xref: /openbmc/linux/kernel/sched/rt.c (revision f7d84fa7)
1 /*
2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3  * policies)
4  */
5 
6 #include "sched.h"
7 
8 #include <linux/slab.h>
9 #include <linux/irq_work.h>
10 
11 int sched_rr_timeslice = RR_TIMESLICE;
12 int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
13 
14 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
15 
16 struct rt_bandwidth def_rt_bandwidth;
17 
18 static 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 
42 void 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 
54 static 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)
77 static void push_irq_work_func(struct irq_work *work);
78 #endif
79 
80 void 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
117 static 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 
124 static 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 
132 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
133 {
134 	return rt_rq->rq;
135 }
136 
137 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
138 {
139 	return rt_se->rt_rq;
140 }
141 
142 static 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 
149 void 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 
167 void 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 
194 int 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 
228 err_free_rq:
229 	kfree(rt_rq);
230 err:
231 	return 0;
232 }
233 
234 #else /* CONFIG_RT_GROUP_SCHED */
235 
236 #define rt_entity_is_task(rt_se) (1)
237 
238 static 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 
243 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
244 {
245 	return container_of(rt_rq, struct rq, rt);
246 }
247 
248 static 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 
255 static 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 
262 void free_rt_sched_group(struct task_group *tg) { }
263 
264 int 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 
272 static void pull_rt_task(struct rq *this_rq);
273 
274 static 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 
280 static inline int rt_overloaded(struct rq *rq)
281 {
282 	return atomic_read(&rq->rd->rto_count);
283 }
284 
285 static 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 
304 static 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 
314 static 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 
327 static 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 
344 static 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 
361 static inline int has_pushable_tasks(struct rq *rq)
362 {
363 	return !plist_head_empty(&rq->rt.pushable_tasks);
364 }
365 
366 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
367 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
368 
369 static void push_rt_tasks(struct rq *);
370 static void pull_rt_task(struct rq *);
371 
372 static 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 
380 static 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 
385 static 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 
396 static 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 
411 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
412 {
413 }
414 
415 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
416 {
417 }
418 
419 static inline
420 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
421 {
422 }
423 
424 static inline
425 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
426 {
427 }
428 
429 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
430 {
431 	return false;
432 }
433 
434 static inline void pull_rt_task(struct rq *this_rq)
435 {
436 }
437 
438 static inline void queue_push_tasks(struct rq *rq)
439 {
440 }
441 #endif /* CONFIG_SMP */
442 
443 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
444 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
445 
446 static 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 
453 static 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 
461 static 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 
466 typedef struct task_group *rt_rq_iter_t;
467 
468 static 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 
489 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
490 {
491 	return rt_se->my_q;
492 }
493 
494 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
495 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
496 
497 static 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 
518 static 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 
531 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
532 {
533 	return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
534 }
535 
536 static 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
549 static inline const struct cpumask *sched_rt_period_mask(void)
550 {
551 	return this_rq()->rd->span;
552 }
553 #else
554 static inline const struct cpumask *sched_rt_period_mask(void)
555 {
556 	return cpu_online_mask;
557 }
558 #endif
559 
560 static inline
561 struct 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 
566 static 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 
573 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
574 {
575 	return rt_rq->rt_runtime;
576 }
577 
578 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
579 {
580 	return ktime_to_ns(def_rt_bandwidth.rt_period);
581 }
582 
583 typedef 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 
591 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
592 {
593 	return NULL;
594 }
595 
596 static 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 
607 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
608 {
609 	dequeue_top_rt_rq(rt_rq);
610 }
611 
612 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
613 {
614 	return rt_rq->rt_throttled;
615 }
616 
617 static inline const struct cpumask *sched_rt_period_mask(void)
618 {
619 	return cpu_online_mask;
620 }
621 
622 static inline
623 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
624 {
625 	return &cpu_rq(cpu)->rt;
626 }
627 
628 static 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 
635 bool 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  */
647 static 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 		}
690 next:
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  */
699 static 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);
766 balanced:
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 
781 static 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 
805 static 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 */
817 static inline void balance_runtime(struct rt_rq *rt_rq) {}
818 #endif /* CONFIG_SMP */
819 
820 static 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 
844 		raw_spin_lock(&rq->lock);
845 		if (rt_rq->rt_time) {
846 			u64 runtime;
847 
848 			raw_spin_lock(&rt_rq->rt_runtime_lock);
849 			if (rt_rq->rt_throttled)
850 				balance_runtime(rt_rq);
851 			runtime = rt_rq->rt_runtime;
852 			rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
853 			if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
854 				rt_rq->rt_throttled = 0;
855 				enqueue = 1;
856 
857 				/*
858 				 * When we're idle and a woken (rt) task is
859 				 * throttled check_preempt_curr() will set
860 				 * skip_update and the time between the wakeup
861 				 * and this unthrottle will get accounted as
862 				 * 'runtime'.
863 				 */
864 				if (rt_rq->rt_nr_running && rq->curr == rq->idle)
865 					rq_clock_skip_update(rq, false);
866 			}
867 			if (rt_rq->rt_time || rt_rq->rt_nr_running)
868 				idle = 0;
869 			raw_spin_unlock(&rt_rq->rt_runtime_lock);
870 		} else if (rt_rq->rt_nr_running) {
871 			idle = 0;
872 			if (!rt_rq_throttled(rt_rq))
873 				enqueue = 1;
874 		}
875 		if (rt_rq->rt_throttled)
876 			throttled = 1;
877 
878 		if (enqueue)
879 			sched_rt_rq_enqueue(rt_rq);
880 		raw_spin_unlock(&rq->lock);
881 	}
882 
883 	if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
884 		return 1;
885 
886 	return idle;
887 }
888 
889 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
890 {
891 #ifdef CONFIG_RT_GROUP_SCHED
892 	struct rt_rq *rt_rq = group_rt_rq(rt_se);
893 
894 	if (rt_rq)
895 		return rt_rq->highest_prio.curr;
896 #endif
897 
898 	return rt_task_of(rt_se)->prio;
899 }
900 
901 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
902 {
903 	u64 runtime = sched_rt_runtime(rt_rq);
904 
905 	if (rt_rq->rt_throttled)
906 		return rt_rq_throttled(rt_rq);
907 
908 	if (runtime >= sched_rt_period(rt_rq))
909 		return 0;
910 
911 	balance_runtime(rt_rq);
912 	runtime = sched_rt_runtime(rt_rq);
913 	if (runtime == RUNTIME_INF)
914 		return 0;
915 
916 	if (rt_rq->rt_time > runtime) {
917 		struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
918 
919 		/*
920 		 * Don't actually throttle groups that have no runtime assigned
921 		 * but accrue some time due to boosting.
922 		 */
923 		if (likely(rt_b->rt_runtime)) {
924 			rt_rq->rt_throttled = 1;
925 			printk_deferred_once("sched: RT throttling activated\n");
926 		} else {
927 			/*
928 			 * In case we did anyway, make it go away,
929 			 * replenishment is a joke, since it will replenish us
930 			 * with exactly 0 ns.
931 			 */
932 			rt_rq->rt_time = 0;
933 		}
934 
935 		if (rt_rq_throttled(rt_rq)) {
936 			sched_rt_rq_dequeue(rt_rq);
937 			return 1;
938 		}
939 	}
940 
941 	return 0;
942 }
943 
944 /*
945  * Update the current task's runtime statistics. Skip current tasks that
946  * are not in our scheduling class.
