2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
8 static inline int rt_overloaded(struct rq
*rq
)
10 return atomic_read(&rq
->rd
->rto_count
);
13 static inline void rt_set_overload(struct rq
*rq
)
18 cpu_set(rq
->cpu
, rq
->rd
->rto_mask
);
20 * Make sure the mask is visible before we set
21 * the overload count. That is checked to determine
22 * if we should look at the mask. It would be a shame
23 * if we looked at the mask, but the mask was not
27 atomic_inc(&rq
->rd
->rto_count
);
30 static inline void rt_clear_overload(struct rq
*rq
)
35 /* the order here really doesn't matter */
36 atomic_dec(&rq
->rd
->rto_count
);
37 cpu_clear(rq
->cpu
, rq
->rd
->rto_mask
);
40 static void update_rt_migration(struct rq
*rq
)
42 if (rq
->rt
.rt_nr_migratory
&& (rq
->rt
.rt_nr_running
> 1)) {
43 if (!rq
->rt
.overloaded
) {
45 rq
->rt
.overloaded
= 1;
47 } else if (rq
->rt
.overloaded
) {
48 rt_clear_overload(rq
);
49 rq
->rt
.overloaded
= 0;
52 #endif /* CONFIG_SMP */
54 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
56 return container_of(rt_se
, struct task_struct
, rt
);
59 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
61 return !list_empty(&rt_se
->run_list
);
64 #ifdef CONFIG_RT_GROUP_SCHED
66 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
71 return rt_rq
->rt_runtime
;
74 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
76 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
79 #define for_each_leaf_rt_rq(rt_rq, rq) \
80 list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
82 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
87 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
92 #define for_each_sched_rt_entity(rt_se) \
93 for (; rt_se; rt_se = rt_se->parent)
95 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
100 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
);
101 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
);
103 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
105 struct sched_rt_entity
*rt_se
= rt_rq
->rt_se
;
107 if (rt_se
&& !on_rt_rq(rt_se
) && rt_rq
->rt_nr_running
) {
108 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
110 enqueue_rt_entity(rt_se
);
111 if (rt_rq
->highest_prio
< curr
->prio
)
116 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
118 struct sched_rt_entity
*rt_se
= rt_rq
->rt_se
;
120 if (rt_se
&& on_rt_rq(rt_se
))
121 dequeue_rt_entity(rt_se
);
124 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
126 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
129 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
131 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
132 struct task_struct
*p
;
135 return !!rt_rq
->rt_nr_boosted
;
137 p
= rt_task_of(rt_se
);
138 return p
->prio
!= p
->normal_prio
;
142 static inline cpumask_t
sched_rt_period_mask(void)
144 return cpu_rq(smp_processor_id())->rd
->span
;
147 static inline cpumask_t
sched_rt_period_mask(void)
149 return cpu_online_map
;
154 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
156 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
159 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
161 return &rt_rq
->tg
->rt_bandwidth
;
164 #else /* !CONFIG_RT_GROUP_SCHED */
166 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
168 return rt_rq
->rt_runtime
;
171 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
173 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
176 #define for_each_leaf_rt_rq(rt_rq, rq) \
177 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
179 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
181 return container_of(rt_rq
, struct rq
, rt
);
184 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
186 struct task_struct
*p
= rt_task_of(rt_se
);
187 struct rq
*rq
= task_rq(p
);
192 #define for_each_sched_rt_entity(rt_se) \
193 for (; rt_se; rt_se = NULL)
195 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
200 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
204 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
208 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
210 return rt_rq
->rt_throttled
;
213 static inline cpumask_t
sched_rt_period_mask(void)
215 return cpu_online_map
;
219 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
221 return &cpu_rq(cpu
)->rt
;
224 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
226 return &def_rt_bandwidth
;
229 #endif /* CONFIG_RT_GROUP_SCHED */
232 static int do_balance_runtime(struct rt_rq
*rt_rq
)
234 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
235 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
