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 * RR_TIMESLICE) / HZ;
12 /* More than 4 hours if BW_SHIFT equals 20. */
13 static const u64 max_rt_runtime = MAX_BW;
15 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
17 struct rt_bandwidth def_rt_bandwidth;
19 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
21 struct rt_bandwidth *rt_b =
22 container_of(timer, struct rt_bandwidth, rt_period_timer);
26 raw_spin_lock(&rt_b->rt_runtime_lock);
28 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
32 raw_spin_unlock(&rt_b->rt_runtime_lock);
33 idle = do_sched_rt_period_timer(rt_b, overrun);
34 raw_spin_lock(&rt_b->rt_runtime_lock);
37 rt_b->rt_period_active = 0;
38 raw_spin_unlock(&rt_b->rt_runtime_lock);
40 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
43 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
45 rt_b->rt_period = ns_to_ktime(period);
46 rt_b->rt_runtime = runtime;
48 raw_spin_lock_init(&rt_b->rt_runtime_lock);
50 hrtimer_init(&rt_b->rt_period_timer, CLOCK_MONOTONIC,
51 HRTIMER_MODE_REL_HARD);
52 rt_b->rt_period_timer.function = sched_rt_period_timer;
55 static inline void do_start_rt_bandwidth(struct rt_bandwidth *rt_b)
57 raw_spin_lock(&rt_b->rt_runtime_lock);
58 if (!rt_b->rt_period_active) {
59 rt_b->rt_period_active = 1;
61 * SCHED_DEADLINE updates the bandwidth, as a run away
62 * RT task with a DL task could hog a CPU. But DL does
63 * not reset the period. If a deadline task was running
64 * without an RT task running, it can cause RT tasks to
65 * throttle when they start up. Kick the timer right away
66 * to update the period.
68 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
69 hrtimer_start_expires(&rt_b->rt_period_timer,
70 HRTIMER_MODE_ABS_PINNED_HARD);
72 raw_spin_unlock(&rt_b->rt_runtime_lock);
75 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
77 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
80 do_start_rt_bandwidth(rt_b);
83 void init_rt_rq(struct rt_rq *rt_rq)
85 struct rt_prio_array *array;
88 array = &rt_rq->active;
89 for (i = 0; i < MAX_RT_PRIO; i++) {
90 INIT_LIST_HEAD(array->queue + i);
91 __clear_bit(i, array->bitmap);
93 /* delimiter for bitsearch: */
94 __set_bit(MAX_RT_PRIO, array->bitmap);
96 #if defined CONFIG_SMP
97 rt_rq->highest_prio.curr = MAX_RT_PRIO;
98 rt_rq->highest_prio.next = MAX_RT_PRIO;
99 rt_rq->rt_nr_migratory = 0;
100 rt_rq->overloaded = 0;
101 plist_head_init(&rt_rq->pushable_tasks);
102 #endif /* CONFIG_SMP */
103 /* We start is dequeued state, because no RT tasks are queued */
104 rt_rq->rt_queued = 0;
107 rt_rq->rt_throttled = 0;
108 rt_rq->rt_runtime = 0;
109 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
112 #ifdef CONFIG_RT_GROUP_SCHED
113 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
115 hrtimer_cancel(&rt_b->rt_period_timer);
118 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
120 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
122 #ifdef CONFIG_SCHED_DEBUG
123 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
125 return container_of(rt_se, struct task_struct, rt);
128 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
133 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
138 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
140 struct rt_rq *rt_rq = rt_se->rt_rq;
145 void free_rt_sched_group(struct task_group *tg)
150 destroy_rt_bandwidth(&tg->rt_bandwidth);
152 for_each_possible_cpu(i) {
163 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
164 struct sched_rt_entity *rt_se, int cpu,
165 struct sched_rt_entity *parent)
167 struct rq *rq = cpu_rq(cpu);
169 rt_rq->highest_prio.curr = MAX_RT_PRIO;
170 rt_rq->rt_nr_boosted = 0;
174 tg->rt_rq[cpu] = rt_rq;
175 tg->rt_se[cpu] = rt_se;
181 rt_se->rt_rq = &rq->rt;
183 rt_se->rt_rq = parent->my_q;
186 rt_se->parent = parent;
187 INIT_LIST_HEAD(&rt_se->run_list);
190 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
193 struct sched_rt_entity *rt_se;
196 tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL);
199 tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL);
203 init_rt_bandwidth(&tg->rt_bandwidth,
204 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
206 for_each_possible_cpu(i) {
207 rt_rq = kzalloc_node(sizeof(struct rt_rq),
208 GFP_KERNEL, cpu_to_node(i));
212 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
213 GFP_KERNEL, cpu_to_node(i));
218 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
219 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
230 #else /* CONFIG_RT_GROUP_SCHED */
232 #define rt_entity_is_task(rt_se) (1)
234 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
236 return container_of(rt_se, struct task_struct, rt);
239 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
241 return container_of(rt_rq, struct rq, rt);
244 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
246 struct task_struct *p = rt_task_of(rt_se);
251 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
253 struct rq *rq = rq_of_rt_se(rt_se);
258 void free_rt_sched_group(struct task_group *tg) { }
260 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
264 #endif /* CONFIG_RT_GROUP_SCHED */
268 static void pull_rt_task(struct rq *this_rq);
270 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
272 /* Try to pull RT tasks here if we lower this rq's prio */
273 return rq->rt.highest_prio.curr > prev->prio;
276 static inline int rt_overloaded(struct rq *rq)
278 return atomic_read(&rq->rd->rto_count);
281 static inline void rt_set_overload(struct rq *rq)
286 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
288 * Make sure the mask is visible before we set
289 * the overload count. That is checked to determine
290 * if we should look at the mask. It would be a shame
291 * if we looked at the mask, but the mask was not
294 * Matched by the barrier in pull_rt_task().
297 atomic_inc(&rq->rd->rto_count);
300 static inline void rt_clear_overload(struct rq *rq)
305 /* the order here really doesn't matter */
306 atomic_dec(&rq->rd->rto_count);
307 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
310 static void update_rt_migration(struct rt_rq *rt_rq)
312 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
313 if (!rt_rq->overloaded) {
314 rt_set_overload(rq_of_rt_rq(rt_rq));
315 rt_rq->overloaded = 1;
317 } else if (rt_rq->overloaded) {
318 rt_clear_overload(rq_of_rt_rq(rt_rq));
319 rt_rq->overloaded = 0;
323 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
325 struct task_struct *p;
327 if (!rt_entity_is_task(rt_se))
330 p = rt_task_of(rt_se);
331 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
333 rt_rq->rt_nr_total++;
334 if (p->nr_cpus_allowed > 1)
335 rt_rq->rt_nr_migratory++;
337 update_rt_migration(rt_rq);
340 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
342 struct task_struct *p;
344 if (!rt_entity_is_task(rt_se))
347 p = rt_task_of(rt_se);
348 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
350 rt_rq->rt_nr_total--;
351 if (p->nr_cpus_allowed > 1)
352 rt_rq->rt_nr_migratory--;
354 update_rt_migration(rt_rq);
357 static inline int has_pushable_tasks(struct rq *rq)
359 return !plist_head_empty(&rq->rt.pushable_tasks);
362 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
363 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
365 static void push_rt_tasks(struct rq *);
366 static void pull_rt_task(struct rq *);
368 static inline void rt_queue_push_tasks(struct rq *rq)
370 if (!has_pushable_tasks(rq))
373 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
376 static inline void rt_queue_pull_task(struct rq *rq)
378 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
381 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
383 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
384 plist_node_init(&p->pushable_tasks, p->prio);
385 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
387 /* Update the highest prio pushable task */
388 if (p->prio < rq->rt.highest_prio.next)
389 rq->rt.highest_prio.next = p->prio;
392 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
394 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
396 /* Update the new highest prio pushable task */
397 if (has_pushable_tasks(rq)) {
398 p = plist_first_entry(&rq->rt.pushable_tasks,
399 struct task_struct, pushable_tasks);
400 rq->rt.highest_prio.next = p->prio;
402 rq->rt.highest_prio.next = MAX_RT_PRIO;
407 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
411 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
416 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
421 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
425 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
430 static inline void pull_rt_task(struct rq *this_rq)
434 static inline void rt_queue_push_tasks(struct rq *rq)
437 #endif /* CONFIG_SMP */
439 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
440 static void dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count);
442 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
447 #ifdef CONFIG_UCLAMP_TASK
449 * Verify the fitness of task @p to run on @cpu taking into account the uclamp
452 * This check is only important for heterogeneous systems where uclamp_min value
453 * is higher than the capacity of a @cpu. For non-heterogeneous system this
454 * function will always return true.
