xref: /linux/kernel/sched/deadline.c (revision 4246b92cf9fb32da8d8b060c92d8302797c6fbea)
1 /*
2  * Deadline Scheduling Class (SCHED_DEADLINE)
3  *
4  * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
5  *
6  * Tasks that periodically executes their instances for less than their
7  * runtime won't miss any of their deadlines.
8  * Tasks that are not periodic or sporadic or that tries to execute more
9  * than their reserved bandwidth will be slowed down (and may potentially
10  * miss some of their deadlines), and won't affect any other task.
11  *
12  * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
13  *                    Juri Lelli <juri.lelli@gmail.com>,
14  *                    Michael Trimarchi <michael@amarulasolutions.com>,
15  *                    Fabio Checconi <fchecconi@gmail.com>
16  */
17 #include "sched.h"
18 
19 #include <linux/slab.h>
20 #include <uapi/linux/sched/types.h>
21 
22 struct dl_bandwidth def_dl_bandwidth;
23 
24 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
25 {
26 	return container_of(dl_se, struct task_struct, dl);
27 }
28 
29 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
30 {
31 	return container_of(dl_rq, struct rq, dl);
32 }
33 
34 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
35 {
36 	struct task_struct *p = dl_task_of(dl_se);
37 	struct rq *rq = task_rq(p);
38 
39 	return &rq->dl;
40 }
41 
42 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
43 {
44 	return !RB_EMPTY_NODE(&dl_se->rb_node);
45 }
46 
47 #ifdef CONFIG_SMP
48 static inline struct dl_bw *dl_bw_of(int i)
49 {
50 	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
51 			 "sched RCU must be held");
52 	return &cpu_rq(i)->rd->dl_bw;
53 }
54 
55 static inline int dl_bw_cpus(int i)
56 {
57 	struct root_domain *rd = cpu_rq(i)->rd;
58 	int cpus = 0;
59 
60 	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
61 			 "sched RCU must be held");
62 	for_each_cpu_and(i, rd->span, cpu_active_mask)
63 		cpus++;
64 
65 	return cpus;
66 }
67 #else
68 static inline struct dl_bw *dl_bw_of(int i)
69 {
70 	return &cpu_rq(i)->dl.dl_bw;
71 }
72 
73 static inline int dl_bw_cpus(int i)
74 {
75 	return 1;
76 }
77 #endif
78 
79 static inline
80 void add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
81 {
82 	u64 old = dl_rq->running_bw;
83 
84 	lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
85 	dl_rq->running_bw += dl_bw;
86 	SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
87 	SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
88 }
89 
90 static inline
91 void sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
92 {
93 	u64 old = dl_rq->running_bw;
94 
95 	lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
96 	dl_rq->running_bw -= dl_bw;
97 	SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
98 	if (dl_rq->running_bw > old)
99 		dl_rq->running_bw = 0;
100 }
101 
102 static inline
103 void add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
104 {
105 	u64 old = dl_rq->this_bw;
106 
107 	lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
108 	dl_rq->this_bw += dl_bw;
109 	SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
110 }
111 
112 static inline
113 void sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
114 {
115 	u64 old = dl_rq->this_bw;
116 
117 	lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
118 	dl_rq->this_bw -= dl_bw;
119 	SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
120 	if (dl_rq->this_bw > old)
121 		dl_rq->this_bw = 0;
122 	SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
123 }
124 
125 void dl_change_utilization(struct task_struct *p, u64 new_bw)
126 {
127 	struct rq *rq;
128 
129 	if (task_on_rq_queued(p))
130 		return;
131 
132 	rq = task_rq(p);
133 	if (p->dl.dl_non_contending) {
134 		sub_running_bw(p->dl.dl_bw, &rq->dl);
135 		p->dl.dl_non_contending = 0;
136 		/*
137 		 * If the timer handler is currently running and the
138 		 * timer cannot be cancelled, inactive_task_timer()
139 		 * will see that dl_not_contending is not set, and
140 		 * will not touch the rq's active utilization,
141 		 * so we are still safe.
142 		 */
143 		if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
144 			put_task_struct(p);
145 	}
146 	sub_rq_bw(p->dl.dl_bw, &rq->dl);
147 	add_rq_bw(new_bw, &rq->dl);
148 }
149 
150 /*
151  * The utilization of a task cannot be immediately removed from
152  * the rq active utilization (running_bw) when the task blocks.
153  * Instead, we have to wait for the so called "0-lag time".
154  *
155  * If a task blocks before the "0-lag time", a timer (the inactive
156  * timer) is armed, and running_bw is decreased when the timer
157  * fires.
158  *
159  * If the task wakes up again before the inactive timer fires,
160  * the timer is cancelled, whereas if the task wakes up after the
161  * inactive timer fired (and running_bw has been decreased) the
162  * task's utilization has to be added to running_bw again.
163  * A flag in the deadline scheduling entity (dl_non_contending)
164  * is used to avoid race conditions between the inactive timer handler
165  * and task wakeups.
166  *
167  * The following diagram shows how running_bw is updated. A task is
168  * "ACTIVE" when its utilization contributes to running_bw; an
169  * "ACTIVE contending" task is in the TASK_RUNNING state, while an
170  * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
171  * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
172  * time already passed, which does not contribute to running_bw anymore.
173  *                              +------------------+
174  *             wakeup           |    ACTIVE        |
175  *          +------------------>+   contending     |
176  *          | add_running_bw    |                  |
177  *          |                   +----+------+------+
178  *          |                        |      ^
179  *          |                dequeue |      |
180  * +--------+-------+                |      |
181  * |                |   t >= 0-lag   |      | wakeup
182  * |    INACTIVE    |<---------------+      |
183  * |                | sub_running_bw |      |
184  * +--------+-------+                |      |
185  *          ^                        |      |
186  *          |              t < 0-lag |      |
187  *          |                        |      |
188  *          |                        V      |
189  *          |                   +----+------+------+
190  *          | sub_running_bw    |    ACTIVE        |
191  *          +-------------------+                  |
192  *            inactive timer    |  non contending  |
193  *            fired             +------------------+
194  *
195  * The task_non_contending() function is invoked when a task
196  * blocks, and checks if the 0-lag time already passed or
197  * not (in the first case, it directly updates running_bw;
198  * in the second case, it arms the inactive timer).
199  *
200  * The task_contending() function is invoked when a task wakes
201  * up, and checks if the task is still in the "ACTIVE non contending"
202  * state or not (in the second case, it updates running_bw).
203  */
204 static void task_non_contending(struct task_struct *p)
205 {
206 	struct sched_dl_entity *dl_se = &p->dl;
207 	struct hrtimer *timer = &dl_se->inactive_timer;
208 	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
209 	struct rq *rq = rq_of_dl_rq(dl_rq);
210 	s64 zerolag_time;
211 
212 	/*
213 	 * If this is a non-deadline task that has been boosted,
214 	 * do nothing
215 	 */
216 	if (dl_se->dl_runtime == 0)
217 		return;
218 
219 	WARN_ON(hrtimer_active(&dl_se->inactive_timer));
220 	WARN_ON(dl_se->dl_non_contending);
221 
222 	zerolag_time = dl_se->deadline -
223 		 div64_long((dl_se->runtime * dl_se->dl_period),
224 			dl_se->dl_runtime);
225 
226 	/*
227 	 * Using relative times instead of the absolute "0-lag time"
228 	 * allows to simplify the code
229 	 */
230 	zerolag_time -= rq_clock(rq);
231 
232 	/*
233 	 * If the "0-lag time" already passed, decrease the active
234 	 * utilization now, instead of starting a timer
235 	 */
236 	if (zerolag_time < 0) {
237 		if (dl_task(p))
238 			sub_running_bw(dl_se->dl_bw, dl_rq);
239 		if (!dl_task(p) || p->state == TASK_DEAD) {
240 			struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
241 
242 			if (p->state == TASK_DEAD)
243 				sub_rq_bw(p->dl.dl_bw, &rq->dl);
244 			raw_spin_lock(&dl_b->lock);
245 			__dl_clear(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
246 			__dl_clear_params(p);
247 			raw_spin_unlock(&dl_b->lock);
248 		}
249 
250 		return;
251 	}
252 
253 	dl_se->dl_non_contending = 1;
254 	get_task_struct(p);
255 	hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL);
256 }
257 
258 static void task_contending(struct sched_dl_entity *dl_se, int flags)
259 {
260 	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
261 
262 	/*
263 	 * If this is a non-deadline task that has been boosted,
264 	 * do nothing
265 	 */
266 	if (dl_se->dl_runtime == 0)
267 		return;
268 
269 	if (flags & ENQUEUE_MIGRATED)
270 		add_rq_bw(dl_se->dl_bw, dl_rq);
271 
272 	if (dl_se->dl_non_contending) {
273 		dl_se->dl_non_contending = 0;
274 		/*
275 		 * If the timer handler is currently running and the
276 		 * timer cannot be cancelled, inactive_task_timer()
277 		 * will see that dl_not_contending is not set, and
278 		 * will not touch the rq's active utilization,
279 		 * so we are still safe.
280 		 */
281 		if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
282 			put_task_struct(dl_task_of(dl_se));
283 	} else {
284 		/*
285 		 * Since "dl_non_contending" is not set, the
286 		 * task's utilization has already been removed from
287 		 * active utilization (either when the task blocked,
288 		 * when the "inactive timer" fired).
289 		 * So, add it back.
290 		 */
291 		add_running_bw(dl_se->dl_bw, dl_rq);
292 	}
293 }
294 
295 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
296 {
297 	struct sched_dl_entity *dl_se = &p->dl;
298 
299 	return dl_rq->root.rb_leftmost == &dl_se->rb_node;
300 }
301 
302 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
303 {
304 	raw_spin_lock_init(&dl_b->dl_runtime_lock);
305 	dl_b->dl_period = period;
306 	dl_b->dl_runtime = runtime;
307 }
308 
309 void init_dl_bw(struct dl_bw *dl_b)
310 {
311 	raw_spin_lock_init(&dl_b->lock);
312 	raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
313 	if (global_rt_runtime() == RUNTIME_INF)
314 		dl_b->bw = -1;
315 	else
316 		dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
317 	raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
318 	dl_b->total_bw = 0;
319 }
320 
321 void init_dl_rq(struct dl_rq *dl_rq)
322 {
323 	dl_rq->root = RB_ROOT_CACHED;
324 
325 #ifdef CONFIG_SMP
326 	/* zero means no -deadline tasks */
327 	dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
328 
329 	dl_rq->dl_nr_migratory = 0;
330 	dl_rq->overloaded = 0;
331 	dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
332 #else
333 	init_dl_bw(&dl_rq->dl_bw);
334 #endif
335 
336 	dl_rq->running_bw = 0;
337 	dl_rq->this_bw = 0;
338 	init_dl_rq_bw_ratio(dl_rq);
339 }
340 
341 #ifdef CONFIG_SMP
342 
343 static inline int dl_overloaded(struct rq *rq)
344 {
345 	return atomic_read(&rq->rd->dlo_count);
346 }
347 
348 static inline void dl_set_overload(struct rq *rq)
349 {
350 	if (!rq->online)
351 		return;
352 
353 	cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
354 	/*
355 	 * Must be visible before the overload count is
356 	 * set (as in sched_rt.c).
