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 task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
106 struct sched_rt_entity
*rt_se
= rt_rq
->rt_se
;
108 if (rt_rq
->rt_nr_running
) {
109 if (rt_se
&& !on_rt_rq(rt_se
))
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
)
202 if (rt_rq
->rt_nr_running
)
203 resched_task(rq_of_rt_rq(rt_rq
)->curr
);
206 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
210 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
212 return rt_rq
->rt_throttled
;
215 static inline cpumask_t
sched_rt_period_mask(void)
217 return cpu_online_map
;
221 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
223 return &cpu_rq(cpu
)->rt
;
226 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
228 return &def_rt_bandwidth
;
231 #endif /* CONFIG_RT_GROUP_SCHED */
235 * We ran out of runtime, see if we can borrow some from our neighbours.
237 static int do_balance_runtime(struct rt_rq
*rt_rq
)
239 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
240 struct root_domain
*rd
= cpu_rq(smp_processor_id())->rd
;
241 int i
, weight
, more
= 0;
244 weight
= cpus_weight(rd
->span
);
246 spin_lock(&rt_b
->rt_runtime_lock
);
247 rt_period
= ktime_to_ns(rt_b
->rt_period
);
248 for_each_cpu_mask_nr(i
, rd
->span
) {
249 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
255 spin_lock(&iter
->rt_runtime_lock
);
257 * Either all rqs have inf runtime and there's nothing to steal
258 * or __disable_runtime() below sets a specific rq to inf to
259 * indicate its been disabled and disalow stealing.
261 if (iter
->rt_runtime
== RUNTIME_INF
)
265 * From runqueues with spare time, take 1/n part of their
266 * spare time, but no more than our period.
268 diff
= iter
->rt_runtime
- iter
->rt_time
;
270 diff
= div_u64((u64
)diff
, weight
);
271 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
272 diff
= rt_period
- rt_rq
->rt_runtime
;
273 iter
->rt_runtime
-= diff
;
274 rt_rq
->rt_runtime
+= diff
;
276 if (rt_rq
->rt_runtime
== rt_period
) {
277 spin_unlock(&iter
->rt_runtime_lock
);
282 spin_unlock(&iter
->rt_runtime_lock
);
284 spin_unlock(&rt_b
->rt_runtime_lock
);
290 * Ensure this RQ takes back all the runtime it lend to its neighbours.
292 static void __disable_runtime(struct rq
*rq
)
294 struct root_domain
*rd
= rq
->rd
;
297 if (unlikely(!scheduler_running
))
300 for_each_leaf_rt_rq(rt_rq
, rq
) {
301 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
305 spin_lock(&rt_b
->rt_runtime_lock
);
306 spin_lock(&rt_rq
->rt_runtime_lock
);
308 * Either we're all inf and nobody needs to borrow, or we're
309 * already disabled and thus have nothing to do, or we have
310 * exactly the right amount of runtime to take out.
312 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
313 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
315 spin_unlock(&rt_rq
->rt_runtime_lock
);
318 * Calculate the difference between what we started out with
319 * and what we current have, that's the amount of runtime
320 * we lend and now have to reclaim.
322 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
325 * Greedy reclaim, take back as much as we can.
327 for_each_cpu_mask(i
, rd
->span
) {
328 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
332 * Can't reclaim from ourselves or disabled runqueues.
334 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
337 spin_lock(&iter
->rt_runtime_lock
);
339 diff
= min_t(s64
, iter
->rt_runtime
, want
);
340 iter
->rt_runtime
-= diff
;
343 iter
->rt_runtime
-= want
;
346 spin_unlock(&iter
->rt_runtime_lock
);
352 spin_lock(&rt_rq
->rt_runtime_lock
);
354 * We cannot be left wanting - that would mean some runtime
355 * leaked out of the system.
360 * Disable all the borrow logic by pretending we have inf
361 * runtime - in which case borrowing doesn't make sense.
