2 * Deadline Scheduling Class (SCHED_DEADLINE)
4 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
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.
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>
19 #include <linux/slab.h>
21 struct dl_bandwidth def_dl_bandwidth;
23 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
25 return container_of(dl_se, struct task_struct, dl);
28 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
30 return container_of(dl_rq, struct rq, dl);
33 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
35 struct task_struct *p = dl_task_of(dl_se);
36 struct rq *rq = task_rq(p);
41 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
43 return !RB_EMPTY_NODE(&dl_se->rb_node);
46 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
48 struct sched_dl_entity *dl_se = &p->dl;
50 return dl_rq->rb_leftmost == &dl_se->rb_node;
53 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
55 raw_spin_lock_init(&dl_b->dl_runtime_lock);
56 dl_b->dl_period = period;
57 dl_b->dl_runtime = runtime;
60 void init_dl_bw(struct dl_bw *dl_b)
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)
67 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
68 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
72 void init_dl_rq(struct dl_rq *dl_rq)
74 dl_rq->rb_root = RB_ROOT;
77 /* zero means no -deadline tasks */
78 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
80 dl_rq->dl_nr_migratory = 0;
81 dl_rq->overloaded = 0;
82 dl_rq->pushable_dl_tasks_root = RB_ROOT;
84 init_dl_bw(&dl_rq->dl_bw);
90 static inline int dl_overloaded(struct rq *rq)
92 return atomic_read(&rq->rd->dlo_count);
95 static inline void dl_set_overload(struct rq *rq)
100 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
102 * Must be visible before the overload count is
103 * set (as in sched_rt.c).
105 * Matched by the barrier in pull_dl_task().
108 atomic_inc(&rq->rd->dlo_count);
111 static inline void dl_clear_overload(struct rq *rq)
116 atomic_dec(&rq->rd->dlo_count);
117 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
120 static void update_dl_migration(struct dl_rq *dl_rq)
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;
127 } else if (dl_rq->overloaded) {
128 dl_clear_overload(rq_of_dl_rq(dl_rq));
129 dl_rq->overloaded = 0;
133 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
135 struct task_struct *p = dl_task_of(dl_se);
137 if (tsk_nr_cpus_allowed(p) > 1)
138 dl_rq->dl_nr_migratory++;
140 update_dl_migration(dl_rq);
143 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
145 struct task_struct *p = dl_task_of(dl_se);
147 if (tsk_nr_cpus_allowed(p) > 1)
148 dl_rq->dl_nr_migratory--;
150 update_dl_migration(dl_rq);
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.
157 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
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;
165 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
169 entry = rb_entry(parent, struct task_struct,
171 if (dl_entity_preempt(&p->dl, &entry->dl))
172 link = &parent->rb_left;
174 link = &parent->rb_right;
180 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
182 rb_link_node(&p->pushable_dl_tasks, parent, link);
183 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
186 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
188 struct dl_rq *dl_rq = &rq->dl;
190 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
193 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
194 struct rb_node *next_node;
196 next_node = rb_next(&p->pushable_dl_tasks);
197 dl_rq->pushable_dl_tasks_leftmost = next_node;
200 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
201 RB_CLEAR_NODE(&p->pushable_dl_tasks);
204 static inline int has_pushable_dl_tasks(struct rq *rq)
206 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
209 static int push_dl_task(struct rq *rq);
211 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
213 return dl_task(prev);
216 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
217 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
219 static void push_dl_tasks(struct rq *);
220 static void pull_dl_task(struct rq *);
222 static inline void queue_push_tasks(struct rq *rq)
224 if (!has_pushable_dl_tasks(rq))
227 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
230 static inline void queue_pull_task(struct rq *rq)
232 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
235 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
237 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
239 struct rq *later_rq = NULL;
240 bool fallback = false;
242 later_rq = find_lock_later_rq(p, rq);
248 * If we cannot preempt any rq, fall back to pick any
252 cpu = cpumask_any_and(cpu_active_mask, tsk_cpus_allowed(p));
253 if (cpu >= nr_cpu_ids) {
255 * Fail to find any suitable cpu.
256 * The task will never come back!
258 BUG_ON(dl_bandwidth_enabled());
261 * If admission control is disabled we
262 * try a little harder to let the task
265 cpu = cpumask_any(cpu_active_mask);
267 later_rq = cpu_rq(cpu);
268 double_lock_balance(rq, later_rq);
272 * By now the task is replenished and enqueued; migrate it.
274 deactivate_task(rq, p, 0);
275 set_task_cpu(p, later_rq->cpu);
276 activate_task(later_rq, p, 0);
279 resched_curr(later_rq);
281 double_unlock_balance(later_rq, rq);
289 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
294 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
299 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
304 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
308 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
313 static inline void pull_dl_task(struct rq *rq)
317 static inline void queue_push_tasks(struct rq *rq)
321 static inline void queue_pull_task(struct rq *rq)
324 #endif /* CONFIG_SMP */
326 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
327 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
328 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
332 * We are being explicitly informed that a new instance is starting,
333 * and this means that:
334 * - the absolute deadline of the entity has to be placed at
335 * current time + relative deadline;
336 * - the runtime of the entity has to be set to the maximum value.
