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 */
47 int sysctl_perf_counter_mlock __read_mostly = 128; /* 'free' kb per counter */
50 * Lock for (sysadmin-configurable) counter reservations:
52 static DEFINE_SPINLOCK(perf_resource_lock);
55 * Architecture provided APIs - weak aliases:
57 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
62 u64 __weak hw_perf_save_disable(void) { return 0; }
63 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
64 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
65 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
66 struct perf_cpu_context *cpuctx,
67 struct perf_counter_context *ctx, int cpu)
72 void __weak perf_counter_print_debug(void) { }
75 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
77 struct perf_counter *group_leader = counter->group_leader;
80 * Depending on whether it is a standalone or sibling counter,
81 * add it straight to the context's counter list, or to the group
82 * leader's sibling list:
84 if (counter->group_leader == counter)
85 list_add_tail(&counter->list_entry, &ctx->counter_list);
87 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
88 group_leader->nr_siblings++;
91 list_add_rcu(&counter->event_entry, &ctx->event_list);
95 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
97 struct perf_counter *sibling, *tmp;
99 list_del_init(&counter->list_entry);
100 list_del_rcu(&counter->event_entry);
102 if (counter->group_leader != counter)
103 counter->group_leader->nr_siblings--;
106 * If this was a group counter with sibling counters then
107 * upgrade the siblings to singleton counters by adding them
108 * to the context list directly:
110 list_for_each_entry_safe(sibling, tmp,
111 &counter->sibling_list, list_entry) {
113 list_move_tail(&sibling->list_entry, &ctx->counter_list);
114 sibling->group_leader = sibling;
119 counter_sched_out(struct perf_counter *counter,
120 struct perf_cpu_context *cpuctx,
121 struct perf_counter_context *ctx)
123 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
126 counter->state = PERF_COUNTER_STATE_INACTIVE;
127 counter->tstamp_stopped = ctx->time;
128 counter->pmu->disable(counter);
131 if (!is_software_counter(counter))
132 cpuctx->active_oncpu--;
134 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
135 cpuctx->exclusive = 0;
139 group_sched_out(struct perf_counter *group_counter,
140 struct perf_cpu_context *cpuctx,
141 struct perf_counter_context *ctx)
143 struct perf_counter *counter;
145 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
148 counter_sched_out(group_counter, cpuctx, ctx);
151 * Schedule out siblings (if any):
153 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
154 counter_sched_out(counter, cpuctx, ctx);
156 if (group_counter->hw_event.exclusive)
157 cpuctx->exclusive = 0;
161 * Cross CPU call to remove a performance counter
163 * We disable the counter on the hardware level first. After that we
164 * remove it from the context list.
166 static void __perf_counter_remove_from_context(void *info)
168 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
169 struct perf_counter *counter = info;
170 struct perf_counter_context *ctx = counter->ctx;
175 * If this is a task context, we need to check whether it is
176 * the current task context of this cpu. If not it has been
177 * scheduled out before the smp call arrived.
179 if (ctx->task && cpuctx->task_ctx != ctx)
182 spin_lock_irqsave(&ctx->lock, flags);
184 counter_sched_out(counter, cpuctx, ctx);
186 counter->task = NULL;
190 * Protect the list operation against NMI by disabling the
191 * counters on a global level. NOP for non NMI based counters.
193 perf_flags = hw_perf_save_disable();
194 list_del_counter(counter, ctx);
195 hw_perf_restore(perf_flags);
199 * Allow more per task counters with respect to the
202 cpuctx->max_pertask =
203 min(perf_max_counters - ctx->nr_counters,
204 perf_max_counters - perf_reserved_percpu);
207 spin_unlock_irqrestore(&ctx->lock, flags);
212 * Remove the counter from a task's (or a CPU's) list of counters.
214 * Must be called with counter->mutex and ctx->mutex held.
216 * CPU counters are removed with a smp call. For task counters we only
217 * call when the task is on a CPU.
219 static void perf_counter_remove_from_context(struct perf_counter *counter)
221 struct perf_counter_context *ctx = counter->ctx;
222 struct task_struct *task = ctx->task;
226 * Per cpu counters are removed via an smp call and
227 * the removal is always sucessful.
229 smp_call_function_single(counter->cpu,
230 __perf_counter_remove_from_context,
236 task_oncpu_function_call(task, __perf_counter_remove_from_context,
239 spin_lock_irq(&ctx->lock);
241 * If the context is active we need to retry the smp call.
243 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
244 spin_unlock_irq(&ctx->lock);
249 * The lock prevents that this context is scheduled in so we
250 * can remove the counter safely, if the call above did not
253 if (!list_empty(&counter->list_entry)) {
255 list_del_counter(counter, ctx);
256 counter->task = NULL;
258 spin_unlock_irq(&ctx->lock);
261 static inline u64 perf_clock(void)
263 return cpu_clock(smp_processor_id());
267 * Update the record of the current time in a context.
269 static void update_context_time(struct perf_counter_context *ctx)
271 u64 now = perf_clock();
273 ctx->time += now - ctx->timestamp;
274 ctx->timestamp = now;
278 * Update the total_time_enabled and total_time_running fields for a counter.
280 static void update_counter_times(struct perf_counter *counter)
282 struct perf_counter_context *ctx = counter->ctx;
285 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
288 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
290 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
291 run_end = counter->tstamp_stopped;
295 counter->total_time_running = run_end - counter->tstamp_running;
299 * Update total_time_enabled and total_time_running for all counters in a group.
301 static void update_group_times(struct perf_counter *leader)
303 struct perf_counter *counter;
305 update_counter_times(leader);
306 list_for_each_entry(counter, &leader->sibling_list, list_entry)
307 update_counter_times(counter);
311 * Cross CPU call to disable a performance counter
313 static void __perf_counter_disable(void *info)
315 struct perf_counter *counter = info;
316 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
317 struct perf_counter_context *ctx = counter->ctx;
321 * If this is a per-task counter, need to check whether this
322 * counter's task is the current task on this cpu.
324 if (ctx->task && cpuctx->task_ctx != ctx)
327 spin_lock_irqsave(&ctx->lock, flags);
330 * If the counter is on, turn it off.
331 * If it is in error state, leave it in error state.
333 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
334 update_context_time(ctx);
335 update_counter_times(counter);
336 if (counter == counter->group_leader)
337 group_sched_out(counter, cpuctx, ctx);
339 counter_sched_out(counter, cpuctx, ctx);
340 counter->state = PERF_COUNTER_STATE_OFF;
343 spin_unlock_irqrestore(&ctx->lock, flags);
349 static void perf_counter_disable(struct perf_counter *counter)
351 struct perf_counter_context *ctx = counter->ctx;
352 struct task_struct *task = ctx->task;
356 * Disable the counter on the cpu that it's on
358 smp_call_function_single(counter->cpu, __perf_counter_disable,
364 task_oncpu_function_call(task, __perf_counter_disable, counter);
366 spin_lock_irq(&ctx->lock);
368 * If the counter is still active, we need to retry the cross-call.
370 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
371 spin_unlock_irq(&ctx->lock);
376 * Since we have the lock this context can't be scheduled
377 * in, so we can change the state safely.
379 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
380 update_counter_times(counter);
381 counter->state = PERF_COUNTER_STATE_OFF;
384 spin_unlock_irq(&ctx->lock);
388 * Disable a counter and all its children.
390 static void perf_counter_disable_family(struct perf_counter *counter)
392 struct perf_counter *child;
394 perf_counter_disable(counter);
397 * Lock the mutex to protect the list of children
399 mutex_lock(&counter->mutex);
400 list_for_each_entry(child, &counter->child_list, child_list)
401 perf_counter_disable(child);
402 mutex_unlock(&counter->mutex);
406 counter_sched_in(struct perf_counter *counter,
407 struct perf_cpu_context *cpuctx,
408 struct perf_counter_context *ctx,
411 if (counter->state <= PERF_COUNTER_STATE_OFF)
414 counter->state = PERF_COUNTER_STATE_ACTIVE;
415 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
417 * The new state must be visible before we turn it on in the hardware:
421 if (counter->pmu->enable(counter)) {
422 counter->state = PERF_COUNTER_STATE_INACTIVE;
427 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
429 if (!is_software_counter(counter))
430 cpuctx->active_oncpu++;
433 if (counter->hw_event.exclusive)
434 cpuctx->exclusive = 1;
440 * Return 1 for a group consisting entirely of software counters,
441 * 0 if the group contains any hardware counters.
443 static int is_software_only_group(struct perf_counter *leader)
445 struct perf_counter *counter;
447 if (!is_software_counter(leader))
450 list_for_each_entry(counter, &leader->sibling_list, list_entry)
451 if (!is_software_counter(counter))
458 * Work out whether we can put this counter group on the CPU now.
460 static int group_can_go_on(struct perf_counter *counter,
461 struct perf_cpu_context *cpuctx,
465 * Groups consisting entirely of software counters can always go on.
