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;
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-1;
98 rt_rq->highest_prio.next = MAX_RT_PRIO-1;
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 unregister_rt_sched_group(struct task_group *tg)
148 destroy_rt_bandwidth(&tg->rt_bandwidth);
152 void free_rt_sched_group(struct task_group *tg)
156 for_each_possible_cpu(i) {
167 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
168 struct sched_rt_entity *rt_se, int cpu,
169 struct sched_rt_entity *parent)
171 struct rq *rq = cpu_rq(cpu);
173 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
174 rt_rq->rt_nr_boosted = 0;
178 tg->rt_rq[cpu] = rt_rq;
179 tg->rt_se[cpu] = rt_se;
185 rt_se->rt_rq = &rq->rt;
187 rt_se->rt_rq = parent->my_q;
190 rt_se->parent = parent;
191 INIT_LIST_HEAD(&rt_se->run_list);
194 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
197 struct sched_rt_entity *rt_se;
200 tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL);
203 tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL);
207 init_rt_bandwidth(&tg->rt_bandwidth,
208 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
210 for_each_possible_cpu(i) {
211 rt_rq = kzalloc_node(sizeof(struct rt_rq),
212 GFP_KERNEL, cpu_to_node(i));
216 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
217 GFP_KERNEL, cpu_to_node(i));
222 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
223 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
234 #else /* CONFIG_RT_GROUP_SCHED */
236 #define rt_entity_is_task(rt_se) (1)
238 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
240 return container_of(rt_se, struct task_struct, rt);
243 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
245 return container_of(rt_rq, struct rq, rt);
248 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
250 struct task_struct *p = rt_task_of(rt_se);
255 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
257 struct rq *rq = rq_of_rt_se(rt_se);
262 void unregister_rt_sched_group(struct task_group *tg) { }
264 void free_rt_sched_group(struct task_group *tg) { }
266 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
270 #endif /* CONFIG_RT_GROUP_SCHED */
274 static void pull_rt_task(struct rq *this_rq);
276 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
278 /* Try to pull RT tasks here if we lower this rq's prio */
279 return rq->online && rq->rt.highest_prio.curr > prev->prio;
282 static inline int rt_overloaded(struct rq *rq)
284 return atomic_read(&rq->rd->rto_count);
287 static inline void rt_set_overload(struct rq *rq)
292 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
294 * Make sure the mask is visible before we set
295 * the overload count. That is checked to determine
296 * if we should look at the mask. It would be a shame
297 * if we looked at the mask, but the mask was not
300 * Matched by the barrier in pull_rt_task().
303 atomic_inc(&rq->rd->rto_count);
306 static inline void rt_clear_overload(struct rq *rq)
311 /* the order here really doesn't matter */
312 atomic_dec(&rq->rd->rto_count);
313 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
316 static void update_rt_migration(struct rt_rq *rt_rq)
318 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
319 if (!rt_rq->overloaded) {
320 rt_set_overload(rq_of_rt_rq(rt_rq));
321 rt_rq->overloaded = 1;
323 } else if (rt_rq->overloaded) {
324 rt_clear_overload(rq_of_rt_rq(rt_rq));
325 rt_rq->overloaded = 0;
329 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
331 struct task_struct *p;
333 if (!rt_entity_is_task(rt_se))
336 p = rt_task_of(rt_se);
337 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
339 rt_rq->rt_nr_total++;
340 if (p->nr_cpus_allowed > 1)
341 rt_rq->rt_nr_migratory++;
343 update_rt_migration(rt_rq);
346 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
348 struct task_struct *p;
350 if (!rt_entity_is_task(rt_se))
353 p = rt_task_of(rt_se);
354 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
356 rt_rq->rt_nr_total--;
357 if (p->nr_cpus_allowed > 1)
358 rt_rq->rt_nr_migratory--;
360 update_rt_migration(rt_rq);
363 static inline int has_pushable_tasks(struct rq *rq)
365 return !plist_head_empty(&rq->rt.pushable_tasks);
368 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
369 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
371 static void push_rt_tasks(struct rq *);
372 static void pull_rt_task(struct rq *);
374 static inline void rt_queue_push_tasks(struct rq *rq)
376 if (!has_pushable_tasks(rq))
379 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
382 static inline void rt_queue_pull_task(struct rq *rq)
384 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
387 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
389 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
390 plist_node_init(&p->pushable_tasks, p->prio);
391 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
393 /* Update the highest prio pushable task */
394 if (p->prio < rq->rt.highest_prio.next)
395 rq->rt.highest_prio.next = p->prio;
398 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
400 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
402 /* Update the new highest prio pushable task */
403 if (has_pushable_tasks(rq)) {
404 p = plist_first_entry(&rq->rt.pushable_tasks,
405 struct task_struct, pushable_tasks);
406 rq->rt.highest_prio.next = p->prio;
408 rq->rt.highest_prio.next = MAX_RT_PRIO-1;
414 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
418 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
423 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
428 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
432 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
437 static inline void pull_rt_task(struct rq *this_rq)
441 static inline void rt_queue_push_tasks(struct rq *rq)
444 #endif /* CONFIG_SMP */
446 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
447 static void dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count);
449 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
454 #ifdef CONFIG_UCLAMP_TASK
456 * Verify the fitness of task @p to run on @cpu taking into account the uclamp
459 * This check is only important for heterogeneous systems where uclamp_min value
460 * is higher than the capacity of a @cpu. For non-heterogeneous system this
461 * function will always return true.
463 * The function will return true if the capacity of the @cpu is >= the
464 * uclamp_min and false otherwise.
