid = current->group_leader->pid,
2342 .tid = current->pid,
2349 perf_counter_mmap_event(&mmap_event);
2352 void perf_counter_munmap(unsigned long addr, unsigned long len,
2353 unsigned long pgoff, struct file *file)
2355 struct perf_mmap_event mmap_event;
2357 if (!atomic_read(&nr_munmap_tracking))
2360 mmap_event = (struct perf_mmap_event){
2363 .header = { .type = PERF_EVENT_MUNMAP, },
2364 .pid = current->group_leader->pid,
2365 .tid = current->pid,
2372 perf_counter_mmap_event(&mmap_event);
2376 * Generic counter overflow handling.
2379 int perf_counter_overflow(struct perf_counter *counter,
2380 int nmi, struct pt_regs *regs, u64 addr)
2382 int events = atomic_read(&counter->event_limit);
2386 * XXX event_limit might not quite work as expected on inherited
2390 counter->pending_kill = POLL_IN;
2391 if (events && atomic_dec_and_test(&counter->event_limit)) {
2393 counter->pending_kill = POLL_HUP;
2395 counter->pending_disable = 1;
2396 perf_pending_queue(&counter->pending,
2397 perf_pending_counter);
2399 perf_counter_disable(counter);
2402 perf_counter_output(counter, nmi, regs, addr);
2407 * Generic software counter infrastructure
2410 static void perf_swcounter_update(struct perf_counter *counter)
2412 struct hw_perf_counter *hwc = &counter->hw;
2417 prev = atomic64_read(&hwc->prev_count);
2418 now = atomic64_read(&hwc->count);
2419 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2424 atomic64_add(delta, &counter->count);
2425 atomic64_sub(delta, &hwc->period_left);
2428 static void perf_swcounter_set_period(struct perf_counter *counter)
2430 struct hw_perf_counter *hwc = &counter->hw;
2431 s64 left = atomic64_read(&hwc->period_left);
2432 s64 period = hwc->irq_period;
2434 if (unlikely(left <= -period)) {
2436 atomic64_set(&hwc->period_left, left);
2439 if (unlikely(left <= 0)) {
2441 atomic64_add(period, &hwc->period_left);
2444 atomic64_set(&hwc->prev_count, -left);
2445 atomic64_set(&hwc->count, -left);
2448 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2450 enum hrtimer_restart ret = HRTIMER_RESTART;
2451 struct perf_counter *counter;
2452 struct pt_regs *regs;
2454 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2455 counter->pmu->read(counter);
2457 regs = get_irq_regs();
2459 * In case we exclude kernel IPs or are somehow not in interrupt
2460 * context, provide the next best thing, the user IP.
2462 if ((counter->hw_event.exclude_kernel || !regs) &&
2463 !counter->hw_event.exclude_user)
2464 regs = task_pt_regs(current);
2467 if (perf_counter_overflow(counter, 0, regs, 0))
2468 ret = HRTIMER_NORESTART;
2471 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2476 static void perf_swcounter_overflow(struct perf_counter *counter,
2477 int nmi, struct pt_regs *regs, u64 addr)
2479 perf_swcounter_update(counter);
2480 perf_swcounter_set_period(counter);
2481 if (perf_counter_overflow(counter, nmi, regs, addr))
2482 /* soft-disable the counter */
2487 static int perf_swcounter_match(struct perf_counter *counter,
2488 enum perf_event_types type,
2489 u32 event, struct pt_regs *regs)
2491 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2494 if (perf_event_raw(&counter->hw_event))
2497 if (perf_event_type(&counter->hw_event) != type)
2500 if (perf_event_id(&counter->hw_event) != event)
2503 if (counter->hw_event.exclude_user && user_mode(regs))
2506 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2512 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2513 int nmi, struct pt_regs *regs, u64 addr)
2515 int neg = atomic64_add_negative(nr, &counter->hw.count);
2516 if (counter->hw.irq_period && !neg)
2517 perf_swcounter_overflow(counter, nmi, regs, addr);
2520 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2521 enum perf_event_types type, u32 event,
2522 u64 nr, int nmi, struct pt_regs *regs,
2525 struct perf_counter *counter;
2527 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2531 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2532 if (perf_swcounter_match(counter, type, event, regs))
2533 perf_swcounter_add(counter, nr, nmi, regs, addr);
2538 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2541 return &cpuctx->recursion[3];
2544 return &cpuctx->recursion[2];
2547 return &cpuctx->recursion[1];
2549 return &cpuctx->recursion[0];
2552 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2553 u64 nr, int nmi, struct pt_regs *regs,
2556 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2557 int *recursion = perf_swcounter_recursion_context(cpuctx);
2565 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2566 nr, nmi, regs, addr);
2567 if (cpuctx->task_ctx) {
2568 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2569 nr, nmi, regs, addr);
2576 put_cpu_var(perf_cpu_context);
2580 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2582 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2585 static void perf_swcounter_read(struct perf_counter *counter)
2587 perf_swcounter_update(counter);
2590 static int perf_swcounter_enable(struct perf_counter *counter)
2592 perf_swcounter_set_period(counter);