363 rt_rq
->rt_runtime
= RUNTIME_INF
;
364 spin_unlock(&rt_rq
->rt_runtime_lock
);
365 spin_unlock(&rt_b
->rt_runtime_lock
);
369 static void disable_runtime(struct rq
*rq
)
373 spin_lock_irqsave(&rq
->lock
, flags
);
374 __disable_runtime(rq
);
375 spin_unlock_irqrestore(&rq
->lock
, flags
);
378 static void __enable_runtime(struct rq
*rq
)
382 if (unlikely(!scheduler_running
))
386 * Reset each runqueue's bandwidth settings
388 for_each_leaf_rt_rq(rt_rq
, rq
) {
389 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
391 spin_lock(&rt_b
->rt_runtime_lock
);
392 spin_lock(&rt_rq
->rt_runtime_lock
);
393 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
395 rt_rq
->rt_throttled
= 0;
396 spin_unlock(&rt_rq
->rt_runtime_lock
);
397 spin_unlock(&rt_b
->rt_runtime_lock
);
401 static void enable_runtime(struct rq
*rq
)
405 spin_lock_irqsave(&rq
->lock
, flags
);
406 __enable_runtime(rq
);
407 spin_unlock_irqrestore(&rq
->lock
, flags
);
410 static int balance_runtime(struct rt_rq
*rt_rq
)
414 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
415 spin_unlock(&rt_rq
->rt_runtime_lock
);
416 more
= do_balance_runtime(rt_rq
);
417 spin_lock(&rt_rq
->rt_runtime_lock
);
422 #else /* !CONFIG_SMP */
423 static inline int balance_runtime(struct rt_rq
*rt_rq
)
427 #endif /* CONFIG_SMP */
429 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
434 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
437 span
= sched_rt_period_mask();
438 for_each_cpu_mask(i
, span
) {
440 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
441 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
443 spin_lock(&rq
->lock
);
444 if (rt_rq
->rt_time
) {
447 spin_lock(&rt_rq
->rt_runtime_lock
);
448 if (rt_rq
->rt_throttled
)
449 balance_runtime(rt_rq
);
450 runtime
= rt_rq
->rt_runtime
;
451 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
452 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
453 rt_rq
->rt_throttled
= 0;
456 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
458 spin_unlock(&rt_rq
->rt_runtime_lock
);
459 } else if (rt_rq
->rt_nr_running
)
463 sched_rt_rq_enqueue(rt_rq
);
464 spin_unlock(&rq
->lock
);
470 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
472 #ifdef CONFIG_RT_GROUP_SCHED
473 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
476 return rt_rq
->highest_prio
;
479 return rt_task_of(rt_se
)->prio
;
482 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
484 u64 runtime
= sched_rt_runtime(rt_rq
);
486 if (rt_rq
->rt_throttled
)
487 return rt_rq_throttled(rt_rq
);
489 if (sched_rt_runtime(rt_rq
) >= sched_rt_period(rt_rq
))
492 balance_runtime(rt_rq
);
493 runtime
= sched_rt_runtime(rt_rq
);
494 if (runtime
== RUNTIME_INF
)
497 if (rt_rq
->rt_time
> runtime
) {
498 rt_rq
->rt_throttled
= 1;
499 if (rt_rq_throttled(rt_rq
)) {
500 sched_rt_rq_dequeue(rt_rq
);
509 * Update the current task's runtime statistics. Skip current tasks that
510 * are not in our scheduling class.
