663b2355a3aa772d8bcc8c90b55a3e0e0e3a6e17
[sfrench/cifs-2.6.git] / kernel / sched / rt.c
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
7 #include "sched.h"
8
9 #include <linux/slab.h>
10 #include <linux/irq_work.h>
11
12 int sched_rr_timeslice = RR_TIMESLICE;
13 int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
14
15 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
16
17 struct rt_bandwidth def_rt_bandwidth;
18
19 static 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
43 void 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,
51                         CLOCK_MONOTONIC, HRTIMER_MODE_REL);
52         rt_b->rt_period_timer.function = sched_rt_period_timer;
53 }
54
55 static 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, HRTIMER_MODE_ABS_PINNED);
73         }
74         raw_spin_unlock(&rt_b->rt_runtime_lock);
75 }
76
77 void init_rt_rq(struct rt_rq *rt_rq)
78 {
79         struct rt_prio_array *array;
80         int i;
81
82         array = &rt_rq->active;
83         for (i = 0; i < MAX_RT_PRIO; i++) {
84                 INIT_LIST_HEAD(array->queue + i);
85                 __clear_bit(i, array->bitmap);
86         }
87         /* delimiter for bitsearch: */
88         __set_bit(MAX_RT_PRIO, array->bitmap);
89
90 #if defined CONFIG_SMP
91         rt_rq->highest_prio.curr = MAX_RT_PRIO;
92         rt_rq->highest_prio.next = MAX_RT_PRIO;
93         rt_rq->rt_nr_migratory = 0;
94         rt_rq->overloaded = 0;
95         plist_head_init(&rt_rq->pushable_tasks);
96 #endif /* CONFIG_SMP */
97         /* We start is dequeued state, because no RT tasks are queued */
98         rt_rq->rt_queued = 0;
99
100         rt_rq->rt_time = 0;
101         rt_rq->rt_throttled = 0;
102         rt_rq->rt_runtime = 0;
103         raw_spin_lock_init(&rt_rq->rt_runtime_lock);
104 }
105
106 #ifdef CONFIG_RT_GROUP_SCHED
107 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
108 {
109         hrtimer_cancel(&rt_b->rt_period_timer);
110 }
111
112 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
113
114 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
115 {
116 #ifdef CONFIG_SCHED_DEBUG
117         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
118 #endif
119         return container_of(rt_se, struct task_struct, rt);
120 }
121
122 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
123 {
124         return rt_rq->rq;
125 }
126
127 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
128 {
129         return rt_se->rt_rq;
130 }
131
132 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
133 {
134         struct rt_rq *rt_rq = rt_se->rt_rq;
135
136         return rt_rq->rq;
137 }
138
139 void free_rt_sched_group(struct task_group *tg)
140 {
141         int i;
142
143         if (tg->rt_se)
144                 destroy_rt_bandwidth(&tg->rt_bandwidth);
145
146         for_each_possible_cpu(i) {
147                 if (tg->rt_rq)
148                         kfree(tg->rt_rq[i]);
149                 if (tg->rt_se)
150                         kfree(tg->rt_se[i]);
151         }
152
153         kfree(tg->rt_rq);
154         kfree(tg->rt_se);
155 }
156
157 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
158                 struct sched_rt_entity *rt_se, int cpu,
159                 struct sched_rt_entity *parent)
160 {
161         struct rq *rq = cpu_rq(cpu);
162
163         rt_rq->highest_prio.curr = MAX_RT_PRIO;
164         rt_rq->rt_nr_boosted = 0;
165         rt_rq->rq = rq;
166         rt_rq->tg = tg;
167
168         tg->rt_rq[cpu] = rt_rq;
169         tg->rt_se[cpu] = rt_se;
170
171         if (!rt_se)
172                 return;
173
174         if (!parent)
175                 rt_se->rt_rq = &rq->rt;
176         else
177                 rt_se->rt_rq = parent->my_q;
178
179         rt_se->my_q = rt_rq;
180         rt_se->parent = parent;
181         INIT_LIST_HEAD(&rt_se->run_list);
182 }
183
184 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
185 {
186         struct rt_rq *rt_rq;
187         struct sched_rt_entity *rt_se;
188         int i;
189
190         tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
191         if (!tg->rt_rq)
192                 goto err;
193         tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
194         if (!tg->rt_se)
195                 goto err;
196
197         init_rt_bandwidth(&tg->rt_bandwidth,
198                         ktime_to_ns(def_rt_bandwidth.rt_period), 0);
199
200         for_each_possible_cpu(i) {
201                 rt_rq = kzalloc_node(sizeof(struct rt_rq),
202                                      GFP_KERNEL, cpu_to_node(i));
203                 if (!rt_rq)
204                         goto err;
205
206                 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
207                                      GFP_KERNEL, cpu_to_node(i));
208                 if (!rt_se)
209                         goto err_free_rq;
210
211                 init_rt_rq(rt_rq);
212                 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
213                 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
214         }
215
216         return 1;
217
218 err_free_rq:
219         kfree(rt_rq);
220 err:
221         return 0;
222 }
223
224 #else /* CONFIG_RT_GROUP_SCHED */
225
226 #define rt_entity_is_task(rt_se) (1)
227
228 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
229 {
230         return container_of(rt_se, struct task_struct, rt);
231 }
232
233 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
234 {
235         return container_of(rt_rq, struct rq, rt);
236 }
237
238 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
239 {
240         struct task_struct *p = rt_task_of(rt_se);
241
242         return task_rq(p);
243 }
244
245 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
246 {
247         struct rq *rq = rq_of_rt_se(rt_se);
248
249         return &rq->rt;
250 }
251
252 void free_rt_sched_group(struct task_group *tg) { }
253
254 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
255 {
256         return 1;
257 }
258 #endif /* CONFIG_RT_GROUP_SCHED */
259
260 #ifdef CONFIG_SMP
261
262 static void pull_rt_task(struct rq *this_rq);
263
264 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
265 {
266         /* Try to pull RT tasks here if we lower this rq's prio */
267         return rq->rt.highest_prio.curr > prev->prio;
268 }
269
270 static inline int rt_overloaded(struct rq *rq)
271 {
272         return atomic_read(&rq->rd->rto_count);
273 }
274
275 static inline void rt_set_overload(struct rq *rq)
276 {
277         if (!rq->online)
278                 return;
279
280         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
281         /*
282          * Make sure the mask is visible before we set
283          * the overload count. That is checked to determine
284          * if we should look at the mask. It would be a shame
285          * if we looked at the mask, but the mask was not
286          * updated yet.
287          *
288          * Matched by the barrier in pull_rt_task().
289          */
290         smp_wmb();
291         atomic_inc(&rq->rd->rto_count);
292 }
293
294 static inline void rt_clear_overload(struct rq *rq)
295 {
296         if (!rq->online)
297                 return;
298
299         /* the order here really doesn't matter */
300         atomic_dec(&rq->rd->rto_count);
301         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
302 }
303
304 static void update_rt_migration(struct rt_rq *rt_rq)
305 {
306         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
307                 if (!rt_rq->overloaded) {
308                         rt_set_overload(rq_of_rt_rq(rt_rq));
309                         rt_rq->overloaded = 1;
310                 }
311         } else if (rt_rq->overloaded) {
312                 rt_clear_overload(rq_of_rt_rq(rt_rq));
313                 rt_rq->overloaded = 0;
314         }
315 }
316
317 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
318 {
319         struct task_struct *p;
320
321         if (!rt_entity_is_task(rt_se))
322                 return;
323
324         p = rt_task_of(rt_se);
325         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
326
327         rt_rq->rt_nr_total++;
328         if (p->nr_cpus_allowed > 1)
329                 rt_rq->rt_nr_migratory++;
330
331         update_rt_migration(rt_rq);
332 }
333
334 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
335 {
336         struct task_struct *p;
337
338         if (!rt_entity_is_task(rt_se))
339                 return;
340
341         p = rt_task_of(rt_se);
342         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
343
344         rt_rq->rt_nr_total--;
345         if (p->nr_cpus_allowed > 1)
346                 rt_rq->rt_nr_migratory--;
347
348         update_rt_migration(rt_rq);
349 }
350
351 static inline int has_pushable_tasks(struct rq *rq)
352 {
353         return !plist_head_empty(&rq->rt.pushable_tasks);
354 }
355
356 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
357 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
358
359 static void push_rt_tasks(struct rq *);
360 static void pull_rt_task(struct rq *);
361
362 static inline void queue_push_tasks(struct rq *rq)
363 {
364         if (!has_pushable_tasks(rq))
365                 return;
366
367         queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
368 }
369
370 static inline void queue_pull_task(struct rq *rq)
371 {
372         queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
373 }
374
375 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
376 {
377         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
378         plist_node_init(&p->pushable_tasks, p->prio);
379         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
380
381         /* Update the highest prio pushable task */
382         if (p->prio < rq->rt.highest_prio.next)
383                 rq->rt.highest_prio.next = p->prio;
384 }
385
386 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
387 {
388         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
389
390         /* Update the new highest prio pushable task */
391         if (has_pushable_tasks(rq)) {
392                 p = plist_first_entry(&rq->rt.pushable_tasks,
393                                       struct task_struct, pushable_tasks);
394                 rq->rt.highest_prio.next = p->prio;
395         } else
396                 rq->rt.highest_prio.next = MAX_RT_PRIO;
397 }
398
399 #else
400
401 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
402 {
403 }
404
405 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
406 {
407 }
408
409 static inline
410 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
411 {
412 }
413
414 static inline
415 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
416 {
417 }
418
419 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
420 {
421         return false;
422 }
423
424 static inline void pull_rt_task(struct rq *this_rq)
425 {
426 }
427
