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 sched_rt_entity *rt_se = rt_rq->rt_se;
107 if (rt_se && !on_rt_rq(rt_se) && rt_rq->rt_nr_running) {
108 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
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 */
234 static int do_balance_runtime(struct rt_rq *rt_rq)
236 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
237 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
238 int i, weight, more = 0;
241 weight = cpus_weight(rd->span);
243 spin_lock(&rt_b->rt_runtime_lock);
244 rt_period = ktime_to_ns(rt_b->rt_period);
245 for_each_cpu_mask_nr(i, rd->span) {
246 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
252 spin_lock(&iter->rt_runtime_lock);
253 if (iter->rt_runtime == RUNTIME_INF)
256 diff = iter->rt_runtime - iter->rt_time;
258 diff = div_u64((u64)diff, weight);
259 if (rt_rq->rt_runtime + diff > rt_period)
260 diff = rt_period - rt_rq->rt_runtime;
261 iter->rt_runtime -= diff;
262 rt_rq->rt_runtime += diff;
264 if (rt_rq->rt_runtime == rt_period) {
265 spin_unlock(&iter->rt_runtime_lock);
270 spin_unlock(&iter->rt_runtime_lock);
272 spin_unlock(&rt_b->rt_runtime_lock);
277 static void __disable_runtime(struct rq *rq)
279 struct root_domain *rd = rq->rd;
282 if (unlikely(!scheduler_running))
285 for_each_leaf_rt_rq(rt_rq, rq) {
286 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
290 spin_lock(&rt_b->rt_runtime_lock);
291 spin_lock(&rt_rq->rt_runtime_lock);
292 if (rt_rq->rt_runtime == RUNTIME_INF ||
293 rt_rq->rt_runtime == rt_b->rt_runtime)
295 spin_unlock(&rt_rq->rt_runtime_lock);
297 want = rt_b->rt_runtime - rt_rq->rt_runtime;
299 for_each_cpu_mask(i, rd->span) {
300 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
303 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
306 spin_lock(&iter->rt_runtime_lock);
308 diff = min_t(s64, iter->rt_runtime, want);
309 iter->rt_runtime -= diff;
312 iter->rt_runtime -= want;
315 spin_unlock(&iter->rt_runtime_lock);
321 spin_lock(&rt_rq->rt_runtime_lock);
324 rt_rq->rt_runtime = RUNTIME_INF;
325 spin_unlock(&rt_rq->rt_runtime_lock);
326 spin_unlock(&rt_b->rt_runtime_lock);
330 static void disable_runtime(struct rq *rq)
334 spin_lock_irqsave(&rq->lock, flags);
335 __disable_runtime(rq);
336 spin_unlock_irqrestore(&rq->lock, flags);
339 static void __enable_runtime(struct rq *rq)
343 if (unlikely(!scheduler_running))
346 for_each_leaf_rt_rq(rt_rq, rq) {
347 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
349 spin_lock(&rt_b->rt_runtime_lock);
350 spin_lock(&rt_rq->rt_runtime_lock);
351 rt_rq->rt_runtime = rt_b->rt_runtime;
353 rt_rq->rt_throttled = 0;
354 spin_unlock(&rt_rq->rt_runtime_lock);
355 spin_unlock(&rt_b->rt_runtime_lock);
359 static void enable_runtime(struct rq *rq)
363 spin_lock_irqsave(&rq->lock, flags);
364 __enable_runtime(rq);
365 spin_unlock_irqrestore(&rq->lock, flags);
368 static int balance_runtime(struct rt_rq *rt_rq)
372 if (rt_rq->rt_time > rt_rq->rt_runtime) {
373 spin_unlock(&rt_rq->rt_runtime_lock);
374 more = do_balance_runtime(rt_rq);
375 spin_lock(&rt_rq->rt_runtime_lock);
380 #else /* !CONFIG_SMP */
381 static inline int balance_runtime(struct rt_rq *rt_rq)
385 #endif /* CONFIG_SMP */
387 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
392 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
395 span = sched_rt_period_mask();
396 for_each_cpu_mask(i, span) {
398 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
399 struct rq *rq = rq_of_rt_rq(rt_rq);
401 spin_lock(&rq->lock);
402 if (rt_rq->rt_time) {
405 spin_lock(&rt_rq->rt_runtime_lock);
406 if (rt_rq->rt_throttled)
407 balance_runtime(rt_rq);
408 runtime = rt_rq->rt_runtime;
409 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
410 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
411 rt_rq->rt_throttled = 0;
414 if (rt_rq->rt_time || rt_rq->rt_nr_running)
416 spin_unlock(&rt_rq->rt_runtime_lock);
