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 * Add a counter from the lists for its context.
116 * Must be called with ctx->mutex and ctx->lock held.
119 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
121 struct perf_counter *group_leader = counter->group_leader;
124 * Depending on whether it is a standalone or sibling counter,
125 * add it straight to the context's counter list, or to the group
126 * leader's sibling list:
128 if (group_leader == counter)
129 list_add_tail(&counter->list_entry, &ctx->counter_list);
131 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
132 group_leader->nr_siblings++;
135 list_add_rcu(&counter->event_entry, &ctx->event_list);
140 * Remove a counter from the lists for its context.
141 * Must be called with ctx->mutex and ctx->lock held.
144 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
146 struct perf_counter *sibling, *tmp;
148 if (list_empty(&counter->list_entry))
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 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;
407 spin_unlock_irqrestore(&ctx->lock, flags);
413 static void perf_counter_disable(struct perf_counter *counter)
415 struct perf_counter_context *ctx = counter->ctx;
416 struct task_struct *task = ctx->task;
420 * Disable the counter on the cpu that it's on
422 smp_call_function_single(counter->cpu, __perf_counter_disable,
428 task_oncpu_function_call(task, __perf_counter_disable, counter);
430 spin_lock_irq(&ctx->lock);
432 * If the counter is still active, we need to retry the cross-call.
434 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
435 spin_unlock_irq(&ctx->lock);
440 * Since we have the lock this context can't be scheduled
441 * in, so we can change the state safely.
443 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
444 update_counter_times(counter);
445 counter->state = PERF_COUNTER_STATE_OFF;
448 spin_unlock_irq(&ctx->lock);
452 counter_sched_in(struct perf_counter *counter,
453 struct perf_cpu_context *cpuctx,
454 struct perf_counter_context *ctx,
457 if (counter->state <= PERF_COUNTER_STATE_OFF)
460 counter->state = PERF_COUNTER_STATE_ACTIVE;
461 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
463 * The new state must be visible before we turn it on in the hardware:
467 if (counter->pmu->enable(counter)) {
468 counter->state = PERF_COUNTER_STATE_INACTIVE;
473 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
475 if (!is_software_counter(counter))
476 cpuctx->active_oncpu++;
479 if (counter->hw_event.exclusive)
480 cpuctx->exclusive = 1;
486 group_sched_in(struct perf_counter *group_counter,
487 struct perf_cpu_context *cpuctx,
488 struct perf_counter_context *ctx,
491 struct perf_counter *counter, *partial_group;
494 if (group_counter->state == PERF_COUNTER_STATE_OFF)
497 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
499 return ret < 0 ? ret : 0;
501 group_counter->prev_state = group_counter->state;
502 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
506 * Schedule in siblings as one group (if any):
508 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
509 counter->prev_state = counter->state;
510 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
511 partial_group = counter;
520 * Groups can be scheduled in as one unit only, so undo any
521 * partial group before returning:
523 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
524 if (counter == partial_group)
526 counter_sched_out(counter, cpuctx, ctx);
528 counter_sched_out(group_counter, cpuctx, ctx);
534 * Return 1 for a group consisting entirely of software counters,
535 * 0 if the group contains any hardware counters.
537 static int is_software_only_group(struct perf_counter *leader)
539 struct perf_counter *counter;
541 if (!is_software_counter(leader))
544 list_for_each_entry(counter, &leader->sibling_list, list_entry)
545 if (!is_software_counter(counter))
552 * Work out whether we can put this counter group on the CPU now.
554 static int group_can_go_on(struct perf_counter *counter,
555 struct perf_cpu_context *cpuctx,
559 * Groups consisting entirely of software counters can always go on.
561 if (is_software_only_group(counter))
564 * If an exclusive group is already on, no other hardware
565 * counters can go on.
567 if (cpuctx->exclusive)
570 * If this group is exclusive and there are already
571 * counters on the CPU, it can't go on.
573 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
576 * Otherwise, try to add it if all previous groups were able
582 static void add_counter_to_ctx(struct perf_counter *counter,
583 struct perf_counter_context *ctx)
585 list_add_counter(counter, ctx);
586 counter->prev_state = PERF_COUNTER_STATE_OFF;
587 counter->tstamp_enabled = ctx->time;
588 counter->tstamp_running = ctx->time;
589 counter->tstamp_stopped = ctx->time;
593 * Cross CPU call to install and enable a performance counter
595 * Must be called with ctx->mutex held
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;
758 * If the counter is in a group and isn't the group leader,
759 * then don't put it on unless the group is on.
761 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
764 if (!group_can_go_on(counter, cpuctx, 1)) {
768 if (counter == leader)
769 err = group_sched_in(counter, cpuctx, ctx,
772 err = counter_sched_in(counter, cpuctx, ctx,
779 * If this counter can't go on and it's part of a
780 * group, then the whole group has to come off.
782 if (leader != counter)
783 group_sched_out(leader, cpuctx, ctx);
784 if (leader->hw_event.pinned) {
785 update_group_times(leader);
786 leader->state = PERF_COUNTER_STATE_ERROR;
791 spin_unlock_irqrestore(&ctx->lock, flags);
797 static void perf_counter_enable(struct perf_counter *counter)
799 struct perf_counter_context *ctx = counter->ctx;
800 struct task_struct *task = ctx->task;
804 * Enable the counter on the cpu that it's on
806 smp_call_function_single(counter->cpu, __perf_counter_enable,
811 spin_lock_irq(&ctx->lock);
812 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
816 * If the counter is in error state, clear that first.
817 * That way, if we see the counter in error state below, we
818 * know that it has gone back into error state, as distinct
819 * from the task having been scheduled away before the
820 * cross-call arrived.
822 if (counter->state == PERF_COUNTER_STATE_ERROR)
823 counter->state = PERF_COUNTER_STATE_OFF;
826 spin_unlock_irq(&ctx->lock);
827 task_oncpu_function_call(task, __perf_counter_enable, counter);
829 spin_lock_irq(&ctx->lock);
832 * If the context is active and the counter is still off,
833 * we need to retry the cross-call.
835 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
839 * Since we have the lock this context can't be scheduled
840 * in, so we can change the state safely.
842 if (counter->state == PERF_COUNTER_STATE_OFF) {
843 counter->state = PERF_COUNTER_STATE_INACTIVE;
844 counter->tstamp_enabled =
845 ctx->time - counter->total_time_enabled;
848 spin_unlock_irq(&ctx->lock);
851 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
854 * not supported on inherited counters
856 if (counter->hw_event.inherit)
859 atomic_add(refresh, &counter->event_limit);
860 perf_counter_enable(counter);
865 void __perf_counter_sched_out(struct perf_counter_context *ctx,
866 struct perf_cpu_context *cpuctx)
868 struct perf_counter *counter;
870 spin_lock(&ctx->lock);
872 if (likely(!ctx->nr_counters))
874 update_context_time(ctx);
877 if (ctx->nr_active) {
878 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
879 if (counter != counter->group_leader)
880 counter_sched_out(counter, cpuctx, ctx);
882 group_sched_out(counter, cpuctx, ctx);
887 spin_unlock(&ctx->lock);
891 * Test whether two contexts are equivalent, i.e. whether they
892 * have both been cloned from the same version of the same context
893 * and they both have the same number of enabled counters.
894 * If the number of enabled counters is the same, then the set
895 * of enabled counters should be the same, because these are both
896 * inherited contexts, therefore we can't access individual counters
897 * in them directly with an fd; we can only enable/disable all
898 * counters via prctl, or enable/disable all counters in a family
899 * via ioctl, which will have the same effect on both contexts.
901 static int context_equiv(struct perf_counter_context *ctx1,
902 struct perf_counter_context *ctx2)
904 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
905 && ctx1->parent_gen == ctx2->parent_gen;
909 * Called from scheduler to remove the counters of the current task,
910 * with interrupts disabled.
912 * We stop each counter and update the counter value in counter->count.
914 * This does not protect us against NMI, but disable()
915 * sets the disabled bit in the control field of counter _before_
916 * accessing the counter control register. If a NMI hits, then it will
917 * not restart the counter.
