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 = 128; /* 'free' kb per counter */
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 u64 __weak hw_perf_save_disable(void) { return 0; }
64 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
65 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
66 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
67 struct perf_cpu_context *cpuctx,
68 struct perf_counter_context *ctx, int cpu)
73 void __weak perf_counter_print_debug(void) { }
76 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
78 struct perf_counter *group_leader = counter->group_leader;
81 * Depending on whether it is a standalone or sibling counter,
82 * add it straight to the context's counter list, or to the group
83 * leader's sibling list:
85 if (group_leader == counter)
86 list_add_tail(&counter->list_entry, &ctx->counter_list);
88 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
89 group_leader->nr_siblings++;
92 list_add_rcu(&counter->event_entry, &ctx->event_list);
96 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
98 struct perf_counter *sibling, *tmp;
100 list_del_init(&counter->list_entry);
101 list_del_rcu(&counter->event_entry);
103 if (counter->group_leader != counter)
104 counter->group_leader->nr_siblings--;
107 * If this was a group counter with sibling counters then
108 * upgrade the siblings to singleton counters by adding them
109 * to the context list directly:
111 list_for_each_entry_safe(sibling, tmp,
112 &counter->sibling_list, list_entry) {
114 list_move_tail(&sibling->list_entry, &ctx->counter_list);
115 sibling->group_leader = sibling;
120 counter_sched_out(struct perf_counter *counter,
121 struct perf_cpu_context *cpuctx,
122 struct perf_counter_context *ctx)
124 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
127 counter->state = PERF_COUNTER_STATE_INACTIVE;
128 counter->tstamp_stopped = ctx->time;
129 counter->pmu->disable(counter);
132 if (!is_software_counter(counter))
133 cpuctx->active_oncpu--;
135 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
136 cpuctx->exclusive = 0;
140 group_sched_out(struct perf_counter *group_counter,
141 struct perf_cpu_context *cpuctx,
142 struct perf_counter_context *ctx)
144 struct perf_counter *counter;
146 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
149 counter_sched_out(group_counter, cpuctx, ctx);
152 * Schedule out siblings (if any):
154 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
155 counter_sched_out(counter, cpuctx, ctx);
157 if (group_counter->hw_event.exclusive)
158 cpuctx->exclusive = 0;
162 * Cross CPU call to remove a performance counter
164 * We disable the counter on the hardware level first. After that we
165 * remove it from the context list.
167 static void __perf_counter_remove_from_context(void *info)
169 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
170 struct perf_counter *counter = info;
171 struct perf_counter_context *ctx = counter->ctx;
176 * If this is a task context, we need to check whether it is
177 * the current task context of this cpu. If not it has been
178 * scheduled out before the smp call arrived.
180 if (ctx->task && cpuctx->task_ctx != ctx)
183 spin_lock_irqsave(&ctx->lock, flags);
185 counter_sched_out(counter, cpuctx, ctx);
187 counter->task = NULL;
191 * Protect the list operation against NMI by disabling the
192 * counters on a global level. NOP for non NMI based counters.
194 perf_flags = hw_perf_save_disable();
195 list_del_counter(counter, ctx);
196 hw_perf_restore(perf_flags);
200 * Allow more per task counters with respect to the
203 cpuctx->max_pertask =
204 min(perf_max_counters - ctx->nr_counters,
205 perf_max_counters - perf_reserved_percpu);
208 spin_unlock_irqrestore(&ctx->lock, flags);
213 * Remove the counter from a task's (or a CPU's) list of counters.
215 * Must be called with counter->mutex and ctx->mutex held.
217 * CPU counters are removed with a smp call. For task counters we only
218 * call when the task is on a CPU.
220 static void perf_counter_remove_from_context(struct perf_counter *counter)
222 struct perf_counter_context *ctx = counter->ctx;
223 struct task_struct *task = ctx->task;
227 * Per cpu counters are removed via an smp call and
228 * the removal is always sucessful.
230 smp_call_function_single(counter->cpu,
231 __perf_counter_remove_from_context,
237 task_oncpu_function_call(task, __perf_counter_remove_from_context,
240 spin_lock_irq(&ctx->lock);
242 * If the context is active we need to retry the smp call.
244 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
245 spin_unlock_irq(&ctx->lock);
250 * The lock prevents that this context is scheduled in so we
251 * can remove the counter safely, if the call above did not
254 if (!list_empty(&counter->list_entry)) {
256 list_del_counter(counter, ctx);
257 counter->task = NULL;
259 spin_unlock_irq(&ctx->lock);
262 static inline u64 perf_clock(void)
264 return cpu_clock(smp_processor_id());
268 * Update the record of the current time in a context.
270 static void update_context_time(struct perf_counter_context *ctx)
272 u64 now = perf_clock();
274 ctx->time += now - ctx->timestamp;
275 ctx->timestamp = now;
279 * Update the total_time_enabled and total_time_running fields for a counter.
281 static void update_counter_times(struct perf_counter *counter)
283 struct perf_counter_context *ctx = counter->ctx;
286 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
289 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
291 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
292 run_end = counter->tstamp_stopped;
296 counter->total_time_running = run_end - counter->tstamp_running;
300 * Update total_time_enabled and total_time_running for all counters in a group.
302 static void update_group_times(struct perf_counter *leader)
304 struct perf_counter *counter;
306 update_counter_times(leader);
307 list_for_each_entry(counter, &leader->sibling_list, list_entry)
308 update_counter_times(counter);
312 * Cross CPU call to disable a performance counter
314 static void __perf_counter_disable(void *info)
316 struct perf_counter *counter = info;
317 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
318 struct perf_counter_context *ctx = counter->ctx;
322 * If this is a per-task counter, need to check whether this
323 * counter's task is the current task on this cpu.
325 if (ctx->task && cpuctx->task_ctx != ctx)
328 spin_lock_irqsave(&ctx->lock, flags);
331 * If the counter is on, turn it off.
332 * If it is in error state, leave it in error state.
334 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
335 update_context_time(ctx);
336 update_counter_times(counter);
337 if (counter == counter->group_leader)
338 group_sched_out(counter, cpuctx, ctx);
340 counter_sched_out(counter, cpuctx, ctx);
341 counter->state = PERF_COUNTER_STATE_OFF;
344 spin_unlock_irqrestore(&ctx->lock, flags);
350 static void perf_counter_disable(struct perf_counter *counter)
352 struct perf_counter_context *ctx = counter->ctx;
353 struct task_struct *task = ctx->task;
357 * Disable the counter on the cpu that it's on
359 smp_call_function_single(counter->cpu, __perf_counter_disable,
365 task_oncpu_function_call(task, __perf_counter_disable, counter);
367 spin_lock_irq(&ctx->lock);
369 * If the counter is still active, we need to retry the cross-call.
371 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
372 spin_unlock_irq(&ctx->lock);
377 * Since we have the lock this context can't be scheduled
378 * in, so we can change the state safely.
380 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
381 update_counter_times(counter);
382 counter->state = PERF_COUNTER_STATE_OFF;
385 spin_unlock_irq(&ctx->lock);
389 counter_sched_in(struct perf_counter *counter,
390 struct perf_cpu_context *cpuctx,
391 struct perf_counter_context *ctx,
394 if (counter->state <= PERF_COUNTER_STATE_OFF)
397 counter->state = PERF_COUNTER_STATE_ACTIVE;
398 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
400 * The new state must be visible before we turn it on in the hardware:
404 if (counter->pmu->enable(counter)) {
405 counter->state = PERF_COUNTER_STATE_INACTIVE;
410 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
412 if (!is_software_counter(counter))
413 cpuctx->active_oncpu++;
416 if (counter->hw_event.exclusive)
417 cpuctx->exclusive = 1;
423 * Return 1 for a group consisting entirely of software counters,
424 * 0 if the group contains any hardware counters.
426 static int is_software_only_group(struct perf_counter *leader)
428 struct perf_counter *counter;
430 if (!is_software_counter(leader))
433 list_for_each_entry(counter, &leader->sibling_list, list_entry)
434 if (!is_software_counter(counter))
441 * Work out whether we can put this counter group on the CPU now.
443 static int group_can_go_on(struct perf_counter *counter,
444 struct perf_cpu_context *cpuctx,
448 * Groups consisting entirely of software counters can always go on.
450 if (is_software_only_group(counter))
453 * If an exclusive group is already on, no other hardware
454 * counters can go on.
456 if (cpuctx->exclusive)
459 * If this group is exclusive and there are already
460 * counters on the CPU, it can't go on.
462 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
465 * Otherwise, try to add it if all previous groups were able
471 static void add_counter_to_ctx(struct perf_counter *counter,
472 struct perf_counter_context *ctx)
474 list_add_counter(counter, ctx);
476 counter->prev_state = PERF_COUNTER_STATE_OFF;
477 counter->tstamp_enabled = ctx->time;
478 counter->tstamp_running = ctx->time;
479 counter->tstamp_stopped = ctx->time;
483 * Cross CPU call to install and enable a performance counter
485 static void __perf_install_in_context(void *info)
487 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
488 struct perf_counter *counter = info;
489 struct perf_counter_context *ctx = counter->ctx;
490 struct perf_counter *leader = counter->group_leader;
491 int cpu = smp_processor_id();
497 * If this is a task context, we need to check whether it is
498 * the current task context of this cpu. If not it has been
499 * scheduled out before the smp call arrived.
