1c4065575fa2c44d91250455db38638d728788e8
[projects/modsched/linux.git] / kernel / sched / cfs / rt.c
1 /*
2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3  * policies)
4  */
5
6 #include "sched.h"
7
8 #include <linux/slab.h>
9
10 int sched_rr_timeslice = RR_TIMESLICE;
11
12 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
13
14 struct rt_bandwidth def_rt_bandwidth;
15
16 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
17 {
18         struct rt_bandwidth *rt_b =
19                 container_of(timer, struct rt_bandwidth, rt_period_timer);
20         ktime_t now;
21         int overrun;
22         int idle = 0;
23
24         for (;;) {
25                 now = hrtimer_cb_get_time(timer);
26                 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
27
28                 if (!overrun)
29                         break;
30
31                 idle = do_sched_rt_period_timer(rt_b, overrun);
32         }
33
34         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
35 }
36
37 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
38 {
39         rt_b->rt_period = ns_to_ktime(period);
40         rt_b->rt_runtime = runtime;
41
42         raw_spin_lock_init(&rt_b->rt_runtime_lock);
43
44         hrtimer_init(&rt_b->rt_period_timer,
45                         CLOCK_MONOTONIC, HRTIMER_MODE_REL);
46         rt_b->rt_period_timer.function = sched_rt_period_timer;
47 }
48
49 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
50 {
51         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
52                 return;
53
54         if (hrtimer_active(&rt_b->rt_period_timer))
55                 return;
56
57         raw_spin_lock(&rt_b->rt_runtime_lock);
58         start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
59         raw_spin_unlock(&rt_b->rt_runtime_lock);
60 }
61
62 void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
63 {
64         struct rt_prio_array *array;
65         int i;
66
67         array = &rt_rq->active;
68         for (i = 0; i < MAX_RT_PRIO; i++) {
69                 INIT_LIST_HEAD(array->queue + i);
70                 __clear_bit(i, array->bitmap);
71         }
72         /* delimiter for bitsearch: */
73         __set_bit(MAX_RT_PRIO, array->bitmap);
74
75 #if defined CONFIG_SMP
76         rt_rq->highest_prio.curr = MAX_RT_PRIO;
77         rt_rq->highest_prio.next = MAX_RT_PRIO;
78         rt_rq->rt_nr_migratory = 0;
79         rt_rq->overloaded = 0;
80         plist_head_init(&rt_rq->pushable_tasks);
81 #endif
82
83         rt_rq->rt_time = 0;
84         rt_rq->rt_throttled = 0;
85         rt_rq->rt_runtime = 0;
86         raw_spin_lock_init(&rt_rq->rt_runtime_lock);
87 }
88
89 #ifdef CONFIG_RT_GROUP_SCHED
90 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
91 {
92         hrtimer_cancel(&rt_b->rt_period_timer);
93 }
94
95 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
96
97 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
98 {
99 #ifdef CONFIG_SCHED_DEBUG
100         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
101 #endif
102         return container_of(rt_se, struct task_struct, rt);
103 }
104
105 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
106 {
107         return rt_rq->rq;
108 }
109
110 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
111 {
112         return rt_se->rt_rq;
113 }
114
115 void free_rt_sched_group(struct task_group *tg)
116 {
117         int i;
118
119         if (tg->rt_se)
120                 destroy_rt_bandwidth(&tg->rt_bandwidth);
121
122         for_each_possible_cpu(i) {
123                 if (tg->rt_rq)
124                         kfree(tg->rt_rq[i]);
125                 if (tg->rt_se)
126                         kfree(tg->rt_se[i]);
127         }
128
129         kfree(tg->rt_rq);
130         kfree(tg->rt_se);
131 }
132
133 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
134                 struct sched_rt_entity *rt_se, int cpu,
135                 struct sched_rt_entity *parent)
136 {
137         struct rq *rq = cpu_rq(cpu);
138
139         rt_rq->highest_prio.curr = MAX_RT_PRIO;
140         rt_rq->rt_nr_boosted = 0;
141         rt_rq->rq = rq;
142         rt_rq->tg = tg;
143
144         tg->rt_rq[cpu] = rt_rq;
145         tg->rt_se[cpu] = rt_se;
146
147         if (!rt_se)
148                 return;
149
150         if (!parent)
151                 rt_se->rt_rq = &rq->rt;
152         else
153                 rt_se->rt_rq = parent->my_q;
154
155         rt_se->my_q = rt_rq;
156         rt_se->parent = parent;
157         INIT_LIST_HEAD(&rt_se->run_list);
158 }
159
160 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
161 {
162         struct rt_rq *rt_rq;
163         struct sched_rt_entity *rt_se;
164         int i;
165
166         tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
167         if (!tg->rt_rq)
168                 goto err;
169         tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
170         if (!tg->rt_se)
171                 goto err;
172
173         init_rt_bandwidth(&tg->rt_bandwidth,
174                         ktime_to_ns(def_rt_bandwidth.rt_period), 0);
175
176         for_each_possible_cpu(i) {
177                 rt_rq = kzalloc_node(sizeof(struct rt_rq),
178                                      GFP_KERNEL, cpu_to_node(i));
179                 if (!rt_rq)
180                         goto err;
181
182                 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
183                                      GFP_KERNEL, cpu_to_node(i));
184                 if (!rt_se)
185                         goto err_free_rq;
186
187                 init_rt_rq(rt_rq, cpu_rq(i));
188                 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
189                 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
190         }
191
192         return 1;
193
194 err_free_rq:
195         kfree(rt_rq);
196 err:
197         return 0;
198 }
199
200 #else /* CONFIG_RT_GROUP_SCHED */
201
202 #define rt_entity_is_task(rt_se) (1)
203
204 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
205 {
206         return container_of(rt_se, struct task_struct, rt);
207 }
208
209 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
210 {
211         return container_of(rt_rq, struct rq, rt);
212 }
213
214 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
215 {
216         struct task_struct *p = rt_task_of(rt_se);
217         struct rq *rq = task_rq(p);
218
219         return &rq->rt;
220 }
221
222 void free_rt_sched_group(struct task_group *tg) { }
223
224 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
225 {
226         return 1;
227 }
228 #endif /* CONFIG_RT_GROUP_SCHED */
229
230 #ifdef CONFIG_SMP
231
232 static inline int rt_overloaded(struct rq *rq)
233 {
234         return atomic_read(&rq->rd->rto_count);
235 }
236
237 static inline void rt_set_overload(struct rq *rq)
238 {
239         if (!rq->online)
240                 return;
241
242         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
243         /*
244          * Make sure the mask is visible before we set
245          * the overload count. That is checked to determine
246          * if we should look at the mask. It would be a shame
247          * if we looked at the mask, but the mask was not
248          * updated yet.
249          *
250          * Matched by the barrier in pull_rt_task().
