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 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
188 list_add_rcu(&rt_rq->leaf_rt_rq_list,
189 &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
192 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
194 list_del_rcu(&rt_rq->leaf_rt_rq_list);
197 #define for_each_leaf_rt_rq(rt_rq, rq) \
198 list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
200 #define for_each_sched_rt_entity(rt_se) \
201 for (; rt_se; rt_se = rt_se->parent)
203 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
208 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
209 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
211 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
213 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
214 struct sched_rt_entity *rt_se;
216 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
218 rt_se = rt_rq->tg->rt_se[cpu];
220 if (rt_rq->rt_nr_running) {
221 if (rt_se && !on_rt_rq(rt_se))
222 enqueue_rt_entity(rt_se, false);
223 if (rt_rq->highest_prio.curr < curr->prio)
228 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
230 struct sched_rt_entity *rt_se;
231 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
233 rt_se = rt_rq->tg->rt_se[cpu];
235 if (rt_se && on_rt_rq(rt_se))
236 dequeue_rt_entity(rt_se);
239 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
241 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
244 static int rt_se_boosted(struct sched_rt_entity *rt_se)
246 struct rt_rq *rt_rq = group_rt_rq(rt_se);
247 struct task_struct *p;
250 return !!rt_rq->rt_nr_boosted;
252 p = rt_task_of(rt_se);
253 return p->prio != p->normal_prio;
257 static inline const struct cpumask *sched_rt_period_mask(void)
259 return cpu_rq(smp_processor_id())->rd->span;
262 static inline const struct cpumask *sched_rt_period_mask(void)
264 return cpu_online_mask;
269 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
271 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
274 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
276 return &rt_rq->tg->rt_bandwidth;
279 #else /* !CONFIG_RT_GROUP_SCHED */
281 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
283 return rt_rq->rt_runtime;
286 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
288 return ktime_to_ns(def_rt_bandwidth.rt_period);
291 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
295 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
299 #define for_each_leaf_rt_rq(rt_rq, rq) \
300 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
302 #define for_each_sched_rt_entity(rt_se) \
303 for (; rt_se; rt_se = NULL)
305 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
310 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
312 if (rt_rq->rt_nr_running)
313 resched_task(rq_of_rt_rq(rt_rq)->curr);
316 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
320 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
322 return rt_rq->rt_throttled;
325 static inline const struct cpumask *sched_rt_period_mask(void)
327 return cpu_online_mask;
331 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
333 return &cpu_rq(cpu)->rt;
336 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
338 return &def_rt_bandwidth;
341 #endif /* CONFIG_RT_GROUP_SCHED */
345 * We ran out of runtime, see if we can borrow some from our neighbours.
347 static int do_balance_runtime(struct rt_rq *rt_rq)
349 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
350 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
351 int i, weight, more = 0;
354 weight = cpumask_weight(rd->span);
356 raw_spin_lock(&rt_b->rt_runtime_lock);
357 rt_period = ktime_to_ns(rt_b->rt_period);
358 for_each_cpu(i, rd->span) {
359 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
365 raw_spin_lock(&iter->rt_runtime_lock);
367 * Either all rqs have inf runtime and there's nothing to steal
368 * or __disable_runtime() below sets a specific rq to inf to
369 * indicate its been disabled and disalow stealing.
371 if (iter->rt_runtime == RUNTIME_INF)
375 * From runqueues with spare time, take 1/n part of their
376 * spare time, but no more than our period.
378 diff = iter->rt_runtime - iter->rt_time;
380 diff = div_u64((u64)diff, weight);
381 if (rt_rq->rt_runtime + diff > rt_period)
382 diff = rt_period - rt_rq->rt_runtime;
383 iter->rt_runtime -= diff;
384 rt_rq->rt_runtime += diff;
386 if (rt_rq->rt_runtime == rt_period) {
387 raw_spin_unlock(&iter->rt_runtime_lock);
392 raw_spin_unlock(&iter->rt_runtime_lock);
394 raw_spin_unlock(&rt_b->rt_runtime_lock);
400 * Ensure this RQ takes back all the runtime it lend to its neighbours.
402 static void __disable_runtime(struct rq *rq)
404 struct root_domain *rd = rq->rd;
407 if (unlikely(!scheduler_running))
410 for_each_leaf_rt_rq(rt_rq, rq) {
411 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
415 raw_spin_lock(&rt_b->rt_runtime_lock);
416 raw_spin_lock(&rt_rq->rt_runtime_lock);
418 * Either we're all inf and nobody needs to borrow, or we're
419 * already disabled and thus have nothing to do, or we have
420 * exactly the right amount of runtime to take out.