501 if (ctx->task && cpuctx->task_ctx != ctx)
504 spin_lock_irqsave(&ctx->lock, flags);
505 update_context_time(ctx);
508 * Protect the list operation against NMI by disabling the
509 * counters on a global level. NOP for non NMI based counters.
511 perf_flags = hw_perf_save_disable();
513 add_counter_to_ctx(counter, ctx);
516 * Don't put the counter on if it is disabled or if
517 * it is in a group and the group isn't on.
519 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
520 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
524 * An exclusive counter can't go on if there are already active
525 * hardware counters, and no hardware counter can go on if there
526 * is already an exclusive counter on.
528 if (!group_can_go_on(counter, cpuctx, 1))
531 err = counter_sched_in(counter, cpuctx, ctx, cpu);
535 * This counter couldn't go on. If it is in a group
536 * then we have to pull the whole group off.
537 * If the counter group is pinned then put it in error state.
539 if (leader != counter)
540 group_sched_out(leader, cpuctx, ctx);
541 if (leader->hw_event.pinned) {
542 update_group_times(leader);
543 leader->state = PERF_COUNTER_STATE_ERROR;
547 if (!err && !ctx->task && cpuctx->max_pertask)
548 cpuctx->max_pertask--;
551 hw_perf_restore(perf_flags);
553 spin_unlock_irqrestore(&ctx->lock, flags);
557 * Attach a performance counter to a context
559 * First we add the counter to the list with the hardware enable bit
560 * in counter->hw_config cleared.
562 * If the counter is attached to a task which is on a CPU we use a smp
563 * call to enable it in the task context. The task might have been
564 * scheduled away, but we check this in the smp call again.
566 * Must be called with ctx->mutex held.
569 perf_install_in_context(struct perf_counter_context *ctx,
570 struct perf_counter *counter,
573 struct task_struct *task = ctx->task;
577 * Per cpu counters are installed via an smp call and
578 * the install is always sucessful.
580 smp_call_function_single(cpu, __perf_install_in_context,
585 counter->task = task;
587 task_oncpu_function_call(task, __perf_install_in_context,
590 spin_lock_irq(&ctx->lock);
592 * we need to retry the smp call.
594 if (ctx->is_active && list_empty(&counter->list_entry)) {
595 spin_unlock_irq(&ctx->lock);
600 * The lock prevents that this context is scheduled in so we
601 * can add the counter safely, if it the call above did not
604 if (list_empty(&counter->list_entry))
605 add_counter_to_ctx(counter, ctx);
606 spin_unlock_irq(&ctx->lock);
610 * Cross CPU call to enable a performance counter
612 static void __perf_counter_enable(void *info)
614 struct perf_counter *counter = info;
615 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
616 struct perf_counter_context *ctx = counter->ctx;
617 struct perf_counter *leader = counter->group_leader;
622 * If this is a per-task counter, need to check whether this
623 * counter's task is the current task on this cpu.
625 if (ctx->task && cpuctx->task_ctx != ctx)
628 spin_lock_irqsave(&ctx->lock, flags);
629 update_context_time(ctx);
631 counter->prev_state = counter->state;
632 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
634 counter->state = PERF_COUNTER_STATE_INACTIVE;
635 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
638 * If the counter is in a group and isn't the group leader,
639 * then don't put it on unless the group is on.
641 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
644 if (!group_can_go_on(counter, cpuctx, 1))
647 err = counter_sched_in(counter, cpuctx, ctx,
652 * If this counter can't go on and it's part of a
653 * group, then the whole group has to come off.
655 if (leader != counter)
656 group_sched_out(leader, cpuctx, ctx);
657 if (leader->hw_event.pinned) {
658 update_group_times(leader);
659 leader->state = PERF_COUNTER_STATE_ERROR;
664 spin_unlock_irqrestore(&ctx->lock, flags);
670 static void perf_counter_enable(struct perf_counter *counter)
672 struct perf_counter_context *ctx = counter->ctx;
673 struct task_struct *task = ctx->task;
677 * Enable the counter on the cpu that it's on
679 smp_call_function_single(counter->cpu, __perf_counter_enable,
684 spin_lock_irq(&ctx->lock);
685 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
689 * If the counter is in error state, clear that first.
690 * That way, if we see the counter in error state below, we
691 * know that it has gone back into error state, as distinct
692 * from the task having been scheduled away before the
693 * cross-call arrived.
695 if (counter->state == PERF_COUNTER_STATE_ERROR)
696 counter->state = PERF_COUNTER_STATE_OFF;
699 spin_unlock_irq(&ctx->lock);
700 task_oncpu_function_call(task, __perf_counter_enable, counter);
702 spin_lock_irq(&ctx->lock);
705 * If the context is active and the counter is still off,
706 * we need to retry the cross-call.
708 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
712 * Since we have the lock this context can't be scheduled
713 * in, so we can change the state safely.
715 if (counter->state == PERF_COUNTER_STATE_OFF) {
716 counter->state = PERF_COUNTER_STATE_INACTIVE;
717 counter->tstamp_enabled =
718 ctx->time - counter->total_time_enabled;
721 spin_unlock_irq(&ctx->lock);
724 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
727 * not supported on inherited counters
729 if (counter->hw_event.inherit)
732 atomic_add(refresh, &counter->event_limit);
733 perf_counter_enable(counter);
738 void __perf_counter_sched_out(struct perf_counter_context *ctx,
739 struct perf_cpu_context *cpuctx)
741 struct perf_counter *counter;
744 spin_lock(&ctx->lock);
746 if (likely(!ctx->nr_counters))
748 update_context_time(ctx);
750 flags = hw_perf_save_disable();
751 if (ctx->nr_active) {
752 list_for_each_entry(counter, &ctx->counter_list, list_entry)
753 group_sched_out(counter, cpuctx, ctx);
755 hw_perf_restore(flags);
757 spin_unlock(&ctx->lock);
761 * Called from scheduler to remove the counters of the current task,
762 * with interrupts disabled.
764 * We stop each counter and update the counter value in counter->count.
766 * This does not protect us against NMI, but disable()
767 * sets the disabled bit in the control field of counter _before_
768 * accessing the counter control register. If a NMI hits, then it will
769 * not restart the counter.
771 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
773 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
774 struct perf_counter_context *ctx = &task->perf_counter_ctx;
775 struct pt_regs *regs;
777 if (likely(!cpuctx->task_ctx))
780 update_context_time(ctx);
782 regs = task_pt_regs(task);
783 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
784 __perf_counter_sched_out(ctx, cpuctx);
786 cpuctx->task_ctx = NULL;
789 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
791 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
795 group_sched_in(struct perf_counter *group_counter,
796 struct perf_cpu_context *cpuctx,
797 struct perf_counter_context *ctx,
800 struct perf_counter *counter, *partial_group;
803 if (group_counter->state == PERF_COUNTER_STATE_OFF)
806 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
808 return ret < 0 ? ret : 0;
810 group_counter->prev_state = group_counter->state;
811 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
815 * Schedule in siblings as one group (if any):
817 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
818 counter->prev_state = counter->state;
819 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
820 partial_group = counter;
829 * Groups can be scheduled in as one unit only, so undo any
830 * partial group before returning:
832 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
833 if (counter == partial_group)
835 counter_sched_out(counter, cpuctx, ctx);
837 counter_sched_out(group_counter, cpuctx, ctx);
843 __perf_counter_sched_in(struct perf_counter_context *ctx,
844 struct perf_cpu_context *cpuctx, int cpu)
846 struct perf_counter *counter;
850 spin_lock(&ctx->lock);
852 if (likely(!ctx->nr_counters))
855 ctx->timestamp = perf_clock();
857 flags = hw_perf_save_disable();
860 * First go through the list and put on any pinned groups
861 * in order to give them the best chance of going on.
863 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
864 if (counter->state <= PERF_COUNTER_STATE_OFF ||
865 !counter->hw_event.pinned)
867 if (counter->cpu != -1 && counter->cpu != cpu)
870 if (group_can_go_on(counter, cpuctx, 1))
871 group_sched_in(counter, cpuctx, ctx, cpu);
874 * If this pinned group hasn't been scheduled,
875 * put it in error state.
877 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
878 update_group_times(counter);
879 counter->state = PERF_COUNTER_STATE_ERROR;
883 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
885 * Ignore counters in OFF or ERROR state, and
886 * ignore pinned counters since we did them already.
888 if (counter->state <= PERF_COUNTER_STATE_OFF ||
889 counter->hw_event.pinned)
893 * Listen to the 'cpu' scheduling filter constraint
896 if (counter->cpu != -1 && counter->cpu != cpu)
899 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
900 if (group_sched_in(counter, cpuctx, ctx, cpu))
904 hw_perf_restore(flags);
906 spin_unlock(&ctx->lock);
910 * Called from scheduler to add the counters of the current task
911 * with interrupts disabled.
913 * We restore the counter value and then enable it.