947  */
948 static void update_curr_rt(struct rq *rq)
949 {
950 	struct task_struct *curr = rq->curr;
951 	struct sched_rt_entity *rt_se = &curr->rt;
952 	u64 delta_exec;
953 
954 	if (curr->sched_class != &rt_sched_class)
955 		return;
956 
957 	delta_exec = rq_clock_task(rq) - curr->se.exec_start;
958 	if (unlikely((s64)delta_exec <= 0))
959 		return;
960 
961 	/* Kick cpufreq (see the comment in kernel/sched/sched.h). */
962 	cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_RT);
963 
964 	schedstat_set(curr->se.statistics.exec_max,
965 		      max(curr->se.statistics.exec_max, delta_exec));
966 
967 	curr->se.sum_exec_runtime += delta_exec;
968 	account_group_exec_runtime(curr, delta_exec);
969 
970 	curr->se.exec_start = rq_clock_task(rq);
971 	cpuacct_charge(curr, delta_exec);
972 
973 	sched_rt_avg_update(rq, delta_exec);
974 
975 	if (!rt_bandwidth_enabled())
976 		return;
977 
978 	for_each_sched_rt_entity(rt_se) {
979 		struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
980 
981 		if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
982 			raw_spin_lock(&rt_rq->rt_runtime_lock);
983 			rt_rq->rt_time += delta_exec;
984 			if (sched_rt_runtime_exceeded(rt_rq))
985 				resched_curr(rq);
986 			raw_spin_unlock(&rt_rq->rt_runtime_lock);
987 		}
988 	}
989 }
990 
991 static void
992 dequeue_top_rt_rq(struct rt_rq *rt_rq)
993 {
994 	struct rq *rq = rq_of_rt_rq(rt_rq);
995 
996 	BUG_ON(&rq->rt != rt_rq);
997 
998 	if (!rt_rq->rt_queued)
999 		return;
1000 
1001 	BUG_ON(!rq->nr_running);
1002 
1003 	sub_nr_running(rq, rt_rq->rt_nr_running);
1004 	rt_rq->rt_queued = 0;
1005 }
1006 
1007 static void
1008 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1009 {
1010 	struct rq *rq = rq_of_rt_rq(rt_rq);
1011 
1012 	BUG_ON(&rq->rt != rt_rq);
1013 
1014 	if (rt_rq->rt_queued)
1015 		return;
1016 	if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1017 		return;
1018 
1019 	add_nr_running(rq, rt_rq->rt_nr_running);
1020 	rt_rq->rt_queued = 1;
1021 }
1022 
1023 #if defined CONFIG_SMP
1024 
1025 static void
1026 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1027 {
1028 	struct rq *rq = rq_of_rt_rq(rt_rq);
1029 
1030 #ifdef CONFIG_RT_GROUP_SCHED
1031 	/*
1032 	 * Change rq's cpupri only if rt_rq is the top queue.
1033 	 */
1034 	if (&rq->rt != rt_rq)
1035 		return;
1036 #endif
1037 	if (rq->online && prio < prev_prio)
1038 		cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1039 }
1040 
1041 static void
1042 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1043 {
1044 	struct rq *rq = rq_of_rt_rq(rt_rq);
1045 
1046 #ifdef CONFIG_RT_GROUP_SCHED
1047 	/*
1048 	 * Change rq's cpupri only if rt_rq is the top queue.
1049 	 */
1050 	if (&rq->rt != rt_rq)
1051 		return;
1052 #endif
1053 	if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1054 		cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1055 }
1056 
1057 #else /* CONFIG_SMP */
1058 
1059 static inline
1060 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1061 static inline
1062 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1063 
1064 #endif /* CONFIG_SMP */
1065 
1066 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1067 static void
1068 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1069 {
1070 	int prev_prio = rt_rq->highest_prio.curr;
1071 
1072 	if (prio < prev_prio)
1073 		rt_rq->highest_prio.curr = prio;
1074 
1075 	inc_rt_prio_smp(rt_rq, prio, prev_prio);
1076 }
1077 
1078 static void
1079 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1080 {
1081 	int prev_prio = rt_rq->highest_prio.curr;
1082 
1083 	if (rt_rq->rt_nr_running) {
1084 
1085 		WARN_ON(prio < prev_prio);
1086 
1087 		/*
1088 		 * This may have been our highest task, and therefore
1089 		 * we may have some recomputation to do
1090 		 */
1091 		if (prio == prev_prio) {
1092 			struct rt_prio_array *array = &rt_rq->active;
1093 
1094 			rt_rq->highest_prio.curr =
1095 				sched_find_first_bit(array->bitmap);
1096 		}
1097 
1098 	} else
1099 		rt_rq->highest_prio.curr = MAX_RT_PRIO;
1100 
1101 	dec_rt_prio_smp(rt_rq, prio, prev_prio);
1102 }
1103 
1104 #else
1105 
1106 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1107 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1108 
1109 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1110 
1111 #ifdef CONFIG_RT_GROUP_SCHED
1112 
1113 static void
1114 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1115 {
1116 	if (rt_se_boosted(rt_se))
1117 		rt_rq->rt_nr_boosted++;
1118 
1119 	if (rt_rq->tg)
1120 		start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1121 }
1122 
1123 static void
1124 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1125 {
1126 	if (rt_se_boosted(rt_se))
1127 		rt_rq->rt_nr_boosted--;
1128 
1129 	WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1130 }
1131 
1132 #else /* CONFIG_RT_GROUP_SCHED */
1133 
1134 static void
1135 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1136 {
1137 	start_rt_bandwidth(&def_rt_bandwidth);
1138 }
1139 
1140 static inline
1141 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1142 
1143 #endif /* CONFIG_RT_GROUP_SCHED */
1144 
1145 static inline
1146 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1147 {
1148 	struct rt_rq *group_rq = group_rt_rq(rt_se);
1149 
1150 	if (group_rq)
1151 		return group_rq->rt_nr_running;
1152 	else
1153 		return 1;
1154 }
1155 
1156 static inline
1157 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1158 {
1159 	struct rt_rq *group_rq = group_rt_rq(rt_se);
1160 	struct task_struct *tsk;
1161 
1162 	if (group_rq)
1163 		return group_rq->rr_nr_running;
1164 
1165 	tsk = rt_task_of(rt_se);
1166 
1167 	return (tsk->policy == SCHED_RR) ? 1 : 0;
1168 }
1169 
1170 static inline
1171 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1172 {
1173 	int prio = rt_se_prio(rt_se);
1174 
1175 	WARN_ON(!rt_prio(prio));
1176 	rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1177 	rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1178 
1179 	inc_rt_prio(rt_rq, prio);
1180 	inc_rt_migration(rt_se, rt_rq);
1181 	inc_rt_group(rt_se, rt_rq);
1182 }
1183 
1184 static inline
1185 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1186 {
1187 	WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1188 	WARN_ON(!rt_rq->rt_nr_running);
1189 	rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1190 	rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1191 
1192 	dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1193 	dec_rt_migration(rt_se, rt_rq);
1194 	dec_rt_group(rt_se, rt_rq);
1195 }
1196 
1197 /*
1198  * Change rt_se->run_list location unless SAVE && !MOVE
1199  *
1200  * assumes ENQUEUE/DEQUEUE flags match
1201  */
1202 static inline bool move_entity(unsigned int flags)
1203 {
1204 	if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1205 		return false;
1206 
1207 	return true;
1208 }
1209 
1210 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1211 {
1212 	list_del_init(&rt_se->run_list);
1213 
1214 	if (list_empty(array->queue + rt_se_prio(rt_se)))
1215 		__clear_bit(rt_se_prio(rt_se), array->bitmap);
1216 
1217 	rt_se->on_list = 0;
1218 }
1219 
1220 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1221 {
1222 	struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1223 	struct rt_prio_array *array = &rt_rq->active;
1224 	struct rt_rq *group_rq = group_rt_rq(rt_se);
1225 	struct list_head *queue = array->queue + rt_se_prio(rt_se);
1226 
1227 	/*
1228 	 * Don't enqueue the group if its throttled, or when empty.
