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#include "pelt.h"
20
21struct dl_bandwidth def_dl_bandwidth;
22
23static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
24{
25 return container_of(dl_se, struct task_struct, dl);
26}
27
28static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
29{
30 return container_of(dl_rq, struct rq, dl);
31}
32
33static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
34{
35 struct task_struct *p = dl_task_of(dl_se);
36 struct rq *rq = task_rq(p);
37
38 return &rq->dl;
39}
40
41static inline int on_dl_rq(struct sched_dl_entity *dl_se)
42{
43 return !RB_EMPTY_NODE(&dl_se->rb_node);
44}
45
46#ifdef CONFIG_SMP
47static inline struct dl_bw *dl_bw_of(int i)
48{
49 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
50 "sched RCU must be held");
51 return &cpu_rq(i)->rd->dl_bw;
52}
53
54static inline int dl_bw_cpus(int i)
55{
56 struct root_domain *rd = cpu_rq(i)->rd;
57 int cpus = 0;
58
59 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
60 "sched RCU must be held");
61 for_each_cpu_and(i, rd->span, cpu_active_mask)
62 cpus++;
63
64 return cpus;
65}
66#else
67static inline struct dl_bw *dl_bw_of(int i)
68{
69 return &cpu_rq(i)->dl.dl_bw;
70}
71
72static inline int dl_bw_cpus(int i)
73{
74 return 1;
75}
76#endif
77
78static inline
79void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
80{
81 u64 old = dl_rq->running_bw;
82
83 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
84 dl_rq->running_bw += dl_bw;
85 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
86 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
87 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
88 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
89}
90
91static inline
92void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
93{
94 u64 old = dl_rq->running_bw;
95
96 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
97 dl_rq->running_bw -= dl_bw;
98 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
99 if (dl_rq->running_bw > old)
100 dl_rq->running_bw = 0;
101 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
102 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
103}
104
105static inline
106void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
107{
108 u64 old = dl_rq->this_bw;
109
110 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
111 dl_rq->this_bw += dl_bw;
112 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
113}
114
115static inline
116void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
117{
118 u64 old = dl_rq->this_bw;
119
120 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
121 dl_rq->this_bw -= dl_bw;
122 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
123 if (dl_rq->this_bw > old)
124 dl_rq->this_bw = 0;
125 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
126}
127
128static inline
129void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
130{
131 if (!dl_entity_is_special(dl_se))
132 __add_rq_bw(dl_se->dl_bw, dl_rq);
133}
134
135static inline
136void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
137{
138 if (!dl_entity_is_special(dl_se))
139 __sub_rq_bw(dl_se->dl_bw, dl_rq);
140}
141
142static inline
143void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
144{
145 if (!dl_entity_is_special(dl_se))
146 __add_running_bw(dl_se->dl_bw, dl_rq);
147}
148
149static inline
150void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
151{
152 if (!dl_entity_is_special(dl_se))
153 __sub_running_bw(dl_se->dl_bw, dl_rq);
154}
155
156void dl_change_utilization(struct task_struct *p, u64 new_bw)
157{
158 struct rq *rq;
159
160 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
161
162 if (task_on_rq_queued(p))
163 return;
164
165 rq = task_rq(p);
166 if (p->dl.dl_non_contending) {
167 sub_running_bw(&p->dl, &rq->dl);
168 p->dl.dl_non_contending = 0;
169 /*
170 * If the timer handler is currently running and the
171 * timer cannot be cancelled, inactive_task_timer()
172 * will see that dl_not_contending is not set, and
173 * will not touch the rq's active utilization,
174 * so we are still safe.
175 */
176 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
177 put_task_struct(p);
178 }
179 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
180 __add_rq_bw(new_bw, &rq->dl);
181}
182
183/*
184 * The utilization of a task cannot be immediately removed from
185 * the rq active utilization (running_bw) when the task blocks.
186 * Instead, we have to wait for the so called "0-lag time".
187 *
188 * If a task blocks before the "0-lag time", a timer (the inactive
189 * timer) is armed, and running_bw is decreased when the timer
190 * fires.
191 *
192 * If the task wakes up again before the inactive timer fires,
193 * the timer is cancelled, whereas if the task wakes up after the
194 * inactive timer fired (and running_bw has been decreased) the
195 * task's utilization has to be added to running_bw again.
196 * A flag in the deadline scheduling entity (dl_non_contending)
197 * is used to avoid race conditions between the inactive timer handler
198 * and task wakeups.
199 *
200 * The following diagram shows how running_bw is updated. A task is
201 * "ACTIVE" when its utilization contributes to running_bw; an
202 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
203 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
204 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
205 * time already passed, which does not contribute to running_bw anymore.
206 * +------------------+
207 * wakeup | ACTIVE |
208 * +------------------>+ contending |
209 * | add_running_bw | |
210 * | +----+------+------+
211 * | | ^
212 * | dequeue | |
213 * +--------+-------+ | |
214 * | | t >= 0-lag | | wakeup
215 * | INACTIVE |<---------------+ |
216 * | | sub_running_bw | |
217 * +--------+-------+ | |
218 * ^ | |
219 * | t < 0-lag | |
220 * | | |
221 * | V |
222 * | +----+------+------+
223 * | sub_running_bw | ACTIVE |
224 * +-------------------+ |
225 * inactive timer | non contending |
226 * fired +------------------+
227 *
228 * The task_non_contending() function is invoked when a task
229 * blocks, and checks if the 0-lag time already passed or
230 * not (in the first case, it directly updates running_bw;
231 * in the second case, it arms the inactive timer).
232 *
233 * The task_contending() function is invoked when a task wakes
234 * up, and checks if the task is still in the "ACTIVE non contending"
235 * state or not (in the second case, it updates running_bw).
236 */
237static void task_non_contending(struct task_struct *p)
238{
239 struct sched_dl_entity *dl_se = &p->dl;
240 struct hrtimer *timer = &dl_se->inactive_timer;
241 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
242 struct rq *rq = rq_of_dl_rq(dl_rq);
243 s64 zerolag_time;
244
245 /*
246 * If this is a non-deadline task that has been boosted,
247 * do nothing
248 */
249 if (dl_se->dl_runtime == 0)
250 return;
251
252 if (dl_entity_is_special(dl_se))
253 return;
254
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) || hrtimer_active(&dl_se->inactive_timer)) {
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_HARD);
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 struct dl_bw *dl_b;
533
534 later_rq = find_lock_later_rq(p, rq);
535 if (!later_rq) {
536 int cpu;
537
538 /*
539 * If we cannot preempt any rq, fall back to pick any
540 * online CPU:
541 */
542 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
543 if (cpu >= nr_cpu_ids) {
544 /*
545 * Failed to find any suitable CPU.
546 * The task will never come back!
547 */
548 BUG_ON(dl_bandwidth_enabled());
549
550 /*
551 * If admission control is disabled we
552 * try a little harder to let the task
553 * run.
554 */
555 cpu = cpumask_any(cpu_active_mask);
556 }
557 later_rq = cpu_rq(cpu);
558 double_lock_balance(rq, later_rq);
559 }
560
561 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
562 /*
563 * Inactive timer is armed (or callback is running, but
564 * waiting for us to release rq locks). In any case, when it
565 * will fire (or continue), it will see running_bw of this
566 * task migrated to later_rq (and correctly handle it).
567 */
568 sub_running_bw(&p->dl, &rq->dl);
569 sub_rq_bw(&p->dl, &rq->dl);
570
571 add_rq_bw(&p->dl, &later_rq->dl);
572 add_running_bw(&p->dl, &later_rq->dl);
573 } else {
574 sub_rq_bw(&p->dl, &rq->dl);
575 add_rq_bw(&p->dl, &later_rq->dl);
576 }
577
578 /*
579 * And we finally need to fixup root_domain(s) bandwidth accounting,
580 * since p is still hanging out in the old (now moved to default) root
581 * domain.
582 */
583 dl_b = &rq->rd->dl_bw;
584 raw_spin_lock(&dl_b->lock);
585 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
586 raw_spin_unlock(&dl_b->lock);
587
588 dl_b = &later_rq->rd->dl_bw;
589 raw_spin_lock(&dl_b->lock);
590 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
591 raw_spin_unlock(&dl_b->lock);
592
593 set_task_cpu(p, later_rq->cpu);
594 double_unlock_balance(later_rq, rq);
595
596 return later_rq;
597}
598
599#else
600
601static inline
602void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
603{
604}
605
606static inline
607void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
608{
609}
610
611static inline
612void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
613{
614}
615
616static inline
617void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
618{
619}
620
621static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
622{
623 return false;
624}
625
626static inline void pull_dl_task(struct rq *rq)
627{
628}
629
630static inline void deadline_queue_push_tasks(struct rq *rq)
631{
632}
633
634static inline void deadline_queue_pull_task(struct rq *rq)
635{
636}
637#endif /* CONFIG_SMP */
638
639static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
640static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
641static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
642
643/*
644 * We are being explicitly informed that a new instance is starting,
645 * and this means that:
646 * - the absolute deadline of the entity has to be placed at
647 * current time + relative deadline;
648 * - the runtime of the entity has to be set to the maximum value.
649 *
650 * The capability of specifying such event is useful whenever a -deadline
651 * entity wants to (try to!) synchronize its behaviour with the scheduler's
652 * one, and to (try to!) reconcile itself with its own scheduling
653 * parameters.
654 */
655static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
656{
657 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
658 struct rq *rq = rq_of_dl_rq(dl_rq);
659
660 WARN_ON(dl_se->dl_boosted);
661 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
662
663 /*
664 * We are racing with the deadline timer. So, do nothing because
665 * the deadline timer handler will take care of properly recharging
666 * the runtime and postponing the deadline
667 */
668 if (dl_se->dl_throttled)
669 return;
670
671 /*
672 * We use the regular wall clock time to set deadlines in the
673 * future; in fact, we must consider execution overheads (time
674 * spent on hardirq context, etc.).
675 */
676 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
677 dl_se->runtime = dl_se->dl_runtime;
678}
679
680/*
681 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
682 * possibility of a entity lasting more than what it declared, and thus
683 * exhausting its runtime.
684 *
685 * Here we are interested in making runtime overrun possible, but we do
686 * not want a entity which is misbehaving to affect the scheduling of all
687 * other entities.
688 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
689 * is used, in order to confine each entity within its own bandwidth.
690 *
691 * This function deals exactly with that, and ensures that when the runtime
692 * of a entity is replenished, its deadline is also postponed. That ensures
693 * the overrunning entity can't interfere with other entity in the system and
694 * can't make them miss their deadlines. Reasons why this kind of overruns
695 * could happen are, typically, a entity voluntarily trying to overcome its
696 * runtime, or it just underestimated it during sched_setattr().
697 */
698static void replenish_dl_entity(struct sched_dl_entity *dl_se,
699 struct sched_dl_entity *pi_se)
700{
701 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
702 struct rq *rq = rq_of_dl_rq(dl_rq);
703
704 BUG_ON(pi_se->dl_runtime <= 0);
705
706 /*
707 * This could be the case for a !-dl task that is boosted.
708 * Just go with full inherited parameters.
709 */
710 if (dl_se->dl_deadline == 0) {
711 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
712 dl_se->runtime = pi_se->dl_runtime;
713 }
714
715 if (dl_se->dl_yielded && dl_se->runtime > 0)
716 dl_se->runtime = 0;
717
718 /*
719 * We keep moving the deadline away until we get some
720 * available runtime for the entity. This ensures correct
721 * handling of situations where the runtime overrun is
722 * arbitrary large.
723 */
724 while (dl_se->runtime <= 0) {
725 dl_se->deadline += pi_se->dl_period;
726 dl_se->runtime += pi_se->dl_runtime;
727 }
728
729 /*
730 * At this point, the deadline really should be "in
731 * the future" with respect to rq->clock. If it's
732 * not, we are, for some reason, lagging too much!
733 * Anyway, after having warn userspace abut that,
734 * we still try to keep the things running by
735 * resetting the deadline and the budget of the
736 * entity.
737 */
738 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
739 printk_deferred_once("sched: DL replenish lagged too much\n");
740 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
741 dl_se->runtime = pi_se->dl_runtime;
742 }
743
744 if (dl_se->dl_yielded)
745 dl_se->dl_yielded = 0;
746 if (dl_se->dl_throttled)
747 dl_se->dl_throttled = 0;
748}
749
750/*
751 * Here we check if --at time t-- an entity (which is probably being
752 * [re]activated or, in general, enqueued) can use its remaining runtime
753 * and its current deadline _without_ exceeding the bandwidth it is
754 * assigned (function returns true if it can't). We are in fact applying
755 * one of the CBS rules: when a task wakes up, if the residual runtime
756 * over residual deadline fits within the allocated bandwidth, then we
757 * can keep the current (absolute) deadline and residual budget without
758 * disrupting the schedulability of the system. Otherwise, we should
759 * refill the runtime and set the deadline a period in the future,
760 * because keeping the current (absolute) deadline of the task would
761 * result in breaking guarantees promised to other tasks (refer to
762 * Documentation/scheduler/sched-deadline.rst for more information).
763 *
764 * This function returns true if:
765 *
766 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
767 *
768 * IOW we can't recycle current parameters.
769 *
770 * Notice that the bandwidth check is done against the deadline. For
771 * task with deadline equal to period this is the same of using
772 * dl_period instead of dl_deadline in the equation above.
773 */
774static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
775 struct sched_dl_entity *pi_se, u64 t)
776{
777 u64 left, right;
778
779 /*
780 * left and right are the two sides of the equation above,
781 * after a bit of shuffling to use multiplications instead
782 * of divisions.
783 *
784 * Note that none of the time values involved in the two
785 * multiplications are absolute: dl_deadline and dl_runtime
786 * are the relative deadline and the maximum runtime of each
787 * instance, runtime is the runtime left for the last instance
788 * and (deadline - t), since t is rq->clock, is the time left
789 * to the (absolute) deadline. Even if overflowing the u64 type
790 * is very unlikely to occur in both cases, here we scale down
791 * as we want to avoid that risk at all. Scaling down by 10
792 * means that we reduce granularity to 1us. We are fine with it,
793 * since this is only a true/false check and, anyway, thinking
794 * of anything below microseconds resolution is actually fiction
795 * (but still we want to give the user that illusion >;).
796 */
797 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
798 right = ((dl_se->deadline - t) >> DL_SCALE) *
799 (pi_se->dl_runtime >> DL_SCALE);
800
801 return dl_time_before(right, left);
802}
803
804/*
805 * Revised wakeup rule [1]: For self-suspending tasks, rather then
806 * re-initializing task's runtime and deadline, the revised wakeup
807 * rule adjusts the task's runtime to avoid the task to overrun its
808 * density.
809 *
810 * Reasoning: a task may overrun the density if:
811 * runtime / (deadline - t) > dl_runtime / dl_deadline
812 *
813 * Therefore, runtime can be adjusted to:
814 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
815 *
816 * In such way that runtime will be equal to the maximum density
817 * the task can use without breaking any rule.
818 *
819 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
820 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
821 */
822static void
823update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
824{
825 u64 laxity = dl_se->deadline - rq_clock(rq);
826
827 /*
828 * If the task has deadline < period, and the deadline is in the past,
829 * it should already be throttled before this check.
830 *
831 * See update_dl_entity() comments for further details.
832 */
833 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
834
835 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
836}
837
838/*
839 * Regarding the deadline, a task with implicit deadline has a relative
840 * deadline == relative period. A task with constrained deadline has a
841 * relative deadline <= relative period.
842 *
843 * We support constrained deadline tasks. However, there are some restrictions
844 * applied only for tasks which do not have an implicit deadline. See
845 * update_dl_entity() to know more about such restrictions.
846 *
847 * The dl_is_implicit() returns true if the task has an implicit deadline.
