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_RT_MUTEXES
47static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
48{
49 return dl_se->pi_se;
50}
51
52static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
53{
54 return pi_of(dl_se) != dl_se;
55}
56#else
57static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
58{
59 return dl_se;
60}
61
62static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
63{
64 return false;
65}
66#endif
67
68#ifdef CONFIG_SMP
69static inline struct dl_bw *dl_bw_of(int i)
70{
71 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
72 "sched RCU must be held");
73 return &cpu_rq(i)->rd->dl_bw;
74}
75
76static inline int dl_bw_cpus(int i)
77{
78 struct root_domain *rd = cpu_rq(i)->rd;
79 int cpus;
80
81 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
82 "sched RCU must be held");
83
84 if (cpumask_subset(rd->span, cpu_active_mask))
85 return cpumask_weight(rd->span);
86
87 cpus = 0;
88
89 for_each_cpu_and(i, rd->span, cpu_active_mask)
90 cpus++;
91
92 return cpus;
93}
94
95static inline unsigned long __dl_bw_capacity(int i)
96{
97 struct root_domain *rd = cpu_rq(i)->rd;
98 unsigned long cap = 0;
99
100 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
101 "sched RCU must be held");
102
103 for_each_cpu_and(i, rd->span, cpu_active_mask)
104 cap += capacity_orig_of(i);
105
106 return cap;
107}
108
109/*
110 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
111 * of the CPU the task is running on rather rd's \Sum CPU capacity.
112 */
113static inline unsigned long dl_bw_capacity(int i)
114{
115 if (!static_branch_unlikely(&sched_asym_cpucapacity) &&
116 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
117 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
118 } else {
119 return __dl_bw_capacity(i);
120 }
121}
122
123static inline bool dl_bw_visited(int cpu, u64 gen)
124{
125 struct root_domain *rd = cpu_rq(cpu)->rd;
126
127 if (rd->visit_gen == gen)
128 return true;
129
130 rd->visit_gen = gen;
131 return false;
132}
133#else
134static inline struct dl_bw *dl_bw_of(int i)
135{
136 return &cpu_rq(i)->dl.dl_bw;
137}
138
139static inline int dl_bw_cpus(int i)
140{
141 return 1;
142}
143
144static inline unsigned long dl_bw_capacity(int i)
145{
146 return SCHED_CAPACITY_SCALE;
147}
148
149static inline bool dl_bw_visited(int cpu, u64 gen)
150{
151 return false;
152}
153#endif
154
155static inline
156void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
157{
158 u64 old = dl_rq->running_bw;
159
160 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
161 dl_rq->running_bw += dl_bw;
162 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
163 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
164 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
165 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
166}
167
168static inline
169void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
170{
171 u64 old = dl_rq->running_bw;
172
173 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
174 dl_rq->running_bw -= dl_bw;
175 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
176 if (dl_rq->running_bw > old)
177 dl_rq->running_bw = 0;
178 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
179 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
180}
181
182static inline
183void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
184{
185 u64 old = dl_rq->this_bw;
186
187 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
188 dl_rq->this_bw += dl_bw;
189 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
190}
191
192static inline
193void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
194{
195 u64 old = dl_rq->this_bw;
196
197 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
198 dl_rq->this_bw -= dl_bw;
199 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
200 if (dl_rq->this_bw > old)
201 dl_rq->this_bw = 0;
202 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
203}
204
205static inline
206void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
207{
208 if (!dl_entity_is_special(dl_se))
209 __add_rq_bw(dl_se->dl_bw, dl_rq);
210}
211
212static inline
213void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
214{
215 if (!dl_entity_is_special(dl_se))
216 __sub_rq_bw(dl_se->dl_bw, dl_rq);
217}
218
219static inline
220void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
221{
222 if (!dl_entity_is_special(dl_se))
223 __add_running_bw(dl_se->dl_bw, dl_rq);
224}
225
226static inline
227void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
228{
229 if (!dl_entity_is_special(dl_se))
230 __sub_running_bw(dl_se->dl_bw, dl_rq);
231}
232
233static void dl_change_utilization(struct task_struct *p, u64 new_bw)
234{
235 struct rq *rq;
236
237 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
238
239 if (task_on_rq_queued(p))
240 return;
241
242 rq = task_rq(p);
243 if (p->dl.dl_non_contending) {
244 sub_running_bw(&p->dl, &rq->dl);
245 p->dl.dl_non_contending = 0;
246 /*
247 * If the timer handler is currently running and the
248 * timer cannot be canceled, inactive_task_timer()
249 * will see that dl_not_contending is not set, and
250 * will not touch the rq's active utilization,
251 * so we are still safe.
252 */
253 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
254 put_task_struct(p);
255 }
256 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
257 __add_rq_bw(new_bw, &rq->dl);
258}
259
260/*
261 * The utilization of a task cannot be immediately removed from
262 * the rq active utilization (running_bw) when the task blocks.
263 * Instead, we have to wait for the so called "0-lag time".
264 *
265 * If a task blocks before the "0-lag time", a timer (the inactive
266 * timer) is armed, and running_bw is decreased when the timer
267 * fires.
268 *
269 * If the task wakes up again before the inactive timer fires,
270 * the timer is canceled, whereas if the task wakes up after the
271 * inactive timer fired (and running_bw has been decreased) the
272 * task's utilization has to be added to running_bw again.
273 * A flag in the deadline scheduling entity (dl_non_contending)
274 * is used to avoid race conditions between the inactive timer handler
275 * and task wakeups.
276 *
277 * The following diagram shows how running_bw is updated. A task is
278 * "ACTIVE" when its utilization contributes to running_bw; an
279 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
280 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
281 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
282 * time already passed, which does not contribute to running_bw anymore.
283 * +------------------+
284 * wakeup | ACTIVE |
285 * +------------------>+ contending |
286 * | add_running_bw | |
287 * | +----+------+------+
288 * | | ^
289 * | dequeue | |
290 * +--------+-------+ | |
291 * | | t >= 0-lag | | wakeup
292 * | INACTIVE |<---------------+ |
293 * | | sub_running_bw | |
294 * +--------+-------+ | |
295 * ^ | |
296 * | t < 0-lag | |
297 * | | |
298 * | V |
299 * | +----+------+------+
300 * | sub_running_bw | ACTIVE |
301 * +-------------------+ |
302 * inactive timer | non contending |
303 * fired +------------------+
304 *
305 * The task_non_contending() function is invoked when a task
306 * blocks, and checks if the 0-lag time already passed or
307 * not (in the first case, it directly updates running_bw;
308 * in the second case, it arms the inactive timer).
309 *
310 * The task_contending() function is invoked when a task wakes
311 * up, and checks if the task is still in the "ACTIVE non contending"
312 * state or not (in the second case, it updates running_bw).
313 */
314static void task_non_contending(struct task_struct *p)
315{
316 struct sched_dl_entity *dl_se = &p->dl;
317 struct hrtimer *timer = &dl_se->inactive_timer;
318 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
319 struct rq *rq = rq_of_dl_rq(dl_rq);
320 s64 zerolag_time;
321
322 /*
323 * If this is a non-deadline task that has been boosted,
324 * do nothing
325 */
326 if (dl_se->dl_runtime == 0)
327 return;
328
329 if (dl_entity_is_special(dl_se))
330 return;
331
332 WARN_ON(dl_se->dl_non_contending);
333
334 zerolag_time = dl_se->deadline -
335 div64_long((dl_se->runtime * dl_se->dl_period),
336 dl_se->dl_runtime);
337
338 /*
339 * Using relative times instead of the absolute "0-lag time"
340 * allows to simplify the code
341 */
342 zerolag_time -= rq_clock(rq);
343
344 /*
345 * If the "0-lag time" already passed, decrease the active
346 * utilization now, instead of starting a timer
347 */
348 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
349 if (dl_task(p))
350 sub_running_bw(dl_se, dl_rq);
351 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
352 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
353
354 if (READ_ONCE(p->__state) == TASK_DEAD)
355 sub_rq_bw(&p->dl, &rq->dl);
356 raw_spin_lock(&dl_b->lock);
357 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
358 __dl_clear_params(p);
359 raw_spin_unlock(&dl_b->lock);
360 }
361
362 return;
363 }
364
365 dl_se->dl_non_contending = 1;
366 get_task_struct(p);
367 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
368}
369
370static void task_contending(struct sched_dl_entity *dl_se, int flags)
371{
372 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
373
374 /*
375 * If this is a non-deadline task that has been boosted,
376 * do nothing
377 */
378 if (dl_se->dl_runtime == 0)
379 return;
380
381 if (flags & ENQUEUE_MIGRATED)
382 add_rq_bw(dl_se, dl_rq);
383
384 if (dl_se->dl_non_contending) {
385 dl_se->dl_non_contending = 0;
386 /*
387 * If the timer handler is currently running and the
388 * timer cannot be canceled, inactive_task_timer()
389 * will see that dl_not_contending is not set, and
390 * will not touch the rq's active utilization,
391 * so we are still safe.
392 */
393 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
394 put_task_struct(dl_task_of(dl_se));
395 } else {
396 /*
397 * Since "dl_non_contending" is not set, the
398 * task's utilization has already been removed from
399 * active utilization (either when the task blocked,
400 * when the "inactive timer" fired).
401 * So, add it back.
402 */
403 add_running_bw(dl_se, dl_rq);
404 }
405}
406
407static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
408{
409 struct sched_dl_entity *dl_se = &p->dl;
410
411 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
412}
413
414static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
415
416void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
417{
418 raw_spin_lock_init(&dl_b->dl_runtime_lock);
419 dl_b->dl_period = period;
420 dl_b->dl_runtime = runtime;
421}
422
423void init_dl_bw(struct dl_bw *dl_b)
424{
425 raw_spin_lock_init(&dl_b->lock);
426 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
427 if (global_rt_runtime() == RUNTIME_INF)
428 dl_b->bw = -1;
429 else
430 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
431 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
432 dl_b->total_bw = 0;
433}
434
435void init_dl_rq(struct dl_rq *dl_rq)
436{
437 dl_rq->root = RB_ROOT_CACHED;
438
439#ifdef CONFIG_SMP
440 /* zero means no -deadline tasks */
441 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
442
443 dl_rq->dl_nr_migratory = 0;
444 dl_rq->overloaded = 0;
445 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
446#else
447 init_dl_bw(&dl_rq->dl_bw);
448#endif
449
450 dl_rq->running_bw = 0;
451 dl_rq->this_bw = 0;
452 init_dl_rq_bw_ratio(dl_rq);
453}
454
455#ifdef CONFIG_SMP
456
457static inline int dl_overloaded(struct rq *rq)
458{
459 return atomic_read(&rq->rd->dlo_count);
460}
461
462static inline void dl_set_overload(struct rq *rq)
463{
464 if (!rq->online)
465 return;
466
467 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
468 /*
469 * Must be visible before the overload count is
470 * set (as in sched_rt.c).
471 *
472 * Matched by the barrier in pull_dl_task().
473 */
474 smp_wmb();
475 atomic_inc(&rq->rd->dlo_count);
476}
477
478static inline void dl_clear_overload(struct rq *rq)
479{
480 if (!rq->online)
481 return;
482
483 atomic_dec(&rq->rd->dlo_count);
484 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
485}
486
487static void update_dl_migration(struct dl_rq *dl_rq)
488{
489 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
490 if (!dl_rq->overloaded) {
491 dl_set_overload(rq_of_dl_rq(dl_rq));
492 dl_rq->overloaded = 1;
493 }
494 } else if (dl_rq->overloaded) {
495 dl_clear_overload(rq_of_dl_rq(dl_rq));
496 dl_rq->overloaded = 0;
497 }
498}
499
500static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
501{
502 struct task_struct *p = dl_task_of(dl_se);
503
504 if (p->nr_cpus_allowed > 1)
505 dl_rq->dl_nr_migratory++;
506
507 update_dl_migration(dl_rq);
508}
509
510static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
511{
512 struct task_struct *p = dl_task_of(dl_se);
513
514 if (p->nr_cpus_allowed > 1)
515 dl_rq->dl_nr_migratory--;
516
517 update_dl_migration(dl_rq);
518}
519
520#define __node_2_pdl(node) \
521 rb_entry((node), struct task_struct, pushable_dl_tasks)
522
523static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
524{
525 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
526}
527
528/*
529 * The list of pushable -deadline task is not a plist, like in
530 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
531 */
532static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
533{
534 struct rb_node *leftmost;
535
536 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
537
538 leftmost = rb_add_cached(&p->pushable_dl_tasks,
539 &rq->dl.pushable_dl_tasks_root,
540 __pushable_less);
541 if (leftmost)
542 rq->dl.earliest_dl.next = p->dl.deadline;
543}
544
545static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
546{
547 struct dl_rq *dl_rq = &rq->dl;
548 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
549 struct rb_node *leftmost;
550
551 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
552 return;
553
554 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
555 if (leftmost)
556 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
557
558 RB_CLEAR_NODE(&p->pushable_dl_tasks);
559}
560
561static inline int has_pushable_dl_tasks(struct rq *rq)
562{
563 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
564}
565
566static int push_dl_task(struct rq *rq);
567
568static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
569{
570 return rq->online && dl_task(prev);
571}
572
573static DEFINE_PER_CPU(struct callback_head, dl_push_head);
574static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
575
576static void push_dl_tasks(struct rq *);
577static void pull_dl_task(struct rq *);
578
579static inline void deadline_queue_push_tasks(struct rq *rq)
580{
581 if (!has_pushable_dl_tasks(rq))
582 return;
583
584 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
585}
586
587static inline void deadline_queue_pull_task(struct rq *rq)
588{
589 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
590}
591
592static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
593
594static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
595{
596 struct rq *later_rq = NULL;
597 struct dl_bw *dl_b;
598
599 later_rq = find_lock_later_rq(p, rq);
600 if (!later_rq) {
601 int cpu;
602
603 /*
604 * If we cannot preempt any rq, fall back to pick any
605 * online CPU:
606 */
607 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
608 if (cpu >= nr_cpu_ids) {
609 /*
610 * Failed to find any suitable CPU.
611 * The task will never come back!
612 */
613 BUG_ON(dl_bandwidth_enabled());
614
615 /*
616 * If admission control is disabled we
617 * try a little harder to let the task
618 * run.
619 */
620 cpu = cpumask_any(cpu_active_mask);
621 }
622 later_rq = cpu_rq(cpu);
623 double_lock_balance(rq, later_rq);
624 }
625
626 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
627 /*
628 * Inactive timer is armed (or callback is running, but
629 * waiting for us to release rq locks). In any case, when it
630 * will fire (or continue), it will see running_bw of this
631 * task migrated to later_rq (and correctly handle it).
632 */
633 sub_running_bw(&p->dl, &rq->dl);
634 sub_rq_bw(&p->dl, &rq->dl);
635
636 add_rq_bw(&p->dl, &later_rq->dl);
637 add_running_bw(&p->dl, &later_rq->dl);
638 } else {
639 sub_rq_bw(&p->dl, &rq->dl);
640 add_rq_bw(&p->dl, &later_rq->dl);
641 }
642
643 /*
644 * And we finally need to fixup root_domain(s) bandwidth accounting,
645 * since p is still hanging out in the old (now moved to default) root
646 * domain.
647 */
648 dl_b = &rq->rd->dl_bw;
649 raw_spin_lock(&dl_b->lock);
650 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
651 raw_spin_unlock(&dl_b->lock);
652
653 dl_b = &later_rq->rd->dl_bw;
654 raw_spin_lock(&dl_b->lock);
655 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
656 raw_spin_unlock(&dl_b->lock);
657
658 set_task_cpu(p, later_rq->cpu);
659 double_unlock_balance(later_rq, rq);
660
661 return later_rq;
662}
663
664#else
665
666static inline
667void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
668{
669}
670
671static inline
672void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
673{
674}
675
676static inline
677void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
678{
679}
680
681static inline
682void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
683{
684}
685
686static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
687{
688 return false;
689}
690
691static inline void pull_dl_task(struct rq *rq)
692{
693}
694
695static inline void deadline_queue_push_tasks(struct rq *rq)
696{
697}
698
699static inline void deadline_queue_pull_task(struct rq *rq)
700{
701}
702#endif /* CONFIG_SMP */
703
704static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
705static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
706static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
707
708/*
709 * We are being explicitly informed that a new instance is starting,
710 * and this means that:
711 * - the absolute deadline of the entity has to be placed at
712 * current time + relative deadline;
713 * - the runtime of the entity has to be set to the maximum value.
714 *
715 * The capability of specifying such event is useful whenever a -deadline
716 * entity wants to (try to!) synchronize its behaviour with the scheduler's
717 * one, and to (try to!) reconcile itself with its own scheduling
718 * parameters.
719 */
720static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
721{
722 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
723 struct rq *rq = rq_of_dl_rq(dl_rq);
724
725 WARN_ON(is_dl_boosted(dl_se));
726 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
727
728 /*
729 * We are racing with the deadline timer. So, do nothing because
730 * the deadline timer handler will take care of properly recharging
731 * the runtime and postponing the deadline
732 */
733 if (dl_se->dl_throttled)
734 return;
735
736 /*
737 * We use the regular wall clock time to set deadlines in the
738 * future; in fact, we must consider execution overheads (time
739 * spent on hardirq context, etc.).
740 */
741 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
742 dl_se->runtime = dl_se->dl_runtime;
743}
744
745/*
746 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
747 * possibility of a entity lasting more than what it declared, and thus
748 * exhausting its runtime.
749 *
750 * Here we are interested in making runtime overrun possible, but we do
751 * not want a entity which is misbehaving to affect the scheduling of all
752 * other entities.
753 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
754 * is used, in order to confine each entity within its own bandwidth.
755 *
756 * This function deals exactly with that, and ensures that when the runtime
757 * of a entity is replenished, its deadline is also postponed. That ensures
758 * the overrunning entity can't interfere with other entity in the system and
759 * can't make them miss their deadlines. Reasons why this kind of overruns
760 * could happen are, typically, a entity voluntarily trying to overcome its
761 * runtime, or it just underestimated it during sched_setattr().
762 */
763static void replenish_dl_entity(struct sched_dl_entity *dl_se)
764{
765 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
766 struct rq *rq = rq_of_dl_rq(dl_rq);
767
768 BUG_ON(pi_of(dl_se)->dl_runtime <= 0);
769
770 /*
771 * This could be the case for a !-dl task that is boosted.
772 * Just go with full inherited parameters.
773 */
774 if (dl_se->dl_deadline == 0) {
775 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
776 dl_se->runtime = pi_of(dl_se)->dl_runtime;
777 }
778
779 if (dl_se->dl_yielded && dl_se->runtime > 0)
780 dl_se->runtime = 0;
781
782 /*
783 * We keep moving the deadline away until we get some
784 * available runtime for the entity. This ensures correct
785 * handling of situations where the runtime overrun is
786 * arbitrary large.
787 */
788 while (dl_se->runtime <= 0) {
789 dl_se->deadline += pi_of(dl_se)->dl_period;
790 dl_se->runtime += pi_of(dl_se)->dl_runtime;
791 }
792
793 /*
794 * At this point, the deadline really should be "in
795 * the future" with respect to rq->clock. If it's
796 * not, we are, for some reason, lagging too much!
797 * Anyway, after having warn userspace abut that,
798 * we still try to keep the things running by
799 * resetting the deadline and the budget of the
800 * entity.
801 */
802 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
803 printk_deferred_once("sched: DL replenish lagged too much\n");
804 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
805 dl_se->runtime = pi_of(dl_se)->dl_runtime;
806 }
807
808 if (dl_se->dl_yielded)
809 dl_se->dl_yielded = 0;
810 if (dl_se->dl_throttled)
811 dl_se->dl_throttled = 0;
812}
813
814/*
815 * Here we check if --at time t-- an entity (which is probably being
816 * [re]activated or, in general, enqueued) can use its remaining runtime
817 * and its current deadline _without_ exceeding the bandwidth it is
818 * assigned (function returns true if it can't). We are in fact applying
819 * one of the CBS rules: when a task wakes up, if the residual runtime
820 * over residual deadline fits within the allocated bandwidth, then we
821 * can keep the current (absolute) deadline and residual budget without
822 * disrupting the schedulability of the system. Otherwise, we should
823 * refill the runtime and set the deadline a period in the future,
824 * because keeping the current (absolute) deadline of the task would
825 * result in breaking guarantees promised to other tasks (refer to
826 * Documentation/scheduler/sched-deadline.rst for more information).
827 *
828 * This function returns true if:
829 *
830 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
831 *
832 * IOW we can't recycle current parameters.
833 *
834 * Notice that the bandwidth check is done against the deadline. For
835 * task with deadline equal to period this is the same of using
836 * dl_period instead of dl_deadline in the equation above.
837 */
838static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
839{
840 u64 left, right;
841
842 /*
843 * left and right are the two sides of the equation above,
844 * after a bit of shuffling to use multiplications instead
845 * of divisions.
846 *
847 * Note that none of the time values involved in the two
848 * multiplications are absolute: dl_deadline and dl_runtime
849 * are the relative deadline and the maximum runtime of each
850 * instance, runtime is the runtime left for the last instance
851 * and (deadline - t), since t is rq->clock, is the time left
852 * to the (absolute) deadline. Even if overflowing the u64 type
853 * is very unlikely to occur in both cases, here we scale down
854 * as we want to avoid that risk at all. Scaling down by 10
855 * means that we reduce granularity to 1us. We are fine with it,
856 * since this is only a true/false check and, anyway, thinking
857 * of anything below microseconds resolution is actually fiction
858 * (but still we want to give the user that illusion >;).
