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 group_sched_in(struct perf_counter *group_counter,
424 struct perf_cpu_context *cpuctx,
425 struct perf_counter_context *ctx,
428 struct perf_counter *counter, *partial_group;
431 if (group_counter->state == PERF_COUNTER_STATE_OFF)
434 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
436 return ret < 0 ? ret : 0;
438 group_counter->prev_state = group_counter->state;
439 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
443 * Schedule in siblings as one group (if any):
445 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
446 counter->prev_state = counter->state;
447 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
448 partial_group = counter;
457 * Groups can be scheduled in as one unit only, so undo any
458 * partial group before returning:
460 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
461 if (counter == partial_group)
463 counter_sched_out(counter, cpuctx, ctx);
465 counter_sched_out(group_counter, cpuctx, ctx);
471 * Return 1 for a group consisting entirely of software counters,
472 * 0 if the group contains any hardware counters.
474 static int is_software_only_group(struct perf_counter *leader)
476 struct perf_counter *counter;
478 if (!is_software_counter(leader))
481 list_for_each_entry(counter, &leader->sibling_list, list_entry)
482 if (!is_software_counter(counter))
489 * Work out whether we can put this counter group on the CPU now.
491 static int group_can_go_on(struct perf_counter *counter,
492 struct perf_cpu_context *cpuctx,
496 * Groups consisting entirely of software counters can always go on.
498 if (is_software_only_group(counter))
501 * If an exclusive group is already on, no other hardware
502 * counters can go on.
504 if (cpuctx->exclusive)
507 * If this group is exclusive and there are already
508 * counters on the CPU, it can't go on.
510 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
513 * Otherwise, try to add it if all previous groups were able
519 static void add_counter_to_ctx(struct perf_counter *counter,
520 struct perf_counter_context *ctx)
522 list_add_counter(counter, ctx);
524 counter->prev_state = PERF_COUNTER_STATE_OFF;
525 counter->tstamp_enabled = ctx->time;
526 counter->tstamp_running = ctx->time;
527 counter->tstamp_stopped = ctx->time;
531 * Cross CPU call to install and enable a performance counter
533 static void __perf_install_in_context(void *info)
535 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
536 struct perf_counter *counter = info;
537 struct perf_counter_context *ctx = counter->ctx;
538 struct perf_counter *leader = counter->group_leader;
539 int cpu = smp_processor_id();
545 * If this is a task context, we need to check whether it is
546 * the current task context of this cpu. If not it has been
547 * scheduled out before the smp call arrived.
549 if (ctx->task && cpuctx->task_ctx != ctx)
552 spin_lock_irqsave(&ctx->lock, flags);
553 update_context_time(ctx);
556 * Protect the list operation against NMI by disabling the
557 * counters on a global level. NOP for non NMI based counters.
559 perf_flags = hw_perf_save_disable();
561 add_counter_to_ctx(counter, ctx);
564 * Don't put the counter on if it is disabled or if
565 * it is in a group and the group isn't on.
567 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
568 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
572 * An exclusive counter can't go on if there are already active
573 * hardware counters, and no hardware counter can go on if there
574 * is already an exclusive counter on.
576 if (!group_can_go_on(counter, cpuctx, 1))
579 err = counter_sched_in(counter, cpuctx, ctx, cpu);
583 * This counter couldn't go on. If it is in a group
584 * then we have to pull the whole group off.
585 * If the counter group is pinned then put it in error state.
587 if (leader != counter)
588 group_sched_out(leader, cpuctx, ctx);
589 if (leader->hw_event.pinned) {
590 update_group_times(leader);
591 leader->state = PERF_COUNTER_STATE_ERROR;
595 if (!err && !ctx->task && cpuctx->max_pertask)
596 cpuctx->max_pertask--;
599 hw_perf_restore(perf_flags);
601 spin_unlock_irqrestore(&ctx->lock, flags);
605 * Attach a performance counter to a context
607 * First we add the counter to the list with the hardware enable bit
608 * in counter->hw_config cleared.
610 * If the counter is attached to a task which is on a CPU we use a smp
611 * call to enable it in the task context. The task might have been
612 * scheduled away, but we check this in the smp call again.
614 * Must be called with ctx->mutex held.
617 perf_install_in_context(struct perf_counter_context *ctx,
618 struct perf_counter *counter,
621 struct task_struct *task = ctx->task;
625 * Per cpu counters are installed via an smp call and
626 * the install is always sucessful.
628 smp_call_function_single(cpu, __perf_install_in_context,
633 counter->task = task;
635 task_oncpu_function_call(task, __perf_install_in_context,
638 spin_lock_irq(&ctx->lock);
640 * we need to retry the smp call.
642 if (ctx->is_active && list_empty(&counter->list_entry)) {
643 spin_unlock_irq(&ctx->lock);
648 * The lock prevents that this context is scheduled in so we
649 * can add the counter safely, if it the call above did not
652 if (list_empty(&counter->list_entry))
653 add_counter_to_ctx(counter, ctx);
654 spin_unlock_irq(&ctx->lock);
658 * Cross CPU call to enable a performance counter
660 static void __perf_counter_enable(void *info)
662 struct perf_counter *counter = info;
663 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
664 struct perf_counter_context *ctx = counter->ctx;
665 struct perf_counter *leader = counter->group_leader;
670 * If this is a per-task counter, need to check whether this
671 * counter's task is the current task on this cpu.
673 if (ctx->task && cpuctx->task_ctx != ctx)
676 spin_lock_irqsave(&ctx->lock, flags);
677 update_context_time(ctx);
679 counter->prev_state = counter->state;
680 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
682 counter->state = PERF_COUNTER_STATE_INACTIVE;
683 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
686 * If the counter is in a group and isn't the group leader,
687 * then don't put it on unless the group is on.
689 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
692 if (!group_can_go_on(counter, cpuctx, 1))
694 else if (counter == leader)
695 err = group_sched_in(counter, cpuctx, ctx,
698 err = counter_sched_in(counter, cpuctx, ctx,
703 * If this counter can't go on and it's part of a
704 * group, then the whole group has to come off.
706 if (leader != counter)
707 group_sched_out(leader, cpuctx, ctx);
708 if (leader->hw_event.pinned) {
709 update_group_times(leader);
710 leader->state = PERF_COUNTER_STATE_ERROR;
715 spin_unlock_irqrestore(&ctx->lock, flags);
721 static void perf_counter_enable(struct perf_counter *counter)
723 struct perf_counter_context *ctx = counter->ctx;
724 struct task_struct *task = ctx->task;
728 * Enable the counter on the cpu that it's on
730 smp_call_function_single(counter->cpu, __perf_counter_enable,
735 spin_lock_irq(&ctx->lock);
736 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
740 * If the counter is in error state, clear that first.
741 * That way, if we see the counter in error state below, we
742 * know that it has gone back into error state, as distinct
743 * from the task having been scheduled away before the
744 * cross-call arrived.
746 if (counter->state == PERF_COUNTER_STATE_ERROR)
747 counter->state = PERF_COUNTER_STATE_OFF;
750 spin_unlock_irq(&ctx->lock);
751 task_oncpu_function_call(task, __perf_counter_enable, counter);
753 spin_lock_irq(&ctx->lock);
756 * If the context is active and the counter is still off,
757 * we need to retry the cross-call.
759 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
763 * Since we have the lock this context can't be scheduled
764 * in, so we can change the state safely.
766 if (counter->state == PERF_COUNTER_STATE_OFF) {
767 counter->state = PERF_COUNTER_STATE_INACTIVE;
768 counter->tstamp_enabled =
769 ctx->time - counter->total_time_enabled;
772 spin_unlock_irq(&ctx->lock);
775 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
778 * not supported on inherited counters
780 if (counter->hw_event.inherit)
783 atomic_add(refresh, &counter->event_limit);
784 perf_counter_enable(counter);
789 void __perf_counter_sched_out(struct perf_counter_context *ctx,
790 struct perf_cpu_context *cpuctx)
792 struct perf_counter *counter;
795 spin_lock(&ctx->lock);
797 if (likely(!ctx->nr_counters))
799 update_context_time(ctx);
801 flags = hw_perf_save_disable();
802 if (ctx->nr_active) {
803 list_for_each_entry(counter, &ctx->counter_list, list_entry)
804 group_sched_out(counter, cpuctx, ctx);
806 hw_perf_restore(flags);
808 spin_unlock(&ctx->lock);
812 * Called from scheduler to remove the counters of the current task,
813 * with interrupts disabled.
