2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
6 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
8 return container_of(rt_se, struct task_struct, rt);
11 #ifdef CONFIG_RT_GROUP_SCHED
13 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
18 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
23 #else /* CONFIG_RT_GROUP_SCHED */
25 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
27 return container_of(rt_rq, struct rq, rt);
30 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
32 struct task_struct *p = rt_task_of(rt_se);
33 struct rq *rq = task_rq(p);
38 #endif /* CONFIG_RT_GROUP_SCHED */
42 static inline int rt_overloaded(struct rq *rq)
44 return atomic_read(&rq->rd->rto_count);
47 static inline void rt_set_overload(struct rq *rq)
52 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
54 * Make sure the mask is visible before we set
55 * the overload count. That is checked to determine
56 * if we should look at the mask. It would be a shame
57 * if we looked at the mask, but the mask was not
61 atomic_inc(&rq->rd->rto_count);
64 static inline void rt_clear_overload(struct rq *rq)
69 /* the order here really doesn't matter */
70 atomic_dec(&rq->rd->rto_count);
71 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
74 static void update_rt_migration(struct rt_rq *rt_rq)
76 if (rt_rq->rt_nr_migratory && (rt_rq->rt_nr_running > 1)) {
77 if (!rt_rq->overloaded) {
78 rt_set_overload(rq_of_rt_rq(rt_rq));
79 rt_rq->overloaded = 1;
81 } else if (rt_rq->overloaded) {
82 rt_clear_overload(rq_of_rt_rq(rt_rq));
83 rt_rq->overloaded = 0;
87 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
89 if (rt_se->nr_cpus_allowed > 1)
90 rt_rq->rt_nr_migratory++;
92 update_rt_migration(rt_rq);
95 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
97 if (rt_se->nr_cpus_allowed > 1)
98 rt_rq->rt_nr_migratory--;
100 update_rt_migration(rt_rq);
103 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
105 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
106 plist_node_init(&p->pushable_tasks, p->prio);
107 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
110 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
112 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
117 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
121 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
126 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
131 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
135 #endif /* CONFIG_SMP */
137 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
139 return !list_empty(&rt_se->run_list);
142 #ifdef CONFIG_RT_GROUP_SCHED
144 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
149 return rt_rq->rt_runtime;
152 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
154 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
157 #define for_each_leaf_rt_rq(rt_rq, rq) \
158 list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
160 #define for_each_sched_rt_entity(rt_se) \
161 for (; rt_se; rt_se = rt_se->parent)
163 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
168 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
169 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
171 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
173 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
174 struct sched_rt_entity *rt_se = rt_rq->rt_se;
176 if (rt_rq->rt_nr_running) {
177 if (rt_se && !on_rt_rq(rt_se))
178 enqueue_rt_entity(rt_se);
179 if (rt_rq->highest_prio.curr < curr->prio)
184 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
186 struct sched_rt_entity *rt_se = rt_rq->rt_se;
188 if (rt_se && on_rt_rq(rt_se))
189 dequeue_rt_entity(rt_se);
192 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
194 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
197 static int rt_se_boosted(struct sched_rt_entity *rt_se)
199 struct rt_rq *rt_rq = group_rt_rq(rt_se);
200 struct task_struct *p;
203 return !!rt_rq->rt_nr_boosted;
205 p = rt_task_of(rt_se);
206 return p->prio != p->normal_prio;
210 static inline const struct cpumask *sched_rt_period_mask(void)
212 return cpu_rq(smp_processor_id())->rd->span;
215 static inline const struct cpumask *sched_rt_period_mask(void)
217 return cpu_online_mask;
222 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
224 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
227 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
229 return &rt_rq->tg->rt_bandwidth;
232 #else /* !CONFIG_RT_GROUP_SCHED */
234 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
236 return rt_rq->rt_runtime;
239 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
241 return ktime_to_ns(def_rt_bandwidth.rt_period);
244 #define for_each_leaf_rt_rq(rt_rq, rq) \
245 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
247 #define for_each_sched_rt_entity(rt_se) \
248 for (; rt_se; rt_se = NULL)
250 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
255 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
257 if (rt_rq->rt_nr_running)
258 resched_task(rq_of_rt_rq(rt_rq)->curr);
261 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
265 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
267 return rt_rq->rt_throttled;
270 static inline const struct cpumask *sched_rt_period_mask(void)
272 return cpu_online_mask;
276 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
278 return &cpu_rq(cpu)->rt;
281 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
283 return &def_rt_bandwidth;
286 #endif /* CONFIG_RT_GROUP_SCHED */
290 * We ran out of runtime, see if we can borrow some from our neighbours.
