Merge branch 'for-3.5-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj...
[firefly-linux-kernel-4.4.55.git] / kernel / sched / 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 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
11
12 struct rt_bandwidth def_rt_bandwidth;
13
14 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
15 {
16         struct rt_bandwidth *rt_b =
17                 container_of(timer, struct rt_bandwidth, rt_period_timer);
18         ktime_t now;
19         int overrun;
20         int idle = 0;
21
22         for (;;) {
23                 now = hrtimer_cb_get_time(timer);
24                 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
25
26                 if (!overrun)
27                         break;
28
29                 idle = do_sched_rt_period_timer(rt_b, overrun);
30         }
31
32         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
33 }
34
35 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
36 {
37         rt_b->rt_period = ns_to_ktime(period);
38         rt_b->rt_runtime = runtime;
39
40         raw_spin_lock_init(&rt_b->rt_runtime_lock);
41
42         hrtimer_init(&rt_b->rt_period_timer,
43                         CLOCK_MONOTONIC, HRTIMER_MODE_REL);
44         rt_b->rt_period_timer.function = sched_rt_period_timer;
45 }
46
47 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
48 {
49         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
50                 return;
51
52         if (hrtimer_active(&rt_b->rt_period_timer))
53                 return;
54
55         raw_spin_lock(&rt_b->rt_runtime_lock);
56         start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
57         raw_spin_unlock(&rt_b->rt_runtime_lock);
58 }
59
60 void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
61 {
62         struct rt_prio_array *array;
63         int i;
64
65         array = &rt_rq->active;
66         for (i = 0; i < MAX_RT_PRIO; i++) {
67                 INIT_LIST_HEAD(array->queue + i);
68                 __clear_bit(i, array->bitmap);
69         }
70         /* delimiter for bitsearch: */
71         __set_bit(MAX_RT_PRIO, array->bitmap);
72
73 #if defined CONFIG_SMP
74         rt_rq->highest_prio.curr = MAX_RT_PRIO;
75         rt_rq->highest_prio.next = MAX_RT_PRIO;
76         rt_rq->rt_nr_migratory = 0;
77         rt_rq->overloaded = 0;
78         plist_head_init(&rt_rq->pushable_tasks);
79 #endif
80
81         rt_rq->rt_time = 0;
82         rt_rq->rt_throttled = 0;
83         rt_rq->rt_runtime = 0;
84         raw_spin_lock_init(&rt_rq->rt_runtime_lock);
85 }
86
87 #ifdef CONFIG_RT_GROUP_SCHED
88 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
89 {
90         hrtimer_cancel(&rt_b->rt_period_timer);
91 }
92
93 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
94
95 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
96 {
97 #ifdef CONFIG_SCHED_DEBUG
98         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
99 #endif
100         return container_of(rt_se, struct task_struct, rt);
101 }
102
103 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
104 {
105         return rt_rq->rq;
106 }
107
108 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
109 {
110         return rt_se->rt_rq;
111 }
112
113 void free_rt_sched_group(struct task_group *tg)
114 {
115         int i;
116
117         if (tg->rt_se)
118                 destroy_rt_bandwidth(&tg->rt_bandwidth);
119
120         for_each_possible_cpu(i) {
121                 if (tg->rt_rq)
122                         kfree(tg->rt_rq[i]);
123                 if (tg->rt_se)
124                         kfree(tg->rt_se[i]);
125         }
126
127         kfree(tg->rt_rq);
128         kfree(tg->rt_se);
129 }
130
131 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
132                 struct sched_rt_entity *rt_se, int cpu,
133                 struct sched_rt_entity *parent)
134 {
135         struct rq *rq = cpu_rq(cpu);
136
137         rt_rq->highest_prio.curr = MAX_RT_PRIO;
138         rt_rq->rt_nr_boosted = 0;
139         rt_rq->rq = rq;
140         rt_rq->tg = tg;
141
142         tg->rt_rq[cpu] = rt_rq;
143         tg->rt_se[cpu] = rt_se;
144
145         if (!rt_se)
146                 return;
147
148         if (!parent)
149                 rt_se->rt_rq = &rq->rt;
150         else
151                 rt_se->rt_rq = parent->my_q;
152
153         rt_se->my_q = rt_rq;
154         rt_se->parent = parent;
155         INIT_LIST_HEAD(&rt_se->run_list);
156 }
157
158 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
159 {
160         struct rt_rq *rt_rq;
161         struct sched_rt_entity *rt_se;
162         int i;
163
164         tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
165         if (!tg->rt_rq)
166                 goto err;
167         tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
168         if (!tg->rt_se)
169                 goto err;
170
171         init_rt_bandwidth(&tg->rt_bandwidth,
172                         ktime_to_ns(def_rt_bandwidth.rt_period), 0);
173
174         for_each_possible_cpu(i) {
175                 rt_rq = kzalloc_node(sizeof(struct rt_rq),
176                                      GFP_KERNEL, cpu_to_node(i));
177                 if (!rt_rq)
178                         goto err;
179
180                 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
181                                      GFP_KERNEL, cpu_to_node(i));
182                 if (!rt_se)
183                         goto err_free_rq;
184
185                 init_rt_rq(rt_rq, cpu_rq(i));
186                 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
187                 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
188         }
189
190         return 1;
191
192 err_free_rq:
193         kfree(rt_rq);
194 err:
195         return 0;
196 }
197
198 #else /* CONFIG_RT_GROUP_SCHED */
199
200 #define rt_entity_is_task(rt_se) (1)
201
202 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
203 {
204         return container_of(rt_se, struct task_struct, rt);
205 }
206
207 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
208 {
209         return container_of(rt_rq, struct rq, rt);
210 }
211
212 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
213 {
214         struct task_struct *p = rt_task_of(rt_se);
215         struct rq *rq = task_rq(p);
216
217         return &rq->rt;
218 }
219
220 void free_rt_sched_group(struct task_group *tg) { }
221
222 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
223 {
224         return 1;
225 }
226 #endif /* CONFIG_RT_GROUP_SCHED */
227
228 #ifdef CONFIG_SMP
229
230 static inline int rt_overloaded(struct rq *rq)
231 {
232         return atomic_read(&rq->rd->rto_count);
233 }
234
235 static inline void rt_set_overload(struct rq *rq)
236 {
237         if (!rq->online)
238                 return;
239
240         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
241         /*
242          * Make sure the mask is visible before we set
243          * the overload count. That is checked to determine
244          * if we should look at the mask. It would be a shame
245          * if we looked at the mask, but the mask was not
246          * updated yet.
