sched/rt: Make update_curr_rt() more accurate
[sfrench/cifs-2.6.git] / kernel / sched / rt.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
4  * policies)
5  */
6
7 #include "sched.h"
8
9 #include <linux/slab.h>
10 #include <linux/irq_work.h>
11
12 int sched_rr_timeslice = RR_TIMESLICE;
13 int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
14
15 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
16
17 struct rt_bandwidth def_rt_bandwidth;
18
19 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
20 {
21         struct rt_bandwidth *rt_b =
22                 container_of(timer, struct rt_bandwidth, rt_period_timer);
23         int idle = 0;
24         int overrun;
25
26         raw_spin_lock(&rt_b->rt_runtime_lock);
27         for (;;) {
28                 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
29                 if (!overrun)
30                         break;
31
32                 raw_spin_unlock(&rt_b->rt_runtime_lock);
33                 idle = do_sched_rt_period_timer(rt_b, overrun);
34                 raw_spin_lock(&rt_b->rt_runtime_lock);
35         }
36         if (idle)
37                 rt_b->rt_period_active = 0;
38         raw_spin_unlock(&rt_b->rt_runtime_lock);
39
40         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
41 }
42
43 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
44 {
45         rt_b->rt_period = ns_to_ktime(period);
46         rt_b->rt_runtime = runtime;
47
48         raw_spin_lock_init(&rt_b->rt_runtime_lock);
49
50         hrtimer_init(&rt_b->rt_period_timer,
51                         CLOCK_MONOTONIC, HRTIMER_MODE_REL);
52         rt_b->rt_period_timer.function = sched_rt_period_timer;
53 }
54
55 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
56 {
57         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
58                 return;
59
60         raw_spin_lock(&rt_b->rt_runtime_lock);
61         if (!rt_b->rt_period_active) {
62                 rt_b->rt_period_active = 1;
63                 /*
64                  * SCHED_DEADLINE updates the bandwidth, as a run away
65                  * RT task with a DL task could hog a CPU. But DL does
66                  * not reset the period. If a deadline task was running
67                  * without an RT task running, it can cause RT tasks to
68                  * throttle when they start up. Kick the timer right away
69                  * to update the period.
70                  */
71                 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
72                 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
73         }
74         raw_spin_unlock(&rt_b->rt_runtime_lock);
75 }
76
77 void init_rt_rq(struct rt_rq *rt_rq)
78 {
79         struct rt_prio_array *array;
80         int i;
81
82         array = &rt_rq->active;
83         for (i = 0; i < MAX_RT_PRIO; i++) {
84                 INIT_LIST_HEAD(array->queue + i);
85                 __clear_bit(i, array->bitmap);
86         }
87         /* delimiter for bitsearch: */
88         __set_bit(MAX_RT_PRIO, array->bitmap);
89
90 #if defined CONFIG_SMP
91         rt_rq->highest_prio.curr = MAX_RT_PRIO;
92         rt_rq->highest_prio.next = MAX_RT_PRIO;
93         rt_rq->rt_nr_migratory = 0;
94         rt_rq->overloaded = 0;
95         plist_head_init(&rt_rq->pushable_tasks);
96 #endif /* CONFIG_SMP */
97         /* We start is dequeued state, because no RT tasks are queued */
98         rt_rq->rt_queued = 0;
99
100         rt_rq->rt_time = 0;
101         rt_rq->rt_throttled = 0;
102         rt_rq->rt_runtime = 0;
103         raw_spin_lock_init(&rt_rq->rt_runtime_lock);
104 }
105
106 #ifdef CONFIG_RT_GROUP_SCHED
107 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
108 {
109         hrtimer_cancel(&rt_b->rt_period_timer);
110 }
111
112 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
113
114 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
115 {
116 #ifdef CONFIG_SCHED_DEBUG
117         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
118 #endif
119         return container_of(rt_se, struct task_struct, rt);
120 }
121
122 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
123 {
124         return rt_rq->rq;
125 }
126
127 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
128 {
129         return rt_se->rt_rq;
130 }
131
132 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
133 {
134         struct rt_rq *rt_rq = rt_se->rt_rq;
135
136         return rt_rq->rq;
137 }
138
139 void free_rt_sched_group(struct task_group *tg)
140 {
141         int i;
142
143         if (tg->rt_se)
144                 destroy_rt_bandwidth(&tg->rt_bandwidth);
145
146         for_each_possible_cpu(i) {
147                 if (tg->rt_rq)
148                         kfree(tg->rt_rq[i]);
149                 if (tg->rt_se)
150                         kfree(tg->rt_se[i]);
151         }
152
153         kfree(tg->rt_rq);
154         kfree(tg->rt_se);
155 }
156
157 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
158                 struct sched_rt_entity *rt_se, int cpu,
159                 struct sched_rt_entity *parent)
160 {
161         struct rq *rq = cpu_rq(cpu);
162
163         rt_rq->highest_prio.curr = MAX_RT_PRIO;
164         rt_rq->rt_nr_boosted = 0;
165         rt_rq->rq = rq;
166         rt_rq->tg = tg;
167
168         tg->rt_rq[cpu] = rt_rq;
169         tg->rt_se[cpu] = rt_se;
170
171         if (!rt_se)
172                 return;
173
174         if (!parent)
175                 rt_se->rt_rq = &rq->rt;
176         else
177                 rt_se->rt_rq = parent->my_q;
178
179         rt_se->my_q = rt_rq;
180         rt_se->parent = parent;
181         INIT_LIST_HEAD(&rt_se->run_list);
182 }
183
184 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
185 {
186         struct rt_rq *rt_rq;
187         struct sched_rt_entity *rt_se;
188         int i;
189
190         tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
191         if (!tg->rt_rq)
192                 goto err;
193         tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
194         if (!tg->rt_se)
195                 goto err;
196
197         init_rt_bandwidth(&tg->rt_bandwidth,
198                         ktime_to_ns(def_rt_bandwidth.rt_period), 0);
199
200         for_each_possible_cpu(i) {
201                 rt_rq = kzalloc_node(sizeof(struct rt_rq),
202                                      GFP_KERNEL, cpu_to_node(i));
203                 if (!rt_rq)
204                         goto err;
205
206                 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
207                                      GFP_KERNEL, cpu_to_node(i));
208                 if (!rt_se)
209                         goto err_free_rq;
210
211                 init_rt_rq(rt_rq);
212                 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
213                 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
214         }
215
216         return 1;
217
218 err_free_rq:
219         kfree(rt_rq);
220 err:
221         return 0;
222 }
223
224 #else /* CONFIG_RT_GROUP_SCHED */
225
226 #define rt_entity_is_task(rt_se) (1)
227
228 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
229 {
230         return container_of(rt_se, struct task_struct, rt);
231 }
232
233 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
234 {
235         return container_of(rt_rq, struct rq, rt);
236 }
237
238 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
239 {
240         struct task_struct *p = rt_task_of(rt_se);
241
242         return task_rq(p);
243 }
244
245 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
246 {
247         struct rq *rq = rq_of_rt_se(rt_se);
248
249         return &rq->rt;
250 }
251
252 void free_rt_sched_group(struct task_group *tg) { }
253
254 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
255 {
256         return 1;
257 }
258 #endif /* CONFIG_RT_GROUP_SCHED */
259
260 #ifdef CONFIG_SMP
261
262 static void pull_rt_task(struct rq *this_rq);
263
264 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
265 {
266         /* Try to pull RT tasks here if we lower this rq's prio */
267         return rq->rt.highest_prio.curr > prev->prio;
268 }
269
270 static inline int rt_overloaded(struct rq *rq)
271 {
272         return atomic_read(&rq->rd->rto_count);
273 }
274
275 static inline void rt_set_overload(struct rq *rq)
276 {
277         if (!rq->online)
278                 return;
279
280         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
281         /*
282          * Make sure the mask is visible before we set
283          * the overload count. That is checked to determine
284          * if we should look at the mask. It would be a shame
285          * if we looked at the mask, but the mask was not
286          * updated yet.
287          *
288          * Matched by the barrier in pull_rt_task().