417 } else if (rt_rq->rt_nr_running)
421 sched_rt_rq_enqueue(rt_rq);
422 spin_unlock(&rq->lock);
428 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
430 #ifdef CONFIG_RT_GROUP_SCHED
431 struct rt_rq *rt_rq = group_rt_rq(rt_se);
434 return rt_rq->highest_prio;
437 return rt_task_of(rt_se)->prio;
440 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
442 u64 runtime = sched_rt_runtime(rt_rq);
444 if (rt_rq->rt_throttled)
445 return rt_rq_throttled(rt_rq);
447 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
450 balance_runtime(rt_rq);
451 runtime = sched_rt_runtime(rt_rq);
452 if (runtime == RUNTIME_INF)
455 if (rt_rq->rt_time > runtime) {
456 rt_rq->rt_throttled = 1;
457 if (rt_rq_throttled(rt_rq)) {
458 sched_rt_rq_dequeue(rt_rq);
467 * Update the current task's runtime statistics. Skip current tasks that
468 * are not in our scheduling class.
470 static void update_curr_rt(struct rq *rq)
472 struct task_struct *curr = rq->curr;
473 struct sched_rt_entity *rt_se = &curr->rt;
474 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
477 if (!task_has_rt_policy(curr))
480 delta_exec = rq->clock - curr->se.exec_start;
481 if (unlikely((s64)delta_exec < 0))
484 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
486 curr->se.sum_exec_runtime += delta_exec;
487 curr->se.exec_start = rq->clock;
488 cpuacct_charge(curr, delta_exec);
490 if (!rt_bandwidth_enabled())
493 for_each_sched_rt_entity(rt_se) {
494 rt_rq = rt_rq_of_se(rt_se);
496 spin_lock(&rt_rq->rt_runtime_lock);
497 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
498 rt_rq->rt_time += delta_exec;
499 if (sched_rt_runtime_exceeded(rt_rq))
502 spin_unlock(&rt_rq->rt_runtime_lock);
507 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
509 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
510 rt_rq->rt_nr_running++;
511 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
512 if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
514 struct rq *rq = rq_of_rt_rq(rt_rq);
517 rt_rq->highest_prio = rt_se_prio(rt_se);
520 cpupri_set(&rq->rd->cpupri, rq->cpu,
526 if (rt_se->nr_cpus_allowed > 1) {
527 struct rq *rq = rq_of_rt_rq(rt_rq);
529 rq->rt.rt_nr_migratory++;
532 update_rt_migration(rq_of_rt_rq(rt_rq));
534 #ifdef CONFIG_RT_GROUP_SCHED
535 if (rt_se_boosted(rt_se))
536 rt_rq->rt_nr_boosted++;
539 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
541 start_rt_bandwidth(&def_rt_bandwidth);
546 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
549 int highest_prio = rt_rq->highest_prio;
552 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
553 WARN_ON(!rt_rq->rt_nr_running);
554 rt_rq->rt_nr_running--;
555 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
556 if (rt_rq->rt_nr_running) {
557 struct rt_prio_array *array;
559 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
560 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
562 array = &rt_rq->active;
563 rt_rq->highest_prio =
564 sched_find_first_bit(array->bitmap);
565 } /* otherwise leave rq->highest prio alone */
567 rt_rq->highest_prio = MAX_RT_PRIO;
570 if (rt_se->nr_cpus_allowed > 1) {
571 struct rq *rq = rq_of_rt_rq(rt_rq);
572 rq->rt.rt_nr_migratory--;
575 if (rt_rq->highest_prio != highest_prio) {
576 struct rq *rq = rq_of_rt_rq(rt_rq);
579 cpupri_set(&rq->rd->cpupri, rq->cpu,
580 rt_rq->highest_prio);
583 update_rt_migration(rq_of_rt_rq(rt_rq));
584 #endif /* CONFIG_SMP */
585 #ifdef CONFIG_RT_GROUP_SCHED
586 if (rt_se_boosted(rt_se))
587 rt_rq->rt_nr_boosted--;
589 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
593 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
595 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
596 struct rt_prio_array *array = &rt_rq->active;
597 struct rt_rq *group_rq = group_rt_rq(rt_se);
598 struct list_head *queue = array->queue + rt_se_prio(rt_se);
601 * Don't enqueue the group if its throttled, or when empty.
