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
)
15 cpu_set(rq
->cpu
, rq
->rd
->rto_mask
);
17 * Make sure the mask is visible before we set
18 * the overload count. That is checked to determine
19 * if we should look at the mask. It would be a shame
20 * if we looked at the mask, but the mask was not
24 atomic_inc(&rq
->rd
->rto_count
);
27 static inline void rt_clear_overload(struct rq
*rq
)
29 /* the order here really doesn't matter */
30 atomic_dec(&rq
->rd
->rto_count
);
31 cpu_clear(rq
->cpu
, rq
->rd
->rto_mask
);
34 static void update_rt_migration(struct rq
*rq
)
36 if (rq
->rt
.rt_nr_migratory
&& (rq
->rt
.rt_nr_running
> 1)) {
37 if (!rq
->rt
.overloaded
) {
39 rq
->rt
.overloaded
= 1;
41 } else if (rq
->rt
.overloaded
) {
42 rt_clear_overload(rq
);
43 rq
->rt
.overloaded
= 0;
46 #endif /* CONFIG_SMP */
48 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
50 return container_of(rt_se
, struct task_struct
, rt
);
53 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
55 return !list_empty(&rt_se
->run_list
);
58 #ifdef CONFIG_RT_GROUP_SCHED
60 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
65 return rt_rq
->rt_runtime
;
68 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
70 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
73 #define for_each_leaf_rt_rq(rt_rq, rq) \
74 list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
76 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
81 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
86 #define for_each_sched_rt_entity(rt_se) \
87 for (; rt_se; rt_se = rt_se->parent)
89 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
94 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
);
95 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
);
97 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
99 struct sched_rt_entity
*rt_se
= rt_rq
->rt_se
;
101 if (rt_se
&& !on_rt_rq(rt_se
) && rt_rq
->rt_nr_running
) {
102 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
104 enqueue_rt_entity(rt_se
);
105 if (rt_rq
->highest_prio
< curr
->prio
)
110 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
112 struct sched_rt_entity
*rt_se
= rt_rq
->rt_se
;
114 if (rt_se
&& on_rt_rq(rt_se
))
115 dequeue_rt_entity(rt_se
);
118 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
120 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
123 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
125 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
126 struct task_struct
*p
;
129 return !!rt_rq
->rt_nr_boosted
;
131 p
= rt_task_of(rt_se
);
132 return p
->prio
!= p
->normal_prio
;
136 static inline cpumask_t
sched_rt_period_mask(void)
138 return cpu_rq(smp_processor_id())->rd
->span
;
141 static inline cpumask_t
sched_rt_period_mask(void)
143 return cpu_online_map
;
148 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
150 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
153 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
155 return &rt_rq
->tg
->rt_bandwidth
;
160 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
162 return rt_rq
->rt_runtime
;
165 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
167 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
170 #define for_each_leaf_rt_rq(rt_rq, rq) \
171 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
173 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
175 return container_of(rt_rq
, struct rq
, rt
);
178 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
180 struct task_struct
*p
= rt_task_of(rt_se
);
181 struct rq
*rq
= task_rq(p
);
186 #define for_each_sched_rt_entity(rt_se) \
187 for (; rt_se; rt_se = NULL)
189 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
194 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
198 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
202 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
204 return rt_rq
->rt_throttled
;
207 static inline cpumask_t
sched_rt_period_mask(void)
209 return cpu_online_map
;
213 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
215 return &cpu_rq(cpu
)->rt
;
218 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
220 return &def_rt_bandwidth
;
225 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
230 if (rt_b
->rt_runtime
== RUNTIME_INF
