575da76a3874a8c1b2ddd0f518e5ecea7a805262
[projects/modsched/linux.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 #include <linux/irq_work.h>
10
11 int sched_rr_timeslice = RR_TIMESLICE;
12
13 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
14
15 struct rt_bandwidth def_rt_bandwidth;
16
17 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
18 {
19         struct rt_bandwidth *rt_b =
20                 container_of(timer, struct rt_bandwidth, rt_period_timer);
21         ktime_t now;
22         int overrun;
23         int idle = 0;
24
25         for (;;) {
26                 now = hrtimer_cb_get_time(timer);
27                 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
28
29                 if (!overrun)
30                         break;
31
32                 idle = do_sched_rt_period_timer(rt_b, overrun);
33         }
34
35         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
36 }
37
38 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
39 {
40         rt_b->rt_period = ns_to_ktime(period);
41         rt_b->rt_runtime = runtime;
42
43         raw_spin_lock_init(&rt_b->rt_runtime_lock);
44
45         hrtimer_init(&rt_b->rt_period_timer,
46                         CLOCK_MONOTONIC, HRTIMER_MODE_REL);
47         rt_b->rt_period_timer.function = sched_rt_period_timer;
48 }
49
50 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
51 {
52         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
53                 return;
54
55         if (hrtimer_active(&rt_b->rt_period_timer))
56                 return;
57
58         raw_spin_lock(&rt_b->rt_runtime_lock);
59         start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
60         raw_spin_unlock(&rt_b->rt_runtime_lock);
61 }
62
63 #ifdef CONFIG_SMP
64 static void push_irq_work_func(struct irq_work *work);
65 #endif
66
67 void init_rt_rq(struct rt_rq *rt_rq)
68 {
69         struct rt_prio_array *array;
70         int i;
71
72         array = &rt_rq->active;
73         for (i = 0; i < MAX_RT_PRIO; i++) {
74                 INIT_LIST_HEAD(array->queue + i);
75                 __clear_bit(i, array->bitmap);
76         }
77         /* delimiter for bitsearch: */
78         __set_bit(MAX_RT_PRIO, array->bitmap);
79
80 #if defined CONFIG_SMP
81         rt_rq->highest_prio.curr = MAX_RT_PRIO;
82         rt_rq->highest_prio.next = MAX_RT_PRIO;
83         rt_rq->rt_nr_migratory = 0;
84         rt_rq->overloaded = 0;
85         plist_head_init(&rt_rq->pushable_tasks);
86
87 #ifdef HAVE_RT_PUSH_IPI
88         rt_rq->push_flags = 0;
89         rt_rq->push_cpu = nr_cpu_ids;
90         raw_spin_lock_init(&rt_rq->push_lock);
91         init_irq_work(&rt_rq->push_work, push_irq_work_func);
92 #endif
93 #endif /* CONFIG_SMP */
94         /* We start is dequeued state, because no RT tasks are queued */
95         rt_rq->rt_queued = 0;
96
97         rt_rq->rt_time = 0;
98         rt_rq->rt_throttled = 0;
99         rt_rq->rt_runtime = 0;
100         raw_spin_lock_init(&rt_rq->rt_runtime_lock);
101 }
102
103 #ifdef CONFIG_RT_GROUP_SCHED
104 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
105 {
106         hrtimer_cancel(&rt_b->rt_period_timer);
107 }
108
109 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
110
111 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
112 {
113 #ifdef CONFIG_SCHED_DEBUG
114         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
115 #endif
116         return container_of(rt_se, struct task_struct, rt);
117 }
118
119 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
120 {
121         return rt_rq->rq;
122 }
123
124 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
125 {
126         return rt_se->rt_rq;
127 }
128
129 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
130 {
131         struct rt_rq *rt_rq = rt_se->rt_rq;
132
133         return rt_rq->rq;
134 }
135
136 void free_rt_sched_group(struct task_group *tg)
137 {
138         int i;
139
140         if (tg->rt_se)
141                 destroy_rt_bandwidth(&tg->rt_bandwidth);
142
143         for_each_possible_cpu(i) {
144                 if (tg->rt_rq)
145                         kfree(tg->rt_rq[i]);
146                 if (tg->rt_se)
147                         kfree(tg->rt_se[i]);
148         }
149
150         kfree(tg->rt_rq);
151         kfree(tg->rt_se);
152 }
153
154 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
155                 struct sched_rt_entity *rt_se, int cpu,
156                 struct sched_rt_entity *parent)
157 {
158         struct rq *rq = cpu_rq(cpu);
159
160         rt_rq->highest_prio.curr = MAX_RT_PRIO;
161         rt_rq->rt_nr_boosted = 0;
162         rt_rq->rq = rq;
163         rt_rq->tg = tg;
164
165         tg->rt_rq[cpu] = rt_rq;
166         tg->rt_se[cpu] = rt_se;
167
168         if (!rt_se)
169                 return;
170
171         if (!parent)
172                 rt_se->rt_rq = &rq->rt;
173         else
174                 rt_se->rt_rq = parent->my_q;
175
176         rt_se->my_q = rt_rq;
177         rt_se->parent = parent;
178         INIT_LIST_HEAD(&rt_se->run_list);
179 }
180
181 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
182 {
183         struct rt_rq *rt_rq;
184         struct sched_rt_entity *rt_se;
185         int i;
186
187         tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
188         if (!tg->rt_rq)
189                 goto err;
190         tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
191         if (!tg->rt_se)
192                 goto err;
193
194         init_rt_bandwidth(&tg->rt_bandwidth,
195                         ktime_to_ns(def_rt_bandwidth.rt_period), 0);
196
197         for_each_possible_cpu(i) {
198                 rt_rq = kzalloc_node(sizeof(struct rt_rq),
199                                      GFP_KERNEL, cpu_to_node(i));
200                 if (!rt_rq)
201                         goto err;
202
203                 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
204                                      GFP_KERNEL, cpu_to_node(i));
205                 if (!rt_se)
206                         goto err_free_rq;
207
208                 init_rt_rq(rt_rq);
209                 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
210                 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
211         }
212
213         return 1;
214
215 err_free_rq:
216         kfree(rt_rq);
217 err:
218         return 0;
219 }
220
221 #else /* CONFIG_RT_GROUP_SCHED */
222
223 #define rt_entity_is_task(rt_se) (1)
224
225 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
226 {
227         return container_of(rt_se, struct task_struct, rt);
228 }
229
230 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
231 {
232         return container_of(rt_rq, struct rq, rt);
233 }
234
235 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
236 {
237         struct task_struct *p = rt_task_of(rt_se);
238
239         return task_rq(p);
240 }
241
242 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
243 {
244         struct rq *rq = rq_of_rt_se(rt_se);
245
246         return &rq->rt;
247 }
248
249 void free_rt_sched_group(struct task_group *tg) { }
250
251 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
252 {
253         return 1;
254 }
255 #endif /* CONFIG_RT_GROUP_SCHED */
256
257 #ifdef CONFIG_SMP
258
259 static int pull_rt_task(struct rq *this_rq);
260
261 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
262 {
263         /* Try to pull RT tasks here if we lower this rq's prio */
264         return rq->rt.highest_prio.curr > prev->prio;
265 }
266
267 static inline int rt_overloaded(struct rq *rq)
268 {
269         return atomic_read(&rq->rd->rto_count);
270 }
271
272 static inline void rt_set_overload(struct rq *rq)
273 {
274         if (!rq->online)
275                 return;
276
277         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
278         /*
279          * Make sure the mask is visible before we set
280          * the overload count. That is checked to determine
281          * if we should look at the mask. It would be a shame
282          * if we looked at the mask, but the mask was not
283          * updated yet.
284          *
285          * Matched by the barrier in pull_rt_task().
