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);
118 void enqueue_pushable_task(struct rq *rq, struct task_struct *p) {}
120 void dequeue_pushable_task(struct rq *rq, struct task_struct *p) {}
122 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
124 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
126 #endif /* CONFIG_SMP */
128 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
130 return !list_empty(&rt_se->run_list);
133 #ifdef CONFIG_RT_GROUP_SCHED
135 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
140 return rt_rq->rt_runtime;
143 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
145 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
148 #define for_each_leaf_rt_rq(rt_rq, rq) \
149 list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
151 #define for_each_sched_rt_entity(rt_se) \
152 for (; rt_se; rt_se = rt_se->parent)
154 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
159 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
160 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
162 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
164 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
165 struct sched_rt_entity *rt_se = rt_rq->rt_se;
167 if (rt_rq->rt_nr_running) {
168 if (rt_se && !on_rt_rq(rt_se))
169 enqueue_rt_entity(rt_se);
170 if (rt_rq->highest_prio.curr < curr->prio)
175 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
177 struct sched_rt_entity *rt_se = rt_rq->rt_se;
179 if (rt_se && on_rt_rq(rt_se))
180 dequeue_rt_entity(rt_se);
183 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
185 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
188 static int rt_se_boosted(struct sched_rt_entity *rt_se)
190 struct rt_rq *rt_rq = group_rt_rq(rt_se);
191 struct task_struct *p;
194 return !!rt_rq->rt_nr_boosted;
196 p = rt_task_of(rt_se);
197 return p->prio != p->normal_prio;
201 static inline const struct cpumask *sched_rt_period_mask(void)
203 return cpu_rq(smp_processor_id())->rd->span;
206 static inline const struct cpumask *sched_rt_period_mask(void)
208 return cpu_online_mask;
213 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
215 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
218 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
220 return &rt_rq->tg->rt_bandwidth;
223 #else /* !CONFIG_RT_GROUP_SCHED */
225 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
227 return rt_rq->rt_runtime;
230 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
232 return ktime_to_ns(def_rt_bandwidth.rt_period);
235 #define for_each_leaf_rt_rq(rt_rq, rq) \
236 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
238 #define for_each_sched_rt_entity(rt_se) \
239 for (; rt_se; rt_se = NULL)
241 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
246 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
248 if (rt_rq->rt_nr_running)
249 resched_task(rq_of_rt_rq(rt_rq)->curr);
252 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
256 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
258 return rt_rq->rt_throttled;
261 static inline const struct cpumask *sched_rt_period_mask(void)
263 return cpu_online_mask;
267 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
269 return &cpu_rq(cpu)->rt;
272 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
274 return &def_rt_bandwidth;
277 #endif /* CONFIG_RT_GROUP_SCHED */
281 * We ran out of runtime, see if we can borrow some from our neighbours.
283 static int do_balance_runtime(struct rt_rq *rt_rq)
285 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
286 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
287 int i, weight, more = 0;
290 weight = cpumask_weight(rd->span);
292 spin_lock(&rt_b->rt_runtime_lock);
293 rt_period = ktime_to_ns(rt_b->rt_period);
294 for_each_cpu(i, rd->span) {
295 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
301 spin_lock(&iter->rt_runtime_lock);
303 * Either all rqs have inf runtime and there's nothing to steal
304 * or __disable_runtime() below sets a specific rq to inf to
305 * indicate its been disabled and disalow stealing.
307 if (iter->rt_runtime == RUNTIME_INF)
311 * From runqueues with spare time, take 1/n part of their
312 * spare time, but no more than our period.
314 diff = iter->rt_runtime - iter->rt_time;
316 diff = div_u64((u64)diff, weight);
317 if (rt_rq->rt_runtime + diff > rt_period)
318 diff = rt_period - rt_rq->rt_runtime;
319 iter->rt_runtime -= diff;
320 rt_rq->rt_runtime += diff;
322 if (rt_rq->rt_runtime == rt_period) {
323 spin_unlock(&iter->rt_runtime_lock);
328 spin_unlock(&iter->rt_runtime_lock);
330 spin_unlock(&rt_b->rt_runtime_lock);
336 * Ensure this RQ takes back all the runtime it lend to its neighbours.
338 static void __disable_runtime(struct rq *rq)
340 struct root_domain *rd = rq->rd;
343 if (unlikely(!scheduler_running))
346 for_each_leaf_rt_rq(rt_rq, rq) {
347 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
351 spin_lock(&rt_b->rt_runtime_lock);
352 spin_lock(&rt_rq->rt_runtime_lock);
354 * Either we're all inf and nobody needs to borrow, or we're
355 * already disabled and thus have nothing to do, or we have
356 * exactly the right amount of runtime to take out.
