2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
6 #ifdef CONFIG_RT_GROUP_SCHED
8 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
10 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
12 #ifdef CONFIG_SCHED_DEBUG
13 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
15 return container_of(rt_se, struct task_struct, rt);
18 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
23 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
28 #else /* CONFIG_RT_GROUP_SCHED */
30 #define rt_entity_is_task(rt_se) (1)
32 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
34 return container_of(rt_se, struct task_struct, rt);
37 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
39 return container_of(rt_rq, struct rq, rt);
42 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
44 struct task_struct *p = rt_task_of(rt_se);
45 struct rq *rq = task_rq(p);
50 #endif /* CONFIG_RT_GROUP_SCHED */
54 static inline int rt_overloaded(struct rq *rq)
56 return atomic_read(&rq->rd->rto_count);
59 static inline void rt_set_overload(struct rq *rq)
64 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
66 * Make sure the mask is visible before we set
67 * the overload count. That is checked to determine
68 * if we should look at the mask. It would be a shame
69 * if we looked at the mask, but the mask was not
73 atomic_inc(&rq->rd->rto_count);
76 static inline void rt_clear_overload(struct rq *rq)
81 /* the order here really doesn't matter */
82 atomic_dec(&rq->rd->rto_count);
83 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
86 static void update_rt_migration(struct rt_rq *rt_rq)
88 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
89 if (!rt_rq->overloaded) {
90 rt_set_overload(rq_of_rt_rq(rt_rq));
91 rt_rq->overloaded = 1;
93 } else if (rt_rq->overloaded) {
94 rt_clear_overload(rq_of_rt_rq(rt_rq));
95 rt_rq->overloaded = 0;
99 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
101 if (!rt_entity_is_task(rt_se))
104 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
106 rt_rq->rt_nr_total++;
107 if (rt_se->nr_cpus_allowed > 1)
108 rt_rq->rt_nr_migratory++;
110 update_rt_migration(rt_rq);
113 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
115 if (!rt_entity_is_task(rt_se))
118 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
120 rt_rq->rt_nr_total--;
121 if (rt_se->nr_cpus_allowed > 1)
122 rt_rq->rt_nr_migratory--;
124 update_rt_migration(rt_rq);
127 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
129 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
130 plist_node_init(&p->pushable_tasks, p->prio);
131 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
134 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
136 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
139 static inline int has_pushable_tasks(struct rq *rq)
141 return !plist_head_empty(&rq->rt.pushable_tasks);
146 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
150 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
155 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
160 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
164 #endif /* CONFIG_SMP */
166 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
168 return !list_empty(&rt_se->run_list);
171 #ifdef CONFIG_RT_GROUP_SCHED
173 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
178 return rt_rq->rt_runtime;
181 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
183 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
186 #define for_each_leaf_rt_rq(rt_rq, rq) \
187 list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
189 #define for_each_sched_rt_entity(rt_se) \
190 for (; rt_se; rt_se = rt_se->parent)
192 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
197 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
198 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
200 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
202 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
203 struct sched_rt_entity *rt_se = rt_rq->rt_se;
205 if (rt_rq->rt_nr_running) {
206 if (rt_se && !on_rt_rq(rt_se))
207 enqueue_rt_entity(rt_se, false);
208 if (rt_rq->highest_prio.curr < curr->prio)
213 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
215 struct sched_rt_entity *rt_se = rt_rq->rt_se;
217 if (rt_se && on_rt_rq(rt_se))
218 dequeue_rt_entity(rt_se);
221 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
223 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
226 static int rt_se_boosted(struct sched_rt_entity *rt_se)
228 struct rt_rq *rt_rq = group_rt_rq(rt_se);
229 struct task_struct *p;
232 return !!rt_rq->rt_nr_boosted;
234 p = rt_task_of(rt_se);
235 return p->prio != p->normal_prio;
239 static inline const struct cpumask *sched_rt_period_mask(void)
241 return cpu_rq(smp_processor_id())->rd->span;
244 static inline const struct cpumask *sched_rt_period_mask(void)
246 return cpu_online_mask;
251 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
253 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
256 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
258 return &rt_rq->tg->rt_bandwidth;
261 #else /* !CONFIG_RT_GROUP_SCHED */
263 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
265 return rt_rq->rt_runtime;
268 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
270 return ktime_to_ns(def_rt_bandwidth.rt_period);
273 #define for_each_leaf_rt_rq(rt_rq, rq) \
274 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
276 #define for_each_sched_rt_entity(rt_se) \
277 for (; rt_se; rt_se = NULL)
279 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
284 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
286 if (rt_rq->rt_nr_running)
287 resched_task(rq_of_rt_rq(rt_rq)->curr);
290 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
294 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
296 return rt_rq->rt_throttled;
299 static inline const struct cpumask *sched_rt_period_mask(void)
301 return cpu_online_mask;
305 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
307 return &cpu_rq(cpu)->rt;
310 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
312 return &def_rt_bandwidth;
315 #endif /* CONFIG_RT_GROUP_SCHED */
319 * We ran out of runtime, see if we can borrow some from our neighbours.
