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// 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
19#include <linux/cpuset.h>
20
21/*
22 * Default limits for DL period; on the top end we guard against small util
23 * tasks still getting ridiculously long effective runtimes, on the bottom end we
24 * guard against timer DoS.
25 */
26static unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
27static unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
28#ifdef CONFIG_SYSCTL
29static struct ctl_table sched_dl_sysctls[] = {
30 {
31 .procname = "sched_deadline_period_max_us",
32 .data = &sysctl_sched_dl_period_max,
33 .maxlen = sizeof(unsigned int),
34 .mode = 0644,
35 .proc_handler = proc_douintvec_minmax,
36 .extra1 = (void *)&sysctl_sched_dl_period_min,
37 },
38 {
39 .procname = "sched_deadline_period_min_us",
40 .data = &sysctl_sched_dl_period_min,
41 .maxlen = sizeof(unsigned int),
42 .mode = 0644,
43 .proc_handler = proc_douintvec_minmax,
44 .extra2 = (void *)&sysctl_sched_dl_period_max,
45 },
46 {}
47};
48
49static int __init sched_dl_sysctl_init(void)
50{
51 register_sysctl_init("kernel", sched_dl_sysctls);
52 return 0;
53}
54late_initcall(sched_dl_sysctl_init);
55#endif
56
57static bool dl_server(struct sched_dl_entity *dl_se)
58{
59 return dl_se->dl_server;
60}
61
62static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
63{
64 BUG_ON(dl_server(dl_se));
65 return container_of(dl_se, struct task_struct, dl);
66}
67
68static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
69{
70 return container_of(dl_rq, struct rq, dl);
71}
72
73static inline struct rq *rq_of_dl_se(struct sched_dl_entity *dl_se)
74{
75 struct rq *rq = dl_se->rq;
76
77 if (!dl_server(dl_se))
78 rq = task_rq(dl_task_of(dl_se));
79
80 return rq;
81}
82
83static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
84{
85 return &rq_of_dl_se(dl_se)->dl;
86}
87
88static inline int on_dl_rq(struct sched_dl_entity *dl_se)
89{
90 return !RB_EMPTY_NODE(&dl_se->rb_node);
91}
92
93#ifdef CONFIG_RT_MUTEXES
94static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
95{
96 return dl_se->pi_se;
97}
98
99static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
100{
101 return pi_of(dl_se) != dl_se;
102}
103#else
104static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
105{
106 return dl_se;
107}
108
109static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
110{
111 return false;
112}
113#endif
114
115#ifdef CONFIG_SMP
116static inline struct dl_bw *dl_bw_of(int i)
117{
118 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
119 "sched RCU must be held");
120 return &cpu_rq(i)->rd->dl_bw;
121}
122
123static inline int dl_bw_cpus(int i)
124{
125 struct root_domain *rd = cpu_rq(i)->rd;
126 int cpus;
127
128 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
129 "sched RCU must be held");
130
131 if (cpumask_subset(rd->span, cpu_active_mask))
132 return cpumask_weight(rd->span);
133
134 cpus = 0;
135
136 for_each_cpu_and(i, rd->span, cpu_active_mask)
137 cpus++;
138
139 return cpus;
140}
141
142static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
143{
144 unsigned long cap = 0;
145 int i;
146
147 for_each_cpu_and(i, mask, cpu_active_mask)
148 cap += arch_scale_cpu_capacity(i);
149
150 return cap;
151}
152
153/*
154 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
155 * of the CPU the task is running on rather rd's \Sum CPU capacity.
156 */
157static inline unsigned long dl_bw_capacity(int i)
158{
159 if (!sched_asym_cpucap_active() &&
160 arch_scale_cpu_capacity(i) == SCHED_CAPACITY_SCALE) {
161 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
162 } else {
163 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
164 "sched RCU must be held");
165
166 return __dl_bw_capacity(cpu_rq(i)->rd->span);
167 }
168}
169
170static inline bool dl_bw_visited(int cpu, u64 gen)
171{
172 struct root_domain *rd = cpu_rq(cpu)->rd;
173
174 if (rd->visit_gen == gen)
175 return true;
176
177 rd->visit_gen = gen;
178 return false;
179}
180
181static inline
182void __dl_update(struct dl_bw *dl_b, s64 bw)
183{
184 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
185 int i;
186
187 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
188 "sched RCU must be held");
189 for_each_cpu_and(i, rd->span, cpu_active_mask) {
190 struct rq *rq = cpu_rq(i);
191
192 rq->dl.extra_bw += bw;
193 }
194}
195#else
196static inline struct dl_bw *dl_bw_of(int i)
197{
198 return &cpu_rq(i)->dl.dl_bw;
199}
200
201static inline int dl_bw_cpus(int i)
202{
203 return 1;
204}
205
206static inline unsigned long dl_bw_capacity(int i)
207{
208 return SCHED_CAPACITY_SCALE;
209}
210
211static inline bool dl_bw_visited(int cpu, u64 gen)
212{
213 return false;
214}
215
216static inline
217void __dl_update(struct dl_bw *dl_b, s64 bw)
218{
219 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
220
221 dl->extra_bw += bw;
222}
223#endif
224
225static inline
226void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
227{
228 dl_b->total_bw -= tsk_bw;
229 __dl_update(dl_b, (s32)tsk_bw / cpus);
230}
231
232static inline
233void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
234{
235 dl_b->total_bw += tsk_bw;
236 __dl_update(dl_b, -((s32)tsk_bw / cpus));
237}
238
239static inline bool
240__dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
241{
242 return dl_b->bw != -1 &&
243 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
244}
245
246static inline
247void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
248{
249 u64 old = dl_rq->running_bw;
250
251 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
252 dl_rq->running_bw += dl_bw;
253 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
254 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
255 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
256 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
257}
258
259static inline
260void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
261{
262 u64 old = dl_rq->running_bw;
263
264 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
265 dl_rq->running_bw -= dl_bw;
266 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
267 if (dl_rq->running_bw > old)
268 dl_rq->running_bw = 0;
269 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
270 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
271}
272
273static inline
274void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
275{
276 u64 old = dl_rq->this_bw;
277
278 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
279 dl_rq->this_bw += dl_bw;
280 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
281}
282
283static inline
284void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
285{
286 u64 old = dl_rq->this_bw;
287
288 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
289 dl_rq->this_bw -= dl_bw;
290 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
291 if (dl_rq->this_bw > old)
292 dl_rq->this_bw = 0;
293 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
294}
295
296static inline
297void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
298{
299 if (!dl_entity_is_special(dl_se))
300 __add_rq_bw(dl_se->dl_bw, dl_rq);
301}
302
303static inline
304void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
305{
306 if (!dl_entity_is_special(dl_se))
307 __sub_rq_bw(dl_se->dl_bw, dl_rq);
308}
309
310static inline
311void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
312{
313 if (!dl_entity_is_special(dl_se))
314 __add_running_bw(dl_se->dl_bw, dl_rq);
315}
316
317static inline
318void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
319{
320 if (!dl_entity_is_special(dl_se))
321 __sub_running_bw(dl_se->dl_bw, dl_rq);
322}
323
324static void dl_change_utilization(struct task_struct *p, u64 new_bw)
325{
326 struct rq *rq;
327
328 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
329
330 if (task_on_rq_queued(p))
331 return;
332
333 rq = task_rq(p);
334 if (p->dl.dl_non_contending) {
335 sub_running_bw(&p->dl, &rq->dl);
336 p->dl.dl_non_contending = 0;
337 /*
338 * If the timer handler is currently running and the
339 * timer cannot be canceled, inactive_task_timer()
340 * will see that dl_not_contending is not set, and
341 * will not touch the rq's active utilization,
342 * so we are still safe.
343 */
344 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
345 put_task_struct(p);
346 }
347 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
348 __add_rq_bw(new_bw, &rq->dl);
349}
350
351static void __dl_clear_params(struct sched_dl_entity *dl_se);
352
353/*
354 * The utilization of a task cannot be immediately removed from
355 * the rq active utilization (running_bw) when the task blocks.
356 * Instead, we have to wait for the so called "0-lag time".
357 *
358 * If a task blocks before the "0-lag time", a timer (the inactive
359 * timer) is armed, and running_bw is decreased when the timer
360 * fires.
361 *
362 * If the task wakes up again before the inactive timer fires,
363 * the timer is canceled, whereas if the task wakes up after the
364 * inactive timer fired (and running_bw has been decreased) the
365 * task's utilization has to be added to running_bw again.
366 * A flag in the deadline scheduling entity (dl_non_contending)
367 * is used to avoid race conditions between the inactive timer handler
368 * and task wakeups.
369 *
370 * The following diagram shows how running_bw is updated. A task is
371 * "ACTIVE" when its utilization contributes to running_bw; an
372 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
373 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
374 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
375 * time already passed, which does not contribute to running_bw anymore.
376 * +------------------+
377 * wakeup | ACTIVE |
378 * +------------------>+ contending |
379 * | add_running_bw | |
380 * | +----+------+------+
381 * | | ^
382 * | dequeue | |
383 * +--------+-------+ | |
384 * | | t >= 0-lag | | wakeup
385 * | INACTIVE |<---------------+ |
386 * | | sub_running_bw | |
387 * +--------+-------+ | |
388 * ^ | |
389 * | t < 0-lag | |
390 * | | |
391 * | V |
392 * | +----+------+------+
393 * | sub_running_bw | ACTIVE |
394 * +-------------------+ |
395 * inactive timer | non contending |
396 * fired +------------------+
397 *
398 * The task_non_contending() function is invoked when a task
399 * blocks, and checks if the 0-lag time already passed or
400 * not (in the first case, it directly updates running_bw;
401 * in the second case, it arms the inactive timer).
402 *
403 * The task_contending() function is invoked when a task wakes
404 * up, and checks if the task is still in the "ACTIVE non contending"
405 * state or not (in the second case, it updates running_bw).
406 */
407static void task_non_contending(struct sched_dl_entity *dl_se)
408{
409 struct hrtimer *timer = &dl_se->inactive_timer;
410 struct rq *rq = rq_of_dl_se(dl_se);
411 struct dl_rq *dl_rq = &rq->dl;
412 s64 zerolag_time;
413
414 /*
415 * If this is a non-deadline task that has been boosted,
416 * do nothing
417 */
418 if (dl_se->dl_runtime == 0)
419 return;
420
421 if (dl_entity_is_special(dl_se))
422 return;
423
424 WARN_ON(dl_se->dl_non_contending);
425
426 zerolag_time = dl_se->deadline -
427 div64_long((dl_se->runtime * dl_se->dl_period),
428 dl_se->dl_runtime);
429
430 /*
431 * Using relative times instead of the absolute "0-lag time"
432 * allows to simplify the code
433 */
434 zerolag_time -= rq_clock(rq);
435
436 /*
437 * If the "0-lag time" already passed, decrease the active
438 * utilization now, instead of starting a timer
439 */
440 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
441 if (dl_server(dl_se)) {
442 sub_running_bw(dl_se, dl_rq);
443 } else {
444 struct task_struct *p = dl_task_of(dl_se);
445
446 if (dl_task(p))
447 sub_running_bw(dl_se, dl_rq);
448
449 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
450 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
451
452 if (READ_ONCE(p->__state) == TASK_DEAD)
453 sub_rq_bw(dl_se, &rq->dl);
454 raw_spin_lock(&dl_b->lock);
455 __dl_sub(dl_b, dl_se->dl_bw, dl_bw_cpus(task_cpu(p)));
456 raw_spin_unlock(&dl_b->lock);
457 __dl_clear_params(dl_se);
458 }
459 }
460
461 return;
462 }
463
464 dl_se->dl_non_contending = 1;
465 if (!dl_server(dl_se))
466 get_task_struct(dl_task_of(dl_se));
467
468 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
469}
470
471static void task_contending(struct sched_dl_entity *dl_se, int flags)
472{
473 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
474
475 /*
476 * If this is a non-deadline task that has been boosted,
477 * do nothing
478 */
479 if (dl_se->dl_runtime == 0)
480 return;
481
482 if (flags & ENQUEUE_MIGRATED)
483 add_rq_bw(dl_se, dl_rq);
484
485 if (dl_se->dl_non_contending) {
486 dl_se->dl_non_contending = 0;
487 /*
488 * If the timer handler is currently running and the
489 * timer cannot be canceled, inactive_task_timer()
490 * will see that dl_not_contending is not set, and
491 * will not touch the rq's active utilization,
492 * so we are still safe.
