1 // SPDX-License-Identifier: GPL-2.0
3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
10 int sched_rr_timeslice = RR_TIMESLICE;
11 int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * 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 inline void do_start_rt_bandwidth(struct rt_bandwidth *rt_b)
55 raw_spin_lock(&rt_b->rt_runtime_lock);
56 if (!rt_b->rt_period_active) {
57 rt_b->rt_period_active = 1;
59 * SCHED_DEADLINE updates the bandwidth, as a run away
60 * RT task with a DL task could hog a CPU. But DL does
61 * not reset the period. If a deadline task was running
62 * without an RT task running, it can cause RT tasks to
63 * throttle when they start up. Kick the timer right away
64 * to update the period.
66 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
67 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
69 raw_spin_unlock(&rt_b->rt_runtime_lock);
72 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
74 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
77 do_start_rt_bandwidth(rt_b);
80 void init_rt_rq(struct rt_rq *rt_rq)
82 struct rt_prio_array *array;
85 array = &rt_rq->active;
86 for (i = 0; i < MAX_RT_PRIO; i++) {
87 INIT_LIST_HEAD(array->queue + i);
88 __clear_bit(i, array->bitmap);
90 /* delimiter for bitsearch: */
91 __set_bit(MAX_RT_PRIO, array->bitmap);
93 #if defined CONFIG_SMP
94 rt_rq->highest_prio.curr = MAX_RT_PRIO;
95 rt_rq->highest_prio.next = MAX_RT_PRIO;
96 rt_rq->rt_nr_migratory = 0;
97 rt_rq->overloaded = 0;
98 plist_head_init(&rt_rq->pushable_tasks);
99 #endif /* CONFIG_SMP */
100 /* We start is dequeued state, because no RT tasks are queued */
101 rt_rq->rt_queued = 0;
104 rt_rq->rt_throttled = 0;
105 rt_rq->rt_runtime = 0;
106 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
109 #ifdef CONFIG_RT_GROUP_SCHED
110 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
112 hrtimer_cancel(&rt_b->rt_period_timer);
115 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
117 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
119 #ifdef CONFIG_SCHED_DEBUG
120 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
122 return container_of(rt_se, struct task_struct, rt);
125 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
130 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
135 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
137 struct rt_rq *rt_rq = rt_se->rt_rq;
142 void free_rt_sched_group(struct task_group *tg)
147 destroy_rt_bandwidth(&tg->rt_bandwidth);
149 for_each_possible_cpu(i) {
160 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
161 struct sched_rt_entity *rt_se, int cpu,
162 struct sched_rt_entity *parent)
164 struct rq *rq = cpu_rq(cpu);
166 rt_rq->highest_prio.curr = MAX_RT_PRIO;
167 rt_rq->rt_nr_boosted = 0;
171 tg->rt_rq[cpu] = rt_rq;
172 tg->rt_se[cpu] = rt_se;
178 rt_se->rt_rq = &rq->rt;
180 rt_se->rt_rq = parent->my_q;
183 rt_se->parent = parent;
184 INIT_LIST_HEAD(&rt_se->run_list);
187 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
190 struct sched_rt_entity *rt_se;
193 tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL);
196 tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL);
200 init_rt_bandwidth(&tg->rt_bandwidth,
201 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
203 for_each_possible_cpu(i) {
204 rt_rq = kzalloc_node(sizeof(struct rt_rq),
205 GFP_KERNEL, cpu_to_node(i));
209 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
210 GFP_KERNEL, cpu_to_node(i));
215 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
216 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
227 #else /* CONFIG_RT_GROUP_SCHED */
229 #define rt_entity_is_task(rt_se) (1)
231 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
233 return container_of(rt_se, struct task_struct, rt);
236 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
238 return container_of(rt_rq, struct rq, rt);
241 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
243 struct task_struct *p = rt_task_of(rt_se);
248 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
250 struct rq *rq = rq_of_rt_se(rt_se);
255 void free_rt_sched_group(struct task_group *tg) { }
257 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
261 #endif /* CONFIG_RT_GROUP_SCHED */
265 static void pull_rt_task(struct rq *this_rq);
267 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
269 /* Try to pull RT tasks here if we lower this rq's prio */
270 return rq->rt.highest_prio.curr > prev->prio;
273 static inline int rt_overloaded(struct rq *rq)
275 return atomic_read(&rq->rd->rto_count);
278 static inline void rt_set_overload(struct rq *rq)
283 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
285 * Make sure the mask is visible before we set
286 * the overload count. That is checked to determine
287 * if we should look at the mask. It would be a shame
288 * if we looked at the mask, but the mask was not
291 * Matched by the barrier in pull_rt_task().
294 atomic_inc(&rq->rd->rto_count);
297 static inline void rt_clear_overload(struct rq *rq)
302 /* the order here really doesn't matter */
303 atomic_dec(&rq->rd->rto_count);
304 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
307 static void update_rt_migration(struct rt_rq *rt_rq)
309 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
310 if (!rt_rq->overloaded) {
311 rt_set_overload(rq_of_rt_rq(rt_rq));
312 rt_rq->overloaded = 1;
314 } else if (rt_rq->overloaded) {
315 rt_clear_overload(rq_of_rt_rq(rt_rq));
316 rt_rq->overloaded = 0;
320 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
322 struct task_struct *p;
324 if (!rt_entity_is_task(rt_se))
327 p = rt_task_of(rt_se);
328 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
330 rt_rq->rt_nr_total++;
331 if (p->nr_cpus_allowed > 1)
332 rt_rq->rt_nr_migratory++;
334 update_rt_migration(rt_rq);
337 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
339 struct task_struct *p;
341 if (!rt_entity_is_task(rt_se))
344 p = rt_task_of(rt_se);
345 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
347 rt_rq->rt_nr_total--;
348 if (p->nr_cpus_allowed > 1)
349 rt_rq->rt_nr_migratory--;
351 update_rt_migration(rt_rq);
354 static inline int has_pushable_tasks(struct rq *rq)
356 return !plist_head_empty(&rq->rt.pushable_tasks);
359 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
360 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
362 static void push_rt_tasks(struct rq *);
363 static void pull_rt_task(struct rq *);
365 static inline void rt_queue_push_tasks(struct rq *rq)
367 if (!has_pushable_tasks(rq))
370 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
373 static inline void rt_queue_pull_task(struct rq *rq)
375 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
378 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
380 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
381 plist_node_init(&p->pushable_tasks, p->prio);
382 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
384 /* Update the highest prio pushable task */
385 if (p->prio < rq->rt.highest_prio.next)
386 rq->rt.highest_prio.next = p->prio;
389 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
391 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
393 /* Update the new highest prio pushable task */
394 if (has_pushable_tasks(rq)) {
395 p = plist_first_entry(&rq->rt.pushable_tasks,
396 struct task_struct, pushable_tasks);
397 rq->rt.highest_prio.next = p->prio;
399 rq->rt.highest_prio.next = MAX_RT_PRIO;
