1 // SPDX-License-Identifier: GPL-2.0
3 * Deadline Scheduling Class (SCHED_DEADLINE)
5 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
7 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't affect any other task.
13 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fchecconi@gmail.com>
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);
47 static inline struct dl_bw *dl_bw_of(int i)
49 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
50 "sched RCU must be held");
51 return &cpu_rq(i)->rd->dl_bw;
54 static inline int dl_bw_cpus(int i)
56 struct root_domain *rd = cpu_rq(i)->rd;
59 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
60 "sched RCU must be held");
61 for_each_cpu_and(i, rd->span, cpu_active_mask)
67 static inline struct dl_bw *dl_bw_of(int i)
69 return &cpu_rq(i)->dl.dl_bw;
72 static inline int dl_bw_cpus(int i)
79 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
81 u64 old = dl_rq->running_bw;
83 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
84 dl_rq->running_bw += dl_bw;
85 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
86 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
87 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
88 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
92 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
94 u64 old = dl_rq->running_bw;
96 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
97 dl_rq->running_bw -= dl_bw;
98 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
99 if (dl_rq->running_bw > old)
100 dl_rq->running_bw = 0;
101 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
102 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
106 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
108 u64 old = dl_rq->this_bw;
110 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
111 dl_rq->this_bw += dl_bw;
112 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
116 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
118 u64 old = dl_rq->this_bw;
120 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
121 dl_rq->this_bw -= dl_bw;
122 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
123 if (dl_rq->this_bw > old)
125 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
129 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
131 if (!dl_entity_is_special(dl_se))
132 __add_rq_bw(dl_se->dl_bw, dl_rq);
136 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
138 if (!dl_entity_is_special(dl_se))
139 __sub_rq_bw(dl_se->dl_bw, dl_rq);
143 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
145 if (!dl_entity_is_special(dl_se))
146 __add_running_bw(dl_se->dl_bw, dl_rq);
150 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
152 if (!dl_entity_is_special(dl_se))
153 __sub_running_bw(dl_se->dl_bw, dl_rq);
156 void dl_change_utilization(struct task_struct *p, u64 new_bw)
160 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
162 if (task_on_rq_queued(p))
166 if (p->dl.dl_non_contending) {
167 sub_running_bw(&p->dl, &rq->dl);
168 p->dl.dl_non_contending = 0;
170 * If the timer handler is currently running and the
171 * timer cannot be cancelled, inactive_task_timer()
172 * will see that dl_not_contending is not set, and
173 * will not touch the rq's active utilization,
174 * so we are still safe.
176 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
179 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
180 __add_rq_bw(new_bw, &rq->dl);
184 * The utilization of a task cannot be immediately removed from
185 * the rq active utilization (running_bw) when the task blocks.
186 * Instead, we have to wait for the so called "0-lag time".
188 * If a task blocks before the "0-lag time", a timer (the inactive
189 * timer) is armed, and running_bw is decreased when the timer
192 * If the task wakes up again before the inactive timer fires,
193 * the timer is cancelled, whereas if the task wakes up after the
194 * inactive timer fired (and running_bw has been decreased) the
195 * task's utilization has to be added to running_bw again.
196 * A flag in the deadline scheduling entity (dl_non_contending)
197 * is used to avoid race conditions between the inactive timer handler
200 * The following diagram shows how running_bw is updated. A task is
201 * "ACTIVE" when its utilization contributes to running_bw; an
202 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
203 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
204 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
205 * time already passed, which does not contribute to running_bw anymore.
206 * +------------------+
208 * +------------------>+ contending |
209 * | add_running_bw | |
210 * | +----+------+------+
213 * +--------+-------+ | |
214 * | | t >= 0-lag | | wakeup
215 * | INACTIVE |<---------------+ |
216 * | | sub_running_bw | |
217 * +--------+-------+ | |
222 * | +----+------+------+
223 * | sub_running_bw | ACTIVE |
224 * +-------------------+ |
225 * inactive timer | non contending |
226 * fired +------------------+
228 * The task_non_contending() function is invoked when a task
229 * blocks, and checks if the 0-lag time already passed or
230 * not (in the first case, it directly updates running_bw;
231 * in the second case, it arms the inactive timer).
233 * The task_contending() function is invoked when a task wakes
234 * up, and checks if the task is still in the "ACTIVE non contending"
235 * state or not (in the second case, it updates running_bw).
237 static void task_non_contending(struct task_struct *p)
239 struct sched_dl_entity *dl_se = &p->dl;
240 struct hrtimer *timer = &dl_se->inactive_timer;
241 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
242 struct rq *rq = rq_of_dl_rq(dl_rq);
246 * If this is a non-deadline task that has been boosted,
249 if (dl_se->dl_runtime == 0)
252 if (dl_entity_is_special(dl_se))
255 WARN_ON(dl_se->dl_non_contending);
257 zerolag_time = dl_se->deadline -
258 div64_long((dl_se->runtime * dl_se->dl_period),
262 * Using relative times instead of the absolute "0-lag time"
263 * allows to simplify the code
265 zerolag_time -= rq_clock(rq);
268 * If the "0-lag time" already passed, decrease the active
269 * utilization now, instead of starting a timer
271 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
273 sub_running_bw(dl_se, dl_rq);
274 if (!dl_task(p) || p->state == TASK_DEAD) {
275 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
277 if (p->state == TASK_DEAD)
278 sub_rq_bw(&p->dl, &rq->dl);
279 raw_spin_lock(&dl_b->lock);
280 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
281 __dl_clear_params(p);
282 raw_spin_unlock(&dl_b->lock);
288 dl_se->dl_non_contending = 1;
290 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL);
293 static void task_contending(struct sched_dl_entity *dl_se, int flags)
295 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
298 * If this is a non-deadline task that has been boosted,
301 if (dl_se->dl_runtime == 0)
304 if (flags & ENQUEUE_MIGRATED)
305 add_rq_bw(dl_se, dl_rq);
307 if (dl_se->dl_non_contending) {
308 dl_se->dl_non_contending = 0;
310 * If the timer handler is currently running and the
311 * timer cannot be cancelled, inactive_task_timer()
312 * will see that dl_not_contending is not set, and
313 * will not touch the rq's active utilization,
314 * so we are still safe.
316 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
317 put_task_struct(dl_task_of(dl_se));
320 * Since "dl_non_contending" is not set, the
321 * task's utilization has already been removed from
322 * active utilization (either when the task blocked,
323 * when the "inactive timer" fired).
326 add_running_bw(dl_se, dl_rq);
330 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
332 struct sched_dl_entity *dl_se = &p->dl;
334 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
337 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
339 raw_spin_lock_init(&dl_b->dl_runtime_lock);
340 dl_b->dl_period = period;
341 dl_b->dl_runtime = runtime;
344 void init_dl_bw(struct dl_bw *dl_b)
346 raw_spin_lock_init(&dl_b->lock);
347 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
348 if (global_rt_runtime() == RUNTIME_INF)
351 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
352 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
356 void init_dl_rq(struct dl_rq *dl_rq)
358 dl_rq->root = RB_ROOT_CACHED;
361 /* zero means no -deadline tasks */
362 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
364 dl_rq->dl_nr_migratory = 0;
365 dl_rq->overloaded = 0;
366 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
368 init_dl_bw(&dl_rq->dl_bw);
371 dl_rq->running_bw = 0;
373 init_dl_rq_bw_ratio(dl_rq);
378 static inline int dl_overloaded(struct rq *rq)
380 return atomic_read(&rq->rd->dlo_count);
383 static inline void dl_set_overload(struct rq *rq)
388 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
390 * Must be visible before the overload count is
391 * set (as in sched_rt.c).
393 * Matched by the barrier in pull_dl_task().
