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 extern unsigned long to_ratio(u64 period, u64 runtime);
62 void init_dl_bw(struct dl_bw *dl_b)
64 raw_spin_lock_init(&dl_b->lock);
65 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
66 if (global_rt_runtime() == RUNTIME_INF)
69 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
70 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
74 void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq)
76 dl_rq->rb_root = RB_ROOT;
79 /* zero means no -deadline tasks */
80 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
82 dl_rq->dl_nr_migratory = 0;
83 dl_rq->overloaded = 0;
84 dl_rq->pushable_dl_tasks_root = RB_ROOT;
86 init_dl_bw(&dl_rq->dl_bw);
92 static inline int dl_overloaded(struct rq *rq)
94 return atomic_read(&rq->rd->dlo_count);
97 static inline void dl_set_overload(struct rq *rq)
102 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
104 * Must be visible before the overload count is
105 * set (as in sched_rt.c).
107 * Matched by the barrier in pull_dl_task().
110 atomic_inc(&rq->rd->dlo_count);
113 static inline void dl_clear_overload(struct rq *rq)
118 atomic_dec(&rq->rd->dlo_count);
119 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
122 static void update_dl_migration(struct dl_rq *dl_rq)
124 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
125 if (!dl_rq->overloaded) {
126 dl_set_overload(rq_of_dl_rq(dl_rq));
127 dl_rq->overloaded = 1;
129 } else if (dl_rq->overloaded) {
130 dl_clear_overload(rq_of_dl_rq(dl_rq));
131 dl_rq->overloaded = 0;
135 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
137 struct task_struct *p = dl_task_of(dl_se);
139 if (p->nr_cpus_allowed > 1)
140 dl_rq->dl_nr_migratory++;
142 update_dl_migration(dl_rq);
145 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
147 struct task_struct *p = dl_task_of(dl_se);
149 if (p->nr_cpus_allowed > 1)
150 dl_rq->dl_nr_migratory--;
152 update_dl_migration(dl_rq);
156 * The list of pushable -deadline task is not a plist, like in
157 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
159 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
161 struct dl_rq *dl_rq = &rq->dl;
162 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node;
163 struct rb_node *parent = NULL;
164 struct task_struct *entry;
167 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
171 entry = rb_entry(parent, struct task_struct,
173 if (dl_entity_preempt(&p->dl, &entry->dl))
174 link = &parent->rb_left;
176 link = &parent->rb_right;
182 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
184 rb_link_node(&p->pushable_dl_tasks, parent, link);
185 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
188 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
190 struct dl_rq *dl_rq = &rq->dl;
192 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
195 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
196 struct rb_node *next_node;
198 next_node = rb_next(&p->pushable_dl_tasks);
199 dl_rq->pushable_dl_tasks_leftmost = next_node;
202 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
203 RB_CLEAR_NODE(&p->pushable_dl_tasks);
206 static inline int has_pushable_dl_tasks(struct rq *rq)
208 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
211 static int push_dl_task(struct rq *rq);
213 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
215 return dl_task(prev);
218 static inline void set_post_schedule(struct rq *rq)
220 rq->post_schedule = has_pushable_dl_tasks(rq);
226 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
231 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
236 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
241 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
245 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
250 static inline int pull_dl_task(struct rq *rq)
255 static inline void set_post_schedule(struct rq *rq)
258 #endif /* CONFIG_SMP */
260 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
261 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
262 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
266 * We are being explicitly informed that a new instance is starting,
267 * and this means that:
268 * - the absolute deadline of the entity has to be placed at
269 * current time + relative deadline;
270 * - the runtime of the entity has to be set to the maximum value.
272 * The capability of specifying such event is useful whenever a -deadline
273 * entity wants to (try to!) synchronize its behaviour with the scheduler's
274 * one, and to (try to!) reconcile itself with its own scheduling
277 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se,
278 struct sched_dl_entity *pi_se)
280 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
281 struct rq *rq = rq_of_dl_rq(dl_rq);
283 WARN_ON(!dl_se->dl_new || dl_se->dl_throttled);
286 * We use the regular wall clock time to set deadlines in the
287 * future; in fact, we must consider execution overheads (time
288 * spent on hardirq context, etc.).
290 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
291 dl_se->runtime = pi_se->dl_runtime;
296 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
297 * possibility of a entity lasting more than what it declared, and thus
298 * exhausting its runtime.
300 * Here we are interested in making runtime overrun possible, but we do
301 * not want a entity which is misbehaving to affect the scheduling of all
303 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
304 * is used, in order to confine each entity within its own bandwidth.
306 * This function deals exactly with that, and ensures that when the runtime
307 * of a entity is replenished, its deadline is also postponed. That ensures
308 * the overrunning entity can't interfere with other entity in the system and
309 * can't make them miss their deadlines. Reasons why this kind of overruns
310 * could happen are, typically, a entity voluntarily trying to overcome its
311 * runtime, or it just underestimated it during sched_setscheduler_ex().