848 */
849static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
850{
851 return dl_se->dl_deadline == dl_se->dl_period;
852}
853
854/*
855 * When a deadline entity is placed in the runqueue, its runtime and deadline
856 * might need to be updated. This is done by a CBS wake up rule. There are two
857 * different rules: 1) the original CBS; and 2) the Revisited CBS.
858 *
859 * When the task is starting a new period, the Original CBS is used. In this
860 * case, the runtime is replenished and a new absolute deadline is set.
861 *
862 * When a task is queued before the begin of the next period, using the
863 * remaining runtime and deadline could make the entity to overflow, see
864 * dl_entity_overflow() to find more about runtime overflow. When such case
865 * is detected, the runtime and deadline need to be updated.
866 *
867 * If the task has an implicit deadline, i.e., deadline == period, the Original
868 * CBS is applied. the runtime is replenished and a new absolute deadline is
869 * set, as in the previous cases.
870 *
871 * However, the Original CBS does not work properly for tasks with
872 * deadline < period, which are said to have a constrained deadline. By
873 * applying the Original CBS, a constrained deadline task would be able to run
874 * runtime/deadline in a period. With deadline < period, the task would
875 * overrun the runtime/period allowed bandwidth, breaking the admission test.
876 *
877 * In order to prevent this misbehave, the Revisited CBS is used for
878 * constrained deadline tasks when a runtime overflow is detected. In the
879 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
880 * the remaining runtime of the task is reduced to avoid runtime overflow.
881 * Please refer to the comments update_dl_revised_wakeup() function to find
882 * more about the Revised CBS rule.
883 */
884static void update_dl_entity(struct sched_dl_entity *dl_se,
885 struct sched_dl_entity *pi_se)
886{
887 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
888 struct rq *rq = rq_of_dl_rq(dl_rq);
889
890 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
891 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
892
893 if (unlikely(!dl_is_implicit(dl_se) &&
894 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
895 !dl_se->dl_boosted)){
896 update_dl_revised_wakeup(dl_se, rq);
897 return;
898 }
899
900 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
901 dl_se->runtime = pi_se->dl_runtime;
902 }
903}
904
905static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
906{
907 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
908}
909
910/*
911 * If the entity depleted all its runtime, and if we want it to sleep
912 * while waiting for some new execution time to become available, we
913 * set the bandwidth replenishment timer to the replenishment instant
914 * and try to activate it.
915 *
916 * Notice that it is important for the caller to know if the timer
917 * actually started or not (i.e., the replenishment instant is in
918 * the future or in the past).
919 */
920static int start_dl_timer(struct task_struct *p)
921{
922 struct sched_dl_entity *dl_se = &p->dl;
923 struct hrtimer *timer = &dl_se->dl_timer;
924 struct rq *rq = task_rq(p);
925 ktime_t now, act;
926 s64 delta;
927
928 lockdep_assert_held(&rq->lock);
929
930 /*
931 * We want the timer to fire at the deadline, but considering
932 * that it is actually coming from rq->clock and not from
933 * hrtimer's time base reading.
934 */
935 act = ns_to_ktime(dl_next_period(dl_se));
936 now = hrtimer_cb_get_time(timer);
937 delta = ktime_to_ns(now) - rq_clock(rq);
938 act = ktime_add_ns(act, delta);
939
940 /*
941 * If the expiry time already passed, e.g., because the value
942 * chosen as the deadline is too small, don't even try to
943 * start the timer in the past!
944 */
945 if (ktime_us_delta(act, now) < 0)
946 return 0;
947
948 /*
949 * !enqueued will guarantee another callback; even if one is already in
950 * progress. This ensures a balanced {get,put}_task_struct().
951 *
952 * The race against __run_timer() clearing the enqueued state is
953 * harmless because we're holding task_rq()->lock, therefore the timer
954 * expiring after we've done the check will wait on its task_rq_lock()
955 * and observe our state.
956 */
957 if (!hrtimer_is_queued(timer)) {
958 get_task_struct(p);
959 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
960 }
961
962 return 1;
963}
964
965/*
966 * This is the bandwidth enforcement timer callback. If here, we know
967 * a task is not on its dl_rq, since the fact that the timer was running
968 * means the task is throttled and needs a runtime replenishment.
969 *
970 * However, what we actually do depends on the fact the task is active,
971 * (it is on its rq) or has been removed from there by a call to
972 * dequeue_task_dl(). In the former case we must issue the runtime
973 * replenishment and add the task back to the dl_rq; in the latter, we just
974 * do nothing but clearing dl_throttled, so that runtime and deadline
975 * updating (and the queueing back to dl_rq) will be done by the
976 * next call to enqueue_task_dl().
977 */
978static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
979{
980 struct sched_dl_entity *dl_se = container_of(timer,
981 struct sched_dl_entity,
982 dl_timer);
983 struct task_struct *p = dl_task_of(dl_se);
984 struct rq_flags rf;
985 struct rq *rq;
986
987 rq = task_rq_lock(p, &rf);
988
989 /*
990 * The task might have changed its scheduling policy to something
991 * different than SCHED_DEADLINE (through switched_from_dl()).
992 */
993 if (!dl_task(p))
994 goto unlock;
995
996 /*
997 * The task might have been boosted by someone else and might be in the
998 * boosting/deboosting path, its not throttled.
999 */
1000 if (dl_se->dl_boosted)
1001 goto unlock;
1002
1003 /*
1004 * Spurious timer due to start_dl_timer() race; or we already received
1005 * a replenishment from rt_mutex_setprio().
1006 */
1007 if (!dl_se->dl_throttled)
1008 goto unlock;
1009
1010 sched_clock_tick();
1011 update_rq_clock(rq);
1012
1013 /*
1014 * If the throttle happened during sched-out; like:
1015 *
1016 * schedule()
1017 * deactivate_task()
1018 * dequeue_task_dl()
1019 * update_curr_dl()
1020 * start_dl_timer()
1021 * __dequeue_task_dl()
1022 * prev->on_rq = 0;
1023 *
1024 * We can be both throttled and !queued. Replenish the counter
1025 * but do not enqueue -- wait for our wakeup to do that.
1026 */
1027 if (!task_on_rq_queued(p)) {
1028 replenish_dl_entity(dl_se, dl_se);
1029 goto unlock;
1030 }
1031
1032#ifdef CONFIG_SMP
1033 if (unlikely(!rq->online)) {
1034 /*
1035 * If the runqueue is no longer available, migrate the
1036 * task elsewhere. This necessarily changes rq.
1037 */
1038 lockdep_unpin_lock(&rq->lock, rf.cookie);
1039 rq = dl_task_offline_migration(rq, p);
1040 rf.cookie = lockdep_pin_lock(&rq->lock);
1041 update_rq_clock(rq);
1042
1043 /*
1044 * Now that the task has been migrated to the new RQ and we
1045 * have that locked, proceed as normal and enqueue the task
1046 * there.
1047 */
1048 }
1049#endif
1050
1051 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1052 if (dl_task(rq->curr))
1053 check_preempt_curr_dl(rq, p, 0);
1054 else
1055 resched_curr(rq);
1056
1057#ifdef CONFIG_SMP
1058 /*
1059 * Queueing this task back might have overloaded rq, check if we need
1060 * to kick someone away.
1061 */
1062 if (has_pushable_dl_tasks(rq)) {
1063 /*
1064 * Nothing relies on rq->lock after this, so its safe to drop
1065 * rq->lock.
1066 */
1067 rq_unpin_lock(rq, &rf);
1068 push_dl_task(rq);
1069 rq_repin_lock(rq, &rf);
1070 }
1071#endif
1072
1073unlock:
1074 task_rq_unlock(rq, p, &rf);
1075
1076 /*
1077 * This can free the task_struct, including this hrtimer, do not touch
1078 * anything related to that after this.
1079 */
1080 put_task_struct(p);
1081
1082 return HRTIMER_NORESTART;
1083}
1084
1085void init_dl_task_timer(struct sched_dl_entity *dl_se)
1086{
1087 struct hrtimer *timer = &dl_se->dl_timer;
1088
1089 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1090 timer->function = dl_task_timer;
1091}
1092
1093/*
1094 * During the activation, CBS checks if it can reuse the current task's
1095 * runtime and period. If the deadline of the task is in the past, CBS
1096 * cannot use the runtime, and so it replenishes the task. This rule
1097 * works fine for implicit deadline tasks (deadline == period), and the
1098 * CBS was designed for implicit deadline tasks. However, a task with
1099 * constrained deadline (deadine < period) might be awakened after the
1100 * deadline, but before the next period. In this case, replenishing the
1101 * task would allow it to run for runtime / deadline. As in this case
1102 * deadline < period, CBS enables a task to run for more than the
1103 * runtime / period. In a very loaded system, this can cause a domino
1104 * effect, making other tasks miss their deadlines.
1105 *
1106 * To avoid this problem, in the activation of a constrained deadline
1107 * task after the deadline but before the next period, throttle the
1108 * task and set the replenishing timer to the begin of the next period,
1109 * unless it is boosted.
1110 */
1111static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1112{
1113 struct task_struct *p = dl_task_of(dl_se);
1114 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1115
1116 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1117 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1118 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1119 return;
1120 dl_se->dl_throttled = 1;
1121 if (dl_se->runtime > 0)
1122 dl_se->runtime = 0;
1123 }
1124}
1125
1126static
1127int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1128{
1129 return (dl_se->runtime <= 0);
1130}
1131
1132extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1133
1134/*
1135 * This function implements the GRUB accounting rule:
1136 * according to the GRUB reclaiming algorithm, the runtime is
1137 * not decreased as "dq = -dt", but as
1138 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1139 * where u is the utilization of the task, Umax is the maximum reclaimable
1140 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1141 * as the difference between the "total runqueue utilization" and the
1142 * runqueue active utilization, and Uextra is the (per runqueue) extra
1143 * reclaimable utilization.
1144 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1145 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1146 * BW_SHIFT.
1147 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1148 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1149 * Since delta is a 64 bit variable, to have an overflow its value
1150 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1151 * So, overflow is not an issue here.
1152 */
1153static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1154{
1155 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1156 u64 u_act;
1157 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1158
1159 /*
1160 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1161 * we compare u_inact + rq->dl.extra_bw with
1162 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1163 * u_inact + rq->dl.extra_bw can be larger than
1164 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1165 * leading to wrong results)
1166 */
1167 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1168 u_act = u_act_min;
1169 else
1170 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1171
1172 return (delta * u_act) >> BW_SHIFT;
1173}
1174
1175/*
1176 * Update the current task's runtime statistics (provided it is still
1177 * a -deadline task and has not been removed from the dl_rq).
1178 */
1179static void update_curr_dl(struct rq *rq)
1180{
1181 struct task_struct *curr = rq->curr;
1182 struct sched_dl_entity *dl_se = &curr->dl;
1183 u64 delta_exec, scaled_delta_exec;
1184 int cpu = cpu_of(rq);
1185 u64 now;
1186
1187 if (!dl_task(curr) || !on_dl_rq(dl_se))
1188 return;
1189
1190 /*
1191 * Consumed budget is computed considering the time as
1192 * observed by schedulable tasks (excluding time spent
1193 * in hardirq context, etc.). Deadlines are instead
1194 * computed using hard walltime. This seems to be the more
1195 * natural solution, but the full ramifications of this
1196 * approach need further study.
1197 */
1198 now = rq_clock_task(rq);
1199 delta_exec = now - curr->se.exec_start;
1200 if (unlikely((s64)delta_exec <= 0)) {
1201 if (unlikely(dl_se->dl_yielded))
1202 goto throttle;
1203 return;
1204 }
1205
1206 schedstat_set(curr->se.statistics.exec_max,
1207 max(curr->se.statistics.exec_max, delta_exec));
1208
1209 curr->se.sum_exec_runtime += delta_exec;
1210 account_group_exec_runtime(curr, delta_exec);
1211
1212 curr->se.exec_start = now;
1213 cgroup_account_cputime(curr, delta_exec);
1214
1215 if (dl_entity_is_special(dl_se))
1216 return;
1217
1218 /*
1219 * For tasks that participate in GRUB, we implement GRUB-PA: the
1220 * spare reclaimed bandwidth is used to clock down frequency.
1221 *
1222 * For the others, we still need to scale reservation parameters
1223 * according to current frequency and CPU maximum capacity.
1224 */
1225 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1226 scaled_delta_exec = grub_reclaim(delta_exec,
1227 rq,
1228 &curr->dl);
1229 } else {
1230 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1231 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1232
1233 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1234 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1235 }
1236
1237 dl_se->runtime -= scaled_delta_exec;
1238
1239throttle:
1240 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1241 dl_se->dl_throttled = 1;
1242
1243 /* If requested, inform the user about runtime overruns. */
1244 if (dl_runtime_exceeded(dl_se) &&
1245 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1246 dl_se->dl_overrun = 1;
1247
1248 __dequeue_task_dl(rq, curr, 0);
1249 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
1250 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1251
1252 if (!is_leftmost(curr, &rq->dl))
1253 resched_curr(rq);
1254 }
1255
1256 /*
1257 * Because -- for now -- we share the rt bandwidth, we need to
1258 * account our runtime there too, otherwise actual rt tasks
1259 * would be able to exceed the shared quota.
1260 *
1261 * Account to the root rt group for now.
1262 *
1263 * The solution we're working towards is having the RT groups scheduled
1264 * using deadline servers -- however there's a few nasties to figure
1265 * out before that can happen.
1266 */
1267 if (rt_bandwidth_enabled()) {
1268 struct rt_rq *rt_rq = &rq->rt;
1269
1270 raw_spin_lock(&rt_rq->rt_runtime_lock);
1271 /*
1272 * We'll let actual RT tasks worry about the overflow here, we
1273 * have our own CBS to keep us inline; only account when RT
1274 * bandwidth is relevant.
1275 */
1276 if (sched_rt_bandwidth_account(rt_rq))
1277 rt_rq->rt_time += delta_exec;
1278 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1279 }
1280}
1281
1282static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1283{
1284 struct sched_dl_entity *dl_se = container_of(timer,
1285 struct sched_dl_entity,
1286 inactive_timer);
1287 struct task_struct *p = dl_task_of(dl_se);
1288 struct rq_flags rf;
1289 struct rq *rq;
1290
1291 rq = task_rq_lock(p, &rf);
1292
1293 sched_clock_tick();
1294 update_rq_clock(rq);
1295
1296 if (!dl_task(p) || p->state == TASK_DEAD) {
1297 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1298
1299 if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1300 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1301 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1302 dl_se->dl_non_contending = 0;
1303 }
1304
1305 raw_spin_lock(&dl_b->lock);
1306 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1307 raw_spin_unlock(&dl_b->lock);
1308 __dl_clear_params(p);
1309
1310 goto unlock;
1311 }
1312 if (dl_se->dl_non_contending == 0)
1313 goto unlock;
1314
1315 sub_running_bw(dl_se, &rq->dl);
1316 dl_se->dl_non_contending = 0;
1317unlock:
1318 task_rq_unlock(rq, p, &rf);
1319 put_task_struct(p);
1320
1321 return HRTIMER_NORESTART;
1322}
1323
1324void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1325{
1326 struct hrtimer *timer = &dl_se->inactive_timer;
1327
1328 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1329 timer->function = inactive_task_timer;
1330}
1331
1332#ifdef CONFIG_SMP
1333
1334static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1335{
1336 struct rq *rq = rq_of_dl_rq(dl_rq);
1337
1338 if (dl_rq->earliest_dl.curr == 0 ||
1339 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1340 dl_rq->earliest_dl.curr = deadline;
1341 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1342 }
1343}
1344
1345static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1346{
1347 struct rq *rq = rq_of_dl_rq(dl_rq);
1348
1349 /*
1350 * Since we may have removed our earliest (and/or next earliest)
1351 * task we must recompute them.