289          */
290         smp_wmb();
291         atomic_inc(&rq->rd->rto_count);
292 }
293
294 static inline void rt_clear_overload(struct rq *rq)
295 {
296         if (!rq->online)
297                 return;
298
299         /* the order here really doesn't matter */
300         atomic_dec(&rq->rd->rto_count);
301         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
302 }
303
304 static void update_rt_migration(struct rt_rq *rt_rq)
305 {
306         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
307                 if (!rt_rq->overloaded) {
308                         rt_set_overload(rq_of_rt_rq(rt_rq));
309                         rt_rq->overloaded = 1;
310                 }
311         } else if (rt_rq->overloaded) {
312                 rt_clear_overload(rq_of_rt_rq(rt_rq));
313                 rt_rq->overloaded = 0;
314         }
315 }
316
317 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
318 {
319         struct task_struct *p;
320
321         if (!rt_entity_is_task(rt_se))
322                 return;
323
324         p = rt_task_of(rt_se);
325         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
326
327         rt_rq->rt_nr_total++;
328         if (p->nr_cpus_allowed > 1)
329                 rt_rq->rt_nr_migratory++;
330
331         update_rt_migration(rt_rq);
332 }
333
334 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
335 {
336         struct task_struct *p;
337
338         if (!rt_entity_is_task(rt_se))
339                 return;
340
341         p = rt_task_of(rt_se);
342         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
343
344         rt_rq->rt_nr_total--;
345         if (p->nr_cpus_allowed > 1)
346                 rt_rq->rt_nr_migratory--;
347
348         update_rt_migration(rt_rq);
349 }
350
351 static inline int has_pushable_tasks(struct rq *rq)
352 {
353         return !plist_head_empty(&rq->rt.pushable_tasks);
354 }
355
356 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
357 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
358
359 static void push_rt_tasks(struct rq *);
360 static void pull_rt_task(struct rq *);
361
362 static inline void queue_push_tasks(struct rq *rq)
363 {
364         if (!has_pushable_tasks(rq))
365                 return;
366
367         queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
368 }
369
370 static inline void queue_pull_task(struct rq *rq)
371 {
372         queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
373 }
374
375 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
376 {
377         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
378         plist_node_init(&p->pushable_tasks, p->prio);
379         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
380
381         /* Update the highest prio pushable task */
382         if (p->prio < rq->rt.highest_prio.next)
383                 rq->rt.highest_prio.next = p->prio;
384 }
385
386 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
387 {
388         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
389
390         /* Update the new highest prio pushable task */
391         if (has_pushable_tasks(rq)) {
392                 p = plist_first_entry(&rq->rt.pushable_tasks,
393                                       struct task_struct, pushable_tasks);
394                 rq->rt.highest_prio.next = p->prio;
395         } else
396                 rq->rt.highest_prio.next = MAX_RT_PRIO;
397 }
398
399 #else
400
401 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
402 {
403 }
404
405 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
406 {
407 }
408
409 static inline
410 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
411 {
412 }
413
414 static inline
415 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
416 {
417 }
418
419 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
420 {
421         return false;
422 }
423
424 static inline void pull_rt_task(struct rq *this_rq)
425 {
426 }
427
428 static inline void queue_push_tasks(struct rq *rq)
429 {
430 }
431 #endif /* CONFIG_SMP */
432
433 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
434 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
435
436 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
437 {
438         return rt_se->on_rq;
439 }
440
441 #ifdef CONFIG_RT_GROUP_SCHED
442
443 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
444 {
445         if (!rt_rq->tg)
446                 return RUNTIME_INF;
447
448         return rt_rq->rt_runtime;
449 }
450
451 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
452 {
453         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
454 }
455
456 typedef struct task_group *rt_rq_iter_t;
457
458 static inline struct task_group *next_task_group(struct task_group *tg)
459 {
460         do {
461                 tg = list_entry_rcu(tg->list.next,
462                         typeof(struct task_group), list);
463         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
464
465         if (&tg->list == &task_groups)
466                 tg = NULL;
467
468         return tg;
469 }
470
471 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
472         for (iter = container_of(&task_groups, typeof(*iter), list);    \
473                 (iter = next_task_group(iter)) &&                       \
474                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
475
476 #define for_each_sched_rt_entity(rt_se) \
477         for (; rt_se; rt_se = rt_se->parent)
478
479 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
480 {
481         return rt_se->my_q;
482 }
483
484 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
485 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
486
487 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
488 {
489         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
490         struct rq *rq = rq_of_rt_rq(rt_rq);
491         struct sched_rt_entity *rt_se;
492
493         int cpu = cpu_of(rq);
494
495         rt_se = rt_rq->tg->rt_se[cpu];
496
497         if (rt_rq->rt_nr_running) {
498                 if (!rt_se)
499                         enqueue_top_rt_rq(rt_rq);
500                 else if (!on_rt_rq(rt_se))
501                         enqueue_rt_entity(rt_se, 0);
502
503                 if (rt_rq->highest_prio.curr < curr->prio)
504                         resched_curr(rq);
505         }
506 }
507
508 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
509 {
510         struct sched_rt_entity *rt_se;
511         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
512
513         rt_se = rt_rq->tg->rt_se[cpu];
514
515         if (!rt_se)
516                 dequeue_top_rt_rq(rt_rq);
517         else if (on_rt_rq(rt_se))
518                 dequeue_rt_entity(rt_se, 0);
519 }
520
521 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
522 {
523         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
524 }
525
526 static int rt_se_boosted(struct sched_rt_entity *rt_se)
527 {
528         struct rt_rq *rt_rq = group_rt_rq(rt_se);
529         struct task_struct *p;
530
531         if (rt_rq)
532                 return !!rt_rq->rt_nr_boosted;
533
534         p = rt_task_of(rt_se);
535         return p->prio != p->normal_prio;
536 }
537
538 #ifdef CONFIG_SMP
539 static inline const struct cpumask *sched_rt_period_mask(void)
540 {
541         return this_rq()->rd->span;
542 }
543 #else
544 static inline const struct cpumask *sched_rt_period_mask(void)
545 {
546         return cpu_online_mask;
547 }
548 #endif
549
550 static inline
551 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
552 {
553         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
554 }
555
556 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
557 {
558         return &rt_rq->tg->rt_bandwidth;
559 }
560
561 #else /* !CONFIG_RT_GROUP_SCHED */
562
563 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
564 {
565         return rt_rq->rt_runtime;
566 }
567
568 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
569 {
570         return ktime_to_ns(def_rt_bandwidth.rt_period);
571 }
572
573 typedef struct rt_rq *rt_rq_iter_t;
574
575 #define for_each_rt_rq(rt_rq, iter, rq) \
576         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
577
578 #define for_each_sched_rt_entity(rt_se) \
579         for (; rt_se; rt_se = NULL)
580
581 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
582 {
583         return NULL;
584 }
585
586 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
587 {
588         struct rq *rq = rq_of_rt_rq(rt_rq);
589
590         if (!rt_rq->rt_nr_running)
591                 return;
592
593         enqueue_top_rt_rq(rt_rq);
594         resched_curr(rq);
595 }
596
597 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
598 {
599         dequeue_top_rt_rq(rt_rq);
600 }
601
602 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
603 {
604         return rt_rq->rt_throttled;
605 }
606
607 static inline const struct cpumask *sched_rt_period_mask(void)
608 {
609         return cpu_online_mask;
610 }
611
612 static inline
613 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
614 {
615         return &cpu_rq(cpu)->rt;
616 }
617
618 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
619 {
620         return &def_rt_bandwidth;
621 }
622
623 #endif /* CONFIG_RT_GROUP_SCHED */
624
625 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
626 {
627         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
628
629         return (hrtimer_active(&rt_b->rt_period_timer) ||
630                 rt_rq->rt_time < rt_b->rt_runtime);
631 }
632
633 #ifdef CONFIG_SMP
634 /*
635  * We ran out of runtime, see if we can borrow some from our neighbours.
636  */
637 static void do_balance_runtime(struct rt_rq *rt_rq)
638 {
639         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
640         struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
641         int i, weight;
642         u64 rt_period;
643
644         weight = cpumask_weight(rd->span);
645
646         raw_spin_lock(&rt_b->rt_runtime_lock);
647         rt_period = ktime_to_ns(rt_b->rt_period);
648         for_each_cpu(i, rd->span) {
649                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
650                 s64 diff;
651
652                 if (iter == rt_rq)
653                         continue;
654
655                 raw_spin_lock(&iter->rt_runtime_lock);
656                 /*
657                  * Either all rqs have inf runtime and there's nothing to steal
658                  * or __disable_runtime() below sets a specific rq to inf to
659                  * indicate its been disabled and disalow stealing.
660                  */
661                 if (iter->rt_runtime == RUNTIME_INF)
662                         goto next;
663
664                 /*
665                  * From runqueues with spare time, take 1/n part of their
666                  * spare time, but no more than our period.
667                  */
668                 diff = iter->rt_runtime - iter->rt_time;
669                 if (diff > 0) {
670                         diff = div_u64((u64)diff, weight);
671                         if (rt_rq->rt_runtime + diff > rt_period)
672                                 diff = rt_period - rt_rq->rt_runtime;
673                         iter->rt_runtime -= diff;
674                         rt_rq->rt_runtime += diff;
675                         if (rt_rq->rt_runtime == rt_period) {
676                                 raw_spin_unlock(&iter->rt_runtime_lock);
677                                 break;
678                         }
679                 }
680 next:
681                 raw_spin_unlock(&iter->rt_runtime_lock);
682         }
683         raw_spin_unlock(&rt_b->rt_runtime_lock);
684 }
685
686 /*
687  * Ensure this RQ takes back all the runtime it lend to its neighbours.
688  */
689 static void __disable_runtime(struct rq *rq)
690 {
691         struct root_domain *rd = rq->rd;
692         rt_rq_iter_t iter;
693         struct rt_rq *rt_rq;
694
695         if (unlikely(!scheduler_running))
696                 return;
697
698         for_each_rt_rq(rt_rq, iter, rq) {
699                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
700                 s64 want;
701                 int i;
702
703                 raw_spin_lock(&rt_b->rt_runtime_lock);
704                 raw_spin_lock(&rt_rq->rt_runtime_lock);
705                 /*
706                  * Either we're all inf and nobody needs to borrow, or we're
707                  * already disabled and thus have nothing to do, or we have
708                  * exactly the right amount of runtime to take out.
709                  */
710                 if (rt_rq->rt_runtime == RUNTIME_INF ||
711                                 rt_rq->rt_runtime == rt_b->rt_runtime)
712                         goto balanced;
713                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
714
715                 /*
716                  * Calculate the difference between what we started out with
717                  * and what we current have, that's the amount of runtime
718                  * we lend and now have to reclaim.
719                  */
720                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
721
722                 /*
723                  * Greedy reclaim, take back as much as we can.
724                  */
725                 for_each_cpu(i, rd->span) {
726                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
727                         s64 diff;
728
729                         /*
730                          * Can't reclaim from ourselves or disabled runqueues.
731                          */
732                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
733                                 continue;
734
735                         raw_spin_lock(&iter->rt_runtime_lock);
736                         if (want > 0) {
737                                 diff = min_t(s64, iter->rt_runtime, want);
738                                 iter->rt_runtime -= diff;
739                                 want -= diff;
740                         } else {
741                                 iter->rt_runtime -= want;
742                                 want -= want;
743                         }
744                         raw_spin_unlock(&iter->rt_runtime_lock);
745
746                         if (!want)
747                                 break;
748                 }
749
750                 raw_spin_lock(&rt_rq->rt_runtime_lock);
751                 /*
752                  * We cannot be left wanting - that would mean some runtime
753                  * leaked out of the system.