493 */
494 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) {
495 if (!dl_server(dl_se))
496 put_task_struct(dl_task_of(dl_se));
497 }
498 } else {
499 /*
500 * Since "dl_non_contending" is not set, the
501 * task's utilization has already been removed from
502 * active utilization (either when the task blocked,
503 * when the "inactive timer" fired).
504 * So, add it back.
505 */
506 add_running_bw(dl_se, dl_rq);
507 }
508}
509
510static inline int is_leftmost(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
511{
512 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
513}
514
515static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
516
517void init_dl_bw(struct dl_bw *dl_b)
518{
519 raw_spin_lock_init(&dl_b->lock);
520 if (global_rt_runtime() == RUNTIME_INF)
521 dl_b->bw = -1;
522 else
523 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
524 dl_b->total_bw = 0;
525}
526
527void init_dl_rq(struct dl_rq *dl_rq)
528{
529 dl_rq->root = RB_ROOT_CACHED;
530
531#ifdef CONFIG_SMP
532 /* zero means no -deadline tasks */
533 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
534
535 dl_rq->overloaded = 0;
536 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
537#else
538 init_dl_bw(&dl_rq->dl_bw);
539#endif
540
541 dl_rq->running_bw = 0;
542 dl_rq->this_bw = 0;
543 init_dl_rq_bw_ratio(dl_rq);
544}
545
546#ifdef CONFIG_SMP
547
548static inline int dl_overloaded(struct rq *rq)
549{
550 return atomic_read(&rq->rd->dlo_count);
551}
552
553static inline void dl_set_overload(struct rq *rq)
554{
555 if (!rq->online)
556 return;
557
558 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
559 /*
560 * Must be visible before the overload count is
561 * set (as in sched_rt.c).
562 *
563 * Matched by the barrier in pull_dl_task().
564 */
565 smp_wmb();
566 atomic_inc(&rq->rd->dlo_count);
567}
568
569static inline void dl_clear_overload(struct rq *rq)
570{
571 if (!rq->online)
572 return;
573
574 atomic_dec(&rq->rd->dlo_count);
575 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
576}
577
578#define __node_2_pdl(node) \
579 rb_entry((node), struct task_struct, pushable_dl_tasks)
580
581static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
582{
583 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
584}
585
586static inline int has_pushable_dl_tasks(struct rq *rq)
587{
588 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
589}
590
591/*
592 * The list of pushable -deadline task is not a plist, like in
593 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
594 */
595static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
596{
597 struct rb_node *leftmost;
598
599 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
600
601 leftmost = rb_add_cached(&p->pushable_dl_tasks,
602 &rq->dl.pushable_dl_tasks_root,
603 __pushable_less);
604 if (leftmost)
605 rq->dl.earliest_dl.next = p->dl.deadline;
606
607 if (!rq->dl.overloaded) {
608 dl_set_overload(rq);
609 rq->dl.overloaded = 1;
610 }
611}
612
613static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
614{
615 struct dl_rq *dl_rq = &rq->dl;
616 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
617 struct rb_node *leftmost;
618
619 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
620 return;
621
622 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
623 if (leftmost)
624 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
625
626 RB_CLEAR_NODE(&p->pushable_dl_tasks);
627
628 if (!has_pushable_dl_tasks(rq) && rq->dl.overloaded) {
629 dl_clear_overload(rq);
630 rq->dl.overloaded = 0;
631 }
632}
633
634static int push_dl_task(struct rq *rq);
635
636static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
637{
638 return rq->online && dl_task(prev);
639}
640
641static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
642static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
643
644static void push_dl_tasks(struct rq *);
645static void pull_dl_task(struct rq *);
646
647static inline void deadline_queue_push_tasks(struct rq *rq)
648{
649 if (!has_pushable_dl_tasks(rq))
650 return;
651
652 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
653}
654
655static inline void deadline_queue_pull_task(struct rq *rq)
656{
657 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
658}
659
660static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
661
662static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
663{
664 struct rq *later_rq = NULL;
665 struct dl_bw *dl_b;
666
667 later_rq = find_lock_later_rq(p, rq);
668 if (!later_rq) {
669 int cpu;
670
671 /*
672 * If we cannot preempt any rq, fall back to pick any
673 * online CPU:
674 */
675 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
676 if (cpu >= nr_cpu_ids) {
677 /*
678 * Failed to find any suitable CPU.
679 * The task will never come back!
680 */
681 WARN_ON_ONCE(dl_bandwidth_enabled());
682
683 /*
684 * If admission control is disabled we
685 * try a little harder to let the task
686 * run.
687 */
688 cpu = cpumask_any(cpu_active_mask);
689 }
690 later_rq = cpu_rq(cpu);
691 double_lock_balance(rq, later_rq);
692 }
693
694 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
695 /*
696 * Inactive timer is armed (or callback is running, but
697 * waiting for us to release rq locks). In any case, when it
698 * will fire (or continue), it will see running_bw of this
699 * task migrated to later_rq (and correctly handle it).
700 */
701 sub_running_bw(&p->dl, &rq->dl);
702 sub_rq_bw(&p->dl, &rq->dl);
703
704 add_rq_bw(&p->dl, &later_rq->dl);
705 add_running_bw(&p->dl, &later_rq->dl);
706 } else {
707 sub_rq_bw(&p->dl, &rq->dl);
708 add_rq_bw(&p->dl, &later_rq->dl);
709 }
710
711 /*
712 * And we finally need to fixup root_domain(s) bandwidth accounting,
713 * since p is still hanging out in the old (now moved to default) root
714 * domain.
715 */
716 dl_b = &rq->rd->dl_bw;
717 raw_spin_lock(&dl_b->lock);
718 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
719 raw_spin_unlock(&dl_b->lock);
720
721 dl_b = &later_rq->rd->dl_bw;
722 raw_spin_lock(&dl_b->lock);
723 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
724 raw_spin_unlock(&dl_b->lock);
725
726 set_task_cpu(p, later_rq->cpu);
727 double_unlock_balance(later_rq, rq);
728
729 return later_rq;
730}
731
732#else
733
734static inline
735void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
736{
737}
738
739static inline
740void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
741{
742}
743
744static inline
745void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
746{
747}
748
749static inline
750void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
751{
752}
753
754static inline void deadline_queue_push_tasks(struct rq *rq)
755{
756}
757
758static inline void deadline_queue_pull_task(struct rq *rq)
759{
760}
761#endif /* CONFIG_SMP */
762
763static void
764enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags);
765static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
766static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags);
767static void wakeup_preempt_dl(struct rq *rq, struct task_struct *p, int flags);
768
769static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
770 struct rq *rq)
771{
772 /* for non-boosted task, pi_of(dl_se) == dl_se */
773 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
774 dl_se->runtime = pi_of(dl_se)->dl_runtime;
775}
776
777/*
778 * We are being explicitly informed that a new instance is starting,
779 * and this means that:
780 * - the absolute deadline of the entity has to be placed at
781 * current time + relative deadline;
782 * - the runtime of the entity has to be set to the maximum value.
783 *
784 * The capability of specifying such event is useful whenever a -deadline
785 * entity wants to (try to!) synchronize its behaviour with the scheduler's
786 * one, and to (try to!) reconcile itself with its own scheduling
787 * parameters.
788 */
789static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
790{
791 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
792 struct rq *rq = rq_of_dl_rq(dl_rq);
793
794 WARN_ON(is_dl_boosted(dl_se));
795 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
796
797 /*
798 * We are racing with the deadline timer. So, do nothing because
799 * the deadline timer handler will take care of properly recharging
800 * the runtime and postponing the deadline
801 */
802 if (dl_se->dl_throttled)
803 return;
804
805 /*
806 * We use the regular wall clock time to set deadlines in the
807 * future; in fact, we must consider execution overheads (time
808 * spent on hardirq context, etc.).
809 */
810 replenish_dl_new_period(dl_se, rq);
811}
812
813/*
814 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
815 * possibility of a entity lasting more than what it declared, and thus
816 * exhausting its runtime.
817 *
818 * Here we are interested in making runtime overrun possible, but we do
819 * not want a entity which is misbehaving to affect the scheduling of all
820 * other entities.
821 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
822 * is used, in order to confine each entity within its own bandwidth.
823 *
824 * This function deals exactly with that, and ensures that when the runtime
825 * of a entity is replenished, its deadline is also postponed. That ensures
826 * the overrunning entity can't interfere with other entity in the system and
827 * can't make them miss their deadlines. Reasons why this kind of overruns
828 * could happen are, typically, a entity voluntarily trying to overcome its
829 * runtime, or it just underestimated it during sched_setattr().
830 */
831static void replenish_dl_entity(struct sched_dl_entity *dl_se)
832{
833 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
834 struct rq *rq = rq_of_dl_rq(dl_rq);
835
836 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
837
838 /*
839 * This could be the case for a !-dl task that is boosted.
840 * Just go with full inherited parameters.
841 */
842 if (dl_se->dl_deadline == 0)
843 replenish_dl_new_period(dl_se, rq);
844
845 if (dl_se->dl_yielded && dl_se->runtime > 0)
846 dl_se->runtime = 0;
847
848 /*
849 * We keep moving the deadline away until we get some
850 * available runtime for the entity. This ensures correct
851 * handling of situations where the runtime overrun is
852 * arbitrary large.
853 */
854 while (dl_se->runtime <= 0) {
855 dl_se->deadline += pi_of(dl_se)->dl_period;
856 dl_se->runtime += pi_of(dl_se)->dl_runtime;
857 }
858
859 /*
860 * At this point, the deadline really should be "in
861 * the future" with respect to rq->clock. If it's
862 * not, we are, for some reason, lagging too much!
863 * Anyway, after having warn userspace abut that,
864 * we still try to keep the things running by
865 * resetting the deadline and the budget of the
866 * entity.
867 */
868 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
869 printk_deferred_once("sched: DL replenish lagged too much\n");
870 replenish_dl_new_period(dl_se, rq);
871 }
872
873 if (dl_se->dl_yielded)
874 dl_se->dl_yielded = 0;
875 if (dl_se->dl_throttled)
876 dl_se->dl_throttled = 0;
877}
878
879/*
880 * Here we check if --at time t-- an entity (which is probably being
881 * [re]activated or, in general, enqueued) can use its remaining runtime
882 * and its current deadline _without_ exceeding the bandwidth it is
883 * assigned (function returns true if it can't). We are in fact applying
884 * one of the CBS rules: when a task wakes up, if the residual runtime
885 * over residual deadline fits within the allocated bandwidth, then we
886 * can keep the current (absolute) deadline and residual budget without
887 * disrupting the schedulability of the system. Otherwise, we should
888 * refill the runtime and set the deadline a period in the future,
889 * because keeping the current (absolute) deadline of the task would
890 * result in breaking guarantees promised to other tasks (refer to
891 * Documentation/scheduler/sched-deadline.rst for more information).
892 *
893 * This function returns true if:
894 *
895 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
896 *
897 * IOW we can't recycle current parameters.
898 *
899 * Notice that the bandwidth check is done against the deadline. For
900 * task with deadline equal to period this is the same of using
901 * dl_period instead of dl_deadline in the equation above.