251          */
252         smp_wmb();
253         atomic_inc(&rq->rd->rto_count);
254 }
255
256 static inline void rt_clear_overload(struct rq *rq)
257 {
258         if (!rq->online)
259                 return;
260
261         /* the order here really doesn't matter */
262         atomic_dec(&rq->rd->rto_count);
263         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
264 }
265
266 static void update_rt_migration(struct rt_rq *rt_rq)
267 {
268         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
269                 if (!rt_rq->overloaded) {
270                         rt_set_overload(rq_of_rt_rq(rt_rq));
271                         rt_rq->overloaded = 1;
272                 }
273         } else if (rt_rq->overloaded) {
274                 rt_clear_overload(rq_of_rt_rq(rt_rq));
275                 rt_rq->overloaded = 0;
276         }
277 }
278
279 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
280 {
281         struct task_struct *p;
282
283         if (!rt_entity_is_task(rt_se))
284                 return;
285
286         p = rt_task_of(rt_se);
287         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
288
289         rt_rq->rt_nr_total++;
290         if (p->nr_cpus_allowed > 1)
291                 rt_rq->rt_nr_migratory++;
292
293         update_rt_migration(rt_rq);
294 }
295
296 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
297 {
298         struct task_struct *p;
299
300         if (!rt_entity_is_task(rt_se))
301                 return;
302
303         p = rt_task_of(rt_se);
304         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
305
306         rt_rq->rt_nr_total--;
307         if (p->nr_cpus_allowed > 1)
308                 rt_rq->rt_nr_migratory--;
309
310         update_rt_migration(rt_rq);
311 }
312
313 static inline int has_pushable_tasks(struct rq *rq)
314 {
315         return !plist_head_empty(&rq->rt.pushable_tasks);
316 }
317
318 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
319 {
320         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
321         plist_node_init(&p->pushable_tasks, p->prio);
322         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
323
324         /* Update the highest prio pushable task */
325         if (p->prio < rq->rt.highest_prio.next)
326                 rq->rt.highest_prio.next = p->prio;
327 }
328
329 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
330 {
331         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
332
333         /* Update the new highest prio pushable task */
334         if (has_pushable_tasks(rq)) {
335                 p = plist_first_entry(&rq->rt.pushable_tasks,
336                                       struct task_struct, pushable_tasks);
337                 rq->rt.highest_prio.next = p->prio;
338         } else
339                 rq->rt.highest_prio.next = MAX_RT_PRIO;
340 }
341
342 #else
343
344 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
345 {
346 }
347
348 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
349 {
350 }
351
352 static inline
353 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
354 {
355 }
356
357 static inline
358 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
359 {
360 }
361
362 #endif /* CONFIG_SMP */
363
364 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
365 {
366         return !list_empty(&rt_se->run_list);
367 }
368
369 #ifdef CONFIG_RT_GROUP_SCHED
370
371 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
372 {
373         if (!rt_rq->tg)
374                 return RUNTIME_INF;
375
376         return rt_rq->rt_runtime;
377 }
378
379 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
380 {
381         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
382 }
383
384 typedef struct task_group *rt_rq_iter_t;
385
386 static inline struct task_group *next_task_group(struct task_group *tg)
387 {
388         do {
389                 tg = list_entry_rcu(tg->list.next,
390                         typeof(struct task_group), list);
391         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
392
393         if (&tg->list == &task_groups)
394                 tg = NULL;
395
396         return tg;
397 }
398
399 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
400         for (iter = container_of(&task_groups, typeof(*iter), list);    \
401                 (iter = next_task_group(iter)) &&                       \
402                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
403
404 #define for_each_sched_rt_entity(rt_se) \
405         for (; rt_se; rt_se = rt_se->parent)
406
407 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
408 {
409         return rt_se->my_q;
410 }
411
412 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
413 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
414
415 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
416 {
417         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
418         struct sched_rt_entity *rt_se;
419
420         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
421
422         rt_se = rt_rq->tg->rt_se[cpu];
423
424         if (rt_rq->rt_nr_running) {
425                 if (rt_se && !on_rt_rq(rt_se))
426                         enqueue_rt_entity(rt_se, false);
427                 if (rt_rq->highest_prio.curr < curr->prio)
428                         resched_task(curr);
429         }
430 }
431
432 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
433 {
434         struct sched_rt_entity *rt_se;
435         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
436
437         rt_se = rt_rq->tg->rt_se[cpu];
438
439         if (rt_se && on_rt_rq(rt_se))
440                 dequeue_rt_entity(rt_se);
441 }
442
443 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
444 {
445         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
446 }
447
448 static int rt_se_boosted(struct sched_rt_entity *rt_se)
449 {
450         struct rt_rq *rt_rq = group_rt_rq(rt_se);
451         struct task_struct *p;
452
453         if (rt_rq)
454                 return !!rt_rq->rt_nr_boosted;
455
456         p = rt_task_of(rt_se);
457         return p->prio != p->normal_prio;
458 }
459
460 #ifdef CONFIG_SMP
461 static inline const struct cpumask *sched_rt_period_mask(void)
462 {
463         return this_rq()->rd->span;
464 }
465 #else
466 static inline const struct cpumask *sched_rt_period_mask(void)
467 {
468         return cpu_online_mask;
469 }
470 #endif
471
472 static inline
473 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
474 {
475         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
476 }
477
478 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
479 {
480         return &rt_rq->tg->rt_bandwidth;
481 }
482
483 #else /* !CONFIG_RT_GROUP_SCHED */
484
485 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
486 {
487         return rt_rq->rt_runtime;
488 }
489
490 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
491 {
492         return ktime_to_ns(def_rt_bandwidth.rt_period);
493 }
494
495 typedef struct rt_rq *rt_rq_iter_t;
496
497 #define for_each_rt_rq(rt_rq, iter, rq) \
498         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
499
500 #define for_each_sched_rt_entity(rt_se) \
501         for (; rt_se; rt_se = NULL)
502
503 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
504 {
505         return NULL;
506 }
507
508 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
509 {
510         if (rt_rq->rt_nr_running)
511                 resched_task(rq_of_rt_rq(rt_rq)->curr);
512 }
513
514 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
515 {
516 }
517
518 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
519 {
520         return rt_rq->rt_throttled;
521 }
522
523 static inline const struct cpumask *sched_rt_period_mask(void)
524 {
525         return cpu_online_mask;
526 }
527
528 static inline
529 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
530 {
531         return &cpu_rq(cpu)->rt;
532 }
533
534 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
535 {
536         return &def_rt_bandwidth;
537 }
538
539 #endif /* CONFIG_RT_GROUP_SCHED */
540
541 #ifdef CONFIG_SMP
542 /*
543  * We ran out of runtime, see if we can borrow some from our neighbours.
544  */
545 static int do_balance_runtime(struct rt_rq *rt_rq)
546 {
547         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
548         struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
549         int i, weight, more = 0;
550         u64 rt_period;
551
552         weight = cpumask_weight(rd->span);
553
554         raw_spin_lock(&rt_b->rt_runtime_lock);
555         rt_period = ktime_to_ns(rt_b->rt_period);
556         for_each_cpu(i, rd->span) {
557                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
558                 s64 diff;
559
560                 if (iter == rt_rq)
561                         continue;
562
563                 raw_spin_lock(&iter->rt_runtime_lock);
564                 /*
565                  * Either all rqs have inf runtime and there's nothing to steal
566                  * or __disable_runtime() below sets a specific rq to inf to
567                  * indicate its been disabled and disalow stealing.
568                  */
569                 if (iter->rt_runtime == RUNTIME_INF)
570                         goto next;
571
572                 /*
573                  * From runqueues with spare time, take 1/n part of their
574                  * spare time, but no more than our period.
