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 rt_rq
->highest_prio
= rt_se_prio(rt_se
);
398 if (rt_se
->nr_cpus_allowed
> 1) {
399 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
400 rq
->rt
.rt_nr_migratory
++;
403 update_rt_migration(rq_of_rt_rq(rt_rq
));
405 #ifdef CONFIG_RT_GROUP_SCHED
406 if (rt_se_boosted(rt_se
))
407 rt_rq
->rt_nr_boosted
++;
410 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
412 start_rt_bandwidth(&def_rt_bandwidth
);
417 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
419 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
420 WARN_ON(!rt_rq
->rt_nr_running
);
421 rt_rq
->rt_nr_running
--;
422 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
423 if (rt_rq
->rt_nr_running
) {
424 struct rt_prio_array
*array
;
426 WARN_ON(rt_se_prio(rt_se
) < rt_rq
->highest_prio
);
427 if (rt_se_prio(rt_se
) == rt_rq
->highest_prio
) {
429 array
= &rt_rq
->active
;
430 rt_rq
->highest_prio
=
431 sched_find_first_bit(array
->bitmap
);
432 } /* otherwise leave rq->highest prio alone */
434 rt_rq
->highest_prio
= MAX_RT_PRIO
;
437 if (rt_se
->nr_cpus_allowed
> 1) {
438 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
439 rq
->rt
.rt_nr_migratory
--;
442 update_rt_migration(rq_of_rt_rq(rt_rq
));
443 #endif /* CONFIG_SMP */
444 #ifdef CONFIG_RT_GROUP_SCHED
445 if (rt_se_boosted(rt_se
))
446 rt_rq
->rt_nr_boosted
--;
448 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
452 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
454 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
455 struct rt_prio_array
*array
= &rt_rq
->active
;
456 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
459 * Don't enqueue the group if its throttled, or when empty.
460 * The latter is a consequence of the former when a child group
461 * get throttled and the current group doesn't have any other
464 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
))
467 list_add_tail(&rt_se
->run_list
, array
->queue
+ rt_se_prio(rt_se
));
468 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
470 inc_rt_tasks(rt_se
, rt_rq
);
473 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
475 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
476 struct rt_prio_array
*array
= &rt_rq
->active
;
478 list_del_init(&rt_se
->run_list
);
479 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
480 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
482 dec_rt_tasks(rt_se
, rt_rq
);
486 * Because the prio of an upper entry depends on the lower
487 * entries, we must remove entries top - down.
489 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
)
491 struct sched_rt_entity
*back
= NULL
;
493 for_each_sched_rt_entity(rt_se
) {
498 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
500 __dequeue_rt_entity(rt_se
);
504 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
506 dequeue_rt_stack(rt_se
);
507 for_each_sched_rt_entity(rt_se
)
508 __enqueue_rt_entity(rt_se
);
511 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
513 dequeue_rt_stack(rt_se
);
515 for_each_sched_rt_entity(rt_se
) {
516 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
518 if (rt_rq
&& rt_rq
->rt_nr_running
)
519 __enqueue_rt_entity(rt_se
);
524 * Adding/removing a task to/from a priority array:
526 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
528 struct sched_rt_entity
*rt_se
= &p
->rt
;
533 enqueue_rt_entity(rt_se
);
536 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
538 struct sched_rt_entity
*rt_se
= &p
->rt
;
541 dequeue_rt_entity(rt_se
);
545 * Put task to the end of the run list without the overhead of dequeue
546 * followed by enqueue.
549 void requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
)
551 struct rt_prio_array
*array
= &rt_rq
->active
;
552 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
555 list_move_tail(&rt_se
->run_list
, queue
);
558 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
560 struct sched_rt_entity
*rt_se
= &p
->rt
;
563 for_each_sched_rt_entity(rt_se
) {
564 rt_rq
= rt_rq_of_se(rt_se
);
565 requeue_rt_entity(rt_rq
, rt_se
);
569 static void yield_task_rt(struct rq
*rq
)
571 requeue_task_rt(rq
, rq
->curr
);
575 static int find_lowest_rq(struct task_struct
*task
);
577 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
579 struct rq
*rq
= task_rq(p
);
582 * If the current task is an RT task, then
583 * try to see if we can wake this RT task up on another
584 * runqueue. Otherwise simply start this RT task
585 * on its current runqueue.
587 * We want to avoid overloading runqueues. Even if
588 * the RT task is of higher priority than the current RT task.