754                  */
755                 BUG_ON(want);
756 balanced:
757                 /*
758                  * Disable all the borrow logic by pretending we have inf
759                  * runtime - in which case borrowing doesn't make sense.
760                  */
761                 rt_rq->rt_runtime = RUNTIME_INF;
762                 rt_rq->rt_throttled = 0;
763                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
764                 raw_spin_unlock(&rt_b->rt_runtime_lock);
765
766                 /* Make rt_rq available for pick_next_task() */
767                 sched_rt_rq_enqueue(rt_rq);
768         }
769 }
770
771 static void __enable_runtime(struct rq *rq)
772 {
773         rt_rq_iter_t iter;
774         struct rt_rq *rt_rq;
775
776         if (unlikely(!scheduler_running))
777                 return;
778
779         /*
780          * Reset each runqueue's bandwidth settings
781          */
782         for_each_rt_rq(rt_rq, iter, rq) {
783                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
784
785                 raw_spin_lock(&rt_b->rt_runtime_lock);
786                 raw_spin_lock(&rt_rq->rt_runtime_lock);
787                 rt_rq->rt_runtime = rt_b->rt_runtime;
788                 rt_rq->rt_time = 0;
789                 rt_rq->rt_throttled = 0;
790                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
791                 raw_spin_unlock(&rt_b->rt_runtime_lock);
792         }
793 }
794
795 static void balance_runtime(struct rt_rq *rt_rq)
796 {
797         if (!sched_feat(RT_RUNTIME_SHARE))
798                 return;
799
800         if (rt_rq->rt_time > rt_rq->rt_runtime) {
801                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
802                 do_balance_runtime(rt_rq);
803                 raw_spin_lock(&rt_rq->rt_runtime_lock);
804         }
805 }
806 #else /* !CONFIG_SMP */
807 static inline void balance_runtime(struct rt_rq *rt_rq) {}
808 #endif /* CONFIG_SMP */
809
810 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
811 {
812         int i, idle = 1, throttled = 0;
813         const struct cpumask *span;
814
815         span = sched_rt_period_mask();
816 #ifdef CONFIG_RT_GROUP_SCHED
817         /*
818          * FIXME: isolated CPUs should really leave the root task group,
819          * whether they are isolcpus or were isolated via cpusets, lest
820          * the timer run on a CPU which does not service all runqueues,
821          * potentially leaving other CPUs indefinitely throttled.  If
822          * isolation is really required, the user will turn the throttle
823          * off to kill the perturbations it causes anyway.  Meanwhile,
824          * this maintains functionality for boot and/or troubleshooting.
825          */
826         if (rt_b == &root_task_group.rt_bandwidth)
827                 span = cpu_online_mask;
828 #endif
829         for_each_cpu(i, span) {
830                 int enqueue = 0;
831                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
832                 struct rq *rq = rq_of_rt_rq(rt_rq);
833                 int skip;
834
835                 /*
836                  * When span == cpu_online_mask, taking each rq->lock
837                  * can be time-consuming. Try to avoid it when possible.
838                  */
839                 raw_spin_lock(&rt_rq->rt_runtime_lock);
840                 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
841                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
842                 if (skip)
843                         continue;
844
845                 raw_spin_lock(&rq->lock);
846                 if (rt_rq->rt_time) {
847                         u64 runtime;
848
849                         raw_spin_lock(&rt_rq->rt_runtime_lock);
850                         if (rt_rq->rt_throttled)
851                                 balance_runtime(rt_rq);
852                         runtime = rt_rq->rt_runtime;
853                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
854                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
855                                 rt_rq->rt_throttled = 0;
856                                 enqueue = 1;
857
858                                 /*
859                                  * When we're idle and a woken (rt) task is
860                                  * throttled check_preempt_curr() will set
861                                  * skip_update and the time between the wakeup
862                                  * and this unthrottle will get accounted as
863                                  * 'runtime'.
864                                  */
865                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
866                                         rq_clock_skip_update(rq, false);
867                         }
868                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
869                                 idle = 0;
870                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
871                 } else if (rt_rq->rt_nr_running) {
872                         idle = 0;
873                         if (!rt_rq_throttled(rt_rq))
874                                 enqueue = 1;
875                 }
876                 if (rt_rq->rt_throttled)
877                         throttled = 1;
878
879                 if (enqueue)
880                         sched_rt_rq_enqueue(rt_rq);
881                 raw_spin_unlock(&rq->lock);
882         }
883
884         if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
885                 return 1;
886
887         return idle;
888 }
889
890 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
891 {
892 #ifdef CONFIG_RT_GROUP_SCHED
893         struct rt_rq *rt_rq = group_rt_rq(rt_se);
894
895         if (rt_rq)
896                 return rt_rq->highest_prio.curr;
897 #endif
898
899         return rt_task_of(rt_se)->prio;
900 }
901
902 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
903 {
904         u64 runtime = sched_rt_runtime(rt_rq);
905
906         if (rt_rq->rt_throttled)
907                 return rt_rq_throttled(rt_rq);
908
909         if (runtime >= sched_rt_period(rt_rq))
910                 return 0;
911
912         balance_runtime(rt_rq);
913         runtime = sched_rt_runtime(rt_rq);
914         if (runtime == RUNTIME_INF)
915                 return 0;
916
917         if (rt_rq->rt_time > runtime) {
918                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
919
920                 /*
921                  * Don't actually throttle groups that have no runtime assigned
922                  * but accrue some time due to boosting.
923                  */
924                 if (likely(rt_b->rt_runtime)) {
925                         rt_rq->rt_throttled = 1;
926                         printk_deferred_once("sched: RT throttling activated\n");
927                 } else {
928                         /*
929                          * In case we did anyway, make it go away,
930                          * replenishment is a joke, since it will replenish us
931                          * with exactly 0 ns.
932                          */
933                         rt_rq->rt_time = 0;
934                 }
935
936                 if (rt_rq_throttled(rt_rq)) {
937                         sched_rt_rq_dequeue(rt_rq);
938                         return 1;
939                 }
940         }
941
942         return 0;
943 }
944
945 /*
946  * Update the current task's runtime statistics. Skip current tasks that
947  * are not in our scheduling class.
948  */
949 static void update_curr_rt(struct rq *rq)
950 {
951         struct task_struct *curr = rq->curr;
952         struct sched_rt_entity *rt_se = &curr->rt;
953         u64 delta_exec;
954         u64 now;
955
956         if (curr->sched_class != &rt_sched_class)
957                 return;
958
959         now = rq_clock_task(rq);
960         delta_exec = now - curr->se.exec_start;
961         if (unlikely((s64)delta_exec <= 0))
962                 return;
963
964         /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
965         cpufreq_update_util(rq, SCHED_CPUFREQ_RT);
966
967         schedstat_set(curr->se.statistics.exec_max,
968                       max(curr->se.statistics.exec_max, delta_exec));
969
970         curr->se.sum_exec_runtime += delta_exec;
971         account_group_exec_runtime(curr, delta_exec);
972
973         curr->se.exec_start = now;
974         cgroup_account_cputime(curr, delta_exec);
975
976         sched_rt_avg_update(rq, delta_exec);
977
978         if (!rt_bandwidth_enabled())
979                 return;
980
981         for_each_sched_rt_entity(rt_se) {
982                 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
983
984                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
985                         raw_spin_lock(&rt_rq->rt_runtime_lock);
986                         rt_rq->rt_time += delta_exec;
987                         if (sched_rt_runtime_exceeded(rt_rq))
988                                 resched_curr(rq);
989                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
990                 }
991         }
992 }
993
994 static void
995 dequeue_top_rt_rq(struct rt_rq *rt_rq)
996 {
997         struct rq *rq = rq_of_rt_rq(rt_rq);
998
999         BUG_ON(&rq->rt != rt_rq);
1000
1001         if (!rt_rq->rt_queued)
1002                 return;
1003
1004         BUG_ON(!rq->nr_running);
1005
1006         sub_nr_running(rq, rt_rq->rt_nr_running);
1007         rt_rq->rt_queued = 0;
1008 }
1009
1010 static void
1011 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1012 {
1013         struct rq *rq = rq_of_rt_rq(rt_rq);
1014
1015         BUG_ON(&rq->rt != rt_rq);
1016
1017         if (rt_rq->rt_queued)
1018                 return;
1019         if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1020                 return;
1021
1022         add_nr_running(rq, rt_rq->rt_nr_running);
1023         rt_rq->rt_queued = 1;
1024 }
1025
1026 #if defined CONFIG_SMP
1027
1028 static void
1029 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1030 {
1031         struct rq *rq = rq_of_rt_rq(rt_rq);
1032
1033 #ifdef CONFIG_RT_GROUP_SCHED
1034         /*
1035          * Change rq's cpupri only if rt_rq is the top queue.
1036          */
1037         if (&rq->rt != rt_rq)
1038                 return;
1039 #endif
1040         if (rq->online && prio < prev_prio)
1041                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1042 }
1043
1044 static void
1045 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1046 {
1047         struct rq *rq = rq_of_rt_rq(rt_rq);
1048
1049 #ifdef CONFIG_RT_GROUP_SCHED
1050         /*
1051          * Change rq's cpupri only if rt_rq is the top queue.
