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