321 static int do_balance_runtime(struct rt_rq *rt_rq)
323 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
324 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
325 int i, weight, more = 0;
328 weight = cpumask_weight(rd->span);
330 raw_spin_lock(&rt_b->rt_runtime_lock);
331 rt_period = ktime_to_ns(rt_b->rt_period);
332 for_each_cpu(i, rd->span) {
333 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
339 raw_spin_lock(&iter->rt_runtime_lock);
341 * Either all rqs have inf runtime and there's nothing to steal
342 * or __disable_runtime() below sets a specific rq to inf to
343 * indicate its been disabled and disalow stealing.
345 if (iter->rt_runtime == RUNTIME_INF)
349 * From runqueues with spare time, take 1/n part of their
350 * spare time, but no more than our period.
352 diff = iter->rt_runtime - iter->rt_time;
354 diff = div_u64((u64)diff, weight);
355 if (rt_rq->rt_runtime + diff > rt_period)
356 diff = rt_period - rt_rq->rt_runtime;
357 iter->rt_runtime -= diff;
358 rt_rq->rt_runtime += diff;
360 if (rt_rq->rt_runtime == rt_period) {
361 raw_spin_unlock(&iter->rt_runtime_lock);
366 raw_spin_unlock(&iter->rt_runtime_lock);
368 raw_spin_unlock(&rt_b->rt_runtime_lock);
374 * Ensure this RQ takes back all the runtime it lend to its neighbours.
376 static void __disable_runtime(struct rq *rq)
378 struct root_domain *rd = rq->rd;
381 if (unlikely(!scheduler_running))
384 for_each_leaf_rt_rq(rt_rq, rq) {
385 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
389 raw_spin_lock(&rt_b->rt_runtime_lock);
390 raw_spin_lock(&rt_rq->rt_runtime_lock);
392 * Either we're all inf and nobody needs to borrow, or we're
393 * already disabled and thus have nothing to do, or we have
394 * exactly the right amount of runtime to take out.
396 if (rt_rq->rt_runtime == RUNTIME_INF ||
397 rt_rq->rt_runtime == rt_b->rt_runtime)
399 raw_spin_unlock(&rt_rq->rt_runtime_lock);
402 * Calculate the difference between what we started out with
403 * and what we current have, that's the amount of runtime
404 * we lend and now have to reclaim.
406 want = rt_b->rt_runtime - rt_rq->rt_runtime;
409 * Greedy reclaim, take back as much as we can.
411 for_each_cpu(i, rd->span) {
412 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
416 * Can't reclaim from ourselves or disabled runqueues.
418 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
421 raw_spin_lock(&iter->rt_runtime_lock);
423 diff = min_t(s64, iter->rt_runtime, want);
424 iter->rt_runtime -= diff;
427 iter->rt_runtime -= want;
430 raw_spin_unlock(&iter->rt_runtime_lock);
436 raw_spin_lock(&rt_rq->rt_runtime_lock);
438 * We cannot be left wanting - that would mean some runtime
439 * leaked out of the system.
444 * Disable all the borrow logic by pretending we have inf
445 * runtime - in which case borrowing doesn't make sense.
