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/dcache.h>
20 #include <linux/percpu.h>
21 #include <linux/ptrace.h>
22 #include <linux/vmstat.h>
23 #include <linux/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.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 */
49 int sysctl_perf_counter_limit __read_mostly = 100000; /* max NMIs per second */
51 static atomic64_t perf_counter_id;
54 * Lock for (sysadmin-configurable) counter reservations:
56 static DEFINE_SPINLOCK(perf_resource_lock);
59 * Architecture provided APIs - weak aliases:
61 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
66 void __weak hw_perf_disable(void) { barrier(); }
67 void __weak hw_perf_enable(void) { barrier(); }
69 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
72 hw_perf_group_sched_in(struct perf_counter *group_leader,
73 struct perf_cpu_context *cpuctx,
74 struct perf_counter_context *ctx, int cpu)
79 void __weak perf_counter_print_debug(void) { }
81 static DEFINE_PER_CPU(int, disable_count);
83 void __perf_disable(void)
85 __get_cpu_var(disable_count)++;
88 bool __perf_enable(void)
90 return !--__get_cpu_var(disable_count);
93 void perf_disable(void)
99 void perf_enable(void)
105 static void get_ctx(struct perf_counter_context *ctx)
107 atomic_inc(&ctx->refcount);
110 static void free_ctx(struct rcu_head *head)
112 struct perf_counter_context *ctx;
114 ctx = container_of(head, struct perf_counter_context, rcu_head);
118 static void put_ctx(struct perf_counter_context *ctx)
120 if (atomic_dec_and_test(&ctx->refcount)) {
122 put_ctx(ctx->parent_ctx);
124 put_task_struct(ctx->task);
125 call_rcu(&ctx->rcu_head, free_ctx);
130 * Get the perf_counter_context for a task and lock it.
131 * This has to cope with with the fact that until it is locked,
132 * the context could get moved to another task.
134 static struct perf_counter_context *
135 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
137 struct perf_counter_context *ctx;
141 ctx = rcu_dereference(task->perf_counter_ctxp);
144 * If this context is a clone of another, it might
145 * get swapped for another underneath us by
146 * perf_counter_task_sched_out, though the
147 * rcu_read_lock() protects us from any context
148 * getting freed. Lock the context and check if it
149 * got swapped before we could get the lock, and retry
150 * if so. If we locked the right context, then it
151 * can't get swapped on us any more.
153 spin_lock_irqsave(&ctx->lock, *flags);
154 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
155 spin_unlock_irqrestore(&ctx->lock, *flags);
164 * Get the context for a task and increment its pin_count so it
165 * can't get swapped to another task. This also increments its
166 * reference count so that the context can't get freed.
168 static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
170 struct perf_counter_context *ctx;
173 ctx = perf_lock_task_context(task, &flags);
177 spin_unlock_irqrestore(&ctx->lock, flags);
182 static void perf_unpin_context(struct perf_counter_context *ctx)
186 spin_lock_irqsave(&ctx->lock, flags);
188 spin_unlock_irqrestore(&ctx->lock, flags);
193 * Add a counter from the lists for its context.
194 * Must be called with ctx->mutex and ctx->lock held.
197 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
199 struct perf_counter *group_leader = counter->group_leader;
202 * Depending on whether it is a standalone or sibling counter,
203 * add it straight to the context's counter list, or to the group
204 * leader's sibling list:
206 if (group_leader == counter)
207 list_add_tail(&counter->list_entry, &ctx->counter_list);
209 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
210 group_leader->nr_siblings++;
213 list_add_rcu(&counter->event_entry, &ctx->event_list);
218 * Remove a counter from the lists for its context.
219 * Must be called with ctx->mutex and ctx->lock held.
222 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
224 struct perf_counter *sibling, *tmp;
226 if (list_empty(&counter->list_entry))
230 list_del_init(&counter->list_entry);
231 list_del_rcu(&counter->event_entry);
233 if (counter->group_leader != counter)
234 counter->group_leader->nr_siblings--;
237 * If this was a group counter with sibling counters then
238 * upgrade the siblings to singleton counters by adding them
239 * to the context list directly:
241 list_for_each_entry_safe(sibling, tmp,
242 &counter->sibling_list, list_entry) {
244 list_move_tail(&sibling->list_entry, &ctx->counter_list);
245 sibling->group_leader = sibling;
250 counter_sched_out(struct perf_counter *counter,
251 struct perf_cpu_context *cpuctx,
252 struct perf_counter_context *ctx)
254 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
257 counter->state = PERF_COUNTER_STATE_INACTIVE;
258 counter->tstamp_stopped = ctx->time;
259 counter->pmu->disable(counter);
262 if (!is_software_counter(counter))
263 cpuctx->active_oncpu--;
265 if (counter->attr.exclusive || !cpuctx->active_oncpu)
266 cpuctx->exclusive = 0;
270 group_sched_out(struct perf_counter *group_counter,
271 struct perf_cpu_context *cpuctx,
272 struct perf_counter_context *ctx)
274 struct perf_counter *counter;
276 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
279 counter_sched_out(group_counter, cpuctx, ctx);
282 * Schedule out siblings (if any):
284 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
285 counter_sched_out(counter, cpuctx, ctx);
287 if (group_counter->attr.exclusive)
288 cpuctx->exclusive = 0;
292 * Cross CPU call to remove a performance counter
294 * We disable the counter on the hardware level first. After that we
295 * remove it from the context list.
297 static void __perf_counter_remove_from_context(void *info)
299 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
300 struct perf_counter *counter = info;
301 struct perf_counter_context *ctx = counter->ctx;
304 * If this is a task context, we need to check whether it is
305 * the current task context of this cpu. If not it has been
306 * scheduled out before the smp call arrived.
308 if (ctx->task && cpuctx->task_ctx != ctx)
311 spin_lock(&ctx->lock);
313 * Protect the list operation against NMI by disabling the
314 * counters on a global level.
318 counter_sched_out(counter, cpuctx, ctx);
320 list_del_counter(counter, ctx);
324 * Allow more per task counters with respect to the
327 cpuctx->max_pertask =
328 min(perf_max_counters - ctx->nr_counters,
329 perf_max_counters - perf_reserved_percpu);
333 spin_unlock(&ctx->lock);
338 * Remove the counter from a task's (or a CPU's) list of counters.
340 * Must be called with ctx->mutex held.
342 * CPU counters are removed with a smp call. For task counters we only
343 * call when the task is on a CPU.
345 * If counter->ctx is a cloned context, callers must make sure that
346 * every task struct that counter->ctx->task could possibly point to
347 * remains valid. This is OK when called from perf_release since
348 * that only calls us on the top-level context, which can't be a clone.
349 * When called from perf_counter_exit_task, it's OK because the
350 * context has been detached from its task.
352 static void perf_counter_remove_from_context(struct perf_counter *counter)
354 struct perf_counter_context *ctx = counter->ctx;
355 struct task_struct *task = ctx->task;
359 * Per cpu counters are removed via an smp call and
360 * the removal is always sucessful.
362 smp_call_function_single(counter->cpu,
363 __perf_counter_remove_from_context,
369 task_oncpu_function_call(task, __perf_counter_remove_from_context,
372 spin_lock_irq(&ctx->lock);
374 * If the context is active we need to retry the smp call.
376 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
377 spin_unlock_irq(&ctx->lock);
382 * The lock prevents that this context is scheduled in so we
383 * can remove the counter safely, if the call above did not
386 if (!list_empty(&counter->list_entry)) {
387 list_del_counter(counter, ctx);
389 spin_unlock_irq(&ctx->lock);
392 static inline u64 perf_clock(void)
394 return cpu_clock(smp_processor_id());
398 * Update the record of the current time in a context.
400 static void update_context_time(struct perf_counter_context *ctx)
402 u64 now = perf_clock();
404 ctx->time += now - ctx->timestamp;
405 ctx->timestamp = now;
409 * Update the total_time_enabled and total_time_running fields for a counter.
411 static void update_counter_times(struct perf_counter *counter)
413 struct perf_counter_context *ctx = counter->ctx;
416 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
419 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
421 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
422 run_end = counter->tstamp_stopped;
426 counter->total_time_running = run_end - counter->tstamp_running;
430 * Update total_time_enabled and total_time_running for all counters in a group.
432 static void update_group_times(struct perf_counter *leader)
434 struct perf_counter *counter;
436 update_counter_times(leader);
437 list_for_each_entry(counter, &leader->sibling_list, list_entry)
438 update_counter_times(counter);
442 * Cross CPU call to disable a performance counter
444 static void __perf_counter_disable(void *info)
446 struct perf_counter *counter = info;
447 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
448 struct perf_counter_context *ctx = counter->ctx;
451 * If this is a per-task counter, need to check whether this
452 * counter's task is the current task on this cpu.
454 if (ctx->task && cpuctx->task_ctx != ctx)
457 spin_lock(&ctx->lock);
460 * If the counter is on, turn it off.
461 * If it is in error state, leave it in error state.
463 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
464 update_context_time(ctx);
465 update_counter_times(counter);
466 if (counter == counter->group_leader)
467 group_sched_out(counter, cpuctx, ctx);
469 counter_sched_out(counter, cpuctx, ctx);
470 counter->state = PERF_COUNTER_STATE_OFF;
473 spin_unlock(&ctx->lock);
479 * If counter->ctx is a cloned context, callers must make sure that
480 * every task struct that counter->ctx->task could possibly point to
481 * remains valid. This condition is satisifed when called through
482 * perf_counter_for_each_child or perf_counter_for_each because they
483 * hold the top-level counter's child_mutex, so any descendant that
484 * goes to exit will block in sync_child_counter.
485 * When called from perf_pending_counter it's OK because counter->ctx
486 * is the current context on this CPU and preemption is disabled,
487 * hence we can't get into perf_counter_task_sched_out for this context.
489 static void perf_counter_disable(struct perf_counter *counter)
491 struct perf_counter_context *ctx = counter->ctx;
492 struct task_struct *task = ctx->task;
496 * Disable the counter on the cpu that it's on
498 smp_call_function_single(counter->cpu, __perf_counter_disable,
504 task_oncpu_function_call(task, __perf_counter_disable, counter);
506 spin_lock_irq(&ctx->lock);
508 * If the counter is still active, we need to retry the cross-call.
510 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
511 spin_unlock_irq(&ctx->lock);
516 * Since we have the lock this context can't be scheduled
517 * in, so we can change the state safely.