575                  */
576                 diff = iter->rt_runtime - iter->rt_time;
577                 if (diff > 0) {
578                         diff = div_u64((u64)diff, weight);
579                         if (rt_rq->rt_runtime + diff > rt_period)
580                                 diff = rt_period - rt_rq->rt_runtime;
581                         iter->rt_runtime -= diff;
582                         rt_rq->rt_runtime += diff;
583                         more = 1;
584                         if (rt_rq->rt_runtime == rt_period) {
585                                 raw_spin_unlock(&iter->rt_runtime_lock);
586                                 break;
587                         }
588                 }
589 next:
590                 raw_spin_unlock(&iter->rt_runtime_lock);
591         }
592         raw_spin_unlock(&rt_b->rt_runtime_lock);
593
594         return more;
595 }
596
597 /*
598  * Ensure this RQ takes back all the runtime it lend to its neighbours.
599  */
600 static void __disable_runtime(struct rq *rq)
601 {
602         struct root_domain *rd = rq->rd;
603         rt_rq_iter_t iter;
604         struct rt_rq *rt_rq;
605
606         if (unlikely(!scheduler_running))
607                 return;
608
609         for_each_rt_rq(rt_rq, iter, rq) {
610                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
611                 s64 want;
612                 int i;
613
614                 raw_spin_lock(&rt_b->rt_runtime_lock);
615                 raw_spin_lock(&rt_rq->rt_runtime_lock);
616                 /*
617                  * Either we're all inf and nobody needs to borrow, or we're
618                  * already disabled and thus have nothing to do, or we have
619                  * exactly the right amount of runtime to take out.
620                  */
621                 if (rt_rq->rt_runtime == RUNTIME_INF ||
622                                 rt_rq->rt_runtime == rt_b->rt_runtime)
623                         goto balanced;
624                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
625
626                 /*
627                  * Calculate the difference between what we started out with
628                  * and what we current have, that's the amount of runtime
629                  * we lend and now have to reclaim.
630                  */
631                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
632
633                 /*
634                  * Greedy reclaim, take back as much as we can.
635                  */
636                 for_each_cpu(i, rd->span) {
637                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
638                         s64 diff;
639
640                         /*
641                          * Can't reclaim from ourselves or disabled runqueues.
642                          */
643                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
644                                 continue;
645
646                         raw_spin_lock(&iter->rt_runtime_lock);
647                         if (want > 0) {
648                                 diff = min_t(s64, iter->rt_runtime, want);
649                                 iter->rt_runtime -= diff;
650                                 want -= diff;
651                         } else {
652                                 iter->rt_runtime -= want;
653                                 want -= want;
654                         }
655                         raw_spin_unlock(&iter->rt_runtime_lock);
656
657                         if (!want)
658                                 break;
659                 }
660
661                 raw_spin_lock(&rt_rq->rt_runtime_lock);
662                 /*
663                  * We cannot be left wanting - that would mean some runtime
664                  * leaked out of the system.
665                  */
666                 BUG_ON(want);
667 balanced:
668                 /*
669                  * Disable all the borrow logic by pretending we have inf
670                  * runtime - in which case borrowing doesn't make sense.
671                  */
672                 rt_rq->rt_runtime = RUNTIME_INF;
673                 rt_rq->rt_throttled = 0;
674                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
675                 raw_spin_unlock(&rt_b->rt_runtime_lock);
676         }
677 }
678
679 static void __enable_runtime(struct rq *rq)
680 {
681         rt_rq_iter_t iter;
682         struct rt_rq *rt_rq;
683
684         if (unlikely(!scheduler_running))
685                 return;
686
687         /*
688          * Reset each runqueue's bandwidth settings
689          */
690         for_each_rt_rq(rt_rq, iter, rq) {
691                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
692
693                 raw_spin_lock(&rt_b->rt_runtime_lock);
694                 raw_spin_lock(&rt_rq->rt_runtime_lock);
695                 rt_rq->rt_runtime = rt_b->rt_runtime;
696                 rt_rq->rt_time = 0;
697                 rt_rq->rt_throttled = 0;
698                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
699                 raw_spin_unlock(&rt_b->rt_runtime_lock);
700         }
701 }
702
703 static int balance_runtime(struct rt_rq *rt_rq)
704 {
705         int more = 0;
706
707         if (!sched_feat(RT_RUNTIME_SHARE))
708                 return more;
709
710         if (rt_rq->rt_time > rt_rq->rt_runtime) {
711                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
712                 more = do_balance_runtime(rt_rq);
713                 raw_spin_lock(&rt_rq->rt_runtime_lock);
714         }
715
716         return more;
717 }
718 #else /* !CONFIG_SMP */
719 static inline int balance_runtime(struct rt_rq *rt_rq)
720 {
721         return 0;
722 }
723 #endif /* CONFIG_SMP */
724
725 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
726 {
727         int i, idle = 1, throttled = 0;
728         const struct cpumask *span;
729
730         span = sched_rt_period_mask();
731 #ifdef CONFIG_RT_GROUP_SCHED
732         /*
733          * FIXME: isolated CPUs should really leave the root task group,
734          * whether they are isolcpus or were isolated via cpusets, lest
735          * the timer run on a CPU which does not service all runqueues,
736          * potentially leaving other CPUs indefinitely throttled.  If
737          * isolation is really required, the user will turn the throttle
738          * off to kill the perturbations it causes anyway.  Meanwhile,
739          * this maintains functionality for boot and/or troubleshooting.
740          */
741         if (rt_b == &root_task_group.rt_bandwidth)
742                 span = cpu_online_mask;
743 #endif
744         for_each_cpu(i, span) {
745                 int enqueue = 0;
746                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
747                 struct rq *rq = rq_of_rt_rq(rt_rq);
748
749                 raw_spin_lock(&rq->lock);
750                 if (rt_rq->rt_time) {
751                         u64 runtime;
752
753                         raw_spin_lock(&rt_rq->rt_runtime_lock);
754                         if (rt_rq->rt_throttled)
755                                 balance_runtime(rt_rq);
756                         runtime = rt_rq->rt_runtime;
757                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
758                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
759                                 rt_rq->rt_throttled = 0;
760                                 enqueue = 1;
761
762                                 /*
763                                  * Force a clock update if the CPU was idle,
764                                  * lest wakeup -> unthrottle time accumulate.
765                                  */
766                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
767                                         rq->skip_clock_update = -1;
768                         }
769                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
770                                 idle = 0;
771                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
772                 } else if (rt_rq->rt_nr_running) {
773                         idle = 0;
774                         if (!rt_rq_throttled(rt_rq))
775                                 enqueue = 1;
776                 }
777                 if (rt_rq->rt_throttled)
778                         throttled = 1;
779
780                 if (enqueue)
781                         sched_rt_rq_enqueue(rt_rq);
782                 raw_spin_unlock(&rq->lock);
783         }
784
785         if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
786                 return 1;
787
788         return idle;
789 }
790
791 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
792 {
793 #ifdef CONFIG_RT_GROUP_SCHED
794         struct rt_rq *rt_rq = group_rt_rq(rt_se);
795
796         if (rt_rq)
797                 return rt_rq->highest_prio.curr;
798 #endif
799
800         return rt_task_of(rt_se)->prio;
801 }
802
803 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
804 {
805         u64 runtime = sched_rt_runtime(rt_rq);
806
807         if (rt_rq->rt_throttled)
808                 return rt_rq_throttled(rt_rq);
809
810         if (runtime >= sched_rt_period(rt_rq))
811                 return 0;
812
813         balance_runtime(rt_rq);
814         runtime = sched_rt_runtime(rt_rq);
815         if (runtime == RUNTIME_INF)
816                 return 0;
817
818         if (rt_rq->rt_time > runtime) {
819                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
820
821                 /*
822                  * Don't actually throttle groups that have no runtime assigned
823                  * but accrue some time due to boosting.
