projects/linux·blob v4.13 · refs

2775 lines · 65.7 KiB

1/*
2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3 * policies)
4 */
5
6#include "sched.h"
7
8#include <linux/slab.h>
9#include <linux/irq_work.h>
10
11int sched_rr_timeslice = RR_TIMESLICE;
12int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
13
14static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
15
16struct rt_bandwidth def_rt_bandwidth;
17
18static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
19{
20 struct rt_bandwidth *rt_b =
21 container_of(timer, struct rt_bandwidth, rt_period_timer);
22 int idle = 0;
23 int overrun;
24
25 raw_spin_lock(&rt_b->rt_runtime_lock);
26 for (;;) {
27 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
28 if (!overrun)
29 break;
30
31 raw_spin_unlock(&rt_b->rt_runtime_lock);
32 idle = do_sched_rt_period_timer(rt_b, overrun);
33 raw_spin_lock(&rt_b->rt_runtime_lock);
34 }
35 if (idle)
36 rt_b->rt_period_active = 0;
37 raw_spin_unlock(&rt_b->rt_runtime_lock);
38
39 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
40}
41
42void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
43{
44 rt_b->rt_period = ns_to_ktime(period);
45 rt_b->rt_runtime = runtime;
46
47 raw_spin_lock_init(&rt_b->rt_runtime_lock);
48
49 hrtimer_init(&rt_b->rt_period_timer,
50 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
51 rt_b->rt_period_timer.function = sched_rt_period_timer;
52}
53
54static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
55{
56 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
57 return;
58
59 raw_spin_lock(&rt_b->rt_runtime_lock);
60 if (!rt_b->rt_period_active) {
61 rt_b->rt_period_active = 1;
62 /*
63 * SCHED_DEADLINE updates the bandwidth, as a run away
64 * RT task with a DL task could hog a CPU. But DL does
65 * not reset the period. If a deadline task was running
66 * without an RT task running, it can cause RT tasks to
67 * throttle when they start up. Kick the timer right away
68 * to update the period.
69 */
70 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
71 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
72 }
73 raw_spin_unlock(&rt_b->rt_runtime_lock);
74}
75
76#if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI)
77static void push_irq_work_func(struct irq_work *work);
78#endif
79
80void init_rt_rq(struct rt_rq *rt_rq)
81{
82 struct rt_prio_array *array;
83 int i;
84
85 array = &rt_rq->active;
86 for (i = 0; i < MAX_RT_PRIO; i++) {
87 INIT_LIST_HEAD(array->queue + i);
88 __clear_bit(i, array->bitmap);
89 }
90 /* delimiter for bitsearch: */
91 __set_bit(MAX_RT_PRIO, array->bitmap);
92
93#if defined CONFIG_SMP
94 rt_rq->highest_prio.curr = MAX_RT_PRIO;
95 rt_rq->highest_prio.next = MAX_RT_PRIO;
96 rt_rq->rt_nr_migratory = 0;
97 rt_rq->overloaded = 0;
98 plist_head_init(&rt_rq->pushable_tasks);
99
100#ifdef HAVE_RT_PUSH_IPI
101 rt_rq->push_flags = 0;
102 rt_rq->push_cpu = nr_cpu_ids;
103 raw_spin_lock_init(&rt_rq->push_lock);
104 init_irq_work(&rt_rq->push_work, push_irq_work_func);
105#endif
106#endif /* CONFIG_SMP */
107 /* We start is dequeued state, because no RT tasks are queued */
108 rt_rq->rt_queued = 0;
109
110 rt_rq->rt_time = 0;
111 rt_rq->rt_throttled = 0;
112 rt_rq->rt_runtime = 0;
113 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
114}
115
116#ifdef CONFIG_RT_GROUP_SCHED
117static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
118{
119 hrtimer_cancel(&rt_b->rt_period_timer);
120}
121
122#define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
123
124static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
125{
126#ifdef CONFIG_SCHED_DEBUG
127 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
128#endif
129 return container_of(rt_se, struct task_struct, rt);
130}
131
132static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
133{
134 return rt_rq->rq;
135}
136
137static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
138{
139 return rt_se->rt_rq;
140}
141
142static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
143{
144 struct rt_rq *rt_rq = rt_se->rt_rq;
145
146 return rt_rq->rq;
147}
148
149void free_rt_sched_group(struct task_group *tg)
150{
151 int i;
152
153 if (tg->rt_se)
154 destroy_rt_bandwidth(&tg->rt_bandwidth);
155
156 for_each_possible_cpu(i) {
157 if (tg->rt_rq)
158 kfree(tg->rt_rq[i]);
159 if (tg->rt_se)
160 kfree(tg->rt_se[i]);
161 }
162
163 kfree(tg->rt_rq);
164 kfree(tg->rt_se);
165}
166
167void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
168 struct sched_rt_entity *rt_se, int cpu,
169 struct sched_rt_entity *parent)
170{
171 struct rq *rq = cpu_rq(cpu);
172
173 rt_rq->highest_prio.curr = MAX_RT_PRIO;
174 rt_rq->rt_nr_boosted = 0;
175 rt_rq->rq = rq;
176 rt_rq->tg = tg;
177
178 tg->rt_rq[cpu] = rt_rq;
179 tg->rt_se[cpu] = rt_se;
180
181 if (!rt_se)
182 return;
183
184 if (!parent)
185 rt_se->rt_rq = &rq->rt;
186 else
187 rt_se->rt_rq = parent->my_q;
188
189 rt_se->my_q = rt_rq;
190 rt_se->parent = parent;
191 INIT_LIST_HEAD(&rt_se->run_list);
192}
193
194int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
195{
196 struct rt_rq *rt_rq;
197 struct sched_rt_entity *rt_se;
198 int i;
199
200 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
201 if (!tg->rt_rq)
202 goto err;
203 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
204 if (!tg->rt_se)
205 goto err;
206
207 init_rt_bandwidth(&tg->rt_bandwidth,
208 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
209
210 for_each_possible_cpu(i) {
211 rt_rq = kzalloc_node(sizeof(struct rt_rq),
212 GFP_KERNEL, cpu_to_node(i));
213 if (!rt_rq)
214 goto err;
215
216 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
217 GFP_KERNEL, cpu_to_node(i));
218 if (!rt_se)
219 goto err_free_rq;
220
221 init_rt_rq(rt_rq);
222 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
223 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
224 }
225
226 return 1;
227
228err_free_rq:
229 kfree(rt_rq);
230err:
231 return 0;
232}
233
234#else /* CONFIG_RT_GROUP_SCHED */
235
236#define rt_entity_is_task(rt_se) (1)
237
238static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
239{
240 return container_of(rt_se, struct task_struct, rt);
241}
242
243static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
244{
245 return container_of(rt_rq, struct rq, rt);
246}
247
248static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
249{
250 struct task_struct *p = rt_task_of(rt_se);
251
252 return task_rq(p);
253}
254
255static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
256{
257 struct rq *rq = rq_of_rt_se(rt_se);
258
259 return &rq->rt;
260}
261
262void free_rt_sched_group(struct task_group *tg) { }
263
264int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
265{
266 return 1;
267}
268#endif /* CONFIG_RT_GROUP_SCHED */
269
270#ifdef CONFIG_SMP
271
272static void pull_rt_task(struct rq *this_rq);
273
274static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
275{
276 /* Try to pull RT tasks here if we lower this rq's prio */
277 return rq->rt.highest_prio.curr > prev->prio;
278}
279
280static inline int rt_overloaded(struct rq *rq)
281{
282 return atomic_read(&rq->rd->rto_count);
283}
284
285static inline void rt_set_overload(struct rq *rq)
286{
287 if (!rq->online)
288 return;
289
290 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
291 /*
292 * Make sure the mask is visible before we set
293 * the overload count. That is checked to determine
294 * if we should look at the mask. It would be a shame
295 * if we looked at the mask, but the mask was not
296 * updated yet.
297 *
298 * Matched by the barrier in pull_rt_task().
