bd2267ad404fa78de092b8bc8f228cfea2dca87b
[projects/modsched/linux.git] / kernel / sched / cfs / rt.c
1 /*
2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3  * policies)
4  */
5
6 #include "sched.h"
7
8 #include <linux/slab.h>
9
10 int sched_rr_timeslice = RR_TIMESLICE;
11
12 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
13
14 struct rt_bandwidth def_rt_bandwidth;
15
16 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
17 {
18         struct rt_bandwidth *rt_b =
19                 container_of(timer, struct rt_bandwidth, rt_period_timer);
20         ktime_t now;
21         int overrun;
22         int idle = 0;
23
24         for (;;) {
25                 now = hrtimer_cb_get_time(timer);
26                 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
27
28                 if (!overrun)
29                         break;
30
31                 idle = do_sched_rt_period_timer(rt_b, overrun);
32         }
33
34         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
35 }
36
37 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
38 {
39         rt_b->rt_period = ns_to_ktime(period);
40         rt_b->rt_runtime = runtime;
41
42         raw_spin_lock_init(&rt_b->rt_runtime_lock);
43
44         hrtimer_init(&rt_b->rt_period_timer,
45                         CLOCK_MONOTONIC, HRTIMER_MODE_REL);
46         rt_b->rt_period_timer.function = sched_rt_period_timer;
47 }
48
49 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
50 {
51         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
52                 return;
53
54         if (hrtimer_active(&rt_b->rt_period_timer))
55                 return;
56
57         raw_spin_lock(&rt_b->rt_runtime_lock);
58         start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
59         raw_spin_unlock(&rt_b->rt_runtime_lock);
60 }
61
62 void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
63 {
64         struct rt_prio_array *array;
65         int i;
66
67         array = &rt_rq->active;
68         for (i = 0; i < MAX_RT_PRIO; i++) {
69                 INIT_LIST_HEAD(array->queue + i);
70                 __clear_bit(i, array->bitmap);
71         }
72         /* delimiter for bitsearch: */
73         __set_bit(MAX_RT_PRIO, array->bitmap);
74
75 #if defined CONFIG_SMP
76         rt_rq->highest_prio.curr = MAX_RT_PRIO;
77         rt_rq->highest_prio.next = MAX_RT_PRIO;
78         rt_rq->rt_nr_migratory = 0;
79         rt_rq->overloaded = 0;
80         plist_head_init(&rt_rq->pushable_tasks);
81 #endif
82
83         rt_rq->rt_time = 0;
84         rt_rq->rt_throttled = 0;
85         rt_rq->rt_runtime = 0;
86         raw_spin_lock_init(&rt_rq->rt_runtime_lock);
87 }
88
89 #ifdef CONFIG_RT_GROUP_SCHED
90 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
91 {
92         hrtimer_cancel(&rt_b->rt_period_timer);
93 }
94
95 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
96
97 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
98 {
99 #ifdef CONFIG_SCHED_DEBUG
100         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
101 #endif
102         return container_of(rt_se, struct task_struct, rt);
103 }
104
105 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
106 {
107         return rt_rq->rq;
108 }
109
110 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
111 {
112         return rt_se->rt_rq;
113 }
114
115 void free_rt_sched_group(struct task_group *tg)
116 {
117         int i;
118
119         if (tg->rt_se)
120                 destroy_rt_bandwidth(&tg->rt_bandwidth);
121
122         for_each_possible_cpu(i) {
123                 if (tg->rt_rq)
124                         kfree(tg->rt_rq[i]);
125                 if (tg->rt_se)
126                         kfree(tg->rt_se[i]);
127         }
128
129         kfree(tg->rt_rq);
130         kfree(tg->rt_se);
131 }
132
133 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
134                 struct sched_rt_entity *rt_se, int cpu,
135                 struct sched_rt_entity *parent)
136 {
137         struct rq *rq = cpu_rq(cpu);
138
139         rt_rq->highest_prio.curr = MAX_RT_PRIO;
140         rt_rq->rt_nr_boosted = 0;
141         rt_rq->rq = rq;
142         rt_rq->tg = tg;
143
144         tg->rt_rq[cpu] = rt_rq;
145         tg->rt_se[cpu] = rt_se;
146
147         if (!rt_se)
148                 return;
149
150         if (!parent)
151                 rt_se->rt_rq = &rq->rt;
152         else
153                 rt_se->rt_rq = parent->my_q;
154
155         rt_se->my_q = rt_rq;
156         rt_se->parent = parent;
157         INIT_LIST_HEAD(&rt_se->run_list);
158 }
159
160 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
161 {
162         struct rt_rq *rt_rq;
163         struct sched_rt_entity *rt_se;
164         int i;
165
166         tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
167         if (!tg->rt_rq)
168                 goto err;
169         tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
170         if (!tg->rt_se)
171                 goto err;
172
173         init_rt_bandwidth(&tg->rt_bandwidth,
174                         ktime_to_ns(def_rt_bandwidth.rt_period), 0);
175
176         for_each_possible_cpu(i) {
177                 rt_rq = kzalloc_node(sizeof(struct rt_rq),
178                                      GFP_KERNEL, cpu_to_node(i));
179                 if (!rt_rq)
180                         goto err;
181
182                 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
183                                      GFP_KERNEL, cpu_to_node(i));
184                 if (!rt_se)
185                         goto err_free_rq;
186
187                 init_rt_rq(rt_rq, cpu_rq(i));
188                 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
189                 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
190         }
191
192         return 1;
193
194 err_free_rq:
195         kfree(rt_rq);
196 err:
197         return 0;
198 }
199
200 #else /* CONFIG_RT_GROUP_SCHED */
201
202 #define rt_entity_is_task(rt_se) (1)
203
204 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
205 {
206         return container_of(rt_se, struct task_struct, rt);
207 }
208
209 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
210 {
211         return container_of(rt_rq, struct rq, rt);
212 }
213
214 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
215 {
216         struct task_struct *p = rt_task_of(rt_se);
217         struct rq *rq = task_rq(p);
218
219         return &rq->rt;
220 }
221
222 void free_rt_sched_group(struct task_group *tg) { }
223
224 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
225 {
226         return 1;
227 }
228 #endif /* CONFIG_RT_GROUP_SCHED */
229
230 #ifdef CONFIG_SMP
231
232 static int pull_rt_task(struct rq *this_rq);
233
234 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
235 {
236         /* Try to pull RT tasks here if we lower this rq's prio */
237         return rq->rt.highest_prio.curr > prev->prio;
238 }
239
240 static inline int rt_overloaded(struct rq *rq)
241 {
242         return atomic_read(&rq->rd->rto_count);
243 }
244
245 static inline void rt_set_overload(struct rq *rq)
246 {
247         if (!rq->online)
248                 return;
249
250         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
251         /*
252          * Make sure the mask is visible before we set
253          * the overload count. That is checked to determine
254          * if we should look at the mask. It would be a shame
255          * if we looked at the mask, but the mask was not
256          * updated yet.
257          *
258          * Matched by the barrier in pull_rt_task().
