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
19#include <linux/cpuset.h>
20
21/*
22 * Default limits for DL period; on the top end we guard against small util
23 * tasks still getting ridiculously long effective runtimes, on the bottom end we
24 * guard against timer DoS.
25 */
26static unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
27static unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
28#ifdef CONFIG_SYSCTL
29static struct ctl_table sched_dl_sysctls[] = {
30 {
31 .procname = "sched_deadline_period_max_us",
32 .data = &sysctl_sched_dl_period_max,
33 .maxlen = sizeof(unsigned int),
34 .mode = 0644,
35 .proc_handler = proc_douintvec_minmax,
36 .extra1 = (void *)&sysctl_sched_dl_period_min,
37 },
38 {
39 .procname = "sched_deadline_period_min_us",
40 .data = &sysctl_sched_dl_period_min,
41 .maxlen = sizeof(unsigned int),
42 .mode = 0644,
43 .proc_handler = proc_douintvec_minmax,
44 .extra2 = (void *)&sysctl_sched_dl_period_max,
45 },
46};
47
48static int __init sched_dl_sysctl_init(void)
49{
50 register_sysctl_init("kernel", sched_dl_sysctls);
51 return 0;
52}
53late_initcall(sched_dl_sysctl_init);
54#endif
55
56static bool dl_server(struct sched_dl_entity *dl_se)
57{
58 return dl_se->dl_server;
59}
60
61static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
62{
63 BUG_ON(dl_server(dl_se));
64 return container_of(dl_se, struct task_struct, dl);
65}
66
67static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
68{
69 return container_of(dl_rq, struct rq, dl);
70}
71
72static inline struct rq *rq_of_dl_se(struct sched_dl_entity *dl_se)
73{
74 struct rq *rq = dl_se->rq;
75
76 if (!dl_server(dl_se))
77 rq = task_rq(dl_task_of(dl_se));
78
79 return rq;
80}
81
82static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
83{
84 return &rq_of_dl_se(dl_se)->dl;
85}
86
87static inline int on_dl_rq(struct sched_dl_entity *dl_se)
88{
89 return !RB_EMPTY_NODE(&dl_se->rb_node);
90}
91
92#ifdef CONFIG_RT_MUTEXES
93static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
94{
95 return dl_se->pi_se;
96}
97
98static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
99{
100 return pi_of(dl_se) != dl_se;
101}
102#else
103static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
104{
105 return dl_se;
106}
107
108static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
109{
110 return false;
111}
112#endif
113
114#ifdef CONFIG_SMP
115static inline struct dl_bw *dl_bw_of(int i)
116{
117 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
118 "sched RCU must be held");
119 return &cpu_rq(i)->rd->dl_bw;
120}
121
122static inline int dl_bw_cpus(int i)
123{
124 struct root_domain *rd = cpu_rq(i)->rd;
125 int cpus;
126
127 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
128 "sched RCU must be held");
129
130 if (cpumask_subset(rd->span, cpu_active_mask))
131 return cpumask_weight(rd->span);
132
133 cpus = 0;
134
135 for_each_cpu_and(i, rd->span, cpu_active_mask)
136 cpus++;
137
138 return cpus;
139}
140
141static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
142{
143 unsigned long cap = 0;
144 int i;
145
146 for_each_cpu_and(i, mask, cpu_active_mask)
147 cap += arch_scale_cpu_capacity(i);
148
149 return cap;
150}
151
152/*
153 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
154 * of the CPU the task is running on rather rd's \Sum CPU capacity.
155 */
156static inline unsigned long dl_bw_capacity(int i)
157{
158 if (!sched_asym_cpucap_active() &&
159 arch_scale_cpu_capacity(i) == SCHED_CAPACITY_SCALE) {
160 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
161 } else {
162 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
163 "sched RCU must be held");
164
165 return __dl_bw_capacity(cpu_rq(i)->rd->span);
166 }
167}
168
169static inline bool dl_bw_visited(int cpu, u64 gen)
170{
171 struct root_domain *rd = cpu_rq(cpu)->rd;
172
173 if (rd->visit_gen == gen)
174 return true;
175
176 rd->visit_gen = gen;
177 return false;
178}
179
180static inline
181void __dl_update(struct dl_bw *dl_b, s64 bw)
182{
183 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
184 int i;
185
186 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
187 "sched RCU must be held");
188 for_each_cpu_and(i, rd->span, cpu_active_mask) {
189 struct rq *rq = cpu_rq(i);
190
191 rq->dl.extra_bw += bw;
192 }
193}
194#else
195static inline struct dl_bw *dl_bw_of(int i)
196{
197 return &cpu_rq(i)->dl.dl_bw;
198}
199
200static inline int dl_bw_cpus(int i)
201{
202 return 1;
203}
204
205static inline unsigned long dl_bw_capacity(int i)
206{
207 return SCHED_CAPACITY_SCALE;
208}
209
210static inline bool dl_bw_visited(int cpu, u64 gen)
211{
212 return false;
213}
214
215static inline
216void __dl_update(struct dl_bw *dl_b, s64 bw)
217{
218 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
219
220 dl->extra_bw += bw;
221}
222#endif
223
224static inline
225void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
226{
227 dl_b->total_bw -= tsk_bw;
228 __dl_update(dl_b, (s32)tsk_bw / cpus);
229}
230
231static inline
232void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
233{
234 dl_b->total_bw += tsk_bw;
235 __dl_update(dl_b, -((s32)tsk_bw / cpus));
236}
237
238static inline bool
239__dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
240{
241 return dl_b->bw != -1 &&
242 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
243}
244
245static inline
246void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
247{
248 u64 old = dl_rq->running_bw;
249
250 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
251 dl_rq->running_bw += dl_bw;
252 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
253 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
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 __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
260{
261 u64 old = dl_rq->running_bw;
262
263 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
264 dl_rq->running_bw -= dl_bw;
265 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
266 if (dl_rq->running_bw > old)
267 dl_rq->running_bw = 0;
268 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
269 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
270}
271
272static inline
273void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
274{
275 u64 old = dl_rq->this_bw;
276
277 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
278 dl_rq->this_bw += dl_bw;
279 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
280}
281
282static inline
283void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
284{
285 u64 old = dl_rq->this_bw;
286
287 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
288 dl_rq->this_bw -= dl_bw;
289 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
290 if (dl_rq->this_bw > old)
291 dl_rq->this_bw = 0;
292 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
293}
294
295static inline
296void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
297{
298 if (!dl_entity_is_special(dl_se))
299 __add_rq_bw(dl_se->dl_bw, dl_rq);
300}
301
302static inline
303void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
304{
305 if (!dl_entity_is_special(dl_se))
306 __sub_rq_bw(dl_se->dl_bw, dl_rq);
307}
308
309static inline
310void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
311{
312 if (!dl_entity_is_special(dl_se))
313 __add_running_bw(dl_se->dl_bw, dl_rq);
314}
315
316static inline
317void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
318{
319 if (!dl_entity_is_special(dl_se))
320 __sub_running_bw(dl_se->dl_bw, dl_rq);
321}
322
323static void dl_rq_change_utilization(struct rq *rq, struct sched_dl_entity *dl_se, u64 new_bw)
324{
325 if (dl_se->dl_non_contending) {
326 sub_running_bw(dl_se, &rq->dl);
327 dl_se->dl_non_contending = 0;
328
329 /*
330 * If the timer handler is currently running and the
331 * timer cannot be canceled, inactive_task_timer()
332 * will see that dl_not_contending is not set, and
333 * will not touch the rq's active utilization,
334 * so we are still safe.
335 */
336 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) {
337 if (!dl_server(dl_se))
338 put_task_struct(dl_task_of(dl_se));
339 }
340 }
341 __sub_rq_bw(dl_se->dl_bw, &rq->dl);
342 __add_rq_bw(new_bw, &rq->dl);
343}
344
345static void dl_change_utilization(struct task_struct *p, u64 new_bw)
346{
347 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
348
349 if (task_on_rq_queued(p))
350 return;
351
352 dl_rq_change_utilization(task_rq(p), &p->dl, new_bw);
353}
354
355static void __dl_clear_params(struct sched_dl_entity *dl_se);
356
357/*
358 * The utilization of a task cannot be immediately removed from
359 * the rq active utilization (running_bw) when the task blocks.
360 * Instead, we have to wait for the so called "0-lag time".
361 *
362 * If a task blocks before the "0-lag time", a timer (the inactive
363 * timer) is armed, and running_bw is decreased when the timer
364 * fires.
365 *
366 * If the task wakes up again before the inactive timer fires,
367 * the timer is canceled, whereas if the task wakes up after the
368 * inactive timer fired (and running_bw has been decreased) the
369 * task's utilization has to be added to running_bw again.
370 * A flag in the deadline scheduling entity (dl_non_contending)
371 * is used to avoid race conditions between the inactive timer handler
372 * and task wakeups.
373 *
374 * The following diagram shows how running_bw is updated. A task is
375 * "ACTIVE" when its utilization contributes to running_bw; an
376 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
377 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
378 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
379 * time already passed, which does not contribute to running_bw anymore.
380 * +------------------+
381 * wakeup | ACTIVE |
382 * +------------------>+ contending |
383 * | add_running_bw | |
384 * | +----+------+------+
385 * | | ^
386 * | dequeue | |
387 * +--------+-------+ | |
388 * | | t >= 0-lag | | wakeup
389 * | INACTIVE |<---------------+ |
390 * | | sub_running_bw | |
391 * +--------+-------+ | |
392 * ^ | |
393 * | t < 0-lag | |
394 * | | |
395 * | V |
396 * | +----+------+------+
397 * | sub_running_bw | ACTIVE |
398 * +-------------------+ |
399 * inactive timer | non contending |
400 * fired +------------------+
401 *
402 * The task_non_contending() function is invoked when a task
403 * blocks, and checks if the 0-lag time already passed or
404 * not (in the first case, it directly updates running_bw;
405 * in the second case, it arms the inactive timer).
406 *
407 * The task_contending() function is invoked when a task wakes
408 * up, and checks if the task is still in the "ACTIVE non contending"
409 * state or not (in the second case, it updates running_bw).
410 */
411static void task_non_contending(struct sched_dl_entity *dl_se)
412{
413 struct hrtimer *timer = &dl_se->inactive_timer;
414 struct rq *rq = rq_of_dl_se(dl_se);
415 struct dl_rq *dl_rq = &rq->dl;
416 s64 zerolag_time;
417
418 /*
419 * If this is a non-deadline task that has been boosted,
420 * do nothing
421 */
422 if (dl_se->dl_runtime == 0)
423 return;
424
425 if (dl_entity_is_special(dl_se))
426 return;
427
428 WARN_ON(dl_se->dl_non_contending);
429
430 zerolag_time = dl_se->deadline -
431 div64_long((dl_se->runtime * dl_se->dl_period),
432 dl_se->dl_runtime);
433
434 /*
435 * Using relative times instead of the absolute "0-lag time"
436 * allows to simplify the code
437 */
438 zerolag_time -= rq_clock(rq);
439
440 /*
441 * If the "0-lag time" already passed, decrease the active
442 * utilization now, instead of starting a timer
443 */
444 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
445 if (dl_server(dl_se)) {
446 sub_running_bw(dl_se, dl_rq);
447 } else {
448 struct task_struct *p = dl_task_of(dl_se);
449
450 if (dl_task(p))
451 sub_running_bw(dl_se, dl_rq);
452
453 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
454 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
455
456 if (READ_ONCE(p->__state) == TASK_DEAD)
457 sub_rq_bw(dl_se, &rq->dl);
458 raw_spin_lock(&dl_b->lock);
459 __dl_sub(dl_b, dl_se->dl_bw, dl_bw_cpus(task_cpu(p)));
460 raw_spin_unlock(&dl_b->lock);
461 __dl_clear_params(dl_se);
462 }
463 }
464
465 return;
466 }
467
468 dl_se->dl_non_contending = 1;
469 if (!dl_server(dl_se))
470 get_task_struct(dl_task_of(dl_se));
471
472 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
473}
474
475static void task_contending(struct sched_dl_entity *dl_se, int flags)
476{
477 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
478
479 /*
480 * If this is a non-deadline task that has been boosted,
481 * do nothing
482 */
483 if (dl_se->dl_runtime == 0)
484 return;
485
486 if (flags & ENQUEUE_MIGRATED)
487 add_rq_bw(dl_se, dl_rq);
488
489 if (dl_se->dl_non_contending) {
490 dl_se->dl_non_contending = 0;
491 /*
492 * If the timer handler is currently running and the
493 * timer cannot be canceled, inactive_task_timer()
494 * will see that dl_not_contending is not set, and
495 * will not touch the rq's active utilization,
496 * so we are still safe.
497 */
498 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) {
499 if (!dl_server(dl_se))
500 put_task_struct(dl_task_of(dl_se));
501 }
502 } else {
503 /*
504 * Since "dl_non_contending" is not set, the
505 * task's utilization has already been removed from
506 * active utilization (either when the task blocked,
507 * when the "inactive timer" fired).
508 * So, add it back.
509 */
510 add_running_bw(dl_se, dl_rq);
511 }
512}
513
514static inline int is_leftmost(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
515{
516 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
517}
518
519static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
520
521void init_dl_bw(struct dl_bw *dl_b)
522{
523 raw_spin_lock_init(&dl_b->lock);
524 if (global_rt_runtime() == RUNTIME_INF)
525 dl_b->bw = -1;
526 else
527 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
528 dl_b->total_bw = 0;
529}
530
531void init_dl_rq(struct dl_rq *dl_rq)
532{
533 dl_rq->root = RB_ROOT_CACHED;
534
535#ifdef CONFIG_SMP
536 /* zero means no -deadline tasks */
537 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
538
539 dl_rq->overloaded = 0;
540 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
541#else
542 init_dl_bw(&dl_rq->dl_bw);
543#endif
544
545 dl_rq->running_bw = 0;
546 dl_rq->this_bw = 0;
547 init_dl_rq_bw_ratio(dl_rq);
548}
549
550#ifdef CONFIG_SMP
551
552static inline int dl_overloaded(struct rq *rq)
553{
554 return atomic_read(&rq->rd->dlo_count);
555}
556
557static inline void dl_set_overload(struct rq *rq)
558{
559 if (!rq->online)
560 return;
561
562 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
563 /*
564 * Must be visible before the overload count is
565 * set (as in sched_rt.c).
566 *
567 * Matched by the barrier in pull_dl_task().
568 */
569 smp_wmb();
570 atomic_inc(&rq->rd->dlo_count);
571}
572
573static inline void dl_clear_overload(struct rq *rq)
574{
575 if (!rq->online)
576 return;
577
578 atomic_dec(&rq->rd->dlo_count);
579 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
580}
581
582#define __node_2_pdl(node) \
583 rb_entry((node), struct task_struct, pushable_dl_tasks)
584
585static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
586{
587 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
588}
589
590static inline int has_pushable_dl_tasks(struct rq *rq)
591{
592 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
593}
594
595/*
596 * The list of pushable -deadline task is not a plist, like in
597 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
598 */
599static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
600{
601 struct rb_node *leftmost;
602
603 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
604
605 leftmost = rb_add_cached(&p->pushable_dl_tasks,
606 &rq->dl.pushable_dl_tasks_root,
607 __pushable_less);
608 if (leftmost)
609 rq->dl.earliest_dl.next = p->dl.deadline;
610
611 if (!rq->dl.overloaded) {
612 dl_set_overload(rq);
613 rq->dl.overloaded = 1;
614 }
615}
616
617static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
618{
619 struct dl_rq *dl_rq = &rq->dl;
620 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
621 struct rb_node *leftmost;
622
623 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
624 return;
625
626 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
627 if (leftmost)
628 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
629
630 RB_CLEAR_NODE(&p->pushable_dl_tasks);
631
632 if (!has_pushable_dl_tasks(rq) && rq->dl.overloaded) {
633 dl_clear_overload(rq);
634 rq->dl.overloaded = 0;
635 }
636}
637
638static int push_dl_task(struct rq *rq);
639
640static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
641{
642 return rq->online && dl_task(prev);
643}
644
645static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
646static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
647
648static void push_dl_tasks(struct rq *);
649static void pull_dl_task(struct rq *);
650
651static inline void deadline_queue_push_tasks(struct rq *rq)
652{
653 if (!has_pushable_dl_tasks(rq))
654 return;
655
656 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
657}
658
659static inline void deadline_queue_pull_task(struct rq *rq)
660{
661 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
662}
663
664static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
665
666static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
667{
668 struct rq *later_rq = NULL;
669 struct dl_bw *dl_b;
670
671 later_rq = find_lock_later_rq(p, rq);
672 if (!later_rq) {
673 int cpu;
674
675 /*
676 * If we cannot preempt any rq, fall back to pick any
677 * online CPU:
678 */
679 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
680 if (cpu >= nr_cpu_ids) {
681 /*
682 * Failed to find any suitable CPU.
683 * The task will never come back!
684 */
685 WARN_ON_ONCE(dl_bandwidth_enabled());
686
687 /*
688 * If admission control is disabled we
689 * try a little harder to let the task
690 * run.
691 */
692 cpu = cpumask_any(cpu_active_mask);
693 }
694 later_rq = cpu_rq(cpu);
695 double_lock_balance(rq, later_rq);
696 }
697
698 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
699 /*
700 * Inactive timer is armed (or callback is running, but
701 * waiting for us to release rq locks). In any case, when it
702 * will fire (or continue), it will see running_bw of this
703 * task migrated to later_rq (and correctly handle it).
704 */
705 sub_running_bw(&p->dl, &rq->dl);
706 sub_rq_bw(&p->dl, &rq->dl);
707
708 add_rq_bw(&p->dl, &later_rq->dl);
709 add_running_bw(&p->dl, &later_rq->dl);
710 } else {
711 sub_rq_bw(&p->dl, &rq->dl);
712 add_rq_bw(&p->dl, &later_rq->dl);
713 }
714
715 /*
716 * And we finally need to fix up root_domain(s) bandwidth accounting,
717 * since p is still hanging out in the old (now moved to default) root
718 * domain.
719 */
720 dl_b = &rq->rd->dl_bw;
721 raw_spin_lock(&dl_b->lock);
722 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
723 raw_spin_unlock(&dl_b->lock);
724
725 dl_b = &later_rq->rd->dl_bw;
726 raw_spin_lock(&dl_b->lock);
727 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
728 raw_spin_unlock(&dl_b->lock);
729
730 set_task_cpu(p, later_rq->cpu);
731 double_unlock_balance(later_rq, rq);
732
733 return later_rq;
734}
735
736#else
737
738static inline
739void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
740{
741}
742
743static inline
744void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
745{
746}
747
748static inline
749void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
750{
751}
752
753static inline
754void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
755{
756}
757
758static inline void deadline_queue_push_tasks(struct rq *rq)
759{
760}
761
762static inline void deadline_queue_pull_task(struct rq *rq)
763{
764}
765#endif /* CONFIG_SMP */
766
767static void
768enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags);
769static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
770static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags);
771static void wakeup_preempt_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 * If it is a deferred reservation, and the server
782 * is not handling an starvation case, defer it.
783 */
784 if (dl_se->dl_defer && !dl_se->dl_defer_running) {
785 dl_se->dl_throttled = 1;
786 dl_se->dl_defer_armed = 1;
787 }
788}
789
790/*
791 * We are being explicitly informed that a new instance is starting,
792 * and this means that:
793 * - the absolute deadline of the entity has to be placed at
794 * current time + relative deadline;
795 * - the runtime of the entity has to be set to the maximum value.
796 *
797 * The capability of specifying such event is useful whenever a -deadline
798 * entity wants to (try to!) synchronize its behaviour with the scheduler's
799 * one, and to (try to!) reconcile itself with its own scheduling
800 * parameters.
801 */
802static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
803{
804 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
805 struct rq *rq = rq_of_dl_rq(dl_rq);
806
807 WARN_ON(is_dl_boosted(dl_se));
808 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
809
810 /*
811 * We are racing with the deadline timer. So, do nothing because
812 * the deadline timer handler will take care of properly recharging
813 * the runtime and postponing the deadline
814 */
815 if (dl_se->dl_throttled)
816 return;
817
818 /*
819 * We use the regular wall clock time to set deadlines in the
820 * future; in fact, we must consider execution overheads (time
821 * spent on hardirq context, etc.).
822 */
823 replenish_dl_new_period(dl_se, rq);
824}
825
826static int start_dl_timer(struct sched_dl_entity *dl_se);
827static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t);
828
829/*
830 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
831 * possibility of a entity lasting more than what it declared, and thus
832 * exhausting its runtime.
833 *
834 * Here we are interested in making runtime overrun possible, but we do
835 * not want a entity which is misbehaving to affect the scheduling of all
836 * other entities.
837 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
838 * is used, in order to confine each entity within its own bandwidth.
839 *
840 * This function deals exactly with that, and ensures that when the runtime
841 * of a entity is replenished, its deadline is also postponed. That ensures
842 * the overrunning entity can't interfere with other entity in the system and
843 * can't make them miss their deadlines. Reasons why this kind of overruns
844 * could happen are, typically, a entity voluntarily trying to overcome its
845 * runtime, or it just underestimated it during sched_setattr().
846 */
847static void replenish_dl_entity(struct sched_dl_entity *dl_se)
848{
849 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
850 struct rq *rq = rq_of_dl_rq(dl_rq);
851
852 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
853
854 /*
855 * This could be the case for a !-dl task that is boosted.
856 * Just go with full inherited parameters.
857 *
858 * Or, it could be the case of a deferred reservation that
859 * was not able to consume its runtime in background and
860 * reached this point with current u > U.
861 *
862 * In both cases, set a new period.
863 */
864 if (dl_se->dl_deadline == 0 ||
865 (dl_se->dl_defer_armed && dl_entity_overflow(dl_se, rq_clock(rq)))) {
866 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
867 dl_se->runtime = pi_of(dl_se)->dl_runtime;
868 }
869
870 if (dl_se->dl_yielded && dl_se->runtime > 0)
871 dl_se->runtime = 0;
872
873 /*
874 * We keep moving the deadline away until we get some
875 * available runtime for the entity. This ensures correct
876 * handling of situations where the runtime overrun is
877 * arbitrary large.
878 */
879 while (dl_se->runtime <= 0) {
880 dl_se->deadline += pi_of(dl_se)->dl_period;
881 dl_se->runtime += pi_of(dl_se)->dl_runtime;
882 }
883
884 /*
885 * At this point, the deadline really should be "in
886 * the future" with respect to rq->clock. If it's
887 * not, we are, for some reason, lagging too much!
888 * Anyway, after having warn userspace abut that,
889 * we still try to keep the things running by
890 * resetting the deadline and the budget of the
891 * entity.
892 */
893 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
894 printk_deferred_once("sched: DL replenish lagged too much\n");
895 replenish_dl_new_period(dl_se, rq);
896 }
897
898 if (dl_se->dl_yielded)
899 dl_se->dl_yielded = 0;
900 if (dl_se->dl_throttled)
901 dl_se->dl_throttled = 0;
902
903 /*
904 * If this is the replenishment of a deferred reservation,
905 * clear the flag and return.
906 */
907 if (dl_se->dl_defer_armed) {
908 dl_se->dl_defer_armed = 0;
909 return;
910 }
911
912 /*
913 * A this point, if the deferred server is not armed, and the deadline
914 * is in the future, if it is not running already, throttle the server
915 * and arm the defer timer.
