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_dl_runtime() == RUNTIME_INF)
69 dl_b->bw = to_ratio(global_dl_period(), global_dl_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_total > 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);
138 dl_rq = &rq_of_dl_rq(dl_rq)->dl;
140 dl_rq->dl_nr_total++;
141 if (p->nr_cpus_allowed > 1)
142 dl_rq->dl_nr_migratory++;
144 update_dl_migration(dl_rq);
147 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
149 struct task_struct *p = dl_task_of(dl_se);
150 dl_rq = &rq_of_dl_rq(dl_rq)->dl;
152 dl_rq->dl_nr_total--;
153 if (p->nr_cpus_allowed > 1)
154 dl_rq->dl_nr_migratory--;
156 update_dl_migration(dl_rq);
160 * The list of pushable -deadline task is not a plist, like in
161 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
163 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
165 struct dl_rq *dl_rq = &rq->dl;
166 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node;
167 struct rb_node *parent = NULL;
168 struct task_struct *entry;
171 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
175 entry = rb_entry(parent, struct task_struct,
177 if (dl_entity_preempt(&p->dl, &entry->dl))
178 link = &parent->rb_left;
180 link = &parent->rb_right;
186 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
188 rb_link_node(&p->pushable_dl_tasks, parent, link);
189 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
192 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
194 struct dl_rq *dl_rq = &rq->dl;
196 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
199 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
200 struct rb_node *next_node;
202 next_node = rb_next(&p->pushable_dl_tasks);
203 dl_rq->pushable_dl_tasks_leftmost = next_node;
206 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
207 RB_CLEAR_NODE(&p->pushable_dl_tasks);
210 static inline int has_pushable_dl_tasks(struct rq *rq)
212 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
215 static int push_dl_task(struct rq *rq);
220 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
225 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
230 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
235 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
239 #endif /* CONFIG_SMP */
241 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
242 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
243 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
247 * We are being explicitly informed that a new instance is starting,
248 * and this means that:
249 * - the absolute deadline of the entity has to be placed at
250 * current time + relative deadline;
251 * - the runtime of the entity has to be set to the maximum value.
253 * The capability of specifying such event is useful whenever a -deadline
254 * entity wants to (try to!) synchronize its behaviour with the scheduler's
255 * one, and to (try to!) reconcile itself with its own scheduling
258 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se,
259 struct sched_dl_entity *pi_se)
261 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
262 struct rq *rq = rq_of_dl_rq(dl_rq);
264 WARN_ON(!dl_se->dl_new || dl_se->dl_throttled);
267 * We use the regular wall clock time to set deadlines in the
268 * future; in fact, we must consider execution overheads (time
269 * spent on hardirq context, etc.).
271 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
272 dl_se->runtime = pi_se->dl_runtime;
277 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
278 * possibility of a entity lasting more than what it declared, and thus
279 * exhausting its runtime.
281 * Here we are interested in making runtime overrun possible, but we do
282 * not want a entity which is misbehaving to affect the scheduling of all
284 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
285 * is used, in order to confine each entity within its own bandwidth.
287 * This function deals exactly with that, and ensures that when the runtime
288 * of a entity is replenished, its deadline is also postponed. That ensures
289 * the overrunning entity can't interfere with other entity in the system and
290 * can't make them miss their deadlines. Reasons why this kind of overruns
291 * could happen are, typically, a entity voluntarily trying to overcome its
292 * runtime, or it just underestimated it during sched_setscheduler_ex().
294 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
295 struct sched_dl_entity *pi_se)
297 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
298 struct rq *rq = rq_of_dl_rq(dl_rq);
300 BUG_ON(pi_se->dl_runtime <= 0);
303 * This could be the case for a !-dl task that is boosted.
304 * Just go with full inherited parameters.
306 if (dl_se->dl_deadline == 0) {
307 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
308 dl_se->runtime = pi_se->dl_runtime;
312 * We keep moving the deadline away until we get some
313 * available runtime for the entity. This ensures correct
314 * handling of situations where the runtime overrun is
317 while (dl_se->runtime <= 0) {
318 dl_se->deadline += pi_se->dl_period;
319 dl_se->runtime += pi_se->dl_runtime;
323 * At this point, the deadline really should be "in
324 * the future" with respect to rq->clock. If it's
325 * not, we are, for some reason, lagging too much!