396 atomic_inc(&rq->rd->dlo_count);
399 static inline void dl_clear_overload(struct rq *rq)
404 atomic_dec(&rq->rd->dlo_count);
405 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
408 static void update_dl_migration(struct dl_rq *dl_rq)
410 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
411 if (!dl_rq->overloaded) {
412 dl_set_overload(rq_of_dl_rq(dl_rq));
413 dl_rq->overloaded = 1;
415 } else if (dl_rq->overloaded) {
416 dl_clear_overload(rq_of_dl_rq(dl_rq));
417 dl_rq->overloaded = 0;
421 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
423 struct task_struct *p = dl_task_of(dl_se);
425 if (p->nr_cpus_allowed > 1)
426 dl_rq->dl_nr_migratory++;
428 update_dl_migration(dl_rq);
431 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
433 struct task_struct *p = dl_task_of(dl_se);
435 if (p->nr_cpus_allowed > 1)
436 dl_rq->dl_nr_migratory--;
438 update_dl_migration(dl_rq);
442 * The list of pushable -deadline task is not a plist, like in
443 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
445 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
447 struct dl_rq *dl_rq = &rq->dl;
448 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node;
449 struct rb_node *parent = NULL;
450 struct task_struct *entry;
451 bool leftmost = true;
453 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
457 entry = rb_entry(parent, struct task_struct,
459 if (dl_entity_preempt(&p->dl, &entry->dl))
460 link = &parent->rb_left;
462 link = &parent->rb_right;
468 dl_rq->earliest_dl.next = p->dl.deadline;
470 rb_link_node(&p->pushable_dl_tasks, parent, link);
471 rb_insert_color_cached(&p->pushable_dl_tasks,
472 &dl_rq->pushable_dl_tasks_root, leftmost);
475 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
477 struct dl_rq *dl_rq = &rq->dl;
479 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
482 if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) {
483 struct rb_node *next_node;
485 next_node = rb_next(&p->pushable_dl_tasks);
487 dl_rq->earliest_dl.next = rb_entry(next_node,
488 struct task_struct, pushable_dl_tasks)->dl.deadline;
492 rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
493 RB_CLEAR_NODE(&p->pushable_dl_tasks);
496 static inline int has_pushable_dl_tasks(struct rq *rq)
498 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
501 static int push_dl_task(struct rq *rq);
503 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
505 return dl_task(prev);
508 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
509 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
511 static void push_dl_tasks(struct rq *);
512 static void pull_dl_task(struct rq *);
514 static inline void deadline_queue_push_tasks(struct rq *rq)
516 if (!has_pushable_dl_tasks(rq))
519 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
522 static inline void deadline_queue_pull_task(struct rq *rq)
524 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
527 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
529 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
531 struct rq *later_rq = NULL;
533 later_rq = find_lock_later_rq(p, rq);
538 * If we cannot preempt any rq, fall back to pick any
541 cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
542 if (cpu >= nr_cpu_ids) {
544 * Failed to find any suitable CPU.
545 * The task will never come back!
547 BUG_ON(dl_bandwidth_enabled());
550 * If admission control is disabled we
551 * try a little harder to let the task
554 cpu = cpumask_any(cpu_active_mask);
556 later_rq = cpu_rq(cpu);
557 double_lock_balance(rq, later_rq);
560 set_task_cpu(p, later_rq->cpu);
561 double_unlock_balance(later_rq, rq);
569 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
574 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
579 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
584 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
588 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
593 static inline void pull_dl_task(struct rq *rq)
597 static inline void deadline_queue_push_tasks(struct rq *rq)
601 static inline void deadline_queue_pull_task(struct rq *rq)
604 #endif /* CONFIG_SMP */
606 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
607 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
608 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
611 * We are being explicitly informed that a new instance is starting,
612 * and this means that:
613 * - the absolute deadline of the entity has to be placed at
614 * current time + relative deadline;
615 * - the runtime of the entity has to be set to the maximum value.
617 * The capability of specifying such event is useful whenever a -deadline
618 * entity wants to (try to!) synchronize its behaviour with the scheduler's
619 * one, and to (try to!) reconcile itself with its own scheduling
622 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
624 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
625 struct rq *rq = rq_of_dl_rq(dl_rq);
627 WARN_ON(dl_se->dl_boosted);
628 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
631 * We are racing with the deadline timer. So, do nothing because
632 * the deadline timer handler will take care of properly recharging
633 * the runtime and postponing the deadline
635 if (dl_se->dl_throttled)
639 * We use the regular wall clock time to set deadlines in the
640 * future; in fact, we must consider execution overheads (time
641 * spent on hardirq context, etc.).
643 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
644 dl_se->runtime = dl_se->dl_runtime;
648 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
649 * possibility of a entity lasting more than what it declared, and thus
650 * exhausting its runtime.
652 * Here we are interested in making runtime overrun possible, but we do
653 * not want a entity which is misbehaving to affect the scheduling of all
655 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
656 * is used, in order to confine each entity within its own bandwidth.
658 * This function deals exactly with that, and ensures that when the runtime
659 * of a entity is replenished, its deadline is also postponed. That ensures
660 * the overrunning entity can't interfere with other entity in the system and
661 * can't make them miss their deadlines. Reasons why this kind of overruns
662 * could happen are, typically, a entity voluntarily trying to overcome its
663 * runtime, or it just underestimated it during sched_setattr().
665 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
666 struct sched_dl_entity *pi_se)
668 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
669 struct rq *rq = rq_of_dl_rq(dl_rq);
671 BUG_ON(pi_se->dl_runtime <= 0);
674 * This could be the case for a !-dl task that is boosted.
675 * Just go with full inherited parameters.
677 if (dl_se->dl_deadline == 0) {
678 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
679 dl_se->runtime = pi_se->dl_runtime;
682 if (dl_se->dl_yielded && dl_se->runtime > 0)
686 * We keep moving the deadline away until we get some
687 * available runtime for the entity. This ensures correct
688 * handling of situations where the runtime overrun is
691 while (dl_se->runtime <= 0) {
692 dl_se->deadline += pi_se->dl_period;
693 dl_se->runtime += pi_se->dl_runtime;
697 * At this point, the deadline really should be "in
698 * the future" with respect to rq->clock. If it's
699 * not, we are, for some reason, lagging too much!
700 * Anyway, after having warn userspace abut that,
701 * we still try to keep the things running by
702 * resetting the deadline and the budget of the
705 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
706 printk_deferred_once("sched: DL replenish lagged too much\n");
707 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
708 dl_se->runtime = pi_se->dl_runtime;
711 if (dl_se->dl_yielded)
712 dl_se->dl_yielded = 0;
713 if (dl_se->dl_throttled)
714 dl_se->dl_throttled = 0;
718 * Here we check if --at time t-- an entity (which is probably being
719 * [re]activated or, in general, enqueued) can use its remaining runtime
720 * and its current deadline _without_ exceeding the bandwidth it is
721 * assigned (function returns true if it can't). We are in fact applying
722 * one of the CBS rules: when a task wakes up, if the residual runtime
723 * over residual deadline fits within the allocated bandwidth, then we
724 * can keep the current (absolute) deadline and residual budget without
725 * disrupting the schedulability of the system. Otherwise, we should
726 * refill the runtime and set the deadline a period in the future,
727 * because keeping the current (absolute) deadline of the task would
728 * result in breaking guarantees promised to other tasks (refer to
729 * Documentation/scheduler/sched-deadline.txt for more information).
731 * This function returns true if:
733 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
735 * IOW we can't recycle current parameters.
737 * Notice that the bandwidth check is done against the deadline. For
738 * task with deadline equal to period this is the same of using
739 * dl_period instead of dl_deadline in the equation above.
741 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
742 struct sched_dl_entity *pi_se, u64 t)
747 * left and right are the two sides of the equation above,
748 * after a bit of shuffling to use multiplications instead
751 * Note that none of the time values involved in the two
752 * multiplications are absolute: dl_deadline and dl_runtime
753 * are the relative deadline and the maximum runtime of each
754 * instance, runtime is the runtime left for the last instance
755 * and (deadline - t), since t is rq->clock, is the time left
756 * to the (absolute) deadline. Even if overflowing the u64 type
757 * is very unlikely to occur in both cases, here we scale down
758 * as we want to avoid that risk at all. Scaling down by 10
759 * means that we reduce granularity to 1us. We are fine with it,
760 * since this is only a true/false check and, anyway, thinking
761 * of anything below microseconds resolution is actually fiction
762 * (but still we want to give the user that illusion >;).