512 static void update_curr_rt(struct rq
*rq
)
514 struct task_struct
*curr
= rq
->curr
;
515 struct sched_rt_entity
*rt_se
= &curr
->rt
;
516 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
519 if (!task_has_rt_policy(curr
))
522 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
523 if (unlikely((s64
)delta_exec
< 0))
526 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
528 curr
->se
.sum_exec_runtime
+= delta_exec
;
529 account_group_exec_runtime(curr
, delta_exec
);
531 curr
->se
.exec_start
= rq
->clock
;
532 cpuacct_charge(curr
, delta_exec
);
534 if (!rt_bandwidth_enabled())
537 for_each_sched_rt_entity(rt_se
) {
538 rt_rq
= rt_rq_of_se(rt_se
);
540 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
541 spin_lock(&rt_rq
->rt_runtime_lock
);
542 rt_rq
->rt_time
+= delta_exec
;
543 if (sched_rt_runtime_exceeded(rt_rq
))
545 spin_unlock(&rt_rq
->rt_runtime_lock
);
551 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
553 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
554 rt_rq
->rt_nr_running
++;
555 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
556 if (rt_se_prio(rt_se
) < rt_rq
->highest_prio
) {
558 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
561 rt_rq
->highest_prio
= rt_se_prio(rt_se
);
564 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
,
570 if (rt_se
->nr_cpus_allowed
> 1) {
571 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
573 rq
->rt
.rt_nr_migratory
++;
576 update_rt_migration(rq_of_rt_rq(rt_rq
));
578 #ifdef CONFIG_RT_GROUP_SCHED
579 if (rt_se_boosted(rt_se
))
580 rt_rq
->rt_nr_boosted
++;
583 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
585 start_rt_bandwidth(&def_rt_bandwidth
);
590 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
593 int highest_prio
= rt_rq
->highest_prio
;
596 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
597 WARN_ON(!rt_rq
->rt_nr_running
);
598 rt_rq
->rt_nr_running
--;
599 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
600 if (rt_rq
->rt_nr_running
) {
601 struct rt_prio_array
*array
;
603 WARN_ON(rt_se_prio(rt_se
) < rt_rq
->highest_prio
);
604 if (rt_se_prio(rt_se
) == rt_rq
->highest_prio
) {
606 array
= &rt_rq
->active
;
607 rt_rq
->highest_prio
=
608 sched_find_first_bit(array
->bitmap
);
609 } /* otherwise leave rq->highest prio alone */
611 rt_rq
->highest_prio
= MAX_RT_PRIO
;
614 if (rt_se
->nr_cpus_allowed
> 1) {
615 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
616 rq
->rt
.rt_nr_migratory
--;
619 if (rt_rq
->highest_prio
!= highest_prio
) {
620 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
623 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
,
624 rt_rq
->highest_prio
);
627 update_rt_migration(rq_of_rt_rq(rt_rq
));
628 #endif /* CONFIG_SMP */
629 #ifdef CONFIG_RT_GROUP_SCHED
630 if (rt_se_boosted(rt_se
))
631 rt_rq
->rt_nr_boosted
--;
633 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
637 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
639 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
640 struct rt_prio_array
*array
= &rt_rq
->active
;
641 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
642 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
645 * Don't enqueue the group if its throttled, or when empty.
646 * The latter is a consequence of the former when a child group
647 * get throttled and the current group doesn't have any other
650 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
))
653 list_add_tail(&rt_se
->run_list
, queue
);
654 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
656 inc_rt_tasks(rt_se
, rt_rq
);
659 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
661 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
662 struct rt_prio_array
*array
= &rt_rq
->active
;
664 list_del_init(&rt_se
->run_list
);
665 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
666 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
668 dec_rt_tasks(rt_se
, rt_rq
);
672 * Because the prio of an upper entry depends on the lower
673 * entries, we must remove entries top - down.
675 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
)
677 struct sched_rt_entity
*back
= NULL
;
679 for_each_sched_rt_entity(rt_se
) {
684 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
686 __dequeue_rt_entity(rt_se
);
690 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
692 dequeue_rt_stack(rt_se
);
693 for_each_sched_rt_entity(rt_se
)
694 __enqueue_rt_entity(rt_se
);
697 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
699 dequeue_rt_stack(rt_se
);
701 for_each_sched_rt_entity(rt_se
) {
702 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
704 if (rt_rq
&& rt_rq
->rt_nr_running
)
705 __enqueue_rt_entity(rt_se
);
710 * Adding/removing a task to/from a priority array:
712 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
714 struct sched_rt_entity
*rt_se
= &p
->rt
;
719 enqueue_rt_entity(rt_se
);
721 inc_cpu_load(rq
, p
->se
.load
.weight
);
724 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
726 struct sched_rt_entity
*rt_se
= &p
->rt
;
729 dequeue_rt_entity(rt_se
);
731 dec_cpu_load(rq
, p
->se
.load
.weight
);
735 * Put task to the end of the run list without the overhead of dequeue
736 * followed by enqueue.