292 static int do_balance_runtime(struct rt_rq *rt_rq)
294 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
295 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
296 int i, weight, more = 0;
299 weight = cpumask_weight(rd->span);
301 spin_lock(&rt_b->rt_runtime_lock);
302 rt_period = ktime_to_ns(rt_b->rt_period);
303 for_each_cpu(i, rd->span) {
304 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
310 spin_lock(&iter->rt_runtime_lock);
312 * Either all rqs have inf runtime and there's nothing to steal
313 * or __disable_runtime() below sets a specific rq to inf to
314 * indicate its been disabled and disalow stealing.
316 if (iter->rt_runtime == RUNTIME_INF)
320 * From runqueues with spare time, take 1/n part of their
321 * spare time, but no more than our period.
323 diff = iter->rt_runtime - iter->rt_time;
325 diff = div_u64((u64)diff, weight);
326 if (rt_rq->rt_runtime + diff > rt_period)
327 diff = rt_period - rt_rq->rt_runtime;
328 iter->rt_runtime -= diff;
329 rt_rq->rt_runtime += diff;
331 if (rt_rq->rt_runtime == rt_period) {
332 spin_unlock(&iter->rt_runtime_lock);
337 spin_unlock(&iter->rt_runtime_lock);
339 spin_unlock(&rt_b->rt_runtime_lock);
345 * Ensure this RQ takes back all the runtime it lend to its neighbours.
347 static void __disable_runtime(struct rq *rq)
349 struct root_domain *rd = rq->rd;
352 if (unlikely(!scheduler_running))
355 for_each_leaf_rt_rq(rt_rq, rq) {
356 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
360 spin_lock(&rt_b->rt_runtime_lock);
361 spin_lock(&rt_rq->rt_runtime_lock);
363 * Either we're all inf and nobody needs to borrow, or we're
364 * already disabled and thus have nothing to do, or we have
365 * exactly the right amount of runtime to take out.
367 if (rt_rq->rt_runtime == RUNTIME_INF ||
368 rt_rq->rt_runtime == rt_b->rt_runtime)
370 spin_unlock(&rt_rq->rt_runtime_lock);
373 * Calculate the difference between what we started out with
374 * and what we current have, that's the amount of runtime
375 * we lend and now have to reclaim.
377 want = rt_b->rt_runtime - rt_rq->rt_runtime;
380 * Greedy reclaim, take back as much as we can.
382 for_each_cpu(i, rd->span) {
383 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
387 * Can't reclaim from ourselves or disabled runqueues.
389 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
392 spin_lock(&iter->rt_runtime_lock);
394 diff = min_t(s64, iter->rt_runtime, want);
395 iter->rt_runtime -= diff;
398 iter->rt_runtime -= want;
401 spin_unlock(&iter->rt_runtime_lock);
407 spin_lock(&rt_rq->rt_runtime_lock);
409 * We cannot be left wanting - that would mean some runtime
410 * leaked out of the system.
415 * Disable all the borrow logic by pretending we have inf
416 * runtime - in which case borrowing doesn't make sense.
418 rt_rq->rt_runtime = RUNTIME_INF;
419 spin_unlock(&rt_rq->rt_runtime_lock);
420 spin_unlock(&rt_b->rt_runtime_lock);
424 static void disable_runtime(struct rq *rq)
428 spin_lock_irqsave(&rq->lock, flags);
429 __disable_runtime(rq);
430 spin_unlock_irqrestore(&rq->lock, flags);
433 static void __enable_runtime(struct rq *rq)
437 if (unlikely(!scheduler_running))
441 * Reset each runqueue's bandwidth settings
443 for_each_leaf_rt_rq(rt_rq, rq) {
444 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
446 spin_lock(&rt_b->rt_runtime_lock);
447 spin_lock(&rt_rq->rt_runtime_lock);
448 rt_rq->rt_runtime = rt_b->rt_runtime;
450 rt_rq->rt_throttled = 0;
451 spin_unlock(&rt_rq->rt_runtime_lock);
452 spin_unlock(&rt_b->rt_runtime_lock);
456 static void enable_runtime(struct rq *rq)
460 spin_lock_irqsave(&rq->lock, flags);
461 __enable_runtime(rq);
462 spin_unlock_irqrestore(&rq->lock, flags);
465 static int balance_runtime(struct rt_rq *rt_rq)
469 if (rt_rq->rt_time > rt_rq->rt_runtime) {
470 spin_unlock(&rt_rq->rt_runtime_lock);
471 more = do_balance_runtime(rt_rq);
472 spin_lock(&rt_rq->rt_runtime_lock);
477 #else /* !CONFIG_SMP */
478 static inline int balance_runtime(struct rt_rq *rt_rq)
482 #endif /* CONFIG_SMP */
484 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
487 const struct cpumask *span;
489 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
492 span = sched_rt_period_mask();
493 for_each_cpu(i, span) {
495 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
496 struct rq *rq = rq_of_rt_rq(rt_rq);
498 spin_lock(&rq->lock);
499 if (rt_rq->rt_time) {
502 spin_lock(&rt_rq->rt_runtime_lock);
503 if (rt_rq->rt_throttled)
504 balance_runtime(rt_rq);
505 runtime = rt_rq->rt_runtime;
