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.
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);
216 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
221 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
226 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
231 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
235 #endif /* CONFIG_SMP */
237 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
238 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
239 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
243 * We are being explicitly informed that a new instance is starting,
244 * and this means that:
245 * - the absolute deadline of the entity has to be placed at
246 * current time + relative deadline;
247 * - the runtime of the entity has to be set to the maximum value.
249 * The capability of specifying such event is useful whenever a -deadline
250 * entity wants to (try to!) synchronize its behaviour with the scheduler's
251 * one, and to (try to!) reconcile itself with its own scheduling
254 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se,
255 struct sched_dl_entity *pi_se)
257 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
258 struct rq *rq = rq_of_dl_rq(dl_rq);
260 WARN_ON(!dl_se->dl_new || dl_se->dl_throttled);
263 * We use the regular wall clock time to set deadlines in the
264 * future; in fact, we must consider execution overheads (time
265 * spent on hardirq context, etc.).
267 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
268 dl_se->runtime = pi_se->dl_runtime;
273 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
274 * possibility of a entity lasting more than what it declared, and thus
275 * exhausting its runtime.
277 * Here we are interested in making runtime overrun possible, but we do
278 * not want a entity which is misbehaving to affect the scheduling of all
280 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
281 * is used, in order to confine each entity within its own bandwidth.
283 * This function deals exactly with that, and ensures that when the runtime
284 * of a entity is replenished, its deadline is also postponed. That ensures
285 * the overrunning entity can't interfere with other entity in the system and
286 * can't make them miss their deadlines. Reasons why this kind of overruns
287 * could happen are, typically, a entity voluntarily trying to overcome its
288 * runtime, or it just underestimated it during sched_setscheduler_ex().
290 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
291 struct sched_dl_entity *pi_se)
293 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
294 struct rq *rq = rq_of_dl_rq(dl_rq);
296 BUG_ON(pi_se->dl_runtime <= 0);
299 * This could be the case for a !-dl task that is boosted.
300 * Just go with full inherited parameters.
302 if (dl_se->dl_deadline == 0) {
303 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
304 dl_se->runtime = pi_se->dl_runtime;
308 * We keep moving the deadline away until we get some
309 * available runtime for the entity. This ensures correct
310 * handling of situations where the runtime overrun is
313 while (dl_se->runtime <= 0) {
314 dl_se->deadline += pi_se->dl_period;
315 dl_se->runtime += pi_se->dl_runtime;
319 * At this point, the deadline really should be "in
320 * the future" with respect to rq->clock. If it's
321 * not, we are, for some reason, lagging too much!
322 * Anyway, after having warn userspace abut that,
323 * we still try to keep the things running by
324 * resetting the deadline and the budget of the
327 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
328 static bool lag_once = false;
332 printk_sched("sched: DL replenish lagged to much\n");
334 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
335 dl_se->runtime = pi_se->dl_runtime;
340 * Here we check if --at time t-- an entity (which is probably being
341 * [re]activated or, in general, enqueued) can use its remaining runtime
342 * and its current deadline _without_ exceeding the bandwidth it is
343 * assigned (function returns true if it can't). We are in fact applying
344 * one of the CBS rules: when a task wakes up, if the residual runtime
345 * over residual deadline fits within the allocated bandwidth, then we
346 * can keep the current (absolute) deadline and residual budget without
347 * disrupting the schedulability of the system. Otherwise, we should
348 * refill the runtime and set the deadline a period in the future,
349 * because keeping the current (absolute) deadline of the task would
350 * result in breaking guarantees promised to other tasks (refer to
351 * Documentation/scheduler/sched-deadline.txt for more informations).
353 * This function returns true if:
355 * runtime / (deadline - t) > dl_runtime / dl_period ,
357 * IOW we can't recycle current parameters.
359 * Notice that the bandwidth check is done against the period. For
360 * task with deadline equal to period this is the same of using
361 * dl_deadline instead of dl_period in the equation above.