467 if (is_software_only_group(counter))
470 * If an exclusive group is already on, no other hardware
471 * counters can go on.
473 if (cpuctx->exclusive)
476 * If this group is exclusive and there are already
477 * counters on the CPU, it can't go on.
479 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
482 * Otherwise, try to add it if all previous groups were able
488 static void add_counter_to_ctx(struct perf_counter *counter,
489 struct perf_counter_context *ctx)
491 list_add_counter(counter, ctx);
493 counter->prev_state = PERF_COUNTER_STATE_OFF;
494 counter->tstamp_enabled = ctx->time;
495 counter->tstamp_running = ctx->time;
496 counter->tstamp_stopped = ctx->time;
500 * Cross CPU call to install and enable a performance counter
502 static void __perf_install_in_context(void *info)
504 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
505 struct perf_counter *counter = info;
506 struct perf_counter_context *ctx = counter->ctx;
507 struct perf_counter *leader = counter->group_leader;
508 int cpu = smp_processor_id();
514 * If this is a task context, we need to check whether it is
515 * the current task context of this cpu. If not it has been
516 * scheduled out before the smp call arrived.
518 if (ctx->task && cpuctx->task_ctx != ctx)
521 spin_lock_irqsave(&ctx->lock, flags);
522 update_context_time(ctx);
525 * Protect the list operation against NMI by disabling the
526 * counters on a global level. NOP for non NMI based counters.
528 perf_flags = hw_perf_save_disable();
530 add_counter_to_ctx(counter, ctx);
533 * Don't put the counter on if it is disabled or if
534 * it is in a group and the group isn't on.
536 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
537 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
541 * An exclusive counter can't go on if there are already active
542 * hardware counters, and no hardware counter can go on if there
543 * is already an exclusive counter on.
545 if (!group_can_go_on(counter, cpuctx, 1))
548 err = counter_sched_in(counter, cpuctx, ctx, cpu);
552 * This counter couldn't go on. If it is in a group
553 * then we have to pull the whole group off.
554 * If the counter group is pinned then put it in error state.
556 if (leader != counter)
557 group_sched_out(leader, cpuctx, ctx);
558 if (leader->hw_event.pinned) {
559 update_group_times(leader);
560 leader->state = PERF_COUNTER_STATE_ERROR;
564 if (!err && !ctx->task && cpuctx->max_pertask)
565 cpuctx->max_pertask--;
568 hw_perf_restore(perf_flags);
570 spin_unlock_irqrestore(&ctx->lock, flags);
574 * Attach a performance counter to a context
576 * First we add the counter to the list with the hardware enable bit
577 * in counter->hw_config cleared.
579 * If the counter is attached to a task which is on a CPU we use a smp
580 * call to enable it in the task context. The task might have been
581 * scheduled away, but we check this in the smp call again.
583 * Must be called with ctx->mutex held.
586 perf_install_in_context(struct perf_counter_context *ctx,
587 struct perf_counter *counter,
590 struct task_struct *task = ctx->task;
594 * Per cpu counters are installed via an smp call and
595 * the install is always sucessful.
597 smp_call_function_single(cpu, __perf_install_in_context,
602 counter->task = task;
604 task_oncpu_function_call(task, __perf_install_in_context,
607 spin_lock_irq(&ctx->lock);
609 * we need to retry the smp call.
611 if (ctx->is_active && list_empty(&counter->list_entry)) {
612 spin_unlock_irq(&ctx->lock);
617 * The lock prevents that this context is scheduled in so we
618 * can add the counter safely, if it the call above did not
621 if (list_empty(&counter->list_entry))
622 add_counter_to_ctx(counter, ctx);
623 spin_unlock_irq(&ctx->lock);
627 * Cross CPU call to enable a performance counter
629 static void __perf_counter_enable(void *info)
631 struct perf_counter *counter = info;
632 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
633 struct perf_counter_context *ctx = counter->ctx;
634 struct perf_counter *leader = counter->group_leader;
639 * If this is a per-task counter, need to check whether this
640 * counter's task is the current task on this cpu.
642 if (ctx->task && cpuctx->task_ctx != ctx)
645 spin_lock_irqsave(&ctx->lock, flags);
646 update_context_time(ctx);
648 counter->prev_state = counter->state;
649 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
651 counter->state = PERF_COUNTER_STATE_INACTIVE;
652 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
655 * If the counter is in a group and isn't the group leader,
656 * then don't put it on unless the group is on.
658 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
661 if (!group_can_go_on(counter, cpuctx, 1))
664 err = counter_sched_in(counter, cpuctx, ctx,
669 * If this counter can't go on and it's part of a
670 * group, then the whole group has to come off.
672 if (leader != counter)
673 group_sched_out(leader, cpuctx, ctx);
674 if (leader->hw_event.pinned) {
675 update_group_times(leader);
676 leader->state = PERF_COUNTER_STATE_ERROR;
681 spin_unlock_irqrestore(&ctx->lock, flags);
687 static void perf_counter_enable(struct perf_counter *counter)
689 struct perf_counter_context *ctx = counter->ctx;
690 struct task_struct *task = ctx->task;
694 * Enable the counter on the cpu that it's on
696 smp_call_function_single(counter->cpu, __perf_counter_enable,
701 spin_lock_irq(&ctx->lock);
702 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
706 * If the counter is in error state, clear that first.
707 * That way, if we see the counter in error state below, we
708 * know that it has gone back into error state, as distinct
709 * from the task having been scheduled away before the
710 * cross-call arrived.
712 if (counter->state == PERF_COUNTER_STATE_ERROR)
713 counter->state = PERF_COUNTER_STATE_OFF;
716 spin_unlock_irq(&ctx->lock);
717 task_oncpu_function_call(task, __perf_counter_enable, counter);
719 spin_lock_irq(&ctx->lock);
722 * If the context is active and the counter is still off,
723 * we need to retry the cross-call.
725 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
729 * Since we have the lock this context can't be scheduled
730 * in, so we can change the state safely.
732 if (counter->state == PERF_COUNTER_STATE_OFF) {
733 counter->state = PERF_COUNTER_STATE_INACTIVE;
734 counter->tstamp_enabled =
735 ctx->time - counter->total_time_enabled;
738 spin_unlock_irq(&ctx->lock);
741 static void perf_counter_refresh(struct perf_counter *counter, int refresh)
743 atomic_add(refresh, &counter->event_limit);
744 perf_counter_enable(counter);
748 * Enable a counter and all its children.
750 static void perf_counter_enable_family(struct perf_counter *counter)
752 struct perf_counter *child;
754 perf_counter_enable(counter);
757 * Lock the mutex to protect the list of children
759 mutex_lock(&counter->mutex);
760 list_for_each_entry(child, &counter->child_list, child_list)
761 perf_counter_enable(child);
762 mutex_unlock(&counter->mutex);
765 void __perf_counter_sched_out(struct perf_counter_context *ctx,
766 struct perf_cpu_context *cpuctx)
768 struct perf_counter *counter;
771 spin_lock(&ctx->lock);
773 if (likely(!ctx->nr_counters))
775 update_context_time(ctx);
777 flags = hw_perf_save_disable();
778 if (ctx->nr_active) {
779 list_for_each_entry(counter, &ctx->counter_list, list_entry)
780 group_sched_out(counter, cpuctx, ctx);
782 hw_perf_restore(flags);
784 spin_unlock(&ctx->lock);
788 * Called from scheduler to remove the counters of the current task,
789 * with interrupts disabled.
791 * We stop each counter and update the counter value in counter->count.
793 * This does not protect us against NMI, but disable()
794 * sets the disabled bit in the control field of counter _before_
795 * accessing the counter control register. If a NMI hits, then it will
796 * not restart the counter.
798 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
800 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
801 struct perf_counter_context *ctx = &task->perf_counter_ctx;
802 struct pt_regs *regs;
804 if (likely(!cpuctx->task_ctx))
807 update_context_time(ctx);
809 regs = task_pt_regs(task);
810 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
811 __perf_counter_sched_out(ctx, cpuctx);
813 cpuctx->task_ctx = NULL;
816 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
818 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
822 group_sched_in(struct perf_counter *group_counter,
823 struct perf_cpu_context *cpuctx,
824 struct perf_counter_context *ctx,
827 struct perf_counter *counter, *partial_group;
830 if (group_counter->state == PERF_COUNTER_STATE_OFF)
833 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
835 return ret < 0 ? ret : 0;
837 group_counter->prev_state = group_counter->state;
838 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
842 * Schedule in siblings as one group (if any):
844 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
845 counter->prev_state = counter->state;
846 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
847 partial_group = counter;
856 * Groups can be scheduled in as one unit only, so undo any
857 * partial group before returning:
859 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
860 if (counter == partial_group)
862 counter_sched_out(counter, cpuctx, ctx);
864 counter_sched_out(group_counter, cpuctx, ctx);
870 __perf_counter_sched_in(struct perf_counter_context *ctx,
871 struct perf_cpu_context *cpuctx, int cpu)
873 struct perf_counter *counter;
877 spin_lock(&ctx->lock);
879 if (likely(!ctx->nr_counters))
882 ctx->timestamp = perf_clock();
884 flags = hw_perf_save_disable();
887 * First go through the list and put on any pinned groups
888 * in order to give them the best chance of going on.
