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 spin_lock(&rt_rq
->rt_runtime_lock
);
378 rt_rq
->rt_time
+= delta_exec
;
379 if (sched_rt_runtime_exceeded(rt_rq
))
381 spin_unlock(&rt_rq
->rt_runtime_lock
);
385 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
387 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
388 rt_rq
->rt_nr_running
++;
389 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
390 if (rt_se_prio(rt_se
) < rt_rq
->highest_prio
)
391 rt_rq
->highest_prio
= rt_se_prio(rt_se
);
394 if (rt_se
->nr_cpus_allowed
> 1) {
395 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
396 rq
->rt
.rt_nr_migratory
++;
399 update_rt_migration(rq_of_rt_rq(rt_rq
));
401 #ifdef CONFIG_RT_GROUP_SCHED
402 if (rt_se_boosted(rt_se
))
403 rt_rq
->rt_nr_boosted
++;
406 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
408 start_rt_bandwidth(&def_rt_bandwidth
);
413 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
415 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
416 WARN_ON(!rt_rq
->rt_nr_running
);
417 rt_rq
->rt_nr_running
--;
418 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
419 if (rt_rq
->rt_nr_running
) {
420 struct rt_prio_array
*array
;
422 WARN_ON(rt_se_prio(rt_se
) < rt_rq
->highest_prio
);
423 if (rt_se_prio(rt_se
) == rt_rq
->highest_prio
) {
425 array
= &rt_rq
->active
;
426 rt_rq
->highest_prio
=
427 sched_find_first_bit(array
->bitmap
);
428 } /* otherwise leave rq->highest prio alone */
430 rt_rq
->highest_prio
= MAX_RT_PRIO
;
433 if (rt_se
->nr_cpus_allowed
> 1) {
434 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
435 rq
->rt
.rt_nr_migratory
--;
438 update_rt_migration(rq_of_rt_rq(rt_rq
));
439 #endif /* CONFIG_SMP */
440 #ifdef CONFIG_RT_GROUP_SCHED
441 if (rt_se_boosted(rt_se
))
442 rt_rq
->rt_nr_boosted
--;
444 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
448 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
)
450 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
451 struct rt_prio_array
*array
= &rt_rq
->active
;
452 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
454 if (group_rq
&& rt_rq_throttled(group_rq
))
457 list_add_tail(&rt_se
->run_list
, array
->queue
+ rt_se_prio(rt_se
));
458 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
460 inc_rt_tasks(rt_se
, rt_rq
);
463 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
)
465 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
466 struct rt_prio_array
*array
= &rt_rq
->active
;
468 list_del_init(&rt_se
->run_list
);
469 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
470 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
472 dec_rt_tasks(rt_se
, rt_rq
);
476 * Because the prio of an upper entry depends on the lower
477 * entries, we must remove entries top - down.
479 * XXX: O(1/2 h^2) because we can only walk up, not down the chain.
480 * doesn't matter much for now, as h=2 for GROUP_SCHED.
482 static void dequeue_rt_stack(struct task_struct
*p
)
484 struct sched_rt_entity
*rt_se
, *top_se
;
487 * dequeue all, top - down.
492 for_each_sched_rt_entity(rt_se
) {
497 dequeue_rt_entity(top_se
);
502 * Adding/removing a task to/from a priority array:
504 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
506 struct sched_rt_entity
*rt_se
= &p
->rt
;
514 * enqueue everybody, bottom - up.
516 for_each_sched_rt_entity(rt_se
)
517 enqueue_rt_entity(rt_se
);
520 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
522 struct sched_rt_entity
*rt_se
= &p
->rt
;
530 * re-enqueue all non-empty rt_rq entities.
532 for_each_sched_rt_entity(rt_se
) {
533 rt_rq
= group_rt_rq(rt_se
);
534 if (rt_rq
&& rt_rq
->rt_nr_running
)
535 enqueue_rt_entity(rt_se
);
540 * Put task to the end of the run list without the overhead of dequeue
541 * followed by enqueue.
544 void requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
)
546 struct rt_prio_array
*array
= &rt_rq
->active
;
548 list_move_tail(&rt_se
->run_list
, array
->queue
+ rt_se_prio(rt_se
));
551 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
553 struct sched_rt_entity
*rt_se
= &p
->rt
;
556 for_each_sched_rt_entity(rt_se
) {
557 rt_rq
= rt_rq_of_se(rt_se
);
558 requeue_rt_entity(rt_rq
, rt_se
);
562 static void yield_task_rt(struct rq
*rq
)
564 requeue_task_rt(rq
, rq
->curr
);
568 static int find_lowest_rq(struct task_struct
*task
);
570 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
572 struct rq
*rq
= task_rq(p
);
575 * If the current task is an RT task, then
576 * try to see if we can wake this RT task up on another
577 * runqueue. Otherwise simply start this RT task
578 * on its current runqueue.