299 */
300 smp_wmb();
301 atomic_inc(&rq->rd->rto_count);
302}
303
304static inline void rt_clear_overload(struct rq *rq)
305{
306 if (!rq->online)
307 return;
308
309 /* the order here really doesn't matter */
310 atomic_dec(&rq->rd->rto_count);
311 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
312}
313
314static void update_rt_migration(struct rt_rq *rt_rq)
315{
316 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
317 if (!rt_rq->overloaded) {
318 rt_set_overload(rq_of_rt_rq(rt_rq));
319 rt_rq->overloaded = 1;
320 }
321 } else if (rt_rq->overloaded) {
322 rt_clear_overload(rq_of_rt_rq(rt_rq));
323 rt_rq->overloaded = 0;
324 }
325}
326
327static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
328{
329 struct task_struct *p;
330
331 if (!rt_entity_is_task(rt_se))
332 return;
333
334 p = rt_task_of(rt_se);
335 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
336
337 rt_rq->rt_nr_total++;
338 if (p->nr_cpus_allowed > 1)
339 rt_rq->rt_nr_migratory++;
340
341 update_rt_migration(rt_rq);
342}
343
344static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
345{
346 struct task_struct *p;
347
348 if (!rt_entity_is_task(rt_se))
349 return;
350
351 p = rt_task_of(rt_se);
352 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
353
354 rt_rq->rt_nr_total--;
355 if (p->nr_cpus_allowed > 1)
356 rt_rq->rt_nr_migratory--;
357
358 update_rt_migration(rt_rq);
359}
360
361static inline int has_pushable_tasks(struct rq *rq)
362{
363 return !plist_head_empty(&rq->rt.pushable_tasks);
364}
365
366static DEFINE_PER_CPU(struct callback_head, rt_push_head);
367static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
368
369static void push_rt_tasks(struct rq *);
370static void pull_rt_task(struct rq *);
371
372static inline void queue_push_tasks(struct rq *rq)
373{
374 if (!has_pushable_tasks(rq))
375 return;
376
377 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
378}
379
380static inline void queue_pull_task(struct rq *rq)
381{
382 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
383}
384
385static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
386{
387 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
388 plist_node_init(&p->pushable_tasks, p->prio);
389 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
390
391 /* Update the highest prio pushable task */
392 if (p->prio < rq->rt.highest_prio.next)
393 rq->rt.highest_prio.next = p->prio;
394}
395
396static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
397{
398 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
399
400 /* Update the new highest prio pushable task */
401 if (has_pushable_tasks(rq)) {
402 p = plist_first_entry(&rq->rt.pushable_tasks,
403 struct task_struct, pushable_tasks);
404 rq->rt.highest_prio.next = p->prio;
405 } else
406 rq->rt.highest_prio.next = MAX_RT_PRIO;
407}
408
409#else
410
411static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
412{
413}
414
415static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
416{
417}
418
419static inline
420void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
421{
422}
423
424static inline
425void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
426{
427}
428
429static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
430{
431 return false;
432}
433
434static inline void pull_rt_task(struct rq *this_rq)
435{
436}
437
438static inline void queue_push_tasks(struct rq *rq)
439{
440}
441#endif /* CONFIG_SMP */
442
443static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
444static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
445
446static inline int on_rt_rq(struct sched_rt_entity *rt_se)
447{
448 return rt_se->on_rq;
449}
450
451#ifdef CONFIG_RT_GROUP_SCHED
452
453static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
454{
455 if (!rt_rq->tg)
456 return RUNTIME_INF;
457
458 return rt_rq->rt_runtime;
459}
460
461static inline u64 sched_rt_period(struct rt_rq *rt_rq)
462{
463 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
464}
465
466typedef struct task_group *rt_rq_iter_t;
467
468static inline struct task_group *next_task_group(struct task_group *tg)
469{
470 do {
471 tg = list_entry_rcu(tg->list.next,
472 typeof(struct task_group), list);
473 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
474
475 if (&tg->list == &task_groups)
476 tg = NULL;
477
478 return tg;
479}
480
481#define for_each_rt_rq(rt_rq, iter, rq) \
482 for (iter = container_of(&task_groups, typeof(*iter), list); \
483 (iter = next_task_group(iter)) && \
484 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
485
486#define for_each_sched_rt_entity(rt_se) \
487 for (; rt_se; rt_se = rt_se->parent)
488
489static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
490{
491 return rt_se->my_q;
492}
493
494static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
495static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
496
497static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
498{
499 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
500 struct rq *rq = rq_of_rt_rq(rt_rq);
501 struct sched_rt_entity *rt_se;
502
503 int cpu = cpu_of(rq);
504
505 rt_se = rt_rq->tg->rt_se[cpu];
506
507 if (rt_rq->rt_nr_running) {
508 if (!rt_se)
509 enqueue_top_rt_rq(rt_rq);
510 else if (!on_rt_rq(rt_se))
511 enqueue_rt_entity(rt_se, 0);
512
513 if (rt_rq->highest_prio.curr < curr->prio)
514 resched_curr(rq);
515 }
516}
517
518static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
519{
520 struct sched_rt_entity *rt_se;
521 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
522
523 rt_se = rt_rq->tg->rt_se[cpu];
524
525 if (!rt_se)
526 dequeue_top_rt_rq(rt_rq);
527 else if (on_rt_rq(rt_se))
528 dequeue_rt_entity(rt_se, 0);
529}
530
531static inline int rt_rq_throttled(struct rt_rq *rt_rq)
532{
533 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
534}
535
536static int rt_se_boosted(struct sched_rt_entity *rt_se)
537{
538 struct rt_rq *rt_rq = group_rt_rq(rt_se);
539 struct task_struct *p;
540
541 if (rt_rq)
542 return !!rt_rq->rt_nr_boosted;
543
544 p = rt_task_of(rt_se);
545 return p->prio != p->normal_prio;
546}
547
548#ifdef CONFIG_SMP
549static inline const struct cpumask *sched_rt_period_mask(void)
550{
551 return this_rq()->rd->span;
552}
553#else
554static inline const struct cpumask *sched_rt_period_mask(void)
555{
556 return cpu_online_mask;
557}
558#endif
559
560static inline
561struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
562{
563 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
564}
565
566static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
567{
568 return &rt_rq->tg->rt_bandwidth;
569}
570
571#else /* !CONFIG_RT_GROUP_SCHED */
572
573static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
574{
575 return rt_rq->rt_runtime;
576}
577
578static inline u64 sched_rt_period(struct rt_rq *rt_rq)
579{
580 return ktime_to_ns(def_rt_bandwidth.rt_period);
581}
582
583typedef struct rt_rq *rt_rq_iter_t;
584
585#define for_each_rt_rq(rt_rq, iter, rq) \
586 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
587
588#define for_each_sched_rt_entity(rt_se) \
589 for (; rt_se; rt_se = NULL)
590
591static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
592{
593 return NULL;
594}
595
596static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
597{
598 struct rq *rq = rq_of_rt_rq(rt_rq);
599
600 if (!rt_rq->rt_nr_running)
601 return;
602
603 enqueue_top_rt_rq(rt_rq);
604 resched_curr(rq);
605}
606
607static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
608{
609 dequeue_top_rt_rq(rt_rq);
610}
611
612static inline int rt_rq_throttled(struct rt_rq *rt_rq)
613{
614 return rt_rq->rt_throttled;
615}
616
617static inline const struct cpumask *sched_rt_period_mask(void)
618{
619 return cpu_online_mask;
620}
621
622static inline
623struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
624{
625 return &cpu_rq(cpu)->rt;
626}
627
628static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
629{
630 return &def_rt_bandwidth;
631}
632
633#endif /* CONFIG_RT_GROUP_SCHED */
634
635bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
636{
637 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
638
639 return (hrtimer_active(&rt_b->rt_period_timer) ||
640 rt_rq->rt_time < rt_b->rt_runtime);
641}
642
643#ifdef CONFIG_SMP
644/*
645 * We ran out of runtime, see if we can borrow some from our neighbours.
646 */
647static void do_balance_runtime(struct rt_rq *rt_rq)
648{
649 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
650 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
651 int i, weight;
652 u64 rt_period;
653
654 weight = cpumask_weight(rd->span);
655
656 raw_spin_lock(&rt_b->rt_runtime_lock);
657 rt_period = ktime_to_ns(rt_b->rt_period);
658 for_each_cpu(i, rd->span) {
659 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
660 s64 diff;
661
662 if (iter == rt_rq)
663 continue;
664
665 raw_spin_lock(&iter->rt_runtime_lock);
666 /*
667 * Either all rqs have inf runtime and there's nothing to steal
668 * or __disable_runtime() below sets a specific rq to inf to
669 * indicate its been disabled and disalow stealing.
670 */
671 if (iter->rt_runtime == RUNTIME_INF)
672 goto next;
673
674 /*
675 * From runqueues with spare time, take 1/n part of their
676 * spare time, but no more than our period.
677 */
678 diff = iter->rt_runtime - iter->rt_time;
679 if (diff > 0) {
680 diff = div_u64((u64)diff, weight);
681 if (rt_rq->rt_runtime + diff > rt_period)
682 diff = rt_period - rt_rq->rt_runtime;
683 iter->rt_runtime -= diff;
684 rt_rq->rt_runtime += diff;
685 if (rt_rq->rt_runtime == rt_period) {
686 raw_spin_unlock(&iter->rt_runtime_lock);
687 break;
688 }
689 }
690next:
691 raw_spin_unlock(&iter->rt_runtime_lock);
692 }
693 raw_spin_unlock(&rt_b->rt_runtime_lock);
694}
695
696/*
697 * Ensure this RQ takes back all the runtime it lend to its neighbours.
698 */
699static void __disable_runtime(struct rq *rq)
700{
701 struct root_domain *rd = rq->rd;
702 rt_rq_iter_t iter;
703 struct rt_rq *rt_rq;
704
705 if (unlikely(!scheduler_running))
706 return;
707
708 for_each_rt_rq(rt_rq, iter, rq) {
709 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
710 s64 want;
711 int i;
712
713 raw_spin_lock(&rt_b->rt_runtime_lock);
714 raw_spin_lock(&rt_rq->rt_runtime_lock);
715 /*
716 * Either we're all inf and nobody needs to borrow, or we're
717 * already disabled and thus have nothing to do, or we have
718 * exactly the right amount of runtime to take out.
719 */
720 if (rt_rq->rt_runtime == RUNTIME_INF ||
721 rt_rq->rt_runtime == rt_b->rt_runtime)
722 goto balanced;
723 raw_spin_unlock(&rt_rq->rt_runtime_lock);
724
725 /*
726 * Calculate the difference between what we started out with
727 * and what we current have, that's the amount of runtime
728 * we lend and now have to reclaim.
729 */
730 want = rt_b->rt_runtime - rt_rq->rt_runtime;
731
732 /*
733 * Greedy reclaim, take back as much as we can.
734 */
735 for_each_cpu(i, rd->span) {
736 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
737 s64 diff;
738
739 /*
740 * Can't reclaim from ourselves or disabled runqueues.
741 */
742 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
743 continue;
744
745 raw_spin_lock(&iter->rt_runtime_lock);
746 if (want > 0) {
747 diff = min_t(s64, iter->rt_runtime, want);
748 iter->rt_runtime -= diff;
749 want -= diff;
750 } else {
751 iter->rt_runtime -= want;
752 want -= want;
753 }
754 raw_spin_unlock(&iter->rt_runtime_lock);
755
756 if (!want)
757 break;
758 }
759
760 raw_spin_lock(&rt_rq->rt_runtime_lock);
761 /*
762 * We cannot be left wanting - that would mean some runtime
763 * leaked out of the system.
764 */
765 BUG_ON(want);
766balanced:
767 /*
768 * Disable all the borrow logic by pretending we have inf
769 * runtime - in which case borrowing doesn't make sense.