363 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
364 struct sched_dl_entity *pi_se, u64 t)
369 * left and right are the two sides of the equation above,
370 * after a bit of shuffling to use multiplications instead
373 * Note that none of the time values involved in the two
374 * multiplications are absolute: dl_deadline and dl_runtime
375 * are the relative deadline and the maximum runtime of each
376 * instance, runtime is the runtime left for the last instance
377 * and (deadline - t), since t is rq->clock, is the time left
378 * to the (absolute) deadline. Even if overflowing the u64 type
379 * is very unlikely to occur in both cases, here we scale down
380 * as we want to avoid that risk at all. Scaling down by 10
381 * means that we reduce granularity to 1us. We are fine with it,
382 * since this is only a true/false check and, anyway, thinking
383 * of anything below microseconds resolution is actually fiction
384 * (but still we want to give the user that illusion >;).
386 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
387 right = ((dl_se->deadline - t) >> DL_SCALE) *
388 (pi_se->dl_runtime >> DL_SCALE);
390 return dl_time_before(right, left);
394 * When a -deadline entity is queued back on the runqueue, its runtime and
395 * deadline might need updating.
397 * The policy here is that we update the deadline of the entity only if:
398 * - the current deadline is in the past,
399 * - using the remaining runtime with the current deadline would make
400 * the entity exceed its bandwidth.
402 static void update_dl_entity(struct sched_dl_entity *dl_se,
403 struct sched_dl_entity *pi_se)
405 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
406 struct rq *rq = rq_of_dl_rq(dl_rq);
409 * The arrival of a new instance needs special treatment, i.e.,
410 * the actual scheduling parameters have to be "renewed".
413 setup_new_dl_entity(dl_se, pi_se);
417 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
418 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
419 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
420 dl_se->runtime = pi_se->dl_runtime;
425 * If the entity depleted all its runtime, and if we want it to sleep
426 * while waiting for some new execution time to become available, we
427 * set the bandwidth enforcement timer to the replenishment instant
428 * and try to activate it.
430 * Notice that it is important for the caller to know if the timer
431 * actually started or not (i.e., the replenishment instant is in
432 * the future or in the past).
434 static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted)
436 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
437 struct rq *rq = rq_of_dl_rq(dl_rq);
446 * We want the timer to fire at the deadline, but considering
447 * that it is actually coming from rq->clock and not from
448 * hrtimer's time base reading.
450 act = ns_to_ktime(dl_se->deadline);
451 now = hrtimer_cb_get_time(&dl_se->dl_timer);
452 delta = ktime_to_ns(now) - rq_clock(rq);
453 act = ktime_add_ns(act, delta);
456 * If the expiry time already passed, e.g., because the value
457 * chosen as the deadline is too small, don't even try to
458 * start the timer in the past!
460 if (ktime_us_delta(act, now) < 0)
463 hrtimer_set_expires(&dl_se->dl_timer, act);
465 soft = hrtimer_get_softexpires(&dl_se->dl_timer);
466 hard = hrtimer_get_expires(&dl_se->dl_timer);
467 range = ktime_to_ns(ktime_sub(hard, soft));
468 __hrtimer_start_range_ns(&dl_se->dl_timer, soft,
469 range, HRTIMER_MODE_ABS, 0);
471 return hrtimer_active(&dl_se->dl_timer);
475 * This is the bandwidth enforcement timer callback. If here, we know
476 * a task is not on its dl_rq, since the fact that the timer was running
477 * means the task is throttled and needs a runtime replenishment.
479 * However, what we actually do depends on the fact the task is active,
480 * (it is on its rq) or has been removed from there by a call to
481 * dequeue_task_dl(). In the former case we must issue the runtime
482 * replenishment and add the task back to the dl_rq; in the latter, we just
483 * do nothing but clearing dl_throttled, so that runtime and deadline
484 * updating (and the queueing back to dl_rq) will be done by the
485 * next call to enqueue_task_dl().
487 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
489 struct sched_dl_entity *dl_se = container_of(timer,
490 struct sched_dl_entity,
492 struct task_struct *p = dl_task_of(dl_se);
493 struct rq *rq = task_rq(p);
494 raw_spin_lock(&rq->lock);
497 * We need to take care of a possible races here. In fact, the
498 * task might have changed its scheduling policy to something
499 * different from SCHED_DEADLINE or changed its reservation
500 * parameters (through sched_setscheduler()).
502 if (!dl_task(p) || dl_se->dl_new)
507 dl_se->dl_throttled = 0;
509 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
510 if (task_has_dl_policy(rq->curr))
511 check_preempt_curr_dl(rq, p, 0);
513 resched_task(rq->curr);
516 * Queueing this task back might have overloaded rq,
517 * check if we need to kick someone away.
