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 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
15 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
20 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
25 #else /* CONFIG_RT_GROUP_SCHED */
27 #define rt_entity_is_task(rt_se) (1)
29 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
31 return container_of(rt_rq, struct rq, rt);
34 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
36 struct task_struct *p = rt_task_of(rt_se);
37 struct rq *rq = task_rq(p);
42 #endif /* CONFIG_RT_GROUP_SCHED */
46 static inline int rt_overloaded(struct rq *rq)
48 return atomic_read(&rq->rd->rto_count);
51 static inline void rt_set_overload(struct rq *rq)
56 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
58 * Make sure the mask is visible before we set
59 * the overload count. That is checked to determine
60 * if we should look at the mask. It would be a shame
61 * if we looked at the mask, but the mask was not
65 atomic_inc(&rq->rd->rto_count);
68 static inline void rt_clear_overload(struct rq *rq)
73 /* the order here really doesn't matter */
74 atomic_dec(&rq->rd->rto_count);
75 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
78 static void update_rt_migration(struct rt_rq *rt_rq)
80 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
81 if (!rt_rq->overloaded) {
82 rt_set_overload(rq_of_rt_rq(rt_rq));
83 rt_rq->overloaded = 1;
85 } else if (rt_rq->overloaded) {
86 rt_clear_overload(rq_of_rt_rq(rt_rq));
87 rt_rq->overloaded = 0;
91 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
93 if (!rt_entity_is_task(rt_se))
96 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
99 if (rt_se->nr_cpus_allowed > 1)
100 rt_rq->rt_nr_migratory++;
102 update_rt_migration(rt_rq);
105 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
107 if (!rt_entity_is_task(rt_se))
110 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
112 rt_rq->rt_nr_total--;
113 if (rt_se->nr_cpus_allowed > 1)
114 rt_rq->rt_nr_migratory--;
116 update_rt_migration(rt_rq);
119 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
121 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
122 plist_node_init(&p->pushable_tasks, p->prio);
123 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
126 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
128 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
133 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
137 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
142 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
147 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
151 #endif /* CONFIG_SMP */
153 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
155 return !list_empty(&rt_se->run_list);
158 #ifdef CONFIG_RT_GROUP_SCHED
160 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
165 return rt_rq->rt_runtime;
168 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
170 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
173 #define for_each_leaf_rt_rq(rt_rq, rq) \
174 list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
176 #define for_each_sched_rt_entity(rt_se) \
177 for (; rt_se; rt_se = rt_se->parent)
179 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
184 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
185 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
187 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
189 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
190 struct sched_rt_entity *rt_se = rt_rq->rt_se;
192 if (rt_rq->rt_nr_running) {
193 if (rt_se && !on_rt_rq(rt_se))
194 enqueue_rt_entity(rt_se);
195 if (rt_rq->highest_prio.curr < curr->prio)
200 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
202 struct sched_rt_entity *rt_se = rt_rq->rt_se;
204 if (rt_se && on_rt_rq(rt_se))
205 dequeue_rt_entity(rt_se);
208 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
210 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
213 static int rt_se_boosted(struct sched_rt_entity *rt_se)
215 struct rt_rq *rt_rq = group_rt_rq(rt_se);
216 struct task_struct *p;
219 return !!rt_rq->rt_nr_boosted;
221 p = rt_task_of(rt_se);
222 return p->prio != p->normal_prio;
226 static inline const struct cpumask *sched_rt_period_mask(void)
228 return cpu_rq(smp_processor_id())->rd->span;
231 static inline const struct cpumask *sched_rt_period_mask(void)
233 return cpu_online_mask;
238 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
240 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
243 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
245 return &rt_rq->tg->rt_bandwidth;
248 #else /* !CONFIG_RT_GROUP_SCHED */
250 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
252 return rt_rq->rt_runtime;
255 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
257 return ktime_to_ns(def_rt_bandwidth.rt_period);
260 #define for_each_leaf_rt_rq(rt_rq, rq) \
261 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
263 #define for_each_sched_rt_entity(rt_se) \
264 for (; rt_se; rt_se = NULL)
266 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
271 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
273 if (rt_rq->rt_nr_running)
274 resched_task(rq_of_rt_rq(rt_rq)->curr);
277 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
281 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
283 return rt_rq->rt_throttled;
286 static inline const struct cpumask *sched_rt_period_mask(void)
288 return cpu_online_mask;
292 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
294 return &cpu_rq(cpu)->rt;
297 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
299 return &def_rt_bandwidth;
302 #endif /* CONFIG_RT_GROUP_SCHED */
306 * We ran out of runtime, see if we can borrow some from our neighbours.