1229 	 * The latter is a consequence of the former when a child group
1230 	 * get throttled and the current group doesn't have any other
1231 	 * active members.
1232 	 */
1233 	if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1234 		if (rt_se->on_list)
1235 			__delist_rt_entity(rt_se, array);
1236 		return;
1237 	}
1238 
1239 	if (move_entity(flags)) {
1240 		WARN_ON_ONCE(rt_se->on_list);
1241 		if (flags & ENQUEUE_HEAD)
1242 			list_add(&rt_se->run_list, queue);
1243 		else
1244 			list_add_tail(&rt_se->run_list, queue);
1245 
1246 		__set_bit(rt_se_prio(rt_se), array->bitmap);
1247 		rt_se->on_list = 1;
1248 	}
1249 	rt_se->on_rq = 1;
1250 
1251 	inc_rt_tasks(rt_se, rt_rq);
1252 }
1253 
1254 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1255 {
1256 	struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1257 	struct rt_prio_array *array = &rt_rq->active;
1258 
1259 	if (move_entity(flags)) {
1260 		WARN_ON_ONCE(!rt_se->on_list);
1261 		__delist_rt_entity(rt_se, array);
1262 	}
1263 	rt_se->on_rq = 0;
1264 
1265 	dec_rt_tasks(rt_se, rt_rq);
1266 }
1267 
1268 /*
1269  * Because the prio of an upper entry depends on the lower
1270  * entries, we must remove entries top - down.
1271  */
1272 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1273 {
1274 	struct sched_rt_entity *back = NULL;
1275 
1276 	for_each_sched_rt_entity(rt_se) {
1277 		rt_se->back = back;
1278 		back = rt_se;
1279 	}
1280 
1281 	dequeue_top_rt_rq(rt_rq_of_se(back));
1282 
1283 	for (rt_se = back; rt_se; rt_se = rt_se->back) {
1284 		if (on_rt_rq(rt_se))
1285 			__dequeue_rt_entity(rt_se, flags);
1286 	}
1287 }
1288 
1289 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1290 {
1291 	struct rq *rq = rq_of_rt_se(rt_se);
1292 
1293 	dequeue_rt_stack(rt_se, flags);
1294 	for_each_sched_rt_entity(rt_se)
1295 		__enqueue_rt_entity(rt_se, flags);
1296 	enqueue_top_rt_rq(&rq->rt);
1297 }
1298 
1299 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1300 {
1301 	struct rq *rq = rq_of_rt_se(rt_se);
1302 
1303 	dequeue_rt_stack(rt_se, flags);
1304 
1305 	for_each_sched_rt_entity(rt_se) {
1306 		struct rt_rq *rt_rq = group_rt_rq(rt_se);
1307 
1308 		if (rt_rq && rt_rq->rt_nr_running)
1309 			__enqueue_rt_entity(rt_se, flags);
1310 	}
1311 	enqueue_top_rt_rq(&rq->rt);
1312 }
1313 
1314 /*
1315  * Adding/removing a task to/from a priority array:
1316  */
1317 static void
1318 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1319 {
1320 	struct sched_rt_entity *rt_se = &p->rt;
1321 
1322 	if (flags & ENQUEUE_WAKEUP)
1323 		rt_se->timeout = 0;
1324 
1325 	enqueue_rt_entity(rt_se, flags);
1326 
1327 	if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1328 		enqueue_pushable_task(rq, p);
1329 }
1330 
1331 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1332 {
1333 	struct sched_rt_entity *rt_se = &p->rt;
1334 
1335 	update_curr_rt(rq);
1336 	dequeue_rt_entity(rt_se, flags);
1337 
1338 	dequeue_pushable_task(rq, p);
1339 }
1340 
1341 /*
1342  * Put task to the head or the end of the run list without the overhead of
1343  * dequeue followed by enqueue.
1344  */
1345 static void
1346 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1347 {
1348 	if (on_rt_rq(rt_se)) {
1349 		struct rt_prio_array *array = &rt_rq->active;
1350 		struct list_head *queue = array->queue + rt_se_prio(rt_se);
1351 
1352 		if (head)
1353 			list_move(&rt_se->run_list, queue);
1354 		else
1355 			list_move_tail(&rt_se->run_list, queue);
1356 	}
1357 }
1358 
1359 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1360 {
1361 	struct sched_rt_entity *rt_se = &p->rt;
1362 	struct rt_rq *rt_rq;
1363 
1364 	for_each_sched_rt_entity(rt_se) {
1365 		rt_rq = rt_rq_of_se(rt_se);
1366 		requeue_rt_entity(rt_rq, rt_se, head);
1367 	}
1368 }
1369 
1370 static void yield_task_rt(struct rq *rq)
1371 {
1372 	requeue_task_rt(rq, rq->curr, 0);
1373 }
1374 
1375 #ifdef CONFIG_SMP
1376 static int find_lowest_rq(struct task_struct *task);
1377 
1378 static int
1379 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1380 {
1381 	struct task_struct *curr;
1382 	struct rq *rq;
1383 
1384 	/* For anything but wake ups, just return the task_cpu */
1385 	if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1386 		goto out;
1387 
1388 	rq = cpu_rq(cpu);
1389 
1390 	rcu_read_lock();
1391 	curr = READ_ONCE(rq->curr); /* unlocked access */
1392 
1393 	/*
1394 	 * If the current task on @p's runqueue is an RT task, then
1395 	 * try to see if we can wake this RT task up on another
1396 	 * runqueue. Otherwise simply start this RT task
1397 	 * on its current runqueue.
1398 	 *
1399 	 * We want to avoid overloading runqueues. If the woken
1400 	 * task is a higher priority, then it will stay on this CPU
1401 	 * and the lower prio task should be moved to another CPU.
1402 	 * Even though this will probably make the lower prio task
1403 	 * lose its cache, we do not want to bounce a higher task
1404 	 * around just because it gave up its CPU, perhaps for a
1405 	 * lock?
1406 	 *
1407 	 * For equal prio tasks, we just let the scheduler sort it out.
1408 	 *
1409 	 * Otherwise, just let it ride on the affined RQ and the
1410 	 * post-schedule router will push the preempted task away
1411 	 *
1412 	 * This test is optimistic, if we get it wrong the load-balancer
1413 	 * will have to sort it out.
