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