602 * The latter is a consequence of the former when a child group
603 * get throttled and the current group doesn't have any other
606 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
609 list_add_tail(&rt_se->run_list, queue);
610 __set_bit(rt_se_prio(rt_se), array->bitmap);
612 inc_rt_tasks(rt_se, rt_rq);
615 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
617 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
618 struct rt_prio_array *array = &rt_rq->active;
620 list_del_init(&rt_se->run_list);
621 if (list_empty(array->queue + rt_se_prio(rt_se)))
622 __clear_bit(rt_se_prio(rt_se), array->bitmap);
624 dec_rt_tasks(rt_se, rt_rq);
628 * Because the prio of an upper entry depends on the lower
629 * entries, we must remove entries top - down.
631 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
633 struct sched_rt_entity *back = NULL;
635 for_each_sched_rt_entity(rt_se) {
640 for (rt_se = back; rt_se; rt_se = rt_se->back) {
642 __dequeue_rt_entity(rt_se);
646 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
648 dequeue_rt_stack(rt_se);
649 for_each_sched_rt_entity(rt_se)
650 __enqueue_rt_entity(rt_se);
653 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
655 dequeue_rt_stack(rt_se);
657 for_each_sched_rt_entity(rt_se) {
658 struct rt_rq *rt_rq = group_rt_rq(rt_se);
660 if (rt_rq && rt_rq->rt_nr_running)
661 __enqueue_rt_entity(rt_se);
666 * Adding/removing a task to/from a priority array:
668 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
670 struct sched_rt_entity *rt_se = &p->rt;
675 enqueue_rt_entity(rt_se);
677 inc_cpu_load(rq, p->se.load.weight);
680 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
682 struct sched_rt_entity *rt_se = &p->rt;
685 dequeue_rt_entity(rt_se);
687 dec_cpu_load(rq, p->se.load.weight);
691 * Put task to the end of the run list without the overhead of dequeue
692 * followed by enqueue.
695 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
697 if (on_rt_rq(rt_se)) {
698 struct rt_prio_array *array = &rt_rq->active;
699 struct list_head *queue = array->queue + rt_se_prio(rt_se);
702 list_move(&rt_se->run_list, queue);
704 list_move_tail(&rt_se->run_list, queue);
708 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
710 struct sched_rt_entity *rt_se = &p->rt;
713 for_each_sched_rt_entity(rt_se) {
714 rt_rq = rt_rq_of_se(rt_se);
715 requeue_rt_entity(rt_rq, rt_se, head);
719 static void yield_task_rt(struct rq *rq)
721 requeue_task_rt(rq, rq->curr, 0);
725 static int find_lowest_rq(struct task_struct *task);
727 static int select_task_rq_rt(struct task_struct *p, int sync)
729 struct rq *rq = task_rq(p);
732 * If the current task is an RT task, then
733 * try to see if we can wake this RT task up on another
734 * runqueue. Otherwise simply start this RT task
735 * on its current runqueue.
737 * We want to avoid overloading runqueues. Even if
738 * the RT task is of higher priority than the current RT task.
739 * RT tasks behave differently than other tasks. If
740 * one gets preempted, we try to push it off to another queue.
