2 * Performance counter core code
4 * Copyright(C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright(C) 2008 Red Hat, Inc., Ingo Molnar
7 * For licencing details see kernel-base/COPYING
11 #include <linux/cpu.h>
12 #include <linux/smp.h>
13 #include <linux/file.h>
14 #include <linux/poll.h>
15 #include <linux/sysfs.h>
16 #include <linux/ptrace.h>
17 #include <linux/percpu.h>
18 #include <linux/uaccess.h>
19 #include <linux/syscalls.h>
20 #include <linux/anon_inodes.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/perf_counter.h>
24 #include <linux/vmstat.h>
25 #include <linux/rculist.h>
26 #include <linux/hardirq.h>
28 #include <asm/irq_regs.h>
31 * Each CPU has a list of per CPU counters:
33 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
35 int perf_max_counters __read_mostly = 1;
36 static int perf_reserved_percpu __read_mostly;
37 static int perf_overcommit __read_mostly = 1;
40 * Mutex for (sysadmin-configurable) counter reservations:
42 static DEFINE_MUTEX(perf_resource_mutex);
45 * Architecture provided APIs - weak aliases:
47 extern __weak const struct hw_perf_counter_ops *
48 hw_perf_counter_init(struct perf_counter *counter)
53 u64 __weak hw_perf_save_disable(void) { return 0; }
54 void __weak hw_perf_restore(u64 ctrl) { barrier(); }
55 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
56 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
57 struct perf_cpu_context *cpuctx,
58 struct perf_counter_context *ctx, int cpu)
63 void __weak perf_counter_print_debug(void) { }
66 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
68 struct perf_counter *group_leader = counter->group_leader;
71 * Depending on whether it is a standalone or sibling counter,
72 * add it straight to the context's counter list, or to the group
73 * leader's sibling list:
75 if (counter->group_leader == counter)
76 list_add_tail(&counter->list_entry, &ctx->counter_list);
78 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
80 list_add_rcu(&counter->event_entry, &ctx->event_list);
84 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
86 struct perf_counter *sibling, *tmp;
88 list_del_init(&counter->list_entry);
89 list_del_rcu(&counter->event_entry);
92 * If this was a group counter with sibling counters then
93 * upgrade the siblings to singleton counters by adding them
94 * to the context list directly:
96 list_for_each_entry_safe(sibling, tmp,
97 &counter->sibling_list, list_entry) {
99 list_move_tail(&sibling->list_entry, &ctx->counter_list);
100 sibling->group_leader = sibling;
105 counter_sched_out(struct perf_counter *counter,
106 struct perf_cpu_context *cpuctx,
107 struct perf_counter_context *ctx)
109 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
112 counter->state = PERF_COUNTER_STATE_INACTIVE;
113 counter->hw_ops->disable(counter);
116 if (!is_software_counter(counter))
117 cpuctx->active_oncpu--;
119 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
120 cpuctx->exclusive = 0;
124 group_sched_out(struct perf_counter *group_counter,
125 struct perf_cpu_context *cpuctx,
126 struct perf_counter_context *ctx)
128 struct perf_counter *counter;
130 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
133 counter_sched_out(group_counter, cpuctx, ctx);
136 * Schedule out siblings (if any):
138 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
139 counter_sched_out(counter, cpuctx, ctx);
141 if (group_counter->hw_event.exclusive)
142 cpuctx->exclusive = 0;
146 * Cross CPU call to remove a performance counter
148 * We disable the counter on the hardware level first. After that we
149 * remove it from the context list.
151 static void __perf_counter_remove_from_context(void *info)
153 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
154 struct perf_counter *counter = info;
155 struct perf_counter_context *ctx = counter->ctx;
160 * If this is a task context, we need to check whether it is
161 * the current task context of this cpu. If not it has been
162 * scheduled out before the smp call arrived.
164 if (ctx->task && cpuctx->task_ctx != ctx)
167 curr_rq_lock_irq_save(&flags);
168 spin_lock(&ctx->lock);
170 counter_sched_out(counter, cpuctx, ctx);
172 counter->task = NULL;
176 * Protect the list operation against NMI by disabling the
177 * counters on a global level. NOP for non NMI based counters.
179 perf_flags = hw_perf_save_disable();
180 list_del_counter(counter, ctx);
181 hw_perf_restore(perf_flags);
185 * Allow more per task counters with respect to the
188 cpuctx->max_pertask =
189 min(perf_max_counters - ctx->nr_counters,
190 perf_max_counters - perf_reserved_percpu);
193 spin_unlock(&ctx->lock);
194 curr_rq_unlock_irq_restore(&flags);
199 * Remove the counter from a task's (or a CPU's) list of counters.
201 * Must be called with counter->mutex and ctx->mutex held.
203 * CPU counters are removed with a smp call. For task counters we only
204 * call when the task is on a CPU.
206 static void perf_counter_remove_from_context(struct perf_counter *counter)
208 struct perf_counter_context *ctx = counter->ctx;
209 struct task_struct *task = ctx->task;
213 * Per cpu counters are removed via an smp call and
214 * the removal is always sucessful.
216 smp_call_function_single(counter->cpu,
217 __perf_counter_remove_from_context,
223 task_oncpu_function_call(task, __perf_counter_remove_from_context,
226 spin_lock_irq(&ctx->lock);
228 * If the context is active we need to retry the smp call.
230 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
231 spin_unlock_irq(&ctx->lock);
236 * The lock prevents that this context is scheduled in so we
237 * can remove the counter safely, if the call above did not
240 if (!list_empty(&counter->list_entry)) {
242 list_del_counter(counter, ctx);
243 counter->task = NULL;
245 spin_unlock_irq(&ctx->lock);
249 * Cross CPU call to disable a performance counter
251 static void __perf_counter_disable(void *info)
253 struct perf_counter *counter = info;
254 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
255 struct perf_counter_context *ctx = counter->ctx;
259 * If this is a per-task counter, need to check whether this
260 * counter's task is the current task on this cpu.
262 if (ctx->task && cpuctx->task_ctx != ctx)
265 curr_rq_lock_irq_save(&flags);
266 spin_lock(&ctx->lock);
269 * If the counter is on, turn it off.
270 * If it is in error state, leave it in error state.
272 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
273 if (counter == counter->group_leader)
274 group_sched_out(counter, cpuctx, ctx);
276 counter_sched_out(counter, cpuctx, ctx);
277 counter->state = PERF_COUNTER_STATE_OFF;
280 spin_unlock(&ctx->lock);
281 curr_rq_unlock_irq_restore(&flags);
287 static void perf_counter_disable(struct perf_counter *counter)
289 struct perf_counter_context *ctx = counter->ctx;
290 struct task_struct *task = ctx->task;
294 * Disable the counter on the cpu that it's on
296 smp_call_function_single(counter->cpu, __perf_counter_disable,
302 task_oncpu_function_call(task, __perf_counter_disable, counter);
304 spin_lock_irq(&ctx->lock);
306 * If the counter is still active, we need to retry the cross-call.
308 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
309 spin_unlock_irq(&ctx->lock);
314 * Since we have the lock this context can't be scheduled
315 * in, so we can change the state safely.