456 * The function will return true if the capacity of the @cpu is >= the
457 * uclamp_min and false otherwise.
459 * Note that uclamp_min will be clamped to uclamp_max if uclamp_min
462 static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
464 unsigned int min_cap;
465 unsigned int max_cap;
466 unsigned int cpu_cap;
468 /* Only heterogeneous systems can benefit from this check */
469 if (!static_branch_unlikely(&sched_asym_cpucapacity))
472 min_cap = uclamp_eff_value(p, UCLAMP_MIN);
473 max_cap = uclamp_eff_value(p, UCLAMP_MAX);
475 cpu_cap = capacity_orig_of(cpu);
477 return cpu_cap >= min(min_cap, max_cap);
480 static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
486 #ifdef CONFIG_RT_GROUP_SCHED
488 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
493 return rt_rq->rt_runtime;
496 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
498 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
501 typedef struct task_group *rt_rq_iter_t;
503 static inline struct task_group *next_task_group(struct task_group *tg)
506 tg = list_entry_rcu(tg->list.next,
507 typeof(struct task_group), list);
508 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
510 if (&tg->list == &task_groups)
516 #define for_each_rt_rq(rt_rq, iter, rq) \
517 for (iter = container_of(&task_groups, typeof(*iter), list); \
518 (iter = next_task_group(iter)) && \
519 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
521 #define for_each_sched_rt_entity(rt_se) \
522 for (; rt_se; rt_se = rt_se->parent)
524 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
529 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
530 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
532 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
534 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
535 struct rq *rq = rq_of_rt_rq(rt_rq);
536 struct sched_rt_entity *rt_se;
538 int cpu = cpu_of(rq);
540 rt_se = rt_rq->tg->rt_se[cpu];
542 if (rt_rq->rt_nr_running) {
544 enqueue_top_rt_rq(rt_rq);
545 else if (!on_rt_rq(rt_se))
546 enqueue_rt_entity(rt_se, 0);
548 if (rt_rq->highest_prio.curr < curr->prio)
553 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
555 struct sched_rt_entity *rt_se;
556 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
558 rt_se = rt_rq->tg->rt_se[cpu];
561 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
562 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
563 cpufreq_update_util(rq_of_rt_rq(rt_rq), 0);
565 else if (on_rt_rq(rt_se))
566 dequeue_rt_entity(rt_se, 0);
569 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
571 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
574 static int rt_se_boosted(struct sched_rt_entity *rt_se)
576 struct rt_rq *rt_rq = group_rt_rq(rt_se);
577 struct task_struct *p;
580 return !!rt_rq->rt_nr_boosted;
582 p = rt_task_of(rt_se);
583 return p->prio != p->normal_prio;
587 static inline const struct cpumask *sched_rt_period_mask(void)
589 return this_rq()->rd->span;
592 static inline const struct cpumask *sched_rt_period_mask(void)
594 return cpu_online_mask;
599 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
601 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
604 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
606 return &rt_rq->tg->rt_bandwidth;
609 #else /* !CONFIG_RT_GROUP_SCHED */
611 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
613 return rt_rq->rt_runtime;
616 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
618 return ktime_to_ns(def_rt_bandwidth.rt_period);
621 typedef struct rt_rq *rt_rq_iter_t;
623 #define for_each_rt_rq(rt_rq, iter, rq) \
624 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
626 #define for_each_sched_rt_entity(rt_se) \
627 for (; rt_se; rt_se = NULL)
629 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
634 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
636 struct rq *rq = rq_of_rt_rq(rt_rq);
638 if (!rt_rq->rt_nr_running)
641 enqueue_top_rt_rq(rt_rq);
645 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
647 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
650 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
652 return rt_rq->rt_throttled;
655 static inline const struct cpumask *sched_rt_period_mask(void)
657 return cpu_online_mask;
661 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
663 return &cpu_rq(cpu)->rt;
666 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
668 return &def_rt_bandwidth;
671 #endif /* CONFIG_RT_GROUP_SCHED */
673 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
675 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
677 return (hrtimer_active(&rt_b->rt_period_timer) ||
678 rt_rq->rt_time < rt_b->rt_runtime);
683 * We ran out of runtime, see if we can borrow some from our neighbours.
685 static void do_balance_runtime(struct rt_rq *rt_rq)
687 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
688 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
692 weight = cpumask_weight(rd->span);
694 raw_spin_lock(&rt_b->rt_runtime_lock);
695 rt_period = ktime_to_ns(rt_b->rt_period);
696 for_each_cpu(i, rd->span) {
697 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
703 raw_spin_lock(&iter->rt_runtime_lock);
705 * Either all rqs have inf runtime and there's nothing to steal
706 * or __disable_runtime() below sets a specific rq to inf to
707 * indicate its been disabled and disalow stealing.
709 if (iter->rt_runtime == RUNTIME_INF)
713 * From runqueues with spare time, take 1/n part of their
714 * spare time, but no more than our period.
716 diff = iter->rt_runtime - iter->rt_time;
718 diff = div_u64((u64)diff, weight);
719 if (rt_rq->rt_runtime + diff > rt_period)
720 diff = rt_period - rt_rq->rt_runtime;
721 iter->rt_runtime -= diff;
722 rt_rq->rt_runtime += diff;
723 if (rt_rq->rt_runtime == rt_period) {
724 raw_spin_unlock(&iter->rt_runtime_lock);
729 raw_spin_unlock(&iter->rt_runtime_lock);
731 raw_spin_unlock(&rt_b->rt_runtime_lock);
735 * Ensure this RQ takes back all the runtime it lend to its neighbours.
737 static void __disable_runtime(struct rq *rq)
739 struct root_domain *rd = rq->rd;
743 if (unlikely(!scheduler_running))
746 for_each_rt_rq(rt_rq, iter, rq) {
747 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
751 raw_spin_lock(&rt_b->rt_runtime_lock);
752 raw_spin_lock(&rt_rq->rt_runtime_lock);
754 * Either we're all inf and nobody needs to borrow, or we're
755 * already disabled and thus have nothing to do, or we have
756 * exactly the right amount of runtime to take out.
758 if (rt_rq->rt_runtime == RUNTIME_INF ||
759 rt_rq->rt_runtime == rt_b->rt_runtime)
761 raw_spin_unlock(&rt_rq->rt_runtime_lock);
764 * Calculate the difference between what we started out with
765 * and what we current have, that's the amount of runtime
766 * we lend and now have to reclaim.
