Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Deadline Scheduling Class (SCHED_DEADLINE)
4 *
5 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6 *
7 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't affect any other task.
12 *
13 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fchecconi@gmail.com>
17 */
18#include "sched.h"
19
20struct dl_bandwidth def_dl_bandwidth;
21
22static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
23{
24 return container_of(dl_se, struct task_struct, dl);
25}
26
27static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
28{
29 return container_of(dl_rq, struct rq, dl);
30}
31
32static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
33{
34 struct task_struct *p = dl_task_of(dl_se);
35 struct rq *rq = task_rq(p);
36
37 return &rq->dl;
38}
39
40static inline int on_dl_rq(struct sched_dl_entity *dl_se)
41{
42 return !RB_EMPTY_NODE(&dl_se->rb_node);
43}
44
45#ifdef CONFIG_SMP
46static inline struct dl_bw *dl_bw_of(int i)
47{
48 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
49 "sched RCU must be held");
50 return &cpu_rq(i)->rd->dl_bw;
51}
52
53static inline int dl_bw_cpus(int i)
54{
55 struct root_domain *rd = cpu_rq(i)->rd;
56 int cpus = 0;
57
58 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
59 "sched RCU must be held");
60 for_each_cpu_and(i, rd->span, cpu_active_mask)
61 cpus++;
62
63 return cpus;
64}
65#else
66static inline struct dl_bw *dl_bw_of(int i)
67{
68 return &cpu_rq(i)->dl.dl_bw;
69}
70
71static inline int dl_bw_cpus(int i)
72{
73 return 1;
74}
75#endif
76
77static inline
78void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
79{
80 u64 old = dl_rq->running_bw;
81
82 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
83 dl_rq->running_bw += dl_bw;
84 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
85 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
86 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
87 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
88}
89
90static inline
91void __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 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
101 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
102}
103
104static inline
105void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
106{
107 u64 old = dl_rq->this_bw;
108
109 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
110 dl_rq->this_bw += dl_bw;
111 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
112}
113
114static inline
115void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
116{
117 u64 old = dl_rq->this_bw;
118
119 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
120 dl_rq->this_bw -= dl_bw;
121 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
122 if (dl_rq->this_bw > old)
123 dl_rq->this_bw = 0;
124 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
125}
126
127static inline
128void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
129{
130 if (!dl_entity_is_special(dl_se))
131 __add_rq_bw(dl_se->dl_bw, dl_rq);
132}
133
134static inline
135void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
136{
137 if (!dl_entity_is_special(dl_se))
138 __sub_rq_bw(dl_se->dl_bw, dl_rq);
139}
140
141static inline
142void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
143{
144 if (!dl_entity_is_special(dl_se))
145 __add_running_bw(dl_se->dl_bw, dl_rq);
146}
147
148static inline
149void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
150{
151 if (!dl_entity_is_special(dl_se))
152 __sub_running_bw(dl_se->dl_bw, dl_rq);
153}
154
155void dl_change_utilization(struct task_struct *p, u64 new_bw)
156{
157 struct rq *rq;
158
159 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
160
161 if (task_on_rq_queued(p))
162 return;
163
164 rq = task_rq(p);
165 if (p->dl.dl_non_contending) {
166 sub_running_bw(&p->dl, &rq->dl);
167 p->dl.dl_non_contending = 0;
168 /*
169 * If the timer handler is currently running and the
170 * timer cannot be cancelled, inactive_task_timer()
171 * will see that dl_not_contending is not set, and
172 * will not touch the rq's active utilization,
173 * so we are still safe.
174 */
175 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
176 put_task_struct(p);
177 }
178 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
179 __add_rq_bw(new_bw, &rq->dl);
180}
181
182/*
183 * The utilization of a task cannot be immediately removed from
184 * the rq active utilization (running_bw) when the task blocks.
185 * Instead, we have to wait for the so called "0-lag time".
186 *
187 * If a task blocks before the "0-lag time", a timer (the inactive
188 * timer) is armed, and running_bw is decreased when the timer
189 * fires.
190 *
191 * If the task wakes up again before the inactive timer fires,
192 * the timer is cancelled, whereas if the task wakes up after the
193 * inactive timer fired (and running_bw has been decreased) the
194 * task's utilization has to be added to running_bw again.
195 * A flag in the deadline scheduling entity (dl_non_contending)
196 * is used to avoid race conditions between the inactive timer handler
197 * and task wakeups.
198 *
199 * The following diagram shows how running_bw is updated. A task is
200 * "ACTIVE" when its utilization contributes to running_bw; an
201 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
202 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
203 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
204 * time already passed, which does not contribute to running_bw anymore.
205 * +------------------+
206 * wakeup | ACTIVE |
207 * +------------------>+ contending |
208 * | add_running_bw | |
209 * | +----+------+------+
210 * | | ^
211 * | dequeue | |
212 * +--------+-------+ | |
213 * | | t >= 0-lag | | wakeup
214 * | INACTIVE |<---------------+ |
215 * | | sub_running_bw | |
216 * +--------+-------+ | |
217 * ^ | |
218 * | t < 0-lag | |
219 * | | |
220 * | V |
221 * | +----+------+------+
222 * | sub_running_bw | ACTIVE |
223 * +-------------------+ |
224 * inactive timer | non contending |
225 * fired +------------------+
226 *
227 * The task_non_contending() function is invoked when a task
228 * blocks, and checks if the 0-lag time already passed or
229 * not (in the first case, it directly updates running_bw;
230 * in the second case, it arms the inactive timer).
231 *
232 * The task_contending() function is invoked when a task wakes
233 * up, and checks if the task is still in the "ACTIVE non contending"
234 * state or not (in the second case, it updates running_bw).
235 */
236static void task_non_contending(struct task_struct *p)
237{
238 struct sched_dl_entity *dl_se = &p->dl;
239 struct hrtimer *timer = &dl_se->inactive_timer;
240 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
241 struct rq *rq = rq_of_dl_rq(dl_rq);
242 s64 zerolag_time;
243
244 /*
245 * If this is a non-deadline task that has been boosted,
246 * do nothing
247 */
248 if (dl_se->dl_runtime == 0)
249 return;
250
251 if (dl_entity_is_special(dl_se))
252 return;
253
254 WARN_ON(hrtimer_active(&dl_se->inactive_timer));
255 WARN_ON(dl_se->dl_non_contending);
256
257 zerolag_time = dl_se->deadline -
258 div64_long((dl_se->runtime * dl_se->dl_period),
259 dl_se->dl_runtime);
260
261 /*
262 * Using relative times instead of the absolute "0-lag time"
263 * allows to simplify the code
264 */
265 zerolag_time -= rq_clock(rq);
266
267 /*
268 * If the "0-lag time" already passed, decrease the active
269 * utilization now, instead of starting a timer
270 */
271 if (zerolag_time < 0) {
272 if (dl_task(p))
273 sub_running_bw(dl_se, dl_rq);
274 if (!dl_task(p) || p->state == TASK_DEAD) {
275 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
276
277 if (p->state == TASK_DEAD)
278 sub_rq_bw(&p->dl, &rq->dl);
279 raw_spin_lock(&dl_b->lock);
280 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
281 __dl_clear_params(p);
282 raw_spin_unlock(&dl_b->lock);
283 }
284
285 return;
286 }
287
288 dl_se->dl_non_contending = 1;
289 get_task_struct(p);
290 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL);
291}
292
293static void task_contending(struct sched_dl_entity *dl_se, int flags)
294{
295 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
296
297 /*
298 * If this is a non-deadline task that has been boosted,
299 * do nothing
300 */
301 if (dl_se->dl_runtime == 0)
302 return;
303
304 if (flags & ENQUEUE_MIGRATED)
305 add_rq_bw(dl_se, dl_rq);
306
307 if (dl_se->dl_non_contending) {
308 dl_se->dl_non_contending = 0;
309 /*
310 * If the timer handler is currently running and the
311 * timer cannot be cancelled, inactive_task_timer()
312 * will see that dl_not_contending is not set, and
313 * will not touch the rq's active utilization,
314 * so we are still safe.
315 */
316 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
317 put_task_struct(dl_task_of(dl_se));
318 } else {
319 /*
320 * Since "dl_non_contending" is not set, the
321 * task's utilization has already been removed from
322 * active utilization (either when the task blocked,
323 * when the "inactive timer" fired).
324 * So, add it back.
325 */
326 add_running_bw(dl_se, dl_rq);
327 }
328}
329
330static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
331{
332 struct sched_dl_entity *dl_se = &p->dl;
333
334 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
335}
336
337void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
338{
339 raw_spin_lock_init(&dl_b->dl_runtime_lock);
340 dl_b->dl_period = period;
341 dl_b->dl_runtime = runtime;
342}
343
344void init_dl_bw(struct dl_bw *dl_b)
345{
346 raw_spin_lock_init(&dl_b->lock);
347 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
348 if (global_rt_runtime() == RUNTIME_INF)
349 dl_b->bw = -1;
350 else
351 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
352 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
353 dl_b->total_bw = 0;
354}
355
356void init_dl_rq(struct dl_rq *dl_rq)
357{
358 dl_rq->root = RB_ROOT_CACHED;
359
360#ifdef CONFIG_SMP
361 /* zero means no -deadline tasks */
362 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
363
364 dl_rq->dl_nr_migratory = 0;
365 dl_rq->overloaded = 0;
366 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
367#else
368 init_dl_bw(&dl_rq->dl_bw);
369#endif
370
371 dl_rq->running_bw = 0;
372 dl_rq->this_bw = 0;
373 init_dl_rq_bw_ratio(dl_rq);
374}
375
376#ifdef CONFIG_SMP
377
378static inline int dl_overloaded(struct rq *rq)
379{
380 return atomic_read(&rq->rd->dlo_count);
381}
382
383static inline void dl_set_overload(struct rq *rq)
384{
385 if (!rq->online)
386 return;
387
388 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
389 /*
390 * Must be visible before the overload count is
391 * set (as in sched_rt.c).
392 *
393 * Matched by the barrier in pull_dl_task().
394 */
395 smp_wmb();
396 atomic_inc(&rq->rd->dlo_count);
397}
398
399static inline void dl_clear_overload(struct rq *rq)
400{
401 if (!rq->online)
402 return;
403
404 atomic_dec(&rq->rd->dlo_count);
405 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
406}
407
408static void update_dl_migration(struct dl_rq *dl_rq)
409{
410 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
411 if (!dl_rq->overloaded) {
412 dl_set_overload(rq_of_dl_rq(dl_rq));
413 dl_rq->overloaded = 1;
414 }
415 } else if (dl_rq->overloaded) {
416 dl_clear_overload(rq_of_dl_rq(dl_rq));
417 dl_rq->overloaded = 0;
418 }
419}
420
421static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
422{
423 struct task_struct *p = dl_task_of(dl_se);
424
425 if (p->nr_cpus_allowed > 1)
426 dl_rq->dl_nr_migratory++;
427
428 update_dl_migration(dl_rq);
429}
430
431static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
432{
433 struct task_struct *p = dl_task_of(dl_se);
434
435 if (p->nr_cpus_allowed > 1)
436 dl_rq->dl_nr_migratory--;
437
438 update_dl_migration(dl_rq);
439}
440
441/*
442 * The list of pushable -deadline task is not a plist, like in
443 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
444 */
445static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
446{
447 struct dl_rq *dl_rq = &rq->dl;
448 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node;
449 struct rb_node *parent = NULL;
450 struct task_struct *entry;
451 bool leftmost = true;
452
453 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
454
455 while (*link) {
456 parent = *link;
457 entry = rb_entry(parent, struct task_struct,
458 pushable_dl_tasks);
459 if (dl_entity_preempt(&p->dl, &entry->dl))
460 link = &parent->rb_left;
461 else {
462 link = &parent->rb_right;
463 leftmost = false;
464 }
465 }
466
467 if (leftmost)
468 dl_rq->earliest_dl.next = p->dl.deadline;
469
470 rb_link_node(&p->pushable_dl_tasks, parent, link);
471 rb_insert_color_cached(&p->pushable_dl_tasks,
472 &dl_rq->pushable_dl_tasks_root, leftmost);
473}
474
475static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
476{
477 struct dl_rq *dl_rq = &rq->dl;
478
479 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
480 return;
481
482 if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) {
483 struct rb_node *next_node;
484
485 next_node = rb_next(&p->pushable_dl_tasks);
486 if (next_node) {
487 dl_rq->earliest_dl.next = rb_entry(next_node,
488 struct task_struct, pushable_dl_tasks)->dl.deadline;
489 }
490 }
491
492 rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
493 RB_CLEAR_NODE(&p->pushable_dl_tasks);
494}
495
496static inline int has_pushable_dl_tasks(struct rq *rq)
497{
498 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
499}
500
501static int push_dl_task(struct rq *rq);
502
503static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
504{
505 return dl_task(prev);
506}
507
508static DEFINE_PER_CPU(struct callback_head, dl_push_head);
509static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
510
511static void push_dl_tasks(struct rq *);
512static void pull_dl_task(struct rq *);
513
514static inline void deadline_queue_push_tasks(struct rq *rq)
515{
516 if (!has_pushable_dl_tasks(rq))
517 return;
518
519 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
520}
521
522static inline void deadline_queue_pull_task(struct rq *rq)
523{
524 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
525}
526
527static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
528
529static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
530{
531 struct rq *later_rq = NULL;
532
533 later_rq = find_lock_later_rq(p, rq);
534 if (!later_rq) {
535 int cpu;
536
537 /*
538 * If we cannot preempt any rq, fall back to pick any
539 * online CPU:
540 */
541 cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
542 if (cpu >= nr_cpu_ids) {
543 /*
544 * Failed to find any suitable CPU.
545 * The task will never come back!
546 */
547 BUG_ON(dl_bandwidth_enabled());
548
549 /*
550 * If admission control is disabled we
551 * try a little harder to let the task
552 * run.
553 */
554 cpu = cpumask_any(cpu_active_mask);
555 }
556 later_rq = cpu_rq(cpu);
557 double_lock_balance(rq, later_rq);
558 }
559
560 set_task_cpu(p, later_rq->cpu);
561 double_unlock_balance(later_rq, rq);
562
563 return later_rq;
564}
565
566#else
567
568static inline
569void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
570{
571}
572
573static inline
574void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
575{
576}
577
578static inline
579void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
580{
581}
582
583static inline
584void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
585{
586}
587
588static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
589{
590 return false;
591}
592
593static inline void pull_dl_task(struct rq *rq)
594{
595}
596
597static inline void deadline_queue_push_tasks(struct rq *rq)
598{
599}
600
601static inline void deadline_queue_pull_task(struct rq *rq)
602{
603}
604#endif /* CONFIG_SMP */
605
606static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
607static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
608static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
609
610/*
611 * We are being explicitly informed that a new instance is starting,
612 * and this means that:
613 * - the absolute deadline of the entity has to be placed at
614 * current time + relative deadline;
615 * - the runtime of the entity has to be set to the maximum value.
616 *
617 * The capability of specifying such event is useful whenever a -deadline
618 * entity wants to (try to!) synchronize its behaviour with the scheduler's
619 * one, and to (try to!) reconcile itself with its own scheduling
620 * parameters.
621 */
622static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
623{
624 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
625 struct rq *rq = rq_of_dl_rq(dl_rq);
626
627 WARN_ON(dl_se->dl_boosted);
628 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
629
630 /*
631 * We are racing with the deadline timer. So, do nothing because
632 * the deadline timer handler will take care of properly recharging
633 * the runtime and postponing the deadline
634 */
635 if (dl_se->dl_throttled)
636 return;
637
638 /*
639 * We use the regular wall clock time to set deadlines in the
640 * future; in fact, we must consider execution overheads (time
641 * spent on hardirq context, etc.).
642 */
643 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
644 dl_se->runtime = dl_se->dl_runtime;
645}
646
647/*
648 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
649 * possibility of a entity lasting more than what it declared, and thus
650 * exhausting its runtime.
651 *
652 * Here we are interested in making runtime overrun possible, but we do
653 * not want a entity which is misbehaving to affect the scheduling of all
654 * other entities.
655 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
656 * is used, in order to confine each entity within its own bandwidth.
657 *
658 * This function deals exactly with that, and ensures that when the runtime
659 * of a entity is replenished, its deadline is also postponed. That ensures
660 * the overrunning entity can't interfere with other entity in the system and
661 * can't make them miss their deadlines. Reasons why this kind of overruns
662 * could happen are, typically, a entity voluntarily trying to overcome its
663 * runtime, or it just underestimated it during sched_setattr().