286          */
287         smp_wmb();
288         atomic_inc(&rq->rd->rto_count);
289 }
290
291 static inline void rt_clear_overload(struct rq *rq)
292 {
293         if (!rq->online)
294                 return;
295
296         /* the order here really doesn't matter */
297         atomic_dec(&rq->rd->rto_count);
298         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
299 }
300
301 static void update_rt_migration(struct rt_rq *rt_rq)
302 {
303         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
304                 if (!rt_rq->overloaded) {
305                         rt_set_overload(rq_of_rt_rq(rt_rq));
306                         rt_rq->overloaded = 1;
307                 }
308         } else if (rt_rq->overloaded) {
309                 rt_clear_overload(rq_of_rt_rq(rt_rq));
310                 rt_rq->overloaded = 0;
311         }
312 }
313
314 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
315 {
316         struct task_struct *p;
317
318         if (!rt_entity_is_task(rt_se))
319                 return;
320
321         p = rt_task_of(rt_se);
322         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
323
324         rt_rq->rt_nr_total++;
325         if (p->nr_cpus_allowed > 1)
326                 rt_rq->rt_nr_migratory++;
327
328         update_rt_migration(rt_rq);
329 }
330
331 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
332 {
333         struct task_struct *p;
334
335         if (!rt_entity_is_task(rt_se))
336                 return;
337
338         p = rt_task_of(rt_se);
339         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
340
341         rt_rq->rt_nr_total--;
342         if (p->nr_cpus_allowed > 1)
343                 rt_rq->rt_nr_migratory--;
344
345         update_rt_migration(rt_rq);
346 }
347
348 static inline int has_pushable_tasks(struct rq *rq)
349 {
350         return !plist_head_empty(&rq->rt.pushable_tasks);
351 }
352
353 static inline void set_post_schedule(struct rq *rq)
354 {
355         /*
356          * We detect this state here so that we can avoid taking the RQ
357          * lock again later if there is no need to push
358          */
359         rq->post_schedule = has_pushable_tasks(rq);
360 }
361
362 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
363 {
364         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
365         plist_node_init(&p->pushable_tasks, p->prio);
366         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
367
368         /* Update the highest prio pushable task */
369         if (p->prio < rq->rt.highest_prio.next)
370                 rq->rt.highest_prio.next = p->prio;
371 }
372
373 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
374 {
375         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
376
377         /* Update the new highest prio pushable task */
378         if (has_pushable_tasks(rq)) {
379                 p = plist_first_entry(&rq->rt.pushable_tasks,
380                                       struct task_struct, pushable_tasks);
381                 rq->rt.highest_prio.next = p->prio;
382         } else
383                 rq->rt.highest_prio.next = MAX_RT_PRIO;
384 }
385
386 #else
387
388 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
389 {
390 }
391
392 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
393 {
394 }
395
396 static inline
397 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
398 {
399 }
400
401 static inline
402 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
403 {
404 }
405
406 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
407 {
408         return false;
409 }
410
411 static inline int pull_rt_task(struct rq *this_rq)
412 {
413         return 0;
414 }
415
416 static inline void set_post_schedule(struct rq *rq)
417 {
418 }
419 #endif /* CONFIG_SMP */
420
421 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
422 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
423
424 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
425 {
426         return !list_empty(&rt_se->run_list);
427 }
428
429 #ifdef CONFIG_RT_GROUP_SCHED
430
431 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
432 {
433         if (!rt_rq->tg)
434                 return RUNTIME_INF;
435
436         return rt_rq->rt_runtime;
437 }
438
439 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
440 {
441         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
442 }
443
444 typedef struct task_group *rt_rq_iter_t;
445
446 static inline struct task_group *next_task_group(struct task_group *tg)
447 {
448         do {
449                 tg = list_entry_rcu(tg->list.next,
450                         typeof(struct task_group), list);
451         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
452
453         if (&tg->list == &task_groups)
454                 tg = NULL;
455
456         return tg;
457 }
458
459 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
460         for (iter = container_of(&task_groups, typeof(*iter), list);    \
461                 (iter = next_task_group(iter)) &&                       \
462                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
463
464 #define for_each_sched_rt_entity(rt_se) \
465         for (; rt_se; rt_se = rt_se->parent)
466
467 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
468 {
469         return rt_se->my_q;
470 }
471
472 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
473 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
474
475 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
476 {
477         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
478         struct rq *rq = rq_of_rt_rq(rt_rq);
479         struct sched_rt_entity *rt_se;
480
481         int cpu = cpu_of(rq);
482
483         rt_se = rt_rq->tg->rt_se[cpu];
484
485         if (rt_rq->rt_nr_running) {
486                 if (!rt_se)
487                         enqueue_top_rt_rq(rt_rq);
488                 else if (!on_rt_rq(rt_se))
489                         enqueue_rt_entity(rt_se, false);
490
491                 if (rt_rq->highest_prio.curr < curr->prio)
492                         resched_curr(rq);
493         }
494 }
495
496 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
497 {
498         struct sched_rt_entity *rt_se;
499         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
500
501         rt_se = rt_rq->tg->rt_se[cpu];
502
503         if (!rt_se)
504                 dequeue_top_rt_rq(rt_rq);
505         else if (on_rt_rq(rt_se))
506                 dequeue_rt_entity(rt_se);
507 }
508
509 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
510 {
511         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
512 }
513
514 static int rt_se_boosted(struct sched_rt_entity *rt_se)
515 {
516         struct rt_rq *rt_rq = group_rt_rq(rt_se);
517         struct task_struct *p;
518
519         if (rt_rq)
520                 return !!rt_rq->rt_nr_boosted;
521
522         p = rt_task_of(rt_se);
523         return p->prio != p->normal_prio;
524 }
525
526 #ifdef CONFIG_SMP
527 static inline const struct cpumask *sched_rt_period_mask(void)
528 {
529         return this_rq()->rd->span;
530 }
531 #else
532 static inline const struct cpumask *sched_rt_period_mask(void)
533 {
534         return cpu_online_mask;
535 }
536 #endif
537
538 static inline
539 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
540 {
541         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
542 }
543
544 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
545 {
546         return &rt_rq->tg->rt_bandwidth;
547 }
548
549 #else /* !CONFIG_RT_GROUP_SCHED */
550
551 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
552 {
553         return rt_rq->rt_runtime;
554 }
555
556 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
557 {
558         return ktime_to_ns(def_rt_bandwidth.rt_period);
559 }
560
561 typedef struct rt_rq *rt_rq_iter_t;
562
563 #define for_each_rt_rq(rt_rq, iter, rq) \
564         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
565
566 #define for_each_sched_rt_entity(rt_se) \
567         for (; rt_se; rt_se = NULL)
568
569 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
570 {
571         return NULL;
572 }
573
574 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
575 {
576         struct rq *rq = rq_of_rt_rq(rt_rq);
577
578         if (!rt_rq->rt_nr_running)
579                 return;
580
581         enqueue_top_rt_rq(rt_rq);
582         resched_curr(rq);
583 }
584
585 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
586 {
587         dequeue_top_rt_rq(rt_rq);
588 }
589
590 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
591 {
592         return rt_rq->rt_throttled;
593 }
594
595 static inline const struct cpumask *sched_rt_period_mask(void)
596 {
597         return cpu_online_mask;
598 }
599
600 static inline
601 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
602 {
603         return &cpu_rq(cpu)->rt;
604 }
605
606 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
607 {
608         return &def_rt_bandwidth;
609 }
610
611 #endif /* CONFIG_RT_GROUP_SCHED */
612
613 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
614 {
615         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
616
617         return (hrtimer_active(&rt_b->rt_period_timer) ||
618                 rt_rq->rt_time < rt_b->rt_runtime);
619 }
620
621 #ifdef CONFIG_SMP
622 /*
623  * We ran out of runtime, see if we can borrow some from our neighbours.
624  */
625 static int do_balance_runtime(struct rt_rq *rt_rq)
626 {
627         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
628         struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
629         int i, weight, more = 0;
630         u64 rt_period;
631
632         weight = cpumask_weight(rd->span);
633
634         raw_spin_lock(&rt_b->rt_runtime_lock);
635         rt_period = ktime_to_ns(rt_b->rt_period);
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                 if (iter == rt_rq)
641                         continue;
642
643                 raw_spin_lock(&iter->rt_runtime_lock);
644                 /*
645                  * Either all rqs have inf runtime and there's nothing to steal
646                  * or __disable_runtime() below sets a specific rq to inf to
647                  * indicate its been disabled and disalow stealing.
648                  */
649                 if (iter->rt_runtime == RUNTIME_INF)
650                         goto next;
651
652                 /*
653                  * From runqueues with spare time, take 1/n part of their
654                  * spare time, but no more than our period.