815 * We stop each counter and update the counter value in counter->count.
817 * This does not protect us against NMI, but disable()
818 * sets the disabled bit in the control field of counter _before_
819 * accessing the counter control register. If a NMI hits, then it will
820 * not restart the counter.
822 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
824 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
825 struct perf_counter_context *ctx = &task->perf_counter_ctx;
826 struct pt_regs *regs;
828 if (likely(!cpuctx->task_ctx))
831 update_context_time(ctx);
833 regs = task_pt_regs(task);
834 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
835 __perf_counter_sched_out(ctx, cpuctx);
837 cpuctx->task_ctx = NULL;
840 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
842 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
844 __perf_counter_sched_out(ctx, cpuctx);
845 cpuctx->task_ctx = NULL;
848 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
850 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
854 __perf_counter_sched_in(struct perf_counter_context *ctx,
855 struct perf_cpu_context *cpuctx, int cpu)
857 struct perf_counter *counter;
861 spin_lock(&ctx->lock);
863 if (likely(!ctx->nr_counters))
866 ctx->timestamp = perf_clock();
868 flags = hw_perf_save_disable();
871 * First go through the list and put on any pinned groups
872 * in order to give them the best chance of going on.
874 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
875 if (counter->state <= PERF_COUNTER_STATE_OFF ||
876 !counter->hw_event.pinned)
878 if (counter->cpu != -1 && counter->cpu != cpu)
881 if (group_can_go_on(counter, cpuctx, 1))
882 group_sched_in(counter, cpuctx, ctx, cpu);
885 * If this pinned group hasn't been scheduled,
886 * put it in error state.
888 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
889 update_group_times(counter);
890 counter->state = PERF_COUNTER_STATE_ERROR;
894 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
896 * Ignore counters in OFF or ERROR state, and
897 * ignore pinned counters since we did them already.
899 if (counter->state <= PERF_COUNTER_STATE_OFF ||
900 counter->hw_event.pinned)
904 * Listen to the 'cpu' scheduling filter constraint
907 if (counter->cpu != -1 && counter->cpu != cpu)
910 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
911 if (group_sched_in(counter, cpuctx, ctx, cpu))
915 hw_perf_restore(flags);
917 spin_unlock(&ctx->lock);
921 * Called from scheduler to add the counters of the current task
922 * with interrupts disabled.
924 * We restore the counter value and then enable it.
926 * This does not protect us against NMI, but enable()
927 * sets the enabled bit in the control field of counter _before_
928 * accessing the counter control register. If a NMI hits, then it will
929 * keep the counter running.
931 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
933 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
934 struct perf_counter_context *ctx = &task->perf_counter_ctx;
936 __perf_counter_sched_in(ctx, cpuctx, cpu);
937 cpuctx->task_ctx = ctx;
940 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
942 struct perf_counter_context *ctx = &cpuctx->ctx;
944 __perf_counter_sched_in(ctx, cpuctx, cpu);
947 int perf_counter_task_disable(void)
949 struct task_struct *curr = current;
950 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
951 struct perf_counter *counter;
955 if (likely(!ctx->nr_counters))
958 local_irq_save(flags);
960 __perf_counter_task_sched_out(ctx);
962 spin_lock(&ctx->lock);
965 * Disable all the counters:
967 perf_flags = hw_perf_save_disable();
969 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
970 if (counter->state != PERF_COUNTER_STATE_ERROR) {
971 update_group_times(counter);
972 counter->state = PERF_COUNTER_STATE_OFF;
976 hw_perf_restore(perf_flags);
978 spin_unlock_irqrestore(&ctx->lock, flags);
983 int perf_counter_task_enable(void)
985 struct task_struct *curr = current;
986 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
987 struct perf_counter *counter;
992 if (likely(!ctx->nr_counters))
995 local_irq_save(flags);
996 cpu = smp_processor_id();
998 __perf_counter_task_sched_out(ctx);
1000 spin_lock(&ctx->lock);
1003 * Disable all the counters:
1005 perf_flags = hw_perf_save_disable();
1007 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1008 if (counter->state > PERF_COUNTER_STATE_OFF)
1010 counter->state = PERF_COUNTER_STATE_INACTIVE;
1011 counter->tstamp_enabled =
1012 ctx->time - counter->total_time_enabled;
1013 counter->hw_event.disabled = 0;
1015 hw_perf_restore(perf_flags);
1017 spin_unlock(&ctx->lock);
1019 perf_counter_task_sched_in(curr, cpu);
1021 local_irq_restore(flags);
1027 * Round-robin a context's counters:
1029 static void rotate_ctx(struct perf_counter_context *ctx)
1031 struct perf_counter *counter;
1034 if (!ctx->nr_counters)
1037 spin_lock(&ctx->lock);
1039 * Rotate the first entry last (works just fine for group counters too):
1041 perf_flags = hw_perf_save_disable();
1042 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1043 list_move_tail(&counter->list_entry, &ctx->counter_list);
1046 hw_perf_restore(perf_flags);
1048 spin_unlock(&ctx->lock);
1051 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1053 struct perf_cpu_context *cpuctx;
1054 struct perf_counter_context *ctx;
1056 if (!atomic_read(&nr_counters))
1059 cpuctx = &per_cpu(perf_cpu_context, cpu);
1060 ctx = &curr->perf_counter_ctx;
1062 perf_counter_cpu_sched_out(cpuctx);
1063 __perf_counter_task_sched_out(ctx);
1065 rotate_ctx(&cpuctx->ctx);
1068 perf_counter_cpu_sched_in(cpuctx, cpu);
1069 perf_counter_task_sched_in(curr, cpu);
1073 * Cross CPU call to read the hardware counter
1075 static void __read(void *info)
1077 struct perf_counter *counter = info;
1078 struct perf_counter_context *ctx = counter->ctx;
1079 unsigned long flags;
1081 local_irq_save(flags);
1083 update_context_time(ctx);
1084 counter->pmu->read(counter);
1085 update_counter_times(counter);
1086 local_irq_restore(flags);
1089 static u64 perf_counter_read(struct perf_counter *counter)
1092 * If counter is enabled and currently active on a CPU, update the
1093 * value in the counter structure:
1095 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1096 smp_call_function_single(counter->oncpu,
1097 __read, counter, 1);
1098 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1099 update_counter_times(counter);
1102 return atomic64_read(&counter->count);
1105 static void put_context(struct perf_counter_context *ctx)
1108 put_task_struct(ctx->task);
1111 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1113 struct perf_cpu_context *cpuctx;
1114 struct perf_counter_context *ctx;
1115 struct task_struct *task;
1118 * If cpu is not a wildcard then this is a percpu counter:
1121 /* Must be root to operate on a CPU counter: */
1122 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1123 return ERR_PTR(-EACCES);
1125 if (cpu < 0 || cpu > num_possible_cpus())
1126 return ERR_PTR(-EINVAL);
1129 * We could be clever and allow to attach a counter to an
1130 * offline CPU and activate it when the CPU comes up, but
1133 if (!cpu_isset(cpu, cpu_online_map))
1134 return ERR_PTR(-ENODEV);
1136 cpuctx = &per_cpu(perf_cpu_context, cpu);
1146 task = find_task_by_vpid(pid);
1148 get_task_struct(task);
1152 return ERR_PTR(-ESRCH);
1154 ctx = &task->perf_counter_ctx;
1157 /* Reuse ptrace permission checks for now. */
1158 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1160 return ERR_PTR(-EACCES);
1166 static void free_counter_rcu(struct rcu_head *head)
1168 struct perf_counter *counter;
1170 counter = container_of(head, struct perf_counter, rcu_head);
1174 static void perf_pending_sync(struct perf_counter *counter);
1176 static void free_counter(struct perf_counter *counter)
1178 perf_pending_sync(counter);
1180 atomic_dec(&nr_counters);
1181 if (counter->hw_event.mmap)
1182 atomic_dec(&nr_mmap_tracking);
1183 if (counter->hw_event.munmap)
1184 atomic_dec(&nr_munmap_tracking);
1185 if (counter->hw_event.comm)
1186 atomic_dec(&nr_comm_tracking);
1188 if (counter->destroy)
1189 counter->destroy(counter);
1191 call_rcu(&counter->rcu_head, free_counter_rcu);
1195 * Called when the last reference to the file is gone.
