2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/ptrace.h>
20 #include <linux/percpu.h>
21 #include <linux/vmstat.h>
22 #include <linux/hardirq.h>
23 #include <linux/rculist.h>
24 #include <linux/uaccess.h>
25 #include <linux/syscalls.h>
26 #include <linux/anon_inodes.h>
27 #include <linux/kernel_stat.h>
28 #include <linux/perf_counter.h>
29 #include <linux/dcache.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_tracking __read_mostly;
44 static atomic_t nr_munmap_tracking __read_mostly;
45 static atomic_t nr_comm_tracking __read_mostly;
47 int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
48 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
51 * Lock for (sysadmin-configurable) counter reservations:
53 static DEFINE_SPINLOCK(perf_resource_lock);
56 * Architecture provided APIs - weak aliases:
58 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
63 void __weak hw_perf_disable(void) { barrier(); }
64 void __weak hw_perf_enable(void) { barrier(); }
66 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
67 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
68 struct perf_cpu_context *cpuctx,
69 struct perf_counter_context *ctx, int cpu)
74 void __weak perf_counter_print_debug(void) { }
76 static DEFINE_PER_CPU(int, disable_count);
78 void __perf_disable(void)
80 __get_cpu_var(disable_count)++;
83 bool __perf_enable(void)
85 return !--__get_cpu_var(disable_count);
88 void perf_disable(void)
94 void perf_enable(void)
100 static void get_ctx(struct perf_counter_context *ctx)
102 atomic_inc(&ctx->refcount);
105 static void put_ctx(struct perf_counter_context *ctx)
107 if (atomic_dec_and_test(&ctx->refcount))
112 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
114 struct perf_counter *group_leader = counter->group_leader;
117 * Depending on whether it is a standalone or sibling counter,
118 * add it straight to the context's counter list, or to the group
119 * leader's sibling list:
121 if (group_leader == counter)
122 list_add_tail(&counter->list_entry, &ctx->counter_list);
124 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
125 group_leader->nr_siblings++;
128 list_add_rcu(&counter->event_entry, &ctx->event_list);
133 * Remove a counter from the lists for its context.
134 * Must be called with counter->mutex and ctx->mutex held.
137 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
139 struct perf_counter *sibling, *tmp;
141 if (list_empty(&counter->list_entry))
145 list_del_init(&counter->list_entry);
146 list_del_rcu(&counter->event_entry);
148 if (counter->group_leader != counter)
149 counter->group_leader->nr_siblings--;
152 * If this was a group counter with sibling counters then
153 * upgrade the siblings to singleton counters by adding them
154 * to the context list directly:
156 list_for_each_entry_safe(sibling, tmp,
157 &counter->sibling_list, list_entry) {
159 list_move_tail(&sibling->list_entry, &ctx->counter_list);
160 sibling->group_leader = sibling;
165 counter_sched_out(struct perf_counter *counter,
166 struct perf_cpu_context *cpuctx,
167 struct perf_counter_context *ctx)
169 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
172 counter->state = PERF_COUNTER_STATE_INACTIVE;
173 counter->tstamp_stopped = ctx->time;
174 counter->pmu->disable(counter);
177 if (!is_software_counter(counter))
178 cpuctx->active_oncpu--;
180 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
181 cpuctx->exclusive = 0;
185 group_sched_out(struct perf_counter *group_counter,
186 struct perf_cpu_context *cpuctx,
187 struct perf_counter_context *ctx)
189 struct perf_counter *counter;
191 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
194 counter_sched_out(group_counter, cpuctx, ctx);
197 * Schedule out siblings (if any):
199 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
200 counter_sched_out(counter, cpuctx, ctx);
202 if (group_counter->hw_event.exclusive)
203 cpuctx->exclusive = 0;
207 * Cross CPU call to remove a performance counter
209 * We disable the counter on the hardware level first. After that we
210 * remove it from the context list.
212 static void __perf_counter_remove_from_context(void *info)
214 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
215 struct perf_counter *counter = info;
216 struct perf_counter_context *ctx = counter->ctx;
220 * If this is a task context, we need to check whether it is
221 * the current task context of this cpu. If not it has been
222 * scheduled out before the smp call arrived.
224 if (ctx->task && cpuctx->task_ctx != ctx)
227 spin_lock_irqsave(&ctx->lock, flags);
229 * Protect the list operation against NMI by disabling the
230 * counters on a global level.
234 counter_sched_out(counter, cpuctx, ctx);
236 list_del_counter(counter, ctx);
240 * Allow more per task counters with respect to the
243 cpuctx->max_pertask =
244 min(perf_max_counters - ctx->nr_counters,
245 perf_max_counters - perf_reserved_percpu);
249 spin_unlock_irqrestore(&ctx->lock, flags);
254 * Remove the counter from a task's (or a CPU's) list of counters.
256 * Must be called with counter->mutex and ctx->mutex held.
258 * CPU counters are removed with a smp call. For task counters we only
259 * call when the task is on a CPU.
261 static void perf_counter_remove_from_context(struct perf_counter *counter)
263 struct perf_counter_context *ctx = counter->ctx;
264 struct task_struct *task = ctx->task;
268 * Per cpu counters are removed via an smp call and
269 * the removal is always sucessful.
271 smp_call_function_single(counter->cpu,
272 __perf_counter_remove_from_context,
278 task_oncpu_function_call(task, __perf_counter_remove_from_context,
281 spin_lock_irq(&ctx->lock);
283 * If the context is active we need to retry the smp call.
285 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
286 spin_unlock_irq(&ctx->lock);
291 * The lock prevents that this context is scheduled in so we
292 * can remove the counter safely, if the call above did not
295 if (!list_empty(&counter->list_entry)) {
296 list_del_counter(counter, ctx);
298 spin_unlock_irq(&ctx->lock);
301 static inline u64 perf_clock(void)
303 return cpu_clock(smp_processor_id());
307 * Update the record of the current time in a context.
309 static void update_context_time(struct perf_counter_context *ctx)
311 u64 now = perf_clock();
313 ctx->time += now - ctx->timestamp;
314 ctx->timestamp = now;
318 * Update the total_time_enabled and total_time_running fields for a counter.
320 static void update_counter_times(struct perf_counter *counter)
322 struct perf_counter_context *ctx = counter->ctx;
325 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
328 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
330 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
331 run_end = counter->tstamp_stopped;
335 counter->total_time_running = run_end - counter->tstamp_running;
339 * Update total_time_enabled and total_time_running for all counters in a group.
341 static void update_group_times(struct perf_counter *leader)
343 struct perf_counter *counter;
345 update_counter_times(leader);
346 list_for_each_entry(counter, &leader->sibling_list, list_entry)
347 update_counter_times(counter);
351 * Cross CPU call to disable a performance counter
353 static void __perf_counter_disable(void *info)
355 struct perf_counter *counter = info;
356 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
357 struct perf_counter_context *ctx = counter->ctx;
361 * If this is a per-task counter, need to check whether this
362 * counter's task is the current task on this cpu.
364 if (ctx->task && cpuctx->task_ctx != ctx)
367 spin_lock_irqsave(&ctx->lock, flags);
370 * If the counter is on, turn it off.
371 * If it is in error state, leave it in error state.
373 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
374 update_context_time(ctx);
375 update_counter_times(counter);
376 if (counter == counter->group_leader)
377 group_sched_out(counter, cpuctx, ctx);
379 counter_sched_out(counter, cpuctx, ctx);
380 counter->state = PERF_COUNTER_STATE_OFF;
383 spin_unlock_irqrestore(&ctx->lock, flags);
389 static void perf_counter_disable(struct perf_counter *counter)
391 struct perf_counter_context *ctx = counter->ctx;
392 struct task_struct *task = ctx->task;
396 * Disable the counter on the cpu that it's on
398 smp_call_function_single(counter->cpu, __perf_counter_disable,
404 task_oncpu_function_call(task, __perf_counter_disable, counter);
406 spin_lock_irq(&ctx->lock);
408 * If the counter is still active, we need to retry the cross-call.
410 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
411 spin_unlock_irq(&ctx->lock);
416 * Since we have the lock this context can't be scheduled
417 * in, so we can change the state safely.
419 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
420 update_counter_times(counter);
421 counter->state = PERF_COUNTER_STATE_OFF;
424 spin_unlock_irq(&ctx->lock);
428 counter_sched_in(struct perf_counter *counter,
429 struct perf_cpu_context *cpuctx,
430 struct perf_counter_context *ctx,
433 if (counter->state <= PERF_COUNTER_STATE_OFF)
436 counter->state = PERF_COUNTER_STATE_ACTIVE;
437 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
439 * The new state must be visible before we turn it on in the hardware:
443 if (counter->pmu->enable(counter)) {
444 counter->state = PERF_COUNTER_STATE_INACTIVE;
449 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
451 if (!is_software_counter(counter))
452 cpuctx->active_oncpu++;
455 if (counter->hw_event.exclusive)
456 cpuctx->exclusive = 1;
462 group_sched_in(struct perf_counter *group_counter,
463 struct perf_cpu_context *cpuctx,
464 struct perf_counter_context *ctx,
467 struct perf_counter *counter, *partial_group;
470 if (group_counter->state == PERF_COUNTER_STATE_OFF)
473 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
475 return ret < 0 ? ret : 0;
477 group_counter->prev_state = group_counter->state;
478 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
482 * Schedule in siblings as one group (if any):
484 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
485 counter->prev_state = counter->state;
486 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
487 partial_group = counter;
496 * Groups can be scheduled in as one unit only, so undo any
497 * partial group before returning:
499 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
500 if (counter == partial_group)
502 counter_sched_out(counter, cpuctx, ctx);
504 counter_sched_out(group_counter, cpuctx, ctx);
510 * Return 1 for a group consisting entirely of software counters,
511 * 0 if the group contains any hardware counters.