404 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
408 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
413 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
418 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
422 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
427 static inline void pull_rt_task(struct rq *this_rq)
431 static inline void rt_queue_push_tasks(struct rq *rq)
434 #endif /* CONFIG_SMP */
436 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
437 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
439 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
444 #ifdef CONFIG_RT_GROUP_SCHED
446 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
451 return rt_rq->rt_runtime;
454 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
456 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
459 typedef struct task_group *rt_rq_iter_t;
461 static inline struct task_group *next_task_group(struct task_group *tg)
464 tg = list_entry_rcu(tg->list.next,
465 typeof(struct task_group), list);
466 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
468 if (&tg->list == &task_groups)
474 #define for_each_rt_rq(rt_rq, iter, rq) \
475 for (iter = container_of(&task_groups, typeof(*iter), list); \
476 (iter = next_task_group(iter)) && \
477 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
479 #define for_each_sched_rt_entity(rt_se) \
480 for (; rt_se; rt_se = rt_se->parent)
482 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
487 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
488 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
490 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
492 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
493 struct rq *rq = rq_of_rt_rq(rt_rq);
494 struct sched_rt_entity *rt_se;
496 int cpu = cpu_of(rq);
498 rt_se = rt_rq->tg->rt_se[cpu];
500 if (rt_rq->rt_nr_running) {
502 enqueue_top_rt_rq(rt_rq);
503 else if (!on_rt_rq(rt_se))
504 enqueue_rt_entity(rt_se, 0);
506 if (rt_rq->highest_prio.curr < curr->prio)
511 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
513 struct sched_rt_entity *rt_se;
514 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
516 rt_se = rt_rq->tg->rt_se[cpu];
519 dequeue_top_rt_rq(rt_rq);
520 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
521 cpufreq_update_util(rq_of_rt_rq(rt_rq), 0);
523 else if (on_rt_rq(rt_se))
524 dequeue_rt_entity(rt_se, 0);
527 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
529 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
532 static int rt_se_boosted(struct sched_rt_entity *rt_se)
534 struct rt_rq *rt_rq = group_rt_rq(rt_se);
535 struct task_struct *p;
538 return !!rt_rq->rt_nr_boosted;
540 p = rt_task_of(rt_se);
541 return p->prio != p->normal_prio;
545 static inline const struct cpumask *sched_rt_period_mask(void)
547 return this_rq()->rd->span;
550 static inline const struct cpumask *sched_rt_period_mask(void)
552 return cpu_online_mask;
557 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
559 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
562 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
564 return &rt_rq->tg->rt_bandwidth;
567 #else /* !CONFIG_RT_GROUP_SCHED */
569 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
571 return rt_rq->rt_runtime;
574 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
576 return ktime_to_ns(def_rt_bandwidth.rt_period);
579 typedef struct rt_rq *rt_rq_iter_t;
581 #define for_each_rt_rq(rt_rq, iter, rq) \
582 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
584 #define for_each_sched_rt_entity(rt_se) \
585 for (; rt_se; rt_se = NULL)
587 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
592 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
594 struct rq *rq = rq_of_rt_rq(rt_rq);
596 if (!rt_rq->rt_nr_running)
599 enqueue_top_rt_rq(rt_rq);
603 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
605 dequeue_top_rt_rq(rt_rq);
608 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
610 return rt_rq->rt_throttled;
613 static inline const struct cpumask *sched_rt_period_mask(void)
615 return cpu_online_mask;
619 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
621 return &cpu_rq(cpu)->rt;
624 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
626 return &def_rt_bandwidth;
629 #endif /* CONFIG_RT_GROUP_SCHED */
631 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
633 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
635 return (hrtimer_active(&rt_b->rt_period_timer) ||
636 rt_rq->rt_time < rt_b->rt_runtime);
641 * We ran out of runtime, see if we can borrow some from our neighbours.
643 static void do_balance_runtime(struct rt_rq *rt_rq)
645 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
646 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
650 weight = cpumask_weight(rd->span);
652 raw_spin_lock(&rt_b->rt_runtime_lock);
653 rt_period = ktime_to_ns(rt_b->rt_period);
654 for_each_cpu(i, rd->span) {
655 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
661 raw_spin_lock(&iter->rt_runtime_lock);
663 * Either all rqs have inf runtime and there's nothing to steal
664 * or __disable_runtime() below sets a specific rq to inf to
665 * indicate its been disabled and disalow stealing.
667 if (iter->rt_runtime == RUNTIME_INF)
671 * From runqueues with spare time, take 1/n part of their
672 * spare time, but no more than our period.
674 diff = iter->rt_runtime - iter->rt_time;
676 diff = div_u64((u64)diff, weight);
677 if (rt_rq->rt_runtime + diff > rt_period)
678 diff = rt_period - rt_rq->rt_runtime;
679 iter->rt_runtime -= diff;
680 rt_rq->rt_runtime += diff;
681 if (rt_rq->rt_runtime == rt_period) {
682 raw_spin_unlock(&iter->rt_runtime_lock);
687 raw_spin_unlock(&iter->rt_runtime_lock);
689 raw_spin_unlock(&rt_b->rt_runtime_lock);
693 * Ensure this RQ takes back all the runtime it lend to its neighbours.
695 static void __disable_runtime(struct rq *rq)
697 struct root_domain *rd = rq->rd;
701 if (unlikely(!scheduler_running))
704 for_each_rt_rq(rt_rq, iter, rq) {
705 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
709 raw_spin_lock(&rt_b->rt_runtime_lock);
710 raw_spin_lock(&rt_rq->rt_runtime_lock);
712 * Either we're all inf and nobody needs to borrow, or we're
713 * already disabled and thus have nothing to do, or we have
714 * exactly the right amount of runtime to take out.
716 if (rt_rq->rt_runtime == RUNTIME_INF ||
717 rt_rq->rt_runtime == rt_b->rt_runtime)
719 raw_spin_unlock(&rt_rq->rt_runtime_lock);
722 * Calculate the difference between what we started out with
723 * and what we current have, that's the amount of runtime
724 * we lend and now have to reclaim.
726 want = rt_b->rt_runtime - rt_rq->rt_runtime;
729 * Greedy reclaim, take back as much as we can.
731 for_each_cpu(i, rd->span) {
732 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
736 * Can't reclaim from ourselves or disabled runqueues.
738 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
741 raw_spin_lock(&iter->rt_runtime_lock);
743 diff = min_t(s64, iter->rt_runtime, want);
744 iter->rt_runtime -= diff;
747 iter->rt_runtime -= want;
750 raw_spin_unlock(&iter->rt_runtime_lock);
756 raw_spin_lock(&rt_rq->rt_runtime_lock);
758 * We cannot be left wanting - that would mean some runtime
759 * leaked out of the system.
764 * Disable all the borrow logic by pretending we have inf
765 * runtime - in which case borrowing doesn't make sense.