902 */
903static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
904{
905 u64 left, right;
906
907 /*
908 * left and right are the two sides of the equation above,
909 * after a bit of shuffling to use multiplications instead
910 * of divisions.
911 *
912 * Note that none of the time values involved in the two
913 * multiplications are absolute: dl_deadline and dl_runtime
914 * are the relative deadline and the maximum runtime of each
915 * instance, runtime is the runtime left for the last instance
916 * and (deadline - t), since t is rq->clock, is the time left
917 * to the (absolute) deadline. Even if overflowing the u64 type
918 * is very unlikely to occur in both cases, here we scale down
919 * as we want to avoid that risk at all. Scaling down by 10
920 * means that we reduce granularity to 1us. We are fine with it,
921 * since this is only a true/false check and, anyway, thinking
922 * of anything below microseconds resolution is actually fiction
923 * (but still we want to give the user that illusion >;).
924 */
925 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
926 right = ((dl_se->deadline - t) >> DL_SCALE) *
927 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
928
929 return dl_time_before(right, left);
930}
931
932/*
933 * Revised wakeup rule [1]: For self-suspending tasks, rather then
934 * re-initializing task's runtime and deadline, the revised wakeup
935 * rule adjusts the task's runtime to avoid the task to overrun its
936 * density.
937 *
938 * Reasoning: a task may overrun the density if:
939 * runtime / (deadline - t) > dl_runtime / dl_deadline
940 *
941 * Therefore, runtime can be adjusted to:
942 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
943 *
944 * In such way that runtime will be equal to the maximum density
945 * the task can use without breaking any rule.
946 *
947 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
948 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
949 */
950static void
951update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
952{
953 u64 laxity = dl_se->deadline - rq_clock(rq);
954
955 /*
956 * If the task has deadline < period, and the deadline is in the past,
957 * it should already be throttled before this check.
958 *
959 * See update_dl_entity() comments for further details.
960 */
961 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
962
963 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
964}
965
966/*
967 * Regarding the deadline, a task with implicit deadline has a relative
968 * deadline == relative period. A task with constrained deadline has a
969 * relative deadline <= relative period.
970 *
971 * We support constrained deadline tasks. However, there are some restrictions
972 * applied only for tasks which do not have an implicit deadline. See
973 * update_dl_entity() to know more about such restrictions.
974 *
975 * The dl_is_implicit() returns true if the task has an implicit deadline.
976 */
977static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
978{
979 return dl_se->dl_deadline == dl_se->dl_period;
980}
981
982/*
983 * When a deadline entity is placed in the runqueue, its runtime and deadline
984 * might need to be updated. This is done by a CBS wake up rule. There are two
985 * different rules: 1) the original CBS; and 2) the Revisited CBS.
986 *
987 * When the task is starting a new period, the Original CBS is used. In this
988 * case, the runtime is replenished and a new absolute deadline is set.
989 *
990 * When a task is queued before the begin of the next period, using the
991 * remaining runtime and deadline could make the entity to overflow, see
992 * dl_entity_overflow() to find more about runtime overflow. When such case
993 * is detected, the runtime and deadline need to be updated.
994 *
995 * If the task has an implicit deadline, i.e., deadline == period, the Original
996 * CBS is applied. the runtime is replenished and a new absolute deadline is
997 * set, as in the previous cases.
998 *
999 * However, the Original CBS does not work properly for tasks with
1000 * deadline < period, which are said to have a constrained deadline. By
1001 * applying the Original CBS, a constrained deadline task would be able to run
1002 * runtime/deadline in a period. With deadline < period, the task would
1003 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1004 *
1005 * In order to prevent this misbehave, the Revisited CBS is used for
1006 * constrained deadline tasks when a runtime overflow is detected. In the
1007 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1008 * the remaining runtime of the task is reduced to avoid runtime overflow.
1009 * Please refer to the comments update_dl_revised_wakeup() function to find
1010 * more about the Revised CBS rule.
1011 */
1012static void update_dl_entity(struct sched_dl_entity *dl_se)
1013{
1014 struct rq *rq = rq_of_dl_se(dl_se);
1015
1016 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1017 dl_entity_overflow(dl_se, rq_clock(rq))) {
1018
1019 if (unlikely(!dl_is_implicit(dl_se) &&
1020 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1021 !is_dl_boosted(dl_se))) {
1022 update_dl_revised_wakeup(dl_se, rq);
1023 return;
1024 }
1025
1026 replenish_dl_new_period(dl_se, rq);
1027 }
1028}
1029
1030static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1031{
1032 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1033}
1034
1035/*
1036 * If the entity depleted all its runtime, and if we want it to sleep
1037 * while waiting for some new execution time to become available, we
1038 * set the bandwidth replenishment timer to the replenishment instant
1039 * and try to activate it.
1040 *
1041 * Notice that it is important for the caller to know if the timer
1042 * actually started or not (i.e., the replenishment instant is in
1043 * the future or in the past).
1044 */
1045static int start_dl_timer(struct sched_dl_entity *dl_se)
1046{
1047 struct hrtimer *timer = &dl_se->dl_timer;
1048 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1049 struct rq *rq = rq_of_dl_rq(dl_rq);
1050 ktime_t now, act;
1051 s64 delta;
1052
1053 lockdep_assert_rq_held(rq);
1054
1055 /*
1056 * We want the timer to fire at the deadline, but considering
1057 * that it is actually coming from rq->clock and not from
1058 * hrtimer's time base reading.
1059 */
1060 act = ns_to_ktime(dl_next_period(dl_se));
1061 now = hrtimer_cb_get_time(timer);
1062 delta = ktime_to_ns(now) - rq_clock(rq);
1063 act = ktime_add_ns(act, delta);
1064
1065 /*
1066 * If the expiry time already passed, e.g., because the value
1067 * chosen as the deadline is too small, don't even try to
1068 * start the timer in the past!
1069 */
1070 if (ktime_us_delta(act, now) < 0)
1071 return 0;
1072
1073 /*
1074 * !enqueued will guarantee another callback; even if one is already in
1075 * progress. This ensures a balanced {get,put}_task_struct().
1076 *
1077 * The race against __run_timer() clearing the enqueued state is
1078 * harmless because we're holding task_rq()->lock, therefore the timer
1079 * expiring after we've done the check will wait on its task_rq_lock()
1080 * and observe our state.
1081 */
1082 if (!hrtimer_is_queued(timer)) {
1083 if (!dl_server(dl_se))
1084 get_task_struct(dl_task_of(dl_se));
1085 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1086 }
1087
1088 return 1;
1089}
1090
1091static void __push_dl_task(struct rq *rq, struct rq_flags *rf)
1092{
1093#ifdef CONFIG_SMP
1094 /*
1095 * Queueing this task back might have overloaded rq, check if we need
1096 * to kick someone away.
1097 */
1098 if (has_pushable_dl_tasks(rq)) {
1099 /*
1100 * Nothing relies on rq->lock after this, so its safe to drop
1101 * rq->lock.
1102 */
1103 rq_unpin_lock(rq, rf);
1104 push_dl_task(rq);
1105 rq_repin_lock(rq, rf);
1106 }
1107#endif
1108}
1109
1110/*
1111 * This is the bandwidth enforcement timer callback. If here, we know
1112 * a task is not on its dl_rq, since the fact that the timer was running
1113 * means the task is throttled and needs a runtime replenishment.
1114 *
1115 * However, what we actually do depends on the fact the task is active,
1116 * (it is on its rq) or has been removed from there by a call to
1117 * dequeue_task_dl(). In the former case we must issue the runtime
1118 * replenishment and add the task back to the dl_rq; in the latter, we just
1119 * do nothing but clearing dl_throttled, so that runtime and deadline
1120 * updating (and the queueing back to dl_rq) will be done by the
1121 * next call to enqueue_task_dl().
1122 */
1123static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1124{
1125 struct sched_dl_entity *dl_se = container_of(timer,
1126 struct sched_dl_entity,
1127 dl_timer);
1128 struct task_struct *p;
1129 struct rq_flags rf;
1130 struct rq *rq;
1131
1132 if (dl_server(dl_se)) {
1133 struct rq *rq = rq_of_dl_se(dl_se);
1134 struct rq_flags rf;
1135
1136 rq_lock(rq, &rf);
1137 if (dl_se->dl_throttled) {
1138 sched_clock_tick();
1139 update_rq_clock(rq);
1140
1141 if (dl_se->server_has_tasks(dl_se)) {
1142 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1143 resched_curr(rq);
1144 __push_dl_task(rq, &rf);
1145 } else {
1146 replenish_dl_entity(dl_se);
1147 }
1148
1149 }
1150 rq_unlock(rq, &rf);
1151
1152 return HRTIMER_NORESTART;
1153 }
1154
1155 p = dl_task_of(dl_se);
1156 rq = task_rq_lock(p, &rf);
1157
1158 /*
1159 * The task might have changed its scheduling policy to something
1160 * different than SCHED_DEADLINE (through switched_from_dl()).
1161 */
1162 if (!dl_task(p))
1163 goto unlock;
1164
1165 /*
1166 * The task might have been boosted by someone else and might be in the
1167 * boosting/deboosting path, its not throttled.
1168 */
1169 if (is_dl_boosted(dl_se))
1170 goto unlock;
1171
1172 /*
1173 * Spurious timer due to start_dl_timer() race; or we already received
1174 * a replenishment from rt_mutex_setprio().
1175 */
1176 if (!dl_se->dl_throttled)
1177 goto unlock;
1178
1179 sched_clock_tick();
1180 update_rq_clock(rq);
1181
1182 /*
1183 * If the throttle happened during sched-out; like:
1184 *
1185 * schedule()
1186 * deactivate_task()
1187 * dequeue_task_dl()
1188 * update_curr_dl()
1189 * start_dl_timer()
1190 * __dequeue_task_dl()
1191 * prev->on_rq = 0;
1192 *
1193 * We can be both throttled and !queued. Replenish the counter
1194 * but do not enqueue -- wait for our wakeup to do that.
1195 */
1196 if (!task_on_rq_queued(p)) {
1197 replenish_dl_entity(dl_se);
1198 goto unlock;
1199 }
1200
1201#ifdef CONFIG_SMP
1202 if (unlikely(!rq->online)) {
1203 /*
1204 * If the runqueue is no longer available, migrate the
1205 * task elsewhere. This necessarily changes rq.
1206 */
1207 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1208 rq = dl_task_offline_migration(rq, p);
1209 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1210 update_rq_clock(rq);
1211
1212 /*
1213 * Now that the task has been migrated to the new RQ and we
1214 * have that locked, proceed as normal and enqueue the task
1215 * there.
1216 */
1217 }
1218#endif
1219
1220 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1221 if (dl_task(rq->curr))
1222 wakeup_preempt_dl(rq, p, 0);
1223 else
1224 resched_curr(rq);
1225
1226 __push_dl_task(rq, &rf);
1227
1228unlock:
1229 task_rq_unlock(rq, p, &rf);
1230
1231 /*
1232 * This can free the task_struct, including this hrtimer, do not touch
1233 * anything related to that after this.
1234 */
1235 put_task_struct(p);
1236
1237 return HRTIMER_NORESTART;
1238}
1239
1240static void init_dl_task_timer(struct sched_dl_entity *dl_se)
1241{
1242 struct hrtimer *timer = &dl_se->dl_timer;
1243
1244 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1245 timer->function = dl_task_timer;
1246}
1247
1248/*
1249 * During the activation, CBS checks if it can reuse the current task's
1250 * runtime and period. If the deadline of the task is in the past, CBS
1251 * cannot use the runtime, and so it replenishes the task. This rule
1252 * works fine for implicit deadline tasks (deadline == period), and the
1253 * CBS was designed for implicit deadline tasks. However, a task with
1254 * constrained deadline (deadline < period) might be awakened after the
1255 * deadline, but before the next period. In this case, replenishing the
1256 * task would allow it to run for runtime / deadline. As in this case
1257 * deadline < period, CBS enables a task to run for more than the
1258 * runtime / period. In a very loaded system, this can cause a domino
1259 * effect, making other tasks miss their deadlines.