357 	 *
358 	 * Matched by the barrier in pull_dl_task().
359 	 */
360 	smp_wmb();
361 	atomic_inc(&rq->rd->dlo_count);
362 }
363 
364 static inline void dl_clear_overload(struct rq *rq)
365 {
366 	if (!rq->online)
367 		return;
368 
369 	atomic_dec(&rq->rd->dlo_count);
370 	cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
371 }
372 
373 static void update_dl_migration(struct dl_rq *dl_rq)
374 {
375 	if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
376 		if (!dl_rq->overloaded) {
377 			dl_set_overload(rq_of_dl_rq(dl_rq));
378 			dl_rq->overloaded = 1;
379 		}
380 	} else if (dl_rq->overloaded) {
381 		dl_clear_overload(rq_of_dl_rq(dl_rq));
382 		dl_rq->overloaded = 0;
383 	}
384 }
385 
386 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
387 {
388 	struct task_struct *p = dl_task_of(dl_se);
389 
390 	if (p->nr_cpus_allowed > 1)
391 		dl_rq->dl_nr_migratory++;
392 
393 	update_dl_migration(dl_rq);
394 }
395 
396 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
397 {
398 	struct task_struct *p = dl_task_of(dl_se);
399 
400 	if (p->nr_cpus_allowed > 1)
401 		dl_rq->dl_nr_migratory--;
402 
403 	update_dl_migration(dl_rq);
404 }
405 
406 /*
407  * The list of pushable -deadline task is not a plist, like in
408  * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
409  */
410 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
411 {
412 	struct dl_rq *dl_rq = &rq->dl;
413 	struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node;
414 	struct rb_node *parent = NULL;
415 	struct task_struct *entry;
416 	bool leftmost = true;
417 
418 	BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
419 
420 	while (*link) {
421 		parent = *link;
422 		entry = rb_entry(parent, struct task_struct,
423 				 pushable_dl_tasks);
424 		if (dl_entity_preempt(&p->dl, &entry->dl))
425 			link = &parent->rb_left;
426 		else {
427 			link = &parent->rb_right;
428 			leftmost = false;
429 		}
430 	}
431 
432 	if (leftmost)
433 		dl_rq->earliest_dl.next = p->dl.deadline;
434 
435 	rb_link_node(&p->pushable_dl_tasks, parent, link);
436 	rb_insert_color_cached(&p->pushable_dl_tasks,
437 			       &dl_rq->pushable_dl_tasks_root, leftmost);
438 }
439 
440 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
441 {
442 	struct dl_rq *dl_rq = &rq->dl;
443 
444 	if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
445 		return;
446 
447 	if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) {
448 		struct rb_node *next_node;
449 
450 		next_node = rb_next(&p->pushable_dl_tasks);
451 		if (next_node) {
452 			dl_rq->earliest_dl.next = rb_entry(next_node,
453 				struct task_struct, pushable_dl_tasks)->dl.deadline;
454 		}
455 	}
456 
457 	rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
458 	RB_CLEAR_NODE(&p->pushable_dl_tasks);
459 }
460 
461 static inline int has_pushable_dl_tasks(struct rq *rq)
462 {
463 	return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
464 }
465 
466 static int push_dl_task(struct rq *rq);
467 
468 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
469 {
470 	return dl_task(prev);
471 }
472 
473 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
474 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
475 
476 static void push_dl_tasks(struct rq *);
477 static void pull_dl_task(struct rq *);
478 
479 static inline void queue_push_tasks(struct rq *rq)
480 {
481 	if (!has_pushable_dl_tasks(rq))
482 		return;
483 
484 	queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
485 }
486 
487 static inline void queue_pull_task(struct rq *rq)
488 {
489 	queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
490 }
491 
492 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
493 
494 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
495 {
496 	struct rq *later_rq = NULL;
497 
498 	later_rq = find_lock_later_rq(p, rq);
499 	if (!later_rq) {
500 		int cpu;
501 
502 		/*
503 		 * If we cannot preempt any rq, fall back to pick any
504 		 * online cpu.
505 		 */
506 		cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
507 		if (cpu >= nr_cpu_ids) {
508 			/*
509 			 * Fail to find any suitable cpu.
510 			 * The task will never come back!
511 			 */
512 			BUG_ON(dl_bandwidth_enabled());
513 
514 			/*
515 			 * If admission control is disabled we
516 			 * try a little harder to let the task
517 			 * run.
518 			 */
519 			cpu = cpumask_any(cpu_active_mask);
520 		}
521 		later_rq = cpu_rq(cpu);
522 		double_lock_balance(rq, later_rq);
523 	}
524 
525 	set_task_cpu(p, later_rq->cpu);
526 	double_unlock_balance(later_rq, rq);
527 
528 	return later_rq;
529 }
530 
531 #else
532 
533 static inline
534 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
535 {
536 }
537 
538 static inline
539 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
540 {
541 }
542 
543 static inline
544 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
545 {
546 }
547 
548 static inline
549 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
550 {
551 }
552 
553 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
554 {
555 	return false;
556 }
557 
558 static inline void pull_dl_task(struct rq *rq)
559 {
560 }
561 
562 static inline void queue_push_tasks(struct rq *rq)
563 {
564 }
565 
566 static inline void queue_pull_task(struct rq *rq)
567 {
568 }
569 #endif /* CONFIG_SMP */
570 
571 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
572 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
573 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
574 				  int flags);
575 
576 /*
577  * We are being explicitly informed that a new instance is starting,
578  * and this means that:
579  *  - the absolute deadline of the entity has to be placed at
580  *    current time + relative deadline;
581  *  - the runtime of the entity has to be set to the maximum value.
582  *
583  * The capability of specifying such event is useful whenever a -deadline
584  * entity wants to (try to!) synchronize its behaviour with the scheduler's
585  * one, and to (try to!) reconcile itself with its own scheduling
586  * parameters.
587  */
588 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
589 {
590 	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
591 	struct rq *rq = rq_of_dl_rq(dl_rq);
592 
593 	WARN_ON(dl_se->dl_boosted);
594 	WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
595 
596 	/*
597 	 * We are racing with the deadline timer. So, do nothing because
598 	 * the deadline timer handler will take care of properly recharging
599 	 * the runtime and postponing the deadline
600 	 */
601 	if (dl_se->dl_throttled)
602 		return;
603 
604 	/*
605 	 * We use the regular wall clock time to set deadlines in the
606 	 * future; in fact, we must consider execution overheads (time
607 	 * spent on hardirq context, etc.).
608 	 */
609 	dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
610 	dl_se->runtime = dl_se->dl_runtime;
611 }
612 
613 /*
614  * Pure Earliest Deadline First (EDF) scheduling does not deal with the
615  * possibility of a entity lasting more than what it declared, and thus
616  * exhausting its runtime.
617  *
618  * Here we are interested in making runtime overrun possible, but we do
619  * not want a entity which is misbehaving to affect the scheduling of all
620  * other entities.
621  * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
622  * is used, in order to confine each entity within its own bandwidth.
623  *
624  * This function deals exactly with that, and ensures that when the runtime
625  * of a entity is replenished, its deadline is also postponed. That ensures
626  * the overrunning entity can't interfere with other entity in the system and
627  * can't make them miss their deadlines. Reasons why this kind of overruns
628  * could happen are, typically, a entity voluntarily trying to overcome its
629  * runtime, or it just underestimated it during sched_setattr().
630  */
631 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
632 				struct sched_dl_entity *pi_se)
633 {
634 	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
635 	struct rq *rq = rq_of_dl_rq(dl_rq);
636 
637 	BUG_ON(pi_se->dl_runtime <= 0);
638 
639 	/*
640 	 * This could be the case for a !-dl task that is boosted.
641 	 * Just go with full inherited parameters.
642 	 */
643 	if (dl_se->dl_deadline == 0) {
644 		dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
645 		dl_se->runtime = pi_se->dl_runtime;
646 	}
647 
648 	if (dl_se->dl_yielded && dl_se->runtime > 0)
649 		dl_se->runtime = 0;
650 
651 	/*
652 	 * We keep moving the deadline away until we get some
653 	 * available runtime for the entity. This ensures correct
654 	 * handling of situations where the runtime overrun is
655 	 * arbitrary large.
656 	 */
657 	while (dl_se->runtime <= 0) {
658 		dl_se->deadline += pi_se->dl_period;
659 		dl_se->runtime += pi_se->dl_runtime;
660 	}
661 
662 	/*
663 	 * At this point, the deadline really should be "in
664 	 * the future" with respect to rq->clock. If it's
665 	 * not, we are, for some reason, lagging too much!
666 	 * Anyway, after having warn userspace abut that,
667 	 * we still try to keep the things running by
668 	 * resetting the deadline and the budget of the
669 	 * entity.
670 	 */
671 	if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
672 		printk_deferred_once("sched: DL replenish lagged too much\n");
673 		dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
674 		dl_se->runtime = pi_se->dl_runtime;
675 	}
676 
677 	if (dl_se->dl_yielded)
678 		dl_se->dl_yielded = 0;
679 	if (dl_se->dl_throttled)
680 		dl_se->dl_throttled = 0;
681 }
682 
683 /*
684  * Here we check if --at time t-- an entity (which is probably being
685  * [re]activated or, in general, enqueued) can use its remaining runtime
686  * and its current deadline _without_ exceeding the bandwidth it is
687  * assigned (function returns true if it can't). We are in fact applying
688  * one of the CBS rules: when a task wakes up, if the residual runtime
689  * over residual deadline fits within the allocated bandwidth, then we
690  * can keep the current (absolute) deadline and residual budget without
691  * disrupting the schedulability of the system. Otherwise, we should
692  * refill the runtime and set the deadline a period in the future,
693  * because keeping the current (absolute) deadline of the task would
694  * result in breaking guarantees promised to other tasks (refer to
695  * Documentation/scheduler/sched-deadline.txt for more informations).
696  *
697  * This function returns true if:
698  *
699  *   runtime / (deadline - t) > dl_runtime / dl_deadline ,
700  *
701  * IOW we can't recycle current parameters.
702  *
703  * Notice that the bandwidth check is done against the deadline. For
704  * task with deadline equal to period this is the same of using
705  * dl_period instead of dl_deadline in the equation above.
706  */
707 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
708 			       struct sched_dl_entity *pi_se, u64 t)
709 {
710 	u64 left, right;
711 
712 	/*
713 	 * left and right are the two sides of the equation above,
714 	 * after a bit of shuffling to use multiplications instead
715 	 * of divisions.
716 	 *
717 	 * Note that none of the time values involved in the two
718 	 * multiplications are absolute: dl_deadline and dl_runtime
719 	 * are the relative deadline and the maximum runtime of each
720 	 * instance, runtime is the runtime left for the last instance
721 	 * and (deadline - t), since t is rq->clock, is the time left
722 	 * to the (absolute) deadline. Even if overflowing the u64 type
723 	 * is very unlikely to occur in both cases, here we scale down
724 	 * as we want to avoid that risk at all. Scaling down by 10
725 	 * means that we reduce granularity to 1us. We are fine with it,
726 	 * since this is only a true/false check and, anyway, thinking
727 	 * of anything below microseconds resolution is actually fiction
728 	 * (but still we want to give the user that illusion >;).
