2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
8 #include <linux/slab.h>
9 #include <linux/irq_work.h>
11 int sched_rr_timeslice = RR_TIMESLICE;
13 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
15 struct rt_bandwidth def_rt_bandwidth;
17 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
19 struct rt_bandwidth *rt_b =
20 container_of(timer, struct rt_bandwidth, rt_period_timer);
24 raw_spin_lock(&rt_b->rt_runtime_lock);
26 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
30 raw_spin_unlock(&rt_b->rt_runtime_lock);
31 idle = do_sched_rt_period_timer(rt_b, overrun);
32 raw_spin_lock(&rt_b->rt_runtime_lock);
35 rt_b->rt_period_active = 0;
36 raw_spin_unlock(&rt_b->rt_runtime_lock);
38 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
41 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
43 rt_b->rt_period = ns_to_ktime(period);
44 rt_b->rt_runtime = runtime;
46 raw_spin_lock_init(&rt_b->rt_runtime_lock);
48 hrtimer_init(&rt_b->rt_period_timer,
49 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
50 rt_b->rt_period_timer.function = sched_rt_period_timer;
53 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
55 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
58 raw_spin_lock(&rt_b->rt_runtime_lock);
59 if (!rt_b->rt_period_active) {
60 rt_b->rt_period_active = 1;
61 hrtimer_forward_now(&rt_b->rt_period_timer, rt_b->rt_period);
62 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
64 raw_spin_unlock(&rt_b->rt_runtime_lock);
68 static void push_irq_work_func(struct irq_work *work);
71 void init_rt_rq(struct rt_rq *rt_rq)
73 struct rt_prio_array *array;
76 array = &rt_rq->active;
77 for (i = 0; i < MAX_RT_PRIO; i++) {
78 INIT_LIST_HEAD(array->queue + i);
79 __clear_bit(i, array->bitmap);
81 /* delimiter for bitsearch: */
82 __set_bit(MAX_RT_PRIO, array->bitmap);
84 #if defined CONFIG_SMP
85 rt_rq->highest_prio.curr = MAX_RT_PRIO;
86 rt_rq->highest_prio.next = MAX_RT_PRIO;
87 rt_rq->rt_nr_migratory = 0;
88 rt_rq->overloaded = 0;
89 plist_head_init(&rt_rq->pushable_tasks);
91 #ifdef HAVE_RT_PUSH_IPI
92 rt_rq->push_flags = 0;
93 rt_rq->push_cpu = nr_cpu_ids;
94 raw_spin_lock_init(&rt_rq->push_lock);
95 init_irq_work(&rt_rq->push_work, push_irq_work_func);
97 #endif /* CONFIG_SMP */
98 /* We start is dequeued state, because no RT tasks are queued */
102 rt_rq->rt_throttled = 0;
103 rt_rq->rt_runtime = 0;
104 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
107 #ifdef CONFIG_RT_GROUP_SCHED
108 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
110 hrtimer_cancel(&rt_b->rt_period_timer);
113 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
115 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
117 #ifdef CONFIG_SCHED_DEBUG
118 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
120 return container_of(rt_se, struct task_struct, rt);
123 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
128 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
133 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
135 struct rt_rq *rt_rq = rt_se->rt_rq;
140 void free_rt_sched_group(struct task_group *tg)
145 destroy_rt_bandwidth(&tg->rt_bandwidth);
147 for_each_possible_cpu(i) {
158 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
159 struct sched_rt_entity *rt_se, int cpu,
160 struct sched_rt_entity *parent)
162 struct rq *rq = cpu_rq(cpu);
164 rt_rq->highest_prio.curr = MAX_RT_PRIO;
165 rt_rq->rt_nr_boosted = 0;
169 tg->rt_rq[cpu] = rt_rq;
170 tg->rt_se[cpu] = rt_se;
176 rt_se->rt_rq = &rq->rt;
178 rt_se->rt_rq = parent->my_q;
181 rt_se->parent = parent;
182 INIT_LIST_HEAD(&rt_se->run_list);
185 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
188 struct sched_rt_entity *rt_se;
191 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
194 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
198 init_rt_bandwidth(&tg->rt_bandwidth,
199 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
201 for_each_possible_cpu(i) {
202 rt_rq = kzalloc_node(sizeof(struct rt_rq),
203 GFP_KERNEL, cpu_to_node(i));
207 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
208 GFP_KERNEL, cpu_to_node(i));
213 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
214 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
225 #else /* CONFIG_RT_GROUP_SCHED */
227 #define rt_entity_is_task(rt_se) (1)
229 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
231 return container_of(rt_se, struct task_struct, rt);
234 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
236 return container_of(rt_rq, struct rq, rt);
239 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
241 struct task_struct *p = rt_task_of(rt_se);
246 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
248 struct rq *rq = rq_of_rt_se(rt_se);
253 void free_rt_sched_group(struct task_group *tg) { }
255 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
259 #endif /* CONFIG_RT_GROUP_SCHED */
263 static int pull_rt_task(struct rq *this_rq);
265 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
267 /* Try to pull RT tasks here if we lower this rq's prio */
268 return rq->rt.highest_prio.curr > prev->prio;
271 static inline int rt_overloaded(struct rq *rq)
273 return atomic_read(&rq->rd->rto_count);
276 static inline void rt_set_overload(struct rq *rq)
281 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
283 * Make sure the mask is visible before we set
284 * the overload count. That is checked to determine
285 * if we should look at the mask. It would be a shame
286 * if we looked at the mask, but the mask was not
289 * Matched by the barrier in pull_rt_task().