317 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
318 counter->state = PERF_COUNTER_STATE_OFF;
320 spin_unlock_irq(&ctx->lock);
324 * Disable a counter and all its children.
326 static void perf_counter_disable_family(struct perf_counter *counter)
328 struct perf_counter *child;
330 perf_counter_disable(counter);
333 * Lock the mutex to protect the list of children
335 mutex_lock(&counter->mutex);
336 list_for_each_entry(child, &counter->child_list, child_list)
337 perf_counter_disable(child);
338 mutex_unlock(&counter->mutex);
342 counter_sched_in(struct perf_counter *counter,
343 struct perf_cpu_context *cpuctx,
344 struct perf_counter_context *ctx,
347 if (counter->state <= PERF_COUNTER_STATE_OFF)
350 counter->state = PERF_COUNTER_STATE_ACTIVE;
351 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
353 * The new state must be visible before we turn it on in the hardware:
357 if (counter->hw_ops->enable(counter)) {
358 counter->state = PERF_COUNTER_STATE_INACTIVE;
363 if (!is_software_counter(counter))
364 cpuctx->active_oncpu++;
367 if (counter->hw_event.exclusive)
368 cpuctx->exclusive = 1;
374 * Return 1 for a group consisting entirely of software counters,
375 * 0 if the group contains any hardware counters.
377 static int is_software_only_group(struct perf_counter *leader)
379 struct perf_counter *counter;
381 if (!is_software_counter(leader))
383 list_for_each_entry(counter, &leader->sibling_list, list_entry)
384 if (!is_software_counter(counter))
390 * Work out whether we can put this counter group on the CPU now.
392 static int group_can_go_on(struct perf_counter *counter,
393 struct perf_cpu_context *cpuctx,
397 * Groups consisting entirely of software counters can always go on.
399 if (is_software_only_group(counter))
402 * If an exclusive group is already on, no other hardware
403 * counters can go on.
405 if (cpuctx->exclusive)
408 * If this group is exclusive and there are already
409 * counters on the CPU, it can't go on.
411 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
414 * Otherwise, try to add it if all previous groups were able
421 * Cross CPU call to install and enable a performance counter
423 static void __perf_install_in_context(void *info)
425 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
426 struct perf_counter *counter = info;
427 struct perf_counter_context *ctx = counter->ctx;
428 struct perf_counter *leader = counter->group_leader;
429 int cpu = smp_processor_id();
435 * If this is a task context, we need to check whether it is
436 * the current task context of this cpu. If not it has been
437 * scheduled out before the smp call arrived.
439 if (ctx->task && cpuctx->task_ctx != ctx)
442 curr_rq_lock_irq_save(&flags);
443 spin_lock(&ctx->lock);
446 * Protect the list operation against NMI by disabling the
447 * counters on a global level. NOP for non NMI based counters.
449 perf_flags = hw_perf_save_disable();
451 list_add_counter(counter, ctx);
453 counter->prev_state = PERF_COUNTER_STATE_OFF;
456 * Don't put the counter on if it is disabled or if
457 * it is in a group and the group isn't on.
459 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
460 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
464 * An exclusive counter can't go on if there are already active
465 * hardware counters, and no hardware counter can go on if there
466 * is already an exclusive counter on.
468 if (!group_can_go_on(counter, cpuctx, 1))
471 err = counter_sched_in(counter, cpuctx, ctx, cpu);
475 * This counter couldn't go on. If it is in a group
476 * then we have to pull the whole group off.
477 * If the counter group is pinned then put it in error state.
479 if (leader != counter)
480 group_sched_out(leader, cpuctx, ctx);
481 if (leader->hw_event.pinned)
482 leader->state = PERF_COUNTER_STATE_ERROR;
485 if (!err && !ctx->task && cpuctx->max_pertask)
486 cpuctx->max_pertask--;
489 hw_perf_restore(perf_flags);
491 spin_unlock(&ctx->lock);
492 curr_rq_unlock_irq_restore(&flags);
496 * Attach a performance counter to a context
498 * First we add the counter to the list with the hardware enable bit
499 * in counter->hw_config cleared.
501 * If the counter is attached to a task which is on a CPU we use a smp
502 * call to enable it in the task context. The task might have been
503 * scheduled away, but we check this in the smp call again.
505 * Must be called with ctx->mutex held.
508 perf_install_in_context(struct perf_counter_context *ctx,
509 struct perf_counter *counter,
512 struct task_struct *task = ctx->task;
516 * Per cpu counters are installed via an smp call and
517 * the install is always sucessful.
519 smp_call_function_single(cpu, __perf_install_in_context,
524 counter->task = task;
526 task_oncpu_function_call(task, __perf_install_in_context,
529 spin_lock_irq(&ctx->lock);
531 * we need to retry the smp call.
533 if (ctx->is_active && list_empty(&counter->list_entry)) {
534 spin_unlock_irq(&ctx->lock);
539 * The lock prevents that this context is scheduled in so we
540 * can add the counter safely, if it the call above did not
543 if (list_empty(&counter->list_entry)) {
544 list_add_counter(counter, ctx);
547 spin_unlock_irq(&ctx->lock);
551 * Cross CPU call to enable a performance counter
553 static void __perf_counter_enable(void *info)
555 struct perf_counter *counter = info;
556 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
557 struct perf_counter_context *ctx = counter->ctx;
558 struct perf_counter *leader = counter->group_leader;
563 * If this is a per-task counter, need to check whether this
564 * counter's task is the current task on this cpu.
566 if (ctx->task && cpuctx->task_ctx != ctx)
569 curr_rq_lock_irq_save(&flags);
570 spin_lock(&ctx->lock);
572 counter->prev_state = counter->state;
573 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
575 counter->state = PERF_COUNTER_STATE_INACTIVE;
578 * If the counter is in a group and isn't the group leader,
579 * then don't put it on unless the group is on.
581 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
584 if (!group_can_go_on(counter, cpuctx, 1))
587 err = counter_sched_in(counter, cpuctx, ctx,
592 * If this counter can't go on and it's part of a
593 * group, then the whole group has to come off.
595 if (leader != counter)
596 group_sched_out(leader, cpuctx, ctx);
597 if (leader->hw_event.pinned)
598 leader->state = PERF_COUNTER_STATE_ERROR;
602 spin_unlock(&ctx->lock);
603 curr_rq_unlock_irq_restore(&flags);
609 static void perf_counter_enable(struct perf_counter *counter)
611 struct perf_counter_context *ctx = counter->ctx;
612 struct task_struct *task = ctx->task;
616 * Enable the counter on the cpu that it's on
618 smp_call_function_single(counter->cpu, __perf_counter_enable,
623 spin_lock_irq(&ctx->lock);
624 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
628 * If the counter is in error state, clear that first.
629 * That way, if we see the counter in error state below, we
630 * know that it has gone back into error state, as distinct
631 * from the task having been scheduled away before the
632 * cross-call arrived.
