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 int perf_counter_refresh(struct perf_counter *counter, int refresh)
744 * not supported on inherited counters
746 if (counter->hw_event.inherit)
749 atomic_add(refresh, &counter->event_limit);
750 perf_counter_enable(counter);
756 * Enable a counter and all its children.
758 static void perf_counter_enable_family(struct perf_counter *counter)
760 struct perf_counter *child;
762 perf_counter_enable(counter);
765 * Lock the mutex to protect the list of children
767 mutex_lock(&counter->mutex);
768 list_for_each_entry(child, &counter->child_list, child_list)
769 perf_counter_enable(child);
770 mutex_unlock(&counter->mutex);
773 void __perf_counter_sched_out(struct perf_counter_context *ctx,
774 struct perf_cpu_context *cpuctx)
776 struct perf_counter *counter;
779 spin_lock(&ctx->lock);
781 if (likely(!ctx->nr_counters))
783 update_context_time(ctx);
785 flags = hw_perf_save_disable();
786 if (ctx->nr_active) {
787 list_for_each_entry(counter, &ctx->counter_list, list_entry)
788 group_sched_out(counter, cpuctx, ctx);
790 hw_perf_restore(flags);
792 spin_unlock(&ctx->lock);
796 * Called from scheduler to remove the counters of the current task,
797 * with interrupts disabled.
799 * We stop each counter and update the counter value in counter->count.
801 * This does not protect us against NMI, but disable()
802 * sets the disabled bit in the control field of counter _before_
803 * accessing the counter control register. If a NMI hits, then it will
804 * not restart the counter.
806 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
808 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
809 struct perf_counter_context *ctx = &task->perf_counter_ctx;
810 struct pt_regs *regs;
812 if (likely(!cpuctx->task_ctx))
815 update_context_time(ctx);
817 regs = task_pt_regs(task);
818 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
819 __perf_counter_sched_out(ctx, cpuctx);
821 cpuctx->task_ctx = NULL;
824 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
826 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
830 group_sched_in(struct perf_counter *group_counter,
831 struct perf_cpu_context *cpuctx,
832 struct perf_counter_context *ctx,
835 struct perf_counter *counter, *partial_group;
838 if (group_counter->state == PERF_COUNTER_STATE_OFF)
841 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
843 return ret < 0 ? ret : 0;
845 group_counter->prev_state = group_counter->state;
846 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
850 * Schedule in siblings as one group (if any):
852 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
853 counter->prev_state = counter->state;
854 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
855 partial_group = counter;
864 * Groups can be scheduled in as one unit only, so undo any
865 * partial group before returning:
867 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
868 if (counter == partial_group)
870 counter_sched_out(counter, cpuctx, ctx);
872 counter_sched_out(group_counter, cpuctx, ctx);
878 __perf_counter_sched_in(struct perf_counter_context *ctx,
879 struct perf_cpu_context *cpuctx, int cpu)
881 struct perf_counter *counter;
885 spin_lock(&ctx->lock);
887 if (likely(!ctx->nr_counters))
890 ctx->timestamp = perf_clock();
892 flags = hw_perf_save_disable();
895 * First go through the list and put on any pinned groups
896 * in order to give them the best chance of going on.
898 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
899 if (counter->state <= PERF_COUNTER_STATE_OFF ||
900 !counter->hw_event.pinned)
902 if (counter->cpu != -1 && counter->cpu != cpu)
905 if (group_can_go_on(counter, cpuctx, 1))
906 group_sched_in(counter, cpuctx, ctx, cpu);
909 * If this pinned group hasn't been scheduled,
910 * put it in error state.
912 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
913 update_group_times(counter);
914 counter->state = PERF_COUNTER_STATE_ERROR;
918 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
920 * Ignore counters in OFF or ERROR state, and
921 * ignore pinned counters since we did them already.
923 if (counter->state <= PERF_COUNTER_STATE_OFF ||
924 counter->hw_event.pinned)
928 * Listen to the 'cpu' scheduling filter constraint
931 if (counter->cpu != -1 && counter->cpu != cpu)
934 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
935 if (group_sched_in(counter, cpuctx, ctx, cpu))
939 hw_perf_restore(flags);
941 spin_unlock(&ctx->lock);
945 * Called from scheduler to add the counters of the current task
946 * with interrupts disabled.
948 * We restore the counter value and then enable it.
950 * This does not protect us against NMI, but enable()
951 * sets the enabled bit in the control field of counter _before_
952 * accessing the counter control register. If a NMI hits, then it will
953 * keep the counter running.
955 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
957 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
958 struct perf_counter_context *ctx = &task->perf_counter_ctx;
960 __perf_counter_sched_in(ctx, cpuctx, cpu);
961 cpuctx->task_ctx = ctx;
964 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
966 struct perf_counter_context *ctx = &cpuctx->ctx;
968 __perf_counter_sched_in(ctx, cpuctx, cpu);
971 int perf_counter_task_disable(void)
973 struct task_struct *curr = current;
974 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
975 struct perf_counter *counter;
980 if (likely(!ctx->nr_counters))
983 local_irq_save(flags);
984 cpu = smp_processor_id();
986 perf_counter_task_sched_out(curr, cpu);
988 spin_lock(&ctx->lock);
991 * Disable all the counters:
993 perf_flags = hw_perf_save_disable();
995 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
996 if (counter->state != PERF_COUNTER_STATE_ERROR) {
997 update_group_times(counter);
998 counter->state = PERF_COUNTER_STATE_OFF;
1002 hw_perf_restore(perf_flags);
1004 spin_unlock_irqrestore(&ctx->lock, flags);
1009 int perf_counter_task_enable(void)
1011 struct task_struct *curr = current;
1012 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1013 struct perf_counter *counter;
1014 unsigned long flags;
1018 if (likely(!ctx->nr_counters))
1021 local_irq_save(flags);
1022 cpu = smp_processor_id();
1024 perf_counter_task_sched_out(curr, cpu);
1026 spin_lock(&ctx->lock);
1029 * Disable all the counters:
1031 perf_flags = hw_perf_save_disable();
1033 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1034 if (counter->state > PERF_COUNTER_STATE_OFF)
1036 counter->state = PERF_COUNTER_STATE_INACTIVE;