422 if (rt_rq->rt_runtime == RUNTIME_INF ||
423 rt_rq->rt_runtime == rt_b->rt_runtime)
425 raw_spin_unlock(&rt_rq->rt_runtime_lock);
428 * Calculate the difference between what we started out with
429 * and what we current have, that's the amount of runtime
430 * we lend and now have to reclaim.
432 want = rt_b->rt_runtime - rt_rq->rt_runtime;
435 * Greedy reclaim, take back as much as we can.
437 for_each_cpu(i, rd->span) {
438 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
442 * Can't reclaim from ourselves or disabled runqueues.
444 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
447 raw_spin_lock(&iter->rt_runtime_lock);
449 diff = min_t(s64, iter->rt_runtime, want);
450 iter->rt_runtime -= diff;
453 iter->rt_runtime -= want;
456 raw_spin_unlock(&iter->rt_runtime_lock);
462 raw_spin_lock(&rt_rq->rt_runtime_lock);
464 * We cannot be left wanting - that would mean some runtime
465 * leaked out of the system.
470 * Disable all the borrow logic by pretending we have inf
471 * runtime - in which case borrowing doesn't make sense.
473 rt_rq->rt_runtime = RUNTIME_INF;
474 raw_spin_unlock(&rt_rq->rt_runtime_lock);
475 raw_spin_unlock(&rt_b->rt_runtime_lock);
479 static void disable_runtime(struct rq *rq)
483 raw_spin_lock_irqsave(&rq->lock, flags);
484 __disable_runtime(rq);
485 raw_spin_unlock_irqrestore(&rq->lock, flags);
488 static void __enable_runtime(struct rq *rq)
492 if (unlikely(!scheduler_running))
496 * Reset each runqueue's bandwidth settings
498 for_each_leaf_rt_rq(rt_rq, rq) {
499 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
501 raw_spin_lock(&rt_b->rt_runtime_lock);
502 raw_spin_lock(&rt_rq->rt_runtime_lock);
503 rt_rq->rt_runtime = rt_b->rt_runtime;
505 rt_rq->rt_throttled = 0;
506 raw_spin_unlock(&rt_rq->rt_runtime_lock);
507 raw_spin_unlock(&rt_b->rt_runtime_lock);
511 static void enable_runtime(struct rq *rq)
515 raw_spin_lock_irqsave(&rq->lock, flags);
516 __enable_runtime(rq);
517 raw_spin_unlock_irqrestore(&rq->lock, flags);
520 static int balance_runtime(struct rt_rq *rt_rq)
524 if (rt_rq->rt_time > rt_rq->rt_runtime) {
525 raw_spin_unlock(&rt_rq->rt_runtime_lock);
526 more = do_balance_runtime(rt_rq);
527 raw_spin_lock(&rt_rq->rt_runtime_lock);
532 #else /* !CONFIG_SMP */
533 static inline int balance_runtime(struct rt_rq *rt_rq)
537 #endif /* CONFIG_SMP */
539 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
542 const struct cpumask *span;
544 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
547 span = sched_rt_period_mask();
548 for_each_cpu(i, span) {
550 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
551 struct rq *rq = rq_of_rt_rq(rt_rq);
553 raw_spin_lock(&rq->lock);
554 if (rt_rq->rt_time) {
557 raw_spin_lock(&rt_rq->rt_runtime_lock);
558 if (rt_rq->rt_throttled)
559 balance_runtime(rt_rq);
560 runtime = rt_rq->rt_runtime;
561 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
562 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
563 rt_rq->rt_throttled = 0;
567 * Force a clock update if the CPU was idle,
568 * lest wakeup -> unthrottle time accumulate.
570 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
571 rq->skip_clock_update = -1;
573 if (rt_rq->rt_time || rt_rq->rt_nr_running)
575 raw_spin_unlock(&rt_rq->rt_runtime_lock);
576 } else if (rt_rq->rt_nr_running) {
578 if (!rt_rq_throttled(rt_rq))
583 sched_rt_rq_enqueue(rt_rq);
584 raw_spin_unlock(&rq->lock);
590 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
592 #ifdef CONFIG_RT_GROUP_SCHED
593 struct rt_rq *rt_rq = group_rt_rq(rt_se);
596 return rt_rq->highest_prio.curr;
599 return rt_task_of(rt_se)->prio;
602 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
604 u64 runtime = sched_rt_runtime(rt_rq);
606 if (rt_rq->rt_throttled)
607 return rt_rq_throttled(rt_rq);
609 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
612 balance_runtime(rt_rq);
613 runtime = sched_rt_runtime(rt_rq);
614 if (runtime == RUNTIME_INF)
617 if (rt_rq->rt_time > runtime) {
618 rt_rq->rt_throttled = 1;
619 if (rt_rq_throttled(rt_rq)) {
620 sched_rt_rq_dequeue(rt_rq);
629 * Update the current task's runtime statistics. Skip current tasks that
630 * are not in our scheduling class.