519 if (has_pushable_dl_tasks(rq))
524 raw_spin_unlock(&rq->lock);
526 return HRTIMER_NORESTART;
529 void init_dl_task_timer(struct sched_dl_entity *dl_se)
531 struct hrtimer *timer = &dl_se->dl_timer;
533 if (hrtimer_active(timer)) {
534 hrtimer_try_to_cancel(timer);
538 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
539 timer->function = dl_task_timer;
543 int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
545 int dmiss = dl_time_before(dl_se->deadline, rq_clock(rq));
546 int rorun = dl_se->runtime <= 0;
548 if (!rorun && !dmiss)
552 * If we are beyond our current deadline and we are still
553 * executing, then we have already used some of the runtime of
554 * the next instance. Thus, if we do not account that, we are
555 * stealing bandwidth from the system at each deadline miss!
558 dl_se->runtime = rorun ? dl_se->runtime : 0;
559 dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
565 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
568 * Update the current task's runtime statistics (provided it is still
569 * a -deadline task and has not been removed from the dl_rq).
571 static void update_curr_dl(struct rq *rq)
573 struct task_struct *curr = rq->curr;
574 struct sched_dl_entity *dl_se = &curr->dl;
577 if (!dl_task(curr) || !on_dl_rq(dl_se))
581 * Consumed budget is computed considering the time as
582 * observed by schedulable tasks (excluding time spent
583 * in hardirq context, etc.). Deadlines are instead
584 * computed using hard walltime. This seems to be the more
585 * natural solution, but the full ramifications of this
586 * approach need further study.
588 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
589 if (unlikely((s64)delta_exec < 0))
592 schedstat_set(curr->se.statistics.exec_max,
593 max(curr->se.statistics.exec_max, delta_exec));
595 curr->se.sum_exec_runtime += delta_exec;
596 account_group_exec_runtime(curr, delta_exec);
598 curr->se.exec_start = rq_clock_task(rq);
599 cpuacct_charge(curr, delta_exec);
601 sched_rt_avg_update(rq, delta_exec);
603 dl_se->runtime -= delta_exec;
604 if (dl_runtime_exceeded(rq, dl_se)) {
605 __dequeue_task_dl(rq, curr, 0);
606 if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted)))
607 dl_se->dl_throttled = 1;
609 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
611 if (!is_leftmost(curr, &rq->dl))
616 * Because -- for now -- we share the rt bandwidth, we need to
617 * account our runtime there too, otherwise actual rt tasks
618 * would be able to exceed the shared quota.
620 * Account to the root rt group for now.
622 * The solution we're working towards is having the RT groups scheduled
623 * using deadline servers -- however there's a few nasties to figure
624 * out before that can happen.
626 if (rt_bandwidth_enabled()) {
627 struct rt_rq *rt_rq = &rq->rt;
629 raw_spin_lock(&rt_rq->rt_runtime_lock);
631 * We'll let actual RT tasks worry about the overflow here, we
632 * have our own CBS to keep us inline; only account when RT
633 * bandwidth is relevant.
635 if (sched_rt_bandwidth_account(rt_rq))
636 rt_rq->rt_time += delta_exec;
637 raw_spin_unlock(&rt_rq->rt_runtime_lock);
643 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu);
645 static inline u64 next_deadline(struct rq *rq)
647 struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu);
649 if (next && dl_prio(next->prio))
650 return next->dl.deadline;
655 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
657 struct rq *rq = rq_of_dl_rq(dl_rq);
659 if (dl_rq->earliest_dl.curr == 0 ||
660 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
662 * If the dl_rq had no -deadline tasks, or if the new task
663 * has shorter deadline than the current one on dl_rq, we
664 * know that the previous earliest becomes our next earliest,
665 * as the new task becomes the earliest itself.
667 dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr;
668 dl_rq->earliest_dl.curr = deadline;
669 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1);
670 } else if (dl_rq->earliest_dl.next == 0 ||
671 dl_time_before(deadline, dl_rq->earliest_dl.next)) {
673 * On the other hand, if the new -deadline task has a
674 * a later deadline than the earliest one on dl_rq, but
675 * it is earlier than the next (if any), we must
676 * recompute the next-earliest.
678 dl_rq->earliest_dl.next = next_deadline(rq);
682 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
684 struct rq *rq = rq_of_dl_rq(dl_rq);
687 * Since we may have removed our earliest (and/or next earliest)
688 * task we must recompute them.