739 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
741 if (on_rt_rq(rt_se
)) {
742 struct rt_prio_array
*array
= &rt_rq
->active
;
743 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
746 list_move(&rt_se
->run_list
, queue
);
748 list_move_tail(&rt_se
->run_list
, queue
);
752 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
754 struct sched_rt_entity
*rt_se
= &p
->rt
;
757 for_each_sched_rt_entity(rt_se
) {
758 rt_rq
= rt_rq_of_se(rt_se
);
759 requeue_rt_entity(rt_rq
, rt_se
, head
);
763 static void yield_task_rt(struct rq
*rq
)
765 requeue_task_rt(rq
, rq
->curr
, 0);
769 static int find_lowest_rq(struct task_struct
*task
);
771 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
773 struct rq
*rq
= task_rq(p
);
776 * If the current task is an RT task, then
777 * try to see if we can wake this RT task up on another
778 * runqueue. Otherwise simply start this RT task
779 * on its current runqueue.
781 * We want to avoid overloading runqueues. Even if
782 * the RT task is of higher priority than the current RT task.
783 * RT tasks behave differently than other tasks. If
784 * one gets preempted, we try to push it off to another queue.
785 * So trying to keep a preempting RT task on the same
786 * cache hot CPU will force the running RT task to
787 * a cold CPU. So we waste all the cache for the lower
788 * RT task in hopes of saving some of a RT task
789 * that is just being woken and probably will have
792 if (unlikely(rt_task(rq
->curr
)) &&
793 (p
->rt
.nr_cpus_allowed
> 1)) {
794 int cpu
= find_lowest_rq(p
);
796 return (cpu
== -1) ? task_cpu(p
) : cpu
;
800 * Otherwise, just let it ride on the affined RQ and the
801 * post-schedule router will push the preempted task away
806 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
810 if (rq
->curr
->rt
.nr_cpus_allowed
== 1)
813 if (p
->rt
.nr_cpus_allowed
!= 1
814 && cpupri_find(&rq
->rd
->cpupri
, p
, &mask
))
817 if (!cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, &mask
))
821 * There appears to be other cpus that can accept
822 * current and none to run 'p', so lets reschedule
823 * to try and push current away:
825 requeue_task_rt(rq
, p
, 1);
826 resched_task(rq
->curr
);
829 #endif /* CONFIG_SMP */
832 * Preempt the current task with a newly woken task if needed:
834 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
, int sync
)
836 if (p
->prio
< rq
->curr
->prio
) {
837 resched_task(rq
->curr
);
845 * - the newly woken task is of equal priority to the current task
846 * - the newly woken task is non-migratable while current is migratable
847 * - current will be preempted on the next reschedule
849 * we should check to see if current can readily move to a different
850 * cpu. If so, we will reschedule to allow the push logic to try
851 * to move current somewhere else, making room for our non-migratable
854 if (p
->prio
== rq
->curr
->prio
&& !need_resched())
855 check_preempt_equal_prio(rq
, p
);
859 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
862 struct rt_prio_array
*array
= &rt_rq
->active
;
863 struct sched_rt_entity
*next
= NULL
;
864 struct list_head
*queue
;
867 idx
= sched_find_first_bit(array
->bitmap
);
868 BUG_ON(idx
>= MAX_RT_PRIO
);
870 queue
= array
->queue
+ idx
;
871 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
876 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
878 struct sched_rt_entity
*rt_se
;
879 struct task_struct
*p
;
884 if (unlikely(!rt_rq
->rt_nr_running
))
887 if (rt_rq_throttled(rt_rq
))
891 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
893 rt_rq
= group_rt_rq(rt_se
);
896 p
= rt_task_of(rt_se
);
897 p
->se
.exec_start
= rq
->clock
;
901 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
904 p
->se
.exec_start
= 0;
909 /* Only try algorithms three times */
910 #define RT_MAX_TRIES 3
912 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