292 atomic_inc(&rq->rd->rto_count);
295 static inline void rt_clear_overload(struct rq *rq)
300 /* the order here really doesn't matter */
301 atomic_dec(&rq->rd->rto_count);
302 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
305 static void update_rt_migration(struct rt_rq *rt_rq)
307 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
308 if (!rt_rq->overloaded) {
309 rt_set_overload(rq_of_rt_rq(rt_rq));
310 rt_rq->overloaded = 1;
312 } else if (rt_rq->overloaded) {
313 rt_clear_overload(rq_of_rt_rq(rt_rq));
314 rt_rq->overloaded = 0;
318 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
320 struct task_struct *p;
322 if (!rt_entity_is_task(rt_se))
325 p = rt_task_of(rt_se);
326 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
328 rt_rq->rt_nr_total++;
329 if (p->nr_cpus_allowed > 1)
330 rt_rq->rt_nr_migratory++;
332 update_rt_migration(rt_rq);
335 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
337 struct task_struct *p;
339 if (!rt_entity_is_task(rt_se))
342 p = rt_task_of(rt_se);
343 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
345 rt_rq->rt_nr_total--;
346 if (p->nr_cpus_allowed > 1)
347 rt_rq->rt_nr_migratory--;
349 update_rt_migration(rt_rq);
352 static inline int has_pushable_tasks(struct rq *rq)
354 return !plist_head_empty(&rq->rt.pushable_tasks);
357 static inline void set_post_schedule(struct rq *rq)
360 * We detect this state here so that we can avoid taking the RQ
361 * lock again later if there is no need to push
363 rq->post_schedule = has_pushable_tasks(rq);
366 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
368 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
369 plist_node_init(&p->pushable_tasks, p->prio);
370 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
372 /* Update the highest prio pushable task */
373 if (p->prio < rq->rt.highest_prio.next)
374 rq->rt.highest_prio.next = p->prio;
377 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
379 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
381 /* Update the new highest prio pushable task */
382 if (has_pushable_tasks(rq)) {
383 p = plist_first_entry(&rq->rt.pushable_tasks,
384 struct task_struct, pushable_tasks);
385 rq->rt.highest_prio.next = p->prio;
387 rq->rt.highest_prio.next = MAX_RT_PRIO;
392 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
396 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
401 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
406 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
410 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
415 static inline int pull_rt_task(struct rq *this_rq)
420 static inline void set_post_schedule(struct rq *rq)
423 #endif /* CONFIG_SMP */
425 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
426 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
428 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
430 return !list_empty(&rt_se->run_list);
433 #ifdef CONFIG_RT_GROUP_SCHED
435 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
440 return rt_rq->rt_runtime;
443 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
445 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
448 typedef struct task_group *rt_rq_iter_t;
450 static inline struct task_group *next_task_group(struct task_group *tg)
453 tg = list_entry_rcu(tg->list.next,
454 typeof(struct task_group), list);
455 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
457 if (&tg->list == &task_groups)
463 #define for_each_rt_rq(rt_rq, iter, rq) \
464 for (iter = container_of(&task_groups, typeof(*iter), list); \
465 (iter = next_task_group(iter)) && \
466 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
468 #define for_each_sched_rt_entity(rt_se) \
469 for (; rt_se; rt_se = rt_se->parent)
471 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
476 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
477 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
479 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
481 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
482 struct rq *rq = rq_of_rt_rq(rt_rq);
483 struct sched_rt_entity *rt_se;
485 int cpu = cpu_of(rq);
487 rt_se = rt_rq->tg->rt_se[cpu];
489 if (rt_rq->rt_nr_running) {
491 enqueue_top_rt_rq(rt_rq);
492 else if (!on_rt_rq(rt_se))
493 enqueue_rt_entity(rt_se, false);
495 if (rt_rq->highest_prio.curr < curr->prio)
500 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
502 struct sched_rt_entity *rt_se;
503 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
505 rt_se = rt_rq->tg->rt_se[cpu];
508 dequeue_top_rt_rq(rt_rq);
509 else if (on_rt_rq(rt_se))
510 dequeue_rt_entity(rt_se);
513 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
515 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
518 static int rt_se_boosted(struct sched_rt_entity *rt_se)
520 struct rt_rq *rt_rq = group_rt_rq(rt_se);
521 struct task_struct *p;
524 return !!rt_rq->rt_nr_boosted;
526 p = rt_task_of(rt_se);
527 return p->prio != p->normal_prio;
531 static inline const struct cpumask *sched_rt_period_mask(void)
533 return this_rq()->rd->span;
536 static inline const struct cpumask *sched_rt_period_mask(void)
538 return cpu_online_mask;
543 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
545 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
548 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
550 return &rt_rq->tg->rt_bandwidth;
553 #else /* !CONFIG_RT_GROUP_SCHED */
555 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
557 return rt_rq->rt_runtime;
560 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
562 return ktime_to_ns(def_rt_bandwidth.rt_period);
565 typedef struct rt_rq *rt_rq_iter_t;
567 #define for_each_rt_rq(rt_rq, iter, rq) \
568 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
570 #define for_each_sched_rt_entity(rt_se) \
571 for (; rt_se; rt_se = NULL)
573 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
578 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
580 struct rq *rq = rq_of_rt_rq(rt_rq);
582 if (!rt_rq->rt_nr_running)
585 enqueue_top_rt_rq(rt_rq);
589 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
591 dequeue_top_rt_rq(rt_rq);
594 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
596 return rt_rq->rt_throttled;
599 static inline const struct cpumask *sched_rt_period_mask(void)
601 return cpu_online_mask;
605 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
607 return &cpu_rq(cpu)->rt;
610 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
612 return &def_rt_bandwidth;
615 #endif /* CONFIG_RT_GROUP_SCHED */
617 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
619 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
621 return (hrtimer_active(&rt_b->rt_period_timer) ||
622 rt_rq->rt_time < rt_b->rt_runtime);
627 * We ran out of runtime, see if we can borrow some from our neighbours.
629 static int do_balance_runtime(struct rt_rq *rt_rq)
631 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
632 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
633 int i, weight, more = 0;
636 weight = cpumask_weight(rd->span);
638 raw_spin_lock(&rt_b->rt_runtime_lock);
639 rt_period = ktime_to_ns(rt_b->rt_period);
640 for_each_cpu(i, rd->span) {
641 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
647 raw_spin_lock(&iter->rt_runtime_lock);
649 * Either all rqs have inf runtime and there's nothing to steal
650 * or __disable_runtime() below sets a specific rq to inf to
651 * indicate its been disabled and disalow stealing.
653 if (iter->rt_runtime == RUNTIME_INF)
657 * From runqueues with spare time, take 1/n part of their
658 * spare time, but no more than our period.