428 static inline void queue_push_tasks(struct rq *rq)
429 {
430 }
431 #endif /* CONFIG_SMP */
432
433 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
434 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
435
436 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
437 {
438         return rt_se->on_rq;
439 }
440
441 #ifdef CONFIG_RT_GROUP_SCHED
442
443 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
444 {
445         if (!rt_rq->tg)
446                 return RUNTIME_INF;
447
448         return rt_rq->rt_runtime;
449 }
450
451 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
452 {
453         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
454 }
455
456 typedef struct task_group *rt_rq_iter_t;
457
458 static inline struct task_group *next_task_group(struct task_group *tg)
459 {
460         do {
461                 tg = list_entry_rcu(tg->list.next,
462                         typeof(struct task_group), list);
463         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
464
465         if (&tg->list == &task_groups)
466                 tg = NULL;
467
468         return tg;
469 }
470
471 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
472         for (iter = container_of(&task_groups, typeof(*iter), list);    \
473                 (iter = next_task_group(iter)) &&                       \
474                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
475
476 #define for_each_sched_rt_entity(rt_se) \
477         for (; rt_se; rt_se = rt_se->parent)
478
479 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
480 {
481         return rt_se->my_q;
482 }
483
484 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
485 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
486
487 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
488 {
489         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
490         struct rq *rq = rq_of_rt_rq(rt_rq);
491         struct sched_rt_entity *rt_se;
492
493         int cpu = cpu_of(rq);
494
495         rt_se = rt_rq->tg->rt_se[cpu];
496
497         if (rt_rq->rt_nr_running) {
498                 if (!rt_se)
499                         enqueue_top_rt_rq(rt_rq);
500                 else if (!on_rt_rq(rt_se))
501                         enqueue_rt_entity(rt_se, 0);
502
503                 if (rt_rq->highest_prio.curr < curr->prio)
504                         resched_curr(rq);
505         }
506 }
507
508 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
509 {
510         struct sched_rt_entity *rt_se;
511         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
512
513         rt_se = rt_rq->tg->rt_se[cpu];
514
515         if (!rt_se)
516                 dequeue_top_rt_rq(rt_rq);
517         else if (on_rt_rq(rt_se))
518                 dequeue_rt_entity(rt_se, 0);
519 }
520
521 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
522 {
523         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
524 }
525
526 static int rt_se_boosted(struct sched_rt_entity *rt_se)
527 {
528         struct rt_rq *rt_rq = group_rt_rq(rt_se);
529         struct task_struct *p;
530
531         if (rt_rq)
532                 return !!rt_rq->rt_nr_boosted;
533
534         p = rt_task_of(rt_se);
535         return p->prio != p->normal_prio;
536 }
537
538 #ifdef CONFIG_SMP
539 static inline const struct cpumask *sched_rt_period_mask(void)
540 {
541         return this_rq()->rd->span;
542 }
543 #else
544 static inline const struct cpumask *sched_rt_period_mask(void)
545 {
546         return cpu_online_mask;
547 }
548 #endif
549
550 static inline
551 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
552 {
553         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
554 }
555
556 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
557 {
558         return &rt_rq->tg->rt_bandwidth;
559 }
560
561 #else /* !CONFIG_RT_GROUP_SCHED */
562
563 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
564 {
565         return rt_rq->rt_runtime;
566 }
567
568 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
569 {
570         return ktime_to_ns(def_rt_bandwidth.rt_period);
571 }
572
573 typedef struct rt_rq *rt_rq_iter_t;
574
575 #define for_each_rt_rq(rt_rq, iter, rq) \
576         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
577
578 #define for_each_sched_rt_entity(rt_se) \
579         for (; rt_se; rt_se = NULL)
580
581 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
582 {
583         return NULL;
584 }
585
586 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
587 {
588         struct rq *rq = rq_of_rt_rq(rt_rq);
589
590         if (!rt_rq->rt_nr_running)
591                 return;
592
593         enqueue_top_rt_rq(rt_rq);
594         resched_curr(rq);
595 }
596
597 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
598 {
599         dequeue_top_rt_rq(rt_rq);
600 }
601
602 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
603 {
604         return rt_rq->rt_throttled;
605 }
606
607 static inline const struct cpumask *sched_rt_period_mask(void)
608 {
609         return cpu_online_mask;
610 }
611
612 static inline
613 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
614 {
615         return &cpu_rq(cpu)->rt;
616 }
617
618 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
619 {
620         return &def_rt_bandwidth;
621 }
622
623 #endif /* CONFIG_RT_GROUP_SCHED */
624
625 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
626 {
627         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
628
629         return (hrtimer_active(&rt_b->rt_period_timer) ||
630                 rt_rq->rt_time < rt_b->rt_runtime);
631 }
632
633 #ifdef CONFIG_SMP
634 /*
635  * We ran out of runtime, see if we can borrow some from our neighbours.
636  */
637 static void do_balance_runtime(struct rt_rq *rt_rq)
638 {
639         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
640         struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
641         int i, weight;
642         u64 rt_period;
643
644         weight = cpumask_weight(rd->span);
645
646         raw_spin_lock(&rt_b->rt_runtime_lock);
647         rt_period = ktime_to_ns(rt_b->rt_period);
648         for_each_cpu(i, rd->span) {
649                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
650                 s64 diff;
651
652                 if (iter == rt_rq)
653                         continue;
654
655                 raw_spin_lock(&iter->rt_runtime_lock);
656                 /*
657                  * Either all rqs have inf runtime and there's nothing to steal
658                  * or __disable_runtime() below sets a specific rq to inf to
659                  * indicate its been disabled and disalow stealing.
660                  */
661                 if (iter->rt_runtime == RUNTIME_INF)
662                         goto next;
663
664                 /*
665                  * From runqueues with spare time, take 1/n part of their
666                  * spare time, but no more than our period.
667                  */
668                 diff = iter->rt_runtime - iter->rt_time;
669                 if (diff > 0) {
670                         diff = div_u64((u64)diff, weight);
671                         if (rt_rq->rt_runtime + diff > rt_period)
672                                 diff = rt_period - rt_rq->rt_runtime;
673                         iter->rt_runtime -= diff;
674                         rt_rq->rt_runtime += diff;
675                         if (rt_rq->rt_runtime == rt_period) {
676                                 raw_spin_unlock(&iter->rt_runtime_lock);
677                                 break;
678                         }
679                 }
680 next:
681                 raw_spin_unlock(&iter->rt_runtime_lock);
682         }
683         raw_spin_unlock(&rt_b->rt_runtime_lock);
684 }
685
686 /*
687  * Ensure this RQ takes back all the runtime it lend to its neighbours.
688  */
689 static void __disable_runtime(struct rq *rq)
690 {
691         struct root_domain *rd = rq->rd;
692         rt_rq_iter_t iter;
693         struct rt_rq *rt_rq;
694
695         if (unlikely(!scheduler_running))
696                 return;
697
698         for_each_rt_rq(rt_rq, iter, rq) {
699                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
700                 s64 want;
701                 int i;
702
703                 raw_spin_lock(&rt_b->rt_runtime_lock);
704                 raw_spin_lock(&rt_rq->rt_runtime_lock);
705                 /*
706                  * Either we're all inf and nobody needs to borrow, or we're
707                  * already disabled and thus have nothing to do, or we have
708                  * exactly the right amount of runtime to take out.
709                  */
710                 if (rt_rq->rt_runtime == RUNTIME_INF ||
711                                 rt_rq->rt_runtime == rt_b->rt_runtime)
712                         goto balanced;
713                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
714
715                 /*
716                  * Calculate the difference between what we started out with
717                  * and what we current have, that's the amount of runtime
718                  * we lend and now have to reclaim.
719                  */
720                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
721
722                 /*
723                  * Greedy reclaim, take back as much as we can.
724                  */
725                 for_each_cpu(i, rd->span) {
726                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
727                         s64 diff;
728
729                         /*
730                          * Can't reclaim from ourselves or disabled runqueues.
731                          */
732                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
733                                 continue;
734
735                         raw_spin_lock(&iter->rt_runtime_lock);
736                         if (want > 0) {
737                                 diff = min_t(s64, iter->rt_runtime, want);
738                                 iter->rt_runtime -= diff;
739                                 want -= diff;
740                         } else {
741                                 iter->rt_runtime -= want;
742                                 want -= want;
743                         }
744                         raw_spin_unlock(&iter->rt_runtime_lock);
745
746                         if (!want)
747                                 break;
748                 }
749
750                 raw_spin_lock(&rt_rq->rt_runtime_lock);
751                 /*
752                  * We cannot be left wanting - that would mean some runtime
753                  * leaked out of the system.
754                  */
755                 BUG_ON(want);
756 balanced:
757                 /*
758                  * Disable all the borrow logic by pretending we have inf
759                  * runtime - in which case borrowing doesn't make sense.