655                  */
656                 diff = iter->rt_runtime - iter->rt_time;
657                 if (diff > 0) {
658                         diff = div_u64((u64)diff, weight);
659                         if (rt_rq->rt_runtime + diff > rt_period)
660                                 diff = rt_period - rt_rq->rt_runtime;
661                         iter->rt_runtime -= diff;
662                         rt_rq->rt_runtime += diff;
663                         more = 1;
664                         if (rt_rq->rt_runtime == rt_period) {
665                                 raw_spin_unlock(&iter->rt_runtime_lock);
666                                 break;
667                         }
668                 }
669 next:
670                 raw_spin_unlock(&iter->rt_runtime_lock);
671         }
672         raw_spin_unlock(&rt_b->rt_runtime_lock);
673
674         return more;
675 }
676
677 /*
678  * Ensure this RQ takes back all the runtime it lend to its neighbours.
679  */
680 static void __disable_runtime(struct rq *rq)
681 {
682         struct root_domain *rd = rq->rd;
683         rt_rq_iter_t iter;
684         struct rt_rq *rt_rq;
685
686         if (unlikely(!scheduler_running))
687                 return;
688
689         for_each_rt_rq(rt_rq, iter, rq) {
690                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
691                 s64 want;
692                 int i;
693
694                 raw_spin_lock(&rt_b->rt_runtime_lock);
695                 raw_spin_lock(&rt_rq->rt_runtime_lock);
696                 /*
697                  * Either we're all inf and nobody needs to borrow, or we're
698                  * already disabled and thus have nothing to do, or we have
699                  * exactly the right amount of runtime to take out.
700                  */
701                 if (rt_rq->rt_runtime == RUNTIME_INF ||
702                                 rt_rq->rt_runtime == rt_b->rt_runtime)
703                         goto balanced;
704                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
705
706                 /*
707                  * Calculate the difference between what we started out with
708                  * and what we current have, that's the amount of runtime
709                  * we lend and now have to reclaim.
710                  */
711                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
712
713                 /*
714                  * Greedy reclaim, take back as much as we can.
715                  */
716                 for_each_cpu(i, rd->span) {
717                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
718                         s64 diff;
719
720                         /*
721                          * Can't reclaim from ourselves or disabled runqueues.
722                          */
723                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
724                                 continue;
725
726                         raw_spin_lock(&iter->rt_runtime_lock);
727                         if (want > 0) {
728                                 diff = min_t(s64, iter->rt_runtime, want);
729                                 iter->rt_runtime -= diff;
730                                 want -= diff;
731                         } else {
732                                 iter->rt_runtime -= want;
733                                 want -= want;
734                         }
735                         raw_spin_unlock(&iter->rt_runtime_lock);
736
737                         if (!want)
738                                 break;
739                 }
740
741                 raw_spin_lock(&rt_rq->rt_runtime_lock);
742                 /*
743                  * We cannot be left wanting - that would mean some runtime
744                  * leaked out of the system.
745                  */
746                 BUG_ON(want);
747 balanced:
748                 /*
749                  * Disable all the borrow logic by pretending we have inf
750                  * runtime - in which case borrowing doesn't make sense.
751                  */
752                 rt_rq->rt_runtime = RUNTIME_INF;
753                 rt_rq->rt_throttled = 0;
754                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
755                 raw_spin_unlock(&rt_b->rt_runtime_lock);
756
757                 /* Make rt_rq available for pick_next_task() */
758                 sched_rt_rq_enqueue(rt_rq);
759         }
760 }
761
762 static void __enable_runtime(struct rq *rq)
763 {
764         rt_rq_iter_t iter;
765         struct rt_rq *rt_rq;
766
767         if (unlikely(!scheduler_running))
768                 return;
769
770         /*
771          * Reset each runqueue's bandwidth settings
772          */
773         for_each_rt_rq(rt_rq, iter, rq) {
774                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
775
776                 raw_spin_lock(&rt_b->rt_runtime_lock);
777                 raw_spin_lock(&rt_rq->rt_runtime_lock);
778                 rt_rq->rt_runtime = rt_b->rt_runtime;
779                 rt_rq->rt_time = 0;
780                 rt_rq->rt_throttled = 0;
781                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
782                 raw_spin_unlock(&rt_b->rt_runtime_lock);
783         }
784 }
785
786 static int balance_runtime(struct rt_rq *rt_rq)
787 {
788         int more = 0;
789
790         if (!sched_feat(RT_RUNTIME_SHARE))
791                 return more;
792
793         if (rt_rq->rt_time > rt_rq->rt_runtime) {
794                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
795                 more = do_balance_runtime(rt_rq);
796                 raw_spin_lock(&rt_rq->rt_runtime_lock);
797         }
798
799         return more;
800 }
801 #else /* !CONFIG_SMP */
802 static inline int balance_runtime(struct rt_rq *rt_rq)
803 {
804         return 0;
805 }
806 #endif /* CONFIG_SMP */
807
808 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
809 {
810         int i, idle = 1, throttled = 0;
811         const struct cpumask *span;
812
813         span = sched_rt_period_mask();
814 #ifdef CONFIG_RT_GROUP_SCHED
815         /*
816          * FIXME: isolated CPUs should really leave the root task group,
817          * whether they are isolcpus or were isolated via cpusets, lest
818          * the timer run on a CPU which does not service all runqueues,
819          * potentially leaving other CPUs indefinitely throttled.  If
820          * isolation is really required, the user will turn the throttle
821          * off to kill the perturbations it causes anyway.  Meanwhile,
822          * this maintains functionality for boot and/or troubleshooting.
823          */
824         if (rt_b == &root_task_group.rt_bandwidth)
825                 span = cpu_online_mask;
826 #endif
827         for_each_cpu(i, span) {
828                 int enqueue = 0;
829                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
830                 struct rq *rq = rq_of_rt_rq(rt_rq);
831
832                 raw_spin_lock(&rq->lock);
833                 if (rt_rq->rt_time) {
834                         u64 runtime;
835
836                         raw_spin_lock(&rt_rq->rt_runtime_lock);
837                         if (rt_rq->rt_throttled)
838                                 balance_runtime(rt_rq);
839                         runtime = rt_rq->rt_runtime;
840                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
841                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
842                                 rt_rq->rt_throttled = 0;
843                                 enqueue = 1;
844
845                                 /*
846                                  * When we're idle and a woken (rt) task is
847                                  * throttled check_preempt_curr() will set
848                                  * skip_update and the time between the wakeup
849                                  * and this unthrottle will get accounted as
850                                  * 'runtime'.
851                                  */
852                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
853                                         rq_clock_skip_update(rq, false);
854                         }
855                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
856                                 idle = 0;
857                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
858                 } else if (rt_rq->rt_nr_running) {
859                         idle = 0;
860                         if (!rt_rq_throttled(rt_rq))
861                                 enqueue = 1;
862                 }
863                 if (rt_rq->rt_throttled)
864                         throttled = 1;
865
866                 if (enqueue)
867                         sched_rt_rq_enqueue(rt_rq);
868                 raw_spin_unlock(&rq->lock);
869         }
870
871         if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
872                 return 1;
873
874         return idle;
875 }
876
877 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
878 {
879 #ifdef CONFIG_RT_GROUP_SCHED
880         struct rt_rq *rt_rq = group_rt_rq(rt_se);
881
882         if (rt_rq)
883                 return rt_rq->highest_prio.curr;
884 #endif
885
886         return rt_task_of(rt_se)->prio;
887 }
888
889 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
890 {
891         u64 runtime = sched_rt_runtime(rt_rq);
892
893         if (rt_rq->rt_throttled)
894                 return rt_rq_throttled(rt_rq);
895
896         if (runtime >= sched_rt_period(rt_rq))
897                 return 0;
898
899         balance_runtime(rt_rq);
900         runtime = sched_rt_runtime(rt_rq);
901         if (runtime == RUNTIME_INF)
902                 return 0;
903
904         if (rt_rq->rt_time > runtime) {
905                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
906
907                 /*
908                  * Don't actually throttle groups that have no runtime assigned
909                  * but accrue some time due to boosting.
910                  */
911                 if (likely(rt_b->rt_runtime)) {
912                         rt_rq->rt_throttled = 1;
913                         printk_deferred_once("sched: RT throttling activated\n");
914                 } else {
915                         /*
916                          * In case we did anyway, make it go away,
917                          * replenishment is a joke, since it will replenish us
918                          * with exactly 0 ns.
919                          */
920                         rt_rq->rt_time = 0;
921                 }
922
923                 if (rt_rq_throttled(rt_rq)) {
924                         sched_rt_rq_dequeue(rt_rq);
925                         return 1;
926                 }
927         }
928
929         return 0;
930 }
931
932 /*
933  * Update the current task's runtime statistics. Skip current tasks that
934  * are not in our scheduling class.