741 * So trying to keep a preempting RT task on the same
742 * cache hot CPU will force the running RT task to
743 * a cold CPU. So we waste all the cache for the lower
744 * RT task in hopes of saving some of a RT task
745 * that is just being woken and probably will have
748 if (unlikely(rt_task(rq->curr)) &&
749 (p->rt.nr_cpus_allowed > 1)) {
750 int cpu = find_lowest_rq(p);
752 return (cpu == -1) ? task_cpu(p) : cpu;
756 * Otherwise, just let it ride on the affined RQ and the
757 * post-schedule router will push the preempted task away
762 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
766 if (rq->curr->rt.nr_cpus_allowed == 1)
769 if (p->rt.nr_cpus_allowed != 1
770 && cpupri_find(&rq->rd->cpupri, p, &mask))
773 if (!cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
777 * There appears to be other cpus that can accept
778 * current and none to run 'p', so lets reschedule
779 * to try and push current away:
781 requeue_task_rt(rq, p, 1);
782 resched_task(rq->curr);
785 #endif /* CONFIG_SMP */
788 * Preempt the current task with a newly woken task if needed:
790 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
792 if (p->prio < rq->curr->prio) {
793 resched_task(rq->curr);
801 * - the newly woken task is of equal priority to the current task
802 * - the newly woken task is non-migratable while current is migratable
803 * - current will be preempted on the next reschedule
805 * we should check to see if current can readily move to a different
806 * cpu. If so, we will reschedule to allow the push logic to try
807 * to move current somewhere else, making room for our non-migratable
810 if (p->prio == rq->curr->prio && !need_resched())
811 check_preempt_equal_prio(rq, p);
815 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
818 struct rt_prio_array *array = &rt_rq->active;
819 struct sched_rt_entity *next = NULL;
820 struct list_head *queue;
823 idx = sched_find_first_bit(array->bitmap);
824 BUG_ON(idx >= MAX_RT_PRIO);
826 queue = array->queue + idx;
827 next = list_entry(queue->next, struct sched_rt_entity, run_list);
832 static struct task_struct *pick_next_task_rt(struct rq *rq)
834 struct sched_rt_entity *rt_se;
835 struct task_struct *p;
840 if (unlikely(!rt_rq->rt_nr_running))
843 if (rt_rq_throttled(rt_rq))
847 rt_se = pick_next_rt_entity(rq, rt_rq);
849 rt_rq = group_rt_rq(rt_se);
852 p = rt_task_of(rt_se);
853 p->se.exec_start = rq->clock;
857 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
860 p->se.exec_start = 0;
865 /* Only try algorithms three times */
866 #define RT_MAX_TRIES 3
868 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
869 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest);
871 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
873 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
875 if (!task_running(rq, p) &&
876 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
877 (p->rt.nr_cpus_allowed > 1))
882 /* Return the second highest RT task, NULL otherwise */
883 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
885 struct task_struct *next = NULL;
886 struct sched_rt_entity *rt_se;
887 struct rt_prio_array *array;
891 for_each_leaf_rt_rq(rt_rq, rq) {
892 array = &rt_rq->active;
893 idx = sched_find_first_bit(array->bitmap);
895 if (idx >= MAX_RT_PRIO)
897 if (next && next->prio < idx)
899 list_for_each_entry(rt_se, array->queue + idx, run_list) {
900 struct task_struct *p = rt_task_of(rt_se);
901 if (pick_rt_task(rq, p, cpu)) {
907 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
915 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
917 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
921 /* "this_cpu" is cheaper to preempt than a remote processor */
922 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
925 first = first_cpu(*mask);
926 if (first != NR_CPUS)
932 static int find_lowest_rq(struct task_struct *task)
934 struct sched_domain *sd;
935 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
936 int this_cpu = smp_processor_id();
937 int cpu = task_cpu(task);
939 if (task->rt.nr_cpus_allowed == 1)
940 return -1; /* No other targets possible */
942 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
943 return -1; /* No targets found */
946 * Only consider CPUs that are usable for migration.
947 * I guess we might want to change cpupri_find() to ignore those
948 * in the first place.
950 cpus_and(*lowest_mask, *lowest_mask, cpu_active_map);
953 * At this point we have built a mask of cpus representing the
954 * lowest priority tasks in the system. Now we want to elect
955 * the best one based on our affinity and topology.
957 * We prioritize the last cpu that the task executed on since
958 * it is most likely cache-hot in that location.
960 if (cpu_isset(cpu, *lowest_mask))
964 * Otherwise, we consult the sched_domains span maps to figure
965 * out which cpu is logically closest to our hot cache data.
