2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
8 static inline int rt_overloaded(struct rq *rq)
10 return atomic_read(&rq->rd->rto_count);
13 static inline void rt_set_overload(struct rq *rq)
18 cpu_set(rq->cpu, rq->rd->rto_mask);
20 * Make sure the mask is visible before we set
21 * the overload count. That is checked to determine
22 * if we should look at the mask. It would be a shame
23 * if we looked at the mask, but the mask was not
27 atomic_inc(&rq->rd->rto_count);
30 static inline void rt_clear_overload(struct rq *rq)
35 /* the order here really doesn't matter */
36 atomic_dec(&rq->rd->rto_count);
37 cpu_clear(rq->cpu, rq->rd->rto_mask);
40 static void update_rt_migration(struct rq *rq)
42 if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
43 if (!rq->rt.overloaded) {
45 rq->rt.overloaded = 1;
47 } else if (rq->rt.overloaded) {
48 rt_clear_overload(rq);
49 rq->rt.overloaded = 0;
52 #endif /* CONFIG_SMP */
54 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
56 return container_of(rt_se, struct task_struct, rt);
59 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
61 return !list_empty(&rt_se->run_list);
64 #ifdef CONFIG_RT_GROUP_SCHED
66 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
71 return rt_rq->rt_runtime;
74 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
76 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
79 #define for_each_leaf_rt_rq(rt_rq, rq) \
80 list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
82 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
87 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
92 #define for_each_sched_rt_entity(rt_se) \
93 for (; rt_se; rt_se = rt_se->parent)
95 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
100 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
101 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
103 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
105 struct sched_rt_entity *rt_se = rt_rq->rt_se;
107 if (rt_se && !on_rt_rq(rt_se) && rt_rq->rt_nr_running) {
108 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
110 enqueue_rt_entity(rt_se);
111 if (rt_rq->highest_prio < curr->prio)
116 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
118 struct sched_rt_entity *rt_se = rt_rq->rt_se;
120 if (rt_se && on_rt_rq(rt_se))
121 dequeue_rt_entity(rt_se);
124 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
126 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
129 static int rt_se_boosted(struct sched_rt_entity *rt_se)
131 struct rt_rq *rt_rq = group_rt_rq(rt_se);
132 struct task_struct *p;
135 return !!rt_rq->rt_nr_boosted;
137 p = rt_task_of(rt_se);
138 return p->prio != p->normal_prio;
142 static inline cpumask_t sched_rt_period_mask(void)
144 return cpu_rq(smp_processor_id())->rd->span;
147 static inline cpumask_t sched_rt_period_mask(void)
149 return cpu_online_map;
154 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
156 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
159 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
161 return &rt_rq->tg->rt_bandwidth;
166 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
168 return rt_rq->rt_runtime;
171 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
173 return ktime_to_ns(def_rt_bandwidth.rt_period);
176 #define for_each_leaf_rt_rq(rt_rq, rq) \
177 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
179 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
181 return container_of(rt_rq, struct rq, rt);
184 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
186 struct task_struct *p = rt_task_of(rt_se);
187 struct rq *rq = task_rq(p);
192 #define for_each_sched_rt_entity(rt_se) \
193 for (; rt_se; rt_se = NULL)
195 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
200 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
204 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
208 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
210 return rt_rq->rt_throttled;
213 static inline cpumask_t sched_rt_period_mask(void)
215 return cpu_online_map;
219 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
221 return &cpu_rq(cpu)->rt;
224 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
226 return &def_rt_bandwidth;
232 static int do_balance_runtime(struct rt_rq *rt_rq)
234 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