513 static int is_software_only_group(struct perf_counter *leader)
515 struct perf_counter *counter;
517 if (!is_software_counter(leader))
520 list_for_each_entry(counter, &leader->sibling_list, list_entry)
521 if (!is_software_counter(counter))
528 * Work out whether we can put this counter group on the CPU now.
530 static int group_can_go_on(struct perf_counter *counter,
531 struct perf_cpu_context *cpuctx,
535 * Groups consisting entirely of software counters can always go on.
537 if (is_software_only_group(counter))
540 * If an exclusive group is already on, no other hardware
541 * counters can go on.
543 if (cpuctx->exclusive)
546 * If this group is exclusive and there are already
547 * counters on the CPU, it can't go on.
549 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
552 * Otherwise, try to add it if all previous groups were able
558 static void add_counter_to_ctx(struct perf_counter *counter,
559 struct perf_counter_context *ctx)
561 list_add_counter(counter, ctx);
562 counter->prev_state = PERF_COUNTER_STATE_OFF;
563 counter->tstamp_enabled = ctx->time;
564 counter->tstamp_running = ctx->time;
565 counter->tstamp_stopped = ctx->time;
569 * Cross CPU call to install and enable a performance counter
571 static void __perf_install_in_context(void *info)
573 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
574 struct perf_counter *counter = info;
575 struct perf_counter_context *ctx = counter->ctx;
576 struct perf_counter *leader = counter->group_leader;
577 int cpu = smp_processor_id();
582 * If this is a task context, we need to check whether it is
583 * the current task context of this cpu. If not it has been
584 * scheduled out before the smp call arrived.
585 * Or possibly this is the right context but it isn't
586 * on this cpu because it had no counters.
588 if (ctx->task && cpuctx->task_ctx != ctx) {
589 if (cpuctx->task_ctx || ctx->task != current)
591 cpuctx->task_ctx = ctx;
594 spin_lock_irqsave(&ctx->lock, flags);
596 update_context_time(ctx);
599 * Protect the list operation against NMI by disabling the
600 * counters on a global level. NOP for non NMI based counters.
604 add_counter_to_ctx(counter, ctx);
607 * Don't put the counter on if it is disabled or if
608 * it is in a group and the group isn't on.
610 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
611 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
615 * An exclusive counter can't go on if there are already active
616 * hardware counters, and no hardware counter can go on if there
617 * is already an exclusive counter on.
619 if (!group_can_go_on(counter, cpuctx, 1))
622 err = counter_sched_in(counter, cpuctx, ctx, cpu);
626 * This counter couldn't go on. If it is in a group
627 * then we have to pull the whole group off.
628 * If the counter group is pinned then put it in error state.
630 if (leader != counter)
631 group_sched_out(leader, cpuctx, ctx);
632 if (leader->hw_event.pinned) {
633 update_group_times(leader);
634 leader->state = PERF_COUNTER_STATE_ERROR;
638 if (!err && !ctx->task && cpuctx->max_pertask)
639 cpuctx->max_pertask--;
644 spin_unlock_irqrestore(&ctx->lock, flags);
648 * Attach a performance counter to a context
650 * First we add the counter to the list with the hardware enable bit
651 * in counter->hw_config cleared.
653 * If the counter is attached to a task which is on a CPU we use a smp
654 * call to enable it in the task context. The task might have been
655 * scheduled away, but we check this in the smp call again.
657 * Must be called with ctx->mutex held.
660 perf_install_in_context(struct perf_counter_context *ctx,
661 struct perf_counter *counter,
664 struct task_struct *task = ctx->task;
668 * Per cpu counters are installed via an smp call and
669 * the install is always sucessful.
671 smp_call_function_single(cpu, __perf_install_in_context,
677 task_oncpu_function_call(task, __perf_install_in_context,
680 spin_lock_irq(&ctx->lock);
682 * we need to retry the smp call.
684 if (ctx->is_active && list_empty(&counter->list_entry)) {
685 spin_unlock_irq(&ctx->lock);
690 * The lock prevents that this context is scheduled in so we
691 * can add the counter safely, if it the call above did not
694 if (list_empty(&counter->list_entry))
695 add_counter_to_ctx(counter, ctx);
696 spin_unlock_irq(&ctx->lock);
700 * Cross CPU call to enable a performance counter
702 static void __perf_counter_enable(void *info)
704 struct perf_counter *counter = info;
705 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
706 struct perf_counter_context *ctx = counter->ctx;
707 struct perf_counter *leader = counter->group_leader;
712 * If this is a per-task counter, need to check whether this
713 * counter's task is the current task on this cpu.
715 if (ctx->task && cpuctx->task_ctx != ctx) {
716 if (cpuctx->task_ctx || ctx->task != current)
718 cpuctx->task_ctx = ctx;
721 spin_lock_irqsave(&ctx->lock, flags);
723 update_context_time(ctx);
725 counter->prev_state = counter->state;
726 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
728 counter->state = PERF_COUNTER_STATE_INACTIVE;
729 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
732 * If the counter is in a group and isn't the group leader,
733 * then don't put it on unless the group is on.
735 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
738 if (!group_can_go_on(counter, cpuctx, 1)) {
742 if (counter == leader)
743 err = group_sched_in(counter, cpuctx, ctx,
746 err = counter_sched_in(counter, cpuctx, ctx,
753 * If this counter can't go on and it's part of a
754 * group, then the whole group has to come off.
756 if (leader != counter)
757 group_sched_out(leader, cpuctx, ctx);
758 if (leader->hw_event.pinned) {
759 update_group_times(leader);
760 leader->state = PERF_COUNTER_STATE_ERROR;
765 spin_unlock_irqrestore(&ctx->lock, flags);
771 static void perf_counter_enable(struct perf_counter *counter)
773 struct perf_counter_context *ctx = counter->ctx;
774 struct task_struct *task = ctx->task;
778 * Enable the counter on the cpu that it's on
780 smp_call_function_single(counter->cpu, __perf_counter_enable,
785 spin_lock_irq(&ctx->lock);
786 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
790 * If the counter is in error state, clear that first.
791 * That way, if we see the counter in error state below, we
792 * know that it has gone back into error state, as distinct
793 * from the task having been scheduled away before the
794 * cross-call arrived.
796 if (counter->state == PERF_COUNTER_STATE_ERROR)
797 counter->state = PERF_COUNTER_STATE_OFF;
800 spin_unlock_irq(&ctx->lock);
801 task_oncpu_function_call(task, __perf_counter_enable, counter);
803 spin_lock_irq(&ctx->lock);
806 * If the context is active and the counter is still off,
807 * we need to retry the cross-call.
809 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
813 * Since we have the lock this context can't be scheduled
814 * in, so we can change the state safely.
816 if (counter->state == PERF_COUNTER_STATE_OFF) {
817 counter->state = PERF_COUNTER_STATE_INACTIVE;
818 counter->tstamp_enabled =
819 ctx->time - counter->total_time_enabled;
822 spin_unlock_irq(&ctx->lock);
825 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
828 * not supported on inherited counters
830 if (counter->hw_event.inherit)
833 atomic_add(refresh, &counter->event_limit);
834 perf_counter_enable(counter);
839 void __perf_counter_sched_out(struct perf_counter_context *ctx,
840 struct perf_cpu_context *cpuctx)
842 struct perf_counter *counter;
844 spin_lock(&ctx->lock);
846 if (likely(!ctx->nr_counters))
848 update_context_time(ctx);
851 if (ctx->nr_active) {
852 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
853 if (counter != counter->group_leader)
854 counter_sched_out(counter, cpuctx, ctx);
856 group_sched_out(counter, cpuctx, ctx);
861 spin_unlock(&ctx->lock);
865 * Called from scheduler to remove the counters of the current task,
866 * with interrupts disabled.
868 * We stop each counter and update the counter value in counter->count.
870 * This does not protect us against NMI, but disable()
871 * sets the disabled bit in the control field of counter _before_
872 * accessing the counter control register. If a NMI hits, then it will
873 * not restart the counter.
875 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
877 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
878 struct perf_counter_context *ctx = task->perf_counter_ctxp;
879 struct pt_regs *regs;
881 if (likely(!ctx || !cpuctx->task_ctx))
884 update_context_time(ctx);
886 regs = task_pt_regs(task);
887 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
888 __perf_counter_sched_out(ctx, cpuctx);
890 cpuctx->task_ctx = NULL;
893 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
895 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
897 if (!cpuctx->task_ctx)
899 __perf_counter_sched_out(ctx, cpuctx);
900 cpuctx->task_ctx = NULL;
903 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
905 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
909 __perf_counter_sched_in(struct perf_counter_context *ctx,
910 struct perf_cpu_context *cpuctx, int cpu)
912 struct perf_counter *counter;
915 spin_lock(&ctx->lock);
917 if (likely(!ctx->nr_counters))
920 ctx->timestamp = perf_clock();
925 * First go through the list and put on any pinned groups
926 * in order to give them the best chance of going on.
928 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
929 if (counter->state <= PERF_COUNTER_STATE_OFF ||
930 !counter->hw_event.pinned)
932 if (counter->cpu != -1 && counter->cpu != cpu)
935 if (counter != counter->group_leader)
936 counter_sched_in(counter, cpuctx, ctx, cpu);
938 if (group_can_go_on(counter, cpuctx, 1))
939 group_sched_in(counter, cpuctx, ctx, cpu);
943 * If this pinned group hasn't been scheduled,
944 * put it in error state.
946 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
947 update_group_times(counter);
948 counter->state = PERF_COUNTER_STATE_ERROR;
952 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
954 * Ignore counters in OFF or ERROR state, and
955 * ignore pinned counters since we did them already.