1260 *
1261 * To avoid this problem, in the activation of a constrained deadline
1262 * task after the deadline but before the next period, throttle the
1263 * task and set the replenishing timer to the begin of the next period,
1264 * unless it is boosted.
1265 */
1266static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1267{
1268 struct rq *rq = rq_of_dl_se(dl_se);
1269
1270 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1271 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1272 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se)))
1273 return;
1274 dl_se->dl_throttled = 1;
1275 if (dl_se->runtime > 0)
1276 dl_se->runtime = 0;
1277 }
1278}
1279
1280static
1281int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1282{
1283 return (dl_se->runtime <= 0);
1284}
1285
1286/*
1287 * This function implements the GRUB accounting rule. According to the
1288 * GRUB reclaiming algorithm, the runtime is not decreased as "dq = -dt",
1289 * but as "dq = -(max{u, (Umax - Uinact - Uextra)} / Umax) dt",
1290 * where u is the utilization of the task, Umax is the maximum reclaimable
1291 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1292 * as the difference between the "total runqueue utilization" and the
1293 * "runqueue active utilization", and Uextra is the (per runqueue) extra
1294 * reclaimable utilization.
1295 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations multiplied
1296 * by 2^BW_SHIFT, the result has to be shifted right by BW_SHIFT.
1297 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT, dl_bw
1298 * is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1299 * Since delta is a 64 bit variable, to have an overflow its value should be
1300 * larger than 2^(64 - 20 - 8), which is more than 64 seconds. So, overflow is
1301 * not an issue here.
1302 */
1303static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1304{
1305 u64 u_act;
1306 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1307
1308 /*
1309 * Instead of computing max{u, (u_max - u_inact - u_extra)}, we
1310 * compare u_inact + u_extra with u_max - u, because u_inact + u_extra
1311 * can be larger than u_max. So, u_max - u_inact - u_extra would be
1312 * negative leading to wrong results.
1313 */
1314 if (u_inact + rq->dl.extra_bw > rq->dl.max_bw - dl_se->dl_bw)
1315 u_act = dl_se->dl_bw;
1316 else
1317 u_act = rq->dl.max_bw - u_inact - rq->dl.extra_bw;
1318
1319 u_act = (u_act * rq->dl.bw_ratio) >> RATIO_SHIFT;
1320 return (delta * u_act) >> BW_SHIFT;
1321}
1322
1323static inline void
1324update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1325 int flags);
1326static void update_curr_dl_se(struct rq *rq, struct sched_dl_entity *dl_se, s64 delta_exec)
1327{
1328 s64 scaled_delta_exec;
1329
1330 if (unlikely(delta_exec <= 0)) {
1331 if (unlikely(dl_se->dl_yielded))
1332 goto throttle;
1333 return;
1334 }
1335
1336 if (dl_entity_is_special(dl_se))
1337 return;
1338
1339 /*
1340 * For tasks that participate in GRUB, we implement GRUB-PA: the
1341 * spare reclaimed bandwidth is used to clock down frequency.
1342 *
1343 * For the others, we still need to scale reservation parameters
1344 * according to current frequency and CPU maximum capacity.
1345 */
1346 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1347 scaled_delta_exec = grub_reclaim(delta_exec, rq, dl_se);
1348 } else {
1349 int cpu = cpu_of(rq);
1350 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1351 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1352
1353 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1354 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1355 }
1356
1357 dl_se->runtime -= scaled_delta_exec;
1358
1359throttle:
1360 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1361 dl_se->dl_throttled = 1;
1362
1363 /* If requested, inform the user about runtime overruns. */
1364 if (dl_runtime_exceeded(dl_se) &&
1365 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1366 dl_se->dl_overrun = 1;
1367
1368 dequeue_dl_entity(dl_se, 0);
1369 if (!dl_server(dl_se)) {
1370 update_stats_dequeue_dl(&rq->dl, dl_se, 0);
1371 dequeue_pushable_dl_task(rq, dl_task_of(dl_se));
1372 }
1373
1374 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se))) {
1375 if (dl_server(dl_se))
1376 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1377 else
1378 enqueue_task_dl(rq, dl_task_of(dl_se), ENQUEUE_REPLENISH);
1379 }
1380
1381 if (!is_leftmost(dl_se, &rq->dl))
1382 resched_curr(rq);
1383 }
1384
1385 /*
1386 * Because -- for now -- we share the rt bandwidth, we need to
1387 * account our runtime there too, otherwise actual rt tasks
1388 * would be able to exceed the shared quota.
1389 *
1390 * Account to the root rt group for now.
1391 *
1392 * The solution we're working towards is having the RT groups scheduled
1393 * using deadline servers -- however there's a few nasties to figure
1394 * out before that can happen.
1395 */
1396 if (rt_bandwidth_enabled()) {
1397 struct rt_rq *rt_rq = &rq->rt;
1398
1399 raw_spin_lock(&rt_rq->rt_runtime_lock);
1400 /*
1401 * We'll let actual RT tasks worry about the overflow here, we
1402 * have our own CBS to keep us inline; only account when RT
1403 * bandwidth is relevant.
1404 */
1405 if (sched_rt_bandwidth_account(rt_rq))
1406 rt_rq->rt_time += delta_exec;
1407 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1408 }
1409}
1410
1411void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec)
1412{
1413 update_curr_dl_se(dl_se->rq, dl_se, delta_exec);
1414}
1415
1416void dl_server_start(struct sched_dl_entity *dl_se)
1417{
1418 if (!dl_server(dl_se)) {
1419 dl_se->dl_server = 1;
1420 setup_new_dl_entity(dl_se);
1421 }
1422 enqueue_dl_entity(dl_se, ENQUEUE_WAKEUP);
1423}
1424
1425void dl_server_stop(struct sched_dl_entity *dl_se)
1426{
1427 dequeue_dl_entity(dl_se, DEQUEUE_SLEEP);
1428}
1429
1430void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
1431 dl_server_has_tasks_f has_tasks,
1432 dl_server_pick_f pick)
1433{
1434 dl_se->rq = rq;
1435 dl_se->server_has_tasks = has_tasks;
1436 dl_se->server_pick = pick;
1437}
1438
1439/*
1440 * Update the current task's runtime statistics (provided it is still
1441 * a -deadline task and has not been removed from the dl_rq).
1442 */
1443static void update_curr_dl(struct rq *rq)
1444{
1445 struct task_struct *curr = rq->curr;
1446 struct sched_dl_entity *dl_se = &curr->dl;
1447 s64 delta_exec;
1448
1449 if (!dl_task(curr) || !on_dl_rq(dl_se))
1450 return;
1451
1452 /*
1453 * Consumed budget is computed considering the time as
1454 * observed by schedulable tasks (excluding time spent
1455 * in hardirq context, etc.). Deadlines are instead
1456 * computed using hard walltime. This seems to be the more
1457 * natural solution, but the full ramifications of this
1458 * approach need further study.
1459 */
1460 delta_exec = update_curr_common(rq);
1461 update_curr_dl_se(rq, dl_se, delta_exec);
1462}
1463
1464static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1465{
1466 struct sched_dl_entity *dl_se = container_of(timer,
1467 struct sched_dl_entity,
1468 inactive_timer);
1469 struct task_struct *p = NULL;
1470 struct rq_flags rf;
1471 struct rq *rq;
1472
1473 if (!dl_server(dl_se)) {
1474 p = dl_task_of(dl_se);
1475 rq = task_rq_lock(p, &rf);
1476 } else {
1477 rq = dl_se->rq;
1478 rq_lock(rq, &rf);
1479 }
1480
1481 sched_clock_tick();
1482 update_rq_clock(rq);
1483
1484 if (dl_server(dl_se))
1485 goto no_task;
1486
1487 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1488 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1489
1490 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1491 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1492 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1493 dl_se->dl_non_contending = 0;
1494 }
1495
1496 raw_spin_lock(&dl_b->lock);
1497 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1498 raw_spin_unlock(&dl_b->lock);
1499 __dl_clear_params(dl_se);
1500
1501 goto unlock;
1502 }
1503
1504no_task:
1505 if (dl_se->dl_non_contending == 0)
1506 goto unlock;
1507
1508 sub_running_bw(dl_se, &rq->dl);
1509 dl_se->dl_non_contending = 0;
1510unlock:
1511
1512 if (!dl_server(dl_se)) {
1513 task_rq_unlock(rq, p, &rf);
1514 put_task_struct(p);
1515 } else {
1516 rq_unlock(rq, &rf);
1517 }
1518
1519 return HRTIMER_NORESTART;
1520}
1521
1522static void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1523{
1524 struct hrtimer *timer = &dl_se->inactive_timer;
1525
1526 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1527 timer->function = inactive_task_timer;
1528}
1529
1530#define __node_2_dle(node) \
1531 rb_entry((node), struct sched_dl_entity, rb_node)
1532
1533#ifdef CONFIG_SMP
1534
1535static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1536{
1537 struct rq *rq = rq_of_dl_rq(dl_rq);
1538
1539 if (dl_rq->earliest_dl.curr == 0 ||
1540 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1541 if (dl_rq->earliest_dl.curr == 0)
1542 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1543 dl_rq->earliest_dl.curr = deadline;
1544 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1545 }
1546}
1547
1548static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1549{
1550 struct rq *rq = rq_of_dl_rq(dl_rq);
1551
1552 /*
1553 * Since we may have removed our earliest (and/or next earliest)
1554 * task we must recompute them.
1555 */
1556 if (!dl_rq->dl_nr_running) {
1557 dl_rq->earliest_dl.curr = 0;
1558 dl_rq->earliest_dl.next = 0;
1559 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1560 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1561 } else {
1562 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1563 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1564
1565 dl_rq->earliest_dl.curr = entry->deadline;
1566 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1567 }
1568}
1569
1570#else
1571
1572static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1573static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1574
1575#endif /* CONFIG_SMP */
1576
1577static inline
1578void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1579{
1580 u64 deadline = dl_se->deadline;
1581
1582 dl_rq->dl_nr_running++;
1583 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1584
1585 inc_dl_deadline(dl_rq, deadline);
1586}
1587
1588static inline
1589void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1590{
1591 WARN_ON(!dl_rq->dl_nr_running);
1592 dl_rq->dl_nr_running--;
1593 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1594
1595 dec_dl_deadline(dl_rq, dl_se->deadline);
1596}
1597
1598static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1599{
1600 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1601}
1602
1603static inline struct sched_statistics *
1604__schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1605{
1606 return &dl_task_of(dl_se)->stats;
1607}
1608
1609static inline void
1610update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1611{
1612 struct sched_statistics *stats;
1613
1614 if (!schedstat_enabled())
1615 return;
1616
1617 stats = __schedstats_from_dl_se(dl_se);
1618 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1619}
1620
1621static inline void
1622update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1623{
1624 struct sched_statistics *stats;
1625
1626 if (!schedstat_enabled())
1627 return;
1628
1629 stats = __schedstats_from_dl_se(dl_se);
1630 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1631}
1632
1633static inline void
1634update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1635{
1636 struct sched_statistics *stats;
1637
1638 if (!schedstat_enabled())
1639 return;
1640
1641 stats = __schedstats_from_dl_se(dl_se);
1642 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1643}
1644
1645static inline void
1646update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1647 int flags)
1648{
1649 if (!schedstat_enabled())
1650 return;
1651
1652 if (flags & ENQUEUE_WAKEUP)
1653 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1654}
1655
1656static inline void
1657update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1658 int flags)
1659{
1660 struct task_struct *p = dl_task_of(dl_se);
1661
1662 if (!schedstat_enabled())
1663 return;
1664
1665 if ((flags & DEQUEUE_SLEEP)) {
1666 unsigned int state;
1667
1668 state = READ_ONCE(p->__state);
1669 if (state & TASK_INTERRUPTIBLE)
1670 __schedstat_set(p->stats.sleep_start,
1671 rq_clock(rq_of_dl_rq(dl_rq)));
1672
1673 if (state & TASK_UNINTERRUPTIBLE)
1674 __schedstat_set(p->stats.block_start,
1675 rq_clock(rq_of_dl_rq(dl_rq)));
1676 }
1677}
1678
1679static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1680{
1681 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1682
1683 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1684
1685 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1686
1687 inc_dl_tasks(dl_se, dl_rq);
1688}
1689
1690static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1691{
1692 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1693
1694 if (RB_EMPTY_NODE(&dl_se->rb_node))
1695 return;
1696
1697 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1698
1699 RB_CLEAR_NODE(&dl_se->rb_node);
1700
1701 dec_dl_tasks(dl_se, dl_rq);
1702}
1703
1704static void
1705enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1706{
1707 WARN_ON_ONCE(on_dl_rq(dl_se));
1708
1709 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1710
1711 /*
1712 * Check if a constrained deadline task was activated
1713 * after the deadline but before the next period.