236 int i
, weight
, more
= 0;
239 weight
= cpus_weight(rd
->span
);
241 spin_lock(&rt_b
->rt_runtime_lock
);
242 rt_period
= ktime_to_ns(rt_b
->rt_period
);
243 for_each_cpu_mask_nr(i
, rd
->span
) {
244 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
250 spin_lock(&iter
->rt_runtime_lock
);
251 if (iter
->rt_runtime
== RUNTIME_INF
)
254 diff
= iter
->rt_runtime
- iter
->rt_time
;
256 diff
= div_u64((u64
)diff
, weight
);
257 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
258 diff
= rt_period
- rt_rq
->rt_runtime
;
259 iter
->rt_runtime
-= diff
;
260 rt_rq
->rt_runtime
+= diff
;
262 if (rt_rq
->rt_runtime
== rt_period
) {
263 spin_unlock(&iter
->rt_runtime_lock
);
268 spin_unlock(&iter
->rt_runtime_lock
);
270 spin_unlock(&rt_b
->rt_runtime_lock
);
275 static void __disable_runtime(struct rq
*rq
)
277 struct root_domain
*rd
= rq
->rd
;
280 if (unlikely(!scheduler_running
))
283 for_each_leaf_rt_rq(rt_rq
, rq
) {
284 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
288 spin_lock(&rt_b
->rt_runtime_lock
);
289 spin_lock(&rt_rq
->rt_runtime_lock
);
290 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
291 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
293 spin_unlock(&rt_rq
->rt_runtime_lock
);
295 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
297 for_each_cpu_mask(i
, rd
->span
) {
298 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
301 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
304 spin_lock(&iter
->rt_runtime_lock
);
306 diff
= min_t(s64
, iter
->rt_runtime
, want
);
307 iter
->rt_runtime
-= diff
;
310 iter
->rt_runtime
-= want
;
313 spin_unlock(&iter
->rt_runtime_lock
);
319 spin_lock(&rt_rq
->rt_runtime_lock
);
322 rt_rq
->rt_runtime
= RUNTIME_INF
;
323 spin_unlock(&rt_rq
->rt_runtime_lock
);
324 spin_unlock(&rt_b
->rt_runtime_lock
);
328 static void disable_runtime(struct rq
*rq
)
332 spin_lock_irqsave(&rq
->lock
, flags
);
333 __disable_runtime(rq
);
334 spin_unlock_irqrestore(&rq
->lock
, flags
);
337 static void __enable_runtime(struct rq
*rq
)
341 if (unlikely(!scheduler_running
))
344 for_each_leaf_rt_rq(rt_rq
, rq
) {
345 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
347 spin_lock(&rt_b
->rt_runtime_lock
);
348 spin_lock(&rt_rq
->rt_runtime_lock
);
349 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
351 spin_unlock(&rt_rq
->rt_runtime_lock
);
352 spin_unlock(&rt_b
->rt_runtime_lock
);
356 static void enable_runtime(struct rq
*rq
)
360 spin_lock_irqsave(&rq
->lock
, flags
);
361 __enable_runtime(rq
);
362 spin_unlock_irqrestore(&rq
->lock
, flags
);
365 static int balance_runtime(struct rt_rq
*rt_rq
)
369 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
370 spin_unlock(&rt_rq
->rt_runtime_lock
);
371 more
= do_balance_runtime(rt_rq
);
372 spin_lock(&rt_rq
->rt_runtime_lock
);
377 #else /* !CONFIG_SMP */
378 static inline int balance_runtime(struct rt_rq
*rt_rq
)
382 #endif /* CONFIG_SMP */
384 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
389 if (rt_b
->rt_runtime
== RUNTIME_INF
)
392 span
= sched_rt_period_mask();
393 for_each_cpu_mask(i
, span
) {
395 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
396 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
398 spin_lock(&rq
->lock
);
399 if (rt_rq
->rt_time
) {
402 spin_lock(&rt_rq
->rt_runtime_lock
);
403 if (rt_rq
->rt_throttled
)
404 balance_runtime(rt_rq
);
405 runtime
= rt_rq
->rt_runtime
;
406 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
407 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
408 rt_rq
->rt_throttled
= 0;
411 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
413 spin_unlock(&rt_rq
->rt_runtime_lock
);
414 } else if (rt_rq
->rt_nr_running
)
418 sched_rt_rq_enqueue(rt_rq
);
419 spin_unlock(&rq
->lock
);
425 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
427 #ifdef CONFIG_RT_GROUP_SCHED
428 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
431 return rt_rq
->highest_prio
;
434 return rt_task_of(rt_se
)->prio
;
437 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
439 u64 runtime
= sched_rt_runtime(rt_rq
);
441 if (rt_rq
->rt_throttled
)
442 return rt_rq_throttled(rt_rq
);
444 if (sched_rt_runtime(rt_rq
) >= sched_rt_period(rt_rq
))
447 balance_runtime(rt_rq
);
448 runtime
= sched_rt_runtime(rt_rq
);
449 if (runtime
== RUNTIME_INF
)
452 if (rt_rq
->rt_time
> runtime
) {
453 rt_rq
->rt_throttled
= 1;
454 if (rt_rq_throttled(rt_rq
)) {
455 sched_rt_rq_dequeue(rt_rq
);
464 * Update the current task's runtime statistics. Skip current tasks that
465 * are not in our scheduling class.