729 	 */
730 	left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
731 	right = ((dl_se->deadline - t) >> DL_SCALE) *
732 		(pi_se->dl_runtime >> DL_SCALE);
733 
734 	return dl_time_before(right, left);
735 }
736 
737 /*
738  * Revised wakeup rule [1]: For self-suspending tasks, rather then
739  * re-initializing task's runtime and deadline, the revised wakeup
740  * rule adjusts the task's runtime to avoid the task to overrun its
741  * density.
742  *
743  * Reasoning: a task may overrun the density if:
744  *    runtime / (deadline - t) > dl_runtime / dl_deadline
745  *
746  * Therefore, runtime can be adjusted to:
747  *     runtime = (dl_runtime / dl_deadline) * (deadline - t)
748  *
749  * In such way that runtime will be equal to the maximum density
750  * the task can use without breaking any rule.
751  *
752  * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
753  * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
754  */
755 static void
756 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
757 {
758 	u64 laxity = dl_se->deadline - rq_clock(rq);
759 
760 	/*
761 	 * If the task has deadline < period, and the deadline is in the past,
762 	 * it should already be throttled before this check.
763 	 *
764 	 * See update_dl_entity() comments for further details.
765 	 */
766 	WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
767 
768 	dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
769 }
770 
771 /*
772  * Regarding the deadline, a task with implicit deadline has a relative
773  * deadline == relative period. A task with constrained deadline has a
774  * relative deadline <= relative period.
775  *
776  * We support constrained deadline tasks. However, there are some restrictions
777  * applied only for tasks which do not have an implicit deadline. See
778  * update_dl_entity() to know more about such restrictions.
779  *
780  * The dl_is_implicit() returns true if the task has an implicit deadline.
781  */
782 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
783 {
784 	return dl_se->dl_deadline == dl_se->dl_period;
785 }
786 
787 /*
788  * When a deadline entity is placed in the runqueue, its runtime and deadline
789  * might need to be updated. This is done by a CBS wake up rule. There are two
790  * different rules: 1) the original CBS; and 2) the Revisited CBS.
791  *
792  * When the task is starting a new period, the Original CBS is used. In this
793  * case, the runtime is replenished and a new absolute deadline is set.
794  *
795  * When a task is queued before the begin of the next period, using the
796  * remaining runtime and deadline could make the entity to overflow, see
797  * dl_entity_overflow() to find more about runtime overflow. When such case
798  * is detected, the runtime and deadline need to be updated.
799  *
800  * If the task has an implicit deadline, i.e., deadline == period, the Original
801  * CBS is applied. the runtime is replenished and a new absolute deadline is
802  * set, as in the previous cases.
803  *
804  * However, the Original CBS does not work properly for tasks with
805  * deadline < period, which are said to have a constrained deadline. By
806  * applying the Original CBS, a constrained deadline task would be able to run
807  * runtime/deadline in a period. With deadline < period, the task would
808  * overrun the runtime/period allowed bandwidth, breaking the admission test.
809  *
810  * In order to prevent this misbehave, the Revisited CBS is used for
811  * constrained deadline tasks when a runtime overflow is detected. In the
812  * Revisited CBS, rather than replenishing & setting a new absolute deadline,
813  * the remaining runtime of the task is reduced to avoid runtime overflow.
814  * Please refer to the comments update_dl_revised_wakeup() function to find
815  * more about the Revised CBS rule.
816  */
817 static void update_dl_entity(struct sched_dl_entity *dl_se,
818 			     struct sched_dl_entity *pi_se)
819 {
820 	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
821 	struct rq *rq = rq_of_dl_rq(dl_rq);
822 
823 	if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
824 	    dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
825 
826 		if (unlikely(!dl_is_implicit(dl_se) &&
827 			     !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
828 			     !dl_se->dl_boosted)){
829 			update_dl_revised_wakeup(dl_se, rq);
830 			return;
831 		}
832 
833 		dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
834 		dl_se->runtime = pi_se->dl_runtime;
835 	}
836 }
837 
838 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
839 {
840 	return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
841 }
842 
843 /*
844  * If the entity depleted all its runtime, and if we want it to sleep
845  * while waiting for some new execution time to become available, we
846  * set the bandwidth replenishment timer to the replenishment instant
847  * and try to activate it.
848  *
849  * Notice that it is important for the caller to know if the timer
850  * actually started or not (i.e., the replenishment instant is in
851  * the future or in the past).
852  */
853 static int start_dl_timer(struct task_struct *p)
854 {
855 	struct sched_dl_entity *dl_se = &p->dl;
856 	struct hrtimer *timer = &dl_se->dl_timer;
857 	struct rq *rq = task_rq(p);
858 	ktime_t now, act;
859 	s64 delta;
860 
861 	lockdep_assert_held(&rq->lock);
862 
863 	/*
864 	 * We want the timer to fire at the deadline, but considering
865 	 * that it is actually coming from rq->clock and not from
866 	 * hrtimer's time base reading.
867 	 */
868 	act = ns_to_ktime(dl_next_period(dl_se));
869 	now = hrtimer_cb_get_time(timer);
870 	delta = ktime_to_ns(now) - rq_clock(rq);
871 	act = ktime_add_ns(act, delta);
872 
873 	/*
874 	 * If the expiry time already passed, e.g., because the value
875 	 * chosen as the deadline is too small, don't even try to
876 	 * start the timer in the past!
877 	 */
878 	if (ktime_us_delta(act, now) < 0)
879 		return 0;
880 
881 	/*
882 	 * !enqueued will guarantee another callback; even if one is already in
883 	 * progress. This ensures a balanced {get,put}_task_struct().
884 	 *
885 	 * The race against __run_timer() clearing the enqueued state is
886 	 * harmless because we're holding task_rq()->lock, therefore the timer
887 	 * expiring after we've done the check will wait on its task_rq_lock()
888 	 * and observe our state.
889 	 */
890 	if (!hrtimer_is_queued(timer)) {
891 		get_task_struct(p);
892 		hrtimer_start(timer, act, HRTIMER_MODE_ABS);
893 	}
894 
895 	return 1;
896 }
897 
898 /*
899  * This is the bandwidth enforcement timer callback. If here, we know
900  * a task is not on its dl_rq, since the fact that the timer was running
901  * means the task is throttled and needs a runtime replenishment.
902  *
903  * However, what we actually do depends on the fact the task is active,
904  * (it is on its rq) or has been removed from there by a call to
905  * dequeue_task_dl(). In the former case we must issue the runtime
906  * replenishment and add the task back to the dl_rq; in the latter, we just
907  * do nothing but clearing dl_throttled, so that runtime and deadline
908  * updating (and the queueing back to dl_rq) will be done by the
909  * next call to enqueue_task_dl().
910  */
911 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
912 {
913 	struct sched_dl_entity *dl_se = container_of(timer,
914 						     struct sched_dl_entity,
915 						     dl_timer);
916 	struct task_struct *p = dl_task_of(dl_se);
917 	struct rq_flags rf;
918 	struct rq *rq;
919 
920 	rq = task_rq_lock(p, &rf);
921 
922 	/*
923 	 * The task might have changed its scheduling policy to something
924 	 * different than SCHED_DEADLINE (through switched_from_dl()).
925 	 */
926 	if (!dl_task(p))
927 		goto unlock;
928 
929 	/*
930 	 * The task might have been boosted by someone else and might be in the
931 	 * boosting/deboosting path, its not throttled.
932 	 */
933 	if (dl_se->dl_boosted)
934 		goto unlock;
935 
936 	/*
937 	 * Spurious timer due to start_dl_timer() race; or we already received
938 	 * a replenishment from rt_mutex_setprio().
939 	 */
940 	if (!dl_se->dl_throttled)
941 		goto unlock;
942 
943 	sched_clock_tick();
944 	update_rq_clock(rq);
945 
946 	/*
947 	 * If the throttle happened during sched-out; like:
948 	 *
949 	 *   schedule()
950 	 *     deactivate_task()
951 	 *       dequeue_task_dl()
952 	 *         update_curr_dl()
953 	 *           start_dl_timer()
954 	 *         __dequeue_task_dl()
955 	 *     prev->on_rq = 0;
956 	 *
957 	 * We can be both throttled and !queued. Replenish the counter
958 	 * but do not enqueue -- wait for our wakeup to do that.
959 	 */
960 	if (!task_on_rq_queued(p)) {
961 		replenish_dl_entity(dl_se, dl_se);
962 		goto unlock;
963 	}
964 
965 #ifdef CONFIG_SMP
966 	if (unlikely(!rq->online)) {
967 		/*
968 		 * If the runqueue is no longer available, migrate the
969 		 * task elsewhere. This necessarily changes rq.
970 		 */
971 		lockdep_unpin_lock(&rq->lock, rf.cookie);
972 		rq = dl_task_offline_migration(rq, p);
973 		rf.cookie = lockdep_pin_lock(&rq->lock);
974 		update_rq_clock(rq);
975 
976 		/*
977 		 * Now that the task has been migrated to the new RQ and we
978 		 * have that locked, proceed as normal and enqueue the task
979 		 * there.
980 		 */
981 	}
982 #endif
983 
984 	enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
985 	if (dl_task(rq->curr))
986 		check_preempt_curr_dl(rq, p, 0);
987 	else
988 		resched_curr(rq);
989 
990 #ifdef CONFIG_SMP
991 	/*
992 	 * Queueing this task back might have overloaded rq, check if we need
993 	 * to kick someone away.
994 	 */
995 	if (has_pushable_dl_tasks(rq)) {
996 		/*
997 		 * Nothing relies on rq->lock after this, so its safe to drop
998 		 * rq->lock.
999 		 */
1000 		rq_unpin_lock(rq, &rf);
1001 		push_dl_task(rq);
1002 		rq_repin_lock(rq, &rf);
1003 	}
1004 #endif
1005 
1006 unlock:
1007 	task_rq_unlock(rq, p, &rf);
1008 
1009 	/*
1010 	 * This can free the task_struct, including this hrtimer, do not touch
1011 	 * anything related to that after this.
1012 	 */
1013 	put_task_struct(p);
1014 
1015 	return HRTIMER_NORESTART;
1016 }
1017 
1018 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1019 {
1020 	struct hrtimer *timer = &dl_se->dl_timer;
1021 
1022 	hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1023 	timer->function = dl_task_timer;
1024 }
1025 
1026 /*
1027  * During the activation, CBS checks if it can reuse the current task's
1028  * runtime and period. If the deadline of the task is in the past, CBS
1029  * cannot use the runtime, and so it replenishes the task. This rule
1030  * works fine for implicit deadline tasks (deadline == period), and the
1031  * CBS was designed for implicit deadline tasks. However, a task with
1032  * constrained deadline (deadine < period) might be awakened after the
1033  * deadline, but before the next period. In this case, replenishing the
1034  * task would allow it to run for runtime / deadline. As in this case
1035  * deadline < period, CBS enables a task to run for more than the
1036  * runtime / period. In a very loaded system, this can cause a domino
1037  * effect, making other tasks miss their deadlines.