1052          */
1053         if (&rq->rt != rt_rq)
1054                 return;
1055 #endif
1056         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1057                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1058 }
1059
1060 #else /* CONFIG_SMP */
1061
1062 static inline
1063 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1064 static inline
1065 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1066
1067 #endif /* CONFIG_SMP */
1068
1069 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1070 static void
1071 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1072 {
1073         int prev_prio = rt_rq->highest_prio.curr;
1074
1075         if (prio < prev_prio)
1076                 rt_rq->highest_prio.curr = prio;
1077
1078         inc_rt_prio_smp(rt_rq, prio, prev_prio);
1079 }
1080
1081 static void
1082 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1083 {
1084         int prev_prio = rt_rq->highest_prio.curr;
1085
1086         if (rt_rq->rt_nr_running) {
1087
1088                 WARN_ON(prio < prev_prio);
1089
1090                 /*
1091                  * This may have been our highest task, and therefore
1092                  * we may have some recomputation to do
1093                  */
1094                 if (prio == prev_prio) {
1095                         struct rt_prio_array *array = &rt_rq->active;
1096
1097                         rt_rq->highest_prio.curr =
1098                                 sched_find_first_bit(array->bitmap);
1099                 }
1100
1101         } else
1102                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1103
1104         dec_rt_prio_smp(rt_rq, prio, prev_prio);
1105 }
1106
1107 #else
1108
1109 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1110 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1111
1112 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1113
1114 #ifdef CONFIG_RT_GROUP_SCHED
1115
1116 static void
1117 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1118 {
1119         if (rt_se_boosted(rt_se))
1120                 rt_rq->rt_nr_boosted++;
1121
1122         if (rt_rq->tg)
1123                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1124 }
1125
1126 static void
1127 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1128 {
1129         if (rt_se_boosted(rt_se))
1130                 rt_rq->rt_nr_boosted--;
1131
1132         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1133 }
1134
1135 #else /* CONFIG_RT_GROUP_SCHED */
1136
1137 static void
1138 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1139 {
1140         start_rt_bandwidth(&def_rt_bandwidth);
1141 }
1142
1143 static inline
1144 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1145
1146 #endif /* CONFIG_RT_GROUP_SCHED */
1147
1148 static inline
1149 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1150 {
1151         struct rt_rq *group_rq = group_rt_rq(rt_se);
1152
1153         if (group_rq)
1154                 return group_rq->rt_nr_running;
1155         else
1156                 return 1;
1157 }
1158
1159 static inline
1160 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1161 {
1162         struct rt_rq *group_rq = group_rt_rq(rt_se);
1163         struct task_struct *tsk;
1164
1165         if (group_rq)
1166                 return group_rq->rr_nr_running;
1167
1168         tsk = rt_task_of(rt_se);
1169
1170         return (tsk->policy == SCHED_RR) ? 1 : 0;
1171 }
1172
1173 static inline
1174 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1175 {
1176         int prio = rt_se_prio(rt_se);
1177
1178         WARN_ON(!rt_prio(prio));
1179         rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1180         rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1181
1182         inc_rt_prio(rt_rq, prio);
1183         inc_rt_migration(rt_se, rt_rq);
1184         inc_rt_group(rt_se, rt_rq);
1185 }
1186
1187 static inline
1188 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1189 {
1190         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1191         WARN_ON(!rt_rq->rt_nr_running);
1192         rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1193         rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1194
1195         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1196         dec_rt_migration(rt_se, rt_rq);
1197         dec_rt_group(rt_se, rt_rq);
1198 }
1199
1200 /*
1201  * Change rt_se->run_list location unless SAVE && !MOVE
1202  *
1203  * assumes ENQUEUE/DEQUEUE flags match
1204  */
1205 static inline bool move_entity(unsigned int flags)
1206 {
1207         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1208                 return false;
1209
1210         return true;
1211 }
1212
1213 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1214 {
1215         list_del_init(&rt_se->run_list);
1216
1217         if (list_empty(array->queue + rt_se_prio(rt_se)))
1218                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1219
1220         rt_se->on_list = 0;
1221 }
1222
1223 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1224 {
1225         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1226         struct rt_prio_array *array = &rt_rq->active;
1227         struct rt_rq *group_rq = group_rt_rq(rt_se);
1228         struct list_head *queue = array->queue + rt_se_prio(rt_se);
1229
1230         /*
1231          * Don't enqueue the group if its throttled, or when empty.
1232          * The latter is a consequence of the former when a child group
1233          * get throttled and the current group doesn't have any other
1234          * active members.
1235          */
1236         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1237                 if (rt_se->on_list)
1238                         __delist_rt_entity(rt_se, array);
1239                 return;
1240         }
1241
1242         if (move_entity(flags)) {
1243                 WARN_ON_ONCE(rt_se->on_list);
1244                 if (flags & ENQUEUE_HEAD)
1245                         list_add(&rt_se->run_list, queue);
1246                 else
1247                         list_add_tail(&rt_se->run_list, queue);
1248
1249                 __set_bit(rt_se_prio(rt_se), array->bitmap);
1250                 rt_se->on_list = 1;
1251         }
1252         rt_se->on_rq = 1;
1253
1254         inc_rt_tasks(rt_se, rt_rq);
1255 }
1256
1257 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1258 {
1259         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1260         struct rt_prio_array *array = &rt_rq->active;
1261
1262         if (move_entity(flags)) {
1263                 WARN_ON_ONCE(!rt_se->on_list);
1264                 __delist_rt_entity(rt_se, array);
1265         }
1266         rt_se->on_rq = 0;
1267
1268         dec_rt_tasks(rt_se, rt_rq);
1269 }
1270
1271 /*
1272  * Because the prio of an upper entry depends on the lower
1273  * entries, we must remove entries top - down.
1274  */
1275 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1276 {
1277         struct sched_rt_entity *back = NULL;
1278
1279         for_each_sched_rt_entity(rt_se) {
1280                 rt_se->back = back;
1281                 back = rt_se;
1282         }
1283
1284         dequeue_top_rt_rq(rt_rq_of_se(back));
1285
1286         for (rt_se = back; rt_se; rt_se = rt_se->back) {
1287                 if (on_rt_rq(rt_se))
1288                         __dequeue_rt_entity(rt_se, flags);
1289         }
1290 }
1291
1292 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1293 {
1294         struct rq *rq = rq_of_rt_se(rt_se);
1295
1296         dequeue_rt_stack(rt_se, flags);
1297         for_each_sched_rt_entity(rt_se)
1298                 __enqueue_rt_entity(rt_se, flags);
1299         enqueue_top_rt_rq(&rq->rt);
1300 }
1301
1302 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1303 {
1304         struct rq *rq = rq_of_rt_se(rt_se);
1305
1306         dequeue_rt_stack(rt_se, flags);
1307
1308         for_each_sched_rt_entity(rt_se) {
1309                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1310
1311                 if (rt_rq && rt_rq->rt_nr_running)
1312                         __enqueue_rt_entity(rt_se, flags);
1313         }
1314         enqueue_top_rt_rq(&rq->rt);
1315 }
1316
1317 /*
1318  * Adding/removing a task to/from a priority array:
1319  */
1320 static void
1321 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1322 {
1323         struct sched_rt_entity *rt_se = &p->rt;
1324
1325         if (flags & ENQUEUE_WAKEUP)
1326                 rt_se->timeout = 0;
1327
1328         enqueue_rt_entity(rt_se, flags);
1329
1330         if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1331                 enqueue_pushable_task(rq, p);
1332 }
1333
1334 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1335 {
1336         struct sched_rt_entity *rt_se = &p->rt;
1337
1338         update_curr_rt(rq);
1339         dequeue_rt_entity(rt_se, flags);
1340
1341         dequeue_pushable_task(rq, p);
1342 }
1343
1344 /*
1345  * Put task to the head or the end of the run list without the overhead of
1346  * dequeue followed by enqueue.
1347  */
1348 static void
1349 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1350 {
1351         if (on_rt_rq(rt_se)) {
1352                 struct rt_prio_array *array = &rt_rq->active;
1353                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1354
1355                 if (head)
1356                         list_move(&rt_se->run_list, queue);
1357                 else
1358                         list_move_tail(&rt_se->run_list, queue);
1359         }
1360 }
1361
1362 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1363 {
1364         struct sched_rt_entity *rt_se = &p->rt;
1365         struct rt_rq *rt_rq;
1366
1367         for_each_sched_rt_entity(rt_se) {
1368                 rt_rq = rt_rq_of_se(rt_se);
1369                 requeue_rt_entity(rt_rq, rt_se, head);
1370         }
1371 }
1372
1373 static void yield_task_rt(struct rq *rq)
1374 {
1375         requeue_task_rt(rq, rq->curr, 0);
1376 }
1377
1378 #ifdef CONFIG_SMP
1379 static int find_lowest_rq(struct task_struct *task);
1380
1381 static int
1382 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1383 {
1384         struct task_struct *curr;
1385         struct rq *rq;
1386
1387         /* For anything but wake ups, just return the task_cpu */
1388         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1389                 goto out;
1390
1391         rq = cpu_rq(cpu);
1392
1393         rcu_read_lock();
1394         curr = READ_ONCE(rq->curr); /* unlocked access */
1395
1396         /*
1397          * If the current task on @p's runqueue is an RT task, then
1398          * try to see if we can wake this RT task up on another
1399          * runqueue. Otherwise simply start this RT task
1400          * on its current runqueue.
1401          *
1402          * We want to avoid overloading runqueues. If the woken
1403          * task is a higher priority, then it will stay on this CPU
1404          * and the lower prio task should be moved to another CPU.
1405          * Even though this will probably make the lower prio task
1406          * lose its cache, we do not want to bounce a higher task
1407          * around just because it gave up its CPU, perhaps for a
1408          * lock?
1409          *
1410          * For equal prio tasks, we just let the scheduler sort it out.
1411          *
1412          * Otherwise, just let it ride on the affined RQ and the
1413          * post-schedule router will push the preempted task away
1414          *
1415          * This test is optimistic, if we get it wrong the load-balancer
1416          * will have to sort it out.
1417          */
1418         if (curr && unlikely(rt_task(curr)) &&
1419             (curr->nr_cpus_allowed < 2 ||
1420              curr->prio <= p->prio)) {
1421                 int target = find_lowest_rq(p);
1422
1423                 /*
1424                  * Don't bother moving it if the destination CPU is
1425                  * not running a lower priority task.