859 */
860 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
861 right = ((dl_se->deadline - t) >> DL_SCALE) *
862 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
863
864 return dl_time_before(right, left);
865}
866
867/*
868 * Revised wakeup rule [1]: For self-suspending tasks, rather then
869 * re-initializing task's runtime and deadline, the revised wakeup
870 * rule adjusts the task's runtime to avoid the task to overrun its
871 * density.
872 *
873 * Reasoning: a task may overrun the density if:
874 * runtime / (deadline - t) > dl_runtime / dl_deadline
875 *
876 * Therefore, runtime can be adjusted to:
877 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
878 *
879 * In such way that runtime will be equal to the maximum density
880 * the task can use without breaking any rule.
881 *
882 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
883 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
884 */
885static void
886update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
887{
888 u64 laxity = dl_se->deadline - rq_clock(rq);
889
890 /*
891 * If the task has deadline < period, and the deadline is in the past,
892 * it should already be throttled before this check.
893 *
894 * See update_dl_entity() comments for further details.
895 */
896 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
897
898 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
899}
900
901/*
902 * Regarding the deadline, a task with implicit deadline has a relative
903 * deadline == relative period. A task with constrained deadline has a
904 * relative deadline <= relative period.
905 *
906 * We support constrained deadline tasks. However, there are some restrictions
907 * applied only for tasks which do not have an implicit deadline. See
908 * update_dl_entity() to know more about such restrictions.
909 *
910 * The dl_is_implicit() returns true if the task has an implicit deadline.
911 */
912static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
913{
914 return dl_se->dl_deadline == dl_se->dl_period;
915}
916
917/*
918 * When a deadline entity is placed in the runqueue, its runtime and deadline
919 * might need to be updated. This is done by a CBS wake up rule. There are two
920 * different rules: 1) the original CBS; and 2) the Revisited CBS.
921 *
922 * When the task is starting a new period, the Original CBS is used. In this
923 * case, the runtime is replenished and a new absolute deadline is set.
924 *
925 * When a task is queued before the begin of the next period, using the
926 * remaining runtime and deadline could make the entity to overflow, see
927 * dl_entity_overflow() to find more about runtime overflow. When such case
928 * is detected, the runtime and deadline need to be updated.
929 *
930 * If the task has an implicit deadline, i.e., deadline == period, the Original
931 * CBS is applied. the runtime is replenished and a new absolute deadline is
932 * set, as in the previous cases.
933 *
934 * However, the Original CBS does not work properly for tasks with
935 * deadline < period, which are said to have a constrained deadline. By
936 * applying the Original CBS, a constrained deadline task would be able to run
937 * runtime/deadline in a period. With deadline < period, the task would
938 * overrun the runtime/period allowed bandwidth, breaking the admission test.
939 *
940 * In order to prevent this misbehave, the Revisited CBS is used for
941 * constrained deadline tasks when a runtime overflow is detected. In the
942 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
943 * the remaining runtime of the task is reduced to avoid runtime overflow.
944 * Please refer to the comments update_dl_revised_wakeup() function to find
945 * more about the Revised CBS rule.
946 */
947static void update_dl_entity(struct sched_dl_entity *dl_se)
948{
949 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
950 struct rq *rq = rq_of_dl_rq(dl_rq);
951
952 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
953 dl_entity_overflow(dl_se, rq_clock(rq))) {
954
955 if (unlikely(!dl_is_implicit(dl_se) &&
956 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
957 !is_dl_boosted(dl_se))) {
958 update_dl_revised_wakeup(dl_se, rq);
959 return;
960 }
961
962 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
963 dl_se->runtime = pi_of(dl_se)->dl_runtime;
964 }
965}
966
967static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
968{
969 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
970}
971
972/*
973 * If the entity depleted all its runtime, and if we want it to sleep
974 * while waiting for some new execution time to become available, we
975 * set the bandwidth replenishment timer to the replenishment instant
976 * and try to activate it.
977 *
978 * Notice that it is important for the caller to know if the timer
979 * actually started or not (i.e., the replenishment instant is in
980 * the future or in the past).
981 */
982static int start_dl_timer(struct task_struct *p)
983{
984 struct sched_dl_entity *dl_se = &p->dl;
985 struct hrtimer *timer = &dl_se->dl_timer;
986 struct rq *rq = task_rq(p);
987 ktime_t now, act;
988 s64 delta;
989
990 lockdep_assert_rq_held(rq);
991
992 /*
993 * We want the timer to fire at the deadline, but considering
994 * that it is actually coming from rq->clock and not from
995 * hrtimer's time base reading.
996 */
997 act = ns_to_ktime(dl_next_period(dl_se));
998 now = hrtimer_cb_get_time(timer);
999 delta = ktime_to_ns(now) - rq_clock(rq);
1000 act = ktime_add_ns(act, delta);
1001
1002 /*
1003 * If the expiry time already passed, e.g., because the value
1004 * chosen as the deadline is too small, don't even try to
1005 * start the timer in the past!
1006 */
1007 if (ktime_us_delta(act, now) < 0)
1008 return 0;
1009
1010 /*
1011 * !enqueued will guarantee another callback; even if one is already in
1012 * progress. This ensures a balanced {get,put}_task_struct().
1013 *
1014 * The race against __run_timer() clearing the enqueued state is
1015 * harmless because we're holding task_rq()->lock, therefore the timer
1016 * expiring after we've done the check will wait on its task_rq_lock()
1017 * and observe our state.
1018 */
1019 if (!hrtimer_is_queued(timer)) {
1020 get_task_struct(p);
1021 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1022 }
1023
1024 return 1;
1025}
1026
1027/*
1028 * This is the bandwidth enforcement timer callback. If here, we know
1029 * a task is not on its dl_rq, since the fact that the timer was running
1030 * means the task is throttled and needs a runtime replenishment.
1031 *
1032 * However, what we actually do depends on the fact the task is active,
1033 * (it is on its rq) or has been removed from there by a call to
1034 * dequeue_task_dl(). In the former case we must issue the runtime
1035 * replenishment and add the task back to the dl_rq; in the latter, we just
1036 * do nothing but clearing dl_throttled, so that runtime and deadline
1037 * updating (and the queueing back to dl_rq) will be done by the
1038 * next call to enqueue_task_dl().
1039 */
1040static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1041{
1042 struct sched_dl_entity *dl_se = container_of(timer,
1043 struct sched_dl_entity,
1044 dl_timer);
1045 struct task_struct *p = dl_task_of(dl_se);
1046 struct rq_flags rf;
1047 struct rq *rq;
1048
1049 rq = task_rq_lock(p, &rf);
1050
1051 /*
1052 * The task might have changed its scheduling policy to something
1053 * different than SCHED_DEADLINE (through switched_from_dl()).
1054 */
1055 if (!dl_task(p))
1056 goto unlock;
1057
1058 /*
1059 * The task might have been boosted by someone else and might be in the
1060 * boosting/deboosting path, its not throttled.
1061 */
1062 if (is_dl_boosted(dl_se))
1063 goto unlock;
1064
1065 /*
1066 * Spurious timer due to start_dl_timer() race; or we already received
1067 * a replenishment from rt_mutex_setprio().
1068 */
1069 if (!dl_se->dl_throttled)
1070 goto unlock;
1071
1072 sched_clock_tick();
1073 update_rq_clock(rq);
1074
1075 /*
1076 * If the throttle happened during sched-out; like:
1077 *
1078 * schedule()
1079 * deactivate_task()
1080 * dequeue_task_dl()
1081 * update_curr_dl()
1082 * start_dl_timer()
1083 * __dequeue_task_dl()
1084 * prev->on_rq = 0;
1085 *
1086 * We can be both throttled and !queued. Replenish the counter
1087 * but do not enqueue -- wait for our wakeup to do that.
1088 */
1089 if (!task_on_rq_queued(p)) {
1090 replenish_dl_entity(dl_se);
1091 goto unlock;
1092 }
1093
1094#ifdef CONFIG_SMP
1095 if (unlikely(!rq->online)) {
1096 /*
1097 * If the runqueue is no longer available, migrate the
1098 * task elsewhere. This necessarily changes rq.
1099 */
1100 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1101 rq = dl_task_offline_migration(rq, p);
1102 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1103 update_rq_clock(rq);
1104
1105 /*
1106 * Now that the task has been migrated to the new RQ and we
1107 * have that locked, proceed as normal and enqueue the task
1108 * there.
1109 */
1110 }
1111#endif
1112
1113 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1114 if (dl_task(rq->curr))
1115 check_preempt_curr_dl(rq, p, 0);
1116 else
1117 resched_curr(rq);
1118
1119#ifdef CONFIG_SMP
1120 /*
1121 * Queueing this task back might have overloaded rq, check if we need
1122 * to kick someone away.
1123 */
1124 if (has_pushable_dl_tasks(rq)) {
1125 /*
1126 * Nothing relies on rq->lock after this, so its safe to drop
1127 * rq->lock.
1128 */
1129 rq_unpin_lock(rq, &rf);
1130 push_dl_task(rq);
1131 rq_repin_lock(rq, &rf);
1132 }
1133#endif
1134
1135unlock:
1136 task_rq_unlock(rq, p, &rf);
1137
1138 /*
1139 * This can free the task_struct, including this hrtimer, do not touch
1140 * anything related to that after this.
1141 */
1142 put_task_struct(p);
1143
1144 return HRTIMER_NORESTART;
1145}
1146
1147void init_dl_task_timer(struct sched_dl_entity *dl_se)
1148{
1149 struct hrtimer *timer = &dl_se->dl_timer;
1150
1151 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1152 timer->function = dl_task_timer;
1153}
1154
1155/*
1156 * During the activation, CBS checks if it can reuse the current task's
1157 * runtime and period. If the deadline of the task is in the past, CBS
1158 * cannot use the runtime, and so it replenishes the task. This rule
1159 * works fine for implicit deadline tasks (deadline == period), and the
1160 * CBS was designed for implicit deadline tasks. However, a task with
1161 * constrained deadline (deadline < period) might be awakened after the
1162 * deadline, but before the next period. In this case, replenishing the
1163 * task would allow it to run for runtime / deadline. As in this case
1164 * deadline < period, CBS enables a task to run for more than the
1165 * runtime / period. In a very loaded system, this can cause a domino
1166 * effect, making other tasks miss their deadlines.
1167 *
1168 * To avoid this problem, in the activation of a constrained deadline
1169 * task after the deadline but before the next period, throttle the
1170 * task and set the replenishing timer to the begin of the next period,
1171 * unless it is boosted.
1172 */
1173static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1174{
1175 struct task_struct *p = dl_task_of(dl_se);
1176 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1177
1178 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1179 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1180 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
1181 return;
1182 dl_se->dl_throttled = 1;
1183 if (dl_se->runtime > 0)
1184 dl_se->runtime = 0;
1185 }
1186}
1187
1188static
1189int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1190{
1191 return (dl_se->runtime <= 0);
1192}
1193
1194extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1195
1196/*
1197 * This function implements the GRUB accounting rule:
1198 * according to the GRUB reclaiming algorithm, the runtime is
1199 * not decreased as "dq = -dt", but as
1200 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1201 * where u is the utilization of the task, Umax is the maximum reclaimable
1202 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1203 * as the difference between the "total runqueue utilization" and the
1204 * runqueue active utilization, and Uextra is the (per runqueue) extra
1205 * reclaimable utilization.
1206 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1207 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1208 * BW_SHIFT.
1209 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT,
1210 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1211 * Since delta is a 64 bit variable, to have an overflow its value
1212 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1213 * So, overflow is not an issue here.
1214 */
1215static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1216{
1217 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1218 u64 u_act;
1219 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1220
1221 /*
1222 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1223 * we compare u_inact + rq->dl.extra_bw with
1224 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1225 * u_inact + rq->dl.extra_bw can be larger than
1226 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1227 * leading to wrong results)
1228 */
1229 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1230 u_act = u_act_min;
1231 else
1232 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1233
1234 return (delta * u_act) >> BW_SHIFT;
1235}
1236
1237/*
1238 * Update the current task's runtime statistics (provided it is still
1239 * a -deadline task and has not been removed from the dl_rq).
1240 */
1241static void update_curr_dl(struct rq *rq)
1242{
1243 struct task_struct *curr = rq->curr;
1244 struct sched_dl_entity *dl_se = &curr->dl;
1245 u64 delta_exec, scaled_delta_exec;
1246 int cpu = cpu_of(rq);
1247 u64 now;
1248
1249 if (!dl_task(curr) || !on_dl_rq(dl_se))
1250 return;
1251
1252 /*
1253 * Consumed budget is computed considering the time as
1254 * observed by schedulable tasks (excluding time spent
1255 * in hardirq context, etc.). Deadlines are instead
1256 * computed using hard walltime. This seems to be the more
1257 * natural solution, but the full ramifications of this
1258 * approach need further study.
1259 */
1260 now = rq_clock_task(rq);
1261 delta_exec = now - curr->se.exec_start;
1262 if (unlikely((s64)delta_exec <= 0)) {
1263 if (unlikely(dl_se->dl_yielded))
1264 goto throttle;
1265 return;
1266 }
1267
1268 schedstat_set(curr->se.statistics.exec_max,
1269 max(curr->se.statistics.exec_max, delta_exec));
1270
1271 curr->se.sum_exec_runtime += delta_exec;
1272 account_group_exec_runtime(curr, delta_exec);
1273
1274 curr->se.exec_start = now;
1275 cgroup_account_cputime(curr, delta_exec);
1276
1277 if (dl_entity_is_special(dl_se))
1278 return;
1279
1280 /*
1281 * For tasks that participate in GRUB, we implement GRUB-PA: the
1282 * spare reclaimed bandwidth is used to clock down frequency.
1283 *
1284 * For the others, we still need to scale reservation parameters
1285 * according to current frequency and CPU maximum capacity.
1286 */
1287 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1288 scaled_delta_exec = grub_reclaim(delta_exec,
1289 rq,
1290 &curr->dl);
1291 } else {
1292 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1293 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1294
1295 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1296 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1297 }
1298
1299 dl_se->runtime -= scaled_delta_exec;
1300
1301throttle:
1302 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1303 dl_se->dl_throttled = 1;
1304
1305 /* If requested, inform the user about runtime overruns. */
1306 if (dl_runtime_exceeded(dl_se) &&
1307 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1308 dl_se->dl_overrun = 1;
1309
1310 __dequeue_task_dl(rq, curr, 0);
1311 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1312 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1313
1314 if (!is_leftmost(curr, &rq->dl))
1315 resched_curr(rq);
1316 }
1317
1318 /*
1319 * Because -- for now -- we share the rt bandwidth, we need to
1320 * account our runtime there too, otherwise actual rt tasks
1321 * would be able to exceed the shared quota.
1322 *
1323 * Account to the root rt group for now.
1324 *
1325 * The solution we're working towards is having the RT groups scheduled
1326 * using deadline servers -- however there's a few nasties to figure
1327 * out before that can happen.
1328 */
1329 if (rt_bandwidth_enabled()) {
1330 struct rt_rq *rt_rq = &rq->rt;
1331
1332 raw_spin_lock(&rt_rq->rt_runtime_lock);
1333 /*
1334 * We'll let actual RT tasks worry about the overflow here, we
1335 * have our own CBS to keep us inline; only account when RT
1336 * bandwidth is relevant.
1337 */
1338 if (sched_rt_bandwidth_account(rt_rq))
1339 rt_rq->rt_time += delta_exec;
1340 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1341 }
1342}
1343
1344static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1345{
1346 struct sched_dl_entity *dl_se = container_of(timer,
1347 struct sched_dl_entity,
1348 inactive_timer);
1349 struct task_struct *p = dl_task_of(dl_se);
1350 struct rq_flags rf;
1351 struct rq *rq;
1352
1353 rq = task_rq_lock(p, &rf);
1354
1355 sched_clock_tick();
1356 update_rq_clock(rq);
1357
1358 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1359 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1360
1361 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1362 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1363 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1364 dl_se->dl_non_contending = 0;
1365 }
1366
1367 raw_spin_lock(&dl_b->lock);
1368 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1369 raw_spin_unlock(&dl_b->lock);
1370 __dl_clear_params(p);
1371
1372 goto unlock;
1373 }
1374 if (dl_se->dl_non_contending == 0)
1375 goto unlock;
1376
1377 sub_running_bw(dl_se, &rq->dl);
1378 dl_se->dl_non_contending = 0;
1379unlock:
1380 task_rq_unlock(rq, p, &rf);
1381 put_task_struct(p);
1382
1383 return HRTIMER_NORESTART;
1384}
1385
1386void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1387{
1388 struct hrtimer *timer = &dl_se->inactive_timer;
1389
1390 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1391 timer->function = inactive_task_timer;
1392}
1393
1394#ifdef CONFIG_SMP
1395
1396static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1397{
1398 struct rq *rq = rq_of_dl_rq(dl_rq);
1399
1400 if (dl_rq->earliest_dl.curr == 0 ||
1401 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1402 if (dl_rq->earliest_dl.curr == 0)
1403 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1404 dl_rq->earliest_dl.curr = deadline;
1405 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1406 }
1407}
1408
1409static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1410{
1411 struct rq *rq = rq_of_dl_rq(dl_rq);
1412
1413 /*
1414 * Since we may have removed our earliest (and/or next earliest)
1415 * task we must recompute them.
1416 */
1417 if (!dl_rq->dl_nr_running) {
1418 dl_rq->earliest_dl.curr = 0;
1419 dl_rq->earliest_dl.next = 0;
1420 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1421 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1422 } else {
1423 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1424 struct sched_dl_entity *entry;
1425
1426 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1427 dl_rq->earliest_dl.curr = entry->deadline;
1428 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1429 }
1430}
1431
1432#else
1433
1434static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1435static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1436
1437#endif /* CONFIG_SMP */
1438
1439static inline
1440void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1441{
1442 int prio = dl_task_of(dl_se)->prio;
1443 u64 deadline = dl_se->deadline;
1444
1445 WARN_ON(!dl_prio(prio));
1446 dl_rq->dl_nr_running++;
1447 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1448
1449 inc_dl_deadline(dl_rq, deadline);
1450 inc_dl_migration(dl_se, dl_rq);
1451}
1452
1453static inline
1454void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1455{
1456 int prio = dl_task_of(dl_se)->prio;
1457
1458 WARN_ON(!dl_prio(prio));
1459 WARN_ON(!dl_rq->dl_nr_running);
1460 dl_rq->dl_nr_running--;
1461 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1462
1463 dec_dl_deadline(dl_rq, dl_se->deadline);
1464 dec_dl_migration(dl_se, dl_rq);
1465}
1466
1467#define __node_2_dle(node) \
1468 rb_entry((node), struct sched_dl_entity, rb_node)
1469
1470static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1471{
1472 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1473}
1474
1475static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1476{
1477 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1478
1479 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1480
1481 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1482
1483 inc_dl_tasks(dl_se, dl_rq);
1484}
1485
1486static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1487{
1488 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1489
1490 if (RB_EMPTY_NODE(&dl_se->rb_node))
1491 return;
1492
1493 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1494
1495 RB_CLEAR_NODE(&dl_se->rb_node);
1496
1497 dec_dl_tasks(dl_se, dl_rq);
1498}
1499
1500static void
1501enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1502{
1503 BUG_ON(on_dl_rq(dl_se));
1504
1505 /*
1506 * If this is a wakeup or a new instance, the scheduling
1507 * parameters of the task might need updating. Otherwise,
1508 * we want a replenishment of its runtime.
1509 */
1510 if (flags & ENQUEUE_WAKEUP) {
1511 task_contending(dl_se, flags);
1512 update_dl_entity(dl_se);
1513 } else if (flags & ENQUEUE_REPLENISH) {
1514 replenish_dl_entity(dl_se);
1515 } else if ((flags & ENQUEUE_RESTORE) &&
1516 dl_time_before(dl_se->deadline,
1517 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1518 setup_new_dl_entity(dl_se);
1519 }
1520
1521 __enqueue_dl_entity(dl_se);
1522}
1523
1524static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1525{
1526 __dequeue_dl_entity(dl_se);
1527}
1528
1529static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1530{
1531 if (is_dl_boosted(&p->dl)) {
1532 /*
1533 * Because of delays in the detection of the overrun of a
1534 * thread's runtime, it might be the case that a thread
1535 * goes to sleep in a rt mutex with negative runtime. As
1536 * a consequence, the thread will be throttled.
1537 *
1538 * While waiting for the mutex, this thread can also be
1539 * boosted via PI, resulting in a thread that is throttled
1540 * and boosted at the same time.
1541 *
1542 * In this case, the boost overrides the throttle.
1543 */
1544 if (p->dl.dl_throttled) {
1545 /*
1546 * The replenish timer needs to be canceled. No
1547 * problem if it fires concurrently: boosted threads
1548 * are ignored in dl_task_timer().
1549 */
1550 hrtimer_try_to_cancel(&p->dl.dl_timer);
1551 p->dl.dl_throttled = 0;
1552 }
1553 } else if (!dl_prio(p->normal_prio)) {
1554 /*
1555 * Special case in which we have a !SCHED_DEADLINE task that is going
1556 * to be deboosted, but exceeds its runtime while doing so. No point in
1557 * replenishing it, as it's going to return back to its original
1558 * scheduling class after this. If it has been throttled, we need to
1559 * clear the flag, otherwise the task may wake up as throttled after
1560 * being boosted again with no means to replenish the runtime and clear
1561 * the throttle.