1352 */
1353 if (!dl_rq->dl_nr_running) {
1354 dl_rq->earliest_dl.curr = 0;
1355 dl_rq->earliest_dl.next = 0;
1356 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1357 } else {
1358 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1359 struct sched_dl_entity *entry;
1360
1361 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1362 dl_rq->earliest_dl.curr = entry->deadline;
1363 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1364 }
1365}
1366
1367#else
1368
1369static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1370static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1371
1372#endif /* CONFIG_SMP */
1373
1374static inline
1375void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1376{
1377 int prio = dl_task_of(dl_se)->prio;
1378 u64 deadline = dl_se->deadline;
1379
1380 WARN_ON(!dl_prio(prio));
1381 dl_rq->dl_nr_running++;
1382 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1383
1384 inc_dl_deadline(dl_rq, deadline);
1385 inc_dl_migration(dl_se, dl_rq);
1386}
1387
1388static inline
1389void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1390{
1391 int prio = dl_task_of(dl_se)->prio;
1392
1393 WARN_ON(!dl_prio(prio));
1394 WARN_ON(!dl_rq->dl_nr_running);
1395 dl_rq->dl_nr_running--;
1396 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1397
1398 dec_dl_deadline(dl_rq, dl_se->deadline);
1399 dec_dl_migration(dl_se, dl_rq);
1400}
1401
1402static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1403{
1404 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1405 struct rb_node **link = &dl_rq->root.rb_root.rb_node;
1406 struct rb_node *parent = NULL;
1407 struct sched_dl_entity *entry;
1408 int leftmost = 1;
1409
1410 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1411
1412 while (*link) {
1413 parent = *link;
1414 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
1415 if (dl_time_before(dl_se->deadline, entry->deadline))
1416 link = &parent->rb_left;
1417 else {
1418 link = &parent->rb_right;
1419 leftmost = 0;
1420 }
1421 }
1422
1423 rb_link_node(&dl_se->rb_node, parent, link);
1424 rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost);
1425
1426 inc_dl_tasks(dl_se, dl_rq);
1427}
1428
1429static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1430{
1431 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1432
1433 if (RB_EMPTY_NODE(&dl_se->rb_node))
1434 return;
1435
1436 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1437 RB_CLEAR_NODE(&dl_se->rb_node);
1438
1439 dec_dl_tasks(dl_se, dl_rq);
1440}
1441
1442static void
1443enqueue_dl_entity(struct sched_dl_entity *dl_se,
1444 struct sched_dl_entity *pi_se, int flags)
1445{
1446 BUG_ON(on_dl_rq(dl_se));
1447
1448 /*
1449 * If this is a wakeup or a new instance, the scheduling
1450 * parameters of the task might need updating. Otherwise,
1451 * we want a replenishment of its runtime.
1452 */
1453 if (flags & ENQUEUE_WAKEUP) {
1454 task_contending(dl_se, flags);
1455 update_dl_entity(dl_se, pi_se);
1456 } else if (flags & ENQUEUE_REPLENISH) {
1457 replenish_dl_entity(dl_se, pi_se);
1458 } else if ((flags & ENQUEUE_RESTORE) &&
1459 dl_time_before(dl_se->deadline,
1460 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1461 setup_new_dl_entity(dl_se);
1462 }
1463
1464 __enqueue_dl_entity(dl_se);
1465}
1466
1467static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1468{
1469 __dequeue_dl_entity(dl_se);
1470}
1471
1472static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1473{
1474 struct task_struct *pi_task = rt_mutex_get_top_task(p);
1475 struct sched_dl_entity *pi_se = &p->dl;
1476
1477 /*
1478 * Use the scheduling parameters of the top pi-waiter task if:
1479 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1480 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1481 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1482 * boosted due to a SCHED_DEADLINE pi-waiter).
1483 * Otherwise we keep our runtime and deadline.
1484 */
1485 if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) {
1486 pi_se = &pi_task->dl;
1487 } else if (!dl_prio(p->normal_prio)) {
1488 /*
1489 * Special case in which we have a !SCHED_DEADLINE task
1490 * that is going to be deboosted, but exceeds its
1491 * runtime while doing so. No point in replenishing
1492 * it, as it's going to return back to its original
1493 * scheduling class after this.
1494 */
1495 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
1496 return;
1497 }
1498
1499 /*
1500 * Check if a constrained deadline task was activated
1501 * after the deadline but before the next period.
1502 * If that is the case, the task will be throttled and
1503 * the replenishment timer will be set to the next period.
1504 */
1505 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1506 dl_check_constrained_dl(&p->dl);
1507
1508 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1509 add_rq_bw(&p->dl, &rq->dl);
1510 add_running_bw(&p->dl, &rq->dl);
1511 }
1512
1513 /*
1514 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1515 * its budget it needs a replenishment and, since it now is on
1516 * its rq, the bandwidth timer callback (which clearly has not
1517 * run yet) will take care of this.
1518 * However, the active utilization does not depend on the fact
1519 * that the task is on the runqueue or not (but depends on the
1520 * task's state - in GRUB parlance, "inactive" vs "active contending").
1521 * In other words, even if a task is throttled its utilization must
1522 * be counted in the active utilization; hence, we need to call
1523 * add_running_bw().
1524 */
1525 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1526 if (flags & ENQUEUE_WAKEUP)
1527 task_contending(&p->dl, flags);
1528
1529 return;
1530 }
1531
1532 enqueue_dl_entity(&p->dl, pi_se, flags);
1533
1534 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1535 enqueue_pushable_dl_task(rq, p);
1536}
1537
1538static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1539{
1540 dequeue_dl_entity(&p->dl);
1541 dequeue_pushable_dl_task(rq, p);
1542}
1543
1544static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1545{
1546 update_curr_dl(rq);
1547 __dequeue_task_dl(rq, p, flags);
1548
1549 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1550 sub_running_bw(&p->dl, &rq->dl);
1551 sub_rq_bw(&p->dl, &rq->dl);
1552 }
1553
1554 /*
1555 * This check allows to start the inactive timer (or to immediately
1556 * decrease the active utilization, if needed) in two cases:
1557 * when the task blocks and when it is terminating
1558 * (p->state == TASK_DEAD). We can handle the two cases in the same
1559 * way, because from GRUB's point of view the same thing is happening
1560 * (the task moves from "active contending" to "active non contending"
1561 * or "inactive")
1562 */
1563 if (flags & DEQUEUE_SLEEP)
1564 task_non_contending(p);
1565}
1566
1567/*
1568 * Yield task semantic for -deadline tasks is:
1569 *
1570 * get off from the CPU until our next instance, with
1571 * a new runtime. This is of little use now, since we
1572 * don't have a bandwidth reclaiming mechanism. Anyway,
1573 * bandwidth reclaiming is planned for the future, and
1574 * yield_task_dl will indicate that some spare budget
1575 * is available for other task instances to use it.
1576 */
1577static void yield_task_dl(struct rq *rq)
1578{
1579 /*
1580 * We make the task go to sleep until its current deadline by
1581 * forcing its runtime to zero. This way, update_curr_dl() stops
1582 * it and the bandwidth timer will wake it up and will give it
1583 * new scheduling parameters (thanks to dl_yielded=1).
1584 */
1585 rq->curr->dl.dl_yielded = 1;
1586
1587 update_rq_clock(rq);
1588 update_curr_dl(rq);
1589 /*
1590 * Tell update_rq_clock() that we've just updated,
1591 * so we don't do microscopic update in schedule()
1592 * and double the fastpath cost.
1593 */
1594 rq_clock_skip_update(rq);
1595}
1596
1597#ifdef CONFIG_SMP
1598
1599static int find_later_rq(struct task_struct *task);
1600
1601static int
1602select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1603{
1604 struct task_struct *curr;
1605 struct rq *rq;
1606
1607 if (sd_flag != SD_BALANCE_WAKE)
1608 goto out;
1609
1610 rq = cpu_rq(cpu);
1611
1612 rcu_read_lock();
1613 curr = READ_ONCE(rq->curr); /* unlocked access */
1614
1615 /*
1616 * If we are dealing with a -deadline task, we must
1617 * decide where to wake it up.
1618 * If it has a later deadline and the current task
1619 * on this rq can't move (provided the waking task
1620 * can!) we prefer to send it somewhere else. On the
1621 * other hand, if it has a shorter deadline, we
1622 * try to make it stay here, it might be important.
1623 */
1624 if (unlikely(dl_task(curr)) &&
1625 (curr->nr_cpus_allowed < 2 ||
1626 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1627 (p->nr_cpus_allowed > 1)) {
1628 int target = find_later_rq(p);
1629
1630 if (target != -1 &&
1631 (dl_time_before(p->dl.deadline,
1632 cpu_rq(target)->dl.earliest_dl.curr) ||
1633 (cpu_rq(target)->dl.dl_nr_running == 0)))
1634 cpu = target;
1635 }
1636 rcu_read_unlock();
1637
1638out:
1639 return cpu;
1640}
1641
1642static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1643{
1644 struct rq *rq;
1645
1646 if (p->state != TASK_WAKING)
1647 return;
1648
1649 rq = task_rq(p);
1650 /*
1651 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1652 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1653 * rq->lock is not... So, lock it
1654 */
1655 raw_spin_lock(&rq->lock);
1656 if (p->dl.dl_non_contending) {
1657 sub_running_bw(&p->dl, &rq->dl);
1658 p->dl.dl_non_contending = 0;
1659 /*
1660 * If the timer handler is currently running and the
1661 * timer cannot be cancelled, inactive_task_timer()
1662 * will see that dl_not_contending is not set, and
1663 * will not touch the rq's active utilization,
1664 * so we are still safe.
1665 */
1666 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1667 put_task_struct(p);
1668 }
1669 sub_rq_bw(&p->dl, &rq->dl);
1670 raw_spin_unlock(&rq->lock);
1671}
1672
1673static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1674{
1675 /*
1676 * Current can't be migrated, useless to reschedule,
1677 * let's hope p can move out.
1678 */
1679 if (rq->curr->nr_cpus_allowed == 1 ||
1680 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1681 return;
1682
1683 /*
1684 * p is migratable, so let's not schedule it and
1685 * see if it is pushed or pulled somewhere else.
1686 */
1687 if (p->nr_cpus_allowed != 1 &&
1688 cpudl_find(&rq->rd->cpudl, p, NULL))
1689 return;
1690
1691 resched_curr(rq);
1692}
1693
1694static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1695{
1696 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1697 /*
1698 * This is OK, because current is on_cpu, which avoids it being
1699 * picked for load-balance and preemption/IRQs are still
1700 * disabled avoiding further scheduler activity on it and we've
1701 * not yet started the picking loop.
1702 */
1703 rq_unpin_lock(rq, rf);
1704 pull_dl_task(rq);
1705 rq_repin_lock(rq, rf);
1706 }
1707
1708 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1709}
1710#endif /* CONFIG_SMP */
1711
1712/*
1713 * Only called when both the current and waking task are -deadline
1714 * tasks.
1715 */
1716static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1717 int flags)
1718{
1719 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1720 resched_curr(rq);
1721 return;
1722 }
1723
1724#ifdef CONFIG_SMP
1725 /*
1726 * In the unlikely case current and p have the same deadline
1727 * let us try to decide what's the best thing to do...
1728 */
1729 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1730 !test_tsk_need_resched(rq->curr))
1731 check_preempt_equal_dl(rq, p);
1732#endif /* CONFIG_SMP */
1733}
1734
1735#ifdef CONFIG_SCHED_HRTICK
1736static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1737{
1738 hrtick_start(rq, p->dl.runtime);
1739}
1740#else /* !CONFIG_SCHED_HRTICK */
1741static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1742{
1743}
1744#endif
1745
1746static void set_next_task_dl(struct rq *rq, struct task_struct *p)
1747{
1748 p->se.exec_start = rq_clock_task(rq);
1749
1750 /* You can't push away the running task */
1751 dequeue_pushable_dl_task(rq, p);
1752
1753 if (hrtick_enabled(rq))
1754 start_hrtick_dl(rq, p);
1755
1756 if (rq->curr->sched_class != &dl_sched_class)
1757 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1758
1759 deadline_queue_push_tasks(rq);
1760}
1761
1762static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1763 struct dl_rq *dl_rq)
1764{
1765 struct rb_node *left = rb_first_cached(&dl_rq->root);
1766
1767 if (!left)
1768 return NULL;
1769
1770 return rb_entry(left, struct sched_dl_entity, rb_node);
1771}
1772
1773static struct task_struct *
1774pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1775{
1776 struct sched_dl_entity *dl_se;
1777 struct dl_rq *dl_rq = &rq->dl;
1778 struct task_struct *p;
1779
1780 WARN_ON_ONCE(prev || rf);
1781
1782 if (!sched_dl_runnable(rq))
1783 return NULL;
1784
1785 dl_se = pick_next_dl_entity(rq, dl_rq);
1786 BUG_ON(!dl_se);
1787 p = dl_task_of(dl_se);
1788 set_next_task_dl(rq, p);
1789 return p;
1790}
1791
1792static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1793{
1794 update_curr_dl(rq);
1795
1796 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1797 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1798 enqueue_pushable_dl_task(rq, p);
1799}
1800
1801/*
1802 * scheduler tick hitting a task of our scheduling class.
1803 *
1804 * NOTE: This function can be called remotely by the tick offload that
1805 * goes along full dynticks. Therefore no local assumption can be made
1806 * and everything must be accessed through the @rq and @curr passed in
1807 * parameters.
1808 */
1809static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1810{
1811 update_curr_dl(rq);
1812
1813 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1814 /*
1815 * Even when we have runtime, update_curr_dl() might have resulted in us
1816 * not being the leftmost task anymore. In that case NEED_RESCHED will
1817 * be set and schedule() will start a new hrtick for the next task.
1818 */
1819 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1820 is_leftmost(p, &rq->dl))
1821 start_hrtick_dl(rq, p);
1822}
1823
1824static void task_fork_dl(struct task_struct *p)
1825{
1826 /*
1827 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1828 * sched_fork()
1829 */
1830}
1831
1832#ifdef CONFIG_SMP
1833
1834/* Only try algorithms three times */
1835#define DL_MAX_TRIES 3
1836
1837static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1838{
1839 if (!task_running(rq, p) &&
1840 cpumask_test_cpu(cpu, p->cpus_ptr))
1841 return 1;
1842 return 0;
1843}
1844
1845/*
1846 * Return the earliest pushable rq's task, which is suitable to be executed
1847 * on the CPU, NULL otherwise:
1848 */
1849static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1850{
1851 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
1852 struct task_struct *p = NULL;
1853
1854 if (!has_pushable_dl_tasks(rq))
1855 return NULL;
1856
1857next_node:
1858 if (next_node) {
1859 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1860
1861 if (pick_dl_task(rq, p, cpu))
1862 return p;
1863
1864 next_node = rb_next(next_node);
1865 goto next_node;
1866 }
1867
1868 return NULL;
1869}
1870
1871static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1872
1873static int find_later_rq(struct task_struct *task)
1874{
1875 struct sched_domain *sd;
1876 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1877 int this_cpu = smp_processor_id();
1878 int cpu = task_cpu(task);
1879
1880 /* Make sure the mask is initialized first */
1881 if (unlikely(!later_mask))
1882 return -1;
1883
1884 if (task->nr_cpus_allowed == 1)
1885 return -1;
1886
1887 /*
1888 * We have to consider system topology and task affinity
1889 * first, then we can look for a suitable CPU.
1890 */
1891 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1892 return -1;
1893
1894 /*
1895 * If we are here, some targets have been found, including
1896 * the most suitable which is, among the runqueues where the
1897 * current tasks have later deadlines than the task's one, the
1898 * rq with the latest possible one.
1899 *
1900 * Now we check how well this matches with task's
1901 * affinity and system topology.
1902 *
1903 * The last CPU where the task run is our first
1904 * guess, since it is most likely cache-hot there.