915 * This does not protect us against NMI, but enable()
916 * sets the enabled bit in the control field of counter _before_
917 * accessing the counter control register. If a NMI hits, then it will
918 * keep the counter running.
920 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
922 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
923 struct perf_counter_context *ctx = &task->perf_counter_ctx;
925 __perf_counter_sched_in(ctx, cpuctx, cpu);
926 cpuctx->task_ctx = ctx;
929 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
931 struct perf_counter_context *ctx = &cpuctx->ctx;
933 __perf_counter_sched_in(ctx, cpuctx, cpu);
936 int perf_counter_task_disable(void)
938 struct task_struct *curr = current;
939 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
940 struct perf_counter *counter;
945 if (likely(!ctx->nr_counters))
948 local_irq_save(flags);
949 cpu = smp_processor_id();
951 perf_counter_task_sched_out(curr, cpu);
953 spin_lock(&ctx->lock);
956 * Disable all the counters:
958 perf_flags = hw_perf_save_disable();
960 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
961 if (counter->state != PERF_COUNTER_STATE_ERROR) {
962 update_group_times(counter);
963 counter->state = PERF_COUNTER_STATE_OFF;
967 hw_perf_restore(perf_flags);
969 spin_unlock_irqrestore(&ctx->lock, flags);
974 int perf_counter_task_enable(void)
976 struct task_struct *curr = current;
977 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
978 struct perf_counter *counter;
983 if (likely(!ctx->nr_counters))
986 local_irq_save(flags);
987 cpu = smp_processor_id();
989 perf_counter_task_sched_out(curr, cpu);
991 spin_lock(&ctx->lock);
994 * Disable all the counters:
996 perf_flags = hw_perf_save_disable();
998 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
999 if (counter->state > PERF_COUNTER_STATE_OFF)
1001 counter->state = PERF_COUNTER_STATE_INACTIVE;
1002 counter->tstamp_enabled =
1003 ctx->time - counter->total_time_enabled;
1004 counter->hw_event.disabled = 0;
1006 hw_perf_restore(perf_flags);
1008 spin_unlock(&ctx->lock);
1010 perf_counter_task_sched_in(curr, cpu);
1012 local_irq_restore(flags);
1018 * Round-robin a context's counters:
1020 static void rotate_ctx(struct perf_counter_context *ctx)
1022 struct perf_counter *counter;
1025 if (!ctx->nr_counters)
1028 spin_lock(&ctx->lock);
1030 * Rotate the first entry last (works just fine for group counters too):
1032 perf_flags = hw_perf_save_disable();
1033 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1034 list_move_tail(&counter->list_entry, &ctx->counter_list);
1037 hw_perf_restore(perf_flags);
1039 spin_unlock(&ctx->lock);
1042 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1044 struct perf_cpu_context *cpuctx;
1045 struct perf_counter_context *ctx;
1047 if (!atomic_read(&nr_counters))
1050 cpuctx = &per_cpu(perf_cpu_context, cpu);
1051 ctx = &curr->perf_counter_ctx;
1053 perf_counter_cpu_sched_out(cpuctx);
1054 perf_counter_task_sched_out(curr, cpu);
1056 rotate_ctx(&cpuctx->ctx);
1059 perf_counter_cpu_sched_in(cpuctx, cpu);
1060 perf_counter_task_sched_in(curr, cpu);
1064 * Cross CPU call to read the hardware counter
1066 static void __read(void *info)
1068 struct perf_counter *counter = info;
1069 struct perf_counter_context *ctx = counter->ctx;
1070 unsigned long flags;
1072 local_irq_save(flags);
1074 update_context_time(ctx);
1075 counter->pmu->read(counter);
1076 update_counter_times(counter);
1077 local_irq_restore(flags);
1080 static u64 perf_counter_read(struct perf_counter *counter)
1083 * If counter is enabled and currently active on a CPU, update the
1084 * value in the counter structure:
1086 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1087 smp_call_function_single(counter->oncpu,
1088 __read, counter, 1);
1089 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1090 update_counter_times(counter);
1093 return atomic64_read(&counter->count);
1096 static void put_context(struct perf_counter_context *ctx)
1099 put_task_struct(ctx->task);
1102 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1104 struct perf_cpu_context *cpuctx;
1105 struct perf_counter_context *ctx;
1106 struct task_struct *task;
1109 * If cpu is not a wildcard then this is a percpu counter:
1112 /* Must be root to operate on a CPU counter: */
1113 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1114 return ERR_PTR(-EACCES);
1116 if (cpu < 0 || cpu > num_possible_cpus())
1117 return ERR_PTR(-EINVAL);
1120 * We could be clever and allow to attach a counter to an
1121 * offline CPU and activate it when the CPU comes up, but
1124 if (!cpu_isset(cpu, cpu_online_map))
1125 return ERR_PTR(-ENODEV);
1127 cpuctx = &per_cpu(perf_cpu_context, cpu);
1137 task = find_task_by_vpid(pid);
1139 get_task_struct(task);
1143 return ERR_PTR(-ESRCH);
1145 ctx = &task->perf_counter_ctx;
1148 /* Reuse ptrace permission checks for now. */
1149 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1151 return ERR_PTR(-EACCES);
1157 static void free_counter_rcu(struct rcu_head *head)
1159 struct perf_counter *counter;
1161 counter = container_of(head, struct perf_counter, rcu_head);
1165 static void perf_pending_sync(struct perf_counter *counter);
1167 static void free_counter(struct perf_counter *counter)
1169 perf_pending_sync(counter);
1171 atomic_dec(&nr_counters);
1172 if (counter->hw_event.mmap)
1173 atomic_dec(&nr_mmap_tracking);
1174 if (counter->hw_event.munmap)
1175 atomic_dec(&nr_munmap_tracking);
1176 if (counter->hw_event.comm)
1177 atomic_dec(&nr_comm_tracking);
1179 if (counter->destroy)
1180 counter->destroy(counter);
1182 call_rcu(&counter->rcu_head, free_counter_rcu);
1186 * Called when the last reference to the file is gone.
1188 static int perf_release(struct inode *inode, struct file *file)
1190 struct perf_counter *counter = file->private_data;
1191 struct perf_counter_context *ctx = counter->ctx;
1193 file->private_data = NULL;
1195 mutex_lock(&ctx->mutex);
1196 mutex_lock(&counter->mutex);
1198 perf_counter_remove_from_context(counter);
1200 mutex_unlock(&counter->mutex);
1201 mutex_unlock(&ctx->mutex);
1203 free_counter(counter);
1210 * Read the performance counter - simple non blocking version for now
1213 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1219 * Return end-of-file for a read on a counter that is in
1220 * error state (i.e. because it was pinned but it couldn't be
1221 * scheduled on to the CPU at some point).
1223 if (counter->state == PERF_COUNTER_STATE_ERROR)
1226 mutex_lock(&counter->mutex);
1227 values[0] = perf_counter_read(counter);
1229 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1230 values[n++] = counter->total_time_enabled +
1231 atomic64_read(&counter->child_total_time_enabled);
1232 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1233 values[n++] = counter->total_time_running +
1234 atomic64_read(&counter->child_total_time_running);
1235 mutex_unlock(&counter->mutex);
1237 if (count < n * sizeof(u64))
1239 count = n * sizeof(u64);
1241 if (copy_to_user(buf, values, count))
1248 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1250 struct perf_counter *counter = file->private_data;
1252 return perf_read_hw(counter, buf, count);
1255 static unsigned int perf_poll(struct file *file, poll_table *wait)
1257 struct perf_counter *counter = file->private_data;
1258 struct perf_mmap_data *data;
1259 unsigned int events = POLL_HUP;
1262 data = rcu_dereference(counter->data);
1264 events = atomic_xchg(&data->poll, 0);
1267 poll_wait(file, &counter->waitq, wait);
1272 static void perf_counter_reset(struct perf_counter *counter)
1274 (void)perf_counter_read(counter);
1275 atomic_set(&counter->count, 0);
1276 perf_counter_update_userpage(counter);
1279 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1280 void (*func)(struct perf_counter *))
1282 struct perf_counter_context *ctx = counter->ctx;
1283 struct perf_counter *sibling;
1285 spin_lock_irq(&ctx->lock);
1286 counter = counter->group_leader;
1289 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1291 spin_unlock_irq(&ctx->lock);
1294 static void perf_counter_for_each_child(struct perf_counter *counter,
1295 void (*func)(struct perf_counter *))
1297 struct perf_counter *child;
1299 mutex_lock(&counter->mutex);
1301 list_for_each_entry(child, &counter->child_list, child_list)
1303 mutex_unlock(&counter->mutex);
1306 static void perf_counter_for_each(struct perf_counter *counter,
1307 void (*func)(struct perf_counter *))
1309 struct perf_counter *child;
1311 mutex_lock(&counter->mutex);
1312 perf_counter_for_each_sibling(counter, func);
1313 list_for_each_entry(child, &counter->child_list, child_list)
1314 perf_counter_for_each_sibling(child, func);
1315 mutex_unlock(&counter->mutex);
1318 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1320 struct perf_counter *counter = file->private_data;
1321 void (*func)(struct perf_counter *);
1325 case PERF_COUNTER_IOC_ENABLE:
1326 func = perf_counter_enable;
1328 case PERF_COUNTER_IOC_DISABLE:
1329 func = perf_counter_disable;
1331 case PERF_COUNTER_IOC_RESET:
1332 func = perf_counter_reset;
1335 case PERF_COUNTER_IOC_REFRESH:
1336 return perf_counter_refresh(counter, arg);
1341 if (flags & PERF_IOC_FLAG_GROUP)
1342 perf_counter_for_each(counter, func);
1344 perf_counter_for_each_child(counter, func);
1350 * Callers need to ensure there can be no nesting of this function, otherwise
1351 * the seqlock logic goes bad. We can not serialize this because the arch
1352 * code calls this from NMI context.