247          */
248         wmb();
249         atomic_inc(&rq->rd->rto_count);
250 }
251
252 static inline void rt_clear_overload(struct rq *rq)
253 {
254         if (!rq->online)
255                 return;
256
257         /* the order here really doesn't matter */
258         atomic_dec(&rq->rd->rto_count);
259         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
260 }
261
262 static void update_rt_migration(struct rt_rq *rt_rq)
263 {
264         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
265                 if (!rt_rq->overloaded) {
266                         rt_set_overload(rq_of_rt_rq(rt_rq));
267                         rt_rq->overloaded = 1;
268                 }
269         } else if (rt_rq->overloaded) {
270                 rt_clear_overload(rq_of_rt_rq(rt_rq));
271                 rt_rq->overloaded = 0;
272         }
273 }
274
275 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
276 {
277         struct task_struct *p;
278
279         if (!rt_entity_is_task(rt_se))
280                 return;
281
282         p = rt_task_of(rt_se);
283         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
284
285         rt_rq->rt_nr_total++;
286         if (p->nr_cpus_allowed > 1)
287                 rt_rq->rt_nr_migratory++;
288
289         update_rt_migration(rt_rq);
290 }
291
292 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
293 {
294         struct task_struct *p;
295
296         if (!rt_entity_is_task(rt_se))
297                 return;
298
299         p = rt_task_of(rt_se);
300         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
301
302         rt_rq->rt_nr_total--;
303         if (p->nr_cpus_allowed > 1)
304                 rt_rq->rt_nr_migratory--;
305
306         update_rt_migration(rt_rq);
307 }
308
309 static inline int has_pushable_tasks(struct rq *rq)
310 {
311         return !plist_head_empty(&rq->rt.pushable_tasks);
312 }
313
314 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
315 {
316         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
317         plist_node_init(&p->pushable_tasks, p->prio);
318         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
319
320         /* Update the highest prio pushable task */
321         if (p->prio < rq->rt.highest_prio.next)
322                 rq->rt.highest_prio.next = p->prio;
323 }
324
325 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
326 {
327         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
328
329         /* Update the new highest prio pushable task */
330         if (has_pushable_tasks(rq)) {
331                 p = plist_first_entry(&rq->rt.pushable_tasks,
332                                       struct task_struct, pushable_tasks);
333                 rq->rt.highest_prio.next = p->prio;
334         } else
335                 rq->rt.highest_prio.next = MAX_RT_PRIO;
336 }
337
338 #else
339
340 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
341 {
342 }
343
344 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
345 {
346 }
347
348 static inline
349 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
350 {
351 }
352
353 static inline
354 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
355 {
356 }
357
358 #endif /* CONFIG_SMP */
359
360 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
361 {
362         return !list_empty(&rt_se->run_list);
363 }
364
365 #ifdef CONFIG_RT_GROUP_SCHED
366
367 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
368 {
369         if (!rt_rq->tg)
370                 return RUNTIME_INF;
371
372         return rt_rq->rt_runtime;
373 }
374
375 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
376 {
377         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
378 }
379
380 typedef struct task_group *rt_rq_iter_t;
381
382 static inline struct task_group *next_task_group(struct task_group *tg)
383 {
384         do {
385                 tg = list_entry_rcu(tg->list.next,
386                         typeof(struct task_group), list);
387         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
388
389         if (&tg->list == &task_groups)
390                 tg = NULL;
391
392         return tg;
393 }
394
395 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
396         for (iter = container_of(&task_groups, typeof(*iter), list);    \
397                 (iter = next_task_group(iter)) &&                       \
398                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
399
400 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
401 {
402         list_add_rcu(&rt_rq->leaf_rt_rq_list,
403                         &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
404 }
405
406 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
407 {
408         list_del_rcu(&rt_rq->leaf_rt_rq_list);
409 }
410
411 #define for_each_leaf_rt_rq(rt_rq, rq) \
412         list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
413
414 #define for_each_sched_rt_entity(rt_se) \
415         for (; rt_se; rt_se = rt_se->parent)
416
417 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
418 {
419         return rt_se->my_q;
420 }
421
422 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
423 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
424
425 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
426 {
427         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
428         struct sched_rt_entity *rt_se;
429
430         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
431
432         rt_se = rt_rq->tg->rt_se[cpu];
433
434         if (rt_rq->rt_nr_running) {
435                 if (rt_se && !on_rt_rq(rt_se))
436                         enqueue_rt_entity(rt_se, false);
437                 if (rt_rq->highest_prio.curr < curr->prio)
438                         resched_task(curr);
439         }
440 }
441
442 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
443 {
444         struct sched_rt_entity *rt_se;
445         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
446
447         rt_se = rt_rq->tg->rt_se[cpu];
448
449         if (rt_se && on_rt_rq(rt_se))
450                 dequeue_rt_entity(rt_se);
451 }
452
453 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
454 {
455         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
456 }
457
458 static int rt_se_boosted(struct sched_rt_entity *rt_se)
459 {
460         struct rt_rq *rt_rq = group_rt_rq(rt_se);
461         struct task_struct *p;
462
463         if (rt_rq)
464                 return !!rt_rq->rt_nr_boosted;
465
466         p = rt_task_of(rt_se);
467         return p->prio != p->normal_prio;
468 }
469
470 #ifdef CONFIG_SMP
471 static inline const struct cpumask *sched_rt_period_mask(void)
472 {
473         return cpu_rq(smp_processor_id())->rd->span;
474 }
475 #else
476 static inline const struct cpumask *sched_rt_period_mask(void)
477 {
478         return cpu_online_mask;
479 }
480 #endif
481
482 static inline
483 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
484 {
485         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
486 }
487
488 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
489 {
490         return &rt_rq->tg->rt_bandwidth;
491 }
492
493 #else /* !CONFIG_RT_GROUP_SCHED */
494
495 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
496 {
497         return rt_rq->rt_runtime;
498 }
499
500 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
501 {
502         return ktime_to_ns(def_rt_bandwidth.rt_period);
503 }
504
505 typedef struct rt_rq *rt_rq_iter_t;
506
507 #define for_each_rt_rq(rt_rq, iter, rq) \
508         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
509
510 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
511 {
512 }
513
514 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
515 {
516 }
517
518 #define for_each_leaf_rt_rq(rt_rq, rq) \
519         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
520
521 #define for_each_sched_rt_entity(rt_se) \
522         for (; rt_se; rt_se = NULL)
523
524 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
525 {
526         return NULL;
527 }
528
529 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
530 {
531         if (rt_rq->rt_nr_running)
532                 resched_task(rq_of_rt_rq(rt_rq)->curr);
533 }
534
535 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
536 {
537 }
538
539 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
540 {
541         return rt_rq->rt_throttled;
542 }
543
544 static inline const struct cpumask *sched_rt_period_mask(void)
545 {
546         return cpu_online_mask;
547 }
548
549 static inline
550 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
551 {
552         return &cpu_rq(cpu)->rt;
553 }
554
555 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
556 {
557         return &def_rt_bandwidth;
558 }
559
560 #endif /* CONFIG_RT_GROUP_SCHED */
561
562 #ifdef CONFIG_SMP
563 /*
564  * We ran out of runtime, see if we can borrow some from our neighbours.
565  */
566 static int do_balance_runtime(struct rt_rq *rt_rq)
567 {
568         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
569         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
570         int i, weight, more = 0;
571         u64 rt_period;
572
573         weight = cpumask_weight(rd->span);
574
575         raw_spin_lock(&rt_b->rt_runtime_lock);
576         rt_period = ktime_to_ns(rt_b->rt_period);
577         for_each_cpu(i, rd->span) {
578                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
579                 s64 diff;
580
581                 if (iter == rt_rq)
582                         continue;
583
584                 raw_spin_lock(&iter->rt_runtime_lock);
585                 /*
586                  * Either all rqs have inf runtime and there's nothing to steal
587                  * or __disable_runtime() below sets a specific rq to inf to
588                  * indicate its been disabled and disalow stealing.
589                  */
590                 if (iter->rt_runtime == RUNTIME_INF)
591                         goto next;
592
593                 /*
594                  * From runqueues with spare time, take 1/n part of their
595                  * spare time, but no more than our period.
