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)) {
109 put_ctx(ctx->parent_ctx);
115 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
117 struct perf_counter *group_leader = counter->group_leader;
120 * Depending on whether it is a standalone or sibling counter,
121 * add it straight to the context's counter list, or to the group
122 * leader's sibling list:
124 if (group_leader == counter)
125 list_add_tail(&counter->list_entry, &ctx->counter_list);
127 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
128 group_leader->nr_siblings++;
131 list_add_rcu(&counter->event_entry, &ctx->event_list);
133 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
138 * Remove a counter from the lists for its context.
139 * Must be called with counter->mutex and ctx->mutex held.
142 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
144 struct perf_counter *sibling, *tmp;
146 if (list_empty(&counter->list_entry))
149 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
152 list_del_init(&counter->list_entry);
153 list_del_rcu(&counter->event_entry);
155 if (counter->group_leader != counter)
156 counter->group_leader->nr_siblings--;
159 * If this was a group counter with sibling counters then
160 * upgrade the siblings to singleton counters by adding them
161 * to the context list directly:
163 list_for_each_entry_safe(sibling, tmp,
164 &counter->sibling_list, list_entry) {
166 list_move_tail(&sibling->list_entry, &ctx->counter_list);
167 sibling->group_leader = sibling;
172 counter_sched_out(struct perf_counter *counter,
173 struct perf_cpu_context *cpuctx,
174 struct perf_counter_context *ctx)
176 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
179 counter->state = PERF_COUNTER_STATE_INACTIVE;
180 counter->tstamp_stopped = ctx->time;
181 counter->pmu->disable(counter);
184 if (!is_software_counter(counter))
185 cpuctx->active_oncpu--;
187 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
188 cpuctx->exclusive = 0;
192 group_sched_out(struct perf_counter *group_counter,
193 struct perf_cpu_context *cpuctx,
194 struct perf_counter_context *ctx)
196 struct perf_counter *counter;
198 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
201 counter_sched_out(group_counter, cpuctx, ctx);
204 * Schedule out siblings (if any):
206 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
207 counter_sched_out(counter, cpuctx, ctx);
209 if (group_counter->hw_event.exclusive)
210 cpuctx->exclusive = 0;
214 * Mark this context as not being a clone of another.
215 * Called when counters are added to or removed from this context.
216 * We also increment our generation number so that anything that
217 * was cloned from this context before this will not match anything
218 * cloned from this context after this.
220 static void unclone_ctx(struct perf_counter_context *ctx)
223 if (!ctx->parent_ctx)
225 put_ctx(ctx->parent_ctx);
226 ctx->parent_ctx = NULL;
230 * Cross CPU call to remove a performance counter
232 * We disable the counter on the hardware level first. After that we
233 * remove it from the context list.
235 static void __perf_counter_remove_from_context(void *info)
237 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
238 struct perf_counter *counter = info;
239 struct perf_counter_context *ctx = counter->ctx;
243 * If this is a task context, we need to check whether it is
244 * the current task context of this cpu. If not it has been
245 * scheduled out before the smp call arrived.
247 if (ctx->task && cpuctx->task_ctx != ctx)
250 spin_lock_irqsave(&ctx->lock, flags);
252 * Protect the list operation against NMI by disabling the
253 * counters on a global level.
257 counter_sched_out(counter, cpuctx, ctx);
259 list_del_counter(counter, ctx);
263 * Allow more per task counters with respect to the
266 cpuctx->max_pertask =
267 min(perf_max_counters - ctx->nr_counters,
268 perf_max_counters - perf_reserved_percpu);
272 spin_unlock_irqrestore(&ctx->lock, flags);
277 * Remove the counter from a task's (or a CPU's) list of counters.
279 * Must be called with counter->mutex and ctx->mutex held.
281 * CPU counters are removed with a smp call. For task counters we only
282 * call when the task is on a CPU.
284 static void perf_counter_remove_from_context(struct perf_counter *counter)
286 struct perf_counter_context *ctx = counter->ctx;
287 struct task_struct *task = ctx->task;
292 * Per cpu counters are removed via an smp call and
293 * the removal is always sucessful.
295 smp_call_function_single(counter->cpu,
296 __perf_counter_remove_from_context,
302 task_oncpu_function_call(task, __perf_counter_remove_from_context,
305 spin_lock_irq(&ctx->lock);
307 * If the context is active we need to retry the smp call.
309 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
310 spin_unlock_irq(&ctx->lock);
315 * The lock prevents that this context is scheduled in so we
316 * can remove the counter safely, if the call above did not
319 if (!list_empty(&counter->list_entry)) {
320 list_del_counter(counter, ctx);
322 spin_unlock_irq(&ctx->lock);
325 static inline u64 perf_clock(void)
327 return cpu_clock(smp_processor_id());
331 * Update the record of the current time in a context.
333 static void update_context_time(struct perf_counter_context *ctx)
335 u64 now = perf_clock();
337 ctx->time += now - ctx->timestamp;
338 ctx->timestamp = now;
342 * Update the total_time_enabled and total_time_running fields for a counter.
344 static void update_counter_times(struct perf_counter *counter)
346 struct perf_counter_context *ctx = counter->ctx;
349 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
352 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
354 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
355 run_end = counter->tstamp_stopped;
359 counter->total_time_running = run_end - counter->tstamp_running;
363 * Update total_time_enabled and total_time_running for all counters in a group.
365 static void update_group_times(struct perf_counter *leader)
367 struct perf_counter *counter;
369 update_counter_times(leader);
370 list_for_each_entry(counter, &leader->sibling_list, list_entry)
371 update_counter_times(counter);
375 * Cross CPU call to disable a performance counter
377 static void __perf_counter_disable(void *info)
379 struct perf_counter *counter = info;
380 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
381 struct perf_counter_context *ctx = counter->ctx;
385 * If this is a per-task counter, need to check whether this
386 * counter's task is the current task on this cpu.
388 if (ctx->task && cpuctx->task_ctx != ctx)
391 spin_lock_irqsave(&ctx->lock, flags);
394 * If the counter is on, turn it off.
395 * If it is in error state, leave it in error state.
397 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
398 update_context_time(ctx);
399 update_counter_times(counter);
400 if (counter == counter->group_leader)
401 group_sched_out(counter, cpuctx, ctx);
403 counter_sched_out(counter, cpuctx, ctx);
404 counter->state = PERF_COUNTER_STATE_OFF;
408 spin_unlock_irqrestore(&ctx->lock, flags);
414 static void perf_counter_disable(struct perf_counter *counter)
416 struct perf_counter_context *ctx = counter->ctx;
417 struct task_struct *task = ctx->task;
421 * Disable the counter on the cpu that it's on
423 smp_call_function_single(counter->cpu, __perf_counter_disable,
429 task_oncpu_function_call(task, __perf_counter_disable, counter);
431 spin_lock_irq(&ctx->lock);
433 * If the counter is still active, we need to retry the cross-call.
435 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
436 spin_unlock_irq(&ctx->lock);
441 * Since we have the lock this context can't be scheduled
442 * in, so we can change the state safely.
444 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
445 update_counter_times(counter);
446 counter->state = PERF_COUNTER_STATE_OFF;
450 spin_unlock_irq(&ctx->lock);
454 counter_sched_in(struct perf_counter *counter,
455 struct perf_cpu_context *cpuctx,
456 struct perf_counter_context *ctx,
459 if (counter->state <= PERF_COUNTER_STATE_OFF)
462 counter->state = PERF_COUNTER_STATE_ACTIVE;
463 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
465 * The new state must be visible before we turn it on in the hardware:
469 if (counter->pmu->enable(counter)) {
470 counter->state = PERF_COUNTER_STATE_INACTIVE;
475 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
477 if (!is_software_counter(counter))
478 cpuctx->active_oncpu++;
481 if (counter->hw_event.exclusive)
482 cpuctx->exclusive = 1;
488 group_sched_in(struct perf_counter *group_counter,
489 struct perf_cpu_context *cpuctx,
490 struct perf_counter_context *ctx,
493 struct perf_counter *counter, *partial_group;
496 if (group_counter->state == PERF_COUNTER_STATE_OFF)
499 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
501 return ret < 0 ? ret : 0;
503 group_counter->prev_state = group_counter->state;
504 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
508 * Schedule in siblings as one group (if any):
510 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
511 counter->prev_state = counter->state;
512 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
513 partial_group = counter;
522 * Groups can be scheduled in as one unit only, so undo any
523 * partial group before returning:
525 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
526 if (counter == partial_group)
528 counter_sched_out(counter, cpuctx, ctx);
530 counter_sched_out(group_counter, cpuctx, ctx);
536 * Return 1 for a group consisting entirely of software counters,
537 * 0 if the group contains any hardware counters.
539 static int is_software_only_group(struct perf_counter *leader)
541 struct perf_counter *counter;
543 if (!is_software_counter(leader))
546 list_for_each_entry(counter, &leader->sibling_list, list_entry)
547 if (!is_software_counter(counter))
554 * Work out whether we can put this counter group on the CPU now.
556 static int group_can_go_on(struct perf_counter *counter,
557 struct perf_cpu_context *cpuctx,
561 * Groups consisting entirely of software counters can always go on.
563 if (is_software_only_group(counter))
566 * If an exclusive group is already on, no other hardware
567 * counters can go on.
569 if (cpuctx->exclusive)
572 * If this group is exclusive and there are already
573 * counters on the CPU, it can't go on.