)
233 span
= sched_rt_period_mask();
234 for_each_cpu_mask(i
, span
) {
236 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
237 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
239 spin_lock(&rq
->lock
);
240 if (rt_rq
->rt_time
) {
243 spin_lock(&rt_rq
->rt_runtime_lock
);
244 runtime
= rt_rq
->rt_runtime
;
245 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
246 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
247 rt_rq
->rt_throttled
= 0;
250 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
252 spin_unlock(&rt_rq
->rt_runtime_lock
);
256 sched_rt_rq_enqueue(rt_rq
);
257 spin_unlock(&rq
->lock
);
264 static int balance_runtime(struct rt_rq
*rt_rq
)
266 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
267 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
268 int i
, weight
, more
= 0;
271 weight
= cpus_weight(rd
->span
);
273 spin_lock(&rt_b
->rt_runtime_lock
);
274 rt_period
= ktime_to_ns(rt_b
->rt_period
);
275 for_each_cpu_mask(i
, rd
->span
) {
276 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
282 spin_lock(&iter
->rt_runtime_lock
);
283 diff
= iter
->rt_runtime
- iter
->rt_time
;
285 do_div(diff
, weight
);
286 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
287 diff
= rt_period
- rt_rq
->rt_runtime
;
288 iter
->rt_runtime
-= diff
;
289 rt_rq
->rt_runtime
+= diff
;
291 if (rt_rq
->rt_runtime
== rt_period
) {
292 spin_unlock(&iter
->rt_runtime_lock
);
296 spin_unlock(&iter
->rt_runtime_lock
);
298 spin_unlock(&rt_b
->rt_runtime_lock
);
304 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
306 #ifdef CONFIG_RT_GROUP_SCHED
307 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
310 return rt_rq
->highest_prio
;
313 return rt_task_of(rt_se
)->prio
;
316 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
318 u64 runtime
= sched_rt_runtime(rt_rq
);
320 if (runtime
== RUNTIME_INF
)
323 if (rt_rq
->rt_throttled
)
324 return rt_rq_throttled(rt_rq
);
326 if (sched_rt_runtime(rt_rq
) >= sched_rt_period(rt_rq
))
330 if (rt_rq
->rt_time
> runtime
) {
333 spin_unlock(&rt_rq
->rt_runtime_lock
);
334 more
= balance_runtime(rt_rq
);
335 spin_lock(&rt_rq
->rt_runtime_lock
);
338 runtime
= sched_rt_runtime(rt_rq
);
342 if (rt_rq
->rt_time
> runtime
) {
343 rt_rq
->rt_throttled
= 1;
344 if (rt_rq_throttled(rt_rq
)) {
345 sched_rt_rq_dequeue(rt_rq
);
354 * Update the current task's runtime statistics. Skip current tasks that
355 * are not in our scheduling class.
357 static void update_curr_rt(struct rq
*rq
)
359 struct task_struct
*curr
= rq
->curr
;
360 struct sched_rt_entity
*rt_se
= &curr
->rt
;
361 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
364 if (!task_has_rt_policy(curr
))
367 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
368 if (unlikely((s64
)delta_exec
< 0))
371 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
373 curr
->se
.sum_exec_runtime
+= delta_exec
;
374 curr
->se
.exec_start
= rq
->clock
;
375 cpuacct_charge(curr
, delta_exec
);
377 for_each_sched_rt_entity(rt_se
) {
378 rt_rq
= rt_rq_of_se(rt_se
);
380 spin_lock(&rt_rq
->rt_runtime_lock
);
381 rt_rq
->rt_time
+= delta_exec
;
382 if (sched_rt_runtime_exceeded(rt_rq
))
384 spin_unlock(&rt_rq
->rt_runtime_lock
);
389 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
391 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
392 rt_rq
->rt_nr_running
++;
393 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
394 if (rt_se_prio(rt_se
) < rt_rq
->highest_prio
) {
395 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
396 rt_rq
->highest_prio
= rt_se_prio(rt_se
);
397 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_se_prio(rt_se
));
401 if (rt_se
->nr_cpus_allowed
> 1) {
402 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
403 rq
->rt
.rt_nr_migratory
++;
406 update_rt_migration(rq_of_rt_rq(rt_rq
));
408 #ifdef CONFIG_RT_GROUP_SCHED
409 if (rt_se_boosted(rt_se
))
410 rt_rq
->rt_nr_boosted
++;
413 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
415 start_rt_bandwidth(&def_rt_bandwidth
);
420 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
423 int highest_prio
= rt_rq
->highest_prio
;
426 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
427 WARN_ON(!rt_rq
->rt_nr_running
);
428 rt_rq
->rt_nr_running
--;
429 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