760                  */
761                 rt_rq->rt_runtime = RUNTIME_INF;
762                 rt_rq->rt_throttled = 0;
763                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
764                 raw_spin_unlock(&rt_b->rt_runtime_lock);
765
766                 /* Make rt_rq available for pick_next_task() */
767                 sched_rt_rq_enqueue(rt_rq);
768         }
769 }
770
771 static void __enable_runtime(struct rq *rq)
772 {
773         rt_rq_iter_t iter;
774         struct rt_rq *rt_rq;
775
776         if (unlikely(!scheduler_running))
777                 return;
778
779         /*
780          * Reset each runqueue's bandwidth settings
781          */
782         for_each_rt_rq(rt_rq, iter, rq) {
783                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
784
785                 raw_spin_lock(&rt_b->rt_runtime_lock);
786                 raw_spin_lock(&rt_rq->rt_runtime_lock);
787                 rt_rq->rt_runtime = rt_b->rt_runtime;
788                 rt_rq->rt_time = 0;
789                 rt_rq->rt_throttled = 0;
790                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
791                 raw_spin_unlock(&rt_b->rt_runtime_lock);
792         }
793 }
794
795 static void balance_runtime(struct rt_rq *rt_rq)
796 {
797         if (!sched_feat(RT_RUNTIME_SHARE))
798                 return;
799
800         if (rt_rq->rt_time > rt_rq->rt_runtime) {
801                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
802                 do_balance_runtime(rt_rq);
803                 raw_spin_lock(&rt_rq->rt_runtime_lock);
804         }
805 }
806 #else /* !CONFIG_SMP */
807 static inline void balance_runtime(struct rt_rq *rt_rq) {}
808 #endif /* CONFIG_SMP */
809
810 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
811 {
812         int i, idle = 1, throttled = 0;
813         const struct cpumask *span;
814
815         span = sched_rt_period_mask();
816 #ifdef CONFIG_RT_GROUP_SCHED
817         /*
818          * FIXME: isolated CPUs should really leave the root task group,
819          * whether they are isolcpus or were isolated via cpusets, lest
820          * the timer run on a CPU which does not service all runqueues,
821          * potentially leaving other CPUs indefinitely throttled.  If
822          * isolation is really required, the user will turn the throttle
823          * off to kill the perturbations it causes anyway.  Meanwhile,
824          * this maintains functionality for boot and/or troubleshooting.
825          */
826         if (rt_b == &root_task_group.rt_bandwidth)
827                 span = cpu_online_mask;
828 #endif
829         for_each_cpu(i, span) {
830                 int enqueue = 0;
831                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
832                 struct rq *rq = rq_of_rt_rq(rt_rq);
833                 int skip;
834
835                 /*
836                  * When span == cpu_online_mask, taking each rq->lock
837                  * can be time-consuming. Try to avoid it when possible.
838                  */
839                 raw_spin_lock(&rt_rq->rt_runtime_lock);
840                 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
841                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
842                 if (skip)
843                         continue;
844
845                 raw_spin_lock(&rq->lock);
846                 if (rt_rq->rt_time) {
847                         u64 runtime;
848
849                         raw_spin_lock(&rt_rq->rt_runtime_lock);
850                         if (rt_rq->rt_throttled)
851                                 balance_runtime(rt_rq);
852                         runtime = rt_rq->rt_runtime;
853                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
854                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
855                                 rt_rq->rt_throttled = 0;
856                                 enqueue = 1;
857
858                                 /*
859                                  * When we're idle and a woken (rt) task is
860                                  * throttled check_preempt_curr() will set
861                                  * skip_update and the time between the wakeup
862                                  * and this unthrottle will get accounted as
863                                  * 'runtime'.
864                                  */
865                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
866                                         rq_clock_skip_update(rq, false);
867                         }
868                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
869                                 idle = 0;
870                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
871                 } else if (rt_rq->rt_nr_running) {
872                         idle = 0;
873                         if (!rt_rq_throttled(rt_rq))
874                                 enqueue = 1;
875                 }
876                 if (rt_rq->rt_throttled)
877                         throttled = 1;
878
879                 if (enqueue)
880                         sched_rt_rq_enqueue(rt_rq);
881                 raw_spin_unlock(&rq->lock);
882         }
883
884         if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
885                 return 1;
886
887         return idle;
888 }
889
890 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
891 {
892 #ifdef CONFIG_RT_GROUP_SCHED
893         struct rt_rq *rt_rq = group_rt_rq(rt_se);
894
895         if (rt_rq)
896                 return rt_rq->highest_prio.curr;
897 #endif
898
899         return rt_task_of(rt_se)->prio;
900 }
901
902 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
903 {
904         u64 runtime = sched_rt_runtime(rt_rq);
905
906         if (rt_rq->rt_throttled)
907                 return rt_rq_throttled(rt_rq);
908
909         if (runtime >= sched_rt_period(rt_rq))
910                 return 0;
911
912         balance_runtime(rt_rq);
913         runtime = sched_rt_runtime(rt_rq);
914         if (runtime == RUNTIME_INF)
915                 return 0;
916
917         if (rt_rq->rt_time > runtime) {
918                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
919
920                 /*
921                  * Don't actually throttle groups that have no runtime assigned
922                  * but accrue some time due to boosting.
923                  */
924                 if (likely(rt_b->rt_runtime)) {
925                         rt_rq->rt_throttled = 1;
926                         printk_deferred_once("sched: RT throttling activated\n");
927                 } else {
928                         /*
929                          * In case we did anyway, make it go away,
930                          * replenishment is a joke, since it will replenish us
931                          * with exactly 0 ns.
932                          */
933                         rt_rq->rt_time = 0;
934                 }
935
936                 if (rt_rq_throttled(rt_rq)) {
937                         sched_rt_rq_dequeue(rt_rq);
938                         return 1;
939                 }
940         }
941
942         return 0;
943 }
944
945 /*
946  * Update the current task's runtime statistics. Skip current tasks that
947  * are not in our scheduling class.
948  */
949 static void update_curr_rt(struct rq *rq)
950 {
951         struct task_struct *curr = rq->curr;
952         struct sched_rt_entity *rt_se = &curr->rt;
953         u64 now = rq_clock_task(rq);
954         u64 delta_exec;
955
956         if (curr->sched_class != &rt_sched_class)
957                 return;
958
959         delta_exec = now - curr->se.exec_start;
960         if (unlikely((s64)delta_exec <= 0))
961                 return;
962
963         /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
964         cpufreq_update_util(rq, SCHED_CPUFREQ_RT);
965
966         schedstat_set(curr->se.statistics.exec_max,
967                       max(curr->se.statistics.exec_max, delta_exec));
968
969         curr->se.sum_exec_runtime += delta_exec;
970         account_group_exec_runtime(curr, delta_exec);
971
972         curr->se.exec_start = now;
973         cgroup_account_cputime(curr, delta_exec);
974
975         sched_rt_avg_update(rq, delta_exec);
976
977         if (!rt_bandwidth_enabled())
978                 return;
979
980         for_each_sched_rt_entity(rt_se) {
981                 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
982
983                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
984                         raw_spin_lock(&rt_rq->rt_runtime_lock);
985                         rt_rq->rt_time += delta_exec;
986                         if (sched_rt_runtime_exceeded(rt_rq))
987                                 resched_curr(rq);
988                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
989                 }
990         }
991 }
992
993 static void
994 dequeue_top_rt_rq(struct rt_rq *rt_rq)
995 {
996         struct rq *rq = rq_of_rt_rq(rt_rq);
997
998         BUG_ON(&rq->rt != rt_rq);
999
1000         if (!rt_rq->rt_queued)
1001                 return;
1002
1003         BUG_ON(!rq->nr_running);
1004
1005         sub_nr_running(rq, rt_rq->rt_nr_running);
1006         rt_rq->rt_queued = 0;
1007 }
1008
1009 static void
1010 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1011 {
1012         struct rq *rq = rq_of_rt_rq(rt_rq);
1013
1014         BUG_ON(&rq->rt != rt_rq);
1015
1016         if (rt_rq->rt_queued)
1017                 return;
1018         if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1019                 return;
1020
1021         add_nr_running(rq, rt_rq->rt_nr_running);
1022         rt_rq->rt_queued = 1;
1023 }
1024
1025 #if defined CONFIG_SMP
1026
1027 static void
1028 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1029 {
1030         struct rq *rq = rq_of_rt_rq(rt_rq);
1031
1032 #ifdef CONFIG_RT_GROUP_SCHED
1033         /*
1034          * Change rq's cpupri only if rt_rq is the top queue.
1035          */
1036         if (&rq->rt != rt_rq)
1037                 return;
1038 #endif
1039         if (rq->online && prio < prev_prio)
1040                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1041 }
1042
1043 static void
1044 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1045 {
1046         struct rq *rq = rq_of_rt_rq(rt_rq);
1047
1048 #ifdef CONFIG_RT_GROUP_SCHED
1049         /*
1050          * Change rq's cpupri only if rt_rq is the top queue.