575 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
578 * Otherwise, try to add it if all previous groups were able
584 static void add_counter_to_ctx(struct perf_counter *counter,
585 struct perf_counter_context *ctx)
587 list_add_counter(counter, ctx);
588 counter->prev_state = PERF_COUNTER_STATE_OFF;
589 counter->tstamp_enabled = ctx->time;
590 counter->tstamp_running = ctx->time;
591 counter->tstamp_stopped = ctx->time;
595 * Cross CPU call to install and enable a performance counter
597 static void __perf_install_in_context(void *info)
599 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
600 struct perf_counter *counter = info;
601 struct perf_counter_context *ctx = counter->ctx;
602 struct perf_counter *leader = counter->group_leader;
603 int cpu = smp_processor_id();
608 * If this is a task context, we need to check whether it is
609 * the current task context of this cpu. If not it has been
610 * scheduled out before the smp call arrived.
611 * Or possibly this is the right context but it isn't
612 * on this cpu because it had no counters.
614 if (ctx->task && cpuctx->task_ctx != ctx) {
615 if (cpuctx->task_ctx || ctx->task != current)
617 cpuctx->task_ctx = ctx;
620 spin_lock_irqsave(&ctx->lock, flags);
622 update_context_time(ctx);
625 * Protect the list operation against NMI by disabling the
626 * counters on a global level. NOP for non NMI based counters.
630 add_counter_to_ctx(counter, ctx);
633 * Don't put the counter on if it is disabled or if
634 * it is in a group and the group isn't on.
636 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
637 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
641 * An exclusive counter can't go on if there are already active
642 * hardware counters, and no hardware counter can go on if there
643 * is already an exclusive counter on.
645 if (!group_can_go_on(counter, cpuctx, 1))
648 err = counter_sched_in(counter, cpuctx, ctx, cpu);
652 * This counter couldn't go on. If it is in a group
653 * then we have to pull the whole group off.
654 * If the counter group is pinned then put it in error state.
656 if (leader != counter)
657 group_sched_out(leader, cpuctx, ctx);
658 if (leader->hw_event.pinned) {
659 update_group_times(leader);
660 leader->state = PERF_COUNTER_STATE_ERROR;
664 if (!err && !ctx->task && cpuctx->max_pertask)
665 cpuctx->max_pertask--;
670 spin_unlock_irqrestore(&ctx->lock, flags);
674 * Attach a performance counter to a context
676 * First we add the counter to the list with the hardware enable bit
677 * in counter->hw_config cleared.
679 * If the counter is attached to a task which is on a CPU we use a smp
680 * call to enable it in the task context. The task might have been
681 * scheduled away, but we check this in the smp call again.
683 * Must be called with ctx->mutex held.
686 perf_install_in_context(struct perf_counter_context *ctx,
687 struct perf_counter *counter,
690 struct task_struct *task = ctx->task;
694 * Per cpu counters are installed via an smp call and
695 * the install is always sucessful.
697 smp_call_function_single(cpu, __perf_install_in_context,
703 task_oncpu_function_call(task, __perf_install_in_context,
706 spin_lock_irq(&ctx->lock);
708 * we need to retry the smp call.
710 if (ctx->is_active && list_empty(&counter->list_entry)) {
711 spin_unlock_irq(&ctx->lock);
716 * The lock prevents that this context is scheduled in so we
717 * can add the counter safely, if it the call above did not
720 if (list_empty(&counter->list_entry))
721 add_counter_to_ctx(counter, ctx);
722 spin_unlock_irq(&ctx->lock);
726 * Cross CPU call to enable a performance counter
728 static void __perf_counter_enable(void *info)
730 struct perf_counter *counter = info;
731 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
732 struct perf_counter_context *ctx = counter->ctx;
733 struct perf_counter *leader = counter->group_leader;
738 * If this is a per-task counter, need to check whether this
739 * counter's task is the current task on this cpu.
741 if (ctx->task && cpuctx->task_ctx != ctx) {
742 if (cpuctx->task_ctx || ctx->task != current)
744 cpuctx->task_ctx = ctx;
747 spin_lock_irqsave(&ctx->lock, flags);
749 update_context_time(ctx);
751 counter->prev_state = counter->state;
752 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
754 counter->state = PERF_COUNTER_STATE_INACTIVE;
755 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
759 * If the counter is in a group and isn't the group leader,
760 * then don't put it on unless the group is on.
762 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
765 if (!group_can_go_on(counter, cpuctx, 1)) {
769 if (counter == leader)
770 err = group_sched_in(counter, cpuctx, ctx,
773 err = counter_sched_in(counter, cpuctx, ctx,
780 * If this counter can't go on and it's part of a
781 * group, then the whole group has to come off.
783 if (leader != counter)
784 group_sched_out(leader, cpuctx, ctx);
785 if (leader->hw_event.pinned) {
786 update_group_times(leader);
787 leader->state = PERF_COUNTER_STATE_ERROR;
792 spin_unlock_irqrestore(&ctx->lock, flags);
798 static void perf_counter_enable(struct perf_counter *counter)
800 struct perf_counter_context *ctx = counter->ctx;
801 struct task_struct *task = ctx->task;
805 * Enable the counter on the cpu that it's on
807 smp_call_function_single(counter->cpu, __perf_counter_enable,
812 spin_lock_irq(&ctx->lock);
813 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
817 * If the counter is in error state, clear that first.
818 * That way, if we see the counter in error state below, we
819 * know that it has gone back into error state, as distinct
820 * from the task having been scheduled away before the
821 * cross-call arrived.
823 if (counter->state == PERF_COUNTER_STATE_ERROR)
824 counter->state = PERF_COUNTER_STATE_OFF;
827 spin_unlock_irq(&ctx->lock);
828 task_oncpu_function_call(task, __perf_counter_enable, counter);
830 spin_lock_irq(&ctx->lock);
833 * If the context is active and the counter is still off,
834 * we need to retry the cross-call.
836 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
840 * Since we have the lock this context can't be scheduled
841 * in, so we can change the state safely.
843 if (counter->state == PERF_COUNTER_STATE_OFF) {
844 counter->state = PERF_COUNTER_STATE_INACTIVE;
845 counter->tstamp_enabled =
846 ctx->time - counter->total_time_enabled;
850 spin_unlock_irq(&ctx->lock);
853 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
856 * not supported on inherited counters
858 if (counter->hw_event.inherit)
861 atomic_add(refresh, &counter->event_limit);
862 perf_counter_enable(counter);
867 void __perf_counter_sched_out(struct perf_counter_context *ctx,
868 struct perf_cpu_context *cpuctx)
870 struct perf_counter *counter;
872 spin_lock(&ctx->lock);
874 if (likely(!ctx->nr_counters))
876 update_context_time(ctx);
879 if (ctx->nr_active) {
880 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
881 if (counter != counter->group_leader)
882 counter_sched_out(counter, cpuctx, ctx);
884 group_sched_out(counter, cpuctx, ctx);
889 spin_unlock(&ctx->lock);
893 * Test whether two contexts are equivalent, i.e. whether they
894 * have both been cloned from the same version of the same context
895 * and they both have the same number of enabled counters.
896 * If the number of enabled counters is the same, then the set
897 * of enabled counters should be the same, because these are both
898 * inherited contexts, therefore we can't access individual counters
899 * in them directly with an fd; we can only enable/disable all
900 * counters via prctl, or enable/disable all counters in a family
901 * via ioctl, which will have the same effect on both contexts.
903 static int context_equiv(struct perf_counter_context *ctx1,
904 struct perf_counter_context *ctx2)
906 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
907 && ctx1->parent_gen == ctx2->parent_gen
908 && ctx1->nr_enabled == ctx2->nr_enabled;
912 * Called from scheduler to remove the counters of the current task,
913 * with interrupts disabled.
915 * We stop each counter and update the counter value in counter->count.
917 * This does not protect us against NMI, but disable()
918 * sets the disabled bit in the control field of counter _before_
919 * accessing the counter control register. If a NMI hits, then it will
920 * not restart the counter.
922 void perf_counter_task_sched_out(struct task_struct *task,
923 struct task_struct *next, int cpu)
925 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
926 struct perf_counter_context *ctx = task->perf_counter_ctxp;
927 struct perf_counter_context *next_ctx;
928 struct pt_regs *regs;
930 if (likely(!ctx || !cpuctx->task_ctx))
933 update_context_time(ctx);
935 regs = task_pt_regs(task);
936 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
938 next_ctx = next->perf_counter_ctxp;
939 if (next_ctx && context_equiv(ctx, next_ctx)) {
940 task->perf_counter_ctxp = next_ctx;
941 next->perf_counter_ctxp = ctx;
943 next_ctx->task = task;
947 __perf_counter_sched_out(ctx, cpuctx);
949 cpuctx->task_ctx = NULL;
952 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
954 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
956 if (!cpuctx->task_ctx)
958 __perf_counter_sched_out(ctx, cpuctx);
959 cpuctx->task_ctx = NULL;
962 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
964 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
968 __perf_counter_sched_in(struct perf_counter_context *ctx,
969 struct perf_cpu_context *cpuctx, int cpu)
971 struct perf_counter *counter;
974 spin_lock(&ctx->lock);
976 if (likely(!ctx->nr_counters))
979 ctx->timestamp = perf_clock();
984 * First go through the list and put on any pinned groups
985 * in order to give them the best chance of going on.
987 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
988 if (counter->state <= PERF_COUNTER_STATE_OFF ||
989 !counter->hw_event.pinned)
991 if (counter->cpu != -1 && counter->cpu != cpu)
994 if (counter != counter->group_leader)
995 counter_sched_in(counter, cpuctx, ctx, cpu);
997 if (group_can_go_on(counter, cpuctx, 1))
998 group_sched_in(counter, cpuctx, ctx, cpu);
1002 * If this pinned group hasn't been scheduled,
1003 * put it in error state.
1005 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1006 update_group_times(counter);
1007 counter->state = PERF_COUNTER_STATE_ERROR;
1011 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1013 * Ignore counters in OFF or ERROR state, and
1014 * ignore pinned counters since we did them already.