935  */
936 static void update_curr_rt(struct rq *rq)
937 {
938         struct task_struct *curr = rq->curr;
939         struct sched_rt_entity *rt_se = &curr->rt;
940         u64 delta_exec;
941
942         if (curr->sched_class != &rt_sched_class)
943                 return;
944
945         delta_exec = rq_clock_task(rq) - curr->se.exec_start;
946         if (unlikely((s64)delta_exec <= 0))
947                 return;
948
949         schedstat_set(curr->se.statistics.exec_max,
950                       max(curr->se.statistics.exec_max, delta_exec));
951
952         curr->se.sum_exec_runtime += delta_exec;
953         account_group_exec_runtime(curr, delta_exec);
954
955         curr->se.exec_start = rq_clock_task(rq);
956         cpuacct_charge(curr, delta_exec);
957
958         sched_rt_avg_update(rq, delta_exec);
959
960         if (!rt_bandwidth_enabled())
961                 return;
962
963         for_each_sched_rt_entity(rt_se) {
964                 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
965
966                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
967                         raw_spin_lock(&rt_rq->rt_runtime_lock);
968                         rt_rq->rt_time += delta_exec;
969                         if (sched_rt_runtime_exceeded(rt_rq))
970                                 resched_curr(rq);
971                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
972                 }
973         }
974 }
975
976 static void
977 dequeue_top_rt_rq(struct rt_rq *rt_rq)
978 {
979         struct rq *rq = rq_of_rt_rq(rt_rq);
980
981         BUG_ON(&rq->rt != rt_rq);
982
983         if (!rt_rq->rt_queued)
984                 return;
985
986         BUG_ON(!rq->nr_running);
987
988         sub_nr_running(rq, rt_rq->rt_nr_running);
989         rt_rq->rt_queued = 0;
990 }
991
992 static void
993 enqueue_top_rt_rq(struct rt_rq *rt_rq)
994 {
995         struct rq *rq = rq_of_rt_rq(rt_rq);
996
997         BUG_ON(&rq->rt != rt_rq);
998
999         if (rt_rq->rt_queued)
1000                 return;
1001         if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1002                 return;
1003
1004         add_nr_running(rq, rt_rq->rt_nr_running);
1005         rt_rq->rt_queued = 1;
1006 }
1007
1008 #if defined CONFIG_SMP
1009
1010 static void
1011 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1012 {
1013         struct rq *rq = rq_of_rt_rq(rt_rq);
1014
1015 #ifdef CONFIG_RT_GROUP_SCHED
1016         /*
1017          * Change rq's cpupri only if rt_rq is the top queue.
1018          */
1019         if (&rq->rt != rt_rq)
1020                 return;
1021 #endif
1022         if (rq->online && prio < prev_prio)
1023                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1024 }
1025
1026 static void
1027 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1028 {
1029         struct rq *rq = rq_of_rt_rq(rt_rq);
1030
1031 #ifdef CONFIG_RT_GROUP_SCHED
1032         /*
1033          * Change rq's cpupri only if rt_rq is the top queue.
1034          */
1035         if (&rq->rt != rt_rq)
1036                 return;
1037 #endif
1038         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1039                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1040 }
1041
1042 #else /* CONFIG_SMP */
1043
1044 static inline
1045 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1046 static inline
1047 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1048
1049 #endif /* CONFIG_SMP */
1050
1051 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1052 static void
1053 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1054 {
1055         int prev_prio = rt_rq->highest_prio.curr;
1056
1057         if (prio < prev_prio)
1058                 rt_rq->highest_prio.curr = prio;
1059
1060         inc_rt_prio_smp(rt_rq, prio, prev_prio);
1061 }
1062
1063 static void
1064 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1065 {
1066         int prev_prio = rt_rq->highest_prio.curr;
1067
1068         if (rt_rq->rt_nr_running) {
1069
1070                 WARN_ON(prio < prev_prio);
1071
1072                 /*
1073                  * This may have been our highest task, and therefore
1074                  * we may have some recomputation to do
1075                  */
1076                 if (prio == prev_prio) {
1077                         struct rt_prio_array *array = &rt_rq->active;
1078
1079                         rt_rq->highest_prio.curr =
1080                                 sched_find_first_bit(array->bitmap);
1081                 }
1082
1083         } else
1084                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1085
1086         dec_rt_prio_smp(rt_rq, prio, prev_prio);
1087 }
1088
1089 #else
1090
1091 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1092 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1093
1094 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1095
1096 #ifdef CONFIG_RT_GROUP_SCHED
1097
1098 static void
1099 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1100 {
1101         if (rt_se_boosted(rt_se))
1102                 rt_rq->rt_nr_boosted++;
1103
1104         if (rt_rq->tg)
1105                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1106 }
1107
1108 static void
1109 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1110 {
1111         if (rt_se_boosted(rt_se))
1112                 rt_rq->rt_nr_boosted--;
1113
1114         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1115 }
1116
1117 #else /* CONFIG_RT_GROUP_SCHED */
1118
1119 static void
1120 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1121 {
1122         start_rt_bandwidth(&def_rt_bandwidth);
1123 }
1124
1125 static inline
1126 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1127
1128 #endif /* CONFIG_RT_GROUP_SCHED */
1129
1130 static inline
1131 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1132 {
1133         struct rt_rq *group_rq = group_rt_rq(rt_se);
1134
1135         if (group_rq)
1136                 return group_rq->rt_nr_running;
1137         else
1138                 return 1;
1139 }
1140
1141 static inline
1142 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1143 {
1144         int prio = rt_se_prio(rt_se);
1145
1146         WARN_ON(!rt_prio(prio));
1147         rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1148
1149         inc_rt_prio(rt_rq, prio);
1150         inc_rt_migration(rt_se, rt_rq);
1151         inc_rt_group(rt_se, rt_rq);
1152 }
1153
1154 static inline
1155 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1156 {
1157         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1158         WARN_ON(!rt_rq->rt_nr_running);
1159         rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1160
1161         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1162         dec_rt_migration(rt_se, rt_rq);
1163         dec_rt_group(rt_se, rt_rq);
1164 }
1165
1166 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1167 {
1168         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1169         struct rt_prio_array *array = &rt_rq->active;
1170         struct rt_rq *group_rq = group_rt_rq(rt_se);
1171         struct list_head *queue = array->queue + rt_se_prio(rt_se);
1172
1173         /*
1174          * Don't enqueue the group if its throttled, or when empty.
1175          * The latter is a consequence of the former when a child group
1176          * get throttled and the current group doesn't have any other
1177          * active members.
1178          */
1179         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
1180                 return;
1181
1182         if (head)
1183                 list_add(&rt_se->run_list, queue);
1184         else
1185                 list_add_tail(&rt_se->run_list, queue);
1186         __set_bit(rt_se_prio(rt_se), array->bitmap);
1187
1188         inc_rt_tasks(rt_se, rt_rq);
1189 }
1190
1191 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
1192 {
1193         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1194         struct rt_prio_array *array = &rt_rq->active;
1195
1196         list_del_init(&rt_se->run_list);
1197         if (list_empty(array->queue + rt_se_prio(rt_se)))
1198                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1199
1200         dec_rt_tasks(rt_se, rt_rq);
1201 }
1202
1203 /*
1204  * Because the prio of an upper entry depends on the lower
1205  * entries, we must remove entries top - down.
1206  */
1207 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
1208 {
1209         struct sched_rt_entity *back = NULL;
1210
1211         for_each_sched_rt_entity(rt_se) {
1212                 rt_se->back = back;
1213                 back = rt_se;
1214         }
1215
1216         dequeue_top_rt_rq(rt_rq_of_se(back));
1217
1218         for (rt_se = back; rt_se; rt_se = rt_se->back) {
1219                 if (on_rt_rq(rt_se))
1220                         __dequeue_rt_entity(rt_se);
1221         }
1222 }
1223
1224 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1225 {
1226         struct rq *rq = rq_of_rt_se(rt_se);
1227
1228         dequeue_rt_stack(rt_se);
1229         for_each_sched_rt_entity(rt_se)
1230                 __enqueue_rt_entity(rt_se, head);
1231         enqueue_top_rt_rq(&rq->rt);
1232 }
1233
1234 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1235 {
1236         struct rq *rq = rq_of_rt_se(rt_se);
1237
1238         dequeue_rt_stack(rt_se);
1239
1240         for_each_sched_rt_entity(rt_se) {
1241                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1242
1243                 if (rt_rq && rt_rq->rt_nr_running)
1244                         __enqueue_rt_entity(rt_se, false);
1245         }
1246         enqueue_top_rt_rq(&rq->rt);
1247 }
1248
1249 /*
1250  * Adding/removing a task to/from a priority array:
1251  */
1252 static void
1253 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1254 {
1255         struct sched_rt_entity *rt_se = &p->rt;
1256
1257         if (flags & ENQUEUE_WAKEUP)
1258                 rt_se->timeout = 0;
1259
1260         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
1261
1262         if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1263                 enqueue_pushable_task(rq, p);
1264 }
1265
1266 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1267 {
1268         struct sched_rt_entity *rt_se = &p->rt;
1269
1270         update_curr_rt(rq);
1271         dequeue_rt_entity(rt_se);
1272
1273         dequeue_pushable_task(rq, p);
1274 }
1275
1276 /*
1277  * Put task to the head or the end of the run list without the overhead of
1278  * dequeue followed by enqueue.