919 void perf_counter_task_sched_out(struct task_struct *task,
920 struct task_struct *next, int cpu)
922 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
923 struct perf_counter_context *ctx = task->perf_counter_ctxp;
924 struct perf_counter_context *next_ctx;
925 struct pt_regs *regs;
927 regs = task_pt_regs(task);
928 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
930 if (likely(!ctx || !cpuctx->task_ctx))
933 update_context_time(ctx);
934 next_ctx = next->perf_counter_ctxp;
935 if (next_ctx && context_equiv(ctx, next_ctx)) {
936 task->perf_counter_ctxp = next_ctx;
937 next->perf_counter_ctxp = ctx;
939 next_ctx->task = task;
943 __perf_counter_sched_out(ctx, cpuctx);
945 cpuctx->task_ctx = NULL;
948 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
950 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
952 if (!cpuctx->task_ctx)
954 __perf_counter_sched_out(ctx, cpuctx);
955 cpuctx->task_ctx = NULL;
958 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
960 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
964 __perf_counter_sched_in(struct perf_counter_context *ctx,
965 struct perf_cpu_context *cpuctx, int cpu)
967 struct perf_counter *counter;
970 spin_lock(&ctx->lock);
972 if (likely(!ctx->nr_counters))
975 ctx->timestamp = perf_clock();
980 * First go through the list and put on any pinned groups
981 * in order to give them the best chance of going on.
983 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
984 if (counter->state <= PERF_COUNTER_STATE_OFF ||
985 !counter->hw_event.pinned)
987 if (counter->cpu != -1 && counter->cpu != cpu)
990 if (counter != counter->group_leader)
991 counter_sched_in(counter, cpuctx, ctx, cpu);
993 if (group_can_go_on(counter, cpuctx, 1))
994 group_sched_in(counter, cpuctx, ctx, cpu);
998 * If this pinned group hasn't been scheduled,
999 * put it in error state.
1001 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1002 update_group_times(counter);
1003 counter->state = PERF_COUNTER_STATE_ERROR;
1007 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1009 * Ignore counters in OFF or ERROR state, and
1010 * ignore pinned counters since we did them already.
1012 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1013 counter->hw_event.pinned)
1017 * Listen to the 'cpu' scheduling filter constraint
1020 if (counter->cpu != -1 && counter->cpu != cpu)
1023 if (counter != counter->group_leader) {
1024 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1027 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1028 if (group_sched_in(counter, cpuctx, ctx, cpu))
1035 spin_unlock(&ctx->lock);
1039 * Called from scheduler to add the counters of the current task
1040 * with interrupts disabled.
1042 * We restore the counter value and then enable it.
1044 * This does not protect us against NMI, but enable()
1045 * sets the enabled bit in the control field of counter _before_
1046 * accessing the counter control register. If a NMI hits, then it will
1047 * keep the counter running.
1049 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1051 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1052 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1056 if (cpuctx->task_ctx == ctx)
1058 __perf_counter_sched_in(ctx, cpuctx, cpu);
1059 cpuctx->task_ctx = ctx;
1062 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1064 struct perf_counter_context *ctx = &cpuctx->ctx;
1066 __perf_counter_sched_in(ctx, cpuctx, cpu);
1069 static void perf_log_period(struct perf_counter *counter, u64 period);
1071 static void perf_adjust_freq(struct perf_counter_context *ctx)
1073 struct perf_counter *counter;
1078 spin_lock(&ctx->lock);
1079 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1080 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1083 if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
1086 events = HZ * counter->hw.interrupts * counter->hw.irq_period;
1087 period = div64_u64(events, counter->hw_event.irq_freq);
1089 delta = (s64)(1 + period - counter->hw.irq_period);
1092 irq_period = counter->hw.irq_period + delta;
1097 perf_log_period(counter, irq_period);
1099 counter->hw.irq_period = irq_period;
1100 counter->hw.interrupts = 0;
1102 spin_unlock(&ctx->lock);
1106 * Round-robin a context's counters:
1108 static void rotate_ctx(struct perf_counter_context *ctx)
1110 struct perf_counter *counter;
1112 if (!ctx->nr_counters)
1115 spin_lock(&ctx->lock);
1117 * Rotate the first entry last (works just fine for group counters too):
1120 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1121 list_move_tail(&counter->list_entry, &ctx->counter_list);
1126 spin_unlock(&ctx->lock);
1129 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1131 struct perf_cpu_context *cpuctx;
1132 struct perf_counter_context *ctx;
1134 if (!atomic_read(&nr_counters))
1137 cpuctx = &per_cpu(perf_cpu_context, cpu);
1138 ctx = curr->perf_counter_ctxp;
1140 perf_adjust_freq(&cpuctx->ctx);
1142 perf_adjust_freq(ctx);
1144 perf_counter_cpu_sched_out(cpuctx);
1146 __perf_counter_task_sched_out(ctx);
1148 rotate_ctx(&cpuctx->ctx);
1152 perf_counter_cpu_sched_in(cpuctx, cpu);
1154 perf_counter_task_sched_in(curr, cpu);
1158 * Cross CPU call to read the hardware counter
1160 static void __read(void *info)
1162 struct perf_counter *counter = info;
1163 struct perf_counter_context *ctx = counter->ctx;
1164 unsigned long flags;
1166 local_irq_save(flags);
1168 update_context_time(ctx);
1169 counter->pmu->read(counter);
1170 update_counter_times(counter);
1171 local_irq_restore(flags);
1174 static u64 perf_counter_read(struct perf_counter *counter)
1177 * If counter is enabled and currently active on a CPU, update the
1178 * value in the counter structure:
1180 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1181 smp_call_function_single(counter->oncpu,
1182 __read, counter, 1);
1183 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1184 update_counter_times(counter);
1187 return atomic64_read(&counter->count);
1191 * Initialize the perf_counter context in a task_struct:
1194 __perf_counter_init_context(struct perf_counter_context *ctx,
1195 struct task_struct *task)
1197 memset(ctx, 0, sizeof(*ctx));
1198 spin_lock_init(&ctx->lock);
1199 mutex_init(&ctx->mutex);
1200 INIT_LIST_HEAD(&ctx->counter_list);
1201 INIT_LIST_HEAD(&ctx->event_list);
1202 atomic_set(&ctx->refcount, 1);
1206 static void put_context(struct perf_counter_context *ctx)
1209 put_task_struct(ctx->task);
1212 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1214 struct perf_cpu_context *cpuctx;
1215 struct perf_counter_context *ctx;
1216 struct perf_counter_context *tctx;
1217 struct task_struct *task;
1220 * If cpu is not a wildcard then this is a percpu counter:
1223 /* Must be root to operate on a CPU counter: */
1224 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1225 return ERR_PTR(-EACCES);
1227 if (cpu < 0 || cpu > num_possible_cpus())
1228 return ERR_PTR(-EINVAL);
1231 * We could be clever and allow to attach a counter to an
1232 * offline CPU and activate it when the CPU comes up, but
1235 if (!cpu_isset(cpu, cpu_online_map))
1236 return ERR_PTR(-ENODEV);
1238 cpuctx = &per_cpu(perf_cpu_context, cpu);
1248 task = find_task_by_vpid(pid);
1250 get_task_struct(task);
1254 return ERR_PTR(-ESRCH);
1256 /* Reuse ptrace permission checks for now. */
1257 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1258 put_task_struct(task);
1259 return ERR_PTR(-EACCES);
1262 ctx = task->perf_counter_ctxp;
1264 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1266 put_task_struct(task);
1267 return ERR_PTR(-ENOMEM);
1269 __perf_counter_init_context(ctx, task);
1271 * Make sure other cpus see correct values for *ctx
1272 * once task->perf_counter_ctxp is visible to them.
1275 tctx = cmpxchg(&task->perf_counter_ctxp, NULL, ctx);
1278 * We raced with some other task; use
1279 * the context they set.
1289 static void free_counter_rcu(struct rcu_head *head)
1291 struct perf_counter *counter;
1293 counter = container_of(head, struct perf_counter, rcu_head);
1294 put_ctx(counter->ctx);
1298 static void perf_pending_sync(struct perf_counter *counter);
1300 static void free_counter(struct perf_counter *counter)
1302 perf_pending_sync(counter);
1304 atomic_dec(&nr_counters);
1305 if (counter->hw_event.mmap)
1306 atomic_dec(&nr_mmap_tracking);
1307 if (counter->hw_event.munmap)
1308 atomic_dec(&nr_munmap_tracking);
1309 if (counter->hw_event.comm)
1310 atomic_dec(&nr_comm_tracking);
1312 if (counter->destroy)
1313 counter->destroy(counter);
1315 call_rcu(&counter->rcu_head, free_counter_rcu);
1319 * Called when the last reference to the file is gone.