466 * Note that uclamp_min will be clamped to uclamp_max if uclamp_min
469 static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
471 unsigned int min_cap;
472 unsigned int max_cap;
473 unsigned int cpu_cap;
475 /* Only heterogeneous systems can benefit from this check */
476 if (!sched_asym_cpucap_active())
479 min_cap = uclamp_eff_value(p, UCLAMP_MIN);
480 max_cap = uclamp_eff_value(p, UCLAMP_MAX);
482 cpu_cap = capacity_orig_of(cpu);
484 return cpu_cap >= min(min_cap, max_cap);
487 static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
493 #ifdef CONFIG_RT_GROUP_SCHED
495 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
500 return rt_rq->rt_runtime;
503 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
505 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
508 typedef struct task_group *rt_rq_iter_t;
510 static inline struct task_group *next_task_group(struct task_group *tg)
513 tg = list_entry_rcu(tg->list.next,
514 typeof(struct task_group), list);
515 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
517 if (&tg->list == &task_groups)
523 #define for_each_rt_rq(rt_rq, iter, rq) \
524 for (iter = container_of(&task_groups, typeof(*iter), list); \
525 (iter = next_task_group(iter)) && \
526 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
528 #define for_each_sched_rt_entity(rt_se) \
529 for (; rt_se; rt_se = rt_se->parent)
531 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
536 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
537 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
539 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
541 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
542 struct rq *rq = rq_of_rt_rq(rt_rq);
543 struct sched_rt_entity *rt_se;
545 int cpu = cpu_of(rq);
547 rt_se = rt_rq->tg->rt_se[cpu];
549 if (rt_rq->rt_nr_running) {
551 enqueue_top_rt_rq(rt_rq);
552 else if (!on_rt_rq(rt_se))
553 enqueue_rt_entity(rt_se, 0);
555 if (rt_rq->highest_prio.curr < curr->prio)
560 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
562 struct sched_rt_entity *rt_se;
563 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
565 rt_se = rt_rq->tg->rt_se[cpu];
568 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
569 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
570 cpufreq_update_util(rq_of_rt_rq(rt_rq), 0);
572 else if (on_rt_rq(rt_se))
573 dequeue_rt_entity(rt_se, 0);
576 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
578 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
581 static int rt_se_boosted(struct sched_rt_entity *rt_se)
583 struct rt_rq *rt_rq = group_rt_rq(rt_se);
584 struct task_struct *p;
587 return !!rt_rq->rt_nr_boosted;
589 p = rt_task_of(rt_se);
590 return p->prio != p->normal_prio;
594 static inline const struct cpumask *sched_rt_period_mask(void)
596 return this_rq()->rd->span;
599 static inline const struct cpumask *sched_rt_period_mask(void)
601 return cpu_online_mask;
606 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
608 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
611 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
613 return &rt_rq->tg->rt_bandwidth;
616 #else /* !CONFIG_RT_GROUP_SCHED */
618 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
620 return rt_rq->rt_runtime;
623 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
625 return ktime_to_ns(def_rt_bandwidth.rt_period);
628 typedef struct rt_rq *rt_rq_iter_t;
630 #define for_each_rt_rq(rt_rq, iter, rq) \
631 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
633 #define for_each_sched_rt_entity(rt_se) \
634 for (; rt_se; rt_se = NULL)
636 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
641 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
643 struct rq *rq = rq_of_rt_rq(rt_rq);
645 if (!rt_rq->rt_nr_running)
648 enqueue_top_rt_rq(rt_rq);
652 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
654 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
657 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
659 return rt_rq->rt_throttled;
662 static inline const struct cpumask *sched_rt_period_mask(void)
664 return cpu_online_mask;
668 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
670 return &cpu_rq(cpu)->rt;
673 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
675 return &def_rt_bandwidth;
678 #endif /* CONFIG_RT_GROUP_SCHED */
680 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
682 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
684 return (hrtimer_active(&rt_b->rt_period_timer) ||
685 rt_rq->rt_time < rt_b->rt_runtime);
690 * We ran out of runtime, see if we can borrow some from our neighbours.
692 static void do_balance_runtime(struct rt_rq *rt_rq)
694 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
695 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
699 weight = cpumask_weight(rd->span);
701 raw_spin_lock(&rt_b->rt_runtime_lock);
702 rt_period = ktime_to_ns(rt_b->rt_period);
703 for_each_cpu(i, rd->span) {
704 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
710 raw_spin_lock(&iter->rt_runtime_lock);
712 * Either all rqs have inf runtime and there's nothing to steal
713 * or __disable_runtime() below sets a specific rq to inf to
714 * indicate its been disabled and disallow stealing.
716 if (iter->rt_runtime == RUNTIME_INF)
720 * From runqueues with spare time, take 1/n part of their
721 * spare time, but no more than our period.
723 diff = iter->rt_runtime - iter->rt_time;
725 diff = div_u64((u64)diff, weight);
726 if (rt_rq->rt_runtime + diff > rt_period)
727 diff = rt_period - rt_rq->rt_runtime;
728 iter->rt_runtime -= diff;
729 rt_rq->rt_runtime += diff;
730 if (rt_rq->rt_runtime == rt_period) {
731 raw_spin_unlock(&iter->rt_runtime_lock);
736 raw_spin_unlock(&iter->rt_runtime_lock);
738 raw_spin_unlock(&rt_b->rt_runtime_lock);
742 * Ensure this RQ takes back all the runtime it lend to its neighbours.
744 static void __disable_runtime(struct rq *rq)
746 struct root_domain *rd = rq->rd;
750 if (unlikely(!scheduler_running))
753 for_each_rt_rq(rt_rq, iter, rq) {
754 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
758 raw_spin_lock(&rt_b->rt_runtime_lock);
759 raw_spin_lock(&rt_rq->rt_runtime_lock);
761 * Either we're all inf and nobody needs to borrow, or we're
762 * already disabled and thus have nothing to do, or we have
763 * exactly the right amount of runtime to take out.
765 if (rt_rq->rt_runtime == RUNTIME_INF ||
766 rt_rq->rt_runtime == rt_b->rt_runtime)
768 raw_spin_unlock(&rt_rq->rt_runtime_lock);
771 * Calculate the difference between what we started out with
772 * and what we current have, that's the amount of runtime
773 * we lend and now have to reclaim.
775 want = rt_b->rt_runtime - rt_rq->rt_runtime;
778 * Greedy reclaim, take back as much as we can.
780 for_each_cpu(i, rd->span) {
781 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
785 * Can't reclaim from ourselves or disabled runqueues.
787 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
790 raw_spin_lock(&iter->rt_runtime_lock);
792 diff = min_t(s64, iter->rt_runtime, want);
793 iter->rt_runtime -= diff;
796 iter->rt_runtime -= want;
799 raw_spin_unlock(&iter->rt_runtime_lock);
805 raw_spin_lock(&rt_rq->rt_runtime_lock);
807 * We cannot be left wanting - that would mean some runtime
808 * leaked out of the system.
813 * Disable all the borrow logic by pretending we have inf
814 * runtime - in which case borrowing doesn't make sense.
816 rt_rq->rt_runtime = RUNTIME_INF;
817 rt_rq->rt_throttled = 0;
818 raw_spin_unlock(&rt_rq->rt_runtime_lock);
819 raw_spin_unlock(&rt_b->rt_runtime_lock);
821 /* Make rt_rq available for pick_next_task() */
822 sched_rt_rq_enqueue(rt_rq);
826 static void __enable_runtime(struct rq *rq)
831 if (unlikely(!scheduler_running))
835 * Reset each runqueue's bandwidth settings
837 for_each_rt_rq(rt_rq, iter, rq) {
838 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
840 raw_spin_lock(&rt_b->rt_runtime_lock);
841 raw_spin_lock(&rt_rq->rt_runtime_lock);
842 rt_rq->rt_runtime = rt_b->rt_runtime;
844 rt_rq->rt_throttled = 0;
845 raw_spin_unlock(&rt_rq->rt_runtime_lock);
846 raw_spin_unlock(&rt_b->rt_runtime_lock);
850 static void balance_runtime(struct rt_rq *rt_rq)
852 if (!sched_feat(RT_RUNTIME_SHARE))
855 if (rt_rq->rt_time > rt_rq->rt_runtime) {
856 raw_spin_unlock(&rt_rq->rt_runtime_lock);
857 do_balance_runtime(rt_rq);
858 raw_spin_lock(&rt_rq->rt_runtime_lock);
861 #else /* !CONFIG_SMP */
862 static inline void balance_runtime(struct rt_rq *rt_rq) {}
863 #endif /* CONFIG_SMP */
865 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
867 int i, idle = 1, throttled = 0;
868 const struct cpumask *span;
870 span = sched_rt_period_mask();
871 #ifdef CONFIG_RT_GROUP_SCHED
873 * FIXME: isolated CPUs should really leave the root task group,
874 * whether they are isolcpus or were isolated via cpusets, lest
875 * the timer run on a CPU which does not service all runqueues,
876 * potentially leaving other CPUs indefinitely throttled. If
877 * isolation is really required, the user will turn the throttle
878 * off to kill the perturbations it causes anyway. Meanwhile,
879 * this maintains functionality for boot and/or troubleshooting.