632 static void update_curr_rt(struct rq *rq)
634 struct task_struct *curr = rq->curr;
635 struct sched_rt_entity *rt_se = &curr->rt;
636 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
639 if (curr->sched_class != &rt_sched_class)
642 delta_exec = rq->clock_task - curr->se.exec_start;
643 if (unlikely((s64)delta_exec < 0))
646 schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
648 curr->se.sum_exec_runtime += delta_exec;
649 account_group_exec_runtime(curr, delta_exec);
651 curr->se.exec_start = rq->clock_task;
652 cpuacct_charge(curr, delta_exec);
654 sched_rt_avg_update(rq, delta_exec);
656 if (!rt_bandwidth_enabled())
659 for_each_sched_rt_entity(rt_se) {
660 rt_rq = rt_rq_of_se(rt_se);
662 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
663 raw_spin_lock(&rt_rq->rt_runtime_lock);
664 rt_rq->rt_time += delta_exec;
665 if (sched_rt_runtime_exceeded(rt_rq))
667 raw_spin_unlock(&rt_rq->rt_runtime_lock);
672 #if defined CONFIG_SMP
674 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
676 static inline int next_prio(struct rq *rq)
678 struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
680 if (next && rt_prio(next->prio))
687 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
689 struct rq *rq = rq_of_rt_rq(rt_rq);
691 if (prio < prev_prio) {
694 * If the new task is higher in priority than anything on the
695 * run-queue, we know that the previous high becomes our
698 rt_rq->highest_prio.next = prev_prio;
701 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
703 } else if (prio == rt_rq->highest_prio.curr)
705 * If the next task is equal in priority to the highest on
706 * the run-queue, then we implicitly know that the next highest
707 * task cannot be any lower than current
709 rt_rq->highest_prio.next = prio;
710 else if (prio < rt_rq->highest_prio.next)
712 * Otherwise, we need to recompute next-highest
714 rt_rq->highest_prio.next = next_prio(rq);
718 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
720 struct rq *rq = rq_of_rt_rq(rt_rq);
722 if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
723 rt_rq->highest_prio.next = next_prio(rq);
725 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
726 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
729 #else /* CONFIG_SMP */
732 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
734 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
736 #endif /* CONFIG_SMP */
738 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
740 inc_rt_prio(struct rt_rq *rt_rq, int prio)
742 int prev_prio = rt_rq->highest_prio.curr;
744 if (prio < prev_prio)
745 rt_rq->highest_prio.curr = prio;
747 inc_rt_prio_smp(rt_rq, prio, prev_prio);
751 dec_rt_prio(struct rt_rq *rt_rq, int prio)
753 int prev_prio = rt_rq->highest_prio.curr;
755 if (rt_rq->rt_nr_running) {
757 WARN_ON(prio < prev_prio);
760 * This may have been our highest task, and therefore
761 * we may have some recomputation to do
763 if (prio == prev_prio) {
764 struct rt_prio_array *array = &rt_rq->active;
766 rt_rq->highest_prio.curr =
767 sched_find_first_bit(array->bitmap);
771 rt_rq->highest_prio.curr = MAX_RT_PRIO;
773 dec_rt_prio_smp(rt_rq, prio, prev_prio);
778 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
779 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
781 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
783 #ifdef CONFIG_RT_GROUP_SCHED
786 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
788 if (rt_se_boosted(rt_se))
789 rt_rq->rt_nr_boosted++;
792 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
796 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
798 if (rt_se_boosted(rt_se))
799 rt_rq->rt_nr_boosted--;
801 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
804 #else /* CONFIG_RT_GROUP_SCHED */
807 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
809 start_rt_bandwidth(&def_rt_bandwidth);
813 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
815 #endif /* CONFIG_RT_GROUP_SCHED */
818 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
820 int prio = rt_se_prio(rt_se);
822 WARN_ON(!rt_prio(prio));
823 rt_rq->rt_nr_running++;
825 inc_rt_prio(rt_rq, prio);
826 inc_rt_migration(rt_se, rt_rq);
827 inc_rt_group(rt_se, rt_rq);
831 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
833 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
834 WARN_ON(!rt_rq->rt_nr_running);
835 rt_rq->rt_nr_running--;
837 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
838 dec_rt_migration(rt_se, rt_rq);
839 dec_rt_group(rt_se, rt_rq);
842 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
844 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
845 struct rt_prio_array *array = &rt_rq->active;
846 struct rt_rq *group_rq = group_rt_rq(rt_se);
847 struct list_head *queue = array->queue + rt_se_prio(rt_se);
850 * Don't enqueue the group if its throttled, or when empty.