664 */
665static void replenish_dl_entity(struct sched_dl_entity *dl_se,
666 struct sched_dl_entity *pi_se)
667{
668 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
669 struct rq *rq = rq_of_dl_rq(dl_rq);
670
671 BUG_ON(pi_se->dl_runtime <= 0);
672
673 /*
674 * This could be the case for a !-dl task that is boosted.
675 * Just go with full inherited parameters.
676 */
677 if (dl_se->dl_deadline == 0) {
678 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
679 dl_se->runtime = pi_se->dl_runtime;
680 }
681
682 if (dl_se->dl_yielded && dl_se->runtime > 0)
683 dl_se->runtime = 0;
684
685 /*
686 * We keep moving the deadline away until we get some
687 * available runtime for the entity. This ensures correct
688 * handling of situations where the runtime overrun is
689 * arbitrary large.
690 */
691 while (dl_se->runtime <= 0) {
692 dl_se->deadline += pi_se->dl_period;
693 dl_se->runtime += pi_se->dl_runtime;
694 }
695
696 /*
697 * At this point, the deadline really should be "in
698 * the future" with respect to rq->clock. If it's
699 * not, we are, for some reason, lagging too much!
700 * Anyway, after having warn userspace abut that,
701 * we still try to keep the things running by
702 * resetting the deadline and the budget of the
703 * entity.
704 */
705 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
706 printk_deferred_once("sched: DL replenish lagged too much\n");
707 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
708 dl_se->runtime = pi_se->dl_runtime;
709 }
710
711 if (dl_se->dl_yielded)
712 dl_se->dl_yielded = 0;
713 if (dl_se->dl_throttled)
714 dl_se->dl_throttled = 0;
715}
716
717/*
718 * Here we check if --at time t-- an entity (which is probably being
719 * [re]activated or, in general, enqueued) can use its remaining runtime
720 * and its current deadline _without_ exceeding the bandwidth it is
721 * assigned (function returns true if it can't). We are in fact applying
722 * one of the CBS rules: when a task wakes up, if the residual runtime
723 * over residual deadline fits within the allocated bandwidth, then we
724 * can keep the current (absolute) deadline and residual budget without
725 * disrupting the schedulability of the system. Otherwise, we should
726 * refill the runtime and set the deadline a period in the future,
727 * because keeping the current (absolute) deadline of the task would
728 * result in breaking guarantees promised to other tasks (refer to
729 * Documentation/scheduler/sched-deadline.txt for more informations).
730 *
731 * This function returns true if:
732 *
733 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
734 *
735 * IOW we can't recycle current parameters.
736 *
737 * Notice that the bandwidth check is done against the deadline. For
738 * task with deadline equal to period this is the same of using
739 * dl_period instead of dl_deadline in the equation above.
740 */
741static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
742 struct sched_dl_entity *pi_se, u64 t)
743{
744 u64 left, right;
745
746 /*
747 * left and right are the two sides of the equation above,
748 * after a bit of shuffling to use multiplications instead
749 * of divisions.
750 *
751 * Note that none of the time values involved in the two
752 * multiplications are absolute: dl_deadline and dl_runtime
753 * are the relative deadline and the maximum runtime of each
754 * instance, runtime is the runtime left for the last instance
755 * and (deadline - t), since t is rq->clock, is the time left
756 * to the (absolute) deadline. Even if overflowing the u64 type
757 * is very unlikely to occur in both cases, here we scale down
758 * as we want to avoid that risk at all. Scaling down by 10
759 * means that we reduce granularity to 1us. We are fine with it,
760 * since this is only a true/false check and, anyway, thinking
761 * of anything below microseconds resolution is actually fiction
762 * (but still we want to give the user that illusion >;).
763 */
764 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
765 right = ((dl_se->deadline - t) >> DL_SCALE) *
766 (pi_se->dl_runtime >> DL_SCALE);
767
768 return dl_time_before(right, left);
769}
770
771/*
772 * Revised wakeup rule [1]: For self-suspending tasks, rather then
773 * re-initializing task's runtime and deadline, the revised wakeup
774 * rule adjusts the task's runtime to avoid the task to overrun its
775 * density.
776 *
777 * Reasoning: a task may overrun the density if:
778 * runtime / (deadline - t) > dl_runtime / dl_deadline
779 *
780 * Therefore, runtime can be adjusted to:
781 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
782 *
783 * In such way that runtime will be equal to the maximum density
784 * the task can use without breaking any rule.
785 *
786 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
787 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
788 */
789static void
790update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
791{
792 u64 laxity = dl_se->deadline - rq_clock(rq);
793
794 /*
795 * If the task has deadline < period, and the deadline is in the past,
796 * it should already be throttled before this check.
797 *
798 * See update_dl_entity() comments for further details.
799 */
800 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
801
802 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
803}
804
805/*
806 * Regarding the deadline, a task with implicit deadline has a relative
807 * deadline == relative period. A task with constrained deadline has a
808 * relative deadline <= relative period.
809 *
810 * We support constrained deadline tasks. However, there are some restrictions
811 * applied only for tasks which do not have an implicit deadline. See
812 * update_dl_entity() to know more about such restrictions.
813 *
814 * The dl_is_implicit() returns true if the task has an implicit deadline.
815 */
816static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
817{
818 return dl_se->dl_deadline == dl_se->dl_period;
819}
820
821/*
822 * When a deadline entity is placed in the runqueue, its runtime and deadline
823 * might need to be updated. This is done by a CBS wake up rule. There are two
824 * different rules: 1) the original CBS; and 2) the Revisited CBS.
825 *
826 * When the task is starting a new period, the Original CBS is used. In this
827 * case, the runtime is replenished and a new absolute deadline is set.
828 *
829 * When a task is queued before the begin of the next period, using the
830 * remaining runtime and deadline could make the entity to overflow, see
831 * dl_entity_overflow() to find more about runtime overflow. When such case
832 * is detected, the runtime and deadline need to be updated.
833 *
834 * If the task has an implicit deadline, i.e., deadline == period, the Original
835 * CBS is applied. the runtime is replenished and a new absolute deadline is
836 * set, as in the previous cases.
837 *
838 * However, the Original CBS does not work properly for tasks with
839 * deadline < period, which are said to have a constrained deadline. By
840 * applying the Original CBS, a constrained deadline task would be able to run
841 * runtime/deadline in a period. With deadline < period, the task would
842 * overrun the runtime/period allowed bandwidth, breaking the admission test.
843 *
844 * In order to prevent this misbehave, the Revisited CBS is used for
845 * constrained deadline tasks when a runtime overflow is detected. In the
846 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
847 * the remaining runtime of the task is reduced to avoid runtime overflow.
848 * Please refer to the comments update_dl_revised_wakeup() function to find
849 * more about the Revised CBS rule.
850 */
851static void update_dl_entity(struct sched_dl_entity *dl_se,
852 struct sched_dl_entity *pi_se)
853{
854 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
855 struct rq *rq = rq_of_dl_rq(dl_rq);
856
857 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
858 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
859
860 if (unlikely(!dl_is_implicit(dl_se) &&
861 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
862 !dl_se->dl_boosted)){
863 update_dl_revised_wakeup(dl_se, rq);
864 return;
865 }
866
867 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
868 dl_se->runtime = pi_se->dl_runtime;
869 }
870}
871
872static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
873{
874 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
875}
876
877/*
878 * If the entity depleted all its runtime, and if we want it to sleep
879 * while waiting for some new execution time to become available, we
880 * set the bandwidth replenishment timer to the replenishment instant
881 * and try to activate it.
882 *
883 * Notice that it is important for the caller to know if the timer
884 * actually started or not (i.e., the replenishment instant is in
885 * the future or in the past).
886 */
887static int start_dl_timer(struct task_struct *p)
888{
889 struct sched_dl_entity *dl_se = &p->dl;
890 struct hrtimer *timer = &dl_se->dl_timer;
891 struct rq *rq = task_rq(p);
892 ktime_t now, act;
893 s64 delta;
894
895 lockdep_assert_held(&rq->lock);
896
897 /*
898 * We want the timer to fire at the deadline, but considering
899 * that it is actually coming from rq->clock and not from
900 * hrtimer's time base reading.
901 */
902 act = ns_to_ktime(dl_next_period(dl_se));
903 now = hrtimer_cb_get_time(timer);
904 delta = ktime_to_ns(now) - rq_clock(rq);
905 act = ktime_add_ns(act, delta);
906
907 /*
908 * If the expiry time already passed, e.g., because the value
909 * chosen as the deadline is too small, don't even try to
910 * start the timer in the past!
911 */
912 if (ktime_us_delta(act, now) < 0)
913 return 0;
914
915 /*
916 * !enqueued will guarantee another callback; even if one is already in
917 * progress. This ensures a balanced {get,put}_task_struct().
918 *
919 * The race against __run_timer() clearing the enqueued state is
920 * harmless because we're holding task_rq()->lock, therefore the timer
921 * expiring after we've done the check will wait on its task_rq_lock()
922 * and observe our state.
923 */
924 if (!hrtimer_is_queued(timer)) {
925 get_task_struct(p);
926 hrtimer_start(timer, act, HRTIMER_MODE_ABS);
927 }
928
929 return 1;
930}
931
932/*
933 * This is the bandwidth enforcement timer callback. If here, we know
934 * a task is not on its dl_rq, since the fact that the timer was running
935 * means the task is throttled and needs a runtime replenishment.
936 *
937 * However, what we actually do depends on the fact the task is active,
938 * (it is on its rq) or has been removed from there by a call to
939 * dequeue_task_dl(). In the former case we must issue the runtime
940 * replenishment and add the task back to the dl_rq; in the latter, we just
941 * do nothing but clearing dl_throttled, so that runtime and deadline
942 * updating (and the queueing back to dl_rq) will be done by the
943 * next call to enqueue_task_dl().
944 */
945static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
946{
947 struct sched_dl_entity *dl_se = container_of(timer,
948 struct sched_dl_entity,
949 dl_timer);
950 struct task_struct *p = dl_task_of(dl_se);
951 struct rq_flags rf;
952 struct rq *rq;
953
954 rq = task_rq_lock(p, &rf);
955
956 /*
957 * The task might have changed its scheduling policy to something
958 * different than SCHED_DEADLINE (through switched_from_dl()).
959 */
960 if (!dl_task(p))
961 goto unlock;
962
963 /*
964 * The task might have been boosted by someone else and might be in the
965 * boosting/deboosting path, its not throttled.
966 */
967 if (dl_se->dl_boosted)
968 goto unlock;
969
970 /*
971 * Spurious timer due to start_dl_timer() race; or we already received
972 * a replenishment from rt_mutex_setprio().
973 */
974 if (!dl_se->dl_throttled)
975 goto unlock;
976
977 sched_clock_tick();
978 update_rq_clock(rq);
979
980 /*
981 * If the throttle happened during sched-out; like:
982 *
983 * schedule()
984 * deactivate_task()
985 * dequeue_task_dl()
986 * update_curr_dl()
987 * start_dl_timer()
988 * __dequeue_task_dl()
989 * prev->on_rq = 0;
990 *
991 * We can be both throttled and !queued. Replenish the counter
992 * but do not enqueue -- wait for our wakeup to do that.
993 */
994 if (!task_on_rq_queued(p)) {
995 replenish_dl_entity(dl_se, dl_se);
996 goto unlock;
997 }
998
999#ifdef CONFIG_SMP
1000 if (unlikely(!rq->online)) {
1001 /*
1002 * If the runqueue is no longer available, migrate the
1003 * task elsewhere. This necessarily changes rq.
1004 */
1005 lockdep_unpin_lock(&rq->lock, rf.cookie);
1006 rq = dl_task_offline_migration(rq, p);
1007 rf.cookie = lockdep_pin_lock(&rq->lock);
1008 update_rq_clock(rq);
1009
1010 /*
1011 * Now that the task has been migrated to the new RQ and we
1012 * have that locked, proceed as normal and enqueue the task
1013 * there.
1014 */
1015 }
1016#endif
1017
1018 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1019 if (dl_task(rq->curr))
1020 check_preempt_curr_dl(rq, p, 0);
1021 else
1022 resched_curr(rq);
1023
1024#ifdef CONFIG_SMP
1025 /*
1026 * Queueing this task back might have overloaded rq, check if we need
1027 * to kick someone away.
1028 */
1029 if (has_pushable_dl_tasks(rq)) {
1030 /*
1031 * Nothing relies on rq->lock after this, so its safe to drop
1032 * rq->lock.
1033 */
1034 rq_unpin_lock(rq, &rf);
1035 push_dl_task(rq);
1036 rq_repin_lock(rq, &rf);
1037 }
1038#endif
1039
1040unlock:
1041 task_rq_unlock(rq, p, &rf);
1042
1043 /*
1044 * This can free the task_struct, including this hrtimer, do not touch
1045 * anything related to that after this.
1046 */
1047 put_task_struct(p);
1048
1049 return HRTIMER_NORESTART;
1050}
1051
1052void init_dl_task_timer(struct sched_dl_entity *dl_se)
1053{
1054 struct hrtimer *timer = &dl_se->dl_timer;
1055
1056 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1057 timer->function = dl_task_timer;
1058}
1059
1060/*
1061 * During the activation, CBS checks if it can reuse the current task's
1062 * runtime and period. If the deadline of the task is in the past, CBS
1063 * cannot use the runtime, and so it replenishes the task. This rule
1064 * works fine for implicit deadline tasks (deadline == period), and the
1065 * CBS was designed for implicit deadline tasks. However, a task with
1066 * constrained deadline (deadine < period) might be awakened after the
1067 * deadline, but before the next period. In this case, replenishing the
1068 * task would allow it to run for runtime / deadline. As in this case
1069 * deadline < period, CBS enables a task to run for more than the
1070 * runtime / period. In a very loaded system, this can cause a domino
1071 * effect, making other tasks miss their deadlines.
1072 *
1073 * To avoid this problem, in the activation of a constrained deadline
1074 * task after the deadline but before the next period, throttle the
1075 * task and set the replenishing timer to the begin of the next period,
1076 * unless it is boosted.
1077 */
1078static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1079{
1080 struct task_struct *p = dl_task_of(dl_se);
1081 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1082
1083 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1084 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1085 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1086 return;
1087 dl_se->dl_throttled = 1;
1088 if (dl_se->runtime > 0)
1089 dl_se->runtime = 0;
1090 }
1091}
1092
1093static
1094int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1095{
1096 return (dl_se->runtime <= 0);
1097}
1098
1099extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1100
1101/*
1102 * This function implements the GRUB accounting rule:
1103 * according to the GRUB reclaiming algorithm, the runtime is
1104 * not decreased as "dq = -dt", but as
1105 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1106 * where u is the utilization of the task, Umax is the maximum reclaimable
1107 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1108 * as the difference between the "total runqueue utilization" and the
1109 * runqueue active utilization, and Uextra is the (per runqueue) extra
1110 * reclaimable utilization.
1111 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1112 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1113 * BW_SHIFT.
1114 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1115 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1116 * Since delta is a 64 bit variable, to have an overflow its value
1117 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1118 * So, overflow is not an issue here.
1119 */
1120static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1121{
1122 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1123 u64 u_act;
1124 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1125
1126 /*
1127 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1128 * we compare u_inact + rq->dl.extra_bw with
1129 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1130 * u_inact + rq->dl.extra_bw can be larger than
1131 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1132 * leading to wrong results)
1133 */
1134 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1135 u_act = u_act_min;
1136 else
1137 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1138
1139 return (delta * u_act) >> BW_SHIFT;
1140}
1141
1142/*
1143 * Update the current task's runtime statistics (provided it is still
1144 * a -deadline task and has not been removed from the dl_rq).
1145 */
1146static void update_curr_dl(struct rq *rq)
1147{
1148 struct task_struct *curr = rq->curr;
1149 struct sched_dl_entity *dl_se = &curr->dl;
1150 u64 delta_exec, scaled_delta_exec;
1151 int cpu = cpu_of(rq);
1152 u64 now;
1153
1154 if (!dl_task(curr) || !on_dl_rq(dl_se))
1155 return;
1156
1157 /*
1158 * Consumed budget is computed considering the time as
1159 * observed by schedulable tasks (excluding time spent
1160 * in hardirq context, etc.). Deadlines are instead
1161 * computed using hard walltime. This seems to be the more
1162 * natural solution, but the full ramifications of this
1163 * approach need further study.