824                  */
825                 if (likely(rt_b->rt_runtime)) {
826                         static bool once = false;
827
828                         rt_rq->rt_throttled = 1;
829
830                         if (!once) {
831                                 once = true;
832                                 printk_sched("sched: RT throttling activated\n");
833                         }
834                 } else {
835                         /*
836                          * In case we did anyway, make it go away,
837                          * replenishment is a joke, since it will replenish us
838                          * with exactly 0 ns.
839                          */
840                         rt_rq->rt_time = 0;
841                 }
842
843                 if (rt_rq_throttled(rt_rq)) {
844                         sched_rt_rq_dequeue(rt_rq);
845                         return 1;
846                 }
847         }
848
849         return 0;
850 }
851
852 /*
853  * Update the current task's runtime statistics. Skip current tasks that
854  * are not in our scheduling class.
855  */
856 static void update_curr_rt(struct rq *rq)
857 {
858         struct task_struct *curr = rq->curr;
859         struct sched_rt_entity *rt_se = &curr->rt;
860         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
861         u64 delta_exec;
862
863         if (curr->sched_class != &rt_sched_class)
864                 return;
865
866         delta_exec = rq_clock_task(rq) - curr->se.exec_start;
867         if (unlikely((s64)delta_exec <= 0))
868                 return;
869
870         schedstat_set(curr->se.statistics.exec_max,
871                       max(curr->se.statistics.exec_max, delta_exec));
872
873         curr->se.sum_exec_runtime += delta_exec;
874         account_group_exec_runtime(curr, delta_exec);
875
876         curr->se.exec_start = rq_clock_task(rq);
877         cpuacct_charge(curr, delta_exec);
878
879         sched_rt_avg_update(rq, delta_exec);
880
881         if (!rt_bandwidth_enabled())
882                 return;
883
884         for_each_sched_rt_entity(rt_se) {
885                 rt_rq = rt_rq_of_se(rt_se);
886
887                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
888                         raw_spin_lock(&rt_rq->rt_runtime_lock);
889                         rt_rq->rt_time += delta_exec;
890                         if (sched_rt_runtime_exceeded(rt_rq))
891                                 resched_task(curr);
892                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
893                 }
894         }
895 }
896
897 #if defined CONFIG_SMP
898
899 static void
900 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
901 {
902         struct rq *rq = rq_of_rt_rq(rt_rq);
903
904 #ifdef CONFIG_RT_GROUP_SCHED
905         /*
906          * Change rq's cpupri only if rt_rq is the top queue.
907          */
908         if (&rq->rt != rt_rq)
909                 return;
910 #endif
911         if (rq->online && prio < prev_prio)
912                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
913 }
914
915 static void
916 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
917 {
918         struct rq *rq = rq_of_rt_rq(rt_rq);
919
920 #ifdef CONFIG_RT_GROUP_SCHED
921         /*
922          * Change rq's cpupri only if rt_rq is the top queue.
923          */
924         if (&rq->rt != rt_rq)
925                 return;
926 #endif
927         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
928                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
929 }
930
931 #else /* CONFIG_SMP */
932
933 static inline
934 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
935 static inline
936 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
937
938 #endif /* CONFIG_SMP */
939
940 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
941 static void
942 inc_rt_prio(struct rt_rq *rt_rq, int prio)
943 {
944         int prev_prio = rt_rq->highest_prio.curr;
945
946         if (prio < prev_prio)
947                 rt_rq->highest_prio.curr = prio;
948
949         inc_rt_prio_smp(rt_rq, prio, prev_prio);
950 }
951
952 static void
953 dec_rt_prio(struct rt_rq *rt_rq, int prio)
954 {
955         int prev_prio = rt_rq->highest_prio.curr;
956
957         if (rt_rq->rt_nr_running) {
958
959                 WARN_ON(prio < prev_prio);
960
961                 /*
962                  * This may have been our highest task, and therefore
963                  * we may have some recomputation to do
964                  */
965                 if (prio == prev_prio) {
966                         struct rt_prio_array *array = &rt_rq->active;
967
968                         rt_rq->highest_prio.curr =
969                                 sched_find_first_bit(array->bitmap);
970                 }
971
972         } else
973                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
974
975         dec_rt_prio_smp(rt_rq, prio, prev_prio);
976 }
977
978 #else
979
980 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
981 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
982
983 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
984
985 #ifdef CONFIG_RT_GROUP_SCHED
986
987 static void
988 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
989 {
990         if (rt_se_boosted(rt_se))
991                 rt_rq->rt_nr_boosted++;
992
993         if (rt_rq->tg)
994                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
995 }
996
997 static void
998 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
999 {
1000         if (rt_se_boosted(rt_se))
1001                 rt_rq->rt_nr_boosted--;
1002
1003         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1004 }
1005
1006 #else /* CONFIG_RT_GROUP_SCHED */
1007
1008 static void
1009 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1010 {
1011         start_rt_bandwidth(&def_rt_bandwidth);
1012 }
1013
1014 static inline
1015 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1016
1017 #endif /* CONFIG_RT_GROUP_SCHED */
1018
1019 static inline
1020 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1021 {
1022         int prio = rt_se_prio(rt_se);
1023
1024         WARN_ON(!rt_prio(prio));
1025         rt_rq->rt_nr_running++;
1026
1027         inc_rt_prio(rt_rq, prio);
1028         inc_rt_migration(rt_se, rt_rq);
1029         inc_rt_group(rt_se, rt_rq);
1030 }
1031
1032 static inline
1033 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1034 {
1035         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1036         WARN_ON(!rt_rq->rt_nr_running);
1037         rt_rq->rt_nr_running--;
1038
1039         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1040         dec_rt_migration(rt_se, rt_rq);
1041         dec_rt_group(rt_se, rt_rq);
1042 }
1043
1044 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1045 {
1046         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1047         struct rt_prio_array *array = &rt_rq->active;
1048         struct rt_rq *group_rq = group_rt_rq(rt_se);
1049         struct list_head *queue = array->queue + rt_se_prio(rt_se);
1050
1051         /*
1052          * Don't enqueue the group if its throttled, or when empty.
1053          * The latter is a consequence of the former when a child group
1054          * get throttled and the current group doesn't have any other
1055          * active members.
