Merge branch 'for-4.15' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup
[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 delta_exec;
954
955         if (curr->sched_class != &rt_sched_class)
956                 return;
957
958         delta_exec = rq_clock_task(rq) - curr->se.exec_start;
959         if (unlikely((s64)delta_exec <= 0))
960                 return;
961
962         /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
963         cpufreq_update_util(rq, SCHED_CPUFREQ_RT);
964
965         schedstat_set(curr->se.statistics.exec_max,
966                       max(curr->se.statistics.exec_max, delta_exec));
967
968         curr->se.sum_exec_runtime += delta_exec;
969         account_group_exec_runtime(curr, delta_exec);
970
971         curr->se.exec_start = rq_clock_task(rq);
972         cgroup_account_cputime(curr, delta_exec);
973
974         sched_rt_avg_update(rq, delta_exec);
975
976         if (!rt_bandwidth_enabled())
977                 return;
978
979         for_each_sched_rt_entity(rt_se) {
980                 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
981
982                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
983                         raw_spin_lock(&rt_rq->rt_runtime_lock);
984                         rt_rq->rt_time += delta_exec;
985                         if (sched_rt_runtime_exceeded(rt_rq))
986                                 resched_curr(rq);
987                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
988                 }
989         }
990 }
991
992 static void
993 dequeue_top_rt_rq(struct rt_rq *rt_rq)
994 {
995         struct rq *rq = rq_of_rt_rq(rt_rq);
996
997         BUG_ON(&rq->rt != rt_rq);
998
999         if (!rt_rq->rt_queued)
1000                 return;
1001
1002         BUG_ON(!rq->nr_running);
1003
1004         sub_nr_running(rq, rt_rq->rt_nr_running);
1005         rt_rq->rt_queued = 0;
1006 }
1007
1008 static void
1009 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1010 {
1011         struct rq *rq = rq_of_rt_rq(rt_rq);
1012
1013         BUG_ON(&rq->rt != rt_rq);
1014
1015         if (rt_rq->rt_queued)
1016                 return;
1017         if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1018                 return;
1019
1020         add_nr_running(rq, rt_rq->rt_nr_running);
1021         rt_rq->rt_queued = 1;
1022 }
1023
1024 #if defined CONFIG_SMP
1025
1026 static void
1027 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1028 {
1029         struct rq *rq = rq_of_rt_rq(rt_rq);
1030
1031 #ifdef CONFIG_RT_GROUP_SCHED
1032         /*
1033          * Change rq's cpupri only if rt_rq is the top queue.
1034          */
1035         if (&rq->rt != rt_rq)
1036                 return;
1037 #endif
1038         if (rq->online && prio < prev_prio)
1039                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1040 }
1041
1042 static void
1043 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1044 {
1045         struct rq *rq = rq_of_rt_rq(rt_rq);
1046
1047 #ifdef CONFIG_RT_GROUP_SCHED
1048         /*
1049          * Change rq's cpupri only if rt_rq is the top queue.
1050          */
1051         if (&rq->rt != rt_rq)
1052                 return;
1053 #endif
1054         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1055                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1056 }
1057
1058 #else /* CONFIG_SMP */
1059
1060 static inline
1061 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1062 static inline
1063 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1064
1065 #endif /* CONFIG_SMP */
1066
1067 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1068 static void
1069 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1070 {
1071         int prev_prio = rt_rq->highest_prio.curr;
1072
1073         if (prio < prev_prio)
1074                 rt_rq->highest_prio.curr = prio;
1075
1076         inc_rt_prio_smp(rt_rq, prio, prev_prio);
1077 }
1078
1079 static void
1080 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1081 {
1082         int prev_prio = rt_rq->highest_prio.curr;
1083
1084         if (rt_rq->rt_nr_running) {
1085
1086                 WARN_ON(prio < prev_prio);
1087
1088                 /*
1089                  * This may have been our highest task, and therefore
1090                  * we may have some recomputation to do
1091                  */
1092                 if (prio == prev_prio) {
1093                         struct rt_prio_array *array = &rt_rq->active;
1094
1095                         rt_rq->highest_prio.curr =
1096                                 sched_find_first_bit(array->bitmap);
1097                 }
1098
1099         } else
1100                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1101
1102         dec_rt_prio_smp(rt_rq, prio, prev_prio);
1103 }
1104
1105 #else
1106
1107 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1108 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1109
1110 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1111
1112 #ifdef CONFIG_RT_GROUP_SCHED
1113
1114 static void
1115 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1116 {
1117         if (rt_se_boosted(rt_se))
1118                 rt_rq->rt_nr_boosted++;
1119
1120         if (rt_rq->tg)
1121                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1122 }
1123
1124 static void
1125 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1126 {
1127         if (rt_se_boosted(rt_se))
1128                 rt_rq->rt_nr_boosted--;
1129
1130         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1131 }
1132
1133 #else /* CONFIG_RT_GROUP_SCHED */
1134
1135 static void
1136 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1137 {
1138         start_rt_bandwidth(&def_rt_bandwidth);
1139 }
1140
1141 static inline
1142 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1143
1144 #endif /* CONFIG_RT_GROUP_SCHED */
1145
1146 static inline
1147 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1148 {
1149         struct rt_rq *group_rq = group_rt_rq(rt_se);
1150
1151         if (group_rq)
1152                 return group_rq->rt_nr_running;
1153         else
1154                 return 1;
1155 }
1156
1157 static inline
1158 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1159 {
1160         struct rt_rq *group_rq = group_rt_rq(rt_se);
1161         struct task_struct *tsk;
1162
1163         if (group_rq)
1164                 return group_rq->rr_nr_running;
1165
1166         tsk = rt_task_of(rt_se);
1167
1168         return (tsk->policy == SCHED_RR) ? 1 : 0;
1169 }
1170
1171 static inline
1172 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1173 {
1174         int prio = rt_se_prio(rt_se);
1175
1176         WARN_ON(!rt_prio(prio));
1177         rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1178         rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1179
1180         inc_rt_prio(rt_rq, prio);
1181         inc_rt_migration(rt_se, rt_rq);
1182         inc_rt_group(rt_se, rt_rq);
1183 }
1184
1185 static inline
1186 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1187 {
1188         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1189         WARN_ON(!rt_rq->rt_nr_running);
1190         rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1191         rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1192
1193         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1194         dec_rt_migration(rt_se, rt_rq);
1195         dec_rt_group(rt_se, rt_rq);
1196 }
1197
1198 /*
1199  * Change rt_se->run_list location unless SAVE && !MOVE
1200  *
1201  * assumes ENQUEUE/DEQUEUE flags match
1202  */
1203 static inline bool move_entity(unsigned int flags)
1204 {
1205         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1206                 return false;
1207
1208         return true;
1209 }
1210
1211 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1212 {
1213         list_del_init(&rt_se->run_list);
1214
1215         if (list_empty(array->queue + rt_se_prio(rt_se)))
1216                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1217
1218         rt_se->on_list = 0;
1219 }
1220
1221 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1222 {
1223         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1224         struct rt_prio_array *array = &rt_rq->active;
1225         struct rt_rq *group_rq = group_rt_rq(rt_se);
1226         struct list_head *queue = array->queue + rt_se_prio(rt_se);
1227
1228         /*
1229          * Don't enqueue the group if its throttled, or when empty.