519 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
520 update_counter_times(counter);
521 counter->state = PERF_COUNTER_STATE_OFF;
524 spin_unlock_irq(&ctx->lock);
528 counter_sched_in(struct perf_counter *counter,
529 struct perf_cpu_context *cpuctx,
530 struct perf_counter_context *ctx,
533 if (counter->state <= PERF_COUNTER_STATE_OFF)
536 counter->state = PERF_COUNTER_STATE_ACTIVE;
537 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
539 * The new state must be visible before we turn it on in the hardware:
543 if (counter->pmu->enable(counter)) {
544 counter->state = PERF_COUNTER_STATE_INACTIVE;
549 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
551 if (!is_software_counter(counter))
552 cpuctx->active_oncpu++;
555 if (counter->attr.exclusive)
556 cpuctx->exclusive = 1;
562 group_sched_in(struct perf_counter *group_counter,
563 struct perf_cpu_context *cpuctx,
564 struct perf_counter_context *ctx,
567 struct perf_counter *counter, *partial_group;
570 if (group_counter->state == PERF_COUNTER_STATE_OFF)
573 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
575 return ret < 0 ? ret : 0;
577 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
581 * Schedule in siblings as one group (if any):
583 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
584 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
585 partial_group = counter;
594 * Groups can be scheduled in as one unit only, so undo any
595 * partial group before returning:
597 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
598 if (counter == partial_group)
600 counter_sched_out(counter, cpuctx, ctx);
602 counter_sched_out(group_counter, cpuctx, ctx);
608 * Return 1 for a group consisting entirely of software counters,
609 * 0 if the group contains any hardware counters.
611 static int is_software_only_group(struct perf_counter *leader)
613 struct perf_counter *counter;
615 if (!is_software_counter(leader))
618 list_for_each_entry(counter, &leader->sibling_list, list_entry)
619 if (!is_software_counter(counter))
626 * Work out whether we can put this counter group on the CPU now.
628 static int group_can_go_on(struct perf_counter *counter,
629 struct perf_cpu_context *cpuctx,
633 * Groups consisting entirely of software counters can always go on.
635 if (is_software_only_group(counter))
638 * If an exclusive group is already on, no other hardware
639 * counters can go on.
641 if (cpuctx->exclusive)
644 * If this group is exclusive and there are already
645 * counters on the CPU, it can't go on.
647 if (counter->attr.exclusive && cpuctx->active_oncpu)
650 * Otherwise, try to add it if all previous groups were able
656 static void add_counter_to_ctx(struct perf_counter *counter,
657 struct perf_counter_context *ctx)
659 list_add_counter(counter, ctx);
660 counter->tstamp_enabled = ctx->time;
661 counter->tstamp_running = ctx->time;
662 counter->tstamp_stopped = ctx->time;
666 * Cross CPU call to install and enable a performance counter
668 * Must be called with ctx->mutex held
670 static void __perf_install_in_context(void *info)
672 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
673 struct perf_counter *counter = info;
674 struct perf_counter_context *ctx = counter->ctx;
675 struct perf_counter *leader = counter->group_leader;
676 int cpu = smp_processor_id();
680 * If this is a task context, we need to check whether it is
681 * the current task context of this cpu. If not it has been
682 * scheduled out before the smp call arrived.
683 * Or possibly this is the right context but it isn't
684 * on this cpu because it had no counters.
686 if (ctx->task && cpuctx->task_ctx != ctx) {
687 if (cpuctx->task_ctx || ctx->task != current)
689 cpuctx->task_ctx = ctx;
692 spin_lock(&ctx->lock);
694 update_context_time(ctx);
697 * Protect the list operation against NMI by disabling the
698 * counters on a global level. NOP for non NMI based counters.
702 add_counter_to_ctx(counter, ctx);
705 * Don't put the counter on if it is disabled or if
706 * it is in a group and the group isn't on.
708 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
709 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
713 * An exclusive counter can't go on if there are already active
714 * hardware counters, and no hardware counter can go on if there
715 * is already an exclusive counter on.
717 if (!group_can_go_on(counter, cpuctx, 1))
720 err = counter_sched_in(counter, cpuctx, ctx, cpu);
724 * This counter couldn't go on. If it is in a group
725 * then we have to pull the whole group off.
726 * If the counter group is pinned then put it in error state.
728 if (leader != counter)
729 group_sched_out(leader, cpuctx, ctx);
730 if (leader->attr.pinned) {
731 update_group_times(leader);
732 leader->state = PERF_COUNTER_STATE_ERROR;
736 if (!err && !ctx->task && cpuctx->max_pertask)
737 cpuctx->max_pertask--;
742 spin_unlock(&ctx->lock);
746 * Attach a performance counter to a context
748 * First we add the counter to the list with the hardware enable bit
749 * in counter->hw_config cleared.
751 * If the counter is attached to a task which is on a CPU we use a smp
752 * call to enable it in the task context. The task might have been
753 * scheduled away, but we check this in the smp call again.
755 * Must be called with ctx->mutex held.
758 perf_install_in_context(struct perf_counter_context *ctx,
759 struct perf_counter *counter,
762 struct task_struct *task = ctx->task;
766 * Per cpu counters are installed via an smp call and
767 * the install is always sucessful.
769 smp_call_function_single(cpu, __perf_install_in_context,
775 task_oncpu_function_call(task, __perf_install_in_context,
778 spin_lock_irq(&ctx->lock);
780 * we need to retry the smp call.
782 if (ctx->is_active && list_empty(&counter->list_entry)) {
783 spin_unlock_irq(&ctx->lock);
788 * The lock prevents that this context is scheduled in so we
789 * can add the counter safely, if it the call above did not
792 if (list_empty(&counter->list_entry))
793 add_counter_to_ctx(counter, ctx);
794 spin_unlock_irq(&ctx->lock);
798 * Cross CPU call to enable a performance counter
800 static void __perf_counter_enable(void *info)
802 struct perf_counter *counter = info;
803 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
804 struct perf_counter_context *ctx = counter->ctx;
805 struct perf_counter *leader = counter->group_leader;
809 * If this is a per-task counter, need to check whether this
810 * counter's task is the current task on this cpu.
812 if (ctx->task && cpuctx->task_ctx != ctx) {
813 if (cpuctx->task_ctx || ctx->task != current)
815 cpuctx->task_ctx = ctx;
818 spin_lock(&ctx->lock);
820 update_context_time(ctx);
822 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
824 counter->state = PERF_COUNTER_STATE_INACTIVE;
825 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
828 * If the counter is in a group and isn't the group leader,
829 * then don't put it on unless the group is on.
831 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
834 if (!group_can_go_on(counter, cpuctx, 1)) {
838 if (counter == leader)
839 err = group_sched_in(counter, cpuctx, ctx,
842 err = counter_sched_in(counter, cpuctx, ctx,
849 * If this counter can't go on and it's part of a
850 * group, then the whole group has to come off.
852 if (leader != counter)
853 group_sched_out(leader, cpuctx, ctx);
854 if (leader->attr.pinned) {
855 update_group_times(leader);
856 leader->state = PERF_COUNTER_STATE_ERROR;
861 spin_unlock(&ctx->lock);
867 * If counter->ctx is a cloned context, callers must make sure that
868 * every task struct that counter->ctx->task could possibly point to
869 * remains valid. This condition is satisfied when called through
870 * perf_counter_for_each_child or perf_counter_for_each as described
871 * for perf_counter_disable.
873 static void perf_counter_enable(struct perf_counter *counter)
875 struct perf_counter_context *ctx = counter->ctx;
876 struct task_struct *task = ctx->task;
880 * Enable the counter on the cpu that it's on
882 smp_call_function_single(counter->cpu, __perf_counter_enable,
887 spin_lock_irq(&ctx->lock);
888 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
892 * If the counter is in error state, clear that first.
893 * That way, if we see the counter in error state below, we
894 * know that it has gone back into error state, as distinct
895 * from the task having been scheduled away before the
896 * cross-call arrived.
898 if (counter->state == PERF_COUNTER_STATE_ERROR)
899 counter->state = PERF_COUNTER_STATE_OFF;
902 spin_unlock_irq(&ctx->lock);
903 task_oncpu_function_call(task, __perf_counter_enable, counter);
905 spin_lock_irq(&ctx->lock);
908 * If the context is active and the counter is still off,
909 * we need to retry the cross-call.
911 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
915 * Since we have the lock this context can't be scheduled
916 * in, so we can change the state safely.
918 if (counter->state == PERF_COUNTER_STATE_OFF) {
919 counter->state = PERF_COUNTER_STATE_INACTIVE;
920 counter->tstamp_enabled =
921 ctx->time - counter->total_time_enabled;
924 spin_unlock_irq(&ctx->lock);
927 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
930 * not supported on inherited counters
932 if (counter->attr.inherit)
935 atomic_add(refresh, &counter->event_limit);
936 perf_counter_enable(counter);
941 void __perf_counter_sched_out(struct perf_counter_context *ctx,
942 struct perf_cpu_context *cpuctx)
944 struct perf_counter *counter;
946 spin_lock(&ctx->lock);
948 if (likely(!ctx->nr_counters))
950 update_context_time(ctx);
953 if (ctx->nr_active) {
954 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
955 if (counter != counter->group_leader)
956 counter_sched_out(counter, cpuctx, ctx);
958 group_sched_out(counter, cpuctx, ctx);
963 spin_unlock(&ctx->lock);
967 * Test whether two contexts are equivalent, i.e. whether they
968 * have both been cloned from the same version of the same context
969 * and they both have the same number of enabled counters.
970 * If the number of enabled counters is the same, then the set
971 * of enabled counters should be the same, because these are both
972 * inherited contexts, therefore we can't access individual counters
973 * in them directly with an fd; we can only enable/disable all
974 * counters via prctl, or enable/disable all counters in a family
975 * via ioctl, which will have the same effect on both contexts.
977 static int context_equiv(struct perf_counter_context *ctx1,
978 struct perf_counter_context *ctx2)
980 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
981 && ctx1->parent_gen == ctx2->parent_gen
982 && !ctx1->pin_count && !ctx2->pin_count;
986 * Called from scheduler to remove the counters of the current task,
987 * with interrupts disabled.
989 * We stop each counter and update the counter value in counter->count.
991 * This does not protect us against NMI, but disable()
992 * sets the disabled bit in the control field of counter _before_
993 * accessing the counter control register. If a NMI hits, then it will
994 * not restart the counter.
996 void perf_counter_task_sched_out(struct task_struct *task,
997 struct task_struct *next, int cpu)
999 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1000 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1001 struct perf_counter_context *next_ctx;
1002 struct perf_counter_context *parent;
1003 struct pt_regs *regs;
1006 regs = task_pt_regs(task);
1007 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
1009 if (likely(!ctx || !cpuctx->task_ctx))
1012 update_context_time(ctx);
1015 parent = rcu_dereference(ctx->parent_ctx);
1016 next_ctx = next->perf_counter_ctxp;
1017 if (parent && next_ctx &&
1018 rcu_dereference(next_ctx->parent_ctx) == parent) {
1020 * Looks like the two contexts are clones, so we might be
1021 * able to optimize the context switch. We lock both
1022 * contexts and check that they are clones under the
1023 * lock (including re-checking that neither has been
1024 * uncloned in the meantime). It doesn't matter which
1025 * order we take the locks because no other cpu could
1026 * be trying to lock both of these tasks.
