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
;
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
;
231 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
236 if (rt_b
->rt_runtime
== RUNTIME_INF
)
239 span
= sched_rt_period_mask();
240 for_each_cpu_mask(i
, span
) {
242 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
243 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
245 spin_lock(&rq
->lock
);
246 if (rt_rq
->rt_time
) {
249 spin_lock(&rt_rq
->rt_runtime_lock
);
250 runtime
= rt_rq
->rt_runtime
;
251 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
252 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
253 rt_rq
->rt_throttled
= 0;
256 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
258 spin_unlock(&rt_rq
->rt_runtime_lock
);
262 sched_rt_rq_enqueue(rt_rq
);
263 spin_unlock(&rq
->lock
);
270 static int balance_runtime(struct rt_rq
*rt_rq
)
272 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
273 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
274 int i
, weight
, more
= 0;
277 weight
= cpus_weight(rd
->span
);
279 spin_lock(&rt_b
->rt_runtime_lock
);
280 rt_period
= ktime_to_ns(rt_b
->rt_period
);
281 for_each_cpu_mask(i
, rd
->span
) {
282 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
288 spin_lock(&iter
->rt_runtime_lock
);
289 if (iter
->rt_runtime
== RUNTIME_INF
)
292 diff
= iter
->rt_runtime
- iter
->rt_time
;
294 do_div(diff
, weight
);
295 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
296 diff
= rt_period
- rt_rq
->rt_runtime
;
297 iter
->rt_runtime
-= diff
;
298 rt_rq
->rt_runtime
+= diff
;
300 if (rt_rq
->rt_runtime
== rt_period
) {
301 spin_unlock(&iter
->rt_runtime_lock
);
306 spin_unlock(&iter
->rt_runtime_lock
);
308 spin_unlock(&rt_b
->rt_runtime_lock
);
313 static void __disable_runtime(struct rq
*rq
)
315 struct root_domain
*rd
= rq
->rd
;
318 if (unlikely(!scheduler_running
))
321 for_each_leaf_rt_rq(rt_rq
, rq
) {
322 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
326 spin_lock(&rt_b
->rt_runtime_lock
);
327 spin_lock(&rt_rq
->rt_runtime_lock
);
328 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
329 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
331 spin_unlock(&rt_rq
->rt_runtime_lock
);
333 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
335 for_each_cpu_mask(i
, rd
->span
) {
336 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
342 spin_lock(&iter
->rt_runtime_lock
);
344 diff
= min_t(s64
, iter
->rt_runtime
, want
);
345 iter
->rt_runtime
-= diff
;
348 iter
->rt_runtime
-= want
;
351 spin_unlock(&iter
->rt_runtime_lock
);
357 spin_lock(&rt_rq
->rt_runtime_lock
);
360 rt_rq
->rt_runtime
= RUNTIME_INF
;
361 spin_unlock(&rt_rq
->rt_runtime_lock
);
362 spin_unlock(&rt_b
->rt_runtime_lock
);
366 static void disable_runtime(struct rq
*rq
)
370 spin_lock_irqsave(&rq
->lock
, flags
);
371 __disable_runtime(rq
);
372 spin_unlock_irqrestore(&rq
->lock
, flags
);
375 static void __enable_runtime(struct rq
*rq
)
377 struct root_domain
*rd
= rq
->rd
;
380 if (unlikely(!scheduler_running
))
383 for_each_leaf_rt_rq(rt_rq
, rq
) {
384 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
386 spin_lock(&rt_b
->rt_runtime_lock
);
387 spin_lock(&rt_rq
->rt_runtime_lock
);
388 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
390 spin_unlock(&rt_rq
->rt_runtime_lock
);
391 spin_unlock(&rt_b
->rt_runtime_lock
);
395 static void enable_runtime(struct rq
*rq
)
399 spin_lock_irqsave(&rq
->lock
, flags
);
400 __enable_runtime(rq
);
401 spin_unlock_irqrestore(&rq
->lock
, flags
);
406 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
408 #ifdef CONFIG_RT_GROUP_SCHED
409 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
412 return rt_rq
->highest_prio
;
415 return rt_task_of(rt_se
)->prio
;
418 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
420 u64 runtime
= sched_rt_runtime(rt_rq
);
422 if (runtime
== RUNTIME_INF
)
425 if (rt_rq
->rt_throttled
)
426 return rt_rq_throttled(rt_rq
);
428 if (sched_rt_runtime(rt_rq
) >= sched_rt_period(rt_rq
))
432 if (rt_rq
->rt_time
> runtime
) {
433 spin_unlock(&rt_rq
->rt_runtime_lock
);
434 balance_runtime(rt_rq
);
435 spin_lock(&rt_rq
->rt_runtime_lock
);
437 runtime
= sched_rt_runtime(rt_rq
);
438 if (runtime
== RUNTIME_INF
)
443 if (rt_rq
->rt_time
> runtime
) {
444 rt_rq
->rt_throttled
= 1;
445 if (rt_rq_throttled(rt_rq
)) {
446 sched_rt_rq_dequeue(rt_rq
);
455 * Update the current task's runtime statistics. Skip current tasks that
456 * are not in our scheduling class.