916 */
917 if (dl_se->dl_defer && !dl_se->dl_defer_running &&
918 dl_time_before(rq_clock(dl_se->rq), dl_se->deadline - dl_se->runtime)) {
919 if (!is_dl_boosted(dl_se) && dl_se->server_has_tasks(dl_se)) {
920
921 /*
922 * Set dl_se->dl_defer_armed and dl_throttled variables to
923 * inform the start_dl_timer() that this is a deferred
924 * activation.
925 */
926 dl_se->dl_defer_armed = 1;
927 dl_se->dl_throttled = 1;
928 if (!start_dl_timer(dl_se)) {
929 /*
930 * If for whatever reason (delays), a previous timer was
931 * queued but not serviced, cancel it and clean the
932 * deferrable server variables intended for start_dl_timer().
933 */
934 hrtimer_try_to_cancel(&dl_se->dl_timer);
935 dl_se->dl_defer_armed = 0;
936 dl_se->dl_throttled = 0;
937 }
938 }
939 }
940}
941
942/*
943 * Here we check if --at time t-- an entity (which is probably being
944 * [re]activated or, in general, enqueued) can use its remaining runtime
945 * and its current deadline _without_ exceeding the bandwidth it is
946 * assigned (function returns true if it can't). We are in fact applying
947 * one of the CBS rules: when a task wakes up, if the residual runtime
948 * over residual deadline fits within the allocated bandwidth, then we
949 * can keep the current (absolute) deadline and residual budget without
950 * disrupting the schedulability of the system. Otherwise, we should
951 * refill the runtime and set the deadline a period in the future,
952 * because keeping the current (absolute) deadline of the task would
953 * result in breaking guarantees promised to other tasks (refer to
954 * Documentation/scheduler/sched-deadline.rst for more information).
955 *
956 * This function returns true if:
957 *
958 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
959 *
960 * IOW we can't recycle current parameters.
961 *
962 * Notice that the bandwidth check is done against the deadline. For
963 * task with deadline equal to period this is the same of using
964 * dl_period instead of dl_deadline in the equation above.
965 */
966static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
967{
968 u64 left, right;
969
970 /*
971 * left and right are the two sides of the equation above,
972 * after a bit of shuffling to use multiplications instead
973 * of divisions.
974 *
975 * Note that none of the time values involved in the two
976 * multiplications are absolute: dl_deadline and dl_runtime
977 * are the relative deadline and the maximum runtime of each
978 * instance, runtime is the runtime left for the last instance
979 * and (deadline - t), since t is rq->clock, is the time left
980 * to the (absolute) deadline. Even if overflowing the u64 type
981 * is very unlikely to occur in both cases, here we scale down
982 * as we want to avoid that risk at all. Scaling down by 10
983 * means that we reduce granularity to 1us. We are fine with it,
984 * since this is only a true/false check and, anyway, thinking
985 * of anything below microseconds resolution is actually fiction
986 * (but still we want to give the user that illusion >;).
987 */
988 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
989 right = ((dl_se->deadline - t) >> DL_SCALE) *
990 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
991
992 return dl_time_before(right, left);
993}
994
995/*
996 * Revised wakeup rule [1]: For self-suspending tasks, rather then
997 * re-initializing task's runtime and deadline, the revised wakeup
998 * rule adjusts the task's runtime to avoid the task to overrun its
999 * density.
1000 *
1001 * Reasoning: a task may overrun the density if:
1002 * runtime / (deadline - t) > dl_runtime / dl_deadline
1003 *
1004 * Therefore, runtime can be adjusted to:
1005 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
1006 *
1007 * In such way that runtime will be equal to the maximum density
1008 * the task can use without breaking any rule.
1009 *
1010 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
1011 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
1012 */
1013static void
1014update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
1015{
1016 u64 laxity = dl_se->deadline - rq_clock(rq);
1017
1018 /*
1019 * If the task has deadline < period, and the deadline is in the past,
1020 * it should already be throttled before this check.
1021 *
1022 * See update_dl_entity() comments for further details.
1023 */
1024 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
1025
1026 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
1027}
1028
1029/*
1030 * Regarding the deadline, a task with implicit deadline has a relative
1031 * deadline == relative period. A task with constrained deadline has a
1032 * relative deadline <= relative period.
1033 *
1034 * We support constrained deadline tasks. However, there are some restrictions
1035 * applied only for tasks which do not have an implicit deadline. See
1036 * update_dl_entity() to know more about such restrictions.
1037 *
1038 * The dl_is_implicit() returns true if the task has an implicit deadline.
1039 */
1040static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
1041{
1042 return dl_se->dl_deadline == dl_se->dl_period;
1043}
1044
1045/*
1046 * When a deadline entity is placed in the runqueue, its runtime and deadline
1047 * might need to be updated. This is done by a CBS wake up rule. There are two
1048 * different rules: 1) the original CBS; and 2) the Revisited CBS.
1049 *
1050 * When the task is starting a new period, the Original CBS is used. In this
1051 * case, the runtime is replenished and a new absolute deadline is set.
1052 *
1053 * When a task is queued before the begin of the next period, using the
1054 * remaining runtime and deadline could make the entity to overflow, see
1055 * dl_entity_overflow() to find more about runtime overflow. When such case
1056 * is detected, the runtime and deadline need to be updated.
1057 *
1058 * If the task has an implicit deadline, i.e., deadline == period, the Original
1059 * CBS is applied. The runtime is replenished and a new absolute deadline is
1060 * set, as in the previous cases.
1061 *
1062 * However, the Original CBS does not work properly for tasks with
1063 * deadline < period, which are said to have a constrained deadline. By
1064 * applying the Original CBS, a constrained deadline task would be able to run
1065 * runtime/deadline in a period. With deadline < period, the task would
1066 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1067 *
1068 * In order to prevent this misbehave, the Revisited CBS is used for
1069 * constrained deadline tasks when a runtime overflow is detected. In the
1070 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1071 * the remaining runtime of the task is reduced to avoid runtime overflow.
1072 * Please refer to the comments update_dl_revised_wakeup() function to find
1073 * more about the Revised CBS rule.
1074 */
1075static void update_dl_entity(struct sched_dl_entity *dl_se)
1076{
1077 struct rq *rq = rq_of_dl_se(dl_se);
1078
1079 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1080 dl_entity_overflow(dl_se, rq_clock(rq))) {
1081
1082 if (unlikely(!dl_is_implicit(dl_se) &&
1083 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1084 !is_dl_boosted(dl_se))) {
1085 update_dl_revised_wakeup(dl_se, rq);
1086 return;
1087 }
1088
1089 replenish_dl_new_period(dl_se, rq);
1090 } else if (dl_server(dl_se) && dl_se->dl_defer) {
1091 /*
1092 * The server can still use its previous deadline, so check if
1093 * it left the dl_defer_running state.
1094 */
1095 if (!dl_se->dl_defer_running) {
1096 dl_se->dl_defer_armed = 1;
1097 dl_se->dl_throttled = 1;
1098 }
1099 }
1100}
1101
1102static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1103{
1104 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1105}
1106
1107/*
1108 * If the entity depleted all its runtime, and if we want it to sleep
1109 * while waiting for some new execution time to become available, we
1110 * set the bandwidth replenishment timer to the replenishment instant
1111 * and try to activate it.
1112 *
1113 * Notice that it is important for the caller to know if the timer
1114 * actually started or not (i.e., the replenishment instant is in
1115 * the future or in the past).
1116 */
1117static int start_dl_timer(struct sched_dl_entity *dl_se)
1118{
1119 struct hrtimer *timer = &dl_se->dl_timer;
1120 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1121 struct rq *rq = rq_of_dl_rq(dl_rq);
1122 ktime_t now, act;
1123 s64 delta;
1124
1125 lockdep_assert_rq_held(rq);
1126
1127 /*
1128 * We want the timer to fire at the deadline, but considering
1129 * that it is actually coming from rq->clock and not from
1130 * hrtimer's time base reading.
1131 *
1132 * The deferred reservation will have its timer set to
1133 * (deadline - runtime). At that point, the CBS rule will decide
1134 * if the current deadline can be used, or if a replenishment is
1135 * required to avoid add too much pressure on the system
1136 * (current u > U).
1137 */
1138 if (dl_se->dl_defer_armed) {
1139 WARN_ON_ONCE(!dl_se->dl_throttled);
1140 act = ns_to_ktime(dl_se->deadline - dl_se->runtime);
1141 } else {
1142 /* act = deadline - rel-deadline + period */
1143 act = ns_to_ktime(dl_next_period(dl_se));
1144 }
1145
1146 now = hrtimer_cb_get_time(timer);
1147 delta = ktime_to_ns(now) - rq_clock(rq);
1148 act = ktime_add_ns(act, delta);
1149
1150 /*
1151 * If the expiry time already passed, e.g., because the value
1152 * chosen as the deadline is too small, don't even try to
1153 * start the timer in the past!
1154 */
1155 if (ktime_us_delta(act, now) < 0)
1156 return 0;
1157
1158 /*
1159 * !enqueued will guarantee another callback; even if one is already in
1160 * progress. This ensures a balanced {get,put}_task_struct().
1161 *
1162 * The race against __run_timer() clearing the enqueued state is
1163 * harmless because we're holding task_rq()->lock, therefore the timer
1164 * expiring after we've done the check will wait on its task_rq_lock()
1165 * and observe our state.
1166 */
1167 if (!hrtimer_is_queued(timer)) {
1168 if (!dl_server(dl_se))
1169 get_task_struct(dl_task_of(dl_se));
1170 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1171 }
1172
1173 return 1;
1174}
1175
1176static void __push_dl_task(struct rq *rq, struct rq_flags *rf)
1177{
1178#ifdef CONFIG_SMP
1179 /*
1180 * Queueing this task back might have overloaded rq, check if we need
1181 * to kick someone away.
1182 */
1183 if (has_pushable_dl_tasks(rq)) {
1184 /*
1185 * Nothing relies on rq->lock after this, so its safe to drop
1186 * rq->lock.
1187 */
1188 rq_unpin_lock(rq, rf);
1189 push_dl_task(rq);
1190 rq_repin_lock(rq, rf);
1191 }
1192#endif
1193}
1194
1195/* a defer timer will not be reset if the runtime consumed was < dl_server_min_res */
1196static const u64 dl_server_min_res = 1 * NSEC_PER_MSEC;
1197
1198static enum hrtimer_restart dl_server_timer(struct hrtimer *timer, struct sched_dl_entity *dl_se)
1199{
1200 struct rq *rq = rq_of_dl_se(dl_se);
1201 u64 fw;
1202
1203 scoped_guard (rq_lock, rq) {
1204 struct rq_flags *rf = &scope.rf;
1205
1206 if (!dl_se->dl_throttled || !dl_se->dl_runtime)
1207 return HRTIMER_NORESTART;
1208
1209 sched_clock_tick();
1210 update_rq_clock(rq);
1211
1212 if (!dl_se->dl_runtime)
1213 return HRTIMER_NORESTART;
1214
1215 if (!dl_se->server_has_tasks(dl_se)) {
1216 replenish_dl_entity(dl_se);
1217 return HRTIMER_NORESTART;
1218 }
1219
1220 if (dl_se->dl_defer_armed) {
1221 /*
1222 * First check if the server could consume runtime in background.
1223 * If so, it is possible to push the defer timer for this amount
1224 * of time. The dl_server_min_res serves as a limit to avoid
1225 * forwarding the timer for a too small amount of time.
1226 */
1227 if (dl_time_before(rq_clock(dl_se->rq),
1228 (dl_se->deadline - dl_se->runtime - dl_server_min_res))) {
1229
1230 /* reset the defer timer */
1231 fw = dl_se->deadline - rq_clock(dl_se->rq) - dl_se->runtime;
1232
1233 hrtimer_forward_now(timer, ns_to_ktime(fw));
1234 return HRTIMER_RESTART;
1235 }
1236
1237 dl_se->dl_defer_running = 1;
1238 }
1239
1240 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1241
1242 if (!dl_task(dl_se->rq->curr) || dl_entity_preempt(dl_se, &dl_se->rq->curr->dl))
1243 resched_curr(rq);
1244
1245 __push_dl_task(rq, rf);
1246 }
1247
1248 return HRTIMER_NORESTART;
1249}
1250
1251/*
1252 * This is the bandwidth enforcement timer callback. If here, we know
1253 * a task is not on its dl_rq, since the fact that the timer was running
1254 * means the task is throttled and needs a runtime replenishment.
1255 *
1256 * However, what we actually do depends on the fact the task is active,
1257 * (it is on its rq) or has been removed from there by a call to
1258 * dequeue_task_dl(). In the former case we must issue the runtime
1259 * replenishment and add the task back to the dl_rq; in the latter, we just
1260 * do nothing but clearing dl_throttled, so that runtime and deadline
1261 * updating (and the queueing back to dl_rq) will be done by the
1262 * next call to enqueue_task_dl().
1263 */
1264static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1265{
1266 struct sched_dl_entity *dl_se = container_of(timer,
1267 struct sched_dl_entity,
1268 dl_timer);
1269 struct task_struct *p;
1270 struct rq_flags rf;
1271 struct rq *rq;
1272
1273 if (dl_server(dl_se))
1274 return dl_server_timer(timer, dl_se);
1275
1276 p = dl_task_of(dl_se);
1277 rq = task_rq_lock(p, &rf);
1278
1279 /*
1280 * The task might have changed its scheduling policy to something
1281 * different than SCHED_DEADLINE (through switched_from_dl()).
1282 */
1283 if (!dl_task(p))
1284 goto unlock;
1285
1286 /*
1287 * The task might have been boosted by someone else and might be in the
1288 * boosting/deboosting path, its not throttled.
1289 */
1290 if (is_dl_boosted(dl_se))
1291 goto unlock;
1292
1293 /*
1294 * Spurious timer due to start_dl_timer() race; or we already received
1295 * a replenishment from rt_mutex_setprio().
1296 */
1297 if (!dl_se->dl_throttled)
1298 goto unlock;
1299
1300 sched_clock_tick();
1301 update_rq_clock(rq);
1302
1303 /*
1304 * If the throttle happened during sched-out; like:
1305 *
1306 * schedule()
1307 * deactivate_task()
1308 * dequeue_task_dl()
1309 * update_curr_dl()
1310 * start_dl_timer()
1311 * __dequeue_task_dl()
1312 * prev->on_rq = 0;
1313 *
1314 * We can be both throttled and !queued. Replenish the counter
1315 * but do not enqueue -- wait for our wakeup to do that.
1316 */
1317 if (!task_on_rq_queued(p)) {
1318 replenish_dl_entity(dl_se);
1319 goto unlock;
1320 }
1321
1322#ifdef CONFIG_SMP
1323 if (unlikely(!rq->online)) {
1324 /*
1325 * If the runqueue is no longer available, migrate the
1326 * task elsewhere. This necessarily changes rq.
1327 */
1328 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1329 rq = dl_task_offline_migration(rq, p);
1330 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1331 update_rq_clock(rq);
1332
1333 /*
1334 * Now that the task has been migrated to the new RQ and we
1335 * have that locked, proceed as normal and enqueue the task
1336 * there.
1337 */
1338 }
1339#endif
1340
1341 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1342 if (dl_task(rq->donor))
1343 wakeup_preempt_dl(rq, p, 0);
1344 else
1345 resched_curr(rq);
1346
1347 __push_dl_task(rq, &rf);
1348
1349unlock:
1350 task_rq_unlock(rq, p, &rf);
1351
1352 /*
1353 * This can free the task_struct, including this hrtimer, do not touch
1354 * anything related to that after this.
1355 */
1356 put_task_struct(p);
1357
1358 return HRTIMER_NORESTART;
1359}
1360
1361static void init_dl_task_timer(struct sched_dl_entity *dl_se)
1362{
1363 struct hrtimer *timer = &dl_se->dl_timer;
1364
1365 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1366 timer->function = dl_task_timer;
1367}
1368
1369/*
1370 * During the activation, CBS checks if it can reuse the current task's
1371 * runtime and period. If the deadline of the task is in the past, CBS
1372 * cannot use the runtime, and so it replenishes the task. This rule
1373 * works fine for implicit deadline tasks (deadline == period), and the
1374 * CBS was designed for implicit deadline tasks. However, a task with
1375 * constrained deadline (deadline < period) might be awakened after the
1376 * deadline, but before the next period. In this case, replenishing the
1377 * task would allow it to run for runtime / deadline. As in this case
1378 * deadline < period, CBS enables a task to run for more than the
1379 * runtime / period. In a very loaded system, this can cause a domino
1380 * effect, making other tasks miss their deadlines.
1381 *
1382 * To avoid this problem, in the activation of a constrained deadline
1383 * task after the deadline but before the next period, throttle the
1384 * task and set the replenishing timer to the begin of the next period,
1385 * unless it is boosted.
1386 */
1387static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1388{
1389 struct rq *rq = rq_of_dl_se(dl_se);
1390
1391 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1392 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1393 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se)))
1394 return;
1395 dl_se->dl_throttled = 1;
1396 if (dl_se->runtime > 0)
1397 dl_se->runtime = 0;
1398 }
1399}
1400
1401static
1402int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1403{
1404 return (dl_se->runtime <= 0);
1405}
1406
1407/*
1408 * This function implements the GRUB accounting rule. According to the
1409 * GRUB reclaiming algorithm, the runtime is not decreased as "dq = -dt",
1410 * but as "dq = -(max{u, (Umax - Uinact - Uextra)} / Umax) dt",
1411 * where u is the utilization of the task, Umax is the maximum reclaimable
1412 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1413 * as the difference between the "total runqueue utilization" and the
1414 * "runqueue active utilization", and Uextra is the (per runqueue) extra
1415 * reclaimable utilization.
1416 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations multiplied
1417 * by 2^BW_SHIFT, the result has to be shifted right by BW_SHIFT.
1418 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT, dl_bw
1419 * is multiplied by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1420 * Since delta is a 64 bit variable, to have an overflow its value should be
1421 * larger than 2^(64 - 20 - 8), which is more than 64 seconds. So, overflow is
1422 * not an issue here.
1423 */
1424static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1425{
1426 u64 u_act;
1427 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1428
1429 /*
1430 * Instead of computing max{u, (u_max - u_inact - u_extra)}, we
1431 * compare u_inact + u_extra with u_max - u, because u_inact + u_extra
1432 * can be larger than u_max. So, u_max - u_inact - u_extra would be
1433 * negative leading to wrong results.
1434 */
1435 if (u_inact + rq->dl.extra_bw > rq->dl.max_bw - dl_se->dl_bw)
1436 u_act = dl_se->dl_bw;
1437 else
1438 u_act = rq->dl.max_bw - u_inact - rq->dl.extra_bw;
1439
1440 u_act = (u_act * rq->dl.bw_ratio) >> RATIO_SHIFT;
1441 return (delta * u_act) >> BW_SHIFT;
1442}
1443
1444s64 dl_scaled_delta_exec(struct rq *rq, struct sched_dl_entity *dl_se, s64 delta_exec)
1445{
1446 s64 scaled_delta_exec;
1447
1448 /*
1449 * For tasks that participate in GRUB, we implement GRUB-PA: the
1450 * spare reclaimed bandwidth is used to clock down frequency.
1451 *
1452 * For the others, we still need to scale reservation parameters
1453 * according to current frequency and CPU maximum capacity.
1454 */
1455 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1456 scaled_delta_exec = grub_reclaim(delta_exec, rq, dl_se);
1457 } else {
1458 int cpu = cpu_of(rq);
1459 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1460 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1461
1462 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1463 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1464 }
1465
1466 return scaled_delta_exec;
1467}
1468
1469static inline void
1470update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1471 int flags);
1472static void update_curr_dl_se(struct rq *rq, struct sched_dl_entity *dl_se, s64 delta_exec)
1473{
1474 s64 scaled_delta_exec;
1475
1476 if (unlikely(delta_exec <= 0)) {
1477 if (unlikely(dl_se->dl_yielded))
1478 goto throttle;
1479 return;
1480 }
1481
1482 if (dl_server(dl_se) && dl_se->dl_throttled && !dl_se->dl_defer)
1483 return;
1484
1485 if (dl_entity_is_special(dl_se))
1486 return;
1487
1488 scaled_delta_exec = dl_scaled_delta_exec(rq, dl_se, delta_exec);
1489
1490 dl_se->runtime -= scaled_delta_exec;
1491
1492 /*
1493 * The fair server can consume its runtime while throttled (not queued/
1494 * running as regular CFS).
1495 *
1496 * If the server consumes its entire runtime in this state. The server
1497 * is not required for the current period. Thus, reset the server by
1498 * starting a new period, pushing the activation.
1499 */
1500 if (dl_se->dl_defer && dl_se->dl_throttled && dl_runtime_exceeded(dl_se)) {
1501 /*
1502 * If the server was previously activated - the starving condition
1503 * took place, it this point it went away because the fair scheduler
1504 * was able to get runtime in background. So return to the initial
1505 * state.
1506 */
1507 dl_se->dl_defer_running = 0;
1508
1509 hrtimer_try_to_cancel(&dl_se->dl_timer);
1510
1511 replenish_dl_new_period(dl_se, dl_se->rq);
1512
1513 /*
1514 * Not being able to start the timer seems problematic. If it could not
1515 * be started for whatever reason, we need to "unthrottle" the DL server
1516 * and queue right away. Otherwise nothing might queue it. That's similar
1517 * to what enqueue_dl_entity() does on start_dl_timer==0. For now, just warn.
1518 */
1519 WARN_ON_ONCE(!start_dl_timer(dl_se));
1520
1521 return;
1522 }
1523
1524throttle:
1525 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1526 dl_se->dl_throttled = 1;
1527
1528 /* If requested, inform the user about runtime overruns. */
1529 if (dl_runtime_exceeded(dl_se) &&
1530 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1531 dl_se->dl_overrun = 1;
1532
1533 dequeue_dl_entity(dl_se, 0);
1534 if (!dl_server(dl_se)) {
1535 update_stats_dequeue_dl(&rq->dl, dl_se, 0);
1536 dequeue_pushable_dl_task(rq, dl_task_of(dl_se));
1537 }
1538
1539 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se))) {
1540 if (dl_server(dl_se))
1541 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1542 else
1543 enqueue_task_dl(rq, dl_task_of(dl_se), ENQUEUE_REPLENISH);
1544 }
1545
1546 if (!is_leftmost(dl_se, &rq->dl))
1547 resched_curr(rq);
1548 }
1549
1550 /*
1551 * The fair server (sole dl_server) does not account for real-time
1552 * workload because it is running fair work.
1553 */
1554 if (dl_se == &rq->fair_server)
1555 return;
1556
1557#ifdef CONFIG_RT_GROUP_SCHED
1558 /*
1559 * Because -- for now -- we share the rt bandwidth, we need to
1560 * account our runtime there too, otherwise actual rt tasks
1561 * would be able to exceed the shared quota.
1562 *
1563 * Account to the root rt group for now.
1564 *
1565 * The solution we're working towards is having the RT groups scheduled
1566 * using deadline servers -- however there's a few nasties to figure
1567 * out before that can happen.
1568 */
1569 if (rt_bandwidth_enabled()) {
1570 struct rt_rq *rt_rq = &rq->rt;
1571
1572 raw_spin_lock(&rt_rq->rt_runtime_lock);
1573 /*
1574 * We'll let actual RT tasks worry about the overflow here, we
1575 * have our own CBS to keep us inline; only account when RT
1576 * bandwidth is relevant.