1230          * The latter is a consequence of the former when a child group
1231          * get throttled and the current group doesn't have any other
1232          * active members.
1233          */
1234         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1235                 if (rt_se->on_list)
1236                         __delist_rt_entity(rt_se, array);
1237                 return;
1238         }
1239
1240         if (move_entity(flags)) {
1241                 WARN_ON_ONCE(rt_se->on_list);
1242                 if (flags & ENQUEUE_HEAD)
1243                         list_add(&rt_se->run_list, queue);
1244                 else
1245                         list_add_tail(&rt_se->run_list, queue);
1246
1247                 __set_bit(rt_se_prio(rt_se), array->bitmap);
1248                 rt_se->on_list = 1;
1249         }
1250         rt_se->on_rq = 1;
1251
1252         inc_rt_tasks(rt_se, rt_rq);
1253 }
1254
1255 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1256 {
1257         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1258         struct rt_prio_array *array = &rt_rq->active;
1259
1260         if (move_entity(flags)) {
1261                 WARN_ON_ONCE(!rt_se->on_list);
1262                 __delist_rt_entity(rt_se, array);
1263         }
1264         rt_se->on_rq = 0;
1265
1266         dec_rt_tasks(rt_se, rt_rq);
1267 }
1268
1269 /*
1270  * Because the prio of an upper entry depends on the lower
1271  * entries, we must remove entries top - down.
1272  */
1273 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1274 {
1275         struct sched_rt_entity *back = NULL;
1276
1277         for_each_sched_rt_entity(rt_se) {
1278                 rt_se->back = back;
1279                 back = rt_se;
1280         }
1281
1282         dequeue_top_rt_rq(rt_rq_of_se(back));
1283
1284         for (rt_se = back; rt_se; rt_se = rt_se->back) {
1285                 if (on_rt_rq(rt_se))
1286                         __dequeue_rt_entity(rt_se, flags);
1287         }
1288 }
1289
1290 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1291 {
1292         struct rq *rq = rq_of_rt_se(rt_se);
1293
1294         dequeue_rt_stack(rt_se, flags);
1295         for_each_sched_rt_entity(rt_se)
1296                 __enqueue_rt_entity(rt_se, flags);
1297         enqueue_top_rt_rq(&rq->rt);
1298 }
1299
1300 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1301 {
1302         struct rq *rq = rq_of_rt_se(rt_se);
1303
1304         dequeue_rt_stack(rt_se, flags);
1305
1306         for_each_sched_rt_entity(rt_se) {
1307                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1308
1309                 if (rt_rq && rt_rq->rt_nr_running)
1310                         __enqueue_rt_entity(rt_se, flags);
1311         }
1312         enqueue_top_rt_rq(&rq->rt);
1313 }
1314
1315 /*
1316  * Adding/removing a task to/from a priority array:
1317  */
1318 static void
1319 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1320 {
1321         struct sched_rt_entity *rt_se = &p->rt;
1322
1323         if (flags & ENQUEUE_WAKEUP)
1324                 rt_se->timeout = 0;
1325
1326         enqueue_rt_entity(rt_se, flags);
1327
1328         if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1329                 enqueue_pushable_task(rq, p);
1330 }
1331
1332 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1333 {
1334         struct sched_rt_entity *rt_se = &p->rt;
1335
1336         update_curr_rt(rq);
1337         dequeue_rt_entity(rt_se, flags);
1338
1339         dequeue_pushable_task(rq, p);
1340 }
1341
1342 /*
1343  * Put task to the head or the end of the run list without the overhead of
1344  * dequeue followed by enqueue.
1345  */
1346 static void
1347 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1348 {
1349         if (on_rt_rq(rt_se)) {
1350                 struct rt_prio_array *array = &rt_rq->active;
1351                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1352
1353                 if (head)
1354                         list_move(&rt_se->run_list, queue);
1355                 else
1356                         list_move_tail(&rt_se->run_list, queue);
1357         }
1358 }
1359
1360 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1361 {
1362         struct sched_rt_entity *rt_se = &p->rt;
1363         struct rt_rq *rt_rq;
1364
1365         for_each_sched_rt_entity(rt_se) {
1366                 rt_rq = rt_rq_of_se(rt_se);
1367                 requeue_rt_entity(rt_rq, rt_se, head);
1368         }
1369 }
1370
1371 static void yield_task_rt(struct rq *rq)
1372 {
1373         requeue_task_rt(rq, rq->curr, 0);
1374 }
1375
1376 #ifdef CONFIG_SMP
1377 static int find_lowest_rq(struct task_struct *task);
1378
1379 static int
1380 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1381 {
1382         struct task_struct *curr;
1383         struct rq *rq;
1384
1385         /* For anything but wake ups, just return the task_cpu */
1386         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1387                 goto out;
1388
1389         rq = cpu_rq(cpu);
1390
1391         rcu_read_lock();
1392         curr = READ_ONCE(rq->curr); /* unlocked access */
1393
1394         /*
1395          * If the current task on @p's runqueue is an RT task, then
1396          * try to see if we can wake this RT task up on another
1397          * runqueue. Otherwise simply start this RT task
1398          * on its current runqueue.