1164 */
1165 now = rq_clock_task(rq);
1166 delta_exec = now - curr->se.exec_start;
1167 if (unlikely((s64)delta_exec <= 0)) {
1168 if (unlikely(dl_se->dl_yielded))
1169 goto throttle;
1170 return;
1171 }
1172
1173 schedstat_set(curr->se.statistics.exec_max,
1174 max(curr->se.statistics.exec_max, delta_exec));
1175
1176 curr->se.sum_exec_runtime += delta_exec;
1177 account_group_exec_runtime(curr, delta_exec);
1178
1179 curr->se.exec_start = now;
1180 cgroup_account_cputime(curr, delta_exec);
1181
1182 sched_rt_avg_update(rq, delta_exec);
1183
1184 if (dl_entity_is_special(dl_se))
1185 return;
1186
1187 /*
1188 * For tasks that participate in GRUB, we implement GRUB-PA: the
1189 * spare reclaimed bandwidth is used to clock down frequency.
1190 *
1191 * For the others, we still need to scale reservation parameters
1192 * according to current frequency and CPU maximum capacity.
1193 */
1194 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1195 scaled_delta_exec = grub_reclaim(delta_exec,
1196 rq,
1197 &curr->dl);
1198 } else {
1199 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1200 unsigned long scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
1201
1202 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1203 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1204 }
1205
1206 dl_se->runtime -= scaled_delta_exec;
1207
1208throttle:
1209 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1210 dl_se->dl_throttled = 1;
1211
1212 /* If requested, inform the user about runtime overruns. */
1213 if (dl_runtime_exceeded(dl_se) &&
1214 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1215 dl_se->dl_overrun = 1;
1216
1217 __dequeue_task_dl(rq, curr, 0);
1218 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
1219 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1220
1221 if (!is_leftmost(curr, &rq->dl))
1222 resched_curr(rq);
1223 }
1224
1225 /*
1226 * Because -- for now -- we share the rt bandwidth, we need to
1227 * account our runtime there too, otherwise actual rt tasks
1228 * would be able to exceed the shared quota.
1229 *
1230 * Account to the root rt group for now.
1231 *
1232 * The solution we're working towards is having the RT groups scheduled
1233 * using deadline servers -- however there's a few nasties to figure
1234 * out before that can happen.
1235 */
1236 if (rt_bandwidth_enabled()) {
1237 struct rt_rq *rt_rq = &rq->rt;
1238
1239 raw_spin_lock(&rt_rq->rt_runtime_lock);
1240 /*
1241 * We'll let actual RT tasks worry about the overflow here, we
1242 * have our own CBS to keep us inline; only account when RT
1243 * bandwidth is relevant.
1244 */
1245 if (sched_rt_bandwidth_account(rt_rq))
1246 rt_rq->rt_time += delta_exec;
1247 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1248 }
1249}
1250
1251static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1252{
1253 struct sched_dl_entity *dl_se = container_of(timer,
1254 struct sched_dl_entity,
1255 inactive_timer);
1256 struct task_struct *p = dl_task_of(dl_se);
1257 struct rq_flags rf;
1258 struct rq *rq;
1259
1260 rq = task_rq_lock(p, &rf);
1261
1262 sched_clock_tick();
1263 update_rq_clock(rq);
1264
1265 if (!dl_task(p) || p->state == TASK_DEAD) {
1266 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1267
1268 if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1269 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1270 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1271 dl_se->dl_non_contending = 0;
1272 }
1273
1274 raw_spin_lock(&dl_b->lock);
1275 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1276 raw_spin_unlock(&dl_b->lock);
1277 __dl_clear_params(p);
1278
1279 goto unlock;
1280 }
1281 if (dl_se->dl_non_contending == 0)
1282 goto unlock;
1283
1284 sub_running_bw(dl_se, &rq->dl);
1285 dl_se->dl_non_contending = 0;
1286unlock:
1287 task_rq_unlock(rq, p, &rf);
1288 put_task_struct(p);
1289
1290 return HRTIMER_NORESTART;
1291}
1292
1293void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1294{
1295 struct hrtimer *timer = &dl_se->inactive_timer;
1296
1297 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1298 timer->function = inactive_task_timer;
1299}
1300
1301#ifdef CONFIG_SMP
1302
1303static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1304{
1305 struct rq *rq = rq_of_dl_rq(dl_rq);
1306
1307 if (dl_rq->earliest_dl.curr == 0 ||
1308 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1309 dl_rq->earliest_dl.curr = deadline;
1310 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1311 }
1312}
1313
1314static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1315{
1316 struct rq *rq = rq_of_dl_rq(dl_rq);
1317
1318 /*
1319 * Since we may have removed our earliest (and/or next earliest)
1320 * task we must recompute them.
1321 */
1322 if (!dl_rq->dl_nr_running) {
1323 dl_rq->earliest_dl.curr = 0;
1324 dl_rq->earliest_dl.next = 0;
1325 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1326 } else {
1327 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1328 struct sched_dl_entity *entry;
1329
1330 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1331 dl_rq->earliest_dl.curr = entry->deadline;
1332 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1333 }
1334}
1335
1336#else
1337
1338static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1339static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1340
1341#endif /* CONFIG_SMP */
1342
1343static inline
1344void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1345{
1346 int prio = dl_task_of(dl_se)->prio;
1347 u64 deadline = dl_se->deadline;
1348
1349 WARN_ON(!dl_prio(prio));
1350 dl_rq->dl_nr_running++;
1351 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1352
1353 inc_dl_deadline(dl_rq, deadline);
1354 inc_dl_migration(dl_se, dl_rq);
1355}
1356
1357static inline
1358void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1359{
1360 int prio = dl_task_of(dl_se)->prio;
1361
1362 WARN_ON(!dl_prio(prio));
1363 WARN_ON(!dl_rq->dl_nr_running);
1364 dl_rq->dl_nr_running--;
1365 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1366
1367 dec_dl_deadline(dl_rq, dl_se->deadline);
1368 dec_dl_migration(dl_se, dl_rq);
1369}
1370
1371static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1372{
1373 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1374 struct rb_node **link = &dl_rq->root.rb_root.rb_node;
1375 struct rb_node *parent = NULL;
1376 struct sched_dl_entity *entry;
1377 int leftmost = 1;
1378
1379 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1380
1381 while (*link) {
1382 parent = *link;
1383 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
1384 if (dl_time_before(dl_se->deadline, entry->deadline))
1385 link = &parent->rb_left;
1386 else {
1387 link = &parent->rb_right;
1388 leftmost = 0;
1389 }
1390 }
1391
1392 rb_link_node(&dl_se->rb_node, parent, link);
1393 rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost);
1394
1395 inc_dl_tasks(dl_se, dl_rq);
1396}
1397
1398static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1399{
1400 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1401
1402 if (RB_EMPTY_NODE(&dl_se->rb_node))
1403 return;
1404
1405 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1406 RB_CLEAR_NODE(&dl_se->rb_node);
1407
1408 dec_dl_tasks(dl_se, dl_rq);
1409}
1410
1411static void
1412enqueue_dl_entity(struct sched_dl_entity *dl_se,
1413 struct sched_dl_entity *pi_se, int flags)
1414{
1415 BUG_ON(on_dl_rq(dl_se));
1416
1417 /*
1418 * If this is a wakeup or a new instance, the scheduling
1419 * parameters of the task might need updating. Otherwise,
1420 * we want a replenishment of its runtime.
1421 */
1422 if (flags & ENQUEUE_WAKEUP) {
1423 task_contending(dl_se, flags);
1424 update_dl_entity(dl_se, pi_se);
1425 } else if (flags & ENQUEUE_REPLENISH) {
1426 replenish_dl_entity(dl_se, pi_se);
1427 } else if ((flags & ENQUEUE_RESTORE) &&
1428 dl_time_before(dl_se->deadline,
1429 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1430 setup_new_dl_entity(dl_se);
1431 }
1432
1433 __enqueue_dl_entity(dl_se);
1434}
1435
1436static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1437{
1438 __dequeue_dl_entity(dl_se);
1439}
1440
1441static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1442{
1443 struct task_struct *pi_task = rt_mutex_get_top_task(p);
1444 struct sched_dl_entity *pi_se = &p->dl;
1445
1446 /*
1447 * Use the scheduling parameters of the top pi-waiter task if:
1448 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1449 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1450 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1451 * boosted due to a SCHED_DEADLINE pi-waiter).
1452 * Otherwise we keep our runtime and deadline.
1453 */
1454 if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) {
1455 pi_se = &pi_task->dl;
1456 } else if (!dl_prio(p->normal_prio)) {
1457 /*
1458 * Special case in which we have a !SCHED_DEADLINE task
1459 * that is going to be deboosted, but exceeds its
1460 * runtime while doing so. No point in replenishing
1461 * it, as it's going to return back to its original
1462 * scheduling class after this.
1463 */
1464 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
1465 return;
1466 }
1467
1468 /*
1469 * Check if a constrained deadline task was activated
1470 * after the deadline but before the next period.
1471 * If that is the case, the task will be throttled and
1472 * the replenishment timer will be set to the next period.
1473 */
1474 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1475 dl_check_constrained_dl(&p->dl);
1476
1477 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1478 add_rq_bw(&p->dl, &rq->dl);
1479 add_running_bw(&p->dl, &rq->dl);
1480 }
1481
1482 /*
1483 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1484 * its budget it needs a replenishment and, since it now is on
1485 * its rq, the bandwidth timer callback (which clearly has not
1486 * run yet) will take care of this.
1487 * However, the active utilization does not depend on the fact
1488 * that the task is on the runqueue or not (but depends on the
1489 * task's state - in GRUB parlance, "inactive" vs "active contending").
1490 * In other words, even if a task is throttled its utilization must
1491 * be counted in the active utilization; hence, we need to call
1492 * add_running_bw().
1493 */
1494 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1495 if (flags & ENQUEUE_WAKEUP)
1496 task_contending(&p->dl, flags);
1497
1498 return;
1499 }
1500
1501 enqueue_dl_entity(&p->dl, pi_se, flags);
1502
1503 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1504 enqueue_pushable_dl_task(rq, p);
1505}
1506
1507static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1508{
1509 dequeue_dl_entity(&p->dl);
1510 dequeue_pushable_dl_task(rq, p);
1511}
1512
1513static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1514{
1515 update_curr_dl(rq);
1516 __dequeue_task_dl(rq, p, flags);
1517
1518 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1519 sub_running_bw(&p->dl, &rq->dl);
1520 sub_rq_bw(&p->dl, &rq->dl);
1521 }
1522
1523 /*
1524 * This check allows to start the inactive timer (or to immediately
1525 * decrease the active utilization, if needed) in two cases:
1526 * when the task blocks and when it is terminating
1527 * (p->state == TASK_DEAD). We can handle the two cases in the same
1528 * way, because from GRUB's point of view the same thing is happening
1529 * (the task moves from "active contending" to "active non contending"
1530 * or "inactive")
1531 */
1532 if (flags & DEQUEUE_SLEEP)
1533 task_non_contending(p);
1534}
1535
1536/*
1537 * Yield task semantic for -deadline tasks is:
1538 *
1539 * get off from the CPU until our next instance, with
1540 * a new runtime. This is of little use now, since we
1541 * don't have a bandwidth reclaiming mechanism. Anyway,
1542 * bandwidth reclaiming is planned for the future, and
1543 * yield_task_dl will indicate that some spare budget
1544 * is available for other task instances to use it.
1545 */
1546static void yield_task_dl(struct rq *rq)
1547{
1548 /*
1549 * We make the task go to sleep until its current deadline by
1550 * forcing its runtime to zero. This way, update_curr_dl() stops
1551 * it and the bandwidth timer will wake it up and will give it
1552 * new scheduling parameters (thanks to dl_yielded=1).
1553 */
1554 rq->curr->dl.dl_yielded = 1;
1555
1556 update_rq_clock(rq);
1557 update_curr_dl(rq);
1558 /*
1559 * Tell update_rq_clock() that we've just updated,
1560 * so we don't do microscopic update in schedule()
1561 * and double the fastpath cost.
1562 */
1563 rq_clock_skip_update(rq);
1564}
1565
1566#ifdef CONFIG_SMP
1567
1568static int find_later_rq(struct task_struct *task);
1569
1570static int
1571select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1572{
1573 struct task_struct *curr;
1574 struct rq *rq;
1575
1576 if (sd_flag != SD_BALANCE_WAKE)
1577 goto out;
1578
1579 rq = cpu_rq(cpu);
1580
1581 rcu_read_lock();
1582 curr = READ_ONCE(rq->curr); /* unlocked access */
1583
1584 /*
1585 * If we are dealing with a -deadline task, we must
1586 * decide where to wake it up.
1587 * If it has a later deadline and the current task
1588 * on this rq can't move (provided the waking task
1589 * can!) we prefer to send it somewhere else. On the
1590 * other hand, if it has a shorter deadline, we
1591 * try to make it stay here, it might be important.
1592 */
1593 if (unlikely(dl_task(curr)) &&
1594 (curr->nr_cpus_allowed < 2 ||
1595 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1596 (p->nr_cpus_allowed > 1)) {
1597 int target = find_later_rq(p);
1598
1599 if (target != -1 &&
1600 (dl_time_before(p->dl.deadline,
1601 cpu_rq(target)->dl.earliest_dl.curr) ||
1602 (cpu_rq(target)->dl.dl_nr_running == 0)))
1603 cpu = target;
1604 }
1605 rcu_read_unlock();
1606
1607out:
1608 return cpu;
1609}
1610
1611static void migrate_task_rq_dl(struct task_struct *p)
1612{
1613 struct rq *rq;
1614
1615 if (p->state != TASK_WAKING)
1616 return;
1617
1618 rq = task_rq(p);
1619 /*
1620 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1621 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1622 * rq->lock is not... So, lock it
1623 */
1624 raw_spin_lock(&rq->lock);
1625 if (p->dl.dl_non_contending) {
1626 sub_running_bw(&p->dl, &rq->dl);
1627 p->dl.dl_non_contending = 0;
1628 /*
1629 * If the timer handler is currently running and the
1630 * timer cannot be cancelled, inactive_task_timer()
1631 * will see that dl_not_contending is not set, and
1632 * will not touch the rq's active utilization,
1633 * so we are still safe.
1634 */
1635 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1636 put_task_struct(p);
1637 }
1638 sub_rq_bw(&p->dl, &rq->dl);
1639 raw_spin_unlock(&rq->lock);
1640}
1641
1642static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1643{
1644 /*
1645 * Current can't be migrated, useless to reschedule,
1646 * let's hope p can move out.
1647 */
1648 if (rq->curr->nr_cpus_allowed == 1 ||
1649 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1650 return;
1651
1652 /*
1653 * p is migratable, so let's not schedule it and
1654 * see if it is pushed or pulled somewhere else.
1655 */
1656 if (p->nr_cpus_allowed != 1 &&
1657 cpudl_find(&rq->rd->cpudl, p, NULL))
1658 return;
1659
1660 resched_curr(rq);
1661}
1662
1663#endif /* CONFIG_SMP */
1664
1665/*
1666 * Only called when both the current and waking task are -deadline
1667 * tasks.
1668 */
1669static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1670 int flags)
1671{
1672 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1673 resched_curr(rq);
1674 return;
1675 }
1676
1677#ifdef CONFIG_SMP
1678 /*
1679 * In the unlikely case current and p have the same deadline
1680 * let us try to decide what's the best thing to do...
1681 */
1682 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1683 !test_tsk_need_resched(rq->curr))
1684 check_preempt_equal_dl(rq, p);
1685#endif /* CONFIG_SMP */
1686}
1687
1688#ifdef CONFIG_SCHED_HRTICK
1689static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1690{
1691 hrtick_start(rq, p->dl.runtime);
1692}
1693#else /* !CONFIG_SCHED_HRTICK */
1694static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1695{
1696}
1697#endif
1698
1699static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1700 struct dl_rq *dl_rq)
1701{
1702 struct rb_node *left = rb_first_cached(&dl_rq->root);
1703
1704 if (!left)
1705 return NULL;
1706
1707 return rb_entry(left, struct sched_dl_entity, rb_node);
1708}
1709
1710static struct task_struct *
1711pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1712{
1713 struct sched_dl_entity *dl_se;
1714 struct task_struct *p;
1715 struct dl_rq *dl_rq;
1716
1717 dl_rq = &rq->dl;
1718
1719 if (need_pull_dl_task(rq, prev)) {
1720 /*
1721 * This is OK, because current is on_cpu, which avoids it being
1722 * picked for load-balance and preemption/IRQs are still
1723 * disabled avoiding further scheduler activity on it and we're
1724 * being very careful to re-start the picking loop.