1037 counter->tstamp_enabled =
1038 ctx->time - counter->total_time_enabled;
1039 counter->hw_event.disabled = 0;
1041 hw_perf_restore(perf_flags);
1043 spin_unlock(&ctx->lock);
1045 perf_counter_task_sched_in(curr, cpu);
1047 local_irq_restore(flags);
1053 * Round-robin a context's counters:
1055 static void rotate_ctx(struct perf_counter_context *ctx)
1057 struct perf_counter *counter;
1060 if (!ctx->nr_counters)
1063 spin_lock(&ctx->lock);
1065 * Rotate the first entry last (works just fine for group counters too):
1067 perf_flags = hw_perf_save_disable();
1068 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1069 list_move_tail(&counter->list_entry, &ctx->counter_list);
1072 hw_perf_restore(perf_flags);
1074 spin_unlock(&ctx->lock);
1077 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1079 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1080 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1082 perf_counter_cpu_sched_out(cpuctx);
1083 perf_counter_task_sched_out(curr, cpu);
1085 rotate_ctx(&cpuctx->ctx);
1088 perf_counter_cpu_sched_in(cpuctx, cpu);
1089 perf_counter_task_sched_in(curr, cpu);
1093 * Cross CPU call to read the hardware counter
1095 static void __read(void *info)
1097 struct perf_counter *counter = info;
1098 struct perf_counter_context *ctx = counter->ctx;
1099 unsigned long flags;
1101 local_irq_save(flags);
1103 update_context_time(ctx);
1104 counter->pmu->read(counter);
1105 update_counter_times(counter);
1106 local_irq_restore(flags);
1109 static u64 perf_counter_read(struct perf_counter *counter)
1112 * If counter is enabled and currently active on a CPU, update the
1113 * value in the counter structure:
1115 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1116 smp_call_function_single(counter->oncpu,
1117 __read, counter, 1);
1118 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1119 update_counter_times(counter);
1122 return atomic64_read(&counter->count);
1125 static void put_context(struct perf_counter_context *ctx)
1128 put_task_struct(ctx->task);
1131 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1133 struct perf_cpu_context *cpuctx;
1134 struct perf_counter_context *ctx;
1135 struct task_struct *task;
1138 * If cpu is not a wildcard then this is a percpu counter:
1141 /* Must be root to operate on a CPU counter: */
1142 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1143 return ERR_PTR(-EACCES);
1145 if (cpu < 0 || cpu > num_possible_cpus())
1146 return ERR_PTR(-EINVAL);
1149 * We could be clever and allow to attach a counter to an
1150 * offline CPU and activate it when the CPU comes up, but
1153 if (!cpu_isset(cpu, cpu_online_map))
1154 return ERR_PTR(-ENODEV);
1156 cpuctx = &per_cpu(perf_cpu_context, cpu);
1166 task = find_task_by_vpid(pid);
1168 get_task_struct(task);
1172 return ERR_PTR(-ESRCH);
1174 ctx = &task->perf_counter_ctx;
1177 /* Reuse ptrace permission checks for now. */
1178 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1180 return ERR_PTR(-EACCES);
1186 static void free_counter_rcu(struct rcu_head *head)
1188 struct perf_counter *counter;
1190 counter = container_of(head, struct perf_counter, rcu_head);
1194 static void perf_pending_sync(struct perf_counter *counter);
1196 static void free_counter(struct perf_counter *counter)
1198 perf_pending_sync(counter);
1200 if (counter->hw_event.mmap)
1201 atomic_dec(&nr_mmap_tracking);
1202 if (counter->hw_event.munmap)
1203 atomic_dec(&nr_munmap_tracking);
1204 if (counter->hw_event.comm)
1205 atomic_dec(&nr_comm_tracking);
1207 if (counter->destroy)
1208 counter->destroy(counter);
1210 call_rcu(&counter->rcu_head, free_counter_rcu);
1214 * Called when the last reference to the file is gone.
1216 static int perf_release(struct inode *inode, struct file *file)
1218 struct perf_counter *counter = file->private_data;
1219 struct perf_counter_context *ctx = counter->ctx;
1221 file->private_data = NULL;
1223 mutex_lock(&ctx->mutex);
1224 mutex_lock(&counter->mutex);
1226 perf_counter_remove_from_context(counter);
1228 mutex_unlock(&counter->mutex);
1229 mutex_unlock(&ctx->mutex);
1231 free_counter(counter);
1238 * Read the performance counter - simple non blocking version for now
1241 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1247 * Return end-of-file for a read on a counter that is in
1248 * error state (i.e. because it was pinned but it couldn't be
1249 * scheduled on to the CPU at some point).
1251 if (counter->state == PERF_COUNTER_STATE_ERROR)
1254 mutex_lock(&counter->mutex);
1255 values[0] = perf_counter_read(counter);
1257 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1258 values[n++] = counter->total_time_enabled +
1259 atomic64_read(&counter->child_total_time_enabled);
1260 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1261 values[n++] = counter->total_time_running +
1262 atomic64_read(&counter->child_total_time_running);
1263 mutex_unlock(&counter->mutex);
1265 if (count < n * sizeof(u64))
1267 count = n * sizeof(u64);
1269 if (copy_to_user(buf, values, count))
1276 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1278 struct perf_counter *counter = file->private_data;
1280 return perf_read_hw(counter, buf, count);
1283 static unsigned int perf_poll(struct file *file, poll_table *wait)
1285 struct perf_counter *counter = file->private_data;
1286 struct perf_mmap_data *data;
1287 unsigned int events = POLL_HUP;
1290 data = rcu_dereference(counter->data);
1292 events = atomic_xchg(&data->poll, 0);
1295 poll_wait(file, &counter->waitq, wait);
1300 static void perf_counter_reset(struct perf_counter *counter)
1302 atomic_set(&counter->count, 0);
1305 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1307 struct perf_counter *counter = file->private_data;
1311 case PERF_COUNTER_IOC_ENABLE:
1312 perf_counter_enable_family(counter);
1314 case PERF_COUNTER_IOC_DISABLE:
1315 perf_counter_disable_family(counter);
1317 case PERF_COUNTER_IOC_REFRESH:
1318 err = perf_counter_refresh(counter, arg);
1320 case PERF_COUNTER_IOC_RESET:
1321 perf_counter_reset(counter);
1330 * Callers need to ensure there can be no nesting of this function, otherwise
1331 * the seqlock logic goes bad. We can not serialize this because the arch
1332 * code calls this from NMI context.
1334 void perf_counter_update_userpage(struct perf_counter *counter)
1336 struct perf_mmap_data *data;
1337 struct perf_counter_mmap_page *userpg;
1340 data = rcu_dereference(counter->data);
1344 userpg = data->user_page;
1347 * Disable preemption so as to not let the corresponding user-space
1348 * spin too long if we get preempted.