589 * RT tasks behave differently than other tasks. If
590 * one gets preempted, we try to push it off to another queue.
591 * So trying to keep a preempting RT task on the same
592 * cache hot CPU will force the running RT task to
593 * a cold CPU. So we waste all the cache for the lower
594 * RT task in hopes of saving some of a RT task
595 * that is just being woken and probably will have
598 if (unlikely(rt_task(rq
->curr
)) &&
599 (p
->rt
.nr_cpus_allowed
> 1)) {
600 int cpu
= find_lowest_rq(p
);
602 return (cpu
== -1) ? task_cpu(p
) : cpu
;
606 * Otherwise, just let it ride on the affined RQ and the
607 * post-schedule router will push the preempted task away
611 #endif /* CONFIG_SMP */
614 * Preempt the current task with a newly woken task if needed:
616 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
618 if (p
->prio
< rq
->curr
->prio
)
619 resched_task(rq
->curr
);
622 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
625 struct rt_prio_array
*array
= &rt_rq
->active
;
626 struct sched_rt_entity
*next
= NULL
;
627 struct list_head
*queue
;
630 idx
= sched_find_first_bit(array
->bitmap
);
631 BUG_ON(idx
>= MAX_RT_PRIO
);
633 queue
= array
->queue
+ idx
;
634 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
639 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
641 struct sched_rt_entity
*rt_se
;
642 struct task_struct
*p
;
647 if (unlikely(!rt_rq
->rt_nr_running
))
650 if (rt_rq_throttled(rt_rq
))
654 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
656 rt_rq
= group_rt_rq(rt_se
);
659 p
= rt_task_of(rt_se
);
660 p
->se
.exec_start
= rq
->clock
;
664 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
667 p
->se
.exec_start
= 0;
672 /* Only try algorithms three times */
673 #define RT_MAX_TRIES 3
675 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
676 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
678 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
680 if (!task_running(rq
, p
) &&
681 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
682 (p
->rt
.nr_cpus_allowed
> 1))
687 /* Return the second highest RT task, NULL otherwise */
688 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
690 struct task_struct
*next
= NULL
;
691 struct sched_rt_entity
*rt_se
;
692 struct rt_prio_array
*array
;
696 for_each_leaf_rt_rq(rt_rq
, rq
) {
697 array
= &rt_rq
->active
;
698 idx
= sched_find_first_bit(array
->bitmap
);
700 if (idx
>= MAX_RT_PRIO
)
702 if (next
&& next
->prio
< idx
)
704 list_for_each_entry(rt_se
, array
->queue
+ idx
, run_list
) {
705 struct task_struct
*p
= rt_task_of(rt_se
);
706 if (pick_rt_task(rq
, p
, cpu
)) {
712 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
720 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
722 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
724 int lowest_prio
= -1;
729 cpus_and(*lowest_mask
, task_rq(task
)->rd
->online
, task
->cpus_allowed
);
732 * Scan each rq for the lowest prio.
734 for_each_cpu_mask(cpu
, *lowest_mask
) {
735 struct rq
*rq
= cpu_rq(cpu
);
737 /* We look for lowest RT prio or non-rt CPU */
738 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
740 * if we already found a low RT queue
741 * and now we found this non-rt queue
742 * clear the mask and set our bit.
743 * Otherwise just return the queue as is
744 * and the count==1 will cause the algorithm
745 * to use the first bit found.
747 if (lowest_cpu
!= -1) {
748 cpus_clear(*lowest_mask
);
749 cpu_set(rq
->cpu
, *lowest_mask
);
754 /* no locking for now */
755 if ((rq
->rt
.highest_prio
> task
->prio
)
756 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
757 if (rq
->rt
.highest_prio
> lowest_prio
) {
758 /* new low - clear old data */
759 lowest_prio
= rq
->rt
.highest_prio
;
765 cpu_clear(cpu
, *lowest_mask
);
769 * Clear out all the set bits that represent
770 * runqueues that were of higher prio than
773 if (lowest_cpu
> 0) {
775 * Perhaps we could add another cpumask op to
776 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
777 * Then that could be optimized to use memset and such.