447 rt_rq->rt_runtime = RUNTIME_INF;
448 raw_spin_unlock(&rt_rq->rt_runtime_lock);
449 raw_spin_unlock(&rt_b->rt_runtime_lock);
453 static void disable_runtime(struct rq *rq)
457 raw_spin_lock_irqsave(&rq->lock, flags);
458 __disable_runtime(rq);
459 raw_spin_unlock_irqrestore(&rq->lock, flags);
462 static void __enable_runtime(struct rq *rq)
466 if (unlikely(!scheduler_running))
470 * Reset each runqueue's bandwidth settings
472 for_each_leaf_rt_rq(rt_rq, rq) {
473 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
475 raw_spin_lock(&rt_b->rt_runtime_lock);
476 raw_spin_lock(&rt_rq->rt_runtime_lock);
477 rt_rq->rt_runtime = rt_b->rt_runtime;
479 rt_rq->rt_throttled = 0;
480 raw_spin_unlock(&rt_rq->rt_runtime_lock);
481 raw_spin_unlock(&rt_b->rt_runtime_lock);
485 static void enable_runtime(struct rq *rq)
489 raw_spin_lock_irqsave(&rq->lock, flags);
490 __enable_runtime(rq);
491 raw_spin_unlock_irqrestore(&rq->lock, flags);
494 static int balance_runtime(struct rt_rq *rt_rq)
498 if (rt_rq->rt_time > rt_rq->rt_runtime) {
499 raw_spin_unlock(&rt_rq->rt_runtime_lock);
500 more = do_balance_runtime(rt_rq);
501 raw_spin_lock(&rt_rq->rt_runtime_lock);
506 #else /* !CONFIG_SMP */
507 static inline int balance_runtime(struct rt_rq *rt_rq)
511 #endif /* CONFIG_SMP */
513 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
516 const struct cpumask *span;
518 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
521 span = sched_rt_period_mask();
522 for_each_cpu(i, span) {
524 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
525 struct rq *rq = rq_of_rt_rq(rt_rq);
527 raw_spin_lock(&rq->lock);
528 if (rt_rq->rt_time) {
531 raw_spin_lock(&rt_rq->rt_runtime_lock);
532 if (rt_rq->rt_throttled)
533 balance_runtime(rt_rq);
534 runtime = rt_rq->rt_runtime;
535 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
536 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
537 rt_rq->rt_throttled = 0;
540 if (rt_rq->rt_time || rt_rq->rt_nr_running)
542 raw_spin_unlock(&rt_rq->rt_runtime_lock);
543 } else if (rt_rq->rt_nr_running)
547 sched_rt_rq_enqueue(rt_rq);
548 raw_spin_unlock(&rq->lock);
554 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
556 #ifdef CONFIG_RT_GROUP_SCHED
557 struct rt_rq *rt_rq = group_rt_rq(rt_se);
560 return rt_rq->highest_prio.curr;
563 return rt_task_of(rt_se)->prio;
566 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
568 u64 runtime = sched_rt_runtime(rt_rq);
570 if (rt_rq->rt_throttled)
571 return rt_rq_throttled(rt_rq);
573 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
576 balance_runtime(rt_rq);
577 runtime = sched_rt_runtime(rt_rq);
578 if (runtime == RUNTIME_INF)
581 if (rt_rq->rt_time > runtime) {
582 rt_rq->rt_throttled = 1;
583 if (rt_rq_throttled(rt_rq)) {
584 sched_rt_rq_dequeue(rt_rq);
593 * Update the current task's runtime statistics. Skip current tasks that
594 * are not in our scheduling class.
596 static void update_curr_rt(struct rq *rq)
598 struct task_struct *curr = rq->curr;
599 struct sched_rt_entity *rt_se = &curr->rt;
600 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
603 if (!task_has_rt_policy(curr))
606 delta_exec = rq->clock - curr->se.exec_start;
607 if (unlikely((s64)delta_exec < 0))
610 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
612 curr->se.sum_exec_runtime += delta_exec;
613 account_group_exec_runtime(curr, delta_exec);
615 curr->se.exec_start = rq->clock;
616 cpuacct_charge(curr, delta_exec);
618 sched_rt_avg_update(rq, delta_exec);
620 if (!rt_bandwidth_enabled())
623 for_each_sched_rt_entity(rt_se) {
624 rt_rq = rt_rq_of_se(rt_se);
626 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
627 raw_spin_lock(&rt_rq->rt_runtime_lock);
628 rt_rq->rt_time += delta_exec;
629 if (sched_rt_runtime_exceeded(rt_rq))
631 raw_spin_unlock(&rt_rq->rt_runtime_lock);
636 #if defined CONFIG_SMP
638 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
640 static inline int next_prio(struct rq *rq)
642 struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
644 if (next && rt_prio(next->prio))
651 inc_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 (prio < prev_prio) {
658 * If the new task is higher in priority than anything on the
659 * run-queue, we know that the previous high becomes our
662 rt_rq->highest_prio.next = prev_prio;
665 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
667 } else if (prio == rt_rq->highest_prio.curr)
669 * If the next task is equal in priority to the highest on
670 * the run-queue, then we implicitly know that the next highest
671 * task cannot be any lower than current
673 rt_rq->highest_prio.next = prio;
674 else if (prio < rt_rq->highest_prio.next)
676 * Otherwise, we need to recompute next-highest
678 rt_rq->highest_prio.next = next_prio(rq);
682 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
684 struct rq *rq = rq_of_rt_rq(rt_rq);