851 * The latter is a consequence of the former when a child group
852 * get throttled and the current group doesn't have any other
855 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
858 if (!rt_rq->rt_nr_running)
859 list_add_leaf_rt_rq(rt_rq);
862 list_add(&rt_se->run_list, queue);
864 list_add_tail(&rt_se->run_list, queue);
865 __set_bit(rt_se_prio(rt_se), array->bitmap);
867 inc_rt_tasks(rt_se, rt_rq);
870 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
872 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
873 struct rt_prio_array *array = &rt_rq->active;
875 list_del_init(&rt_se->run_list);
876 if (list_empty(array->queue + rt_se_prio(rt_se)))
877 __clear_bit(rt_se_prio(rt_se), array->bitmap);
879 dec_rt_tasks(rt_se, rt_rq);
880 if (!rt_rq->rt_nr_running)
881 list_del_leaf_rt_rq(rt_rq);
885 * Because the prio of an upper entry depends on the lower
886 * entries, we must remove entries top - down.
888 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
890 struct sched_rt_entity *back = NULL;
892 for_each_sched_rt_entity(rt_se) {
897 for (rt_se = back; rt_se; rt_se = rt_se->back) {
899 __dequeue_rt_entity(rt_se);
903 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
905 dequeue_rt_stack(rt_se);
906 for_each_sched_rt_entity(rt_se)
907 __enqueue_rt_entity(rt_se, head);
910 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
912 dequeue_rt_stack(rt_se);
914 for_each_sched_rt_entity(rt_se) {
915 struct rt_rq *rt_rq = group_rt_rq(rt_se);
917 if (rt_rq && rt_rq->rt_nr_running)
918 __enqueue_rt_entity(rt_se, false);
923 * Adding/removing a task to/from a priority array:
926 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
928 struct sched_rt_entity *rt_se = &p->rt;
930 if (flags & ENQUEUE_WAKEUP)
933 enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
935 if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
936 enqueue_pushable_task(rq, p);
939 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
941 struct sched_rt_entity *rt_se = &p->rt;
944 dequeue_rt_entity(rt_se);
946 dequeue_pushable_task(rq, p);
950 * Put task to the end of the run list without the overhead of dequeue
951 * followed by enqueue.
954 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
956 if (on_rt_rq(rt_se)) {
957 struct rt_prio_array *array = &rt_rq->active;
958 struct list_head *queue = array->queue + rt_se_prio(rt_se);
961 list_move(&rt_se->run_list, queue);
963 list_move_tail(&rt_se->run_list, queue);
967 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
969 struct sched_rt_entity *rt_se = &p->rt;
972 for_each_sched_rt_entity(rt_se) {
973 rt_rq = rt_rq_of_se(rt_se);
974 requeue_rt_entity(rt_rq, rt_se, head);
978 static void yield_task_rt(struct rq *rq)
980 requeue_task_rt(rq, rq->curr, 0);
984 static int find_lowest_rq(struct task_struct *task);
987 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
989 struct task_struct *curr;
993 if (sd_flag != SD_BALANCE_WAKE)
994 return smp_processor_id();
1000 curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1003 * If the current task on @p's runqueue is an RT task, then
1004 * try to see if we can wake this RT task up on another
1005 * runqueue. Otherwise simply start this RT task
1006 * on its current runqueue.
1008 * We want to avoid overloading runqueues. If the woken
1009 * task is a higher priority, then it will stay on this CPU
1010 * and the lower prio task should be moved to another CPU.
1011 * Even though this will probably make the lower prio task
1012 * lose its cache, we do not want to bounce a higher task
1013 * around just because it gave up its CPU, perhaps for a
1016 * For equal prio tasks, we just let the scheduler sort it out.