596                  */
597                 diff = iter->rt_runtime - iter->rt_time;
598                 if (diff > 0) {
599                         diff = div_u64((u64)diff, weight);
600                         if (rt_rq->rt_runtime + diff > rt_period)
601                                 diff = rt_period - rt_rq->rt_runtime;
602                         iter->rt_runtime -= diff;
603                         rt_rq->rt_runtime += diff;
604                         more = 1;
605                         if (rt_rq->rt_runtime == rt_period) {
606                                 raw_spin_unlock(&iter->rt_runtime_lock);
607                                 break;
608                         }
609                 }
610 next:
611                 raw_spin_unlock(&iter->rt_runtime_lock);
612         }
613         raw_spin_unlock(&rt_b->rt_runtime_lock);
614
615         return more;
616 }
617
618 /*
619  * Ensure this RQ takes back all the runtime it lend to its neighbours.
620  */
621 static void __disable_runtime(struct rq *rq)
622 {
623         struct root_domain *rd = rq->rd;
624         rt_rq_iter_t iter;
625         struct rt_rq *rt_rq;
626
627         if (unlikely(!scheduler_running))
628                 return;
629
630         for_each_rt_rq(rt_rq, iter, rq) {
631                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
632                 s64 want;
633                 int i;
634
635                 raw_spin_lock(&rt_b->rt_runtime_lock);
636                 raw_spin_lock(&rt_rq->rt_runtime_lock);
637                 /*
638                  * Either we're all inf and nobody needs to borrow, or we're
639                  * already disabled and thus have nothing to do, or we have
640                  * exactly the right amount of runtime to take out.
641                  */
642                 if (rt_rq->rt_runtime == RUNTIME_INF ||
643                                 rt_rq->rt_runtime == rt_b->rt_runtime)
644                         goto balanced;
645                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
646
647                 /*
648                  * Calculate the difference between what we started out with
649                  * and what we current have, that's the amount of runtime
650                  * we lend and now have to reclaim.
651                  */
652                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
653
654                 /*
655                  * Greedy reclaim, take back as much as we can.
656                  */
657                 for_each_cpu(i, rd->span) {
658                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
659                         s64 diff;
660
661                         /*
662                          * Can't reclaim from ourselves or disabled runqueues.
663                          */
664                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
665                                 continue;
666
667                         raw_spin_lock(&iter->rt_runtime_lock);
668                         if (want > 0) {
669                                 diff = min_t(s64, iter->rt_runtime, want);
670                                 iter->rt_runtime -= diff;
671                                 want -= diff;
672                         } else {
673                                 iter->rt_runtime -= want;
674                                 want -= want;
675                         }
676                         raw_spin_unlock(&iter->rt_runtime_lock);
677
678                         if (!want)
679                                 break;
680                 }
681
682                 raw_spin_lock(&rt_rq->rt_runtime_lock);
683                 /*
684                  * We cannot be left wanting - that would mean some runtime
685                  * leaked out of the system.
686                  */
687                 BUG_ON(want);
688 balanced:
689                 /*
690                  * Disable all the borrow logic by pretending we have inf
691                  * runtime - in which case borrowing doesn't make sense.
692                  */
693                 rt_rq->rt_runtime = RUNTIME_INF;
694                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
695                 raw_spin_unlock(&rt_b->rt_runtime_lock);
696         }
697 }
698
699 static void disable_runtime(struct rq *rq)
700 {
701         unsigned long flags;
702
703         raw_spin_lock_irqsave(&rq->lock, flags);
704         __disable_runtime(rq);
705         raw_spin_unlock_irqrestore(&rq->lock, flags);
706 }
707
708 static void __enable_runtime(struct rq *rq)
709 {
710         rt_rq_iter_t iter;
711         struct rt_rq *rt_rq;
712
713         if (unlikely(!scheduler_running))
714                 return;
715
716         /*
717          * Reset each runqueue's bandwidth settings
718          */
719         for_each_rt_rq(rt_rq, iter, rq) {
720                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
721
722                 raw_spin_lock(&rt_b->rt_runtime_lock);
723                 raw_spin_lock(&rt_rq->rt_runtime_lock);
724                 rt_rq->rt_runtime = rt_b->rt_runtime;
725                 rt_rq->rt_time = 0;
726                 rt_rq->rt_throttled = 0;
727                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
728                 raw_spin_unlock(&rt_b->rt_runtime_lock);
729         }
730 }
731
732 static void enable_runtime(struct rq *rq)
733 {
734         unsigned long flags;
735
736         raw_spin_lock_irqsave(&rq->lock, flags);
737         __enable_runtime(rq);
738         raw_spin_unlock_irqrestore(&rq->lock, flags);
739 }
740
741 int update_runtime(struct notifier_block *nfb, unsigned long action, void *hcpu)
742 {
743         int cpu = (int)(long)hcpu;
744
745         switch (action) {
746         case CPU_DOWN_PREPARE:
747         case CPU_DOWN_PREPARE_FROZEN:
748                 disable_runtime(cpu_rq(cpu));
749                 return NOTIFY_OK;
750
751         case CPU_DOWN_FAILED:
752         case CPU_DOWN_FAILED_FROZEN:
753         case CPU_ONLINE:
754         case CPU_ONLINE_FROZEN:
755                 enable_runtime(cpu_rq(cpu));
756                 return NOTIFY_OK;
757
758         default:
759                 return NOTIFY_DONE;
760         }
761 }
762
763 static int balance_runtime(struct rt_rq *rt_rq)
764 {
765         int more = 0;
766
767         if (!sched_feat(RT_RUNTIME_SHARE))
768                 return more;
769
770         if (rt_rq->rt_time > rt_rq->rt_runtime) {
771                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
772                 more = do_balance_runtime(rt_rq);
773                 raw_spin_lock(&rt_rq->rt_runtime_lock);
774         }
775
776         return more;
777 }
778 #else /* !CONFIG_SMP */
779 static inline int balance_runtime(struct rt_rq *rt_rq)
780 {
781         return 0;
782 }
783 #endif /* CONFIG_SMP */
784
785 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
786 {
787         int i, idle = 1, throttled = 0;
788         const struct cpumask *span;
789
790         span = sched_rt_period_mask();
791         for_each_cpu(i, span) {
792                 int enqueue = 0;
793                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
794                 struct rq *rq = rq_of_rt_rq(rt_rq);
795
796                 raw_spin_lock(&rq->lock);
797                 if (rt_rq->rt_time) {
798                         u64 runtime;
799
800                         raw_spin_lock(&rt_rq->rt_runtime_lock);
801                         if (rt_rq->rt_throttled)
802                                 balance_runtime(rt_rq);
803                         runtime = rt_rq->rt_runtime;
804                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
805                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
806                                 rt_rq->rt_throttled = 0;
807                                 enqueue = 1;
808
809                                 /*
810                                  * Force a clock update if the CPU was idle,
811                                  * lest wakeup -> unthrottle time accumulate.