2596 static void perf_swcounter_disable(struct perf_counter *counter)
2598 perf_swcounter_update(counter);
2601 static const struct pmu perf_ops_generic = {
2602 .enable = perf_swcounter_enable,
2603 .disable = perf_swcounter_disable,
2604 .read = perf_swcounter_read,
2608 * Software counter: cpu wall time clock
2611 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2613 int cpu = raw_smp_processor_id();
2617 now = cpu_clock(cpu);
2618 prev = atomic64_read(&counter->hw.prev_count);
2619 atomic64_set(&counter->hw.prev_count, now);
2620 atomic64_add(now - prev, &counter->count);
2623 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2625 struct hw_perf_counter *hwc = &counter->hw;
2626 int cpu = raw_smp_processor_id();
2628 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2629 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2630 hwc->hrtimer.function = perf_swcounter_hrtimer;
2631 if (hwc->irq_period) {
2632 __hrtimer_start_range_ns(&hwc->hrtimer,
2633 ns_to_ktime(hwc->irq_period), 0,
2634 HRTIMER_MODE_REL, 0);
2640 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2642 hrtimer_cancel(&counter->hw.hrtimer);
2643 cpu_clock_perf_counter_update(counter);
2646 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2648 cpu_clock_perf_counter_update(counter);
2651 static const struct pmu perf_ops_cpu_clock = {
2652 .enable = cpu_clock_perf_counter_enable,
2653 .disable = cpu_clock_perf_counter_disable,
2654 .read = cpu_clock_perf_counter_read,
2658 * Software counter: task time clock
2661 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2666 prev = atomic64_xchg(&counter->hw.prev_count, now);
2668 atomic64_add(delta, &counter->count);
2671 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2673 struct hw_perf_counter *hwc = &counter->hw;
2676 now = counter->ctx->time;
2678 atomic64_set(&hwc->prev_count, now);
2679 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2680 hwc->hrtimer.function = perf_swcounter_hrtimer;
2681 if (hwc->irq_period) {
2682 __hrtimer_start_range_ns(&hwc->hrtimer,
2683 ns_to_ktime(hwc->irq_period), 0,
2684 HRTIMER_MODE_REL, 0);
2690 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2692 hrtimer_cancel(&counter->hw.hrtimer);
2693 task_clock_perf_counter_update(counter, counter->ctx->time);
2697 static void task_clock_perf_counter_read(struct perf_counter *counter)
2702 update_context_time(counter->ctx);
2703 time = counter->ctx->time;
2705 u64 now = perf_clock();
2706 u64 delta = now - counter->ctx->timestamp;
2707 time = counter->ctx->time + delta;
2710 task_clock_perf_counter_update(counter, time);
2713 static const struct pmu perf_ops_task_clock = {
2714 .enable = task_clock_perf_counter_enable,
2715 .disable = task_clock_perf_counter_disable,
2716 .read = task_clock_perf_counter_read,
2720 * Software counter: cpu migrations
2723 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2725 struct task_struct *curr = counter->ctx->task;
2728 return curr->se.nr_migrations;
2729 return cpu_nr_migrations(smp_processor_id());
2732 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2737 prev = atomic64_read(&counter->hw.prev_count);
2738 now = get_cpu_migrations(counter);
2740 atomic64_set(&counter->hw.prev_count, now);
2744 atomic64_add(delta, &counter->count);
2747 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2749 cpu_migrations_perf_counter_update(counter);
2752 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2754 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2755 atomic64_set(&counter->hw.prev_count,
2756 get_cpu_migrations(counter));
2760 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2762 cpu_migrations_perf_counter_update(counter);
2765 static const struct pmu perf_ops_cpu_migrations = {
2766 .enable = cpu_migrations_perf_counter_enable,
2767 .disable = cpu_migrations_perf_counter_disable,
2768 .read = cpu_migrations_perf_counter_read,
2771 #ifdef CONFIG_EVENT_PROFILE
2772 void perf_tpcounter_event(int event_id)
2774 struct pt_regs *regs = get_irq_regs();
2777 regs = task_pt_regs(current);
2779 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2781 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2783 extern int ftrace_profile_enable(int);
2784 extern void ftrace_profile_disable(int);
2786 static void tp_perf_counter_destroy(struct perf_counter *counter)
2788 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2791 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2793 int event_id = perf_event_id(&counter->hw_event);
2796 ret = ftrace_profile_enable(event_id);
2800 counter->destroy = tp_perf_counter_destroy;
2801 counter->hw.irq_period = counter->hw_event.irq_period;
2803 return &perf_ops_generic;
2806 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2812 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2814 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2815 const struct pmu *pmu = NULL;
2816 struct hw_perf_counter *hwc = &counter->hw;
2819 * Software counters (currently) can't in general distinguish
2820 * between user, kernel and hypervisor events.