326 * Anyway, after having warn userspace abut that,
327 * we still try to keep the things running by
328 * resetting the deadline and the budget of the
331 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
332 static bool lag_once = false;
336 printk_sched("sched: DL replenish lagged to much\n");
338 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
339 dl_se->runtime = pi_se->dl_runtime;
344 * Here we check if --at time t-- an entity (which is probably being
345 * [re]activated or, in general, enqueued) can use its remaining runtime
346 * and its current deadline _without_ exceeding the bandwidth it is
347 * assigned (function returns true if it can't). We are in fact applying
348 * one of the CBS rules: when a task wakes up, if the residual runtime
349 * over residual deadline fits within the allocated bandwidth, then we
350 * can keep the current (absolute) deadline and residual budget without
351 * disrupting the schedulability of the system. Otherwise, we should
352 * refill the runtime and set the deadline a period in the future,
353 * because keeping the current (absolute) deadline of the task would
354 * result in breaking guarantees promised to other tasks.
356 * This function returns true if:
358 * runtime / (deadline - t) > dl_runtime / dl_period ,
360 * IOW we can't recycle current parameters.
362 * Notice that the bandwidth check is done against the period. For
363 * task with deadline equal to period this is the same of using
364 * dl_deadline instead of dl_period in the equation above.
366 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
367 struct sched_dl_entity *pi_se, u64 t)
372 * left and right are the two sides of the equation above,
373 * after a bit of shuffling to use multiplications instead
376 * Note that none of the time values involved in the two
377 * multiplications are absolute: dl_deadline and dl_runtime
378 * are the relative deadline and the maximum runtime of each
379 * instance, runtime is the runtime left for the last instance
380 * and (deadline - t), since t is rq->clock, is the time left
381 * to the (absolute) deadline. Even if overflowing the u64 type
382 * is very unlikely to occur in both cases, here we scale down
383 * as we want to avoid that risk at all. Scaling down by 10
384 * means that we reduce granularity to 1us. We are fine with it,
385 * since this is only a true/false check and, anyway, thinking
386 * of anything below microseconds resolution is actually fiction
387 * (but still we want to give the user that illusion >;).
389 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
390 right = ((dl_se->deadline - t) >> DL_SCALE) *
391 (pi_se->dl_runtime >> DL_SCALE);
393 return dl_time_before(right, left);
397 * When a -deadline entity is queued back on the runqueue, its runtime and
398 * deadline might need updating.
400 * The policy here is that we update the deadline of the entity only if:
401 * - the current deadline is in the past,
402 * - using the remaining runtime with the current deadline would make
403 * the entity exceed its bandwidth.
405 static void update_dl_entity(struct sched_dl_entity *dl_se,
406 struct sched_dl_entity *pi_se)
408 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
409 struct rq *rq = rq_of_dl_rq(dl_rq);
412 * The arrival of a new instance needs special treatment, i.e.,
413 * the actual scheduling parameters have to be "renewed".
416 setup_new_dl_entity(dl_se, pi_se);
420 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
421 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
422 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
423 dl_se->runtime = pi_se->dl_runtime;
428 * If the entity depleted all its runtime, and if we want it to sleep
429 * while waiting for some new execution time to become available, we
430 * set the bandwidth enforcement timer to the replenishment instant
431 * and try to activate it.
433 * Notice that it is important for the caller to know if the timer
434 * actually started or not (i.e., the replenishment instant is in
435 * the future or in the past).
437 static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted)
439 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
440 struct rq *rq = rq_of_dl_rq(dl_rq);
449 * We want the timer to fire at the deadline, but considering
450 * that it is actually coming from rq->clock and not from
451 * hrtimer's time base reading.
453 act = ns_to_ktime(dl_se->deadline);
454 now = hrtimer_cb_get_time(&dl_se->dl_timer);
455 delta = ktime_to_ns(now) - rq_clock(rq);
456 act = ktime_add_ns(act, delta);
459 * If the expiry time already passed, e.g., because the value
460 * chosen as the deadline is too small, don't even try to
461 * start the timer in the past!
463 if (ktime_us_delta(act, now) < 0)
466 hrtimer_set_expires(&dl_se->dl_timer, act);
468 soft = hrtimer_get_softexpires(&dl_se->dl_timer);
469 hard = hrtimer_get_expires(&dl_se->dl_timer);
470 range = ktime_to_ns(ktime_sub(hard, soft));
471 __hrtimer_start_range_ns(&dl_se->dl_timer, soft,
472 range, HRTIMER_MODE_ABS, 0);
474 return hrtimer_active(&dl_se->dl_timer);
478 * This is the bandwidth enforcement timer callback. If here, we know
479 * a task is not on its dl_rq, since the fact that the timer was running
480 * means the task is throttled and needs a runtime replenishment.