768 want = rt_b->rt_runtime - rt_rq->rt_runtime;
771 * Greedy reclaim, take back as much as we can.
773 for_each_cpu(i, rd->span) {
774 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
778 * Can't reclaim from ourselves or disabled runqueues.
780 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
783 raw_spin_lock(&iter->rt_runtime_lock);
785 diff = min_t(s64, iter->rt_runtime, want);
786 iter->rt_runtime -= diff;
789 iter->rt_runtime -= want;
792 raw_spin_unlock(&iter->rt_runtime_lock);
798 raw_spin_lock(&rt_rq->rt_runtime_lock);
800 * We cannot be left wanting - that would mean some runtime
801 * leaked out of the system.
806 * Disable all the borrow logic by pretending we have inf
807 * runtime - in which case borrowing doesn't make sense.
809 rt_rq->rt_runtime = RUNTIME_INF;
810 rt_rq->rt_throttled = 0;
811 raw_spin_unlock(&rt_rq->rt_runtime_lock);
812 raw_spin_unlock(&rt_b->rt_runtime_lock);
814 /* Make rt_rq available for pick_next_task() */
815 sched_rt_rq_enqueue(rt_rq);
819 static void __enable_runtime(struct rq *rq)
824 if (unlikely(!scheduler_running))
828 * Reset each runqueue's bandwidth settings
830 for_each_rt_rq(rt_rq, iter, rq) {
831 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
833 raw_spin_lock(&rt_b->rt_runtime_lock);
834 raw_spin_lock(&rt_rq->rt_runtime_lock);
835 rt_rq->rt_runtime = rt_b->rt_runtime;
837 rt_rq->rt_throttled = 0;
838 raw_spin_unlock(&rt_rq->rt_runtime_lock);
839 raw_spin_unlock(&rt_b->rt_runtime_lock);
843 static void balance_runtime(struct rt_rq *rt_rq)
845 if (!sched_feat(RT_RUNTIME_SHARE))
848 if (rt_rq->rt_time > rt_rq->rt_runtime) {
849 raw_spin_unlock(&rt_rq->rt_runtime_lock);
850 do_balance_runtime(rt_rq);
851 raw_spin_lock(&rt_rq->rt_runtime_lock);
854 #else /* !CONFIG_SMP */
855 static inline void balance_runtime(struct rt_rq *rt_rq) {}
856 #endif /* CONFIG_SMP */
858 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
860 int i, idle = 1, throttled = 0;
861 const struct cpumask *span;
863 span = sched_rt_period_mask();
864 #ifdef CONFIG_RT_GROUP_SCHED
866 * FIXME: isolated CPUs should really leave the root task group,
867 * whether they are isolcpus or were isolated via cpusets, lest
868 * the timer run on a CPU which does not service all runqueues,
869 * potentially leaving other CPUs indefinitely throttled. If
870 * isolation is really required, the user will turn the throttle
871 * off to kill the perturbations it causes anyway. Meanwhile,
872 * this maintains functionality for boot and/or troubleshooting.
874 if (rt_b == &root_task_group.rt_bandwidth)
875 span = cpu_online_mask;
877 for_each_cpu(i, span) {
879 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
880 struct rq *rq = rq_of_rt_rq(rt_rq);
884 * When span == cpu_online_mask, taking each rq->lock
885 * can be time-consuming. Try to avoid it when possible.
887 raw_spin_lock(&rt_rq->rt_runtime_lock);
888 if (!sched_feat(RT_RUNTIME_SHARE) && rt_rq->rt_runtime != RUNTIME_INF)
889 rt_rq->rt_runtime = rt_b->rt_runtime;
890 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
891 raw_spin_unlock(&rt_rq->rt_runtime_lock);
895 raw_spin_lock(&rq->lock);
898 if (rt_rq->rt_time) {
901 raw_spin_lock(&rt_rq->rt_runtime_lock);
902 if (rt_rq->rt_throttled)
903 balance_runtime(rt_rq);
904 runtime = rt_rq->rt_runtime;
905 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
906 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
907 rt_rq->rt_throttled = 0;
911 * When we're idle and a woken (rt) task is
912 * throttled check_preempt_curr() will set
913 * skip_update and the time between the wakeup
914 * and this unthrottle will get accounted as
917 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
918 rq_clock_cancel_skipupdate(rq);
920 if (rt_rq->rt_time || rt_rq->rt_nr_running)
922 raw_spin_unlock(&rt_rq->rt_runtime_lock);
923 } else if (rt_rq->rt_nr_running) {
925 if (!rt_rq_throttled(rt_rq))
928 if (rt_rq->rt_throttled)
932 sched_rt_rq_enqueue(rt_rq);
933 raw_spin_unlock(&rq->lock);
936 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
942 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
944 #ifdef CONFIG_RT_GROUP_SCHED
945 struct rt_rq *rt_rq = group_rt_rq(rt_se);
948 return rt_rq->highest_prio.curr;
951 return rt_task_of(rt_se)->prio;
954 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
956 u64 runtime = sched_rt_runtime(rt_rq);
958 if (rt_rq->rt_throttled)
959 return rt_rq_throttled(rt_rq);
961 if (runtime >= sched_rt_period(rt_rq))
964 balance_runtime(rt_rq);
965 runtime = sched_rt_runtime(rt_rq);
966 if (runtime == RUNTIME_INF)
969 if (rt_rq->rt_time > runtime) {
970 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
973 * Don't actually throttle groups that have no runtime assigned
974 * but accrue some time due to boosting.
976 if (likely(rt_b->rt_runtime)) {
977 rt_rq->rt_throttled = 1;
978 printk_deferred_once("sched: RT throttling activated\n");
981 * In case we did anyway, make it go away,
982 * replenishment is a joke, since it will replenish us
988 if (rt_rq_throttled(rt_rq)) {
989 sched_rt_rq_dequeue(rt_rq);
998 * Update the current task's runtime statistics. Skip current tasks that
999 * are not in our scheduling class.
1001 static void update_curr_rt(struct rq *rq)
1003 struct task_struct *curr = rq->curr;
1004 struct sched_rt_entity *rt_se = &curr->rt;
1008 if (curr->sched_class != &rt_sched_class)
1011 now = rq_clock_task(rq);
1012 delta_exec = now - curr->se.exec_start;
1013 if (unlikely((s64)delta_exec <= 0))
1016 schedstat_set(curr->se.statistics.exec_max,
1017 max(curr->se.statistics.exec_max, delta_exec));
1019 curr->se.sum_exec_runtime += delta_exec;
1020 account_group_exec_runtime(curr, delta_exec);
1022 curr->se.exec_start = now;
1023 cgroup_account_cputime(curr, delta_exec);
1025 if (!rt_bandwidth_enabled())
1028 for_each_sched_rt_entity(rt_se) {
1029 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1032 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
1033 raw_spin_lock(&rt_rq->rt_runtime_lock);
1034 rt_rq->rt_time += delta_exec;
1035 exceeded = sched_rt_runtime_exceeded(rt_rq);
1038 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1040 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq));
1046 dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count)
1048 struct rq *rq = rq_of_rt_rq(rt_rq);
1050 BUG_ON(&rq->rt != rt_rq);
1052 if (!rt_rq->rt_queued)
1055 BUG_ON(!rq->nr_running);
1057 sub_nr_running(rq, count);
1058 rt_rq->rt_queued = 0;
1063 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1065 struct rq *rq = rq_of_rt_rq(rt_rq);
1067 BUG_ON(&rq->rt != rt_rq);
1069 if (rt_rq->rt_queued)
1072 if (rt_rq_throttled(rt_rq))
1075 if (rt_rq->rt_nr_running) {
1076 add_nr_running(rq, rt_rq->rt_nr_running);
1077 rt_rq->rt_queued = 1;
1080 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1081 cpufreq_update_util(rq, 0);
1084 #if defined CONFIG_SMP
1087 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1089 struct rq *rq = rq_of_rt_rq(rt_rq);
1091 #ifdef CONFIG_RT_GROUP_SCHED
1093 * Change rq's cpupri only if rt_rq is the top queue.