467 static void update_curr_rt(struct rq
*rq
)
469 struct task_struct
*curr
= rq
->curr
;
470 struct sched_rt_entity
*rt_se
= &curr
->rt
;
471 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
474 if (!task_has_rt_policy(curr
))
477 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
478 if (unlikely((s64
)delta_exec
< 0))
481 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
483 curr
->se
.sum_exec_runtime
+= delta_exec
;
484 curr
->se
.exec_start
= rq
->clock
;
485 cpuacct_charge(curr
, delta_exec
);
487 for_each_sched_rt_entity(rt_se
) {
488 rt_rq
= rt_rq_of_se(rt_se
);
490 spin_lock(&rt_rq
->rt_runtime_lock
);
491 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
492 rt_rq
->rt_time
+= delta_exec
;
493 if (sched_rt_runtime_exceeded(rt_rq
))
496 spin_unlock(&rt_rq
->rt_runtime_lock
);
501 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
503 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
504 rt_rq
->rt_nr_running
++;
505 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
506 if (rt_se_prio(rt_se
) < rt_rq
->highest_prio
) {
508 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
511 rt_rq
->highest_prio
= rt_se_prio(rt_se
);
514 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
,
520 if (rt_se
->nr_cpus_allowed
> 1) {
521 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
523 rq
->rt
.rt_nr_migratory
++;
526 update_rt_migration(rq_of_rt_rq(rt_rq
));
528 #ifdef CONFIG_RT_GROUP_SCHED
529 if (rt_se_boosted(rt_se
))
530 rt_rq
->rt_nr_boosted
++;
533 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
535 start_rt_bandwidth(&def_rt_bandwidth
);
540 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
543 int highest_prio
= rt_rq
->highest_prio
;
546 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
547 WARN_ON(!rt_rq
->rt_nr_running
);
548 rt_rq
->rt_nr_running
--;
549 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
550 if (rt_rq
->rt_nr_running
) {
551 struct rt_prio_array
*array
;
553 WARN_ON(rt_se_prio(rt_se
) < rt_rq
->highest_prio
);
554 if (rt_se_prio(rt_se
) == rt_rq
->highest_prio
) {
556 array
= &rt_rq
->active
;
557 rt_rq
->highest_prio
=
558 sched_find_first_bit(array
->bitmap
);
559 } /* otherwise leave rq->highest prio alone */
561 rt_rq
->highest_prio
= MAX_RT_PRIO
;
564 if (rt_se
->nr_cpus_allowed
> 1) {
565 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
566 rq
->rt
.rt_nr_migratory
--;
569 if (rt_rq
->highest_prio
!= highest_prio
) {
570 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
573 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
,
574 rt_rq
->highest_prio
);
577 update_rt_migration(rq_of_rt_rq(rt_rq
));
578 #endif /* CONFIG_SMP */
579 #ifdef CONFIG_RT_GROUP_SCHED
580 if (rt_se_boosted(rt_se
))
581 rt_rq
->rt_nr_boosted
--;
583 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
587 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
589 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
590 struct rt_prio_array
*array
= &rt_rq
->active
;
591 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
592 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
595 * Don't enqueue the group if its throttled, or when empty.