1714 * If that is the case, the task will be throttled and
1715 * the replenishment timer will be set to the next period.
1716 */
1717 if (!dl_se->dl_throttled && !dl_is_implicit(dl_se))
1718 dl_check_constrained_dl(dl_se);
1719
1720 if (flags & (ENQUEUE_RESTORE|ENQUEUE_MIGRATING)) {
1721 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1722
1723 add_rq_bw(dl_se, dl_rq);
1724 add_running_bw(dl_se, dl_rq);
1725 }
1726
1727 /*
1728 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1729 * its budget it needs a replenishment and, since it now is on
1730 * its rq, the bandwidth timer callback (which clearly has not
1731 * run yet) will take care of this.
1732 * However, the active utilization does not depend on the fact
1733 * that the task is on the runqueue or not (but depends on the
1734 * task's state - in GRUB parlance, "inactive" vs "active contending").
1735 * In other words, even if a task is throttled its utilization must
1736 * be counted in the active utilization; hence, we need to call
1737 * add_running_bw().
1738 */
1739 if (dl_se->dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1740 if (flags & ENQUEUE_WAKEUP)
1741 task_contending(dl_se, flags);
1742
1743 return;
1744 }
1745
1746 /*
1747 * If this is a wakeup or a new instance, the scheduling
1748 * parameters of the task might need updating. Otherwise,
1749 * we want a replenishment of its runtime.
1750 */
1751 if (flags & ENQUEUE_WAKEUP) {
1752 task_contending(dl_se, flags);
1753 update_dl_entity(dl_se);
1754 } else if (flags & ENQUEUE_REPLENISH) {
1755 replenish_dl_entity(dl_se);
1756 } else if ((flags & ENQUEUE_RESTORE) &&
1757 dl_time_before(dl_se->deadline, rq_clock(rq_of_dl_se(dl_se)))) {
1758 setup_new_dl_entity(dl_se);
1759 }
1760
1761 __enqueue_dl_entity(dl_se);
1762}
1763
1764static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1765{
1766 __dequeue_dl_entity(dl_se);
1767
1768 if (flags & (DEQUEUE_SAVE|DEQUEUE_MIGRATING)) {
1769 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1770
1771 sub_running_bw(dl_se, dl_rq);
1772 sub_rq_bw(dl_se, dl_rq);
1773 }
1774
1775 /*
1776 * This check allows to start the inactive timer (or to immediately
1777 * decrease the active utilization, if needed) in two cases:
1778 * when the task blocks and when it is terminating
1779 * (p->state == TASK_DEAD). We can handle the two cases in the same
1780 * way, because from GRUB's point of view the same thing is happening
1781 * (the task moves from "active contending" to "active non contending"
1782 * or "inactive")
1783 */
1784 if (flags & DEQUEUE_SLEEP)
1785 task_non_contending(dl_se);
1786}
1787
1788static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1789{
1790 if (is_dl_boosted(&p->dl)) {
1791 /*
1792 * Because of delays in the detection of the overrun of a
1793 * thread's runtime, it might be the case that a thread
1794 * goes to sleep in a rt mutex with negative runtime. As
1795 * a consequence, the thread will be throttled.
1796 *
1797 * While waiting for the mutex, this thread can also be
1798 * boosted via PI, resulting in a thread that is throttled
1799 * and boosted at the same time.
1800 *
1801 * In this case, the boost overrides the throttle.
1802 */
1803 if (p->dl.dl_throttled) {
1804 /*
1805 * The replenish timer needs to be canceled. No
1806 * problem if it fires concurrently: boosted threads
1807 * are ignored in dl_task_timer().
1808 */
1809 hrtimer_try_to_cancel(&p->dl.dl_timer);
1810 p->dl.dl_throttled = 0;
1811 }
1812 } else if (!dl_prio(p->normal_prio)) {
1813 /*
1814 * Special case in which we have a !SCHED_DEADLINE task that is going
1815 * to be deboosted, but exceeds its runtime while doing so. No point in
1816 * replenishing it, as it's going to return back to its original
1817 * scheduling class after this. If it has been throttled, we need to
1818 * clear the flag, otherwise the task may wake up as throttled after
1819 * being boosted again with no means to replenish the runtime and clear
1820 * the throttle.
1821 */
1822 p->dl.dl_throttled = 0;
1823 if (!(flags & ENQUEUE_REPLENISH))
1824 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
1825 task_pid_nr(p));
1826
1827 return;
1828 }
1829
1830 check_schedstat_required();
1831 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
1832
1833 if (p->on_rq == TASK_ON_RQ_MIGRATING)
1834 flags |= ENQUEUE_MIGRATING;
1835
1836 enqueue_dl_entity(&p->dl, flags);
1837
1838 if (dl_server(&p->dl))
1839 return;
1840
1841 if (!task_current(rq, p) && !p->dl.dl_throttled && p->nr_cpus_allowed > 1)
1842 enqueue_pushable_dl_task(rq, p);
1843}
1844
1845static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1846{
1847 update_curr_dl(rq);
1848
1849 if (p->on_rq == TASK_ON_RQ_MIGRATING)
1850 flags |= DEQUEUE_MIGRATING;
1851
1852 dequeue_dl_entity(&p->dl, flags);
1853 if (!p->dl.dl_throttled && !dl_server(&p->dl))
1854 dequeue_pushable_dl_task(rq, p);
1855}
1856
1857/*
1858 * Yield task semantic for -deadline tasks is:
1859 *
1860 * get off from the CPU until our next instance, with
1861 * a new runtime. This is of little use now, since we
1862 * don't have a bandwidth reclaiming mechanism. Anyway,
1863 * bandwidth reclaiming is planned for the future, and
1864 * yield_task_dl will indicate that some spare budget
1865 * is available for other task instances to use it.
1866 */
1867static void yield_task_dl(struct rq *rq)
1868{
1869 /*
1870 * We make the task go to sleep until its current deadline by
1871 * forcing its runtime to zero. This way, update_curr_dl() stops
1872 * it and the bandwidth timer will wake it up and will give it
1873 * new scheduling parameters (thanks to dl_yielded=1).
1874 */
1875 rq->curr->dl.dl_yielded = 1;
1876
1877 update_rq_clock(rq);
1878 update_curr_dl(rq);
1879 /*
1880 * Tell update_rq_clock() that we've just updated,
1881 * so we don't do microscopic update in schedule()
1882 * and double the fastpath cost.
1883 */
1884 rq_clock_skip_update(rq);
1885}
1886
1887#ifdef CONFIG_SMP
1888
1889static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
1890 struct rq *rq)
1891{
1892 return (!rq->dl.dl_nr_running ||
1893 dl_time_before(p->dl.deadline,
1894 rq->dl.earliest_dl.curr));
1895}
1896
1897static int find_later_rq(struct task_struct *task);
1898
1899static int
1900select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1901{
1902 struct task_struct *curr;
1903 bool select_rq;
1904 struct rq *rq;
1905
1906 if (!(flags & WF_TTWU))
1907 goto out;
1908
1909 rq = cpu_rq(cpu);
1910
1911 rcu_read_lock();
1912 curr = READ_ONCE(rq->curr); /* unlocked access */
1913
1914 /*
1915 * If we are dealing with a -deadline task, we must
1916 * decide where to wake it up.
1917 * If it has a later deadline and the current task
1918 * on this rq can't move (provided the waking task
1919 * can!) we prefer to send it somewhere else. On the
1920 * other hand, if it has a shorter deadline, we
1921 * try to make it stay here, it might be important.
1922 */
1923 select_rq = unlikely(dl_task(curr)) &&
1924 (curr->nr_cpus_allowed < 2 ||
1925 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1926 p->nr_cpus_allowed > 1;
1927
1928 /*
1929 * Take the capacity of the CPU into account to
1930 * ensure it fits the requirement of the task.
1931 */
1932 if (sched_asym_cpucap_active())
1933 select_rq |= !dl_task_fits_capacity(p, cpu);
1934
1935 if (select_rq) {
1936 int target = find_later_rq(p);
1937
1938 if (target != -1 &&
1939 dl_task_is_earliest_deadline(p, cpu_rq(target)))
1940 cpu = target;
1941 }
1942 rcu_read_unlock();
1943
1944out:
1945 return cpu;
1946}
1947
1948static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1949{
1950 struct rq_flags rf;
1951 struct rq *rq;
1952
1953 if (READ_ONCE(p->__state) != TASK_WAKING)
1954 return;
1955
1956 rq = task_rq(p);
1957 /*
1958 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1959 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1960 * rq->lock is not... So, lock it
1961 */
1962 rq_lock(rq, &rf);
1963 if (p->dl.dl_non_contending) {
1964 update_rq_clock(rq);
1965 sub_running_bw(&p->dl, &rq->dl);
1966 p->dl.dl_non_contending = 0;
1967 /*
1968 * If the timer handler is currently running and the
1969 * timer cannot be canceled, inactive_task_timer()
1970 * will see that dl_not_contending is not set, and
1971 * will not touch the rq's active utilization,
1972 * so we are still safe.
1973 */
1974 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1975 put_task_struct(p);
1976 }
1977 sub_rq_bw(&p->dl, &rq->dl);
1978 rq_unlock(rq, &rf);
1979}
1980
1981static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1982{
1983 /*
1984 * Current can't be migrated, useless to reschedule,
1985 * let's hope p can move out.
1986 */
1987 if (rq->curr->nr_cpus_allowed == 1 ||
1988 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1989 return;
1990
1991 /*
1992 * p is migratable, so let's not schedule it and
1993 * see if it is pushed or pulled somewhere else.
1994 */
1995 if (p->nr_cpus_allowed != 1 &&
1996 cpudl_find(&rq->rd->cpudl, p, NULL))
1997 return;
1998
1999 resched_curr(rq);
2000}
2001
2002static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
2003{
2004 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
2005 /*
2006 * This is OK, because current is on_cpu, which avoids it being
2007 * picked for load-balance and preemption/IRQs are still
2008 * disabled avoiding further scheduler activity on it and we've
2009 * not yet started the picking loop.
2010 */
2011 rq_unpin_lock(rq, rf);
2012 pull_dl_task(rq);
2013 rq_repin_lock(rq, rf);
2014 }
2015
2016 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
2017}
2018#endif /* CONFIG_SMP */
2019
2020/*
2021 * Only called when both the current and waking task are -deadline
2022 * tasks.
2023 */
2024static void wakeup_preempt_dl(struct rq *rq, struct task_struct *p,
2025 int flags)
2026{
2027 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
2028 resched_curr(rq);
2029 return;
2030 }
2031
2032#ifdef CONFIG_SMP
2033 /*
2034 * In the unlikely case current and p have the same deadline
2035 * let us try to decide what's the best thing to do...