1095 if (&rq->rt != rt_rq)
1098 if (rq->online && prio < prev_prio)
1099 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1103 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1105 struct rq *rq = rq_of_rt_rq(rt_rq);
1107 #ifdef CONFIG_RT_GROUP_SCHED
1109 * Change rq's cpupri only if rt_rq is the top queue.
1111 if (&rq->rt != rt_rq)
1114 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1115 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1118 #else /* CONFIG_SMP */
1121 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1123 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1125 #endif /* CONFIG_SMP */
1127 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1129 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1131 int prev_prio = rt_rq->highest_prio.curr;
1133 if (prio < prev_prio)
1134 rt_rq->highest_prio.curr = prio;
1136 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1140 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1142 int prev_prio = rt_rq->highest_prio.curr;
1144 if (rt_rq->rt_nr_running) {
1146 WARN_ON(prio < prev_prio);
1149 * This may have been our highest task, and therefore
1150 * we may have some recomputation to do
1152 if (prio == prev_prio) {
1153 struct rt_prio_array *array = &rt_rq->active;
1155 rt_rq->highest_prio.curr =
1156 sched_find_first_bit(array->bitmap);
1160 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1162 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1167 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1168 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1170 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1172 #ifdef CONFIG_RT_GROUP_SCHED
1175 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1177 if (rt_se_boosted(rt_se))
1178 rt_rq->rt_nr_boosted++;
1181 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1185 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1187 if (rt_se_boosted(rt_se))
1188 rt_rq->rt_nr_boosted--;
1190 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1193 #else /* CONFIG_RT_GROUP_SCHED */
1196 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1198 start_rt_bandwidth(&def_rt_bandwidth);
1202 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1204 #endif /* CONFIG_RT_GROUP_SCHED */
1207 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1209 struct rt_rq *group_rq = group_rt_rq(rt_se);
1212 return group_rq->rt_nr_running;
1218 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1220 struct rt_rq *group_rq = group_rt_rq(rt_se);
1221 struct task_struct *tsk;
1224 return group_rq->rr_nr_running;
1226 tsk = rt_task_of(rt_se);
1228 return (tsk->policy == SCHED_RR) ? 1 : 0;
1232 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1234 int prio = rt_se_prio(rt_se);
1236 WARN_ON(!rt_prio(prio));
1237 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1238 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1240 inc_rt_prio(rt_rq, prio);
1241 inc_rt_migration(rt_se, rt_rq);
1242 inc_rt_group(rt_se, rt_rq);
1246 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1248 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1249 WARN_ON(!rt_rq->rt_nr_running);
1250 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1251 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1253 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1254 dec_rt_migration(rt_se, rt_rq);
1255 dec_rt_group(rt_se, rt_rq);
1259 * Change rt_se->run_list location unless SAVE && !MOVE
1261 * assumes ENQUEUE/DEQUEUE flags match
1263 static inline bool move_entity(unsigned int flags)
1265 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1271 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1273 list_del_init(&rt_se->run_list);
1275 if (list_empty(array->queue + rt_se_prio(rt_se)))
1276 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1281 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1283 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1284 struct rt_prio_array *array = &rt_rq->active;
1285 struct rt_rq *group_rq = group_rt_rq(rt_se);
1286 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1289 * Don't enqueue the group if its throttled, or when empty.
1290 * The latter is a consequence of the former when a child group
1291 * get throttled and the current group doesn't have any other
1294 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1296 __delist_rt_entity(rt_se, array);
1300 if (move_entity(flags)) {
1301 WARN_ON_ONCE(rt_se->on_list);
1302 if (flags & ENQUEUE_HEAD)
1303 list_add(&rt_se->run_list, queue);
1305 list_add_tail(&rt_se->run_list, queue);
1307 __set_bit(rt_se_prio(rt_se), array->bitmap);
1312 inc_rt_tasks(rt_se, rt_rq);
1315 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1317 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1318 struct rt_prio_array *array = &rt_rq->active;
1320 if (move_entity(flags)) {
1321 WARN_ON_ONCE(!rt_se->on_list);
1322 __delist_rt_entity(rt_se, array);
1326 dec_rt_tasks(rt_se, rt_rq);
1330 * Because the prio of an upper entry depends on the lower
1331 * entries, we must remove entries top - down.
1333 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1335 struct sched_rt_entity *back = NULL;
1336 unsigned int rt_nr_running;
1338 for_each_sched_rt_entity(rt_se) {
1343 rt_nr_running = rt_rq_of_se(back)->rt_nr_running;
1345 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1346 if (on_rt_rq(rt_se))
1347 __dequeue_rt_entity(rt_se, flags);
1350 dequeue_top_rt_rq(rt_rq_of_se(back), rt_nr_running);
1353 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1355 struct rq *rq = rq_of_rt_se(rt_se);
1357 dequeue_rt_stack(rt_se, flags);
1358 for_each_sched_rt_entity(rt_se)
1359 __enqueue_rt_entity(rt_se, flags);
1360 enqueue_top_rt_rq(&rq->rt);
1363 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1365 struct rq *rq = rq_of_rt_se(rt_se);
1367 dequeue_rt_stack(rt_se, flags);
1369 for_each_sched_rt_entity(rt_se) {
1370 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1372 if (rt_rq && rt_rq->rt_nr_running)
1373 __enqueue_rt_entity(rt_se, flags);
1375 enqueue_top_rt_rq(&rq->rt);
1379 * Adding/removing a task to/from a priority array:
1382 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1384 struct sched_rt_entity *rt_se = &p->rt;
1386 if (flags & ENQUEUE_WAKEUP)
1389 enqueue_rt_entity(rt_se, flags);
1391 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1392 enqueue_pushable_task(rq, p);
1395 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1397 struct sched_rt_entity *rt_se = &p->rt;
1400 dequeue_rt_entity(rt_se, flags);
1402 dequeue_pushable_task(rq, p);
1406 * Put task to the head or the end of the run list without the overhead of
1407 * dequeue followed by enqueue.
1410 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1412 if (on_rt_rq(rt_se)) {
1413 struct rt_prio_array *array = &rt_rq->active;
1414 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1417 list_move(&rt_se->run_list, queue);
1419 list_move_tail(&rt_se->run_list, queue);
1423 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1425 struct sched_rt_entity *rt_se = &p->rt;
1426 struct rt_rq *rt_rq;
1428 for_each_sched_rt_entity(rt_se) {
1429 rt_rq = rt_rq_of_se(rt_se);
1430 requeue_rt_entity(rt_rq, rt_se, head);
1434 static void yield_task_rt(struct rq *rq)
1436 requeue_task_rt(rq, rq->curr, 0);
1440 static int find_lowest_rq(struct task_struct *task);
1443 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1445 struct task_struct *curr;
1449 /* For anything but wake ups, just return the task_cpu */
1450 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1456 curr = READ_ONCE(rq->curr); /* unlocked access */
1459 * If the current task on @p's runqueue is an RT task, then
1460 * try to see if we can wake this RT task up on another
1461 * runqueue. Otherwise simply start this RT task
1462 * on its current runqueue.