1028 spin_lock(&ctx->lock);
1029 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1030 if (context_equiv(ctx, next_ctx)) {
1032 * XXX do we need a memory barrier of sorts
1033 * wrt to rcu_dereference() of perf_counter_ctxp
1035 task->perf_counter_ctxp = next_ctx;
1036 next->perf_counter_ctxp = ctx;
1038 next_ctx->task = task;
1041 spin_unlock(&next_ctx->lock);
1042 spin_unlock(&ctx->lock);
1047 __perf_counter_sched_out(ctx, cpuctx);
1048 cpuctx->task_ctx = NULL;
1053 * Called with IRQs disabled
1055 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1057 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1059 if (!cpuctx->task_ctx)
1062 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1065 __perf_counter_sched_out(ctx, cpuctx);
1066 cpuctx->task_ctx = NULL;
1070 * Called with IRQs disabled
1072 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1074 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1078 __perf_counter_sched_in(struct perf_counter_context *ctx,
1079 struct perf_cpu_context *cpuctx, int cpu)
1081 struct perf_counter *counter;
1084 spin_lock(&ctx->lock);
1086 if (likely(!ctx->nr_counters))
1089 ctx->timestamp = perf_clock();
1094 * First go through the list and put on any pinned groups
1095 * in order to give them the best chance of going on.
1097 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1098 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1099 !counter->attr.pinned)
1101 if (counter->cpu != -1 && counter->cpu != cpu)
1104 if (counter != counter->group_leader)
1105 counter_sched_in(counter, cpuctx, ctx, cpu);
1107 if (group_can_go_on(counter, cpuctx, 1))
1108 group_sched_in(counter, cpuctx, ctx, cpu);
1112 * If this pinned group hasn't been scheduled,
1113 * put it in error state.
1115 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1116 update_group_times(counter);
1117 counter->state = PERF_COUNTER_STATE_ERROR;
1121 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1123 * Ignore counters in OFF or ERROR state, and
1124 * ignore pinned counters since we did them already.
1126 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1127 counter->attr.pinned)
1131 * Listen to the 'cpu' scheduling filter constraint
1134 if (counter->cpu != -1 && counter->cpu != cpu)
1137 if (counter != counter->group_leader) {
1138 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1141 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1142 if (group_sched_in(counter, cpuctx, ctx, cpu))
1149 spin_unlock(&ctx->lock);
1153 * Called from scheduler to add the counters of the current task
1154 * with interrupts disabled.
1156 * We restore the counter value and then enable it.
1158 * This does not protect us against NMI, but enable()
1159 * sets the enabled bit in the control field of counter _before_
1160 * accessing the counter control register. If a NMI hits, then it will
1161 * keep the counter running.
1163 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1165 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1166 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1170 if (cpuctx->task_ctx == ctx)
1172 __perf_counter_sched_in(ctx, cpuctx, cpu);
1173 cpuctx->task_ctx = ctx;
1176 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1178 struct perf_counter_context *ctx = &cpuctx->ctx;
1180 __perf_counter_sched_in(ctx, cpuctx, cpu);
1183 #define MAX_INTERRUPTS (~0ULL)
1185 static void perf_log_throttle(struct perf_counter *counter, int enable);
1186 static void perf_log_period(struct perf_counter *counter, u64 period);
1188 static void perf_adjust_freq(struct perf_counter_context *ctx)
1190 struct perf_counter *counter;
1191 u64 interrupts, sample_period;
1195 spin_lock(&ctx->lock);
1196 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1197 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1200 interrupts = counter->hw.interrupts;
1201 counter->hw.interrupts = 0;
1203 if (interrupts == MAX_INTERRUPTS) {
1204 perf_log_throttle(counter, 1);
1205 counter->pmu->unthrottle(counter);
1206 interrupts = 2*sysctl_perf_counter_limit/HZ;
1209 if (!counter->attr.freq || !counter->attr.sample_freq)
1212 events = HZ * interrupts * counter->hw.sample_period;
1213 period = div64_u64(events, counter->attr.sample_freq);
1215 delta = (s64)(1 + period - counter->hw.sample_period);
1218 sample_period = counter->hw.sample_period + delta;
1223 perf_log_period(counter, sample_period);
1225 counter->hw.sample_period = sample_period;
1227 spin_unlock(&ctx->lock);
1231 * Round-robin a context's counters:
1233 static void rotate_ctx(struct perf_counter_context *ctx)
1235 struct perf_counter *counter;
1237 if (!ctx->nr_counters)
1240 spin_lock(&ctx->lock);
1242 * Rotate the first entry last (works just fine for group counters too):
1245 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1246 list_move_tail(&counter->list_entry, &ctx->counter_list);
1251 spin_unlock(&ctx->lock);
1254 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1256 struct perf_cpu_context *cpuctx;
1257 struct perf_counter_context *ctx;
1259 if (!atomic_read(&nr_counters))
1262 cpuctx = &per_cpu(perf_cpu_context, cpu);
1263 ctx = curr->perf_counter_ctxp;
1265 perf_adjust_freq(&cpuctx->ctx);
1267 perf_adjust_freq(ctx);
1269 perf_counter_cpu_sched_out(cpuctx);
1271 __perf_counter_task_sched_out(ctx);
1273 rotate_ctx(&cpuctx->ctx);
1277 perf_counter_cpu_sched_in(cpuctx, cpu);
1279 perf_counter_task_sched_in(curr, cpu);
1283 * Cross CPU call to read the hardware counter
1285 static void __read(void *info)
1287 struct perf_counter *counter = info;
1288 struct perf_counter_context *ctx = counter->ctx;
1289 unsigned long flags;
1291 local_irq_save(flags);
1293 update_context_time(ctx);
1294 counter->pmu->read(counter);
1295 update_counter_times(counter);
1296 local_irq_restore(flags);
1299 static u64 perf_counter_read(struct perf_counter *counter)
1302 * If counter is enabled and currently active on a CPU, update the
1303 * value in the counter structure:
1305 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1306 smp_call_function_single(counter->oncpu,
1307 __read, counter, 1);
1308 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1309 update_counter_times(counter);
1312 return atomic64_read(&counter->count);
1316 * Initialize the perf_counter context in a task_struct:
1319 __perf_counter_init_context(struct perf_counter_context *ctx,
1320 struct task_struct *task)
1322 memset(ctx, 0, sizeof(*ctx));
1323 spin_lock_init(&ctx->lock);
1324 mutex_init(&ctx->mutex);
1325 INIT_LIST_HEAD(&ctx->counter_list);
1326 INIT_LIST_HEAD(&ctx->event_list);
1327 atomic_set(&ctx->refcount, 1);
1331 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1333 struct perf_counter_context *parent_ctx;
1334 struct perf_counter_context *ctx;
1335 struct perf_cpu_context *cpuctx;
1336 struct task_struct *task;
1337 unsigned long flags;
1341 * If cpu is not a wildcard then this is a percpu counter:
1344 /* Must be root to operate on a CPU counter: */
1345 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1346 return ERR_PTR(-EACCES);
1348 if (cpu < 0 || cpu > num_possible_cpus())
1349 return ERR_PTR(-EINVAL);
1352 * We could be clever and allow to attach a counter to an
1353 * offline CPU and activate it when the CPU comes up, but
1356 if (!cpu_isset(cpu, cpu_online_map))
1357 return ERR_PTR(-ENODEV);
1359 cpuctx = &per_cpu(perf_cpu_context, cpu);
1370 task = find_task_by_vpid(pid);
1372 get_task_struct(task);
1376 return ERR_PTR(-ESRCH);
1379 * Can't attach counters to a dying task.
1382 if (task->flags & PF_EXITING)
1385 /* Reuse ptrace permission checks for now. */
1387 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1391 ctx = perf_lock_task_context(task, &flags);
1393 parent_ctx = ctx->parent_ctx;
1395 put_ctx(parent_ctx);
1396 ctx->parent_ctx = NULL; /* no longer a clone */
1399 * Get an extra reference before dropping the lock so that
1400 * this context won't get freed if the task exits.
1403 spin_unlock_irqrestore(&ctx->lock, flags);
1407 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1411 __perf_counter_init_context(ctx, task);
1413 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1415 * We raced with some other task; use
1416 * the context they set.
1421 get_task_struct(task);
1424 put_task_struct(task);
1428 put_task_struct(task);
1429 return ERR_PTR(err);
1432 static void free_counter_rcu(struct rcu_head *head)
1434 struct perf_counter *counter;
1436 counter = container_of(head, struct perf_counter, rcu_head);
1438 put_pid_ns(counter->ns);
1442 static void perf_pending_sync(struct perf_counter *counter);
1444 static void free_counter(struct perf_counter *counter)
1446 perf_pending_sync(counter);
1448 atomic_dec(&nr_counters);
1449 if (counter->attr.mmap)
1450 atomic_dec(&nr_mmap_tracking);
1451 if (counter->attr.munmap)
1452 atomic_dec(&nr_munmap_tracking);
1453 if (counter->attr.comm)
1454 atomic_dec(&nr_comm_tracking);
1456 if (counter->destroy)
1457 counter->destroy(counter);
1459 put_ctx(counter->ctx);
1460 call_rcu(&counter->rcu_head, free_counter_rcu);
1464 * Called when the last reference to the file is gone.
1466 static int perf_release(struct inode *inode, struct file *file)
1468 struct perf_counter *counter = file->private_data;
1469 struct perf_counter_context *ctx = counter->ctx;
1471 file->private_data = NULL;
1473 WARN_ON_ONCE(ctx->parent_ctx);
1474 mutex_lock(&ctx->mutex);
1475 perf_counter_remove_from_context(counter);
1476 mutex_unlock(&ctx->mutex);
1478 mutex_lock(&counter->owner->perf_counter_mutex);
1479 list_del_init(&counter->owner_entry);
1480 mutex_unlock(&counter->owner->perf_counter_mutex);
1481 put_task_struct(counter->owner);
1483 free_counter(counter);
1489 * Read the performance counter - simple non blocking version for now
1492 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1498 * Return end-of-file for a read on a counter that is in
1499 * error state (i.e. because it was pinned but it couldn't be
1500 * scheduled on to the CPU at some point).
1502 if (counter->state == PERF_COUNTER_STATE_ERROR)
1505 WARN_ON_ONCE(counter->ctx->parent_ctx);
1506 mutex_lock(&counter->child_mutex);
1507 values[0] = perf_counter_read(counter);
1509 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1510 values[n++] = counter->total_time_enabled +
1511 atomic64_read(&counter->child_total_time_enabled);
1512 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1513 values[n++] = counter->total_time_running +
1514 atomic64_read(&counter->child_total_time_running);
1515 if (counter->attr.read_format & PERF_FORMAT_ID)
1516 values[n++] = counter->id;
1517 mutex_unlock(&counter->child_mutex);
1519 if (count < n * sizeof(u64))
1521 count = n * sizeof(u64);
1523 if (copy_to_user(buf, values, count))
1530 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1532 struct perf_counter *counter = file->private_data;
1534 return perf_read_hw(counter, buf, count);
1537 static unsigned int perf_poll(struct file *file, poll_table *wait)
1539 struct perf_counter *counter = file->private_data;
1540 struct perf_mmap_data *data;
1541 unsigned int events = POLL_HUP;
1544 data = rcu_dereference(counter->data);
1546 events = atomic_xchg(&data->poll, 0);
1549 poll_wait(file, &counter->waitq, wait);
1554 static void perf_counter_reset(struct perf_counter *counter)
1556 (void)perf_counter_read(counter);
1557 atomic64_set(&counter->count, 0);
1558 perf_counter_update_userpage(counter);
1561 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1562 void (*func)(struct perf_counter *))
1564 struct perf_counter_context *ctx = counter->ctx;
1565 struct perf_counter *sibling;
1567 WARN_ON_ONCE(ctx->parent_ctx);
1568 mutex_lock(&ctx->mutex);
1569 counter = counter->group_leader;
1572 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1574 mutex_unlock(&ctx->mutex);
1578 * Holding the top-level counter's child_mutex means that any
1579 * descendant process that has inherited this counter will block
1580 * in sync_child_counter if it goes to exit, thus satisfying the
1581 * task existence requirements of perf_counter_enable/disable.