313 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
314 struct sched_dl_entity *pi_se)
316 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
317 struct rq *rq = rq_of_dl_rq(dl_rq);
319 BUG_ON(pi_se->dl_runtime <= 0);
322 * This could be the case for a !-dl task that is boosted.
323 * Just go with full inherited parameters.
325 if (dl_se->dl_deadline == 0) {
326 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
327 dl_se->runtime = pi_se->dl_runtime;
331 * We keep moving the deadline away until we get some
332 * available runtime for the entity. This ensures correct
333 * handling of situations where the runtime overrun is
336 while (dl_se->runtime <= 0) {
337 dl_se->deadline += pi_se->dl_period;
338 dl_se->runtime += pi_se->dl_runtime;
342 * At this point, the deadline really should be "in
343 * the future" with respect to rq->clock. If it's
344 * not, we are, for some reason, lagging too much!
345 * Anyway, after having warn userspace abut that,
346 * we still try to keep the things running by
347 * resetting the deadline and the budget of the
350 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
351 static bool lag_once = false;
355 printk_sched("sched: DL replenish lagged to much\n");
357 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
358 dl_se->runtime = pi_se->dl_runtime;
363 * Here we check if --at time t-- an entity (which is probably being
364 * [re]activated or, in general, enqueued) can use its remaining runtime
365 * and its current deadline _without_ exceeding the bandwidth it is
366 * assigned (function returns true if it can't). We are in fact applying
367 * one of the CBS rules: when a task wakes up, if the residual runtime
368 * over residual deadline fits within the allocated bandwidth, then we
369 * can keep the current (absolute) deadline and residual budget without
370 * disrupting the schedulability of the system. Otherwise, we should
371 * refill the runtime and set the deadline a period in the future,
372 * because keeping the current (absolute) deadline of the task would
373 * result in breaking guarantees promised to other tasks (refer to
374 * Documentation/scheduler/sched-deadline.txt for more informations).
376 * This function returns true if:
378 * runtime / (deadline - t) > dl_runtime / dl_period ,
380 * IOW we can't recycle current parameters.
382 * Notice that the bandwidth check is done against the period. For
383 * task with deadline equal to period this is the same of using
384 * dl_deadline instead of dl_period in the equation above.
386 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
387 struct sched_dl_entity *pi_se, u64 t)
392 * left and right are the two sides of the equation above,
393 * after a bit of shuffling to use multiplications instead
396 * Note that none of the time values involved in the two
397 * multiplications are absolute: dl_deadline and dl_runtime
398 * are the relative deadline and the maximum runtime of each
399 * instance, runtime is the runtime left for the last instance
400 * and (deadline - t), since t is rq->clock, is the time left
401 * to the (absolute) deadline. Even if overflowing the u64 type
402 * is very unlikely to occur in both cases, here we scale down
403 * as we want to avoid that risk at all. Scaling down by 10
404 * means that we reduce granularity to 1us. We are fine with it,
405 * since this is only a true/false check and, anyway, thinking
406 * of anything below microseconds resolution is actually fiction
407 * (but still we want to give the user that illusion >;).
409 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
410 right = ((dl_se->deadline - t) >> DL_SCALE) *
411 (pi_se->dl_runtime >> DL_SCALE);
413 return dl_time_before(right, left);
417 * When a -deadline entity is queued back on the runqueue, its runtime and
418 * deadline might need updating.
420 * The policy here is that we update the deadline of the entity only if:
421 * - the current deadline is in the past,
422 * - using the remaining runtime with the current deadline would make
423 * the entity exceed its bandwidth.
425 static void update_dl_entity(struct sched_dl_entity *dl_se,
426 struct sched_dl_entity *pi_se)
428 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
429 struct rq *rq = rq_of_dl_rq(dl_rq);
432 * The arrival of a new instance needs special treatment, i.e.,
433 * the actual scheduling parameters have to be "renewed".
436 setup_new_dl_entity(dl_se, pi_se);
440 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
441 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
442 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
443 dl_se->runtime = pi_se->dl_runtime;
448 * If the entity depleted all its runtime, and if we want it to sleep
449 * while waiting for some new execution time to become available, we
450 * set the bandwidth enforcement timer to the replenishment instant
451 * and try to activate it.
453 * Notice that it is important for the caller to know if the timer
454 * actually started or not (i.e., the replenishment instant is in
455 * the future or in the past).
457 static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted)
459 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
460 struct rq *rq = rq_of_dl_rq(dl_rq);
469 * We want the timer to fire at the deadline, but considering
470 * that it is actually coming from rq->clock and not from
471 * hrtimer's time base reading.