690 if (!dl_rq->dl_nr_running) {
691 dl_rq->earliest_dl.curr = 0;
692 dl_rq->earliest_dl.next = 0;
693 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
695 struct rb_node *leftmost = dl_rq->rb_leftmost;
696 struct sched_dl_entity *entry;
698 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
699 dl_rq->earliest_dl.curr = entry->deadline;
700 dl_rq->earliest_dl.next = next_deadline(rq);
701 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1);
707 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
708 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
710 #endif /* CONFIG_SMP */
713 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
715 int prio = dl_task_of(dl_se)->prio;
716 u64 deadline = dl_se->deadline;
718 WARN_ON(!dl_prio(prio));
719 dl_rq->dl_nr_running++;
720 inc_nr_running(rq_of_dl_rq(dl_rq));
722 inc_dl_deadline(dl_rq, deadline);
723 inc_dl_migration(dl_se, dl_rq);
727 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
729 int prio = dl_task_of(dl_se)->prio;
731 WARN_ON(!dl_prio(prio));
732 WARN_ON(!dl_rq->dl_nr_running);
733 dl_rq->dl_nr_running--;
734 dec_nr_running(rq_of_dl_rq(dl_rq));
736 dec_dl_deadline(dl_rq, dl_se->deadline);
737 dec_dl_migration(dl_se, dl_rq);
740 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
742 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
743 struct rb_node **link = &dl_rq->rb_root.rb_node;
744 struct rb_node *parent = NULL;
745 struct sched_dl_entity *entry;
748 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
752 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
753 if (dl_time_before(dl_se->deadline, entry->deadline))
754 link = &parent->rb_left;
756 link = &parent->rb_right;
762 dl_rq->rb_leftmost = &dl_se->rb_node;
764 rb_link_node(&dl_se->rb_node, parent, link);
765 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
767 inc_dl_tasks(dl_se, dl_rq);
770 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
772 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
774 if (RB_EMPTY_NODE(&dl_se->rb_node))
777 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
778 struct rb_node *next_node;
780 next_node = rb_next(&dl_se->rb_node);
781 dl_rq->rb_leftmost = next_node;
784 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
785 RB_CLEAR_NODE(&dl_se->rb_node);
787 dec_dl_tasks(dl_se, dl_rq);
791 enqueue_dl_entity(struct sched_dl_entity *dl_se,
792 struct sched_dl_entity *pi_se, int flags)
794 BUG_ON(on_dl_rq(dl_se));
797 * If this is a wakeup or a new instance, the scheduling
798 * parameters of the task might need updating. Otherwise,
799 * we want a replenishment of its runtime.
801 if (!dl_se->dl_new && flags & ENQUEUE_REPLENISH)
802 replenish_dl_entity(dl_se, pi_se);
804 update_dl_entity(dl_se, pi_se);
806 __enqueue_dl_entity(dl_se);
809 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
811 __dequeue_dl_entity(dl_se);
814 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
816 struct task_struct *pi_task = rt_mutex_get_top_task(p);
817 struct sched_dl_entity *pi_se = &p->dl;
820 * Use the scheduling parameters of the top pi-waiter
821 * task if we have one and its (relative) deadline is
822 * smaller than our one... OTW we keep our runtime and
825 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio))
826 pi_se = &pi_task->dl;
829 * If p is throttled, we do nothing. In fact, if it exhausted
830 * its budget it needs a replenishment and, since it now is on
831 * its rq, the bandwidth timer callback (which clearly has not
832 * run yet) will take care of this.
834 if (p->dl.dl_throttled)
837 enqueue_dl_entity(&p->dl, pi_se, flags);
839 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
840 enqueue_pushable_dl_task(rq, p);
843 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
845 dequeue_dl_entity(&p->dl);
846 dequeue_pushable_dl_task(rq, p);
849 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
852 __dequeue_task_dl(rq, p, flags);
856 * Yield task semantic for -deadline tasks is:
858 * get off from the CPU until our next instance, with
859 * a new runtime. This is of little use now, since we
860 * don't have a bandwidth reclaiming mechanism. Anyway,
861 * bandwidth reclaiming is planned for the future, and
862 * yield_task_dl will indicate that some spare budget
863 * is available for other task instances to use it.