890 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
891 if (counter->state <= PERF_COUNTER_STATE_OFF ||
892 !counter->hw_event.pinned)
894 if (counter->cpu != -1 && counter->cpu != cpu)
897 if (group_can_go_on(counter, cpuctx, 1))
898 group_sched_in(counter, cpuctx, ctx, cpu);
901 * If this pinned group hasn't been scheduled,
902 * put it in error state.
904 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
905 update_group_times(counter);
906 counter->state = PERF_COUNTER_STATE_ERROR;
910 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
912 * Ignore counters in OFF or ERROR state, and
913 * ignore pinned counters since we did them already.
915 if (counter->state <= PERF_COUNTER_STATE_OFF ||
916 counter->hw_event.pinned)
920 * Listen to the 'cpu' scheduling filter constraint
923 if (counter->cpu != -1 && counter->cpu != cpu)
926 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
927 if (group_sched_in(counter, cpuctx, ctx, cpu))
931 hw_perf_restore(flags);
933 spin_unlock(&ctx->lock);
937 * Called from scheduler to add the counters of the current task
938 * with interrupts disabled.
940 * We restore the counter value and then enable it.
942 * This does not protect us against NMI, but enable()
943 * sets the enabled bit in the control field of counter _before_
944 * accessing the counter control register. If a NMI hits, then it will
945 * keep the counter running.
947 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
949 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
950 struct perf_counter_context *ctx = &task->perf_counter_ctx;
952 __perf_counter_sched_in(ctx, cpuctx, cpu);
953 cpuctx->task_ctx = ctx;
956 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
958 struct perf_counter_context *ctx = &cpuctx->ctx;
960 __perf_counter_sched_in(ctx, cpuctx, cpu);
963 int perf_counter_task_disable(void)
965 struct task_struct *curr = current;
966 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
967 struct perf_counter *counter;
972 if (likely(!ctx->nr_counters))
975 local_irq_save(flags);
976 cpu = smp_processor_id();
978 perf_counter_task_sched_out(curr, cpu);
980 spin_lock(&ctx->lock);
983 * Disable all the counters:
985 perf_flags = hw_perf_save_disable();
987 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
988 if (counter->state != PERF_COUNTER_STATE_ERROR) {
989 update_group_times(counter);
990 counter->state = PERF_COUNTER_STATE_OFF;
994 hw_perf_restore(perf_flags);
996 spin_unlock_irqrestore(&ctx->lock, flags);
1001 int perf_counter_task_enable(void)
1003 struct task_struct *curr = current;
1004 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1005 struct perf_counter *counter;
1006 unsigned long flags;
1010 if (likely(!ctx->nr_counters))
1013 local_irq_save(flags);
1014 cpu = smp_processor_id();
1016 perf_counter_task_sched_out(curr, cpu);
1018 spin_lock(&ctx->lock);
1021 * Disable all the counters:
1023 perf_flags = hw_perf_save_disable();
1025 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1026 if (counter->state > PERF_COUNTER_STATE_OFF)
1028 counter->state = PERF_COUNTER_STATE_INACTIVE;
1029 counter->tstamp_enabled =
1030 ctx->time - counter->total_time_enabled;
1031 counter->hw_event.disabled = 0;
1033 hw_perf_restore(perf_flags);
1035 spin_unlock(&ctx->lock);
1037 perf_counter_task_sched_in(curr, cpu);
1039 local_irq_restore(flags);
1045 * Round-robin a context's counters:
1047 static void rotate_ctx(struct perf_counter_context *ctx)
1049 struct perf_counter *counter;
1052 if (!ctx->nr_counters)
1055 spin_lock(&ctx->lock);
1057 * Rotate the first entry last (works just fine for group counters too):
1059 perf_flags = hw_perf_save_disable();
1060 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1061 list_move_tail(&counter->list_entry, &ctx->counter_list);
1064 hw_perf_restore(perf_flags);
1066 spin_unlock(&ctx->lock);
1069 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1071 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1072 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1074 perf_counter_cpu_sched_out(cpuctx);
1075 perf_counter_task_sched_out(curr, cpu);
1077 rotate_ctx(&cpuctx->ctx);
1080 perf_counter_cpu_sched_in(cpuctx, cpu);
1081 perf_counter_task_sched_in(curr, cpu);
1085 * Cross CPU call to read the hardware counter
1087 static void __read(void *info)
1089 struct perf_counter *counter = info;
1090 struct perf_counter_context *ctx = counter->ctx;
1091 unsigned long flags;
1093 local_irq_save(flags);
1095 update_context_time(ctx);
1096 counter->pmu->read(counter);
1097 update_counter_times(counter);
1098 local_irq_restore(flags);
1101 static u64 perf_counter_read(struct perf_counter *counter)
1104 * If counter is enabled and currently active on a CPU, update the
1105 * value in the counter structure:
1107 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1108 smp_call_function_single(counter->oncpu,
1109 __read, counter, 1);
1110 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1111 update_counter_times(counter);
1114 return atomic64_read(&counter->count);
1117 static void put_context(struct perf_counter_context *ctx)
1120 put_task_struct(ctx->task);
1123 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1125 struct perf_cpu_context *cpuctx;
1126 struct perf_counter_context *ctx;
1127 struct task_struct *task;
1130 * If cpu is not a wildcard then this is a percpu counter:
1133 /* Must be root to operate on a CPU counter: */
1134 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1135 return ERR_PTR(-EACCES);
1137 if (cpu < 0 || cpu > num_possible_cpus())
1138 return ERR_PTR(-EINVAL);
1141 * We could be clever and allow to attach a counter to an
1142 * offline CPU and activate it when the CPU comes up, but
1145 if (!cpu_isset(cpu, cpu_online_map))
1146 return ERR_PTR(-ENODEV);
1148 cpuctx = &per_cpu(perf_cpu_context, cpu);
1158 task = find_task_by_vpid(pid);
1160 get_task_struct(task);
1164 return ERR_PTR(-ESRCH);
1166 ctx = &task->perf_counter_ctx;
1169 /* Reuse ptrace permission checks for now. */
1170 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1172 return ERR_PTR(-EACCES);
1178 static void free_counter_rcu(struct rcu_head *head)
1180 struct perf_counter *counter;
1182 counter = container_of(head, struct perf_counter, rcu_head);
1186 static void perf_pending_sync(struct perf_counter *counter);
1188 static void free_counter(struct perf_counter *counter)
1190 perf_pending_sync(counter);
1192 if (counter->hw_event.mmap)
1193 atomic_dec(&nr_mmap_tracking);
1194 if (counter->hw_event.munmap)
1195 atomic_dec(&nr_munmap_tracking);
1196 if (counter->hw_event.comm)
1197 atomic_dec(&nr_comm_tracking);
1199 if (counter->destroy)
1200 counter->destroy(counter);
1202 call_rcu(&counter->rcu_head, free_counter_rcu);
1206 * Called when the last reference to the file is gone.
1208 static int perf_release(struct inode *inode, struct file *file)
1210 struct perf_counter *counter = file->private_data;
1211 struct perf_counter_context *ctx = counter->ctx;
1213 file->private_data = NULL;
1215 mutex_lock(&ctx->mutex);
1216 mutex_lock(&counter->mutex);
1218 perf_counter_remove_from_context(counter);
1220 mutex_unlock(&counter->mutex);
1221 mutex_unlock(&ctx->mutex);
1223 free_counter(counter);
1230 * Read the performance counter - simple non blocking version for now
1233 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1239 * Return end-of-file for a read on a counter that is in
1240 * error state (i.e. because it was pinned but it couldn't be
1241 * scheduled on to the CPU at some point).
1243 if (counter->state == PERF_COUNTER_STATE_ERROR)
1246 mutex_lock(&counter->mutex);
1247 values[0] = perf_counter_read(counter);
1249 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1250 values[n++] = counter->total_time_enabled +
1251 atomic64_read(&counter->child_total_time_enabled);
1252 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1253 values[n++] = counter->total_time_running +
1254 atomic64_read(&counter->child_total_time_running);
1255 mutex_unlock(&counter->mutex);
1257 if (count < n * sizeof(u64))
1259 count = n * sizeof(u64);
1261 if (copy_to_user(buf, values, count))
1268 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1270 struct perf_counter *counter = file->private_data;
1272 return perf_read_hw(counter, buf, count);
1275 static unsigned int perf_poll(struct file *file, poll_table *wait)
1277 struct perf_counter *counter = file->private_data;
1278 struct perf_mmap_data *data;
1279 unsigned int events = POLL_HUP;
1282 data = rcu_dereference(counter->data);
1284 events = atomic_xchg(&data->poll, 0);
1287 poll_wait(file, &counter->waitq, wait);
1292 static void perf_counter_reset(struct perf_counter *counter)
1294 atomic_set(&counter->count, 0);
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);
1312 case PERF_COUNTER_IOC_RESET:
1313 perf_counter_reset(counter);
1322 * Callers need to ensure there can be no nesting of this function, otherwise
1323 * the seqlock logic goes bad. We can not serialize this because the arch
1324 * code calls this from NMI context.