779 for_each_cpu_mask(cpu
, *lowest_mask
) {
780 if (cpu
>= lowest_cpu
)
782 cpu_clear(cpu
, *lowest_mask
);
789 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
793 /* "this_cpu" is cheaper to preempt than a remote processor */
794 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
797 first
= first_cpu(*mask
);
798 if (first
!= NR_CPUS
)
804 static int find_lowest_rq(struct task_struct
*task
)
806 struct sched_domain
*sd
;
807 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
808 int this_cpu
= smp_processor_id();
809 int cpu
= task_cpu(task
);
810 int count
= find_lowest_cpus(task
, lowest_mask
);
813 return -1; /* No targets found */
816 * There is no sense in performing an optimal search if only one
820 return first_cpu(*lowest_mask
);
823 * At this point we have built a mask of cpus representing the
824 * lowest priority tasks in the system. Now we want to elect
825 * the best one based on our affinity and topology.
827 * We prioritize the last cpu that the task executed on since
828 * it is most likely cache-hot in that location.
830 if (cpu_isset(cpu
, *lowest_mask
))
834 * Otherwise, we consult the sched_domains span maps to figure
835 * out which cpu is logically closest to our hot cache data.
838 this_cpu
= -1; /* Skip this_cpu opt if the same */
840 for_each_domain(cpu
, sd
) {
841 if (sd
->flags
& SD_WAKE_AFFINE
) {
842 cpumask_t domain_mask
;
845 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
847 best_cpu
= pick_optimal_cpu(this_cpu
,
855 * And finally, if there were no matches within the domains
856 * just give the caller *something* to work with from the compatible
859 return pick_optimal_cpu(this_cpu
, lowest_mask
);
862 /* Will lock the rq it finds */
863 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
865 struct rq
*lowest_rq
= NULL
;
869 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
870 cpu
= find_lowest_rq(task
);
872 if ((cpu
== -1) || (cpu
== rq
->cpu
))
875 lowest_rq
= cpu_rq(cpu
);
877 /* if the prio of this runqueue changed, try again */
878 if (double_lock_balance(rq
, lowest_rq
)) {
880 * We had to unlock the run queue. In
881 * the mean time, task could have
882 * migrated already or had its affinity changed.
883 * Also make sure that it wasn't scheduled on its rq.
885 if (unlikely(task_rq(task
) != rq
||
886 !cpu_isset(lowest_rq
->cpu
,
887 task
->cpus_allowed
) ||
888 task_running(rq
, task
) ||
891 spin_unlock(&lowest_rq
->lock
);
897 /* If this rq is still suitable use it. */
898 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
902 spin_unlock(&lowest_rq
->lock
);
910 * If the current CPU has more than one RT task, see if the non
911 * running task can migrate over to a CPU that is running a task
912 * of lesser priority.
914 static int push_rt_task(struct rq
*rq
)
916 struct task_struct
*next_task
;
917 struct rq
*lowest_rq
;
919 int paranoid
= RT_MAX_TRIES
;
921 if (!rq
->rt
.overloaded
)
924 next_task
= pick_next_highest_task_rt(rq
, -1);
929 if (unlikely(next_task
== rq
->curr
)) {
935 * It's possible that the next_task slipped in of
936 * higher priority than current. If that's the case
937 * just reschedule current.
939 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
940 resched_task(rq
->curr
);
944 /* We might release rq lock */
945 get_task_struct(next_task
);
947 /* find_lock_lowest_rq locks the rq if found */
948 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
950 struct task_struct
*task
;
952 * find lock_lowest_rq releases rq->lock
953 * so it is possible that next_task has changed.
954 * If it has, then try again.
956 task
= pick_next_highest_task_rt(rq
, -1);
957 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
958 put_task_struct(next_task
);
965 deactivate_task(rq
, next_task
, 0);
966 set_task_cpu(next_task
, lowest_rq
->cpu
);
967 activate_task(lowest_rq
, next_task
, 0);
969 resched_task(lowest_rq
->curr
);
971 spin_unlock(&lowest_rq
->lock
);
975 put_task_struct(next_task
);
981 * TODO: Currently we just use the second highest prio task on
982 * the queue, and stop when it can't migrate (or there's
983 * no more RT tasks). There may be a case where a lower
984 * priority RT task has a different affinity than the
985 * higher RT task. In this case the lower RT task could
986 * possibly be able to migrate where as the higher priority
987 * RT task could not. We currently ignore this issue.
988 * Enhancements are welcome!