968 this_cpu = -1; /* Skip this_cpu opt if the same */
970 for_each_domain(cpu, sd) {
971 if (sd->flags & SD_WAKE_AFFINE) {
972 cpumask_t domain_mask;
975 cpus_and(domain_mask, sd->span, *lowest_mask);
977 best_cpu = pick_optimal_cpu(this_cpu,
985 * And finally, if there were no matches within the domains
986 * just give the caller *something* to work with from the compatible
989 return pick_optimal_cpu(this_cpu, lowest_mask);
992 /* Will lock the rq it finds */
993 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
995 struct rq *lowest_rq = NULL;
999 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1000 cpu = find_lowest_rq(task);
1002 if ((cpu == -1) || (cpu == rq->cpu))
1005 lowest_rq = cpu_rq(cpu);
1007 /* if the prio of this runqueue changed, try again */
1008 if (double_lock_balance(rq, lowest_rq)) {
1010 * We had to unlock the run queue. In
1011 * the mean time, task could have
1012 * migrated already or had its affinity changed.
1013 * Also make sure that it wasn't scheduled on its rq.
1015 if (unlikely(task_rq(task) != rq ||
1016 !cpu_isset(lowest_rq->cpu,
1017 task->cpus_allowed) ||
1018 task_running(rq, task) ||
1021 spin_unlock(&lowest_rq->lock);
1027 /* If this rq is still suitable use it. */
1028 if (lowest_rq->rt.highest_prio > task->prio)
1032 double_unlock_balance(rq, lowest_rq);
1040 * If the current CPU has more than one RT task, see if the non
1041 * running task can migrate over to a CPU that is running a task
1042 * of lesser priority.
1044 static int push_rt_task(struct rq *rq)
1046 struct task_struct *next_task;
1047 struct rq *lowest_rq;
1049 int paranoid = RT_MAX_TRIES;
1051 if (!rq->rt.overloaded)
1054 next_task = pick_next_highest_task_rt(rq, -1);
1059 if (unlikely(next_task == rq->curr)) {
1065 * It's possible that the next_task slipped in of
1066 * higher priority than current. If that's the case
1067 * just reschedule current.
1069 if (unlikely(next_task->prio < rq->curr->prio)) {
1070 resched_task(rq->curr);
1074 /* We might release rq lock */
1075 get_task_struct(next_task);
1077 /* find_lock_lowest_rq locks the rq if found */
1078 lowest_rq = find_lock_lowest_rq(next_task, rq);
1080 struct task_struct *task;
1082 * find lock_lowest_rq releases rq->lock
1083 * so it is possible that next_task has changed.
1084 * If it has, then try again.
1086 task = pick_next_highest_task_rt(rq, -1);
1087 if (unlikely(task != next_task) && task && paranoid--) {
1088 put_task_struct(next_task);
1095 deactivate_task(rq, next_task, 0);
1096 set_task_cpu(next_task, lowest_rq->cpu);
1097 activate_task(lowest_rq, next_task, 0);
1099 resched_task(lowest_rq->curr);
1101 double_unlock_balance(rq, lowest_rq);
1105 put_task_struct(next_task);
1111 * TODO: Currently we just use the second highest prio task on
1112 * the queue, and stop when it can't migrate (or there's
1113 * no more RT tasks). There may be a case where a lower
1114 * priority RT task has a different affinity than the
1115 * higher RT task. In this case the lower RT task could
1116 * possibly be able to migrate where as the higher priority
1117 * RT task could not. We currently ignore this issue.
1118 * Enhancements are welcome!
1120 static void push_rt_tasks(struct rq *rq)
1122 /* push_rt_task will return true if it moved an RT */
1123 while (push_rt_task(rq))
1127 static int pull_rt_task(struct rq *this_rq)
1129 int this_cpu = this_rq->cpu, ret = 0, cpu;
1130 struct task_struct *p, *next;
1133 if (likely(!rt_overloaded(this_rq)))
1136 next = pick_next_task_rt(this_rq);
1138 for_each_cpu_mask_nr(cpu, this_rq->rd->rto_mask) {
1139 if (this_cpu == cpu)
1142 src_rq = cpu_rq(cpu);
1144 * We can potentially drop this_rq's lock in
1145 * double_lock_balance, and another CPU could
1146 * steal our next task - hence we must cause
1147 * the caller to recalculate the next task
1150 if (double_lock_balance(this_rq, src_rq)) {
1151 struct task_struct *old_next = next;
1153 next = pick_next_task_rt(this_rq);
1154 if (next != old_next)
1159 * Are there still pullable RT tasks?