1399          *
1400          * We want to avoid overloading runqueues. If the woken
1401          * task is a higher priority, then it will stay on this CPU
1402          * and the lower prio task should be moved to another CPU.
1403          * Even though this will probably make the lower prio task
1404          * lose its cache, we do not want to bounce a higher task
1405          * around just because it gave up its CPU, perhaps for a
1406          * lock?
1407          *
1408          * For equal prio tasks, we just let the scheduler sort it out.
1409          *
1410          * Otherwise, just let it ride on the affined RQ and the
1411          * post-schedule router will push the preempted task away
1412          *
1413          * This test is optimistic, if we get it wrong the load-balancer
1414          * will have to sort it out.
1415          */
1416         if (curr && unlikely(rt_task(curr)) &&
1417             (curr->nr_cpus_allowed < 2 ||
1418              curr->prio <= p->prio)) {
1419                 int target = find_lowest_rq(p);
1420
1421                 /*
1422                  * Don't bother moving it if the destination CPU is
1423                  * not running a lower priority task.
1424                  */
1425                 if (target != -1 &&
1426                     p->prio < cpu_rq(target)->rt.highest_prio.curr)
1427                         cpu = target;
1428         }
1429         rcu_read_unlock();
1430
1431 out:
1432         return cpu;
1433 }
1434
1435 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1436 {
1437         /*
1438          * Current can't be migrated, useless to reschedule,
1439          * let's hope p can move out.
1440          */
1441         if (rq->curr->nr_cpus_allowed == 1 ||
1442             !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1443                 return;
1444
1445         /*
1446          * p is migratable, so let's not schedule it and
1447          * see if it is pushed or pulled somewhere else.
1448          */
1449         if (p->nr_cpus_allowed != 1
1450             && cpupri_find(&rq->rd->cpupri, p, NULL))
1451                 return;
1452
1453         /*
1454          * There appears to be other cpus that can accept
1455          * current and none to run 'p', so lets reschedule
1456          * to try and push current away:
1457          */
1458         requeue_task_rt(rq, p, 1);
1459         resched_curr(rq);
1460 }
1461
1462 #endif /* CONFIG_SMP */
1463
1464 /*
1465  * Preempt the current task with a newly woken task if needed:
1466  */
1467 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1468 {
1469         if (p->prio < rq->curr->prio) {
1470                 resched_curr(rq);
1471                 return;
1472         }
1473
1474 #ifdef CONFIG_SMP
1475         /*
1476          * If:
1477          *
1478          * - the newly woken task is of equal priority to the current task
1479          * - the newly woken task is non-migratable while current is migratable
1480          * - current will be preempted on the next reschedule
1481          *
1482          * we should check to see if current can readily move to a different
1483          * cpu.  If so, we will reschedule to allow the push logic to try
1484          * to move current somewhere else, making room for our non-migratable
1485          * task.
1486          */
1487         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1488                 check_preempt_equal_prio(rq, p);
1489 #endif
1490 }
1491
1492 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1493                                                    struct rt_rq *rt_rq)
1494 {
1495         struct rt_prio_array *array = &rt_rq->active;
1496         struct sched_rt_entity *next = NULL;
1497         struct list_head *queue;
1498         int idx;
1499
1500         idx = sched_find_first_bit(array->bitmap);
1501         BUG_ON(idx >= MAX_RT_PRIO);
1502
1503         queue = array->queue + idx;
1504         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1505
1506         return next;
1507 }
1508
1509 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1510 {
1511         struct sched_rt_entity *rt_se;
1512         struct task_struct *p;
1513         struct rt_rq *rt_rq  = &rq->rt;
1514
1515         do {
1516                 rt_se = pick_next_rt_entity(rq, rt_rq);
1517                 BUG_ON(!rt_se);
1518                 rt_rq = group_rt_rq(rt_se);
1519         } while (rt_rq);
1520
1521         p = rt_task_of(rt_se);
1522         p->se.exec_start = rq_clock_task(rq);
1523
1524         return p;
1525 }
1526
1527 static struct task_struct *
1528 pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1529 {
1530         struct task_struct *p;
1531         struct rt_rq *rt_rq = &rq->rt;
1532
1533         if (need_pull_rt_task(rq, prev)) {
1534                 /*
1535                  * This is OK, because current is on_cpu, which avoids it being
1536                  * picked for load-balance and preemption/IRQs are still
1537                  * disabled avoiding further scheduler activity on it and we're
1538                  * being very careful to re-start the picking loop.
1539                  */
1540                 rq_unpin_lock(rq, rf);
1541                 pull_rt_task(rq);
1542                 rq_repin_lock(rq, rf);
1543                 /*
1544                  * pull_rt_task() can drop (and re-acquire) rq->lock; this
1545                  * means a dl or stop task can slip in, in which case we need
1546                  * to re-start task selection.
1547                  */
1548                 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1549                              rq->dl.dl_nr_running))
1550                         return RETRY_TASK;
1551         }
1552
1553         /*
1554          * We may dequeue prev's rt_rq in put_prev_task().
1555          * So, we update time before rt_nr_running check.