580 * We want to avoid overloading runqueues. Even if
581 * the RT task is of higher priority than the current RT task.
582 * RT tasks behave differently than other tasks. If
583 * one gets preempted, we try to push it off to another queue.
584 * So trying to keep a preempting RT task on the same
585 * cache hot CPU will force the running RT task to
586 * a cold CPU. So we waste all the cache for the lower
587 * RT task in hopes of saving some of a RT task
588 * that is just being woken and probably will have
591 if (unlikely(rt_task(rq
->curr
)) &&
592 (p
->rt
.nr_cpus_allowed
> 1)) {
593 int cpu
= find_lowest_rq(p
);
595 return (cpu
== -1) ? task_cpu(p
) : cpu
;
599 * Otherwise, just let it ride on the affined RQ and the
600 * post-schedule router will push the preempted task away
604 #endif /* CONFIG_SMP */
607 * Preempt the current task with a newly woken task if needed:
609 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
611 if (p
->prio
< rq
->curr
->prio
)
612 resched_task(rq
->curr
);
615 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
618 struct rt_prio_array
*array
= &rt_rq
->active
;
619 struct sched_rt_entity
*next
= NULL
;
620 struct list_head
*queue
;
623 idx
= sched_find_first_bit(array
->bitmap
);
624 BUG_ON(idx
>= MAX_RT_PRIO
);
626 queue
= array
->queue
+ idx
;
627 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
632 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
634 struct sched_rt_entity
*rt_se
;
635 struct task_struct
*p
;
640 if (unlikely(!rt_rq
->rt_nr_running
))
643 if (rt_rq_throttled(rt_rq
))
647 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
649 rt_rq
= group_rt_rq(rt_se
);
652 p
= rt_task_of(rt_se
);
653 p
->se
.exec_start
= rq
->clock
;
657 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
660 p
->se
.exec_start
= 0;
665 /* Only try algorithms three times */
666 #define RT_MAX_TRIES 3
668 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
669 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
671 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
673 if (!task_running(rq
, p
) &&
674 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
675 (p
->rt
.nr_cpus_allowed
> 1))
680 /* Return the second highest RT task, NULL otherwise */
681 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
683 struct task_struct
*next
= NULL
;
684 struct sched_rt_entity
*rt_se
;
685 struct rt_prio_array
*array
;
689 for_each_leaf_rt_rq(rt_rq
, rq
) {
690 array
= &rt_rq
->active
;
691 idx
= sched_find_first_bit(array
->bitmap
);
693 if (idx
>= MAX_RT_PRIO
)
695 if (next
&& next
->prio
< idx
)
697 list_for_each_entry(rt_se
, array
->queue
+ idx
, run_list
) {
698 struct task_struct
*p
= rt_task_of(rt_se
);
699 if (pick_rt_task(rq
, p
, cpu
)) {
705 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
713 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
715 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
717 int lowest_prio
= -1;
722 cpus_and(*lowest_mask
, task_rq(task
)->rd
->online
, task
->cpus_allowed
);
725 * Scan each rq for the lowest prio.
727 for_each_cpu_mask(cpu
, *lowest_mask
) {
728 struct rq
*rq
= cpu_rq(cpu
);
730 /* We look for lowest RT prio or non-rt CPU */
731 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
733 * if we already found a low RT queue
734 * and now we found this non-rt queue
735 * clear the mask and set our bit.
736 * Otherwise just return the queue as is
737 * and the count==1 will cause the algorithm
738 * to use the first bit found.
740 if (lowest_cpu
!= -1) {
741 cpus_clear(*lowest_mask
);
742 cpu_set(rq
->cpu
, *lowest_mask
);
747 /* no locking for now */
748 if ((rq
->rt
.highest_prio
> task
->prio
)
749 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
750 if (rq
->rt
.highest_prio
> lowest_prio
) {
751 /* new low - clear old data */
752 lowest_prio
= rq
->rt
.highest_prio
;
758 cpu_clear(cpu
, *lowest_mask
);
762 * Clear out all the set bits that represent
763 * runqueues that were of higher prio than
766 if (lowest_cpu
> 0) {
768 * Perhaps we could add another cpumask op to
769 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
770 * Then that could be optimized to use memset and such.