1414 	 */
1415 	if (curr && unlikely(rt_task(curr)) &&
1416 	    (curr->nr_cpus_allowed < 2 ||
1417 	     curr->prio <= p->prio)) {
1418 		int target = find_lowest_rq(p);
1419 
1420 		/*
1421 		 * Don't bother moving it if the destination CPU is
1422 		 * not running a lower priority task.
1423 		 */
1424 		if (target != -1 &&
1425 		    p->prio < cpu_rq(target)->rt.highest_prio.curr)
1426 			cpu = target;
1427 	}
1428 	rcu_read_unlock();
1429 
1430 out:
1431 	return cpu;
1432 }
1433 
1434 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1435 {
1436 	/*
1437 	 * Current can't be migrated, useless to reschedule,
1438 	 * let's hope p can move out.
1439 	 */
1440 	if (rq->curr->nr_cpus_allowed == 1 ||
1441 	    !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1442 		return;
1443 
1444 	/*
1445 	 * p is migratable, so let's not schedule it and
1446 	 * see if it is pushed or pulled somewhere else.
1447 	 */
1448 	if (p->nr_cpus_allowed != 1
1449 	    && cpupri_find(&rq->rd->cpupri, p, NULL))
1450 		return;
1451 
1452 	/*
1453 	 * There appears to be other cpus that can accept
1454 	 * current and none to run 'p', so lets reschedule
1455 	 * to try and push current away:
1456 	 */
1457 	requeue_task_rt(rq, p, 1);
1458 	resched_curr(rq);
1459 }
1460 
1461 #endif /* CONFIG_SMP */
1462 
1463 /*
1464  * Preempt the current task with a newly woken task if needed:
1465  */
1466 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1467 {
1468 	if (p->prio < rq->curr->prio) {
1469 		resched_curr(rq);
1470 		return;
1471 	}
1472 
1473 #ifdef CONFIG_SMP
1474 	/*
1475 	 * If:
1476 	 *
1477 	 * - the newly woken task is of equal priority to the current task
1478 	 * - the newly woken task is non-migratable while current is migratable
1479 	 * - current will be preempted on the next reschedule
1480 	 *
1481 	 * we should check to see if current can readily move to a different
1482 	 * cpu.  If so, we will reschedule to allow the push logic to try
1483 	 * to move current somewhere else, making room for our non-migratable
1484 	 * task.
1485 	 */
1486 	if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1487 		check_preempt_equal_prio(rq, p);
1488 #endif
1489 }
1490 
1491 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1492 						   struct rt_rq *rt_rq)
1493 {
1494 	struct rt_prio_array *array = &rt_rq->active;
1495 	struct sched_rt_entity *next = NULL;
1496 	struct list_head *queue;
1497 	int idx;
1498 
1499 	idx = sched_find_first_bit(array->bitmap);
1500 	BUG_ON(idx >= MAX_RT_PRIO);
1501 
1502 	queue = array->queue + idx;
1503 	next = list_entry(queue->next, struct sched_rt_entity, run_list);
1504 
1505 	return next;
1506 }
1507 
1508 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1509 {
1510 	struct sched_rt_entity *rt_se;
1511 	struct task_struct *p;
1512 	struct rt_rq *rt_rq  = &rq->rt;
1513 
1514 	do {
1515 		rt_se = pick_next_rt_entity(rq, rt_rq);
1516 		BUG_ON(!rt_se);
1517 		rt_rq = group_rt_rq(rt_se);
1518 	} while (rt_rq);
1519 
1520 	p = rt_task_of(rt_se);
1521 	p->se.exec_start = rq_clock_task(rq);
1522 
1523 	return p;
1524 }
1525 
1526 static struct task_struct *
1527 pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1528 {
1529 	struct task_struct *p;
1530 	struct rt_rq *rt_rq = &rq->rt;
1531 
1532 	if (need_pull_rt_task(rq, prev)) {
1533 		/*
1534 		 * This is OK, because current is on_cpu, which avoids it being
1535 		 * picked for load-balance and preemption/IRQs are still
1536 		 * disabled avoiding further scheduler activity on it and we're
1537 		 * being very careful to re-start the picking loop.
1538 		 */
1539 		rq_unpin_lock(rq, rf);
1540 		pull_rt_task(rq);
1541 		rq_repin_lock(rq, rf);
1542 		/*
1543 		 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1544 		 * means a dl or stop task can slip in, in which case we need
1545 		 * to re-start task selection.
1546 		 */
1547 		if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1548 			     rq->dl.dl_nr_running))
1549 			return RETRY_TASK;
1550 	}
1551 
1552 	/*
1553 	 * We may dequeue prev's rt_rq in put_prev_task().
1554 	 * So, we update time before rt_nr_running check.
1555 	 */
1556 	if (prev->sched_class == &rt_sched_class)
1557 		update_curr_rt(rq);
1558 
1559 	if (!rt_rq->rt_queued)
1560 		return NULL;
1561 
1562 	put_prev_task(rq, prev);
1563 
1564 	p = _pick_next_task_rt(rq);
1565 
1566 	/* The running task is never eligible for pushing */
1567 	dequeue_pushable_task(rq, p);
1568 
1569 	queue_push_tasks(rq);
1570 
1571 	return p;
1572 }
1573 
1574 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1575 {
1576 	update_curr_rt(rq);
1577 
1578 	/*
1579 	 * The previous task needs to be made eligible for pushing
1580 	 * if it is still active
1581 	 */
1582 	if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1583 		enqueue_pushable_task(rq, p);
1584 }
1585 
1586 #ifdef CONFIG_SMP
1587 
1588 /* Only try algorithms three times */
1589 #define RT_MAX_TRIES 3
1590 
1591 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1592 {
1593 	if (!task_running(rq, p) &&
1594 	    cpumask_test_cpu(cpu, &p->cpus_allowed))
1595 		return 1;
1596 	return 0;
1597 }
1598 
1599 /*
1600  * Return the highest pushable rq's task, which is suitable to be executed
1601  * on the cpu, NULL otherwise
1602  */
1603 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1604 {
1605 	struct plist_head *head = &rq->rt.pushable_tasks;
1606 	struct task_struct *p;
1607 
1608 	if (!has_pushable_tasks(rq))
1609 		return NULL;
1610 
1611 	plist_for_each_entry(p, head, pushable_tasks) {
1612 		if (pick_rt_task(rq, p, cpu))
1613 			return p;
1614 	}
1615 
1616 	return NULL;
1617 }
1618 
1619 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1620 
1621 static int find_lowest_rq(struct task_struct *task)
1622 {
1623 	struct sched_domain *sd;
1624 	struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1625 	int this_cpu = smp_processor_id();
1626 	int cpu      = task_cpu(task);
1627 
1628 	/* Make sure the mask is initialized first */
1629 	if (unlikely(!lowest_mask))
1630 		return -1;
1631 
1632 	if (task->nr_cpus_allowed == 1)
1633 		return -1; /* No other targets possible */
1634 
1635 	if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1636 		return -1; /* No targets found */
1637 
1638 	/*
1639 	 * At this point we have built a mask of cpus representing the
1640 	 * lowest priority tasks in the system.  Now we want to elect
1641 	 * the best one based on our affinity and topology.