1018 * Otherwise, just let it ride on the affined RQ and the
1019 * post-schedule router will push the preempted task away
1021 * This test is optimistic, if we get it wrong the load-balancer
1022 * will have to sort it out.
1024 if (curr && unlikely(rt_task(curr)) &&
1025 (curr->rt.nr_cpus_allowed < 2 ||
1026 curr->prio < p->prio) &&
1027 (p->rt.nr_cpus_allowed > 1)) {
1028 int target = find_lowest_rq(p);
1038 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1040 if (rq->curr->rt.nr_cpus_allowed == 1)
1043 if (p->rt.nr_cpus_allowed != 1
1044 && cpupri_find(&rq->rd->cpupri, p, NULL))
1047 if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1051 * There appears to be other cpus that can accept
1052 * current and none to run 'p', so lets reschedule
1053 * to try and push current away:
1055 requeue_task_rt(rq, p, 1);
1056 resched_task(rq->curr);
1059 #endif /* CONFIG_SMP */
1062 * Preempt the current task with a newly woken task if needed:
1064 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1066 if (p->prio < rq->curr->prio) {
1067 resched_task(rq->curr);
1075 * - the newly woken task is of equal priority to the current task
1076 * - the newly woken task is non-migratable while current is migratable
1077 * - current will be preempted on the next reschedule
1079 * we should check to see if current can readily move to a different
1080 * cpu. If so, we will reschedule to allow the push logic to try
1081 * to move current somewhere else, making room for our non-migratable
1084 if (p->prio == rq->curr->prio && !need_resched())
1085 check_preempt_equal_prio(rq, p);
1089 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1090 struct rt_rq *rt_rq)
1092 struct rt_prio_array *array = &rt_rq->active;
1093 struct sched_rt_entity *next = NULL;
1094 struct list_head *queue;
1097 idx = sched_find_first_bit(array->bitmap);
1098 BUG_ON(idx >= MAX_RT_PRIO);
1100 queue = array->queue + idx;
1101 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1106 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1108 struct sched_rt_entity *rt_se;
1109 struct task_struct *p;
1110 struct rt_rq *rt_rq;
1114 if (unlikely(!rt_rq->rt_nr_running))
1117 if (rt_rq_throttled(rt_rq))
1121 rt_se = pick_next_rt_entity(rq, rt_rq);
1123 rt_rq = group_rt_rq(rt_se);
1126 p = rt_task_of(rt_se);
1127 p->se.exec_start = rq->clock_task;
1132 static struct task_struct *pick_next_task_rt(struct rq *rq)
1134 struct task_struct *p = _pick_next_task_rt(rq);
1136 /* The running task is never eligible for pushing */
1138 dequeue_pushable_task(rq, p);
1142 * We detect this state here so that we can avoid taking the RQ
1143 * lock again later if there is no need to push
1145 rq->post_schedule = has_pushable_tasks(rq);
1151 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1154 p->se.exec_start = 0;
1157 * The previous task needs to be made eligible for pushing
1158 * if it is still active
1160 if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1161 enqueue_pushable_task(rq, p);
1166 /* Only try algorithms three times */
1167 #define RT_MAX_TRIES 3
1169 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1171 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1173 if (!task_running(rq, p) &&
1174 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1175 (p->rt.nr_cpus_allowed > 1))
1180 /* Return the second highest RT task, NULL otherwise */
1181 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1183 struct task_struct *next = NULL;
1184 struct sched_rt_entity *rt_se;
1185 struct rt_prio_array *array;
1186 struct rt_rq *rt_rq;
1189 for_each_leaf_rt_rq(rt_rq, rq) {
1190 array = &rt_rq->active;
1191 idx = sched_find_first_bit(array->bitmap);
1193 if (idx >= MAX_RT_PRIO)
1195 if (next && next->prio < idx)
1197 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1198 struct task_struct *p;
1200 if (!rt_entity_is_task(rt_se))
1203 p = rt_task_of(rt_se);
1204 if (pick_rt_task(rq, p, cpu)) {
1210 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1218 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1220 static int find_lowest_rq(struct task_struct *task)
1222 struct sched_domain *sd;
1223 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1224 int this_cpu = smp_processor_id();
1225 int cpu = task_cpu(task);
1227 if (task->rt.nr_cpus_allowed == 1)
1228 return -1; /* No other targets possible */
1230 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1231 return -1; /* No targets found */
1234 * At this point we have built a mask of cpus representing the
1235 * lowest priority tasks in the system. Now we want to elect
1236 * the best one based on our affinity and topology.