358 if (rt_rq->rt_runtime == RUNTIME_INF ||
359 rt_rq->rt_runtime == rt_b->rt_runtime)
361 spin_unlock(&rt_rq->rt_runtime_lock);
364 * Calculate the difference between what we started out with
365 * and what we current have, that's the amount of runtime
366 * we lend and now have to reclaim.
368 want = rt_b->rt_runtime - rt_rq->rt_runtime;
371 * Greedy reclaim, take back as much as we can.
373 for_each_cpu(i, rd->span) {
374 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
378 * Can't reclaim from ourselves or disabled runqueues.
380 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
383 spin_lock(&iter->rt_runtime_lock);
385 diff = min_t(s64, iter->rt_runtime, want);
386 iter->rt_runtime -= diff;
389 iter->rt_runtime -= want;
392 spin_unlock(&iter->rt_runtime_lock);
398 spin_lock(&rt_rq->rt_runtime_lock);
400 * We cannot be left wanting - that would mean some runtime
401 * leaked out of the system.
406 * Disable all the borrow logic by pretending we have inf
407 * runtime - in which case borrowing doesn't make sense.
409 rt_rq->rt_runtime = RUNTIME_INF;
410 spin_unlock(&rt_rq->rt_runtime_lock);
411 spin_unlock(&rt_b->rt_runtime_lock);
415 static void disable_runtime(struct rq *rq)
419 spin_lock_irqsave(&rq->lock, flags);
420 __disable_runtime(rq);
421 spin_unlock_irqrestore(&rq->lock, flags);
424 static void __enable_runtime(struct rq *rq)
428 if (unlikely(!scheduler_running))
432 * Reset each runqueue's bandwidth settings
434 for_each_leaf_rt_rq(rt_rq, rq) {
435 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
437 spin_lock(&rt_b->rt_runtime_lock);
438 spin_lock(&rt_rq->rt_runtime_lock);
439 rt_rq->rt_runtime = rt_b->rt_runtime;
441 rt_rq->rt_throttled = 0;
442 spin_unlock(&rt_rq->rt_runtime_lock);
443 spin_unlock(&rt_b->rt_runtime_lock);
447 static void enable_runtime(struct rq *rq)
451 spin_lock_irqsave(&rq->lock, flags);
452 __enable_runtime(rq);
453 spin_unlock_irqrestore(&rq->lock, flags);
456 static int balance_runtime(struct rt_rq *rt_rq)
460 if (rt_rq->rt_time > rt_rq->rt_runtime) {
461 spin_unlock(&rt_rq->rt_runtime_lock);
462 more = do_balance_runtime(rt_rq);
463 spin_lock(&rt_rq->rt_runtime_lock);
468 #else /* !CONFIG_SMP */
469 static inline int balance_runtime(struct rt_rq *rt_rq)
473 #endif /* CONFIG_SMP */
475 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
478 const struct cpumask *span;
480 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
483 span = sched_rt_period_mask();
484 for_each_cpu(i, span) {
486 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
487 struct rq *rq = rq_of_rt_rq(rt_rq);
489 spin_lock(&rq->lock);
490 if (rt_rq->rt_time) {
493 spin_lock(&rt_rq->rt_runtime_lock);
494 if (rt_rq->rt_throttled)
495 balance_runtime(rt_rq);
496 runtime = rt_rq->rt_runtime;
497 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
498 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
499 rt_rq->rt_throttled = 0;
502 if (rt_rq->rt_time || rt_rq->rt_nr_running)
504 spin_unlock(&rt_rq->rt_runtime_lock);
505 } else if (rt_rq->rt_nr_running)
509 sched_rt_rq_enqueue(rt_rq);
510 spin_unlock(&rq->lock);
516 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
518 #ifdef CONFIG_RT_GROUP_SCHED
519 struct rt_rq *rt_rq = group_rt_rq(rt_se);
522 return rt_rq->highest_prio.curr;
525 return rt_task_of(rt_se)->prio;
528 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
530 u64 runtime = sched_rt_runtime(rt_rq);
532 if (rt_rq->rt_throttled)
533 return rt_rq_throttled(rt_rq);
535 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
538 balance_runtime(rt_rq);
539 runtime = sched_rt_runtime(rt_rq);
540 if (runtime == RUNTIME_INF)
543 if (rt_rq->rt_time > runtime) {
544 rt_rq->rt_throttled = 1;
545 if (rt_rq_throttled(rt_rq)) {
546 sched_rt_rq_dequeue(rt_rq);
555 * Update the current task's runtime statistics. Skip current tasks that
556 * are not in our scheduling class.