235 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
236 int i, weight, more = 0;
239 weight = cpus_weight(rd->span);
241 spin_lock(&rt_b->rt_runtime_lock);
242 rt_period = ktime_to_ns(rt_b->rt_period);
243 for_each_cpu_mask(i, rd->span) {
244 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
250 spin_lock(&iter->rt_runtime_lock);
251 if (iter->rt_runtime == RUNTIME_INF)
254 diff = iter->rt_runtime - iter->rt_time;
256 do_div(diff, weight);
257 if (rt_rq->rt_runtime + diff > rt_period)
258 diff = rt_period - rt_rq->rt_runtime;
259 iter->rt_runtime -= diff;
260 rt_rq->rt_runtime += diff;
262 if (rt_rq->rt_runtime == rt_period) {
263 spin_unlock(&iter->rt_runtime_lock);
268 spin_unlock(&iter->rt_runtime_lock);
270 spin_unlock(&rt_b->rt_runtime_lock);
275 static void __disable_runtime(struct rq *rq)
277 struct root_domain *rd = rq->rd;
280 if (unlikely(!scheduler_running))
283 for_each_leaf_rt_rq(rt_rq, rq) {
284 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
288 spin_lock(&rt_b->rt_runtime_lock);
289 spin_lock(&rt_rq->rt_runtime_lock);
290 if (rt_rq->rt_runtime == RUNTIME_INF ||
291 rt_rq->rt_runtime == rt_b->rt_runtime)
293 spin_unlock(&rt_rq->rt_runtime_lock);
295 want = rt_b->rt_runtime - rt_rq->rt_runtime;
297 for_each_cpu_mask(i, rd->span) {
298 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
304 spin_lock(&iter->rt_runtime_lock);
306 diff = min_t(s64, iter->rt_runtime, want);
307 iter->rt_runtime -= diff;
310 iter->rt_runtime -= want;
313 spin_unlock(&iter->rt_runtime_lock);
319 spin_lock(&rt_rq->rt_runtime_lock);
322 rt_rq->rt_runtime = RUNTIME_INF;
323 spin_unlock(&rt_rq->rt_runtime_lock);
324 spin_unlock(&rt_b->rt_runtime_lock);
328 static void disable_runtime(struct rq *rq)
332 spin_lock_irqsave(&rq->lock, flags);
333 __disable_runtime(rq);
334 spin_unlock_irqrestore(&rq->lock, flags);
337 static void __enable_runtime(struct rq *rq)
339 struct root_domain *rd = rq->rd;
342 if (unlikely(!scheduler_running))
345 for_each_leaf_rt_rq(rt_rq, rq) {
346 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
348 spin_lock(&rt_b->rt_runtime_lock);
349 spin_lock(&rt_rq->rt_runtime_lock);
350 rt_rq->rt_runtime = rt_b->rt_runtime;
352 spin_unlock(&rt_rq->rt_runtime_lock);
353 spin_unlock(&rt_b->rt_runtime_lock);
357 static void enable_runtime(struct rq *rq)
361 spin_lock_irqsave(&rq->lock, flags);
362 __enable_runtime(rq);
363 spin_unlock_irqrestore(&rq->lock, flags);
366 static int balance_runtime(struct rt_rq *rt_rq)
370 if (rt_rq->rt_time > rt_rq->rt_runtime) {
371 spin_unlock(&rt_rq->rt_runtime_lock);
372 more = do_balance_runtime(rt_rq);
373 spin_lock(&rt_rq->rt_runtime_lock);
379 static inline int balance_runtime(struct rt_rq *rt_rq)
385 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
390 if (rt_b->rt_runtime == RUNTIME_INF)
393 span = sched_rt_period_mask();
394 for_each_cpu_mask(i, span) {
396 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
397 struct rq *rq = rq_of_rt_rq(rt_rq);
399 spin_lock(&rq->lock);
400 if (rt_rq->rt_time) {
403 spin_lock(&rt_rq->rt_runtime_lock);
404 if (rt_rq->rt_throttled)
405 balance_runtime(rt_rq);
406 runtime = rt_rq->rt_runtime;
407 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
408 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
409 rt_rq->rt_throttled = 0;
412 if (rt_rq->rt_time || rt_rq->rt_nr_running)
414 spin_unlock(&rt_rq->rt_runtime_lock);
415 } else if (rt_rq->rt_nr_running)
419 sched_rt_rq_enqueue(rt_rq);
420 spin_unlock(&rq->lock);
426 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
428 #ifdef CONFIG_RT_GROUP_SCHED
429 struct rt_rq *rt_rq = group_rt_rq(rt_se);
432 return rt_rq->highest_prio;
435 return rt_task_of(rt_se)->prio;
438 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
440 u64 runtime = sched_rt_runtime(rt_rq);
442 if (runtime == RUNTIME_INF)
445 if (rt_rq->rt_throttled)
446 return rt_rq_throttled(rt_rq);
448 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
451 balance_runtime(rt_rq);
452 runtime = sched_rt_runtime(rt_rq);
453 if (runtime == RUNTIME_INF)
456 if (rt_rq->rt_time > runtime) {
457 rt_rq->rt_throttled = 1;
458 if (rt_rq_throttled(rt_rq)) {
459 sched_rt_rq_dequeue(rt_rq);
468 * Update the current task's runtime statistics. Skip current tasks that
469 * are not in our scheduling class.