1642 	 *
1643 	 * We prioritize the last cpu that the task executed on since
1644 	 * it is most likely cache-hot in that location.
1645 	 */
1646 	if (cpumask_test_cpu(cpu, lowest_mask))
1647 		return cpu;
1648 
1649 	/*
1650 	 * Otherwise, we consult the sched_domains span maps to figure
1651 	 * out which cpu is logically closest to our hot cache data.
1652 	 */
1653 	if (!cpumask_test_cpu(this_cpu, lowest_mask))
1654 		this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1655 
1656 	rcu_read_lock();
1657 	for_each_domain(cpu, sd) {
1658 		if (sd->flags & SD_WAKE_AFFINE) {
1659 			int best_cpu;
1660 
1661 			/*
1662 			 * "this_cpu" is cheaper to preempt than a
1663 			 * remote processor.
1664 			 */
1665 			if (this_cpu != -1 &&
1666 			    cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1667 				rcu_read_unlock();
1668 				return this_cpu;
1669 			}
1670 
1671 			best_cpu = cpumask_first_and(lowest_mask,
1672 						     sched_domain_span(sd));
1673 			if (best_cpu < nr_cpu_ids) {
1674 				rcu_read_unlock();
1675 				return best_cpu;
1676 			}
1677 		}
1678 	}
1679 	rcu_read_unlock();
1680 
1681 	/*
1682 	 * And finally, if there were no matches within the domains
1683 	 * just give the caller *something* to work with from the compatible
1684 	 * locations.
1685 	 */
1686 	if (this_cpu != -1)
1687 		return this_cpu;
1688 
1689 	cpu = cpumask_any(lowest_mask);
1690 	if (cpu < nr_cpu_ids)
1691 		return cpu;
1692 	return -1;
1693 }
1694 
1695 /* Will lock the rq it finds */
1696 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1697 {
1698 	struct rq *lowest_rq = NULL;
1699 	int tries;
1700 	int cpu;
1701 
1702 	for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1703 		cpu = find_lowest_rq(task);
1704 
1705 		if ((cpu == -1) || (cpu == rq->cpu))
1706 			break;
1707 
1708 		lowest_rq = cpu_rq(cpu);
1709 
1710 		if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1711 			/*
1712 			 * Target rq has tasks of equal or higher priority,
1713 			 * retrying does not release any lock and is unlikely
1714 			 * to yield a different result.
1715 			 */
1716 			lowest_rq = NULL;
1717 			break;
1718 		}
1719 
1720 		/* if the prio of this runqueue changed, try again */
1721 		if (double_lock_balance(rq, lowest_rq)) {
1722 			/*
1723 			 * We had to unlock the run queue. In
1724 			 * the mean time, task could have
1725 			 * migrated already or had its affinity changed.
1726 			 * Also make sure that it wasn't scheduled on its rq.
1727 			 */
1728 			if (unlikely(task_rq(task) != rq ||
1729 				     !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) ||
1730 				     task_running(rq, task) ||
1731 				     !rt_task(task) ||
1732 				     !task_on_rq_queued(task))) {
1733 
1734 				double_unlock_balance(rq, lowest_rq);
1735 				lowest_rq = NULL;
1736 				break;
1737 			}
1738 		}
1739 
1740 		/* If this rq is still suitable use it. */
1741 		if (lowest_rq->rt.highest_prio.curr > task->prio)
1742 			break;
1743 
1744 		/* try again */
1745 		double_unlock_balance(rq, lowest_rq);
1746 		lowest_rq = NULL;
1747 	}
1748 
1749 	return lowest_rq;
1750 }
1751 
1752 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1753 {
1754 	struct task_struct *p;
1755 
1756 	if (!has_pushable_tasks(rq))
1757 		return NULL;
1758 
1759 	p = plist_first_entry(&rq->rt.pushable_tasks,
1760 			      struct task_struct, pushable_tasks);
1761 
1762 	BUG_ON(rq->cpu != task_cpu(p));
1763 	BUG_ON(task_current(rq, p));
1764 	BUG_ON(p->nr_cpus_allowed <= 1);
1765 
1766 	BUG_ON(!task_on_rq_queued(p));
1767 	BUG_ON(!rt_task(p));
1768 
1769 	return p;
1770 }
1771 
1772 /*
1773  * If the current CPU has more than one RT task, see if the non
1774  * running task can migrate over to a CPU that is running a task
1775  * of lesser priority.
1776  */
1777 static int push_rt_task(struct rq *rq)
1778 {
1779 	struct task_struct *next_task;
1780 	struct rq *lowest_rq;
1781 	int ret = 0;
1782 
1783 	if (!rq->rt.overloaded)
1784 		return 0;
1785 
1786 	next_task = pick_next_pushable_task(rq);
1787 	if (!next_task)
1788 		return 0;
1789 
1790 retry:
1791 	if (unlikely(next_task == rq->curr)) {
1792 		WARN_ON(1);
1793 		return 0;
1794 	}
1795 
1796 	/*
1797 	 * It's possible that the next_task slipped in of
1798 	 * higher priority than current. If that's the case
1799 	 * just reschedule current.
1800 	 */
1801 	if (unlikely(next_task->prio < rq->curr->prio)) {
1802 		resched_curr(rq);
1803 		return 0;
1804 	}
1805 
1806 	/* We might release rq lock */
1807 	get_task_struct(next_task);
1808 
1809 	/* find_lock_lowest_rq locks the rq if found */
1810 	lowest_rq = find_lock_lowest_rq(next_task, rq);
1811 	if (!lowest_rq) {
1812 		struct task_struct *task;
1813 		/*
1814 		 * find_lock_lowest_rq releases rq->lock
1815 		 * so it is possible that next_task has migrated.
1816 		 *
1817 		 * We need to make sure that the task is still on the same
1818 		 * run-queue and is also still the next task eligible for
1819 		 * pushing.
1820 		 */
1821 		task = pick_next_pushable_task(rq);
1822 		if (task_cpu(next_task) == rq->cpu && task == next_task) {
1823 			/*
1824 			 * The task hasn't migrated, and is still the next
1825 			 * eligible task, but we failed to find a run-queue
1826 			 * to push it to.  Do not retry in this case, since
1827 			 * other cpus will pull from us when ready.
1828 			 */
1829 			goto out;
1830 		}
1831 
1832 		if (!task)
1833 			/* No more tasks, just exit */
1834 			goto out;
1835 
1836 		/*
1837 		 * Something has shifted, try again.
1838 		 */
1839 		put_task_struct(next_task);
1840 		next_task = task;
1841 		goto retry;
1842 	}
1843 
1844 	deactivate_task(rq, next_task, 0);
1845 	set_task_cpu(next_task, lowest_rq->cpu);
1846 	activate_task(lowest_rq, next_task, 0);
1847 	ret = 1;
1848 
1849 	resched_curr(lowest_rq);
1850 
1851 	double_unlock_balance(rq, lowest_rq);
1852 
1853 out:
1854 	put_task_struct(next_task);
1855 
1856 	return ret;
1857 }
1858 
1859 static void push_rt_tasks(struct rq *rq)
1860 {
1861 	/* push_rt_task will return true if it moved an RT */
1862 	while (push_rt_task(rq))
1863 		;
1864 }
1865 
1866 #ifdef HAVE_RT_PUSH_IPI
1867 /*
1868  * The search for the next cpu always starts at rq->cpu and ends
1869  * when we reach rq->cpu again. It will never return rq->cpu.