634 if (counter->state == PERF_COUNTER_STATE_ERROR)
635 counter->state = PERF_COUNTER_STATE_OFF;
638 spin_unlock_irq(&ctx->lock);
639 task_oncpu_function_call(task, __perf_counter_enable, counter);
641 spin_lock_irq(&ctx->lock);
644 * If the context is active and the counter is still off,
645 * we need to retry the cross-call.
647 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
651 * Since we have the lock this context can't be scheduled
652 * in, so we can change the state safely.
654 if (counter->state == PERF_COUNTER_STATE_OFF)
655 counter->state = PERF_COUNTER_STATE_INACTIVE;
657 spin_unlock_irq(&ctx->lock);
661 * Enable a counter and all its children.
663 static void perf_counter_enable_family(struct perf_counter *counter)
665 struct perf_counter *child;
667 perf_counter_enable(counter);
670 * Lock the mutex to protect the list of children
672 mutex_lock(&counter->mutex);
673 list_for_each_entry(child, &counter->child_list, child_list)
674 perf_counter_enable(child);
675 mutex_unlock(&counter->mutex);
678 void __perf_counter_sched_out(struct perf_counter_context *ctx,
679 struct perf_cpu_context *cpuctx)
681 struct perf_counter *counter;
684 spin_lock(&ctx->lock);
686 if (likely(!ctx->nr_counters))
689 flags = hw_perf_save_disable();
690 if (ctx->nr_active) {
691 list_for_each_entry(counter, &ctx->counter_list, list_entry)
692 group_sched_out(counter, cpuctx, ctx);
694 hw_perf_restore(flags);
696 spin_unlock(&ctx->lock);
700 * Called from scheduler to remove the counters of the current task,
701 * with interrupts disabled.
703 * We stop each counter and update the counter value in counter->count.
705 * This does not protect us against NMI, but disable()
706 * sets the disabled bit in the control field of counter _before_
707 * accessing the counter control register. If a NMI hits, then it will
708 * not restart the counter.
710 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
712 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
713 struct perf_counter_context *ctx = &task->perf_counter_ctx;
714 struct pt_regs *regs;
716 if (likely(!cpuctx->task_ctx))
719 regs = task_pt_regs(task);
720 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs);
721 __perf_counter_sched_out(ctx, cpuctx);
723 cpuctx->task_ctx = NULL;
726 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
728 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
732 group_sched_in(struct perf_counter *group_counter,
733 struct perf_cpu_context *cpuctx,
734 struct perf_counter_context *ctx,
737 struct perf_counter *counter, *partial_group;
740 if (group_counter->state == PERF_COUNTER_STATE_OFF)
743 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
745 return ret < 0 ? ret : 0;
747 group_counter->prev_state = group_counter->state;
748 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
752 * Schedule in siblings as one group (if any):
754 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
755 counter->prev_state = counter->state;
756 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
757 partial_group = counter;
766 * Groups can be scheduled in as one unit only, so undo any
767 * partial group before returning:
769 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
770 if (counter == partial_group)
772 counter_sched_out(counter, cpuctx, ctx);
774 counter_sched_out(group_counter, cpuctx, ctx);
780 __perf_counter_sched_in(struct perf_counter_context *ctx,
781 struct perf_cpu_context *cpuctx, int cpu)
783 struct perf_counter *counter;
787 spin_lock(&ctx->lock);
789 if (likely(!ctx->nr_counters))
792 flags = hw_perf_save_disable();
795 * First go through the list and put on any pinned groups
796 * in order to give them the best chance of going on.
798 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
799 if (counter->state <= PERF_COUNTER_STATE_OFF ||
800 !counter->hw_event.pinned)
802 if (counter->cpu != -1 && counter->cpu != cpu)
805 if (group_can_go_on(counter, cpuctx, 1))
806 group_sched_in(counter, cpuctx, ctx, cpu);
809 * If this pinned group hasn't been scheduled,
810 * put it in error state.
812 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
813 counter->state = PERF_COUNTER_STATE_ERROR;
816 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
818 * Ignore counters in OFF or ERROR state, and
819 * ignore pinned counters since we did them already.
821 if (counter->state <= PERF_COUNTER_STATE_OFF ||
822 counter->hw_event.pinned)
826 * Listen to the 'cpu' scheduling filter constraint
829 if (counter->cpu != -1 && counter->cpu != cpu)
832 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
833 if (group_sched_in(counter, cpuctx, ctx, cpu))
837 hw_perf_restore(flags);
839 spin_unlock(&ctx->lock);
843 * Called from scheduler to add the counters of the current task
844 * with interrupts disabled.
846 * We restore the counter value and then enable it.
848 * This does not protect us against NMI, but enable()
849 * sets the enabled bit in the control field of counter _before_
850 * accessing the counter control register. If a NMI hits, then it will
851 * keep the counter running.
853 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
855 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
856 struct perf_counter_context *ctx = &task->perf_counter_ctx;
858 __perf_counter_sched_in(ctx, cpuctx, cpu);
859 cpuctx->task_ctx = ctx;
862 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
864 struct perf_counter_context *ctx = &cpuctx->ctx;
866 __perf_counter_sched_in(ctx, cpuctx, cpu);
869 int perf_counter_task_disable(void)
871 struct task_struct *curr = current;
872 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
873 struct perf_counter *counter;
878 if (likely(!ctx->nr_counters))
881 curr_rq_lock_irq_save(&flags);
882 cpu = smp_processor_id();
884 /* force the update of the task clock: */
885 __task_delta_exec(curr, 1);
887 perf_counter_task_sched_out(curr, cpu);
889 spin_lock(&ctx->lock);
892 * Disable all the counters:
894 perf_flags = hw_perf_save_disable();
896 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
897 if (counter->state != PERF_COUNTER_STATE_ERROR)
898 counter->state = PERF_COUNTER_STATE_OFF;
901 hw_perf_restore(perf_flags);
903 spin_unlock(&ctx->lock);
905 curr_rq_unlock_irq_restore(&flags);
910 int perf_counter_task_enable(void)
912 struct task_struct *curr = current;
913 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
914 struct perf_counter *counter;
919 if (likely(!ctx->nr_counters))
922 curr_rq_lock_irq_save(&flags);
923 cpu = smp_processor_id();
925 /* force the update of the task clock: */
926 __task_delta_exec(curr, 1);
928 perf_counter_task_sched_out(curr, cpu);
930 spin_lock(&ctx->lock);
933 * Disable all the counters:
935 perf_flags = hw_perf_save_disable();
937 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
938 if (counter->state > PERF_COUNTER_STATE_OFF)
940 counter->state = PERF_COUNTER_STATE_INACTIVE;
941 counter->hw_event.disabled = 0;
943 hw_perf_restore(perf_flags);
945 spin_unlock(&ctx->lock);
947 perf_counter_task_sched_in(curr, cpu);
949 curr_rq_unlock_irq_restore(&flags);
955 * Round-robin a context's counters:
957 static void rotate_ctx(struct perf_counter_context *ctx)
959 struct perf_counter *counter;
962 if (!ctx->nr_counters)
965 spin_lock(&ctx->lock);
967 * Rotate the first entry last (works just fine for group counters too):
969 perf_flags = hw_perf_save_disable();
970 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
971 list_move_tail(&counter->list_entry, &ctx->counter_list);
974 hw_perf_restore(perf_flags);
976 spin_unlock(&ctx->lock);
979 void perf_counter_task_tick(struct task_struct *curr, int cpu)
981 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
982 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
983 const int rotate_percpu = 0;
986 perf_counter_cpu_sched_out(cpuctx);
987 perf_counter_task_sched_out(curr, cpu);
990 rotate_ctx(&cpuctx->ctx);
994 perf_counter_cpu_sched_in(cpuctx, cpu);
995 perf_counter_task_sched_in(curr, cpu);
999 * Cross CPU call to read the hardware counter
1001 static void __read(void *info)
1003 struct perf_counter *counter = info;
1004 unsigned long flags;
1006 curr_rq_lock_irq_save(&flags);
1007 counter->hw_ops->read(counter);
1008 curr_rq_unlock_irq_restore(&flags);
1011 static u64 perf_counter_read(struct perf_counter *counter)
1014 * If counter is enabled and currently active on a CPU, update the
1015 * value in the counter structure:
1017 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1018 smp_call_function_single(counter->oncpu,
1019 __read, counter, 1);
1022 return atomic64_read(&counter->count);
1026 * Cross CPU call to switch performance data pointers
1028 static void __perf_switch_irq_data(void *info)
1030 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1031 struct perf_counter *counter = info;
1032 struct perf_counter_context *ctx = counter->ctx;
1033 struct perf_data *oldirqdata = counter->irqdata;
1036 * If this is a task context, we need to check whether it is
1037 * the current task context of this cpu. If not it has been
1038 * scheduled out before the smp call arrived.