660 diff = iter->rt_runtime - iter->rt_time;
662 diff = div_u64((u64)diff, weight);
663 if (rt_rq->rt_runtime + diff > rt_period)
664 diff = rt_period - rt_rq->rt_runtime;
665 iter->rt_runtime -= diff;
666 rt_rq->rt_runtime += diff;
668 if (rt_rq->rt_runtime == rt_period) {
669 raw_spin_unlock(&iter->rt_runtime_lock);
674 raw_spin_unlock(&iter->rt_runtime_lock);
676 raw_spin_unlock(&rt_b->rt_runtime_lock);
682 * Ensure this RQ takes back all the runtime it lend to its neighbours.
684 static void __disable_runtime(struct rq *rq)
686 struct root_domain *rd = rq->rd;
690 if (unlikely(!scheduler_running))
693 for_each_rt_rq(rt_rq, iter, rq) {
694 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
698 raw_spin_lock(&rt_b->rt_runtime_lock);
699 raw_spin_lock(&rt_rq->rt_runtime_lock);
701 * Either we're all inf and nobody needs to borrow, or we're
702 * already disabled and thus have nothing to do, or we have
703 * exactly the right amount of runtime to take out.
705 if (rt_rq->rt_runtime == RUNTIME_INF ||
706 rt_rq->rt_runtime == rt_b->rt_runtime)
708 raw_spin_unlock(&rt_rq->rt_runtime_lock);
711 * Calculate the difference between what we started out with
712 * and what we current have, that's the amount of runtime
713 * we lend and now have to reclaim.
715 want = rt_b->rt_runtime - rt_rq->rt_runtime;
718 * Greedy reclaim, take back as much as we can.
720 for_each_cpu(i, rd->span) {
721 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
725 * Can't reclaim from ourselves or disabled runqueues.
727 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
730 raw_spin_lock(&iter->rt_runtime_lock);
732 diff = min_t(s64, iter->rt_runtime, want);
733 iter->rt_runtime -= diff;
736 iter->rt_runtime -= want;
739 raw_spin_unlock(&iter->rt_runtime_lock);
745 raw_spin_lock(&rt_rq->rt_runtime_lock);
747 * We cannot be left wanting - that would mean some runtime
748 * leaked out of the system.
753 * Disable all the borrow logic by pretending we have inf
754 * runtime - in which case borrowing doesn't make sense.
756 rt_rq->rt_runtime = RUNTIME_INF;
757 rt_rq->rt_throttled = 0;
758 raw_spin_unlock(&rt_rq->rt_runtime_lock);
759 raw_spin_unlock(&rt_b->rt_runtime_lock);
761 /* Make rt_rq available for pick_next_task() */
762 sched_rt_rq_enqueue(rt_rq);
766 static void __enable_runtime(struct rq *rq)
771 if (unlikely(!scheduler_running))
775 * Reset each runqueue's bandwidth settings
777 for_each_rt_rq(rt_rq, iter, rq) {
778 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
780 raw_spin_lock(&rt_b->rt_runtime_lock);
781 raw_spin_lock(&rt_rq->rt_runtime_lock);
782 rt_rq->rt_runtime = rt_b->rt_runtime;
784 rt_rq->rt_throttled = 0;
785 raw_spin_unlock(&rt_rq->rt_runtime_lock);
786 raw_spin_unlock(&rt_b->rt_runtime_lock);
790 static int balance_runtime(struct rt_rq *rt_rq)
794 if (!sched_feat(RT_RUNTIME_SHARE))
797 if (rt_rq->rt_time > rt_rq->rt_runtime) {
798 raw_spin_unlock(&rt_rq->rt_runtime_lock);
799 more = do_balance_runtime(rt_rq);
800 raw_spin_lock(&rt_rq->rt_runtime_lock);
805 #else /* !CONFIG_SMP */
806 static inline int balance_runtime(struct rt_rq *rt_rq)
810 #endif /* CONFIG_SMP */
812 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
814 int i, idle = 1, throttled = 0;
815 const struct cpumask *span;
817 span = sched_rt_period_mask();
818 #ifdef CONFIG_RT_GROUP_SCHED
820 * FIXME: isolated CPUs should really leave the root task group,
821 * whether they are isolcpus or were isolated via cpusets, lest
822 * the timer run on a CPU which does not service all runqueues,
823 * potentially leaving other CPUs indefinitely throttled. If
824 * isolation is really required, the user will turn the throttle
825 * off to kill the perturbations it causes anyway. Meanwhile,
826 * this maintains functionality for boot and/or troubleshooting.
828 if (rt_b == &root_task_group.rt_bandwidth)
829 span = cpu_online_mask;
831 for_each_cpu(i, span) {
833 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
834 struct rq *rq = rq_of_rt_rq(rt_rq);
836 raw_spin_lock(&rq->lock);
837 if (rt_rq->rt_time) {
840 raw_spin_lock(&rt_rq->rt_runtime_lock);
841 if (rt_rq->rt_throttled)
842 balance_runtime(rt_rq);
843 runtime = rt_rq->rt_runtime;
844 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
845 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
846 rt_rq->rt_throttled = 0;
850 * When we're idle and a woken (rt) task is
851 * throttled check_preempt_curr() will set
852 * skip_update and the time between the wakeup
853 * and this unthrottle will get accounted as
856 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
857 rq_clock_skip_update(rq, false);
859 if (rt_rq->rt_time || rt_rq->rt_nr_running)
861 raw_spin_unlock(&rt_rq->rt_runtime_lock);
862 } else if (rt_rq->rt_nr_running) {
864 if (!rt_rq_throttled(rt_rq))
867 if (rt_rq->rt_throttled)
871 sched_rt_rq_enqueue(rt_rq);
872 raw_spin_unlock(&rq->lock);
875 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
881 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
883 #ifdef CONFIG_RT_GROUP_SCHED
884 struct rt_rq *rt_rq = group_rt_rq(rt_se);
887 return rt_rq->highest_prio.curr;
890 return rt_task_of(rt_se)->prio;
893 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
895 u64 runtime = sched_rt_runtime(rt_rq);
897 if (rt_rq->rt_throttled)
898 return rt_rq_throttled(rt_rq);
900 if (runtime >= sched_rt_period(rt_rq))
903 balance_runtime(rt_rq);
904 runtime = sched_rt_runtime(rt_rq);
905 if (runtime == RUNTIME_INF)
908 if (rt_rq->rt_time > runtime) {
909 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
912 * Don't actually throttle groups that have no runtime assigned
913 * but accrue some time due to boosting.