1197 static int perf_release(struct inode *inode, struct file *file)
1199 struct perf_counter *counter = file->private_data;
1200 struct perf_counter_context *ctx = counter->ctx;
1202 file->private_data = NULL;
1204 mutex_lock(&ctx->mutex);
1205 mutex_lock(&counter->mutex);
1207 perf_counter_remove_from_context(counter);
1209 mutex_unlock(&counter->mutex);
1210 mutex_unlock(&ctx->mutex);
1212 free_counter(counter);
1219 * Read the performance counter - simple non blocking version for now
1222 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1228 * Return end-of-file for a read on a counter that is in
1229 * error state (i.e. because it was pinned but it couldn't be
1230 * scheduled on to the CPU at some point).
1232 if (counter->state == PERF_COUNTER_STATE_ERROR)
1235 mutex_lock(&counter->mutex);
1236 values[0] = perf_counter_read(counter);
1238 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1239 values[n++] = counter->total_time_enabled +
1240 atomic64_read(&counter->child_total_time_enabled);
1241 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1242 values[n++] = counter->total_time_running +
1243 atomic64_read(&counter->child_total_time_running);
1244 mutex_unlock(&counter->mutex);
1246 if (count < n * sizeof(u64))
1248 count = n * sizeof(u64);
1250 if (copy_to_user(buf, values, count))
1257 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1259 struct perf_counter *counter = file->private_data;
1261 return perf_read_hw(counter, buf, count);
1264 static unsigned int perf_poll(struct file *file, poll_table *wait)
1266 struct perf_counter *counter = file->private_data;
1267 struct perf_mmap_data *data;
1268 unsigned int events = POLL_HUP;
1271 data = rcu_dereference(counter->data);
1273 events = atomic_xchg(&data->poll, 0);
1276 poll_wait(file, &counter->waitq, wait);
1281 static void perf_counter_reset(struct perf_counter *counter)
1283 (void)perf_counter_read(counter);
1284 atomic64_set(&counter->count, 0);
1285 perf_counter_update_userpage(counter);
1288 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1289 void (*func)(struct perf_counter *))
1291 struct perf_counter_context *ctx = counter->ctx;
1292 struct perf_counter *sibling;
1294 spin_lock_irq(&ctx->lock);
1295 counter = counter->group_leader;
1298 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1300 spin_unlock_irq(&ctx->lock);
1303 static void perf_counter_for_each_child(struct perf_counter *counter,
1304 void (*func)(struct perf_counter *))
1306 struct perf_counter *child;
1308 mutex_lock(&counter->mutex);
1310 list_for_each_entry(child, &counter->child_list, child_list)
1312 mutex_unlock(&counter->mutex);
1315 static void perf_counter_for_each(struct perf_counter *counter,
1316 void (*func)(struct perf_counter *))
1318 struct perf_counter *child;
1320 mutex_lock(&counter->mutex);
1321 perf_counter_for_each_sibling(counter, func);
1322 list_for_each_entry(child, &counter->child_list, child_list)
1323 perf_counter_for_each_sibling(child, func);
1324 mutex_unlock(&counter->mutex);
1327 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1329 struct perf_counter *counter = file->private_data;
1330 void (*func)(struct perf_counter *);
1334 case PERF_COUNTER_IOC_ENABLE:
1335 func = perf_counter_enable;
1337 case PERF_COUNTER_IOC_DISABLE:
1338 func = perf_counter_disable;
1340 case PERF_COUNTER_IOC_RESET:
1341 func = perf_counter_reset;
1344 case PERF_COUNTER_IOC_REFRESH:
1345 return perf_counter_refresh(counter, arg);
1350 if (flags & PERF_IOC_FLAG_GROUP)
1351 perf_counter_for_each(counter, func);
1353 perf_counter_for_each_child(counter, func);
1359 * Callers need to ensure there can be no nesting of this function, otherwise
1360 * the seqlock logic goes bad. We can not serialize this because the arch
1361 * code calls this from NMI context.
1363 void perf_counter_update_userpage(struct perf_counter *counter)
1365 struct perf_mmap_data *data;
1366 struct perf_counter_mmap_page *userpg;
1369 data = rcu_dereference(counter->data);
1373 userpg = data->user_page;
1376 * Disable preemption so as to not let the corresponding user-space
1377 * spin too long if we get preempted.
1382 userpg->index = counter->hw.idx;
1383 userpg->offset = atomic64_read(&counter->count);
1384 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1385 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1394 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1396 struct perf_counter *counter = vma->vm_file->private_data;
1397 struct perf_mmap_data *data;
1398 int ret = VM_FAULT_SIGBUS;
1401 data = rcu_dereference(counter->data);
1405 if (vmf->pgoff == 0) {
1406 vmf->page = virt_to_page(data->user_page);
1408 int nr = vmf->pgoff - 1;
1410 if ((unsigned)nr > data->nr_pages)
1413 vmf->page = virt_to_page(data->data_pages[nr]);
1415 get_page(vmf->page);
1423 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1425 struct perf_mmap_data *data;
1429 WARN_ON(atomic_read(&counter->mmap_count));
1431 size = sizeof(struct perf_mmap_data);
1432 size += nr_pages * sizeof(void *);
1434 data = kzalloc(size, GFP_KERNEL);
1438 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1439 if (!data->user_page)
1440 goto fail_user_page;
1442 for (i = 0; i < nr_pages; i++) {
1443 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1444 if (!data->data_pages[i])
1445 goto fail_data_pages;
1448 data->nr_pages = nr_pages;
1449 atomic_set(&data->lock, -1);
1451 rcu_assign_pointer(counter->data, data);
1456 for (i--; i >= 0; i--)
1457 free_page((unsigned long)data->data_pages[i]);
1459 free_page((unsigned long)data->user_page);
1468 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1470 struct perf_mmap_data *data = container_of(rcu_head,
1471 struct perf_mmap_data, rcu_head);
1474 free_page((unsigned long)data->user_page);
1475 for (i = 0; i < data->nr_pages; i++)
1476 free_page((unsigned long)data->data_pages[i]);
1480 static void perf_mmap_data_free(struct perf_counter *counter)
1482 struct perf_mmap_data *data = counter->data;
1484 WARN_ON(atomic_read(&counter->mmap_count));
1486 rcu_assign_pointer(counter->data, NULL);
1487 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1490 static void perf_mmap_open(struct vm_area_struct *vma)
1492 struct perf_counter *counter = vma->vm_file->private_data;
1494 atomic_inc(&counter->mmap_count);
1497 static void perf_mmap_close(struct vm_area_struct *vma)
1499 struct perf_counter *counter = vma->vm_file->private_data;
1501 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1502 &counter->mmap_mutex)) {
1503 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1504 perf_mmap_data_free(counter);
1505 mutex_unlock(&counter->mmap_mutex);
1509 static struct vm_operations_struct perf_mmap_vmops = {
1510 .open = perf_mmap_open,
1511 .close = perf_mmap_close,
1512 .fault = perf_mmap_fault,
1515 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1517 struct perf_counter *counter = file->private_data;
1518 unsigned long vma_size;
1519 unsigned long nr_pages;
1520 unsigned long locked, lock_limit;
1524 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1527 vma_size = vma->vm_end - vma->vm_start;
1528 nr_pages = (vma_size / PAGE_SIZE) - 1;
1531 * If we have data pages ensure they're a power-of-two number, so we
1532 * can do bitmasks instead of modulo.