596 * The latter is a consequence of the former when a child group
597 * get throttled and the current group doesn't have any other
600 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
))
603 list_add_tail(&rt_se
->run_list
, queue
);
604 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
606 inc_rt_tasks(rt_se
, rt_rq
);
609 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
611 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
612 struct rt_prio_array
*array
= &rt_rq
->active
;
614 list_del_init(&rt_se
->run_list
);
615 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
616 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
618 dec_rt_tasks(rt_se
, rt_rq
);
622 * Because the prio of an upper entry depends on the lower
623 * entries, we must remove entries top - down.
625 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
)
627 struct sched_rt_entity
*back
= NULL
;
629 for_each_sched_rt_entity(rt_se
) {
634 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
636 __dequeue_rt_entity(rt_se
);
640 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
642 dequeue_rt_stack(rt_se
);
643 for_each_sched_rt_entity(rt_se
)
644 __enqueue_rt_entity(rt_se
);
647 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
649 dequeue_rt_stack(rt_se
);
651 for_each_sched_rt_entity(rt_se
) {
652 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
654 if (rt_rq
&& rt_rq
->rt_nr_running
)
655 __enqueue_rt_entity(rt_se
);
660 * Adding/removing a task to/from a priority array:
662 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
664 struct sched_rt_entity
*rt_se
= &p
->rt
;
669 enqueue_rt_entity(rt_se
);
671 inc_cpu_load(rq
, p
->se
.load
.weight
);
674 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
676 struct sched_rt_entity
*rt_se
= &p
->rt
;
679 dequeue_rt_entity(rt_se
);
681 dec_cpu_load(rq
, p
->se
.load
.weight
);
685 * Put task to the end of the run list without the overhead of dequeue
686 * followed by enqueue.
689 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
691 if (on_rt_rq(rt_se
)) {
692 struct rt_prio_array
*array
= &rt_rq
->active
;
693 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
696 list_move(&rt_se
->run_list
, queue
);
698 list_move_tail(&rt_se
->run_list
, queue
);
702 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
704 struct sched_rt_entity
*rt_se
= &p
->rt
;
707 for_each_sched_rt_entity(rt_se
) {
708 rt_rq
= rt_rq_of_se(rt_se
);
709 requeue_rt_entity(rt_rq
, rt_se
, head
);
713 static void yield_task_rt(struct rq
*rq
)
715 requeue_task_rt(rq
, rq
->curr
, 0);
719 static int find_lowest_rq(struct task_struct
*task
);
721 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
723 struct rq
*rq
= task_rq(p
);
726 * If the current task is an RT task, then
727 * try to see if we can wake this RT task up on another
728 * runqueue. Otherwise simply start this RT task
729 * on its current runqueue.
731 * We want to avoid overloading runqueues. Even if
732 * the RT task is of higher priority than the current RT task.
733 * RT tasks behave differently than other tasks. If
734 * one gets preempted, we try to push it off to another queue.
735 * So trying to keep a preempting RT task on the same
736 * cache hot CPU will force the running RT task to
737 * a cold CPU. So we waste all the cache for the lower
738 * RT task in hopes of saving some of a RT task
739 * that is just being woken and probably will have
742 if (unlikely(rt_task(rq
->curr
)) &&
743 (p
->rt
.nr_cpus_allowed
> 1)) {
744 int cpu
= find_lowest_rq(p
);
746 return (cpu
== -1) ? task_cpu(p
) : cpu
;
750 * Otherwise, just let it ride on the affined RQ and the
751 * post-schedule router will push the preempted task away
756 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
760 if (rq
->curr
->rt
.nr_cpus_allowed
== 1)
763 if (p
->rt
.nr_cpus_allowed
!= 1
764 && cpupri_find(&rq
->rd
->cpupri
, p
, &mask
))
767 if (!cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, &mask
))
771 * There appears to be other cpus that can accept
772 * current and none to run 'p', so lets reschedule
773 * to try and push current away:
775 requeue_task_rt(rq
, p
, 1);
776 resched_task(rq
->curr
);
779 #endif /* CONFIG_SMP */
782 * Preempt the current task with a newly woken task if needed:
784 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
786 if (p
->prio
< rq
->curr
->prio
) {
787 resched_task(rq
->curr
);
795 * - the newly woken task is of equal priority to the current task
796 * - the newly woken task is non-migratable while current is migratable
