1 // SPDX-License-Identifier: GPL-2.0
3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
7 int sched_rr_timeslice = RR_TIMESLICE;
8 /* More than 4 hours if BW_SHIFT equals 20. */
9 static const u64 max_rt_runtime = MAX_BW;
11 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
13 struct rt_bandwidth def_rt_bandwidth;
16 * period over which we measure -rt task CPU usage in us.
19 unsigned int sysctl_sched_rt_period = 1000000;
22 * part of the period that we allow rt tasks to run in us.
25 int sysctl_sched_rt_runtime = 950000;
28 static int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
29 static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
30 size_t *lenp, loff_t *ppos);
31 static int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
32 size_t *lenp, loff_t *ppos);
33 static struct ctl_table sched_rt_sysctls[] = {
35 .procname = "sched_rt_period_us",
36 .data = &sysctl_sched_rt_period,
37 .maxlen = sizeof(unsigned int),
39 .proc_handler = sched_rt_handler,
42 .procname = "sched_rt_runtime_us",
43 .data = &sysctl_sched_rt_runtime,
44 .maxlen = sizeof(int),
46 .proc_handler = sched_rt_handler,
49 .procname = "sched_rr_timeslice_ms",
50 .data = &sysctl_sched_rr_timeslice,
51 .maxlen = sizeof(int),
53 .proc_handler = sched_rr_handler,
58 static int __init sched_rt_sysctl_init(void)
60 register_sysctl_init("kernel", sched_rt_sysctls);
63 late_initcall(sched_rt_sysctl_init);
66 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
68 struct rt_bandwidth *rt_b =
69 container_of(timer, struct rt_bandwidth, rt_period_timer);
73 raw_spin_lock(&rt_b->rt_runtime_lock);
75 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
79 raw_spin_unlock(&rt_b->rt_runtime_lock);
80 idle = do_sched_rt_period_timer(rt_b, overrun);
81 raw_spin_lock(&rt_b->rt_runtime_lock);
84 rt_b->rt_period_active = 0;
85 raw_spin_unlock(&rt_b->rt_runtime_lock);
87 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
90 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
92 rt_b->rt_period = ns_to_ktime(period);
93 rt_b->rt_runtime = runtime;
95 raw_spin_lock_init(&rt_b->rt_runtime_lock);
97 hrtimer_init(&rt_b->rt_period_timer, CLOCK_MONOTONIC,
98 HRTIMER_MODE_REL_HARD);
99 rt_b->rt_period_timer.function = sched_rt_period_timer;
102 static inline void do_start_rt_bandwidth(struct rt_bandwidth *rt_b)
104 raw_spin_lock(&rt_b->rt_runtime_lock);
105 if (!rt_b->rt_period_active) {
106 rt_b->rt_period_active = 1;
108 * SCHED_DEADLINE updates the bandwidth, as a run away
109 * RT task with a DL task could hog a CPU. But DL does
110 * not reset the period. If a deadline task was running
111 * without an RT task running, it can cause RT tasks to
112 * throttle when they start up. Kick the timer right away
113 * to update the period.
115 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
116 hrtimer_start_expires(&rt_b->rt_period_timer,
117 HRTIMER_MODE_ABS_PINNED_HARD);
119 raw_spin_unlock(&rt_b->rt_runtime_lock);
122 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
124 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
127 do_start_rt_bandwidth(rt_b);
130 void init_rt_rq(struct rt_rq *rt_rq)
132 struct rt_prio_array *array;
135 array = &rt_rq->active;
136 for (i = 0; i < MAX_RT_PRIO; i++) {
137 INIT_LIST_HEAD(array->queue + i);
138 __clear_bit(i, array->bitmap);
140 /* delimiter for bitsearch: */
141 __set_bit(MAX_RT_PRIO, array->bitmap);
143 #if defined CONFIG_SMP
144 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
145 rt_rq->highest_prio.next = MAX_RT_PRIO-1;
146 rt_rq->rt_nr_migratory = 0;
147 rt_rq->overloaded = 0;
148 plist_head_init(&rt_rq->pushable_tasks);
149 #endif /* CONFIG_SMP */
150 /* We start is dequeued state, because no RT tasks are queued */
151 rt_rq->rt_queued = 0;
154 rt_rq->rt_throttled = 0;
155 rt_rq->rt_runtime = 0;
156 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
159 #ifdef CONFIG_RT_GROUP_SCHED
160 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
162 hrtimer_cancel(&rt_b->rt_period_timer);
165 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
167 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
169 #ifdef CONFIG_SCHED_DEBUG
170 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
172 return container_of(rt_se, struct task_struct, rt);
175 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
180 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
185 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
187 struct rt_rq *rt_rq = rt_se->rt_rq;
192 void unregister_rt_sched_group(struct task_group *tg)
195 destroy_rt_bandwidth(&tg->rt_bandwidth);
199 void free_rt_sched_group(struct task_group *tg)
203 for_each_possible_cpu(i) {
214 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
215 struct sched_rt_entity *rt_se, int cpu,
216 struct sched_rt_entity *parent)
218 struct rq *rq = cpu_rq(cpu);
220 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
221 rt_rq->rt_nr_boosted = 0;
225 tg->rt_rq[cpu] = rt_rq;
226 tg->rt_se[cpu] = rt_se;
232 rt_se->rt_rq = &rq->rt;
234 rt_se->rt_rq = parent->my_q;
237 rt_se->parent = parent;
238 INIT_LIST_HEAD(&rt_se->run_list);
241 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
244 struct sched_rt_entity *rt_se;
247 tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL);
250 tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL);
254 init_rt_bandwidth(&tg->rt_bandwidth,
255 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
257 for_each_possible_cpu(i) {
258 rt_rq = kzalloc_node(sizeof(struct rt_rq),
259 GFP_KERNEL, cpu_to_node(i));
263 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
264 GFP_KERNEL, cpu_to_node(i));
269 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
270 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
281 #else /* CONFIG_RT_GROUP_SCHED */
283 #define rt_entity_is_task(rt_se) (1)
285 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
287 return container_of(rt_se, struct task_struct, rt);
290 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
292 return container_of(rt_rq, struct rq, rt);
295 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
297 struct task_struct *p = rt_task_of(rt_se);
302 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
304 struct rq *rq = rq_of_rt_se(rt_se);
309 void unregister_rt_sched_group(struct task_group *tg) { }
311 void free_rt_sched_group(struct task_group *tg) { }
313 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
317 #endif /* CONFIG_RT_GROUP_SCHED */
321 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
323 /* Try to pull RT tasks here if we lower this rq's prio */
324 return rq->online && rq->rt.highest_prio.curr > prev->prio;
327 static inline int rt_overloaded(struct rq *rq)
329 return atomic_read(&rq->rd->rto_count);
332 static inline void rt_set_overload(struct rq *rq)
337 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
339 * Make sure the mask is visible before we set
340 * the overload count. That is checked to determine
341 * if we should look at the mask. It would be a shame
342 * if we looked at the mask, but the mask was not
345 * Matched by the barrier in pull_rt_task().