1326 void perf_counter_update_userpage(struct perf_counter *counter)
1328 struct perf_mmap_data *data;
1329 struct perf_counter_mmap_page *userpg;
1332 data = rcu_dereference(counter->data);
1336 userpg = data->user_page;
1339 * Disable preemption so as to not let the corresponding user-space
1340 * spin too long if we get preempted.
1345 userpg->index = counter->hw.idx;
1346 userpg->offset = atomic64_read(&counter->count);
1347 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1348 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1357 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1359 struct perf_counter *counter = vma->vm_file->private_data;
1360 struct perf_mmap_data *data;
1361 int ret = VM_FAULT_SIGBUS;
1364 data = rcu_dereference(counter->data);
1368 if (vmf->pgoff == 0) {
1369 vmf->page = virt_to_page(data->user_page);
1371 int nr = vmf->pgoff - 1;
1373 if ((unsigned)nr > data->nr_pages)
1376 vmf->page = virt_to_page(data->data_pages[nr]);
1378 get_page(vmf->page);
1386 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1388 struct perf_mmap_data *data;
1392 WARN_ON(atomic_read(&counter->mmap_count));
1394 size = sizeof(struct perf_mmap_data);
1395 size += nr_pages * sizeof(void *);
1397 data = kzalloc(size, GFP_KERNEL);
1401 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1402 if (!data->user_page)
1403 goto fail_user_page;
1405 for (i = 0; i < nr_pages; i++) {
1406 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1407 if (!data->data_pages[i])
1408 goto fail_data_pages;
1411 data->nr_pages = nr_pages;
1412 atomic_set(&data->lock, -1);
1414 rcu_assign_pointer(counter->data, data);
1419 for (i--; i >= 0; i--)
1420 free_page((unsigned long)data->data_pages[i]);
1422 free_page((unsigned long)data->user_page);
1431 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1433 struct perf_mmap_data *data = container_of(rcu_head,
1434 struct perf_mmap_data, rcu_head);
1437 free_page((unsigned long)data->user_page);
1438 for (i = 0; i < data->nr_pages; i++)
1439 free_page((unsigned long)data->data_pages[i]);
1443 static void perf_mmap_data_free(struct perf_counter *counter)
1445 struct perf_mmap_data *data = counter->data;
1447 WARN_ON(atomic_read(&counter->mmap_count));
1449 rcu_assign_pointer(counter->data, NULL);
1450 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1453 static void perf_mmap_open(struct vm_area_struct *vma)
1455 struct perf_counter *counter = vma->vm_file->private_data;
1457 atomic_inc(&counter->mmap_count);
1460 static void perf_mmap_close(struct vm_area_struct *vma)
1462 struct perf_counter *counter = vma->vm_file->private_data;
1464 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1465 &counter->mmap_mutex)) {
1466 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1467 perf_mmap_data_free(counter);
1468 mutex_unlock(&counter->mmap_mutex);
1472 static struct vm_operations_struct perf_mmap_vmops = {
1473 .open = perf_mmap_open,
1474 .close = perf_mmap_close,
1475 .fault = perf_mmap_fault,
1478 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1480 struct perf_counter *counter = file->private_data;
1481 unsigned long vma_size;
1482 unsigned long nr_pages;
1483 unsigned long locked, lock_limit;
1487 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1490 vma_size = vma->vm_end - vma->vm_start;
1491 nr_pages = (vma_size / PAGE_SIZE) - 1;
1494 * If we have data pages ensure they're a power-of-two number, so we
1495 * can do bitmasks instead of modulo.
1497 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1500 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1503 if (vma->vm_pgoff != 0)
1506 mutex_lock(&counter->mmap_mutex);
1507 if (atomic_inc_not_zero(&counter->mmap_count)) {
1508 if (nr_pages != counter->data->nr_pages)
1513 extra = nr_pages /* + 1 only account the data pages */;
1514 extra -= sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1518 locked = vma->vm_mm->locked_vm + extra;
1520 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1521 lock_limit >>= PAGE_SHIFT;
1523 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1528 WARN_ON(counter->data);
1529 ret = perf_mmap_data_alloc(counter, nr_pages);
1533 atomic_set(&counter->mmap_count, 1);
1534 vma->vm_mm->locked_vm += extra;
1535 counter->data->nr_locked = extra;
1537 mutex_unlock(&counter->mmap_mutex);
1539 vma->vm_flags &= ~VM_MAYWRITE;
1540 vma->vm_flags |= VM_RESERVED;
1541 vma->vm_ops = &perf_mmap_vmops;
1546 static int perf_fasync(int fd, struct file *filp, int on)
1548 struct perf_counter *counter = filp->private_data;
1549 struct inode *inode = filp->f_path.dentry->d_inode;
1552 mutex_lock(&inode->i_mutex);
1553 retval = fasync_helper(fd, filp, on, &counter->fasync);
1554 mutex_unlock(&inode->i_mutex);
1562 static const struct file_operations perf_fops = {
1563 .release = perf_release,
1566 .unlocked_ioctl = perf_ioctl,
1567 .compat_ioctl = perf_ioctl,
1569 .fasync = perf_fasync,
1573 * Perf counter wakeup
1575 * If there's data, ensure we set the poll() state and publish everything
1576 * to user-space before waking everybody up.
1579 void perf_counter_wakeup(struct perf_counter *counter)
1581 wake_up_all(&counter->waitq);
1583 if (counter->pending_kill) {
1584 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1585 counter->pending_kill = 0;
1592 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1594 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1595 * single linked list and use cmpxchg() to add entries lockless.
1598 static void perf_pending_counter(struct perf_pending_entry *entry)
1600 struct perf_counter *counter = container_of(entry,
1601 struct perf_counter, pending);
1603 if (counter->pending_disable) {
1604 counter->pending_disable = 0;
1605 perf_counter_disable(counter);
1608 if (counter->pending_wakeup) {
1609 counter->pending_wakeup = 0;
1610 perf_counter_wakeup(counter);
1614 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1616 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1620 static void perf_pending_queue(struct perf_pending_entry *entry,
1621 void (*func)(struct perf_pending_entry *))
1623 struct perf_pending_entry **head;
1625 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1630 head = &get_cpu_var(perf_pending_head);
1633 entry->next = *head;
1634 } while (cmpxchg(head, entry->next, entry) != entry->next);
1636 set_perf_counter_pending();
1638 put_cpu_var(perf_pending_head);
1641 static int __perf_pending_run(void)
1643 struct perf_pending_entry *list;
1646 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1647 while (list != PENDING_TAIL) {
1648 void (*func)(struct perf_pending_entry *);
1649 struct perf_pending_entry *entry = list;
1656 * Ensure we observe the unqueue before we issue the wakeup,
1657 * so that we won't be waiting forever.
1658 * -- see perf_not_pending().
1669 static inline int perf_not_pending(struct perf_counter *counter)
1672 * If we flush on whatever cpu we run, there is a chance we don't
1676 __perf_pending_run();
1680 * Ensure we see the proper queue state before going to sleep
1681 * so that we do not miss the wakeup. -- see perf_pending_handle()
1684 return counter->pending.next == NULL;
1687 static void perf_pending_sync(struct perf_counter *counter)
1689 wait_event(counter->waitq, perf_not_pending(counter));
1692 void perf_counter_do_pending(void)
1694 __perf_pending_run();
1698 * Callchain support -- arch specific
1701 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1710 struct perf_output_handle {
1711 struct perf_counter *counter;
1712 struct perf_mmap_data *data;
1713 unsigned int offset;
1718 unsigned long flags;
1721 static void perf_output_wakeup(struct perf_output_handle *handle)
1723 atomic_set(&handle->data->poll, POLL_IN);
1726 handle->counter->pending_wakeup = 1;
1727 perf_pending_queue(&handle->counter->pending,
1728 perf_pending_counter);
1730 perf_counter_wakeup(handle->counter);
1734 * Curious locking construct.
1736 * We need to ensure a later event doesn't publish a head when a former
1737 * event isn't done writing. However since we need to deal with NMIs we
1738 * cannot fully serialize things.
1740 * What we do is serialize between CPUs so we only have to deal with NMI
1741 * nesting on a single CPU.