1426                  */
1427                 if (target != -1 &&
1428                     p->prio < cpu_rq(target)->rt.highest_prio.curr)
1429                         cpu = target;
1430         }
1431         rcu_read_unlock();
1432
1433 out:
1434         return cpu;
1435 }
1436
1437 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1438 {
1439         /*
1440          * Current can't be migrated, useless to reschedule,
1441          * let's hope p can move out.
1442          */
1443         if (rq->curr->nr_cpus_allowed == 1 ||
1444             !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1445                 return;
1446
1447         /*
1448          * p is migratable, so let's not schedule it and
1449          * see if it is pushed or pulled somewhere else.
1450          */
1451         if (p->nr_cpus_allowed != 1
1452             && cpupri_find(&rq->rd->cpupri, p, NULL))
1453                 return;
1454
1455         /*
1456          * There appears to be other cpus that can accept
1457          * current and none to run 'p', so lets reschedule
1458          * to try and push current away:
1459          */
1460         requeue_task_rt(rq, p, 1);
1461         resched_curr(rq);
1462 }
1463
1464 #endif /* CONFIG_SMP */
1465
1466 /*
1467  * Preempt the current task with a newly woken task if needed:
1468  */
1469 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1470 {
1471         if (p->prio < rq->curr->prio) {
1472                 resched_curr(rq);
1473                 return;
1474         }
1475
1476 #ifdef CONFIG_SMP
1477         /*
1478          * If:
1479          *
1480          * - the newly woken task is of equal priority to the current task
1481          * - the newly woken task is non-migratable while current is migratable
1482          * - current will be preempted on the next reschedule
1483          *
1484          * we should check to see if current can readily move to a different
1485          * cpu.  If so, we will reschedule to allow the push logic to try
1486          * to move current somewhere else, making room for our non-migratable
1487          * task.
1488          */
1489         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1490                 check_preempt_equal_prio(rq, p);
1491 #endif
1492 }
1493
1494 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1495                                                    struct rt_rq *rt_rq)
1496 {
1497         struct rt_prio_array *array = &rt_rq->active;
1498         struct sched_rt_entity *next = NULL;
1499         struct list_head *queue;
1500         int idx;
1501
1502         idx = sched_find_first_bit(array->bitmap);
1503         BUG_ON(idx >= MAX_RT_PRIO);
1504
1505         queue = array->queue + idx;
1506         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1507
1508         return next;
1509 }
1510
1511 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1512 {
1513         struct sched_rt_entity *rt_se;
1514         struct task_struct *p;
1515         struct rt_rq *rt_rq  = &rq->rt;
1516
1517         do {
1518                 rt_se = pick_next_rt_entity(rq, rt_rq);
1519                 BUG_ON(!rt_se);
1520                 rt_rq = group_rt_rq(rt_se);
1521         } while (rt_rq);
1522
1523         p = rt_task_of(rt_se);
1524         p->se.exec_start = rq_clock_task(rq);
1525
1526         return p;
1527 }
1528
1529 static struct task_struct *
1530 pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1531 {
1532         struct task_struct *p;
1533         struct rt_rq *rt_rq = &rq->rt;
1534
1535         if (need_pull_rt_task(rq, prev)) {
1536                 /*
1537                  * This is OK, because current is on_cpu, which avoids it being
1538                  * picked for load-balance and preemption/IRQs are still
1539                  * disabled avoiding further scheduler activity on it and we're
1540                  * being very careful to re-start the picking loop.
1541                  */
1542                 rq_unpin_lock(rq, rf);
1543                 pull_rt_task(rq);
1544                 rq_repin_lock(rq, rf);
1545                 /*
1546                  * pull_rt_task() can drop (and re-acquire) rq->lock; this
1547                  * means a dl or stop task can slip in, in which case we need
1548                  * to re-start task selection.
1549                  */
1550                 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1551                              rq->dl.dl_nr_running))
1552                         return RETRY_TASK;
1553         }
1554
1555         /*
1556          * We may dequeue prev's rt_rq in put_prev_task().
1557          * So, we update time before rt_nr_running check.
1558          */
1559         if (prev->sched_class == &rt_sched_class)
1560                 update_curr_rt(rq);
1561
1562         if (!rt_rq->rt_queued)
1563                 return NULL;
1564
1565         put_prev_task(rq, prev);
1566
1567         p = _pick_next_task_rt(rq);
1568
1569         /* The running task is never eligible for pushing */
1570         dequeue_pushable_task(rq, p);
1571
1572         queue_push_tasks(rq);
1573
1574         return p;
1575 }
1576
1577 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1578 {
1579         update_curr_rt(rq);
1580
1581         /*
1582          * The previous task needs to be made eligible for pushing
1583          * if it is still active
1584          */
1585         if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1586                 enqueue_pushable_task(rq, p);
1587 }
1588
1589 #ifdef CONFIG_SMP
1590
1591 /* Only try algorithms three times */
1592 #define RT_MAX_TRIES 3
1593
1594 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1595 {
1596         if (!task_running(rq, p) &&
1597             cpumask_test_cpu(cpu, &p->cpus_allowed))
1598                 return 1;
1599         return 0;
1600 }
1601
1602 /*
1603  * Return the highest pushable rq's task, which is suitable to be executed
1604  * on the cpu, NULL otherwise
1605  */
1606 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1607 {
1608         struct plist_head *head = &rq->rt.pushable_tasks;
1609         struct task_struct *p;
1610
1611         if (!has_pushable_tasks(rq))
1612                 return NULL;
1613
1614         plist_for_each_entry(p, head, pushable_tasks) {
1615                 if (pick_rt_task(rq, p, cpu))
1616                         return p;
1617         }
1618
1619         return NULL;
1620 }
1621
1622 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1623
1624 static int find_lowest_rq(struct task_struct *task)
1625 {
1626         struct sched_domain *sd;
1627         struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1628         int this_cpu = smp_processor_id();
1629         int cpu      = task_cpu(task);
1630
1631         /* Make sure the mask is initialized first */
1632         if (unlikely(!lowest_mask))
1633                 return -1;
1634
1635         if (task->nr_cpus_allowed == 1)
1636                 return -1; /* No other targets possible */
1637
1638         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1639                 return -1; /* No targets found */
1640
1641         /*
1642          * At this point we have built a mask of cpus representing the
1643          * lowest priority tasks in the system.  Now we want to elect
1644          * the best one based on our affinity and topology.
1645          *
1646          * We prioritize the last cpu that the task executed on since
1647          * it is most likely cache-hot in that location.
1648          */
1649         if (cpumask_test_cpu(cpu, lowest_mask))
1650                 return cpu;
1651
1652         /*
1653          * Otherwise, we consult the sched_domains span maps to figure
1654          * out which cpu is logically closest to our hot cache data.
1655          */
1656         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1657                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1658
1659         rcu_read_lock();
1660         for_each_domain(cpu, sd) {
1661                 if (sd->flags & SD_WAKE_AFFINE) {
1662                         int best_cpu;
1663
1664                         /*
1665                          * "this_cpu" is cheaper to preempt than a
1666                          * remote processor.
1667                          */
1668                         if (this_cpu != -1 &&
1669                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1670                                 rcu_read_unlock();
1671                                 return this_cpu;
1672                         }
1673
1674                         best_cpu = cpumask_first_and(lowest_mask,
1675                                                      sched_domain_span(sd));
1676                         if (best_cpu < nr_cpu_ids) {
1677                                 rcu_read_unlock();
1678                                 return best_cpu;
1679                         }
1680                 }
1681         }
1682         rcu_read_unlock();
1683
1684         /*
1685          * And finally, if there were no matches within the domains
1686          * just give the caller *something* to work with from the compatible
1687          * locations.
1688          */
1689         if (this_cpu != -1)
1690                 return this_cpu;
1691
1692         cpu = cpumask_any(lowest_mask);
1693         if (cpu < nr_cpu_ids)
1694                 return cpu;
1695         return -1;
1696 }
1697
1698 /* Will lock the rq it finds */
1699 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1700 {
1701         struct rq *lowest_rq = NULL;
1702         int tries;
1703         int cpu;
1704
1705         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1706                 cpu = find_lowest_rq(task);
1707
1708                 if ((cpu == -1) || (cpu == rq->cpu))
1709                         break;
1710
1711                 lowest_rq = cpu_rq(cpu);
1712
1713                 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1714                         /*
1715                          * Target rq has tasks of equal or higher priority,
1716                          * retrying does not release any lock and is unlikely
1717                          * to yield a different result.
1718                          */
1719                         lowest_rq = NULL;
1720                         break;
1721                 }
1722
1723                 /* if the prio of this runqueue changed, try again */
1724                 if (double_lock_balance(rq, lowest_rq)) {
1725                         /*
1726                          * We had to unlock the run queue. In
1727                          * the mean time, task could have
1728                          * migrated already or had its affinity changed.
1729                          * Also make sure that it wasn't scheduled on its rq.
1730                          */
1731                         if (unlikely(task_rq(task) != rq ||
1732                                      !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) ||
1733                                      task_running(rq, task) ||
1734                                      !rt_task(task) ||
1735                                      !task_on_rq_queued(task))) {
1736
1737                                 double_unlock_balance(rq, lowest_rq);
1738                                 lowest_rq = NULL;
1739                                 break;
1740                         }
1741                 }
1742
1743                 /* If this rq is still suitable use it. */
1744                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1745                         break;
1746
1747                 /* try again */
1748                 double_unlock_balance(rq, lowest_rq);
1749                 lowest_rq = NULL;
1750         }
1751
1752         return lowest_rq;
1753 }
1754
1755 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1756 {
1757         struct task_struct *p;
1758
1759         if (!has_pushable_tasks(rq))
1760                 return NULL;
1761
1762         p = plist_first_entry(&rq->rt.pushable_tasks,
1763                               struct task_struct, pushable_tasks);
1764
1765         BUG_ON(rq->cpu != task_cpu(p));
1766         BUG_ON(task_current(rq, p));
1767         BUG_ON(p->nr_cpus_allowed <= 1);
1768
1769         BUG_ON(!task_on_rq_queued(p));
1770         BUG_ON(!rt_task(p));
1771
1772         return p;
1773 }
1774
1775 /*
1776  * If the current CPU has more than one RT task, see if the non
1777  * running task can migrate over to a CPU that is running a task
1778  * of lesser priority.
