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);
458 if (group_rq && rt_rq_throttled(group_rq))
461 if (rt_se->nr_cpus_allowed == 1)
462 list_add_tail(&rt_se->run_list,
463 array->xqueue + rt_se_prio(rt_se));
465 list_add_tail(&rt_se->run_list,
466 array->squeue + 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->squeue + rt_se_prio(rt_se))
480 && list_empty(array->xqueue + rt_se_prio(rt_se)))
481 __clear_bit(rt_se_prio(rt_se), array->bitmap);
483 dec_rt_tasks(rt_se, rt_rq);
487 * Because the prio of an upper entry depends on the lower
488 * entries, we must remove entries top - down.
490 static void dequeue_rt_stack(struct task_struct *p)
492 struct sched_rt_entity *rt_se, *back = NULL;
495 for_each_sched_rt_entity(rt_se) {
500 for (rt_se = back; rt_se; rt_se = rt_se->back) {
502 dequeue_rt_entity(rt_se);
507 * Adding/removing a task to/from a priority array:
509 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
511 struct sched_rt_entity *rt_se = &p->rt;
519 * enqueue everybody, bottom - up.
521 for_each_sched_rt_entity(rt_se)
522 enqueue_rt_entity(rt_se);
525 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
527 struct sched_rt_entity *rt_se = &p->rt;
535 * re-enqueue all non-empty rt_rq entities.
537 for_each_sched_rt_entity(rt_se) {
538 rt_rq = group_rt_rq(rt_se);
539 if (rt_rq && rt_rq->rt_nr_running)
540 enqueue_rt_entity(rt_se);
545 * Put task to the end of the run list without the overhead of dequeue
546 * followed by enqueue.
548 * Note: We always enqueue the task to the shared-queue, regardless of its
549 * previous position w.r.t. exclusive vs shared. This is so that exclusive RR
550 * tasks fairly round-robin with all tasks on the runqueue, not just other
554 void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
556 struct rt_prio_array *array = &rt_rq->active;
558 list_del_init(&rt_se->run_list);
559 list_add_tail(&rt_se->run_list, array->squeue + rt_se_prio(rt_se));
562 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
564 struct sched_rt_entity *rt_se = &p->rt;
567 for_each_sched_rt_entity(rt_se) {
568 rt_rq = rt_rq_of_se(rt_se);
569 requeue_rt_entity(rt_rq, rt_se);
573 static void yield_task_rt(struct rq *rq)
575 requeue_task_rt(rq, rq->curr);
579 static int find_lowest_rq(struct task_struct *task);
581 static int select_task_rq_rt(struct task_struct *p, int sync)
583 struct rq *rq = task_rq(p);
586 * If the current task is an RT task, then
587 * try to see if we can wake this RT task up on another
588 * runqueue. Otherwise simply start this RT task
589 * on its current runqueue.
591 * We want to avoid overloading runqueues. Even if
592 * the RT task is of higher priority than the current RT task.
593 * RT tasks behave differently than other tasks. If
594 * one gets preempted, we try to push it off to another queue.