1743 * We only publish the head (and generate a wakeup) when the outer-most
1746 static void perf_output_lock(struct perf_output_handle *handle)
1748 struct perf_mmap_data *data = handle->data;
1753 local_irq_save(handle->flags);
1754 cpu = smp_processor_id();
1756 if (in_nmi() && atomic_read(&data->lock) == cpu)
1759 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1765 static void perf_output_unlock(struct perf_output_handle *handle)
1767 struct perf_mmap_data *data = handle->data;
1770 data->done_head = data->head;
1772 if (!handle->locked)
1777 * The xchg implies a full barrier that ensures all writes are done
1778 * before we publish the new head, matched by a rmb() in userspace when
1779 * reading this position.
1781 while ((head = atomic_xchg(&data->done_head, 0)))
1782 data->user_page->data_head = head;
1785 * NMI can happen here, which means we can miss a done_head update.
1788 cpu = atomic_xchg(&data->lock, -1);
1789 WARN_ON_ONCE(cpu != smp_processor_id());
1792 * Therefore we have to validate we did not indeed do so.
1794 if (unlikely(atomic_read(&data->done_head))) {
1796 * Since we had it locked, we can lock it again.
1798 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1804 if (atomic_xchg(&data->wakeup, 0))
1805 perf_output_wakeup(handle);
1807 local_irq_restore(handle->flags);
1810 static int perf_output_begin(struct perf_output_handle *handle,
1811 struct perf_counter *counter, unsigned int size,
1812 int nmi, int overflow)
1814 struct perf_mmap_data *data;
1815 unsigned int offset, head;
1818 data = rcu_dereference(counter->data);
1822 handle->data = data;
1823 handle->counter = counter;
1825 handle->overflow = overflow;
1827 if (!data->nr_pages)
1830 perf_output_lock(handle);
1833 offset = head = atomic_read(&data->head);
1835 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1837 handle->offset = offset;
1838 handle->head = head;
1840 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
1841 atomic_set(&data->wakeup, 1);
1846 perf_output_wakeup(handle);
1853 static void perf_output_copy(struct perf_output_handle *handle,
1854 void *buf, unsigned int len)
1856 unsigned int pages_mask;
1857 unsigned int offset;
1861 offset = handle->offset;
1862 pages_mask = handle->data->nr_pages - 1;
1863 pages = handle->data->data_pages;
1866 unsigned int page_offset;
1869 nr = (offset >> PAGE_SHIFT) & pages_mask;
1870 page_offset = offset & (PAGE_SIZE - 1);
1871 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1873 memcpy(pages[nr] + page_offset, buf, size);
1880 handle->offset = offset;
1882 WARN_ON_ONCE(handle->offset > handle->head);
1885 #define perf_output_put(handle, x) \
1886 perf_output_copy((handle), &(x), sizeof(x))
1888 static void perf_output_end(struct perf_output_handle *handle)
1890 struct perf_counter *counter = handle->counter;
1891 struct perf_mmap_data *data = handle->data;
1893 int wakeup_events = counter->hw_event.wakeup_events;
1895 if (handle->overflow && wakeup_events) {
1896 int events = atomic_inc_return(&data->events);
1897 if (events >= wakeup_events) {
1898 atomic_sub(wakeup_events, &data->events);
1899 atomic_set(&data->wakeup, 1);
1903 perf_output_unlock(handle);
1907 static void perf_counter_output(struct perf_counter *counter,
1908 int nmi, struct pt_regs *regs, u64 addr)
1911 u64 record_type = counter->hw_event.record_type;
1912 struct perf_output_handle handle;
1913 struct perf_event_header header;
1922 struct perf_callchain_entry *callchain = NULL;
1923 int callchain_size = 0;
1927 header.size = sizeof(header);
1929 header.misc = PERF_EVENT_MISC_OVERFLOW;
1930 header.misc |= user_mode(regs) ?
1931 PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
1933 if (record_type & PERF_RECORD_IP) {
1934 ip = instruction_pointer(regs);
1935 header.type |= PERF_RECORD_IP;
1936 header.size += sizeof(ip);
1939 if (record_type & PERF_RECORD_TID) {
1940 /* namespace issues */
1941 tid_entry.pid = current->group_leader->pid;
1942 tid_entry.tid = current->pid;
1944 header.type |= PERF_RECORD_TID;
1945 header.size += sizeof(tid_entry);
1948 if (record_type & PERF_RECORD_TIME) {
1950 * Maybe do better on x86 and provide cpu_clock_nmi()
1952 time = sched_clock();
1954 header.type |= PERF_RECORD_TIME;
1955 header.size += sizeof(u64);
1958 if (record_type & PERF_RECORD_ADDR) {
1959 header.type |= PERF_RECORD_ADDR;
1960 header.size += sizeof(u64);
1963 if (record_type & PERF_RECORD_GROUP) {
1964 header.type |= PERF_RECORD_GROUP;
1965 header.size += sizeof(u64) +
1966 counter->nr_siblings * sizeof(group_entry);
1969 if (record_type & PERF_RECORD_CALLCHAIN) {
1970 callchain = perf_callchain(regs);
1973 callchain_size = (1 + callchain->nr) * sizeof(u64);
1975 header.type |= PERF_RECORD_CALLCHAIN;
1976 header.size += callchain_size;
1980 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
1984 perf_output_put(&handle, header);
1986 if (record_type & PERF_RECORD_IP)
1987 perf_output_put(&handle, ip);
1989 if (record_type & PERF_RECORD_TID)
1990 perf_output_put(&handle, tid_entry);
1992 if (record_type & PERF_RECORD_TIME)
1993 perf_output_put(&handle, time);
1995 if (record_type & PERF_RECORD_ADDR)
1996 perf_output_put(&handle, addr);
1998 if (record_type & PERF_RECORD_GROUP) {
1999 struct perf_counter *leader, *sub;
2000 u64 nr = counter->nr_siblings;
2002 perf_output_put(&handle, nr);
2004 leader = counter->group_leader;
2005 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2007 sub->pmu->read(sub);
2009 group_entry.event = sub->hw_event.config;
2010 group_entry.counter = atomic64_read(&sub->count);
2012 perf_output_put(&handle, group_entry);
2017 perf_output_copy(&handle, callchain, callchain_size);
2019 perf_output_end(&handle);
2026 struct perf_comm_event {
2027 struct task_struct *task;
2032 struct perf_event_header header;
2039 static void perf_counter_comm_output(struct perf_counter *counter,
2040 struct perf_comm_event *comm_event)
2042 struct perf_output_handle handle;
2043 int size = comm_event->event.header.size;
2044 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2049 perf_output_put(&handle, comm_event->event);
2050 perf_output_copy(&handle, comm_event->comm,
2051 comm_event->comm_size);
2052 perf_output_end(&handle);
2055 static int perf_counter_comm_match(struct perf_counter *counter,
2056 struct perf_comm_event *comm_event)
2058 if (counter->hw_event.comm &&
2059 comm_event->event.header.type == PERF_EVENT_COMM)
2065 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2066 struct perf_comm_event *comm_event)
2068 struct perf_counter *counter;
2070 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2074 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2075 if (perf_counter_comm_match(counter, comm_event))
2076 perf_counter_comm_output(counter, comm_event);
2081 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2083 struct perf_cpu_context *cpuctx;
2085 char *comm = comm_event->task->comm;
2087 size = ALIGN(strlen(comm)+1, sizeof(u64));
2089 comm_event->comm = comm;
2090 comm_event->comm_size = size;
2092 comm_event->event.header.size = sizeof(comm_event->event) + size;
2094 cpuctx = &get_cpu_var(perf_cpu_context);
2095 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2096 put_cpu_var(perf_cpu_context);
2098 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
2101 void perf_counter_comm(struct task_struct *task)
2103 struct perf_comm_event comm_event;
2105 if (!atomic_read(&nr_comm_tracking))
2108 comm_event = (struct perf_comm_event){
2111 .header = { .type = PERF_EVENT_COMM, },
2112 .pid = task->group_leader->pid,
2117 perf_counter_comm_event(&comm_event);
2124 struct perf_mmap_event {
2130 struct perf_event_header header;
2140 static void perf_counter_mmap_output(struct perf_counter *counter,
2141 struct perf_mmap_event *mmap_event)
2143 struct perf_output_handle handle;
2144 int size = mmap_event->event.header.size;
2145 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2150 perf_output_put(&handle, mmap_event->event);
2151 perf_output_copy(&handle, mmap_event->file_name,
2152 mmap_event->file_size);
2153 perf_output_end(&handle);
2156 static int perf_counter_mmap_match(struct perf_counter *counter,
2157 struct perf_mmap_event *mmap_event)
2159 if (counter->hw_event.mmap &&
2160 mmap_event->event.header.type == PERF_EVENT_MMAP)
2163 if (counter->hw_event.munmap &&
2164 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2170 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2171 struct perf_mmap_event *mmap_event)
2173 struct perf_counter *counter;
2175 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2179 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2180 if (perf_counter_mmap_match(counter, mmap_event))
2181 perf_counter_mmap_output(counter, mmap_event);
2186 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2188 struct perf_cpu_context *cpuctx;
2189 struct file *file = mmap_event->file;
2196 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2198 name = strncpy(tmp, "//enomem", sizeof(tmp));
2201 name = d_path(&file->f_path, buf, PATH_MAX);
2203 name = strncpy(tmp, "//toolong", sizeof(tmp));
2207 name = strncpy(tmp, "//anon", sizeof(tmp));
2212 size = ALIGN(strlen(name)+1, sizeof(u64));
2214 mmap_event->file_name = name;
2215 mmap_event->file_size = size;
2217 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2219 cpuctx = &get_cpu_var(perf_cpu_context);
2220 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2221 put_cpu_var(perf_cpu_context);
2223 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2228 void perf_counter_mmap(unsigned long addr, unsigned long len,
2229 unsigned long pgoff, struct file *file)
2231 struct perf_mmap_event mmap_event;
2233 if (!atomic_read(&nr_mmap_tracking))
2236 mmap_event = (struct perf_mmap_event){
2239 .header = { .type = PERF_EVENT_MMAP, },
2240 .pid = current->group_leader->pid,
2241 .tid = current->pid,
2248 perf_counter_mmap_event(&mmap_event);
2251 void perf_counter_munmap(unsigned long addr, unsigned long len,
2252 unsigned long pgoff, struct file *file)
2254 struct perf_mmap_event mmap_event;
2256 if (!atomic_read(&nr_munmap_tracking))
2259 mmap_event = (struct perf_mmap_event){
2262 .header = { .type = PERF_EVENT_MUNMAP, },
2263 .pid = current->group_leader->pid,
2264 .tid = current->pid,
2271 perf_counter_mmap_event(&mmap_event);
2275 * Generic counter overflow handling.