1583 static void perf_counter_for_each_child(struct perf_counter *counter,
1584 void (*func)(struct perf_counter *))
1586 struct perf_counter *child;
1588 WARN_ON_ONCE(counter->ctx->parent_ctx);
1589 mutex_lock(&counter->child_mutex);
1591 list_for_each_entry(child, &counter->child_list, child_list)
1593 mutex_unlock(&counter->child_mutex);
1596 static void perf_counter_for_each(struct perf_counter *counter,
1597 void (*func)(struct perf_counter *))
1599 struct perf_counter *child;
1601 WARN_ON_ONCE(counter->ctx->parent_ctx);
1602 mutex_lock(&counter->child_mutex);
1603 perf_counter_for_each_sibling(counter, func);
1604 list_for_each_entry(child, &counter->child_list, child_list)
1605 perf_counter_for_each_sibling(child, func);
1606 mutex_unlock(&counter->child_mutex);
1609 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1611 struct perf_counter_context *ctx = counter->ctx;
1616 if (!counter->attr.sample_period)
1619 size = copy_from_user(&value, arg, sizeof(value));
1620 if (size != sizeof(value))
1626 spin_lock_irq(&ctx->lock);
1627 if (counter->attr.freq) {
1628 if (value > sysctl_perf_counter_limit) {
1633 counter->attr.sample_freq = value;
1635 counter->attr.sample_period = value;
1636 counter->hw.sample_period = value;
1638 perf_log_period(counter, value);
1641 spin_unlock_irq(&ctx->lock);
1646 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1648 struct perf_counter *counter = file->private_data;
1649 void (*func)(struct perf_counter *);
1653 case PERF_COUNTER_IOC_ENABLE:
1654 func = perf_counter_enable;
1656 case PERF_COUNTER_IOC_DISABLE:
1657 func = perf_counter_disable;
1659 case PERF_COUNTER_IOC_RESET:
1660 func = perf_counter_reset;
1663 case PERF_COUNTER_IOC_REFRESH:
1664 return perf_counter_refresh(counter, arg);
1666 case PERF_COUNTER_IOC_PERIOD:
1667 return perf_counter_period(counter, (u64 __user *)arg);
1673 if (flags & PERF_IOC_FLAG_GROUP)
1674 perf_counter_for_each(counter, func);
1676 perf_counter_for_each_child(counter, func);
1681 int perf_counter_task_enable(void)
1683 struct perf_counter *counter;
1685 mutex_lock(¤t->perf_counter_mutex);
1686 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1687 perf_counter_for_each_child(counter, perf_counter_enable);
1688 mutex_unlock(¤t->perf_counter_mutex);
1693 int perf_counter_task_disable(void)
1695 struct perf_counter *counter;
1697 mutex_lock(¤t->perf_counter_mutex);
1698 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1699 perf_counter_for_each_child(counter, perf_counter_disable);
1700 mutex_unlock(¤t->perf_counter_mutex);
1706 * Callers need to ensure there can be no nesting of this function, otherwise
1707 * the seqlock logic goes bad. We can not serialize this because the arch
1708 * code calls this from NMI context.
1710 void perf_counter_update_userpage(struct perf_counter *counter)
1712 struct perf_counter_mmap_page *userpg;
1713 struct perf_mmap_data *data;
1716 data = rcu_dereference(counter->data);
1720 userpg = data->user_page;
1723 * Disable preemption so as to not let the corresponding user-space
1724 * spin too long if we get preempted.
1729 userpg->index = counter->hw.idx;
1730 userpg->offset = atomic64_read(&counter->count);
1731 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1732 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1741 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1743 struct perf_counter *counter = vma->vm_file->private_data;
1744 struct perf_mmap_data *data;
1745 int ret = VM_FAULT_SIGBUS;
1748 data = rcu_dereference(counter->data);
1752 if (vmf->pgoff == 0) {
1753 vmf->page = virt_to_page(data->user_page);
1755 int nr = vmf->pgoff - 1;
1757 if ((unsigned)nr > data->nr_pages)
1760 vmf->page = virt_to_page(data->data_pages[nr]);
1762 get_page(vmf->page);
1770 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1772 struct perf_mmap_data *data;
1776 WARN_ON(atomic_read(&counter->mmap_count));
1778 size = sizeof(struct perf_mmap_data);
1779 size += nr_pages * sizeof(void *);
1781 data = kzalloc(size, GFP_KERNEL);
1785 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1786 if (!data->user_page)
1787 goto fail_user_page;
1789 for (i = 0; i < nr_pages; i++) {
1790 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1791 if (!data->data_pages[i])
1792 goto fail_data_pages;
1795 data->nr_pages = nr_pages;
1796 atomic_set(&data->lock, -1);
1798 rcu_assign_pointer(counter->data, data);
1803 for (i--; i >= 0; i--)
1804 free_page((unsigned long)data->data_pages[i]);
1806 free_page((unsigned long)data->user_page);
1815 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1817 struct perf_mmap_data *data;
1820 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
1822 free_page((unsigned long)data->user_page);
1823 for (i = 0; i < data->nr_pages; i++)
1824 free_page((unsigned long)data->data_pages[i]);
1828 static void perf_mmap_data_free(struct perf_counter *counter)
1830 struct perf_mmap_data *data = counter->data;
1832 WARN_ON(atomic_read(&counter->mmap_count));
1834 rcu_assign_pointer(counter->data, NULL);
1835 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1838 static void perf_mmap_open(struct vm_area_struct *vma)
1840 struct perf_counter *counter = vma->vm_file->private_data;
1842 atomic_inc(&counter->mmap_count);
1845 static void perf_mmap_close(struct vm_area_struct *vma)
1847 struct perf_counter *counter = vma->vm_file->private_data;
1849 WARN_ON_ONCE(counter->ctx->parent_ctx);
1850 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
1851 struct user_struct *user = current_user();
1853 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
1854 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1855 perf_mmap_data_free(counter);
1856 mutex_unlock(&counter->mmap_mutex);
1860 static struct vm_operations_struct perf_mmap_vmops = {
1861 .open = perf_mmap_open,
1862 .close = perf_mmap_close,
1863 .fault = perf_mmap_fault,
1866 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1868 struct perf_counter *counter = file->private_data;
1869 unsigned long user_locked, user_lock_limit;
1870 struct user_struct *user = current_user();
1871 unsigned long locked, lock_limit;
1872 unsigned long vma_size;
1873 unsigned long nr_pages;
1874 long user_extra, extra;
1877 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1880 vma_size = vma->vm_end - vma->vm_start;
1881 nr_pages = (vma_size / PAGE_SIZE) - 1;
1884 * If we have data pages ensure they're a power-of-two number, so we
1885 * can do bitmasks instead of modulo.
1887 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1890 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1893 if (vma->vm_pgoff != 0)
1896 WARN_ON_ONCE(counter->ctx->parent_ctx);
1897 mutex_lock(&counter->mmap_mutex);
1898 if (atomic_inc_not_zero(&counter->mmap_count)) {
1899 if (nr_pages != counter->data->nr_pages)
1904 user_extra = nr_pages + 1;
1905 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1908 * Increase the limit linearly with more CPUs:
1910 user_lock_limit *= num_online_cpus();
1912 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
1915 if (user_locked > user_lock_limit)
1916 extra = user_locked - user_lock_limit;
1918 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1919 lock_limit >>= PAGE_SHIFT;
1920 locked = vma->vm_mm->locked_vm + extra;
1922 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1927 WARN_ON(counter->data);
1928 ret = perf_mmap_data_alloc(counter, nr_pages);
1932 atomic_set(&counter->mmap_count, 1);
1933 atomic_long_add(user_extra, &user->locked_vm);
1934 vma->vm_mm->locked_vm += extra;
1935 counter->data->nr_locked = extra;
1937 mutex_unlock(&counter->mmap_mutex);
1939 vma->vm_flags &= ~VM_MAYWRITE;
1940 vma->vm_flags |= VM_RESERVED;
1941 vma->vm_ops = &perf_mmap_vmops;
1946 static int perf_fasync(int fd, struct file *filp, int on)
1948 struct inode *inode = filp->f_path.dentry->d_inode;
1949 struct perf_counter *counter = filp->private_data;
1952 mutex_lock(&inode->i_mutex);
1953 retval = fasync_helper(fd, filp, on, &counter->fasync);
1954 mutex_unlock(&inode->i_mutex);
1962 static const struct file_operations perf_fops = {
1963 .release = perf_release,
1966 .unlocked_ioctl = perf_ioctl,
1967 .compat_ioctl = perf_ioctl,
1969 .fasync = perf_fasync,
1973 * Perf counter wakeup
1975 * If there's data, ensure we set the poll() state and publish everything
1976 * to user-space before waking everybody up.
1979 void perf_counter_wakeup(struct perf_counter *counter)
1981 wake_up_all(&counter->waitq);
1983 if (counter->pending_kill) {
1984 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1985 counter->pending_kill = 0;
1992 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1994 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1995 * single linked list and use cmpxchg() to add entries lockless.
1998 static void perf_pending_counter(struct perf_pending_entry *entry)
2000 struct perf_counter *counter = container_of(entry,
2001 struct perf_counter, pending);
2003 if (counter->pending_disable) {
2004 counter->pending_disable = 0;
2005 perf_counter_disable(counter);
2008 if (counter->pending_wakeup) {
2009 counter->pending_wakeup = 0;
2010 perf_counter_wakeup(counter);
2014 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2016 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2020 static void perf_pending_queue(struct perf_pending_entry *entry,
2021 void (*func)(struct perf_pending_entry *))
2023 struct perf_pending_entry **head;
2025 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2030 head = &get_cpu_var(perf_pending_head);
2033 entry->next = *head;
2034 } while (cmpxchg(head, entry->next, entry) != entry->next);
2036 set_perf_counter_pending();
2038 put_cpu_var(perf_pending_head);
2041 static int __perf_pending_run(void)
2043 struct perf_pending_entry *list;
2046 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2047 while (list != PENDING_TAIL) {
2048 void (*func)(struct perf_pending_entry *);
2049 struct perf_pending_entry *entry = list;
2056 * Ensure we observe the unqueue before we issue the wakeup,
2057 * so that we won't be waiting forever.