348 atomic_inc(&rq->rd->rto_count);
351 static inline void rt_clear_overload(struct rq *rq)
356 /* the order here really doesn't matter */
357 atomic_dec(&rq->rd->rto_count);
358 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
361 static void update_rt_migration(struct rt_rq *rt_rq)
363 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
364 if (!rt_rq->overloaded) {
365 rt_set_overload(rq_of_rt_rq(rt_rq));
366 rt_rq->overloaded = 1;
368 } else if (rt_rq->overloaded) {
369 rt_clear_overload(rq_of_rt_rq(rt_rq));
370 rt_rq->overloaded = 0;
374 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
376 struct task_struct *p;
378 if (!rt_entity_is_task(rt_se))
381 p = rt_task_of(rt_se);
382 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
384 rt_rq->rt_nr_total++;
385 if (p->nr_cpus_allowed > 1)
386 rt_rq->rt_nr_migratory++;
388 update_rt_migration(rt_rq);
391 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
393 struct task_struct *p;
395 if (!rt_entity_is_task(rt_se))
398 p = rt_task_of(rt_se);
399 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
401 rt_rq->rt_nr_total--;
402 if (p->nr_cpus_allowed > 1)
403 rt_rq->rt_nr_migratory--;
405 update_rt_migration(rt_rq);
408 static inline int has_pushable_tasks(struct rq *rq)
410 return !plist_head_empty(&rq->rt.pushable_tasks);
413 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
414 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
416 static void push_rt_tasks(struct rq *);
417 static void pull_rt_task(struct rq *);
419 static inline void rt_queue_push_tasks(struct rq *rq)
421 if (!has_pushable_tasks(rq))
424 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
427 static inline void rt_queue_pull_task(struct rq *rq)
429 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
432 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
434 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
435 plist_node_init(&p->pushable_tasks, p->prio);
436 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
438 /* Update the highest prio pushable task */
439 if (p->prio < rq->rt.highest_prio.next)
440 rq->rt.highest_prio.next = p->prio;
443 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
445 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
447 /* Update the new highest prio pushable task */
448 if (has_pushable_tasks(rq)) {
449 p = plist_first_entry(&rq->rt.pushable_tasks,
450 struct task_struct, pushable_tasks);
451 rq->rt.highest_prio.next = p->prio;
453 rq->rt.highest_prio.next = MAX_RT_PRIO-1;
459 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
463 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
468 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
473 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
477 static inline void rt_queue_push_tasks(struct rq *rq)
480 #endif /* CONFIG_SMP */
482 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
483 static void dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count);
485 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
490 #ifdef CONFIG_UCLAMP_TASK
492 * Verify the fitness of task @p to run on @cpu taking into account the uclamp
495 * This check is only important for heterogeneous systems where uclamp_min value
496 * is higher than the capacity of a @cpu. For non-heterogeneous system this
497 * function will always return true.
499 * The function will return true if the capacity of the @cpu is >= the
500 * uclamp_min and false otherwise.
502 * Note that uclamp_min will be clamped to uclamp_max if uclamp_min
505 static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
507 unsigned int min_cap;
508 unsigned int max_cap;
509 unsigned int cpu_cap;
511 /* Only heterogeneous systems can benefit from this check */
512 if (!static_branch_unlikely(&sched_asym_cpucapacity))
515 min_cap = uclamp_eff_value(p, UCLAMP_MIN);
516 max_cap = uclamp_eff_value(p, UCLAMP_MAX);
518 cpu_cap = capacity_orig_of(cpu);
520 return cpu_cap >= min(min_cap, max_cap);
523 static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
529 #ifdef CONFIG_RT_GROUP_SCHED
531 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
536 return rt_rq->rt_runtime;
539 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
541 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
544 typedef struct task_group *rt_rq_iter_t;
546 static inline struct task_group *next_task_group(struct task_group *tg)
549 tg = list_entry_rcu(tg->list.next,
550 typeof(struct task_group), list);
551 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
553 if (&tg->list == &task_groups)
559 #define for_each_rt_rq(rt_rq, iter, rq) \
560 for (iter = container_of(&task_groups, typeof(*iter), list); \
561 (iter = next_task_group(iter)) && \
562 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
564 #define for_each_sched_rt_entity(rt_se) \
565 for (; rt_se; rt_se = rt_se->parent)
567 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
572 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
573 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
575 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
577 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
578 struct rq *rq = rq_of_rt_rq(rt_rq);
579 struct sched_rt_entity *rt_se;
581 int cpu = cpu_of(rq);
583 rt_se = rt_rq->tg->rt_se[cpu];
585 if (rt_rq->rt_nr_running) {
587 enqueue_top_rt_rq(rt_rq);
588 else if (!on_rt_rq(rt_se))
589 enqueue_rt_entity(rt_se, 0);
591 if (rt_rq->highest_prio.curr < curr->prio)
596 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
598 struct sched_rt_entity *rt_se;
599 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
601 rt_se = rt_rq->tg->rt_se[cpu];
604 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
605 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
606 cpufreq_update_util(rq_of_rt_rq(rt_rq), 0);
608 else if (on_rt_rq(rt_se))
609 dequeue_rt_entity(rt_se, 0);
612 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
614 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
617 static int rt_se_boosted(struct sched_rt_entity *rt_se)
619 struct rt_rq *rt_rq = group_rt_rq(rt_se);
620 struct task_struct *p;
623 return !!rt_rq->rt_nr_boosted;
625 p = rt_task_of(rt_se);
626 return p->prio != p->normal_prio;
630 static inline const struct cpumask *sched_rt_period_mask(void)
632 return this_rq()->rd->span;
635 static inline const struct cpumask *sched_rt_period_mask(void)
637 return cpu_online_mask;
642 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
644 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
647 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
649 return &rt_rq->tg->rt_bandwidth;
652 #else /* !CONFIG_RT_GROUP_SCHED */
654 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
656 return rt_rq->rt_runtime;
659 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
661 return ktime_to_ns(def_rt_bandwidth.rt_period);
664 typedef struct rt_rq *rt_rq_iter_t;
666 #define for_each_rt_rq(rt_rq, iter, rq) \
667 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
669 #define for_each_sched_rt_entity(rt_se) \
670 for (; rt_se; rt_se = NULL)
672 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
677 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
679 struct rq *rq = rq_of_rt_rq(rt_rq);
681 if (!rt_rq->rt_nr_running)
684 enqueue_top_rt_rq(rt_rq);
688 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
690 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
693 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
695 return rt_rq->rt_throttled;
698 static inline const struct cpumask *sched_rt_period_mask(void)
700 return cpu_online_mask;
704 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
706 return &cpu_rq(cpu)->rt;
709 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
711 return &def_rt_bandwidth;
714 #endif /* CONFIG_RT_GROUP_SCHED */
716 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
718 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
720 return (hrtimer_active(&rt_b->rt_period_timer) ||
721 rt_rq->rt_time < rt_b->rt_runtime);
726 * We ran out of runtime, see if we can borrow some from our neighbours.
728 static void do_balance_runtime(struct rt_rq *rt_rq)
730 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
731 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
735 weight = cpumask_weight(rd->span);
737 raw_spin_lock(&rt_b->rt_runtime_lock);
738 rt_period = ktime_to_ns(rt_b->rt_period);
739 for_each_cpu(i, rd->span) {
740 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
746 raw_spin_lock(&iter->rt_runtime_lock);
748 * Either all rqs have inf runtime and there's nothing to steal
749 * or __disable_runtime() below sets a specific rq to inf to
750 * indicate its been disabled and disallow stealing.
752 if (iter->rt_runtime == RUNTIME_INF)
756 * From runqueues with spare time, take 1/n part of their
757 * spare time, but no more than our period.
759 diff = iter->rt_runtime - iter->rt_time;
761 diff = div_u64((u64)diff, weight);
762 if (rt_rq->rt_runtime + diff > rt_period)
763 diff = rt_period - rt_rq->rt_runtime;
764 iter->rt_runtime -= diff;
765 rt_rq->rt_runtime += diff;
766 if (rt_rq->rt_runtime == rt_period) {
767 raw_spin_unlock(&iter->rt_runtime_lock);
772 raw_spin_unlock(&iter->rt_runtime_lock);
774 raw_spin_unlock(&rt_b->rt_runtime_lock);
778 * Ensure this RQ takes back all the runtime it lend to its neighbours.
780 static void __disable_runtime(struct rq *rq)
782 struct root_domain *rd = rq->rd;
786 if (unlikely(!scheduler_running))
789 for_each_rt_rq(rt_rq, iter, rq) {
790 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
794 raw_spin_lock(&rt_b->rt_runtime_lock);
795 raw_spin_lock(&rt_rq->rt_runtime_lock);
797 * Either we're all inf and nobody needs to borrow, or we're
798 * already disabled and thus have nothing to do, or we have
799 * exactly the right amount of runtime to take out.
801 if (rt_rq->rt_runtime == RUNTIME_INF ||
802 rt_rq->rt_runtime == rt_b->rt_runtime)
804 raw_spin_unlock(&rt_rq->rt_runtime_lock);
807 * Calculate the difference between what we started out with
808 * and what we current have, that's the amount of runtime
809 * we lend and now have to reclaim.
811 want = rt_b->rt_runtime - rt_rq->rt_runtime;
814 * Greedy reclaim, take back as much as we can.
816 for_each_cpu(i, rd->span) {
817 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
821 * Can't reclaim from ourselves or disabled runqueues.
823 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
826 raw_spin_lock(&iter->rt_runtime_lock);
828 diff = min_t(s64, iter->rt_runtime, want);
829 iter->rt_runtime -= diff;
832 iter->rt_runtime -= want;
835 raw_spin_unlock(&iter->rt_runtime_lock);
841 raw_spin_lock(&rt_rq->rt_runtime_lock);
843 * We cannot be left wanting - that would mean some runtime
844 * leaked out of the system.