506 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
507 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
508 rt_rq->rt_throttled = 0;
511 if (rt_rq->rt_time || rt_rq->rt_nr_running)
513 spin_unlock(&rt_rq->rt_runtime_lock);
514 } else if (rt_rq->rt_nr_running)
518 sched_rt_rq_enqueue(rt_rq);
519 spin_unlock(&rq->lock);
525 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
527 #ifdef CONFIG_RT_GROUP_SCHED
528 struct rt_rq *rt_rq = group_rt_rq(rt_se);
531 return rt_rq->highest_prio.curr;
534 return rt_task_of(rt_se)->prio;
537 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
539 u64 runtime = sched_rt_runtime(rt_rq);
541 if (rt_rq->rt_throttled)
542 return rt_rq_throttled(rt_rq);
544 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
547 balance_runtime(rt_rq);
548 runtime = sched_rt_runtime(rt_rq);
549 if (runtime == RUNTIME_INF)
552 if (rt_rq->rt_time > runtime) {
553 rt_rq->rt_throttled = 1;
554 if (rt_rq_throttled(rt_rq)) {
555 sched_rt_rq_dequeue(rt_rq);
564 * Update the current task's runtime statistics. Skip current tasks that
565 * are not in our scheduling class.
567 static void update_curr_rt(struct rq *rq)
569 struct task_struct *curr = rq->curr;
570 struct sched_rt_entity *rt_se = &curr->rt;
571 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
574 if (!task_has_rt_policy(curr))
577 delta_exec = rq->clock - curr->se.exec_start;
578 if (unlikely((s64)delta_exec < 0))
581 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
583 curr->se.sum_exec_runtime += delta_exec;
584 account_group_exec_runtime(curr, delta_exec);
586 curr->se.exec_start = rq->clock;
587 cpuacct_charge(curr, delta_exec);
589 if (!rt_bandwidth_enabled())
592 for_each_sched_rt_entity(rt_se) {
593 rt_rq = rt_rq_of_se(rt_se);
595 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
596 spin_lock(&rt_rq->rt_runtime_lock);
597 rt_rq->rt_time += delta_exec;
598 if (sched_rt_runtime_exceeded(rt_rq))
600 spin_unlock(&rt_rq->rt_runtime_lock);
605 #if defined CONFIG_SMP
607 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
609 static inline int next_prio(struct rq *rq)
611 struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
613 if (next && rt_prio(next->prio))
620 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
622 struct rq *rq = rq_of_rt_rq(rt_rq);
624 if (prio < prev_prio) {
627 * If the new task is higher in priority than anything on the
628 * run-queue, we know that the previous high becomes our
631 rt_rq->highest_prio.next = prev_prio;
634 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
636 } else if (prio == rt_rq->highest_prio.curr)
638 * If the next task is equal in priority to the highest on
639 * the run-queue, then we implicitly know that the next highest
640 * task cannot be any lower than current
642 rt_rq->highest_prio.next = prio;
643 else if (prio < rt_rq->highest_prio.next)
645 * Otherwise, we need to recompute next-highest
647 rt_rq->highest_prio.next = next_prio(rq);
651 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
653 struct rq *rq = rq_of_rt_rq(rt_rq);
655 if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
656 rt_rq->highest_prio.next = next_prio(rq);
658 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
659 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
662 #else /* CONFIG_SMP */
665 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
667 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
669 #endif /* CONFIG_SMP */
671 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
673 inc_rt_prio(struct rt_rq *rt_rq, int prio)
675 int prev_prio = rt_rq->highest_prio.curr;
677 if (prio < prev_prio)
678 rt_rq->highest_prio.curr = prio;
680 inc_rt_prio_smp(rt_rq, prio, prev_prio);
684 dec_rt_prio(struct rt_rq *rt_rq, int prio)
686 int prev_prio = rt_rq->highest_prio.curr;
688 if (rt_rq->rt_nr_running) {
690 WARN_ON(prio < prev_prio);
693 * This may have been our highest task, and therefore
694 * we may have some recomputation to do
696 if (prio == prev_prio) {
697 struct rt_prio_array *array = &rt_rq->active;
699 rt_rq->highest_prio.curr =
700 sched_find_first_bit(array->bitmap);
704 rt_rq->highest_prio.curr = MAX_RT_PRIO;
706 dec_rt_prio_smp(rt_rq, prio, prev_prio);
711 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
712 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
714 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
716 #ifdef CONFIG_RT_GROUP_SCHED
719 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
721 if (rt_se_boosted(rt_se))
722 rt_rq->rt_nr_boosted++;