1464 * We want to avoid overloading runqueues. If the woken
1465 * task is a higher priority, then it will stay on this CPU
1466 * and the lower prio task should be moved to another CPU.
1467 * Even though this will probably make the lower prio task
1468 * lose its cache, we do not want to bounce a higher task
1469 * around just because it gave up its CPU, perhaps for a
1472 * For equal prio tasks, we just let the scheduler sort it out.
1474 * Otherwise, just let it ride on the affined RQ and the
1475 * post-schedule router will push the preempted task away
1477 * This test is optimistic, if we get it wrong the load-balancer
1478 * will have to sort it out.
1480 * We take into account the capacity of the CPU to ensure it fits the
1481 * requirement of the task - which is only important on heterogeneous
1482 * systems like big.LITTLE.
1485 unlikely(rt_task(curr)) &&
1486 (curr->nr_cpus_allowed < 2 || curr->prio <= p->prio);
1488 if (test || !rt_task_fits_capacity(p, cpu)) {
1489 int target = find_lowest_rq(p);
1492 * Bail out if we were forcing a migration to find a better
1493 * fitting CPU but our search failed.
1495 if (!test && target != -1 && !rt_task_fits_capacity(p, target))
1499 * Don't bother moving it if the destination CPU is
1500 * not running a lower priority task.
1503 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1514 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1517 * Current can't be migrated, useless to reschedule,
1518 * let's hope p can move out.
1520 if (rq->curr->nr_cpus_allowed == 1 ||
1521 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1525 * p is migratable, so let's not schedule it and
1526 * see if it is pushed or pulled somewhere else.
1528 if (p->nr_cpus_allowed != 1 &&
1529 cpupri_find(&rq->rd->cpupri, p, NULL))
1533 * There appear to be other CPUs that can accept
1534 * the current task but none can run 'p', so lets reschedule
1535 * to try and push the current task away:
1537 requeue_task_rt(rq, p, 1);
1541 static int balance_rt(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1543 if (!on_rt_rq(&p->rt) && need_pull_rt_task(rq, p)) {
1545 * This is OK, because current is on_cpu, which avoids it being
1546 * picked for load-balance and preemption/IRQs are still
1547 * disabled avoiding further scheduler activity on it and we've
1548 * not yet started the picking loop.
1550 rq_unpin_lock(rq, rf);
1552 rq_repin_lock(rq, rf);
1555 return sched_stop_runnable(rq) || sched_dl_runnable(rq) || sched_rt_runnable(rq);
1557 #endif /* CONFIG_SMP */
1560 * Preempt the current task with a newly woken task if needed:
1562 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1564 if (p->prio < rq->curr->prio) {
1573 * - the newly woken task is of equal priority to the current task
1574 * - the newly woken task is non-migratable while current is migratable
1575 * - current will be preempted on the next reschedule
1577 * we should check to see if current can readily move to a different
1578 * cpu. If so, we will reschedule to allow the push logic to try
1579 * to move current somewhere else, making room for our non-migratable
1582 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1583 check_preempt_equal_prio(rq, p);
1587 static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool first)
1589 p->se.exec_start = rq_clock_task(rq);
1591 /* The running task is never eligible for pushing */
1592 dequeue_pushable_task(rq, p);
1598 * If prev task was rt, put_prev_task() has already updated the
1599 * utilization. We only care of the case where we start to schedule a
1602 if (rq->curr->sched_class != &rt_sched_class)
1603 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1605 rt_queue_push_tasks(rq);
1608 static struct sched_rt_entity *pick_next_rt_entity(struct rt_rq *rt_rq)
1610 struct rt_prio_array *array = &rt_rq->active;
1611 struct sched_rt_entity *next = NULL;
1612 struct list_head *queue;
1615 idx = sched_find_first_bit(array->bitmap);
1616 BUG_ON(idx >= MAX_RT_PRIO);
1618 queue = array->queue + idx;
1619 if (SCHED_WARN_ON(list_empty(queue)))
1621 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1626 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1628 struct sched_rt_entity *rt_se;
1629 struct rt_rq *rt_rq = &rq->rt;
1632 rt_se = pick_next_rt_entity(rt_rq);
1633 if (unlikely(!rt_se))
1635 rt_rq = group_rt_rq(rt_se);
1638 return rt_task_of(rt_se);
1641 static struct task_struct *pick_next_task_rt(struct rq *rq)
1643 struct task_struct *p;
1645 if (!sched_rt_runnable(rq))
1648 p = _pick_next_task_rt(rq);
1649 set_next_task_rt(rq, p, true);
1653 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1657 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1660 * The previous task needs to be made eligible for pushing
1661 * if it is still active
1663 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1664 enqueue_pushable_task(rq, p);
1669 /* Only try algorithms three times */
1670 #define RT_MAX_TRIES 3
1672 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1674 if (!task_running(rq, p) &&
1675 cpumask_test_cpu(cpu, p->cpus_ptr))
1682 * Return the highest pushable rq's task, which is suitable to be executed
1683 * on the CPU, NULL otherwise
1685 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1687 struct plist_head *head = &rq->rt.pushable_tasks;
1688 struct task_struct *p;
1690 if (!has_pushable_tasks(rq))
1693 plist_for_each_entry(p, head, pushable_tasks) {
1694 if (pick_rt_task(rq, p, cpu))
1701 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1703 static int find_lowest_rq(struct task_struct *task)
1705 struct sched_domain *sd;
1706 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1707 int this_cpu = smp_processor_id();
1708 int cpu = task_cpu(task);
1711 /* Make sure the mask is initialized first */
1712 if (unlikely(!lowest_mask))
1715 if (task->nr_cpus_allowed == 1)
1716 return -1; /* No other targets possible */
1719 * If we're on asym system ensure we consider the different capacities
1720 * of the CPUs when searching for the lowest_mask.
1722 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
1724 ret = cpupri_find_fitness(&task_rq(task)->rd->cpupri,
1726 rt_task_fits_capacity);
1729 ret = cpupri_find(&task_rq(task)->rd->cpupri,
1734 return -1; /* No targets found */
1737 * At this point we have built a mask of CPUs representing the
1738 * lowest priority tasks in the system. Now we want to elect
1739 * the best one based on our affinity and topology.
1741 * We prioritize the last CPU that the task executed on since
1742 * it is most likely cache-hot in that location.
1744 if (cpumask_test_cpu(cpu, lowest_mask))
1748 * Otherwise, we consult the sched_domains span maps to figure
1749 * out which CPU is logically closest to our hot cache data.
1751 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1752 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1755 for_each_domain(cpu, sd) {
1756 if (sd->flags & SD_WAKE_AFFINE) {
1760 * "this_cpu" is cheaper to preempt than a
1763 if (this_cpu != -1 &&
1764 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1769 best_cpu = cpumask_first_and(lowest_mask,
1770 sched_domain_span(sd));
1771 if (best_cpu < nr_cpu_ids) {
1780 * And finally, if there were no matches within the domains
1781 * just give the caller *something* to work with from the compatible
1787 cpu = cpumask_any(lowest_mask);
1788 if (cpu < nr_cpu_ids)
1794 /* Will lock the rq it finds */
1795 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1797 struct rq *lowest_rq = NULL;
1801 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1802 cpu = find_lowest_rq(task);
1804 if ((cpu == -1) || (cpu == rq->cpu))
1807 lowest_rq = cpu_rq(cpu);
1809 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1811 * Target rq has tasks of equal or higher priority,
1812 * retrying does not release any lock and is unlikely
1813 * to yield a different result.