1038  *
1039  * To avoid this problem, in the activation of a constrained deadline
1040  * task after the deadline but before the next period, throttle the
1041  * task and set the replenishing timer to the begin of the next period,
1042  * unless it is boosted.
1043  */
1044 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1045 {
1046 	struct task_struct *p = dl_task_of(dl_se);
1047 	struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1048 
1049 	if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1050 	    dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1051 		if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1052 			return;
1053 		dl_se->dl_throttled = 1;
1054 		if (dl_se->runtime > 0)
1055 			dl_se->runtime = 0;
1056 	}
1057 }
1058 
1059 static
1060 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1061 {
1062 	return (dl_se->runtime <= 0);
1063 }
1064 
1065 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1066 
1067 /*
1068  * This function implements the GRUB accounting rule:
1069  * according to the GRUB reclaiming algorithm, the runtime is
1070  * not decreased as "dq = -dt", but as
1071  * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1072  * where u is the utilization of the task, Umax is the maximum reclaimable
1073  * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1074  * as the difference between the "total runqueue utilization" and the
1075  * runqueue active utilization, and Uextra is the (per runqueue) extra
1076  * reclaimable utilization.
1077  * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1078  * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1079  * BW_SHIFT.
1080  * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1081  * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1082  * Since delta is a 64 bit variable, to have an overflow its value
1083  * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1084  * So, overflow is not an issue here.
1085  */
1086 u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1087 {
1088 	u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1089 	u64 u_act;
1090 	u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1091 
1092 	/*
1093 	 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1094 	 * we compare u_inact + rq->dl.extra_bw with
1095 	 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1096 	 * u_inact + rq->dl.extra_bw can be larger than
1097 	 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1098 	 * leading to wrong results)
1099 	 */
1100 	if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1101 		u_act = u_act_min;
1102 	else
1103 		u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1104 
1105 	return (delta * u_act) >> BW_SHIFT;
1106 }
1107 
1108 /*
1109  * Update the current task's runtime statistics (provided it is still
1110  * a -deadline task and has not been removed from the dl_rq).
1111  */
1112 static void update_curr_dl(struct rq *rq)
1113 {
1114 	struct task_struct *curr = rq->curr;
1115 	struct sched_dl_entity *dl_se = &curr->dl;
1116 	u64 delta_exec;
1117 
1118 	if (!dl_task(curr) || !on_dl_rq(dl_se))
1119 		return;
1120 
1121 	/*
1122 	 * Consumed budget is computed considering the time as
1123 	 * observed by schedulable tasks (excluding time spent
1124 	 * in hardirq context, etc.). Deadlines are instead
1125 	 * computed using hard walltime. This seems to be the more
1126 	 * natural solution, but the full ramifications of this
1127 	 * approach need further study.
1128 	 */
1129 	delta_exec = rq_clock_task(rq) - curr->se.exec_start;
1130 	if (unlikely((s64)delta_exec <= 0)) {
1131 		if (unlikely(dl_se->dl_yielded))
1132 			goto throttle;
1133 		return;
1134 	}
1135 
1136 	/* kick cpufreq (see the comment in kernel/sched/sched.h). */
1137 	cpufreq_update_util(rq, SCHED_CPUFREQ_DL);
1138 
1139 	schedstat_set(curr->se.statistics.exec_max,
1140 		      max(curr->se.statistics.exec_max, delta_exec));
1141 
1142 	curr->se.sum_exec_runtime += delta_exec;
1143 	account_group_exec_runtime(curr, delta_exec);
1144 
1145 	curr->se.exec_start = rq_clock_task(rq);
1146 	cpuacct_charge(curr, delta_exec);
1147 
1148 	sched_rt_avg_update(rq, delta_exec);
1149 
1150 	if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM))
1151 		delta_exec = grub_reclaim(delta_exec, rq, &curr->dl);
1152 	dl_se->runtime -= delta_exec;
1153 
1154 throttle:
1155 	if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1156 		dl_se->dl_throttled = 1;
1157 		__dequeue_task_dl(rq, curr, 0);
1158 		if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
1159 			enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1160 
1161 		if (!is_leftmost(curr, &rq->dl))
1162 			resched_curr(rq);
1163 	}
1164 
1165 	/*
1166 	 * Because -- for now -- we share the rt bandwidth, we need to
1167 	 * account our runtime there too, otherwise actual rt tasks
1168 	 * would be able to exceed the shared quota.
1169 	 *
1170 	 * Account to the root rt group for now.
1171 	 *
1172 	 * The solution we're working towards is having the RT groups scheduled
1173 	 * using deadline servers -- however there's a few nasties to figure
1174 	 * out before that can happen.
1175 	 */
1176 	if (rt_bandwidth_enabled()) {
1177 		struct rt_rq *rt_rq = &rq->rt;
1178 
1179 		raw_spin_lock(&rt_rq->rt_runtime_lock);
1180 		/*
1181 		 * We'll let actual RT tasks worry about the overflow here, we
1182 		 * have our own CBS to keep us inline; only account when RT
1183 		 * bandwidth is relevant.
1184 		 */
1185 		if (sched_rt_bandwidth_account(rt_rq))
1186 			rt_rq->rt_time += delta_exec;
1187 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
1188 	}
1189 }
1190 
1191 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1192 {
1193 	struct sched_dl_entity *dl_se = container_of(timer,
1194 						     struct sched_dl_entity,
1195 						     inactive_timer);
1196 	struct task_struct *p = dl_task_of(dl_se);
1197 	struct rq_flags rf;
1198 	struct rq *rq;
1199 
1200 	rq = task_rq_lock(p, &rf);
1201 
1202 	if (!dl_task(p) || p->state == TASK_DEAD) {
1203 		struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1204 
1205 		if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1206 			sub_running_bw(p->dl.dl_bw, dl_rq_of_se(&p->dl));
1207 			sub_rq_bw(p->dl.dl_bw, dl_rq_of_se(&p->dl));
1208 			dl_se->dl_non_contending = 0;
1209 		}
1210 
1211 		raw_spin_lock(&dl_b->lock);
1212 		__dl_clear(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1213 		raw_spin_unlock(&dl_b->lock);
1214 		__dl_clear_params(p);
1215 
1216 		goto unlock;
1217 	}
1218 	if (dl_se->dl_non_contending == 0)
1219 		goto unlock;
1220 
1221 	sched_clock_tick();
1222 	update_rq_clock(rq);
1223 
1224 	sub_running_bw(dl_se->dl_bw, &rq->dl);
1225 	dl_se->dl_non_contending = 0;
1226 unlock:
1227 	task_rq_unlock(rq, p, &rf);
1228 	put_task_struct(p);
1229 
1230 	return HRTIMER_NORESTART;
1231 }
1232 
1233 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1234 {
1235 	struct hrtimer *timer = &dl_se->inactive_timer;
1236 
1237 	hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1238 	timer->function = inactive_task_timer;
1239 }
1240 
1241 #ifdef CONFIG_SMP
1242 
1243 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1244 {
1245 	struct rq *rq = rq_of_dl_rq(dl_rq);
1246 
1247 	if (dl_rq->earliest_dl.curr == 0 ||
1248 	    dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1249 		dl_rq->earliest_dl.curr = deadline;
1250 		cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1251 	}
1252 }
1253 
1254 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1255 {
1256 	struct rq *rq = rq_of_dl_rq(dl_rq);
1257 
1258 	/*
1259 	 * Since we may have removed our earliest (and/or next earliest)
1260 	 * task we must recompute them.
1261 	 */
1262 	if (!dl_rq->dl_nr_running) {
1263 		dl_rq->earliest_dl.curr = 0;
1264 		dl_rq->earliest_dl.next = 0;
1265 		cpudl_clear(&rq->rd->cpudl, rq->cpu);
1266 	} else {
1267 		struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1268 		struct sched_dl_entity *entry;
1269 
1270 		entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1271 		dl_rq->earliest_dl.curr = entry->deadline;
1272 		cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1273 	}
1274 }
1275 
1276 #else
1277 
1278 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1279 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1280 
1281 #endif /* CONFIG_SMP */
1282 
1283 static inline
1284 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1285 {
1286 	int prio = dl_task_of(dl_se)->prio;
1287 	u64 deadline = dl_se->deadline;
1288 
1289 	WARN_ON(!dl_prio(prio));
1290 	dl_rq->dl_nr_running++;
1291 	add_nr_running(rq_of_dl_rq(dl_rq), 1);
1292 
1293 	inc_dl_deadline(dl_rq, deadline);
1294 	inc_dl_migration(dl_se, dl_rq);
1295 }
1296 
1297 static inline
1298 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1299 {
1300 	int prio = dl_task_of(dl_se)->prio;
1301 
1302 	WARN_ON(!dl_prio(prio));
1303 	WARN_ON(!dl_rq->dl_nr_running);
1304 	dl_rq->dl_nr_running--;
1305 	sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1306 
1307 	dec_dl_deadline(dl_rq, dl_se->deadline);
1308 	dec_dl_migration(dl_se, dl_rq);
1309 }
1310 
1311 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1312 {
1313 	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1314 	struct rb_node **link = &dl_rq->root.rb_root.rb_node;
1315 	struct rb_node *parent = NULL;
1316 	struct sched_dl_entity *entry;
1317 	int leftmost = 1;
1318 
1319 	BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1320 
1321 	while (*link) {
1322 		parent = *link;
1323 		entry = rb_entry(parent, struct sched_dl_entity, rb_node);
1324 		if (dl_time_before(dl_se->deadline, entry->deadline))
1325 			link = &parent->rb_left;
1326 		else {
1327 			link = &parent->rb_right;
1328 			leftmost = 0;
1329 		}
1330 	}
1331 
1332 	rb_link_node(&dl_se->rb_node, parent, link);
1333 	rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost);
1334 
1335 	inc_dl_tasks(dl_se, dl_rq);
1336 }
1337 
1338 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1339 {
1340 	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1341 
1342 	if (RB_EMPTY_NODE(&dl_se->rb_node))
1343 		return;
1344 
1345 	rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1346 	RB_CLEAR_NODE(&dl_se->rb_node);
1347 
1348 	dec_dl_tasks(dl_se, dl_rq);
1349 }
1350 
1351 static void
1352 enqueue_dl_entity(struct sched_dl_entity *dl_se,
1353 		  struct sched_dl_entity *pi_se, int flags)
1354 {
1355 	BUG_ON(on_dl_rq(dl_se));
1356 
1357 	/*
1358 	 * If this is a wakeup or a new instance, the scheduling
1359 	 * parameters of the task might need updating. Otherwise,
1360 	 * we want a replenishment of its runtime.