725 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
729 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
731 if (rt_se_boosted(rt_se))
732 rt_rq->rt_nr_boosted--;
734 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
737 #else /* CONFIG_RT_GROUP_SCHED */
740 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
742 start_rt_bandwidth(&def_rt_bandwidth);
746 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
748 #endif /* CONFIG_RT_GROUP_SCHED */
751 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
753 int prio = rt_se_prio(rt_se);
755 WARN_ON(!rt_prio(prio));
756 rt_rq->rt_nr_running++;
758 inc_rt_prio(rt_rq, prio);
759 inc_rt_migration(rt_se, rt_rq);
760 inc_rt_group(rt_se, rt_rq);
764 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
766 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
767 WARN_ON(!rt_rq->rt_nr_running);
768 rt_rq->rt_nr_running--;
770 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
771 dec_rt_migration(rt_se, rt_rq);
772 dec_rt_group(rt_se, rt_rq);
775 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
777 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
778 struct rt_prio_array *array = &rt_rq->active;
779 struct rt_rq *group_rq = group_rt_rq(rt_se);
780 struct list_head *queue = array->queue + rt_se_prio(rt_se);
783 * Don't enqueue the group if its throttled, or when empty.
784 * The latter is a consequence of the former when a child group
785 * get throttled and the current group doesn't have any other
788 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
791 list_add_tail(&rt_se->run_list, queue);
792 __set_bit(rt_se_prio(rt_se), array->bitmap);
794 inc_rt_tasks(rt_se, rt_rq);
797 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
799 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
800 struct rt_prio_array *array = &rt_rq->active;
802 list_del_init(&rt_se->run_list);
803 if (list_empty(array->queue + rt_se_prio(rt_se)))
804 __clear_bit(rt_se_prio(rt_se), array->bitmap);
806 dec_rt_tasks(rt_se, rt_rq);
810 * Because the prio of an upper entry depends on the lower
811 * entries, we must remove entries top - down.
813 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
815 struct sched_rt_entity *back = NULL;
817 for_each_sched_rt_entity(rt_se) {
822 for (rt_se = back; rt_se; rt_se = rt_se->back) {
824 __dequeue_rt_entity(rt_se);
828 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
830 dequeue_rt_stack(rt_se);
831 for_each_sched_rt_entity(rt_se)
832 __enqueue_rt_entity(rt_se);
835 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
837 dequeue_rt_stack(rt_se);
839 for_each_sched_rt_entity(rt_se) {
840 struct rt_rq *rt_rq = group_rt_rq(rt_se);
842 if (rt_rq && rt_rq->rt_nr_running)
843 __enqueue_rt_entity(rt_se);
848 * Adding/removing a task to/from a priority array:
850 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
852 struct sched_rt_entity *rt_se = &p->rt;
857 enqueue_rt_entity(rt_se);
859 if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
860 enqueue_pushable_task(rq, p);
862 inc_cpu_load(rq, p->se.load.weight);
865 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
867 struct sched_rt_entity *rt_se = &p->rt;
870 dequeue_rt_entity(rt_se);
872 dequeue_pushable_task(rq, p);
874 dec_cpu_load(rq, p->se.load.weight);
878 * Put task to the end of the run list without the overhead of dequeue
879 * followed by enqueue.
882 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
884 if (on_rt_rq(rt_se)) {
885 struct rt_prio_array *array = &rt_rq->active;
886 struct list_head *queue = array->queue + rt_se_prio(rt_se);
889 list_move(&rt_se->run_list, queue);
891 list_move_tail(&rt_se->run_list, queue);
895 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
897 struct sched_rt_entity *rt_se = &p->rt;
900 for_each_sched_rt_entity(rt_se) {
901 rt_rq = rt_rq_of_se(rt_se);
902 requeue_rt_entity(rt_rq, rt_se, head);
906 static void yield_task_rt(struct rq *rq)
908 requeue_task_rt(rq, rq->curr, 0);
912 static int find_lowest_rq(struct task_struct *task);
914 static int select_task_rq_rt(struct task_struct *p, int sync)
916 struct rq *rq = task_rq(p);
919 * If the current task is an RT task, then
920 * try to see if we can wake this RT task up on another
921 * runqueue. Otherwise simply start this RT task
922 * on its current runqueue.
924 * We want to avoid overloading runqueues. Even if
925 * the RT task is of higher priority than the current RT task.
926 * RT tasks behave differently than other tasks. If
927 * one gets preempted, we try to push it off to another queue.