957 if (counter->state <= PERF_COUNTER_STATE_OFF ||
958 counter->hw_event.pinned)
962 * Listen to the 'cpu' scheduling filter constraint
965 if (counter->cpu != -1 && counter->cpu != cpu)
968 if (counter != counter->group_leader) {
969 if (counter_sched_in(counter, cpuctx, ctx, cpu))
972 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
973 if (group_sched_in(counter, cpuctx, ctx, cpu))
980 spin_unlock(&ctx->lock);
984 * Called from scheduler to add the counters of the current task
985 * with interrupts disabled.
987 * We restore the counter value and then enable it.
989 * This does not protect us against NMI, but enable()
990 * sets the enabled bit in the control field of counter _before_
991 * accessing the counter control register. If a NMI hits, then it will
992 * keep the counter running.
994 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
996 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
997 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1001 __perf_counter_sched_in(ctx, cpuctx, cpu);
1002 cpuctx->task_ctx = ctx;
1005 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1007 struct perf_counter_context *ctx = &cpuctx->ctx;
1009 __perf_counter_sched_in(ctx, cpuctx, cpu);
1012 int perf_counter_task_disable(void)
1014 struct task_struct *curr = current;
1015 struct perf_counter_context *ctx = curr->perf_counter_ctxp;
1016 struct perf_counter *counter;
1017 unsigned long flags;
1019 if (!ctx || !ctx->nr_counters)
1022 local_irq_save(flags);
1024 __perf_counter_task_sched_out(ctx);
1026 spin_lock(&ctx->lock);
1029 * Disable all the counters:
1033 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1034 if (counter->state != PERF_COUNTER_STATE_ERROR) {
1035 update_group_times(counter);
1036 counter->state = PERF_COUNTER_STATE_OFF;
1042 spin_unlock_irqrestore(&ctx->lock, flags);
1047 int perf_counter_task_enable(void)
1049 struct task_struct *curr = current;
1050 struct perf_counter_context *ctx = curr->perf_counter_ctxp;
1051 struct perf_counter *counter;
1052 unsigned long flags;
1055 if (!ctx || !ctx->nr_counters)
1058 local_irq_save(flags);
1059 cpu = smp_processor_id();
1061 __perf_counter_task_sched_out(ctx);
1063 spin_lock(&ctx->lock);
1066 * Disable all the counters:
1070 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1071 if (counter->state > PERF_COUNTER_STATE_OFF)
1073 counter->state = PERF_COUNTER_STATE_INACTIVE;
1074 counter->tstamp_enabled =
1075 ctx->time - counter->total_time_enabled;
1076 counter->hw_event.disabled = 0;
1080 spin_unlock(&ctx->lock);
1082 perf_counter_task_sched_in(curr, cpu);
1084 local_irq_restore(flags);
1089 static void perf_log_period(struct perf_counter *counter, u64 period);
1091 static void perf_adjust_freq(struct perf_counter_context *ctx)
1093 struct perf_counter *counter;
1098 spin_lock(&ctx->lock);
1099 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1100 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1103 if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
1106 events = HZ * counter->hw.interrupts * counter->hw.irq_period;
1107 period = div64_u64(events, counter->hw_event.irq_freq);
1109 delta = (s64)(1 + period - counter->hw.irq_period);
1112 irq_period = counter->hw.irq_period + delta;
1117 perf_log_period(counter, irq_period);
1119 counter->hw.irq_period = irq_period;
1120 counter->hw.interrupts = 0;
1122 spin_unlock(&ctx->lock);
1126 * Round-robin a context's counters:
1128 static void rotate_ctx(struct perf_counter_context *ctx)
1130 struct perf_counter *counter;
1132 if (!ctx->nr_counters)
1135 spin_lock(&ctx->lock);
1137 * Rotate the first entry last (works just fine for group counters too):
1140 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1141 list_move_tail(&counter->list_entry, &ctx->counter_list);
1146 spin_unlock(&ctx->lock);
1149 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1151 struct perf_cpu_context *cpuctx;
1152 struct perf_counter_context *ctx;
1154 if (!atomic_read(&nr_counters))
1157 cpuctx = &per_cpu(perf_cpu_context, cpu);
1158 ctx = curr->perf_counter_ctxp;
1160 perf_adjust_freq(&cpuctx->ctx);
1162 perf_adjust_freq(ctx);
1164 perf_counter_cpu_sched_out(cpuctx);
1166 __perf_counter_task_sched_out(ctx);
1168 rotate_ctx(&cpuctx->ctx);
1172 perf_counter_cpu_sched_in(cpuctx, cpu);
1174 perf_counter_task_sched_in(curr, cpu);
1178 * Cross CPU call to read the hardware counter
1180 static void __read(void *info)
1182 struct perf_counter *counter = info;
1183 struct perf_counter_context *ctx = counter->ctx;
1184 unsigned long flags;
1186 local_irq_save(flags);
1188 update_context_time(ctx);
1189 counter->pmu->read(counter);
1190 update_counter_times(counter);
1191 local_irq_restore(flags);
1194 static u64 perf_counter_read(struct perf_counter *counter)
1197 * If counter is enabled and currently active on a CPU, update the
1198 * value in the counter structure:
1200 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1201 smp_call_function_single(counter->oncpu,
1202 __read, counter, 1);
1203 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1204 update_counter_times(counter);
1207 return atomic64_read(&counter->count);
1211 * Initialize the perf_counter context in a task_struct:
1214 __perf_counter_init_context(struct perf_counter_context *ctx,
1215 struct task_struct *task)
1217 memset(ctx, 0, sizeof(*ctx));
1218 spin_lock_init(&ctx->lock);
1219 mutex_init(&ctx->mutex);
1220 INIT_LIST_HEAD(&ctx->counter_list);
1221 INIT_LIST_HEAD(&ctx->event_list);
1222 atomic_set(&ctx->refcount, 1);
1226 static void put_context(struct perf_counter_context *ctx)
1229 put_task_struct(ctx->task);
1232 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1234 struct perf_cpu_context *cpuctx;
1235 struct perf_counter_context *ctx;
1236 struct perf_counter_context *tctx;
1237 struct task_struct *task;
1240 * If cpu is not a wildcard then this is a percpu counter:
1243 /* Must be root to operate on a CPU counter: */
1244 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1245 return ERR_PTR(-EACCES);
1247 if (cpu < 0 || cpu > num_possible_cpus())
1248 return ERR_PTR(-EINVAL);
1251 * We could be clever and allow to attach a counter to an
1252 * offline CPU and activate it when the CPU comes up, but
1255 if (!cpu_isset(cpu, cpu_online_map))
1256 return ERR_PTR(-ENODEV);
1258 cpuctx = &per_cpu(perf_cpu_context, cpu);
1268 task = find_task_by_vpid(pid);
1270 get_task_struct(task);
1274 return ERR_PTR(-ESRCH);
1276 /* Reuse ptrace permission checks for now. */
1277 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1278 put_task_struct(task);
1279 return ERR_PTR(-EACCES);
1282 ctx = task->perf_counter_ctxp;
1284 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1286 put_task_struct(task);
1287 return ERR_PTR(-ENOMEM);
1289 __perf_counter_init_context(ctx, task);
1291 * Make sure other cpus see correct values for *ctx
1292 * once task->perf_counter_ctxp is visible to them.
1295 tctx = cmpxchg(&task->perf_counter_ctxp, NULL, ctx);
1298 * We raced with some other task; use
1299 * the context they set.
1309 static void free_counter_rcu(struct rcu_head *head)
1311 struct perf_counter *counter;
1313 counter = container_of(head, struct perf_counter, rcu_head);
1314 put_ctx(counter->ctx);
1318 static void perf_pending_sync(struct perf_counter *counter);
1320 static void free_counter(struct perf_counter *counter)
1322 perf_pending_sync(counter);
1324 atomic_dec(&nr_counters);
1325 if (counter->hw_event.mmap)
1326 atomic_dec(&nr_mmap_tracking);
1327 if (counter->hw_event.munmap)
1328 atomic_dec(&nr_munmap_tracking);
1329 if (counter->hw_event.comm)
1330 atomic_dec(&nr_comm_tracking);
1332 if (counter->destroy)
1333 counter->destroy(counter);
1335 call_rcu(&counter->rcu_head, free_counter_rcu);
1339 * Called when the last reference to the file is gone.
1341 static int perf_release(struct inode *inode, struct file *file)
1343 struct perf_counter *counter = file->private_data;
1344 struct perf_counter_context *ctx = counter->ctx;
1346 file->private_data = NULL;
1348 mutex_lock(&ctx->mutex);
1349 mutex_lock(&counter->mutex);
1351 perf_counter_remove_from_context(counter);
1353 mutex_unlock(&counter->mutex);
1354 mutex_unlock(&ctx->mutex);
1356 free_counter(counter);
1363 * Read the performance counter - simple non blocking version for now
1366 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1372 * Return end-of-file for a read on a counter that is in
1373 * error state (i.e. because it was pinned but it couldn't be
1374 * scheduled on to the CPU at some point).