764 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
765 right = ((dl_se->deadline - t) >> DL_SCALE) *
766 (pi_se->dl_runtime >> DL_SCALE);
768 return dl_time_before(right, left);
772 * Revised wakeup rule [1]: For self-suspending tasks, rather then
773 * re-initializing task's runtime and deadline, the revised wakeup
774 * rule adjusts the task's runtime to avoid the task to overrun its
777 * Reasoning: a task may overrun the density if:
778 * runtime / (deadline - t) > dl_runtime / dl_deadline
780 * Therefore, runtime can be adjusted to:
781 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
783 * In such way that runtime will be equal to the maximum density
784 * the task can use without breaking any rule.
786 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
787 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
790 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
792 u64 laxity = dl_se->deadline - rq_clock(rq);
795 * If the task has deadline < period, and the deadline is in the past,
796 * it should already be throttled before this check.
798 * See update_dl_entity() comments for further details.
800 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
802 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
806 * Regarding the deadline, a task with implicit deadline has a relative
807 * deadline == relative period. A task with constrained deadline has a
808 * relative deadline <= relative period.
810 * We support constrained deadline tasks. However, there are some restrictions
811 * applied only for tasks which do not have an implicit deadline. See
812 * update_dl_entity() to know more about such restrictions.
814 * The dl_is_implicit() returns true if the task has an implicit deadline.
816 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
818 return dl_se->dl_deadline == dl_se->dl_period;
822 * When a deadline entity is placed in the runqueue, its runtime and deadline
823 * might need to be updated. This is done by a CBS wake up rule. There are two
824 * different rules: 1) the original CBS; and 2) the Revisited CBS.
826 * When the task is starting a new period, the Original CBS is used. In this
827 * case, the runtime is replenished and a new absolute deadline is set.
829 * When a task is queued before the begin of the next period, using the
830 * remaining runtime and deadline could make the entity to overflow, see
831 * dl_entity_overflow() to find more about runtime overflow. When such case
832 * is detected, the runtime and deadline need to be updated.
834 * If the task has an implicit deadline, i.e., deadline == period, the Original
835 * CBS is applied. the runtime is replenished and a new absolute deadline is
836 * set, as in the previous cases.
838 * However, the Original CBS does not work properly for tasks with
839 * deadline < period, which are said to have a constrained deadline. By
840 * applying the Original CBS, a constrained deadline task would be able to run
841 * runtime/deadline in a period. With deadline < period, the task would
842 * overrun the runtime/period allowed bandwidth, breaking the admission test.
844 * In order to prevent this misbehave, the Revisited CBS is used for
845 * constrained deadline tasks when a runtime overflow is detected. In the
846 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
847 * the remaining runtime of the task is reduced to avoid runtime overflow.
848 * Please refer to the comments update_dl_revised_wakeup() function to find
849 * more about the Revised CBS rule.
851 static void update_dl_entity(struct sched_dl_entity *dl_se,
852 struct sched_dl_entity *pi_se)
854 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
855 struct rq *rq = rq_of_dl_rq(dl_rq);
857 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
858 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
860 if (unlikely(!dl_is_implicit(dl_se) &&
861 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
862 !dl_se->dl_boosted)){
863 update_dl_revised_wakeup(dl_se, rq);
867 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
868 dl_se->runtime = pi_se->dl_runtime;
872 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
874 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
878 * If the entity depleted all its runtime, and if we want it to sleep
879 * while waiting for some new execution time to become available, we
880 * set the bandwidth replenishment timer to the replenishment instant
881 * and try to activate it.
883 * Notice that it is important for the caller to know if the timer
884 * actually started or not (i.e., the replenishment instant is in
885 * the future or in the past).
887 static int start_dl_timer(struct task_struct *p)
889 struct sched_dl_entity *dl_se = &p->dl;
890 struct hrtimer *timer = &dl_se->dl_timer;
891 struct rq *rq = task_rq(p);
895 lockdep_assert_held(&rq->lock);
898 * We want the timer to fire at the deadline, but considering
899 * that it is actually coming from rq->clock and not from
900 * hrtimer's time base reading.
902 act = ns_to_ktime(dl_next_period(dl_se));
903 now = hrtimer_cb_get_time(timer);
904 delta = ktime_to_ns(now) - rq_clock(rq);
905 act = ktime_add_ns(act, delta);
908 * If the expiry time already passed, e.g., because the value
909 * chosen as the deadline is too small, don't even try to
910 * start the timer in the past!
912 if (ktime_us_delta(act, now) < 0)
916 * !enqueued will guarantee another callback; even if one is already in
917 * progress. This ensures a balanced {get,put}_task_struct().
919 * The race against __run_timer() clearing the enqueued state is
920 * harmless because we're holding task_rq()->lock, therefore the timer
921 * expiring after we've done the check will wait on its task_rq_lock()
922 * and observe our state.
924 if (!hrtimer_is_queued(timer)) {
926 hrtimer_start(timer, act, HRTIMER_MODE_ABS);
933 * This is the bandwidth enforcement timer callback. If here, we know
934 * a task is not on its dl_rq, since the fact that the timer was running
935 * means the task is throttled and needs a runtime replenishment.
937 * However, what we actually do depends on the fact the task is active,
938 * (it is on its rq) or has been removed from there by a call to
939 * dequeue_task_dl(). In the former case we must issue the runtime
940 * replenishment and add the task back to the dl_rq; in the latter, we just
941 * do nothing but clearing dl_throttled, so that runtime and deadline
942 * updating (and the queueing back to dl_rq) will be done by the
943 * next call to enqueue_task_dl().
945 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
947 struct sched_dl_entity *dl_se = container_of(timer,
948 struct sched_dl_entity,
950 struct task_struct *p = dl_task_of(dl_se);
954 rq = task_rq_lock(p, &rf);
957 * The task might have changed its scheduling policy to something
958 * different than SCHED_DEADLINE (through switched_from_dl()).
964 * The task might have been boosted by someone else and might be in the
965 * boosting/deboosting path, its not throttled.
967 if (dl_se->dl_boosted)
971 * Spurious timer due to start_dl_timer() race; or we already received
972 * a replenishment from rt_mutex_setprio().
974 if (!dl_se->dl_throttled)
981 * If the throttle happened during sched-out; like:
988 * __dequeue_task_dl()
991 * We can be both throttled and !queued. Replenish the counter
992 * but do not enqueue -- wait for our wakeup to do that.
994 if (!task_on_rq_queued(p)) {
995 replenish_dl_entity(dl_se, dl_se);
1000 if (unlikely(!rq->online)) {
1002 * If the runqueue is no longer available, migrate the
1003 * task elsewhere. This necessarily changes rq.
1005 lockdep_unpin_lock(&rq->lock, rf.cookie);
1006 rq = dl_task_offline_migration(rq, p);
1007 rf.cookie = lockdep_pin_lock(&rq->lock);
1008 update_rq_clock(rq);
1011 * Now that the task has been migrated to the new RQ and we
1012 * have that locked, proceed as normal and enqueue the task
1018 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1019 if (dl_task(rq->curr))
1020 check_preempt_curr_dl(rq, p, 0);
1026 * Queueing this task back might have overloaded rq, check if we need
1027 * to kick someone away.
1029 if (has_pushable_dl_tasks(rq)) {
1031 * Nothing relies on rq->lock after this, so its safe to drop
1034 rq_unpin_lock(rq, &rf);
1036 rq_repin_lock(rq, &rf);
1041 task_rq_unlock(rq, p, &rf);
1044 * This can free the task_struct, including this hrtimer, do not touch
1045 * anything related to that after this.