913 static inline void double_unlock_balance(struct rq
*this_rq
,
916 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
918 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
920 if (!task_running(rq
, p
) &&
921 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
922 (p
->rt
.nr_cpus_allowed
> 1))
927 /* Return the second highest RT task, NULL otherwise */
928 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
930 struct task_struct
*next
= NULL
;
931 struct sched_rt_entity
*rt_se
;
932 struct rt_prio_array
*array
;
936 for_each_leaf_rt_rq(rt_rq
, rq
) {
937 array
= &rt_rq
->active
;
938 idx
= sched_find_first_bit(array
->bitmap
);
940 if (idx
>= MAX_RT_PRIO
)
942 if (next
&& next
->prio
< idx
)
944 list_for_each_entry(rt_se
, array
->queue
+ idx
, run_list
) {
945 struct task_struct
*p
= rt_task_of(rt_se
);
946 if (pick_rt_task(rq
, p
, cpu
)) {
952 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
960 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
962 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
966 /* "this_cpu" is cheaper to preempt than a remote processor */
967 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
970 first
= first_cpu(*mask
);
971 if (first
!= NR_CPUS
)
977 static int find_lowest_rq(struct task_struct
*task
)
979 struct sched_domain
*sd
;
980 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
981 int this_cpu
= smp_processor_id();
982 int cpu
= task_cpu(task
);
984 if (task
->rt
.nr_cpus_allowed
== 1)
985 return -1; /* No other targets possible */
987 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
988 return -1; /* No targets found */
991 * Only consider CPUs that are usable for migration.
992 * I guess we might want to change cpupri_find() to ignore those
993 * in the first place.
995 cpus_and(*lowest_mask
, *lowest_mask
, cpu_active_map
);
998 * At this point we have built a mask of cpus representing the
999 * lowest priority tasks in the system. Now we want to elect
1000 * the best one based on our affinity and topology.
1002 * We prioritize the last cpu that the task executed on since
1003 * it is most likely cache-hot in that location.
1005 if (cpu_isset(cpu
, *lowest_mask
))
1009 * Otherwise, we consult the sched_domains span maps to figure
1010 * out which cpu is logically closest to our hot cache data.
1012 if (this_cpu
== cpu
)
1013 this_cpu
= -1; /* Skip this_cpu opt if the same */
1015 for_each_domain(cpu
, sd
) {
1016 if (sd
->flags
& SD_WAKE_AFFINE
) {
1017 cpumask_t domain_mask
;
1020 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
1022 best_cpu
= pick_optimal_cpu(this_cpu
,
1030 * And finally, if there were no matches within the domains
1031 * just give the caller *something* to work with from the compatible
1034 return pick_optimal_cpu(this_cpu
, lowest_mask
);
1037 /* Will lock the rq it finds */
1038 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
1040 struct rq
*lowest_rq
= NULL
;
1044 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
1045 cpu
= find_lowest_rq(task
);
1047 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1050 lowest_rq
= cpu_rq(cpu
);
1052 /* if the prio of this runqueue changed, try again */
1053 if (double_lock_balance(rq
, lowest_rq
)) {
1055 * We had to unlock the run queue. In
1056 * the mean time, task could have
1057 * migrated already or had its affinity changed.
1058 * Also make sure that it wasn't scheduled on its rq.
1060 if (unlikely(task_rq(task
) != rq
||
1061 !cpu_isset(lowest_rq
->cpu
,
1062 task
->cpus_allowed
) ||
1063 task_running(rq
, task
) ||
1066 spin_unlock(&lowest_rq
->lock
);
1072 /* If this rq is still suitable use it. */
1073 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
1077 double_unlock_balance(rq
, lowest_rq
);
1085 * If the current CPU has more than one RT task, see if the non
1086 * running task can migrate over to a CPU that is running a task
1087 * of lesser priority.