1562 */
1563 p->dl.dl_throttled = 0;
1564 BUG_ON(!is_dl_boosted(&p->dl) || flags != ENQUEUE_REPLENISH);
1565 return;
1566 }
1567
1568 /*
1569 * Check if a constrained deadline task was activated
1570 * after the deadline but before the next period.
1571 * If that is the case, the task will be throttled and
1572 * the replenishment timer will be set to the next period.
1573 */
1574 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1575 dl_check_constrained_dl(&p->dl);
1576
1577 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1578 add_rq_bw(&p->dl, &rq->dl);
1579 add_running_bw(&p->dl, &rq->dl);
1580 }
1581
1582 /*
1583 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1584 * its budget it needs a replenishment and, since it now is on
1585 * its rq, the bandwidth timer callback (which clearly has not
1586 * run yet) will take care of this.
1587 * However, the active utilization does not depend on the fact
1588 * that the task is on the runqueue or not (but depends on the
1589 * task's state - in GRUB parlance, "inactive" vs "active contending").
1590 * In other words, even if a task is throttled its utilization must
1591 * be counted in the active utilization; hence, we need to call
1592 * add_running_bw().
1593 */
1594 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1595 if (flags & ENQUEUE_WAKEUP)
1596 task_contending(&p->dl, flags);
1597
1598 return;
1599 }
1600
1601 enqueue_dl_entity(&p->dl, flags);
1602
1603 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1604 enqueue_pushable_dl_task(rq, p);
1605}
1606
1607static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1608{
1609 dequeue_dl_entity(&p->dl);
1610 dequeue_pushable_dl_task(rq, p);
1611}
1612
1613static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1614{
1615 update_curr_dl(rq);
1616 __dequeue_task_dl(rq, p, flags);
1617
1618 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1619 sub_running_bw(&p->dl, &rq->dl);
1620 sub_rq_bw(&p->dl, &rq->dl);
1621 }
1622
1623 /*
1624 * This check allows to start the inactive timer (or to immediately
1625 * decrease the active utilization, if needed) in two cases:
1626 * when the task blocks and when it is terminating
1627 * (p->state == TASK_DEAD). We can handle the two cases in the same
1628 * way, because from GRUB's point of view the same thing is happening
1629 * (the task moves from "active contending" to "active non contending"
1630 * or "inactive")
1631 */
1632 if (flags & DEQUEUE_SLEEP)
1633 task_non_contending(p);
1634}
1635
1636/*
1637 * Yield task semantic for -deadline tasks is:
1638 *
1639 * get off from the CPU until our next instance, with
1640 * a new runtime. This is of little use now, since we
1641 * don't have a bandwidth reclaiming mechanism. Anyway,
1642 * bandwidth reclaiming is planned for the future, and
1643 * yield_task_dl will indicate that some spare budget
1644 * is available for other task instances to use it.
1645 */
1646static void yield_task_dl(struct rq *rq)
1647{
1648 /*
1649 * We make the task go to sleep until its current deadline by
1650 * forcing its runtime to zero. This way, update_curr_dl() stops
1651 * it and the bandwidth timer will wake it up and will give it
1652 * new scheduling parameters (thanks to dl_yielded=1).
1653 */
1654 rq->curr->dl.dl_yielded = 1;
1655
1656 update_rq_clock(rq);
1657 update_curr_dl(rq);
1658 /*
1659 * Tell update_rq_clock() that we've just updated,
1660 * so we don't do microscopic update in schedule()
1661 * and double the fastpath cost.
1662 */
1663 rq_clock_skip_update(rq);
1664}
1665
1666#ifdef CONFIG_SMP
1667
1668static int find_later_rq(struct task_struct *task);
1669
1670static int
1671select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1672{
1673 struct task_struct *curr;
1674 bool select_rq;
1675 struct rq *rq;
1676
1677 if (!(flags & WF_TTWU))
1678 goto out;
1679
1680 rq = cpu_rq(cpu);
1681
1682 rcu_read_lock();
1683 curr = READ_ONCE(rq->curr); /* unlocked access */
1684
1685 /*
1686 * If we are dealing with a -deadline task, we must
1687 * decide where to wake it up.
1688 * If it has a later deadline and the current task
1689 * on this rq can't move (provided the waking task
1690 * can!) we prefer to send it somewhere else. On the
1691 * other hand, if it has a shorter deadline, we
1692 * try to make it stay here, it might be important.
1693 */
1694 select_rq = unlikely(dl_task(curr)) &&
1695 (curr->nr_cpus_allowed < 2 ||
1696 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1697 p->nr_cpus_allowed > 1;
1698
1699 /*
1700 * Take the capacity of the CPU into account to
1701 * ensure it fits the requirement of the task.
1702 */
1703 if (static_branch_unlikely(&sched_asym_cpucapacity))
1704 select_rq |= !dl_task_fits_capacity(p, cpu);
1705
1706 if (select_rq) {
1707 int target = find_later_rq(p);
1708
1709 if (target != -1 &&
1710 (dl_time_before(p->dl.deadline,
1711 cpu_rq(target)->dl.earliest_dl.curr) ||
1712 (cpu_rq(target)->dl.dl_nr_running == 0)))
1713 cpu = target;
1714 }
1715 rcu_read_unlock();
1716
1717out:
1718 return cpu;
1719}
1720
1721static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1722{
1723 struct rq *rq;
1724
1725 if (READ_ONCE(p->__state) != TASK_WAKING)
1726 return;
1727
1728 rq = task_rq(p);
1729 /*
1730 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1731 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1732 * rq->lock is not... So, lock it
1733 */
1734 raw_spin_rq_lock(rq);
1735 if (p->dl.dl_non_contending) {
1736 update_rq_clock(rq);
1737 sub_running_bw(&p->dl, &rq->dl);
1738 p->dl.dl_non_contending = 0;
1739 /*
1740 * If the timer handler is currently running and the
1741 * timer cannot be canceled, inactive_task_timer()
1742 * will see that dl_not_contending is not set, and
1743 * will not touch the rq's active utilization,
1744 * so we are still safe.
1745 */
1746 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1747 put_task_struct(p);
1748 }
1749 sub_rq_bw(&p->dl, &rq->dl);
1750 raw_spin_rq_unlock(rq);
1751}
1752
1753static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1754{
1755 /*
1756 * Current can't be migrated, useless to reschedule,
1757 * let's hope p can move out.
1758 */
1759 if (rq->curr->nr_cpus_allowed == 1 ||
1760 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1761 return;
1762
1763 /*
1764 * p is migratable, so let's not schedule it and
1765 * see if it is pushed or pulled somewhere else.
1766 */
1767 if (p->nr_cpus_allowed != 1 &&
1768 cpudl_find(&rq->rd->cpudl, p, NULL))
1769 return;
1770
1771 resched_curr(rq);
1772}
1773
1774static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1775{
1776 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1777 /*
1778 * This is OK, because current is on_cpu, which avoids it being
1779 * picked for load-balance and preemption/IRQs are still
1780 * disabled avoiding further scheduler activity on it and we've
1781 * not yet started the picking loop.
1782 */
1783 rq_unpin_lock(rq, rf);
1784 pull_dl_task(rq);
1785 rq_repin_lock(rq, rf);
1786 }
1787
1788 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1789}
1790#endif /* CONFIG_SMP */
1791
1792/*
1793 * Only called when both the current and waking task are -deadline
1794 * tasks.
1795 */
1796static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1797 int flags)
1798{
1799 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1800 resched_curr(rq);
1801 return;
1802 }
1803
1804#ifdef CONFIG_SMP
1805 /*
1806 * In the unlikely case current and p have the same deadline
1807 * let us try to decide what's the best thing to do...
1808 */
1809 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1810 !test_tsk_need_resched(rq->curr))
1811 check_preempt_equal_dl(rq, p);
1812#endif /* CONFIG_SMP */
1813}
1814
1815#ifdef CONFIG_SCHED_HRTICK
1816static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1817{
1818 hrtick_start(rq, p->dl.runtime);
1819}
1820#else /* !CONFIG_SCHED_HRTICK */
1821static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1822{
1823}
1824#endif
1825
1826static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
1827{
1828 p->se.exec_start = rq_clock_task(rq);
1829
1830 /* You can't push away the running task */
1831 dequeue_pushable_dl_task(rq, p);
1832
1833 if (!first)
1834 return;
1835
1836 if (hrtick_enabled_dl(rq))
1837 start_hrtick_dl(rq, p);
1838
1839 if (rq->curr->sched_class != &dl_sched_class)
1840 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1841
1842 deadline_queue_push_tasks(rq);
1843}
1844
1845static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1846 struct dl_rq *dl_rq)
1847{
1848 struct rb_node *left = rb_first_cached(&dl_rq->root);
1849
1850 if (!left)
1851 return NULL;
1852
1853 return rb_entry(left, struct sched_dl_entity, rb_node);
1854}
1855
1856static struct task_struct *pick_task_dl(struct rq *rq)
1857{
1858 struct sched_dl_entity *dl_se;
1859 struct dl_rq *dl_rq = &rq->dl;
1860 struct task_struct *p;
1861
1862 if (!sched_dl_runnable(rq))
1863 return NULL;
1864
1865 dl_se = pick_next_dl_entity(rq, dl_rq);
1866 BUG_ON(!dl_se);
1867 p = dl_task_of(dl_se);
1868
1869 return p;
1870}
1871
1872static struct task_struct *pick_next_task_dl(struct rq *rq)
1873{
1874 struct task_struct *p;
1875
1876 p = pick_task_dl(rq);
1877 if (p)
1878 set_next_task_dl(rq, p, true);
1879
1880 return p;
1881}
1882
1883static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1884{
1885 update_curr_dl(rq);
1886
1887 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1888 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1889 enqueue_pushable_dl_task(rq, p);
1890}
1891
1892/*
1893 * scheduler tick hitting a task of our scheduling class.
1894 *
1895 * NOTE: This function can be called remotely by the tick offload that
1896 * goes along full dynticks. Therefore no local assumption can be made
1897 * and everything must be accessed through the @rq and @curr passed in
1898 * parameters.
1899 */
1900static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1901{
1902 update_curr_dl(rq);
1903
1904 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1905 /*
1906 * Even when we have runtime, update_curr_dl() might have resulted in us
1907 * not being the leftmost task anymore. In that case NEED_RESCHED will
1908 * be set and schedule() will start a new hrtick for the next task.
1909 */
1910 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
1911 is_leftmost(p, &rq->dl))
1912 start_hrtick_dl(rq, p);
1913}
1914
1915static void task_fork_dl(struct task_struct *p)
1916{
1917 /*
1918 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1919 * sched_fork()
1920 */
1921}
1922
1923#ifdef CONFIG_SMP
1924
1925/* Only try algorithms three times */
1926#define DL_MAX_TRIES 3
1927
1928static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1929{
1930 if (!task_running(rq, p) &&
1931 cpumask_test_cpu(cpu, &p->cpus_mask))
1932 return 1;
1933 return 0;
1934}
1935
1936/*
1937 * Return the earliest pushable rq's task, which is suitable to be executed
1938 * on the CPU, NULL otherwise:
1939 */
1940static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1941{
1942 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
1943 struct task_struct *p = NULL;
1944
1945 if (!has_pushable_dl_tasks(rq))
1946 return NULL;
1947
1948next_node:
1949 if (next_node) {
1950 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1951
1952 if (pick_dl_task(rq, p, cpu))
1953 return p;
1954
1955 next_node = rb_next(next_node);
1956 goto next_node;
1957 }
1958
1959 return NULL;
1960}
1961
1962static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1963
1964static int find_later_rq(struct task_struct *task)
1965{
1966 struct sched_domain *sd;
1967 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1968 int this_cpu = smp_processor_id();
1969 int cpu = task_cpu(task);
1970
1971 /* Make sure the mask is initialized first */
1972 if (unlikely(!later_mask))
1973 return -1;
1974
1975 if (task->nr_cpus_allowed == 1)
1976 return -1;
1977
1978 /*
1979 * We have to consider system topology and task affinity
1980 * first, then we can look for a suitable CPU.
1981 */
1982 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1983 return -1;
1984
1985 /*
1986 * If we are here, some targets have been found, including
1987 * the most suitable which is, among the runqueues where the
1988 * current tasks have later deadlines than the task's one, the
1989 * rq with the latest possible one.
1990 *
1991 * Now we check how well this matches with task's
1992 * affinity and system topology.
1993 *
1994 * The last CPU where the task run is our first
1995 * guess, since it is most likely cache-hot there.
1996 */
1997 if (cpumask_test_cpu(cpu, later_mask))
1998 return cpu;
1999 /*
2000 * Check if this_cpu is to be skipped (i.e., it is
2001 * not in the mask) or not.
2002 */
2003 if (!cpumask_test_cpu(this_cpu, later_mask))
2004 this_cpu = -1;
2005
2006 rcu_read_lock();
2007 for_each_domain(cpu, sd) {
2008 if (sd->flags & SD_WAKE_AFFINE) {
2009 int best_cpu;
2010
2011 /*
2012 * If possible, preempting this_cpu is
2013 * cheaper than migrating.
2014 */
2015 if (this_cpu != -1 &&
2016 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2017 rcu_read_unlock();
2018 return this_cpu;
2019 }
2020
2021 best_cpu = cpumask_any_and_distribute(later_mask,
2022 sched_domain_span(sd));
2023 /*
2024 * Last chance: if a CPU being in both later_mask
2025 * and current sd span is valid, that becomes our
2026 * choice. Of course, the latest possible CPU is
2027 * already under consideration through later_mask.
2028 */
2029 if (best_cpu < nr_cpu_ids) {
2030 rcu_read_unlock();
2031 return best_cpu;
2032 }
2033 }
2034 }
2035 rcu_read_unlock();
2036
2037 /*
2038 * At this point, all our guesses failed, we just return
2039 * 'something', and let the caller sort the things out.
2040 */
2041 if (this_cpu != -1)
2042 return this_cpu;
2043
2044 cpu = cpumask_any_distribute(later_mask);
2045 if (cpu < nr_cpu_ids)
2046 return cpu;
2047
2048 return -1;
2049}
2050
2051/* Locks the rq it finds */
2052static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2053{
2054 struct rq *later_rq = NULL;
2055 int tries;
2056 int cpu;
2057
2058 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2059 cpu = find_later_rq(task);
2060
2061 if ((cpu == -1) || (cpu == rq->cpu))
2062 break;
2063
2064 later_rq = cpu_rq(cpu);
2065
2066 if (later_rq->dl.dl_nr_running &&
2067 !dl_time_before(task->dl.deadline,
2068 later_rq->dl.earliest_dl.curr)) {
2069 /*
2070 * Target rq has tasks of equal or earlier deadline,
2071 * retrying does not release any lock and is unlikely
2072 * to yield a different result.
2073 */
2074 later_rq = NULL;
2075 break;
2076 }
2077
2078 /* Retry if something changed. */
2079 if (double_lock_balance(rq, later_rq)) {
2080 if (unlikely(task_rq(task) != rq ||
2081 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2082 task_running(rq, task) ||
2083 !dl_task(task) ||
2084 !task_on_rq_queued(task))) {
2085 double_unlock_balance(rq, later_rq);
2086 later_rq = NULL;
2087 break;
2088 }
2089 }
2090
2091 /*
2092 * If the rq we found has no -deadline task, or
2093 * its earliest one has a later deadline than our
2094 * task, the rq is a good one.
2095 */
2096 if (!later_rq->dl.dl_nr_running ||
2097 dl_time_before(task->dl.deadline,
2098 later_rq->dl.earliest_dl.curr))
2099 break;
2100
2101 /* Otherwise we try again. */
2102 double_unlock_balance(rq, later_rq);
2103 later_rq = NULL;
2104 }
2105
2106 return later_rq;
2107}
2108
2109static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2110{
2111 struct task_struct *p;
2112
2113 if (!has_pushable_dl_tasks(rq))
2114 return NULL;
2115
2116 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
2117 struct task_struct, pushable_dl_tasks);
2118
2119 BUG_ON(rq->cpu != task_cpu(p));
2120 BUG_ON(task_current(rq, p));
2121 BUG_ON(p->nr_cpus_allowed <= 1);
2122
2123 BUG_ON(!task_on_rq_queued(p));
2124 BUG_ON(!dl_task(p));
2125
2126 return p;
2127}
2128
2129/*
2130 * See if the non running -deadline tasks on this rq
2131 * can be sent to some other CPU where they can preempt
2132 * and start executing.
2133 */
2134static int push_dl_task(struct rq *rq)
2135{
2136 struct task_struct *next_task;
2137 struct rq *later_rq;
2138 int ret = 0;
2139
2140 if (!rq->dl.overloaded)
2141 return 0;
2142
2143 next_task = pick_next_pushable_dl_task(rq);
2144 if (!next_task)
2145 return 0;
2146
2147retry:
2148 if (is_migration_disabled(next_task))
2149 return 0;
2150
2151 if (WARN_ON(next_task == rq->curr))
2152 return 0;
2153
2154 /*
2155 * If next_task preempts rq->curr, and rq->curr
2156 * can move away, it makes sense to just reschedule
2157 * without going further in pushing next_task.
2158 */
2159 if (dl_task(rq->curr) &&
2160 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2161 rq->curr->nr_cpus_allowed > 1) {
2162 resched_curr(rq);
2163 return 0;
2164 }
2165
2166 /* We might release rq lock */
2167 get_task_struct(next_task);
2168
2169 /* Will lock the rq it'll find */
2170 later_rq = find_lock_later_rq(next_task, rq);
2171 if (!later_rq) {
2172 struct task_struct *task;
2173
2174 /*
2175 * We must check all this again, since
2176 * find_lock_later_rq releases rq->lock and it is
2177 * then possible that next_task has migrated.
2178 */
2179 task = pick_next_pushable_dl_task(rq);
2180 if (task == next_task) {
2181 /*
2182 * The task is still there. We don't try
2183 * again, some other CPU will pull it when ready.
2184 */
2185 goto out;
2186 }
2187
2188 if (!task)
2189 /* No more tasks */
2190 goto out;
2191
2192 put_task_struct(next_task);
2193 next_task = task;
2194 goto retry;
2195 }
2196
2197 deactivate_task(rq, next_task, 0);
2198 set_task_cpu(next_task, later_rq->cpu);
2199
2200 /*
2201 * Update the later_rq clock here, because the clock is used
2202 * by the cpufreq_update_util() inside __add_running_bw().
2203 */
2204 update_rq_clock(later_rq);
2205 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
2206 ret = 1;
2207
2208 resched_curr(later_rq);
2209
2210 double_unlock_balance(rq, later_rq);
2211
2212out:
2213 put_task_struct(next_task);
2214
2215 return ret;
2216}
2217
2218static void push_dl_tasks(struct rq *rq)
2219{
2220 /* push_dl_task() will return true if it moved a -deadline task */
2221 while (push_dl_task(rq))
2222 ;
2223}
2224
2225static void pull_dl_task(struct rq *this_rq)
2226{
2227 int this_cpu = this_rq->cpu, cpu;
2228 struct task_struct *p, *push_task;
2229 bool resched = false;
2230 struct rq *src_rq;
2231 u64 dmin = LONG_MAX;
2232
2233 if (likely(!dl_overloaded(this_rq)))
2234 return;
2235
2236 /*
2237 * Match the barrier from dl_set_overloaded; this guarantees that if we
2238 * see overloaded we must also see the dlo_mask bit.
2239 */
2240 smp_rmb();
2241
2242 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2243 if (this_cpu == cpu)
2244 continue;
2245
2246 src_rq = cpu_rq(cpu);
2247
2248 /*
2249 * It looks racy, abd it is! However, as in sched_rt.c,
2250 * we are fine with this.
2251 */
2252 if (this_rq->dl.dl_nr_running &&
2253 dl_time_before(this_rq->dl.earliest_dl.curr,
2254 src_rq->dl.earliest_dl.next))
2255 continue;
2256
2257 /* Might drop this_rq->lock */
2258 push_task = NULL;
2259 double_lock_balance(this_rq, src_rq);
2260
2261 /*
2262 * If there are no more pullable tasks on the
2263 * rq, we're done with it.
2264 */
2265 if (src_rq->dl.dl_nr_running <= 1)
2266 goto skip;
2267
2268 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2269
2270 /*
2271 * We found a task to be pulled if:
2272 * - it preempts our current (if there's one),
2273 * - it will preempt the last one we pulled (if any).
2274 */
2275 if (p && dl_time_before(p->dl.deadline, dmin) &&
2276 (!this_rq->dl.dl_nr_running ||
2277 dl_time_before(p->dl.deadline,
2278 this_rq->dl.earliest_dl.curr))) {
2279 WARN_ON(p == src_rq->curr);
2280 WARN_ON(!task_on_rq_queued(p));
2281
2282 /*
2283 * Then we pull iff p has actually an earlier
2284 * deadline than the current task of its runqueue.