686 if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
687 rt_rq->highest_prio.next = next_prio(rq);
689 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
690 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
693 #else /* CONFIG_SMP */
696 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
698 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
700 #endif /* CONFIG_SMP */
702 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
704 inc_rt_prio(struct rt_rq *rt_rq, int prio)
706 int prev_prio = rt_rq->highest_prio.curr;
708 if (prio < prev_prio)
709 rt_rq->highest_prio.curr = prio;
711 inc_rt_prio_smp(rt_rq, prio, prev_prio);
715 dec_rt_prio(struct rt_rq *rt_rq, int prio)
717 int prev_prio = rt_rq->highest_prio.curr;
719 if (rt_rq->rt_nr_running) {
721 WARN_ON(prio < prev_prio);
724 * This may have been our highest task, and therefore
725 * we may have some recomputation to do
727 if (prio == prev_prio) {
728 struct rt_prio_array *array = &rt_rq->active;
730 rt_rq->highest_prio.curr =
731 sched_find_first_bit(array->bitmap);
735 rt_rq->highest_prio.curr = MAX_RT_PRIO;
737 dec_rt_prio_smp(rt_rq, prio, prev_prio);
742 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
743 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
745 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
747 #ifdef CONFIG_RT_GROUP_SCHED
750 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
752 if (rt_se_boosted(rt_se))
753 rt_rq->rt_nr_boosted++;
756 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
760 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
762 if (rt_se_boosted(rt_se))
763 rt_rq->rt_nr_boosted--;
765 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
768 #else /* CONFIG_RT_GROUP_SCHED */
771 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
773 start_rt_bandwidth(&def_rt_bandwidth);
777 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
779 #endif /* CONFIG_RT_GROUP_SCHED */
782 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
784 int prio = rt_se_prio(rt_se);
786 WARN_ON(!rt_prio(prio));
787 rt_rq->rt_nr_running++;
789 inc_rt_prio(rt_rq, prio);
790 inc_rt_migration(rt_se, rt_rq);
791 inc_rt_group(rt_se, rt_rq);
795 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
797 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
798 WARN_ON(!rt_rq->rt_nr_running);
799 rt_rq->rt_nr_running--;
801 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
802 dec_rt_migration(rt_se, rt_rq);
803 dec_rt_group(rt_se, rt_rq);
806 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
808 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
809 struct rt_prio_array *array = &rt_rq->active;
810 struct rt_rq *group_rq = group_rt_rq(rt_se);
811 struct list_head *queue = array->queue + rt_se_prio(rt_se);
814 * Don't enqueue the group if its throttled, or when empty.
815 * The latter is a consequence of the former when a child group
816 * get throttled and the current group doesn't have any other
819 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
823 list_add(&rt_se->run_list, queue);
825 list_add_tail(&rt_se->run_list, queue);
826 __set_bit(rt_se_prio(rt_se), array->bitmap);
828 inc_rt_tasks(rt_se, rt_rq);
831 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
833 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
834 struct rt_prio_array *array = &rt_rq->active;
836 list_del_init(&rt_se->run_list);
837 if (list_empty(array->queue + rt_se_prio(rt_se)))
838 __clear_bit(rt_se_prio(rt_se), array->bitmap);
840 dec_rt_tasks(rt_se, rt_rq);
844 * Because the prio of an upper entry depends on the lower
845 * entries, we must remove entries top - down.
847 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
849 struct sched_rt_entity *back = NULL;
851 for_each_sched_rt_entity(rt_se) {
856 for (rt_se = back; rt_se; rt_se = rt_se->back) {
858 __dequeue_rt_entity(rt_se);
862 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
864 dequeue_rt_stack(rt_se);
865 for_each_sched_rt_entity(rt_se)
866 __enqueue_rt_entity(rt_se, head);
869 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
871 dequeue_rt_stack(rt_se);
873 for_each_sched_rt_entity(rt_se) {
874 struct rt_rq *rt_rq = group_rt_rq(rt_se);
876 if (rt_rq && rt_rq->rt_nr_running)
877 __enqueue_rt_entity(rt_se, false);
882 * Adding/removing a task to/from a priority array:
885 enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup, bool head)
887 struct sched_rt_entity *rt_se = &p->rt;
892 enqueue_rt_entity(rt_se, head);
894 if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
895 enqueue_pushable_task(rq, p);
898 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
900 struct sched_rt_entity *rt_se = &p->rt;
903 dequeue_rt_entity(rt_se);
905 dequeue_pushable_task(rq, p);
909 * Put task to the end of the run list without the overhead of dequeue
910 * followed by enqueue.