458 static void update_curr_rt(struct rq
*rq
)
460 struct task_struct
*curr
= rq
->curr
;
461 struct sched_rt_entity
*rt_se
= &curr
->rt
;
462 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
465 if (!task_has_rt_policy(curr
))
468 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
469 if (unlikely((s64
)delta_exec
< 0))
472 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
474 curr
->se
.sum_exec_runtime
+= delta_exec
;
475 curr
->se
.exec_start
= rq
->clock
;
476 cpuacct_charge(curr
, delta_exec
);
478 for_each_sched_rt_entity(rt_se
) {
479 rt_rq
= rt_rq_of_se(rt_se
);
481 spin_lock(&rt_rq
->rt_runtime_lock
);
482 rt_rq
->rt_time
+= delta_exec
;
483 if (sched_rt_runtime_exceeded(rt_rq
))
485 spin_unlock(&rt_rq
->rt_runtime_lock
);
490 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
492 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
493 rt_rq
->rt_nr_running
++;
494 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
495 if (rt_se_prio(rt_se
) < rt_rq
->highest_prio
) {
496 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
498 rt_rq
->highest_prio
= rt_se_prio(rt_se
);
501 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
,
507 if (rt_se
->nr_cpus_allowed
> 1) {
508 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
510 rq
->rt
.rt_nr_migratory
++;
513 update_rt_migration(rq_of_rt_rq(rt_rq
));
515 #ifdef CONFIG_RT_GROUP_SCHED
516 if (rt_se_boosted(rt_se
))
517 rt_rq
->rt_nr_boosted
++;
520 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
522 start_rt_bandwidth(&def_rt_bandwidth
);
527 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
530 int highest_prio
= rt_rq
->highest_prio
;
533 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
534 WARN_ON(!rt_rq
->rt_nr_running
);
535 rt_rq
->rt_nr_running
--;
536 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
537 if (rt_rq
->rt_nr_running
) {
538 struct rt_prio_array
*array
;
540 WARN_ON(rt_se_prio(rt_se
) < rt_rq
->highest_prio
);
541 if (rt_se_prio(rt_se
) == rt_rq
->highest_prio
) {
543 array
= &rt_rq
->active
;
544 rt_rq
->highest_prio
=
545 sched_find_first_bit(array
->bitmap
);
546 } /* otherwise leave rq->highest prio alone */
548 rt_rq
->highest_prio
= MAX_RT_PRIO
;
551 if (rt_se
->nr_cpus_allowed
> 1) {
552 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
553 rq
->rt
.rt_nr_migratory
--;
556 if (rt_rq
->highest_prio
!= highest_prio
) {
557 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
560 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
,
561 rt_rq
->highest_prio
);
564 update_rt_migration(rq_of_rt_rq(rt_rq
));
565 #endif /* CONFIG_SMP */
566 #ifdef CONFIG_RT_GROUP_SCHED
567 if (rt_se_boosted(rt_se
))
568 rt_rq
->rt_nr_boosted
--;
570 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
574 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
576 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
577 struct rt_prio_array
*array
= &rt_rq
->active
;
578 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
580 if (group_rq
&& rt_rq_throttled(group_rq
))
583 if (rt_se
->nr_cpus_allowed
== 1)
584 list_add_tail(&rt_se
->run_list
,
585 array
->xqueue
+ rt_se_prio(rt_se
));
587 list_add_tail(&rt_se
->run_list
,
588 array
->squeue
+ rt_se_prio(rt_se
));
590 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
592 inc_rt_tasks(rt_se
, rt_rq
);
595 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
597 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
598 struct rt_prio_array
*array
= &rt_rq
->active
;
600 list_del_init(&rt_se
->run_list
);
601 if (list_empty(array
->squeue
+ rt_se_prio(rt_se
))
602 && list_empty(array
->xqueue
+ rt_se_prio(rt_se
)))
603 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
605 dec_rt_tasks(rt_se
, rt_rq
);
609 * Because the prio of an upper entry depends on the lower
610 * entries, we must remove entries top - down.