2285 */
2286 if (dl_time_before(p->dl.deadline,
2287 src_rq->curr->dl.deadline))
2288 goto skip;
2289
2290 if (is_migration_disabled(p)) {
2291 push_task = get_push_task(src_rq);
2292 } else {
2293 deactivate_task(src_rq, p, 0);
2294 set_task_cpu(p, this_cpu);
2295 activate_task(this_rq, p, 0);
2296 dmin = p->dl.deadline;
2297 resched = true;
2298 }
2299
2300 /* Is there any other task even earlier? */
2301 }
2302skip:
2303 double_unlock_balance(this_rq, src_rq);
2304
2305 if (push_task) {
2306 raw_spin_rq_unlock(this_rq);
2307 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2308 push_task, &src_rq->push_work);
2309 raw_spin_rq_lock(this_rq);
2310 }
2311 }
2312
2313 if (resched)
2314 resched_curr(this_rq);
2315}
2316
2317/*
2318 * Since the task is not running and a reschedule is not going to happen
2319 * anytime soon on its runqueue, we try pushing it away now.
2320 */
2321static void task_woken_dl(struct rq *rq, struct task_struct *p)
2322{
2323 if (!task_running(rq, p) &&
2324 !test_tsk_need_resched(rq->curr) &&
2325 p->nr_cpus_allowed > 1 &&
2326 dl_task(rq->curr) &&
2327 (rq->curr->nr_cpus_allowed < 2 ||
2328 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2329 push_dl_tasks(rq);
2330 }
2331}
2332
2333static void set_cpus_allowed_dl(struct task_struct *p,
2334 const struct cpumask *new_mask,
2335 u32 flags)
2336{
2337 struct root_domain *src_rd;
2338 struct rq *rq;
2339
2340 BUG_ON(!dl_task(p));
2341
2342 rq = task_rq(p);
2343 src_rd = rq->rd;
2344 /*
2345 * Migrating a SCHED_DEADLINE task between exclusive
2346 * cpusets (different root_domains) entails a bandwidth
2347 * update. We already made space for us in the destination
2348 * domain (see cpuset_can_attach()).
2349 */
2350 if (!cpumask_intersects(src_rd->span, new_mask)) {
2351 struct dl_bw *src_dl_b;
2352
2353 src_dl_b = dl_bw_of(cpu_of(rq));
2354 /*
2355 * We now free resources of the root_domain we are migrating
2356 * off. In the worst case, sched_setattr() may temporary fail
2357 * until we complete the update.
2358 */
2359 raw_spin_lock(&src_dl_b->lock);
2360 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2361 raw_spin_unlock(&src_dl_b->lock);
2362 }
2363
2364 set_cpus_allowed_common(p, new_mask, flags);
2365}
2366
2367/* Assumes rq->lock is held */
2368static void rq_online_dl(struct rq *rq)
2369{
2370 if (rq->dl.overloaded)
2371 dl_set_overload(rq);
2372
2373 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2374 if (rq->dl.dl_nr_running > 0)
2375 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2376}
2377
2378/* Assumes rq->lock is held */
2379static void rq_offline_dl(struct rq *rq)
2380{
2381 if (rq->dl.overloaded)
2382 dl_clear_overload(rq);
2383
2384 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2385 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2386}
2387
2388void __init init_sched_dl_class(void)
2389{
2390 unsigned int i;
2391
2392 for_each_possible_cpu(i)
2393 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2394 GFP_KERNEL, cpu_to_node(i));
2395}
2396
2397void dl_add_task_root_domain(struct task_struct *p)
2398{
2399 struct rq_flags rf;
2400 struct rq *rq;
2401 struct dl_bw *dl_b;
2402
2403 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2404 if (!dl_task(p)) {
2405 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2406 return;
2407 }
2408
2409 rq = __task_rq_lock(p, &rf);
2410
2411 dl_b = &rq->rd->dl_bw;
2412 raw_spin_lock(&dl_b->lock);
2413
2414 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2415
2416 raw_spin_unlock(&dl_b->lock);
2417
2418 task_rq_unlock(rq, p, &rf);
2419}
2420
2421void dl_clear_root_domain(struct root_domain *rd)
2422{
2423 unsigned long flags;
2424
2425 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2426 rd->dl_bw.total_bw = 0;
2427 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2428}
2429
2430#endif /* CONFIG_SMP */
2431
2432static void switched_from_dl(struct rq *rq, struct task_struct *p)
2433{
2434 /*
2435 * task_non_contending() can start the "inactive timer" (if the 0-lag
2436 * time is in the future). If the task switches back to dl before
2437 * the "inactive timer" fires, it can continue to consume its current
2438 * runtime using its current deadline. If it stays outside of
2439 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2440 * will reset the task parameters.
2441 */
2442 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2443 task_non_contending(p);
2444
2445 if (!task_on_rq_queued(p)) {
2446 /*
2447 * Inactive timer is armed. However, p is leaving DEADLINE and
2448 * might migrate away from this rq while continuing to run on
2449 * some other class. We need to remove its contribution from
2450 * this rq running_bw now, or sub_rq_bw (below) will complain.
2451 */
2452 if (p->dl.dl_non_contending)
2453 sub_running_bw(&p->dl, &rq->dl);
2454 sub_rq_bw(&p->dl, &rq->dl);
2455 }
2456
2457 /*
2458 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2459 * at the 0-lag time, because the task could have been migrated
2460 * while SCHED_OTHER in the meanwhile.
2461 */
2462 if (p->dl.dl_non_contending)
2463 p->dl.dl_non_contending = 0;
2464
2465 /*
2466 * Since this might be the only -deadline task on the rq,
2467 * this is the right place to try to pull some other one
2468 * from an overloaded CPU, if any.
2469 */
2470 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2471 return;
2472
2473 deadline_queue_pull_task(rq);
2474}
2475
2476/*
2477 * When switching to -deadline, we may overload the rq, then
2478 * we try to push someone off, if possible.
2479 */
2480static void switched_to_dl(struct rq *rq, struct task_struct *p)
2481{
2482 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2483 put_task_struct(p);
2484
2485 /* If p is not queued we will update its parameters at next wakeup. */
2486 if (!task_on_rq_queued(p)) {
2487 add_rq_bw(&p->dl, &rq->dl);
2488
2489 return;
2490 }
2491
2492 if (rq->curr != p) {
2493#ifdef CONFIG_SMP
2494 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2495 deadline_queue_push_tasks(rq);
2496#endif
2497 if (dl_task(rq->curr))
2498 check_preempt_curr_dl(rq, p, 0);
2499 else
2500 resched_curr(rq);
2501 } else {
2502 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2503 }
2504}
2505
2506/*
2507 * If the scheduling parameters of a -deadline task changed,
2508 * a push or pull operation might be needed.
2509 */
2510static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2511 int oldprio)
2512{
2513 if (task_on_rq_queued(p) || task_current(rq, p)) {
2514#ifdef CONFIG_SMP
2515 /*
2516 * This might be too much, but unfortunately
2517 * we don't have the old deadline value, and
2518 * we can't argue if the task is increasing
2519 * or lowering its prio, so...
2520 */
2521 if (!rq->dl.overloaded)
2522 deadline_queue_pull_task(rq);
2523
2524 /*
2525 * If we now have a earlier deadline task than p,
2526 * then reschedule, provided p is still on this
2527 * runqueue.
2528 */
2529 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2530 resched_curr(rq);
2531#else
2532 /*
2533 * Again, we don't know if p has a earlier
2534 * or later deadline, so let's blindly set a
2535 * (maybe not needed) rescheduling point.
2536 */
2537 resched_curr(rq);
2538#endif /* CONFIG_SMP */
2539 }
2540}
2541
2542DEFINE_SCHED_CLASS(dl) = {
2543
2544 .enqueue_task = enqueue_task_dl,
2545 .dequeue_task = dequeue_task_dl,
2546 .yield_task = yield_task_dl,
2547
2548 .check_preempt_curr = check_preempt_curr_dl,
2549
2550 .pick_next_task = pick_next_task_dl,
2551 .put_prev_task = put_prev_task_dl,
2552 .set_next_task = set_next_task_dl,
2553
2554#ifdef CONFIG_SMP
2555 .balance = balance_dl,
2556 .pick_task = pick_task_dl,
2557 .select_task_rq = select_task_rq_dl,
2558 .migrate_task_rq = migrate_task_rq_dl,
2559 .set_cpus_allowed = set_cpus_allowed_dl,
2560 .rq_online = rq_online_dl,
2561 .rq_offline = rq_offline_dl,
2562 .task_woken = task_woken_dl,
2563 .find_lock_rq = find_lock_later_rq,
2564#endif
2565
2566 .task_tick = task_tick_dl,
2567 .task_fork = task_fork_dl,
2568
2569 .prio_changed = prio_changed_dl,
2570 .switched_from = switched_from_dl,
2571 .switched_to = switched_to_dl,
2572
2573 .update_curr = update_curr_dl,
2574};
2575
2576/* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2577static u64 dl_generation;
2578
2579int sched_dl_global_validate(void)
2580{
2581 u64 runtime = global_rt_runtime();
2582 u64 period = global_rt_period();
2583 u64 new_bw = to_ratio(period, runtime);
2584 u64 gen = ++dl_generation;
2585 struct dl_bw *dl_b;
2586 int cpu, cpus, ret = 0;
2587 unsigned long flags;
2588
2589 /*
2590 * Here we want to check the bandwidth not being set to some
2591 * value smaller than the currently allocated bandwidth in
2592 * any of the root_domains.
2593 */
2594 for_each_possible_cpu(cpu) {
2595 rcu_read_lock_sched();
2596
2597 if (dl_bw_visited(cpu, gen))
2598 goto next;
2599
2600 dl_b = dl_bw_of(cpu);
2601 cpus = dl_bw_cpus(cpu);
2602
2603 raw_spin_lock_irqsave(&dl_b->lock, flags);
2604 if (new_bw * cpus < dl_b->total_bw)
2605 ret = -EBUSY;
2606 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2607
2608next:
2609 rcu_read_unlock_sched();
2610
2611 if (ret)
2612 break;
2613 }
2614
2615 return ret;
2616}
2617
2618static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2619{
2620 if (global_rt_runtime() == RUNTIME_INF) {
2621 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2622 dl_rq->extra_bw = 1 << BW_SHIFT;
2623 } else {
2624 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2625 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2626 dl_rq->extra_bw = to_ratio(global_rt_period(),
2627 global_rt_runtime());
2628 }
2629}
2630
2631void sched_dl_do_global(void)
2632{
2633 u64 new_bw = -1;
2634 u64 gen = ++dl_generation;
2635 struct dl_bw *dl_b;
2636 int cpu;
2637 unsigned long flags;
2638
2639 def_dl_bandwidth.dl_period = global_rt_period();
2640 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2641
2642 if (global_rt_runtime() != RUNTIME_INF)
2643 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2644
2645 for_each_possible_cpu(cpu) {
2646 rcu_read_lock_sched();
2647
2648 if (dl_bw_visited(cpu, gen)) {
2649 rcu_read_unlock_sched();
2650 continue;
2651 }
2652
2653 dl_b = dl_bw_of(cpu);
2654
2655 raw_spin_lock_irqsave(&dl_b->lock, flags);
2656 dl_b->bw = new_bw;
2657 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2658
2659 rcu_read_unlock_sched();
2660 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2661 }
2662}
2663
2664/*
2665 * We must be sure that accepting a new task (or allowing changing the
2666 * parameters of an existing one) is consistent with the bandwidth
2667 * constraints. If yes, this function also accordingly updates the currently
2668 * allocated bandwidth to reflect the new situation.
2669 *
2670 * This function is called while holding p's rq->lock.
2671 */
2672int sched_dl_overflow(struct task_struct *p, int policy,
2673 const struct sched_attr *attr)
2674{
2675 u64 period = attr->sched_period ?: attr->sched_deadline;
2676 u64 runtime = attr->sched_runtime;
2677 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2678 int cpus, err = -1, cpu = task_cpu(p);
2679 struct dl_bw *dl_b = dl_bw_of(cpu);
2680 unsigned long cap;
2681
2682 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2683 return 0;
2684
2685 /* !deadline task may carry old deadline bandwidth */
2686 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2687 return 0;
2688
2689 /*
2690 * Either if a task, enters, leave, or stays -deadline but changes
2691 * its parameters, we may need to update accordingly the total
2692 * allocated bandwidth of the container.
2693 */
2694 raw_spin_lock(&dl_b->lock);
2695 cpus = dl_bw_cpus(cpu);
2696 cap = dl_bw_capacity(cpu);
2697
2698 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2699 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2700 if (hrtimer_active(&p->dl.inactive_timer))
2701 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2702 __dl_add(dl_b, new_bw, cpus);
2703 err = 0;
2704 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2705 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2706 /*
2707 * XXX this is slightly incorrect: when the task
2708 * utilization decreases, we should delay the total
2709 * utilization change until the task's 0-lag point.
2710 * But this would require to set the task's "inactive
2711 * timer" when the task is not inactive.
2712 */
2713 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2714 __dl_add(dl_b, new_bw, cpus);
2715 dl_change_utilization(p, new_bw);
2716 err = 0;
2717 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2718 /*
2719 * Do not decrease the total deadline utilization here,
2720 * switched_from_dl() will take care to do it at the correct
2721 * (0-lag) time.
2722 */
2723 err = 0;
2724 }
2725 raw_spin_unlock(&dl_b->lock);
2726
2727 return err;
2728}
2729
2730/*
2731 * This function initializes the sched_dl_entity of a newly becoming
2732 * SCHED_DEADLINE task.
2733 *
2734 * Only the static values are considered here, the actual runtime and the
2735 * absolute deadline will be properly calculated when the task is enqueued
2736 * for the first time with its new policy.
2737 */
2738void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2739{
2740 struct sched_dl_entity *dl_se = &p->dl;
2741
2742 dl_se->dl_runtime = attr->sched_runtime;
2743 dl_se->dl_deadline = attr->sched_deadline;
2744 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2745 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
2746 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2747 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2748}
2749
2750void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2751{
2752 struct sched_dl_entity *dl_se = &p->dl;
2753
2754 attr->sched_priority = p->rt_priority;
2755 attr->sched_runtime = dl_se->dl_runtime;
2756 attr->sched_deadline = dl_se->dl_deadline;
2757 attr->sched_period = dl_se->dl_period;
2758 attr->sched_flags &= ~SCHED_DL_FLAGS;
2759 attr->sched_flags |= dl_se->flags;
2760}
2761
2762/*
2763 * Default limits for DL period; on the top end we guard against small util
2764 * tasks still getting ridiculously long effective runtimes, on the bottom end we
2765 * guard against timer DoS.
2766 */
2767unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
2768unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
2769
2770/*
2771 * This function validates the new parameters of a -deadline task.
2772 * We ask for the deadline not being zero, and greater or equal
2773 * than the runtime, as well as the period of being zero or
2774 * greater than deadline. Furthermore, we have to be sure that
2775 * user parameters are above the internal resolution of 1us (we
2776 * check sched_runtime only since it is always the smaller one) and
2777 * below 2^63 ns (we have to check both sched_deadline and
2778 * sched_period, as the latter can be zero).
2779 */
2780bool __checkparam_dl(const struct sched_attr *attr)
2781{
2782 u64 period, max, min;
2783
2784 /* special dl tasks don't actually use any parameter */
2785 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2786 return true;
2787
2788 /* deadline != 0 */
2789 if (attr->sched_deadline == 0)
2790 return false;
2791
2792 /*
2793 * Since we truncate DL_SCALE bits, make sure we're at least
2794 * that big.
2795 */
2796 if (attr->sched_runtime < (1ULL << DL_SCALE))
2797 return false;
2798
2799 /*
2800 * Since we use the MSB for wrap-around and sign issues, make
2801 * sure it's not set (mind that period can be equal to zero).
2802 */
2803 if (attr->sched_deadline & (1ULL << 63) ||
2804 attr->sched_period & (1ULL << 63))
2805 return false;
2806
2807 period = attr->sched_period;
2808 if (!period)
2809 period = attr->sched_deadline;
2810
2811 /* runtime <= deadline <= period (if period != 0) */
2812 if (period < attr->sched_deadline ||
2813 attr->sched_deadline < attr->sched_runtime)
2814 return false;
2815
2816 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2817 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2818
2819 if (period < min || period > max)
2820 return false;
2821
2822 return true;
2823}
2824
2825/*
2826 * This function clears the sched_dl_entity static params.
2827 */
2828void __dl_clear_params(struct task_struct *p)
2829{
2830 struct sched_dl_entity *dl_se = &p->dl;
2831
2832 dl_se->dl_runtime = 0;
2833 dl_se->dl_deadline = 0;
2834 dl_se->dl_period = 0;
2835 dl_se->flags = 0;
2836 dl_se->dl_bw = 0;
2837 dl_se->dl_density = 0;
2838
2839 dl_se->dl_throttled = 0;
2840 dl_se->dl_yielded = 0;
2841 dl_se->dl_non_contending = 0;
2842 dl_se->dl_overrun = 0;
2843
2844#ifdef CONFIG_RT_MUTEXES
2845 dl_se->pi_se = dl_se;
2846#endif
2847}
2848
2849bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2850{
2851 struct sched_dl_entity *dl_se = &p->dl;
2852
2853 if (dl_se->dl_runtime != attr->sched_runtime ||
2854 dl_se->dl_deadline != attr->sched_deadline ||
2855 dl_se->dl_period != attr->sched_period ||
2856 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
2857 return true;
2858
2859 return false;
2860}
2861
2862#ifdef CONFIG_SMP
2863int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2864{
2865 unsigned long flags, cap;
2866 unsigned int dest_cpu;
2867 struct dl_bw *dl_b;
2868 bool overflow;
2869 int ret;
2870
2871 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2872
2873 rcu_read_lock_sched();
2874 dl_b = dl_bw_of(dest_cpu);
2875 raw_spin_lock_irqsave(&dl_b->lock, flags);
2876 cap = dl_bw_capacity(dest_cpu);
2877 overflow = __dl_overflow(dl_b, cap, 0, p->dl.dl_bw);
2878 if (overflow) {
2879 ret = -EBUSY;
2880 } else {
2881 /*
2882 * We reserve space for this task in the destination
2883 * root_domain, as we can't fail after this point.
2884 * We will free resources in the source root_domain
2885 * later on (see set_cpus_allowed_dl()).
2886 */
2887 int cpus = dl_bw_cpus(dest_cpu);
2888
2889 __dl_add(dl_b, p->dl.dl_bw, cpus);
2890 ret = 0;
2891 }
2892 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2893 rcu_read_unlock_sched();
2894
2895 return ret;
2896}
2897
2898int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2899 const struct cpumask *trial)
2900{
2901 int ret = 1, trial_cpus;
2902 struct dl_bw *cur_dl_b;
2903 unsigned long flags;
2904
2905 rcu_read_lock_sched();
2906 cur_dl_b = dl_bw_of(cpumask_any(cur));
2907 trial_cpus = cpumask_weight(trial);
2908
2909 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2910 if (cur_dl_b->bw != -1 &&
2911 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2912 ret = 0;
2913 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2914 rcu_read_unlock_sched();
2915
2916 return ret;
2917}
2918
2919bool dl_cpu_busy(unsigned int cpu)
2920{
2921 unsigned long flags, cap;
2922 struct dl_bw *dl_b;
2923 bool overflow;
2924
2925 rcu_read_lock_sched();
2926 dl_b = dl_bw_of(cpu);
2927 raw_spin_lock_irqsave(&dl_b->lock, flags);
2928 cap = dl_bw_capacity(cpu);
2929 overflow = __dl_overflow(dl_b, cap, 0, 0);
2930 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2931 rcu_read_unlock_sched();
2932
2933 return overflow;
2934}
2935#endif
2936
2937#ifdef CONFIG_SCHED_DEBUG
2938void print_dl_stats(struct seq_file *m, int cpu)
2939{
2940 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2941}
2942#endif /* CONFIG_SCHED_DEBUG */
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Deadline Scheduling Class (SCHED_DEADLINE)
4 *
5 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6 *
7 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't affect any other task.
12 *
13 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fchecconi@gmail.com>
17 */
18
19/*
20 * Default limits for DL period; on the top end we guard against small util
21 * tasks still getting ridiculously long effective runtimes, on the bottom end we
22 * guard against timer DoS.