1353 userpg->index = counter->hw.idx;
1354 userpg->offset = atomic64_read(&counter->count);
1355 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1356 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1365 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1367 struct perf_counter *counter = vma->vm_file->private_data;
1368 struct perf_mmap_data *data;
1369 int ret = VM_FAULT_SIGBUS;
1372 data = rcu_dereference(counter->data);
1376 if (vmf->pgoff == 0) {
1377 vmf->page = virt_to_page(data->user_page);
1379 int nr = vmf->pgoff - 1;
1381 if ((unsigned)nr > data->nr_pages)
1384 vmf->page = virt_to_page(data->data_pages[nr]);
1386 get_page(vmf->page);
1394 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1396 struct perf_mmap_data *data;
1400 WARN_ON(atomic_read(&counter->mmap_count));
1402 size = sizeof(struct perf_mmap_data);
1403 size += nr_pages * sizeof(void *);
1405 data = kzalloc(size, GFP_KERNEL);
1409 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1410 if (!data->user_page)
1411 goto fail_user_page;
1413 for (i = 0; i < nr_pages; i++) {
1414 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1415 if (!data->data_pages[i])
1416 goto fail_data_pages;
1419 data->nr_pages = nr_pages;
1420 atomic_set(&data->lock, -1);
1422 rcu_assign_pointer(counter->data, data);
1427 for (i--; i >= 0; i--)
1428 free_page((unsigned long)data->data_pages[i]);
1430 free_page((unsigned long)data->user_page);
1439 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1441 struct perf_mmap_data *data = container_of(rcu_head,
1442 struct perf_mmap_data, rcu_head);
1445 free_page((unsigned long)data->user_page);
1446 for (i = 0; i < data->nr_pages; i++)
1447 free_page((unsigned long)data->data_pages[i]);
1451 static void perf_mmap_data_free(struct perf_counter *counter)
1453 struct perf_mmap_data *data = counter->data;
1455 WARN_ON(atomic_read(&counter->mmap_count));
1457 rcu_assign_pointer(counter->data, NULL);
1458 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1461 static void perf_mmap_open(struct vm_area_struct *vma)
1463 struct perf_counter *counter = vma->vm_file->private_data;
1465 atomic_inc(&counter->mmap_count);
1468 static void perf_mmap_close(struct vm_area_struct *vma)
1470 struct perf_counter *counter = vma->vm_file->private_data;
1472 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1473 &counter->mmap_mutex)) {
1474 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1475 perf_mmap_data_free(counter);
1476 mutex_unlock(&counter->mmap_mutex);
1480 static struct vm_operations_struct perf_mmap_vmops = {
1481 .open = perf_mmap_open,
1482 .close = perf_mmap_close,
1483 .fault = perf_mmap_fault,
1486 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1488 struct perf_counter *counter = file->private_data;
1489 unsigned long vma_size;
1490 unsigned long nr_pages;
1491 unsigned long locked, lock_limit;
1495 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1498 vma_size = vma->vm_end - vma->vm_start;
1499 nr_pages = (vma_size / PAGE_SIZE) - 1;
1502 * If we have data pages ensure they're a power-of-two number, so we
1503 * can do bitmasks instead of modulo.
1505 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1508 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1511 if (vma->vm_pgoff != 0)
1514 mutex_lock(&counter->mmap_mutex);
1515 if (atomic_inc_not_zero(&counter->mmap_count)) {
1516 if (nr_pages != counter->data->nr_pages)
1521 extra = nr_pages /* + 1 only account the data pages */;
1522 extra -= sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1526 locked = vma->vm_mm->locked_vm + extra;
1528 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1529 lock_limit >>= PAGE_SHIFT;
1531 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1536 WARN_ON(counter->data);
1537 ret = perf_mmap_data_alloc(counter, nr_pages);
1541 atomic_set(&counter->mmap_count, 1);
1542 vma->vm_mm->locked_vm += extra;
1543 counter->data->nr_locked = extra;
1545 mutex_unlock(&counter->mmap_mutex);
1547 vma->vm_flags &= ~VM_MAYWRITE;
1548 vma->vm_flags |= VM_RESERVED;
1549 vma->vm_ops = &perf_mmap_vmops;
1554 static int perf_fasync(int fd, struct file *filp, int on)
1556 struct perf_counter *counter = filp->private_data;
1557 struct inode *inode = filp->f_path.dentry->d_inode;
1560 mutex_lock(&inode->i_mutex);
1561 retval = fasync_helper(fd, filp, on, &counter->fasync);
1562 mutex_unlock(&inode->i_mutex);
1570 static const struct file_operations perf_fops = {
1571 .release = perf_release,
1574 .unlocked_ioctl = perf_ioctl,
1575 .compat_ioctl = perf_ioctl,
1577 .fasync = perf_fasync,
1581 * Perf counter wakeup
1583 * If there's data, ensure we set the poll() state and publish everything
1584 * to user-space before waking everybody up.
1587 void perf_counter_wakeup(struct perf_counter *counter)
1589 wake_up_all(&counter->waitq);
1591 if (counter->pending_kill) {
1592 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1593 counter->pending_kill = 0;
1600 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1602 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1603 * single linked list and use cmpxchg() to add entries lockless.
1606 static void perf_pending_counter(struct perf_pending_entry *entry)
1608 struct perf_counter *counter = container_of(entry,
1609 struct perf_counter, pending);
1611 if (counter->pending_disable) {
1612 counter->pending_disable = 0;
1613 perf_counter_disable(counter);
1616 if (counter->pending_wakeup) {
1617 counter->pending_wakeup = 0;
1618 perf_counter_wakeup(counter);
1622 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1624 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1628 static void perf_pending_queue(struct perf_pending_entry *entry,
1629 void (*func)(struct perf_pending_entry *))
1631 struct perf_pending_entry **head;
1633 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1638 head = &get_cpu_var(perf_pending_head);
1641 entry->next = *head;
1642 } while (cmpxchg(head, entry->next, entry) != entry->next);
1644 set_perf_counter_pending();
1646 put_cpu_var(perf_pending_head);
1649 static int __perf_pending_run(void)
1651 struct perf_pending_entry *list;
1654 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1655 while (list != PENDING_TAIL) {
1656 void (*func)(struct perf_pending_entry *);
1657 struct perf_pending_entry *entry = list;
1664 * Ensure we observe the unqueue before we issue the wakeup,
1665 * so that we won't be waiting forever.
1666 * -- see perf_not_pending().
1677 static inline int perf_not_pending(struct perf_counter *counter)
1680 * If we flush on whatever cpu we run, there is a chance we don't
1684 __perf_pending_run();
1688 * Ensure we see the proper queue state before going to sleep
1689 * so that we do not miss the wakeup. -- see perf_pending_handle()
1692 return counter->pending.next == NULL;
1695 static void perf_pending_sync(struct perf_counter *counter)
1697 wait_event(counter->waitq, perf_not_pending(counter));
1700 void perf_counter_do_pending(void)
1702 __perf_pending_run();
1706 * Callchain support -- arch specific
1709 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1718 struct perf_output_handle {
1719 struct perf_counter *counter;
1720 struct perf_mmap_data *data;
1721 unsigned int offset;
1726 unsigned long flags;
1729 static void perf_output_wakeup(struct perf_output_handle *handle)
1731 atomic_set(&handle->data->poll, POLL_IN);
1734 handle->counter->pending_wakeup = 1;
1735 perf_pending_queue(&handle->counter->pending,
1736 perf_pending_counter);
1738 perf_counter_wakeup(handle->counter);
1742 * Curious locking construct.
1744 * We need to ensure a later event doesn't publish a head when a former
1745 * event isn't done writing. However since we need to deal with NMIs we
1746 * cannot fully serialize things.
1748 * What we do is serialize between CPUs so we only have to deal with NMI
1749 * nesting on a single CPU.
1751 * We only publish the head (and generate a wakeup) when the outer-most
1754 static void perf_output_lock(struct perf_output_handle *handle)
1756 struct perf_mmap_data *data = handle->data;
1761 local_irq_save(handle->flags);
1762 cpu = smp_processor_id();
1764 if (in_nmi() && atomic_read(&data->lock) == cpu)
1767 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1773 static void perf_output_unlock(struct perf_output_handle *handle)
1775 struct perf_mmap_data *data = handle->data;
1778 data->done_head = data->head;
1780 if (!handle->locked)
1785 * The xchg implies a full barrier that ensures all writes are done
1786 * before we publish the new head, matched by a rmb() in userspace when
1787 * reading this position.