928 * So trying to keep a preempting RT task on the same
929 * cache hot CPU will force the running RT task to
930 * a cold CPU. So we waste all the cache for the lower
931 * RT task in hopes of saving some of a RT task
932 * that is just being woken and probably will have
935 if (unlikely(rt_task(rq->curr)) &&
936 (p->rt.nr_cpus_allowed > 1)) {
937 int cpu = find_lowest_rq(p);
939 return (cpu == -1) ? task_cpu(p) : cpu;
943 * Otherwise, just let it ride on the affined RQ and the
944 * post-schedule router will push the preempted task away
949 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
951 if (rq->curr->rt.nr_cpus_allowed == 1)
954 if (p->rt.nr_cpus_allowed != 1
955 && cpupri_find(&rq->rd->cpupri, p, NULL))
958 if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
962 * There appears to be other cpus that can accept
963 * current and none to run 'p', so lets reschedule
964 * to try and push current away:
966 requeue_task_rt(rq, p, 1);
967 resched_task(rq->curr);
970 #endif /* CONFIG_SMP */
973 * Preempt the current task with a newly woken task if needed:
975 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
977 if (p->prio < rq->curr->prio) {
978 resched_task(rq->curr);
986 * - the newly woken task is of equal priority to the current task
987 * - the newly woken task is non-migratable while current is migratable
988 * - current will be preempted on the next reschedule
990 * we should check to see if current can readily move to a different
991 * cpu. If so, we will reschedule to allow the push logic to try
992 * to move current somewhere else, making room for our non-migratable
995 if (p->prio == rq->curr->prio && !need_resched())
996 check_preempt_equal_prio(rq, p);
1000 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1001 struct rt_rq *rt_rq)
1003 struct rt_prio_array *array = &rt_rq->active;
1004 struct sched_rt_entity *next = NULL;
1005 struct list_head *queue;
1008 idx = sched_find_first_bit(array->bitmap);
1009 BUG_ON(idx >= MAX_RT_PRIO);
1011 queue = array->queue + idx;
1012 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1017 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1019 struct sched_rt_entity *rt_se;
1020 struct task_struct *p;
1021 struct rt_rq *rt_rq;
1025 if (unlikely(!rt_rq->rt_nr_running))
1028 if (rt_rq_throttled(rt_rq))
1032 rt_se = pick_next_rt_entity(rq, rt_rq);
1034 rt_rq = group_rt_rq(rt_se);
1037 p = rt_task_of(rt_se);
1038 p->se.exec_start = rq->clock;
1043 static struct task_struct *pick_next_task_rt(struct rq *rq)
1045 struct task_struct *p = _pick_next_task_rt(rq);
1047 /* The running task is never eligible for pushing */
1049 dequeue_pushable_task(rq, p);
1054 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1057 p->se.exec_start = 0;
1060 * The previous task needs to be made eligible for pushing
1061 * if it is still active
1063 if (p->se.on_rq && p->rt.nr_cpus_allowed > 1)
1064 enqueue_pushable_task(rq, p);
1069 /* Only try algorithms three times */
1070 #define RT_MAX_TRIES 3
1072 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1074 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1076 if (!task_running(rq, p) &&
1077 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1078 (p->rt.nr_cpus_allowed > 1))
1083 /* Return the second highest RT task, NULL otherwise */
1084 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1086 struct task_struct *next = NULL;
1087 struct sched_rt_entity *rt_se;
1088 struct rt_prio_array *array;
1089 struct rt_rq *rt_rq;
1092 for_each_leaf_rt_rq(rt_rq, rq) {
1093 array = &rt_rq->active;
1094 idx = sched_find_first_bit(array->bitmap);
1096 if (idx >= MAX_RT_PRIO)
1098 if (next && next->prio < idx)
1100 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1101 struct task_struct *p = rt_task_of(rt_se);
1102 if (pick_rt_task(rq, p, cpu)) {
1108 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1116 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1118 static inline int pick_optimal_cpu(int this_cpu,
1119 const struct cpumask *mask)
1123 /* "this_cpu" is cheaper to preempt than a remote processor */
1124 if ((this_cpu != -1) && cpumask_test_cpu(this_cpu, mask))
1127 first = cpumask_first(mask);
1128 if (first < nr_cpu_ids)
1134 static int find_lowest_rq(struct task_struct *task)
1136 struct sched_domain *sd;
1137 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1138 int this_cpu = smp_processor_id();
1139 int cpu = task_cpu(task);
1140 cpumask_var_t domain_mask;
1142 if (task->rt.nr_cpus_allowed == 1)
1143 return -1; /* No other targets possible */
1145 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1146 return -1; /* No targets found */
1149 * Only consider CPUs that are usable for migration.