473 act = ns_to_ktime(dl_se->deadline);
474 now = hrtimer_cb_get_time(&dl_se->dl_timer);
475 delta = ktime_to_ns(now) - rq_clock(rq);
476 act = ktime_add_ns(act, delta);
479 * If the expiry time already passed, e.g., because the value
480 * chosen as the deadline is too small, don't even try to
481 * start the timer in the past!
483 if (ktime_us_delta(act, now) < 0)
486 hrtimer_set_expires(&dl_se->dl_timer, act);
488 soft = hrtimer_get_softexpires(&dl_se->dl_timer);
489 hard = hrtimer_get_expires(&dl_se->dl_timer);
490 range = ktime_to_ns(ktime_sub(hard, soft));
491 __hrtimer_start_range_ns(&dl_se->dl_timer, soft,
492 range, HRTIMER_MODE_ABS, 0);
494 return hrtimer_active(&dl_se->dl_timer);
498 * This is the bandwidth enforcement timer callback. If here, we know
499 * a task is not on its dl_rq, since the fact that the timer was running
500 * means the task is throttled and needs a runtime replenishment.
502 * However, what we actually do depends on the fact the task is active,
503 * (it is on its rq) or has been removed from there by a call to
504 * dequeue_task_dl(). In the former case we must issue the runtime
505 * replenishment and add the task back to the dl_rq; in the latter, we just
506 * do nothing but clearing dl_throttled, so that runtime and deadline
507 * updating (and the queueing back to dl_rq) will be done by the
508 * next call to enqueue_task_dl().
510 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
512 struct sched_dl_entity *dl_se = container_of(timer,
513 struct sched_dl_entity,
515 struct task_struct *p = dl_task_of(dl_se);
516 struct rq *rq = task_rq(p);
517 raw_spin_lock(&rq->lock);
520 * We need to take care of a possible races here. In fact, the
521 * task might have changed its scheduling policy to something
522 * different from SCHED_DEADLINE or changed its reservation
523 * parameters (through sched_setscheduler()).
525 if (!dl_task(p) || dl_se->dl_new)
530 dl_se->dl_throttled = 0;
531 dl_se->dl_yielded = 0;
533 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
534 if (task_has_dl_policy(rq->curr))
535 check_preempt_curr_dl(rq, p, 0);
537 resched_task(rq->curr);
540 * Queueing this task back might have overloaded rq,
541 * check if we need to kick someone away.
543 if (has_pushable_dl_tasks(rq))
548 raw_spin_unlock(&rq->lock);
550 return HRTIMER_NORESTART;
553 void init_dl_task_timer(struct sched_dl_entity *dl_se)
555 struct hrtimer *timer = &dl_se->dl_timer;
557 if (hrtimer_active(timer)) {
558 hrtimer_try_to_cancel(timer);
562 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
563 timer->function = dl_task_timer;
567 int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
569 int dmiss = dl_time_before(dl_se->deadline, rq_clock(rq));
570 int rorun = dl_se->runtime <= 0;
572 if (!rorun && !dmiss)
576 * If we are beyond our current deadline and we are still
577 * executing, then we have already used some of the runtime of
578 * the next instance. Thus, if we do not account that, we are
579 * stealing bandwidth from the system at each deadline miss!
582 dl_se->runtime = rorun ? dl_se->runtime : 0;
583 dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
589 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
592 * Update the current task's runtime statistics (provided it is still
593 * a -deadline task and has not been removed from the dl_rq).
595 static void update_curr_dl(struct rq *rq)
597 struct task_struct *curr = rq->curr;
598 struct sched_dl_entity *dl_se = &curr->dl;
601 if (!dl_task(curr) || !on_dl_rq(dl_se))
605 * Consumed budget is computed considering the time as
606 * observed by schedulable tasks (excluding time spent
607 * in hardirq context, etc.). Deadlines are instead
608 * computed using hard walltime. This seems to be the more
609 * natural solution, but the full ramifications of this
610 * approach need further study.
612 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
613 if (unlikely((s64)delta_exec <= 0))
616 schedstat_set(curr->se.statistics.exec_max,
617 max(curr->se.statistics.exec_max, delta_exec));
619 curr->se.sum_exec_runtime += delta_exec;
620 account_group_exec_runtime(curr, delta_exec);
622 curr->se.exec_start = rq_clock_task(rq);
623 cpuacct_charge(curr, delta_exec);
625 sched_rt_avg_update(rq, delta_exec);
627 dl_se->runtime -= delta_exec;
628 if (dl_runtime_exceeded(rq, dl_se)) {
629 __dequeue_task_dl(rq, curr, 0);
630 if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted)))
631 dl_se->dl_throttled = 1;
633 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
635 if (!is_leftmost(curr, &rq->dl))
640 * Because -- for now -- we share the rt bandwidth, we need to
641 * account our runtime there too, otherwise actual rt tasks
642 * would be able to exceed the shared quota.