1789 while ((head = atomic_xchg(&data->done_head, 0)))
1790 data->user_page->data_head = head;
1793 * NMI can happen here, which means we can miss a done_head update.
1796 cpu = atomic_xchg(&data->lock, -1);
1797 WARN_ON_ONCE(cpu != smp_processor_id());
1800 * Therefore we have to validate we did not indeed do so.
1802 if (unlikely(atomic_read(&data->done_head))) {
1804 * Since we had it locked, we can lock it again.
1806 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1812 if (atomic_xchg(&data->wakeup, 0))
1813 perf_output_wakeup(handle);
1815 local_irq_restore(handle->flags);
1818 static int perf_output_begin(struct perf_output_handle *handle,
1819 struct perf_counter *counter, unsigned int size,
1820 int nmi, int overflow)
1822 struct perf_mmap_data *data;
1823 unsigned int offset, head;
1826 * For inherited counters we send all the output towards the parent.
1828 if (counter->parent)
1829 counter = counter->parent;
1832 data = rcu_dereference(counter->data);
1836 handle->data = data;
1837 handle->counter = counter;
1839 handle->overflow = overflow;
1841 if (!data->nr_pages)
1844 perf_output_lock(handle);
1847 offset = head = atomic_read(&data->head);
1849 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1851 handle->offset = offset;
1852 handle->head = head;
1854 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
1855 atomic_set(&data->wakeup, 1);
1860 perf_output_wakeup(handle);
1867 static void perf_output_copy(struct perf_output_handle *handle,
1868 void *buf, unsigned int len)
1870 unsigned int pages_mask;
1871 unsigned int offset;
1875 offset = handle->offset;
1876 pages_mask = handle->data->nr_pages - 1;
1877 pages = handle->data->data_pages;
1880 unsigned int page_offset;
1883 nr = (offset >> PAGE_SHIFT) & pages_mask;
1884 page_offset = offset & (PAGE_SIZE - 1);
1885 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1887 memcpy(pages[nr] + page_offset, buf, size);
1894 handle->offset = offset;
1896 WARN_ON_ONCE(handle->offset > handle->head);
1899 #define perf_output_put(handle, x) \
1900 perf_output_copy((handle), &(x), sizeof(x))
1902 static void perf_output_end(struct perf_output_handle *handle)
1904 struct perf_counter *counter = handle->counter;
1905 struct perf_mmap_data *data = handle->data;
1907 int wakeup_events = counter->hw_event.wakeup_events;
1909 if (handle->overflow && wakeup_events) {
1910 int events = atomic_inc_return(&data->events);
1911 if (events >= wakeup_events) {
1912 atomic_sub(wakeup_events, &data->events);
1913 atomic_set(&data->wakeup, 1);
1917 perf_output_unlock(handle);
1921 static void perf_counter_output(struct perf_counter *counter,
1922 int nmi, struct pt_regs *regs, u64 addr)
1925 u64 record_type = counter->hw_event.record_type;
1926 struct perf_output_handle handle;
1927 struct perf_event_header header;
1936 struct perf_callchain_entry *callchain = NULL;
1937 int callchain_size = 0;
1941 header.size = sizeof(header);
1943 header.misc = PERF_EVENT_MISC_OVERFLOW;
1944 header.misc |= user_mode(regs) ?
1945 PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
1947 if (record_type & PERF_RECORD_IP) {
1948 ip = instruction_pointer(regs);
1949 header.type |= PERF_RECORD_IP;
1950 header.size += sizeof(ip);
1953 if (record_type & PERF_RECORD_TID) {
1954 /* namespace issues */
1955 tid_entry.pid = current->group_leader->pid;
1956 tid_entry.tid = current->pid;
1958 header.type |= PERF_RECORD_TID;
1959 header.size += sizeof(tid_entry);
1962 if (record_type & PERF_RECORD_TIME) {
1964 * Maybe do better on x86 and provide cpu_clock_nmi()
1966 time = sched_clock();
1968 header.type |= PERF_RECORD_TIME;
1969 header.size += sizeof(u64);
1972 if (record_type & PERF_RECORD_ADDR) {
1973 header.type |= PERF_RECORD_ADDR;
1974 header.size += sizeof(u64);
1977 if (record_type & PERF_RECORD_GROUP) {
1978 header.type |= PERF_RECORD_GROUP;
1979 header.size += sizeof(u64) +
1980 counter->nr_siblings * sizeof(group_entry);
1983 if (record_type & PERF_RECORD_CALLCHAIN) {
1984 callchain = perf_callchain(regs);
1987 callchain_size = (1 + callchain->nr) * sizeof(u64);
1989 header.type |= PERF_RECORD_CALLCHAIN;
1990 header.size += callchain_size;
1994 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
1998 perf_output_put(&handle, header);
2000 if (record_type & PERF_RECORD_IP)
2001 perf_output_put(&handle, ip);
2003 if (record_type & PERF_RECORD_TID)
2004 perf_output_put(&handle, tid_entry);
2006 if (record_type & PERF_RECORD_TIME)
2007 perf_output_put(&handle, time);
2009 if (record_type & PERF_RECORD_ADDR)
2010 perf_output_put(&handle, addr);
2013 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2015 if (record_type & PERF_RECORD_GROUP) {
2016 struct perf_counter *leader, *sub;
2017 u64 nr = counter->nr_siblings;
2019 perf_output_put(&handle, nr);
2021 leader = counter->group_leader;
2022 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2024 sub->pmu->read(sub);
2026 group_entry.event = sub->hw_event.config;
2027 group_entry.counter = atomic64_read(&sub->count);
2029 perf_output_put(&handle, group_entry);
2034 perf_output_copy(&handle, callchain, callchain_size);
2036 perf_output_end(&handle);
2043 struct perf_comm_event {
2044 struct task_struct *task;
2049 struct perf_event_header header;
2056 static void perf_counter_comm_output(struct perf_counter *counter,
2057 struct perf_comm_event *comm_event)
2059 struct perf_output_handle handle;
2060 int size = comm_event->event.header.size;
2061 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2066 perf_output_put(&handle, comm_event->event);
2067 perf_output_copy(&handle, comm_event->comm,
2068 comm_event->comm_size);
2069 perf_output_end(&handle);
2072 static int perf_counter_comm_match(struct perf_counter *counter,
2073 struct perf_comm_event *comm_event)
2075 if (counter->hw_event.comm &&