338 * The capability of specifying such event is useful whenever a -deadline
339 * entity wants to (try to!) synchronize its behaviour with the scheduler's
340 * one, and to (try to!) reconcile itself with its own scheduling
343 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se,
344 struct sched_dl_entity *pi_se)
346 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
347 struct rq *rq = rq_of_dl_rq(dl_rq);
349 WARN_ON(!dl_se->dl_new || dl_se->dl_throttled);
352 * We use the regular wall clock time to set deadlines in the
353 * future; in fact, we must consider execution overheads (time
354 * spent on hardirq context, etc.).
356 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
357 dl_se->runtime = pi_se->dl_runtime;
362 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
363 * possibility of a entity lasting more than what it declared, and thus
364 * exhausting its runtime.
366 * Here we are interested in making runtime overrun possible, but we do
367 * not want a entity which is misbehaving to affect the scheduling of all
369 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
370 * is used, in order to confine each entity within its own bandwidth.
372 * This function deals exactly with that, and ensures that when the runtime
373 * of a entity is replenished, its deadline is also postponed. That ensures
374 * the overrunning entity can't interfere with other entity in the system and
375 * can't make them miss their deadlines. Reasons why this kind of overruns
376 * could happen are, typically, a entity voluntarily trying to overcome its
377 * runtime, or it just underestimated it during sched_setattr().
379 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
380 struct sched_dl_entity *pi_se)
382 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
383 struct rq *rq = rq_of_dl_rq(dl_rq);
385 BUG_ON(pi_se->dl_runtime <= 0);
388 * This could be the case for a !-dl task that is boosted.
389 * Just go with full inherited parameters.
391 if (dl_se->dl_deadline == 0) {
392 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
393 dl_se->runtime = pi_se->dl_runtime;
397 * We keep moving the deadline away until we get some
398 * available runtime for the entity. This ensures correct
399 * handling of situations where the runtime overrun is
402 while (dl_se->runtime <= 0) {
403 dl_se->deadline += pi_se->dl_period;
404 dl_se->runtime += pi_se->dl_runtime;
408 * At this point, the deadline really should be "in
409 * the future" with respect to rq->clock. If it's
410 * not, we are, for some reason, lagging too much!
411 * Anyway, after having warn userspace abut that,
412 * we still try to keep the things running by
413 * resetting the deadline and the budget of the
416 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
417 printk_deferred_once("sched: DL replenish lagged to much\n");
418 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
419 dl_se->runtime = pi_se->dl_runtime;
422 if (dl_se->dl_yielded)
423 dl_se->dl_yielded = 0;
424 if (dl_se->dl_throttled)
425 dl_se->dl_throttled = 0;
429 * Here we check if --at time t-- an entity (which is probably being
430 * [re]activated or, in general, enqueued) can use its remaining runtime
431 * and its current deadline _without_ exceeding the bandwidth it is
432 * assigned (function returns true if it can't). We are in fact applying
433 * one of the CBS rules: when a task wakes up, if the residual runtime
434 * over residual deadline fits within the allocated bandwidth, then we
435 * can keep the current (absolute) deadline and residual budget without
436 * disrupting the schedulability of the system. Otherwise, we should
437 * refill the runtime and set the deadline a period in the future,
438 * because keeping the current (absolute) deadline of the task would
439 * result in breaking guarantees promised to other tasks (refer to
440 * Documentation/scheduler/sched-deadline.txt for more informations).
442 * This function returns true if:
444 * runtime / (deadline - t) > dl_runtime / dl_period ,
446 * IOW we can't recycle current parameters.
448 * Notice that the bandwidth check is done against the period. For
449 * task with deadline equal to period this is the same of using
450 * dl_deadline instead of dl_period in the equation above.
452 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
453 struct sched_dl_entity *pi_se, u64 t)
458 * left and right are the two sides of the equation above,
459 * after a bit of shuffling to use multiplications instead
462 * Note that none of the time values involved in the two
463 * multiplications are absolute: dl_deadline and dl_runtime
464 * are the relative deadline and the maximum runtime of each
465 * instance, runtime is the runtime left for the last instance
466 * and (deadline - t), since t is rq->clock, is the time left
467 * to the (absolute) deadline. Even if overflowing the u64 type
468 * is very unlikely to occur in both cases, here we scale down
469 * as we want to avoid that risk at all. Scaling down by 10
470 * means that we reduce granularity to 1us. We are fine with it,
471 * since this is only a true/false check and, anyway, thinking
472 * of anything below microseconds resolution is actually fiction
473 * (but still we want to give the user that illusion >;).
475 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
476 right = ((dl_se->deadline - t) >> DL_SCALE) *
477 (pi_se->dl_runtime >> DL_SCALE);
479 return dl_time_before(right, left);
483 * When a -deadline entity is queued back on the runqueue, its runtime and
484 * deadline might need updating.
486 * The policy here is that we update the deadline of the entity only if:
487 * - the current deadline is in the past,
488 * - using the remaining runtime with the current deadline would make
489 * the entity exceed its bandwidth.
491 static void update_dl_entity(struct sched_dl_entity *dl_se,
492 struct sched_dl_entity *pi_se)
494 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
495 struct rq *rq = rq_of_dl_rq(dl_rq);
498 * The arrival of a new instance needs special treatment, i.e.,
499 * the actual scheduling parameters have to be "renewed".