767 rt_rq->rt_runtime = RUNTIME_INF;
768 rt_rq->rt_throttled = 0;
769 raw_spin_unlock(&rt_rq->rt_runtime_lock);
770 raw_spin_unlock(&rt_b->rt_runtime_lock);
772 /* Make rt_rq available for pick_next_task() */
773 sched_rt_rq_enqueue(rt_rq);
777 static void __enable_runtime(struct rq *rq)
782 if (unlikely(!scheduler_running))
786 * Reset each runqueue's bandwidth settings
788 for_each_rt_rq(rt_rq, iter, rq) {
789 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
791 raw_spin_lock(&rt_b->rt_runtime_lock);
792 raw_spin_lock(&rt_rq->rt_runtime_lock);
793 rt_rq->rt_runtime = rt_b->rt_runtime;
795 rt_rq->rt_throttled = 0;
796 raw_spin_unlock(&rt_rq->rt_runtime_lock);
797 raw_spin_unlock(&rt_b->rt_runtime_lock);
801 static void balance_runtime(struct rt_rq *rt_rq)
803 if (!sched_feat(RT_RUNTIME_SHARE))
806 if (rt_rq->rt_time > rt_rq->rt_runtime) {
807 raw_spin_unlock(&rt_rq->rt_runtime_lock);
808 do_balance_runtime(rt_rq);
809 raw_spin_lock(&rt_rq->rt_runtime_lock);
812 #else /* !CONFIG_SMP */
813 static inline void balance_runtime(struct rt_rq *rt_rq) {}
814 #endif /* CONFIG_SMP */
816 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
818 int i, idle = 1, throttled = 0;
819 const struct cpumask *span;
821 span = sched_rt_period_mask();
822 #ifdef CONFIG_RT_GROUP_SCHED
824 * FIXME: isolated CPUs should really leave the root task group,
825 * whether they are isolcpus or were isolated via cpusets, lest
826 * the timer run on a CPU which does not service all runqueues,
827 * potentially leaving other CPUs indefinitely throttled. If
828 * isolation is really required, the user will turn the throttle
829 * off to kill the perturbations it causes anyway. Meanwhile,
830 * this maintains functionality for boot and/or troubleshooting.
832 if (rt_b == &root_task_group.rt_bandwidth)
833 span = cpu_online_mask;
835 for_each_cpu(i, span) {
837 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
838 struct rq *rq = rq_of_rt_rq(rt_rq);
842 * When span == cpu_online_mask, taking each rq->lock
843 * can be time-consuming. Try to avoid it when possible.
845 raw_spin_lock(&rt_rq->rt_runtime_lock);
846 if (!sched_feat(RT_RUNTIME_SHARE) && rt_rq->rt_runtime != RUNTIME_INF)
847 rt_rq->rt_runtime = rt_b->rt_runtime;
848 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
849 raw_spin_unlock(&rt_rq->rt_runtime_lock);
853 raw_spin_lock(&rq->lock);
856 if (rt_rq->rt_time) {
859 raw_spin_lock(&rt_rq->rt_runtime_lock);
860 if (rt_rq->rt_throttled)
861 balance_runtime(rt_rq);
862 runtime = rt_rq->rt_runtime;
863 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
864 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
865 rt_rq->rt_throttled = 0;
869 * When we're idle and a woken (rt) task is
870 * throttled check_preempt_curr() will set
871 * skip_update and the time between the wakeup
872 * and this unthrottle will get accounted as
875 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
876 rq_clock_cancel_skipupdate(rq);
878 if (rt_rq->rt_time || rt_rq->rt_nr_running)
880 raw_spin_unlock(&rt_rq->rt_runtime_lock);
881 } else if (rt_rq->rt_nr_running) {
883 if (!rt_rq_throttled(rt_rq))
886 if (rt_rq->rt_throttled)
890 sched_rt_rq_enqueue(rt_rq);
891 raw_spin_unlock(&rq->lock);
894 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
900 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
902 #ifdef CONFIG_RT_GROUP_SCHED
903 struct rt_rq *rt_rq = group_rt_rq(rt_se);
906 return rt_rq->highest_prio.curr;
909 return rt_task_of(rt_se)->prio;
912 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
914 u64 runtime = sched_rt_runtime(rt_rq);
916 if (rt_rq->rt_throttled)
917 return rt_rq_throttled(rt_rq);
919 if (runtime >= sched_rt_period(rt_rq))
922 balance_runtime(rt_rq);
923 runtime = sched_rt_runtime(rt_rq);
924 if (runtime == RUNTIME_INF)
927 if (rt_rq->rt_time > runtime) {
928 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
931 * Don't actually throttle groups that have no runtime assigned
932 * but accrue some time due to boosting.
934 if (likely(rt_b->rt_runtime)) {
935 rt_rq->rt_throttled = 1;
936 printk_deferred_once("sched: RT throttling activated\n");
939 * In case we did anyway, make it go away,
940 * replenishment is a joke, since it will replenish us
946 if (rt_rq_throttled(rt_rq)) {
947 sched_rt_rq_dequeue(rt_rq);
956 * Update the current task's runtime statistics. Skip current tasks that
957 * are not in our scheduling class.
959 static void update_curr_rt(struct rq *rq)
961 struct task_struct *curr = rq->curr;
962 struct sched_rt_entity *rt_se = &curr->rt;
966 if (curr->sched_class != &rt_sched_class)
969 now = rq_clock_task(rq);
970 delta_exec = now - curr->se.exec_start;
971 if (unlikely((s64)delta_exec <= 0))
974 schedstat_set(curr->se.statistics.exec_max,
975 max(curr->se.statistics.exec_max, delta_exec));
977 curr->se.sum_exec_runtime += delta_exec;
978 account_group_exec_runtime(curr, delta_exec);
980 curr->se.exec_start = now;
981 cgroup_account_cputime(curr, delta_exec);
983 if (!rt_bandwidth_enabled())
986 for_each_sched_rt_entity(rt_se) {
987 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
990 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
991 raw_spin_lock(&rt_rq->rt_runtime_lock);
992 rt_rq->rt_time += delta_exec;
993 exceeded = sched_rt_runtime_exceeded(rt_rq);
996 raw_spin_unlock(&rt_rq->rt_runtime_lock);
998 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq));
1004 dequeue_top_rt_rq(struct rt_rq *rt_rq)
1006 struct rq *rq = rq_of_rt_rq(rt_rq);
1008 BUG_ON(&rq->rt != rt_rq);
1010 if (!rt_rq->rt_queued)
1013 BUG_ON(!rq->nr_running);
1015 sub_nr_running(rq, rt_rq->rt_nr_running);
1016 rt_rq->rt_queued = 0;
1021 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1023 struct rq *rq = rq_of_rt_rq(rt_rq);
1025 BUG_ON(&rq->rt != rt_rq);
1027 if (rt_rq->rt_queued)
1030 if (rt_rq_throttled(rt_rq))
1033 if (rt_rq->rt_nr_running) {
1034 add_nr_running(rq, rt_rq->rt_nr_running);
1035 rt_rq->rt_queued = 1;
1038 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1039 cpufreq_update_util(rq, 0);
1042 #if defined CONFIG_SMP
1045 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1047 struct rq *rq = rq_of_rt_rq(rt_rq);
1049 #ifdef CONFIG_RT_GROUP_SCHED
1051 * Change rq's cpupri only if rt_rq is the top queue.
1053 if (&rq->rt != rt_rq)
1056 if (rq->online && prio < prev_prio)
1057 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1061 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1063 struct rq *rq = rq_of_rt_rq(rt_rq);
1065 #ifdef CONFIG_RT_GROUP_SCHED
1067 * Change rq's cpupri only if rt_rq is the top queue.