812                                  */
813                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
814                                         rq->skip_clock_update = -1;
815                         }
816                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
817                                 idle = 0;
818                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
819                 } else if (rt_rq->rt_nr_running) {
820                         idle = 0;
821                         if (!rt_rq_throttled(rt_rq))
822                                 enqueue = 1;
823                 }
824                 if (rt_rq->rt_throttled)
825                         throttled = 1;
826
827                 if (enqueue)
828                         sched_rt_rq_enqueue(rt_rq);
829                 raw_spin_unlock(&rq->lock);
830         }
831
832         if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
833                 return 1;
834
835         return idle;
836 }
837
838 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
839 {
840 #ifdef CONFIG_RT_GROUP_SCHED
841         struct rt_rq *rt_rq = group_rt_rq(rt_se);
842
843         if (rt_rq)
844                 return rt_rq->highest_prio.curr;
845 #endif
846
847         return rt_task_of(rt_se)->prio;
848 }
849
850 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
851 {
852         u64 runtime = sched_rt_runtime(rt_rq);
853
854         if (rt_rq->rt_throttled)
855                 return rt_rq_throttled(rt_rq);
856
857         if (runtime >= sched_rt_period(rt_rq))
858                 return 0;
859
860         balance_runtime(rt_rq);
861         runtime = sched_rt_runtime(rt_rq);
862         if (runtime == RUNTIME_INF)
863                 return 0;
864
865         if (rt_rq->rt_time > runtime) {
866                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
867
868                 /*
869                  * Don't actually throttle groups that have no runtime assigned
870                  * but accrue some time due to boosting.
871                  */
872                 if (likely(rt_b->rt_runtime)) {
873                         static bool once = false;
874
875                         rt_rq->rt_throttled = 1;
876
877                         if (!once) {
878                                 once = true;
879                                 printk_sched("sched: RT throttling activated\n");
880                         }
881                 } else {
882                         /*
883                          * In case we did anyway, make it go away,
884                          * replenishment is a joke, since it will replenish us
885                          * with exactly 0 ns.
886                          */
887                         rt_rq->rt_time = 0;
888                 }
889
890                 if (rt_rq_throttled(rt_rq)) {
891                         sched_rt_rq_dequeue(rt_rq);
892                         return 1;
893                 }
894         }
895
896         return 0;
897 }
898
899 /*
900  * Update the current task's runtime statistics. Skip current tasks that
901  * are not in our scheduling class.
902  */
903 static void update_curr_rt(struct rq *rq)
904 {
905         struct task_struct *curr = rq->curr;
906         struct sched_rt_entity *rt_se = &curr->rt;
907         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
908         u64 delta_exec;
909
910         if (curr->sched_class != &rt_sched_class)
911                 return;
912
913         delta_exec = rq->clock_task - curr->se.exec_start;
914         if (unlikely((s64)delta_exec < 0))
915                 delta_exec = 0;
916
917         schedstat_set(curr->se.statistics.exec_max,
918                       max(curr->se.statistics.exec_max, delta_exec));
919
920         curr->se.sum_exec_runtime += delta_exec;
921         account_group_exec_runtime(curr, delta_exec);
922
923         curr->se.exec_start = rq->clock_task;
924         cpuacct_charge(curr, delta_exec);
925
926         sched_rt_avg_update(rq, delta_exec);
927
928         if (!rt_bandwidth_enabled())
929                 return;
930
931         for_each_sched_rt_entity(rt_se) {
932                 rt_rq = rt_rq_of_se(rt_se);
933
934                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
935                         raw_spin_lock(&rt_rq->rt_runtime_lock);
936                         rt_rq->rt_time += delta_exec;
937                         if (sched_rt_runtime_exceeded(rt_rq))
938                                 resched_task(curr);
939                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
940                 }
941         }
942 }
943
944 #if defined CONFIG_SMP
945
946 static void
947 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
948 {
949         struct rq *rq = rq_of_rt_rq(rt_rq);
950
951         if (rq->online && prio < prev_prio)
952                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
953 }
954
955 static void
956 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
957 {
958         struct rq *rq = rq_of_rt_rq(rt_rq);
959
960         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
961                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
962 }
963
964 #else /* CONFIG_SMP */
965
966 static inline
967 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
968 static inline
969 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
970
971 #endif /* CONFIG_SMP */
972
973 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
974 static void
975 inc_rt_prio(struct rt_rq *rt_rq, int prio)
976 {
977         int prev_prio = rt_rq->highest_prio.curr;
978
979         if (prio < prev_prio)
980                 rt_rq->highest_prio.curr = prio;
981
982         inc_rt_prio_smp(rt_rq, prio, prev_prio);
983 }
984
985 static void
986 dec_rt_prio(struct rt_rq *rt_rq, int prio)
987 {
988         int prev_prio = rt_rq->highest_prio.curr;
989
990         if (rt_rq->rt_nr_running) {
991
992                 WARN_ON(prio < prev_prio);
993
994                 /*
995                  * This may have been our highest task, and therefore
996                  * we may have some recomputation to do
997                  */
998                 if (prio == prev_prio) {
999                         struct rt_prio_array *array = &rt_rq->active;
1000
1001                         rt_rq->highest_prio.curr =
1002                                 sched_find_first_bit(array->bitmap);
1003                 }
1004
1005         } else
1006                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1007
1008         dec_rt_prio_smp(rt_rq, prio, prev_prio);
1009 }
1010
1011 #else
1012
1013 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1014 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1015
1016 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1017
1018 #ifdef CONFIG_RT_GROUP_SCHED
1019
1020 static void
1021 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1022 {
1023         if (rt_se_boosted(rt_se))
1024                 rt_rq->rt_nr_boosted++;
1025
1026         if (rt_rq->tg)
1027                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1028 }
1029
1030 static void
1031 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1032 {
1033         if (rt_se_boosted(rt_se))
1034                 rt_rq->rt_nr_boosted--;
1035
1036         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1037 }
1038
1039 #else /* CONFIG_RT_GROUP_SCHED */
1040
1041 static void
1042 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1043 {
1044         start_rt_bandwidth(&def_rt_bandwidth);
1045 }
1046
1047 static inline
1048 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1049
1050 #endif /* CONFIG_RT_GROUP_SCHED */
1051
1052 static inline
1053 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1054 {
1055         int prio = rt_se_prio(rt_se);
1056
1057         WARN_ON(!rt_prio(prio));
1058         rt_rq->rt_nr_running++;
1059
1060         inc_rt_prio(rt_rq, prio);
1061         inc_rt_migration(rt_se, rt_rq);
1062         inc_rt_group(rt_se, rt_rq);
1063 }
1064
1065 static inline
1066 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1067 {
1068         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1069         WARN_ON(!rt_rq->rt_nr_running);
1070         rt_rq->rt_nr_running--;
1071
1072         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1073         dec_rt_migration(rt_se, rt_rq);
1074         dec_rt_group(rt_se, rt_rq);
1075 }
1076
1077 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1078 {
1079         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1080         struct rt_prio_array *array = &rt_rq->active;
1081         struct rt_rq *group_rq = group_rt_rq(rt_se);
1082         struct list_head *queue = array->queue + rt_se_prio(rt_se);
1083
1084         /*
1085          * Don't enqueue the group if its throttled, or when empty.
1086          * The latter is a consequence of the former when a child group
1087          * get throttled and the current group doesn't have any other
1088          * active members.