990 static void push_rt_tasks(struct rq
*rq
)
992 /* push_rt_task will return true if it moved an RT */
993 while (push_rt_task(rq
))
997 static int pull_rt_task(struct rq
*this_rq
)
999 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
1000 struct task_struct
*p
, *next
;
1003 if (likely(!rt_overloaded(this_rq
)))
1006 next
= pick_next_task_rt(this_rq
);
1008 for_each_cpu_mask(cpu
, this_rq
->rd
->rto_mask
) {
1009 if (this_cpu
== cpu
)
1012 src_rq
= cpu_rq(cpu
);
1014 * We can potentially drop this_rq's lock in
1015 * double_lock_balance, and another CPU could
1016 * steal our next task - hence we must cause
1017 * the caller to recalculate the next task
1020 if (double_lock_balance(this_rq
, src_rq
)) {
1021 struct task_struct
*old_next
= next
;
1023 next
= pick_next_task_rt(this_rq
);
1024 if (next
!= old_next
)
1029 * Are there still pullable RT tasks?
1031 if (src_rq
->rt
.rt_nr_running
<= 1)
1034 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
1037 * Do we have an RT task that preempts
1038 * the to-be-scheduled task?
1040 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
1041 WARN_ON(p
== src_rq
->curr
);
1042 WARN_ON(!p
->se
.on_rq
);
1045 * There's a chance that p is higher in priority
1046 * than what's currently running on its cpu.
1047 * This is just that p is wakeing up and hasn't
1048 * had a chance to schedule. We only pull
1049 * p if it is lower in priority than the
1050 * current task on the run queue or
1051 * this_rq next task is lower in prio than
1052 * the current task on that rq.
1054 if (p
->prio
< src_rq
->curr
->prio
||
1055 (next
&& next
->prio
< src_rq
->curr
->prio
))
1060 deactivate_task(src_rq
, p
, 0);
1061 set_task_cpu(p
, this_cpu
);
1062 activate_task(this_rq
, p
, 0);
1064 * We continue with the search, just in
1065 * case there's an even higher prio task
1066 * in another runqueue. (low likelyhood
1069 * Update next so that we won't pick a task
1070 * on another cpu with a priority lower (or equal)
1071 * than the one we just picked.
1077 spin_unlock(&src_rq
->lock
);
1083 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
1085 /* Try to pull RT tasks here if we lower this rq's prio */
1086 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
1090 static void post_schedule_rt(struct rq
*rq
)
1093 * If we have more than one rt_task queued, then
1094 * see if we can push the other rt_tasks off to other CPUS.
1095 * Note we may release the rq lock, and since
1096 * the lock was owned by prev, we need to release it
1097 * first via finish_lock_switch and then reaquire it here.
1099 if (unlikely(rq
->rt
.overloaded
)) {
1100 spin_lock_irq(&rq
->lock
);
1102 spin_unlock_irq(&rq
->lock
);
1107 * If we are not running and we are not going to reschedule soon, we should
1108 * try to push tasks away now
1110 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
1112 if (!task_running(rq
, p
) &&
1113 !test_tsk_need_resched(rq
->curr
) &&
1118 static unsigned long
1119 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1120 unsigned long max_load_move
,
1121 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1122 int *all_pinned
, int *this_best_prio
)
1124 /* don't touch RT tasks */
1129 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1130 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1132 /* don't touch RT tasks */
1136 static void set_cpus_allowed_rt(struct task_struct
*p
,
1137 const cpumask_t
*new_mask
)
1139 int weight
= cpus_weight(*new_mask
);
1141 BUG_ON(!rt_task(p
));
1144 * Update the migration status of the RQ if we have an RT task
1145 * which is running AND changing its weight value.
1147 if (p
->se
.on_rq
&& (weight
!= p
->rt
.nr_cpus_allowed
)) {
1148 struct rq
*rq
= task_rq(p
);
1150 if ((p
->rt
.nr_cpus_allowed
<= 1) && (weight
> 1)) {
1151 rq
->rt
.rt_nr_migratory
++;
1152 } else if ((p
->rt
.nr_cpus_allowed
> 1) && (weight
<= 1)) {
1153 BUG_ON(!rq
->rt
.rt_nr_migratory
);
1154 rq
->rt
.rt_nr_migratory
--;
1157 update_rt_migration(rq
);
1160 p
->cpus_allowed
= *new_mask
;
1161 p
->rt
.nr_cpus_allowed
= weight
;
1164 /* Assumes rq->lock is held */
1165 static void join_domain_rt(struct rq
*rq
)
1167 if (rq
->rt
.overloaded
)
1168 rt_set_overload(rq
);
1171 /* Assumes rq->lock is held */
1172 static void leave_domain_rt(struct rq
*rq
)
1174 if (rq
->rt
.overloaded
)
1175 rt_clear_overload(rq
);
1179 * When switch from the rt queue, we bring ourselves to a position
1180 * that we might want to pull RT tasks from other runqueues.