1041 if (cpuctx->task_ctx != ctx)
1043 spin_lock(&ctx->lock);
1046 /* Change the pointer NMI safe */
1047 atomic_long_set((atomic_long_t *)&counter->irqdata,
1048 (unsigned long) counter->usrdata);
1049 counter->usrdata = oldirqdata;
1052 spin_unlock(&ctx->lock);
1055 static struct perf_data *perf_switch_irq_data(struct perf_counter *counter)
1057 struct perf_counter_context *ctx = counter->ctx;
1058 struct perf_data *oldirqdata = counter->irqdata;
1059 struct task_struct *task = ctx->task;
1062 smp_call_function_single(counter->cpu,
1063 __perf_switch_irq_data,
1065 return counter->usrdata;
1069 spin_lock_irq(&ctx->lock);
1070 if (counter->state != PERF_COUNTER_STATE_ACTIVE) {
1071 counter->irqdata = counter->usrdata;
1072 counter->usrdata = oldirqdata;
1073 spin_unlock_irq(&ctx->lock);
1076 spin_unlock_irq(&ctx->lock);
1077 task_oncpu_function_call(task, __perf_switch_irq_data, counter);
1078 /* Might have failed, because task was scheduled out */
1079 if (counter->irqdata == oldirqdata)
1082 return counter->usrdata;
1085 static void put_context(struct perf_counter_context *ctx)
1088 put_task_struct(ctx->task);
1091 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1093 struct perf_cpu_context *cpuctx;
1094 struct perf_counter_context *ctx;
1095 struct task_struct *task;
1098 * If cpu is not a wildcard then this is a percpu counter:
1101 /* Must be root to operate on a CPU counter: */
1102 if (!capable(CAP_SYS_ADMIN))
1103 return ERR_PTR(-EACCES);
1105 if (cpu < 0 || cpu > num_possible_cpus())
1106 return ERR_PTR(-EINVAL);
1109 * We could be clever and allow to attach a counter to an
1110 * offline CPU and activate it when the CPU comes up, but
1113 if (!cpu_isset(cpu, cpu_online_map))
1114 return ERR_PTR(-ENODEV);
1116 cpuctx = &per_cpu(perf_cpu_context, cpu);
1126 task = find_task_by_vpid(pid);
1128 get_task_struct(task);
1132 return ERR_PTR(-ESRCH);
1134 ctx = &task->perf_counter_ctx;
1137 /* Reuse ptrace permission checks for now. */
1138 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1140 return ERR_PTR(-EACCES);
1146 static void free_counter_rcu(struct rcu_head *head)
1148 struct perf_counter *counter;
1150 counter = container_of(head, struct perf_counter, rcu_head);
1154 static void free_counter(struct perf_counter *counter)
1156 if (counter->destroy)
1157 counter->destroy(counter);
1159 call_rcu(&counter->rcu_head, free_counter_rcu);
1163 * Called when the last reference to the file is gone.
1165 static int perf_release(struct inode *inode, struct file *file)
1167 struct perf_counter *counter = file->private_data;
1168 struct perf_counter_context *ctx = counter->ctx;
1170 file->private_data = NULL;
1172 mutex_lock(&ctx->mutex);
1173 mutex_lock(&counter->mutex);
1175 perf_counter_remove_from_context(counter);
1177 mutex_unlock(&counter->mutex);
1178 mutex_unlock(&ctx->mutex);
1180 free_counter(counter);
1187 * Read the performance counter - simple non blocking version for now
1190 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1194 if (count != sizeof(cntval))
1198 * Return end-of-file for a read on a counter that is in
1199 * error state (i.e. because it was pinned but it couldn't be
1200 * scheduled on to the CPU at some point).