1279  */
1280 static void
1281 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1282 {
1283         if (on_rt_rq(rt_se)) {
1284                 struct rt_prio_array *array = &rt_rq->active;
1285                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1286
1287                 if (head)
1288                         list_move(&rt_se->run_list, queue);
1289                 else
1290                         list_move_tail(&rt_se->run_list, queue);
1291         }
1292 }
1293
1294 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1295 {
1296         struct sched_rt_entity *rt_se = &p->rt;
1297         struct rt_rq *rt_rq;
1298
1299         for_each_sched_rt_entity(rt_se) {
1300                 rt_rq = rt_rq_of_se(rt_se);
1301                 requeue_rt_entity(rt_rq, rt_se, head);
1302         }
1303 }
1304
1305 static void yield_task_rt(struct rq *rq)
1306 {
1307         requeue_task_rt(rq, rq->curr, 0);
1308 }
1309
1310 #ifdef CONFIG_SMP
1311 static int find_lowest_rq(struct task_struct *task);
1312
1313 static int
1314 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1315 {
1316         struct task_struct *curr;
1317         struct rq *rq;
1318
1319         /* For anything but wake ups, just return the task_cpu */
1320         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1321                 goto out;
1322
1323         rq = cpu_rq(cpu);
1324
1325         rcu_read_lock();
1326         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1327
1328         /*
1329          * If the current task on @p's runqueue is an RT task, then
1330          * try to see if we can wake this RT task up on another
1331          * runqueue. Otherwise simply start this RT task
1332          * on its current runqueue.
1333          *
1334          * We want to avoid overloading runqueues. If the woken
1335          * task is a higher priority, then it will stay on this CPU
1336          * and the lower prio task should be moved to another CPU.
1337          * Even though this will probably make the lower prio task
1338          * lose its cache, we do not want to bounce a higher task
1339          * around just because it gave up its CPU, perhaps for a
1340          * lock?
1341          *
1342          * For equal prio tasks, we just let the scheduler sort it out.
1343          *
1344          * Otherwise, just let it ride on the affined RQ and the
1345          * post-schedule router will push the preempted task away
1346          *
1347          * This test is optimistic, if we get it wrong the load-balancer
1348          * will have to sort it out.
1349          */
1350         if (curr && unlikely(rt_task(curr)) &&
1351             (curr->nr_cpus_allowed < 2 ||
1352              curr->prio <= p->prio)) {
1353                 int target = find_lowest_rq(p);
1354
1355                 /*
1356                  * Don't bother moving it if the destination CPU is
1357                  * not running a lower priority task.
1358                  */
1359                 if (target != -1 &&
1360                     p->prio < cpu_rq(target)->rt.highest_prio.curr)
1361                         cpu = target;
1362         }
1363         rcu_read_unlock();
1364
1365 out:
1366         return cpu;
1367 }
1368
1369 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1370 {
1371         /*
1372          * Current can't be migrated, useless to reschedule,
1373          * let's hope p can move out.
1374          */
1375         if (rq->curr->nr_cpus_allowed == 1 ||
1376             !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1377                 return;
1378
1379         /*
1380          * p is migratable, so let's not schedule it and
1381          * see if it is pushed or pulled somewhere else.
1382          */
1383         if (p->nr_cpus_allowed != 1
1384             && cpupri_find(&rq->rd->cpupri, p, NULL))
1385                 return;
1386
1387         /*
1388          * There appears to be other cpus that can accept
1389          * current and none to run 'p', so lets reschedule
1390          * to try and push current away:
1391          */
1392         requeue_task_rt(rq, p, 1);
1393         resched_curr(rq);
1394 }
1395
1396 #endif /* CONFIG_SMP */
1397
1398 /*
1399  * Preempt the current task with a newly woken task if needed:
1400  */
1401 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1402 {
1403         if (p->prio < rq->curr->prio) {
1404                 resched_curr(rq);
1405                 return;
1406         }
1407
1408 #ifdef CONFIG_SMP
1409         /*
1410          * If:
1411          *
1412          * - the newly woken task is of equal priority to the current task
1413          * - the newly woken task is non-migratable while current is migratable
1414          * - current will be preempted on the next reschedule
1415          *
1416          * we should check to see if current can readily move to a different
1417          * cpu.  If so, we will reschedule to allow the push logic to try
1418          * to move current somewhere else, making room for our non-migratable
1419          * task.
1420          */
1421         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1422                 check_preempt_equal_prio(rq, p);
1423 #endif
1424 }
1425
1426 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1427                                                    struct rt_rq *rt_rq)
1428 {
1429         struct rt_prio_array *array = &rt_rq->active;
1430         struct sched_rt_entity *next = NULL;
1431         struct list_head *queue;
1432         int idx;
1433
1434         idx = sched_find_first_bit(array->bitmap);
1435         BUG_ON(idx >= MAX_RT_PRIO);
1436
1437         queue = array->queue + idx;
1438         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1439
1440         return next;
1441 }
1442
1443 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1444 {
1445         struct sched_rt_entity *rt_se;
1446         struct task_struct *p;
1447         struct rt_rq *rt_rq  = &rq->rt;
1448
1449         do {
1450                 rt_se = pick_next_rt_entity(rq, rt_rq);
1451                 BUG_ON(!rt_se);
1452                 rt_rq = group_rt_rq(rt_se);
1453         } while (rt_rq);
1454
1455         p = rt_task_of(rt_se);
1456         p->se.exec_start = rq_clock_task(rq);
1457
1458         return p;
1459 }
1460
1461 static struct task_struct *
1462 pick_next_task_rt(struct rq *rq, struct task_struct *prev)
1463 {
1464         struct task_struct *p;
1465         struct rt_rq *rt_rq = &rq->rt;
1466
1467         if (need_pull_rt_task(rq, prev)) {
1468                 pull_rt_task(rq);
1469                 /*
1470                  * pull_rt_task() can drop (and re-acquire) rq->lock; this
1471                  * means a dl or stop task can slip in, in which case we need
1472                  * to re-start task selection.
1473                  */
1474                 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1475                              rq->dl.dl_nr_running))
1476                         return RETRY_TASK;
1477         }
1478
1479         /*
1480          * We may dequeue prev's rt_rq in put_prev_task().
1481          * So, we update time before rt_nr_running check.
1482          */
1483         if (prev->sched_class == &rt_sched_class)
1484                 update_curr_rt(rq);
1485
1486         if (!rt_rq->rt_queued)
1487                 return NULL;
1488
1489         put_prev_task(rq, prev);
1490
1491         p = _pick_next_task_rt(rq);
1492
1493         /* The running task is never eligible for pushing */
1494         dequeue_pushable_task(rq, p);
1495
1496         set_post_schedule(rq);
1497
1498         return p;
1499 }
1500
1501 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1502 {
1503         update_curr_rt(rq);
1504
1505         /*
1506          * The previous task needs to be made eligible for pushing
1507          * if it is still active
1508          */
1509         if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1510                 enqueue_pushable_task(rq, p);
1511 }
1512
1513 #ifdef CONFIG_SMP
1514
1515 /* Only try algorithms three times */
1516 #define RT_MAX_TRIES 3
1517
1518 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1519 {
1520         if (!task_running(rq, p) &&
1521             cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1522                 return 1;
1523         return 0;
1524 }
1525
1526 /*
1527  * Return the highest pushable rq's task, which is suitable to be executed
1528  * on the cpu, NULL otherwise
1529  */
1530 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1531 {
1532         struct plist_head *head = &rq->rt.pushable_tasks;
1533         struct task_struct *p;
1534
1535         if (!has_pushable_tasks(rq))
1536                 return NULL;
1537
1538         plist_for_each_entry(p, head, pushable_tasks) {
1539                 if (pick_rt_task(rq, p, cpu))
1540                         return p;
1541         }
1542
1543         return NULL;
1544 }
1545
1546 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1547
1548 static int find_lowest_rq(struct task_struct *task)
1549 {
1550         struct sched_domain *sd;
1551         struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1552         int this_cpu = smp_processor_id();
1553         int cpu      = task_cpu(task);
1554
1555         /* Make sure the mask is initialized first */
1556         if (unlikely(!lowest_mask))
1557                 return -1;
1558
1559         if (task->nr_cpus_allowed == 1)
1560                 return -1; /* No other targets possible */
1561
1562         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1563                 return -1; /* No targets found */
1564
1565         /*
1566          * At this point we have built a mask of cpus representing the
1567          * lowest priority tasks in the system.  Now we want to elect
1568          * the best one based on our affinity and topology.