1051          */
1052         if (&rq->rt != rt_rq)
1053                 return;
1054 #endif
1055         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1056                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1057 }
1058
1059 #else /* CONFIG_SMP */
1060
1061 static inline
1062 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1063 static inline
1064 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1065
1066 #endif /* CONFIG_SMP */
1067
1068 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1069 static void
1070 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1071 {
1072         int prev_prio = rt_rq->highest_prio.curr;
1073
1074         if (prio < prev_prio)
1075                 rt_rq->highest_prio.curr = prio;
1076
1077         inc_rt_prio_smp(rt_rq, prio, prev_prio);
1078 }
1079
1080 static void
1081 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1082 {
1083         int prev_prio = rt_rq->highest_prio.curr;
1084
1085         if (rt_rq->rt_nr_running) {
1086
1087                 WARN_ON(prio < prev_prio);
1088
1089                 /*
1090                  * This may have been our highest task, and therefore
1091                  * we may have some recomputation to do
1092                  */
1093                 if (prio == prev_prio) {
1094                         struct rt_prio_array *array = &rt_rq->active;
1095
1096                         rt_rq->highest_prio.curr =
1097                                 sched_find_first_bit(array->bitmap);
1098                 }
1099
1100         } else
1101                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1102
1103         dec_rt_prio_smp(rt_rq, prio, prev_prio);
1104 }
1105
1106 #else
1107
1108 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1109 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1110
1111 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1112
1113 #ifdef CONFIG_RT_GROUP_SCHED
1114
1115 static void
1116 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1117 {
1118         if (rt_se_boosted(rt_se))
1119                 rt_rq->rt_nr_boosted++;
1120
1121         if (rt_rq->tg)
1122                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1123 }
1124
1125 static void
1126 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1127 {
1128         if (rt_se_boosted(rt_se))
1129                 rt_rq->rt_nr_boosted--;
1130
1131         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1132 }
1133
1134 #else /* CONFIG_RT_GROUP_SCHED */
1135
1136 static void
1137 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1138 {
1139         start_rt_bandwidth(&def_rt_bandwidth);
1140 }
1141
1142 static inline
1143 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1144
1145 #endif /* CONFIG_RT_GROUP_SCHED */
1146
1147 static inline
1148 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1149 {
1150         struct rt_rq *group_rq = group_rt_rq(rt_se);
1151
1152         if (group_rq)
1153                 return group_rq->rt_nr_running;
1154         else
1155                 return 1;
1156 }
1157
1158 static inline
1159 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1160 {
1161         struct rt_rq *group_rq = group_rt_rq(rt_se);
1162         struct task_struct *tsk;
1163
1164         if (group_rq)
1165                 return group_rq->rr_nr_running;
1166
1167         tsk = rt_task_of(rt_se);
1168
1169         return (tsk->policy == SCHED_RR) ? 1 : 0;
1170 }
1171
1172 static inline
1173 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1174 {
1175         int prio = rt_se_prio(rt_se);
1176
1177         WARN_ON(!rt_prio(prio));
1178         rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1179         rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1180
1181         inc_rt_prio(rt_rq, prio);
1182         inc_rt_migration(rt_se, rt_rq);
1183         inc_rt_group(rt_se, rt_rq);
1184 }
1185
1186 static inline
1187 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1188 {
1189         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1190         WARN_ON(!rt_rq->rt_nr_running);
1191         rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1192         rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1193
1194         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1195         dec_rt_migration(rt_se, rt_rq);
1196         dec_rt_group(rt_se, rt_rq);
1197 }
1198
1199 /*
1200  * Change rt_se->run_list location unless SAVE && !MOVE
1201  *
1202  * assumes ENQUEUE/DEQUEUE flags match
1203  */
1204 static inline bool move_entity(unsigned int flags)
1205 {
1206         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1207                 return false;
1208
1209         return true;
1210 }
1211
1212 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1213 {
1214         list_del_init(&rt_se->run_list);
1215
1216         if (list_empty(array->queue + rt_se_prio(rt_se)))
1217                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1218
1219         rt_se->on_list = 0;
1220 }
1221
1222 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1223 {
1224         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1225         struct rt_prio_array *array = &rt_rq->active;
1226         struct rt_rq *group_rq = group_rt_rq(rt_se);
1227         struct list_head *queue = array->queue + rt_se_prio(rt_se);
1228
1229         /*
1230          * Don't enqueue the group if its throttled, or when empty.
1231          * The latter is a consequence of the former when a child group
1232          * get throttled and the current group doesn't have any other
1233          * active members.
1234          */
1235         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1236                 if (rt_se->on_list)
1237                         __delist_rt_entity(rt_se, array);
1238                 return;
1239         }
1240
1241         if (move_entity(flags)) {
1242                 WARN_ON_ONCE(rt_se->on_list);
1243                 if (flags & ENQUEUE_HEAD)
1244                         list_add(&rt_se->run_list, queue);
1245                 else
1246                         list_add_tail(&rt_se->run_list, queue);
1247
1248                 __set_bit(rt_se_prio(rt_se), array->bitmap);
1249                 rt_se->on_list = 1;
1250         }
1251         rt_se->on_rq = 1;
1252
1253         inc_rt_tasks(rt_se, rt_rq);
1254 }
1255
1256 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1257 {
1258         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1259         struct rt_prio_array *array = &rt_rq->active;
1260
1261         if (move_entity(flags)) {
1262                 WARN_ON_ONCE(!rt_se->on_list);
1263                 __delist_rt_entity(rt_se, array);
1264         }
1265         rt_se->on_rq = 0;
1266
1267         dec_rt_tasks(rt_se, rt_rq);
1268 }
1269
1270 /*
1271  * Because the prio of an upper entry depends on the lower
1272  * entries, we must remove entries top - down.
1273  */
1274 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1275 {
1276         struct sched_rt_entity *back = NULL;
1277
1278         for_each_sched_rt_entity(rt_se) {
1279                 rt_se->back = back;
1280                 back = rt_se;
1281         }
1282
1283         dequeue_top_rt_rq(rt_rq_of_se(back));
1284
1285         for (rt_se = back; rt_se; rt_se = rt_se->back) {
1286                 if (on_rt_rq(rt_se))
1287                         __dequeue_rt_entity(rt_se, flags);
1288         }
1289 }
1290
1291 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1292 {
1293         struct rq *rq = rq_of_rt_se(rt_se);
1294
1295         dequeue_rt_stack(rt_se, flags);
1296         for_each_sched_rt_entity(rt_se)
1297                 __enqueue_rt_entity(rt_se, flags);
1298         enqueue_top_rt_rq(&rq->rt);
1299 }
1300
1301 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1302 {
1303         struct rq *rq = rq_of_rt_se(rt_se);
1304
1305         dequeue_rt_stack(rt_se, flags);
1306
1307         for_each_sched_rt_entity(rt_se) {
1308                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1309
1310                 if (rt_rq && rt_rq->rt_nr_running)
1311                         __enqueue_rt_entity(rt_se, flags);
1312         }
1313         enqueue_top_rt_rq(&rq->rt);
1314 }
1315
1316 /*
1317  * Adding/removing a task to/from a priority array:
1318  */
1319 static void
1320 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1321 {
1322         struct sched_rt_entity *rt_se = &p->rt;
1323
1324         if (flags & ENQUEUE_WAKEUP)
1325                 rt_se->timeout = 0;
1326
1327         enqueue_rt_entity(rt_se, flags);
1328
1329         if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1330                 enqueue_pushable_task(rq, p);
1331 }
1332
1333 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1334 {
1335         struct sched_rt_entity *rt_se = &p->rt;
1336
1337         update_curr_rt(rq);
1338         dequeue_rt_entity(rt_se, flags);
1339
1340         dequeue_pushable_task(rq, p);
1341 }
1342
1343 /*
1344  * Put task to the head or the end of the run list without the overhead of
1345  * dequeue followed by enqueue.
1346  */
1347 static void
1348 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1349 {
1350         if (on_rt_rq(rt_se)) {
1351                 struct rt_prio_array *array = &rt_rq->active;
1352                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1353
1354                 if (head)
1355                         list_move(&rt_se->run_list, queue);
1356                 else
1357                         list_move_tail(&rt_se->run_list, queue);
1358         }
1359 }
1360
1361 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1362 {
1363         struct sched_rt_entity *rt_se = &p->rt;
1364         struct rt_rq *rt_rq;
1365
1366         for_each_sched_rt_entity(rt_se) {
1367                 rt_rq = rt_rq_of_se(rt_se);
1368                 requeue_rt_entity(rt_rq, rt_se, head);
1369         }
1370 }
1371
1372 static void yield_task_rt(struct rq *rq)
1373 {
1374         requeue_task_rt(rq, rq->curr, 0);
1375 }
1376
1377 #ifdef CONFIG_SMP
1378 static int find_lowest_rq(struct task_struct *task);
1379
1380 static int
1381 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1382 {
1383         struct task_struct *curr;
1384         struct rq *rq;
1385
1386         /* For anything but wake ups, just return the task_cpu */
1387         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1388                 goto out;
1389
1390         rq = cpu_rq(cpu);
1391
1392         rcu_read_lock();
1393         curr = READ_ONCE(rq->curr); /* unlocked access */
1394
1395         /*
1396          * If the current task on @p's runqueue is an RT task, then
1397          * try to see if we can wake this RT task up on another
1398          * runqueue. Otherwise simply start this RT task
1399          * on its current runqueue.
1400          *
1401          * We want to avoid overloading runqueues. If the woken
1402          * task is a higher priority, then it will stay on this CPU
1403          * and the lower prio task should be moved to another CPU.
1404          * Even though this will probably make the lower prio task
1405          * lose its cache, we do not want to bounce a higher task
1406          * around just because it gave up its CPU, perhaps for a
1407          * lock?
1408          *
1409          * For equal prio tasks, we just let the scheduler sort it out.
1410          *
1411          * Otherwise, just let it ride on the affined RQ and the
1412          * post-schedule router will push the preempted task away
1413          *
1414          * This test is optimistic, if we get it wrong the load-balancer
1415          * will have to sort it out.
1416          */
1417         if (curr && unlikely(rt_task(curr)) &&
1418             (curr->nr_cpus_allowed < 2 ||
1419              curr->prio <= p->prio)) {
1420                 int target = find_lowest_rq(p);
1421
1422                 /*
1423                  * Don't bother moving it if the destination CPU is
1424                  * not running a lower priority task.
1425                  */
1426                 if (target != -1 &&
1427                     p->prio < cpu_rq(target)->rt.highest_prio.curr)
1428                         cpu = target;
1429         }
1430         rcu_read_unlock();
1431
1432 out:
1433         return cpu;
1434 }
1435
1436 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1437 {
1438         /*
1439          * Current can't be migrated, useless to reschedule,
1440          * let's hope p can move out.
1441          */
1442         if (rq->curr->nr_cpus_allowed == 1 ||
1443             !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1444                 return;
1445
1446         /*
1447          * p is migratable, so let's not schedule it and
1448          * see if it is pushed or pulled somewhere else.