2821 * However, context switches and cpu migrations are considered
2822 * to be kernel events, and page faults are never hypervisor
2825 switch (perf_event_id(&counter->hw_event)) {
2826 case PERF_COUNT_CPU_CLOCK:
2827 pmu = &perf_ops_cpu_clock;
2829 if (hw_event->irq_period && hw_event->irq_period < 10000)
2830 hw_event->irq_period = 10000;
2832 case PERF_COUNT_TASK_CLOCK:
2834 * If the user instantiates this as a per-cpu counter,
2835 * use the cpu_clock counter instead.
2837 if (counter->ctx->task)
2838 pmu = &perf_ops_task_clock;
2840 pmu = &perf_ops_cpu_clock;
2842 if (hw_event->irq_period && hw_event->irq_period < 10000)
2843 hw_event->irq_period = 10000;
2845 case PERF_COUNT_PAGE_FAULTS:
2846 case PERF_COUNT_PAGE_FAULTS_MIN:
2847 case PERF_COUNT_PAGE_FAULTS_MAJ:
2848 case PERF_COUNT_CONTEXT_SWITCHES:
2849 pmu = &perf_ops_generic;
2851 case PERF_COUNT_CPU_MIGRATIONS:
2852 if (!counter->hw_event.exclude_kernel)
2853 pmu = &perf_ops_cpu_migrations;
2858 hwc->irq_period = hw_event->irq_period;
2864 * Allocate and initialize a counter structure
2866 static struct perf_counter *
2867 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2869 struct perf_counter_context *ctx,
2870 struct perf_counter *group_leader,
2873 const struct pmu *pmu;
2874 struct perf_counter *counter;
2877 counter = kzalloc(sizeof(*counter), gfpflags);
2879 return ERR_PTR(-ENOMEM);
2882 * Single counters are their own group leaders, with an
2883 * empty sibling list:
2886 group_leader = counter;
2888 mutex_init(&counter->mutex);
2889 INIT_LIST_HEAD(&counter->list_entry);
2890 INIT_LIST_HEAD(&counter->event_entry);
2891 INIT_LIST_HEAD(&counter->sibling_list);
2892 init_waitqueue_head(&counter->waitq);
2894 mutex_init(&counter->mmap_mutex);
2896 INIT_LIST_HEAD(&counter->child_list);
2899 counter->hw_event = *hw_event;
2900 counter->group_leader = group_leader;
2901 counter->pmu = NULL;
2904 counter->state = PERF_COUNTER_STATE_INACTIVE;
2905 if (hw_event->disabled)
2906 counter->state = PERF_COUNTER_STATE_OFF;
2911 * we currently do not support PERF_RECORD_GROUP on inherited counters
2913 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
2916 if (perf_event_raw(hw_event)) {
2917 pmu = hw_perf_counter_init(counter);
2921 switch (perf_event_type(hw_event)) {
2922 case PERF_TYPE_HARDWARE:
2923 pmu = hw_perf_counter_init(counter);
2926 case PERF_TYPE_SOFTWARE:
2927 pmu = sw_perf_counter_init(counter);
2930 case PERF_TYPE_TRACEPOINT:
2931 pmu = tp_perf_counter_init(counter);
2938 else if (IS_ERR(pmu))
2943 return ERR_PTR(err);
2948 atomic_inc(&nr_counters);
2949 if (counter->hw_event.mmap)
2950 atomic_inc(&nr_mmap_tracking);
2951 if (counter->hw_event.munmap)
2952 atomic_inc(&nr_munmap_tracking);
2953 if (counter->hw_event.comm)
2954 atomic_inc(&nr_comm_tracking);
2960 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2962 * @hw_event_uptr: event type attributes for monitoring/sampling
2965 * @group_fd: group leader counter fd
2967 SYSCALL_DEFINE5(perf_counter_open,
2968 const struct perf_counter_hw_event __user *, hw_event_uptr,
2969 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2971 struct perf_counter *counter, *group_leader;
2972 struct perf_counter_hw_event hw_event;
2973 struct perf_counter_context *ctx;
2974 struct file *counter_file = NULL;
2975 struct file *group_file = NULL;
2976 int fput_needed = 0;
2977 int fput_needed2 = 0;
2980 /* for future expandability... */
2984 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2988 * Get the target context (task or percpu):