1577 */
1578 if (sched_rt_bandwidth_account(rt_rq))
1579 rt_rq->rt_time += delta_exec;
1580 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1581 }
1582#endif
1583}
1584
1585/*
1586 * In the non-defer mode, the idle time is not accounted, as the
1587 * server provides a guarantee.
1588 *
1589 * If the dl_server is in defer mode, the idle time is also considered
1590 * as time available for the fair server, avoiding a penalty for the
1591 * rt scheduler that did not consumed that time.
1592 */
1593void dl_server_update_idle_time(struct rq *rq, struct task_struct *p)
1594{
1595 s64 delta_exec, scaled_delta_exec;
1596
1597 if (!rq->fair_server.dl_defer)
1598 return;
1599
1600 /* no need to discount more */
1601 if (rq->fair_server.runtime < 0)
1602 return;
1603
1604 delta_exec = rq_clock_task(rq) - p->se.exec_start;
1605 if (delta_exec < 0)
1606 return;
1607
1608 scaled_delta_exec = dl_scaled_delta_exec(rq, &rq->fair_server, delta_exec);
1609
1610 rq->fair_server.runtime -= scaled_delta_exec;
1611
1612 if (rq->fair_server.runtime < 0) {
1613 rq->fair_server.dl_defer_running = 0;
1614 rq->fair_server.runtime = 0;
1615 }
1616
1617 p->se.exec_start = rq_clock_task(rq);
1618}
1619
1620void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec)
1621{
1622 /* 0 runtime = fair server disabled */
1623 if (dl_se->dl_runtime)
1624 update_curr_dl_se(dl_se->rq, dl_se, delta_exec);
1625}
1626
1627void dl_server_start(struct sched_dl_entity *dl_se)
1628{
1629 struct rq *rq = dl_se->rq;
1630
1631 /*
1632 * XXX: the apply do not work fine at the init phase for the
1633 * fair server because things are not yet set. We need to improve
1634 * this before getting generic.
1635 */
1636 if (!dl_server(dl_se)) {
1637 u64 runtime = 50 * NSEC_PER_MSEC;
1638 u64 period = 1000 * NSEC_PER_MSEC;
1639
1640 dl_server_apply_params(dl_se, runtime, period, 1);
1641
1642 dl_se->dl_server = 1;
1643 dl_se->dl_defer = 1;
1644 setup_new_dl_entity(dl_se);
1645 }
1646
1647 if (!dl_se->dl_runtime)
1648 return;
1649
1650 dl_se->dl_server_active = 1;
1651 enqueue_dl_entity(dl_se, ENQUEUE_WAKEUP);
1652 if (!dl_task(dl_se->rq->curr) || dl_entity_preempt(dl_se, &rq->curr->dl))
1653 resched_curr(dl_se->rq);
1654}
1655
1656void dl_server_stop(struct sched_dl_entity *dl_se)
1657{
1658 if (!dl_se->dl_runtime)
1659 return;
1660
1661 dequeue_dl_entity(dl_se, DEQUEUE_SLEEP);
1662 hrtimer_try_to_cancel(&dl_se->dl_timer);
1663 dl_se->dl_defer_armed = 0;
1664 dl_se->dl_throttled = 0;
1665 dl_se->dl_server_active = 0;
1666}
1667
1668void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
1669 dl_server_has_tasks_f has_tasks,
1670 dl_server_pick_f pick_task)
1671{
1672 dl_se->rq = rq;
1673 dl_se->server_has_tasks = has_tasks;
1674 dl_se->server_pick_task = pick_task;
1675}
1676
1677void __dl_server_attach_root(struct sched_dl_entity *dl_se, struct rq *rq)
1678{
1679 u64 new_bw = dl_se->dl_bw;
1680 int cpu = cpu_of(rq);
1681 struct dl_bw *dl_b;
1682
1683 dl_b = dl_bw_of(cpu_of(rq));
1684 guard(raw_spinlock)(&dl_b->lock);
1685
1686 if (!dl_bw_cpus(cpu))
1687 return;
1688
1689 __dl_add(dl_b, new_bw, dl_bw_cpus(cpu));
1690}
1691
1692int dl_server_apply_params(struct sched_dl_entity *dl_se, u64 runtime, u64 period, bool init)
1693{
1694 u64 old_bw = init ? 0 : to_ratio(dl_se->dl_period, dl_se->dl_runtime);
1695 u64 new_bw = to_ratio(period, runtime);
1696 struct rq *rq = dl_se->rq;
1697 int cpu = cpu_of(rq);
1698 struct dl_bw *dl_b;
1699 unsigned long cap;
1700 int retval = 0;
1701 int cpus;
1702
1703 dl_b = dl_bw_of(cpu);
1704 guard(raw_spinlock)(&dl_b->lock);
1705
1706 cpus = dl_bw_cpus(cpu);
1707 cap = dl_bw_capacity(cpu);
1708
1709 if (__dl_overflow(dl_b, cap, old_bw, new_bw))
1710 return -EBUSY;
1711
1712 if (init) {
1713 __add_rq_bw(new_bw, &rq->dl);
1714 __dl_add(dl_b, new_bw, cpus);
1715 } else {
1716 __dl_sub(dl_b, dl_se->dl_bw, cpus);
1717 __dl_add(dl_b, new_bw, cpus);
1718
1719 dl_rq_change_utilization(rq, dl_se, new_bw);
1720 }
1721
1722 dl_se->dl_runtime = runtime;
1723 dl_se->dl_deadline = period;
1724 dl_se->dl_period = period;
1725
1726 dl_se->runtime = 0;
1727 dl_se->deadline = 0;
1728
1729 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
1730 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
1731
1732 return retval;
1733}
1734
1735/*
1736 * Update the current task's runtime statistics (provided it is still
1737 * a -deadline task and has not been removed from the dl_rq).
1738 */
1739static void update_curr_dl(struct rq *rq)
1740{
1741 struct task_struct *donor = rq->donor;
1742 struct sched_dl_entity *dl_se = &donor->dl;
1743 s64 delta_exec;
1744
1745 if (!dl_task(donor) || !on_dl_rq(dl_se))
1746 return;
1747
1748 /*
1749 * Consumed budget is computed considering the time as
1750 * observed by schedulable tasks (excluding time spent
1751 * in hardirq context, etc.). Deadlines are instead
1752 * computed using hard walltime. This seems to be the more
1753 * natural solution, but the full ramifications of this
1754 * approach need further study.
1755 */
1756 delta_exec = update_curr_common(rq);
1757 update_curr_dl_se(rq, dl_se, delta_exec);
1758}
1759
1760static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1761{
1762 struct sched_dl_entity *dl_se = container_of(timer,
1763 struct sched_dl_entity,
1764 inactive_timer);
1765 struct task_struct *p = NULL;
1766 struct rq_flags rf;
1767 struct rq *rq;
1768
1769 if (!dl_server(dl_se)) {
1770 p = dl_task_of(dl_se);
1771 rq = task_rq_lock(p, &rf);
1772 } else {
1773 rq = dl_se->rq;
1774 rq_lock(rq, &rf);
1775 }
1776
1777 sched_clock_tick();
1778 update_rq_clock(rq);
1779
1780 if (dl_server(dl_se))
1781 goto no_task;
1782
1783 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1784 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1785
1786 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1787 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1788 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1789 dl_se->dl_non_contending = 0;
1790 }
1791
1792 raw_spin_lock(&dl_b->lock);
1793 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1794 raw_spin_unlock(&dl_b->lock);
1795 __dl_clear_params(dl_se);
1796
1797 goto unlock;
1798 }
1799
1800no_task:
1801 if (dl_se->dl_non_contending == 0)
1802 goto unlock;
1803
1804 sub_running_bw(dl_se, &rq->dl);
1805 dl_se->dl_non_contending = 0;
1806unlock:
1807
1808 if (!dl_server(dl_se)) {
1809 task_rq_unlock(rq, p, &rf);
1810 put_task_struct(p);
1811 } else {
1812 rq_unlock(rq, &rf);
1813 }
1814
1815 return HRTIMER_NORESTART;
1816}
1817
1818static void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1819{
1820 struct hrtimer *timer = &dl_se->inactive_timer;
1821
1822 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1823 timer->function = inactive_task_timer;
1824}
1825
1826#define __node_2_dle(node) \
1827 rb_entry((node), struct sched_dl_entity, rb_node)
1828
1829#ifdef CONFIG_SMP
1830
1831static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1832{
1833 struct rq *rq = rq_of_dl_rq(dl_rq);
1834
1835 if (dl_rq->earliest_dl.curr == 0 ||
1836 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1837 if (dl_rq->earliest_dl.curr == 0)
1838 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1839 dl_rq->earliest_dl.curr = deadline;
1840 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1841 }
1842}
1843
1844static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1845{
1846 struct rq *rq = rq_of_dl_rq(dl_rq);
1847
1848 /*
1849 * Since we may have removed our earliest (and/or next earliest)
1850 * task we must recompute them.
1851 */
1852 if (!dl_rq->dl_nr_running) {
1853 dl_rq->earliest_dl.curr = 0;
1854 dl_rq->earliest_dl.next = 0;
1855 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1856 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1857 } else {
1858 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1859 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1860
1861 dl_rq->earliest_dl.curr = entry->deadline;
1862 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1863 }
1864}
1865
1866#else
1867
1868static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1869static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1870
1871#endif /* CONFIG_SMP */
1872
1873static inline
1874void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1875{
1876 u64 deadline = dl_se->deadline;
1877
1878 dl_rq->dl_nr_running++;
1879 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1880
1881 inc_dl_deadline(dl_rq, deadline);
1882}
1883
1884static inline
1885void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1886{
1887 WARN_ON(!dl_rq->dl_nr_running);
1888 dl_rq->dl_nr_running--;
1889 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1890
1891 dec_dl_deadline(dl_rq, dl_se->deadline);
1892}
1893
1894static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1895{
1896 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1897}
1898
1899static __always_inline struct sched_statistics *
1900__schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1901{
1902 if (!schedstat_enabled())
1903 return NULL;
1904
1905 if (dl_server(dl_se))
1906 return NULL;
1907
1908 return &dl_task_of(dl_se)->stats;
1909}
1910
1911static inline void
1912update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1913{
1914 struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1915 if (stats)
1916 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1917}
1918
1919static inline void
1920update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1921{
1922 struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1923 if (stats)
1924 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1925}
1926
1927static inline void
1928update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1929{
1930 struct sched_statistics *stats = __schedstats_from_dl_se(dl_se);
1931 if (stats)
1932 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1933}
1934
1935static inline void
1936update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1937 int flags)
1938{
1939 if (!schedstat_enabled())
1940 return;
1941
1942 if (flags & ENQUEUE_WAKEUP)
1943 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1944}
1945
1946static inline void
1947update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1948 int flags)
1949{
1950 struct task_struct *p = dl_task_of(dl_se);
1951
1952 if (!schedstat_enabled())
1953 return;
1954
1955 if ((flags & DEQUEUE_SLEEP)) {
1956 unsigned int state;
1957
1958 state = READ_ONCE(p->__state);
1959 if (state & TASK_INTERRUPTIBLE)
1960 __schedstat_set(p->stats.sleep_start,
1961 rq_clock(rq_of_dl_rq(dl_rq)));
1962
1963 if (state & TASK_UNINTERRUPTIBLE)
1964 __schedstat_set(p->stats.block_start,
1965 rq_clock(rq_of_dl_rq(dl_rq)));
1966 }
1967}
1968
1969static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1970{
1971 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1972
1973 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1974
1975 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1976
1977 inc_dl_tasks(dl_se, dl_rq);
1978}
1979
1980static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1981{
1982 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1983
1984 if (RB_EMPTY_NODE(&dl_se->rb_node))
1985 return;
1986
1987 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1988
1989 RB_CLEAR_NODE(&dl_se->rb_node);
1990
1991 dec_dl_tasks(dl_se, dl_rq);
1992}
1993
1994static void
1995enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1996{
1997 WARN_ON_ONCE(on_dl_rq(dl_se));
1998
1999 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
2000
2001 /*
2002 * Check if a constrained deadline task was activated
2003 * after the deadline but before the next period.
2004 * If that is the case, the task will be throttled and
2005 * the replenishment timer will be set to the next period.
2006 */
2007 if (!dl_se->dl_throttled && !dl_is_implicit(dl_se))
2008 dl_check_constrained_dl(dl_se);
2009
2010 if (flags & (ENQUEUE_RESTORE|ENQUEUE_MIGRATING)) {
2011 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
2012
2013 add_rq_bw(dl_se, dl_rq);
2014 add_running_bw(dl_se, dl_rq);
2015 }
2016
2017 /*
2018 * If p is throttled, we do not enqueue it. In fact, if it exhausted
2019 * its budget it needs a replenishment and, since it now is on
2020 * its rq, the bandwidth timer callback (which clearly has not
2021 * run yet) will take care of this.
2022 * However, the active utilization does not depend on the fact
2023 * that the task is on the runqueue or not (but depends on the
2024 * task's state - in GRUB parlance, "inactive" vs "active contending").
2025 * In other words, even if a task is throttled its utilization must
2026 * be counted in the active utilization; hence, we need to call
2027 * add_running_bw().
2028 */
2029 if (!dl_se->dl_defer && dl_se->dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
2030 if (flags & ENQUEUE_WAKEUP)
2031 task_contending(dl_se, flags);
2032
2033 return;
2034 }
2035
2036 /*
2037 * If this is a wakeup or a new instance, the scheduling
2038 * parameters of the task might need updating. Otherwise,
2039 * we want a replenishment of its runtime.
2040 */
2041 if (flags & ENQUEUE_WAKEUP) {
2042 task_contending(dl_se, flags);
2043 update_dl_entity(dl_se);
2044 } else if (flags & ENQUEUE_REPLENISH) {
2045 replenish_dl_entity(dl_se);
2046 } else if ((flags & ENQUEUE_RESTORE) &&
2047 !is_dl_boosted(dl_se) &&
2048 dl_time_before(dl_se->deadline, rq_clock(rq_of_dl_se(dl_se)))) {
2049 setup_new_dl_entity(dl_se);
2050 }
2051
2052 /*
2053 * If the reservation is still throttled, e.g., it got replenished but is a
2054 * deferred task and still got to wait, don't enqueue.
2055 */
2056 if (dl_se->dl_throttled && start_dl_timer(dl_se))
2057 return;
2058
2059 /*
2060 * We're about to enqueue, make sure we're not ->dl_throttled!
2061 * In case the timer was not started, say because the defer time
2062 * has passed, mark as not throttled and mark unarmed.
2063 * Also cancel earlier timers, since letting those run is pointless.
2064 */
2065 if (dl_se->dl_throttled) {
2066 hrtimer_try_to_cancel(&dl_se->dl_timer);
2067 dl_se->dl_defer_armed = 0;
2068 dl_se->dl_throttled = 0;
2069 }
2070
2071 __enqueue_dl_entity(dl_se);
2072}
2073
2074static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags)
2075{
2076 __dequeue_dl_entity(dl_se);
2077
2078 if (flags & (DEQUEUE_SAVE|DEQUEUE_MIGRATING)) {
2079 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
2080
2081 sub_running_bw(dl_se, dl_rq);
2082 sub_rq_bw(dl_se, dl_rq);
2083 }
2084
2085 /*
2086 * This check allows to start the inactive timer (or to immediately
2087 * decrease the active utilization, if needed) in two cases:
2088 * when the task blocks and when it is terminating
2089 * (p->state == TASK_DEAD). We can handle the two cases in the same
2090 * way, because from GRUB's point of view the same thing is happening
2091 * (the task moves from "active contending" to "active non contending"
2092 * or "inactive")
2093 */
2094 if (flags & DEQUEUE_SLEEP)
2095 task_non_contending(dl_se);
2096}
2097
2098static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2099{
2100 if (is_dl_boosted(&p->dl)) {
2101 /*
2102 * Because of delays in the detection of the overrun of a
2103 * thread's runtime, it might be the case that a thread
2104 * goes to sleep in a rt mutex with negative runtime. As
2105 * a consequence, the thread will be throttled.
2106 *
2107 * While waiting for the mutex, this thread can also be
2108 * boosted via PI, resulting in a thread that is throttled
2109 * and boosted at the same time.
2110 *
2111 * In this case, the boost overrides the throttle.
2112 */
2113 if (p->dl.dl_throttled) {
2114 /*
2115 * The replenish timer needs to be canceled. No
2116 * problem if it fires concurrently: boosted threads
2117 * are ignored in dl_task_timer().
2118 *
2119 * If the timer callback was running (hrtimer_try_to_cancel == -1),
2120 * it will eventually call put_task_struct().
2121 */
2122 if (hrtimer_try_to_cancel(&p->dl.dl_timer) == 1 &&
2123 !dl_server(&p->dl))
2124 put_task_struct(p);
2125 p->dl.dl_throttled = 0;
2126 }
2127 } else if (!dl_prio(p->normal_prio)) {
2128 /*
2129 * Special case in which we have a !SCHED_DEADLINE task that is going
2130 * to be deboosted, but exceeds its runtime while doing so. No point in
2131 * replenishing it, as it's going to return back to its original
2132 * scheduling class after this. If it has been throttled, we need to
2133 * clear the flag, otherwise the task may wake up as throttled after
2134 * being boosted again with no means to replenish the runtime and clear
2135 * the throttle.
2136 */
2137 p->dl.dl_throttled = 0;
2138 if (!(flags & ENQUEUE_REPLENISH))
2139 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
2140 task_pid_nr(p));
2141
2142 return;
2143 }
2144
2145 check_schedstat_required();
2146 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
2147
2148 if (p->on_rq == TASK_ON_RQ_MIGRATING)
2149 flags |= ENQUEUE_MIGRATING;
2150
2151 enqueue_dl_entity(&p->dl, flags);
2152
2153 if (dl_server(&p->dl))
2154 return;
2155
2156 if (!task_current(rq, p) && !p->dl.dl_throttled && p->nr_cpus_allowed > 1)
2157 enqueue_pushable_dl_task(rq, p);
2158}
2159
2160static bool dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2161{
2162 update_curr_dl(rq);
2163
2164 if (p->on_rq == TASK_ON_RQ_MIGRATING)
2165 flags |= DEQUEUE_MIGRATING;
2166
2167 dequeue_dl_entity(&p->dl, flags);
2168 if (!p->dl.dl_throttled && !dl_server(&p->dl))
2169 dequeue_pushable_dl_task(rq, p);
2170
2171 return true;
2172}
2173
2174/*
2175 * Yield task semantic for -deadline tasks is:
2176 *
2177 * get off from the CPU until our next instance, with
2178 * a new runtime. This is of little use now, since we
2179 * don't have a bandwidth reclaiming mechanism. Anyway,
2180 * bandwidth reclaiming is planned for the future, and
2181 * yield_task_dl will indicate that some spare budget
2182 * is available for other task instances to use it.
2183 */
2184static void yield_task_dl(struct rq *rq)
2185{
2186 /*
2187 * We make the task go to sleep until its current deadline by
2188 * forcing its runtime to zero. This way, update_curr_dl() stops
2189 * it and the bandwidth timer will wake it up and will give it
2190 * new scheduling parameters (thanks to dl_yielded=1).
2191 */
2192 rq->curr->dl.dl_yielded = 1;
2193
2194 update_rq_clock(rq);
2195 update_curr_dl(rq);
2196 /*
2197 * Tell update_rq_clock() that we've just updated,
2198 * so we don't do microscopic update in schedule()
2199 * and double the fastpath cost.
2200 */
2201 rq_clock_skip_update(rq);
2202}
2203
2204#ifdef CONFIG_SMP
2205
2206static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
2207 struct rq *rq)
2208{
2209 return (!rq->dl.dl_nr_running ||
2210 dl_time_before(p->dl.deadline,
2211 rq->dl.earliest_dl.curr));
2212}
2213
2214static int find_later_rq(struct task_struct *task);
2215
2216static int
2217select_task_rq_dl(struct task_struct *p, int cpu, int flags)
2218{
2219 struct task_struct *curr, *donor;
2220 bool select_rq;
2221 struct rq *rq;
2222
2223 if (!(flags & WF_TTWU))
2224 goto out;
2225
2226 rq = cpu_rq(cpu);
2227
2228 rcu_read_lock();
2229 curr = READ_ONCE(rq->curr); /* unlocked access */
2230 donor = READ_ONCE(rq->donor);
2231
2232 /*
2233 * If we are dealing with a -deadline task, we must
2234 * decide where to wake it up.
2235 * If it has a later deadline and the current task
2236 * on this rq can't move (provided the waking task
2237 * can!) we prefer to send it somewhere else. On the
2238 * other hand, if it has a shorter deadline, we
2239 * try to make it stay here, it might be important.
2240 */
2241 select_rq = unlikely(dl_task(donor)) &&
2242 (curr->nr_cpus_allowed < 2 ||
2243 !dl_entity_preempt(&p->dl, &donor->dl)) &&
2244 p->nr_cpus_allowed > 1;
2245
2246 /*
2247 * Take the capacity of the CPU into account to
2248 * ensure it fits the requirement of the task.
2249 */
2250 if (sched_asym_cpucap_active())
2251 select_rq |= !dl_task_fits_capacity(p, cpu);
2252
2253 if (select_rq) {
2254 int target = find_later_rq(p);
2255
2256 if (target != -1 &&
2257 dl_task_is_earliest_deadline(p, cpu_rq(target)))
2258 cpu = target;
2259 }
2260 rcu_read_unlock();
2261
2262out:
2263 return cpu;
2264}
2265
2266static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
2267{
2268 struct rq_flags rf;
2269 struct rq *rq;
2270
2271 if (READ_ONCE(p->__state) != TASK_WAKING)
2272 return;
2273
2274 rq = task_rq(p);
2275 /*
2276 * Since p->state == TASK_WAKING, set_task_cpu() has been called
2277 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
2278 * rq->lock is not... So, lock it
2279 */
2280 rq_lock(rq, &rf);
2281 if (p->dl.dl_non_contending) {
2282 update_rq_clock(rq);
2283 sub_running_bw(&p->dl, &rq->dl);
2284 p->dl.dl_non_contending = 0;
2285 /*
2286 * If the timer handler is currently running and the
2287 * timer cannot be canceled, inactive_task_timer()
2288 * will see that dl_not_contending is not set, and
2289 * will not touch the rq's active utilization,
2290 * so we are still safe.
2291 */
2292 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2293 put_task_struct(p);
2294 }
2295 sub_rq_bw(&p->dl, &rq->dl);
2296 rq_unlock(rq, &rf);
2297}
2298
2299static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
2300{
2301 /*
2302 * Current can't be migrated, useless to reschedule,
2303 * let's hope p can move out.
2304 */
2305 if (rq->curr->nr_cpus_allowed == 1 ||
2306 !cpudl_find(&rq->rd->cpudl, rq->donor, NULL))
2307 return;
2308
2309 /*
2310 * p is migratable, so let's not schedule it and
2311 * see if it is pushed or pulled somewhere else.
2312 */
2313 if (p->nr_cpus_allowed != 1 &&
2314 cpudl_find(&rq->rd->cpudl, p, NULL))
2315 return;
2316
2317 resched_curr(rq);
2318}
2319
2320static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
2321{
2322 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
2323 /*
2324 * This is OK, because current is on_cpu, which avoids it being
2325 * picked for load-balance and preemption/IRQs are still
2326 * disabled avoiding further scheduler activity on it and we've
2327 * not yet started the picking loop.