1376 if (counter->state == PERF_COUNTER_STATE_ERROR)
1379 mutex_lock(&counter->mutex);
1380 values[0] = perf_counter_read(counter);
1382 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1383 values[n++] = counter->total_time_enabled +
1384 atomic64_read(&counter->child_total_time_enabled);
1385 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1386 values[n++] = counter->total_time_running +
1387 atomic64_read(&counter->child_total_time_running);
1388 mutex_unlock(&counter->mutex);
1390 if (count < n * sizeof(u64))
1392 count = n * sizeof(u64);
1394 if (copy_to_user(buf, values, count))
1401 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1403 struct perf_counter *counter = file->private_data;
1405 return perf_read_hw(counter, buf, count);
1408 static unsigned int perf_poll(struct file *file, poll_table *wait)
1410 struct perf_counter *counter = file->private_data;
1411 struct perf_mmap_data *data;
1412 unsigned int events = POLL_HUP;
1415 data = rcu_dereference(counter->data);
1417 events = atomic_xchg(&data->poll, 0);
1420 poll_wait(file, &counter->waitq, wait);
1425 static void perf_counter_reset(struct perf_counter *counter)
1427 (void)perf_counter_read(counter);
1428 atomic64_set(&counter->count, 0);
1429 perf_counter_update_userpage(counter);
1432 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1433 void (*func)(struct perf_counter *))
1435 struct perf_counter_context *ctx = counter->ctx;
1436 struct perf_counter *sibling;
1438 spin_lock_irq(&ctx->lock);
1439 counter = counter->group_leader;
1442 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1444 spin_unlock_irq(&ctx->lock);
1447 static void perf_counter_for_each_child(struct perf_counter *counter,
1448 void (*func)(struct perf_counter *))
1450 struct perf_counter *child;
1452 mutex_lock(&counter->mutex);
1454 list_for_each_entry(child, &counter->child_list, child_list)
1456 mutex_unlock(&counter->mutex);
1459 static void perf_counter_for_each(struct perf_counter *counter,
1460 void (*func)(struct perf_counter *))
1462 struct perf_counter *child;
1464 mutex_lock(&counter->mutex);
1465 perf_counter_for_each_sibling(counter, func);
1466 list_for_each_entry(child, &counter->child_list, child_list)
1467 perf_counter_for_each_sibling(child, func);
1468 mutex_unlock(&counter->mutex);
1471 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1473 struct perf_counter *counter = file->private_data;
1474 void (*func)(struct perf_counter *);
1478 case PERF_COUNTER_IOC_ENABLE:
1479 func = perf_counter_enable;
1481 case PERF_COUNTER_IOC_DISABLE:
1482 func = perf_counter_disable;
1484 case PERF_COUNTER_IOC_RESET:
1485 func = perf_counter_reset;
1488 case PERF_COUNTER_IOC_REFRESH:
1489 return perf_counter_refresh(counter, arg);
1494 if (flags & PERF_IOC_FLAG_GROUP)
1495 perf_counter_for_each(counter, func);
1497 perf_counter_for_each_child(counter, func);
1503 * Callers need to ensure there can be no nesting of this function, otherwise
1504 * the seqlock logic goes bad. We can not serialize this because the arch
1505 * code calls this from NMI context.
1507 void perf_counter_update_userpage(struct perf_counter *counter)
1509 struct perf_mmap_data *data;
1510 struct perf_counter_mmap_page *userpg;
1513 data = rcu_dereference(counter->data);
1517 userpg = data->user_page;
1520 * Disable preemption so as to not let the corresponding user-space
1521 * spin too long if we get preempted.
1526 userpg->index = counter->hw.idx;
1527 userpg->offset = atomic64_read(&counter->count);
1528 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1529 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1538 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1540 struct perf_counter *counter = vma->vm_file->private_data;
1541 struct perf_mmap_data *data;
1542 int ret = VM_FAULT_SIGBUS;
1545 data = rcu_dereference(counter->data);
1549 if (vmf->pgoff == 0) {
1550 vmf->page = virt_to_page(data->user_page);
1552 int nr = vmf->pgoff - 1;
1554 if ((unsigned)nr > data->nr_pages)
1557 vmf->page = virt_to_page(data->data_pages[nr]);
1559 get_page(vmf->page);
1567 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1569 struct perf_mmap_data *data;
1573 WARN_ON(atomic_read(&counter->mmap_count));
1575 size = sizeof(struct perf_mmap_data);
1576 size += nr_pages * sizeof(void *);
1578 data = kzalloc(size, GFP_KERNEL);
1582 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1583 if (!data->user_page)
1584 goto fail_user_page;
1586 for (i = 0; i < nr_pages; i++) {
1587 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1588 if (!data->data_pages[i])
1589 goto fail_data_pages;
1592 data->nr_pages = nr_pages;
1593 atomic_set(&data->lock, -1);
1595 rcu_assign_pointer(counter->data, data);
1600 for (i--; i >= 0; i--)
1601 free_page((unsigned long)data->data_pages[i]);
1603 free_page((unsigned long)data->user_page);
1612 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1614 struct perf_mmap_data *data = container_of(rcu_head,
1615 struct perf_mmap_data, rcu_head);
1618 free_page((unsigned long)data->user_page);
1619 for (i = 0; i < data->nr_pages; i++)
1620 free_page((unsigned long)data->data_pages[i]);
1624 static void perf_mmap_data_free(struct perf_counter *counter)
1626 struct perf_mmap_data *data = counter->data;
1628 WARN_ON(atomic_read(&counter->mmap_count));
1630 rcu_assign_pointer(counter->data, NULL);
1631 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1634 static void perf_mmap_open(struct vm_area_struct *vma)
1636 struct perf_counter *counter = vma->vm_file->private_data;
1638 atomic_inc(&counter->mmap_count);
1641 static void perf_mmap_close(struct vm_area_struct *vma)
1643 struct perf_counter *counter = vma->vm_file->private_data;
1645 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1646 &counter->mmap_mutex)) {
1647 struct user_struct *user = current_user();
1649 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
1650 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1651 perf_mmap_data_free(counter);
1652 mutex_unlock(&counter->mmap_mutex);
1656 static struct vm_operations_struct perf_mmap_vmops = {
1657 .open = perf_mmap_open,
1658 .close = perf_mmap_close,
1659 .fault = perf_mmap_fault,
1662 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1664 struct perf_counter *counter = file->private_data;
1665 struct user_struct *user = current_user();
1666 unsigned long vma_size;
1667 unsigned long nr_pages;
1668 unsigned long user_locked, user_lock_limit;
1669 unsigned long locked, lock_limit;
1670 long user_extra, extra;
1673 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1676 vma_size = vma->vm_end - vma->vm_start;
1677 nr_pages = (vma_size / PAGE_SIZE) - 1;
1680 * If we have data pages ensure they're a power-of-two number, so we
1681 * can do bitmasks instead of modulo.
1683 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1686 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1689 if (vma->vm_pgoff != 0)
1692 mutex_lock(&counter->mmap_mutex);
1693 if (atomic_inc_not_zero(&counter->mmap_count)) {
1694 if (nr_pages != counter->data->nr_pages)
1699 user_extra = nr_pages + 1;
1700 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1701 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
1704 if (user_locked > user_lock_limit)
1705 extra = user_locked - user_lock_limit;
1707 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1708 lock_limit >>= PAGE_SHIFT;
1709 locked = vma->vm_mm->locked_vm + extra;
1711 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1716 WARN_ON(counter->data);
1717 ret = perf_mmap_data_alloc(counter, nr_pages);
1721 atomic_set(&counter->mmap_count, 1);
1722 atomic_long_add(user_extra, &user->locked_vm);
1723 vma->vm_mm->locked_vm += extra;
1724 counter->data->nr_locked = extra;
1726 mutex_unlock(&counter->mmap_mutex);
1728 vma->vm_flags &= ~VM_MAYWRITE;
1729 vma->vm_flags |= VM_RESERVED;
1730 vma->vm_ops = &perf_mmap_vmops;
1735 static int perf_fasync(int fd, struct file *filp, int on)
1737 struct perf_counter *counter = filp->private_data;
1738 struct inode *inode = filp->f_path.dentry->d_inode;
1741 mutex_lock(&inode->i_mutex);
1742 retval = fasync_helper(fd, filp, on, &counter->fasync);
1743 mutex_unlock(&inode->i_mutex);
1751 static const struct file_operations perf_fops = {
1752 .release = perf_release,
1755 .unlocked_ioctl = perf_ioctl,
1756 .compat_ioctl = perf_ioctl,
1758 .fasync = perf_fasync,
1762 * Perf counter wakeup
1764 * If there's data, ensure we set the poll() state and publish everything
1765 * to user-space before waking everybody up.
1768 void perf_counter_wakeup(struct perf_counter *counter)
1770 wake_up_all(&counter->waitq);
1772 if (counter->pending_kill) {
1773 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1774 counter->pending_kill = 0;
1781 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1783 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1784 * single linked list and use cmpxchg() to add entries lockless.
1787 static void perf_pending_counter(struct perf_pending_entry *entry)
1789 struct perf_counter *counter = container_of(entry,
1790 struct perf_counter, pending);
1792 if (counter->pending_disable) {
1793 counter->pending_disable = 0;
1794 perf_counter_disable(counter);
1797 if (counter->pending_wakeup) {
1798 counter->pending_wakeup = 0;
1799 perf_counter_wakeup(counter);
1803 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1805 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1809 static void perf_pending_queue(struct perf_pending_entry *entry,
1810 void (*func)(struct perf_pending_entry *))
1812 struct perf_pending_entry **head;
1814 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1819 head = &get_cpu_var(perf_pending_head);
1822 entry->next = *head;
1823 } while (cmpxchg(head, entry->next, entry) != entry->next);
1825 set_perf_counter_pending();
1827 put_cpu_var(perf_pending_head);
1830 static int __perf_pending_run(void)
1832 struct perf_pending_entry *list;
1835 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1836 while (list != PENDING_TAIL) {
1837 void (*func)(struct perf_pending_entry *);
1838 struct perf_pending_entry *entry = list;
1845 * Ensure we observe the unqueue before we issue the wakeup,
1846 * so that we won't be waiting forever.
1847 * -- see perf_not_pending().