797 * - current will be preempted on the next reschedule
799 * we should check to see if current can readily move to a different
800 * cpu. If so, we will reschedule to allow the push logic to try
801 * to move current somewhere else, making room for our non-migratable
804 if (p
->prio
== rq
->curr
->prio
&& !need_resched())
805 check_preempt_equal_prio(rq
, p
);
809 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
812 struct rt_prio_array
*array
= &rt_rq
->active
;
813 struct sched_rt_entity
*next
= NULL
;
814 struct list_head
*queue
;
817 idx
= sched_find_first_bit(array
->bitmap
);
818 BUG_ON(idx
>= MAX_RT_PRIO
);
820 queue
= array
->queue
+ idx
;
821 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
826 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
828 struct sched_rt_entity
*rt_se
;
829 struct task_struct
*p
;
834 if (unlikely(!rt_rq
->rt_nr_running
))
837 if (rt_rq_throttled(rt_rq
))
841 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
843 rt_rq
= group_rt_rq(rt_se
);
846 p
= rt_task_of(rt_se
);
847 p
->se
.exec_start
= rq
->clock
;
851 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
854 p
->se
.exec_start
= 0;
859 /* Only try algorithms three times */
860 #define RT_MAX_TRIES 3
862 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
863 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
);
865 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
867 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
869 if (!task_running(rq
, p
) &&
870 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
871 (p
->rt
.nr_cpus_allowed
> 1))
876 /* Return the second highest RT task, NULL otherwise */
877 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
879 struct task_struct
*next
= NULL
;
880 struct sched_rt_entity
*rt_se
;
881 struct rt_prio_array
*array
;
885 for_each_leaf_rt_rq(rt_rq
, rq
) {
886 array
= &rt_rq
->active
;
887 idx
= sched_find_first_bit(array
->bitmap
);
889 if (idx
>= MAX_RT_PRIO
)
891 if (next
&& next
->prio
< idx
)
893 list_for_each_entry(rt_se
, array
->queue
+ idx
, run_list
) {
894 struct task_struct
*p
= rt_task_of(rt_se
);
895 if (pick_rt_task(rq
, p
, cpu
)) {
901 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
909 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
911 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
915 /* "this_cpu" is cheaper to preempt than a remote processor */
916 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
919 first
= first_cpu(*mask
);
920 if (first
!= NR_CPUS
)
926 static int find_lowest_rq(struct task_struct
*task
)
928 struct sched_domain
*sd
;
929 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
930 int this_cpu
= smp_processor_id();
931 int cpu
= task_cpu(task
);
933 if (task
->rt
.nr_cpus_allowed
== 1)
934 return -1; /* No other targets possible */
936 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
937 return -1; /* No targets found */
940 * Only consider CPUs that are usable for migration.
941 * I guess we might want to change cpupri_find() to ignore those
942 * in the first place.
944 cpus_and(*lowest_mask
, *lowest_mask
, cpu_active_map
);
947 * At this point we have built a mask of cpus representing the
948 * lowest priority tasks in the system. Now we want to elect
949 * the best one based on our affinity and topology.
951 * We prioritize the last cpu that the task executed on since
952 * it is most likely cache-hot in that location.
954 if (cpu_isset(cpu
, *lowest_mask
))
958 * Otherwise, we consult the sched_domains span maps to figure
959 * out which cpu is logically closest to our hot cache data.
962 this_cpu
= -1; /* Skip this_cpu opt if the same */
964 for_each_domain(cpu
, sd
) {
965 if (sd
->flags
& SD_WAKE_AFFINE
) {
966 cpumask_t domain_mask
;
969 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
971 best_cpu
= pick_optimal_cpu(this_cpu
,
979 * And finally, if there were no matches within the domains
980 * just give the caller *something* to work with from the compatible
983 return pick_optimal_cpu(this_cpu
, lowest_mask
);
986 /* Will lock the rq it finds */
987 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
989 struct rq
*lowest_rq
= NULL
;
993 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
994 cpu
= find_lowest_rq(task
);
996 if ((cpu
== -1) || (cpu
== rq
->cpu
))
999 lowest_rq
= cpu_rq(cpu
);
1001 /* if the prio of this runqueue changed, try again */
1002 if (double_lock_balance(rq
, lowest_rq
)) {
1004 * We had to unlock the run queue. In
1005 * the mean time, task could have
1006 * migrated already or had its affinity changed.