1361 	 */
1362 	if (flags & ENQUEUE_WAKEUP) {
1363 		task_contending(dl_se, flags);
1364 		update_dl_entity(dl_se, pi_se);
1365 	} else if (flags & ENQUEUE_REPLENISH) {
1366 		replenish_dl_entity(dl_se, pi_se);
1367 	}
1368 
1369 	__enqueue_dl_entity(dl_se);
1370 }
1371 
1372 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1373 {
1374 	__dequeue_dl_entity(dl_se);
1375 }
1376 
1377 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1378 {
1379 	struct task_struct *pi_task = rt_mutex_get_top_task(p);
1380 	struct sched_dl_entity *pi_se = &p->dl;
1381 
1382 	/*
1383 	 * Use the scheduling parameters of the top pi-waiter task if:
1384 	 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1385 	 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1386 	 *   smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1387 	 *   boosted due to a SCHED_DEADLINE pi-waiter).
1388 	 * Otherwise we keep our runtime and deadline.
1389 	 */
1390 	if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) {
1391 		pi_se = &pi_task->dl;
1392 	} else if (!dl_prio(p->normal_prio)) {
1393 		/*
1394 		 * Special case in which we have a !SCHED_DEADLINE task
1395 		 * that is going to be deboosted, but exceeds its
1396 		 * runtime while doing so. No point in replenishing
1397 		 * it, as it's going to return back to its original
1398 		 * scheduling class after this.
1399 		 */
1400 		BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
1401 		return;
1402 	}
1403 
1404 	/*
1405 	 * Check if a constrained deadline task was activated
1406 	 * after the deadline but before the next period.
1407 	 * If that is the case, the task will be throttled and
1408 	 * the replenishment timer will be set to the next period.
1409 	 */
1410 	if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1411 		dl_check_constrained_dl(&p->dl);
1412 
1413 	if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1414 		add_rq_bw(p->dl.dl_bw, &rq->dl);
1415 		add_running_bw(p->dl.dl_bw, &rq->dl);
1416 	}
1417 
1418 	/*
1419 	 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1420 	 * its budget it needs a replenishment and, since it now is on
1421 	 * its rq, the bandwidth timer callback (which clearly has not
1422 	 * run yet) will take care of this.
1423 	 * However, the active utilization does not depend on the fact
1424 	 * that the task is on the runqueue or not (but depends on the
1425 	 * task's state - in GRUB parlance, "inactive" vs "active contending").
1426 	 * In other words, even if a task is throttled its utilization must
1427 	 * be counted in the active utilization; hence, we need to call
1428 	 * add_running_bw().
1429 	 */
1430 	if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1431 		if (flags & ENQUEUE_WAKEUP)
1432 			task_contending(&p->dl, flags);
1433 
1434 		return;
1435 	}
1436 
1437 	enqueue_dl_entity(&p->dl, pi_se, flags);
1438 
1439 	if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1440 		enqueue_pushable_dl_task(rq, p);
1441 }
1442 
1443 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1444 {
1445 	dequeue_dl_entity(&p->dl);
1446 	dequeue_pushable_dl_task(rq, p);
1447 }
1448 
1449 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1450 {
1451 	update_curr_dl(rq);
1452 	__dequeue_task_dl(rq, p, flags);
1453 
1454 	if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1455 		sub_running_bw(p->dl.dl_bw, &rq->dl);
1456 		sub_rq_bw(p->dl.dl_bw, &rq->dl);
1457 	}
1458 
1459 	/*
1460 	 * This check allows to start the inactive timer (or to immediately
1461 	 * decrease the active utilization, if needed) in two cases:
1462 	 * when the task blocks and when it is terminating
1463 	 * (p->state == TASK_DEAD). We can handle the two cases in the same
1464 	 * way, because from GRUB's point of view the same thing is happening
1465 	 * (the task moves from "active contending" to "active non contending"
1466 	 * or "inactive")
1467 	 */
1468 	if (flags & DEQUEUE_SLEEP)
1469 		task_non_contending(p);
1470 }
1471 
1472 /*
1473  * Yield task semantic for -deadline tasks is:
1474  *
1475  *   get off from the CPU until our next instance, with
1476  *   a new runtime. This is of little use now, since we
1477  *   don't have a bandwidth reclaiming mechanism. Anyway,
1478  *   bandwidth reclaiming is planned for the future, and
1479  *   yield_task_dl will indicate that some spare budget
1480  *   is available for other task instances to use it.
1481  */
1482 static void yield_task_dl(struct rq *rq)
1483 {
1484 	/*
1485 	 * We make the task go to sleep until its current deadline by
1486 	 * forcing its runtime to zero. This way, update_curr_dl() stops
1487 	 * it and the bandwidth timer will wake it up and will give it
1488 	 * new scheduling parameters (thanks to dl_yielded=1).
1489 	 */
1490 	rq->curr->dl.dl_yielded = 1;
1491 
1492 	update_rq_clock(rq);
1493 	update_curr_dl(rq);
1494 	/*
1495 	 * Tell update_rq_clock() that we've just updated,
1496 	 * so we don't do microscopic update in schedule()
1497 	 * and double the fastpath cost.
1498 	 */
1499 	rq_clock_skip_update(rq, true);
1500 }
1501 
1502 #ifdef CONFIG_SMP
1503 
1504 static int find_later_rq(struct task_struct *task);
1505 
1506 static int
1507 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1508 {
1509 	struct task_struct *curr;
1510 	struct rq *rq;
1511 
1512 	if (sd_flag != SD_BALANCE_WAKE)
1513 		goto out;
1514 
1515 	rq = cpu_rq(cpu);
1516 
1517 	rcu_read_lock();
1518 	curr = READ_ONCE(rq->curr); /* unlocked access */
1519 
1520 	/*
1521 	 * If we are dealing with a -deadline task, we must
1522 	 * decide where to wake it up.
1523 	 * If it has a later deadline and the current task
1524 	 * on this rq can't move (provided the waking task
1525 	 * can!) we prefer to send it somewhere else. On the
1526 	 * other hand, if it has a shorter deadline, we
1527 	 * try to make it stay here, it might be important.
1528 	 */
1529 	if (unlikely(dl_task(curr)) &&
1530 	    (curr->nr_cpus_allowed < 2 ||
1531 	     !dl_entity_preempt(&p->dl, &curr->dl)) &&
1532 	    (p->nr_cpus_allowed > 1)) {
1533 		int target = find_later_rq(p);
1534 
1535 		if (target != -1 &&
1536 				(dl_time_before(p->dl.deadline,
1537 					cpu_rq(target)->dl.earliest_dl.curr) ||
1538 				(cpu_rq(target)->dl.dl_nr_running == 0)))
1539 			cpu = target;
1540 	}
1541 	rcu_read_unlock();
1542 
1543 out:
1544 	return cpu;
1545 }
1546 
1547 static void migrate_task_rq_dl(struct task_struct *p)
1548 {
1549 	struct rq *rq;
1550 
1551 	if (p->state != TASK_WAKING)
1552 		return;
1553 
1554 	rq = task_rq(p);
1555 	/*
1556 	 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1557 	 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1558 	 * rq->lock is not... So, lock it
1559 	 */
1560 	raw_spin_lock(&rq->lock);
1561 	if (p->dl.dl_non_contending) {
1562 		sub_running_bw(p->dl.dl_bw, &rq->dl);
1563 		p->dl.dl_non_contending = 0;
1564 		/*
1565 		 * If the timer handler is currently running and the
1566 		 * timer cannot be cancelled, inactive_task_timer()
1567 		 * will see that dl_not_contending is not set, and
1568 		 * will not touch the rq's active utilization,
1569 		 * so we are still safe.
1570 		 */
1571 		if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1572 			put_task_struct(p);
1573 	}
1574 	sub_rq_bw(p->dl.dl_bw, &rq->dl);
1575 	raw_spin_unlock(&rq->lock);
1576 }
1577 
1578 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1579 {
1580 	/*
1581 	 * Current can't be migrated, useless to reschedule,
1582 	 * let's hope p can move out.
1583 	 */
1584 	if (rq->curr->nr_cpus_allowed == 1 ||
1585 	    !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1586 		return;
1587 
1588 	/*
1589 	 * p is migratable, so let's not schedule it and
1590 	 * see if it is pushed or pulled somewhere else.
1591 	 */
1592 	if (p->nr_cpus_allowed != 1 &&
1593 	    cpudl_find(&rq->rd->cpudl, p, NULL))
1594 		return;
1595 
1596 	resched_curr(rq);
1597 }
1598 
1599 #endif /* CONFIG_SMP */
1600 
1601 /*
1602  * Only called when both the current and waking task are -deadline
1603  * tasks.
1604  */
1605 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1606 				  int flags)
1607 {
1608 	if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1609 		resched_curr(rq);
1610 		return;
1611 	}
1612 
1613 #ifdef CONFIG_SMP
1614 	/*
1615 	 * In the unlikely case current and p have the same deadline
1616 	 * let us try to decide what's the best thing to do...
1617 	 */
1618 	if ((p->dl.deadline == rq->curr->dl.deadline) &&
1619 	    !test_tsk_need_resched(rq->curr))
1620 		check_preempt_equal_dl(rq, p);
1621 #endif /* CONFIG_SMP */
1622 }
1623 
1624 #ifdef CONFIG_SCHED_HRTICK
1625 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1626 {
1627 	hrtick_start(rq, p->dl.runtime);
1628 }
1629 #else /* !CONFIG_SCHED_HRTICK */
1630 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1631 {
1632 }
1633 #endif
1634 
1635 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1636 						   struct dl_rq *dl_rq)
1637 {
1638 	struct rb_node *left = rb_first_cached(&dl_rq->root);
1639 
1640 	if (!left)
1641 		return NULL;
1642 
1643 	return rb_entry(left, struct sched_dl_entity, rb_node);
1644 }
1645 
1646 static struct task_struct *
1647 pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1648 {
1649 	struct sched_dl_entity *dl_se;
1650 	struct task_struct *p;
1651 	struct dl_rq *dl_rq;
1652 
1653 	dl_rq = &rq->dl;
1654 
1655 	if (need_pull_dl_task(rq, prev)) {
1656 		/*
1657 		 * This is OK, because current is on_cpu, which avoids it being
1658 		 * picked for load-balance and preemption/IRQs are still
1659 		 * disabled avoiding further scheduler activity on it and we're
1660 		 * being very careful to re-start the picking loop.
1661 		 */
1662 		rq_unpin_lock(rq, rf);
1663 		pull_dl_task(rq);
1664 		rq_repin_lock(rq, rf);
1665 		/*
1666 		 * pull_dl_task() can drop (and re-acquire) rq->lock; this
1667 		 * means a stop task can slip in, in which case we need to
1668 		 * re-start task selection.
1669 		 */
1670 		if (rq->stop && task_on_rq_queued(rq->stop))
1671 			return RETRY_TASK;
1672 	}
1673 
1674 	/*
1675 	 * When prev is DL, we may throttle it in put_prev_task().
1676 	 * So, we update time before we check for dl_nr_running.