259          */
260         smp_wmb();
261         atomic_inc(&rq->rd->rto_count);
262 }
263
264 static inline void rt_clear_overload(struct rq *rq)
265 {
266         if (!rq->online)
267                 return;
268
269         /* the order here really doesn't matter */
270         atomic_dec(&rq->rd->rto_count);
271         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
272 }
273
274 static void update_rt_migration(struct rt_rq *rt_rq)
275 {
276         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
277                 if (!rt_rq->overloaded) {
278                         rt_set_overload(rq_of_rt_rq(rt_rq));
279                         rt_rq->overloaded = 1;
280                 }
281         } else if (rt_rq->overloaded) {
282                 rt_clear_overload(rq_of_rt_rq(rt_rq));
283                 rt_rq->overloaded = 0;
284         }
285 }
286
287 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
288 {
289         struct task_struct *p;
290
291         if (!rt_entity_is_task(rt_se))
292                 return;
293
294         p = rt_task_of(rt_se);
295         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
296
297         rt_rq->rt_nr_total++;
298         if (p->nr_cpus_allowed > 1)
299                 rt_rq->rt_nr_migratory++;
300
301         update_rt_migration(rt_rq);
302 }
303
304 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
305 {
306         struct task_struct *p;
307
308         if (!rt_entity_is_task(rt_se))
309                 return;
310
311         p = rt_task_of(rt_se);
312         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
313
314         rt_rq->rt_nr_total--;
315         if (p->nr_cpus_allowed > 1)
316                 rt_rq->rt_nr_migratory--;
317
318         update_rt_migration(rt_rq);
319 }
320
321 static inline int has_pushable_tasks(struct rq *rq)
322 {
323         return !plist_head_empty(&rq->rt.pushable_tasks);
324 }
325
326 static inline void set_post_schedule(struct rq *rq)
327 {
328         /*
329          * We detect this state here so that we can avoid taking the RQ
330          * lock again later if there is no need to push
331          */
332         rq->post_schedule = has_pushable_tasks(rq);
333 }
334
335 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
336 {
337         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
338         plist_node_init(&p->pushable_tasks, p->prio);
339         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
340
341         /* Update the highest prio pushable task */
342         if (p->prio < rq->rt.highest_prio.next)
343                 rq->rt.highest_prio.next = p->prio;
344 }
345
346 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
347 {
348         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
349
350         /* Update the new highest prio pushable task */
351         if (has_pushable_tasks(rq)) {
352                 p = plist_first_entry(&rq->rt.pushable_tasks,
353                                       struct task_struct, pushable_tasks);
354                 rq->rt.highest_prio.next = p->prio;
355         } else
356                 rq->rt.highest_prio.next = MAX_RT_PRIO;
357 }
358
359 #else
360
361 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
362 {
363 }
364
365 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
366 {
367 }
368
369 static inline
370 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
371 {
372 }
373
374 static inline
375 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
376 {
377 }
378
379 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
380 {
381         return false;
382 }
383
384 static inline int pull_rt_task(struct rq *this_rq)
385 {
386         return 0;
387 }
388
389 static inline void set_post_schedule(struct rq *rq)
390 {
391 }
392 #endif /* CONFIG_SMP */
393
394 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
395 {
396         return !list_empty(&rt_se->run_list);
397 }
398
399 #ifdef CONFIG_RT_GROUP_SCHED
400
401 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
402 {
403         if (!rt_rq->tg)
404                 return RUNTIME_INF;
405
406         return rt_rq->rt_runtime;
407 }
408
409 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
410 {
411         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
412 }
413
414 typedef struct task_group *rt_rq_iter_t;
415
416 static inline struct task_group *next_task_group(struct task_group *tg)
417 {
418         do {
419                 tg = list_entry_rcu(tg->list.next,
420                         typeof(struct task_group), list);
421         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
422
423         if (&tg->list == &task_groups)
424                 tg = NULL;
425
426         return tg;
427 }
428
429 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
430         for (iter = container_of(&task_groups, typeof(*iter), list);    \
431                 (iter = next_task_group(iter)) &&                       \
432                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
433
434 #define for_each_sched_rt_entity(rt_se) \
435         for (; rt_se; rt_se = rt_se->parent)
436
437 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
438 {
439         return rt_se->my_q;
440 }
441
442 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
443 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
444
445 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
446 {
447         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
448         struct sched_rt_entity *rt_se;
449
450         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
451
452         rt_se = rt_rq->tg->rt_se[cpu];
453
454         if (rt_rq->rt_nr_running) {
455                 if (rt_se && !on_rt_rq(rt_se))
456                         enqueue_rt_entity(rt_se, false);
457                 if (rt_rq->highest_prio.curr < curr->prio)
458                         resched_task(curr);
459         }
460 }
461
462 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
463 {
464         struct sched_rt_entity *rt_se;
465         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
466
467         rt_se = rt_rq->tg->rt_se[cpu];
468
469         if (rt_se && on_rt_rq(rt_se))
470                 dequeue_rt_entity(rt_se);
471 }
472
473 static int rt_se_boosted(struct sched_rt_entity *rt_se)
474 {
475         struct rt_rq *rt_rq = group_rt_rq(rt_se);
476         struct task_struct *p;
477
478         if (rt_rq)
479                 return !!rt_rq->rt_nr_boosted;
480
481         p = rt_task_of(rt_se);
482         return p->prio != p->normal_prio;
483 }
484
485 #ifdef CONFIG_SMP
486 static inline const struct cpumask *sched_rt_period_mask(void)
487 {
488         return this_rq()->rd->span;
489 }
490 #else
491 static inline const struct cpumask *sched_rt_period_mask(void)
492 {
493         return cpu_online_mask;
494 }
495 #endif
496
497 static inline
498 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
499 {
500         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
501 }
502
503 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
504 {
505         return &rt_rq->tg->rt_bandwidth;
506 }
507
508 #else /* !CONFIG_RT_GROUP_SCHED */
509
510 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
511 {
512         return rt_rq->rt_runtime;
513 }
514
515 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
516 {
517         return ktime_to_ns(def_rt_bandwidth.rt_period);
518 }
519
520 typedef struct rt_rq *rt_rq_iter_t;
521
522 #define for_each_rt_rq(rt_rq, iter, rq) \
523         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
524
525 #define for_each_sched_rt_entity(rt_se) \
526         for (; rt_se; rt_se = NULL)
527
528 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
529 {
530         return NULL;
531 }
532
533 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
534 {
535         if (rt_rq->rt_nr_running)
536                 resched_task(rq_of_rt_rq(rt_rq)->curr);
537 }
538
539 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
540 {
541 }
542
543 static inline const struct cpumask *sched_rt_period_mask(void)
544 {
545         return cpu_online_mask;
546 }
547
548 static inline
549 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
550 {
551         return &cpu_rq(cpu)->rt;
552 }
553
554 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
555 {
556         return &def_rt_bandwidth;
557 }
558
559 #endif /* CONFIG_RT_GROUP_SCHED */
560
561 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
562 {
563         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
564
565         return (hrtimer_active(&rt_b->rt_period_timer) ||
566                 rt_rq->rt_time < rt_b->rt_runtime);
567 }
568
569 #ifdef CONFIG_SMP
570 /*
571  * We ran out of runtime, see if we can borrow some from our neighbours.
572  */
573 static int do_balance_runtime(struct rt_rq *rt_rq)
574 {
575         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
576         struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
577         int i, weight, more = 0;
578         u64 rt_period;
579
580         weight = cpumask_weight(rd->span);
581
582         raw_spin_lock(&rt_b->rt_runtime_lock);
583         rt_period = ktime_to_ns(rt_b->rt_period);
584         for_each_cpu(i, rd->span) {
585                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
586                 s64 diff;
587
588                 if (iter == rt_rq)
589                         continue;
590
591                 raw_spin_lock(&iter->rt_runtime_lock);
592                 /*
593                  * Either all rqs have inf runtime and there's nothing to steal
594                  * or __disable_runtime() below sets a specific rq to inf to
595                  * indicate its been disabled and disalow stealing.