1016 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1017 counter->hw_event.pinned)
1021 * Listen to the 'cpu' scheduling filter constraint
1024 if (counter->cpu != -1 && counter->cpu != cpu)
1027 if (counter != counter->group_leader) {
1028 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1031 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1032 if (group_sched_in(counter, cpuctx, ctx, cpu))
1039 spin_unlock(&ctx->lock);
1043 * Called from scheduler to add the counters of the current task
1044 * with interrupts disabled.
1046 * We restore the counter value and then enable it.
1048 * This does not protect us against NMI, but enable()
1049 * sets the enabled bit in the control field of counter _before_
1050 * accessing the counter control register. If a NMI hits, then it will
1051 * keep the counter running.
1053 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1055 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1056 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1060 if (cpuctx->task_ctx == ctx)
1062 __perf_counter_sched_in(ctx, cpuctx, cpu);
1063 cpuctx->task_ctx = ctx;
1066 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1068 struct perf_counter_context *ctx = &cpuctx->ctx;
1070 __perf_counter_sched_in(ctx, cpuctx, cpu);
1073 int perf_counter_task_disable(void)
1075 struct task_struct *curr = current;
1076 struct perf_counter_context *ctx = curr->perf_counter_ctxp;
1077 struct perf_counter *counter;
1078 unsigned long flags;
1080 if (!ctx || !ctx->nr_counters)
1083 local_irq_save(flags);
1085 __perf_counter_task_sched_out(ctx);
1087 spin_lock(&ctx->lock);
1090 * Disable all the counters:
1094 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1095 if (counter->state != PERF_COUNTER_STATE_ERROR) {
1096 update_group_times(counter);
1097 counter->state = PERF_COUNTER_STATE_OFF;
1103 spin_unlock_irqrestore(&ctx->lock, flags);
1108 int perf_counter_task_enable(void)
1110 struct task_struct *curr = current;
1111 struct perf_counter_context *ctx = curr->perf_counter_ctxp;
1112 struct perf_counter *counter;
1113 unsigned long flags;
1116 if (!ctx || !ctx->nr_counters)
1119 local_irq_save(flags);
1120 cpu = smp_processor_id();
1122 __perf_counter_task_sched_out(ctx);
1124 spin_lock(&ctx->lock);
1127 * Disable all the counters:
1131 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1132 if (counter->state > PERF_COUNTER_STATE_OFF)
1134 counter->state = PERF_COUNTER_STATE_INACTIVE;
1135 counter->tstamp_enabled =
1136 ctx->time - counter->total_time_enabled;
1137 counter->hw_event.disabled = 0;
1141 spin_unlock(&ctx->lock);
1143 perf_counter_task_sched_in(curr, cpu);
1145 local_irq_restore(flags);
1150 static void perf_log_period(struct perf_counter *counter, u64 period);
1152 static void perf_adjust_freq(struct perf_counter_context *ctx)
1154 struct perf_counter *counter;
1159 spin_lock(&ctx->lock);
1160 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1161 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1164 if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
1167 events = HZ * counter->hw.interrupts * counter->hw.irq_period;
1168 period = div64_u64(events, counter->hw_event.irq_freq);
1170 delta = (s64)(1 + period - counter->hw.irq_period);
1173 irq_period = counter->hw.irq_period + delta;
1178 perf_log_period(counter, irq_period);
1180 counter->hw.irq_period = irq_period;
1181 counter->hw.interrupts = 0;
1183 spin_unlock(&ctx->lock);
1187 * Round-robin a context's counters:
1189 static void rotate_ctx(struct perf_counter_context *ctx)
1191 struct perf_counter *counter;
1193 if (!ctx->nr_counters)
1196 spin_lock(&ctx->lock);
1198 * Rotate the first entry last (works just fine for group counters too):
1201 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1202 list_move_tail(&counter->list_entry, &ctx->counter_list);
1207 spin_unlock(&ctx->lock);
1210 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1212 struct perf_cpu_context *cpuctx;
1213 struct perf_counter_context *ctx;
1215 if (!atomic_read(&nr_counters))
1218 cpuctx = &per_cpu(perf_cpu_context, cpu);
1219 ctx = curr->perf_counter_ctxp;
1221 perf_adjust_freq(&cpuctx->ctx);
1223 perf_adjust_freq(ctx);
1225 perf_counter_cpu_sched_out(cpuctx);
1227 __perf_counter_task_sched_out(ctx);
1229 rotate_ctx(&cpuctx->ctx);
1233 perf_counter_cpu_sched_in(cpuctx, cpu);
1235 perf_counter_task_sched_in(curr, cpu);
1239 * Cross CPU call to read the hardware counter
1241 static void __read(void *info)
1243 struct perf_counter *counter = info;
1244 struct perf_counter_context *ctx = counter->ctx;
1245 unsigned long flags;
1247 local_irq_save(flags);
1249 update_context_time(ctx);
1250 counter->pmu->read(counter);
1251 update_counter_times(counter);
1252 local_irq_restore(flags);
1255 static u64 perf_counter_read(struct perf_counter *counter)
1258 * If counter is enabled and currently active on a CPU, update the
1259 * value in the counter structure:
1261 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1262 smp_call_function_single(counter->oncpu,
1263 __read, counter, 1);
1264 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1265 update_counter_times(counter);
1268 return atomic64_read(&counter->count);
1272 * Initialize the perf_counter context in a task_struct:
1275 __perf_counter_init_context(struct perf_counter_context *ctx,
1276 struct task_struct *task)
1278 memset(ctx, 0, sizeof(*ctx));
1279 spin_lock_init(&ctx->lock);
1280 mutex_init(&ctx->mutex);
1281 INIT_LIST_HEAD(&ctx->counter_list);
1282 INIT_LIST_HEAD(&ctx->event_list);
1283 atomic_set(&ctx->refcount, 1);
1287 static void put_context(struct perf_counter_context *ctx)
1290 put_task_struct(ctx->task);
1293 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1295 struct perf_cpu_context *cpuctx;
1296 struct perf_counter_context *ctx;
1297 struct perf_counter_context *tctx;
1298 struct task_struct *task;
1301 * If cpu is not a wildcard then this is a percpu counter:
1304 /* Must be root to operate on a CPU counter: */
1305 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1306 return ERR_PTR(-EACCES);
1308 if (cpu < 0 || cpu > num_possible_cpus())
1309 return ERR_PTR(-EINVAL);
1312 * We could be clever and allow to attach a counter to an
1313 * offline CPU and activate it when the CPU comes up, but
1316 if (!cpu_isset(cpu, cpu_online_map))
1317 return ERR_PTR(-ENODEV);
1319 cpuctx = &per_cpu(perf_cpu_context, cpu);
1329 task = find_task_by_vpid(pid);
1331 get_task_struct(task);
1335 return ERR_PTR(-ESRCH);
1337 /* Reuse ptrace permission checks for now. */
1338 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1339 put_task_struct(task);
1340 return ERR_PTR(-EACCES);
1343 ctx = task->perf_counter_ctxp;
1345 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1347 put_task_struct(task);
1348 return ERR_PTR(-ENOMEM);
1350 __perf_counter_init_context(ctx, task);
1352 * Make sure other cpus see correct values for *ctx
1353 * once task->perf_counter_ctxp is visible to them.
1356 tctx = cmpxchg(&task->perf_counter_ctxp, NULL, ctx);
1359 * We raced with some other task; use
1360 * the context they set.
1370 static void free_counter_rcu(struct rcu_head *head)
1372 struct perf_counter *counter;
1374 counter = container_of(head, struct perf_counter, rcu_head);
1375 put_ctx(counter->ctx);
1379 static void perf_pending_sync(struct perf_counter *counter);
1381 static void free_counter(struct perf_counter *counter)
1383 perf_pending_sync(counter);
1385 atomic_dec(&nr_counters);
1386 if (counter->hw_event.mmap)
1387 atomic_dec(&nr_mmap_tracking);
1388 if (counter->hw_event.munmap)
1389 atomic_dec(&nr_munmap_tracking);
1390 if (counter->hw_event.comm)
1391 atomic_dec(&nr_comm_tracking);
1393 if (counter->destroy)
1394 counter->destroy(counter);
1396 call_rcu(&counter->rcu_head, free_counter_rcu);
1400 * Called when the last reference to the file is gone.
1402 static int perf_release(struct inode *inode, struct file *file)
1404 struct perf_counter *counter = file->private_data;
1405 struct perf_counter_context *ctx = counter->ctx;
1407 file->private_data = NULL;
1409 mutex_lock(&ctx->mutex);
1410 mutex_lock(&counter->mutex);
1412 perf_counter_remove_from_context(counter);
1414 mutex_unlock(&counter->mutex);
1415 mutex_unlock(&ctx->mutex);
1417 free_counter(counter);
1424 * Read the performance counter - simple non blocking version for now
1427 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1433 * Return end-of-file for a read on a counter that is in
1434 * error state (i.e. because it was pinned but it couldn't be
1435 * scheduled on to the CPU at some point).