482 * However, what we actually do depends on the fact the task is active,
483 * (it is on its rq) or has been removed from there by a call to
484 * dequeue_task_dl(). In the former case we must issue the runtime
485 * replenishment and add the task back to the dl_rq; in the latter, we just
486 * do nothing but clearing dl_throttled, so that runtime and deadline
487 * updating (and the queueing back to dl_rq) will be done by the
488 * next call to enqueue_task_dl().
490 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
492 struct sched_dl_entity *dl_se = container_of(timer,
493 struct sched_dl_entity,
495 struct task_struct *p = dl_task_of(dl_se);
496 struct rq *rq = task_rq(p);
497 raw_spin_lock(&rq->lock);
500 * We need to take care of a possible races here. In fact, the
501 * task might have changed its scheduling policy to something
502 * different from SCHED_DEADLINE or changed its reservation
503 * parameters (through sched_setscheduler()).
505 if (!dl_task(p) || dl_se->dl_new)
510 dl_se->dl_throttled = 0;
512 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
513 if (task_has_dl_policy(rq->curr))
514 check_preempt_curr_dl(rq, p, 0);
516 resched_task(rq->curr);
519 * Queueing this task back might have overloaded rq,
520 * check if we need to kick someone away.
522 if (has_pushable_dl_tasks(rq))
527 raw_spin_unlock(&rq->lock);
529 return HRTIMER_NORESTART;
532 void init_dl_task_timer(struct sched_dl_entity *dl_se)
534 struct hrtimer *timer = &dl_se->dl_timer;
536 if (hrtimer_active(timer)) {
537 hrtimer_try_to_cancel(timer);
541 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
542 timer->function = dl_task_timer;
546 int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
548 int dmiss = dl_time_before(dl_se->deadline, rq_clock(rq));
549 int rorun = dl_se->runtime <= 0;
551 if (!rorun && !dmiss)
555 * If we are beyond our current deadline and we are still
556 * executing, then we have already used some of the runtime of
557 * the next instance. Thus, if we do not account that, we are
558 * stealing bandwidth from the system at each deadline miss!
561 dl_se->runtime = rorun ? dl_se->runtime : 0;
562 dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
569 * Update the current task's runtime statistics (provided it is still
570 * a -deadline task and has not been removed from the dl_rq).
572 static void update_curr_dl(struct rq *rq)
574 struct task_struct *curr = rq->curr;
575 struct sched_dl_entity *dl_se = &curr->dl;
578 if (!dl_task(curr) || !on_dl_rq(dl_se))
582 * Consumed budget is computed considering the time as
583 * observed by schedulable tasks (excluding time spent
584 * in hardirq context, etc.). Deadlines are instead
585 * computed using hard walltime. This seems to be the more
586 * natural solution, but the full ramifications of this
587 * approach need further study.
589 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
590 if (unlikely((s64)delta_exec < 0))
593 schedstat_set(curr->se.statistics.exec_max,
594 max(curr->se.statistics.exec_max, delta_exec));
596 curr->se.sum_exec_runtime += delta_exec;
597 account_group_exec_runtime(curr, delta_exec);
599 curr->se.exec_start = rq_clock_task(rq);
600 cpuacct_charge(curr, delta_exec);
602 sched_rt_avg_update(rq, delta_exec);
604 dl_se->runtime -= delta_exec;
605 if (dl_runtime_exceeded(rq, dl_se)) {
606 __dequeue_task_dl(rq, curr, 0);
607 if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted)))
608 dl_se->dl_throttled = 1;
610 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
612 if (!is_leftmost(curr, &rq->dl))
619 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu);
621 static inline u64 next_deadline(struct rq *rq)
623 struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu);
625 if (next && dl_prio(next->prio))
626 return next->dl.deadline;
631 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
633 struct rq *rq = rq_of_dl_rq(dl_rq);
635 if (dl_rq->earliest_dl.curr == 0 ||
636 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
638 * If the dl_rq had no -deadline tasks, or if the new task
639 * has shorter deadline than the current one on dl_rq, we
640 * know that the previous earliest becomes our next earliest,
641 * as the new task becomes the earliest itself.