2036 */
2037 if ((p->dl.deadline == rq->curr->dl.deadline) &&
2038 !test_tsk_need_resched(rq->curr))
2039 check_preempt_equal_dl(rq, p);
2040#endif /* CONFIG_SMP */
2041}
2042
2043#ifdef CONFIG_SCHED_HRTICK
2044static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2045{
2046 hrtick_start(rq, dl_se->runtime);
2047}
2048#else /* !CONFIG_SCHED_HRTICK */
2049static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2050{
2051}
2052#endif
2053
2054static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
2055{
2056 struct sched_dl_entity *dl_se = &p->dl;
2057 struct dl_rq *dl_rq = &rq->dl;
2058
2059 p->se.exec_start = rq_clock_task(rq);
2060 if (on_dl_rq(&p->dl))
2061 update_stats_wait_end_dl(dl_rq, dl_se);
2062
2063 /* You can't push away the running task */
2064 dequeue_pushable_dl_task(rq, p);
2065
2066 if (!first)
2067 return;
2068
2069 if (rq->curr->sched_class != &dl_sched_class)
2070 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2071
2072 deadline_queue_push_tasks(rq);
2073}
2074
2075static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
2076{
2077 struct rb_node *left = rb_first_cached(&dl_rq->root);
2078
2079 if (!left)
2080 return NULL;
2081
2082 return __node_2_dle(left);
2083}
2084
2085static struct task_struct *pick_task_dl(struct rq *rq)
2086{
2087 struct sched_dl_entity *dl_se;
2088 struct dl_rq *dl_rq = &rq->dl;
2089 struct task_struct *p;
2090
2091again:
2092 if (!sched_dl_runnable(rq))
2093 return NULL;
2094
2095 dl_se = pick_next_dl_entity(dl_rq);
2096 WARN_ON_ONCE(!dl_se);
2097
2098 if (dl_server(dl_se)) {
2099 p = dl_se->server_pick(dl_se);
2100 if (!p) {
2101 WARN_ON_ONCE(1);
2102 dl_se->dl_yielded = 1;
2103 update_curr_dl_se(rq, dl_se, 0);
2104 goto again;
2105 }
2106 p->dl_server = dl_se;
2107 } else {
2108 p = dl_task_of(dl_se);
2109 }
2110
2111 return p;
2112}
2113
2114static struct task_struct *pick_next_task_dl(struct rq *rq)
2115{
2116 struct task_struct *p;
2117
2118 p = pick_task_dl(rq);
2119 if (!p)
2120 return p;
2121
2122 if (!p->dl_server)
2123 set_next_task_dl(rq, p, true);
2124
2125 if (hrtick_enabled(rq))
2126 start_hrtick_dl(rq, &p->dl);
2127
2128 return p;
2129}
2130
2131static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2132{
2133 struct sched_dl_entity *dl_se = &p->dl;
2134 struct dl_rq *dl_rq = &rq->dl;
2135
2136 if (on_dl_rq(&p->dl))
2137 update_stats_wait_start_dl(dl_rq, dl_se);
2138
2139 update_curr_dl(rq);
2140
2141 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2142 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2143 enqueue_pushable_dl_task(rq, p);
2144}
2145
2146/*
2147 * scheduler tick hitting a task of our scheduling class.
2148 *
2149 * NOTE: This function can be called remotely by the tick offload that
2150 * goes along full dynticks. Therefore no local assumption can be made
2151 * and everything must be accessed through the @rq and @curr passed in
2152 * parameters.
2153 */
2154static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2155{
2156 update_curr_dl(rq);
2157
2158 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2159 /*
2160 * Even when we have runtime, update_curr_dl() might have resulted in us
2161 * not being the leftmost task anymore. In that case NEED_RESCHED will
2162 * be set and schedule() will start a new hrtick for the next task.
2163 */
2164 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2165 is_leftmost(&p->dl, &rq->dl))
2166 start_hrtick_dl(rq, &p->dl);
2167}
2168
2169static void task_fork_dl(struct task_struct *p)
2170{
2171 /*
2172 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2173 * sched_fork()
2174 */
2175}
2176
2177#ifdef CONFIG_SMP
2178
2179/* Only try algorithms three times */
2180#define DL_MAX_TRIES 3
2181
2182static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2183{
2184 if (!task_on_cpu(rq, p) &&
2185 cpumask_test_cpu(cpu, &p->cpus_mask))
2186 return 1;
2187 return 0;
2188}
2189
2190/*
2191 * Return the earliest pushable rq's task, which is suitable to be executed
2192 * on the CPU, NULL otherwise:
2193 */
2194static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2195{
2196 struct task_struct *p = NULL;
2197 struct rb_node *next_node;
2198
2199 if (!has_pushable_dl_tasks(rq))
2200 return NULL;
2201
2202 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2203
2204next_node:
2205 if (next_node) {
2206 p = __node_2_pdl(next_node);
2207
2208 if (pick_dl_task(rq, p, cpu))
2209 return p;
2210
2211 next_node = rb_next(next_node);
2212 goto next_node;
2213 }
2214
2215 return NULL;
2216}
2217
2218static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2219
2220static int find_later_rq(struct task_struct *task)
2221{
2222 struct sched_domain *sd;
2223 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2224 int this_cpu = smp_processor_id();
2225 int cpu = task_cpu(task);
2226
2227 /* Make sure the mask is initialized first */
2228 if (unlikely(!later_mask))
2229 return -1;
2230
2231 if (task->nr_cpus_allowed == 1)
2232 return -1;
2233
2234 /*
2235 * We have to consider system topology and task affinity
2236 * first, then we can look for a suitable CPU.
2237 */
2238 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2239 return -1;
2240
2241 /*
2242 * If we are here, some targets have been found, including
2243 * the most suitable which is, among the runqueues where the
2244 * current tasks have later deadlines than the task's one, the
2245 * rq with the latest possible one.
2246 *
2247 * Now we check how well this matches with task's
2248 * affinity and system topology.
2249 *
2250 * The last CPU where the task run is our first
2251 * guess, since it is most likely cache-hot there.
2252 */
2253 if (cpumask_test_cpu(cpu, later_mask))
2254 return cpu;
2255 /*
2256 * Check if this_cpu is to be skipped (i.e., it is
2257 * not in the mask) or not.
2258 */
2259 if (!cpumask_test_cpu(this_cpu, later_mask))
2260 this_cpu = -1;
2261
2262 rcu_read_lock();
2263 for_each_domain(cpu, sd) {
2264 if (sd->flags & SD_WAKE_AFFINE) {
2265 int best_cpu;
2266
2267 /*
2268 * If possible, preempting this_cpu is
2269 * cheaper than migrating.
2270 */
2271 if (this_cpu != -1 &&
2272 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2273 rcu_read_unlock();
2274 return this_cpu;
2275 }
2276
2277 best_cpu = cpumask_any_and_distribute(later_mask,
2278 sched_domain_span(sd));
2279 /*
2280 * Last chance: if a CPU being in both later_mask
2281 * and current sd span is valid, that becomes our
2282 * choice. Of course, the latest possible CPU is
2283 * already under consideration through later_mask.
2284 */
2285 if (best_cpu < nr_cpu_ids) {
2286 rcu_read_unlock();
2287 return best_cpu;
2288 }
2289 }
2290 }
2291 rcu_read_unlock();
2292
2293 /*
2294 * At this point, all our guesses failed, we just return
2295 * 'something', and let the caller sort the things out.
2296 */
2297 if (this_cpu != -1)
2298 return this_cpu;
2299
2300 cpu = cpumask_any_distribute(later_mask);
2301 if (cpu < nr_cpu_ids)
2302 return cpu;
2303
2304 return -1;
2305}
2306
2307/* Locks the rq it finds */
2308static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2309{
2310 struct rq *later_rq = NULL;
2311 int tries;
2312 int cpu;
2313
2314 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2315 cpu = find_later_rq(task);
2316
2317 if ((cpu == -1) || (cpu == rq->cpu))
2318 break;
2319
2320 later_rq = cpu_rq(cpu);
2321
2322 if (!dl_task_is_earliest_deadline(task, later_rq)) {
2323 /*
2324 * Target rq has tasks of equal or earlier deadline,
2325 * retrying does not release any lock and is unlikely
2326 * to yield a different result.
2327 */
2328 later_rq = NULL;
2329 break;
2330 }
2331
2332 /* Retry if something changed. */
2333 if (double_lock_balance(rq, later_rq)) {
2334 if (unlikely(task_rq(task) != rq ||
2335 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2336 task_on_cpu(rq, task) ||
2337 !dl_task(task) ||
2338 is_migration_disabled(task) ||
2339 !task_on_rq_queued(task))) {
2340 double_unlock_balance(rq, later_rq);
2341 later_rq = NULL;
2342 break;
2343 }
2344 }
2345
2346 /*
2347 * If the rq we found has no -deadline task, or
2348 * its earliest one has a later deadline than our
2349 * task, the rq is a good one.
2350 */
2351 if (dl_task_is_earliest_deadline(task, later_rq))
2352 break;
2353
2354 /* Otherwise we try again. */
2355 double_unlock_balance(rq, later_rq);
2356 later_rq = NULL;
2357 }
2358
2359 return later_rq;
2360}
2361
2362static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2363{
2364 struct task_struct *p;
2365
2366 if (!has_pushable_dl_tasks(rq))
2367 return NULL;
2368
2369 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2370
2371 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2372 WARN_ON_ONCE(task_current(rq, p));
2373 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2374
2375 WARN_ON_ONCE(!task_on_rq_queued(p));
2376 WARN_ON_ONCE(!dl_task(p));
2377
2378 return p;
2379}
2380
2381/*
2382 * See if the non running -deadline tasks on this rq
2383 * can be sent to some other CPU where they can preempt
2384 * and start executing.
2385 */
2386static int push_dl_task(struct rq *rq)
2387{
2388 struct task_struct *next_task;
2389 struct rq *later_rq;
2390 int ret = 0;
2391
2392 next_task = pick_next_pushable_dl_task(rq);
2393 if (!next_task)
2394 return 0;
2395
2396retry:
2397 /*
2398 * If next_task preempts rq->curr, and rq->curr
2399 * can move away, it makes sense to just reschedule
2400 * without going further in pushing next_task.
2401 */
2402 if (dl_task(rq->curr) &&
2403 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2404 rq->curr->nr_cpus_allowed > 1) {
2405 resched_curr(rq);
2406 return 0;
2407 }
2408
2409 if (is_migration_disabled(next_task))
2410 return 0;
2411
2412 if (WARN_ON(next_task == rq->curr))
2413 return 0;
2414
2415 /* We might release rq lock */
2416 get_task_struct(next_task);
2417
2418 /* Will lock the rq it'll find */
2419 later_rq = find_lock_later_rq(next_task, rq);
2420 if (!later_rq) {
2421 struct task_struct *task;
2422
2423 /*
2424 * We must check all this again, since
2425 * find_lock_later_rq releases rq->lock and it is
2426 * then possible that next_task has migrated.
2427 */
2428 task = pick_next_pushable_dl_task(rq);
2429 if (task == next_task) {
2430 /*
2431 * The task is still there. We don't try
2432 * again, some other CPU will pull it when ready.