1437 if (counter->state == PERF_COUNTER_STATE_ERROR)
1440 mutex_lock(&counter->mutex);
1441 values[0] = perf_counter_read(counter);
1443 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1444 values[n++] = counter->total_time_enabled +
1445 atomic64_read(&counter->child_total_time_enabled);
1446 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1447 values[n++] = counter->total_time_running +
1448 atomic64_read(&counter->child_total_time_running);
1449 mutex_unlock(&counter->mutex);
1451 if (count < n * sizeof(u64))
1453 count = n * sizeof(u64);
1455 if (copy_to_user(buf, values, count))
1462 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1464 struct perf_counter *counter = file->private_data;
1466 return perf_read_hw(counter, buf, count);
1469 static unsigned int perf_poll(struct file *file, poll_table *wait)
1471 struct perf_counter *counter = file->private_data;
1472 struct perf_mmap_data *data;
1473 unsigned int events = POLL_HUP;
1476 data = rcu_dereference(counter->data);
1478 events = atomic_xchg(&data->poll, 0);
1481 poll_wait(file, &counter->waitq, wait);
1486 static void perf_counter_reset(struct perf_counter *counter)
1488 (void)perf_counter_read(counter);
1489 atomic64_set(&counter->count, 0);
1490 perf_counter_update_userpage(counter);
1493 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1494 void (*func)(struct perf_counter *))
1496 struct perf_counter_context *ctx = counter->ctx;
1497 struct perf_counter *sibling;
1499 spin_lock_irq(&ctx->lock);
1500 counter = counter->group_leader;
1503 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1505 spin_unlock_irq(&ctx->lock);
1508 static void perf_counter_for_each_child(struct perf_counter *counter,
1509 void (*func)(struct perf_counter *))
1511 struct perf_counter *child;
1513 mutex_lock(&counter->mutex);
1515 list_for_each_entry(child, &counter->child_list, child_list)
1517 mutex_unlock(&counter->mutex);
1520 static void perf_counter_for_each(struct perf_counter *counter,
1521 void (*func)(struct perf_counter *))
1523 struct perf_counter *child;
1525 mutex_lock(&counter->mutex);
1526 perf_counter_for_each_sibling(counter, func);
1527 list_for_each_entry(child, &counter->child_list, child_list)
1528 perf_counter_for_each_sibling(child, func);
1529 mutex_unlock(&counter->mutex);
1532 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1534 struct perf_counter *counter = file->private_data;
1535 void (*func)(struct perf_counter *);
1539 case PERF_COUNTER_IOC_ENABLE:
1540 func = perf_counter_enable;
1542 case PERF_COUNTER_IOC_DISABLE:
1543 func = perf_counter_disable;
1545 case PERF_COUNTER_IOC_RESET:
1546 func = perf_counter_reset;
1549 case PERF_COUNTER_IOC_REFRESH:
1550 return perf_counter_refresh(counter, arg);
1555 if (flags & PERF_IOC_FLAG_GROUP)
1556 perf_counter_for_each(counter, func);
1558 perf_counter_for_each_child(counter, func);
1564 * Callers need to ensure there can be no nesting of this function, otherwise
1565 * the seqlock logic goes bad. We can not serialize this because the arch
1566 * code calls this from NMI context.
1568 void perf_counter_update_userpage(struct perf_counter *counter)
1570 struct perf_mmap_data *data;
1571 struct perf_counter_mmap_page *userpg;
1574 data = rcu_dereference(counter->data);
1578 userpg = data->user_page;
1581 * Disable preemption so as to not let the corresponding user-space
1582 * spin too long if we get preempted.
1587 userpg->index = counter->hw.idx;
1588 userpg->offset = atomic64_read(&counter->count);
1589 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1590 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1599 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1601 struct perf_counter *counter = vma->vm_file->private_data;
1602 struct perf_mmap_data *data;
1603 int ret = VM_FAULT_SIGBUS;
1606 data = rcu_dereference(counter->data);
1610 if (vmf->pgoff == 0) {
1611 vmf->page = virt_to_page(data->user_page);
1613 int nr = vmf->pgoff - 1;
1615 if ((unsigned)nr > data->nr_pages)
1618 vmf->page = virt_to_page(data->data_pages[nr]);
1620 get_page(vmf->page);
1628 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1630 struct perf_mmap_data *data;
1634 WARN_ON(atomic_read(&counter->mmap_count));
1636 size = sizeof(struct perf_mmap_data);
1637 size += nr_pages * sizeof(void *);
1639 data = kzalloc(size, GFP_KERNEL);
1643 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1644 if (!data->user_page)
1645 goto fail_user_page;
1647 for (i = 0; i < nr_pages; i++) {
1648 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1649 if (!data->data_pages[i])
1650 goto fail_data_pages;
1653 data->nr_pages = nr_pages;
1654 atomic_set(&data->lock, -1);
1656 rcu_assign_pointer(counter->data, data);
1661 for (i--; i >= 0; i--)
1662 free_page((unsigned long)data->data_pages[i]);
1664 free_page((unsigned long)data->user_page);
1673 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1675 struct perf_mmap_data *data = container_of(rcu_head,
1676 struct perf_mmap_data, rcu_head);
1679 free_page((unsigned long)data->user_page);
1680 for (i = 0; i < data->nr_pages; i++)
1681 free_page((unsigned long)data->data_pages[i]);
1685 static void perf_mmap_data_free(struct perf_counter *counter)
1687 struct perf_mmap_data *data = counter->data;
1689 WARN_ON(atomic_read(&counter->mmap_count));
1691 rcu_assign_pointer(counter->data, NULL);
1692 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1695 static void perf_mmap_open(struct vm_area_struct *vma)
1697 struct perf_counter *counter = vma->vm_file->private_data;
1699 atomic_inc(&counter->mmap_count);
1702 static void perf_mmap_close(struct vm_area_struct *vma)
1704 struct perf_counter *counter = vma->vm_file->private_data;
1706 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1707 &counter->mmap_mutex)) {
1708 struct user_struct *user = current_user();
1710 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
1711 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1712 perf_mmap_data_free(counter);
1713 mutex_unlock(&counter->mmap_mutex);
1717 static struct vm_operations_struct perf_mmap_vmops = {
1718 .open = perf_mmap_open,
1719 .close = perf_mmap_close,
1720 .fault = perf_mmap_fault,
1723 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1725 struct perf_counter *counter = file->private_data;
1726 struct user_struct *user = current_user();
1727 unsigned long vma_size;
1728 unsigned long nr_pages;
1729 unsigned long user_locked, user_lock_limit;
1730 unsigned long locked, lock_limit;
1731 long user_extra, extra;
1734 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1737 vma_size = vma->vm_end - vma->vm_start;
1738 nr_pages = (vma_size / PAGE_SIZE) - 1;
1741 * If we have data pages ensure they're a power-of-two number, so we
1742 * can do bitmasks instead of modulo.
1744 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1747 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1750 if (vma->vm_pgoff != 0)
1753 mutex_lock(&counter->mmap_mutex);
1754 if (atomic_inc_not_zero(&counter->mmap_count)) {
1755 if (nr_pages != counter->data->nr_pages)
1760 user_extra = nr_pages + 1;
1761 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1762 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
1765 if (user_locked > user_lock_limit)
1766 extra = user_locked - user_lock_limit;
1768 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1769 lock_limit >>= PAGE_SHIFT;
1770 locked = vma->vm_mm->locked_vm + extra;
1772 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1777 WARN_ON(counter->data);
1778 ret = perf_mmap_data_alloc(counter, nr_pages);
1782 atomic_set(&counter->mmap_count, 1);
1783 atomic_long_add(user_extra, &user->locked_vm);
1784 vma->vm_mm->locked_vm += extra;
1785 counter->data->nr_locked = extra;
1787 mutex_unlock(&counter->mmap_mutex);
1789 vma->vm_flags &= ~VM_MAYWRITE;
1790 vma->vm_flags |= VM_RESERVED;
1791 vma->vm_ops = &perf_mmap_vmops;
1796 static int perf_fasync(int fd, struct file *filp, int on)
1798 struct perf_counter *counter = filp->private_data;
1799 struct inode *inode = filp->f_path.dentry->d_inode;
1802 mutex_lock(&inode->i_mutex);
1803 retval = fasync_helper(fd, filp, on, &counter->fasync);
1804 mutex_unlock(&inode->i_mutex);
1812 static const struct file_operations perf_fops = {
1813 .release = perf_release,
1816 .unlocked_ioctl = perf_ioctl,
1817 .compat_ioctl = perf_ioctl,
1819 .fasync = perf_fasync,
1823 * Perf counter wakeup
1825 * If there's data, ensure we set the poll() state and publish everything
1826 * to user-space before waking everybody up.
1829 void perf_counter_wakeup(struct perf_counter *counter)
1831 wake_up_all(&counter->waitq);
1833 if (counter->pending_kill) {
1834 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1835 counter->pending_kill = 0;
1842 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1844 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1845 * single linked list and use cmpxchg() to add entries lockless.
1848 static void perf_pending_counter(struct perf_pending_entry *entry)
1850 struct perf_counter *counter = container_of(entry,
1851 struct perf_counter, pending);
1853 if (counter->pending_disable) {
1854 counter->pending_disable = 0;
1855 perf_counter_disable(counter);
1858 if (counter->pending_wakeup) {
1859 counter->pending_wakeup = 0;
1860 perf_counter_wakeup(counter);
1864 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1866 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1870 static void perf_pending_queue(struct perf_pending_entry *entry,
1871 void (*func)(struct perf_pending_entry *))
1873 struct perf_pending_entry **head;
1875 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1880 head = &get_cpu_var(perf_pending_head);
1883 entry->next = *head;
1884 } while (cmpxchg(head, entry->next, entry) != entry->next);
1886 set_perf_counter_pending();
1888 put_cpu_var(perf_pending_head);
1891 static int __perf_pending_run(void)
1893 struct perf_pending_entry *list;
1896 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1897 while (list != PENDING_TAIL) {
1898 void (*func)(struct perf_pending_entry *);
1899 struct perf_pending_entry *entry = list;
1906 * Ensure we observe the unqueue before we issue the wakeup,
1907 * so that we won't be waiting forever.
1908 * -- see perf_not_pending().