471 static void update_curr_rt(struct rq *rq)
473 struct task_struct *curr = rq->curr;
474 struct sched_rt_entity *rt_se = &curr->rt;
475 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
478 if (!task_has_rt_policy(curr))
481 delta_exec = rq->clock - curr->se.exec_start;
482 if (unlikely((s64)delta_exec < 0))
485 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
487 curr->se.sum_exec_runtime += delta_exec;
488 curr->se.exec_start = rq->clock;
489 cpuacct_charge(curr, delta_exec);
491 for_each_sched_rt_entity(rt_se) {
492 rt_rq = rt_rq_of_se(rt_se);
494 spin_lock(&rt_rq->rt_runtime_lock);
495 rt_rq->rt_time += delta_exec;
496 if (sched_rt_runtime_exceeded(rt_rq))
498 spin_unlock(&rt_rq->rt_runtime_lock);
503 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
505 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
506 rt_rq->rt_nr_running++;
507 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
508 if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
509 struct rq *rq = rq_of_rt_rq(rt_rq);
511 rt_rq->highest_prio = rt_se_prio(rt_se);
514 cpupri_set(&rq->rd->cpupri, rq->cpu,
520 if (rt_se->nr_cpus_allowed > 1) {
521 struct rq *rq = rq_of_rt_rq(rt_rq);
523 rq->rt.rt_nr_migratory++;
526 update_rt_migration(rq_of_rt_rq(rt_rq));
528 #ifdef CONFIG_RT_GROUP_SCHED
529 if (rt_se_boosted(rt_se))
530 rt_rq->rt_nr_boosted++;
533 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
535 start_rt_bandwidth(&def_rt_bandwidth);
540 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
543 int highest_prio = rt_rq->highest_prio;
546 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
547 WARN_ON(!rt_rq->rt_nr_running);
548 rt_rq->rt_nr_running--;
549 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
550 if (rt_rq->rt_nr_running) {
551 struct rt_prio_array *array;
553 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
554 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
556 array = &rt_rq->active;
557 rt_rq->highest_prio =
558 sched_find_first_bit(array->bitmap);
559 } /* otherwise leave rq->highest prio alone */
561 rt_rq->highest_prio = MAX_RT_PRIO;
564 if (rt_se->nr_cpus_allowed > 1) {
565 struct rq *rq = rq_of_rt_rq(rt_rq);
566 rq->rt.rt_nr_migratory--;
569 if (rt_rq->highest_prio != highest_prio) {
570 struct rq *rq = rq_of_rt_rq(rt_rq);
573 cpupri_set(&rq->rd->cpupri, rq->cpu,
574 rt_rq->highest_prio);
577 update_rt_migration(rq_of_rt_rq(rt_rq));
578 #endif /* CONFIG_SMP */
579 #ifdef CONFIG_RT_GROUP_SCHED
580 if (rt_se_boosted(rt_se))
581 rt_rq->rt_nr_boosted--;
583 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
587 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
589 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
590 struct rt_prio_array *array = &rt_rq->active;
591 struct rt_rq *group_rq = group_rt_rq(rt_se);
592 struct list_head *queue = array->queue + rt_se_prio(rt_se);
595 * Don't enqueue the group if its throttled, or when empty.