595 * So trying to keep a preempting RT task on the same
596 * cache hot CPU will force the running RT task to
597 * a cold CPU. So we waste all the cache for the lower
598 * RT task in hopes of saving some of a RT task
599 * that is just being woken and probably will have
602 if (unlikely(rt_task(rq->curr)) &&
603 (p->rt.nr_cpus_allowed > 1)) {
604 int cpu = find_lowest_rq(p);
606 return (cpu == -1) ? task_cpu(p) : cpu;
610 * Otherwise, just let it ride on the affined RQ and the
611 * post-schedule router will push the preempted task away
615 #endif /* CONFIG_SMP */
617 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
618 struct rt_rq *rt_rq);
621 * Preempt the current task with a newly woken task if needed:
623 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
625 if (p->prio < rq->curr->prio) {
626 resched_task(rq->curr);
634 * - the newly woken task is of equal priority to the current task
635 * - the newly woken task is non-migratable while current is migratable
636 * - current will be preempted on the next reschedule
638 * we should check to see if current can readily move to a different
639 * cpu. If so, we will reschedule to allow the push logic to try
640 * to move current somewhere else, making room for our non-migratable
643 if((p->prio == rq->curr->prio)
644 && p->rt.nr_cpus_allowed == 1
645 && rq->curr->rt.nr_cpus_allowed != 1
646 && pick_next_rt_entity(rq, &rq->rt) != &rq->curr->rt) {
649 if (cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
651 * There appears to be other cpus that can accept
652 * current, so lets reschedule to try and push it away
654 resched_task(rq->curr);
659 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
662 struct rt_prio_array *array = &rt_rq->active;
663 struct sched_rt_entity *next = NULL;
664 struct list_head *queue;
667 idx = sched_find_first_bit(array->bitmap);
668 BUG_ON(idx >= MAX_RT_PRIO);
670 queue = array->xqueue + idx;
671 if (!list_empty(queue))
672 next = list_entry(queue->next, struct sched_rt_entity,
675 queue = array->squeue + idx;
676 next = list_entry(queue->next, struct sched_rt_entity,
683 static struct task_struct *pick_next_task_rt(struct rq *rq)
685 struct sched_rt_entity *rt_se;
686 struct task_struct *p;
691 if (unlikely(!rt_rq->rt_nr_running))
694 if (rt_rq_throttled(rt_rq))
698 rt_se = pick_next_rt_entity(rq, rt_rq);
700 rt_rq = group_rt_rq(rt_se);
703 p = rt_task_of(rt_se);
704 p->se.exec_start = rq->clock;
708 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
711 p->se.exec_start = 0;
716 /* Only try algorithms three times */
717 #define RT_MAX_TRIES 3
719 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
720 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
722 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
724 if (!task_running(rq, p) &&
725 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
726 (p->rt.nr_cpus_allowed > 1))
731 /* Return the second highest RT task, NULL otherwise */
732 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
734 struct task_struct *next = NULL;
735 struct sched_rt_entity *rt_se;
736 struct rt_prio_array *array;
740 for_each_leaf_rt_rq(rt_rq, rq) {
741 array = &rt_rq->active;
742 idx = sched_find_first_bit(array->bitmap);
744 if (idx >= MAX_RT_PRIO)
746 if (next && next->prio < idx)
748 list_for_each_entry(rt_se, array->squeue + idx, run_list) {
749 struct task_struct *p = rt_task_of(rt_se);
750 if (pick_rt_task(rq, p, cpu)) {
756 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
764 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
766 static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
768 int lowest_prio = -1;
773 cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed);
776 * Scan each rq for the lowest prio.
778 for_each_cpu_mask(cpu, *lowest_mask) {
779 struct rq *rq = cpu_rq(cpu);
781 /* We look for lowest RT prio or non-rt CPU */
782 if (rq->rt.highest_prio >= MAX_RT_PRIO) {
784 * if we already found a low RT queue
785 * and now we found this non-rt queue
786 * clear the mask and set our bit.
787 * Otherwise just return the queue as is
788 * and the count==1 will cause the algorithm
789 * to use the first bit found.
791 if (lowest_cpu != -1) {
792 cpus_clear(*lowest_mask);
793 cpu_set(rq->cpu, *lowest_mask);
798 /* no locking for now */
799 if ((rq->rt.highest_prio > task->prio)
800 && (rq->rt.highest_prio >= lowest_prio)) {
801 if (rq->rt.highest_prio > lowest_prio) {
802 /* new low - clear old data */
803 lowest_prio = rq->rt.highest_prio;
809 cpu_clear(cpu, *lowest_mask);
813 * Clear out all the set bits that represent
814 * runqueues that were of higher prio than
817 if (lowest_cpu > 0) {
819 * Perhaps we could add another cpumask op to
820 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
821 * Then that could be optimized to use memset and such.
823 for_each_cpu_mask(cpu, *lowest_mask) {
824 if (cpu >= lowest_cpu)
826 cpu_clear(cpu, *lowest_mask);
833 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
837 /* "this_cpu" is cheaper to preempt than a remote processor */
838 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
841 first = first_cpu(*mask);
842 if (first != NR_CPUS)
848 static int find_lowest_rq(struct task_struct *task)
850 struct sched_domain *sd;
851 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
852 int this_cpu = smp_processor_id();
853 int cpu = task_cpu(task);
854 int count = find_lowest_cpus(task, lowest_mask);
857 return -1; /* No targets found */
860 * There is no sense in performing an optimal search if only one
864 return first_cpu(*lowest_mask);
867 * At this point we have built a mask of cpus representing the
868 * lowest priority tasks in the system. Now we want to elect
869 * the best one based on our affinity and topology.