849 * Disable all the borrow logic by pretending we have inf
850 * runtime - in which case borrowing doesn't make sense.
852 rt_rq->rt_runtime = RUNTIME_INF;
853 rt_rq->rt_throttled = 0;
854 raw_spin_unlock(&rt_rq->rt_runtime_lock);
855 raw_spin_unlock(&rt_b->rt_runtime_lock);
857 /* Make rt_rq available for pick_next_task() */
858 sched_rt_rq_enqueue(rt_rq);
862 static void __enable_runtime(struct rq *rq)
867 if (unlikely(!scheduler_running))
871 * Reset each runqueue's bandwidth settings
873 for_each_rt_rq(rt_rq, iter, rq) {
874 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
876 raw_spin_lock(&rt_b->rt_runtime_lock);
877 raw_spin_lock(&rt_rq->rt_runtime_lock);
878 rt_rq->rt_runtime = rt_b->rt_runtime;
880 rt_rq->rt_throttled = 0;
881 raw_spin_unlock(&rt_rq->rt_runtime_lock);
882 raw_spin_unlock(&rt_b->rt_runtime_lock);
886 static void balance_runtime(struct rt_rq *rt_rq)
888 if (!sched_feat(RT_RUNTIME_SHARE))
891 if (rt_rq->rt_time > rt_rq->rt_runtime) {
892 raw_spin_unlock(&rt_rq->rt_runtime_lock);
893 do_balance_runtime(rt_rq);
894 raw_spin_lock(&rt_rq->rt_runtime_lock);
897 #else /* !CONFIG_SMP */
898 static inline void balance_runtime(struct rt_rq *rt_rq) {}
899 #endif /* CONFIG_SMP */
901 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
903 int i, idle = 1, throttled = 0;
904 const struct cpumask *span;
906 span = sched_rt_period_mask();
907 #ifdef CONFIG_RT_GROUP_SCHED
909 * FIXME: isolated CPUs should really leave the root task group,
910 * whether they are isolcpus or were isolated via cpusets, lest
911 * the timer run on a CPU which does not service all runqueues,
912 * potentially leaving other CPUs indefinitely throttled. If
913 * isolation is really required, the user will turn the throttle
914 * off to kill the perturbations it causes anyway. Meanwhile,
915 * this maintains functionality for boot and/or troubleshooting.
917 if (rt_b == &root_task_group.rt_bandwidth)
918 span = cpu_online_mask;
920 for_each_cpu(i, span) {
922 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
923 struct rq *rq = rq_of_rt_rq(rt_rq);
928 * When span == cpu_online_mask, taking each rq->lock
929 * can be time-consuming. Try to avoid it when possible.
931 raw_spin_lock(&rt_rq->rt_runtime_lock);
932 if (!sched_feat(RT_RUNTIME_SHARE) && rt_rq->rt_runtime != RUNTIME_INF)
933 rt_rq->rt_runtime = rt_b->rt_runtime;
934 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
935 raw_spin_unlock(&rt_rq->rt_runtime_lock);
942 if (rt_rq->rt_time) {
945 raw_spin_lock(&rt_rq->rt_runtime_lock);
946 if (rt_rq->rt_throttled)
947 balance_runtime(rt_rq);
948 runtime = rt_rq->rt_runtime;
949 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
950 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
951 rt_rq->rt_throttled = 0;
955 * When we're idle and a woken (rt) task is
956 * throttled check_preempt_curr() will set
957 * skip_update and the time between the wakeup
958 * and this unthrottle will get accounted as
961 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
962 rq_clock_cancel_skipupdate(rq);
964 if (rt_rq->rt_time || rt_rq->rt_nr_running)
966 raw_spin_unlock(&rt_rq->rt_runtime_lock);
967 } else if (rt_rq->rt_nr_running) {
969 if (!rt_rq_throttled(rt_rq))
972 if (rt_rq->rt_throttled)
976 sched_rt_rq_enqueue(rt_rq);
980 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
986 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
988 #ifdef CONFIG_RT_GROUP_SCHED
989 struct rt_rq *rt_rq = group_rt_rq(rt_se);
992 return rt_rq->highest_prio.curr;
995 return rt_task_of(rt_se)->prio;
998 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
1000 u64 runtime = sched_rt_runtime(rt_rq);
1002 if (rt_rq->rt_throttled)
1003 return rt_rq_throttled(rt_rq);
1005 if (runtime >= sched_rt_period(rt_rq))
1008 balance_runtime(rt_rq);
1009 runtime = sched_rt_runtime(rt_rq);
1010 if (runtime == RUNTIME_INF)
1013 if (rt_rq->rt_time > runtime) {
1014 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
1017 * Don't actually throttle groups that have no runtime assigned
1018 * but accrue some time due to boosting.
1020 if (likely(rt_b->rt_runtime)) {
1021 rt_rq->rt_throttled = 1;
1022 printk_deferred_once("sched: RT throttling activated\n");
1025 * In case we did anyway, make it go away,
1026 * replenishment is a joke, since it will replenish us
1027 * with exactly 0 ns.
1032 if (rt_rq_throttled(rt_rq)) {
1033 sched_rt_rq_dequeue(rt_rq);
1042 * Update the current task's runtime statistics. Skip current tasks that
1043 * are not in our scheduling class.
1045 static void update_curr_rt(struct rq *rq)
1047 struct task_struct *curr = rq->curr;
1048 struct sched_rt_entity *rt_se = &curr->rt;
1052 if (curr->sched_class != &rt_sched_class)
1055 now = rq_clock_task(rq);
1056 delta_exec = now - curr->se.exec_start;
1057 if (unlikely((s64)delta_exec <= 0))
1060 schedstat_set(curr->stats.exec_max,
1061 max(curr->stats.exec_max, delta_exec));
1063 trace_sched_stat_runtime(curr, delta_exec, 0);
1065 curr->se.sum_exec_runtime += delta_exec;
1066 account_group_exec_runtime(curr, delta_exec);
1068 curr->se.exec_start = now;
1069 cgroup_account_cputime(curr, delta_exec);
1071 if (!rt_bandwidth_enabled())
1074 for_each_sched_rt_entity(rt_se) {
1075 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1078 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
1079 raw_spin_lock(&rt_rq->rt_runtime_lock);
1080 rt_rq->rt_time += delta_exec;
1081 exceeded = sched_rt_runtime_exceeded(rt_rq);
1084 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1086 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq));
1092 dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count)
1094 struct rq *rq = rq_of_rt_rq(rt_rq);
1096 BUG_ON(&rq->rt != rt_rq);
1098 if (!rt_rq->rt_queued)
1101 BUG_ON(!rq->nr_running);
1103 sub_nr_running(rq, count);
1104 rt_rq->rt_queued = 0;
1109 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1111 struct rq *rq = rq_of_rt_rq(rt_rq);
1113 BUG_ON(&rq->rt != rt_rq);
1115 if (rt_rq->rt_queued)
1118 if (rt_rq_throttled(rt_rq))
1121 if (rt_rq->rt_nr_running) {
1122 add_nr_running(rq, rt_rq->rt_nr_running);
1123 rt_rq->rt_queued = 1;
1126 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1127 cpufreq_update_util(rq, 0);
1130 #if defined CONFIG_SMP
1133 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1135 struct rq *rq = rq_of_rt_rq(rt_rq);
1137 #ifdef CONFIG_RT_GROUP_SCHED
1139 * Change rq's cpupri only if rt_rq is the top queue.
1141 if (&rq->rt != rt_rq)
1144 if (rq->online && prio < prev_prio)
1145 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1149 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1151 struct rq *rq = rq_of_rt_rq(rt_rq);
1153 #ifdef CONFIG_RT_GROUP_SCHED
1155 * Change rq's cpupri only if rt_rq is the top queue.