1870  * This returns the next cpu to check, or nr_cpu_ids if the loop
1871  * is complete.
1872  *
1873  * rq->rt.push_cpu holds the last cpu returned by this function,
1874  * or if this is the first instance, it must hold rq->cpu.
1875  */
1876 static int rto_next_cpu(struct rq *rq)
1877 {
1878 	int prev_cpu = rq->rt.push_cpu;
1879 	int cpu;
1880 
1881 	cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1882 
1883 	/*
1884 	 * If the previous cpu is less than the rq's CPU, then it already
1885 	 * passed the end of the mask, and has started from the beginning.
1886 	 * We end if the next CPU is greater or equal to rq's CPU.
1887 	 */
1888 	if (prev_cpu < rq->cpu) {
1889 		if (cpu >= rq->cpu)
1890 			return nr_cpu_ids;
1891 
1892 	} else if (cpu >= nr_cpu_ids) {
1893 		/*
1894 		 * We passed the end of the mask, start at the beginning.
1895 		 * If the result is greater or equal to the rq's CPU, then
1896 		 * the loop is finished.
1897 		 */
1898 		cpu = cpumask_first(rq->rd->rto_mask);
1899 		if (cpu >= rq->cpu)
1900 			return nr_cpu_ids;
1901 	}
1902 	rq->rt.push_cpu = cpu;
1903 
1904 	/* Return cpu to let the caller know if the loop is finished or not */
1905 	return cpu;
1906 }
1907 
1908 static int find_next_push_cpu(struct rq *rq)
1909 {
1910 	struct rq *next_rq;
1911 	int cpu;
1912 
1913 	while (1) {
1914 		cpu = rto_next_cpu(rq);
1915 		if (cpu >= nr_cpu_ids)
1916 			break;
1917 		next_rq = cpu_rq(cpu);
1918 
1919 		/* Make sure the next rq can push to this rq */
1920 		if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1921 			break;
1922 	}
1923 
1924 	return cpu;
1925 }
1926 
1927 #define RT_PUSH_IPI_EXECUTING		1
1928 #define RT_PUSH_IPI_RESTART		2
1929 
1930 /*
1931  * When a high priority task schedules out from a CPU and a lower priority
1932  * task is scheduled in, a check is made to see if there's any RT tasks
1933  * on other CPUs that are waiting to run because a higher priority RT task
1934  * is currently running on its CPU. In this case, the CPU with multiple RT
1935  * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1936  * up that may be able to run one of its non-running queued RT tasks.
1937  *
1938  * On large CPU boxes, there's the case that several CPUs could schedule
1939  * a lower priority task at the same time, in which case it will look for
1940  * any overloaded CPUs that it could pull a task from. To do this, the runqueue
1941  * lock must be taken from that overloaded CPU. Having 10s of CPUs all fighting
1942  * for a single overloaded CPU's runqueue lock can produce a large latency.
1943  * (This has actually been observed on large boxes running cyclictest).
1944  * Instead of taking the runqueue lock of the overloaded CPU, each of the
1945  * CPUs that scheduled a lower priority task simply sends an IPI to the
1946  * overloaded CPU. An IPI is much cheaper than taking an runqueue lock with
1947  * lots of contention. The overloaded CPU will look to push its non-running
1948  * RT task off, and if it does, it can then ignore the other IPIs coming
1949  * in, and just pass those IPIs off to any other overloaded CPU.
1950  *
1951  * When a CPU schedules a lower priority task, it only sends an IPI to
1952  * the "next" CPU that has overloaded RT tasks. This prevents IPI storms,
1953  * as having 10 CPUs scheduling lower priority tasks and 10 CPUs with
1954  * RT overloaded tasks, would cause 100 IPIs to go out at once.
1955  *
1956  * The overloaded RT CPU, when receiving an IPI, will try to push off its
1957  * overloaded RT tasks and then send an IPI to the next CPU that has
1958  * overloaded RT tasks. This stops when all CPUs with overloaded RT tasks
1959  * have completed. Just because a CPU may have pushed off its own overloaded
1960  * RT task does not mean it should stop sending the IPI around to other
1961  * overloaded CPUs. There may be another RT task waiting to run on one of
1962  * those CPUs that are of higher priority than the one that was just
1963  * pushed.
1964  *
1965  * An optimization that could possibly be made is to make a CPU array similar
1966  * to the cpupri array mask of all running RT tasks, but for the overloaded
1967  * case, then the IPI could be sent to only the CPU with the highest priority
1968  * RT task waiting, and that CPU could send off further IPIs to the CPU with
1969  * the next highest waiting task. Since the overloaded case is much less likely
1970  * to happen, the complexity of this implementation may not be worth it.
1971  * Instead, just send an IPI around to all overloaded CPUs.
1972  *
1973  * The rq->rt.push_flags holds the status of the IPI that is going around.
1974  * A run queue can only send out a single IPI at a time. The possible flags
1975  * for rq->rt.push_flags are:
1976  *
1977  *    (None or zero):		No IPI is going around for the current rq
1978  *    RT_PUSH_IPI_EXECUTING:	An IPI for the rq is being passed around
1979  *    RT_PUSH_IPI_RESTART:	The priority of the running task for the rq
1980  *				has changed, and the IPI should restart
1981  *				circulating the overloaded CPUs again.
1982  *
1983  * rq->rt.push_cpu contains the CPU that is being sent the IPI. It is updated
1984  * before sending to the next CPU.
1985  *
1986  * Instead of having all CPUs that schedule a lower priority task send
1987  * an IPI to the same "first" CPU in the RT overload mask, they send it
1988  * to the next overloaded CPU after their own CPU. This helps distribute
1989  * the work when there's more than one overloaded CPU and multiple CPUs
1990  * scheduling in lower priority tasks.
1991  *
1992  * When a rq schedules a lower priority task than what was currently
1993  * running, the next CPU with overloaded RT tasks is examined first.
1994  * That is, if CPU 1 and 5 are overloaded, and CPU 3 schedules a lower
1995  * priority task, it will send an IPI first to CPU 5, then CPU 5 will
1996  * send to CPU 1 if it is still overloaded. CPU 1 will clear the
1997  * rq->rt.push_flags if RT_PUSH_IPI_RESTART is not set.
1998  *
1999  * The first CPU to notice IPI_RESTART is set, will clear that flag and then
2000  * send an IPI to the next overloaded CPU after the rq->cpu and not the next
2001  * CPU after push_cpu. That is, if CPU 1, 4 and 5 are overloaded when CPU 3
2002  * schedules a lower priority task, and the IPI_RESTART gets set while the
2003  * handling is being done on CPU 5, it will clear the flag and send it back to
2004  * CPU 4 instead of CPU 1.
2005  *
2006  * Note, the above logic can be disabled by turning off the sched_feature
2007  * RT_PUSH_IPI. Then the rq lock of the overloaded CPU will simply be
2008  * taken by the CPU requesting a pull and the waiting RT task will be pulled
2009  * by that CPU. This may be fine for machines with few CPUs.