2076 comm_event->event.header.type == PERF_EVENT_COMM)
2082 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2083 struct perf_comm_event *comm_event)
2085 struct perf_counter *counter;
2087 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2091 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2092 if (perf_counter_comm_match(counter, comm_event))
2093 perf_counter_comm_output(counter, comm_event);
2098 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2100 struct perf_cpu_context *cpuctx;
2102 char *comm = comm_event->task->comm;
2104 size = ALIGN(strlen(comm)+1, sizeof(u64));
2106 comm_event->comm = comm;
2107 comm_event->comm_size = size;
2109 comm_event->event.header.size = sizeof(comm_event->event) + size;
2111 cpuctx = &get_cpu_var(perf_cpu_context);
2112 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2113 put_cpu_var(perf_cpu_context);
2115 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
2118 void perf_counter_comm(struct task_struct *task)
2120 struct perf_comm_event comm_event;
2122 if (!atomic_read(&nr_comm_tracking))
2125 comm_event = (struct perf_comm_event){
2128 .header = { .type = PERF_EVENT_COMM, },
2129 .pid = task->group_leader->pid,
2134 perf_counter_comm_event(&comm_event);
2141 struct perf_mmap_event {
2147 struct perf_event_header header;
2157 static void perf_counter_mmap_output(struct perf_counter *counter,
2158 struct perf_mmap_event *mmap_event)
2160 struct perf_output_handle handle;
2161 int size = mmap_event->event.header.size;
2162 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2167 perf_output_put(&handle, mmap_event->event);
2168 perf_output_copy(&handle, mmap_event->file_name,
2169 mmap_event->file_size);
2170 perf_output_end(&handle);
2173 static int perf_counter_mmap_match(struct perf_counter *counter,
2174 struct perf_mmap_event *mmap_event)
2176 if (counter->hw_event.mmap &&
2177 mmap_event->event.header.type == PERF_EVENT_MMAP)
2180 if (counter->hw_event.munmap &&
2181 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2187 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2188 struct perf_mmap_event *mmap_event)
2190 struct perf_counter *counter;
2192 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2196 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2197 if (perf_counter_mmap_match(counter, mmap_event))
2198 perf_counter_mmap_output(counter, mmap_event);
2203 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2205 struct perf_cpu_context *cpuctx;
2206 struct file *file = mmap_event->file;
2213 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2215 name = strncpy(tmp, "//enomem", sizeof(tmp));
2218 name = d_path(&file->f_path, buf, PATH_MAX);
2220 name = strncpy(tmp, "//toolong", sizeof(tmp));
2224 name = strncpy(tmp, "//anon", sizeof(tmp));
2229 size = ALIGN(strlen(name)+1, sizeof(u64));
2231 mmap_event->file_name = name;
2232 mmap_event->file_size = size;
2234 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2236 cpuctx = &get_cpu_var(perf_cpu_context);
2237 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2238 put_cpu_var(perf_cpu_context);
2240 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2245 void perf_counter_mmap(unsigned long addr, unsigned long len,
2246 unsigned long pgoff, struct file *file)
2248 struct perf_mmap_event mmap_event;
2250 if (!atomic_read(&nr_mmap_tracking))
2253 mmap_event = (struct perf_mmap_event){
2256 .header = { .type = PERF_EVENT_MMAP, },
2257 .pid = current->group_leader->pid,
2258 .tid = current->pid,
2265 perf_counter_mmap_event(&mmap_event);
2268 void perf_counter_munmap(unsigned long addr, unsigned long len,
2269 unsigned long pgoff, struct file *file)
2271 struct perf_mmap_event mmap_event;
2273 if (!atomic_read(&nr_munmap_tracking))
2276 mmap_event = (struct perf_mmap_event){
2279 .header = { .type = PERF_EVENT_MUNMAP, },
2280 .pid = current->group_leader->pid,
2281 .tid = current->pid,
2288 perf_counter_mmap_event(&mmap_event);
2292 * Generic counter overflow handling.
2295 int perf_counter_overflow(struct perf_counter *counter,
2296 int nmi, struct pt_regs *regs, u64 addr)
2298 int events = atomic_read(&counter->event_limit);
2302 * XXX event_limit might not quite work as expected on inherited
2306 counter->pending_kill = POLL_IN;
2307 if (events && atomic_dec_and_test(&counter->event_limit)) {
2309 counter->pending_kill = POLL_HUP;
2311 counter->pending_disable = 1;
2312 perf_pending_queue(&counter->pending,
2313 perf_pending_counter);
2315 perf_counter_disable(counter);
2318 perf_counter_output(counter, nmi, regs, addr);
2323 * Generic software counter infrastructure
2326 static void perf_swcounter_update(struct perf_counter *counter)
2328 struct hw_perf_counter *hwc = &counter->hw;
2333 prev = atomic64_read(&hwc->prev_count);
2334 now = atomic64_read(&hwc->count);
2335 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2340 atomic64_add(delta, &counter->count);
2341 atomic64_sub(delta, &hwc->period_left);
2344 static void perf_swcounter_set_period(struct perf_counter *counter)
2346 struct hw_perf_counter *hwc = &counter->hw;
2347 s64 left = atomic64_read(&hwc->period_left);
2348 s64 period = hwc->irq_period;
2350 if (unlikely(left <= -period)) {
2352 atomic64_set(&hwc->period_left, left);
2355 if (unlikely(left <= 0)) {
2357 atomic64_add(period, &hwc->period_left);
2360 atomic64_set(&hwc->prev_count, -left);
2361 atomic64_set(&hwc->count, -left);
2364 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2366 enum hrtimer_restart ret = HRTIMER_RESTART;
2367 struct perf_counter *counter;
2368 struct pt_regs *regs;
2370 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2371 counter->pmu->read(counter);
2373 regs = get_irq_regs();
2375 * In case we exclude kernel IPs or are somehow not in interrupt
2376 * context, provide the next best thing, the user IP.