1069 if (&rq->rt != rt_rq)
1072 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1073 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1076 #else /* CONFIG_SMP */
1079 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1081 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1083 #endif /* CONFIG_SMP */
1085 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1087 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1089 int prev_prio = rt_rq->highest_prio.curr;
1091 if (prio < prev_prio)
1092 rt_rq->highest_prio.curr = prio;
1094 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1098 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1100 int prev_prio = rt_rq->highest_prio.curr;
1102 if (rt_rq->rt_nr_running) {
1104 WARN_ON(prio < prev_prio);
1107 * This may have been our highest task, and therefore
1108 * we may have some recomputation to do
1110 if (prio == prev_prio) {
1111 struct rt_prio_array *array = &rt_rq->active;
1113 rt_rq->highest_prio.curr =
1114 sched_find_first_bit(array->bitmap);
1118 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1120 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1125 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1126 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1128 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1130 #ifdef CONFIG_RT_GROUP_SCHED
1133 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1135 if (rt_se_boosted(rt_se))
1136 rt_rq->rt_nr_boosted++;
1139 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1143 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1145 if (rt_se_boosted(rt_se))
1146 rt_rq->rt_nr_boosted--;
1148 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1151 #else /* CONFIG_RT_GROUP_SCHED */
1154 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1156 start_rt_bandwidth(&def_rt_bandwidth);
1160 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1162 #endif /* CONFIG_RT_GROUP_SCHED */
1165 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1167 struct rt_rq *group_rq = group_rt_rq(rt_se);
1170 return group_rq->rt_nr_running;
1176 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1178 struct rt_rq *group_rq = group_rt_rq(rt_se);
1179 struct task_struct *tsk;
1182 return group_rq->rr_nr_running;
1184 tsk = rt_task_of(rt_se);
1186 return (tsk->policy == SCHED_RR) ? 1 : 0;
1190 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1192 int prio = rt_se_prio(rt_se);
1194 WARN_ON(!rt_prio(prio));
1195 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1196 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1198 inc_rt_prio(rt_rq, prio);
1199 inc_rt_migration(rt_se, rt_rq);
1200 inc_rt_group(rt_se, rt_rq);
1204 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1206 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1207 WARN_ON(!rt_rq->rt_nr_running);
1208 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1209 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1211 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1212 dec_rt_migration(rt_se, rt_rq);
1213 dec_rt_group(rt_se, rt_rq);
1217 * Change rt_se->run_list location unless SAVE && !MOVE
1219 * assumes ENQUEUE/DEQUEUE flags match
1221 static inline bool move_entity(unsigned int flags)
1223 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1229 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1231 list_del_init(&rt_se->run_list);
1233 if (list_empty(array->queue + rt_se_prio(rt_se)))
1234 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1239 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1241 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1242 struct rt_prio_array *array = &rt_rq->active;
1243 struct rt_rq *group_rq = group_rt_rq(rt_se);
1244 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1247 * Don't enqueue the group if its throttled, or when empty.
1248 * The latter is a consequence of the former when a child group
1249 * get throttled and the current group doesn't have any other
1252 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1254 __delist_rt_entity(rt_se, array);
1258 if (move_entity(flags)) {
1259 WARN_ON_ONCE(rt_se->on_list);
1260 if (flags & ENQUEUE_HEAD)
1261 list_add(&rt_se->run_list, queue);
1263 list_add_tail(&rt_se->run_list, queue);
1265 __set_bit(rt_se_prio(rt_se), array->bitmap);
1270 inc_rt_tasks(rt_se, rt_rq);
1273 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1275 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1276 struct rt_prio_array *array = &rt_rq->active;
1278 if (move_entity(flags)) {
1279 WARN_ON_ONCE(!rt_se->on_list);
1280 __delist_rt_entity(rt_se, array);
1284 dec_rt_tasks(rt_se, rt_rq);
1288 * Because the prio of an upper entry depends on the lower
1289 * entries, we must remove entries top - down.
1291 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1293 struct sched_rt_entity *back = NULL;
1295 for_each_sched_rt_entity(rt_se) {
1300 dequeue_top_rt_rq(rt_rq_of_se(back));
1302 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1303 if (on_rt_rq(rt_se))
1304 __dequeue_rt_entity(rt_se, flags);
1308 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1310 struct rq *rq = rq_of_rt_se(rt_se);
1312 dequeue_rt_stack(rt_se, flags);
1313 for_each_sched_rt_entity(rt_se)
1314 __enqueue_rt_entity(rt_se, flags);
1315 enqueue_top_rt_rq(&rq->rt);
1318 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1320 struct rq *rq = rq_of_rt_se(rt_se);
1322 dequeue_rt_stack(rt_se, flags);
1324 for_each_sched_rt_entity(rt_se) {
1325 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1327 if (rt_rq && rt_rq->rt_nr_running)
1328 __enqueue_rt_entity(rt_se, flags);
1330 enqueue_top_rt_rq(&rq->rt);
1334 * Adding/removing a task to/from a priority array:
1337 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1339 struct sched_rt_entity *rt_se = &p->rt;
1341 if (flags & ENQUEUE_WAKEUP)
1344 enqueue_rt_entity(rt_se, flags);
1346 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1347 enqueue_pushable_task(rq, p);
1350 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1352 struct sched_rt_entity *rt_se = &p->rt;
1355 dequeue_rt_entity(rt_se, flags);
1357 dequeue_pushable_task(rq, p);
1361 * Put task to the head or the end of the run list without the overhead of
1362 * dequeue followed by enqueue.
1365 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1367 if (on_rt_rq(rt_se)) {
1368 struct rt_prio_array *array = &rt_rq->active;
1369 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1372 list_move(&rt_se->run_list, queue);
1374 list_move_tail(&rt_se->run_list, queue);
1378 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1380 struct sched_rt_entity *rt_se = &p->rt;
1381 struct rt_rq *rt_rq;
1383 for_each_sched_rt_entity(rt_se) {
1384 rt_rq = rt_rq_of_se(rt_se);
1385 requeue_rt_entity(rt_rq, rt_se, head);
1389 static void yield_task_rt(struct rq *rq)
1391 requeue_task_rt(rq, rq->curr, 0);
1395 static int find_lowest_rq(struct task_struct *task);
1398 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1400 struct task_struct *curr;
1403 /* For anything but wake ups, just return the task_cpu */
1404 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1410 curr = READ_ONCE(rq->curr); /* unlocked access */
1413 * If the current task on @p's runqueue is an RT task, then
1414 * try to see if we can wake this RT task up on another
1415 * runqueue. Otherwise simply start this RT task
1416 * on its current runqueue.
1418 * We want to avoid overloading runqueues. If the woken
1419 * task is a higher priority, then it will stay on this CPU
1420 * and the lower prio task should be moved to another CPU.
1421 * Even though this will probably make the lower prio task
1422 * lose its cache, we do not want to bounce a higher task
1423 * around just because it gave up its CPU, perhaps for a
1426 * For equal prio tasks, we just let the scheduler sort it out.
1428 * Otherwise, just let it ride on the affined RQ and the
1429 * post-schedule router will push the preempted task away
1431 * This test is optimistic, if we get it wrong the load-balancer
1432 * will have to sort it out.