430 if (rt_rq
->rt_nr_running
) {
431 struct rt_prio_array
*array
;
433 WARN_ON(rt_se_prio(rt_se
) < rt_rq
->highest_prio
);
434 if (rt_se_prio(rt_se
) == rt_rq
->highest_prio
) {
436 array
= &rt_rq
->active
;
437 rt_rq
->highest_prio
=
438 sched_find_first_bit(array
->bitmap
);
439 } /* otherwise leave rq->highest prio alone */
441 rt_rq
->highest_prio
= MAX_RT_PRIO
;
444 if (rt_se
->nr_cpus_allowed
> 1) {
445 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
446 rq
->rt
.rt_nr_migratory
--;
449 if (rt_rq
->highest_prio
!= highest_prio
) {
450 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
451 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_rq
->highest_prio
);
454 update_rt_migration(rq_of_rt_rq(rt_rq
));
455 #endif /* CONFIG_SMP */
456 #ifdef CONFIG_RT_GROUP_SCHED
457 if (rt_se_boosted(rt_se
))
458 rt_rq
->rt_nr_boosted
--;
460 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
464 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
466 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
467 struct rt_prio_array
*array
= &rt_rq
->active
;
468 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
470 if (group_rq
&& rt_rq_throttled(group_rq
))
473 if (rt_se
->nr_cpus_allowed
== 1)
474 list_add_tail(&rt_se
->run_list
,
475 array
->xqueue
+ rt_se_prio(rt_se
));
477 list_add_tail(&rt_se
->run_list
,
478 array
->squeue
+ rt_se_prio(rt_se
));
480 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
482 inc_rt_tasks(rt_se
, rt_rq
);
485 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
487 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
488 struct rt_prio_array
*array
= &rt_rq
->active
;
490 list_del_init(&rt_se
->run_list
);
491 if (list_empty(array
->squeue
+ rt_se_prio(rt_se
))
492 && list_empty(array
->xqueue
+ rt_se_prio(rt_se
)))
493 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
495 dec_rt_tasks(rt_se
, rt_rq
);
499 * Because the prio of an upper entry depends on the lower
500 * entries, we must remove entries top - down.
502 static void dequeue_rt_stack(struct task_struct
*p
)
504 struct sched_rt_entity
*rt_se
, *back
= NULL
;
507 for_each_sched_rt_entity(rt_se
) {
512 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
514 dequeue_rt_entity(rt_se
);
519 * Adding/removing a task to/from a priority array:
521 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
523 struct sched_rt_entity
*rt_se
= &p
->rt
;
531 * enqueue everybody, bottom - up.
533 for_each_sched_rt_entity(rt_se
)
534 enqueue_rt_entity(rt_se
);
537 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
539 struct sched_rt_entity
*rt_se
= &p
->rt
;
547 * re-enqueue all non-empty rt_rq entities.
549 for_each_sched_rt_entity(rt_se
) {
550 rt_rq
= group_rt_rq(rt_se
);
551 if (rt_rq
&& rt_rq
->rt_nr_running
)
552 enqueue_rt_entity(rt_se
);
557 * Put task to the end of the run list without the overhead of dequeue
558 * followed by enqueue.
560 * Note: We always enqueue the task to the shared-queue, regardless of its
561 * previous position w.r.t. exclusive vs shared. This is so that exclusive RR
562 * tasks fairly round-robin with all tasks on the runqueue, not just other
566 void requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
)
568 struct rt_prio_array
*array
= &rt_rq
->active
;
570 list_del_init(&rt_se
->run_list
);
571 list_add_tail(&rt_se
->run_list
, array
->squeue
+ rt_se_prio(rt_se
));
574 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
576 struct sched_rt_entity
*rt_se
= &p
->rt
;
579 for_each_sched_rt_entity(rt_se
) {
580 rt_rq
= rt_rq_of_se(rt_se
);
581 requeue_rt_entity(rt_rq
, rt_se
);
585 static void yield_task_rt(struct rq
*rq
)
587 requeue_task_rt(rq
, rq
->curr
);
591 static int find_lowest_rq(struct task_struct
*task
);
593 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
595 struct rq
*rq
= task_rq(p
);
598 * If the current task is an RT task, then
599 * try to see if we can wake this RT task up on another
600 * runqueue. Otherwise simply start this RT task
601 * on its current runqueue.
603 * We want to avoid overloading runqueues. Even if
604 * the RT task is of higher priority than the current RT task.
605 * RT tasks behave differently than other tasks. If
606 * one gets preempted, we try to push it off to another queue.