1354 void perf_counter_update_userpage(struct perf_counter *counter)
1356 struct perf_mmap_data *data;
1357 struct perf_counter_mmap_page *userpg;
1360 data = rcu_dereference(counter->data);
1364 userpg = data->user_page;
1367 * Disable preemption so as to not let the corresponding user-space
1368 * spin too long if we get preempted.
1373 userpg->index = counter->hw.idx;
1374 userpg->offset = atomic64_read(&counter->count);
1375 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1376 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1385 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1387 struct perf_counter *counter = vma->vm_file->private_data;
1388 struct perf_mmap_data *data;
1389 int ret = VM_FAULT_SIGBUS;
1392 data = rcu_dereference(counter->data);
1396 if (vmf->pgoff == 0) {
1397 vmf->page = virt_to_page(data->user_page);
1399 int nr = vmf->pgoff - 1;
1401 if ((unsigned)nr > data->nr_pages)
1404 vmf->page = virt_to_page(data->data_pages[nr]);
1406 get_page(vmf->page);
1414 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1416 struct perf_mmap_data *data;
1420 WARN_ON(atomic_read(&counter->mmap_count));
1422 size = sizeof(struct perf_mmap_data);
1423 size += nr_pages * sizeof(void *);
1425 data = kzalloc(size, GFP_KERNEL);
1429 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1430 if (!data->user_page)
1431 goto fail_user_page;
1433 for (i = 0; i < nr_pages; i++) {
1434 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1435 if (!data->data_pages[i])
1436 goto fail_data_pages;
1439 data->nr_pages = nr_pages;
1440 atomic_set(&data->lock, -1);
1442 rcu_assign_pointer(counter->data, data);
1447 for (i--; i >= 0; i--)
1448 free_page((unsigned long)data->data_pages[i]);
1450 free_page((unsigned long)data->user_page);
1459 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1461 struct perf_mmap_data *data = container_of(rcu_head,
1462 struct perf_mmap_data, rcu_head);
1465 free_page((unsigned long)data->user_page);
1466 for (i = 0; i < data->nr_pages; i++)
1467 free_page((unsigned long)data->data_pages[i]);
1471 static void perf_mmap_data_free(struct perf_counter *counter)
1473 struct perf_mmap_data *data = counter->data;
1475 WARN_ON(atomic_read(&counter->mmap_count));
1477 rcu_assign_pointer(counter->data, NULL);
1478 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1481 static void perf_mmap_open(struct vm_area_struct *vma)
1483 struct perf_counter *counter = vma->vm_file->private_data;
1485 atomic_inc(&counter->mmap_count);
1488 static void perf_mmap_close(struct vm_area_struct *vma)
1490 struct perf_counter *counter = vma->vm_file->private_data;
1492 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1493 &counter->mmap_mutex)) {
1494 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1495 perf_mmap_data_free(counter);
1496 mutex_unlock(&counter->mmap_mutex);
1500 static struct vm_operations_struct perf_mmap_vmops = {
1501 .open = perf_mmap_open,
1502 .close = perf_mmap_close,
1503 .fault = perf_mmap_fault,
1506 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1508 struct perf_counter *counter = file->private_data;
1509 unsigned long vma_size;
1510 unsigned long nr_pages;
1511 unsigned long locked, lock_limit;
1515 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1518 vma_size = vma->vm_end - vma->vm_start;
1519 nr_pages = (vma_size / PAGE_SIZE) - 1;
1522 * If we have data pages ensure they're a power-of-two number, so we
1523 * can do bitmasks instead of modulo.
1525 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1528 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1531 if (vma->vm_pgoff != 0)
1534 mutex_lock(&counter->mmap_mutex);
1535 if (atomic_inc_not_zero(&counter->mmap_count)) {
1536 if (nr_pages != counter->data->nr_pages)
1541 extra = nr_pages /* + 1 only account the data pages */;
1542 extra -= sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1546 locked = vma->vm_mm->locked_vm + extra;
1548 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1549 lock_limit >>= PAGE_SHIFT;
1551 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1556 WARN_ON(counter->data);
1557 ret = perf_mmap_data_alloc(counter, nr_pages);
1561 atomic_set(&counter->mmap_count, 1);
1562 vma->vm_mm->locked_vm += extra;
1563 counter->data->nr_locked = extra;
1565 mutex_unlock(&counter->mmap_mutex);
1567 vma->vm_flags &= ~VM_MAYWRITE;
1568 vma->vm_flags |= VM_RESERVED;
1569 vma->vm_ops = &perf_mmap_vmops;
1574 static int perf_fasync(int fd, struct file *filp, int on)
1576 struct perf_counter *counter = filp->private_data;
1577 struct inode *inode = filp->f_path.dentry->d_inode;
1580 mutex_lock(&inode->i_mutex);
1581 retval = fasync_helper(fd, filp, on, &counter->fasync);
1582 mutex_unlock(&inode->i_mutex);
1590 static const struct file_operations perf_fops = {
1591 .release = perf_release,
1594 .unlocked_ioctl = perf_ioctl,
1595 .compat_ioctl = perf_ioctl,
1597 .fasync = perf_fasync,
1601 * Perf counter wakeup
1603 * If there's data, ensure we set the poll() state and publish everything
1604 * to user-space before waking everybody up.
1607 void perf_counter_wakeup(struct perf_counter *counter)
1609 wake_up_all(&counter->waitq);
1611 if (counter->pending_kill) {
1612 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1613 counter->pending_kill = 0;
1620 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1622 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1623 * single linked list and use cmpxchg() to add entries lockless.
1626 static void perf_pending_counter(struct perf_pending_entry *entry)
1628 struct perf_counter *counter = container_of(entry,
1629 struct perf_counter, pending);
1631 if (counter->pending_disable) {
1632 counter->pending_disable = 0;
1633 perf_counter_disable(counter);
1636 if (counter->pending_wakeup) {
1637 counter->pending_wakeup = 0;
1638 perf_counter_wakeup(counter);
1642 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1644 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1648 static void perf_pending_queue(struct perf_pending_entry *entry,
1649 void (*func)(struct perf_pending_entry *))
1651 struct perf_pending_entry **head;
1653 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1658 head = &get_cpu_var(perf_pending_head);
1661 entry->next = *head;
1662 } while (cmpxchg(head, entry->next, entry) != entry->next);
1664 set_perf_counter_pending();
1666 put_cpu_var(perf_pending_head);
1669 static int __perf_pending_run(void)
1671 struct perf_pending_entry *list;
1674 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1675 while (list != PENDING_TAIL) {
1676 void (*func)(struct perf_pending_entry *);
1677 struct perf_pending_entry *entry = list;
1684 * Ensure we observe the unqueue before we issue the wakeup,
1685 * so that we won't be waiting forever.
1686 * -- see perf_not_pending().
1697 static inline int perf_not_pending(struct perf_counter *counter)
1700 * If we flush on whatever cpu we run, there is a chance we don't
1704 __perf_pending_run();
1708 * Ensure we see the proper queue state before going to sleep
1709 * so that we do not miss the wakeup. -- see perf_pending_handle()
1712 return counter->pending.next == NULL;
1715 static void perf_pending_sync(struct perf_counter *counter)
1717 wait_event(counter->waitq, perf_not_pending(counter));
1720 void perf_counter_do_pending(void)
1722 __perf_pending_run();
1726 * Callchain support -- arch specific
1729 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1738 struct perf_output_handle {
1739 struct perf_counter *counter;
1740 struct perf_mmap_data *data;
1741 unsigned int offset;
1746 unsigned long flags;
1749 static void perf_output_wakeup(struct perf_output_handle *handle)
1751 atomic_set(&handle->data->poll, POLL_IN);
1754 handle->counter->pending_wakeup = 1;
1755 perf_pending_queue(&handle->counter->pending,
1756 perf_pending_counter);
1758 perf_counter_wakeup(handle->counter);
1762 * Curious locking construct.
1764 * We need to ensure a later event doesn't publish a head when a former
1765 * event isn't done writing. However since we need to deal with NMIs we
1766 * cannot fully serialize things.
1768 * What we do is serialize between CPUs so we only have to deal with NMI
1769 * nesting on a single CPU.
1771 * We only publish the head (and generate a wakeup) when the outer-most
1774 static void perf_output_lock(struct perf_output_handle *handle)
1776 struct perf_mmap_data *data = handle->data;
1781 local_irq_save(handle->flags);
1782 cpu = smp_processor_id();
1784 if (in_nmi() && atomic_read(&data->lock) == cpu)
1787 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1793 static void perf_output_unlock(struct perf_output_handle *handle)
1795 struct perf_mmap_data *data = handle->data;
1798 data->done_head = data->head;
1800 if (!handle->locked)
1805 * The xchg implies a full barrier that ensures all writes are done
1806 * before we publish the new head, matched by a rmb() in userspace when
1807 * reading this position.