502 setup_new_dl_entity(dl_se, pi_se);
506 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
507 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
508 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
509 dl_se->runtime = pi_se->dl_runtime;
514 * If the entity depleted all its runtime, and if we want it to sleep
515 * while waiting for some new execution time to become available, we
516 * set the bandwidth enforcement timer to the replenishment instant
517 * and try to activate it.
519 * Notice that it is important for the caller to know if the timer
520 * actually started or not (i.e., the replenishment instant is in
521 * the future or in the past).
523 static int start_dl_timer(struct task_struct *p)
525 struct sched_dl_entity *dl_se = &p->dl;
526 struct hrtimer *timer = &dl_se->dl_timer;
527 struct rq *rq = task_rq(p);
531 lockdep_assert_held(&rq->lock);
534 * We want the timer to fire at the deadline, but considering
535 * that it is actually coming from rq->clock and not from
536 * hrtimer's time base reading.
538 act = ns_to_ktime(dl_se->deadline);
539 now = hrtimer_cb_get_time(timer);
540 delta = ktime_to_ns(now) - rq_clock(rq);
541 act = ktime_add_ns(act, delta);
544 * If the expiry time already passed, e.g., because the value
545 * chosen as the deadline is too small, don't even try to
546 * start the timer in the past!
548 if (ktime_us_delta(act, now) < 0)
552 * !enqueued will guarantee another callback; even if one is already in
553 * progress. This ensures a balanced {get,put}_task_struct().
555 * The race against __run_timer() clearing the enqueued state is
556 * harmless because we're holding task_rq()->lock, therefore the timer
557 * expiring after we've done the check will wait on its task_rq_lock()
558 * and observe our state.
560 if (!hrtimer_is_queued(timer)) {
562 hrtimer_start(timer, act, HRTIMER_MODE_ABS);
569 * This is the bandwidth enforcement timer callback. If here, we know
570 * a task is not on its dl_rq, since the fact that the timer was running
571 * means the task is throttled and needs a runtime replenishment.
573 * However, what we actually do depends on the fact the task is active,
574 * (it is on its rq) or has been removed from there by a call to
575 * dequeue_task_dl(). In the former case we must issue the runtime
576 * replenishment and add the task back to the dl_rq; in the latter, we just
577 * do nothing but clearing dl_throttled, so that runtime and deadline
578 * updating (and the queueing back to dl_rq) will be done by the
579 * next call to enqueue_task_dl().
581 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
583 struct sched_dl_entity *dl_se = container_of(timer,
584 struct sched_dl_entity,
586 struct task_struct *p = dl_task_of(dl_se);
590 rq = task_rq_lock(p, &flags);
593 * The task might have changed its scheduling policy to something
594 * different than SCHED_DEADLINE (through switched_fromd_dl()).
597 __dl_clear_params(p);
602 * This is possible if switched_from_dl() raced against a running
603 * callback that took the above !dl_task() path and we've since then
604 * switched back into SCHED_DEADLINE.
606 * There's nothing to do except drop our task reference.
612 * The task might have been boosted by someone else and might be in the
613 * boosting/deboosting path, its not throttled.
615 if (dl_se->dl_boosted)
619 * Spurious timer due to start_dl_timer() race; or we already received
620 * a replenishment from rt_mutex_setprio().
622 if (!dl_se->dl_throttled)
629 * If the throttle happened during sched-out; like:
636 * __dequeue_task_dl()
639 * We can be both throttled and !queued. Replenish the counter
640 * but do not enqueue -- wait for our wakeup to do that.
642 if (!task_on_rq_queued(p)) {
643 replenish_dl_entity(dl_se, dl_se);
647 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
648 if (dl_task(rq->curr))
649 check_preempt_curr_dl(rq, p, 0);
655 * Perform balancing operations here; after the replenishments. We
656 * cannot drop rq->lock before this, otherwise the assertion in
657 * start_dl_timer() about not missing updates is not true.
659 * If we find that the rq the task was on is no longer available, we
660 * need to select a new rq.
662 * XXX figure out if select_task_rq_dl() deals with offline cpus.
664 if (unlikely(!rq->online))
665 rq = dl_task_offline_migration(rq, p);
668 * Queueing this task back might have overloaded rq, check if we need
669 * to kick someone away.
671 if (has_pushable_dl_tasks(rq)) {
673 * Nothing relies on rq->lock after this, so its safe to drop
676 lockdep_unpin_lock(&rq->lock);
678 lockdep_pin_lock(&rq->lock);
683 task_rq_unlock(rq, p, &flags);
686 * This can free the task_struct, including this hrtimer, do not touch
687 * anything related to that after this.
691 return HRTIMER_NORESTART;
694 void init_dl_task_timer(struct sched_dl_entity *dl_se)
696 struct hrtimer *timer = &dl_se->dl_timer;
698 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
699 timer->function = dl_task_timer;
704 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
706 return (dl_se->runtime <= 0);
709 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
712 * Update the current task's runtime statistics (provided it is still
713 * a -deadline task and has not been removed from the dl_rq).
715 static void update_curr_dl(struct rq *rq)
717 struct task_struct *curr = rq->curr;
718 struct sched_dl_entity *dl_se = &curr->dl;
721 if (!dl_task(curr) || !on_dl_rq(dl_se))
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.