644 * Account to the root rt group for now.
646 * The solution we're working towards is having the RT groups scheduled
647 * using deadline servers -- however there's a few nasties to figure
648 * out before that can happen.
650 if (rt_bandwidth_enabled()) {
651 struct rt_rq *rt_rq = &rq->rt;
653 raw_spin_lock(&rt_rq->rt_runtime_lock);
655 * We'll let actual RT tasks worry about the overflow here, we
656 * have our own CBS to keep us inline; only account when RT
657 * bandwidth is relevant.
659 if (sched_rt_bandwidth_account(rt_rq))
660 rt_rq->rt_time += delta_exec;
661 raw_spin_unlock(&rt_rq->rt_runtime_lock);
667 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu);
669 static inline u64 next_deadline(struct rq *rq)
671 struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu);
673 if (next && dl_prio(next->prio))
674 return next->dl.deadline;
679 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
681 struct rq *rq = rq_of_dl_rq(dl_rq);
683 if (dl_rq->earliest_dl.curr == 0 ||
684 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
686 * If the dl_rq had no -deadline tasks, or if the new task
687 * has shorter deadline than the current one on dl_rq, we
688 * know that the previous earliest becomes our next earliest,
689 * as the new task becomes the earliest itself.
691 dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr;
692 dl_rq->earliest_dl.curr = deadline;
693 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1);
694 } else if (dl_rq->earliest_dl.next == 0 ||
695 dl_time_before(deadline, dl_rq->earliest_dl.next)) {
697 * On the other hand, if the new -deadline task has a
698 * a later deadline than the earliest one on dl_rq, but
699 * it is earlier than the next (if any), we must
700 * recompute the next-earliest.
702 dl_rq->earliest_dl.next = next_deadline(rq);
706 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
708 struct rq *rq = rq_of_dl_rq(dl_rq);
711 * Since we may have removed our earliest (and/or next earliest)
712 * task we must recompute them.
714 if (!dl_rq->dl_nr_running) {
715 dl_rq->earliest_dl.curr = 0;
716 dl_rq->earliest_dl.next = 0;
717 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
719 struct rb_node *leftmost = dl_rq->rb_leftmost;
720 struct sched_dl_entity *entry;
722 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
723 dl_rq->earliest_dl.curr = entry->deadline;
724 dl_rq->earliest_dl.next = next_deadline(rq);
725 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1);
731 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
732 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
734 #endif /* CONFIG_SMP */
737 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
739 int prio = dl_task_of(dl_se)->prio;
740 u64 deadline = dl_se->deadline;
742 WARN_ON(!dl_prio(prio));
743 dl_rq->dl_nr_running++;
744 inc_nr_running(rq_of_dl_rq(dl_rq));
746 inc_dl_deadline(dl_rq, deadline);
747 inc_dl_migration(dl_se, dl_rq);
751 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
753 int prio = dl_task_of(dl_se)->prio;
755 WARN_ON(!dl_prio(prio));
756 WARN_ON(!dl_rq->dl_nr_running);
757 dl_rq->dl_nr_running--;
758 dec_nr_running(rq_of_dl_rq(dl_rq));
760 dec_dl_deadline(dl_rq, dl_se->deadline);
761 dec_dl_migration(dl_se, dl_rq);
764 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
766 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
767 struct rb_node **link = &dl_rq->rb_root.rb_node;
768 struct rb_node *parent = NULL;
769 struct sched_dl_entity *entry;
772 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
776 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
777 if (dl_time_before(dl_se->deadline, entry->deadline))
778 link = &parent->rb_left;
780 link = &parent->rb_right;
786 dl_rq->rb_leftmost = &dl_se->rb_node;
788 rb_link_node(&dl_se->rb_node, parent, link);
789 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
791 inc_dl_tasks(dl_se, dl_rq);
794 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
796 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
798 if (RB_EMPTY_NODE(&dl_se->rb_node))
801 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
802 struct rb_node *next_node;
804 next_node = rb_next(&dl_se->rb_node);
805 dl_rq->rb_leftmost = next_node;
808 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
809 RB_CLEAR_NODE(&dl_se->rb_node);
811 dec_dl_tasks(dl_se, dl_rq);
815 enqueue_dl_entity(struct sched_dl_entity *dl_se,
816 struct sched_dl_entity *pi_se, int flags)
818 BUG_ON(on_dl_rq(dl_se));
821 * If this is a wakeup or a new instance, the scheduling
822 * parameters of the task might need updating. Otherwise,
823 * we want a replenishment of its runtime.
