Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
4 * policies)
5 */
6
7int sched_rr_timeslice = RR_TIMESLICE;
8/* More than 4 hours if BW_SHIFT equals 20. */
9static const u64 max_rt_runtime = MAX_BW;
10
11static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
12
13struct rt_bandwidth def_rt_bandwidth;
14
15/*
16 * period over which we measure -rt task CPU usage in us.
17 * default: 1s
18 */
19unsigned int sysctl_sched_rt_period = 1000000;
20
21/*
22 * part of the period that we allow rt tasks to run in us.
23 * default: 0.95s
24 */
25int sysctl_sched_rt_runtime = 950000;
26
27#ifdef CONFIG_SYSCTL
28static int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
29static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
30 size_t *lenp, loff_t *ppos);
31static int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
32 size_t *lenp, loff_t *ppos);
33static struct ctl_table sched_rt_sysctls[] = {
34 {
35 .procname = "sched_rt_period_us",
36 .data = &sysctl_sched_rt_period,
37 .maxlen = sizeof(unsigned int),
38 .mode = 0644,
39 .proc_handler = sched_rt_handler,
40 },
41 {
42 .procname = "sched_rt_runtime_us",
43 .data = &sysctl_sched_rt_runtime,
44 .maxlen = sizeof(int),
45 .mode = 0644,
46 .proc_handler = sched_rt_handler,
47 },
48 {
49 .procname = "sched_rr_timeslice_ms",
50 .data = &sysctl_sched_rr_timeslice,
51 .maxlen = sizeof(int),
52 .mode = 0644,
53 .proc_handler = sched_rr_handler,
54 },
55 {}
56};
57
58static int __init sched_rt_sysctl_init(void)
59{
60 register_sysctl_init("kernel", sched_rt_sysctls);
61 return 0;
62}
63late_initcall(sched_rt_sysctl_init);
64#endif
65
66static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
67{
68 struct rt_bandwidth *rt_b =
69 container_of(timer, struct rt_bandwidth, rt_period_timer);
70 int idle = 0;
71 int overrun;
72
73 raw_spin_lock(&rt_b->rt_runtime_lock);
74 for (;;) {
75 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
76 if (!overrun)
77 break;
78
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);
82 }
83 if (idle)
84 rt_b->rt_period_active = 0;
85 raw_spin_unlock(&rt_b->rt_runtime_lock);
86
87 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
88}
89
90void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
91{
92 rt_b->rt_period = ns_to_ktime(period);
93 rt_b->rt_runtime = runtime;
94
95 raw_spin_lock_init(&rt_b->rt_runtime_lock);
96
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;
100}
101
102static inline void do_start_rt_bandwidth(struct rt_bandwidth *rt_b)
103{
104 raw_spin_lock(&rt_b->rt_runtime_lock);
105 if (!rt_b->rt_period_active) {
106 rt_b->rt_period_active = 1;
107 /*
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.
114 */
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);
118 }
119 raw_spin_unlock(&rt_b->rt_runtime_lock);
120}
121
122static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
123{
124 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
125 return;
126
127 do_start_rt_bandwidth(rt_b);
128}
129
130void init_rt_rq(struct rt_rq *rt_rq)
131{
132 struct rt_prio_array *array;
133 int i;
134
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);
139 }
140 /* delimiter for bitsearch: */
141 __set_bit(MAX_RT_PRIO, array->bitmap);
142
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;
152
153 rt_rq->rt_time = 0;
154 rt_rq->rt_throttled = 0;
155 rt_rq->rt_runtime = 0;
156 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
157}
158
159#ifdef CONFIG_RT_GROUP_SCHED
160static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
161{
162 hrtimer_cancel(&rt_b->rt_period_timer);
163}
164
165#define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
166
167static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
168{
169#ifdef CONFIG_SCHED_DEBUG
170 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
171#endif
172 return container_of(rt_se, struct task_struct, rt);
173}
174
175static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
176{
177 return rt_rq->rq;
178}
179
180static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
181{
182 return rt_se->rt_rq;
183}
184
185static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
186{
187 struct rt_rq *rt_rq = rt_se->rt_rq;
188
189 return rt_rq->rq;
190}
191
192void unregister_rt_sched_group(struct task_group *tg)
193{
194 if (tg->rt_se)
195 destroy_rt_bandwidth(&tg->rt_bandwidth);
196
197}
198
199void free_rt_sched_group(struct task_group *tg)
200{
201 int i;
202
203 for_each_possible_cpu(i) {
204 if (tg->rt_rq)
205 kfree(tg->rt_rq[i]);
206 if (tg->rt_se)
207 kfree(tg->rt_se[i]);
208 }
209
210 kfree(tg->rt_rq);
211 kfree(tg->rt_se);
212}
213
214void 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)
217{
218 struct rq *rq = cpu_rq(cpu);
219
220 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
221 rt_rq->rt_nr_boosted = 0;
222 rt_rq->rq = rq;
223 rt_rq->tg = tg;
224
225 tg->rt_rq[cpu] = rt_rq;
226 tg->rt_se[cpu] = rt_se;
227
228 if (!rt_se)
229 return;
230
231 if (!parent)
232 rt_se->rt_rq = &rq->rt;
233 else
234 rt_se->rt_rq = parent->my_q;
235
236 rt_se->my_q = rt_rq;
237 rt_se->parent = parent;
238 INIT_LIST_HEAD(&rt_se->run_list);
239}
240
241int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
242{
243 struct rt_rq *rt_rq;
244 struct sched_rt_entity *rt_se;
245 int i;
246
247 tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL);
248 if (!tg->rt_rq)
249 goto err;
250 tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL);
251 if (!tg->rt_se)
252 goto err;
253
254 init_rt_bandwidth(&tg->rt_bandwidth,
255 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
256
257 for_each_possible_cpu(i) {
258 rt_rq = kzalloc_node(sizeof(struct rt_rq),
259 GFP_KERNEL, cpu_to_node(i));
260 if (!rt_rq)
261 goto err;
262
263 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
264 GFP_KERNEL, cpu_to_node(i));
265 if (!rt_se)
266 goto err_free_rq;
267
268 init_rt_rq(rt_rq);
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]);
271 }
272
273 return 1;
274
275err_free_rq:
276 kfree(rt_rq);
277err:
278 return 0;
279}
280
281#else /* CONFIG_RT_GROUP_SCHED */
282
283#define rt_entity_is_task(rt_se) (1)
284
285static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
286{
287 return container_of(rt_se, struct task_struct, rt);
288}
289
290static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
291{
292 return container_of(rt_rq, struct rq, rt);
293}
294
295static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
296{
297 struct task_struct *p = rt_task_of(rt_se);
298
299 return task_rq(p);
300}
301
302static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
303{
304 struct rq *rq = rq_of_rt_se(rt_se);
305
306 return &rq->rt;
307}
308
309void unregister_rt_sched_group(struct task_group *tg) { }
310
311void free_rt_sched_group(struct task_group *tg) { }
312
313int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
314{
315 return 1;
316}
317#endif /* CONFIG_RT_GROUP_SCHED */
318
319#ifdef CONFIG_SMP
320
321static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
322{
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;
325}
326
327static inline int rt_overloaded(struct rq *rq)
328{
329 return atomic_read(&rq->rd->rto_count);
330}
331
332static inline void rt_set_overload(struct rq *rq)
333{
334 if (!rq->online)
335 return;
336
337 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
338 /*
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
343 * updated yet.
344 *
345 * Matched by the barrier in pull_rt_task().
346 */
347 smp_wmb();
348 atomic_inc(&rq->rd->rto_count);
349}
350
351static inline void rt_clear_overload(struct rq *rq)
352{
353 if (!rq->online)
354 return;
355
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);
359}
360
361static void update_rt_migration(struct rt_rq *rt_rq)
362{
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;
367 }
368 } else if (rt_rq->overloaded) {
369 rt_clear_overload(rq_of_rt_rq(rt_rq));
370 rt_rq->overloaded = 0;
371 }
372}
373
374static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
375{
376 struct task_struct *p;
377
378 if (!rt_entity_is_task(rt_se))
379 return;
380
381 p = rt_task_of(rt_se);
382 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
383
384 rt_rq->rt_nr_total++;
385 if (p->nr_cpus_allowed > 1)
386 rt_rq->rt_nr_migratory++;
387
388 update_rt_migration(rt_rq);
389}
390
391static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
392{
393 struct task_struct *p;
394
395 if (!rt_entity_is_task(rt_se))
396 return;
397
398 p = rt_task_of(rt_se);
399 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
400
401 rt_rq->rt_nr_total--;
402 if (p->nr_cpus_allowed > 1)
403 rt_rq->rt_nr_migratory--;
404
405 update_rt_migration(rt_rq);
406}
407
408static inline int has_pushable_tasks(struct rq *rq)
409{
410 return !plist_head_empty(&rq->rt.pushable_tasks);
411}
412
413static DEFINE_PER_CPU(struct balance_callback, rt_push_head);
414static DEFINE_PER_CPU(struct balance_callback, rt_pull_head);
415
416static void push_rt_tasks(struct rq *);
417static void pull_rt_task(struct rq *);
418
419static inline void rt_queue_push_tasks(struct rq *rq)
420{
421 if (!has_pushable_tasks(rq))
422 return;
423
424 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
425}
426
427static inline void rt_queue_pull_task(struct rq *rq)
428{
429 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
430}
431
432static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
433{
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);
437
438 /* Update the highest prio pushable task */
439 if (p->prio < rq->rt.highest_prio.next)
440 rq->rt.highest_prio.next = p->prio;
441}
442
443static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
444{
445 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
446
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;
452 } else {
453 rq->rt.highest_prio.next = MAX_RT_PRIO-1;
454 }
455}
456
457#else
458
459static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
460{
461}
462
463static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
464{
465}
466
467static inline
468void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
469{
470}
471
472static inline
473void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
474{
475}
476
477static inline void rt_queue_push_tasks(struct rq *rq)
478{
479}
480#endif /* CONFIG_SMP */
481
482static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
483static void dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count);
484
485static inline int on_rt_rq(struct sched_rt_entity *rt_se)
486{
487 return rt_se->on_rq;
488}
489
490#ifdef CONFIG_UCLAMP_TASK
491/*
492 * Verify the fitness of task @p to run on @cpu taking into account the uclamp
493 * settings.
494 *
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.
498 *
499 * The function will return true if the capacity of the @cpu is >= the
500 * uclamp_min and false otherwise.
501 *
502 * Note that uclamp_min will be clamped to uclamp_max if uclamp_min
503 * > uclamp_max.
504 */
505static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
506{
507 unsigned int min_cap;
508 unsigned int max_cap;
509 unsigned int cpu_cap;
510
511 /* Only heterogeneous systems can benefit from this check */
512 if (!sched_asym_cpucap_active())
513 return true;
514
515 min_cap = uclamp_eff_value(p, UCLAMP_MIN);
516 max_cap = uclamp_eff_value(p, UCLAMP_MAX);
517
518 cpu_cap = capacity_orig_of(cpu);
519
520 return cpu_cap >= min(min_cap, max_cap);
521}
522#else
523static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
524{
525 return true;
526}
527#endif
528
529#ifdef CONFIG_RT_GROUP_SCHED
530
531static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
532{
533 if (!rt_rq->tg)
534 return RUNTIME_INF;
535
536 return rt_rq->rt_runtime;
537}
538
539static inline u64 sched_rt_period(struct rt_rq *rt_rq)
540{
541 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
542}
543
544typedef struct task_group *rt_rq_iter_t;
545
546static inline struct task_group *next_task_group(struct task_group *tg)
547{
548 do {
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));
552
553 if (&tg->list == &task_groups)
554 tg = NULL;
555
556 return tg;
557}
558
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)]);)
563
564#define for_each_sched_rt_entity(rt_se) \
565 for (; rt_se; rt_se = rt_se->parent)
566
567static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
568{
569 return rt_se->my_q;
570}
571
572static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
573static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
574
575static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
576{
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;
580
581 int cpu = cpu_of(rq);
582
583 rt_se = rt_rq->tg->rt_se[cpu];
584
585 if (rt_rq->rt_nr_running) {
586 if (!rt_se)
587 enqueue_top_rt_rq(rt_rq);
588 else if (!on_rt_rq(rt_se))
589 enqueue_rt_entity(rt_se, 0);
590
591 if (rt_rq->highest_prio.curr < curr->prio)
592 resched_curr(rq);
593 }
594}
595
596static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
597{
598 struct sched_rt_entity *rt_se;
599 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
600
601 rt_se = rt_rq->tg->rt_se[cpu];
602
603 if (!rt_se) {
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);
607 }
608 else if (on_rt_rq(rt_se))
609 dequeue_rt_entity(rt_se, 0);
610}
611
612static inline int rt_rq_throttled(struct rt_rq *rt_rq)
613{
614 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
615}
616
617static int rt_se_boosted(struct sched_rt_entity *rt_se)
618{
619 struct rt_rq *rt_rq = group_rt_rq(rt_se);
620 struct task_struct *p;
621
622 if (rt_rq)
623 return !!rt_rq->rt_nr_boosted;
624
625 p = rt_task_of(rt_se);
626 return p->prio != p->normal_prio;
627}
628
629#ifdef CONFIG_SMP
630static inline const struct cpumask *sched_rt_period_mask(void)
631{
632 return this_rq()->rd->span;
633}
634#else
635static inline const struct cpumask *sched_rt_period_mask(void)
636{
637 return cpu_online_mask;
638}
639#endif
640
641static inline
642struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
643{
644 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
645}
646
647static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
648{
649 return &rt_rq->tg->rt_bandwidth;
650}
651
652#else /* !CONFIG_RT_GROUP_SCHED */
653
654static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
655{
656 return rt_rq->rt_runtime;
657}
658
659static inline u64 sched_rt_period(struct rt_rq *rt_rq)
660{
661 return ktime_to_ns(def_rt_bandwidth.rt_period);
662}
663
664typedef struct rt_rq *rt_rq_iter_t;
665
666#define for_each_rt_rq(rt_rq, iter, rq) \
667 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
668
669#define for_each_sched_rt_entity(rt_se) \
670 for (; rt_se; rt_se = NULL)
671
672static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
673{
674 return NULL;
675}
676
677static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
678{
679 struct rq *rq = rq_of_rt_rq(rt_rq);
680
681 if (!rt_rq->rt_nr_running)
682 return;
683
684 enqueue_top_rt_rq(rt_rq);
685 resched_curr(rq);
686}
687
688static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
689{
690 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
691}
692
693static inline int rt_rq_throttled(struct rt_rq *rt_rq)
694{
695 return rt_rq->rt_throttled;
696}
697
698static inline const struct cpumask *sched_rt_period_mask(void)
699{
700 return cpu_online_mask;
701}
702
703static inline
704struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
705{
706 return &cpu_rq(cpu)->rt;
707}
708
709static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
710{
711 return &def_rt_bandwidth;
712}
713
714#endif /* CONFIG_RT_GROUP_SCHED */
715
716bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
717{
718 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
719
720 return (hrtimer_active(&rt_b->rt_period_timer) ||
721 rt_rq->rt_time < rt_b->rt_runtime);
722}
723
724#ifdef CONFIG_SMP
725/*
726 * We ran out of runtime, see if we can borrow some from our neighbours.
727 */
728static void do_balance_runtime(struct rt_rq *rt_rq)
729{
730 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
731 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
732 int i, weight;
733 u64 rt_period;
734
735 weight = cpumask_weight(rd->span);
736
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);
741 s64 diff;
742
743 if (iter == rt_rq)
744 continue;
745
746 raw_spin_lock(&iter->rt_runtime_lock);
747 /*
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.
751 */
752 if (iter->rt_runtime == RUNTIME_INF)
753 goto next;
754
755 /*
756 * From runqueues with spare time, take 1/n part of their
757 * spare time, but no more than our period.
758 */
759 diff = iter->rt_runtime - iter->rt_time;
760 if (diff > 0) {
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);
768 break;
769 }
770 }
771next:
772 raw_spin_unlock(&iter->rt_runtime_lock);
773 }
774 raw_spin_unlock(&rt_b->rt_runtime_lock);
775}
776
777/*
778 * Ensure this RQ takes back all the runtime it lend to its neighbours.
779 */
780static void __disable_runtime(struct rq *rq)
781{
782 struct root_domain *rd = rq->rd;
783 rt_rq_iter_t iter;
784 struct rt_rq *rt_rq;
785
786 if (unlikely(!scheduler_running))
787 return;
788
789 for_each_rt_rq(rt_rq, iter, rq) {
790 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
791 s64 want;
792 int i;
793
794 raw_spin_lock(&rt_b->rt_runtime_lock);
795 raw_spin_lock(&rt_rq->rt_runtime_lock);
796 /*
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.
800 */
801 if (rt_rq->rt_runtime == RUNTIME_INF ||
802 rt_rq->rt_runtime == rt_b->rt_runtime)
803 goto balanced;
804 raw_spin_unlock(&rt_rq->rt_runtime_lock);
805
806 /*
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.
810 */
811 want = rt_b->rt_runtime - rt_rq->rt_runtime;
812
813 /*
814 * Greedy reclaim, take back as much as we can.
815 */
816 for_each_cpu(i, rd->span) {
817 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
818 s64 diff;
819
820 /*
821 * Can't reclaim from ourselves or disabled runqueues.
822 */
823 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
824 continue;
825
826 raw_spin_lock(&iter->rt_runtime_lock);
827 if (want > 0) {
828 diff = min_t(s64, iter->rt_runtime, want);
829 iter->rt_runtime -= diff;
830 want -= diff;
831 } else {
832 iter->rt_runtime -= want;
833 want -= want;
834 }
835 raw_spin_unlock(&iter->rt_runtime_lock);
836
837 if (!want)
838 break;
839 }
840
841 raw_spin_lock(&rt_rq->rt_runtime_lock);
842 /*
843 * We cannot be left wanting - that would mean some runtime
844 * leaked out of the system.
845 */
846 WARN_ON_ONCE(want);
847balanced:
848 /*
849 * Disable all the borrow logic by pretending we have inf
850 * runtime - in which case borrowing doesn't make sense.
851 */
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);
856
857 /* Make rt_rq available for pick_next_task() */
858 sched_rt_rq_enqueue(rt_rq);
859 }
860}
861
862static void __enable_runtime(struct rq *rq)
863{
864 rt_rq_iter_t iter;
865 struct rt_rq *rt_rq;
866
867 if (unlikely(!scheduler_running))
868 return;
869
870 /*
871 * Reset each runqueue's bandwidth settings
872 */
873 for_each_rt_rq(rt_rq, iter, rq) {
874 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
875
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;
879 rt_rq->rt_time = 0;
880 rt_rq->rt_throttled = 0;
881 raw_spin_unlock(&rt_rq->rt_runtime_lock);
882 raw_spin_unlock(&rt_b->rt_runtime_lock);
883 }
884}
885
886static void balance_runtime(struct rt_rq *rt_rq)
887{
888 if (!sched_feat(RT_RUNTIME_SHARE))
889 return;
890
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);
895 }
896}
897#else /* !CONFIG_SMP */
898static inline void balance_runtime(struct rt_rq *rt_rq) {}
899#endif /* CONFIG_SMP */
900
901static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
902{
903 int i, idle = 1, throttled = 0;
904 const struct cpumask *span;
905
906 span = sched_rt_period_mask();
907#ifdef CONFIG_RT_GROUP_SCHED
908 /*
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.
916 */
917 if (rt_b == &root_task_group.rt_bandwidth)
918 span = cpu_online_mask;
919#endif
920 for_each_cpu(i, span) {
921 int enqueue = 0;
922 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
923 struct rq *rq = rq_of_rt_rq(rt_rq);
924 struct rq_flags rf;
925 int skip;
926
927 /*
928 * When span == cpu_online_mask, taking each rq->lock
929 * can be time-consuming. Try to avoid it when possible.
930 */
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);
936 if (skip)
937 continue;
938
939 rq_lock(rq, &rf);
940 update_rq_clock(rq);
941
942 if (rt_rq->rt_time) {
943 u64 runtime;
944
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;
952 enqueue = 1;
953
954 /*
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
959 * 'runtime'.
960 */
961 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
962 rq_clock_cancel_skipupdate(rq);
963 }
964 if (rt_rq->rt_time || rt_rq->rt_nr_running)
965 idle = 0;
966 raw_spin_unlock(&rt_rq->rt_runtime_lock);
967 } else if (rt_rq->rt_nr_running) {
968 idle = 0;
969 if (!rt_rq_throttled(rt_rq))
970 enqueue = 1;
971 }
972 if (rt_rq->rt_throttled)
973 throttled = 1;
974
975 if (enqueue)
976 sched_rt_rq_enqueue(rt_rq);
977 rq_unlock(rq, &rf);
978 }
979
980 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
981 return 1;
982
983 return idle;
984}
985
986static inline int rt_se_prio(struct sched_rt_entity *rt_se)
987{
988#ifdef CONFIG_RT_GROUP_SCHED
989 struct rt_rq *rt_rq = group_rt_rq(rt_se);
990
991 if (rt_rq)
992 return rt_rq->highest_prio.curr;
993#endif
994
995 return rt_task_of(rt_se)->prio;
996}
997
998static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
999{
1000 u64 runtime = sched_rt_runtime(rt_rq);
1001
1002 if (rt_rq->rt_throttled)
1003 return rt_rq_throttled(rt_rq);
1004
1005 if (runtime >= sched_rt_period(rt_rq))
1006 return 0;
1007
1008 balance_runtime(rt_rq);
1009 runtime = sched_rt_runtime(rt_rq);
1010 if (runtime == RUNTIME_INF)
1011 return 0;
1012
1013 if (rt_rq->rt_time > runtime) {
1014 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
1015
1016 /*
1017 * Don't actually throttle groups that have no runtime assigned
1018 * but accrue some time due to boosting.
