projects/linux·blob v5.8 · refs

2799 lines · 65.3 KiB

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