1056          */
1057         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
1058                 return;
1059
1060         if (head)
1061                 list_add(&rt_se->run_list, queue);
1062         else
1063                 list_add_tail(&rt_se->run_list, queue);
1064         __set_bit(rt_se_prio(rt_se), array->bitmap);
1065
1066         inc_rt_tasks(rt_se, rt_rq);
1067 }
1068
1069 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
1070 {
1071         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1072         struct rt_prio_array *array = &rt_rq->active;
1073
1074         list_del_init(&rt_se->run_list);
1075         if (list_empty(array->queue + rt_se_prio(rt_se)))
1076                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1077
1078         dec_rt_tasks(rt_se, rt_rq);
1079 }
1080
1081 /*
1082  * Because the prio of an upper entry depends on the lower
1083  * entries, we must remove entries top - down.
1084  */
1085 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
1086 {
1087         struct sched_rt_entity *back = NULL;
1088
1089         for_each_sched_rt_entity(rt_se) {
1090                 rt_se->back = back;
1091                 back = rt_se;
1092         }
1093
1094         for (rt_se = back; rt_se; rt_se = rt_se->back) {
1095                 if (on_rt_rq(rt_se))
1096                         __dequeue_rt_entity(rt_se);
1097         }
1098 }
1099
1100 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1101 {
1102         dequeue_rt_stack(rt_se);
1103         for_each_sched_rt_entity(rt_se)
1104                 __enqueue_rt_entity(rt_se, head);
1105 }
1106
1107 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1108 {
1109         dequeue_rt_stack(rt_se);
1110
1111         for_each_sched_rt_entity(rt_se) {
1112                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1113
1114                 if (rt_rq && rt_rq->rt_nr_running)
1115                         __enqueue_rt_entity(rt_se, false);
1116         }
1117 }
1118
1119 /*
1120  * Adding/removing a task to/from a priority array:
1121  */
1122 static void
1123 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1124 {
1125         struct sched_rt_entity *rt_se = &p->rt;
1126
1127         if (flags & ENQUEUE_WAKEUP)
1128                 rt_se->timeout = 0;
1129
1130         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
1131
1132         if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1133                 enqueue_pushable_task(rq, p);
1134
1135         inc_nr_running(rq);
1136 }
1137
1138 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1139 {
1140         struct sched_rt_entity *rt_se = &p->rt;
1141
1142         update_curr_rt(rq);
1143         dequeue_rt_entity(rt_se);
1144
1145         dequeue_pushable_task(rq, p);
1146
1147         dec_nr_running(rq);
1148 }
1149
1150 /*
1151  * Put task to the head or the end of the run list without the overhead of
1152  * dequeue followed by enqueue.
1153  */
1154 static void
1155 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1156 {
1157         if (on_rt_rq(rt_se)) {
1158                 struct rt_prio_array *array = &rt_rq->active;
1159                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1160
1161                 if (head)
1162                         list_move(&rt_se->run_list, queue);
1163                 else
1164                         list_move_tail(&rt_se->run_list, queue);
1165         }
1166 }
1167
1168 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1169 {
1170         struct sched_rt_entity *rt_se = &p->rt;
1171         struct rt_rq *rt_rq;
1172
1173         for_each_sched_rt_entity(rt_se) {
1174                 rt_rq = rt_rq_of_se(rt_se);
1175                 requeue_rt_entity(rt_rq, rt_se, head);
1176         }
1177 }
1178
1179 static void yield_task_rt(struct rq *rq)
1180 {
1181         requeue_task_rt(rq, rq->curr, 0);
1182 }
1183
1184 #ifdef CONFIG_SMP
1185 static int find_lowest_rq(struct task_struct *task);
1186
1187 static int
1188 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1189 {
1190         struct task_struct *curr;
1191         struct rq *rq;
1192
1193         if (p->nr_cpus_allowed == 1)
1194                 goto out;
1195
1196         /* For anything but wake ups, just return the task_cpu */
1197         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1198                 goto out;
1199
1200         rq = cpu_rq(cpu);
1201
1202         rcu_read_lock();
1203         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1204
1205         /*
1206          * If the current task on @p's runqueue is an RT task, then
1207          * try to see if we can wake this RT task up on another
1208          * runqueue. Otherwise simply start this RT task
1209          * on its current runqueue.
1210          *
1211          * We want to avoid overloading runqueues. If the woken
1212          * task is a higher priority, then it will stay on this CPU
1213          * and the lower prio task should be moved to another CPU.
1214          * Even though this will probably make the lower prio task
1215          * lose its cache, we do not want to bounce a higher task
1216          * around just because it gave up its CPU, perhaps for a
1217          * lock?
1218          *
1219          * For equal prio tasks, we just let the scheduler sort it out.
1220          *
1221          * Otherwise, just let it ride on the affined RQ and the
1222          * post-schedule router will push the preempted task away
1223          *
1224          * This test is optimistic, if we get it wrong the load-balancer
1225          * will have to sort it out.
1226          */
1227         if (curr && unlikely(rt_task(curr)) &&
1228             (curr->nr_cpus_allowed < 2 ||
1229              curr->prio <= p->prio)) {
1230                 int target = find_lowest_rq(p);
1231
1232                 if (target != -1)
1233                         cpu = target;
1234         }
1235         rcu_read_unlock();
1236
1237 out:
1238         return cpu;
1239 }
1240
1241 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1242 {
1243         if (rq->curr->nr_cpus_allowed == 1)
1244                 return;
1245
1246         if (p->nr_cpus_allowed != 1
1247             && cpupri_find(&rq->rd->cpupri, p, NULL))
1248                 return;
1249
1250         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1251                 return;
1252
1253         /*
1254          * There appears to be other cpus that can accept
1255          * current and none to run 'p', so lets reschedule
1256          * to try and push current away:
1257          */
1258         requeue_task_rt(rq, p, 1);
1259         resched_task(rq->curr);
1260 }
1261
1262 #endif /* CONFIG_SMP */
1263
1264 /*
1265  * Preempt the current task with a newly woken task if needed:
1266  */
1267 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1268 {
1269         if (p->prio < rq->curr->prio) {
1270                 resched_task(rq->curr);
1271                 return;
1272         }
1273
1274 #ifdef CONFIG_SMP
1275         /*
1276          * If:
1277          *
1278          * - the newly woken task is of equal priority to the current task
1279          * - the newly woken task is non-migratable while current is migratable
1280          * - current will be preempted on the next reschedule
1281          *
1282          * we should check to see if current can readily move to a different
1283          * cpu.  If so, we will reschedule to allow the push logic to try
1284          * to move current somewhere else, making room for our non-migratable
1285          * task.