881 if (rt_b == &root_task_group.rt_bandwidth)
882 span = cpu_online_mask;
884 for_each_cpu(i, span) {
886 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
887 struct rq *rq = rq_of_rt_rq(rt_rq);
892 * When span == cpu_online_mask, taking each rq->lock
893 * can be time-consuming. Try to avoid it when possible.
895 raw_spin_lock(&rt_rq->rt_runtime_lock);
896 if (!sched_feat(RT_RUNTIME_SHARE) && rt_rq->rt_runtime != RUNTIME_INF)
897 rt_rq->rt_runtime = rt_b->rt_runtime;
898 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
899 raw_spin_unlock(&rt_rq->rt_runtime_lock);
906 if (rt_rq->rt_time) {
909 raw_spin_lock(&rt_rq->rt_runtime_lock);
910 if (rt_rq->rt_throttled)
911 balance_runtime(rt_rq);
912 runtime = rt_rq->rt_runtime;
913 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
914 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
915 rt_rq->rt_throttled = 0;
919 * When we're idle and a woken (rt) task is
920 * throttled check_preempt_curr() will set
921 * skip_update and the time between the wakeup
922 * and this unthrottle will get accounted as
925 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
926 rq_clock_cancel_skipupdate(rq);
928 if (rt_rq->rt_time || rt_rq->rt_nr_running)
930 raw_spin_unlock(&rt_rq->rt_runtime_lock);
931 } else if (rt_rq->rt_nr_running) {
933 if (!rt_rq_throttled(rt_rq))
936 if (rt_rq->rt_throttled)
940 sched_rt_rq_enqueue(rt_rq);
944 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
950 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
952 #ifdef CONFIG_RT_GROUP_SCHED
953 struct rt_rq *rt_rq = group_rt_rq(rt_se);
956 return rt_rq->highest_prio.curr;
959 return rt_task_of(rt_se)->prio;
962 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
964 u64 runtime = sched_rt_runtime(rt_rq);
966 if (rt_rq->rt_throttled)
967 return rt_rq_throttled(rt_rq);
969 if (runtime >= sched_rt_period(rt_rq))
972 balance_runtime(rt_rq);
973 runtime = sched_rt_runtime(rt_rq);
974 if (runtime == RUNTIME_INF)
977 if (rt_rq->rt_time > runtime) {
978 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
981 * Don't actually throttle groups that have no runtime assigned
982 * but accrue some time due to boosting.
984 if (likely(rt_b->rt_runtime)) {
985 rt_rq->rt_throttled = 1;
986 printk_deferred_once("sched: RT throttling activated\n");
989 * In case we did anyway, make it go away,
990 * replenishment is a joke, since it will replenish us
996 if (rt_rq_throttled(rt_rq)) {
997 sched_rt_rq_dequeue(rt_rq);
1006 * Update the current task's runtime statistics. Skip current tasks that
1007 * are not in our scheduling class.
1009 static void update_curr_rt(struct rq *rq)
1011 struct task_struct *curr = rq->curr;
1012 struct sched_rt_entity *rt_se = &curr->rt;
1016 if (curr->sched_class != &rt_sched_class)
1019 now = rq_clock_task(rq);
1020 delta_exec = now - curr->se.exec_start;
1021 if (unlikely((s64)delta_exec <= 0))
1024 schedstat_set(curr->se.statistics.exec_max,
1025 max(curr->se.statistics.exec_max, delta_exec));
1027 curr->se.sum_exec_runtime += delta_exec;
1028 account_group_exec_runtime(curr, delta_exec);
1030 curr->se.exec_start = now;
1031 cgroup_account_cputime(curr, delta_exec);
1033 if (!rt_bandwidth_enabled())
1036 for_each_sched_rt_entity(rt_se) {
1037 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1040 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
1041 raw_spin_lock(&rt_rq->rt_runtime_lock);
1042 rt_rq->rt_time += delta_exec;
1043 exceeded = sched_rt_runtime_exceeded(rt_rq);
1046 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1048 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq));
1054 dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count)
1056 struct rq *rq = rq_of_rt_rq(rt_rq);
1058 BUG_ON(&rq->rt != rt_rq);
1060 if (!rt_rq->rt_queued)
1063 BUG_ON(!rq->nr_running);
1065 sub_nr_running(rq, count);
1066 rt_rq->rt_queued = 0;
1071 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1073 struct rq *rq = rq_of_rt_rq(rt_rq);
1075 BUG_ON(&rq->rt != rt_rq);
1077 if (rt_rq->rt_queued)
1080 if (rt_rq_throttled(rt_rq))
1083 if (rt_rq->rt_nr_running) {
1084 add_nr_running(rq, rt_rq->rt_nr_running);
1085 rt_rq->rt_queued = 1;
1088 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1089 cpufreq_update_util(rq, 0);
1092 #if defined CONFIG_SMP
1095 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1097 struct rq *rq = rq_of_rt_rq(rt_rq);
1099 #ifdef CONFIG_RT_GROUP_SCHED
1101 * Change rq's cpupri only if rt_rq is the top queue.
1103 if (&rq->rt != rt_rq)
1106 if (rq->online && prio < prev_prio)
1107 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1111 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1113 struct rq *rq = rq_of_rt_rq(rt_rq);
1115 #ifdef CONFIG_RT_GROUP_SCHED
1117 * Change rq's cpupri only if rt_rq is the top queue.
1119 if (&rq->rt != rt_rq)
1122 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1123 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1126 #else /* CONFIG_SMP */
1129 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1131 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1133 #endif /* CONFIG_SMP */
1135 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1137 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1139 int prev_prio = rt_rq->highest_prio.curr;
1141 if (prio < prev_prio)
1142 rt_rq->highest_prio.curr = prio;
1144 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1148 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1150 int prev_prio = rt_rq->highest_prio.curr;
1152 if (rt_rq->rt_nr_running) {
1154 WARN_ON(prio < prev_prio);
1157 * This may have been our highest task, and therefore
1158 * we may have some recomputation to do
1160 if (prio == prev_prio) {
1161 struct rt_prio_array *array = &rt_rq->active;
1163 rt_rq->highest_prio.curr =
1164 sched_find_first_bit(array->bitmap);
1168 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
1171 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1176 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1177 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1179 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1181 #ifdef CONFIG_RT_GROUP_SCHED
1184 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1186 if (rt_se_boosted(rt_se))
1187 rt_rq->rt_nr_boosted++;
1190 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1194 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1196 if (rt_se_boosted(rt_se))
1197 rt_rq->rt_nr_boosted--;
1199 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1202 #else /* CONFIG_RT_GROUP_SCHED */
1205 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1207 start_rt_bandwidth(&def_rt_bandwidth);
1211 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1213 #endif /* CONFIG_RT_GROUP_SCHED */
1216 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1218 struct rt_rq *group_rq = group_rt_rq(rt_se);
1221 return group_rq->rt_nr_running;
1227 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1229 struct rt_rq *group_rq = group_rt_rq(rt_se);
1230 struct task_struct *tsk;
1233 return group_rq->rr_nr_running;
1235 tsk = rt_task_of(rt_se);
1237 return (tsk->policy == SCHED_RR) ? 1 : 0;
1241 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1243 int prio = rt_se_prio(rt_se);
1245 WARN_ON(!rt_prio(prio));
1246 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1247 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1249 inc_rt_prio(rt_rq, prio);
1250 inc_rt_migration(rt_se, rt_rq);
1251 inc_rt_group(rt_se, rt_rq);
1255 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1257 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1258 WARN_ON(!rt_rq->rt_nr_running);
1259 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1260 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1262 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1263 dec_rt_migration(rt_se, rt_rq);
1264 dec_rt_group(rt_se, rt_rq);
1268 * Change rt_se->run_list location unless SAVE && !MOVE
1270 * assumes ENQUEUE/DEQUEUE flags match
1272 static inline bool move_entity(unsigned int flags)
1274 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1280 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1282 list_del_init(&rt_se->run_list);
1284 if (list_empty(array->queue + rt_se_prio(rt_se)))
1285 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1290 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1292 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1293 struct rt_prio_array *array = &rt_rq->active;
1294 struct rt_rq *group_rq = group_rt_rq(rt_se);
1295 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1298 * Don't enqueue the group if its throttled, or when empty.