23 */
24static unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
25static unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
26#ifdef CONFIG_SYSCTL
27static struct ctl_table sched_dl_sysctls[] = {
28 {
29 .procname = "sched_deadline_period_max_us",
30 .data = &sysctl_sched_dl_period_max,
31 .maxlen = sizeof(unsigned int),
32 .mode = 0644,
33 .proc_handler = proc_douintvec_minmax,
34 .extra1 = (void *)&sysctl_sched_dl_period_min,
35 },
36 {
37 .procname = "sched_deadline_period_min_us",
38 .data = &sysctl_sched_dl_period_min,
39 .maxlen = sizeof(unsigned int),
40 .mode = 0644,
41 .proc_handler = proc_douintvec_minmax,
42 .extra2 = (void *)&sysctl_sched_dl_period_max,
43 },
44 {}
45};
46
47static int __init sched_dl_sysctl_init(void)
48{
49 register_sysctl_init("kernel", sched_dl_sysctls);
50 return 0;
51}
52late_initcall(sched_dl_sysctl_init);
53#endif
54
55static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
56{
57 return container_of(dl_se, struct task_struct, dl);
58}
59
60static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
61{
62 return container_of(dl_rq, struct rq, dl);
63}
64
65static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
66{
67 struct task_struct *p = dl_task_of(dl_se);
68 struct rq *rq = task_rq(p);
69
70 return &rq->dl;
71}
72
73static inline int on_dl_rq(struct sched_dl_entity *dl_se)
74{
75 return !RB_EMPTY_NODE(&dl_se->rb_node);
76}
77
78#ifdef CONFIG_RT_MUTEXES
79static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
80{
81 return dl_se->pi_se;
82}
83
84static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
85{
86 return pi_of(dl_se) != dl_se;
87}
88#else
89static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
90{
91 return dl_se;
92}
93
94static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
95{
96 return false;
97}
98#endif
99
100#ifdef CONFIG_SMP
101static inline struct dl_bw *dl_bw_of(int i)
102{
103 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
104 "sched RCU must be held");
105 return &cpu_rq(i)->rd->dl_bw;
106}
107
108static inline int dl_bw_cpus(int i)
109{
110 struct root_domain *rd = cpu_rq(i)->rd;
111 int cpus;
112
113 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
114 "sched RCU must be held");
115
116 if (cpumask_subset(rd->span, cpu_active_mask))
117 return cpumask_weight(rd->span);
118
119 cpus = 0;
120
121 for_each_cpu_and(i, rd->span, cpu_active_mask)
122 cpus++;
123
124 return cpus;
125}
126
127static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
128{
129 unsigned long cap = 0;
130 int i;
131
132 for_each_cpu_and(i, mask, cpu_active_mask)
133 cap += capacity_orig_of(i);
134
135 return cap;
136}
137
138/*
139 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
140 * of the CPU the task is running on rather rd's \Sum CPU capacity.
141 */
142static inline unsigned long dl_bw_capacity(int i)
143{
144 if (!sched_asym_cpucap_active() &&
145 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
146 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
147 } else {
148 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
149 "sched RCU must be held");
150
151 return __dl_bw_capacity(cpu_rq(i)->rd->span);
152 }
153}
154
155static inline bool dl_bw_visited(int cpu, u64 gen)
156{
157 struct root_domain *rd = cpu_rq(cpu)->rd;
158
159 if (rd->visit_gen == gen)
160 return true;
161
162 rd->visit_gen = gen;
163 return false;
164}
165
166static inline
167void __dl_update(struct dl_bw *dl_b, s64 bw)
168{
169 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
170 int i;
171
172 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
173 "sched RCU must be held");
174 for_each_cpu_and(i, rd->span, cpu_active_mask) {
175 struct rq *rq = cpu_rq(i);
176
177 rq->dl.extra_bw += bw;
178 }
179}
180#else
181static inline struct dl_bw *dl_bw_of(int i)
182{
183 return &cpu_rq(i)->dl.dl_bw;
184}
185
186static inline int dl_bw_cpus(int i)
187{
188 return 1;
189}
190
191static inline unsigned long dl_bw_capacity(int i)
192{
193 return SCHED_CAPACITY_SCALE;
194}
195
196static inline bool dl_bw_visited(int cpu, u64 gen)
197{
198 return false;
199}
200
201static inline
202void __dl_update(struct dl_bw *dl_b, s64 bw)
203{
204 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
205
206 dl->extra_bw += bw;
207}
208#endif
209
210static inline
211void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
212{
213 dl_b->total_bw -= tsk_bw;
214 __dl_update(dl_b, (s32)tsk_bw / cpus);
215}
216
217static inline
218void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
219{
220 dl_b->total_bw += tsk_bw;
221 __dl_update(dl_b, -((s32)tsk_bw / cpus));
222}
223
224static inline bool
225__dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
226{
227 return dl_b->bw != -1 &&
228 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
229}
230
231static inline
232void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
233{
234 u64 old = dl_rq->running_bw;
235
236 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
237 dl_rq->running_bw += dl_bw;
238 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
239 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
240 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
241 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
242}
243
244static inline
245void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
246{
247 u64 old = dl_rq->running_bw;
248
249 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
250 dl_rq->running_bw -= dl_bw;
251 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
252 if (dl_rq->running_bw > old)
253 dl_rq->running_bw = 0;
254 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
255 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
256}
257
258static inline
259void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
260{
261 u64 old = dl_rq->this_bw;
262
263 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
264 dl_rq->this_bw += dl_bw;
265 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
266}
267
268static inline
269void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
270{
271 u64 old = dl_rq->this_bw;
272
273 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
274 dl_rq->this_bw -= dl_bw;
275 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
276 if (dl_rq->this_bw > old)
277 dl_rq->this_bw = 0;
278 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
279}
280
281static inline
282void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
283{
284 if (!dl_entity_is_special(dl_se))
285 __add_rq_bw(dl_se->dl_bw, dl_rq);
286}
287
288static inline
289void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
290{
291 if (!dl_entity_is_special(dl_se))
292 __sub_rq_bw(dl_se->dl_bw, dl_rq);
293}
294
295static inline
296void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
297{
298 if (!dl_entity_is_special(dl_se))
299 __add_running_bw(dl_se->dl_bw, dl_rq);
300}
301
302static inline
303void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
304{
305 if (!dl_entity_is_special(dl_se))
306 __sub_running_bw(dl_se->dl_bw, dl_rq);
307}
308
309static void dl_change_utilization(struct task_struct *p, u64 new_bw)
310{
311 struct rq *rq;
312
313 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
314
315 if (task_on_rq_queued(p))
316 return;
317
318 rq = task_rq(p);
319 if (p->dl.dl_non_contending) {
320 sub_running_bw(&p->dl, &rq->dl);
321 p->dl.dl_non_contending = 0;
322 /*
323 * If the timer handler is currently running and the
324 * timer cannot be canceled, inactive_task_timer()
325 * will see that dl_not_contending is not set, and
326 * will not touch the rq's active utilization,
327 * so we are still safe.
328 */
329 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
330 put_task_struct(p);
331 }
332 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
333 __add_rq_bw(new_bw, &rq->dl);
334}
335
336/*
337 * The utilization of a task cannot be immediately removed from
338 * the rq active utilization (running_bw) when the task blocks.
339 * Instead, we have to wait for the so called "0-lag time".
340 *
341 * If a task blocks before the "0-lag time", a timer (the inactive
342 * timer) is armed, and running_bw is decreased when the timer
343 * fires.
344 *
345 * If the task wakes up again before the inactive timer fires,
346 * the timer is canceled, whereas if the task wakes up after the
347 * inactive timer fired (and running_bw has been decreased) the
348 * task's utilization has to be added to running_bw again.
349 * A flag in the deadline scheduling entity (dl_non_contending)
350 * is used to avoid race conditions between the inactive timer handler
351 * and task wakeups.
352 *
353 * The following diagram shows how running_bw is updated. A task is
354 * "ACTIVE" when its utilization contributes to running_bw; an
355 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
356 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
357 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
358 * time already passed, which does not contribute to running_bw anymore.
359 * +------------------+
360 * wakeup | ACTIVE |
361 * +------------------>+ contending |
362 * | add_running_bw | |
363 * | +----+------+------+
364 * | | ^
365 * | dequeue | |
366 * +--------+-------+ | |
367 * | | t >= 0-lag | | wakeup
368 * | INACTIVE |<---------------+ |
369 * | | sub_running_bw | |
370 * +--------+-------+ | |
371 * ^ | |
372 * | t < 0-lag | |
373 * | | |
374 * | V |
375 * | +----+------+------+
376 * | sub_running_bw | ACTIVE |
377 * +-------------------+ |
378 * inactive timer | non contending |
379 * fired +------------------+
380 *
381 * The task_non_contending() function is invoked when a task
382 * blocks, and checks if the 0-lag time already passed or
383 * not (in the first case, it directly updates running_bw;
384 * in the second case, it arms the inactive timer).
385 *
386 * The task_contending() function is invoked when a task wakes
387 * up, and checks if the task is still in the "ACTIVE non contending"
388 * state or not (in the second case, it updates running_bw).
389 */
390static void task_non_contending(struct task_struct *p)
391{
392 struct sched_dl_entity *dl_se = &p->dl;
393 struct hrtimer *timer = &dl_se->inactive_timer;
394 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
395 struct rq *rq = rq_of_dl_rq(dl_rq);
396 s64 zerolag_time;
397
398 /*
399 * If this is a non-deadline task that has been boosted,
400 * do nothing
401 */
402 if (dl_se->dl_runtime == 0)
403 return;
404
405 if (dl_entity_is_special(dl_se))
406 return;
407
408 WARN_ON(dl_se->dl_non_contending);
409
410 zerolag_time = dl_se->deadline -
411 div64_long((dl_se->runtime * dl_se->dl_period),
412 dl_se->dl_runtime);
413
414 /*
415 * Using relative times instead of the absolute "0-lag time"
416 * allows to simplify the code
417 */
418 zerolag_time -= rq_clock(rq);
419
420 /*
421 * If the "0-lag time" already passed, decrease the active
422 * utilization now, instead of starting a timer
423 */
424 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
425 if (dl_task(p))
426 sub_running_bw(dl_se, dl_rq);
427 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
428 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
429
430 if (READ_ONCE(p->__state) == TASK_DEAD)
431 sub_rq_bw(&p->dl, &rq->dl);
432 raw_spin_lock(&dl_b->lock);
433 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
434 raw_spin_unlock(&dl_b->lock);
435 __dl_clear_params(p);
436 }
437
438 return;
439 }
440
441 dl_se->dl_non_contending = 1;
442 get_task_struct(p);
443 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
444}
445
446static void task_contending(struct sched_dl_entity *dl_se, int flags)
447{
448 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
449
450 /*
451 * If this is a non-deadline task that has been boosted,
452 * do nothing
453 */
454 if (dl_se->dl_runtime == 0)
455 return;
456
457 if (flags & ENQUEUE_MIGRATED)
458 add_rq_bw(dl_se, dl_rq);
459
460 if (dl_se->dl_non_contending) {
461 dl_se->dl_non_contending = 0;
462 /*
463 * If the timer handler is currently running and the
464 * timer cannot be canceled, inactive_task_timer()
465 * will see that dl_not_contending is not set, and
466 * will not touch the rq's active utilization,
467 * so we are still safe.
468 */
469 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
470 put_task_struct(dl_task_of(dl_se));
471 } else {
472 /*
473 * Since "dl_non_contending" is not set, the
474 * task's utilization has already been removed from
475 * active utilization (either when the task blocked,
476 * when the "inactive timer" fired).
477 * So, add it back.
478 */
479 add_running_bw(dl_se, dl_rq);
480 }
481}
482
483static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
484{
485 struct sched_dl_entity *dl_se = &p->dl;
486
487 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
488}
489
490static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
491
492void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
493{
494 raw_spin_lock_init(&dl_b->dl_runtime_lock);
495 dl_b->dl_period = period;
496 dl_b->dl_runtime = runtime;
497}
498
499void init_dl_bw(struct dl_bw *dl_b)
500{
501 raw_spin_lock_init(&dl_b->lock);
502 if (global_rt_runtime() == RUNTIME_INF)
503 dl_b->bw = -1;
504 else
505 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
506 dl_b->total_bw = 0;
507}
508
509void init_dl_rq(struct dl_rq *dl_rq)
510{
511 dl_rq->root = RB_ROOT_CACHED;
512
513#ifdef CONFIG_SMP
514 /* zero means no -deadline tasks */
515 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
516
517 dl_rq->dl_nr_migratory = 0;
518 dl_rq->overloaded = 0;
519 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
520#else
521 init_dl_bw(&dl_rq->dl_bw);
522#endif
523
524 dl_rq->running_bw = 0;
525 dl_rq->this_bw = 0;
526 init_dl_rq_bw_ratio(dl_rq);
527}
528
529#ifdef CONFIG_SMP
530
531static inline int dl_overloaded(struct rq *rq)
532{
533 return atomic_read(&rq->rd->dlo_count);
534}
535
536static inline void dl_set_overload(struct rq *rq)
537{
538 if (!rq->online)
539 return;
540
541 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
542 /*
543 * Must be visible before the overload count is
544 * set (as in sched_rt.c).
545 *
546 * Matched by the barrier in pull_dl_task().
547 */
548 smp_wmb();
549 atomic_inc(&rq->rd->dlo_count);
550}
551
552static inline void dl_clear_overload(struct rq *rq)
553{
554 if (!rq->online)
555 return;
556
557 atomic_dec(&rq->rd->dlo_count);
558 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
559}
560
561static void update_dl_migration(struct dl_rq *dl_rq)
562{
563 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
564 if (!dl_rq->overloaded) {
565 dl_set_overload(rq_of_dl_rq(dl_rq));
566 dl_rq->overloaded = 1;
567 }
568 } else if (dl_rq->overloaded) {
569 dl_clear_overload(rq_of_dl_rq(dl_rq));
570 dl_rq->overloaded = 0;
571 }
572}
573
574static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
575{
576 struct task_struct *p = dl_task_of(dl_se);
577
578 if (p->nr_cpus_allowed > 1)
579 dl_rq->dl_nr_migratory++;
580
581 update_dl_migration(dl_rq);
582}
583
584static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
585{
586 struct task_struct *p = dl_task_of(dl_se);
587
588 if (p->nr_cpus_allowed > 1)
589 dl_rq->dl_nr_migratory--;
590
591 update_dl_migration(dl_rq);
592}
593
594#define __node_2_pdl(node) \
595 rb_entry((node), struct task_struct, pushable_dl_tasks)
596
597static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
598{
599 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
600}
601
602/*
603 * The list of pushable -deadline task is not a plist, like in
604 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
605 */
606static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
607{
608 struct rb_node *leftmost;
609
610 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
611
612 leftmost = rb_add_cached(&p->pushable_dl_tasks,
613 &rq->dl.pushable_dl_tasks_root,
614 __pushable_less);
615 if (leftmost)
616 rq->dl.earliest_dl.next = p->dl.deadline;
617}
618
619static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
620{
621 struct dl_rq *dl_rq = &rq->dl;
622 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
623 struct rb_node *leftmost;
624
625 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
626 return;
627
628 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
629 if (leftmost)
630 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
631
632 RB_CLEAR_NODE(&p->pushable_dl_tasks);
633}
634
635static inline int has_pushable_dl_tasks(struct rq *rq)
636{
637 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
638}
639
640static int push_dl_task(struct rq *rq);
641
642static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
643{
644 return rq->online && dl_task(prev);
645}
646
647static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
648static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
649
650static void push_dl_tasks(struct rq *);
651static void pull_dl_task(struct rq *);
652
653static inline void deadline_queue_push_tasks(struct rq *rq)
654{
655 if (!has_pushable_dl_tasks(rq))
656 return;
657
658 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
659}
660
661static inline void deadline_queue_pull_task(struct rq *rq)
662{
663 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
664}
665
666static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
667
668static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
669{
670 struct rq *later_rq = NULL;
671 struct dl_bw *dl_b;
672
673 later_rq = find_lock_later_rq(p, rq);
674 if (!later_rq) {
675 int cpu;
676
677 /*
678 * If we cannot preempt any rq, fall back to pick any
679 * online CPU:
680 */
681 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
682 if (cpu >= nr_cpu_ids) {
683 /*
684 * Failed to find any suitable CPU.
685 * The task will never come back!
686 */
687 WARN_ON_ONCE(dl_bandwidth_enabled());
688
689 /*
690 * If admission control is disabled we
691 * try a little harder to let the task
692 * run.
693 */
694 cpu = cpumask_any(cpu_active_mask);
695 }
696 later_rq = cpu_rq(cpu);
697 double_lock_balance(rq, later_rq);
698 }
699
700 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
701 /*
702 * Inactive timer is armed (or callback is running, but
703 * waiting for us to release rq locks). In any case, when it
704 * will fire (or continue), it will see running_bw of this
705 * task migrated to later_rq (and correctly handle it).
706 */
707 sub_running_bw(&p->dl, &rq->dl);
708 sub_rq_bw(&p->dl, &rq->dl);
709
710 add_rq_bw(&p->dl, &later_rq->dl);
711 add_running_bw(&p->dl, &later_rq->dl);
712 } else {
713 sub_rq_bw(&p->dl, &rq->dl);
714 add_rq_bw(&p->dl, &later_rq->dl);
715 }
716
717 /*
718 * And we finally need to fixup root_domain(s) bandwidth accounting,
719 * since p is still hanging out in the old (now moved to default) root
720 * domain.
721 */
722 dl_b = &rq->rd->dl_bw;
723 raw_spin_lock(&dl_b->lock);
724 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
725 raw_spin_unlock(&dl_b->lock);
726
727 dl_b = &later_rq->rd->dl_bw;
728 raw_spin_lock(&dl_b->lock);
729 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
730 raw_spin_unlock(&dl_b->lock);
731
732 set_task_cpu(p, later_rq->cpu);
733 double_unlock_balance(later_rq, rq);
734
735 return later_rq;
736}
737
738#else
739
740static inline
741void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
742{
743}
744
745static inline
746void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
747{
748}
749
750static inline
751void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
752{
753}
754
755static inline
756void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
757{
758}
759
760static inline void deadline_queue_push_tasks(struct rq *rq)
761{
762}
763
764static inline void deadline_queue_pull_task(struct rq *rq)
765{
766}
767#endif /* CONFIG_SMP */
768
769static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
770static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
771static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
772
773static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
774 struct rq *rq)
775{
776 /* for non-boosted task, pi_of(dl_se) == dl_se */
777 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
778 dl_se->runtime = pi_of(dl_se)->dl_runtime;
779}
780
781/*
782 * We are being explicitly informed that a new instance is starting,
783 * and this means that:
784 * - the absolute deadline of the entity has to be placed at
785 * current time + relative deadline;
786 * - the runtime of the entity has to be set to the maximum value.
787 *
788 * The capability of specifying such event is useful whenever a -deadline
789 * entity wants to (try to!) synchronize its behaviour with the scheduler's
790 * one, and to (try to!) reconcile itself with its own scheduling
791 * parameters.
792 */
793static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
794{
795 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
796 struct rq *rq = rq_of_dl_rq(dl_rq);
797
798 WARN_ON(is_dl_boosted(dl_se));
799 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
800
801 /*
802 * We are racing with the deadline timer. So, do nothing because
803 * the deadline timer handler will take care of properly recharging
804 * the runtime and postponing the deadline
805 */
806 if (dl_se->dl_throttled)
807 return;
808
809 /*
810 * We use the regular wall clock time to set deadlines in the
811 * future; in fact, we must consider execution overheads (time
812 * spent on hardirq context, etc.).
813 */
814 replenish_dl_new_period(dl_se, rq);
815}
816
817/*
818 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
819 * possibility of a entity lasting more than what it declared, and thus
820 * exhausting its runtime.
821 *
822 * Here we are interested in making runtime overrun possible, but we do
823 * not want a entity which is misbehaving to affect the scheduling of all
824 * other entities.
825 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
826 * is used, in order to confine each entity within its own bandwidth.
827 *
828 * This function deals exactly with that, and ensures that when the runtime
829 * of a entity is replenished, its deadline is also postponed. That ensures
830 * the overrunning entity can't interfere with other entity in the system and
831 * can't make them miss their deadlines. Reasons why this kind of overruns
832 * could happen are, typically, a entity voluntarily trying to overcome its
833 * runtime, or it just underestimated it during sched_setattr().
834 */
835static void replenish_dl_entity(struct sched_dl_entity *dl_se)
836{
837 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
838 struct rq *rq = rq_of_dl_rq(dl_rq);
839
840 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
841
842 /*
843 * This could be the case for a !-dl task that is boosted.
844 * Just go with full inherited parameters.
845 */
846 if (dl_se->dl_deadline == 0)
847 replenish_dl_new_period(dl_se, rq);
848
849 if (dl_se->dl_yielded && dl_se->runtime > 0)
850 dl_se->runtime = 0;
851
852 /*
853 * We keep moving the deadline away until we get some
854 * available runtime for the entity. This ensures correct
855 * handling of situations where the runtime overrun is
856 * arbitrary large.
857 */
858 while (dl_se->runtime <= 0) {
859 dl_se->deadline += pi_of(dl_se)->dl_period;
860 dl_se->runtime += pi_of(dl_se)->dl_runtime;
861 }
862
863 /*
864 * At this point, the deadline really should be "in
865 * the future" with respect to rq->clock. If it's
866 * not, we are, for some reason, lagging too much!
867 * Anyway, after having warn userspace abut that,
868 * we still try to keep the things running by
869 * resetting the deadline and the budget of the
870 * entity.
871 */
872 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
873 printk_deferred_once("sched: DL replenish lagged too much\n");
874 replenish_dl_new_period(dl_se, rq);
875 }
876
877 if (dl_se->dl_yielded)
878 dl_se->dl_yielded = 0;
879 if (dl_se->dl_throttled)
880 dl_se->dl_throttled = 0;
881}
882
883/*
884 * Here we check if --at time t-- an entity (which is probably being
885 * [re]activated or, in general, enqueued) can use its remaining runtime
886 * and its current deadline _without_ exceeding the bandwidth it is
887 * assigned (function returns true if it can't). We are in fact applying
888 * one of the CBS rules: when a task wakes up, if the residual runtime
889 * over residual deadline fits within the allocated bandwidth, then we
890 * can keep the current (absolute) deadline and residual budget without
891 * disrupting the schedulability of the system. Otherwise, we should
892 * refill the runtime and set the deadline a period in the future,
893 * because keeping the current (absolute) deadline of the task would
894 * result in breaking guarantees promised to other tasks (refer to
895 * Documentation/scheduler/sched-deadline.rst for more information).
896 *
897 * This function returns true if:
898 *
899 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
900 *
901 * IOW we can't recycle current parameters.