607 * So trying to keep a preempting RT task on the same
608 * cache hot CPU will force the running RT task to
609 * a cold CPU. So we waste all the cache for the lower
610 * RT task in hopes of saving some of a RT task
611 * that is just being woken and probably will have
614 if (unlikely(rt_task(rq
->curr
)) &&
615 (p
->rt
.nr_cpus_allowed
> 1)) {
616 int cpu
= find_lowest_rq(p
);
618 return (cpu
== -1) ? task_cpu(p
) : cpu
;
622 * Otherwise, just let it ride on the affined RQ and the
623 * post-schedule router will push the preempted task away
627 #endif /* CONFIG_SMP */
629 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
630 struct rt_rq
*rt_rq
);
633 * Preempt the current task with a newly woken task if needed:
635 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
637 if (p
->prio
< rq
->curr
->prio
) {
638 resched_task(rq
->curr
);
646 * - the newly woken task is of equal priority to the current task
647 * - the newly woken task is non-migratable while current is migratable
648 * - current will be preempted on the next reschedule
650 * we should check to see if current can readily move to a different
651 * cpu. If so, we will reschedule to allow the push logic to try
652 * to move current somewhere else, making room for our non-migratable
655 if((p
->prio
== rq
->curr
->prio
)
656 && p
->rt
.nr_cpus_allowed
== 1
657 && rq
->curr
->rt
.nr_cpus_allowed
!= 1
658 && pick_next_rt_entity(rq
, &rq
->rt
) != &rq
->curr
->rt
) {
661 if (cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, &mask
))
663 * There appears to be other cpus that can accept
664 * current, so lets reschedule to try and push it away
666 resched_task(rq
->curr
);
671 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
674 struct rt_prio_array
*array
= &rt_rq
->active
;
675 struct sched_rt_entity
*next
= NULL
;
676 struct list_head
*queue
;
679 idx
= sched_find_first_bit(array
->bitmap
);
680 BUG_ON(idx
>= MAX_RT_PRIO
);
682 queue
= array
->xqueue
+ idx
;
683 if (!list_empty(queue
))
684 next
= list_entry(queue
->next
, struct sched_rt_entity
,
687 queue
= array
->squeue
+ idx
;
688 next
= list_entry(queue
->next
, struct sched_rt_entity
,
695 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
697 struct sched_rt_entity
*rt_se
;
698 struct task_struct
*p
;
703 if (unlikely(!rt_rq
->rt_nr_running
))
706 if (rt_rq_throttled(rt_rq
))
710 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
712 rt_rq
= group_rt_rq(rt_se
);
715 p
= rt_task_of(rt_se
);
716 p
->se
.exec_start
= rq
->clock
;
720 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
723 p
->se
.exec_start
= 0;
728 /* Only try algorithms three times */
729 #define RT_MAX_TRIES 3
731 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
732 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
734 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
736 if (!task_running(rq
, p
) &&
737 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
738 (p
->rt
.nr_cpus_allowed
> 1))
743 /* Return the second highest RT task, NULL otherwise */
744 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
746 struct task_struct
*next
= NULL
;
747 struct sched_rt_entity
*rt_se
;
748 struct rt_prio_array
*array
;
752 for_each_leaf_rt_rq(rt_rq
, rq
) {
753 array
= &rt_rq
->active
;
754 idx
= sched_find_first_bit(array
->bitmap
);
756 if (idx
>= MAX_RT_PRIO
)
758 if (next
&& next
->prio
< idx
)
760 list_for_each_entry(rt_se
, array
->squeue
+ idx
, run_list
) {
761 struct task_struct
*p
= rt_task_of(rt_se
);
762 if (pick_rt_task(rq
, p
, cpu
)) {
768 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
776 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
778 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
782 /* "this_cpu" is cheaper to preempt than a remote processor */
783 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
786 first
= first_cpu(*mask
);
787 if (first
!= NR_CPUS
)
793 static int find_lowest_rq(struct task_struct
*task
)
795 struct sched_domain
*sd
;
796 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
797 int this_cpu
= smp_processor_id();
798 int cpu
= task_cpu(task
);
800 if (task
->rt
.nr_cpus_allowed
== 1)
801 return -1; /* No other targets possible */
803 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
804 return -1; /* No targets found */
807 * At this point we have built a mask of cpus representing the
808 * lowest priority tasks in the system. Now we want to elect
809 * the best one based on our affinity and topology.
811 * We prioritize the last cpu that the task executed on since
812 * it is most likely cache-hot in that location.