1049 return HRTIMER_NORESTART;
1052 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1054 struct hrtimer *timer = &dl_se->dl_timer;
1056 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1057 timer->function = dl_task_timer;
1061 * During the activation, CBS checks if it can reuse the current task's
1062 * runtime and period. If the deadline of the task is in the past, CBS
1063 * cannot use the runtime, and so it replenishes the task. This rule
1064 * works fine for implicit deadline tasks (deadline == period), and the
1065 * CBS was designed for implicit deadline tasks. However, a task with
1066 * constrained deadline (deadine < period) might be awakened after the
1067 * deadline, but before the next period. In this case, replenishing the
1068 * task would allow it to run for runtime / deadline. As in this case
1069 * deadline < period, CBS enables a task to run for more than the
1070 * runtime / period. In a very loaded system, this can cause a domino
1071 * effect, making other tasks miss their deadlines.
1073 * To avoid this problem, in the activation of a constrained deadline
1074 * task after the deadline but before the next period, throttle the
1075 * task and set the replenishing timer to the begin of the next period,
1076 * unless it is boosted.
1078 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1080 struct task_struct *p = dl_task_of(dl_se);
1081 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1083 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1084 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1085 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1087 dl_se->dl_throttled = 1;
1088 if (dl_se->runtime > 0)
1094 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1096 return (dl_se->runtime <= 0);
1099 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1102 * This function implements the GRUB accounting rule:
1103 * according to the GRUB reclaiming algorithm, the runtime is
1104 * not decreased as "dq = -dt", but as
1105 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1106 * where u is the utilization of the task, Umax is the maximum reclaimable
1107 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1108 * as the difference between the "total runqueue utilization" and the
1109 * runqueue active utilization, and Uextra is the (per runqueue) extra
1110 * reclaimable utilization.
1111 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1112 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1114 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1115 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1116 * Since delta is a 64 bit variable, to have an overflow its value
1117 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1118 * So, overflow is not an issue here.
1120 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1122 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1124 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1127 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1128 * we compare u_inact + rq->dl.extra_bw with
1129 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1130 * u_inact + rq->dl.extra_bw can be larger than
1131 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1132 * leading to wrong results)
1134 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1137 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1139 return (delta * u_act) >> BW_SHIFT;
1143 * Update the current task's runtime statistics (provided it is still
1144 * a -deadline task and has not been removed from the dl_rq).
1146 static void update_curr_dl(struct rq *rq)
1148 struct task_struct *curr = rq->curr;
1149 struct sched_dl_entity *dl_se = &curr->dl;
1150 u64 delta_exec, scaled_delta_exec;
1151 int cpu = cpu_of(rq);
1154 if (!dl_task(curr) || !on_dl_rq(dl_se))
1158 * Consumed budget is computed considering the time as
1159 * observed by schedulable tasks (excluding time spent
1160 * in hardirq context, etc.). Deadlines are instead
1161 * computed using hard walltime. This seems to be the more
1162 * natural solution, but the full ramifications of this
1163 * approach need further study.
1165 now = rq_clock_task(rq);
1166 delta_exec = now - curr->se.exec_start;
1167 if (unlikely((s64)delta_exec <= 0)) {
1168 if (unlikely(dl_se->dl_yielded))
1173 schedstat_set(curr->se.statistics.exec_max,
1174 max(curr->se.statistics.exec_max, delta_exec));
1176 curr->se.sum_exec_runtime += delta_exec;
1177 account_group_exec_runtime(curr, delta_exec);
1179 curr->se.exec_start = now;
1180 cgroup_account_cputime(curr, delta_exec);
1182 if (dl_entity_is_special(dl_se))
1186 * For tasks that participate in GRUB, we implement GRUB-PA: the
1187 * spare reclaimed bandwidth is used to clock down frequency.
1189 * For the others, we still need to scale reservation parameters
1190 * according to current frequency and CPU maximum capacity.
1192 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1193 scaled_delta_exec = grub_reclaim(delta_exec,
1197 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1198 unsigned long scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
1200 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1201 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1204 dl_se->runtime -= scaled_delta_exec;
1207 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1208 dl_se->dl_throttled = 1;
1210 /* If requested, inform the user about runtime overruns. */
1211 if (dl_runtime_exceeded(dl_se) &&
1212 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1213 dl_se->dl_overrun = 1;
1215 __dequeue_task_dl(rq, curr, 0);
1216 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
1217 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1219 if (!is_leftmost(curr, &rq->dl))
1224 * Because -- for now -- we share the rt bandwidth, we need to
1225 * account our runtime there too, otherwise actual rt tasks
1226 * would be able to exceed the shared quota.
1228 * Account to the root rt group for now.
1230 * The solution we're working towards is having the RT groups scheduled
1231 * using deadline servers -- however there's a few nasties to figure
1232 * out before that can happen.
1234 if (rt_bandwidth_enabled()) {
1235 struct rt_rq *rt_rq = &rq->rt;
1237 raw_spin_lock(&rt_rq->rt_runtime_lock);
1239 * We'll let actual RT tasks worry about the overflow here, we
1240 * have our own CBS to keep us inline; only account when RT
1241 * bandwidth is relevant.
1243 if (sched_rt_bandwidth_account(rt_rq))
1244 rt_rq->rt_time += delta_exec;
1245 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1249 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1251 struct sched_dl_entity *dl_se = container_of(timer,
1252 struct sched_dl_entity,
1254 struct task_struct *p = dl_task_of(dl_se);
1258 rq = task_rq_lock(p, &rf);
1261 update_rq_clock(rq);
1263 if (!dl_task(p) || p->state == TASK_DEAD) {
1264 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1266 if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1267 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1268 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1269 dl_se->dl_non_contending = 0;
1272 raw_spin_lock(&dl_b->lock);
1273 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1274 raw_spin_unlock(&dl_b->lock);
1275 __dl_clear_params(p);
1279 if (dl_se->dl_non_contending == 0)
1282 sub_running_bw(dl_se, &rq->dl);
1283 dl_se->dl_non_contending = 0;
1285 task_rq_unlock(rq, p, &rf);
1288 return HRTIMER_NORESTART;
1291 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1293 struct hrtimer *timer = &dl_se->inactive_timer;
1295 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1296 timer->function = inactive_task_timer;
1301 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1303 struct rq *rq = rq_of_dl_rq(dl_rq);
1305 if (dl_rq->earliest_dl.curr == 0 ||
1306 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1307 dl_rq->earliest_dl.curr = deadline;
1308 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1312 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1314 struct rq *rq = rq_of_dl_rq(dl_rq);
1317 * Since we may have removed our earliest (and/or next earliest)
1318 * task we must recompute them.
1320 if (!dl_rq->dl_nr_running) {
1321 dl_rq->earliest_dl.curr = 0;
1322 dl_rq->earliest_dl.next = 0;
1323 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1325 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1326 struct sched_dl_entity *entry;
1328 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1329 dl_rq->earliest_dl.curr = entry->deadline;
1330 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1336 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1337 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1339 #endif /* CONFIG_SMP */
1342 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1344 int prio = dl_task_of(dl_se)->prio;
1345 u64 deadline = dl_se->deadline;
1347 WARN_ON(!dl_prio(prio));
1348 dl_rq->dl_nr_running++;
1349 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1351 inc_dl_deadline(dl_rq, deadline);
1352 inc_dl_migration(dl_se, dl_rq);
1356 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1358 int prio = dl_task_of(dl_se)->prio;
1360 WARN_ON(!dl_prio(prio));
1361 WARN_ON(!dl_rq->dl_nr_running);
1362 dl_rq->dl_nr_running--;
1363 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1365 dec_dl_deadline(dl_rq, dl_se->deadline);
1366 dec_dl_migration(dl_se, dl_rq);
1369 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1371 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1372 struct rb_node **link = &dl_rq->root.rb_root.rb_node;
1373 struct rb_node *parent = NULL;
1374 struct sched_dl_entity *entry;
1377 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1381 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
1382 if (dl_time_before(dl_se->deadline, entry->deadline))
1383 link = &parent->rb_left;
1385 link = &parent->rb_right;
1390 rb_link_node(&dl_se->rb_node, parent, link);
1391 rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost);
1393 inc_dl_tasks(dl_se, dl_rq);
1396 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1398 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1400 if (RB_EMPTY_NODE(&dl_se->rb_node))
1403 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1404 RB_CLEAR_NODE(&dl_se->rb_node);
1406 dec_dl_tasks(dl_se, dl_rq);
1410 enqueue_dl_entity(struct sched_dl_entity *dl_se,
1411 struct sched_dl_entity *pi_se, int flags)
1413 BUG_ON(on_dl_rq(dl_se));
1416 * If this is a wakeup or a new instance, the scheduling
1417 * parameters of the task might need updating. Otherwise,
1418 * we want a replenishment of its runtime.