2990 ctx = find_get_context(pid, cpu);
2992 return PTR_ERR(ctx);
2995 * Look up the group leader (we will attach this counter to it):
2997 group_leader = NULL;
2998 if (group_fd != -1) {
3000 group_file = fget_light(group_fd, &fput_needed);
3002 goto err_put_context;
3003 if (group_file->f_op != &perf_fops)
3004 goto err_put_context;
3006 group_leader = group_file->private_data;
3008 * Do not allow a recursive hierarchy (this new sibling
3009 * becoming part of another group-sibling):
3011 if (group_leader->group_leader != group_leader)
3012 goto err_put_context;
3014 * Do not allow to attach to a group in a different
3015 * task or CPU context:
3017 if (group_leader->ctx != ctx)
3018 goto err_put_context;
3020 * Only a group leader can be exclusive or pinned
3022 if (hw_event.exclusive || hw_event.pinned)
3023 goto err_put_context;
3026 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
3028 ret = PTR_ERR(counter);
3029 if (IS_ERR(counter))
3030 goto err_put_context;
3032 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3034 goto err_free_put_context;
3036 counter_file = fget_light(ret, &fput_needed2);
3038 goto err_free_put_context;
3040 counter->filp = counter_file;
3041 mutex_lock(&ctx->mutex);
3042 perf_install_in_context(ctx, counter, cpu);
3043 mutex_unlock(&ctx->mutex);
3045 fput_light(counter_file, fput_needed2);
3048 fput_light(group_file, fput_needed);
3052 err_free_put_context:
3062 * Initialize the perf_counter context in a task_struct:
3065 __perf_counter_init_context(struct perf_counter_context *ctx,
3066 struct task_struct *task)
3068 memset(ctx, 0, sizeof(*ctx));
3069 spin_lock_init(&ctx->lock);
3070 mutex_init(&ctx->mutex);
3071 INIT_LIST_HEAD(&ctx->counter_list);
3072 INIT_LIST_HEAD(&ctx->event_list);
3077 * inherit a counter from parent task to child task:
3079 static struct perf_counter *
3080 inherit_counter(struct perf_counter *parent_counter,
3081 struct task_struct *parent,
3082 struct perf_counter_context *parent_ctx,
3083 struct task_struct *child,
3084 struct perf_counter *group_leader,
3085 struct perf_counter_context *child_ctx)
3087 struct perf_counter *child_counter;
3090 * Instead of creating recursive hierarchies of counters,
3091 * we link inherited counters back to the original parent,
3092 * which has a filp for sure, which we use as the reference
3095 if (parent_counter->parent)
3096 parent_counter = parent_counter->parent;
3098 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3099 parent_counter->cpu, child_ctx,
3100 group_leader, GFP_KERNEL);
3101 if (IS_ERR(child_counter))
3102 return child_counter;
3105 * Link it up in the child's context:
3107 child_counter->task = child;
3108 add_counter_to_ctx(child_counter, child_ctx);
3110 child_counter->parent = parent_counter;
3112 * inherit into child's child as well:
3114 child_counter->hw_event.inherit = 1;
3117 * Get a reference to the parent filp - we will fput it
3118 * when the child counter exits. This is safe to do because
3119 * we are in the parent and we know that the filp still
3120 * exists and has a nonzero count:
3122 atomic_long_inc(&parent_counter->filp->f_count);
3125 * Link this into the parent counter's child list
3127 mutex_lock(&parent_counter->mutex);
3128 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3131 * Make the child state follow the state of the parent counter,
3132 * not its hw_event.disabled bit. We hold the parent's mutex,
3133 * so we won't race with perf_counter_{en,dis}able_family.