1007 * Also make sure that it wasn't scheduled on its rq.
1009 if (unlikely(task_rq(task
) != rq
||
1010 !cpu_isset(lowest_rq
->cpu
,
1011 task
->cpus_allowed
) ||
1012 task_running(rq
, task
) ||
1015 spin_unlock(&lowest_rq
->lock
);
1021 /* If this rq is still suitable use it. */
1022 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
1026 double_unlock_balance(rq
, lowest_rq
);
1034 * If the current CPU has more than one RT task, see if the non
1035 * running task can migrate over to a CPU that is running a task
1036 * of lesser priority.
1038 static int push_rt_task(struct rq
*rq
)
1040 struct task_struct
*next_task
;
1041 struct rq
*lowest_rq
;
1043 int paranoid
= RT_MAX_TRIES
;
1045 if (!rq
->rt
.overloaded
)
1048 next_task
= pick_next_highest_task_rt(rq
, -1);
1053 if (unlikely(next_task
== rq
->curr
)) {
1059 * It's possible that the next_task slipped in of
1060 * higher priority than current. If that's the case
1061 * just reschedule current.
1063 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1064 resched_task(rq
->curr
);
1068 /* We might release rq lock */
1069 get_task_struct(next_task
);
1071 /* find_lock_lowest_rq locks the rq if found */
1072 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1074 struct task_struct
*task
;
1076 * find lock_lowest_rq releases rq->lock
1077 * so it is possible that next_task has changed.
1078 * If it has, then try again.
1080 task
= pick_next_highest_task_rt(rq
, -1);
1081 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
1082 put_task_struct(next_task
);
1089 deactivate_task(rq
, next_task
, 0);
1090 set_task_cpu(next_task
, lowest_rq
->cpu
);
1091 activate_task(lowest_rq
, next_task
, 0);
1093 resched_task(lowest_rq
->curr
);
1095 double_unlock_balance(rq
, lowest_rq
);
1099 put_task_struct(next_task
);
1105 * TODO: Currently we just use the second highest prio task on
1106 * the queue, and stop when it can't migrate (or there's
1107 * no more RT tasks). There may be a case where a lower
1108 * priority RT task has a different affinity than the
1109 * higher RT task. In this case the lower RT task could
1110 * possibly be able to migrate where as the higher priority
1111 * RT task could not. We currently ignore this issue.
1112 * Enhancements are welcome!
1114 static void push_rt_tasks(struct rq
*rq
)
1116 /* push_rt_task will return true if it moved an RT */
1117 while (push_rt_task(rq
))
1121 static int pull_rt_task(struct rq
*this_rq
)
1123 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
1124 struct task_struct
*p
, *next
;
1127 if (likely(!rt_overloaded(this_rq
)))
1130 next
= pick_next_task_rt(this_rq
);
1132 for_each_cpu_mask_nr(cpu
, this_rq
->rd
->rto_mask
) {
1133 if (this_cpu
== cpu
)
1136 src_rq
= cpu_rq(cpu
);
1138 * We can potentially drop this_rq's lock in
1139 * double_lock_balance, and another CPU could
1140 * steal our next task - hence we must cause
1141 * the caller to recalculate the next task
1144 if (double_lock_balance(this_rq
, src_rq
)) {
1145 struct task_struct
*old_next
= next
;
1147 next
= pick_next_task_rt(this_rq
);
1148 if (next
!= old_next
)
1153 * Are there still pullable RT tasks?
1155 if (src_rq
->rt
.rt_nr_running
<= 1)
1158 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
1161 * Do we have an RT task that preempts
1162 * the to-be-scheduled task?
1164 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
1165 WARN_ON(p
== src_rq
->curr
);
1166 WARN_ON(!p
->se
.on_rq
);
1169 * There's a chance that p is higher in priority
1170 * than what's currently running on its cpu.
1171 * This is just that p is wakeing up and hasn't
1172 * had a chance to schedule. We only pull
1173 * p if it is lower in priority than the
1174 * current task on the run queue or
1175 * this_rq next task is lower in prio than
1176 * the current task on that rq.