558 static void update_curr_rt(struct rq *rq)
560 struct task_struct *curr = rq->curr;
561 struct sched_rt_entity *rt_se = &curr->rt;
562 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
565 if (!task_has_rt_policy(curr))
568 delta_exec = rq->clock - curr->se.exec_start;
569 if (unlikely((s64)delta_exec < 0))
572 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
574 curr->se.sum_exec_runtime += delta_exec;
575 account_group_exec_runtime(curr, delta_exec);
577 curr->se.exec_start = rq->clock;
578 cpuacct_charge(curr, delta_exec);
580 if (!rt_bandwidth_enabled())
583 for_each_sched_rt_entity(rt_se) {
584 rt_rq = rt_rq_of_se(rt_se);
586 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
587 spin_lock(&rt_rq->rt_runtime_lock);
588 rt_rq->rt_time += delta_exec;
589 if (sched_rt_runtime_exceeded(rt_rq))
591 spin_unlock(&rt_rq->rt_runtime_lock);
596 #if defined CONFIG_SMP
598 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
600 static inline int next_prio(struct rq *rq)
602 struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
604 if (next && rt_prio(next->prio))
611 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
613 struct rq *rq = rq_of_rt_rq(rt_rq);
615 if (prio < prev_prio) {
618 * If the new task is higher in priority than anything on the
619 * run-queue, we know that the previous high becomes our
622 rt_rq->highest_prio.next = prev_prio;
625 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
627 } else if (prio == rt_rq->highest_prio.curr)
629 * If the next task is equal in priority to the highest on
630 * the run-queue, then we implicitly know that the next highest
631 * task cannot be any lower than current
633 rt_rq->highest_prio.next = prio;
634 else if (prio < rt_rq->highest_prio.next)
636 * Otherwise, we need to recompute next-highest
638 rt_rq->highest_prio.next = next_prio(rq);
642 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
644 struct rq *rq = rq_of_rt_rq(rt_rq);
646 if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
647 rt_rq->highest_prio.next = next_prio(rq);
649 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
650 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
653 #else /* CONFIG_SMP */
656 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
658 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
660 #endif /* CONFIG_SMP */
662 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
664 inc_rt_prio(struct rt_rq *rt_rq, int prio)
666 int prev_prio = rt_rq->highest_prio.curr;
668 if (prio < prev_prio)
669 rt_rq->highest_prio.curr = prio;
671 inc_rt_prio_smp(rt_rq, prio, prev_prio);
675 dec_rt_prio(struct rt_rq *rt_rq, int prio)
677 int prev_prio = rt_rq->highest_prio.curr;
679 if (rt_rq->rt_nr_running) {
681 WARN_ON(prio < prev_prio);
684 * This may have been our highest task, and therefore
685 * we may have some recomputation to do
687 if (prio == prev_prio) {
688 struct rt_prio_array *array = &rt_rq->active;
690 rt_rq->highest_prio.curr =
691 sched_find_first_bit(array->bitmap);
695 rt_rq->highest_prio.curr = MAX_RT_PRIO;
697 dec_rt_prio_smp(rt_rq, prio, prev_prio);
702 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
703 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
705 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
707 #ifdef CONFIG_RT_GROUP_SCHED
710 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
712 if (rt_se_boosted(rt_se))
713 rt_rq->rt_nr_boosted++;
716 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
720 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
722 if (rt_se_boosted(rt_se))
723 rt_rq->rt_nr_boosted--;
725 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
728 #else /* CONFIG_RT_GROUP_SCHED */
731 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
733 start_rt_bandwidth(&def_rt_bandwidth);
737 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
739 #endif /* CONFIG_RT_GROUP_SCHED */
742 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
744 int prio = rt_se_prio(rt_se);
746 WARN_ON(!rt_prio(prio));
747 rt_rq->rt_nr_running++;
749 inc_rt_prio(rt_rq, prio);
750 inc_rt_migration(rt_se, rt_rq);
751 inc_rt_group(rt_se, rt_rq);
755 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
757 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
758 WARN_ON(!rt_rq->rt_nr_running);
759 rt_rq->rt_nr_running--;
761 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
762 dec_rt_migration(rt_se, rt_rq);
763 dec_rt_group(rt_se, rt_rq);
766 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
768 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
769 struct rt_prio_array *array = &rt_rq->active;
770 struct rt_rq *group_rq = group_rt_rq(rt_se);
771 struct list_head *queue = array->queue + rt_se_prio(rt_se);
774 * Don't enqueue the group if its throttled, or when empty.