1534 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1537 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1540 if (vma->vm_pgoff != 0)
1543 mutex_lock(&counter->mmap_mutex);
1544 if (atomic_inc_not_zero(&counter->mmap_count)) {
1545 if (nr_pages != counter->data->nr_pages)
1550 extra = nr_pages /* + 1 only account the data pages */;
1551 extra -= sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1555 locked = vma->vm_mm->locked_vm + extra;
1557 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1558 lock_limit >>= PAGE_SHIFT;
1560 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1565 WARN_ON(counter->data);
1566 ret = perf_mmap_data_alloc(counter, nr_pages);
1570 atomic_set(&counter->mmap_count, 1);
1571 vma->vm_mm->locked_vm += extra;
1572 counter->data->nr_locked = extra;
1574 mutex_unlock(&counter->mmap_mutex);
1576 vma->vm_flags &= ~VM_MAYWRITE;
1577 vma->vm_flags |= VM_RESERVED;
1578 vma->vm_ops = &perf_mmap_vmops;
1583 static int perf_fasync(int fd, struct file *filp, int on)
1585 struct perf_counter *counter = filp->private_data;
1586 struct inode *inode = filp->f_path.dentry->d_inode;
1589 mutex_lock(&inode->i_mutex);
1590 retval = fasync_helper(fd, filp, on, &counter->fasync);
1591 mutex_unlock(&inode->i_mutex);
1599 static const struct file_operations perf_fops = {
1600 .release = perf_release,
1603 .unlocked_ioctl = perf_ioctl,
1604 .compat_ioctl = perf_ioctl,
1606 .fasync = perf_fasync,
1610 * Perf counter wakeup
1612 * If there's data, ensure we set the poll() state and publish everything
1613 * to user-space before waking everybody up.
1616 void perf_counter_wakeup(struct perf_counter *counter)
1618 wake_up_all(&counter->waitq);
1620 if (counter->pending_kill) {
1621 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1622 counter->pending_kill = 0;
1629 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1631 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1632 * single linked list and use cmpxchg() to add entries lockless.
1635 static void perf_pending_counter(struct perf_pending_entry *entry)
1637 struct perf_counter *counter = container_of(entry,
1638 struct perf_counter, pending);
1640 if (counter->pending_disable) {
1641 counter->pending_disable = 0;
1642 perf_counter_disable(counter);
1645 if (counter->pending_wakeup) {
1646 counter->pending_wakeup = 0;
1647 perf_counter_wakeup(counter);
1651 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1653 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1657 static void perf_pending_queue(struct perf_pending_entry *entry,
1658 void (*func)(struct perf_pending_entry *))
1660 struct perf_pending_entry **head;
1662 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1667 head = &get_cpu_var(perf_pending_head);
1670 entry->next = *head;
1671 } while (cmpxchg(head, entry->next, entry) != entry->next);
1673 set_perf_counter_pending();
1675 put_cpu_var(perf_pending_head);
1678 static int __perf_pending_run(void)
1680 struct perf_pending_entry *list;
1683 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1684 while (list != PENDING_TAIL) {
1685 void (*func)(struct perf_pending_entry *);
1686 struct perf_pending_entry *entry = list;
1693 * Ensure we observe the unqueue before we issue the wakeup,
1694 * so that we won't be waiting forever.
1695 * -- see perf_not_pending().
1706 static inline int perf_not_pending(struct perf_counter *counter)
1709 * If we flush on whatever cpu we run, there is a chance we don't
1713 __perf_pending_run();
1717 * Ensure we see the proper queue state before going to sleep
1718 * so that we do not miss the wakeup. -- see perf_pending_handle()
1721 return counter->pending.next == NULL;
1724 static void perf_pending_sync(struct perf_counter *counter)
1726 wait_event(counter->waitq, perf_not_pending(counter));
1729 void perf_counter_do_pending(void)
1731 __perf_pending_run();
1735 * Callchain support -- arch specific
1738 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1747 struct perf_output_handle {
1748 struct perf_counter *counter;
1749 struct perf_mmap_data *data;
1750 unsigned int offset;
1755 unsigned long flags;
1758 static void perf_output_wakeup(struct perf_output_handle *handle)
1760 atomic_set(&handle->data->poll, POLL_IN);
1763 handle->counter->pending_wakeup = 1;
1764 perf_pending_queue(&handle->counter->pending,
1765 perf_pending_counter);
1767 perf_counter_wakeup(handle->counter);
1771 * Curious locking construct.
1773 * We need to ensure a later event doesn't publish a head when a former
1774 * event isn't done writing. However since we need to deal with NMIs we
1775 * cannot fully serialize things.
1777 * What we do is serialize between CPUs so we only have to deal with NMI
1778 * nesting on a single CPU.
1780 * We only publish the head (and generate a wakeup) when the outer-most
1783 static void perf_output_lock(struct perf_output_handle *handle)
1785 struct perf_mmap_data *data = handle->data;
1790 local_irq_save(handle->flags);
1791 cpu = smp_processor_id();
1793 if (in_nmi() && atomic_read(&data->lock) == cpu)
1796 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1802 static void perf_output_unlock(struct perf_output_handle *handle)
1804 struct perf_mmap_data *data = handle->data;
1807 data->done_head = data->head;
1809 if (!handle->locked)
1814 * The xchg implies a full barrier that ensures all writes are done
1815 * before we publish the new head, matched by a rmb() in userspace when
1816 * reading this position.
1818 while ((head = atomic_xchg(&data->done_head, 0)))
1819 data->user_page->data_head = head;
1822 * NMI can happen here, which means we can miss a done_head update.
1825 cpu = atomic_xchg(&data->lock, -1);
1826 WARN_ON_ONCE(cpu != smp_processor_id());
1829 * Therefore we have to validate we did not indeed do so.
1831 if (unlikely(atomic_read(&data->done_head))) {
1833 * Since we had it locked, we can lock it again.
1835 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1841 if (atomic_xchg(&data->wakeup, 0))
1842 perf_output_wakeup(handle);
1844 local_irq_restore(handle->flags);
1847 static int perf_output_begin(struct perf_output_handle *handle,
1848 struct perf_counter *counter, unsigned int size,
1849 int nmi, int overflow)
1851 struct perf_mmap_data *data;
1852 unsigned int offset, head;
1855 * For inherited counters we send all the output towards the parent.
1857 if (counter->parent)
1858 counter = counter->parent;
1861 data = rcu_dereference(counter->data);
1865 handle->data = data;
1866 handle->counter = counter;
1868 handle->overflow = overflow;
1870 if (!data->nr_pages)
1873 perf_output_lock(handle);
1876 offset = head = atomic_read(&data->head);
1878 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1880 handle->offset = offset;
1881 handle->head = head;
1883 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
1884 atomic_set(&data->wakeup, 1);
1889 perf_output_wakeup(handle);
1896 static void perf_output_copy(struct perf_output_handle *handle,
1897 void *buf, unsigned int len)
1899 unsigned int pages_mask;
1900 unsigned int offset;
1904 offset = handle->offset;
1905 pages_mask = handle->data->nr_pages - 1;
1906 pages = handle->data->data_pages;
1909 unsigned int page_offset;
1912 nr = (offset >> PAGE_SHIFT) & pages_mask;
1913 page_offset = offset & (PAGE_SIZE - 1);
1914 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1916 memcpy(pages[nr] + page_offset, buf, size);
1923 handle->offset = offset;
1925 WARN_ON_ONCE(handle->offset > handle->head);
1928 #define perf_output_put(handle, x) \
1929 perf_output_copy((handle), &(x), sizeof(x))
1931 static void perf_output_end(struct perf_output_handle *handle)
1933 struct perf_counter *counter = handle->counter;
1934 struct perf_mmap_data *data = handle->data;
1936 int wakeup_events = counter->hw_event.wakeup_events;
1938 if (handle->overflow && wakeup_events) {
1939 int events = atomic_inc_return(&data->events);
1940 if (events >= wakeup_events) {
1941 atomic_sub(wakeup_events, &data->events);
1942 atomic_set(&data->wakeup, 1);
1946 perf_output_unlock(handle);
1950 static void perf_counter_output(struct perf_counter *counter,
1951 int nmi, struct pt_regs *regs, u64 addr)
1954 u64 record_type = counter->hw_event.record_type;
1955 struct perf_output_handle handle;
1956 struct perf_event_header header;
1965 struct perf_callchain_entry *callchain = NULL;
1966 int callchain_size = 0;
1973 header.size = sizeof(header);
1975 header.misc = PERF_EVENT_MISC_OVERFLOW;
1976 header.misc |= user_mode(regs) ?