2433 */
2434 goto out;
2435 }
2436
2437 if (!task)
2438 /* No more tasks */
2439 goto out;
2440
2441 put_task_struct(next_task);
2442 next_task = task;
2443 goto retry;
2444 }
2445
2446 deactivate_task(rq, next_task, 0);
2447 set_task_cpu(next_task, later_rq->cpu);
2448 activate_task(later_rq, next_task, 0);
2449 ret = 1;
2450
2451 resched_curr(later_rq);
2452
2453 double_unlock_balance(rq, later_rq);
2454
2455out:
2456 put_task_struct(next_task);
2457
2458 return ret;
2459}
2460
2461static void push_dl_tasks(struct rq *rq)
2462{
2463 /* push_dl_task() will return true if it moved a -deadline task */
2464 while (push_dl_task(rq))
2465 ;
2466}
2467
2468static void pull_dl_task(struct rq *this_rq)
2469{
2470 int this_cpu = this_rq->cpu, cpu;
2471 struct task_struct *p, *push_task;
2472 bool resched = false;
2473 struct rq *src_rq;
2474 u64 dmin = LONG_MAX;
2475
2476 if (likely(!dl_overloaded(this_rq)))
2477 return;
2478
2479 /*
2480 * Match the barrier from dl_set_overloaded; this guarantees that if we
2481 * see overloaded we must also see the dlo_mask bit.
2482 */
2483 smp_rmb();
2484
2485 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2486 if (this_cpu == cpu)
2487 continue;
2488
2489 src_rq = cpu_rq(cpu);
2490
2491 /*
2492 * It looks racy, abd it is! However, as in sched_rt.c,
2493 * we are fine with this.
2494 */
2495 if (this_rq->dl.dl_nr_running &&
2496 dl_time_before(this_rq->dl.earliest_dl.curr,
2497 src_rq->dl.earliest_dl.next))
2498 continue;
2499
2500 /* Might drop this_rq->lock */
2501 push_task = NULL;
2502 double_lock_balance(this_rq, src_rq);
2503
2504 /*
2505 * If there are no more pullable tasks on the
2506 * rq, we're done with it.
2507 */
2508 if (src_rq->dl.dl_nr_running <= 1)
2509 goto skip;
2510
2511 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2512
2513 /*
2514 * We found a task to be pulled if:
2515 * - it preempts our current (if there's one),
2516 * - it will preempt the last one we pulled (if any).
2517 */
2518 if (p && dl_time_before(p->dl.deadline, dmin) &&
2519 dl_task_is_earliest_deadline(p, this_rq)) {
2520 WARN_ON(p == src_rq->curr);
2521 WARN_ON(!task_on_rq_queued(p));
2522
2523 /*
2524 * Then we pull iff p has actually an earlier
2525 * deadline than the current task of its runqueue.
2526 */
2527 if (dl_time_before(p->dl.deadline,
2528 src_rq->curr->dl.deadline))
2529 goto skip;
2530
2531 if (is_migration_disabled(p)) {
2532 push_task = get_push_task(src_rq);
2533 } else {
2534 deactivate_task(src_rq, p, 0);
2535 set_task_cpu(p, this_cpu);
2536 activate_task(this_rq, p, 0);
2537 dmin = p->dl.deadline;
2538 resched = true;
2539 }
2540
2541 /* Is there any other task even earlier? */
2542 }
2543skip:
2544 double_unlock_balance(this_rq, src_rq);
2545
2546 if (push_task) {
2547 preempt_disable();
2548 raw_spin_rq_unlock(this_rq);
2549 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2550 push_task, &src_rq->push_work);
2551 preempt_enable();
2552 raw_spin_rq_lock(this_rq);
2553 }
2554 }
2555
2556 if (resched)
2557 resched_curr(this_rq);
2558}
2559
2560/*
2561 * Since the task is not running and a reschedule is not going to happen
2562 * anytime soon on its runqueue, we try pushing it away now.
2563 */
2564static void task_woken_dl(struct rq *rq, struct task_struct *p)
2565{
2566 if (!task_on_cpu(rq, p) &&
2567 !test_tsk_need_resched(rq->curr) &&
2568 p->nr_cpus_allowed > 1 &&
2569 dl_task(rq->curr) &&
2570 (rq->curr->nr_cpus_allowed < 2 ||
2571 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2572 push_dl_tasks(rq);
2573 }
2574}
2575
2576static void set_cpus_allowed_dl(struct task_struct *p,
2577 struct affinity_context *ctx)
2578{
2579 struct root_domain *src_rd;
2580 struct rq *rq;
2581
2582 WARN_ON_ONCE(!dl_task(p));
2583
2584 rq = task_rq(p);
2585 src_rd = rq->rd;
2586 /*
2587 * Migrating a SCHED_DEADLINE task between exclusive
2588 * cpusets (different root_domains) entails a bandwidth
2589 * update. We already made space for us in the destination
2590 * domain (see cpuset_can_attach()).
2591 */
2592 if (!cpumask_intersects(src_rd->span, ctx->new_mask)) {
2593 struct dl_bw *src_dl_b;
2594
2595 src_dl_b = dl_bw_of(cpu_of(rq));
2596 /*
2597 * We now free resources of the root_domain we are migrating
2598 * off. In the worst case, sched_setattr() may temporary fail
2599 * until we complete the update.
2600 */
2601 raw_spin_lock(&src_dl_b->lock);
2602 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2603 raw_spin_unlock(&src_dl_b->lock);
2604 }
2605
2606 set_cpus_allowed_common(p, ctx);
2607}
2608
2609/* Assumes rq->lock is held */
2610static void rq_online_dl(struct rq *rq)
2611{
2612 if (rq->dl.overloaded)
2613 dl_set_overload(rq);
2614
2615 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2616 if (rq->dl.dl_nr_running > 0)
2617 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2618}
2619
2620/* Assumes rq->lock is held */
2621static void rq_offline_dl(struct rq *rq)
2622{
2623 if (rq->dl.overloaded)
2624 dl_clear_overload(rq);
2625
2626 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2627 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2628}
2629
2630void __init init_sched_dl_class(void)
2631{
2632 unsigned int i;
2633
2634 for_each_possible_cpu(i)
2635 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2636 GFP_KERNEL, cpu_to_node(i));
2637}
2638
2639void dl_add_task_root_domain(struct task_struct *p)
2640{
2641 struct rq_flags rf;
2642 struct rq *rq;
2643 struct dl_bw *dl_b;
2644
2645 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2646 if (!dl_task(p)) {
2647 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2648 return;
2649 }
2650
2651 rq = __task_rq_lock(p, &rf);
2652
2653 dl_b = &rq->rd->dl_bw;
2654 raw_spin_lock(&dl_b->lock);
2655
2656 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2657
2658 raw_spin_unlock(&dl_b->lock);
2659
2660 task_rq_unlock(rq, p, &rf);
2661}
2662
2663void dl_clear_root_domain(struct root_domain *rd)
2664{
2665 unsigned long flags;
2666
2667 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2668 rd->dl_bw.total_bw = 0;
2669 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2670}
2671
2672#endif /* CONFIG_SMP */
2673
2674static void switched_from_dl(struct rq *rq, struct task_struct *p)
2675{
2676 /*
2677 * task_non_contending() can start the "inactive timer" (if the 0-lag
2678 * time is in the future). If the task switches back to dl before
2679 * the "inactive timer" fires, it can continue to consume its current
2680 * runtime using its current deadline. If it stays outside of
2681 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2682 * will reset the task parameters.
2683 */
2684 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2685 task_non_contending(&p->dl);
2686
2687 /*
2688 * In case a task is setscheduled out from SCHED_DEADLINE we need to
2689 * keep track of that on its cpuset (for correct bandwidth tracking).
2690 */
2691 dec_dl_tasks_cs(p);
2692
2693 if (!task_on_rq_queued(p)) {
2694 /*
2695 * Inactive timer is armed. However, p is leaving DEADLINE and
2696 * might migrate away from this rq while continuing to run on
2697 * some other class. We need to remove its contribution from
2698 * this rq running_bw now, or sub_rq_bw (below) will complain.
2699 */
2700 if (p->dl.dl_non_contending)
2701 sub_running_bw(&p->dl, &rq->dl);
2702 sub_rq_bw(&p->dl, &rq->dl);
2703 }
2704
2705 /*
2706 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2707 * at the 0-lag time, because the task could have been migrated
2708 * while SCHED_OTHER in the meanwhile.
2709 */
2710 if (p->dl.dl_non_contending)
2711 p->dl.dl_non_contending = 0;
2712
2713 /*
2714 * Since this might be the only -deadline task on the rq,
2715 * this is the right place to try to pull some other one
2716 * from an overloaded CPU, if any.
2717 */
2718 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2719 return;
2720
2721 deadline_queue_pull_task(rq);
2722}
2723
2724/*
2725 * When switching to -deadline, we may overload the rq, then
2726 * we try to push someone off, if possible.
2727 */
2728static void switched_to_dl(struct rq *rq, struct task_struct *p)
2729{
2730 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2731 put_task_struct(p);
2732
2733 /*
2734 * In case a task is setscheduled to SCHED_DEADLINE we need to keep
2735 * track of that on its cpuset (for correct bandwidth tracking).
2736 */
2737 inc_dl_tasks_cs(p);
2738
2739 /* If p is not queued we will update its parameters at next wakeup. */
2740 if (!task_on_rq_queued(p)) {
2741 add_rq_bw(&p->dl, &rq->dl);
2742
2743 return;
2744 }
2745
2746 if (rq->curr != p) {
2747#ifdef CONFIG_SMP
2748 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2749 deadline_queue_push_tasks(rq);
2750#endif
2751 if (dl_task(rq->curr))
2752 wakeup_preempt_dl(rq, p, 0);
2753 else
2754 resched_curr(rq);
2755 } else {
2756 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2757 }
2758}
2759
2760/*
2761 * If the scheduling parameters of a -deadline task changed,
2762 * a push or pull operation might be needed.
2763 */
2764static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2765 int oldprio)
2766{
2767 if (!task_on_rq_queued(p))
2768 return;
2769
2770#ifdef CONFIG_SMP
2771 /*
2772 * This might be too much, but unfortunately
2773 * we don't have the old deadline value, and
2774 * we can't argue if the task is increasing
2775 * or lowering its prio, so...
2776 */
2777 if (!rq->dl.overloaded)
2778 deadline_queue_pull_task(rq);
2779
2780 if (task_current(rq, p)) {
2781 /*
2782 * If we now have a earlier deadline task than p,
2783 * then reschedule, provided p is still on this
2784 * runqueue.
2785 */
2786 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2787 resched_curr(rq);
2788 } else {
2789 /*
2790 * Current may not be deadline in case p was throttled but we
2791 * have just replenished it (e.g. rt_mutex_setprio()).
2792 *
2793 * Otherwise, if p was given an earlier deadline, reschedule.
2794 */
2795 if (!dl_task(rq->curr) ||
2796 dl_time_before(p->dl.deadline, rq->curr->dl.deadline))
2797 resched_curr(rq);
2798 }
2799#else
2800 /*
2801 * We don't know if p has a earlier or later deadline, so let's blindly
2802 * set a (maybe not needed) rescheduling point.
2803 */
2804 resched_curr(rq);
2805#endif
2806}
2807
2808#ifdef CONFIG_SCHED_CORE
2809static int task_is_throttled_dl(struct task_struct *p, int cpu)
2810{
2811 return p->dl.dl_throttled;
2812}
2813#endif
2814
2815DEFINE_SCHED_CLASS(dl) = {
2816
2817 .enqueue_task = enqueue_task_dl,
2818 .dequeue_task = dequeue_task_dl,
2819 .yield_task = yield_task_dl,
2820
2821 .wakeup_preempt = wakeup_preempt_dl,
2822
2823 .pick_next_task = pick_next_task_dl,
2824 .put_prev_task = put_prev_task_dl,
2825 .set_next_task = set_next_task_dl,
2826
2827#ifdef CONFIG_SMP
2828 .balance = balance_dl,
2829 .pick_task = pick_task_dl,
2830 .select_task_rq = select_task_rq_dl,
2831 .migrate_task_rq = migrate_task_rq_dl,
2832 .set_cpus_allowed = set_cpus_allowed_dl,
2833 .rq_online = rq_online_dl,
2834 .rq_offline = rq_offline_dl,
2835 .task_woken = task_woken_dl,
2836 .find_lock_rq = find_lock_later_rq,
2837#endif
2838
2839 .task_tick = task_tick_dl,
2840 .task_fork = task_fork_dl,
2841
2842 .prio_changed = prio_changed_dl,
2843 .switched_from = switched_from_dl,
2844 .switched_to = switched_to_dl,
2845
2846 .update_curr = update_curr_dl,
2847#ifdef CONFIG_SCHED_CORE
2848 .task_is_throttled = task_is_throttled_dl,
2849#endif
2850};
2851
2852/* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2853static u64 dl_generation;
2854
2855int sched_dl_global_validate(void)
2856{
2857 u64 runtime = global_rt_runtime();
2858 u64 period = global_rt_period();
2859 u64 new_bw = to_ratio(period, runtime);
2860 u64 gen = ++dl_generation;
2861 struct dl_bw *dl_b;
2862 int cpu, cpus, ret = 0;
2863 unsigned long flags;
2864
2865 /*
2866 * Here we want to check the bandwidth not being set to some
2867 * value smaller than the currently allocated bandwidth in
2868 * any of the root_domains.