1178 if (p
->prio
< src_rq
->curr
->prio
||
1179 (next
&& next
->prio
< src_rq
->curr
->prio
))
1184 deactivate_task(src_rq
, p
, 0);
1185 set_task_cpu(p
, this_cpu
);
1186 activate_task(this_rq
, p
, 0);
1188 * We continue with the search, just in
1189 * case there's an even higher prio task
1190 * in another runqueue. (low likelyhood
1193 * Update next so that we won't pick a task
1194 * on another cpu with a priority lower (or equal)
1195 * than the one we just picked.
1201 double_unlock_balance(this_rq
, src_rq
);
1207 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
1209 /* Try to pull RT tasks here if we lower this rq's prio */
1210 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
1214 static void post_schedule_rt(struct rq
*rq
)
1217 * If we have more than one rt_task queued, then
1218 * see if we can push the other rt_tasks off to other CPUS.
1219 * Note we may release the rq lock, and since
1220 * the lock was owned by prev, we need to release it
1221 * first via finish_lock_switch and then reaquire it here.
1223 if (unlikely(rq
->rt
.overloaded
)) {
1224 spin_lock_irq(&rq
->lock
);
1226 spin_unlock_irq(&rq
->lock
);
1231 * If we are not running and we are not going to reschedule soon, we should
1232 * try to push tasks away now
1234 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
1236 if (!task_running(rq
, p
) &&
1237 !test_tsk_need_resched(rq
->curr
) &&
1242 static unsigned long
1243 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1244 unsigned long max_load_move
,
1245 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1246 int *all_pinned
, int *this_best_prio
)
1248 /* don't touch RT tasks */
1253 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1254 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1256 /* don't touch RT tasks */
1260 static void set_cpus_allowed_rt(struct task_struct
*p
,
1261 const cpumask_t
*new_mask
)
1263 int weight
= cpus_weight(*new_mask
);
1265 BUG_ON(!rt_task(p
));
1268 * Update the migration status of the RQ if we have an RT task
1269 * which is running AND changing its weight value.
1271 if (p
->se
.on_rq
&& (weight
!= p
->rt
.nr_cpus_allowed
)) {
1272 struct rq
*rq
= task_rq(p
);
1274 if ((p
->rt
.nr_cpus_allowed
<= 1) && (weight
> 1)) {
1275 rq
->rt
.rt_nr_migratory
++;
1276 } else if ((p
->rt
.nr_cpus_allowed
> 1) && (weight
<= 1)) {
1277 BUG_ON(!rq
->rt
.rt_nr_migratory
);
1278 rq
->rt
.rt_nr_migratory
--;
1281 update_rt_migration(rq
);
1284 p
->cpus_allowed
= *new_mask
;
1285 p
->rt
.nr_cpus_allowed
= weight
;
1288 /* Assumes rq->lock is held */
1289 static void rq_online_rt(struct rq
*rq
)
1291 if (rq
->rt
.overloaded
)
1292 rt_set_overload(rq
);
1294 __enable_runtime(rq
);
1296 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
);
1299 /* Assumes rq->lock is held */
1300 static void rq_offline_rt(struct rq
*rq
)
1302 if (rq
->rt
.overloaded
)
1303 rt_clear_overload(rq
);
1305 __disable_runtime(rq
);
1307 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
1311 * When switch from the rt queue, we bring ourselves to a position
1312 * that we might want to pull RT tasks from other runqueues.
1314 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
,
1318 * If there are other RT tasks then we will reschedule
1319 * and the scheduling of the other RT tasks will handle
1320 * the balancing. But if we are the last RT task
1321 * we may need to handle the pulling of RT tasks
1324 if (!rq
->rt
.rt_nr_running
)
1327 #endif /* CONFIG_SMP */
1330 * When switching a task to RT, we may overload the runqueue
1331 * with RT tasks. In this case we try to push them off to
1334 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
,
1337 int check_resched
= 1;
1340 * If we are already running, then there's nothing
1341 * that needs to be done. But if we are not running
1342 * we may need to preempt the current running task.
1343 * If that current running task is also an RT task
1344 * then see if we can move to another run queue.