1725 */
1726 rq_unpin_lock(rq, rf);
1727 pull_dl_task(rq);
1728 rq_repin_lock(rq, rf);
1729 /*
1730 * pull_dl_task() can drop (and re-acquire) rq->lock; this
1731 * means a stop task can slip in, in which case we need to
1732 * re-start task selection.
1733 */
1734 if (rq->stop && task_on_rq_queued(rq->stop))
1735 return RETRY_TASK;
1736 }
1737
1738 /*
1739 * When prev is DL, we may throttle it in put_prev_task().
1740 * So, we update time before we check for dl_nr_running.
1741 */
1742 if (prev->sched_class == &dl_sched_class)
1743 update_curr_dl(rq);
1744
1745 if (unlikely(!dl_rq->dl_nr_running))
1746 return NULL;
1747
1748 put_prev_task(rq, prev);
1749
1750 dl_se = pick_next_dl_entity(rq, dl_rq);
1751 BUG_ON(!dl_se);
1752
1753 p = dl_task_of(dl_se);
1754 p->se.exec_start = rq_clock_task(rq);
1755
1756 /* Running task will never be pushed. */
1757 dequeue_pushable_dl_task(rq, p);
1758
1759 if (hrtick_enabled(rq))
1760 start_hrtick_dl(rq, p);
1761
1762 deadline_queue_push_tasks(rq);
1763
1764 return p;
1765}
1766
1767static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1768{
1769 update_curr_dl(rq);
1770
1771 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1772 enqueue_pushable_dl_task(rq, p);
1773}
1774
1775/*
1776 * scheduler tick hitting a task of our scheduling class.
1777 *
1778 * NOTE: This function can be called remotely by the tick offload that
1779 * goes along full dynticks. Therefore no local assumption can be made
1780 * and everything must be accessed through the @rq and @curr passed in
1781 * parameters.
1782 */
1783static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1784{
1785 update_curr_dl(rq);
1786
1787 /*
1788 * Even when we have runtime, update_curr_dl() might have resulted in us
1789 * not being the leftmost task anymore. In that case NEED_RESCHED will
1790 * be set and schedule() will start a new hrtick for the next task.
1791 */
1792 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1793 is_leftmost(p, &rq->dl))
1794 start_hrtick_dl(rq, p);
1795}
1796
1797static void task_fork_dl(struct task_struct *p)
1798{
1799 /*
1800 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1801 * sched_fork()
1802 */
1803}
1804
1805static void set_curr_task_dl(struct rq *rq)
1806{
1807 struct task_struct *p = rq->curr;
1808
1809 p->se.exec_start = rq_clock_task(rq);
1810
1811 /* You can't push away the running task */
1812 dequeue_pushable_dl_task(rq, p);
1813}
1814
1815#ifdef CONFIG_SMP
1816
1817/* Only try algorithms three times */
1818#define DL_MAX_TRIES 3
1819
1820static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1821{
1822 if (!task_running(rq, p) &&
1823 cpumask_test_cpu(cpu, &p->cpus_allowed))
1824 return 1;
1825 return 0;
1826}
1827
1828/*
1829 * Return the earliest pushable rq's task, which is suitable to be executed
1830 * on the CPU, NULL otherwise:
1831 */
1832static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1833{
1834 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
1835 struct task_struct *p = NULL;
1836
1837 if (!has_pushable_dl_tasks(rq))
1838 return NULL;
1839
1840next_node:
1841 if (next_node) {
1842 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1843
1844 if (pick_dl_task(rq, p, cpu))
1845 return p;
1846
1847 next_node = rb_next(next_node);
1848 goto next_node;
1849 }
1850
1851 return NULL;
1852}
1853
1854static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1855
1856static int find_later_rq(struct task_struct *task)
1857{
1858 struct sched_domain *sd;
1859 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1860 int this_cpu = smp_processor_id();
1861 int cpu = task_cpu(task);
1862
1863 /* Make sure the mask is initialized first */
1864 if (unlikely(!later_mask))
1865 return -1;
1866
1867 if (task->nr_cpus_allowed == 1)
1868 return -1;
1869
1870 /*
1871 * We have to consider system topology and task affinity
1872 * first, then we can look for a suitable CPU.
1873 */
1874 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1875 return -1;
1876
1877 /*
1878 * If we are here, some targets have been found, including
1879 * the most suitable which is, among the runqueues where the
1880 * current tasks have later deadlines than the task's one, the
1881 * rq with the latest possible one.
1882 *
1883 * Now we check how well this matches with task's
1884 * affinity and system topology.
1885 *
1886 * The last CPU where the task run is our first
1887 * guess, since it is most likely cache-hot there.
1888 */
1889 if (cpumask_test_cpu(cpu, later_mask))
1890 return cpu;
1891 /*
1892 * Check if this_cpu is to be skipped (i.e., it is
1893 * not in the mask) or not.
1894 */
1895 if (!cpumask_test_cpu(this_cpu, later_mask))
1896 this_cpu = -1;
1897
1898 rcu_read_lock();
1899 for_each_domain(cpu, sd) {
1900 if (sd->flags & SD_WAKE_AFFINE) {
1901 int best_cpu;
1902
1903 /*
1904 * If possible, preempting this_cpu is
1905 * cheaper than migrating.
1906 */
1907 if (this_cpu != -1 &&
1908 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1909 rcu_read_unlock();
1910 return this_cpu;
1911 }
1912
1913 best_cpu = cpumask_first_and(later_mask,
1914 sched_domain_span(sd));
1915 /*
1916 * Last chance: if a CPU being in both later_mask
1917 * and current sd span is valid, that becomes our
1918 * choice. Of course, the latest possible CPU is
1919 * already under consideration through later_mask.
1920 */
1921 if (best_cpu < nr_cpu_ids) {
1922 rcu_read_unlock();
1923 return best_cpu;
1924 }
1925 }
1926 }
1927 rcu_read_unlock();
1928
1929 /*
1930 * At this point, all our guesses failed, we just return
1931 * 'something', and let the caller sort the things out.
1932 */
1933 if (this_cpu != -1)
1934 return this_cpu;
1935
1936 cpu = cpumask_any(later_mask);
1937 if (cpu < nr_cpu_ids)
1938 return cpu;
1939
1940 return -1;
1941}
1942
1943/* Locks the rq it finds */
1944static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1945{
1946 struct rq *later_rq = NULL;
1947 int tries;
1948 int cpu;
1949
1950 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1951 cpu = find_later_rq(task);
1952
1953 if ((cpu == -1) || (cpu == rq->cpu))
1954 break;
1955
1956 later_rq = cpu_rq(cpu);
1957
1958 if (later_rq->dl.dl_nr_running &&
1959 !dl_time_before(task->dl.deadline,
1960 later_rq->dl.earliest_dl.curr)) {
1961 /*
1962 * Target rq has tasks of equal or earlier deadline,
1963 * retrying does not release any lock and is unlikely
1964 * to yield a different result.
1965 */
1966 later_rq = NULL;
1967 break;
1968 }
1969
1970 /* Retry if something changed. */
1971 if (double_lock_balance(rq, later_rq)) {
1972 if (unlikely(task_rq(task) != rq ||
1973 !cpumask_test_cpu(later_rq->cpu, &task->cpus_allowed) ||
1974 task_running(rq, task) ||
1975 !dl_task(task) ||
1976 !task_on_rq_queued(task))) {
1977 double_unlock_balance(rq, later_rq);
1978 later_rq = NULL;
1979 break;
1980 }
1981 }
1982
1983 /*
1984 * If the rq we found has no -deadline task, or
1985 * its earliest one has a later deadline than our
1986 * task, the rq is a good one.
1987 */
1988 if (!later_rq->dl.dl_nr_running ||
1989 dl_time_before(task->dl.deadline,
1990 later_rq->dl.earliest_dl.curr))
1991 break;
1992
1993 /* Otherwise we try again. */
1994 double_unlock_balance(rq, later_rq);
1995 later_rq = NULL;
1996 }
1997
1998 return later_rq;
1999}
2000
2001static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2002{
2003 struct task_struct *p;
2004
2005 if (!has_pushable_dl_tasks(rq))
2006 return NULL;
2007
2008 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
2009 struct task_struct, pushable_dl_tasks);
2010
2011 BUG_ON(rq->cpu != task_cpu(p));
2012 BUG_ON(task_current(rq, p));
2013 BUG_ON(p->nr_cpus_allowed <= 1);
2014
2015 BUG_ON(!task_on_rq_queued(p));
2016 BUG_ON(!dl_task(p));
2017
2018 return p;
2019}
2020
2021/*
2022 * See if the non running -deadline tasks on this rq
2023 * can be sent to some other CPU where they can preempt
2024 * and start executing.
2025 */
2026static int push_dl_task(struct rq *rq)
2027{
2028 struct task_struct *next_task;
2029 struct rq *later_rq;
2030 int ret = 0;
2031
2032 if (!rq->dl.overloaded)
2033 return 0;
2034
2035 next_task = pick_next_pushable_dl_task(rq);
2036 if (!next_task)
2037 return 0;
2038
2039retry:
2040 if (unlikely(next_task == rq->curr)) {
2041 WARN_ON(1);
2042 return 0;
2043 }
2044
2045 /*
2046 * If next_task preempts rq->curr, and rq->curr
2047 * can move away, it makes sense to just reschedule
2048 * without going further in pushing next_task.
2049 */
2050 if (dl_task(rq->curr) &&
2051 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2052 rq->curr->nr_cpus_allowed > 1) {
2053 resched_curr(rq);
2054 return 0;
2055 }
2056
2057 /* We might release rq lock */
2058 get_task_struct(next_task);
2059
2060 /* Will lock the rq it'll find */
2061 later_rq = find_lock_later_rq(next_task, rq);
2062 if (!later_rq) {
2063 struct task_struct *task;
2064
2065 /*
2066 * We must check all this again, since
2067 * find_lock_later_rq releases rq->lock and it is
2068 * then possible that next_task has migrated.
2069 */
2070 task = pick_next_pushable_dl_task(rq);
2071 if (task == next_task) {
2072 /*
2073 * The task is still there. We don't try
2074 * again, some other CPU will pull it when ready.
2075 */
2076 goto out;
2077 }
2078
2079 if (!task)
2080 /* No more tasks */
2081 goto out;
2082
2083 put_task_struct(next_task);
2084 next_task = task;
2085 goto retry;
2086 }
2087
2088 deactivate_task(rq, next_task, 0);
2089 sub_running_bw(&next_task->dl, &rq->dl);
2090 sub_rq_bw(&next_task->dl, &rq->dl);
2091 set_task_cpu(next_task, later_rq->cpu);
2092 add_rq_bw(&next_task->dl, &later_rq->dl);
2093 add_running_bw(&next_task->dl, &later_rq->dl);
2094 activate_task(later_rq, next_task, 0);
2095 ret = 1;
2096
2097 resched_curr(later_rq);
2098
2099 double_unlock_balance(rq, later_rq);
2100
2101out:
2102 put_task_struct(next_task);
2103
2104 return ret;
2105}
2106
2107static void push_dl_tasks(struct rq *rq)
2108{
2109 /* push_dl_task() will return true if it moved a -deadline task */
2110 while (push_dl_task(rq))
2111 ;
2112}
2113
2114static void pull_dl_task(struct rq *this_rq)
2115{
2116 int this_cpu = this_rq->cpu, cpu;
2117 struct task_struct *p;
2118 bool resched = false;
2119 struct rq *src_rq;
2120 u64 dmin = LONG_MAX;
2121
2122 if (likely(!dl_overloaded(this_rq)))
2123 return;
2124
2125 /*
2126 * Match the barrier from dl_set_overloaded; this guarantees that if we
2127 * see overloaded we must also see the dlo_mask bit.
2128 */
2129 smp_rmb();
2130
2131 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2132 if (this_cpu == cpu)
2133 continue;
2134
2135 src_rq = cpu_rq(cpu);
2136
2137 /*
2138 * It looks racy, abd it is! However, as in sched_rt.c,
2139 * we are fine with this.
2140 */
2141 if (this_rq->dl.dl_nr_running &&
2142 dl_time_before(this_rq->dl.earliest_dl.curr,
2143 src_rq->dl.earliest_dl.next))
2144 continue;
2145
2146 /* Might drop this_rq->lock */
2147 double_lock_balance(this_rq, src_rq);
2148
2149 /*
2150 * If there are no more pullable tasks on the
2151 * rq, we're done with it.
2152 */
2153 if (src_rq->dl.dl_nr_running <= 1)
2154 goto skip;
2155
2156 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2157
2158 /*
2159 * We found a task to be pulled if:
2160 * - it preempts our current (if there's one),
2161 * - it will preempt the last one we pulled (if any).
2162 */
2163 if (p && dl_time_before(p->dl.deadline, dmin) &&
2164 (!this_rq->dl.dl_nr_running ||
2165 dl_time_before(p->dl.deadline,
2166 this_rq->dl.earliest_dl.curr))) {
2167 WARN_ON(p == src_rq->curr);
2168 WARN_ON(!task_on_rq_queued(p));
2169
2170 /*
2171 * Then we pull iff p has actually an earlier
2172 * deadline than the current task of its runqueue.
2173 */
2174 if (dl_time_before(p->dl.deadline,
2175 src_rq->curr->dl.deadline))
2176 goto skip;
2177
2178 resched = true;
2179
2180 deactivate_task(src_rq, p, 0);
2181 sub_running_bw(&p->dl, &src_rq->dl);
2182 sub_rq_bw(&p->dl, &src_rq->dl);
2183 set_task_cpu(p, this_cpu);
2184 add_rq_bw(&p->dl, &this_rq->dl);
2185 add_running_bw(&p->dl, &this_rq->dl);
2186 activate_task(this_rq, p, 0);
2187 dmin = p->dl.deadline;
2188
2189 /* Is there any other task even earlier? */
2190 }
2191skip:
2192 double_unlock_balance(this_rq, src_rq);
2193 }
2194
2195 if (resched)
2196 resched_curr(this_rq);
2197}
2198
2199/*
2200 * Since the task is not running and a reschedule is not going to happen
2201 * anytime soon on its runqueue, we try pushing it away now.
2202 */
2203static void task_woken_dl(struct rq *rq, struct task_struct *p)
2204{
2205 if (!task_running(rq, p) &&
2206 !test_tsk_need_resched(rq->curr) &&
2207 p->nr_cpus_allowed > 1 &&
2208 dl_task(rq->curr) &&
2209 (rq->curr->nr_cpus_allowed < 2 ||
2210 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2211 push_dl_tasks(rq);
2212 }
2213}
2214
2215static void set_cpus_allowed_dl(struct task_struct *p,
2216 const struct cpumask *new_mask)
2217{
2218 struct root_domain *src_rd;
2219 struct rq *rq;
2220
2221 BUG_ON(!dl_task(p));
2222
2223 rq = task_rq(p);
2224 src_rd = rq->rd;
2225 /*
2226 * Migrating a SCHED_DEADLINE task between exclusive
2227 * cpusets (different root_domains) entails a bandwidth
2228 * update. We already made space for us in the destination
2229 * domain (see cpuset_can_attach()).
2230 */
2231 if (!cpumask_intersects(src_rd->span, new_mask)) {
2232 struct dl_bw *src_dl_b;
2233
2234 src_dl_b = dl_bw_of(cpu_of(rq));
2235 /*
2236 * We now free resources of the root_domain we are migrating
2237 * off. In the worst case, sched_setattr() may temporary fail
2238 * until we complete the update.
2239 */
2240 raw_spin_lock(&src_dl_b->lock);
2241 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2242 raw_spin_unlock(&src_dl_b->lock);
2243 }
2244
2245 set_cpus_allowed_common(p, new_mask);
2246}
2247
2248/* Assumes rq->lock is held */
2249static void rq_online_dl(struct rq *rq)
2250{
2251 if (rq->dl.overloaded)
2252 dl_set_overload(rq);
2253
2254 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2255 if (rq->dl.dl_nr_running > 0)
2256 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2257}
2258
2259/* Assumes rq->lock is held */
2260static void rq_offline_dl(struct rq *rq)
2261{
2262 if (rq->dl.overloaded)
2263 dl_clear_overload(rq);
2264
2265 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2266 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2267}
2268
2269void __init init_sched_dl_class(void)
2270{
2271 unsigned int i;
2272
2273 for_each_possible_cpu(i)
2274 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2275 GFP_KERNEL, cpu_to_node(i));
2276}
2277
2278#endif /* CONFIG_SMP */
2279
2280static void switched_from_dl(struct rq *rq, struct task_struct *p)
2281{
2282 /*
2283 * task_non_contending() can start the "inactive timer" (if the 0-lag
2284 * time is in the future). If the task switches back to dl before
2285 * the "inactive timer" fires, it can continue to consume its current
2286 * runtime using its current deadline. If it stays outside of
2287 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2288 * will reset the task parameters.