2278 int perf_counter_overflow(struct perf_counter *counter,
2279 int nmi, struct pt_regs *regs, u64 addr)
2281 int events = atomic_read(&counter->event_limit);
2284 counter->pending_kill = POLL_IN;
2285 if (events && atomic_dec_and_test(&counter->event_limit)) {
2287 counter->pending_kill = POLL_HUP;
2289 counter->pending_disable = 1;
2290 perf_pending_queue(&counter->pending,
2291 perf_pending_counter);
2293 perf_counter_disable(counter);
2296 perf_counter_output(counter, nmi, regs, addr);
2301 * Generic software counter infrastructure
2304 static void perf_swcounter_update(struct perf_counter *counter)
2306 struct hw_perf_counter *hwc = &counter->hw;
2311 prev = atomic64_read(&hwc->prev_count);
2312 now = atomic64_read(&hwc->count);
2313 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2318 atomic64_add(delta, &counter->count);
2319 atomic64_sub(delta, &hwc->period_left);
2322 static void perf_swcounter_set_period(struct perf_counter *counter)
2324 struct hw_perf_counter *hwc = &counter->hw;
2325 s64 left = atomic64_read(&hwc->period_left);
2326 s64 period = hwc->irq_period;
2328 if (unlikely(left <= -period)) {
2330 atomic64_set(&hwc->period_left, left);
2333 if (unlikely(left <= 0)) {
2335 atomic64_add(period, &hwc->period_left);
2338 atomic64_set(&hwc->prev_count, -left);
2339 atomic64_set(&hwc->count, -left);
2342 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2344 enum hrtimer_restart ret = HRTIMER_RESTART;
2345 struct perf_counter *counter;
2346 struct pt_regs *regs;
2348 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2349 counter->pmu->read(counter);
2351 regs = get_irq_regs();
2353 * In case we exclude kernel IPs or are somehow not in interrupt
2354 * context, provide the next best thing, the user IP.
2356 if ((counter->hw_event.exclude_kernel || !regs) &&
2357 !counter->hw_event.exclude_user)
2358 regs = task_pt_regs(current);
2361 if (perf_counter_overflow(counter, 0, regs, 0))
2362 ret = HRTIMER_NORESTART;
2365 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2370 static void perf_swcounter_overflow(struct perf_counter *counter,
2371 int nmi, struct pt_regs *regs, u64 addr)
2373 perf_swcounter_update(counter);
2374 perf_swcounter_set_period(counter);
2375 if (perf_counter_overflow(counter, nmi, regs, addr))
2376 /* soft-disable the counter */
2381 static int perf_swcounter_match(struct perf_counter *counter,
2382 enum perf_event_types type,
2383 u32 event, struct pt_regs *regs)
2385 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2388 if (perf_event_raw(&counter->hw_event))
2391 if (perf_event_type(&counter->hw_event) != type)
2394 if (perf_event_id(&counter->hw_event) != event)
2397 if (counter->hw_event.exclude_user && user_mode(regs))
2400 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2406 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2407 int nmi, struct pt_regs *regs, u64 addr)
2409 int neg = atomic64_add_negative(nr, &counter->hw.count);
2410 if (counter->hw.irq_period && !neg)
2411 perf_swcounter_overflow(counter, nmi, regs, addr);
2414 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2415 enum perf_event_types type, u32 event,
2416 u64 nr, int nmi, struct pt_regs *regs,
2419 struct perf_counter *counter;
2421 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2425 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2426 if (perf_swcounter_match(counter, type, event, regs))
2427 perf_swcounter_add(counter, nr, nmi, regs, addr);
2432 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2435 return &cpuctx->recursion[3];
2438 return &cpuctx->recursion[2];
2441 return &cpuctx->recursion[1];
2443 return &cpuctx->recursion[0];
2446 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2447 u64 nr, int nmi, struct pt_regs *regs,
2450 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2451 int *recursion = perf_swcounter_recursion_context(cpuctx);
2459 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2460 nr, nmi, regs, addr);
2461 if (cpuctx->task_ctx) {
2462 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2463 nr, nmi, regs, addr);
2470 put_cpu_var(perf_cpu_context);
2474 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2476 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2479 static void perf_swcounter_read(struct perf_counter *counter)
2481 perf_swcounter_update(counter);
2484 static int perf_swcounter_enable(struct perf_counter *counter)
2486 perf_swcounter_set_period(counter);
2490 static void perf_swcounter_disable(struct perf_counter *counter)
2492 perf_swcounter_update(counter);
2495 static const struct pmu perf_ops_generic = {
2496 .enable = perf_swcounter_enable,
2497 .disable = perf_swcounter_disable,
2498 .read = perf_swcounter_read,
2502 * Software counter: cpu wall time clock
2505 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2507 int cpu = raw_smp_processor_id();
2511 now = cpu_clock(cpu);
2512 prev = atomic64_read(&counter->hw.prev_count);
2513 atomic64_set(&counter->hw.prev_count, now);
2514 atomic64_add(now - prev, &counter->count);
2517 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2519 struct hw_perf_counter *hwc = &counter->hw;
2520 int cpu = raw_smp_processor_id();
2522 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2523 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2524 hwc->hrtimer.function = perf_swcounter_hrtimer;
2525 if (hwc->irq_period) {
2526 __hrtimer_start_range_ns(&hwc->hrtimer,
2527 ns_to_ktime(hwc->irq_period), 0,
2528 HRTIMER_MODE_REL, 0);
2534 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2536 hrtimer_cancel(&counter->hw.hrtimer);
2537 cpu_clock_perf_counter_update(counter);
2540 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2542 cpu_clock_perf_counter_update(counter);
2545 static const struct pmu perf_ops_cpu_clock = {
2546 .enable = cpu_clock_perf_counter_enable,
2547 .disable = cpu_clock_perf_counter_disable,
2548 .read = cpu_clock_perf_counter_read,
2552 * Software counter: task time clock
2555 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2560 prev = atomic64_xchg(&counter->hw.prev_count, now);
2562 atomic64_add(delta, &counter->count);
2565 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2567 struct hw_perf_counter *hwc = &counter->hw;
2570 now = counter->ctx->time;
2572 atomic64_set(&hwc->prev_count, now);
2573 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2574 hwc->hrtimer.function = perf_swcounter_hrtimer;
2575 if (hwc->irq_period) {
2576 __hrtimer_start_range_ns(&hwc->hrtimer,
2577 ns_to_ktime(hwc->irq_period), 0,
2578 HRTIMER_MODE_REL, 0);
2584 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2586 hrtimer_cancel(&counter->hw.hrtimer);
2587 task_clock_perf_counter_update(counter, counter->ctx->time);
2591 static void task_clock_perf_counter_read(struct perf_counter *counter)
2596 update_context_time(counter->ctx);
2597 time = counter->ctx->time;
2599 u64 now = perf_clock();
2600 u64 delta = now - counter->ctx->timestamp;
2601 time = counter->ctx->time + delta;
2604 task_clock_perf_counter_update(counter, time);
2607 static const struct pmu perf_ops_task_clock = {
2608 .enable = task_clock_perf_counter_enable,
2609 .disable = task_clock_perf_counter_disable,
2610 .read = task_clock_perf_counter_read,
2614 * Software counter: cpu migrations
2617 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2619 struct task_struct *curr = counter->ctx->task;
2622 return curr->se.nr_migrations;
2623 return cpu_nr_migrations(smp_processor_id());
2626 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2631 prev = atomic64_read(&counter->hw.prev_count);
2632 now = get_cpu_migrations(counter);
2634 atomic64_set(&counter->hw.prev_count, now);
2638 atomic64_add(delta, &counter->count);
2641 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2643 cpu_migrations_perf_counter_update(counter);
2646 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2648 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2649 atomic64_set(&counter->hw.prev_count,
2650 get_cpu_migrations(counter));
2654 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2656 cpu_migrations_perf_counter_update(counter);
2659 static const struct pmu perf_ops_cpu_migrations = {
2660 .enable = cpu_migrations_perf_counter_enable,
2661 .disable = cpu_migrations_perf_counter_disable,
2662 .read = cpu_migrations_perf_counter_read,
2665 #ifdef CONFIG_EVENT_PROFILE
2666 void perf_tpcounter_event(int event_id)
2668 struct pt_regs *regs = get_irq_regs();
2671 regs = task_pt_regs(current);
2673 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2675 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2677 extern int ftrace_profile_enable(int);
2678 extern void ftrace_profile_disable(int);
2680 static void tp_perf_counter_destroy(struct perf_counter *counter)
2682 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2685 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2687 int event_id = perf_event_id(&counter->hw_event);
2690 ret = ftrace_profile_enable(event_id);
2694 counter->destroy = tp_perf_counter_destroy;
2695 counter->hw.irq_period = counter->hw_event.irq_period;
2697 return &perf_ops_generic;
2700 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2706 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2708 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2709 const struct pmu *pmu = NULL;
2710 struct hw_perf_counter *hwc = &counter->hw;
2713 * Software counters (currently) can't in general distinguish
2714 * between user, kernel and hypervisor events.