1977 PERF_EVENT_MISC_USER : PERF_EVENT_MISC_KERNEL;
1979 if (record_type & PERF_RECORD_IP) {
1980 ip = instruction_pointer(regs);
1981 header.type |= PERF_RECORD_IP;
1982 header.size += sizeof(ip);
1985 if (record_type & PERF_RECORD_TID) {
1986 /* namespace issues */
1987 tid_entry.pid = current->group_leader->pid;
1988 tid_entry.tid = current->pid;
1990 header.type |= PERF_RECORD_TID;
1991 header.size += sizeof(tid_entry);
1994 if (record_type & PERF_RECORD_TIME) {
1996 * Maybe do better on x86 and provide cpu_clock_nmi()
1998 time = sched_clock();
2000 header.type |= PERF_RECORD_TIME;
2001 header.size += sizeof(u64);
2004 if (record_type & PERF_RECORD_ADDR) {
2005 header.type |= PERF_RECORD_ADDR;
2006 header.size += sizeof(u64);
2009 if (record_type & PERF_RECORD_CONFIG) {
2010 header.type |= PERF_RECORD_CONFIG;
2011 header.size += sizeof(u64);
2014 if (record_type & PERF_RECORD_CPU) {
2015 header.type |= PERF_RECORD_CPU;
2016 header.size += sizeof(cpu_entry);
2018 cpu_entry.cpu = raw_smp_processor_id();
2021 if (record_type & PERF_RECORD_GROUP) {
2022 header.type |= PERF_RECORD_GROUP;
2023 header.size += sizeof(u64) +
2024 counter->nr_siblings * sizeof(group_entry);
2027 if (record_type & PERF_RECORD_CALLCHAIN) {
2028 callchain = perf_callchain(regs);
2031 callchain_size = (1 + callchain->nr) * sizeof(u64);
2033 header.type |= PERF_RECORD_CALLCHAIN;
2034 header.size += callchain_size;
2038 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2042 perf_output_put(&handle, header);
2044 if (record_type & PERF_RECORD_IP)
2045 perf_output_put(&handle, ip);
2047 if (record_type & PERF_RECORD_TID)
2048 perf_output_put(&handle, tid_entry);
2050 if (record_type & PERF_RECORD_TIME)
2051 perf_output_put(&handle, time);
2053 if (record_type & PERF_RECORD_ADDR)
2054 perf_output_put(&handle, addr);
2056 if (record_type & PERF_RECORD_CONFIG)
2057 perf_output_put(&handle, counter->hw_event.config);
2059 if (record_type & PERF_RECORD_CPU)
2060 perf_output_put(&handle, cpu_entry);
2063 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2065 if (record_type & PERF_RECORD_GROUP) {
2066 struct perf_counter *leader, *sub;
2067 u64 nr = counter->nr_siblings;
2069 perf_output_put(&handle, nr);
2071 leader = counter->group_leader;
2072 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2074 sub->pmu->read(sub);
2076 group_entry.event = sub->hw_event.config;
2077 group_entry.counter = atomic64_read(&sub->count);
2079 perf_output_put(&handle, group_entry);
2084 perf_output_copy(&handle, callchain, callchain_size);
2086 perf_output_end(&handle);
2093 struct perf_comm_event {
2094 struct task_struct *task;
2099 struct perf_event_header header;
2106 static void perf_counter_comm_output(struct perf_counter *counter,
2107 struct perf_comm_event *comm_event)
2109 struct perf_output_handle handle;
2110 int size = comm_event->event.header.size;
2111 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2116 perf_output_put(&handle, comm_event->event);
2117 perf_output_copy(&handle, comm_event->comm,
2118 comm_event->comm_size);
2119 perf_output_end(&handle);
2122 static int perf_counter_comm_match(struct perf_counter *counter,
2123 struct perf_comm_event *comm_event)
2125 if (counter->hw_event.comm &&
2126 comm_event->event.header.type == PERF_EVENT_COMM)
2132 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2133 struct perf_comm_event *comm_event)
2135 struct perf_counter *counter;
2137 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2141 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2142 if (perf_counter_comm_match(counter, comm_event))
2143 perf_counter_comm_output(counter, comm_event);
2148 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2150 struct perf_cpu_context *cpuctx;
2152 char *comm = comm_event->task->comm;
2154 size = ALIGN(strlen(comm)+1, sizeof(u64));
2156 comm_event->comm = comm;
2157 comm_event->comm_size = size;
2159 comm_event->event.header.size = sizeof(comm_event->event) + size;
2161 cpuctx = &get_cpu_var(perf_cpu_context);
2162 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2163 put_cpu_var(perf_cpu_context);
2165 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
2168 void perf_counter_comm(struct task_struct *task)
2170 struct perf_comm_event comm_event;
2172 if (!atomic_read(&nr_comm_tracking))
2175 comm_event = (struct perf_comm_event){
2178 .header = { .type = PERF_EVENT_COMM, },
2179 .pid = task->group_leader->pid,
2184 perf_counter_comm_event(&comm_event);
2191 struct perf_mmap_event {
2197 struct perf_event_header header;
2207 static void perf_counter_mmap_output(struct perf_counter *counter,
2208 struct perf_mmap_event *mmap_event)
2210 struct perf_output_handle handle;
2211 int size = mmap_event->event.header.size;
2212 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2217 perf_output_put(&handle, mmap_event->event);
2218 perf_output_copy(&handle, mmap_event->file_name,
2219 mmap_event->file_size);
2220 perf_output_end(&handle);
2223 static int perf_counter_mmap_match(struct perf_counter *counter,
2224 struct perf_mmap_event *mmap_event)
2226 if (counter->hw_event.mmap &&
2227 mmap_event->event.header.type == PERF_EVENT_MMAP)
2230 if (counter->hw_event.munmap &&
2231 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2237 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2238 struct perf_mmap_event *mmap_event)
2240 struct perf_counter *counter;
2242 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2246 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2247 if (perf_counter_mmap_match(counter, mmap_event))
2248 perf_counter_mmap_output(counter, mmap_event);
2253 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2255 struct perf_cpu_context *cpuctx;
2256 struct file *file = mmap_event->file;
2263 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2265 name = strncpy(tmp, "//enomem", sizeof(tmp));
2268 name = d_path(&file->f_path, buf, PATH_MAX);
2270 name = strncpy(tmp, "//toolong", sizeof(tmp));
2274 name = strncpy(tmp, "//anon", sizeof(tmp));
2279 size = ALIGN(strlen(name)+1, sizeof(u64));
2281 mmap_event->file_name = name;
2282 mmap_event->file_size = size;
2284 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2286 cpuctx = &get_cpu_var(perf_cpu_context);
2287 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2288 put_cpu_var(perf_cpu_context);
2290 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2295 void perf_counter_mmap(unsigned long addr, unsigned long len,
2296 unsigned long pgoff, struct file *file)
2298 struct perf_mmap_event mmap_event;
2300 if (!atomic_read(&nr_mmap_tracking))
2303 mmap_event = (struct perf_mmap_event){
2306 .header = { .type = PERF_EVENT_MMAP, },
2307 .pid = current->group_leader->pid,
2308 .tid = current->pid,
2315 perf_counter_mmap_event(&mmap_event);
2318 void perf_counter_munmap(unsigned long addr, unsigned long len,
2319 unsigned long pgoff, struct file *file)
2321 struct perf_mmap_event mmap_event;
2323 if (!atomic_read(&nr_munmap_tracking))
2326 mmap_event = (struct perf_mmap_event){
2329 .header = { .type = PERF_EVENT_MUNMAP, },
2330 .pid = current->group_leader->pid,
2331 .tid = current->pid,
2338 perf_counter_mmap_event(&mmap_event);
2342 * Generic counter overflow handling.