612 static void dequeue_rt_stack(struct task_struct
*p
)
614 struct sched_rt_entity
*rt_se
, *back
= NULL
;
617 for_each_sched_rt_entity(rt_se
) {
622 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
624 dequeue_rt_entity(rt_se
);
629 * Adding/removing a task to/from a priority array:
631 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
633 struct sched_rt_entity
*rt_se
= &p
->rt
;
641 * enqueue everybody, bottom - up.
643 for_each_sched_rt_entity(rt_se
)
644 enqueue_rt_entity(rt_se
);
647 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
649 struct sched_rt_entity
*rt_se
= &p
->rt
;
657 * re-enqueue all non-empty rt_rq entities.
659 for_each_sched_rt_entity(rt_se
) {
660 rt_rq
= group_rt_rq(rt_se
);
661 if (rt_rq
&& rt_rq
->rt_nr_running
)
662 enqueue_rt_entity(rt_se
);
667 * Put task to the end of the run list without the overhead of dequeue
668 * followed by enqueue.
670 * Note: We always enqueue the task to the shared-queue, regardless of its
671 * previous position w.r.t. exclusive vs shared. This is so that exclusive RR
672 * tasks fairly round-robin with all tasks on the runqueue, not just other
676 void requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
)
678 struct rt_prio_array
*array
= &rt_rq
->active
;
680 list_del_init(&rt_se
->run_list
);
681 list_add_tail(&rt_se
->run_list
, array
->squeue
+ rt_se_prio(rt_se
));
684 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
686 struct sched_rt_entity
*rt_se
= &p
->rt
;
689 for_each_sched_rt_entity(rt_se
) {
690 rt_rq
= rt_rq_of_se(rt_se
);
691 requeue_rt_entity(rt_rq
, rt_se
);
695 static void yield_task_rt(struct rq
*rq
)
697 requeue_task_rt(rq
, rq
->curr
);
701 static int find_lowest_rq(struct task_struct
*task
);
703 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
705 struct rq
*rq
= task_rq(p
);
708 * If the current task is an RT task, then
709 * try to see if we can wake this RT task up on another
710 * runqueue. Otherwise simply start this RT task
711 * on its current runqueue.
713 * We want to avoid overloading runqueues. Even if
714 * the RT task is of higher priority than the current RT task.
715 * RT tasks behave differently than other tasks. If
716 * one gets preempted, we try to push it off to another queue.