2010  */
2011 static void tell_cpu_to_push(struct rq *rq)
2012 {
2013 	int cpu;
2014 
2015 	if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
2016 		raw_spin_lock(&rq->rt.push_lock);
2017 		/* Make sure it's still executing */
2018 		if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
2019 			/*
2020 			 * Tell the IPI to restart the loop as things have
2021 			 * changed since it started.
2022 			 */
2023 			rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
2024 			raw_spin_unlock(&rq->rt.push_lock);
2025 			return;
2026 		}
2027 		raw_spin_unlock(&rq->rt.push_lock);
2028 	}
2029 
2030 	/* When here, there's no IPI going around */
2031 
2032 	rq->rt.push_cpu = rq->cpu;
2033 	cpu = find_next_push_cpu(rq);
2034 	if (cpu >= nr_cpu_ids)
2035 		return;
2036 
2037 	rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
2038 
2039 	irq_work_queue_on(&rq->rt.push_work, cpu);
2040 }
2041 
2042 /* Called from hardirq context */
2043 static void try_to_push_tasks(void *arg)
2044 {
2045 	struct rt_rq *rt_rq = arg;
2046 	struct rq *rq, *src_rq;
2047 	int this_cpu;
2048 	int cpu;
2049 
2050 	this_cpu = rt_rq->push_cpu;
2051 
2052 	/* Paranoid check */
2053 	BUG_ON(this_cpu != smp_processor_id());
2054 
2055 	rq = cpu_rq(this_cpu);
2056 	src_rq = rq_of_rt_rq(rt_rq);
2057 
2058 again:
2059 	if (has_pushable_tasks(rq)) {
2060 		raw_spin_lock(&rq->lock);
2061 		push_rt_task(rq);
2062 		raw_spin_unlock(&rq->lock);
2063 	}
2064 
2065 	/* Pass the IPI to the next rt overloaded queue */
2066 	raw_spin_lock(&rt_rq->push_lock);
2067 	/*
2068 	 * If the source queue changed since the IPI went out,
2069 	 * we need to restart the search from that CPU again.
2070 	 */
2071 	if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
2072 		rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
2073 		rt_rq->push_cpu = src_rq->cpu;
2074 	}
2075 
2076 	cpu = find_next_push_cpu(src_rq);
2077 
2078 	if (cpu >= nr_cpu_ids)
2079 		rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
2080 	raw_spin_unlock(&rt_rq->push_lock);
2081 
2082 	if (cpu >= nr_cpu_ids)
2083 		return;
2084 
2085 	/*
2086 	 * It is possible that a restart caused this CPU to be
2087 	 * chosen again. Don't bother with an IPI, just see if we
2088 	 * have more to push.
2089 	 */
2090 	if (unlikely(cpu == rq->cpu))
2091 		goto again;
2092 
2093 	/* Try the next RT overloaded CPU */
2094 	irq_work_queue_on(&rt_rq->push_work, cpu);
2095 }
2096 
2097 static void push_irq_work_func(struct irq_work *work)
2098 {
2099 	struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
2100 
2101 	try_to_push_tasks(rt_rq);
2102 }
2103 #endif /* HAVE_RT_PUSH_IPI */
2104 
2105 static void pull_rt_task(struct rq *this_rq)
2106 {
2107 	int this_cpu = this_rq->cpu, cpu;
2108 	bool resched = false;
2109 	struct task_struct *p;
2110 	struct rq *src_rq;
2111 
2112 	if (likely(!rt_overloaded(this_rq)))
2113 		return;
2114 
2115 	/*
2116 	 * Match the barrier from rt_set_overloaded; this guarantees that if we
2117 	 * see overloaded we must also see the rto_mask bit.
2118 	 */
2119 	smp_rmb();
2120 
2121 #ifdef HAVE_RT_PUSH_IPI
2122 	if (sched_feat(RT_PUSH_IPI)) {
2123 		tell_cpu_to_push(this_rq);
2124 		return;
2125 	}
2126 #endif
2127 
2128 	for_each_cpu(cpu, this_rq->rd->rto_mask) {
2129 		if (this_cpu == cpu)
2130 			continue;
2131 
2132 		src_rq = cpu_rq(cpu);
2133 
2134 		/*
2135 		 * Don't bother taking the src_rq->lock if the next highest
2136 		 * task is known to be lower-priority than our current task.
2137 		 * This may look racy, but if this value is about to go
2138 		 * logically higher, the src_rq will push this task away.
2139 		 * And if its going logically lower, we do not care
2140 		 */
2141 		if (src_rq->rt.highest_prio.next >=
2142 		    this_rq->rt.highest_prio.curr)
2143 			continue;
2144 
2145 		/*
2146 		 * We can potentially drop this_rq's lock in
2147 		 * double_lock_balance, and another CPU could
2148 		 * alter this_rq
2149 		 */
2150 		double_lock_balance(this_rq, src_rq);
2151 
2152 		/*
2153 		 * We can pull only a task, which is pushable
2154 		 * on its rq, and no others.
2155 		 */
2156 		p = pick_highest_pushable_task(src_rq, this_cpu);
2157 
2158 		/*
2159 		 * Do we have an RT task that preempts
2160 		 * the to-be-scheduled task?
2161 		 */
2162 		if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2163 			WARN_ON(p == src_rq->curr);
2164 			WARN_ON(!task_on_rq_queued(p));
2165 
2166 			/*
2167 			 * There's a chance that p is higher in priority
2168 			 * than what's currently running on its cpu.
2169 			 * This is just that p is wakeing up and hasn't
2170 			 * had a chance to schedule. We only pull
2171 			 * p if it is lower in priority than the
2172 			 * current task on the run queue
2173 			 */
2174 			if (p->prio < src_rq->curr->prio)
2175 				goto skip;
2176 
2177 			resched = true;
2178 
2179 			deactivate_task(src_rq, p, 0);
2180 			set_task_cpu(p, this_cpu);
2181 			activate_task(this_rq, p, 0);
2182 			/*
2183 			 * We continue with the search, just in
2184 			 * case there's an even higher prio task
2185 			 * in another runqueue. (low likelihood
2186 			 * but possible)
2187 			 */
2188 		}
2189 skip:
2190 		double_unlock_balance(this_rq, src_rq);
2191 	}
2192 
2193 	if (resched)
2194 		resched_curr(this_rq);
2195 }
2196 
2197 /*
2198  * If we are not running and we are not going to reschedule soon, we should
2199  * try to push tasks away now
2200  */
2201 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2202 {
2203 	if (!task_running(rq, p) &&
2204 	    !test_tsk_need_resched(rq->curr) &&
2205 	    p->nr_cpus_allowed > 1 &&
2206 	    (dl_task(rq->curr) || rt_task(rq->curr)) &&
2207 	    (rq->curr->nr_cpus_allowed < 2 ||
2208 	     rq->curr->prio <= p->prio))
2209 		push_rt_tasks(rq);
2210 }
2211 
2212 /* Assumes rq->lock is held */
2213 static void rq_online_rt(struct rq *rq)
2214 {
2215 	if (rq->rt.overloaded)
2216 		rt_set_overload(rq);
2217 
2218 	__enable_runtime(rq);
2219 
2220 	cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2221 }
2222 
2223 /* Assumes rq->lock is held */
2224 static void rq_offline_rt(struct rq *rq)
2225 {
2226 	if (rq->rt.overloaded)
2227 		rt_clear_overload(rq);
2228 
2229 	__disable_runtime(rq);
2230 
2231 	cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2232 }
2233 
2234 /*
2235  * When switch from the rt queue, we bring ourselves to a position
2236  * that we might want to pull RT tasks from other runqueues.