1919 static inline int perf_not_pending(struct perf_counter *counter)
1922 * If we flush on whatever cpu we run, there is a chance we don't
1926 __perf_pending_run();
1930 * Ensure we see the proper queue state before going to sleep
1931 * so that we do not miss the wakeup. -- see perf_pending_handle()
1934 return counter->pending.next == NULL;
1937 static void perf_pending_sync(struct perf_counter *counter)
1939 wait_event(counter->waitq, perf_not_pending(counter));
1942 void perf_counter_do_pending(void)
1944 __perf_pending_run();
1948 * Callchain support -- arch specific
1951 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1960 struct perf_output_handle {
1961 struct perf_counter *counter;
1962 struct perf_mmap_data *data;
1963 unsigned int offset;
1968 unsigned long flags;
1971 static void perf_output_wakeup(struct perf_output_handle *handle)
1973 atomic_set(&handle->data->poll, POLL_IN);
1976 handle->counter->pending_wakeup = 1;
1977 perf_pending_queue(&handle->counter->pending,
1978 perf_pending_counter);
1980 perf_counter_wakeup(handle->counter);
1984 * Curious locking construct.
1986 * We need to ensure a later event doesn't publish a head when a former
1987 * event isn't done writing. However since we need to deal with NMIs we
1988 * cannot fully serialize things.
1990 * What we do is serialize between CPUs so we only have to deal with NMI
1991 * nesting on a single CPU.
1993 * We only publish the head (and generate a wakeup) when the outer-most
1996 static void perf_output_lock(struct perf_output_handle *handle)
1998 struct perf_mmap_data *data = handle->data;
2003 local_irq_save(handle->flags);
2004 cpu = smp_processor_id();
2006 if (in_nmi() && atomic_read(&data->lock) == cpu)
2009 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2015 static void perf_output_unlock(struct perf_output_handle *handle)
2017 struct perf_mmap_data *data = handle->data;
2020 data->done_head = data->head;
2022 if (!handle->locked)
2027 * The xchg implies a full barrier that ensures all writes are done
2028 * before we publish the new head, matched by a rmb() in userspace when
2029 * reading this position.
2031 while ((head = atomic_xchg(&data->done_head, 0)))
2032 data->user_page->data_head = head;
2035 * NMI can happen here, which means we can miss a done_head update.
2038 cpu = atomic_xchg(&data->lock, -1);
2039 WARN_ON_ONCE(cpu != smp_processor_id());
2042 * Therefore we have to validate we did not indeed do so.
2044 if (unlikely(atomic_read(&data->done_head))) {
2046 * Since we had it locked, we can lock it again.
2048 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2054 if (atomic_xchg(&data->wakeup, 0))
2055 perf_output_wakeup(handle);
2057 local_irq_restore(handle->flags);
2060 static int perf_output_begin(struct perf_output_handle *handle,
2061 struct perf_counter *counter, unsigned int size,
2062 int nmi, int overflow)
2064 struct perf_mmap_data *data;
2065 unsigned int offset, head;
2068 * For inherited counters we send all the output towards the parent.
2070 if (counter->parent)
2071 counter = counter->parent;
2074 data = rcu_dereference(counter->data);
2078 handle->data = data;
2079 handle->counter = counter;
2081 handle->overflow = overflow;
2083 if (!data->nr_pages)
2086 perf_output_lock(handle);
2089 offset = head = atomic_read(&data->head);
2091 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
2093 handle->offset = offset;
2094 handle->head = head;
2096 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2097 atomic_set(&data->wakeup, 1);
2102 perf_output_wakeup(handle);
2109 static void perf_output_copy(struct perf_output_handle *handle,
2110 void *buf, unsigned int len)
2112 unsigned int pages_mask;
2113 unsigned int offset;
2117 offset = handle->offset;
2118 pages_mask = handle->data->nr_pages - 1;
2119 pages = handle->data->data_pages;
2122 unsigned int page_offset;
2125 nr = (offset >> PAGE_SHIFT) & pages_mask;
2126 page_offset = offset & (PAGE_SIZE - 1);
2127 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2129 memcpy(pages[nr] + page_offset, buf, size);
2136 handle->offset = offset;
2139 * Check we didn't copy past our reservation window, taking the
2140 * possible unsigned int wrap into account.
2142 WARN_ON_ONCE(((int)(handle->head - handle->offset)) < 0);
2145 #define perf_output_put(handle, x) \
2146 perf_output_copy((handle), &(x), sizeof(x))
2148 static void perf_output_end(struct perf_output_handle *handle)
2150 struct perf_counter *counter = handle->counter;
2151 struct perf_mmap_data *data = handle->data;
2153 int wakeup_events = counter->hw_event.wakeup_events;
2155 if (handle->overflow && wakeup_events) {
2156 int events = atomic_inc_return(&data->events);
2157 if (events >= wakeup_events) {
2158 atomic_sub(wakeup_events, &data->events);
2159 atomic_set(&data->wakeup, 1);
2163 perf_output_unlock(handle);
2167 static void perf_counter_output(struct perf_counter *counter,
2168 int nmi, struct pt_regs *regs, u64 addr)
2171 u64 record_type = counter->hw_event.record_type;
2172 struct perf_output_handle handle;
2173 struct perf_event_header header;
2182 struct perf_callchain_entry *callchain = NULL;
2183 int callchain_size = 0;
2190 header.size = sizeof(header);
2192 header.misc = PERF_EVENT_MISC_OVERFLOW;
2193 header.misc |= perf_misc_flags(regs);
2195 if (record_type & PERF_RECORD_IP) {
2196 ip = perf_instruction_pointer(regs);
2197 header.type |= PERF_RECORD_IP;
2198 header.size += sizeof(ip);
2201 if (record_type & PERF_RECORD_TID) {
2202 /* namespace issues */
2203 tid_entry.pid = current->group_leader->pid;
2204 tid_entry.tid = current->pid;
2206 header.type |= PERF_RECORD_TID;
2207 header.size += sizeof(tid_entry);
2210 if (record_type & PERF_RECORD_TIME) {
2212 * Maybe do better on x86 and provide cpu_clock_nmi()
2214 time = sched_clock();
2216 header.type |= PERF_RECORD_TIME;
2217 header.size += sizeof(u64);
2220 if (record_type & PERF_RECORD_ADDR) {
2221 header.type |= PERF_RECORD_ADDR;
2222 header.size += sizeof(u64);
2225 if (record_type & PERF_RECORD_CONFIG) {
2226 header.type |= PERF_RECORD_CONFIG;
2227 header.size += sizeof(u64);
2230 if (record_type & PERF_RECORD_CPU) {
2231 header.type |= PERF_RECORD_CPU;
2232 header.size += sizeof(cpu_entry);
2234 cpu_entry.cpu = raw_smp_processor_id();
2237 if (record_type & PERF_RECORD_GROUP) {
2238 header.type |= PERF_RECORD_GROUP;
2239 header.size += sizeof(u64) +
2240 counter->nr_siblings * sizeof(group_entry);
2243 if (record_type & PERF_RECORD_CALLCHAIN) {
2244 callchain = perf_callchain(regs);
2247 callchain_size = (1 + callchain->nr) * sizeof(u64);
2249 header.type |= PERF_RECORD_CALLCHAIN;
2250 header.size += callchain_size;
2254 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2258 perf_output_put(&handle, header);
2260 if (record_type & PERF_RECORD_IP)
2261 perf_output_put(&handle, ip);
2263 if (record_type & PERF_RECORD_TID)
2264 perf_output_put(&handle, tid_entry);
2266 if (record_type & PERF_RECORD_TIME)
2267 perf_output_put(&handle, time);
2269 if (record_type & PERF_RECORD_ADDR)
2270 perf_output_put(&handle, addr);
2272 if (record_type & PERF_RECORD_CONFIG)
2273 perf_output_put(&handle, counter->hw_event.config);
2275 if (record_type & PERF_RECORD_CPU)
2276 perf_output_put(&handle, cpu_entry);
2279 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2281 if (record_type & PERF_RECORD_GROUP) {
2282 struct perf_counter *leader, *sub;
2283 u64 nr = counter->nr_siblings;
2285 perf_output_put(&handle, nr);
2287 leader = counter->group_leader;
2288 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2290 sub->pmu->read(sub);
2292 group_entry.event = sub->hw_event.config;
2293 group_entry.counter = atomic64_read(&sub->count);
2295 perf_output_put(&handle, group_entry);
2300 perf_output_copy(&handle, callchain, callchain_size);
2302 perf_output_end(&handle);
2309 struct perf_comm_event {
2310 struct task_struct *task;
2315 struct perf_event_header header;
2322 static void perf_counter_comm_output(struct perf_counter *counter,
2323 struct perf_comm_event *comm_event)
2325 struct perf_output_handle handle;
2326 int size = comm_event->event.header.size;
2327 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2332 perf_output_put(&handle, comm_event->event);
2333 perf_output_copy(&handle, comm_event->comm,
2334 comm_event->comm_size);
2335 perf_output_end(&handle);
2338 static int perf_counter_comm_match(struct perf_counter *counter,
2339 struct perf_comm_event *comm_event)
2341 if (counter->hw_event.comm &&
2342 comm_event->event.header.type == PERF_EVENT_COMM)
2348 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2349 struct perf_comm_event *comm_event)
2351 struct perf_counter *counter;
2353 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2357 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2358 if (perf_counter_comm_match(counter, comm_event))
2359 perf_counter_comm_output(counter, comm_event);
2364 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2366 struct perf_cpu_context *cpuctx;
2368 char *comm = comm_event->task->comm;
2370 size = ALIGN(strlen(comm)+1, sizeof(u64));
2372 comm_event->comm = comm;
2373 comm_event->comm_size = size;
2375 comm_event->event.header.size = sizeof(comm_event->event) + size;
2377 cpuctx = &get_cpu_var(perf_cpu_context);
2378 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2379 put_cpu_var(perf_cpu_context);
2381 perf_counter_comm_ctx(current->perf_counter_ctxp, comm_event);
2384 void perf_counter_comm(struct task_struct *task)
2386 struct perf_comm_event comm_event;
2388 if (!atomic_read(&nr_comm_tracking))