732 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
733 if (unlikely((s64)delta_exec <= 0))
736 schedstat_set(curr->se.statistics.exec_max,
737 max(curr->se.statistics.exec_max, delta_exec));
739 curr->se.sum_exec_runtime += delta_exec;
740 account_group_exec_runtime(curr, delta_exec);
742 curr->se.exec_start = rq_clock_task(rq);
743 cpuacct_charge(curr, delta_exec);
745 sched_rt_avg_update(rq, delta_exec);
747 dl_se->runtime -= dl_se->dl_yielded ? 0 : delta_exec;
748 if (dl_runtime_exceeded(dl_se)) {
749 dl_se->dl_throttled = 1;
750 __dequeue_task_dl(rq, curr, 0);
751 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
752 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
754 if (!is_leftmost(curr, &rq->dl))
759 * Because -- for now -- we share the rt bandwidth, we need to
760 * account our runtime there too, otherwise actual rt tasks
761 * would be able to exceed the shared quota.
763 * Account to the root rt group for now.
765 * The solution we're working towards is having the RT groups scheduled
766 * using deadline servers -- however there's a few nasties to figure
767 * out before that can happen.
769 if (rt_bandwidth_enabled()) {
770 struct rt_rq *rt_rq = &rq->rt;
772 raw_spin_lock(&rt_rq->rt_runtime_lock);
774 * We'll let actual RT tasks worry about the overflow here, we
775 * have our own CBS to keep us inline; only account when RT
776 * bandwidth is relevant.
778 if (sched_rt_bandwidth_account(rt_rq))
779 rt_rq->rt_time += delta_exec;
780 raw_spin_unlock(&rt_rq->rt_runtime_lock);
786 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu);
788 static inline u64 next_deadline(struct rq *rq)
790 struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu);
792 if (next && dl_prio(next->prio))
793 return next->dl.deadline;
798 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
800 struct rq *rq = rq_of_dl_rq(dl_rq);
802 if (dl_rq->earliest_dl.curr == 0 ||
803 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
805 * If the dl_rq had no -deadline tasks, or if the new task
806 * has shorter deadline than the current one on dl_rq, we
807 * know that the previous earliest becomes our next earliest,
808 * as the new task becomes the earliest itself.
810 dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr;
811 dl_rq->earliest_dl.curr = deadline;
812 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1);
813 } else if (dl_rq->earliest_dl.next == 0 ||
814 dl_time_before(deadline, dl_rq->earliest_dl.next)) {
816 * On the other hand, if the new -deadline task has a
817 * a later deadline than the earliest one on dl_rq, but
818 * it is earlier than the next (if any), we must
819 * recompute the next-earliest.
821 dl_rq->earliest_dl.next = next_deadline(rq);
825 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
827 struct rq *rq = rq_of_dl_rq(dl_rq);
830 * Since we may have removed our earliest (and/or next earliest)
831 * task we must recompute them.
833 if (!dl_rq->dl_nr_running) {
834 dl_rq->earliest_dl.curr = 0;
835 dl_rq->earliest_dl.next = 0;
836 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
838 struct rb_node *leftmost = dl_rq->rb_leftmost;
839 struct sched_dl_entity *entry;
841 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
842 dl_rq->earliest_dl.curr = entry->deadline;
843 dl_rq->earliest_dl.next = next_deadline(rq);
844 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1);
850 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
851 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
853 #endif /* CONFIG_SMP */
856 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
858 int prio = dl_task_of(dl_se)->prio;
859 u64 deadline = dl_se->deadline;
861 WARN_ON(!dl_prio(prio));
862 dl_rq->dl_nr_running++;
863 add_nr_running(rq_of_dl_rq(dl_rq), 1);
865 inc_dl_deadline(dl_rq, deadline);
866 inc_dl_migration(dl_se, dl_rq);
870 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
872 int prio = dl_task_of(dl_se)->prio;
874 WARN_ON(!dl_prio(prio));
875 WARN_ON(!dl_rq->dl_nr_running);
876 dl_rq->dl_nr_running--;
877 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
879 dec_dl_deadline(dl_rq, dl_se->deadline);
880 dec_dl_migration(dl_se, dl_rq);
883 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
885 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
886 struct rb_node **link = &dl_rq->rb_root.rb_node;
887 struct rb_node *parent = NULL;
888 struct sched_dl_entity *entry;
891 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
895 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
896 if (dl_time_before(dl_se->deadline, entry->deadline))
897 link = &parent->rb_left;
899 link = &parent->rb_right;
905 dl_rq->rb_leftmost = &dl_se->rb_node;
907 rb_link_node(&dl_se->rb_node, parent, link);
908 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
910 inc_dl_tasks(dl_se, dl_rq);
913 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
915 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
917 if (RB_EMPTY_NODE(&dl_se->rb_node))
920 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
921 struct rb_node *next_node;
923 next_node = rb_next(&dl_se->rb_node);
924 dl_rq->rb_leftmost = next_node;
927 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
928 RB_CLEAR_NODE(&dl_se->rb_node);
930 dec_dl_tasks(dl_se, dl_rq);
934 enqueue_dl_entity(struct sched_dl_entity *dl_se,
935 struct sched_dl_entity *pi_se, int flags)
937 BUG_ON(on_dl_rq(dl_se));
940 * If this is a wakeup or a new instance, the scheduling
941 * parameters of the task might need updating. Otherwise,
942 * we want a replenishment of its runtime.