913 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
915 if (on_rt_rq(rt_se)) {
916 struct rt_prio_array *array = &rt_rq->active;
917 struct list_head *queue = array->queue + rt_se_prio(rt_se);
920 list_move(&rt_se->run_list, queue);
922 list_move_tail(&rt_se->run_list, queue);
926 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
928 struct sched_rt_entity *rt_se = &p->rt;
931 for_each_sched_rt_entity(rt_se) {
932 rt_rq = rt_rq_of_se(rt_se);
933 requeue_rt_entity(rt_rq, rt_se, head);
937 static void yield_task_rt(struct rq *rq)
939 requeue_task_rt(rq, rq->curr, 0);
943 static int find_lowest_rq(struct task_struct *task);
945 static int select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
947 struct rq *rq = task_rq(p);
949 if (sd_flag != SD_BALANCE_WAKE)
950 return smp_processor_id();
953 * If the current task is an RT task, then
954 * try to see if we can wake this RT task up on another
955 * runqueue. Otherwise simply start this RT task
956 * on its current runqueue.
958 * We want to avoid overloading runqueues. Even if
959 * the RT task is of higher priority than the current RT task.
960 * RT tasks behave differently than other tasks. If
961 * one gets preempted, we try to push it off to another queue.
962 * So trying to keep a preempting RT task on the same
963 * cache hot CPU will force the running RT task to
964 * a cold CPU. So we waste all the cache for the lower
965 * RT task in hopes of saving some of a RT task
966 * that is just being woken and probably will have
969 if (unlikely(rt_task(rq->curr)) &&
970 (p->rt.nr_cpus_allowed > 1)) {
971 int cpu = find_lowest_rq(p);
973 return (cpu == -1) ? task_cpu(p) : cpu;
977 * Otherwise, just let it ride on the affined RQ and the
978 * post-schedule router will push the preempted task away
983 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
985 if (rq->curr->rt.nr_cpus_allowed == 1)
988 if (p->rt.nr_cpus_allowed != 1
989 && cpupri_find(&rq->rd->cpupri, p, NULL))
992 if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
996 * There appears to be other cpus that can accept
997 * current and none to run 'p', so lets reschedule
998 * to try and push current away:
1000 requeue_task_rt(rq, p, 1);
1001 resched_task(rq->curr);
1004 #endif /* CONFIG_SMP */
1007 * Preempt the current task with a newly woken task if needed:
1009 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1011 if (p->prio < rq->curr->prio) {
1012 resched_task(rq->curr);
1020 * - the newly woken task is of equal priority to the current task
1021 * - the newly woken task is non-migratable while current is migratable
1022 * - current will be preempted on the next reschedule
1024 * we should check to see if current can readily move to a different
1025 * cpu. If so, we will reschedule to allow the push logic to try
1026 * to move current somewhere else, making room for our non-migratable
1029 if (p->prio == rq->curr->prio && !need_resched())
1030 check_preempt_equal_prio(rq, p);
1034 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1035 struct rt_rq *rt_rq)
1037 struct rt_prio_array *array = &rt_rq->active;
1038 struct sched_rt_entity *next = NULL;
1039 struct list_head *queue;
1042 idx = sched_find_first_bit(array->bitmap);
1043 BUG_ON(idx >= MAX_RT_PRIO);
1045 queue = array->queue + idx;
1046 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1051 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1053 struct sched_rt_entity *rt_se;
1054 struct task_struct *p;
1055 struct rt_rq *rt_rq;
1059 if (unlikely(!rt_rq->rt_nr_running))
1062 if (rt_rq_throttled(rt_rq))
1066 rt_se = pick_next_rt_entity(rq, rt_rq);
1068 rt_rq = group_rt_rq(rt_se);
1071 p = rt_task_of(rt_se);
1072 p->se.exec_start = rq->clock;
1077 static struct task_struct *pick_next_task_rt(struct rq *rq)
1079 struct task_struct *p = _pick_next_task_rt(rq);
1081 /* The running task is never eligible for pushing */
1083 dequeue_pushable_task(rq, p);
1087 * We detect this state here so that we can avoid taking the RQ
1088 * lock again later if there is no need to push
1090 rq->post_schedule = has_pushable_tasks(rq);
1096 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1099 p->se.exec_start = 0;
1102 * The previous task needs to be made eligible for pushing
1103 * if it is still active
1105 if (p->se.on_rq && p->rt.nr_cpus_allowed > 1)
1106 enqueue_pushable_task(rq, p);
1111 /* Only try algorithms three times */
1112 #define RT_MAX_TRIES 3
1114 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1116 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1118 if (!task_running(rq, p) &&
1119 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1120 (p->rt.nr_cpus_allowed > 1))
1125 /* Return the second highest RT task, NULL otherwise */
1126 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1128 struct task_struct *next = NULL;
1129 struct sched_rt_entity *rt_se;
1130 struct rt_prio_array *array;
1131 struct rt_rq *rt_rq;
1134 for_each_leaf_rt_rq(rt_rq, rq) {
1135 array = &rt_rq->active;
1136 idx = sched_find_first_bit(array->bitmap);
1138 if (idx >= MAX_RT_PRIO)
1140 if (next && next->prio < idx)
1142 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1143 struct task_struct *p = rt_task_of(rt_se);
1144 if (pick_rt_task(rq, p, cpu)) {
1150 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1158 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1160 static int find_lowest_rq(struct task_struct *task)
1162 struct sched_domain *sd;
1163 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1164 int this_cpu = smp_processor_id();
1165 int cpu = task_cpu(task);
1167 if (task->rt.nr_cpus_allowed == 1)
1168 return -1; /* No other targets possible */
1170 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1171 return -1; /* No targets found */
1174 * At this point we have built a mask of cpus representing the
1175 * lowest priority tasks in the system. Now we want to elect
1176 * the best one based on our affinity and topology.