596                  */
597                 if (iter->rt_runtime == RUNTIME_INF)
598                         goto next;
599
600                 /*
601                  * From runqueues with spare time, take 1/n part of their
602                  * spare time, but no more than our period.
603                  */
604                 diff = iter->rt_runtime - iter->rt_time;
605                 if (diff > 0) {
606                         diff = div_u64((u64)diff, weight);
607                         if (rt_rq->rt_runtime + diff > rt_period)
608                                 diff = rt_period - rt_rq->rt_runtime;
609                         iter->rt_runtime -= diff;
610                         rt_rq->rt_runtime += diff;
611                         more = 1;
612                         if (rt_rq->rt_runtime == rt_period) {
613                                 raw_spin_unlock(&iter->rt_runtime_lock);
614                                 break;
615                         }
616                 }
617 next:
618                 raw_spin_unlock(&iter->rt_runtime_lock);
619         }
620         raw_spin_unlock(&rt_b->rt_runtime_lock);
621
622         return more;
623 }
624
625 /*
626  * Ensure this RQ takes back all the runtime it lend to its neighbours.
627  */
628 static void __disable_runtime(struct rq *rq)
629 {
630         struct root_domain *rd = rq->rd;
631         rt_rq_iter_t iter;
632         struct rt_rq *rt_rq;
633
634         if (unlikely(!scheduler_running))
635                 return;
636
637         for_each_rt_rq(rt_rq, iter, rq) {
638                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
639                 s64 want;
640                 int i;
641
642                 raw_spin_lock(&rt_b->rt_runtime_lock);
643                 raw_spin_lock(&rt_rq->rt_runtime_lock);
644                 /*
645                  * Either we're all inf and nobody needs to borrow, or we're
646                  * already disabled and thus have nothing to do, or we have
647                  * exactly the right amount of runtime to take out.
648                  */
649                 if (rt_rq->rt_runtime == RUNTIME_INF ||
650                                 rt_rq->rt_runtime == rt_b->rt_runtime)
651                         goto balanced;
652                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
653
654                 /*
655                  * Calculate the difference between what we started out with
656                  * and what we current have, that's the amount of runtime
657                  * we lend and now have to reclaim.
658                  */
659                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
660
661                 /*
662                  * Greedy reclaim, take back as much as we can.
663                  */
664                 for_each_cpu(i, rd->span) {
665                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
666                         s64 diff;
667
668                         /*
669                          * Can't reclaim from ourselves or disabled runqueues.
670                          */
671                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
672                                 continue;
673
674                         raw_spin_lock(&iter->rt_runtime_lock);
675                         if (want > 0) {
676                                 diff = min_t(s64, iter->rt_runtime, want);
677                                 iter->rt_runtime -= diff;
678                                 want -= diff;
679                         } else {
680                                 iter->rt_runtime -= want;
681                                 want -= want;
682                         }
683                         raw_spin_unlock(&iter->rt_runtime_lock);
684
685                         if (!want)
686                                 break;
687                 }
688
689                 raw_spin_lock(&rt_rq->rt_runtime_lock);
690                 /*
691                  * We cannot be left wanting - that would mean some runtime
692                  * leaked out of the system.
693                  */
694                 BUG_ON(want);
695 balanced:
696                 /*
697                  * Disable all the borrow logic by pretending we have inf
698                  * runtime - in which case borrowing doesn't make sense.
699                  */
700                 rt_rq->rt_runtime = RUNTIME_INF;
701                 rt_rq->rt_throttled = 0;
702                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
703                 raw_spin_unlock(&rt_b->rt_runtime_lock);
704         }
705 }
706
707 static void __enable_runtime(struct rq *rq)
708 {
709         rt_rq_iter_t iter;
710         struct rt_rq *rt_rq;
711
712         if (unlikely(!scheduler_running))
713                 return;
714
715         /*
716          * Reset each runqueue's bandwidth settings
717          */
718         for_each_rt_rq(rt_rq, iter, rq) {
719                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
720
721                 raw_spin_lock(&rt_b->rt_runtime_lock);
722                 raw_spin_lock(&rt_rq->rt_runtime_lock);
723                 rt_rq->rt_runtime = rt_b->rt_runtime;
724                 rt_rq->rt_time = 0;
725                 rt_rq->rt_throttled = 0;
726                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
727                 raw_spin_unlock(&rt_b->rt_runtime_lock);
728         }
729 }
730
731 static int balance_runtime(struct rt_rq *rt_rq)
732 {
733         int more = 0;
734
735         if (!sched_feat(RT_RUNTIME_SHARE))
736                 return more;
737
738         if (rt_rq->rt_time > rt_rq->rt_runtime) {
739                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
740                 more = do_balance_runtime(rt_rq);
741                 raw_spin_lock(&rt_rq->rt_runtime_lock);
742         }
743
744         return more;
745 }
746 #else /* !CONFIG_SMP */
747 static inline int balance_runtime(struct rt_rq *rt_rq)
748 {
749         return 0;
750 }
751 #endif /* CONFIG_SMP */
752
753 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
754 {
755         int i, idle = 1, throttled = 0;
756         const struct cpumask *span;
757
758         span = sched_rt_period_mask();
759 #ifdef CONFIG_RT_GROUP_SCHED
760         /*
761          * FIXME: isolated CPUs should really leave the root task group,
762          * whether they are isolcpus or were isolated via cpusets, lest
763          * the timer run on a CPU which does not service all runqueues,
764          * potentially leaving other CPUs indefinitely throttled.  If
765          * isolation is really required, the user will turn the throttle
766          * off to kill the perturbations it causes anyway.  Meanwhile,
767          * this maintains functionality for boot and/or troubleshooting.
768          */
769         if (rt_b == &root_task_group.rt_bandwidth)
770                 span = cpu_online_mask;
771 #endif
772         for_each_cpu(i, span) {
773                 int enqueue = 0;
774                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
775                 struct rq *rq = rq_of_rt_rq(rt_rq);
776
777                 raw_spin_lock(&rq->lock);
778                 if (rt_rq->rt_time) {
779                         u64 runtime;
780
781                         raw_spin_lock(&rt_rq->rt_runtime_lock);
782                         if (rt_rq->rt_throttled)
783                                 balance_runtime(rt_rq);
784                         runtime = rt_rq->rt_runtime;
785                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
786                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
787                                 rt_rq->rt_throttled = 0;
788                                 enqueue = 1;
789
790                                 /*
791                                  * Force a clock update if the CPU was idle,
792                                  * lest wakeup -> unthrottle time accumulate.