717 * So trying to keep a preempting RT task on the same
718 * cache hot CPU will force the running RT task to
719 * a cold CPU. So we waste all the cache for the lower
720 * RT task in hopes of saving some of a RT task
721 * that is just being woken and probably will have
724 if (unlikely(rt_task(rq
->curr
)) &&
725 (p
->rt
.nr_cpus_allowed
> 1)) {
726 int cpu
= find_lowest_rq(p
);
728 return (cpu
== -1) ? task_cpu(p
) : cpu
;
732 * Otherwise, just let it ride on the affined RQ and the
733 * post-schedule router will push the preempted task away
737 #endif /* CONFIG_SMP */
739 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
740 struct rt_rq
*rt_rq
);
743 * Preempt the current task with a newly woken task if needed:
745 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
747 if (p
->prio
< rq
->curr
->prio
) {
748 resched_task(rq
->curr
);
756 * - the newly woken task is of equal priority to the current task
757 * - the newly woken task is non-migratable while current is migratable
758 * - current will be preempted on the next reschedule
760 * we should check to see if current can readily move to a different
761 * cpu. If so, we will reschedule to allow the push logic to try
762 * to move current somewhere else, making room for our non-migratable
765 if((p
->prio
== rq
->curr
->prio
)
766 && p
->rt
.nr_cpus_allowed
== 1
767 && rq
->curr
->rt
.nr_cpus_allowed
!= 1
768 && pick_next_rt_entity(rq
, &rq
->rt
) != &rq
->curr
->rt
) {
771 if (cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, &mask
))
773 * There appears to be other cpus that can accept
774 * current, so lets reschedule to try and push it away
776 resched_task(rq
->curr
);
781 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
784 struct rt_prio_array
*array
= &rt_rq
->active
;
785 struct sched_rt_entity
*next
= NULL
;
786 struct list_head
*queue
;
789 idx
= sched_find_first_bit(array
->bitmap
);
790 BUG_ON(idx
>= MAX_RT_PRIO
);
792 queue
= array
->xqueue
+ idx
;
793 if (!list_empty(queue
))
794 next
= list_entry(queue
->next
, struct sched_rt_entity
,
797 queue
= array
->squeue
+ idx
;
798 next
= list_entry(queue
->next
, struct sched_rt_entity
,
805 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
807 struct sched_rt_entity
*rt_se
;
808 struct task_struct
*p
;
813 if (unlikely(!rt_rq
->rt_nr_running
))
816 if (rt_rq_throttled(rt_rq
))
820 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
822 rt_rq
= group_rt_rq(rt_se
);
825 p
= rt_task_of(rt_se
);
826 p
->se
.exec_start
= rq
->clock
;
830 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
833 p
->se
.exec_start
= 0;
838 /* Only try algorithms three times */
839 #define RT_MAX_TRIES 3
841 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
842 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
844 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
846 if (!task_running(rq
, p
) &&
847 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
848 (p
->rt
.nr_cpus_allowed
> 1))
853 /* Return the second highest RT task, NULL otherwise */
854 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
856 struct task_struct
*next
= NULL
;
857 struct sched_rt_entity
*rt_se
;
858 struct rt_prio_array
*array
;
862 for_each_leaf_rt_rq(rt_rq
, rq
) {
863 array
= &rt_rq
->active
;
864 idx
= sched_find_first_bit(array
->bitmap
);
866 if (idx
>= MAX_RT_PRIO
)
868 if (next
&& next
->prio
< idx
)
870 list_for_each_entry(rt_se
, array
->squeue
+ idx
, run_list
) {
871 struct task_struct
*p
= rt_task_of(rt_se
);
872 if (pick_rt_task(rq
, p
, cpu
)) {
878 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
886 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
888 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
892 /* "this_cpu" is cheaper to preempt than a remote processor */
893 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
896 first
= first_cpu(*mask
);
897 if (first
!= NR_CPUS
)
903 static int find_lowest_rq(struct task_struct
*task
)
905 struct sched_domain
*sd
;
906 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
907 int this_cpu
= smp_processor_id();
908 int cpu
= task_cpu(task
);
910 if (task
->rt
.nr_cpus_allowed
== 1)
911 return -1; /* No other targets possible */
913 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
914 return -1; /* No targets found */
917 * At this point we have built a mask of cpus representing the
918 * lowest priority tasks in the system. Now we want to elect
919 * the best one based on our affinity and topology.
921 * We prioritize the last cpu that the task executed on since
922 * it is most likely cache-hot in that location.
924 if (cpu_isset(cpu
, *lowest_mask
))
928 * Otherwise, we consult the sched_domains span maps to figure
929 * out which cpu is logically closest to our hot cache data.