1858 static inline int perf_not_pending(struct perf_counter *counter)
1861 * If we flush on whatever cpu we run, there is a chance we don't
1865 __perf_pending_run();
1869 * Ensure we see the proper queue state before going to sleep
1870 * so that we do not miss the wakeup. -- see perf_pending_handle()
1873 return counter->pending.next == NULL;
1876 static void perf_pending_sync(struct perf_counter *counter)
1878 wait_event(counter->waitq, perf_not_pending(counter));
1881 void perf_counter_do_pending(void)
1883 __perf_pending_run();
1887 * Callchain support -- arch specific
1890 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1899 struct perf_output_handle {
1900 struct perf_counter *counter;
1901 struct perf_mmap_data *data;
1902 unsigned int offset;
1907 unsigned long flags;
1910 static void perf_output_wakeup(struct perf_output_handle *handle)
1912 atomic_set(&handle->data->poll, POLL_IN);
1915 handle->counter->pending_wakeup = 1;
1916 perf_pending_queue(&handle->counter->pending,
1917 perf_pending_counter);
1919 perf_counter_wakeup(handle->counter);
1923 * Curious locking construct.
1925 * We need to ensure a later event doesn't publish a head when a former
1926 * event isn't done writing. However since we need to deal with NMIs we
1927 * cannot fully serialize things.
1929 * What we do is serialize between CPUs so we only have to deal with NMI
1930 * nesting on a single CPU.
1932 * We only publish the head (and generate a wakeup) when the outer-most
1935 static void perf_output_lock(struct perf_output_handle *handle)
1937 struct perf_mmap_data *data = handle->data;
1942 local_irq_save(handle->flags);
1943 cpu = smp_processor_id();
1945 if (in_nmi() && atomic_read(&data->lock) == cpu)
1948 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1954 static void perf_output_unlock(struct perf_output_handle *handle)
1956 struct perf_mmap_data *data = handle->data;
1959 data->done_head = data->head;
1961 if (!handle->locked)
1966 * The xchg implies a full barrier that ensures all writes are done
1967 * before we publish the new head, matched by a rmb() in userspace when
1968 * reading this position.
1970 while ((head = atomic_xchg(&data->done_head, 0)))
1971 data->user_page->data_head = head;
1974 * NMI can happen here, which means we can miss a done_head update.
1977 cpu = atomic_xchg(&data->lock, -1);
1978 WARN_ON_ONCE(cpu != smp_processor_id());
1981 * Therefore we have to validate we did not indeed do so.
1983 if (unlikely(atomic_read(&data->done_head))) {
1985 * Since we had it locked, we can lock it again.
1987 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1993 if (atomic_xchg(&data->wakeup, 0))
1994 perf_output_wakeup(handle);
1996 local_irq_restore(handle->flags);
1999 static int perf_output_begin(struct perf_output_handle *handle,
2000 struct perf_counter *counter, unsigned int size,
2001 int nmi, int overflow)
2003 struct perf_mmap_data *data;
2004 unsigned int offset, head;
2007 * For inherited counters we send all the output towards the parent.
2009 if (counter->parent)
2010 counter = counter->parent;
2013 data = rcu_dereference(counter->data);
2017 handle->data = data;
2018 handle->counter = counter;
2020 handle->overflow = overflow;
2022 if (!data->nr_pages)
2025 perf_output_lock(handle);
2028 offset = head = atomic_read(&data->head);
2030 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
2032 handle->offset = offset;
2033 handle->head = head;
2035 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2036 atomic_set(&data->wakeup, 1);
2041 perf_output_wakeup(handle);
2048 static void perf_output_copy(struct perf_output_handle *handle,
2049 void *buf, unsigned int len)
2051 unsigned int pages_mask;
2052 unsigned int offset;
2056 offset = handle->offset;
2057 pages_mask = handle->data->nr_pages - 1;
2058 pages = handle->data->data_pages;
2061 unsigned int page_offset;
2064 nr = (offset >> PAGE_SHIFT) & pages_mask;
2065 page_offset = offset & (PAGE_SIZE - 1);
2066 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2068 memcpy(pages[nr] + page_offset, buf, size);
2075 handle->offset = offset;
2078 * Check we didn't copy past our reservation window, taking the
2079 * possible unsigned int wrap into account.
2081 WARN_ON_ONCE(((int)(handle->head - handle->offset)) < 0);
2084 #define perf_output_put(handle, x) \
2085 perf_output_copy((handle), &(x), sizeof(x))
2087 static void perf_output_end(struct perf_output_handle *handle)
2089 struct perf_counter *counter = handle->counter;
2090 struct perf_mmap_data *data = handle->data;
2092 int wakeup_events = counter->hw_event.wakeup_events;
2094 if (handle->overflow && wakeup_events) {
2095 int events = atomic_inc_return(&data->events);
2096 if (events >= wakeup_events) {
2097 atomic_sub(wakeup_events, &data->events);
2098 atomic_set(&data->wakeup, 1);
2102 perf_output_unlock(handle);
2106 static void perf_counter_output(struct perf_counter *counter,
2107 int nmi, struct pt_regs *regs, u64 addr)
2110 u64 record_type = counter->hw_event.record_type;
2111 struct perf_output_handle handle;
2112 struct perf_event_header header;
2121 struct perf_callchain_entry *callchain = NULL;
2122 int callchain_size = 0;
2129 header.size = sizeof(header);
2131 header.misc = PERF_EVENT_MISC_OVERFLOW;
2132 header.misc |= perf_misc_flags(regs);
2134 if (record_type & PERF_RECORD_IP) {
2135 ip = perf_instruction_pointer(regs);
2136 header.type |= PERF_RECORD_IP;
2137 header.size += sizeof(ip);
2140 if (record_type & PERF_RECORD_TID) {
2141 /* namespace issues */
2142 tid_entry.pid = current->group_leader->pid;
2143 tid_entry.tid = current->pid;
2145 header.type |= PERF_RECORD_TID;
2146 header.size += sizeof(tid_entry);
2149 if (record_type & PERF_RECORD_TIME) {
2151 * Maybe do better on x86 and provide cpu_clock_nmi()
2153 time = sched_clock();
2155 header.type |= PERF_RECORD_TIME;
2156 header.size += sizeof(u64);
2159 if (record_type & PERF_RECORD_ADDR) {
2160 header.type |= PERF_RECORD_ADDR;
2161 header.size += sizeof(u64);
2164 if (record_type & PERF_RECORD_CONFIG) {
2165 header.type |= PERF_RECORD_CONFIG;
2166 header.size += sizeof(u64);
2169 if (record_type & PERF_RECORD_CPU) {
2170 header.type |= PERF_RECORD_CPU;
2171 header.size += sizeof(cpu_entry);
2173 cpu_entry.cpu = raw_smp_processor_id();
2176 if (record_type & PERF_RECORD_GROUP) {
2177 header.type |= PERF_RECORD_GROUP;
2178 header.size += sizeof(u64) +
2179 counter->nr_siblings * sizeof(group_entry);
2182 if (record_type & PERF_RECORD_CALLCHAIN) {
2183 callchain = perf_callchain(regs);
2186 callchain_size = (1 + callchain->nr) * sizeof(u64);
2188 header.type |= PERF_RECORD_CALLCHAIN;
2189 header.size += callchain_size;
2193 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2197 perf_output_put(&handle, header);
2199 if (record_type & PERF_RECORD_IP)
2200 perf_output_put(&handle, ip);
2202 if (record_type & PERF_RECORD_TID)
2203 perf_output_put(&handle, tid_entry);
2205 if (record_type & PERF_RECORD_TIME)
2206 perf_output_put(&handle, time);
2208 if (record_type & PERF_RECORD_ADDR)
2209 perf_output_put(&handle, addr);
2211 if (record_type & PERF_RECORD_CONFIG)
2212 perf_output_put(&handle, counter->hw_event.config);
2214 if (record_type & PERF_RECORD_CPU)
2215 perf_output_put(&handle, cpu_entry);
2218 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2220 if (record_type & PERF_RECORD_GROUP) {
2221 struct perf_counter *leader, *sub;
2222 u64 nr = counter->nr_siblings;
2224 perf_output_put(&handle, nr);
2226 leader = counter->group_leader;
2227 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2229 sub->pmu->read(sub);
2231 group_entry.event = sub->hw_event.config;
2232 group_entry.counter = atomic64_read(&sub->count);
2234 perf_output_put(&handle, group_entry);
2239 perf_output_copy(&handle, callchain, callchain_size);
2241 perf_output_end(&handle);
2248 struct perf_comm_event {
2249 struct task_struct *task;
2254 struct perf_event_header header;
2261 static void perf_counter_comm_output(struct perf_counter *counter,
2262 struct perf_comm_event *comm_event)
2264 struct perf_output_handle handle;
2265 int size = comm_event->event.header.size;
2266 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2271 perf_output_put(&handle, comm_event->event);
2272 perf_output_copy(&handle, comm_event->comm,
2273 comm_event->comm_size);
2274 perf_output_end(&handle);
2277 static int perf_counter_comm_match(struct perf_counter *counter,
2278 struct perf_comm_event *comm_event)
2280 if (counter->hw_event.comm &&
2281 comm_event->event.header.type == PERF_EVENT_COMM)
2287 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2288 struct perf_comm_event *comm_event)
2290 struct perf_counter *counter;
2292 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2296 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2297 if (perf_counter_comm_match(counter, comm_event))
2298 perf_counter_comm_output(counter, comm_event);
2303 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2305 struct perf_cpu_context *cpuctx;
2307 char *comm = comm_event->task->comm;
2309 size = ALIGN(strlen(comm)+1, sizeof(u64));
2311 comm_event->comm = comm;
2312 comm_event->comm_size = size;
2314 comm_event->event.header.size = sizeof(comm_event->event) + size;
2316 cpuctx = &get_cpu_var(perf_cpu_context);
2317 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2318 put_cpu_var(perf_cpu_context);
2320 perf_counter_comm_ctx(current->perf_counter_ctxp, comm_event);
2323 void perf_counter_comm(struct task_struct *task)
2325 struct perf_comm_event comm_event;
2327 if (!atomic_read(&nr_comm_tracking))
2329 if (!current->perf_counter_ctxp)
2332 comm_event = (struct perf_comm_event){
2335 .header = { .type = PERF_EVENT_COMM, },
2336 .pid = task->group_leader->pid,
2341 perf_counter_comm_event(&comm_event);