2345 int perf_counter_overflow(struct perf_counter *counter,
2346 int nmi, struct pt_regs *regs, u64 addr)
2348 int events = atomic_read(&counter->event_limit);
2352 * XXX event_limit might not quite work as expected on inherited
2356 counter->pending_kill = POLL_IN;
2357 if (events && atomic_dec_and_test(&counter->event_limit)) {
2359 counter->pending_kill = POLL_HUP;
2361 counter->pending_disable = 1;
2362 perf_pending_queue(&counter->pending,
2363 perf_pending_counter);
2365 perf_counter_disable(counter);
2368 perf_counter_output(counter, nmi, regs, addr);
2373 * Generic software counter infrastructure
2376 static void perf_swcounter_update(struct perf_counter *counter)
2378 struct hw_perf_counter *hwc = &counter->hw;
2383 prev = atomic64_read(&hwc->prev_count);
2384 now = atomic64_read(&hwc->count);
2385 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2390 atomic64_add(delta, &counter->count);
2391 atomic64_sub(delta, &hwc->period_left);
2394 static void perf_swcounter_set_period(struct perf_counter *counter)
2396 struct hw_perf_counter *hwc = &counter->hw;
2397 s64 left = atomic64_read(&hwc->period_left);
2398 s64 period = hwc->irq_period;
2400 if (unlikely(left <= -period)) {
2402 atomic64_set(&hwc->period_left, left);
2405 if (unlikely(left <= 0)) {
2407 atomic64_add(period, &hwc->period_left);
2410 atomic64_set(&hwc->prev_count, -left);
2411 atomic64_set(&hwc->count, -left);
2414 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2416 enum hrtimer_restart ret = HRTIMER_RESTART;
2417 struct perf_counter *counter;
2418 struct pt_regs *regs;
2420 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2421 counter->pmu->read(counter);
2423 regs = get_irq_regs();
2425 * In case we exclude kernel IPs or are somehow not in interrupt
2426 * context, provide the next best thing, the user IP.
2428 if ((counter->hw_event.exclude_kernel || !regs) &&
2429 !counter->hw_event.exclude_user)
2430 regs = task_pt_regs(current);
2433 if (perf_counter_overflow(counter, 0, regs, 0))
2434 ret = HRTIMER_NORESTART;
2437 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
2442 static void perf_swcounter_overflow(struct perf_counter *counter,
2443 int nmi, struct pt_regs *regs, u64 addr)
2445 perf_swcounter_update(counter);
2446 perf_swcounter_set_period(counter);
2447 if (perf_counter_overflow(counter, nmi, regs, addr))
2448 /* soft-disable the counter */
2453 static int perf_swcounter_match(struct perf_counter *counter,
2454 enum perf_event_types type,
2455 u32 event, struct pt_regs *regs)
2457 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2460 if (perf_event_raw(&counter->hw_event))
2463 if (perf_event_type(&counter->hw_event) != type)
2466 if (perf_event_id(&counter->hw_event) != event)
2469 if (counter->hw_event.exclude_user && user_mode(regs))
2472 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2478 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2479 int nmi, struct pt_regs *regs, u64 addr)
2481 int neg = atomic64_add_negative(nr, &counter->hw.count);
2482 if (counter->hw.irq_period && !neg)
2483 perf_swcounter_overflow(counter, nmi, regs, addr);
2486 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2487 enum perf_event_types type, u32 event,
2488 u64 nr, int nmi, struct pt_regs *regs,
2491 struct perf_counter *counter;
2493 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2497 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2498 if (perf_swcounter_match(counter, type, event, regs))
2499 perf_swcounter_add(counter, nr, nmi, regs, addr);
2504 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2507 return &cpuctx->recursion[3];
2510 return &cpuctx->recursion[2];
2513 return &cpuctx->recursion[1];
2515 return &cpuctx->recursion[0];
2518 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2519 u64 nr, int nmi, struct pt_regs *regs,
2522 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2523 int *recursion = perf_swcounter_recursion_context(cpuctx);
2531 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2532 nr, nmi, regs, addr);
2533 if (cpuctx->task_ctx) {
2534 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2535 nr, nmi, regs, addr);
2542 put_cpu_var(perf_cpu_context);
2546 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2548 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2551 static void perf_swcounter_read(struct perf_counter *counter)
2553 perf_swcounter_update(counter);
2556 static int perf_swcounter_enable(struct perf_counter *counter)
2558 perf_swcounter_set_period(counter);
2562 static void perf_swcounter_disable(struct perf_counter *counter)
2564 perf_swcounter_update(counter);
2567 static const struct pmu perf_ops_generic = {
2568 .enable = perf_swcounter_enable,
2569 .disable = perf_swcounter_disable,
2570 .read = perf_swcounter_read,
2574 * Software counter: cpu wall time clock
2577 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2579 int cpu = raw_smp_processor_id();
2583 now = cpu_clock(cpu);
2584 prev = atomic64_read(&counter->hw.prev_count);
2585 atomic64_set(&counter->hw.prev_count, now);
2586 atomic64_add(now - prev, &counter->count);
2589 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2591 struct hw_perf_counter *hwc = &counter->hw;
2592 int cpu = raw_smp_processor_id();
2594 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2595 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2596 hwc->hrtimer.function = perf_swcounter_hrtimer;
2597 if (hwc->irq_period) {
2598 __hrtimer_start_range_ns(&hwc->hrtimer,
2599 ns_to_ktime(hwc->irq_period), 0,
2600 HRTIMER_MODE_REL, 0);
2606 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2608 hrtimer_cancel(&counter->hw.hrtimer);
2609 cpu_clock_perf_counter_update(counter);
2612 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2614 cpu_clock_perf_counter_update(counter);
2617 static const struct pmu perf_ops_cpu_clock = {
2618 .enable = cpu_clock_perf_counter_enable,
2619 .disable = cpu_clock_perf_counter_disable,
2620 .read = cpu_clock_perf_counter_read,
2624 * Software counter: task time clock
2627 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2632 prev = atomic64_xchg(&counter->hw.prev_count, now);
2634 atomic64_add(delta, &counter->count);
2637 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2639 struct hw_perf_counter *hwc = &counter->hw;
2642 now = counter->ctx->time;
2644 atomic64_set(&hwc->prev_count, now);
2645 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2646 hwc->hrtimer.function = perf_swcounter_hrtimer;
2647 if (hwc->irq_period) {
2648 __hrtimer_start_range_ns(&hwc->hrtimer,
2649 ns_to_ktime(hwc->irq_period), 0,
2650 HRTIMER_MODE_REL, 0);
2656 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2658 hrtimer_cancel(&counter->hw.hrtimer);
2659 task_clock_perf_counter_update(counter, counter->ctx->time);
2663 static void task_clock_perf_counter_read(struct perf_counter *counter)
2668 update_context_time(counter->ctx);
2669 time = counter->ctx->time;
2671 u64 now = perf_clock();
2672 u64 delta = now - counter->ctx->timestamp;
2673 time = counter->ctx->time + delta;
2676 task_clock_perf_counter_update(counter, time);
2679 static const struct pmu perf_ops_task_clock = {
2680 .enable = task_clock_perf_counter_enable,
2681 .disable = task_clock_perf_counter_disable,
2682 .read = task_clock_perf_counter_read,
2686 * Software counter: cpu migrations
2689 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2691 struct task_struct *curr = counter->ctx->task;
2694 return curr->se.nr_migrations;
2695 return cpu_nr_migrations(smp_processor_id());
2698 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2703 prev = atomic64_read(&counter->hw.prev_count);
2704 now = get_cpu_migrations(counter);
2706 atomic64_set(&counter->hw.prev_count, now);
2710 atomic64_add(delta, &counter->count);
2713 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2715 cpu_migrations_perf_counter_update(counter);
2718 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2720 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2721 atomic64_set(&counter->hw.prev_count,
2722 get_cpu_migrations(counter));
2726 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2728 cpu_migrations_perf_counter_update(counter);
2731 static const struct pmu perf_ops_cpu_migrations = {
2732 .enable = cpu_migrations_perf_counter_enable,
2733 .disable = cpu_migrations_perf_counter_disable,
2734 .read = cpu_migrations_perf_counter_read,
2737 #ifdef CONFIG_EVENT_PROFILE
2738 void perf_tpcounter_event(int event_id)
2740 struct pt_regs *regs = get_irq_regs();
2743 regs = task_pt_regs(current);
2745 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2747 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2749 extern int ftrace_profile_enable(int);
2750 extern void ftrace_profile_disable(int);
2752 static void tp_perf_counter_destroy(struct perf_counter *counter)
2754 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2757 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2759 int event_id = perf_event_id(&counter->hw_event);
2762 ret = ftrace_profile_enable(event_id);
2766 counter->destroy = tp_perf_counter_destroy;
2767 counter->hw.irq_period = counter->hw_event.irq_period;
2769 return &perf_ops_generic;
2772 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2778 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2780 struct perf_counter_hw_event *hw_event = &counter->hw_event;
2781 const struct pmu *pmu = NULL;
2782 struct hw_perf_counter *hwc = &counter->hw;
2785 * Software counters (currently) can't in general distinguish
2786 * between user, kernel and hypervisor events.