2378 if ((counter->hw_event.exclude_kernel || !regs) &&
2379 !counter->hw_event.exclude_user)
2380 regs = task_pt_regs(current);
2383 if (perf_counter_overflow(counter, 0, regs, 0))
2384 ret = HRTIMER_NORESTART;
2387 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2392 static void perf_swcounter_overflow(struct perf_counter *counter,
2393 int nmi, struct pt_regs *regs, u64 addr)
2395 perf_swcounter_update(counter);
2396 perf_swcounter_set_period(counter);
2397 if (perf_counter_overflow(counter, nmi, regs, addr))
2398 /* soft-disable the counter */
2403 static int perf_swcounter_match(struct perf_counter *counter,
2404 enum perf_event_types type,
2405 u32 event, struct pt_regs *regs)
2407 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2410 if (perf_event_raw(&counter->hw_event))
2413 if (perf_event_type(&counter->hw_event) != type)
2416 if (perf_event_id(&counter->hw_event) != event)
2419 if (counter->hw_event.exclude_user && user_mode(regs))
2422 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2428 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2429 int nmi, struct pt_regs *regs, u64 addr)
2431 int neg = atomic64_add_negative(nr, &counter->hw.count);
2432 if (counter->hw.irq_period && !neg)
2433 perf_swcounter_overflow(counter, nmi, regs, addr);
2436 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2437 enum perf_event_types type, u32 event,
2438 u64 nr, int nmi, struct pt_regs *regs,
2441 struct perf_counter *counter;
2443 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2447 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2448 if (perf_swcounter_match(counter, type, event, regs))
2449 perf_swcounter_add(counter, nr, nmi, regs, addr);
2454 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2457 return &cpuctx->recursion[3];
2460 return &cpuctx->recursion[2];
2463 return &cpuctx->recursion[1];
2465 return &cpuctx->recursion[0];
2468 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2469 u64 nr, int nmi, struct pt_regs *regs,
2472 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2473 int *recursion = perf_swcounter_recursion_context(cpuctx);
2481 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2482 nr, nmi, regs, addr);
2483 if (cpuctx->task_ctx) {
2484 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2485 nr, nmi, regs, addr);
2492 put_cpu_var(perf_cpu_context);
2496 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2498 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2501 static void perf_swcounter_read(struct perf_counter *counter)
2503 perf_swcounter_update(counter);
2506 static int perf_swcounter_enable(struct perf_counter *counter)
2508 perf_swcounter_set_period(counter);
2512 static void perf_swcounter_disable(struct perf_counter *counter)
2514 perf_swcounter_update(counter);
2517 static const struct pmu perf_ops_generic = {
2518 .enable = perf_swcounter_enable,
2519 .disable = perf_swcounter_disable,
2520 .read = perf_swcounter_read,
2524 * Software counter: cpu wall time clock
2527 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2529 int cpu = raw_smp_processor_id();
2533 now = cpu_clock(cpu);
2534 prev = atomic64_read(&counter->hw.prev_count);
2535 atomic64_set(&counter->hw.prev_count, now);
2536 atomic64_add(now - prev, &counter->count);
2539 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2541 struct hw_perf_counter *hwc = &counter->hw;
2542 int cpu = raw_smp_processor_id();
2544 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2545 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2546 hwc->hrtimer.function = perf_swcounter_hrtimer;
2547 if (hwc->irq_period) {
2548 __hrtimer_start_range_ns(&hwc->hrtimer,
2549 ns_to_ktime(hwc->irq_period), 0,
2550 HRTIMER_MODE_REL, 0);
2556 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2558 hrtimer_cancel(&counter->hw.hrtimer);
2559 cpu_clock_perf_counter_update(counter);
2562 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2564 cpu_clock_perf_counter_update(counter);
2567 static const struct pmu perf_ops_cpu_clock = {
2568 .enable = cpu_clock_perf_counter_enable,
2569 .disable = cpu_clock_perf_counter_disable,
2570 .read = cpu_clock_perf_counter_read,
2574 * Software counter: task time clock
2577 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2582 prev = atomic64_xchg(&counter->hw.prev_count, now);
2584 atomic64_add(delta, &counter->count);
2587 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2589 struct hw_perf_counter *hwc = &counter->hw;
2592 now = counter->ctx->time;
2594 atomic64_set(&hwc->prev_count, now);
2595 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2596 hwc->hrtimer.function = perf_swcounter_hrtimer;
2597 if (hwc->irq_period) {
2598 __hrtimer_start_range_ns(&hwc->hrtimer,
2599 ns_to_ktime(hwc->irq_period), 0,
2600 HRTIMER_MODE_REL, 0);
2606 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2608 hrtimer_cancel(&counter->hw.hrtimer);
2609 task_clock_perf_counter_update(counter, counter->ctx->time);
2613 static void task_clock_perf_counter_read(struct perf_counter *counter)
2618 update_context_time(counter->ctx);
2619 time = counter->ctx->time;
2621 u64 now = perf_clock();
2622 u64 delta = now - counter->ctx->timestamp;
2623 time = counter->ctx->time + delta;
2626 task_clock_perf_counter_update(counter, time);
2629 static const struct pmu perf_ops_task_clock = {
2630 .enable = task_clock_perf_counter_enable,
2631 .disable = task_clock_perf_counter_disable,
2632 .read = task_clock_perf_counter_read,
2636 * Software counter: cpu migrations
2639 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2641 struct task_struct *curr = counter->ctx->task;
2644 return curr->se.nr_migrations;
2645 return cpu_nr_migrations(smp_processor_id());
2648 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2653 prev = atomic64_read(&counter->hw.prev_count);
2654 now = get_cpu_migrations(counter);
2656 atomic64_set(&counter->hw.prev_count, now);
2660 atomic64_add(delta, &counter->count);
2663 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2665 cpu_migrations_perf_counter_update(counter);
2668 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2670 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2671 atomic64_set(&counter->hw.prev_count,
2672 get_cpu_migrations(counter));
2676 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2678 cpu_migrations_perf_counter_update(counter);
2681 static const struct pmu perf_ops_cpu_migrations = {
2682 .enable = cpu_migrations_perf_counter_enable,
2683 .disable = cpu_migrations_perf_counter_disable,
2684 .read = cpu_migrations_perf_counter_read,
2687 #ifdef CONFIG_EVENT_PROFILE
2688 void perf_tpcounter_event(int event_id)
2690 struct pt_regs *regs = get_irq_regs();
2693 regs = task_pt_regs(current);
2695 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2697 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2699 extern int ftrace_profile_enable(int);
2700 extern void ftrace_profile_disable(int);
2702 static void tp_perf_counter_destroy(struct perf_counter *counter)
2704 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2707 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2709 int event_id = perf_event_id(&counter->hw_event);
2712 ret = ftrace_profile_enable(event_id);
2716 counter->destroy = tp_perf_counter_destroy;
2717 counter->hw.irq_period = counter->hw_event.irq_period;
2719 return &perf_ops_generic;
2722 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2728 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2730 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2731 const struct pmu *pmu = NULL;
2732 struct hw_perf_counter *hwc = &counter->hw;
2735 * Software counters (currently) can't in general distinguish
2736 * between user, kernel and hypervisor events.
2737 * However, context switches and cpu migrations are considered
2738 * to be kernel events, and page faults are never hypervisor
2741 switch (perf_event_id(&counter->hw_event)) {
2742 case PERF_COUNT_CPU_CLOCK:
2743 pmu = &perf_ops_cpu_clock;
2745 if (hw_event->irq_period && hw_event->irq_period < 10000)
2746 hw_event->irq_period = 10000;
2748 case PERF_COUNT_TASK_CLOCK:
2750 * If the user instantiates this as a per-cpu counter,
2751 * use the cpu_clock counter instead.