1286          */
1287         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1288                 check_preempt_equal_prio(rq, p);
1289 #endif
1290 }
1291
1292 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1293                                                    struct rt_rq *rt_rq)
1294 {
1295         struct rt_prio_array *array = &rt_rq->active;
1296         struct sched_rt_entity *next = NULL;
1297         struct list_head *queue;
1298         int idx;
1299
1300         idx = sched_find_first_bit(array->bitmap);
1301         BUG_ON(idx >= MAX_RT_PRIO);
1302
1303         queue = array->queue + idx;
1304         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1305
1306         return next;
1307 }
1308
1309 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1310 {
1311         struct sched_rt_entity *rt_se;
1312         struct task_struct *p;
1313         struct rt_rq *rt_rq;
1314
1315         rt_rq = &rq->rt;
1316
1317         if (!rt_rq->rt_nr_running)
1318                 return NULL;
1319
1320         if (rt_rq_throttled(rt_rq))
1321                 return NULL;
1322
1323         do {
1324                 rt_se = pick_next_rt_entity(rq, rt_rq);
1325                 BUG_ON(!rt_se);
1326                 rt_rq = group_rt_rq(rt_se);
1327         } while (rt_rq);
1328
1329         p = rt_task_of(rt_se);
1330         p->se.exec_start = rq_clock_task(rq);
1331
1332         return p;
1333 }
1334
1335 static struct task_struct *pick_next_task_rt(struct rq *rq)
1336 {
1337         struct task_struct *p = _pick_next_task_rt(rq);
1338
1339         /* The running task is never eligible for pushing */
1340         if (p)
1341                 dequeue_pushable_task(rq, p);
1342
1343 #ifdef CONFIG_SMP
1344         /*
1345          * We detect this state here so that we can avoid taking the RQ
1346          * lock again later if there is no need to push
1347          */
1348         rq->post_schedule = has_pushable_tasks(rq);
1349 #endif
1350
1351         return p;
1352 }
1353
1354 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1355 {
1356         update_curr_rt(rq);
1357
1358         /*
1359          * The previous task needs to be made eligible for pushing
1360          * if it is still active
1361          */
1362         if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1363                 enqueue_pushable_task(rq, p);
1364 }
1365
1366 #ifdef CONFIG_SMP
1367
1368 /* Only try algorithms three times */
1369 #define RT_MAX_TRIES 3
1370
1371 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1372 {
1373         if (!task_running(rq, p) &&
1374             cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1375                 return 1;
1376         return 0;
1377 }
1378
1379 /*
1380  * Return the highest pushable rq's task, which is suitable to be executed
1381  * on the cpu, NULL otherwise
1382  */
1383 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1384 {
1385         struct plist_head *head = &rq->rt.pushable_tasks;
1386         struct task_struct *p;
1387
1388         if (!has_pushable_tasks(rq))
1389                 return NULL;
1390
1391         plist_for_each_entry(p, head, pushable_tasks) {
1392                 if (pick_rt_task(rq, p, cpu))
1393                         return p;
1394         }
1395
1396         return NULL;
1397 }
1398
1399 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1400
1401 static int find_lowest_rq(struct task_struct *task)
1402 {
1403         struct sched_domain *sd;
1404         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1405         int this_cpu = smp_processor_id();
1406         int cpu      = task_cpu(task);
1407
1408         /* Make sure the mask is initialized first */
1409         if (unlikely(!lowest_mask))
1410                 return -1;
1411
1412         if (task->nr_cpus_allowed == 1)
1413                 return -1; /* No other targets possible */
1414
1415         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1416                 return -1; /* No targets found */
1417
1418         /*
1419          * At this point we have built a mask of cpus representing the
1420          * lowest priority tasks in the system.  Now we want to elect
1421          * the best one based on our affinity and topology.
1422          *
1423          * We prioritize the last cpu that the task executed on since
1424          * it is most likely cache-hot in that location.
1425          */
1426         if (cpumask_test_cpu(cpu, lowest_mask))
1427                 return cpu;
1428
1429         /*
1430          * Otherwise, we consult the sched_domains span maps to figure
1431          * out which cpu is logically closest to our hot cache data.
1432          */
1433         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1434                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1435
1436         rcu_read_lock();
1437         for_each_domain(cpu, sd) {
1438                 if (sd->flags & SD_WAKE_AFFINE) {
1439                         int best_cpu;
1440
1441                         /*
1442                          * "this_cpu" is cheaper to preempt than a
1443                          * remote processor.
1444                          */
1445                         if (this_cpu != -1 &&
1446                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1447                                 rcu_read_unlock();
1448                                 return this_cpu;
1449                         }
1450
1451                         best_cpu = cpumask_first_and(lowest_mask,
1452                                                      sched_domain_span(sd));
1453                         if (best_cpu < nr_cpu_ids) {
1454                                 rcu_read_unlock();
1455                                 return best_cpu;
1456                         }
1457                 }
1458         }
1459         rcu_read_unlock();
1460
1461         /*
1462          * And finally, if there were no matches within the domains
1463          * just give the caller *something* to work with from the compatible
1464          * locations.
1465          */
1466         if (this_cpu != -1)
1467                 return this_cpu;
1468
1469         cpu = cpumask_any(lowest_mask);
1470         if (cpu < nr_cpu_ids)
1471                 return cpu;
1472         return -1;
1473 }
1474
1475 /* Will lock the rq it finds */
1476 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1477 {
1478         struct rq *lowest_rq = NULL;
1479         int tries;
1480         int cpu;
1481
1482         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1483                 cpu = find_lowest_rq(task);
1484
1485                 if ((cpu == -1) || (cpu == rq->cpu))
1486                         break;
1487
1488                 lowest_rq = cpu_rq(cpu);
1489
1490                 /* if the prio of this runqueue changed, try again */
1491                 if (double_lock_balance(rq, lowest_rq)) {
1492                         /*
1493                          * We had to unlock the run queue. In
1494                          * the mean time, task could have
1495                          * migrated already or had its affinity changed.
1496                          * Also make sure that it wasn't scheduled on its rq.
1497                          */
1498                         if (unlikely(task_rq(task) != rq ||
1499                                      !cpumask_test_cpu(lowest_rq->cpu,
1500                                                        tsk_cpus_allowed(task)) ||
1501                                      task_running(rq, task) ||
1502                                      !task->on_rq)) {
1503
1504                                 double_unlock_balance(rq, lowest_rq);
1505                                 lowest_rq = NULL;
1506                                 break;
1507                         }
1508                 }
1509
1510                 /* If this rq is still suitable use it. */
1511                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1512                         break;
1513
1514                 /* try again */
1515                 double_unlock_balance(rq, lowest_rq);
1516                 lowest_rq = NULL;
1517         }
1518
1519         return lowest_rq;
1520 }
1521
1522 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1523 {
1524         struct task_struct *p;
1525
1526         if (!has_pushable_tasks(rq))
1527                 return NULL;
1528
1529         p = plist_first_entry(&rq->rt.pushable_tasks,
1530                               struct task_struct, pushable_tasks);
1531
1532         BUG_ON(rq->cpu != task_cpu(p));
1533         BUG_ON(task_current(rq, p));
1534         BUG_ON(p->nr_cpus_allowed <= 1);
1535
1536         BUG_ON(!p->on_rq);
1537         BUG_ON(!rt_task(p));
1538
1539         return p;
1540 }
1541
1542 /*
1543  * If the current CPU has more than one RT task, see if the non
1544  * running task can migrate over to a CPU that is running a task
1545  * of lesser priority.
1546  */
1547 static int push_rt_task(struct rq *rq)
1548 {
1549         struct task_struct *next_task;
1550         struct rq *lowest_rq;
1551         int ret = 0;
1552
1553         if (!rq->rt.overloaded)
1554                 return 0;
1555
1556         next_task = pick_next_pushable_task(rq);
1557         if (!next_task)
1558                 return 0;
1559
1560 retry:
1561         if (unlikely(next_task == rq->curr)) {
1562                 WARN_ON(1);
1563                 return 0;
1564         }
1565
1566         /*
1567          * It's possible that the next_task slipped in of
1568          * higher priority than current. If that's the case
1569          * just reschedule current.