1569          *
1570          * We prioritize the last cpu that the task executed on since
1571          * it is most likely cache-hot in that location.
1572          */
1573         if (cpumask_test_cpu(cpu, lowest_mask))
1574                 return cpu;
1575
1576         /*
1577          * Otherwise, we consult the sched_domains span maps to figure
1578          * out which cpu is logically closest to our hot cache data.
1579          */
1580         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1581                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1582
1583         rcu_read_lock();
1584         for_each_domain(cpu, sd) {
1585                 if (sd->flags & SD_WAKE_AFFINE) {
1586                         int best_cpu;
1587
1588                         /*
1589                          * "this_cpu" is cheaper to preempt than a
1590                          * remote processor.
1591                          */
1592                         if (this_cpu != -1 &&
1593                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1594                                 rcu_read_unlock();
1595                                 return this_cpu;
1596                         }
1597
1598                         best_cpu = cpumask_first_and(lowest_mask,
1599                                                      sched_domain_span(sd));
1600                         if (best_cpu < nr_cpu_ids) {
1601                                 rcu_read_unlock();
1602                                 return best_cpu;
1603                         }
1604                 }
1605         }
1606         rcu_read_unlock();
1607
1608         /*
1609          * And finally, if there were no matches within the domains
1610          * just give the caller *something* to work with from the compatible
1611          * locations.
1612          */
1613         if (this_cpu != -1)
1614                 return this_cpu;
1615
1616         cpu = cpumask_any(lowest_mask);
1617         if (cpu < nr_cpu_ids)
1618                 return cpu;
1619         return -1;
1620 }
1621
1622 /* Will lock the rq it finds */
1623 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1624 {
1625         struct rq *lowest_rq = NULL;
1626         int tries;
1627         int cpu;
1628
1629         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1630                 cpu = find_lowest_rq(task);
1631
1632                 if ((cpu == -1) || (cpu == rq->cpu))
1633                         break;
1634
1635                 lowest_rq = cpu_rq(cpu);
1636
1637                 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1638                         /*
1639                          * Target rq has tasks of equal or higher priority,
1640                          * retrying does not release any lock and is unlikely
1641                          * to yield a different result.
1642                          */
1643                         lowest_rq = NULL;
1644                         break;
1645                 }
1646
1647                 /* if the prio of this runqueue changed, try again */
1648                 if (double_lock_balance(rq, lowest_rq)) {
1649                         /*
1650                          * We had to unlock the run queue. In
1651                          * the mean time, task could have
1652                          * migrated already or had its affinity changed.
1653                          * Also make sure that it wasn't scheduled on its rq.
1654                          */
1655                         if (unlikely(task_rq(task) != rq ||
1656                                      !cpumask_test_cpu(lowest_rq->cpu,
1657                                                        tsk_cpus_allowed(task)) ||
1658                                      task_running(rq, task) ||
1659                                      !task_on_rq_queued(task))) {
1660
1661                                 double_unlock_balance(rq, lowest_rq);
1662                                 lowest_rq = NULL;
1663                                 break;
1664                         }
1665                 }
1666
1667                 /* If this rq is still suitable use it. */
1668                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1669                         break;
1670
1671                 /* try again */
1672                 double_unlock_balance(rq, lowest_rq);
1673                 lowest_rq = NULL;
1674         }
1675
1676         return lowest_rq;
1677 }
1678
1679 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1680 {
1681         struct task_struct *p;
1682
1683         if (!has_pushable_tasks(rq))
1684                 return NULL;
1685
1686         p = plist_first_entry(&rq->rt.pushable_tasks,
1687                               struct task_struct, pushable_tasks);
1688
1689         BUG_ON(rq->cpu != task_cpu(p));
1690         BUG_ON(task_current(rq, p));
1691         BUG_ON(p->nr_cpus_allowed <= 1);
1692
1693         BUG_ON(!task_on_rq_queued(p));
1694         BUG_ON(!rt_task(p));
1695
1696         return p;
1697 }
1698
1699 /*
1700  * If the current CPU has more than one RT task, see if the non
1701  * running task can migrate over to a CPU that is running a task
1702  * of lesser priority.
1703  */
1704 static int push_rt_task(struct rq *rq)
1705 {
1706         struct task_struct *next_task;
1707         struct rq *lowest_rq;
1708         int ret = 0;
1709
1710         if (!rq->rt.overloaded)
1711                 return 0;
1712
1713         next_task = pick_next_pushable_task(rq);
1714         if (!next_task)
1715                 return 0;
1716
1717 retry:
1718         if (unlikely(next_task == rq->curr)) {
1719                 WARN_ON(1);
1720                 return 0;
1721         }
1722
1723         /*
1724          * It's possible that the next_task slipped in of
1725          * higher priority than current. If that's the case
1726          * just reschedule current.
1727          */
1728         if (unlikely(next_task->prio < rq->curr->prio)) {
1729                 resched_curr(rq);
1730                 return 0;
1731         }
1732
1733         /* We might release rq lock */
1734         get_task_struct(next_task);
1735
1736         /* find_lock_lowest_rq locks the rq if found */
1737         lowest_rq = find_lock_lowest_rq(next_task, rq);
1738         if (!lowest_rq) {
1739                 struct task_struct *task;
1740                 /*
1741                  * find_lock_lowest_rq releases rq->lock
1742                  * so it is possible that next_task has migrated.
1743                  *
1744                  * We need to make sure that the task is still on the same
1745                  * run-queue and is also still the next task eligible for
1746                  * pushing.
1747                  */
1748                 task = pick_next_pushable_task(rq);
1749                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1750                         /*
1751                          * The task hasn't migrated, and is still the next
1752                          * eligible task, but we failed to find a run-queue
1753                          * to push it to.  Do not retry in this case, since
1754                          * other cpus will pull from us when ready.
1755                          */
1756                         goto out;
1757                 }
1758
1759                 if (!task)
1760                         /* No more tasks, just exit */
1761                         goto out;
1762
1763                 /*
1764                  * Something has shifted, try again.
1765                  */
1766                 put_task_struct(next_task);
1767                 next_task = task;
1768                 goto retry;
1769         }
1770
1771         deactivate_task(rq, next_task, 0);
1772         set_task_cpu(next_task, lowest_rq->cpu);
1773         activate_task(lowest_rq, next_task, 0);
1774         ret = 1;
1775
1776         resched_curr(lowest_rq);
1777
1778         double_unlock_balance(rq, lowest_rq);
1779
1780 out:
1781         put_task_struct(next_task);
1782
1783         return ret;
1784 }
1785
1786 static void push_rt_tasks(struct rq *rq)
1787 {
1788         /* push_rt_task will return true if it moved an RT */
1789         while (push_rt_task(rq))
1790                 ;
1791 }
1792
1793 #ifdef HAVE_RT_PUSH_IPI
1794 /*
1795  * The search for the next cpu always starts at rq->cpu and ends
1796  * when we reach rq->cpu again. It will never return rq->cpu.
1797  * This returns the next cpu to check, or nr_cpu_ids if the loop
1798  * is complete.
1799  *
1800  * rq->rt.push_cpu holds the last cpu returned by this function,
1801  * or if this is the first instance, it must hold rq->cpu.
1802  */
1803 static int rto_next_cpu(struct rq *rq)
1804 {
1805         int prev_cpu = rq->rt.push_cpu;
1806         int cpu;
1807
1808         cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1809
1810         /*
1811          * If the previous cpu is less than the rq's CPU, then it already
1812          * passed the end of the mask, and has started from the beginning.