1348 if (rq
->rt
.overloaded
&& push_rt_task(rq
) &&
1349 /* Don't resched if we changed runqueues */
1352 #endif /* CONFIG_SMP */
1353 if (check_resched
&& p
->prio
< rq
->curr
->prio
)
1354 resched_task(rq
->curr
);
1359 * Priority of the task has changed. This may cause
1360 * us to initiate a push or pull.
1362 static void prio_changed_rt(struct rq
*rq
, struct task_struct
*p
,
1363 int oldprio
, int running
)
1368 * If our priority decreases while running, we
1369 * may need to pull tasks to this runqueue.
1371 if (oldprio
< p
->prio
)
1374 * If there's a higher priority task waiting to run
1375 * then reschedule. Note, the above pull_rt_task
1376 * can release the rq lock and p could migrate.
1377 * Only reschedule if p is still on the same runqueue.
1379 if (p
->prio
> rq
->rt
.highest_prio
&& rq
->curr
== p
)
1382 /* For UP simply resched on drop of prio */
1383 if (oldprio
< p
->prio
)
1385 #endif /* CONFIG_SMP */
1388 * This task is not running, but if it is
1389 * greater than the current running task
1392 if (p
->prio
< rq
->curr
->prio
)
1393 resched_task(rq
->curr
);
1397 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
1399 unsigned long soft
, hard
;
1404 soft
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_cur
;
1405 hard
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_max
;
1407 if (soft
!= RLIM_INFINITY
) {
1411 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
1412 if (p
->rt
.timeout
> next
)
1413 p
->it_sched_expires
= p
->se
.sum_exec_runtime
;
1417 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
1424 * RR tasks need a special form of timeslice management.
1425 * FIFO tasks have no timeslices.
1427 if (p
->policy
!= SCHED_RR
)
1430 if (--p
->rt
.time_slice
)
1433 p
->rt
.time_slice
= DEF_TIMESLICE
;
1436 * Requeue to the end of queue if we are not the only element
1439 if (p
->rt
.run_list
.prev
!= p
->rt
.run_list
.next
) {
1440 requeue_task_rt(rq
, p
, 0);
1441 set_tsk_need_resched(p
);
1445 static void set_curr_task_rt(struct rq
*rq
)
1447 struct task_struct
*p
= rq
->curr
;
1449 p
->se
.exec_start
= rq
->clock
;
1452 static const struct sched_class rt_sched_class
= {
1453 .next
= &fair_sched_class
,
1454 .enqueue_task
= enqueue_task_rt
,
1455 .dequeue_task
= dequeue_task_rt
,
1456 .yield_task
= yield_task_rt
,
1458 .select_task_rq
= select_task_rq_rt
,
1459 #endif /* CONFIG_SMP */
1461 .check_preempt_curr
= check_preempt_curr_rt
,
1463 .pick_next_task
= pick_next_task_rt
,
1464 .put_prev_task
= put_prev_task_rt
,
1467 .load_balance
= load_balance_rt
,
1468 .move_one_task
= move_one_task_rt
,
1469 .set_cpus_allowed
= set_cpus_allowed_rt
,
1470 .rq_online
= rq_online_rt
,
1471 .rq_offline
= rq_offline_rt
,
1472 .pre_schedule
= pre_schedule_rt
,
1473 .post_schedule
= post_schedule_rt
,
1474 .task_wake_up
= task_wake_up_rt
,
1475 .switched_from
= switched_from_rt
,
1478 .set_curr_task
= set_curr_task_rt
,
1479 .task_tick
= task_tick_rt
,
1481 .prio_changed
= prio_changed_rt
,
1482 .switched_to
= switched_to_rt
,
1485 #ifdef CONFIG_SCHED_DEBUG
1486 extern void print_rt_rq(struct seq_file
*m
, int cpu
, struct rt_rq
*rt_rq
);
1488 static void print_rt_stats(struct seq_file
*m
, int cpu
)
1490 struct rt_rq
*rt_rq
;
1493 for_each_leaf_rt_rq(rt_rq
, cpu_rq(cpu
))
1494 print_rt_rq(m
, cpu
, rt_rq
);
1497 #endif /* CONFIG_SCHED_DEBUG */