596 * The latter is a consequence of the former when a child group
597 * get throttled and the current group doesn't have any other
600 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
603 if (rt_se->nr_cpus_allowed == 1)
604 list_add(&rt_se->run_list, queue);
606 list_add_tail(&rt_se->run_list, queue);
608 __set_bit(rt_se_prio(rt_se), array->bitmap);
610 inc_rt_tasks(rt_se, rt_rq);
613 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
615 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
616 struct rt_prio_array *array = &rt_rq->active;
618 list_del_init(&rt_se->run_list);
619 if (list_empty(array->queue + rt_se_prio(rt_se)))
620 __clear_bit(rt_se_prio(rt_se), array->bitmap);
622 dec_rt_tasks(rt_se, rt_rq);
626 * Because the prio of an upper entry depends on the lower
627 * entries, we must remove entries top - down.
629 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
631 struct sched_rt_entity *back = NULL;
633 for_each_sched_rt_entity(rt_se) {
638 for (rt_se = back; rt_se; rt_se = rt_se->back) {
640 __dequeue_rt_entity(rt_se);
644 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
646 dequeue_rt_stack(rt_se);
647 for_each_sched_rt_entity(rt_se)
648 __enqueue_rt_entity(rt_se);
651 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
653 dequeue_rt_stack(rt_se);
655 for_each_sched_rt_entity(rt_se) {
656 struct rt_rq *rt_rq = group_rt_rq(rt_se);
658 if (rt_rq && rt_rq->rt_nr_running)
659 __enqueue_rt_entity(rt_se);
664 * Adding/removing a task to/from a priority array:
666 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
668 struct sched_rt_entity *rt_se = &p->rt;
673 enqueue_rt_entity(rt_se);
676 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
678 struct sched_rt_entity *rt_se = &p->rt;
681 dequeue_rt_entity(rt_se);
685 * Put task to the end of the run list without the overhead of dequeue
686 * followed by enqueue.
689 void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
691 struct rt_prio_array *array = &rt_rq->active;
692 struct list_head *queue = array->queue + rt_se_prio(rt_se);
694 if (on_rt_rq(rt_se)) {
695 list_del_init(&rt_se->run_list);
696 list_add_tail(&rt_se->run_list,
697 array->queue + rt_se_prio(rt_se));
701 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
703 struct sched_rt_entity *rt_se = &p->rt;
706 for_each_sched_rt_entity(rt_se) {
707 rt_rq = rt_rq_of_se(rt_se);
708 requeue_rt_entity(rt_rq, rt_se);
712 static void yield_task_rt(struct rq *rq)
714 requeue_task_rt(rq, rq->curr);
718 static int find_lowest_rq(struct task_struct *task);
720 static int select_task_rq_rt(struct task_struct *p, int sync)
722 struct rq *rq = task_rq(p);
725 * If the current task is an RT task, then
726 * try to see if we can wake this RT task up on another
727 * runqueue. Otherwise simply start this RT task
728 * on its current runqueue.
730 * We want to avoid overloading runqueues. Even if
731 * the RT task is of higher priority than the current RT task.
732 * RT tasks behave differently than other tasks. If
733 * one gets preempted, we try to push it off to another queue.