814 if (cpu_isset(cpu
, *lowest_mask
))
818 * Otherwise, we consult the sched_domains span maps to figure
819 * out which cpu is logically closest to our hot cache data.
822 this_cpu
= -1; /* Skip this_cpu opt if the same */
824 for_each_domain(cpu
, sd
) {
825 if (sd
->flags
& SD_WAKE_AFFINE
) {
826 cpumask_t domain_mask
;
829 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
831 best_cpu
= pick_optimal_cpu(this_cpu
,
839 * And finally, if there were no matches within the domains
840 * just give the caller *something* to work with from the compatible
843 return pick_optimal_cpu(this_cpu
, lowest_mask
);
846 /* Will lock the rq it finds */
847 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
849 struct rq
*lowest_rq
= NULL
;
853 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
854 cpu
= find_lowest_rq(task
);
856 if ((cpu
== -1) || (cpu
== rq
->cpu
))
859 lowest_rq
= cpu_rq(cpu
);
861 /* if the prio of this runqueue changed, try again */
862 if (double_lock_balance(rq
, lowest_rq
)) {
864 * We had to unlock the run queue. In
865 * the mean time, task could have
866 * migrated already or had its affinity changed.
867 * Also make sure that it wasn't scheduled on its rq.
869 if (unlikely(task_rq(task
) != rq
||
870 !cpu_isset(lowest_rq
->cpu
,
871 task
->cpus_allowed
) ||
872 task_running(rq
, task
) ||
875 spin_unlock(&lowest_rq
->lock
);
881 /* If this rq is still suitable use it. */
882 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
886 spin_unlock(&lowest_rq
->lock
);
894 * If the current CPU has more than one RT task, see if the non
895 * running task can migrate over to a CPU that is running a task
896 * of lesser priority.
898 static int push_rt_task(struct rq
*rq
)
900 struct task_struct
*next_task
;
901 struct rq
*lowest_rq
;
903 int paranoid
= RT_MAX_TRIES
;
905 if (!rq
->rt
.overloaded
)
908 next_task
= pick_next_highest_task_rt(rq
, -1);
913 if (unlikely(next_task
== rq
->curr
)) {
919 * It's possible that the next_task slipped in of
920 * higher priority than current. If that's the case
921 * just reschedule current.
923 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
924 resched_task(rq
->curr
);
928 /* We might release rq lock */
929 get_task_struct(next_task
);
931 /* find_lock_lowest_rq locks the rq if found */
932 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
934 struct task_struct
*task
;
936 * find lock_lowest_rq releases rq->lock
937 * so it is possible that next_task has changed.
938 * If it has, then try again.
940 task
= pick_next_highest_task_rt(rq
, -1);
941 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
942 put_task_struct(next_task
);
949 deactivate_task(rq
, next_task
, 0);
950 set_task_cpu(next_task
, lowest_rq
->cpu
);
951 activate_task(lowest_rq
, next_task
, 0);
953 resched_task(lowest_rq
->curr
);
955 spin_unlock(&lowest_rq
->lock
);
959 put_task_struct(next_task
);
965 * TODO: Currently we just use the second highest prio task on
966 * the queue, and stop when it can't migrate (or there's
967 * no more RT tasks). There may be a case where a lower
968 * priority RT task has a different affinity than the
969 * higher RT task. In this case the lower RT task could
970 * possibly be able to migrate where as the higher priority
971 * RT task could not. We currently ignore this issue.
972 * Enhancements are welcome!
974 static void push_rt_tasks(struct rq
*rq
)
976 /* push_rt_task will return true if it moved an RT */
977 while (push_rt_task(rq
))
981 static int pull_rt_task(struct rq
*this_rq
)
983 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
984 struct task_struct
*p
, *next
;
987 if (likely(!rt_overloaded(this_rq
)))
990 next
= pick_next_task_rt(this_rq
);
992 for_each_cpu_mask(cpu
, this_rq
->rd
->rto_mask
) {
996 src_rq
= cpu_rq(cpu
);
998 * We can potentially drop this_rq's lock in
999 * double_lock_balance, and another CPU could
1000 * steal our next task - hence we must cause
1001 * the caller to recalculate the next task
1004 if (double_lock_balance(this_rq
, src_rq
)) {
1005 struct task_struct
*old_next
= next
;
1007 next
= pick_next_task_rt(this_rq
);
1008 if (next
!= old_next
)
1013 * Are there still pullable RT tasks?