915 if (likely(rt_b->rt_runtime)) {
916 rt_rq->rt_throttled = 1;
917 printk_deferred_once("sched: RT throttling activated\n");
920 * In case we did anyway, make it go away,
921 * replenishment is a joke, since it will replenish us
927 if (rt_rq_throttled(rt_rq)) {
928 sched_rt_rq_dequeue(rt_rq);
937 * Update the current task's runtime statistics. Skip current tasks that
938 * are not in our scheduling class.
940 static void update_curr_rt(struct rq *rq)
942 struct task_struct *curr = rq->curr;
943 struct sched_rt_entity *rt_se = &curr->rt;
946 if (curr->sched_class != &rt_sched_class)
949 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
950 if (unlikely((s64)delta_exec <= 0))
953 schedstat_set(curr->se.statistics.exec_max,
954 max(curr->se.statistics.exec_max, delta_exec));
956 curr->se.sum_exec_runtime += delta_exec;
957 account_group_exec_runtime(curr, delta_exec);
959 curr->se.exec_start = rq_clock_task(rq);
960 cpuacct_charge(curr, delta_exec);
962 sched_rt_avg_update(rq, delta_exec);
964 if (!rt_bandwidth_enabled())
967 for_each_sched_rt_entity(rt_se) {
968 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
970 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
971 raw_spin_lock(&rt_rq->rt_runtime_lock);
972 rt_rq->rt_time += delta_exec;
973 if (sched_rt_runtime_exceeded(rt_rq))
975 raw_spin_unlock(&rt_rq->rt_runtime_lock);
981 dequeue_top_rt_rq(struct rt_rq *rt_rq)
983 struct rq *rq = rq_of_rt_rq(rt_rq);
985 BUG_ON(&rq->rt != rt_rq);
987 if (!rt_rq->rt_queued)
990 BUG_ON(!rq->nr_running);
992 sub_nr_running(rq, rt_rq->rt_nr_running);
993 rt_rq->rt_queued = 0;
997 enqueue_top_rt_rq(struct rt_rq *rt_rq)
999 struct rq *rq = rq_of_rt_rq(rt_rq);
1001 BUG_ON(&rq->rt != rt_rq);
1003 if (rt_rq->rt_queued)
1005 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1008 add_nr_running(rq, rt_rq->rt_nr_running);
1009 rt_rq->rt_queued = 1;
1012 #if defined CONFIG_SMP
1015 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1017 struct rq *rq = rq_of_rt_rq(rt_rq);
1019 #ifdef CONFIG_RT_GROUP_SCHED
1021 * Change rq's cpupri only if rt_rq is the top queue.
1023 if (&rq->rt != rt_rq)
1026 if (rq->online && prio < prev_prio)
1027 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1031 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1033 struct rq *rq = rq_of_rt_rq(rt_rq);
1035 #ifdef CONFIG_RT_GROUP_SCHED
1037 * Change rq's cpupri only if rt_rq is the top queue.
1039 if (&rq->rt != rt_rq)
1042 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1043 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1046 #else /* CONFIG_SMP */
1049 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1051 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1053 #endif /* CONFIG_SMP */
1055 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1057 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1059 int prev_prio = rt_rq->highest_prio.curr;
1061 if (prio < prev_prio)
1062 rt_rq->highest_prio.curr = prio;
1064 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1068 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1070 int prev_prio = rt_rq->highest_prio.curr;
1072 if (rt_rq->rt_nr_running) {
1074 WARN_ON(prio < prev_prio);
1077 * This may have been our highest task, and therefore
1078 * we may have some recomputation to do
1080 if (prio == prev_prio) {
1081 struct rt_prio_array *array = &rt_rq->active;
1083 rt_rq->highest_prio.curr =
1084 sched_find_first_bit(array->bitmap);
1088 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1090 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1095 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1096 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1098 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1100 #ifdef CONFIG_RT_GROUP_SCHED
1103 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1105 if (rt_se_boosted(rt_se))
1106 rt_rq->rt_nr_boosted++;
1109 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1113 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1115 if (rt_se_boosted(rt_se))
1116 rt_rq->rt_nr_boosted--;
1118 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1121 #else /* CONFIG_RT_GROUP_SCHED */
1124 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1126 start_rt_bandwidth(&def_rt_bandwidth);
1130 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1132 #endif /* CONFIG_RT_GROUP_SCHED */
1135 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1137 struct rt_rq *group_rq = group_rt_rq(rt_se);
1140 return group_rq->rt_nr_running;
1146 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1148 int prio = rt_se_prio(rt_se);
1150 WARN_ON(!rt_prio(prio));
1151 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1153 inc_rt_prio(rt_rq, prio);
1154 inc_rt_migration(rt_se, rt_rq);
1155 inc_rt_group(rt_se, rt_rq);
1159 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1161 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1162 WARN_ON(!rt_rq->rt_nr_running);
1163 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1165 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1166 dec_rt_migration(rt_se, rt_rq);
1167 dec_rt_group(rt_se, rt_rq);
1170 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1172 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1173 struct rt_prio_array *array = &rt_rq->active;
1174 struct rt_rq *group_rq = group_rt_rq(rt_se);
1175 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1178 * Don't enqueue the group if its throttled, or when empty.
1179 * The latter is a consequence of the former when a child group
1180 * get throttled and the current group doesn't have any other
1183 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
1187 list_add(&rt_se->run_list, queue);
1189 list_add_tail(&rt_se->run_list, queue);
1190 __set_bit(rt_se_prio(rt_se), array->bitmap);
1192 inc_rt_tasks(rt_se, rt_rq);
1195 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
1197 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1198 struct rt_prio_array *array = &rt_rq->active;
1200 list_del_init(&rt_se->run_list);
1201 if (list_empty(array->queue + rt_se_prio(rt_se)))
1202 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1204 dec_rt_tasks(rt_se, rt_rq);
1208 * Because the prio of an upper entry depends on the lower
1209 * entries, we must remove entries top - down.