1677 	 */
1678 	if (prev->sched_class == &dl_sched_class)
1679 		update_curr_dl(rq);
1680 
1681 	if (unlikely(!dl_rq->dl_nr_running))
1682 		return NULL;
1683 
1684 	put_prev_task(rq, prev);
1685 
1686 	dl_se = pick_next_dl_entity(rq, dl_rq);
1687 	BUG_ON(!dl_se);
1688 
1689 	p = dl_task_of(dl_se);
1690 	p->se.exec_start = rq_clock_task(rq);
1691 
1692 	/* Running task will never be pushed. */
1693        dequeue_pushable_dl_task(rq, p);
1694 
1695 	if (hrtick_enabled(rq))
1696 		start_hrtick_dl(rq, p);
1697 
1698 	queue_push_tasks(rq);
1699 
1700 	return p;
1701 }
1702 
1703 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1704 {
1705 	update_curr_dl(rq);
1706 
1707 	if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1708 		enqueue_pushable_dl_task(rq, p);
1709 }
1710 
1711 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1712 {
1713 	update_curr_dl(rq);
1714 
1715 	/*
1716 	 * Even when we have runtime, update_curr_dl() might have resulted in us
1717 	 * not being the leftmost task anymore. In that case NEED_RESCHED will
1718 	 * be set and schedule() will start a new hrtick for the next task.
1719 	 */
1720 	if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1721 	    is_leftmost(p, &rq->dl))
1722 		start_hrtick_dl(rq, p);
1723 }
1724 
1725 static void task_fork_dl(struct task_struct *p)
1726 {
1727 	/*
1728 	 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1729 	 * sched_fork()
1730 	 */
1731 }
1732 
1733 static void set_curr_task_dl(struct rq *rq)
1734 {
1735 	struct task_struct *p = rq->curr;
1736 
1737 	p->se.exec_start = rq_clock_task(rq);
1738 
1739 	/* You can't push away the running task */
1740 	dequeue_pushable_dl_task(rq, p);
1741 }
1742 
1743 #ifdef CONFIG_SMP
1744 
1745 /* Only try algorithms three times */
1746 #define DL_MAX_TRIES 3
1747 
1748 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1749 {
1750 	if (!task_running(rq, p) &&
1751 	    cpumask_test_cpu(cpu, &p->cpus_allowed))
1752 		return 1;
1753 	return 0;
1754 }
1755 
1756 /*
1757  * Return the earliest pushable rq's task, which is suitable to be executed
1758  * on the CPU, NULL otherwise:
1759  */
1760 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1761 {
1762 	struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
1763 	struct task_struct *p = NULL;
1764 
1765 	if (!has_pushable_dl_tasks(rq))
1766 		return NULL;
1767 
1768 next_node:
1769 	if (next_node) {
1770 		p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1771 
1772 		if (pick_dl_task(rq, p, cpu))
1773 			return p;
1774 
1775 		next_node = rb_next(next_node);
1776 		goto next_node;
1777 	}
1778 
1779 	return NULL;
1780 }
1781 
1782 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1783 
1784 static int find_later_rq(struct task_struct *task)
1785 {
1786 	struct sched_domain *sd;
1787 	struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1788 	int this_cpu = smp_processor_id();
1789 	int cpu = task_cpu(task);
1790 
1791 	/* Make sure the mask is initialized first */
1792 	if (unlikely(!later_mask))
1793 		return -1;
1794 
1795 	if (task->nr_cpus_allowed == 1)
1796 		return -1;
1797 
1798 	/*
1799 	 * We have to consider system topology and task affinity
1800 	 * first, then we can look for a suitable cpu.
1801 	 */
1802 	if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1803 		return -1;
1804 
1805 	/*
1806 	 * If we are here, some targets have been found, including
1807 	 * the most suitable which is, among the runqueues where the
1808 	 * current tasks have later deadlines than the task's one, the
1809 	 * rq with the latest possible one.
1810 	 *
1811 	 * Now we check how well this matches with task's
1812 	 * affinity and system topology.
1813 	 *
1814 	 * The last cpu where the task run is our first
1815 	 * guess, since it is most likely cache-hot there.
1816 	 */
1817 	if (cpumask_test_cpu(cpu, later_mask))
1818 		return cpu;
1819 	/*
1820 	 * Check if this_cpu is to be skipped (i.e., it is
1821 	 * not in the mask) or not.
1822 	 */
1823 	if (!cpumask_test_cpu(this_cpu, later_mask))
1824 		this_cpu = -1;
1825 
1826 	rcu_read_lock();
1827 	for_each_domain(cpu, sd) {
1828 		if (sd->flags & SD_WAKE_AFFINE) {
1829 			int best_cpu;
1830 
1831 			/*
1832 			 * If possible, preempting this_cpu is
1833 			 * cheaper than migrating.
1834 			 */
1835 			if (this_cpu != -1 &&
1836 			    cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1837 				rcu_read_unlock();
1838 				return this_cpu;
1839 			}
1840 
1841 			best_cpu = cpumask_first_and(later_mask,
1842 							sched_domain_span(sd));
1843 			/*
1844 			 * Last chance: if a cpu being in both later_mask
1845 			 * and current sd span is valid, that becomes our
1846 			 * choice. Of course, the latest possible cpu is
1847 			 * already under consideration through later_mask.
1848 			 */
1849 			if (best_cpu < nr_cpu_ids) {
1850 				rcu_read_unlock();
1851 				return best_cpu;
1852 			}
1853 		}
1854 	}
1855 	rcu_read_unlock();
1856 
1857 	/*
1858 	 * At this point, all our guesses failed, we just return
1859 	 * 'something', and let the caller sort the things out.
1860 	 */
1861 	if (this_cpu != -1)
1862 		return this_cpu;
1863 
1864 	cpu = cpumask_any(later_mask);
1865 	if (cpu < nr_cpu_ids)
1866 		return cpu;
1867 
1868 	return -1;
1869 }
1870 
1871 /* Locks the rq it finds */
1872 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1873 {
1874 	struct rq *later_rq = NULL;
1875 	int tries;
1876 	int cpu;
1877 
1878 	for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1879 		cpu = find_later_rq(task);
1880 
1881 		if ((cpu == -1) || (cpu == rq->cpu))
1882 			break;
1883 
1884 		later_rq = cpu_rq(cpu);
1885 
1886 		if (later_rq->dl.dl_nr_running &&
1887 		    !dl_time_before(task->dl.deadline,
1888 					later_rq->dl.earliest_dl.curr)) {
1889 			/*
1890 			 * Target rq has tasks of equal or earlier deadline,
1891 			 * retrying does not release any lock and is unlikely
1892 			 * to yield a different result.
1893 			 */
1894 			later_rq = NULL;
1895 			break;
1896 		}
1897 
1898 		/* Retry if something changed. */
1899 		if (double_lock_balance(rq, later_rq)) {
1900 			if (unlikely(task_rq(task) != rq ||
1901 				     !cpumask_test_cpu(later_rq->cpu, &task->cpus_allowed) ||
1902 				     task_running(rq, task) ||
1903 				     !dl_task(task) ||
1904 				     !task_on_rq_queued(task))) {
1905 				double_unlock_balance(rq, later_rq);
1906 				later_rq = NULL;
1907 				break;
1908 			}
1909 		}
1910 
1911 		/*
1912 		 * If the rq we found has no -deadline task, or
1913 		 * its earliest one has a later deadline than our
1914 		 * task, the rq is a good one.
1915 		 */
1916 		if (!later_rq->dl.dl_nr_running ||
1917 		    dl_time_before(task->dl.deadline,
1918 				   later_rq->dl.earliest_dl.curr))
1919 			break;
1920 
1921 		/* Otherwise we try again. */
1922 		double_unlock_balance(rq, later_rq);
1923 		later_rq = NULL;
1924 	}
1925 
1926 	return later_rq;
1927 }
1928 
1929 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1930 {
1931 	struct task_struct *p;
1932 
1933 	if (!has_pushable_dl_tasks(rq))
1934 		return NULL;
1935 
1936 	p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
1937 		     struct task_struct, pushable_dl_tasks);
1938 
1939 	BUG_ON(rq->cpu != task_cpu(p));
1940 	BUG_ON(task_current(rq, p));
1941 	BUG_ON(p->nr_cpus_allowed <= 1);
1942 
1943 	BUG_ON(!task_on_rq_queued(p));
1944 	BUG_ON(!dl_task(p));
1945 
1946 	return p;
1947 }
1948 
1949 /*
1950  * See if the non running -deadline tasks on this rq
1951  * can be sent to some other CPU where they can preempt
1952  * and start executing.
1953  */
1954 static int push_dl_task(struct rq *rq)
1955 {
1956 	struct task_struct *next_task;
1957 	struct rq *later_rq;
1958 	int ret = 0;
1959 
1960 	if (!rq->dl.overloaded)
1961 		return 0;
1962 
1963 	next_task = pick_next_pushable_dl_task(rq);
1964 	if (!next_task)
1965 		return 0;
1966 
1967 retry:
1968 	if (unlikely(next_task == rq->curr)) {
1969 		WARN_ON(1);
1970 		return 0;
1971 	}
1972 
1973 	/*
1974 	 * If next_task preempts rq->curr, and rq->curr
1975 	 * can move away, it makes sense to just reschedule
1976 	 * without going further in pushing next_task.
1977 	 */
1978 	if (dl_task(rq->curr) &&
1979 	    dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1980 	    rq->curr->nr_cpus_allowed > 1) {
1981 		resched_curr(rq);
1982 		return 0;
1983 	}
1984 
1985 	/* We might release rq lock */
1986 	get_task_struct(next_task);
1987 
1988 	/* Will lock the rq it'll find */
1989 	later_rq = find_lock_later_rq(next_task, rq);
1990 	if (!later_rq) {
1991 		struct task_struct *task;
1992 
1993 		/*
1994 		 * We must check all this again, since
1995 		 * find_lock_later_rq releases rq->lock and it is
1996 		 * then possible that next_task has migrated.
1997 		 */
1998 		task = pick_next_pushable_dl_task(rq);
1999 		if (task == next_task) {
2000 			/*
2001 			 * The task is still there. We don't try
2002 			 * again, some other cpu will pull it when ready.
2003 			 */
2004 			goto out;
2005 		}
2006 
2007 		if (!task)
2008 			/* No more tasks */
2009 			goto out;
2010 
2011 		put_task_struct(next_task);
2012 		next_task = task;
2013 		goto retry;
2014 	}
2015 
2016 	deactivate_task(rq, next_task, 0);
2017 	sub_running_bw(next_task->dl.dl_bw, &rq->dl);
2018 	sub_rq_bw(next_task->dl.dl_bw, &rq->dl);
2019 	set_task_cpu(next_task, later_rq->cpu);
2020 	add_rq_bw(next_task->dl.dl_bw, &later_rq->dl);
2021 	add_running_bw(next_task->dl.dl_bw, &later_rq->dl);
2022 	activate_task(later_rq, next_task, 0);
2023 	ret = 1;
2024 
2025 	resched_curr(later_rq);
2026 
2027 	double_unlock_balance(rq, later_rq);
2028 
2029 out:
2030 	put_task_struct(next_task);
2031 
2032 	return ret;
2033 }
2034 
2035 static void push_dl_tasks(struct rq *rq)
2036 {
2037 	/* push_dl_task() will return true if it moved a -deadline task */
2038 	while (push_dl_task(rq))
2039 		;
2040 }
2041 
2042 static void pull_dl_task(struct rq *this_rq)
2043 {
2044 	int this_cpu = this_rq->cpu, cpu;
2045 	struct task_struct *p;
2046 	bool resched = false;
2047 	struct rq *src_rq;
2048 	u64 dmin = LONG_MAX;
2049 
2050 	if (likely(!dl_overloaded(this_rq)))
2051 		return;
2052 
2053 	/*
2054 	 * Match the barrier from dl_set_overloaded; this guarantees that if we
2055 	 * see overloaded we must also see the dlo_mask bit.