2753 if (counter->ctx->task)
2754 pmu = &perf_ops_task_clock;
2756 pmu = &perf_ops_cpu_clock;
2758 if (hw_event->irq_period && hw_event->irq_period < 10000)
2759 hw_event->irq_period = 10000;
2761 case PERF_COUNT_PAGE_FAULTS:
2762 case PERF_COUNT_PAGE_FAULTS_MIN:
2763 case PERF_COUNT_PAGE_FAULTS_MAJ:
2764 case PERF_COUNT_CONTEXT_SWITCHES:
2765 pmu = &perf_ops_generic;
2767 case PERF_COUNT_CPU_MIGRATIONS:
2768 if (!counter->hw_event.exclude_kernel)
2769 pmu = &perf_ops_cpu_migrations;
2774 hwc->irq_period = hw_event->irq_period;
2780 * Allocate and initialize a counter structure
2782 static struct perf_counter *
2783 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2785 struct perf_counter_context *ctx,
2786 struct perf_counter *group_leader,
2789 const struct pmu *pmu;
2790 struct perf_counter *counter;
2793 counter = kzalloc(sizeof(*counter), gfpflags);
2795 return ERR_PTR(-ENOMEM);
2798 * Single counters are their own group leaders, with an
2799 * empty sibling list:
2802 group_leader = counter;
2804 mutex_init(&counter->mutex);
2805 INIT_LIST_HEAD(&counter->list_entry);
2806 INIT_LIST_HEAD(&counter->event_entry);
2807 INIT_LIST_HEAD(&counter->sibling_list);
2808 init_waitqueue_head(&counter->waitq);
2810 mutex_init(&counter->mmap_mutex);
2812 INIT_LIST_HEAD(&counter->child_list);
2815 counter->hw_event = *hw_event;
2816 counter->group_leader = group_leader;
2817 counter->pmu = NULL;
2820 counter->state = PERF_COUNTER_STATE_INACTIVE;
2821 if (hw_event->disabled)
2822 counter->state = PERF_COUNTER_STATE_OFF;
2827 * we currently do not support PERF_RECORD_GROUP on inherited counters
2829 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
2832 if (perf_event_raw(hw_event)) {
2833 pmu = hw_perf_counter_init(counter);
2837 switch (perf_event_type(hw_event)) {
2838 case PERF_TYPE_HARDWARE:
2839 pmu = hw_perf_counter_init(counter);
2842 case PERF_TYPE_SOFTWARE:
2843 pmu = sw_perf_counter_init(counter);
2846 case PERF_TYPE_TRACEPOINT:
2847 pmu = tp_perf_counter_init(counter);
2854 else if (IS_ERR(pmu))
2859 return ERR_PTR(err);
2864 if (counter->hw_event.mmap)
2865 atomic_inc(&nr_mmap_tracking);
2866 if (counter->hw_event.munmap)
2867 atomic_inc(&nr_munmap_tracking);
2868 if (counter->hw_event.comm)
2869 atomic_inc(&nr_comm_tracking);
2875 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2877 * @hw_event_uptr: event type attributes for monitoring/sampling
2880 * @group_fd: group leader counter fd
2882 SYSCALL_DEFINE5(perf_counter_open,
2883 const struct perf_counter_hw_event __user *, hw_event_uptr,
2884 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2886 struct perf_counter *counter, *group_leader;
2887 struct perf_counter_hw_event hw_event;
2888 struct perf_counter_context *ctx;
2889 struct file *counter_file = NULL;
2890 struct file *group_file = NULL;
2891 int fput_needed = 0;
2892 int fput_needed2 = 0;
2895 /* for future expandability... */
2899 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2903 * Get the target context (task or percpu):
2905 ctx = find_get_context(pid, cpu);
2907 return PTR_ERR(ctx);
2910 * Look up the group leader (we will attach this counter to it):
2912 group_leader = NULL;
2913 if (group_fd != -1) {
2915 group_file = fget_light(group_fd, &fput_needed);
2917 goto err_put_context;
2918 if (group_file->f_op != &perf_fops)
2919 goto err_put_context;
2921 group_leader = group_file->private_data;
2923 * Do not allow a recursive hierarchy (this new sibling
2924 * becoming part of another group-sibling):
2926 if (group_leader->group_leader != group_leader)
2927 goto err_put_context;
2929 * Do not allow to attach to a group in a different
2930 * task or CPU context:
2932 if (group_leader->ctx != ctx)
2933 goto err_put_context;
2935 * Only a group leader can be exclusive or pinned
2937 if (hw_event.exclusive || hw_event.pinned)
2938 goto err_put_context;
2941 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2943 ret = PTR_ERR(counter);
2944 if (IS_ERR(counter))
2945 goto err_put_context;
2947 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2949 goto err_free_put_context;
2951 counter_file = fget_light(ret, &fput_needed2);
2953 goto err_free_put_context;
2955 counter->filp = counter_file;
2956 mutex_lock(&ctx->mutex);
2957 perf_install_in_context(ctx, counter, cpu);
2958 mutex_unlock(&ctx->mutex);
2960 fput_light(counter_file, fput_needed2);
2963 fput_light(group_file, fput_needed);
2967 err_free_put_context:
2977 * Initialize the perf_counter context in a task_struct:
2980 __perf_counter_init_context(struct perf_counter_context *ctx,
2981 struct task_struct *task)
2983 memset(ctx, 0, sizeof(*ctx));
2984 spin_lock_init(&ctx->lock);
2985 mutex_init(&ctx->mutex);
2986 INIT_LIST_HEAD(&ctx->counter_list);
2987 INIT_LIST_HEAD(&ctx->event_list);
2992 * inherit a counter from parent task to child task:
2994 static struct perf_counter *
2995 inherit_counter(struct perf_counter *parent_counter,
2996 struct task_struct *parent,
2997 struct perf_counter_context *parent_ctx,
2998 struct task_struct *child,
2999 struct perf_counter *group_leader,
3000 struct perf_counter_context *child_ctx)
3002 struct perf_counter *child_counter;
3005 * Instead of creating recursive hierarchies of counters,
3006 * we link inherited counters back to the original parent,
3007 * which has a filp for sure, which we use as the reference
3010 if (parent_counter->parent)
3011 parent_counter = parent_counter->parent;
3013 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3014 parent_counter->cpu, child_ctx,
3015 group_leader, GFP_KERNEL);
3016 if (IS_ERR(child_counter))
3017 return child_counter;
3020 * Link it up in the child's context:
3022 child_counter->task = child;
3023 add_counter_to_ctx(child_counter, child_ctx);
3025 child_counter->parent = parent_counter;
3027 * inherit into child's child as well:
3029 child_counter->hw_event.inherit = 1;
3032 * Get a reference to the parent filp - we will fput it
3033 * when the child counter exits. This is safe to do because
3034 * we are in the parent and we know that the filp still
3035 * exists and has a nonzero count:
3037 atomic_long_inc(&parent_counter->filp->f_count);
3040 * Link this into the parent counter's child list
3042 mutex_lock(&parent_counter->mutex);
3043 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3046 * Make the child state follow the state of the parent counter,
3047 * not its hw_event.disabled bit. We hold the parent's mutex,
3048 * so we won't race with perf_counter_{en,dis}able_family.
3050 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3051 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3053 child_counter->state = PERF_COUNTER_STATE_OFF;
3055 mutex_unlock(&parent_counter->mutex);
3057 return child_counter;
3060 static int inherit_group(struct perf_counter *parent_counter,
3061 struct task_struct *parent,
3062 struct perf_counter_context *parent_ctx,
3063 struct task_struct *child,
3064 struct perf_counter_context *child_ctx)
3066 struct perf_counter *leader;
3067 struct perf_counter *sub;
3068 struct perf_counter *child_ctr;
3070 leader = inherit_counter(parent_counter, parent, parent_ctx,
3071 child, NULL, child_ctx);
3073 return PTR_ERR(leader);
3074 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3075 child_ctr = inherit_counter(sub, parent, parent_ctx,
3076 child, leader, child_ctx);
3077 if (IS_ERR(child_ctr))
3078 return PTR_ERR(child_ctr);
3083 static void sync_child_counter(struct perf_counter *child_counter,
3084 struct perf_counter *parent_counter)
3086 u64 parent_val, child_val;
3088 parent_val = atomic64_read(&parent_counter->count);
3089 child_val = atomic64_read(&child_counter->count);
3092 * Add back the child's count to the parent's count:
3094 atomic64_add(child_val, &parent_counter->count);
3095 atomic64_add(child_counter->total_time_enabled,
3096 &parent_counter->child_total_time_enabled);
3097 atomic64_add(child_counter->total_time_running,
3098 &parent_counter->child_total_time_running);
3101 * Remove this counter from the parent's list
3103 mutex_lock(&parent_counter->mutex);
3104 list_del_init(&child_counter->child_list);
3105 mutex_unlock(&parent_counter->mutex);
3108 * Release the parent counter, if this was the last
3111 fput(parent_counter->filp);
3115 __perf_counter_exit_task(struct task_struct *child,
3116 struct perf_counter *child_counter,
3117 struct perf_counter_context *child_ctx)
3119 struct perf_counter *parent_counter;
3120 struct perf_counter *sub, *tmp;
3123 * If we do not self-reap then we have to wait for the
3124 * child task to unschedule (it will happen for sure),
3125 * so that its counter is at its final count. (This
3126 * condition triggers rarely - child tasks usually get
3127 * off their CPU before the parent has a chance to
3128 * get this far into the reaping action)
3130 if (child != current) {
3131 wait_task_inactive(child, 0);
3132 list_del_init(&child_counter->list_entry);
3133 update_counter_times(child_counter);
3135 struct perf_cpu_context *cpuctx;
3136 unsigned long flags;
3140 * Disable and unlink this counter.