1321 static int perf_release(struct inode *inode, struct file *file)
1323 struct perf_counter *counter = file->private_data;
1324 struct perf_counter_context *ctx = counter->ctx;
1326 file->private_data = NULL;
1328 mutex_lock(&ctx->mutex);
1329 perf_counter_remove_from_context(counter);
1330 mutex_unlock(&ctx->mutex);
1332 mutex_lock(&counter->owner->perf_counter_mutex);
1333 list_del_init(&counter->owner_entry);
1334 mutex_unlock(&counter->owner->perf_counter_mutex);
1335 put_task_struct(counter->owner);
1337 free_counter(counter);
1344 * Read the performance counter - simple non blocking version for now
1347 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1353 * Return end-of-file for a read on a counter that is in
1354 * error state (i.e. because it was pinned but it couldn't be
1355 * scheduled on to the CPU at some point).
1357 if (counter->state == PERF_COUNTER_STATE_ERROR)
1360 mutex_lock(&counter->child_mutex);
1361 values[0] = perf_counter_read(counter);
1363 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1364 values[n++] = counter->total_time_enabled +
1365 atomic64_read(&counter->child_total_time_enabled);
1366 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1367 values[n++] = counter->total_time_running +
1368 atomic64_read(&counter->child_total_time_running);
1369 mutex_unlock(&counter->child_mutex);
1371 if (count < n * sizeof(u64))
1373 count = n * sizeof(u64);
1375 if (copy_to_user(buf, values, count))
1382 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1384 struct perf_counter *counter = file->private_data;
1386 return perf_read_hw(counter, buf, count);
1389 static unsigned int perf_poll(struct file *file, poll_table *wait)
1391 struct perf_counter *counter = file->private_data;
1392 struct perf_mmap_data *data;
1393 unsigned int events = POLL_HUP;
1396 data = rcu_dereference(counter->data);
1398 events = atomic_xchg(&data->poll, 0);
1401 poll_wait(file, &counter->waitq, wait);
1406 static void perf_counter_reset(struct perf_counter *counter)
1408 (void)perf_counter_read(counter);
1409 atomic64_set(&counter->count, 0);
1410 perf_counter_update_userpage(counter);
1413 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1414 void (*func)(struct perf_counter *))
1416 struct perf_counter_context *ctx = counter->ctx;
1417 struct perf_counter *sibling;
1419 mutex_lock(&ctx->mutex);
1420 counter = counter->group_leader;
1423 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1425 mutex_unlock(&ctx->mutex);
1428 static void perf_counter_for_each_child(struct perf_counter *counter,
1429 void (*func)(struct perf_counter *))
1431 struct perf_counter *child;
1433 mutex_lock(&counter->child_mutex);
1435 list_for_each_entry(child, &counter->child_list, child_list)
1437 mutex_unlock(&counter->child_mutex);
1440 static void perf_counter_for_each(struct perf_counter *counter,
1441 void (*func)(struct perf_counter *))
1443 struct perf_counter *child;
1445 mutex_lock(&counter->child_mutex);
1446 perf_counter_for_each_sibling(counter, func);
1447 list_for_each_entry(child, &counter->child_list, child_list)
1448 perf_counter_for_each_sibling(child, func);
1449 mutex_unlock(&counter->child_mutex);
1452 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1454 struct perf_counter *counter = file->private_data;
1455 void (*func)(struct perf_counter *);
1459 case PERF_COUNTER_IOC_ENABLE:
1460 func = perf_counter_enable;
1462 case PERF_COUNTER_IOC_DISABLE:
1463 func = perf_counter_disable;
1465 case PERF_COUNTER_IOC_RESET:
1466 func = perf_counter_reset;
1469 case PERF_COUNTER_IOC_REFRESH:
1470 return perf_counter_refresh(counter, arg);
1475 if (flags & PERF_IOC_FLAG_GROUP)
1476 perf_counter_for_each(counter, func);
1478 perf_counter_for_each_child(counter, func);
1483 int perf_counter_task_enable(void)
1485 struct perf_counter *counter;
1487 mutex_lock(¤t->perf_counter_mutex);
1488 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1489 perf_counter_for_each_child(counter, perf_counter_enable);
1490 mutex_unlock(¤t->perf_counter_mutex);
1495 int perf_counter_task_disable(void)
1497 struct perf_counter *counter;
1499 mutex_lock(¤t->perf_counter_mutex);
1500 list_for_each_entry(counter, ¤t->perf_counter_list, owner_entry)
1501 perf_counter_for_each_child(counter, perf_counter_disable);
1502 mutex_unlock(¤t->perf_counter_mutex);
1508 * Callers need to ensure there can be no nesting of this function, otherwise
1509 * the seqlock logic goes bad. We can not serialize this because the arch
1510 * code calls this from NMI context.
1512 void perf_counter_update_userpage(struct perf_counter *counter)
1514 struct perf_mmap_data *data;
1515 struct perf_counter_mmap_page *userpg;
1518 data = rcu_dereference(counter->data);
1522 userpg = data->user_page;
1525 * Disable preemption so as to not let the corresponding user-space
1526 * spin too long if we get preempted.
1531 userpg->index = counter->hw.idx;
1532 userpg->offset = atomic64_read(&counter->count);
1533 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1534 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1543 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1545 struct perf_counter *counter = vma->vm_file->private_data;
1546 struct perf_mmap_data *data;
1547 int ret = VM_FAULT_SIGBUS;
1550 data = rcu_dereference(counter->data);
1554 if (vmf->pgoff == 0) {
1555 vmf->page = virt_to_page(data->user_page);
1557 int nr = vmf->pgoff - 1;
1559 if ((unsigned)nr > data->nr_pages)
1562 vmf->page = virt_to_page(data->data_pages[nr]);
1564 get_page(vmf->page);
1572 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1574 struct perf_mmap_data *data;
1578 WARN_ON(atomic_read(&counter->mmap_count));
1580 size = sizeof(struct perf_mmap_data);
1581 size += nr_pages * sizeof(void *);
1583 data = kzalloc(size, GFP_KERNEL);
1587 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1588 if (!data->user_page)
1589 goto fail_user_page;
1591 for (i = 0; i < nr_pages; i++) {
1592 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1593 if (!data->data_pages[i])
1594 goto fail_data_pages;
1597 data->nr_pages = nr_pages;
1598 atomic_set(&data->lock, -1);
1600 rcu_assign_pointer(counter->data, data);
1605 for (i--; i >= 0; i--)
1606 free_page((unsigned long)data->data_pages[i]);
1608 free_page((unsigned long)data->user_page);
1617 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1619 struct perf_mmap_data *data = container_of(rcu_head,
1620 struct perf_mmap_data, rcu_head);
1623 free_page((unsigned long)data->user_page);
1624 for (i = 0; i < data->nr_pages; i++)
1625 free_page((unsigned long)data->data_pages[i]);
1629 static void perf_mmap_data_free(struct perf_counter *counter)
1631 struct perf_mmap_data *data = counter->data;
1633 WARN_ON(atomic_read(&counter->mmap_count));
1635 rcu_assign_pointer(counter->data, NULL);
1636 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1639 static void perf_mmap_open(struct vm_area_struct *vma)
1641 struct perf_counter *counter = vma->vm_file->private_data;
1643 atomic_inc(&counter->mmap_count);
1646 static void perf_mmap_close(struct vm_area_struct *vma)
1648 struct perf_counter *counter = vma->vm_file->private_data;
1650 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1651 &counter->mmap_mutex)) {
1652 struct user_struct *user = current_user();
1654 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
1655 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1656 perf_mmap_data_free(counter);
1657 mutex_unlock(&counter->mmap_mutex);
1661 static struct vm_operations_struct perf_mmap_vmops = {
1662 .open = perf_mmap_open,
1663 .close = perf_mmap_close,
1664 .fault = perf_mmap_fault,
1667 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1669 struct perf_counter *counter = file->private_data;
1670 struct user_struct *user = current_user();
1671 unsigned long vma_size;
1672 unsigned long nr_pages;
1673 unsigned long user_locked, user_lock_limit;
1674 unsigned long locked, lock_limit;
1675 long user_extra, extra;
1678 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1681 vma_size = vma->vm_end - vma->vm_start;
1682 nr_pages = (vma_size / PAGE_SIZE) - 1;
1685 * If we have data pages ensure they're a power-of-two number, so we
1686 * can do bitmasks instead of modulo.