1819 /* if the prio of this runqueue changed, try again */
1820 if (double_lock_balance(rq, lowest_rq)) {
1822 * We had to unlock the run queue. In
1823 * the mean time, task could have
1824 * migrated already or had its affinity changed.
1825 * Also make sure that it wasn't scheduled on its rq.
1827 if (unlikely(task_rq(task) != rq ||
1828 !cpumask_test_cpu(lowest_rq->cpu, task->cpus_ptr) ||
1829 task_running(rq, task) ||
1831 !task_on_rq_queued(task))) {
1833 double_unlock_balance(rq, lowest_rq);
1839 /* If this rq is still suitable use it. */
1840 if (lowest_rq->rt.highest_prio.curr > task->prio)
1844 double_unlock_balance(rq, lowest_rq);
1851 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1853 struct task_struct *p;
1855 if (!has_pushable_tasks(rq))
1858 p = plist_first_entry(&rq->rt.pushable_tasks,
1859 struct task_struct, pushable_tasks);
1861 BUG_ON(rq->cpu != task_cpu(p));
1862 BUG_ON(task_current(rq, p));
1863 BUG_ON(p->nr_cpus_allowed <= 1);
1865 BUG_ON(!task_on_rq_queued(p));
1866 BUG_ON(!rt_task(p));
1872 * If the current CPU has more than one RT task, see if the non
1873 * running task can migrate over to a CPU that is running a task
1874 * of lesser priority.
1876 static int push_rt_task(struct rq *rq)
1878 struct task_struct *next_task;
1879 struct rq *lowest_rq;
1882 if (!rq->rt.overloaded)
1885 next_task = pick_next_pushable_task(rq);
1890 if (WARN_ON(next_task == rq->curr))
1894 * It's possible that the next_task slipped in of
1895 * higher priority than current. If that's the case
1896 * just reschedule current.
1898 if (unlikely(next_task->prio < rq->curr->prio)) {
1903 /* We might release rq lock */
1904 get_task_struct(next_task);
1906 /* find_lock_lowest_rq locks the rq if found */
1907 lowest_rq = find_lock_lowest_rq(next_task, rq);
1909 struct task_struct *task;
1911 * find_lock_lowest_rq releases rq->lock
1912 * so it is possible that next_task has migrated.
1914 * We need to make sure that the task is still on the same
1915 * run-queue and is also still the next task eligible for
1918 task = pick_next_pushable_task(rq);
1919 if (task == next_task) {
1921 * The task hasn't migrated, and is still the next
1922 * eligible task, but we failed to find a run-queue
1923 * to push it to. Do not retry in this case, since
1924 * other CPUs will pull from us when ready.
1930 /* No more tasks, just exit */
1934 * Something has shifted, try again.
1936 put_task_struct(next_task);
1941 deactivate_task(rq, next_task, 0);
1942 set_task_cpu(next_task, lowest_rq->cpu);
1943 activate_task(lowest_rq, next_task, 0);
1946 resched_curr(lowest_rq);
1948 double_unlock_balance(rq, lowest_rq);
1951 put_task_struct(next_task);
1956 static void push_rt_tasks(struct rq *rq)
1958 /* push_rt_task will return true if it moved an RT */
1959 while (push_rt_task(rq))
1963 #ifdef HAVE_RT_PUSH_IPI
1966 * When a high priority task schedules out from a CPU and a lower priority
1967 * task is scheduled in, a check is made to see if there's any RT tasks
1968 * on other CPUs that are waiting to run because a higher priority RT task
1969 * is currently running on its CPU. In this case, the CPU with multiple RT
1970 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1971 * up that may be able to run one of its non-running queued RT tasks.
1973 * All CPUs with overloaded RT tasks need to be notified as there is currently
1974 * no way to know which of these CPUs have the highest priority task waiting
1975 * to run. Instead of trying to take a spinlock on each of these CPUs,
1976 * which has shown to cause large latency when done on machines with many
1977 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1978 * RT tasks waiting to run.
1980 * Just sending an IPI to each of the CPUs is also an issue, as on large
1981 * count CPU machines, this can cause an IPI storm on a CPU, especially
1982 * if its the only CPU with multiple RT tasks queued, and a large number
1983 * of CPUs scheduling a lower priority task at the same time.
1985 * Each root domain has its own irq work function that can iterate over
1986 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1987 * tassk must be checked if there's one or many CPUs that are lowering
1988 * their priority, there's a single irq work iterator that will try to
1989 * push off RT tasks that are waiting to run.
1991 * When a CPU schedules a lower priority task, it will kick off the
1992 * irq work iterator that will jump to each CPU with overloaded RT tasks.
1993 * As it only takes the first CPU that schedules a lower priority task
1994 * to start the process, the rto_start variable is incremented and if
1995 * the atomic result is one, then that CPU will try to take the rto_lock.
1996 * This prevents high contention on the lock as the process handles all
1997 * CPUs scheduling lower priority tasks.
1999 * All CPUs that are scheduling a lower priority task will increment the
2000 * rt_loop_next variable. This will make sure that the irq work iterator
2001 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
2002 * priority task, even if the iterator is in the middle of a scan. Incrementing
2003 * the rt_loop_next will cause the iterator to perform another scan.
2006 static int rto_next_cpu(struct root_domain *rd)
2012 * When starting the IPI RT pushing, the rto_cpu is set to -1,
2013 * rt_next_cpu() will simply return the first CPU found in
2016 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
2017 * will return the next CPU found in the rto_mask.
2019 * If there are no more CPUs left in the rto_mask, then a check is made
2020 * against rto_loop and rto_loop_next. rto_loop is only updated with
2021 * the rto_lock held, but any CPU may increment the rto_loop_next
2022 * without any locking.
2026 /* When rto_cpu is -1 this acts like cpumask_first() */
2027 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
2031 if (cpu < nr_cpu_ids)
2037 * ACQUIRE ensures we see the @rto_mask changes
2038 * made prior to the @next value observed.
2040 * Matches WMB in rt_set_overload().
2042 next = atomic_read_acquire(&rd->rto_loop_next);
2044 if (rd->rto_loop == next)
2047 rd->rto_loop = next;
2053 static inline bool rto_start_trylock(atomic_t *v)
2055 return !atomic_cmpxchg_acquire(v, 0, 1);
2058 static inline void rto_start_unlock(atomic_t *v)
2060 atomic_set_release(v, 0);
2063 static void tell_cpu_to_push(struct rq *rq)
2067 /* Keep the loop going if the IPI is currently active */
2068 atomic_inc(&rq->rd->rto_loop_next);
2070 /* Only one CPU can initiate a loop at a time */
2071 if (!rto_start_trylock(&rq->rd->rto_loop_start))
2074 raw_spin_lock(&rq->rd->rto_lock);
2077 * The rto_cpu is updated under the lock, if it has a valid CPU
2078 * then the IPI is still running and will continue due to the
2079 * update to loop_next, and nothing needs to be done here.
2080 * Otherwise it is finishing up and an ipi needs to be sent.
2082 if (rq->rd->rto_cpu < 0)
2083 cpu = rto_next_cpu(rq->rd);
2085 raw_spin_unlock(&rq->rd->rto_lock);
2087 rto_start_unlock(&rq->rd->rto_loop_start);
2090 /* Make sure the rd does not get freed while pushing */
2091 sched_get_rd(rq->rd);
2092 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2096 /* Called from hardirq context */
2097 void rto_push_irq_work_func(struct irq_work *work)
2099 struct root_domain *rd =
2100 container_of(work, struct root_domain, rto_push_work);
2107 * We do not need to grab the lock to check for has_pushable_tasks.
2108 * When it gets updated, a check is made if a push is possible.