3142 * Be careful about zapping the list - IRQ/NMI context
3143 * could still be processing it:
3145 local_irq_save(flags);
3146 perf_flags = hw_perf_save_disable();
3148 cpuctx = &__get_cpu_var(perf_cpu_context);
3150 group_sched_out(child_counter, cpuctx, child_ctx);
3151 update_counter_times(child_counter);
3153 list_del_init(&child_counter->list_entry);
3155 child_ctx->nr_counters--;
3157 hw_perf_restore(perf_flags);
3158 local_irq_restore(flags);
3161 parent_counter = child_counter->parent;
3163 * It can happen that parent exits first, and has counters
3164 * that are still around due to the child reference. These
3165 * counters need to be zapped - but otherwise linger.
3167 if (parent_counter) {
3168 sync_child_counter(child_counter, parent_counter);
3169 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
3172 sync_child_counter(sub, sub->parent);
3176 free_counter(child_counter);
3181 * When a child task exits, feed back counter values to parent counters.
3183 * Note: we may be running in child context, but the PID is not hashed
3184 * anymore so new counters will not be added.
3186 void perf_counter_exit_task(struct task_struct *child)
3188 struct perf_counter *child_counter, *tmp;
3189 struct perf_counter_context *child_ctx;
3191 child_ctx = &child->perf_counter_ctx;
3193 if (likely(!child_ctx->nr_counters))
3196 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3198 __perf_counter_exit_task(child, child_counter, child_ctx);
3202 * Initialize the perf_counter context in task_struct
3204 void perf_counter_init_task(struct task_struct *child)
3206 struct perf_counter_context *child_ctx, *parent_ctx;
3207 struct perf_counter *counter;
3208 struct task_struct *parent = current;
3210 child_ctx = &child->perf_counter_ctx;
3211 parent_ctx = &parent->perf_counter_ctx;
3213 __perf_counter_init_context(child_ctx, child);
3216 * This is executed from the parent task context, so inherit
3217 * counters that have been marked for cloning:
3220 if (likely(!parent_ctx->nr_counters))
3224 * Lock the parent list. No need to lock the child - not PID
3225 * hashed yet and not running, so nobody can access it.
3227 mutex_lock(&parent_ctx->mutex);
3230 * We dont have to disable NMIs - we are only looking at
3231 * the list, not manipulating it:
3233 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
3234 if (!counter->hw_event.inherit)
3237 if (inherit_group(counter, parent,
3238 parent_ctx, child, child_ctx))
3242 mutex_unlock(&parent_ctx->mutex);
3245 static void __cpuinit perf_counter_init_cpu(int cpu)
3247 struct perf_cpu_context *cpuctx;
3249 cpuctx = &per_cpu(perf_cpu_context, cpu);
3250 __perf_counter_init_context(&cpuctx->ctx, NULL);
3252 spin_lock(&perf_resource_lock);
3253 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3254 spin_unlock(&perf_resource_lock);
3256 hw_perf_counter_setup(cpu);
3259 #ifdef CONFIG_HOTPLUG_CPU
3260 static void __perf_counter_exit_cpu(void *info)
3262 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3263 struct perf_counter_context *ctx = &cpuctx->ctx;
3264 struct perf_counter *counter, *tmp;
3266 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3267 __perf_counter_remove_from_context(counter);
3269 static void perf_counter_exit_cpu(int cpu)
3271 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3272 struct perf_counter_context *ctx = &cpuctx->ctx;
3274 mutex_lock(&ctx->mutex);
3275 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3276 mutex_unlock(&ctx->mutex);
3279 static inline void perf_counter_exit_cpu(int cpu) { }
3282 static int __cpuinit
3283 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3285 unsigned int cpu = (long)hcpu;
3289 case CPU_UP_PREPARE:
3290 case CPU_UP_PREPARE_FROZEN:
3291 perf_counter_init_cpu(cpu);
3294 case CPU_DOWN_PREPARE:
3295 case CPU_DOWN_PREPARE_FROZEN:
3296 perf_counter_exit_cpu(cpu);
3306 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3307 .notifier_call = perf_cpu_notify,
3310 void __init perf_counter_init(void)
3312 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3313 (void *)(long)smp_processor_id());
3314 register_cpu_notifier(&perf_cpu_nb);
3317 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3319 return sprintf(buf, "%d\n", perf_reserved_percpu);
3323 perf_set_reserve_percpu(struct sysdev_class *class,
3327 struct perf_cpu_context *cpuctx;
3331 err = strict_strtoul(buf, 10, &val);
3334 if (val > perf_max_counters)
3337 spin_lock(&perf_resource_lock);
3338 perf_reserved_percpu = val;
3339 for_each_online_cpu(cpu) {
3340 cpuctx = &per_cpu(perf_cpu_context, cpu);
3341 spin_lock_irq(&cpuctx->ctx.lock);
3342 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3343 perf_max_counters - perf_reserved_percpu);
3344 cpuctx->max_pertask = mpt;
3345 spin_unlock_irq(&cpuctx->ctx.lock);
3347 spin_unlock(&perf_resource_lock);
3352 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3354 return sprintf(buf, "%d\n", perf_overcommit);
3358 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3363 err = strict_strtoul(buf, 10, &val);
3369 spin_lock(&perf_resource_lock);
3370 perf_overcommit = val;
3371 spin_unlock(&perf_resource_lock);
3376 static SYSDEV_CLASS_ATTR(
3379 perf_show_reserve_percpu,
3380 perf_set_reserve_percpu
3383 static SYSDEV_CLASS_ATTR(
3386 perf_show_overcommit,
3390 static struct attribute *perfclass_attrs[] = {
3391 &attr_reserve_percpu.attr,
3392 &attr_overcommit.attr,
3396 static struct attribute_group perfclass_attr_group = {
3397 .attrs = perfclass_attrs,
3398 .name = "perf_counters",
3401 static int __init perf_counter_sysfs_init(void)
3403 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3404 &perfclass_attr_group);
3406 device_initcall(perf_counter_sysfs_init);