775 * The latter is a consequence of the former when a child group
776 * get throttled and the current group doesn't have any other
779 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
782 list_add_tail(&rt_se->run_list, queue);
783 __set_bit(rt_se_prio(rt_se), array->bitmap);
785 inc_rt_tasks(rt_se, rt_rq);
788 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
790 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
791 struct rt_prio_array *array = &rt_rq->active;
793 list_del_init(&rt_se->run_list);
794 if (list_empty(array->queue + rt_se_prio(rt_se)))
795 __clear_bit(rt_se_prio(rt_se), array->bitmap);
797 dec_rt_tasks(rt_se, rt_rq);
801 * Because the prio of an upper entry depends on the lower
802 * entries, we must remove entries top - down.
804 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
806 struct sched_rt_entity *back = NULL;
808 for_each_sched_rt_entity(rt_se) {
813 for (rt_se = back; rt_se; rt_se = rt_se->back) {
815 __dequeue_rt_entity(rt_se);
819 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
821 dequeue_rt_stack(rt_se);
822 for_each_sched_rt_entity(rt_se)
823 __enqueue_rt_entity(rt_se);
826 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
828 dequeue_rt_stack(rt_se);
830 for_each_sched_rt_entity(rt_se) {
831 struct rt_rq *rt_rq = group_rt_rq(rt_se);
833 if (rt_rq && rt_rq->rt_nr_running)
834 __enqueue_rt_entity(rt_se);
839 * Adding/removing a task to/from a priority array:
841 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
843 struct sched_rt_entity *rt_se = &p->rt;
848 enqueue_rt_entity(rt_se);
850 if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
851 enqueue_pushable_task(rq, p);
853 inc_cpu_load(rq, p->se.load.weight);
856 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
858 struct sched_rt_entity *rt_se = &p->rt;
861 dequeue_rt_entity(rt_se);
863 dequeue_pushable_task(rq, p);
865 dec_cpu_load(rq, p->se.load.weight);
869 * Put task to the end of the run list without the overhead of dequeue
870 * followed by enqueue.
873 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
875 if (on_rt_rq(rt_se)) {
876 struct rt_prio_array *array = &rt_rq->active;
877 struct list_head *queue = array->queue + rt_se_prio(rt_se);
880 list_move(&rt_se->run_list, queue);
882 list_move_tail(&rt_se->run_list, queue);
886 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
888 struct sched_rt_entity *rt_se = &p->rt;
891 for_each_sched_rt_entity(rt_se) {
892 rt_rq = rt_rq_of_se(rt_se);
893 requeue_rt_entity(rt_rq, rt_se, head);
897 static void yield_task_rt(struct rq *rq)
899 requeue_task_rt(rq, rq->curr, 0);
903 static int find_lowest_rq(struct task_struct *task);
905 static int select_task_rq_rt(struct task_struct *p, int sync)
907 struct rq *rq = task_rq(p);
910 * If the current task is an RT task, then
911 * try to see if we can wake this RT task up on another
912 * runqueue. Otherwise simply start this RT task
913 * on its current runqueue.
915 * We want to avoid overloading runqueues. Even if
916 * the RT task is of higher priority than the current RT task.
917 * RT tasks behave differently than other tasks. If
918 * one gets preempted, we try to push it off to another queue.