1015 if (src_rq
->rt
.rt_nr_running
<= 1)
1018 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
1021 * Do we have an RT task that preempts
1022 * the to-be-scheduled task?
1024 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
1025 WARN_ON(p
== src_rq
->curr
);
1026 WARN_ON(!p
->se
.on_rq
);
1029 * There's a chance that p is higher in priority
1030 * than what's currently running on its cpu.
1031 * This is just that p is wakeing up and hasn't
1032 * had a chance to schedule. We only pull
1033 * p if it is lower in priority than the
1034 * current task on the run queue or
1035 * this_rq next task is lower in prio than
1036 * the current task on that rq.
1038 if (p
->prio
< src_rq
->curr
->prio
||
1039 (next
&& next
->prio
< src_rq
->curr
->prio
))
1044 deactivate_task(src_rq
, p
, 0);
1045 set_task_cpu(p
, this_cpu
);
1046 activate_task(this_rq
, p
, 0);
1048 * We continue with the search, just in
1049 * case there's an even higher prio task
1050 * in another runqueue. (low likelyhood
1053 * Update next so that we won't pick a task
1054 * on another cpu with a priority lower (or equal)
1055 * than the one we just picked.
1061 spin_unlock(&src_rq
->lock
);
1067 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
1069 /* Try to pull RT tasks here if we lower this rq's prio */
1070 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
1074 static void post_schedule_rt(struct rq
*rq
)
1077 * If we have more than one rt_task queued, then
1078 * see if we can push the other rt_tasks off to other CPUS.
1079 * Note we may release the rq lock, and since
1080 * the lock was owned by prev, we need to release it
1081 * first via finish_lock_switch and then reaquire it here.
1083 if (unlikely(rq
->rt
.overloaded
)) {
1084 spin_lock_irq(&rq
->lock
);
1086 spin_unlock_irq(&rq
->lock
);
1091 * If we are not running and we are not going to reschedule soon, we should
1092 * try to push tasks away now
1094 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
1096 if (!task_running(rq
, p
) &&
1097 !test_tsk_need_resched(rq
->curr
) &&
1102 static unsigned long
1103 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1104 unsigned long max_load_move
,
1105 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1106 int *all_pinned
, int *this_best_prio
)
1108 /* don't touch RT tasks */
1113 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1114 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1116 /* don't touch RT tasks */
1120 static void set_cpus_allowed_rt(struct task_struct
*p
,
1121 const cpumask_t
*new_mask
)
1123 int weight
= cpus_weight(*new_mask
);
1125 BUG_ON(!rt_task(p
));
1128 * Update the migration status of the RQ if we have an RT task
1129 * which is running AND changing its weight value.
1131 if (p
->se
.on_rq
&& (weight
!= p
->rt
.nr_cpus_allowed
)) {
1132 struct rq
*rq
= task_rq(p
);
1134 if ((p
->rt
.nr_cpus_allowed
<= 1) && (weight
> 1)) {
1135 rq
->rt
.rt_nr_migratory
++;
1136 } else if ((p
->rt
.nr_cpus_allowed
> 1) && (weight
<= 1)) {
1137 BUG_ON(!rq
->rt
.rt_nr_migratory
);
1138 rq
->rt
.rt_nr_migratory
--;
1141 update_rt_migration(rq
);
1143 if (unlikely(weight
== 1 || p
->rt
.nr_cpus_allowed
== 1))
1145 * If either the new or old weight is a "1", we need
1146 * to requeue to properly move between shared and
1149 requeue_task_rt(rq
, p
);
1152 p
->cpus_allowed
= *new_mask
;
1153 p
->rt
.nr_cpus_allowed
= weight
;
1156 /* Assumes rq->lock is held */
1157 static void join_domain_rt(struct rq
*rq
)
1159 if (rq
->rt
.overloaded
)
1160 rt_set_overload(rq
);
1162 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
);
1165 /* Assumes rq->lock is held */
1166 static void leave_domain_rt(struct rq
*rq
)
1168 if (rq
->rt
.overloaded
)
1169 rt_clear_overload(rq
);
1171 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
1175 * When switch from the rt queue, we bring ourselves to a position
1176 * that we might want to pull RT tasks from other runqueues.