793                                  */
794                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
795                                         rq->skip_clock_update = -1;
796                         }
797                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
798                                 idle = 0;
799                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
800                 } else if (rt_rq->rt_nr_running) {
801                         idle = 0;
802                         if (!rt_rq_throttled(rt_rq))
803                                 enqueue = 1;
804                 }
805                 if (rt_rq->rt_throttled)
806                         throttled = 1;
807
808                 if (enqueue)
809                         sched_rt_rq_enqueue(rt_rq);
810                 raw_spin_unlock(&rq->lock);
811         }
812
813         if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
814                 return 1;
815
816         return idle;
817 }
818
819 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
820 {
821 #ifdef CONFIG_RT_GROUP_SCHED
822         struct rt_rq *rt_rq = group_rt_rq(rt_se);
823
824         if (rt_rq)
825                 return rt_rq->highest_prio.curr;
826 #endif
827
828         return rt_task_of(rt_se)->prio;
829 }
830
831 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
832 {
833         u64 runtime = sched_rt_runtime(rt_rq);
834
835         if (rt_rq->rt_throttled)
836                 return rt_rq_throttled(rt_rq);
837
838         if (runtime >= sched_rt_period(rt_rq))
839                 return 0;
840
841         balance_runtime(rt_rq);
842         runtime = sched_rt_runtime(rt_rq);
843         if (runtime == RUNTIME_INF)
844                 return 0;
845
846         if (rt_rq->rt_time > runtime) {
847                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
848
849                 /*
850                  * Don't actually throttle groups that have no runtime assigned
851                  * but accrue some time due to boosting.
852                  */
853                 if (likely(rt_b->rt_runtime)) {
854                         static bool once = false;
855
856                         rt_rq->rt_throttled = 1;
857
858                         if (!once) {
859                                 once = true;
860                                 printk_sched("sched: RT throttling activated\n");
861                         }
862                 } else {
863                         /*
864                          * In case we did anyway, make it go away,
865                          * replenishment is a joke, since it will replenish us
866                          * with exactly 0 ns.
867                          */
868                         rt_rq->rt_time = 0;
869                 }
870
871                 if (rt_rq_throttled(rt_rq)) {
872                         sched_rt_rq_dequeue(rt_rq);
873                         return 1;
874                 }
875         }
876
877         return 0;
878 }
879
880 /*
881  * Update the current task's runtime statistics. Skip current tasks that
882  * are not in our scheduling class.
883  */
884 static void update_curr_rt(struct rq *rq)
885 {
886         struct task_struct *curr = rq->curr;
887         struct sched_rt_entity *rt_se = &curr->rt;
888         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
889         u64 delta_exec;
890
891         if (curr->sched_class != &rt_sched_class)
892                 return;
893
894         delta_exec = rq_clock_task(rq) - curr->se.exec_start;
895         if (unlikely((s64)delta_exec <= 0))
896                 return;
897
898         schedstat_set(curr->se.statistics.exec_max,
899                       max(curr->se.statistics.exec_max, delta_exec));
900
901         curr->se.sum_exec_runtime += delta_exec;
902         account_group_exec_runtime(curr, delta_exec);
903
904         curr->se.exec_start = rq_clock_task(rq);
905         cpuacct_charge(curr, delta_exec);
906
907         sched_rt_avg_update(rq, delta_exec);
908
909         if (!rt_bandwidth_enabled())
910                 return;
911
912         for_each_sched_rt_entity(rt_se) {
913                 rt_rq = rt_rq_of_se(rt_se);
914
915                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
916                         raw_spin_lock(&rt_rq->rt_runtime_lock);
917                         rt_rq->rt_time += delta_exec;
918                         if (sched_rt_runtime_exceeded(rt_rq))
919                                 resched_task(curr);
920                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
921                 }
922         }
923 }
924
925 #if defined CONFIG_SMP
926
927 static void
928 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
929 {
930         struct rq *rq = rq_of_rt_rq(rt_rq);
931
932 #ifdef CONFIG_RT_GROUP_SCHED
933         /*
934          * Change rq's cpupri only if rt_rq is the top queue.
935          */
936         if (&rq->rt != rt_rq)
937                 return;
938 #endif
939         if (rq->online && prio < prev_prio)
940                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
941 }
942
943 static void
944 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
945 {
946         struct rq *rq = rq_of_rt_rq(rt_rq);
947
948 #ifdef CONFIG_RT_GROUP_SCHED
949         /*
950          * Change rq's cpupri only if rt_rq is the top queue.
951          */
952         if (&rq->rt != rt_rq)
953                 return;
954 #endif
955         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
956                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
957 }
958
959 #else /* CONFIG_SMP */
960
961 static inline
962 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
963 static inline
964 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
965
966 #endif /* CONFIG_SMP */
967
968 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
969 static void
970 inc_rt_prio(struct rt_rq *rt_rq, int prio)
971 {
972         int prev_prio = rt_rq->highest_prio.curr;
973
974         if (prio < prev_prio)
975                 rt_rq->highest_prio.curr = prio;
976
977         inc_rt_prio_smp(rt_rq, prio, prev_prio);
978 }
979
980 static void
981 dec_rt_prio(struct rt_rq *rt_rq, int prio)
982 {
983         int prev_prio = rt_rq->highest_prio.curr;
984
985         if (rt_rq->rt_nr_running) {
986
987                 WARN_ON(prio < prev_prio);
988
989                 /*
990                  * This may have been our highest task, and therefore
991                  * we may have some recomputation to do
992                  */
993                 if (prio == prev_prio) {
994                         struct rt_prio_array *array = &rt_rq->active;
995
996                         rt_rq->highest_prio.curr =
997                                 sched_find_first_bit(array->bitmap);
998                 }
999
1000         } else
1001                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1002
1003         dec_rt_prio_smp(rt_rq, prio, prev_prio);
1004 }
1005
1006 #else
1007
1008 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1009 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1010
1011 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1012
1013 #ifdef CONFIG_RT_GROUP_SCHED
1014
1015 static void
1016 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1017 {
1018         if (rt_se_boosted(rt_se))
1019                 rt_rq->rt_nr_boosted++;
1020
1021         if (rt_rq->tg)
1022                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1023 }
1024
1025 static void
1026 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1027 {
1028         if (rt_se_boosted(rt_se))
1029                 rt_rq->rt_nr_boosted--;
1030
1031         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1032 }
1033
1034 #else /* CONFIG_RT_GROUP_SCHED */
1035
1036 static void
1037 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1038 {
1039         start_rt_bandwidth(&def_rt_bandwidth);
1040 }
1041
1042 static inline
1043 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1044
1045 #endif /* CONFIG_RT_GROUP_SCHED */
1046
1047 static inline
1048 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1049 {
1050         int prio = rt_se_prio(rt_se);
1051
1052         WARN_ON(!rt_prio(prio));
1053         rt_rq->rt_nr_running++;
1054
1055         inc_rt_prio(rt_rq, prio);
1056         inc_rt_migration(rt_se, rt_rq);
1057         inc_rt_group(rt_se, rt_rq);
1058 }
1059
1060 static inline
1061 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1062 {
1063         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1064         WARN_ON(!rt_rq->rt_nr_running);
1065         rt_rq->rt_nr_running--;
1066
1067         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1068         dec_rt_migration(rt_se, rt_rq);
1069         dec_rt_group(rt_se, rt_rq);
1070 }
1071
1072 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1073 {
1074         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1075         struct rt_prio_array *array = &rt_rq->active;
1076         struct rt_rq *group_rq = group_rt_rq(rt_se);
1077         struct list_head *queue = array->queue + rt_se_prio(rt_se);
1078
1079         /*
1080          * Don't enqueue the group if its throttled, or when empty.
1081          * The latter is a consequence of the former when a child group
1082          * get throttled and the current group doesn't have any other
1083          * active members.