1202 if (counter->state == PERF_COUNTER_STATE_ERROR)
1205 mutex_lock(&counter->mutex);
1206 cntval = perf_counter_read(counter);
1207 mutex_unlock(&counter->mutex);
1209 return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
1213 perf_copy_usrdata(struct perf_data *usrdata, char __user *buf, size_t count)
1218 count = min(count, (size_t)usrdata->len);
1219 if (copy_to_user(buf, usrdata->data + usrdata->rd_idx, count))
1222 /* Adjust the counters */
1223 usrdata->len -= count;
1225 usrdata->rd_idx = 0;
1227 usrdata->rd_idx += count;
1233 perf_read_irq_data(struct perf_counter *counter,
1238 struct perf_data *irqdata, *usrdata;
1239 DECLARE_WAITQUEUE(wait, current);
1242 irqdata = counter->irqdata;
1243 usrdata = counter->usrdata;
1245 if (usrdata->len + irqdata->len >= count)
1251 spin_lock_irq(&counter->waitq.lock);
1252 __add_wait_queue(&counter->waitq, &wait);
1254 set_current_state(TASK_INTERRUPTIBLE);
1255 if (usrdata->len + irqdata->len >= count)
1258 if (signal_pending(current))
1261 if (counter->state == PERF_COUNTER_STATE_ERROR)
1264 spin_unlock_irq(&counter->waitq.lock);
1266 spin_lock_irq(&counter->waitq.lock);
1268 __remove_wait_queue(&counter->waitq, &wait);
1269 __set_current_state(TASK_RUNNING);
1270 spin_unlock_irq(&counter->waitq.lock);
1272 if (usrdata->len + irqdata->len < count &&
1273 counter->state != PERF_COUNTER_STATE_ERROR)
1274 return -ERESTARTSYS;
1276 mutex_lock(&counter->mutex);
1278 /* Drain pending data first: */
1279 res = perf_copy_usrdata(usrdata, buf, count);
1280 if (res < 0 || res == count)
1283 /* Switch irq buffer: */
1284 usrdata = perf_switch_irq_data(counter);
1285 res2 = perf_copy_usrdata(usrdata, buf + res, count - res);
1293 mutex_unlock(&counter->mutex);
1299 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1301 struct perf_counter *counter = file->private_data;
1303 switch (counter->hw_event.record_type) {
1304 case PERF_RECORD_SIMPLE:
1305 return perf_read_hw(counter, buf, count);
1307 case PERF_RECORD_IRQ:
1308 case PERF_RECORD_GROUP:
1309 return perf_read_irq_data(counter, buf, count,
1310 file->f_flags & O_NONBLOCK);
1315 static unsigned int perf_poll(struct file *file, poll_table *wait)
1317 struct perf_counter *counter = file->private_data;
1318 unsigned int events = 0;
1319 unsigned long flags;
1321 poll_wait(file, &counter->waitq, wait);
1323 spin_lock_irqsave(&counter->waitq.lock, flags);
1324 if (counter->usrdata->len || counter->irqdata->len)
1326 spin_unlock_irqrestore(&counter->waitq.lock, flags);
1331 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1333 struct perf_counter *counter = file->private_data;
1337 case PERF_COUNTER_IOC_ENABLE:
1338 perf_counter_enable_family(counter);
1340 case PERF_COUNTER_IOC_DISABLE:
1341 perf_counter_disable_family(counter);
1349 static const struct file_operations perf_fops = {
1350 .release = perf_release,
1353 .unlocked_ioctl = perf_ioctl,
1354 .compat_ioctl = perf_ioctl,
1361 static void perf_counter_store_irq(struct perf_counter *counter, u64 data)
1363 struct perf_data *irqdata = counter->irqdata;
1365 if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
1368 u64 *p = (u64 *) &irqdata->data[irqdata->len];
1371 irqdata->len += sizeof(u64);
1375 static void perf_counter_handle_group(struct perf_counter *counter)
1377 struct perf_counter *leader, *sub;
1379 leader = counter->group_leader;
1380 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1382 sub->hw_ops->read(sub);
1383 perf_counter_store_irq(counter, sub->hw_event.config);
1384 perf_counter_store_irq(counter, atomic64_read(&sub->count));
1388 void perf_counter_output(struct perf_counter *counter,
1389 int nmi, struct pt_regs *regs)
1391 switch (counter->hw_event.record_type) {
1392 case PERF_RECORD_SIMPLE:
1395 case PERF_RECORD_IRQ:
1396 perf_counter_store_irq(counter, instruction_pointer(regs));
1399 case PERF_RECORD_GROUP:
1400 perf_counter_handle_group(counter);
1405 counter->wakeup_pending = 1;
1406 set_perf_counter_pending();
1408 wake_up(&counter->waitq);
1412 * Generic software counter infrastructure
1415 static void perf_swcounter_update(struct perf_counter *counter)
1417 struct hw_perf_counter *hwc = &counter->hw;
1422 prev = atomic64_read(&hwc->prev_count);
1423 now = atomic64_read(&hwc->count);
1424 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
1429 atomic64_add(delta, &counter->count);
1430 atomic64_sub(delta, &hwc->period_left);
1433 static void perf_swcounter_set_period(struct perf_counter *counter)
1435 struct hw_perf_counter *hwc = &counter->hw;
1436 s64 left = atomic64_read(&hwc->period_left);
1437 s64 period = hwc->irq_period;
1439 if (unlikely(left <= -period)) {
1441 atomic64_set(&hwc->period_left, left);
1444 if (unlikely(left <= 0)) {
1446 atomic64_add(period, &hwc->period_left);
1449 atomic64_set(&hwc->prev_count, -left);
1450 atomic64_set(&hwc->count, -left);
1453 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
1455 struct perf_counter *counter;
1456 struct pt_regs *regs;
1458 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
1459 counter->hw_ops->read(counter);
1461 regs = get_irq_regs();
1463 * In case we exclude kernel IPs or are somehow not in interrupt
1464 * context, provide the next best thing, the user IP.
1466 if ((counter->hw_event.exclude_kernel || !regs) &&
1467 !counter->hw_event.exclude_user)
1468 regs = task_pt_regs(current);
1471 perf_counter_output(counter, 0, regs);
1473 hrtimer_forward_now(hrtimer, ns_to_ktime(counter->hw.irq_period));
1475 return HRTIMER_RESTART;
1478 static void perf_swcounter_overflow(struct perf_counter *counter,
1479 int nmi, struct pt_regs *regs)
1481 perf_swcounter_update(counter);
1482 perf_swcounter_set_period(counter);
1483 perf_counter_output(counter, nmi, regs);
1486 static int perf_swcounter_match(struct perf_counter *counter,
1487 enum perf_event_types type,
1488 u32 event, struct pt_regs *regs)
1490 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1493 if (perf_event_raw(&counter->hw_event))
1496 if (perf_event_type(&counter->hw_event) != type)
1499 if (perf_event_id(&counter->hw_event) != event)
1502 if (counter->hw_event.exclude_user && user_mode(regs))
1505 if (counter->hw_event.exclude_kernel && !user_mode(regs))
1511 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
1512 int nmi, struct pt_regs *regs)
1514 int neg = atomic64_add_negative(nr, &counter->hw.count);
1515 if (counter->hw.irq_period && !neg)
1516 perf_swcounter_overflow(counter, nmi, regs);
1519 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
1520 enum perf_event_types type, u32 event,
1521 u64 nr, int nmi, struct pt_regs *regs)
1523 struct perf_counter *counter;
1525 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
1529 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
1530 if (perf_swcounter_match(counter, type, event, regs))
1531 perf_swcounter_add(counter, nr, nmi, regs);