770 */
771 rt_rq->rt_runtime = RUNTIME_INF;
772 rt_rq->rt_throttled = 0;
773 raw_spin_unlock(&rt_rq->rt_runtime_lock);
774 raw_spin_unlock(&rt_b->rt_runtime_lock);
775
776 /* Make rt_rq available for pick_next_task() */
777 sched_rt_rq_enqueue(rt_rq);
778 }
779}
780
781static void __enable_runtime(struct rq *rq)
782{
783 rt_rq_iter_t iter;
784 struct rt_rq *rt_rq;
785
786 if (unlikely(!scheduler_running))
787 return;
788
789 /*
790 * Reset each runqueue's bandwidth settings
791 */
792 for_each_rt_rq(rt_rq, iter, rq) {
793 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
794
795 raw_spin_lock(&rt_b->rt_runtime_lock);
796 raw_spin_lock(&rt_rq->rt_runtime_lock);
797 rt_rq->rt_runtime = rt_b->rt_runtime;
798 rt_rq->rt_time = 0;
799 rt_rq->rt_throttled = 0;
800 raw_spin_unlock(&rt_rq->rt_runtime_lock);
801 raw_spin_unlock(&rt_b->rt_runtime_lock);
802 }
803}
804
805static void balance_runtime(struct rt_rq *rt_rq)
806{
807 if (!sched_feat(RT_RUNTIME_SHARE))
808 return;
809
810 if (rt_rq->rt_time > rt_rq->rt_runtime) {
811 raw_spin_unlock(&rt_rq->rt_runtime_lock);
812 do_balance_runtime(rt_rq);
813 raw_spin_lock(&rt_rq->rt_runtime_lock);
814 }
815}
816#else /* !CONFIG_SMP */
817static inline void balance_runtime(struct rt_rq *rt_rq) {}
818#endif /* CONFIG_SMP */
819
820static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
821{
822 int i, idle = 1, throttled = 0;
823 const struct cpumask *span;
824
825 span = sched_rt_period_mask();
826#ifdef CONFIG_RT_GROUP_SCHED
827 /*
828 * FIXME: isolated CPUs should really leave the root task group,
829 * whether they are isolcpus or were isolated via cpusets, lest
830 * the timer run on a CPU which does not service all runqueues,
831 * potentially leaving other CPUs indefinitely throttled. If
832 * isolation is really required, the user will turn the throttle
833 * off to kill the perturbations it causes anyway. Meanwhile,
834 * this maintains functionality for boot and/or troubleshooting.
835 */
836 if (rt_b == &root_task_group.rt_bandwidth)
837 span = cpu_online_mask;
838#endif
839 for_each_cpu(i, span) {
840 int enqueue = 0;
841 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
842 struct rq *rq = rq_of_rt_rq(rt_rq);
843 int skip;
844
845 /*
846 * When span == cpu_online_mask, taking each rq->lock
847 * can be time-consuming. Try to avoid it when possible.
848 */
849 raw_spin_lock(&rt_rq->rt_runtime_lock);
850 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running;
851 raw_spin_unlock(&rt_rq->rt_runtime_lock);
852 if (skip)
853 continue;
854
855 raw_spin_lock(&rq->lock);
856 if (rt_rq->rt_time) {
857 u64 runtime;
858
859 raw_spin_lock(&rt_rq->rt_runtime_lock);
860 if (rt_rq->rt_throttled)
861 balance_runtime(rt_rq);
862 runtime = rt_rq->rt_runtime;
863 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
864 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
865 rt_rq->rt_throttled = 0;
866 enqueue = 1;
867
868 /*
869 * When we're idle and a woken (rt) task is
870 * throttled check_preempt_curr() will set
871 * skip_update and the time between the wakeup
872 * and this unthrottle will get accounted as
873 * 'runtime'.
874 */
875 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
876 rq_clock_skip_update(rq, false);
877 }
878 if (rt_rq->rt_time || rt_rq->rt_nr_running)
879 idle = 0;
880 raw_spin_unlock(&rt_rq->rt_runtime_lock);
881 } else if (rt_rq->rt_nr_running) {
882 idle = 0;
883 if (!rt_rq_throttled(rt_rq))
884 enqueue = 1;
885 }
886 if (rt_rq->rt_throttled)
887 throttled = 1;
888
889 if (enqueue)
890 sched_rt_rq_enqueue(rt_rq);
891 raw_spin_unlock(&rq->lock);
892 }
893
894 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
895 return 1;
896
897 return idle;
898}
899
900static inline int rt_se_prio(struct sched_rt_entity *rt_se)
901{
902#ifdef CONFIG_RT_GROUP_SCHED
903 struct rt_rq *rt_rq = group_rt_rq(rt_se);
904
905 if (rt_rq)
906 return rt_rq->highest_prio.curr;
907#endif
908
909 return rt_task_of(rt_se)->prio;
910}
911
912static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
913{
914 u64 runtime = sched_rt_runtime(rt_rq);
915
916 if (rt_rq->rt_throttled)
917 return rt_rq_throttled(rt_rq);
918
919 if (runtime >= sched_rt_period(rt_rq))
920 return 0;
921
922 balance_runtime(rt_rq);
923 runtime = sched_rt_runtime(rt_rq);
924 if (runtime == RUNTIME_INF)
925 return 0;
926
927 if (rt_rq->rt_time > runtime) {
928 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
929
930 /*
931 * Don't actually throttle groups that have no runtime assigned
932 * but accrue some time due to boosting.
933 */
934 if (likely(rt_b->rt_runtime)) {
935 rt_rq->rt_throttled = 1;
936 printk_deferred_once("sched: RT throttling activated\n");
937 } else {
938 /*
939 * In case we did anyway, make it go away,
940 * replenishment is a joke, since it will replenish us
941 * with exactly 0 ns.
942 */
943 rt_rq->rt_time = 0;
944 }
945
946 if (rt_rq_throttled(rt_rq)) {
947 sched_rt_rq_dequeue(rt_rq);
948 return 1;
949 }
950 }
951
952 return 0;
953}
954
955/*
956 * Update the current task's runtime statistics. Skip current tasks that
957 * are not in our scheduling class.
958 */
959static void update_curr_rt(struct rq *rq)
960{
961 struct task_struct *curr = rq->curr;
962 struct sched_rt_entity *rt_se = &curr->rt;
963 u64 delta_exec;
964
965 if (curr->sched_class != &rt_sched_class)
966 return;
967
968 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
969 if (unlikely((s64)delta_exec <= 0))
970 return;
971
972 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
973 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_RT);
974
975 schedstat_set(curr->se.statistics.exec_max,
976 max(curr->se.statistics.exec_max, delta_exec));
977
978 curr->se.sum_exec_runtime += delta_exec;
979 account_group_exec_runtime(curr, delta_exec);
980
981 curr->se.exec_start = rq_clock_task(rq);
982 cpuacct_charge(curr, delta_exec);
983
984 sched_rt_avg_update(rq, delta_exec);
985
986 if (!rt_bandwidth_enabled())
987 return;
988
989 for_each_sched_rt_entity(rt_se) {
990 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
991
992 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
993 raw_spin_lock(&rt_rq->rt_runtime_lock);
994 rt_rq->rt_time += delta_exec;
995 if (sched_rt_runtime_exceeded(rt_rq))
996 resched_curr(rq);
997 raw_spin_unlock(&rt_rq->rt_runtime_lock);
998 }
999 }
1000}
1001
1002static void
1003dequeue_top_rt_rq(struct rt_rq *rt_rq)
1004{
1005 struct rq *rq = rq_of_rt_rq(rt_rq);
1006
1007 BUG_ON(&rq->rt != rt_rq);
1008
1009 if (!rt_rq->rt_queued)
1010 return;
1011
1012 BUG_ON(!rq->nr_running);
1013
1014 sub_nr_running(rq, rt_rq->rt_nr_running);
1015 rt_rq->rt_queued = 0;
1016}
1017
1018static void
1019enqueue_top_rt_rq(struct rt_rq *rt_rq)
1020{
1021 struct rq *rq = rq_of_rt_rq(rt_rq);
1022
1023 BUG_ON(&rq->rt != rt_rq);
1024
1025 if (rt_rq->rt_queued)
1026 return;
1027 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1028 return;
1029
1030 add_nr_running(rq, rt_rq->rt_nr_running);
1031 rt_rq->rt_queued = 1;
1032}
1033
1034#if defined CONFIG_SMP
1035
1036static void
1037inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1038{
1039 struct rq *rq = rq_of_rt_rq(rt_rq);
1040
1041#ifdef CONFIG_RT_GROUP_SCHED
1042 /*
1043 * Change rq's cpupri only if rt_rq is the top queue.
1044 */
1045 if (&rq->rt != rt_rq)
1046 return;
1047#endif
1048 if (rq->online && prio < prev_prio)
1049 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1050}
1051
1052static void
1053dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1054{
1055 struct rq *rq = rq_of_rt_rq(rt_rq);
1056
1057#ifdef CONFIG_RT_GROUP_SCHED
1058 /*
1059 * Change rq's cpupri only if rt_rq is the top queue.