871 * We prioritize the last cpu that the task executed on since
872 * it is most likely cache-hot in that location.
874 if (cpu_isset(cpu, *lowest_mask))
878 * Otherwise, we consult the sched_domains span maps to figure
879 * out which cpu is logically closest to our hot cache data.
882 this_cpu = -1; /* Skip this_cpu opt if the same */
884 for_each_domain(cpu, sd) {
885 if (sd->flags & SD_WAKE_AFFINE) {
886 cpumask_t domain_mask;
889 cpus_and(domain_mask, sd->span, *lowest_mask);
891 best_cpu = pick_optimal_cpu(this_cpu,
899 * And finally, if there were no matches within the domains
900 * just give the caller *something* to work with from the compatible
903 return pick_optimal_cpu(this_cpu, lowest_mask);
906 /* Will lock the rq it finds */
907 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
909 struct rq *lowest_rq = NULL;
913 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
914 cpu = find_lowest_rq(task);
916 if ((cpu == -1) || (cpu == rq->cpu))
919 lowest_rq = cpu_rq(cpu);
921 /* if the prio of this runqueue changed, try again */
922 if (double_lock_balance(rq, lowest_rq)) {
924 * We had to unlock the run queue. In
925 * the mean time, task could have
926 * migrated already or had its affinity changed.
927 * Also make sure that it wasn't scheduled on its rq.
929 if (unlikely(task_rq(task) != rq ||
930 !cpu_isset(lowest_rq->cpu,
931 task->cpus_allowed) ||
932 task_running(rq, task) ||
935 spin_unlock(&lowest_rq->lock);
941 /* If this rq is still suitable use it. */
942 if (lowest_rq->rt.highest_prio > task->prio)
946 spin_unlock(&lowest_rq->lock);
954 * If the current CPU has more than one RT task, see if the non
955 * running task can migrate over to a CPU that is running a task
956 * of lesser priority.
958 static int push_rt_task(struct rq *rq)
960 struct task_struct *next_task;
961 struct rq *lowest_rq;
963 int paranoid = RT_MAX_TRIES;
965 if (!rq->rt.overloaded)
968 next_task = pick_next_highest_task_rt(rq, -1);
973 if (unlikely(next_task == rq->curr)) {
979 * It's possible that the next_task slipped in of
980 * higher priority than current. If that's the case
981 * just reschedule current.
983 if (unlikely(next_task->prio < rq->curr->prio)) {
984 resched_task(rq->curr);
988 /* We might release rq lock */
989 get_task_struct(next_task);
991 /* find_lock_lowest_rq locks the rq if found */
992 lowest_rq = find_lock_lowest_rq(next_task, rq);
994 struct task_struct *task;
996 * find lock_lowest_rq releases rq->lock
997 * so it is possible that next_task has changed.
998 * If it has, then try again.
1000 task = pick_next_highest_task_rt(rq, -1);
1001 if (unlikely(task != next_task) && task && paranoid--) {
1002 put_task_struct(next_task);
1009 deactivate_task(rq, next_task, 0);
1010 set_task_cpu(next_task, lowest_rq->cpu);
1011 activate_task(lowest_rq, next_task, 0);
1013 resched_task(lowest_rq->curr);
1015 spin_unlock(&lowest_rq->lock);
1019 put_task_struct(next_task);
1025 * TODO: Currently we just use the second highest prio task on
1026 * the queue, and stop when it can't migrate (or there's
1027 * no more RT tasks). There may be a case where a lower
1028 * priority RT task has a different affinity than the
1029 * higher RT task. In this case the lower RT task could
1030 * possibly be able to migrate where as the higher priority
1031 * RT task could not. We currently ignore this issue.
1032 * Enhancements are welcome!