1536 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
1539 return &cpuctx->recursion[3];
1542 return &cpuctx->recursion[2];
1545 return &cpuctx->recursion[1];
1547 return &cpuctx->recursion[0];
1550 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
1551 u64 nr, int nmi, struct pt_regs *regs)
1553 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
1554 int *recursion = perf_swcounter_recursion_context(cpuctx);
1562 perf_swcounter_ctx_event(&cpuctx->ctx, type, event, nr, nmi, regs);
1563 if (cpuctx->task_ctx) {
1564 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
1572 put_cpu_var(perf_cpu_context);
1575 void perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)
1577 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs);
1580 static void perf_swcounter_read(struct perf_counter *counter)
1582 perf_swcounter_update(counter);
1585 static int perf_swcounter_enable(struct perf_counter *counter)
1587 perf_swcounter_set_period(counter);
1591 static void perf_swcounter_disable(struct perf_counter *counter)
1593 perf_swcounter_update(counter);
1596 static const struct hw_perf_counter_ops perf_ops_generic = {
1597 .enable = perf_swcounter_enable,
1598 .disable = perf_swcounter_disable,
1599 .read = perf_swcounter_read,
1603 * Software counter: cpu wall time clock
1606 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
1608 int cpu = raw_smp_processor_id();
1612 now = cpu_clock(cpu);
1613 prev = atomic64_read(&counter->hw.prev_count);
1614 atomic64_set(&counter->hw.prev_count, now);
1615 atomic64_add(now - prev, &counter->count);
1618 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
1620 struct hw_perf_counter *hwc = &counter->hw;
1621 int cpu = raw_smp_processor_id();
1623 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
1624 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1625 hwc->hrtimer.function = perf_swcounter_hrtimer;
1626 if (hwc->irq_period) {
1627 __hrtimer_start_range_ns(&hwc->hrtimer,
1628 ns_to_ktime(hwc->irq_period), 0,
1629 HRTIMER_MODE_REL, 0);
1635 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
1637 hrtimer_cancel(&counter->hw.hrtimer);
1638 cpu_clock_perf_counter_update(counter);
1641 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
1643 cpu_clock_perf_counter_update(counter);
1646 static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
1647 .enable = cpu_clock_perf_counter_enable,
1648 .disable = cpu_clock_perf_counter_disable,
1649 .read = cpu_clock_perf_counter_read,
1653 * Software counter: task time clock
1657 * Called from within the scheduler:
1659 static u64 task_clock_perf_counter_val(struct perf_counter *counter, int update)
1661 struct task_struct *curr = counter->task;
1664 delta = __task_delta_exec(curr, update);
1666 return curr->se.sum_exec_runtime + delta;
1669 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
1674 prev = atomic64_read(&counter->hw.prev_count);
1676 atomic64_set(&counter->hw.prev_count, now);
1680 atomic64_add(delta, &counter->count);
1683 static int task_clock_perf_counter_enable(struct perf_counter *counter)
1685 struct hw_perf_counter *hwc = &counter->hw;
1687 atomic64_set(&hwc->prev_count, task_clock_perf_counter_val(counter, 0));
1688 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1689 hwc->hrtimer.function = perf_swcounter_hrtimer;
1690 if (hwc->irq_period) {
1691 __hrtimer_start_range_ns(&hwc->hrtimer,
1692 ns_to_ktime(hwc->irq_period), 0,
1693 HRTIMER_MODE_REL, 0);
1699 static void task_clock_perf_counter_disable(struct perf_counter *counter)
1701 hrtimer_cancel(&counter->hw.hrtimer);
1702 task_clock_perf_counter_update(counter,
1703 task_clock_perf_counter_val(counter, 0));
1706 static void task_clock_perf_counter_read(struct perf_counter *counter)
1708 task_clock_perf_counter_update(counter,
1709 task_clock_perf_counter_val(counter, 1));
1712 static const struct hw_perf_counter_ops perf_ops_task_clock = {
1713 .enable = task_clock_perf_counter_enable,
1714 .disable = task_clock_perf_counter_disable,
1715 .read = task_clock_perf_counter_read,
1719 * Software counter: cpu migrations
1722 static inline u64 get_cpu_migrations(struct perf_counter *counter)
1724 struct task_struct *curr = counter->ctx->task;
1727 return curr->se.nr_migrations;
1728 return cpu_nr_migrations(smp_processor_id());
1731 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
1736 prev = atomic64_read(&counter->hw.prev_count);
1737 now = get_cpu_migrations(counter);
1739 atomic64_set(&counter->hw.prev_count, now);
1743 atomic64_add(delta, &counter->count);
1746 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
1748 cpu_migrations_perf_counter_update(counter);
1751 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
1753 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
1754 atomic64_set(&counter->hw.prev_count,
1755 get_cpu_migrations(counter));
1759 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
1761 cpu_migrations_perf_counter_update(counter);
1764 static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
1765 .enable = cpu_migrations_perf_counter_enable,
1766 .disable = cpu_migrations_perf_counter_disable,
1767 .read = cpu_migrations_perf_counter_read,
1770 #ifdef CONFIG_EVENT_PROFILE
1771 void perf_tpcounter_event(int event_id)
1773 struct pt_regs *regs = get_irq_regs();
1776 regs = task_pt_regs(current);
1778 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs);
1781 extern int ftrace_profile_enable(int);
1782 extern void ftrace_profile_disable(int);
1784 static void tp_perf_counter_destroy(struct perf_counter *counter)
1786 ftrace_profile_disable(perf_event_id(&counter->hw_event));
1789 static const struct hw_perf_counter_ops *
1790 tp_perf_counter_init(struct perf_counter *counter)
1792 int event_id = perf_event_id(&counter->hw_event);
1795 ret = ftrace_profile_enable(event_id);
1799 counter->destroy = tp_perf_counter_destroy;
1800 counter->hw.irq_period = counter->hw_event.irq_period;
1802 return &perf_ops_generic;
1805 static const struct hw_perf_counter_ops *
1806 tp_perf_counter_init(struct perf_counter *counter)
1812 static const struct hw_perf_counter_ops *
1813 sw_perf_counter_init(struct perf_counter *counter)
1815 struct perf_counter_hw_event *hw_event = &counter->hw_event;
1816 const struct hw_perf_counter_ops *hw_ops = NULL;
1817 struct hw_perf_counter *hwc = &counter->hw;
1820 * Software counters (currently) can't in general distinguish
1821 * between user, kernel and hypervisor events.
1822 * However, context switches and cpu migrations are considered
1823 * to be kernel events, and page faults are never hypervisor
1826 switch (perf_event_id(&counter->hw_event)) {
1827 case PERF_COUNT_CPU_CLOCK:
1828 hw_ops = &perf_ops_cpu_clock;
1830 if (hw_event->irq_period && hw_event->irq_period < 10000)
1831 hw_event->irq_period = 10000;
1833 case PERF_COUNT_TASK_CLOCK:
1835 * If the user instantiates this as a per-cpu counter,
1836 * use the cpu_clock counter instead.