2289 */
2290 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2291 task_non_contending(p);
2292
2293 if (!task_on_rq_queued(p))
2294 sub_rq_bw(&p->dl, &rq->dl);
2295
2296 /*
2297 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2298 * at the 0-lag time, because the task could have been migrated
2299 * while SCHED_OTHER in the meanwhile.
2300 */
2301 if (p->dl.dl_non_contending)
2302 p->dl.dl_non_contending = 0;
2303
2304 /*
2305 * Since this might be the only -deadline task on the rq,
2306 * this is the right place to try to pull some other one
2307 * from an overloaded CPU, if any.
2308 */
2309 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2310 return;
2311
2312 deadline_queue_pull_task(rq);
2313}
2314
2315/*
2316 * When switching to -deadline, we may overload the rq, then
2317 * we try to push someone off, if possible.
2318 */
2319static void switched_to_dl(struct rq *rq, struct task_struct *p)
2320{
2321 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2322 put_task_struct(p);
2323
2324 /* If p is not queued we will update its parameters at next wakeup. */
2325 if (!task_on_rq_queued(p)) {
2326 add_rq_bw(&p->dl, &rq->dl);
2327
2328 return;
2329 }
2330
2331 if (rq->curr != p) {
2332#ifdef CONFIG_SMP
2333 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2334 deadline_queue_push_tasks(rq);
2335#endif
2336 if (dl_task(rq->curr))
2337 check_preempt_curr_dl(rq, p, 0);
2338 else
2339 resched_curr(rq);
2340 }
2341}
2342
2343/*
2344 * If the scheduling parameters of a -deadline task changed,
2345 * a push or pull operation might be needed.
2346 */
2347static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2348 int oldprio)
2349{
2350 if (task_on_rq_queued(p) || rq->curr == p) {
2351#ifdef CONFIG_SMP
2352 /*
2353 * This might be too much, but unfortunately
2354 * we don't have the old deadline value, and
2355 * we can't argue if the task is increasing
2356 * or lowering its prio, so...
2357 */
2358 if (!rq->dl.overloaded)
2359 deadline_queue_pull_task(rq);
2360
2361 /*
2362 * If we now have a earlier deadline task than p,
2363 * then reschedule, provided p is still on this
2364 * runqueue.
2365 */
2366 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2367 resched_curr(rq);
2368#else
2369 /*
2370 * Again, we don't know if p has a earlier
2371 * or later deadline, so let's blindly set a
2372 * (maybe not needed) rescheduling point.
2373 */
2374 resched_curr(rq);
2375#endif /* CONFIG_SMP */
2376 }
2377}
2378
2379const struct sched_class dl_sched_class = {
2380 .next = &rt_sched_class,
2381 .enqueue_task = enqueue_task_dl,
2382 .dequeue_task = dequeue_task_dl,
2383 .yield_task = yield_task_dl,
2384
2385 .check_preempt_curr = check_preempt_curr_dl,
2386
2387 .pick_next_task = pick_next_task_dl,
2388 .put_prev_task = put_prev_task_dl,
2389
2390#ifdef CONFIG_SMP
2391 .select_task_rq = select_task_rq_dl,
2392 .migrate_task_rq = migrate_task_rq_dl,
2393 .set_cpus_allowed = set_cpus_allowed_dl,
2394 .rq_online = rq_online_dl,
2395 .rq_offline = rq_offline_dl,
2396 .task_woken = task_woken_dl,
2397#endif
2398
2399 .set_curr_task = set_curr_task_dl,
2400 .task_tick = task_tick_dl,
2401 .task_fork = task_fork_dl,
2402
2403 .prio_changed = prio_changed_dl,
2404 .switched_from = switched_from_dl,
2405 .switched_to = switched_to_dl,
2406
2407 .update_curr = update_curr_dl,
2408};
2409
2410int sched_dl_global_validate(void)
2411{
2412 u64 runtime = global_rt_runtime();
2413 u64 period = global_rt_period();
2414 u64 new_bw = to_ratio(period, runtime);
2415 struct dl_bw *dl_b;
2416 int cpu, ret = 0;
2417 unsigned long flags;
2418
2419 /*
2420 * Here we want to check the bandwidth not being set to some
2421 * value smaller than the currently allocated bandwidth in
2422 * any of the root_domains.
2423 *
2424 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2425 * cycling on root_domains... Discussion on different/better
2426 * solutions is welcome!
2427 */
2428 for_each_possible_cpu(cpu) {
2429 rcu_read_lock_sched();
2430 dl_b = dl_bw_of(cpu);
2431
2432 raw_spin_lock_irqsave(&dl_b->lock, flags);
2433 if (new_bw < dl_b->total_bw)
2434 ret = -EBUSY;
2435 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2436
2437 rcu_read_unlock_sched();
2438
2439 if (ret)
2440 break;
2441 }
2442
2443 return ret;
2444}
2445
2446void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2447{
2448 if (global_rt_runtime() == RUNTIME_INF) {
2449 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2450 dl_rq->extra_bw = 1 << BW_SHIFT;
2451 } else {
2452 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2453 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2454 dl_rq->extra_bw = to_ratio(global_rt_period(),
2455 global_rt_runtime());
2456 }
2457}
2458
2459void sched_dl_do_global(void)
2460{
2461 u64 new_bw = -1;
2462 struct dl_bw *dl_b;
2463 int cpu;
2464 unsigned long flags;
2465
2466 def_dl_bandwidth.dl_period = global_rt_period();
2467 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2468
2469 if (global_rt_runtime() != RUNTIME_INF)
2470 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2471
2472 /*
2473 * FIXME: As above...
2474 */
2475 for_each_possible_cpu(cpu) {
2476 rcu_read_lock_sched();
2477 dl_b = dl_bw_of(cpu);
2478
2479 raw_spin_lock_irqsave(&dl_b->lock, flags);
2480 dl_b->bw = new_bw;
2481 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2482
2483 rcu_read_unlock_sched();
2484 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2485 }
2486}
2487
2488/*
2489 * We must be sure that accepting a new task (or allowing changing the
2490 * parameters of an existing one) is consistent with the bandwidth
2491 * constraints. If yes, this function also accordingly updates the currently
2492 * allocated bandwidth to reflect the new situation.
2493 *
2494 * This function is called while holding p's rq->lock.
2495 */
2496int sched_dl_overflow(struct task_struct *p, int policy,
2497 const struct sched_attr *attr)
2498{
2499 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2500 u64 period = attr->sched_period ?: attr->sched_deadline;
2501 u64 runtime = attr->sched_runtime;
2502 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2503 int cpus, err = -1;
2504
2505 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2506 return 0;
2507
2508 /* !deadline task may carry old deadline bandwidth */
2509 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2510 return 0;
2511
2512 /*
2513 * Either if a task, enters, leave, or stays -deadline but changes
2514 * its parameters, we may need to update accordingly the total
2515 * allocated bandwidth of the container.
2516 */
2517 raw_spin_lock(&dl_b->lock);
2518 cpus = dl_bw_cpus(task_cpu(p));
2519 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2520 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2521 if (hrtimer_active(&p->dl.inactive_timer))
2522 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2523 __dl_add(dl_b, new_bw, cpus);
2524 err = 0;
2525 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2526 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2527 /*
2528 * XXX this is slightly incorrect: when the task
2529 * utilization decreases, we should delay the total
2530 * utilization change until the task's 0-lag point.
2531 * But this would require to set the task's "inactive
2532 * timer" when the task is not inactive.
2533 */
2534 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2535 __dl_add(dl_b, new_bw, cpus);
2536 dl_change_utilization(p, new_bw);
2537 err = 0;
2538 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2539 /*
2540 * Do not decrease the total deadline utilization here,
2541 * switched_from_dl() will take care to do it at the correct
2542 * (0-lag) time.
2543 */
2544 err = 0;
2545 }
2546 raw_spin_unlock(&dl_b->lock);
2547
2548 return err;
2549}
2550
2551/*
2552 * This function initializes the sched_dl_entity of a newly becoming
2553 * SCHED_DEADLINE task.
2554 *
2555 * Only the static values are considered here, the actual runtime and the
2556 * absolute deadline will be properly calculated when the task is enqueued
2557 * for the first time with its new policy.
2558 */
2559void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2560{
2561 struct sched_dl_entity *dl_se = &p->dl;
2562
2563 dl_se->dl_runtime = attr->sched_runtime;
2564 dl_se->dl_deadline = attr->sched_deadline;
2565 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2566 dl_se->flags = attr->sched_flags;
2567 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2568 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2569}
2570
2571void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2572{
2573 struct sched_dl_entity *dl_se = &p->dl;
2574
2575 attr->sched_priority = p->rt_priority;
2576 attr->sched_runtime = dl_se->dl_runtime;
2577 attr->sched_deadline = dl_se->dl_deadline;
2578 attr->sched_period = dl_se->dl_period;
2579 attr->sched_flags = dl_se->flags;
2580}
2581
2582/*
2583 * This function validates the new parameters of a -deadline task.
2584 * We ask for the deadline not being zero, and greater or equal
2585 * than the runtime, as well as the period of being zero or
2586 * greater than deadline. Furthermore, we have to be sure that
2587 * user parameters are above the internal resolution of 1us (we
2588 * check sched_runtime only since it is always the smaller one) and
2589 * below 2^63 ns (we have to check both sched_deadline and
2590 * sched_period, as the latter can be zero).
2591 */
2592bool __checkparam_dl(const struct sched_attr *attr)
2593{
2594 /* special dl tasks don't actually use any parameter */
2595 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2596 return true;
2597
2598 /* deadline != 0 */
2599 if (attr->sched_deadline == 0)
2600 return false;
2601
2602 /*
2603 * Since we truncate DL_SCALE bits, make sure we're at least
2604 * that big.
2605 */
2606 if (attr->sched_runtime < (1ULL << DL_SCALE))
2607 return false;
2608
2609 /*
2610 * Since we use the MSB for wrap-around and sign issues, make
2611 * sure it's not set (mind that period can be equal to zero).
2612 */
2613 if (attr->sched_deadline & (1ULL << 63) ||
2614 attr->sched_period & (1ULL << 63))
2615 return false;
2616
2617 /* runtime <= deadline <= period (if period != 0) */
2618 if ((attr->sched_period != 0 &&
2619 attr->sched_period < attr->sched_deadline) ||
2620 attr->sched_deadline < attr->sched_runtime)
2621 return false;
2622
2623 return true;
2624}
2625
2626/*
2627 * This function clears the sched_dl_entity static params.
2628 */
2629void __dl_clear_params(struct task_struct *p)
2630{
2631 struct sched_dl_entity *dl_se = &p->dl;
2632
2633 dl_se->dl_runtime = 0;
2634 dl_se->dl_deadline = 0;
2635 dl_se->dl_period = 0;
2636 dl_se->flags = 0;
2637 dl_se->dl_bw = 0;
2638 dl_se->dl_density = 0;
2639
2640 dl_se->dl_throttled = 0;
2641 dl_se->dl_yielded = 0;
2642 dl_se->dl_non_contending = 0;
2643 dl_se->dl_overrun = 0;
2644}
2645
2646bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2647{
2648 struct sched_dl_entity *dl_se = &p->dl;
2649
2650 if (dl_se->dl_runtime != attr->sched_runtime ||
2651 dl_se->dl_deadline != attr->sched_deadline ||
2652 dl_se->dl_period != attr->sched_period ||
2653 dl_se->flags != attr->sched_flags)
2654 return true;
2655
2656 return false;
2657}
2658
2659#ifdef CONFIG_SMP
2660int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2661{
2662 unsigned int dest_cpu;
2663 struct dl_bw *dl_b;
2664 bool overflow;
2665 int cpus, ret;
2666 unsigned long flags;
2667
2668 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2669
2670 rcu_read_lock_sched();
2671 dl_b = dl_bw_of(dest_cpu);
2672 raw_spin_lock_irqsave(&dl_b->lock, flags);
2673 cpus = dl_bw_cpus(dest_cpu);
2674 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
2675 if (overflow) {
2676 ret = -EBUSY;
2677 } else {
2678 /*
2679 * We reserve space for this task in the destination
2680 * root_domain, as we can't fail after this point.
2681 * We will free resources in the source root_domain
2682 * later on (see set_cpus_allowed_dl()).
2683 */
2684 __dl_add(dl_b, p->dl.dl_bw, cpus);
2685 ret = 0;
2686 }
2687 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2688 rcu_read_unlock_sched();
2689
2690 return ret;
2691}
2692
2693int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2694 const struct cpumask *trial)
2695{
2696 int ret = 1, trial_cpus;
2697 struct dl_bw *cur_dl_b;
2698 unsigned long flags;
2699
2700 rcu_read_lock_sched();
2701 cur_dl_b = dl_bw_of(cpumask_any(cur));
2702 trial_cpus = cpumask_weight(trial);
2703
2704 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2705 if (cur_dl_b->bw != -1 &&
2706 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2707 ret = 0;
2708 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2709 rcu_read_unlock_sched();
2710
2711 return ret;
2712}
2713
2714bool dl_cpu_busy(unsigned int cpu)
2715{
2716 unsigned long flags;
2717 struct dl_bw *dl_b;
2718 bool overflow;
2719 int cpus;
2720
2721 rcu_read_lock_sched();
2722 dl_b = dl_bw_of(cpu);
2723 raw_spin_lock_irqsave(&dl_b->lock, flags);
2724 cpus = dl_bw_cpus(cpu);
2725 overflow = __dl_overflow(dl_b, cpus, 0, 0);
2726 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2727 rcu_read_unlock_sched();
2728
2729 return overflow;
2730}
2731#endif
2732
2733#ifdef CONFIG_SCHED_DEBUG
2734void print_dl_stats(struct seq_file *m, int cpu)
2735{
2736 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2737}
2738#endif /* CONFIG_SCHED_DEBUG */
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
21struct dl_bandwidth def_dl_bandwidth;
22
23static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
24{
25 return container_of(dl_se, struct task_struct, dl);
26}
27
28static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
29{
30 return container_of(dl_rq, struct rq, dl);
31}
32
33static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
34{
35 struct task_struct *p = dl_task_of(dl_se);
36 struct rq *rq = task_rq(p);
37
38 return &rq->dl;
39}
40
41static inline int on_dl_rq(struct sched_dl_entity *dl_se)
42{
43 return !RB_EMPTY_NODE(&dl_se->rb_node);
44}
45
46static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
47{
48 struct sched_dl_entity *dl_se = &p->dl;
49
50 return dl_rq->rb_leftmost == &dl_se->rb_node;
51}
52
53void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
54{
55 raw_spin_lock_init(&dl_b->dl_runtime_lock);
56 dl_b->dl_period = period;
57 dl_b->dl_runtime = runtime;
58}
59
60void init_dl_bw(struct dl_bw *dl_b)
61{
62 raw_spin_lock_init(&dl_b->lock);
63 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
64 if (global_rt_runtime() == RUNTIME_INF)
65 dl_b->bw = -1;
66 else
67 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
68 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
69 dl_b->total_bw = 0;
70}
71
72void init_dl_rq(struct dl_rq *dl_rq)
73{
74 dl_rq->rb_root = RB_ROOT;
75
76#ifdef CONFIG_SMP
77 /* zero means no -deadline tasks */
78 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
79
80 dl_rq->dl_nr_migratory = 0;
81 dl_rq->overloaded = 0;
82 dl_rq->pushable_dl_tasks_root = RB_ROOT;
83#else
84 init_dl_bw(&dl_rq->dl_bw);
85#endif
86}
87
88#ifdef CONFIG_SMP
89
90static inline int dl_overloaded(struct rq *rq)
91{
92 return atomic_read(&rq->rd->dlo_count);
93}
94
95static inline void dl_set_overload(struct rq *rq)
96{
97 if (!rq->online)
98 return;
99
100 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
101 /*
102 * Must be visible before the overload count is
103 * set (as in sched_rt.c).
104 *
105 * Matched by the barrier in pull_dl_task().