1238 * We prioritize the last cpu that the task executed on since
1239 * it is most likely cache-hot in that location.
1241 if (cpumask_test_cpu(cpu, lowest_mask))
1245 * Otherwise, we consult the sched_domains span maps to figure
1246 * out which cpu is logically closest to our hot cache data.
1248 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1249 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1251 for_each_domain(cpu, sd) {
1252 if (sd->flags & SD_WAKE_AFFINE) {
1256 * "this_cpu" is cheaper to preempt than a
1259 if (this_cpu != -1 &&
1260 cpumask_test_cpu(this_cpu, sched_domain_span(sd)))
1263 best_cpu = cpumask_first_and(lowest_mask,
1264 sched_domain_span(sd));
1265 if (best_cpu < nr_cpu_ids)
1271 * And finally, if there were no matches within the domains
1272 * just give the caller *something* to work with from the compatible
1278 cpu = cpumask_any(lowest_mask);
1279 if (cpu < nr_cpu_ids)
1284 /* Will lock the rq it finds */
1285 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1287 struct rq *lowest_rq = NULL;
1291 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1292 cpu = find_lowest_rq(task);
1294 if ((cpu == -1) || (cpu == rq->cpu))
1297 lowest_rq = cpu_rq(cpu);
1299 /* if the prio of this runqueue changed, try again */
1300 if (double_lock_balance(rq, lowest_rq)) {
1302 * We had to unlock the run queue. In
1303 * the mean time, task could have
1304 * migrated already or had its affinity changed.
1305 * Also make sure that it wasn't scheduled on its rq.
1307 if (unlikely(task_rq(task) != rq ||
1308 !cpumask_test_cpu(lowest_rq->cpu,
1309 &task->cpus_allowed) ||
1310 task_running(rq, task) ||
1313 raw_spin_unlock(&lowest_rq->lock);
1319 /* If this rq is still suitable use it. */
1320 if (lowest_rq->rt.highest_prio.curr > task->prio)
1324 double_unlock_balance(rq, lowest_rq);
1331 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1333 struct task_struct *p;
1335 if (!has_pushable_tasks(rq))
1338 p = plist_first_entry(&rq->rt.pushable_tasks,
1339 struct task_struct, pushable_tasks);
1341 BUG_ON(rq->cpu != task_cpu(p));
1342 BUG_ON(task_current(rq, p));
1343 BUG_ON(p->rt.nr_cpus_allowed <= 1);
1346 BUG_ON(!rt_task(p));
1352 * If the current CPU has more than one RT task, see if the non
1353 * running task can migrate over to a CPU that is running a task
1354 * of lesser priority.
1356 static int push_rt_task(struct rq *rq)
1358 struct task_struct *next_task;
1359 struct rq *lowest_rq;
1361 if (!rq->rt.overloaded)
1364 next_task = pick_next_pushable_task(rq);
1369 if (unlikely(next_task == rq->curr)) {
1375 * It's possible that the next_task slipped in of
1376 * higher priority than current. If that's the case
1377 * just reschedule current.
1379 if (unlikely(next_task->prio < rq->curr->prio)) {
1380 resched_task(rq->curr);
1384 /* We might release rq lock */
1385 get_task_struct(next_task);
1387 /* find_lock_lowest_rq locks the rq if found */
1388 lowest_rq = find_lock_lowest_rq(next_task, rq);
1390 struct task_struct *task;
1392 * find lock_lowest_rq releases rq->lock
1393 * so it is possible that next_task has migrated.
1395 * We need to make sure that the task is still on the same
1396 * run-queue and is also still the next task eligible for
1399 task = pick_next_pushable_task(rq);
1400 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1402 * If we get here, the task hasn't moved at all, but
1403 * it has failed to push. We will not try again,
1404 * since the other cpus will pull from us when they
1407 dequeue_pushable_task(rq, next_task);
1412 /* No more tasks, just exit */
1416 * Something has shifted, try again.
1418 put_task_struct(next_task);
1423 deactivate_task(rq, next_task, 0);
1424 set_task_cpu(next_task, lowest_rq->cpu);
1425 activate_task(lowest_rq, next_task, 0);
1427 resched_task(lowest_rq->curr);
1429 double_unlock_balance(rq, lowest_rq);
1432 put_task_struct(next_task);
1437 static void push_rt_tasks(struct rq *rq)
1439 /* push_rt_task will return true if it moved an RT */
1440 while (push_rt_task(rq))
1444 static int pull_rt_task(struct rq *this_rq)
1446 int this_cpu = this_rq->cpu, ret = 0, cpu;
1447 struct task_struct *p;
1450 if (likely(!rt_overloaded(this_rq)))
1453 for_each_cpu(cpu, this_rq->rd->rto_mask) {
1454 if (this_cpu == cpu)
1457 src_rq = cpu_rq(cpu);
1460 * Don't bother taking the src_rq->lock if the next highest
1461 * task is known to be lower-priority than our current task.