1060 */
1061 if (&rq->rt != rt_rq)
1062 return;
1063#endif
1064 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1065 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1066}
1067
1068#else /* CONFIG_SMP */
1069
1070static inline
1071void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1072static inline
1073void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1074
1075#endif /* CONFIG_SMP */
1076
1077#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1078static void
1079inc_rt_prio(struct rt_rq *rt_rq, int prio)
1080{
1081 int prev_prio = rt_rq->highest_prio.curr;
1082
1083 if (prio < prev_prio)
1084 rt_rq->highest_prio.curr = prio;
1085
1086 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1087}
1088
1089static void
1090dec_rt_prio(struct rt_rq *rt_rq, int prio)
1091{
1092 int prev_prio = rt_rq->highest_prio.curr;
1093
1094 if (rt_rq->rt_nr_running) {
1095
1096 WARN_ON(prio < prev_prio);
1097
1098 /*
1099 * This may have been our highest task, and therefore
1100 * we may have some recomputation to do
1101 */
1102 if (prio == prev_prio) {
1103 struct rt_prio_array *array = &rt_rq->active;
1104
1105 rt_rq->highest_prio.curr =
1106 sched_find_first_bit(array->bitmap);
1107 }
1108
1109 } else
1110 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1111
1112 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1113}
1114
1115#else
1116
1117static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1118static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1119
1120#endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1121
1122#ifdef CONFIG_RT_GROUP_SCHED
1123
1124static void
1125inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1126{
1127 if (rt_se_boosted(rt_se))
1128 rt_rq->rt_nr_boosted++;
1129
1130 if (rt_rq->tg)
1131 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1132}
1133
1134static void
1135dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1136{
1137 if (rt_se_boosted(rt_se))
1138 rt_rq->rt_nr_boosted--;
1139
1140 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1141}
1142
1143#else /* CONFIG_RT_GROUP_SCHED */
1144
1145static void
1146inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1147{
1148 start_rt_bandwidth(&def_rt_bandwidth);
1149}
1150
1151static inline
1152void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1153
1154#endif /* CONFIG_RT_GROUP_SCHED */
1155
1156static inline
1157unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1158{
1159 struct rt_rq *group_rq = group_rt_rq(rt_se);
1160
1161 if (group_rq)
1162 return group_rq->rt_nr_running;
1163 else
1164 return 1;
1165}
1166
1167static inline
1168unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1169{
1170 struct rt_rq *group_rq = group_rt_rq(rt_se);
1171 struct task_struct *tsk;
1172
1173 if (group_rq)
1174 return group_rq->rr_nr_running;
1175
1176 tsk = rt_task_of(rt_se);
1177
1178 return (tsk->policy == SCHED_RR) ? 1 : 0;
1179}
1180
1181static inline
1182void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1183{
1184 int prio = rt_se_prio(rt_se);
1185
1186 WARN_ON(!rt_prio(prio));
1187 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1188 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1189
1190 inc_rt_prio(rt_rq, prio);
1191 inc_rt_migration(rt_se, rt_rq);
1192 inc_rt_group(rt_se, rt_rq);
1193}
1194
1195static inline
1196void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1197{
1198 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1199 WARN_ON(!rt_rq->rt_nr_running);
1200 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1201 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1202
1203 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1204 dec_rt_migration(rt_se, rt_rq);
1205 dec_rt_group(rt_se, rt_rq);
1206}
1207
1208/*
1209 * Change rt_se->run_list location unless SAVE && !MOVE
1210 *
1211 * assumes ENQUEUE/DEQUEUE flags match
1212 */
1213static inline bool move_entity(unsigned int flags)
1214{
1215 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1216 return false;
1217
1218 return true;
1219}
1220
1221static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1222{
1223 list_del_init(&rt_se->run_list);
1224
1225 if (list_empty(array->queue + rt_se_prio(rt_se)))
1226 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1227
1228 rt_se->on_list = 0;
1229}
1230
1231static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1232{
1233 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1234 struct rt_prio_array *array = &rt_rq->active;
1235 struct rt_rq *group_rq = group_rt_rq(rt_se);
1236 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1237
1238 /*
1239 * Don't enqueue the group if its throttled, or when empty.
1240 * The latter is a consequence of the former when a child group
1241 * get throttled and the current group doesn't have any other
1242 * active members.
1243 */
1244 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1245 if (rt_se->on_list)
1246 __delist_rt_entity(rt_se, array);
1247 return;
1248 }
1249
1250 if (move_entity(flags)) {
1251 WARN_ON_ONCE(rt_se->on_list);
1252 if (flags & ENQUEUE_HEAD)
1253 list_add(&rt_se->run_list, queue);
1254 else
1255 list_add_tail(&rt_se->run_list, queue);
1256
1257 __set_bit(rt_se_prio(rt_se), array->bitmap);
1258 rt_se->on_list = 1;
1259 }
1260 rt_se->on_rq = 1;
1261
1262 inc_rt_tasks(rt_se, rt_rq);
1263}
1264
1265static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1266{
1267 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1268 struct rt_prio_array *array = &rt_rq->active;
1269
1270 if (move_entity(flags)) {
1271 WARN_ON_ONCE(!rt_se->on_list);
1272 __delist_rt_entity(rt_se, array);
1273 }
1274 rt_se->on_rq = 0;
1275
1276 dec_rt_tasks(rt_se, rt_rq);
1277}
1278
1279/*
1280 * Because the prio of an upper entry depends on the lower
1281 * entries, we must remove entries top - down.
1282 */
1283static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1284{
1285 struct sched_rt_entity *back = NULL;
1286
1287 for_each_sched_rt_entity(rt_se) {
1288 rt_se->back = back;
1289 back = rt_se;
1290 }
1291
1292 dequeue_top_rt_rq(rt_rq_of_se(back));
1293
1294 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1295 if (on_rt_rq(rt_se))
1296 __dequeue_rt_entity(rt_se, flags);
1297 }
1298}
1299
1300static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1301{
1302 struct rq *rq = rq_of_rt_se(rt_se);
1303
1304 dequeue_rt_stack(rt_se, flags);
1305 for_each_sched_rt_entity(rt_se)
1306 __enqueue_rt_entity(rt_se, flags);
1307 enqueue_top_rt_rq(&rq->rt);
1308}
1309
1310static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1311{
1312 struct rq *rq = rq_of_rt_se(rt_se);
1313
1314 dequeue_rt_stack(rt_se, flags);
1315
1316 for_each_sched_rt_entity(rt_se) {
1317 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1318
1319 if (rt_rq && rt_rq->rt_nr_running)
1320 __enqueue_rt_entity(rt_se, flags);
1321 }
1322 enqueue_top_rt_rq(&rq->rt);
1323}
1324
1325/*
1326 * Adding/removing a task to/from a priority array:
1327 */
1328static void
1329enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1330{
1331 struct sched_rt_entity *rt_se = &p->rt;
1332
1333 if (flags & ENQUEUE_WAKEUP)
1334 rt_se->timeout = 0;
1335
1336 enqueue_rt_entity(rt_se, flags);
1337
1338 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1339 enqueue_pushable_task(rq, p);
1340}
1341
1342static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1343{
1344 struct sched_rt_entity *rt_se = &p->rt;
1345
1346 update_curr_rt(rq);
1347 dequeue_rt_entity(rt_se, flags);
1348
1349 dequeue_pushable_task(rq, p);
1350}
1351
1352/*
1353 * Put task to the head or the end of the run list without the overhead of
1354 * dequeue followed by enqueue.
1355 */
1356static void
1357requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1358{
1359 if (on_rt_rq(rt_se)) {
1360 struct rt_prio_array *array = &rt_rq->active;
1361 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1362
1363 if (head)
1364 list_move(&rt_se->run_list, queue);
1365 else
1366 list_move_tail(&rt_se->run_list, queue);
1367 }
1368}
1369
1370static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1371{
1372 struct sched_rt_entity *rt_se = &p->rt;
1373 struct rt_rq *rt_rq;
1374
1375 for_each_sched_rt_entity(rt_se) {
1376 rt_rq = rt_rq_of_se(rt_se);
1377 requeue_rt_entity(rt_rq, rt_se, head);
1378 }
1379}
1380
1381static void yield_task_rt(struct rq *rq)
1382{
1383 requeue_task_rt(rq, rq->curr, 0);
1384}
1385
1386#ifdef CONFIG_SMP
1387static int find_lowest_rq(struct task_struct *task);
1388
1389static int
1390select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1391{
1392 struct task_struct *curr;
1393 struct rq *rq;
1394
1395 /* For anything but wake ups, just return the task_cpu */
1396 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1397 goto out;
1398
1399 rq = cpu_rq(cpu);
1400
1401 rcu_read_lock();
1402 curr = READ_ONCE(rq->curr); /* unlocked access */
1403
1404 /*
1405 * If the current task on @p's runqueue is an RT task, then
1406 * try to see if we can wake this RT task up on another
1407 * runqueue. Otherwise simply start this RT task
1408 * on its current runqueue.
1409 *
1410 * We want to avoid overloading runqueues. If the woken
1411 * task is a higher priority, then it will stay on this CPU
1412 * and the lower prio task should be moved to another CPU.
1413 * Even though this will probably make the lower prio task
1414 * lose its cache, we do not want to bounce a higher task
1415 * around just because it gave up its CPU, perhaps for a
1416 * lock?
1417 *
1418 * For equal prio tasks, we just let the scheduler sort it out.
1419 *
1420 * Otherwise, just let it ride on the affined RQ and the
1421 * post-schedule router will push the preempted task away
1422 *
1423 * This test is optimistic, if we get it wrong the load-balancer
1424 * will have to sort it out.
1425 */
1426 if (curr && unlikely(rt_task(curr)) &&
1427 (curr->nr_cpus_allowed < 2 ||
1428 curr->prio <= p->prio)) {
1429 int target = find_lowest_rq(p);
1430
1431 /*
1432 * Don't bother moving it if the destination CPU is
1433 * not running a lower priority task.
1434 */
1435 if (target != -1 &&
1436 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1437 cpu = target;
1438 }
1439 rcu_read_unlock();
1440
1441out:
1442 return cpu;
1443}
1444
1445static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1446{
1447 /*
1448 * Current can't be migrated, useless to reschedule,
1449 * let's hope p can move out.
1450 */
1451 if (rq->curr->nr_cpus_allowed == 1 ||
1452 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1453 return;
1454
1455 /*
1456 * p is migratable, so let's not schedule it and
1457 * see if it is pushed or pulled somewhere else.
1458 */
1459 if (p->nr_cpus_allowed != 1
1460 && cpupri_find(&rq->rd->cpupri, p, NULL))
1461 return;
1462
1463 /*
1464 * There appears to be other cpus that can accept
1465 * current and none to run 'p', so lets reschedule
1466 * to try and push current away:
1467 */
1468 requeue_task_rt(rq, p, 1);
1469 resched_curr(rq);
1470}
1471
1472#endif /* CONFIG_SMP */
1473
1474/*
1475 * Preempt the current task with a newly woken task if needed:
1476 */
1477static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1478{
1479 if (p->prio < rq->curr->prio) {
1480 resched_curr(rq);
1481 return;
1482 }
1483
1484#ifdef CONFIG_SMP
1485 /*
1486 * If:
1487 *
1488 * - the newly woken task is of equal priority to the current task
1489 * - the newly woken task is non-migratable while current is migratable
1490 * - current will be preempted on the next reschedule
1491 *
1492 * we should check to see if current can readily move to a different
1493 * cpu. If so, we will reschedule to allow the push logic to try
1494 * to move current somewhere else, making room for our non-migratable
1495 * task.