2390 if (!current->perf_counter_ctxp)
2393 comm_event = (struct perf_comm_event){
2396 .header = { .type = PERF_EVENT_COMM, },
2397 .pid = task->group_leader->pid,
2402 perf_counter_comm_event(&comm_event);
2409 struct perf_mmap_event {
2415 struct perf_event_header header;
2425 static void perf_counter_mmap_output(struct perf_counter *counter,
2426 struct perf_mmap_event *mmap_event)
2428 struct perf_output_handle handle;
2429 int size = mmap_event->event.header.size;
2430 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2435 perf_output_put(&handle, mmap_event->event);
2436 perf_output_copy(&handle, mmap_event->file_name,
2437 mmap_event->file_size);
2438 perf_output_end(&handle);
2441 static int perf_counter_mmap_match(struct perf_counter *counter,
2442 struct perf_mmap_event *mmap_event)
2444 if (counter->hw_event.mmap &&
2445 mmap_event->event.header.type == PERF_EVENT_MMAP)
2448 if (counter->hw_event.munmap &&
2449 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2455 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2456 struct perf_mmap_event *mmap_event)
2458 struct perf_counter *counter;
2460 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2464 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2465 if (perf_counter_mmap_match(counter, mmap_event))
2466 perf_counter_mmap_output(counter, mmap_event);
2471 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2473 struct perf_cpu_context *cpuctx;
2474 struct file *file = mmap_event->file;
2481 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2483 name = strncpy(tmp, "//enomem", sizeof(tmp));
2486 name = d_path(&file->f_path, buf, PATH_MAX);
2488 name = strncpy(tmp, "//toolong", sizeof(tmp));
2492 name = strncpy(tmp, "//anon", sizeof(tmp));
2497 size = ALIGN(strlen(name)+1, sizeof(u64));
2499 mmap_event->file_name = name;
2500 mmap_event->file_size = size;
2502 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2504 cpuctx = &get_cpu_var(perf_cpu_context);
2505 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2506 put_cpu_var(perf_cpu_context);
2508 perf_counter_mmap_ctx(current->perf_counter_ctxp, mmap_event);
2513 void perf_counter_mmap(unsigned long addr, unsigned long len,
2514 unsigned long pgoff, struct file *file)
2516 struct perf_mmap_event mmap_event;
2518 if (!atomic_read(&nr_mmap_tracking))
2520 if (!current->perf_counter_ctxp)
2523 mmap_event = (struct perf_mmap_event){
2526 .header = { .type = PERF_EVENT_MMAP, },
2527 .pid = current->group_leader->pid,
2528 .tid = current->pid,
2535 perf_counter_mmap_event(&mmap_event);
2538 void perf_counter_munmap(unsigned long addr, unsigned long len,
2539 unsigned long pgoff, struct file *file)
2541 struct perf_mmap_event mmap_event;
2543 if (!atomic_read(&nr_munmap_tracking))
2546 mmap_event = (struct perf_mmap_event){
2549 .header = { .type = PERF_EVENT_MUNMAP, },
2550 .pid = current->group_leader->pid,
2551 .tid = current->pid,
2558 perf_counter_mmap_event(&mmap_event);
2565 static void perf_log_period(struct perf_counter *counter, u64 period)
2567 struct perf_output_handle handle;
2571 struct perf_event_header header;
2576 .type = PERF_EVENT_PERIOD,
2578 .size = sizeof(freq_event),
2580 .time = sched_clock(),
2584 if (counter->hw.irq_period == period)
2587 ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
2591 perf_output_put(&handle, freq_event);
2592 perf_output_end(&handle);
2596 * Generic counter overflow handling.
2599 int perf_counter_overflow(struct perf_counter *counter,
2600 int nmi, struct pt_regs *regs, u64 addr)
2602 int events = atomic_read(&counter->event_limit);
2605 counter->hw.interrupts++;
2608 * XXX event_limit might not quite work as expected on inherited
2612 counter->pending_kill = POLL_IN;
2613 if (events && atomic_dec_and_test(&counter->event_limit)) {
2615 counter->pending_kill = POLL_HUP;
2617 counter->pending_disable = 1;
2618 perf_pending_queue(&counter->pending,
2619 perf_pending_counter);
2621 perf_counter_disable(counter);
2624 perf_counter_output(counter, nmi, regs, addr);
2629 * Generic software counter infrastructure
2632 static void perf_swcounter_update(struct perf_counter *counter)
2634 struct hw_perf_counter *hwc = &counter->hw;
2639 prev = atomic64_read(&hwc->prev_count);
2640 now = atomic64_read(&hwc->count);
2641 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2646 atomic64_add(delta, &counter->count);
2647 atomic64_sub(delta, &hwc->period_left);
2650 static void perf_swcounter_set_period(struct perf_counter *counter)
2652 struct hw_perf_counter *hwc = &counter->hw;
2653 s64 left = atomic64_read(&hwc->period_left);
2654 s64 period = hwc->irq_period;
2656 if (unlikely(left <= -period)) {
2658 atomic64_set(&hwc->period_left, left);
2661 if (unlikely(left <= 0)) {
2663 atomic64_add(period, &hwc->period_left);
2666 atomic64_set(&hwc->prev_count, -left);
2667 atomic64_set(&hwc->count, -left);
2670 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2672 enum hrtimer_restart ret = HRTIMER_RESTART;
2673 struct perf_counter *counter;
2674 struct pt_regs *regs;
2677 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2678 counter->pmu->read(counter);
2680 regs = get_irq_regs();
2682 * In case we exclude kernel IPs or are somehow not in interrupt
2683 * context, provide the next best thing, the user IP.
2685 if ((counter->hw_event.exclude_kernel || !regs) &&
2686 !counter->hw_event.exclude_user)
2687 regs = task_pt_regs(current);
2690 if (perf_counter_overflow(counter, 0, regs, 0))
2691 ret = HRTIMER_NORESTART;
2694 period = max_t(u64, 10000, counter->hw.irq_period);
2695 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
2700 static void perf_swcounter_overflow(struct perf_counter *counter,
2701 int nmi, struct pt_regs *regs, u64 addr)
2703 perf_swcounter_update(counter);
2704 perf_swcounter_set_period(counter);
2705 if (perf_counter_overflow(counter, nmi, regs, addr))
2706 /* soft-disable the counter */
2711 static int perf_swcounter_match(struct perf_counter *counter,
2712 enum perf_event_types type,
2713 u32 event, struct pt_regs *regs)
2715 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2718 if (perf_event_raw(&counter->hw_event))
2721 if (perf_event_type(&counter->hw_event) != type)
2724 if (perf_event_id(&counter->hw_event) != event)
2727 if (counter->hw_event.exclude_user && user_mode(regs))
2730 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2736 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2737 int nmi, struct pt_regs *regs, u64 addr)
2739 int neg = atomic64_add_negative(nr, &counter->hw.count);
2740 if (counter->hw.irq_period && !neg)
2741 perf_swcounter_overflow(counter, nmi, regs, addr);
2744 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2745 enum perf_event_types type, u32 event,
2746 u64 nr, int nmi, struct pt_regs *regs,
2749 struct perf_counter *counter;
2751 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2755 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2756 if (perf_swcounter_match(counter, type, event, regs))
2757 perf_swcounter_add(counter, nr, nmi, regs, addr);
2762 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2765 return &cpuctx->recursion[3];
2768 return &cpuctx->recursion[2];
2771 return &cpuctx->recursion[1];
2773 return &cpuctx->recursion[0];
2776 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2777 u64 nr, int nmi, struct pt_regs *regs,
2780 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2781 int *recursion = perf_swcounter_recursion_context(cpuctx);
2789 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2790 nr, nmi, regs, addr);
2791 if (cpuctx->task_ctx) {
2792 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2793 nr, nmi, regs, addr);
2800 put_cpu_var(perf_cpu_context);
2804 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2806 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2809 static void perf_swcounter_read(struct perf_counter *counter)
2811 perf_swcounter_update(counter);
2814 static int perf_swcounter_enable(struct perf_counter *counter)
2816 perf_swcounter_set_period(counter);
2820 static void perf_swcounter_disable(struct perf_counter *counter)
2822 perf_swcounter_update(counter);
2825 static const struct pmu perf_ops_generic = {
2826 .enable = perf_swcounter_enable,
2827 .disable = perf_swcounter_disable,
2828 .read = perf_swcounter_read,
2832 * Software counter: cpu wall time clock
2835 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2837 int cpu = raw_smp_processor_id();
2841 now = cpu_clock(cpu);
2842 prev = atomic64_read(&counter->hw.prev_count);
2843 atomic64_set(&counter->hw.prev_count, now);
2844 atomic64_add(now - prev, &counter->count);
2847 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2849 struct hw_perf_counter *hwc = &counter->hw;
2850 int cpu = raw_smp_processor_id();
2852 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2853 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2854 hwc->hrtimer.function = perf_swcounter_hrtimer;
2855 if (hwc->irq_period) {
2856 u64 period = max_t(u64, 10000, hwc->irq_period);
2857 __hrtimer_start_range_ns(&hwc->hrtimer,
2858 ns_to_ktime(period), 0,
2859 HRTIMER_MODE_REL, 0);
2865 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2867 if (counter->hw.irq_period)
2868 hrtimer_cancel(&counter->hw.hrtimer);
2869 cpu_clock_perf_counter_update(counter);
2872 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2874 cpu_clock_perf_counter_update(counter);
2877 static const struct pmu perf_ops_cpu_clock = {
2878 .enable = cpu_clock_perf_counter_enable,
2879 .disable = cpu_clock_perf_counter_disable,