2715 * However, context switches and cpu migrations are considered
2716 * to be kernel events, and page faults are never hypervisor
2719 switch (perf_event_id(&counter->hw_event)) {
2720 case PERF_COUNT_CPU_CLOCK:
2721 pmu = &perf_ops_cpu_clock;
2723 if (hw_event->irq_period && hw_event->irq_period < 10000)
2724 hw_event->irq_period = 10000;
2726 case PERF_COUNT_TASK_CLOCK:
2728 * If the user instantiates this as a per-cpu counter,
2729 * use the cpu_clock counter instead.
2731 if (counter->ctx->task)
2732 pmu = &perf_ops_task_clock;
2734 pmu = &perf_ops_cpu_clock;
2736 if (hw_event->irq_period && hw_event->irq_period < 10000)
2737 hw_event->irq_period = 10000;
2739 case PERF_COUNT_PAGE_FAULTS:
2740 case PERF_COUNT_PAGE_FAULTS_MIN:
2741 case PERF_COUNT_PAGE_FAULTS_MAJ:
2742 case PERF_COUNT_CONTEXT_SWITCHES:
2743 pmu = &perf_ops_generic;
2745 case PERF_COUNT_CPU_MIGRATIONS:
2746 if (!counter->hw_event.exclude_kernel)
2747 pmu = &perf_ops_cpu_migrations;
2752 hwc->irq_period = hw_event->irq_period;
2758 * Allocate and initialize a counter structure
2760 static struct perf_counter *
2761 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2763 struct perf_counter_context *ctx,
2764 struct perf_counter *group_leader,
2767 const struct pmu *pmu;
2768 struct perf_counter *counter;
2771 counter = kzalloc(sizeof(*counter), gfpflags);
2773 return ERR_PTR(-ENOMEM);
2776 * Single counters are their own group leaders, with an
2777 * empty sibling list:
2780 group_leader = counter;
2782 mutex_init(&counter->mutex);
2783 INIT_LIST_HEAD(&counter->list_entry);
2784 INIT_LIST_HEAD(&counter->event_entry);
2785 INIT_LIST_HEAD(&counter->sibling_list);
2786 init_waitqueue_head(&counter->waitq);
2788 mutex_init(&counter->mmap_mutex);
2790 INIT_LIST_HEAD(&counter->child_list);
2793 counter->hw_event = *hw_event;
2794 counter->group_leader = group_leader;
2795 counter->pmu = NULL;
2798 counter->state = PERF_COUNTER_STATE_INACTIVE;
2799 if (hw_event->disabled)
2800 counter->state = PERF_COUNTER_STATE_OFF;
2804 if (perf_event_raw(hw_event)) {
2805 pmu = hw_perf_counter_init(counter);
2809 switch (perf_event_type(hw_event)) {
2810 case PERF_TYPE_HARDWARE:
2811 pmu = hw_perf_counter_init(counter);
2814 case PERF_TYPE_SOFTWARE:
2815 pmu = sw_perf_counter_init(counter);
2818 case PERF_TYPE_TRACEPOINT:
2819 pmu = tp_perf_counter_init(counter);
2826 else if (IS_ERR(pmu))
2831 return ERR_PTR(err);
2836 if (counter->hw_event.mmap)
2837 atomic_inc(&nr_mmap_tracking);
2838 if (counter->hw_event.munmap)
2839 atomic_inc(&nr_munmap_tracking);
2840 if (counter->hw_event.comm)
2841 atomic_inc(&nr_comm_tracking);
2847 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2849 * @hw_event_uptr: event type attributes for monitoring/sampling
2852 * @group_fd: group leader counter fd
2854 SYSCALL_DEFINE5(perf_counter_open,
2855 const struct perf_counter_hw_event __user *, hw_event_uptr,
2856 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2858 struct perf_counter *counter, *group_leader;
2859 struct perf_counter_hw_event hw_event;
2860 struct perf_counter_context *ctx;
2861 struct file *counter_file = NULL;
2862 struct file *group_file = NULL;
2863 int fput_needed = 0;
2864 int fput_needed2 = 0;
2867 /* for future expandability... */
2871 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2875 * Get the target context (task or percpu):
2877 ctx = find_get_context(pid, cpu);
2879 return PTR_ERR(ctx);
2882 * Look up the group leader (we will attach this counter to it):
2884 group_leader = NULL;
2885 if (group_fd != -1) {
2887 group_file = fget_light(group_fd, &fput_needed);
2889 goto err_put_context;
2890 if (group_file->f_op != &perf_fops)
2891 goto err_put_context;
2893 group_leader = group_file->private_data;
2895 * Do not allow a recursive hierarchy (this new sibling
2896 * becoming part of another group-sibling):
2898 if (group_leader->group_leader != group_leader)
2899 goto err_put_context;
2901 * Do not allow to attach to a group in a different
2902 * task or CPU context:
2904 if (group_leader->ctx != ctx)
2905 goto err_put_context;
2907 * Only a group leader can be exclusive or pinned
2909 if (hw_event.exclusive || hw_event.pinned)
2910 goto err_put_context;
2913 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2915 ret = PTR_ERR(counter);
2916 if (IS_ERR(counter))
2917 goto err_put_context;
2919 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2921 goto err_free_put_context;
2923 counter_file = fget_light(ret, &fput_needed2);
2925 goto err_free_put_context;
2927 counter->filp = counter_file;
2928 mutex_lock(&ctx->mutex);
2929 perf_install_in_context(ctx, counter, cpu);
2930 mutex_unlock(&ctx->mutex);
2932 fput_light(counter_file, fput_needed2);
2935 fput_light(group_file, fput_needed);
2939 err_free_put_context:
2949 * Initialize the perf_counter context in a task_struct:
2952 __perf_counter_init_context(struct perf_counter_context *ctx,
2953 struct task_struct *task)
2955 memset(ctx, 0, sizeof(*ctx));
2956 spin_lock_init(&ctx->lock);
2957 mutex_init(&ctx->mutex);
2958 INIT_LIST_HEAD(&ctx->counter_list);
2959 INIT_LIST_HEAD(&ctx->event_list);
2964 * inherit a counter from parent task to child task:
2966 static struct perf_counter *
2967 inherit_counter(struct perf_counter *parent_counter,
2968 struct task_struct *parent,
2969 struct perf_counter_context *parent_ctx,
2970 struct task_struct *child,
2971 struct perf_counter *group_leader,
2972 struct perf_counter_context *child_ctx)
2974 struct perf_counter *child_counter;
2977 * Instead of creating recursive hierarchies of counters,
2978 * we link inherited counters back to the original parent,
2979 * which has a filp for sure, which we use as the reference
2982 if (parent_counter->parent)
2983 parent_counter = parent_counter->parent;
2985 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2986 parent_counter->cpu, child_ctx,
2987 group_leader, GFP_KERNEL);
2988 if (IS_ERR(child_counter))
2989 return child_counter;
2992 * Link it up in the child's context:
2994 child_counter->task = child;
2995 add_counter_to_ctx(child_counter, child_ctx);
2997 child_counter->parent = parent_counter;
2999 * inherit into child's child as well:
3001 child_counter->hw_event.inherit = 1;
3004 * Get a reference to the parent filp - we will fput it
3005 * when the child counter exits. This is safe to do because
3006 * we are in the parent and we know that the filp still
3007 * exists and has a nonzero count:
3009 atomic_long_inc(&parent_counter->filp->f_count);
3012 * Link this into the parent counter's child list
3014 mutex_lock(&parent_counter->mutex);
3015 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3018 * Make the child state follow the state of the parent counter,
3019 * not its hw_event.disabled bit. We hold the parent's mutex,
3020 * so we won't race with perf_counter_{en,dis}able_family.