1779  */
1780 static int push_rt_task(struct rq *rq)
1781 {
1782         struct task_struct *next_task;
1783         struct rq *lowest_rq;
1784         int ret = 0;
1785
1786         if (!rq->rt.overloaded)
1787                 return 0;
1788
1789         next_task = pick_next_pushable_task(rq);
1790         if (!next_task)
1791                 return 0;
1792
1793 retry:
1794         if (unlikely(next_task == rq->curr)) {
1795                 WARN_ON(1);
1796                 return 0;
1797         }
1798
1799         /*
1800          * It's possible that the next_task slipped in of
1801          * higher priority than current. If that's the case
1802          * just reschedule current.
1803          */
1804         if (unlikely(next_task->prio < rq->curr->prio)) {
1805                 resched_curr(rq);
1806                 return 0;
1807         }
1808
1809         /* We might release rq lock */
1810         get_task_struct(next_task);
1811
1812         /* find_lock_lowest_rq locks the rq if found */
1813         lowest_rq = find_lock_lowest_rq(next_task, rq);
1814         if (!lowest_rq) {
1815                 struct task_struct *task;
1816                 /*
1817                  * find_lock_lowest_rq releases rq->lock
1818                  * so it is possible that next_task has migrated.
1819                  *
1820                  * We need to make sure that the task is still on the same
1821                  * run-queue and is also still the next task eligible for
1822                  * pushing.
1823                  */
1824                 task = pick_next_pushable_task(rq);
1825                 if (task == next_task) {
1826                         /*
1827                          * The task hasn't migrated, and is still the next
1828                          * eligible task, but we failed to find a run-queue
1829                          * to push it to.  Do not retry in this case, since
1830                          * other cpus will pull from us when ready.
1831                          */
1832                         goto out;
1833                 }
1834
1835                 if (!task)
1836                         /* No more tasks, just exit */
1837                         goto out;
1838
1839                 /*
1840                  * Something has shifted, try again.
1841                  */
1842                 put_task_struct(next_task);
1843                 next_task = task;
1844                 goto retry;
1845         }
1846
1847         deactivate_task(rq, next_task, 0);
1848         set_task_cpu(next_task, lowest_rq->cpu);
1849         activate_task(lowest_rq, next_task, 0);
1850         ret = 1;
1851
1852         resched_curr(lowest_rq);
1853
1854         double_unlock_balance(rq, lowest_rq);
1855
1856 out:
1857         put_task_struct(next_task);
1858
1859         return ret;
1860 }
1861
1862 static void push_rt_tasks(struct rq *rq)
1863 {
1864         /* push_rt_task will return true if it moved an RT */
1865         while (push_rt_task(rq))
1866                 ;
1867 }
1868
1869 #ifdef HAVE_RT_PUSH_IPI
1870
1871 /*
1872  * When a high priority task schedules out from a CPU and a lower priority
1873  * task is scheduled in, a check is made to see if there's any RT tasks
1874  * on other CPUs that are waiting to run because a higher priority RT task
1875  * is currently running on its CPU. In this case, the CPU with multiple RT
1876  * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1877  * up that may be able to run one of its non-running queued RT tasks.
1878  *
1879  * All CPUs with overloaded RT tasks need to be notified as there is currently
1880  * no way to know which of these CPUs have the highest priority task waiting
1881  * to run. Instead of trying to take a spinlock on each of these CPUs,
1882  * which has shown to cause large latency when done on machines with many
1883  * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1884  * RT tasks waiting to run.
1885  *
1886  * Just sending an IPI to each of the CPUs is also an issue, as on large
1887  * count CPU machines, this can cause an IPI storm on a CPU, especially
1888  * if its the only CPU with multiple RT tasks queued, and a large number
1889  * of CPUs scheduling a lower priority task at the same time.
1890  *
1891  * Each root domain has its own irq work function that can iterate over
1892  * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1893  * tassk must be checked if there's one or many CPUs that are lowering
1894  * their priority, there's a single irq work iterator that will try to
1895  * push off RT tasks that are waiting to run.
1896  *
1897  * When a CPU schedules a lower priority task, it will kick off the
1898  * irq work iterator that will jump to each CPU with overloaded RT tasks.
1899  * As it only takes the first CPU that schedules a lower priority task
1900  * to start the process, the rto_start variable is incremented and if
1901  * the atomic result is one, then that CPU will try to take the rto_lock.
1902  * This prevents high contention on the lock as the process handles all
1903  * CPUs scheduling lower priority tasks.
1904  *
1905  * All CPUs that are scheduling a lower priority task will increment the
1906  * rt_loop_next variable. This will make sure that the irq work iterator
1907  * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1908  * priority task, even if the iterator is in the middle of a scan. Incrementing
1909  * the rt_loop_next will cause the iterator to perform another scan.
1910  *
1911  */
1912 static int rto_next_cpu(struct root_domain *rd)
1913 {
1914         int next;
1915         int cpu;
1916
1917         /*
1918          * When starting the IPI RT pushing, the rto_cpu is set to -1,
1919          * rt_next_cpu() will simply return the first CPU found in
1920          * the rto_mask.
1921          *
1922          * If rto_next_cpu() is called with rto_cpu is a valid cpu, it
1923          * will return the next CPU found in the rto_mask.
1924          *
1925          * If there are no more CPUs left in the rto_mask, then a check is made
1926          * against rto_loop and rto_loop_next. rto_loop is only updated with
1927          * the rto_lock held, but any CPU may increment the rto_loop_next
1928          * without any locking.
1929          */
1930         for (;;) {
1931
1932                 /* When rto_cpu is -1 this acts like cpumask_first() */
1933                 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask);
1934
1935                 rd->rto_cpu = cpu;
1936
1937                 if (cpu < nr_cpu_ids)
1938                         return cpu;
1939
1940                 rd->rto_cpu = -1;
1941
1942                 /*
1943                  * ACQUIRE ensures we see the @rto_mask changes
1944                  * made prior to the @next value observed.
1945                  *
1946                  * Matches WMB in rt_set_overload().
1947                  */
1948                 next = atomic_read_acquire(&rd->rto_loop_next);
1949
1950                 if (rd->rto_loop == next)
1951                         break;
1952
1953                 rd->rto_loop = next;
1954         }
1955
1956         return -1;
1957 }
1958
1959 static inline bool rto_start_trylock(atomic_t *v)
1960 {
1961         return !atomic_cmpxchg_acquire(v, 0, 1);
1962 }
1963
1964 static inline void rto_start_unlock(atomic_t *v)
1965 {
1966         atomic_set_release(v, 0);
1967 }
1968
1969 static void tell_cpu_to_push(struct rq *rq)
1970 {
1971         int cpu = -1;
1972
1973         /* Keep the loop going if the IPI is currently active */
1974         atomic_inc(&rq->rd->rto_loop_next);
1975
1976         /* Only one CPU can initiate a loop at a time */
1977         if (!rto_start_trylock(&rq->rd->rto_loop_start))
1978                 return;
1979
1980         raw_spin_lock(&rq->rd->rto_lock);
1981
1982         /*
1983          * The rto_cpu is updated under the lock, if it has a valid cpu
1984          * then the IPI is still running and will continue due to the
1985          * update to loop_next, and nothing needs to be done here.
1986          * Otherwise it is finishing up and an ipi needs to be sent.
1987          */
1988         if (rq->rd->rto_cpu < 0)
1989                 cpu = rto_next_cpu(rq->rd);
1990
1991         raw_spin_unlock(&rq->rd->rto_lock);
1992
1993         rto_start_unlock(&rq->rd->rto_loop_start);
1994
1995         if (cpu >= 0) {
1996                 /* Make sure the rd does not get freed while pushing */
1997                 sched_get_rd(rq->rd);
1998                 irq_work_queue_on(&rq->rd->rto_push_work, cpu);
1999         }
2000 }
2001
2002 /* Called from hardirq context */
2003 void rto_push_irq_work_func(struct irq_work *work)
2004 {
2005         struct root_domain *rd =
2006                 container_of(work, struct root_domain, rto_push_work);
2007         struct rq *rq;
2008         int cpu;
2009
2010         rq = this_rq();
2011
2012         /*
2013          * We do not need to grab the lock to check for has_pushable_tasks.
2014          * When it gets updated, a check is made if a push is possible.
2015          */
2016         if (has_pushable_tasks(rq)) {
2017                 raw_spin_lock(&rq->lock);
2018                 push_rt_tasks(rq);
2019                 raw_spin_unlock(&rq->lock);
2020         }
2021
2022         raw_spin_lock(&rd->rto_lock);
2023
2024         /* Pass the IPI to the next rt overloaded queue */
2025         cpu = rto_next_cpu(rd);
2026
2027         raw_spin_unlock(&rd->rto_lock);
2028
2029         if (cpu < 0) {
2030                 sched_put_rd(rd);
2031                 return;
2032         }
2033
2034         /* Try the next RT overloaded CPU */
2035         irq_work_queue_on(&rd->rto_push_work, cpu);
2036 }
2037 #endif /* HAVE_RT_PUSH_IPI */
2038
2039 static void pull_rt_task(struct rq *this_rq)
2040 {
2041         int this_cpu = this_rq->cpu, cpu;
2042         bool resched = false;
2043         struct task_struct *p;
2044         struct rq *src_rq;
2045         int rt_overload_count = rt_overloaded(this_rq);
2046
2047         if (likely(!rt_overload_count))
2048                 return;
2049
2050         /*
2051          * Match the barrier from rt_set_overloaded; this guarantees that if we
2052          * see overloaded we must also see the rto_mask bit.