1449          */
1450         if (p->nr_cpus_allowed != 1
1451             && cpupri_find(&rq->rd->cpupri, p, NULL))
1452                 return;
1453
1454         /*
1455          * There appears to be other cpus that can accept
1456          * current and none to run 'p', so lets reschedule
1457          * to try and push current away:
1458          */
1459         requeue_task_rt(rq, p, 1);
1460         resched_curr(rq);
1461 }
1462
1463 #endif /* CONFIG_SMP */
1464
1465 /*
1466  * Preempt the current task with a newly woken task if needed:
1467  */
1468 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1469 {
1470         if (p->prio < rq->curr->prio) {
1471                 resched_curr(rq);
1472                 return;
1473         }
1474
1475 #ifdef CONFIG_SMP
1476         /*
1477          * If:
1478          *
1479          * - the newly woken task is of equal priority to the current task
1480          * - the newly woken task is non-migratable while current is migratable
1481          * - current will be preempted on the next reschedule
1482          *
1483          * we should check to see if current can readily move to a different
1484          * cpu.  If so, we will reschedule to allow the push logic to try
1485          * to move current somewhere else, making room for our non-migratable
1486          * task.
1487          */
1488         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1489                 check_preempt_equal_prio(rq, p);
1490 #endif
1491 }
1492
1493 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1494                                                    struct rt_rq *rt_rq)
1495 {
1496         struct rt_prio_array *array = &rt_rq->active;
1497         struct sched_rt_entity *next = NULL;
1498         struct list_head *queue;
1499         int idx;
1500
1501         idx = sched_find_first_bit(array->bitmap);
1502         BUG_ON(idx >= MAX_RT_PRIO);
1503
1504         queue = array->queue + idx;
1505         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1506
1507         return next;
1508 }
1509
1510 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1511 {
1512         struct sched_rt_entity *rt_se;
1513         struct task_struct *p;
1514         struct rt_rq *rt_rq  = &rq->rt;
1515
1516         do {
1517                 rt_se = pick_next_rt_entity(rq, rt_rq);
1518                 BUG_ON(!rt_se);
1519                 rt_rq = group_rt_rq(rt_se);
1520         } while (rt_rq);
1521
1522         p = rt_task_of(rt_se);
1523         p->se.exec_start = rq_clock_task(rq);
1524
1525         return p;
1526 }
1527
1528 static struct task_struct *
1529 pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1530 {
1531         struct task_struct *p;
1532         struct rt_rq *rt_rq = &rq->rt;
1533
1534         if (need_pull_rt_task(rq, prev)) {
1535                 /*
1536                  * This is OK, because current is on_cpu, which avoids it being
1537                  * picked for load-balance and preemption/IRQs are still
1538                  * disabled avoiding further scheduler activity on it and we're
1539                  * being very careful to re-start the picking loop.
1540                  */
1541                 rq_unpin_lock(rq, rf);
1542                 pull_rt_task(rq);
1543                 rq_repin_lock(rq, rf);
1544                 /*
1545                  * pull_rt_task() can drop (and re-acquire) rq->lock; this
1546                  * means a dl or stop task can slip in, in which case we need
1547                  * to re-start task selection.
1548                  */
1549                 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1550                              rq->dl.dl_nr_running))
1551                         return RETRY_TASK;
1552         }
1553
1554         /*
1555          * We may dequeue prev's rt_rq in put_prev_task().
1556          * So, we update time before rt_nr_running check.
1557          */
1558         if (prev->sched_class == &rt_sched_class)
1559                 update_curr_rt(rq);
1560
1561         if (!rt_rq->rt_queued)
1562                 return NULL;
1563
1564         put_prev_task(rq, prev);
1565
1566         p = _pick_next_task_rt(rq);
1567
1568         /* The running task is never eligible for pushing */
1569         dequeue_pushable_task(rq, p);
1570
1571         queue_push_tasks(rq);
1572
1573         return p;
1574 }
1575
1576 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1577 {
1578         update_curr_rt(rq);
1579
1580         /*
1581          * The previous task needs to be made eligible for pushing
1582          * if it is still active
1583          */
1584         if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1585                 enqueue_pushable_task(rq, p);
1586 }
1587
1588 #ifdef CONFIG_SMP
1589
1590 /* Only try algorithms three times */
1591 #define RT_MAX_TRIES 3
1592
1593 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1594 {
1595         if (!task_running(rq, p) &&
1596             cpumask_test_cpu(cpu, &p->cpus_allowed))
1597                 return 1;
1598         return 0;
1599 }
1600
1601 /*
1602  * Return the highest pushable rq's task, which is suitable to be executed
1603  * on the cpu, NULL otherwise
1604  */
1605 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1606 {
1607         struct plist_head *head = &rq->rt.pushable_tasks;
1608         struct task_struct *p;
1609
1610         if (!has_pushable_tasks(rq))
1611                 return NULL;
1612
1613         plist_for_each_entry(p, head, pushable_tasks) {
1614                 if (pick_rt_task(rq, p, cpu))
1615                         return p;
1616         }
1617
1618         return NULL;
1619 }
1620
1621 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1622
1623 static int find_lowest_rq(struct task_struct *task)
1624 {
1625         struct sched_domain *sd;
1626         struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1627         int this_cpu = smp_processor_id();
1628         int cpu      = task_cpu(task);
1629
1630         /* Make sure the mask is initialized first */
1631         if (unlikely(!lowest_mask))
1632                 return -1;
1633
1634         if (task->nr_cpus_allowed == 1)
1635                 return -1; /* No other targets possible */
1636
1637         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1638                 return -1; /* No targets found */
1639
1640         /*
1641          * At this point we have built a mask of cpus representing the
1642          * lowest priority tasks in the system.  Now we want to elect
1643          * the best one based on our affinity and topology.
1644          *
1645          * We prioritize the last cpu that the task executed on since
1646          * it is most likely cache-hot in that location.
1647          */
1648         if (cpumask_test_cpu(cpu, lowest_mask))
1649                 return cpu;
1650
1651         /*
1652          * Otherwise, we consult the sched_domains span maps to figure
1653          * out which cpu is logically closest to our hot cache data.
1654          */
1655         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1656                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1657
1658         rcu_read_lock();
1659         for_each_domain(cpu, sd) {
1660                 if (sd->flags & SD_WAKE_AFFINE) {
1661                         int best_cpu;
1662
1663                         /*
1664                          * "this_cpu" is cheaper to preempt than a
1665                          * remote processor.
1666                          */
1667                         if (this_cpu != -1 &&
1668                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1669                                 rcu_read_unlock();
1670                                 return this_cpu;
1671                         }
1672
1673                         best_cpu = cpumask_first_and(lowest_mask,
1674                                                      sched_domain_span(sd));
1675                         if (best_cpu < nr_cpu_ids) {
1676                                 rcu_read_unlock();
1677                                 return best_cpu;
1678                         }
1679                 }
1680         }
1681         rcu_read_unlock();
1682
1683         /*
1684          * And finally, if there were no matches within the domains
1685          * just give the caller *something* to work with from the compatible
1686          * locations.
1687          */
1688         if (this_cpu != -1)
1689                 return this_cpu;
1690
1691         cpu = cpumask_any(lowest_mask);
1692         if (cpu < nr_cpu_ids)
1693                 return cpu;
1694         return -1;
1695 }
1696
1697 /* Will lock the rq it finds */
1698 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1699 {
1700         struct rq *lowest_rq = NULL;
1701         int tries;
1702         int cpu;
1703
1704         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1705                 cpu = find_lowest_rq(task);
1706
1707                 if ((cpu == -1) || (cpu == rq->cpu))
1708                         break;
1709
1710                 lowest_rq = cpu_rq(cpu);
1711
1712                 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1713                         /*
1714                          * Target rq has tasks of equal or higher priority,
1715                          * retrying does not release any lock and is unlikely
1716                          * to yield a different result.
1717                          */
1718                         lowest_rq = NULL;
1719                         break;
1720                 }
1721
1722                 /* if the prio of this runqueue changed, try again */
1723                 if (double_lock_balance(rq, lowest_rq)) {
1724                         /*
1725                          * We had to unlock the run queue. In
1726                          * the mean time, task could have
1727                          * migrated already or had its affinity changed.
1728                          * Also make sure that it wasn't scheduled on its rq.
1729                          */
1730                         if (unlikely(task_rq(task) != rq ||
1731                                      !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) ||
1732                                      task_running(rq, task) ||
1733                                      !rt_task(task) ||
1734                                      !task_on_rq_queued(task))) {
1735
1736                                 double_unlock_balance(rq, lowest_rq);
1737                                 lowest_rq = NULL;
1738                                 break;
1739                         }
1740                 }
1741
1742                 /* If this rq is still suitable use it. */
1743                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1744                         break;
1745
1746                 /* try again */
1747                 double_unlock_balance(rq, lowest_rq);
1748                 lowest_rq = NULL;
1749         }
1750
1751         return lowest_rq;
1752 }
1753
1754 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1755 {
1756         struct task_struct *p;
1757
1758         if (!has_pushable_tasks(rq))
1759                 return NULL;
1760
1761         p = plist_first_entry(&rq->rt.pushable_tasks,
1762                               struct task_struct, pushable_tasks);
1763
1764         BUG_ON(rq->cpu != task_cpu(p));
1765         BUG_ON(task_current(rq, p));
1766         BUG_ON(p->nr_cpus_allowed <= 1);
1767
1768         BUG_ON(!task_on_rq_queued(p));
1769         BUG_ON(!rt_task(p));
1770
1771         return p;
1772 }
1773
1774 /*
1775  * If the current CPU has more than one RT task, see if the non
1776  * running task can migrate over to a CPU that is running a task
1777  * of lesser priority.
