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