1809 while ((head = atomic_xchg(&data->done_head, 0)))
1810 data->user_page->data_head = head;
1813 * NMI can happen here, which means we can miss a done_head update.
1816 cpu = atomic_xchg(&data->lock, -1);
1817 WARN_ON_ONCE(cpu != smp_processor_id());
1820 * Therefore we have to validate we did not indeed do so.
1822 if (unlikely(atomic_read(&data->done_head))) {
1824 * Since we had it locked, we can lock it again.
1826 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1832 if (atomic_xchg(&data->wakeup, 0))
1833 perf_output_wakeup(handle);
1835 local_irq_restore(handle->flags);
1838 static int perf_output_begin(struct perf_output_handle *handle,
1839 struct perf_counter *counter, unsigned int size,
1840 int nmi, int overflow)
1842 struct perf_mmap_data *data;
1843 unsigned int offset, head;
1846 * For inherited counters we send all the output towards the parent.
1848 if (counter->parent)
1849 counter = counter->parent;
1852 data = rcu_dereference(counter->data);
1856 handle->data = data;
1857 handle->counter = counter;
1859 handle->overflow = overflow;
1861 if (!data->nr_pages)
1864 perf_output_lock(handle);
1867 offset = head = atomic_read(&data->head);
1869 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1871 handle->offset = offset;
1872 handle->head = head;
1874 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
1875 atomic_set(&data->wakeup, 1);
1880 perf_output_wakeup(handle);
1887 static void perf_output_copy(struct perf_output_handle *handle,
1888 void *buf, unsigned int len)
1890 unsigned int pages_mask;
1891 unsigned int offset;
1895 offset = handle->offset;
1896 pages_mask = handle->data->nr_pages - 1;
1897 pages = handle->data->data_pages;
1900 unsigned int page_offset;
1903 nr = (offset >> PAGE_SHIFT) & pages_mask;
1904 page_offset = offset & (PAGE_SIZE - 1);
1905 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1907 memcpy(pages[nr] + page_offset, buf, size);
1914 handle->offset = offset;
1916 WARN_ON_ONCE(handle->offset > handle->head);
1919 #define perf_output_put(handle, x) \
1920 perf_output_copy((handle), &(x), sizeof(x))
1922 static void perf_output_end(struct perf_output_handle *handle)
1924 struct perf_counter *counter = handle->counter;
1925 struct perf_mmap_data *data = handle->data;
1927 int wakeup_events = counter->hw_event.wakeup_events;
1929 if (handle->overflow && wakeup_events) {
1930 int events = atomic_inc_return(&data->events);
1931 if (events >= wakeup_events) {
1932 atomic_sub(wakeup_events, &data->events);
1933 atomic_set(&data->wakeup, 1);
1937 perf_output_unlock(handle);
1941 static void perf_counter_output(struct perf_counter *counter,
1942 int nmi, struct pt_regs *regs, u64 addr)
1945 u64 record_type = counter->hw_event.record_type;
1946 struct perf_output_handle handle;
1947 struct perf_event_header header;
1956 struct perf_callchain_entry *callchain = NULL;
1957 int callchain_size = 0;
1964 header.size = sizeof(header);
1966 header.misc = PERF_EVENT_MISC_OVERFLOW;
1967 header.misc |= user_mode(regs) ?
1968 PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
1970 if (record_type & PERF_RECORD_IP) {
1971 ip = instruction_pointer(regs);
1972 header.type |= PERF_RECORD_IP;
1973 header.size += sizeof(ip);
1976 if (record_type & PERF_RECORD_TID) {
1977 /* namespace issues */
1978 tid_entry.pid = current->group_leader->pid;
1979 tid_entry.tid = current->pid;
1981 header.type |= PERF_RECORD_TID;
1982 header.size += sizeof(tid_entry);
1985 if (record_type & PERF_RECORD_TIME) {
1987 * Maybe do better on x86 and provide cpu_clock_nmi()
1989 time = sched_clock();
1991 header.type |= PERF_RECORD_TIME;
1992 header.size += sizeof(u64);
1995 if (record_type & PERF_RECORD_ADDR) {
1996 header.type |= PERF_RECORD_ADDR;
1997 header.size += sizeof(u64);
2000 if (record_type & PERF_RECORD_CONFIG) {
2001 header.type |= PERF_RECORD_CONFIG;
2002 header.size += sizeof(u64);
2005 if (record_type & PERF_RECORD_CPU) {
2006 header.type |= PERF_RECORD_CPU;
2007 header.size += sizeof(cpu_entry);
2009 cpu_entry.cpu = raw_smp_processor_id();
2012 if (record_type & PERF_RECORD_GROUP) {
2013 header.type |= PERF_RECORD_GROUP;
2014 header.size += sizeof(u64) +
2015 counter->nr_siblings * sizeof(group_entry);
2018 if (record_type & PERF_RECORD_CALLCHAIN) {
2019 callchain = perf_callchain(regs);
2022 callchain_size = (1 + callchain->nr) * sizeof(u64);
2024 header.type |= PERF_RECORD_CALLCHAIN;
2025 header.size += callchain_size;
2029 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2033 perf_output_put(&handle, header);
2035 if (record_type & PERF_RECORD_IP)
2036 perf_output_put(&handle, ip);
2038 if (record_type & PERF_RECORD_TID)
2039 perf_output_put(&handle, tid_entry);
2041 if (record_type & PERF_RECORD_TIME)
2042 perf_output_put(&handle, time);
2044 if (record_type & PERF_RECORD_ADDR)
2045 perf_output_put(&handle, addr);
2047 if (record_type & PERF_RECORD_CONFIG)
2048 perf_output_put(&handle, counter->hw_event.config);
2050 if (record_type & PERF_RECORD_CPU)
2051 perf_output_put(&handle, cpu_entry);
2054 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2056 if (record_type & PERF_RECORD_GROUP) {
2057 struct perf_counter *leader, *sub;
2058 u64 nr = counter->nr_siblings;
2060 perf_output_put(&handle, nr);
2062 leader = counter->group_leader;
2063 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2065 sub->pmu->read(sub);
2067 group_entry.event = sub->hw_event.config;
2068 group_entry.counter = atomic64_read(&sub->count);
2070 perf_output_put(&handle, group_entry);
2075 perf_output_copy(&handle, callchain, callchain_size);
2077 perf_output_end(&handle);
2084 struct perf_comm_event {
2085 struct task_struct *task;
2090 struct perf_event_header header;
2097 static void perf_counter_comm_output(struct perf_counter *counter,
2098 struct perf_comm_event *comm_event)
2100 struct perf_output_handle handle;
2101 int size = comm_event->event.header.size;
2102 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2107 perf_output_put(&handle, comm_event->event);
2108 perf_output_copy(&handle, comm_event->comm,
2109 comm_event->comm_size);
2110 perf_output_end(&handle);
2113 static int perf_counter_comm_match(struct perf_counter *counter,
2114 struct perf_comm_event *comm_event)
2116 if (counter->hw_event.comm &&
2117 comm_event->event.header.type == PERF_EVENT_COMM)
2123 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2124 struct perf_comm_event *comm_event)
2126 struct perf_counter *counter;
2128 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2132 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2133 if (perf_counter_comm_match(counter, comm_event))
2134 perf_counter_comm_output(counter, comm_event);
2139 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2141 struct perf_cpu_context *cpuctx;
2143 char *comm = comm_event->task->comm;
2145 size = ALIGN(strlen(comm)+1, sizeof(u64));
2147 comm_event->comm = comm;
2148 comm_event->comm_size = size;
2150 comm_event->event.header.size = sizeof(comm_event->event) + size;
2152 cpuctx = &get_cpu_var(perf_cpu_context);
2153 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2154 put_cpu_var(perf_cpu_context);
2156 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
2159 void perf_counter_comm(struct task_struct *task)
2161 struct perf_comm_event comm_event;
2163 if (!atomic_read(&nr_comm_tracking))
2166 comm_event = (struct perf_comm_event){
2169 .header = { .type = PERF_EVENT_COMM, },
2170 .pid = task->group_leader->pid,
2175 perf_counter_comm_event(&comm_event);
2182 struct perf_mmap_event {
2188 struct perf_event_header header;
2198 static void perf_counter_mmap_output(struct perf_counter *counter,
2199 struct perf_mmap_event *mmap_event)
2201 struct perf_output_handle handle;
2202 int size = mmap_event->event.header.size;
2203 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2208 perf_output_put(&handle, mmap_event->event);
2209 perf_output_copy(&handle, mmap_event->file_name,
2210 mmap_event->file_size);
2211 perf_output_end(&handle);
2214 static int perf_counter_mmap_match(struct perf_counter *counter,
2215 struct perf_mmap_event *mmap_event)
2217 if (counter->hw_event.mmap &&
2218 mmap_event->event.header.type == PERF_EVENT_MMAP)
2221 if (counter->hw_event.munmap &&
2222 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2228 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2229 struct perf_mmap_event *mmap_event)
2231 struct perf_counter *counter;
2233 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2237 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2238 if (perf_counter_mmap_match(counter, mmap_event))
2239 perf_counter_mmap_output(counter, mmap_event);
2244 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2246 struct perf_cpu_context *cpuctx;
2247 struct file *file = mmap_event->file;
2254 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2256 name = strncpy(tmp, "//enomem", sizeof(tmp));
2259 name = d_path(&file->f_path, buf, PATH_MAX);
2261 name = strncpy(tmp, "//toolong", sizeof(tmp));
2265 name = strncpy(tmp, "//anon", sizeof(tmp));
2270 size = ALIGN(strlen(name)+1, sizeof(u64));
2272 mmap_event->file_name = name;
2273 mmap_event->file_size = size;
2275 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2277 cpuctx = &get_cpu_var(perf_cpu_context);
2278 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2279 put_cpu_var(perf_cpu_context);
2281 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2286 void perf_counter_mmap(unsigned long addr, unsigned long len,
2287 unsigned long pgoff, struct file *file)
2289 struct perf_mmap_event mmap_event;
2291 if (!atomic_read(&nr_mmap_tracking))
2294 mmap_event = (struct perf_mmap_event){
2297 .header = { .type = PERF_EVENT_MMAP, },
2298 .pid = current->group_leader->pid,
2299 .tid = current->pid,
2306 perf_counter_mmap_event(&mmap_event);
2309 void perf_counter_munmap(unsigned long addr, unsigned long len,
2310 unsigned long pgoff, struct file *file)
2312 struct perf_mmap_event mmap_event;
2314 if (!atomic_read(&nr_munmap_tracking))
2317 mmap_event = (struct perf_mmap_event){
2320 .header = { .type = PERF_EVENT_MUNMAP, },
2321 .pid = current->group_leader->pid,
2322 .tid = current->pid,
2329 perf_counter_mmap_event(&mmap_event);
2333 * Generic counter overflow handling.