308 static int do_balance_runtime(struct rt_rq *rt_rq)
310 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
311 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
312 int i, weight, more = 0;
315 weight = cpumask_weight(rd->span);
317 spin_lock(&rt_b->rt_runtime_lock);
318 rt_period = ktime_to_ns(rt_b->rt_period);
319 for_each_cpu(i, rd->span) {
320 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
326 spin_lock(&iter->rt_runtime_lock);
328 * Either all rqs have inf runtime and there's nothing to steal
329 * or __disable_runtime() below sets a specific rq to inf to
330 * indicate its been disabled and disalow stealing.
332 if (iter->rt_runtime == RUNTIME_INF)
336 * From runqueues with spare time, take 1/n part of their
337 * spare time, but no more than our period.
339 diff = iter->rt_runtime - iter->rt_time;
341 diff = div_u64((u64)diff, weight);
342 if (rt_rq->rt_runtime + diff > rt_period)
343 diff = rt_period - rt_rq->rt_runtime;
344 iter->rt_runtime -= diff;
345 rt_rq->rt_runtime += diff;
347 if (rt_rq->rt_runtime == rt_period) {
348 spin_unlock(&iter->rt_runtime_lock);
353 spin_unlock(&iter->rt_runtime_lock);
355 spin_unlock(&rt_b->rt_runtime_lock);
361 * Ensure this RQ takes back all the runtime it lend to its neighbours.
363 static void __disable_runtime(struct rq *rq)
365 struct root_domain *rd = rq->rd;
368 if (unlikely(!scheduler_running))
371 for_each_leaf_rt_rq(rt_rq, rq) {
372 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
376 spin_lock(&rt_b->rt_runtime_lock);
377 spin_lock(&rt_rq->rt_runtime_lock);
379 * Either we're all inf and nobody needs to borrow, or we're
380 * already disabled and thus have nothing to do, or we have
381 * exactly the right amount of runtime to take out.
383 if (rt_rq->rt_runtime == RUNTIME_INF ||
384 rt_rq->rt_runtime == rt_b->rt_runtime)
386 spin_unlock(&rt_rq->rt_runtime_lock);
389 * Calculate the difference between what we started out with
390 * and what we current have, that's the amount of runtime
391 * we lend and now have to reclaim.
393 want = rt_b->rt_runtime - rt_rq->rt_runtime;
396 * Greedy reclaim, take back as much as we can.
398 for_each_cpu(i, rd->span) {
399 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
403 * Can't reclaim from ourselves or disabled runqueues.
405 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
408 spin_lock(&iter->rt_runtime_lock);
410 diff = min_t(s64, iter->rt_runtime, want);
411 iter->rt_runtime -= diff;
414 iter->rt_runtime -= want;
417 spin_unlock(&iter->rt_runtime_lock);
423 spin_lock(&rt_rq->rt_runtime_lock);
425 * We cannot be left wanting - that would mean some runtime
426 * leaked out of the system.
431 * Disable all the borrow logic by pretending we have inf
432 * runtime - in which case borrowing doesn't make sense.
434 rt_rq->rt_runtime = RUNTIME_INF;
435 spin_unlock(&rt_rq->rt_runtime_lock);
436 spin_unlock(&rt_b->rt_runtime_lock);
440 static void disable_runtime(struct rq *rq)
444 spin_lock_irqsave(&rq->lock, flags);
445 __disable_runtime(rq);
446 spin_unlock_irqrestore(&rq->lock, flags);
449 static void __enable_runtime(struct rq *rq)
453 if (unlikely(!scheduler_running))
457 * Reset each runqueue's bandwidth settings
459 for_each_leaf_rt_rq(rt_rq, rq) {
460 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
462 spin_lock(&rt_b->rt_runtime_lock);
463 spin_lock(&rt_rq->rt_runtime_lock);
464 rt_rq->rt_runtime = rt_b->rt_runtime;
466 rt_rq->rt_throttled = 0;
467 spin_unlock(&rt_rq->rt_runtime_lock);
468 spin_unlock(&rt_b->rt_runtime_lock);
472 static void enable_runtime(struct rq *rq)
476 spin_lock_irqsave(&rq->lock, flags);
477 __enable_runtime(rq);
478 spin_unlock_irqrestore(&rq->lock, flags);
481 static int balance_runtime(struct rt_rq *rt_rq)
485 if (rt_rq->rt_time > rt_rq->rt_runtime) {
486 spin_unlock(&rt_rq->rt_runtime_lock);
487 more = do_balance_runtime(rt_rq);
488 spin_lock(&rt_rq->rt_runtime_lock);
493 #else /* !CONFIG_SMP */
494 static inline int balance_runtime(struct rt_rq *rt_rq)
498 #endif /* CONFIG_SMP */
500 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
503 const struct cpumask *span;
505 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
508 span = sched_rt_period_mask();
509 for_each_cpu(i, span) {
511 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
512 struct rq *rq = rq_of_rt_rq(rt_rq);
514 spin_lock(&rq->lock);
515 if (rt_rq->rt_time) {
518 spin_lock(&rt_rq->rt_runtime_lock);
519 if (rt_rq->rt_throttled)
520 balance_runtime(rt_rq);