2348 struct perf_mmap_event {
2354 struct perf_event_header header;
2364 static void perf_counter_mmap_output(struct perf_counter *counter,
2365 struct perf_mmap_event *mmap_event)
2367 struct perf_output_handle handle;
2368 int size = mmap_event->event.header.size;
2369 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2374 perf_output_put(&handle, mmap_event->event);
2375 perf_output_copy(&handle, mmap_event->file_name,
2376 mmap_event->file_size);
2377 perf_output_end(&handle);
2380 static int perf_counter_mmap_match(struct perf_counter *counter,
2381 struct perf_mmap_event *mmap_event)
2383 if (counter->hw_event.mmap &&
2384 mmap_event->event.header.type == PERF_EVENT_MMAP)
2387 if (counter->hw_event.munmap &&
2388 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2394 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2395 struct perf_mmap_event *mmap_event)
2397 struct perf_counter *counter;
2399 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2403 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2404 if (perf_counter_mmap_match(counter, mmap_event))
2405 perf_counter_mmap_output(counter, mmap_event);
2410 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2412 struct perf_cpu_context *cpuctx;
2413 struct file *file = mmap_event->file;
2420 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2422 name = strncpy(tmp, "//enomem", sizeof(tmp));
2425 name = d_path(&file->f_path, buf, PATH_MAX);
2427 name = strncpy(tmp, "//toolong", sizeof(tmp));
2431 name = strncpy(tmp, "//anon", sizeof(tmp));
2436 size = ALIGN(strlen(name)+1, sizeof(u64));
2438 mmap_event->file_name = name;
2439 mmap_event->file_size = size;
2441 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2443 cpuctx = &get_cpu_var(perf_cpu_context);
2444 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2445 put_cpu_var(perf_cpu_context);
2447 perf_counter_mmap_ctx(current->perf_counter_ctxp, mmap_event);
2452 void perf_counter_mmap(unsigned long addr, unsigned long len,
2453 unsigned long pgoff, struct file *file)
2455 struct perf_mmap_event mmap_event;
2457 if (!atomic_read(&nr_mmap_tracking))
2459 if (!current->perf_counter_ctxp)
2462 mmap_event = (struct perf_mmap_event){
2465 .header = { .type = PERF_EVENT_MMAP, },
2466 .pid = current->group_leader->pid,
2467 .tid = current->pid,
2474 perf_counter_mmap_event(&mmap_event);
2477 void perf_counter_munmap(unsigned long addr, unsigned long len,
2478 unsigned long pgoff, struct file *file)
2480 struct perf_mmap_event mmap_event;
2482 if (!atomic_read(&nr_munmap_tracking))
2485 mmap_event = (struct perf_mmap_event){
2488 .header = { .type = PERF_EVENT_MUNMAP, },
2489 .pid = current->group_leader->pid,
2490 .tid = current->pid,
2497 perf_counter_mmap_event(&mmap_event);
2504 static void perf_log_period(struct perf_counter *counter, u64 period)
2506 struct perf_output_handle handle;
2510 struct perf_event_header header;
2515 .type = PERF_EVENT_PERIOD,
2517 .size = sizeof(freq_event),
2519 .time = sched_clock(),
2523 if (counter->hw.irq_period == period)
2526 ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
2530 perf_output_put(&handle, freq_event);
2531 perf_output_end(&handle);
2535 * Generic counter overflow handling.
2538 int perf_counter_overflow(struct perf_counter *counter,
2539 int nmi, struct pt_regs *regs, u64 addr)
2541 int events = atomic_read(&counter->event_limit);
2544 counter->hw.interrupts++;
2547 * XXX event_limit might not quite work as expected on inherited
2551 counter->pending_kill = POLL_IN;
2552 if (events && atomic_dec_and_test(&counter->event_limit)) {
2554 counter->pending_kill = POLL_HUP;
2556 counter->pending_disable = 1;
2557 perf_pending_queue(&counter->pending,
2558 perf_pending_counter);
2560 perf_counter_disable(counter);
2563 perf_counter_output(counter, nmi, regs, addr);
2568 * Generic software counter infrastructure
2571 static void perf_swcounter_update(struct perf_counter *counter)
2573 struct hw_perf_counter *hwc = &counter->hw;
2578 prev = atomic64_read(&hwc->prev_count);
2579 now = atomic64_read(&hwc->count);
2580 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2585 atomic64_add(delta, &counter->count);
2586 atomic64_sub(delta, &hwc->period_left);
2589 static void perf_swcounter_set_period(struct perf_counter *counter)
2591 struct hw_perf_counter *hwc = &counter->hw;
2592 s64 left = atomic64_read(&hwc->period_left);
2593 s64 period = hwc->irq_period;
2595 if (unlikely(left <= -period)) {
2597 atomic64_set(&hwc->period_left, left);
2600 if (unlikely(left <= 0)) {
2602 atomic64_add(period, &hwc->period_left);
2605 atomic64_set(&hwc->prev_count, -left);
2606 atomic64_set(&hwc->count, -left);
2609 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2611 enum hrtimer_restart ret = HRTIMER_RESTART;
2612 struct perf_counter *counter;
2613 struct pt_regs *regs;
2616 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2617 counter->pmu->read(counter);
2619 regs = get_irq_regs();
2621 * In case we exclude kernel IPs or are somehow not in interrupt
2622 * context, provide the next best thing, the user IP.
2624 if ((counter->hw_event.exclude_kernel || !regs) &&
2625 !counter->hw_event.exclude_user)
2626 regs = task_pt_regs(current);
2629 if (perf_counter_overflow(counter, 0, regs, 0))
2630 ret = HRTIMER_NORESTART;
2633 period = max_t(u64, 10000, counter->hw.irq_period);
2634 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
2639 static void perf_swcounter_overflow(struct perf_counter *counter,
2640 int nmi, struct pt_regs *regs, u64 addr)
2642 perf_swcounter_update(counter);
2643 perf_swcounter_set_period(counter);
2644 if (perf_counter_overflow(counter, nmi, regs, addr))
2645 /* soft-disable the counter */
2650 static int perf_swcounter_match(struct perf_counter *counter,
2651 enum perf_event_types type,
2652 u32 event, struct pt_regs *regs)
2654 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2657 if (perf_event_raw(&counter->hw_event))
2660 if (perf_event_type(&counter->hw_event) != type)
2663 if (perf_event_id(&counter->hw_event) != event)
2666 if (counter->hw_event.exclude_user && user_mode(regs))
2669 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2675 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2676 int nmi, struct pt_regs *regs, u64 addr)
2678 int neg = atomic64_add_negative(nr, &counter->hw.count);
2679 if (counter->hw.irq_period && !neg)
2680 perf_swcounter_overflow(counter, nmi, regs, addr);
2683 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2684 enum perf_event_types type, u32 event,
2685 u64 nr, int nmi, struct pt_regs *regs,
2688 struct perf_counter *counter;
2690 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2694 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2695 if (perf_swcounter_match(counter, type, event, regs))
2696 perf_swcounter_add(counter, nr, nmi, regs, addr);
2701 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2704 return &cpuctx->recursion[3];
2707 return &cpuctx->recursion[2];
2710 return &cpuctx->recursion[1];
2712 return &cpuctx->recursion[0];
2715 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2716 u64 nr, int nmi, struct pt_regs *regs,
2719 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2720 int *recursion = perf_swcounter_recursion_context(cpuctx);
2728 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2729 nr, nmi, regs, addr);
2730 if (cpuctx->task_ctx) {
2731 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2732 nr, nmi, regs, addr);
2739 put_cpu_var(perf_cpu_context);
2743 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2745 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2748 static void perf_swcounter_read(struct perf_counter *counter)
2750 perf_swcounter_update(counter);
2753 static int perf_swcounter_enable(struct perf_counter *counter)
2755 perf_swcounter_set_period(counter);
2759 static void perf_swcounter_disable(struct perf_counter *counter)
2761 perf_swcounter_update(counter);
2764 static const struct pmu perf_ops_generic = {
2765 .enable = perf_swcounter_enable,
2766 .disable = perf_swcounter_disable,
2767 .read = perf_swcounter_read,
2771 * Software counter: cpu wall time clock
2774 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2776 int cpu = raw_smp_processor_id();
2780 now = cpu_clock(cpu);
2781 prev = atomic64_read(&counter->hw.prev_count);
2782 atomic64_set(&counter->hw.prev_count, now);
2783 atomic64_add(now - prev, &counter->count);
2786 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2788 struct hw_perf_counter *hwc = &counter->hw;
2789 int cpu = raw_smp_processor_id();
2791 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2792 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2793 hwc->hrtimer.function = perf_swcounter_hrtimer;
2794 if (hwc->irq_period) {
2795 u64 period = max_t(u64, 10000, hwc->irq_period);
2796 __hrtimer_start_range_ns(&hwc->hrtimer,
2797 ns_to_ktime(period), 0,
2798 HRTIMER_MODE_REL, 0);
2804 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2806 if (counter->hw.irq_period)
2807 hrtimer_cancel(&counter->hw.hrtimer);
2808 cpu_clock_perf_counter_update(counter);
2811 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2813 cpu_clock_perf_counter_update(counter);
2816 static const struct pmu perf_ops_cpu_clock = {
2817 .enable = cpu_clock_perf_counter_enable,
2818 .disable = cpu_clock_perf_counter_disable,
2819 .read = cpu_clock_perf_counter_read,
2823 * Software counter: task time clock
2826 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2831 prev = atomic64_xchg(&counter->hw.prev_count, now);
2833 atomic64_add(delta, &counter->count);
2836 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2838 struct hw_perf_counter *hwc = &counter->hw;
2841 now = counter->ctx->time;
2843 atomic64_set(&hwc->prev_count, now);