1157 if (&rq->rt != rt_rq)
1160 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1161 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1164 #else /* CONFIG_SMP */
1167 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1169 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1171 #endif /* CONFIG_SMP */
1173 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1175 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1177 int prev_prio = rt_rq->highest_prio.curr;
1179 if (prio < prev_prio)
1180 rt_rq->highest_prio.curr = prio;
1182 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1186 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1188 int prev_prio = rt_rq->highest_prio.curr;
1190 if (rt_rq->rt_nr_running) {
1192 WARN_ON(prio < prev_prio);
1195 * This may have been our highest task, and therefore
1196 * we may have some recomputation to do
1198 if (prio == prev_prio) {
1199 struct rt_prio_array *array = &rt_rq->active;
1201 rt_rq->highest_prio.curr =
1202 sched_find_first_bit(array->bitmap);
1206 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
1209 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1214 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1215 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1217 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1219 #ifdef CONFIG_RT_GROUP_SCHED
1222 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1224 if (rt_se_boosted(rt_se))
1225 rt_rq->rt_nr_boosted++;
1228 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1232 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1234 if (rt_se_boosted(rt_se))
1235 rt_rq->rt_nr_boosted--;
1237 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1240 #else /* CONFIG_RT_GROUP_SCHED */
1243 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1245 start_rt_bandwidth(&def_rt_bandwidth);
1249 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1251 #endif /* CONFIG_RT_GROUP_SCHED */
1254 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1256 struct rt_rq *group_rq = group_rt_rq(rt_se);
1259 return group_rq->rt_nr_running;
1265 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1267 struct rt_rq *group_rq = group_rt_rq(rt_se);
1268 struct task_struct *tsk;
1271 return group_rq->rr_nr_running;
1273 tsk = rt_task_of(rt_se);
1275 return (tsk->policy == SCHED_RR) ? 1 : 0;
1279 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1281 int prio = rt_se_prio(rt_se);
1283 WARN_ON(!rt_prio(prio));
1284 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1285 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1287 inc_rt_prio(rt_rq, prio);
1288 inc_rt_migration(rt_se, rt_rq);
1289 inc_rt_group(rt_se, rt_rq);
1293 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1295 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1296 WARN_ON(!rt_rq->rt_nr_running);
1297 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1298 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1300 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1301 dec_rt_migration(rt_se, rt_rq);
1302 dec_rt_group(rt_se, rt_rq);
1306 * Change rt_se->run_list location unless SAVE && !MOVE
1308 * assumes ENQUEUE/DEQUEUE flags match
1310 static inline bool move_entity(unsigned int flags)
1312 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1318 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1320 list_del_init(&rt_se->run_list);
1322 if (list_empty(array->queue + rt_se_prio(rt_se)))
1323 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1328 static inline struct sched_statistics *
1329 __schedstats_from_rt_se(struct sched_rt_entity *rt_se)
1331 #ifdef CONFIG_RT_GROUP_SCHED
1332 /* schedstats is not supported for rt group. */
1333 if (!rt_entity_is_task(rt_se))
1337 return &rt_task_of(rt_se)->stats;
1341 update_stats_wait_start_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1343 struct sched_statistics *stats;
1344 struct task_struct *p = NULL;
1346 if (!schedstat_enabled())
1349 if (rt_entity_is_task(rt_se))
1350 p = rt_task_of(rt_se);
1352 stats = __schedstats_from_rt_se(rt_se);
1356 __update_stats_wait_start(rq_of_rt_rq(rt_rq), p, stats);
1360 update_stats_enqueue_sleeper_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1362 struct sched_statistics *stats;
1363 struct task_struct *p = NULL;
1365 if (!schedstat_enabled())
1368 if (rt_entity_is_task(rt_se))
1369 p = rt_task_of(rt_se);
1371 stats = __schedstats_from_rt_se(rt_se);
1375 __update_stats_enqueue_sleeper(rq_of_rt_rq(rt_rq), p, stats);
1379 update_stats_enqueue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1382 if (!schedstat_enabled())
1385 if (flags & ENQUEUE_WAKEUP)
1386 update_stats_enqueue_sleeper_rt(rt_rq, rt_se);
1390 update_stats_wait_end_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1392 struct sched_statistics *stats;
1393 struct task_struct *p = NULL;
1395 if (!schedstat_enabled())
1398 if (rt_entity_is_task(rt_se))
1399 p = rt_task_of(rt_se);
1401 stats = __schedstats_from_rt_se(rt_se);
1405 __update_stats_wait_end(rq_of_rt_rq(rt_rq), p, stats);
1409 update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1412 struct task_struct *p = NULL;
1414 if (!schedstat_enabled())
1417 if (rt_entity_is_task(rt_se))
1418 p = rt_task_of(rt_se);
1420 if ((flags & DEQUEUE_SLEEP) && p) {
1423 state = READ_ONCE(p->__state);
1424 if (state & TASK_INTERRUPTIBLE)
1425 __schedstat_set(p->stats.sleep_start,
1426 rq_clock(rq_of_rt_rq(rt_rq)));
1428 if (state & TASK_UNINTERRUPTIBLE)
1429 __schedstat_set(p->stats.block_start,
1430 rq_clock(rq_of_rt_rq(rt_rq)));
1434 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1436 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1437 struct rt_prio_array *array = &rt_rq->active;
1438 struct rt_rq *group_rq = group_rt_rq(rt_se);
1439 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1442 * Don't enqueue the group if its throttled, or when empty.
1443 * The latter is a consequence of the former when a child group
1444 * get throttled and the current group doesn't have any other
1447 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1449 __delist_rt_entity(rt_se, array);
1453 if (move_entity(flags)) {
1454 WARN_ON_ONCE(rt_se->on_list);
1455 if (flags & ENQUEUE_HEAD)
1456 list_add(&rt_se->run_list, queue);
1458 list_add_tail(&rt_se->run_list, queue);
1460 __set_bit(rt_se_prio(rt_se), array->bitmap);
1465 inc_rt_tasks(rt_se, rt_rq);
1468 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1470 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1471 struct rt_prio_array *array = &rt_rq->active;
1473 if (move_entity(flags)) {
1474 WARN_ON_ONCE(!rt_se->on_list);
1475 __delist_rt_entity(rt_se, array);
1479 dec_rt_tasks(rt_se, rt_rq);
1483 * Because the prio of an upper entry depends on the lower
1484 * entries, we must remove entries top - down.
1486 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1488 struct sched_rt_entity *back = NULL;
1489 unsigned int rt_nr_running;
1491 for_each_sched_rt_entity(rt_se) {
1496 rt_nr_running = rt_rq_of_se(back)->rt_nr_running;
1498 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1499 if (on_rt_rq(rt_se))
1500 __dequeue_rt_entity(rt_se, flags);
1503 dequeue_top_rt_rq(rt_rq_of_se(back), rt_nr_running);
1506 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1508 struct rq *rq = rq_of_rt_se(rt_se);
1510 update_stats_enqueue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1512 dequeue_rt_stack(rt_se, flags);
1513 for_each_sched_rt_entity(rt_se)
1514 __enqueue_rt_entity(rt_se, flags);
1515 enqueue_top_rt_rq(&rq->rt);
1518 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1520 struct rq *rq = rq_of_rt_se(rt_se);
1522 update_stats_dequeue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1524 dequeue_rt_stack(rt_se, flags);
1526 for_each_sched_rt_entity(rt_se) {
1527 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1529 if (rt_rq && rt_rq->rt_nr_running)
1530 __enqueue_rt_entity(rt_se, flags);
1532 enqueue_top_rt_rq(&rq->rt);
1536 * Adding/removing a task to/from a priority array:
1539 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1541 struct sched_rt_entity *rt_se = &p->rt;
1543 if (flags & ENQUEUE_WAKEUP)
1546 check_schedstat_required();
1547 update_stats_wait_start_rt(rt_rq_of_se(rt_se), rt_se);
1549 enqueue_rt_entity(rt_se, flags);
1551 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1552 enqueue_pushable_task(rq, p);
1555 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1557 struct sched_rt_entity *rt_se = &p->rt;
1560 dequeue_rt_entity(rt_se, flags);
1562 dequeue_pushable_task(rq, p);
1566 * Put task to the head or the end of the run list without the overhead of
1567 * dequeue followed by enqueue.