106 */
107 smp_wmb();
108 atomic_inc(&rq->rd->dlo_count);
109}
110
111static inline void dl_clear_overload(struct rq *rq)
112{
113 if (!rq->online)
114 return;
115
116 atomic_dec(&rq->rd->dlo_count);
117 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
118}
119
120static void update_dl_migration(struct dl_rq *dl_rq)
121{
122 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
123 if (!dl_rq->overloaded) {
124 dl_set_overload(rq_of_dl_rq(dl_rq));
125 dl_rq->overloaded = 1;
126 }
127 } else if (dl_rq->overloaded) {
128 dl_clear_overload(rq_of_dl_rq(dl_rq));
129 dl_rq->overloaded = 0;
130 }
131}
132
133static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
134{
135 struct task_struct *p = dl_task_of(dl_se);
136
137 if (tsk_nr_cpus_allowed(p) > 1)
138 dl_rq->dl_nr_migratory++;
139
140 update_dl_migration(dl_rq);
141}
142
143static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
144{
145 struct task_struct *p = dl_task_of(dl_se);
146
147 if (tsk_nr_cpus_allowed(p) > 1)
148 dl_rq->dl_nr_migratory--;
149
150 update_dl_migration(dl_rq);
151}
152
153/*
154 * The list of pushable -deadline task is not a plist, like in
155 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
156 */
157static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
158{
159 struct dl_rq *dl_rq = &rq->dl;
160 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node;
161 struct rb_node *parent = NULL;
162 struct task_struct *entry;
163 int leftmost = 1;
164
165 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
166
167 while (*link) {
168 parent = *link;
169 entry = rb_entry(parent, struct task_struct,
170 pushable_dl_tasks);
171 if (dl_entity_preempt(&p->dl, &entry->dl))
172 link = &parent->rb_left;
173 else {
174 link = &parent->rb_right;
175 leftmost = 0;
176 }
177 }
178
179 if (leftmost) {
180 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
181 dl_rq->earliest_dl.next = p->dl.deadline;
182 }
183
184 rb_link_node(&p->pushable_dl_tasks, parent, link);
185 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
186}
187
188static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
189{
190 struct dl_rq *dl_rq = &rq->dl;
191
192 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
193 return;
194
195 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
196 struct rb_node *next_node;
197
198 next_node = rb_next(&p->pushable_dl_tasks);
199 dl_rq->pushable_dl_tasks_leftmost = next_node;
200 if (next_node) {
201 dl_rq->earliest_dl.next = rb_entry(next_node,
202 struct task_struct, pushable_dl_tasks)->dl.deadline;
203 }
204 }
205
206 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
207 RB_CLEAR_NODE(&p->pushable_dl_tasks);
208}
209
210static inline int has_pushable_dl_tasks(struct rq *rq)
211{
212 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
213}
214
215static int push_dl_task(struct rq *rq);
216
217static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
218{
219 return dl_task(prev);
220}
221
222static DEFINE_PER_CPU(struct callback_head, dl_push_head);
223static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
224
225static void push_dl_tasks(struct rq *);
226static void pull_dl_task(struct rq *);
227
228static inline void queue_push_tasks(struct rq *rq)
229{
230 if (!has_pushable_dl_tasks(rq))
231 return;
232
233 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
234}
235
236static inline void queue_pull_task(struct rq *rq)
237{
238 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
239}
240
241static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
242
243static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
244{
245 struct rq *later_rq = NULL;
246
247 later_rq = find_lock_later_rq(p, rq);
248 if (!later_rq) {
249 int cpu;
250
251 /*
252 * If we cannot preempt any rq, fall back to pick any
253 * online cpu.
254 */
255 cpu = cpumask_any_and(cpu_active_mask, tsk_cpus_allowed(p));
256 if (cpu >= nr_cpu_ids) {
257 /*
258 * Fail to find any suitable cpu.
259 * The task will never come back!
260 */
261 BUG_ON(dl_bandwidth_enabled());
262
263 /*
264 * If admission control is disabled we
265 * try a little harder to let the task
266 * run.
267 */
268 cpu = cpumask_any(cpu_active_mask);
269 }
270 later_rq = cpu_rq(cpu);
271 double_lock_balance(rq, later_rq);
272 }
273
274 set_task_cpu(p, later_rq->cpu);
275 double_unlock_balance(later_rq, rq);
276
277 return later_rq;
278}
279
280#else
281
282static inline
283void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
284{
285}
286
287static inline
288void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
289{
290}
291
292static inline
293void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
294{
295}
296
297static inline
298void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
299{
300}
301
302static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
303{
304 return false;
305}
306
307static inline void pull_dl_task(struct rq *rq)
308{
309}
310
311static inline void queue_push_tasks(struct rq *rq)
312{
313}
314
315static inline void queue_pull_task(struct rq *rq)
316{
317}
318#endif /* CONFIG_SMP */
319
320static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
321static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
322static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
323 int flags);
324
325/*
326 * We are being explicitly informed that a new instance is starting,
327 * and this means that:
328 * - the absolute deadline of the entity has to be placed at
329 * current time + relative deadline;
330 * - the runtime of the entity has to be set to the maximum value.
331 *
332 * The capability of specifying such event is useful whenever a -deadline
333 * entity wants to (try to!) synchronize its behaviour with the scheduler's
334 * one, and to (try to!) reconcile itself with its own scheduling
335 * parameters.
336 */
337static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
338{
339 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
340 struct rq *rq = rq_of_dl_rq(dl_rq);
341
342 WARN_ON(dl_se->dl_boosted);
343 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
344
345 /*
346 * We are racing with the deadline timer. So, do nothing because
347 * the deadline timer handler will take care of properly recharging
348 * the runtime and postponing the deadline
349 */
350 if (dl_se->dl_throttled)
351 return;
352
353 /*
354 * We use the regular wall clock time to set deadlines in the
355 * future; in fact, we must consider execution overheads (time
356 * spent on hardirq context, etc.).
357 */
358 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
359 dl_se->runtime = dl_se->dl_runtime;
360}
361
362/*
363 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
364 * possibility of a entity lasting more than what it declared, and thus
365 * exhausting its runtime.
366 *
367 * Here we are interested in making runtime overrun possible, but we do
368 * not want a entity which is misbehaving to affect the scheduling of all
369 * other entities.
370 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
371 * is used, in order to confine each entity within its own bandwidth.
372 *
373 * This function deals exactly with that, and ensures that when the runtime
374 * of a entity is replenished, its deadline is also postponed. That ensures
375 * the overrunning entity can't interfere with other entity in the system and
376 * can't make them miss their deadlines. Reasons why this kind of overruns
377 * could happen are, typically, a entity voluntarily trying to overcome its
378 * runtime, or it just underestimated it during sched_setattr().
379 */
380static void replenish_dl_entity(struct sched_dl_entity *dl_se,
381 struct sched_dl_entity *pi_se)
382{
383 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
384 struct rq *rq = rq_of_dl_rq(dl_rq);
385
386 BUG_ON(pi_se->dl_runtime <= 0);
387
388 /*
389 * This could be the case for a !-dl task that is boosted.
390 * Just go with full inherited parameters.
391 */
392 if (dl_se->dl_deadline == 0) {
393 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
394 dl_se->runtime = pi_se->dl_runtime;
395 }
396
397 if (dl_se->dl_yielded && dl_se->runtime > 0)
398 dl_se->runtime = 0;
399
400 /*
401 * We keep moving the deadline away until we get some
402 * available runtime for the entity. This ensures correct
403 * handling of situations where the runtime overrun is
404 * arbitrary large.
405 */
406 while (dl_se->runtime <= 0) {
407 dl_se->deadline += pi_se->dl_period;
408 dl_se->runtime += pi_se->dl_runtime;
409 }
410
411 /*
412 * At this point, the deadline really should be "in
413 * the future" with respect to rq->clock. If it's
414 * not, we are, for some reason, lagging too much!
415 * Anyway, after having warn userspace abut that,
416 * we still try to keep the things running by
417 * resetting the deadline and the budget of the
418 * entity.
419 */
420 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
421 printk_deferred_once("sched: DL replenish lagged too much\n");
422 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
423 dl_se->runtime = pi_se->dl_runtime;
424 }
425
426 if (dl_se->dl_yielded)
427 dl_se->dl_yielded = 0;
428 if (dl_se->dl_throttled)
429 dl_se->dl_throttled = 0;
430}
431
432/*
433 * Here we check if --at time t-- an entity (which is probably being
434 * [re]activated or, in general, enqueued) can use its remaining runtime
435 * and its current deadline _without_ exceeding the bandwidth it is
436 * assigned (function returns true if it can't). We are in fact applying
437 * one of the CBS rules: when a task wakes up, if the residual runtime
438 * over residual deadline fits within the allocated bandwidth, then we
439 * can keep the current (absolute) deadline and residual budget without
440 * disrupting the schedulability of the system. Otherwise, we should
441 * refill the runtime and set the deadline a period in the future,
442 * because keeping the current (absolute) deadline of the task would
443 * result in breaking guarantees promised to other tasks (refer to
444 * Documentation/scheduler/sched-deadline.txt for more informations).
445 *
446 * This function returns true if:
447 *
448 * runtime / (deadline - t) > dl_runtime / dl_period ,
449 *
450 * IOW we can't recycle current parameters.
451 *
452 * Notice that the bandwidth check is done against the period. For
453 * task with deadline equal to period this is the same of using
454 * dl_deadline instead of dl_period in the equation above.
455 */
456static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
457 struct sched_dl_entity *pi_se, u64 t)
458{
459 u64 left, right;
460
461 /*
462 * left and right are the two sides of the equation above,
463 * after a bit of shuffling to use multiplications instead
464 * of divisions.
465 *
466 * Note that none of the time values involved in the two
467 * multiplications are absolute: dl_deadline and dl_runtime
468 * are the relative deadline and the maximum runtime of each
469 * instance, runtime is the runtime left for the last instance
470 * and (deadline - t), since t is rq->clock, is the time left
471 * to the (absolute) deadline. Even if overflowing the u64 type
472 * is very unlikely to occur in both cases, here we scale down
473 * as we want to avoid that risk at all. Scaling down by 10
474 * means that we reduce granularity to 1us. We are fine with it,
475 * since this is only a true/false check and, anyway, thinking
476 * of anything below microseconds resolution is actually fiction
477 * (but still we want to give the user that illusion >;).
478 */
479 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
480 right = ((dl_se->deadline - t) >> DL_SCALE) *
481 (pi_se->dl_runtime >> DL_SCALE);
482
483 return dl_time_before(right, left);
484}
485
486/*
487 * When a -deadline entity is queued back on the runqueue, its runtime and
488 * deadline might need updating.
489 *
490 * The policy here is that we update the deadline of the entity only if:
491 * - the current deadline is in the past,
492 * - using the remaining runtime with the current deadline would make
493 * the entity exceed its bandwidth.
494 */
495static void update_dl_entity(struct sched_dl_entity *dl_se,
496 struct sched_dl_entity *pi_se)
497{
498 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
499 struct rq *rq = rq_of_dl_rq(dl_rq);
500
501 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
502 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
503 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
504 dl_se->runtime = pi_se->dl_runtime;
505 }
506}
507
508/*
509 * If the entity depleted all its runtime, and if we want it to sleep
510 * while waiting for some new execution time to become available, we
511 * set the bandwidth enforcement timer to the replenishment instant
512 * and try to activate it.
513 *
514 * Notice that it is important for the caller to know if the timer
515 * actually started or not (i.e., the replenishment instant is in
516 * the future or in the past).
517 */
518static int start_dl_timer(struct task_struct *p)
519{
520 struct sched_dl_entity *dl_se = &p->dl;
521 struct hrtimer *timer = &dl_se->dl_timer;
522 struct rq *rq = task_rq(p);
523 ktime_t now, act;
524 s64 delta;
525
526 lockdep_assert_held(&rq->lock);
527
528 /*
529 * We want the timer to fire at the deadline, but considering
530 * that it is actually coming from rq->clock and not from
531 * hrtimer's time base reading.
532 */
533 act = ns_to_ktime(dl_se->deadline);
534 now = hrtimer_cb_get_time(timer);
535 delta = ktime_to_ns(now) - rq_clock(rq);
536 act = ktime_add_ns(act, delta);
537
538 /*
539 * If the expiry time already passed, e.g., because the value
540 * chosen as the deadline is too small, don't even try to
541 * start the timer in the past!
542 */
543 if (ktime_us_delta(act, now) < 0)
544 return 0;
545
546 /*
547 * !enqueued will guarantee another callback; even if one is already in
548 * progress. This ensures a balanced {get,put}_task_struct().
549 *
550 * The race against __run_timer() clearing the enqueued state is
551 * harmless because we're holding task_rq()->lock, therefore the timer
552 * expiring after we've done the check will wait on its task_rq_lock()
553 * and observe our state.
554 */
555 if (!hrtimer_is_queued(timer)) {
556 get_task_struct(p);
557 hrtimer_start(timer, act, HRTIMER_MODE_ABS);
558 }
559
560 return 1;
561}
562
563/*
564 * This is the bandwidth enforcement timer callback. If here, we know
565 * a task is not on its dl_rq, since the fact that the timer was running
566 * means the task is throttled and needs a runtime replenishment.
567 *
568 * However, what we actually do depends on the fact the task is active,
569 * (it is on its rq) or has been removed from there by a call to
570 * dequeue_task_dl(). In the former case we must issue the runtime
571 * replenishment and add the task back to the dl_rq; in the latter, we just
572 * do nothing but clearing dl_throttled, so that runtime and deadline
573 * updating (and the queueing back to dl_rq) will be done by the
574 * next call to enqueue_task_dl().
575 */
576static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
577{
578 struct sched_dl_entity *dl_se = container_of(timer,
579 struct sched_dl_entity,
580 dl_timer);
581 struct task_struct *p = dl_task_of(dl_se);
582 struct rq_flags rf;
583 struct rq *rq;
584
585 rq = task_rq_lock(p, &rf);
586
587 /*
588 * The task might have changed its scheduling policy to something
589 * different than SCHED_DEADLINE (through switched_from_dl()).
590 */
591 if (!dl_task(p)) {
592 __dl_clear_params(p);
593 goto unlock;
594 }
595
596 /*
597 * The task might have been boosted by someone else and might be in the
598 * boosting/deboosting path, its not throttled.
599 */
600 if (dl_se->dl_boosted)
601 goto unlock;
602
603 /*
604 * Spurious timer due to start_dl_timer() race; or we already received
605 * a replenishment from rt_mutex_setprio().
606 */
607 if (!dl_se->dl_throttled)
608 goto unlock;
609
610 sched_clock_tick();
611 update_rq_clock(rq);
612
613 /*
614 * If the throttle happened during sched-out; like:
615 *
616 * schedule()
617 * deactivate_task()
618 * dequeue_task_dl()
619 * update_curr_dl()
620 * start_dl_timer()
621 * __dequeue_task_dl()
622 * prev->on_rq = 0;
623 *
624 * We can be both throttled and !queued. Replenish the counter
625 * but do not enqueue -- wait for our wakeup to do that.
626 */
627 if (!task_on_rq_queued(p)) {
628 replenish_dl_entity(dl_se, dl_se);
629 goto unlock;
630 }
631
632#ifdef CONFIG_SMP
633 if (unlikely(!rq->online)) {
634 /*
635 * If the runqueue is no longer available, migrate the
636 * task elsewhere. This necessarily changes rq.
637 */
638 lockdep_unpin_lock(&rq->lock, rf.cookie);
639 rq = dl_task_offline_migration(rq, p);
640 rf.cookie = lockdep_pin_lock(&rq->lock);
641
642 /*
643 * Now that the task has been migrated to the new RQ and we
644 * have that locked, proceed as normal and enqueue the task
645 * there.
646 */
647 }
648#endif
649
650 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
651 if (dl_task(rq->curr))
652 check_preempt_curr_dl(rq, p, 0);
653 else
654 resched_curr(rq);
655
656#ifdef CONFIG_SMP
657 /*
658 * Queueing this task back might have overloaded rq, check if we need
659 * to kick someone away.
660 */
661 if (has_pushable_dl_tasks(rq)) {
662 /*
663 * Nothing relies on rq->lock after this, so its safe to drop
664 * rq->lock.
665 */
666 lockdep_unpin_lock(&rq->lock, rf.cookie);
667 push_dl_task(rq);
668 lockdep_repin_lock(&rq->lock, rf.cookie);
669 }
670#endif
671
672unlock:
673 task_rq_unlock(rq, p, &rf);
674
675 /*
676 * This can free the task_struct, including this hrtimer, do not touch
677 * anything related to that after this.
678 */
679 put_task_struct(p);
680
681 return HRTIMER_NORESTART;
682}
683
684void init_dl_task_timer(struct sched_dl_entity *dl_se)
685{
686 struct hrtimer *timer = &dl_se->dl_timer;
687
688 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
689 timer->function = dl_task_timer;
690}
691
692static
693int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
694{
695 return (dl_se->runtime <= 0);
696}
697
698extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
699
700/*
701 * Update the current task's runtime statistics (provided it is still
702 * a -deadline task and has not been removed from the dl_rq).