1778  */
1779 static int push_rt_task(struct rq *rq)
1780 {
1781         struct task_struct *next_task;
1782         struct rq *lowest_rq;
1783         int ret = 0;
1784
1785         if (!rq->rt.overloaded)
1786                 return 0;
1787
1788         next_task = pick_next_pushable_task(rq);
1789         if (!next_task)
1790                 return 0;
1791
1792 retry:
1793         if (unlikely(next_task == rq->curr)) {
1794                 WARN_ON(1);
1795                 return 0;
1796         }
1797
1798         /*
1799          * It's possible that the next_task slipped in of
1800          * higher priority than current. If that's the case
1801          * just reschedule current.
1802          */
1803         if (unlikely(next_task->prio < rq->curr->prio)) {
1804                 resched_curr(rq);
1805                 return 0;
1806         }
1807
1808         /* We might release rq lock */
1809         get_task_struct(next_task);
1810
1811         /* find_lock_lowest_rq locks the rq if found */
1812         lowest_rq = find_lock_lowest_rq(next_task, rq);
1813         if (!lowest_rq) {
1814                 struct task_struct *task;
1815                 /*
1816                  * find_lock_lowest_rq releases rq->lock
1817                  * so it is possible that next_task has migrated.
1818                  *
1819                  * We need to make sure that the task is still on the same
1820                  * run-queue and is also still the next task eligible for
1821                  * pushing.
1822                  */
1823                 task = pick_next_pushable_task(rq);
1824                 if (task == next_task) {
1825                         /*
1826                          * The task hasn't migrated, and is still the next
1827                          * eligible task, but we failed to find a run-queue
1828                          * to push it to.  Do not retry in this case, since
1829                          * other cpus will pull from us when ready.
1830                          */
1831                         goto out;
1832                 }
1833
1834                 if (!task)
1835                         /* No more tasks, just exit */
1836                         goto out;
1837
1838                 /*
1839                  * Something has shifted, try again.
1840                  */
1841                 put_task_struct(next_task);
1842                 next_task = task;
1843                 goto retry;
1844         }
1845
1846         deactivate_task(rq, next_task, 0);
1847         set_task_cpu(next_task, lowest_rq->cpu);
1848         activate_task(lowest_rq, next_task, 0);
1849         ret = 1;
1850
1851         resched_curr(lowest_rq);
1852
1853         double_unlock_balance(rq, lowest_rq);
1854
1855 out:
1856         put_task_struct(next_task);
1857
1858         return ret;
1859 }
1860
1861 static void push_rt_tasks(struct rq *rq)
1862 {
1863         /* push_rt_task will return true if it moved an RT */
1864         while (push_rt_task(rq))
1865                 ;
1866 }
1867
1868 #ifdef HAVE_RT_PUSH_IPI
1869
1870 /*
1871  * When a high priority task schedules out from a CPU and a lower priority
1872  * task is scheduled in, a check is made to see if there's any RT tasks
1873  * on other CPUs that are waiting to run because a higher priority RT task
1874  * is currently running on its CPU. In this case, the CPU with multiple RT
1875  * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1876  * up that may be able to run one of its non-running queued RT tasks.
1877  *
1878  * All CPUs with overloaded RT tasks need to be notified as there is currently
1879  * no way to know which of these CPUs have the highest priority task waiting
1880  * to run. Instead of trying to take a spinlock on each of these CPUs,
1881  * which has shown to cause large latency when done on machines with many
1882  * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1883  * RT tasks waiting to run.
1884  *
1885  * Just sending an IPI to each of the CPUs is also an issue, as on large
1886  * count CPU machines, this can cause an IPI storm on a CPU, especially
1887  * if its the only CPU with multiple RT tasks queued, and a large number
1888  * of CPUs scheduling a lower priority task at the same time.
1889  *
1890  * Each root domain has its own irq work function that can iterate over
1891  * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1892  * tassk must be checked if there's one or many CPUs that are lowering
1893  * their priority, there's a single irq work iterator that will try to
1894  * push off RT tasks that are waiting to run.
1895  *
1896  * When a CPU schedules a lower priority task, it will kick off the
1897  * irq work iterator that will jump to each CPU with overloaded RT tasks.
1898  * As it only takes the first CPU that schedules a lower priority task
1899  * to start the process, the rto_start variable is incremented and if
1900  * the atomic result is one, then that CPU will try to take the rto_lock.
1901  * This prevents high contention on the lock as the process handles all
1902  * CPUs scheduling lower priority tasks.
1903  *
1904  * All CPUs that are scheduling a lower priority task will increment the
1905  * rt_loop_next variable. This will make sure that the irq work iterator
1906  * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1907  * priority task, even if the iterator is in the middle of a scan. Incrementing
1908  * the rt_loop_next will cause the iterator to perform another scan.
1909  *
1910  */
1911 static int rto_next_cpu(struct root_domain *rd)
1912 {
1913         int next;
1914         int cpu;
1915
1916         /*
1917          * When starting the IPI RT pushing, the rto_cpu is set to -1,
1918          * rt_next_cpu() will simply return the first CPU found in
1919          * the rto_mask.
1920          *
1921          * If rto_next_cpu() is called with rto_cpu is a valid cpu, it
1922          * will return the next CPU found in the rto_mask.
1923          *
1924          * If there are no more CPUs left in the rto_mask, then a check is made
1925          * against rto_loop and rto_loop_next. rto_loop is only updated with
1926          * the rto_lock held, but any CPU may increment the rto_loop_next
1927          * without any locking.
1928          */
1929         for (;;) {
1930
1931                 /* When rto_cpu is -1 this acts like cpumask_first() */
1932                 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
1933
1934                 rd->rto_cpu = cpu;
1935
1936                 if (cpu < nr_cpu_ids)
1937                         return cpu;
1938
1939                 rd->rto_cpu = -1;
1940
1941                 /*
1942                  * ACQUIRE ensures we see the @rto_mask changes
1943                  * made prior to the @next value observed.
1944                  *
1945                  * Matches WMB in rt_set_overload().
1946                  */
1947                 next = atomic_read_acquire(&rd->rto_loop_next);
1948
1949                 if (rd->rto_loop == next)
1950                         break;
1951
1952                 rd->rto_loop = next;
1953         }
1954
1955         return -1;
1956 }
1957
1958 static inline bool rto_start_trylock(atomic_t *v)
1959 {
1960         return !atomic_cmpxchg_acquire(v, 0, 1);
1961 }
1962
1963 static inline void rto_start_unlock(atomic_t *v)
1964 {
1965         atomic_set_release(v, 0);
1966 }
1967
1968 static void tell_cpu_to_push(struct rq *rq)
1969 {
1970         int cpu = -1;
1971
1972         /* Keep the loop going if the IPI is currently active */
1973         atomic_inc(&rq->rd->rto_loop_next);
1974
1975         /* Only one CPU can initiate a loop at a time */
1976         if (!rto_start_trylock(&rq->rd->rto_loop_start))
1977                 return;
1978
1979         raw_spin_lock(&rq->rd->rto_lock);
1980
1981         /*
1982          * The rto_cpu is updated under the lock, if it has a valid cpu
1983          * then the IPI is still running and will continue due to the
1984          * update to loop_next, and nothing needs to be done here.
1985          * Otherwise it is finishing up and an ipi needs to be sent.
1986          */
1987         if (rq->rd->rto_cpu < 0)
1988                 cpu = rto_next_cpu(rq->rd);
1989
1990         raw_spin_unlock(&rq->rd->rto_lock);
1991
1992         rto_start_unlock(&rq->rd->rto_loop_start);
1993
1994         if (cpu >= 0) {
1995                 /* Make sure the rd does not get freed while pushing */
1996                 sched_get_rd(rq->rd);
1997                 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
1998         }
1999 }
2000
2001 /* Called from hardirq context */
2002 void rto_push_irq_work_func(struct irq_work *work)
2003 {
2004         struct root_domain *rd =
2005                 container_of(work, struct root_domain, rto_push_work);
2006         struct rq *rq;
2007         int cpu;
2008
2009         rq = this_rq();
2010
2011         /*
2012          * We do not need to grab the lock to check for has_pushable_tasks.
2013          * When it gets updated, a check is made if a push is possible.
2014          */
2015         if (has_pushable_tasks(rq)) {
2016                 raw_spin_lock(&rq->lock);
2017                 push_rt_tasks(rq);
2018                 raw_spin_unlock(&rq->lock);
2019         }
2020
2021         raw_spin_lock(&rd->rto_lock);
2022
2023         /* Pass the IPI to the next rt overloaded queue */
2024         cpu = rto_next_cpu(rd);
2025
2026         raw_spin_unlock(&rd->rto_lock);
2027
2028         if (cpu < 0) {
2029                 sched_put_rd(rd);
2030                 return;
2031         }
2032
2033         /* Try the next RT overloaded CPU */
2034         irq_work_queue_on(&rd->rto_push_work, cpu);
2035 }
2036 #endif /* HAVE_RT_PUSH_IPI */
2037
2038 static void pull_rt_task(struct rq *this_rq)
2039 {
2040         int this_cpu = this_rq->cpu, cpu;
2041         bool resched = false;
2042         struct task_struct *p;
2043         struct rq *src_rq;
2044         int rt_overload_count = rt_overloaded(this_rq);
2045
2046         if (likely(!rt_overload_count))
2047                 return;
2048
2049         /*
2050          * Match the barrier from rt_set_overloaded; this guarantees that if we
2051          * see overloaded we must also see the rto_mask bit.