2328 */
2329 rq_unpin_lock(rq, rf);
2330 pull_dl_task(rq);
2331 rq_repin_lock(rq, rf);
2332 }
2333
2334 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
2335}
2336#endif /* CONFIG_SMP */
2337
2338/*
2339 * Only called when both the current and waking task are -deadline
2340 * tasks.
2341 */
2342static void wakeup_preempt_dl(struct rq *rq, struct task_struct *p,
2343 int flags)
2344{
2345 if (dl_entity_preempt(&p->dl, &rq->donor->dl)) {
2346 resched_curr(rq);
2347 return;
2348 }
2349
2350#ifdef CONFIG_SMP
2351 /*
2352 * In the unlikely case current and p have the same deadline
2353 * let us try to decide what's the best thing to do...
2354 */
2355 if ((p->dl.deadline == rq->donor->dl.deadline) &&
2356 !test_tsk_need_resched(rq->curr))
2357 check_preempt_equal_dl(rq, p);
2358#endif /* CONFIG_SMP */
2359}
2360
2361#ifdef CONFIG_SCHED_HRTICK
2362static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2363{
2364 hrtick_start(rq, dl_se->runtime);
2365}
2366#else /* !CONFIG_SCHED_HRTICK */
2367static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2368{
2369}
2370#endif
2371
2372static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
2373{
2374 struct sched_dl_entity *dl_se = &p->dl;
2375 struct dl_rq *dl_rq = &rq->dl;
2376
2377 p->se.exec_start = rq_clock_task(rq);
2378 if (on_dl_rq(&p->dl))
2379 update_stats_wait_end_dl(dl_rq, dl_se);
2380
2381 /* You can't push away the running task */
2382 dequeue_pushable_dl_task(rq, p);
2383
2384 if (!first)
2385 return;
2386
2387 if (rq->donor->sched_class != &dl_sched_class)
2388 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2389
2390 deadline_queue_push_tasks(rq);
2391
2392 if (hrtick_enabled_dl(rq))
2393 start_hrtick_dl(rq, &p->dl);
2394}
2395
2396static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
2397{
2398 struct rb_node *left = rb_first_cached(&dl_rq->root);
2399
2400 if (!left)
2401 return NULL;
2402
2403 return __node_2_dle(left);
2404}
2405
2406/*
2407 * __pick_next_task_dl - Helper to pick the next -deadline task to run.
2408 * @rq: The runqueue to pick the next task from.
2409 */
2410static struct task_struct *__pick_task_dl(struct rq *rq)
2411{
2412 struct sched_dl_entity *dl_se;
2413 struct dl_rq *dl_rq = &rq->dl;
2414 struct task_struct *p;
2415
2416again:
2417 if (!sched_dl_runnable(rq))
2418 return NULL;
2419
2420 dl_se = pick_next_dl_entity(dl_rq);
2421 WARN_ON_ONCE(!dl_se);
2422
2423 if (dl_server(dl_se)) {
2424 p = dl_se->server_pick_task(dl_se);
2425 if (!p) {
2426 if (dl_server_active(dl_se)) {
2427 dl_se->dl_yielded = 1;
2428 update_curr_dl_se(rq, dl_se, 0);
2429 }
2430 goto again;
2431 }
2432 rq->dl_server = dl_se;
2433 } else {
2434 p = dl_task_of(dl_se);
2435 }
2436
2437 return p;
2438}
2439
2440static struct task_struct *pick_task_dl(struct rq *rq)
2441{
2442 return __pick_task_dl(rq);
2443}
2444
2445static void put_prev_task_dl(struct rq *rq, struct task_struct *p, struct task_struct *next)
2446{
2447 struct sched_dl_entity *dl_se = &p->dl;
2448 struct dl_rq *dl_rq = &rq->dl;
2449
2450 if (on_dl_rq(&p->dl))
2451 update_stats_wait_start_dl(dl_rq, dl_se);
2452
2453 update_curr_dl(rq);
2454
2455 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2456 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2457 enqueue_pushable_dl_task(rq, p);
2458}
2459
2460/*
2461 * scheduler tick hitting a task of our scheduling class.
2462 *
2463 * NOTE: This function can be called remotely by the tick offload that
2464 * goes along full dynticks. Therefore no local assumption can be made
2465 * and everything must be accessed through the @rq and @curr passed in
2466 * parameters.
2467 */
2468static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2469{
2470 update_curr_dl(rq);
2471
2472 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2473 /*
2474 * Even when we have runtime, update_curr_dl() might have resulted in us
2475 * not being the leftmost task anymore. In that case NEED_RESCHED will
2476 * be set and schedule() will start a new hrtick for the next task.
2477 */
2478 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2479 is_leftmost(&p->dl, &rq->dl))
2480 start_hrtick_dl(rq, &p->dl);
2481}
2482
2483static void task_fork_dl(struct task_struct *p)
2484{
2485 /*
2486 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2487 * sched_fork()
2488 */
2489}
2490
2491#ifdef CONFIG_SMP
2492
2493/* Only try algorithms three times */
2494#define DL_MAX_TRIES 3
2495
2496/*
2497 * Return the earliest pushable rq's task, which is suitable to be executed
2498 * on the CPU, NULL otherwise:
2499 */
2500static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2501{
2502 struct task_struct *p = NULL;
2503 struct rb_node *next_node;
2504
2505 if (!has_pushable_dl_tasks(rq))
2506 return NULL;
2507
2508 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2509
2510next_node:
2511 if (next_node) {
2512 p = __node_2_pdl(next_node);
2513
2514 if (task_is_pushable(rq, p, cpu))
2515 return p;
2516
2517 next_node = rb_next(next_node);
2518 goto next_node;
2519 }
2520
2521 return NULL;
2522}
2523
2524static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2525
2526static int find_later_rq(struct task_struct *task)
2527{
2528 struct sched_domain *sd;
2529 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2530 int this_cpu = smp_processor_id();
2531 int cpu = task_cpu(task);
2532
2533 /* Make sure the mask is initialized first */
2534 if (unlikely(!later_mask))
2535 return -1;
2536
2537 if (task->nr_cpus_allowed == 1)
2538 return -1;
2539
2540 /*
2541 * We have to consider system topology and task affinity
2542 * first, then we can look for a suitable CPU.
2543 */
2544 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2545 return -1;
2546
2547 /*
2548 * If we are here, some targets have been found, including
2549 * the most suitable which is, among the runqueues where the
2550 * current tasks have later deadlines than the task's one, the
2551 * rq with the latest possible one.
2552 *
2553 * Now we check how well this matches with task's
2554 * affinity and system topology.
2555 *
2556 * The last CPU where the task run is our first
2557 * guess, since it is most likely cache-hot there.
2558 */
2559 if (cpumask_test_cpu(cpu, later_mask))
2560 return cpu;
2561 /*
2562 * Check if this_cpu is to be skipped (i.e., it is
2563 * not in the mask) or not.
2564 */
2565 if (!cpumask_test_cpu(this_cpu, later_mask))
2566 this_cpu = -1;
2567
2568 rcu_read_lock();
2569 for_each_domain(cpu, sd) {
2570 if (sd->flags & SD_WAKE_AFFINE) {
2571 int best_cpu;
2572
2573 /*
2574 * If possible, preempting this_cpu is
2575 * cheaper than migrating.
2576 */
2577 if (this_cpu != -1 &&
2578 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2579 rcu_read_unlock();
2580 return this_cpu;
2581 }
2582
2583 best_cpu = cpumask_any_and_distribute(later_mask,
2584 sched_domain_span(sd));
2585 /*
2586 * Last chance: if a CPU being in both later_mask
2587 * and current sd span is valid, that becomes our
2588 * choice. Of course, the latest possible CPU is
2589 * already under consideration through later_mask.
2590 */
2591 if (best_cpu < nr_cpu_ids) {
2592 rcu_read_unlock();
2593 return best_cpu;
2594 }
2595 }
2596 }
2597 rcu_read_unlock();
2598
2599 /*
2600 * At this point, all our guesses failed, we just return
2601 * 'something', and let the caller sort the things out.
2602 */
2603 if (this_cpu != -1)
2604 return this_cpu;
2605
2606 cpu = cpumask_any_distribute(later_mask);
2607 if (cpu < nr_cpu_ids)
2608 return cpu;
2609
2610 return -1;
2611}
2612
2613/* Locks the rq it finds */
2614static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2615{
2616 struct rq *later_rq = NULL;
2617 int tries;
2618 int cpu;
2619
2620 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2621 cpu = find_later_rq(task);
2622
2623 if ((cpu == -1) || (cpu == rq->cpu))
2624 break;
2625
2626 later_rq = cpu_rq(cpu);
2627
2628 if (!dl_task_is_earliest_deadline(task, later_rq)) {
2629 /*
2630 * Target rq has tasks of equal or earlier deadline,
2631 * retrying does not release any lock and is unlikely
2632 * to yield a different result.
2633 */
2634 later_rq = NULL;
2635 break;
2636 }
2637
2638 /* Retry if something changed. */
2639 if (double_lock_balance(rq, later_rq)) {
2640 if (unlikely(task_rq(task) != rq ||
2641 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2642 task_on_cpu(rq, task) ||
2643 !dl_task(task) ||
2644 is_migration_disabled(task) ||
2645 !task_on_rq_queued(task))) {
2646 double_unlock_balance(rq, later_rq);
2647 later_rq = NULL;
2648 break;
2649 }
2650 }
2651
2652 /*
2653 * If the rq we found has no -deadline task, or
2654 * its earliest one has a later deadline than our
2655 * task, the rq is a good one.
2656 */
2657 if (dl_task_is_earliest_deadline(task, later_rq))
2658 break;
2659
2660 /* Otherwise we try again. */
2661 double_unlock_balance(rq, later_rq);
2662 later_rq = NULL;
2663 }
2664
2665 return later_rq;
2666}
2667
2668static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2669{
2670 struct task_struct *p;
2671
2672 if (!has_pushable_dl_tasks(rq))
2673 return NULL;
2674
2675 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2676
2677 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2678 WARN_ON_ONCE(task_current(rq, p));
2679 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2680
2681 WARN_ON_ONCE(!task_on_rq_queued(p));
2682 WARN_ON_ONCE(!dl_task(p));
2683
2684 return p;
2685}
2686
2687/*
2688 * See if the non running -deadline tasks on this rq
2689 * can be sent to some other CPU where they can preempt
2690 * and start executing.
2691 */
2692static int push_dl_task(struct rq *rq)
2693{
2694 struct task_struct *next_task;
2695 struct rq *later_rq;
2696 int ret = 0;
2697
2698 next_task = pick_next_pushable_dl_task(rq);
2699 if (!next_task)
2700 return 0;
2701
2702retry:
2703 /*
2704 * If next_task preempts rq->curr, and rq->curr
2705 * can move away, it makes sense to just reschedule
2706 * without going further in pushing next_task.
2707 */
2708 if (dl_task(rq->donor) &&
2709 dl_time_before(next_task->dl.deadline, rq->donor->dl.deadline) &&
2710 rq->curr->nr_cpus_allowed > 1) {
2711 resched_curr(rq);
2712 return 0;
2713 }
2714
2715 if (is_migration_disabled(next_task))
2716 return 0;
2717
2718 if (WARN_ON(next_task == rq->curr))
2719 return 0;
2720
2721 /* We might release rq lock */
2722 get_task_struct(next_task);
2723
2724 /* Will lock the rq it'll find */
2725 later_rq = find_lock_later_rq(next_task, rq);
2726 if (!later_rq) {
2727 struct task_struct *task;
2728
2729 /*
2730 * We must check all this again, since
2731 * find_lock_later_rq releases rq->lock and it is
2732 * then possible that next_task has migrated.
2733 */
2734 task = pick_next_pushable_dl_task(rq);
2735 if (task == next_task) {
2736 /*
2737 * The task is still there. We don't try
2738 * again, some other CPU will pull it when ready.
2739 */
2740 goto out;
2741 }
2742
2743 if (!task)
2744 /* No more tasks */
2745 goto out;
2746
2747 put_task_struct(next_task);
2748 next_task = task;
2749 goto retry;
2750 }
2751
2752 move_queued_task_locked(rq, later_rq, next_task);
2753 ret = 1;
2754
2755 resched_curr(later_rq);
2756
2757 double_unlock_balance(rq, later_rq);
2758
2759out:
2760 put_task_struct(next_task);
2761
2762 return ret;
2763}
2764
2765static void push_dl_tasks(struct rq *rq)
2766{
2767 /* push_dl_task() will return true if it moved a -deadline task */
2768 while (push_dl_task(rq))
2769 ;
2770}
2771
2772static void pull_dl_task(struct rq *this_rq)
2773{
2774 int this_cpu = this_rq->cpu, cpu;
2775 struct task_struct *p, *push_task;
2776 bool resched = false;
2777 struct rq *src_rq;
2778 u64 dmin = LONG_MAX;
2779
2780 if (likely(!dl_overloaded(this_rq)))
2781 return;
2782
2783 /*
2784 * Match the barrier from dl_set_overloaded; this guarantees that if we
2785 * see overloaded we must also see the dlo_mask bit.
2786 */
2787 smp_rmb();
2788
2789 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2790 if (this_cpu == cpu)
2791 continue;
2792
2793 src_rq = cpu_rq(cpu);
2794
2795 /*
2796 * It looks racy, and it is! However, as in sched_rt.c,
2797 * we are fine with this.
2798 */
2799 if (this_rq->dl.dl_nr_running &&
2800 dl_time_before(this_rq->dl.earliest_dl.curr,
2801 src_rq->dl.earliest_dl.next))
2802 continue;
2803
2804 /* Might drop this_rq->lock */
2805 push_task = NULL;
2806 double_lock_balance(this_rq, src_rq);
2807
2808 /*
2809 * If there are no more pullable tasks on the
2810 * rq, we're done with it.
2811 */
2812 if (src_rq->dl.dl_nr_running <= 1)
2813 goto skip;
2814
2815 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2816
2817 /*
2818 * We found a task to be pulled if:
2819 * - it preempts our current (if there's one),
2820 * - it will preempt the last one we pulled (if any).
2821 */
2822 if (p && dl_time_before(p->dl.deadline, dmin) &&
2823 dl_task_is_earliest_deadline(p, this_rq)) {
2824 WARN_ON(p == src_rq->curr);
2825 WARN_ON(!task_on_rq_queued(p));
2826
2827 /*
2828 * Then we pull iff p has actually an earlier
2829 * deadline than the current task of its runqueue.
2830 */
2831 if (dl_time_before(p->dl.deadline,
2832 src_rq->donor->dl.deadline))
2833 goto skip;
2834
2835 if (is_migration_disabled(p)) {
2836 push_task = get_push_task(src_rq);
2837 } else {
2838 move_queued_task_locked(src_rq, this_rq, p);
2839 dmin = p->dl.deadline;
2840 resched = true;
2841 }
2842
2843 /* Is there any other task even earlier? */
2844 }
2845skip:
2846 double_unlock_balance(this_rq, src_rq);
2847
2848 if (push_task) {
2849 preempt_disable();
2850 raw_spin_rq_unlock(this_rq);
2851 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2852 push_task, &src_rq->push_work);
2853 preempt_enable();
2854 raw_spin_rq_lock(this_rq);
2855 }
2856 }
2857
2858 if (resched)
2859 resched_curr(this_rq);
2860}
2861
2862/*
2863 * Since the task is not running and a reschedule is not going to happen
2864 * anytime soon on its runqueue, we try pushing it away now.
2865 */
2866static void task_woken_dl(struct rq *rq, struct task_struct *p)
2867{
2868 if (!task_on_cpu(rq, p) &&
2869 !test_tsk_need_resched(rq->curr) &&
2870 p->nr_cpus_allowed > 1 &&
2871 dl_task(rq->donor) &&
2872 (rq->curr->nr_cpus_allowed < 2 ||
2873 !dl_entity_preempt(&p->dl, &rq->donor->dl))) {
2874 push_dl_tasks(rq);
2875 }
2876}
2877
2878static void set_cpus_allowed_dl(struct task_struct *p,
2879 struct affinity_context *ctx)
2880{
2881 struct root_domain *src_rd;
2882 struct rq *rq;
2883
2884 WARN_ON_ONCE(!dl_task(p));
2885
2886 rq = task_rq(p);
2887 src_rd = rq->rd;
2888 /*
2889 * Migrating a SCHED_DEADLINE task between exclusive
2890 * cpusets (different root_domains) entails a bandwidth
2891 * update. We already made space for us in the destination
2892 * domain (see cpuset_can_attach()).
2893 */
2894 if (!cpumask_intersects(src_rd->span, ctx->new_mask)) {
2895 struct dl_bw *src_dl_b;
2896
2897 src_dl_b = dl_bw_of(cpu_of(rq));
2898 /*
2899 * We now free resources of the root_domain we are migrating
2900 * off. In the worst case, sched_setattr() may temporary fail
2901 * until we complete the update.
2902 */
2903 raw_spin_lock(&src_dl_b->lock);
2904 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2905 raw_spin_unlock(&src_dl_b->lock);
2906 }
2907
2908 set_cpus_allowed_common(p, ctx);
2909}
2910
2911/* Assumes rq->lock is held */
2912static void rq_online_dl(struct rq *rq)
2913{
2914 if (rq->dl.overloaded)
2915 dl_set_overload(rq);
2916
2917 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2918 if (rq->dl.dl_nr_running > 0)
2919 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2920}
2921
2922/* Assumes rq->lock is held */
2923static void rq_offline_dl(struct rq *rq)
2924{
2925 if (rq->dl.overloaded)
2926 dl_clear_overload(rq);
2927
2928 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2929 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2930}
2931
2932void __init init_sched_dl_class(void)
2933{
2934 unsigned int i;
2935
2936 for_each_possible_cpu(i)
2937 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2938 GFP_KERNEL, cpu_to_node(i));
2939}
2940
2941void dl_add_task_root_domain(struct task_struct *p)
2942{
2943 struct rq_flags rf;
2944 struct rq *rq;
2945 struct dl_bw *dl_b;
2946
2947 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2948 if (!dl_task(p)) {
2949 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2950 return;
2951 }
2952
2953 rq = __task_rq_lock(p, &rf);
2954
2955 dl_b = &rq->rd->dl_bw;
2956 raw_spin_lock(&dl_b->lock);
2957
2958 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2959
2960 raw_spin_unlock(&dl_b->lock);
2961
2962 task_rq_unlock(rq, p, &rf);
2963}
2964
2965void dl_clear_root_domain(struct root_domain *rd)
2966{
2967 unsigned long flags;
2968
2969 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2970 rd->dl_bw.total_bw = 0;
2971 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2972}
2973
2974#endif /* CONFIG_SMP */
2975
2976static void switched_from_dl(struct rq *rq, struct task_struct *p)
2977{
2978 /*
2979 * task_non_contending() can start the "inactive timer" (if the 0-lag
2980 * time is in the future). If the task switches back to dl before
2981 * the "inactive timer" fires, it can continue to consume its current
2982 * runtime using its current deadline. If it stays outside of
2983 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2984 * will reset the task parameters.
2985 */
2986 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2987 task_non_contending(&p->dl);
2988
2989 /*
2990 * In case a task is setscheduled out from SCHED_DEADLINE we need to
2991 * keep track of that on its cpuset (for correct bandwidth tracking).
2992 */
2993 dec_dl_tasks_cs(p);
2994
2995 if (!task_on_rq_queued(p)) {
2996 /*
2997 * Inactive timer is armed. However, p is leaving DEADLINE and
2998 * might migrate away from this rq while continuing to run on
2999 * some other class. We need to remove its contribution from
3000 * this rq running_bw now, or sub_rq_bw (below) will complain.
3001 */
3002 if (p->dl.dl_non_contending)
3003 sub_running_bw(&p->dl, &rq->dl);
3004 sub_rq_bw(&p->dl, &rq->dl);
3005 }
3006
3007 /*
3008 * We cannot use inactive_task_timer() to invoke sub_running_bw()
3009 * at the 0-lag time, because the task could have been migrated
3010 * while SCHED_OTHER in the meanwhile.
3011 */
3012 if (p->dl.dl_non_contending)
3013 p->dl.dl_non_contending = 0;
3014
3015 /*
3016 * Since this might be the only -deadline task on the rq,
3017 * this is the right place to try to pull some other one
3018 * from an overloaded CPU, if any.
3019 */
3020 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
3021 return;
3022
3023 deadline_queue_pull_task(rq);
3024}
3025
3026/*
3027 * When switching to -deadline, we may overload the rq, then
3028 * we try to push someone off, if possible.
3029 */
3030static void switched_to_dl(struct rq *rq, struct task_struct *p)
3031{
3032 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
3033 put_task_struct(p);
3034
3035 /*
3036 * In case a task is setscheduled to SCHED_DEADLINE we need to keep
3037 * track of that on its cpuset (for correct bandwidth tracking).
3038 */
3039 inc_dl_tasks_cs(p);
3040
3041 /* If p is not queued we will update its parameters at next wakeup. */
3042 if (!task_on_rq_queued(p)) {
3043 add_rq_bw(&p->dl, &rq->dl);
3044
3045 return;
3046 }
3047
3048 if (rq->donor != p) {
3049#ifdef CONFIG_SMP
3050 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
3051 deadline_queue_push_tasks(rq);
3052#endif
3053 if (dl_task(rq->donor))
3054 wakeup_preempt_dl(rq, p, 0);
3055 else
3056 resched_curr(rq);
3057 } else {
3058 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
3059 }
3060}
3061
3062/*
3063 * If the scheduling parameters of a -deadline task changed,
3064 * a push or pull operation might be needed.
3065 */
3066static void prio_changed_dl(struct rq *rq, struct task_struct *p,
3067 int oldprio)
3068{
3069 if (!task_on_rq_queued(p))
3070 return;
3071
3072#ifdef CONFIG_SMP
3073 /*
3074 * This might be too much, but unfortunately
3075 * we don't have the old deadline value, and
3076 * we can't argue if the task is increasing
3077 * or lowering its prio, so...
3078 */
3079 if (!rq->dl.overloaded)
3080 deadline_queue_pull_task(rq);
3081
3082 if (task_current_donor(rq, p)) {
3083 /*
3084 * If we now have a earlier deadline task than p,
3085 * then reschedule, provided p is still on this
3086 * runqueue.
3087 */
3088 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
3089 resched_curr(rq);
3090 } else {
3091 /*
3092 * Current may not be deadline in case p was throttled but we
3093 * have just replenished it (e.g. rt_mutex_setprio()).
3094 *
3095 * Otherwise, if p was given an earlier deadline, reschedule.
3096 */
3097 if (!dl_task(rq->curr) ||
3098 dl_time_before(p->dl.deadline, rq->curr->dl.deadline))
3099 resched_curr(rq);
3100 }
3101#else
3102 /*
3103 * We don't know if p has a earlier or later deadline, so let's blindly
3104 * set a (maybe not needed) rescheduling point.