1211 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
1213 struct sched_rt_entity *back = NULL;
1215 for_each_sched_rt_entity(rt_se) {
1220 dequeue_top_rt_rq(rt_rq_of_se(back));
1222 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1223 if (on_rt_rq(rt_se))
1224 __dequeue_rt_entity(rt_se);
1228 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1230 struct rq *rq = rq_of_rt_se(rt_se);
1232 dequeue_rt_stack(rt_se);
1233 for_each_sched_rt_entity(rt_se)
1234 __enqueue_rt_entity(rt_se, head);
1235 enqueue_top_rt_rq(&rq->rt);
1238 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1240 struct rq *rq = rq_of_rt_se(rt_se);
1242 dequeue_rt_stack(rt_se);
1244 for_each_sched_rt_entity(rt_se) {
1245 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1247 if (rt_rq && rt_rq->rt_nr_running)
1248 __enqueue_rt_entity(rt_se, false);
1250 enqueue_top_rt_rq(&rq->rt);
1254 * Adding/removing a task to/from a priority array:
1257 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1259 struct sched_rt_entity *rt_se = &p->rt;
1261 if (flags & ENQUEUE_WAKEUP)
1264 enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
1266 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1267 enqueue_pushable_task(rq, p);
1270 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1272 struct sched_rt_entity *rt_se = &p->rt;
1275 dequeue_rt_entity(rt_se);
1277 dequeue_pushable_task(rq, p);
1281 * Put task to the head or the end of the run list without the overhead of
1282 * dequeue followed by enqueue.
1285 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1287 if (on_rt_rq(rt_se)) {
1288 struct rt_prio_array *array = &rt_rq->active;
1289 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1292 list_move(&rt_se->run_list, queue);
1294 list_move_tail(&rt_se->run_list, queue);
1298 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1300 struct sched_rt_entity *rt_se = &p->rt;
1301 struct rt_rq *rt_rq;
1303 for_each_sched_rt_entity(rt_se) {
1304 rt_rq = rt_rq_of_se(rt_se);
1305 requeue_rt_entity(rt_rq, rt_se, head);
1309 static void yield_task_rt(struct rq *rq)
1311 requeue_task_rt(rq, rq->curr, 0);
1315 static int find_lowest_rq(struct task_struct *task);
1318 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1320 struct task_struct *curr;
1323 /* For anything but wake ups, just return the task_cpu */
1324 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1330 curr = READ_ONCE(rq->curr); /* unlocked access */
1333 * If the current task on @p's runqueue is an RT task, then
1334 * try to see if we can wake this RT task up on another
1335 * runqueue. Otherwise simply start this RT task
1336 * on its current runqueue.
1338 * We want to avoid overloading runqueues. If the woken
1339 * task is a higher priority, then it will stay on this CPU
1340 * and the lower prio task should be moved to another CPU.
1341 * Even though this will probably make the lower prio task
1342 * lose its cache, we do not want to bounce a higher task
1343 * around just because it gave up its CPU, perhaps for a
1346 * For equal prio tasks, we just let the scheduler sort it out.
1348 * Otherwise, just let it ride on the affined RQ and the
1349 * post-schedule router will push the preempted task away
1351 * This test is optimistic, if we get it wrong the load-balancer
1352 * will have to sort it out.
1354 if (curr && unlikely(rt_task(curr)) &&
1355 (curr->nr_cpus_allowed < 2 ||
1356 curr->prio <= p->prio)) {
1357 int target = find_lowest_rq(p);
1360 * Don't bother moving it if the destination CPU is
1361 * not running a lower priority task.
1364 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1373 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1376 * Current can't be migrated, useless to reschedule,
1377 * let's hope p can move out.
1379 if (rq->curr->nr_cpus_allowed == 1 ||
1380 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1384 * p is migratable, so let's not schedule it and
1385 * see if it is pushed or pulled somewhere else.
1387 if (p->nr_cpus_allowed != 1
1388 && cpupri_find(&rq->rd->cpupri, p, NULL))
1392 * There appears to be other cpus that can accept
1393 * current and none to run 'p', so lets reschedule
1394 * to try and push current away:
1396 requeue_task_rt(rq, p, 1);
1400 #endif /* CONFIG_SMP */
1403 * Preempt the current task with a newly woken task if needed:
1405 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1407 if (p->prio < rq->curr->prio) {
1416 * - the newly woken task is of equal priority to the current task
1417 * - the newly woken task is non-migratable while current is migratable
1418 * - current will be preempted on the next reschedule
1420 * we should check to see if current can readily move to a different
1421 * cpu. If so, we will reschedule to allow the push logic to try
1422 * to move current somewhere else, making room for our non-migratable
1425 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1426 check_preempt_equal_prio(rq, p);
1430 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1431 struct rt_rq *rt_rq)
1433 struct rt_prio_array *array = &rt_rq->active;
1434 struct sched_rt_entity *next = NULL;
1435 struct list_head *queue;
1438 idx = sched_find_first_bit(array->bitmap);
1439 BUG_ON(idx >= MAX_RT_PRIO);
1441 queue = array->queue + idx;
1442 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1447 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1449 struct sched_rt_entity *rt_se;
1450 struct task_struct *p;
1451 struct rt_rq *rt_rq = &rq->rt;
1454 rt_se = pick_next_rt_entity(rq, rt_rq);
1456 rt_rq = group_rt_rq(rt_se);
1459 p = rt_task_of(rt_se);
1460 p->se.exec_start = rq_clock_task(rq);
1465 static struct task_struct *
1466 pick_next_task_rt(struct rq *rq, struct task_struct *prev)
1468 struct task_struct *p;
1469 struct rt_rq *rt_rq = &rq->rt;
1471 if (need_pull_rt_task(rq, prev)) {
1474 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1475 * means a dl or stop task can slip in, in which case we need
1476 * to re-start task selection.
1478 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1479 rq->dl.dl_nr_running))
1484 * We may dequeue prev's rt_rq in put_prev_task().
1485 * So, we update time before rt_nr_running check.