1089 static int push_rt_task(struct rq
*rq
)
1091 struct task_struct
*next_task
;
1092 struct rq
*lowest_rq
;
1094 int paranoid
= RT_MAX_TRIES
;
1096 if (!rq
->rt
.overloaded
)
1099 next_task
= pick_next_highest_task_rt(rq
, -1);
1104 if (unlikely(next_task
== rq
->curr
)) {
1110 * It's possible that the next_task slipped in of
1111 * higher priority than current. If that's the case
1112 * just reschedule current.
1114 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1115 resched_task(rq
->curr
);
1119 /* We might release rq lock */
1120 get_task_struct(next_task
);
1122 /* find_lock_lowest_rq locks the rq if found */
1123 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1125 struct task_struct
*task
;
1127 * find lock_lowest_rq releases rq->lock
1128 * so it is possible that next_task has changed.
1129 * If it has, then try again.
1131 task
= pick_next_highest_task_rt(rq
, -1);
1132 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
1133 put_task_struct(next_task
);
1140 deactivate_task(rq
, next_task
, 0);
1141 set_task_cpu(next_task
, lowest_rq
->cpu
);
1142 activate_task(lowest_rq
, next_task
, 0);
1144 resched_task(lowest_rq
->curr
);
1146 double_unlock_balance(rq
, lowest_rq
);
1150 put_task_struct(next_task
);
1156 * TODO: Currently we just use the second highest prio task on
1157 * the queue, and stop when it can't migrate (or there's
1158 * no more RT tasks). There may be a case where a lower
1159 * priority RT task has a different affinity than the
1160 * higher RT task. In this case the lower RT task could
1161 * possibly be able to migrate where as the higher priority
1162 * RT task could not. We currently ignore this issue.
1163 * Enhancements are welcome!
1165 static void push_rt_tasks(struct rq
*rq
)
1167 /* push_rt_task will return true if it moved an RT */
1168 while (push_rt_task(rq
))
1172 static int pull_rt_task(struct rq
*this_rq
)
1174 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
1175 struct task_struct
*p
, *next
;
1178 if (likely(!rt_overloaded(this_rq
)))
1181 next
= pick_next_task_rt(this_rq
);
1183 for_each_cpu_mask_nr(cpu
, this_rq
->rd
->rto_mask
) {
1184 if (this_cpu
== cpu
)
1187 src_rq
= cpu_rq(cpu
);
1189 * We can potentially drop this_rq's lock in
1190 * double_lock_balance, and another CPU could
1191 * steal our next task - hence we must cause
1192 * the caller to recalculate the next task
1195 if (double_lock_balance(this_rq
, src_rq
)) {
1196 struct task_struct
*old_next
= next
;
1198 next
= pick_next_task_rt(this_rq
);
1199 if (next
!= old_next
)
1204 * Are there still pullable RT tasks?
1206 if (src_rq
->rt
.rt_nr_running
<= 1)
1209 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
1212 * Do we have an RT task that preempts
1213 * the to-be-scheduled task?
1215 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
1216 WARN_ON(p
== src_rq
->curr
);
1217 WARN_ON(!p
->se
.on_rq
);
1220 * There's a chance that p is higher in priority
1221 * than what's currently running on its cpu.
1222 * This is just that p is wakeing up and hasn't
1223 * had a chance to schedule. We only pull
1224 * p if it is lower in priority than the
1225 * current task on the run queue or
1226 * this_rq next task is lower in prio than
1227 * the current task on that rq.
1229 if (p
->prio
< src_rq
->curr
->prio
||
1230 (next
&& next
->prio
< src_rq
->curr
->prio
))
1235 deactivate_task(src_rq
, p
, 0);
1236 set_task_cpu(p
, this_cpu
);
1237 activate_task(this_rq
, p
, 0);
1239 * We continue with the search, just in
1240 * case there's an even higher prio task
1241 * in another runqueue. (low likelyhood
1244 * Update next so that we won't pick a task
1245 * on another cpu with a priority lower (or equal)
1246 * than the one we just picked.