734 * So trying to keep a preempting RT task on the same
735 * cache hot CPU will force the running RT task to
736 * a cold CPU. So we waste all the cache for the lower
737 * RT task in hopes of saving some of a RT task
738 * that is just being woken and probably will have
741 if (unlikely(rt_task(rq->curr)) &&
742 (p->rt.nr_cpus_allowed > 1)) {
743 int cpu = find_lowest_rq(p);
745 return (cpu == -1) ? task_cpu(p) : cpu;
749 * Otherwise, just let it ride on the affined RQ and the
750 * post-schedule router will push the preempted task away
754 #endif /* CONFIG_SMP */
757 * Preempt the current task with a newly woken task if needed:
759 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
761 if (p->prio < rq->curr->prio) {
762 resched_task(rq->curr);
770 * - the newly woken task is of equal priority to the current task
771 * - the newly woken task is non-migratable while current is migratable
772 * - current will be preempted on the next reschedule
774 * we should check to see if current can readily move to a different
775 * cpu. If so, we will reschedule to allow the push logic to try
776 * to move current somewhere else, making room for our non-migratable
779 if((p->prio == rq->curr->prio)
780 && p->rt.nr_cpus_allowed == 1
781 && rq->curr->rt.nr_cpus_allowed != 1) {
784 if (cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
786 * There appears to be other cpus that can accept
787 * current, so lets reschedule to try and push it away
789 resched_task(rq->curr);
794 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
797 struct rt_prio_array *array = &rt_rq->active;
798 struct sched_rt_entity *next = NULL;
799 struct list_head *queue;
802 idx = sched_find_first_bit(array->bitmap);
803 BUG_ON(idx >= MAX_RT_PRIO);
805 queue = array->queue + idx;
806 next = list_entry(queue->next, struct sched_rt_entity, run_list);
811 static struct task_struct *pick_next_task_rt(struct rq *rq)
813 struct sched_rt_entity *rt_se;
814 struct task_struct *p;
819 if (unlikely(!rt_rq->rt_nr_running))
822 if (rt_rq_throttled(rt_rq))
826 rt_se = pick_next_rt_entity(rq, rt_rq);
828 rt_rq = group_rt_rq(rt_se);
831 p = rt_task_of(rt_se);
832 p->se.exec_start = rq->clock;
836 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
839 p->se.exec_start = 0;
844 /* Only try algorithms three times */
845 #define RT_MAX_TRIES 3
847 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
848 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
850 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
852 if (!task_running(rq, p) &&
853 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
854 (p->rt.nr_cpus_allowed > 1))
859 /* Return the second highest RT task, NULL otherwise */
860 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
862 struct task_struct *next = NULL;
863 struct sched_rt_entity *rt_se;
864 struct rt_prio_array *array;
868 for_each_leaf_rt_rq(rt_rq, rq) {
869 array = &rt_rq->active;
870 idx = sched_find_first_bit(array->bitmap);
872 if (idx >= MAX_RT_PRIO)
874 if (next && next->prio < idx)
876 list_for_each_entry(rt_se, array->queue + idx, run_list) {
877 struct task_struct *p = rt_task_of(rt_se);
878 if (pick_rt_task(rq, p, cpu)) {
884 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
892 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
894 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
898 /* "this_cpu" is cheaper to preempt than a remote processor */
899 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
902 first = first_cpu(*mask);
903 if (first != NR_CPUS)
909 static int find_lowest_rq(struct task_struct *task)
911 struct sched_domain *sd;
912 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
913 int this_cpu = smp_processor_id();
914 int cpu = task_cpu(task);
916 if (task->rt.nr_cpus_allowed == 1)
917 return -1; /* No other targets possible */
919 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
920 return -1; /* No targets found */
923 * At this point we have built a mask of cpus representing the
924 * lowest priority tasks in the system. Now we want to elect
925 * the best one based on our affinity and topology.
927 * We prioritize the last cpu that the task executed on since
928 * it is most likely cache-hot in that location.
930 if (cpu_isset(cpu, *lowest_mask))
934 * Otherwise, we consult the sched_domains span maps to figure
935 * out which cpu is logically closest to our hot cache data.