1487 if (prev->sched_class == &rt_sched_class)
1490 if (!rt_rq->rt_queued)
1493 put_prev_task(rq, prev);
1495 p = _pick_next_task_rt(rq);
1497 /* The running task is never eligible for pushing */
1498 dequeue_pushable_task(rq, p);
1500 set_post_schedule(rq);
1505 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1510 * The previous task needs to be made eligible for pushing
1511 * if it is still active
1513 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1514 enqueue_pushable_task(rq, p);
1519 /* Only try algorithms three times */
1520 #define RT_MAX_TRIES 3
1522 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1524 if (!task_running(rq, p) &&
1525 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1531 * Return the highest pushable rq's task, which is suitable to be executed
1532 * on the cpu, NULL otherwise
1534 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1536 struct plist_head *head = &rq->rt.pushable_tasks;
1537 struct task_struct *p;
1539 if (!has_pushable_tasks(rq))
1542 plist_for_each_entry(p, head, pushable_tasks) {
1543 if (pick_rt_task(rq, p, cpu))
1550 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1552 static int find_lowest_rq(struct task_struct *task)
1554 struct sched_domain *sd;
1555 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1556 int this_cpu = smp_processor_id();
1557 int cpu = task_cpu(task);
1559 /* Make sure the mask is initialized first */
1560 if (unlikely(!lowest_mask))
1563 if (task->nr_cpus_allowed == 1)
1564 return -1; /* No other targets possible */
1566 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1567 return -1; /* No targets found */
1570 * At this point we have built a mask of cpus representing the
1571 * lowest priority tasks in the system. Now we want to elect
1572 * the best one based on our affinity and topology.
1574 * We prioritize the last cpu that the task executed on since
1575 * it is most likely cache-hot in that location.
1577 if (cpumask_test_cpu(cpu, lowest_mask))
1581 * Otherwise, we consult the sched_domains span maps to figure
1582 * out which cpu is logically closest to our hot cache data.
1584 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1585 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1588 for_each_domain(cpu, sd) {
1589 if (sd->flags & SD_WAKE_AFFINE) {
1593 * "this_cpu" is cheaper to preempt than a
1596 if (this_cpu != -1 &&
1597 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1602 best_cpu = cpumask_first_and(lowest_mask,
1603 sched_domain_span(sd));
1604 if (best_cpu < nr_cpu_ids) {
1613 * And finally, if there were no matches within the domains
1614 * just give the caller *something* to work with from the compatible
1620 cpu = cpumask_any(lowest_mask);
1621 if (cpu < nr_cpu_ids)
1626 /* Will lock the rq it finds */
1627 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1629 struct rq *lowest_rq = NULL;
1633 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1634 cpu = find_lowest_rq(task);
1636 if ((cpu == -1) || (cpu == rq->cpu))
1639 lowest_rq = cpu_rq(cpu);
1641 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1643 * Target rq has tasks of equal or higher priority,
1644 * retrying does not release any lock and is unlikely
1645 * to yield a different result.
1651 /* if the prio of this runqueue changed, try again */
1652 if (double_lock_balance(rq, lowest_rq)) {
1654 * We had to unlock the run queue. In
1655 * the mean time, task could have
1656 * migrated already or had its affinity changed.
1657 * Also make sure that it wasn't scheduled on its rq.
1659 if (unlikely(task_rq(task) != rq ||
1660 !cpumask_test_cpu(lowest_rq->cpu,
1661 tsk_cpus_allowed(task)) ||
1662 task_running(rq, task) ||
1663 !task_on_rq_queued(task))) {
1665 double_unlock_balance(rq, lowest_rq);
1671 /* If this rq is still suitable use it. */
1672 if (lowest_rq->rt.highest_prio.curr > task->prio)
1676 double_unlock_balance(rq, lowest_rq);
1683 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1685 struct task_struct *p;
1687 if (!has_pushable_tasks(rq))
1690 p = plist_first_entry(&rq->rt.pushable_tasks,
1691 struct task_struct, pushable_tasks);
1693 BUG_ON(rq->cpu != task_cpu(p));
1694 BUG_ON(task_current(rq, p));
1695 BUG_ON(p->nr_cpus_allowed <= 1);
1697 BUG_ON(!task_on_rq_queued(p));
1698 BUG_ON(!rt_task(p));
1704 * If the current CPU has more than one RT task, see if the non
1705 * running task can migrate over to a CPU that is running a task
1706 * of lesser priority.
1708 static int push_rt_task(struct rq *rq)
1710 struct task_struct *next_task;
1711 struct rq *lowest_rq;
1714 if (!rq->rt.overloaded)
1717 next_task = pick_next_pushable_task(rq);
1722 if (unlikely(next_task == rq->curr)) {
1728 * It's possible that the next_task slipped in of
1729 * higher priority than current. If that's the case
1730 * just reschedule current.
1732 if (unlikely(next_task->prio < rq->curr->prio)) {
1737 /* We might release rq lock */
1738 get_task_struct(next_task);
1740 /* find_lock_lowest_rq locks the rq if found */
1741 lowest_rq = find_lock_lowest_rq(next_task, rq);
1743 struct task_struct *task;
1745 * find_lock_lowest_rq releases rq->lock
1746 * so it is possible that next_task has migrated.
1748 * We need to make sure that the task is still on the same
1749 * run-queue and is also still the next task eligible for
1752 task = pick_next_pushable_task(rq);
1753 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1755 * The task hasn't migrated, and is still the next
1756 * eligible task, but we failed to find a run-queue
1757 * to push it to. Do not retry in this case, since
1758 * other cpus will pull from us when ready.
1764 /* No more tasks, just exit */
1768 * Something has shifted, try again.
1770 put_task_struct(next_task);
1775 deactivate_task(rq, next_task, 0);
1776 set_task_cpu(next_task, lowest_rq->cpu);
1777 activate_task(lowest_rq, next_task, 0);
1780 resched_curr(lowest_rq);
1782 double_unlock_balance(rq, lowest_rq);
1785 put_task_struct(next_task);
1790 static void push_rt_tasks(struct rq *rq)
1792 /* push_rt_task will return true if it moved an RT */
1793 while (push_rt_task(rq))
1797 #ifdef HAVE_RT_PUSH_IPI
1799 * The search for the next cpu always starts at rq->cpu and ends
1800 * when we reach rq->cpu again. It will never return rq->cpu.
1801 * This returns the next cpu to check, or nr_cpu_ids if the loop
1804 * rq->rt.push_cpu holds the last cpu returned by this function,
1805 * or if this is the first instance, it must hold rq->cpu.