521 runtime = rt_rq->rt_runtime;
522 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
523 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
524 rt_rq->rt_throttled = 0;
527 if (rt_rq->rt_time || rt_rq->rt_nr_running)
529 spin_unlock(&rt_rq->rt_runtime_lock);
530 } else if (rt_rq->rt_nr_running)
534 sched_rt_rq_enqueue(rt_rq);
535 spin_unlock(&rq->lock);
541 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
543 #ifdef CONFIG_RT_GROUP_SCHED
544 struct rt_rq *rt_rq = group_rt_rq(rt_se);
547 return rt_rq->highest_prio.curr;
550 return rt_task_of(rt_se)->prio;
553 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
555 u64 runtime = sched_rt_runtime(rt_rq);
557 if (rt_rq->rt_throttled)
558 return rt_rq_throttled(rt_rq);
560 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
563 balance_runtime(rt_rq);
564 runtime = sched_rt_runtime(rt_rq);
565 if (runtime == RUNTIME_INF)
568 if (rt_rq->rt_time > runtime) {
569 rt_rq->rt_throttled = 1;
570 if (rt_rq_throttled(rt_rq)) {
571 sched_rt_rq_dequeue(rt_rq);
580 * Update the current task's runtime statistics. Skip current tasks that
581 * are not in our scheduling class.
583 static void update_curr_rt(struct rq *rq)
585 struct task_struct *curr = rq->curr;
586 struct sched_rt_entity *rt_se = &curr->rt;
587 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
590 if (!task_has_rt_policy(curr))
593 delta_exec = rq->clock - curr->se.exec_start;
594 if (unlikely((s64)delta_exec < 0))
597 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
599 curr->se.sum_exec_runtime += delta_exec;
600 account_group_exec_runtime(curr, delta_exec);
602 curr->se.exec_start = rq->clock;
603 cpuacct_charge(curr, delta_exec);
605 if (!rt_bandwidth_enabled())
608 for_each_sched_rt_entity(rt_se) {
609 rt_rq = rt_rq_of_se(rt_se);
611 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
612 spin_lock(&rt_rq->rt_runtime_lock);
613 rt_rq->rt_time += delta_exec;
614 if (sched_rt_runtime_exceeded(rt_rq))
616 spin_unlock(&rt_rq->rt_runtime_lock);
621 #if defined CONFIG_SMP
623 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
625 static inline int next_prio(struct rq *rq)
627 struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
629 if (next && rt_prio(next->prio))
636 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
638 struct rq *rq = rq_of_rt_rq(rt_rq);
640 if (prio < prev_prio) {
643 * If the new task is higher in priority than anything on the
644 * run-queue, we know that the previous high becomes our
647 rt_rq->highest_prio.next = prev_prio;
650 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
652 } else if (prio == rt_rq->highest_prio.curr)
654 * If the next task is equal in priority to the highest on
655 * the run-queue, then we implicitly know that the next highest
656 * task cannot be any lower than current
658 rt_rq->highest_prio.next = prio;
659 else if (prio < rt_rq->highest_prio.next)
661 * Otherwise, we need to recompute next-highest
663 rt_rq->highest_prio.next = next_prio(rq);
667 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
669 struct rq *rq = rq_of_rt_rq(rt_rq);
671 if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
672 rt_rq->highest_prio.next = next_prio(rq);
674 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
675 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
678 #else /* CONFIG_SMP */
681 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
683 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
685 #endif /* CONFIG_SMP */
687 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
689 inc_rt_prio(struct rt_rq *rt_rq, int prio)
691 int prev_prio = rt_rq->highest_prio.curr;
693 if (prio < prev_prio)
694 rt_rq->highest_prio.curr = prio;
696 inc_rt_prio_smp(rt_rq, prio, prev_prio);
700 dec_rt_prio(struct rt_rq *rt_rq, int prio)
702 int prev_prio = rt_rq->highest_prio.curr;
704 if (rt_rq->rt_nr_running) {
706 WARN_ON(prio < prev_prio);
709 * This may have been our highest task, and therefore
710 * we may have some recomputation to do
712 if (prio == prev_prio) {
713 struct rt_prio_array *array = &rt_rq->active;
715 rt_rq->highest_prio.curr =
716 sched_find_first_bit(array->bitmap);
720 rt_rq->highest_prio.curr = MAX_RT_PRIO;
722 dec_rt_prio_smp(rt_rq, prio, prev_prio);
727 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
728 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
730 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
732 #ifdef CONFIG_RT_GROUP_SCHED
735 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