919 * So trying to keep a preempting RT task on the same
920 * cache hot CPU will force the running RT task to
921 * a cold CPU. So we waste all the cache for the lower
922 * RT task in hopes of saving some of a RT task
923 * that is just being woken and probably will have
926 if (unlikely(rt_task(rq->curr)) &&
927 (p->rt.nr_cpus_allowed > 1)) {
928 int cpu = find_lowest_rq(p);
930 return (cpu == -1) ? task_cpu(p) : cpu;
934 * Otherwise, just let it ride on the affined RQ and the
935 * post-schedule router will push the preempted task away
940 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
944 if (rq->curr->rt.nr_cpus_allowed == 1)
947 if (!alloc_cpumask_var(&mask, GFP_ATOMIC))
950 if (p->rt.nr_cpus_allowed != 1
951 && cpupri_find(&rq->rd->cpupri, p, mask))
954 if (!cpupri_find(&rq->rd->cpupri, rq->curr, mask))
958 * There appears to be other cpus that can accept
959 * current and none to run 'p', so lets reschedule
960 * to try and push current away:
962 requeue_task_rt(rq, p, 1);
963 resched_task(rq->curr);
965 free_cpumask_var(mask);
968 #endif /* CONFIG_SMP */
971 * Preempt the current task with a newly woken task if needed:
973 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
975 if (p->prio < rq->curr->prio) {
976 resched_task(rq->curr);
984 * - the newly woken task is of equal priority to the current task
985 * - the newly woken task is non-migratable while current is migratable
986 * - current will be preempted on the next reschedule
988 * we should check to see if current can readily move to a different
989 * cpu. If so, we will reschedule to allow the push logic to try
990 * to move current somewhere else, making room for our non-migratable
993 if (p->prio == rq->curr->prio && !need_resched())
994 check_preempt_equal_prio(rq, p);
998 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1001 struct rt_prio_array *array = &rt_rq->active;
1002 struct sched_rt_entity *next = NULL;
1003 struct list_head *queue;
1006 idx = sched_find_first_bit(array->bitmap);
1007 BUG_ON(idx >= MAX_RT_PRIO);
1009 queue = array->queue + idx;
1010 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1015 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1017 struct sched_rt_entity *rt_se;
1018 struct task_struct *p;
1019 struct rt_rq *rt_rq;
1023 if (unlikely(!rt_rq->rt_nr_running))
1026 if (rt_rq_throttled(rt_rq))
1030 rt_se = pick_next_rt_entity(rq, rt_rq);
1032 rt_rq = group_rt_rq(rt_se);
1035 p = rt_task_of(rt_se);
1036 p->se.exec_start = rq->clock;
1041 static struct task_struct *pick_next_task_rt(struct rq *rq)
1043 struct task_struct *p = _pick_next_task_rt(rq);
1045 /* The running task is never eligible for pushing */
1047 dequeue_pushable_task(rq, p);
1052 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1055 p->se.exec_start = 0;
1058 * The previous task needs to be made eligible for pushing
1059 * if it is still active
1061 if (p->se.on_rq && p->rt.nr_cpus_allowed > 1)
1062 enqueue_pushable_task(rq, p);
1067 /* Only try algorithms three times */
1068 #define RT_MAX_TRIES 3
1070 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1072 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1074 if (!task_running(rq, p) &&
1075 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1076 (p->rt.nr_cpus_allowed > 1))
1081 /* Return the second highest RT task, NULL otherwise */
1082 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1084 struct task_struct *next = NULL;
1085 struct sched_rt_entity *rt_se;
1086 struct rt_prio_array *array;
1087 struct rt_rq *rt_rq;
1090 for_each_leaf_rt_rq(rt_rq, rq) {
1091 array = &rt_rq->active;
1092 idx = sched_find_first_bit(array->bitmap);
1094 if (idx >= MAX_RT_PRIO)
1096 if (next && next->prio < idx)
1098 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1099 struct task_struct *p = rt_task_of(rt_se);
1100 if (pick_rt_task(rq, p, cpu)) {
1106 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1114 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1116 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
1120 /* "this_cpu" is cheaper to preempt than a remote processor */
1121 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
1124 first = first_cpu(*mask);
1125 if (first != NR_CPUS)
1131 static int find_lowest_rq(struct task_struct *task)
1133 struct sched_domain *sd;
1134 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1135 int this_cpu = smp_processor_id();
1136 int cpu = task_cpu(task);
1138 if (task->rt.nr_cpus_allowed == 1)
1139 return -1; /* No other targets possible */
1141 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1142 return -1; /* No targets found */
1145 * Only consider CPUs that are usable for migration.
1146 * I guess we might want to change cpupri_find() to ignore those
1147 * in the first place.
1149 cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
1152 * At this point we have built a mask of cpus representing the
1153 * lowest priority tasks in the system. Now we want to elect
1154 * the best one based on our affinity and topology.
1156 * We prioritize the last cpu that the task executed on since
1157 * it is most likely cache-hot in that location.
1159 if (cpumask_test_cpu(cpu, lowest_mask))
1163 * Otherwise, we consult the sched_domains span maps to figure
1164 * out which cpu is logically closest to our hot cache data.