1299 * The latter is a consequence of the former when a child group
1300 * get throttled and the current group doesn't have any other
1303 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1305 __delist_rt_entity(rt_se, array);
1309 if (move_entity(flags)) {
1310 WARN_ON_ONCE(rt_se->on_list);
1311 if (flags & ENQUEUE_HEAD)
1312 list_add(&rt_se->run_list, queue);
1314 list_add_tail(&rt_se->run_list, queue);
1316 __set_bit(rt_se_prio(rt_se), array->bitmap);
1321 inc_rt_tasks(rt_se, rt_rq);
1324 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1326 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1327 struct rt_prio_array *array = &rt_rq->active;
1329 if (move_entity(flags)) {
1330 WARN_ON_ONCE(!rt_se->on_list);
1331 __delist_rt_entity(rt_se, array);
1335 dec_rt_tasks(rt_se, rt_rq);
1339 * Because the prio of an upper entry depends on the lower
1340 * entries, we must remove entries top - down.
1342 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1344 struct sched_rt_entity *back = NULL;
1345 unsigned int rt_nr_running;
1347 for_each_sched_rt_entity(rt_se) {
1352 rt_nr_running = rt_rq_of_se(back)->rt_nr_running;
1354 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1355 if (on_rt_rq(rt_se))
1356 __dequeue_rt_entity(rt_se, flags);
1359 dequeue_top_rt_rq(rt_rq_of_se(back), rt_nr_running);
1362 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1364 struct rq *rq = rq_of_rt_se(rt_se);
1366 dequeue_rt_stack(rt_se, flags);
1367 for_each_sched_rt_entity(rt_se)
1368 __enqueue_rt_entity(rt_se, flags);
1369 enqueue_top_rt_rq(&rq->rt);
1372 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1374 struct rq *rq = rq_of_rt_se(rt_se);
1376 dequeue_rt_stack(rt_se, flags);
1378 for_each_sched_rt_entity(rt_se) {
1379 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1381 if (rt_rq && rt_rq->rt_nr_running)
1382 __enqueue_rt_entity(rt_se, flags);
1384 enqueue_top_rt_rq(&rq->rt);
1388 * Adding/removing a task to/from a priority array:
1391 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1393 struct sched_rt_entity *rt_se = &p->rt;
1395 if (flags & ENQUEUE_WAKEUP)
1398 enqueue_rt_entity(rt_se, flags);
1400 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1401 enqueue_pushable_task(rq, p);
1404 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1406 struct sched_rt_entity *rt_se = &p->rt;
1409 dequeue_rt_entity(rt_se, flags);
1411 dequeue_pushable_task(rq, p);
1415 * Put task to the head or the end of the run list without the overhead of
1416 * dequeue followed by enqueue.
1419 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1421 if (on_rt_rq(rt_se)) {
1422 struct rt_prio_array *array = &rt_rq->active;
1423 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1426 list_move(&rt_se->run_list, queue);
1428 list_move_tail(&rt_se->run_list, queue);
1432 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1434 struct sched_rt_entity *rt_se = &p->rt;
1435 struct rt_rq *rt_rq;
1437 for_each_sched_rt_entity(rt_se) {
1438 rt_rq = rt_rq_of_se(rt_se);
1439 requeue_rt_entity(rt_rq, rt_se, head);
1443 static void yield_task_rt(struct rq *rq)
1445 requeue_task_rt(rq, rq->curr, 0);
1449 static int find_lowest_rq(struct task_struct *task);
1452 select_task_rq_rt(struct task_struct *p, int cpu, int flags)
1454 struct task_struct *curr;
1458 /* For anything but wake ups, just return the task_cpu */
1459 if (!(flags & (WF_TTWU | WF_FORK)))
1465 curr = READ_ONCE(rq->curr); /* unlocked access */
1468 * If the current task on @p's runqueue is an RT task, then
1469 * try to see if we can wake this RT task up on another
1470 * runqueue. Otherwise simply start this RT task
1471 * on its current runqueue.
1473 * We want to avoid overloading runqueues. If the woken
1474 * task is a higher priority, then it will stay on this CPU
1475 * and the lower prio task should be moved to another CPU.
1476 * Even though this will probably make the lower prio task
1477 * lose its cache, we do not want to bounce a higher task
1478 * around just because it gave up its CPU, perhaps for a
1481 * For equal prio tasks, we just let the scheduler sort it out.
1483 * Otherwise, just let it ride on the affined RQ and the
1484 * post-schedule router will push the preempted task away
1486 * This test is optimistic, if we get it wrong the load-balancer
1487 * will have to sort it out.
1489 * We take into account the capacity of the CPU to ensure it fits the
1490 * requirement of the task - which is only important on heterogeneous
1491 * systems like big.LITTLE.
1494 unlikely(rt_task(curr)) &&
1495 (curr->nr_cpus_allowed < 2 || curr->prio <= p->prio);
1497 if (test || !rt_task_fits_capacity(p, cpu)) {
1498 int target = find_lowest_rq(p);
1501 * Bail out if we were forcing a migration to find a better
1502 * fitting CPU but our search failed.
1504 if (!test && target != -1 && !rt_task_fits_capacity(p, target))
1508 * Don't bother moving it if the destination CPU is
1509 * not running a lower priority task.
1512 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1523 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1526 * Current can't be migrated, useless to reschedule,
1527 * let's hope p can move out.
1529 if (rq->curr->nr_cpus_allowed == 1 ||
1530 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1534 * p is migratable, so let's not schedule it and
1535 * see if it is pushed or pulled somewhere else.
1537 if (p->nr_cpus_allowed != 1 &&
1538 cpupri_find(&rq->rd->cpupri, p, NULL))
1542 * There appear to be other CPUs that can accept
1543 * the current task but none can run 'p', so lets reschedule
1544 * to try and push the current task away:
1546 requeue_task_rt(rq, p, 1);
1550 static int balance_rt(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1552 if (!on_rt_rq(&p->rt) && need_pull_rt_task(rq, p)) {
1554 * This is OK, because current is on_cpu, which avoids it being
1555 * picked for load-balance and preemption/IRQs are still
1556 * disabled avoiding further scheduler activity on it and we've
1557 * not yet started the picking loop.