825 if (!dl_se->dl_new && flags & ENQUEUE_REPLENISH)
826 replenish_dl_entity(dl_se, pi_se);
828 update_dl_entity(dl_se, pi_se);
830 __enqueue_dl_entity(dl_se);
833 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
835 __dequeue_dl_entity(dl_se);
838 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
840 struct task_struct *pi_task = rt_mutex_get_top_task(p);
841 struct sched_dl_entity *pi_se = &p->dl;
844 * Use the scheduling parameters of the top pi-waiter
845 * task if we have one and its (relative) deadline is
846 * smaller than our one... OTW we keep our runtime and
849 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio))
850 pi_se = &pi_task->dl;
853 * If p is throttled, we do nothing. In fact, if it exhausted
854 * its budget it needs a replenishment and, since it now is on
855 * its rq, the bandwidth timer callback (which clearly has not
856 * run yet) will take care of this.
858 if (p->dl.dl_throttled)
861 enqueue_dl_entity(&p->dl, pi_se, flags);
863 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
864 enqueue_pushable_dl_task(rq, p);
867 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
869 dequeue_dl_entity(&p->dl);
870 dequeue_pushable_dl_task(rq, p);
873 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
876 __dequeue_task_dl(rq, p, flags);
880 * Yield task semantic for -deadline tasks is:
882 * get off from the CPU until our next instance, with
883 * a new runtime. This is of little use now, since we
884 * don't have a bandwidth reclaiming mechanism. Anyway,
885 * bandwidth reclaiming is planned for the future, and
886 * yield_task_dl will indicate that some spare budget
887 * is available for other task instances to use it.
889 static void yield_task_dl(struct rq *rq)
891 struct task_struct *p = rq->curr;
894 * We make the task go to sleep until its current deadline by
895 * forcing its runtime to zero. This way, update_curr_dl() stops
896 * it and the bandwidth timer will wake it up and will give it
897 * new scheduling parameters (thanks to dl_yielded=1).
899 if (p->dl.runtime > 0) {
900 rq->curr->dl.dl_yielded = 1;
908 static int find_later_rq(struct task_struct *task);
911 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
913 struct task_struct *curr;
916 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
922 curr = ACCESS_ONCE(rq->curr); /* unlocked access */
925 * If we are dealing with a -deadline task, we must
926 * decide where to wake it up.
927 * If it has a later deadline and the current task
928 * on this rq can't move (provided the waking task
929 * can!) we prefer to send it somewhere else. On the
930 * other hand, if it has a shorter deadline, we
931 * try to make it stay here, it might be important.
933 if (unlikely(dl_task(curr)) &&
934 (curr->nr_cpus_allowed < 2 ||
935 !dl_entity_preempt(&p->dl, &curr->dl)) &&
936 (p->nr_cpus_allowed > 1)) {
937 int target = find_later_rq(p);
948 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
951 * Current can't be migrated, useless to reschedule,
952 * let's hope p can move out.
954 if (rq->curr->nr_cpus_allowed == 1 ||
955 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
959 * p is migratable, so let's not schedule it and
960 * see if it is pushed or pulled somewhere else.
962 if (p->nr_cpus_allowed != 1 &&
963 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
966 resched_task(rq->curr);
969 static int pull_dl_task(struct rq *this_rq);
971 #endif /* CONFIG_SMP */
974 * Only called when both the current and waking task are -deadline
977 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
980 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
981 resched_task(rq->curr);
987 * In the unlikely case current and p have the same deadline
988 * let us try to decide what's the best thing to do...
990 if ((p->dl.deadline == rq->curr->dl.deadline) &&
991 !test_tsk_need_resched(rq->curr))
992 check_preempt_equal_dl(rq, p);
993 #endif /* CONFIG_SMP */
996 #ifdef CONFIG_SCHED_HRTICK
997 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
999 s64 delta = p->dl.dl_runtime - p->dl.runtime;
1002 hrtick_start(rq, p->dl.runtime);
1006 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1007 struct dl_rq *dl_rq)
1009 struct rb_node *left = dl_rq->rb_leftmost;
1014 return rb_entry(left, struct sched_dl_entity, rb_node);
1017 struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev)
1019 struct sched_dl_entity *dl_se;
1020 struct task_struct *p;
1021 struct dl_rq *dl_rq;
1025 if (need_pull_dl_task(rq, prev)) {
1028 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1029 * means a stop task can slip in, in which case we need to
1030 * re-start task selection.
1032 if (rq->stop && rq->stop->on_rq)
1037 * When prev is DL, we may throttle it in put_prev_task().
1038 * So, we update time before we check for dl_nr_running.