3135 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3136 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3138 child_counter->state = PERF_COUNTER_STATE_OFF;
3140 mutex_unlock(&parent_counter->mutex);
3142 return child_counter;
3145 static int inherit_group(struct perf_counter *parent_counter,
3146 struct task_struct *parent,
3147 struct perf_counter_context *parent_ctx,
3148 struct task_struct *child,
3149 struct perf_counter_context *child_ctx)
3151 struct perf_counter *leader;
3152 struct perf_counter *sub;
3153 struct perf_counter *child_ctr;
3155 leader = inherit_counter(parent_counter, parent, parent_ctx,
3156 child, NULL, child_ctx);
3158 return PTR_ERR(leader);
3159 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3160 child_ctr = inherit_counter(sub, parent, parent_ctx,
3161 child, leader, child_ctx);
3162 if (IS_ERR(child_ctr))
3163 return PTR_ERR(child_ctr);
3168 static void sync_child_counter(struct perf_counter *child_counter,
3169 struct perf_counter *parent_counter)
3171 u64 parent_val, child_val;
3173 parent_val = atomic64_read(&parent_counter->count);
3174 child_val = atomic64_read(&child_counter->count);
3177 * Add back the child's count to the parent's count:
3179 atomic64_add(child_val, &parent_counter->count);
3180 atomic64_add(child_counter->total_time_enabled,
3181 &parent_counter->child_total_time_enabled);
3182 atomic64_add(child_counter->total_time_running,
3183 &parent_counter->child_total_time_running);
3186 * Remove this counter from the parent's list
3188 mutex_lock(&parent_counter->mutex);
3189 list_del_init(&child_counter->child_list);
3190 mutex_unlock(&parent_counter->mutex);
3193 * Release the parent counter, if this was the last
3196 fput(parent_counter->filp);
3200 __perf_counter_exit_task(struct task_struct *child,
3201 struct perf_counter *child_counter,
3202 struct perf_counter_context *child_ctx)
3204 struct perf_counter *parent_counter;
3205 struct perf_counter *sub, *tmp;
3208 * If we do not self-reap then we have to wait for the
3209 * child task to unschedule (it will happen for sure),
3210 * so that its counter is at its final count. (This
3211 * condition triggers rarely - child tasks usually get
3212 * off their CPU before the parent has a chance to
3213 * get this far into the reaping action)
3215 if (child != current) {
3216 wait_task_inactive(child, 0);
3217 list_del_init(&child_counter->list_entry);
3218 update_counter_times(child_counter);
3220 struct perf_cpu_context *cpuctx;
3221 unsigned long flags;
3224 * Disable and unlink this counter.
3226 * Be careful about zapping the list - IRQ/NMI context
3227 * could still be processing it:
3229 local_irq_save(flags);
3232 cpuctx = &__get_cpu_var(perf_cpu_context);
3234 group_sched_out(child_counter, cpuctx, child_ctx);
3235 update_counter_times(child_counter);
3237 list_del_init(&child_counter->list_entry);
3239 child_ctx->nr_counters--;
3242 local_irq_restore(flags);
3245 parent_counter = child_counter->parent;
3247 * It can happen that parent exits first, and has counters
3248 * that are still around due to the child reference. These
3249 * counters need to be zapped - but otherwise linger.
3251 if (parent_counter) {
3252 sync_child_counter(child_counter, parent_counter);
3253 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
3256 sync_child_counter(sub, sub->parent);
3260 free_counter(child_counter);
3265 * When a child task exits, feed back counter values to parent counters.
3267 * Note: we may be running in child context, but the PID is not hashed
3268 * anymore so new counters will not be added.
3270 void perf_counter_exit_task(struct task_struct *child)
3272 struct perf_counter *child_counter, *tmp;
3273 struct perf_counter_context *child_ctx;
3275 child_ctx = &child->perf_counter_ctx;
3277 if (likely(!child_ctx->nr_counters))
3280 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3282 __perf_counter_exit_task(child, child_counter, child_ctx);
3286 * Initialize the perf_counter context in task_struct
3288 void perf_counter_init_task(struct task_struct *child)
3290 struct perf_counter_context *child_ctx, *parent_ctx;
3291 struct perf_counter *counter;
3292 struct task_struct *parent = current;
3294 child_ctx = &child->perf_counter_ctx;
3295 parent_ctx = &parent->perf_counter_ctx;
3297 __perf_counter_init_context(child_ctx, child);
3300 * This is executed from the parent task context, so inherit
3301 * counters that have been marked for cloning:
3304 if (likely(!parent_ctx->nr_counters))
3308 * Lock the parent list. No need to lock the child - not PID
3309 * hashed yet and not running, so nobody can access it.