2052          */
2053         smp_rmb();
2054
2055         /* If we are the only overloaded CPU do nothing */
2056         if (rt_overload_count == 1 &&
2057             cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2058                 return;
2059
2060 #ifdef HAVE_RT_PUSH_IPI
2061         if (sched_feat(RT_PUSH_IPI)) {
2062                 tell_cpu_to_push(this_rq);
2063                 return;
2064         }
2065 #endif
2066
2067         for_each_cpu(cpu, this_rq->rd->rto_mask) {
2068                 if (this_cpu == cpu)
2069                         continue;
2070
2071                 src_rq = cpu_rq(cpu);
2072
2073                 /*
2074                  * Don't bother taking the src_rq->lock if the next highest
2075                  * task is known to be lower-priority than our current task.
2076                  * This may look racy, but if this value is about to go
2077                  * logically higher, the src_rq will push this task away.
2078                  * And if its going logically lower, we do not care
2079                  */
2080                 if (src_rq->rt.highest_prio.next >=
2081                     this_rq->rt.highest_prio.curr)
2082                         continue;
2083
2084                 /*
2085                  * We can potentially drop this_rq's lock in
2086                  * double_lock_balance, and another CPU could
2087                  * alter this_rq
2088                  */
2089                 double_lock_balance(this_rq, src_rq);
2090
2091                 /*
2092                  * We can pull only a task, which is pushable
2093                  * on its rq, and no others.
2094                  */
2095                 p = pick_highest_pushable_task(src_rq, this_cpu);
2096
2097                 /*
2098                  * Do we have an RT task that preempts
2099                  * the to-be-scheduled task?
2100                  */
2101                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2102                         WARN_ON(p == src_rq->curr);
2103                         WARN_ON(!task_on_rq_queued(p));
2104
2105                         /*
2106                          * There's a chance that p is higher in priority
2107                          * than what's currently running on its cpu.
2108                          * This is just that p is wakeing up and hasn't
2109                          * had a chance to schedule. We only pull
2110                          * p if it is lower in priority than the
2111                          * current task on the run queue
2112                          */
2113                         if (p->prio < src_rq->curr->prio)
2114                                 goto skip;
2115
2116                         resched = true;
2117
2118                         deactivate_task(src_rq, p, 0);
2119                         set_task_cpu(p, this_cpu);
2120                         activate_task(this_rq, p, 0);
2121                         /*
2122                          * We continue with the search, just in
2123                          * case there's an even higher prio task
2124                          * in another runqueue. (low likelihood
2125                          * but possible)
2126                          */
2127                 }
2128 skip:
2129                 double_unlock_balance(this_rq, src_rq);
2130         }
2131
2132         if (resched)
2133                 resched_curr(this_rq);
2134 }
2135
2136 /*
2137  * If we are not running and we are not going to reschedule soon, we should
2138  * try to push tasks away now
2139  */
2140 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2141 {
2142         if (!task_running(rq, p) &&
2143             !test_tsk_need_resched(rq->curr) &&
2144             p->nr_cpus_allowed > 1 &&
2145             (dl_task(rq->curr) || rt_task(rq->curr)) &&
2146             (rq->curr->nr_cpus_allowed < 2 ||
2147              rq->curr->prio <= p->prio))
2148                 push_rt_tasks(rq);
2149 }
2150
2151 /* Assumes rq->lock is held */
2152 static void rq_online_rt(struct rq *rq)
2153 {
2154         if (rq->rt.overloaded)
2155                 rt_set_overload(rq);
2156
2157         __enable_runtime(rq);
2158
2159         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2160 }
2161
2162 /* Assumes rq->lock is held */
2163 static void rq_offline_rt(struct rq *rq)
2164 {
2165         if (rq->rt.overloaded)
2166                 rt_clear_overload(rq);
2167
2168         __disable_runtime(rq);
2169
2170         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2171 }
2172
2173 /*
2174  * When switch from the rt queue, we bring ourselves to a position
2175  * that we might want to pull RT tasks from other runqueues.
2176  */
2177 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2178 {
2179         /*
2180          * If there are other RT tasks then we will reschedule
2181          * and the scheduling of the other RT tasks will handle
2182          * the balancing. But if we are the last RT task
2183          * we may need to handle the pulling of RT tasks
2184          * now.
2185          */
2186         if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2187                 return;
2188
2189         queue_pull_task(rq);
2190 }
2191
2192 void __init init_sched_rt_class(void)
2193 {
2194         unsigned int i;
2195
2196         for_each_possible_cpu(i) {
2197                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2198                                         GFP_KERNEL, cpu_to_node(i));
2199         }
2200 }
2201 #endif /* CONFIG_SMP */
2202
2203 /*
2204  * When switching a task to RT, we may overload the runqueue
2205  * with RT tasks. In this case we try to push them off to
2206  * other runqueues.
2207  */
2208 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2209 {
2210         /*
2211          * If we are already running, then there's nothing
2212          * that needs to be done. But if we are not running
2213          * we may need to preempt the current running task.
2214          * If that current running task is also an RT task
2215          * then see if we can move to another run queue.
2216          */
2217         if (task_on_rq_queued(p) && rq->curr != p) {
2218 #ifdef CONFIG_SMP
2219                 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2220                         queue_push_tasks(rq);
2221 #endif /* CONFIG_SMP */
2222                 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2223                         resched_curr(rq);
2224         }
2225 }
2226
2227 /*
2228  * Priority of the task has changed. This may cause
2229  * us to initiate a push or pull.
2230  */
2231 static void
2232 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2233 {
2234         if (!task_on_rq_queued(p))
2235                 return;
2236
2237         if (rq->curr == p) {
2238 #ifdef CONFIG_SMP
2239                 /*
2240                  * If our priority decreases while running, we
2241                  * may need to pull tasks to this runqueue.
2242                  */
2243                 if (oldprio < p->prio)
2244                         queue_pull_task(rq);
2245
2246                 /*
2247                  * If there's a higher priority task waiting to run
2248                  * then reschedule.
2249                  */
2250                 if (p->prio > rq->rt.highest_prio.curr)
2251                         resched_curr(rq);
2252 #else
2253                 /* For UP simply resched on drop of prio */
2254                 if (oldprio < p->prio)
2255                         resched_curr(rq);
2256 #endif /* CONFIG_SMP */
2257         } else {
2258                 /*
2259                  * This task is not running, but if it is
2260                  * greater than the current running task
2261                  * then reschedule.
2262                  */
2263                 if (p->prio < rq->curr->prio)
2264                         resched_curr(rq);
2265         }
2266 }
2267
2268 #ifdef CONFIG_POSIX_TIMERS
2269 static void watchdog(struct rq *rq, struct task_struct *p)
2270 {
2271         unsigned long soft, hard;
2272
2273         /* max may change after cur was read, this will be fixed next tick */
2274         soft = task_rlimit(p, RLIMIT_RTTIME);
2275         hard = task_rlimit_max(p, RLIMIT_RTTIME);
2276
2277         if (soft != RLIM_INFINITY) {
2278                 unsigned long next;
2279
2280                 if (p->rt.watchdog_stamp != jiffies) {
2281                         p->rt.timeout++;
2282                         p->rt.watchdog_stamp = jiffies;
2283                 }
2284
2285                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2286                 if (p->rt.timeout > next)
2287                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2288         }
2289 }
2290 #else
2291 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2292 #endif
2293
2294 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2295 {
2296         struct sched_rt_entity *rt_se = &p->rt;
2297
2298         update_curr_rt(rq);
2299
2300         watchdog(rq, p);
2301
2302         /*
2303          * RR tasks need a special form of timeslice management.
2304          * FIFO tasks have no timeslices.
2305          */
2306         if (p->policy != SCHED_RR)
2307                 return;
2308
2309         if (--p->rt.time_slice)
2310                 return;
2311
2312         p->rt.time_slice = sched_rr_timeslice;
2313
2314         /*
2315          * Requeue to the end of queue if we (and all of our ancestors) are not
2316          * the only element on the queue
2317          */
2318         for_each_sched_rt_entity(rt_se) {
2319                 if (rt_se->run_list.prev != rt_se->run_list.next) {
2320                         requeue_task_rt(rq, p, 0);
2321                         resched_curr(rq);
2322                         return;
2323                 }
2324         }
2325 }
2326
2327 static void set_curr_task_rt(struct rq *rq)
2328 {
2329         struct task_struct *p = rq->curr;
2330
2331         p->se.exec_start = rq_clock_task(rq);
2332
2333         /* The running task is never eligible for pushing */
2334         dequeue_pushable_task(rq, p);
2335 }
2336
2337 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2338 {
2339         /*
2340          * Time slice is 0 for SCHED_FIFO tasks
2341          */
2342         if (task->policy == SCHED_RR)
2343                 return sched_rr_timeslice;
2344         else
2345                 return 0;
2346 }
2347
2348 const struct sched_class rt_sched_class = {
2349         .next                   = &fair_sched_class,
2350         .enqueue_task           = enqueue_task_rt,
2351         .dequeue_task           = dequeue_task_rt,
2352         .yield_task             = yield_task_rt,
2353
2354         .check_preempt_curr     = check_preempt_curr_rt,
2355
2356         .pick_next_task         = pick_next_task_rt,
2357         .put_prev_task          = put_prev_task_rt,
2358
2359 #ifdef CONFIG_SMP
2360         .select_task_rq         = select_task_rq_rt,
2361
2362         .set_cpus_allowed       = set_cpus_allowed_common,
2363         .rq_online              = rq_online_rt,
2364         .rq_offline             = rq_offline_rt,
2365         .task_woken             = task_woken_rt,
2366         .switched_from          = switched_from_rt,
2367 #endif
2368
2369         .set_curr_task          = set_curr_task_rt,
2370         .task_tick              = task_tick_rt,
2371
2372         .get_rr_interval        = get_rr_interval_rt,
2373
2374         .prio_changed           = prio_changed_rt,
2375         .switched_to            = switched_to_rt,
2376
2377         .update_curr            = update_curr_rt,
2378 };
2379
2380 #ifdef CONFIG_RT_GROUP_SCHED
2381 /*
2382  * Ensure that the real time constraints are schedulable.