703 */
704static void update_curr_dl(struct rq *rq)
705{
706 struct task_struct *curr = rq->curr;
707 struct sched_dl_entity *dl_se = &curr->dl;
708 u64 delta_exec;
709
710 if (!dl_task(curr) || !on_dl_rq(dl_se))
711 return;
712
713 /*
714 * Consumed budget is computed considering the time as
715 * observed by schedulable tasks (excluding time spent
716 * in hardirq context, etc.). Deadlines are instead
717 * computed using hard walltime. This seems to be the more
718 * natural solution, but the full ramifications of this
719 * approach need further study.
720 */
721 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
722 if (unlikely((s64)delta_exec <= 0)) {
723 if (unlikely(dl_se->dl_yielded))
724 goto throttle;
725 return;
726 }
727
728 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
729 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_DL);
730
731 schedstat_set(curr->se.statistics.exec_max,
732 max(curr->se.statistics.exec_max, delta_exec));
733
734 curr->se.sum_exec_runtime += delta_exec;
735 account_group_exec_runtime(curr, delta_exec);
736
737 curr->se.exec_start = rq_clock_task(rq);
738 cpuacct_charge(curr, delta_exec);
739
740 sched_rt_avg_update(rq, delta_exec);
741
742 dl_se->runtime -= delta_exec;
743
744throttle:
745 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
746 dl_se->dl_throttled = 1;
747 __dequeue_task_dl(rq, curr, 0);
748 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
749 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
750
751 if (!is_leftmost(curr, &rq->dl))
752 resched_curr(rq);
753 }
754
755 /*
756 * Because -- for now -- we share the rt bandwidth, we need to
757 * account our runtime there too, otherwise actual rt tasks
758 * would be able to exceed the shared quota.
759 *
760 * Account to the root rt group for now.
761 *
762 * The solution we're working towards is having the RT groups scheduled
763 * using deadline servers -- however there's a few nasties to figure
764 * out before that can happen.
765 */
766 if (rt_bandwidth_enabled()) {
767 struct rt_rq *rt_rq = &rq->rt;
768
769 raw_spin_lock(&rt_rq->rt_runtime_lock);
770 /*
771 * We'll let actual RT tasks worry about the overflow here, we
772 * have our own CBS to keep us inline; only account when RT
773 * bandwidth is relevant.
774 */
775 if (sched_rt_bandwidth_account(rt_rq))
776 rt_rq->rt_time += delta_exec;
777 raw_spin_unlock(&rt_rq->rt_runtime_lock);
778 }
779}
780
781#ifdef CONFIG_SMP
782
783static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
784{
785 struct rq *rq = rq_of_dl_rq(dl_rq);
786
787 if (dl_rq->earliest_dl.curr == 0 ||
788 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
789 dl_rq->earliest_dl.curr = deadline;
790 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
791 }
792}
793
794static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
795{
796 struct rq *rq = rq_of_dl_rq(dl_rq);
797
798 /*
799 * Since we may have removed our earliest (and/or next earliest)
800 * task we must recompute them.
801 */
802 if (!dl_rq->dl_nr_running) {
803 dl_rq->earliest_dl.curr = 0;
804 dl_rq->earliest_dl.next = 0;
805 cpudl_clear(&rq->rd->cpudl, rq->cpu);
806 } else {
807 struct rb_node *leftmost = dl_rq->rb_leftmost;
808 struct sched_dl_entity *entry;
809
810 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
811 dl_rq->earliest_dl.curr = entry->deadline;
812 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
813 }
814}
815
816#else
817
818static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
819static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
820
821#endif /* CONFIG_SMP */
822
823static inline
824void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
825{
826 int prio = dl_task_of(dl_se)->prio;
827 u64 deadline = dl_se->deadline;
828
829 WARN_ON(!dl_prio(prio));
830 dl_rq->dl_nr_running++;
831 add_nr_running(rq_of_dl_rq(dl_rq), 1);
832
833 inc_dl_deadline(dl_rq, deadline);
834 inc_dl_migration(dl_se, dl_rq);
835}
836
837static inline
838void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
839{
840 int prio = dl_task_of(dl_se)->prio;
841
842 WARN_ON(!dl_prio(prio));
843 WARN_ON(!dl_rq->dl_nr_running);
844 dl_rq->dl_nr_running--;
845 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
846
847 dec_dl_deadline(dl_rq, dl_se->deadline);
848 dec_dl_migration(dl_se, dl_rq);
849}
850
851static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
852{
853 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
854 struct rb_node **link = &dl_rq->rb_root.rb_node;
855 struct rb_node *parent = NULL;
856 struct sched_dl_entity *entry;
857 int leftmost = 1;
858
859 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
860
861 while (*link) {
862 parent = *link;
863 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
864 if (dl_time_before(dl_se->deadline, entry->deadline))
865 link = &parent->rb_left;
866 else {
867 link = &parent->rb_right;
868 leftmost = 0;
869 }
870 }
871
872 if (leftmost)
873 dl_rq->rb_leftmost = &dl_se->rb_node;
874
875 rb_link_node(&dl_se->rb_node, parent, link);
876 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
877
878 inc_dl_tasks(dl_se, dl_rq);
879}
880
881static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
882{
883 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
884
885 if (RB_EMPTY_NODE(&dl_se->rb_node))
886 return;
887
888 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
889 struct rb_node *next_node;
890
891 next_node = rb_next(&dl_se->rb_node);
892 dl_rq->rb_leftmost = next_node;
893 }
894
895 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
896 RB_CLEAR_NODE(&dl_se->rb_node);
897
898 dec_dl_tasks(dl_se, dl_rq);
899}
900
901static void
902enqueue_dl_entity(struct sched_dl_entity *dl_se,
903 struct sched_dl_entity *pi_se, int flags)
904{
905 BUG_ON(on_dl_rq(dl_se));
906
907 /*
908 * If this is a wakeup or a new instance, the scheduling
909 * parameters of the task might need updating. Otherwise,
910 * we want a replenishment of its runtime.
911 */
912 if (flags & ENQUEUE_WAKEUP)
913 update_dl_entity(dl_se, pi_se);
914 else if (flags & ENQUEUE_REPLENISH)
915 replenish_dl_entity(dl_se, pi_se);
916
917 __enqueue_dl_entity(dl_se);
918}
919
920static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
921{
922 __dequeue_dl_entity(dl_se);
923}
924
925static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
926{
927 struct task_struct *pi_task = rt_mutex_get_top_task(p);
928 struct sched_dl_entity *pi_se = &p->dl;
929
930 /*
931 * Use the scheduling parameters of the top pi-waiter
932 * task if we have one and its (absolute) deadline is
933 * smaller than our one... OTW we keep our runtime and
934 * deadline.
935 */
936 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) {
937 pi_se = &pi_task->dl;
938 } else if (!dl_prio(p->normal_prio)) {
939 /*
940 * Special case in which we have a !SCHED_DEADLINE task
941 * that is going to be deboosted, but exceedes its
942 * runtime while doing so. No point in replenishing
943 * it, as it's going to return back to its original
944 * scheduling class after this.
945 */
946 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
947 return;
948 }
949
950 /*
951 * If p is throttled, we do nothing. In fact, if it exhausted
952 * its budget it needs a replenishment and, since it now is on
953 * its rq, the bandwidth timer callback (which clearly has not
954 * run yet) will take care of this.
955 */
956 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH))
957 return;
958
959 enqueue_dl_entity(&p->dl, pi_se, flags);
960
961 if (!task_current(rq, p) && tsk_nr_cpus_allowed(p) > 1)
962 enqueue_pushable_dl_task(rq, p);
963}
964
965static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
966{
967 dequeue_dl_entity(&p->dl);
968 dequeue_pushable_dl_task(rq, p);
969}
970
971static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
972{
973 update_curr_dl(rq);
974 __dequeue_task_dl(rq, p, flags);
975}
976
977/*
978 * Yield task semantic for -deadline tasks is:
979 *
980 * get off from the CPU until our next instance, with
981 * a new runtime. This is of little use now, since we
982 * don't have a bandwidth reclaiming mechanism. Anyway,
983 * bandwidth reclaiming is planned for the future, and
984 * yield_task_dl will indicate that some spare budget
985 * is available for other task instances to use it.
986 */
987static void yield_task_dl(struct rq *rq)
988{
989 /*
990 * We make the task go to sleep until its current deadline by
991 * forcing its runtime to zero. This way, update_curr_dl() stops
992 * it and the bandwidth timer will wake it up and will give it
993 * new scheduling parameters (thanks to dl_yielded=1).
994 */
995 rq->curr->dl.dl_yielded = 1;
996
997 update_rq_clock(rq);
998 update_curr_dl(rq);
999 /*
1000 * Tell update_rq_clock() that we've just updated,
1001 * so we don't do microscopic update in schedule()
1002 * and double the fastpath cost.
1003 */
1004 rq_clock_skip_update(rq, true);
1005}
1006
1007#ifdef CONFIG_SMP
1008
1009static int find_later_rq(struct task_struct *task);
1010
1011static int
1012select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1013{
1014 struct task_struct *curr;
1015 struct rq *rq;
1016
1017 if (sd_flag != SD_BALANCE_WAKE)
1018 goto out;
1019
1020 rq = cpu_rq(cpu);
1021
1022 rcu_read_lock();
1023 curr = READ_ONCE(rq->curr); /* unlocked access */
1024
1025 /*
1026 * If we are dealing with a -deadline task, we must
1027 * decide where to wake it up.
1028 * If it has a later deadline and the current task
1029 * on this rq can't move (provided the waking task
1030 * can!) we prefer to send it somewhere else. On the
1031 * other hand, if it has a shorter deadline, we
1032 * try to make it stay here, it might be important.
1033 */
1034 if (unlikely(dl_task(curr)) &&
1035 (tsk_nr_cpus_allowed(curr) < 2 ||
1036 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1037 (tsk_nr_cpus_allowed(p) > 1)) {
1038 int target = find_later_rq(p);
1039
1040 if (target != -1 &&
1041 (dl_time_before(p->dl.deadline,
1042 cpu_rq(target)->dl.earliest_dl.curr) ||
1043 (cpu_rq(target)->dl.dl_nr_running == 0)))
1044 cpu = target;
1045 }
1046 rcu_read_unlock();
1047
1048out:
1049 return cpu;
1050}
1051
1052static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1053{
1054 /*
1055 * Current can't be migrated, useless to reschedule,
1056 * let's hope p can move out.
1057 */
1058 if (tsk_nr_cpus_allowed(rq->curr) == 1 ||
1059 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
1060 return;
1061
1062 /*
1063 * p is migratable, so let's not schedule it and
1064 * see if it is pushed or pulled somewhere else.
1065 */
1066 if (tsk_nr_cpus_allowed(p) != 1 &&
1067 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
1068 return;
1069
1070 resched_curr(rq);
1071}
1072
1073#endif /* CONFIG_SMP */
1074
1075/*
1076 * Only called when both the current and waking task are -deadline
1077 * tasks.
1078 */
1079static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1080 int flags)
1081{
1082 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1083 resched_curr(rq);
1084 return;
1085 }
1086
1087#ifdef CONFIG_SMP
1088 /*
1089 * In the unlikely case current and p have the same deadline
1090 * let us try to decide what's the best thing to do...
1091 */
1092 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1093 !test_tsk_need_resched(rq->curr))
1094 check_preempt_equal_dl(rq, p);
1095#endif /* CONFIG_SMP */
1096}
1097
1098#ifdef CONFIG_SCHED_HRTICK
1099static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1100{
1101 hrtick_start(rq, p->dl.runtime);
1102}
1103#else /* !CONFIG_SCHED_HRTICK */
1104static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1105{
1106}
1107#endif
1108
1109static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1110 struct dl_rq *dl_rq)
1111{
1112 struct rb_node *left = dl_rq->rb_leftmost;
1113
1114 if (!left)
1115 return NULL;
1116
1117 return rb_entry(left, struct sched_dl_entity, rb_node);
1118}
1119
1120struct task_struct *
1121pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
1122{
1123 struct sched_dl_entity *dl_se;
1124 struct task_struct *p;
1125 struct dl_rq *dl_rq;
1126
1127 dl_rq = &rq->dl;
1128
1129 if (need_pull_dl_task(rq, prev)) {
1130 /*
1131 * This is OK, because current is on_cpu, which avoids it being
1132 * picked for load-balance and preemption/IRQs are still
1133 * disabled avoiding further scheduler activity on it and we're
1134 * being very careful to re-start the picking loop.
1135 */
1136 lockdep_unpin_lock(&rq->lock, cookie);
1137 pull_dl_task(rq);
1138 lockdep_repin_lock(&rq->lock, cookie);
1139 /*
1140 * pull_dl_task() can drop (and re-acquire) rq->lock; this
1141 * means a stop task can slip in, in which case we need to
1142 * re-start task selection.
1143 */
1144 if (rq->stop && task_on_rq_queued(rq->stop))
1145 return RETRY_TASK;
1146 }
1147
1148 /*
1149 * When prev is DL, we may throttle it in put_prev_task().
1150 * So, we update time before we check for dl_nr_running.
1151 */
1152 if (prev->sched_class == &dl_sched_class)
1153 update_curr_dl(rq);
1154
1155 if (unlikely(!dl_rq->dl_nr_running))
1156 return NULL;
1157
1158 put_prev_task(rq, prev);
1159
1160 dl_se = pick_next_dl_entity(rq, dl_rq);
1161 BUG_ON(!dl_se);
1162
1163 p = dl_task_of(dl_se);
1164 p->se.exec_start = rq_clock_task(rq);
1165
1166 /* Running task will never be pushed. */
1167 dequeue_pushable_dl_task(rq, p);
1168
1169 if (hrtick_enabled(rq))
1170 start_hrtick_dl(rq, p);
1171
1172 queue_push_tasks(rq);
1173
1174 return p;
1175}
1176
1177static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1178{
1179 update_curr_dl(rq);
1180
1181 if (on_dl_rq(&p->dl) && tsk_nr_cpus_allowed(p) > 1)
1182 enqueue_pushable_dl_task(rq, p);
1183}
1184
1185static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1186{
1187 update_curr_dl(rq);
1188
1189 /*
1190 * Even when we have runtime, update_curr_dl() might have resulted in us
1191 * not being the leftmost task anymore. In that case NEED_RESCHED will
1192 * be set and schedule() will start a new hrtick for the next task.
1193 */
1194 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1195 is_leftmost(p, &rq->dl))
1196 start_hrtick_dl(rq, p);
1197}
1198
1199static void task_fork_dl(struct task_struct *p)
1200{
1201 /*
1202 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1203 * sched_fork()
1204 */
1205}
1206
1207static void task_dead_dl(struct task_struct *p)
1208{
1209 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1210
1211 /*
1212 * Since we are TASK_DEAD we won't slip out of the domain!
1213 */
1214 raw_spin_lock_irq(&dl_b->lock);
1215 /* XXX we should retain the bw until 0-lag */
1216 dl_b->total_bw -= p->dl.dl_bw;
1217 raw_spin_unlock_irq(&dl_b->lock);
1218}
1219
1220static void set_curr_task_dl(struct rq *rq)
1221{
1222 struct task_struct *p = rq->curr;
1223
1224 p->se.exec_start = rq_clock_task(rq);
1225
1226 /* You can't push away the running task */
1227 dequeue_pushable_dl_task(rq, p);
1228}
1229
1230#ifdef CONFIG_SMP
1231
1232/* Only try algorithms three times */
1233#define DL_MAX_TRIES 3
1234
1235static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1236{
1237 if (!task_running(rq, p) &&
1238 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1239 return 1;
1240 return 0;
1241}
1242
1243/*
1244 * Return the earliest pushable rq's task, which is suitable to be executed
1245 * on the CPU, NULL otherwise:
1246 */
1247static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1248{
1249 struct rb_node *next_node = rq->dl.pushable_dl_tasks_leftmost;
1250 struct task_struct *p = NULL;
1251
1252 if (!has_pushable_dl_tasks(rq))
1253 return NULL;
1254
1255next_node:
1256 if (next_node) {
1257 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1258
1259 if (pick_dl_task(rq, p, cpu))
1260 return p;
1261
1262 next_node = rb_next(next_node);
1263 goto next_node;
1264 }
1265
1266 return NULL;
1267}
1268
1269static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1270
1271static int find_later_rq(struct task_struct *task)
1272{
1273 struct sched_domain *sd;
1274 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1275 int this_cpu = smp_processor_id();
1276 int best_cpu, cpu = task_cpu(task);
1277
1278 /* Make sure the mask is initialized first */
1279 if (unlikely(!later_mask))
1280 return -1;
1281
1282 if (tsk_nr_cpus_allowed(task) == 1)
1283 return -1;
1284
1285 /*
1286 * We have to consider system topology and task affinity
1287 * first, then we can look for a suitable cpu.