1019 */
1020 if (likely(rt_b->rt_runtime)) {
1021 rt_rq->rt_throttled = 1;
1022 printk_deferred_once("sched: RT throttling activated\n");
1023 } else {
1024 /*
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.
1028 */
1029 rt_rq->rt_time = 0;
1030 }
1031
1032 if (rt_rq_throttled(rt_rq)) {
1033 sched_rt_rq_dequeue(rt_rq);
1034 return 1;
1035 }
1036 }
1037
1038 return 0;
1039}
1040
1041/*
1042 * Update the current task's runtime statistics. Skip current tasks that
1043 * are not in our scheduling class.
1044 */
1045static void update_curr_rt(struct rq *rq)
1046{
1047 struct task_struct *curr = rq->curr;
1048 struct sched_rt_entity *rt_se = &curr->rt;
1049 u64 delta_exec;
1050 u64 now;
1051
1052 if (curr->sched_class != &rt_sched_class)
1053 return;
1054
1055 now = rq_clock_task(rq);
1056 delta_exec = now - curr->se.exec_start;
1057 if (unlikely((s64)delta_exec <= 0))
1058 return;
1059
1060 schedstat_set(curr->stats.exec_max,
1061 max(curr->stats.exec_max, delta_exec));
1062
1063 trace_sched_stat_runtime(curr, delta_exec, 0);
1064
1065 update_current_exec_runtime(curr, now, delta_exec);
1066
1067 if (!rt_bandwidth_enabled())
1068 return;
1069
1070 for_each_sched_rt_entity(rt_se) {
1071 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1072 int exceeded;
1073
1074 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
1075 raw_spin_lock(&rt_rq->rt_runtime_lock);
1076 rt_rq->rt_time += delta_exec;
1077 exceeded = sched_rt_runtime_exceeded(rt_rq);
1078 if (exceeded)
1079 resched_curr(rq);
1080 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1081 if (exceeded)
1082 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq));
1083 }
1084 }
1085}
1086
1087static void
1088dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count)
1089{
1090 struct rq *rq = rq_of_rt_rq(rt_rq);
1091
1092 BUG_ON(&rq->rt != rt_rq);
1093
1094 if (!rt_rq->rt_queued)
1095 return;
1096
1097 BUG_ON(!rq->nr_running);
1098
1099 sub_nr_running(rq, count);
1100 rt_rq->rt_queued = 0;
1101
1102}
1103
1104static void
1105enqueue_top_rt_rq(struct rt_rq *rt_rq)
1106{
1107 struct rq *rq = rq_of_rt_rq(rt_rq);
1108
1109 BUG_ON(&rq->rt != rt_rq);
1110
1111 if (rt_rq->rt_queued)
1112 return;
1113
1114 if (rt_rq_throttled(rt_rq))
1115 return;
1116
1117 if (rt_rq->rt_nr_running) {
1118 add_nr_running(rq, rt_rq->rt_nr_running);
1119 rt_rq->rt_queued = 1;
1120 }
1121
1122 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1123 cpufreq_update_util(rq, 0);
1124}
1125
1126#if defined CONFIG_SMP
1127
1128static void
1129inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1130{
1131 struct rq *rq = rq_of_rt_rq(rt_rq);
1132
1133#ifdef CONFIG_RT_GROUP_SCHED
1134 /*
1135 * Change rq's cpupri only if rt_rq is the top queue.
1136 */
1137 if (&rq->rt != rt_rq)
1138 return;
1139#endif
1140 if (rq->online && prio < prev_prio)
1141 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1142}
1143
1144static void
1145dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1146{
1147 struct rq *rq = rq_of_rt_rq(rt_rq);
1148
1149#ifdef CONFIG_RT_GROUP_SCHED
1150 /*
1151 * Change rq's cpupri only if rt_rq is the top queue.
1152 */
1153 if (&rq->rt != rt_rq)
1154 return;
1155#endif
1156 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1157 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1158}
1159
1160#else /* CONFIG_SMP */
1161
1162static inline
1163void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1164static inline
1165void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1166
1167#endif /* CONFIG_SMP */
1168
1169#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1170static void
1171inc_rt_prio(struct rt_rq *rt_rq, int prio)
1172{
1173 int prev_prio = rt_rq->highest_prio.curr;
1174
1175 if (prio < prev_prio)
1176 rt_rq->highest_prio.curr = prio;
1177
1178 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1179}
1180
1181static void
1182dec_rt_prio(struct rt_rq *rt_rq, int prio)
1183{
1184 int prev_prio = rt_rq->highest_prio.curr;
1185
1186 if (rt_rq->rt_nr_running) {
1187
1188 WARN_ON(prio < prev_prio);
1189
1190 /*
1191 * This may have been our highest task, and therefore
1192 * we may have some recomputation to do
1193 */
1194 if (prio == prev_prio) {
1195 struct rt_prio_array *array = &rt_rq->active;
1196
1197 rt_rq->highest_prio.curr =
1198 sched_find_first_bit(array->bitmap);
1199 }
1200
1201 } else {
1202 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
1203 }
1204
1205 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1206}
1207
1208#else
1209
1210static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1211static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1212
1213#endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1214
1215#ifdef CONFIG_RT_GROUP_SCHED
1216
1217static void
1218inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1219{
1220 if (rt_se_boosted(rt_se))
1221 rt_rq->rt_nr_boosted++;
1222
1223 if (rt_rq->tg)
1224 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1225}
1226
1227static void
1228dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1229{
1230 if (rt_se_boosted(rt_se))
1231 rt_rq->rt_nr_boosted--;
1232
1233 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1234}
1235
1236#else /* CONFIG_RT_GROUP_SCHED */
1237
1238static void
1239inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1240{
1241 start_rt_bandwidth(&def_rt_bandwidth);
1242}
1243
1244static inline
1245void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1246
1247#endif /* CONFIG_RT_GROUP_SCHED */
1248
1249static inline
1250unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1251{
1252 struct rt_rq *group_rq = group_rt_rq(rt_se);
1253
1254 if (group_rq)
1255 return group_rq->rt_nr_running;
1256 else
1257 return 1;
1258}
1259
1260static inline
1261unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1262{
1263 struct rt_rq *group_rq = group_rt_rq(rt_se);
1264 struct task_struct *tsk;
1265
1266 if (group_rq)
1267 return group_rq->rr_nr_running;
1268
1269 tsk = rt_task_of(rt_se);
1270
1271 return (tsk->policy == SCHED_RR) ? 1 : 0;
1272}
1273
1274static inline
1275void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1276{
1277 int prio = rt_se_prio(rt_se);
1278
1279 WARN_ON(!rt_prio(prio));
1280 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1281 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1282
1283 inc_rt_prio(rt_rq, prio);
1284 inc_rt_migration(rt_se, rt_rq);
1285 inc_rt_group(rt_se, rt_rq);
1286}
1287
1288static inline
1289void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1290{
1291 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1292 WARN_ON(!rt_rq->rt_nr_running);
1293 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1294 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1295
1296 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1297 dec_rt_migration(rt_se, rt_rq);
1298 dec_rt_group(rt_se, rt_rq);
1299}
1300
1301/*
1302 * Change rt_se->run_list location unless SAVE && !MOVE
1303 *
1304 * assumes ENQUEUE/DEQUEUE flags match
1305 */
1306static inline bool move_entity(unsigned int flags)
1307{
1308 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1309 return false;
1310
1311 return true;
1312}
1313
1314static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1315{
1316 list_del_init(&rt_se->run_list);
1317
1318 if (list_empty(array->queue + rt_se_prio(rt_se)))
1319 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1320
1321 rt_se->on_list = 0;
1322}
1323
1324static inline struct sched_statistics *
1325__schedstats_from_rt_se(struct sched_rt_entity *rt_se)
1326{
1327#ifdef CONFIG_RT_GROUP_SCHED
1328 /* schedstats is not supported for rt group. */
1329 if (!rt_entity_is_task(rt_se))
1330 return NULL;
1331#endif
1332
1333 return &rt_task_of(rt_se)->stats;
1334}
1335
1336static inline void
1337update_stats_wait_start_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1338{
1339 struct sched_statistics *stats;
1340 struct task_struct *p = NULL;
1341
1342 if (!schedstat_enabled())
1343 return;
1344
1345 if (rt_entity_is_task(rt_se))
1346 p = rt_task_of(rt_se);
1347
1348 stats = __schedstats_from_rt_se(rt_se);
1349 if (!stats)
1350 return;
1351
1352 __update_stats_wait_start(rq_of_rt_rq(rt_rq), p, stats);
1353}
1354
1355static inline void
1356update_stats_enqueue_sleeper_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1357{
1358 struct sched_statistics *stats;
1359 struct task_struct *p = NULL;
1360
1361 if (!schedstat_enabled())
1362 return;
1363
1364 if (rt_entity_is_task(rt_se))
1365 p = rt_task_of(rt_se);
1366
1367 stats = __schedstats_from_rt_se(rt_se);
1368 if (!stats)
1369 return;
1370
1371 __update_stats_enqueue_sleeper(rq_of_rt_rq(rt_rq), p, stats);
1372}
1373
1374static inline void
1375update_stats_enqueue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1376 int flags)
1377{
1378 if (!schedstat_enabled())
1379 return;
1380
1381 if (flags & ENQUEUE_WAKEUP)
1382 update_stats_enqueue_sleeper_rt(rt_rq, rt_se);
1383}
1384
1385static inline void
1386update_stats_wait_end_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1387{
1388 struct sched_statistics *stats;
1389 struct task_struct *p = NULL;
1390
1391 if (!schedstat_enabled())
1392 return;
1393
1394 if (rt_entity_is_task(rt_se))
1395 p = rt_task_of(rt_se);
1396
1397 stats = __schedstats_from_rt_se(rt_se);
1398 if (!stats)
1399 return;
1400
1401 __update_stats_wait_end(rq_of_rt_rq(rt_rq), p, stats);
1402}
1403
1404static inline void
1405update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1406 int flags)
1407{
1408 struct task_struct *p = NULL;
1409
1410 if (!schedstat_enabled())
1411 return;
1412
1413 if (rt_entity_is_task(rt_se))
1414 p = rt_task_of(rt_se);
1415
1416 if ((flags & DEQUEUE_SLEEP) && p) {
1417 unsigned int state;
1418
1419 state = READ_ONCE(p->__state);
1420 if (state & TASK_INTERRUPTIBLE)
1421 __schedstat_set(p->stats.sleep_start,
1422 rq_clock(rq_of_rt_rq(rt_rq)));
1423
1424 if (state & TASK_UNINTERRUPTIBLE)
1425 __schedstat_set(p->stats.block_start,
1426 rq_clock(rq_of_rt_rq(rt_rq)));
1427 }
1428}
1429
1430static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1431{
1432 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1433 struct rt_prio_array *array = &rt_rq->active;
1434 struct rt_rq *group_rq = group_rt_rq(rt_se);
1435 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1436
1437 /*
1438 * Don't enqueue the group if its throttled, or when empty.
1439 * The latter is a consequence of the former when a child group
1440 * get throttled and the current group doesn't have any other
1441 * active members.
1442 */
1443 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1444 if (rt_se->on_list)
1445 __delist_rt_entity(rt_se, array);
1446 return;
1447 }
1448
1449 if (move_entity(flags)) {
1450 WARN_ON_ONCE(rt_se->on_list);
1451 if (flags & ENQUEUE_HEAD)
1452 list_add(&rt_se->run_list, queue);
1453 else
1454 list_add_tail(&rt_se->run_list, queue);
1455
1456 __set_bit(rt_se_prio(rt_se), array->bitmap);
1457 rt_se->on_list = 1;
1458 }
1459 rt_se->on_rq = 1;
1460
1461 inc_rt_tasks(rt_se, rt_rq);
1462}
1463
1464static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1465{
1466 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1467 struct rt_prio_array *array = &rt_rq->active;
1468
1469 if (move_entity(flags)) {
1470 WARN_ON_ONCE(!rt_se->on_list);
1471 __delist_rt_entity(rt_se, array);
1472 }
1473 rt_se->on_rq = 0;
1474
1475 dec_rt_tasks(rt_se, rt_rq);
1476}
1477
1478/*
1479 * Because the prio of an upper entry depends on the lower
1480 * entries, we must remove entries top - down.
1481 */
1482static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1483{
1484 struct sched_rt_entity *back = NULL;
1485 unsigned int rt_nr_running;
1486
1487 for_each_sched_rt_entity(rt_se) {
1488 rt_se->back = back;
1489 back = rt_se;
1490 }
1491
1492 rt_nr_running = rt_rq_of_se(back)->rt_nr_running;
1493
1494 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1495 if (on_rt_rq(rt_se))
1496 __dequeue_rt_entity(rt_se, flags);
1497 }
1498
1499 dequeue_top_rt_rq(rt_rq_of_se(back), rt_nr_running);
1500}
1501
1502static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1503{
1504 struct rq *rq = rq_of_rt_se(rt_se);
1505
1506 update_stats_enqueue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1507
1508 dequeue_rt_stack(rt_se, flags);
1509 for_each_sched_rt_entity(rt_se)
1510 __enqueue_rt_entity(rt_se, flags);
1511 enqueue_top_rt_rq(&rq->rt);
1512}
1513
1514static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1515{
1516 struct rq *rq = rq_of_rt_se(rt_se);
1517
1518 update_stats_dequeue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1519
1520 dequeue_rt_stack(rt_se, flags);
1521
1522 for_each_sched_rt_entity(rt_se) {
1523 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1524
1525 if (rt_rq && rt_rq->rt_nr_running)
1526 __enqueue_rt_entity(rt_se, flags);
1527 }
1528 enqueue_top_rt_rq(&rq->rt);
1529}
1530
1531/*
1532 * Adding/removing a task to/from a priority array:
1533 */
1534static void
1535enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1536{
1537 struct sched_rt_entity *rt_se = &p->rt;
1538
1539 if (flags & ENQUEUE_WAKEUP)
1540 rt_se->timeout = 0;
1541
1542 check_schedstat_required();
1543 update_stats_wait_start_rt(rt_rq_of_se(rt_se), rt_se);
1544
1545 enqueue_rt_entity(rt_se, flags);
1546
1547 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1548 enqueue_pushable_task(rq, p);
1549}
1550
1551static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1552{
1553 struct sched_rt_entity *rt_se = &p->rt;
1554
1555 update_curr_rt(rq);
1556 dequeue_rt_entity(rt_se, flags);
1557
1558 dequeue_pushable_task(rq, p);
1559}
1560
1561/*
1562 * Put task to the head or the end of the run list without the overhead of
1563 * dequeue followed by enqueue.
1564 */
1565static void
1566requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1567{
1568 if (on_rt_rq(rt_se)) {
1569 struct rt_prio_array *array = &rt_rq->active;
1570 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1571
1572 if (head)
1573 list_move(&rt_se->run_list, queue);
1574 else
1575 list_move_tail(&rt_se->run_list, queue);
1576 }
1577}
1578
1579static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1580{
1581 struct sched_rt_entity *rt_se = &p->rt;
1582 struct rt_rq *rt_rq;
1583
1584 for_each_sched_rt_entity(rt_se) {
1585 rt_rq = rt_rq_of_se(rt_se);
1586 requeue_rt_entity(rt_rq, rt_se, head);
1587 }
1588}
1589
1590static void yield_task_rt(struct rq *rq)
1591{
1592 requeue_task_rt(rq, rq->curr, 0);
1593}
1594
1595#ifdef CONFIG_SMP
1596static int find_lowest_rq(struct task_struct *task);
1597
1598static int
1599select_task_rq_rt(struct task_struct *p, int cpu, int flags)
1600{
1601 struct task_struct *curr;
1602 struct rq *rq;
1603 bool test;
1604
1605 /* For anything but wake ups, just return the task_cpu */
1606 if (!(flags & (WF_TTWU | WF_FORK)))
1607 goto out;
1608
1609 rq = cpu_rq(cpu);
1610
1611 rcu_read_lock();
1612 curr = READ_ONCE(rq->curr); /* unlocked access */
1613
1614 /*
1615 * If the current task on @p's runqueue is an RT task, then
1616 * try to see if we can wake this RT task up on another
1617 * runqueue. Otherwise simply start this RT task
1618 * on its current runqueue.
1619 *
1620 * We want to avoid overloading runqueues. If the woken
1621 * task is a higher priority, then it will stay on this CPU
1622 * and the lower prio task should be moved to another CPU.
1623 * Even though this will probably make the lower prio task
1624 * lose its cache, we do not want to bounce a higher task
1625 * around just because it gave up its CPU, perhaps for a
1626 * lock?
1627 *
1628 * For equal prio tasks, we just let the scheduler sort it out.
1629 *
1630 * Otherwise, just let it ride on the affined RQ and the
1631 * post-schedule router will push the preempted task away
1632 *
1633 * This test is optimistic, if we get it wrong the load-balancer
1634 * will have to sort it out.
1635 *
1636 * We take into account the capacity of the CPU to ensure it fits the
1637 * requirement of the task - which is only important on heterogeneous
1638 * systems like big.LITTLE.
1639 */
1640 test = curr &&
1641 unlikely(rt_task(curr)) &&
1642 (curr->nr_cpus_allowed < 2 || curr->prio <= p->prio);
1643
1644 if (test || !rt_task_fits_capacity(p, cpu)) {
1645 int target = find_lowest_rq(p);
1646
1647 /*
1648 * Bail out if we were forcing a migration to find a better
1649 * fitting CPU but our search failed.
1650 */
1651 if (!test && target != -1 && !rt_task_fits_capacity(p, target))
1652 goto out_unlock;
1653
1654 /*
1655 * Don't bother moving it if the destination CPU is
1656 * not running a lower priority task.
1657 */
1658 if (target != -1 &&
1659 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1660 cpu = target;
1661 }
1662
1663out_unlock:
1664 rcu_read_unlock();
1665
1666out:
1667 return cpu;
1668}
1669
1670static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1671{
1672 /*
1673 * Current can't be migrated, useless to reschedule,
1674 * let's hope p can move out.
1675 */
1676 if (rq->curr->nr_cpus_allowed == 1 ||
1677 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1678 return;
1679
1680 /*
1681 * p is migratable, so let's not schedule it and
1682 * see if it is pushed or pulled somewhere else.
1683 */
1684 if (p->nr_cpus_allowed != 1 &&
1685 cpupri_find(&rq->rd->cpupri, p, NULL))
1686 return;
1687
1688 /*
1689 * There appear to be other CPUs that can accept
1690 * the current task but none can run 'p', so lets reschedule
1691 * to try and push the current task away:
1692 */
1693 requeue_task_rt(rq, p, 1);
1694 resched_curr(rq);
1695}
1696
1697static int balance_rt(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1698{
1699 if (!on_rt_rq(&p->rt) && need_pull_rt_task(rq, p)) {
1700 /*
1701 * This is OK, because current is on_cpu, which avoids it being
1702 * picked for load-balance and preemption/IRQs are still
1703 * disabled avoiding further scheduler activity on it and we've
1704 * not yet started the picking loop.
1705 */
1706 rq_unpin_lock(rq, rf);
1707 pull_rt_task(rq);
1708 rq_repin_lock(rq, rf);
1709 }
1710
1711 return sched_stop_runnable(rq) || sched_dl_runnable(rq) || sched_rt_runnable(rq);
1712}
1713#endif /* CONFIG_SMP */
1714
1715/*
1716 * Preempt the current task with a newly woken task if needed:
1717 */
1718static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1719{
1720 if (p->prio < rq->curr->prio) {
1721 resched_curr(rq);
1722 return;
1723 }
1724
1725#ifdef CONFIG_SMP
1726 /*
1727 * If:
1728 *
1729 * - the newly woken task is of equal priority to the current task
1730 * - the newly woken task is non-migratable while current is migratable
1731 * - current will be preempted on the next reschedule
1732 *
1733 * we should check to see if current can readily move to a different
1734 * cpu. If so, we will reschedule to allow the push logic to try
1735 * to move current somewhere else, making room for our non-migratable
1736 * task.