2336 int perf_counter_overflow(struct perf_counter *counter,
2337 int nmi, struct pt_regs *regs, u64 addr)
2339 int events = atomic_read(&counter->event_limit);
2343 * XXX event_limit might not quite work as expected on inherited
2347 counter->pending_kill = POLL_IN;
2348 if (events && atomic_dec_and_test(&counter->event_limit)) {
2350 counter->pending_kill = POLL_HUP;
2352 counter->pending_disable = 1;
2353 perf_pending_queue(&counter->pending,
2354 perf_pending_counter);
2356 perf_counter_disable(counter);
2359 perf_counter_output(counter, nmi, regs, addr);
2364 * Generic software counter infrastructure
2367 static void perf_swcounter_update(struct perf_counter *counter)
2369 struct hw_perf_counter *hwc = &counter->hw;
2374 prev = atomic64_read(&hwc->prev_count);
2375 now = atomic64_read(&hwc->count);
2376 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2381 atomic64_add(delta, &counter->count);
2382 atomic64_sub(delta, &hwc->period_left);
2385 static void perf_swcounter_set_period(struct perf_counter *counter)
2387 struct hw_perf_counter *hwc = &counter->hw;
2388 s64 left = atomic64_read(&hwc->period_left);
2389 s64 period = hwc->irq_period;
2391 if (unlikely(left <= -period)) {
2393 atomic64_set(&hwc->period_left, left);
2396 if (unlikely(left <= 0)) {
2398 atomic64_add(period, &hwc->period_left);
2401 atomic64_set(&hwc->prev_count, -left);
2402 atomic64_set(&hwc->count, -left);
2405 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2407 enum hrtimer_restart ret = HRTIMER_RESTART;
2408 struct perf_counter *counter;
2409 struct pt_regs *regs;
2411 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2412 counter->pmu->read(counter);
2414 regs = get_irq_regs();
2416 * In case we exclude kernel IPs or are somehow not in interrupt
2417 * context, provide the next best thing, the user IP.
2419 if ((counter->hw_event.exclude_kernel || !regs) &&
2420 !counter->hw_event.exclude_user)
2421 regs = task_pt_regs(current);
2424 if (perf_counter_overflow(counter, 0, regs, 0))
2425 ret = HRTIMER_NORESTART;
2428 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2433 static void perf_swcounter_overflow(struct perf_counter *counter,
2434 int nmi, struct pt_regs *regs, u64 addr)
2436 perf_swcounter_update(counter);
2437 perf_swcounter_set_period(counter);
2438 if (perf_counter_overflow(counter, nmi, regs, addr))
2439 /* soft-disable the counter */
2444 static int perf_swcounter_match(struct perf_counter *counter,
2445 enum perf_event_types type,
2446 u32 event, struct pt_regs *regs)
2448 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2451 if (perf_event_raw(&counter->hw_event))
2454 if (perf_event_type(&counter->hw_event) != type)
2457 if (perf_event_id(&counter->hw_event) != event)
2460 if (counter->hw_event.exclude_user && user_mode(regs))
2463 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2469 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2470 int nmi, struct pt_regs *regs, u64 addr)
2472 int neg = atomic64_add_negative(nr, &counter->hw.count);
2473 if (counter->hw.irq_period && !neg)
2474 perf_swcounter_overflow(counter, nmi, regs, addr);
2477 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2478 enum perf_event_types type, u32 event,
2479 u64 nr, int nmi, struct pt_regs *regs,
2482 struct perf_counter *counter;
2484 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2488 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2489 if (perf_swcounter_match(counter, type, event, regs))
2490 perf_swcounter_add(counter, nr, nmi, regs, addr);
2495 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2498 return &cpuctx->recursion[3];
2501 return &cpuctx->recursion[2];
2504 return &cpuctx->recursion[1];
2506 return &cpuctx->recursion[0];
2509 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2510 u64 nr, int nmi, struct pt_regs *regs,
2513 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2514 int *recursion = perf_swcounter_recursion_context(cpuctx);
2522 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2523 nr, nmi, regs, addr);
2524 if (cpuctx->task_ctx) {
2525 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2526 nr, nmi, regs, addr);
2533 put_cpu_var(perf_cpu_context);
2537 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2539 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2542 static void perf_swcounter_read(struct perf_counter *counter)
2544 perf_swcounter_update(counter);
2547 static int perf_swcounter_enable(struct perf_counter *counter)
2549 perf_swcounter_set_period(counter);
2553 static void perf_swcounter_disable(struct perf_counter *counter)
2555 perf_swcounter_update(counter);
2558 static const struct pmu perf_ops_generic = {
2559 .enable = perf_swcounter_enable,
2560 .disable = perf_swcounter_disable,
2561 .read = perf_swcounter_read,
2565 * Software counter: cpu wall time clock
2568 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2570 int cpu = raw_smp_processor_id();
2574 now = cpu_clock(cpu);
2575 prev = atomic64_read(&counter->hw.prev_count);
2576 atomic64_set(&counter->hw.prev_count, now);
2577 atomic64_add(now - prev, &counter->count);
2580 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2582 struct hw_perf_counter *hwc = &counter->hw;
2583 int cpu = raw_smp_processor_id();
2585 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2586 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2587 hwc->hrtimer.function = perf_swcounter_hrtimer;
2588 if (hwc->irq_period) {
2589 __hrtimer_start_range_ns(&hwc->hrtimer,
2590 ns_to_ktime(hwc->irq_period), 0,
2591 HRTIMER_MODE_REL, 0);
2597 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2599 hrtimer_cancel(&counter->hw.hrtimer);
2600 cpu_clock_perf_counter_update(counter);
2603 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2605 cpu_clock_perf_counter_update(counter);
2608 static const struct pmu perf_ops_cpu_clock = {
2609 .enable = cpu_clock_perf_counter_enable,
2610 .disable = cpu_clock_perf_counter_disable,
2611 .read = cpu_clock_perf_counter_read,
2615 * Software counter: task time clock
2618 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2623 prev = atomic64_xchg(&counter->hw.prev_count, now);
2625 atomic64_add(delta, &counter->count);
2628 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2630 struct hw_perf_counter *hwc = &counter->hw;
2633 now = counter->ctx->time;
2635 atomic64_set(&hwc->prev_count, now);
2636 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2637 hwc->hrtimer.function = perf_swcounter_hrtimer;
2638 if (hwc->irq_period) {
2639 __hrtimer_start_range_ns(&hwc->hrtimer,
2640 ns_to_ktime(hwc->irq_period), 0,
2641 HRTIMER_MODE_REL, 0);
2647 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2649 hrtimer_cancel(&counter->hw.hrtimer);
2650 task_clock_perf_counter_update(counter, counter->ctx->time);
2654 static void task_clock_perf_counter_read(struct perf_counter *counter)
2659 update_context_time(counter->ctx);
2660 time = counter->ctx->time;
2662 u64 now = perf_clock();
2663 u64 delta = now - counter->ctx->timestamp;
2664 time = counter->ctx->time + delta;
2667 task_clock_perf_counter_update(counter, time);
2670 static const struct pmu perf_ops_task_clock = {
2671 .enable = task_clock_perf_counter_enable,
2672 .disable = task_clock_perf_counter_disable,
2673 .read = task_clock_perf_counter_read,
2677 * Software counter: cpu migrations
2680 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2682 struct task_struct *curr = counter->ctx->task;
2685 return curr->se.nr_migrations;
2686 return cpu_nr_migrations(smp_processor_id());
2689 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2694 prev = atomic64_read(&counter->hw.prev_count);
2695 now = get_cpu_migrations(counter);
2697 atomic64_set(&counter->hw.prev_count, now);
2701 atomic64_add(delta, &counter->count);
2704 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2706 cpu_migrations_perf_counter_update(counter);
2709 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2711 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2712 atomic64_set(&counter->hw.prev_count,
2713 get_cpu_migrations(counter));
2717 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2719 cpu_migrations_perf_counter_update(counter);
2722 static const struct pmu perf_ops_cpu_migrations = {
2723 .enable = cpu_migrations_perf_counter_enable,
2724 .disable = cpu_migrations_perf_counter_disable,
2725 .read = cpu_migrations_perf_counter_read,
2728 #ifdef CONFIG_EVENT_PROFILE
2729 void perf_tpcounter_event(int event_id)
2731 struct pt_regs *regs = get_irq_regs();
2734 regs = task_pt_regs(current);
2736 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2738 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2740 extern int ftrace_profile_enable(int);
2741 extern void ftrace_profile_disable(int);
2743 static void tp_perf_counter_destroy(struct perf_counter *counter)
2745 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2748 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2750 int event_id = perf_event_id(&counter->hw_event);
2753 ret = ftrace_profile_enable(event_id);
2757 counter->destroy = tp_perf_counter_destroy;
2758 counter->hw.irq_period = counter->hw_event.irq_period;
2760 return &perf_ops_generic;
2763 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2769 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2771 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2772 const struct pmu *pmu = NULL;
2773 struct hw_perf_counter *hwc = &counter->hw;
2776 * Software counters (currently) can't in general distinguish
2777 * between user, kernel and hypervisor events.