1178 * We prioritize the last cpu that the task executed on since
1179 * it is most likely cache-hot in that location.
1181 if (cpumask_test_cpu(cpu, lowest_mask))
1185 * Otherwise, we consult the sched_domains span maps to figure
1186 * out which cpu is logically closest to our hot cache data.
1188 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1189 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1191 for_each_domain(cpu, sd) {
1192 if (sd->flags & SD_WAKE_AFFINE) {
1196 * "this_cpu" is cheaper to preempt than a
1199 if (this_cpu != -1 &&
1200 cpumask_test_cpu(this_cpu, sched_domain_span(sd)))
1203 best_cpu = cpumask_first_and(lowest_mask,
1204 sched_domain_span(sd));
1205 if (best_cpu < nr_cpu_ids)
1211 * And finally, if there were no matches within the domains
1212 * just give the caller *something* to work with from the compatible
1218 cpu = cpumask_any(lowest_mask);
1219 if (cpu < nr_cpu_ids)
1224 /* Will lock the rq it finds */
1225 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1227 struct rq *lowest_rq = NULL;
1231 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1232 cpu = find_lowest_rq(task);
1234 if ((cpu == -1) || (cpu == rq->cpu))
1237 lowest_rq = cpu_rq(cpu);
1239 /* if the prio of this runqueue changed, try again */
1240 if (double_lock_balance(rq, lowest_rq)) {
1242 * We had to unlock the run queue. In
1243 * the mean time, task could have
1244 * migrated already or had its affinity changed.
1245 * Also make sure that it wasn't scheduled on its rq.
1247 if (unlikely(task_rq(task) != rq ||
1248 !cpumask_test_cpu(lowest_rq->cpu,
1249 &task->cpus_allowed) ||
1250 task_running(rq, task) ||
1253 raw_spin_unlock(&lowest_rq->lock);
1259 /* If this rq is still suitable use it. */
1260 if (lowest_rq->rt.highest_prio.curr > task->prio)
1264 double_unlock_balance(rq, lowest_rq);
1271 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1273 struct task_struct *p;
1275 if (!has_pushable_tasks(rq))
1278 p = plist_first_entry(&rq->rt.pushable_tasks,
1279 struct task_struct, pushable_tasks);
1281 BUG_ON(rq->cpu != task_cpu(p));
1282 BUG_ON(task_current(rq, p));
1283 BUG_ON(p->rt.nr_cpus_allowed <= 1);
1285 BUG_ON(!p->se.on_rq);
1286 BUG_ON(!rt_task(p));
1292 * If the current CPU has more than one RT task, see if the non
1293 * running task can migrate over to a CPU that is running a task
1294 * of lesser priority.
1296 static int push_rt_task(struct rq *rq)
1298 struct task_struct *next_task;
1299 struct rq *lowest_rq;
1301 if (!rq->rt.overloaded)
1304 next_task = pick_next_pushable_task(rq);
1309 if (unlikely(next_task == rq->curr)) {
1315 * It's possible that the next_task slipped in of
1316 * higher priority than current. If that's the case
1317 * just reschedule current.
1319 if (unlikely(next_task->prio < rq->curr->prio)) {
1320 resched_task(rq->curr);
1324 /* We might release rq lock */
1325 get_task_struct(next_task);
1327 /* find_lock_lowest_rq locks the rq if found */
1328 lowest_rq = find_lock_lowest_rq(next_task, rq);
1330 struct task_struct *task;
1332 * find lock_lowest_rq releases rq->lock
1333 * so it is possible that next_task has migrated.
1335 * We need to make sure that the task is still on the same
1336 * run-queue and is also still the next task eligible for
1339 task = pick_next_pushable_task(rq);
1340 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1342 * If we get here, the task hasnt moved at all, but
1343 * it has failed to push. We will not try again,
1344 * since the other cpus will pull from us when they
1347 dequeue_pushable_task(rq, next_task);
1352 /* No more tasks, just exit */
1356 * Something has shifted, try again.