1040 if (prev->sched_class == &dl_sched_class)
1043 if (unlikely(!dl_rq->dl_nr_running))
1046 put_prev_task(rq, prev);
1048 dl_se = pick_next_dl_entity(rq, dl_rq);
1051 p = dl_task_of(dl_se);
1052 p->se.exec_start = rq_clock_task(rq);
1054 /* Running task will never be pushed. */
1055 dequeue_pushable_dl_task(rq, p);
1057 #ifdef CONFIG_SCHED_HRTICK
1058 if (hrtick_enabled(rq))
1059 start_hrtick_dl(rq, p);
1062 set_post_schedule(rq);
1067 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1071 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1072 enqueue_pushable_dl_task(rq, p);
1075 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1079 #ifdef CONFIG_SCHED_HRTICK
1080 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0)
1081 start_hrtick_dl(rq, p);
1085 static void task_fork_dl(struct task_struct *p)
1088 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1093 static void task_dead_dl(struct task_struct *p)
1095 struct hrtimer *timer = &p->dl.dl_timer;
1096 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1099 * Since we are TASK_DEAD we won't slip out of the domain!
1101 raw_spin_lock_irq(&dl_b->lock);
1102 dl_b->total_bw -= p->dl.dl_bw;
1103 raw_spin_unlock_irq(&dl_b->lock);
1105 hrtimer_cancel(timer);
1108 static void set_curr_task_dl(struct rq *rq)
1110 struct task_struct *p = rq->curr;
1112 p->se.exec_start = rq_clock_task(rq);
1114 /* You can't push away the running task */
1115 dequeue_pushable_dl_task(rq, p);
1120 /* Only try algorithms three times */
1121 #define DL_MAX_TRIES 3
1123 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1125 if (!task_running(rq, p) &&
1126 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1127 (p->nr_cpus_allowed > 1))
1133 /* Returns the second earliest -deadline task, NULL otherwise */
1134 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu)
1136 struct rb_node *next_node = rq->dl.rb_leftmost;
1137 struct sched_dl_entity *dl_se;
1138 struct task_struct *p = NULL;
1141 next_node = rb_next(next_node);
1143 dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
1144 p = dl_task_of(dl_se);
1146 if (pick_dl_task(rq, p, cpu))
1155 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1157 static int find_later_rq(struct task_struct *task)
1159 struct sched_domain *sd;
1160 struct cpumask *later_mask = __get_cpu_var(local_cpu_mask_dl);
1161 int this_cpu = smp_processor_id();
1162 int best_cpu, cpu = task_cpu(task);
1164 /* Make sure the mask is initialized first */
1165 if (unlikely(!later_mask))
1168 if (task->nr_cpus_allowed == 1)
1171 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
1177 * If we are here, some target has been found,
1178 * the most suitable of which is cached in best_cpu.
1179 * This is, among the runqueues where the current tasks
1180 * have later deadlines than the task's one, the rq
1181 * with the latest possible one.
1183 * Now we check how well this matches with task's
1184 * affinity and system topology.
1186 * The last cpu where the task run is our first
1187 * guess, since it is most likely cache-hot there.
1189 if (cpumask_test_cpu(cpu, later_mask))
1192 * Check if this_cpu is to be skipped (i.e., it is
1193 * not in the mask) or not.
1195 if (!cpumask_test_cpu(this_cpu, later_mask))
1199 for_each_domain(cpu, sd) {
1200 if (sd->flags & SD_WAKE_AFFINE) {
1203 * If possible, preempting this_cpu is
1204 * cheaper than migrating.
1206 if (this_cpu != -1 &&
1207 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1213 * Last chance: if best_cpu is valid and is
1214 * in the mask, that becomes our choice.
1216 if (best_cpu < nr_cpu_ids &&
1217 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
1226 * At this point, all our guesses failed, we just return
1227 * 'something', and let the caller sort the things out.
1232 cpu = cpumask_any(later_mask);
1233 if (cpu < nr_cpu_ids)
1239 /* Locks the rq it finds */
1240 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1242 struct rq *later_rq = NULL;
1246 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1247 cpu = find_later_rq(task);
1249 if ((cpu == -1) || (cpu == rq->cpu))
1252 later_rq = cpu_rq(cpu);
1254 /* Retry if something changed. */
1255 if (double_lock_balance(rq, later_rq)) {
1256 if (unlikely(task_rq(task) != rq ||
1257 !cpumask_test_cpu(later_rq->cpu,
1258 &task->cpus_allowed) ||
1259 task_running(rq, task) || !task->on_rq)) {
1260 double_unlock_balance(rq, later_rq);
1267 * If the rq we found has no -deadline task, or
1268 * its earliest one has a later deadline than our
1269 * task, the rq is a good one.