1084          */
1085         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
1086                 return;
1087
1088         if (head)
1089                 list_add(&rt_se->run_list, queue);
1090         else
1091                 list_add_tail(&rt_se->run_list, queue);
1092         __set_bit(rt_se_prio(rt_se), array->bitmap);
1093
1094         inc_rt_tasks(rt_se, rt_rq);
1095 }
1096
1097 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
1098 {
1099         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1100         struct rt_prio_array *array = &rt_rq->active;
1101
1102         list_del_init(&rt_se->run_list);
1103         if (list_empty(array->queue + rt_se_prio(rt_se)))
1104                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1105
1106         dec_rt_tasks(rt_se, rt_rq);
1107 }
1108
1109 /*
1110  * Because the prio of an upper entry depends on the lower
1111  * entries, we must remove entries top - down.
1112  */
1113 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
1114 {
1115         struct sched_rt_entity *back = NULL;
1116
1117         for_each_sched_rt_entity(rt_se) {
1118                 rt_se->back = back;
1119                 back = rt_se;
1120         }
1121
1122         for (rt_se = back; rt_se; rt_se = rt_se->back) {
1123                 if (on_rt_rq(rt_se))
1124                         __dequeue_rt_entity(rt_se);
1125         }
1126 }
1127
1128 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1129 {
1130         dequeue_rt_stack(rt_se);
1131         for_each_sched_rt_entity(rt_se)
1132                 __enqueue_rt_entity(rt_se, head);
1133 }
1134
1135 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1136 {
1137         dequeue_rt_stack(rt_se);
1138
1139         for_each_sched_rt_entity(rt_se) {
1140                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1141
1142                 if (rt_rq && rt_rq->rt_nr_running)
1143                         __enqueue_rt_entity(rt_se, false);
1144         }
1145 }
1146
1147 /*
1148  * Adding/removing a task to/from a priority array:
1149  */
1150 static void
1151 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1152 {
1153         struct sched_rt_entity *rt_se = &p->rt;
1154
1155         if (flags & ENQUEUE_WAKEUP)
1156                 rt_se->timeout = 0;
1157
1158         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
1159
1160         if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1161                 enqueue_pushable_task(rq, p);
1162
1163         inc_nr_running(rq);
1164 }
1165
1166 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1167 {
1168         struct sched_rt_entity *rt_se = &p->rt;
1169
1170         update_curr_rt(rq);
1171         dequeue_rt_entity(rt_se);
1172
1173         dequeue_pushable_task(rq, p);
1174
1175         dec_nr_running(rq);
1176 }
1177
1178 /*
1179  * Put task to the head or the end of the run list without the overhead of
1180  * dequeue followed by enqueue.
1181  */
1182 static void
1183 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1184 {
1185         if (on_rt_rq(rt_se)) {
1186                 struct rt_prio_array *array = &rt_rq->active;
1187                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1188
1189                 if (head)
1190                         list_move(&rt_se->run_list, queue);
1191                 else
1192                         list_move_tail(&rt_se->run_list, queue);
1193         }
1194 }
1195
1196 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1197 {
1198         struct sched_rt_entity *rt_se = &p->rt;
1199         struct rt_rq *rt_rq;
1200
1201         for_each_sched_rt_entity(rt_se) {
1202                 rt_rq = rt_rq_of_se(rt_se);
1203                 requeue_rt_entity(rt_rq, rt_se, head);
1204         }
1205 }
1206
1207 static void yield_task_rt(struct rq *rq)
1208 {
1209         requeue_task_rt(rq, rq->curr, 0);
1210 }
1211
1212 #ifdef CONFIG_SMP
1213 static int find_lowest_rq(struct task_struct *task);
1214
1215 static int
1216 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1217 {
1218         struct task_struct *curr;
1219         struct rq *rq;
1220
1221         if (p->nr_cpus_allowed == 1)
1222                 goto out;
1223
1224         /* For anything but wake ups, just return the task_cpu */
1225         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1226                 goto out;
1227
1228         rq = cpu_rq(cpu);
1229
1230         rcu_read_lock();
1231         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1232
1233         /*
1234          * If the current task on @p's runqueue is an RT task, then
1235          * try to see if we can wake this RT task up on another
1236          * runqueue. Otherwise simply start this RT task
1237          * on its current runqueue.
1238          *
1239          * We want to avoid overloading runqueues. If the woken
1240          * task is a higher priority, then it will stay on this CPU
1241          * and the lower prio task should be moved to another CPU.
1242          * Even though this will probably make the lower prio task
1243          * lose its cache, we do not want to bounce a higher task
1244          * around just because it gave up its CPU, perhaps for a
1245          * lock?
1246          *
1247          * For equal prio tasks, we just let the scheduler sort it out.
1248          *
1249          * Otherwise, just let it ride on the affined RQ and the
1250          * post-schedule router will push the preempted task away
1251          *
1252          * This test is optimistic, if we get it wrong the load-balancer
1253          * will have to sort it out.
1254          */
1255         if (curr && unlikely(rt_task(curr)) &&
1256             (curr->nr_cpus_allowed < 2 ||
1257              curr->prio <= p->prio)) {
1258                 int target = find_lowest_rq(p);
1259
1260                 if (target != -1)
1261                         cpu = target;
1262         }
1263         rcu_read_unlock();
1264
1265 out:
1266         return cpu;
1267 }
1268
1269 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1270 {
1271         if (rq->curr->nr_cpus_allowed == 1)
1272                 return;
1273
1274         if (p->nr_cpus_allowed != 1
1275             && cpupri_find(&rq->rd->cpupri, p, NULL))
1276                 return;
1277
1278         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1279                 return;
1280
1281         /*
1282          * There appears to be other cpus that can accept
1283          * current and none to run 'p', so lets reschedule
1284          * to try and push current away:
1285          */
1286         requeue_task_rt(rq, p, 1);
1287         resched_task(rq->curr);
1288 }
1289
1290 #endif /* CONFIG_SMP */
1291
1292 /*
1293  * Preempt the current task with a newly woken task if needed:
1294  */
1295 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1296 {
1297         if (p->prio < rq->curr->prio) {
1298                 resched_task(rq->curr);
1299                 return;
1300         }
1301
1302 #ifdef CONFIG_SMP
1303         /*
1304          * If:
1305          *
1306          * - the newly woken task is of equal priority to the current task
1307          * - the newly woken task is non-migratable while current is migratable
1308          * - current will be preempted on the next reschedule
1309          *
1310          * we should check to see if current can readily move to a different
1311          * cpu.  If so, we will reschedule to allow the push logic to try
1312          * to move current somewhere else, making room for our non-migratable
1313          * task.