2844 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2845 hwc->hrtimer.function = perf_swcounter_hrtimer;
2846 if (hwc->irq_period) {
2847 u64 period = max_t(u64, 10000, hwc->irq_period);
2848 __hrtimer_start_range_ns(&hwc->hrtimer,
2849 ns_to_ktime(period), 0,
2850 HRTIMER_MODE_REL, 0);
2856 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2858 if (counter->hw.irq_period)
2859 hrtimer_cancel(&counter->hw.hrtimer);
2860 task_clock_perf_counter_update(counter, counter->ctx->time);
2864 static void task_clock_perf_counter_read(struct perf_counter *counter)
2869 update_context_time(counter->ctx);
2870 time = counter->ctx->time;
2872 u64 now = perf_clock();
2873 u64 delta = now - counter->ctx->timestamp;
2874 time = counter->ctx->time + delta;
2877 task_clock_perf_counter_update(counter, time);
2880 static const struct pmu perf_ops_task_clock = {
2881 .enable = task_clock_perf_counter_enable,
2882 .disable = task_clock_perf_counter_disable,
2883 .read = task_clock_perf_counter_read,
2887 * Software counter: cpu migrations
2890 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2892 struct task_struct *curr = counter->ctx->task;
2895 return curr->se.nr_migrations;
2896 return cpu_nr_migrations(smp_processor_id());
2899 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2904 prev = atomic64_read(&counter->hw.prev_count);
2905 now = get_cpu_migrations(counter);
2907 atomic64_set(&counter->hw.prev_count, now);
2911 atomic64_add(delta, &counter->count);
2914 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2916 cpu_migrations_perf_counter_update(counter);
2919 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2921 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2922 atomic64_set(&counter->hw.prev_count,
2923 get_cpu_migrations(counter));
2927 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2929 cpu_migrations_perf_counter_update(counter);
2932 static const struct pmu perf_ops_cpu_migrations = {
2933 .enable = cpu_migrations_perf_counter_enable,
2934 .disable = cpu_migrations_perf_counter_disable,
2935 .read = cpu_migrations_perf_counter_read,
2938 #ifdef CONFIG_EVENT_PROFILE
2939 void perf_tpcounter_event(int event_id)
2941 struct pt_regs *regs = get_irq_regs();
2944 regs = task_pt_regs(current);
2946 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2948 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2950 extern int ftrace_profile_enable(int);
2951 extern void ftrace_profile_disable(int);
2953 static void tp_perf_counter_destroy(struct perf_counter *counter)
2955 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2958 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2960 int event_id = perf_event_id(&counter->hw_event);
2963 ret = ftrace_profile_enable(event_id);
2967 counter->destroy = tp_perf_counter_destroy;
2968 counter->hw.irq_period = counter->hw_event.irq_period;
2970 return &perf_ops_generic;
2973 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2979 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2981 const struct pmu *pmu = NULL;
2984 * Software counters (currently) can't in general distinguish
2985 * between user, kernel and hypervisor events.
2986 * However, context switches and cpu migrations are considered
2987 * to be kernel events, and page faults are never hypervisor
2990 switch (perf_event_id(&counter->hw_event)) {
2991 case PERF_COUNT_CPU_CLOCK:
2992 pmu = &perf_ops_cpu_clock;
2995 case PERF_COUNT_TASK_CLOCK:
2997 * If the user instantiates this as a per-cpu counter,
2998 * use the cpu_clock counter instead.
3000 if (counter->ctx->task)
3001 pmu = &perf_ops_task_clock;
3003 pmu = &perf_ops_cpu_clock;
3006 case PERF_COUNT_PAGE_FAULTS:
3007 case PERF_COUNT_PAGE_FAULTS_MIN:
3008 case PERF_COUNT_PAGE_FAULTS_MAJ:
3009 case PERF_COUNT_CONTEXT_SWITCHES:
3010 pmu = &perf_ops_generic;
3012 case PERF_COUNT_CPU_MIGRATIONS:
3013 if (!counter->hw_event.exclude_kernel)
3014 pmu = &perf_ops_cpu_migrations;
3022 * Allocate and initialize a counter structure
3024 static struct perf_counter *
3025 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
3027 struct perf_counter_context *ctx,
3028 struct perf_counter *group_leader,
3031 const struct pmu *pmu;
3032 struct perf_counter *counter;
3033 struct hw_perf_counter *hwc;
3036 counter = kzalloc(sizeof(*counter), gfpflags);
3038 return ERR_PTR(-ENOMEM);
3041 * Single counters are their own group leaders, with an
3042 * empty sibling list:
3045 group_leader = counter;
3047 mutex_init(&counter->mutex);
3048 INIT_LIST_HEAD(&counter->list_entry);
3049 INIT_LIST_HEAD(&counter->event_entry);
3050 INIT_LIST_HEAD(&counter->sibling_list);
3051 init_waitqueue_head(&counter->waitq);
3053 mutex_init(&counter->mmap_mutex);
3055 INIT_LIST_HEAD(&counter->child_list);
3058 counter->hw_event = *hw_event;
3059 counter->group_leader = group_leader;
3060 counter->pmu = NULL;
3064 counter->state = PERF_COUNTER_STATE_INACTIVE;
3065 if (hw_event->disabled)
3066 counter->state = PERF_COUNTER_STATE_OFF;
3071 if (hw_event->freq && hw_event->irq_freq)
3072 hwc->irq_period = div64_u64(TICK_NSEC, hw_event->irq_freq);
3074 hwc->irq_period = hw_event->irq_period;
3077 * we currently do not support PERF_RECORD_GROUP on inherited counters
3079 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
3082 if (perf_event_raw(hw_event)) {
3083 pmu = hw_perf_counter_init(counter);
3087 switch (perf_event_type(hw_event)) {
3088 case PERF_TYPE_HARDWARE:
3089 pmu = hw_perf_counter_init(counter);
3092 case PERF_TYPE_SOFTWARE:
3093 pmu = sw_perf_counter_init(counter);
3096 case PERF_TYPE_TRACEPOINT:
3097 pmu = tp_perf_counter_init(counter);
3104 else if (IS_ERR(pmu))
3109 return ERR_PTR(err);
3114 atomic_inc(&nr_counters);
3115 if (counter->hw_event.mmap)
3116 atomic_inc(&nr_mmap_tracking);
3117 if (counter->hw_event.munmap)
3118 atomic_inc(&nr_munmap_tracking);
3119 if (counter->hw_event.comm)
3120 atomic_inc(&nr_comm_tracking);
3126 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3128 * @hw_event_uptr: event type attributes for monitoring/sampling
3131 * @group_fd: group leader counter fd
3133 SYSCALL_DEFINE5(perf_counter_open,
3134 const struct perf_counter_hw_event __user *, hw_event_uptr,
3135 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3137 struct perf_counter *counter, *group_leader;
3138 struct perf_counter_hw_event hw_event;
3139 struct perf_counter_context *ctx;
3140 struct file *counter_file = NULL;
3141 struct file *group_file = NULL;
3142 int fput_needed = 0;
3143 int fput_needed2 = 0;
3146 /* for future expandability... */
3150 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
3154 * Get the target context (task or percpu):
3156 ctx = find_get_context(pid, cpu);
3158 return PTR_ERR(ctx);
3161 * Look up the group leader (we will attach this counter to it):
3163 group_leader = NULL;
3164 if (group_fd != -1) {
3166 group_file = fget_light(group_fd, &fput_needed);
3168 goto err_put_context;
3169 if (group_file->f_op != &perf_fops)
3170 goto err_put_context;
3172 group_leader = group_file->private_data;
3174 * Do not allow a recursive hierarchy (this new sibling
3175 * becoming part of another group-sibling):
3177 if (group_leader->group_leader != group_leader)
3178 goto err_put_context;
3180 * Do not allow to attach to a group in a different
3181 * task or CPU context:
3183 if (group_leader->ctx != ctx)
3184 goto err_put_context;
3186 * Only a group leader can be exclusive or pinned
3188 if (hw_event.exclusive || hw_event.pinned)
3189 goto err_put_context;
3192 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
3194 ret = PTR_ERR(counter);
3195 if (IS_ERR(counter))
3196 goto err_put_context;
3198 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3200 goto err_free_put_context;
3202 counter_file = fget_light(ret, &fput_needed2);
3204 goto err_free_put_context;
3206 counter->filp = counter_file;
3207 mutex_lock(&ctx->mutex);
3208 perf_install_in_context(ctx, counter, cpu);
3209 mutex_unlock(&ctx->mutex);
3211 fput_light(counter_file, fput_needed2);
3214 fput_light(group_file, fput_needed);
3218 err_free_put_context:
3228 * inherit a counter from parent task to child task:
3230 static struct perf_counter *
3231 inherit_counter(struct perf_counter *parent_counter,
3232 struct task_struct *parent,
3233 struct perf_counter_context *parent_ctx,
3234 struct task_struct *child,
3235 struct perf_counter *group_leader,
3236 struct perf_counter_context *child_ctx)
3238 struct perf_counter *child_counter;
3241 * Instead of creating recursive hierarchies of counters,
3242 * we link inherited counters back to the original parent,
3243 * which has a filp for sure, which we use as the reference
3246 if (parent_counter->parent)
3247 parent_counter = parent_counter->parent;
3249 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3250 parent_counter->cpu, child_ctx,
3251 group_leader, GFP_KERNEL);
3252 if (IS_ERR(child_counter))
3253 return child_counter;
3256 * Link it up in the child's context:
3258 add_counter_to_ctx(child_counter, child_ctx);
3260 child_counter->parent = parent_counter;
3262 * inherit into child's child as well:
3264 child_counter->hw_event.inherit = 1;
3267 * Get a reference to the parent filp - we will fput it
3268 * when the child counter exits. This is safe to do because
3269 * we are in the parent and we know that the filp still
3270 * exists and has a nonzero count:
3272 atomic_long_inc(&parent_counter->filp->f_count);
3275 * Link this into the parent counter's child list
3277 mutex_lock(&parent_counter->mutex);
3278 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3281 * Make the child state follow the state of the parent counter,
3282 * not its hw_event.disabled bit. We hold the parent's mutex,
3283 * so we won't race with perf_counter_{en,dis}able_family.