3022 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3023 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3025 child_counter->state = PERF_COUNTER_STATE_OFF;
3027 mutex_unlock(&parent_counter->mutex);
3029 return child_counter;
3032 static int inherit_group(struct perf_counter *parent_counter,
3033 struct task_struct *parent,
3034 struct perf_counter_context *parent_ctx,
3035 struct task_struct *child,
3036 struct perf_counter_context *child_ctx)
3038 struct perf_counter *leader;
3039 struct perf_counter *sub;
3040 struct perf_counter *child_ctr;
3042 leader = inherit_counter(parent_counter, parent, parent_ctx,
3043 child, NULL, child_ctx);
3045 return PTR_ERR(leader);
3046 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3047 child_ctr = inherit_counter(sub, parent, parent_ctx,
3048 child, leader, child_ctx);
3049 if (IS_ERR(child_ctr))
3050 return PTR_ERR(child_ctr);
3055 static void sync_child_counter(struct perf_counter *child_counter,
3056 struct perf_counter *parent_counter)
3058 u64 parent_val, child_val;
3060 parent_val = atomic64_read(&parent_counter->count);
3061 child_val = atomic64_read(&child_counter->count);
3064 * Add back the child's count to the parent's count:
3066 atomic64_add(child_val, &parent_counter->count);
3067 atomic64_add(child_counter->total_time_enabled,
3068 &parent_counter->child_total_time_enabled);
3069 atomic64_add(child_counter->total_time_running,
3070 &parent_counter->child_total_time_running);
3073 * Remove this counter from the parent's list
3075 mutex_lock(&parent_counter->mutex);
3076 list_del_init(&child_counter->child_list);
3077 mutex_unlock(&parent_counter->mutex);
3080 * Release the parent counter, if this was the last
3083 fput(parent_counter->filp);
3087 __perf_counter_exit_task(struct task_struct *child,
3088 struct perf_counter *child_counter,
3089 struct perf_counter_context *child_ctx)
3091 struct perf_counter *parent_counter;
3092 struct perf_counter *sub, *tmp;
3095 * If we do not self-reap then we have to wait for the
3096 * child task to unschedule (it will happen for sure),
3097 * so that its counter is at its final count. (This
3098 * condition triggers rarely - child tasks usually get
3099 * off their CPU before the parent has a chance to
3100 * get this far into the reaping action)
3102 if (child != current) {
3103 wait_task_inactive(child, 0);
3104 list_del_init(&child_counter->list_entry);
3105 update_counter_times(child_counter);
3107 struct perf_cpu_context *cpuctx;
3108 unsigned long flags;
3112 * Disable and unlink this counter.
3114 * Be careful about zapping the list - IRQ/NMI context
3115 * could still be processing it:
3117 local_irq_save(flags);
3118 perf_flags = hw_perf_save_disable();
3120 cpuctx = &__get_cpu_var(perf_cpu_context);
3122 group_sched_out(child_counter, cpuctx, child_ctx);
3123 update_counter_times(child_counter);
3125 list_del_init(&child_counter->list_entry);
3127 child_ctx->nr_counters--;
3129 hw_perf_restore(perf_flags);
3130 local_irq_restore(flags);
3133 parent_counter = child_counter->parent;
3135 * It can happen that parent exits first, and has counters
3136 * that are still around due to the child reference. These
3137 * counters need to be zapped - but otherwise linger.
3139 if (parent_counter) {
3140 sync_child_counter(child_counter, parent_counter);
3141 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
3144 sync_child_counter(sub, sub->parent);
3148 free_counter(child_counter);
3153 * When a child task exits, feed back counter values to parent counters.
3155 * Note: we may be running in child context, but the PID is not hashed
3156 * anymore so new counters will not be added.
3158 void perf_counter_exit_task(struct task_struct *child)
3160 struct perf_counter *child_counter, *tmp;
3161 struct perf_counter_context *child_ctx;
3163 child_ctx = &child->perf_counter_ctx;
3165 if (likely(!child_ctx->nr_counters))
3168 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3170 __perf_counter_exit_task(child, child_counter, child_ctx);
3174 * Initialize the perf_counter context in task_struct
3176 void perf_counter_init_task(struct task_struct *child)
3178 struct perf_counter_context *child_ctx, *parent_ctx;
3179 struct perf_counter *counter;
3180 struct task_struct *parent = current;
3182 child_ctx = &child->perf_counter_ctx;
3183 parent_ctx = &parent->perf_counter_ctx;
3185 __perf_counter_init_context(child_ctx, child);
3188 * This is executed from the parent task context, so inherit
3189 * counters that have been marked for cloning:
3192 if (likely(!parent_ctx->nr_counters))
3196 * Lock the parent list. No need to lock the child - not PID
3197 * hashed yet and not running, so nobody can access it.
3199 mutex_lock(&parent_ctx->mutex);
3202 * We dont have to disable NMIs - we are only looking at
3203 * the list, not manipulating it:
3205 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
3206 if (!counter->hw_event.inherit)
3209 if (inherit_group(counter, parent,
3210 parent_ctx, child, child_ctx))
3214 mutex_unlock(&parent_ctx->mutex);
3217 static void __cpuinit perf_counter_init_cpu(int cpu)
3219 struct perf_cpu_context *cpuctx;
3221 cpuctx = &per_cpu(perf_cpu_context, cpu);
3222 __perf_counter_init_context(&cpuctx->ctx, NULL);
3224 spin_lock(&perf_resource_lock);
3225 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3226 spin_unlock(&perf_resource_lock);
3228 hw_perf_counter_setup(cpu);
3231 #ifdef CONFIG_HOTPLUG_CPU
3232 static void __perf_counter_exit_cpu(void *info)
3234 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3235 struct perf_counter_context *ctx = &cpuctx->ctx;
3236 struct perf_counter *counter, *tmp;
3238 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3239 __perf_counter_remove_from_context(counter);
3241 static void perf_counter_exit_cpu(int cpu)
3243 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3244 struct perf_counter_context *ctx = &cpuctx->ctx;
3246 mutex_lock(&ctx->mutex);
3247 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3248 mutex_unlock(&ctx->mutex);
3251 static inline void perf_counter_exit_cpu(int cpu) { }
3254 static int __cpuinit
3255 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3257 unsigned int cpu = (long)hcpu;
3261 case CPU_UP_PREPARE:
3262 case CPU_UP_PREPARE_FROZEN:
3263 perf_counter_init_cpu(cpu);
3266 case CPU_DOWN_PREPARE:
3267 case CPU_DOWN_PREPARE_FROZEN:
3268 perf_counter_exit_cpu(cpu);
3278 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3279 .notifier_call = perf_cpu_notify,
3282 void __init perf_counter_init(void)
3284 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3285 (void *)(long)smp_processor_id());
3286 register_cpu_notifier(&perf_cpu_nb);
3289 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3291 return sprintf(buf, "%d\n", perf_reserved_percpu);
3295 perf_set_reserve_percpu(struct sysdev_class *class,
3299 struct perf_cpu_context *cpuctx;
3303 err = strict_strtoul(buf, 10, &val);
3306 if (val > perf_max_counters)
3309 spin_lock(&perf_resource_lock);
3310 perf_reserved_percpu = val;
3311 for_each_online_cpu(cpu) {
3312 cpuctx = &per_cpu(perf_cpu_context, cpu);
3313 spin_lock_irq(&cpuctx->ctx.lock);
3314 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3315 perf_max_counters - perf_reserved_percpu);
3316 cpuctx->max_pertask = mpt;
3317 spin_unlock_irq(&cpuctx->ctx.lock);
3319 spin_unlock(&perf_resource_lock);
3324 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3326 return sprintf(buf, "%d\n", perf_overcommit);
3330 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3335 err = strict_strtoul(buf, 10, &val);
3341 spin_lock(&perf_resource_lock);
3342 perf_overcommit = val;
3343 spin_unlock(&perf_resource_lock);
3348 static SYSDEV_CLASS_ATTR(
3351 perf_show_reserve_percpu,
3352 perf_set_reserve_percpu
3355 static SYSDEV_CLASS_ATTR(
3358 perf_show_overcommit,
3362 static struct attribute *perfclass_attrs[] = {
3363 &attr_reserve_percpu.attr,
3364 &attr_overcommit.attr,
3368 static struct attribute_group perfclass_attr_group = {
3369 .attrs = perfclass_attrs,
3370 .name = "perf_counters",
3373 static int __init perf_counter_sysfs_init(void)
3375 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3376 &perfclass_attr_group);
3378 device_initcall(perf_counter_sysfs_init);