1314          */
1315         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1316                 check_preempt_equal_prio(rq, p);
1317 #endif
1318 }
1319
1320 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1321                                                    struct rt_rq *rt_rq)
1322 {
1323         struct rt_prio_array *array = &rt_rq->active;
1324         struct sched_rt_entity *next = NULL;
1325         struct list_head *queue;
1326         int idx;
1327
1328         idx = sched_find_first_bit(array->bitmap);
1329         BUG_ON(idx >= MAX_RT_PRIO);
1330
1331         queue = array->queue + idx;
1332         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1333
1334         return next;
1335 }
1336
1337 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1338 {
1339         struct sched_rt_entity *rt_se;
1340         struct task_struct *p;
1341         struct rt_rq *rt_rq  = &rq->rt;
1342
1343         do {
1344                 rt_se = pick_next_rt_entity(rq, rt_rq);
1345                 BUG_ON(!rt_se);
1346                 rt_rq = group_rt_rq(rt_se);
1347         } while (rt_rq);
1348
1349         p = rt_task_of(rt_se);
1350         p->se.exec_start = rq_clock_task(rq);
1351
1352         return p;
1353 }
1354
1355 static struct task_struct *
1356 pick_next_task_rt(struct rq *rq, struct task_struct *prev)
1357 {
1358         struct task_struct *p;
1359         struct rt_rq *rt_rq = &rq->rt;
1360
1361         if (need_pull_rt_task(rq, prev)) {
1362                 pull_rt_task(rq);
1363                 /*
1364                  * pull_rt_task() can drop (and re-acquire) rq->lock; this
1365                  * means a dl or stop task can slip in, in which case we need
1366                  * to re-start task selection.
1367                  */
1368                 if (unlikely((rq->stop && rq->stop->on_rq) ||
1369                              rq->dl.dl_nr_running))
1370                         return RETRY_TASK;
1371         }
1372
1373         /*
1374          * We may dequeue prev's rt_rq in put_prev_task().
1375          * So, we update time before rt_nr_running check.
1376          */
1377         if (prev->sched_class == &rt_sched_class)
1378                 update_curr_rt(rq);
1379
1380         if (!rt_rq->rt_nr_running)
1381                 return NULL;
1382
1383         if (rt_rq_throttled(rt_rq))
1384                 return NULL;
1385
1386         put_prev_task(rq, prev);
1387
1388         p = _pick_next_task_rt(rq);
1389
1390         /* The running task is never eligible for pushing */
1391         if (p)
1392                 dequeue_pushable_task(rq, p);
1393
1394         set_post_schedule(rq);
1395
1396         return p;
1397 }
1398
1399 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1400 {
1401         update_curr_rt(rq);
1402
1403         /*
1404          * The previous task needs to be made eligible for pushing
1405          * if it is still active
1406          */
1407         if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1408                 enqueue_pushable_task(rq, p);
1409 }
1410
1411 #ifdef CONFIG_SMP
1412
1413 /* Only try algorithms three times */
1414 #define RT_MAX_TRIES 3
1415
1416 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1417 {
1418         if (!task_running(rq, p) &&
1419             cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1420                 return 1;
1421         return 0;
1422 }
1423
1424 /*
1425  * Return the highest pushable rq's task, which is suitable to be executed
1426  * on the cpu, NULL otherwise
1427  */
1428 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1429 {
1430         struct plist_head *head = &rq->rt.pushable_tasks;
1431         struct task_struct *p;
1432
1433         if (!has_pushable_tasks(rq))
1434                 return NULL;
1435
1436         plist_for_each_entry(p, head, pushable_tasks) {
1437                 if (pick_rt_task(rq, p, cpu))
1438                         return p;
1439         }
1440
1441         return NULL;
1442 }
1443
1444 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1445
1446 static int find_lowest_rq(struct task_struct *task)
1447 {
1448         struct sched_domain *sd;
1449         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1450         int this_cpu = smp_processor_id();
1451         int cpu      = task_cpu(task);
1452
1453         /* Make sure the mask is initialized first */
1454         if (unlikely(!lowest_mask))
1455                 return -1;
1456
1457         if (task->nr_cpus_allowed == 1)
1458                 return -1; /* No other targets possible */
1459
1460         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1461                 return -1; /* No targets found */
1462
1463         /*
1464          * At this point we have built a mask of cpus representing the
1465          * lowest priority tasks in the system.  Now we want to elect
1466          * the best one based on our affinity and topology.
1467          *
1468          * We prioritize the last cpu that the task executed on since
1469          * it is most likely cache-hot in that location.
1470          */
1471         if (cpumask_test_cpu(cpu, lowest_mask))
1472                 return cpu;
1473
1474         /*
1475          * Otherwise, we consult the sched_domains span maps to figure
1476          * out which cpu is logically closest to our hot cache data.
1477          */
1478         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1479                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1480
1481         rcu_read_lock();
1482         for_each_domain(cpu, sd) {
1483                 if (sd->flags & SD_WAKE_AFFINE) {
1484                         int best_cpu;
1485
1486                         /*
1487                          * "this_cpu" is cheaper to preempt than a
1488                          * remote processor.
1489                          */
1490                         if (this_cpu != -1 &&
1491                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1492                                 rcu_read_unlock();
1493                                 return this_cpu;
1494                         }
1495
1496                         best_cpu = cpumask_first_and(lowest_mask,
1497                                                      sched_domain_span(sd));
1498                         if (best_cpu < nr_cpu_ids) {
1499                                 rcu_read_unlock();
1500                                 return best_cpu;
1501                         }
1502                 }
1503         }
1504         rcu_read_unlock();
1505
1506         /*
1507          * And finally, if there were no matches within the domains
1508          * just give the caller *something* to work with from the compatible
1509          * locations.
1510          */
1511         if (this_cpu != -1)
1512                 return this_cpu;
1513
1514         cpu = cpumask_any(lowest_mask);
1515         if (cpu < nr_cpu_ids)
1516                 return cpu;
1517         return -1;
1518 }
1519
1520 /* Will lock the rq it finds */
1521 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1522 {
1523         struct rq *lowest_rq = NULL;
1524         int tries;
1525         int cpu;
1526
1527         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1528                 cpu = find_lowest_rq(task);
1529
1530                 if ((cpu == -1) || (cpu == rq->cpu))
1531                         break;
1532
1533                 lowest_rq = cpu_rq(cpu);
1534
1535                 /* if the prio of this runqueue changed, try again */
1536                 if (double_lock_balance(rq, lowest_rq)) {
1537                         /*
1538                          * We had to unlock the run queue. In
1539                          * the mean time, task could have
1540                          * migrated already or had its affinity changed.
1541                          * Also make sure that it wasn't scheduled on its rq.
1542                          */
1543                         if (unlikely(task_rq(task) != rq ||
1544                                      !cpumask_test_cpu(lowest_rq->cpu,
1545                                                        tsk_cpus_allowed(task)) ||
1546                                      task_running(rq, task) ||
1547                                      !task->on_rq)) {
1548
1549                                 double_unlock_balance(rq, lowest_rq);
1550                                 lowest_rq = NULL;
1551                                 break;
1552                         }
1553                 }
1554
1555                 /* If this rq is still suitable use it. */
1556                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1557                         break;
1558
1559                 /* try again */
1560                 double_unlock_balance(rq, lowest_rq);
1561                 lowest_rq = NULL;
1562         }
1563
1564         return lowest_rq;
1565 }
1566
1567 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1568 {
1569         struct task_struct *p;
1570
1571         if (!has_pushable_tasks(rq))
1572                 return NULL;
1573
1574         p = plist_first_entry(&rq->rt.pushable_tasks,
1575                               struct task_struct, pushable_tasks);
1576
1577         BUG_ON(rq->cpu != task_cpu(p));
1578         BUG_ON(task_current(rq, p));
1579         BUG_ON(p->nr_cpus_allowed <= 1);
1580
1581         BUG_ON(!p->on_rq);
1582         BUG_ON(!rt_task(p));
1583
1584         return p;
1585 }
1586
1587 /*
1588  * If the current CPU has more than one RT task, see if the non
1589  * running task can migrate over to a CPU that is running a task
1590  * of lesser priority.