3105 */
3106 resched_curr(rq);
3107#endif
3108}
3109
3110#ifdef CONFIG_SCHED_CORE
3111static int task_is_throttled_dl(struct task_struct *p, int cpu)
3112{
3113 return p->dl.dl_throttled;
3114}
3115#endif
3116
3117DEFINE_SCHED_CLASS(dl) = {
3118
3119 .enqueue_task = enqueue_task_dl,
3120 .dequeue_task = dequeue_task_dl,
3121 .yield_task = yield_task_dl,
3122
3123 .wakeup_preempt = wakeup_preempt_dl,
3124
3125 .pick_task = pick_task_dl,
3126 .put_prev_task = put_prev_task_dl,
3127 .set_next_task = set_next_task_dl,
3128
3129#ifdef CONFIG_SMP
3130 .balance = balance_dl,
3131 .select_task_rq = select_task_rq_dl,
3132 .migrate_task_rq = migrate_task_rq_dl,
3133 .set_cpus_allowed = set_cpus_allowed_dl,
3134 .rq_online = rq_online_dl,
3135 .rq_offline = rq_offline_dl,
3136 .task_woken = task_woken_dl,
3137 .find_lock_rq = find_lock_later_rq,
3138#endif
3139
3140 .task_tick = task_tick_dl,
3141 .task_fork = task_fork_dl,
3142
3143 .prio_changed = prio_changed_dl,
3144 .switched_from = switched_from_dl,
3145 .switched_to = switched_to_dl,
3146
3147 .update_curr = update_curr_dl,
3148#ifdef CONFIG_SCHED_CORE
3149 .task_is_throttled = task_is_throttled_dl,
3150#endif
3151};
3152
3153/* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
3154static u64 dl_generation;
3155
3156int sched_dl_global_validate(void)
3157{
3158 u64 runtime = global_rt_runtime();
3159 u64 period = global_rt_period();
3160 u64 new_bw = to_ratio(period, runtime);
3161 u64 gen = ++dl_generation;
3162 struct dl_bw *dl_b;
3163 int cpu, cpus, ret = 0;
3164 unsigned long flags;
3165
3166 /*
3167 * Here we want to check the bandwidth not being set to some
3168 * value smaller than the currently allocated bandwidth in
3169 * any of the root_domains.
3170 */
3171 for_each_possible_cpu(cpu) {
3172 rcu_read_lock_sched();
3173
3174 if (dl_bw_visited(cpu, gen))
3175 goto next;
3176
3177 dl_b = dl_bw_of(cpu);
3178 cpus = dl_bw_cpus(cpu);
3179
3180 raw_spin_lock_irqsave(&dl_b->lock, flags);
3181 if (new_bw * cpus < dl_b->total_bw)
3182 ret = -EBUSY;
3183 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3184
3185next:
3186 rcu_read_unlock_sched();
3187
3188 if (ret)
3189 break;
3190 }
3191
3192 return ret;
3193}
3194
3195static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
3196{
3197 if (global_rt_runtime() == RUNTIME_INF) {
3198 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
3199 dl_rq->max_bw = dl_rq->extra_bw = 1 << BW_SHIFT;
3200 } else {
3201 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
3202 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
3203 dl_rq->max_bw = dl_rq->extra_bw =
3204 to_ratio(global_rt_period(), global_rt_runtime());
3205 }
3206}
3207
3208void sched_dl_do_global(void)
3209{
3210 u64 new_bw = -1;
3211 u64 gen = ++dl_generation;
3212 struct dl_bw *dl_b;
3213 int cpu;
3214 unsigned long flags;
3215
3216 if (global_rt_runtime() != RUNTIME_INF)
3217 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
3218
3219 for_each_possible_cpu(cpu) {
3220 rcu_read_lock_sched();
3221
3222 if (dl_bw_visited(cpu, gen)) {
3223 rcu_read_unlock_sched();
3224 continue;
3225 }
3226
3227 dl_b = dl_bw_of(cpu);
3228
3229 raw_spin_lock_irqsave(&dl_b->lock, flags);
3230 dl_b->bw = new_bw;
3231 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3232
3233 rcu_read_unlock_sched();
3234 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
3235 }
3236}
3237
3238/*
3239 * We must be sure that accepting a new task (or allowing changing the
3240 * parameters of an existing one) is consistent with the bandwidth
3241 * constraints. If yes, this function also accordingly updates the currently
3242 * allocated bandwidth to reflect the new situation.
3243 *
3244 * This function is called while holding p's rq->lock.
3245 */
3246int sched_dl_overflow(struct task_struct *p, int policy,
3247 const struct sched_attr *attr)
3248{
3249 u64 period = attr->sched_period ?: attr->sched_deadline;
3250 u64 runtime = attr->sched_runtime;
3251 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
3252 int cpus, err = -1, cpu = task_cpu(p);
3253 struct dl_bw *dl_b = dl_bw_of(cpu);
3254 unsigned long cap;
3255
3256 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3257 return 0;
3258
3259 /* !deadline task may carry old deadline bandwidth */
3260 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
3261 return 0;
3262
3263 /*
3264 * Either if a task, enters, leave, or stays -deadline but changes
3265 * its parameters, we may need to update accordingly the total
3266 * allocated bandwidth of the container.
3267 */
3268 raw_spin_lock(&dl_b->lock);
3269 cpus = dl_bw_cpus(cpu);
3270 cap = dl_bw_capacity(cpu);
3271
3272 if (dl_policy(policy) && !task_has_dl_policy(p) &&
3273 !__dl_overflow(dl_b, cap, 0, new_bw)) {
3274 if (hrtimer_active(&p->dl.inactive_timer))
3275 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3276 __dl_add(dl_b, new_bw, cpus);
3277 err = 0;
3278 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
3279 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
3280 /*
3281 * XXX this is slightly incorrect: when the task
3282 * utilization decreases, we should delay the total
3283 * utilization change until the task's 0-lag point.
3284 * But this would require to set the task's "inactive
3285 * timer" when the task is not inactive.
3286 */
3287 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3288 __dl_add(dl_b, new_bw, cpus);
3289 dl_change_utilization(p, new_bw);
3290 err = 0;
3291 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
3292 /*
3293 * Do not decrease the total deadline utilization here,
3294 * switched_from_dl() will take care to do it at the correct
3295 * (0-lag) time.
3296 */
3297 err = 0;
3298 }
3299 raw_spin_unlock(&dl_b->lock);
3300
3301 return err;
3302}
3303
3304/*
3305 * This function initializes the sched_dl_entity of a newly becoming
3306 * SCHED_DEADLINE task.
3307 *
3308 * Only the static values are considered here, the actual runtime and the
3309 * absolute deadline will be properly calculated when the task is enqueued
3310 * for the first time with its new policy.
3311 */
3312void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3313{
3314 struct sched_dl_entity *dl_se = &p->dl;
3315
3316 dl_se->dl_runtime = attr->sched_runtime;
3317 dl_se->dl_deadline = attr->sched_deadline;
3318 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3319 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
3320 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3321 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3322}
3323
3324void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3325{
3326 struct sched_dl_entity *dl_se = &p->dl;
3327
3328 attr->sched_priority = p->rt_priority;
3329 attr->sched_runtime = dl_se->dl_runtime;
3330 attr->sched_deadline = dl_se->dl_deadline;
3331 attr->sched_period = dl_se->dl_period;
3332 attr->sched_flags &= ~SCHED_DL_FLAGS;
3333 attr->sched_flags |= dl_se->flags;
3334}
3335
3336/*
3337 * This function validates the new parameters of a -deadline task.
3338 * We ask for the deadline not being zero, and greater or equal
3339 * than the runtime, as well as the period of being zero or
3340 * greater than deadline. Furthermore, we have to be sure that
3341 * user parameters are above the internal resolution of 1us (we
3342 * check sched_runtime only since it is always the smaller one) and
3343 * below 2^63 ns (we have to check both sched_deadline and
3344 * sched_period, as the latter can be zero).
3345 */
3346bool __checkparam_dl(const struct sched_attr *attr)
3347{
3348 u64 period, max, min;
3349
3350 /* special dl tasks don't actually use any parameter */
3351 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3352 return true;
3353
3354 /* deadline != 0 */
3355 if (attr->sched_deadline == 0)
3356 return false;
3357
3358 /*
3359 * Since we truncate DL_SCALE bits, make sure we're at least
3360 * that big.
3361 */
3362 if (attr->sched_runtime < (1ULL << DL_SCALE))
3363 return false;
3364
3365 /*
3366 * Since we use the MSB for wrap-around and sign issues, make
3367 * sure it's not set (mind that period can be equal to zero).
3368 */
3369 if (attr->sched_deadline & (1ULL << 63) ||
3370 attr->sched_period & (1ULL << 63))
3371 return false;
3372
3373 period = attr->sched_period;
3374 if (!period)
3375 period = attr->sched_deadline;
3376
3377 /* runtime <= deadline <= period (if period != 0) */
3378 if (period < attr->sched_deadline ||
3379 attr->sched_deadline < attr->sched_runtime)
3380 return false;
3381
3382 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
3383 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
3384
3385 if (period < min || period > max)
3386 return false;
3387
3388 return true;
3389}
3390
3391/*
3392 * This function clears the sched_dl_entity static params.
3393 */
3394static void __dl_clear_params(struct sched_dl_entity *dl_se)
3395{
3396 dl_se->dl_runtime = 0;
3397 dl_se->dl_deadline = 0;
3398 dl_se->dl_period = 0;
3399 dl_se->flags = 0;
3400 dl_se->dl_bw = 0;
3401 dl_se->dl_density = 0;
3402
3403 dl_se->dl_throttled = 0;
3404 dl_se->dl_yielded = 0;
3405 dl_se->dl_non_contending = 0;
3406 dl_se->dl_overrun = 0;
3407 dl_se->dl_server = 0;
3408
3409#ifdef CONFIG_RT_MUTEXES
3410 dl_se->pi_se = dl_se;
3411#endif
3412}
3413
3414void init_dl_entity(struct sched_dl_entity *dl_se)
3415{
3416 RB_CLEAR_NODE(&dl_se->rb_node);
3417 init_dl_task_timer(dl_se);
3418 init_dl_inactive_task_timer(dl_se);
3419 __dl_clear_params(dl_se);
3420}
3421
3422bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
3423{
3424 struct sched_dl_entity *dl_se = &p->dl;
3425
3426 if (dl_se->dl_runtime != attr->sched_runtime ||
3427 dl_se->dl_deadline != attr->sched_deadline ||
3428 dl_se->dl_period != attr->sched_period ||
3429 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3430 return true;
3431
3432 return false;
3433}
3434
3435#ifdef CONFIG_SMP
3436int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3437 const struct cpumask *trial)
3438{
3439 unsigned long flags, cap;
3440 struct dl_bw *cur_dl_b;
3441 int ret = 1;
3442
3443 rcu_read_lock_sched();
3444 cur_dl_b = dl_bw_of(cpumask_any(cur));
3445 cap = __dl_bw_capacity(trial);
3446 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3447 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3448 ret = 0;
3449 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3450 rcu_read_unlock_sched();
3451
3452 return ret;
3453}
3454
3455enum dl_bw_request {
3456 dl_bw_req_check_overflow = 0,
3457 dl_bw_req_alloc,
3458 dl_bw_req_free
3459};
3460
3461static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
3462{
3463 unsigned long flags;
3464 struct dl_bw *dl_b;
3465 bool overflow = 0;
3466
3467 rcu_read_lock_sched();
3468 dl_b = dl_bw_of(cpu);
3469 raw_spin_lock_irqsave(&dl_b->lock, flags);
3470
3471 if (req == dl_bw_req_free) {
3472 __dl_sub(dl_b, dl_bw, dl_bw_cpus(cpu));
3473 } else {
3474 unsigned long cap = dl_bw_capacity(cpu);
3475
3476 overflow = __dl_overflow(dl_b, cap, 0, dl_bw);
3477
3478 if (req == dl_bw_req_alloc && !overflow) {
3479 /*
3480 * We reserve space in the destination
3481 * root_domain, as we can't fail after this point.
3482 * We will free resources in the source root_domain
3483 * later on (see set_cpus_allowed_dl()).
3484 */
3485 __dl_add(dl_b, dl_bw, dl_bw_cpus(cpu));
3486 }
3487 }
3488
3489 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3490 rcu_read_unlock_sched();
3491
3492 return overflow ? -EBUSY : 0;
3493}
3494
3495int dl_bw_check_overflow(int cpu)
3496{
3497 return dl_bw_manage(dl_bw_req_check_overflow, cpu, 0);
3498}
3499
3500int dl_bw_alloc(int cpu, u64 dl_bw)
3501{
3502 return dl_bw_manage(dl_bw_req_alloc, cpu, dl_bw);
3503}
3504
3505void dl_bw_free(int cpu, u64 dl_bw)
3506{
3507 dl_bw_manage(dl_bw_req_free, cpu, dl_bw);
3508}
3509#endif
3510
3511#ifdef CONFIG_SCHED_DEBUG
3512void print_dl_stats(struct seq_file *m, int cpu)
3513{
3514 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3515}
3516#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#include "sched.h"
19#include "pelt.h"
20
21struct dl_bandwidth def_dl_bandwidth;
22
23static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
24{
25 return container_of(dl_se, struct task_struct, dl);
26}
27
28static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
29{
30 return container_of(dl_rq, struct rq, dl);
31}
32
33static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
34{
35 struct task_struct *p = dl_task_of(dl_se);
36 struct rq *rq = task_rq(p);
37
38 return &rq->dl;
39}
40
41static inline int on_dl_rq(struct sched_dl_entity *dl_se)
42{
43 return !RB_EMPTY_NODE(&dl_se->rb_node);
44}
45
46#ifdef CONFIG_SMP
47static inline struct dl_bw *dl_bw_of(int i)
48{
49 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
50 "sched RCU must be held");
51 return &cpu_rq(i)->rd->dl_bw;
52}
53
54static inline int dl_bw_cpus(int i)
55{
56 struct root_domain *rd = cpu_rq(i)->rd;
57 int cpus = 0;
58
59 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
60 "sched RCU must be held");
61 for_each_cpu_and(i, rd->span, cpu_active_mask)
62 cpus++;
63
64 return cpus;
65}
66#else
67static inline struct dl_bw *dl_bw_of(int i)
68{
69 return &cpu_rq(i)->dl.dl_bw;
70}
71
72static inline int dl_bw_cpus(int i)
73{
74 return 1;
75}
76#endif
77
78static inline
79void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
80{
81 u64 old = dl_rq->running_bw;
82
83 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
84 dl_rq->running_bw += dl_bw;
85 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
86 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
87 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
88 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
89}
90
91static inline
92void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
93{
94 u64 old = dl_rq->running_bw;
95
96 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
97 dl_rq->running_bw -= dl_bw;
98 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
99 if (dl_rq->running_bw > old)
100 dl_rq->running_bw = 0;
101 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
102 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
103}
104
105static inline
106void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
107{
108 u64 old = dl_rq->this_bw;
109
110 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
111 dl_rq->this_bw += dl_bw;
112 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
113}
114
115static inline
116void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
117{
118 u64 old = dl_rq->this_bw;
119
120 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
121 dl_rq->this_bw -= dl_bw;
122 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
123 if (dl_rq->this_bw > old)
124 dl_rq->this_bw = 0;
125 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
126}
127
128static inline
129void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
130{
131 if (!dl_entity_is_special(dl_se))
132 __add_rq_bw(dl_se->dl_bw, dl_rq);
133}
134
135static inline
136void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
137{
138 if (!dl_entity_is_special(dl_se))
139 __sub_rq_bw(dl_se->dl_bw, dl_rq);
140}
141
142static inline
143void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
144{
145 if (!dl_entity_is_special(dl_se))
146 __add_running_bw(dl_se->dl_bw, dl_rq);
147}
148
149static inline
150void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
151{
152 if (!dl_entity_is_special(dl_se))
153 __sub_running_bw(dl_se->dl_bw, dl_rq);
154}
155
156void dl_change_utilization(struct task_struct *p, u64 new_bw)
157{
158 struct rq *rq;
159
160 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
161
162 if (task_on_rq_queued(p))
163 return;
164
165 rq = task_rq(p);
166 if (p->dl.dl_non_contending) {
167 sub_running_bw(&p->dl, &rq->dl);
168 p->dl.dl_non_contending = 0;
169 /*
170 * If the timer handler is currently running and the
171 * timer cannot be cancelled, inactive_task_timer()
172 * will see that dl_not_contending is not set, and
173 * will not touch the rq's active utilization,
174 * so we are still safe.
175 */
176 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
177 put_task_struct(p);
178 }
179 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
180 __add_rq_bw(new_bw, &rq->dl);
181}
182
183/*
184 * The utilization of a task cannot be immediately removed from
185 * the rq active utilization (running_bw) when the task blocks.
186 * Instead, we have to wait for the so called "0-lag time".
187 *
188 * If a task blocks before the "0-lag time", a timer (the inactive
189 * timer) is armed, and running_bw is decreased when the timer
190 * fires.
191 *
192 * If the task wakes up again before the inactive timer fires,
193 * the timer is cancelled, whereas if the task wakes up after the
194 * inactive timer fired (and running_bw has been decreased) the
195 * task's utilization has to be added to running_bw again.
196 * A flag in the deadline scheduling entity (dl_non_contending)
197 * is used to avoid race conditions between the inactive timer handler
198 * and task wakeups.
199 *
200 * The following diagram shows how running_bw is updated. A task is
201 * "ACTIVE" when its utilization contributes to running_bw; an
202 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
203 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
204 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
205 * time already passed, which does not contribute to running_bw anymore.
206 * +------------------+
207 * wakeup | ACTIVE |
208 * +------------------>+ contending |
209 * | add_running_bw | |
210 * | +----+------+------+
211 * | | ^
212 * | dequeue | |
213 * +--------+-------+ | |
214 * | | t >= 0-lag | | wakeup
215 * | INACTIVE |<---------------+ |
216 * | | sub_running_bw | |
217 * +--------+-------+ | |
218 * ^ | |
219 * | t < 0-lag | |
220 * | | |
221 * | V |
222 * | +----+------+------+
223 * | sub_running_bw | ACTIVE |
224 * +-------------------+ |
225 * inactive timer | non contending |
226 * fired +------------------+
227 *
228 * The task_non_contending() function is invoked when a task
229 * blocks, and checks if the 0-lag time already passed or
230 * not (in the first case, it directly updates running_bw;
231 * in the second case, it arms the inactive timer).
232 *
233 * The task_contending() function is invoked when a task wakes
234 * up, and checks if the task is still in the "ACTIVE non contending"
235 * state or not (in the second case, it updates running_bw).
236 */
237static void task_non_contending(struct task_struct *p)
238{
239 struct sched_dl_entity *dl_se = &p->dl;
240 struct hrtimer *timer = &dl_se->inactive_timer;
241 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
242 struct rq *rq = rq_of_dl_rq(dl_rq);
243 s64 zerolag_time;
244
245 /*
246 * If this is a non-deadline task that has been boosted,
247 * do nothing
248 */
249 if (dl_se->dl_runtime == 0)
250 return;
251
252 if (dl_entity_is_special(dl_se))
253 return;
254
255 WARN_ON(dl_se->dl_non_contending);
256
257 zerolag_time = dl_se->deadline -
258 div64_long((dl_se->runtime * dl_se->dl_period),
259 dl_se->dl_runtime);
260
261 /*
262 * Using relative times instead of the absolute "0-lag time"
263 * allows to simplify the code
264 */
265 zerolag_time -= rq_clock(rq);
266
267 /*
268 * If the "0-lag time" already passed, decrease the active
269 * utilization now, instead of starting a timer
270 */
271 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
272 if (dl_task(p))
273 sub_running_bw(dl_se, dl_rq);
274 if (!dl_task(p) || p->state == TASK_DEAD) {
275 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
276
277 if (p->state == TASK_DEAD)
278 sub_rq_bw(&p->dl, &rq->dl);
279 raw_spin_lock(&dl_b->lock);
280 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
281 __dl_clear_params(p);
282 raw_spin_unlock(&dl_b->lock);
283 }
284
285 return;
286 }
287
288 dl_se->dl_non_contending = 1;
289 get_task_struct(p);
290 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
291}
292
293static void task_contending(struct sched_dl_entity *dl_se, int flags)
294{
295 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
296
297 /*
298 * If this is a non-deadline task that has been boosted,
299 * do nothing
300 */
301 if (dl_se->dl_runtime == 0)
302 return;
303
304 if (flags & ENQUEUE_MIGRATED)
305 add_rq_bw(dl_se, dl_rq);
306
307 if (dl_se->dl_non_contending) {
308 dl_se->dl_non_contending = 0;
309 /*
310 * If the timer handler is currently running and the
311 * timer cannot be cancelled, inactive_task_timer()
312 * will see that dl_not_contending is not set, and
313 * will not touch the rq's active utilization,
314 * so we are still safe.
315 */
316 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
317 put_task_struct(dl_task_of(dl_se));
318 } else {
319 /*
320 * Since "dl_non_contending" is not set, the
321 * task's utilization has already been removed from
322 * active utilization (either when the task blocked,
323 * when the "inactive timer" fired).
324 * So, add it back.
325 */
326 add_running_bw(dl_se, dl_rq);
327 }
328}
329
330static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
331{
332 struct sched_dl_entity *dl_se = &p->dl;
333
334 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
335}
336
337void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
338{
339 raw_spin_lock_init(&dl_b->dl_runtime_lock);
340 dl_b->dl_period = period;
341 dl_b->dl_runtime = runtime;
342}
343
344void init_dl_bw(struct dl_bw *dl_b)
345{
346 raw_spin_lock_init(&dl_b->lock);
347 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
348 if (global_rt_runtime() == RUNTIME_INF)
349 dl_b->bw = -1;
350 else
351 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
352 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
353 dl_b->total_bw = 0;
354}
355
356void init_dl_rq(struct dl_rq *dl_rq)
357{
358 dl_rq->root = RB_ROOT_CACHED;
359
360#ifdef CONFIG_SMP
361 /* zero means no -deadline tasks */
362 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
363
364 dl_rq->dl_nr_migratory = 0;
365 dl_rq->overloaded = 0;
366 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
367#else
368 init_dl_bw(&dl_rq->dl_bw);
369#endif
370
371 dl_rq->running_bw = 0;
372 dl_rq->this_bw = 0;
373 init_dl_rq_bw_ratio(dl_rq);
374}
375
376#ifdef CONFIG_SMP
377
378static inline int dl_overloaded(struct rq *rq)
379{
380 return atomic_read(&rq->rd->dlo_count);
381}
382
383static inline void dl_set_overload(struct rq *rq)
384{
385 if (!rq->online)
386 return;
387
388 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
389 /*
390 * Must be visible before the overload count is
391 * set (as in sched_rt.c).
392 *
393 * Matched by the barrier in pull_dl_task().