2383  */
2384 static DEFINE_MUTEX(rt_constraints_mutex);
2385
2386 /* Must be called with tasklist_lock held */
2387 static inline int tg_has_rt_tasks(struct task_group *tg)
2388 {
2389         struct task_struct *g, *p;
2390
2391         /*
2392          * Autogroups do not have RT tasks; see autogroup_create().
2393          */
2394         if (task_group_is_autogroup(tg))
2395                 return 0;
2396
2397         for_each_process_thread(g, p) {
2398                 if (rt_task(p) && task_group(p) == tg)
2399                         return 1;
2400         }
2401
2402         return 0;
2403 }
2404
2405 struct rt_schedulable_data {
2406         struct task_group *tg;
2407         u64 rt_period;
2408         u64 rt_runtime;
2409 };
2410
2411 static int tg_rt_schedulable(struct task_group *tg, void *data)
2412 {
2413         struct rt_schedulable_data *d = data;
2414         struct task_group *child;
2415         unsigned long total, sum = 0;
2416         u64 period, runtime;
2417
2418         period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2419         runtime = tg->rt_bandwidth.rt_runtime;
2420
2421         if (tg == d->tg) {
2422                 period = d->rt_period;
2423                 runtime = d->rt_runtime;
2424         }
2425
2426         /*
2427          * Cannot have more runtime than the period.
2428          */
2429         if (runtime > period && runtime != RUNTIME_INF)
2430                 return -EINVAL;
2431
2432         /*
2433          * Ensure we don't starve existing RT tasks.
2434          */
2435         if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
2436                 return -EBUSY;
2437
2438         total = to_ratio(period, runtime);
2439
2440         /*
2441          * Nobody can have more than the global setting allows.
2442          */
2443         if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2444                 return -EINVAL;
2445
2446         /*
2447          * The sum of our children's runtime should not exceed our own.
2448          */
2449         list_for_each_entry_rcu(child, &tg->children, siblings) {
2450                 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2451                 runtime = child->rt_bandwidth.rt_runtime;
2452
2453                 if (child == d->tg) {
2454                         period = d->rt_period;
2455                         runtime = d->rt_runtime;
2456                 }
2457
2458                 sum += to_ratio(period, runtime);
2459         }
2460
2461         if (sum > total)
2462                 return -EINVAL;
2463
2464         return 0;
2465 }
2466
2467 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2468 {
2469         int ret;
2470
2471         struct rt_schedulable_data data = {
2472                 .tg = tg,
2473                 .rt_period = period,
2474                 .rt_runtime = runtime,
2475         };
2476
2477         rcu_read_lock();
2478         ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2479         rcu_read_unlock();
2480
2481         return ret;
2482 }
2483
2484 static int tg_set_rt_bandwidth(struct task_group *tg,
2485                 u64 rt_period, u64 rt_runtime)
2486 {
2487         int i, err = 0;
2488
2489         /*
2490          * Disallowing the root group RT runtime is BAD, it would disallow the
2491          * kernel creating (and or operating) RT threads.
2492          */
2493         if (tg == &root_task_group && rt_runtime == 0)
2494                 return -EINVAL;
2495
2496         /* No period doesn't make any sense. */
2497         if (rt_period == 0)
2498                 return -EINVAL;
2499
2500         mutex_lock(&rt_constraints_mutex);
2501         read_lock(&tasklist_lock);
2502         err = __rt_schedulable(tg, rt_period, rt_runtime);
2503         if (err)
2504                 goto unlock;
2505
2506         raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2507         tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2508         tg->rt_bandwidth.rt_runtime = rt_runtime;
2509
2510         for_each_possible_cpu(i) {
2511                 struct rt_rq *rt_rq = tg->rt_rq[i];
2512
2513                 raw_spin_lock(&rt_rq->rt_runtime_lock);
2514                 rt_rq->rt_runtime = rt_runtime;
2515                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2516         }
2517         raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2518 unlock:
2519         read_unlock(&tasklist_lock);
2520         mutex_unlock(&rt_constraints_mutex);
2521
2522         return err;
2523 }
2524
2525 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2526 {
2527         u64 rt_runtime, rt_period;
2528
2529         rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2530         rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2531         if (rt_runtime_us < 0)
2532                 rt_runtime = RUNTIME_INF;
2533
2534         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2535 }
2536
2537 long sched_group_rt_runtime(struct task_group *tg)
2538 {
2539         u64 rt_runtime_us;
2540
2541         if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2542                 return -1;
2543
2544         rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2545         do_div(rt_runtime_us, NSEC_PER_USEC);
2546         return rt_runtime_us;
2547 }
2548
2549 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2550 {
2551         u64 rt_runtime, rt_period;
2552
2553         rt_period = rt_period_us * NSEC_PER_USEC;
2554         rt_runtime = tg->rt_bandwidth.rt_runtime;
2555
2556         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2557 }
2558
2559 long sched_group_rt_period(struct task_group *tg)
2560 {
2561         u64 rt_period_us;
2562
2563         rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2564         do_div(rt_period_us, NSEC_PER_USEC);
2565         return rt_period_us;
2566 }
2567
2568 static int sched_rt_global_constraints(void)
2569 {
2570         int ret = 0;
2571
2572         mutex_lock(&rt_constraints_mutex);
2573         read_lock(&tasklist_lock);
2574         ret = __rt_schedulable(NULL, 0, 0);
2575         read_unlock(&tasklist_lock);
2576         mutex_unlock(&rt_constraints_mutex);
2577
2578         return ret;
2579 }
2580
2581 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2582 {
2583         /* Don't accept realtime tasks when there is no way for them to run */
2584         if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2585                 return 0;
2586
2587         return 1;
2588 }
2589
2590 #else /* !CONFIG_RT_GROUP_SCHED */
2591 static int sched_rt_global_constraints(void)
2592 {
2593         unsigned long flags;
2594         int i;
2595
2596         raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2597         for_each_possible_cpu(i) {
2598                 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2599
2600                 raw_spin_lock(&rt_rq->rt_runtime_lock);
2601                 rt_rq->rt_runtime = global_rt_runtime();
2602                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2603         }
2604         raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2605
2606         return 0;
2607 }
2608 #endif /* CONFIG_RT_GROUP_SCHED */
2609
2610 static int sched_rt_global_validate(void)
2611 {
2612         if (sysctl_sched_rt_period <= 0)
2613                 return -EINVAL;
2614
2615         if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2616                 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
2617                 return -EINVAL;
2618
2619         return 0;
2620 }
2621
2622 static void sched_rt_do_global(void)
2623 {
2624         def_rt_bandwidth.rt_runtime = global_rt_runtime();
2625         def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2626 }
2627
2628 int sched_rt_handler(struct ctl_table *table, int write,
2629                 void __user *buffer, size_t *lenp,
2630                 loff_t *ppos)
2631 {
2632         int old_period, old_runtime;
2633         static DEFINE_MUTEX(mutex);
2634         int ret;
2635
2636         mutex_lock(&mutex);
2637         old_period = sysctl_sched_rt_period;
2638         old_runtime = sysctl_sched_rt_runtime;
2639
2640         ret = proc_dointvec(table, write, buffer, lenp, ppos);
2641
2642         if (!ret && write) {
2643                 ret = sched_rt_global_validate();
2644                 if (ret)
2645                         goto undo;
2646
2647                 ret = sched_dl_global_validate();
2648                 if (ret)
2649                         goto undo;
2650
2651                 ret = sched_rt_global_constraints();
2652                 if (ret)
2653                         goto undo;
2654
2655                 sched_rt_do_global();
2656                 sched_dl_do_global();
2657         }
2658         if (0) {
2659 undo:
2660                 sysctl_sched_rt_period = old_period;
2661                 sysctl_sched_rt_runtime = old_runtime;
2662         }
2663         mutex_unlock(&mutex);
2664
2665         return ret;
2666 }
2667
2668 int sched_rr_handler(struct ctl_table *table, int write,
2669                 void __user *buffer, size_t *lenp,
2670                 loff_t *ppos)
2671 {
2672         int ret;
2673         static DEFINE_MUTEX(mutex);
2674
2675         mutex_lock(&mutex);
2676         ret = proc_dointvec(table, write, buffer, lenp, ppos);
2677         /*
2678          * Make sure that internally we keep jiffies.
2679          * Also, writing zero resets the timeslice to default:
2680          */
2681         if (!ret && write) {
2682                 sched_rr_timeslice =
2683                         sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2684                         msecs_to_jiffies(sysctl_sched_rr_timeslice);
2685         }
2686         mutex_unlock(&mutex);
2687         return ret;
2688 }
2689
2690 #ifdef CONFIG_SCHED_DEBUG
2691 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2692
2693 void print_rt_stats(struct seq_file *m, int cpu)
2694 {
2695         rt_rq_iter_t iter;
2696         struct rt_rq *rt_rq;
2697
2698         rcu_read_lock();
2699         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2700                 print_rt_rq(m, cpu, rt_rq);
2701         rcu_read_unlock();
2702 }
2703 #endif /* CONFIG_SCHED_DEBUG */