865 static void yield_task_dl(struct rq *rq)
867 struct task_struct *p = rq->curr;
870 * We make the task go to sleep until its current deadline by
871 * forcing its runtime to zero. This way, update_curr_dl() stops
872 * it and the bandwidth timer will wake it up and will give it
873 * new scheduling parameters (thanks to dl_new=1).
875 if (p->dl.runtime > 0) {
876 rq->curr->dl.dl_new = 1;
884 static int find_later_rq(struct task_struct *task);
887 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
889 struct task_struct *curr;
892 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
898 curr = ACCESS_ONCE(rq->curr); /* unlocked access */
901 * If we are dealing with a -deadline task, we must
902 * decide where to wake it up.
903 * If it has a later deadline and the current task
904 * on this rq can't move (provided the waking task
905 * can!) we prefer to send it somewhere else. On the
906 * other hand, if it has a shorter deadline, we
907 * try to make it stay here, it might be important.
909 if (unlikely(dl_task(curr)) &&
910 (curr->nr_cpus_allowed < 2 ||
911 !dl_entity_preempt(&p->dl, &curr->dl)) &&
912 (p->nr_cpus_allowed > 1)) {
913 int target = find_later_rq(p);
924 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
927 * Current can't be migrated, useless to reschedule,
928 * let's hope p can move out.
930 if (rq->curr->nr_cpus_allowed == 1 ||
931 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
935 * p is migratable, so let's not schedule it and
936 * see if it is pushed or pulled somewhere else.
938 if (p->nr_cpus_allowed != 1 &&
939 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
942 resched_task(rq->curr);
945 #endif /* CONFIG_SMP */
948 * Only called when both the current and waking task are -deadline
951 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
954 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
955 resched_task(rq->curr);
961 * In the unlikely case current and p have the same deadline
962 * let us try to decide what's the best thing to do...
964 if ((p->dl.deadline == rq->curr->dl.deadline) &&
965 !test_tsk_need_resched(rq->curr))
966 check_preempt_equal_dl(rq, p);
967 #endif /* CONFIG_SMP */
970 #ifdef CONFIG_SCHED_HRTICK
971 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
973 s64 delta = p->dl.dl_runtime - p->dl.runtime;
976 hrtick_start(rq, p->dl.runtime);
980 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
983 struct rb_node *left = dl_rq->rb_leftmost;
988 return rb_entry(left, struct sched_dl_entity, rb_node);
991 struct task_struct *pick_next_task_dl(struct rq *rq)
993 struct sched_dl_entity *dl_se;
994 struct task_struct *p;
999 if (unlikely(!dl_rq->dl_nr_running))
1002 dl_se = pick_next_dl_entity(rq, dl_rq);
1005 p = dl_task_of(dl_se);
1006 p->se.exec_start = rq_clock_task(rq);
1008 /* Running task will never be pushed. */
1009 dequeue_pushable_dl_task(rq, p);
1011 #ifdef CONFIG_SCHED_HRTICK
1012 if (hrtick_enabled(rq))
1013 start_hrtick_dl(rq, p);
1017 rq->post_schedule = has_pushable_dl_tasks(rq);
1018 #endif /* CONFIG_SMP */
1023 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1027 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1028 enqueue_pushable_dl_task(rq, p);
1031 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1035 #ifdef CONFIG_SCHED_HRTICK
1036 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0)
1037 start_hrtick_dl(rq, p);
1041 static void task_fork_dl(struct task_struct *p)
1044 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1049 static void task_dead_dl(struct task_struct *p)
1051 struct hrtimer *timer = &p->dl.dl_timer;
1052 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1055 * Since we are TASK_DEAD we won't slip out of the domain!