1178 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
,
1182 * If there are other RT tasks then we will reschedule
1183 * and the scheduling of the other RT tasks will handle
1184 * the balancing. But if we are the last RT task
1185 * we may need to handle the pulling of RT tasks
1188 if (!rq
->rt
.rt_nr_running
)
1191 #endif /* CONFIG_SMP */
1194 * When switching a task to RT, we may overload the runqueue
1195 * with RT tasks. In this case we try to push them off to
1198 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
,
1201 int check_resched
= 1;
1204 * If we are already running, then there's nothing
1205 * that needs to be done. But if we are not running
1206 * we may need to preempt the current running task.
1207 * If that current running task is also an RT task
1208 * then see if we can move to another run queue.
1212 if (rq
->rt
.overloaded
&& push_rt_task(rq
) &&
1213 /* Don't resched if we changed runqueues */
1216 #endif /* CONFIG_SMP */
1217 if (check_resched
&& p
->prio
< rq
->curr
->prio
)
1218 resched_task(rq
->curr
);
1223 * Priority of the task has changed. This may cause
1224 * us to initiate a push or pull.
1226 static void prio_changed_rt(struct rq
*rq
, struct task_struct
*p
,
1227 int oldprio
, int running
)
1232 * If our priority decreases while running, we
1233 * may need to pull tasks to this runqueue.
1235 if (oldprio
< p
->prio
)
1238 * If there's a higher priority task waiting to run
1239 * then reschedule. Note, the above pull_rt_task
1240 * can release the rq lock and p could migrate.
1241 * Only reschedule if p is still on the same runqueue.
1243 if (p
->prio
> rq
->rt
.highest_prio
&& rq
->curr
== p
)
1246 /* For UP simply resched on drop of prio */
1247 if (oldprio
< p
->prio
)
1249 #endif /* CONFIG_SMP */
1252 * This task is not running, but if it is
1253 * greater than the current running task
1256 if (p
->prio
< rq
->curr
->prio
)
1257 resched_task(rq
->curr
);
1261 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
1263 unsigned long soft
, hard
;
1268 soft
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_cur
;
1269 hard
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_max
;
1271 if (soft
!= RLIM_INFINITY
) {
1275 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
1276 if (p
->rt
.timeout
> next
)
1277 p
->it_sched_expires
= p
->se
.sum_exec_runtime
;
1281 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
1288 * RR tasks need a special form of timeslice management.
1289 * FIFO tasks have no timeslices.
1291 if (p
->policy
!= SCHED_RR
)
1294 if (--p
->rt
.time_slice
)
1297 p
->rt
.time_slice
= DEF_TIMESLICE
;
1300 * Requeue to the end of queue if we are not the only element
1303 if (p
->rt
.run_list
.prev
!= p
->rt
.run_list
.next
) {
1304 requeue_task_rt(rq
, p
);
1305 set_tsk_need_resched(p
);
1309 static void set_curr_task_rt(struct rq
*rq
)
1311 struct task_struct
*p
= rq
->curr
;
1313 p
->se
.exec_start
= rq
->clock
;
1316 static const struct sched_class rt_sched_class
= {
1317 .next
= &fair_sched_class
,
1318 .enqueue_task
= enqueue_task_rt
,
1319 .dequeue_task
= dequeue_task_rt
,
1320 .yield_task
= yield_task_rt
,
1322 .select_task_rq
= select_task_rq_rt
,
1323 #endif /* CONFIG_SMP */
1325 .check_preempt_curr
= check_preempt_curr_rt
,
1327 .pick_next_task
= pick_next_task_rt
,
1328 .put_prev_task
= put_prev_task_rt
,
1331 .load_balance
= load_balance_rt
,
1332 .move_one_task
= move_one_task_rt
,
1333 .set_cpus_allowed
= set_cpus_allowed_rt
,
1334 .join_domain
= join_domain_rt
,
1335 .leave_domain
= leave_domain_rt
,
1336 .pre_schedule
= pre_schedule_rt
,
1337 .post_schedule
= post_schedule_rt
,
1338 .task_wake_up
= task_wake_up_rt
,
1339 .switched_from
= switched_from_rt
,
1342 .set_curr_task
= set_curr_task_rt
,
1343 .task_tick
= task_tick_rt
,
1345 .prio_changed
= prio_changed_rt
,
1346 .switched_to
= switched_to_rt
,