1556          */
1557         if (prev->sched_class == &rt_sched_class)
1558                 update_curr_rt(rq);
1559
1560         if (!rt_rq->rt_queued)
1561                 return NULL;
1562
1563         put_prev_task(rq, prev);
1564
1565         p = _pick_next_task_rt(rq);
1566
1567         /* The running task is never eligible for pushing */
1568         dequeue_pushable_task(rq, p);
1569
1570         queue_push_tasks(rq);
1571
1572         return p;
1573 }
1574
1575 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1576 {
1577         update_curr_rt(rq);
1578
1579         /*
1580          * The previous task needs to be made eligible for pushing
1581          * if it is still active
1582          */
1583         if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1584                 enqueue_pushable_task(rq, p);
1585 }
1586
1587 #ifdef CONFIG_SMP
1588
1589 /* Only try algorithms three times */
1590 #define RT_MAX_TRIES 3
1591
1592 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1593 {
1594         if (!task_running(rq, p) &&
1595             cpumask_test_cpu(cpu, &p->cpus_allowed))
1596                 return 1;
1597         return 0;
1598 }
1599
1600 /*
1601  * Return the highest pushable rq's task, which is suitable to be executed
1602  * on the cpu, NULL otherwise
1603  */
1604 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1605 {
1606         struct plist_head *head = &rq->rt.pushable_tasks;
1607         struct task_struct *p;
1608
1609         if (!has_pushable_tasks(rq))
1610                 return NULL;
1611
1612         plist_for_each_entry(p, head, pushable_tasks) {
1613                 if (pick_rt_task(rq, p, cpu))
1614                         return p;
1615         }
1616
1617         return NULL;
1618 }
1619
1620 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1621
1622 static int find_lowest_rq(struct task_struct *task)
1623 {
1624         struct sched_domain *sd;
1625         struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1626         int this_cpu = smp_processor_id();
1627         int cpu      = task_cpu(task);
1628
1629         /* Make sure the mask is initialized first */
1630         if (unlikely(!lowest_mask))
1631                 return -1;
1632
1633         if (task->nr_cpus_allowed == 1)
1634                 return -1; /* No other targets possible */
1635
1636         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1637                 return -1; /* No targets found */
1638
1639         /*
1640          * At this point we have built a mask of cpus representing the
1641          * lowest priority tasks in the system.  Now we want to elect
1642          * the best one based on our affinity and topology.
1643          *
1644          * We prioritize the last cpu that the task executed on since
1645          * it is most likely cache-hot in that location.
1646          */
1647         if (cpumask_test_cpu(cpu, lowest_mask))
1648                 return cpu;
1649
1650         /*
1651          * Otherwise, we consult the sched_domains span maps to figure
1652          * out which cpu is logically closest to our hot cache data.
1653          */
1654         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1655                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1656
1657         rcu_read_lock();
1658         for_each_domain(cpu, sd) {
1659                 if (sd->flags & SD_WAKE_AFFINE) {
1660                         int best_cpu;
1661
1662                         /*
1663                          * "this_cpu" is cheaper to preempt than a
1664                          * remote processor.
1665                          */
1666                         if (this_cpu != -1 &&
1667                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1668                                 rcu_read_unlock();
1669                                 return this_cpu;
1670                         }
1671
1672                         best_cpu = cpumask_first_and(lowest_mask,
1673                                                      sched_domain_span(sd));
1674                         if (best_cpu < nr_cpu_ids) {
1675                                 rcu_read_unlock();
1676                                 return best_cpu;
1677                         }
1678                 }
1679         }
1680         rcu_read_unlock();
1681
1682         /*
1683          * And finally, if there were no matches within the domains
1684          * just give the caller *something* to work with from the compatible
1685          * locations.
1686          */
1687         if (this_cpu != -1)
1688                 return this_cpu;
1689
1690         cpu = cpumask_any(lowest_mask);
1691         if (cpu < nr_cpu_ids)
1692                 return cpu;
1693         return -1;
1694 }
1695
1696 /* Will lock the rq it finds */
1697 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1698 {
1699         struct rq *lowest_rq = NULL;
1700         int tries;
1701         int cpu;
1702
1703         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1704                 cpu = find_lowest_rq(task);
1705
1706                 if ((cpu == -1) || (cpu == rq->cpu))
1707                         break;
1708
1709                 lowest_rq = cpu_rq(cpu);
1710
1711                 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1712                         /*
1713                          * Target rq has tasks of equal or higher priority,
1714                          * retrying does not release any lock and is unlikely
1715                          * to yield a different result.
1716                          */
1717                         lowest_rq = NULL;
1718                         break;
1719                 }
1720
1721                 /* if the prio of this runqueue changed, try again */
1722                 if (double_lock_balance(rq, lowest_rq)) {
1723                         /*
1724                          * We had to unlock the run queue. In
1725                          * the mean time, task could have
1726                          * migrated already or had its affinity changed.
1727                          * Also make sure that it wasn't scheduled on its rq.
1728                          */
1729                         if (unlikely(task_rq(task) != rq ||
1730                                      !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) ||
1731                                      task_running(rq, task) ||
1732                                      !rt_task(task) ||
1733                                      !task_on_rq_queued(task))) {
1734
1735                                 double_unlock_balance(rq, lowest_rq);
1736                                 lowest_rq = NULL;
1737                                 break;
1738                         }
1739                 }
1740
1741                 /* If this rq is still suitable use it. */
1742                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1743                         break;
1744
1745                 /* try again */
1746                 double_unlock_balance(rq, lowest_rq);
1747                 lowest_rq = NULL;
1748         }
1749
1750         return lowest_rq;
1751 }
1752
1753 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1754 {
1755         struct task_struct *p;
1756
1757         if (!has_pushable_tasks(rq))
1758                 return NULL;
1759
1760         p = plist_first_entry(&rq->rt.pushable_tasks,
1761                               struct task_struct, pushable_tasks);
1762
1763         BUG_ON(rq->cpu != task_cpu(p));
1764         BUG_ON(task_current(rq, p));
1765         BUG_ON(p->nr_cpus_allowed <= 1);
1766
1767         BUG_ON(!task_on_rq_queued(p));
1768         BUG_ON(!rt_task(p));
1769
1770         return p;
1771 }
1772
1773 /*
1774  * If the current CPU has more than one RT task, see if the non
1775  * running task can migrate over to a CPU that is running a task
1776  * of lesser priority.
1777  */
1778 static int push_rt_task(struct rq *rq)
1779 {
1780         struct task_struct *next_task;
1781         struct rq *lowest_rq;
1782         int ret = 0;
1783
1784         if (!rq->rt.overloaded)
1785                 return 0;
1786
1787         next_task = pick_next_pushable_task(rq);
1788         if (!next_task)
1789                 return 0;
1790
1791 retry:
1792         if (unlikely(next_task == rq->curr)) {
1793                 WARN_ON(1);
1794                 return 0;
1795         }
1796
1797         /*
1798          * It's possible that the next_task slipped in of
1799          * higher priority than current. If that's the case
1800          * just reschedule current.
1801          */
1802         if (unlikely(next_task->prio < rq->curr->prio)) {
1803                 resched_curr(rq);
1804                 return 0;
1805         }
1806
1807         /* We might release rq lock */
1808         get_task_struct(next_task);
1809
1810         /* find_lock_lowest_rq locks the rq if found */
1811         lowest_rq = find_lock_lowest_rq(next_task, rq);
1812         if (!lowest_rq) {
1813                 struct task_struct *task;
1814                 /*
1815                  * find_lock_lowest_rq releases rq->lock
1816                  * so it is possible that next_task has migrated.