2053          */
2054         smp_rmb();
2055
2056         /* If we are the only overloaded CPU do nothing */
2057         if (rt_overload_count == 1 &&
2058             cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask))
2059                 return;
2060
2061 #ifdef HAVE_RT_PUSH_IPI
2062         if (sched_feat(RT_PUSH_IPI)) {
2063                 tell_cpu_to_push(this_rq);
2064                 return;
2065         }
2066 #endif
2067
2068         for_each_cpu(cpu, this_rq->rd->rto_mask) {
2069                 if (this_cpu == cpu)
2070                         continue;
2071
2072                 src_rq = cpu_rq(cpu);
2073
2074                 /*
2075                  * Don't bother taking the src_rq->lock if the next highest
2076                  * task is known to be lower-priority than our current task.
2077                  * This may look racy, but if this value is about to go
2078                  * logically higher, the src_rq will push this task away.
2079                  * And if its going logically lower, we do not care
2080                  */
2081                 if (src_rq->rt.highest_prio.next >=
2082                     this_rq->rt.highest_prio.curr)
2083                         continue;
2084
2085                 /*
2086                  * We can potentially drop this_rq's lock in
2087                  * double_lock_balance, and another CPU could
2088                  * alter this_rq
2089                  */
2090                 double_lock_balance(this_rq, src_rq);
2091
2092                 /*
2093                  * We can pull only a task, which is pushable
2094                  * on its rq, and no others.
2095                  */
2096                 p = pick_highest_pushable_task(src_rq, this_cpu);
2097
2098                 /*
2099                  * Do we have an RT task that preempts
2100                  * the to-be-scheduled task?
2101                  */
2102                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2103                         WARN_ON(p == src_rq->curr);
2104                         WARN_ON(!task_on_rq_queued(p));
2105
2106                         /*
2107                          * There's a chance that p is higher in priority
2108                          * than what's currently running on its cpu.
2109                          * This is just that p is wakeing up and hasn't
2110                          * had a chance to schedule. We only pull
2111                          * p if it is lower in priority than the
2112                          * current task on the run queue
2113                          */
2114                         if (p->prio < src_rq->curr->prio)
2115                                 goto skip;
2116
2117                         resched = true;
2118
2119                         deactivate_task(src_rq, p, 0);
2120                         set_task_cpu(p, this_cpu);
2121                         activate_task(this_rq, p, 0);
2122                         /*
2123                          * We continue with the search, just in
2124                          * case there's an even higher prio task
2125                          * in another runqueue. (low likelihood
2126                          * but possible)
2127                          */
2128                 }
2129 skip:
2130                 double_unlock_balance(this_rq, src_rq);
2131         }
2132
2133         if (resched)
2134                 resched_curr(this_rq);
2135 }
2136
2137 /*
2138  * If we are not running and we are not going to reschedule soon, we should
2139  * try to push tasks away now
2140  */
2141 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2142 {
2143         if (!task_running(rq, p) &&
2144             !test_tsk_need_resched(rq->curr) &&
2145             p->nr_cpus_allowed > 1 &&
2146             (dl_task(rq->curr) || rt_task(rq->curr)) &&
2147             (rq->curr->nr_cpus_allowed < 2 ||
2148              rq->curr->prio <= p->prio))
2149                 push_rt_tasks(rq);
2150 }
2151
2152 /* Assumes rq->lock is held */
2153 static void rq_online_rt(struct rq *rq)
2154 {
2155         if (rq->rt.overloaded)
2156                 rt_set_overload(rq);
2157
2158         __enable_runtime(rq);
2159
2160         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2161 }
2162
2163 /* Assumes rq->lock is held */
2164 static void rq_offline_rt(struct rq *rq)
2165 {
2166         if (rq->rt.overloaded)
2167                 rt_clear_overload(rq);
2168
2169         __disable_runtime(rq);
2170
2171         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2172 }
2173
2174 /*
2175  * When switch from the rt queue, we bring ourselves to a position
2176  * that we might want to pull RT tasks from other runqueues.
2177  */
2178 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2179 {
2180         /*
2181          * If there are other RT tasks then we will reschedule
2182          * and the scheduling of the other RT tasks will handle
2183          * the balancing. But if we are the last RT task
2184          * we may need to handle the pulling of RT tasks
2185          * now.
2186          */
2187         if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2188                 return;
2189
2190         queue_pull_task(rq);
2191 }
2192
2193 void __init init_sched_rt_class(void)
2194 {
2195         unsigned int i;
2196
2197         for_each_possible_cpu(i) {
2198                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2199                                         GFP_KERNEL, cpu_to_node(i));
2200         }
2201 }
2202 #endif /* CONFIG_SMP */
2203
2204 /*
2205  * When switching a task to RT, we may overload the runqueue
2206  * with RT tasks. In this case we try to push them off to
2207  * other runqueues.
2208  */
2209 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2210 {
2211         /*
2212          * If we are already running, then there's nothing
2213          * that needs to be done. But if we are not running
2214          * we may need to preempt the current running task.
2215          * If that current running task is also an RT task
2216          * then see if we can move to another run queue.
2217          */
2218         if (task_on_rq_queued(p) && rq->curr != p) {
2219 #ifdef CONFIG_SMP
2220                 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2221                         queue_push_tasks(rq);
2222 #endif /* CONFIG_SMP */
2223                 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq)))
2224                         resched_curr(rq);
2225         }
2226 }
2227
2228 /*
2229  * Priority of the task has changed. This may cause
2230  * us to initiate a push or pull.
2231  */
2232 static void
2233 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2234 {
2235         if (!task_on_rq_queued(p))
2236                 return;
2237
2238         if (rq->curr == p) {
2239 #ifdef CONFIG_SMP
2240                 /*
2241                  * If our priority decreases while running, we
2242                  * may need to pull tasks to this runqueue.
2243                  */
2244                 if (oldprio < p->prio)
2245                         queue_pull_task(rq);
2246
2247                 /*
2248                  * If there's a higher priority task waiting to run
2249                  * then reschedule.
2250                  */
2251                 if (p->prio > rq->rt.highest_prio.curr)
2252                         resched_curr(rq);
2253 #else
2254                 /* For UP simply resched on drop of prio */
2255                 if (oldprio < p->prio)
2256                         resched_curr(rq);
2257 #endif /* CONFIG_SMP */
2258         } else {
2259                 /*
2260                  * This task is not running, but if it is
2261                  * greater than the current running task
2262                  * then reschedule.
2263                  */
2264                 if (p->prio < rq->curr->prio)
2265                         resched_curr(rq);
2266         }
2267 }
2268
2269 #ifdef CONFIG_POSIX_TIMERS
2270 static void watchdog(struct rq *rq, struct task_struct *p)
2271 {
2272         unsigned long soft, hard;
2273
2274         /* max may change after cur was read, this will be fixed next tick */
2275         soft = task_rlimit(p, RLIMIT_RTTIME);
2276         hard = task_rlimit_max(p, RLIMIT_RTTIME);
2277
2278         if (soft != RLIM_INFINITY) {
2279                 unsigned long next;
2280
2281                 if (p->rt.watchdog_stamp != jiffies) {
2282                         p->rt.timeout++;
2283                         p->rt.watchdog_stamp = jiffies;
2284                 }
2285
2286                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2287                 if (p->rt.timeout > next)
2288                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2289         }
2290 }
2291 #else
2292 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2293 #endif
2294
2295 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2296 {
2297         struct sched_rt_entity *rt_se = &p->rt;
2298
2299         update_curr_rt(rq);
2300
2301         watchdog(rq, p);
2302
2303         /*
2304          * RR tasks need a special form of timeslice management.
2305          * FIFO tasks have no timeslices.
2306          */
2307         if (p->policy != SCHED_RR)
2308                 return;
2309
2310         if (--p->rt.time_slice)
2311                 return;
2312
2313         p->rt.time_slice = sched_rr_timeslice;
2314
2315         /*
2316          * Requeue to the end of queue if we (and all of our ancestors) are not
2317          * the only element on the queue
2318          */
2319         for_each_sched_rt_entity(rt_se) {
2320                 if (rt_se->run_list.prev != rt_se->run_list.next) {
2321                         requeue_task_rt(rq, p, 0);
2322                         resched_curr(rq);
2323                         return;
2324                 }
2325         }
2326 }
2327
2328 static void set_curr_task_rt(struct rq *rq)
2329 {
2330         struct task_struct *p = rq->curr;
2331
2332         p->se.exec_start = rq_clock_task(rq);
2333
2334         /* The running task is never eligible for pushing */
2335         dequeue_pushable_task(rq, p);
2336 }
2337
2338 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2339 {
2340         /*
2341          * Time slice is 0 for SCHED_FIFO tasks
2342          */
2343         if (task->policy == SCHED_RR)
2344                 return sched_rr_timeslice;
2345         else
2346                 return 0;
2347 }
2348
2349 const struct sched_class rt_sched_class = {
2350         .next                   = &fair_sched_class,
2351         .enqueue_task           = enqueue_task_rt,
2352         .dequeue_task           = dequeue_task_rt,
2353         .yield_task             = yield_task_rt,
2354
2355         .check_preempt_curr     = check_preempt_curr_rt,
2356
2357         .pick_next_task         = pick_next_task_rt,
2358         .put_prev_task          = put_prev_task_rt,
2359
2360 #ifdef CONFIG_SMP
2361         .select_task_rq         = select_task_rq_rt,
2362
2363         .set_cpus_allowed       = set_cpus_allowed_common,
2364         .rq_online              = rq_online_rt,
2365         .rq_offline             = rq_offline_rt,
2366         .task_woken             = task_woken_rt,
2367         .switched_from          = switched_from_rt,
2368 #endif
2369
2370         .set_curr_task          = set_curr_task_rt,
2371         .task_tick              = task_tick_rt,
2372
2373         .get_rr_interval        = get_rr_interval_rt,
2374
2375         .prio_changed           = prio_changed_rt,
2376         .switched_to            = switched_to_rt,
2377
2378         .update_curr            = update_curr_rt,
2379 };
2380
2381 #ifdef CONFIG_RT_GROUP_SCHED
2382 /*
2383  * Ensure that the real time constraints are schedulable.