1591  */
1592 static int push_rt_task(struct rq *rq)
1593 {
1594         struct task_struct *next_task;
1595         struct rq *lowest_rq;
1596         int ret = 0;
1597
1598         if (!rq->rt.overloaded)
1599                 return 0;
1600
1601         next_task = pick_next_pushable_task(rq);
1602         if (!next_task)
1603                 return 0;
1604
1605 retry:
1606         if (unlikely(next_task == rq->curr)) {
1607                 WARN_ON(1);
1608                 return 0;
1609         }
1610
1611         /*
1612          * It's possible that the next_task slipped in of
1613          * higher priority than current. If that's the case
1614          * just reschedule current.
1615          */
1616         if (unlikely(next_task->prio < rq->curr->prio)) {
1617                 resched_task(rq->curr);
1618                 return 0;
1619         }
1620
1621         /* We might release rq lock */
1622         get_task_struct(next_task);
1623
1624         /* find_lock_lowest_rq locks the rq if found */
1625         lowest_rq = find_lock_lowest_rq(next_task, rq);
1626         if (!lowest_rq) {
1627                 struct task_struct *task;
1628                 /*
1629                  * find_lock_lowest_rq releases rq->lock
1630                  * so it is possible that next_task has migrated.
1631                  *
1632                  * We need to make sure that the task is still on the same
1633                  * run-queue and is also still the next task eligible for
1634                  * pushing.
1635                  */
1636                 task = pick_next_pushable_task(rq);
1637                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1638                         /*
1639                          * The task hasn't migrated, and is still the next
1640                          * eligible task, but we failed to find a run-queue
1641                          * to push it to.  Do not retry in this case, since
1642                          * other cpus will pull from us when ready.
1643                          */
1644                         goto out;
1645                 }
1646
1647                 if (!task)
1648                         /* No more tasks, just exit */
1649                         goto out;
1650
1651                 /*
1652                  * Something has shifted, try again.
1653                  */
1654                 put_task_struct(next_task);
1655                 next_task = task;
1656                 goto retry;
1657         }
1658
1659         deactivate_task(rq, next_task, 0);
1660         set_task_cpu(next_task, lowest_rq->cpu);
1661         activate_task(lowest_rq, next_task, 0);
1662         ret = 1;
1663
1664         resched_task(lowest_rq->curr);
1665
1666         double_unlock_balance(rq, lowest_rq);
1667
1668 out:
1669         put_task_struct(next_task);
1670
1671         return ret;
1672 }
1673
1674 static void push_rt_tasks(struct rq *rq)
1675 {
1676         /* push_rt_task will return true if it moved an RT */
1677         while (push_rt_task(rq))
1678                 ;
1679 }
1680
1681 static int pull_rt_task(struct rq *this_rq)
1682 {
1683         int this_cpu = this_rq->cpu, ret = 0, cpu;
1684         struct task_struct *p;
1685         struct rq *src_rq;
1686
1687         if (likely(!rt_overloaded(this_rq)))
1688                 return 0;
1689
1690         /*
1691          * Match the barrier from rt_set_overloaded; this guarantees that if we
1692          * see overloaded we must also see the rto_mask bit.
1693          */
1694         smp_rmb();
1695
1696         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1697                 if (this_cpu == cpu)
1698                         continue;
1699
1700                 src_rq = cpu_rq(cpu);
1701
1702                 /*
1703                  * Don't bother taking the src_rq->lock if the next highest
1704                  * task is known to be lower-priority than our current task.
1705                  * This may look racy, but if this value is about to go
1706                  * logically higher, the src_rq will push this task away.
1707                  * And if its going logically lower, we do not care
1708                  */
1709                 if (src_rq->rt.highest_prio.next >=
1710                     this_rq->rt.highest_prio.curr)
1711                         continue;
1712
1713                 /*
1714                  * We can potentially drop this_rq's lock in
1715                  * double_lock_balance, and another CPU could
1716                  * alter this_rq
1717                  */
1718                 double_lock_balance(this_rq, src_rq);
1719
1720                 /*
1721                  * We can pull only a task, which is pushable
1722                  * on its rq, and no others.
1723                  */
1724                 p = pick_highest_pushable_task(src_rq, this_cpu);
1725
1726                 /*
1727                  * Do we have an RT task that preempts
1728                  * the to-be-scheduled task?
1729                  */
1730                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1731                         WARN_ON(p == src_rq->curr);
1732                         WARN_ON(!p->on_rq);
1733
1734                         /*
1735                          * There's a chance that p is higher in priority
1736                          * than what's currently running on its cpu.
1737                          * This is just that p is wakeing up and hasn't
1738                          * had a chance to schedule. We only pull
1739                          * p if it is lower in priority than the
1740                          * current task on the run queue
1741                          */
1742                         if (p->prio < src_rq->curr->prio)
1743                                 goto skip;
1744
1745                         ret = 1;
1746
1747                         deactivate_task(src_rq, p, 0);
1748                         set_task_cpu(p, this_cpu);
1749                         activate_task(this_rq, p, 0);
1750                         /*
1751                          * We continue with the search, just in
1752                          * case there's an even higher prio task
1753                          * in another runqueue. (low likelihood
1754                          * but possible)
1755                          */
1756                 }
1757 skip:
1758                 double_unlock_balance(this_rq, src_rq);
1759         }
1760
1761         return ret;
1762 }
1763
1764 static void post_schedule_rt(struct rq *rq)
1765 {
1766         push_rt_tasks(rq);
1767 }
1768
1769 /*
1770  * If we are not running and we are not going to reschedule soon, we should
1771  * try to push tasks away now
1772  */
1773 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1774 {
1775         if (!task_running(rq, p) &&
1776             !test_tsk_need_resched(rq->curr) &&
1777             has_pushable_tasks(rq) &&
1778             p->nr_cpus_allowed > 1 &&
1779             (dl_task(rq->curr) || rt_task(rq->curr)) &&
1780             (rq->curr->nr_cpus_allowed < 2 ||
1781              rq->curr->prio <= p->prio))
1782                 push_rt_tasks(rq);
1783 }
1784
1785 static void set_cpus_allowed_rt(struct task_struct *p,
1786                                 const struct cpumask *new_mask)
1787 {
1788         struct rq *rq;
1789         int weight;
1790
1791         BUG_ON(!rt_task(p));
1792
1793         if (!p->on_rq)
1794                 return;
1795
1796         weight = cpumask_weight(new_mask);
1797
1798         /*
1799          * Only update if the process changes its state from whether it
1800          * can migrate or not.
1801          */
1802         if ((p->nr_cpus_allowed > 1) == (weight > 1))
1803                 return;
1804
1805         rq = task_rq(p);
1806
1807         /*
1808          * The process used to be able to migrate OR it can now migrate
1809          */
1810         if (weight <= 1) {
1811                 if (!task_current(rq, p))
1812                         dequeue_pushable_task(rq, p);
1813                 BUG_ON(!rq->rt.rt_nr_migratory);
1814                 rq->rt.rt_nr_migratory--;
1815         } else {
1816                 if (!task_current(rq, p))
1817                         enqueue_pushable_task(rq, p);
1818                 rq->rt.rt_nr_migratory++;
1819         }
1820
1821         update_rt_migration(&rq->rt);
1822 }
1823
1824 /* Assumes rq->lock is held */
1825 static void rq_online_rt(struct rq *rq)
1826 {
1827         if (rq->rt.overloaded)
1828                 rt_set_overload(rq);
1829
1830         __enable_runtime(rq);
1831
1832         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1833 }
1834
1835 /* Assumes rq->lock is held */
1836 static void rq_offline_rt(struct rq *rq)
1837 {
1838         if (rq->rt.overloaded)
1839                 rt_clear_overload(rq);
1840
1841         __disable_runtime(rq);
1842
1843         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1844 }
1845
1846 /*
1847  * When switch from the rt queue, we bring ourselves to a position
1848  * that we might want to pull RT tasks from other runqueues.