1150 * I guess we might want to change cpupri_find() to ignore those
1151 * in the first place.
1153 cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
1156 * At this point we have built a mask of cpus representing the
1157 * lowest priority tasks in the system. Now we want to elect
1158 * the best one based on our affinity and topology.
1160 * We prioritize the last cpu that the task executed on since
1161 * it is most likely cache-hot in that location.
1163 if (cpumask_test_cpu(cpu, lowest_mask))
1167 * Otherwise, we consult the sched_domains span maps to figure
1168 * out which cpu is logically closest to our hot cache data.
1170 if (this_cpu == cpu)
1171 this_cpu = -1; /* Skip this_cpu opt if the same */
1173 if (alloc_cpumask_var(&domain_mask, GFP_ATOMIC)) {
1174 for_each_domain(cpu, sd) {
1175 if (sd->flags & SD_WAKE_AFFINE) {
1178 cpumask_and(domain_mask,
1179 sched_domain_span(sd),
1182 best_cpu = pick_optimal_cpu(this_cpu,
1185 if (best_cpu != -1) {
1186 free_cpumask_var(domain_mask);
1191 free_cpumask_var(domain_mask);
1195 * And finally, if there were no matches within the domains
1196 * just give the caller *something* to work with from the compatible
1199 return pick_optimal_cpu(this_cpu, lowest_mask);
1202 /* Will lock the rq it finds */
1203 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1205 struct rq *lowest_rq = NULL;
1209 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1210 cpu = find_lowest_rq(task);
1212 if ((cpu == -1) || (cpu == rq->cpu))
1215 lowest_rq = cpu_rq(cpu);
1217 /* if the prio of this runqueue changed, try again */
1218 if (double_lock_balance(rq, lowest_rq)) {
1220 * We had to unlock the run queue. In
1221 * the mean time, task could have
1222 * migrated already or had its affinity changed.
1223 * Also make sure that it wasn't scheduled on its rq.
1225 if (unlikely(task_rq(task) != rq ||
1226 !cpumask_test_cpu(lowest_rq->cpu,
1227 &task->cpus_allowed) ||
1228 task_running(rq, task) ||
1231 spin_unlock(&lowest_rq->lock);
1237 /* If this rq is still suitable use it. */
1238 if (lowest_rq->rt.highest_prio.curr > task->prio)
1242 double_unlock_balance(rq, lowest_rq);
1249 static inline int has_pushable_tasks(struct rq *rq)
1251 return !plist_head_empty(&rq->rt.pushable_tasks);
1254 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1256 struct task_struct *p;
1258 if (!has_pushable_tasks(rq))
1261 p = plist_first_entry(&rq->rt.pushable_tasks,
1262 struct task_struct, pushable_tasks);
1264 BUG_ON(rq->cpu != task_cpu(p));
1265 BUG_ON(task_current(rq, p));
1266 BUG_ON(p->rt.nr_cpus_allowed <= 1);
1268 BUG_ON(!p->se.on_rq);
1269 BUG_ON(!rt_task(p));
1275 * If the current CPU has more than one RT task, see if the non
1276 * running task can migrate over to a CPU that is running a task
1277 * of lesser priority.
1279 static int push_rt_task(struct rq *rq)
1281 struct task_struct *next_task;
1282 struct rq *lowest_rq;
1284 if (!rq->rt.overloaded)
1287 next_task = pick_next_pushable_task(rq);
1292 if (unlikely(next_task == rq->curr)) {
1298 * It's possible that the next_task slipped in of
1299 * higher priority than current. If that's the case
1300 * just reschedule current.
1302 if (unlikely(next_task->prio < rq->curr->prio)) {
1303 resched_task(rq->curr);
1307 /* We might release rq lock */
1308 get_task_struct(next_task);
1310 /* find_lock_lowest_rq locks the rq if found */
1311 lowest_rq = find_lock_lowest_rq(next_task, rq);
1313 struct task_struct *task;
1315 * find lock_lowest_rq releases rq->lock
1316 * so it is possible that next_task has migrated.
1318 * We need to make sure that the task is still on the same
1319 * run-queue and is also still the next task eligible for
1322 task = pick_next_pushable_task(rq);
1323 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1325 * If we get here, the task hasnt moved at all, but
1326 * it has failed to push. We will not try again,
1327 * since the other cpus will pull from us when they
1330 dequeue_pushable_task(rq, next_task);
1335 /* No more tasks, just exit */
1339 * Something has shifted, try again.