1737 */
1738 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1739 check_preempt_equal_prio(rq, p);
1740#endif
1741}
1742
1743static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool first)
1744{
1745 struct sched_rt_entity *rt_se = &p->rt;
1746 struct rt_rq *rt_rq = &rq->rt;
1747
1748 p->se.exec_start = rq_clock_task(rq);
1749 if (on_rt_rq(&p->rt))
1750 update_stats_wait_end_rt(rt_rq, rt_se);
1751
1752 /* The running task is never eligible for pushing */
1753 dequeue_pushable_task(rq, p);
1754
1755 if (!first)
1756 return;
1757
1758 /*
1759 * If prev task was rt, put_prev_task() has already updated the
1760 * utilization. We only care of the case where we start to schedule a
1761 * rt task
1762 */
1763 if (rq->curr->sched_class != &rt_sched_class)
1764 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1765
1766 rt_queue_push_tasks(rq);
1767}
1768
1769static struct sched_rt_entity *pick_next_rt_entity(struct rt_rq *rt_rq)
1770{
1771 struct rt_prio_array *array = &rt_rq->active;
1772 struct sched_rt_entity *next = NULL;
1773 struct list_head *queue;
1774 int idx;
1775
1776 idx = sched_find_first_bit(array->bitmap);
1777 BUG_ON(idx >= MAX_RT_PRIO);
1778
1779 queue = array->queue + idx;
1780 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1781
1782 return next;
1783}
1784
1785static struct task_struct *_pick_next_task_rt(struct rq *rq)
1786{
1787 struct sched_rt_entity *rt_se;
1788 struct rt_rq *rt_rq = &rq->rt;
1789
1790 do {
1791 rt_se = pick_next_rt_entity(rt_rq);
1792 BUG_ON(!rt_se);
1793 rt_rq = group_rt_rq(rt_se);
1794 } while (rt_rq);
1795
1796 return rt_task_of(rt_se);
1797}
1798
1799static struct task_struct *pick_task_rt(struct rq *rq)
1800{
1801 struct task_struct *p;
1802
1803 if (!sched_rt_runnable(rq))
1804 return NULL;
1805
1806 p = _pick_next_task_rt(rq);
1807
1808 return p;
1809}
1810
1811static struct task_struct *pick_next_task_rt(struct rq *rq)
1812{
1813 struct task_struct *p = pick_task_rt(rq);
1814
1815 if (p)
1816 set_next_task_rt(rq, p, true);
1817
1818 return p;
1819}
1820
1821static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1822{
1823 struct sched_rt_entity *rt_se = &p->rt;
1824 struct rt_rq *rt_rq = &rq->rt;
1825
1826 if (on_rt_rq(&p->rt))
1827 update_stats_wait_start_rt(rt_rq, rt_se);
1828
1829 update_curr_rt(rq);
1830
1831 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1832
1833 /*
1834 * The previous task needs to be made eligible for pushing
1835 * if it is still active
1836 */
1837 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1838 enqueue_pushable_task(rq, p);
1839}
1840
1841#ifdef CONFIG_SMP
1842
1843/* Only try algorithms three times */
1844#define RT_MAX_TRIES 3
1845
1846static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1847{
1848 if (!task_on_cpu(rq, p) &&
1849 cpumask_test_cpu(cpu, &p->cpus_mask))
1850 return 1;
1851
1852 return 0;
1853}
1854
1855/*
1856 * Return the highest pushable rq's task, which is suitable to be executed
1857 * on the CPU, NULL otherwise
1858 */
1859static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1860{
1861 struct plist_head *head = &rq->rt.pushable_tasks;
1862 struct task_struct *p;
1863
1864 if (!has_pushable_tasks(rq))
1865 return NULL;
1866
1867 plist_for_each_entry(p, head, pushable_tasks) {
1868 if (pick_rt_task(rq, p, cpu))
1869 return p;
1870 }
1871
1872 return NULL;
1873}
1874
1875static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1876
1877static int find_lowest_rq(struct task_struct *task)
1878{
1879 struct sched_domain *sd;
1880 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1881 int this_cpu = smp_processor_id();
1882 int cpu = task_cpu(task);
1883 int ret;
1884
1885 /* Make sure the mask is initialized first */
1886 if (unlikely(!lowest_mask))
1887 return -1;
1888
1889 if (task->nr_cpus_allowed == 1)
1890 return -1; /* No other targets possible */
1891
1892 /*
1893 * If we're on asym system ensure we consider the different capacities
1894 * of the CPUs when searching for the lowest_mask.
1895 */
1896 if (sched_asym_cpucap_active()) {
1897
1898 ret = cpupri_find_fitness(&task_rq(task)->rd->cpupri,
1899 task, lowest_mask,
1900 rt_task_fits_capacity);
1901 } else {
1902
1903 ret = cpupri_find(&task_rq(task)->rd->cpupri,
1904 task, lowest_mask);
1905 }
1906
1907 if (!ret)
1908 return -1; /* No targets found */
1909
1910 /*
1911 * At this point we have built a mask of CPUs representing the
1912 * lowest priority tasks in the system. Now we want to elect
1913 * the best one based on our affinity and topology.
1914 *
1915 * We prioritize the last CPU that the task executed on since
1916 * it is most likely cache-hot in that location.
1917 */
1918 if (cpumask_test_cpu(cpu, lowest_mask))
1919 return cpu;
1920
1921 /*
1922 * Otherwise, we consult the sched_domains span maps to figure
1923 * out which CPU is logically closest to our hot cache data.
1924 */
1925 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1926 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1927
1928 rcu_read_lock();
1929 for_each_domain(cpu, sd) {
1930 if (sd->flags & SD_WAKE_AFFINE) {
1931 int best_cpu;
1932
1933 /*
1934 * "this_cpu" is cheaper to preempt than a
1935 * remote processor.
1936 */
1937 if (this_cpu != -1 &&
1938 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1939 rcu_read_unlock();
1940 return this_cpu;
1941 }
1942
1943 best_cpu = cpumask_any_and_distribute(lowest_mask,
1944 sched_domain_span(sd));
1945 if (best_cpu < nr_cpu_ids) {
1946 rcu_read_unlock();
1947 return best_cpu;
1948 }
1949 }
1950 }
1951 rcu_read_unlock();
1952
1953 /*
1954 * And finally, if there were no matches within the domains
1955 * just give the caller *something* to work with from the compatible
1956 * locations.
1957 */
1958 if (this_cpu != -1)
1959 return this_cpu;
1960
1961 cpu = cpumask_any_distribute(lowest_mask);
1962 if (cpu < nr_cpu_ids)
1963 return cpu;
1964
1965 return -1;
1966}
1967
1968/* Will lock the rq it finds */
1969static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1970{
1971 struct rq *lowest_rq = NULL;
1972 int tries;
1973 int cpu;
1974
1975 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1976 cpu = find_lowest_rq(task);
1977
1978 if ((cpu == -1) || (cpu == rq->cpu))
1979 break;
1980
1981 lowest_rq = cpu_rq(cpu);
1982
1983 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1984 /*
1985 * Target rq has tasks of equal or higher priority,
1986 * retrying does not release any lock and is unlikely
1987 * to yield a different result.
1988 */
1989 lowest_rq = NULL;
1990 break;
1991 }
1992
1993 /* if the prio of this runqueue changed, try again */
1994 if (double_lock_balance(rq, lowest_rq)) {
1995 /*
1996 * We had to unlock the run queue. In
1997 * the mean time, task could have
1998 * migrated already or had its affinity changed.
1999 * Also make sure that it wasn't scheduled on its rq.
2000 */
2001 if (unlikely(task_rq(task) != rq ||
2002 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_mask) ||
2003 task_on_cpu(rq, task) ||
2004 !rt_task(task) ||
2005 !task_on_rq_queued(task))) {
2006
2007 double_unlock_balance(rq, lowest_rq);
2008 lowest_rq = NULL;
2009 break;
2010 }
2011 }
2012
2013 /* If this rq is still suitable use it. */
2014 if (lowest_rq->rt.highest_prio.curr > task->prio)
2015 break;
2016
2017 /* try again */
2018 double_unlock_balance(rq, lowest_rq);
2019 lowest_rq = NULL;
2020 }
2021
2022 return lowest_rq;
2023}
2024
2025static struct task_struct *pick_next_pushable_task(struct rq *rq)
2026{
2027 struct task_struct *p;
2028
2029 if (!has_pushable_tasks(rq))
2030 return NULL;
2031
2032 p = plist_first_entry(&rq->rt.pushable_tasks,
2033 struct task_struct, pushable_tasks);
2034
2035 BUG_ON(rq->cpu != task_cpu(p));
2036 BUG_ON(task_current(rq, p));
2037 BUG_ON(p->nr_cpus_allowed <= 1);
2038
2039 BUG_ON(!task_on_rq_queued(p));
2040 BUG_ON(!rt_task(p));
2041
2042 return p;
2043}
2044
2045/*
2046 * If the current CPU has more than one RT task, see if the non
2047 * running task can migrate over to a CPU that is running a task
2048 * of lesser priority.
2049 */
2050static int push_rt_task(struct rq *rq, bool pull)
2051{
2052 struct task_struct *next_task;
2053 struct rq *lowest_rq;
2054 int ret = 0;
2055
2056 if (!rq->rt.overloaded)
2057 return 0;
2058
2059 next_task = pick_next_pushable_task(rq);
2060 if (!next_task)
2061 return 0;
2062
2063retry:
2064 /*
2065 * It's possible that the next_task slipped in of
2066 * higher priority than current. If that's the case
2067 * just reschedule current.
2068 */
2069 if (unlikely(next_task->prio < rq->curr->prio)) {
2070 resched_curr(rq);
2071 return 0;
2072 }
2073
2074 if (is_migration_disabled(next_task)) {
2075 struct task_struct *push_task = NULL;
2076 int cpu;
2077
2078 if (!pull || rq->push_busy)
2079 return 0;
2080
2081 /*
2082 * Invoking find_lowest_rq() on anything but an RT task doesn't
2083 * make sense. Per the above priority check, curr has to
2084 * be of higher priority than next_task, so no need to
2085 * reschedule when bailing out.
2086 *
2087 * Note that the stoppers are masqueraded as SCHED_FIFO
2088 * (cf. sched_set_stop_task()), so we can't rely on rt_task().
2089 */
2090 if (rq->curr->sched_class != &rt_sched_class)
2091 return 0;
2092
2093 cpu = find_lowest_rq(rq->curr);
2094 if (cpu == -1 || cpu == rq->cpu)
2095 return 0;
2096
2097 /*
2098 * Given we found a CPU with lower priority than @next_task,
2099 * therefore it should be running. However we cannot migrate it
2100 * to this other CPU, instead attempt to push the current
2101 * running task on this CPU away.
2102 */
2103 push_task = get_push_task(rq);
2104 if (push_task) {
2105 raw_spin_rq_unlock(rq);
2106 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2107 push_task, &rq->push_work);
2108 raw_spin_rq_lock(rq);
2109 }
2110
2111 return 0;
2112 }
2113
2114 if (WARN_ON(next_task == rq->curr))
2115 return 0;
2116
2117 /* We might release rq lock */
2118 get_task_struct(next_task);
2119
2120 /* find_lock_lowest_rq locks the rq if found */
2121 lowest_rq = find_lock_lowest_rq(next_task, rq);
2122 if (!lowest_rq) {
2123 struct task_struct *task;
2124 /*
2125 * find_lock_lowest_rq releases rq->lock
2126 * so it is possible that next_task has migrated.
2127 *
2128 * We need to make sure that the task is still on the same
2129 * run-queue and is also still the next task eligible for
2130 * pushing.
2131 */
2132 task = pick_next_pushable_task(rq);
2133 if (task == next_task) {
2134 /*
2135 * The task hasn't migrated, and is still the next
2136 * eligible task, but we failed to find a run-queue
2137 * to push it to. Do not retry in this case, since
2138 * other CPUs will pull from us when ready.
2139 */
2140 goto out;
2141 }
2142
2143 if (!task)
2144 /* No more tasks, just exit */
2145 goto out;
2146
2147 /*
2148 * Something has shifted, try again.
2149 */
2150 put_task_struct(next_task);
2151 next_task = task;
2152 goto retry;
2153 }
2154
2155 deactivate_task(rq, next_task, 0);
2156 set_task_cpu(next_task, lowest_rq->cpu);
2157 activate_task(lowest_rq, next_task, 0);
2158 resched_curr(lowest_rq);
2159 ret = 1;
2160
2161 double_unlock_balance(rq, lowest_rq);
2162out:
2163 put_task_struct(next_task);
2164
2165 return ret;
2166}
2167
2168static void push_rt_tasks(struct rq *rq)
2169{
2170 /* push_rt_task will return true if it moved an RT */
2171 while (push_rt_task(rq, false))
2172 ;
2173}
2174
2175#ifdef HAVE_RT_PUSH_IPI
2176
2177/*
2178 * When a high priority task schedules out from a CPU and a lower priority
2179 * task is scheduled in, a check is made to see if there's any RT tasks
2180 * on other CPUs that are waiting to run because a higher priority RT task
2181 * is currently running on its CPU. In this case, the CPU with multiple RT
2182 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
2183 * up that may be able to run one of its non-running queued RT tasks.
2184 *
2185 * All CPUs with overloaded RT tasks need to be notified as there is currently
2186 * no way to know which of these CPUs have the highest priority task waiting
2187 * to run. Instead of trying to take a spinlock on each of these CPUs,
2188 * which has shown to cause large latency when done on machines with many
2189 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
2190 * RT tasks waiting to run.
2191 *
2192 * Just sending an IPI to each of the CPUs is also an issue, as on large
2193 * count CPU machines, this can cause an IPI storm on a CPU, especially
2194 * if its the only CPU with multiple RT tasks queued, and a large number
2195 * of CPUs scheduling a lower priority task at the same time.
2196 *
2197 * Each root domain has its own irq work function that can iterate over
2198 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
2199 * task must be checked if there's one or many CPUs that are lowering
2200 * their priority, there's a single irq work iterator that will try to
2201 * push off RT tasks that are waiting to run.
2202 *
2203 * When a CPU schedules a lower priority task, it will kick off the
2204 * irq work iterator that will jump to each CPU with overloaded RT tasks.
2205 * As it only takes the first CPU that schedules a lower priority task
2206 * to start the process, the rto_start variable is incremented and if
2207 * the atomic result is one, then that CPU will try to take the rto_lock.
2208 * This prevents high contention on the lock as the process handles all
2209 * CPUs scheduling lower priority tasks.
2210 *
2211 * All CPUs that are scheduling a lower priority task will increment the
2212 * rt_loop_next variable. This will make sure that the irq work iterator
2213 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
2214 * priority task, even if the iterator is in the middle of a scan. Incrementing
2215 * the rt_loop_next will cause the iterator to perform another scan.
2216 *
2217 */
2218static int rto_next_cpu(struct root_domain *rd)
2219{
2220 int next;
2221 int cpu;
2222
2223 /*
2224 * When starting the IPI RT pushing, the rto_cpu is set to -1,
2225 * rt_next_cpu() will simply return the first CPU found in
2226 * the rto_mask.
2227 *
2228 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
2229 * will return the next CPU found in the rto_mask.
2230 *
2231 * If there are no more CPUs left in the rto_mask, then a check is made
2232 * against rto_loop and rto_loop_next. rto_loop is only updated with
2233 * the rto_lock held, but any CPU may increment the rto_loop_next
2234 * without any locking.
2235 */
2236 for (;;) {
2237
2238 /* When rto_cpu is -1 this acts like cpumask_first() */
2239 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
2240
2241 rd->rto_cpu = cpu;
2242
2243 if (cpu < nr_cpu_ids)
2244 return cpu;
2245
2246 rd->rto_cpu = -1;
2247
2248 /*
2249 * ACQUIRE ensures we see the @rto_mask changes
2250 * made prior to the @next value observed.
2251 *
2252 * Matches WMB in rt_set_overload().
2253 */
2254 next = atomic_read_acquire(&rd->rto_loop_next);
2255
2256 if (rd->rto_loop == next)
2257 break;
2258
2259 rd->rto_loop = next;
2260 }
2261
2262 return -1;
2263}
2264
2265static inline bool rto_start_trylock(atomic_t *v)
2266{
2267 return !atomic_cmpxchg_acquire(v, 0, 1);
2268}
2269
2270static inline void rto_start_unlock(atomic_t *v)
2271{
2272 atomic_set_release(v, 0);
2273}
2274
2275static void tell_cpu_to_push(struct rq *rq)
2276{
2277 int cpu = -1;
2278
2279 /* Keep the loop going if the IPI is currently active */
2280 atomic_inc(&rq->rd->rto_loop_next);
2281
2282 /* Only one CPU can initiate a loop at a time */
2283 if (!rto_start_trylock(&rq->rd->rto_loop_start))
2284 return;
2285
2286 raw_spin_lock(&rq->rd->rto_lock);
2287
2288 /*
2289 * The rto_cpu is updated under the lock, if it has a valid CPU
2290 * then the IPI is still running and will continue due to the
2291 * update to loop_next, and nothing needs to be done here.
2292 * Otherwise it is finishing up and an ipi needs to be sent.
2293 */
2294 if (rq->rd->rto_cpu < 0)
2295 cpu = rto_next_cpu(rq->rd);
2296
2297 raw_spin_unlock(&rq->rd->rto_lock);
2298
2299 rto_start_unlock(&rq->rd->rto_loop_start);
2300
2301 if (cpu >= 0) {
2302 /* Make sure the rd does not get freed while pushing */
2303 sched_get_rd(rq->rd);
2304 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2305 }
2306}
2307
2308/* Called from hardirq context */
2309void rto_push_irq_work_func(struct irq_work *work)
2310{
2311 struct root_domain *rd =
2312 container_of(work, struct root_domain, rto_push_work);
2313 struct rq *rq;
2314 int cpu;
2315
2316 rq = this_rq();
2317
2318 /*
2319 * We do not need to grab the lock to check for has_pushable_tasks.
2320 * When it gets updated, a check is made if a push is possible.
2321 */
2322 if (has_pushable_tasks(rq)) {
2323 raw_spin_rq_lock(rq);
2324 while (push_rt_task(rq, true))
2325 ;
2326 raw_spin_rq_unlock(rq);
2327 }
2328
2329 raw_spin_lock(&rd->rto_lock);
2330
2331 /* Pass the IPI to the next rt overloaded queue */
2332 cpu = rto_next_cpu(rd);
2333
2334 raw_spin_unlock(&rd->rto_lock);
2335
2336 if (cpu < 0) {
2337 sched_put_rd(rd);
2338 return;
2339 }
2340
2341 /* Try the next RT overloaded CPU */
2342 irq_work_queue_on(&rd->rto_push_work, cpu);
2343}
2344#endif /* HAVE_RT_PUSH_IPI */
2345
2346static void pull_rt_task(struct rq *this_rq)
2347{
2348 int this_cpu = this_rq->cpu, cpu;
2349 bool resched = false;
2350 struct task_struct *p, *push_task;
2351 struct rq *src_rq;
2352 int rt_overload_count = rt_overloaded(this_rq);
2353
2354 if (likely(!rt_overload_count))
2355 return;
2356
2357 /*
2358 * Match the barrier from rt_set_overloaded; this guarantees that if we
2359 * see overloaded we must also see the rto_mask bit.
2360 */
2361 smp_rmb();
2362
2363 /* If we are the only overloaded CPU do nothing */
2364 if (rt_overload_count == 1 &&
2365 cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2366 return;
2367
2368#ifdef HAVE_RT_PUSH_IPI
2369 if (sched_feat(RT_PUSH_IPI)) {
2370 tell_cpu_to_push(this_rq);
2371 return;
2372 }
2373#endif
2374
2375 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2376 if (this_cpu == cpu)
2377 continue;
2378
2379 src_rq = cpu_rq(cpu);
2380
2381 /*
2382 * Don't bother taking the src_rq->lock if the next highest
2383 * task is known to be lower-priority than our current task.
2384 * This may look racy, but if this value is about to go
2385 * logically higher, the src_rq will push this task away.
2386 * And if its going logically lower, we do not care
2387 */
2388 if (src_rq->rt.highest_prio.next >=
2389 this_rq->rt.highest_prio.curr)
2390 continue;
2391
2392 /*
2393 * We can potentially drop this_rq's lock in
2394 * double_lock_balance, and another CPU could
2395 * alter this_rq
2396 */
2397 push_task = NULL;
2398 double_lock_balance(this_rq, src_rq);
2399
2400 /*
2401 * We can pull only a task, which is pushable
2402 * on its rq, and no others.
2403 */
2404 p = pick_highest_pushable_task(src_rq, this_cpu);
2405
2406 /*
2407 * Do we have an RT task that preempts
2408 * the to-be-scheduled task?
2409 */
2410 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2411 WARN_ON(p == src_rq->curr);
2412 WARN_ON(!task_on_rq_queued(p));
2413
2414 /*
2415 * There's a chance that p is higher in priority
2416 * than what's currently running on its CPU.