1813          * We end if the next CPU is greater or equal to rq's CPU.
1814          */
1815         if (prev_cpu < rq->cpu) {
1816                 if (cpu >= rq->cpu)
1817                         return nr_cpu_ids;
1818
1819         } else if (cpu >= nr_cpu_ids) {
1820                 /*
1821                  * We passed the end of the mask, start at the beginning.
1822                  * If the result is greater or equal to the rq's CPU, then
1823                  * the loop is finished.
1824                  */
1825                 cpu = cpumask_first(rq->rd->rto_mask);
1826                 if (cpu >= rq->cpu)
1827                         return nr_cpu_ids;
1828         }
1829         rq->rt.push_cpu = cpu;
1830
1831         /* Return cpu to let the caller know if the loop is finished or not */
1832         return cpu;
1833 }
1834
1835 static int find_next_push_cpu(struct rq *rq)
1836 {
1837         struct rq *next_rq;
1838         int cpu;
1839
1840         while (1) {
1841                 cpu = rto_next_cpu(rq);
1842                 if (cpu >= nr_cpu_ids)
1843                         break;
1844                 next_rq = cpu_rq(cpu);
1845
1846                 /* Make sure the next rq can push to this rq */
1847                 if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1848                         break;
1849         }
1850
1851         return cpu;
1852 }
1853
1854 #define RT_PUSH_IPI_EXECUTING           1
1855 #define RT_PUSH_IPI_RESTART             2
1856
1857 static void tell_cpu_to_push(struct rq *rq)
1858 {
1859         int cpu;
1860
1861         if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1862                 raw_spin_lock(&rq->rt.push_lock);
1863                 /* Make sure it's still executing */
1864                 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1865                         /*
1866                          * Tell the IPI to restart the loop as things have
1867                          * changed since it started.
1868                          */
1869                         rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
1870                         raw_spin_unlock(&rq->rt.push_lock);
1871                         return;
1872                 }
1873                 raw_spin_unlock(&rq->rt.push_lock);
1874         }
1875
1876         /* When here, there's no IPI going around */
1877
1878         rq->rt.push_cpu = rq->cpu;
1879         cpu = find_next_push_cpu(rq);
1880         if (cpu >= nr_cpu_ids)
1881                 return;
1882
1883         rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
1884
1885         irq_work_queue_on(&rq->rt.push_work, cpu);
1886 }
1887
1888 /* Called from hardirq context */
1889 static void try_to_push_tasks(void *arg)
1890 {
1891         struct rt_rq *rt_rq = arg;
1892         struct rq *rq, *src_rq;
1893         int this_cpu;
1894         int cpu;
1895
1896         this_cpu = rt_rq->push_cpu;
1897
1898         /* Paranoid check */
1899         BUG_ON(this_cpu != smp_processor_id());
1900
1901         rq = cpu_rq(this_cpu);
1902         src_rq = rq_of_rt_rq(rt_rq);
1903
1904 again:
1905         if (has_pushable_tasks(rq)) {
1906                 raw_spin_lock(&rq->lock);
1907                 push_rt_task(rq);
1908                 raw_spin_unlock(&rq->lock);
1909         }
1910
1911         /* Pass the IPI to the next rt overloaded queue */
1912         raw_spin_lock(&rt_rq->push_lock);
1913         /*
1914          * If the source queue changed since the IPI went out,
1915          * we need to restart the search from that CPU again.
1916          */
1917         if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
1918                 rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
1919                 rt_rq->push_cpu = src_rq->cpu;
1920         }
1921
1922         cpu = find_next_push_cpu(src_rq);
1923
1924         if (cpu >= nr_cpu_ids)
1925                 rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
1926         raw_spin_unlock(&rt_rq->push_lock);
1927
1928         if (cpu >= nr_cpu_ids)
1929                 return;
1930
1931         /*
1932          * It is possible that a restart caused this CPU to be
1933          * chosen again. Don't bother with an IPI, just see if we
1934          * have more to push.
1935          */
1936         if (unlikely(cpu == rq->cpu))
1937                 goto again;
1938
1939         /* Try the next RT overloaded CPU */
1940         irq_work_queue_on(&rt_rq->push_work, cpu);
1941 }
1942
1943 static void push_irq_work_func(struct irq_work *work)
1944 {
1945         struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
1946
1947         try_to_push_tasks(rt_rq);
1948 }
1949 #endif /* HAVE_RT_PUSH_IPI */
1950
1951 static int pull_rt_task(struct rq *this_rq)
1952 {
1953         int this_cpu = this_rq->cpu, ret = 0, cpu;
1954         struct task_struct *p;
1955         struct rq *src_rq;
1956
1957         if (likely(!rt_overloaded(this_rq)))
1958                 return 0;
1959
1960         /*
1961          * Match the barrier from rt_set_overloaded; this guarantees that if we
1962          * see overloaded we must also see the rto_mask bit.
1963          */
1964         smp_rmb();
1965
1966 #ifdef HAVE_RT_PUSH_IPI
1967         if (sched_feat(RT_PUSH_IPI)) {
1968                 tell_cpu_to_push(this_rq);
1969                 return 0;
1970         }
1971 #endif
1972
1973         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1974                 if (this_cpu == cpu)
1975                         continue;
1976
1977                 src_rq = cpu_rq(cpu);
1978
1979                 /*
1980                  * Don't bother taking the src_rq->lock if the next highest
1981                  * task is known to be lower-priority than our current task.
1982                  * This may look racy, but if this value is about to go
1983                  * logically higher, the src_rq will push this task away.
1984                  * And if its going logically lower, we do not care
1985                  */
1986                 if (src_rq->rt.highest_prio.next >=
1987                     this_rq->rt.highest_prio.curr)
1988                         continue;
1989
1990                 /*
1991                  * We can potentially drop this_rq's lock in
1992                  * double_lock_balance, and another CPU could
1993                  * alter this_rq
1994                  */
1995                 double_lock_balance(this_rq, src_rq);
1996
1997                 /*
1998                  * We can pull only a task, which is pushable
1999                  * on its rq, and no others.
2000                  */
2001                 p = pick_highest_pushable_task(src_rq, this_cpu);
2002
2003                 /*
2004                  * Do we have an RT task that preempts
2005                  * the to-be-scheduled task?
2006                  */
2007                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2008                         WARN_ON(p == src_rq->curr);
2009                         WARN_ON(!task_on_rq_queued(p));
2010
2011                         /*
2012                          * There's a chance that p is higher in priority
2013                          * than what's currently running on its cpu.
2014                          * This is just that p is wakeing up and hasn't
2015                          * had a chance to schedule. We only pull
2016                          * p if it is lower in priority than the
2017                          * current task on the run queue
2018                          */
2019                         if (p->prio < src_rq->curr->prio)
2020                                 goto skip;
2021
2022                         ret = 1;
2023
2024                         deactivate_task(src_rq, p, 0);
2025                         set_task_cpu(p, this_cpu);
2026                         activate_task(this_rq, p, 0);
2027                         /*
2028                          * We continue with the search, just in
2029                          * case there's an even higher prio task
2030                          * in another runqueue. (low likelihood
2031                          * but possible)
2032                          */
2033                 }
2034 skip:
2035                 double_unlock_balance(this_rq, src_rq);
2036         }
2037
2038         return ret;
2039 }
2040
2041 static void post_schedule_rt(struct rq *rq)
2042 {
2043         push_rt_tasks(rq);
2044 }
2045
2046 /*
2047  * If we are not running and we are not going to reschedule soon, we should
2048  * try to push tasks away now
2049  */
2050 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2051 {
2052         if (!task_running(rq, p) &&
2053             !test_tsk_need_resched(rq->curr) &&
2054             has_pushable_tasks(rq) &&
2055             p->nr_cpus_allowed > 1 &&
2056             (dl_task(rq->curr) || rt_task(rq->curr)) &&
2057             (rq->curr->nr_cpus_allowed < 2 ||
2058              rq->curr->prio <= p->prio))
2059                 push_rt_tasks(rq);
2060 }
2061
2062 static void set_cpus_allowed_rt(struct task_struct *p,
2063                                 const struct cpumask *new_mask)
2064 {
2065         struct rq *rq;
2066         int weight;
2067
2068         BUG_ON(!rt_task(p));
2069
2070         if (!task_on_rq_queued(p))
2071                 return;
2072
2073         weight = cpumask_weight(new_mask);
2074
2075         /*
2076          * Only update if the process changes its state from whether it
2077          * can migrate or not.
