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