2058 * -- see perf_not_pending().
2069 static inline int perf_not_pending(struct perf_counter *counter)
2072 * If we flush on whatever cpu we run, there is a chance we don't
2076 __perf_pending_run();
2080 * Ensure we see the proper queue state before going to sleep
2081 * so that we do not miss the wakeup. -- see perf_pending_handle()
2084 return counter->pending.next == NULL;
2087 static void perf_pending_sync(struct perf_counter *counter)
2089 wait_event(counter->waitq, perf_not_pending(counter));
2092 void perf_counter_do_pending(void)
2094 __perf_pending_run();
2098 * Callchain support -- arch specific
2101 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2110 struct perf_output_handle {
2111 struct perf_counter *counter;
2112 struct perf_mmap_data *data;
2114 unsigned long offset;
2118 unsigned long flags;
2121 static void perf_output_wakeup(struct perf_output_handle *handle)
2123 atomic_set(&handle->data->poll, POLL_IN);
2126 handle->counter->pending_wakeup = 1;
2127 perf_pending_queue(&handle->counter->pending,
2128 perf_pending_counter);
2130 perf_counter_wakeup(handle->counter);
2134 * Curious locking construct.
2136 * We need to ensure a later event doesn't publish a head when a former
2137 * event isn't done writing. However since we need to deal with NMIs we
2138 * cannot fully serialize things.
2140 * What we do is serialize between CPUs so we only have to deal with NMI
2141 * nesting on a single CPU.
2143 * We only publish the head (and generate a wakeup) when the outer-most
2146 static void perf_output_lock(struct perf_output_handle *handle)
2148 struct perf_mmap_data *data = handle->data;
2153 local_irq_save(handle->flags);
2154 cpu = smp_processor_id();
2156 if (in_nmi() && atomic_read(&data->lock) == cpu)
2159 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2165 static void perf_output_unlock(struct perf_output_handle *handle)
2167 struct perf_mmap_data *data = handle->data;
2171 data->done_head = data->head;
2173 if (!handle->locked)
2178 * The xchg implies a full barrier that ensures all writes are done
2179 * before we publish the new head, matched by a rmb() in userspace when
2180 * reading this position.
2182 while ((head = atomic_long_xchg(&data->done_head, 0)))
2183 data->user_page->data_head = head;
2186 * NMI can happen here, which means we can miss a done_head update.
2189 cpu = atomic_xchg(&data->lock, -1);
2190 WARN_ON_ONCE(cpu != smp_processor_id());
2193 * Therefore we have to validate we did not indeed do so.
2195 if (unlikely(atomic_long_read(&data->done_head))) {
2197 * Since we had it locked, we can lock it again.
2199 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2205 if (atomic_xchg(&data->wakeup, 0))
2206 perf_output_wakeup(handle);
2208 local_irq_restore(handle->flags);
2211 static int perf_output_begin(struct perf_output_handle *handle,
2212 struct perf_counter *counter, unsigned int size,
2213 int nmi, int overflow)
2215 struct perf_mmap_data *data;
2216 unsigned int offset, head;
2219 * For inherited counters we send all the output towards the parent.
2221 if (counter->parent)
2222 counter = counter->parent;
2225 data = rcu_dereference(counter->data);
2229 handle->data = data;
2230 handle->counter = counter;
2232 handle->overflow = overflow;
2234 if (!data->nr_pages)
2237 perf_output_lock(handle);
2240 offset = head = atomic_read(&data->head);
2242 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2244 handle->offset = offset;
2245 handle->head = head;
2247 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2248 atomic_set(&data->wakeup, 1);
2253 perf_output_wakeup(handle);
2260 static void perf_output_copy(struct perf_output_handle *handle,
2261 void *buf, unsigned int len)
2263 unsigned int pages_mask;
2264 unsigned int offset;
2268 offset = handle->offset;
2269 pages_mask = handle->data->nr_pages - 1;
2270 pages = handle->data->data_pages;
2273 unsigned int page_offset;
2276 nr = (offset >> PAGE_SHIFT) & pages_mask;
2277 page_offset = offset & (PAGE_SIZE - 1);
2278 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2280 memcpy(pages[nr] + page_offset, buf, size);
2287 handle->offset = offset;
2290 * Check we didn't copy past our reservation window, taking the
2291 * possible unsigned int wrap into account.
2293 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2296 #define perf_output_put(handle, x) \
2297 perf_output_copy((handle), &(x), sizeof(x))
2299 static void perf_output_end(struct perf_output_handle *handle)
2301 struct perf_counter *counter = handle->counter;
2302 struct perf_mmap_data *data = handle->data;
2304 int wakeup_events = counter->attr.wakeup_events;
2306 if (handle->overflow && wakeup_events) {
2307 int events = atomic_inc_return(&data->events);
2308 if (events >= wakeup_events) {
2309 atomic_sub(wakeup_events, &data->events);
2310 atomic_set(&data->wakeup, 1);
2314 perf_output_unlock(handle);
2318 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2321 * only top level counters have the pid namespace they were created in
2323 if (counter->parent)
2324 counter = counter->parent;
2326 return task_tgid_nr_ns(p, counter->ns);
2329 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2332 * only top level counters have the pid namespace they were created in
2334 if (counter->parent)
2335 counter = counter->parent;
2337 return task_pid_nr_ns(p, counter->ns);
2340 static void perf_counter_output(struct perf_counter *counter,
2341 int nmi, struct pt_regs *regs, u64 addr)
2344 u64 sample_type = counter->attr.sample_type;
2345 struct perf_output_handle handle;
2346 struct perf_event_header header;
2355 struct perf_callchain_entry *callchain = NULL;
2356 int callchain_size = 0;
2363 header.size = sizeof(header);
2365 header.misc = PERF_EVENT_MISC_OVERFLOW;
2366 header.misc |= perf_misc_flags(regs);
2368 if (sample_type & PERF_SAMPLE_IP) {
2369 ip = perf_instruction_pointer(regs);
2370 header.type |= PERF_SAMPLE_IP;
2371 header.size += sizeof(ip);
2374 if (sample_type & PERF_SAMPLE_TID) {
2375 /* namespace issues */
2376 tid_entry.pid = perf_counter_pid(counter, current);
2377 tid_entry.tid = perf_counter_tid(counter, current);
2379 header.type |= PERF_SAMPLE_TID;
2380 header.size += sizeof(tid_entry);
2383 if (sample_type & PERF_SAMPLE_TIME) {
2385 * Maybe do better on x86 and provide cpu_clock_nmi()
2387 time = sched_clock();
2389 header.type |= PERF_SAMPLE_TIME;
2390 header.size += sizeof(u64);
2393 if (sample_type & PERF_SAMPLE_ADDR) {
2394 header.type |= PERF_SAMPLE_ADDR;
2395 header.size += sizeof(u64);
2398 if (sample_type & PERF_SAMPLE_CONFIG) {
2399 header.type |= PERF_SAMPLE_CONFIG;
2400 header.size += sizeof(u64);
2403 if (sample_type & PERF_SAMPLE_CPU) {
2404 header.type |= PERF_SAMPLE_CPU;
2405 header.size += sizeof(cpu_entry);
2407 cpu_entry.cpu = raw_smp_processor_id();
2410 if (sample_type & PERF_SAMPLE_GROUP) {
2411 header.type |= PERF_SAMPLE_GROUP;
2412 header.size += sizeof(u64) +
2413 counter->nr_siblings * sizeof(group_entry);
2416 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2417 callchain = perf_callchain(regs);
2420 callchain_size = (1 + callchain->nr) * sizeof(u64);
2422 header.type |= PERF_SAMPLE_CALLCHAIN;
2423 header.size += callchain_size;
2427 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2431 perf_output_put(&handle, header);
2433 if (sample_type & PERF_SAMPLE_IP)
2434 perf_output_put(&handle, ip);
2436 if (sample_type & PERF_SAMPLE_TID)
2437 perf_output_put(&handle, tid_entry);
2439 if (sample_type & PERF_SAMPLE_TIME)
2440 perf_output_put(&handle, time);
2442 if (sample_type & PERF_SAMPLE_ADDR)
2443 perf_output_put(&handle, addr);
2445 if (sample_type & PERF_SAMPLE_CONFIG)
2446 perf_output_put(&handle, counter->attr.config);
2448 if (sample_type & PERF_SAMPLE_CPU)
2449 perf_output_put(&handle, cpu_entry);
2452 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2454 if (sample_type & PERF_SAMPLE_GROUP) {
2455 struct perf_counter *leader, *sub;
2456 u64 nr = counter->nr_siblings;
2458 perf_output_put(&handle, nr);
2460 leader = counter->group_leader;
2461 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2463 sub->pmu->read(sub);
2465 group_entry.id = sub->id;
2466 group_entry.counter = atomic64_read(&sub->count);
2468 perf_output_put(&handle, group_entry);
2473 perf_output_copy(&handle, callchain, callchain_size);
2475 perf_output_end(&handle);
2482 struct perf_comm_event {
2483 struct task_struct *task;
2488 struct perf_event_header header;
2495 static void perf_counter_comm_output(struct perf_counter *counter,
2496 struct perf_comm_event *comm_event)
2498 struct perf_output_handle handle;
2499 int size = comm_event->event.header.size;
2500 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2505 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
2506 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
2508 perf_output_put(&handle, comm_event->event);
2509 perf_output_copy(&handle, comm_event->comm,
2510 comm_event->comm_size);
2511 perf_output_end(&handle);
2514 static int perf_counter_comm_match(struct perf_counter *counter,
2515 struct perf_comm_event *comm_event)
2517 if (counter->attr.comm &&
2518 comm_event->event.header.type == PERF_EVENT_COMM)
2524 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2525 struct perf_comm_event *comm_event)
2527 struct perf_counter *counter;
2529 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2533 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2534 if (perf_counter_comm_match(counter, comm_event))
2535 perf_counter_comm_output(counter, comm_event);
2540 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2542 struct perf_cpu_context *cpuctx;
2543 struct perf_counter_context *ctx;
2545 char *comm = comm_event->task->comm;
2547 size = ALIGN(strlen(comm)+1, sizeof(u64));
2549 comm_event->comm = comm;
2550 comm_event->comm_size = size;
2552 comm_event->event.header.size = sizeof(comm_event->event) + size;
2554 cpuctx = &get_cpu_var(perf_cpu_context);
2555 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2556 put_cpu_var(perf_cpu_context);
2560 * doesn't really matter which of the child contexts the
2561 * events ends up in.