772 for_each_cpu_mask(cpu
, *lowest_mask
) {
773 if (cpu
>= lowest_cpu
)
775 cpu_clear(cpu
, *lowest_mask
);
782 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
786 /* "this_cpu" is cheaper to preempt than a remote processor */
787 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
790 first
= first_cpu(*mask
);
791 if (first
!= NR_CPUS
)
797 static int find_lowest_rq(struct task_struct
*task
)
799 struct sched_domain
*sd
;
800 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
801 int this_cpu
= smp_processor_id();
802 int cpu
= task_cpu(task
);
803 int count
= find_lowest_cpus(task
, lowest_mask
);
806 return -1; /* No targets found */
809 * There is no sense in performing an optimal search if only one
813 return first_cpu(*lowest_mask
);
816 * At this point we have built a mask of cpus representing the
817 * lowest priority tasks in the system. Now we want to elect
818 * the best one based on our affinity and topology.
820 * We prioritize the last cpu that the task executed on since
821 * it is most likely cache-hot in that location.
823 if (cpu_isset(cpu
, *lowest_mask
))
827 * Otherwise, we consult the sched_domains span maps to figure
828 * out which cpu is logically closest to our hot cache data.
831 this_cpu
= -1; /* Skip this_cpu opt if the same */
833 for_each_domain(cpu
, sd
) {
834 if (sd
->flags
& SD_WAKE_AFFINE
) {
835 cpumask_t domain_mask
;
838 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
840 best_cpu
= pick_optimal_cpu(this_cpu
,
848 * And finally, if there were no matches within the domains
849 * just give the caller *something* to work with from the compatible
852 return pick_optimal_cpu(this_cpu
, lowest_mask
);
855 /* Will lock the rq it finds */
856 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
858 struct rq
*lowest_rq
= NULL
;
862 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
863 cpu
= find_lowest_rq(task
);
865 if ((cpu
== -1) || (cpu
== rq
->cpu
))
868 lowest_rq
= cpu_rq(cpu
);
870 /* if the prio of this runqueue changed, try again */
871 if (double_lock_balance(rq
, lowest_rq
)) {
873 * We had to unlock the run queue. In
874 * the mean time, task could have
875 * migrated already or had its affinity changed.
876 * Also make sure that it wasn't scheduled on its rq.
878 if (unlikely(task_rq(task
) != rq
||
879 !cpu_isset(lowest_rq
->cpu
,
880 task
->cpus_allowed
) ||
881 task_running(rq
, task
) ||
884 spin_unlock(&lowest_rq
->lock
);
890 /* If this rq is still suitable use it. */
891 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
895 spin_unlock(&lowest_rq
->lock
);
903 * If the current CPU has more than one RT task, see if the non
904 * running task can migrate over to a CPU that is running a task
905 * of lesser priority.
907 static int push_rt_task(struct rq
*rq
)
909 struct task_struct
*next_task
;
910 struct rq
*lowest_rq
;
912 int paranoid
= RT_MAX_TRIES
;
914 if (!rq
->rt
.overloaded
)
917 next_task
= pick_next_highest_task_rt(rq
, -1);
922 if (unlikely(next_task
== rq
->curr
)) {
928 * It's possible that the next_task slipped in of
929 * higher priority than current. If that's the case
930 * just reschedule current.
932 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
933 resched_task(rq
->curr
);
937 /* We might release rq lock */
938 get_task_struct(next_task
);
940 /* find_lock_lowest_rq locks the rq if found */
941 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
943 struct task_struct
*task
;
945 * find lock_lowest_rq releases rq->lock
946 * so it is possible that next_task has changed.
947 * If it has, then try again.
949 task
= pick_next_highest_task_rt(rq
, -1);
950 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
951 put_task_struct(next_task
);
958 deactivate_task(rq
, next_task
, 0);
959 set_task_cpu(next_task
, lowest_rq
->cpu
);
960 activate_task(lowest_rq
, next_task
, 0);
962 resched_task(lowest_rq
->curr
);
964 spin_unlock(&lowest_rq
->lock
);
968 put_task_struct(next_task
);
974 * TODO: Currently we just use the second highest prio task on
975 * the queue, and stop when it can't migrate (or there's
976 * no more RT tasks). There may be a case where a lower
977 * priority RT task has a different affinity than the
978 * higher RT task. In this case the lower RT task could
979 * possibly be able to migrate where as the higher priority
980 * RT task could not. We currently ignore this issue.
981 * Enhancements are welcome!