938 this_cpu = -1; /* Skip this_cpu opt if the same */
940 for_each_domain(cpu, sd) {
941 if (sd->flags & SD_WAKE_AFFINE) {
942 cpumask_t domain_mask;
945 cpus_and(domain_mask, sd->span, *lowest_mask);
947 best_cpu = pick_optimal_cpu(this_cpu,
955 * And finally, if there were no matches within the domains
956 * just give the caller *something* to work with from the compatible
959 return pick_optimal_cpu(this_cpu, lowest_mask);
962 /* Will lock the rq it finds */
963 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
965 struct rq *lowest_rq = NULL;
969 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
970 cpu = find_lowest_rq(task);
972 if ((cpu == -1) || (cpu == rq->cpu))
975 lowest_rq = cpu_rq(cpu);
977 /* if the prio of this runqueue changed, try again */
978 if (double_lock_balance(rq, lowest_rq)) {
980 * We had to unlock the run queue. In
981 * the mean time, task could have
982 * migrated already or had its affinity changed.
983 * Also make sure that it wasn't scheduled on its rq.
985 if (unlikely(task_rq(task) != rq ||
986 !cpu_isset(lowest_rq->cpu,
987 task->cpus_allowed) ||
988 task_running(rq, task) ||
991 spin_unlock(&lowest_rq->lock);
997 /* If this rq is still suitable use it. */
998 if (lowest_rq->rt.highest_prio > task->prio)
1002 spin_unlock(&lowest_rq->lock);
1010 * If the current CPU has more than one RT task, see if the non
1011 * running task can migrate over to a CPU that is running a task
1012 * of lesser priority.
1014 static int push_rt_task(struct rq *rq)
1016 struct task_struct *next_task;
1017 struct rq *lowest_rq;
1019 int paranoid = RT_MAX_TRIES;
1021 if (!rq->rt.overloaded)
1024 next_task = pick_next_highest_task_rt(rq, -1);
1029 if (unlikely(next_task == rq->curr)) {
1035 * It's possible that the next_task slipped in of
1036 * higher priority than current. If that's the case
1037 * just reschedule current.
1039 if (unlikely(next_task->prio < rq->curr->prio)) {
1040 resched_task(rq->curr);
1044 /* We might release rq lock */
1045 get_task_struct(next_task);
1047 /* find_lock_lowest_rq locks the rq if found */
1048 lowest_rq = find_lock_lowest_rq(next_task, rq);
1050 struct task_struct *task;
1052 * find lock_lowest_rq releases rq->lock
1053 * so it is possible that next_task has changed.
1054 * If it has, then try again.
1056 task = pick_next_highest_task_rt(rq, -1);
1057 if (unlikely(task != next_task) && task && paranoid--) {
1058 put_task_struct(next_task);
1065 deactivate_task(rq, next_task, 0);
1066 set_task_cpu(next_task, lowest_rq->cpu);
1067 activate_task(lowest_rq, next_task, 0);
1069 resched_task(lowest_rq->curr);
1071 spin_unlock(&lowest_rq->lock);
1075 put_task_struct(next_task);
1081 * TODO: Currently we just use the second highest prio task on
1082 * the queue, and stop when it can't migrate (or there's
1083 * no more RT tasks). There may be a case where a lower
1084 * priority RT task has a different affinity than the
1085 * higher RT task. In this case the lower RT task could
1086 * possibly be able to migrate where as the higher priority
1087 * RT task could not. We currently ignore this issue.
1088 * Enhancements are welcome!
1090 static void push_rt_tasks(struct rq *rq)
1092 /* push_rt_task will return true if it moved an RT */
1093 while (push_rt_task(rq))
1097 static int pull_rt_task(struct rq *this_rq)
1099 int this_cpu = this_rq->cpu, ret = 0, cpu;
1100 struct task_struct *p, *next;
1103 if (likely(!rt_overloaded(this_rq)))
1106 next = pick_next_task_rt(this_rq);
1108 for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
1109 if (this_cpu == cpu)
1112 src_rq = cpu_rq(cpu);
1114 * We can potentially drop this_rq's lock in
1115 * double_lock_balance, and another CPU could
1116 * steal our next task - hence we must cause
1117 * the caller to recalculate the next task
1120 if (double_lock_balance(this_rq, src_rq)) {
1121 struct task_struct *old_next = next;
1123 next = pick_next_task_rt(this_rq);
1124 if (next != old_next)
1129 * Are there still pullable RT tasks?