2563 ctx = rcu_dereference(current->perf_counter_ctxp);
2565 perf_counter_comm_ctx(ctx, comm_event);
2569 void perf_counter_comm(struct task_struct *task)
2571 struct perf_comm_event comm_event;
2573 if (!atomic_read(&nr_comm_tracking))
2576 comm_event = (struct perf_comm_event){
2579 .header = { .type = PERF_EVENT_COMM, },
2583 perf_counter_comm_event(&comm_event);
2590 struct perf_mmap_event {
2596 struct perf_event_header header;
2606 static void perf_counter_mmap_output(struct perf_counter *counter,
2607 struct perf_mmap_event *mmap_event)
2609 struct perf_output_handle handle;
2610 int size = mmap_event->event.header.size;
2611 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2616 mmap_event->event.pid = perf_counter_pid(counter, current);
2617 mmap_event->event.tid = perf_counter_tid(counter, current);
2619 perf_output_put(&handle, mmap_event->event);
2620 perf_output_copy(&handle, mmap_event->file_name,
2621 mmap_event->file_size);
2622 perf_output_end(&handle);
2625 static int perf_counter_mmap_match(struct perf_counter *counter,
2626 struct perf_mmap_event *mmap_event)
2628 if (counter->attr.mmap &&
2629 mmap_event->event.header.type == PERF_EVENT_MMAP)
2632 if (counter->attr.munmap &&
2633 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2639 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2640 struct perf_mmap_event *mmap_event)
2642 struct perf_counter *counter;
2644 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2648 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2649 if (perf_counter_mmap_match(counter, mmap_event))
2650 perf_counter_mmap_output(counter, mmap_event);
2655 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2657 struct perf_cpu_context *cpuctx;
2658 struct perf_counter_context *ctx;
2659 struct file *file = mmap_event->file;
2666 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2668 name = strncpy(tmp, "//enomem", sizeof(tmp));
2671 name = d_path(&file->f_path, buf, PATH_MAX);
2673 name = strncpy(tmp, "//toolong", sizeof(tmp));
2677 name = strncpy(tmp, "//anon", sizeof(tmp));
2682 size = ALIGN(strlen(name)+1, sizeof(u64));
2684 mmap_event->file_name = name;
2685 mmap_event->file_size = size;
2687 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2689 cpuctx = &get_cpu_var(perf_cpu_context);
2690 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2691 put_cpu_var(perf_cpu_context);
2695 * doesn't really matter which of the child contexts the
2696 * events ends up in.
2698 ctx = rcu_dereference(current->perf_counter_ctxp);
2700 perf_counter_mmap_ctx(ctx, mmap_event);
2706 void perf_counter_mmap(unsigned long addr, unsigned long len,
2707 unsigned long pgoff, struct file *file)
2709 struct perf_mmap_event mmap_event;
2711 if (!atomic_read(&nr_mmap_tracking))
2714 mmap_event = (struct perf_mmap_event){
2717 .header = { .type = PERF_EVENT_MMAP, },
2724 perf_counter_mmap_event(&mmap_event);
2727 void perf_counter_munmap(unsigned long addr, unsigned long len,
2728 unsigned long pgoff, struct file *file)
2730 struct perf_mmap_event mmap_event;
2732 if (!atomic_read(&nr_munmap_tracking))
2735 mmap_event = (struct perf_mmap_event){
2738 .header = { .type = PERF_EVENT_MUNMAP, },
2745 perf_counter_mmap_event(&mmap_event);
2749 * Log sample_period changes so that analyzing tools can re-normalize the
2753 static void perf_log_period(struct perf_counter *counter, u64 period)
2755 struct perf_output_handle handle;
2759 struct perf_event_header header;
2764 .type = PERF_EVENT_PERIOD,
2766 .size = sizeof(freq_event),
2768 .time = sched_clock(),
2772 if (counter->hw.sample_period == period)
2775 ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
2779 perf_output_put(&handle, freq_event);
2780 perf_output_end(&handle);
2784 * IRQ throttle logging
2787 static void perf_log_throttle(struct perf_counter *counter, int enable)
2789 struct perf_output_handle handle;
2793 struct perf_event_header header;
2795 } throttle_event = {
2797 .type = PERF_EVENT_THROTTLE + 1,
2799 .size = sizeof(throttle_event),
2801 .time = sched_clock(),
2804 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
2808 perf_output_put(&handle, throttle_event);
2809 perf_output_end(&handle);
2813 * Generic counter overflow handling.
2816 int perf_counter_overflow(struct perf_counter *counter,
2817 int nmi, struct pt_regs *regs, u64 addr)
2819 int events = atomic_read(&counter->event_limit);
2820 int throttle = counter->pmu->unthrottle != NULL;
2824 counter->hw.interrupts++;
2826 if (counter->hw.interrupts != MAX_INTERRUPTS) {
2827 counter->hw.interrupts++;
2828 if (HZ*counter->hw.interrupts > (u64)sysctl_perf_counter_limit) {
2829 counter->hw.interrupts = MAX_INTERRUPTS;
2830 perf_log_throttle(counter, 0);
2835 * Keep re-disabling counters even though on the previous
2836 * pass we disabled it - just in case we raced with a
2837 * sched-in and the counter got enabled again:
2844 * XXX event_limit might not quite work as expected on inherited
2848 counter->pending_kill = POLL_IN;
2849 if (events && atomic_dec_and_test(&counter->event_limit)) {
2851 counter->pending_kill = POLL_HUP;
2853 counter->pending_disable = 1;
2854 perf_pending_queue(&counter->pending,
2855 perf_pending_counter);
2857 perf_counter_disable(counter);
2860 perf_counter_output(counter, nmi, regs, addr);
2865 * Generic software counter infrastructure
2868 static void perf_swcounter_update(struct perf_counter *counter)
2870 struct hw_perf_counter *hwc = &counter->hw;
2875 prev = atomic64_read(&hwc->prev_count);
2876 now = atomic64_read(&hwc->count);
2877 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2882 atomic64_add(delta, &counter->count);
2883 atomic64_sub(delta, &hwc->period_left);
2886 static void perf_swcounter_set_period(struct perf_counter *counter)
2888 struct hw_perf_counter *hwc = &counter->hw;
2889 s64 left = atomic64_read(&hwc->period_left);
2890 s64 period = hwc->sample_period;
2892 if (unlikely(left <= -period)) {
2894 atomic64_set(&hwc->period_left, left);
2897 if (unlikely(left <= 0)) {
2899 atomic64_add(period, &hwc->period_left);
2902 atomic64_set(&hwc->prev_count, -left);
2903 atomic64_set(&hwc->count, -left);
2906 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2908 enum hrtimer_restart ret = HRTIMER_RESTART;
2909 struct perf_counter *counter;
2910 struct pt_regs *regs;
2913 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2914 counter->pmu->read(counter);
2916 regs = get_irq_regs();
2918 * In case we exclude kernel IPs or are somehow not in interrupt
2919 * context, provide the next best thing, the user IP.
2921 if ((counter->attr.exclude_kernel || !regs) &&
2922 !counter->attr.exclude_user)
2923 regs = task_pt_regs(current);
2926 if (perf_counter_overflow(counter, 0, regs, 0))
2927 ret = HRTIMER_NORESTART;
2930 period = max_t(u64, 10000, counter->hw.sample_period);
2931 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
2936 static void perf_swcounter_overflow(struct perf_counter *counter,
2937 int nmi, struct pt_regs *regs, u64 addr)
2939 perf_swcounter_update(counter);
2940 perf_swcounter_set_period(counter);
2941 if (perf_counter_overflow(counter, nmi, regs, addr))
2942 /* soft-disable the counter */
2947 static int perf_swcounter_is_counting(struct perf_counter *counter)
2949 struct perf_counter_context *ctx;
2950 unsigned long flags;
2953 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
2956 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
2960 * If the counter is inactive, it could be just because
2961 * its task is scheduled out, or because it's in a group
2962 * which could not go on the PMU. We want to count in
2963 * the first case but not the second. If the context is
2964 * currently active then an inactive software counter must
2965 * be the second case. If it's not currently active then
2966 * we need to know whether the counter was active when the
2967 * context was last active, which we can determine by
2968 * comparing counter->tstamp_stopped with ctx->time.
2970 * We are within an RCU read-side critical section,
2971 * which protects the existence of *ctx.
2974 spin_lock_irqsave(&ctx->lock, flags);
2976 /* Re-check state now we have the lock */
2977 if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
2978 counter->ctx->is_active ||
2979 counter->tstamp_stopped < ctx->time)
2981 spin_unlock_irqrestore(&ctx->lock, flags);
2985 static int perf_swcounter_match(struct perf_counter *counter,
2986 enum perf_event_types type,
2987 u32 event, struct pt_regs *regs)
2991 event_config = ((u64) type << PERF_COUNTER_TYPE_SHIFT) | event;
2993 if (!perf_swcounter_is_counting(counter))
2996 if (counter->attr.config != event_config)
3000 if (counter->attr.exclude_user && user_mode(regs))
3003 if (counter->attr.exclude_kernel && !user_mode(regs))
3010 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3011 int nmi, struct pt_regs *regs, u64 addr)
3013 int neg = atomic64_add_negative(nr, &counter->hw.count);
3015 if (counter->hw.sample_period && !neg && regs)
3016 perf_swcounter_overflow(counter, nmi, regs, addr);
3019 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3020 enum perf_event_types type, u32 event,
3021 u64 nr, int nmi, struct pt_regs *regs,
3024 struct perf_counter *counter;
3026 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3030 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3031 if (perf_swcounter_match(counter, type, event, regs))
3032 perf_swcounter_add(counter, nr, nmi, regs, addr);
3037 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3040 return &cpuctx->recursion[3];
3043 return &cpuctx->recursion[2];
3046 return &cpuctx->recursion[1];
3048 return &cpuctx->recursion[0];
3051 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
3052 u64 nr, int nmi, struct pt_regs *regs,
3055 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3056 int *recursion = perf_swcounter_recursion_context(cpuctx);
3057 struct perf_counter_context *ctx;
3065 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3066 nr, nmi, regs, addr);
3069 * doesn't really matter which of the child contexts the
3070 * events ends up in.