983 static void push_rt_tasks(struct rq
*rq
)
985 /* push_rt_task will return true if it moved an RT */
986 while (push_rt_task(rq
))
990 static int pull_rt_task(struct rq
*this_rq
)
992 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
993 struct task_struct
*p
, *next
;
996 if (likely(!rt_overloaded(this_rq
)))
999 next
= pick_next_task_rt(this_rq
);
1001 for_each_cpu_mask(cpu
, this_rq
->rd
->rto_mask
) {
1002 if (this_cpu
== cpu
)
1005 src_rq
= cpu_rq(cpu
);
1007 * We can potentially drop this_rq's lock in
1008 * double_lock_balance, and another CPU could
1009 * steal our next task - hence we must cause
1010 * the caller to recalculate the next task
1013 if (double_lock_balance(this_rq
, src_rq
)) {
1014 struct task_struct
*old_next
= next
;
1016 next
= pick_next_task_rt(this_rq
);
1017 if (next
!= old_next
)
1022 * Are there still pullable RT tasks?
1024 if (src_rq
->rt
.rt_nr_running
<= 1)
1027 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
1030 * Do we have an RT task that preempts
1031 * the to-be-scheduled task?
1033 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
1034 WARN_ON(p
== src_rq
->curr
);
1035 WARN_ON(!p
->se
.on_rq
);
1038 * There's a chance that p is higher in priority
1039 * than what's currently running on its cpu.
1040 * This is just that p is wakeing up and hasn't
1041 * had a chance to schedule. We only pull
1042 * p if it is lower in priority than the
1043 * current task on the run queue or
1044 * this_rq next task is lower in prio than
1045 * the current task on that rq.
1047 if (p
->prio
< src_rq
->curr
->prio
||
1048 (next
&& next
->prio
< src_rq
->curr
->prio
))
1053 deactivate_task(src_rq
, p
, 0);
1054 set_task_cpu(p
, this_cpu
);
1055 activate_task(this_rq
, p
, 0);
1057 * We continue with the search, just in
1058 * case there's an even higher prio task
1059 * in another runqueue. (low likelyhood
1062 * Update next so that we won't pick a task
1063 * on another cpu with a priority lower (or equal)
1064 * than the one we just picked.
1070 spin_unlock(&src_rq
->lock
);
1076 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
1078 /* Try to pull RT tasks here if we lower this rq's prio */
1079 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
1083 static void post_schedule_rt(struct rq
*rq
)
1086 * If we have more than one rt_task queued, then
1087 * see if we can push the other rt_tasks off to other CPUS.
1088 * Note we may release the rq lock, and since
1089 * the lock was owned by prev, we need to release it
1090 * first via finish_lock_switch and then reaquire it here.
1092 if (unlikely(rq
->rt
.overloaded
)) {
1093 spin_lock_irq(&rq
->lock
);
1095 spin_unlock_irq(&rq
->lock
);
1100 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
1102 if (!task_running(rq
, p
) &&
1103 (p
->prio
>= rq
->rt
.highest_prio
) &&
1108 static unsigned long
1109 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1110 unsigned long max_load_move
,
1111 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1112 int *all_pinned
, int *this_best_prio
)
1114 /* don't touch RT tasks */
1119 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1120 struct sched_domain
*sd
, enum cpu_idle_type idle
)
1122 /* don't touch RT tasks */
1126 static void set_cpus_allowed_rt(struct task_struct
*p
, cpumask_t
*new_mask
)
1128 int weight
= cpus_weight(*new_mask
);
1130 BUG_ON(!rt_task(p
));
1133 * Update the migration status of the RQ if we have an RT task
1134 * which is running AND changing its weight value.
1136 if (p
->se
.on_rq
&& (weight
!= p
->rt
.nr_cpus_allowed
)) {
1137 struct rq
*rq
= task_rq(p
);
1139 if ((p
->rt
.nr_cpus_allowed
<= 1) && (weight
> 1)) {
1140 rq
->rt
.rt_nr_migratory
++;
1141 } else if ((p
->rt
.nr_cpus_allowed
> 1) && (weight
<= 1)) {
1142 BUG_ON(!rq
->rt
.rt_nr_migratory
);
1143 rq
->rt
.rt_nr_migratory
--;
1146 update_rt_migration(rq
);
1149 p
->cpus_allowed
= *new_mask
;
1150 p
->rt
.nr_cpus_allowed
= weight
;
1153 /* Assumes rq->lock is held */
1154 static void join_domain_rt(struct rq
*rq
)
1156 if (rq
->rt
.overloaded
)
1157 rt_set_overload(rq
);
1160 /* Assumes rq->lock is held */
1161 static void leave_domain_rt(struct rq
*rq
)
1163 if (rq
->rt
.overloaded
)
1164 rt_clear_overload(rq
);
1168 * When switch from the rt queue, we bring ourselves to a position
1169 * that we might want to pull RT tasks from other runqueues.