737 if (rt_se_boosted(rt_se))
738 rt_rq->rt_nr_boosted++;
741 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
745 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
747 if (rt_se_boosted(rt_se))
748 rt_rq->rt_nr_boosted--;
750 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
753 #else /* CONFIG_RT_GROUP_SCHED */
756 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
758 start_rt_bandwidth(&def_rt_bandwidth);
762 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
764 #endif /* CONFIG_RT_GROUP_SCHED */
767 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
769 int prio = rt_se_prio(rt_se);
771 WARN_ON(!rt_prio(prio));
772 rt_rq->rt_nr_running++;
774 inc_rt_prio(rt_rq, prio);
775 inc_rt_migration(rt_se, rt_rq);
776 inc_rt_group(rt_se, rt_rq);
780 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
782 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
783 WARN_ON(!rt_rq->rt_nr_running);
784 rt_rq->rt_nr_running--;
786 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
787 dec_rt_migration(rt_se, rt_rq);
788 dec_rt_group(rt_se, rt_rq);
791 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
793 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
794 struct rt_prio_array *array = &rt_rq->active;
795 struct rt_rq *group_rq = group_rt_rq(rt_se);
796 struct list_head *queue = array->queue + rt_se_prio(rt_se);
799 * Don't enqueue the group if its throttled, or when empty.
800 * The latter is a consequence of the former when a child group
801 * get throttled and the current group doesn't have any other
804 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
807 list_add_tail(&rt_se->run_list, queue);
808 __set_bit(rt_se_prio(rt_se), array->bitmap);
810 inc_rt_tasks(rt_se, rt_rq);
813 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
815 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
816 struct rt_prio_array *array = &rt_rq->active;
818 list_del_init(&rt_se->run_list);
819 if (list_empty(array->queue + rt_se_prio(rt_se)))
820 __clear_bit(rt_se_prio(rt_se), array->bitmap);
822 dec_rt_tasks(rt_se, rt_rq);
826 * Because the prio of an upper entry depends on the lower
827 * entries, we must remove entries top - down.
829 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
831 struct sched_rt_entity *back = NULL;
833 for_each_sched_rt_entity(rt_se) {
838 for (rt_se = back; rt_se; rt_se = rt_se->back) {
840 __dequeue_rt_entity(rt_se);
844 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
846 dequeue_rt_stack(rt_se);
847 for_each_sched_rt_entity(rt_se)
848 __enqueue_rt_entity(rt_se);
851 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
853 dequeue_rt_stack(rt_se);
855 for_each_sched_rt_entity(rt_se) {
856 struct rt_rq *rt_rq = group_rt_rq(rt_se);
858 if (rt_rq && rt_rq->rt_nr_running)
859 __enqueue_rt_entity(rt_se);
864 * Adding/removing a task to/from a priority array:
866 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
868 struct sched_rt_entity *rt_se = &p->rt;
873 enqueue_rt_entity(rt_se);
875 if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
876 enqueue_pushable_task(rq, p);
878 inc_cpu_load(rq, p->se.load.weight);
881 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
883 struct sched_rt_entity *rt_se = &p->rt;
886 dequeue_rt_entity(rt_se);
888 dequeue_pushable_task(rq, p);
890 dec_cpu_load(rq, p->se.load.weight);
894 * Put task to the end of the run list without the overhead of dequeue
895 * followed by enqueue.
898 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
900 if (on_rt_rq(rt_se)) {
901 struct rt_prio_array *array = &rt_rq->active;
902 struct list_head *queue = array->queue + rt_se_prio(rt_se);
905 list_move(&rt_se->run_list, queue);
907 list_move_tail(&rt_se->run_list, queue);
911 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
913 struct sched_rt_entity *rt_se = &p->rt;
916 for_each_sched_rt_entity(rt_se) {
917 rt_rq = rt_rq_of_se(rt_se);
918 requeue_rt_entity(rt_rq, rt_se, head);
922 static void yield_task_rt(struct rq *rq)
924 requeue_task_rt(rq, rq->curr, 0);
928 static int find_lowest_rq(struct task_struct *task);
930 static int select_task_rq_rt(struct task_struct *p, int sync)
932 struct rq *rq = task_rq(p);
935 * If the current task is an RT task, then
936 * try to see if we can wake this RT task up on another
937 * runqueue. Otherwise simply start this RT task
938 * on its current runqueue.
940 * We want to avoid overloading runqueues. Even if
941 * the RT task is of higher priority than the current RT task.