1807 static int rto_next_cpu(struct rq *rq)
1809 int prev_cpu = rq->rt.push_cpu;
1812 cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1815 * If the previous cpu is less than the rq's CPU, then it already
1816 * passed the end of the mask, and has started from the beginning.
1817 * We end if the next CPU is greater or equal to rq's CPU.
1819 if (prev_cpu < rq->cpu) {
1823 } else if (cpu >= nr_cpu_ids) {
1825 * We passed the end of the mask, start at the beginning.
1826 * If the result is greater or equal to the rq's CPU, then
1827 * the loop is finished.
1829 cpu = cpumask_first(rq->rd->rto_mask);
1833 rq->rt.push_cpu = cpu;
1835 /* Return cpu to let the caller know if the loop is finished or not */
1839 static int find_next_push_cpu(struct rq *rq)
1845 cpu = rto_next_cpu(rq);
1846 if (cpu >= nr_cpu_ids)
1848 next_rq = cpu_rq(cpu);
1850 /* Make sure the next rq can push to this rq */
1851 if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1858 #define RT_PUSH_IPI_EXECUTING 1
1859 #define RT_PUSH_IPI_RESTART 2
1861 static void tell_cpu_to_push(struct rq *rq)
1865 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1866 raw_spin_lock(&rq->rt.push_lock);
1867 /* Make sure it's still executing */
1868 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1870 * Tell the IPI to restart the loop as things have
1871 * changed since it started.
1873 rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
1874 raw_spin_unlock(&rq->rt.push_lock);
1877 raw_spin_unlock(&rq->rt.push_lock);
1880 /* When here, there's no IPI going around */
1882 rq->rt.push_cpu = rq->cpu;
1883 cpu = find_next_push_cpu(rq);
1884 if (cpu >= nr_cpu_ids)
1887 rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
1889 irq_work_queue_on(&rq->rt.push_work, cpu);
1892 /* Called from hardirq context */
1893 static void try_to_push_tasks(void *arg)
1895 struct rt_rq *rt_rq = arg;
1896 struct rq *rq, *src_rq;
1900 this_cpu = rt_rq->push_cpu;
1902 /* Paranoid check */
1903 BUG_ON(this_cpu != smp_processor_id());
1905 rq = cpu_rq(this_cpu);
1906 src_rq = rq_of_rt_rq(rt_rq);
1909 if (has_pushable_tasks(rq)) {
1910 raw_spin_lock(&rq->lock);
1912 raw_spin_unlock(&rq->lock);
1915 /* Pass the IPI to the next rt overloaded queue */
1916 raw_spin_lock(&rt_rq->push_lock);
1918 * If the source queue changed since the IPI went out,
1919 * we need to restart the search from that CPU again.
1921 if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
1922 rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
1923 rt_rq->push_cpu = src_rq->cpu;
1926 cpu = find_next_push_cpu(src_rq);
1928 if (cpu >= nr_cpu_ids)
1929 rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
1930 raw_spin_unlock(&rt_rq->push_lock);
1932 if (cpu >= nr_cpu_ids)
1936 * It is possible that a restart caused this CPU to be
1937 * chosen again. Don't bother with an IPI, just see if we
1938 * have more to push.
1940 if (unlikely(cpu == rq->cpu))
1943 /* Try the next RT overloaded CPU */
1944 irq_work_queue_on(&rt_rq->push_work, cpu);
1947 static void push_irq_work_func(struct irq_work *work)
1949 struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
1951 try_to_push_tasks(rt_rq);
1953 #endif /* HAVE_RT_PUSH_IPI */
1955 static int pull_rt_task(struct rq *this_rq)
1957 int this_cpu = this_rq->cpu, ret = 0, cpu;
1958 struct task_struct *p;
1961 if (likely(!rt_overloaded(this_rq)))
1965 * Match the barrier from rt_set_overloaded; this guarantees that if we
1966 * see overloaded we must also see the rto_mask bit.
1970 #ifdef HAVE_RT_PUSH_IPI
1971 if (sched_feat(RT_PUSH_IPI)) {
1972 tell_cpu_to_push(this_rq);
1977 for_each_cpu(cpu, this_rq->rd->rto_mask) {
1978 if (this_cpu == cpu)
1981 src_rq = cpu_rq(cpu);
1984 * Don't bother taking the src_rq->lock if the next highest
1985 * task is known to be lower-priority than our current task.
1986 * This may look racy, but if this value is about to go
1987 * logically higher, the src_rq will push this task away.
1988 * And if its going logically lower, we do not care
1990 if (src_rq->rt.highest_prio.next >=
1991 this_rq->rt.highest_prio.curr)
1995 * We can potentially drop this_rq's lock in
1996 * double_lock_balance, and another CPU could
1999 double_lock_balance(this_rq, src_rq);
2002 * We can pull only a task, which is pushable
2003 * on its rq, and no others.
2005 p = pick_highest_pushable_task(src_rq, this_cpu);
2008 * Do we have an RT task that preempts
2009 * the to-be-scheduled task?
2011 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2012 WARN_ON(p == src_rq->curr);
2013 WARN_ON(!task_on_rq_queued(p));
2016 * There's a chance that p is higher in priority
2017 * than what's currently running on its cpu.
2018 * This is just that p is wakeing up and hasn't
2019 * had a chance to schedule. We only pull
2020 * p if it is lower in priority than the
2021 * current task on the run queue
2023 if (p->prio < src_rq->curr->prio)
2028 deactivate_task(src_rq, p, 0);
2029 set_task_cpu(p, this_cpu);
2030 activate_task(this_rq, p, 0);
2032 * We continue with the search, just in
2033 * case there's an even higher prio task
2034 * in another runqueue. (low likelihood
2039 double_unlock_balance(this_rq, src_rq);
2045 static void post_schedule_rt(struct rq *rq)
2051 * If we are not running and we are not going to reschedule soon, we should
2052 * try to push tasks away now
2054 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2056 if (!task_running(rq, p) &&
2057 !test_tsk_need_resched(rq->curr) &&
2058 has_pushable_tasks(rq) &&
2059 p->nr_cpus_allowed > 1 &&
2060 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2061 (rq->curr->nr_cpus_allowed < 2 ||
2062 rq->curr->prio <= p->prio))
2066 static void set_cpus_allowed_rt(struct task_struct *p,
2067 const struct cpumask *new_mask)
2072 BUG_ON(!rt_task(p));
2074 if (!task_on_rq_queued(p))
2077 weight = cpumask_weight(new_mask);
2080 * Only update if the process changes its state from whether it
2081 * can migrate or not.