1817                  *
1818                  * We need to make sure that the task is still on the same
1819                  * run-queue and is also still the next task eligible for
1820                  * pushing.
1821                  */
1822                 task = pick_next_pushable_task(rq);
1823                 if (task == next_task) {
1824                         /*
1825                          * The task hasn't migrated, and is still the next
1826                          * eligible task, but we failed to find a run-queue
1827                          * to push it to.  Do not retry in this case, since
1828                          * other cpus will pull from us when ready.
1829                          */
1830                         goto out;
1831                 }
1832
1833                 if (!task)
1834                         /* No more tasks, just exit */
1835                         goto out;
1836
1837                 /*
1838                  * Something has shifted, try again.
1839                  */
1840                 put_task_struct(next_task);
1841                 next_task = task;
1842                 goto retry;
1843         }
1844
1845         deactivate_task(rq, next_task, 0);
1846         set_task_cpu(next_task, lowest_rq->cpu);
1847         activate_task(lowest_rq, next_task, 0);
1848         ret = 1;
1849
1850         resched_curr(lowest_rq);
1851
1852         double_unlock_balance(rq, lowest_rq);
1853
1854 out:
1855         put_task_struct(next_task);
1856
1857         return ret;
1858 }
1859
1860 static void push_rt_tasks(struct rq *rq)
1861 {
1862         /* push_rt_task will return true if it moved an RT */
1863         while (push_rt_task(rq))
1864                 ;
1865 }
1866
1867 #ifdef HAVE_RT_PUSH_IPI
1868
1869 /*
1870  * When a high priority task schedules out from a CPU and a lower priority
1871  * task is scheduled in, a check is made to see if there's any RT tasks
1872  * on other CPUs that are waiting to run because a higher priority RT task
1873  * is currently running on its CPU. In this case, the CPU with multiple RT
1874  * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1875  * up that may be able to run one of its non-running queued RT tasks.
1876  *
1877  * All CPUs with overloaded RT tasks need to be notified as there is currently
1878  * no way to know which of these CPUs have the highest priority task waiting
1879  * to run. Instead of trying to take a spinlock on each of these CPUs,
1880  * which has shown to cause large latency when done on machines with many
1881  * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1882  * RT tasks waiting to run.
1883  *
1884  * Just sending an IPI to each of the CPUs is also an issue, as on large
1885  * count CPU machines, this can cause an IPI storm on a CPU, especially
1886  * if its the only CPU with multiple RT tasks queued, and a large number
1887  * of CPUs scheduling a lower priority task at the same time.
1888  *
1889  * Each root domain has its own irq work function that can iterate over
1890  * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1891  * tassk must be checked if there's one or many CPUs that are lowering
1892  * their priority, there's a single irq work iterator that will try to
1893  * push off RT tasks that are waiting to run.
1894  *
1895  * When a CPU schedules a lower priority task, it will kick off the
1896  * irq work iterator that will jump to each CPU with overloaded RT tasks.
1897  * As it only takes the first CPU that schedules a lower priority task
1898  * to start the process, the rto_start variable is incremented and if
1899  * the atomic result is one, then that CPU will try to take the rto_lock.
1900  * This prevents high contention on the lock as the process handles all
1901  * CPUs scheduling lower priority tasks.
1902  *
1903  * All CPUs that are scheduling a lower priority task will increment the
1904  * rt_loop_next variable. This will make sure that the irq work iterator
1905  * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1906  * priority task, even if the iterator is in the middle of a scan. Incrementing
1907  * the rt_loop_next will cause the iterator to perform another scan.
1908  *
1909  */
1910 static int rto_next_cpu(struct rq *rq)
1911 {
1912         struct root_domain *rd = rq->rd;
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);
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                 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
1996 }
1997
1998 /* Called from hardirq context */
1999 void rto_push_irq_work_func(struct irq_work *work)
2000 {
2001         struct rq *rq;
2002         int cpu;
2003
2004         rq = this_rq();
2005
2006         /*
2007          * We do not need to grab the lock to check for has_pushable_tasks.
2008          * When it gets updated, a check is made if a push is possible.
2009          */
2010         if (has_pushable_tasks(rq)) {
2011                 raw_spin_lock(&rq->lock);
2012                 push_rt_tasks(rq);
2013                 raw_spin_unlock(&rq->lock);
2014         }
2015
2016         raw_spin_lock(&rq->rd->rto_lock);
2017
2018         /* Pass the IPI to the next rt overloaded queue */
2019         cpu = rto_next_cpu(rq);
2020
2021         raw_spin_unlock(&rq->rd->rto_lock);
2022
2023         if (cpu < 0)
2024                 return;
2025
2026         /* Try the next RT overloaded CPU */
2027         irq_work_queue_on(&rq->rd->rto_push_work, cpu);
2028 }
2029 #endif /* HAVE_RT_PUSH_IPI */
2030
2031 static void pull_rt_task(struct rq *this_rq)
2032 {
2033         int this_cpu = this_rq->cpu, cpu;
2034         bool resched = false;
2035         struct task_struct *p;
2036         struct rq *src_rq;
2037
2038         if (likely(!rt_overloaded(this_rq)))
2039                 return;
2040
2041         /*
2042          * Match the barrier from rt_set_overloaded; this guarantees that if we
2043          * see overloaded we must also see the rto_mask bit.
2044          */
2045         smp_rmb();
2046
2047 #ifdef HAVE_RT_PUSH_IPI
2048         if (sched_feat(RT_PUSH_IPI)) {
2049                 tell_cpu_to_push(this_rq);
2050                 return;
2051         }
2052 #endif
2053
2054         for_each_cpu(cpu, this_rq->rd->rto_mask) {
2055                 if (this_cpu == cpu)
2056                         continue;
2057
2058                 src_rq = cpu_rq(cpu);
2059
2060                 /*
2061                  * Don't bother taking the src_rq->lock if the next highest
2062                  * task is known to be lower-priority than our current task.