1570          */
1571         if (unlikely(next_task->prio < rq->curr->prio)) {
1572                 resched_task(rq->curr);
1573                 return 0;
1574         }
1575
1576         /* We might release rq lock */
1577         get_task_struct(next_task);
1578
1579         /* find_lock_lowest_rq locks the rq if found */
1580         lowest_rq = find_lock_lowest_rq(next_task, rq);
1581         if (!lowest_rq) {
1582                 struct task_struct *task;
1583                 /*
1584                  * find_lock_lowest_rq releases rq->lock
1585                  * so it is possible that next_task has migrated.
1586                  *
1587                  * We need to make sure that the task is still on the same
1588                  * run-queue and is also still the next task eligible for
1589                  * pushing.
1590                  */
1591                 task = pick_next_pushable_task(rq);
1592                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1593                         /*
1594                          * The task hasn't migrated, and is still the next
1595                          * eligible task, but we failed to find a run-queue
1596                          * to push it to.  Do not retry in this case, since
1597                          * other cpus will pull from us when ready.
1598                          */
1599                         goto out;
1600                 }
1601
1602                 if (!task)
1603                         /* No more tasks, just exit */
1604                         goto out;
1605
1606                 /*
1607                  * Something has shifted, try again.
1608                  */
1609                 put_task_struct(next_task);
1610                 next_task = task;
1611                 goto retry;
1612         }
1613
1614         deactivate_task(rq, next_task, 0);
1615         set_task_cpu(next_task, lowest_rq->cpu);
1616         activate_task(lowest_rq, next_task, 0);
1617         ret = 1;
1618
1619         resched_task(lowest_rq->curr);
1620
1621         double_unlock_balance(rq, lowest_rq);
1622
1623 out:
1624         put_task_struct(next_task);
1625
1626         return ret;
1627 }
1628
1629 static void push_rt_tasks(struct rq *rq)
1630 {
1631         /* push_rt_task will return true if it moved an RT */
1632         while (push_rt_task(rq))
1633                 ;
1634 }
1635
1636 static int pull_rt_task(struct rq *this_rq)
1637 {
1638         int this_cpu = this_rq->cpu, ret = 0, cpu;
1639         struct task_struct *p;
1640         struct rq *src_rq;
1641
1642         if (likely(!rt_overloaded(this_rq)))
1643                 return 0;
1644
1645         /*
1646          * Match the barrier from rt_set_overloaded; this guarantees that if we
1647          * see overloaded we must also see the rto_mask bit.
1648          */
1649         smp_rmb();
1650
1651         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1652                 if (this_cpu == cpu)
1653                         continue;
1654
1655                 src_rq = cpu_rq(cpu);
1656
1657                 /*
1658                  * Don't bother taking the src_rq->lock if the next highest
1659                  * task is known to be lower-priority than our current task.
1660                  * This may look racy, but if this value is about to go
1661                  * logically higher, the src_rq will push this task away.
1662                  * And if its going logically lower, we do not care
1663                  */
1664                 if (src_rq->rt.highest_prio.next >=
1665                     this_rq->rt.highest_prio.curr)
1666                         continue;
1667
1668                 /*
1669                  * We can potentially drop this_rq's lock in
1670                  * double_lock_balance, and another CPU could
1671                  * alter this_rq
1672                  */
1673                 double_lock_balance(this_rq, src_rq);
1674
1675                 /*
1676                  * We can pull only a task, which is pushable
1677                  * on its rq, and no others.
1678                  */
1679                 p = pick_highest_pushable_task(src_rq, this_cpu);
1680
1681                 /*
1682                  * Do we have an RT task that preempts
1683                  * the to-be-scheduled task?
1684                  */
1685                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1686                         WARN_ON(p == src_rq->curr);
1687                         WARN_ON(!p->on_rq);
1688
1689                         /*
1690                          * There's a chance that p is higher in priority
1691                          * than what's currently running on its cpu.
1692                          * This is just that p is wakeing up and hasn't
1693                          * had a chance to schedule. We only pull
1694                          * p if it is lower in priority than the
1695                          * current task on the run queue
1696                          */
1697                         if (p->prio < src_rq->curr->prio)
1698                                 goto skip;
1699
1700                         ret = 1;
1701
1702                         deactivate_task(src_rq, p, 0);
1703                         set_task_cpu(p, this_cpu);
1704                         activate_task(this_rq, p, 0);
1705                         /*
1706                          * We continue with the search, just in
1707                          * case there's an even higher prio task
1708                          * in another runqueue. (low likelihood
1709                          * but possible)
1710                          */
1711                 }
1712 skip:
1713                 double_unlock_balance(this_rq, src_rq);
1714         }
1715
1716         return ret;
1717 }
1718
1719 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1720 {
1721         /* Try to pull RT tasks here if we lower this rq's prio */
1722         if (rq->rt.highest_prio.curr > prev->prio)
1723                 pull_rt_task(rq);
1724 }
1725
1726 static void post_schedule_rt(struct rq *rq)
1727 {
1728         push_rt_tasks(rq);
1729 }
1730
1731 /*
1732  * If we are not running and we are not going to reschedule soon, we should
1733  * try to push tasks away now
1734  */
1735 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1736 {
1737         if (!task_running(rq, p) &&
1738             !test_tsk_need_resched(rq->curr) &&
1739             has_pushable_tasks(rq) &&
1740             p->nr_cpus_allowed > 1 &&
1741             rt_task(rq->curr) &&
1742             (rq->curr->nr_cpus_allowed < 2 ||
1743              rq->curr->prio <= p->prio))
1744                 push_rt_tasks(rq);
1745 }
1746
1747 static void set_cpus_allowed_rt(struct task_struct *p,
1748                                 const struct cpumask *new_mask)
1749 {
1750         struct rq *rq;
1751         int weight;
1752
1753         BUG_ON(!rt_task(p));
1754
1755         if (!p->on_rq)
1756                 return;
1757
1758         weight = cpumask_weight(new_mask);
1759
1760         /*
1761          * Only update if the process changes its state from whether it
1762          * can migrate or not.
1763          */
1764         if ((p->nr_cpus_allowed > 1) == (weight > 1))
1765                 return;
1766
1767         rq = task_rq(p);
1768
1769         /*
1770          * The process used to be able to migrate OR it can now migrate
1771          */
1772         if (weight <= 1) {
1773                 if (!task_current(rq, p))
1774                         dequeue_pushable_task(rq, p);
1775                 BUG_ON(!rq->rt.rt_nr_migratory);
1776                 rq->rt.rt_nr_migratory--;
1777         } else {
1778                 if (!task_current(rq, p))
1779                         enqueue_pushable_task(rq, p);
1780                 rq->rt.rt_nr_migratory++;
1781         }
1782
1783         update_rt_migration(&rq->rt);
1784 }
1785
1786 /* Assumes rq->lock is held */
1787 static void rq_online_rt(struct rq *rq)
1788 {
1789         if (rq->rt.overloaded)
1790                 rt_set_overload(rq);
1791
1792         __enable_runtime(rq);
1793
1794         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1795 }
1796
1797 /* Assumes rq->lock is held */
1798 static void rq_offline_rt(struct rq *rq)
1799 {
1800         if (rq->rt.overloaded)
1801                 rt_clear_overload(rq);
1802
1803         __disable_runtime(rq);
1804
1805         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1806 }
1807
1808 /*
1809  * When switch from the rt queue, we bring ourselves to a position
1810  * that we might want to pull RT tasks from other runqueues.