643 dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr;
644 dl_rq->earliest_dl.curr = deadline;
645 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1);
646 } else if (dl_rq->earliest_dl.next == 0 ||
647 dl_time_before(deadline, dl_rq->earliest_dl.next)) {
649 * On the other hand, if the new -deadline task has a
650 * a later deadline than the earliest one on dl_rq, but
651 * it is earlier than the next (if any), we must
652 * recompute the next-earliest.
654 dl_rq->earliest_dl.next = next_deadline(rq);
658 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
660 struct rq *rq = rq_of_dl_rq(dl_rq);
663 * Since we may have removed our earliest (and/or next earliest)
664 * task we must recompute them.
666 if (!dl_rq->dl_nr_running) {
667 dl_rq->earliest_dl.curr = 0;
668 dl_rq->earliest_dl.next = 0;
669 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
671 struct rb_node *leftmost = dl_rq->rb_leftmost;
672 struct sched_dl_entity *entry;
674 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
675 dl_rq->earliest_dl.curr = entry->deadline;
676 dl_rq->earliest_dl.next = next_deadline(rq);
677 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1);
683 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
684 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
686 #endif /* CONFIG_SMP */
689 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
691 int prio = dl_task_of(dl_se)->prio;
692 u64 deadline = dl_se->deadline;
694 WARN_ON(!dl_prio(prio));
695 dl_rq->dl_nr_running++;
697 inc_dl_deadline(dl_rq, deadline);
698 inc_dl_migration(dl_se, dl_rq);
702 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
704 int prio = dl_task_of(dl_se)->prio;
706 WARN_ON(!dl_prio(prio));
707 WARN_ON(!dl_rq->dl_nr_running);
708 dl_rq->dl_nr_running--;
710 dec_dl_deadline(dl_rq, dl_se->deadline);
711 dec_dl_migration(dl_se, dl_rq);
714 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
716 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
717 struct rb_node **link = &dl_rq->rb_root.rb_node;
718 struct rb_node *parent = NULL;
719 struct sched_dl_entity *entry;
722 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
726 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
727 if (dl_time_before(dl_se->deadline, entry->deadline))
728 link = &parent->rb_left;
730 link = &parent->rb_right;
736 dl_rq->rb_leftmost = &dl_se->rb_node;
738 rb_link_node(&dl_se->rb_node, parent, link);
739 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
741 inc_dl_tasks(dl_se, dl_rq);
744 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
746 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
748 if (RB_EMPTY_NODE(&dl_se->rb_node))
751 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
752 struct rb_node *next_node;
754 next_node = rb_next(&dl_se->rb_node);
755 dl_rq->rb_leftmost = next_node;
758 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
759 RB_CLEAR_NODE(&dl_se->rb_node);
761 dec_dl_tasks(dl_se, dl_rq);
765 enqueue_dl_entity(struct sched_dl_entity *dl_se,
766 struct sched_dl_entity *pi_se, int flags)
768 BUG_ON(on_dl_rq(dl_se));
771 * If this is a wakeup or a new instance, the scheduling
772 * parameters of the task might need updating. Otherwise,
773 * we want a replenishment of its runtime.
775 if (!dl_se->dl_new && flags & ENQUEUE_REPLENISH)
776 replenish_dl_entity(dl_se, pi_se);
778 update_dl_entity(dl_se, pi_se);
780 __enqueue_dl_entity(dl_se);
783 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
785 __dequeue_dl_entity(dl_se);
788 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
790 struct task_struct *pi_task = rt_mutex_get_top_task(p);
791 struct sched_dl_entity *pi_se = &p->dl;
794 * Use the scheduling parameters of the top pi-waiter
795 * task if we have one and its (relative) deadline is
796 * smaller than our one... OTW we keep our runtime and
799 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio))
800 pi_se = &pi_task->dl;
803 * If p is throttled, we do nothing. In fact, if it exhausted
804 * its budget it needs a replenishment and, since it now is on
805 * its rq, the bandwidth timer callback (which clearly has not
806 * run yet) will take care of this.
808 if (p->dl.dl_throttled)
811 enqueue_dl_entity(&p->dl, pi_se, flags);
813 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
814 enqueue_pushable_dl_task(rq, p);
819 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
821 dequeue_dl_entity(&p->dl);
822 dequeue_pushable_dl_task(rq, p);
825 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
828 __dequeue_task_dl(rq, p, flags);
834 * Yield task semantic for -deadline tasks is:
836 * get off from the CPU until our next instance, with
837 * a new runtime. This is of little use now, since we
838 * don't have a bandwidth reclaiming mechanism. Anyway,
839 * bandwidth reclaiming is planned for the future, and
840 * yield_task_dl will indicate that some spare budget
841 * is available for other task instances to use it.