2110 if (has_pushable_tasks(rq)) {
2111 raw_spin_lock(&rq->lock);
2113 raw_spin_unlock(&rq->lock);
2116 raw_spin_lock(&rd->rto_lock);
2118 /* Pass the IPI to the next rt overloaded queue */
2119 cpu = rto_next_cpu(rd);
2121 raw_spin_unlock(&rd->rto_lock);
2128 /* Try the next RT overloaded CPU */
2129 irq_work_queue_on(&rd->rto_push_work, cpu);
2131 #endif /* HAVE_RT_PUSH_IPI */
2133 static void pull_rt_task(struct rq *this_rq)
2135 int this_cpu = this_rq->cpu, cpu;
2136 bool resched = false;
2137 struct task_struct *p;
2139 int rt_overload_count = rt_overloaded(this_rq);
2141 if (likely(!rt_overload_count))
2145 * Match the barrier from rt_set_overloaded; this guarantees that if we
2146 * see overloaded we must also see the rto_mask bit.
2150 /* If we are the only overloaded CPU do nothing */
2151 if (rt_overload_count == 1 &&
2152 cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2155 #ifdef HAVE_RT_PUSH_IPI
2156 if (sched_feat(RT_PUSH_IPI)) {
2157 tell_cpu_to_push(this_rq);
2162 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2163 if (this_cpu == cpu)
2166 src_rq = cpu_rq(cpu);
2169 * Don't bother taking the src_rq->lock if the next highest
2170 * task is known to be lower-priority than our current task.
2171 * This may look racy, but if this value is about to go
2172 * logically higher, the src_rq will push this task away.
2173 * And if its going logically lower, we do not care
2175 if (src_rq->rt.highest_prio.next >=
2176 this_rq->rt.highest_prio.curr)
2180 * We can potentially drop this_rq's lock in
2181 * double_lock_balance, and another CPU could
2184 double_lock_balance(this_rq, src_rq);
2187 * We can pull only a task, which is pushable
2188 * on its rq, and no others.
2190 p = pick_highest_pushable_task(src_rq, this_cpu);
2193 * Do we have an RT task that preempts
2194 * the to-be-scheduled task?
2196 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2197 WARN_ON(p == src_rq->curr);
2198 WARN_ON(!task_on_rq_queued(p));
2201 * There's a chance that p is higher in priority
2202 * than what's currently running on its CPU.
2203 * This is just that p is wakeing up and hasn't
2204 * had a chance to schedule. We only pull
2205 * p if it is lower in priority than the
2206 * current task on the run queue
2208 if (p->prio < src_rq->curr->prio)
2213 deactivate_task(src_rq, p, 0);
2214 set_task_cpu(p, this_cpu);
2215 activate_task(this_rq, p, 0);
2217 * We continue with the search, just in
2218 * case there's an even higher prio task
2219 * in another runqueue. (low likelihood
2224 double_unlock_balance(this_rq, src_rq);
2228 resched_curr(this_rq);
2232 * If we are not running and we are not going to reschedule soon, we should
2233 * try to push tasks away now
2235 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2237 bool need_to_push = !task_running(rq, p) &&
2238 !test_tsk_need_resched(rq->curr) &&
2239 p->nr_cpus_allowed > 1 &&
2240 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2241 (rq->curr->nr_cpus_allowed < 2 ||
2242 rq->curr->prio <= p->prio);
2248 /* Assumes rq->lock is held */
2249 static void rq_online_rt(struct rq *rq)
2251 if (rq->rt.overloaded)
2252 rt_set_overload(rq);
2254 __enable_runtime(rq);
2256 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2259 /* Assumes rq->lock is held */
2260 static void rq_offline_rt(struct rq *rq)
2262 if (rq->rt.overloaded)
2263 rt_clear_overload(rq);
2265 __disable_runtime(rq);
2267 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2271 * When switch from the rt queue, we bring ourselves to a position
2272 * that we might want to pull RT tasks from other runqueues.
2274 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2277 * If there are other RT tasks then we will reschedule
2278 * and the scheduling of the other RT tasks will handle
2279 * the balancing. But if we are the last RT task
2280 * we may need to handle the pulling of RT tasks
2283 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2286 rt_queue_pull_task(rq);
2289 void __init init_sched_rt_class(void)
2293 for_each_possible_cpu(i) {
2294 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2295 GFP_KERNEL, cpu_to_node(i));
2298 #endif /* CONFIG_SMP */
2301 * When switching a task to RT, we may overload the runqueue
2302 * with RT tasks. In this case we try to push them off to
2305 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2308 * If we are running, update the avg_rt tracking, as the running time
2309 * will now on be accounted into the latter.
2311 if (task_current(rq, p)) {
2312 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2317 * If we are not running we may need to preempt the current
2318 * running task. If that current running task is also an RT task
2319 * then see if we can move to another run queue.
2321 if (task_on_rq_queued(p)) {
2323 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2324 rt_queue_push_tasks(rq);
2325 #endif /* CONFIG_SMP */
2326 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2332 * Priority of the task has changed. This may cause
2333 * us to initiate a push or pull.
2336 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2338 if (!task_on_rq_queued(p))
2341 if (rq->curr == p) {
2344 * If our priority decreases while running, we
2345 * may need to pull tasks to this runqueue.
2347 if (oldprio < p->prio)
2348 rt_queue_pull_task(rq);
2351 * If there's a higher priority task waiting to run
2354 if (p->prio > rq->rt.highest_prio.curr)
2357 /* For UP simply resched on drop of prio */
2358 if (oldprio < p->prio)
2360 #endif /* CONFIG_SMP */
2363 * This task is not running, but if it is
2364 * greater than the current running task
2367 if (p->prio < rq->curr->prio)
2372 #ifdef CONFIG_POSIX_TIMERS
2373 static void watchdog(struct rq *rq, struct task_struct *p)
2375 unsigned long soft, hard;
2377 /* max may change after cur was read, this will be fixed next tick */
2378 soft = task_rlimit(p, RLIMIT_RTTIME);
2379 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2381 if (soft != RLIM_INFINITY) {
2384 if (p->rt.watchdog_stamp != jiffies) {
2386 p->rt.watchdog_stamp = jiffies;
2389 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2390 if (p->rt.timeout > next) {
2391 posix_cputimers_rt_watchdog(&p->posix_cputimers,
2392 p->se.sum_exec_runtime);
2397 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2401 * scheduler tick hitting a task of our scheduling class.
2403 * NOTE: This function can be called remotely by the tick offload that
2404 * goes along full dynticks. Therefore no local assumption can be made
2405 * and everything must be accessed through the @rq and @curr passed in
2408 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2410 struct sched_rt_entity *rt_se = &p->rt;
2413 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2418 * RR tasks need a special form of timeslice management.
2419 * FIFO tasks have no timeslices.