944 if (dl_se->dl_new || flags & ENQUEUE_WAKEUP)
945 update_dl_entity(dl_se, pi_se);
946 else if (flags & ENQUEUE_REPLENISH)
947 replenish_dl_entity(dl_se, pi_se);
949 __enqueue_dl_entity(dl_se);
952 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
954 __dequeue_dl_entity(dl_se);
957 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
959 struct task_struct *pi_task = rt_mutex_get_top_task(p);
960 struct sched_dl_entity *pi_se = &p->dl;
963 * Use the scheduling parameters of the top pi-waiter
964 * task if we have one and its (absolute) deadline is
965 * smaller than our one... OTW we keep our runtime and
968 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) {
969 pi_se = &pi_task->dl;
970 } else if (!dl_prio(p->normal_prio)) {
972 * Special case in which we have a !SCHED_DEADLINE task
973 * that is going to be deboosted, but exceedes its
974 * runtime while doing so. No point in replenishing
975 * it, as it's going to return back to its original
976 * scheduling class after this.
978 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
983 * If p is throttled, we do nothing. In fact, if it exhausted
984 * its budget it needs a replenishment and, since it now is on
985 * its rq, the bandwidth timer callback (which clearly has not
986 * run yet) will take care of this.
988 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH))
991 enqueue_dl_entity(&p->dl, pi_se, flags);
993 if (!task_current(rq, p) && tsk_nr_cpus_allowed(p) > 1)
994 enqueue_pushable_dl_task(rq, p);
997 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
999 dequeue_dl_entity(&p->dl);
1000 dequeue_pushable_dl_task(rq, p);
1003 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1006 __dequeue_task_dl(rq, p, flags);
1010 * Yield task semantic for -deadline tasks is:
1012 * get off from the CPU until our next instance, with
1013 * a new runtime. This is of little use now, since we
1014 * don't have a bandwidth reclaiming mechanism. Anyway,
1015 * bandwidth reclaiming is planned for the future, and
1016 * yield_task_dl will indicate that some spare budget
1017 * is available for other task instances to use it.
1019 static void yield_task_dl(struct rq *rq)
1021 struct task_struct *p = rq->curr;
1024 * We make the task go to sleep until its current deadline by
1025 * forcing its runtime to zero. This way, update_curr_dl() stops
1026 * it and the bandwidth timer will wake it up and will give it
1027 * new scheduling parameters (thanks to dl_yielded=1).
1029 if (p->dl.runtime > 0) {
1030 rq->curr->dl.dl_yielded = 1;
1033 update_rq_clock(rq);
1036 * Tell update_rq_clock() that we've just updated,
1037 * so we don't do microscopic update in schedule()
1038 * and double the fastpath cost.
1040 rq_clock_skip_update(rq, true);
1045 static int find_later_rq(struct task_struct *task);
1048 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1050 struct task_struct *curr;
1053 if (sd_flag != SD_BALANCE_WAKE)
1059 curr = READ_ONCE(rq->curr); /* unlocked access */
1062 * If we are dealing with a -deadline task, we must
1063 * decide where to wake it up.
1064 * If it has a later deadline and the current task
1065 * on this rq can't move (provided the waking task
1066 * can!) we prefer to send it somewhere else. On the
1067 * other hand, if it has a shorter deadline, we
1068 * try to make it stay here, it might be important.
1070 if (unlikely(dl_task(curr)) &&
1071 (tsk_nr_cpus_allowed(curr) < 2 ||
1072 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1073 (tsk_nr_cpus_allowed(p) > 1)) {
1074 int target = find_later_rq(p);
1077 (dl_time_before(p->dl.deadline,
1078 cpu_rq(target)->dl.earliest_dl.curr) ||
1079 (cpu_rq(target)->dl.dl_nr_running == 0)))
1088 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1091 * Current can't be migrated, useless to reschedule,
1092 * let's hope p can move out.
1094 if (tsk_nr_cpus_allowed(rq->curr) == 1 ||
1095 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
1099 * p is migratable, so let's not schedule it and
1100 * see if it is pushed or pulled somewhere else.
1102 if (tsk_nr_cpus_allowed(p) != 1 &&
1103 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
1109 #endif /* CONFIG_SMP */
1112 * Only called when both the current and waking task are -deadline
1115 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1118 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1125 * In the unlikely case current and p have the same deadline
1126 * let us try to decide what's the best thing to do...
1128 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1129 !test_tsk_need_resched(rq->curr))
1130 check_preempt_equal_dl(rq, p);
1131 #endif /* CONFIG_SMP */
1134 #ifdef CONFIG_SCHED_HRTICK
1135 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1137 hrtick_start(rq, p->dl.runtime);
1139 #else /* !CONFIG_SCHED_HRTICK */
1140 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1145 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1146 struct dl_rq *dl_rq)
1148 struct rb_node *left = dl_rq->rb_leftmost;
1153 return rb_entry(left, struct sched_dl_entity, rb_node);
1156 struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev)
1158 struct sched_dl_entity *dl_se;
1159 struct task_struct *p;
1160 struct dl_rq *dl_rq;
1164 if (need_pull_dl_task(rq, prev)) {
1166 * This is OK, because current is on_cpu, which avoids it being
1167 * picked for load-balance and preemption/IRQs are still
1168 * disabled avoiding further scheduler activity on it and we're
1169 * being very careful to re-start the picking loop.