2880 .read = cpu_clock_perf_counter_read,
2884 * Software counter: task time clock
2887 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2892 prev = atomic64_xchg(&counter->hw.prev_count, now);
2894 atomic64_add(delta, &counter->count);
2897 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2899 struct hw_perf_counter *hwc = &counter->hw;
2902 now = counter->ctx->time;
2904 atomic64_set(&hwc->prev_count, now);
2905 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2906 hwc->hrtimer.function = perf_swcounter_hrtimer;
2907 if (hwc->irq_period) {
2908 u64 period = max_t(u64, 10000, hwc->irq_period);
2909 __hrtimer_start_range_ns(&hwc->hrtimer,
2910 ns_to_ktime(period), 0,
2911 HRTIMER_MODE_REL, 0);
2917 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2919 if (counter->hw.irq_period)
2920 hrtimer_cancel(&counter->hw.hrtimer);
2921 task_clock_perf_counter_update(counter, counter->ctx->time);
2925 static void task_clock_perf_counter_read(struct perf_counter *counter)
2930 update_context_time(counter->ctx);
2931 time = counter->ctx->time;
2933 u64 now = perf_clock();
2934 u64 delta = now - counter->ctx->timestamp;
2935 time = counter->ctx->time + delta;
2938 task_clock_perf_counter_update(counter, time);
2941 static const struct pmu perf_ops_task_clock = {
2942 .enable = task_clock_perf_counter_enable,
2943 .disable = task_clock_perf_counter_disable,
2944 .read = task_clock_perf_counter_read,
2948 * Software counter: cpu migrations
2951 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2953 struct task_struct *curr = counter->ctx->task;
2956 return curr->se.nr_migrations;
2957 return cpu_nr_migrations(smp_processor_id());
2960 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2965 prev = atomic64_read(&counter->hw.prev_count);
2966 now = get_cpu_migrations(counter);
2968 atomic64_set(&counter->hw.prev_count, now);
2972 atomic64_add(delta, &counter->count);
2975 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2977 cpu_migrations_perf_counter_update(counter);
2980 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2982 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2983 atomic64_set(&counter->hw.prev_count,
2984 get_cpu_migrations(counter));
2988 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2990 cpu_migrations_perf_counter_update(counter);
2993 static const struct pmu perf_ops_cpu_migrations = {
2994 .enable = cpu_migrations_perf_counter_enable,
2995 .disable = cpu_migrations_perf_counter_disable,
2996 .read = cpu_migrations_perf_counter_read,
2999 #ifdef CONFIG_EVENT_PROFILE
3000 void perf_tpcounter_event(int event_id)
3002 struct pt_regs *regs = get_irq_regs();
3005 regs = task_pt_regs(current);
3007 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
3009 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3011 extern int ftrace_profile_enable(int);
3012 extern void ftrace_profile_disable(int);
3014 static void tp_perf_counter_destroy(struct perf_counter *counter)
3016 ftrace_profile_disable(perf_event_id(&counter->hw_event));
3019 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3021 int event_id = perf_event_id(&counter->hw_event);
3024 ret = ftrace_profile_enable(event_id);
3028 counter->destroy = tp_perf_counter_destroy;
3029 counter->hw.irq_period = counter->hw_event.irq_period;
3031 return &perf_ops_generic;
3034 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3040 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3042 const struct pmu *pmu = NULL;
3045 * Software counters (currently) can't in general distinguish
3046 * between user, kernel and hypervisor events.
3047 * However, context switches and cpu migrations are considered
3048 * to be kernel events, and page faults are never hypervisor
3051 switch (perf_event_id(&counter->hw_event)) {
3052 case PERF_COUNT_CPU_CLOCK:
3053 pmu = &perf_ops_cpu_clock;
3056 case PERF_COUNT_TASK_CLOCK:
3058 * If the user instantiates this as a per-cpu counter,
3059 * use the cpu_clock counter instead.
3061 if (counter->ctx->task)
3062 pmu = &perf_ops_task_clock;
3064 pmu = &perf_ops_cpu_clock;
3067 case PERF_COUNT_PAGE_FAULTS:
3068 case PERF_COUNT_PAGE_FAULTS_MIN:
3069 case PERF_COUNT_PAGE_FAULTS_MAJ:
3070 case PERF_COUNT_CONTEXT_SWITCHES:
3071 pmu = &perf_ops_generic;
3073 case PERF_COUNT_CPU_MIGRATIONS:
3074 if (!counter->hw_event.exclude_kernel)
3075 pmu = &perf_ops_cpu_migrations;
3083 * Allocate and initialize a counter structure
3085 static struct perf_counter *
3086 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
3088 struct perf_counter_context *ctx,
3089 struct perf_counter *group_leader,
3092 const struct pmu *pmu;
3093 struct perf_counter *counter;
3094 struct hw_perf_counter *hwc;
3097 counter = kzalloc(sizeof(*counter), gfpflags);
3099 return ERR_PTR(-ENOMEM);
3102 * Single counters are their own group leaders, with an
3103 * empty sibling list:
3106 group_leader = counter;
3108 mutex_init(&counter->mutex);
3109 INIT_LIST_HEAD(&counter->list_entry);
3110 INIT_LIST_HEAD(&counter->event_entry);
3111 INIT_LIST_HEAD(&counter->sibling_list);
3112 init_waitqueue_head(&counter->waitq);
3114 mutex_init(&counter->mmap_mutex);
3116 INIT_LIST_HEAD(&counter->child_list);
3119 counter->hw_event = *hw_event;
3120 counter->group_leader = group_leader;
3121 counter->pmu = NULL;
3125 counter->state = PERF_COUNTER_STATE_INACTIVE;
3126 if (hw_event->disabled)
3127 counter->state = PERF_COUNTER_STATE_OFF;
3132 if (hw_event->freq && hw_event->irq_freq)
3133 hwc->irq_period = div64_u64(TICK_NSEC, hw_event->irq_freq);
3135 hwc->irq_period = hw_event->irq_period;
3138 * we currently do not support PERF_RECORD_GROUP on inherited counters
3140 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
3143 if (perf_event_raw(hw_event)) {
3144 pmu = hw_perf_counter_init(counter);
3148 switch (perf_event_type(hw_event)) {
3149 case PERF_TYPE_HARDWARE:
3150 pmu = hw_perf_counter_init(counter);
3153 case PERF_TYPE_SOFTWARE:
3154 pmu = sw_perf_counter_init(counter);
3157 case PERF_TYPE_TRACEPOINT:
3158 pmu = tp_perf_counter_init(counter);
3165 else if (IS_ERR(pmu))
3170 return ERR_PTR(err);
3175 atomic_inc(&nr_counters);
3176 if (counter->hw_event.mmap)
3177 atomic_inc(&nr_mmap_tracking);
3178 if (counter->hw_event.munmap)
3179 atomic_inc(&nr_munmap_tracking);
3180 if (counter->hw_event.comm)
3181 atomic_inc(&nr_comm_tracking);
3187 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3189 * @hw_event_uptr: event type attributes for monitoring/sampling
3192 * @group_fd: group leader counter fd
3194 SYSCALL_DEFINE5(perf_counter_open,
3195 const struct perf_counter_hw_event __user *, hw_event_uptr,
3196 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3198 struct perf_counter *counter, *group_leader;
3199 struct perf_counter_hw_event hw_event;
3200 struct perf_counter_context *ctx;
3201 struct file *counter_file = NULL;
3202 struct file *group_file = NULL;
3203 int fput_needed = 0;
3204 int fput_needed2 = 0;
3207 /* for future expandability... */
3211 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
3215 * Get the target context (task or percpu):
3217 ctx = find_get_context(pid, cpu);
3219 return PTR_ERR(ctx);
3222 * Look up the group leader (we will attach this counter to it):
3224 group_leader = NULL;
3225 if (group_fd != -1) {
3227 group_file = fget_light(group_fd, &fput_needed);
3229 goto err_put_context;
3230 if (group_file->f_op != &perf_fops)
3231 goto err_put_context;
3233 group_leader = group_file->private_data;
3235 * Do not allow a recursive hierarchy (this new sibling
3236 * becoming part of another group-sibling):
3238 if (group_leader->group_leader != group_leader)
3239 goto err_put_context;
3241 * Do not allow to attach to a group in a different
3242 * task or CPU context:
3244 if (group_leader->ctx != ctx)
3245 goto err_put_context;
3247 * Only a group leader can be exclusive or pinned
3249 if (hw_event.exclusive || hw_event.pinned)
3250 goto err_put_context;
3253 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
3255 ret = PTR_ERR(counter);
3256 if (IS_ERR(counter))
3257 goto err_put_context;
3259 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3261 goto err_free_put_context;
3263 counter_file = fget_light(ret, &fput_needed2);
3265 goto err_free_put_context;
3267 counter->filp = counter_file;
3268 mutex_lock(&ctx->mutex);
3269 perf_install_in_context(ctx, counter, cpu);
3270 mutex_unlock(&ctx->mutex);
3272 fput_light(counter_file, fput_needed2);
3275 fput_light(group_file, fput_needed);
3279 err_free_put_context:
3289 * inherit a counter from parent task to child task:
3291 static struct perf_counter *
3292 inherit_counter(struct perf_counter *parent_counter,
3293 struct task_struct *parent,
3294 struct perf_counter_context *parent_ctx,
3295 struct task_struct *child,
3296 struct perf_counter *group_leader,
3297 struct perf_counter_context *child_ctx)
3299 struct perf_counter *child_counter;
3302 * Instead of creating recursive hierarchies of counters,
3303 * we link inherited counters back to the original parent,
3304 * which has a filp for sure, which we use as the reference
3307 if (parent_counter->parent)
3308 parent_counter = parent_counter->parent;
3310 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3311 parent_counter->cpu, child_ctx,
3312 group_leader, GFP_KERNEL);
3313 if (IS_ERR(child_counter))
3314 return child_counter;
3317 * Make the child state follow the state of the parent counter,
3318 * not its hw_event.disabled bit. We hold the parent's mutex,
3319 * so we won't race with perf_counter_{en,dis}able_family.