1131 if (src_rq->rt.rt_nr_running <= 1)
1134 p = pick_next_highest_task_rt(src_rq, this_cpu);
1137 * Do we have an RT task that preempts
1138 * the to-be-scheduled task?
1140 if (p && (!next || (p->prio < next->prio))) {
1141 WARN_ON(p == src_rq->curr);
1142 WARN_ON(!p->se.on_rq);
1145 * There's a chance that p is higher in priority
1146 * than what's currently running on its cpu.
1147 * This is just that p is wakeing up and hasn't
1148 * had a chance to schedule. We only pull
1149 * p if it is lower in priority than the
1150 * current task on the run queue or
1151 * this_rq next task is lower in prio than
1152 * the current task on that rq.
1154 if (p->prio < src_rq->curr->prio ||
1155 (next && next->prio < src_rq->curr->prio))
1160 deactivate_task(src_rq, p, 0);
1161 set_task_cpu(p, this_cpu);
1162 activate_task(this_rq, p, 0);
1164 * We continue with the search, just in
1165 * case there's an even higher prio task
1166 * in another runqueue. (low likelyhood
1169 * Update next so that we won't pick a task
1170 * on another cpu with a priority lower (or equal)
1171 * than the one we just picked.
1177 spin_unlock(&src_rq->lock);
1183 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1185 /* Try to pull RT tasks here if we lower this rq's prio */
1186 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1190 static void post_schedule_rt(struct rq *rq)
1193 * If we have more than one rt_task queued, then
1194 * see if we can push the other rt_tasks off to other CPUS.
1195 * Note we may release the rq lock, and since
1196 * the lock was owned by prev, we need to release it
1197 * first via finish_lock_switch and then reaquire it here.
1199 if (unlikely(rq->rt.overloaded)) {
1200 spin_lock_irq(&rq->lock);
1202 spin_unlock_irq(&rq->lock);
1207 * If we are not running and we are not going to reschedule soon, we should
1208 * try to push tasks away now
1210 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1212 if (!task_running(rq, p) &&
1213 !test_tsk_need_resched(rq->curr) &&
1218 static unsigned long
1219 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1220 unsigned long max_load_move,
1221 struct sched_domain *sd, enum cpu_idle_type idle,
1222 int *all_pinned, int *this_best_prio)
1224 /* don't touch RT tasks */
1229 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1230 struct sched_domain *sd, enum cpu_idle_type idle)
1232 /* don't touch RT tasks */
1236 static void set_cpus_allowed_rt(struct task_struct *p,
1237 const cpumask_t *new_mask)
1239 int weight = cpus_weight(*new_mask);
1241 BUG_ON(!rt_task(p));
1244 * Update the migration status of the RQ if we have an RT task
1245 * which is running AND changing its weight value.
1247 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1248 struct rq *rq = task_rq(p);
1250 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1251 rq->rt.rt_nr_migratory++;
1252 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1253 BUG_ON(!rq->rt.rt_nr_migratory);
1254 rq->rt.rt_nr_migratory--;
1257 update_rt_migration(rq);
1260 p->cpus_allowed = *new_mask;
1261 p->rt.nr_cpus_allowed = weight;
1264 /* Assumes rq->lock is held */
1265 static void rq_online_rt(struct rq *rq)
1267 if (rq->rt.overloaded)
1268 rt_set_overload(rq);
1270 __enable_runtime(rq);
1272 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1275 /* Assumes rq->lock is held */
1276 static void rq_offline_rt(struct rq *rq)
1278 if (rq->rt.overloaded)
1279 rt_clear_overload(rq);
1281 __disable_runtime(rq);
1283 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1287 * When switch from the rt queue, we bring ourselves to a position
1288 * that we might want to pull RT tasks from other runqueues.