1171 lockdep_unpin_lock(&rq->lock);
1173 lockdep_pin_lock(&rq->lock);
1175 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1176 * means a stop task can slip in, in which case we need to
1177 * re-start task selection.
1179 if (rq->stop && task_on_rq_queued(rq->stop))
1184 * When prev is DL, we may throttle it in put_prev_task().
1185 * So, we update time before we check for dl_nr_running.
1187 if (prev->sched_class == &dl_sched_class)
1190 if (unlikely(!dl_rq->dl_nr_running))
1193 put_prev_task(rq, prev);
1195 dl_se = pick_next_dl_entity(rq, dl_rq);
1198 p = dl_task_of(dl_se);
1199 p->se.exec_start = rq_clock_task(rq);
1201 /* Running task will never be pushed. */
1202 dequeue_pushable_dl_task(rq, p);
1204 if (hrtick_enabled(rq))
1205 start_hrtick_dl(rq, p);
1207 queue_push_tasks(rq);
1212 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1216 if (on_dl_rq(&p->dl) && tsk_nr_cpus_allowed(p) > 1)
1217 enqueue_pushable_dl_task(rq, p);
1220 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1225 * Even when we have runtime, update_curr_dl() might have resulted in us
1226 * not being the leftmost task anymore. In that case NEED_RESCHED will
1227 * be set and schedule() will start a new hrtick for the next task.
1229 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1230 is_leftmost(p, &rq->dl))
1231 start_hrtick_dl(rq, p);
1234 static void task_fork_dl(struct task_struct *p)
1237 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1242 static void task_dead_dl(struct task_struct *p)
1244 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1247 * Since we are TASK_DEAD we won't slip out of the domain!
1249 raw_spin_lock_irq(&dl_b->lock);
1250 /* XXX we should retain the bw until 0-lag */
1251 dl_b->total_bw -= p->dl.dl_bw;
1252 raw_spin_unlock_irq(&dl_b->lock);
1255 static void set_curr_task_dl(struct rq *rq)
1257 struct task_struct *p = rq->curr;
1259 p->se.exec_start = rq_clock_task(rq);
1261 /* You can't push away the running task */
1262 dequeue_pushable_dl_task(rq, p);
1267 /* Only try algorithms three times */
1268 #define DL_MAX_TRIES 3
1270 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1272 if (!task_running(rq, p) &&
1273 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1278 /* Returns the second earliest -deadline task, NULL otherwise */
1279 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu)
1281 struct rb_node *next_node = rq->dl.rb_leftmost;
1282 struct sched_dl_entity *dl_se;
1283 struct task_struct *p = NULL;
1286 next_node = rb_next(next_node);
1288 dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
1289 p = dl_task_of(dl_se);
1291 if (pick_dl_task(rq, p, cpu))
1301 * Return the earliest pushable rq's task, which is suitable to be executed
1302 * on the CPU, NULL otherwise:
1304 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1306 struct rb_node *next_node = rq->dl.pushable_dl_tasks_leftmost;
1307 struct task_struct *p = NULL;
1309 if (!has_pushable_dl_tasks(rq))
1314 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1316 if (pick_dl_task(rq, p, cpu))
1319 next_node = rb_next(next_node);
1326 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1328 static int find_later_rq(struct task_struct *task)
1330 struct sched_domain *sd;
1331 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1332 int this_cpu = smp_processor_id();
1333 int best_cpu, cpu = task_cpu(task);
1335 /* Make sure the mask is initialized first */
1336 if (unlikely(!later_mask))
1339 if (tsk_nr_cpus_allowed(task) == 1)
1343 * We have to consider system topology and task affinity
1344 * first, then we can look for a suitable cpu.
1346 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
1352 * If we are here, some target has been found,
1353 * the most suitable of which is cached in best_cpu.
1354 * This is, among the runqueues where the current tasks
1355 * have later deadlines than the task's one, the rq
1356 * with the latest possible one.
1358 * Now we check how well this matches with task's
1359 * affinity and system topology.
1361 * The last cpu where the task run is our first
1362 * guess, since it is most likely cache-hot there.
1364 if (cpumask_test_cpu(cpu, later_mask))
1367 * Check if this_cpu is to be skipped (i.e., it is
1368 * not in the mask) or not.
1370 if (!cpumask_test_cpu(this_cpu, later_mask))
1374 for_each_domain(cpu, sd) {
1375 if (sd->flags & SD_WAKE_AFFINE) {
1378 * If possible, preempting this_cpu is
1379 * cheaper than migrating.
1381 if (this_cpu != -1 &&
1382 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1388 * Last chance: if best_cpu is valid and is
1389 * in the mask, that becomes our choice.
1391 if (best_cpu < nr_cpu_ids &&
1392 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
1401 * At this point, all our guesses failed, we just return
1402 * 'something', and let the caller sort the things out.
1407 cpu = cpumask_any(later_mask);
1408 if (cpu < nr_cpu_ids)
1414 /* Locks the rq it finds */
1415 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1417 struct rq *later_rq = NULL;
1421 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1422 cpu = find_later_rq(task);
1424 if ((cpu == -1) || (cpu == rq->cpu))
1427 later_rq = cpu_rq(cpu);
1429 if (later_rq->dl.dl_nr_running &&
1430 !dl_time_before(task->dl.deadline,
1431 later_rq->dl.earliest_dl.curr)) {
1433 * Target rq has tasks of equal or earlier deadline,
1434 * retrying does not release any lock and is unlikely
1435 * to yield a different result.