1838 if (counter->ctx->task)
1839 hw_ops = &perf_ops_task_clock;
1841 hw_ops = &perf_ops_cpu_clock;
1843 if (hw_event->irq_period && hw_event->irq_period < 10000)
1844 hw_event->irq_period = 10000;
1846 case PERF_COUNT_PAGE_FAULTS:
1847 case PERF_COUNT_PAGE_FAULTS_MIN:
1848 case PERF_COUNT_PAGE_FAULTS_MAJ:
1849 case PERF_COUNT_CONTEXT_SWITCHES:
1850 hw_ops = &perf_ops_generic;
1852 case PERF_COUNT_CPU_MIGRATIONS:
1853 if (!counter->hw_event.exclude_kernel)
1854 hw_ops = &perf_ops_cpu_migrations;
1859 hwc->irq_period = hw_event->irq_period;
1865 * Allocate and initialize a counter structure
1867 static struct perf_counter *
1868 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
1870 struct perf_counter_context *ctx,
1871 struct perf_counter *group_leader,
1874 const struct hw_perf_counter_ops *hw_ops;
1875 struct perf_counter *counter;
1877 counter = kzalloc(sizeof(*counter), gfpflags);
1882 * Single counters are their own group leaders, with an
1883 * empty sibling list:
1886 group_leader = counter;
1888 mutex_init(&counter->mutex);
1889 INIT_LIST_HEAD(&counter->list_entry);
1890 INIT_LIST_HEAD(&counter->event_entry);
1891 INIT_LIST_HEAD(&counter->sibling_list);
1892 init_waitqueue_head(&counter->waitq);
1894 INIT_LIST_HEAD(&counter->child_list);
1896 counter->irqdata = &counter->data[0];
1897 counter->usrdata = &counter->data[1];
1899 counter->hw_event = *hw_event;
1900 counter->wakeup_pending = 0;
1901 counter->group_leader = group_leader;
1902 counter->hw_ops = NULL;
1905 counter->state = PERF_COUNTER_STATE_INACTIVE;
1906 if (hw_event->disabled)
1907 counter->state = PERF_COUNTER_STATE_OFF;
1911 if (perf_event_raw(hw_event)) {
1912 hw_ops = hw_perf_counter_init(counter);
1916 switch (perf_event_type(hw_event)) {
1917 case PERF_TYPE_HARDWARE:
1918 hw_ops = hw_perf_counter_init(counter);
1921 case PERF_TYPE_SOFTWARE:
1922 hw_ops = sw_perf_counter_init(counter);
1925 case PERF_TYPE_TRACEPOINT:
1926 hw_ops = tp_perf_counter_init(counter);
1935 counter->hw_ops = hw_ops;
1941 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
1943 * @hw_event_uptr: event type attributes for monitoring/sampling
1946 * @group_fd: group leader counter fd
1948 SYSCALL_DEFINE5(perf_counter_open,
1949 const struct perf_counter_hw_event __user *, hw_event_uptr,
1950 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
1952 struct perf_counter *counter, *group_leader;
1953 struct perf_counter_hw_event hw_event;
1954 struct perf_counter_context *ctx;
1955 struct file *counter_file = NULL;
1956 struct file *group_file = NULL;
1957 int fput_needed = 0;
1958 int fput_needed2 = 0;
1961 /* for future expandability... */
1965 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
1969 * Get the target context (task or percpu):
1971 ctx = find_get_context(pid, cpu);
1973 return PTR_ERR(ctx);
1976 * Look up the group leader (we will attach this counter to it):
1978 group_leader = NULL;
1979 if (group_fd != -1) {
1981 group_file = fget_light(group_fd, &fput_needed);
1983 goto err_put_context;
1984 if (group_file->f_op != &perf_fops)
1985 goto err_put_context;
1987 group_leader = group_file->private_data;
1989 * Do not allow a recursive hierarchy (this new sibling
1990 * becoming part of another group-sibling):
1992 if (group_leader->group_leader != group_leader)
1993 goto err_put_context;
1995 * Do not allow to attach to a group in a different
1996 * task or CPU context:
1998 if (group_leader->ctx != ctx)
1999 goto err_put_context;
2001 * Only a group leader can be exclusive or pinned
2003 if (hw_event.exclusive || hw_event.pinned)
2004 goto err_put_context;
2008 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
2011 goto err_put_context;
2013 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
2015 goto err_free_put_context;
2017 counter_file = fget_light(ret, &fput_needed2);
2019 goto err_free_put_context;
2021 counter->filp = counter_file;
2022 mutex_lock(&ctx->mutex);
2023 perf_install_in_context(ctx, counter, cpu);
2024 mutex_unlock(&ctx->mutex);
2026 fput_light(counter_file, fput_needed2);
2029 fput_light(group_file, fput_needed);
2033 err_free_put_context:
2043 * Initialize the perf_counter context in a task_struct:
2046 __perf_counter_init_context(struct perf_counter_context *ctx,
2047 struct task_struct *task)
2049 memset(ctx, 0, sizeof(*ctx));
2050 spin_lock_init(&ctx->lock);
2051 mutex_init(&ctx->mutex);
2052 INIT_LIST_HEAD(&ctx->counter_list);
2053 INIT_LIST_HEAD(&ctx->event_list);
2058 * inherit a counter from parent task to child task:
2060 static struct perf_counter *
2061 inherit_counter(struct perf_counter *parent_counter,
2062 struct task_struct *parent,
2063 struct perf_counter_context *parent_ctx,
2064 struct task_struct *child,
2065 struct perf_counter *group_leader,
2066 struct perf_counter_context *child_ctx)
2068 struct perf_counter *child_counter;
2071 * Instead of creating recursive hierarchies of counters,
2072 * we link inherited counters back to the original parent,
2073 * which has a filp for sure, which we use as the reference
2076 if (parent_counter->parent)
2077 parent_counter = parent_counter->parent;
2079 child_counter = perf_counter_alloc(&parent_counter->hw_event,
2080 parent_counter->cpu, child_ctx,
2081 group_leader, GFP_KERNEL);
2086 * Link it up in the child's context:
2088 child_counter->task = child;
2089 list_add_counter(child_counter, child_ctx);
2090 child_ctx->nr_counters++;
2092 child_counter->parent = parent_counter;
2094 * inherit into child's child as well:
2096 child_counter->hw_event.inherit = 1;
2099 * Get a reference to the parent filp - we will fput it
2100 * when the child counter exits. This is safe to do because
2101 * we are in the parent and we know that the filp still
2102 * exists and has a nonzero count:
2104 atomic_long_inc(&parent_counter->filp->f_count);
2107 * Link this into the parent counter's child list
2109 mutex_lock(&parent_counter->mutex);
2110 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
2113 * Make the child state follow the state of the parent counter,
2114 * not its hw_event.disabled bit. We hold the parent's mutex,
2115 * so we won't race with perf_counter_{en,dis}able_family.
2117 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
2118 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
2120 child_counter->state = PERF_COUNTER_STATE_OFF;
2122 mutex_unlock(&parent_counter->mutex);
2124 return child_counter;
2127 static int inherit_group(struct perf_counter *parent_counter,
2128 struct task_struct *parent,
2129 struct perf_counter_context *parent_ctx,
2130 struct task_struct *child,
2131 struct perf_counter_context *child_ctx)
2133 struct perf_counter *leader;
2134 struct perf_counter *sub;
2136 leader = inherit_counter(parent_counter, parent, parent_ctx,
2137 child, NULL, child_ctx);
2140 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
2141 if (!inherit_counter(sub, parent, parent_ctx,
2142 child, leader, child_ctx))
2148 static void sync_child_counter(struct perf_counter *child_counter,
2149 struct perf_counter *parent_counter)
2151 u64 parent_val, child_val;
2153 parent_val = atomic64_read(&parent_counter->count);
2154 child_val = atomic64_read(&child_counter->count);
2157 * Add back the child's count to the parent's count:
2159 atomic64_add(child_val, &parent_counter->count);
2162 * Remove this counter from the parent's list
2164 mutex_lock(&parent_counter->mutex);
2165 list_del_init(&child_counter->child_list);
2166 mutex_unlock(&parent_counter->mutex);
2169 * Release the parent counter, if this was the last
2172 fput(parent_counter->filp);
2176 __perf_counter_exit_task(struct task_struct *child,
2177 struct perf_counter *child_counter,
2178 struct perf_counter_context *child_ctx)
2180 struct perf_counter *parent_counter;
2181 struct perf_counter *sub, *tmp;
2184 * If we do not self-reap then we have to wait for the
2185 * child task to unschedule (it will happen for sure),
2186 * so that its counter is at its final count. (This
2187 * condition triggers rarely - child tasks usually get
2188 * off their CPU before the parent has a chance to
2189 * get this far into the reaping action)
2191 if (child != current) {
2192 wait_task_inactive(child, 0);
2193 list_del_init(&child_counter->list_entry);
2195 struct perf_cpu_context *cpuctx;
2196 unsigned long flags;
2200 * Disable and unlink this counter.