932 this_cpu
= -1; /* Skip this_cpu opt if the same */
934 for_each_domain(cpu
, sd
) {
935 if (sd
->flags
& SD_WAKE_AFFINE
) {
936 cpumask_t domain_mask
;
939 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
941 best_cpu
= pick_optimal_cpu(this_cpu
,
949 * And finally, if there were no matches within the domains
950 * just give the caller *something* to work with from the compatible
953 return pick_optimal_cpu(this_cpu
, lowest_mask
);
956 /* Will lock the rq it finds */
957 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
959 struct rq
*lowest_rq
= NULL
;
963 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
964 cpu
= find_lowest_rq(task
);
966 if ((cpu
== -1) || (cpu
== rq
->cpu
))
969 lowest_rq
= cpu_rq(cpu
);
971 /* if the prio of this runqueue changed, try again */
972 if (double_lock_balance(rq
, lowest_rq
)) {
974 * We had to unlock the run queue. In
975 * the mean time, task could have
976 * migrated already or had its affinity changed.
977 * Also make sure that it wasn't scheduled on its rq.
979 if (unlikely(task_rq(task
) != rq
||
980 !cpu_isset(lowest_rq
->cpu
,
981 task
->cpus_allowed
) ||
982 task_running(rq
, task
) ||
985 spin_unlock(&lowest_rq
->lock
);
991 /* If this rq is still suitable use it. */
992 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
996 spin_unlock(&lowest_rq
->lock
);
1004 * If the current CPU has more than one RT task, see if the non
1005 * running task can migrate over to a CPU that is running a task
1006 * of lesser priority.
1008 static int push_rt_task(struct rq
*rq
)
1010 struct task_struct
*next_task
;
1011 struct rq
*lowest_rq
;
1013 int paranoid
= RT_MAX_TRIES
;
1015 if (!rq
->rt
.overloaded
)
1018 next_task
= pick_next_highest_task_rt(rq
, -1);
1023 if (unlikely(next_task
== rq
->curr
)) {
1029 * It's possible that the next_task slipped in of
1030 * higher priority than current. If that's the case
1031 * just reschedule current.
1033 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1034 resched_task(rq
->curr
);
1038 /* We might release rq lock */
1039 get_task_struct(next_task
);
1041 /* find_lock_lowest_rq locks the rq if found */
1042 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1044 struct task_struct
*task
;
1046 * find lock_lowest_rq releases rq->lock
1047 * so it is possible that next_task has changed.
1048 * If it has, then try again.
1050 task
= pick_next_highest_task_rt(rq
, -1);
1051 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
1052 put_task_struct(next_task
);
1059 deactivate_task(rq
, next_task
, 0);
1060 set_task_cpu(next_task
, lowest_rq
->cpu
);
1061 activate_task(lowest_rq
, next_task
, 0);
1063 resched_task(lowest_rq
->curr
);
1065 spin_unlock(&lowest_rq
->lock
);
1069 put_task_struct(next_task
);
1075 * TODO: Currently we just use the second highest prio task on
1076 * the queue, and stop when it can't migrate (or there's
1077 * no more RT tasks). There may be a case where a lower
1078 * priority RT task has a different affinity than the
1079 * higher RT task. In this case the lower RT task could
1080 * possibly be able to migrate where as the higher priority
1081 * RT task could not. We currently ignore this issue.
1082 * Enhancements are welcome!
1084 static void push_rt_tasks(struct rq
*rq
)
1086 /* push_rt_task will return true if it moved an RT */
1087 while (push_rt_task(rq
))
1091 static int pull_rt_task(struct rq
*this_rq
)
1093 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
1094 struct task_struct
*p
, *next
;
1097 if (likely(!rt_overloaded(this_rq
)))
1100 next
= pick_next_task_rt(this_rq
);
1102 for_each_cpu_mask(cpu
, this_rq
->rd
->rto_mask
) {
1103 if (this_cpu
== cpu
)
1106 src_rq
= cpu_rq(cpu
);
1108 * We can potentially drop this_rq's lock in
1109 * double_lock_balance, and another CPU could
1110 * steal our next task - hence we must cause
1111 * the caller to recalculate the next task
1114 if (double_lock_balance(this_rq
, src_rq
)) {
1115 struct task_struct
*old_next
= next
;
1117 next
= pick_next_task_rt(this_rq
);
1118 if (next
!= old_next
)
1123 * Are there still pullable RT tasks?
1125 if (src_rq
->rt
.rt_nr_running
<= 1)
1128 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
1131 * Do we have an RT task that preempts
1132 * the to-be-scheduled task?