1182 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
,
1186 * If there are other RT tasks then we will reschedule
1187 * and the scheduling of the other RT tasks will handle
1188 * the balancing. But if we are the last RT task
1189 * we may need to handle the pulling of RT tasks
1192 if (!rq
->rt
.rt_nr_running
)
1195 #endif /* CONFIG_SMP */
1198 * When switching a task to RT, we may overload the runqueue
1199 * with RT tasks. In this case we try to push them off to
1202 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
,
1205 int check_resched
= 1;
1208 * If we are already running, then there's nothing
1209 * that needs to be done. But if we are not running
1210 * we may need to preempt the current running task.
1211 * If that current running task is also an RT task
1212 * then see if we can move to another run queue.
1216 if (rq
->rt
.overloaded
&& push_rt_task(rq
) &&
1217 /* Don't resched if we changed runqueues */
1220 #endif /* CONFIG_SMP */
1221 if (check_resched
&& p
->prio
< rq
->curr
->prio
)
1222 resched_task(rq
->curr
);
1227 * Priority of the task has changed. This may cause
1228 * us to initiate a push or pull.
1230 static void prio_changed_rt(struct rq
*rq
, struct task_struct
*p
,
1231 int oldprio
, int running
)
1236 * If our priority decreases while running, we
1237 * may need to pull tasks to this runqueue.
1239 if (oldprio
< p
->prio
)
1242 * If there's a higher priority task waiting to run
1243 * then reschedule. Note, the above pull_rt_task
1244 * can release the rq lock and p could migrate.
1245 * Only reschedule if p is still on the same runqueue.
1247 if (p
->prio
> rq
->rt
.highest_prio
&& rq
->curr
== p
)
1250 /* For UP simply resched on drop of prio */
1251 if (oldprio
< p
->prio
)
1253 #endif /* CONFIG_SMP */
1256 * This task is not running, but if it is
1257 * greater than the current running task
1260 if (p
->prio
< rq
->curr
->prio
)
1261 resched_task(rq
->curr
);
1265 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
1267 unsigned long soft
, hard
;
1272 soft
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_cur
;
1273 hard
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_max
;
1275 if (soft
!= RLIM_INFINITY
) {
1279 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
1280 if (p
->rt
.timeout
> next
)
1281 p
->it_sched_expires
= p
->se
.sum_exec_runtime
;
1285 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
1292 * RR tasks need a special form of timeslice management.
1293 * FIFO tasks have no timeslices.
1295 if (p
->policy
!= SCHED_RR
)
1298 if (--p
->rt
.time_slice
)
1301 p
->rt
.time_slice
= DEF_TIMESLICE
;
1304 * Requeue to the end of queue if we are not the only element
1307 if (p
->rt
.run_list
.prev
!= p
->rt
.run_list
.next
) {
1308 requeue_task_rt(rq
, p
);
1309 set_tsk_need_resched(p
);
1313 static void set_curr_task_rt(struct rq
*rq
)
1315 struct task_struct
*p
= rq
->curr
;
1317 p
->se
.exec_start
= rq
->clock
;
1320 static const struct sched_class rt_sched_class
= {
1321 .next
= &fair_sched_class
,
1322 .enqueue_task
= enqueue_task_rt
,
1323 .dequeue_task
= dequeue_task_rt
,
1324 .yield_task
= yield_task_rt
,
1326 .select_task_rq
= select_task_rq_rt
,
1327 #endif /* CONFIG_SMP */
1329 .check_preempt_curr
= check_preempt_curr_rt
,
1331 .pick_next_task
= pick_next_task_rt
,
1332 .put_prev_task
= put_prev_task_rt
,
1335 .load_balance
= load_balance_rt
,
1336 .move_one_task
= move_one_task_rt
,
1337 .set_cpus_allowed
= set_cpus_allowed_rt
,
1338 .join_domain
= join_domain_rt
,
1339 .leave_domain
= leave_domain_rt
,
1340 .pre_schedule
= pre_schedule_rt
,
1341 .post_schedule
= post_schedule_rt
,
1342 .task_wake_up
= task_wake_up_rt
,
1343 .switched_from
= switched_from_rt
,
1346 .set_curr_task
= set_curr_task_rt
,
1347 .task_tick
= task_tick_rt
,
1349 .prio_changed
= prio_changed_rt
,
1350 .switched_to
= switched_to_rt
,