2083 if ((p->nr_cpus_allowed > 1) == (weight > 1))
2089 * The process used to be able to migrate OR it can now migrate
2092 if (!task_current(rq, p))
2093 dequeue_pushable_task(rq, p);
2094 BUG_ON(!rq->rt.rt_nr_migratory);
2095 rq->rt.rt_nr_migratory--;
2097 if (!task_current(rq, p))
2098 enqueue_pushable_task(rq, p);
2099 rq->rt.rt_nr_migratory++;
2102 update_rt_migration(&rq->rt);
2105 /* Assumes rq->lock is held */
2106 static void rq_online_rt(struct rq *rq)
2108 if (rq->rt.overloaded)
2109 rt_set_overload(rq);
2111 __enable_runtime(rq);
2113 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2116 /* Assumes rq->lock is held */
2117 static void rq_offline_rt(struct rq *rq)
2119 if (rq->rt.overloaded)
2120 rt_clear_overload(rq);
2122 __disable_runtime(rq);
2124 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2128 * When switch from the rt queue, we bring ourselves to a position
2129 * that we might want to pull RT tasks from other runqueues.
2131 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2134 * If there are other RT tasks then we will reschedule
2135 * and the scheduling of the other RT tasks will handle
2136 * the balancing. But if we are the last RT task
2137 * we may need to handle the pulling of RT tasks
2140 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2143 if (pull_rt_task(rq))
2147 void __init init_sched_rt_class(void)
2151 for_each_possible_cpu(i) {
2152 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2153 GFP_KERNEL, cpu_to_node(i));
2156 #endif /* CONFIG_SMP */
2159 * When switching a task to RT, we may overload the runqueue
2160 * with RT tasks. In this case we try to push them off to
2163 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2165 int check_resched = 1;
2168 * If we are already running, then there's nothing
2169 * that needs to be done. But if we are not running
2170 * we may need to preempt the current running task.
2171 * If that current running task is also an RT task
2172 * then see if we can move to another run queue.
2174 if (task_on_rq_queued(p) && rq->curr != p) {
2176 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded &&
2177 /* Don't resched if we changed runqueues */
2178 push_rt_task(rq) && rq != task_rq(p))
2180 #endif /* CONFIG_SMP */
2181 if (check_resched && p->prio < rq->curr->prio)
2187 * Priority of the task has changed. This may cause
2188 * us to initiate a push or pull.
2191 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2193 if (!task_on_rq_queued(p))
2196 if (rq->curr == p) {
2199 * If our priority decreases while running, we
2200 * may need to pull tasks to this runqueue.
2202 if (oldprio < p->prio)
2205 * If there's a higher priority task waiting to run
2206 * then reschedule. Note, the above pull_rt_task
2207 * can release the rq lock and p could migrate.
2208 * Only reschedule if p is still on the same runqueue.
2210 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
2213 /* For UP simply resched on drop of prio */
2214 if (oldprio < p->prio)
2216 #endif /* CONFIG_SMP */
2219 * This task is not running, but if it is
2220 * greater than the current running task
2223 if (p->prio < rq->curr->prio)
2228 static void watchdog(struct rq *rq, struct task_struct *p)
2230 unsigned long soft, hard;
2232 /* max may change after cur was read, this will be fixed next tick */
2233 soft = task_rlimit(p, RLIMIT_RTTIME);
2234 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2236 if (soft != RLIM_INFINITY) {
2239 if (p->rt.watchdog_stamp != jiffies) {
2241 p->rt.watchdog_stamp = jiffies;
2244 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2245 if (p->rt.timeout > next)
2246 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2250 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2252 struct sched_rt_entity *rt_se = &p->rt;
2259 * RR tasks need a special form of timeslice management.
2260 * FIFO tasks have no timeslices.
2262 if (p->policy != SCHED_RR)
2265 if (--p->rt.time_slice)
2268 p->rt.time_slice = sched_rr_timeslice;
2271 * Requeue to the end of queue if we (and all of our ancestors) are not
2272 * the only element on the queue
2274 for_each_sched_rt_entity(rt_se) {
2275 if (rt_se->run_list.prev != rt_se->run_list.next) {
2276 requeue_task_rt(rq, p, 0);
2283 static void set_curr_task_rt(struct rq *rq)
2285 struct task_struct *p = rq->curr;
2287 p->se.exec_start = rq_clock_task(rq);
2289 /* The running task is never eligible for pushing */
2290 dequeue_pushable_task(rq, p);
2293 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2296 * Time slice is 0 for SCHED_FIFO tasks
2298 if (task->policy == SCHED_RR)
2299 return sched_rr_timeslice;
2304 const struct sched_class rt_sched_class = {
2305 .next = &fair_sched_class,
2306 .enqueue_task = enqueue_task_rt,
2307 .dequeue_task = dequeue_task_rt,
2308 .yield_task = yield_task_rt,
2310 .check_preempt_curr = check_preempt_curr_rt,
2312 .pick_next_task = pick_next_task_rt,
2313 .put_prev_task = put_prev_task_rt,
2316 .select_task_rq = select_task_rq_rt,
2318 .set_cpus_allowed = set_cpus_allowed_rt,
2319 .rq_online = rq_online_rt,
2320 .rq_offline = rq_offline_rt,
2321 .post_schedule = post_schedule_rt,
2322 .task_woken = task_woken_rt,
2323 .switched_from = switched_from_rt,
2326 .set_curr_task = set_curr_task_rt,
2327 .task_tick = task_tick_rt,
2329 .get_rr_interval = get_rr_interval_rt,
2331 .prio_changed = prio_changed_rt,
2332 .switched_to = switched_to_rt,
2334 .update_curr = update_curr_rt,
2337 #ifdef CONFIG_SCHED_DEBUG
2338 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2340 void print_rt_stats(struct seq_file *m, int cpu)
2343 struct rt_rq *rt_rq;
2346 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2347 print_rt_rq(m, cpu, rt_rq);
2350 #endif /* CONFIG_SCHED_DEBUG */