1570 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1572 if (on_rt_rq(rt_se)) {
1573 struct rt_prio_array *array = &rt_rq->active;
1574 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1577 list_move(&rt_se->run_list, queue);
1579 list_move_tail(&rt_se->run_list, queue);
1583 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1585 struct sched_rt_entity *rt_se = &p->rt;
1586 struct rt_rq *rt_rq;
1588 for_each_sched_rt_entity(rt_se) {
1589 rt_rq = rt_rq_of_se(rt_se);
1590 requeue_rt_entity(rt_rq, rt_se, head);
1594 static void yield_task_rt(struct rq *rq)
1596 requeue_task_rt(rq, rq->curr, 0);
1600 static int find_lowest_rq(struct task_struct *task);
1603 select_task_rq_rt(struct task_struct *p, int cpu, int flags)
1605 struct task_struct *curr;
1609 /* For anything but wake ups, just return the task_cpu */
1610 if (!(flags & (WF_TTWU | WF_FORK)))
1616 curr = READ_ONCE(rq->curr); /* unlocked access */
1619 * If the current task on @p's runqueue is an RT task, then
1620 * try to see if we can wake this RT task up on another
1621 * runqueue. Otherwise simply start this RT task
1622 * on its current runqueue.
1624 * We want to avoid overloading runqueues. If the woken
1625 * task is a higher priority, then it will stay on this CPU
1626 * and the lower prio task should be moved to another CPU.
1627 * Even though this will probably make the lower prio task
1628 * lose its cache, we do not want to bounce a higher task
1629 * around just because it gave up its CPU, perhaps for a
1632 * For equal prio tasks, we just let the scheduler sort it out.
1634 * Otherwise, just let it ride on the affined RQ and the
1635 * post-schedule router will push the preempted task away
1637 * This test is optimistic, if we get it wrong the load-balancer
1638 * will have to sort it out.
1640 * We take into account the capacity of the CPU to ensure it fits the
1641 * requirement of the task - which is only important on heterogeneous
1642 * systems like big.LITTLE.
1645 unlikely(rt_task(curr)) &&
1646 (curr->nr_cpus_allowed < 2 || curr->prio <= p->prio);
1648 if (test || !rt_task_fits_capacity(p, cpu)) {
1649 int target = find_lowest_rq(p);
1652 * Bail out if we were forcing a migration to find a better
1653 * fitting CPU but our search failed.
1655 if (!test && target != -1 && !rt_task_fits_capacity(p, target))
1659 * Don't bother moving it if the destination CPU is
1660 * not running a lower priority task.
1663 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1674 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1677 * Current can't be migrated, useless to reschedule,
1678 * let's hope p can move out.
1680 if (rq->curr->nr_cpus_allowed == 1 ||
1681 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1685 * p is migratable, so let's not schedule it and
1686 * see if it is pushed or pulled somewhere else.
1688 if (p->nr_cpus_allowed != 1 &&
1689 cpupri_find(&rq->rd->cpupri, p, NULL))
1693 * There appear to be other CPUs that can accept
1694 * the current task but none can run 'p', so lets reschedule
1695 * to try and push the current task away:
1697 requeue_task_rt(rq, p, 1);
1701 static int balance_rt(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1703 if (!on_rt_rq(&p->rt) && need_pull_rt_task(rq, p)) {
1705 * This is OK, because current is on_cpu, which avoids it being
1706 * picked for load-balance and preemption/IRQs are still
1707 * disabled avoiding further scheduler activity on it and we've
1708 * not yet started the picking loop.
1710 rq_unpin_lock(rq, rf);
1712 rq_repin_lock(rq, rf);
1715 return sched_stop_runnable(rq) || sched_dl_runnable(rq) || sched_rt_runnable(rq);
1717 #endif /* CONFIG_SMP */
1720 * Preempt the current task with a newly woken task if needed:
1722 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1724 if (p->prio < rq->curr->prio) {
1733 * - the newly woken task is of equal priority to the current task
1734 * - the newly woken task is non-migratable while current is migratable
1735 * - current will be preempted on the next reschedule
1737 * we should check to see if current can readily move to a different
1738 * cpu. If so, we will reschedule to allow the push logic to try
1739 * to move current somewhere else, making room for our non-migratable
1742 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1743 check_preempt_equal_prio(rq, p);
1747 static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool first)
1749 struct sched_rt_entity *rt_se = &p->rt;
1750 struct rt_rq *rt_rq = &rq->rt;
1752 p->se.exec_start = rq_clock_task(rq);
1753 if (on_rt_rq(&p->rt))
1754 update_stats_wait_end_rt(rt_rq, rt_se);
1756 /* The running task is never eligible for pushing */
1757 dequeue_pushable_task(rq, p);
1763 * If prev task was rt, put_prev_task() has already updated the
1764 * utilization. We only care of the case where we start to schedule a
1767 if (rq->curr->sched_class != &rt_sched_class)
1768 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1770 rt_queue_push_tasks(rq);
1773 static struct sched_rt_entity *pick_next_rt_entity(struct rt_rq *rt_rq)
1775 struct rt_prio_array *array = &rt_rq->active;
1776 struct sched_rt_entity *next = NULL;
1777 struct list_head *queue;
1780 idx = sched_find_first_bit(array->bitmap);
1781 BUG_ON(idx >= MAX_RT_PRIO);
1783 queue = array->queue + idx;
1784 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1789 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1791 struct sched_rt_entity *rt_se;
1792 struct rt_rq *rt_rq = &rq->rt;
1795 rt_se = pick_next_rt_entity(rt_rq);
1797 rt_rq = group_rt_rq(rt_se);
1800 return rt_task_of(rt_se);
1803 static struct task_struct *pick_task_rt(struct rq *rq)
1805 struct task_struct *p;
1807 if (!sched_rt_runnable(rq))
1810 p = _pick_next_task_rt(rq);
1815 static struct task_struct *pick_next_task_rt(struct rq *rq)
1817 struct task_struct *p = pick_task_rt(rq);
1820 set_next_task_rt(rq, p, true);
1825 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1827 struct sched_rt_entity *rt_se = &p->rt;
1828 struct rt_rq *rt_rq = &rq->rt;
1830 if (on_rt_rq(&p->rt))
1831 update_stats_wait_start_rt(rt_rq, rt_se);
1835 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1838 * The previous task needs to be made eligible for pushing
1839 * if it is still active
1841 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1842 enqueue_pushable_task(rq, p);
1847 /* Only try algorithms three times */
1848 #define RT_MAX_TRIES 3
1850 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1852 if (!task_running(rq, p) &&
1853 cpumask_test_cpu(cpu, &p->cpus_mask))
1860 * Return the highest pushable rq's task, which is suitable to be executed
1861 * on the CPU, NULL otherwise
1863 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1865 struct plist_head *head = &rq->rt.pushable_tasks;
1866 struct task_struct *p;
1868 if (!has_pushable_tasks(rq))
1871 plist_for_each_entry(p, head, pushable_tasks) {
1872 if (pick_rt_task(rq, p, cpu))
1879 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1881 static int find_lowest_rq(struct task_struct *task)
1883 struct sched_domain *sd;
1884 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1885 int this_cpu = smp_processor_id();
1886 int cpu = task_cpu(task);
1889 /* Make sure the mask is initialized first */
1890 if (unlikely(!lowest_mask))
1893 if (task->nr_cpus_allowed == 1)
1894 return -1; /* No other targets possible */
1897 * If we're on asym system ensure we consider the different capacities
1898 * of the CPUs when searching for the lowest_mask.
1900 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
1902 ret = cpupri_find_fitness(&task_rq(task)->rd->cpupri,
1904 rt_task_fits_capacity);
1907 ret = cpupri_find(&task_rq(task)->rd->cpupri,
1912 return -1; /* No targets found */
1915 * At this point we have built a mask of CPUs representing the
1916 * lowest priority tasks in the system. Now we want to elect
1917 * the best one based on our affinity and topology.
1919 * We prioritize the last CPU that the task executed on since
1920 * it is most likely cache-hot in that location.
1922 if (cpumask_test_cpu(cpu, lowest_mask))
1926 * Otherwise, we consult the sched_domains span maps to figure
1927 * out which CPU is logically closest to our hot cache data.
1929 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1930 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1933 for_each_domain(cpu, sd) {
1934 if (sd->flags & SD_WAKE_AFFINE) {
1938 * "this_cpu" is cheaper to preempt than a
1941 if (this_cpu != -1 &&
1942 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1947 best_cpu = cpumask_any_and_distribute(lowest_mask,
1948 sched_domain_span(sd));
1949 if (best_cpu < nr_cpu_ids) {
1958 * And finally, if there were no matches within the domains
1959 * just give the caller *something* to work with from the compatible
1965 cpu = cpumask_any_distribute(lowest_mask);
1966 if (cpu < nr_cpu_ids)
1972 /* Will lock the rq it finds */
1973 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1975 struct rq *lowest_rq = NULL;
1979 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1980 cpu = find_lowest_rq(task);
1982 if ((cpu == -1) || (cpu == rq->cpu))
1985 lowest_rq = cpu_rq(cpu);
1987 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1989 * Target rq has tasks of equal or higher priority,
1990 * retrying does not release any lock and is unlikely
1991 * to yield a different result.