1441 /* Retry if something changed. */
1442 if (double_lock_balance(rq, later_rq)) {
1443 if (unlikely(task_rq(task) != rq ||
1444 !cpumask_test_cpu(later_rq->cpu,
1445 tsk_cpus_allowed(task)) ||
1446 task_running(rq, task) ||
1447 !task_on_rq_queued(task))) {
1448 double_unlock_balance(rq, later_rq);
1455 * If the rq we found has no -deadline task, or
1456 * its earliest one has a later deadline than our
1457 * task, the rq is a good one.
1459 if (!later_rq->dl.dl_nr_running ||
1460 dl_time_before(task->dl.deadline,
1461 later_rq->dl.earliest_dl.curr))
1464 /* Otherwise we try again. */
1465 double_unlock_balance(rq, later_rq);
1472 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1474 struct task_struct *p;
1476 if (!has_pushable_dl_tasks(rq))
1479 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
1480 struct task_struct, pushable_dl_tasks);
1482 BUG_ON(rq->cpu != task_cpu(p));
1483 BUG_ON(task_current(rq, p));
1484 BUG_ON(tsk_nr_cpus_allowed(p) <= 1);
1486 BUG_ON(!task_on_rq_queued(p));
1487 BUG_ON(!dl_task(p));
1493 * See if the non running -deadline tasks on this rq
1494 * can be sent to some other CPU where they can preempt
1495 * and start executing.
1497 static int push_dl_task(struct rq *rq)
1499 struct task_struct *next_task;
1500 struct rq *later_rq;
1503 if (!rq->dl.overloaded)
1506 next_task = pick_next_pushable_dl_task(rq);
1511 if (unlikely(next_task == rq->curr)) {
1517 * If next_task preempts rq->curr, and rq->curr
1518 * can move away, it makes sense to just reschedule
1519 * without going further in pushing next_task.
1521 if (dl_task(rq->curr) &&
1522 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1523 tsk_nr_cpus_allowed(rq->curr) > 1) {
1528 /* We might release rq lock */
1529 get_task_struct(next_task);
1531 /* Will lock the rq it'll find */
1532 later_rq = find_lock_later_rq(next_task, rq);
1534 struct task_struct *task;
1537 * We must check all this again, since
1538 * find_lock_later_rq releases rq->lock and it is
1539 * then possible that next_task has migrated.
1541 task = pick_next_pushable_dl_task(rq);
1542 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1544 * The task is still there. We don't try
1545 * again, some other cpu will pull it when ready.
1554 put_task_struct(next_task);
1559 deactivate_task(rq, next_task, 0);
1560 set_task_cpu(next_task, later_rq->cpu);
1561 activate_task(later_rq, next_task, 0);
1564 resched_curr(later_rq);
1566 double_unlock_balance(rq, later_rq);
1569 put_task_struct(next_task);
1574 static void push_dl_tasks(struct rq *rq)
1576 /* push_dl_task() will return true if it moved a -deadline task */
1577 while (push_dl_task(rq))
1581 static void pull_dl_task(struct rq *this_rq)
1583 int this_cpu = this_rq->cpu, cpu;
1584 struct task_struct *p;
1585 bool resched = false;
1587 u64 dmin = LONG_MAX;
1589 if (likely(!dl_overloaded(this_rq)))
1593 * Match the barrier from dl_set_overloaded; this guarantees that if we
1594 * see overloaded we must also see the dlo_mask bit.
1598 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
1599 if (this_cpu == cpu)
1602 src_rq = cpu_rq(cpu);
1605 * It looks racy, abd it is! However, as in sched_rt.c,
1606 * we are fine with this.
1608 if (this_rq->dl.dl_nr_running &&
1609 dl_time_before(this_rq->dl.earliest_dl.curr,
1610 src_rq->dl.earliest_dl.next))
1613 /* Might drop this_rq->lock */
1614 double_lock_balance(this_rq, src_rq);
1617 * If there are no more pullable tasks on the
1618 * rq, we're done with it.
1620 if (src_rq->dl.dl_nr_running <= 1)
1623 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
1626 * We found a task to be pulled if:
1627 * - it preempts our current (if there's one),
1628 * - it will preempt the last one we pulled (if any).
1630 if (p && dl_time_before(p->dl.deadline, dmin) &&
1631 (!this_rq->dl.dl_nr_running ||
1632 dl_time_before(p->dl.deadline,
1633 this_rq->dl.earliest_dl.curr))) {
1634 WARN_ON(p == src_rq->curr);
1635 WARN_ON(!task_on_rq_queued(p));
1638 * Then we pull iff p has actually an earlier
1639 * deadline than the current task of its runqueue.
1641 if (dl_time_before(p->dl.deadline,
1642 src_rq->curr->dl.deadline))
1647 deactivate_task(src_rq, p, 0);
1648 set_task_cpu(p, this_cpu);
1649 activate_task(this_rq, p, 0);
1650 dmin = p->dl.deadline;
1652 /* Is there any other task even earlier? */
1655 double_unlock_balance(this_rq, src_rq);
1659 resched_curr(this_rq);
1663 * Since the task is not running and a reschedule is not going to happen
1664 * anytime soon on its runqueue, we try pushing it away now.