3072 ctx = rcu_dereference(current->perf_counter_ctxp);
3074 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, regs, addr);
3081 put_cpu_var(perf_cpu_context);
3085 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
3087 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
3090 static void perf_swcounter_read(struct perf_counter *counter)
3092 perf_swcounter_update(counter);
3095 static int perf_swcounter_enable(struct perf_counter *counter)
3097 perf_swcounter_set_period(counter);
3101 static void perf_swcounter_disable(struct perf_counter *counter)
3103 perf_swcounter_update(counter);
3106 static const struct pmu perf_ops_generic = {
3107 .enable = perf_swcounter_enable,
3108 .disable = perf_swcounter_disable,
3109 .read = perf_swcounter_read,
3113 * Software counter: cpu wall time clock
3116 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3118 int cpu = raw_smp_processor_id();
3122 now = cpu_clock(cpu);
3123 prev = atomic64_read(&counter->hw.prev_count);
3124 atomic64_set(&counter->hw.prev_count, now);
3125 atomic64_add(now - prev, &counter->count);
3128 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3130 struct hw_perf_counter *hwc = &counter->hw;
3131 int cpu = raw_smp_processor_id();
3133 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3134 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3135 hwc->hrtimer.function = perf_swcounter_hrtimer;
3136 if (hwc->sample_period) {
3137 u64 period = max_t(u64, 10000, hwc->sample_period);
3138 __hrtimer_start_range_ns(&hwc->hrtimer,
3139 ns_to_ktime(period), 0,
3140 HRTIMER_MODE_REL, 0);
3146 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3148 if (counter->hw.sample_period)
3149 hrtimer_cancel(&counter->hw.hrtimer);
3150 cpu_clock_perf_counter_update(counter);
3153 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3155 cpu_clock_perf_counter_update(counter);
3158 static const struct pmu perf_ops_cpu_clock = {
3159 .enable = cpu_clock_perf_counter_enable,
3160 .disable = cpu_clock_perf_counter_disable,
3161 .read = cpu_clock_perf_counter_read,
3165 * Software counter: task time clock
3168 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3173 prev = atomic64_xchg(&counter->hw.prev_count, now);
3175 atomic64_add(delta, &counter->count);
3178 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3180 struct hw_perf_counter *hwc = &counter->hw;
3183 now = counter->ctx->time;
3185 atomic64_set(&hwc->prev_count, now);
3186 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3187 hwc->hrtimer.function = perf_swcounter_hrtimer;
3188 if (hwc->sample_period) {
3189 u64 period = max_t(u64, 10000, hwc->sample_period);
3190 __hrtimer_start_range_ns(&hwc->hrtimer,
3191 ns_to_ktime(period), 0,
3192 HRTIMER_MODE_REL, 0);
3198 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3200 if (counter->hw.sample_period)
3201 hrtimer_cancel(&counter->hw.hrtimer);
3202 task_clock_perf_counter_update(counter, counter->ctx->time);
3206 static void task_clock_perf_counter_read(struct perf_counter *counter)
3211 update_context_time(counter->ctx);
3212 time = counter->ctx->time;
3214 u64 now = perf_clock();
3215 u64 delta = now - counter->ctx->timestamp;
3216 time = counter->ctx->time + delta;
3219 task_clock_perf_counter_update(counter, time);
3222 static const struct pmu perf_ops_task_clock = {
3223 .enable = task_clock_perf_counter_enable,
3224 .disable = task_clock_perf_counter_disable,
3225 .read = task_clock_perf_counter_read,
3229 * Software counter: cpu migrations
3231 void perf_counter_task_migration(struct task_struct *task, int cpu)
3233 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3234 struct perf_counter_context *ctx;
3236 perf_swcounter_ctx_event(&cpuctx->ctx, PERF_TYPE_SOFTWARE,
3237 PERF_COUNT_CPU_MIGRATIONS,
3240 ctx = perf_pin_task_context(task);
3242 perf_swcounter_ctx_event(ctx, PERF_TYPE_SOFTWARE,
3243 PERF_COUNT_CPU_MIGRATIONS,
3245 perf_unpin_context(ctx);
3249 #ifdef CONFIG_EVENT_PROFILE
3250 void perf_tpcounter_event(int event_id)
3252 struct pt_regs *regs = get_irq_regs();
3255 regs = task_pt_regs(current);
3257 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
3259 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3261 extern int ftrace_profile_enable(int);
3262 extern void ftrace_profile_disable(int);
3264 static void tp_perf_counter_destroy(struct perf_counter *counter)
3266 ftrace_profile_disable(perf_event_id(&counter->attr));
3269 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3271 int event_id = perf_event_id(&counter->attr);
3274 ret = ftrace_profile_enable(event_id);
3278 counter->destroy = tp_perf_counter_destroy;
3279 counter->hw.sample_period = counter->attr.sample_period;
3281 return &perf_ops_generic;
3284 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3290 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3292 const struct pmu *pmu = NULL;
3295 * Software counters (currently) can't in general distinguish
3296 * between user, kernel and hypervisor events.
3297 * However, context switches and cpu migrations are considered
3298 * to be kernel events, and page faults are never hypervisor
3301 switch (perf_event_id(&counter->attr)) {
3302 case PERF_COUNT_CPU_CLOCK:
3303 pmu = &perf_ops_cpu_clock;
3306 case PERF_COUNT_TASK_CLOCK:
3308 * If the user instantiates this as a per-cpu counter,
3309 * use the cpu_clock counter instead.
3311 if (counter->ctx->task)
3312 pmu = &perf_ops_task_clock;
3314 pmu = &perf_ops_cpu_clock;
3317 case PERF_COUNT_PAGE_FAULTS:
3318 case PERF_COUNT_PAGE_FAULTS_MIN:
3319 case PERF_COUNT_PAGE_FAULTS_MAJ:
3320 case PERF_COUNT_CONTEXT_SWITCHES:
3321 case PERF_COUNT_CPU_MIGRATIONS:
3322 pmu = &perf_ops_generic;
3330 * Allocate and initialize a counter structure
3332 static struct perf_counter *
3333 perf_counter_alloc(struct perf_counter_attr *attr,
3335 struct perf_counter_context *ctx,
3336 struct perf_counter *group_leader,
3339 const struct pmu *pmu;
3340 struct perf_counter *counter;
3341 struct hw_perf_counter *hwc;
3344 counter = kzalloc(sizeof(*counter), gfpflags);
3346 return ERR_PTR(-ENOMEM);
3349 * Single counters are their own group leaders, with an
3350 * empty sibling list:
3353 group_leader = counter;
3355 mutex_init(&counter->child_mutex);
3356 INIT_LIST_HEAD(&counter->child_list);
3358 INIT_LIST_HEAD(&counter->list_entry);
3359 INIT_LIST_HEAD(&counter->event_entry);
3360 INIT_LIST_HEAD(&counter->sibling_list);
3361 init_waitqueue_head(&counter->waitq);
3363 mutex_init(&counter->mmap_mutex);
3366 counter->attr = *attr;
3367 counter->group_leader = group_leader;
3368 counter->pmu = NULL;
3370 counter->oncpu = -1;
3372 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
3373 counter->id = atomic64_inc_return(&perf_counter_id);
3375 counter->state = PERF_COUNTER_STATE_INACTIVE;
3378 counter->state = PERF_COUNTER_STATE_OFF;
3383 if (attr->freq && attr->sample_freq)
3384 hwc->sample_period = div64_u64(TICK_NSEC, attr->sample_freq);
3386 hwc->sample_period = attr->sample_period;
3389 * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3391 if (attr->inherit && (attr->sample_type & PERF_SAMPLE_GROUP))
3394 if (perf_event_raw(attr)) {
3395 pmu = hw_perf_counter_init(counter);
3399 switch (perf_event_type(attr)) {
3400 case PERF_TYPE_HARDWARE:
3401 pmu = hw_perf_counter_init(counter);
3404 case PERF_TYPE_SOFTWARE:
3405 pmu = sw_perf_counter_init(counter);
3408 case PERF_TYPE_TRACEPOINT:
3409 pmu = tp_perf_counter_init(counter);
3416 else if (IS_ERR(pmu))
3421 put_pid_ns(counter->ns);
3423 return ERR_PTR(err);
3428 atomic_inc(&nr_counters);
3429 if (counter->attr.mmap)
3430 atomic_inc(&nr_mmap_tracking);
3431 if (counter->attr.munmap)
3432 atomic_inc(&nr_munmap_tracking);
3433 if (counter->attr.comm)
3434 atomic_inc(&nr_comm_tracking);
3440 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3442 * @attr_uptr: event type attributes for monitoring/sampling
3445 * @group_fd: group leader counter fd
3447 SYSCALL_DEFINE5(perf_counter_open,
3448 const struct perf_counter_attr __user *, attr_uptr,
3449 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3451 struct perf_counter *counter, *group_leader;
3452 struct perf_counter_attr attr;
3453 struct perf_counter_context *ctx;
3454 struct file *counter_file = NULL;
3455 struct file *group_file = NULL;
3456 int fput_needed = 0;
3457 int fput_needed2 = 0;
3460 /* for future expandability... */
3464 if (copy_from_user(&attr, attr_uptr, sizeof(attr)) != 0)
3468 * Get the target context (task or percpu):
3470 ctx = find_get_context(pid, cpu);
3472 return PTR_ERR(ctx);
3475 * Look up the group leader (we will attach this counter to it):
3477 group_leader = NULL;
3478 if (group_fd != -1) {
3480 group_file = fget_light(group_fd, &fput_needed);
3482 goto err_put_context;
3483 if (group_file->f_op != &perf_fops)
3484 goto err_put_context;
3486 group_leader = group_file->private_data;
3488 * Do not allow a recursive hierarchy (this new sibling
3489 * becoming part of another group-sibling):
3491 if (group_leader->group_leader != group_leader)
3492 goto err_put_context;
3494 * Do not allow to attach to a group in a different
3495 * task or CPU context:
3497 if (group_leader->ctx != ctx)
3498 goto err_put_context;
3500 * Only a group leader can be exclusive or pinned
3502 if (attr.exclusive || attr.pinned)
3503 goto err_put_context;
3506 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
3508 ret = PTR_ERR(counter);
3509 if (IS_ERR(counter))
3510 goto err_put_context;
3512 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3514 goto err_free_put_context;
3516 counter_file = fget_light(ret, &fput_needed2);
3518 goto err_free_put_context;
3520 counter->filp = counter_file;
3521 WARN_ON_ONCE(ctx->parent_ctx);
3522 mutex_lock(&ctx->mutex);
3523 perf_install_in_context(ctx, counter, cpu);
3525 mutex_unlock(&ctx->mutex);
3527 counter->owner = current;
3528 get_task_struct(current);
3529 mutex_lock(¤t->perf_counter_mutex);
3530 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
3531 mutex_unlock(¤t->perf_counter_mutex);
3533 fput_light(counter_file, fput_needed2);
3536 fput_light(group_file, fput_needed);
3540 err_free_put_context:
3550 * inherit a counter from parent task to child task:
3552 static struct perf_counter *
3553 inherit_counter(struct perf_counter *parent_counter,
3554 struct task_struct *parent,
3555 struct perf_counter_context *parent_ctx,
3556 struct task_struct *child,
3557 struct perf_counter *group_leader,
3558 struct perf_counter_context *child_ctx)
3560 struct perf_counter *child_counter;
3563 * Instead of creating recursive hierarchies of counters,
3564 * we link inherited counters back to the original parent,
3565 * which has a filp for sure, which we use as the reference
3568 if (parent_counter->parent)
3569 parent_counter = parent_counter->parent;
3571 child_counter = perf_counter_alloc(&parent_counter->attr,
3572 parent_counter->cpu, child_ctx,
3573 group_leader, GFP_KERNEL);
3574 if (IS_ERR(child_counter))
3575 return child_counter;
3579 * Make the child state follow the state of the parent counter,
3580 * not its attr.disabled bit. We hold the parent's mutex,
3581 * so we won't race with perf_counter_{en, dis}able_family.