1161 if (src_rq->rt.rt_nr_running <= 1)
1164 p = pick_next_highest_task_rt(src_rq, this_cpu);
1167 * Do we have an RT task that preempts
1168 * the to-be-scheduled task?
1170 if (p && (!next || (p->prio < next->prio))) {
1171 WARN_ON(p == src_rq->curr);
1172 WARN_ON(!p->se.on_rq);
1175 * There's a chance that p is higher in priority
1176 * than what's currently running on its cpu.
1177 * This is just that p is wakeing up and hasn't
1178 * had a chance to schedule. We only pull
1179 * p if it is lower in priority than the
1180 * current task on the run queue or
1181 * this_rq next task is lower in prio than
1182 * the current task on that rq.
1184 if (p->prio < src_rq->curr->prio ||
1185 (next && next->prio < src_rq->curr->prio))
1190 deactivate_task(src_rq, p, 0);
1191 set_task_cpu(p, this_cpu);
1192 activate_task(this_rq, p, 0);
1194 * We continue with the search, just in
1195 * case there's an even higher prio task
1196 * in another runqueue. (low likelyhood
1199 * Update next so that we won't pick a task
1200 * on another cpu with a priority lower (or equal)
1201 * than the one we just picked.
1207 double_unlock_balance(this_rq, src_rq);
1213 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1215 /* Try to pull RT tasks here if we lower this rq's prio */
1216 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1220 static void post_schedule_rt(struct rq *rq)
1223 * If we have more than one rt_task queued, then
1224 * see if we can push the other rt_tasks off to other CPUS.
1225 * Note we may release the rq lock, and since
1226 * the lock was owned by prev, we need to release it
1227 * first via finish_lock_switch and then reaquire it here.
1229 if (unlikely(rq->rt.overloaded)) {
1230 spin_lock_irq(&rq->lock);
1232 spin_unlock_irq(&rq->lock);
1237 * If we are not running and we are not going to reschedule soon, we should
1238 * try to push tasks away now
1240 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1242 if (!task_running(rq, p) &&
1243 !test_tsk_need_resched(rq->curr) &&
1248 static unsigned long
1249 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1250 unsigned long max_load_move,
1251 struct sched_domain *sd, enum cpu_idle_type idle,
1252 int *all_pinned, int *this_best_prio)
1254 /* don't touch RT tasks */
1259 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1260 struct sched_domain *sd, enum cpu_idle_type idle)
1262 /* don't touch RT tasks */
1266 static void set_cpus_allowed_rt(struct task_struct *p,
1267 const cpumask_t *new_mask)
1269 int weight = cpus_weight(*new_mask);
1271 BUG_ON(!rt_task(p));
1274 * Update the migration status of the RQ if we have an RT task
1275 * which is running AND changing its weight value.
1277 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1278 struct rq *rq = task_rq(p);
1280 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1281 rq->rt.rt_nr_migratory++;
1282 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1283 BUG_ON(!rq->rt.rt_nr_migratory);
1284 rq->rt.rt_nr_migratory--;
1287 update_rt_migration(rq);
1290 p->cpus_allowed = *new_mask;
1291 p->rt.nr_cpus_allowed = weight;
1294 /* Assumes rq->lock is held */
1295 static void rq_online_rt(struct rq *rq)
1297 if (rq->rt.overloaded)
1298 rt_set_overload(rq);
1300 __enable_runtime(rq);
1302 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1305 /* Assumes rq->lock is held */
1306 static void rq_offline_rt(struct rq *rq)
1308 if (rq->rt.overloaded)
1309 rt_clear_overload(rq);
1311 __disable_runtime(rq);
1313 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1317 * When switch from the rt queue, we bring ourselves to a position
1318 * that we might want to pull RT tasks from other runqueues.