1434 if (curr && unlikely(rt_task(curr)) &&
1435 (curr->nr_cpus_allowed < 2 ||
1436 curr->prio <= p->prio)) {
1437 int target = find_lowest_rq(p);
1440 * Don't bother moving it if the destination CPU is
1441 * not running a lower priority task.
1444 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1453 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1456 * Current can't be migrated, useless to reschedule,
1457 * let's hope p can move out.
1459 if (rq->curr->nr_cpus_allowed == 1 ||
1460 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1464 * p is migratable, so let's not schedule it and
1465 * see if it is pushed or pulled somewhere else.
1467 if (p->nr_cpus_allowed != 1
1468 && cpupri_find(&rq->rd->cpupri, p, NULL))
1472 * There appear to be other CPUs that can accept
1473 * the current task but none can run 'p', so lets reschedule
1474 * to try and push the current task away:
1476 requeue_task_rt(rq, p, 1);
1480 #endif /* CONFIG_SMP */
1483 * Preempt the current task with a newly woken task if needed:
1485 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1487 if (p->prio < rq->curr->prio) {
1496 * - the newly woken task is of equal priority to the current task
1497 * - the newly woken task is non-migratable while current is migratable
1498 * - current will be preempted on the next reschedule
1500 * we should check to see if current can readily move to a different
1501 * cpu. If so, we will reschedule to allow the push logic to try
1502 * to move current somewhere else, making room for our non-migratable
1505 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1506 check_preempt_equal_prio(rq, p);
1510 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1511 struct rt_rq *rt_rq)
1513 struct rt_prio_array *array = &rt_rq->active;
1514 struct sched_rt_entity *next = NULL;
1515 struct list_head *queue;
1518 idx = sched_find_first_bit(array->bitmap);
1519 BUG_ON(idx >= MAX_RT_PRIO);
1521 queue = array->queue + idx;
1522 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1527 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1529 struct sched_rt_entity *rt_se;
1530 struct task_struct *p;
1531 struct rt_rq *rt_rq = &rq->rt;
1534 rt_se = pick_next_rt_entity(rq, rt_rq);
1536 rt_rq = group_rt_rq(rt_se);
1539 p = rt_task_of(rt_se);
1540 p->se.exec_start = rq_clock_task(rq);
1545 static struct task_struct *
1546 pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1548 struct task_struct *p;
1549 struct rt_rq *rt_rq = &rq->rt;
1551 if (need_pull_rt_task(rq, prev)) {
1553 * This is OK, because current is on_cpu, which avoids it being
1554 * picked for load-balance and preemption/IRQs are still
1555 * disabled avoiding further scheduler activity on it and we're
1556 * being very careful to re-start the picking loop.
1558 rq_unpin_lock(rq, rf);
1560 rq_repin_lock(rq, rf);
1562 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1563 * means a dl or stop task can slip in, in which case we need
1564 * to re-start task selection.
1566 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1567 rq->dl.dl_nr_running))
1572 * We may dequeue prev's rt_rq in put_prev_task().
1573 * So, we update time before rt_nr_running check.
1575 if (prev->sched_class == &rt_sched_class)
1578 if (!rt_rq->rt_queued)
1581 put_prev_task(rq, prev);
1583 p = _pick_next_task_rt(rq);
1585 /* The running task is never eligible for pushing */
1586 dequeue_pushable_task(rq, p);
1588 rt_queue_push_tasks(rq);
1591 * If prev task was rt, put_prev_task() has already updated the
1592 * utilization. We only care of the case where we start to schedule a
1595 if (rq->curr->sched_class != &rt_sched_class)
1596 update_rt_rq_load_avg(rq_clock_task(rq), rq, 0);
1601 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1605 update_rt_rq_load_avg(rq_clock_task(rq), rq, 1);
1608 * The previous task needs to be made eligible for pushing
1609 * if it is still active
1611 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1612 enqueue_pushable_task(rq, p);
1617 /* Only try algorithms three times */
1618 #define RT_MAX_TRIES 3
1620 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1622 if (!task_running(rq, p) &&
1623 cpumask_test_cpu(cpu, &p->cpus_allowed))
1630 * Return the highest pushable rq's task, which is suitable to be executed
1631 * on the CPU, NULL otherwise
1633 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1635 struct plist_head *head = &rq->rt.pushable_tasks;
1636 struct task_struct *p;
1638 if (!has_pushable_tasks(rq))
1641 plist_for_each_entry(p, head, pushable_tasks) {
1642 if (pick_rt_task(rq, p, cpu))
1649 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1651 static int find_lowest_rq(struct task_struct *task)
1653 struct sched_domain *sd;
1654 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1655 int this_cpu = smp_processor_id();
1656 int cpu = task_cpu(task);
1658 /* Make sure the mask is initialized first */
1659 if (unlikely(!lowest_mask))
1662 if (task->nr_cpus_allowed == 1)
1663 return -1; /* No other targets possible */
1665 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1666 return -1; /* No targets found */
1669 * At this point we have built a mask of CPUs representing the
1670 * lowest priority tasks in the system. Now we want to elect
1671 * the best one based on our affinity and topology.
1673 * We prioritize the last CPU that the task executed on since
1674 * it is most likely cache-hot in that location.
1676 if (cpumask_test_cpu(cpu, lowest_mask))
1680 * Otherwise, we consult the sched_domains span maps to figure
1681 * out which CPU is logically closest to our hot cache data.
1683 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1684 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1687 for_each_domain(cpu, sd) {
1688 if (sd->flags & SD_WAKE_AFFINE) {
1692 * "this_cpu" is cheaper to preempt than a
1695 if (this_cpu != -1 &&
1696 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1701 best_cpu = cpumask_first_and(lowest_mask,
1702 sched_domain_span(sd));
1703 if (best_cpu < nr_cpu_ids) {
1712 * And finally, if there were no matches within the domains
1713 * just give the caller *something* to work with from the compatible
1719 cpu = cpumask_any(lowest_mask);
1720 if (cpu < nr_cpu_ids)
1726 /* Will lock the rq it finds */
1727 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1729 struct rq *lowest_rq = NULL;
1733 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1734 cpu = find_lowest_rq(task);
1736 if ((cpu == -1) || (cpu == rq->cpu))
1739 lowest_rq = cpu_rq(cpu);
1741 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1743 * Target rq has tasks of equal or higher priority,
1744 * retrying does not release any lock and is unlikely
1745 * to yield a different result.
1751 /* if the prio of this runqueue changed, try again */
1752 if (double_lock_balance(rq, lowest_rq)) {
1754 * We had to unlock the run queue. In
1755 * the mean time, task could have
1756 * migrated already or had its affinity changed.
1757 * Also make sure that it wasn't scheduled on its rq.
1759 if (unlikely(task_rq(task) != rq ||
1760 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) ||
1761 task_running(rq, task) ||
1763 !task_on_rq_queued(task))) {
1765 double_unlock_balance(rq, lowest_rq);
1771 /* If this rq is still suitable use it. */
1772 if (lowest_rq->rt.highest_prio.curr > task->prio)
1776 double_unlock_balance(rq, lowest_rq);
1783 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1785 struct task_struct *p;
1787 if (!has_pushable_tasks(rq))
1790 p = plist_first_entry(&rq->rt.pushable_tasks,
1791 struct task_struct, pushable_tasks);
1793 BUG_ON(rq->cpu != task_cpu(p));
1794 BUG_ON(task_current(rq, p));
1795 BUG_ON(p->nr_cpus_allowed <= 1);
1797 BUG_ON(!task_on_rq_queued(p));
1798 BUG_ON(!rt_task(p));
1804 * If the current CPU has more than one RT task, see if the non
1805 * running task can migrate over to a CPU that is running a task
1806 * of lesser priority.