1166 if (this_cpu == cpu)
1167 this_cpu = -1; /* Skip this_cpu opt if the same */
1169 for_each_domain(cpu, sd) {
1170 if (sd->flags & SD_WAKE_AFFINE) {
1171 cpumask_t domain_mask;
1174 cpumask_and(&domain_mask, sched_domain_span(sd),
1177 best_cpu = pick_optimal_cpu(this_cpu,
1185 * And finally, if there were no matches within the domains
1186 * just give the caller *something* to work with from the compatible
1189 return pick_optimal_cpu(this_cpu, lowest_mask);
1192 /* Will lock the rq it finds */
1193 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1195 struct rq *lowest_rq = NULL;
1199 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1200 cpu = find_lowest_rq(task);
1202 if ((cpu == -1) || (cpu == rq->cpu))
1205 lowest_rq = cpu_rq(cpu);
1207 /* if the prio of this runqueue changed, try again */
1208 if (double_lock_balance(rq, lowest_rq)) {
1210 * We had to unlock the run queue. In
1211 * the mean time, task could have
1212 * migrated already or had its affinity changed.
1213 * Also make sure that it wasn't scheduled on its rq.
1215 if (unlikely(task_rq(task) != rq ||
1216 !cpumask_test_cpu(lowest_rq->cpu,
1217 &task->cpus_allowed) ||
1218 task_running(rq, task) ||
1221 spin_unlock(&lowest_rq->lock);
1227 /* If this rq is still suitable use it. */
1228 if (lowest_rq->rt.highest_prio.curr > task->prio)
1232 double_unlock_balance(rq, lowest_rq);
1239 static inline int has_pushable_tasks(struct rq *rq)
1241 return !plist_head_empty(&rq->rt.pushable_tasks);
1244 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1246 struct task_struct *p;
1248 if (!has_pushable_tasks(rq))
1251 p = plist_first_entry(&rq->rt.pushable_tasks,
1252 struct task_struct, pushable_tasks);
1254 BUG_ON(rq->cpu != task_cpu(p));
1255 BUG_ON(task_current(rq, p));
1256 BUG_ON(p->rt.nr_cpus_allowed <= 1);
1258 BUG_ON(!p->se.on_rq);
1259 BUG_ON(!rt_task(p));
1265 * If the current CPU has more than one RT task, see if the non
1266 * running task can migrate over to a CPU that is running a task
1267 * of lesser priority.
1269 static int push_rt_task(struct rq *rq)
1271 struct task_struct *next_task;
1272 struct rq *lowest_rq;
1274 if (!rq->rt.overloaded)
1277 next_task = pick_next_pushable_task(rq);
1282 if (unlikely(next_task == rq->curr)) {
1288 * It's possible that the next_task slipped in of
1289 * higher priority than current. If that's the case
1290 * just reschedule current.
1292 if (unlikely(next_task->prio < rq->curr->prio)) {
1293 resched_task(rq->curr);
1297 /* We might release rq lock */
1298 get_task_struct(next_task);
1300 /* find_lock_lowest_rq locks the rq if found */
1301 lowest_rq = find_lock_lowest_rq(next_task, rq);
1303 struct task_struct *task;
1305 * find lock_lowest_rq releases rq->lock
1306 * so it is possible that next_task has migrated.
1308 * We need to make sure that the task is still on the same
1309 * run-queue and is also still the next task eligible for
1312 task = pick_next_pushable_task(rq);
1313 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1315 * If we get here, the task hasnt moved at all, but
1316 * it has failed to push. We will not try again,
1317 * since the other cpus will pull from us when they
1320 dequeue_pushable_task(rq, next_task);
1325 /* No more tasks, just exit */
1329 * Something has shifted, try again.
1331 put_task_struct(next_task);
1336 deactivate_task(rq, next_task, 0);
1337 set_task_cpu(next_task, lowest_rq->cpu);
1338 activate_task(lowest_rq, next_task, 0);
1340 resched_task(lowest_rq->curr);
1342 double_unlock_balance(rq, lowest_rq);
1345 put_task_struct(next_task);
1350 static void push_rt_tasks(struct rq *rq)
1352 /* push_rt_task will return true if it moved an RT */
1353 while (push_rt_task(rq))
1357 static int pull_rt_task(struct rq *this_rq)
1359 int this_cpu = this_rq->cpu, ret = 0, cpu;
1360 struct task_struct *p;
1363 if (likely(!rt_overloaded(this_rq)))
1366 for_each_cpu(cpu, this_rq->rd->rto_mask) {
1367 if (this_cpu == cpu)
1370 src_rq = cpu_rq(cpu);
1373 * Don't bother taking the src_rq->lock if the next highest
1374 * task is known to be lower-priority than our current task.