1134 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
1135 WARN_ON(p
== src_rq
->curr
);
1136 WARN_ON(!p
->se
.on_rq
);
1139 * There's a chance that p is higher in priority
1140 * than what's currently running on its cpu.
1141 * This is just that p is wakeing up and hasn't
1142 * had a chance to schedule. We only pull
1143 * p if it is lower in priority than the
1144 * current task on the run queue or
1145 * this_rq next task is lower in prio than
1146 * the current task on that rq.
1148 if (p
->prio
< src_rq
->curr
->prio
||
1149 (next
&& next
->prio
< src_rq
->curr
->prio
))
1154 deactivate_task(src_rq
, p
, 0);
1155 set_task_cpu(p
, this_cpu
);
1156 activate_task(this_rq
, p
, 0);
1158 * We continue with the search, just in
1159 * case there's an even higher prio task
1160 * in another runqueue. (low likelyhood
1163 * Update next so that we won't pick a task
1164 * on another cpu with a priority lower (or equal)
1165 * than the one we just picked.
1171 spin_unlock(&src_rq
->lock
);
1177 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
1179 /* Try to pull RT tasks here if we lower this rq's prio */
1180 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
1184 static void post_schedule_rt(struct rq
*rq
)
1187 * If we have more than one rt_task queued, then
1188 * see if we can push the other rt_tasks off to other CPUS.
1189 * Note we may release the rq lock, and since
1190 * the lock was owned by prev, we need to release it
1191 * first via finish_lock_switch and then reaquire it here.
1193 if (unlikely(rq
->rt
.overloaded
)) {
1194 spin_lock_irq(&rq
->lock
);
1196 spin_unlock_irq(&rq
->lock
);
1201 * If we are not running and we are not going to reschedule soon, we should
1202 * try to push tasks away now
1204 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
1206 if (!task_running(rq
, p
) &&
1207 !test_tsk_need_resched(rq
->curr
) &&
1212 static unsigned long
1213 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1214 unsigned long max_load_move
,
1215 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1216 int *all_pinned
, int *this_best_prio
)
1218 /* don't touch RT tasks */
1223 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1224 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1226 /* don't touch RT tasks */
1230 static void set_cpus_allowed_rt(struct task_struct
*p
,
1231 const cpumask_t
*new_mask
)
1233 int weight
= cpus_weight(*new_mask
);
1235 BUG_ON(!rt_task(p
));
1238 * Update the migration status of the RQ if we have an RT task
1239 * which is running AND changing its weight value.
1241 if (p
->se
.on_rq
&& (weight
!= p
->rt
.nr_cpus_allowed
)) {
1242 struct rq
*rq
= task_rq(p
);
1244 if ((p
->rt
.nr_cpus_allowed
<= 1) && (weight
> 1)) {
1245 rq
->rt
.rt_nr_migratory
++;
1246 } else if ((p
->rt
.nr_cpus_allowed
> 1) && (weight
<= 1)) {
1247 BUG_ON(!rq
->rt
.rt_nr_migratory
);
1248 rq
->rt
.rt_nr_migratory
--;
1251 update_rt_migration(rq
);
1253 if (unlikely(weight
== 1 || p
->rt
.nr_cpus_allowed
== 1))
1255 * If either the new or old weight is a "1", we need
1256 * to requeue to properly move between shared and
1259 requeue_task_rt(rq
, p
);
1262 p
->cpus_allowed
= *new_mask
;
1263 p
->rt
.nr_cpus_allowed
= weight
;
1266 /* Assumes rq->lock is held */
1267 static void rq_online_rt(struct rq
*rq
)
1269 if (rq
->rt
.overloaded
)
1270 rt_set_overload(rq
);
1272 __enable_runtime(rq
);
1274 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
);
1277 /* Assumes rq->lock is held */
1278 static void rq_offline_rt(struct rq
*rq
)
1280 if (rq
->rt
.overloaded
)
1281 rt_clear_overload(rq
);
1283 __disable_runtime(rq
);
1285 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
1289 * When switch from the rt queue, we bring ourselves to a position
1290 * that we might want to pull RT tasks from other runqueues.