2417 * This is just that p is waking up and hasn't
2418 * had a chance to schedule. We only pull
2419 * p if it is lower in priority than the
2420 * current task on the run queue
2421 */
2422 if (p->prio < src_rq->curr->prio)
2423 goto skip;
2424
2425 if (is_migration_disabled(p)) {
2426 push_task = get_push_task(src_rq);
2427 } else {
2428 deactivate_task(src_rq, p, 0);
2429 set_task_cpu(p, this_cpu);
2430 activate_task(this_rq, p, 0);
2431 resched = true;
2432 }
2433 /*
2434 * We continue with the search, just in
2435 * case there's an even higher prio task
2436 * in another runqueue. (low likelihood
2437 * but possible)
2438 */
2439 }
2440skip:
2441 double_unlock_balance(this_rq, src_rq);
2442
2443 if (push_task) {
2444 raw_spin_rq_unlock(this_rq);
2445 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2446 push_task, &src_rq->push_work);
2447 raw_spin_rq_lock(this_rq);
2448 }
2449 }
2450
2451 if (resched)
2452 resched_curr(this_rq);
2453}
2454
2455/*
2456 * If we are not running and we are not going to reschedule soon, we should
2457 * try to push tasks away now
2458 */
2459static void task_woken_rt(struct rq *rq, struct task_struct *p)
2460{
2461 bool need_to_push = !task_on_cpu(rq, p) &&
2462 !test_tsk_need_resched(rq->curr) &&
2463 p->nr_cpus_allowed > 1 &&
2464 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2465 (rq->curr->nr_cpus_allowed < 2 ||
2466 rq->curr->prio <= p->prio);
2467
2468 if (need_to_push)
2469 push_rt_tasks(rq);
2470}
2471
2472/* Assumes rq->lock is held */
2473static void rq_online_rt(struct rq *rq)
2474{
2475 if (rq->rt.overloaded)
2476 rt_set_overload(rq);
2477
2478 __enable_runtime(rq);
2479
2480 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2481}
2482
2483/* Assumes rq->lock is held */
2484static void rq_offline_rt(struct rq *rq)
2485{
2486 if (rq->rt.overloaded)
2487 rt_clear_overload(rq);
2488
2489 __disable_runtime(rq);
2490
2491 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2492}
2493
2494/*
2495 * When switch from the rt queue, we bring ourselves to a position
2496 * that we might want to pull RT tasks from other runqueues.
2497 */
2498static void switched_from_rt(struct rq *rq, struct task_struct *p)
2499{
2500 /*
2501 * If there are other RT tasks then we will reschedule
2502 * and the scheduling of the other RT tasks will handle
2503 * the balancing. But if we are the last RT task
2504 * we may need to handle the pulling of RT tasks
2505 * now.
2506 */
2507 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2508 return;
2509
2510 rt_queue_pull_task(rq);
2511}
2512
2513void __init init_sched_rt_class(void)
2514{
2515 unsigned int i;
2516
2517 for_each_possible_cpu(i) {
2518 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2519 GFP_KERNEL, cpu_to_node(i));
2520 }
2521}
2522#endif /* CONFIG_SMP */
2523
2524/*
2525 * When switching a task to RT, we may overload the runqueue
2526 * with RT tasks. In this case we try to push them off to
2527 * other runqueues.
2528 */
2529static void switched_to_rt(struct rq *rq, struct task_struct *p)
2530{
2531 /*
2532 * If we are running, update the avg_rt tracking, as the running time
2533 * will now on be accounted into the latter.
2534 */
2535 if (task_current(rq, p)) {
2536 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2537 return;
2538 }
2539
2540 /*
2541 * If we are not running we may need to preempt the current
2542 * running task. If that current running task is also an RT task
2543 * then see if we can move to another run queue.
2544 */
2545 if (task_on_rq_queued(p)) {
2546#ifdef CONFIG_SMP
2547 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2548 rt_queue_push_tasks(rq);
2549#endif /* CONFIG_SMP */
2550 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2551 resched_curr(rq);
2552 }
2553}
2554
2555/*
2556 * Priority of the task has changed. This may cause
2557 * us to initiate a push or pull.
2558 */
2559static void
2560prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2561{
2562 if (!task_on_rq_queued(p))
2563 return;
2564
2565 if (task_current(rq, p)) {
2566#ifdef CONFIG_SMP
2567 /*
2568 * If our priority decreases while running, we
2569 * may need to pull tasks to this runqueue.
2570 */
2571 if (oldprio < p->prio)
2572 rt_queue_pull_task(rq);
2573
2574 /*
2575 * If there's a higher priority task waiting to run
2576 * then reschedule.
2577 */
2578 if (p->prio > rq->rt.highest_prio.curr)
2579 resched_curr(rq);
2580#else
2581 /* For UP simply resched on drop of prio */
2582 if (oldprio < p->prio)
2583 resched_curr(rq);
2584#endif /* CONFIG_SMP */
2585 } else {
2586 /*
2587 * This task is not running, but if it is
2588 * greater than the current running task
2589 * then reschedule.
2590 */
2591 if (p->prio < rq->curr->prio)
2592 resched_curr(rq);
2593 }
2594}
2595
2596#ifdef CONFIG_POSIX_TIMERS
2597static void watchdog(struct rq *rq, struct task_struct *p)
2598{
2599 unsigned long soft, hard;
2600
2601 /* max may change after cur was read, this will be fixed next tick */
2602 soft = task_rlimit(p, RLIMIT_RTTIME);
2603 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2604
2605 if (soft != RLIM_INFINITY) {
2606 unsigned long next;
2607
2608 if (p->rt.watchdog_stamp != jiffies) {
2609 p->rt.timeout++;
2610 p->rt.watchdog_stamp = jiffies;
2611 }
2612
2613 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2614 if (p->rt.timeout > next) {
2615 posix_cputimers_rt_watchdog(&p->posix_cputimers,
2616 p->se.sum_exec_runtime);
2617 }
2618 }
2619}
2620#else
2621static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2622#endif
2623
2624/*
2625 * scheduler tick hitting a task of our scheduling class.
2626 *
2627 * NOTE: This function can be called remotely by the tick offload that
2628 * goes along full dynticks. Therefore no local assumption can be made
2629 * and everything must be accessed through the @rq and @curr passed in
2630 * parameters.
2631 */
2632static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2633{
2634 struct sched_rt_entity *rt_se = &p->rt;
2635
2636 update_curr_rt(rq);
2637 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2638
2639 watchdog(rq, p);
2640
2641 /*
2642 * RR tasks need a special form of timeslice management.
2643 * FIFO tasks have no timeslices.
2644 */
2645 if (p->policy != SCHED_RR)
2646 return;
2647
2648 if (--p->rt.time_slice)
2649 return;
2650
2651 p->rt.time_slice = sched_rr_timeslice;
2652
2653 /*
2654 * Requeue to the end of queue if we (and all of our ancestors) are not
2655 * the only element on the queue
2656 */
2657 for_each_sched_rt_entity(rt_se) {
2658 if (rt_se->run_list.prev != rt_se->run_list.next) {
2659 requeue_task_rt(rq, p, 0);
2660 resched_curr(rq);
2661 return;
2662 }
2663 }
2664}
2665
2666static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2667{
2668 /*
2669 * Time slice is 0 for SCHED_FIFO tasks
2670 */
2671 if (task->policy == SCHED_RR)
2672 return sched_rr_timeslice;
2673 else
2674 return 0;
2675}
2676
2677DEFINE_SCHED_CLASS(rt) = {
2678
2679 .enqueue_task = enqueue_task_rt,
2680 .dequeue_task = dequeue_task_rt,
2681 .yield_task = yield_task_rt,
2682
2683 .check_preempt_curr = check_preempt_curr_rt,
2684
2685 .pick_next_task = pick_next_task_rt,
2686 .put_prev_task = put_prev_task_rt,
2687 .set_next_task = set_next_task_rt,
2688
2689#ifdef CONFIG_SMP
2690 .balance = balance_rt,
2691 .pick_task = pick_task_rt,
2692 .select_task_rq = select_task_rq_rt,
2693 .set_cpus_allowed = set_cpus_allowed_common,
2694 .rq_online = rq_online_rt,
2695 .rq_offline = rq_offline_rt,
2696 .task_woken = task_woken_rt,
2697 .switched_from = switched_from_rt,
2698 .find_lock_rq = find_lock_lowest_rq,
2699#endif
2700
2701 .task_tick = task_tick_rt,
2702
2703 .get_rr_interval = get_rr_interval_rt,
2704
2705 .prio_changed = prio_changed_rt,
2706 .switched_to = switched_to_rt,
2707
2708 .update_curr = update_curr_rt,
2709
2710#ifdef CONFIG_UCLAMP_TASK
2711 .uclamp_enabled = 1,
2712#endif
2713};
2714
2715#ifdef CONFIG_RT_GROUP_SCHED
2716/*
2717 * Ensure that the real time constraints are schedulable.
2718 */
2719static DEFINE_MUTEX(rt_constraints_mutex);
2720
2721static inline int tg_has_rt_tasks(struct task_group *tg)
2722{
2723 struct task_struct *task;
2724 struct css_task_iter it;
2725 int ret = 0;
2726
2727 /*
2728 * Autogroups do not have RT tasks; see autogroup_create().
2729 */
2730 if (task_group_is_autogroup(tg))
2731 return 0;
2732
2733 css_task_iter_start(&tg->css, 0, &it);
2734 while (!ret && (task = css_task_iter_next(&it)))
2735 ret |= rt_task(task);
2736 css_task_iter_end(&it);
2737
2738 return ret;
2739}
2740
2741struct rt_schedulable_data {
2742 struct task_group *tg;
2743 u64 rt_period;
2744 u64 rt_runtime;
2745};
2746
2747static int tg_rt_schedulable(struct task_group *tg, void *data)
2748{
2749 struct rt_schedulable_data *d = data;
2750 struct task_group *child;
2751 unsigned long total, sum = 0;
2752 u64 period, runtime;
2753
2754 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2755 runtime = tg->rt_bandwidth.rt_runtime;
2756
2757 if (tg == d->tg) {
2758 period = d->rt_period;
2759 runtime = d->rt_runtime;
2760 }
2761
2762 /*
2763 * Cannot have more runtime than the period.
2764 */
2765 if (runtime > period && runtime != RUNTIME_INF)
2766 return -EINVAL;
2767
2768 /*
2769 * Ensure we don't starve existing RT tasks if runtime turns zero.
2770 */
2771 if (rt_bandwidth_enabled() && !runtime &&
2772 tg->rt_bandwidth.rt_runtime && tg_has_rt_tasks(tg))
2773 return -EBUSY;
2774
2775 total = to_ratio(period, runtime);
2776
2777 /*
2778 * Nobody can have more than the global setting allows.
2779 */
2780 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2781 return -EINVAL;
2782
2783 /*
2784 * The sum of our children's runtime should not exceed our own.
2785 */
2786 list_for_each_entry_rcu(child, &tg->children, siblings) {
2787 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2788 runtime = child->rt_bandwidth.rt_runtime;
2789
2790 if (child == d->tg) {
2791 period = d->rt_period;
2792 runtime = d->rt_runtime;
2793 }
2794
2795 sum += to_ratio(period, runtime);
2796 }
2797
2798 if (sum > total)
2799 return -EINVAL;
2800
2801 return 0;
2802}
2803
2804static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2805{
2806 int ret;
2807
2808 struct rt_schedulable_data data = {
2809 .tg = tg,
2810 .rt_period = period,
2811 .rt_runtime = runtime,
2812 };
2813
2814 rcu_read_lock();
2815 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2816 rcu_read_unlock();
2817
2818 return ret;
2819}
2820
2821static int tg_set_rt_bandwidth(struct task_group *tg,
2822 u64 rt_period, u64 rt_runtime)
2823{
2824 int i, err = 0;
2825
2826 /*
2827 * Disallowing the root group RT runtime is BAD, it would disallow the
2828 * kernel creating (and or operating) RT threads.
2829 */
2830 if (tg == &root_task_group && rt_runtime == 0)
2831 return -EINVAL;
2832
2833 /* No period doesn't make any sense. */
2834 if (rt_period == 0)
2835 return -EINVAL;
2836
2837 /*
2838 * Bound quota to defend quota against overflow during bandwidth shift.
2839 */
2840 if (rt_runtime != RUNTIME_INF && rt_runtime > max_rt_runtime)
2841 return -EINVAL;
2842
2843 mutex_lock(&rt_constraints_mutex);
2844 err = __rt_schedulable(tg, rt_period, rt_runtime);
2845 if (err)
2846 goto unlock;
2847
2848 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2849 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2850 tg->rt_bandwidth.rt_runtime = rt_runtime;
2851
2852 for_each_possible_cpu(i) {
2853 struct rt_rq *rt_rq = tg->rt_rq[i];
2854
2855 raw_spin_lock(&rt_rq->rt_runtime_lock);
2856 rt_rq->rt_runtime = rt_runtime;
2857 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2858 }
2859 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2860unlock:
2861 mutex_unlock(&rt_constraints_mutex);
2862
2863 return err;
2864}
2865
2866int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2867{
2868 u64 rt_runtime, rt_period;
2869
2870 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2871 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2872 if (rt_runtime_us < 0)
2873 rt_runtime = RUNTIME_INF;
2874 else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC)
2875 return -EINVAL;
2876
2877 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2878}
2879
2880long sched_group_rt_runtime(struct task_group *tg)
2881{
2882 u64 rt_runtime_us;
2883
2884 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2885 return -1;
2886
2887 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2888 do_div(rt_runtime_us, NSEC_PER_USEC);
2889 return rt_runtime_us;
2890}
2891
2892int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2893{
2894 u64 rt_runtime, rt_period;
2895
2896 if (rt_period_us > U64_MAX / NSEC_PER_USEC)
2897 return -EINVAL;
2898
2899 rt_period = rt_period_us * NSEC_PER_USEC;
2900 rt_runtime = tg->rt_bandwidth.rt_runtime;
2901
2902 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2903}
2904
2905long sched_group_rt_period(struct task_group *tg)
2906{
2907 u64 rt_period_us;
2908
2909 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2910 do_div(rt_period_us, NSEC_PER_USEC);
2911 return rt_period_us;
2912}
2913
2914#ifdef CONFIG_SYSCTL
2915static int sched_rt_global_constraints(void)
2916{
2917 int ret = 0;
2918
2919 mutex_lock(&rt_constraints_mutex);
2920 ret = __rt_schedulable(NULL, 0, 0);
2921 mutex_unlock(&rt_constraints_mutex);
2922
2923 return ret;
2924}
2925#endif /* CONFIG_SYSCTL */
2926
2927int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2928{
2929 /* Don't accept realtime tasks when there is no way for them to run */
2930 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2931 return 0;
2932
2933 return 1;
2934}
2935
2936#else /* !CONFIG_RT_GROUP_SCHED */
2937
2938#ifdef CONFIG_SYSCTL
2939static int sched_rt_global_constraints(void)
2940{
2941 unsigned long flags;
2942 int i;
2943
2944 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2945 for_each_possible_cpu(i) {
2946 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2947
2948 raw_spin_lock(&rt_rq->rt_runtime_lock);
2949 rt_rq->rt_runtime = global_rt_runtime();
2950 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2951 }
2952 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2953
2954 return 0;
2955}
2956#endif /* CONFIG_SYSCTL */
2957#endif /* CONFIG_RT_GROUP_SCHED */
2958
2959#ifdef CONFIG_SYSCTL
2960static int sched_rt_global_validate(void)
2961{
2962 if (sysctl_sched_rt_period <= 0)
2963 return -EINVAL;
2964
2965 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2966 ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
2967 ((u64)sysctl_sched_rt_runtime *
2968 NSEC_PER_USEC > max_rt_runtime)))
2969 return -EINVAL;
2970
2971 return 0;
2972}
2973
2974static void sched_rt_do_global(void)
2975{
2976 unsigned long flags;
2977
2978 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2979 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2980 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2981 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2982}
2983
2984static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
2985 size_t *lenp, loff_t *ppos)
2986{
2987 int old_period, old_runtime;
2988 static DEFINE_MUTEX(mutex);
2989 int ret;
2990
2991 mutex_lock(&mutex);
2992 old_period = sysctl_sched_rt_period;
2993 old_runtime = sysctl_sched_rt_runtime;
2994
2995 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2996
2997 if (!ret && write) {
2998 ret = sched_rt_global_validate();
2999 if (ret)
3000 goto undo;
3001
3002 ret = sched_dl_global_validate();
3003 if (ret)
3004 goto undo;
3005
3006 ret = sched_rt_global_constraints();
3007 if (ret)
3008 goto undo;
3009
3010 sched_rt_do_global();
3011 sched_dl_do_global();
3012 }
3013 if (0) {
3014undo:
3015 sysctl_sched_rt_period = old_period;
3016 sysctl_sched_rt_runtime = old_runtime;
3017 }
3018 mutex_unlock(&mutex);
3019
3020 return ret;
3021}
3022
3023static int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
3024 size_t *lenp, loff_t *ppos)
3025{
3026 int ret;
3027 static DEFINE_MUTEX(mutex);
3028
3029 mutex_lock(&mutex);
3030 ret = proc_dointvec(table, write, buffer, lenp, ppos);
3031 /*
3032 * Make sure that internally we keep jiffies.
3033 * Also, writing zero resets the timeslice to default:
3034 */
3035 if (!ret && write) {
3036 sched_rr_timeslice =
3037 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
3038 msecs_to_jiffies(sysctl_sched_rr_timeslice);
3039 }
3040 mutex_unlock(&mutex);
3041
3042 return ret;
3043}
3044#endif /* CONFIG_SYSCTL */
3045
3046#ifdef CONFIG_SCHED_DEBUG
3047void print_rt_stats(struct seq_file *m, int cpu)
3048{
3049 rt_rq_iter_t iter;
3050 struct rt_rq *rt_rq;
3051
3052 rcu_read_lock();
3053 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
3054 print_rt_rq(m, cpu, rt_rq);
3055 rcu_read_unlock();
3056}
3057#endif /* CONFIG_SCHED_DEBUG */
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
4 * policies)
5 */
6
7int sched_rr_timeslice = RR_TIMESLICE;
8/* More than 4 hours if BW_SHIFT equals 20. */
9static const u64 max_rt_runtime = MAX_BW;
10
11static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
12
13struct rt_bandwidth def_rt_bandwidth;
14
15/*
16 * period over which we measure -rt task CPU usage in us.
17 * default: 1s
18 */
19int sysctl_sched_rt_period = 1000000;
20
21/*
22 * part of the period that we allow rt tasks to run in us.
23 * default: 0.95s
24 */
25int sysctl_sched_rt_runtime = 950000;
26
27#ifdef CONFIG_SYSCTL
28static int sysctl_sched_rr_timeslice = (MSEC_PER_SEC * RR_TIMESLICE) / HZ;
29static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
30 size_t *lenp, loff_t *ppos);
31static int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
32 size_t *lenp, loff_t *ppos);
33static struct ctl_table sched_rt_sysctls[] = {
34 {
35 .procname = "sched_rt_period_us",
36 .data = &sysctl_sched_rt_period,
37 .maxlen = sizeof(int),
38 .mode = 0644,
39 .proc_handler = sched_rt_handler,
40 .extra1 = SYSCTL_ONE,
41 .extra2 = SYSCTL_INT_MAX,
42 },
43 {
44 .procname = "sched_rt_runtime_us",
45 .data = &sysctl_sched_rt_runtime,
46 .maxlen = sizeof(int),
47 .mode = 0644,
48 .proc_handler = sched_rt_handler,
49 .extra1 = SYSCTL_NEG_ONE,
50 .extra2 = (void *)&sysctl_sched_rt_period,
51 },
52 {
53 .procname = "sched_rr_timeslice_ms",
54 .data = &sysctl_sched_rr_timeslice,
55 .maxlen = sizeof(int),
56 .mode = 0644,
57 .proc_handler = sched_rr_handler,
58 },
59 {}
60};
61
62static int __init sched_rt_sysctl_init(void)
63{
64 register_sysctl_init("kernel", sched_rt_sysctls);
65 return 0;
66}
67late_initcall(sched_rt_sysctl_init);
68#endif
69
70static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
71{
72 struct rt_bandwidth *rt_b =
73 container_of(timer, struct rt_bandwidth, rt_period_timer);
74 int idle = 0;
75 int overrun;
76
77 raw_spin_lock(&rt_b->rt_runtime_lock);
78 for (;;) {
79 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
80 if (!overrun)
81 break;
82
83 raw_spin_unlock(&rt_b->rt_runtime_lock);
84 idle = do_sched_rt_period_timer(rt_b, overrun);
85 raw_spin_lock(&rt_b->rt_runtime_lock);
86 }
87 if (idle)
88 rt_b->rt_period_active = 0;
89 raw_spin_unlock(&rt_b->rt_runtime_lock);
90
91 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
92}
93
94void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
95{
96 rt_b->rt_period = ns_to_ktime(period);
97 rt_b->rt_runtime = runtime;
98
99 raw_spin_lock_init(&rt_b->rt_runtime_lock);
100
101 hrtimer_init(&rt_b->rt_period_timer, CLOCK_MONOTONIC,
102 HRTIMER_MODE_REL_HARD);
103 rt_b->rt_period_timer.function = sched_rt_period_timer;
104}
105
106static inline void do_start_rt_bandwidth(struct rt_bandwidth *rt_b)
107{
108 raw_spin_lock(&rt_b->rt_runtime_lock);
109 if (!rt_b->rt_period_active) {
110 rt_b->rt_period_active = 1;
111 /*
112 * SCHED_DEADLINE updates the bandwidth, as a run away
113 * RT task with a DL task could hog a CPU. But DL does
114 * not reset the period. If a deadline task was running
115 * without an RT task running, it can cause RT tasks to
116 * throttle when they start up. Kick the timer right away
117 * to update the period.