843 static void yield_task_dl(struct rq *rq)
845 struct task_struct *p = rq->curr;
848 * We make the task go to sleep until its current deadline by
849 * forcing its runtime to zero. This way, update_curr_dl() stops
850 * it and the bandwidth timer will wake it up and will give it
851 * new scheduling parameters (thanks to dl_new=1).
853 if (p->dl.runtime > 0) {
854 rq->curr->dl.dl_new = 1;
862 static int find_later_rq(struct task_struct *task);
865 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
867 struct task_struct *curr;
870 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
876 curr = ACCESS_ONCE(rq->curr); /* unlocked access */
879 * If we are dealing with a -deadline task, we must
880 * decide where to wake it up.
881 * If it has a later deadline and the current task
882 * on this rq can't move (provided the waking task
883 * can!) we prefer to send it somewhere else. On the
884 * other hand, if it has a shorter deadline, we
885 * try to make it stay here, it might be important.
887 if (unlikely(dl_task(curr)) &&
888 (curr->nr_cpus_allowed < 2 ||
889 !dl_entity_preempt(&p->dl, &curr->dl)) &&
890 (p->nr_cpus_allowed > 1)) {
891 int target = find_later_rq(p);
902 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
905 * Current can't be migrated, useless to reschedule,
906 * let's hope p can move out.
908 if (rq->curr->nr_cpus_allowed == 1 ||
909 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
913 * p is migratable, so let's not schedule it and
914 * see if it is pushed or pulled somewhere else.
916 if (p->nr_cpus_allowed != 1 &&
917 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
920 resched_task(rq->curr);
923 #endif /* CONFIG_SMP */
926 * Only called when both the current and waking task are -deadline
929 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
932 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
933 resched_task(rq->curr);
939 * In the unlikely case current and p have the same deadline
940 * let us try to decide what's the best thing to do...
942 if ((p->dl.deadline == rq->curr->dl.deadline) &&
943 !test_tsk_need_resched(rq->curr))
944 check_preempt_equal_dl(rq, p);
945 #endif /* CONFIG_SMP */
948 #ifdef CONFIG_SCHED_HRTICK
949 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
951 s64 delta = p->dl.dl_runtime - p->dl.runtime;
954 hrtick_start(rq, p->dl.runtime);
958 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
961 struct rb_node *left = dl_rq->rb_leftmost;
966 return rb_entry(left, struct sched_dl_entity, rb_node);
969 struct task_struct *pick_next_task_dl(struct rq *rq)
971 struct sched_dl_entity *dl_se;
972 struct task_struct *p;
977 if (unlikely(!dl_rq->dl_nr_running))
980 dl_se = pick_next_dl_entity(rq, dl_rq);
983 p = dl_task_of(dl_se);
984 p->se.exec_start = rq_clock_task(rq);
986 /* Running task will never be pushed. */
988 dequeue_pushable_dl_task(rq, p);
990 #ifdef CONFIG_SCHED_HRTICK
991 if (hrtick_enabled(rq))
992 start_hrtick_dl(rq, p);
996 rq->post_schedule = has_pushable_dl_tasks(rq);
997 #endif /* CONFIG_SMP */
1002 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1006 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1007 enqueue_pushable_dl_task(rq, p);
1010 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1014 #ifdef CONFIG_SCHED_HRTICK
1015 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0)
1016 start_hrtick_dl(rq, p);
1020 static void task_fork_dl(struct task_struct *p)
1023 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1028 static void task_dead_dl(struct task_struct *p)
1030 struct hrtimer *timer = &p->dl.dl_timer;
1031 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1034 * Since we are TASK_DEAD we won't slip out of the domain!