942 * RT tasks behave differently than other tasks. If
943 * one gets preempted, we try to push it off to another queue.
944 * So trying to keep a preempting RT task on the same
945 * cache hot CPU will force the running RT task to
946 * a cold CPU. So we waste all the cache for the lower
947 * RT task in hopes of saving some of a RT task
948 * that is just being woken and probably will have
951 if (unlikely(rt_task(rq->curr)) &&
952 (p->rt.nr_cpus_allowed > 1)) {
953 int cpu = find_lowest_rq(p);
955 return (cpu == -1) ? task_cpu(p) : cpu;
959 * Otherwise, just let it ride on the affined RQ and the
960 * post-schedule router will push the preempted task away
965 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
967 if (rq->curr->rt.nr_cpus_allowed == 1)
970 if (p->rt.nr_cpus_allowed != 1
971 && cpupri_find(&rq->rd->cpupri, p, NULL))
974 if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
978 * There appears to be other cpus that can accept
979 * current and none to run 'p', so lets reschedule
980 * to try and push current away:
982 requeue_task_rt(rq, p, 1);
983 resched_task(rq->curr);
986 #endif /* CONFIG_SMP */
989 * Preempt the current task with a newly woken task if needed:
991 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
993 if (p->prio < rq->curr->prio) {
994 resched_task(rq->curr);
1002 * - the newly woken task is of equal priority to the current task
1003 * - the newly woken task is non-migratable while current is migratable
1004 * - current will be preempted on the next reschedule
1006 * we should check to see if current can readily move to a different
1007 * cpu. If so, we will reschedule to allow the push logic to try
1008 * to move current somewhere else, making room for our non-migratable
1011 if (p->prio == rq->curr->prio && !need_resched())
1012 check_preempt_equal_prio(rq, p);
1016 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1017 struct rt_rq *rt_rq)
1019 struct rt_prio_array *array = &rt_rq->active;
1020 struct sched_rt_entity *next = NULL;
1021 struct list_head *queue;
1024 idx = sched_find_first_bit(array->bitmap);
1025 BUG_ON(idx >= MAX_RT_PRIO);
1027 queue = array->queue + idx;
1028 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1033 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1035 struct sched_rt_entity *rt_se;
1036 struct task_struct *p;
1037 struct rt_rq *rt_rq;
1041 if (unlikely(!rt_rq->rt_nr_running))
1044 if (rt_rq_throttled(rt_rq))
1048 rt_se = pick_next_rt_entity(rq, rt_rq);
1050 rt_rq = group_rt_rq(rt_se);
1053 p = rt_task_of(rt_se);
1054 p->se.exec_start = rq->clock;
1059 static inline int has_pushable_tasks(struct rq *rq)
1061 return !plist_head_empty(&rq->rt.pushable_tasks);
1064 static struct task_struct *pick_next_task_rt(struct rq *rq)
1066 struct task_struct *p = _pick_next_task_rt(rq);
1068 /* The running task is never eligible for pushing */
1070 dequeue_pushable_task(rq, p);
1073 * We detect this state here so that we can avoid taking the RQ
1074 * lock again later if there is no need to push
1076 rq->post_schedule = has_pushable_tasks(rq);
1081 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1084 p->se.exec_start = 0;
1087 * The previous task needs to be made eligible for pushing
1088 * if it is still active
1090 if (p->se.on_rq && p->rt.nr_cpus_allowed > 1)
1091 enqueue_pushable_task(rq, p);
1096 /* Only try algorithms three times */
1097 #define RT_MAX_TRIES 3
1099 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1101 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1103 if (!task_running(rq, p) &&
1104 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1105 (p->rt.nr_cpus_allowed > 1))
1110 /* Return the second highest RT task, NULL otherwise */
1111 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1113 struct task_struct *next = NULL;
1114 struct sched_rt_entity *rt_se;
1115 struct rt_prio_array *array;
1116 struct rt_rq *rt_rq;
1119 for_each_leaf_rt_rq(rt_rq, rq) {
1120 array = &rt_rq->active;
1121 idx = sched_find_first_bit(array->bitmap);
1123 if (idx >= MAX_RT_PRIO)
1125 if (next && next->prio < idx)
1127 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1128 struct task_struct *p = rt_task_of(rt_se);
1129 if (pick_rt_task(rq, p, cpu)) {
1135 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1143 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1145 static inline int pick_optimal_cpu(int this_cpu,
1146 const struct cpumask *mask)
1150 /* "this_cpu" is cheaper to preempt than a remote processor */
1151 if ((this_cpu != -1) && cpumask_test_cpu(this_cpu, mask))
1154 first = cpumask_first(mask);
1155 if (first < nr_cpu_ids)
1161 static int find_lowest_rq(struct task_struct *task)
1163 struct sched_domain *sd;
1164 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1165 int this_cpu = smp_processor_id();
1166 int cpu = task_cpu(task);
1167 cpumask_var_t domain_mask;
1169 if (task->rt.nr_cpus_allowed == 1)
1170 return -1; /* No other targets possible */
1172 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1173 return -1; /* No targets found */
1176 * Only consider CPUs that are usable for migration.