902 *
903 * Notice that the bandwidth check is done against the deadline. For
904 * task with deadline equal to period this is the same of using
905 * dl_period instead of dl_deadline in the equation above.
906 */
907static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
908{
909 u64 left, right;
910
911 /*
912 * left and right are the two sides of the equation above,
913 * after a bit of shuffling to use multiplications instead
914 * of divisions.
915 *
916 * Note that none of the time values involved in the two
917 * multiplications are absolute: dl_deadline and dl_runtime
918 * are the relative deadline and the maximum runtime of each
919 * instance, runtime is the runtime left for the last instance
920 * and (deadline - t), since t is rq->clock, is the time left
921 * to the (absolute) deadline. Even if overflowing the u64 type
922 * is very unlikely to occur in both cases, here we scale down
923 * as we want to avoid that risk at all. Scaling down by 10
924 * means that we reduce granularity to 1us. We are fine with it,
925 * since this is only a true/false check and, anyway, thinking
926 * of anything below microseconds resolution is actually fiction
927 * (but still we want to give the user that illusion >;).
928 */
929 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
930 right = ((dl_se->deadline - t) >> DL_SCALE) *
931 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
932
933 return dl_time_before(right, left);
934}
935
936/*
937 * Revised wakeup rule [1]: For self-suspending tasks, rather then
938 * re-initializing task's runtime and deadline, the revised wakeup
939 * rule adjusts the task's runtime to avoid the task to overrun its
940 * density.
941 *
942 * Reasoning: a task may overrun the density if:
943 * runtime / (deadline - t) > dl_runtime / dl_deadline
944 *
945 * Therefore, runtime can be adjusted to:
946 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
947 *
948 * In such way that runtime will be equal to the maximum density
949 * the task can use without breaking any rule.
950 *
951 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
952 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
953 */
954static void
955update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
956{
957 u64 laxity = dl_se->deadline - rq_clock(rq);
958
959 /*
960 * If the task has deadline < period, and the deadline is in the past,
961 * it should already be throttled before this check.
962 *
963 * See update_dl_entity() comments for further details.
964 */
965 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
966
967 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
968}
969
970/*
971 * Regarding the deadline, a task with implicit deadline has a relative
972 * deadline == relative period. A task with constrained deadline has a
973 * relative deadline <= relative period.
974 *
975 * We support constrained deadline tasks. However, there are some restrictions
976 * applied only for tasks which do not have an implicit deadline. See
977 * update_dl_entity() to know more about such restrictions.
978 *
979 * The dl_is_implicit() returns true if the task has an implicit deadline.
980 */
981static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
982{
983 return dl_se->dl_deadline == dl_se->dl_period;
984}
985
986/*
987 * When a deadline entity is placed in the runqueue, its runtime and deadline
988 * might need to be updated. This is done by a CBS wake up rule. There are two
989 * different rules: 1) the original CBS; and 2) the Revisited CBS.
990 *
991 * When the task is starting a new period, the Original CBS is used. In this
992 * case, the runtime is replenished and a new absolute deadline is set.
993 *
994 * When a task is queued before the begin of the next period, using the
995 * remaining runtime and deadline could make the entity to overflow, see
996 * dl_entity_overflow() to find more about runtime overflow. When such case
997 * is detected, the runtime and deadline need to be updated.
998 *
999 * If the task has an implicit deadline, i.e., deadline == period, the Original
1000 * CBS is applied. the runtime is replenished and a new absolute deadline is
1001 * set, as in the previous cases.
1002 *
1003 * However, the Original CBS does not work properly for tasks with
1004 * deadline < period, which are said to have a constrained deadline. By
1005 * applying the Original CBS, a constrained deadline task would be able to run
1006 * runtime/deadline in a period. With deadline < period, the task would
1007 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1008 *
1009 * In order to prevent this misbehave, the Revisited CBS is used for
1010 * constrained deadline tasks when a runtime overflow is detected. In the
1011 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1012 * the remaining runtime of the task is reduced to avoid runtime overflow.
1013 * Please refer to the comments update_dl_revised_wakeup() function to find
1014 * more about the Revised CBS rule.
1015 */
1016static void update_dl_entity(struct sched_dl_entity *dl_se)
1017{
1018 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1019 struct rq *rq = rq_of_dl_rq(dl_rq);
1020
1021 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1022 dl_entity_overflow(dl_se, rq_clock(rq))) {
1023
1024 if (unlikely(!dl_is_implicit(dl_se) &&
1025 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1026 !is_dl_boosted(dl_se))) {
1027 update_dl_revised_wakeup(dl_se, rq);
1028 return;
1029 }
1030
1031 replenish_dl_new_period(dl_se, rq);
1032 }
1033}
1034
1035static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1036{
1037 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1038}
1039
1040/*
1041 * If the entity depleted all its runtime, and if we want it to sleep
1042 * while waiting for some new execution time to become available, we
1043 * set the bandwidth replenishment timer to the replenishment instant
1044 * and try to activate it.
1045 *
1046 * Notice that it is important for the caller to know if the timer
1047 * actually started or not (i.e., the replenishment instant is in
1048 * the future or in the past).
1049 */
1050static int start_dl_timer(struct task_struct *p)
1051{
1052 struct sched_dl_entity *dl_se = &p->dl;
1053 struct hrtimer *timer = &dl_se->dl_timer;
1054 struct rq *rq = task_rq(p);
1055 ktime_t now, act;
1056 s64 delta;
1057
1058 lockdep_assert_rq_held(rq);
1059
1060 /*
1061 * We want the timer to fire at the deadline, but considering
1062 * that it is actually coming from rq->clock and not from
1063 * hrtimer's time base reading.
1064 */
1065 act = ns_to_ktime(dl_next_period(dl_se));
1066 now = hrtimer_cb_get_time(timer);
1067 delta = ktime_to_ns(now) - rq_clock(rq);
1068 act = ktime_add_ns(act, delta);
1069
1070 /*
1071 * If the expiry time already passed, e.g., because the value
1072 * chosen as the deadline is too small, don't even try to
1073 * start the timer in the past!
1074 */
1075 if (ktime_us_delta(act, now) < 0)
1076 return 0;
1077
1078 /*
1079 * !enqueued will guarantee another callback; even if one is already in
1080 * progress. This ensures a balanced {get,put}_task_struct().
1081 *
1082 * The race against __run_timer() clearing the enqueued state is
1083 * harmless because we're holding task_rq()->lock, therefore the timer
1084 * expiring after we've done the check will wait on its task_rq_lock()
1085 * and observe our state.
1086 */
1087 if (!hrtimer_is_queued(timer)) {
1088 get_task_struct(p);
1089 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1090 }
1091
1092 return 1;
1093}
1094
1095/*
1096 * This is the bandwidth enforcement timer callback. If here, we know
1097 * a task is not on its dl_rq, since the fact that the timer was running
1098 * means the task is throttled and needs a runtime replenishment.
1099 *
1100 * However, what we actually do depends on the fact the task is active,
1101 * (it is on its rq) or has been removed from there by a call to
1102 * dequeue_task_dl(). In the former case we must issue the runtime
1103 * replenishment and add the task back to the dl_rq; in the latter, we just
1104 * do nothing but clearing dl_throttled, so that runtime and deadline
1105 * updating (and the queueing back to dl_rq) will be done by the
1106 * next call to enqueue_task_dl().
1107 */
1108static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1109{
1110 struct sched_dl_entity *dl_se = container_of(timer,
1111 struct sched_dl_entity,
1112 dl_timer);
1113 struct task_struct *p = dl_task_of(dl_se);
1114 struct rq_flags rf;
1115 struct rq *rq;
1116
1117 rq = task_rq_lock(p, &rf);
1118
1119 /*
1120 * The task might have changed its scheduling policy to something
1121 * different than SCHED_DEADLINE (through switched_from_dl()).
1122 */
1123 if (!dl_task(p))
1124 goto unlock;
1125
1126 /*
1127 * The task might have been boosted by someone else and might be in the
1128 * boosting/deboosting path, its not throttled.
1129 */
1130 if (is_dl_boosted(dl_se))
1131 goto unlock;
1132
1133 /*
1134 * Spurious timer due to start_dl_timer() race; or we already received
1135 * a replenishment from rt_mutex_setprio().
1136 */
1137 if (!dl_se->dl_throttled)
1138 goto unlock;
1139
1140 sched_clock_tick();
1141 update_rq_clock(rq);
1142
1143 /*
1144 * If the throttle happened during sched-out; like:
1145 *
1146 * schedule()
1147 * deactivate_task()
1148 * dequeue_task_dl()
1149 * update_curr_dl()
1150 * start_dl_timer()
1151 * __dequeue_task_dl()
1152 * prev->on_rq = 0;
1153 *
1154 * We can be both throttled and !queued. Replenish the counter
1155 * but do not enqueue -- wait for our wakeup to do that.
1156 */
1157 if (!task_on_rq_queued(p)) {
1158 replenish_dl_entity(dl_se);
1159 goto unlock;
1160 }
1161
1162#ifdef CONFIG_SMP
1163 if (unlikely(!rq->online)) {
1164 /*
1165 * If the runqueue is no longer available, migrate the
1166 * task elsewhere. This necessarily changes rq.
1167 */
1168 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1169 rq = dl_task_offline_migration(rq, p);
1170 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1171 update_rq_clock(rq);
1172
1173 /*
1174 * Now that the task has been migrated to the new RQ and we
1175 * have that locked, proceed as normal and enqueue the task
1176 * there.
1177 */
1178 }
1179#endif
1180
1181 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1182 if (dl_task(rq->curr))
1183 check_preempt_curr_dl(rq, p, 0);
1184 else
1185 resched_curr(rq);
1186
1187#ifdef CONFIG_SMP
1188 /*
1189 * Queueing this task back might have overloaded rq, check if we need
1190 * to kick someone away.
1191 */
1192 if (has_pushable_dl_tasks(rq)) {
1193 /*
1194 * Nothing relies on rq->lock after this, so its safe to drop
1195 * rq->lock.
1196 */
1197 rq_unpin_lock(rq, &rf);
1198 push_dl_task(rq);
1199 rq_repin_lock(rq, &rf);
1200 }
1201#endif
1202
1203unlock:
1204 task_rq_unlock(rq, p, &rf);
1205
1206 /*
1207 * This can free the task_struct, including this hrtimer, do not touch
1208 * anything related to that after this.
1209 */
1210 put_task_struct(p);
1211
1212 return HRTIMER_NORESTART;
1213}
1214
1215void init_dl_task_timer(struct sched_dl_entity *dl_se)
1216{
1217 struct hrtimer *timer = &dl_se->dl_timer;
1218
1219 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1220 timer->function = dl_task_timer;
1221}
1222
1223/*
1224 * During the activation, CBS checks if it can reuse the current task's
1225 * runtime and period. If the deadline of the task is in the past, CBS
1226 * cannot use the runtime, and so it replenishes the task. This rule
1227 * works fine for implicit deadline tasks (deadline == period), and the
1228 * CBS was designed for implicit deadline tasks. However, a task with
1229 * constrained deadline (deadline < period) might be awakened after the
1230 * deadline, but before the next period. In this case, replenishing the
1231 * task would allow it to run for runtime / deadline. As in this case
1232 * deadline < period, CBS enables a task to run for more than the
1233 * runtime / period. In a very loaded system, this can cause a domino
1234 * effect, making other tasks miss their deadlines.
1235 *
1236 * To avoid this problem, in the activation of a constrained deadline
1237 * task after the deadline but before the next period, throttle the
1238 * task and set the replenishing timer to the begin of the next period,
1239 * unless it is boosted.
1240 */
1241static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1242{
1243 struct task_struct *p = dl_task_of(dl_se);
1244 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1245
1246 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1247 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1248 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
1249 return;
1250 dl_se->dl_throttled = 1;
1251 if (dl_se->runtime > 0)
1252 dl_se->runtime = 0;
1253 }
1254}
1255
1256static
1257int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1258{
1259 return (dl_se->runtime <= 0);
1260}
1261
1262/*
1263 * This function implements the GRUB accounting rule:
1264 * according to the GRUB reclaiming algorithm, the runtime is
1265 * not decreased as "dq = -dt", but as
1266 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1267 * where u is the utilization of the task, Umax is the maximum reclaimable
1268 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1269 * as the difference between the "total runqueue utilization" and the
1270 * runqueue active utilization, and Uextra is the (per runqueue) extra
1271 * reclaimable utilization.
1272 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1273 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1274 * BW_SHIFT.
1275 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT,
1276 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1277 * Since delta is a 64 bit variable, to have an overflow its value
1278 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1279 * So, overflow is not an issue here.
1280 */
1281static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1282{
1283 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1284 u64 u_act;
1285 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1286
1287 /*
1288 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1289 * we compare u_inact + rq->dl.extra_bw with
1290 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1291 * u_inact + rq->dl.extra_bw can be larger than
1292 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1293 * leading to wrong results)
1294 */
1295 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1296 u_act = u_act_min;
1297 else
1298 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1299
1300 return (delta * u_act) >> BW_SHIFT;
1301}
1302
1303/*
1304 * Update the current task's runtime statistics (provided it is still
1305 * a -deadline task and has not been removed from the dl_rq).
1306 */
1307static void update_curr_dl(struct rq *rq)
1308{
1309 struct task_struct *curr = rq->curr;
1310 struct sched_dl_entity *dl_se = &curr->dl;
1311 u64 delta_exec, scaled_delta_exec;
1312 int cpu = cpu_of(rq);
1313 u64 now;
1314
1315 if (!dl_task(curr) || !on_dl_rq(dl_se))
1316 return;
1317
1318 /*
1319 * Consumed budget is computed considering the time as
1320 * observed by schedulable tasks (excluding time spent
1321 * in hardirq context, etc.). Deadlines are instead
1322 * computed using hard walltime. This seems to be the more
1323 * natural solution, but the full ramifications of this
1324 * approach need further study.
1325 */
1326 now = rq_clock_task(rq);
1327 delta_exec = now - curr->se.exec_start;
1328 if (unlikely((s64)delta_exec <= 0)) {
1329 if (unlikely(dl_se->dl_yielded))
1330 goto throttle;
1331 return;
1332 }
1333
1334 schedstat_set(curr->stats.exec_max,
1335 max(curr->stats.exec_max, delta_exec));
1336
1337 trace_sched_stat_runtime(curr, delta_exec, 0);
1338
1339 update_current_exec_runtime(curr, now, delta_exec);
1340
1341 if (dl_entity_is_special(dl_se))
1342 return;
1343
1344 /*
1345 * For tasks that participate in GRUB, we implement GRUB-PA: the
1346 * spare reclaimed bandwidth is used to clock down frequency.
1347 *
1348 * For the others, we still need to scale reservation parameters
1349 * according to current frequency and CPU maximum capacity.
1350 */
1351 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1352 scaled_delta_exec = grub_reclaim(delta_exec,
1353 rq,
1354 &curr->dl);
1355 } else {
1356 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1357 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1358
1359 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1360 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1361 }
1362
1363 dl_se->runtime -= scaled_delta_exec;
1364
1365throttle:
1366 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1367 dl_se->dl_throttled = 1;
1368
1369 /* If requested, inform the user about runtime overruns. */
1370 if (dl_runtime_exceeded(dl_se) &&
1371 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1372 dl_se->dl_overrun = 1;
1373
1374 __dequeue_task_dl(rq, curr, 0);
1375 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1376 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1377
1378 if (!is_leftmost(curr, &rq->dl))
1379 resched_curr(rq);
1380 }
1381
1382 /*
1383 * Because -- for now -- we share the rt bandwidth, we need to
1384 * account our runtime there too, otherwise actual rt tasks
1385 * would be able to exceed the shared quota.
1386 *
1387 * Account to the root rt group for now.
1388 *
1389 * The solution we're working towards is having the RT groups scheduled
1390 * using deadline servers -- however there's a few nasties to figure
1391 * out before that can happen.
1392 */
1393 if (rt_bandwidth_enabled()) {
1394 struct rt_rq *rt_rq = &rq->rt;
1395
1396 raw_spin_lock(&rt_rq->rt_runtime_lock);
1397 /*
1398 * We'll let actual RT tasks worry about the overflow here, we
1399 * have our own CBS to keep us inline; only account when RT
1400 * bandwidth is relevant.
1401 */
1402 if (sched_rt_bandwidth_account(rt_rq))
1403 rt_rq->rt_time += delta_exec;
1404 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1405 }
1406}
1407
1408static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1409{
1410 struct sched_dl_entity *dl_se = container_of(timer,
1411 struct sched_dl_entity,
1412 inactive_timer);
1413 struct task_struct *p = dl_task_of(dl_se);
1414 struct rq_flags rf;
1415 struct rq *rq;
1416
1417 rq = task_rq_lock(p, &rf);
1418
1419 sched_clock_tick();
1420 update_rq_clock(rq);
1421
1422 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1423 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1424
1425 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1426 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1427 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1428 dl_se->dl_non_contending = 0;
1429 }
1430
1431 raw_spin_lock(&dl_b->lock);
1432 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1433 raw_spin_unlock(&dl_b->lock);
1434 __dl_clear_params(p);
1435
1436 goto unlock;
1437 }
1438 if (dl_se->dl_non_contending == 0)
1439 goto unlock;
1440
1441 sub_running_bw(dl_se, &rq->dl);
1442 dl_se->dl_non_contending = 0;
1443unlock:
1444 task_rq_unlock(rq, p, &rf);
1445 put_task_struct(p);
1446
1447 return HRTIMER_NORESTART;
1448}
1449
1450void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1451{
1452 struct hrtimer *timer = &dl_se->inactive_timer;
1453
1454 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1455 timer->function = inactive_task_timer;
1456}
1457
1458#define __node_2_dle(node) \
1459 rb_entry((node), struct sched_dl_entity, rb_node)
1460
1461#ifdef CONFIG_SMP
1462
1463static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1464{
1465 struct rq *rq = rq_of_dl_rq(dl_rq);
1466
1467 if (dl_rq->earliest_dl.curr == 0 ||
1468 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1469 if (dl_rq->earliest_dl.curr == 0)
1470 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1471 dl_rq->earliest_dl.curr = deadline;
1472 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1473 }
1474}
1475
1476static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1477{
1478 struct rq *rq = rq_of_dl_rq(dl_rq);
1479
1480 /*
1481 * Since we may have removed our earliest (and/or next earliest)
1482 * task we must recompute them.
1483 */
1484 if (!dl_rq->dl_nr_running) {
1485 dl_rq->earliest_dl.curr = 0;
1486 dl_rq->earliest_dl.next = 0;
1487 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1488 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1489 } else {
1490 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1491 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1492
1493 dl_rq->earliest_dl.curr = entry->deadline;
1494 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1495 }
1496}
1497
1498#else
1499
1500static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1501static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1502
1503#endif /* CONFIG_SMP */
1504
1505static inline
1506void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1507{
1508 int prio = dl_task_of(dl_se)->prio;
1509 u64 deadline = dl_se->deadline;
1510
1511 WARN_ON(!dl_prio(prio));
1512 dl_rq->dl_nr_running++;
1513 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1514
1515 inc_dl_deadline(dl_rq, deadline);
1516 inc_dl_migration(dl_se, dl_rq);
1517}
1518
1519static inline
1520void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1521{
1522 int prio = dl_task_of(dl_se)->prio;
1523
1524 WARN_ON(!dl_prio(prio));
1525 WARN_ON(!dl_rq->dl_nr_running);
1526 dl_rq->dl_nr_running--;
1527 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1528
1529 dec_dl_deadline(dl_rq, dl_se->deadline);
1530 dec_dl_migration(dl_se, dl_rq);
1531}
1532
1533static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1534{
1535 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1536}
1537
1538static inline struct sched_statistics *
1539__schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1540{
1541 return &dl_task_of(dl_se)->stats;
1542}
1543
1544static inline void
1545update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1546{
1547 struct sched_statistics *stats;
1548
1549 if (!schedstat_enabled())
1550 return;
1551
1552 stats = __schedstats_from_dl_se(dl_se);
1553 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1554}
1555
1556static inline void
1557update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1558{
1559 struct sched_statistics *stats;
1560
1561 if (!schedstat_enabled())
1562 return;
1563
1564 stats = __schedstats_from_dl_se(dl_se);
1565 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1566}
1567
1568static inline void
1569update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1570{
1571 struct sched_statistics *stats;
1572
1573 if (!schedstat_enabled())
1574 return;
1575
1576 stats = __schedstats_from_dl_se(dl_se);
1577 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1578}
1579
1580static inline void
1581update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1582 int flags)
1583{
1584 if (!schedstat_enabled())
1585 return;
1586
1587 if (flags & ENQUEUE_WAKEUP)
1588 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1589}
1590
1591static inline void
1592update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1593 int flags)
1594{
1595 struct task_struct *p = dl_task_of(dl_se);
1596
1597 if (!schedstat_enabled())
1598 return;
1599
1600 if ((flags & DEQUEUE_SLEEP)) {
1601 unsigned int state;
1602
1603 state = READ_ONCE(p->__state);
1604 if (state & TASK_INTERRUPTIBLE)
1605 __schedstat_set(p->stats.sleep_start,
1606 rq_clock(rq_of_dl_rq(dl_rq)));
1607
1608 if (state & TASK_UNINTERRUPTIBLE)
1609 __schedstat_set(p->stats.block_start,
1610 rq_clock(rq_of_dl_rq(dl_rq)));
1611 }
1612}
1613
1614static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1615{
1616 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1617
1618 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1619
1620 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1621
1622 inc_dl_tasks(dl_se, dl_rq);
1623}
1624
1625static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1626{
1627 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1628
1629 if (RB_EMPTY_NODE(&dl_se->rb_node))
1630 return;
1631
1632 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1633
1634 RB_CLEAR_NODE(&dl_se->rb_node);
1635
1636 dec_dl_tasks(dl_se, dl_rq);
1637}
1638
1639static void
1640enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1641{
1642 WARN_ON_ONCE(on_dl_rq(dl_se));
1643
1644 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1645
1646 /*
1647 * If this is a wakeup or a new instance, the scheduling
1648 * parameters of the task might need updating. Otherwise,
1649 * we want a replenishment of its runtime.