1252 double_unlock_balance(this_rq
, src_rq
);
1258 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
1260 /* Try to pull RT tasks here if we lower this rq's prio */
1261 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
1265 static void post_schedule_rt(struct rq
*rq
)
1268 * If we have more than one rt_task queued, then
1269 * see if we can push the other rt_tasks off to other CPUS.
1270 * Note we may release the rq lock, and since
1271 * the lock was owned by prev, we need to release it
1272 * first via finish_lock_switch and then reaquire it here.
1274 if (unlikely(rq
->rt
.overloaded
)) {
1275 spin_lock_irq(&rq
->lock
);
1277 spin_unlock_irq(&rq
->lock
);
1282 * If we are not running and we are not going to reschedule soon, we should
1283 * try to push tasks away now
1285 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
1287 if (!task_running(rq
, p
) &&
1288 !test_tsk_need_resched(rq
->curr
) &&
1293 static unsigned long
1294 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1295 unsigned long max_load_move
,
1296 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1297 int *all_pinned
, int *this_best_prio
)
1299 /* don't touch RT tasks */
1304 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1305 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1307 /* don't touch RT tasks */
1311 static void set_cpus_allowed_rt(struct task_struct
*p
,
1312 const cpumask_t
*new_mask
)
1314 int weight
= cpus_weight(*new_mask
);
1316 BUG_ON(!rt_task(p
));
1319 * Update the migration status of the RQ if we have an RT task
1320 * which is running AND changing its weight value.
1322 if (p
->se
.on_rq
&& (weight
!= p
->rt
.nr_cpus_allowed
)) {
1323 struct rq
*rq
= task_rq(p
);
1325 if ((p
->rt
.nr_cpus_allowed
<= 1) && (weight
> 1)) {
1326 rq
->rt
.rt_nr_migratory
++;
1327 } else if ((p
->rt
.nr_cpus_allowed
> 1) && (weight
<= 1)) {
1328 BUG_ON(!rq
->rt
.rt_nr_migratory
);
1329 rq
->rt
.rt_nr_migratory
--;
1332 update_rt_migration(rq
);
1335 p
->cpus_allowed
= *new_mask
;
1336 p
->rt
.nr_cpus_allowed
= weight
;
1339 /* Assumes rq->lock is held */
1340 static void rq_online_rt(struct rq
*rq
)
1342 if (rq
->rt
.overloaded
)
1343 rt_set_overload(rq
);
1345 __enable_runtime(rq
);
1347 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
);
1350 /* Assumes rq->lock is held */
1351 static void rq_offline_rt(struct rq
*rq
)
1353 if (rq
->rt
.overloaded
)
1354 rt_clear_overload(rq
);
1356 __disable_runtime(rq
);
1358 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
1362 * When switch from the rt queue, we bring ourselves to a position
1363 * that we might want to pull RT tasks from other runqueues.
1365 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
,
1369 * If there are other RT tasks then we will reschedule
1370 * and the scheduling of the other RT tasks will handle
1371 * the balancing. But if we are the last RT task
1372 * we may need to handle the pulling of RT tasks
1375 if (!rq
->rt
.rt_nr_running
)
1378 #endif /* CONFIG_SMP */
1381 * When switching a task to RT, we may overload the runqueue
1382 * with RT tasks. In this case we try to push them off to
1385 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
,
1388 int check_resched
= 1;
1391 * If we are already running, then there's nothing
1392 * that needs to be done. But if we are not running
1393 * we may need to preempt the current running task.
1394 * If that current running task is also an RT task
1395 * then see if we can move to another run queue.