1688 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1691 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1694 if (vma->vm_pgoff != 0)
1697 mutex_lock(&counter->mmap_mutex);
1698 if (atomic_inc_not_zero(&counter->mmap_count)) {
1699 if (nr_pages != counter->data->nr_pages)
1704 user_extra = nr_pages + 1;
1705 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1708 * Increase the limit linearly with more CPUs:
1710 user_lock_limit *= num_online_cpus();
1712 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
1715 if (user_locked > user_lock_limit)
1716 extra = user_locked - user_lock_limit;
1718 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1719 lock_limit >>= PAGE_SHIFT;
1720 locked = vma->vm_mm->locked_vm + extra;
1722 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1727 WARN_ON(counter->data);
1728 ret = perf_mmap_data_alloc(counter, nr_pages);
1732 atomic_set(&counter->mmap_count, 1);
1733 atomic_long_add(user_extra, &user->locked_vm);
1734 vma->vm_mm->locked_vm += extra;
1735 counter->data->nr_locked = extra;
1737 mutex_unlock(&counter->mmap_mutex);
1739 vma->vm_flags &= ~VM_MAYWRITE;
1740 vma->vm_flags |= VM_RESERVED;
1741 vma->vm_ops = &perf_mmap_vmops;
1746 static int perf_fasync(int fd, struct file *filp, int on)
1748 struct perf_counter *counter = filp->private_data;
1749 struct inode *inode = filp->f_path.dentry->d_inode;
1752 mutex_lock(&inode->i_mutex);
1753 retval = fasync_helper(fd, filp, on, &counter->fasync);
1754 mutex_unlock(&inode->i_mutex);
1762 static const struct file_operations perf_fops = {
1763 .release = perf_release,
1766 .unlocked_ioctl = perf_ioctl,
1767 .compat_ioctl = perf_ioctl,
1769 .fasync = perf_fasync,
1773 * Perf counter wakeup
1775 * If there's data, ensure we set the poll() state and publish everything
1776 * to user-space before waking everybody up.
1779 void perf_counter_wakeup(struct perf_counter *counter)
1781 wake_up_all(&counter->waitq);
1783 if (counter->pending_kill) {
1784 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1785 counter->pending_kill = 0;
1792 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1794 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1795 * single linked list and use cmpxchg() to add entries lockless.
1798 static void perf_pending_counter(struct perf_pending_entry *entry)
1800 struct perf_counter *counter = container_of(entry,
1801 struct perf_counter, pending);
1803 if (counter->pending_disable) {
1804 counter->pending_disable = 0;
1805 perf_counter_disable(counter);
1808 if (counter->pending_wakeup) {
1809 counter->pending_wakeup = 0;
1810 perf_counter_wakeup(counter);
1814 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1816 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1820 static void perf_pending_queue(struct perf_pending_entry *entry,
1821 void (*func)(struct perf_pending_entry *))
1823 struct perf_pending_entry **head;
1825 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1830 head = &get_cpu_var(perf_pending_head);
1833 entry->next = *head;
1834 } while (cmpxchg(head, entry->next, entry) != entry->next);
1836 set_perf_counter_pending();
1838 put_cpu_var(perf_pending_head);
1841 static int __perf_pending_run(void)
1843 struct perf_pending_entry *list;
1846 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1847 while (list != PENDING_TAIL) {
1848 void (*func)(struct perf_pending_entry *);
1849 struct perf_pending_entry *entry = list;
1856 * Ensure we observe the unqueue before we issue the wakeup,
1857 * so that we won't be waiting forever.
1858 * -- see perf_not_pending().
1869 static inline int perf_not_pending(struct perf_counter *counter)
1872 * If we flush on whatever cpu we run, there is a chance we don't
1876 __perf_pending_run();
1880 * Ensure we see the proper queue state before going to sleep
1881 * so that we do not miss the wakeup. -- see perf_pending_handle()
1884 return counter->pending.next == NULL;
1887 static void perf_pending_sync(struct perf_counter *counter)
1889 wait_event(counter->waitq, perf_not_pending(counter));
1892 void perf_counter_do_pending(void)
1894 __perf_pending_run();
1898 * Callchain support -- arch specific
1901 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1910 struct perf_output_handle {
1911 struct perf_counter *counter;
1912 struct perf_mmap_data *data;
1913 unsigned int offset;
1918 unsigned long flags;
1921 static void perf_output_wakeup(struct perf_output_handle *handle)
1923 atomic_set(&handle->data->poll, POLL_IN);
1926 handle->counter->pending_wakeup = 1;
1927 perf_pending_queue(&handle->counter->pending,
1928 perf_pending_counter);
1930 perf_counter_wakeup(handle->counter);
1934 * Curious locking construct.
1936 * We need to ensure a later event doesn't publish a head when a former
1937 * event isn't done writing. However since we need to deal with NMIs we
1938 * cannot fully serialize things.
1940 * What we do is serialize between CPUs so we only have to deal with NMI
1941 * nesting on a single CPU.
1943 * We only publish the head (and generate a wakeup) when the outer-most
1946 static void perf_output_lock(struct perf_output_handle *handle)
1948 struct perf_mmap_data *data = handle->data;
1953 local_irq_save(handle->flags);
1954 cpu = smp_processor_id();
1956 if (in_nmi() && atomic_read(&data->lock) == cpu)
1959 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1965 static void perf_output_unlock(struct perf_output_handle *handle)
1967 struct perf_mmap_data *data = handle->data;
1970 data->done_head = data->head;
1972 if (!handle->locked)
1977 * The xchg implies a full barrier that ensures all writes are done
1978 * before we publish the new head, matched by a rmb() in userspace when
1979 * reading this position.
1981 while ((head = atomic_xchg(&data->done_head, 0)))
1982 data->user_page->data_head = head;
1985 * NMI can happen here, which means we can miss a done_head update.
1988 cpu = atomic_xchg(&data->lock, -1);
1989 WARN_ON_ONCE(cpu != smp_processor_id());
1992 * Therefore we have to validate we did not indeed do so.
1994 if (unlikely(atomic_read(&data->done_head))) {
1996 * Since we had it locked, we can lock it again.
1998 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2004 if (atomic_xchg(&data->wakeup, 0))
2005 perf_output_wakeup(handle);
2007 local_irq_restore(handle->flags);
2010 static int perf_output_begin(struct perf_output_handle *handle,
2011 struct perf_counter *counter, unsigned int size,
2012 int nmi, int overflow)
2014 struct perf_mmap_data *data;
2015 unsigned int offset, head;
2018 * For inherited counters we send all the output towards the parent.
2020 if (counter->parent)
2021 counter = counter->parent;
2024 data = rcu_dereference(counter->data);
2028 handle->data = data;
2029 handle->counter = counter;
2031 handle->overflow = overflow;
2033 if (!data->nr_pages)
2036 perf_output_lock(handle);
2039 offset = head = atomic_read(&data->head);
2041 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
2043 handle->offset = offset;
2044 handle->head = head;
2046 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2047 atomic_set(&data->wakeup, 1);
2052 perf_output_wakeup(handle);
2059 static void perf_output_copy(struct perf_output_handle *handle,
2060 void *buf, unsigned int len)
2062 unsigned int pages_mask;
2063 unsigned int offset;
2067 offset = handle->offset;
2068 pages_mask = handle->data->nr_pages - 1;
2069 pages = handle->data->data_pages;
2072 unsigned int page_offset;
2075 nr = (offset >> PAGE_SHIFT) & pages_mask;
2076 page_offset = offset & (PAGE_SIZE - 1);
2077 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2079 memcpy(pages[nr] + page_offset, buf, size);
2086 handle->offset = offset;
2089 * Check we didn't copy past our reservation window, taking the
2090 * possible unsigned int wrap into account.