2787 * However, context switches and cpu migrations are considered
2788 * to be kernel events, and page faults are never hypervisor
2791 switch (perf_event_id(&counter->hw_event)) {
2792 case PERF_COUNT_CPU_CLOCK:
2793 pmu = &perf_ops_cpu_clock;
2795 if (hw_event->irq_period && hw_event->irq_period < 10000)
2796 hw_event->irq_period = 10000;
2798 case PERF_COUNT_TASK_CLOCK:
2800 * If the user instantiates this as a per-cpu counter,
2801 * use the cpu_clock counter instead.
2803 if (counter->ctx->task)
2804 pmu = &perf_ops_task_clock;
2806 pmu = &perf_ops_cpu_clock;
2808 if (hw_event->irq_period && hw_event->irq_period < 10000)
2809 hw_event->irq_period = 10000;
2811 case PERF_COUNT_PAGE_FAULTS:
2812 case PERF_COUNT_PAGE_FAULTS_MIN:
2813 case PERF_COUNT_PAGE_FAULTS_MAJ:
2814 case PERF_COUNT_CONTEXT_SWITCHES:
2815 pmu = &perf_ops_generic;
2817 case PERF_COUNT_CPU_MIGRATIONS:
2818 if (!counter->hw_event.exclude_kernel)
2819 pmu = &perf_ops_cpu_migrations;
2824 hwc->irq_period = hw_event->irq_period;
2830 * Allocate and initialize a counter structure
2832 static struct perf_counter *
2833 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2835 struct perf_counter_context *ctx,
2836 struct perf_counter *group_leader,
2839 const struct pmu *pmu;
2840 struct perf_counter *counter;
2843 counter = kzalloc(sizeof(*counter), gfpflags);
2845 return ERR_PTR(-ENOMEM);
2848 * Single counters are their own group leaders, with an
2849 * empty sibling list:
2852 group_leader = counter;
2854 mutex_init(&counter->mutex);
2855 INIT_LIST_HEAD(&counter->list_entry);
2856 INIT_LIST_HEAD(&counter->event_entry);
2857 INIT_LIST_HEAD(&counter->sibling_list);
2858 init_waitqueue_head(&counter->waitq);
2860 mutex_init(&counter->mmap_mutex);
2862 INIT_LIST_HEAD(&counter->child_list);
2865 counter->hw_event = *hw_event;
2866 counter->group_leader = group_leader;
2867 counter->pmu = NULL;
2870 counter->state = PERF_COUNTER_STATE_INACTIVE;
2871 if (hw_event->disabled)
2872 counter->state = PERF_COUNTER_STATE_OFF;
2877 * we currently do not support PERF_RECORD_GROUP on inherited counters
2879 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
2882 if (perf_event_raw(hw_event)) {
2883 pmu = hw_perf_counter_init(counter);
2887 switch (perf_event_type(hw_event)) {
2888 case PERF_TYPE_HARDWARE:
2889 pmu = hw_perf_counter_init(counter);
2892 case PERF_TYPE_SOFTWARE:
2893 pmu = sw_perf_counter_init(counter);
2896 case PERF_TYPE_TRACEPOINT:
2897 pmu = tp_perf_counter_init(counter);
2904 else if (IS_ERR(pmu))
2909 return ERR_PTR(err);
2914 atomic_inc(&nr_counters);
2915 if (counter->hw_event.mmap)
2916 atomic_inc(&nr_mmap_tracking);
2917 if (counter->hw_event.munmap)
2918 atomic_inc(&nr_munmap_tracking);
2919 if (counter->hw_event.comm)
2920 atomic_inc(&nr_comm_tracking);
2926 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
2928 * @hw_event_uptr: event type attributes for monitoring/sampling
2931 * @group_fd: group leader counter fd
2933 SYSCALL_DEFINE5(perf_counter_open,
2934 const struct perf_counter_hw_event __user *, hw_event_uptr,
2935 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
2937 struct perf_counter *counter, *group_leader;
2938 struct perf_counter_hw_event hw_event;
2939 struct perf_counter_context *ctx;
2940 struct file *counter_file = NULL;
2941 struct file *group_file = NULL;
2942 int fput_needed = 0;
2943 int fput_needed2 = 0;
2946 /* for future expandability... */
2950 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
2954 * Get the target context (task or percpu):
2956 ctx = find_get_context(pid, cpu);
2958 return PTR_ERR(ctx);
2961 * Look up the group leader (we will attach this counter to it):
2963 group_leader = NULL;
2964 if (group_fd != -1) {
2966 group_file = fget_light(group_fd, &fput_needed);
2968 goto err_put_context;
2969 if (group_file->f_op != &perf_fops)
2970 goto err_put_context;
2972 group_leader = group_file->private_data;
2974 * Do not allow a recursive hierarchy (this new sibling
2975 * becoming part of another group-sibling):
2977 if (group_leader->group_leader != group_leader)
2978 goto err_put_context;
2980 * Do not allow to attach to a group in a different
2981 * task or CPU context:
2983 if (group_leader->ctx != ctx)
2984 goto err_put_context;
2986 * Only a group leader can be exclusive or pinned
2988 if (hw_event.exclusive || hw_event.pinned)
2989 goto err_put_context;
2992 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2994 ret = PTR_ERR(counter);
2995 if (IS_ERR(counter))
2996 goto err_put_context;
2998 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3000 goto err_free_put_context;
3002 counter_file = fget_light(ret, &fput_needed2);
3004 goto err_free_put_context;
3006 counter->filp = counter_file;
3007 mutex_lock(&ctx->mutex);
3008 perf_install_in_context(ctx, counter, cpu);
3009 mutex_unlock(&ctx->mutex);
3011 fput_light(counter_file, fput_needed2);
3014 fput_light(group_file, fput_needed);
3018 err_free_put_context:
3028 * Initialize the perf_counter context in a task_struct:
3031 __perf_counter_init_context(struct perf_counter_context *ctx,
3032 struct task_struct *task)
3034 memset(ctx, 0, sizeof(*ctx));
3035 spin_lock_init(&ctx->lock);
3036 mutex_init(&ctx->mutex);
3037 INIT_LIST_HEAD(&ctx->counter_list);
3038 INIT_LIST_HEAD(&ctx->event_list);
3043 * inherit a counter from parent task to child task:
3045 static struct perf_counter *
3046 inherit_counter(struct perf_counter *parent_counter,
3047 struct task_struct *parent,
3048 struct perf_counter_context *parent_ctx,
3049 struct task_struct *child,
3050 struct perf_counter *group_leader,
3051 struct perf_counter_context *child_ctx)
3053 struct perf_counter *child_counter;
3056 * Instead of creating recursive hierarchies of counters,
3057 * we link inherited counters back to the original parent,
3058 * which has a filp for sure, which we use as the reference
3061 if (parent_counter->parent)
3062 parent_counter = parent_counter->parent;
3064 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3065 parent_counter->cpu, child_ctx,
3066 group_leader, GFP_KERNEL);
3067 if (IS_ERR(child_counter))
3068 return child_counter;
3071 * Link it up in the child's context:
3073 child_counter->task = child;
3074 add_counter_to_ctx(child_counter, child_ctx);
3076 child_counter->parent = parent_counter;
3078 * inherit into child's child as well:
3080 child_counter->hw_event.inherit = 1;
3083 * Get a reference to the parent filp - we will fput it
3084 * when the child counter exits. This is safe to do because
3085 * we are in the parent and we know that the filp still
3086 * exists and has a nonzero count:
3088 atomic_long_inc(&parent_counter->filp->f_count);
3091 * Link this into the parent counter's child list
3093 mutex_lock(&parent_counter->mutex);
3094 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3097 * Make the child state follow the state of the parent counter,
3098 * not its hw_event.disabled bit. We hold the parent's mutex,
3099 * so we won't race with perf_counter_{en,dis}able_family.