1341 put_task_struct(next_task);
1346 deactivate_task(rq, next_task, 0);
1347 set_task_cpu(next_task, lowest_rq->cpu);
1348 activate_task(lowest_rq, next_task, 0);
1350 resched_task(lowest_rq->curr);
1352 double_unlock_balance(rq, lowest_rq);
1355 put_task_struct(next_task);
1360 static void push_rt_tasks(struct rq *rq)
1362 /* push_rt_task will return true if it moved an RT */
1363 while (push_rt_task(rq))
1367 static int pull_rt_task(struct rq *this_rq)
1369 int this_cpu = this_rq->cpu, ret = 0, cpu;
1370 struct task_struct *p;
1373 if (likely(!rt_overloaded(this_rq)))
1376 for_each_cpu(cpu, this_rq->rd->rto_mask) {
1377 if (this_cpu == cpu)
1380 src_rq = cpu_rq(cpu);
1383 * Don't bother taking the src_rq->lock if the next highest
1384 * task is known to be lower-priority than our current task.
1385 * This may look racy, but if this value is about to go
1386 * logically higher, the src_rq will push this task away.
1387 * And if its going logically lower, we do not care
1389 if (src_rq->rt.highest_prio.next >=
1390 this_rq->rt.highest_prio.curr)
1394 * We can potentially drop this_rq's lock in
1395 * double_lock_balance, and another CPU could
1398 double_lock_balance(this_rq, src_rq);
1401 * Are there still pullable RT tasks?
1403 if (src_rq->rt.rt_nr_running <= 1)
1406 p = pick_next_highest_task_rt(src_rq, this_cpu);
1409 * Do we have an RT task that preempts
1410 * the to-be-scheduled task?
1412 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1413 WARN_ON(p == src_rq->curr);
1414 WARN_ON(!p->se.on_rq);
1417 * There's a chance that p is higher in priority
1418 * than what's currently running on its cpu.
1419 * This is just that p is wakeing up and hasn't
1420 * had a chance to schedule. We only pull
1421 * p if it is lower in priority than the
1422 * current task on the run queue
1424 if (p->prio < src_rq->curr->prio)
1429 deactivate_task(src_rq, p, 0);
1430 set_task_cpu(p, this_cpu);
1431 activate_task(this_rq, p, 0);
1433 * We continue with the search, just in
1434 * case there's an even higher prio task
1435 * in another runqueue. (low likelyhood
1440 double_unlock_balance(this_rq, src_rq);
1446 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1448 /* Try to pull RT tasks here if we lower this rq's prio */
1449 if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1454 * assumes rq->lock is held
1456 static int needs_post_schedule_rt(struct rq *rq)
1458 return has_pushable_tasks(rq);
1461 static void post_schedule_rt(struct rq *rq)
1464 * This is only called if needs_post_schedule_rt() indicates that
1465 * we need to push tasks away
1467 spin_lock_irq(&rq->lock);
1469 spin_unlock_irq(&rq->lock);
1473 * If we are not running and we are not going to reschedule soon, we should
1474 * try to push tasks away now
1476 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1478 if (!task_running(rq, p) &&
1479 !test_tsk_need_resched(rq->curr) &&
1480 has_pushable_tasks(rq) &&
1481 p->rt.nr_cpus_allowed > 1)
1485 static unsigned long
1486 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1487 unsigned long max_load_move,
1488 struct sched_domain *sd, enum cpu_idle_type idle,
1489 int *all_pinned, int *this_best_prio)
1491 /* don't touch RT tasks */
1496 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1497 struct sched_domain *sd, enum cpu_idle_type idle)
1499 /* don't touch RT tasks */
1503 static void set_cpus_allowed_rt(struct task_struct *p,
1504 const struct cpumask *new_mask)
1506 int weight = cpumask_weight(new_mask);
1508 BUG_ON(!rt_task(p));
1511 * Update the migration status of the RQ if we have an RT task
1512 * which is running AND changing its weight value.
1514 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1515 struct rq *rq = task_rq(p);
1517 if (!task_current(rq, p)) {
1519 * Make sure we dequeue this task from the pushable list
1520 * before going further. It will either remain off of
1521 * the list because we are no longer pushable, or it
1524 if (p->rt.nr_cpus_allowed > 1)
1525 dequeue_pushable_task(rq, p);
1528 * Requeue if our weight is changing and still > 1
1531 enqueue_pushable_task(rq, p);
1535 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1536 rq->rt.rt_nr_migratory++;
1537 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1538 BUG_ON(!rq->rt.rt_nr_migratory);
1539 rq->rt.rt_nr_migratory--;
1542 update_rt_migration(&rq->rt);
1545 cpumask_copy(&p->cpus_allowed, new_mask);
1546 p->rt.nr_cpus_allowed = weight;
1549 /* Assumes rq->lock is held */
1550 static void rq_online_rt(struct rq *rq)
1552 if (rq->rt.overloaded)
1553 rt_set_overload(rq);
1555 __enable_runtime(rq);
1557 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1560 /* Assumes rq->lock is held */
1561 static void rq_offline_rt(struct rq *rq)
1563 if (rq->rt.overloaded)
1564 rt_clear_overload(rq);
1566 __disable_runtime(rq);
1568 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1572 * When switch from the rt queue, we bring ourselves to a position
1573 * that we might want to pull RT tasks from other runqueues.