1462 * This may look racy, but if this value is about to go
1463 * logically higher, the src_rq will push this task away.
1464 * And if its going logically lower, we do not care
1466 if (src_rq->rt.highest_prio.next >=
1467 this_rq->rt.highest_prio.curr)
1471 * We can potentially drop this_rq's lock in
1472 * double_lock_balance, and another CPU could
1475 double_lock_balance(this_rq, src_rq);
1478 * Are there still pullable RT tasks?
1480 if (src_rq->rt.rt_nr_running <= 1)
1483 p = pick_next_highest_task_rt(src_rq, this_cpu);
1486 * Do we have an RT task that preempts
1487 * the to-be-scheduled task?
1489 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1490 WARN_ON(p == src_rq->curr);
1494 * There's a chance that p is higher in priority
1495 * than what's currently running on its cpu.
1496 * This is just that p is wakeing up and hasn't
1497 * had a chance to schedule. We only pull
1498 * p if it is lower in priority than the
1499 * current task on the run queue
1501 if (p->prio < src_rq->curr->prio)
1506 deactivate_task(src_rq, p, 0);
1507 set_task_cpu(p, this_cpu);
1508 activate_task(this_rq, p, 0);
1510 * We continue with the search, just in
1511 * case there's an even higher prio task
1512 * in another runqueue. (low likelihood
1517 double_unlock_balance(this_rq, src_rq);
1523 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1525 /* Try to pull RT tasks here if we lower this rq's prio */
1526 if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1530 static void post_schedule_rt(struct rq *rq)
1536 * If we are not running and we are not going to reschedule soon, we should
1537 * try to push tasks away now
1539 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1541 if (!task_running(rq, p) &&
1542 !test_tsk_need_resched(rq->curr) &&
1543 has_pushable_tasks(rq) &&
1544 p->rt.nr_cpus_allowed > 1 &&
1545 rt_task(rq->curr) &&
1546 (rq->curr->rt.nr_cpus_allowed < 2 ||
1547 rq->curr->prio < p->prio))
1551 static void set_cpus_allowed_rt(struct task_struct *p,
1552 const struct cpumask *new_mask)
1554 int weight = cpumask_weight(new_mask);
1556 BUG_ON(!rt_task(p));
1559 * Update the migration status of the RQ if we have an RT task
1560 * which is running AND changing its weight value.
1562 if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1563 struct rq *rq = task_rq(p);
1565 if (!task_current(rq, p)) {
1567 * Make sure we dequeue this task from the pushable list
1568 * before going further. It will either remain off of
1569 * the list because we are no longer pushable, or it
1572 if (p->rt.nr_cpus_allowed > 1)
1573 dequeue_pushable_task(rq, p);
1576 * Requeue if our weight is changing and still > 1
1579 enqueue_pushable_task(rq, p);
1583 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1584 rq->rt.rt_nr_migratory++;
1585 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1586 BUG_ON(!rq->rt.rt_nr_migratory);
1587 rq->rt.rt_nr_migratory--;
1590 update_rt_migration(&rq->rt);
1593 cpumask_copy(&p->cpus_allowed, new_mask);
1594 p->rt.nr_cpus_allowed = weight;
1597 /* Assumes rq->lock is held */
1598 static void rq_online_rt(struct rq *rq)
1600 if (rq->rt.overloaded)
1601 rt_set_overload(rq);
1603 __enable_runtime(rq);
1605 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1608 /* Assumes rq->lock is held */
1609 static void rq_offline_rt(struct rq *rq)
1611 if (rq->rt.overloaded)
1612 rt_clear_overload(rq);
1614 __disable_runtime(rq);
1616 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1620 * When switch from the rt queue, we bring ourselves to a position
1621 * that we might want to pull RT tasks from other runqueues.