1496 */
1497 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1498 check_preempt_equal_prio(rq, p);
1499#endif
1500}
1501
1502static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1503 struct rt_rq *rt_rq)
1504{
1505 struct rt_prio_array *array = &rt_rq->active;
1506 struct sched_rt_entity *next = NULL;
1507 struct list_head *queue;
1508 int idx;
1509
1510 idx = sched_find_first_bit(array->bitmap);
1511 BUG_ON(idx >= MAX_RT_PRIO);
1512
1513 queue = array->queue + idx;
1514 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1515
1516 return next;
1517}
1518
1519static struct task_struct *_pick_next_task_rt(struct rq *rq)
1520{
1521 struct sched_rt_entity *rt_se;
1522 struct task_struct *p;
1523 struct rt_rq *rt_rq = &rq->rt;
1524
1525 do {
1526 rt_se = pick_next_rt_entity(rq, rt_rq);
1527 BUG_ON(!rt_se);
1528 rt_rq = group_rt_rq(rt_se);
1529 } while (rt_rq);
1530
1531 p = rt_task_of(rt_se);
1532 p->se.exec_start = rq_clock_task(rq);
1533
1534 return p;
1535}
1536
1537static struct task_struct *
1538pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1539{
1540 struct task_struct *p;
1541 struct rt_rq *rt_rq = &rq->rt;
1542
1543 if (need_pull_rt_task(rq, prev)) {
1544 /*
1545 * This is OK, because current is on_cpu, which avoids it being
1546 * picked for load-balance and preemption/IRQs are still
1547 * disabled avoiding further scheduler activity on it and we're
1548 * being very careful to re-start the picking loop.
1549 */
1550 rq_unpin_lock(rq, rf);
1551 pull_rt_task(rq);
1552 rq_repin_lock(rq, rf);
1553 /*
1554 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1555 * means a dl or stop task can slip in, in which case we need
1556 * to re-start task selection.
1557 */
1558 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1559 rq->dl.dl_nr_running))
1560 return RETRY_TASK;
1561 }
1562
1563 /*
1564 * We may dequeue prev's rt_rq in put_prev_task().
1565 * So, we update time before rt_nr_running check.
1566 */
1567 if (prev->sched_class == &rt_sched_class)
1568 update_curr_rt(rq);
1569
1570 if (!rt_rq->rt_queued)
1571 return NULL;
1572
1573 put_prev_task(rq, prev);
1574
1575 p = _pick_next_task_rt(rq);
1576
1577 /* The running task is never eligible for pushing */
1578 dequeue_pushable_task(rq, p);
1579
1580 queue_push_tasks(rq);
1581
1582 return p;
1583}
1584
1585static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1586{
1587 update_curr_rt(rq);
1588
1589 /*
1590 * The previous task needs to be made eligible for pushing
1591 * if it is still active
1592 */
1593 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1594 enqueue_pushable_task(rq, p);
1595}
1596
1597#ifdef CONFIG_SMP
1598
1599/* Only try algorithms three times */
1600#define RT_MAX_TRIES 3
1601
1602static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1603{
1604 if (!task_running(rq, p) &&
1605 cpumask_test_cpu(cpu, &p->cpus_allowed))
1606 return 1;
1607 return 0;
1608}
1609
1610/*
1611 * Return the highest pushable rq's task, which is suitable to be executed
1612 * on the cpu, NULL otherwise
1613 */
1614static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1615{
1616 struct plist_head *head = &rq->rt.pushable_tasks;
1617 struct task_struct *p;
1618
1619 if (!has_pushable_tasks(rq))
1620 return NULL;
1621
1622 plist_for_each_entry(p, head, pushable_tasks) {
1623 if (pick_rt_task(rq, p, cpu))
1624 return p;
1625 }
1626
1627 return NULL;
1628}
1629
1630static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1631
1632static int find_lowest_rq(struct task_struct *task)
1633{
1634 struct sched_domain *sd;
1635 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1636 int this_cpu = smp_processor_id();
1637 int cpu = task_cpu(task);
1638
1639 /* Make sure the mask is initialized first */
1640 if (unlikely(!lowest_mask))
1641 return -1;
1642
1643 if (task->nr_cpus_allowed == 1)
1644 return -1; /* No other targets possible */
1645
1646 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1647 return -1; /* No targets found */
1648
1649 /*
1650 * At this point we have built a mask of cpus representing the
1651 * lowest priority tasks in the system. Now we want to elect
1652 * the best one based on our affinity and topology.
1653 *
1654 * We prioritize the last cpu that the task executed on since
1655 * it is most likely cache-hot in that location.
1656 */
1657 if (cpumask_test_cpu(cpu, lowest_mask))
1658 return cpu;
1659
1660 /*
1661 * Otherwise, we consult the sched_domains span maps to figure
1662 * out which cpu is logically closest to our hot cache data.
1663 */
1664 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1665 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1666
1667 rcu_read_lock();
1668 for_each_domain(cpu, sd) {
1669 if (sd->flags & SD_WAKE_AFFINE) {
1670 int best_cpu;
1671
1672 /*
1673 * "this_cpu" is cheaper to preempt than a
1674 * remote processor.
1675 */
1676 if (this_cpu != -1 &&
1677 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1678 rcu_read_unlock();
1679 return this_cpu;
1680 }
1681
1682 best_cpu = cpumask_first_and(lowest_mask,
1683 sched_domain_span(sd));
1684 if (best_cpu < nr_cpu_ids) {
1685 rcu_read_unlock();
1686 return best_cpu;
1687 }
1688 }
1689 }
1690 rcu_read_unlock();
1691
1692 /*
1693 * And finally, if there were no matches within the domains
1694 * just give the caller *something* to work with from the compatible
1695 * locations.
1696 */
1697 if (this_cpu != -1)
1698 return this_cpu;
1699
1700 cpu = cpumask_any(lowest_mask);
1701 if (cpu < nr_cpu_ids)
1702 return cpu;
1703 return -1;
1704}
1705
1706/* Will lock the rq it finds */
1707static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1708{
1709 struct rq *lowest_rq = NULL;
1710 int tries;
1711 int cpu;
1712
1713 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1714 cpu = find_lowest_rq(task);
1715
1716 if ((cpu == -1) || (cpu == rq->cpu))
1717 break;
1718
1719 lowest_rq = cpu_rq(cpu);
1720
1721 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1722 /*
1723 * Target rq has tasks of equal or higher priority,
1724 * retrying does not release any lock and is unlikely
1725 * to yield a different result.
1726 */
1727 lowest_rq = NULL;
1728 break;
1729 }
1730
1731 /* if the prio of this runqueue changed, try again */
1732 if (double_lock_balance(rq, lowest_rq)) {
1733 /*
1734 * We had to unlock the run queue. In
1735 * the mean time, task could have
1736 * migrated already or had its affinity changed.
1737 * Also make sure that it wasn't scheduled on its rq.
1738 */
1739 if (unlikely(task_rq(task) != rq ||
1740 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) ||
1741 task_running(rq, task) ||
1742 !rt_task(task) ||
1743 !task_on_rq_queued(task))) {
1744
1745 double_unlock_balance(rq, lowest_rq);
1746 lowest_rq = NULL;
1747 break;
1748 }
1749 }
1750
1751 /* If this rq is still suitable use it. */
1752 if (lowest_rq->rt.highest_prio.curr > task->prio)
1753 break;
1754
1755 /* try again */
1756 double_unlock_balance(rq, lowest_rq);
1757 lowest_rq = NULL;
1758 }
1759
1760 return lowest_rq;
1761}
1762
1763static struct task_struct *pick_next_pushable_task(struct rq *rq)
1764{
1765 struct task_struct *p;
1766
1767 if (!has_pushable_tasks(rq))
1768 return NULL;
1769
1770 p = plist_first_entry(&rq->rt.pushable_tasks,
1771 struct task_struct, pushable_tasks);
1772
1773 BUG_ON(rq->cpu != task_cpu(p));
1774 BUG_ON(task_current(rq, p));
1775 BUG_ON(p->nr_cpus_allowed <= 1);
1776
1777 BUG_ON(!task_on_rq_queued(p));
1778 BUG_ON(!rt_task(p));
1779
1780 return p;
1781}
1782
1783/*
1784 * If the current CPU has more than one RT task, see if the non
1785 * running task can migrate over to a CPU that is running a task
1786 * of lesser priority.
1787 */
1788static int push_rt_task(struct rq *rq)
1789{
1790 struct task_struct *next_task;
1791 struct rq *lowest_rq;
1792 int ret = 0;
1793
1794 if (!rq->rt.overloaded)
1795 return 0;
1796
1797 next_task = pick_next_pushable_task(rq);
1798 if (!next_task)
1799 return 0;
1800
1801retry:
1802 if (unlikely(next_task == rq->curr)) {
1803 WARN_ON(1);
1804 return 0;
1805 }
1806
1807 /*
1808 * It's possible that the next_task slipped in of
1809 * higher priority than current. If that's the case
1810 * just reschedule current.
1811 */
1812 if (unlikely(next_task->prio < rq->curr->prio)) {
1813 resched_curr(rq);
1814 return 0;
1815 }
1816
1817 /* We might release rq lock */
1818 get_task_struct(next_task);
1819
1820 /* find_lock_lowest_rq locks the rq if found */
1821 lowest_rq = find_lock_lowest_rq(next_task, rq);
1822 if (!lowest_rq) {
1823 struct task_struct *task;
1824 /*
1825 * find_lock_lowest_rq releases rq->lock
1826 * so it is possible that next_task has migrated.