2063                  * This may look racy, but if this value is about to go
2064                  * logically higher, the src_rq will push this task away.
2065                  * And if its going logically lower, we do not care
2066                  */
2067                 if (src_rq->rt.highest_prio.next >=
2068                     this_rq->rt.highest_prio.curr)
2069                         continue;
2070
2071                 /*
2072                  * We can potentially drop this_rq's lock in
2073                  * double_lock_balance, and another CPU could
2074                  * alter this_rq
2075                  */
2076                 double_lock_balance(this_rq, src_rq);
2077
2078                 /*
2079                  * We can pull only a task, which is pushable
2080                  * on its rq, and no others.
2081                  */
2082                 p = pick_highest_pushable_task(src_rq, this_cpu);
2083
2084                 /*
2085                  * Do we have an RT task that preempts
2086                  * the to-be-scheduled task?
2087                  */
2088                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2089                         WARN_ON(p == src_rq->curr);
2090                         WARN_ON(!task_on_rq_queued(p));
2091
2092                         /*
2093                          * There's a chance that p is higher in priority
2094                          * than what's currently running on its cpu.
2095                          * This is just that p is wakeing up and hasn't
2096                          * had a chance to schedule. We only pull
2097                          * p if it is lower in priority than the
2098                          * current task on the run queue
2099                          */
2100                         if (p->prio < src_rq->curr->prio)
2101                                 goto skip;
2102
2103                         resched = true;
2104
2105                         deactivate_task(src_rq, p, 0);
2106                         set_task_cpu(p, this_cpu);
2107                         activate_task(this_rq, p, 0);
2108                         /*
2109                          * We continue with the search, just in
2110                          * case there's an even higher prio task
2111                          * in another runqueue. (low likelihood
2112                          * but possible)
2113                          */
2114                 }
2115 skip:
2116                 double_unlock_balance(this_rq, src_rq);
2117         }
2118
2119         if (resched)
2120                 resched_curr(this_rq);
2121 }
2122
2123 /*
2124  * If we are not running and we are not going to reschedule soon, we should
2125  * try to push tasks away now
2126  */
2127 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2128 {
2129         if (!task_running(rq, p) &&
2130             !test_tsk_need_resched(rq->curr) &&
2131             p->nr_cpus_allowed > 1 &&
2132             (dl_task(rq->curr) || rt_task(rq->curr)) &&
2133             (rq->curr->nr_cpus_allowed < 2 ||
2134              rq->curr->prio <= p->prio))
2135                 push_rt_tasks(rq);
2136 }
2137
2138 /* Assumes rq->lock is held */
2139 static void rq_online_rt(struct rq *rq)
2140 {
2141         if (rq->rt.overloaded)
2142                 rt_set_overload(rq);
2143
2144         __enable_runtime(rq);
2145
2146         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2147 }
2148
2149 /* Assumes rq->lock is held */
2150 static void rq_offline_rt(struct rq *rq)
2151 {
2152         if (rq->rt.overloaded)
2153                 rt_clear_overload(rq);
2154
2155         __disable_runtime(rq);
2156
2157         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2158 }
2159
2160 /*
2161  * When switch from the rt queue, we bring ourselves to a position
2162  * that we might want to pull RT tasks from other runqueues.
2163  */
2164 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2165 {
2166         /*
2167          * If there are other RT tasks then we will reschedule
2168          * and the scheduling of the other RT tasks will handle
2169          * the balancing. But if we are the last RT task
2170          * we may need to handle the pulling of RT tasks
2171          * now.
2172          */
2173         if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2174                 return;
2175
2176         queue_pull_task(rq);
2177 }
2178
2179 void __init init_sched_rt_class(void)
2180 {
2181         unsigned int i;
2182
2183         for_each_possible_cpu(i) {
2184                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2185                                         GFP_KERNEL, cpu_to_node(i));
2186         }
2187 }
2188 #endif /* CONFIG_SMP */
2189
2190 /*
2191  * When switching a task to RT, we may overload the runqueue
2192  * with RT tasks. In this case we try to push them off to
2193  * other runqueues.
2194  */
2195 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2196 {
2197         /*
2198          * If we are already running, then there's nothing
2199          * that needs to be done. But if we are not running
2200          * we may need to preempt the current running task.
2201          * If that current running task is also an RT task
2202          * then see if we can move to another run queue.
2203          */
2204         if (task_on_rq_queued(p) && rq->curr != p) {
2205 #ifdef CONFIG_SMP
2206                 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2207                         queue_push_tasks(rq);
2208 #endif /* CONFIG_SMP */
2209                 if (p->prio < rq->curr->prio)
2210                         resched_curr(rq);
2211         }
2212 }
2213
2214 /*
2215  * Priority of the task has changed. This may cause
2216  * us to initiate a push or pull.
2217  */
2218 static void
2219 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2220 {
2221         if (!task_on_rq_queued(p))
2222                 return;
2223
2224         if (rq->curr == p) {
2225 #ifdef CONFIG_SMP
2226                 /*
2227                  * If our priority decreases while running, we
2228                  * may need to pull tasks to this runqueue.
2229                  */
2230                 if (oldprio < p->prio)
2231                         queue_pull_task(rq);
2232
2233                 /*
2234                  * If there's a higher priority task waiting to run
2235                  * then reschedule.
2236                  */
2237                 if (p->prio > rq->rt.highest_prio.curr)
2238                         resched_curr(rq);
2239 #else
2240                 /* For UP simply resched on drop of prio */
2241                 if (oldprio < p->prio)
2242                         resched_curr(rq);
2243 #endif /* CONFIG_SMP */
2244         } else {
2245                 /*
2246                  * This task is not running, but if it is
2247                  * greater than the current running task
2248                  * then reschedule.