3101 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3102 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3104 child_counter->state = PERF_COUNTER_STATE_OFF;
3106 mutex_unlock(&parent_counter->mutex);
3108 return child_counter;
3111 static int inherit_group(struct perf_counter *parent_counter,
3112 struct task_struct *parent,
3113 struct perf_counter_context *parent_ctx,
3114 struct task_struct *child,
3115 struct perf_counter_context *child_ctx)
3117 struct perf_counter *leader;
3118 struct perf_counter *sub;
3119 struct perf_counter *child_ctr;
3121 leader = inherit_counter(parent_counter, parent, parent_ctx,
3122 child, NULL, child_ctx);
3124 return PTR_ERR(leader);
3125 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3126 child_ctr = inherit_counter(sub, parent, parent_ctx,
3127 child, leader, child_ctx);
3128 if (IS_ERR(child_ctr))
3129 return PTR_ERR(child_ctr);
3134 static void sync_child_counter(struct perf_counter *child_counter,
3135 struct perf_counter *parent_counter)
3137 u64 parent_val, child_val;
3139 parent_val = atomic64_read(&parent_counter->count);
3140 child_val = atomic64_read(&child_counter->count);
3143 * Add back the child's count to the parent's count:
3145 atomic64_add(child_val, &parent_counter->count);
3146 atomic64_add(child_counter->total_time_enabled,
3147 &parent_counter->child_total_time_enabled);
3148 atomic64_add(child_counter->total_time_running,
3149 &parent_counter->child_total_time_running);
3152 * Remove this counter from the parent's list
3154 mutex_lock(&parent_counter->mutex);
3155 list_del_init(&child_counter->child_list);
3156 mutex_unlock(&parent_counter->mutex);
3159 * Release the parent counter, if this was the last
3162 fput(parent_counter->filp);
3166 __perf_counter_exit_task(struct task_struct *child,
3167 struct perf_counter *child_counter,
3168 struct perf_counter_context *child_ctx)
3170 struct perf_counter *parent_counter;
3171 struct perf_counter *sub, *tmp;
3174 * If we do not self-reap then we have to wait for the
3175 * child task to unschedule (it will happen for sure),
3176 * so that its counter is at its final count. (This
3177 * condition triggers rarely - child tasks usually get
3178 * off their CPU before the parent has a chance to
3179 * get this far into the reaping action)
3181 if (child != current) {
3182 wait_task_inactive(child, 0);
3183 list_del_init(&child_counter->list_entry);
3184 update_counter_times(child_counter);
3186 struct perf_cpu_context *cpuctx;
3187 unsigned long flags;
3191 * Disable and unlink this counter.
3193 * Be careful about zapping the list - IRQ/NMI context
3194 * could still be processing it:
3196 local_irq_save(flags);
3197 perf_flags = hw_perf_save_disable();
3199 cpuctx = &__get_cpu_var(perf_cpu_context);
3201 group_sched_out(child_counter, cpuctx, child_ctx);
3202 update_counter_times(child_counter);
3204 list_del_init(&child_counter->list_entry);
3206 child_ctx->nr_counters--;
3208 hw_perf_restore(perf_flags);
3209 local_irq_restore(flags);
3212 parent_counter = child_counter->parent;
3214 * It can happen that parent exits first, and has counters
3215 * that are still around due to the child reference. These
3216 * counters need to be zapped - but otherwise linger.
3218 if (parent_counter) {
3219 sync_child_counter(child_counter, parent_counter);
3220 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
3223 sync_child_counter(sub, sub->parent);
3227 free_counter(child_counter);
3232 * When a child task exits, feed back counter values to parent counters.
3234 * Note: we may be running in child context, but the PID is not hashed
3235 * anymore so new counters will not be added.
3237 void perf_counter_exit_task(struct task_struct *child)
3239 struct perf_counter *child_counter, *tmp;
3240 struct perf_counter_context *child_ctx;
3242 child_ctx = &child->perf_counter_ctx;
3244 if (likely(!child_ctx->nr_counters))
3247 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3249 __perf_counter_exit_task(child, child_counter, child_ctx);
3253 * Initialize the perf_counter context in task_struct
3255 void perf_counter_init_task(struct task_struct *child)
3257 struct perf_counter_context *child_ctx, *parent_ctx;
3258 struct perf_counter *counter;
3259 struct task_struct *parent = current;
3261 child_ctx = &child->perf_counter_ctx;
3262 parent_ctx = &parent->perf_counter_ctx;
3264 __perf_counter_init_context(child_ctx, child);
3267 * This is executed from the parent task context, so inherit
3268 * counters that have been marked for cloning:
3271 if (likely(!parent_ctx->nr_counters))
3275 * Lock the parent list. No need to lock the child - not PID
3276 * hashed yet and not running, so nobody can access it.
3278 mutex_lock(&parent_ctx->mutex);
3281 * We dont have to disable NMIs - we are only looking at
3282 * the list, not manipulating it:
3284 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
3285 if (!counter->hw_event.inherit)
3288 if (inherit_group(counter, parent,
3289 parent_ctx, child, child_ctx))
3293 mutex_unlock(&parent_ctx->mutex);
3296 static void __cpuinit perf_counter_init_cpu(int cpu)
3298 struct perf_cpu_context *cpuctx;
3300 cpuctx = &per_cpu(perf_cpu_context, cpu);
3301 __perf_counter_init_context(&cpuctx->ctx, NULL);
3303 spin_lock(&perf_resource_lock);
3304 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3305 spin_unlock(&perf_resource_lock);
3307 hw_perf_counter_setup(cpu);
3310 #ifdef CONFIG_HOTPLUG_CPU
3311 static void __perf_counter_exit_cpu(void *info)
3313 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3314 struct perf_counter_context *ctx = &cpuctx->ctx;
3315 struct perf_counter *counter, *tmp;
3317 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3318 __perf_counter_remove_from_context(counter);
3320 static void perf_counter_exit_cpu(int cpu)
3322 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3323 struct perf_counter_context *ctx = &cpuctx->ctx;
3325 mutex_lock(&ctx->mutex);
3326 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3327 mutex_unlock(&ctx->mutex);
3330 static inline void perf_counter_exit_cpu(int cpu) { }
3333 static int __cpuinit
3334 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3336 unsigned int cpu = (long)hcpu;
3340 case CPU_UP_PREPARE:
3341 case CPU_UP_PREPARE_FROZEN:
3342 perf_counter_init_cpu(cpu);
3345 case CPU_DOWN_PREPARE:
3346 case CPU_DOWN_PREPARE_FROZEN:
3347 perf_counter_exit_cpu(cpu);
3357 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3358 .notifier_call = perf_cpu_notify,
3361 void __init perf_counter_init(void)
3363 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3364 (void *)(long)smp_processor_id());
3365 register_cpu_notifier(&perf_cpu_nb);
3368 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3370 return sprintf(buf, "%d\n", perf_reserved_percpu);
3374 perf_set_reserve_percpu(struct sysdev_class *class,
3378 struct perf_cpu_context *cpuctx;
3382 err = strict_strtoul(buf, 10, &val);
3385 if (val > perf_max_counters)
3388 spin_lock(&perf_resource_lock);
3389 perf_reserved_percpu = val;
3390 for_each_online_cpu(cpu) {
3391 cpuctx = &per_cpu(perf_cpu_context, cpu);
3392 spin_lock_irq(&cpuctx->ctx.lock);
3393 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3394 perf_max_counters - perf_reserved_percpu);
3395 cpuctx->max_pertask = mpt;
3396 spin_unlock_irq(&cpuctx->ctx.lock);
3398 spin_unlock(&perf_resource_lock);
3403 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3405 return sprintf(buf, "%d\n", perf_overcommit);
3409 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3414 err = strict_strtoul(buf, 10, &val);
3420 spin_lock(&perf_resource_lock);
3421 perf_overcommit = val;
3422 spin_unlock(&perf_resource_lock);
3427 static SYSDEV_CLASS_ATTR(
3430 perf_show_reserve_percpu,
3431 perf_set_reserve_percpu
3434 static SYSDEV_CLASS_ATTR(
3437 perf_show_overcommit,
3441 static struct attribute *perfclass_attrs[] = {
3442 &attr_reserve_percpu.attr,
3443 &attr_overcommit.attr,
3447 static struct attribute_group perfclass_attr_group = {
3448 .attrs = perfclass_attrs,
3449 .name = "perf_counters",
3452 static int __init perf_counter_sysfs_init(void)
3454 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3455 &perfclass_attr_group);
3457 device_initcall(perf_counter_sysfs_init);