2092 WARN_ON_ONCE(((int)(handle->head - handle->offset)) < 0);
2095 #define perf_output_put(handle, x) \
2096 perf_output_copy((handle), &(x), sizeof(x))
2098 static void perf_output_end(struct perf_output_handle *handle)
2100 struct perf_counter *counter = handle->counter;
2101 struct perf_mmap_data *data = handle->data;
2103 int wakeup_events = counter->hw_event.wakeup_events;
2105 if (handle->overflow && wakeup_events) {
2106 int events = atomic_inc_return(&data->events);
2107 if (events >= wakeup_events) {
2108 atomic_sub(wakeup_events, &data->events);
2109 atomic_set(&data->wakeup, 1);
2113 perf_output_unlock(handle);
2117 static void perf_counter_output(struct perf_counter *counter,
2118 int nmi, struct pt_regs *regs, u64 addr)
2121 u64 record_type = counter->hw_event.record_type;
2122 struct perf_output_handle handle;
2123 struct perf_event_header header;
2132 struct perf_callchain_entry *callchain = NULL;
2133 int callchain_size = 0;
2140 header.size = sizeof(header);
2142 header.misc = PERF_EVENT_MISC_OVERFLOW;
2143 header.misc |= perf_misc_flags(regs);
2145 if (record_type & PERF_RECORD_IP) {
2146 ip = perf_instruction_pointer(regs);
2147 header.type |= PERF_RECORD_IP;
2148 header.size += sizeof(ip);
2151 if (record_type & PERF_RECORD_TID) {
2152 /* namespace issues */
2153 tid_entry.pid = current->group_leader->pid;
2154 tid_entry.tid = current->pid;
2156 header.type |= PERF_RECORD_TID;
2157 header.size += sizeof(tid_entry);
2160 if (record_type & PERF_RECORD_TIME) {
2162 * Maybe do better on x86 and provide cpu_clock_nmi()
2164 time = sched_clock();
2166 header.type |= PERF_RECORD_TIME;
2167 header.size += sizeof(u64);
2170 if (record_type & PERF_RECORD_ADDR) {
2171 header.type |= PERF_RECORD_ADDR;
2172 header.size += sizeof(u64);
2175 if (record_type & PERF_RECORD_CONFIG) {
2176 header.type |= PERF_RECORD_CONFIG;
2177 header.size += sizeof(u64);
2180 if (record_type & PERF_RECORD_CPU) {
2181 header.type |= PERF_RECORD_CPU;
2182 header.size += sizeof(cpu_entry);
2184 cpu_entry.cpu = raw_smp_processor_id();
2187 if (record_type & PERF_RECORD_GROUP) {
2188 header.type |= PERF_RECORD_GROUP;
2189 header.size += sizeof(u64) +
2190 counter->nr_siblings * sizeof(group_entry);
2193 if (record_type & PERF_RECORD_CALLCHAIN) {
2194 callchain = perf_callchain(regs);
2197 callchain_size = (1 + callchain->nr) * sizeof(u64);
2199 header.type |= PERF_RECORD_CALLCHAIN;
2200 header.size += callchain_size;
2204 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2208 perf_output_put(&handle, header);
2210 if (record_type & PERF_RECORD_IP)
2211 perf_output_put(&handle, ip);
2213 if (record_type & PERF_RECORD_TID)
2214 perf_output_put(&handle, tid_entry);
2216 if (record_type & PERF_RECORD_TIME)
2217 perf_output_put(&handle, time);
2219 if (record_type & PERF_RECORD_ADDR)
2220 perf_output_put(&handle, addr);
2222 if (record_type & PERF_RECORD_CONFIG)
2223 perf_output_put(&handle, counter->hw_event.config);
2225 if (record_type & PERF_RECORD_CPU)
2226 perf_output_put(&handle, cpu_entry);
2229 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2231 if (record_type & PERF_RECORD_GROUP) {
2232 struct perf_counter *leader, *sub;
2233 u64 nr = counter->nr_siblings;
2235 perf_output_put(&handle, nr);
2237 leader = counter->group_leader;
2238 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2240 sub->pmu->read(sub);
2242 group_entry.event = sub->hw_event.config;
2243 group_entry.counter = atomic64_read(&sub->count);
2245 perf_output_put(&handle, group_entry);
2250 perf_output_copy(&handle, callchain, callchain_size);
2252 perf_output_end(&handle);
2259 struct perf_comm_event {
2260 struct task_struct *task;
2265 struct perf_event_header header;
2272 static void perf_counter_comm_output(struct perf_counter *counter,
2273 struct perf_comm_event *comm_event)
2275 struct perf_output_handle handle;
2276 int size = comm_event->event.header.size;
2277 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2282 perf_output_put(&handle, comm_event->event);
2283 perf_output_copy(&handle, comm_event->comm,
2284 comm_event->comm_size);
2285 perf_output_end(&handle);
2288 static int perf_counter_comm_match(struct perf_counter *counter,
2289 struct perf_comm_event *comm_event)
2291 if (counter->hw_event.comm &&
2292 comm_event->event.header.type == PERF_EVENT_COMM)
2298 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2299 struct perf_comm_event *comm_event)
2301 struct perf_counter *counter;
2303 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2307 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2308 if (perf_counter_comm_match(counter, comm_event))
2309 perf_counter_comm_output(counter, comm_event);
2314 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2316 struct perf_cpu_context *cpuctx;
2318 char *comm = comm_event->task->comm;
2320 size = ALIGN(strlen(comm)+1, sizeof(u64));
2322 comm_event->comm = comm;
2323 comm_event->comm_size = size;
2325 comm_event->event.header.size = sizeof(comm_event->event) + size;
2327 cpuctx = &get_cpu_var(perf_cpu_context);
2328 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2329 put_cpu_var(perf_cpu_context);
2331 perf_counter_comm_ctx(current->perf_counter_ctxp, comm_event);
2334 void perf_counter_comm(struct task_struct *task)
2336 struct perf_comm_event comm_event;
2338 if (!atomic_read(&nr_comm_tracking))
2340 if (!current->perf_counter_ctxp)
2343 comm_event = (struct perf_comm_event){
2346 .header = { .type = PERF_EVENT_COMM, },
2347 .pid = task->group_leader->pid,
2352 perf_counter_comm_event(&comm_event);
2359 struct perf_mmap_event {
2365 struct perf_event_header header;
2375 static void perf_counter_mmap_output(struct perf_counter *counter,
2376 struct perf_mmap_event *mmap_event)
2378 struct perf_output_handle handle;
2379 int size = mmap_event->event.header.size;
2380 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2385 perf_output_put(&handle, mmap_event->event);
2386 perf_output_copy(&handle, mmap_event->file_name,
2387 mmap_event->file_size);
2388 perf_output_end(&handle);
2391 static int perf_counter_mmap_match(struct perf_counter *counter,
2392 struct perf_mmap_event *mmap_event)
2394 if (counter->hw_event.mmap &&
2395 mmap_event->event.header.type == PERF_EVENT_MMAP)
2398 if (counter->hw_event.munmap &&
2399 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2405 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2406 struct perf_mmap_event *mmap_event)
2408 struct perf_counter *counter;
2410 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2414 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2415 if (perf_counter_mmap_match(counter, mmap_event))
2416 perf_counter_mmap_output(counter, mmap_event);
2421 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2423 struct perf_cpu_context *cpuctx;
2424 struct file *file = mmap_event->file;
2431 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2433 name = strncpy(tmp, "//enomem", sizeof(tmp));
2436 name = d_path(&file->f_path, buf, PATH_MAX);
2438 name = strncpy(tmp, "//toolong", sizeof(tmp));
2442 name = strncpy(tmp, "//anon", sizeof(tmp));
2447 size = ALIGN(strlen(name)+1, sizeof(u64));
2449 mmap_event->file_name = name;
2450 mmap_event->file_size = size;
2452 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2454 cpuctx = &get_cpu_var(perf_cpu_context);
2455 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2456 put_cpu_var(perf_cpu_context);
2458 perf_counter_mmap_ctx(current->perf_counter_ctxp, mmap_event);
2463 void perf_counter_mmap(unsigned long addr, unsigned long len,
2464 unsigned long pgoff, struct file *file)
2466 struct perf_mmap_event mmap_event;
2468 if (!atomic_read(&nr_mmap_tracking))
2470 if (!current->perf_counter_ctxp)
2473 mmap_event = (struct perf_mmap_event){
2476 .header = { .type = PERF_EVENT_MMAP, },
2477 .pid = current->group_leader->pid,
2478 .tid = current->pid,
2485 perf_counter_mmap_event(&mmap_event);
2488 void perf_counter_munmap(unsigned long addr, unsigned long len,
2489 unsigned long pgoff, struct file *file)
2491 struct perf_mmap_event mmap_event;
2493 if (!atomic_read(&nr_munmap_tracking))
2496 mmap_event = (struct perf_mmap_event){
2499 .header = { .type = PERF_EVENT_MUNMAP, },
2500 .pid = current->group_leader->pid,
2501 .tid = current->pid,
2508 perf_counter_mmap_event(&mmap_event);
2512 * Log irq_period changes so that analyzing tools can re-normalize the
2516 static void perf_log_period(struct perf_counter *counter, u64 period)
2518 struct perf_output_handle handle;
2522 struct perf_event_header header;
2527 .type = PERF_EVENT_PERIOD,
2529 .size = sizeof(freq_event),
2531 .time = sched_clock(),
2535 if (counter->hw.irq_period == period)
2538 ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
2542 perf_output_put(&handle, freq_event);
2543 perf_output_end(&handle);
2547 * Generic counter overflow handling.