1905 */
1906 if (cpumask_test_cpu(cpu, later_mask))
1907 return cpu;
1908 /*
1909 * Check if this_cpu is to be skipped (i.e., it is
1910 * not in the mask) or not.
1911 */
1912 if (!cpumask_test_cpu(this_cpu, later_mask))
1913 this_cpu = -1;
1914
1915 rcu_read_lock();
1916 for_each_domain(cpu, sd) {
1917 if (sd->flags & SD_WAKE_AFFINE) {
1918 int best_cpu;
1919
1920 /*
1921 * If possible, preempting this_cpu is
1922 * cheaper than migrating.
1923 */
1924 if (this_cpu != -1 &&
1925 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1926 rcu_read_unlock();
1927 return this_cpu;
1928 }
1929
1930 best_cpu = cpumask_first_and(later_mask,
1931 sched_domain_span(sd));
1932 /*
1933 * Last chance: if a CPU being in both later_mask
1934 * and current sd span is valid, that becomes our
1935 * choice. Of course, the latest possible CPU is
1936 * already under consideration through later_mask.
1937 */
1938 if (best_cpu < nr_cpu_ids) {
1939 rcu_read_unlock();
1940 return best_cpu;
1941 }
1942 }
1943 }
1944 rcu_read_unlock();
1945
1946 /*
1947 * At this point, all our guesses failed, we just return
1948 * 'something', and let the caller sort the things out.
1949 */
1950 if (this_cpu != -1)
1951 return this_cpu;
1952
1953 cpu = cpumask_any(later_mask);
1954 if (cpu < nr_cpu_ids)
1955 return cpu;
1956
1957 return -1;
1958}
1959
1960/* Locks the rq it finds */
1961static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1962{
1963 struct rq *later_rq = NULL;
1964 int tries;
1965 int cpu;
1966
1967 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1968 cpu = find_later_rq(task);
1969
1970 if ((cpu == -1) || (cpu == rq->cpu))
1971 break;
1972
1973 later_rq = cpu_rq(cpu);
1974
1975 if (later_rq->dl.dl_nr_running &&
1976 !dl_time_before(task->dl.deadline,
1977 later_rq->dl.earliest_dl.curr)) {
1978 /*
1979 * Target rq has tasks of equal or earlier deadline,
1980 * retrying does not release any lock and is unlikely
1981 * to yield a different result.
1982 */
1983 later_rq = NULL;
1984 break;
1985 }
1986
1987 /* Retry if something changed. */
1988 if (double_lock_balance(rq, later_rq)) {
1989 if (unlikely(task_rq(task) != rq ||
1990 !cpumask_test_cpu(later_rq->cpu, task->cpus_ptr) ||
1991 task_running(rq, task) ||
1992 !dl_task(task) ||
1993 !task_on_rq_queued(task))) {
1994 double_unlock_balance(rq, later_rq);
1995 later_rq = NULL;
1996 break;
1997 }
1998 }
1999
2000 /*
2001 * If the rq we found has no -deadline task, or
2002 * its earliest one has a later deadline than our
2003 * task, the rq is a good one.
2004 */
2005 if (!later_rq->dl.dl_nr_running ||
2006 dl_time_before(task->dl.deadline,
2007 later_rq->dl.earliest_dl.curr))
2008 break;
2009
2010 /* Otherwise we try again. */
2011 double_unlock_balance(rq, later_rq);
2012 later_rq = NULL;
2013 }
2014
2015 return later_rq;
2016}
2017
2018static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2019{
2020 struct task_struct *p;
2021
2022 if (!has_pushable_dl_tasks(rq))
2023 return NULL;
2024
2025 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
2026 struct task_struct, pushable_dl_tasks);
2027
2028 BUG_ON(rq->cpu != task_cpu(p));
2029 BUG_ON(task_current(rq, p));
2030 BUG_ON(p->nr_cpus_allowed <= 1);
2031
2032 BUG_ON(!task_on_rq_queued(p));
2033 BUG_ON(!dl_task(p));
2034
2035 return p;
2036}
2037
2038/*
2039 * See if the non running -deadline tasks on this rq
2040 * can be sent to some other CPU where they can preempt
2041 * and start executing.
2042 */
2043static int push_dl_task(struct rq *rq)
2044{
2045 struct task_struct *next_task;
2046 struct rq *later_rq;
2047 int ret = 0;
2048
2049 if (!rq->dl.overloaded)
2050 return 0;
2051
2052 next_task = pick_next_pushable_dl_task(rq);
2053 if (!next_task)
2054 return 0;
2055
2056retry:
2057 if (WARN_ON(next_task == rq->curr))
2058 return 0;
2059
2060 /*
2061 * If next_task preempts rq->curr, and rq->curr
2062 * can move away, it makes sense to just reschedule
2063 * without going further in pushing next_task.
2064 */
2065 if (dl_task(rq->curr) &&
2066 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2067 rq->curr->nr_cpus_allowed > 1) {
2068 resched_curr(rq);
2069 return 0;
2070 }
2071
2072 /* We might release rq lock */
2073 get_task_struct(next_task);
2074
2075 /* Will lock the rq it'll find */
2076 later_rq = find_lock_later_rq(next_task, rq);
2077 if (!later_rq) {
2078 struct task_struct *task;
2079
2080 /*
2081 * We must check all this again, since
2082 * find_lock_later_rq releases rq->lock and it is
2083 * then possible that next_task has migrated.
2084 */
2085 task = pick_next_pushable_dl_task(rq);
2086 if (task == next_task) {
2087 /*
2088 * The task is still there. We don't try
2089 * again, some other CPU will pull it when ready.
2090 */
2091 goto out;
2092 }
2093
2094 if (!task)
2095 /* No more tasks */
2096 goto out;
2097
2098 put_task_struct(next_task);
2099 next_task = task;
2100 goto retry;
2101 }
2102
2103 deactivate_task(rq, next_task, 0);
2104 set_task_cpu(next_task, later_rq->cpu);
2105
2106 /*
2107 * Update the later_rq clock here, because the clock is used
2108 * by the cpufreq_update_util() inside __add_running_bw().
2109 */
2110 update_rq_clock(later_rq);
2111 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
2112 ret = 1;
2113
2114 resched_curr(later_rq);
2115
2116 double_unlock_balance(rq, later_rq);
2117
2118out:
2119 put_task_struct(next_task);
2120
2121 return ret;
2122}
2123
2124static void push_dl_tasks(struct rq *rq)
2125{
2126 /* push_dl_task() will return true if it moved a -deadline task */
2127 while (push_dl_task(rq))
2128 ;
2129}
2130
2131static void pull_dl_task(struct rq *this_rq)
2132{
2133 int this_cpu = this_rq->cpu, cpu;
2134 struct task_struct *p;
2135 bool resched = false;
2136 struct rq *src_rq;
2137 u64 dmin = LONG_MAX;
2138
2139 if (likely(!dl_overloaded(this_rq)))
2140 return;
2141
2142 /*
2143 * Match the barrier from dl_set_overloaded; this guarantees that if we
2144 * see overloaded we must also see the dlo_mask bit.
2145 */
2146 smp_rmb();
2147
2148 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2149 if (this_cpu == cpu)
2150 continue;
2151
2152 src_rq = cpu_rq(cpu);
2153
2154 /*
2155 * It looks racy, abd it is! However, as in sched_rt.c,
2156 * we are fine with this.
2157 */
2158 if (this_rq->dl.dl_nr_running &&
2159 dl_time_before(this_rq->dl.earliest_dl.curr,
2160 src_rq->dl.earliest_dl.next))
2161 continue;
2162
2163 /* Might drop this_rq->lock */
2164 double_lock_balance(this_rq, src_rq);
2165
2166 /*
2167 * If there are no more pullable tasks on the
2168 * rq, we're done with it.
2169 */
2170 if (src_rq->dl.dl_nr_running <= 1)
2171 goto skip;
2172
2173 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2174
2175 /*
2176 * We found a task to be pulled if:
2177 * - it preempts our current (if there's one),
2178 * - it will preempt the last one we pulled (if any).
2179 */
2180 if (p && dl_time_before(p->dl.deadline, dmin) &&
2181 (!this_rq->dl.dl_nr_running ||
2182 dl_time_before(p->dl.deadline,
2183 this_rq->dl.earliest_dl.curr))) {
2184 WARN_ON(p == src_rq->curr);
2185 WARN_ON(!task_on_rq_queued(p));
2186
2187 /*
2188 * Then we pull iff p has actually an earlier
2189 * deadline than the current task of its runqueue.
2190 */
2191 if (dl_time_before(p->dl.deadline,
2192 src_rq->curr->dl.deadline))
2193 goto skip;
2194
2195 resched = true;
2196
2197 deactivate_task(src_rq, p, 0);
2198 set_task_cpu(p, this_cpu);
2199 activate_task(this_rq, p, 0);
2200 dmin = p->dl.deadline;
2201
2202 /* Is there any other task even earlier? */
2203 }
2204skip:
2205 double_unlock_balance(this_rq, src_rq);
2206 }
2207
2208 if (resched)
2209 resched_curr(this_rq);
2210}
2211
2212/*
2213 * Since the task is not running and a reschedule is not going to happen
2214 * anytime soon on its runqueue, we try pushing it away now.
2215 */
2216static void task_woken_dl(struct rq *rq, struct task_struct *p)
2217{
2218 if (!task_running(rq, p) &&
2219 !test_tsk_need_resched(rq->curr) &&
2220 p->nr_cpus_allowed > 1 &&
2221 dl_task(rq->curr) &&
2222 (rq->curr->nr_cpus_allowed < 2 ||
2223 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2224 push_dl_tasks(rq);
2225 }
2226}
2227
2228static void set_cpus_allowed_dl(struct task_struct *p,
2229 const struct cpumask *new_mask)
2230{
2231 struct root_domain *src_rd;
2232 struct rq *rq;
2233
2234 BUG_ON(!dl_task(p));
2235
2236 rq = task_rq(p);
2237 src_rd = rq->rd;
2238 /*
2239 * Migrating a SCHED_DEADLINE task between exclusive
2240 * cpusets (different root_domains) entails a bandwidth
2241 * update. We already made space for us in the destination
2242 * domain (see cpuset_can_attach()).
2243 */
2244 if (!cpumask_intersects(src_rd->span, new_mask)) {
2245 struct dl_bw *src_dl_b;
2246
2247 src_dl_b = dl_bw_of(cpu_of(rq));
2248 /*
2249 * We now free resources of the root_domain we are migrating
2250 * off. In the worst case, sched_setattr() may temporary fail
2251 * until we complete the update.
2252 */
2253 raw_spin_lock(&src_dl_b->lock);
2254 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2255 raw_spin_unlock(&src_dl_b->lock);
2256 }
2257
2258 set_cpus_allowed_common(p, new_mask);
2259}
2260
2261/* Assumes rq->lock is held */
2262static void rq_online_dl(struct rq *rq)
2263{
2264 if (rq->dl.overloaded)
2265 dl_set_overload(rq);
2266
2267 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2268 if (rq->dl.dl_nr_running > 0)
2269 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2270}
2271
2272/* Assumes rq->lock is held */
2273static void rq_offline_dl(struct rq *rq)
2274{
2275 if (rq->dl.overloaded)
2276 dl_clear_overload(rq);
2277
2278 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2279 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2280}
2281
2282void __init init_sched_dl_class(void)
2283{
2284 unsigned int i;
2285
2286 for_each_possible_cpu(i)
2287 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2288 GFP_KERNEL, cpu_to_node(i));
2289}
2290
2291void dl_add_task_root_domain(struct task_struct *p)
2292{
2293 struct rq_flags rf;
2294 struct rq *rq;
2295 struct dl_bw *dl_b;
2296
2297 rq = task_rq_lock(p, &rf);
2298 if (!dl_task(p))
2299 goto unlock;
2300
2301 dl_b = &rq->rd->dl_bw;
2302 raw_spin_lock(&dl_b->lock);
2303
2304 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2305
2306 raw_spin_unlock(&dl_b->lock);
2307
2308unlock:
2309 task_rq_unlock(rq, p, &rf);
2310}
2311
2312void dl_clear_root_domain(struct root_domain *rd)
2313{
2314 unsigned long flags;
2315
2316 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2317 rd->dl_bw.total_bw = 0;
2318 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2319}
2320
2321#endif /* CONFIG_SMP */
2322
2323static void switched_from_dl(struct rq *rq, struct task_struct *p)
2324{
2325 /*
2326 * task_non_contending() can start the "inactive timer" (if the 0-lag
2327 * time is in the future). If the task switches back to dl before
2328 * the "inactive timer" fires, it can continue to consume its current
2329 * runtime using its current deadline. If it stays outside of
2330 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2331 * will reset the task parameters.
2332 */
2333 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2334 task_non_contending(p);
2335
2336 if (!task_on_rq_queued(p)) {
2337 /*
2338 * Inactive timer is armed. However, p is leaving DEADLINE and
2339 * might migrate away from this rq while continuing to run on
2340 * some other class. We need to remove its contribution from
2341 * this rq running_bw now, or sub_rq_bw (below) will complain.
2342 */
2343 if (p->dl.dl_non_contending)
2344 sub_running_bw(&p->dl, &rq->dl);
2345 sub_rq_bw(&p->dl, &rq->dl);
2346 }
2347
2348 /*
2349 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2350 * at the 0-lag time, because the task could have been migrated
2351 * while SCHED_OTHER in the meanwhile.
2352 */
2353 if (p->dl.dl_non_contending)
2354 p->dl.dl_non_contending = 0;
2355
2356 /*
2357 * Since this might be the only -deadline task on the rq,
2358 * this is the right place to try to pull some other one
2359 * from an overloaded CPU, if any.
2360 */
2361 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2362 return;
2363
2364 deadline_queue_pull_task(rq);
2365}
2366
2367/*
2368 * When switching to -deadline, we may overload the rq, then
2369 * we try to push someone off, if possible.
2370 */
2371static void switched_to_dl(struct rq *rq, struct task_struct *p)
2372{
2373 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2374 put_task_struct(p);
2375
2376 /* If p is not queued we will update its parameters at next wakeup. */
2377 if (!task_on_rq_queued(p)) {
2378 add_rq_bw(&p->dl, &rq->dl);
2379
2380 return;
2381 }
2382
2383 if (rq->curr != p) {
2384#ifdef CONFIG_SMP
2385 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2386 deadline_queue_push_tasks(rq);
2387#endif
2388 if (dl_task(rq->curr))
2389 check_preempt_curr_dl(rq, p, 0);
2390 else
2391 resched_curr(rq);
2392 }
2393}
2394
2395/*
2396 * If the scheduling parameters of a -deadline task changed,
2397 * a push or pull operation might be needed.
2398 */
2399static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2400 int oldprio)
2401{
2402 if (task_on_rq_queued(p) || rq->curr == p) {
2403#ifdef CONFIG_SMP
2404 /*
2405 * This might be too much, but unfortunately
2406 * we don't have the old deadline value, and
2407 * we can't argue if the task is increasing
2408 * or lowering its prio, so...
2409 */
2410 if (!rq->dl.overloaded)
2411 deadline_queue_pull_task(rq);
2412
2413 /*
2414 * If we now have a earlier deadline task than p,
2415 * then reschedule, provided p is still on this
2416 * runqueue.
2417 */
2418 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2419 resched_curr(rq);
2420#else
2421 /*
2422 * Again, we don't know if p has a earlier
2423 * or later deadline, so let's blindly set a
2424 * (maybe not needed) rescheduling point.