118 */
119 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
120 hrtimer_start_expires(&rt_b->rt_period_timer,
121 HRTIMER_MODE_ABS_PINNED_HARD);
122 }
123 raw_spin_unlock(&rt_b->rt_runtime_lock);
124}
125
126static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
127{
128 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
129 return;
130
131 do_start_rt_bandwidth(rt_b);
132}
133
134void init_rt_rq(struct rt_rq *rt_rq)
135{
136 struct rt_prio_array *array;
137 int i;
138
139 array = &rt_rq->active;
140 for (i = 0; i < MAX_RT_PRIO; i++) {
141 INIT_LIST_HEAD(array->queue + i);
142 __clear_bit(i, array->bitmap);
143 }
144 /* delimiter for bitsearch: */
145 __set_bit(MAX_RT_PRIO, array->bitmap);
146
147#if defined CONFIG_SMP
148 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
149 rt_rq->highest_prio.next = MAX_RT_PRIO-1;
150 rt_rq->overloaded = 0;
151 plist_head_init(&rt_rq->pushable_tasks);
152#endif /* CONFIG_SMP */
153 /* We start is dequeued state, because no RT tasks are queued */
154 rt_rq->rt_queued = 0;
155
156 rt_rq->rt_time = 0;
157 rt_rq->rt_throttled = 0;
158 rt_rq->rt_runtime = 0;
159 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
160}
161
162#ifdef CONFIG_RT_GROUP_SCHED
163static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
164{
165 hrtimer_cancel(&rt_b->rt_period_timer);
166}
167
168#define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
169
170static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
171{
172#ifdef CONFIG_SCHED_DEBUG
173 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
174#endif
175 return container_of(rt_se, struct task_struct, rt);
176}
177
178static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
179{
180 return rt_rq->rq;
181}
182
183static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
184{
185 return rt_se->rt_rq;
186}
187
188static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
189{
190 struct rt_rq *rt_rq = rt_se->rt_rq;
191
192 return rt_rq->rq;
193}
194
195void unregister_rt_sched_group(struct task_group *tg)
196{
197 if (tg->rt_se)
198 destroy_rt_bandwidth(&tg->rt_bandwidth);
199
200}
201
202void free_rt_sched_group(struct task_group *tg)
203{
204 int i;
205
206 for_each_possible_cpu(i) {
207 if (tg->rt_rq)
208 kfree(tg->rt_rq[i]);
209 if (tg->rt_se)
210 kfree(tg->rt_se[i]);
211 }
212
213 kfree(tg->rt_rq);
214 kfree(tg->rt_se);
215}
216
217void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
218 struct sched_rt_entity *rt_se, int cpu,
219 struct sched_rt_entity *parent)
220{
221 struct rq *rq = cpu_rq(cpu);
222
223 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
224 rt_rq->rt_nr_boosted = 0;
225 rt_rq->rq = rq;
226 rt_rq->tg = tg;
227
228 tg->rt_rq[cpu] = rt_rq;
229 tg->rt_se[cpu] = rt_se;
230
231 if (!rt_se)
232 return;
233
234 if (!parent)
235 rt_se->rt_rq = &rq->rt;
236 else
237 rt_se->rt_rq = parent->my_q;
238
239 rt_se->my_q = rt_rq;
240 rt_se->parent = parent;
241 INIT_LIST_HEAD(&rt_se->run_list);
242}
243
244int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
245{
246 struct rt_rq *rt_rq;
247 struct sched_rt_entity *rt_se;
248 int i;
249
250 tg->rt_rq = kcalloc(nr_cpu_ids, sizeof(rt_rq), GFP_KERNEL);
251 if (!tg->rt_rq)
252 goto err;
253 tg->rt_se = kcalloc(nr_cpu_ids, sizeof(rt_se), GFP_KERNEL);
254 if (!tg->rt_se)
255 goto err;
256
257 init_rt_bandwidth(&tg->rt_bandwidth,
258 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
259
260 for_each_possible_cpu(i) {
261 rt_rq = kzalloc_node(sizeof(struct rt_rq),
262 GFP_KERNEL, cpu_to_node(i));
263 if (!rt_rq)
264 goto err;
265
266 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
267 GFP_KERNEL, cpu_to_node(i));
268 if (!rt_se)
269 goto err_free_rq;
270
271 init_rt_rq(rt_rq);
272 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
273 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
274 }
275
276 return 1;
277
278err_free_rq:
279 kfree(rt_rq);
280err:
281 return 0;
282}
283
284#else /* CONFIG_RT_GROUP_SCHED */
285
286#define rt_entity_is_task(rt_se) (1)
287
288static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
289{
290 return container_of(rt_se, struct task_struct, rt);
291}
292
293static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
294{
295 return container_of(rt_rq, struct rq, rt);
296}
297
298static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
299{
300 struct task_struct *p = rt_task_of(rt_se);
301
302 return task_rq(p);
303}
304
305static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
306{
307 struct rq *rq = rq_of_rt_se(rt_se);
308
309 return &rq->rt;
310}
311
312void unregister_rt_sched_group(struct task_group *tg) { }
313
314void free_rt_sched_group(struct task_group *tg) { }
315
316int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
317{
318 return 1;
319}
320#endif /* CONFIG_RT_GROUP_SCHED */
321
322#ifdef CONFIG_SMP
323
324static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
325{
326 /* Try to pull RT tasks here if we lower this rq's prio */
327 return rq->online && rq->rt.highest_prio.curr > prev->prio;
328}
329
330static inline int rt_overloaded(struct rq *rq)
331{
332 return atomic_read(&rq->rd->rto_count);
333}
334
335static inline void rt_set_overload(struct rq *rq)
336{
337 if (!rq->online)
338 return;
339
340 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
341 /*
342 * Make sure the mask is visible before we set
343 * the overload count. That is checked to determine
344 * if we should look at the mask. It would be a shame
345 * if we looked at the mask, but the mask was not
346 * updated yet.
347 *
348 * Matched by the barrier in pull_rt_task().
349 */
350 smp_wmb();
351 atomic_inc(&rq->rd->rto_count);
352}
353
354static inline void rt_clear_overload(struct rq *rq)
355{
356 if (!rq->online)
357 return;
358
359 /* the order here really doesn't matter */
360 atomic_dec(&rq->rd->rto_count);
361 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
362}
363
364static inline int has_pushable_tasks(struct rq *rq)
365{
366 return !plist_head_empty(&rq->rt.pushable_tasks);
367}
368
369static DEFINE_PER_CPU(struct balance_callback, rt_push_head);
370static DEFINE_PER_CPU(struct balance_callback, rt_pull_head);
371
372static void push_rt_tasks(struct rq *);
373static void pull_rt_task(struct rq *);
374
375static inline void rt_queue_push_tasks(struct rq *rq)
376{
377 if (!has_pushable_tasks(rq))
378 return;
379
380 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
381}
382
383static inline void rt_queue_pull_task(struct rq *rq)
384{
385 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
386}
387
388static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
389{
390 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
391 plist_node_init(&p->pushable_tasks, p->prio);
392 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
393
394 /* Update the highest prio pushable task */
395 if (p->prio < rq->rt.highest_prio.next)
396 rq->rt.highest_prio.next = p->prio;
397
398 if (!rq->rt.overloaded) {
399 rt_set_overload(rq);
400 rq->rt.overloaded = 1;
401 }
402}
403
404static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
405{
406 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
407
408 /* Update the new highest prio pushable task */
409 if (has_pushable_tasks(rq)) {
410 p = plist_first_entry(&rq->rt.pushable_tasks,
411 struct task_struct, pushable_tasks);
412 rq->rt.highest_prio.next = p->prio;
413 } else {
414 rq->rt.highest_prio.next = MAX_RT_PRIO-1;
415
416 if (rq->rt.overloaded) {
417 rt_clear_overload(rq);
418 rq->rt.overloaded = 0;
419 }
420 }
421}
422
423#else
424
425static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
426{
427}
428
429static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
430{
431}
432
433static inline void rt_queue_push_tasks(struct rq *rq)
434{
435}
436#endif /* CONFIG_SMP */
437
438static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
439static void dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count);
440
441static inline int on_rt_rq(struct sched_rt_entity *rt_se)
442{
443 return rt_se->on_rq;
444}
445
446#ifdef CONFIG_UCLAMP_TASK
447/*
448 * Verify the fitness of task @p to run on @cpu taking into account the uclamp
449 * settings.
450 *
451 * This check is only important for heterogeneous systems where uclamp_min value
452 * is higher than the capacity of a @cpu. For non-heterogeneous system this
453 * function will always return true.
454 *
455 * The function will return true if the capacity of the @cpu is >= the
456 * uclamp_min and false otherwise.
457 *
458 * Note that uclamp_min will be clamped to uclamp_max if uclamp_min
459 * > uclamp_max.
460 */
461static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
462{
463 unsigned int min_cap;
464 unsigned int max_cap;
465 unsigned int cpu_cap;
466
467 /* Only heterogeneous systems can benefit from this check */
468 if (!sched_asym_cpucap_active())
469 return true;
470
471 min_cap = uclamp_eff_value(p, UCLAMP_MIN);
472 max_cap = uclamp_eff_value(p, UCLAMP_MAX);
473
474 cpu_cap = arch_scale_cpu_capacity(cpu);
475
476 return cpu_cap >= min(min_cap, max_cap);
477}
478#else
479static inline bool rt_task_fits_capacity(struct task_struct *p, int cpu)
480{
481 return true;
482}
483#endif
484
485#ifdef CONFIG_RT_GROUP_SCHED
486
487static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
488{
489 if (!rt_rq->tg)
490 return RUNTIME_INF;
491
492 return rt_rq->rt_runtime;
493}
494
495static inline u64 sched_rt_period(struct rt_rq *rt_rq)
496{
497 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
498}
499
500typedef struct task_group *rt_rq_iter_t;
501
502static inline struct task_group *next_task_group(struct task_group *tg)
503{
504 do {
505 tg = list_entry_rcu(tg->list.next,
506 typeof(struct task_group), list);
507 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
508
509 if (&tg->list == &task_groups)
510 tg = NULL;
511
512 return tg;
513}
514
515#define for_each_rt_rq(rt_rq, iter, rq) \
516 for (iter = container_of(&task_groups, typeof(*iter), list); \
517 (iter = next_task_group(iter)) && \
518 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
519
520#define for_each_sched_rt_entity(rt_se) \
521 for (; rt_se; rt_se = rt_se->parent)
522
523static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
524{
525 return rt_se->my_q;
526}
527
528static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
529static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
530
531static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
532{
533 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
534 struct rq *rq = rq_of_rt_rq(rt_rq);
535 struct sched_rt_entity *rt_se;
536
537 int cpu = cpu_of(rq);
538
539 rt_se = rt_rq->tg->rt_se[cpu];
540
541 if (rt_rq->rt_nr_running) {
542 if (!rt_se)
543 enqueue_top_rt_rq(rt_rq);
544 else if (!on_rt_rq(rt_se))
545 enqueue_rt_entity(rt_se, 0);
546
547 if (rt_rq->highest_prio.curr < curr->prio)
548 resched_curr(rq);
549 }
550}
551
552static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
553{
554 struct sched_rt_entity *rt_se;
555 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
556
557 rt_se = rt_rq->tg->rt_se[cpu];
558
559 if (!rt_se) {
560 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
561 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
562 cpufreq_update_util(rq_of_rt_rq(rt_rq), 0);
563 }
564 else if (on_rt_rq(rt_se))
565 dequeue_rt_entity(rt_se, 0);
566}
567
568static inline int rt_rq_throttled(struct rt_rq *rt_rq)
569{
570 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
571}
572
573static int rt_se_boosted(struct sched_rt_entity *rt_se)
574{
575 struct rt_rq *rt_rq = group_rt_rq(rt_se);
576 struct task_struct *p;
577
578 if (rt_rq)
579 return !!rt_rq->rt_nr_boosted;
580
581 p = rt_task_of(rt_se);
582 return p->prio != p->normal_prio;
583}
584
585#ifdef CONFIG_SMP
586static inline const struct cpumask *sched_rt_period_mask(void)
587{
588 return this_rq()->rd->span;
589}
590#else
591static inline const struct cpumask *sched_rt_period_mask(void)
592{
593 return cpu_online_mask;
594}
595#endif
596
597static inline
598struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
599{
600 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
601}
602
603static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
604{
605 return &rt_rq->tg->rt_bandwidth;
606}
607
608#else /* !CONFIG_RT_GROUP_SCHED */
609
610static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
611{
612 return rt_rq->rt_runtime;
613}
614
615static inline u64 sched_rt_period(struct rt_rq *rt_rq)
616{
617 return ktime_to_ns(def_rt_bandwidth.rt_period);
618}
619
620typedef struct rt_rq *rt_rq_iter_t;
621
622#define for_each_rt_rq(rt_rq, iter, rq) \
623 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
624
625#define for_each_sched_rt_entity(rt_se) \
626 for (; rt_se; rt_se = NULL)
627
628static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
629{
630 return NULL;
631}
632
633static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
634{
635 struct rq *rq = rq_of_rt_rq(rt_rq);
636
637 if (!rt_rq->rt_nr_running)
638 return;
639
640 enqueue_top_rt_rq(rt_rq);
641 resched_curr(rq);
642}
643
644static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
645{
646 dequeue_top_rt_rq(rt_rq, rt_rq->rt_nr_running);
647}
648
649static inline int rt_rq_throttled(struct rt_rq *rt_rq)
650{
651 return rt_rq->rt_throttled;
652}
653
654static inline const struct cpumask *sched_rt_period_mask(void)
655{
656 return cpu_online_mask;
657}
658
659static inline
660struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
661{
662 return &cpu_rq(cpu)->rt;
663}
664
665static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
666{
667 return &def_rt_bandwidth;
668}
669
670#endif /* CONFIG_RT_GROUP_SCHED */
671
672bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
673{
674 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
675
676 return (hrtimer_active(&rt_b->rt_period_timer) ||
677 rt_rq->rt_time < rt_b->rt_runtime);
678}
679
680#ifdef CONFIG_SMP
681/*
682 * We ran out of runtime, see if we can borrow some from our neighbours.
683 */
684static void do_balance_runtime(struct rt_rq *rt_rq)
685{
686 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
687 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
688 int i, weight;
689 u64 rt_period;
690
691 weight = cpumask_weight(rd->span);
692
693 raw_spin_lock(&rt_b->rt_runtime_lock);
694 rt_period = ktime_to_ns(rt_b->rt_period);
695 for_each_cpu(i, rd->span) {
696 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
697 s64 diff;
698
699 if (iter == rt_rq)
700 continue;
701
702 raw_spin_lock(&iter->rt_runtime_lock);
703 /*
704 * Either all rqs have inf runtime and there's nothing to steal
705 * or __disable_runtime() below sets a specific rq to inf to
706 * indicate its been disabled and disallow stealing.
707 */
708 if (iter->rt_runtime == RUNTIME_INF)
709 goto next;
710
711 /*
712 * From runqueues with spare time, take 1/n part of their
713 * spare time, but no more than our period.
714 */
715 diff = iter->rt_runtime - iter->rt_time;
716 if (diff > 0) {
717 diff = div_u64((u64)diff, weight);
718 if (rt_rq->rt_runtime + diff > rt_period)
719 diff = rt_period - rt_rq->rt_runtime;
720 iter->rt_runtime -= diff;
721 rt_rq->rt_runtime += diff;
722 if (rt_rq->rt_runtime == rt_period) {
723 raw_spin_unlock(&iter->rt_runtime_lock);
724 break;
725 }
726 }
727next:
728 raw_spin_unlock(&iter->rt_runtime_lock);
729 }
730 raw_spin_unlock(&rt_b->rt_runtime_lock);
731}
732
733/*
734 * Ensure this RQ takes back all the runtime it lend to its neighbours.
735 */
736static void __disable_runtime(struct rq *rq)
737{
738 struct root_domain *rd = rq->rd;
739 rt_rq_iter_t iter;
740 struct rt_rq *rt_rq;
741
742 if (unlikely(!scheduler_running))
743 return;
744
745 for_each_rt_rq(rt_rq, iter, rq) {
746 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
747 s64 want;
748 int i;
749
750 raw_spin_lock(&rt_b->rt_runtime_lock);
751 raw_spin_lock(&rt_rq->rt_runtime_lock);
752 /*
753 * Either we're all inf and nobody needs to borrow, or we're
754 * already disabled and thus have nothing to do, or we have
755 * exactly the right amount of runtime to take out.
756 */
757 if (rt_rq->rt_runtime == RUNTIME_INF ||
758 rt_rq->rt_runtime == rt_b->rt_runtime)
759 goto balanced;
760 raw_spin_unlock(&rt_rq->rt_runtime_lock);
761
762 /*
763 * Calculate the difference between what we started out with
764 * and what we current have, that's the amount of runtime
765 * we lend and now have to reclaim.
766 */
767 want = rt_b->rt_runtime - rt_rq->rt_runtime;
768
769 /*
770 * Greedy reclaim, take back as much as we can.
771 */
772 for_each_cpu(i, rd->span) {
773 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
774 s64 diff;
775
776 /*
777 * Can't reclaim from ourselves or disabled runqueues.
778 */
779 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
780 continue;
781
782 raw_spin_lock(&iter->rt_runtime_lock);
783 if (want > 0) {
784 diff = min_t(s64, iter->rt_runtime, want);
785 iter->rt_runtime -= diff;
786 want -= diff;
787 } else {
788 iter->rt_runtime -= want;
789 want -= want;
790 }
791 raw_spin_unlock(&iter->rt_runtime_lock);
792
793 if (!want)
794 break;
795 }
796
797 raw_spin_lock(&rt_rq->rt_runtime_lock);
798 /*
799 * We cannot be left wanting - that would mean some runtime
800 * leaked out of the system.
801 */
802 WARN_ON_ONCE(want);
803balanced:
804 /*
805 * Disable all the borrow logic by pretending we have inf
806 * runtime - in which case borrowing doesn't make sense.
807 */
808 rt_rq->rt_runtime = RUNTIME_INF;
809 rt_rq->rt_throttled = 0;
810 raw_spin_unlock(&rt_rq->rt_runtime_lock);
811 raw_spin_unlock(&rt_b->rt_runtime_lock);
812
813 /* Make rt_rq available for pick_next_task() */
814 sched_rt_rq_enqueue(rt_rq);
815 }
816}
817
818static void __enable_runtime(struct rq *rq)
819{
820 rt_rq_iter_t iter;
821 struct rt_rq *rt_rq;
822
823 if (unlikely(!scheduler_running))
824 return;
825
826 /*
827 * Reset each runqueue's bandwidth settings
828 */
829 for_each_rt_rq(rt_rq, iter, rq) {
830 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
831
832 raw_spin_lock(&rt_b->rt_runtime_lock);
833 raw_spin_lock(&rt_rq->rt_runtime_lock);
834 rt_rq->rt_runtime = rt_b->rt_runtime;
835 rt_rq->rt_time = 0;
836 rt_rq->rt_throttled = 0;
837 raw_spin_unlock(&rt_rq->rt_runtime_lock);
838 raw_spin_unlock(&rt_b->rt_runtime_lock);
839 }
840}
841
842static void balance_runtime(struct rt_rq *rt_rq)
843{
844 if (!sched_feat(RT_RUNTIME_SHARE))
845 return;
846
847 if (rt_rq->rt_time > rt_rq->rt_runtime) {
848 raw_spin_unlock(&rt_rq->rt_runtime_lock);
849 do_balance_runtime(rt_rq);
850 raw_spin_lock(&rt_rq->rt_runtime_lock);
851 }
852}
853#else /* !CONFIG_SMP */
854static inline void balance_runtime(struct rt_rq *rt_rq) {}
855#endif /* CONFIG_SMP */
856
857static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
858{
859 int i, idle = 1, throttled = 0;
860 const struct cpumask *span;
861
862 span = sched_rt_period_mask();
863#ifdef CONFIG_RT_GROUP_SCHED
864 /*
865 * FIXME: isolated CPUs should really leave the root task group,
866 * whether they are isolcpus or were isolated via cpusets, lest
867 * the timer run on a CPU which does not service all runqueues,
868 * potentially leaving other CPUs indefinitely throttled. If
869 * isolation is really required, the user will turn the throttle
870 * off to kill the perturbations it causes anyway. Meanwhile,
871 * this maintains functionality for boot and/or troubleshooting.
872 */
873 if (rt_b == &root_task_group.rt_bandwidth)
874 span = cpu_online_mask;
875#endif
876 for_each_cpu(i, span) {
877 int enqueue = 0;
878 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
879 struct rq *rq = rq_of_rt_rq(rt_rq);
880 struct rq_flags rf;
881 int skip;
882
883 /*
884 * When span == cpu_online_mask, taking each rq->lock
885 * can be time-consuming. Try to avoid it when possible.