1034 static void push_rt_tasks(struct rq *rq)
1036 /* push_rt_task will return true if it moved an RT */
1037 while (push_rt_task(rq))
1041 static int pull_rt_task(struct rq *this_rq)
1043 int this_cpu = this_rq->cpu, ret = 0, cpu;
1044 struct task_struct *p, *next;
1047 if (likely(!rt_overloaded(this_rq)))
1050 next = pick_next_task_rt(this_rq);
1052 for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
1053 if (this_cpu == cpu)
1056 src_rq = cpu_rq(cpu);
1058 * We can potentially drop this_rq's lock in
1059 * double_lock_balance, and another CPU could
1060 * steal our next task - hence we must cause
1061 * the caller to recalculate the next task
1064 if (double_lock_balance(this_rq, src_rq)) {
1065 struct task_struct *old_next = next;
1067 next = pick_next_task_rt(this_rq);
1068 if (next != old_next)
1073 * Are there still pullable RT tasks?
1075 if (src_rq->rt.rt_nr_running <= 1)
1078 p = pick_next_highest_task_rt(src_rq, this_cpu);
1081 * Do we have an RT task that preempts
1082 * the to-be-scheduled task?
1084 if (p && (!next || (p->prio < next->prio))) {
1085 WARN_ON(p == src_rq->curr);
1086 WARN_ON(!p->se.on_rq);
1089 * There's a chance that p is higher in priority
1090 * than what's currently running on its cpu.
1091 * This is just that p is wakeing up and hasn't
1092 * had a chance to schedule. We only pull
1093 * p if it is lower in priority than the
1094 * current task on the run queue or
1095 * this_rq next task is lower in prio than
1096 * the current task on that rq.
1098 if (p->prio < src_rq->curr->prio ||
1099 (next && next->prio < src_rq->curr->prio))
1104 deactivate_task(src_rq, p, 0);
1105 set_task_cpu(p, this_cpu);
1106 activate_task(this_rq, p, 0);
1108 * We continue with the search, just in
1109 * case there's an even higher prio task
1110 * in another runqueue. (low likelyhood
1113 * Update next so that we won't pick a task
1114 * on another cpu with a priority lower (or equal)
1115 * than the one we just picked.
1121 spin_unlock(&src_rq->lock);
1127 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1129 /* Try to pull RT tasks here if we lower this rq's prio */
1130 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1134 static void post_schedule_rt(struct rq *rq)
1137 * If we have more than one rt_task queued, then
1138 * see if we can push the other rt_tasks off to other CPUS.
1139 * Note we may release the rq lock, and since
1140 * the lock was owned by prev, we need to release it
1141 * first via finish_lock_switch and then reaquire it here.
1143 if (unlikely(rq->rt.overloaded)) {
1144 spin_lock_irq(&rq->lock);
1146 spin_unlock_irq(&rq->lock);
1151 * If we are not running and we are not going to reschedule soon, we should
1152 * try to push tasks away now
1154 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1156 if (!task_running(rq, p) &&
1157 !test_tsk_need_resched(rq->curr) &&
1162 static unsigned long
1163 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1164 unsigned long max_load_move,
1165 struct sched_domain *sd, enum cpu_idle_type idle,
1166 int *all_pinned, int *this_best_prio)
1168 /* don't touch RT tasks */
1173 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1174 struct sched_domain *sd, enum cpu_idle_type idle)
1176 /* don't touch RT tasks */
1180 static void set_cpus_allowed_rt(struct task_struct *p,
1181 const cpumask_t *new_mask)
1183 int weight = cpus_weight(*new_mask);
1185 BUG_ON(!rt_task(p));
1188 * Update the migration status of the RQ if we have an RT task
1189 * which is running AND changing its weight value.
1191 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1192 struct rq *rq = task_rq(p);
1194 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1195 rq->rt.rt_nr_migratory++;
1196 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1197 BUG_ON(!rq->rt.rt_nr_migratory);
1198 rq->rt.rt_nr_migratory--;
1201 update_rt_migration(rq);
1203 if (unlikely(weight == 1 || p->rt.nr_cpus_allowed == 1))
1205 * If either the new or old weight is a "1", we need
1206 * to requeue to properly move between shared and
1209 requeue_task_rt(rq, p);
1212 p->cpus_allowed = *new_mask;
1213 p->rt.nr_cpus_allowed = weight;
1216 /* Assumes rq->lock is held */
1217 static void join_domain_rt(struct rq *rq)
1219 if (rq->rt.overloaded)
1220 rt_set_overload(rq);
1223 /* Assumes rq->lock is held */
1224 static void leave_domain_rt(struct rq *rq)
1226 if (rq->rt.overloaded)
1227 rt_clear_overload(rq);
1231 * When switch from the rt queue, we bring ourselves to a position
1232 * that we might want to pull RT tasks from other runqueues.