1997 /* if the prio of this runqueue changed, try again */
1998 if (double_lock_balance(rq, lowest_rq)) {
2000 * We had to unlock the run queue. In
2001 * the mean time, task could have
2002 * migrated already or had its affinity changed.
2003 * Also make sure that it wasn't scheduled on its rq.
2005 if (unlikely(task_rq(task) != rq ||
2006 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_mask) ||
2007 task_running(rq, task) ||
2009 !task_on_rq_queued(task))) {
2011 double_unlock_balance(rq, lowest_rq);
2017 /* If this rq is still suitable use it. */
2018 if (lowest_rq->rt.highest_prio.curr > task->prio)
2022 double_unlock_balance(rq, lowest_rq);
2029 static struct task_struct *pick_next_pushable_task(struct rq *rq)
2031 struct task_struct *p;
2033 if (!has_pushable_tasks(rq))
2036 p = plist_first_entry(&rq->rt.pushable_tasks,
2037 struct task_struct, pushable_tasks);
2039 BUG_ON(rq->cpu != task_cpu(p));
2040 BUG_ON(task_current(rq, p));
2041 BUG_ON(p->nr_cpus_allowed <= 1);
2043 BUG_ON(!task_on_rq_queued(p));
2044 BUG_ON(!rt_task(p));
2050 * If the current CPU has more than one RT task, see if the non
2051 * running task can migrate over to a CPU that is running a task
2052 * of lesser priority.
2054 static int push_rt_task(struct rq *rq, bool pull)
2056 struct task_struct *next_task;
2057 struct rq *lowest_rq;
2060 if (!rq->rt.overloaded)
2063 next_task = pick_next_pushable_task(rq);
2069 * It's possible that the next_task slipped in of
2070 * higher priority than current. If that's the case
2071 * just reschedule current.
2073 if (unlikely(next_task->prio < rq->curr->prio)) {
2078 if (is_migration_disabled(next_task)) {
2079 struct task_struct *push_task = NULL;
2082 if (!pull || rq->push_busy)
2086 * Invoking find_lowest_rq() on anything but an RT task doesn't
2087 * make sense. Per the above priority check, curr has to
2088 * be of higher priority than next_task, so no need to
2089 * reschedule when bailing out.
2091 * Note that the stoppers are masqueraded as SCHED_FIFO
2092 * (cf. sched_set_stop_task()), so we can't rely on rt_task().
2094 if (rq->curr->sched_class != &rt_sched_class)
2097 cpu = find_lowest_rq(rq->curr);
2098 if (cpu == -1 || cpu == rq->cpu)
2102 * Given we found a CPU with lower priority than @next_task,
2103 * therefore it should be running. However we cannot migrate it
2104 * to this other CPU, instead attempt to push the current
2105 * running task on this CPU away.
2107 push_task = get_push_task(rq);
2109 raw_spin_rq_unlock(rq);
2110 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2111 push_task, &rq->push_work);
2112 raw_spin_rq_lock(rq);
2118 if (WARN_ON(next_task == rq->curr))
2121 /* We might release rq lock */
2122 get_task_struct(next_task);
2124 /* find_lock_lowest_rq locks the rq if found */
2125 lowest_rq = find_lock_lowest_rq(next_task, rq);
2127 struct task_struct *task;
2129 * find_lock_lowest_rq releases rq->lock
2130 * so it is possible that next_task has migrated.
2132 * We need to make sure that the task is still on the same
2133 * run-queue and is also still the next task eligible for
2136 task = pick_next_pushable_task(rq);
2137 if (task == next_task) {
2139 * The task hasn't migrated, and is still the next
2140 * eligible task, but we failed to find a run-queue
2141 * to push it to. Do not retry in this case, since
2142 * other CPUs will pull from us when ready.
2148 /* No more tasks, just exit */
2152 * Something has shifted, try again.
2154 put_task_struct(next_task);
2159 deactivate_task(rq, next_task, 0);
2160 set_task_cpu(next_task, lowest_rq->cpu);
2161 activate_task(lowest_rq, next_task, 0);
2162 resched_curr(lowest_rq);
2165 double_unlock_balance(rq, lowest_rq);
2167 put_task_struct(next_task);
2172 static void push_rt_tasks(struct rq *rq)
2174 /* push_rt_task will return true if it moved an RT */
2175 while (push_rt_task(rq, false))
2179 #ifdef HAVE_RT_PUSH_IPI
2182 * When a high priority task schedules out from a CPU and a lower priority
2183 * task is scheduled in, a check is made to see if there's any RT tasks
2184 * on other CPUs that are waiting to run because a higher priority RT task
2185 * is currently running on its CPU. In this case, the CPU with multiple RT
2186 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
2187 * up that may be able to run one of its non-running queued RT tasks.
2189 * All CPUs with overloaded RT tasks need to be notified as there is currently
2190 * no way to know which of these CPUs have the highest priority task waiting
2191 * to run. Instead of trying to take a spinlock on each of these CPUs,
2192 * which has shown to cause large latency when done on machines with many
2193 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
2194 * RT tasks waiting to run.
2196 * Just sending an IPI to each of the CPUs is also an issue, as on large
2197 * count CPU machines, this can cause an IPI storm on a CPU, especially
2198 * if its the only CPU with multiple RT tasks queued, and a large number
2199 * of CPUs scheduling a lower priority task at the same time.
2201 * Each root domain has its own irq work function that can iterate over
2202 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
2203 * task must be checked if there's one or many CPUs that are lowering
2204 * their priority, there's a single irq work iterator that will try to
2205 * push off RT tasks that are waiting to run.
2207 * When a CPU schedules a lower priority task, it will kick off the
2208 * irq work iterator that will jump to each CPU with overloaded RT tasks.
2209 * As it only takes the first CPU that schedules a lower priority task
2210 * to start the process, the rto_start variable is incremented and if
2211 * the atomic result is one, then that CPU will try to take the rto_lock.
2212 * This prevents high contention on the lock as the process handles all
2213 * CPUs scheduling lower priority tasks.
2215 * All CPUs that are scheduling a lower priority task will increment the
2216 * rt_loop_next variable. This will make sure that the irq work iterator
2217 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
2218 * priority task, even if the iterator is in the middle of a scan. Incrementing
2219 * the rt_loop_next will cause the iterator to perform another scan.
2222 static int rto_next_cpu(struct root_domain *rd)
2228 * When starting the IPI RT pushing, the rto_cpu is set to -1,
2229 * rt_next_cpu() will simply return the first CPU found in
2232 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
2233 * will return the next CPU found in the rto_mask.
2235 * If there are no more CPUs left in the rto_mask, then a check is made
2236 * against rto_loop and rto_loop_next. rto_loop is only updated with
2237 * the rto_lock held, but any CPU may increment the rto_loop_next
2238 * without any locking.
2242 /* When rto_cpu is -1 this acts like cpumask_first() */
2243 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
2247 if (cpu < nr_cpu_ids)
2253 * ACQUIRE ensures we see the @rto_mask changes
2254 * made prior to the @next value observed.
2256 * Matches WMB in rt_set_overload().
2258 next = atomic_read_acquire(&rd->rto_loop_next);
2260 if (rd->rto_loop == next)
2263 rd->rto_loop = next;
2269 static inline bool rto_start_trylock(atomic_t *v)
2271 return !atomic_cmpxchg_acquire(v, 0, 1);
2274 static inline void rto_start_unlock(atomic_t *v)
2276 atomic_set_release(v, 0);
2279 static void tell_cpu_to_push(struct rq *rq)
2283 /* Keep the loop going if the IPI is currently active */
2284 atomic_inc(&rq->rd->rto_loop_next);
2286 /* Only one CPU can initiate a loop at a time */
2287 if (!rto_start_trylock(&rq->rd->rto_loop_start))
2290 raw_spin_lock(&rq->rd->rto_lock);
2293 * The rto_cpu is updated under the lock, if it has a valid CPU
2294 * then the IPI is still running and will continue due to the
2295 * update to loop_next, and nothing needs to be done here.