3583 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3584 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3586 child_counter->state = PERF_COUNTER_STATE_OFF;
3589 * Link it up in the child's context:
3591 add_counter_to_ctx(child_counter, child_ctx);
3593 child_counter->parent = parent_counter;
3595 * inherit into child's child as well:
3597 child_counter->attr.inherit = 1;
3600 * Get a reference to the parent filp - we will fput it
3601 * when the child counter exits. This is safe to do because
3602 * we are in the parent and we know that the filp still
3603 * exists and has a nonzero count:
3605 atomic_long_inc(&parent_counter->filp->f_count);
3608 * Link this into the parent counter's child list
3610 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
3611 mutex_lock(&parent_counter->child_mutex);
3612 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3613 mutex_unlock(&parent_counter->child_mutex);
3615 return child_counter;
3618 static int inherit_group(struct perf_counter *parent_counter,
3619 struct task_struct *parent,
3620 struct perf_counter_context *parent_ctx,
3621 struct task_struct *child,
3622 struct perf_counter_context *child_ctx)
3624 struct perf_counter *leader;
3625 struct perf_counter *sub;
3626 struct perf_counter *child_ctr;
3628 leader = inherit_counter(parent_counter, parent, parent_ctx,
3629 child, NULL, child_ctx);
3631 return PTR_ERR(leader);
3632 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3633 child_ctr = inherit_counter(sub, parent, parent_ctx,
3634 child, leader, child_ctx);
3635 if (IS_ERR(child_ctr))
3636 return PTR_ERR(child_ctr);
3641 static void sync_child_counter(struct perf_counter *child_counter,
3642 struct perf_counter *parent_counter)
3646 child_val = atomic64_read(&child_counter->count);
3649 * Add back the child's count to the parent's count:
3651 atomic64_add(child_val, &parent_counter->count);
3652 atomic64_add(child_counter->total_time_enabled,
3653 &parent_counter->child_total_time_enabled);
3654 atomic64_add(child_counter->total_time_running,
3655 &parent_counter->child_total_time_running);
3658 * Remove this counter from the parent's list
3660 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
3661 mutex_lock(&parent_counter->child_mutex);
3662 list_del_init(&child_counter->child_list);
3663 mutex_unlock(&parent_counter->child_mutex);
3666 * Release the parent counter, if this was the last
3669 fput(parent_counter->filp);
3673 __perf_counter_exit_task(struct perf_counter *child_counter,
3674 struct perf_counter_context *child_ctx)
3676 struct perf_counter *parent_counter;
3678 update_counter_times(child_counter);
3679 perf_counter_remove_from_context(child_counter);
3681 parent_counter = child_counter->parent;
3683 * It can happen that parent exits first, and has counters
3684 * that are still around due to the child reference. These
3685 * counters need to be zapped - but otherwise linger.
3687 if (parent_counter) {
3688 sync_child_counter(child_counter, parent_counter);
3689 free_counter(child_counter);
3694 * When a child task exits, feed back counter values to parent counters.
3696 void perf_counter_exit_task(struct task_struct *child)
3698 struct perf_counter *child_counter, *tmp;
3699 struct perf_counter_context *child_ctx;
3700 unsigned long flags;
3702 if (likely(!child->perf_counter_ctxp))
3705 local_irq_save(flags);
3707 * We can't reschedule here because interrupts are disabled,
3708 * and either child is current or it is a task that can't be
3709 * scheduled, so we are now safe from rescheduling changing
3712 child_ctx = child->perf_counter_ctxp;
3713 __perf_counter_task_sched_out(child_ctx);
3716 * Take the context lock here so that if find_get_context is
3717 * reading child->perf_counter_ctxp, we wait until it has
3718 * incremented the context's refcount before we do put_ctx below.
3720 spin_lock(&child_ctx->lock);
3721 child->perf_counter_ctxp = NULL;
3722 if (child_ctx->parent_ctx) {
3724 * This context is a clone; unclone it so it can't get
3725 * swapped to another process while we're removing all
3726 * the counters from it.
3728 put_ctx(child_ctx->parent_ctx);
3729 child_ctx->parent_ctx = NULL;
3731 spin_unlock(&child_ctx->lock);
3732 local_irq_restore(flags);
3734 mutex_lock(&child_ctx->mutex);
3737 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3739 __perf_counter_exit_task(child_counter, child_ctx);
3742 * If the last counter was a group counter, it will have appended all
3743 * its siblings to the list, but we obtained 'tmp' before that which
3744 * will still point to the list head terminating the iteration.
3746 if (!list_empty(&child_ctx->counter_list))
3749 mutex_unlock(&child_ctx->mutex);
3755 * free an unexposed, unused context as created by inheritance by
3756 * init_task below, used by fork() in case of fail.
3758 void perf_counter_free_task(struct task_struct *task)
3760 struct perf_counter_context *ctx = task->perf_counter_ctxp;
3761 struct perf_counter *counter, *tmp;
3766 mutex_lock(&ctx->mutex);
3768 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
3769 struct perf_counter *parent = counter->parent;
3771 if (WARN_ON_ONCE(!parent))
3774 mutex_lock(&parent->child_mutex);
3775 list_del_init(&counter->child_list);
3776 mutex_unlock(&parent->child_mutex);
3780 list_del_counter(counter, ctx);
3781 free_counter(counter);
3784 if (!list_empty(&ctx->counter_list))
3787 mutex_unlock(&ctx->mutex);
3793 * Initialize the perf_counter context in task_struct
3795 int perf_counter_init_task(struct task_struct *child)
3797 struct perf_counter_context *child_ctx, *parent_ctx;
3798 struct perf_counter_context *cloned_ctx;
3799 struct perf_counter *counter;
3800 struct task_struct *parent = current;
3801 int inherited_all = 1;
3804 child->perf_counter_ctxp = NULL;
3806 mutex_init(&child->perf_counter_mutex);
3807 INIT_LIST_HEAD(&child->perf_counter_list);
3809 if (likely(!parent->perf_counter_ctxp))
3813 * This is executed from the parent task context, so inherit
3814 * counters that have been marked for cloning.
3815 * First allocate and initialize a context for the child.
3818 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
3822 __perf_counter_init_context(child_ctx, child);
3823 child->perf_counter_ctxp = child_ctx;
3824 get_task_struct(child);
3827 * If the parent's context is a clone, pin it so it won't get
3830 parent_ctx = perf_pin_task_context(parent);
3833 * No need to check if parent_ctx != NULL here; since we saw
3834 * it non-NULL earlier, the only reason for it to become NULL
3835 * is if we exit, and since we're currently in the middle of
3836 * a fork we can't be exiting at the same time.
3840 * Lock the parent list. No need to lock the child - not PID
3841 * hashed yet and not running, so nobody can access it.
3843 mutex_lock(&parent_ctx->mutex);
3846 * We dont have to disable NMIs - we are only looking at
3847 * the list, not manipulating it:
3849 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
3850 if (counter != counter->group_leader)
3853 if (!counter->attr.inherit) {
3858 ret = inherit_group(counter, parent, parent_ctx,
3866 if (inherited_all) {
3868 * Mark the child context as a clone of the parent
3869 * context, or of whatever the parent is a clone of.
3870 * Note that if the parent is a clone, it could get
3871 * uncloned at any point, but that doesn't matter
3872 * because the list of counters and the generation
3873 * count can't have changed since we took the mutex.
3875 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
3877 child_ctx->parent_ctx = cloned_ctx;
3878 child_ctx->parent_gen = parent_ctx->parent_gen;
3880 child_ctx->parent_ctx = parent_ctx;
3881 child_ctx->parent_gen = parent_ctx->generation;
3883 get_ctx(child_ctx->parent_ctx);
3886 mutex_unlock(&parent_ctx->mutex);
3888 perf_unpin_context(parent_ctx);
3893 static void __cpuinit perf_counter_init_cpu(int cpu)
3895 struct perf_cpu_context *cpuctx;
3897 cpuctx = &per_cpu(perf_cpu_context, cpu);
3898 __perf_counter_init_context(&cpuctx->ctx, NULL);
3900 spin_lock(&perf_resource_lock);
3901 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3902 spin_unlock(&perf_resource_lock);
3904 hw_perf_counter_setup(cpu);
3907 #ifdef CONFIG_HOTPLUG_CPU
3908 static void __perf_counter_exit_cpu(void *info)
3910 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3911 struct perf_counter_context *ctx = &cpuctx->ctx;
3912 struct perf_counter *counter, *tmp;
3914 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3915 __perf_counter_remove_from_context(counter);
3917 static void perf_counter_exit_cpu(int cpu)
3919 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3920 struct perf_counter_context *ctx = &cpuctx->ctx;
3922 mutex_lock(&ctx->mutex);
3923 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3924 mutex_unlock(&ctx->mutex);
3927 static inline void perf_counter_exit_cpu(int cpu) { }
3930 static int __cpuinit
3931 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3933 unsigned int cpu = (long)hcpu;
3937 case CPU_UP_PREPARE:
3938 case CPU_UP_PREPARE_FROZEN:
3939 perf_counter_init_cpu(cpu);
3942 case CPU_DOWN_PREPARE:
3943 case CPU_DOWN_PREPARE_FROZEN:
3944 perf_counter_exit_cpu(cpu);
3955 * This has to have a higher priority than migration_notifier in sched.c.
3957 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3958 .notifier_call = perf_cpu_notify,
3962 void __init perf_counter_init(void)
3964 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3965 (void *)(long)smp_processor_id());
3966 register_cpu_notifier(&perf_cpu_nb);
3969 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3971 return sprintf(buf, "%d\n", perf_reserved_percpu);
3975 perf_set_reserve_percpu(struct sysdev_class *class,
3979 struct perf_cpu_context *cpuctx;
3983 err = strict_strtoul(buf, 10, &val);
3986 if (val > perf_max_counters)
3989 spin_lock(&perf_resource_lock);
3990 perf_reserved_percpu = val;
3991 for_each_online_cpu(cpu) {
3992 cpuctx = &per_cpu(perf_cpu_context, cpu);
3993 spin_lock_irq(&cpuctx->ctx.lock);
3994 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3995 perf_max_counters - perf_reserved_percpu);
3996 cpuctx->max_pertask = mpt;
3997 spin_unlock_irq(&cpuctx->ctx.lock);
3999 spin_unlock(&perf_resource_lock);
4004 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4006 return sprintf(buf, "%d\n", perf_overcommit);
4010 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4015 err = strict_strtoul(buf, 10, &val);
4021 spin_lock(&perf_resource_lock);
4022 perf_overcommit = val;
4023 spin_unlock(&perf_resource_lock);
4028 static SYSDEV_CLASS_ATTR(
4031 perf_show_reserve_percpu,
4032 perf_set_reserve_percpu
4035 static SYSDEV_CLASS_ATTR(
4038 perf_show_overcommit,
4042 static struct attribute *perfclass_attrs[] = {
4043 &attr_reserve_percpu.attr,
4044 &attr_overcommit.attr,
4048 static struct attribute_group perfclass_attr_group = {
4049 .attrs = perfclass_attrs,
4050 .name = "perf_counters",
4053 static int __init perf_counter_sysfs_init(void)
4055 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4056 &perfclass_attr_group);
4058 device_initcall(perf_counter_sysfs_init);