1089          */
1090         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
1091                 return;
1092
1093         if (!rt_rq->rt_nr_running)
1094                 list_add_leaf_rt_rq(rt_rq);
1095
1096         if (head)
1097                 list_add(&rt_se->run_list, queue);
1098         else
1099                 list_add_tail(&rt_se->run_list, queue);
1100         __set_bit(rt_se_prio(rt_se), array->bitmap);
1101
1102         inc_rt_tasks(rt_se, rt_rq);
1103 }
1104
1105 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
1106 {
1107         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1108         struct rt_prio_array *array = &rt_rq->active;
1109
1110         list_del_init(&rt_se->run_list);
1111         if (list_empty(array->queue + rt_se_prio(rt_se)))
1112                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1113
1114         dec_rt_tasks(rt_se, rt_rq);
1115         if (!rt_rq->rt_nr_running)
1116                 list_del_leaf_rt_rq(rt_rq);
1117 }
1118
1119 /*
1120  * Because the prio of an upper entry depends on the lower
1121  * entries, we must remove entries top - down.
1122  */
1123 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
1124 {
1125         struct sched_rt_entity *back = NULL;
1126
1127         for_each_sched_rt_entity(rt_se) {
1128                 rt_se->back = back;
1129                 back = rt_se;
1130         }
1131
1132         for (rt_se = back; rt_se; rt_se = rt_se->back) {
1133                 if (on_rt_rq(rt_se))
1134                         __dequeue_rt_entity(rt_se);
1135         }
1136 }
1137
1138 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1139 {
1140         dequeue_rt_stack(rt_se);
1141         for_each_sched_rt_entity(rt_se)
1142                 __enqueue_rt_entity(rt_se, head);
1143 }
1144
1145 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1146 {
1147         dequeue_rt_stack(rt_se);
1148
1149         for_each_sched_rt_entity(rt_se) {
1150                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1151
1152                 if (rt_rq && rt_rq->rt_nr_running)
1153                         __enqueue_rt_entity(rt_se, false);
1154         }
1155 }
1156
1157 /*
1158  * Adding/removing a task to/from a priority array:
1159  */
1160 static void
1161 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1162 {
1163         struct sched_rt_entity *rt_se = &p->rt;
1164
1165         if (flags & ENQUEUE_WAKEUP)
1166                 rt_se->timeout = 0;
1167
1168         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
1169
1170         if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1171                 enqueue_pushable_task(rq, p);
1172
1173         inc_nr_running(rq);
1174 }
1175
1176 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1177 {
1178         struct sched_rt_entity *rt_se = &p->rt;
1179
1180         update_curr_rt(rq);
1181         dequeue_rt_entity(rt_se);
1182
1183         dequeue_pushable_task(rq, p);
1184
1185         dec_nr_running(rq);
1186 }
1187
1188 /*
1189  * Put task to the head or the end of the run list without the overhead of
1190  * dequeue followed by enqueue.
1191  */
1192 static void
1193 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1194 {
1195         if (on_rt_rq(rt_se)) {
1196                 struct rt_prio_array *array = &rt_rq->active;
1197                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1198
1199                 if (head)
1200                         list_move(&rt_se->run_list, queue);
1201                 else
1202                         list_move_tail(&rt_se->run_list, queue);
1203         }
1204 }
1205
1206 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1207 {
1208         struct sched_rt_entity *rt_se = &p->rt;
1209         struct rt_rq *rt_rq;
1210
1211         for_each_sched_rt_entity(rt_se) {
1212                 rt_rq = rt_rq_of_se(rt_se);
1213                 requeue_rt_entity(rt_rq, rt_se, head);
1214         }
1215 }
1216
1217 static void yield_task_rt(struct rq *rq)
1218 {
1219         requeue_task_rt(rq, rq->curr, 0);
1220 }
1221
1222 #ifdef CONFIG_SMP
1223 static int find_lowest_rq(struct task_struct *task);
1224
1225 static int
1226 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
1227 {
1228         struct task_struct *curr;
1229         struct rq *rq;
1230         int cpu;
1231
1232         cpu = task_cpu(p);
1233
1234         if (p->nr_cpus_allowed == 1)
1235                 goto out;
1236
1237         /* For anything but wake ups, just return the task_cpu */
1238         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1239                 goto out;
1240
1241         rq = cpu_rq(cpu);
1242
1243         rcu_read_lock();
1244         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1245
1246         /*
1247          * If the current task on @p's runqueue is an RT task, then
1248          * try to see if we can wake this RT task up on another
1249          * runqueue. Otherwise simply start this RT task
1250          * on its current runqueue.
1251          *
1252          * We want to avoid overloading runqueues. If the woken
1253          * task is a higher priority, then it will stay on this CPU
1254          * and the lower prio task should be moved to another CPU.
1255          * Even though this will probably make the lower prio task
1256          * lose its cache, we do not want to bounce a higher task
1257          * around just because it gave up its CPU, perhaps for a
1258          * lock?
1259          *
1260          * For equal prio tasks, we just let the scheduler sort it out.
1261          *
1262          * Otherwise, just let it ride on the affined RQ and the
1263          * post-schedule router will push the preempted task away
1264          *
1265          * This test is optimistic, if we get it wrong the load-balancer
1266          * will have to sort it out.
1267          */
1268         if (curr && unlikely(rt_task(curr)) &&
1269             (curr->nr_cpus_allowed < 2 ||
1270              curr->prio <= p->prio) &&
1271             (p->nr_cpus_allowed > 1)) {
1272                 int target = find_lowest_rq(p);
1273
1274                 if (target != -1)
1275                         cpu = target;
1276         }
1277         rcu_read_unlock();
1278
1279 out:
1280         return cpu;
1281 }
1282
1283 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1284 {
1285         if (rq->curr->nr_cpus_allowed == 1)
1286                 return;
1287
1288         if (p->nr_cpus_allowed != 1
1289             && cpupri_find(&rq->rd->cpupri, p, NULL))
1290                 return;
1291
1292         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1293                 return;
1294
1295         /*
1296          * There appears to be other cpus that can accept
1297          * current and none to run 'p', so lets reschedule
1298          * to try and push current away:
1299          */
1300         requeue_task_rt(rq, p, 1);
1301         resched_task(rq->curr);
1302 }
1303
1304 #endif /* CONFIG_SMP */
1305
1306 /*
1307  * Preempt the current task with a newly woken task if needed:
1308  */
1309 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1310 {
1311         if (p->prio < rq->curr->prio) {
1312                 resched_task(rq->curr);
1313                 return;
1314         }
1315
1316 #ifdef CONFIG_SMP
1317         /*
1318          * If:
1319          *
1320          * - the newly woken task is of equal priority to the current task
1321          * - the newly woken task is non-migratable while current is migratable
1322          * - current will be preempted on the next reschedule
1323          *
1324          * we should check to see if current can readily move to a different
1325          * cpu.  If so, we will reschedule to allow the push logic to try
1326          * to move current somewhere else, making room for our non-migratable
1327          * task.