1057 raw_spin_lock_irq(&dl_b->lock);
1058 dl_b->total_bw -= p->dl.dl_bw;
1059 raw_spin_unlock_irq(&dl_b->lock);
1061 hrtimer_cancel(timer);
1064 static void set_curr_task_dl(struct rq *rq)
1066 struct task_struct *p = rq->curr;
1068 p->se.exec_start = rq_clock_task(rq);
1070 /* You can't push away the running task */
1071 dequeue_pushable_dl_task(rq, p);
1076 /* Only try algorithms three times */
1077 #define DL_MAX_TRIES 3
1079 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1081 if (!task_running(rq, p) &&
1082 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1083 (p->nr_cpus_allowed > 1))
1089 /* Returns the second earliest -deadline task, NULL otherwise */
1090 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu)
1092 struct rb_node *next_node = rq->dl.rb_leftmost;
1093 struct sched_dl_entity *dl_se;
1094 struct task_struct *p = NULL;
1097 next_node = rb_next(next_node);
1099 dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
1100 p = dl_task_of(dl_se);
1102 if (pick_dl_task(rq, p, cpu))
1111 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1113 static int find_later_rq(struct task_struct *task)
1115 struct sched_domain *sd;
1116 struct cpumask *later_mask = __get_cpu_var(local_cpu_mask_dl);
1117 int this_cpu = smp_processor_id();
1118 int best_cpu, cpu = task_cpu(task);
1120 /* Make sure the mask is initialized first */
1121 if (unlikely(!later_mask))
1124 if (task->nr_cpus_allowed == 1)
1127 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
1133 * If we are here, some target has been found,
1134 * the most suitable of which is cached in best_cpu.
1135 * This is, among the runqueues where the current tasks
1136 * have later deadlines than the task's one, the rq
1137 * with the latest possible one.
1139 * Now we check how well this matches with task's
1140 * affinity and system topology.
1142 * The last cpu where the task run is our first
1143 * guess, since it is most likely cache-hot there.
1145 if (cpumask_test_cpu(cpu, later_mask))
1148 * Check if this_cpu is to be skipped (i.e., it is
1149 * not in the mask) or not.
1151 if (!cpumask_test_cpu(this_cpu, later_mask))
1155 for_each_domain(cpu, sd) {
1156 if (sd->flags & SD_WAKE_AFFINE) {
1159 * If possible, preempting this_cpu is
1160 * cheaper than migrating.
1162 if (this_cpu != -1 &&
1163 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1169 * Last chance: if best_cpu is valid and is
1170 * in the mask, that becomes our choice.
1172 if (best_cpu < nr_cpu_ids &&
1173 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
1182 * At this point, all our guesses failed, we just return
1183 * 'something', and let the caller sort the things out.
1188 cpu = cpumask_any(later_mask);
1189 if (cpu < nr_cpu_ids)
1195 /* Locks the rq it finds */
1196 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1198 struct rq *later_rq = NULL;
1202 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1203 cpu = find_later_rq(task);
1205 if ((cpu == -1) || (cpu == rq->cpu))
1208 later_rq = cpu_rq(cpu);
1210 /* Retry if something changed. */
1211 if (double_lock_balance(rq, later_rq)) {
1212 if (unlikely(task_rq(task) != rq ||
1213 !cpumask_test_cpu(later_rq->cpu,
1214 &task->cpus_allowed) ||
1215 task_running(rq, task) || !task->on_rq)) {
1216 double_unlock_balance(rq, later_rq);
1223 * If the rq we found has no -deadline task, or
1224 * its earliest one has a later deadline than our
1225 * task, the rq is a good one.
1227 if (!later_rq->dl.dl_nr_running ||
1228 dl_time_before(task->dl.deadline,
1229 later_rq->dl.earliest_dl.curr))
1232 /* Otherwise we try again. */
1233 double_unlock_balance(rq, later_rq);
1240 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1242 struct task_struct *p;
1244 if (!has_pushable_dl_tasks(rq))
1247 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
1248 struct task_struct, pushable_dl_tasks);
1250 BUG_ON(rq->cpu != task_cpu(p));
1251 BUG_ON(task_current(rq, p));
1252 BUG_ON(p->nr_cpus_allowed <= 1);
1255 BUG_ON(!dl_task(p));
1261 * See if the non running -deadline tasks on this rq
1262 * can be sent to some other CPU where they can preempt
1263 * and start executing.
1265 static int push_dl_task(struct rq *rq)
1267 struct task_struct *next_task;
1268 struct rq *later_rq;
1270 if (!rq->dl.overloaded)
1273 next_task = pick_next_pushable_dl_task(rq);
1278 if (unlikely(next_task == rq->curr)) {
1284 * If next_task preempts rq->curr, and rq->curr
1285 * can move away, it makes sense to just reschedule
1286 * without going further in pushing next_task.