1420 if (flags & ENQUEUE_WAKEUP) {
1421 task_contending(dl_se, flags);
1422 update_dl_entity(dl_se, pi_se);
1423 } else if (flags & ENQUEUE_REPLENISH) {
1424 replenish_dl_entity(dl_se, pi_se);
1425 } else if ((flags & ENQUEUE_RESTORE) &&
1426 dl_time_before(dl_se->deadline,
1427 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1428 setup_new_dl_entity(dl_se);
1431 __enqueue_dl_entity(dl_se);
1434 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1436 __dequeue_dl_entity(dl_se);
1439 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1441 struct task_struct *pi_task = rt_mutex_get_top_task(p);
1442 struct sched_dl_entity *pi_se = &p->dl;
1445 * Use the scheduling parameters of the top pi-waiter task if:
1446 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1447 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1448 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1449 * boosted due to a SCHED_DEADLINE pi-waiter).
1450 * Otherwise we keep our runtime and deadline.
1452 if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) {
1453 pi_se = &pi_task->dl;
1454 } else if (!dl_prio(p->normal_prio)) {
1456 * Special case in which we have a !SCHED_DEADLINE task
1457 * that is going to be deboosted, but exceeds its
1458 * runtime while doing so. No point in replenishing
1459 * it, as it's going to return back to its original
1460 * scheduling class after this.
1462 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
1467 * Check if a constrained deadline task was activated
1468 * after the deadline but before the next period.
1469 * If that is the case, the task will be throttled and
1470 * the replenishment timer will be set to the next period.
1472 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1473 dl_check_constrained_dl(&p->dl);
1475 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1476 add_rq_bw(&p->dl, &rq->dl);
1477 add_running_bw(&p->dl, &rq->dl);
1481 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1482 * its budget it needs a replenishment and, since it now is on
1483 * its rq, the bandwidth timer callback (which clearly has not
1484 * run yet) will take care of this.
1485 * However, the active utilization does not depend on the fact
1486 * that the task is on the runqueue or not (but depends on the
1487 * task's state - in GRUB parlance, "inactive" vs "active contending").
1488 * In other words, even if a task is throttled its utilization must
1489 * be counted in the active utilization; hence, we need to call
1492 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1493 if (flags & ENQUEUE_WAKEUP)
1494 task_contending(&p->dl, flags);
1499 enqueue_dl_entity(&p->dl, pi_se, flags);
1501 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1502 enqueue_pushable_dl_task(rq, p);
1505 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1507 dequeue_dl_entity(&p->dl);
1508 dequeue_pushable_dl_task(rq, p);
1511 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1514 __dequeue_task_dl(rq, p, flags);
1516 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1517 sub_running_bw(&p->dl, &rq->dl);
1518 sub_rq_bw(&p->dl, &rq->dl);
1522 * This check allows to start the inactive timer (or to immediately
1523 * decrease the active utilization, if needed) in two cases:
1524 * when the task blocks and when it is terminating
1525 * (p->state == TASK_DEAD). We can handle the two cases in the same
1526 * way, because from GRUB's point of view the same thing is happening
1527 * (the task moves from "active contending" to "active non contending"
1530 if (flags & DEQUEUE_SLEEP)
1531 task_non_contending(p);
1535 * Yield task semantic for -deadline tasks is:
1537 * get off from the CPU until our next instance, with
1538 * a new runtime. This is of little use now, since we
1539 * don't have a bandwidth reclaiming mechanism. Anyway,
1540 * bandwidth reclaiming is planned for the future, and
1541 * yield_task_dl will indicate that some spare budget
1542 * is available for other task instances to use it.
1544 static void yield_task_dl(struct rq *rq)
1547 * We make the task go to sleep until its current deadline by
1548 * forcing its runtime to zero. This way, update_curr_dl() stops
1549 * it and the bandwidth timer will wake it up and will give it
1550 * new scheduling parameters (thanks to dl_yielded=1).
1552 rq->curr->dl.dl_yielded = 1;
1554 update_rq_clock(rq);
1557 * Tell update_rq_clock() that we've just updated,
1558 * so we don't do microscopic update in schedule()
1559 * and double the fastpath cost.
1561 rq_clock_skip_update(rq);
1566 static int find_later_rq(struct task_struct *task);
1569 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1571 struct task_struct *curr;
1574 if (sd_flag != SD_BALANCE_WAKE)
1580 curr = READ_ONCE(rq->curr); /* unlocked access */
1583 * If we are dealing with a -deadline task, we must
1584 * decide where to wake it up.
1585 * If it has a later deadline and the current task
1586 * on this rq can't move (provided the waking task
1587 * can!) we prefer to send it somewhere else. On the
1588 * other hand, if it has a shorter deadline, we
1589 * try to make it stay here, it might be important.
1591 if (unlikely(dl_task(curr)) &&
1592 (curr->nr_cpus_allowed < 2 ||
1593 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1594 (p->nr_cpus_allowed > 1)) {
1595 int target = find_later_rq(p);
1598 (dl_time_before(p->dl.deadline,
1599 cpu_rq(target)->dl.earliest_dl.curr) ||
1600 (cpu_rq(target)->dl.dl_nr_running == 0)))
1609 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1613 if (p->state != TASK_WAKING)
1618 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1619 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1620 * rq->lock is not... So, lock it
1622 raw_spin_lock(&rq->lock);
1623 if (p->dl.dl_non_contending) {
1624 sub_running_bw(&p->dl, &rq->dl);
1625 p->dl.dl_non_contending = 0;
1627 * If the timer handler is currently running and the
1628 * timer cannot be cancelled, inactive_task_timer()
1629 * will see that dl_not_contending is not set, and
1630 * will not touch the rq's active utilization,
1631 * so we are still safe.
1633 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1636 sub_rq_bw(&p->dl, &rq->dl);
1637 raw_spin_unlock(&rq->lock);
1640 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1643 * Current can't be migrated, useless to reschedule,
1644 * let's hope p can move out.
1646 if (rq->curr->nr_cpus_allowed == 1 ||
1647 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1651 * p is migratable, so let's not schedule it and
1652 * see if it is pushed or pulled somewhere else.
1654 if (p->nr_cpus_allowed != 1 &&
1655 cpudl_find(&rq->rd->cpudl, p, NULL))
1661 #endif /* CONFIG_SMP */
1664 * Only called when both the current and waking task are -deadline
1667 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1670 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1677 * In the unlikely case current and p have the same deadline
1678 * let us try to decide what's the best thing to do...
1680 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1681 !test_tsk_need_resched(rq->curr))
1682 check_preempt_equal_dl(rq, p);
1683 #endif /* CONFIG_SMP */
1686 #ifdef CONFIG_SCHED_HRTICK
1687 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1689 hrtick_start(rq, p->dl.runtime);
1691 #else /* !CONFIG_SCHED_HRTICK */
1692 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1697 static inline void set_next_task(struct rq *rq, struct task_struct *p)
1699 p->se.exec_start = rq_clock_task(rq);
1701 /* You can't push away the running task */
1702 dequeue_pushable_dl_task(rq, p);
1705 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1706 struct dl_rq *dl_rq)
1708 struct rb_node *left = rb_first_cached(&dl_rq->root);
1713 return rb_entry(left, struct sched_dl_entity, rb_node);
1716 static struct task_struct *
1717 pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1719 struct sched_dl_entity *dl_se;
1720 struct task_struct *p;
1721 struct dl_rq *dl_rq;
1725 if (need_pull_dl_task(rq, prev)) {
1727 * This is OK, because current is on_cpu, which avoids it being
1728 * picked for load-balance and preemption/IRQs are still
1729 * disabled avoiding further scheduler activity on it and we're
1730 * being very careful to re-start the picking loop.