1320 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1324 * If there are other RT tasks then we will reschedule
1325 * and the scheduling of the other RT tasks will handle
1326 * the balancing. But if we are the last RT task
1327 * we may need to handle the pulling of RT tasks
1330 if (!rq->rt.rt_nr_running)
1333 #endif /* CONFIG_SMP */
1336 * When switching a task to RT, we may overload the runqueue
1337 * with RT tasks. In this case we try to push them off to
1340 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1343 int check_resched = 1;
1346 * If we are already running, then there's nothing
1347 * that needs to be done. But if we are not running
1348 * we may need to preempt the current running task.
1349 * If that current running task is also an RT task
1350 * then see if we can move to another run queue.
1354 if (rq->rt.overloaded && push_rt_task(rq) &&
1355 /* Don't resched if we changed runqueues */
1358 #endif /* CONFIG_SMP */
1359 if (check_resched && p->prio < rq->curr->prio)
1360 resched_task(rq->curr);
1365 * Priority of the task has changed. This may cause
1366 * us to initiate a push or pull.
1368 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1369 int oldprio, int running)
1374 * If our priority decreases while running, we
1375 * may need to pull tasks to this runqueue.
1377 if (oldprio < p->prio)
1380 * If there's a higher priority task waiting to run
1381 * then reschedule. Note, the above pull_rt_task
1382 * can release the rq lock and p could migrate.
1383 * Only reschedule if p is still on the same runqueue.
1385 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1388 /* For UP simply resched on drop of prio */
1389 if (oldprio < p->prio)
1391 #endif /* CONFIG_SMP */
1394 * This task is not running, but if it is
1395 * greater than the current running task
1398 if (p->prio < rq->curr->prio)
1399 resched_task(rq->curr);
1403 static void watchdog(struct rq *rq, struct task_struct *p)
1405 unsigned long soft, hard;
1410 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1411 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1413 if (soft != RLIM_INFINITY) {
1417 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1418 if (p->rt.timeout > next)
1419 p->it_sched_expires = p->se.sum_exec_runtime;
1423 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1430 * RR tasks need a special form of timeslice management.
1431 * FIFO tasks have no timeslices.
1433 if (p->policy != SCHED_RR)
1436 if (--p->rt.time_slice)
1439 p->rt.time_slice = DEF_TIMESLICE;
1442 * Requeue to the end of queue if we are not the only element
1445 if (p->rt.run_list.prev != p->rt.run_list.next) {
1446 requeue_task_rt(rq, p, 0);
1447 set_tsk_need_resched(p);
1451 static void set_curr_task_rt(struct rq *rq)
1453 struct task_struct *p = rq->curr;
1455 p->se.exec_start = rq->clock;
1458 static const struct sched_class rt_sched_class = {
1459 .next = &fair_sched_class,
1460 .enqueue_task = enqueue_task_rt,
1461 .dequeue_task = dequeue_task_rt,
1462 .yield_task = yield_task_rt,
1464 .select_task_rq = select_task_rq_rt,
1465 #endif /* CONFIG_SMP */
1467 .check_preempt_curr = check_preempt_curr_rt,
1469 .pick_next_task = pick_next_task_rt,
1470 .put_prev_task = put_prev_task_rt,
1473 .load_balance = load_balance_rt,
1474 .move_one_task = move_one_task_rt,
1475 .set_cpus_allowed = set_cpus_allowed_rt,
1476 .rq_online = rq_online_rt,
1477 .rq_offline = rq_offline_rt,
1478 .pre_schedule = pre_schedule_rt,
1479 .post_schedule = post_schedule_rt,
1480 .task_wake_up = task_wake_up_rt,
1481 .switched_from = switched_from_rt,
1484 .set_curr_task = set_curr_task_rt,
1485 .task_tick = task_tick_rt,
1487 .prio_changed = prio_changed_rt,
1488 .switched_to = switched_to_rt,
1491 #ifdef CONFIG_SCHED_DEBUG
1492 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1494 static void print_rt_stats(struct seq_file *m, int cpu)
1496 struct rt_rq *rt_rq;
1499 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1500 print_rt_rq(m, cpu, rt_rq);
1503 #endif /* CONFIG_SCHED_DEBUG */