1036 raw_spin_lock_irq(&dl_b->lock);
1037 dl_b->total_bw -= p->dl.dl_bw;
1038 raw_spin_unlock_irq(&dl_b->lock);
1040 hrtimer_cancel(timer);
1043 static void set_curr_task_dl(struct rq *rq)
1045 struct task_struct *p = rq->curr;
1047 p->se.exec_start = rq_clock_task(rq);
1049 /* You can't push away the running task */
1050 dequeue_pushable_dl_task(rq, p);
1055 /* Only try algorithms three times */
1056 #define DL_MAX_TRIES 3
1058 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1060 if (!task_running(rq, p) &&
1061 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1062 (p->nr_cpus_allowed > 1))
1068 /* Returns the second earliest -deadline task, NULL otherwise */
1069 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu)
1071 struct rb_node *next_node = rq->dl.rb_leftmost;
1072 struct sched_dl_entity *dl_se;
1073 struct task_struct *p = NULL;
1076 next_node = rb_next(next_node);
1078 dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
1079 p = dl_task_of(dl_se);
1081 if (pick_dl_task(rq, p, cpu))
1090 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1092 static int find_later_rq(struct task_struct *task)
1094 struct sched_domain *sd;
1095 struct cpumask *later_mask = __get_cpu_var(local_cpu_mask_dl);
1096 int this_cpu = smp_processor_id();
1097 int best_cpu, cpu = task_cpu(task);
1099 /* Make sure the mask is initialized first */
1100 if (unlikely(!later_mask))
1103 if (task->nr_cpus_allowed == 1)
1106 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
1112 * If we are here, some target has been found,
1113 * the most suitable of which is cached in best_cpu.
1114 * This is, among the runqueues where the current tasks
1115 * have later deadlines than the task's one, the rq
1116 * with the latest possible one.
1118 * Now we check how well this matches with task's
1119 * affinity and system topology.
1121 * The last cpu where the task run is our first
1122 * guess, since it is most likely cache-hot there.
1124 if (cpumask_test_cpu(cpu, later_mask))
1127 * Check if this_cpu is to be skipped (i.e., it is
1128 * not in the mask) or not.
1130 if (!cpumask_test_cpu(this_cpu, later_mask))
1134 for_each_domain(cpu, sd) {
1135 if (sd->flags & SD_WAKE_AFFINE) {
1138 * If possible, preempting this_cpu is
1139 * cheaper than migrating.
1141 if (this_cpu != -1 &&
1142 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1148 * Last chance: if best_cpu is valid and is
1149 * in the mask, that becomes our choice.
1151 if (best_cpu < nr_cpu_ids &&
1152 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
1161 * At this point, all our guesses failed, we just return
1162 * 'something', and let the caller sort the things out.
1167 cpu = cpumask_any(later_mask);
1168 if (cpu < nr_cpu_ids)
1174 /* Locks the rq it finds */
1175 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1177 struct rq *later_rq = NULL;
1181 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1182 cpu = find_later_rq(task);
1184 if ((cpu == -1) || (cpu == rq->cpu))
1187 later_rq = cpu_rq(cpu);
1189 /* Retry if something changed. */
1190 if (double_lock_balance(rq, later_rq)) {
1191 if (unlikely(task_rq(task) != rq ||
1192 !cpumask_test_cpu(later_rq->cpu,
1193 &task->cpus_allowed) ||
1194 task_running(rq, task) || !task->on_rq)) {
1195 double_unlock_balance(rq, later_rq);
1202 * If the rq we found has no -deadline task, or
1203 * its earliest one has a later deadline than our
1204 * task, the rq is a good one.
1206 if (!later_rq->dl.dl_nr_running ||
1207 dl_time_before(task->dl.deadline,
1208 later_rq->dl.earliest_dl.curr))
1211 /* Otherwise we try again. */
1212 double_unlock_balance(rq, later_rq);
1219 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1221 struct task_struct *p;
1223 if (!has_pushable_dl_tasks(rq))
1226 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
1227 struct task_struct, pushable_dl_tasks);
1229 BUG_ON(rq->cpu != task_cpu(p));
1230 BUG_ON(task_current(rq, p));
1231 BUG_ON(p->nr_cpus_allowed <= 1);
1234 BUG_ON(!dl_task(p));
1240 * See if the non running -deadline tasks on this rq
1241 * can be sent to some other CPU where they can preempt
1242 * and start executing.
1244 static int push_dl_task(struct rq *rq)
1246 struct task_struct *next_task;
1247 struct rq *later_rq;
1249 if (!rq->dl.overloaded)
1252 next_task = pick_next_pushable_dl_task(rq);
1257 if (unlikely(next_task == rq->curr)) {
1263 * If next_task preempts rq->curr, and rq->curr
1264 * can move away, it makes sense to just reschedule
1265 * without going further in pushing next_task.