2421 if (p->policy != SCHED_RR)
2424 if (--p->rt.time_slice)
2427 p->rt.time_slice = sched_rr_timeslice;
2430 * Requeue to the end of queue if we (and all of our ancestors) are not
2431 * the only element on the queue
2433 for_each_sched_rt_entity(rt_se) {
2434 if (rt_se->run_list.prev != rt_se->run_list.next) {
2435 requeue_task_rt(rq, p, 0);
2442 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2445 * Time slice is 0 for SCHED_FIFO tasks
2447 if (task->policy == SCHED_RR)
2448 return sched_rr_timeslice;
2453 const struct sched_class rt_sched_class
2454 __section("__rt_sched_class") = {
2455 .enqueue_task = enqueue_task_rt,
2456 .dequeue_task = dequeue_task_rt,
2457 .yield_task = yield_task_rt,
2459 .check_preempt_curr = check_preempt_curr_rt,
2461 .pick_next_task = pick_next_task_rt,
2462 .put_prev_task = put_prev_task_rt,
2463 .set_next_task = set_next_task_rt,
2466 .balance = balance_rt,
2467 .select_task_rq = select_task_rq_rt,
2468 .set_cpus_allowed = set_cpus_allowed_common,
2469 .rq_online = rq_online_rt,
2470 .rq_offline = rq_offline_rt,
2471 .task_woken = task_woken_rt,
2472 .switched_from = switched_from_rt,
2475 .task_tick = task_tick_rt,
2477 .get_rr_interval = get_rr_interval_rt,
2479 .prio_changed = prio_changed_rt,
2480 .switched_to = switched_to_rt,
2482 .update_curr = update_curr_rt,
2484 #ifdef CONFIG_UCLAMP_TASK
2485 .uclamp_enabled = 1,
2489 #ifdef CONFIG_RT_GROUP_SCHED
2491 * Ensure that the real time constraints are schedulable.
2493 static DEFINE_MUTEX(rt_constraints_mutex);
2495 static inline int tg_has_rt_tasks(struct task_group *tg)
2497 struct task_struct *task;
2498 struct css_task_iter it;
2502 * Autogroups do not have RT tasks; see autogroup_create().
2504 if (task_group_is_autogroup(tg))
2507 css_task_iter_start(&tg->css, 0, &it);
2508 while (!ret && (task = css_task_iter_next(&it)))
2509 ret |= rt_task(task);
2510 css_task_iter_end(&it);
2515 struct rt_schedulable_data {
2516 struct task_group *tg;
2521 static int tg_rt_schedulable(struct task_group *tg, void *data)
2523 struct rt_schedulable_data *d = data;
2524 struct task_group *child;
2525 unsigned long total, sum = 0;
2526 u64 period, runtime;
2528 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2529 runtime = tg->rt_bandwidth.rt_runtime;
2532 period = d->rt_period;
2533 runtime = d->rt_runtime;
2537 * Cannot have more runtime than the period.
2539 if (runtime > period && runtime != RUNTIME_INF)
2543 * Ensure we don't starve existing RT tasks if runtime turns zero.
2545 if (rt_bandwidth_enabled() && !runtime &&
2546 tg->rt_bandwidth.rt_runtime && tg_has_rt_tasks(tg))
2549 total = to_ratio(period, runtime);
2552 * Nobody can have more than the global setting allows.
2554 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2558 * The sum of our children's runtime should not exceed our own.
2560 list_for_each_entry_rcu(child, &tg->children, siblings) {
2561 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2562 runtime = child->rt_bandwidth.rt_runtime;
2564 if (child == d->tg) {
2565 period = d->rt_period;
2566 runtime = d->rt_runtime;
2569 sum += to_ratio(period, runtime);
2578 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2582 struct rt_schedulable_data data = {
2584 .rt_period = period,
2585 .rt_runtime = runtime,
2589 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2595 static int tg_set_rt_bandwidth(struct task_group *tg,
2596 u64 rt_period, u64 rt_runtime)
2601 * Disallowing the root group RT runtime is BAD, it would disallow the
2602 * kernel creating (and or operating) RT threads.
2604 if (tg == &root_task_group && rt_runtime == 0)
2607 /* No period doesn't make any sense. */
2612 * Bound quota to defend quota against overflow during bandwidth shift.
2614 if (rt_runtime != RUNTIME_INF && rt_runtime > max_rt_runtime)
2617 mutex_lock(&rt_constraints_mutex);
2618 err = __rt_schedulable(tg, rt_period, rt_runtime);
2622 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2623 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2624 tg->rt_bandwidth.rt_runtime = rt_runtime;
2626 for_each_possible_cpu(i) {
2627 struct rt_rq *rt_rq = tg->rt_rq[i];
2629 raw_spin_lock(&rt_rq->rt_runtime_lock);
2630 rt_rq->rt_runtime = rt_runtime;
2631 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2633 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2635 mutex_unlock(&rt_constraints_mutex);
2640 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2642 u64 rt_runtime, rt_period;
2644 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2645 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2646 if (rt_runtime_us < 0)
2647 rt_runtime = RUNTIME_INF;
2648 else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC)
2651 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2654 long sched_group_rt_runtime(struct task_group *tg)
2658 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2661 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2662 do_div(rt_runtime_us, NSEC_PER_USEC);
2663 return rt_runtime_us;
2666 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2668 u64 rt_runtime, rt_period;
2670 if (rt_period_us > U64_MAX / NSEC_PER_USEC)
2673 rt_period = rt_period_us * NSEC_PER_USEC;
2674 rt_runtime = tg->rt_bandwidth.rt_runtime;
2676 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2679 long sched_group_rt_period(struct task_group *tg)
2683 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2684 do_div(rt_period_us, NSEC_PER_USEC);
2685 return rt_period_us;
2688 static int sched_rt_global_constraints(void)
2692 mutex_lock(&rt_constraints_mutex);
2693 ret = __rt_schedulable(NULL, 0, 0);
2694 mutex_unlock(&rt_constraints_mutex);
2699 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2701 /* Don't accept realtime tasks when there is no way for them to run */
2702 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2708 #else /* !CONFIG_RT_GROUP_SCHED */
2709 static int sched_rt_global_constraints(void)
2711 unsigned long flags;
2714 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2715 for_each_possible_cpu(i) {
2716 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2718 raw_spin_lock(&rt_rq->rt_runtime_lock);
2719 rt_rq->rt_runtime = global_rt_runtime();
2720 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2722 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2726 #endif /* CONFIG_RT_GROUP_SCHED */
2728 static int sched_rt_global_validate(void)
2730 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2731 ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
2732 ((u64)sysctl_sched_rt_runtime *
2733 NSEC_PER_USEC > max_rt_runtime)))
2739 static void sched_rt_do_global(void)
2741 unsigned long flags;
2743 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2744 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2745 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2746 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2749 int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
2750 size_t *lenp, loff_t *ppos)
2752 int old_period, old_runtime;
2753 static DEFINE_MUTEX(mutex);
2757 old_period = sysctl_sched_rt_period;
2758 old_runtime = sysctl_sched_rt_runtime;
2760 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2762 if (!ret && write) {
2763 ret = sched_rt_global_validate();
2767 ret = sched_dl_global_validate();
2771 ret = sched_rt_global_constraints();
2775 sched_rt_do_global();
2776 sched_dl_do_global();
2780 sysctl_sched_rt_period = old_period;
2781 sysctl_sched_rt_runtime = old_runtime;
2783 mutex_unlock(&mutex);
2788 int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
2789 size_t *lenp, loff_t *ppos)
2792 static DEFINE_MUTEX(mutex);
2795 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2797 * Make sure that internally we keep jiffies.
2798 * Also, writing zero resets the timeslice to default:
2800 if (!ret && write) {
2801 sched_rr_timeslice =
2802 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2803 msecs_to_jiffies(sysctl_sched_rr_timeslice);
2805 if (sysctl_sched_rr_timeslice <= 0)
2806 sysctl_sched_rr_timeslice = jiffies_to_msecs(RR_TIMESLICE);
2808 mutex_unlock(&mutex);
2813 #ifdef CONFIG_SCHED_DEBUG
2814 void print_rt_stats(struct seq_file *m, int cpu)
2817 struct rt_rq *rt_rq;
2820 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2821 print_rt_rq(m, cpu, rt_rq);
2824 #endif /* CONFIG_SCHED_DEBUG */