1328          */
1329         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1330                 check_preempt_equal_prio(rq, p);
1331 #endif
1332 }
1333
1334 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1335                                                    struct rt_rq *rt_rq)
1336 {
1337         struct rt_prio_array *array = &rt_rq->active;
1338         struct sched_rt_entity *next = NULL;
1339         struct list_head *queue;
1340         int idx;
1341
1342         idx = sched_find_first_bit(array->bitmap);
1343         BUG_ON(idx >= MAX_RT_PRIO);
1344
1345         queue = array->queue + idx;
1346         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1347
1348         return next;
1349 }
1350
1351 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1352 {
1353         struct sched_rt_entity *rt_se;
1354         struct task_struct *p;
1355         struct rt_rq *rt_rq;
1356
1357         rt_rq = &rq->rt;
1358
1359         if (!rt_rq->rt_nr_running)
1360                 return NULL;
1361
1362         if (rt_rq_throttled(rt_rq))
1363                 return NULL;
1364
1365         do {
1366                 rt_se = pick_next_rt_entity(rq, rt_rq);
1367                 BUG_ON(!rt_se);
1368                 rt_rq = group_rt_rq(rt_se);
1369         } while (rt_rq);
1370
1371         p = rt_task_of(rt_se);
1372         p->se.exec_start = rq->clock_task;
1373
1374         return p;
1375 }
1376
1377 static struct task_struct *pick_next_task_rt(struct rq *rq)
1378 {
1379         struct task_struct *p = _pick_next_task_rt(rq);
1380
1381         /* The running task is never eligible for pushing */
1382         if (p)
1383                 dequeue_pushable_task(rq, p);
1384
1385 #ifdef CONFIG_SMP
1386         /*
1387          * We detect this state here so that we can avoid taking the RQ
1388          * lock again later if there is no need to push
1389          */
1390         rq->post_schedule = has_pushable_tasks(rq);
1391 #endif
1392
1393         return p;
1394 }
1395
1396 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1397 {
1398         update_curr_rt(rq);
1399
1400         /*
1401          * The previous task needs to be made eligible for pushing
1402          * if it is still active
1403          */
1404         if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1405                 enqueue_pushable_task(rq, p);
1406 }
1407
1408 #ifdef CONFIG_SMP
1409
1410 /* Only try algorithms three times */
1411 #define RT_MAX_TRIES 3
1412
1413 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1414 {
1415         if (!task_running(rq, p) &&
1416             (cpu < 0 || cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) &&
1417             (p->nr_cpus_allowed > 1))
1418                 return 1;
1419         return 0;
1420 }
1421
1422 /* Return the second highest RT task, NULL otherwise */
1423 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1424 {
1425         struct task_struct *next = NULL;
1426         struct sched_rt_entity *rt_se;
1427         struct rt_prio_array *array;
1428         struct rt_rq *rt_rq;
1429         int idx;
1430
1431         for_each_leaf_rt_rq(rt_rq, rq) {
1432                 array = &rt_rq->active;
1433                 idx = sched_find_first_bit(array->bitmap);
1434 next_idx:
1435                 if (idx >= MAX_RT_PRIO)
1436                         continue;
1437                 if (next && next->prio <= idx)
1438                         continue;
1439                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1440                         struct task_struct *p;
1441
1442                         if (!rt_entity_is_task(rt_se))
1443                                 continue;
1444
1445                         p = rt_task_of(rt_se);
1446                         if (pick_rt_task(rq, p, cpu)) {
1447                                 next = p;
1448                                 break;
1449                         }
1450                 }
1451                 if (!next) {
1452                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1453                         goto next_idx;
1454                 }
1455         }
1456
1457         return next;
1458 }
1459
1460 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1461
1462 static int find_lowest_rq(struct task_struct *task)
1463 {
1464         struct sched_domain *sd;
1465         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1466         int this_cpu = smp_processor_id();
1467         int cpu      = task_cpu(task);
1468
1469         /* Make sure the mask is initialized first */
1470         if (unlikely(!lowest_mask))
1471                 return -1;
1472
1473         if (task->nr_cpus_allowed == 1)
1474                 return -1; /* No other targets possible */
1475
1476         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1477                 return -1; /* No targets found */
1478
1479         /*
1480          * At this point we have built a mask of cpus representing the
1481          * lowest priority tasks in the system.  Now we want to elect
1482          * the best one based on our affinity and topology.
1483          *
1484          * We prioritize the last cpu that the task executed on since
1485          * it is most likely cache-hot in that location.
1486          */
1487         if (cpumask_test_cpu(cpu, lowest_mask))
1488                 return cpu;
1489
1490         /*
1491          * Otherwise, we consult the sched_domains span maps to figure
1492          * out which cpu is logically closest to our hot cache data.
1493          */
1494         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1495                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1496
1497         rcu_read_lock();
1498         for_each_domain(cpu, sd) {
1499                 if (sd->flags & SD_WAKE_AFFINE) {
1500                         int best_cpu;
1501
1502                         /*
1503                          * "this_cpu" is cheaper to preempt than a
1504                          * remote processor.
1505                          */
1506                         if (this_cpu != -1 &&
1507                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1508                                 rcu_read_unlock();
1509                                 return this_cpu;
1510                         }
1511
1512                         best_cpu = cpumask_first_and(lowest_mask,
1513                                                      sched_domain_span(sd));
1514                         if (best_cpu < nr_cpu_ids) {
1515                                 rcu_read_unlock();
1516                                 return best_cpu;
1517                         }
1518                 }
1519         }
1520         rcu_read_unlock();
1521
1522         /*
1523          * And finally, if there were no matches within the domains
1524          * just give the caller *something* to work with from the compatible
1525          * locations.
1526          */
1527         if (this_cpu != -1)
1528                 return this_cpu;
1529
1530         cpu = cpumask_any(lowest_mask);
1531         if (cpu < nr_cpu_ids)
1532                 return cpu;
1533         return -1;
1534 }
1535
1536 /* Will lock the rq it finds */
1537 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1538 {
1539         struct rq *lowest_rq = NULL;
1540         int tries;
1541         int cpu;
1542
1543         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1544                 cpu = find_lowest_rq(task);
1545
1546                 if ((cpu == -1) || (cpu == rq->cpu))
1547                         break;
1548
1549                 lowest_rq = cpu_rq(cpu);
1550
1551                 /* if the prio of this runqueue changed, try again */
1552                 if (double_lock_balance(rq, lowest_rq)) {
1553                         /*
1554                          * We had to unlock the run queue. In
1555                          * the mean time, task could have
1556                          * migrated already or had its affinity changed.
1557                          * Also make sure that it wasn't scheduled on its rq.
1558                          */
1559                         if (unlikely(task_rq(task) != rq ||
1560                                      !cpumask_test_cpu(lowest_rq->cpu,
1561                                                        tsk_cpus_allowed(task)) ||
1562                                      task_running(rq, task) ||
1563                                      !task->on_rq)) {
1564
1565                                 raw_spin_unlock(&lowest_rq->lock);
1566                                 lowest_rq = NULL;
1567                                 break;
1568                         }
1569                 }
1570
1571                 /* If this rq is still suitable use it. */
1572                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1573                         break;
1574
1575                 /* try again */
1576                 double_unlock_balance(rq, lowest_rq);
1577                 lowest_rq = NULL;
1578         }
1579
1580         return lowest_rq;
1581 }
1582
1583 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1584 {
1585         struct task_struct *p;
1586
1587         if (!has_pushable_tasks(rq))
1588                 return NULL;
1589
1590         p = plist_first_entry(&rq->rt.pushable_tasks,
1591                               struct task_struct, pushable_tasks);
1592
1593         BUG_ON(rq->cpu != task_cpu(p));
1594         BUG_ON(task_current(rq, p));
1595         BUG_ON(p->nr_cpus_allowed <= 1);
1596
1597         BUG_ON(!p->on_rq);
1598         BUG_ON(!rt_task(p));
1599
1600         return p;
1601 }
1602
1603 /*
1604  * If the current CPU has more than one RT task, see if the non
1605  * running task can migrate over to a CPU that is running a task
1606  * of lesser priority.