1292 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
,
1296 * If there are other RT tasks then we will reschedule
1297 * and the scheduling of the other RT tasks will handle
1298 * the balancing. But if we are the last RT task
1299 * we may need to handle the pulling of RT tasks
1302 if (!rq
->rt
.rt_nr_running
)
1305 #endif /* CONFIG_SMP */
1308 * When switching a task to RT, we may overload the runqueue
1309 * with RT tasks. In this case we try to push them off to
1312 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
,
1315 int check_resched
= 1;
1318 * If we are already running, then there's nothing
1319 * that needs to be done. But if we are not running
1320 * we may need to preempt the current running task.
1321 * If that current running task is also an RT task
1322 * then see if we can move to another run queue.
1326 if (rq
->rt
.overloaded
&& push_rt_task(rq
) &&
1327 /* Don't resched if we changed runqueues */
1330 #endif /* CONFIG_SMP */
1331 if (check_resched
&& p
->prio
< rq
->curr
->prio
)
1332 resched_task(rq
->curr
);
1337 * Priority of the task has changed. This may cause
1338 * us to initiate a push or pull.
1340 static void prio_changed_rt(struct rq
*rq
, struct task_struct
*p
,
1341 int oldprio
, int running
)
1346 * If our priority decreases while running, we
1347 * may need to pull tasks to this runqueue.
1349 if (oldprio
< p
->prio
)
1352 * If there's a higher priority task waiting to run
1353 * then reschedule. Note, the above pull_rt_task
1354 * can release the rq lock and p could migrate.
1355 * Only reschedule if p is still on the same runqueue.
1357 if (p
->prio
> rq
->rt
.highest_prio
&& rq
->curr
== p
)
1360 /* For UP simply resched on drop of prio */
1361 if (oldprio
< p
->prio
)
1363 #endif /* CONFIG_SMP */
1366 * This task is not running, but if it is
1367 * greater than the current running task
1370 if (p
->prio
< rq
->curr
->prio
)
1371 resched_task(rq
->curr
);
1375 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
1377 unsigned long soft
, hard
;
1382 soft
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_cur
;
1383 hard
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_max
;
1385 if (soft
!= RLIM_INFINITY
) {
1389 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
1390 if (p
->rt
.timeout
> next
)
1391 p
->it_sched_expires
= p
->se
.sum_exec_runtime
;
1395 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
1402 * RR tasks need a special form of timeslice management.
1403 * FIFO tasks have no timeslices.
1405 if (p
->policy
!= SCHED_RR
)
1408 if (--p
->rt
.time_slice
)
1411 p
->rt
.time_slice
= DEF_TIMESLICE
;
1414 * Requeue to the end of queue if we are not the only element
1417 if (p
->rt
.run_list
.prev
!= p
->rt
.run_list
.next
) {
1418 requeue_task_rt(rq
, p
);
1419 set_tsk_need_resched(p
);
1423 static void set_curr_task_rt(struct rq
*rq
)
1425 struct task_struct
*p
= rq
->curr
;
1427 p
->se
.exec_start
= rq
->clock
;
1430 static const struct sched_class rt_sched_class
= {
1431 .next
= &fair_sched_class
,
1432 .enqueue_task
= enqueue_task_rt
,
1433 .dequeue_task
= dequeue_task_rt
,
1434 .yield_task
= yield_task_rt
,
1436 .select_task_rq
= select_task_rq_rt
,
1437 #endif /* CONFIG_SMP */
1439 .check_preempt_curr
= check_preempt_curr_rt
,
1441 .pick_next_task
= pick_next_task_rt
,
1442 .put_prev_task
= put_prev_task_rt
,
1445 .load_balance
= load_balance_rt
,
1446 .move_one_task
= move_one_task_rt
,
1447 .set_cpus_allowed
= set_cpus_allowed_rt
,
1448 .rq_online
= rq_online_rt
,
1449 .rq_offline
= rq_offline_rt
,
1450 .pre_schedule
= pre_schedule_rt
,
1451 .post_schedule
= post_schedule_rt
,
1452 .task_wake_up
= task_wake_up_rt
,
1453 .switched_from
= switched_from_rt
,
1456 .set_curr_task
= set_curr_task_rt
,
1457 .task_tick
= task_tick_rt
,
1459 .prio_changed
= prio_changed_rt
,
1460 .switched_to
= switched_to_rt
,