1559 rq_unpin_lock(rq, rf);
1561 rq_repin_lock(rq, rf);
1564 return sched_stop_runnable(rq) || sched_dl_runnable(rq) || sched_rt_runnable(rq);
1566 #endif /* CONFIG_SMP */
1569 * Preempt the current task with a newly woken task if needed:
1571 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1573 if (p->prio < rq->curr->prio) {
1582 * - the newly woken task is of equal priority to the current task
1583 * - the newly woken task is non-migratable while current is migratable
1584 * - current will be preempted on the next reschedule
1586 * we should check to see if current can readily move to a different
1587 * cpu. If so, we will reschedule to allow the push logic to try
1588 * to move current somewhere else, making room for our non-migratable
1591 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1592 check_preempt_equal_prio(rq, p);
1596 static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool first)
1598 p->se.exec_start = rq_clock_task(rq);
1600 /* The running task is never eligible for pushing */
1601 dequeue_pushable_task(rq, p);
1607 * If prev task was rt, put_prev_task() has already updated the
1608 * utilization. We only care of the case where we start to schedule a
1611 if (rq->curr->sched_class != &rt_sched_class)
1612 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1614 rt_queue_push_tasks(rq);
1617 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1618 struct rt_rq *rt_rq)
1620 struct rt_prio_array *array = &rt_rq->active;
1621 struct sched_rt_entity *next = NULL;
1622 struct list_head *queue;
1625 idx = sched_find_first_bit(array->bitmap);
1626 BUG_ON(idx >= MAX_RT_PRIO);
1628 queue = array->queue + idx;
1629 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1634 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1636 struct sched_rt_entity *rt_se;
1637 struct rt_rq *rt_rq = &rq->rt;
1640 rt_se = pick_next_rt_entity(rq, rt_rq);
1642 rt_rq = group_rt_rq(rt_se);
1645 return rt_task_of(rt_se);
1648 static struct task_struct *pick_task_rt(struct rq *rq)
1650 struct task_struct *p;
1652 if (!sched_rt_runnable(rq))
1655 p = _pick_next_task_rt(rq);
1660 static struct task_struct *pick_next_task_rt(struct rq *rq)
1662 struct task_struct *p = pick_task_rt(rq);
1665 set_next_task_rt(rq, p, true);
1670 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1674 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1677 * The previous task needs to be made eligible for pushing
1678 * if it is still active
1680 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1681 enqueue_pushable_task(rq, p);
1686 /* Only try algorithms three times */
1687 #define RT_MAX_TRIES 3
1689 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1691 if (!task_running(rq, p) &&
1692 cpumask_test_cpu(cpu, &p->cpus_mask))
1699 * Return the highest pushable rq's task, which is suitable to be executed
1700 * on the CPU, NULL otherwise
1702 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1704 struct plist_head *head = &rq->rt.pushable_tasks;
1705 struct task_struct *p;
1707 if (!has_pushable_tasks(rq))
1710 plist_for_each_entry(p, head, pushable_tasks) {
1711 if (pick_rt_task(rq, p, cpu))
1718 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1720 static int find_lowest_rq(struct task_struct *task)
1722 struct sched_domain *sd;
1723 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1724 int this_cpu = smp_processor_id();
1725 int cpu = task_cpu(task);
1728 /* Make sure the mask is initialized first */
1729 if (unlikely(!lowest_mask))
1732 if (task->nr_cpus_allowed == 1)
1733 return -1; /* No other targets possible */
1736 * If we're on asym system ensure we consider the different capacities
1737 * of the CPUs when searching for the lowest_mask.
1739 if (sched_asym_cpucap_active()) {
1741 ret = cpupri_find_fitness(&task_rq(task)->rd->cpupri,
1743 rt_task_fits_capacity);
1746 ret = cpupri_find(&task_rq(task)->rd->cpupri,
1751 return -1; /* No targets found */
1754 * At this point we have built a mask of CPUs representing the
1755 * lowest priority tasks in the system. Now we want to elect
1756 * the best one based on our affinity and topology.
1758 * We prioritize the last CPU that the task executed on since
1759 * it is most likely cache-hot in that location.
1761 if (cpumask_test_cpu(cpu, lowest_mask))
1765 * Otherwise, we consult the sched_domains span maps to figure
1766 * out which CPU is logically closest to our hot cache data.
1768 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1769 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1772 for_each_domain(cpu, sd) {
1773 if (sd->flags & SD_WAKE_AFFINE) {
1777 * "this_cpu" is cheaper to preempt than a
1780 if (this_cpu != -1 &&
1781 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1786 best_cpu = cpumask_any_and_distribute(lowest_mask,
1787 sched_domain_span(sd));
1788 if (best_cpu < nr_cpu_ids) {
1797 * And finally, if there were no matches within the domains
1798 * just give the caller *something* to work with from the compatible
1804 cpu = cpumask_any_distribute(lowest_mask);
1805 if (cpu < nr_cpu_ids)
1811 /* Will lock the rq it finds */
1812 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1814 struct rq *lowest_rq = NULL;
1818 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1819 cpu = find_lowest_rq(task);
1821 if ((cpu == -1) || (cpu == rq->cpu))
1824 lowest_rq = cpu_rq(cpu);
1826 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1828 * Target rq has tasks of equal or higher priority,
1829 * retrying does not release any lock and is unlikely
1830 * to yield a different result.
1836 /* if the prio of this runqueue changed, try again */
1837 if (double_lock_balance(rq, lowest_rq)) {
1839 * We had to unlock the run queue. In
1840 * the mean time, task could have
1841 * migrated already or had its affinity changed.
1842 * Also make sure that it wasn't scheduled on its rq.
1844 if (unlikely(task_rq(task) != rq ||
1845 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_mask) ||
1846 task_running(rq, task) ||
1848 !task_on_rq_queued(task))) {
1850 double_unlock_balance(rq, lowest_rq);
1856 /* If this rq is still suitable use it. */
1857 if (lowest_rq->rt.highest_prio.curr > task->prio)
1861 double_unlock_balance(rq, lowest_rq);
1868 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1870 struct task_struct *p;
1872 if (!has_pushable_tasks(rq))
1875 p = plist_first_entry(&rq->rt.pushable_tasks,
1876 struct task_struct, pushable_tasks);
1878 BUG_ON(rq->cpu != task_cpu(p));
1879 BUG_ON(task_current(rq, p));
1880 BUG_ON(p->nr_cpus_allowed <= 1);
1882 BUG_ON(!task_on_rq_queued(p));
1883 BUG_ON(!rt_task(p));
1889 * If the current CPU has more than one RT task, see if the non
1890 * running task can migrate over to a CPU that is running a task
1891 * of lesser priority.
1893 static int push_rt_task(struct rq *rq, bool pull)
1895 struct task_struct *next_task;
1896 struct rq *lowest_rq;
1899 if (!rq->rt.overloaded)
1902 next_task = pick_next_pushable_task(rq);
1908 * It's possible that the next_task slipped in of
1909 * higher priority than current. If that's the case
1910 * just reschedule current.
1912 if (unlikely(next_task->prio < rq->curr->prio)) {
1917 if (is_migration_disabled(next_task)) {
1918 struct task_struct *push_task = NULL;
1921 if (!pull || rq->push_busy)
1925 * Invoking find_lowest_rq() on anything but an RT task doesn't
1926 * make sense. Per the above priority check, curr has to
1927 * be of higher priority than next_task, so no need to
1928 * reschedule when bailing out.