1375 * This may look racy, but if this value is about to go
1376 * logically higher, the src_rq will push this task away.
1377 * And if its going logically lower, we do not care
1379 if (src_rq->rt.highest_prio.next >=
1380 this_rq->rt.highest_prio.curr)
1384 * We can potentially drop this_rq's lock in
1385 * double_lock_balance, and another CPU could
1388 double_lock_balance(this_rq, src_rq);
1391 * Are there still pullable RT tasks?
1393 if (src_rq->rt.rt_nr_running <= 1)
1396 p = pick_next_highest_task_rt(src_rq, this_cpu);
1399 * Do we have an RT task that preempts
1400 * the to-be-scheduled task?
1402 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1403 WARN_ON(p == src_rq->curr);
1404 WARN_ON(!p->se.on_rq);
1407 * There's a chance that p is higher in priority
1408 * than what's currently running on its cpu.
1409 * This is just that p is wakeing up and hasn't
1410 * had a chance to schedule. We only pull
1411 * p if it is lower in priority than the
1412 * current task on the run queue
1414 if (p->prio < src_rq->curr->prio)
1419 deactivate_task(src_rq, p, 0);
1420 set_task_cpu(p, this_cpu);
1421 activate_task(this_rq, p, 0);
1423 * We continue with the search, just in
1424 * case there's an even higher prio task
1425 * in another runqueue. (low likelyhood
1430 double_unlock_balance(this_rq, src_rq);
1436 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1438 /* Try to pull RT tasks here if we lower this rq's prio */
1439 if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1444 * assumes rq->lock is held
1446 static int needs_post_schedule_rt(struct rq *rq)
1448 return has_pushable_tasks(rq);
1451 static void post_schedule_rt(struct rq *rq)
1454 * This is only called if needs_post_schedule_rt() indicates that
1455 * we need to push tasks away
1457 spin_lock_irq(&rq->lock);
1459 spin_unlock_irq(&rq->lock);
1463 * If we are not running and we are not going to reschedule soon, we should
1464 * try to push tasks away now
1466 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1468 if (!task_running(rq, p) &&
1469 !test_tsk_need_resched(rq->curr) &&
1470 has_pushable_tasks(rq) &&
1471 p->rt.nr_cpus_allowed > 1)
1475 static unsigned long
1476 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1477 unsigned long max_load_move,
1478 struct sched_domain *sd, enum cpu_idle_type idle,
1479 int *all_pinned, int *this_best_prio)
1481 /* don't touch RT tasks */
1486 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1487 struct sched_domain *sd, enum cpu_idle_type idle)
1489 /* don't touch RT tasks */
1493 static void set_cpus_allowed_rt(struct task_struct *p,
1494 const struct cpumask *new_mask)
1496 int weight = cpumask_weight(new_mask);
1498 BUG_ON(!rt_task(p));
1501 * Update the migration status of the RQ if we have an RT task
1502 * which is running AND changing its weight value.
1504 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1505 struct rq *rq = task_rq(p);
1507 if (!task_current(rq, p)) {
1509 * Make sure we dequeue this task from the pushable list
1510 * before going further. It will either remain off of
1511 * the list because we are no longer pushable, or it
1514 if (p->rt.nr_cpus_allowed > 1)
1515 dequeue_pushable_task(rq, p);
1518 * Requeue if our weight is changing and still > 1
1521 enqueue_pushable_task(rq, p);
1525 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1526 rq->rt.rt_nr_migratory++;
1527 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1528 BUG_ON(!rq->rt.rt_nr_migratory);
1529 rq->rt.rt_nr_migratory--;
1532 update_rt_migration(&rq->rt);
1535 cpumask_copy(&p->cpus_allowed, new_mask);
1536 p->rt.nr_cpus_allowed = weight;
1539 /* Assumes rq->lock is held */
1540 static void rq_online_rt(struct rq *rq)
1542 if (rq->rt.overloaded)
1543 rt_set_overload(rq);
1545 __enable_runtime(rq);
1547 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1550 /* Assumes rq->lock is held */
1551 static void rq_offline_rt(struct rq *rq)
1553 if (rq->rt.overloaded)
1554 rt_clear_overload(rq);
1556 __disable_runtime(rq);
1558 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1562 * When switch from the rt queue, we bring ourselves to a position
1563 * that we might want to pull RT tasks from other runqueues.