1808 static int push_rt_task(struct rq *rq)
1810 struct task_struct *next_task;
1811 struct rq *lowest_rq;
1814 if (!rq->rt.overloaded)
1817 next_task = pick_next_pushable_task(rq);
1822 if (unlikely(next_task == rq->curr)) {
1828 * It's possible that the next_task slipped in of
1829 * higher priority than current. If that's the case
1830 * just reschedule current.
1832 if (unlikely(next_task->prio < rq->curr->prio)) {
1837 /* We might release rq lock */
1838 get_task_struct(next_task);
1840 /* find_lock_lowest_rq locks the rq if found */
1841 lowest_rq = find_lock_lowest_rq(next_task, rq);
1843 struct task_struct *task;
1845 * find_lock_lowest_rq releases rq->lock
1846 * so it is possible that next_task has migrated.
1848 * We need to make sure that the task is still on the same
1849 * run-queue and is also still the next task eligible for
1852 task = pick_next_pushable_task(rq);
1853 if (task == next_task) {
1855 * The task hasn't migrated, and is still the next
1856 * eligible task, but we failed to find a run-queue
1857 * to push it to. Do not retry in this case, since
1858 * other CPUs will pull from us when ready.
1864 /* No more tasks, just exit */
1868 * Something has shifted, try again.
1870 put_task_struct(next_task);
1875 deactivate_task(rq, next_task, 0);
1876 set_task_cpu(next_task, lowest_rq->cpu);
1877 activate_task(lowest_rq, next_task, 0);
1880 resched_curr(lowest_rq);
1882 double_unlock_balance(rq, lowest_rq);
1885 put_task_struct(next_task);
1890 static void push_rt_tasks(struct rq *rq)
1892 /* push_rt_task will return true if it moved an RT */
1893 while (push_rt_task(rq))
1897 #ifdef HAVE_RT_PUSH_IPI
1900 * When a high priority task schedules out from a CPU and a lower priority
1901 * task is scheduled in, a check is made to see if there's any RT tasks
1902 * on other CPUs that are waiting to run because a higher priority RT task
1903 * is currently running on its CPU. In this case, the CPU with multiple RT
1904 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1905 * up that may be able to run one of its non-running queued RT tasks.
1907 * All CPUs with overloaded RT tasks need to be notified as there is currently
1908 * no way to know which of these CPUs have the highest priority task waiting
1909 * to run. Instead of trying to take a spinlock on each of these CPUs,
1910 * which has shown to cause large latency when done on machines with many
1911 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1912 * RT tasks waiting to run.
1914 * Just sending an IPI to each of the CPUs is also an issue, as on large
1915 * count CPU machines, this can cause an IPI storm on a CPU, especially
1916 * if its the only CPU with multiple RT tasks queued, and a large number
1917 * of CPUs scheduling a lower priority task at the same time.
1919 * Each root domain has its own irq work function that can iterate over
1920 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1921 * tassk must be checked if there's one or many CPUs that are lowering
1922 * their priority, there's a single irq work iterator that will try to
1923 * push off RT tasks that are waiting to run.
1925 * When a CPU schedules a lower priority task, it will kick off the
1926 * irq work iterator that will jump to each CPU with overloaded RT tasks.
1927 * As it only takes the first CPU that schedules a lower priority task
1928 * to start the process, the rto_start variable is incremented and if
1929 * the atomic result is one, then that CPU will try to take the rto_lock.
1930 * This prevents high contention on the lock as the process handles all
1931 * CPUs scheduling lower priority tasks.
1933 * All CPUs that are scheduling a lower priority task will increment the
1934 * rt_loop_next variable. This will make sure that the irq work iterator
1935 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1936 * priority task, even if the iterator is in the middle of a scan. Incrementing
1937 * the rt_loop_next will cause the iterator to perform another scan.
1940 static int rto_next_cpu(struct root_domain *rd)
1946 * When starting the IPI RT pushing, the rto_cpu is set to -1,
1947 * rt_next_cpu() will simply return the first CPU found in
1950 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
1951 * will return the next CPU found in the rto_mask.
1953 * If there are no more CPUs left in the rto_mask, then a check is made
1954 * against rto_loop and rto_loop_next. rto_loop is only updated with
1955 * the rto_lock held, but any CPU may increment the rto_loop_next
1956 * without any locking.
1960 /* When rto_cpu is -1 this acts like cpumask_first() */
1961 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
1965 if (cpu < nr_cpu_ids)
1971 * ACQUIRE ensures we see the @rto_mask changes
1972 * made prior to the @next value observed.
1974 * Matches WMB in rt_set_overload().
1976 next = atomic_read_acquire(&rd->rto_loop_next);
1978 if (rd->rto_loop == next)
1981 rd->rto_loop = next;
1987 static inline bool rto_start_trylock(atomic_t *v)
1989 return !atomic_cmpxchg_acquire(v, 0, 1);
1992 static inline void rto_start_unlock(atomic_t *v)
1994 atomic_set_release(v, 0);
1997 static void tell_cpu_to_push(struct rq *rq)
2001 /* Keep the loop going if the IPI is currently active */
2002 atomic_inc(&rq->rd->rto_loop_next);
2004 /* Only one CPU can initiate a loop at a time */
2005 if (!rto_start_trylock(&rq->rd->rto_loop_start))
2008 raw_spin_lock(&rq->rd->rto_lock);
2011 * The rto_cpu is updated under the lock, if it has a valid CPU
2012 * then the IPI is still running and will continue due to the
2013 * update to loop_next, and nothing needs to be done here.
2014 * Otherwise it is finishing up and an ipi needs to be sent.
2016 if (rq->rd->rto_cpu < 0)
2017 cpu = rto_next_cpu(rq->rd);
2019 raw_spin_unlock(&rq->rd->rto_lock);
2021 rto_start_unlock(&rq->rd->rto_loop_start);
2024 /* Make sure the rd does not get freed while pushing */
2025 sched_get_rd(rq->rd);
2026 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2030 /* Called from hardirq context */
2031 void rto_push_irq_work_func(struct irq_work *work)
2033 struct root_domain *rd =
2034 container_of(work, struct root_domain, rto_push_work);
2041 * We do not need to grab the lock to check for has_pushable_tasks.
2042 * When it gets updated, a check is made if a push is possible.