1827 *
1828 * We need to make sure that the task is still on the same
1829 * run-queue and is also still the next task eligible for
1830 * pushing.
1831 */
1832 task = pick_next_pushable_task(rq);
1833 if (task == next_task) {
1834 /*
1835 * The task hasn't migrated, and is still the next
1836 * eligible task, but we failed to find a run-queue
1837 * to push it to. Do not retry in this case, since
1838 * other cpus will pull from us when ready.
1839 */
1840 goto out;
1841 }
1842
1843 if (!task)
1844 /* No more tasks, just exit */
1845 goto out;
1846
1847 /*
1848 * Something has shifted, try again.
1849 */
1850 put_task_struct(next_task);
1851 next_task = task;
1852 goto retry;
1853 }
1854
1855 deactivate_task(rq, next_task, 0);
1856 set_task_cpu(next_task, lowest_rq->cpu);
1857 activate_task(lowest_rq, next_task, 0);
1858 ret = 1;
1859
1860 resched_curr(lowest_rq);
1861
1862 double_unlock_balance(rq, lowest_rq);
1863
1864out:
1865 put_task_struct(next_task);
1866
1867 return ret;
1868}
1869
1870static void push_rt_tasks(struct rq *rq)
1871{
1872 /* push_rt_task will return true if it moved an RT */
1873 while (push_rt_task(rq))
1874 ;
1875}
1876
1877#ifdef HAVE_RT_PUSH_IPI
1878/*
1879 * The search for the next cpu always starts at rq->cpu and ends
1880 * when we reach rq->cpu again. It will never return rq->cpu.
1881 * This returns the next cpu to check, or nr_cpu_ids if the loop
1882 * is complete.
1883 *
1884 * rq->rt.push_cpu holds the last cpu returned by this function,
1885 * or if this is the first instance, it must hold rq->cpu.
1886 */
1887static int rto_next_cpu(struct rq *rq)
1888{
1889 int prev_cpu = rq->rt.push_cpu;
1890 int cpu;
1891
1892 cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1893
1894 /*
1895 * If the previous cpu is less than the rq's CPU, then it already
1896 * passed the end of the mask, and has started from the beginning.
1897 * We end if the next CPU is greater or equal to rq's CPU.
1898 */
1899 if (prev_cpu < rq->cpu) {
1900 if (cpu >= rq->cpu)
1901 return nr_cpu_ids;
1902
1903 } else if (cpu >= nr_cpu_ids) {
1904 /*
1905 * We passed the end of the mask, start at the beginning.
1906 * If the result is greater or equal to the rq's CPU, then
1907 * the loop is finished.
1908 */
1909 cpu = cpumask_first(rq->rd->rto_mask);
1910 if (cpu >= rq->cpu)
1911 return nr_cpu_ids;
1912 }
1913 rq->rt.push_cpu = cpu;
1914
1915 /* Return cpu to let the caller know if the loop is finished or not */
1916 return cpu;
1917}
1918
1919static int find_next_push_cpu(struct rq *rq)
1920{
1921 struct rq *next_rq;
1922 int cpu;
1923
1924 while (1) {
1925 cpu = rto_next_cpu(rq);
1926 if (cpu >= nr_cpu_ids)
1927 break;
1928 next_rq = cpu_rq(cpu);
1929
1930 /* Make sure the next rq can push to this rq */
1931 if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1932 break;
1933 }
1934
1935 return cpu;
1936}
1937
1938#define RT_PUSH_IPI_EXECUTING 1
1939#define RT_PUSH_IPI_RESTART 2
1940
1941/*
1942 * When a high priority task schedules out from a CPU and a lower priority
1943 * task is scheduled in, a check is made to see if there's any RT tasks
1944 * on other CPUs that are waiting to run because a higher priority RT task
1945 * is currently running on its CPU. In this case, the CPU with multiple RT
1946 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1947 * up that may be able to run one of its non-running queued RT tasks.
1948 *
1949 * On large CPU boxes, there's the case that several CPUs could schedule
1950 * a lower priority task at the same time, in which case it will look for
1951 * any overloaded CPUs that it could pull a task from. To do this, the runqueue
1952 * lock must be taken from that overloaded CPU. Having 10s of CPUs all fighting
1953 * for a single overloaded CPU's runqueue lock can produce a large latency.
1954 * (This has actually been observed on large boxes running cyclictest).
1955 * Instead of taking the runqueue lock of the overloaded CPU, each of the
1956 * CPUs that scheduled a lower priority task simply sends an IPI to the
1957 * overloaded CPU. An IPI is much cheaper than taking an runqueue lock with
1958 * lots of contention. The overloaded CPU will look to push its non-running
1959 * RT task off, and if it does, it can then ignore the other IPIs coming
1960 * in, and just pass those IPIs off to any other overloaded CPU.
1961 *
1962 * When a CPU schedules a lower priority task, it only sends an IPI to
1963 * the "next" CPU that has overloaded RT tasks. This prevents IPI storms,
1964 * as having 10 CPUs scheduling lower priority tasks and 10 CPUs with
1965 * RT overloaded tasks, would cause 100 IPIs to go out at once.
1966 *
1967 * The overloaded RT CPU, when receiving an IPI, will try to push off its
1968 * overloaded RT tasks and then send an IPI to the next CPU that has
1969 * overloaded RT tasks. This stops when all CPUs with overloaded RT tasks
1970 * have completed. Just because a CPU may have pushed off its own overloaded
1971 * RT task does not mean it should stop sending the IPI around to other
1972 * overloaded CPUs. There may be another RT task waiting to run on one of
1973 * those CPUs that are of higher priority than the one that was just
1974 * pushed.
1975 *
1976 * An optimization that could possibly be made is to make a CPU array similar
1977 * to the cpupri array mask of all running RT tasks, but for the overloaded
1978 * case, then the IPI could be sent to only the CPU with the highest priority
1979 * RT task waiting, and that CPU could send off further IPIs to the CPU with
1980 * the next highest waiting task. Since the overloaded case is much less likely
1981 * to happen, the complexity of this implementation may not be worth it.
1982 * Instead, just send an IPI around to all overloaded CPUs.
1983 *
1984 * The rq->rt.push_flags holds the status of the IPI that is going around.
1985 * A run queue can only send out a single IPI at a time. The possible flags
1986 * for rq->rt.push_flags are:
1987 *
1988 * (None or zero): No IPI is going around for the current rq
1989 * RT_PUSH_IPI_EXECUTING: An IPI for the rq is being passed around
1990 * RT_PUSH_IPI_RESTART: The priority of the running task for the rq
1991 * has changed, and the IPI should restart
1992 * circulating the overloaded CPUs again.
1993 *
1994 * rq->rt.push_cpu contains the CPU that is being sent the IPI. It is updated
1995 * before sending to the next CPU.
1996 *
1997 * Instead of having all CPUs that schedule a lower priority task send
1998 * an IPI to the same "first" CPU in the RT overload mask, they send it
1999 * to the next overloaded CPU after their own CPU. This helps distribute
2000 * the work when there's more than one overloaded CPU and multiple CPUs
2001 * scheduling in lower priority tasks.
2002 *
2003 * When a rq schedules a lower priority task than what was currently
2004 * running, the next CPU with overloaded RT tasks is examined first.
2005 * That is, if CPU 1 and 5 are overloaded, and CPU 3 schedules a lower
2006 * priority task, it will send an IPI first to CPU 5, then CPU 5 will
2007 * send to CPU 1 if it is still overloaded. CPU 1 will clear the
2008 * rq->rt.push_flags if RT_PUSH_IPI_RESTART is not set.
2009 *
2010 * The first CPU to notice IPI_RESTART is set, will clear that flag and then
2011 * send an IPI to the next overloaded CPU after the rq->cpu and not the next
2012 * CPU after push_cpu. That is, if CPU 1, 4 and 5 are overloaded when CPU 3
2013 * schedules a lower priority task, and the IPI_RESTART gets set while the
2014 * handling is being done on CPU 5, it will clear the flag and send it back to
2015 * CPU 4 instead of CPU 1.
2016 *
2017 * Note, the above logic can be disabled by turning off the sched_feature
2018 * RT_PUSH_IPI. Then the rq lock of the overloaded CPU will simply be
2019 * taken by the CPU requesting a pull and the waiting RT task will be pulled
2020 * by that CPU. This may be fine for machines with few CPUs.
2021 */
2022static void tell_cpu_to_push(struct rq *rq)
2023{
2024 int cpu;
2025
2026 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
2027 raw_spin_lock(&rq->rt.push_lock);
2028 /* Make sure it's still executing */
2029 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
2030 /*
2031 * Tell the IPI to restart the loop as things have
2032 * changed since it started.
2033 */
2034 rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
2035 raw_spin_unlock(&rq->rt.push_lock);
2036 return;
2037 }
2038 raw_spin_unlock(&rq->rt.push_lock);
2039 }
2040
2041 /* When here, there's no IPI going around */
2042
2043 rq->rt.push_cpu = rq->cpu;
2044 cpu = find_next_push_cpu(rq);
2045 if (cpu >= nr_cpu_ids)
2046 return;
2047
2048 rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
2049
2050 irq_work_queue_on(&rq->rt.push_work, cpu);
2051}
2052
2053/* Called from hardirq context */
2054static void try_to_push_tasks(void *arg)
2055{
2056 struct rt_rq *rt_rq = arg;
2057 struct rq *rq, *src_rq;
2058 int this_cpu;
2059 int cpu;
2060
2061 this_cpu = rt_rq->push_cpu;
2062
2063 /* Paranoid check */
2064 BUG_ON(this_cpu != smp_processor_id());
2065
2066 rq = cpu_rq(this_cpu);
2067 src_rq = rq_of_rt_rq(rt_rq);
2068
2069again:
2070 if (has_pushable_tasks(rq)) {
2071 raw_spin_lock(&rq->lock);
2072 push_rt_task(rq);
2073 raw_spin_unlock(&rq->lock);
2074 }
2075
2076 /* Pass the IPI to the next rt overloaded queue */
2077 raw_spin_lock(&rt_rq->push_lock);
2078 /*
2079 * If the source queue changed since the IPI went out,
2080 * we need to restart the search from that CPU again.