1177 * I guess we might want to change cpupri_find() to ignore those
1178 * in the first place.
1180 cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
1183 * At this point we have built a mask of cpus representing the
1184 * lowest priority tasks in the system. Now we want to elect
1185 * the best one based on our affinity and topology.
1187 * We prioritize the last cpu that the task executed on since
1188 * it is most likely cache-hot in that location.
1190 if (cpumask_test_cpu(cpu, lowest_mask))
1194 * Otherwise, we consult the sched_domains span maps to figure
1195 * out which cpu is logically closest to our hot cache data.
1197 if (this_cpu == cpu)
1198 this_cpu = -1; /* Skip this_cpu opt if the same */
1200 if (alloc_cpumask_var(&domain_mask, GFP_ATOMIC)) {
1201 for_each_domain(cpu, sd) {
1202 if (sd->flags & SD_WAKE_AFFINE) {
1205 cpumask_and(domain_mask,
1206 sched_domain_span(sd),
1209 best_cpu = pick_optimal_cpu(this_cpu,
1212 if (best_cpu != -1) {
1213 free_cpumask_var(domain_mask);
1218 free_cpumask_var(domain_mask);
1222 * And finally, if there were no matches within the domains
1223 * just give the caller *something* to work with from the compatible
1226 return pick_optimal_cpu(this_cpu, lowest_mask);
1229 /* Will lock the rq it finds */
1230 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1232 struct rq *lowest_rq = NULL;
1236 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1237 cpu = find_lowest_rq(task);
1239 if ((cpu == -1) || (cpu == rq->cpu))
1242 lowest_rq = cpu_rq(cpu);
1244 /* if the prio of this runqueue changed, try again */
1245 if (double_lock_balance(rq, lowest_rq)) {
1247 * We had to unlock the run queue. In
1248 * the mean time, task could have
1249 * migrated already or had its affinity changed.
1250 * Also make sure that it wasn't scheduled on its rq.
1252 if (unlikely(task_rq(task) != rq ||
1253 !cpumask_test_cpu(lowest_rq->cpu,
1254 &task->cpus_allowed) ||
1255 task_running(rq, task) ||
1258 spin_unlock(&lowest_rq->lock);
1264 /* If this rq is still suitable use it. */
1265 if (lowest_rq->rt.highest_prio.curr > task->prio)
1269 double_unlock_balance(rq, lowest_rq);
1276 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1278 struct task_struct *p;
1280 if (!has_pushable_tasks(rq))
1283 p = plist_first_entry(&rq->rt.pushable_tasks,
1284 struct task_struct, pushable_tasks);
1286 BUG_ON(rq->cpu != task_cpu(p));
1287 BUG_ON(task_current(rq, p));
1288 BUG_ON(p->rt.nr_cpus_allowed <= 1);
1290 BUG_ON(!p->se.on_rq);
1291 BUG_ON(!rt_task(p));
1297 * If the current CPU has more than one RT task, see if the non
1298 * running task can migrate over to a CPU that is running a task
1299 * of lesser priority.
1301 static int push_rt_task(struct rq *rq)
1303 struct task_struct *next_task;
1304 struct rq *lowest_rq;
1306 if (!rq->rt.overloaded)
1309 next_task = pick_next_pushable_task(rq);
1314 if (unlikely(next_task == rq->curr)) {
1320 * It's possible that the next_task slipped in of
1321 * higher priority than current. If that's the case
1322 * just reschedule current.
1324 if (unlikely(next_task->prio < rq->curr->prio)) {
1325 resched_task(rq->curr);
1329 /* We might release rq lock */
1330 get_task_struct(next_task);
1332 /* find_lock_lowest_rq locks the rq if found */
1333 lowest_rq = find_lock_lowest_rq(next_task, rq);
1335 struct task_struct *task;
1337 * find lock_lowest_rq releases rq->lock
1338 * so it is possible that next_task has migrated.
1340 * We need to make sure that the task is still on the same
1341 * run-queue and is also still the next task eligible for
1344 task = pick_next_pushable_task(rq);
1345 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1347 * If we get here, the task hasnt moved at all, but
1348 * it has failed to push. We will not try again,
1349 * since the other cpus will pull from us when they
1352 dequeue_pushable_task(rq, next_task);
1357 /* No more tasks, just exit */
1361 * Something has shifted, try again.