1171 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
,
1175 * If there are other RT tasks then we will reschedule
1176 * and the scheduling of the other RT tasks will handle
1177 * the balancing. But if we are the last RT task
1178 * we may need to handle the pulling of RT tasks
1181 if (!rq
->rt
.rt_nr_running
)
1184 #endif /* CONFIG_SMP */
1187 * When switching a task to RT, we may overload the runqueue
1188 * with RT tasks. In this case we try to push them off to
1191 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
,
1194 int check_resched
= 1;
1197 * If we are already running, then there's nothing
1198 * that needs to be done. But if we are not running
1199 * we may need to preempt the current running task.
1200 * If that current running task is also an RT task
1201 * then see if we can move to another run queue.
1205 if (rq
->rt
.overloaded
&& push_rt_task(rq
) &&
1206 /* Don't resched if we changed runqueues */
1209 #endif /* CONFIG_SMP */
1210 if (check_resched
&& p
->prio
< rq
->curr
->prio
)
1211 resched_task(rq
->curr
);
1216 * Priority of the task has changed. This may cause
1217 * us to initiate a push or pull.
1219 static void prio_changed_rt(struct rq
*rq
, struct task_struct
*p
,
1220 int oldprio
, int running
)
1225 * If our priority decreases while running, we
1226 * may need to pull tasks to this runqueue.
1228 if (oldprio
< p
->prio
)
1231 * If there's a higher priority task waiting to run
1232 * then reschedule. Note, the above pull_rt_task
1233 * can release the rq lock and p could migrate.
1234 * Only reschedule if p is still on the same runqueue.
1236 if (p
->prio
> rq
->rt
.highest_prio
&& rq
->curr
== p
)
1239 /* For UP simply resched on drop of prio */
1240 if (oldprio
< p
->prio
)
1242 #endif /* CONFIG_SMP */
1245 * This task is not running, but if it is
1246 * greater than the current running task
1249 if (p
->prio
< rq
->curr
->prio
)
1250 resched_task(rq
->curr
);
1254 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
1256 unsigned long soft
, hard
;
1261 soft
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_cur
;
1262 hard
= p
->signal
->rlim
[RLIMIT_RTTIME
].rlim_max
;
1264 if (soft
!= RLIM_INFINITY
) {
1268 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
1269 if (p
->rt
.timeout
> next
)
1270 p
->it_sched_expires
= p
->se
.sum_exec_runtime
;
1274 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
1281 * RR tasks need a special form of timeslice management.
1282 * FIFO tasks have no timeslices.
1284 if (p
->policy
!= SCHED_RR
)
1287 if (--p
->rt
.time_slice
)
1290 p
->rt
.time_slice
= DEF_TIMESLICE
;
1293 * Requeue to the end of queue if we are not the only element
1296 if (p
->rt
.run_list
.prev
!= p
->rt
.run_list
.next
) {
1297 requeue_task_rt(rq
, p
);
1298 set_tsk_need_resched(p
);
1302 static void set_curr_task_rt(struct rq
*rq
)
1304 struct task_struct
*p
= rq
->curr
;
1306 p
->se
.exec_start
= rq
->clock
;
1309 const struct sched_class rt_sched_class
= {
1310 .next
= &fair_sched_class
,
1311 .enqueue_task
= enqueue_task_rt
,
1312 .dequeue_task
= dequeue_task_rt
,
1313 .yield_task
= yield_task_rt
,
1315 .select_task_rq
= select_task_rq_rt
,
1316 #endif /* CONFIG_SMP */
1318 .check_preempt_curr
= check_preempt_curr_rt
,
1320 .pick_next_task
= pick_next_task_rt
,
1321 .put_prev_task
= put_prev_task_rt
,
1324 .load_balance
= load_balance_rt
,
1325 .move_one_task
= move_one_task_rt
,
1326 .set_cpus_allowed
= set_cpus_allowed_rt
,
1327 .join_domain
= join_domain_rt
,
1328 .leave_domain
= leave_domain_rt
,
1329 .pre_schedule
= pre_schedule_rt
,
1330 .post_schedule
= post_schedule_rt
,
1331 .task_wake_up
= task_wake_up_rt
,
1332 .switched_from
= switched_from_rt
,
1335 .set_curr_task
= set_curr_task_rt
,
1336 .task_tick
= task_tick_rt
,
1338 .prio_changed
= prio_changed_rt
,
1339 .switched_to
= switched_to_rt
,