2425 */
2426 resched_curr(rq);
2427#endif /* CONFIG_SMP */
2428 }
2429}
2430
2431const struct sched_class dl_sched_class = {
2432 .next = &rt_sched_class,
2433 .enqueue_task = enqueue_task_dl,
2434 .dequeue_task = dequeue_task_dl,
2435 .yield_task = yield_task_dl,
2436
2437 .check_preempt_curr = check_preempt_curr_dl,
2438
2439 .pick_next_task = pick_next_task_dl,
2440 .put_prev_task = put_prev_task_dl,
2441 .set_next_task = set_next_task_dl,
2442
2443#ifdef CONFIG_SMP
2444 .balance = balance_dl,
2445 .select_task_rq = select_task_rq_dl,
2446 .migrate_task_rq = migrate_task_rq_dl,
2447 .set_cpus_allowed = set_cpus_allowed_dl,
2448 .rq_online = rq_online_dl,
2449 .rq_offline = rq_offline_dl,
2450 .task_woken = task_woken_dl,
2451#endif
2452
2453 .task_tick = task_tick_dl,
2454 .task_fork = task_fork_dl,
2455
2456 .prio_changed = prio_changed_dl,
2457 .switched_from = switched_from_dl,
2458 .switched_to = switched_to_dl,
2459
2460 .update_curr = update_curr_dl,
2461};
2462
2463int sched_dl_global_validate(void)
2464{
2465 u64 runtime = global_rt_runtime();
2466 u64 period = global_rt_period();
2467 u64 new_bw = to_ratio(period, runtime);
2468 struct dl_bw *dl_b;
2469 int cpu, ret = 0;
2470 unsigned long flags;
2471
2472 /*
2473 * Here we want to check the bandwidth not being set to some
2474 * value smaller than the currently allocated bandwidth in
2475 * any of the root_domains.
2476 *
2477 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2478 * cycling on root_domains... Discussion on different/better
2479 * solutions is welcome!
2480 */
2481 for_each_possible_cpu(cpu) {
2482 rcu_read_lock_sched();
2483 dl_b = dl_bw_of(cpu);
2484
2485 raw_spin_lock_irqsave(&dl_b->lock, flags);
2486 if (new_bw < dl_b->total_bw)
2487 ret = -EBUSY;
2488 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2489
2490 rcu_read_unlock_sched();
2491
2492 if (ret)
2493 break;
2494 }
2495
2496 return ret;
2497}
2498
2499void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2500{
2501 if (global_rt_runtime() == RUNTIME_INF) {
2502 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2503 dl_rq->extra_bw = 1 << BW_SHIFT;
2504 } else {
2505 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2506 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2507 dl_rq->extra_bw = to_ratio(global_rt_period(),
2508 global_rt_runtime());
2509 }
2510}
2511
2512void sched_dl_do_global(void)
2513{
2514 u64 new_bw = -1;
2515 struct dl_bw *dl_b;
2516 int cpu;
2517 unsigned long flags;
2518
2519 def_dl_bandwidth.dl_period = global_rt_period();
2520 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2521
2522 if (global_rt_runtime() != RUNTIME_INF)
2523 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2524
2525 /*
2526 * FIXME: As above...
2527 */
2528 for_each_possible_cpu(cpu) {
2529 rcu_read_lock_sched();
2530 dl_b = dl_bw_of(cpu);
2531
2532 raw_spin_lock_irqsave(&dl_b->lock, flags);
2533 dl_b->bw = new_bw;
2534 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2535
2536 rcu_read_unlock_sched();
2537 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2538 }
2539}
2540
2541/*
2542 * We must be sure that accepting a new task (or allowing changing the
2543 * parameters of an existing one) is consistent with the bandwidth
2544 * constraints. If yes, this function also accordingly updates the currently
2545 * allocated bandwidth to reflect the new situation.
2546 *
2547 * This function is called while holding p's rq->lock.
2548 */
2549int sched_dl_overflow(struct task_struct *p, int policy,
2550 const struct sched_attr *attr)
2551{
2552 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2553 u64 period = attr->sched_period ?: attr->sched_deadline;
2554 u64 runtime = attr->sched_runtime;
2555 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2556 int cpus, err = -1;
2557
2558 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2559 return 0;
2560
2561 /* !deadline task may carry old deadline bandwidth */
2562 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2563 return 0;
2564
2565 /*
2566 * Either if a task, enters, leave, or stays -deadline but changes
2567 * its parameters, we may need to update accordingly the total
2568 * allocated bandwidth of the container.
2569 */
2570 raw_spin_lock(&dl_b->lock);
2571 cpus = dl_bw_cpus(task_cpu(p));
2572 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2573 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2574 if (hrtimer_active(&p->dl.inactive_timer))
2575 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2576 __dl_add(dl_b, new_bw, cpus);
2577 err = 0;
2578 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2579 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2580 /*
2581 * XXX this is slightly incorrect: when the task
2582 * utilization decreases, we should delay the total
2583 * utilization change until the task's 0-lag point.
2584 * But this would require to set the task's "inactive
2585 * timer" when the task is not inactive.
2586 */
2587 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2588 __dl_add(dl_b, new_bw, cpus);
2589 dl_change_utilization(p, new_bw);
2590 err = 0;
2591 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2592 /*
2593 * Do not decrease the total deadline utilization here,
2594 * switched_from_dl() will take care to do it at the correct
2595 * (0-lag) time.
2596 */
2597 err = 0;
2598 }
2599 raw_spin_unlock(&dl_b->lock);
2600
2601 return err;
2602}
2603
2604/*
2605 * This function initializes the sched_dl_entity of a newly becoming
2606 * SCHED_DEADLINE task.
2607 *
2608 * Only the static values are considered here, the actual runtime and the
2609 * absolute deadline will be properly calculated when the task is enqueued
2610 * for the first time with its new policy.
2611 */
2612void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2613{
2614 struct sched_dl_entity *dl_se = &p->dl;
2615
2616 dl_se->dl_runtime = attr->sched_runtime;
2617 dl_se->dl_deadline = attr->sched_deadline;
2618 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2619 dl_se->flags = attr->sched_flags;
2620 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2621 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2622}
2623
2624void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2625{
2626 struct sched_dl_entity *dl_se = &p->dl;
2627
2628 attr->sched_priority = p->rt_priority;
2629 attr->sched_runtime = dl_se->dl_runtime;
2630 attr->sched_deadline = dl_se->dl_deadline;
2631 attr->sched_period = dl_se->dl_period;
2632 attr->sched_flags = dl_se->flags;
2633}
2634
2635/*
2636 * This function validates the new parameters of a -deadline task.
2637 * We ask for the deadline not being zero, and greater or equal
2638 * than the runtime, as well as the period of being zero or
2639 * greater than deadline. Furthermore, we have to be sure that
2640 * user parameters are above the internal resolution of 1us (we
2641 * check sched_runtime only since it is always the smaller one) and
2642 * below 2^63 ns (we have to check both sched_deadline and
2643 * sched_period, as the latter can be zero).
2644 */
2645bool __checkparam_dl(const struct sched_attr *attr)
2646{
2647 /* special dl tasks don't actually use any parameter */
2648 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2649 return true;
2650
2651 /* deadline != 0 */
2652 if (attr->sched_deadline == 0)
2653 return false;
2654
2655 /*
2656 * Since we truncate DL_SCALE bits, make sure we're at least
2657 * that big.
2658 */
2659 if (attr->sched_runtime < (1ULL << DL_SCALE))
2660 return false;
2661
2662 /*
2663 * Since we use the MSB for wrap-around and sign issues, make
2664 * sure it's not set (mind that period can be equal to zero).
2665 */
2666 if (attr->sched_deadline & (1ULL << 63) ||
2667 attr->sched_period & (1ULL << 63))
2668 return false;
2669
2670 /* runtime <= deadline <= period (if period != 0) */
2671 if ((attr->sched_period != 0 &&
2672 attr->sched_period < attr->sched_deadline) ||
2673 attr->sched_deadline < attr->sched_runtime)
2674 return false;
2675
2676 return true;
2677}
2678
2679/*
2680 * This function clears the sched_dl_entity static params.
2681 */
2682void __dl_clear_params(struct task_struct *p)
2683{
2684 struct sched_dl_entity *dl_se = &p->dl;
2685
2686 dl_se->dl_runtime = 0;
2687 dl_se->dl_deadline = 0;
2688 dl_se->dl_period = 0;
2689 dl_se->flags = 0;
2690 dl_se->dl_bw = 0;
2691 dl_se->dl_density = 0;
2692
2693 dl_se->dl_throttled = 0;
2694 dl_se->dl_yielded = 0;
2695 dl_se->dl_non_contending = 0;
2696 dl_se->dl_overrun = 0;
2697}
2698
2699bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2700{
2701 struct sched_dl_entity *dl_se = &p->dl;
2702
2703 if (dl_se->dl_runtime != attr->sched_runtime ||
2704 dl_se->dl_deadline != attr->sched_deadline ||
2705 dl_se->dl_period != attr->sched_period ||
2706 dl_se->flags != attr->sched_flags)
2707 return true;
2708
2709 return false;
2710}
2711
2712#ifdef CONFIG_SMP
2713int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2714{
2715 unsigned int dest_cpu;
2716 struct dl_bw *dl_b;
2717 bool overflow;
2718 int cpus, ret;
2719 unsigned long flags;
2720
2721 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2722
2723 rcu_read_lock_sched();
2724 dl_b = dl_bw_of(dest_cpu);
2725 raw_spin_lock_irqsave(&dl_b->lock, flags);
2726 cpus = dl_bw_cpus(dest_cpu);
2727 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
2728 if (overflow) {
2729 ret = -EBUSY;
2730 } else {
2731 /*
2732 * We reserve space for this task in the destination
2733 * root_domain, as we can't fail after this point.
2734 * We will free resources in the source root_domain
2735 * later on (see set_cpus_allowed_dl()).
2736 */
2737 __dl_add(dl_b, p->dl.dl_bw, cpus);
2738 ret = 0;
2739 }
2740 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2741 rcu_read_unlock_sched();
2742
2743 return ret;
2744}
2745
2746int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2747 const struct cpumask *trial)
2748{
2749 int ret = 1, trial_cpus;
2750 struct dl_bw *cur_dl_b;
2751 unsigned long flags;
2752
2753 rcu_read_lock_sched();
2754 cur_dl_b = dl_bw_of(cpumask_any(cur));
2755 trial_cpus = cpumask_weight(trial);
2756
2757 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2758 if (cur_dl_b->bw != -1 &&
2759 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2760 ret = 0;
2761 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2762 rcu_read_unlock_sched();
2763
2764 return ret;
2765}
2766
2767bool dl_cpu_busy(unsigned int cpu)
2768{
2769 unsigned long flags;
2770 struct dl_bw *dl_b;
2771 bool overflow;
2772 int cpus;
2773
2774 rcu_read_lock_sched();
2775 dl_b = dl_bw_of(cpu);
2776 raw_spin_lock_irqsave(&dl_b->lock, flags);
2777 cpus = dl_bw_cpus(cpu);
2778 overflow = __dl_overflow(dl_b, cpus, 0, 0);
2779 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2780 rcu_read_unlock_sched();
2781
2782 return overflow;
2783}
2784#endif
2785
2786#ifdef CONFIG_SCHED_DEBUG
2787void print_dl_stats(struct seq_file *m, int cpu)
2788{
2789 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2790}
2791#endif /* CONFIG_SCHED_DEBUG */
1/*
2 * Deadline Scheduling Class (SCHED_DEADLINE)
3 *
4 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
5 *
6 * Tasks that periodically executes their instances for less than their
7 * runtime won't miss any of their deadlines.
8 * Tasks that are not periodic or sporadic or that tries to execute more
9 * than their reserved bandwidth will be slowed down (and may potentially
10 * miss some of their deadlines), and won't affect any other task.
11 *
12 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
13 * Juri Lelli <juri.lelli@gmail.com>,
14 * Michael Trimarchi <michael@amarulasolutions.com>,
15 * Fabio Checconi <fchecconi@gmail.com>
16 */
17#include "sched.h"
18
19#include <linux/slab.h>
20
21struct dl_bandwidth def_dl_bandwidth;
22
23static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
24{
25 return container_of(dl_se, struct task_struct, dl);
26}
27
28static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
29{
30 return container_of(dl_rq, struct rq, dl);
31}
32
33static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
34{
35 struct task_struct *p = dl_task_of(dl_se);
36 struct rq *rq = task_rq(p);
37
38 return &rq->dl;
39}
40
41static inline int on_dl_rq(struct sched_dl_entity *dl_se)
42{
43 return !RB_EMPTY_NODE(&dl_se->rb_node);
44}
45
46static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
47{
48 struct sched_dl_entity *dl_se = &p->dl;
49
50 return dl_rq->rb_leftmost == &dl_se->rb_node;
51}
52
53void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
54{
55 raw_spin_lock_init(&dl_b->dl_runtime_lock);
56 dl_b->dl_period = period;
57 dl_b->dl_runtime = runtime;
58}
59
60void init_dl_bw(struct dl_bw *dl_b)
61{
62 raw_spin_lock_init(&dl_b->lock);
63 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
64 if (global_rt_runtime() == RUNTIME_INF)
65 dl_b->bw = -1;
66 else
67 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
68 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
69 dl_b->total_bw = 0;
70}
71
72void init_dl_rq(struct dl_rq *dl_rq)
73{
74 dl_rq->rb_root = RB_ROOT;
75
76#ifdef CONFIG_SMP
77 /* zero means no -deadline tasks */
78 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
79
80 dl_rq->dl_nr_migratory = 0;
81 dl_rq->overloaded = 0;
82 dl_rq->pushable_dl_tasks_root = RB_ROOT;
83#else
84 init_dl_bw(&dl_rq->dl_bw);
85#endif
86}
87
88#ifdef CONFIG_SMP
89
90static inline int dl_overloaded(struct rq *rq)
91{
92 return atomic_read(&rq->rd->dlo_count);
93}
94
95static inline void dl_set_overload(struct rq *rq)
96{
97 if (!rq->online)
98 return;
99
100 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
101 /*
102 * Must be visible before the overload count is
103 * set (as in sched_rt.c).
104 *
105 * Matched by the barrier in pull_dl_task().
106 */
107 smp_wmb();
108 atomic_inc(&rq->rd->dlo_count);
109}
110
111static inline void dl_clear_overload(struct rq *rq)
112{
113 if (!rq->online)
114 return;
115
116 atomic_dec(&rq->rd->dlo_count);
117 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
118}
119
120static void update_dl_migration(struct dl_rq *dl_rq)
121{
122 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
123 if (!dl_rq->overloaded) {
124 dl_set_overload(rq_of_dl_rq(dl_rq));
125 dl_rq->overloaded = 1;
126 }
127 } else if (dl_rq->overloaded) {
128 dl_clear_overload(rq_of_dl_rq(dl_rq));
129 dl_rq->overloaded = 0;
130 }
131}
132
133static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
134{
135 struct task_struct *p = dl_task_of(dl_se);
136
137 if (p->nr_cpus_allowed > 1)
138 dl_rq->dl_nr_migratory++;
139
140 update_dl_migration(dl_rq);
141}
142
143static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
144{
145 struct task_struct *p = dl_task_of(dl_se);
146
147 if (p->nr_cpus_allowed > 1)
148 dl_rq->dl_nr_migratory--;
149
150 update_dl_migration(dl_rq);
151}
152
153/*
154 * The list of pushable -deadline task is not a plist, like in
155 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
156 */
157static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
158{
159 struct dl_rq *dl_rq = &rq->dl;
160 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node;
161 struct rb_node *parent = NULL;
162 struct task_struct *entry;
163 int leftmost = 1;
164
165 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
166
167 while (*link) {
168 parent = *link;
169 entry = rb_entry(parent, struct task_struct,
170 pushable_dl_tasks);
171 if (dl_entity_preempt(&p->dl, &entry->dl))
172 link = &parent->rb_left;
173 else {
174 link = &parent->rb_right;
175 leftmost = 0;
176 }
177 }
178
179 if (leftmost) {
180 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
181 dl_rq->earliest_dl.next = p->dl.deadline;
182 }
183
184 rb_link_node(&p->pushable_dl_tasks, parent, link);
185 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
186}
187
188static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
189{
190 struct dl_rq *dl_rq = &rq->dl;
191
192 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
193 return;
194
195 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
196 struct rb_node *next_node;
197
198 next_node = rb_next(&p->pushable_dl_tasks);
199 dl_rq->pushable_dl_tasks_leftmost = next_node;
200 if (next_node) {
201 dl_rq->earliest_dl.next = rb_entry(next_node,
202 struct task_struct, pushable_dl_tasks)->dl.deadline;
203 }
204 }
205
206 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
207 RB_CLEAR_NODE(&p->pushable_dl_tasks);
208}
209
210static inline int has_pushable_dl_tasks(struct rq *rq)
211{
212 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
213}
214
215static int push_dl_task(struct rq *rq);
216
217static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
218{
219 return dl_task(prev);
220}
221
222static DEFINE_PER_CPU(struct callback_head, dl_push_head);
223static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
224
225static void push_dl_tasks(struct rq *);
226static void pull_dl_task(struct rq *);
227
228static inline void queue_push_tasks(struct rq *rq)
229{
230 if (!has_pushable_dl_tasks(rq))
231 return;
232
233 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
234}
235
236static inline void queue_pull_task(struct rq *rq)
237{
238 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
239}
240
241static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
242
243static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
244{
245 struct rq *later_rq = NULL;
246 bool fallback = false;
247
248 later_rq = find_lock_later_rq(p, rq);
249
250 if (!later_rq) {
251 int cpu;
252
253 /*
254 * If we cannot preempt any rq, fall back to pick any
255 * online cpu.