1267 if (dl_task(rq->curr) &&
1268 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1269 rq->curr->nr_cpus_allowed > 1) {
1270 resched_task(rq->curr);
1274 /* We might release rq lock */
1275 get_task_struct(next_task);
1277 /* Will lock the rq it'll find */
1278 later_rq = find_lock_later_rq(next_task, rq);
1280 struct task_struct *task;
1283 * We must check all this again, since
1284 * find_lock_later_rq releases rq->lock and it is
1285 * then possible that next_task has migrated.
1287 task = pick_next_pushable_dl_task(rq);
1288 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1290 * The task is still there. We don't try
1291 * again, some other cpu will pull it when ready.
1293 dequeue_pushable_dl_task(rq, next_task);
1301 put_task_struct(next_task);
1306 deactivate_task(rq, next_task, 0);
1307 set_task_cpu(next_task, later_rq->cpu);
1308 activate_task(later_rq, next_task, 0);
1310 resched_task(later_rq->curr);
1312 double_unlock_balance(rq, later_rq);
1315 put_task_struct(next_task);
1320 static void push_dl_tasks(struct rq *rq)
1322 /* Terminates as it moves a -deadline task */
1323 while (push_dl_task(rq))
1327 static int pull_dl_task(struct rq *this_rq)
1329 int this_cpu = this_rq->cpu, ret = 0, cpu;
1330 struct task_struct *p;
1332 u64 dmin = LONG_MAX;
1334 if (likely(!dl_overloaded(this_rq)))
1338 * Match the barrier from dl_set_overloaded; this guarantees that if we
1339 * see overloaded we must also see the dlo_mask bit.
1343 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
1344 if (this_cpu == cpu)
1347 src_rq = cpu_rq(cpu);
1350 * It looks racy, abd it is! However, as in sched_rt.c,
1351 * we are fine with this.
1353 if (this_rq->dl.dl_nr_running &&
1354 dl_time_before(this_rq->dl.earliest_dl.curr,
1355 src_rq->dl.earliest_dl.next))
1358 /* Might drop this_rq->lock */
1359 double_lock_balance(this_rq, src_rq);
1362 * If there are no more pullable tasks on the
1363 * rq, we're done with it.
1365 if (src_rq->dl.dl_nr_running <= 1)
1368 p = pick_next_earliest_dl_task(src_rq, this_cpu);
1371 * We found a task to be pulled if:
1372 * - it preempts our current (if there's one),
1373 * - it will preempt the last one we pulled (if any).
1375 if (p && dl_time_before(p->dl.deadline, dmin) &&
1376 (!this_rq->dl.dl_nr_running ||
1377 dl_time_before(p->dl.deadline,
1378 this_rq->dl.earliest_dl.curr))) {
1379 WARN_ON(p == src_rq->curr);
1383 * Then we pull iff p has actually an earlier
1384 * deadline than the current task of its runqueue.
1386 if (dl_time_before(p->dl.deadline,
1387 src_rq->curr->dl.deadline))
1392 deactivate_task(src_rq, p, 0);
1393 set_task_cpu(p, this_cpu);
1394 activate_task(this_rq, p, 0);
1395 dmin = p->dl.deadline;
1397 /* Is there any other task even earlier? */
1400 double_unlock_balance(this_rq, src_rq);
1406 static void pre_schedule_dl(struct rq *rq, struct task_struct *prev)
1408 /* Try to pull other tasks here */
1413 static void post_schedule_dl(struct rq *rq)
1419 * Since the task is not running and a reschedule is not going to happen
1420 * anytime soon on its runqueue, we try pushing it away now.
1422 static void task_woken_dl(struct rq *rq, struct task_struct *p)
1424 if (!task_running(rq, p) &&
1425 !test_tsk_need_resched(rq->curr) &&
1426 has_pushable_dl_tasks(rq) &&
1427 p->nr_cpus_allowed > 1 &&
1428 dl_task(rq->curr) &&
1429 (rq->curr->nr_cpus_allowed < 2 ||
1430 dl_entity_preempt(&rq->curr->dl, &p->dl))) {
1435 static void set_cpus_allowed_dl(struct task_struct *p,
1436 const struct cpumask *new_mask)
1441 BUG_ON(!dl_task(p));
1444 * Update only if the task is actually running (i.e.,
1445 * it is on the rq AND it is not throttled).