1363 put_task_struct(next_task);
1368 deactivate_task(rq, next_task, 0);
1369 set_task_cpu(next_task, lowest_rq->cpu);
1370 activate_task(lowest_rq, next_task, 0);
1372 resched_task(lowest_rq->curr);
1374 double_unlock_balance(rq, lowest_rq);
1377 put_task_struct(next_task);
1382 static void push_rt_tasks(struct rq *rq)
1384 /* push_rt_task will return true if it moved an RT */
1385 while (push_rt_task(rq))
1389 static int pull_rt_task(struct rq *this_rq)
1391 int this_cpu = this_rq->cpu, ret = 0, cpu;
1392 struct task_struct *p;
1395 if (likely(!rt_overloaded(this_rq)))
1398 for_each_cpu(cpu, this_rq->rd->rto_mask) {
1399 if (this_cpu == cpu)
1402 src_rq = cpu_rq(cpu);
1405 * Don't bother taking the src_rq->lock if the next highest
1406 * task is known to be lower-priority than our current task.
1407 * This may look racy, but if this value is about to go
1408 * logically higher, the src_rq will push this task away.
1409 * And if its going logically lower, we do not care
1411 if (src_rq->rt.highest_prio.next >=
1412 this_rq->rt.highest_prio.curr)
1416 * We can potentially drop this_rq's lock in
1417 * double_lock_balance, and another CPU could
1420 double_lock_balance(this_rq, src_rq);
1423 * Are there still pullable RT tasks?
1425 if (src_rq->rt.rt_nr_running <= 1)
1428 p = pick_next_highest_task_rt(src_rq, this_cpu);
1431 * Do we have an RT task that preempts
1432 * the to-be-scheduled task?
1434 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1435 WARN_ON(p == src_rq->curr);
1436 WARN_ON(!p->se.on_rq);
1439 * There's a chance that p is higher in priority
1440 * than what's currently running on its cpu.
1441 * This is just that p is wakeing up and hasn't
1442 * had a chance to schedule. We only pull
1443 * p if it is lower in priority than the
1444 * current task on the run queue
1446 if (p->prio < src_rq->curr->prio)
1451 deactivate_task(src_rq, p, 0);
1452 set_task_cpu(p, this_cpu);
1453 activate_task(this_rq, p, 0);
1455 * We continue with the search, just in
1456 * case there's an even higher prio task
1457 * in another runqueue. (low likelyhood
1462 double_unlock_balance(this_rq, src_rq);
1468 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1470 /* Try to pull RT tasks here if we lower this rq's prio */
1471 if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1475 static void post_schedule_rt(struct rq *rq)
1481 * If we are not running and we are not going to reschedule soon, we should
1482 * try to push tasks away now
1484 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1486 if (!task_running(rq, p) &&
1487 !test_tsk_need_resched(rq->curr) &&
1488 has_pushable_tasks(rq) &&
1489 p->rt.nr_cpus_allowed > 1)
1493 static unsigned long
1494 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1495 unsigned long max_load_move,
1496 struct sched_domain *sd, enum cpu_idle_type idle,
1497 int *all_pinned, int *this_best_prio)
1499 /* don't touch RT tasks */
1504 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1505 struct sched_domain *sd, enum cpu_idle_type idle)
1507 /* don't touch RT tasks */
1511 static void set_cpus_allowed_rt(struct task_struct *p,
1512 const struct cpumask *new_mask)
1514 int weight = cpumask_weight(new_mask);
1516 BUG_ON(!rt_task(p));
1519 * Update the migration status of the RQ if we have an RT task
1520 * which is running AND changing its weight value.
1522 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1523 struct rq *rq = task_rq(p);
1525 if (!task_current(rq, p)) {
1527 * Make sure we dequeue this task from the pushable list
1528 * before going further. It will either remain off of
1529 * the list because we are no longer pushable, or it
1532 if (p->rt.nr_cpus_allowed > 1)
1533 dequeue_pushable_task(rq, p);
1536 * Requeue if our weight is changing and still > 1
1539 enqueue_pushable_task(rq, p);
1543 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1544 rq->rt.rt_nr_migratory++;
1545 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1546 BUG_ON(!rq->rt.rt_nr_migratory);
1547 rq->rt.rt_nr_migratory--;
1550 update_rt_migration(&rq->rt);
1553 cpumask_copy(&p->cpus_allowed, new_mask);
1554 p->rt.nr_cpus_allowed = weight;
1557 /* Assumes rq->lock is held */
1558 static void rq_online_rt(struct rq *rq)
1560 if (rq->rt.overloaded)
1561 rt_set_overload(rq);
1563 __enable_runtime(rq);
1565 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1568 /* Assumes rq->lock is held */
1569 static void rq_offline_rt(struct rq *rq)
1571 if (rq->rt.overloaded)
1572 rt_clear_overload(rq);
1574 __disable_runtime(rq);
1576 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1580 * When switch from the rt queue, we bring ourselves to a position
1581 * that we might want to pull RT tasks from other runqueues.
