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