256 */
257 fallback = true;
258 cpu = cpumask_any_and(cpu_active_mask, tsk_cpus_allowed(p));
259 if (cpu >= nr_cpu_ids) {
260 /*
261 * Fail to find any suitable cpu.
262 * The task will never come back!
263 */
264 BUG_ON(dl_bandwidth_enabled());
265
266 /*
267 * If admission control is disabled we
268 * try a little harder to let the task
269 * run.
270 */
271 cpu = cpumask_any(cpu_active_mask);
272 }
273 later_rq = cpu_rq(cpu);
274 double_lock_balance(rq, later_rq);
275 }
276
277 /*
278 * By now the task is replenished and enqueued; migrate it.
279 */
280 deactivate_task(rq, p, 0);
281 set_task_cpu(p, later_rq->cpu);
282 activate_task(later_rq, p, 0);
283
284 if (!fallback)
285 resched_curr(later_rq);
286
287 double_unlock_balance(later_rq, rq);
288
289 return later_rq;
290}
291
292#else
293
294static inline
295void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
296{
297}
298
299static inline
300void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
301{
302}
303
304static inline
305void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
306{
307}
308
309static inline
310void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
311{
312}
313
314static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
315{
316 return false;
317}
318
319static inline void pull_dl_task(struct rq *rq)
320{
321}
322
323static inline void queue_push_tasks(struct rq *rq)
324{
325}
326
327static inline void queue_pull_task(struct rq *rq)
328{
329}
330#endif /* CONFIG_SMP */
331
332static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
333static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
334static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
335 int flags);
336
337/*
338 * We are being explicitly informed that a new instance is starting,
339 * and this means that:
340 * - the absolute deadline of the entity has to be placed at
341 * current time + relative deadline;
342 * - the runtime of the entity has to be set to the maximum value.
343 *
344 * The capability of specifying such event is useful whenever a -deadline
345 * entity wants to (try to!) synchronize its behaviour with the scheduler's
346 * one, and to (try to!) reconcile itself with its own scheduling
347 * parameters.
348 */
349static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se,
350 struct sched_dl_entity *pi_se)
351{
352 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
353 struct rq *rq = rq_of_dl_rq(dl_rq);
354
355 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
356
357 /*
358 * We are racing with the deadline timer. So, do nothing because
359 * the deadline timer handler will take care of properly recharging
360 * the runtime and postponing the deadline
361 */
362 if (dl_se->dl_throttled)
363 return;
364
365 /*
366 * We use the regular wall clock time to set deadlines in the
367 * future; in fact, we must consider execution overheads (time
368 * spent on hardirq context, etc.).
369 */
370 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
371 dl_se->runtime = pi_se->dl_runtime;
372}
373
374/*
375 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
376 * possibility of a entity lasting more than what it declared, and thus
377 * exhausting its runtime.
378 *
379 * Here we are interested in making runtime overrun possible, but we do
380 * not want a entity which is misbehaving to affect the scheduling of all
381 * other entities.
382 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
383 * is used, in order to confine each entity within its own bandwidth.
384 *
385 * This function deals exactly with that, and ensures that when the runtime
386 * of a entity is replenished, its deadline is also postponed. That ensures
387 * the overrunning entity can't interfere with other entity in the system and
388 * can't make them miss their deadlines. Reasons why this kind of overruns
389 * could happen are, typically, a entity voluntarily trying to overcome its
390 * runtime, or it just underestimated it during sched_setattr().
391 */
392static void replenish_dl_entity(struct sched_dl_entity *dl_se,
393 struct sched_dl_entity *pi_se)
394{
395 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
396 struct rq *rq = rq_of_dl_rq(dl_rq);
397
398 BUG_ON(pi_se->dl_runtime <= 0);
399
400 /*
401 * This could be the case for a !-dl task that is boosted.
402 * Just go with full inherited parameters.
403 */
404 if (dl_se->dl_deadline == 0) {
405 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
406 dl_se->runtime = pi_se->dl_runtime;
407 }
408
409 if (dl_se->dl_yielded && dl_se->runtime > 0)
410 dl_se->runtime = 0;
411
412 /*
413 * We keep moving the deadline away until we get some
414 * available runtime for the entity. This ensures correct
415 * handling of situations where the runtime overrun is
416 * arbitrary large.
417 */
418 while (dl_se->runtime <= 0) {
419 dl_se->deadline += pi_se->dl_period;
420 dl_se->runtime += pi_se->dl_runtime;
421 }
422
423 /*
424 * At this point, the deadline really should be "in
425 * the future" with respect to rq->clock. If it's
426 * not, we are, for some reason, lagging too much!
427 * Anyway, after having warn userspace abut that,
428 * we still try to keep the things running by
429 * resetting the deadline and the budget of the
430 * entity.
431 */
432 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
433 printk_deferred_once("sched: DL replenish lagged too much\n");
434 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
435 dl_se->runtime = pi_se->dl_runtime;
436 }
437
438 if (dl_se->dl_yielded)
439 dl_se->dl_yielded = 0;
440 if (dl_se->dl_throttled)
441 dl_se->dl_throttled = 0;
442}
443
444/*
445 * Here we check if --at time t-- an entity (which is probably being
446 * [re]activated or, in general, enqueued) can use its remaining runtime
447 * and its current deadline _without_ exceeding the bandwidth it is
448 * assigned (function returns true if it can't). We are in fact applying
449 * one of the CBS rules: when a task wakes up, if the residual runtime
450 * over residual deadline fits within the allocated bandwidth, then we
451 * can keep the current (absolute) deadline and residual budget without
452 * disrupting the schedulability of the system. Otherwise, we should
453 * refill the runtime and set the deadline a period in the future,
454 * because keeping the current (absolute) deadline of the task would
455 * result in breaking guarantees promised to other tasks (refer to
456 * Documentation/scheduler/sched-deadline.txt for more informations).
457 *
458 * This function returns true if:
459 *
460 * runtime / (deadline - t) > dl_runtime / dl_period ,
461 *
462 * IOW we can't recycle current parameters.
463 *
464 * Notice that the bandwidth check is done against the period. For
465 * task with deadline equal to period this is the same of using
466 * dl_deadline instead of dl_period in the equation above.
467 */
468static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
469 struct sched_dl_entity *pi_se, u64 t)
470{
471 u64 left, right;
472
473 /*
474 * left and right are the two sides of the equation above,
475 * after a bit of shuffling to use multiplications instead
476 * of divisions.
477 *
478 * Note that none of the time values involved in the two
479 * multiplications are absolute: dl_deadline and dl_runtime
480 * are the relative deadline and the maximum runtime of each
481 * instance, runtime is the runtime left for the last instance
482 * and (deadline - t), since t is rq->clock, is the time left
483 * to the (absolute) deadline. Even if overflowing the u64 type
484 * is very unlikely to occur in both cases, here we scale down
485 * as we want to avoid that risk at all. Scaling down by 10
486 * means that we reduce granularity to 1us. We are fine with it,
487 * since this is only a true/false check and, anyway, thinking
488 * of anything below microseconds resolution is actually fiction
489 * (but still we want to give the user that illusion >;).
490 */
491 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
492 right = ((dl_se->deadline - t) >> DL_SCALE) *
493 (pi_se->dl_runtime >> DL_SCALE);
494
495 return dl_time_before(right, left);
496}
497
498/*
499 * When a -deadline entity is queued back on the runqueue, its runtime and
500 * deadline might need updating.
501 *
502 * The policy here is that we update the deadline of the entity only if:
503 * - the current deadline is in the past,
504 * - using the remaining runtime with the current deadline would make
505 * the entity exceed its bandwidth.
506 */
507static void update_dl_entity(struct sched_dl_entity *dl_se,
508 struct sched_dl_entity *pi_se)
509{
510 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
511 struct rq *rq = rq_of_dl_rq(dl_rq);
512
513 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
514 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
515 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
516 dl_se->runtime = pi_se->dl_runtime;
517 }
518}
519
520/*
521 * If the entity depleted all its runtime, and if we want it to sleep
522 * while waiting for some new execution time to become available, we
523 * set the bandwidth enforcement timer to the replenishment instant
524 * and try to activate it.
525 *
526 * Notice that it is important for the caller to know if the timer
527 * actually started or not (i.e., the replenishment instant is in
528 * the future or in the past).
529 */
530static int start_dl_timer(struct task_struct *p)
531{
532 struct sched_dl_entity *dl_se = &p->dl;
533 struct hrtimer *timer = &dl_se->dl_timer;
534 struct rq *rq = task_rq(p);
535 ktime_t now, act;
536 s64 delta;
537
538 lockdep_assert_held(&rq->lock);
539
540 /*
541 * We want the timer to fire at the deadline, but considering
542 * that it is actually coming from rq->clock and not from
543 * hrtimer's time base reading.
544 */
545 act = ns_to_ktime(dl_se->deadline);
546 now = hrtimer_cb_get_time(timer);
547 delta = ktime_to_ns(now) - rq_clock(rq);
548 act = ktime_add_ns(act, delta);
549
550 /*
551 * If the expiry time already passed, e.g., because the value
552 * chosen as the deadline is too small, don't even try to
553 * start the timer in the past!
554 */
555 if (ktime_us_delta(act, now) < 0)
556 return 0;
557
558 /*
559 * !enqueued will guarantee another callback; even if one is already in
560 * progress. This ensures a balanced {get,put}_task_struct().
561 *
562 * The race against __run_timer() clearing the enqueued state is
563 * harmless because we're holding task_rq()->lock, therefore the timer
564 * expiring after we've done the check will wait on its task_rq_lock()
565 * and observe our state.
566 */
567 if (!hrtimer_is_queued(timer)) {
568 get_task_struct(p);
569 hrtimer_start(timer, act, HRTIMER_MODE_ABS);
570 }
571
572 return 1;
573}
574
575/*
576 * This is the bandwidth enforcement timer callback. If here, we know
577 * a task is not on its dl_rq, since the fact that the timer was running
578 * means the task is throttled and needs a runtime replenishment.
579 *
580 * However, what we actually do depends on the fact the task is active,
581 * (it is on its rq) or has been removed from there by a call to
582 * dequeue_task_dl(). In the former case we must issue the runtime
583 * replenishment and add the task back to the dl_rq; in the latter, we just
584 * do nothing but clearing dl_throttled, so that runtime and deadline
585 * updating (and the queueing back to dl_rq) will be done by the
586 * next call to enqueue_task_dl().
587 */
588static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
589{
590 struct sched_dl_entity *dl_se = container_of(timer,
591 struct sched_dl_entity,
592 dl_timer);
593 struct task_struct *p = dl_task_of(dl_se);
594 unsigned long flags;
595 struct rq *rq;
596
597 rq = task_rq_lock(p, &flags);
598
599 /*
600 * The task might have changed its scheduling policy to something
601 * different than SCHED_DEADLINE (through switched_fromd_dl()).
602 */
603 if (!dl_task(p)) {
604 __dl_clear_params(p);
605 goto unlock;
606 }
607
608 /*
609 * The task might have been boosted by someone else and might be in the
610 * boosting/deboosting path, its not throttled.
611 */
612 if (dl_se->dl_boosted)
613 goto unlock;
614
615 /*
616 * Spurious timer due to start_dl_timer() race; or we already received
617 * a replenishment from rt_mutex_setprio().
618 */
619 if (!dl_se->dl_throttled)
620 goto unlock;
621
622 sched_clock_tick();
623 update_rq_clock(rq);
624
625 /*
626 * If the throttle happened during sched-out; like:
627 *
628 * schedule()
629 * deactivate_task()
630 * dequeue_task_dl()
631 * update_curr_dl()
632 * start_dl_timer()
633 * __dequeue_task_dl()
634 * prev->on_rq = 0;
635 *
636 * We can be both throttled and !queued. Replenish the counter
637 * but do not enqueue -- wait for our wakeup to do that.
638 */
639 if (!task_on_rq_queued(p)) {
640 replenish_dl_entity(dl_se, dl_se);
641 goto unlock;
642 }
643
644 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
645 if (dl_task(rq->curr))
646 check_preempt_curr_dl(rq, p, 0);
647 else
648 resched_curr(rq);
649
650#ifdef CONFIG_SMP
651 /*
652 * Perform balancing operations here; after the replenishments. We
653 * cannot drop rq->lock before this, otherwise the assertion in
654 * start_dl_timer() about not missing updates is not true.
655 *
656 * If we find that the rq the task was on is no longer available, we
657 * need to select a new rq.
658 *
659 * XXX figure out if select_task_rq_dl() deals with offline cpus.
660 */
661 if (unlikely(!rq->online))
662 rq = dl_task_offline_migration(rq, p);
663
664 /*
665 * Queueing this task back might have overloaded rq, check if we need
666 * to kick someone away.
667 */
668 if (has_pushable_dl_tasks(rq)) {
669 /*
670 * Nothing relies on rq->lock after this, so its safe to drop
671 * rq->lock.
672 */
673 lockdep_unpin_lock(&rq->lock);
674 push_dl_task(rq);
675 lockdep_pin_lock(&rq->lock);
676 }
677#endif
678
679unlock:
680 task_rq_unlock(rq, p, &flags);
681
682 /*
683 * This can free the task_struct, including this hrtimer, do not touch
684 * anything related to that after this.
685 */
686 put_task_struct(p);
687
688 return HRTIMER_NORESTART;
689}
690
691void init_dl_task_timer(struct sched_dl_entity *dl_se)
692{
693 struct hrtimer *timer = &dl_se->dl_timer;
694
695 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
696 timer->function = dl_task_timer;
697}
698
699static
700int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
701{
702 return (dl_se->runtime <= 0);
703}
704
705extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
706
707/*
708 * Update the current task's runtime statistics (provided it is still
709 * a -deadline task and has not been removed from the dl_rq).
710 */
711static void update_curr_dl(struct rq *rq)
712{
713 struct task_struct *curr = rq->curr;
714 struct sched_dl_entity *dl_se = &curr->dl;
715 u64 delta_exec;
716
717 if (!dl_task(curr) || !on_dl_rq(dl_se))
718 return;
719
720 /* Kick cpufreq (see the comment in linux/cpufreq.h). */
721 if (cpu_of(rq) == smp_processor_id())
722 cpufreq_trigger_update(rq_clock(rq));
723
724 /*
725 * Consumed budget is computed considering the time as
726 * observed by schedulable tasks (excluding time spent
727 * in hardirq context, etc.). Deadlines are instead
728 * computed using hard walltime. This seems to be the more
729 * natural solution, but the full ramifications of this
730 * approach need further study.