2056 	 */
2057 	smp_rmb();
2058 
2059 	for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2060 		if (this_cpu == cpu)
2061 			continue;
2062 
2063 		src_rq = cpu_rq(cpu);
2064 
2065 		/*
2066 		 * It looks racy, abd it is! However, as in sched_rt.c,
2067 		 * we are fine with this.
2068 		 */
2069 		if (this_rq->dl.dl_nr_running &&
2070 		    dl_time_before(this_rq->dl.earliest_dl.curr,
2071 				   src_rq->dl.earliest_dl.next))
2072 			continue;
2073 
2074 		/* Might drop this_rq->lock */
2075 		double_lock_balance(this_rq, src_rq);
2076 
2077 		/*
2078 		 * If there are no more pullable tasks on the
2079 		 * rq, we're done with it.
2080 		 */
2081 		if (src_rq->dl.dl_nr_running <= 1)
2082 			goto skip;
2083 
2084 		p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2085 
2086 		/*
2087 		 * We found a task to be pulled if:
2088 		 *  - it preempts our current (if there's one),
2089 		 *  - it will preempt the last one we pulled (if any).
2090 		 */
2091 		if (p && dl_time_before(p->dl.deadline, dmin) &&
2092 		    (!this_rq->dl.dl_nr_running ||
2093 		     dl_time_before(p->dl.deadline,
2094 				    this_rq->dl.earliest_dl.curr))) {
2095 			WARN_ON(p == src_rq->curr);
2096 			WARN_ON(!task_on_rq_queued(p));
2097 
2098 			/*
2099 			 * Then we pull iff p has actually an earlier
2100 			 * deadline than the current task of its runqueue.
2101 			 */
2102 			if (dl_time_before(p->dl.deadline,
2103 					   src_rq->curr->dl.deadline))
2104 				goto skip;
2105 
2106 			resched = true;
2107 
2108 			deactivate_task(src_rq, p, 0);
2109 			sub_running_bw(p->dl.dl_bw, &src_rq->dl);
2110 			sub_rq_bw(p->dl.dl_bw, &src_rq->dl);
2111 			set_task_cpu(p, this_cpu);
2112 			add_rq_bw(p->dl.dl_bw, &this_rq->dl);
2113 			add_running_bw(p->dl.dl_bw, &this_rq->dl);
2114 			activate_task(this_rq, p, 0);
2115 			dmin = p->dl.deadline;
2116 
2117 			/* Is there any other task even earlier? */
2118 		}
2119 skip:
2120 		double_unlock_balance(this_rq, src_rq);
2121 	}
2122 
2123 	if (resched)
2124 		resched_curr(this_rq);
2125 }
2126 
2127 /*
2128  * Since the task is not running and a reschedule is not going to happen
2129  * anytime soon on its runqueue, we try pushing it away now.
2130  */
2131 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2132 {
2133 	if (!task_running(rq, p) &&
2134 	    !test_tsk_need_resched(rq->curr) &&
2135 	    p->nr_cpus_allowed > 1 &&
2136 	    dl_task(rq->curr) &&
2137 	    (rq->curr->nr_cpus_allowed < 2 ||
2138 	     !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2139 		push_dl_tasks(rq);
2140 	}
2141 }
2142 
2143 static void set_cpus_allowed_dl(struct task_struct *p,
2144 				const struct cpumask *new_mask)
2145 {
2146 	struct root_domain *src_rd;
2147 	struct rq *rq;
2148 
2149 	BUG_ON(!dl_task(p));
2150 
2151 	rq = task_rq(p);
2152 	src_rd = rq->rd;
2153 	/*
2154 	 * Migrating a SCHED_DEADLINE task between exclusive
2155 	 * cpusets (different root_domains) entails a bandwidth
2156 	 * update. We already made space for us in the destination
2157 	 * domain (see cpuset_can_attach()).
2158 	 */
2159 	if (!cpumask_intersects(src_rd->span, new_mask)) {
2160 		struct dl_bw *src_dl_b;
2161 
2162 		src_dl_b = dl_bw_of(cpu_of(rq));
2163 		/*
2164 		 * We now free resources of the root_domain we are migrating
2165 		 * off. In the worst case, sched_setattr() may temporary fail
2166 		 * until we complete the update.
2167 		 */
2168 		raw_spin_lock(&src_dl_b->lock);
2169 		__dl_clear(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2170 		raw_spin_unlock(&src_dl_b->lock);
2171 	}
2172 
2173 	set_cpus_allowed_common(p, new_mask);
2174 }
2175 
2176 /* Assumes rq->lock is held */
2177 static void rq_online_dl(struct rq *rq)
2178 {
2179 	if (rq->dl.overloaded)
2180 		dl_set_overload(rq);
2181 
2182 	cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2183 	if (rq->dl.dl_nr_running > 0)
2184 		cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2185 }
2186 
2187 /* Assumes rq->lock is held */
2188 static void rq_offline_dl(struct rq *rq)
2189 {
2190 	if (rq->dl.overloaded)
2191 		dl_clear_overload(rq);
2192 
2193 	cpudl_clear(&rq->rd->cpudl, rq->cpu);
2194 	cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2195 }
2196 
2197 void __init init_sched_dl_class(void)
2198 {
2199 	unsigned int i;
2200 
2201 	for_each_possible_cpu(i)
2202 		zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2203 					GFP_KERNEL, cpu_to_node(i));
2204 }
2205 
2206 #endif /* CONFIG_SMP */
2207 
2208 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2209 {
2210 	/*
2211 	 * task_non_contending() can start the "inactive timer" (if the 0-lag
2212 	 * time is in the future). If the task switches back to dl before
2213 	 * the "inactive timer" fires, it can continue to consume its current
2214 	 * runtime using its current deadline. If it stays outside of
2215 	 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2216 	 * will reset the task parameters.
2217 	 */
2218 	if (task_on_rq_queued(p) && p->dl.dl_runtime)
2219 		task_non_contending(p);
2220 
2221 	if (!task_on_rq_queued(p))
2222 		sub_rq_bw(p->dl.dl_bw, &rq->dl);
2223 
2224 	/*
2225 	 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2226 	 * at the 0-lag time, because the task could have been migrated
2227 	 * while SCHED_OTHER in the meanwhile.
2228 	 */
2229 	if (p->dl.dl_non_contending)
2230 		p->dl.dl_non_contending = 0;
2231 
2232 	/*
2233 	 * Since this might be the only -deadline task on the rq,
2234 	 * this is the right place to try to pull some other one
2235 	 * from an overloaded cpu, if any.
2236 	 */
2237 	if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2238 		return;
2239 
2240 	queue_pull_task(rq);
2241 }
2242 
2243 /*
2244  * When switching to -deadline, we may overload the rq, then
2245  * we try to push someone off, if possible.
2246  */
2247 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2248 {
2249 	if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2250 		put_task_struct(p);
2251 
2252 	/* If p is not queued we will update its parameters at next wakeup. */
2253 	if (!task_on_rq_queued(p)) {
2254 		add_rq_bw(p->dl.dl_bw, &rq->dl);
2255 
2256 		return;
2257 	}
2258 	/*
2259 	 * If p is boosted we already updated its params in
2260 	 * rt_mutex_setprio()->enqueue_task(..., ENQUEUE_REPLENISH),
2261 	 * p's deadline being now already after rq_clock(rq).
2262 	 */
2263 	if (dl_time_before(p->dl.deadline, rq_clock(rq)))
2264 		setup_new_dl_entity(&p->dl);
2265 
2266 	if (rq->curr != p) {
2267 #ifdef CONFIG_SMP
2268 		if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2269 			queue_push_tasks(rq);
2270 #endif
2271 		if (dl_task(rq->curr))
2272 			check_preempt_curr_dl(rq, p, 0);
2273 		else
2274 			resched_curr(rq);
2275 	}
2276 }
2277 
2278 /*
2279  * If the scheduling parameters of a -deadline task changed,
2280  * a push or pull operation might be needed.
2281  */
2282 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2283 			    int oldprio)
2284 {
2285 	if (task_on_rq_queued(p) || rq->curr == p) {
2286 #ifdef CONFIG_SMP
2287 		/*
2288 		 * This might be too much, but unfortunately
2289 		 * we don't have the old deadline value, and
2290 		 * we can't argue if the task is increasing
2291 		 * or lowering its prio, so...
2292 		 */
2293 		if (!rq->dl.overloaded)
2294 			queue_pull_task(rq);
2295 
2296 		/*
2297 		 * If we now have a earlier deadline task than p,
2298 		 * then reschedule, provided p is still on this
2299 		 * runqueue.
2300 		 */
2301 		if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2302 			resched_curr(rq);
2303 #else
2304 		/*
2305 		 * Again, we don't know if p has a earlier
2306 		 * or later deadline, so let's blindly set a
2307 		 * (maybe not needed) rescheduling point.
2308 		 */
2309 		resched_curr(rq);
2310 #endif /* CONFIG_SMP */
2311 	}
2312 }
2313 
2314 const struct sched_class dl_sched_class = {
2315 	.next			= &rt_sched_class,
2316 	.enqueue_task		= enqueue_task_dl,
2317 	.dequeue_task		= dequeue_task_dl,
2318 	.yield_task		= yield_task_dl,
2319 
2320 	.check_preempt_curr	= check_preempt_curr_dl,
2321 
2322 	.pick_next_task		= pick_next_task_dl,
2323 	.put_prev_task		= put_prev_task_dl,
2324 
2325 #ifdef CONFIG_SMP
2326 	.select_task_rq		= select_task_rq_dl,
2327 	.migrate_task_rq	= migrate_task_rq_dl,
2328 	.set_cpus_allowed       = set_cpus_allowed_dl,
2329 	.rq_online              = rq_online_dl,
2330 	.rq_offline             = rq_offline_dl,
2331 	.task_woken		= task_woken_dl,
2332 #endif
2333 
2334 	.set_curr_task		= set_curr_task_dl,
2335 	.task_tick		= task_tick_dl,
2336 	.task_fork              = task_fork_dl,
2337 
2338 	.prio_changed           = prio_changed_dl,
2339 	.switched_from		= switched_from_dl,
2340 	.switched_to		= switched_to_dl,
2341 
2342 	.update_curr		= update_curr_dl,
2343 };
2344 
2345 int sched_dl_global_validate(void)
2346 {
2347 	u64 runtime = global_rt_runtime();
2348 	u64 period = global_rt_period();
2349 	u64 new_bw = to_ratio(period, runtime);
2350 	struct dl_bw *dl_b;
2351 	int cpu, ret = 0;
2352 	unsigned long flags;
2353 
2354 	/*
2355 	 * Here we want to check the bandwidth not being set to some
2356 	 * value smaller than the currently allocated bandwidth in
2357 	 * any of the root_domains.