1930 * Note that the stoppers are masqueraded as SCHED_FIFO
1931 * (cf. sched_set_stop_task()), so we can't rely on rt_task().
1933 if (rq->curr->sched_class != &rt_sched_class)
1936 cpu = find_lowest_rq(rq->curr);
1937 if (cpu == -1 || cpu == rq->cpu)
1941 * Given we found a CPU with lower priority than @next_task,
1942 * therefore it should be running. However we cannot migrate it
1943 * to this other CPU, instead attempt to push the current
1944 * running task on this CPU away.
1946 push_task = get_push_task(rq);
1948 raw_spin_rq_unlock(rq);
1949 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
1950 push_task, &rq->push_work);
1951 raw_spin_rq_lock(rq);
1957 if (WARN_ON(next_task == rq->curr))
1960 /* We might release rq lock */
1961 get_task_struct(next_task);
1963 /* find_lock_lowest_rq locks the rq if found */
1964 lowest_rq = find_lock_lowest_rq(next_task, rq);
1966 struct task_struct *task;
1968 * find_lock_lowest_rq releases rq->lock
1969 * so it is possible that next_task has migrated.
1971 * We need to make sure that the task is still on the same
1972 * run-queue and is also still the next task eligible for
1975 task = pick_next_pushable_task(rq);
1976 if (task == next_task) {
1978 * The task hasn't migrated, and is still the next
1979 * eligible task, but we failed to find a run-queue
1980 * to push it to. Do not retry in this case, since
1981 * other CPUs will pull from us when ready.
1987 /* No more tasks, just exit */
1991 * Something has shifted, try again.
1993 put_task_struct(next_task);
1998 deactivate_task(rq, next_task, 0);
1999 set_task_cpu(next_task, lowest_rq->cpu);
2000 activate_task(lowest_rq, next_task, 0);
2001 resched_curr(lowest_rq);
2004 double_unlock_balance(rq, lowest_rq);
2006 put_task_struct(next_task);
2011 static void push_rt_tasks(struct rq *rq)
2013 /* push_rt_task will return true if it moved an RT */
2014 while (push_rt_task(rq, false))
2018 #ifdef HAVE_RT_PUSH_IPI
2021 * When a high priority task schedules out from a CPU and a lower priority
2022 * task is scheduled in, a check is made to see if there's any RT tasks
2023 * on other CPUs that are waiting to run because a higher priority RT task
2024 * is currently running on its CPU. In this case, the CPU with multiple RT
2025 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
2026 * up that may be able to run one of its non-running queued RT tasks.
2028 * All CPUs with overloaded RT tasks need to be notified as there is currently
2029 * no way to know which of these CPUs have the highest priority task waiting
2030 * to run. Instead of trying to take a spinlock on each of these CPUs,
2031 * which has shown to cause large latency when done on machines with many
2032 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
2033 * RT tasks waiting to run.
2035 * Just sending an IPI to each of the CPUs is also an issue, as on large
2036 * count CPU machines, this can cause an IPI storm on a CPU, especially
2037 * if its the only CPU with multiple RT tasks queued, and a large number
2038 * of CPUs scheduling a lower priority task at the same time.
2040 * Each root domain has its own irq work function that can iterate over
2041 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
2042 * task must be checked if there's one or many CPUs that are lowering
2043 * their priority, there's a single irq work iterator that will try to
2044 * push off RT tasks that are waiting to run.
2046 * When a CPU schedules a lower priority task, it will kick off the
2047 * irq work iterator that will jump to each CPU with overloaded RT tasks.
2048 * As it only takes the first CPU that schedules a lower priority task
2049 * to start the process, the rto_start variable is incremented and if
2050 * the atomic result is one, then that CPU will try to take the rto_lock.
2051 * This prevents high contention on the lock as the process handles all
2052 * CPUs scheduling lower priority tasks.
2054 * All CPUs that are scheduling a lower priority task will increment the
2055 * rt_loop_next variable. This will make sure that the irq work iterator
2056 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
2057 * priority task, even if the iterator is in the middle of a scan. Incrementing
2058 * the rt_loop_next will cause the iterator to perform another scan.
2061 static int rto_next_cpu(struct root_domain *rd)
2067 * When starting the IPI RT pushing, the rto_cpu is set to -1,
2068 * rt_next_cpu() will simply return the first CPU found in
2071 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
2072 * will return the next CPU found in the rto_mask.
2074 * If there are no more CPUs left in the rto_mask, then a check is made
2075 * against rto_loop and rto_loop_next. rto_loop is only updated with
2076 * the rto_lock held, but any CPU may increment the rto_loop_next
2077 * without any locking.
2081 /* When rto_cpu is -1 this acts like cpumask_first() */
2082 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
2086 if (cpu < nr_cpu_ids)
2092 * ACQUIRE ensures we see the @rto_mask changes
2093 * made prior to the @next value observed.
2095 * Matches WMB in rt_set_overload().
2097 next = atomic_read_acquire(&rd->rto_loop_next);
2099 if (rd->rto_loop == next)
2102 rd->rto_loop = next;
2108 static inline bool rto_start_trylock(atomic_t *v)
2110 return !atomic_cmpxchg_acquire(v, 0, 1);
2113 static inline void rto_start_unlock(atomic_t *v)
2115 atomic_set_release(v, 0);
2118 static void tell_cpu_to_push(struct rq *rq)
2122 /* Keep the loop going if the IPI is currently active */
2123 atomic_inc(&rq->rd->rto_loop_next);
2125 /* Only one CPU can initiate a loop at a time */
2126 if (!rto_start_trylock(&rq->rd->rto_loop_start))
2129 raw_spin_lock(&rq->rd->rto_lock);
2132 * The rto_cpu is updated under the lock, if it has a valid CPU
2133 * then the IPI is still running and will continue due to the
2134 * update to loop_next, and nothing needs to be done here.
2135 * Otherwise it is finishing up and an ipi needs to be sent.
2137 if (rq->rd->rto_cpu < 0)
2138 cpu = rto_next_cpu(rq->rd);
2140 raw_spin_unlock(&rq->rd->rto_lock);
2142 rto_start_unlock(&rq->rd->rto_loop_start);
2145 /* Make sure the rd does not get freed while pushing */
2146 sched_get_rd(rq->rd);
2147 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2151 /* Called from hardirq context */
2152 void rto_push_irq_work_func(struct irq_work *work)
2154 struct root_domain *rd =
2155 container_of(work, struct root_domain, rto_push_work);
2162 * We do not need to grab the lock to check for has_pushable_tasks.
2163 * When it gets updated, a check is made if a push is possible.
2165 if (has_pushable_tasks(rq)) {
2166 raw_spin_rq_lock(rq);
2167 while (push_rt_task(rq, true))
2169 raw_spin_rq_unlock(rq);
2172 raw_spin_lock(&rd->rto_lock);
2174 /* Pass the IPI to the next rt overloaded queue */
2175 cpu = rto_next_cpu(rd);
2177 raw_spin_unlock(&rd->rto_lock);
2184 /* Try the next RT overloaded CPU */
2185 irq_work_queue_on(&rd->rto_push_work, cpu);
2187 #endif /* HAVE_RT_PUSH_IPI */
2189 static void pull_rt_task(struct rq *this_rq)
2191 int this_cpu = this_rq->cpu, cpu;
2192 bool resched = false;
2193 struct task_struct *p, *push_task;
2195 int rt_overload_count = rt_overloaded(this_rq);
2197 if (likely(!rt_overload_count))
2201 * Match the barrier from rt_set_overloaded; this guarantees that if we
2202 * see overloaded we must also see the rto_mask bit.