1732 rq_unpin_lock(rq, rf);
1734 rq_repin_lock(rq, rf);
1736 * pull_dl_task() can drop (and re-acquire) rq->lock; this
1737 * means a stop task can slip in, in which case we need to
1738 * re-start task selection.
1740 if (rq->stop && task_on_rq_queued(rq->stop))
1745 * When prev is DL, we may throttle it in put_prev_task().
1746 * So, we update time before we check for dl_nr_running.
1748 if (prev->sched_class == &dl_sched_class)
1751 if (unlikely(!dl_rq->dl_nr_running))
1754 put_prev_task(rq, prev);
1756 dl_se = pick_next_dl_entity(rq, dl_rq);
1759 p = dl_task_of(dl_se);
1761 set_next_task(rq, p);
1763 if (hrtick_enabled(rq))
1764 start_hrtick_dl(rq, p);
1766 deadline_queue_push_tasks(rq);
1768 if (rq->curr->sched_class != &dl_sched_class)
1769 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1774 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1778 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1779 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1780 enqueue_pushable_dl_task(rq, p);
1784 * scheduler tick hitting a task of our scheduling class.
1786 * NOTE: This function can be called remotely by the tick offload that
1787 * goes along full dynticks. Therefore no local assumption can be made
1788 * and everything must be accessed through the @rq and @curr passed in
1791 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1795 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1797 * Even when we have runtime, update_curr_dl() might have resulted in us
1798 * not being the leftmost task anymore. In that case NEED_RESCHED will
1799 * be set and schedule() will start a new hrtick for the next task.
1801 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1802 is_leftmost(p, &rq->dl))
1803 start_hrtick_dl(rq, p);
1806 static void task_fork_dl(struct task_struct *p)
1809 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1814 static void set_curr_task_dl(struct rq *rq)
1816 set_next_task(rq, rq->curr);
1821 /* Only try algorithms three times */
1822 #define DL_MAX_TRIES 3
1824 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1826 if (!task_running(rq, p) &&
1827 cpumask_test_cpu(cpu, &p->cpus_allowed))
1833 * Return the earliest pushable rq's task, which is suitable to be executed
1834 * on the CPU, NULL otherwise:
1836 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1838 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
1839 struct task_struct *p = NULL;
1841 if (!has_pushable_dl_tasks(rq))
1846 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1848 if (pick_dl_task(rq, p, cpu))
1851 next_node = rb_next(next_node);
1858 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1860 static int find_later_rq(struct task_struct *task)
1862 struct sched_domain *sd;
1863 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1864 int this_cpu = smp_processor_id();
1865 int cpu = task_cpu(task);
1867 /* Make sure the mask is initialized first */
1868 if (unlikely(!later_mask))
1871 if (task->nr_cpus_allowed == 1)
1875 * We have to consider system topology and task affinity
1876 * first, then we can look for a suitable CPU.
1878 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1882 * If we are here, some targets have been found, including
1883 * the most suitable which is, among the runqueues where the
1884 * current tasks have later deadlines than the task's one, the
1885 * rq with the latest possible one.
1887 * Now we check how well this matches with task's
1888 * affinity and system topology.
1890 * The last CPU where the task run is our first
1891 * guess, since it is most likely cache-hot there.
1893 if (cpumask_test_cpu(cpu, later_mask))
1896 * Check if this_cpu is to be skipped (i.e., it is
1897 * not in the mask) or not.
1899 if (!cpumask_test_cpu(this_cpu, later_mask))
1903 for_each_domain(cpu, sd) {
1904 if (sd->flags & SD_WAKE_AFFINE) {
1908 * If possible, preempting this_cpu is
1909 * cheaper than migrating.
1911 if (this_cpu != -1 &&
1912 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1917 best_cpu = cpumask_first_and(later_mask,
1918 sched_domain_span(sd));
1920 * Last chance: if a CPU being in both later_mask
1921 * and current sd span is valid, that becomes our
1922 * choice. Of course, the latest possible CPU is
1923 * already under consideration through later_mask.
1925 if (best_cpu < nr_cpu_ids) {
1934 * At this point, all our guesses failed, we just return
1935 * 'something', and let the caller sort the things out.
1940 cpu = cpumask_any(later_mask);
1941 if (cpu < nr_cpu_ids)
1947 /* Locks the rq it finds */
1948 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1950 struct rq *later_rq = NULL;
1954 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1955 cpu = find_later_rq(task);
1957 if ((cpu == -1) || (cpu == rq->cpu))
1960 later_rq = cpu_rq(cpu);
1962 if (later_rq->dl.dl_nr_running &&
1963 !dl_time_before(task->dl.deadline,
1964 later_rq->dl.earliest_dl.curr)) {
1966 * Target rq has tasks of equal or earlier deadline,
1967 * retrying does not release any lock and is unlikely
1968 * to yield a different result.
1974 /* Retry if something changed. */
1975 if (double_lock_balance(rq, later_rq)) {
1976 if (unlikely(task_rq(task) != rq ||
1977 !cpumask_test_cpu(later_rq->cpu, &task->cpus_allowed) ||
1978 task_running(rq, task) ||
1980 !task_on_rq_queued(task))) {
1981 double_unlock_balance(rq, later_rq);
1988 * If the rq we found has no -deadline task, or
1989 * its earliest one has a later deadline than our
1990 * task, the rq is a good one.
1992 if (!later_rq->dl.dl_nr_running ||
1993 dl_time_before(task->dl.deadline,
1994 later_rq->dl.earliest_dl.curr))
1997 /* Otherwise we try again. */
1998 double_unlock_balance(rq, later_rq);
2005 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2007 struct task_struct *p;
2009 if (!has_pushable_dl_tasks(rq))
2012 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
2013 struct task_struct, pushable_dl_tasks);
2015 BUG_ON(rq->cpu != task_cpu(p));
2016 BUG_ON(task_current(rq, p));
2017 BUG_ON(p->nr_cpus_allowed <= 1);
2019 BUG_ON(!task_on_rq_queued(p));
2020 BUG_ON(!dl_task(p));
2026 * See if the non running -deadline tasks on this rq
2027 * can be sent to some other CPU where they can preempt
2028 * and start executing.
2030 static int push_dl_task(struct rq *rq)
2032 struct task_struct *next_task;
2033 struct rq *later_rq;
2036 if (!rq->dl.overloaded)
2039 next_task = pick_next_pushable_dl_task(rq);
2044 if (WARN_ON(next_task == rq->curr))
2048 * If next_task preempts rq->curr, and rq->curr
2049 * can move away, it makes sense to just reschedule
2050 * without going further in pushing next_task.
2052 if (dl_task(rq->curr) &&
2053 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2054 rq->curr->nr_cpus_allowed > 1) {
2059 /* We might release rq lock */
2060 get_task_struct(next_task);
2062 /* Will lock the rq it'll find */
2063 later_rq = find_lock_later_rq(next_task, rq);
2065 struct task_struct *task;
2068 * We must check all this again, since
2069 * find_lock_later_rq releases rq->lock and it is
2070 * then possible that next_task has migrated.
2072 task = pick_next_pushable_dl_task(rq);
2073 if (task == next_task) {
2075 * The task is still there. We don't try
2076 * again, some other CPU will pull it when ready.
2085 put_task_struct(next_task);
2090 deactivate_task(rq, next_task, 0);
2091 sub_running_bw(&next_task->dl, &rq->dl);
2092 sub_rq_bw(&next_task->dl, &rq->dl);
2093 set_task_cpu(next_task, later_rq->cpu);
2094 add_rq_bw(&next_task->dl, &later_rq->dl);
2097 * Update the later_rq clock here, because the clock is used
2098 * by the cpufreq_update_util() inside __add_running_bw().