1358 put_task_struct(next_task);
1363 deactivate_task(rq, next_task, 0);
1364 set_task_cpu(next_task, lowest_rq->cpu);
1365 activate_task(lowest_rq, next_task, 0);
1367 resched_task(lowest_rq->curr);
1369 double_unlock_balance(rq, lowest_rq);
1372 put_task_struct(next_task);
1377 static void push_rt_tasks(struct rq *rq)
1379 /* push_rt_task will return true if it moved an RT */
1380 while (push_rt_task(rq))
1384 static int pull_rt_task(struct rq *this_rq)
1386 int this_cpu = this_rq->cpu, ret = 0, cpu;
1387 struct task_struct *p;
1390 if (likely(!rt_overloaded(this_rq)))
1393 for_each_cpu(cpu, this_rq->rd->rto_mask) {
1394 if (this_cpu == cpu)
1397 src_rq = cpu_rq(cpu);
1400 * Don't bother taking the src_rq->lock if the next highest
1401 * task is known to be lower-priority than our current task.
1402 * This may look racy, but if this value is about to go
1403 * logically higher, the src_rq will push this task away.
1404 * And if its going logically lower, we do not care
1406 if (src_rq->rt.highest_prio.next >=
1407 this_rq->rt.highest_prio.curr)
1411 * We can potentially drop this_rq's lock in
1412 * double_lock_balance, and another CPU could
1415 double_lock_balance(this_rq, src_rq);
1418 * Are there still pullable RT tasks?
1420 if (src_rq->rt.rt_nr_running <= 1)
1423 p = pick_next_highest_task_rt(src_rq, this_cpu);
1426 * Do we have an RT task that preempts
1427 * the to-be-scheduled task?
1429 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1430 WARN_ON(p == src_rq->curr);
1431 WARN_ON(!p->se.on_rq);
1434 * There's a chance that p is higher in priority
1435 * than what's currently running on its cpu.
1436 * This is just that p is wakeing up and hasn't
1437 * had a chance to schedule. We only pull
1438 * p if it is lower in priority than the
1439 * current task on the run queue
1441 if (p->prio < src_rq->curr->prio)
1446 deactivate_task(src_rq, p, 0);
1447 set_task_cpu(p, this_cpu);
1448 activate_task(this_rq, p, 0);
1450 * We continue with the search, just in
1451 * case there's an even higher prio task
1452 * in another runqueue. (low likelyhood
1457 double_unlock_balance(this_rq, src_rq);
1463 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1465 /* Try to pull RT tasks here if we lower this rq's prio */
1466 if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1470 static void post_schedule_rt(struct rq *rq)
1476 * If we are not running and we are not going to reschedule soon, we should
1477 * try to push tasks away now
1479 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1481 if (!task_running(rq, p) &&
1482 !test_tsk_need_resched(rq->curr) &&
1483 has_pushable_tasks(rq) &&
1484 p->rt.nr_cpus_allowed > 1)
1488 static void set_cpus_allowed_rt(struct task_struct *p,
1489 const struct cpumask *new_mask)
1491 int weight = cpumask_weight(new_mask);
1493 BUG_ON(!rt_task(p));
1496 * Update the migration status of the RQ if we have an RT task
1497 * which is running AND changing its weight value.
1499 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1500 struct rq *rq = task_rq(p);
1502 if (!task_current(rq, p)) {
1504 * Make sure we dequeue this task from the pushable list
1505 * before going further. It will either remain off of
1506 * the list because we are no longer pushable, or it
1509 if (p->rt.nr_cpus_allowed > 1)
1510 dequeue_pushable_task(rq, p);
1513 * Requeue if our weight is changing and still > 1
1516 enqueue_pushable_task(rq, p);
1520 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1521 rq->rt.rt_nr_migratory++;
1522 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1523 BUG_ON(!rq->rt.rt_nr_migratory);
1524 rq->rt.rt_nr_migratory--;
1527 update_rt_migration(&rq->rt);
1530 cpumask_copy(&p->cpus_allowed, new_mask);
1531 p->rt.nr_cpus_allowed = weight;
1534 /* Assumes rq->lock is held */
1535 static void rq_online_rt(struct rq *rq)
1537 if (rq->rt.overloaded)
1538 rt_set_overload(rq);
1540 __enable_runtime(rq);
1542 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1545 /* Assumes rq->lock is held */
1546 static void rq_offline_rt(struct rq *rq)
1548 if (rq->rt.overloaded)
1549 rt_clear_overload(rq);
1551 __disable_runtime(rq);
1553 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1557 * When switch from the rt queue, we bring ourselves to a position
1558 * that we might want to pull RT tasks from other runqueues.