1234 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1238 * If there are other RT tasks then we will reschedule
1239 * and the scheduling of the other RT tasks will handle
1240 * the balancing. But if we are the last RT task
1241 * we may need to handle the pulling of RT tasks
1244 if (!rq->rt.rt_nr_running)
1247 #endif /* CONFIG_SMP */
1250 * When switching a task to RT, we may overload the runqueue
1251 * with RT tasks. In this case we try to push them off to
1254 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1257 int check_resched = 1;
1260 * If we are already running, then there's nothing
1261 * that needs to be done. But if we are not running
1262 * we may need to preempt the current running task.
1263 * If that current running task is also an RT task
1264 * then see if we can move to another run queue.
1268 if (rq->rt.overloaded && push_rt_task(rq) &&
1269 /* Don't resched if we changed runqueues */
1272 #endif /* CONFIG_SMP */
1273 if (check_resched && p->prio < rq->curr->prio)
1274 resched_task(rq->curr);
1279 * Priority of the task has changed. This may cause
1280 * us to initiate a push or pull.
1282 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1283 int oldprio, int running)
1288 * If our priority decreases while running, we
1289 * may need to pull tasks to this runqueue.
1291 if (oldprio < p->prio)
1294 * If there's a higher priority task waiting to run
1295 * then reschedule. Note, the above pull_rt_task
1296 * can release the rq lock and p could migrate.
1297 * Only reschedule if p is still on the same runqueue.
1299 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1302 /* For UP simply resched on drop of prio */
1303 if (oldprio < p->prio)
1305 #endif /* CONFIG_SMP */
1308 * This task is not running, but if it is
1309 * greater than the current running task
1312 if (p->prio < rq->curr->prio)
1313 resched_task(rq->curr);
1317 static void watchdog(struct rq *rq, struct task_struct *p)
1319 unsigned long soft, hard;
1324 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1325 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1327 if (soft != RLIM_INFINITY) {
1331 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1332 if (p->rt.timeout > next)
1333 p->it_sched_expires = p->se.sum_exec_runtime;
1337 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1344 * RR tasks need a special form of timeslice management.
1345 * FIFO tasks have no timeslices.
1347 if (p->policy != SCHED_RR)
1350 if (--p->rt.time_slice)
1353 p->rt.time_slice = DEF_TIMESLICE;
1356 * Requeue to the end of queue if we are not the only element
1359 if (p->rt.run_list.prev != p->rt.run_list.next) {
1360 requeue_task_rt(rq, p);
1361 set_tsk_need_resched(p);
1365 static void set_curr_task_rt(struct rq *rq)
1367 struct task_struct *p = rq->curr;
1369 p->se.exec_start = rq->clock;
1372 static const struct sched_class rt_sched_class = {
1373 .next = &fair_sched_class,
1374 .enqueue_task = enqueue_task_rt,
1375 .dequeue_task = dequeue_task_rt,
1376 .yield_task = yield_task_rt,
1378 .select_task_rq = select_task_rq_rt,
1379 #endif /* CONFIG_SMP */
1381 .check_preempt_curr = check_preempt_curr_rt,
1383 .pick_next_task = pick_next_task_rt,
1384 .put_prev_task = put_prev_task_rt,
1387 .load_balance = load_balance_rt,
1388 .move_one_task = move_one_task_rt,
1389 .set_cpus_allowed = set_cpus_allowed_rt,
1390 .join_domain = join_domain_rt,
1391 .leave_domain = leave_domain_rt,
1392 .pre_schedule = pre_schedule_rt,
1393 .post_schedule = post_schedule_rt,
1394 .task_wake_up = task_wake_up_rt,
1395 .switched_from = switched_from_rt,
1398 .set_curr_task = set_curr_task_rt,
1399 .task_tick = task_tick_rt,
1401 .prio_changed = prio_changed_rt,
1402 .switched_to = switched_to_rt,