394 */
395 smp_wmb();
396 atomic_inc(&rq->rd->dlo_count);
397}
398
399static inline void dl_clear_overload(struct rq *rq)
400{
401 if (!rq->online)
402 return;
403
404 atomic_dec(&rq->rd->dlo_count);
405 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
406}
407
408static void update_dl_migration(struct dl_rq *dl_rq)
409{
410 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
411 if (!dl_rq->overloaded) {
412 dl_set_overload(rq_of_dl_rq(dl_rq));
413 dl_rq->overloaded = 1;
414 }
415 } else if (dl_rq->overloaded) {
416 dl_clear_overload(rq_of_dl_rq(dl_rq));
417 dl_rq->overloaded = 0;
418 }
419}
420
421static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
422{
423 struct task_struct *p = dl_task_of(dl_se);
424
425 if (p->nr_cpus_allowed > 1)
426 dl_rq->dl_nr_migratory++;
427
428 update_dl_migration(dl_rq);
429}
430
431static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
432{
433 struct task_struct *p = dl_task_of(dl_se);
434
435 if (p->nr_cpus_allowed > 1)
436 dl_rq->dl_nr_migratory--;
437
438 update_dl_migration(dl_rq);
439}
440
441/*
442 * The list of pushable -deadline task is not a plist, like in
443 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
444 */
445static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
446{
447 struct dl_rq *dl_rq = &rq->dl;
448 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node;
449 struct rb_node *parent = NULL;
450 struct task_struct *entry;
451 bool leftmost = true;
452
453 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
454
455 while (*link) {
456 parent = *link;
457 entry = rb_entry(parent, struct task_struct,
458 pushable_dl_tasks);
459 if (dl_entity_preempt(&p->dl, &entry->dl))
460 link = &parent->rb_left;
461 else {
462 link = &parent->rb_right;
463 leftmost = false;
464 }
465 }
466
467 if (leftmost)
468 dl_rq->earliest_dl.next = p->dl.deadline;
469
470 rb_link_node(&p->pushable_dl_tasks, parent, link);
471 rb_insert_color_cached(&p->pushable_dl_tasks,
472 &dl_rq->pushable_dl_tasks_root, leftmost);
473}
474
475static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
476{
477 struct dl_rq *dl_rq = &rq->dl;
478
479 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
480 return;
481
482 if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) {
483 struct rb_node *next_node;
484
485 next_node = rb_next(&p->pushable_dl_tasks);
486 if (next_node) {
487 dl_rq->earliest_dl.next = rb_entry(next_node,
488 struct task_struct, pushable_dl_tasks)->dl.deadline;
489 }
490 }
491
492 rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
493 RB_CLEAR_NODE(&p->pushable_dl_tasks);
494}
495
496static inline int has_pushable_dl_tasks(struct rq *rq)
497{
498 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
499}
500
501static int push_dl_task(struct rq *rq);
502
503static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
504{
505 return dl_task(prev);
506}
507
508static DEFINE_PER_CPU(struct callback_head, dl_push_head);
509static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
510
511static void push_dl_tasks(struct rq *);
512static void pull_dl_task(struct rq *);
513
514static inline void deadline_queue_push_tasks(struct rq *rq)
515{
516 if (!has_pushable_dl_tasks(rq))
517 return;
518
519 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
520}
521
522static inline void deadline_queue_pull_task(struct rq *rq)
523{
524 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
525}
526
527static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
528
529static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
530{
531 struct rq *later_rq = NULL;
532 struct dl_bw *dl_b;
533
534 later_rq = find_lock_later_rq(p, rq);
535 if (!later_rq) {
536 int cpu;
537
538 /*
539 * If we cannot preempt any rq, fall back to pick any
540 * online CPU:
541 */
542 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
543 if (cpu >= nr_cpu_ids) {
544 /*
545 * Failed to find any suitable CPU.
546 * The task will never come back!
547 */
548 BUG_ON(dl_bandwidth_enabled());
549
550 /*
551 * If admission control is disabled we
552 * try a little harder to let the task
553 * run.
554 */
555 cpu = cpumask_any(cpu_active_mask);
556 }
557 later_rq = cpu_rq(cpu);
558 double_lock_balance(rq, later_rq);
559 }
560
561 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
562 /*
563 * Inactive timer is armed (or callback is running, but
564 * waiting for us to release rq locks). In any case, when it
565 * will fire (or continue), it will see running_bw of this
566 * task migrated to later_rq (and correctly handle it).
567 */
568 sub_running_bw(&p->dl, &rq->dl);
569 sub_rq_bw(&p->dl, &rq->dl);
570
571 add_rq_bw(&p->dl, &later_rq->dl);
572 add_running_bw(&p->dl, &later_rq->dl);
573 } else {
574 sub_rq_bw(&p->dl, &rq->dl);
575 add_rq_bw(&p->dl, &later_rq->dl);
576 }
577
578 /*
579 * And we finally need to fixup root_domain(s) bandwidth accounting,
580 * since p is still hanging out in the old (now moved to default) root
581 * domain.
582 */
583 dl_b = &rq->rd->dl_bw;
584 raw_spin_lock(&dl_b->lock);
585 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
586 raw_spin_unlock(&dl_b->lock);
587
588 dl_b = &later_rq->rd->dl_bw;
589 raw_spin_lock(&dl_b->lock);
590 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
591 raw_spin_unlock(&dl_b->lock);
592
593 set_task_cpu(p, later_rq->cpu);
594 double_unlock_balance(later_rq, rq);
595
596 return later_rq;
597}
598
599#else
600
601static inline
602void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
603{
604}
605
606static inline
607void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
608{
609}
610
611static inline
612void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
613{
614}
615
616static inline
617void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
618{
619}
620
621static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
622{
623 return false;
624}
625
626static inline void pull_dl_task(struct rq *rq)
627{
628}
629
630static inline void deadline_queue_push_tasks(struct rq *rq)
631{
632}
633
634static inline void deadline_queue_pull_task(struct rq *rq)
635{
636}
637#endif /* CONFIG_SMP */
638
639static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
640static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
641static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
642
643/*
644 * We are being explicitly informed that a new instance is starting,
645 * and this means that:
646 * - the absolute deadline of the entity has to be placed at
647 * current time + relative deadline;
648 * - the runtime of the entity has to be set to the maximum value.
649 *
650 * The capability of specifying such event is useful whenever a -deadline
651 * entity wants to (try to!) synchronize its behaviour with the scheduler's
652 * one, and to (try to!) reconcile itself with its own scheduling
653 * parameters.
654 */
655static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
656{
657 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
658 struct rq *rq = rq_of_dl_rq(dl_rq);
659
660 WARN_ON(dl_se->dl_boosted);
661 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
662
663 /*
664 * We are racing with the deadline timer. So, do nothing because
665 * the deadline timer handler will take care of properly recharging
666 * the runtime and postponing the deadline
667 */
668 if (dl_se->dl_throttled)
669 return;
670
671 /*
672 * We use the regular wall clock time to set deadlines in the
673 * future; in fact, we must consider execution overheads (time
674 * spent on hardirq context, etc.).
675 */
676 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
677 dl_se->runtime = dl_se->dl_runtime;
678}
679
680/*
681 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
682 * possibility of a entity lasting more than what it declared, and thus
683 * exhausting its runtime.
684 *
685 * Here we are interested in making runtime overrun possible, but we do
686 * not want a entity which is misbehaving to affect the scheduling of all
687 * other entities.
688 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
689 * is used, in order to confine each entity within its own bandwidth.
690 *
691 * This function deals exactly with that, and ensures that when the runtime
692 * of a entity is replenished, its deadline is also postponed. That ensures
693 * the overrunning entity can't interfere with other entity in the system and
694 * can't make them miss their deadlines. Reasons why this kind of overruns
695 * could happen are, typically, a entity voluntarily trying to overcome its
696 * runtime, or it just underestimated it during sched_setattr().
697 */
698static void replenish_dl_entity(struct sched_dl_entity *dl_se,
699 struct sched_dl_entity *pi_se)
700{
701 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
702 struct rq *rq = rq_of_dl_rq(dl_rq);
703
704 BUG_ON(pi_se->dl_runtime <= 0);
705
706 /*
707 * This could be the case for a !-dl task that is boosted.
708 * Just go with full inherited parameters.
709 */
710 if (dl_se->dl_deadline == 0) {
711 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
712 dl_se->runtime = pi_se->dl_runtime;
713 }
714
715 if (dl_se->dl_yielded && dl_se->runtime > 0)
716 dl_se->runtime = 0;
717
718 /*
719 * We keep moving the deadline away until we get some
720 * available runtime for the entity. This ensures correct
721 * handling of situations where the runtime overrun is
722 * arbitrary large.
723 */
724 while (dl_se->runtime <= 0) {
725 dl_se->deadline += pi_se->dl_period;
726 dl_se->runtime += pi_se->dl_runtime;
727 }
728
729 /*
730 * At this point, the deadline really should be "in
731 * the future" with respect to rq->clock. If it's
732 * not, we are, for some reason, lagging too much!
733 * Anyway, after having warn userspace abut that,
734 * we still try to keep the things running by
735 * resetting the deadline and the budget of the
736 * entity.
737 */
738 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
739 printk_deferred_once("sched: DL replenish lagged too much\n");
740 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
741 dl_se->runtime = pi_se->dl_runtime;
742 }
743
744 if (dl_se->dl_yielded)
745 dl_se->dl_yielded = 0;
746 if (dl_se->dl_throttled)
747 dl_se->dl_throttled = 0;
748}
749
750/*
751 * Here we check if --at time t-- an entity (which is probably being
752 * [re]activated or, in general, enqueued) can use its remaining runtime
753 * and its current deadline _without_ exceeding the bandwidth it is
754 * assigned (function returns true if it can't). We are in fact applying
755 * one of the CBS rules: when a task wakes up, if the residual runtime
756 * over residual deadline fits within the allocated bandwidth, then we
757 * can keep the current (absolute) deadline and residual budget without
758 * disrupting the schedulability of the system. Otherwise, we should
759 * refill the runtime and set the deadline a period in the future,
760 * because keeping the current (absolute) deadline of the task would
761 * result in breaking guarantees promised to other tasks (refer to
762 * Documentation/scheduler/sched-deadline.rst for more information).
763 *
764 * This function returns true if:
765 *
766 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
767 *
768 * IOW we can't recycle current parameters.
769 *
770 * Notice that the bandwidth check is done against the deadline. For
771 * task with deadline equal to period this is the same of using
772 * dl_period instead of dl_deadline in the equation above.
773 */
774static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
775 struct sched_dl_entity *pi_se, u64 t)
776{
777 u64 left, right;
778
779 /*
780 * left and right are the two sides of the equation above,
781 * after a bit of shuffling to use multiplications instead
782 * of divisions.
783 *
784 * Note that none of the time values involved in the two
785 * multiplications are absolute: dl_deadline and dl_runtime
786 * are the relative deadline and the maximum runtime of each
787 * instance, runtime is the runtime left for the last instance
788 * and (deadline - t), since t is rq->clock, is the time left
789 * to the (absolute) deadline. Even if overflowing the u64 type
790 * is very unlikely to occur in both cases, here we scale down
791 * as we want to avoid that risk at all. Scaling down by 10
792 * means that we reduce granularity to 1us. We are fine with it,
793 * since this is only a true/false check and, anyway, thinking
794 * of anything below microseconds resolution is actually fiction
795 * (but still we want to give the user that illusion >;).
796 */
797 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
798 right = ((dl_se->deadline - t) >> DL_SCALE) *
799 (pi_se->dl_runtime >> DL_SCALE);
800
801 return dl_time_before(right, left);
802}
803
804/*
805 * Revised wakeup rule [1]: For self-suspending tasks, rather then
806 * re-initializing task's runtime and deadline, the revised wakeup
807 * rule adjusts the task's runtime to avoid the task to overrun its
808 * density.
809 *
810 * Reasoning: a task may overrun the density if:
811 * runtime / (deadline - t) > dl_runtime / dl_deadline
812 *
813 * Therefore, runtime can be adjusted to:
814 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
815 *
816 * In such way that runtime will be equal to the maximum density
817 * the task can use without breaking any rule.
818 *
819 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
820 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
821 */
822static void
823update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
824{
825 u64 laxity = dl_se->deadline - rq_clock(rq);
826
827 /*
828 * If the task has deadline < period, and the deadline is in the past,
829 * it should already be throttled before this check.
830 *
831 * See update_dl_entity() comments for further details.
832 */
833 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
834
835 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
836}
837
838/*
839 * Regarding the deadline, a task with implicit deadline has a relative
840 * deadline == relative period. A task with constrained deadline has a
841 * relative deadline <= relative period.
842 *
843 * We support constrained deadline tasks. However, there are some restrictions
844 * applied only for tasks which do not have an implicit deadline. See
845 * update_dl_entity() to know more about such restrictions.
846 *
847 * The dl_is_implicit() returns true if the task has an implicit deadline.
848 */
849static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
850{
851 return dl_se->dl_deadline == dl_se->dl_period;
852}
853
854/*
855 * When a deadline entity is placed in the runqueue, its runtime and deadline
856 * might need to be updated. This is done by a CBS wake up rule. There are two
857 * different rules: 1) the original CBS; and 2) the Revisited CBS.
858 *
859 * When the task is starting a new period, the Original CBS is used. In this
860 * case, the runtime is replenished and a new absolute deadline is set.
861 *
862 * When a task is queued before the begin of the next period, using the
863 * remaining runtime and deadline could make the entity to overflow, see
864 * dl_entity_overflow() to find more about runtime overflow. When such case
865 * is detected, the runtime and deadline need to be updated.
866 *
867 * If the task has an implicit deadline, i.e., deadline == period, the Original
868 * CBS is applied. the runtime is replenished and a new absolute deadline is
869 * set, as in the previous cases.
870 *
871 * However, the Original CBS does not work properly for tasks with
872 * deadline < period, which are said to have a constrained deadline. By
873 * applying the Original CBS, a constrained deadline task would be able to run
874 * runtime/deadline in a period. With deadline < period, the task would
875 * overrun the runtime/period allowed bandwidth, breaking the admission test.
876 *
877 * In order to prevent this misbehave, the Revisited CBS is used for
878 * constrained deadline tasks when a runtime overflow is detected. In the
879 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
880 * the remaining runtime of the task is reduced to avoid runtime overflow.
881 * Please refer to the comments update_dl_revised_wakeup() function to find
882 * more about the Revised CBS rule.
883 */
884static void update_dl_entity(struct sched_dl_entity *dl_se,
885 struct sched_dl_entity *pi_se)
886{
887 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
888 struct rq *rq = rq_of_dl_rq(dl_rq);
889
890 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
891 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
892
893 if (unlikely(!dl_is_implicit(dl_se) &&
894 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
895 !dl_se->dl_boosted)){
896 update_dl_revised_wakeup(dl_se, rq);
897 return;
898 }
899
900 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
901 dl_se->runtime = pi_se->dl_runtime;
902 }
903}
904
905static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
906{
907 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
908}
909
910/*
911 * If the entity depleted all its runtime, and if we want it to sleep
912 * while waiting for some new execution time to become available, we
913 * set the bandwidth replenishment timer to the replenishment instant
914 * and try to activate it.
915 *
916 * Notice that it is important for the caller to know if the timer
917 * actually started or not (i.e., the replenishment instant is in
918 * the future or in the past).
919 */
920static int start_dl_timer(struct task_struct *p)
921{
922 struct sched_dl_entity *dl_se = &p->dl;
923 struct hrtimer *timer = &dl_se->dl_timer;
924 struct rq *rq = task_rq(p);
925 ktime_t now, act;
926 s64 delta;
927
928 lockdep_assert_held(&rq->lock);
929
930 /*
931 * We want the timer to fire at the deadline, but considering
932 * that it is actually coming from rq->clock and not from
933 * hrtimer's time base reading.
934 */
935 act = ns_to_ktime(dl_next_period(dl_se));
936 now = hrtimer_cb_get_time(timer);
937 delta = ktime_to_ns(now) - rq_clock(rq);
938 act = ktime_add_ns(act, delta);
939
940 /*
941 * If the expiry time already passed, e.g., because the value
942 * chosen as the deadline is too small, don't even try to
943 * start the timer in the past!
944 */
945 if (ktime_us_delta(act, now) < 0)
946 return 0;
947
948 /*
949 * !enqueued will guarantee another callback; even if one is already in
950 * progress. This ensures a balanced {get,put}_task_struct().
951 *
952 * The race against __run_timer() clearing the enqueued state is
953 * harmless because we're holding task_rq()->lock, therefore the timer
954 * expiring after we've done the check will wait on its task_rq_lock()
955 * and observe our state.
956 */
957 if (!hrtimer_is_queued(timer)) {
958 get_task_struct(p);
959 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
960 }
961
962 return 1;
963}
964
965/*
966 * This is the bandwidth enforcement timer callback. If here, we know
967 * a task is not on its dl_rq, since the fact that the timer was running
968 * means the task is throttled and needs a runtime replenishment.
969 *
970 * However, what we actually do depends on the fact the task is active,
971 * (it is on its rq) or has been removed from there by a call to
972 * dequeue_task_dl(). In the former case we must issue the runtime
973 * replenishment and add the task back to the dl_rq; in the latter, we just
974 * do nothing but clearing dl_throttled, so that runtime and deadline
975 * updating (and the queueing back to dl_rq) will be done by the
976 * next call to enqueue_task_dl().
977 */
978static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
979{
980 struct sched_dl_entity *dl_se = container_of(timer,
981 struct sched_dl_entity,
982 dl_timer);
983 struct task_struct *p = dl_task_of(dl_se);
984 struct rq_flags rf;
985 struct rq *rq;
986
987 rq = task_rq_lock(p, &rf);
988
989 /*
990 * The task might have changed its scheduling policy to something
991 * different than SCHED_DEADLINE (through switched_from_dl()).
992 */
993 if (!dl_task(p))
994 goto unlock;
995
996 /*
997 * The task might have been boosted by someone else and might be in the
998 * boosting/deboosting path, its not throttled.
999 */
1000 if (dl_se->dl_boosted)
1001 goto unlock;
1002
1003 /*
1004 * Spurious timer due to start_dl_timer() race; or we already received
1005 * a replenishment from rt_mutex_setprio().
1006 */
1007 if (!dl_se->dl_throttled)
1008 goto unlock;
1009
1010 sched_clock_tick();
1011 update_rq_clock(rq);
1012
1013 /*
1014 * If the throttle happened during sched-out; like:
1015 *
1016 * schedule()
1017 * deactivate_task()
1018 * dequeue_task_dl()
1019 * update_curr_dl()
1020 * start_dl_timer()
1021 * __dequeue_task_dl()
1022 * prev->on_rq = 0;
1023 *
1024 * We can be both throttled and !queued. Replenish the counter
1025 * but do not enqueue -- wait for our wakeup to do that.
1026 */
1027 if (!task_on_rq_queued(p)) {
1028 replenish_dl_entity(dl_se, dl_se);
1029 goto unlock;
1030 }
1031
1032#ifdef CONFIG_SMP
1033 if (unlikely(!rq->online)) {
1034 /*
1035 * If the runqueue is no longer available, migrate the
1036 * task elsewhere. This necessarily changes rq.
1037 */
1038 lockdep_unpin_lock(&rq->lock, rf.cookie);
1039 rq = dl_task_offline_migration(rq, p);
1040 rf.cookie = lockdep_pin_lock(&rq->lock);
1041 update_rq_clock(rq);
1042
1043 /*
1044 * Now that the task has been migrated to the new RQ and we
1045 * have that locked, proceed as normal and enqueue the task
1046 * there.
1047 */
1048 }
1049#endif
1050
1051 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1052 if (dl_task(rq->curr))
1053 check_preempt_curr_dl(rq, p, 0);
1054 else
1055 resched_curr(rq);
1056
1057#ifdef CONFIG_SMP
1058 /*
1059 * Queueing this task back might have overloaded rq, check if we need
1060 * to kick someone away.
1061 */
1062 if (has_pushable_dl_tasks(rq)) {
1063 /*
1064 * Nothing relies on rq->lock after this, so its safe to drop
1065 * rq->lock.
1066 */
1067 rq_unpin_lock(rq, &rf);
1068 push_dl_task(rq);
1069 rq_repin_lock(rq, &rf);
1070 }
1071#endif
1072
1073unlock:
1074 task_rq_unlock(rq, p, &rf);
1075
1076 /*
1077 * This can free the task_struct, including this hrtimer, do not touch
1078 * anything related to that after this.
1079 */
1080 put_task_struct(p);
1081
1082 return HRTIMER_NORESTART;
1083}
1084
1085void init_dl_task_timer(struct sched_dl_entity *dl_se)
1086{
1087 struct hrtimer *timer = &dl_se->dl_timer;
1088
1089 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1090 timer->function = dl_task_timer;
1091}
1092
1093/*
1094 * During the activation, CBS checks if it can reuse the current task's
1095 * runtime and period. If the deadline of the task is in the past, CBS
1096 * cannot use the runtime, and so it replenishes the task. This rule
1097 * works fine for implicit deadline tasks (deadline == period), and the
1098 * CBS was designed for implicit deadline tasks. However, a task with
1099 * constrained deadline (deadine < period) might be awakened after the
1100 * deadline, but before the next period. In this case, replenishing the
1101 * task would allow it to run for runtime / deadline. As in this case
1102 * deadline < period, CBS enables a task to run for more than the
1103 * runtime / period. In a very loaded system, this can cause a domino
1104 * effect, making other tasks miss their deadlines.
1105 *
1106 * To avoid this problem, in the activation of a constrained deadline
1107 * task after the deadline but before the next period, throttle the
1108 * task and set the replenishing timer to the begin of the next period,
1109 * unless it is boosted.
1110 */
1111static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1112{
1113 struct task_struct *p = dl_task_of(dl_se);
1114 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1115
1116 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1117 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1118 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1119 return;
1120 dl_se->dl_throttled = 1;
1121 if (dl_se->runtime > 0)
1122 dl_se->runtime = 0;
1123 }
1124}
1125
1126static
1127int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1128{
1129 return (dl_se->runtime <= 0);
1130}
1131
1132extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1133
1134/*
1135 * This function implements the GRUB accounting rule:
1136 * according to the GRUB reclaiming algorithm, the runtime is
1137 * not decreased as "dq = -dt", but as
1138 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1139 * where u is the utilization of the task, Umax is the maximum reclaimable
1140 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1141 * as the difference between the "total runqueue utilization" and the
1142 * runqueue active utilization, and Uextra is the (per runqueue) extra
1143 * reclaimable utilization.
1144 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1145 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1146 * BW_SHIFT.
1147 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1148 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1149 * Since delta is a 64 bit variable, to have an overflow its value
1150 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1151 * So, overflow is not an issue here.
1152 */
1153static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1154{
1155 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1156 u64 u_act;
1157 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1158
1159 /*
1160 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1161 * we compare u_inact + rq->dl.extra_bw with
1162 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1163 * u_inact + rq->dl.extra_bw can be larger than
1164 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1165 * leading to wrong results)
1166 */
1167 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1168 u_act = u_act_min;
1169 else
1170 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1171
1172 return (delta * u_act) >> BW_SHIFT;
1173}
1174
1175/*
1176 * Update the current task's runtime statistics (provided it is still
1177 * a -deadline task and has not been removed from the dl_rq).
1178 */
1179static void update_curr_dl(struct rq *rq)
1180{
1181 struct task_struct *curr = rq->curr;
1182 struct sched_dl_entity *dl_se = &curr->dl;
1183 u64 delta_exec, scaled_delta_exec;
1184 int cpu = cpu_of(rq);
1185 u64 now;
1186
1187 if (!dl_task(curr) || !on_dl_rq(dl_se))
1188 return;
1189
1190 /*
1191 * Consumed budget is computed considering the time as
1192 * observed by schedulable tasks (excluding time spent
1193 * in hardirq context, etc.). Deadlines are instead
1194 * computed using hard walltime. This seems to be the more
1195 * natural solution, but the full ramifications of this
1196 * approach need further study.