2081 */
2082 if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
2083 rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
2084 rt_rq->push_cpu = src_rq->cpu;
2085 }
2086
2087 cpu = find_next_push_cpu(src_rq);
2088
2089 if (cpu >= nr_cpu_ids)
2090 rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
2091 raw_spin_unlock(&rt_rq->push_lock);
2092
2093 if (cpu >= nr_cpu_ids)
2094 return;
2095
2096 /*
2097 * It is possible that a restart caused this CPU to be
2098 * chosen again. Don't bother with an IPI, just see if we
2099 * have more to push.
2100 */
2101 if (unlikely(cpu == rq->cpu))
2102 goto again;
2103
2104 /* Try the next RT overloaded CPU */
2105 irq_work_queue_on(&rt_rq->push_work, cpu);
2106}
2107
2108static void push_irq_work_func(struct irq_work *work)
2109{
2110 struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
2111
2112 try_to_push_tasks(rt_rq);
2113}
2114#endif /* HAVE_RT_PUSH_IPI */
2115
2116static void pull_rt_task(struct rq *this_rq)
2117{
2118 int this_cpu = this_rq->cpu, cpu;
2119 bool resched = false;
2120 struct task_struct *p;
2121 struct rq *src_rq;
2122
2123 if (likely(!rt_overloaded(this_rq)))
2124 return;
2125
2126 /*
2127 * Match the barrier from rt_set_overloaded; this guarantees that if we
2128 * see overloaded we must also see the rto_mask bit.
2129 */
2130 smp_rmb();
2131
2132#ifdef HAVE_RT_PUSH_IPI
2133 if (sched_feat(RT_PUSH_IPI)) {
2134 tell_cpu_to_push(this_rq);
2135 return;
2136 }
2137#endif
2138
2139 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2140 if (this_cpu == cpu)
2141 continue;
2142
2143 src_rq = cpu_rq(cpu);
2144
2145 /*
2146 * Don't bother taking the src_rq->lock if the next highest
2147 * task is known to be lower-priority than our current task.
2148 * This may look racy, but if this value is about to go
2149 * logically higher, the src_rq will push this task away.
2150 * And if its going logically lower, we do not care
2151 */
2152 if (src_rq->rt.highest_prio.next >=
2153 this_rq->rt.highest_prio.curr)
2154 continue;
2155
2156 /*
2157 * We can potentially drop this_rq's lock in
2158 * double_lock_balance, and another CPU could
2159 * alter this_rq
2160 */
2161 double_lock_balance(this_rq, src_rq);
2162
2163 /*
2164 * We can pull only a task, which is pushable
2165 * on its rq, and no others.
2166 */
2167 p = pick_highest_pushable_task(src_rq, this_cpu);
2168
2169 /*
2170 * Do we have an RT task that preempts
2171 * the to-be-scheduled task?
2172 */
2173 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2174 WARN_ON(p == src_rq->curr);
2175 WARN_ON(!task_on_rq_queued(p));
2176
2177 /*
2178 * There's a chance that p is higher in priority
2179 * than what's currently running on its cpu.
2180 * This is just that p is wakeing up and hasn't
2181 * had a chance to schedule. We only pull
2182 * p if it is lower in priority than the
2183 * current task on the run queue
2184 */
2185 if (p->prio < src_rq->curr->prio)
2186 goto skip;
2187
2188 resched = true;
2189
2190 deactivate_task(src_rq, p, 0);
2191 set_task_cpu(p, this_cpu);
2192 activate_task(this_rq, p, 0);
2193 /*
2194 * We continue with the search, just in
2195 * case there's an even higher prio task
2196 * in another runqueue. (low likelihood
2197 * but possible)
2198 */
2199 }
2200skip:
2201 double_unlock_balance(this_rq, src_rq);
2202 }
2203
2204 if (resched)
2205 resched_curr(this_rq);
2206}
2207
2208/*
2209 * If we are not running and we are not going to reschedule soon, we should
2210 * try to push tasks away now
2211 */
2212static void task_woken_rt(struct rq *rq, struct task_struct *p)
2213{
2214 if (!task_running(rq, p) &&
2215 !test_tsk_need_resched(rq->curr) &&
2216 p->nr_cpus_allowed > 1 &&
2217 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2218 (rq->curr->nr_cpus_allowed < 2 ||
2219 rq->curr->prio <= p->prio))
2220 push_rt_tasks(rq);
2221}
2222
2223/* Assumes rq->lock is held */
2224static void rq_online_rt(struct rq *rq)
2225{
2226 if (rq->rt.overloaded)
2227 rt_set_overload(rq);
2228
2229 __enable_runtime(rq);
2230
2231 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2232}
2233
2234/* Assumes rq->lock is held */
2235static void rq_offline_rt(struct rq *rq)
2236{
2237 if (rq->rt.overloaded)
2238 rt_clear_overload(rq);
2239
2240 __disable_runtime(rq);
2241
2242 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2243}
2244
2245/*
2246 * When switch from the rt queue, we bring ourselves to a position
2247 * that we might want to pull RT tasks from other runqueues.
2248 */
2249static void switched_from_rt(struct rq *rq, struct task_struct *p)
2250{
2251 /*
2252 * If there are other RT tasks then we will reschedule
2253 * and the scheduling of the other RT tasks will handle
2254 * the balancing. But if we are the last RT task
2255 * we may need to handle the pulling of RT tasks
2256 * now.
2257 */
2258 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2259 return;
2260
2261 queue_pull_task(rq);
2262}
2263
2264void __init init_sched_rt_class(void)
2265{
2266 unsigned int i;
2267
2268 for_each_possible_cpu(i) {
2269 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2270 GFP_KERNEL, cpu_to_node(i));
2271 }
2272}
2273#endif /* CONFIG_SMP */
2274
2275/*
2276 * When switching a task to RT, we may overload the runqueue
2277 * with RT tasks. In this case we try to push them off to
2278 * other runqueues.
2279 */
2280static void switched_to_rt(struct rq *rq, struct task_struct *p)
2281{
2282 /*
2283 * If we are already running, then there's nothing
2284 * that needs to be done. But if we are not running
2285 * we may need to preempt the current running task.
2286 * If that current running task is also an RT task
2287 * then see if we can move to another run queue.
2288 */
2289 if (task_on_rq_queued(p) && rq->curr != p) {
2290#ifdef CONFIG_SMP
2291 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2292 queue_push_tasks(rq);
2293#endif /* CONFIG_SMP */
2294 if (p->prio < rq->curr->prio)
2295 resched_curr(rq);
2296 }
2297}
2298
2299/*
2300 * Priority of the task has changed. This may cause
2301 * us to initiate a push or pull.
2302 */
2303static void
2304prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2305{
2306 if (!task_on_rq_queued(p))
2307 return;
2308
2309 if (rq->curr == p) {
2310#ifdef CONFIG_SMP
2311 /*
2312 * If our priority decreases while running, we
2313 * may need to pull tasks to this runqueue.
2314 */
2315 if (oldprio < p->prio)
2316 queue_pull_task(rq);
2317
2318 /*
2319 * If there's a higher priority task waiting to run
2320 * then reschedule.
2321 */
2322 if (p->prio > rq->rt.highest_prio.curr)
2323 resched_curr(rq);
2324#else
2325 /* For UP simply resched on drop of prio */
2326 if (oldprio < p->prio)
2327 resched_curr(rq);
2328#endif /* CONFIG_SMP */
2329 } else {
2330 /*
2331 * This task is not running, but if it is
2332 * greater than the current running task
2333 * then reschedule.
2334 */
2335 if (p->prio < rq->curr->prio)
2336 resched_curr(rq);
2337 }
2338}
2339
2340#ifdef CONFIG_POSIX_TIMERS
2341static void watchdog(struct rq *rq, struct task_struct *p)
2342{
2343 unsigned long soft, hard;
2344
2345 /* max may change after cur was read, this will be fixed next tick */
2346 soft = task_rlimit(p, RLIMIT_RTTIME);
2347 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2348
2349 if (soft != RLIM_INFINITY) {
2350 unsigned long next;
2351
2352 if (p->rt.watchdog_stamp != jiffies) {
2353 p->rt.timeout++;
2354 p->rt.watchdog_stamp = jiffies;
2355 }
2356
2357 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2358 if (p->rt.timeout > next)
2359 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2360 }
2361}
2362#else
2363static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2364#endif
2365
2366static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2367{
2368 struct sched_rt_entity *rt_se = &p->rt;
2369
2370 update_curr_rt(rq);
2371
2372 watchdog(rq, p);
2373
2374 /*
2375 * RR tasks need a special form of timeslice management.
2376 * FIFO tasks have no timeslices.
2377 */
2378 if (p->policy != SCHED_RR)
2379 return;
2380
2381 if (--p->rt.time_slice)
2382 return;
2383
2384 p->rt.time_slice = sched_rr_timeslice;
2385
2386 /*
2387 * Requeue to the end of queue if we (and all of our ancestors) are not
2388 * the only element on the queue
2389 */
2390 for_each_sched_rt_entity(rt_se) {
2391 if (rt_se->run_list.prev != rt_se->run_list.next) {
2392 requeue_task_rt(rq, p, 0);
2393 resched_curr(rq);
2394 return;
2395 }
2396 }
2397}
2398
2399static void set_curr_task_rt(struct rq *rq)
2400{
2401 struct task_struct *p = rq->curr;
2402
2403 p->se.exec_start = rq_clock_task(rq);
2404
2405 /* The running task is never eligible for pushing */
2406 dequeue_pushable_task(rq, p);
2407}
2408
2409static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2410{
2411 /*
2412 * Time slice is 0 for SCHED_FIFO tasks
2413 */
2414 if (task->policy == SCHED_RR)
2415 return sched_rr_timeslice;
2416 else
2417 return 0;
2418}
2419
2420const struct sched_class rt_sched_class = {
2421 .next = &fair_sched_class,
2422 .enqueue_task = enqueue_task_rt,
2423 .dequeue_task = dequeue_task_rt,
2424 .yield_task = yield_task_rt,
2425
2426 .check_preempt_curr = check_preempt_curr_rt,
2427
2428 .pick_next_task = pick_next_task_rt,
2429 .put_prev_task = put_prev_task_rt,
2430
2431#ifdef CONFIG_SMP
2432 .select_task_rq = select_task_rq_rt,
2433
2434 .set_cpus_allowed = set_cpus_allowed_common,
2435 .rq_online = rq_online_rt,
2436 .rq_offline = rq_offline_rt,
2437 .task_woken = task_woken_rt,
2438 .switched_from = switched_from_rt,
2439#endif
2440
2441 .set_curr_task = set_curr_task_rt,
2442 .task_tick = task_tick_rt,
2443
2444 .get_rr_interval = get_rr_interval_rt,
2445
2446 .prio_changed = prio_changed_rt,
2447 .switched_to = switched_to_rt,
2448
2449 .update_curr = update_curr_rt,
2450};
2451
2452#ifdef CONFIG_RT_GROUP_SCHED
2453/*
2454 * Ensure that the real time constraints are schedulable.