1583 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1587 * If there are other RT tasks then we will reschedule
1588 * and the scheduling of the other RT tasks will handle
1589 * the balancing. But if we are the last RT task
1590 * we may need to handle the pulling of RT tasks
1593 if (!rq->rt.rt_nr_running)
1597 static inline void init_sched_rt_class(void)
1601 for_each_possible_cpu(i)
1602 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1603 GFP_KERNEL, cpu_to_node(i));
1605 #endif /* CONFIG_SMP */
1608 * When switching a task to RT, we may overload the runqueue
1609 * with RT tasks. In this case we try to push them off to
1612 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1615 int check_resched = 1;
1618 * If we are already running, then there's nothing
1619 * that needs to be done. But if we are not running
1620 * we may need to preempt the current running task.
1621 * If that current running task is also an RT task
1622 * then see if we can move to another run queue.
1626 if (rq->rt.overloaded && push_rt_task(rq) &&
1627 /* Don't resched if we changed runqueues */
1630 #endif /* CONFIG_SMP */
1631 if (check_resched && p->prio < rq->curr->prio)
1632 resched_task(rq->curr);
1637 * Priority of the task has changed. This may cause
1638 * us to initiate a push or pull.
1640 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1641 int oldprio, int running)
1646 * If our priority decreases while running, we
1647 * may need to pull tasks to this runqueue.
1649 if (oldprio < p->prio)
1652 * If there's a higher priority task waiting to run
1653 * then reschedule. Note, the above pull_rt_task
1654 * can release the rq lock and p could migrate.
1655 * Only reschedule if p is still on the same runqueue.
1657 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1660 /* For UP simply resched on drop of prio */
1661 if (oldprio < p->prio)
1663 #endif /* CONFIG_SMP */
1666 * This task is not running, but if it is
1667 * greater than the current running task
1670 if (p->prio < rq->curr->prio)
1671 resched_task(rq->curr);
1675 static void watchdog(struct rq *rq, struct task_struct *p)
1677 unsigned long soft, hard;
1682 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1683 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1685 if (soft != RLIM_INFINITY) {
1689 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1690 if (p->rt.timeout > next)
1691 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1695 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1702 * RR tasks need a special form of timeslice management.
1703 * FIFO tasks have no timeslices.
1705 if (p->policy != SCHED_RR)
1708 if (--p->rt.time_slice)
1711 p->rt.time_slice = DEF_TIMESLICE;
1714 * Requeue to the end of queue if we are not the only element
1717 if (p->rt.run_list.prev != p->rt.run_list.next) {
1718 requeue_task_rt(rq, p, 0);
1719 set_tsk_need_resched(p);
1723 static void set_curr_task_rt(struct rq *rq)
1725 struct task_struct *p = rq->curr;
1727 p->se.exec_start = rq->clock;
1729 /* The running task is never eligible for pushing */
1730 dequeue_pushable_task(rq, p);
1733 static const struct sched_class rt_sched_class = {
1734 .next = &fair_sched_class,
1735 .enqueue_task = enqueue_task_rt,
1736 .dequeue_task = dequeue_task_rt,
1737 .yield_task = yield_task_rt,
1739 .check_preempt_curr = check_preempt_curr_rt,
1741 .pick_next_task = pick_next_task_rt,
1742 .put_prev_task = put_prev_task_rt,
1745 .select_task_rq = select_task_rq_rt,
1747 .load_balance = load_balance_rt,
1748 .move_one_task = move_one_task_rt,
1749 .set_cpus_allowed = set_cpus_allowed_rt,
1750 .rq_online = rq_online_rt,
1751 .rq_offline = rq_offline_rt,
1752 .pre_schedule = pre_schedule_rt,
1753 .post_schedule = post_schedule_rt,
1754 .task_wake_up = task_wake_up_rt,
1755 .switched_from = switched_from_rt,
1758 .set_curr_task = set_curr_task_rt,
1759 .task_tick = task_tick_rt,
1761 .prio_changed = prio_changed_rt,
1762 .switched_to = switched_to_rt,
1765 #ifdef CONFIG_SCHED_DEBUG
1766 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1768 static void print_rt_stats(struct seq_file *m, int cpu)
1770 struct rt_rq *rt_rq;
1773 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1774 print_rt_rq(m, cpu, rt_rq);
1777 #endif /* CONFIG_SCHED_DEBUG */