2550 int perf_counter_overflow(struct perf_counter *counter,
2551 int nmi, struct pt_regs *regs, u64 addr)
2553 int events = atomic_read(&counter->event_limit);
2556 counter->hw.interrupts++;
2559 * XXX event_limit might not quite work as expected on inherited
2563 counter->pending_kill = POLL_IN;
2564 if (events && atomic_dec_and_test(&counter->event_limit)) {
2566 counter->pending_kill = POLL_HUP;
2568 counter->pending_disable = 1;
2569 perf_pending_queue(&counter->pending,
2570 perf_pending_counter);
2572 perf_counter_disable(counter);
2575 perf_counter_output(counter, nmi, regs, addr);
2580 * Generic software counter infrastructure
2583 static void perf_swcounter_update(struct perf_counter *counter)
2585 struct hw_perf_counter *hwc = &counter->hw;
2590 prev = atomic64_read(&hwc->prev_count);
2591 now = atomic64_read(&hwc->count);
2592 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2597 atomic64_add(delta, &counter->count);
2598 atomic64_sub(delta, &hwc->period_left);
2601 static void perf_swcounter_set_period(struct perf_counter *counter)
2603 struct hw_perf_counter *hwc = &counter->hw;
2604 s64 left = atomic64_read(&hwc->period_left);
2605 s64 period = hwc->irq_period;
2607 if (unlikely(left <= -period)) {
2609 atomic64_set(&hwc->period_left, left);
2612 if (unlikely(left <= 0)) {
2614 atomic64_add(period, &hwc->period_left);
2617 atomic64_set(&hwc->prev_count, -left);
2618 atomic64_set(&hwc->count, -left);
2621 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2623 enum hrtimer_restart ret = HRTIMER_RESTART;
2624 struct perf_counter *counter;
2625 struct pt_regs *regs;
2628 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2629 counter->pmu->read(counter);
2631 regs = get_irq_regs();
2633 * In case we exclude kernel IPs or are somehow not in interrupt
2634 * context, provide the next best thing, the user IP.
2636 if ((counter->hw_event.exclude_kernel || !regs) &&
2637 !counter->hw_event.exclude_user)
2638 regs = task_pt_regs(current);
2641 if (perf_counter_overflow(counter, 0, regs, 0))
2642 ret = HRTIMER_NORESTART;
2645 period = max_t(u64, 10000, counter->hw.irq_period);
2646 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
2651 static void perf_swcounter_overflow(struct perf_counter *counter,
2652 int nmi, struct pt_regs *regs, u64 addr)
2654 perf_swcounter_update(counter);
2655 perf_swcounter_set_period(counter);
2656 if (perf_counter_overflow(counter, nmi, regs, addr))
2657 /* soft-disable the counter */
2662 static int perf_swcounter_match(struct perf_counter *counter,
2663 enum perf_event_types type,
2664 u32 event, struct pt_regs *regs)
2666 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2669 if (perf_event_raw(&counter->hw_event))
2672 if (perf_event_type(&counter->hw_event) != type)
2675 if (perf_event_id(&counter->hw_event) != event)
2678 if (counter->hw_event.exclude_user && user_mode(regs))
2681 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2687 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2688 int nmi, struct pt_regs *regs, u64 addr)
2690 int neg = atomic64_add_negative(nr, &counter->hw.count);
2691 if (counter->hw.irq_period && !neg)
2692 perf_swcounter_overflow(counter, nmi, regs, addr);
2695 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2696 enum perf_event_types type, u32 event,
2697 u64 nr, int nmi, struct pt_regs *regs,
2700 struct perf_counter *counter;
2702 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2706 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2707 if (perf_swcounter_match(counter, type, event, regs))
2708 perf_swcounter_add(counter, nr, nmi, regs, addr);
2713 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2716 return &cpuctx->recursion[3];
2719 return &cpuctx->recursion[2];
2722 return &cpuctx->recursion[1];
2724 return &cpuctx->recursion[0];
2727 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2728 u64 nr, int nmi, struct pt_regs *regs,
2731 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2732 int *recursion = perf_swcounter_recursion_context(cpuctx);
2740 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2741 nr, nmi, regs, addr);
2742 if (cpuctx->task_ctx) {
2743 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2744 nr, nmi, regs, addr);
2751 put_cpu_var(perf_cpu_context);
2755 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2757 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2760 static void perf_swcounter_read(struct perf_counter *counter)
2762 perf_swcounter_update(counter);
2765 static int perf_swcounter_enable(struct perf_counter *counter)
2767 perf_swcounter_set_period(counter);
2771 static void perf_swcounter_disable(struct perf_counter *counter)
2773 perf_swcounter_update(counter);
2776 static const struct pmu perf_ops_generic = {
2777 .enable = perf_swcounter_enable,
2778 .disable = perf_swcounter_disable,
2779 .read = perf_swcounter_read,
2783 * Software counter: cpu wall time clock
2786 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2788 int cpu = raw_smp_processor_id();
2792 now = cpu_clock(cpu);
2793 prev = atomic64_read(&counter->hw.prev_count);
2794 atomic64_set(&counter->hw.prev_count, now);
2795 atomic64_add(now - prev, &counter->count);
2798 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2800 struct hw_perf_counter *hwc = &counter->hw;
2801 int cpu = raw_smp_processor_id();
2803 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2804 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2805 hwc->hrtimer.function = perf_swcounter_hrtimer;
2806 if (hwc->irq_period) {
2807 u64 period = max_t(u64, 10000, hwc->irq_period);
2808 __hrtimer_start_range_ns(&hwc->hrtimer,
2809 ns_to_ktime(period), 0,
2810 HRTIMER_MODE_REL, 0);
2816 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2818 if (counter->hw.irq_period)
2819 hrtimer_cancel(&counter->hw.hrtimer);
2820 cpu_clock_perf_counter_update(counter);
2823 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2825 cpu_clock_perf_counter_update(counter);
2828 static const struct pmu perf_ops_cpu_clock = {
2829 .enable = cpu_clock_perf_counter_enable,
2830 .disable = cpu_clock_perf_counter_disable,
2831 .read = cpu_clock_perf_counter_read,
2835 * Software counter: task time clock
2838 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2843 prev = atomic64_xchg(&counter->hw.prev_count, now);
2845 atomic64_add(delta, &counter->count);
2848 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2850 struct hw_perf_counter *hwc = &counter->hw;
2853 now = counter->ctx->time;
2855 atomic64_set(&hwc->prev_count, now);
2856 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2857 hwc->hrtimer.function = perf_swcounter_hrtimer;
2858 if (hwc->irq_period) {
2859 u64 period = max_t(u64, 10000, hwc->irq_period);
2860 __hrtimer_start_range_ns(&hwc->hrtimer,
2861 ns_to_ktime(period), 0,
2862 HRTIMER_MODE_REL, 0);
2868 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2870 if (counter->hw.irq_period)
2871 hrtimer_cancel(&counter->hw.hrtimer);
2872 task_clock_perf_counter_update(counter, counter->ctx->time);
2876 static void task_clock_perf_counter_read(struct perf_counter *counter)
2881 update_context_time(counter->ctx);
2882 time = counter->ctx->time;
2884 u64 now = perf_clock();
2885 u64 delta = now - counter->ctx->timestamp;
2886 time = counter->ctx->time + delta;
2889 task_clock_perf_counter_update(counter, time);
2892 static const struct pmu perf_ops_task_clock = {
2893 .enable = task_clock_perf_counter_enable,
2894 .disable = task_clock_perf_counter_disable,
2895 .read = task_clock_perf_counter_read,
2899 * Software counter: cpu migrations
2902 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2904 struct task_struct *curr = counter->ctx->task;
2907 return curr->se.nr_migrations;
2908 return cpu_nr_migrations(smp_processor_id());
2911 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2916 prev = atomic64_read(&counter->hw.prev_count);
2917 now = get_cpu_migrations(counter);
2919 atomic64_set(&counter->hw.prev_count, now);
2923 atomic64_add(delta, &counter->count);
2926 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2928 cpu_migrations_perf_counter_update(counter);
2931 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2933 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2934 atomic64_set(&counter->hw.prev_count,
2935 get_cpu_migrations(counter));
2939 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2941 cpu_migrations_perf_counter_update(counter);
2944 static const struct pmu perf_ops_cpu_migrations = {
2945 .enable = cpu_migrations_perf_counter_enable,
2946 .disable = cpu_migrations_perf_counter_disable,
2947 .read = cpu_migrations_perf_counter_read,
2950 #ifdef CONFIG_EVENT_PROFILE
2951 void perf_tpcounter_event(int event_id)
2953 struct pt_regs *regs = get_irq_regs();
2956 regs = task_pt_regs(current);
2958 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2960 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2962 extern int ftrace_profile_enable(int);
2963 extern void ftrace_profile_disable(int);
2965 static void tp_perf_counter_destroy(struct perf_counter *counter)
2967 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2970 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2972 int event_id = perf_event_id(&counter->hw_event);
2975 ret = ftrace_profile_enable(event_id);
2979 counter->destroy = tp_perf_counter_destroy;
2980 counter->hw.irq_period = counter->hw_event.irq_period;
2982 return &perf_ops_generic;
2985 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2991 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2993 const struct pmu *pmu = NULL;
2996 * Software counters (currently) can't in general distinguish
2997 * between user, kernel and hypervisor events.