1399 if (rq
->rt
.overloaded
&& push_rt_task(rq
) &&
1400 /* Don't resched if we changed runqueues */
1403 #endif /* CONFIG_SMP */
1404 if (check_resched
&& p
->prio
< rq
->curr
->prio
)
1405 resched_task(rq
->curr
);
1410 * Priority of the task has changed. This may cause
1411 * us to initiate a push or pull.
1413 static void prio_changed_rt(struct rq
*rq
, struct task_struct
*p
,
1414 int oldprio
, int running
)
1419 * If our priority decreases while running, we
1420 * may need to pull tasks to this runqueue.
1422 if (oldprio
< p
->prio
)
1425 * If there's a higher priority task waiting to run
1426 * then reschedule. Note, the above pull_rt_task
1427 * can release the rq lock and p could migrate.
1428 * Only reschedule if p is still on the same runqueue.
1430 if (p
->prio
> rq
->rt
.highest_prio
&& rq
->curr
== p
)
1433 /* For UP simply resched on drop of prio */
1434 if (oldprio
< p
->prio
)
1436 #endif /* CONFIG_SMP */
1439 * This task is not running, but if it is
1440 * greater than the current running task
1443 if (p
->prio
< rq
->curr
->prio
)
1444 resched_task(rq
->curr
);
1448 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
1450 unsigned long soft
, hard
;
1455 soft
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_cur
;
1456 hard
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_max
;
1458 if (soft
!= RLIM_INFINITY
) {
1462 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
1463 if (p
->rt
.timeout
> next
)
1464 p
->cputime_expires
.sched_exp
= p
->se
.sum_exec_runtime
;
1468 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
1475 * RR tasks need a special form of timeslice management.
1476 * FIFO tasks have no timeslices.
1478 if (p
->policy
!= SCHED_RR
)
1481 if (--p
->rt
.time_slice
)
1484 p
->rt
.time_slice
= DEF_TIMESLICE
;
1487 * Requeue to the end of queue if we are not the only element
1490 if (p
->rt
.run_list
.prev
!= p
->rt
.run_list
.next
) {
1491 requeue_task_rt(rq
, p
, 0);
1492 set_tsk_need_resched(p
);
1496 static void set_curr_task_rt(struct rq
*rq
)
1498 struct task_struct
*p
= rq
->curr
;
1500 p
->se
.exec_start
= rq
->clock
;
1503 static const struct sched_class rt_sched_class
= {
1504 .next
= &fair_sched_class
,
1505 .enqueue_task
= enqueue_task_rt
,
1506 .dequeue_task
= dequeue_task_rt
,
1507 .yield_task
= yield_task_rt
,
1509 .check_preempt_curr
= check_preempt_curr_rt
,
1511 .pick_next_task
= pick_next_task_rt
,
1512 .put_prev_task
= put_prev_task_rt
,
1515 .select_task_rq
= select_task_rq_rt
,
1517 .load_balance
= load_balance_rt
,
1518 .move_one_task
= move_one_task_rt
,
1519 .set_cpus_allowed
= set_cpus_allowed_rt
,
1520 .rq_online
= rq_online_rt
,
1521 .rq_offline
= rq_offline_rt
,
1522 .pre_schedule
= pre_schedule_rt
,
1523 .post_schedule
= post_schedule_rt
,
1524 .task_wake_up
= task_wake_up_rt
,
1525 .switched_from
= switched_from_rt
,
1528 .set_curr_task
= set_curr_task_rt
,
1529 .task_tick
= task_tick_rt
,
1531 .prio_changed
= prio_changed_rt
,
1532 .switched_to
= switched_to_rt
,
1535 #ifdef CONFIG_SCHED_DEBUG
1536 extern void print_rt_rq(struct seq_file
*m
, int cpu
, struct rt_rq
*rt_rq
);
1538 static void print_rt_stats(struct seq_file
*m
, int cpu
)
1540 struct rt_rq
*rt_rq
;
1543 for_each_leaf_rt_rq(rt_rq
, cpu_rq(cpu
))
1544 print_rt_rq(m
, cpu
, rt_rq
);
1547 #endif /* CONFIG_SCHED_DEBUG */