1288 */
1289 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
1290 task, later_mask);
1291 if (best_cpu == -1)
1292 return -1;
1293
1294 /*
1295 * If we are here, some target has been found,
1296 * the most suitable of which is cached in best_cpu.
1297 * This is, among the runqueues where the current tasks
1298 * have later deadlines than the task's one, the rq
1299 * with the latest possible one.
1300 *
1301 * Now we check how well this matches with task's
1302 * affinity and system topology.
1303 *
1304 * The last cpu where the task run is our first
1305 * guess, since it is most likely cache-hot there.
1306 */
1307 if (cpumask_test_cpu(cpu, later_mask))
1308 return cpu;
1309 /*
1310 * Check if this_cpu is to be skipped (i.e., it is
1311 * not in the mask) or not.
1312 */
1313 if (!cpumask_test_cpu(this_cpu, later_mask))
1314 this_cpu = -1;
1315
1316 rcu_read_lock();
1317 for_each_domain(cpu, sd) {
1318 if (sd->flags & SD_WAKE_AFFINE) {
1319
1320 /*
1321 * If possible, preempting this_cpu is
1322 * cheaper than migrating.
1323 */
1324 if (this_cpu != -1 &&
1325 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1326 rcu_read_unlock();
1327 return this_cpu;
1328 }
1329
1330 /*
1331 * Last chance: if best_cpu is valid and is
1332 * in the mask, that becomes our choice.
1333 */
1334 if (best_cpu < nr_cpu_ids &&
1335 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
1336 rcu_read_unlock();
1337 return best_cpu;
1338 }
1339 }
1340 }
1341 rcu_read_unlock();
1342
1343 /*
1344 * At this point, all our guesses failed, we just return
1345 * 'something', and let the caller sort the things out.
1346 */
1347 if (this_cpu != -1)
1348 return this_cpu;
1349
1350 cpu = cpumask_any(later_mask);
1351 if (cpu < nr_cpu_ids)
1352 return cpu;
1353
1354 return -1;
1355}
1356
1357/* Locks the rq it finds */
1358static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1359{
1360 struct rq *later_rq = NULL;
1361 int tries;
1362 int cpu;
1363
1364 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1365 cpu = find_later_rq(task);
1366
1367 if ((cpu == -1) || (cpu == rq->cpu))
1368 break;
1369
1370 later_rq = cpu_rq(cpu);
1371
1372 if (later_rq->dl.dl_nr_running &&
1373 !dl_time_before(task->dl.deadline,
1374 later_rq->dl.earliest_dl.curr)) {
1375 /*
1376 * Target rq has tasks of equal or earlier deadline,
1377 * retrying does not release any lock and is unlikely
1378 * to yield a different result.
1379 */
1380 later_rq = NULL;
1381 break;
1382 }
1383
1384 /* Retry if something changed. */
1385 if (double_lock_balance(rq, later_rq)) {
1386 if (unlikely(task_rq(task) != rq ||
1387 !cpumask_test_cpu(later_rq->cpu,
1388 tsk_cpus_allowed(task)) ||
1389 task_running(rq, task) ||
1390 !dl_task(task) ||
1391 !task_on_rq_queued(task))) {
1392 double_unlock_balance(rq, later_rq);
1393 later_rq = NULL;
1394 break;
1395 }
1396 }
1397
1398 /*
1399 * If the rq we found has no -deadline task, or
1400 * its earliest one has a later deadline than our
1401 * task, the rq is a good one.
1402 */
1403 if (!later_rq->dl.dl_nr_running ||
1404 dl_time_before(task->dl.deadline,
1405 later_rq->dl.earliest_dl.curr))
1406 break;
1407
1408 /* Otherwise we try again. */
1409 double_unlock_balance(rq, later_rq);
1410 later_rq = NULL;
1411 }
1412
1413 return later_rq;
1414}
1415
1416static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1417{
1418 struct task_struct *p;
1419
1420 if (!has_pushable_dl_tasks(rq))
1421 return NULL;
1422
1423 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
1424 struct task_struct, pushable_dl_tasks);
1425
1426 BUG_ON(rq->cpu != task_cpu(p));
1427 BUG_ON(task_current(rq, p));
1428 BUG_ON(tsk_nr_cpus_allowed(p) <= 1);
1429
1430 BUG_ON(!task_on_rq_queued(p));
1431 BUG_ON(!dl_task(p));
1432
1433 return p;
1434}
1435
1436/*
1437 * See if the non running -deadline tasks on this rq
1438 * can be sent to some other CPU where they can preempt
1439 * and start executing.
1440 */
1441static int push_dl_task(struct rq *rq)
1442{
1443 struct task_struct *next_task;
1444 struct rq *later_rq;
1445 int ret = 0;
1446
1447 if (!rq->dl.overloaded)
1448 return 0;
1449
1450 next_task = pick_next_pushable_dl_task(rq);
1451 if (!next_task)
1452 return 0;
1453
1454retry:
1455 if (unlikely(next_task == rq->curr)) {
1456 WARN_ON(1);
1457 return 0;
1458 }
1459
1460 /*
1461 * If next_task preempts rq->curr, and rq->curr
1462 * can move away, it makes sense to just reschedule
1463 * without going further in pushing next_task.
1464 */
1465 if (dl_task(rq->curr) &&
1466 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1467 tsk_nr_cpus_allowed(rq->curr) > 1) {
1468 resched_curr(rq);
1469 return 0;
1470 }
1471
1472 /* We might release rq lock */
1473 get_task_struct(next_task);
1474
1475 /* Will lock the rq it'll find */
1476 later_rq = find_lock_later_rq(next_task, rq);
1477 if (!later_rq) {
1478 struct task_struct *task;
1479
1480 /*
1481 * We must check all this again, since
1482 * find_lock_later_rq releases rq->lock and it is
1483 * then possible that next_task has migrated.
1484 */
1485 task = pick_next_pushable_dl_task(rq);
1486 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1487 /*
1488 * The task is still there. We don't try
1489 * again, some other cpu will pull it when ready.
1490 */
1491 goto out;
1492 }
1493
1494 if (!task)
1495 /* No more tasks */
1496 goto out;
1497
1498 put_task_struct(next_task);
1499 next_task = task;
1500 goto retry;
1501 }
1502
1503 deactivate_task(rq, next_task, 0);
1504 set_task_cpu(next_task, later_rq->cpu);
1505 activate_task(later_rq, next_task, 0);
1506 ret = 1;
1507
1508 resched_curr(later_rq);
1509
1510 double_unlock_balance(rq, later_rq);
1511
1512out:
1513 put_task_struct(next_task);
1514
1515 return ret;
1516}
1517
1518static void push_dl_tasks(struct rq *rq)
1519{
1520 /* push_dl_task() will return true if it moved a -deadline task */
1521 while (push_dl_task(rq))
1522 ;
1523}
1524
1525static void pull_dl_task(struct rq *this_rq)
1526{
1527 int this_cpu = this_rq->cpu, cpu;
1528 struct task_struct *p;
1529 bool resched = false;
1530 struct rq *src_rq;
1531 u64 dmin = LONG_MAX;
1532
1533 if (likely(!dl_overloaded(this_rq)))
1534 return;
1535
1536 /*
1537 * Match the barrier from dl_set_overloaded; this guarantees that if we
1538 * see overloaded we must also see the dlo_mask bit.
1539 */
1540 smp_rmb();
1541
1542 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
1543 if (this_cpu == cpu)
1544 continue;
1545
1546 src_rq = cpu_rq(cpu);
1547
1548 /*
1549 * It looks racy, abd it is! However, as in sched_rt.c,
1550 * we are fine with this.
1551 */
1552 if (this_rq->dl.dl_nr_running &&
1553 dl_time_before(this_rq->dl.earliest_dl.curr,
1554 src_rq->dl.earliest_dl.next))
1555 continue;
1556
1557 /* Might drop this_rq->lock */
1558 double_lock_balance(this_rq, src_rq);
1559
1560 /*
1561 * If there are no more pullable tasks on the
1562 * rq, we're done with it.
1563 */
1564 if (src_rq->dl.dl_nr_running <= 1)
1565 goto skip;
1566
1567 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
1568
1569 /*
1570 * We found a task to be pulled if:
1571 * - it preempts our current (if there's one),
1572 * - it will preempt the last one we pulled (if any).
1573 */
1574 if (p && dl_time_before(p->dl.deadline, dmin) &&
1575 (!this_rq->dl.dl_nr_running ||
1576 dl_time_before(p->dl.deadline,
1577 this_rq->dl.earliest_dl.curr))) {
1578 WARN_ON(p == src_rq->curr);
1579 WARN_ON(!task_on_rq_queued(p));
1580
1581 /*
1582 * Then we pull iff p has actually an earlier
1583 * deadline than the current task of its runqueue.
1584 */
1585 if (dl_time_before(p->dl.deadline,
1586 src_rq->curr->dl.deadline))
1587 goto skip;
1588
1589 resched = true;
1590
1591 deactivate_task(src_rq, p, 0);
1592 set_task_cpu(p, this_cpu);
1593 activate_task(this_rq, p, 0);
1594 dmin = p->dl.deadline;
1595
1596 /* Is there any other task even earlier? */
1597 }
1598skip:
1599 double_unlock_balance(this_rq, src_rq);
1600 }
1601
1602 if (resched)
1603 resched_curr(this_rq);
1604}
1605
1606/*
1607 * Since the task is not running and a reschedule is not going to happen
1608 * anytime soon on its runqueue, we try pushing it away now.
1609 */
1610static void task_woken_dl(struct rq *rq, struct task_struct *p)
1611{
1612 if (!task_running(rq, p) &&
1613 !test_tsk_need_resched(rq->curr) &&
1614 tsk_nr_cpus_allowed(p) > 1 &&
1615 dl_task(rq->curr) &&
1616 (tsk_nr_cpus_allowed(rq->curr) < 2 ||
1617 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
1618 push_dl_tasks(rq);
1619 }
1620}
1621
1622static void set_cpus_allowed_dl(struct task_struct *p,
1623 const struct cpumask *new_mask)
1624{
1625 struct root_domain *src_rd;
1626 struct rq *rq;
1627
1628 BUG_ON(!dl_task(p));
1629
1630 rq = task_rq(p);
1631 src_rd = rq->rd;
1632 /*
1633 * Migrating a SCHED_DEADLINE task between exclusive
1634 * cpusets (different root_domains) entails a bandwidth
1635 * update. We already made space for us in the destination
1636 * domain (see cpuset_can_attach()).
1637 */
1638 if (!cpumask_intersects(src_rd->span, new_mask)) {
1639 struct dl_bw *src_dl_b;
1640
1641 src_dl_b = dl_bw_of(cpu_of(rq));
1642 /*
1643 * We now free resources of the root_domain we are migrating
1644 * off. In the worst case, sched_setattr() may temporary fail
1645 * until we complete the update.
1646 */
1647 raw_spin_lock(&src_dl_b->lock);
1648 __dl_clear(src_dl_b, p->dl.dl_bw);
1649 raw_spin_unlock(&src_dl_b->lock);
1650 }
1651
1652 set_cpus_allowed_common(p, new_mask);
1653}
1654
1655/* Assumes rq->lock is held */
1656static void rq_online_dl(struct rq *rq)
1657{
1658 if (rq->dl.overloaded)
1659 dl_set_overload(rq);
1660
1661 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
1662 if (rq->dl.dl_nr_running > 0)
1663 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
1664}
1665
1666/* Assumes rq->lock is held */
1667static void rq_offline_dl(struct rq *rq)
1668{
1669 if (rq->dl.overloaded)
1670 dl_clear_overload(rq);
1671
1672 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1673 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
1674}
1675
1676void __init init_sched_dl_class(void)
1677{
1678 unsigned int i;
1679
1680 for_each_possible_cpu(i)
1681 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
1682 GFP_KERNEL, cpu_to_node(i));
1683}
1684
1685#endif /* CONFIG_SMP */
1686
1687static void switched_from_dl(struct rq *rq, struct task_struct *p)
1688{
1689 /*
1690 * Start the deadline timer; if we switch back to dl before this we'll
1691 * continue consuming our current CBS slice. If we stay outside of
1692 * SCHED_DEADLINE until the deadline passes, the timer will reset the
1693 * task.
1694 */
1695 if (!start_dl_timer(p))
1696 __dl_clear_params(p);
1697
1698 /*
1699 * Since this might be the only -deadline task on the rq,
1700 * this is the right place to try to pull some other one
1701 * from an overloaded cpu, if any.
1702 */
1703 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
1704 return;
1705
1706 queue_pull_task(rq);
1707}
1708
1709/*
1710 * When switching to -deadline, we may overload the rq, then
1711 * we try to push someone off, if possible.
1712 */
1713static void switched_to_dl(struct rq *rq, struct task_struct *p)
1714{
1715
1716 /* If p is not queued we will update its parameters at next wakeup. */
1717 if (!task_on_rq_queued(p))
1718 return;
1719
1720 /*
1721 * If p is boosted we already updated its params in
1722 * rt_mutex_setprio()->enqueue_task(..., ENQUEUE_REPLENISH),
1723 * p's deadline being now already after rq_clock(rq).
1724 */
1725 if (dl_time_before(p->dl.deadline, rq_clock(rq)))
1726 setup_new_dl_entity(&p->dl);
1727
1728 if (rq->curr != p) {
1729#ifdef CONFIG_SMP
1730 if (tsk_nr_cpus_allowed(p) > 1 && rq->dl.overloaded)
1731 queue_push_tasks(rq);
1732#endif
1733 if (dl_task(rq->curr))
1734 check_preempt_curr_dl(rq, p, 0);
1735 else
1736 resched_curr(rq);
1737 }
1738}
1739
1740/*
1741 * If the scheduling parameters of a -deadline task changed,
1742 * a push or pull operation might be needed.
1743 */
1744static void prio_changed_dl(struct rq *rq, struct task_struct *p,
1745 int oldprio)
1746{
1747 if (task_on_rq_queued(p) || rq->curr == p) {
1748#ifdef CONFIG_SMP
1749 /*
1750 * This might be too much, but unfortunately
1751 * we don't have the old deadline value, and
1752 * we can't argue if the task is increasing
1753 * or lowering its prio, so...
1754 */
1755 if (!rq->dl.overloaded)
1756 queue_pull_task(rq);
1757
1758 /*
1759 * If we now have a earlier deadline task than p,
1760 * then reschedule, provided p is still on this
1761 * runqueue.
1762 */
1763 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
1764 resched_curr(rq);
1765#else
1766 /*
1767 * Again, we don't know if p has a earlier
1768 * or later deadline, so let's blindly set a
1769 * (maybe not needed) rescheduling point.
1770 */
1771 resched_curr(rq);
1772#endif /* CONFIG_SMP */
1773 }
1774}
1775
1776const struct sched_class dl_sched_class = {
1777 .next = &rt_sched_class,
1778 .enqueue_task = enqueue_task_dl,
1779 .dequeue_task = dequeue_task_dl,
1780 .yield_task = yield_task_dl,
1781
1782 .check_preempt_curr = check_preempt_curr_dl,
1783
1784 .pick_next_task = pick_next_task_dl,
1785 .put_prev_task = put_prev_task_dl,
1786
1787#ifdef CONFIG_SMP
1788 .select_task_rq = select_task_rq_dl,
1789 .set_cpus_allowed = set_cpus_allowed_dl,
1790 .rq_online = rq_online_dl,
1791 .rq_offline = rq_offline_dl,
1792 .task_woken = task_woken_dl,
1793#endif
1794
1795 .set_curr_task = set_curr_task_dl,
1796 .task_tick = task_tick_dl,
1797 .task_fork = task_fork_dl,
1798 .task_dead = task_dead_dl,
1799
1800 .prio_changed = prio_changed_dl,
1801 .switched_from = switched_from_dl,
1802 .switched_to = switched_to_dl,
1803
1804 .update_curr = update_curr_dl,
1805};
1806
1807#ifdef CONFIG_SCHED_DEBUG
1808extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
1809
1810void print_dl_stats(struct seq_file *m, int cpu)
1811{
1812 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
1813}
1814#endif /* CONFIG_SCHED_DEBUG */