2998 * However, context switches and cpu migrations are considered
2999 * to be kernel events, and page faults are never hypervisor
3002 switch (perf_event_id(&counter->hw_event)) {
3003 case PERF_COUNT_CPU_CLOCK:
3004 pmu = &perf_ops_cpu_clock;
3007 case PERF_COUNT_TASK_CLOCK:
3009 * If the user instantiates this as a per-cpu counter,
3010 * use the cpu_clock counter instead.
3012 if (counter->ctx->task)
3013 pmu = &perf_ops_task_clock;
3015 pmu = &perf_ops_cpu_clock;
3018 case PERF_COUNT_PAGE_FAULTS:
3019 case PERF_COUNT_PAGE_FAULTS_MIN:
3020 case PERF_COUNT_PAGE_FAULTS_MAJ:
3021 case PERF_COUNT_CONTEXT_SWITCHES:
3022 pmu = &perf_ops_generic;
3024 case PERF_COUNT_CPU_MIGRATIONS:
3025 if (!counter->hw_event.exclude_kernel)
3026 pmu = &perf_ops_cpu_migrations;
3034 * Allocate and initialize a counter structure
3036 static struct perf_counter *
3037 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
3039 struct perf_counter_context *ctx,
3040 struct perf_counter *group_leader,
3043 const struct pmu *pmu;
3044 struct perf_counter *counter;
3045 struct hw_perf_counter *hwc;
3048 counter = kzalloc(sizeof(*counter), gfpflags);
3050 return ERR_PTR(-ENOMEM);
3053 * Single counters are their own group leaders, with an
3054 * empty sibling list:
3057 group_leader = counter;
3059 mutex_init(&counter->child_mutex);
3060 INIT_LIST_HEAD(&counter->child_list);
3062 INIT_LIST_HEAD(&counter->list_entry);
3063 INIT_LIST_HEAD(&counter->event_entry);
3064 INIT_LIST_HEAD(&counter->sibling_list);
3065 init_waitqueue_head(&counter->waitq);
3067 mutex_init(&counter->mmap_mutex);
3070 counter->hw_event = *hw_event;
3071 counter->group_leader = group_leader;
3072 counter->pmu = NULL;
3076 counter->state = PERF_COUNTER_STATE_INACTIVE;
3077 if (hw_event->disabled)
3078 counter->state = PERF_COUNTER_STATE_OFF;
3083 if (hw_event->freq && hw_event->irq_freq)
3084 hwc->irq_period = div64_u64(TICK_NSEC, hw_event->irq_freq);
3086 hwc->irq_period = hw_event->irq_period;
3089 * we currently do not support PERF_RECORD_GROUP on inherited counters
3091 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
3094 if (perf_event_raw(hw_event)) {
3095 pmu = hw_perf_counter_init(counter);
3099 switch (perf_event_type(hw_event)) {
3100 case PERF_TYPE_HARDWARE:
3101 pmu = hw_perf_counter_init(counter);
3104 case PERF_TYPE_SOFTWARE:
3105 pmu = sw_perf_counter_init(counter);
3108 case PERF_TYPE_TRACEPOINT:
3109 pmu = tp_perf_counter_init(counter);
3116 else if (IS_ERR(pmu))
3121 return ERR_PTR(err);
3126 atomic_inc(&nr_counters);
3127 if (counter->hw_event.mmap)
3128 atomic_inc(&nr_mmap_tracking);
3129 if (counter->hw_event.munmap)
3130 atomic_inc(&nr_munmap_tracking);
3131 if (counter->hw_event.comm)
3132 atomic_inc(&nr_comm_tracking);
3138 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3140 * @hw_event_uptr: event type attributes for monitoring/sampling
3143 * @group_fd: group leader counter fd
3145 SYSCALL_DEFINE5(perf_counter_open,
3146 const struct perf_counter_hw_event __user *, hw_event_uptr,
3147 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3149 struct perf_counter *counter, *group_leader;
3150 struct perf_counter_hw_event hw_event;
3151 struct perf_counter_context *ctx;
3152 struct file *counter_file = NULL;
3153 struct file *group_file = NULL;
3154 int fput_needed = 0;
3155 int fput_needed2 = 0;
3158 /* for future expandability... */
3162 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
3166 * Get the target context (task or percpu):
3168 ctx = find_get_context(pid, cpu);
3170 return PTR_ERR(ctx);
3173 * Look up the group leader (we will attach this counter to it):
3175 group_leader = NULL;
3176 if (group_fd != -1) {
3178 group_file = fget_light(group_fd, &fput_needed);
3180 goto err_put_context;
3181 if (group_file->f_op != &perf_fops)
3182 goto err_put_context;
3184 group_leader = group_file->private_data;
3186 * Do not allow a recursive hierarchy (this new sibling
3187 * becoming part of another group-sibling):
3189 if (group_leader->group_leader != group_leader)
3190 goto err_put_context;
3192 * Do not allow to attach to a group in a different
3193 * task or CPU context:
3195 if (group_leader->ctx != ctx)
3196 goto err_put_context;
3198 * Only a group leader can be exclusive or pinned
3200 if (hw_event.exclusive || hw_event.pinned)
3201 goto err_put_context;
3204 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
3206 ret = PTR_ERR(counter);
3207 if (IS_ERR(counter))
3208 goto err_put_context;
3210 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3212 goto err_free_put_context;
3214 counter_file = fget_light(ret, &fput_needed2);
3216 goto err_free_put_context;
3218 counter->filp = counter_file;
3219 mutex_lock(&ctx->mutex);
3220 perf_install_in_context(ctx, counter, cpu);
3221 mutex_unlock(&ctx->mutex);
3223 counter->owner = current;
3224 get_task_struct(current);
3225 mutex_lock(¤t->perf_counter_mutex);
3226 list_add_tail(&counter->owner_entry, ¤t->perf_counter_list);
3227 mutex_unlock(¤t->perf_counter_mutex);
3229 fput_light(counter_file, fput_needed2);
3232 fput_light(group_file, fput_needed);
3236 err_free_put_context:
3246 * inherit a counter from parent task to child task:
3248 static struct perf_counter *
3249 inherit_counter(struct perf_counter *parent_counter,
3250 struct task_struct *parent,
3251 struct perf_counter_context *parent_ctx,
3252 struct task_struct *child,
3253 struct perf_counter *group_leader,
3254 struct perf_counter_context *child_ctx)
3256 struct perf_counter *child_counter;
3259 * Instead of creating recursive hierarchies of counters,
3260 * we link inherited counters back to the original parent,
3261 * which has a filp for sure, which we use as the reference
3264 if (parent_counter->parent)
3265 parent_counter = parent_counter->parent;
3267 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3268 parent_counter->cpu, child_ctx,
3269 group_leader, GFP_KERNEL);
3270 if (IS_ERR(child_counter))
3271 return child_counter;
3274 * Make the child state follow the state of the parent counter,
3275 * not its hw_event.disabled bit. We hold the parent's mutex,
3276 * so we won't race with perf_counter_{en,dis}able_family.