1290 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1294 * If there are other RT tasks then we will reschedule
1295 * and the scheduling of the other RT tasks will handle
1296 * the balancing. But if we are the last RT task
1297 * we may need to handle the pulling of RT tasks
1300 if (!rq->rt.rt_nr_running)
1303 #endif /* CONFIG_SMP */
1306 * When switching a task to RT, we may overload the runqueue
1307 * with RT tasks. In this case we try to push them off to
1310 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1313 int check_resched = 1;
1316 * If we are already running, then there's nothing
1317 * that needs to be done. But if we are not running
1318 * we may need to preempt the current running task.
1319 * If that current running task is also an RT task
1320 * then see if we can move to another run queue.
1324 if (rq->rt.overloaded && push_rt_task(rq) &&
1325 /* Don't resched if we changed runqueues */
1328 #endif /* CONFIG_SMP */
1329 if (check_resched && p->prio < rq->curr->prio)
1330 resched_task(rq->curr);
1335 * Priority of the task has changed. This may cause
1336 * us to initiate a push or pull.
1338 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1339 int oldprio, int running)
1344 * If our priority decreases while running, we
1345 * may need to pull tasks to this runqueue.
1347 if (oldprio < p->prio)
1350 * If there's a higher priority task waiting to run
1351 * then reschedule. Note, the above pull_rt_task
1352 * can release the rq lock and p could migrate.
1353 * Only reschedule if p is still on the same runqueue.
1355 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1358 /* For UP simply resched on drop of prio */
1359 if (oldprio < p->prio)
1361 #endif /* CONFIG_SMP */
1364 * This task is not running, but if it is
1365 * greater than the current running task
1368 if (p->prio < rq->curr->prio)
1369 resched_task(rq->curr);
1373 static void watchdog(struct rq *rq, struct task_struct *p)
1375 unsigned long soft, hard;
1380 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1381 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1383 if (soft != RLIM_INFINITY) {
1387 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1388 if (p->rt.timeout > next)
1389 p->it_sched_expires = p->se.sum_exec_runtime;
1393 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1400 * RR tasks need a special form of timeslice management.
1401 * FIFO tasks have no timeslices.
1403 if (p->policy != SCHED_RR)
1406 if (--p->rt.time_slice)
1409 p->rt.time_slice = DEF_TIMESLICE;
1412 * Requeue to the end of queue if we are not the only element
1415 if (p->rt.run_list.prev != p->rt.run_list.next) {
1416 requeue_task_rt(rq, p);
1417 set_tsk_need_resched(p);
1421 static void set_curr_task_rt(struct rq *rq)
1423 struct task_struct *p = rq->curr;
1425 p->se.exec_start = rq->clock;
1428 static const struct sched_class rt_sched_class = {
1429 .next = &fair_sched_class,
1430 .enqueue_task = enqueue_task_rt,
1431 .dequeue_task = dequeue_task_rt,
1432 .yield_task = yield_task_rt,
1434 .select_task_rq = select_task_rq_rt,
1435 #endif /* CONFIG_SMP */
1437 .check_preempt_curr = check_preempt_curr_rt,
1439 .pick_next_task = pick_next_task_rt,
1440 .put_prev_task = put_prev_task_rt,
1443 .load_balance = load_balance_rt,
1444 .move_one_task = move_one_task_rt,
1445 .set_cpus_allowed = set_cpus_allowed_rt,
1446 .rq_online = rq_online_rt,
1447 .rq_offline = rq_offline_rt,
1448 .pre_schedule = pre_schedule_rt,
1449 .post_schedule = post_schedule_rt,
1450 .task_wake_up = task_wake_up_rt,
1451 .switched_from = switched_from_rt,
1454 .set_curr_task = set_curr_task_rt,
1455 .task_tick = task_tick_rt,
1457 .prio_changed = prio_changed_rt,
1458 .switched_to = switched_to_rt,
1461 #ifdef CONFIG_SCHED_DEBUG
1462 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1464 static void print_rt_stats(struct seq_file *m, int cpu)
1466 struct rt_rq *rt_rq;
1469 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1470 print_rt_rq(m, cpu, rt_rq);