1650 */
1651 if (flags & ENQUEUE_WAKEUP) {
1652 task_contending(dl_se, flags);
1653 update_dl_entity(dl_se);
1654 } else if (flags & ENQUEUE_REPLENISH) {
1655 replenish_dl_entity(dl_se);
1656 } else if ((flags & ENQUEUE_RESTORE) &&
1657 dl_time_before(dl_se->deadline,
1658 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1659 setup_new_dl_entity(dl_se);
1660 }
1661
1662 __enqueue_dl_entity(dl_se);
1663}
1664
1665static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1666{
1667 __dequeue_dl_entity(dl_se);
1668}
1669
1670static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1671{
1672 if (is_dl_boosted(&p->dl)) {
1673 /*
1674 * Because of delays in the detection of the overrun of a
1675 * thread's runtime, it might be the case that a thread
1676 * goes to sleep in a rt mutex with negative runtime. As
1677 * a consequence, the thread will be throttled.
1678 *
1679 * While waiting for the mutex, this thread can also be
1680 * boosted via PI, resulting in a thread that is throttled
1681 * and boosted at the same time.
1682 *
1683 * In this case, the boost overrides the throttle.
1684 */
1685 if (p->dl.dl_throttled) {
1686 /*
1687 * The replenish timer needs to be canceled. No
1688 * problem if it fires concurrently: boosted threads
1689 * are ignored in dl_task_timer().
1690 */
1691 hrtimer_try_to_cancel(&p->dl.dl_timer);
1692 p->dl.dl_throttled = 0;
1693 }
1694 } else if (!dl_prio(p->normal_prio)) {
1695 /*
1696 * Special case in which we have a !SCHED_DEADLINE task that is going
1697 * to be deboosted, but exceeds its runtime while doing so. No point in
1698 * replenishing it, as it's going to return back to its original
1699 * scheduling class after this. If it has been throttled, we need to
1700 * clear the flag, otherwise the task may wake up as throttled after
1701 * being boosted again with no means to replenish the runtime and clear
1702 * the throttle.
1703 */
1704 p->dl.dl_throttled = 0;
1705 if (!(flags & ENQUEUE_REPLENISH))
1706 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
1707 task_pid_nr(p));
1708
1709 return;
1710 }
1711
1712 /*
1713 * Check if a constrained deadline task was activated
1714 * after the deadline but before the next period.
1715 * If that is the case, the task will be throttled and
1716 * the replenishment timer will be set to the next period.
1717 */
1718 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1719 dl_check_constrained_dl(&p->dl);
1720
1721 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1722 add_rq_bw(&p->dl, &rq->dl);
1723 add_running_bw(&p->dl, &rq->dl);
1724 }
1725
1726 /*
1727 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1728 * its budget it needs a replenishment and, since it now is on
1729 * its rq, the bandwidth timer callback (which clearly has not
1730 * run yet) will take care of this.
1731 * However, the active utilization does not depend on the fact
1732 * that the task is on the runqueue or not (but depends on the
1733 * task's state - in GRUB parlance, "inactive" vs "active contending").
1734 * In other words, even if a task is throttled its utilization must
1735 * be counted in the active utilization; hence, we need to call
1736 * add_running_bw().
1737 */
1738 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1739 if (flags & ENQUEUE_WAKEUP)
1740 task_contending(&p->dl, flags);
1741
1742 return;
1743 }
1744
1745 check_schedstat_required();
1746 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
1747
1748 enqueue_dl_entity(&p->dl, flags);
1749
1750 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1751 enqueue_pushable_dl_task(rq, p);
1752}
1753
1754static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1755{
1756 update_stats_dequeue_dl(&rq->dl, &p->dl, flags);
1757 dequeue_dl_entity(&p->dl);
1758 dequeue_pushable_dl_task(rq, p);
1759}
1760
1761static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1762{
1763 update_curr_dl(rq);
1764 __dequeue_task_dl(rq, p, flags);
1765
1766 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1767 sub_running_bw(&p->dl, &rq->dl);
1768 sub_rq_bw(&p->dl, &rq->dl);
1769 }
1770
1771 /*
1772 * This check allows to start the inactive timer (or to immediately
1773 * decrease the active utilization, if needed) in two cases:
1774 * when the task blocks and when it is terminating
1775 * (p->state == TASK_DEAD). We can handle the two cases in the same
1776 * way, because from GRUB's point of view the same thing is happening
1777 * (the task moves from "active contending" to "active non contending"
1778 * or "inactive")
1779 */
1780 if (flags & DEQUEUE_SLEEP)
1781 task_non_contending(p);
1782}
1783
1784/*
1785 * Yield task semantic for -deadline tasks is:
1786 *
1787 * get off from the CPU until our next instance, with
1788 * a new runtime. This is of little use now, since we
1789 * don't have a bandwidth reclaiming mechanism. Anyway,
1790 * bandwidth reclaiming is planned for the future, and
1791 * yield_task_dl will indicate that some spare budget
1792 * is available for other task instances to use it.
1793 */
1794static void yield_task_dl(struct rq *rq)
1795{
1796 /*
1797 * We make the task go to sleep until its current deadline by
1798 * forcing its runtime to zero. This way, update_curr_dl() stops
1799 * it and the bandwidth timer will wake it up and will give it
1800 * new scheduling parameters (thanks to dl_yielded=1).
1801 */
1802 rq->curr->dl.dl_yielded = 1;
1803
1804 update_rq_clock(rq);
1805 update_curr_dl(rq);
1806 /*
1807 * Tell update_rq_clock() that we've just updated,
1808 * so we don't do microscopic update in schedule()
1809 * and double the fastpath cost.
1810 */
1811 rq_clock_skip_update(rq);
1812}
1813
1814#ifdef CONFIG_SMP
1815
1816static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
1817 struct rq *rq)
1818{
1819 return (!rq->dl.dl_nr_running ||
1820 dl_time_before(p->dl.deadline,
1821 rq->dl.earliest_dl.curr));
1822}
1823
1824static int find_later_rq(struct task_struct *task);
1825
1826static int
1827select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1828{
1829 struct task_struct *curr;
1830 bool select_rq;
1831 struct rq *rq;
1832
1833 if (!(flags & WF_TTWU))
1834 goto out;
1835
1836 rq = cpu_rq(cpu);
1837
1838 rcu_read_lock();
1839 curr = READ_ONCE(rq->curr); /* unlocked access */
1840
1841 /*
1842 * If we are dealing with a -deadline task, we must
1843 * decide where to wake it up.
1844 * If it has a later deadline and the current task
1845 * on this rq can't move (provided the waking task
1846 * can!) we prefer to send it somewhere else. On the
1847 * other hand, if it has a shorter deadline, we
1848 * try to make it stay here, it might be important.
1849 */
1850 select_rq = unlikely(dl_task(curr)) &&
1851 (curr->nr_cpus_allowed < 2 ||
1852 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1853 p->nr_cpus_allowed > 1;
1854
1855 /*
1856 * Take the capacity of the CPU into account to
1857 * ensure it fits the requirement of the task.
1858 */
1859 if (sched_asym_cpucap_active())
1860 select_rq |= !dl_task_fits_capacity(p, cpu);
1861
1862 if (select_rq) {
1863 int target = find_later_rq(p);
1864
1865 if (target != -1 &&
1866 dl_task_is_earliest_deadline(p, cpu_rq(target)))
1867 cpu = target;
1868 }
1869 rcu_read_unlock();
1870
1871out:
1872 return cpu;
1873}
1874
1875static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1876{
1877 struct rq_flags rf;
1878 struct rq *rq;
1879
1880 if (READ_ONCE(p->__state) != TASK_WAKING)
1881 return;
1882
1883 rq = task_rq(p);
1884 /*
1885 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1886 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1887 * rq->lock is not... So, lock it
1888 */
1889 rq_lock(rq, &rf);
1890 if (p->dl.dl_non_contending) {
1891 update_rq_clock(rq);
1892 sub_running_bw(&p->dl, &rq->dl);
1893 p->dl.dl_non_contending = 0;
1894 /*
1895 * If the timer handler is currently running and the
1896 * timer cannot be canceled, inactive_task_timer()
1897 * will see that dl_not_contending is not set, and
1898 * will not touch the rq's active utilization,
1899 * so we are still safe.
1900 */
1901 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1902 put_task_struct(p);
1903 }
1904 sub_rq_bw(&p->dl, &rq->dl);
1905 rq_unlock(rq, &rf);
1906}
1907
1908static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1909{
1910 /*
1911 * Current can't be migrated, useless to reschedule,
1912 * let's hope p can move out.
1913 */
1914 if (rq->curr->nr_cpus_allowed == 1 ||
1915 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1916 return;
1917
1918 /*
1919 * p is migratable, so let's not schedule it and
1920 * see if it is pushed or pulled somewhere else.
1921 */
1922 if (p->nr_cpus_allowed != 1 &&
1923 cpudl_find(&rq->rd->cpudl, p, NULL))
1924 return;
1925
1926 resched_curr(rq);
1927}
1928
1929static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1930{
1931 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1932 /*
1933 * This is OK, because current is on_cpu, which avoids it being
1934 * picked for load-balance and preemption/IRQs are still
1935 * disabled avoiding further scheduler activity on it and we've
1936 * not yet started the picking loop.
1937 */
1938 rq_unpin_lock(rq, rf);
1939 pull_dl_task(rq);
1940 rq_repin_lock(rq, rf);
1941 }
1942
1943 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1944}
1945#endif /* CONFIG_SMP */
1946
1947/*
1948 * Only called when both the current and waking task are -deadline
1949 * tasks.
1950 */
1951static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1952 int flags)
1953{
1954 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1955 resched_curr(rq);
1956 return;
1957 }
1958
1959#ifdef CONFIG_SMP
1960 /*
1961 * In the unlikely case current and p have the same deadline
1962 * let us try to decide what's the best thing to do...
1963 */
1964 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1965 !test_tsk_need_resched(rq->curr))
1966 check_preempt_equal_dl(rq, p);
1967#endif /* CONFIG_SMP */
1968}
1969
1970#ifdef CONFIG_SCHED_HRTICK
1971static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1972{
1973 hrtick_start(rq, p->dl.runtime);
1974}
1975#else /* !CONFIG_SCHED_HRTICK */
1976static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1977{
1978}
1979#endif
1980
1981static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
1982{
1983 struct sched_dl_entity *dl_se = &p->dl;
1984 struct dl_rq *dl_rq = &rq->dl;
1985
1986 p->se.exec_start = rq_clock_task(rq);
1987 if (on_dl_rq(&p->dl))
1988 update_stats_wait_end_dl(dl_rq, dl_se);
1989
1990 /* You can't push away the running task */
1991 dequeue_pushable_dl_task(rq, p);
1992
1993 if (!first)
1994 return;
1995
1996 if (hrtick_enabled_dl(rq))
1997 start_hrtick_dl(rq, p);
1998
1999 if (rq->curr->sched_class != &dl_sched_class)
2000 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2001
2002 deadline_queue_push_tasks(rq);
2003}
2004
2005static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
2006{
2007 struct rb_node *left = rb_first_cached(&dl_rq->root);
2008
2009 if (!left)
2010 return NULL;
2011
2012 return __node_2_dle(left);
2013}
2014
2015static struct task_struct *pick_task_dl(struct rq *rq)
2016{
2017 struct sched_dl_entity *dl_se;
2018 struct dl_rq *dl_rq = &rq->dl;
2019 struct task_struct *p;
2020
2021 if (!sched_dl_runnable(rq))
2022 return NULL;
2023
2024 dl_se = pick_next_dl_entity(dl_rq);
2025 WARN_ON_ONCE(!dl_se);
2026 p = dl_task_of(dl_se);
2027
2028 return p;
2029}
2030
2031static struct task_struct *pick_next_task_dl(struct rq *rq)
2032{
2033 struct task_struct *p;
2034
2035 p = pick_task_dl(rq);
2036 if (p)
2037 set_next_task_dl(rq, p, true);
2038
2039 return p;
2040}
2041
2042static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2043{
2044 struct sched_dl_entity *dl_se = &p->dl;
2045 struct dl_rq *dl_rq = &rq->dl;
2046
2047 if (on_dl_rq(&p->dl))
2048 update_stats_wait_start_dl(dl_rq, dl_se);
2049
2050 update_curr_dl(rq);
2051
2052 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2053 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2054 enqueue_pushable_dl_task(rq, p);
2055}
2056
2057/*
2058 * scheduler tick hitting a task of our scheduling class.
2059 *
2060 * NOTE: This function can be called remotely by the tick offload that
2061 * goes along full dynticks. Therefore no local assumption can be made
2062 * and everything must be accessed through the @rq and @curr passed in
2063 * parameters.
2064 */
2065static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2066{
2067 update_curr_dl(rq);
2068
2069 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2070 /*
2071 * Even when we have runtime, update_curr_dl() might have resulted in us
2072 * not being the leftmost task anymore. In that case NEED_RESCHED will
2073 * be set and schedule() will start a new hrtick for the next task.
2074 */
2075 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2076 is_leftmost(p, &rq->dl))
2077 start_hrtick_dl(rq, p);
2078}
2079
2080static void task_fork_dl(struct task_struct *p)
2081{
2082 /*
2083 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2084 * sched_fork()
2085 */
2086}
2087
2088#ifdef CONFIG_SMP
2089
2090/* Only try algorithms three times */
2091#define DL_MAX_TRIES 3
2092
2093static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2094{
2095 if (!task_on_cpu(rq, p) &&
2096 cpumask_test_cpu(cpu, &p->cpus_mask))
2097 return 1;
2098 return 0;
2099}
2100
2101/*
2102 * Return the earliest pushable rq's task, which is suitable to be executed
2103 * on the CPU, NULL otherwise:
2104 */
2105static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2106{
2107 struct task_struct *p = NULL;
2108 struct rb_node *next_node;
2109
2110 if (!has_pushable_dl_tasks(rq))
2111 return NULL;
2112
2113 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2114
2115next_node:
2116 if (next_node) {
2117 p = __node_2_pdl(next_node);
2118
2119 if (pick_dl_task(rq, p, cpu))
2120 return p;
2121
2122 next_node = rb_next(next_node);
2123 goto next_node;
2124 }
2125
2126 return NULL;
2127}
2128
2129static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2130
2131static int find_later_rq(struct task_struct *task)
2132{
2133 struct sched_domain *sd;
2134 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2135 int this_cpu = smp_processor_id();
2136 int cpu = task_cpu(task);
2137
2138 /* Make sure the mask is initialized first */
2139 if (unlikely(!later_mask))
2140 return -1;
2141
2142 if (task->nr_cpus_allowed == 1)
2143 return -1;
2144
2145 /*
2146 * We have to consider system topology and task affinity
2147 * first, then we can look for a suitable CPU.
2148 */
2149 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2150 return -1;
2151
2152 /*
2153 * If we are here, some targets have been found, including
2154 * the most suitable which is, among the runqueues where the
2155 * current tasks have later deadlines than the task's one, the
2156 * rq with the latest possible one.
2157 *
2158 * Now we check how well this matches with task's
2159 * affinity and system topology.
2160 *
2161 * The last CPU where the task run is our first
2162 * guess, since it is most likely cache-hot there.
2163 */
2164 if (cpumask_test_cpu(cpu, later_mask))
2165 return cpu;
2166 /*
2167 * Check if this_cpu is to be skipped (i.e., it is
2168 * not in the mask) or not.
2169 */
2170 if (!cpumask_test_cpu(this_cpu, later_mask))
2171 this_cpu = -1;
2172
2173 rcu_read_lock();
2174 for_each_domain(cpu, sd) {
2175 if (sd->flags & SD_WAKE_AFFINE) {
2176 int best_cpu;
2177
2178 /*
2179 * If possible, preempting this_cpu is
2180 * cheaper than migrating.
2181 */
2182 if (this_cpu != -1 &&
2183 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2184 rcu_read_unlock();
2185 return this_cpu;
2186 }
2187
2188 best_cpu = cpumask_any_and_distribute(later_mask,
2189 sched_domain_span(sd));
2190 /*
2191 * Last chance: if a CPU being in both later_mask
2192 * and current sd span is valid, that becomes our
2193 * choice. Of course, the latest possible CPU is
2194 * already under consideration through later_mask.
2195 */
2196 if (best_cpu < nr_cpu_ids) {
2197 rcu_read_unlock();
2198 return best_cpu;
2199 }
2200 }
2201 }
2202 rcu_read_unlock();
2203
2204 /*
2205 * At this point, all our guesses failed, we just return
2206 * 'something', and let the caller sort the things out.
2207 */
2208 if (this_cpu != -1)
2209 return this_cpu;
2210
2211 cpu = cpumask_any_distribute(later_mask);
2212 if (cpu < nr_cpu_ids)
2213 return cpu;
2214
2215 return -1;
2216}
2217
2218/* Locks the rq it finds */
2219static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2220{
2221 struct rq *later_rq = NULL;
2222 int tries;
2223 int cpu;
2224
2225 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2226 cpu = find_later_rq(task);
2227
2228 if ((cpu == -1) || (cpu == rq->cpu))
2229 break;
2230
2231 later_rq = cpu_rq(cpu);
2232
2233 if (!dl_task_is_earliest_deadline(task, later_rq)) {
2234 /*
2235 * Target rq has tasks of equal or earlier deadline,
2236 * retrying does not release any lock and is unlikely
2237 * to yield a different result.
2238 */
2239 later_rq = NULL;
2240 break;
2241 }
2242
2243 /* Retry if something changed. */
2244 if (double_lock_balance(rq, later_rq)) {
2245 if (unlikely(task_rq(task) != rq ||
2246 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2247 task_on_cpu(rq, task) ||
2248 !dl_task(task) ||
2249 !task_on_rq_queued(task))) {
2250 double_unlock_balance(rq, later_rq);
2251 later_rq = NULL;
2252 break;
2253 }
2254 }
2255
2256 /*
2257 * If the rq we found has no -deadline task, or
2258 * its earliest one has a later deadline than our
2259 * task, the rq is a good one.
2260 */
2261 if (dl_task_is_earliest_deadline(task, later_rq))
2262 break;
2263
2264 /* Otherwise we try again. */
2265 double_unlock_balance(rq, later_rq);
2266 later_rq = NULL;
2267 }
2268
2269 return later_rq;
2270}
2271
2272static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2273{
2274 struct task_struct *p;
2275
2276 if (!has_pushable_dl_tasks(rq))
2277 return NULL;
2278
2279 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2280
2281 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2282 WARN_ON_ONCE(task_current(rq, p));
2283 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2284
2285 WARN_ON_ONCE(!task_on_rq_queued(p));
2286 WARN_ON_ONCE(!dl_task(p));
2287
2288 return p;
2289}
2290
2291/*
2292 * See if the non running -deadline tasks on this rq
2293 * can be sent to some other CPU where they can preempt
2294 * and start executing.
2295 */
2296static int push_dl_task(struct rq *rq)
2297{
2298 struct task_struct *next_task;
2299 struct rq *later_rq;
2300 int ret = 0;
2301
2302 if (!rq->dl.overloaded)
2303 return 0;
2304
2305 next_task = pick_next_pushable_dl_task(rq);
2306 if (!next_task)
2307 return 0;
2308
2309retry:
2310 /*
2311 * If next_task preempts rq->curr, and rq->curr
2312 * can move away, it makes sense to just reschedule
2313 * without going further in pushing next_task.
2314 */
2315 if (dl_task(rq->curr) &&
2316 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2317 rq->curr->nr_cpus_allowed > 1) {
2318 resched_curr(rq);
2319 return 0;
2320 }
2321
2322 if (is_migration_disabled(next_task))
2323 return 0;
2324
2325 if (WARN_ON(next_task == rq->curr))
2326 return 0;
2327
2328 /* We might release rq lock */
2329 get_task_struct(next_task);
2330
2331 /* Will lock the rq it'll find */
2332 later_rq = find_lock_later_rq(next_task, rq);
2333 if (!later_rq) {
2334 struct task_struct *task;
2335
2336 /*
2337 * We must check all this again, since
2338 * find_lock_later_rq releases rq->lock and it is
2339 * then possible that next_task has migrated.
2340 */
2341 task = pick_next_pushable_dl_task(rq);
2342 if (task == next_task) {
2343 /*
2344 * The task is still there. We don't try
2345 * again, some other CPU will pull it when ready.