1565 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1569 * If there are other RT tasks then we will reschedule
1570 * and the scheduling of the other RT tasks will handle
1571 * the balancing. But if we are the last RT task
1572 * we may need to handle the pulling of RT tasks
1575 if (!rq->rt.rt_nr_running)
1579 static inline void init_sched_rt_class(void)
1583 for_each_possible_cpu(i)
1584 alloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1585 GFP_KERNEL, cpu_to_node(i));
1587 #endif /* CONFIG_SMP */
1590 * When switching a task to RT, we may overload the runqueue
1591 * with RT tasks. In this case we try to push them off to
1594 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1597 int check_resched = 1;
1600 * If we are already running, then there's nothing
1601 * that needs to be done. But if we are not running
1602 * we may need to preempt the current running task.
1603 * If that current running task is also an RT task
1604 * then see if we can move to another run queue.
1608 if (rq->rt.overloaded && push_rt_task(rq) &&
1609 /* Don't resched if we changed runqueues */
1612 #endif /* CONFIG_SMP */
1613 if (check_resched && p->prio < rq->curr->prio)
1614 resched_task(rq->curr);
1619 * Priority of the task has changed. This may cause
1620 * us to initiate a push or pull.
1622 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1623 int oldprio, int running)
1628 * If our priority decreases while running, we
1629 * may need to pull tasks to this runqueue.
1631 if (oldprio < p->prio)
1634 * If there's a higher priority task waiting to run
1635 * then reschedule. Note, the above pull_rt_task
1636 * can release the rq lock and p could migrate.
1637 * Only reschedule if p is still on the same runqueue.
1639 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1642 /* For UP simply resched on drop of prio */
1643 if (oldprio < p->prio)
1645 #endif /* CONFIG_SMP */
1648 * This task is not running, but if it is
1649 * greater than the current running task
1652 if (p->prio < rq->curr->prio)
1653 resched_task(rq->curr);
1657 static void watchdog(struct rq *rq, struct task_struct *p)
1659 unsigned long soft, hard;
1664 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1665 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1667 if (soft != RLIM_INFINITY) {
1671 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1672 if (p->rt.timeout > next)
1673 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1677 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1684 * RR tasks need a special form of timeslice management.
1685 * FIFO tasks have no timeslices.
1687 if (p->policy != SCHED_RR)
1690 if (--p->rt.time_slice)
1693 p->rt.time_slice = DEF_TIMESLICE;
1696 * Requeue to the end of queue if we are not the only element
1699 if (p->rt.run_list.prev != p->rt.run_list.next) {
1700 requeue_task_rt(rq, p, 0);
1701 set_tsk_need_resched(p);
1705 static void set_curr_task_rt(struct rq *rq)
1707 struct task_struct *p = rq->curr;
1709 p->se.exec_start = rq->clock;
1711 /* The running task is never eligible for pushing */
1712 dequeue_pushable_task(rq, p);
1715 static const struct sched_class rt_sched_class = {
1716 .next = &fair_sched_class,
1717 .enqueue_task = enqueue_task_rt,
1718 .dequeue_task = dequeue_task_rt,
1719 .yield_task = yield_task_rt,
1721 .check_preempt_curr = check_preempt_curr_rt,
1723 .pick_next_task = pick_next_task_rt,
1724 .put_prev_task = put_prev_task_rt,
1727 .select_task_rq = select_task_rq_rt,
1729 .load_balance = load_balance_rt,
1730 .move_one_task = move_one_task_rt,
1731 .set_cpus_allowed = set_cpus_allowed_rt,
1732 .rq_online = rq_online_rt,
1733 .rq_offline = rq_offline_rt,
1734 .pre_schedule = pre_schedule_rt,
1735 .needs_post_schedule = needs_post_schedule_rt,
1736 .post_schedule = post_schedule_rt,
1737 .task_wake_up = task_wake_up_rt,
1738 .switched_from = switched_from_rt,
1741 .set_curr_task = set_curr_task_rt,
1742 .task_tick = task_tick_rt,
1744 .prio_changed = prio_changed_rt,
1745 .switched_to = switched_to_rt,
1748 #ifdef CONFIG_SCHED_DEBUG
1749 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1751 static void print_rt_stats(struct seq_file *m, int cpu)
1753 struct rt_rq *rt_rq;
1756 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1757 print_rt_rq(m, cpu, rt_rq);
1760 #endif /* CONFIG_SCHED_DEBUG */