2869 */
2870 for_each_possible_cpu(cpu) {
2871 rcu_read_lock_sched();
2872
2873 if (dl_bw_visited(cpu, gen))
2874 goto next;
2875
2876 dl_b = dl_bw_of(cpu);
2877 cpus = dl_bw_cpus(cpu);
2878
2879 raw_spin_lock_irqsave(&dl_b->lock, flags);
2880 if (new_bw * cpus < dl_b->total_bw)
2881 ret = -EBUSY;
2882 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2883
2884next:
2885 rcu_read_unlock_sched();
2886
2887 if (ret)
2888 break;
2889 }
2890
2891 return ret;
2892}
2893
2894static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2895{
2896 if (global_rt_runtime() == RUNTIME_INF) {
2897 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2898 dl_rq->max_bw = dl_rq->extra_bw = 1 << BW_SHIFT;
2899 } else {
2900 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2901 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2902 dl_rq->max_bw = dl_rq->extra_bw =
2903 to_ratio(global_rt_period(), global_rt_runtime());
2904 }
2905}
2906
2907void sched_dl_do_global(void)
2908{
2909 u64 new_bw = -1;
2910 u64 gen = ++dl_generation;
2911 struct dl_bw *dl_b;
2912 int cpu;
2913 unsigned long flags;
2914
2915 if (global_rt_runtime() != RUNTIME_INF)
2916 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2917
2918 for_each_possible_cpu(cpu) {
2919 rcu_read_lock_sched();
2920
2921 if (dl_bw_visited(cpu, gen)) {
2922 rcu_read_unlock_sched();
2923 continue;
2924 }
2925
2926 dl_b = dl_bw_of(cpu);
2927
2928 raw_spin_lock_irqsave(&dl_b->lock, flags);
2929 dl_b->bw = new_bw;
2930 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2931
2932 rcu_read_unlock_sched();
2933 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2934 }
2935}
2936
2937/*
2938 * We must be sure that accepting a new task (or allowing changing the
2939 * parameters of an existing one) is consistent with the bandwidth
2940 * constraints. If yes, this function also accordingly updates the currently
2941 * allocated bandwidth to reflect the new situation.
2942 *
2943 * This function is called while holding p's rq->lock.
2944 */
2945int sched_dl_overflow(struct task_struct *p, int policy,
2946 const struct sched_attr *attr)
2947{
2948 u64 period = attr->sched_period ?: attr->sched_deadline;
2949 u64 runtime = attr->sched_runtime;
2950 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2951 int cpus, err = -1, cpu = task_cpu(p);
2952 struct dl_bw *dl_b = dl_bw_of(cpu);
2953 unsigned long cap;
2954
2955 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2956 return 0;
2957
2958 /* !deadline task may carry old deadline bandwidth */
2959 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2960 return 0;
2961
2962 /*
2963 * Either if a task, enters, leave, or stays -deadline but changes
2964 * its parameters, we may need to update accordingly the total
2965 * allocated bandwidth of the container.
2966 */
2967 raw_spin_lock(&dl_b->lock);
2968 cpus = dl_bw_cpus(cpu);
2969 cap = dl_bw_capacity(cpu);
2970
2971 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2972 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2973 if (hrtimer_active(&p->dl.inactive_timer))
2974 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2975 __dl_add(dl_b, new_bw, cpus);
2976 err = 0;
2977 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2978 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2979 /*
2980 * XXX this is slightly incorrect: when the task
2981 * utilization decreases, we should delay the total
2982 * utilization change until the task's 0-lag point.
2983 * But this would require to set the task's "inactive
2984 * timer" when the task is not inactive.
2985 */
2986 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2987 __dl_add(dl_b, new_bw, cpus);
2988 dl_change_utilization(p, new_bw);
2989 err = 0;
2990 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2991 /*
2992 * Do not decrease the total deadline utilization here,
2993 * switched_from_dl() will take care to do it at the correct
2994 * (0-lag) time.
2995 */
2996 err = 0;
2997 }
2998 raw_spin_unlock(&dl_b->lock);
2999
3000 return err;
3001}
3002
3003/*
3004 * This function initializes the sched_dl_entity of a newly becoming
3005 * SCHED_DEADLINE task.
3006 *
3007 * Only the static values are considered here, the actual runtime and the
3008 * absolute deadline will be properly calculated when the task is enqueued
3009 * for the first time with its new policy.
3010 */
3011void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3012{
3013 struct sched_dl_entity *dl_se = &p->dl;
3014
3015 dl_se->dl_runtime = attr->sched_runtime;
3016 dl_se->dl_deadline = attr->sched_deadline;
3017 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3018 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
3019 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3020 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3021}
3022
3023void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3024{
3025 struct sched_dl_entity *dl_se = &p->dl;
3026
3027 attr->sched_priority = p->rt_priority;
3028 attr->sched_runtime = dl_se->dl_runtime;
3029 attr->sched_deadline = dl_se->dl_deadline;
3030 attr->sched_period = dl_se->dl_period;
3031 attr->sched_flags &= ~SCHED_DL_FLAGS;
3032 attr->sched_flags |= dl_se->flags;
3033}
3034
3035/*
3036 * This function validates the new parameters of a -deadline task.
3037 * We ask for the deadline not being zero, and greater or equal
3038 * than the runtime, as well as the period of being zero or
3039 * greater than deadline. Furthermore, we have to be sure that
3040 * user parameters are above the internal resolution of 1us (we
3041 * check sched_runtime only since it is always the smaller one) and
3042 * below 2^63 ns (we have to check both sched_deadline and
3043 * sched_period, as the latter can be zero).
3044 */
3045bool __checkparam_dl(const struct sched_attr *attr)
3046{
3047 u64 period, max, min;
3048
3049 /* special dl tasks don't actually use any parameter */
3050 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3051 return true;
3052
3053 /* deadline != 0 */
3054 if (attr->sched_deadline == 0)
3055 return false;
3056
3057 /*
3058 * Since we truncate DL_SCALE bits, make sure we're at least
3059 * that big.
3060 */
3061 if (attr->sched_runtime < (1ULL << DL_SCALE))
3062 return false;
3063
3064 /*
3065 * Since we use the MSB for wrap-around and sign issues, make
3066 * sure it's not set (mind that period can be equal to zero).
3067 */
3068 if (attr->sched_deadline & (1ULL << 63) ||
3069 attr->sched_period & (1ULL << 63))
3070 return false;
3071
3072 period = attr->sched_period;
3073 if (!period)
3074 period = attr->sched_deadline;
3075
3076 /* runtime <= deadline <= period (if period != 0) */
3077 if (period < attr->sched_deadline ||
3078 attr->sched_deadline < attr->sched_runtime)
3079 return false;
3080
3081 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
3082 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
3083
3084 if (period < min || period > max)
3085 return false;
3086
3087 return true;
3088}
3089
3090/*
3091 * This function clears the sched_dl_entity static params.
3092 */
3093static void __dl_clear_params(struct sched_dl_entity *dl_se)
3094{
3095 dl_se->dl_runtime = 0;
3096 dl_se->dl_deadline = 0;
3097 dl_se->dl_period = 0;
3098 dl_se->flags = 0;
3099 dl_se->dl_bw = 0;
3100 dl_se->dl_density = 0;
3101
3102 dl_se->dl_throttled = 0;
3103 dl_se->dl_yielded = 0;
3104 dl_se->dl_non_contending = 0;
3105 dl_se->dl_overrun = 0;
3106 dl_se->dl_server = 0;
3107
3108#ifdef CONFIG_RT_MUTEXES
3109 dl_se->pi_se = dl_se;
3110#endif
3111}
3112
3113void init_dl_entity(struct sched_dl_entity *dl_se)
3114{
3115 RB_CLEAR_NODE(&dl_se->rb_node);
3116 init_dl_task_timer(dl_se);
3117 init_dl_inactive_task_timer(dl_se);
3118 __dl_clear_params(dl_se);
3119}
3120
3121bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
3122{
3123 struct sched_dl_entity *dl_se = &p->dl;
3124
3125 if (dl_se->dl_runtime != attr->sched_runtime ||
3126 dl_se->dl_deadline != attr->sched_deadline ||
3127 dl_se->dl_period != attr->sched_period ||
3128 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3129 return true;
3130
3131 return false;
3132}
3133
3134#ifdef CONFIG_SMP
3135int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3136 const struct cpumask *trial)
3137{
3138 unsigned long flags, cap;
3139 struct dl_bw *cur_dl_b;
3140 int ret = 1;
3141
3142 rcu_read_lock_sched();
3143 cur_dl_b = dl_bw_of(cpumask_any(cur));
3144 cap = __dl_bw_capacity(trial);
3145 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3146 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3147 ret = 0;
3148 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3149 rcu_read_unlock_sched();
3150
3151 return ret;
3152}
3153
3154enum dl_bw_request {
3155 dl_bw_req_check_overflow = 0,
3156 dl_bw_req_alloc,
3157 dl_bw_req_free
3158};
3159
3160static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
3161{
3162 unsigned long flags;
3163 struct dl_bw *dl_b;
3164 bool overflow = 0;
3165
3166 rcu_read_lock_sched();
3167 dl_b = dl_bw_of(cpu);
3168 raw_spin_lock_irqsave(&dl_b->lock, flags);
3169
3170 if (req == dl_bw_req_free) {
3171 __dl_sub(dl_b, dl_bw, dl_bw_cpus(cpu));
3172 } else {
3173 unsigned long cap = dl_bw_capacity(cpu);
3174
3175 overflow = __dl_overflow(dl_b, cap, 0, dl_bw);
3176
3177 if (req == dl_bw_req_alloc && !overflow) {
3178 /*
3179 * We reserve space in the destination
3180 * root_domain, as we can't fail after this point.
3181 * We will free resources in the source root_domain
3182 * later on (see set_cpus_allowed_dl()).
3183 */
3184 __dl_add(dl_b, dl_bw, dl_bw_cpus(cpu));
3185 }
3186 }
3187
3188 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3189 rcu_read_unlock_sched();
3190
3191 return overflow ? -EBUSY : 0;
3192}
3193
3194int dl_bw_check_overflow(int cpu)
3195{
3196 return dl_bw_manage(dl_bw_req_check_overflow, cpu, 0);
3197}
3198
3199int dl_bw_alloc(int cpu, u64 dl_bw)
3200{
3201 return dl_bw_manage(dl_bw_req_alloc, cpu, dl_bw);
3202}
3203
3204void dl_bw_free(int cpu, u64 dl_bw)
3205{
3206 dl_bw_manage(dl_bw_req_free, cpu, dl_bw);
3207}
3208#endif
3209
3210#ifdef CONFIG_SCHED_DEBUG
3211void print_dl_stats(struct seq_file *m, int cpu)
3212{
3213 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3214}
3215#endif /* CONFIG_SCHED_DEBUG */