2778 * However, context switches and cpu migrations are considered
2779 * to be kernel events, and page faults are never hypervisor
2782 switch (perf_event_id(&counter->hw_event)) {
2783 case PERF_COUNT_CPU_CLOCK:
2784 pmu = &perf_ops_cpu_clock;
2786 if (hw_event->irq_period && hw_event->irq_period < 10000)
2787 hw_event->irq_period = 10000;
2789 case PERF_COUNT_TASK_CLOCK:
2791 * If the user instantiates this as a per-cpu counter,
2792 * use the cpu_clock counter instead.
2794 if (counter->ctx->task)
2795 pmu = &perf_ops_task_clock;
2797 pmu = &perf_ops_cpu_clock;
2799 if (hw_event->irq_period && hw_event->irq_period < 10000)
2800 hw_event->irq_period = 10000;
2802 case PERF_COUNT_PAGE_FAULTS:
2803 case PERF_COUNT_PAGE_FAULTS_MIN:
2804 case PERF_COUNT_PAGE_FAULTS_MAJ:
2805 case PERF_COUNT_CONTEXT_SWITCHES:
2806 pmu = &perf_ops_generic;
2808 case PERF_COUNT_CPU_MIGRATIONS:
2809 if (!counter->hw_event.exclude_kernel)
2810 pmu = &perf_ops_cpu_migrations;
2815 hwc->irq_period = hw_event->irq_period;
2821 * Allocate and initialize a counter structure
2823 static struct perf_counter *
2824 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2826 struct perf_counter_context *ctx,
2827 struct perf_counter *group_leader,
2830 const struct pmu *pmu;
2831 struct perf_counter *counter;
2834 counter = kzalloc(sizeof(*counter), gfpflags);
2836 return ERR_PTR(-ENOMEM);
2839 * Single counters are their own group leaders, with an
2840 * empty sibling list:
2843 group_leader = counter;
2845 mutex_init(&counter->mutex);
2846 INIT_LIST_HEAD(&counter->list_entry);
2847 INIT_LIST_HEAD(&counter->event_entry);
2848 INIT_LIST_HEAD(&counter->sibling_list);
2849 init_waitqueue_head(&counter->waitq);
2851 mutex_init(&counter->mmap_mutex);
2853 INIT_LIST_HEAD(&counter->child_list);
2856 counter->hw_event = *hw_event;
2857 counter->group_leader = group_leader;
2858 counter->pmu = NULL;
2861 counter->state = PERF_COUNTER_STATE_INACTIVE;
2862 if (hw_event->disabled)
2863 counter->state = PERF_COUNTER_STATE_OFF;
2868 * we currently do not support PERF_RECORD_GROUP on inherited counters
2870 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
2873 if (perf_event_raw(hw_event)) {
2874 pmu = hw_perf_counter_init(counter);
2878 switch (perf_event_type(hw_event)) {
2879 case PERF_TYPE_HARDWARE:
2880 pmu = hw_perf_counter_init(counter);
2883 case PERF_TYPE_SOFTWARE:
2884 pmu = sw_perf_counter_init(counter);
2887 case PERF_TYPE_TRACEPOINT:
2888 pmu = tp_perf_counter_init(counter);
2895 else if (IS_ERR(pmu))
2900 return ERR_PTR(err);
2905 atomic_inc(&nr_counters);
2906 if (counter->hw_event.mmap)
2907 atomic_inc(&nr_mmap_tracking);
2908 if (counter->hw_event.munmap)
2909 atomic_inc(&nr_munmap_tracking);
2910 if (counter->hw_event.comm)
2911 atomic_inc(&nr_comm_tracking);
2917 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2919 * @hw_event_uptr: event type attributes for monitoring/sampling
2922 * @group_fd: group leader counter fd
2924 SYSCALL_DEFINE5(perf_counter_open,
2925 const struct perf_counter_hw_event __user *, hw_event_uptr,
2926 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2928 struct perf_counter *counter, *group_leader;
2929 struct perf_counter_hw_event hw_event;
2930 struct perf_counter_context *ctx;
2931 struct file *counter_file = NULL;
2932 struct file *group_file = NULL;
2933 int fput_needed = 0;
2934 int fput_needed2 = 0;
2937 /* for future expandability... */
2941 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2945 * Get the target context (task or percpu):
2947 ctx = find_get_context(pid, cpu);
2949 return PTR_ERR(ctx);
2952 * Look up the group leader (we will attach this counter to it):
2954 group_leader = NULL;
2955 if (group_fd != -1) {
2957 group_file = fget_light(group_fd, &fput_needed);
2959 goto err_put_context;
2960 if (group_file->f_op != &perf_fops)
2961 goto err_put_context;
2963 group_leader = group_file->private_data;
2965 * Do not allow a recursive hierarchy (this new sibling
2966 * becoming part of another group-sibling):
2968 if (group_leader->group_leader != group_leader)
2969 goto err_put_context;
2971 * Do not allow to attach to a group in a different
2972 * task or CPU context:
2974 if (group_leader->ctx != ctx)
2975 goto err_put_context;
2977 * Only a group leader can be exclusive or pinned
2979 if (hw_event.exclusive || hw_event.pinned)
2980 goto err_put_context;
2983 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2985 ret = PTR_ERR(counter);
2986 if (IS_ERR(counter))
2987 goto err_put_context;
2989 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2991 goto err_free_put_context;
2993 counter_file = fget_light(ret, &fput_needed2);
2995 goto err_free_put_context;
2997 counter->filp = counter_file;
2998 mutex_lock(&ctx->mutex);
2999 perf_install_in_context(ctx, counter, cpu);
3000 mutex_unlock(&ctx->mutex);
3002 fput_light(counter_file, fput_needed2);
3005 fput_light(group_file, fput_needed);
3009 err_free_put_context:
3019 * Initialize the perf_counter context in a task_struct:
3022 __perf_counter_init_context(struct perf_counter_context *ctx,
3023 struct task_struct *task)
3025 memset(ctx, 0, sizeof(*ctx));
3026 spin_lock_init(&ctx->lock);
3027 mutex_init(&ctx->mutex);
3028 INIT_LIST_HEAD(&ctx->counter_list);
3029 INIT_LIST_HEAD(&ctx->event_list);
3034 * inherit a counter from parent task to child task:
3036 static struct perf_counter *
3037 inherit_counter(struct perf_counter *parent_counter,
3038 struct task_struct *parent,
3039 struct perf_counter_context *parent_ctx,
3040 struct task_struct *child,
3041 struct perf_counter *group_leader,
3042 struct perf_counter_context *child_ctx)
3044 struct perf_counter *child_counter;
3047 * Instead of creating recursive hierarchies of counters,
3048 * we link inherited counters back to the original parent,
3049 * which has a filp for sure, which we use as the reference
3052 if (parent_counter->parent)
3053 parent_counter = parent_counter->parent;
3055 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3056 parent_counter->cpu, child_ctx,
3057 group_leader, GFP_KERNEL);
3058 if (IS_ERR(child_counter))
3059 return child_counter;
3062 * Link it up in the child's context:
3064 child_counter->task = child;
3065 add_counter_to_ctx(child_counter, child_ctx);
3067 child_counter->parent = parent_counter;
3069 * inherit into child's child as well:
3071 child_counter->hw_event.inherit = 1;
3074 * Get a reference to the parent filp - we will fput it
3075 * when the child counter exits. This is safe to do because
3076 * we are in the parent and we know that the filp still
3077 * exists and has a nonzero count:
3079 atomic_long_inc(&parent_counter->filp->f_count);
3082 * Link this into the parent counter's child list
3084 mutex_lock(&parent_counter->mutex);
3085 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3088 * Make the child state follow the state of the parent counter,
3089 * not its hw_event.disabled bit. We hold the parent's mutex,
3090 * so we won't race with perf_counter_{en,dis}able_family.