2296 * Otherwise it is finishing up and an ipi needs to be sent.
2298 if (rq->rd->rto_cpu < 0)
2299 cpu = rto_next_cpu(rq->rd);
2301 raw_spin_unlock(&rq->rd->rto_lock);
2303 rto_start_unlock(&rq->rd->rto_loop_start);
2306 /* Make sure the rd does not get freed while pushing */
2307 sched_get_rd(rq->rd);
2308 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2312 /* Called from hardirq context */
2313 void rto_push_irq_work_func(struct irq_work *work)
2315 struct root_domain *rd =
2316 container_of(work, struct root_domain, rto_push_work);
2323 * We do not need to grab the lock to check for has_pushable_tasks.
2324 * When it gets updated, a check is made if a push is possible.
2326 if (has_pushable_tasks(rq)) {
2327 raw_spin_rq_lock(rq);
2328 while (push_rt_task(rq, true))
2330 raw_spin_rq_unlock(rq);
2333 raw_spin_lock(&rd->rto_lock);
2335 /* Pass the IPI to the next rt overloaded queue */
2336 cpu = rto_next_cpu(rd);
2338 raw_spin_unlock(&rd->rto_lock);
2345 /* Try the next RT overloaded CPU */
2346 irq_work_queue_on(&rd->rto_push_work, cpu);
2348 #endif /* HAVE_RT_PUSH_IPI */
2350 static void pull_rt_task(struct rq *this_rq)
2352 int this_cpu = this_rq->cpu, cpu;
2353 bool resched = false;
2354 struct task_struct *p, *push_task;
2356 int rt_overload_count = rt_overloaded(this_rq);
2358 if (likely(!rt_overload_count))
2362 * Match the barrier from rt_set_overloaded; this guarantees that if we
2363 * see overloaded we must also see the rto_mask bit.
2367 /* If we are the only overloaded CPU do nothing */
2368 if (rt_overload_count == 1 &&
2369 cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2372 #ifdef HAVE_RT_PUSH_IPI
2373 if (sched_feat(RT_PUSH_IPI)) {
2374 tell_cpu_to_push(this_rq);
2379 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2380 if (this_cpu == cpu)
2383 src_rq = cpu_rq(cpu);
2386 * Don't bother taking the src_rq->lock if the next highest
2387 * task is known to be lower-priority than our current task.
2388 * This may look racy, but if this value is about to go
2389 * logically higher, the src_rq will push this task away.
2390 * And if its going logically lower, we do not care
2392 if (src_rq->rt.highest_prio.next >=
2393 this_rq->rt.highest_prio.curr)
2397 * We can potentially drop this_rq's lock in
2398 * double_lock_balance, and another CPU could
2402 double_lock_balance(this_rq, src_rq);
2405 * We can pull only a task, which is pushable
2406 * on its rq, and no others.
2408 p = pick_highest_pushable_task(src_rq, this_cpu);
2411 * Do we have an RT task that preempts
2412 * the to-be-scheduled task?
2414 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2415 WARN_ON(p == src_rq->curr);
2416 WARN_ON(!task_on_rq_queued(p));
2419 * There's a chance that p is higher in priority
2420 * than what's currently running on its CPU.
2421 * This is just that p is waking up and hasn't
2422 * had a chance to schedule. We only pull
2423 * p if it is lower in priority than the
2424 * current task on the run queue
2426 if (p->prio < src_rq->curr->prio)
2429 if (is_migration_disabled(p)) {
2430 push_task = get_push_task(src_rq);
2432 deactivate_task(src_rq, p, 0);
2433 set_task_cpu(p, this_cpu);
2434 activate_task(this_rq, p, 0);
2438 * We continue with the search, just in
2439 * case there's an even higher prio task
2440 * in another runqueue. (low likelihood
2445 double_unlock_balance(this_rq, src_rq);
2448 raw_spin_rq_unlock(this_rq);
2449 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2450 push_task, &src_rq->push_work);
2451 raw_spin_rq_lock(this_rq);
2456 resched_curr(this_rq);
2460 * If we are not running and we are not going to reschedule soon, we should
2461 * try to push tasks away now
2463 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2465 bool need_to_push = !task_running(rq, p) &&
2466 !test_tsk_need_resched(rq->curr) &&
2467 p->nr_cpus_allowed > 1 &&
2468 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2469 (rq->curr->nr_cpus_allowed < 2 ||
2470 rq->curr->prio <= p->prio);
2476 /* Assumes rq->lock is held */
2477 static void rq_online_rt(struct rq *rq)
2479 if (rq->rt.overloaded)
2480 rt_set_overload(rq);
2482 __enable_runtime(rq);
2484 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2487 /* Assumes rq->lock is held */
2488 static void rq_offline_rt(struct rq *rq)
2490 if (rq->rt.overloaded)
2491 rt_clear_overload(rq);
2493 __disable_runtime(rq);
2495 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2499 * When switch from the rt queue, we bring ourselves to a position
2500 * that we might want to pull RT tasks from other runqueues.
2502 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2505 * If there are other RT tasks then we will reschedule
2506 * and the scheduling of the other RT tasks will handle
2507 * the balancing. But if we are the last RT task
2508 * we may need to handle the pulling of RT tasks
2511 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2514 rt_queue_pull_task(rq);
2517 void __init init_sched_rt_class(void)
2521 for_each_possible_cpu(i) {
2522 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2523 GFP_KERNEL, cpu_to_node(i));
2526 #endif /* CONFIG_SMP */
2529 * When switching a task to RT, we may overload the runqueue
2530 * with RT tasks. In this case we try to push them off to
2533 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2536 * If we are running, update the avg_rt tracking, as the running time
2537 * will now on be accounted into the latter.
2539 if (task_current(rq, p)) {
2540 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2545 * If we are not running we may need to preempt the current
2546 * running task. If that current running task is also an RT task
2547 * then see if we can move to another run queue.
2549 if (task_on_rq_queued(p)) {
2551 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2552 rt_queue_push_tasks(rq);
2553 #endif /* CONFIG_SMP */
2554 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2560 * Priority of the task has changed. This may cause
2561 * us to initiate a push or pull.
2564 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2566 if (!task_on_rq_queued(p))
2569 if (task_current(rq, p)) {
2572 * If our priority decreases while running, we
2573 * may need to pull tasks to this runqueue.
2575 if (oldprio < p->prio)
2576 rt_queue_pull_task(rq);
2579 * If there's a higher priority task waiting to run
2582 if (p->prio > rq->rt.highest_prio.curr)
2585 /* For UP simply resched on drop of prio */
2586 if (oldprio < p->prio)
2588 #endif /* CONFIG_SMP */
2591 * This task is not running, but if it is
2592 * greater than the current running task
2595 if (p->prio < rq->curr->prio)
2600 #ifdef CONFIG_POSIX_TIMERS
2601 static void watchdog(struct rq *rq, struct task_struct *p)
2603 unsigned long soft, hard;
2605 /* max may change after cur was read, this will be fixed next tick */
2606 soft = task_rlimit(p, RLIMIT_RTTIME);
2607 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2609 if (soft != RLIM_INFINITY) {
2612 if (p->rt.watchdog_stamp != jiffies) {
2614 p->rt.watchdog_stamp = jiffies;
2617 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2618 if (p->rt.timeout > next) {
2619 posix_cputimers_rt_watchdog(&p->posix_cputimers,
2620 p->se.sum_exec_runtime);
2625 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2629 * scheduler tick hitting a task of our scheduling class.
2631 * NOTE: This function can be called remotely by the tick offload that
2632 * goes along full dynticks. Therefore no local assumption can be made
2633 * and everything must be accessed through the @rq and @curr passed in
2636 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2638 struct sched_rt_entity *rt_se = &p->rt;
2641 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2646 * RR tasks need a special form of timeslice management.
2647 * FIFO tasks have no timeslices.