1607  */
1608 static int push_rt_task(struct rq *rq)
1609 {
1610         struct task_struct *next_task;
1611         struct rq *lowest_rq;
1612         int ret = 0;
1613
1614         if (!rq->rt.overloaded)
1615                 return 0;
1616
1617         next_task = pick_next_pushable_task(rq);
1618         if (!next_task)
1619                 return 0;
1620
1621 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1622        if (unlikely(task_running(rq, next_task)))
1623                return 0;
1624 #endif
1625
1626 retry:
1627         if (unlikely(next_task == rq->curr)) {
1628                 WARN_ON(1);
1629                 return 0;
1630         }
1631
1632         /*
1633          * It's possible that the next_task slipped in of
1634          * higher priority than current. If that's the case
1635          * just reschedule current.
1636          */
1637         if (unlikely(next_task->prio < rq->curr->prio)) {
1638                 resched_task(rq->curr);
1639                 return 0;
1640         }
1641
1642         /* We might release rq lock */
1643         get_task_struct(next_task);
1644
1645         /* find_lock_lowest_rq locks the rq if found */
1646         lowest_rq = find_lock_lowest_rq(next_task, rq);
1647         if (!lowest_rq) {
1648                 struct task_struct *task;
1649                 /*
1650                  * find_lock_lowest_rq releases rq->lock
1651                  * so it is possible that next_task has migrated.
1652                  *
1653                  * We need to make sure that the task is still on the same
1654                  * run-queue and is also still the next task eligible for
1655                  * pushing.
1656                  */
1657                 task = pick_next_pushable_task(rq);
1658                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1659                         /*
1660                          * The task hasn't migrated, and is still the next
1661                          * eligible task, but we failed to find a run-queue
1662                          * to push it to.  Do not retry in this case, since
1663                          * other cpus will pull from us when ready.
1664                          */
1665                         goto out;
1666                 }
1667
1668                 if (!task)
1669                         /* No more tasks, just exit */
1670                         goto out;
1671
1672                 /*
1673                  * Something has shifted, try again.
1674                  */
1675                 put_task_struct(next_task);
1676                 next_task = task;
1677                 goto retry;
1678         }
1679
1680         deactivate_task(rq, next_task, 0);
1681         set_task_cpu(next_task, lowest_rq->cpu);
1682         activate_task(lowest_rq, next_task, 0);
1683         ret = 1;
1684
1685         resched_task(lowest_rq->curr);
1686
1687         double_unlock_balance(rq, lowest_rq);
1688
1689 out:
1690         put_task_struct(next_task);
1691
1692         return ret;
1693 }
1694
1695 static void push_rt_tasks(struct rq *rq)
1696 {
1697         /* push_rt_task will return true if it moved an RT */
1698         while (push_rt_task(rq))
1699                 ;
1700 }
1701
1702 static int pull_rt_task(struct rq *this_rq)
1703 {
1704         int this_cpu = this_rq->cpu, ret = 0, cpu;
1705         struct task_struct *p;
1706         struct rq *src_rq;
1707
1708         if (likely(!rt_overloaded(this_rq)))
1709                 return 0;
1710
1711         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1712                 if (this_cpu == cpu)
1713                         continue;
1714
1715                 src_rq = cpu_rq(cpu);
1716
1717                 /*
1718                  * Don't bother taking the src_rq->lock if the next highest
1719                  * task is known to be lower-priority than our current task.
1720                  * This may look racy, but if this value is about to go
1721                  * logically higher, the src_rq will push this task away.
1722                  * And if its going logically lower, we do not care
1723                  */
1724                 if (src_rq->rt.highest_prio.next >=
1725                     this_rq->rt.highest_prio.curr)
1726                         continue;
1727
1728                 /*
1729                  * We can potentially drop this_rq's lock in
1730                  * double_lock_balance, and another CPU could
1731                  * alter this_rq
1732                  */
1733                 double_lock_balance(this_rq, src_rq);
1734
1735                 /*
1736                  * Are there still pullable RT tasks?
1737                  */
1738                 if (src_rq->rt.rt_nr_running <= 1)
1739                         goto skip;
1740
1741                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1742
1743                 /*
1744                  * Do we have an RT task that preempts
1745                  * the to-be-scheduled task?
1746                  */
1747                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1748                         WARN_ON(p == src_rq->curr);
1749                         WARN_ON(!p->on_rq);
1750
1751                         /*
1752                          * There's a chance that p is higher in priority
1753                          * than what's currently running on its cpu.
1754                          * This is just that p is wakeing up and hasn't
1755                          * had a chance to schedule. We only pull
1756                          * p if it is lower in priority than the
1757                          * current task on the run queue
1758                          */
1759                         if (p->prio < src_rq->curr->prio)
1760                                 goto skip;
1761
1762                         ret = 1;
1763
1764                         deactivate_task(src_rq, p, 0);
1765                         set_task_cpu(p, this_cpu);
1766                         activate_task(this_rq, p, 0);
1767                         /*
1768                          * We continue with the search, just in
1769                          * case there's an even higher prio task
1770                          * in another runqueue. (low likelihood
1771                          * but possible)
1772                          */
1773                 }
1774 skip:
1775                 double_unlock_balance(this_rq, src_rq);
1776         }
1777
1778         return ret;
1779 }
1780
1781 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1782 {
1783         /* Try to pull RT tasks here if we lower this rq's prio */
1784         if (rq->rt.highest_prio.curr > prev->prio)
1785                 pull_rt_task(rq);
1786 }
1787
1788 static void post_schedule_rt(struct rq *rq)
1789 {
1790         push_rt_tasks(rq);
1791 }
1792
1793 /*
1794  * If we are not running and we are not going to reschedule soon, we should
1795  * try to push tasks away now
1796  */
1797 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1798 {
1799         if (!task_running(rq, p) &&
1800             !test_tsk_need_resched(rq->curr) &&
1801             has_pushable_tasks(rq) &&
1802             p->nr_cpus_allowed > 1 &&
1803             rt_task(rq->curr) &&
1804             (rq->curr->nr_cpus_allowed < 2 ||
1805              rq->curr->prio <= p->prio))
1806                 push_rt_tasks(rq);
1807 }
1808
1809 static void set_cpus_allowed_rt(struct task_struct *p,
1810                                 const struct cpumask *new_mask)
1811 {
1812         struct rq *rq;
1813         int weight;
1814
1815         BUG_ON(!rt_task(p));
1816
1817         if (!p->on_rq)
1818                 return;
1819
1820         weight = cpumask_weight(new_mask);
1821
1822         /*
1823          * Only update if the process changes its state from whether it
1824          * can migrate or not.
1825          */
1826         if ((p->nr_cpus_allowed > 1) == (weight > 1))
1827                 return;
1828
1829         rq = task_rq(p);
1830
1831         /*
1832          * The process used to be able to migrate OR it can now migrate
1833          */
1834         if (weight <= 1) {
1835                 if (!task_current(rq, p))
1836                         dequeue_pushable_task(rq, p);
1837                 BUG_ON(!rq->rt.rt_nr_migratory);
1838                 rq->rt.rt_nr_migratory--;
1839         } else {
1840                 if (!task_current(rq, p))
1841                         enqueue_pushable_task(rq, p);
1842                 rq->rt.rt_nr_migratory++;
1843         }
1844
1845         update_rt_migration(&rq->rt);
1846 }
1847
1848 /* Assumes rq->lock is held */
1849 static void rq_online_rt(struct rq *rq)
1850 {
1851         if (rq->rt.overloaded)
1852                 rt_set_overload(rq);
1853
1854         __enable_runtime(rq);
1855
1856         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1857 }
1858
1859 /* Assumes rq->lock is held */
1860 static void rq_offline_rt(struct rq *rq)
1861 {
1862         if (rq->rt.overloaded)
1863                 rt_clear_overload(rq);
1864
1865         __disable_runtime(rq);
1866
1867         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1868 }
1869
1870 /*
1871  * When switch from the rt queue, we bring ourselves to a position
1872  * that we might want to pull RT tasks from other runqueues.