1811  */
1812 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1813 {
1814         /*
1815          * If there are other RT tasks then we will reschedule
1816          * and the scheduling of the other RT tasks will handle
1817          * the balancing. But if we are the last RT task
1818          * we may need to handle the pulling of RT tasks
1819          * now.
1820          */
1821         if (!p->on_rq || rq->rt.rt_nr_running)
1822                 return;
1823
1824         if (pull_rt_task(rq))
1825                 resched_task(rq->curr);
1826 }
1827
1828 void init_sched_rt_class(void)
1829 {
1830         unsigned int i;
1831
1832         for_each_possible_cpu(i) {
1833                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1834                                         GFP_KERNEL, cpu_to_node(i));
1835         }
1836 }
1837 #endif /* CONFIG_SMP */
1838
1839 /*
1840  * When switching a task to RT, we may overload the runqueue
1841  * with RT tasks. In this case we try to push them off to
1842  * other runqueues.
1843  */
1844 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1845 {
1846         int check_resched = 1;
1847
1848         /*
1849          * If we are already running, then there's nothing
1850          * that needs to be done. But if we are not running
1851          * we may need to preempt the current running task.
1852          * If that current running task is also an RT task
1853          * then see if we can move to another run queue.
1854          */
1855         if (p->on_rq && rq->curr != p) {
1856 #ifdef CONFIG_SMP
1857                 if (rq->rt.overloaded && push_rt_task(rq) &&
1858                     /* Don't resched if we changed runqueues */
1859                     rq != task_rq(p))
1860                         check_resched = 0;
1861 #endif /* CONFIG_SMP */
1862                 if (check_resched && p->prio < rq->curr->prio)
1863                         resched_task(rq->curr);
1864         }
1865 }
1866
1867 /*
1868  * Priority of the task has changed. This may cause
1869  * us to initiate a push or pull.
1870  */
1871 static void
1872 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1873 {
1874         if (!p->on_rq)
1875                 return;
1876
1877         if (rq->curr == p) {
1878 #ifdef CONFIG_SMP
1879                 /*
1880                  * If our priority decreases while running, we
1881                  * may need to pull tasks to this runqueue.
1882                  */
1883                 if (oldprio < p->prio)
1884                         pull_rt_task(rq);
1885                 /*
1886                  * If there's a higher priority task waiting to run
1887                  * then reschedule. Note, the above pull_rt_task
1888                  * can release the rq lock and p could migrate.
1889                  * Only reschedule if p is still on the same runqueue.
1890                  */
1891                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1892                         resched_task(p);
1893 #else
1894                 /* For UP simply resched on drop of prio */
1895                 if (oldprio < p->prio)
1896                         resched_task(p);
1897 #endif /* CONFIG_SMP */
1898         } else {
1899                 /*
1900                  * This task is not running, but if it is
1901                  * greater than the current running task
1902                  * then reschedule.
1903                  */
1904                 if (p->prio < rq->curr->prio)
1905                         resched_task(rq->curr);
1906         }
1907 }
1908
1909 static void watchdog(struct rq *rq, struct task_struct *p)
1910 {
1911         unsigned long soft, hard;
1912
1913         /* max may change after cur was read, this will be fixed next tick */
1914         soft = task_rlimit(p, RLIMIT_RTTIME);
1915         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1916
1917         if (soft != RLIM_INFINITY) {
1918                 unsigned long next;
1919
1920                 if (p->rt.watchdog_stamp != jiffies) {
1921                         p->rt.timeout++;
1922                         p->rt.watchdog_stamp = jiffies;
1923                 }
1924
1925                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1926                 if (p->rt.timeout > next)
1927                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1928         }
1929 }
1930
1931 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1932 {
1933         struct sched_rt_entity *rt_se = &p->rt;
1934
1935         update_curr_rt(rq);
1936
1937         watchdog(rq, p);
1938
1939         /*
1940          * RR tasks need a special form of timeslice management.
1941          * FIFO tasks have no timeslices.
1942          */
1943         if (p->policy != SCHED_RR)
1944                 return;
1945
1946         if (--p->rt.time_slice)
1947                 return;
1948
1949         p->rt.time_slice = sched_rr_timeslice;
1950
1951         /*
1952          * Requeue to the end of queue if we (and all of our ancestors) are not
1953          * the only element on the queue
1954          */
1955         for_each_sched_rt_entity(rt_se) {
1956                 if (rt_se->run_list.prev != rt_se->run_list.next) {
1957                         requeue_task_rt(rq, p, 0);
1958                         set_tsk_need_resched(p);
1959                         return;
1960                 }
1961         }
1962 }
1963
1964 static void set_curr_task_rt(struct rq *rq)
1965 {
1966         struct task_struct *p = rq->curr;
1967
1968         p->se.exec_start = rq_clock_task(rq);
1969
1970         /* The running task is never eligible for pushing */
1971         dequeue_pushable_task(rq, p);
1972 }
1973
1974 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1975 {
1976         /*
1977          * Time slice is 0 for SCHED_FIFO tasks
1978          */
1979         if (task->policy == SCHED_RR)
1980                 return sched_rr_timeslice;
1981         else
1982                 return 0;
1983 }
1984
1985 const struct sched_class rt_sched_class = {
1986         .next                   = &fair_sched_class,
1987         .enqueue_task           = enqueue_task_rt,
1988         .dequeue_task           = dequeue_task_rt,
1989         .yield_task             = yield_task_rt,
1990
1991         .check_preempt_curr     = check_preempt_curr_rt,
1992
1993         .pick_next_task         = pick_next_task_rt,
1994         .put_prev_task          = put_prev_task_rt,
1995
1996 #ifdef CONFIG_SMP
1997         .select_task_rq         = select_task_rq_rt,
1998
1999         .set_cpus_allowed       = set_cpus_allowed_rt,
2000         .rq_online              = rq_online_rt,
2001         .rq_offline             = rq_offline_rt,
2002         .pre_schedule           = pre_schedule_rt,
2003         .post_schedule          = post_schedule_rt,
2004         .task_woken             = task_woken_rt,
2005         .switched_from          = switched_from_rt,
2006 #endif
2007
2008         .set_curr_task          = set_curr_task_rt,
2009         .task_tick              = task_tick_rt,
2010
2011         .get_rr_interval        = get_rr_interval_rt,
2012
2013         .prio_changed           = prio_changed_rt,
2014         .switched_to            = switched_to_rt,
2015 };
2016
2017 #ifdef CONFIG_SCHED_DEBUG
2018 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2019
2020 void print_rt_stats(struct seq_file *m, int cpu)
2021 {
2022         rt_rq_iter_t iter;
2023         struct rt_rq *rt_rq;
2024
2025         rcu_read_lock();
2026         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2027                 print_rt_rq(m, cpu, rt_rq);
2028         rcu_read_unlock();
2029 }
2030 #endif /* CONFIG_SCHED_DEBUG */