2202 * Be careful about zapping the list - IRQ/NMI context
2203 * could still be processing it:
2205 curr_rq_lock_irq_save(&flags);
2206 perf_flags = hw_perf_save_disable();
2208 cpuctx = &__get_cpu_var(perf_cpu_context);
2210 group_sched_out(child_counter, cpuctx, child_ctx);
2212 list_del_init(&child_counter->list_entry);
2214 child_ctx->nr_counters--;
2216 hw_perf_restore(perf_flags);
2217 curr_rq_unlock_irq_restore(&flags);
2220 parent_counter = child_counter->parent;
2222 * It can happen that parent exits first, and has counters
2223 * that are still around due to the child reference. These
2224 * counters need to be zapped - but otherwise linger.
2226 if (parent_counter) {
2227 sync_child_counter(child_counter, parent_counter);
2228 list_for_each_entry_safe(sub, tmp, &child_counter->sibling_list,
2231 sync_child_counter(sub, sub->parent);
2235 free_counter(child_counter);
2240 * When a child task exits, feed back counter values to parent counters.
2242 * Note: we may be running in child context, but the PID is not hashed
2243 * anymore so new counters will not be added.
2245 void perf_counter_exit_task(struct task_struct *child)
2247 struct perf_counter *child_counter, *tmp;
2248 struct perf_counter_context *child_ctx;
2250 child_ctx = &child->perf_counter_ctx;
2252 if (likely(!child_ctx->nr_counters))
2255 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
2257 __perf_counter_exit_task(child, child_counter, child_ctx);
2261 * Initialize the perf_counter context in task_struct
2263 void perf_counter_init_task(struct task_struct *child)
2265 struct perf_counter_context *child_ctx, *parent_ctx;
2266 struct perf_counter *counter;
2267 struct task_struct *parent = current;
2269 child_ctx = &child->perf_counter_ctx;
2270 parent_ctx = &parent->perf_counter_ctx;
2272 __perf_counter_init_context(child_ctx, child);
2275 * This is executed from the parent task context, so inherit
2276 * counters that have been marked for cloning:
2279 if (likely(!parent_ctx->nr_counters))
2283 * Lock the parent list. No need to lock the child - not PID
2284 * hashed yet and not running, so nobody can access it.
2286 mutex_lock(&parent_ctx->mutex);
2289 * We dont have to disable NMIs - we are only looking at
2290 * the list, not manipulating it:
2292 list_for_each_entry(counter, &parent_ctx->counter_list, list_entry) {
2293 if (!counter->hw_event.inherit)
2296 if (inherit_group(counter, parent,
2297 parent_ctx, child, child_ctx))
2301 mutex_unlock(&parent_ctx->mutex);
2304 static void __cpuinit perf_counter_init_cpu(int cpu)
2306 struct perf_cpu_context *cpuctx;
2308 cpuctx = &per_cpu(perf_cpu_context, cpu);
2309 __perf_counter_init_context(&cpuctx->ctx, NULL);
2311 mutex_lock(&perf_resource_mutex);
2312 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
2313 mutex_unlock(&perf_resource_mutex);
2315 hw_perf_counter_setup(cpu);
2318 #ifdef CONFIG_HOTPLUG_CPU
2319 static void __perf_counter_exit_cpu(void *info)
2321 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
2322 struct perf_counter_context *ctx = &cpuctx->ctx;
2323 struct perf_counter *counter, *tmp;
2325 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
2326 __perf_counter_remove_from_context(counter);
2328 static void perf_counter_exit_cpu(int cpu)
2330 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
2331 struct perf_counter_context *ctx = &cpuctx->ctx;
2333 mutex_lock(&ctx->mutex);
2334 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
2335 mutex_unlock(&ctx->mutex);
2338 static inline void perf_counter_exit_cpu(int cpu) { }
2341 static int __cpuinit
2342 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
2344 unsigned int cpu = (long)hcpu;
2348 case CPU_UP_PREPARE:
2349 case CPU_UP_PREPARE_FROZEN:
2350 perf_counter_init_cpu(cpu);
2353 case CPU_DOWN_PREPARE:
2354 case CPU_DOWN_PREPARE_FROZEN:
2355 perf_counter_exit_cpu(cpu);
2365 static struct notifier_block __cpuinitdata perf_cpu_nb = {
2366 .notifier_call = perf_cpu_notify,
2369 static int __init perf_counter_init(void)
2371 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
2372 (void *)(long)smp_processor_id());
2373 register_cpu_notifier(&perf_cpu_nb);
2377 early_initcall(perf_counter_init);
2379 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
2381 return sprintf(buf, "%d\n", perf_reserved_percpu);
2385 perf_set_reserve_percpu(struct sysdev_class *class,
2389 struct perf_cpu_context *cpuctx;
2393 err = strict_strtoul(buf, 10, &val);
2396 if (val > perf_max_counters)
2399 mutex_lock(&perf_resource_mutex);
2400 perf_reserved_percpu = val;
2401 for_each_online_cpu(cpu) {
2402 cpuctx = &per_cpu(perf_cpu_context, cpu);
2403 spin_lock_irq(&cpuctx->ctx.lock);
2404 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
2405 perf_max_counters - perf_reserved_percpu);
2406 cpuctx->max_pertask = mpt;
2407 spin_unlock_irq(&cpuctx->ctx.lock);
2409 mutex_unlock(&perf_resource_mutex);
2414 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
2416 return sprintf(buf, "%d\n", perf_overcommit);
2420 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
2425 err = strict_strtoul(buf, 10, &val);
2431 mutex_lock(&perf_resource_mutex);
2432 perf_overcommit = val;
2433 mutex_unlock(&perf_resource_mutex);
2438 static SYSDEV_CLASS_ATTR(
2441 perf_show_reserve_percpu,
2442 perf_set_reserve_percpu
2445 static SYSDEV_CLASS_ATTR(
2448 perf_show_overcommit,
2452 static struct attribute *perfclass_attrs[] = {
2453 &attr_reserve_percpu.attr,
2454 &attr_overcommit.attr,
2458 static struct attribute_group perfclass_attr_group = {
2459 .attrs = perfclass_attrs,
2460 .name = "perf_counters",
2463 static int __init perf_counter_sysfs_init(void)
2465 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
2466 &perfclass_attr_group);
2468 device_initcall(perf_counter_sysfs_init);