2358 	 *
2359 	 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2360 	 * cycling on root_domains... Discussion on different/better
2361 	 * solutions is welcome!
2362 	 */
2363 	for_each_possible_cpu(cpu) {
2364 		rcu_read_lock_sched();
2365 		dl_b = dl_bw_of(cpu);
2366 
2367 		raw_spin_lock_irqsave(&dl_b->lock, flags);
2368 		if (new_bw < dl_b->total_bw)
2369 			ret = -EBUSY;
2370 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2371 
2372 		rcu_read_unlock_sched();
2373 
2374 		if (ret)
2375 			break;
2376 	}
2377 
2378 	return ret;
2379 }
2380 
2381 void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2382 {
2383 	if (global_rt_runtime() == RUNTIME_INF) {
2384 		dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2385 		dl_rq->extra_bw = 1 << BW_SHIFT;
2386 	} else {
2387 		dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2388 			  global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2389 		dl_rq->extra_bw = to_ratio(global_rt_period(),
2390 						    global_rt_runtime());
2391 	}
2392 }
2393 
2394 void sched_dl_do_global(void)
2395 {
2396 	u64 new_bw = -1;
2397 	struct dl_bw *dl_b;
2398 	int cpu;
2399 	unsigned long flags;
2400 
2401 	def_dl_bandwidth.dl_period = global_rt_period();
2402 	def_dl_bandwidth.dl_runtime = global_rt_runtime();
2403 
2404 	if (global_rt_runtime() != RUNTIME_INF)
2405 		new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2406 
2407 	/*
2408 	 * FIXME: As above...
2409 	 */
2410 	for_each_possible_cpu(cpu) {
2411 		rcu_read_lock_sched();
2412 		dl_b = dl_bw_of(cpu);
2413 
2414 		raw_spin_lock_irqsave(&dl_b->lock, flags);
2415 		dl_b->bw = new_bw;
2416 		raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2417 
2418 		rcu_read_unlock_sched();
2419 		init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2420 	}
2421 }
2422 
2423 /*
2424  * We must be sure that accepting a new task (or allowing changing the
2425  * parameters of an existing one) is consistent with the bandwidth
2426  * constraints. If yes, this function also accordingly updates the currently
2427  * allocated bandwidth to reflect the new situation.
2428  *
2429  * This function is called while holding p's rq->lock.
2430  */
2431 int sched_dl_overflow(struct task_struct *p, int policy,
2432 		      const struct sched_attr *attr)
2433 {
2434 	struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2435 	u64 period = attr->sched_period ?: attr->sched_deadline;
2436 	u64 runtime = attr->sched_runtime;
2437 	u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2438 	int cpus, err = -1;
2439 
2440 	/* !deadline task may carry old deadline bandwidth */
2441 	if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2442 		return 0;
2443 
2444 	/*
2445 	 * Either if a task, enters, leave, or stays -deadline but changes
2446 	 * its parameters, we may need to update accordingly the total
2447 	 * allocated bandwidth of the container.
2448 	 */
2449 	raw_spin_lock(&dl_b->lock);
2450 	cpus = dl_bw_cpus(task_cpu(p));
2451 	if (dl_policy(policy) && !task_has_dl_policy(p) &&
2452 	    !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2453 		if (hrtimer_active(&p->dl.inactive_timer))
2454 			__dl_clear(dl_b, p->dl.dl_bw, cpus);
2455 		__dl_add(dl_b, new_bw, cpus);
2456 		err = 0;
2457 	} else if (dl_policy(policy) && task_has_dl_policy(p) &&
2458 		   !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2459 		/*
2460 		 * XXX this is slightly incorrect: when the task
2461 		 * utilization decreases, we should delay the total
2462 		 * utilization change until the task's 0-lag point.
2463 		 * But this would require to set the task's "inactive
2464 		 * timer" when the task is not inactive.
2465 		 */
2466 		__dl_clear(dl_b, p->dl.dl_bw, cpus);
2467 		__dl_add(dl_b, new_bw, cpus);
2468 		dl_change_utilization(p, new_bw);
2469 		err = 0;
2470 	} else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2471 		/*
2472 		 * Do not decrease the total deadline utilization here,
2473 		 * switched_from_dl() will take care to do it at the correct
2474 		 * (0-lag) time.
2475 		 */
2476 		err = 0;
2477 	}
2478 	raw_spin_unlock(&dl_b->lock);
2479 
2480 	return err;
2481 }
2482 
2483 /*
2484  * This function initializes the sched_dl_entity of a newly becoming
2485  * SCHED_DEADLINE task.
2486  *
2487  * Only the static values are considered here, the actual runtime and the
2488  * absolute deadline will be properly calculated when the task is enqueued
2489  * for the first time with its new policy.
2490  */
2491 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2492 {
2493 	struct sched_dl_entity *dl_se = &p->dl;
2494 
2495 	dl_se->dl_runtime = attr->sched_runtime;
2496 	dl_se->dl_deadline = attr->sched_deadline;
2497 	dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2498 	dl_se->flags = attr->sched_flags;
2499 	dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2500 	dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2501 }
2502 
2503 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2504 {
2505 	struct sched_dl_entity *dl_se = &p->dl;
2506 
2507 	attr->sched_priority = p->rt_priority;
2508 	attr->sched_runtime = dl_se->dl_runtime;
2509 	attr->sched_deadline = dl_se->dl_deadline;
2510 	attr->sched_period = dl_se->dl_period;
2511 	attr->sched_flags = dl_se->flags;
2512 }
2513 
2514 /*
2515  * This function validates the new parameters of a -deadline task.
2516  * We ask for the deadline not being zero, and greater or equal
2517  * than the runtime, as well as the period of being zero or
2518  * greater than deadline. Furthermore, we have to be sure that
2519  * user parameters are above the internal resolution of 1us (we
2520  * check sched_runtime only since it is always the smaller one) and
2521  * below 2^63 ns (we have to check both sched_deadline and
2522  * sched_period, as the latter can be zero).
2523  */
2524 bool __checkparam_dl(const struct sched_attr *attr)
2525 {
2526 	/* deadline != 0 */
2527 	if (attr->sched_deadline == 0)
2528 		return false;
2529 
2530 	/*
2531 	 * Since we truncate DL_SCALE bits, make sure we're at least
2532 	 * that big.
2533 	 */
2534 	if (attr->sched_runtime < (1ULL << DL_SCALE))
2535 		return false;
2536 
2537 	/*
2538 	 * Since we use the MSB for wrap-around and sign issues, make
2539 	 * sure it's not set (mind that period can be equal to zero).
2540 	 */
2541 	if (attr->sched_deadline & (1ULL << 63) ||
2542 	    attr->sched_period & (1ULL << 63))
2543 		return false;
2544 
2545 	/* runtime <= deadline <= period (if period != 0) */
2546 	if ((attr->sched_period != 0 &&
2547 	     attr->sched_period < attr->sched_deadline) ||
2548 	    attr->sched_deadline < attr->sched_runtime)
2549 		return false;
2550 
2551 	return true;
2552 }
2553 
2554 /*
2555  * This function clears the sched_dl_entity static params.
2556  */
2557 void __dl_clear_params(struct task_struct *p)
2558 {
2559 	struct sched_dl_entity *dl_se = &p->dl;
2560 
2561 	dl_se->dl_runtime = 0;
2562 	dl_se->dl_deadline = 0;
2563 	dl_se->dl_period = 0;
2564 	dl_se->flags = 0;
2565 	dl_se->dl_bw = 0;
2566 	dl_se->dl_density = 0;
2567 
2568 	dl_se->dl_throttled = 0;
2569 	dl_se->dl_yielded = 0;
2570 	dl_se->dl_non_contending = 0;
2571 }
2572 
2573 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2574 {
2575 	struct sched_dl_entity *dl_se = &p->dl;
2576 
2577 	if (dl_se->dl_runtime != attr->sched_runtime ||
2578 	    dl_se->dl_deadline != attr->sched_deadline ||
2579 	    dl_se->dl_period != attr->sched_period ||
2580 	    dl_se->flags != attr->sched_flags)
2581 		return true;
2582 
2583 	return false;
2584 }
2585 
2586 #ifdef CONFIG_SMP
2587 int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2588 {
2589 	unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
2590 							cs_cpus_allowed);
2591 	struct dl_bw *dl_b;
2592 	bool overflow;
2593 	int cpus, ret;
2594 	unsigned long flags;
2595 
2596 	rcu_read_lock_sched();
2597 	dl_b = dl_bw_of(dest_cpu);
2598 	raw_spin_lock_irqsave(&dl_b->lock, flags);
2599 	cpus = dl_bw_cpus(dest_cpu);
2600 	overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
2601 	if (overflow)
2602 		ret = -EBUSY;
2603 	else {
2604 		/*
2605 		 * We reserve space for this task in the destination
2606 		 * root_domain, as we can't fail after this point.
2607 		 * We will free resources in the source root_domain
2608 		 * later on (see set_cpus_allowed_dl()).
2609 		 */
2610 		__dl_add(dl_b, p->dl.dl_bw, cpus);
2611 		ret = 0;
2612 	}
2613 	raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2614 	rcu_read_unlock_sched();
2615 	return ret;
2616 }
2617 
2618 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2619 				 const struct cpumask *trial)
2620 {
2621 	int ret = 1, trial_cpus;
2622 	struct dl_bw *cur_dl_b;
2623 	unsigned long flags;
2624 
2625 	rcu_read_lock_sched();
2626 	cur_dl_b = dl_bw_of(cpumask_any(cur));
2627 	trial_cpus = cpumask_weight(trial);
2628 
2629 	raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2630 	if (cur_dl_b->bw != -1 &&
2631 	    cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2632 		ret = 0;
2633 	raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2634 	rcu_read_unlock_sched();
2635 	return ret;
2636 }
2637 
2638 bool dl_cpu_busy(unsigned int cpu)
2639 {
2640 	unsigned long flags;
2641 	struct dl_bw *dl_b;
2642 	bool overflow;
2643 	int cpus;
2644 
2645 	rcu_read_lock_sched();
2646 	dl_b = dl_bw_of(cpu);
2647 	raw_spin_lock_irqsave(&dl_b->lock, flags);
2648 	cpus = dl_bw_cpus(cpu);
2649 	overflow = __dl_overflow(dl_b, cpus, 0, 0);
2650 	raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2651 	rcu_read_unlock_sched();
2652 	return overflow;
2653 }
2654 #endif
2655 
2656 #ifdef CONFIG_SCHED_DEBUG
2657 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2658 
2659 void print_dl_stats(struct seq_file *m, int cpu)
2660 {
2661 	print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2662 }
2663 #endif /* CONFIG_SCHED_DEBUG */
2664