2346 */
2347 goto out;
2348 }
2349
2350 if (!task)
2351 /* No more tasks */
2352 goto out;
2353
2354 put_task_struct(next_task);
2355 next_task = task;
2356 goto retry;
2357 }
2358
2359 deactivate_task(rq, next_task, 0);
2360 set_task_cpu(next_task, later_rq->cpu);
2361 activate_task(later_rq, next_task, 0);
2362 ret = 1;
2363
2364 resched_curr(later_rq);
2365
2366 double_unlock_balance(rq, later_rq);
2367
2368out:
2369 put_task_struct(next_task);
2370
2371 return ret;
2372}
2373
2374static void push_dl_tasks(struct rq *rq)
2375{
2376 /* push_dl_task() will return true if it moved a -deadline task */
2377 while (push_dl_task(rq))
2378 ;
2379}
2380
2381static void pull_dl_task(struct rq *this_rq)
2382{
2383 int this_cpu = this_rq->cpu, cpu;
2384 struct task_struct *p, *push_task;
2385 bool resched = false;
2386 struct rq *src_rq;
2387 u64 dmin = LONG_MAX;
2388
2389 if (likely(!dl_overloaded(this_rq)))
2390 return;
2391
2392 /*
2393 * Match the barrier from dl_set_overloaded; this guarantees that if we
2394 * see overloaded we must also see the dlo_mask bit.
2395 */
2396 smp_rmb();
2397
2398 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2399 if (this_cpu == cpu)
2400 continue;
2401
2402 src_rq = cpu_rq(cpu);
2403
2404 /*
2405 * It looks racy, abd it is! However, as in sched_rt.c,
2406 * we are fine with this.
2407 */
2408 if (this_rq->dl.dl_nr_running &&
2409 dl_time_before(this_rq->dl.earliest_dl.curr,
2410 src_rq->dl.earliest_dl.next))
2411 continue;
2412
2413 /* Might drop this_rq->lock */
2414 push_task = NULL;
2415 double_lock_balance(this_rq, src_rq);
2416
2417 /*
2418 * If there are no more pullable tasks on the
2419 * rq, we're done with it.
2420 */
2421 if (src_rq->dl.dl_nr_running <= 1)
2422 goto skip;
2423
2424 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2425
2426 /*
2427 * We found a task to be pulled if:
2428 * - it preempts our current (if there's one),
2429 * - it will preempt the last one we pulled (if any).
2430 */
2431 if (p && dl_time_before(p->dl.deadline, dmin) &&
2432 dl_task_is_earliest_deadline(p, this_rq)) {
2433 WARN_ON(p == src_rq->curr);
2434 WARN_ON(!task_on_rq_queued(p));
2435
2436 /*
2437 * Then we pull iff p has actually an earlier
2438 * deadline than the current task of its runqueue.
2439 */
2440 if (dl_time_before(p->dl.deadline,
2441 src_rq->curr->dl.deadline))
2442 goto skip;
2443
2444 if (is_migration_disabled(p)) {
2445 push_task = get_push_task(src_rq);
2446 } else {
2447 deactivate_task(src_rq, p, 0);
2448 set_task_cpu(p, this_cpu);
2449 activate_task(this_rq, p, 0);
2450 dmin = p->dl.deadline;
2451 resched = true;
2452 }
2453
2454 /* Is there any other task even earlier? */
2455 }
2456skip:
2457 double_unlock_balance(this_rq, src_rq);
2458
2459 if (push_task) {
2460 raw_spin_rq_unlock(this_rq);
2461 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2462 push_task, &src_rq->push_work);
2463 raw_spin_rq_lock(this_rq);
2464 }
2465 }
2466
2467 if (resched)
2468 resched_curr(this_rq);
2469}
2470
2471/*
2472 * Since the task is not running and a reschedule is not going to happen
2473 * anytime soon on its runqueue, we try pushing it away now.
2474 */
2475static void task_woken_dl(struct rq *rq, struct task_struct *p)
2476{
2477 if (!task_on_cpu(rq, p) &&
2478 !test_tsk_need_resched(rq->curr) &&
2479 p->nr_cpus_allowed > 1 &&
2480 dl_task(rq->curr) &&
2481 (rq->curr->nr_cpus_allowed < 2 ||
2482 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2483 push_dl_tasks(rq);
2484 }
2485}
2486
2487static void set_cpus_allowed_dl(struct task_struct *p,
2488 struct affinity_context *ctx)
2489{
2490 struct root_domain *src_rd;
2491 struct rq *rq;
2492
2493 WARN_ON_ONCE(!dl_task(p));
2494
2495 rq = task_rq(p);
2496 src_rd = rq->rd;
2497 /*
2498 * Migrating a SCHED_DEADLINE task between exclusive
2499 * cpusets (different root_domains) entails a bandwidth
2500 * update. We already made space for us in the destination
2501 * domain (see cpuset_can_attach()).
2502 */
2503 if (!cpumask_intersects(src_rd->span, ctx->new_mask)) {
2504 struct dl_bw *src_dl_b;
2505
2506 src_dl_b = dl_bw_of(cpu_of(rq));
2507 /*
2508 * We now free resources of the root_domain we are migrating
2509 * off. In the worst case, sched_setattr() may temporary fail
2510 * until we complete the update.
2511 */
2512 raw_spin_lock(&src_dl_b->lock);
2513 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2514 raw_spin_unlock(&src_dl_b->lock);
2515 }
2516
2517 set_cpus_allowed_common(p, ctx);
2518}
2519
2520/* Assumes rq->lock is held */
2521static void rq_online_dl(struct rq *rq)
2522{
2523 if (rq->dl.overloaded)
2524 dl_set_overload(rq);
2525
2526 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2527 if (rq->dl.dl_nr_running > 0)
2528 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2529}
2530
2531/* Assumes rq->lock is held */
2532static void rq_offline_dl(struct rq *rq)
2533{
2534 if (rq->dl.overloaded)
2535 dl_clear_overload(rq);
2536
2537 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2538 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2539}
2540
2541void __init init_sched_dl_class(void)
2542{
2543 unsigned int i;
2544
2545 for_each_possible_cpu(i)
2546 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2547 GFP_KERNEL, cpu_to_node(i));
2548}
2549
2550void dl_add_task_root_domain(struct task_struct *p)
2551{
2552 struct rq_flags rf;
2553 struct rq *rq;
2554 struct dl_bw *dl_b;
2555
2556 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2557 if (!dl_task(p)) {
2558 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2559 return;
2560 }
2561
2562 rq = __task_rq_lock(p, &rf);
2563
2564 dl_b = &rq->rd->dl_bw;
2565 raw_spin_lock(&dl_b->lock);
2566
2567 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2568
2569 raw_spin_unlock(&dl_b->lock);
2570
2571 task_rq_unlock(rq, p, &rf);
2572}
2573
2574void dl_clear_root_domain(struct root_domain *rd)
2575{
2576 unsigned long flags;
2577
2578 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2579 rd->dl_bw.total_bw = 0;
2580 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2581}
2582
2583#endif /* CONFIG_SMP */
2584
2585static void switched_from_dl(struct rq *rq, struct task_struct *p)
2586{
2587 /*
2588 * task_non_contending() can start the "inactive timer" (if the 0-lag
2589 * time is in the future). If the task switches back to dl before
2590 * the "inactive timer" fires, it can continue to consume its current
2591 * runtime using its current deadline. If it stays outside of
2592 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2593 * will reset the task parameters.
2594 */
2595 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2596 task_non_contending(p);
2597
2598 if (!task_on_rq_queued(p)) {
2599 /*
2600 * Inactive timer is armed. However, p is leaving DEADLINE and
2601 * might migrate away from this rq while continuing to run on
2602 * some other class. We need to remove its contribution from
2603 * this rq running_bw now, or sub_rq_bw (below) will complain.
2604 */
2605 if (p->dl.dl_non_contending)
2606 sub_running_bw(&p->dl, &rq->dl);
2607 sub_rq_bw(&p->dl, &rq->dl);
2608 }
2609
2610 /*
2611 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2612 * at the 0-lag time, because the task could have been migrated
2613 * while SCHED_OTHER in the meanwhile.
2614 */
2615 if (p->dl.dl_non_contending)
2616 p->dl.dl_non_contending = 0;
2617
2618 /*
2619 * Since this might be the only -deadline task on the rq,
2620 * this is the right place to try to pull some other one
2621 * from an overloaded CPU, if any.
2622 */
2623 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2624 return;
2625
2626 deadline_queue_pull_task(rq);
2627}
2628
2629/*
2630 * When switching to -deadline, we may overload the rq, then
2631 * we try to push someone off, if possible.
2632 */
2633static void switched_to_dl(struct rq *rq, struct task_struct *p)
2634{
2635 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2636 put_task_struct(p);
2637
2638 /* If p is not queued we will update its parameters at next wakeup. */
2639 if (!task_on_rq_queued(p)) {
2640 add_rq_bw(&p->dl, &rq->dl);
2641
2642 return;
2643 }
2644
2645 if (rq->curr != p) {
2646#ifdef CONFIG_SMP
2647 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2648 deadline_queue_push_tasks(rq);
2649#endif
2650 if (dl_task(rq->curr))
2651 check_preempt_curr_dl(rq, p, 0);
2652 else
2653 resched_curr(rq);
2654 } else {
2655 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2656 }
2657}
2658
2659/*
2660 * If the scheduling parameters of a -deadline task changed,
2661 * a push or pull operation might be needed.
2662 */
2663static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2664 int oldprio)
2665{
2666 if (task_on_rq_queued(p) || task_current(rq, p)) {
2667#ifdef CONFIG_SMP
2668 /*
2669 * This might be too much, but unfortunately
2670 * we don't have the old deadline value, and
2671 * we can't argue if the task is increasing
2672 * or lowering its prio, so...
2673 */
2674 if (!rq->dl.overloaded)
2675 deadline_queue_pull_task(rq);
2676
2677 /*
2678 * If we now have a earlier deadline task than p,
2679 * then reschedule, provided p is still on this
2680 * runqueue.
2681 */
2682 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2683 resched_curr(rq);
2684#else
2685 /*
2686 * Again, we don't know if p has a earlier
2687 * or later deadline, so let's blindly set a
2688 * (maybe not needed) rescheduling point.
2689 */
2690 resched_curr(rq);
2691#endif /* CONFIG_SMP */
2692 }
2693}
2694
2695DEFINE_SCHED_CLASS(dl) = {
2696
2697 .enqueue_task = enqueue_task_dl,
2698 .dequeue_task = dequeue_task_dl,
2699 .yield_task = yield_task_dl,
2700
2701 .check_preempt_curr = check_preempt_curr_dl,
2702
2703 .pick_next_task = pick_next_task_dl,
2704 .put_prev_task = put_prev_task_dl,
2705 .set_next_task = set_next_task_dl,
2706
2707#ifdef CONFIG_SMP
2708 .balance = balance_dl,
2709 .pick_task = pick_task_dl,
2710 .select_task_rq = select_task_rq_dl,
2711 .migrate_task_rq = migrate_task_rq_dl,
2712 .set_cpus_allowed = set_cpus_allowed_dl,
2713 .rq_online = rq_online_dl,
2714 .rq_offline = rq_offline_dl,
2715 .task_woken = task_woken_dl,
2716 .find_lock_rq = find_lock_later_rq,
2717#endif
2718
2719 .task_tick = task_tick_dl,
2720 .task_fork = task_fork_dl,
2721
2722 .prio_changed = prio_changed_dl,
2723 .switched_from = switched_from_dl,
2724 .switched_to = switched_to_dl,
2725
2726 .update_curr = update_curr_dl,
2727};
2728
2729/* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2730static u64 dl_generation;
2731
2732int sched_dl_global_validate(void)
2733{
2734 u64 runtime = global_rt_runtime();
2735 u64 period = global_rt_period();
2736 u64 new_bw = to_ratio(period, runtime);
2737 u64 gen = ++dl_generation;
2738 struct dl_bw *dl_b;
2739 int cpu, cpus, ret = 0;
2740 unsigned long flags;
2741
2742 /*
2743 * Here we want to check the bandwidth not being set to some
2744 * value smaller than the currently allocated bandwidth in
2745 * any of the root_domains.
2746 */
2747 for_each_possible_cpu(cpu) {
2748 rcu_read_lock_sched();
2749
2750 if (dl_bw_visited(cpu, gen))
2751 goto next;
2752
2753 dl_b = dl_bw_of(cpu);
2754 cpus = dl_bw_cpus(cpu);
2755
2756 raw_spin_lock_irqsave(&dl_b->lock, flags);
2757 if (new_bw * cpus < dl_b->total_bw)
2758 ret = -EBUSY;
2759 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2760
2761next:
2762 rcu_read_unlock_sched();
2763
2764 if (ret)
2765 break;
2766 }
2767
2768 return ret;
2769}
2770
2771static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2772{
2773 if (global_rt_runtime() == RUNTIME_INF) {
2774 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2775 dl_rq->extra_bw = 1 << BW_SHIFT;
2776 } else {
2777 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2778 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2779 dl_rq->extra_bw = to_ratio(global_rt_period(),
2780 global_rt_runtime());
2781 }
2782}
2783
2784void sched_dl_do_global(void)
2785{
2786 u64 new_bw = -1;
2787 u64 gen = ++dl_generation;
2788 struct dl_bw *dl_b;
2789 int cpu;
2790 unsigned long flags;
2791
2792 if (global_rt_runtime() != RUNTIME_INF)
2793 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2794
2795 for_each_possible_cpu(cpu) {
2796 rcu_read_lock_sched();
2797
2798 if (dl_bw_visited(cpu, gen)) {
2799 rcu_read_unlock_sched();
2800 continue;
2801 }
2802
2803 dl_b = dl_bw_of(cpu);
2804
2805 raw_spin_lock_irqsave(&dl_b->lock, flags);
2806 dl_b->bw = new_bw;
2807 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2808
2809 rcu_read_unlock_sched();
2810 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2811 }
2812}
2813
2814/*
2815 * We must be sure that accepting a new task (or allowing changing the
2816 * parameters of an existing one) is consistent with the bandwidth
2817 * constraints. If yes, this function also accordingly updates the currently
2818 * allocated bandwidth to reflect the new situation.
2819 *
2820 * This function is called while holding p's rq->lock.
2821 */
2822int sched_dl_overflow(struct task_struct *p, int policy,
2823 const struct sched_attr *attr)
2824{
2825 u64 period = attr->sched_period ?: attr->sched_deadline;
2826 u64 runtime = attr->sched_runtime;
2827 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2828 int cpus, err = -1, cpu = task_cpu(p);
2829 struct dl_bw *dl_b = dl_bw_of(cpu);
2830 unsigned long cap;
2831
2832 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2833 return 0;
2834
2835 /* !deadline task may carry old deadline bandwidth */
2836 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2837 return 0;
2838
2839 /*
2840 * Either if a task, enters, leave, or stays -deadline but changes
2841 * its parameters, we may need to update accordingly the total
2842 * allocated bandwidth of the container.
2843 */
2844 raw_spin_lock(&dl_b->lock);
2845 cpus = dl_bw_cpus(cpu);
2846 cap = dl_bw_capacity(cpu);
2847
2848 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2849 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2850 if (hrtimer_active(&p->dl.inactive_timer))
2851 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2852 __dl_add(dl_b, new_bw, cpus);
2853 err = 0;
2854 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2855 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2856 /*
2857 * XXX this is slightly incorrect: when the task
2858 * utilization decreases, we should delay the total
2859 * utilization change until the task's 0-lag point.
2860 * But this would require to set the task's "inactive
2861 * timer" when the task is not inactive.
2862 */
2863 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2864 __dl_add(dl_b, new_bw, cpus);
2865 dl_change_utilization(p, new_bw);
2866 err = 0;
2867 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2868 /*
2869 * Do not decrease the total deadline utilization here,
2870 * switched_from_dl() will take care to do it at the correct
2871 * (0-lag) time.
2872 */
2873 err = 0;
2874 }
2875 raw_spin_unlock(&dl_b->lock);
2876
2877 return err;
2878}
2879
2880/*
2881 * This function initializes the sched_dl_entity of a newly becoming
2882 * SCHED_DEADLINE task.
2883 *
2884 * Only the static values are considered here, the actual runtime and the
2885 * absolute deadline will be properly calculated when the task is enqueued
2886 * for the first time with its new policy.
2887 */
2888void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2889{
2890 struct sched_dl_entity *dl_se = &p->dl;
2891
2892 dl_se->dl_runtime = attr->sched_runtime;
2893 dl_se->dl_deadline = attr->sched_deadline;
2894 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2895 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
2896 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2897 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2898}
2899
2900void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2901{
2902 struct sched_dl_entity *dl_se = &p->dl;
2903
2904 attr->sched_priority = p->rt_priority;
2905 attr->sched_runtime = dl_se->dl_runtime;
2906 attr->sched_deadline = dl_se->dl_deadline;
2907 attr->sched_period = dl_se->dl_period;
2908 attr->sched_flags &= ~SCHED_DL_FLAGS;
2909 attr->sched_flags |= dl_se->flags;
2910}
2911
2912/*
2913 * This function validates the new parameters of a -deadline task.
2914 * We ask for the deadline not being zero, and greater or equal
2915 * than the runtime, as well as the period of being zero or
2916 * greater than deadline. Furthermore, we have to be sure that
2917 * user parameters are above the internal resolution of 1us (we
2918 * check sched_runtime only since it is always the smaller one) and
2919 * below 2^63 ns (we have to check both sched_deadline and
2920 * sched_period, as the latter can be zero).
2921 */
2922bool __checkparam_dl(const struct sched_attr *attr)
2923{
2924 u64 period, max, min;
2925
2926 /* special dl tasks don't actually use any parameter */
2927 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2928 return true;
2929
2930 /* deadline != 0 */
2931 if (attr->sched_deadline == 0)
2932 return false;
2933
2934 /*
2935 * Since we truncate DL_SCALE bits, make sure we're at least
2936 * that big.
2937 */
2938 if (attr->sched_runtime < (1ULL << DL_SCALE))
2939 return false;
2940
2941 /*
2942 * Since we use the MSB for wrap-around and sign issues, make
2943 * sure it's not set (mind that period can be equal to zero).
2944 */
2945 if (attr->sched_deadline & (1ULL << 63) ||
2946 attr->sched_period & (1ULL << 63))
2947 return false;
2948
2949 period = attr->sched_period;
2950 if (!period)
2951 period = attr->sched_deadline;
2952
2953 /* runtime <= deadline <= period (if period != 0) */
2954 if (period < attr->sched_deadline ||
2955 attr->sched_deadline < attr->sched_runtime)
2956 return false;
2957
2958 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2959 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2960
2961 if (period < min || period > max)
2962 return false;
2963
2964 return true;
2965}
2966
2967/*
2968 * This function clears the sched_dl_entity static params.
2969 */
2970void __dl_clear_params(struct task_struct *p)
2971{
2972 struct sched_dl_entity *dl_se = &p->dl;
2973
2974 dl_se->dl_runtime = 0;
2975 dl_se->dl_deadline = 0;
2976 dl_se->dl_period = 0;
2977 dl_se->flags = 0;
2978 dl_se->dl_bw = 0;
2979 dl_se->dl_density = 0;
2980
2981 dl_se->dl_throttled = 0;
2982 dl_se->dl_yielded = 0;
2983 dl_se->dl_non_contending = 0;
2984 dl_se->dl_overrun = 0;
2985
2986#ifdef CONFIG_RT_MUTEXES
2987 dl_se->pi_se = dl_se;
2988#endif
2989}
2990
2991bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2992{
2993 struct sched_dl_entity *dl_se = &p->dl;
2994
2995 if (dl_se->dl_runtime != attr->sched_runtime ||
2996 dl_se->dl_deadline != attr->sched_deadline ||
2997 dl_se->dl_period != attr->sched_period ||
2998 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
2999 return true;
3000
3001 return false;
3002}
3003
3004#ifdef CONFIG_SMP
3005int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3006 const struct cpumask *trial)
3007{
3008 unsigned long flags, cap;
3009 struct dl_bw *cur_dl_b;
3010 int ret = 1;
3011
3012 rcu_read_lock_sched();
3013 cur_dl_b = dl_bw_of(cpumask_any(cur));
3014 cap = __dl_bw_capacity(trial);
3015 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3016 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3017 ret = 0;
3018 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3019 rcu_read_unlock_sched();
3020
3021 return ret;
3022}
3023
3024int dl_cpu_busy(int cpu, struct task_struct *p)
3025{
3026 unsigned long flags, cap;
3027 struct dl_bw *dl_b;
3028 bool overflow;
3029
3030 rcu_read_lock_sched();
3031 dl_b = dl_bw_of(cpu);
3032 raw_spin_lock_irqsave(&dl_b->lock, flags);
3033 cap = dl_bw_capacity(cpu);
3034 overflow = __dl_overflow(dl_b, cap, 0, p ? p->dl.dl_bw : 0);
3035
3036 if (!overflow && p) {
3037 /*
3038 * We reserve space for this task in the destination
3039 * root_domain, as we can't fail after this point.
3040 * We will free resources in the source root_domain
3041 * later on (see set_cpus_allowed_dl()).
3042 */
3043 __dl_add(dl_b, p->dl.dl_bw, dl_bw_cpus(cpu));
3044 }
3045
3046 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3047 rcu_read_unlock_sched();
3048
3049 return overflow ? -EBUSY : 0;
3050}
3051#endif
3052
3053#ifdef CONFIG_SCHED_DEBUG
3054void print_dl_stats(struct seq_file *m, int cpu)
3055{
3056 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3057}
3058#endif /* CONFIG_SCHED_DEBUG */