1271 if (!later_rq->dl.dl_nr_running ||
1272 dl_time_before(task->dl.deadline,
1273 later_rq->dl.earliest_dl.curr))
1276 /* Otherwise we try again. */
1277 double_unlock_balance(rq, later_rq);
1284 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1286 struct task_struct *p;
1288 if (!has_pushable_dl_tasks(rq))
1291 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
1292 struct task_struct, pushable_dl_tasks);
1294 BUG_ON(rq->cpu != task_cpu(p));
1295 BUG_ON(task_current(rq, p));
1296 BUG_ON(p->nr_cpus_allowed <= 1);
1299 BUG_ON(!dl_task(p));
1305 * See if the non running -deadline tasks on this rq
1306 * can be sent to some other CPU where they can preempt
1307 * and start executing.
1309 static int push_dl_task(struct rq *rq)
1311 struct task_struct *next_task;
1312 struct rq *later_rq;
1314 if (!rq->dl.overloaded)
1317 next_task = pick_next_pushable_dl_task(rq);
1322 if (unlikely(next_task == rq->curr)) {
1328 * If next_task preempts rq->curr, and rq->curr
1329 * can move away, it makes sense to just reschedule
1330 * without going further in pushing next_task.
1332 if (dl_task(rq->curr) &&
1333 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1334 rq->curr->nr_cpus_allowed > 1) {
1335 resched_task(rq->curr);
1339 /* We might release rq lock */
1340 get_task_struct(next_task);
1342 /* Will lock the rq it'll find */
1343 later_rq = find_lock_later_rq(next_task, rq);
1345 struct task_struct *task;
1348 * We must check all this again, since
1349 * find_lock_later_rq releases rq->lock and it is
1350 * then possible that next_task has migrated.
1352 task = pick_next_pushable_dl_task(rq);
1353 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1355 * The task is still there. We don't try
1356 * again, some other cpu will pull it when ready.
1358 dequeue_pushable_dl_task(rq, next_task);
1366 put_task_struct(next_task);
1371 deactivate_task(rq, next_task, 0);
1372 set_task_cpu(next_task, later_rq->cpu);
1373 activate_task(later_rq, next_task, 0);
1375 resched_task(later_rq->curr);
1377 double_unlock_balance(rq, later_rq);
1380 put_task_struct(next_task);
1385 static void push_dl_tasks(struct rq *rq)
1387 /* Terminates as it moves a -deadline task */
1388 while (push_dl_task(rq))
1392 static int pull_dl_task(struct rq *this_rq)
1394 int this_cpu = this_rq->cpu, ret = 0, cpu;
1395 struct task_struct *p;
1397 u64 dmin = LONG_MAX;
1399 if (likely(!dl_overloaded(this_rq)))
1403 * Match the barrier from dl_set_overloaded; this guarantees that if we
1404 * see overloaded we must also see the dlo_mask bit.
1408 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
1409 if (this_cpu == cpu)
1412 src_rq = cpu_rq(cpu);
1415 * It looks racy, abd it is! However, as in sched_rt.c,
1416 * we are fine with this.
1418 if (this_rq->dl.dl_nr_running &&
1419 dl_time_before(this_rq->dl.earliest_dl.curr,
1420 src_rq->dl.earliest_dl.next))
1423 /* Might drop this_rq->lock */
1424 double_lock_balance(this_rq, src_rq);
1427 * If there are no more pullable tasks on the
1428 * rq, we're done with it.
1430 if (src_rq->dl.dl_nr_running <= 1)
1433 p = pick_next_earliest_dl_task(src_rq, this_cpu);
1436 * We found a task to be pulled if:
1437 * - it preempts our current (if there's one),
1438 * - it will preempt the last one we pulled (if any).
1440 if (p && dl_time_before(p->dl.deadline, dmin) &&
1441 (!this_rq->dl.dl_nr_running ||
1442 dl_time_before(p->dl.deadline,
1443 this_rq->dl.earliest_dl.curr))) {
1444 WARN_ON(p == src_rq->curr);
1448 * Then we pull iff p has actually an earlier
1449 * deadline than the current task of its runqueue.
1451 if (dl_time_before(p->dl.deadline,
1452 src_rq->curr->dl.deadline))
1457 deactivate_task(src_rq, p, 0);
1458 set_task_cpu(p, this_cpu);
1459 activate_task(this_rq, p, 0);
1460 dmin = p->dl.deadline;
1462 /* Is there any other task even earlier? */
1465 double_unlock_balance(this_rq, src_rq);
1471 static void post_schedule_dl(struct rq *rq)
1477 * Since the task is not running and a reschedule is not going to happen
1478 * anytime soon on its runqueue, we try pushing it away now.
1480 static void task_woken_dl(struct rq *rq, struct task_struct *p)
1482 if (!task_running(rq, p) &&
1483 !test_tsk_need_resched(rq->curr) &&
1484 has_pushable_dl_tasks(rq) &&
1485 p->nr_cpus_allowed > 1 &&
1486 dl_task(rq->curr) &&
1487 (rq->curr->nr_cpus_allowed < 2 ||
1488 dl_entity_preempt(&rq->curr->dl, &p->dl))) {
1493 static void set_cpus_allowed_dl(struct task_struct *p,
1494 const struct cpumask *new_mask)
1499 BUG_ON(!dl_task(p));
1502 * Update only if the task is actually running (i.e.,
1503 * it is on the rq AND it is not throttled).