2206 /* If we are the only overloaded CPU do nothing */
2207 if (rt_overload_count == 1 &&
2208 cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2211 #ifdef HAVE_RT_PUSH_IPI
2212 if (sched_feat(RT_PUSH_IPI)) {
2213 tell_cpu_to_push(this_rq);
2218 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2219 if (this_cpu == cpu)
2222 src_rq = cpu_rq(cpu);
2225 * Don't bother taking the src_rq->lock if the next highest
2226 * task is known to be lower-priority than our current task.
2227 * This may look racy, but if this value is about to go
2228 * logically higher, the src_rq will push this task away.
2229 * And if its going logically lower, we do not care
2231 if (src_rq->rt.highest_prio.next >=
2232 this_rq->rt.highest_prio.curr)
2236 * We can potentially drop this_rq's lock in
2237 * double_lock_balance, and another CPU could
2241 double_lock_balance(this_rq, src_rq);
2244 * We can pull only a task, which is pushable
2245 * on its rq, and no others.
2247 p = pick_highest_pushable_task(src_rq, this_cpu);
2250 * Do we have an RT task that preempts
2251 * the to-be-scheduled task?
2253 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2254 WARN_ON(p == src_rq->curr);
2255 WARN_ON(!task_on_rq_queued(p));
2258 * There's a chance that p is higher in priority
2259 * than what's currently running on its CPU.
2260 * This is just that p is waking up and hasn't
2261 * had a chance to schedule. We only pull
2262 * p if it is lower in priority than the
2263 * current task on the run queue
2265 if (p->prio < src_rq->curr->prio)
2268 if (is_migration_disabled(p)) {
2269 push_task = get_push_task(src_rq);
2271 deactivate_task(src_rq, p, 0);
2272 set_task_cpu(p, this_cpu);
2273 activate_task(this_rq, p, 0);
2277 * We continue with the search, just in
2278 * case there's an even higher prio task
2279 * in another runqueue. (low likelihood
2284 double_unlock_balance(this_rq, src_rq);
2287 raw_spin_rq_unlock(this_rq);
2288 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2289 push_task, &src_rq->push_work);
2290 raw_spin_rq_lock(this_rq);
2295 resched_curr(this_rq);
2299 * If we are not running and we are not going to reschedule soon, we should
2300 * try to push tasks away now
2302 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2304 bool need_to_push = !task_running(rq, p) &&
2305 !test_tsk_need_resched(rq->curr) &&
2306 p->nr_cpus_allowed > 1 &&
2307 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2308 (rq->curr->nr_cpus_allowed < 2 ||
2309 rq->curr->prio <= p->prio);
2315 /* Assumes rq->lock is held */
2316 static void rq_online_rt(struct rq *rq)
2318 if (rq->rt.overloaded)
2319 rt_set_overload(rq);
2321 __enable_runtime(rq);
2323 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2326 /* Assumes rq->lock is held */
2327 static void rq_offline_rt(struct rq *rq)
2329 if (rq->rt.overloaded)
2330 rt_clear_overload(rq);
2332 __disable_runtime(rq);
2334 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2338 * When switch from the rt queue, we bring ourselves to a position
2339 * that we might want to pull RT tasks from other runqueues.
2341 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2344 * If there are other RT tasks then we will reschedule
2345 * and the scheduling of the other RT tasks will handle
2346 * the balancing. But if we are the last RT task
2347 * we may need to handle the pulling of RT tasks
2350 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2353 rt_queue_pull_task(rq);
2356 void __init init_sched_rt_class(void)
2360 for_each_possible_cpu(i) {
2361 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2362 GFP_KERNEL, cpu_to_node(i));
2365 #endif /* CONFIG_SMP */
2368 * When switching a task to RT, we may overload the runqueue
2369 * with RT tasks. In this case we try to push them off to
2372 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2375 * If we are running, update the avg_rt tracking, as the running time
2376 * will now on be accounted into the latter.
2378 if (task_current(rq, p)) {
2379 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2384 * If we are not running we may need to preempt the current
2385 * running task. If that current running task is also an RT task
2386 * then see if we can move to another run queue.
2388 if (task_on_rq_queued(p)) {
2390 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2391 rt_queue_push_tasks(rq);
2392 #endif /* CONFIG_SMP */
2393 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2399 * Priority of the task has changed. This may cause
2400 * us to initiate a push or pull.
2403 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2405 if (!task_on_rq_queued(p))
2408 if (task_current(rq, p)) {
2411 * If our priority decreases while running, we
2412 * may need to pull tasks to this runqueue.
2414 if (oldprio < p->prio)
2415 rt_queue_pull_task(rq);
2418 * If there's a higher priority task waiting to run
2421 if (p->prio > rq->rt.highest_prio.curr)
2424 /* For UP simply resched on drop of prio */
2425 if (oldprio < p->prio)
2427 #endif /* CONFIG_SMP */
2430 * This task is not running, but if it is
2431 * greater than the current running task
2434 if (p->prio < rq->curr->prio)
2439 #ifdef CONFIG_POSIX_TIMERS
2440 static void watchdog(struct rq *rq, struct task_struct *p)
2442 unsigned long soft, hard;
2444 /* max may change after cur was read, this will be fixed next tick */
2445 soft = task_rlimit(p, RLIMIT_RTTIME);
2446 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2448 if (soft != RLIM_INFINITY) {
2451 if (p->rt.watchdog_stamp != jiffies) {
2453 p->rt.watchdog_stamp = jiffies;
2456 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2457 if (p->rt.timeout > next) {
2458 posix_cputimers_rt_watchdog(&p->posix_cputimers,
2459 p->se.sum_exec_runtime);
2464 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2468 * scheduler tick hitting a task of our scheduling class.
2470 * NOTE: This function can be called remotely by the tick offload that
2471 * goes along full dynticks. Therefore no local assumption can be made
2472 * and everything must be accessed through the @rq and @curr passed in
2475 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2477 struct sched_rt_entity *rt_se = &p->rt;
2480 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2485 * RR tasks need a special form of timeslice management.
2486 * FIFO tasks have no timeslices.
2488 if (p->policy != SCHED_RR)
2491 if (--p->rt.time_slice)
2494 p->rt.time_slice = sched_rr_timeslice;
2497 * Requeue to the end of queue if we (and all of our ancestors) are not
2498 * the only element on the queue
2500 for_each_sched_rt_entity(rt_se) {
2501 if (rt_se->run_list.prev != rt_se->run_list.next) {
2502 requeue_task_rt(rq, p, 0);
2509 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2512 * Time slice is 0 for SCHED_FIFO tasks
2514 if (task->policy == SCHED_RR)
2515 return sched_rr_timeslice;
2520 DEFINE_SCHED_CLASS(rt) = {
2522 .enqueue_task = enqueue_task_rt,
2523 .dequeue_task = dequeue_task_rt,
2524 .yield_task = yield_task_rt,
2526 .check_preempt_curr = check_preempt_curr_rt,
2528 .pick_next_task = pick_next_task_rt,
2529 .put_prev_task = put_prev_task_rt,
2530 .set_next_task = set_next_task_rt,
2533 .balance = balance_rt,
2534 .pick_task = pick_task_rt,
2535 .select_task_rq = select_task_rq_rt,
2536 .set_cpus_allowed = set_cpus_allowed_common,
2537 .rq_online = rq_online_rt,
2538 .rq_offline = rq_offline_rt,
2539 .task_woken = task_woken_rt,
2540 .switched_from = switched_from_rt,
2541 .find_lock_rq = find_lock_lowest_rq,
2544 .task_tick = task_tick_rt,
2546 .get_rr_interval = get_rr_interval_rt,
2548 .prio_changed = prio_changed_rt,
2549 .switched_to = switched_to_rt,
2551 .update_curr = update_curr_rt,
2553 #ifdef CONFIG_UCLAMP_TASK
2554 .uclamp_enabled = 1,
2558 #ifdef CONFIG_RT_GROUP_SCHED
2560 * Ensure that the real time constraints are schedulable.