2100 update_rq_clock(later_rq);
2101 add_running_bw(&next_task->dl, &later_rq->dl);
2102 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
2105 resched_curr(later_rq);
2107 double_unlock_balance(rq, later_rq);
2110 put_task_struct(next_task);
2115 static void push_dl_tasks(struct rq *rq)
2117 /* push_dl_task() will return true if it moved a -deadline task */
2118 while (push_dl_task(rq))
2122 static void pull_dl_task(struct rq *this_rq)
2124 int this_cpu = this_rq->cpu, cpu;
2125 struct task_struct *p;
2126 bool resched = false;
2128 u64 dmin = LONG_MAX;
2130 if (likely(!dl_overloaded(this_rq)))
2134 * Match the barrier from dl_set_overloaded; this guarantees that if we
2135 * see overloaded we must also see the dlo_mask bit.
2139 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2140 if (this_cpu == cpu)
2143 src_rq = cpu_rq(cpu);
2146 * It looks racy, abd it is! However, as in sched_rt.c,
2147 * we are fine with this.
2149 if (this_rq->dl.dl_nr_running &&
2150 dl_time_before(this_rq->dl.earliest_dl.curr,
2151 src_rq->dl.earliest_dl.next))
2154 /* Might drop this_rq->lock */
2155 double_lock_balance(this_rq, src_rq);
2158 * If there are no more pullable tasks on the
2159 * rq, we're done with it.
2161 if (src_rq->dl.dl_nr_running <= 1)
2164 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2167 * We found a task to be pulled if:
2168 * - it preempts our current (if there's one),
2169 * - it will preempt the last one we pulled (if any).
2171 if (p && dl_time_before(p->dl.deadline, dmin) &&
2172 (!this_rq->dl.dl_nr_running ||
2173 dl_time_before(p->dl.deadline,
2174 this_rq->dl.earliest_dl.curr))) {
2175 WARN_ON(p == src_rq->curr);
2176 WARN_ON(!task_on_rq_queued(p));
2179 * Then we pull iff p has actually an earlier
2180 * deadline than the current task of its runqueue.
2182 if (dl_time_before(p->dl.deadline,
2183 src_rq->curr->dl.deadline))
2188 deactivate_task(src_rq, p, 0);
2189 sub_running_bw(&p->dl, &src_rq->dl);
2190 sub_rq_bw(&p->dl, &src_rq->dl);
2191 set_task_cpu(p, this_cpu);
2192 add_rq_bw(&p->dl, &this_rq->dl);
2193 add_running_bw(&p->dl, &this_rq->dl);
2194 activate_task(this_rq, p, 0);
2195 dmin = p->dl.deadline;
2197 /* Is there any other task even earlier? */
2200 double_unlock_balance(this_rq, src_rq);
2204 resched_curr(this_rq);
2208 * Since the task is not running and a reschedule is not going to happen
2209 * anytime soon on its runqueue, we try pushing it away now.
2211 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2213 if (!task_running(rq, p) &&
2214 !test_tsk_need_resched(rq->curr) &&
2215 p->nr_cpus_allowed > 1 &&
2216 dl_task(rq->curr) &&
2217 (rq->curr->nr_cpus_allowed < 2 ||
2218 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2223 static void set_cpus_allowed_dl(struct task_struct *p,
2224 const struct cpumask *new_mask)
2226 struct root_domain *src_rd;
2229 BUG_ON(!dl_task(p));
2234 * Migrating a SCHED_DEADLINE task between exclusive
2235 * cpusets (different root_domains) entails a bandwidth
2236 * update. We already made space for us in the destination
2237 * domain (see cpuset_can_attach()).
2239 if (!cpumask_intersects(src_rd->span, new_mask)) {
2240 struct dl_bw *src_dl_b;
2242 src_dl_b = dl_bw_of(cpu_of(rq));
2244 * We now free resources of the root_domain we are migrating
2245 * off. In the worst case, sched_setattr() may temporary fail
2246 * until we complete the update.
2248 raw_spin_lock(&src_dl_b->lock);
2249 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2250 raw_spin_unlock(&src_dl_b->lock);
2253 set_cpus_allowed_common(p, new_mask);
2256 /* Assumes rq->lock is held */
2257 static void rq_online_dl(struct rq *rq)
2259 if (rq->dl.overloaded)
2260 dl_set_overload(rq);
2262 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2263 if (rq->dl.dl_nr_running > 0)
2264 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2267 /* Assumes rq->lock is held */
2268 static void rq_offline_dl(struct rq *rq)
2270 if (rq->dl.overloaded)
2271 dl_clear_overload(rq);
2273 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2274 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2277 void __init init_sched_dl_class(void)
2281 for_each_possible_cpu(i)
2282 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2283 GFP_KERNEL, cpu_to_node(i));
2286 #endif /* CONFIG_SMP */
2288 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2291 * task_non_contending() can start the "inactive timer" (if the 0-lag
2292 * time is in the future). If the task switches back to dl before
2293 * the "inactive timer" fires, it can continue to consume its current
2294 * runtime using its current deadline. If it stays outside of
2295 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2296 * will reset the task parameters.
2298 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2299 task_non_contending(p);
2301 if (!task_on_rq_queued(p)) {
2303 * Inactive timer is armed. However, p is leaving DEADLINE and
2304 * might migrate away from this rq while continuing to run on
2305 * some other class. We need to remove its contribution from
2306 * this rq running_bw now, or sub_rq_bw (below) will complain.
2308 if (p->dl.dl_non_contending)
2309 sub_running_bw(&p->dl, &rq->dl);
2310 sub_rq_bw(&p->dl, &rq->dl);
2314 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2315 * at the 0-lag time, because the task could have been migrated
2316 * while SCHED_OTHER in the meanwhile.
2318 if (p->dl.dl_non_contending)
2319 p->dl.dl_non_contending = 0;
2322 * Since this might be the only -deadline task on the rq,
2323 * this is the right place to try to pull some other one
2324 * from an overloaded CPU, if any.
2326 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2329 deadline_queue_pull_task(rq);
2333 * When switching to -deadline, we may overload the rq, then
2334 * we try to push someone off, if possible.
2336 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2338 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2341 /* If p is not queued we will update its parameters at next wakeup. */
2342 if (!task_on_rq_queued(p)) {
2343 add_rq_bw(&p->dl, &rq->dl);
2348 if (rq->curr != p) {
2350 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2351 deadline_queue_push_tasks(rq);
2353 if (dl_task(rq->curr))
2354 check_preempt_curr_dl(rq, p, 0);
2361 * If the scheduling parameters of a -deadline task changed,
2362 * a push or pull operation might be needed.
2364 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2367 if (task_on_rq_queued(p) || rq->curr == p) {
2370 * This might be too much, but unfortunately
2371 * we don't have the old deadline value, and
2372 * we can't argue if the task is increasing
2373 * or lowering its prio, so...
2375 if (!rq->dl.overloaded)
2376 deadline_queue_pull_task(rq);
2379 * If we now have a earlier deadline task than p,
2380 * then reschedule, provided p is still on this
2383 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2387 * Again, we don't know if p has a earlier
2388 * or later deadline, so let's blindly set a
2389 * (maybe not needed) rescheduling point.
2392 #endif /* CONFIG_SMP */
2396 const struct sched_class dl_sched_class = {
2397 .next = &rt_sched_class,
2398 .enqueue_task = enqueue_task_dl,
2399 .dequeue_task = dequeue_task_dl,
2400 .yield_task = yield_task_dl,
2402 .check_preempt_curr = check_preempt_curr_dl,
2404 .pick_next_task = pick_next_task_dl,
2405 .put_prev_task = put_prev_task_dl,
2408 .select_task_rq = select_task_rq_dl,
2409 .migrate_task_rq = migrate_task_rq_dl,
2410 .set_cpus_allowed = set_cpus_allowed_dl,
2411 .rq_online = rq_online_dl,
2412 .rq_offline = rq_offline_dl,
2413 .task_woken = task_woken_dl,
2416 .set_curr_task = set_curr_task_dl,
2417 .task_tick = task_tick_dl,
2418 .task_fork = task_fork_dl,
2420 .prio_changed = prio_changed_dl,
2421 .switched_from = switched_from_dl,
2422 .switched_to = switched_to_dl,
2424 .update_curr = update_curr_dl,
2427 int sched_dl_global_validate(void)
2429 u64 runtime = global_rt_runtime();
2430 u64 period = global_rt_period();
2431 u64 new_bw = to_ratio(period, runtime);
2434 unsigned long flags;
2437 * Here we want to check the bandwidth not being set to some
2438 * value smaller than the currently allocated bandwidth in
2439 * any of the root_domains.