1505 if (!on_dl_rq(&p->dl))
1508 weight = cpumask_weight(new_mask);
1511 * Only update if the process changes its state from whether it
1512 * can migrate or not.
1514 if ((p->nr_cpus_allowed > 1) == (weight > 1))
1520 * The process used to be able to migrate OR it can now migrate
1523 if (!task_current(rq, p))
1524 dequeue_pushable_dl_task(rq, p);
1525 BUG_ON(!rq->dl.dl_nr_migratory);
1526 rq->dl.dl_nr_migratory--;
1528 if (!task_current(rq, p))
1529 enqueue_pushable_dl_task(rq, p);
1530 rq->dl.dl_nr_migratory++;
1533 update_dl_migration(&rq->dl);
1536 /* Assumes rq->lock is held */
1537 static void rq_online_dl(struct rq *rq)
1539 if (rq->dl.overloaded)
1540 dl_set_overload(rq);
1542 if (rq->dl.dl_nr_running > 0)
1543 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1);
1546 /* Assumes rq->lock is held */
1547 static void rq_offline_dl(struct rq *rq)
1549 if (rq->dl.overloaded)
1550 dl_clear_overload(rq);
1552 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
1555 void init_sched_dl_class(void)
1559 for_each_possible_cpu(i)
1560 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
1561 GFP_KERNEL, cpu_to_node(i));
1564 #endif /* CONFIG_SMP */
1566 static void switched_from_dl(struct rq *rq, struct task_struct *p)
1568 if (hrtimer_active(&p->dl.dl_timer) && !dl_policy(p->policy))
1569 hrtimer_try_to_cancel(&p->dl.dl_timer);
1573 * Since this might be the only -deadline task on the rq,
1574 * this is the right place to try to pull some other one
1575 * from an overloaded cpu, if any.
1577 if (!rq->dl.dl_nr_running)
1583 * When switching to -deadline, we may overload the rq, then
1584 * we try to push someone off, if possible.
1586 static void switched_to_dl(struct rq *rq, struct task_struct *p)
1588 int check_resched = 1;
1591 * If p is throttled, don't consider the possibility
1592 * of preempting rq->curr, the check will be done right
1593 * after its runtime will get replenished.
1595 if (unlikely(p->dl.dl_throttled))
1598 if (p->on_rq && rq->curr != p) {
1600 if (rq->dl.overloaded && push_dl_task(rq) && rq != task_rq(p))
1601 /* Only reschedule if pushing failed */
1603 #endif /* CONFIG_SMP */
1604 if (check_resched && task_has_dl_policy(rq->curr))
1605 check_preempt_curr_dl(rq, p, 0);
1610 * If the scheduling parameters of a -deadline task changed,
1611 * a push or pull operation might be needed.
1613 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
1616 if (p->on_rq || rq->curr == p) {
1619 * This might be too much, but unfortunately
1620 * we don't have the old deadline value, and
1621 * we can't argue if the task is increasing
1622 * or lowering its prio, so...
1624 if (!rq->dl.overloaded)
1628 * If we now have a earlier deadline task than p,
1629 * then reschedule, provided p is still on this
1632 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) &&
1637 * Again, we don't know if p has a earlier
1638 * or later deadline, so let's blindly set a
1639 * (maybe not needed) rescheduling point.
1642 #endif /* CONFIG_SMP */
1644 switched_to_dl(rq, p);
1647 const struct sched_class dl_sched_class = {
1648 .next = &rt_sched_class,
1649 .enqueue_task = enqueue_task_dl,
1650 .dequeue_task = dequeue_task_dl,
1651 .yield_task = yield_task_dl,
1653 .check_preempt_curr = check_preempt_curr_dl,
1655 .pick_next_task = pick_next_task_dl,
1656 .put_prev_task = put_prev_task_dl,
1659 .select_task_rq = select_task_rq_dl,
1660 .set_cpus_allowed = set_cpus_allowed_dl,
1661 .rq_online = rq_online_dl,
1662 .rq_offline = rq_offline_dl,
1663 .post_schedule = post_schedule_dl,
1664 .task_woken = task_woken_dl,
1667 .set_curr_task = set_curr_task_dl,
1668 .task_tick = task_tick_dl,
1669 .task_fork = task_fork_dl,
1670 .task_dead = task_dead_dl,
1672 .prio_changed = prio_changed_dl,
1673 .switched_from = switched_from_dl,
1674 .switched_to = switched_to_dl,