1666 static void task_woken_dl(struct rq *rq, struct task_struct *p)
1668 if (!task_running(rq, p) &&
1669 !test_tsk_need_resched(rq->curr) &&
1670 tsk_nr_cpus_allowed(p) > 1 &&
1671 dl_task(rq->curr) &&
1672 (tsk_nr_cpus_allowed(rq->curr) < 2 ||
1673 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
1678 static void set_cpus_allowed_dl(struct task_struct *p,
1679 const struct cpumask *new_mask)
1681 struct root_domain *src_rd;
1684 BUG_ON(!dl_task(p));
1689 * Migrating a SCHED_DEADLINE task between exclusive
1690 * cpusets (different root_domains) entails a bandwidth
1691 * update. We already made space for us in the destination
1692 * domain (see cpuset_can_attach()).
1694 if (!cpumask_intersects(src_rd->span, new_mask)) {
1695 struct dl_bw *src_dl_b;
1697 src_dl_b = dl_bw_of(cpu_of(rq));
1699 * We now free resources of the root_domain we are migrating
1700 * off. In the worst case, sched_setattr() may temporary fail
1701 * until we complete the update.
1703 raw_spin_lock(&src_dl_b->lock);
1704 __dl_clear(src_dl_b, p->dl.dl_bw);
1705 raw_spin_unlock(&src_dl_b->lock);
1708 set_cpus_allowed_common(p, new_mask);
1711 /* Assumes rq->lock is held */
1712 static void rq_online_dl(struct rq *rq)
1714 if (rq->dl.overloaded)
1715 dl_set_overload(rq);
1717 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
1718 if (rq->dl.dl_nr_running > 0)
1719 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1);
1722 /* Assumes rq->lock is held */
1723 static void rq_offline_dl(struct rq *rq)
1725 if (rq->dl.overloaded)
1726 dl_clear_overload(rq);
1728 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
1729 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
1732 void __init init_sched_dl_class(void)
1736 for_each_possible_cpu(i)
1737 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
1738 GFP_KERNEL, cpu_to_node(i));
1741 #endif /* CONFIG_SMP */
1743 static void switched_from_dl(struct rq *rq, struct task_struct *p)
1746 * Start the deadline timer; if we switch back to dl before this we'll
1747 * continue consuming our current CBS slice. If we stay outside of
1748 * SCHED_DEADLINE until the deadline passes, the timer will reset the
1751 if (!start_dl_timer(p))
1752 __dl_clear_params(p);
1755 * Since this might be the only -deadline task on the rq,
1756 * this is the right place to try to pull some other one
1757 * from an overloaded cpu, if any.
1759 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
1762 queue_pull_task(rq);
1766 * When switching to -deadline, we may overload the rq, then
1767 * we try to push someone off, if possible.
1769 static void switched_to_dl(struct rq *rq, struct task_struct *p)
1771 if (task_on_rq_queued(p) && rq->curr != p) {
1773 if (tsk_nr_cpus_allowed(p) > 1 && rq->dl.overloaded)
1774 queue_push_tasks(rq);
1776 if (dl_task(rq->curr))
1777 check_preempt_curr_dl(rq, p, 0);
1785 * If the scheduling parameters of a -deadline task changed,
1786 * a push or pull operation might be needed.
1788 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
1791 if (task_on_rq_queued(p) || rq->curr == p) {
1794 * This might be too much, but unfortunately
1795 * we don't have the old deadline value, and
1796 * we can't argue if the task is increasing
1797 * or lowering its prio, so...
1799 if (!rq->dl.overloaded)
1800 queue_pull_task(rq);
1803 * If we now have a earlier deadline task than p,
1804 * then reschedule, provided p is still on this
1807 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
1811 * Again, we don't know if p has a earlier
1812 * or later deadline, so let's blindly set a
1813 * (maybe not needed) rescheduling point.
1816 #endif /* CONFIG_SMP */
1818 switched_to_dl(rq, p);
1821 const struct sched_class dl_sched_class = {
1822 .next = &rt_sched_class,
1823 .enqueue_task = enqueue_task_dl,
1824 .dequeue_task = dequeue_task_dl,
1825 .yield_task = yield_task_dl,
1827 .check_preempt_curr = check_preempt_curr_dl,
1829 .pick_next_task = pick_next_task_dl,
1830 .put_prev_task = put_prev_task_dl,
1833 .select_task_rq = select_task_rq_dl,
1834 .set_cpus_allowed = set_cpus_allowed_dl,
1835 .rq_online = rq_online_dl,
1836 .rq_offline = rq_offline_dl,
1837 .task_woken = task_woken_dl,
1840 .set_curr_task = set_curr_task_dl,
1841 .task_tick = task_tick_dl,
1842 .task_fork = task_fork_dl,
1843 .task_dead = task_dead_dl,
1845 .prio_changed = prio_changed_dl,
1846 .switched_from = switched_from_dl,
1847 .switched_to = switched_to_dl,
1849 .update_curr = update_curr_dl,
1852 #ifdef CONFIG_SCHED_DEBUG
1853 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
1855 void print_dl_stats(struct seq_file *m, int cpu)
1857 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
1859 #endif /* CONFIG_SCHED_DEBUG */