2384  */
2385 static DEFINE_MUTEX(rt_constraints_mutex);
2386
2387 /* Must be called with tasklist_lock held */
2388 static inline int tg_has_rt_tasks(struct task_group *tg)
2389 {
2390         struct task_struct *g, *p;
2391
2392         /*
2393          * Autogroups do not have RT tasks; see autogroup_create().
2394          */
2395         if (task_group_is_autogroup(tg))
2396                 return 0;
2397
2398         for_each_process_thread(g, p) {
2399                 if (rt_task(p) && task_group(p) == tg)
2400                         return 1;
2401         }
2402
2403         return 0;
2404 }
2405
2406 struct rt_schedulable_data {
2407         struct task_group *tg;
2408         u64 rt_period;
2409         u64 rt_runtime;
2410 };
2411
2412 static int tg_rt_schedulable(struct task_group *tg, void *data)
2413 {
2414         struct rt_schedulable_data *d = data;
2415         struct task_group *child;
2416         unsigned long total, sum = 0;
2417         u64 period, runtime;
2418
2419         period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2420         runtime = tg->rt_bandwidth.rt_runtime;
2421
2422         if (tg == d->tg) {
2423                 period = d->rt_period;
2424                 runtime = d->rt_runtime;
2425         }
2426
2427         /*
2428          * Cannot have more runtime than the period.
2429          */
2430         if (runtime > period && runtime != RUNTIME_INF)
2431                 return -EINVAL;
2432
2433         /*
2434          * Ensure we don't starve existing RT tasks.
2435          */
2436         if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
2437                 return -EBUSY;
2438
2439         total = to_ratio(period, runtime);
2440
2441         /*
2442          * Nobody can have more than the global setting allows.
2443          */
2444         if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2445                 return -EINVAL;
2446
2447         /*
2448          * The sum of our children's runtime should not exceed our own.
2449          */
2450         list_for_each_entry_rcu(child, &tg->children, siblings) {
2451                 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2452                 runtime = child->rt_bandwidth.rt_runtime;
2453
2454                 if (child == d->tg) {
2455                         period = d->rt_period;
2456                         runtime = d->rt_runtime;
2457                 }
2458
2459                 sum += to_ratio(period, runtime);
2460         }
2461
2462         if (sum > total)
2463                 return -EINVAL;
2464
2465         return 0;
2466 }
2467
2468 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2469 {
2470         int ret;
2471
2472         struct rt_schedulable_data data = {
2473                 .tg = tg,
2474                 .rt_period = period,
2475                 .rt_runtime = runtime,
2476         };
2477
2478         rcu_read_lock();
2479         ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2480         rcu_read_unlock();
2481
2482         return ret;
2483 }
2484
2485 static int tg_set_rt_bandwidth(struct task_group *tg,
2486                 u64 rt_period, u64 rt_runtime)
2487 {
2488         int i, err = 0;
2489
2490         /*
2491          * Disallowing the root group RT runtime is BAD, it would disallow the
2492          * kernel creating (and or operating) RT threads.
2493          */
2494         if (tg == &root_task_group && rt_runtime == 0)
2495                 return -EINVAL;
2496
2497         /* No period doesn't make any sense. */
2498         if (rt_period == 0)
2499                 return -EINVAL;
2500
2501         mutex_lock(&rt_constraints_mutex);
2502         read_lock(&tasklist_lock);
2503         err = __rt_schedulable(tg, rt_period, rt_runtime);
2504         if (err)
2505                 goto unlock;
2506
2507         raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2508         tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2509         tg->rt_bandwidth.rt_runtime = rt_runtime;
2510
2511         for_each_possible_cpu(i) {
2512                 struct rt_rq *rt_rq = tg->rt_rq[i];
2513
2514                 raw_spin_lock(&rt_rq->rt_runtime_lock);
2515                 rt_rq->rt_runtime = rt_runtime;
2516                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2517         }
2518         raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2519 unlock:
2520         read_unlock(&tasklist_lock);
2521         mutex_unlock(&rt_constraints_mutex);
2522
2523         return err;
2524 }
2525
2526 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2527 {
2528         u64 rt_runtime, rt_period;
2529
2530         rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2531         rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2532         if (rt_runtime_us < 0)
2533                 rt_runtime = RUNTIME_INF;
2534
2535         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2536 }
2537
2538 long sched_group_rt_runtime(struct task_group *tg)
2539 {
2540         u64 rt_runtime_us;
2541
2542         if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2543                 return -1;
2544
2545         rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2546         do_div(rt_runtime_us, NSEC_PER_USEC);
2547         return rt_runtime_us;
2548 }
2549
2550 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2551 {
2552         u64 rt_runtime, rt_period;
2553
2554         rt_period = rt_period_us * NSEC_PER_USEC;
2555         rt_runtime = tg->rt_bandwidth.rt_runtime;
2556
2557         return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2558 }
2559
2560 long sched_group_rt_period(struct task_group *tg)
2561 {
2562         u64 rt_period_us;
2563
2564         rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2565         do_div(rt_period_us, NSEC_PER_USEC);
2566         return rt_period_us;
2567 }
2568
2569 static int sched_rt_global_constraints(void)
2570 {
2571         int ret = 0;
2572
2573         mutex_lock(&rt_constraints_mutex);
2574         read_lock(&tasklist_lock);
2575         ret = __rt_schedulable(NULL, 0, 0);
2576         read_unlock(&tasklist_lock);
2577         mutex_unlock(&rt_constraints_mutex);
2578
2579         return ret;
2580 }
2581
2582 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2583 {
2584         /* Don't accept realtime tasks when there is no way for them to run */
2585         if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2586                 return 0;
2587
2588         return 1;
2589 }
2590
2591 #else /* !CONFIG_RT_GROUP_SCHED */
2592 static int sched_rt_global_constraints(void)
2593 {
2594         unsigned long flags;
2595         int i;
2596
2597         raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2598         for_each_possible_cpu(i) {
2599                 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2600
2601                 raw_spin_lock(&rt_rq->rt_runtime_lock);
2602                 rt_rq->rt_runtime = global_rt_runtime();
2603                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2604         }
2605         raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2606
2607         return 0;
2608 }
2609 #endif /* CONFIG_RT_GROUP_SCHED */
2610
2611 static int sched_rt_global_validate(void)
2612 {
2613         if (sysctl_sched_rt_period <= 0)
2614                 return -EINVAL;
2615
2616         if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2617                 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
2618                 return -EINVAL;
2619
2620         return 0;
2621 }
2622
2623 static void sched_rt_do_global(void)
2624 {
2625         def_rt_bandwidth.rt_runtime = global_rt_runtime();
2626         def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2627 }
2628
2629 int sched_rt_handler(struct ctl_table *table, int write,
2630                 void __user *buffer, size_t *lenp,
2631                 loff_t *ppos)
2632 {
2633         int old_period, old_runtime;
2634         static DEFINE_MUTEX(mutex);
2635         int ret;
2636
2637         mutex_lock(&mutex);
2638         old_period = sysctl_sched_rt_period;
2639         old_runtime = sysctl_sched_rt_runtime;
2640
2641         ret = proc_dointvec(table, write, buffer, lenp, ppos);
2642
2643         if (!ret && write) {
2644                 ret = sched_rt_global_validate();
2645                 if (ret)
2646                         goto undo;
2647
2648                 ret = sched_dl_global_validate();
2649                 if (ret)
2650                         goto undo;
2651
2652                 ret = sched_rt_global_constraints();
2653                 if (ret)
2654                         goto undo;
2655
2656                 sched_rt_do_global();
2657                 sched_dl_do_global();
2658         }
2659         if (0) {
2660 undo:
2661                 sysctl_sched_rt_period = old_period;
2662                 sysctl_sched_rt_runtime = old_runtime;
2663         }
2664         mutex_unlock(&mutex);
2665
2666         return ret;
2667 }
2668
2669 int sched_rr_handler(struct ctl_table *table, int write,
2670                 void __user *buffer, size_t *lenp,
2671                 loff_t *ppos)
2672 {
2673         int ret;
2674         static DEFINE_MUTEX(mutex);
2675
2676         mutex_lock(&mutex);
2677         ret = proc_dointvec(table, write, buffer, lenp, ppos);
2678         /*
2679          * Make sure that internally we keep jiffies.
2680          * Also, writing zero resets the timeslice to default:
2681          */
2682         if (!ret && write) {
2683                 sched_rr_timeslice =
2684                         sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2685                         msecs_to_jiffies(sysctl_sched_rr_timeslice);
2686         }
2687         mutex_unlock(&mutex);
2688         return ret;
2689 }
2690
2691 #ifdef CONFIG_SCHED_DEBUG
2692 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2693
2694 void print_rt_stats(struct seq_file *m, int cpu)
2695 {
2696         rt_rq_iter_t iter;
2697         struct rt_rq *rt_rq;
2698
2699         rcu_read_lock();
2700         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2701                 print_rt_rq(m, cpu, rt_rq);
2702         rcu_read_unlock();
2703 }
2704 #endif /* CONFIG_SCHED_DEBUG */