2249                  */
2250                 if (p->prio < rq->curr->prio)
2251                         resched_curr(rq);
2252         }
2253 }
2254
2255 #ifdef CONFIG_POSIX_TIMERS
2256 static void watchdog(struct rq *rq, struct task_struct *p)
2257 {
2258         unsigned long soft, hard;
2259
2260         /* max may change after cur was read, this will be fixed next tick */
2261         soft = task_rlimit(p, RLIMIT_RTTIME);
2262         hard = task_rlimit_max(p, RLIMIT_RTTIME);
2263
2264         if (soft != RLIM_INFINITY) {
2265                 unsigned long next;
2266
2267                 if (p->rt.watchdog_stamp != jiffies) {
2268                         p->rt.timeout++;
2269                         p->rt.watchdog_stamp = jiffies;
2270                 }
2271
2272                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2273                 if (p->rt.timeout > next)
2274                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2275         }
2276 }
2277 #else
2278 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2279 #endif
2280
2281 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2282 {
2283         struct sched_rt_entity *rt_se = &p->rt;
2284
2285         update_curr_rt(rq);
2286
2287         watchdog(rq, p);
2288
2289         /*
2290          * RR tasks need a special form of timeslice management.
2291          * FIFO tasks have no timeslices.
2292          */
2293         if (p->policy != SCHED_RR)
2294                 return;
2295
2296         if (--p->rt.time_slice)
2297                 return;
2298
2299         p->rt.time_slice = sched_rr_timeslice;
2300
2301         /*
2302          * Requeue to the end of queue if we (and all of our ancestors) are not
2303          * the only element on the queue
2304          */
2305         for_each_sched_rt_entity(rt_se) {
2306                 if (rt_se->run_list.prev != rt_se->run_list.next) {
2307                         requeue_task_rt(rq, p, 0);
2308                         resched_curr(rq);
2309                         return;
2310                 }
2311         }
2312 }
2313
2314 static void set_curr_task_rt(struct rq *rq)
2315 {
2316         struct task_struct *p = rq->curr;
2317
2318         p->se.exec_start = rq_clock_task(rq);
2319
2320         /* The running task is never eligible for pushing */
2321         dequeue_pushable_task(rq, p);
2322 }
2323
2324 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2325 {
2326         /*
2327          * Time slice is 0 for SCHED_FIFO tasks
2328          */
2329         if (task->policy == SCHED_RR)
2330                 return sched_rr_timeslice;
2331         else
2332                 return 0;
2333 }
2334
2335 const struct sched_class rt_sched_class = {
2336         .next                   = &fair_sched_class,
2337         .enqueue_task           = enqueue_task_rt,
2338         .dequeue_task           = dequeue_task_rt,
2339         .yield_task             = yield_task_rt,
2340
2341         .check_preempt_curr     = check_preempt_curr_rt,
2342
2343         .pick_next_task         = pick_next_task_rt,
2344         .put_prev_task          = put_prev_task_rt,
2345
2346 #ifdef CONFIG_SMP
2347         .select_task_rq         = select_task_rq_rt,
2348
2349         .set_cpus_allowed       = set_cpus_allowed_common,
2350         .rq_online              = rq_online_rt,
2351         .rq_offline             = rq_offline_rt,
2352         .task_woken             = task_woken_rt,
2353         .switched_from          = switched_from_rt,
2354 #endif
2355
2356         .set_curr_task          = set_curr_task_rt,
2357         .task_tick              = task_tick_rt,
2358
2359         .get_rr_interval        = get_rr_interval_rt,
2360
2361         .prio_changed           = prio_changed_rt,
2362         .switched_to            = switched_to_rt,
2363
2364         .update_curr            = update_curr_rt,
2365 };
2366
2367 #ifdef CONFIG_RT_GROUP_SCHED
2368 /*
2369  * Ensure that the real time constraints are schedulable.
2370  */
2371 static DEFINE_MUTEX(rt_constraints_mutex);
2372
2373 /* Must be called with tasklist_lock held */
2374 static inline int tg_has_rt_tasks(struct task_group *tg)
2375 {
2376         struct task_struct *g, *p;
2377
2378         /*
2379          * Autogroups do not have RT tasks; see autogroup_create().
2380          */
2381         if (task_group_is_autogroup(tg))
2382                 return 0;
2383
2384         for_each_process_thread(g, p) {
2385                 if (rt_task(p) && task_group(p) == tg)
2386                         return 1;
2387         }
2388
2389         return 0;
2390 }
2391
2392 struct rt_schedulable_data {
2393         struct task_group *tg;
2394         u64 rt_period;
2395         u64 rt_runtime;
2396 };
2397
2398 static int tg_rt_schedulable(struct task_group *tg, void *data)
2399 {
2400         struct rt_schedulable_data *d = data;
2401         struct task_group *child;
2402         unsigned long total, sum = 0;
2403         u64 period, runtime;
2404
2405         period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2406         runtime = tg->rt_bandwidth.rt_runtime;
2407
2408         if (tg == d->tg) {
2409                 period = d->rt_period;
2410                 runtime = d->rt_runtime;
2411         }
2412
2413         /*
2414          * Cannot have more runtime than the period.
2415          */
2416         if (runtime > period && runtime != RUNTIME_INF)
2417                 return -EINVAL;
2418
2419         /*
2420          * Ensure we don't starve existing RT tasks.
2421          */
2422         if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
2423                 return -EBUSY;
2424
2425         total = to_ratio(period, runtime);
2426
2427         /*
2428          * Nobody can have more than the global setting allows.
2429          */
2430         if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2431                 return -EINVAL;
2432
2433         /*
2434          * The sum of our children's runtime should not exceed our own.
2435          */
2436         list_for_each_entry_rcu(child, &tg->children, siblings) {
2437                 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2438                 runtime = child->rt_bandwidth.rt_runtime;
2439
2440                 if (child == d->tg) {
2441                         period = d->rt_period;
2442                         runtime = d->rt_runtime;
2443                 }
2444
2445                 sum += to_ratio(period, runtime);
2446         }
2447
2448         if (sum > total)
2449                 return -EINVAL;
2450
2451         return 0;
2452 }
2453
2454 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2455 {
2456         int ret;
2457
2458         struct rt_schedulable_data data = {
2459                 .tg = tg,
2460                 .rt_period = period,
2461                 .rt_runtime = runtime,
2462         };
2463
2464         rcu_read_lock();
2465         ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2466         rcu_read_unlock();
2467
2468         return ret;
2469 }
2470
2471 static int tg_set_rt_bandwidth(struct task_group *tg,
2472                 u64 rt_period, u64 rt_runtime)
2473 {
2474         int i, err = 0;
2475
2476         /*
2477          * Disallowing the root group RT runtime is BAD, it would disallow the
2478          * kernel creating (and or operating) RT threads.