2562 static DEFINE_MUTEX(rt_constraints_mutex);
2564 static inline int tg_has_rt_tasks(struct task_group *tg)
2566 struct task_struct *task;
2567 struct css_task_iter it;
2571 * Autogroups do not have RT tasks; see autogroup_create().
2573 if (task_group_is_autogroup(tg))
2576 css_task_iter_start(&tg->css, 0, &it);
2577 while (!ret && (task = css_task_iter_next(&it)))
2578 ret |= rt_task(task);
2579 css_task_iter_end(&it);
2584 struct rt_schedulable_data {
2585 struct task_group *tg;
2590 static int tg_rt_schedulable(struct task_group *tg, void *data)
2592 struct rt_schedulable_data *d = data;
2593 struct task_group *child;
2594 unsigned long total, sum = 0;
2595 u64 period, runtime;
2597 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2598 runtime = tg->rt_bandwidth.rt_runtime;
2601 period = d->rt_period;
2602 runtime = d->rt_runtime;
2606 * Cannot have more runtime than the period.
2608 if (runtime > period && runtime != RUNTIME_INF)
2612 * Ensure we don't starve existing RT tasks if runtime turns zero.
2614 if (rt_bandwidth_enabled() && !runtime &&
2615 tg->rt_bandwidth.rt_runtime && tg_has_rt_tasks(tg))
2618 total = to_ratio(period, runtime);
2621 * Nobody can have more than the global setting allows.
2623 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2627 * The sum of our children's runtime should not exceed our own.
2629 list_for_each_entry_rcu(child, &tg->children, siblings) {
2630 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2631 runtime = child->rt_bandwidth.rt_runtime;
2633 if (child == d->tg) {
2634 period = d->rt_period;
2635 runtime = d->rt_runtime;
2638 sum += to_ratio(period, runtime);
2647 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2651 struct rt_schedulable_data data = {
2653 .rt_period = period,
2654 .rt_runtime = runtime,
2658 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2664 static int tg_set_rt_bandwidth(struct task_group *tg,
2665 u64 rt_period, u64 rt_runtime)
2670 * Disallowing the root group RT runtime is BAD, it would disallow the
2671 * kernel creating (and or operating) RT threads.
2673 if (tg == &root_task_group && rt_runtime == 0)
2676 /* No period doesn't make any sense. */
2681 * Bound quota to defend quota against overflow during bandwidth shift.
2683 if (rt_runtime != RUNTIME_INF && rt_runtime > max_rt_runtime)
2686 mutex_lock(&rt_constraints_mutex);
2687 err = __rt_schedulable(tg, rt_period, rt_runtime);
2691 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2692 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2693 tg->rt_bandwidth.rt_runtime = rt_runtime;
2695 for_each_possible_cpu(i) {
2696 struct rt_rq *rt_rq = tg->rt_rq[i];
2698 raw_spin_lock(&rt_rq->rt_runtime_lock);
2699 rt_rq->rt_runtime = rt_runtime;
2700 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2702 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2704 mutex_unlock(&rt_constraints_mutex);
2709 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2711 u64 rt_runtime, rt_period;
2713 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2714 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2715 if (rt_runtime_us < 0)
2716 rt_runtime = RUNTIME_INF;
2717 else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC)
2720 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2723 long sched_group_rt_runtime(struct task_group *tg)
2727 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2730 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2731 do_div(rt_runtime_us, NSEC_PER_USEC);
2732 return rt_runtime_us;
2735 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2737 u64 rt_runtime, rt_period;
2739 if (rt_period_us > U64_MAX / NSEC_PER_USEC)
2742 rt_period = rt_period_us * NSEC_PER_USEC;
2743 rt_runtime = tg->rt_bandwidth.rt_runtime;
2745 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2748 long sched_group_rt_period(struct task_group *tg)
2752 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2753 do_div(rt_period_us, NSEC_PER_USEC);
2754 return rt_period_us;
2757 static int sched_rt_global_constraints(void)
2761 mutex_lock(&rt_constraints_mutex);
2762 ret = __rt_schedulable(NULL, 0, 0);
2763 mutex_unlock(&rt_constraints_mutex);
2768 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2770 /* Don't accept realtime tasks when there is no way for them to run */
2771 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2777 #else /* !CONFIG_RT_GROUP_SCHED */
2778 static int sched_rt_global_constraints(void)
2780 unsigned long flags;
2783 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2784 for_each_possible_cpu(i) {
2785 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2787 raw_spin_lock(&rt_rq->rt_runtime_lock);
2788 rt_rq->rt_runtime = global_rt_runtime();
2789 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2791 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2795 #endif /* CONFIG_RT_GROUP_SCHED */
2797 static int sched_rt_global_validate(void)
2799 if (sysctl_sched_rt_period <= 0)
2802 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2803 ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
2804 ((u64)sysctl_sched_rt_runtime *
2805 NSEC_PER_USEC > max_rt_runtime)))
2811 static void sched_rt_do_global(void)
2813 unsigned long flags;
2815 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2816 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2817 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2818 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2821 int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
2822 size_t *lenp, loff_t *ppos)
2824 int old_period, old_runtime;
2825 static DEFINE_MUTEX(mutex);
2829 old_period = sysctl_sched_rt_period;
2830 old_runtime = sysctl_sched_rt_runtime;
2832 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2834 if (!ret && write) {
2835 ret = sched_rt_global_validate();
2839 ret = sched_dl_global_validate();
2843 ret = sched_rt_global_constraints();
2847 sched_rt_do_global();
2848 sched_dl_do_global();
2852 sysctl_sched_rt_period = old_period;
2853 sysctl_sched_rt_runtime = old_runtime;
2855 mutex_unlock(&mutex);
2860 int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
2861 size_t *lenp, loff_t *ppos)
2864 static DEFINE_MUTEX(mutex);
2867 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2869 * Make sure that internally we keep jiffies.
2870 * Also, writing zero resets the timeslice to default:
2872 if (!ret && write) {
2873 sched_rr_timeslice =
2874 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2875 msecs_to_jiffies(sysctl_sched_rr_timeslice);
2877 mutex_unlock(&mutex);
2882 #ifdef CONFIG_SCHED_DEBUG
2883 void print_rt_stats(struct seq_file *m, int cpu)
2886 struct rt_rq *rt_rq;
2889 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2890 print_rt_rq(m, cpu, rt_rq);
2893 #endif /* CONFIG_SCHED_DEBUG */