2237  */
2238 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2239 {
2240 	/*
2241 	 * If there are other RT tasks then we will reschedule
2242 	 * and the scheduling of the other RT tasks will handle
2243 	 * the balancing. But if we are the last RT task
2244 	 * we may need to handle the pulling of RT tasks
2245 	 * now.
2246 	 */
2247 	if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2248 		return;
2249 
2250 	queue_pull_task(rq);
2251 }
2252 
2253 void __init init_sched_rt_class(void)
2254 {
2255 	unsigned int i;
2256 
2257 	for_each_possible_cpu(i) {
2258 		zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2259 					GFP_KERNEL, cpu_to_node(i));
2260 	}
2261 }
2262 #endif /* CONFIG_SMP */
2263 
2264 /*
2265  * When switching a task to RT, we may overload the runqueue
2266  * with RT tasks. In this case we try to push them off to
2267  * other runqueues.
2268  */
2269 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2270 {
2271 	/*
2272 	 * If we are already running, then there's nothing
2273 	 * that needs to be done. But if we are not running
2274 	 * we may need to preempt the current running task.
2275 	 * If that current running task is also an RT task
2276 	 * then see if we can move to another run queue.
2277 	 */
2278 	if (task_on_rq_queued(p) && rq->curr != p) {
2279 #ifdef CONFIG_SMP
2280 		if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2281 			queue_push_tasks(rq);
2282 #endif /* CONFIG_SMP */
2283 		if (p->prio < rq->curr->prio)
2284 			resched_curr(rq);
2285 	}
2286 }
2287 
2288 /*
2289  * Priority of the task has changed. This may cause
2290  * us to initiate a push or pull.
2291  */
2292 static void
2293 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2294 {
2295 	if (!task_on_rq_queued(p))
2296 		return;
2297 
2298 	if (rq->curr == p) {
2299 #ifdef CONFIG_SMP
2300 		/*
2301 		 * If our priority decreases while running, we
2302 		 * may need to pull tasks to this runqueue.
2303 		 */
2304 		if (oldprio < p->prio)
2305 			queue_pull_task(rq);
2306 
2307 		/*
2308 		 * If there's a higher priority task waiting to run
2309 		 * then reschedule.
2310 		 */
2311 		if (p->prio > rq->rt.highest_prio.curr)
2312 			resched_curr(rq);
2313 #else
2314 		/* For UP simply resched on drop of prio */
2315 		if (oldprio < p->prio)
2316 			resched_curr(rq);
2317 #endif /* CONFIG_SMP */
2318 	} else {
2319 		/*
2320 		 * This task is not running, but if it is
2321 		 * greater than the current running task
2322 		 * then reschedule.
2323 		 */
2324 		if (p->prio < rq->curr->prio)
2325 			resched_curr(rq);
2326 	}
2327 }
2328 
2329 #ifdef CONFIG_POSIX_TIMERS
2330 static void watchdog(struct rq *rq, struct task_struct *p)
2331 {
2332 	unsigned long soft, hard;
2333 
2334 	/* max may change after cur was read, this will be fixed next tick */
2335 	soft = task_rlimit(p, RLIMIT_RTTIME);
2336 	hard = task_rlimit_max(p, RLIMIT_RTTIME);
2337 
2338 	if (soft != RLIM_INFINITY) {
2339 		unsigned long next;
2340 
2341 		if (p->rt.watchdog_stamp != jiffies) {
2342 			p->rt.timeout++;
2343 			p->rt.watchdog_stamp = jiffies;
2344 		}
2345 
2346 		next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2347 		if (p->rt.timeout > next)
2348 			p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2349 	}
2350 }
2351 #else
2352 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2353 #endif
2354 
2355 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2356 {
2357 	struct sched_rt_entity *rt_se = &p->rt;
2358 
2359 	update_curr_rt(rq);
2360 
2361 	watchdog(rq, p);
2362 
2363 	/*
2364 	 * RR tasks need a special form of timeslice management.
2365 	 * FIFO tasks have no timeslices.
2366 	 */
2367 	if (p->policy != SCHED_RR)
2368 		return;
2369 
2370 	if (--p->rt.time_slice)
2371 		return;
2372 
2373 	p->rt.time_slice = sched_rr_timeslice;
2374 
2375 	/*
2376 	 * Requeue to the end of queue if we (and all of our ancestors) are not
2377 	 * the only element on the queue
2378 	 */
2379 	for_each_sched_rt_entity(rt_se) {
2380 		if (rt_se->run_list.prev != rt_se->run_list.next) {
2381 			requeue_task_rt(rq, p, 0);
2382 			resched_curr(rq);
2383 			return;
2384 		}
2385 	}
2386 }
2387 
2388 static void set_curr_task_rt(struct rq *rq)
2389 {
2390 	struct task_struct *p = rq->curr;
2391 
2392 	p->se.exec_start = rq_clock_task(rq);
2393 
2394 	/* The running task is never eligible for pushing */
2395 	dequeue_pushable_task(rq, p);
2396 }
2397 
2398 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2399 {
2400 	/*
2401 	 * Time slice is 0 for SCHED_FIFO tasks
2402 	 */
2403 	if (task->policy == SCHED_RR)
2404 		return sched_rr_timeslice;
2405 	else
2406 		return 0;
2407 }
2408 
2409 const struct sched_class rt_sched_class = {
2410 	.next			= &fair_sched_class,
2411 	.enqueue_task		= enqueue_task_rt,
2412 	.dequeue_task		= dequeue_task_rt,
2413 	.yield_task		= yield_task_rt,
2414 
2415 	.check_preempt_curr	= check_preempt_curr_rt,
2416 
2417 	.pick_next_task		= pick_next_task_rt,
2418 	.put_prev_task		= put_prev_task_rt,
2419 
2420 #ifdef CONFIG_SMP
2421 	.select_task_rq		= select_task_rq_rt,
2422 
2423 	.set_cpus_allowed       = set_cpus_allowed_common,
2424 	.rq_online              = rq_online_rt,
2425 	.rq_offline             = rq_offline_rt,
2426 	.task_woken		= task_woken_rt,
2427 	.switched_from		= switched_from_rt,
2428 #endif
2429 
2430 	.set_curr_task          = set_curr_task_rt,
2431 	.task_tick		= task_tick_rt,
2432 
2433 	.get_rr_interval	= get_rr_interval_rt,
2434 
2435 	.prio_changed		= prio_changed_rt,
2436 	.switched_to		= switched_to_rt,
2437 
2438 	.update_curr		= update_curr_rt,
2439 };
2440 
2441 #ifdef CONFIG_SCHED_DEBUG
2442 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2443 
2444 void print_rt_stats(struct seq_file *m, int cpu)
2445 {
2446 	rt_rq_iter_t iter;
2447 	struct rt_rq *rt_rq;
2448 
2449 	rcu_read_lock();
2450 	for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2451 		print_rt_rq(m, cpu, rt_rq);
2452 	rcu_read_unlock();
2453 }
2454 #endif /* CONFIG_SCHED_DEBUG */
2455