886 */
887 raw_spin_lock(&rt_rq->rt_runtime_lock);
888 if (!sched_feat(RT_RUNTIME_SHARE) && rt_rq->rt_runtime != RUNTIME_INF)
889 rt_rq->rt_runtime = rt_b->rt_runtime;
890 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
891 raw_spin_unlock(&rt_rq->rt_runtime_lock);
892 if (skip)
893 continue;
894
895 rq_lock(rq, &rf);
896 update_rq_clock(rq);
897
898 if (rt_rq->rt_time) {
899 u64 runtime;
900
901 raw_spin_lock(&rt_rq->rt_runtime_lock);
902 if (rt_rq->rt_throttled)
903 balance_runtime(rt_rq);
904 runtime = rt_rq->rt_runtime;
905 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
906 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
907 rt_rq->rt_throttled = 0;
908 enqueue = 1;
909
910 /*
911 * When we're idle and a woken (rt) task is
912 * throttled wakeup_preempt() will set
913 * skip_update and the time between the wakeup
914 * and this unthrottle will get accounted as
915 * 'runtime'.
916 */
917 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
918 rq_clock_cancel_skipupdate(rq);
919 }
920 if (rt_rq->rt_time || rt_rq->rt_nr_running)
921 idle = 0;
922 raw_spin_unlock(&rt_rq->rt_runtime_lock);
923 } else if (rt_rq->rt_nr_running) {
924 idle = 0;
925 if (!rt_rq_throttled(rt_rq))
926 enqueue = 1;
927 }
928 if (rt_rq->rt_throttled)
929 throttled = 1;
930
931 if (enqueue)
932 sched_rt_rq_enqueue(rt_rq);
933 rq_unlock(rq, &rf);
934 }
935
936 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
937 return 1;
938
939 return idle;
940}
941
942static inline int rt_se_prio(struct sched_rt_entity *rt_se)
943{
944#ifdef CONFIG_RT_GROUP_SCHED
945 struct rt_rq *rt_rq = group_rt_rq(rt_se);
946
947 if (rt_rq)
948 return rt_rq->highest_prio.curr;
949#endif
950
951 return rt_task_of(rt_se)->prio;
952}
953
954static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
955{
956 u64 runtime = sched_rt_runtime(rt_rq);
957
958 if (rt_rq->rt_throttled)
959 return rt_rq_throttled(rt_rq);
960
961 if (runtime >= sched_rt_period(rt_rq))
962 return 0;
963
964 balance_runtime(rt_rq);
965 runtime = sched_rt_runtime(rt_rq);
966 if (runtime == RUNTIME_INF)
967 return 0;
968
969 if (rt_rq->rt_time > runtime) {
970 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
971
972 /*
973 * Don't actually throttle groups that have no runtime assigned
974 * but accrue some time due to boosting.
975 */
976 if (likely(rt_b->rt_runtime)) {
977 rt_rq->rt_throttled = 1;
978 printk_deferred_once("sched: RT throttling activated\n");
979 } else {
980 /*
981 * In case we did anyway, make it go away,
982 * replenishment is a joke, since it will replenish us
983 * with exactly 0 ns.
984 */
985 rt_rq->rt_time = 0;
986 }
987
988 if (rt_rq_throttled(rt_rq)) {
989 sched_rt_rq_dequeue(rt_rq);
990 return 1;
991 }
992 }
993
994 return 0;
995}
996
997/*
998 * Update the current task's runtime statistics. Skip current tasks that
999 * are not in our scheduling class.
1000 */
1001static void update_curr_rt(struct rq *rq)
1002{
1003 struct task_struct *curr = rq->curr;
1004 struct sched_rt_entity *rt_se = &curr->rt;
1005 s64 delta_exec;
1006
1007 if (curr->sched_class != &rt_sched_class)
1008 return;
1009
1010 delta_exec = update_curr_common(rq);
1011 if (unlikely(delta_exec <= 0))
1012 return;
1013
1014 if (!rt_bandwidth_enabled())
1015 return;
1016
1017 for_each_sched_rt_entity(rt_se) {
1018 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1019 int exceeded;
1020
1021 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
1022 raw_spin_lock(&rt_rq->rt_runtime_lock);
1023 rt_rq->rt_time += delta_exec;
1024 exceeded = sched_rt_runtime_exceeded(rt_rq);
1025 if (exceeded)
1026 resched_curr(rq);
1027 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1028 if (exceeded)
1029 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq));
1030 }
1031 }
1032}
1033
1034static void
1035dequeue_top_rt_rq(struct rt_rq *rt_rq, unsigned int count)
1036{
1037 struct rq *rq = rq_of_rt_rq(rt_rq);
1038
1039 BUG_ON(&rq->rt != rt_rq);
1040
1041 if (!rt_rq->rt_queued)
1042 return;
1043
1044 BUG_ON(!rq->nr_running);
1045
1046 sub_nr_running(rq, count);
1047 rt_rq->rt_queued = 0;
1048
1049}
1050
1051static void
1052enqueue_top_rt_rq(struct rt_rq *rt_rq)
1053{
1054 struct rq *rq = rq_of_rt_rq(rt_rq);
1055
1056 BUG_ON(&rq->rt != rt_rq);
1057
1058 if (rt_rq->rt_queued)
1059 return;
1060
1061 if (rt_rq_throttled(rt_rq))
1062 return;
1063
1064 if (rt_rq->rt_nr_running) {
1065 add_nr_running(rq, rt_rq->rt_nr_running);
1066 rt_rq->rt_queued = 1;
1067 }
1068
1069 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1070 cpufreq_update_util(rq, 0);
1071}
1072
1073#if defined CONFIG_SMP
1074
1075static void
1076inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1077{
1078 struct rq *rq = rq_of_rt_rq(rt_rq);
1079
1080#ifdef CONFIG_RT_GROUP_SCHED
1081 /*
1082 * Change rq's cpupri only if rt_rq is the top queue.
1083 */
1084 if (&rq->rt != rt_rq)
1085 return;
1086#endif
1087 if (rq->online && prio < prev_prio)
1088 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1089}
1090
1091static void
1092dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1093{
1094 struct rq *rq = rq_of_rt_rq(rt_rq);
1095
1096#ifdef CONFIG_RT_GROUP_SCHED
1097 /*
1098 * Change rq's cpupri only if rt_rq is the top queue.
1099 */
1100 if (&rq->rt != rt_rq)
1101 return;
1102#endif
1103 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1104 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1105}
1106
1107#else /* CONFIG_SMP */
1108
1109static inline
1110void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1111static inline
1112void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1113
1114#endif /* CONFIG_SMP */
1115
1116#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1117static void
1118inc_rt_prio(struct rt_rq *rt_rq, int prio)
1119{
1120 int prev_prio = rt_rq->highest_prio.curr;
1121
1122 if (prio < prev_prio)
1123 rt_rq->highest_prio.curr = prio;
1124
1125 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1126}
1127
1128static void
1129dec_rt_prio(struct rt_rq *rt_rq, int prio)
1130{
1131 int prev_prio = rt_rq->highest_prio.curr;
1132
1133 if (rt_rq->rt_nr_running) {
1134
1135 WARN_ON(prio < prev_prio);
1136
1137 /*
1138 * This may have been our highest task, and therefore
1139 * we may have some recomputation to do
1140 */
1141 if (prio == prev_prio) {
1142 struct rt_prio_array *array = &rt_rq->active;
1143
1144 rt_rq->highest_prio.curr =
1145 sched_find_first_bit(array->bitmap);
1146 }
1147
1148 } else {
1149 rt_rq->highest_prio.curr = MAX_RT_PRIO-1;
1150 }
1151
1152 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1153}
1154
1155#else
1156
1157static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1158static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1159
1160#endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1161
1162#ifdef CONFIG_RT_GROUP_SCHED
1163
1164static void
1165inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1166{
1167 if (rt_se_boosted(rt_se))
1168 rt_rq->rt_nr_boosted++;
1169
1170 if (rt_rq->tg)
1171 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1172}
1173
1174static void
1175dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1176{
1177 if (rt_se_boosted(rt_se))
1178 rt_rq->rt_nr_boosted--;
1179
1180 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1181}
1182
1183#else /* CONFIG_RT_GROUP_SCHED */
1184
1185static void
1186inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1187{
1188 start_rt_bandwidth(&def_rt_bandwidth);
1189}
1190
1191static inline
1192void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1193
1194#endif /* CONFIG_RT_GROUP_SCHED */
1195
1196static inline
1197unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1198{
1199 struct rt_rq *group_rq = group_rt_rq(rt_se);
1200
1201 if (group_rq)
1202 return group_rq->rt_nr_running;
1203 else
1204 return 1;
1205}
1206
1207static inline
1208unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1209{
1210 struct rt_rq *group_rq = group_rt_rq(rt_se);
1211 struct task_struct *tsk;
1212
1213 if (group_rq)
1214 return group_rq->rr_nr_running;
1215
1216 tsk = rt_task_of(rt_se);
1217
1218 return (tsk->policy == SCHED_RR) ? 1 : 0;
1219}
1220
1221static inline
1222void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1223{
1224 int prio = rt_se_prio(rt_se);
1225
1226 WARN_ON(!rt_prio(prio));
1227 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1228 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1229
1230 inc_rt_prio(rt_rq, prio);
1231 inc_rt_group(rt_se, rt_rq);
1232}
1233
1234static inline
1235void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1236{
1237 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1238 WARN_ON(!rt_rq->rt_nr_running);
1239 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1240 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1241
1242 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1243 dec_rt_group(rt_se, rt_rq);
1244}
1245
1246/*
1247 * Change rt_se->run_list location unless SAVE && !MOVE
1248 *
1249 * assumes ENQUEUE/DEQUEUE flags match
1250 */
1251static inline bool move_entity(unsigned int flags)
1252{
1253 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1254 return false;
1255
1256 return true;
1257}
1258
1259static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1260{
1261 list_del_init(&rt_se->run_list);
1262
1263 if (list_empty(array->queue + rt_se_prio(rt_se)))
1264 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1265
1266 rt_se->on_list = 0;
1267}
1268
1269static inline struct sched_statistics *
1270__schedstats_from_rt_se(struct sched_rt_entity *rt_se)
1271{
1272#ifdef CONFIG_RT_GROUP_SCHED
1273 /* schedstats is not supported for rt group. */
1274 if (!rt_entity_is_task(rt_se))
1275 return NULL;
1276#endif
1277
1278 return &rt_task_of(rt_se)->stats;
1279}
1280
1281static inline void
1282update_stats_wait_start_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1283{
1284 struct sched_statistics *stats;
1285 struct task_struct *p = NULL;
1286
1287 if (!schedstat_enabled())
1288 return;
1289
1290 if (rt_entity_is_task(rt_se))
1291 p = rt_task_of(rt_se);
1292
1293 stats = __schedstats_from_rt_se(rt_se);
1294 if (!stats)
1295 return;
1296
1297 __update_stats_wait_start(rq_of_rt_rq(rt_rq), p, stats);
1298}
1299
1300static inline void
1301update_stats_enqueue_sleeper_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1302{
1303 struct sched_statistics *stats;
1304 struct task_struct *p = NULL;
1305
1306 if (!schedstat_enabled())
1307 return;
1308
1309 if (rt_entity_is_task(rt_se))
1310 p = rt_task_of(rt_se);
1311
1312 stats = __schedstats_from_rt_se(rt_se);
1313 if (!stats)
1314 return;
1315
1316 __update_stats_enqueue_sleeper(rq_of_rt_rq(rt_rq), p, stats);
1317}
1318
1319static inline void
1320update_stats_enqueue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1321 int flags)
1322{
1323 if (!schedstat_enabled())
1324 return;
1325
1326 if (flags & ENQUEUE_WAKEUP)
1327 update_stats_enqueue_sleeper_rt(rt_rq, rt_se);
1328}
1329
1330static inline void
1331update_stats_wait_end_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
1332{
1333 struct sched_statistics *stats;
1334 struct task_struct *p = NULL;
1335
1336 if (!schedstat_enabled())
1337 return;
1338
1339 if (rt_entity_is_task(rt_se))
1340 p = rt_task_of(rt_se);
1341
1342 stats = __schedstats_from_rt_se(rt_se);
1343 if (!stats)
1344 return;
1345
1346 __update_stats_wait_end(rq_of_rt_rq(rt_rq), p, stats);
1347}
1348
1349static inline void
1350update_stats_dequeue_rt(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se,
1351 int flags)
1352{
1353 struct task_struct *p = NULL;
1354
1355 if (!schedstat_enabled())
1356 return;
1357
1358 if (rt_entity_is_task(rt_se))
1359 p = rt_task_of(rt_se);
1360
1361 if ((flags & DEQUEUE_SLEEP) && p) {
1362 unsigned int state;
1363
1364 state = READ_ONCE(p->__state);
1365 if (state & TASK_INTERRUPTIBLE)
1366 __schedstat_set(p->stats.sleep_start,
1367 rq_clock(rq_of_rt_rq(rt_rq)));
1368
1369 if (state & TASK_UNINTERRUPTIBLE)
1370 __schedstat_set(p->stats.block_start,
1371 rq_clock(rq_of_rt_rq(rt_rq)));
1372 }
1373}
1374
1375static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1376{
1377 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1378 struct rt_prio_array *array = &rt_rq->active;
1379 struct rt_rq *group_rq = group_rt_rq(rt_se);
1380 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1381
1382 /*
1383 * Don't enqueue the group if its throttled, or when empty.
1384 * The latter is a consequence of the former when a child group
1385 * get throttled and the current group doesn't have any other
1386 * active members.
1387 */
1388 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1389 if (rt_se->on_list)
1390 __delist_rt_entity(rt_se, array);
1391 return;
1392 }
1393
1394 if (move_entity(flags)) {
1395 WARN_ON_ONCE(rt_se->on_list);
1396 if (flags & ENQUEUE_HEAD)
1397 list_add(&rt_se->run_list, queue);
1398 else
1399 list_add_tail(&rt_se->run_list, queue);
1400
1401 __set_bit(rt_se_prio(rt_se), array->bitmap);
1402 rt_se->on_list = 1;
1403 }
1404 rt_se->on_rq = 1;
1405
1406 inc_rt_tasks(rt_se, rt_rq);
1407}
1408
1409static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1410{
1411 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1412 struct rt_prio_array *array = &rt_rq->active;
1413
1414 if (move_entity(flags)) {
1415 WARN_ON_ONCE(!rt_se->on_list);
1416 __delist_rt_entity(rt_se, array);
1417 }
1418 rt_se->on_rq = 0;
1419
1420 dec_rt_tasks(rt_se, rt_rq);
1421}
1422
1423/*
1424 * Because the prio of an upper entry depends on the lower
1425 * entries, we must remove entries top - down.
1426 */
1427static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1428{
1429 struct sched_rt_entity *back = NULL;
1430 unsigned int rt_nr_running;
1431
1432 for_each_sched_rt_entity(rt_se) {
1433 rt_se->back = back;
1434 back = rt_se;
1435 }
1436
1437 rt_nr_running = rt_rq_of_se(back)->rt_nr_running;
1438
1439 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1440 if (on_rt_rq(rt_se))
1441 __dequeue_rt_entity(rt_se, flags);
1442 }
1443
1444 dequeue_top_rt_rq(rt_rq_of_se(back), rt_nr_running);
1445}
1446
1447static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1448{
1449 struct rq *rq = rq_of_rt_se(rt_se);
1450
1451 update_stats_enqueue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1452
1453 dequeue_rt_stack(rt_se, flags);
1454 for_each_sched_rt_entity(rt_se)
1455 __enqueue_rt_entity(rt_se, flags);
1456 enqueue_top_rt_rq(&rq->rt);
1457}
1458
1459static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1460{
1461 struct rq *rq = rq_of_rt_se(rt_se);
1462
1463 update_stats_dequeue_rt(rt_rq_of_se(rt_se), rt_se, flags);
1464
1465 dequeue_rt_stack(rt_se, flags);
1466
1467 for_each_sched_rt_entity(rt_se) {
1468 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1469
1470 if (rt_rq && rt_rq->rt_nr_running)
1471 __enqueue_rt_entity(rt_se, flags);
1472 }
1473 enqueue_top_rt_rq(&rq->rt);
1474}
1475
1476/*
1477 * Adding/removing a task to/from a priority array:
1478 */
1479static void
1480enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1481{
1482 struct sched_rt_entity *rt_se = &p->rt;
1483
1484 if (flags & ENQUEUE_WAKEUP)
1485 rt_se->timeout = 0;
1486
1487 check_schedstat_required();
1488 update_stats_wait_start_rt(rt_rq_of_se(rt_se), rt_se);
1489
1490 enqueue_rt_entity(rt_se, flags);
1491
1492 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1493 enqueue_pushable_task(rq, p);
1494}
1495
1496static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1497{
1498 struct sched_rt_entity *rt_se = &p->rt;
1499
1500 update_curr_rt(rq);
1501 dequeue_rt_entity(rt_se, flags);
1502
1503 dequeue_pushable_task(rq, p);
1504}
1505
1506/*
1507 * Put task to the head or the end of the run list without the overhead of
1508 * dequeue followed by enqueue.
1509 */
1510static void
1511requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1512{
1513 if (on_rt_rq(rt_se)) {
1514 struct rt_prio_array *array = &rt_rq->active;
1515 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1516
1517 if (head)
1518 list_move(&rt_se->run_list, queue);
1519 else
1520 list_move_tail(&rt_se->run_list, queue);
1521 }
1522}
1523
1524static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1525{
1526 struct sched_rt_entity *rt_se = &p->rt;
1527 struct rt_rq *rt_rq;
1528
1529 for_each_sched_rt_entity(rt_se) {
1530 rt_rq = rt_rq_of_se(rt_se);
1531 requeue_rt_entity(rt_rq, rt_se, head);
1532 }
1533}
1534
1535static void yield_task_rt(struct rq *rq)
1536{
1537 requeue_task_rt(rq, rq->curr, 0);
1538}
1539
1540#ifdef CONFIG_SMP
1541static int find_lowest_rq(struct task_struct *task);
1542
1543static int
1544select_task_rq_rt(struct task_struct *p, int cpu, int flags)
1545{
1546 struct task_struct *curr;
1547 struct rq *rq;
1548 bool test;
1549
1550 /* For anything but wake ups, just return the task_cpu */
1551 if (!(flags & (WF_TTWU | WF_FORK)))
1552 goto out;
1553
1554 rq = cpu_rq(cpu);
1555
1556 rcu_read_lock();
1557 curr = READ_ONCE(rq->curr); /* unlocked access */
1558
1559 /*
1560 * If the current task on @p's runqueue is an RT task, then
1561 * try to see if we can wake this RT task up on another
1562 * runqueue. Otherwise simply start this RT task
1563 * on its current runqueue.
1564 *
1565 * We want to avoid overloading runqueues. If the woken
1566 * task is a higher priority, then it will stay on this CPU
1567 * and the lower prio task should be moved to another CPU.
1568 * Even though this will probably make the lower prio task
1569 * lose its cache, we do not want to bounce a higher task
1570 * around just because it gave up its CPU, perhaps for a
1571 * lock?
1572 *
1573 * For equal prio tasks, we just let the scheduler sort it out.
1574 *
1575 * Otherwise, just let it ride on the affined RQ and the
1576 * post-schedule router will push the preempted task away
1577 *
1578 * This test is optimistic, if we get it wrong the load-balancer
1579 * will have to sort it out.
1580 *
1581 * We take into account the capacity of the CPU to ensure it fits the
1582 * requirement of the task - which is only important on heterogeneous
1583 * systems like big.LITTLE.
1584 */
1585 test = curr &&
1586 unlikely(rt_task(curr)) &&
1587 (curr->nr_cpus_allowed < 2 || curr->prio <= p->prio);
1588
1589 if (test || !rt_task_fits_capacity(p, cpu)) {
1590 int target = find_lowest_rq(p);
1591
1592 /*
1593 * Bail out if we were forcing a migration to find a better
1594 * fitting CPU but our search failed.
1595 */
1596 if (!test && target != -1 && !rt_task_fits_capacity(p, target))
1597 goto out_unlock;
1598
1599 /*
1600 * Don't bother moving it if the destination CPU is
1601 * not running a lower priority task.
1602 */
1603 if (target != -1 &&
1604 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1605 cpu = target;
1606 }
1607
1608out_unlock:
1609 rcu_read_unlock();
1610
1611out:
1612 return cpu;
1613}
1614
1615static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1616{
1617 /*
1618 * Current can't be migrated, useless to reschedule,
1619 * let's hope p can move out.
1620 */
1621 if (rq->curr->nr_cpus_allowed == 1 ||
1622 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1623 return;
1624
1625 /*
1626 * p is migratable, so let's not schedule it and
1627 * see if it is pushed or pulled somewhere else.