3278 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3279 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3281 child_counter->state = PERF_COUNTER_STATE_OFF;
3284 * Link it up in the child's context:
3286 add_counter_to_ctx(child_counter, child_ctx);
3288 child_counter->parent = parent_counter;
3290 * inherit into child's child as well:
3292 child_counter->hw_event.inherit = 1;
3295 * Get a reference to the parent filp - we will fput it
3296 * when the child counter exits. This is safe to do because
3297 * we are in the parent and we know that the filp still
3298 * exists and has a nonzero count:
3300 atomic_long_inc(&parent_counter->filp->f_count);
3303 * Link this into the parent counter's child list
3305 mutex_lock(&parent_counter->child_mutex);
3306 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3307 mutex_unlock(&parent_counter->child_mutex);
3309 return child_counter;
3312 static int inherit_group(struct perf_counter *parent_counter,
3313 struct task_struct *parent,
3314 struct perf_counter_context *parent_ctx,
3315 struct task_struct *child,
3316 struct perf_counter_context *child_ctx)
3318 struct perf_counter *leader;
3319 struct perf_counter *sub;
3320 struct perf_counter *child_ctr;
3322 leader = inherit_counter(parent_counter, parent, parent_ctx,
3323 child, NULL, child_ctx);
3325 return PTR_ERR(leader);
3326 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3327 child_ctr = inherit_counter(sub, parent, parent_ctx,
3328 child, leader, child_ctx);
3329 if (IS_ERR(child_ctr))
3330 return PTR_ERR(child_ctr);
3335 static void sync_child_counter(struct perf_counter *child_counter,
3336 struct perf_counter *parent_counter)
3340 child_val = atomic64_read(&child_counter->count);
3343 * Add back the child's count to the parent's count:
3345 atomic64_add(child_val, &parent_counter->count);
3346 atomic64_add(child_counter->total_time_enabled,
3347 &parent_counter->child_total_time_enabled);
3348 atomic64_add(child_counter->total_time_running,
3349 &parent_counter->child_total_time_running);
3352 * Remove this counter from the parent's list
3354 mutex_lock(&parent_counter->child_mutex);
3355 list_del_init(&child_counter->child_list);
3356 mutex_unlock(&parent_counter->child_mutex);
3359 * Release the parent counter, if this was the last
3362 fput(parent_counter->filp);
3366 __perf_counter_exit_task(struct task_struct *child,
3367 struct perf_counter *child_counter,
3368 struct perf_counter_context *child_ctx)
3370 struct perf_counter *parent_counter;
3372 update_counter_times(child_counter);
3373 perf_counter_remove_from_context(child_counter);
3375 parent_counter = child_counter->parent;
3377 * It can happen that parent exits first, and has counters
3378 * that are still around due to the child reference. These
3379 * counters need to be zapped - but otherwise linger.
3381 if (parent_counter) {
3382 sync_child_counter(child_counter, parent_counter);
3383 free_counter(child_counter);
3388 * When a child task exits, feed back counter values to parent counters.
3390 * Note: we may be running in child context, but the PID is not hashed
3391 * anymore so new counters will not be added.
3392 * (XXX not sure that is true when we get called from flush_old_exec.
3395 void perf_counter_exit_task(struct task_struct *child)
3397 struct perf_counter *child_counter, *tmp;
3398 struct perf_counter_context *child_ctx;
3399 unsigned long flags;
3401 WARN_ON_ONCE(child != current);
3403 child_ctx = child->perf_counter_ctxp;
3405 if (likely(!child_ctx))
3408 local_irq_save(flags);
3409 __perf_counter_task_sched_out(child_ctx);
3410 child->perf_counter_ctxp = NULL;
3411 local_irq_restore(flags);
3413 mutex_lock(&child_ctx->mutex);
3416 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3418 __perf_counter_exit_task(child, child_counter, child_ctx);
3421 * If the last counter was a group counter, it will have appended all
3422 * its siblings to the list, but we obtained 'tmp' before that which
3423 * will still point to the list head terminating the iteration.
3425 if (!list_empty(&child_ctx->counter_list))
3428 mutex_unlock(&child_ctx->mutex);
3434 * Initialize the perf_counter context in task_struct
3436 int perf_counter_init_task(struct task_struct *child)
3438 struct perf_counter_context *child_ctx, *parent_ctx;
3439 struct perf_counter *counter;
3440 struct task_struct *parent = current;
3441 int inherited_all = 1;
3444 child->perf_counter_ctxp = NULL;
3446 mutex_init(&child->perf_counter_mutex);
3447 INIT_LIST_HEAD(&child->perf_counter_list);
3449 parent_ctx = parent->perf_counter_ctxp;
3450 if (likely(!parent_ctx || !parent_ctx->nr_counters))
3454 * This is executed from the parent task context, so inherit
3455 * counters that have been marked for cloning.
3456 * First allocate and initialize a context for the child.
3459 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
3463 __perf_counter_init_context(child_ctx, child);
3464 child->perf_counter_ctxp = child_ctx;
3467 * Lock the parent list. No need to lock the child - not PID
3468 * hashed yet and not running, so nobody can access it.
3470 mutex_lock(&parent_ctx->mutex);
3473 * We dont have to disable NMIs - we are only looking at
3474 * the list, not manipulating it:
3476 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
3477 if (counter != counter->group_leader)
3480 if (!counter->hw_event.inherit) {
3485 ret = inherit_group(counter, parent, parent_ctx,
3493 if (inherited_all) {
3495 * Mark the child context as a clone of the parent
3496 * context, or of whatever the parent is a clone of.
3498 if (parent_ctx->parent_ctx) {
3499 child_ctx->parent_ctx = parent_ctx->parent_ctx;
3500 child_ctx->parent_gen = parent_ctx->parent_gen;
3502 child_ctx->parent_ctx = parent_ctx;
3503 child_ctx->parent_gen = parent_ctx->generation;
3505 get_ctx(child_ctx->parent_ctx);
3508 mutex_unlock(&parent_ctx->mutex);
3513 static void __cpuinit perf_counter_init_cpu(int cpu)
3515 struct perf_cpu_context *cpuctx;
3517 cpuctx = &per_cpu(perf_cpu_context, cpu);
3518 __perf_counter_init_context(&cpuctx->ctx, NULL);
3520 spin_lock(&perf_resource_lock);
3521 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3522 spin_unlock(&perf_resource_lock);
3524 hw_perf_counter_setup(cpu);
3527 #ifdef CONFIG_HOTPLUG_CPU
3528 static void __perf_counter_exit_cpu(void *info)
3530 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3531 struct perf_counter_context *ctx = &cpuctx->ctx;
3532 struct perf_counter *counter, *tmp;
3534 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3535 __perf_counter_remove_from_context(counter);
3537 static void perf_counter_exit_cpu(int cpu)
3539 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3540 struct perf_counter_context *ctx = &cpuctx->ctx;
3542 mutex_lock(&ctx->mutex);
3543 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3544 mutex_unlock(&ctx->mutex);
3547 static inline void perf_counter_exit_cpu(int cpu) { }
3550 static int __cpuinit
3551 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3553 unsigned int cpu = (long)hcpu;
3557 case CPU_UP_PREPARE:
3558 case CPU_UP_PREPARE_FROZEN:
3559 perf_counter_init_cpu(cpu);
3562 case CPU_DOWN_PREPARE:
3563 case CPU_DOWN_PREPARE_FROZEN:
3564 perf_counter_exit_cpu(cpu);
3574 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3575 .notifier_call = perf_cpu_notify,
3578 void __init perf_counter_init(void)
3580 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3581 (void *)(long)smp_processor_id());
3582 register_cpu_notifier(&perf_cpu_nb);
3585 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3587 return sprintf(buf, "%d\n", perf_reserved_percpu);
3591 perf_set_reserve_percpu(struct sysdev_class *class,
3595 struct perf_cpu_context *cpuctx;
3599 err = strict_strtoul(buf, 10, &val);
3602 if (val > perf_max_counters)
3605 spin_lock(&perf_resource_lock);
3606 perf_reserved_percpu = val;
3607 for_each_online_cpu(cpu) {
3608 cpuctx = &per_cpu(perf_cpu_context, cpu);
3609 spin_lock_irq(&cpuctx->ctx.lock);
3610 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3611 perf_max_counters - perf_reserved_percpu);
3612 cpuctx->max_pertask = mpt;
3613 spin_unlock_irq(&cpuctx->ctx.lock);
3615 spin_unlock(&perf_resource_lock);
3620 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3622 return sprintf(buf, "%d\n", perf_overcommit);
3626 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3631 err = strict_strtoul(buf, 10, &val);
3637 spin_lock(&perf_resource_lock);
3638 perf_overcommit = val;
3639 spin_unlock(&perf_resource_lock);
3644 static SYSDEV_CLASS_ATTR(
3647 perf_show_reserve_percpu,
3648 perf_set_reserve_percpu
3651 static SYSDEV_CLASS_ATTR(
3654 perf_show_overcommit,
3658 static struct attribute *perfclass_attrs[] = {
3659 &attr_reserve_percpu.attr,
3660 &attr_overcommit.attr,
3664 static struct attribute_group perfclass_attr_group = {
3665 .attrs = perfclass_attrs,
3666 .name = "perf_counters",
3669 static int __init perf_counter_sysfs_init(void)
3671 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3672 &perfclass_attr_group);
3674 device_initcall(perf_counter_sysfs_init);