1560 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1564 * If there are other RT tasks then we will reschedule
1565 * and the scheduling of the other RT tasks will handle
1566 * the balancing. But if we are the last RT task
1567 * we may need to handle the pulling of RT tasks
1570 if (!rq->rt.rt_nr_running)
1574 static inline void init_sched_rt_class(void)
1578 for_each_possible_cpu(i)
1579 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1580 GFP_KERNEL, cpu_to_node(i));
1582 #endif /* CONFIG_SMP */
1585 * When switching a task to RT, we may overload the runqueue
1586 * with RT tasks. In this case we try to push them off to
1589 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1592 int check_resched = 1;
1595 * If we are already running, then there's nothing
1596 * that needs to be done. But if we are not running
1597 * we may need to preempt the current running task.
1598 * If that current running task is also an RT task
1599 * then see if we can move to another run queue.
1603 if (rq->rt.overloaded && push_rt_task(rq) &&
1604 /* Don't resched if we changed runqueues */
1607 #endif /* CONFIG_SMP */
1608 if (check_resched && p->prio < rq->curr->prio)
1609 resched_task(rq->curr);
1614 * Priority of the task has changed. This may cause
1615 * us to initiate a push or pull.
1617 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1618 int oldprio, int running)
1623 * If our priority decreases while running, we
1624 * may need to pull tasks to this runqueue.
1626 if (oldprio < p->prio)
1629 * If there's a higher priority task waiting to run
1630 * then reschedule. Note, the above pull_rt_task
1631 * can release the rq lock and p could migrate.
1632 * Only reschedule if p is still on the same runqueue.
1634 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1637 /* For UP simply resched on drop of prio */
1638 if (oldprio < p->prio)
1640 #endif /* CONFIG_SMP */
1643 * This task is not running, but if it is
1644 * greater than the current running task
1647 if (p->prio < rq->curr->prio)
1648 resched_task(rq->curr);
1652 static void watchdog(struct rq *rq, struct task_struct *p)
1654 unsigned long soft, hard;
1659 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1660 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1662 if (soft != RLIM_INFINITY) {
1666 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1667 if (p->rt.timeout > next)
1668 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1672 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1679 * RR tasks need a special form of timeslice management.
1680 * FIFO tasks have no timeslices.
1682 if (p->policy != SCHED_RR)
1685 if (--p->rt.time_slice)
1688 p->rt.time_slice = DEF_TIMESLICE;
1691 * Requeue to the end of queue if we are not the only element
1694 if (p->rt.run_list.prev != p->rt.run_list.next) {
1695 requeue_task_rt(rq, p, 0);
1696 set_tsk_need_resched(p);
1700 static void set_curr_task_rt(struct rq *rq)
1702 struct task_struct *p = rq->curr;
1704 p->se.exec_start = rq->clock;
1706 /* The running task is never eligible for pushing */
1707 dequeue_pushable_task(rq, p);
1710 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1713 * Time slice is 0 for SCHED_FIFO tasks
1715 if (task->policy == SCHED_RR)
1716 return DEF_TIMESLICE;
1721 static const struct sched_class rt_sched_class = {
1722 .next = &fair_sched_class,
1723 .enqueue_task = enqueue_task_rt,
1724 .dequeue_task = dequeue_task_rt,
1725 .yield_task = yield_task_rt,
1727 .check_preempt_curr = check_preempt_curr_rt,
1729 .pick_next_task = pick_next_task_rt,
1730 .put_prev_task = put_prev_task_rt,
1733 .select_task_rq = select_task_rq_rt,
1735 .set_cpus_allowed = set_cpus_allowed_rt,
1736 .rq_online = rq_online_rt,
1737 .rq_offline = rq_offline_rt,
1738 .pre_schedule = pre_schedule_rt,
1739 .post_schedule = post_schedule_rt,
1740 .task_woken = task_woken_rt,
1741 .switched_from = switched_from_rt,
1744 .set_curr_task = set_curr_task_rt,
1745 .task_tick = task_tick_rt,
1747 .get_rr_interval = get_rr_interval_rt,
1749 .prio_changed = prio_changed_rt,
1750 .switched_to = switched_to_rt,
1753 #ifdef CONFIG_SCHED_DEBUG
1754 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1756 static void print_rt_stats(struct seq_file *m, int cpu)
1758 struct rt_rq *rt_rq;
1761 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1762 print_rt_rq(m, cpu, rt_rq);
1765 #endif /* CONFIG_SCHED_DEBUG */