3285 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3286 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3288 child_counter->state = PERF_COUNTER_STATE_OFF;
3290 mutex_unlock(&parent_counter->mutex);
3292 return child_counter;
3295 static int inherit_group(struct perf_counter *parent_counter,
3296 struct task_struct *parent,
3297 struct perf_counter_context *parent_ctx,
3298 struct task_struct *child,
3299 struct perf_counter_context *child_ctx)
3301 struct perf_counter *leader;
3302 struct perf_counter *sub;
3303 struct perf_counter *child_ctr;
3305 leader = inherit_counter(parent_counter, parent, parent_ctx,
3306 child, NULL, child_ctx);
3308 return PTR_ERR(leader);
3309 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3310 child_ctr = inherit_counter(sub, parent, parent_ctx,
3311 child, leader, child_ctx);
3312 if (IS_ERR(child_ctr))
3313 return PTR_ERR(child_ctr);
3318 static void sync_child_counter(struct perf_counter *child_counter,
3319 struct perf_counter *parent_counter)
3323 child_val = atomic64_read(&child_counter->count);
3326 * Add back the child's count to the parent's count:
3328 atomic64_add(child_val, &parent_counter->count);
3329 atomic64_add(child_counter->total_time_enabled,
3330 &parent_counter->child_total_time_enabled);
3331 atomic64_add(child_counter->total_time_running,
3332 &parent_counter->child_total_time_running);
3335 * Remove this counter from the parent's list
3337 mutex_lock(&parent_counter->mutex);
3338 list_del_init(&child_counter->child_list);
3339 mutex_unlock(&parent_counter->mutex);
3342 * Release the parent counter, if this was the last
3345 fput(parent_counter->filp);
3349 __perf_counter_exit_task(struct task_struct *child,
3350 struct perf_counter *child_counter,
3351 struct perf_counter_context *child_ctx)
3353 struct perf_counter *parent_counter;
3356 * Protect against concurrent operations on child_counter
3357 * due its fd getting closed, etc.
3359 mutex_lock(&child_counter->mutex);
3361 update_counter_times(child_counter);
3362 list_del_counter(child_counter, child_ctx);
3364 mutex_unlock(&child_counter->mutex);
3366 parent_counter = child_counter->parent;
3368 * It can happen that parent exits first, and has counters
3369 * that are still around due to the child reference. These
3370 * counters need to be zapped - but otherwise linger.
3372 if (parent_counter) {
3373 sync_child_counter(child_counter, parent_counter);
3374 free_counter(child_counter);
3379 * When a child task exits, feed back counter values to parent counters.
3381 * Note: we may be running in child context, but the PID is not hashed
3382 * anymore so new counters will not be added.
3383 * (XXX not sure that is true when we get called from flush_old_exec.
3386 void perf_counter_exit_task(struct task_struct *child)
3388 struct perf_counter *child_counter, *tmp;
3389 struct perf_counter_context *child_ctx;
3390 unsigned long flags;
3392 WARN_ON_ONCE(child != current);
3394 child_ctx = child->perf_counter_ctxp;
3396 if (likely(!child_ctx))
3399 local_irq_save(flags);
3400 __perf_counter_task_sched_out(child_ctx);
3401 child->perf_counter_ctxp = NULL;
3402 local_irq_restore(flags);
3404 mutex_lock(&child_ctx->mutex);
3407 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3409 __perf_counter_exit_task(child, child_counter, child_ctx);
3412 * If the last counter was a group counter, it will have appended all
3413 * its siblings to the list, but we obtained 'tmp' before that which
3414 * will still point to the list head terminating the iteration.
3416 if (!list_empty(&child_ctx->counter_list))
3419 mutex_unlock(&child_ctx->mutex);
3425 * Initialize the perf_counter context in task_struct
3427 void perf_counter_init_task(struct task_struct *child)
3429 struct perf_counter_context *child_ctx, *parent_ctx;
3430 struct perf_counter *counter;
3431 struct task_struct *parent = current;
3433 child->perf_counter_ctxp = NULL;
3436 * This is executed from the parent task context, so inherit
3437 * counters that have been marked for cloning.
3438 * First allocate and initialize a context for the child.
3441 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
3445 parent_ctx = parent->perf_counter_ctxp;
3446 if (likely(!parent_ctx || !parent_ctx->nr_counters))
3449 __perf_counter_init_context(child_ctx, child);
3450 child->perf_counter_ctxp = child_ctx;
3453 * Lock the parent list. No need to lock the child - not PID
3454 * hashed yet and not running, so nobody can access it.
3456 mutex_lock(&parent_ctx->mutex);
3459 * We dont have to disable NMIs - we are only looking at
3460 * the list, not manipulating it:
3462 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
3463 if (counter != counter->group_leader)
3466 if (!counter->hw_event.inherit)
3469 if (inherit_group(counter, parent,
3470 parent_ctx, child, child_ctx))
3474 mutex_unlock(&parent_ctx->mutex);
3477 static void __cpuinit perf_counter_init_cpu(int cpu)
3479 struct perf_cpu_context *cpuctx;
3481 cpuctx = &per_cpu(perf_cpu_context, cpu);
3482 __perf_counter_init_context(&cpuctx->ctx, NULL);
3484 spin_lock(&perf_resource_lock);
3485 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3486 spin_unlock(&perf_resource_lock);
3488 hw_perf_counter_setup(cpu);
3491 #ifdef CONFIG_HOTPLUG_CPU
3492 static void __perf_counter_exit_cpu(void *info)
3494 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3495 struct perf_counter_context *ctx = &cpuctx->ctx;
3496 struct perf_counter *counter, *tmp;
3498 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3499 __perf_counter_remove_from_context(counter);
3501 static void perf_counter_exit_cpu(int cpu)
3503 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3504 struct perf_counter_context *ctx = &cpuctx->ctx;
3506 mutex_lock(&ctx->mutex);
3507 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3508 mutex_unlock(&ctx->mutex);
3511 static inline void perf_counter_exit_cpu(int cpu) { }
3514 static int __cpuinit
3515 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3517 unsigned int cpu = (long)hcpu;
3521 case CPU_UP_PREPARE:
3522 case CPU_UP_PREPARE_FROZEN:
3523 perf_counter_init_cpu(cpu);
3526 case CPU_DOWN_PREPARE:
3527 case CPU_DOWN_PREPARE_FROZEN:
3528 perf_counter_exit_cpu(cpu);
3538 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3539 .notifier_call = perf_cpu_notify,
3542 void __init perf_counter_init(void)
3544 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3545 (void *)(long)smp_processor_id());
3546 register_cpu_notifier(&perf_cpu_nb);
3549 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3551 return sprintf(buf, "%d\n", perf_reserved_percpu);
3555 perf_set_reserve_percpu(struct sysdev_class *class,
3559 struct perf_cpu_context *cpuctx;
3563 err = strict_strtoul(buf, 10, &val);
3566 if (val > perf_max_counters)
3569 spin_lock(&perf_resource_lock);
3570 perf_reserved_percpu = val;
3571 for_each_online_cpu(cpu) {
3572 cpuctx = &per_cpu(perf_cpu_context, cpu);
3573 spin_lock_irq(&cpuctx->ctx.lock);
3574 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3575 perf_max_counters - perf_reserved_percpu);
3576 cpuctx->max_pertask = mpt;
3577 spin_unlock_irq(&cpuctx->ctx.lock);
3579 spin_unlock(&perf_resource_lock);
3584 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3586 return sprintf(buf, "%d\n", perf_overcommit);
3590 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3595 err = strict_strtoul(buf, 10, &val);
3601 spin_lock(&perf_resource_lock);
3602 perf_overcommit = val;
3603 spin_unlock(&perf_resource_lock);
3608 static SYSDEV_CLASS_ATTR(
3611 perf_show_reserve_percpu,
3612 perf_set_reserve_percpu
3615 static SYSDEV_CLASS_ATTR(
3618 perf_show_overcommit,
3622 static struct attribute *perfclass_attrs[] = {
3623 &attr_reserve_percpu.attr,
3624 &attr_overcommit.attr,
3628 static struct attribute_group perfclass_attr_group = {
3629 .attrs = perfclass_attrs,
3630 .name = "perf_counters",
3633 static int __init perf_counter_sysfs_init(void)
3635 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3636 &perfclass_attr_group);
3638 device_initcall(perf_counter_sysfs_init);