731 */
732 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
733 if (unlikely((s64)delta_exec <= 0)) {
734 if (unlikely(dl_se->dl_yielded))
735 goto throttle;
736 return;
737 }
738
739 schedstat_set(curr->se.statistics.exec_max,
740 max(curr->se.statistics.exec_max, delta_exec));
741
742 curr->se.sum_exec_runtime += delta_exec;
743 account_group_exec_runtime(curr, delta_exec);
744
745 curr->se.exec_start = rq_clock_task(rq);
746 cpuacct_charge(curr, delta_exec);
747
748 sched_rt_avg_update(rq, delta_exec);
749
750 dl_se->runtime -= delta_exec;
751
752throttle:
753 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
754 dl_se->dl_throttled = 1;
755 __dequeue_task_dl(rq, curr, 0);
756 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
757 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
758
759 if (!is_leftmost(curr, &rq->dl))
760 resched_curr(rq);
761 }
762
763 /*
764 * Because -- for now -- we share the rt bandwidth, we need to
765 * account our runtime there too, otherwise actual rt tasks
766 * would be able to exceed the shared quota.
767 *
768 * Account to the root rt group for now.
769 *
770 * The solution we're working towards is having the RT groups scheduled
771 * using deadline servers -- however there's a few nasties to figure
772 * out before that can happen.
773 */
774 if (rt_bandwidth_enabled()) {
775 struct rt_rq *rt_rq = &rq->rt;
776
777 raw_spin_lock(&rt_rq->rt_runtime_lock);
778 /*
779 * We'll let actual RT tasks worry about the overflow here, we
780 * have our own CBS to keep us inline; only account when RT
781 * bandwidth is relevant.
782 */
783 if (sched_rt_bandwidth_account(rt_rq))
784 rt_rq->rt_time += delta_exec;
785 raw_spin_unlock(&rt_rq->rt_runtime_lock);
786 }
787}
788
789#ifdef CONFIG_SMP
790
791static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
792{
793 struct rq *rq = rq_of_dl_rq(dl_rq);
794
795 if (dl_rq->earliest_dl.curr == 0 ||
796 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
797 dl_rq->earliest_dl.curr = deadline;
798 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1);
799 }
800}
801
802static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
803{
804 struct rq *rq = rq_of_dl_rq(dl_rq);
805
806 /*
807 * Since we may have removed our earliest (and/or next earliest)
808 * task we must recompute them.
809 */
810 if (!dl_rq->dl_nr_running) {
811 dl_rq->earliest_dl.curr = 0;
812 dl_rq->earliest_dl.next = 0;
813 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
814 } else {
815 struct rb_node *leftmost = dl_rq->rb_leftmost;
816 struct sched_dl_entity *entry;
817
818 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
819 dl_rq->earliest_dl.curr = entry->deadline;
820 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1);
821 }
822}
823
824#else
825
826static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
827static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
828
829#endif /* CONFIG_SMP */
830
831static inline
832void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
833{
834 int prio = dl_task_of(dl_se)->prio;
835 u64 deadline = dl_se->deadline;
836
837 WARN_ON(!dl_prio(prio));
838 dl_rq->dl_nr_running++;
839 add_nr_running(rq_of_dl_rq(dl_rq), 1);
840
841 inc_dl_deadline(dl_rq, deadline);
842 inc_dl_migration(dl_se, dl_rq);
843}
844
845static inline
846void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
847{
848 int prio = dl_task_of(dl_se)->prio;
849
850 WARN_ON(!dl_prio(prio));
851 WARN_ON(!dl_rq->dl_nr_running);
852 dl_rq->dl_nr_running--;
853 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
854
855 dec_dl_deadline(dl_rq, dl_se->deadline);
856 dec_dl_migration(dl_se, dl_rq);
857}
858
859static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
860{
861 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
862 struct rb_node **link = &dl_rq->rb_root.rb_node;
863 struct rb_node *parent = NULL;
864 struct sched_dl_entity *entry;
865 int leftmost = 1;
866
867 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
868
869 while (*link) {
870 parent = *link;
871 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
872 if (dl_time_before(dl_se->deadline, entry->deadline))
873 link = &parent->rb_left;
874 else {
875 link = &parent->rb_right;
876 leftmost = 0;
877 }
878 }
879
880 if (leftmost)
881 dl_rq->rb_leftmost = &dl_se->rb_node;
882
883 rb_link_node(&dl_se->rb_node, parent, link);
884 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
885
886 inc_dl_tasks(dl_se, dl_rq);
887}
888
889static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
890{
891 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
892
893 if (RB_EMPTY_NODE(&dl_se->rb_node))
894 return;
895
896 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
897 struct rb_node *next_node;
898
899 next_node = rb_next(&dl_se->rb_node);
900 dl_rq->rb_leftmost = next_node;
901 }
902
903 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
904 RB_CLEAR_NODE(&dl_se->rb_node);
905
906 dec_dl_tasks(dl_se, dl_rq);
907}
908
909static void
910enqueue_dl_entity(struct sched_dl_entity *dl_se,
911 struct sched_dl_entity *pi_se, int flags)
912{
913 BUG_ON(on_dl_rq(dl_se));
914
915 /*
916 * If this is a wakeup or a new instance, the scheduling
917 * parameters of the task might need updating. Otherwise,
918 * we want a replenishment of its runtime.
919 */
920 if (flags & ENQUEUE_WAKEUP)
921 update_dl_entity(dl_se, pi_se);
922 else if (flags & ENQUEUE_REPLENISH)
923 replenish_dl_entity(dl_se, pi_se);
924
925 __enqueue_dl_entity(dl_se);
926}
927
928static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
929{
930 __dequeue_dl_entity(dl_se);
931}
932
933static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
934{
935 struct task_struct *pi_task = rt_mutex_get_top_task(p);
936 struct sched_dl_entity *pi_se = &p->dl;
937
938 /*
939 * Use the scheduling parameters of the top pi-waiter
940 * task if we have one and its (absolute) deadline is
941 * smaller than our one... OTW we keep our runtime and
942 * deadline.
943 */
944 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) {
945 pi_se = &pi_task->dl;
946 } else if (!dl_prio(p->normal_prio)) {
947 /*
948 * Special case in which we have a !SCHED_DEADLINE task
949 * that is going to be deboosted, but exceedes its
950 * runtime while doing so. No point in replenishing
951 * it, as it's going to return back to its original
952 * scheduling class after this.
953 */
954 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
955 return;
956 }
957
958 /*
959 * If p is throttled, we do nothing. In fact, if it exhausted
960 * its budget it needs a replenishment and, since it now is on
961 * its rq, the bandwidth timer callback (which clearly has not
962 * run yet) will take care of this.
963 */
964 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH))
965 return;
966
967 enqueue_dl_entity(&p->dl, pi_se, flags);
968
969 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
970 enqueue_pushable_dl_task(rq, p);
971}
972
973static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
974{
975 dequeue_dl_entity(&p->dl);
976 dequeue_pushable_dl_task(rq, p);
977}
978
979static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
980{
981 update_curr_dl(rq);
982 __dequeue_task_dl(rq, p, flags);
983}
984
985/*
986 * Yield task semantic for -deadline tasks is:
987 *
988 * get off from the CPU until our next instance, with
989 * a new runtime. This is of little use now, since we
990 * don't have a bandwidth reclaiming mechanism. Anyway,
991 * bandwidth reclaiming is planned for the future, and
992 * yield_task_dl will indicate that some spare budget
993 * is available for other task instances to use it.
994 */
995static void yield_task_dl(struct rq *rq)
996{
997 /*
998 * We make the task go to sleep until its current deadline by
999 * forcing its runtime to zero. This way, update_curr_dl() stops
1000 * it and the bandwidth timer will wake it up and will give it
1001 * new scheduling parameters (thanks to dl_yielded=1).
1002 */
1003 rq->curr->dl.dl_yielded = 1;
1004
1005 update_rq_clock(rq);
1006 update_curr_dl(rq);
1007 /*
1008 * Tell update_rq_clock() that we've just updated,
1009 * so we don't do microscopic update in schedule()
1010 * and double the fastpath cost.
1011 */
1012 rq_clock_skip_update(rq, true);
1013}
1014
1015#ifdef CONFIG_SMP
1016
1017static int find_later_rq(struct task_struct *task);
1018
1019static int
1020select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1021{
1022 struct task_struct *curr;
1023 struct rq *rq;
1024
1025 if (sd_flag != SD_BALANCE_WAKE)
1026 goto out;
1027
1028 rq = cpu_rq(cpu);
1029
1030 rcu_read_lock();
1031 curr = READ_ONCE(rq->curr); /* unlocked access */
1032
1033 /*
1034 * If we are dealing with a -deadline task, we must
1035 * decide where to wake it up.
1036 * If it has a later deadline and the current task
1037 * on this rq can't move (provided the waking task
1038 * can!) we prefer to send it somewhere else. On the
1039 * other hand, if it has a shorter deadline, we
1040 * try to make it stay here, it might be important.
1041 */
1042 if (unlikely(dl_task(curr)) &&
1043 (curr->nr_cpus_allowed < 2 ||
1044 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1045 (p->nr_cpus_allowed > 1)) {
1046 int target = find_later_rq(p);
1047
1048 if (target != -1 &&
1049 (dl_time_before(p->dl.deadline,
1050 cpu_rq(target)->dl.earliest_dl.curr) ||
1051 (cpu_rq(target)->dl.dl_nr_running == 0)))
1052 cpu = target;
1053 }
1054 rcu_read_unlock();
1055
1056out:
1057 return cpu;
1058}
1059
1060static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1061{
1062 /*
1063 * Current can't be migrated, useless to reschedule,
1064 * let's hope p can move out.
1065 */
1066 if (rq->curr->nr_cpus_allowed == 1 ||
1067 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
1068 return;
1069
1070 /*
1071 * p is migratable, so let's not schedule it and
1072 * see if it is pushed or pulled somewhere else.
1073 */
1074 if (p->nr_cpus_allowed != 1 &&
1075 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
1076 return;
1077
1078 resched_curr(rq);
1079}
1080
1081#endif /* CONFIG_SMP */
1082
1083/*
1084 * Only called when both the current and waking task are -deadline
1085 * tasks.
1086 */
1087static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1088 int flags)
1089{
1090 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1091 resched_curr(rq);
1092 return;
1093 }
1094
1095#ifdef CONFIG_SMP
1096 /*
1097 * In the unlikely case current and p have the same deadline
1098 * let us try to decide what's the best thing to do...
1099 */
1100 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1101 !test_tsk_need_resched(rq->curr))
1102 check_preempt_equal_dl(rq, p);
1103#endif /* CONFIG_SMP */
1104}
1105
1106#ifdef CONFIG_SCHED_HRTICK
1107static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1108{
1109 hrtick_start(rq, p->dl.runtime);
1110}
1111#else /* !CONFIG_SCHED_HRTICK */
1112static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1113{
1114}
1115#endif
1116
1117static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1118 struct dl_rq *dl_rq)
1119{
1120 struct rb_node *left = dl_rq->rb_leftmost;
1121
1122 if (!left)
1123 return NULL;
1124
1125 return rb_entry(left, struct sched_dl_entity, rb_node);
1126}
1127
1128struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev)
1129{
1130 struct sched_dl_entity *dl_se;
1131 struct task_struct *p;
1132 struct dl_rq *dl_rq;
1133
1134 dl_rq = &rq->dl;
1135
1136 if (need_pull_dl_task(rq, prev)) {
1137 /*
1138 * This is OK, because current is on_cpu, which avoids it being
1139 * picked for load-balance and preemption/IRQs are still
1140 * disabled avoiding further scheduler activity on it and we're
1141 * being very careful to re-start the picking loop.
1142 */
1143 lockdep_unpin_lock(&rq->lock);
1144 pull_dl_task(rq);
1145 lockdep_pin_lock(&rq->lock);
1146 /*
1147 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1148 * means a stop task can slip in, in which case we need to
1149 * re-start task selection.
1150 */
1151 if (rq->stop && task_on_rq_queued(rq->stop))
1152 return RETRY_TASK;
1153 }
1154
1155 /*
1156 * When prev is DL, we may throttle it in put_prev_task().
1157 * So, we update time before we check for dl_nr_running.
1158 */
1159 if (prev->sched_class == &dl_sched_class)
1160 update_curr_dl(rq);
1161
1162 if (unlikely(!dl_rq->dl_nr_running))
1163 return NULL;
1164
1165 put_prev_task(rq, prev);
1166
1167 dl_se = pick_next_dl_entity(rq, dl_rq);
1168 BUG_ON(!dl_se);
1169
1170 p = dl_task_of(dl_se);
1171 p->se.exec_start = rq_clock_task(rq);
1172
1173 /* Running task will never be pushed. */
1174 dequeue_pushable_dl_task(rq, p);
1175
1176 if (hrtick_enabled(rq))
1177 start_hrtick_dl(rq, p);
1178
1179 queue_push_tasks(rq);
1180
1181 return p;
1182}
1183
1184static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1185{
1186 update_curr_dl(rq);
1187
1188 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1189 enqueue_pushable_dl_task(rq, p);
1190}
1191
1192static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1193{
1194 update_curr_dl(rq);
1195
1196 /*
1197 * Even when we have runtime, update_curr_dl() might have resulted in us
1198 * not being the leftmost task anymore. In that case NEED_RESCHED will
1199 * be set and schedule() will start a new hrtick for the next task.
1200 */
1201 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1202 is_leftmost(p, &rq->dl))
1203 start_hrtick_dl(rq, p);
1204}
1205
1206static void task_fork_dl(struct task_struct *p)
1207{
1208 /*
1209 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1210 * sched_fork()
1211 */
1212}
1213
1214static void task_dead_dl(struct task_struct *p)
1215{
1216 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1217
1218 /*
1219 * Since we are TASK_DEAD we won't slip out of the domain!
1220 */
1221 raw_spin_lock_irq(&dl_b->lock);
1222 /* XXX we should retain the bw until 0-lag */
1223 dl_b->total_bw -= p->dl.dl_bw;
1224 raw_spin_unlock_irq(&dl_b->lock);
1225}
1226
1227static void set_curr_task_dl(struct rq *rq)
1228{
1229 struct task_struct *p = rq->curr;
1230
1231 p->se.exec_start = rq_clock_task(rq);
1232
1233 /* You can't push away the running task */
1234 dequeue_pushable_dl_task(rq, p);
1235}
1236
1237#ifdef CONFIG_SMP
1238
1239/* Only try algorithms three times */
1240#define DL_MAX_TRIES 3
1241
1242static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1243{
1244 if (!task_running(rq, p) &&
1245 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1246 return 1;
1247 return 0;
1248}
1249
1250/*
1251 * Return the earliest pushable rq's task, which is suitable to be executed
1252 * on the CPU, NULL otherwise:
1253 */
1254static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1255{
1256 struct rb_node *next_node = rq->dl.pushable_dl_tasks_leftmost;
1257 struct task_struct *p = NULL;
1258
1259 if (!has_pushable_dl_tasks(rq))
1260 return NULL;
1261
1262next_node:
1263 if (next_node) {
1264 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1265
1266 if (pick_dl_task(rq, p, cpu))
1267 return p;
1268
1269 next_node = rb_next(next_node);
1270 goto next_node;
1271 }
1272
1273 return NULL;
1274}
1275
1276static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1277
1278static int find_later_rq(struct task_struct *task)
1279{
1280 struct sched_domain *sd;
1281 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1282 int this_cpu = smp_processor_id();
1283 int best_cpu, cpu = task_cpu(task);
1284
1285 /* Make sure the mask is initialized first */
1286 if (unlikely(!later_mask))
1287 return -1;
1288
1289 if (task->nr_cpus_allowed == 1)
1290 return -1;
1291
1292 /*
1293 * We have to consider system topology and task affinity
1294 * first, then we can look for a suitable cpu.