1288 if (dl_task(rq->curr) &&
1289 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1290 rq->curr->nr_cpus_allowed > 1) {
1291 resched_task(rq->curr);
1295 /* We might release rq lock */
1296 get_task_struct(next_task);
1298 /* Will lock the rq it'll find */
1299 later_rq = find_lock_later_rq(next_task, rq);
1301 struct task_struct *task;
1304 * We must check all this again, since
1305 * find_lock_later_rq releases rq->lock and it is
1306 * then possible that next_task has migrated.
1308 task = pick_next_pushable_dl_task(rq);
1309 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1311 * The task is still there. We don't try
1312 * again, some other cpu will pull it when ready.
1314 dequeue_pushable_dl_task(rq, next_task);
1322 put_task_struct(next_task);
1327 deactivate_task(rq, next_task, 0);
1328 set_task_cpu(next_task, later_rq->cpu);
1329 activate_task(later_rq, next_task, 0);
1331 resched_task(later_rq->curr);
1333 double_unlock_balance(rq, later_rq);
1336 put_task_struct(next_task);
1341 static void push_dl_tasks(struct rq *rq)
1343 /* Terminates as it moves a -deadline task */
1344 while (push_dl_task(rq))
1348 static int pull_dl_task(struct rq *this_rq)
1350 int this_cpu = this_rq->cpu, ret = 0, cpu;
1351 struct task_struct *p;
1353 u64 dmin = LONG_MAX;
1355 if (likely(!dl_overloaded(this_rq)))
1359 * Match the barrier from dl_set_overloaded; this guarantees that if we
1360 * see overloaded we must also see the dlo_mask bit.
1364 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
1365 if (this_cpu == cpu)
1368 src_rq = cpu_rq(cpu);
1371 * It looks racy, abd it is! However, as in sched_rt.c,
1372 * we are fine with this.
1374 if (this_rq->dl.dl_nr_running &&
1375 dl_time_before(this_rq->dl.earliest_dl.curr,
1376 src_rq->dl.earliest_dl.next))
1379 /* Might drop this_rq->lock */
1380 double_lock_balance(this_rq, src_rq);
1383 * If there are no more pullable tasks on the
1384 * rq, we're done with it.
1386 if (src_rq->dl.dl_nr_running <= 1)
1389 p = pick_next_earliest_dl_task(src_rq, this_cpu);
1392 * We found a task to be pulled if:
1393 * - it preempts our current (if there's one),
1394 * - it will preempt the last one we pulled (if any).
1396 if (p && dl_time_before(p->dl.deadline, dmin) &&
1397 (!this_rq->dl.dl_nr_running ||
1398 dl_time_before(p->dl.deadline,
1399 this_rq->dl.earliest_dl.curr))) {
1400 WARN_ON(p == src_rq->curr);
1404 * Then we pull iff p has actually an earlier
1405 * deadline than the current task of its runqueue.
1407 if (dl_time_before(p->dl.deadline,
1408 src_rq->curr->dl.deadline))
1413 deactivate_task(src_rq, p, 0);
1414 set_task_cpu(p, this_cpu);
1415 activate_task(this_rq, p, 0);
1416 dmin = p->dl.deadline;
1418 /* Is there any other task even earlier? */
1421 double_unlock_balance(this_rq, src_rq);
1427 static void pre_schedule_dl(struct rq *rq, struct task_struct *prev)
1429 /* Try to pull other tasks here */
1434 static void post_schedule_dl(struct rq *rq)
1440 * Since the task is not running and a reschedule is not going to happen
1441 * anytime soon on its runqueue, we try pushing it away now.
1443 static void task_woken_dl(struct rq *rq, struct task_struct *p)
1445 if (!task_running(rq, p) &&
1446 !test_tsk_need_resched(rq->curr) &&
1447 has_pushable_dl_tasks(rq) &&
1448 p->nr_cpus_allowed > 1 &&
1449 dl_task(rq->curr) &&
1450 (rq->curr->nr_cpus_allowed < 2 ||
1451 dl_entity_preempt(&rq->curr->dl, &p->dl))) {
1456 static void set_cpus_allowed_dl(struct task_struct *p,
1457 const struct cpumask *new_mask)
1462 BUG_ON(!dl_task(p));
1465 * Update only if the task is actually running (i.e.,
1466 * it is on the rq AND it is not throttled).
1468 if (!on_dl_rq(&p->dl))
1471 weight = cpumask_weight(new_mask);
1474 * Only update if the process changes its state from whether it
1475 * can migrate or not.