1447 if (!on_dl_rq(&p->dl))
1450 weight = cpumask_weight(new_mask);
1453 * Only update if the process changes its state from whether it
1454 * can migrate or not.
1456 if ((p->nr_cpus_allowed > 1) == (weight > 1))
1462 * The process used to be able to migrate OR it can now migrate
1465 if (!task_current(rq, p))
1466 dequeue_pushable_dl_task(rq, p);
1467 BUG_ON(!rq->dl.dl_nr_migratory);
1468 rq->dl.dl_nr_migratory--;
1470 if (!task_current(rq, p))
1471 enqueue_pushable_dl_task(rq, p);
1472 rq->dl.dl_nr_migratory++;
1475 update_dl_migration(&rq->dl);
1478 /* Assumes rq->lock is held */
1479 static void rq_online_dl(struct rq *rq)
1481 if (rq->dl.overloaded)
1482 dl_set_overload(rq);
1484 if (rq->dl.dl_nr_running > 0)
1485 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1);
1488 /* Assumes rq->lock is held */
1489 static void rq_offline_dl(struct rq *rq)
1491 if (rq->dl.overloaded)
1492 dl_clear_overload(rq);
1494 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
1497 void init_sched_dl_class(void)
1501 for_each_possible_cpu(i)
1502 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
1503 GFP_KERNEL, cpu_to_node(i));
1506 #endif /* CONFIG_SMP */
1508 static void switched_from_dl(struct rq *rq, struct task_struct *p)
1510 if (hrtimer_active(&p->dl.dl_timer) && !dl_policy(p->policy))
1511 hrtimer_try_to_cancel(&p->dl.dl_timer);
1515 * Since this might be the only -deadline task on the rq,
1516 * this is the right place to try to pull some other one
1517 * from an overloaded cpu, if any.
1519 if (!rq->dl.dl_nr_running)
1525 * When switching to -deadline, we may overload the rq, then
1526 * we try to push someone off, if possible.
1528 static void switched_to_dl(struct rq *rq, struct task_struct *p)
1530 int check_resched = 1;
1533 * If p is throttled, don't consider the possibility
1534 * of preempting rq->curr, the check will be done right
1535 * after its runtime will get replenished.
1537 if (unlikely(p->dl.dl_throttled))
1540 if (p->on_rq || rq->curr != p) {
1542 if (rq->dl.overloaded && push_dl_task(rq) && rq != task_rq(p))
1543 /* Only reschedule if pushing failed */
1545 #endif /* CONFIG_SMP */
1546 if (check_resched && task_has_dl_policy(rq->curr))
1547 check_preempt_curr_dl(rq, p, 0);
1552 * If the scheduling parameters of a -deadline task changed,
1553 * a push or pull operation might be needed.
1555 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
1558 if (p->on_rq || rq->curr == p) {
1561 * This might be too much, but unfortunately
1562 * we don't have the old deadline value, and
1563 * we can't argue if the task is increasing
1564 * or lowering its prio, so...
1566 if (!rq->dl.overloaded)
1570 * If we now have a earlier deadline task than p,
1571 * then reschedule, provided p is still on this
1574 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) &&
1579 * Again, we don't know if p has a earlier
1580 * or later deadline, so let's blindly set a
1581 * (maybe not needed) rescheduling point.
1584 #endif /* CONFIG_SMP */
1586 switched_to_dl(rq, p);
1589 const struct sched_class dl_sched_class = {
1590 .next = &rt_sched_class,
1591 .enqueue_task = enqueue_task_dl,
1592 .dequeue_task = dequeue_task_dl,
1593 .yield_task = yield_task_dl,
1595 .check_preempt_curr = check_preempt_curr_dl,
1597 .pick_next_task = pick_next_task_dl,
1598 .put_prev_task = put_prev_task_dl,
1601 .select_task_rq = select_task_rq_dl,
1602 .set_cpus_allowed = set_cpus_allowed_dl,
1603 .rq_online = rq_online_dl,
1604 .rq_offline = rq_offline_dl,
1605 .pre_schedule = pre_schedule_dl,
1606 .post_schedule = post_schedule_dl,
1607 .task_woken = task_woken_dl,
1610 .set_curr_task = set_curr_task_dl,
1611 .task_tick = task_tick_dl,
1612 .task_fork = task_fork_dl,
1613 .task_dead = task_dead_dl,
1615 .prio_changed = prio_changed_dl,
1616 .switched_from = switched_from_dl,
1617 .switched_to = switched_to_dl,