2441 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2442 * cycling on root_domains... Discussion on different/better
2443 * solutions is welcome!
2445 for_each_possible_cpu(cpu) {
2446 rcu_read_lock_sched();
2447 dl_b = dl_bw_of(cpu);
2449 raw_spin_lock_irqsave(&dl_b->lock, flags);
2450 if (new_bw < dl_b->total_bw)
2452 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2454 rcu_read_unlock_sched();
2463 void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2465 if (global_rt_runtime() == RUNTIME_INF) {
2466 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2467 dl_rq->extra_bw = 1 << BW_SHIFT;
2469 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2470 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2471 dl_rq->extra_bw = to_ratio(global_rt_period(),
2472 global_rt_runtime());
2476 void sched_dl_do_global(void)
2481 unsigned long flags;
2483 def_dl_bandwidth.dl_period = global_rt_period();
2484 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2486 if (global_rt_runtime() != RUNTIME_INF)
2487 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2490 * FIXME: As above...
2492 for_each_possible_cpu(cpu) {
2493 rcu_read_lock_sched();
2494 dl_b = dl_bw_of(cpu);
2496 raw_spin_lock_irqsave(&dl_b->lock, flags);
2498 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2500 rcu_read_unlock_sched();
2501 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2506 * We must be sure that accepting a new task (or allowing changing the
2507 * parameters of an existing one) is consistent with the bandwidth
2508 * constraints. If yes, this function also accordingly updates the currently
2509 * allocated bandwidth to reflect the new situation.
2511 * This function is called while holding p's rq->lock.
2513 int sched_dl_overflow(struct task_struct *p, int policy,
2514 const struct sched_attr *attr)
2516 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2517 u64 period = attr->sched_period ?: attr->sched_deadline;
2518 u64 runtime = attr->sched_runtime;
2519 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2522 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2525 /* !deadline task may carry old deadline bandwidth */
2526 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2530 * Either if a task, enters, leave, or stays -deadline but changes
2531 * its parameters, we may need to update accordingly the total
2532 * allocated bandwidth of the container.
2534 raw_spin_lock(&dl_b->lock);
2535 cpus = dl_bw_cpus(task_cpu(p));
2536 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2537 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2538 if (hrtimer_active(&p->dl.inactive_timer))
2539 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2540 __dl_add(dl_b, new_bw, cpus);
2542 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2543 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2545 * XXX this is slightly incorrect: when the task
2546 * utilization decreases, we should delay the total
2547 * utilization change until the task's 0-lag point.
2548 * But this would require to set the task's "inactive
2549 * timer" when the task is not inactive.
2551 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2552 __dl_add(dl_b, new_bw, cpus);
2553 dl_change_utilization(p, new_bw);
2555 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2557 * Do not decrease the total deadline utilization here,
2558 * switched_from_dl() will take care to do it at the correct
2563 raw_spin_unlock(&dl_b->lock);
2569 * This function initializes the sched_dl_entity of a newly becoming
2570 * SCHED_DEADLINE task.
2572 * Only the static values are considered here, the actual runtime and the
2573 * absolute deadline will be properly calculated when the task is enqueued
2574 * for the first time with its new policy.
2576 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2578 struct sched_dl_entity *dl_se = &p->dl;
2580 dl_se->dl_runtime = attr->sched_runtime;
2581 dl_se->dl_deadline = attr->sched_deadline;
2582 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2583 dl_se->flags = attr->sched_flags;
2584 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2585 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2588 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2590 struct sched_dl_entity *dl_se = &p->dl;
2592 attr->sched_priority = p->rt_priority;
2593 attr->sched_runtime = dl_se->dl_runtime;
2594 attr->sched_deadline = dl_se->dl_deadline;
2595 attr->sched_period = dl_se->dl_period;
2596 attr->sched_flags = dl_se->flags;
2600 * This function validates the new parameters of a -deadline task.
2601 * We ask for the deadline not being zero, and greater or equal
2602 * than the runtime, as well as the period of being zero or
2603 * greater than deadline. Furthermore, we have to be sure that
2604 * user parameters are above the internal resolution of 1us (we
2605 * check sched_runtime only since it is always the smaller one) and
2606 * below 2^63 ns (we have to check both sched_deadline and
2607 * sched_period, as the latter can be zero).
2609 bool __checkparam_dl(const struct sched_attr *attr)
2611 /* special dl tasks don't actually use any parameter */
2612 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2616 if (attr->sched_deadline == 0)
2620 * Since we truncate DL_SCALE bits, make sure we're at least
2623 if (attr->sched_runtime < (1ULL << DL_SCALE))
2627 * Since we use the MSB for wrap-around and sign issues, make
2628 * sure it's not set (mind that period can be equal to zero).
2630 if (attr->sched_deadline & (1ULL << 63) ||
2631 attr->sched_period & (1ULL << 63))
2634 /* runtime <= deadline <= period (if period != 0) */
2635 if ((attr->sched_period != 0 &&
2636 attr->sched_period < attr->sched_deadline) ||
2637 attr->sched_deadline < attr->sched_runtime)
2644 * This function clears the sched_dl_entity static params.
2646 void __dl_clear_params(struct task_struct *p)
2648 struct sched_dl_entity *dl_se = &p->dl;
2650 dl_se->dl_runtime = 0;
2651 dl_se->dl_deadline = 0;
2652 dl_se->dl_period = 0;
2655 dl_se->dl_density = 0;
2657 dl_se->dl_throttled = 0;
2658 dl_se->dl_yielded = 0;
2659 dl_se->dl_non_contending = 0;
2660 dl_se->dl_overrun = 0;
2663 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2665 struct sched_dl_entity *dl_se = &p->dl;
2667 if (dl_se->dl_runtime != attr->sched_runtime ||
2668 dl_se->dl_deadline != attr->sched_deadline ||
2669 dl_se->dl_period != attr->sched_period ||
2670 dl_se->flags != attr->sched_flags)
2677 int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2679 unsigned int dest_cpu;
2683 unsigned long flags;
2685 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2687 rcu_read_lock_sched();
2688 dl_b = dl_bw_of(dest_cpu);
2689 raw_spin_lock_irqsave(&dl_b->lock, flags);
2690 cpus = dl_bw_cpus(dest_cpu);
2691 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
2696 * We reserve space for this task in the destination
2697 * root_domain, as we can't fail after this point.
2698 * We will free resources in the source root_domain
2699 * later on (see set_cpus_allowed_dl()).
2701 __dl_add(dl_b, p->dl.dl_bw, cpus);
2704 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2705 rcu_read_unlock_sched();
2710 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2711 const struct cpumask *trial)
2713 int ret = 1, trial_cpus;
2714 struct dl_bw *cur_dl_b;
2715 unsigned long flags;
2717 rcu_read_lock_sched();
2718 cur_dl_b = dl_bw_of(cpumask_any(cur));
2719 trial_cpus = cpumask_weight(trial);
2721 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2722 if (cur_dl_b->bw != -1 &&
2723 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2725 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2726 rcu_read_unlock_sched();
2731 bool dl_cpu_busy(unsigned int cpu)
2733 unsigned long flags;
2738 rcu_read_lock_sched();
2739 dl_b = dl_bw_of(cpu);
2740 raw_spin_lock_irqsave(&dl_b->lock, flags);
2741 cpus = dl_bw_cpus(cpu);
2742 overflow = __dl_overflow(dl_b, cpus, 0, 0);
2743 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2744 rcu_read_unlock_sched();
2750 #ifdef CONFIG_SCHED_DEBUG
2751 void print_dl_stats(struct seq_file *m, int cpu)
2753 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2755 #endif /* CONFIG_SCHED_DEBUG */