1477 if ((p->nr_cpus_allowed > 1) == (weight > 1))
1483 * The process used to be able to migrate OR it can now migrate
1486 if (!task_current(rq, p))
1487 dequeue_pushable_dl_task(rq, p);
1488 BUG_ON(!rq->dl.dl_nr_migratory);
1489 rq->dl.dl_nr_migratory--;
1491 if (!task_current(rq, p))
1492 enqueue_pushable_dl_task(rq, p);
1493 rq->dl.dl_nr_migratory++;
1496 update_dl_migration(&rq->dl);
1499 /* Assumes rq->lock is held */
1500 static void rq_online_dl(struct rq *rq)
1502 if (rq->dl.overloaded)
1503 dl_set_overload(rq);
1505 if (rq->dl.dl_nr_running > 0)
1506 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1);
1509 /* Assumes rq->lock is held */
1510 static void rq_offline_dl(struct rq *rq)
1512 if (rq->dl.overloaded)
1513 dl_clear_overload(rq);
1515 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
1518 void init_sched_dl_class(void)
1522 for_each_possible_cpu(i)
1523 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
1524 GFP_KERNEL, cpu_to_node(i));
1527 #endif /* CONFIG_SMP */
1529 static void switched_from_dl(struct rq *rq, struct task_struct *p)
1531 if (hrtimer_active(&p->dl.dl_timer) && !dl_policy(p->policy))
1532 hrtimer_try_to_cancel(&p->dl.dl_timer);
1536 * Since this might be the only -deadline task on the rq,
1537 * this is the right place to try to pull some other one
1538 * from an overloaded cpu, if any.
1540 if (!rq->dl.dl_nr_running)
1546 * When switching to -deadline, we may overload the rq, then
1547 * we try to push someone off, if possible.
1549 static void switched_to_dl(struct rq *rq, struct task_struct *p)
1551 int check_resched = 1;
1554 * If p is throttled, don't consider the possibility
1555 * of preempting rq->curr, the check will be done right
1556 * after its runtime will get replenished.
1558 if (unlikely(p->dl.dl_throttled))
1561 if (p->on_rq || rq->curr != p) {
1563 if (rq->dl.overloaded && push_dl_task(rq) && rq != task_rq(p))
1564 /* Only reschedule if pushing failed */
1566 #endif /* CONFIG_SMP */
1567 if (check_resched && task_has_dl_policy(rq->curr))
1568 check_preempt_curr_dl(rq, p, 0);
1573 * If the scheduling parameters of a -deadline task changed,
1574 * a push or pull operation might be needed.
1576 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
1579 if (p->on_rq || rq->curr == p) {
1582 * This might be too much, but unfortunately
1583 * we don't have the old deadline value, and
1584 * we can't argue if the task is increasing
1585 * or lowering its prio, so...
1587 if (!rq->dl.overloaded)
1591 * If we now have a earlier deadline task than p,
1592 * then reschedule, provided p is still on this
1595 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) &&
1600 * Again, we don't know if p has a earlier
1601 * or later deadline, so let's blindly set a
1602 * (maybe not needed) rescheduling point.
1605 #endif /* CONFIG_SMP */
1607 switched_to_dl(rq, p);
1610 const struct sched_class dl_sched_class = {
1611 .next = &rt_sched_class,
1612 .enqueue_task = enqueue_task_dl,
1613 .dequeue_task = dequeue_task_dl,
1614 .yield_task = yield_task_dl,
1616 .check_preempt_curr = check_preempt_curr_dl,
1618 .pick_next_task = pick_next_task_dl,
1619 .put_prev_task = put_prev_task_dl,
1622 .select_task_rq = select_task_rq_dl,
1623 .set_cpus_allowed = set_cpus_allowed_dl,
1624 .rq_online = rq_online_dl,
1625 .rq_offline = rq_offline_dl,
1626 .pre_schedule = pre_schedule_dl,
1627 .post_schedule = post_schedule_dl,
1628 .task_woken = task_woken_dl,
1631 .set_curr_task = set_curr_task_dl,
1632 .task_tick = task_tick_dl,
1633 .task_fork = task_fork_dl,
1634 .task_dead = task_dead_dl,
1636 .prio_changed = prio_changed_dl,
1637 .switched_from = switched_from_dl,
1638 .switched_to = switched_to_dl,