2455 */
2456static DEFINE_MUTEX(rt_constraints_mutex);
2457
2458/* Must be called with tasklist_lock held */
2459static inline int tg_has_rt_tasks(struct task_group *tg)
2460{
2461 struct task_struct *g, *p;
2462
2463 /*
2464 * Autogroups do not have RT tasks; see autogroup_create().
2465 */
2466 if (task_group_is_autogroup(tg))
2467 return 0;
2468
2469 for_each_process_thread(g, p) {
2470 if (rt_task(p) && task_group(p) == tg)
2471 return 1;
2472 }
2473
2474 return 0;
2475}
2476
2477struct rt_schedulable_data {
2478 struct task_group *tg;
2479 u64 rt_period;
2480 u64 rt_runtime;
2481};
2482
2483static int tg_rt_schedulable(struct task_group *tg, void *data)
2484{
2485 struct rt_schedulable_data *d = data;
2486 struct task_group *child;
2487 unsigned long total, sum = 0;
2488 u64 period, runtime;
2489
2490 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2491 runtime = tg->rt_bandwidth.rt_runtime;
2492
2493 if (tg == d->tg) {
2494 period = d->rt_period;
2495 runtime = d->rt_runtime;
2496 }
2497
2498 /*
2499 * Cannot have more runtime than the period.
2500 */
2501 if (runtime > period && runtime != RUNTIME_INF)
2502 return -EINVAL;
2503
2504 /*
2505 * Ensure we don't starve existing RT tasks.
2506 */
2507 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
2508 return -EBUSY;
2509
2510 total = to_ratio(period, runtime);
2511
2512 /*
2513 * Nobody can have more than the global setting allows.
2514 */
2515 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
2516 return -EINVAL;
2517
2518 /*
2519 * The sum of our children's runtime should not exceed our own.
2520 */
2521 list_for_each_entry_rcu(child, &tg->children, siblings) {
2522 period = ktime_to_ns(child->rt_bandwidth.rt_period);
2523 runtime = child->rt_bandwidth.rt_runtime;
2524
2525 if (child == d->tg) {
2526 period = d->rt_period;
2527 runtime = d->rt_runtime;
2528 }
2529
2530 sum += to_ratio(period, runtime);
2531 }
2532
2533 if (sum > total)
2534 return -EINVAL;
2535
2536 return 0;
2537}
2538
2539static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
2540{
2541 int ret;
2542
2543 struct rt_schedulable_data data = {
2544 .tg = tg,
2545 .rt_period = period,
2546 .rt_runtime = runtime,
2547 };
2548
2549 rcu_read_lock();
2550 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
2551 rcu_read_unlock();
2552
2553 return ret;
2554}
2555
2556static int tg_set_rt_bandwidth(struct task_group *tg,
2557 u64 rt_period, u64 rt_runtime)
2558{
2559 int i, err = 0;
2560
2561 /*
2562 * Disallowing the root group RT runtime is BAD, it would disallow the
2563 * kernel creating (and or operating) RT threads.
2564 */
2565 if (tg == &root_task_group && rt_runtime == 0)
2566 return -EINVAL;
2567
2568 /* No period doesn't make any sense. */
2569 if (rt_period == 0)
2570 return -EINVAL;
2571
2572 mutex_lock(&rt_constraints_mutex);
2573 read_lock(&tasklist_lock);
2574 err = __rt_schedulable(tg, rt_period, rt_runtime);
2575 if (err)
2576 goto unlock;
2577
2578 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2579 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
2580 tg->rt_bandwidth.rt_runtime = rt_runtime;
2581
2582 for_each_possible_cpu(i) {
2583 struct rt_rq *rt_rq = tg->rt_rq[i];
2584
2585 raw_spin_lock(&rt_rq->rt_runtime_lock);
2586 rt_rq->rt_runtime = rt_runtime;
2587 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2588 }
2589 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
2590unlock:
2591 read_unlock(&tasklist_lock);
2592 mutex_unlock(&rt_constraints_mutex);
2593
2594 return err;
2595}
2596
2597int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
2598{
2599 u64 rt_runtime, rt_period;
2600
2601 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
2602 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
2603 if (rt_runtime_us < 0)
2604 rt_runtime = RUNTIME_INF;
2605
2606 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2607}
2608
2609long sched_group_rt_runtime(struct task_group *tg)
2610{
2611 u64 rt_runtime_us;
2612
2613 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
2614 return -1;
2615
2616 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
2617 do_div(rt_runtime_us, NSEC_PER_USEC);
2618 return rt_runtime_us;
2619}
2620
2621int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
2622{
2623 u64 rt_runtime, rt_period;
2624
2625 rt_period = rt_period_us * NSEC_PER_USEC;
2626 rt_runtime = tg->rt_bandwidth.rt_runtime;
2627
2628 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
2629}
2630
2631long sched_group_rt_period(struct task_group *tg)
2632{
2633 u64 rt_period_us;
2634
2635 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
2636 do_div(rt_period_us, NSEC_PER_USEC);
2637 return rt_period_us;
2638}
2639
2640static int sched_rt_global_constraints(void)
2641{
2642 int ret = 0;
2643
2644 mutex_lock(&rt_constraints_mutex);
2645 read_lock(&tasklist_lock);
2646 ret = __rt_schedulable(NULL, 0, 0);
2647 read_unlock(&tasklist_lock);
2648 mutex_unlock(&rt_constraints_mutex);
2649
2650 return ret;
2651}
2652
2653int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
2654{
2655 /* Don't accept realtime tasks when there is no way for them to run */
2656 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
2657 return 0;
2658
2659 return 1;
2660}
2661
2662#else /* !CONFIG_RT_GROUP_SCHED */
2663static int sched_rt_global_constraints(void)
2664{
2665 unsigned long flags;
2666 int i;
2667
2668 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
2669 for_each_possible_cpu(i) {
2670 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
2671
2672 raw_spin_lock(&rt_rq->rt_runtime_lock);
2673 rt_rq->rt_runtime = global_rt_runtime();
2674 raw_spin_unlock(&rt_rq->rt_runtime_lock);
2675 }
2676 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
2677
2678 return 0;
2679}
2680#endif /* CONFIG_RT_GROUP_SCHED */
2681
2682static int sched_rt_global_validate(void)
2683{
2684 if (sysctl_sched_rt_period <= 0)
2685 return -EINVAL;
2686
2687 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
2688 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
2689 return -EINVAL;
2690
2691 return 0;
2692}
2693
2694static void sched_rt_do_global(void)
2695{
2696 def_rt_bandwidth.rt_runtime = global_rt_runtime();
2697 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
2698}
2699
2700int sched_rt_handler(struct ctl_table *table, int write,
2701 void __user *buffer, size_t *lenp,
2702 loff_t *ppos)
2703{
2704 int old_period, old_runtime;
2705 static DEFINE_MUTEX(mutex);
2706 int ret;
2707
2708 mutex_lock(&mutex);
2709 old_period = sysctl_sched_rt_period;
2710 old_runtime = sysctl_sched_rt_runtime;
2711
2712 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2713
2714 if (!ret && write) {
2715 ret = sched_rt_global_validate();
2716 if (ret)
2717 goto undo;
2718
2719 ret = sched_dl_global_validate();
2720 if (ret)
2721 goto undo;
2722
2723 ret = sched_rt_global_constraints();
2724 if (ret)
2725 goto undo;
2726
2727 sched_rt_do_global();
2728 sched_dl_do_global();
2729 }
2730 if (0) {
2731undo:
2732 sysctl_sched_rt_period = old_period;
2733 sysctl_sched_rt_runtime = old_runtime;
2734 }
2735 mutex_unlock(&mutex);
2736
2737 return ret;
2738}
2739
2740int sched_rr_handler(struct ctl_table *table, int write,
2741 void __user *buffer, size_t *lenp,
2742 loff_t *ppos)
2743{
2744 int ret;
2745 static DEFINE_MUTEX(mutex);
2746
2747 mutex_lock(&mutex);
2748 ret = proc_dointvec(table, write, buffer, lenp, ppos);
2749 /*
2750 * Make sure that internally we keep jiffies.
2751 * Also, writing zero resets the timeslice to default:
2752 */
2753 if (!ret && write) {
2754 sched_rr_timeslice =
2755 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
2756 msecs_to_jiffies(sysctl_sched_rr_timeslice);
2757 }
2758 mutex_unlock(&mutex);
2759 return ret;
2760}
2761
2762#ifdef CONFIG_SCHED_DEBUG
2763extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2764
2765void print_rt_stats(struct seq_file *m, int cpu)
2766{
2767 rt_rq_iter_t iter;
2768 struct rt_rq *rt_rq;
2769
2770 rcu_read_lock();
2771 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2772 print_rt_rq(m, cpu, rt_rq);
2773 rcu_read_unlock();
2774}
2775#endif /* CONFIG_SCHED_DEBUG */