2078          */
2079         if ((p->nr_cpus_allowed > 1) == (weight > 1))
2080                 return;
2081
2082         rq = task_rq(p);
2083
2084         /*
2085          * The process used to be able to migrate OR it can now migrate
2086          */
2087         if (weight <= 1) {
2088                 if (!task_current(rq, p))
2089                         dequeue_pushable_task(rq, p);
2090                 BUG_ON(!rq->rt.rt_nr_migratory);
2091                 rq->rt.rt_nr_migratory--;
2092         } else {
2093                 if (!task_current(rq, p))
2094                         enqueue_pushable_task(rq, p);
2095                 rq->rt.rt_nr_migratory++;
2096         }
2097
2098         update_rt_migration(&rq->rt);
2099 }
2100
2101 /* Assumes rq->lock is held */
2102 static void rq_online_rt(struct rq *rq)
2103 {
2104         if (rq->rt.overloaded)
2105                 rt_set_overload(rq);
2106
2107         __enable_runtime(rq);
2108
2109         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2110 }
2111
2112 /* Assumes rq->lock is held */
2113 static void rq_offline_rt(struct rq *rq)
2114 {
2115         if (rq->rt.overloaded)
2116                 rt_clear_overload(rq);
2117
2118         __disable_runtime(rq);
2119
2120         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2121 }
2122
2123 /*
2124  * When switch from the rt queue, we bring ourselves to a position
2125  * that we might want to pull RT tasks from other runqueues.
2126  */
2127 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2128 {
2129         /*
2130          * If there are other RT tasks then we will reschedule
2131          * and the scheduling of the other RT tasks will handle
2132          * the balancing. But if we are the last RT task
2133          * we may need to handle the pulling of RT tasks
2134          * now.
2135          */
2136         if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2137                 return;
2138
2139         if (pull_rt_task(rq))
2140                 resched_curr(rq);
2141 }
2142
2143 void __init init_sched_rt_class(void)
2144 {
2145         unsigned int i;
2146
2147         for_each_possible_cpu(i) {
2148                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2149                                         GFP_KERNEL, cpu_to_node(i));
2150         }
2151 }
2152 #endif /* CONFIG_SMP */
2153
2154 /*
2155  * When switching a task to RT, we may overload the runqueue
2156  * with RT tasks. In this case we try to push them off to
2157  * other runqueues.
2158  */
2159 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2160 {
2161         int check_resched = 1;
2162
2163         /*
2164          * If we are already running, then there's nothing
2165          * that needs to be done. But if we are not running
2166          * we may need to preempt the current running task.
2167          * If that current running task is also an RT task
2168          * then see if we can move to another run queue.
2169          */
2170         if (task_on_rq_queued(p) && rq->curr != p) {
2171 #ifdef CONFIG_SMP
2172                 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded &&
2173                     /* Don't resched if we changed runqueues */
2174                     push_rt_task(rq) && rq != task_rq(p))
2175                         check_resched = 0;
2176 #endif /* CONFIG_SMP */
2177                 if (check_resched && p->prio < rq->curr->prio)
2178                         resched_curr(rq);
2179         }
2180 }
2181
2182 /*
2183  * Priority of the task has changed. This may cause
2184  * us to initiate a push or pull.
2185  */
2186 static void
2187 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2188 {
2189         if (!task_on_rq_queued(p))
2190                 return;
2191
2192         if (rq->curr == p) {
2193 #ifdef CONFIG_SMP
2194                 /*
2195                  * If our priority decreases while running, we
2196                  * may need to pull tasks to this runqueue.
2197                  */
2198                 if (oldprio < p->prio)
2199                         pull_rt_task(rq);
2200                 /*
2201                  * If there's a higher priority task waiting to run
2202                  * then reschedule. Note, the above pull_rt_task
2203                  * can release the rq lock and p could migrate.
2204                  * Only reschedule if p is still on the same runqueue.
2205                  */
2206                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
2207                         resched_curr(rq);
2208 #else
2209                 /* For UP simply resched on drop of prio */
2210                 if (oldprio < p->prio)
2211                         resched_curr(rq);
2212 #endif /* CONFIG_SMP */
2213         } else {
2214                 /*
2215                  * This task is not running, but if it is
2216                  * greater than the current running task
2217                  * then reschedule.
2218                  */
2219                 if (p->prio < rq->curr->prio)
2220                         resched_curr(rq);
2221         }
2222 }
2223
2224 static void watchdog(struct rq *rq, struct task_struct *p)
2225 {
2226         unsigned long soft, hard;
2227
2228         /* max may change after cur was read, this will be fixed next tick */
2229         soft = task_rlimit(p, RLIMIT_RTTIME);
2230         hard = task_rlimit_max(p, RLIMIT_RTTIME);
2231
2232         if (soft != RLIM_INFINITY) {
2233                 unsigned long next;
2234
2235                 if (p->rt.watchdog_stamp != jiffies) {
2236                         p->rt.timeout++;
2237                         p->rt.watchdog_stamp = jiffies;
2238                 }
2239
2240                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2241                 if (p->rt.timeout > next)
2242                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2243         }
2244 }
2245
2246 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2247 {
2248         struct sched_rt_entity *rt_se = &p->rt;
2249
2250         update_curr_rt(rq);
2251
2252         watchdog(rq, p);
2253
2254         /*
2255          * RR tasks need a special form of timeslice management.
2256          * FIFO tasks have no timeslices.
2257          */
2258         if (p->policy != SCHED_RR)
2259                 return;
2260
2261         if (--p->rt.time_slice)
2262                 return;
2263
2264         p->rt.time_slice = sched_rr_timeslice;
2265
2266         /*
2267          * Requeue to the end of queue if we (and all of our ancestors) are not
2268          * the only element on the queue
2269          */
2270         for_each_sched_rt_entity(rt_se) {
2271                 if (rt_se->run_list.prev != rt_se->run_list.next) {
2272                         requeue_task_rt(rq, p, 0);
2273                         resched_curr(rq);
2274                         return;
2275                 }
2276         }
2277 }
2278
2279 static void set_curr_task_rt(struct rq *rq)
2280 {
2281         struct task_struct *p = rq->curr;
2282
2283         p->se.exec_start = rq_clock_task(rq);
2284
2285         /* The running task is never eligible for pushing */
2286         dequeue_pushable_task(rq, p);
2287 }
2288
2289 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2290 {
2291         /*
2292          * Time slice is 0 for SCHED_FIFO tasks
2293          */
2294         if (task->policy == SCHED_RR)
2295                 return sched_rr_timeslice;
2296         else
2297                 return 0;
2298 }
2299
2300 const struct sched_class rt_sched_class = {
2301         .next                   = &fair_sched_class,
2302         .enqueue_task           = enqueue_task_rt,
2303         .dequeue_task           = dequeue_task_rt,
2304         .yield_task             = yield_task_rt,
2305
2306         .check_preempt_curr     = check_preempt_curr_rt,
2307
2308         .pick_next_task         = pick_next_task_rt,
2309         .put_prev_task          = put_prev_task_rt,
2310
2311 #ifdef CONFIG_SMP
2312         .select_task_rq         = select_task_rq_rt,
2313
2314         .set_cpus_allowed       = set_cpus_allowed_rt,
2315         .rq_online              = rq_online_rt,
2316         .rq_offline             = rq_offline_rt,
2317         .post_schedule          = post_schedule_rt,
2318         .task_woken             = task_woken_rt,
2319         .switched_from          = switched_from_rt,
2320 #endif
2321
2322         .set_curr_task          = set_curr_task_rt,
2323         .task_tick              = task_tick_rt,
2324
2325         .get_rr_interval        = get_rr_interval_rt,
2326
2327         .prio_changed           = prio_changed_rt,
2328         .switched_to            = switched_to_rt,
2329
2330         .update_curr            = update_curr_rt,
2331 };
2332
2333 #ifdef CONFIG_SCHED_DEBUG
2334 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2335
2336 void print_rt_stats(struct seq_file *m, int cpu)
2337 {
2338         rt_rq_iter_t iter;
2339         struct rt_rq *rt_rq;
2340
2341         rcu_read_lock();
2342         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2343                 print_rt_rq(m, cpu, rt_rq);
2344         rcu_read_unlock();
2345 }
2346 #endif /* CONFIG_SCHED_DEBUG */