3321 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3322 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3324 child_counter->state = PERF_COUNTER_STATE_OFF;
3327 * Link it up in the child's context:
3329 add_counter_to_ctx(child_counter, child_ctx);
3331 child_counter->parent = parent_counter;
3333 * inherit into child's child as well:
3335 child_counter->hw_event.inherit = 1;
3338 * Get a reference to the parent filp - we will fput it
3339 * when the child counter exits. This is safe to do because
3340 * we are in the parent and we know that the filp still
3341 * exists and has a nonzero count:
3343 atomic_long_inc(&parent_counter->filp->f_count);
3346 * Link this into the parent counter's child list
3348 mutex_lock(&parent_counter->mutex);
3349 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3351 mutex_unlock(&parent_counter->mutex);
3353 return child_counter;
3356 static int inherit_group(struct perf_counter *parent_counter,
3357 struct task_struct *parent,
3358 struct perf_counter_context *parent_ctx,
3359 struct task_struct *child,
3360 struct perf_counter_context *child_ctx)
3362 struct perf_counter *leader;
3363 struct perf_counter *sub;
3364 struct perf_counter *child_ctr;
3366 leader = inherit_counter(parent_counter, parent, parent_ctx,
3367 child, NULL, child_ctx);
3369 return PTR_ERR(leader);
3370 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3371 child_ctr = inherit_counter(sub, parent, parent_ctx,
3372 child, leader, child_ctx);
3373 if (IS_ERR(child_ctr))
3374 return PTR_ERR(child_ctr);
3379 static void sync_child_counter(struct perf_counter *child_counter,
3380 struct perf_counter *parent_counter)
3384 child_val = atomic64_read(&child_counter->count);
3387 * Add back the child's count to the parent's count:
3389 atomic64_add(child_val, &parent_counter->count);
3390 atomic64_add(child_counter->total_time_enabled,
3391 &parent_counter->child_total_time_enabled);
3392 atomic64_add(child_counter->total_time_running,
3393 &parent_counter->child_total_time_running);
3396 * Remove this counter from the parent's list
3398 mutex_lock(&parent_counter->mutex);
3399 list_del_init(&child_counter->child_list);
3400 mutex_unlock(&parent_counter->mutex);
3403 * Release the parent counter, if this was the last
3406 fput(parent_counter->filp);
3410 __perf_counter_exit_task(struct task_struct *child,
3411 struct perf_counter *child_counter,
3412 struct perf_counter_context *child_ctx)
3414 struct perf_counter *parent_counter;
3417 * Protect against concurrent operations on child_counter
3418 * due its fd getting closed, etc.
3420 mutex_lock(&child_counter->mutex);
3422 update_counter_times(child_counter);
3423 list_del_counter(child_counter, child_ctx);
3425 mutex_unlock(&child_counter->mutex);
3427 parent_counter = child_counter->parent;
3429 * It can happen that parent exits first, and has counters
3430 * that are still around due to the child reference. These
3431 * counters need to be zapped - but otherwise linger.
3433 if (parent_counter) {
3434 sync_child_counter(child_counter, parent_counter);
3435 free_counter(child_counter);
3440 * When a child task exits, feed back counter values to parent counters.
3442 * Note: we may be running in child context, but the PID is not hashed
3443 * anymore so new counters will not be added.
3444 * (XXX not sure that is true when we get called from flush_old_exec.
3447 void perf_counter_exit_task(struct task_struct *child)
3449 struct perf_counter *child_counter, *tmp;
3450 struct perf_counter_context *child_ctx;
3451 unsigned long flags;
3453 WARN_ON_ONCE(child != current);
3455 child_ctx = child->perf_counter_ctxp;
3457 if (likely(!child_ctx))
3460 local_irq_save(flags);
3461 __perf_counter_task_sched_out(child_ctx);
3462 child->perf_counter_ctxp = NULL;
3463 local_irq_restore(flags);
3465 mutex_lock(&child_ctx->mutex);
3468 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3470 __perf_counter_exit_task(child, child_counter, child_ctx);
3473 * If the last counter was a group counter, it will have appended all
3474 * its siblings to the list, but we obtained 'tmp' before that which
3475 * will still point to the list head terminating the iteration.
3477 if (!list_empty(&child_ctx->counter_list))
3480 mutex_unlock(&child_ctx->mutex);
3486 * Initialize the perf_counter context in task_struct
3488 void perf_counter_init_task(struct task_struct *child)
3490 struct perf_counter_context *child_ctx, *parent_ctx;
3491 struct perf_counter *counter;
3492 struct task_struct *parent = current;
3493 int inherited_all = 1;
3495 child->perf_counter_ctxp = NULL;
3498 * This is executed from the parent task context, so inherit
3499 * counters that have been marked for cloning.
3500 * First allocate and initialize a context for the child.
3503 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
3507 parent_ctx = parent->perf_counter_ctxp;
3508 if (likely(!parent_ctx || !parent_ctx->nr_counters))
3511 __perf_counter_init_context(child_ctx, child);
3512 child->perf_counter_ctxp = child_ctx;
3515 * Lock the parent list. No need to lock the child - not PID
3516 * hashed yet and not running, so nobody can access it.
3518 mutex_lock(&parent_ctx->mutex);
3521 * We dont have to disable NMIs - we are only looking at
3522 * the list, not manipulating it:
3524 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
3525 if (counter != counter->group_leader)
3528 if (!counter->hw_event.inherit) {
3533 if (inherit_group(counter, parent,
3534 parent_ctx, child, child_ctx)) {
3540 if (inherited_all) {
3542 * Mark the child context as a clone of the parent
3543 * context, or of whatever the parent is a clone of.
3545 if (parent_ctx->parent_ctx) {
3546 child_ctx->parent_ctx = parent_ctx->parent_ctx;
3547 child_ctx->parent_gen = parent_ctx->parent_gen;
3549 child_ctx->parent_ctx = parent_ctx;
3550 child_ctx->parent_gen = parent_ctx->generation;
3552 get_ctx(child_ctx->parent_ctx);
3555 mutex_unlock(&parent_ctx->mutex);
3558 static void __cpuinit perf_counter_init_cpu(int cpu)
3560 struct perf_cpu_context *cpuctx;
3562 cpuctx = &per_cpu(perf_cpu_context, cpu);
3563 __perf_counter_init_context(&cpuctx->ctx, NULL);
3565 spin_lock(&perf_resource_lock);
3566 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3567 spin_unlock(&perf_resource_lock);
3569 hw_perf_counter_setup(cpu);
3572 #ifdef CONFIG_HOTPLUG_CPU
3573 static void __perf_counter_exit_cpu(void *info)
3575 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3576 struct perf_counter_context *ctx = &cpuctx->ctx;
3577 struct perf_counter *counter, *tmp;
3579 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3580 __perf_counter_remove_from_context(counter);
3582 static void perf_counter_exit_cpu(int cpu)
3584 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3585 struct perf_counter_context *ctx = &cpuctx->ctx;
3587 mutex_lock(&ctx->mutex);
3588 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3589 mutex_unlock(&ctx->mutex);
3592 static inline void perf_counter_exit_cpu(int cpu) { }
3595 static int __cpuinit
3596 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3598 unsigned int cpu = (long)hcpu;
3602 case CPU_UP_PREPARE:
3603 case CPU_UP_PREPARE_FROZEN:
3604 perf_counter_init_cpu(cpu);
3607 case CPU_DOWN_PREPARE:
3608 case CPU_DOWN_PREPARE_FROZEN:
3609 perf_counter_exit_cpu(cpu);
3619 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3620 .notifier_call = perf_cpu_notify,
3623 void __init perf_counter_init(void)
3625 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3626 (void *)(long)smp_processor_id());
3627 register_cpu_notifier(&perf_cpu_nb);
3630 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3632 return sprintf(buf, "%d\n", perf_reserved_percpu);
3636 perf_set_reserve_percpu(struct sysdev_class *class,
3640 struct perf_cpu_context *cpuctx;
3644 err = strict_strtoul(buf, 10, &val);
3647 if (val > perf_max_counters)
3650 spin_lock(&perf_resource_lock);
3651 perf_reserved_percpu = val;
3652 for_each_online_cpu(cpu) {
3653 cpuctx = &per_cpu(perf_cpu_context, cpu);
3654 spin_lock_irq(&cpuctx->ctx.lock);
3655 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3656 perf_max_counters - perf_reserved_percpu);
3657 cpuctx->max_pertask = mpt;
3658 spin_unlock_irq(&cpuctx->ctx.lock);
3660 spin_unlock(&perf_resource_lock);
3665 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3667 return sprintf(buf, "%d\n", perf_overcommit);
3671 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3676 err = strict_strtoul(buf, 10, &val);
3682 spin_lock(&perf_resource_lock);
3683 perf_overcommit = val;
3684 spin_unlock(&perf_resource_lock);
3689 static SYSDEV_CLASS_ATTR(
3692 perf_show_reserve_percpu,
3693 perf_set_reserve_percpu
3696 static SYSDEV_CLASS_ATTR(
3699 perf_show_overcommit,
3703 static struct attribute *perfclass_attrs[] = {
3704 &attr_reserve_percpu.attr,
3705 &attr_overcommit.attr,
3709 static struct attribute_group perfclass_attr_group = {
3710 .attrs = perfclass_attrs,
3711 .name = "perf_counters",
3714 static int __init perf_counter_sysfs_init(void)
3716 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3717 &perfclass_attr_group);
3719 device_initcall(perf_counter_sysfs_init);