1623 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1626 * If there are other RT tasks then we will reschedule
1627 * and the scheduling of the other RT tasks will handle
1628 * the balancing. But if we are the last RT task
1629 * we may need to handle the pulling of RT tasks
1632 if (p->on_rq && !rq->rt.rt_nr_running)
1636 static inline void init_sched_rt_class(void)
1640 for_each_possible_cpu(i)
1641 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1642 GFP_KERNEL, cpu_to_node(i));
1644 #endif /* CONFIG_SMP */
1647 * When switching a task to RT, we may overload the runqueue
1648 * with RT tasks. In this case we try to push them off to
1651 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1653 int check_resched = 1;
1656 * If we are already running, then there's nothing
1657 * that needs to be done. But if we are not running
1658 * we may need to preempt the current running task.
1659 * If that current running task is also an RT task
1660 * then see if we can move to another run queue.
1662 if (p->on_rq && rq->curr != p) {
1664 if (rq->rt.overloaded && push_rt_task(rq) &&
1665 /* Don't resched if we changed runqueues */
1668 #endif /* CONFIG_SMP */
1669 if (check_resched && p->prio < rq->curr->prio)
1670 resched_task(rq->curr);
1675 * Priority of the task has changed. This may cause
1676 * us to initiate a push or pull.
1679 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1684 if (rq->curr == p) {
1687 * If our priority decreases while running, we
1688 * may need to pull tasks to this runqueue.
1690 if (oldprio < p->prio)
1693 * If there's a higher priority task waiting to run
1694 * then reschedule. Note, the above pull_rt_task
1695 * can release the rq lock and p could migrate.
1696 * Only reschedule if p is still on the same runqueue.
1698 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1701 /* For UP simply resched on drop of prio */
1702 if (oldprio < p->prio)
1704 #endif /* CONFIG_SMP */
1707 * This task is not running, but if it is
1708 * greater than the current running task
1711 if (p->prio < rq->curr->prio)
1712 resched_task(rq->curr);
1716 static void watchdog(struct rq *rq, struct task_struct *p)
1718 unsigned long soft, hard;
1720 /* max may change after cur was read, this will be fixed next tick */
1721 soft = task_rlimit(p, RLIMIT_RTTIME);
1722 hard = task_rlimit_max(p, RLIMIT_RTTIME);
1724 if (soft != RLIM_INFINITY) {
1728 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1729 if (p->rt.timeout > next)
1730 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1734 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1741 * RR tasks need a special form of timeslice management.
1742 * FIFO tasks have no timeslices.
1744 if (p->policy != SCHED_RR)
1747 if (--p->rt.time_slice)
1750 p->rt.time_slice = DEF_TIMESLICE;
1753 * Requeue to the end of queue if we are not the only element
1756 if (p->rt.run_list.prev != p->rt.run_list.next) {
1757 requeue_task_rt(rq, p, 0);
1758 set_tsk_need_resched(p);
1762 static void set_curr_task_rt(struct rq *rq)
1764 struct task_struct *p = rq->curr;
1766 p->se.exec_start = rq->clock_task;
1768 /* The running task is never eligible for pushing */
1769 dequeue_pushable_task(rq, p);
1772 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1775 * Time slice is 0 for SCHED_FIFO tasks
1777 if (task->policy == SCHED_RR)
1778 return DEF_TIMESLICE;
1783 static const struct sched_class rt_sched_class = {
1784 .next = &fair_sched_class,
1785 .enqueue_task = enqueue_task_rt,
1786 .dequeue_task = dequeue_task_rt,
1787 .yield_task = yield_task_rt,
1789 .check_preempt_curr = check_preempt_curr_rt,
1791 .pick_next_task = pick_next_task_rt,
1792 .put_prev_task = put_prev_task_rt,
1795 .select_task_rq = select_task_rq_rt,
1797 .set_cpus_allowed = set_cpus_allowed_rt,
1798 .rq_online = rq_online_rt,
1799 .rq_offline = rq_offline_rt,
1800 .pre_schedule = pre_schedule_rt,
1801 .post_schedule = post_schedule_rt,
1802 .task_woken = task_woken_rt,
1803 .switched_from = switched_from_rt,
1806 .set_curr_task = set_curr_task_rt,
1807 .task_tick = task_tick_rt,
1809 .get_rr_interval = get_rr_interval_rt,
1811 .prio_changed = prio_changed_rt,
1812 .switched_to = switched_to_rt,
1815 #ifdef CONFIG_SCHED_DEBUG
1816 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1818 static void print_rt_stats(struct seq_file *m, int cpu)
1820 struct rt_rq *rt_rq;
1823 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1824 print_rt_rq(m, cpu, rt_rq);
1827 #endif /* CONFIG_SCHED_DEBUG */