1628 */
1629 if (p->nr_cpus_allowed != 1 &&
1630 cpupri_find(&rq->rd->cpupri, p, NULL))
1631 return;
1632
1633 /*
1634 * There appear to be other CPUs that can accept
1635 * the current task but none can run 'p', so lets reschedule
1636 * to try and push the current task away:
1637 */
1638 requeue_task_rt(rq, p, 1);
1639 resched_curr(rq);
1640}
1641
1642static int balance_rt(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1643{
1644 if (!on_rt_rq(&p->rt) && need_pull_rt_task(rq, p)) {
1645 /*
1646 * This is OK, because current is on_cpu, which avoids it being
1647 * picked for load-balance and preemption/IRQs are still
1648 * disabled avoiding further scheduler activity on it and we've
1649 * not yet started the picking loop.
1650 */
1651 rq_unpin_lock(rq, rf);
1652 pull_rt_task(rq);
1653 rq_repin_lock(rq, rf);
1654 }
1655
1656 return sched_stop_runnable(rq) || sched_dl_runnable(rq) || sched_rt_runnable(rq);
1657}
1658#endif /* CONFIG_SMP */
1659
1660/*
1661 * Preempt the current task with a newly woken task if needed:
1662 */
1663static void wakeup_preempt_rt(struct rq *rq, struct task_struct *p, int flags)
1664{
1665 if (p->prio < rq->curr->prio) {
1666 resched_curr(rq);
1667 return;
1668 }
1669
1670#ifdef CONFIG_SMP
1671 /*
1672 * If:
1673 *
1674 * - the newly woken task is of equal priority to the current task
1675 * - the newly woken task is non-migratable while current is migratable
1676 * - current will be preempted on the next reschedule
1677 *
1678 * we should check to see if current can readily move to a different
1679 * cpu. If so, we will reschedule to allow the push logic to try
1680 * to move current somewhere else, making room for our non-migratable
1681 * task.
1682 */
1683 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1684 check_preempt_equal_prio(rq, p);
1685#endif
1686}
1687
1688static inline void set_next_task_rt(struct rq *rq, struct task_struct *p, bool first)
1689{
1690 struct sched_rt_entity *rt_se = &p->rt;
1691 struct rt_rq *rt_rq = &rq->rt;
1692
1693 p->se.exec_start = rq_clock_task(rq);
1694 if (on_rt_rq(&p->rt))
1695 update_stats_wait_end_rt(rt_rq, rt_se);
1696
1697 /* The running task is never eligible for pushing */
1698 dequeue_pushable_task(rq, p);
1699
1700 if (!first)
1701 return;
1702
1703 /*
1704 * If prev task was rt, put_prev_task() has already updated the
1705 * utilization. We only care of the case where we start to schedule a
1706 * rt task
1707 */
1708 if (rq->curr->sched_class != &rt_sched_class)
1709 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1710
1711 rt_queue_push_tasks(rq);
1712}
1713
1714static struct sched_rt_entity *pick_next_rt_entity(struct rt_rq *rt_rq)
1715{
1716 struct rt_prio_array *array = &rt_rq->active;
1717 struct sched_rt_entity *next = NULL;
1718 struct list_head *queue;
1719 int idx;
1720
1721 idx = sched_find_first_bit(array->bitmap);
1722 BUG_ON(idx >= MAX_RT_PRIO);
1723
1724 queue = array->queue + idx;
1725 if (SCHED_WARN_ON(list_empty(queue)))
1726 return NULL;
1727 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1728
1729 return next;
1730}
1731
1732static struct task_struct *_pick_next_task_rt(struct rq *rq)
1733{
1734 struct sched_rt_entity *rt_se;
1735 struct rt_rq *rt_rq = &rq->rt;
1736
1737 do {
1738 rt_se = pick_next_rt_entity(rt_rq);
1739 if (unlikely(!rt_se))
1740 return NULL;
1741 rt_rq = group_rt_rq(rt_se);
1742 } while (rt_rq);
1743
1744 return rt_task_of(rt_se);
1745}
1746
1747static struct task_struct *pick_task_rt(struct rq *rq)
1748{
1749 struct task_struct *p;
1750
1751 if (!sched_rt_runnable(rq))
1752 return NULL;
1753
1754 p = _pick_next_task_rt(rq);
1755
1756 return p;
1757}
1758
1759static struct task_struct *pick_next_task_rt(struct rq *rq)
1760{
1761 struct task_struct *p = pick_task_rt(rq);
1762
1763 if (p)
1764 set_next_task_rt(rq, p, true);
1765
1766 return p;
1767}
1768
1769static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1770{
1771 struct sched_rt_entity *rt_se = &p->rt;
1772 struct rt_rq *rt_rq = &rq->rt;
1773
1774 if (on_rt_rq(&p->rt))
1775 update_stats_wait_start_rt(rt_rq, rt_se);
1776
1777 update_curr_rt(rq);
1778
1779 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1780
1781 /*
1782 * The previous task needs to be made eligible for pushing
1783 * if it is still active
1784 */
1785 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1786 enqueue_pushable_task(rq, p);
1787}
1788
1789#ifdef CONFIG_SMP
1790
1791/* Only try algorithms three times */
1792#define RT_MAX_TRIES 3
1793
1794static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1795{
1796 if (!task_on_cpu(rq, p) &&
1797 cpumask_test_cpu(cpu, &p->cpus_mask))
1798 return 1;
1799
1800 return 0;
1801}
1802
1803/*
1804 * Return the highest pushable rq's task, which is suitable to be executed
1805 * on the CPU, NULL otherwise
1806 */
1807static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1808{
1809 struct plist_head *head = &rq->rt.pushable_tasks;
1810 struct task_struct *p;
1811
1812 if (!has_pushable_tasks(rq))
1813 return NULL;
1814
1815 plist_for_each_entry(p, head, pushable_tasks) {
1816 if (pick_rt_task(rq, p, cpu))
1817 return p;
1818 }
1819
1820 return NULL;
1821}
1822
1823static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1824
1825static int find_lowest_rq(struct task_struct *task)
1826{
1827 struct sched_domain *sd;
1828 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1829 int this_cpu = smp_processor_id();
1830 int cpu = task_cpu(task);
1831 int ret;
1832
1833 /* Make sure the mask is initialized first */
1834 if (unlikely(!lowest_mask))
1835 return -1;
1836
1837 if (task->nr_cpus_allowed == 1)
1838 return -1; /* No other targets possible */
1839
1840 /*
1841 * If we're on asym system ensure we consider the different capacities
1842 * of the CPUs when searching for the lowest_mask.
1843 */
1844 if (sched_asym_cpucap_active()) {
1845
1846 ret = cpupri_find_fitness(&task_rq(task)->rd->cpupri,
1847 task, lowest_mask,
1848 rt_task_fits_capacity);
1849 } else {
1850
1851 ret = cpupri_find(&task_rq(task)->rd->cpupri,
1852 task, lowest_mask);
1853 }
1854
1855 if (!ret)
1856 return -1; /* No targets found */
1857
1858 /*
1859 * At this point we have built a mask of CPUs representing the
1860 * lowest priority tasks in the system. Now we want to elect
1861 * the best one based on our affinity and topology.
1862 *
1863 * We prioritize the last CPU that the task executed on since
1864 * it is most likely cache-hot in that location.
1865 */
1866 if (cpumask_test_cpu(cpu, lowest_mask))
1867 return cpu;
1868
1869 /*
1870 * Otherwise, we consult the sched_domains span maps to figure
1871 * out which CPU is logically closest to our hot cache data.
1872 */
1873 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1874 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1875
1876 rcu_read_lock();
1877 for_each_domain(cpu, sd) {
1878 if (sd->flags & SD_WAKE_AFFINE) {
1879 int best_cpu;
1880
1881 /*
1882 * "this_cpu" is cheaper to preempt than a
1883 * remote processor.
1884 */
1885 if (this_cpu != -1 &&
1886 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1887 rcu_read_unlock();
1888 return this_cpu;
1889 }
1890
1891 best_cpu = cpumask_any_and_distribute(lowest_mask,
1892 sched_domain_span(sd));
1893 if (best_cpu < nr_cpu_ids) {
1894 rcu_read_unlock();
1895 return best_cpu;
1896 }
1897 }
1898 }
1899 rcu_read_unlock();
1900
1901 /*
1902 * And finally, if there were no matches within the domains
1903 * just give the caller *something* to work with from the compatible
1904 * locations.
1905 */
1906 if (this_cpu != -1)
1907 return this_cpu;
1908
1909 cpu = cpumask_any_distribute(lowest_mask);
1910 if (cpu < nr_cpu_ids)
1911 return cpu;
1912
1913 return -1;
1914}
1915
1916/* Will lock the rq it finds */
1917static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1918{
1919 struct rq *lowest_rq = NULL;
1920 int tries;
1921 int cpu;
1922
1923 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1924 cpu = find_lowest_rq(task);
1925
1926 if ((cpu == -1) || (cpu == rq->cpu))
1927 break;
1928
1929 lowest_rq = cpu_rq(cpu);
1930
1931 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1932 /*
1933 * Target rq has tasks of equal or higher priority,
1934 * retrying does not release any lock and is unlikely
1935 * to yield a different result.
1936 */
1937 lowest_rq = NULL;
1938 break;
1939 }
1940
1941 /* if the prio of this runqueue changed, try again */
1942 if (double_lock_balance(rq, lowest_rq)) {
1943 /*
1944 * We had to unlock the run queue. In
1945 * the mean time, task could have
1946 * migrated already or had its affinity changed.
1947 * Also make sure that it wasn't scheduled on its rq.
1948 * It is possible the task was scheduled, set
1949 * "migrate_disabled" and then got preempted, so we must
1950 * check the task migration disable flag here too.
1951 */
1952 if (unlikely(task_rq(task) != rq ||
1953 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_mask) ||
1954 task_on_cpu(rq, task) ||
1955 !rt_task(task) ||
1956 is_migration_disabled(task) ||
1957 !task_on_rq_queued(task))) {
1958
1959 double_unlock_balance(rq, lowest_rq);
1960 lowest_rq = NULL;
1961 break;
1962 }
1963 }
1964
1965 /* If this rq is still suitable use it. */
1966 if (lowest_rq->rt.highest_prio.curr > task->prio)
1967 break;
1968
1969 /* try again */
1970 double_unlock_balance(rq, lowest_rq);
1971 lowest_rq = NULL;
1972 }
1973
1974 return lowest_rq;
1975}
1976
1977static struct task_struct *pick_next_pushable_task(struct rq *rq)
1978{
1979 struct task_struct *p;
1980
1981 if (!has_pushable_tasks(rq))
1982 return NULL;
1983
1984 p = plist_first_entry(&rq->rt.pushable_tasks,
1985 struct task_struct, pushable_tasks);
1986
1987 BUG_ON(rq->cpu != task_cpu(p));
1988 BUG_ON(task_current(rq, p));
1989 BUG_ON(p->nr_cpus_allowed <= 1);
1990
1991 BUG_ON(!task_on_rq_queued(p));
1992 BUG_ON(!rt_task(p));
1993
1994 return p;
1995}
1996
1997/*
1998 * If the current CPU has more than one RT task, see if the non
1999 * running task can migrate over to a CPU that is running a task
2000 * of lesser priority.
2001 */
2002static int push_rt_task(struct rq *rq, bool pull)
2003{
2004 struct task_struct *next_task;
2005 struct rq *lowest_rq;
2006 int ret = 0;
2007
2008 if (!rq->rt.overloaded)
2009 return 0;
2010
2011 next_task = pick_next_pushable_task(rq);
2012 if (!next_task)
2013 return 0;
2014
2015retry:
2016 /*
2017 * It's possible that the next_task slipped in of
2018 * higher priority than current. If that's the case
2019 * just reschedule current.
2020 */
2021 if (unlikely(next_task->prio < rq->curr->prio)) {
2022 resched_curr(rq);
2023 return 0;
2024 }
2025
2026 if (is_migration_disabled(next_task)) {
2027 struct task_struct *push_task = NULL;
2028 int cpu;
2029
2030 if (!pull || rq->push_busy)
2031 return 0;
2032
2033 /*
2034 * Invoking find_lowest_rq() on anything but an RT task doesn't
2035 * make sense. Per the above priority check, curr has to
2036 * be of higher priority than next_task, so no need to
2037 * reschedule when bailing out.
2038 *
2039 * Note that the stoppers are masqueraded as SCHED_FIFO
2040 * (cf. sched_set_stop_task()), so we can't rely on rt_task().
2041 */
2042 if (rq->curr->sched_class != &rt_sched_class)
2043 return 0;
2044
2045 cpu = find_lowest_rq(rq->curr);
2046 if (cpu == -1 || cpu == rq->cpu)
2047 return 0;
2048
2049 /*
2050 * Given we found a CPU with lower priority than @next_task,
2051 * therefore it should be running. However we cannot migrate it
2052 * to this other CPU, instead attempt to push the current
2053 * running task on this CPU away.
2054 */
2055 push_task = get_push_task(rq);
2056 if (push_task) {
2057 preempt_disable();
2058 raw_spin_rq_unlock(rq);
2059 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2060 push_task, &rq->push_work);
2061 preempt_enable();
2062 raw_spin_rq_lock(rq);
2063 }
2064
2065 return 0;
2066 }
2067
2068 if (WARN_ON(next_task == rq->curr))
2069 return 0;
2070
2071 /* We might release rq lock */
2072 get_task_struct(next_task);
2073
2074 /* find_lock_lowest_rq locks the rq if found */
2075 lowest_rq = find_lock_lowest_rq(next_task, rq);
2076 if (!lowest_rq) {
2077 struct task_struct *task;
2078 /*
2079 * find_lock_lowest_rq releases rq->lock
2080 * so it is possible that next_task has migrated.
2081 *
2082 * We need to make sure that the task is still on the same
2083 * run-queue and is also still the next task eligible for
2084 * pushing.
2085 */
2086 task = pick_next_pushable_task(rq);
2087 if (task == next_task) {
2088 /*
2089 * The task hasn't migrated, and is still the next
2090 * eligible task, but we failed to find a run-queue
2091 * to push it to. Do not retry in this case, since
2092 * other CPUs will pull from us when ready.
2093 */
2094 goto out;
2095 }
2096
2097 if (!task)
2098 /* No more tasks, just exit */
2099 goto out;
2100
2101 /*
2102 * Something has shifted, try again.
2103 */
2104 put_task_struct(next_task);
2105 next_task = task;
2106 goto retry;
2107 }
2108
2109 deactivate_task(rq, next_task, 0);
2110 set_task_cpu(next_task, lowest_rq->cpu);
2111 activate_task(lowest_rq, next_task, 0);
2112 resched_curr(lowest_rq);
2113 ret = 1;
2114
2115 double_unlock_balance(rq, lowest_rq);
2116out:
2117 put_task_struct(next_task);
2118
2119 return ret;
2120}
2121
2122static void push_rt_tasks(struct rq *rq)
2123{
2124 /* push_rt_task will return true if it moved an RT */
2125 while (push_rt_task(rq, false))
2126 ;
2127}
2128
2129#ifdef HAVE_RT_PUSH_IPI
2130
2131/*
2132 * When a high priority task schedules out from a CPU and a lower priority
2133 * task is scheduled in, a check is made to see if there's any RT tasks
2134 * on other CPUs that are waiting to run because a higher priority RT task
2135 * is currently running on its CPU. In this case, the CPU with multiple RT
2136 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
2137 * up that may be able to run one of its non-running queued RT tasks.
2138 *
2139 * All CPUs with overloaded RT tasks need to be notified as there is currently
2140 * no way to know which of these CPUs have the highest priority task waiting
2141 * to run. Instead of trying to take a spinlock on each of these CPUs,
2142 * which has shown to cause large latency when done on machines with many
2143 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
2144 * RT tasks waiting to run.
2145 *
2146 * Just sending an IPI to each of the CPUs is also an issue, as on large
2147 * count CPU machines, this can cause an IPI storm on a CPU, especially
2148 * if its the only CPU with multiple RT tasks queued, and a large number
2149 * of CPUs scheduling a lower priority task at the same time.
2150 *
2151 * Each root domain has its own irq work function that can iterate over
2152 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
2153 * task must be checked if there's one or many CPUs that are lowering
2154 * their priority, there's a single irq work iterator that will try to
2155 * push off RT tasks that are waiting to run.
2156 *
2157 * When a CPU schedules a lower priority task, it will kick off the
2158 * irq work iterator that will jump to each CPU with overloaded RT tasks.
2159 * As it only takes the first CPU that schedules a lower priority task
2160 * to start the process, the rto_start variable is incremented and if
2161 * the atomic result is one, then that CPU will try to take the rto_lock.
2162 * This prevents high contention on the lock as the process handles all
2163 * CPUs scheduling lower priority tasks.
2164 *
2165 * All CPUs that are scheduling a lower priority task will increment the
2166 * rt_loop_next variable. This will make sure that the irq work iterator
2167 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
2168 * priority task, even if the iterator is in the middle of a scan. Incrementing
2169 * the rt_loop_next will cause the iterator to perform another scan.
2170 *
2171 */
2172static int rto_next_cpu(struct root_domain *rd)
2173{
2174 int next;
2175 int cpu;
2176
2177 /*
2178 * When starting the IPI RT pushing, the rto_cpu is set to -1,
2179 * rt_next_cpu() will simply return the first CPU found in
2180 * the rto_mask.
2181 *
2182 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
2183 * will return the next CPU found in the rto_mask.
2184 *
2185 * If there are no more CPUs left in the rto_mask, then a check is made
2186 * against rto_loop and rto_loop_next. rto_loop is only updated with
2187 * the rto_lock held, but any CPU may increment the rto_loop_next
2188 * without any locking.
2189 */
2190 for (;;) {
2191
2192 /* When rto_cpu is -1 this acts like cpumask_first() */
2193 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
2194
2195 rd->rto_cpu = cpu;
2196
2197 if (cpu < nr_cpu_ids)
2198 return cpu;
2199
2200 rd->rto_cpu = -1;
2201
2202 /*
2203 * ACQUIRE ensures we see the @rto_mask changes
2204 * made prior to the @next value observed.
2205 *
2206 * Matches WMB in rt_set_overload().
2207 */
2208 next = atomic_read_acquire(&rd->rto_loop_next);
2209
2210 if (rd->rto_loop == next)
2211 break;
2212
2213 rd->rto_loop = next;
2214 }
2215
2216 return -1;
2217}
2218
2219static inline bool rto_start_trylock(atomic_t *v)
2220{
2221 return !atomic_cmpxchg_acquire(v, 0, 1);
2222}
2223
2224static inline void rto_start_unlock(atomic_t *v)
2225{
2226 atomic_set_release(v, 0);
2227}
2228
2229static void tell_cpu_to_push(struct rq *rq)
2230{
2231 int cpu = -1;
2232
2233 /* Keep the loop going if the IPI is currently active */
2234 atomic_inc(&rq->rd->rto_loop_next);
2235
2236 /* Only one CPU can initiate a loop at a time */
2237 if (!rto_start_trylock(&rq->rd->rto_loop_start))
2238 return;
2239
2240 raw_spin_lock(&rq->rd->rto_lock);
2241
2242 /*
2243 * The rto_cpu is updated under the lock, if it has a valid CPU
2244 * then the IPI is still running and will continue due to the
2245 * update to loop_next, and nothing needs to be done here.
2246 * Otherwise it is finishing up and an ipi needs to be sent.
2247 */
2248 if (rq->rd->rto_cpu < 0)
2249 cpu = rto_next_cpu(rq->rd);
2250
2251 raw_spin_unlock(&rq->rd->rto_lock);
2252
2253 rto_start_unlock(&rq->rd->rto_loop_start);
2254
2255 if (cpu >= 0) {
2256 /* Make sure the rd does not get freed while pushing */
2257 sched_get_rd(rq->rd);
2258 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2259 }
2260}
2261
2262/* Called from hardirq context */
2263void rto_push_irq_work_func(struct irq_work *work)
2264{
2265 struct root_domain *rd =
2266 container_of(work, struct root_domain, rto_push_work);
2267 struct rq *rq;
2268 int cpu;
2269
2270 rq = this_rq();
2271
2272 /*
2273 * We do not need to grab the lock to check for has_pushable_tasks.
2274 * When it gets updated, a check is made if a push is possible.
2275 */
2276 if (has_pushable_tasks(rq)) {
2277 raw_spin_rq_lock(rq);
2278 while (push_rt_task(rq, true))
2279 ;
2280 raw_spin_rq_unlock(rq);
2281 }
2282
2283 raw_spin_lock(&rd->rto_lock);
2284
2285 /* Pass the IPI to the next rt overloaded queue */
2286 cpu = rto_next_cpu(rd);
2287
2288 raw_spin_unlock(&rd->rto_lock);
2289
2290 if (cpu < 0) {
2291 sched_put_rd(rd);
2292 return;
2293 }
2294
2295 /* Try the next RT overloaded CPU */
2296 irq_work_queue_on(&rd->rto_push_work, cpu);
2297}
2298#endif /* HAVE_RT_PUSH_IPI */
2299
2300static void pull_rt_task(struct rq *this_rq)
2301{
2302 int this_cpu = this_rq->cpu, cpu;
2303 bool resched = false;
2304 struct task_struct *p, *push_task;
2305 struct rq *src_rq;
2306 int rt_overload_count = rt_overloaded(this_rq);
2307
2308 if (likely(!rt_overload_count))
2309 return;
2310
2311 /*
2312 * Match the barrier from rt_set_overloaded; this guarantees that if we
2313 * see overloaded we must also see the rto_mask bit.