2479          */
2480         if (tg == &root_task_group && rt_runtime == 0)
2481                 return -EINVAL;
2482
2483         /* No period doesn't make any sense. */
2484         if (rt_period == 0)
2485                 return -EINVAL;
2486
2487         mutex_lock(&rt_constraints_mutex);
2488         read_lock(&tasklist_lock);
2489         err = __rt_schedulable(tg, rt_period, rt_runtime);
2490         if (err)
2491                 goto unlock;
2492
2493         raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2494         tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2495         tg->rt_bandwidth.rt_runtime = rt_runtime;
2496
2497         for_each_possible_cpu(i) {
2498                 struct rt_rq *rt_rq = tg->rt_rq[i];
2499
2500                 raw_spin_lock(&rt_rq->rt_runtime_lock);
2501                 rt_rq->rt_runtime = rt_runtime;
2502                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2503         }
2504         raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2505 unlock:
2506         read_unlock(&tasklist_lock);
2507         mutex_unlock(&rt_constraints_mutex);
2508
2509         return err;
2510 }
2511
2512 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2513 {
2514         u64 rt_runtime, rt_period;
2515
2516         rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2517         rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2518         if (rt_runtime_us < 0)
2519                 rt_runtime = RUNTIME_INF;
2520
2521         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2522 }
2523
2524 long sched_group_rt_runtime(struct task_group *tg)
2525 {
2526         u64 rt_runtime_us;
2527
2528         if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2529                 return -1;
2530
2531         rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2532         do_div(rt_runtime_us, NSEC_PER_USEC);
2533         return rt_runtime_us;
2534 }
2535
2536 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2537 {
2538         u64 rt_runtime, rt_period;
2539
2540         rt_period = rt_period_us * NSEC_PER_USEC;
2541         rt_runtime = tg->rt_bandwidth.rt_runtime;
2542
2543         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2544 }
2545
2546 long sched_group_rt_period(struct task_group *tg)
2547 {
2548         u64 rt_period_us;
2549
2550         rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2551         do_div(rt_period_us, NSEC_PER_USEC);
2552         return rt_period_us;
2553 }
2554
2555 static int sched_rt_global_constraints(void)
2556 {
2557         int ret = 0;
2558
2559         mutex_lock(&rt_constraints_mutex);
2560         read_lock(&tasklist_lock);
2561         ret = __rt_schedulable(NULL, 0, 0);
2562         read_unlock(&tasklist_lock);
2563         mutex_unlock(&rt_constraints_mutex);
2564
2565         return ret;
2566 }
2567
2568 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2569 {
2570         /* Don't accept realtime tasks when there is no way for them to run */
2571         if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2572                 return 0;
2573
2574         return 1;
2575 }
2576
2577 #else /* !CONFIG_RT_GROUP_SCHED */
2578 static int sched_rt_global_constraints(void)
2579 {
2580         unsigned long flags;
2581         int i;
2582
2583         raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2584         for_each_possible_cpu(i) {
2585                 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2586
2587                 raw_spin_lock(&rt_rq->rt_runtime_lock);
2588                 rt_rq->rt_runtime = global_rt_runtime();
2589                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2590         }
2591         raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2592
2593         return 0;
2594 }
2595 #endif /* CONFIG_RT_GROUP_SCHED */
2596
2597 static int sched_rt_global_validate(void)
2598 {
2599         if (sysctl_sched_rt_period <= 0)
2600                 return -EINVAL;
2601
2602         if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2603                 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
2604                 return -EINVAL;
2605
2606         return 0;
2607 }
2608
2609 static void sched_rt_do_global(void)
2610 {
2611         def_rt_bandwidth.rt_runtime = global_rt_runtime();
2612         def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2613 }
2614
2615 int sched_rt_handler(struct ctl_table *table, int write,
2616                 void __user *buffer, size_t *lenp,
2617                 loff_t *ppos)
2618 {
2619         int old_period, old_runtime;
2620         static DEFINE_MUTEX(mutex);
2621         int ret;
2622
2623         mutex_lock(&mutex);
2624         old_period = sysctl_sched_rt_period;
2625         old_runtime = sysctl_sched_rt_runtime;
2626
2627         ret = proc_dointvec(table, write, buffer, lenp, ppos);
2628
2629         if (!ret && write) {
2630                 ret = sched_rt_global_validate();
2631                 if (ret)
2632                         goto undo;
2633
2634                 ret = sched_dl_global_validate();
2635                 if (ret)
2636                         goto undo;
2637
2638                 ret = sched_rt_global_constraints();
2639                 if (ret)
2640                         goto undo;
2641
2642                 sched_rt_do_global();
2643                 sched_dl_do_global();
2644         }
2645         if (0) {
2646 undo:
2647                 sysctl_sched_rt_period = old_period;
2648                 sysctl_sched_rt_runtime = old_runtime;
2649         }
2650         mutex_unlock(&mutex);
2651
2652         return ret;
2653 }
2654
2655 int sched_rr_handler(struct ctl_table *table, int write,
2656                 void __user *buffer, size_t *lenp,
2657                 loff_t *ppos)
2658 {
2659         int ret;
2660         static DEFINE_MUTEX(mutex);
2661
2662         mutex_lock(&mutex);
2663         ret = proc_dointvec(table, write, buffer, lenp, ppos);
2664         /*
2665          * Make sure that internally we keep jiffies.
2666          * Also, writing zero resets the timeslice to default:
2667          */
2668         if (!ret && write) {
2669                 sched_rr_timeslice =
2670                         sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2671                         msecs_to_jiffies(sysctl_sched_rr_timeslice);
2672         }
2673         mutex_unlock(&mutex);
2674         return ret;
2675 }
2676
2677 #ifdef CONFIG_SCHED_DEBUG
2678 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2679
2680 void print_rt_stats(struct seq_file *m, int cpu)
2681 {
2682         rt_rq_iter_t iter;
2683         struct rt_rq *rt_rq;
2684
2685         rcu_read_lock();
2686         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2687                 print_rt_rq(m, cpu, rt_rq);
2688         rcu_read_unlock();
2689 }
2690 #endif /* CONFIG_SCHED_DEBUG */