1873  */
1874 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1875 {
1876         /*
1877          * If there are other RT tasks then we will reschedule
1878          * and the scheduling of the other RT tasks will handle
1879          * the balancing. But if we are the last RT task
1880          * we may need to handle the pulling of RT tasks
1881          * now.
1882          */
1883         if (p->on_rq && !rq->rt.rt_nr_running)
1884                 pull_rt_task(rq);
1885 }
1886
1887 void init_sched_rt_class(void)
1888 {
1889         unsigned int i;
1890
1891         for_each_possible_cpu(i) {
1892                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1893                                         GFP_KERNEL, cpu_to_node(i));
1894         }
1895 }
1896 #endif /* CONFIG_SMP */
1897
1898 /*
1899  * When switching a task to RT, we may overload the runqueue
1900  * with RT tasks. In this case we try to push them off to
1901  * other runqueues.
1902  */
1903 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1904 {
1905         int check_resched = 1;
1906
1907         /*
1908          * If we are already running, then there's nothing
1909          * that needs to be done. But if we are not running
1910          * we may need to preempt the current running task.
1911          * If that current running task is also an RT task
1912          * then see if we can move to another run queue.
1913          */
1914         if (p->on_rq && rq->curr != p) {
1915 #ifdef CONFIG_SMP
1916                 if (rq->rt.overloaded && push_rt_task(rq) &&
1917                     /* Don't resched if we changed runqueues */
1918                     rq != task_rq(p))
1919                         check_resched = 0;
1920 #endif /* CONFIG_SMP */
1921                 if (check_resched && p->prio < rq->curr->prio)
1922                         resched_task(rq->curr);
1923         }
1924 }
1925
1926 /*
1927  * Priority of the task has changed. This may cause
1928  * us to initiate a push or pull.
1929  */
1930 static void
1931 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1932 {
1933         if (!p->on_rq)
1934                 return;
1935
1936         if (rq->curr == p) {
1937 #ifdef CONFIG_SMP
1938                 /*
1939                  * If our priority decreases while running, we
1940                  * may need to pull tasks to this runqueue.
1941                  */
1942                 if (oldprio < p->prio)
1943                         pull_rt_task(rq);
1944                 /*
1945                  * If there's a higher priority task waiting to run
1946                  * then reschedule. Note, the above pull_rt_task
1947                  * can release the rq lock and p could migrate.
1948                  * Only reschedule if p is still on the same runqueue.
1949                  */
1950                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1951                         resched_task(p);
1952 #else
1953                 /* For UP simply resched on drop of prio */
1954                 if (oldprio < p->prio)
1955                         resched_task(p);
1956 #endif /* CONFIG_SMP */
1957         } else {
1958                 /*
1959                  * This task is not running, but if it is
1960                  * greater than the current running task
1961                  * then reschedule.
1962                  */
1963                 if (p->prio < rq->curr->prio)
1964                         resched_task(rq->curr);
1965         }
1966 }
1967
1968 static void watchdog(struct rq *rq, struct task_struct *p)
1969 {
1970         unsigned long soft, hard;
1971
1972         /* max may change after cur was read, this will be fixed next tick */
1973         soft = task_rlimit(p, RLIMIT_RTTIME);
1974         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1975
1976         if (soft != RLIM_INFINITY) {
1977                 unsigned long next;
1978
1979                 p->rt.timeout++;
1980                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1981                 if (p->rt.timeout > next)
1982                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1983         }
1984 }
1985
1986 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1987 {
1988         struct sched_rt_entity *rt_se = &p->rt;
1989
1990         update_curr_rt(rq);
1991
1992         watchdog(rq, p);
1993
1994         /*
1995          * RR tasks need a special form of timeslice management.
1996          * FIFO tasks have no timeslices.
1997          */
1998         if (p->policy != SCHED_RR)
1999                 return;
2000
2001         if (--p->rt.time_slice)
2002                 return;
2003
2004         p->rt.time_slice = RR_TIMESLICE;
2005
2006         /*
2007          * Requeue to the end of queue if we (and all of our ancestors) are the
2008          * only element on the queue
2009          */
2010         for_each_sched_rt_entity(rt_se) {
2011                 if (rt_se->run_list.prev != rt_se->run_list.next) {
2012                         requeue_task_rt(rq, p, 0);
2013                         set_tsk_need_resched(p);
2014                         return;
2015                 }
2016         }
2017 }
2018
2019 static void set_curr_task_rt(struct rq *rq)
2020 {
2021         struct task_struct *p = rq->curr;
2022
2023         p->se.exec_start = rq->clock_task;
2024
2025         /* The running task is never eligible for pushing */
2026         dequeue_pushable_task(rq, p);
2027 }
2028
2029 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2030 {
2031         /*
2032          * Time slice is 0 for SCHED_FIFO tasks
2033          */
2034         if (task->policy == SCHED_RR)
2035                 return RR_TIMESLICE;
2036         else
2037                 return 0;
2038 }
2039
2040 const struct sched_class rt_sched_class = {
2041         .next                   = &fair_sched_class,
2042         .enqueue_task           = enqueue_task_rt,
2043         .dequeue_task           = dequeue_task_rt,
2044         .yield_task             = yield_task_rt,
2045
2046         .check_preempt_curr     = check_preempt_curr_rt,
2047
2048         .pick_next_task         = pick_next_task_rt,
2049         .put_prev_task          = put_prev_task_rt,
2050
2051 #ifdef CONFIG_SMP
2052         .select_task_rq         = select_task_rq_rt,
2053
2054         .set_cpus_allowed       = set_cpus_allowed_rt,
2055         .rq_online              = rq_online_rt,
2056         .rq_offline             = rq_offline_rt,
2057         .pre_schedule           = pre_schedule_rt,
2058         .post_schedule          = post_schedule_rt,
2059         .task_woken             = task_woken_rt,
2060         .switched_from          = switched_from_rt,
2061 #endif
2062
2063         .set_curr_task          = set_curr_task_rt,
2064         .task_tick              = task_tick_rt,
2065
2066         .get_rr_interval        = get_rr_interval_rt,
2067
2068         .prio_changed           = prio_changed_rt,
2069         .switched_to            = switched_to_rt,
2070 };
2071
2072 #ifdef CONFIG_SCHED_DEBUG
2073 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2074
2075 void print_rt_stats(struct seq_file *m, int cpu)
2076 {
2077         rt_rq_iter_t iter;
2078         struct rt_rq *rt_rq;
2079
2080         rcu_read_lock();
2081         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2082                 print_rt_rq(m, cpu, rt_rq);
2083         rcu_read_unlock();
2084 }
2085 #endif /* CONFIG_SCHED_DEBUG */