3311 mutex_lock(&parent_ctx->mutex);
3314 * We dont have to disable NMIs - we are only looking at
3315 * the list, not manipulating it:
3317 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
3318 if (!counter->hw_event.inherit)
3321 if (inherit_group(counter, parent,
3322 parent_ctx, child, child_ctx))
3326 mutex_unlock(&parent_ctx->mutex);
3329 static void __cpuinit perf_counter_init_cpu(int cpu)
3331 struct perf_cpu_context *cpuctx;
3333 cpuctx = &per_cpu(perf_cpu_context, cpu);
3334 __perf_counter_init_context(&cpuctx->ctx, NULL);
3336 spin_lock(&perf_resource_lock);
3337 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3338 spin_unlock(&perf_resource_lock);
3340 hw_perf_counter_setup(cpu);
3343 #ifdef CONFIG_HOTPLUG_CPU
3344 static void __perf_counter_exit_cpu(void *info)
3346 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3347 struct perf_counter_context *ctx = &cpuctx->ctx;
3348 struct perf_counter *counter, *tmp;
3350 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3351 __perf_counter_remove_from_context(counter);
3353 static void perf_counter_exit_cpu(int cpu)
3355 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3356 struct perf_counter_context *ctx = &cpuctx->ctx;
3358 mutex_lock(&ctx->mutex);
3359 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3360 mutex_unlock(&ctx->mutex);
3363 static inline void perf_counter_exit_cpu(int cpu) { }
3366 static int __cpuinit
3367 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3369 unsigned int cpu = (long)hcpu;
3373 case CPU_UP_PREPARE:
3374 case CPU_UP_PREPARE_FROZEN:
3375 perf_counter_init_cpu(cpu);
3378 case CPU_DOWN_PREPARE:
3379 case CPU_DOWN_PREPARE_FROZEN:
3380 perf_counter_exit_cpu(cpu);
3390 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3391 .notifier_call = perf_cpu_notify,
3394 void __init perf_counter_init(void)
3396 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3397 (void *)(long)smp_processor_id());
3398 register_cpu_notifier(&perf_cpu_nb);
3401 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3403 return sprintf(buf, "%d\n", perf_reserved_percpu);
3407 perf_set_reserve_percpu(struct sysdev_class *class,
3411 struct perf_cpu_context *cpuctx;
3415 err = strict_strtoul(buf, 10, &val);
3418 if (val > perf_max_counters)
3421 spin_lock(&perf_resource_lock);
3422 perf_reserved_percpu = val;
3423 for_each_online_cpu(cpu) {
3424 cpuctx = &per_cpu(perf_cpu_context, cpu);
3425 spin_lock_irq(&cpuctx->ctx.lock);
3426 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3427 perf_max_counters - perf_reserved_percpu);
3428 cpuctx->max_pertask = mpt;
3429 spin_unlock_irq(&cpuctx->ctx.lock);
3431 spin_unlock(&perf_resource_lock);
3436 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3438 return sprintf(buf, "%d\n", perf_overcommit);
3442 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3447 err = strict_strtoul(buf, 10, &val);
3453 spin_lock(&perf_resource_lock);
3454 perf_overcommit = val;
3455 spin_unlock(&perf_resource_lock);
3460 static SYSDEV_CLASS_ATTR(
3463 perf_show_reserve_percpu,
3464 perf_set_reserve_percpu
3467 static SYSDEV_CLASS_ATTR(
3470 perf_show_overcommit,
3474 static struct attribute *perfclass_attrs[] = {
3475 &attr_reserve_percpu.attr,
3476 &attr_overcommit.attr,
3480 static struct attribute_group perfclass_attr_group = {
3481 .attrs = perfclass_attrs,
3482 .name = "perf_counters",
3485 static int __init perf_counter_sysfs_init(void)
3487 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3488 &perfclass_attr_group);
3490 device_initcall(perf_counter_sysfs_init);