1575 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1579 * If there are other RT tasks then we will reschedule
1580 * and the scheduling of the other RT tasks will handle
1581 * the balancing. But if we are the last RT task
1582 * we may need to handle the pulling of RT tasks
1585 if (!rq->rt.rt_nr_running)
1589 static inline void init_sched_rt_class(void)
1593 for_each_possible_cpu(i)
1594 alloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1595 GFP_KERNEL, cpu_to_node(i));
1597 #endif /* CONFIG_SMP */
1600 * When switching a task to RT, we may overload the runqueue
1601 * with RT tasks. In this case we try to push them off to
1604 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1607 int check_resched = 1;
1610 * If we are already running, then there's nothing
1611 * that needs to be done. But if we are not running
1612 * we may need to preempt the current running task.
1613 * If that current running task is also an RT task
1614 * then see if we can move to another run queue.
1618 if (rq->rt.overloaded && push_rt_task(rq) &&
1619 /* Don't resched if we changed runqueues */
1622 #endif /* CONFIG_SMP */
1623 if (check_resched && p->prio < rq->curr->prio)
1624 resched_task(rq->curr);
1629 * Priority of the task has changed. This may cause
1630 * us to initiate a push or pull.
1632 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1633 int oldprio, int running)
1638 * If our priority decreases while running, we
1639 * may need to pull tasks to this runqueue.
1641 if (oldprio < p->prio)
1644 * If there's a higher priority task waiting to run
1645 * then reschedule. Note, the above pull_rt_task
1646 * can release the rq lock and p could migrate.
1647 * Only reschedule if p is still on the same runqueue.
1649 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1652 /* For UP simply resched on drop of prio */
1653 if (oldprio < p->prio)
1655 #endif /* CONFIG_SMP */
1658 * This task is not running, but if it is
1659 * greater than the current running task
1662 if (p->prio < rq->curr->prio)
1663 resched_task(rq->curr);
1667 static void watchdog(struct rq *rq, struct task_struct *p)
1669 unsigned long soft, hard;
1674 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1675 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1677 if (soft != RLIM_INFINITY) {
1681 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1682 if (p->rt.timeout > next)
1683 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1687 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1694 * RR tasks need a special form of timeslice management.
1695 * FIFO tasks have no timeslices.
1697 if (p->policy != SCHED_RR)
1700 if (--p->rt.time_slice)
1703 p->rt.time_slice = DEF_TIMESLICE;
1706 * Requeue to the end of queue if we are not the only element
1709 if (p->rt.run_list.prev != p->rt.run_list.next) {
1710 requeue_task_rt(rq, p, 0);
1711 set_tsk_need_resched(p);
1715 static void set_curr_task_rt(struct rq *rq)
1717 struct task_struct *p = rq->curr;
1719 p->se.exec_start = rq->clock;
1721 /* The running task is never eligible for pushing */
1722 dequeue_pushable_task(rq, p);
1725 static const struct sched_class rt_sched_class = {
1726 .next = &fair_sched_class,
1727 .enqueue_task = enqueue_task_rt,
1728 .dequeue_task = dequeue_task_rt,
1729 .yield_task = yield_task_rt,
1731 .check_preempt_curr = check_preempt_curr_rt,
1733 .pick_next_task = pick_next_task_rt,
1734 .put_prev_task = put_prev_task_rt,
1737 .select_task_rq = select_task_rq_rt,
1739 .load_balance = load_balance_rt,
1740 .move_one_task = move_one_task_rt,
1741 .set_cpus_allowed = set_cpus_allowed_rt,
1742 .rq_online = rq_online_rt,
1743 .rq_offline = rq_offline_rt,
1744 .pre_schedule = pre_schedule_rt,
1745 .needs_post_schedule = needs_post_schedule_rt,
1746 .post_schedule = post_schedule_rt,
1747 .task_wake_up = task_wake_up_rt,
1748 .switched_from = switched_from_rt,
1751 .set_curr_task = set_curr_task_rt,
1752 .task_tick = task_tick_rt,
1754 .prio_changed = prio_changed_rt,
1755 .switched_to = switched_to_rt,
1758 #ifdef CONFIG_SCHED_DEBUG
1759 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1761 static void print_rt_stats(struct seq_file *m, int cpu)
1763 struct rt_rq *rt_rq;
1766 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1767 print_rt_rq(m, cpu, rt_rq);
1770 #endif /* CONFIG_SCHED_DEBUG */