1197 */
1198 now = rq_clock_task(rq);
1199 delta_exec = now - curr->se.exec_start;
1200 if (unlikely((s64)delta_exec <= 0)) {
1201 if (unlikely(dl_se->dl_yielded))
1202 goto throttle;
1203 return;
1204 }
1205
1206 schedstat_set(curr->se.statistics.exec_max,
1207 max(curr->se.statistics.exec_max, delta_exec));
1208
1209 curr->se.sum_exec_runtime += delta_exec;
1210 account_group_exec_runtime(curr, delta_exec);
1211
1212 curr->se.exec_start = now;
1213 cgroup_account_cputime(curr, delta_exec);
1214
1215 if (dl_entity_is_special(dl_se))
1216 return;
1217
1218 /*
1219 * For tasks that participate in GRUB, we implement GRUB-PA: the
1220 * spare reclaimed bandwidth is used to clock down frequency.
1221 *
1222 * For the others, we still need to scale reservation parameters
1223 * according to current frequency and CPU maximum capacity.
1224 */
1225 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1226 scaled_delta_exec = grub_reclaim(delta_exec,
1227 rq,
1228 &curr->dl);
1229 } else {
1230 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1231 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1232
1233 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1234 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1235 }
1236
1237 dl_se->runtime -= scaled_delta_exec;
1238
1239throttle:
1240 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1241 dl_se->dl_throttled = 1;
1242
1243 /* If requested, inform the user about runtime overruns. */
1244 if (dl_runtime_exceeded(dl_se) &&
1245 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1246 dl_se->dl_overrun = 1;
1247
1248 __dequeue_task_dl(rq, curr, 0);
1249 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
1250 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1251
1252 if (!is_leftmost(curr, &rq->dl))
1253 resched_curr(rq);
1254 }
1255
1256 /*
1257 * Because -- for now -- we share the rt bandwidth, we need to
1258 * account our runtime there too, otherwise actual rt tasks
1259 * would be able to exceed the shared quota.
1260 *
1261 * Account to the root rt group for now.
1262 *
1263 * The solution we're working towards is having the RT groups scheduled
1264 * using deadline servers -- however there's a few nasties to figure
1265 * out before that can happen.
1266 */
1267 if (rt_bandwidth_enabled()) {
1268 struct rt_rq *rt_rq = &rq->rt;
1269
1270 raw_spin_lock(&rt_rq->rt_runtime_lock);
1271 /*
1272 * We'll let actual RT tasks worry about the overflow here, we
1273 * have our own CBS to keep us inline; only account when RT
1274 * bandwidth is relevant.
1275 */
1276 if (sched_rt_bandwidth_account(rt_rq))
1277 rt_rq->rt_time += delta_exec;
1278 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1279 }
1280}
1281
1282static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1283{
1284 struct sched_dl_entity *dl_se = container_of(timer,
1285 struct sched_dl_entity,
1286 inactive_timer);
1287 struct task_struct *p = dl_task_of(dl_se);
1288 struct rq_flags rf;
1289 struct rq *rq;
1290
1291 rq = task_rq_lock(p, &rf);
1292
1293 sched_clock_tick();
1294 update_rq_clock(rq);
1295
1296 if (!dl_task(p) || p->state == TASK_DEAD) {
1297 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1298
1299 if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1300 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1301 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1302 dl_se->dl_non_contending = 0;
1303 }
1304
1305 raw_spin_lock(&dl_b->lock);
1306 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1307 raw_spin_unlock(&dl_b->lock);
1308 __dl_clear_params(p);
1309
1310 goto unlock;
1311 }
1312 if (dl_se->dl_non_contending == 0)
1313 goto unlock;
1314
1315 sub_running_bw(dl_se, &rq->dl);
1316 dl_se->dl_non_contending = 0;
1317unlock:
1318 task_rq_unlock(rq, p, &rf);
1319 put_task_struct(p);
1320
1321 return HRTIMER_NORESTART;
1322}
1323
1324void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1325{
1326 struct hrtimer *timer = &dl_se->inactive_timer;
1327
1328 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1329 timer->function = inactive_task_timer;
1330}
1331
1332#ifdef CONFIG_SMP
1333
1334static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1335{
1336 struct rq *rq = rq_of_dl_rq(dl_rq);
1337
1338 if (dl_rq->earliest_dl.curr == 0 ||
1339 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1340 dl_rq->earliest_dl.curr = deadline;
1341 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1342 }
1343}
1344
1345static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1346{
1347 struct rq *rq = rq_of_dl_rq(dl_rq);
1348
1349 /*
1350 * Since we may have removed our earliest (and/or next earliest)
1351 * task we must recompute them.
1352 */
1353 if (!dl_rq->dl_nr_running) {
1354 dl_rq->earliest_dl.curr = 0;
1355 dl_rq->earliest_dl.next = 0;
1356 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1357 } else {
1358 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1359 struct sched_dl_entity *entry;
1360
1361 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1362 dl_rq->earliest_dl.curr = entry->deadline;
1363 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1364 }
1365}
1366
1367#else
1368
1369static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1370static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1371
1372#endif /* CONFIG_SMP */
1373
1374static inline
1375void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1376{
1377 int prio = dl_task_of(dl_se)->prio;
1378 u64 deadline = dl_se->deadline;
1379
1380 WARN_ON(!dl_prio(prio));
1381 dl_rq->dl_nr_running++;
1382 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1383
1384 inc_dl_deadline(dl_rq, deadline);
1385 inc_dl_migration(dl_se, dl_rq);
1386}
1387
1388static inline
1389void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1390{
1391 int prio = dl_task_of(dl_se)->prio;
1392
1393 WARN_ON(!dl_prio(prio));
1394 WARN_ON(!dl_rq->dl_nr_running);
1395 dl_rq->dl_nr_running--;
1396 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1397
1398 dec_dl_deadline(dl_rq, dl_se->deadline);
1399 dec_dl_migration(dl_se, dl_rq);
1400}
1401
1402static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1403{
1404 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1405 struct rb_node **link = &dl_rq->root.rb_root.rb_node;
1406 struct rb_node *parent = NULL;
1407 struct sched_dl_entity *entry;
1408 int leftmost = 1;
1409
1410 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1411
1412 while (*link) {
1413 parent = *link;
1414 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
1415 if (dl_time_before(dl_se->deadline, entry->deadline))
1416 link = &parent->rb_left;
1417 else {
1418 link = &parent->rb_right;
1419 leftmost = 0;
1420 }
1421 }
1422
1423 rb_link_node(&dl_se->rb_node, parent, link);
1424 rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost);
1425
1426 inc_dl_tasks(dl_se, dl_rq);
1427}
1428
1429static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1430{
1431 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1432
1433 if (RB_EMPTY_NODE(&dl_se->rb_node))
1434 return;
1435
1436 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1437 RB_CLEAR_NODE(&dl_se->rb_node);
1438
1439 dec_dl_tasks(dl_se, dl_rq);
1440}
1441
1442static void
1443enqueue_dl_entity(struct sched_dl_entity *dl_se,
1444 struct sched_dl_entity *pi_se, int flags)
1445{
1446 BUG_ON(on_dl_rq(dl_se));
1447
1448 /*
1449 * If this is a wakeup or a new instance, the scheduling
1450 * parameters of the task might need updating. Otherwise,
1451 * we want a replenishment of its runtime.
1452 */
1453 if (flags & ENQUEUE_WAKEUP) {
1454 task_contending(dl_se, flags);
1455 update_dl_entity(dl_se, pi_se);
1456 } else if (flags & ENQUEUE_REPLENISH) {
1457 replenish_dl_entity(dl_se, pi_se);
1458 } else if ((flags & ENQUEUE_RESTORE) &&
1459 dl_time_before(dl_se->deadline,
1460 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1461 setup_new_dl_entity(dl_se);
1462 }
1463
1464 __enqueue_dl_entity(dl_se);
1465}
1466
1467static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1468{
1469 __dequeue_dl_entity(dl_se);
1470}
1471
1472static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1473{
1474 struct task_struct *pi_task = rt_mutex_get_top_task(p);
1475 struct sched_dl_entity *pi_se = &p->dl;
1476
1477 /*
1478 * Use the scheduling parameters of the top pi-waiter task if:
1479 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1480 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1481 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1482 * boosted due to a SCHED_DEADLINE pi-waiter).
1483 * Otherwise we keep our runtime and deadline.
1484 */
1485 if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) {
1486 pi_se = &pi_task->dl;
1487 } else if (!dl_prio(p->normal_prio)) {
1488 /*
1489 * Special case in which we have a !SCHED_DEADLINE task
1490 * that is going to be deboosted, but exceeds its
1491 * runtime while doing so. No point in replenishing
1492 * it, as it's going to return back to its original
1493 * scheduling class after this.
1494 */
1495 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
1496 return;
1497 }
1498
1499 /*
1500 * Check if a constrained deadline task was activated
1501 * after the deadline but before the next period.
1502 * If that is the case, the task will be throttled and
1503 * the replenishment timer will be set to the next period.
1504 */
1505 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1506 dl_check_constrained_dl(&p->dl);
1507
1508 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1509 add_rq_bw(&p->dl, &rq->dl);
1510 add_running_bw(&p->dl, &rq->dl);
1511 }
1512
1513 /*
1514 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1515 * its budget it needs a replenishment and, since it now is on
1516 * its rq, the bandwidth timer callback (which clearly has not
1517 * run yet) will take care of this.
1518 * However, the active utilization does not depend on the fact
1519 * that the task is on the runqueue or not (but depends on the
1520 * task's state - in GRUB parlance, "inactive" vs "active contending").
1521 * In other words, even if a task is throttled its utilization must
1522 * be counted in the active utilization; hence, we need to call
1523 * add_running_bw().
1524 */
1525 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1526 if (flags & ENQUEUE_WAKEUP)
1527 task_contending(&p->dl, flags);
1528
1529 return;
1530 }
1531
1532 enqueue_dl_entity(&p->dl, pi_se, flags);
1533
1534 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1535 enqueue_pushable_dl_task(rq, p);
1536}
1537
1538static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1539{
1540 dequeue_dl_entity(&p->dl);
1541 dequeue_pushable_dl_task(rq, p);
1542}
1543
1544static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1545{
1546 update_curr_dl(rq);
1547 __dequeue_task_dl(rq, p, flags);
1548
1549 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1550 sub_running_bw(&p->dl, &rq->dl);
1551 sub_rq_bw(&p->dl, &rq->dl);
1552 }
1553
1554 /*
1555 * This check allows to start the inactive timer (or to immediately
1556 * decrease the active utilization, if needed) in two cases:
1557 * when the task blocks and when it is terminating
1558 * (p->state == TASK_DEAD). We can handle the two cases in the same
1559 * way, because from GRUB's point of view the same thing is happening
1560 * (the task moves from "active contending" to "active non contending"
1561 * or "inactive")
1562 */
1563 if (flags & DEQUEUE_SLEEP)
1564 task_non_contending(p);
1565}
1566
1567/*
1568 * Yield task semantic for -deadline tasks is:
1569 *
1570 * get off from the CPU until our next instance, with
1571 * a new runtime. This is of little use now, since we
1572 * don't have a bandwidth reclaiming mechanism. Anyway,
1573 * bandwidth reclaiming is planned for the future, and
1574 * yield_task_dl will indicate that some spare budget
1575 * is available for other task instances to use it.
1576 */
1577static void yield_task_dl(struct rq *rq)
1578{
1579 /*
1580 * We make the task go to sleep until its current deadline by
1581 * forcing its runtime to zero. This way, update_curr_dl() stops
1582 * it and the bandwidth timer will wake it up and will give it
1583 * new scheduling parameters (thanks to dl_yielded=1).
1584 */
1585 rq->curr->dl.dl_yielded = 1;
1586
1587 update_rq_clock(rq);
1588 update_curr_dl(rq);
1589 /*
1590 * Tell update_rq_clock() that we've just updated,
1591 * so we don't do microscopic update in schedule()
1592 * and double the fastpath cost.
1593 */
1594 rq_clock_skip_update(rq);
1595}
1596
1597#ifdef CONFIG_SMP
1598
1599static int find_later_rq(struct task_struct *task);
1600
1601static int
1602select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1603{
1604 struct task_struct *curr;
1605 struct rq *rq;
1606
1607 if (sd_flag != SD_BALANCE_WAKE)
1608 goto out;
1609
1610 rq = cpu_rq(cpu);
1611
1612 rcu_read_lock();
1613 curr = READ_ONCE(rq->curr); /* unlocked access */
1614
1615 /*
1616 * If we are dealing with a -deadline task, we must
1617 * decide where to wake it up.
1618 * If it has a later deadline and the current task
1619 * on this rq can't move (provided the waking task
1620 * can!) we prefer to send it somewhere else. On the
1621 * other hand, if it has a shorter deadline, we
1622 * try to make it stay here, it might be important.
1623 */
1624 if (unlikely(dl_task(curr)) &&
1625 (curr->nr_cpus_allowed < 2 ||
1626 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1627 (p->nr_cpus_allowed > 1)) {
1628 int target = find_later_rq(p);
1629
1630 if (target != -1 &&
1631 (dl_time_before(p->dl.deadline,
1632 cpu_rq(target)->dl.earliest_dl.curr) ||
1633 (cpu_rq(target)->dl.dl_nr_running == 0)))
1634 cpu = target;
1635 }
1636 rcu_read_unlock();
1637
1638out:
1639 return cpu;
1640}
1641
1642static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1643{
1644 struct rq *rq;
1645
1646 if (p->state != TASK_WAKING)
1647 return;
1648
1649 rq = task_rq(p);
1650 /*
1651 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1652 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1653 * rq->lock is not... So, lock it
1654 */
1655 raw_spin_lock(&rq->lock);
1656 if (p->dl.dl_non_contending) {
1657 sub_running_bw(&p->dl, &rq->dl);
1658 p->dl.dl_non_contending = 0;
1659 /*
1660 * If the timer handler is currently running and the
1661 * timer cannot be cancelled, inactive_task_timer()
1662 * will see that dl_not_contending is not set, and
1663 * will not touch the rq's active utilization,
1664 * so we are still safe.
1665 */
1666 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1667 put_task_struct(p);
1668 }
1669 sub_rq_bw(&p->dl, &rq->dl);
1670 raw_spin_unlock(&rq->lock);
1671}
1672
1673static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1674{
1675 /*
1676 * Current can't be migrated, useless to reschedule,
1677 * let's hope p can move out.
1678 */
1679 if (rq->curr->nr_cpus_allowed == 1 ||
1680 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1681 return;
1682
1683 /*
1684 * p is migratable, so let's not schedule it and
1685 * see if it is pushed or pulled somewhere else.
1686 */
1687 if (p->nr_cpus_allowed != 1 &&
1688 cpudl_find(&rq->rd->cpudl, p, NULL))
1689 return;
1690
1691 resched_curr(rq);
1692}
1693
1694static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1695{
1696 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1697 /*
1698 * This is OK, because current is on_cpu, which avoids it being
1699 * picked for load-balance and preemption/IRQs are still
1700 * disabled avoiding further scheduler activity on it and we've
1701 * not yet started the picking loop.
1702 */
1703 rq_unpin_lock(rq, rf);
1704 pull_dl_task(rq);
1705 rq_repin_lock(rq, rf);
1706 }
1707
1708 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1709}
1710#endif /* CONFIG_SMP */
1711
1712/*
1713 * Only called when both the current and waking task are -deadline
1714 * tasks.
1715 */
1716static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1717 int flags)
1718{
1719 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1720 resched_curr(rq);
1721 return;
1722 }
1723
1724#ifdef CONFIG_SMP
1725 /*
1726 * In the unlikely case current and p have the same deadline
1727 * let us try to decide what's the best thing to do...
1728 */
1729 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1730 !test_tsk_need_resched(rq->curr))
1731 check_preempt_equal_dl(rq, p);
1732#endif /* CONFIG_SMP */
1733}
1734
1735#ifdef CONFIG_SCHED_HRTICK
1736static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1737{
1738 hrtick_start(rq, p->dl.runtime);
1739}
1740#else /* !CONFIG_SCHED_HRTICK */
1741static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1742{
1743}
1744#endif
1745
1746static void set_next_task_dl(struct rq *rq, struct task_struct *p)
1747{
1748 p->se.exec_start = rq_clock_task(rq);
1749
1750 /* You can't push away the running task */
1751 dequeue_pushable_dl_task(rq, p);
1752
1753 if (hrtick_enabled(rq))
1754 start_hrtick_dl(rq, p);
1755
1756 if (rq->curr->sched_class != &dl_sched_class)
1757 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1758
1759 deadline_queue_push_tasks(rq);
1760}
1761
1762static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1763 struct dl_rq *dl_rq)
1764{
1765 struct rb_node *left = rb_first_cached(&dl_rq->root);
1766
1767 if (!left)
1768 return NULL;
1769
1770 return rb_entry(left, struct sched_dl_entity, rb_node);
1771}
1772
1773static struct task_struct *
1774pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1775{
1776 struct sched_dl_entity *dl_se;
1777 struct dl_rq *dl_rq = &rq->dl;
1778 struct task_struct *p;
1779
1780 WARN_ON_ONCE(prev || rf);
1781
1782 if (!sched_dl_runnable(rq))
1783 return NULL;
1784
1785 dl_se = pick_next_dl_entity(rq, dl_rq);
1786 BUG_ON(!dl_se);
1787 p = dl_task_of(dl_se);
1788 set_next_task_dl(rq, p);
1789 return p;
1790}
1791
1792static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1793{
1794 update_curr_dl(rq);
1795
1796 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1797 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1798 enqueue_pushable_dl_task(rq, p);
1799}
1800
1801/*
1802 * scheduler tick hitting a task of our scheduling class.
1803 *
1804 * NOTE: This function can be called remotely by the tick offload that
1805 * goes along full dynticks. Therefore no local assumption can be made
1806 * and everything must be accessed through the @rq and @curr passed in
1807 * parameters.
1808 */
1809static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1810{
1811 update_curr_dl(rq);
1812
1813 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1814 /*
1815 * Even when we have runtime, update_curr_dl() might have resulted in us
1816 * not being the leftmost task anymore. In that case NEED_RESCHED will
1817 * be set and schedule() will start a new hrtick for the next task.
1818 */
1819 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1820 is_leftmost(p, &rq->dl))
1821 start_hrtick_dl(rq, p);
1822}
1823
1824static void task_fork_dl(struct task_struct *p)
1825{
1826 /*
1827 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1828 * sched_fork()
1829 */
1830}
1831
1832#ifdef CONFIG_SMP
1833
1834/* Only try algorithms three times */
1835#define DL_MAX_TRIES 3
1836
1837static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1838{
1839 if (!task_running(rq, p) &&
1840 cpumask_test_cpu(cpu, p->cpus_ptr))
1841 return 1;
1842 return 0;
1843}
1844
1845/*
1846 * Return the earliest pushable rq's task, which is suitable to be executed
1847 * on the CPU, NULL otherwise:
1848 */
1849static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1850{
1851 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
1852 struct task_struct *p = NULL;
1853
1854 if (!has_pushable_dl_tasks(rq))
1855 return NULL;
1856
1857next_node:
1858 if (next_node) {
1859 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1860
1861 if (pick_dl_task(rq, p, cpu))
1862 return p;
1863
1864 next_node = rb_next(next_node);
1865 goto next_node;
1866 }
1867
1868 return NULL;
1869}
1870
1871static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1872
1873static int find_later_rq(struct task_struct *task)
1874{
1875 struct sched_domain *sd;
1876 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1877 int this_cpu = smp_processor_id();
1878 int cpu = task_cpu(task);
1879
1880 /* Make sure the mask is initialized first */
1881 if (unlikely(!later_mask))
1882 return -1;
1883
1884 if (task->nr_cpus_allowed == 1)
1885 return -1;
1886
1887 /*
1888 * We have to consider system topology and task affinity
1889 * first, then we can look for a suitable CPU.
1890 */
1891 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1892 return -1;
1893
1894 /*
1895 * If we are here, some targets have been found, including
1896 * the most suitable which is, among the runqueues where the
1897 * current tasks have later deadlines than the task's one, the
1898 * rq with the latest possible one.
1899 *
1900 * Now we check how well this matches with task's
1901 * affinity and system topology.
1902 *
1903 * The last CPU where the task run is our first
1904 * guess, since it is most likely cache-hot there.
1905 */
1906 if (cpumask_test_cpu(cpu, later_mask))
1907 return cpu;
1908 /*
1909 * Check if this_cpu is to be skipped (i.e., it is
1910 * not in the mask) or not.
1911 */
1912 if (!cpumask_test_cpu(this_cpu, later_mask))
1913 this_cpu = -1;
1914
1915 rcu_read_lock();
1916 for_each_domain(cpu, sd) {
1917 if (sd->flags & SD_WAKE_AFFINE) {
1918 int best_cpu;
1919
1920 /*
1921 * If possible, preempting this_cpu is
1922 * cheaper than migrating.
1923 */
1924 if (this_cpu != -1 &&
1925 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1926 rcu_read_unlock();
1927 return this_cpu;
1928 }
1929
1930 best_cpu = cpumask_first_and(later_mask,
1931 sched_domain_span(sd));
1932 /*
1933 * Last chance: if a CPU being in both later_mask
1934 * and current sd span is valid, that becomes our
1935 * choice. Of course, the latest possible CPU is
1936 * already under consideration through later_mask.
1937 */
1938 if (best_cpu < nr_cpu_ids) {
1939 rcu_read_unlock();
1940 return best_cpu;
1941 }
1942 }
1943 }
1944 rcu_read_unlock();
1945
1946 /*
1947 * At this point, all our guesses failed, we just return
1948 * 'something', and let the caller sort the things out.
1949 */
1950 if (this_cpu != -1)
1951 return this_cpu;
1952
1953 cpu = cpumask_any(later_mask);
1954 if (cpu < nr_cpu_ids)
1955 return cpu;
1956
1957 return -1;
1958}
1959
1960/* Locks the rq it finds */
1961static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1962{
1963 struct rq *later_rq = NULL;
1964 int tries;
1965 int cpu;
1966
1967 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1968 cpu = find_later_rq(task);
1969
1970 if ((cpu == -1) || (cpu == rq->cpu))
1971 break;
1972
1973 later_rq = cpu_rq(cpu);
1974
1975 if (later_rq->dl.dl_nr_running &&
1976 !dl_time_before(task->dl.deadline,
1977 later_rq->dl.earliest_dl.curr)) {
1978 /*
1979 * Target rq has tasks of equal or earlier deadline,
1980 * retrying does not release any lock and is unlikely
1981 * to yield a different result.
1982 */
1983 later_rq = NULL;
1984 break;
1985 }
1986
1987 /* Retry if something changed. */
1988 if (double_lock_balance(rq, later_rq)) {
1989 if (unlikely(task_rq(task) != rq ||
1990 !cpumask_test_cpu(later_rq->cpu, task->cpus_ptr) ||
1991 task_running(rq, task) ||
1992 !dl_task(task) ||
1993 !task_on_rq_queued(task))) {
1994 double_unlock_balance(rq, later_rq);
1995 later_rq = NULL;
1996 break;
1997 }
1998 }
1999
2000 /*
2001 * If the rq we found has no -deadline task, or
2002 * its earliest one has a later deadline than our
2003 * task, the rq is a good one.