2314 */
2315 smp_rmb();
2316
2317 /* If we are the only overloaded CPU do nothing */
2318 if (rt_overload_count == 1 &&
2319 cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2320 return;
2321
2322#ifdef HAVE_RT_PUSH_IPI
2323 if (sched_feat(RT_PUSH_IPI)) {
2324 tell_cpu_to_push(this_rq);
2325 return;
2326 }
2327#endif
2328
2329 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2330 if (this_cpu == cpu)
2331 continue;
2332
2333 src_rq = cpu_rq(cpu);
2334
2335 /*
2336 * Don't bother taking the src_rq->lock if the next highest
2337 * task is known to be lower-priority than our current task.
2338 * This may look racy, but if this value is about to go
2339 * logically higher, the src_rq will push this task away.
2340 * And if its going logically lower, we do not care
2341 */
2342 if (src_rq->rt.highest_prio.next >=
2343 this_rq->rt.highest_prio.curr)
2344 continue;
2345
2346 /*
2347 * We can potentially drop this_rq's lock in
2348 * double_lock_balance, and another CPU could
2349 * alter this_rq
2350 */
2351 push_task = NULL;
2352 double_lock_balance(this_rq, src_rq);
2353
2354 /*
2355 * We can pull only a task, which is pushable
2356 * on its rq, and no others.
2357 */
2358 p = pick_highest_pushable_task(src_rq, this_cpu);
2359
2360 /*
2361 * Do we have an RT task that preempts
2362 * the to-be-scheduled task?
2363 */
2364 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2365 WARN_ON(p == src_rq->curr);
2366 WARN_ON(!task_on_rq_queued(p));
2367
2368 /*
2369 * There's a chance that p is higher in priority
2370 * than what's currently running on its CPU.
2371 * This is just that p is waking up and hasn't
2372 * had a chance to schedule. We only pull
2373 * p if it is lower in priority than the
2374 * current task on the run queue
2375 */
2376 if (p->prio < src_rq->curr->prio)
2377 goto skip;
2378
2379 if (is_migration_disabled(p)) {
2380 push_task = get_push_task(src_rq);
2381 } else {
2382 deactivate_task(src_rq, p, 0);
2383 set_task_cpu(p, this_cpu);
2384 activate_task(this_rq, p, 0);
2385 resched = true;
2386 }
2387 /*
2388 * We continue with the search, just in
2389 * case there's an even higher prio task
2390 * in another runqueue. (low likelihood
2391 * but possible)
2392 */
2393 }
2394skip:
2395 double_unlock_balance(this_rq, src_rq);
2396
2397 if (push_task) {
2398 preempt_disable();
2399 raw_spin_rq_unlock(this_rq);
2400 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2401 push_task, &src_rq->push_work);
2402 preempt_enable();
2403 raw_spin_rq_lock(this_rq);
2404 }
2405 }
2406
2407 if (resched)
2408 resched_curr(this_rq);
2409}
2410
2411/*
2412 * If we are not running and we are not going to reschedule soon, we should
2413 * try to push tasks away now
2414 */
2415static void task_woken_rt(struct rq *rq, struct task_struct *p)
2416{
2417 bool need_to_push = !task_on_cpu(rq, p) &&
2418 !test_tsk_need_resched(rq->curr) &&
2419 p->nr_cpus_allowed > 1 &&
2420 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2421 (rq->curr->nr_cpus_allowed < 2 ||
2422 rq->curr->prio <= p->prio);
2423
2424 if (need_to_push)
2425 push_rt_tasks(rq);
2426}
2427
2428/* Assumes rq->lock is held */
2429static void rq_online_rt(struct rq *rq)
2430{
2431 if (rq->rt.overloaded)
2432 rt_set_overload(rq);
2433
2434 __enable_runtime(rq);
2435
2436 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2437}
2438
2439/* Assumes rq->lock is held */
2440static void rq_offline_rt(struct rq *rq)
2441{
2442 if (rq->rt.overloaded)
2443 rt_clear_overload(rq);
2444
2445 __disable_runtime(rq);
2446
2447 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2448}
2449
2450/*
2451 * When switch from the rt queue, we bring ourselves to a position
2452 * that we might want to pull RT tasks from other runqueues.
2453 */
2454static void switched_from_rt(struct rq *rq, struct task_struct *p)
2455{
2456 /*
2457 * If there are other RT tasks then we will reschedule
2458 * and the scheduling of the other RT tasks will handle
2459 * the balancing. But if we are the last RT task
2460 * we may need to handle the pulling of RT tasks
2461 * now.
2462 */
2463 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2464 return;
2465
2466 rt_queue_pull_task(rq);
2467}
2468
2469void __init init_sched_rt_class(void)
2470{
2471 unsigned int i;
2472
2473 for_each_possible_cpu(i) {
2474 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2475 GFP_KERNEL, cpu_to_node(i));
2476 }
2477}
2478#endif /* CONFIG_SMP */
2479
2480/*
2481 * When switching a task to RT, we may overload the runqueue
2482 * with RT tasks. In this case we try to push them off to
2483 * other runqueues.
2484 */
2485static void switched_to_rt(struct rq *rq, struct task_struct *p)
2486{
2487 /*
2488 * If we are running, update the avg_rt tracking, as the running time
2489 * will now on be accounted into the latter.
2490 */
2491 if (task_current(rq, p)) {
2492 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2493 return;
2494 }
2495
2496 /*
2497 * If we are not running we may need to preempt the current
2498 * running task. If that current running task is also an RT task
2499 * then see if we can move to another run queue.
2500 */
2501 if (task_on_rq_queued(p)) {
2502#ifdef CONFIG_SMP
2503 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2504 rt_queue_push_tasks(rq);
2505#endif /* CONFIG_SMP */
2506 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2507 resched_curr(rq);
2508 }
2509}
2510
2511/*
2512 * Priority of the task has changed. This may cause
2513 * us to initiate a push or pull.
2514 */
2515static void
2516prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2517{
2518 if (!task_on_rq_queued(p))
2519 return;
2520
2521 if (task_current(rq, p)) {
2522#ifdef CONFIG_SMP
2523 /*
2524 * If our priority decreases while running, we
2525 * may need to pull tasks to this runqueue.
2526 */
2527 if (oldprio < p->prio)
2528 rt_queue_pull_task(rq);
2529
2530 /*
2531 * If there's a higher priority task waiting to run
2532 * then reschedule.
2533 */
2534 if (p->prio > rq->rt.highest_prio.curr)
2535 resched_curr(rq);
2536#else
2537 /* For UP simply resched on drop of prio */
2538 if (oldprio < p->prio)
2539 resched_curr(rq);
2540#endif /* CONFIG_SMP */
2541 } else {
2542 /*
2543 * This task is not running, but if it is
2544 * greater than the current running task
2545 * then reschedule.
2546 */
2547 if (p->prio < rq->curr->prio)
2548 resched_curr(rq);
2549 }
2550}
2551
2552#ifdef CONFIG_POSIX_TIMERS
2553static void watchdog(struct rq *rq, struct task_struct *p)
2554{
2555 unsigned long soft, hard;
2556
2557 /* max may change after cur was read, this will be fixed next tick */
2558 soft = task_rlimit(p, RLIMIT_RTTIME);
2559 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2560
2561 if (soft != RLIM_INFINITY) {
2562 unsigned long next;
2563
2564 if (p->rt.watchdog_stamp != jiffies) {
2565 p->rt.timeout++;
2566 p->rt.watchdog_stamp = jiffies;
2567 }
2568
2569 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2570 if (p->rt.timeout > next) {
2571 posix_cputimers_rt_watchdog(&p->posix_cputimers,
2572 p->se.sum_exec_runtime);
2573 }
2574 }
2575}
2576#else
2577static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2578#endif
2579
2580/*
2581 * scheduler tick hitting a task of our scheduling class.
2582 *
2583 * NOTE: This function can be called remotely by the tick offload that
2584 * goes along full dynticks. Therefore no local assumption can be made
2585 * and everything must be accessed through the @rq and @curr passed in
2586 * parameters.
2587 */
2588static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2589{
2590 struct sched_rt_entity *rt_se = &p->rt;
2591
2592 update_curr_rt(rq);
2593 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2594
2595 watchdog(rq, p);
2596
2597 /*
2598 * RR tasks need a special form of timeslice management.
2599 * FIFO tasks have no timeslices.
2600 */
2601 if (p->policy != SCHED_RR)
2602 return;
2603
2604 if (--p->rt.time_slice)
2605 return;
2606
2607 p->rt.time_slice = sched_rr_timeslice;
2608
2609 /*
2610 * Requeue to the end of queue if we (and all of our ancestors) are not
2611 * the only element on the queue
2612 */
2613 for_each_sched_rt_entity(rt_se) {
2614 if (rt_se->run_list.prev != rt_se->run_list.next) {
2615 requeue_task_rt(rq, p, 0);
2616 resched_curr(rq);
2617 return;
2618 }
2619 }
2620}
2621
2622static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2623{
2624 /*
2625 * Time slice is 0 for SCHED_FIFO tasks
2626 */
2627 if (task->policy == SCHED_RR)
2628 return sched_rr_timeslice;
2629 else
2630 return 0;
2631}
2632
2633#ifdef CONFIG_SCHED_CORE
2634static int task_is_throttled_rt(struct task_struct *p, int cpu)
2635{
2636 struct rt_rq *rt_rq;
2637
2638#ifdef CONFIG_RT_GROUP_SCHED
2639 rt_rq = task_group(p)->rt_rq[cpu];
2640#else
2641 rt_rq = &cpu_rq(cpu)->rt;
2642#endif
2643
2644 return rt_rq_throttled(rt_rq);
2645}
2646#endif
2647
2648DEFINE_SCHED_CLASS(rt) = {
2649
2650 .enqueue_task = enqueue_task_rt,
2651 .dequeue_task = dequeue_task_rt,
2652 .yield_task = yield_task_rt,
2653
2654 .wakeup_preempt = wakeup_preempt_rt,
2655
2656 .pick_next_task = pick_next_task_rt,
2657 .put_prev_task = put_prev_task_rt,
2658 .set_next_task = set_next_task_rt,
2659
2660#ifdef CONFIG_SMP
2661 .balance = balance_rt,
2662 .pick_task = pick_task_rt,
2663 .select_task_rq = select_task_rq_rt,
2664 .set_cpus_allowed = set_cpus_allowed_common,
2665 .rq_online = rq_online_rt,
2666 .rq_offline = rq_offline_rt,
2667 .task_woken = task_woken_rt,
2668 .switched_from = switched_from_rt,
2669 .find_lock_rq = find_lock_lowest_rq,
2670#endif
2671
2672 .task_tick = task_tick_rt,
2673
2674 .get_rr_interval = get_rr_interval_rt,
2675
2676 .prio_changed = prio_changed_rt,
2677 .switched_to = switched_to_rt,
2678
2679 .update_curr = update_curr_rt,
2680
2681#ifdef CONFIG_SCHED_CORE
2682 .task_is_throttled = task_is_throttled_rt,
2683#endif
2684
2685#ifdef CONFIG_UCLAMP_TASK
2686 .uclamp_enabled = 1,
2687#endif
2688};
2689
2690#ifdef CONFIG_RT_GROUP_SCHED
2691/*
2692 * Ensure that the real time constraints are schedulable.
2693 */
2694static DEFINE_MUTEX(rt_constraints_mutex);
2695
2696static inline int tg_has_rt_tasks(struct task_group *tg)
2697{
2698 struct task_struct *task;
2699 struct css_task_iter it;
2700 int ret = 0;
2701
2702 /*
2703 * Autogroups do not have RT tasks; see autogroup_create().
2704 */
2705 if (task_group_is_autogroup(tg))
2706 return 0;
2707
2708 css_task_iter_start(&tg->css, 0, &it);
2709 while (!ret && (task = css_task_iter_next(&it)))
2710 ret |= rt_task(task);
2711 css_task_iter_end(&it);
2712
2713 return ret;
2714}
2715
2716struct rt_schedulable_data {
2717 struct task_group *tg;
2718 u64 rt_period;
2719 u64 rt_runtime;
2720};
2721
2722static int tg_rt_schedulable(struct task_group *tg, void *data)
2723{
2724 struct rt_schedulable_data *d = data;
2725 struct task_group *child;
2726 unsigned long total, sum = 0;
2727 u64 period, runtime;
2728
2729 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2730 runtime = tg->rt_bandwidth.rt_runtime;
2731
2732 if (tg == d->tg) {
2733 period = d->rt_period;
2734 runtime = d->rt_runtime;
2735 }
2736
2737 /*
2738 * Cannot have more runtime than the period.
2739 */
2740 if (runtime > period && runtime != RUNTIME_INF)
2741 return -EINVAL;
2742
2743 /*
2744 * Ensure we don't starve existing RT tasks if runtime turns zero.
2745 */
2746 if (rt_bandwidth_enabled() && !runtime &&
2747 tg->rt_bandwidth.rt_runtime && tg_has_rt_tasks(tg))
2748 return -EBUSY;
2749
2750 total = to_ratio(period, runtime);
2751
2752 /*
2753 * Nobody can have more than the global setting allows.
2754 */
2755 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2756 return -EINVAL;
2757
2758 /*
2759 * The sum of our children's runtime should not exceed our own.
2760 */
2761 list_for_each_entry_rcu(child, &tg->children, siblings) {
2762 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2763 runtime = child->rt_bandwidth.rt_runtime;
2764
2765 if (child == d->tg) {
2766 period = d->rt_period;
2767 runtime = d->rt_runtime;
2768 }
2769
2770 sum += to_ratio(period, runtime);
2771 }
2772
2773 if (sum > total)
2774 return -EINVAL;
2775
2776 return 0;
2777}
2778
2779static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2780{
2781 int ret;
2782
2783 struct rt_schedulable_data data = {
2784 .tg = tg,
2785 .rt_period = period,
2786 .rt_runtime = runtime,
2787 };
2788
2789 rcu_read_lock();
2790 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2791 rcu_read_unlock();
2792
2793 return ret;
2794}
2795
2796static int tg_set_rt_bandwidth(struct task_group *tg,
2797 u64 rt_period, u64 rt_runtime)
2798{
2799 int i, err = 0;
2800
2801 /*
2802 * Disallowing the root group RT runtime is BAD, it would disallow the
2803 * kernel creating (and or operating) RT threads.
2804 */
2805 if (tg == &root_task_group && rt_runtime == 0)
2806 return -EINVAL;
2807
2808 /* No period doesn't make any sense. */
2809 if (rt_period == 0)
2810 return -EINVAL;
2811
2812 /*
2813 * Bound quota to defend quota against overflow during bandwidth shift.
2814 */
2815 if (rt_runtime != RUNTIME_INF && rt_runtime > max_rt_runtime)
2816 return -EINVAL;
2817
2818 mutex_lock(&rt_constraints_mutex);
2819 err = __rt_schedulable(tg, rt_period, rt_runtime);
2820 if (err)
2821 goto unlock;
2822
2823 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2824 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2825 tg->rt_bandwidth.rt_runtime = rt_runtime;
2826
2827 for_each_possible_cpu(i) {
2828 struct rt_rq *rt_rq = tg->rt_rq[i];
2829
2830 raw_spin_lock(&rt_rq->rt_runtime_lock);
2831 rt_rq->rt_runtime = rt_runtime;
2832 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2833 }
2834 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2835unlock:
2836 mutex_unlock(&rt_constraints_mutex);
2837
2838 return err;
2839}
2840
2841int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2842{
2843 u64 rt_runtime, rt_period;
2844
2845 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2846 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2847 if (rt_runtime_us < 0)
2848 rt_runtime = RUNTIME_INF;
2849 else if ((u64)rt_runtime_us > U64_MAX / NSEC_PER_USEC)
2850 return -EINVAL;
2851
2852 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2853}
2854
2855long sched_group_rt_runtime(struct task_group *tg)
2856{
2857 u64 rt_runtime_us;
2858
2859 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2860 return -1;
2861
2862 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2863 do_div(rt_runtime_us, NSEC_PER_USEC);
2864 return rt_runtime_us;
2865}
2866
2867int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2868{
2869 u64 rt_runtime, rt_period;
2870
2871 if (rt_period_us > U64_MAX / NSEC_PER_USEC)
2872 return -EINVAL;
2873
2874 rt_period = rt_period_us * NSEC_PER_USEC;
2875 rt_runtime = tg->rt_bandwidth.rt_runtime;
2876
2877 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2878}
2879
2880long sched_group_rt_period(struct task_group *tg)
2881{
2882 u64 rt_period_us;
2883
2884 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2885 do_div(rt_period_us, NSEC_PER_USEC);
2886 return rt_period_us;
2887}
2888
2889#ifdef CONFIG_SYSCTL
2890static int sched_rt_global_constraints(void)
2891{
2892 int ret = 0;
2893
2894 mutex_lock(&rt_constraints_mutex);
2895 ret = __rt_schedulable(NULL, 0, 0);
2896 mutex_unlock(&rt_constraints_mutex);
2897
2898 return ret;
2899}
2900#endif /* CONFIG_SYSCTL */
2901
2902int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2903{
2904 /* Don't accept realtime tasks when there is no way for them to run */
2905 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2906 return 0;
2907
2908 return 1;
2909}
2910
2911#else /* !CONFIG_RT_GROUP_SCHED */
2912
2913#ifdef CONFIG_SYSCTL
2914static int sched_rt_global_constraints(void)
2915{
2916 unsigned long flags;
2917 int i;
2918
2919 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2920 for_each_possible_cpu(i) {
2921 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2922
2923 raw_spin_lock(&rt_rq->rt_runtime_lock);
2924 rt_rq->rt_runtime = global_rt_runtime();
2925 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2926 }
2927 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2928
2929 return 0;
2930}
2931#endif /* CONFIG_SYSCTL */
2932#endif /* CONFIG_RT_GROUP_SCHED */
2933
2934#ifdef CONFIG_SYSCTL
2935static int sched_rt_global_validate(void)
2936{
2937 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2938 ((sysctl_sched_rt_runtime > sysctl_sched_rt_period) ||
2939 ((u64)sysctl_sched_rt_runtime *
2940 NSEC_PER_USEC > max_rt_runtime)))
2941 return -EINVAL;
2942
2943 return 0;
2944}
2945
2946static void sched_rt_do_global(void)
2947{
2948 unsigned long flags;
2949
2950 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2951 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2952 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2953 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2954}
2955
2956static int sched_rt_handler(struct ctl_table *table, int write, void *buffer,
2957 size_t *lenp, loff_t *ppos)
2958{
2959 int old_period, old_runtime;
2960 static DEFINE_MUTEX(mutex);
2961 int ret;
2962
2963 mutex_lock(&mutex);
2964 old_period = sysctl_sched_rt_period;
2965 old_runtime = sysctl_sched_rt_runtime;
2966
2967 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2968
2969 if (!ret && write) {
2970 ret = sched_rt_global_validate();
2971 if (ret)
2972 goto undo;
2973
2974 ret = sched_dl_global_validate();
2975 if (ret)
2976 goto undo;
2977
2978 ret = sched_rt_global_constraints();
2979 if (ret)
2980 goto undo;
2981
2982 sched_rt_do_global();
2983 sched_dl_do_global();
2984 }
2985 if (0) {
2986undo:
2987 sysctl_sched_rt_period = old_period;
2988 sysctl_sched_rt_runtime = old_runtime;
2989 }
2990 mutex_unlock(&mutex);
2991
2992 return ret;
2993}
2994
2995static int sched_rr_handler(struct ctl_table *table, int write, void *buffer,
2996 size_t *lenp, loff_t *ppos)
2997{
2998 int ret;
2999 static DEFINE_MUTEX(mutex);
3000
3001 mutex_lock(&mutex);
3002 ret = proc_dointvec(table, write, buffer, lenp, ppos);
3003 /*
3004 * Make sure that internally we keep jiffies.
3005 * Also, writing zero resets the timeslice to default:
3006 */
3007 if (!ret && write) {
3008 sched_rr_timeslice =
3009 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
3010 msecs_to_jiffies(sysctl_sched_rr_timeslice);
3011
3012 if (sysctl_sched_rr_timeslice <= 0)
3013 sysctl_sched_rr_timeslice = jiffies_to_msecs(RR_TIMESLICE);
3014 }
3015 mutex_unlock(&mutex);
3016
3017 return ret;
3018}
3019#endif /* CONFIG_SYSCTL */
3020
3021#ifdef CONFIG_SCHED_DEBUG
3022void print_rt_stats(struct seq_file *m, int cpu)
3023{
3024 rt_rq_iter_t iter;
3025 struct rt_rq *rt_rq;
3026
3027 rcu_read_lock();
3028 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
3029 print_rt_rq(m, cpu, rt_rq);
3030 rcu_read_unlock();
3031}
3032#endif /* CONFIG_SCHED_DEBUG */