1295 */
1296 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
1297 task, later_mask);
1298 if (best_cpu == -1)
1299 return -1;
1300
1301 /*
1302 * If we are here, some target has been found,
1303 * the most suitable of which is cached in best_cpu.
1304 * This is, among the runqueues where the current tasks
1305 * have later deadlines than the task's one, the rq
1306 * with the latest possible one.
1307 *
1308 * Now we check how well this matches with task's
1309 * affinity and system topology.
1310 *
1311 * The last cpu where the task run is our first
1312 * guess, since it is most likely cache-hot there.
1313 */
1314 if (cpumask_test_cpu(cpu, later_mask))
1315 return cpu;
1316 /*
1317 * Check if this_cpu is to be skipped (i.e., it is
1318 * not in the mask) or not.
1319 */
1320 if (!cpumask_test_cpu(this_cpu, later_mask))
1321 this_cpu = -1;
1322
1323 rcu_read_lock();
1324 for_each_domain(cpu, sd) {
1325 if (sd->flags & SD_WAKE_AFFINE) {
1326
1327 /*
1328 * If possible, preempting this_cpu is
1329 * cheaper than migrating.
1330 */
1331 if (this_cpu != -1 &&
1332 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1333 rcu_read_unlock();
1334 return this_cpu;
1335 }
1336
1337 /*
1338 * Last chance: if best_cpu is valid and is
1339 * in the mask, that becomes our choice.
1340 */
1341 if (best_cpu < nr_cpu_ids &&
1342 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
1343 rcu_read_unlock();
1344 return best_cpu;
1345 }
1346 }
1347 }
1348 rcu_read_unlock();
1349
1350 /*
1351 * At this point, all our guesses failed, we just return
1352 * 'something', and let the caller sort the things out.
1353 */
1354 if (this_cpu != -1)
1355 return this_cpu;
1356
1357 cpu = cpumask_any(later_mask);
1358 if (cpu < nr_cpu_ids)
1359 return cpu;
1360
1361 return -1;
1362}
1363
1364/* Locks the rq it finds */
1365static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1366{
1367 struct rq *later_rq = NULL;
1368 int tries;
1369 int cpu;
1370
1371 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1372 cpu = find_later_rq(task);
1373
1374 if ((cpu == -1) || (cpu == rq->cpu))
1375 break;
1376
1377 later_rq = cpu_rq(cpu);
1378
1379 if (later_rq->dl.dl_nr_running &&
1380 !dl_time_before(task->dl.deadline,
1381 later_rq->dl.earliest_dl.curr)) {
1382 /*
1383 * Target rq has tasks of equal or earlier deadline,
1384 * retrying does not release any lock and is unlikely
1385 * to yield a different result.
1386 */
1387 later_rq = NULL;
1388 break;
1389 }
1390
1391 /* Retry if something changed. */
1392 if (double_lock_balance(rq, later_rq)) {
1393 if (unlikely(task_rq(task) != rq ||
1394 !cpumask_test_cpu(later_rq->cpu,
1395 &task->cpus_allowed) ||
1396 task_running(rq, task) ||
1397 !dl_task(task) ||
1398 !task_on_rq_queued(task))) {
1399 double_unlock_balance(rq, later_rq);
1400 later_rq = NULL;
1401 break;
1402 }
1403 }
1404
1405 /*
1406 * If the rq we found has no -deadline task, or
1407 * its earliest one has a later deadline than our
1408 * task, the rq is a good one.
1409 */
1410 if (!later_rq->dl.dl_nr_running ||
1411 dl_time_before(task->dl.deadline,
1412 later_rq->dl.earliest_dl.curr))
1413 break;
1414
1415 /* Otherwise we try again. */
1416 double_unlock_balance(rq, later_rq);
1417 later_rq = NULL;
1418 }
1419
1420 return later_rq;
1421}
1422
1423static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1424{
1425 struct task_struct *p;
1426
1427 if (!has_pushable_dl_tasks(rq))
1428 return NULL;
1429
1430 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
1431 struct task_struct, pushable_dl_tasks);
1432
1433 BUG_ON(rq->cpu != task_cpu(p));
1434 BUG_ON(task_current(rq, p));
1435 BUG_ON(p->nr_cpus_allowed <= 1);
1436
1437 BUG_ON(!task_on_rq_queued(p));
1438 BUG_ON(!dl_task(p));
1439
1440 return p;
1441}
1442
1443/*
1444 * See if the non running -deadline tasks on this rq
1445 * can be sent to some other CPU where they can preempt
1446 * and start executing.
1447 */
1448static int push_dl_task(struct rq *rq)
1449{
1450 struct task_struct *next_task;
1451 struct rq *later_rq;
1452 int ret = 0;
1453
1454 if (!rq->dl.overloaded)
1455 return 0;
1456
1457 next_task = pick_next_pushable_dl_task(rq);
1458 if (!next_task)
1459 return 0;
1460
1461retry:
1462 if (unlikely(next_task == rq->curr)) {
1463 WARN_ON(1);
1464 return 0;
1465 }
1466
1467 /*
1468 * If next_task preempts rq->curr, and rq->curr
1469 * can move away, it makes sense to just reschedule
1470 * without going further in pushing next_task.
1471 */
1472 if (dl_task(rq->curr) &&
1473 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1474 rq->curr->nr_cpus_allowed > 1) {
1475 resched_curr(rq);
1476 return 0;
1477 }
1478
1479 /* We might release rq lock */
1480 get_task_struct(next_task);
1481
1482 /* Will lock the rq it'll find */
1483 later_rq = find_lock_later_rq(next_task, rq);
1484 if (!later_rq) {
1485 struct task_struct *task;
1486
1487 /*
1488 * We must check all this again, since
1489 * find_lock_later_rq releases rq->lock and it is
1490 * then possible that next_task has migrated.
1491 */
1492 task = pick_next_pushable_dl_task(rq);
1493 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1494 /*
1495 * The task is still there. We don't try
1496 * again, some other cpu will pull it when ready.
1497 */
1498 goto out;
1499 }
1500
1501 if (!task)
1502 /* No more tasks */
1503 goto out;
1504
1505 put_task_struct(next_task);
1506 next_task = task;
1507 goto retry;
1508 }
1509
1510 deactivate_task(rq, next_task, 0);
1511 set_task_cpu(next_task, later_rq->cpu);
1512 activate_task(later_rq, next_task, 0);
1513 ret = 1;
1514
1515 resched_curr(later_rq);
1516
1517 double_unlock_balance(rq, later_rq);
1518
1519out:
1520 put_task_struct(next_task);
1521
1522 return ret;
1523}
1524
1525static void push_dl_tasks(struct rq *rq)
1526{
1527 /* push_dl_task() will return true if it moved a -deadline task */
1528 while (push_dl_task(rq))
1529 ;
1530}
1531
1532static void pull_dl_task(struct rq *this_rq)
1533{
1534 int this_cpu = this_rq->cpu, cpu;
1535 struct task_struct *p;
1536 bool resched = false;
1537 struct rq *src_rq;
1538 u64 dmin = LONG_MAX;
1539
1540 if (likely(!dl_overloaded(this_rq)))
1541 return;
1542
1543 /*
1544 * Match the barrier from dl_set_overloaded; this guarantees that if we
1545 * see overloaded we must also see the dlo_mask bit.
1546 */
1547 smp_rmb();
1548
1549 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
1550 if (this_cpu == cpu)
1551 continue;
1552
1553 src_rq = cpu_rq(cpu);
1554
1555 /*
1556 * It looks racy, abd it is! However, as in sched_rt.c,
1557 * we are fine with this.
1558 */
1559 if (this_rq->dl.dl_nr_running &&
1560 dl_time_before(this_rq->dl.earliest_dl.curr,
1561 src_rq->dl.earliest_dl.next))
1562 continue;
1563
1564 /* Might drop this_rq->lock */
1565 double_lock_balance(this_rq, src_rq);
1566
1567 /*
1568 * If there are no more pullable tasks on the
1569 * rq, we're done with it.
1570 */
1571 if (src_rq->dl.dl_nr_running <= 1)
1572 goto skip;
1573
1574 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
1575
1576 /*
1577 * We found a task to be pulled if:
1578 * - it preempts our current (if there's one),
1579 * - it will preempt the last one we pulled (if any).
1580 */
1581 if (p && dl_time_before(p->dl.deadline, dmin) &&
1582 (!this_rq->dl.dl_nr_running ||
1583 dl_time_before(p->dl.deadline,
1584 this_rq->dl.earliest_dl.curr))) {
1585 WARN_ON(p == src_rq->curr);
1586 WARN_ON(!task_on_rq_queued(p));
1587
1588 /*
1589 * Then we pull iff p has actually an earlier
1590 * deadline than the current task of its runqueue.
1591 */
1592 if (dl_time_before(p->dl.deadline,
1593 src_rq->curr->dl.deadline))
1594 goto skip;
1595
1596 resched = true;
1597
1598 deactivate_task(src_rq, p, 0);
1599 set_task_cpu(p, this_cpu);
1600 activate_task(this_rq, p, 0);
1601 dmin = p->dl.deadline;
1602
1603 /* Is there any other task even earlier? */
1604 }
1605skip:
1606 double_unlock_balance(this_rq, src_rq);
1607 }
1608
1609 if (resched)
1610 resched_curr(this_rq);
1611}
1612
1613/*
1614 * Since the task is not running and a reschedule is not going to happen
1615 * anytime soon on its runqueue, we try pushing it away now.
1616 */
1617static void task_woken_dl(struct rq *rq, struct task_struct *p)
1618{
1619 if (!task_running(rq, p) &&
1620 !test_tsk_need_resched(rq->curr) &&
1621 p->nr_cpus_allowed > 1 &&
1622 dl_task(rq->curr) &&
1623 (rq->curr->nr_cpus_allowed < 2 ||
1624 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
1625 push_dl_tasks(rq);
1626 }
1627}
1628
1629static void set_cpus_allowed_dl(struct task_struct *p,
1630 const struct cpumask *new_mask)
1631{
1632 struct root_domain *src_rd;
1633 struct rq *rq;
1634
1635 BUG_ON(!dl_task(p));
1636
1637 rq = task_rq(p);
1638 src_rd = rq->rd;
1639 /*
1640 * Migrating a SCHED_DEADLINE task between exclusive
1641 * cpusets (different root_domains) entails a bandwidth
1642 * update. We already made space for us in the destination
1643 * domain (see cpuset_can_attach()).
1644 */
1645 if (!cpumask_intersects(src_rd->span, new_mask)) {
1646 struct dl_bw *src_dl_b;
1647
1648 src_dl_b = dl_bw_of(cpu_of(rq));
1649 /*
1650 * We now free resources of the root_domain we are migrating
1651 * off. In the worst case, sched_setattr() may temporary fail
1652 * until we complete the update.
1653 */
1654 raw_spin_lock(&src_dl_b->lock);
1655 __dl_clear(src_dl_b, p->dl.dl_bw);
1656 raw_spin_unlock(&src_dl_b->lock);
1657 }
1658
1659 set_cpus_allowed_common(p, new_mask);
1660}
1661
1662/* Assumes rq->lock is held */
1663static void rq_online_dl(struct rq *rq)
1664{
1665 if (rq->dl.overloaded)
1666 dl_set_overload(rq);
1667
1668 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
1669 if (rq->dl.dl_nr_running > 0)
1670 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1);
1671}
1672
1673/* Assumes rq->lock is held */
1674static void rq_offline_dl(struct rq *rq)
1675{
1676 if (rq->dl.overloaded)
1677 dl_clear_overload(rq);
1678
1679 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
1680 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
1681}
1682
1683void __init init_sched_dl_class(void)
1684{
1685 unsigned int i;
1686
1687 for_each_possible_cpu(i)
1688 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
1689 GFP_KERNEL, cpu_to_node(i));
1690}
1691
1692#endif /* CONFIG_SMP */
1693
1694static void switched_from_dl(struct rq *rq, struct task_struct *p)
1695{
1696 /*
1697 * Start the deadline timer; if we switch back to dl before this we'll
1698 * continue consuming our current CBS slice. If we stay outside of
1699 * SCHED_DEADLINE until the deadline passes, the timer will reset the
1700 * task.
1701 */
1702 if (!start_dl_timer(p))
1703 __dl_clear_params(p);
1704
1705 /*
1706 * Since this might be the only -deadline task on the rq,
1707 * this is the right place to try to pull some other one
1708 * from an overloaded cpu, if any.
1709 */
1710 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
1711 return;
1712
1713 queue_pull_task(rq);
1714}
1715
1716/*
1717 * When switching to -deadline, we may overload the rq, then
1718 * we try to push someone off, if possible.
1719 */
1720static void switched_to_dl(struct rq *rq, struct task_struct *p)
1721{
1722 if (dl_time_before(p->dl.deadline, rq_clock(rq)))
1723 setup_new_dl_entity(&p->dl, &p->dl);
1724
1725 if (task_on_rq_queued(p) && rq->curr != p) {
1726#ifdef CONFIG_SMP
1727 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
1728 queue_push_tasks(rq);
1729#else
1730 if (dl_task(rq->curr))
1731 check_preempt_curr_dl(rq, p, 0);
1732 else
1733 resched_curr(rq);
1734#endif
1735 }
1736}
1737
1738/*
1739 * If the scheduling parameters of a -deadline task changed,
1740 * a push or pull operation might be needed.
1741 */
1742static void prio_changed_dl(struct rq *rq, struct task_struct *p,
1743 int oldprio)
1744{
1745 if (task_on_rq_queued(p) || rq->curr == p) {
1746#ifdef CONFIG_SMP
1747 /*
1748 * This might be too much, but unfortunately
1749 * we don't have the old deadline value, and
1750 * we can't argue if the task is increasing
1751 * or lowering its prio, so...
1752 */
1753 if (!rq->dl.overloaded)
1754 queue_pull_task(rq);
1755
1756 /*
1757 * If we now have a earlier deadline task than p,
1758 * then reschedule, provided p is still on this
1759 * runqueue.
1760 */
1761 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
1762 resched_curr(rq);
1763#else
1764 /*
1765 * Again, we don't know if p has a earlier
1766 * or later deadline, so let's blindly set a
1767 * (maybe not needed) rescheduling point.
1768 */
1769 resched_curr(rq);
1770#endif /* CONFIG_SMP */
1771 }
1772}
1773
1774const struct sched_class dl_sched_class = {
1775 .next = &rt_sched_class,
1776 .enqueue_task = enqueue_task_dl,
1777 .dequeue_task = dequeue_task_dl,
1778 .yield_task = yield_task_dl,
1779
1780 .check_preempt_curr = check_preempt_curr_dl,
1781
1782 .pick_next_task = pick_next_task_dl,
1783 .put_prev_task = put_prev_task_dl,
1784
1785#ifdef CONFIG_SMP
1786 .select_task_rq = select_task_rq_dl,
1787 .set_cpus_allowed = set_cpus_allowed_dl,
1788 .rq_online = rq_online_dl,
1789 .rq_offline = rq_offline_dl,
1790 .task_woken = task_woken_dl,
1791#endif
1792
1793 .set_curr_task = set_curr_task_dl,
1794 .task_tick = task_tick_dl,
1795 .task_fork = task_fork_dl,
1796 .task_dead = task_dead_dl,
1797
1798 .prio_changed = prio_changed_dl,
1799 .switched_from = switched_from_dl,
1800 .switched_to = switched_to_dl,
1801
1802 .update_curr = update_curr_dl,
1803};
1804
1805#ifdef CONFIG_SCHED_DEBUG
1806extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
1807
1808void print_dl_stats(struct seq_file *m, int cpu)
1809{
1810 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
1811}
1812#endif /* CONFIG_SCHED_DEBUG */