3092 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3093 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3095 child_counter->state = PERF_COUNTER_STATE_OFF;
3097 mutex_unlock(&parent_counter->mutex);
3099 return child_counter;
3102 static int inherit_group(struct perf_counter *parent_counter,
3103 struct task_struct *parent,
3104 struct perf_counter_context *parent_ctx,
3105 struct task_struct *child,
3106 struct perf_counter_context *child_ctx)
3108 struct perf_counter *leader;
3109 struct perf_counter *sub;
3110 struct perf_counter *child_ctr;
3112 leader = inherit_counter(parent_counter, parent, parent_ctx,
3113 child, NULL, child_ctx);
3115 return PTR_ERR(leader);
3116 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3117 child_ctr = inherit_counter(sub, parent, parent_ctx,
3118 child, leader, child_ctx);
3119 if (IS_ERR(child_ctr))
3120 return PTR_ERR(child_ctr);
3125 static void sync_child_counter(struct perf_counter *child_counter,
3126 struct perf_counter *parent_counter)
3128 u64 parent_val, child_val;
3130 parent_val = atomic64_read(&parent_counter->count);
3131 child_val = atomic64_read(&child_counter->count);
3134 * Add back the child's count to the parent's count:
3136 atomic64_add(child_val, &parent_counter->count);
3137 atomic64_add(child_counter->total_time_enabled,
3138 &parent_counter->child_total_time_enabled);
3139 atomic64_add(child_counter->total_time_running,
3140 &parent_counter->child_total_time_running);
3143 * Remove this counter from the parent's list
3145 mutex_lock(&parent_counter->mutex);
3146 list_del_init(&child_counter->child_list);
3147 mutex_unlock(&parent_counter->mutex);
3150 * Release the parent counter, if this was the last
3153 fput(parent_counter->filp);
3157 __perf_counter_exit_task(struct task_struct *child,
3158 struct perf_counter *child_counter,
3159 struct perf_counter_context *child_ctx)
3161 struct perf_counter *parent_counter;
3162 struct perf_counter *sub, *tmp;
3165 * If we do not self-reap then we have to wait for the
3166 * child task to unschedule (it will happen for sure),
3167 * so that its counter is at its final count. (This
3168 * condition triggers rarely - child tasks usually get
3169 * off their CPU before the parent has a chance to
3170 * get this far into the reaping action)
3172 if (child != current) {
3173 wait_task_inactive(child, 0);
3174 list_del_init(&child_counter->list_entry);
3175 update_counter_times(child_counter);
3177 struct perf_cpu_context *cpuctx;
3178 unsigned long flags;
3182 * Disable and unlink this counter.
3184 * Be careful about zapping the list - IRQ/NMI context
3185 * could still be processing it:
3187 local_irq_save(flags);
3188 perf_flags = hw_perf_save_disable();
3190 cpuctx = &__get_cpu_var(perf_cpu_context);
3192 group_sched_out(child_counter, cpuctx, child_ctx);
3193 update_counter_times(child_counter);
3195 list_del_init(&child_counter->list_entry);
3197 child_ctx->nr_counters--;
3199 hw_perf_restore(perf_flags);
3200 local_irq_restore(flags);
3203 parent_counter = child_counter->parent;
3205 * It can happen that parent exits first, and has counters
3206 * that are still around due to the child reference. These
3207 * counters need to be zapped - but otherwise linger.
3209 if (parent_counter) {
3210 sync_child_counter(child_counter, parent_counter);
3211 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
3214 sync_child_counter(sub, sub->parent);
3218 free_counter(child_counter);
3223 * When a child task exits, feed back counter values to parent counters.
3225 * Note: we may be running in child context, but the PID is not hashed
3226 * anymore so new counters will not be added.
3228 void perf_counter_exit_task(struct task_struct *child)
3230 struct perf_counter *child_counter, *tmp;
3231 struct perf_counter_context *child_ctx;
3233 child_ctx = &child->perf_counter_ctx;
3235 if (likely(!child_ctx->nr_counters))
3238 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3240 __perf_counter_exit_task(child, child_counter, child_ctx);
3244 * Initialize the perf_counter context in task_struct
3246 void perf_counter_init_task(struct task_struct *child)
3248 struct perf_counter_context *child_ctx, *parent_ctx;
3249 struct perf_counter *counter;
3250 struct task_struct *parent = current;
3252 child_ctx = &child->perf_counter_ctx;
3253 parent_ctx = &parent->perf_counter_ctx;
3255 __perf_counter_init_context(child_ctx, child);
3258 * This is executed from the parent task context, so inherit
3259 * counters that have been marked for cloning:
3262 if (likely(!parent_ctx->nr_counters))
3266 * Lock the parent list. No need to lock the child - not PID
3267 * hashed yet and not running, so nobody can access it.
3269 mutex_lock(&parent_ctx->mutex);
3272 * We dont have to disable NMIs - we are only looking at
3273 * the list, not manipulating it:
3275 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
3276 if (!counter->hw_event.inherit)
3279 if (inherit_group(counter, parent,
3280 parent_ctx, child, child_ctx))
3284 mutex_unlock(&parent_ctx->mutex);
3287 static void __cpuinit perf_counter_init_cpu(int cpu)
3289 struct perf_cpu_context *cpuctx;
3291 cpuctx = &per_cpu(perf_cpu_context, cpu);
3292 __perf_counter_init_context(&cpuctx->ctx, NULL);
3294 spin_lock(&perf_resource_lock);
3295 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3296 spin_unlock(&perf_resource_lock);
3298 hw_perf_counter_setup(cpu);
3301 #ifdef CONFIG_HOTPLUG_CPU
3302 static void __perf_counter_exit_cpu(void *info)
3304 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3305 struct perf_counter_context *ctx = &cpuctx->ctx;
3306 struct perf_counter *counter, *tmp;
3308 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3309 __perf_counter_remove_from_context(counter);
3311 static void perf_counter_exit_cpu(int cpu)
3313 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3314 struct perf_counter_context *ctx = &cpuctx->ctx;
3316 mutex_lock(&ctx->mutex);
3317 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3318 mutex_unlock(&ctx->mutex);
3321 static inline void perf_counter_exit_cpu(int cpu) { }
3324 static int __cpuinit
3325 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3327 unsigned int cpu = (long)hcpu;
3331 case CPU_UP_PREPARE:
3332 case CPU_UP_PREPARE_FROZEN:
3333 perf_counter_init_cpu(cpu);
3336 case CPU_DOWN_PREPARE:
3337 case CPU_DOWN_PREPARE_FROZEN:
3338 perf_counter_exit_cpu(cpu);
3348 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3349 .notifier_call = perf_cpu_notify,
3352 void __init perf_counter_init(void)
3354 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3355 (void *)(long)smp_processor_id());
3356 register_cpu_notifier(&perf_cpu_nb);
3359 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3361 return sprintf(buf, "%d\n", perf_reserved_percpu);
3365 perf_set_reserve_percpu(struct sysdev_class *class,
3369 struct perf_cpu_context *cpuctx;
3373 err = strict_strtoul(buf, 10, &val);
3376 if (val > perf_max_counters)
3379 spin_lock(&perf_resource_lock);
3380 perf_reserved_percpu = val;
3381 for_each_online_cpu(cpu) {
3382 cpuctx = &per_cpu(perf_cpu_context, cpu);
3383 spin_lock_irq(&cpuctx->ctx.lock);
3384 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3385 perf_max_counters - perf_reserved_percpu);
3386 cpuctx->max_pertask = mpt;
3387 spin_unlock_irq(&cpuctx->ctx.lock);
3389 spin_unlock(&perf_resource_lock);
3394 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3396 return sprintf(buf, "%d\n", perf_overcommit);
3400 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3405 err = strict_strtoul(buf, 10, &val);
3411 spin_lock(&perf_resource_lock);
3412 perf_overcommit = val;
3413 spin_unlock(&perf_resource_lock);
3418 static SYSDEV_CLASS_ATTR(
3421 perf_show_reserve_percpu,
3422 perf_set_reserve_percpu
3425 static SYSDEV_CLASS_ATTR(
3428 perf_show_overcommit,
3432 static struct attribute *perfclass_attrs[] = {
3433 &attr_reserve_percpu.attr,
3434 &attr_overcommit.attr,
3438 static struct attribute_group perfclass_attr_group = {
3439 .attrs = perfclass_attrs,
3440 .name = "perf_counters",
3443 static int __init perf_counter_sysfs_init(void)
3445 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3446 &perfclass_attr_group);
3448 device_initcall(perf_counter_sysfs_init);