1849  */
1850 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1851 {
1852         /*
1853          * If there are other RT tasks then we will reschedule
1854          * and the scheduling of the other RT tasks will handle
1855          * the balancing. But if we are the last RT task
1856          * we may need to handle the pulling of RT tasks
1857          * now.
1858          */
1859         if (!p->on_rq || rq->rt.rt_nr_running)
1860                 return;
1861
1862         if (pull_rt_task(rq))
1863                 resched_task(rq->curr);
1864 }
1865
1866 void __init init_sched_rt_class(void)
1867 {
1868         unsigned int i;
1869
1870         for_each_possible_cpu(i) {
1871                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1872                                         GFP_KERNEL, cpu_to_node(i));
1873         }
1874 }
1875 #endif /* CONFIG_SMP */
1876
1877 /*
1878  * When switching a task to RT, we may overload the runqueue
1879  * with RT tasks. In this case we try to push them off to
1880  * other runqueues.
1881  */
1882 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1883 {
1884         int check_resched = 1;
1885
1886         /*
1887          * If we are already running, then there's nothing
1888          * that needs to be done. But if we are not running
1889          * we may need to preempt the current running task.
1890          * If that current running task is also an RT task
1891          * then see if we can move to another run queue.
1892          */
1893         if (p->on_rq && rq->curr != p) {
1894 #ifdef CONFIG_SMP
1895                 if (rq->rt.overloaded && push_rt_task(rq) &&
1896                     /* Don't resched if we changed runqueues */
1897                     rq != task_rq(p))
1898                         check_resched = 0;
1899 #endif /* CONFIG_SMP */
1900                 if (check_resched && p->prio < rq->curr->prio)
1901                         resched_task(rq->curr);
1902         }
1903 }
1904
1905 /*
1906  * Priority of the task has changed. This may cause
1907  * us to initiate a push or pull.
1908  */
1909 static void
1910 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1911 {
1912         if (!p->on_rq)
1913                 return;
1914
1915         if (rq->curr == p) {
1916 #ifdef CONFIG_SMP
1917                 /*
1918                  * If our priority decreases while running, we
1919                  * may need to pull tasks to this runqueue.
1920                  */
1921                 if (oldprio < p->prio)
1922                         pull_rt_task(rq);
1923                 /*
1924                  * If there's a higher priority task waiting to run
1925                  * then reschedule. Note, the above pull_rt_task
1926                  * can release the rq lock and p could migrate.
1927                  * Only reschedule if p is still on the same runqueue.
1928                  */
1929                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1930                         resched_task(p);
1931 #else
1932                 /* For UP simply resched on drop of prio */
1933                 if (oldprio < p->prio)
1934                         resched_task(p);
1935 #endif /* CONFIG_SMP */
1936         } else {
1937                 /*
1938                  * This task is not running, but if it is
1939                  * greater than the current running task
1940                  * then reschedule.
1941                  */
1942                 if (p->prio < rq->curr->prio)
1943                         resched_task(rq->curr);
1944         }
1945 }
1946
1947 static void watchdog(struct rq *rq, struct task_struct *p)
1948 {
1949         unsigned long soft, hard;
1950
1951         /* max may change after cur was read, this will be fixed next tick */
1952         soft = task_rlimit(p, RLIMIT_RTTIME);
1953         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1954
1955         if (soft != RLIM_INFINITY) {
1956                 unsigned long next;
1957
1958                 if (p->rt.watchdog_stamp != jiffies) {
1959                         p->rt.timeout++;
1960                         p->rt.watchdog_stamp = jiffies;
1961                 }
1962
1963                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1964                 if (p->rt.timeout > next)
1965                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1966         }
1967 }
1968
1969 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1970 {
1971         struct sched_rt_entity *rt_se = &p->rt;
1972
1973         update_curr_rt(rq);
1974
1975         watchdog(rq, p);
1976
1977         /*
1978          * RR tasks need a special form of timeslice management.
1979          * FIFO tasks have no timeslices.
1980          */
1981         if (p->policy != SCHED_RR)
1982                 return;
1983
1984         if (--p->rt.time_slice)
1985                 return;
1986
1987         p->rt.time_slice = sched_rr_timeslice;
1988
1989         /*
1990          * Requeue to the end of queue if we (and all of our ancestors) are not
1991          * the only element on the queue
1992          */
1993         for_each_sched_rt_entity(rt_se) {
1994                 if (rt_se->run_list.prev != rt_se->run_list.next) {
1995                         requeue_task_rt(rq, p, 0);
1996                         set_tsk_need_resched(p);
1997                         return;
1998                 }
1999         }
2000 }
2001
2002 static void set_curr_task_rt(struct rq *rq)
2003 {
2004         struct task_struct *p = rq->curr;
2005
2006         p->se.exec_start = rq_clock_task(rq);
2007
2008         /* The running task is never eligible for pushing */
2009         dequeue_pushable_task(rq, p);
2010 }
2011
2012 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2013 {
2014         /*
2015          * Time slice is 0 for SCHED_FIFO tasks
2016          */
2017         if (task->policy == SCHED_RR)
2018                 return sched_rr_timeslice;
2019         else
2020                 return 0;
2021 }
2022
2023 const struct sched_class rt_sched_class = {
2024         .next                   = &fair_sched_class,
2025         .enqueue_task           = enqueue_task_rt,
2026         .dequeue_task           = dequeue_task_rt,
2027         .yield_task             = yield_task_rt,
2028
2029         .check_preempt_curr     = check_preempt_curr_rt,
2030
2031         .pick_next_task         = pick_next_task_rt,
2032         .put_prev_task          = put_prev_task_rt,
2033
2034 #ifdef CONFIG_SMP
2035         .select_task_rq         = select_task_rq_rt,
2036
2037         .set_cpus_allowed       = set_cpus_allowed_rt,
2038         .rq_online              = rq_online_rt,
2039         .rq_offline             = rq_offline_rt,
2040         .post_schedule          = post_schedule_rt,
2041         .task_woken             = task_woken_rt,
2042         .switched_from          = switched_from_rt,
2043 #endif
2044
2045         .set_curr_task          = set_curr_task_rt,
2046         .task_tick              = task_tick_rt,
2047
2048         .get_rr_interval        = get_rr_interval_rt,
2049
2050         .prio_changed           = prio_changed_rt,
2051         .switched_to            = switched_to_rt,
2052 };
2053
2054 #ifdef CONFIG_SCHED_DEBUG
2055 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2056
2057 void print_rt_stats(struct seq_file *m, int cpu)
2058 {
2059         rt_rq_iter_t iter;
2060         struct rt_rq *rt_rq;
2061
2062         rcu_read_lock();
2063         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2064                 print_rt_rq(m, cpu, rt_rq);
2065         rcu_read_unlock();
2066 }
2067 #endif /* CONFIG_SCHED_DEBUG */