2004 */
2005 if (!later_rq->dl.dl_nr_running ||
2006 dl_time_before(task->dl.deadline,
2007 later_rq->dl.earliest_dl.curr))
2008 break;
2009
2010 /* Otherwise we try again. */
2011 double_unlock_balance(rq, later_rq);
2012 later_rq = NULL;
2013 }
2014
2015 return later_rq;
2016}
2017
2018static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2019{
2020 struct task_struct *p;
2021
2022 if (!has_pushable_dl_tasks(rq))
2023 return NULL;
2024
2025 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
2026 struct task_struct, pushable_dl_tasks);
2027
2028 BUG_ON(rq->cpu != task_cpu(p));
2029 BUG_ON(task_current(rq, p));
2030 BUG_ON(p->nr_cpus_allowed <= 1);
2031
2032 BUG_ON(!task_on_rq_queued(p));
2033 BUG_ON(!dl_task(p));
2034
2035 return p;
2036}
2037
2038/*
2039 * See if the non running -deadline tasks on this rq
2040 * can be sent to some other CPU where they can preempt
2041 * and start executing.
2042 */
2043static int push_dl_task(struct rq *rq)
2044{
2045 struct task_struct *next_task;
2046 struct rq *later_rq;
2047 int ret = 0;
2048
2049 if (!rq->dl.overloaded)
2050 return 0;
2051
2052 next_task = pick_next_pushable_dl_task(rq);
2053 if (!next_task)
2054 return 0;
2055
2056retry:
2057 if (WARN_ON(next_task == rq->curr))
2058 return 0;
2059
2060 /*
2061 * If next_task preempts rq->curr, and rq->curr
2062 * can move away, it makes sense to just reschedule
2063 * without going further in pushing next_task.
2064 */
2065 if (dl_task(rq->curr) &&
2066 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2067 rq->curr->nr_cpus_allowed > 1) {
2068 resched_curr(rq);
2069 return 0;
2070 }
2071
2072 /* We might release rq lock */
2073 get_task_struct(next_task);
2074
2075 /* Will lock the rq it'll find */
2076 later_rq = find_lock_later_rq(next_task, rq);
2077 if (!later_rq) {
2078 struct task_struct *task;
2079
2080 /*
2081 * We must check all this again, since
2082 * find_lock_later_rq releases rq->lock and it is
2083 * then possible that next_task has migrated.
2084 */
2085 task = pick_next_pushable_dl_task(rq);
2086 if (task == next_task) {
2087 /*
2088 * The task is still there. We don't try
2089 * again, some other CPU will pull it when ready.
2090 */
2091 goto out;
2092 }
2093
2094 if (!task)
2095 /* No more tasks */
2096 goto out;
2097
2098 put_task_struct(next_task);
2099 next_task = task;
2100 goto retry;
2101 }
2102
2103 deactivate_task(rq, next_task, 0);
2104 set_task_cpu(next_task, later_rq->cpu);
2105
2106 /*
2107 * Update the later_rq clock here, because the clock is used
2108 * by the cpufreq_update_util() inside __add_running_bw().
2109 */
2110 update_rq_clock(later_rq);
2111 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
2112 ret = 1;
2113
2114 resched_curr(later_rq);
2115
2116 double_unlock_balance(rq, later_rq);
2117
2118out:
2119 put_task_struct(next_task);
2120
2121 return ret;
2122}
2123
2124static void push_dl_tasks(struct rq *rq)
2125{
2126 /* push_dl_task() will return true if it moved a -deadline task */
2127 while (push_dl_task(rq))
2128 ;
2129}
2130
2131static void pull_dl_task(struct rq *this_rq)
2132{
2133 int this_cpu = this_rq->cpu, cpu;
2134 struct task_struct *p;
2135 bool resched = false;
2136 struct rq *src_rq;
2137 u64 dmin = LONG_MAX;
2138
2139 if (likely(!dl_overloaded(this_rq)))
2140 return;
2141
2142 /*
2143 * Match the barrier from dl_set_overloaded; this guarantees that if we
2144 * see overloaded we must also see the dlo_mask bit.
2145 */
2146 smp_rmb();
2147
2148 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2149 if (this_cpu == cpu)
2150 continue;
2151
2152 src_rq = cpu_rq(cpu);
2153
2154 /*
2155 * It looks racy, abd it is! However, as in sched_rt.c,
2156 * we are fine with this.
2157 */
2158 if (this_rq->dl.dl_nr_running &&
2159 dl_time_before(this_rq->dl.earliest_dl.curr,
2160 src_rq->dl.earliest_dl.next))
2161 continue;
2162
2163 /* Might drop this_rq->lock */
2164 double_lock_balance(this_rq, src_rq);
2165
2166 /*
2167 * If there are no more pullable tasks on the
2168 * rq, we're done with it.
2169 */
2170 if (src_rq->dl.dl_nr_running <= 1)
2171 goto skip;
2172
2173 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2174
2175 /*
2176 * We found a task to be pulled if:
2177 * - it preempts our current (if there's one),
2178 * - it will preempt the last one we pulled (if any).
2179 */
2180 if (p && dl_time_before(p->dl.deadline, dmin) &&
2181 (!this_rq->dl.dl_nr_running ||
2182 dl_time_before(p->dl.deadline,
2183 this_rq->dl.earliest_dl.curr))) {
2184 WARN_ON(p == src_rq->curr);
2185 WARN_ON(!task_on_rq_queued(p));
2186
2187 /*
2188 * Then we pull iff p has actually an earlier
2189 * deadline than the current task of its runqueue.
2190 */
2191 if (dl_time_before(p->dl.deadline,
2192 src_rq->curr->dl.deadline))
2193 goto skip;
2194
2195 resched = true;
2196
2197 deactivate_task(src_rq, p, 0);
2198 set_task_cpu(p, this_cpu);
2199 activate_task(this_rq, p, 0);
2200 dmin = p->dl.deadline;
2201
2202 /* Is there any other task even earlier? */
2203 }
2204skip:
2205 double_unlock_balance(this_rq, src_rq);
2206 }
2207
2208 if (resched)
2209 resched_curr(this_rq);
2210}
2211
2212/*
2213 * Since the task is not running and a reschedule is not going to happen
2214 * anytime soon on its runqueue, we try pushing it away now.
2215 */
2216static void task_woken_dl(struct rq *rq, struct task_struct *p)
2217{
2218 if (!task_running(rq, p) &&
2219 !test_tsk_need_resched(rq->curr) &&
2220 p->nr_cpus_allowed > 1 &&
2221 dl_task(rq->curr) &&
2222 (rq->curr->nr_cpus_allowed < 2 ||
2223 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2224 push_dl_tasks(rq);
2225 }
2226}
2227
2228static void set_cpus_allowed_dl(struct task_struct *p,
2229 const struct cpumask *new_mask)
2230{
2231 struct root_domain *src_rd;
2232 struct rq *rq;
2233
2234 BUG_ON(!dl_task(p));
2235
2236 rq = task_rq(p);
2237 src_rd = rq->rd;
2238 /*
2239 * Migrating a SCHED_DEADLINE task between exclusive
2240 * cpusets (different root_domains) entails a bandwidth
2241 * update. We already made space for us in the destination
2242 * domain (see cpuset_can_attach()).
2243 */
2244 if (!cpumask_intersects(src_rd->span, new_mask)) {
2245 struct dl_bw *src_dl_b;
2246
2247 src_dl_b = dl_bw_of(cpu_of(rq));
2248 /*
2249 * We now free resources of the root_domain we are migrating
2250 * off. In the worst case, sched_setattr() may temporary fail
2251 * until we complete the update.
2252 */
2253 raw_spin_lock(&src_dl_b->lock);
2254 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2255 raw_spin_unlock(&src_dl_b->lock);
2256 }
2257
2258 set_cpus_allowed_common(p, new_mask);
2259}
2260
2261/* Assumes rq->lock is held */
2262static void rq_online_dl(struct rq *rq)
2263{
2264 if (rq->dl.overloaded)
2265 dl_set_overload(rq);
2266
2267 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2268 if (rq->dl.dl_nr_running > 0)
2269 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2270}
2271
2272/* Assumes rq->lock is held */
2273static void rq_offline_dl(struct rq *rq)
2274{
2275 if (rq->dl.overloaded)
2276 dl_clear_overload(rq);
2277
2278 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2279 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2280}
2281
2282void __init init_sched_dl_class(void)
2283{
2284 unsigned int i;
2285
2286 for_each_possible_cpu(i)
2287 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2288 GFP_KERNEL, cpu_to_node(i));
2289}
2290
2291void dl_add_task_root_domain(struct task_struct *p)
2292{
2293 struct rq_flags rf;
2294 struct rq *rq;
2295 struct dl_bw *dl_b;
2296
2297 rq = task_rq_lock(p, &rf);
2298 if (!dl_task(p))
2299 goto unlock;
2300
2301 dl_b = &rq->rd->dl_bw;
2302 raw_spin_lock(&dl_b->lock);
2303
2304 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2305
2306 raw_spin_unlock(&dl_b->lock);
2307
2308unlock:
2309 task_rq_unlock(rq, p, &rf);
2310}
2311
2312void dl_clear_root_domain(struct root_domain *rd)
2313{
2314 unsigned long flags;
2315
2316 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2317 rd->dl_bw.total_bw = 0;
2318 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2319}
2320
2321#endif /* CONFIG_SMP */
2322
2323static void switched_from_dl(struct rq *rq, struct task_struct *p)
2324{
2325 /*
2326 * task_non_contending() can start the "inactive timer" (if the 0-lag
2327 * time is in the future). If the task switches back to dl before
2328 * the "inactive timer" fires, it can continue to consume its current
2329 * runtime using its current deadline. If it stays outside of
2330 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2331 * will reset the task parameters.
2332 */
2333 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2334 task_non_contending(p);
2335
2336 if (!task_on_rq_queued(p)) {
2337 /*
2338 * Inactive timer is armed. However, p is leaving DEADLINE and
2339 * might migrate away from this rq while continuing to run on
2340 * some other class. We need to remove its contribution from
2341 * this rq running_bw now, or sub_rq_bw (below) will complain.
2342 */
2343 if (p->dl.dl_non_contending)
2344 sub_running_bw(&p->dl, &rq->dl);
2345 sub_rq_bw(&p->dl, &rq->dl);
2346 }
2347
2348 /*
2349 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2350 * at the 0-lag time, because the task could have been migrated
2351 * while SCHED_OTHER in the meanwhile.
2352 */
2353 if (p->dl.dl_non_contending)
2354 p->dl.dl_non_contending = 0;
2355
2356 /*
2357 * Since this might be the only -deadline task on the rq,
2358 * this is the right place to try to pull some other one
2359 * from an overloaded CPU, if any.
2360 */
2361 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2362 return;
2363
2364 deadline_queue_pull_task(rq);
2365}
2366
2367/*
2368 * When switching to -deadline, we may overload the rq, then
2369 * we try to push someone off, if possible.
2370 */
2371static void switched_to_dl(struct rq *rq, struct task_struct *p)
2372{
2373 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2374 put_task_struct(p);
2375
2376 /* If p is not queued we will update its parameters at next wakeup. */
2377 if (!task_on_rq_queued(p)) {
2378 add_rq_bw(&p->dl, &rq->dl);
2379
2380 return;
2381 }
2382
2383 if (rq->curr != p) {
2384#ifdef CONFIG_SMP
2385 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2386 deadline_queue_push_tasks(rq);
2387#endif
2388 if (dl_task(rq->curr))
2389 check_preempt_curr_dl(rq, p, 0);
2390 else
2391 resched_curr(rq);
2392 }
2393}
2394
2395/*
2396 * If the scheduling parameters of a -deadline task changed,
2397 * a push or pull operation might be needed.
2398 */
2399static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2400 int oldprio)
2401{
2402 if (task_on_rq_queued(p) || rq->curr == p) {
2403#ifdef CONFIG_SMP
2404 /*
2405 * This might be too much, but unfortunately
2406 * we don't have the old deadline value, and
2407 * we can't argue if the task is increasing
2408 * or lowering its prio, so...
2409 */
2410 if (!rq->dl.overloaded)
2411 deadline_queue_pull_task(rq);
2412
2413 /*
2414 * If we now have a earlier deadline task than p,
2415 * then reschedule, provided p is still on this
2416 * runqueue.
2417 */
2418 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2419 resched_curr(rq);
2420#else
2421 /*
2422 * Again, we don't know if p has a earlier
2423 * or later deadline, so let's blindly set a
2424 * (maybe not needed) rescheduling point.
2425 */
2426 resched_curr(rq);
2427#endif /* CONFIG_SMP */
2428 }
2429}
2430
2431const struct sched_class dl_sched_class = {
2432 .next = &rt_sched_class,
2433 .enqueue_task = enqueue_task_dl,
2434 .dequeue_task = dequeue_task_dl,
2435 .yield_task = yield_task_dl,
2436
2437 .check_preempt_curr = check_preempt_curr_dl,
2438
2439 .pick_next_task = pick_next_task_dl,
2440 .put_prev_task = put_prev_task_dl,
2441 .set_next_task = set_next_task_dl,
2442
2443#ifdef CONFIG_SMP
2444 .balance = balance_dl,
2445 .select_task_rq = select_task_rq_dl,
2446 .migrate_task_rq = migrate_task_rq_dl,
2447 .set_cpus_allowed = set_cpus_allowed_dl,
2448 .rq_online = rq_online_dl,
2449 .rq_offline = rq_offline_dl,
2450 .task_woken = task_woken_dl,
2451#endif
2452
2453 .task_tick = task_tick_dl,
2454 .task_fork = task_fork_dl,
2455
2456 .prio_changed = prio_changed_dl,
2457 .switched_from = switched_from_dl,
2458 .switched_to = switched_to_dl,
2459
2460 .update_curr = update_curr_dl,
2461};
2462
2463int sched_dl_global_validate(void)
2464{
2465 u64 runtime = global_rt_runtime();
2466 u64 period = global_rt_period();
2467 u64 new_bw = to_ratio(period, runtime);
2468 struct dl_bw *dl_b;
2469 int cpu, ret = 0;
2470 unsigned long flags;
2471
2472 /*
2473 * Here we want to check the bandwidth not being set to some
2474 * value smaller than the currently allocated bandwidth in
2475 * any of the root_domains.
2476 *
2477 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2478 * cycling on root_domains... Discussion on different/better
2479 * solutions is welcome!
2480 */
2481 for_each_possible_cpu(cpu) {
2482 rcu_read_lock_sched();
2483 dl_b = dl_bw_of(cpu);
2484
2485 raw_spin_lock_irqsave(&dl_b->lock, flags);
2486 if (new_bw < dl_b->total_bw)
2487 ret = -EBUSY;
2488 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2489
2490 rcu_read_unlock_sched();
2491
2492 if (ret)
2493 break;
2494 }
2495
2496 return ret;
2497}
2498
2499void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2500{
2501 if (global_rt_runtime() == RUNTIME_INF) {
2502 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2503 dl_rq->extra_bw = 1 << BW_SHIFT;
2504 } else {
2505 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2506 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2507 dl_rq->extra_bw = to_ratio(global_rt_period(),
2508 global_rt_runtime());
2509 }
2510}
2511
2512void sched_dl_do_global(void)
2513{
2514 u64 new_bw = -1;
2515 struct dl_bw *dl_b;
2516 int cpu;
2517 unsigned long flags;
2518
2519 def_dl_bandwidth.dl_period = global_rt_period();
2520 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2521
2522 if (global_rt_runtime() != RUNTIME_INF)
2523 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2524
2525 /*
2526 * FIXME: As above...
2527 */
2528 for_each_possible_cpu(cpu) {
2529 rcu_read_lock_sched();
2530 dl_b = dl_bw_of(cpu);
2531
2532 raw_spin_lock_irqsave(&dl_b->lock, flags);
2533 dl_b->bw = new_bw;
2534 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2535
2536 rcu_read_unlock_sched();
2537 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2538 }
2539}
2540
2541/*
2542 * We must be sure that accepting a new task (or allowing changing the
2543 * parameters of an existing one) is consistent with the bandwidth
2544 * constraints. If yes, this function also accordingly updates the currently
2545 * allocated bandwidth to reflect the new situation.
2546 *
2547 * This function is called while holding p's rq->lock.
2548 */
2549int sched_dl_overflow(struct task_struct *p, int policy,
2550 const struct sched_attr *attr)
2551{
2552 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2553 u64 period = attr->sched_period ?: attr->sched_deadline;
2554 u64 runtime = attr->sched_runtime;
2555 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2556 int cpus, err = -1;
2557
2558 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2559 return 0;
2560
2561 /* !deadline task may carry old deadline bandwidth */
2562 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2563 return 0;
2564
2565 /*
2566 * Either if a task, enters, leave, or stays -deadline but changes
2567 * its parameters, we may need to update accordingly the total
2568 * allocated bandwidth of the container.
2569 */
2570 raw_spin_lock(&dl_b->lock);
2571 cpus = dl_bw_cpus(task_cpu(p));
2572 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2573 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2574 if (hrtimer_active(&p->dl.inactive_timer))
2575 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2576 __dl_add(dl_b, new_bw, cpus);
2577 err = 0;
2578 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2579 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2580 /*
2581 * XXX this is slightly incorrect: when the task
2582 * utilization decreases, we should delay the total
2583 * utilization change until the task's 0-lag point.
2584 * But this would require to set the task's "inactive
2585 * timer" when the task is not inactive.
2586 */
2587 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2588 __dl_add(dl_b, new_bw, cpus);
2589 dl_change_utilization(p, new_bw);
2590 err = 0;
2591 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2592 /*
2593 * Do not decrease the total deadline utilization here,
2594 * switched_from_dl() will take care to do it at the correct
2595 * (0-lag) time.
2596 */
2597 err = 0;
2598 }
2599 raw_spin_unlock(&dl_b->lock);
2600
2601 return err;
2602}
2603
2604/*
2605 * This function initializes the sched_dl_entity of a newly becoming
2606 * SCHED_DEADLINE task.
2607 *
2608 * Only the static values are considered here, the actual runtime and the
2609 * absolute deadline will be properly calculated when the task is enqueued
2610 * for the first time with its new policy.
2611 */
2612void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2613{
2614 struct sched_dl_entity *dl_se = &p->dl;
2615
2616 dl_se->dl_runtime = attr->sched_runtime;
2617 dl_se->dl_deadline = attr->sched_deadline;
2618 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2619 dl_se->flags = attr->sched_flags;
2620 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2621 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2622}
2623
2624void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2625{
2626 struct sched_dl_entity *dl_se = &p->dl;
2627
2628 attr->sched_priority = p->rt_priority;
2629 attr->sched_runtime = dl_se->dl_runtime;
2630 attr->sched_deadline = dl_se->dl_deadline;
2631 attr->sched_period = dl_se->dl_period;
2632 attr->sched_flags = dl_se->flags;
2633}
2634
2635/*
2636 * This function validates the new parameters of a -deadline task.
2637 * We ask for the deadline not being zero, and greater or equal
2638 * than the runtime, as well as the period of being zero or
2639 * greater than deadline. Furthermore, we have to be sure that
2640 * user parameters are above the internal resolution of 1us (we
2641 * check sched_runtime only since it is always the smaller one) and
2642 * below 2^63 ns (we have to check both sched_deadline and
2643 * sched_period, as the latter can be zero).
2644 */
2645bool __checkparam_dl(const struct sched_attr *attr)
2646{
2647 /* special dl tasks don't actually use any parameter */
2648 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2649 return true;
2650
2651 /* deadline != 0 */
2652 if (attr->sched_deadline == 0)
2653 return false;
2654
2655 /*
2656 * Since we truncate DL_SCALE bits, make sure we're at least
2657 * that big.
2658 */
2659 if (attr->sched_runtime < (1ULL << DL_SCALE))
2660 return false;
2661
2662 /*
2663 * Since we use the MSB for wrap-around and sign issues, make
2664 * sure it's not set (mind that period can be equal to zero).
2665 */
2666 if (attr->sched_deadline & (1ULL << 63) ||
2667 attr->sched_period & (1ULL << 63))
2668 return false;
2669
2670 /* runtime <= deadline <= period (if period != 0) */
2671 if ((attr->sched_period != 0 &&
2672 attr->sched_period < attr->sched_deadline) ||
2673 attr->sched_deadline < attr->sched_runtime)
2674 return false;
2675
2676 return true;
2677}
2678
2679/*
2680 * This function clears the sched_dl_entity static params.
2681 */
2682void __dl_clear_params(struct task_struct *p)
2683{
2684 struct sched_dl_entity *dl_se = &p->dl;
2685
2686 dl_se->dl_runtime = 0;
2687 dl_se->dl_deadline = 0;
2688 dl_se->dl_period = 0;
2689 dl_se->flags = 0;
2690 dl_se->dl_bw = 0;
2691 dl_se->dl_density = 0;
2692
2693 dl_se->dl_throttled = 0;
2694 dl_se->dl_yielded = 0;
2695 dl_se->dl_non_contending = 0;
2696 dl_se->dl_overrun = 0;
2697}
2698
2699bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2700{
2701 struct sched_dl_entity *dl_se = &p->dl;
2702
2703 if (dl_se->dl_runtime != attr->sched_runtime ||
2704 dl_se->dl_deadline != attr->sched_deadline ||
2705 dl_se->dl_period != attr->sched_period ||
2706 dl_se->flags != attr->sched_flags)
2707 return true;
2708
2709 return false;
2710}
2711
2712#ifdef CONFIG_SMP
2713int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2714{
2715 unsigned int dest_cpu;
2716 struct dl_bw *dl_b;
2717 bool overflow;
2718 int cpus, ret;
2719 unsigned long flags;
2720
2721 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2722
2723 rcu_read_lock_sched();
2724 dl_b = dl_bw_of(dest_cpu);
2725 raw_spin_lock_irqsave(&dl_b->lock, flags);
2726 cpus = dl_bw_cpus(dest_cpu);
2727 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
2728 if (overflow) {
2729 ret = -EBUSY;
2730 } else {
2731 /*
2732 * We reserve space for this task in the destination
2733 * root_domain, as we can't fail after this point.
2734 * We will free resources in the source root_domain
2735 * later on (see set_cpus_allowed_dl()).
2736 */
2737 __dl_add(dl_b, p->dl.dl_bw, cpus);
2738 ret = 0;
2739 }
2740 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2741 rcu_read_unlock_sched();
2742
2743 return ret;
2744}
2745
2746int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2747 const struct cpumask *trial)
2748{
2749 int ret = 1, trial_cpus;
2750 struct dl_bw *cur_dl_b;
2751 unsigned long flags;
2752
2753 rcu_read_lock_sched();
2754 cur_dl_b = dl_bw_of(cpumask_any(cur));
2755 trial_cpus = cpumask_weight(trial);
2756
2757 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2758 if (cur_dl_b->bw != -1 &&
2759 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2760 ret = 0;
2761 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2762 rcu_read_unlock_sched();
2763
2764 return ret;
2765}
2766
2767bool dl_cpu_busy(unsigned int cpu)
2768{
2769 unsigned long flags;
2770 struct dl_bw *dl_b;
2771 bool overflow;
2772 int cpus;
2773
2774 rcu_read_lock_sched();
2775 dl_b = dl_bw_of(cpu);
2776 raw_spin_lock_irqsave(&dl_b->lock, flags);
2777 cpus = dl_bw_cpus(cpu);
2778 overflow = __dl_overflow(dl_b, cpus, 0, 0);
2779 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2780 rcu_read_unlock_sched();
2781
2782 return overflow;
2783}
2784#endif
2785
2786#ifdef CONFIG_SCHED_DEBUG
2787void print_dl_stats(struct seq_file *m, int cpu)
2788{
2789 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2790}
2791#endif /* CONFIG_SCHED_DEBUG */