2044 if (has_pushable_tasks(rq)) {
2045 raw_spin_lock(&rq->lock);
2047 raw_spin_unlock(&rq->lock);
2050 raw_spin_lock(&rd->rto_lock);
2052 /* Pass the IPI to the next rt overloaded queue */
2053 cpu = rto_next_cpu(rd);
2055 raw_spin_unlock(&rd->rto_lock);
2062 /* Try the next RT overloaded CPU */
2063 irq_work_queue_on(&rd->rto_push_work, cpu);
2065 #endif /* HAVE_RT_PUSH_IPI */
2067 static void pull_rt_task(struct rq *this_rq)
2069 int this_cpu = this_rq->cpu, cpu;
2070 bool resched = false;
2071 struct task_struct *p;
2073 int rt_overload_count = rt_overloaded(this_rq);
2075 if (likely(!rt_overload_count))
2079 * Match the barrier from rt_set_overloaded; this guarantees that if we
2080 * see overloaded we must also see the rto_mask bit.
2084 /* If we are the only overloaded CPU do nothing */
2085 if (rt_overload_count == 1 &&
2086 cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2089 #ifdef HAVE_RT_PUSH_IPI
2090 if (sched_feat(RT_PUSH_IPI)) {
2091 tell_cpu_to_push(this_rq);
2096 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2097 if (this_cpu == cpu)
2100 src_rq = cpu_rq(cpu);
2103 * Don't bother taking the src_rq->lock if the next highest
2104 * task is known to be lower-priority than our current task.
2105 * This may look racy, but if this value is about to go
2106 * logically higher, the src_rq will push this task away.
2107 * And if its going logically lower, we do not care
2109 if (src_rq->rt.highest_prio.next >=
2110 this_rq->rt.highest_prio.curr)
2114 * We can potentially drop this_rq's lock in
2115 * double_lock_balance, and another CPU could
2118 double_lock_balance(this_rq, src_rq);
2121 * We can pull only a task, which is pushable
2122 * on its rq, and no others.
2124 p = pick_highest_pushable_task(src_rq, this_cpu);
2127 * Do we have an RT task that preempts
2128 * the to-be-scheduled task?
2130 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2131 WARN_ON(p == src_rq->curr);
2132 WARN_ON(!task_on_rq_queued(p));
2135 * There's a chance that p is higher in priority
2136 * than what's currently running on its CPU.
2137 * This is just that p is wakeing up and hasn't
2138 * had a chance to schedule. We only pull
2139 * p if it is lower in priority than the
2140 * current task on the run queue
2142 if (p->prio < src_rq->curr->prio)
2147 deactivate_task(src_rq, p, 0);
2148 set_task_cpu(p, this_cpu);
2149 activate_task(this_rq, p, 0);
2151 * We continue with the search, just in
2152 * case there's an even higher prio task
2153 * in another runqueue. (low likelihood
2158 double_unlock_balance(this_rq, src_rq);
2162 resched_curr(this_rq);
2166 * If we are not running and we are not going to reschedule soon, we should
2167 * try to push tasks away now
2169 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2171 if (!task_running(rq, p) &&
2172 !test_tsk_need_resched(rq->curr) &&
2173 p->nr_cpus_allowed > 1 &&
2174 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2175 (rq->curr->nr_cpus_allowed < 2 ||
2176 rq->curr->prio <= p->prio))
2180 /* Assumes rq->lock is held */
2181 static void rq_online_rt(struct rq *rq)
2183 if (rq->rt.overloaded)
2184 rt_set_overload(rq);
2186 __enable_runtime(rq);
2188 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2191 /* Assumes rq->lock is held */
2192 static void rq_offline_rt(struct rq *rq)
2194 if (rq->rt.overloaded)
2195 rt_clear_overload(rq);
2197 __disable_runtime(rq);
2199 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2203 * When switch from the rt queue, we bring ourselves to a position
2204 * that we might want to pull RT tasks from other runqueues.
2206 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2209 * If there are other RT tasks then we will reschedule
2210 * and the scheduling of the other RT tasks will handle
2211 * the balancing. But if we are the last RT task
2212 * we may need to handle the pulling of RT tasks
2215 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2218 rt_queue_pull_task(rq);
2221 void __init init_sched_rt_class(void)
2225 for_each_possible_cpu(i) {
2226 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2227 GFP_KERNEL, cpu_to_node(i));
2230 #endif /* CONFIG_SMP */
2233 * When switching a task to RT, we may overload the runqueue
2234 * with RT tasks. In this case we try to push them off to
2237 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2240 * If we are already running, then there's nothing
2241 * that needs to be done. But if we are not running
2242 * we may need to preempt the current running task.
2243 * If that current running task is also an RT task
2244 * then see if we can move to another run queue.
2246 if (task_on_rq_queued(p) && rq->curr != p) {
2248 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2249 rt_queue_push_tasks(rq);
2250 #endif /* CONFIG_SMP */
2251 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2257 * Priority of the task has changed. This may cause
2258 * us to initiate a push or pull.
2261 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2263 if (!task_on_rq_queued(p))
2266 if (rq->curr == p) {
2269 * If our priority decreases while running, we
2270 * may need to pull tasks to this runqueue.
2272 if (oldprio < p->prio)
2273 rt_queue_pull_task(rq);
2276 * If there's a higher priority task waiting to run
2279 if (p->prio > rq->rt.highest_prio.curr)
2282 /* For UP simply resched on drop of prio */
2283 if (oldprio < p->prio)
2285 #endif /* CONFIG_SMP */
2288 * This task is not running, but if it is
2289 * greater than the current running task
2292 if (p->prio < rq->curr->prio)
2297 #ifdef CONFIG_POSIX_TIMERS
2298 static void watchdog(struct rq *rq, struct task_struct *p)
2300 unsigned long soft, hard;
2302 /* max may change after cur was read, this will be fixed next tick */
2303 soft = task_rlimit(p, RLIMIT_RTTIME);
2304 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2306 if (soft != RLIM_INFINITY) {
2309 if (p->rt.watchdog_stamp != jiffies) {
2311 p->rt.watchdog_stamp = jiffies;
2314 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2315 if (p->rt.timeout > next)
2316 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2320 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2324 * scheduler tick hitting a task of our scheduling class.
2326 * NOTE: This function can be called remotely by the tick offload that
2327 * goes along full dynticks. Therefore no local assumption can be made
2328 * and everything must be accessed through the @rq and @curr passed in
2331 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2333 struct sched_rt_entity *rt_se = &p->rt;
2336 update_rt_rq_load_avg(rq_clock_task(rq), rq, 1);
2341 * RR tasks need a special form of timeslice management.
2342 * FIFO tasks have no timeslices.