2649 if (p->policy != SCHED_RR)
2652 if (--p->rt.time_slice)
2655 p->rt.time_slice = sched_rr_timeslice;
2658 * Requeue to the end of queue if we (and all of our ancestors) are not
2659 * the only element on the queue
2661 for_each_sched_rt_entity(rt_se) {
2662 if (rt_se->run_list.prev != rt_se->run_list.next) {
2663 requeue_task_rt(rq, p, 0);
2670 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2673 * Time slice is 0 for SCHED_FIFO tasks
2675 if (task->policy == SCHED_RR)
2676 return sched_rr_timeslice;
2681 DEFINE_SCHED_CLASS(rt) = {
2683 .enqueue_task = enqueue_task_rt,
2684 .dequeue_task = dequeue_task_rt,
2685 .yield_task = yield_task_rt,
2687 .check_preempt_curr = check_preempt_curr_rt,
2689 .pick_next_task = pick_next_task_rt,
2690 .put_prev_task = put_prev_task_rt,
2691 .set_next_task = set_next_task_rt,
2694 .balance = balance_rt,
2695 .pick_task = pick_task_rt,
2696 .select_task_rq = select_task_rq_rt,
2697 .set_cpus_allowed = set_cpus_allowed_common,
2698 .rq_online = rq_online_rt,
2699 .rq_offline = rq_offline_rt,
2700 .task_woken = task_woken_rt,
2701 .switched_from = switched_from_rt,
2702 .find_lock_rq = find_lock_lowest_rq,
2705 .task_tick = task_tick_rt,
2707 .get_rr_interval = get_rr_interval_rt,
2709 .prio_changed = prio_changed_rt,
2710 .switched_to = switched_to_rt,
2712 .update_curr = update_curr_rt,
2714 #ifdef CONFIG_UCLAMP_TASK
2715 .uclamp_enabled = 1,
2719 #ifdef CONFIG_RT_GROUP_SCHED
2721 * Ensure that the real time constraints are schedulable.
2723 static DEFINE_MUTEX(rt_constraints_mutex);
2725 static inline int tg_has_rt_tasks(struct task_group *tg)
2727 struct task_struct *task;
2728 struct css_task_iter it;
2732 * Autogroups do not have RT tasks; see autogroup_create().
2734 if (task_group_is_autogroup(tg))
2737 css_task_iter_start(&tg->css, 0, &it);
2738 while (!ret && (task = css_task_iter_next(&it)))
2739 ret |= rt_task(task);
2740 css_task_iter_end(&it);
2745 struct rt_schedulable_data {
2746 struct task_group *tg;
2751 static int tg_rt_schedulable(struct task_group *tg, void *data)
2753 struct rt_schedulable_data *d = data;
2754 struct task_group *child;
2755 unsigned long total, sum = 0;
2756 u64 period, runtime;
2758 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2759 runtime = tg->rt_bandwidth.rt_runtime;
2762 period = d->rt_period;
2763 runtime = d->rt_runtime;
2767 * Cannot have more runtime than the period.
2769 if (runtime > period && runtime != RUNTIME_INF)
2773 * Ensure we don't starve existing RT tasks if runtime turns zero.
2775 if (rt_bandwidth_enabled() && !runtime &&
2776 tg->rt_bandwidth.rt_runtime && tg_has_rt_tasks(tg))
2779 total = to_ratio(period, runtime);
2782 * Nobody can have more than the global setting allows.
2784 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2788 * The sum of our children's runtime should not exceed our own.
2790 list_for_each_entry_rcu(child, &tg->children, siblings) {
2791 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2792 runtime = child->rt_bandwidth.rt_runtime;
2794 if (child == d->tg) {
2795 period = d->rt_period;
2796 runtime = d->rt_runtime;
2799 sum += to_ratio(period, runtime);
2808 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2812 struct rt_schedulable_data data = {
2814 .rt_period = period,
2815 .rt_runtime = runtime,
2819 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2825 static int tg_set_rt_bandwidth(struct task_group *tg,
2826 u64 rt_period, u64 rt_runtime)
2831 * Disallowing the root group RT runtime is BAD, it would disallow the
2832 * kernel creating (and or operating) RT threads.
2834 if (tg == &root_task_group && rt_runtime == 0)
2837 /* No period doesn't make any sense. */
2842 * Bound quota to defend quota against overflow during bandwidth shift.
2844 if (rt_runtime != RUNTIME_INF && rt_runtime > max_rt_runtime)
2847 mutex_lock(&rt_constraints_mutex);
2848 err = __rt_schedulable(tg, rt_period, rt_runtime);
2852 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2853 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2854 tg->rt_bandwidth.rt_runtime = rt_runtime;
2856 for_each_possible_cpu(i) {
2857 struct rt_rq *rt_rq = tg->rt_rq[i];
2859 raw_spin_lock(&rt_rq->rt_runtime_lock);
2860 rt_rq->rt_runtime = rt_runtime;
2861 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2863 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2865 mutex_unlock(&rt_constraints_mutex);
2870 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2872 u64 rt_runtime, rt_period;
2874 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2875 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2876 if (rt_runtime_us < 0)
2877 rt_runtime = RUNTIME_INF;
2878 else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC)
2881 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2884 long sched_group_rt_runtime(struct task_group *tg)
2888 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2891 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2892 do_div(rt_runtime_us, NSEC_PER_USEC);
2893 return rt_runtime_us;
2896 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2898 u64 rt_runtime, rt_period;
2900 if (rt_period_us > U64_MAX / NSEC_PER_USEC)
2903 rt_period = rt_period_us * NSEC_PER_USEC;
2904 rt_runtime = tg->rt_bandwidth.rt_runtime;
2906 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2909 long sched_group_rt_period(struct task_group *tg)
2913 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2914 do_div(rt_period_us, NSEC_PER_USEC);
2915 return rt_period_us;
2918 #ifdef CONFIG_SYSCTL
2919 static int sched_rt_global_constraints(void)
2923 mutex_lock(&rt_constraints_mutex);
2924 ret = __rt_schedulable(NULL, 0, 0);
2925 mutex_unlock(&rt_constraints_mutex);
2929 #endif /* CONFIG_SYSCTL */
2931 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2933 /* Don't accept realtime tasks when there is no way for them to run */
2934 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2940 #else /* !CONFIG_RT_GROUP_SCHED */
2942 #ifdef CONFIG_SYSCTL
2943 static int sched_rt_global_constraints(void)
2945 unsigned long flags;
2948 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2949 for_each_possible_cpu(i) {
2950 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2952 raw_spin_lock(&rt_rq->rt_runtime_lock);
2953 rt_rq->rt_runtime = global_rt_runtime();
2954 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2956 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2960 #endif /* CONFIG_SYSCTL */
2961 #endif /* CONFIG_RT_GROUP_SCHED */
2963 #ifdef CONFIG_SYSCTL
2964 static int sched_rt_global_validate(void)
2966 if (sysctl_sched_rt_period <= 0)
2969 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2970 ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
2971 ((u64)sysctl_sched_rt_runtime *
2972 NSEC_PER_USEC > max_rt_runtime)))
2978 static void sched_rt_do_global(void)
2980 unsigned long flags;
2982 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2983 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2984 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2985 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2988 static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
2989 size_t *lenp, loff_t *ppos)
2991 int old_period, old_runtime;
2992 static DEFINE_MUTEX(mutex);
2996 old_period = sysctl_sched_rt_period;
2997 old_runtime = sysctl_sched_rt_runtime;
2999 ret = proc_dointvec(table, write, buffer, lenp, ppos);
3001 if (!ret && write) {
3002 ret = sched_rt_global_validate();
3006 ret = sched_dl_global_validate();
3010 ret = sched_rt_global_constraints();
3014 sched_rt_do_global();
3015 sched_dl_do_global();
3019 sysctl_sched_rt_period = old_period;
3020 sysctl_sched_rt_runtime = old_runtime;
3022 mutex_unlock(&mutex);
3027 static int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
3028 size_t *lenp, loff_t *ppos)
3031 static DEFINE_MUTEX(mutex);
3034 ret = proc_dointvec(table, write, buffer, lenp, ppos);
3036 * Make sure that internally we keep jiffies.
3037 * Also, writing zero resets the timeslice to default:
3039 if (!ret && write) {
3040 sched_rr_timeslice =
3041 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
3042 msecs_to_jiffies(sysctl_sched_rr_timeslice);
3044 mutex_unlock(&mutex);
3048 #endif /* CONFIG_SYSCTL */
3050 #ifdef CONFIG_SCHED_DEBUG
3051 void print_rt_stats(struct seq_file *m, int cpu)
3054 struct rt_rq *rt_rq;
3057 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
3058 print_rt_rq(m, cpu, rt_rq);
3061 #endif /* CONFIG_SCHED_DEBUG */