2344 if (p->policy != SCHED_RR)
2347 if (--p->rt.time_slice)
2350 p->rt.time_slice = sched_rr_timeslice;
2353 * Requeue to the end of queue if we (and all of our ancestors) are not
2354 * the only element on the queue
2356 for_each_sched_rt_entity(rt_se) {
2357 if (rt_se->run_list.prev != rt_se->run_list.next) {
2358 requeue_task_rt(rq, p, 0);
2365 static void set_curr_task_rt(struct rq *rq)
2367 struct task_struct *p = rq->curr;
2369 p->se.exec_start = rq_clock_task(rq);
2371 /* The running task is never eligible for pushing */
2372 dequeue_pushable_task(rq, p);
2375 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2378 * Time slice is 0 for SCHED_FIFO tasks
2380 if (task->policy == SCHED_RR)
2381 return sched_rr_timeslice;
2386 const struct sched_class rt_sched_class = {
2387 .next = &fair_sched_class,
2388 .enqueue_task = enqueue_task_rt,
2389 .dequeue_task = dequeue_task_rt,
2390 .yield_task = yield_task_rt,
2392 .check_preempt_curr = check_preempt_curr_rt,
2394 .pick_next_task = pick_next_task_rt,
2395 .put_prev_task = put_prev_task_rt,
2398 .select_task_rq = select_task_rq_rt,
2400 .set_cpus_allowed = set_cpus_allowed_common,
2401 .rq_online = rq_online_rt,
2402 .rq_offline = rq_offline_rt,
2403 .task_woken = task_woken_rt,
2404 .switched_from = switched_from_rt,
2407 .set_curr_task = set_curr_task_rt,
2408 .task_tick = task_tick_rt,
2410 .get_rr_interval = get_rr_interval_rt,
2412 .prio_changed = prio_changed_rt,
2413 .switched_to = switched_to_rt,
2415 .update_curr = update_curr_rt,
2418 #ifdef CONFIG_RT_GROUP_SCHED
2420 * Ensure that the real time constraints are schedulable.
2422 static DEFINE_MUTEX(rt_constraints_mutex);
2424 /* Must be called with tasklist_lock held */
2425 static inline int tg_has_rt_tasks(struct task_group *tg)
2427 struct task_struct *g, *p;
2430 * Autogroups do not have RT tasks; see autogroup_create().
2432 if (task_group_is_autogroup(tg))
2435 for_each_process_thread(g, p) {
2436 if (rt_task(p) && task_group(p) == tg)
2443 struct rt_schedulable_data {
2444 struct task_group *tg;
2449 static int tg_rt_schedulable(struct task_group *tg, void *data)
2451 struct rt_schedulable_data *d = data;
2452 struct task_group *child;
2453 unsigned long total, sum = 0;
2454 u64 period, runtime;
2456 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2457 runtime = tg->rt_bandwidth.rt_runtime;
2460 period = d->rt_period;
2461 runtime = d->rt_runtime;
2465 * Cannot have more runtime than the period.
2467 if (runtime > period && runtime != RUNTIME_INF)
2471 * Ensure we don't starve existing RT tasks.
2473 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
2476 total = to_ratio(period, runtime);
2479 * Nobody can have more than the global setting allows.
2481 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2485 * The sum of our children's runtime should not exceed our own.
2487 list_for_each_entry_rcu(child, &tg->children, siblings) {
2488 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2489 runtime = child->rt_bandwidth.rt_runtime;
2491 if (child == d->tg) {
2492 period = d->rt_period;
2493 runtime = d->rt_runtime;
2496 sum += to_ratio(period, runtime);
2505 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2509 struct rt_schedulable_data data = {
2511 .rt_period = period,
2512 .rt_runtime = runtime,
2516 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2522 static int tg_set_rt_bandwidth(struct task_group *tg,
2523 u64 rt_period, u64 rt_runtime)
2528 * Disallowing the root group RT runtime is BAD, it would disallow the
2529 * kernel creating (and or operating) RT threads.
2531 if (tg == &root_task_group && rt_runtime == 0)
2534 /* No period doesn't make any sense. */
2538 mutex_lock(&rt_constraints_mutex);
2539 read_lock(&tasklist_lock);
2540 err = __rt_schedulable(tg, rt_period, rt_runtime);
2544 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2545 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2546 tg->rt_bandwidth.rt_runtime = rt_runtime;
2548 for_each_possible_cpu(i) {
2549 struct rt_rq *rt_rq = tg->rt_rq[i];
2551 raw_spin_lock(&rt_rq->rt_runtime_lock);
2552 rt_rq->rt_runtime = rt_runtime;
2553 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2555 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2557 read_unlock(&tasklist_lock);
2558 mutex_unlock(&rt_constraints_mutex);
2563 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2565 u64 rt_runtime, rt_period;
2567 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2568 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2569 if (rt_runtime_us < 0)
2570 rt_runtime = RUNTIME_INF;
2571 else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC)
2574 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2577 long sched_group_rt_runtime(struct task_group *tg)
2581 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2584 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2585 do_div(rt_runtime_us, NSEC_PER_USEC);
2586 return rt_runtime_us;
2589 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2591 u64 rt_runtime, rt_period;
2593 if (rt_period_us > U64_MAX / NSEC_PER_USEC)
2596 rt_period = rt_period_us * NSEC_PER_USEC;
2597 rt_runtime = tg->rt_bandwidth.rt_runtime;
2599 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2602 long sched_group_rt_period(struct task_group *tg)
2606 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2607 do_div(rt_period_us, NSEC_PER_USEC);
2608 return rt_period_us;
2611 static int sched_rt_global_constraints(void)
2615 mutex_lock(&rt_constraints_mutex);
2616 read_lock(&tasklist_lock);
2617 ret = __rt_schedulable(NULL, 0, 0);
2618 read_unlock(&tasklist_lock);
2619 mutex_unlock(&rt_constraints_mutex);
2624 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2626 /* Don't accept realtime tasks when there is no way for them to run */
2627 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2633 #else /* !CONFIG_RT_GROUP_SCHED */
2634 static int sched_rt_global_constraints(void)
2636 unsigned long flags;
2639 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2640 for_each_possible_cpu(i) {
2641 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2643 raw_spin_lock(&rt_rq->rt_runtime_lock);
2644 rt_rq->rt_runtime = global_rt_runtime();
2645 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2647 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2651 #endif /* CONFIG_RT_GROUP_SCHED */
2653 static int sched_rt_global_validate(void)
2655 if (sysctl_sched_rt_period <= 0)
2658 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2659 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
2665 static void sched_rt_do_global(void)
2667 unsigned long flags;
2669 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2670 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2671 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2672 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2675 int sched_rt_handler(struct ctl_table *table, int write,
2676 void __user *buffer, size_t *lenp,
2679 int old_period, old_runtime;
2680 static DEFINE_MUTEX(mutex);
2684 old_period = sysctl_sched_rt_period;
2685 old_runtime = sysctl_sched_rt_runtime;
2687 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2689 if (!ret && write) {
2690 ret = sched_rt_global_validate();
2694 ret = sched_dl_global_validate();
2698 ret = sched_rt_global_constraints();
2702 sched_rt_do_global();
2703 sched_dl_do_global();
2707 sysctl_sched_rt_period = old_period;
2708 sysctl_sched_rt_runtime = old_runtime;
2710 mutex_unlock(&mutex);
2715 int sched_rr_handler(struct ctl_table *table, int write,
2716 void __user *buffer, size_t *lenp,
2720 static DEFINE_MUTEX(mutex);
2723 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2725 * Make sure that internally we keep jiffies.
2726 * Also, writing zero resets the timeslice to default:
2728 if (!ret && write) {
2729 sched_rr_timeslice =
2730 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2731 msecs_to_jiffies(sysctl_sched_rr_timeslice);
2733 mutex_unlock(&mutex);
2738 #ifdef CONFIG_SCHED_DEBUG
2739 void print_rt_stats(struct seq_file *m, int cpu)
2742 struct rt_rq *rt_rq;
2745 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2746 print_rt_rq(m, cpu, rt_rq);
2749 #endif /* CONFIG_SCHED_DEBUG */