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(hrtimer_active(&dl_se->inactive_timer));
256 WARN_ON(dl_se->dl_non_contending);
258 zerolag_time = dl_se->deadline -
259 div64_long((dl_se->runtime * dl_se->dl_period),
263 * Using relative times instead of the absolute "0-lag time"
264 * allows to simplify the code
266 zerolag_time -= rq_clock(rq);
269 * If the "0-lag time" already passed, decrease the active
270 * utilization now, instead of starting a timer
272 if (zerolag_time < 0) {
274 sub_running_bw(dl_se, dl_rq);
275 if (!dl_task(p) || p->state == TASK_DEAD) {
276 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
278 if (p->state == TASK_DEAD)
279 sub_rq_bw(&p->dl, &rq->dl);
280 raw_spin_lock(&dl_b->lock);
281 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
282 __dl_clear_params(p);
283 raw_spin_unlock(&dl_b->lock);
289 dl_se->dl_non_contending = 1;
291 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL);
294 static void task_contending(struct sched_dl_entity *dl_se, int flags)
296 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
299 * If this is a non-deadline task that has been boosted,
302 if (dl_se->dl_runtime == 0)
305 if (flags & ENQUEUE_MIGRATED)
306 add_rq_bw(dl_se, dl_rq);
308 if (dl_se->dl_non_contending) {
309 dl_se->dl_non_contending = 0;
311 * If the timer handler is currently running and the
312 * timer cannot be cancelled, inactive_task_timer()
313 * will see that dl_not_contending is not set, and
314 * will not touch the rq's active utilization,
315 * so we are still safe.
317 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
318 put_task_struct(dl_task_of(dl_se));
321 * Since "dl_non_contending" is not set, the
322 * task's utilization has already been removed from
323 * active utilization (either when the task blocked,
324 * when the "inactive timer" fired).
327 add_running_bw(dl_se, dl_rq);
331 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
333 struct sched_dl_entity *dl_se = &p->dl;
335 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
338 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
340 raw_spin_lock_init(&dl_b->dl_runtime_lock);
341 dl_b->dl_period = period;
342 dl_b->dl_runtime = runtime;
345 void init_dl_bw(struct dl_bw *dl_b)
347 raw_spin_lock_init(&dl_b->lock);
348 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
349 if (global_rt_runtime() == RUNTIME_INF)
352 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
353 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
357 void init_dl_rq(struct dl_rq *dl_rq)
359 dl_rq->root = RB_ROOT_CACHED;
362 /* zero means no -deadline tasks */
363 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
365 dl_rq->dl_nr_migratory = 0;
366 dl_rq->overloaded = 0;
367 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
369 init_dl_bw(&dl_rq->dl_bw);
372 dl_rq->running_bw = 0;
374 init_dl_rq_bw_ratio(dl_rq);
379 static inline int dl_overloaded(struct rq *rq)
381 return atomic_read(&rq->rd->dlo_count);
384 static inline void dl_set_overload(struct rq *rq)
389 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
391 * Must be visible before the overload count is
392 * set (as in sched_rt.c).
394 * Matched by the barrier in pull_dl_task().
397 atomic_inc(&rq->rd->dlo_count);
400 static inline void dl_clear_overload(struct rq *rq)
405 atomic_dec(&rq->rd->dlo_count);
406 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
409 static void update_dl_migration(struct dl_rq *dl_rq)
411 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
412 if (!dl_rq->overloaded) {
413 dl_set_overload(rq_of_dl_rq(dl_rq));
414 dl_rq->overloaded = 1;
416 } else if (dl_rq->overloaded) {
417 dl_clear_overload(rq_of_dl_rq(dl_rq));
418 dl_rq->overloaded = 0;
422 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
424 struct task_struct *p = dl_task_of(dl_se);
426 if (p->nr_cpus_allowed > 1)
427 dl_rq->dl_nr_migratory++;
429 update_dl_migration(dl_rq);
432 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
434 struct task_struct *p = dl_task_of(dl_se);
436 if (p->nr_cpus_allowed > 1)
437 dl_rq->dl_nr_migratory--;
439 update_dl_migration(dl_rq);
443 * The list of pushable -deadline task is not a plist, like in
444 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
446 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
448 struct dl_rq *dl_rq = &rq->dl;
449 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node;
450 struct rb_node *parent = NULL;
451 struct task_struct *entry;
452 bool leftmost = true;
454 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
458 entry = rb_entry(parent, struct task_struct,
460 if (dl_entity_preempt(&p->dl, &entry->dl))
461 link = &parent->rb_left;
463 link = &parent->rb_right;
469 dl_rq->earliest_dl.next = p->dl.deadline;
471 rb_link_node(&p->pushable_dl_tasks, parent, link);
472 rb_insert_color_cached(&p->pushable_dl_tasks,
473 &dl_rq->pushable_dl_tasks_root, leftmost);
476 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
478 struct dl_rq *dl_rq = &rq->dl;
480 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
483 if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) {
484 struct rb_node *next_node;
486 next_node = rb_next(&p->pushable_dl_tasks);
488 dl_rq->earliest_dl.next = rb_entry(next_node,
489 struct task_struct, pushable_dl_tasks)->dl.deadline;
493 rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
494 RB_CLEAR_NODE(&p->pushable_dl_tasks);
497 static inline int has_pushable_dl_tasks(struct rq *rq)
499 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
502 static int push_dl_task(struct rq *rq);
504 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
506 return dl_task(prev);
509 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
510 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
512 static void push_dl_tasks(struct rq *);
513 static void pull_dl_task(struct rq *);
515 static inline void deadline_queue_push_tasks(struct rq *rq)
517 if (!has_pushable_dl_tasks(rq))
520 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
523 static inline void deadline_queue_pull_task(struct rq *rq)
525 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
528 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
530 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
532 struct rq *later_rq = NULL;
534 later_rq = find_lock_later_rq(p, rq);
539 * If we cannot preempt any rq, fall back to pick any
542 cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
543 if (cpu >= nr_cpu_ids) {
545 * Failed to find any suitable CPU.
546 * The task will never come back!
548 BUG_ON(dl_bandwidth_enabled());
551 * If admission control is disabled we
552 * try a little harder to let the task
555 cpu = cpumask_any(cpu_active_mask);
557 later_rq = cpu_rq(cpu);
558 double_lock_balance(rq, later_rq);
561 set_task_cpu(p, later_rq->cpu);
562 double_unlock_balance(later_rq, rq);
570 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
575 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
580 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
585 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
589 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
594 static inline void pull_dl_task(struct rq *rq)
598 static inline void deadline_queue_push_tasks(struct rq *rq)
602 static inline void deadline_queue_pull_task(struct rq *rq)
605 #endif /* CONFIG_SMP */
607 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
608 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
609 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
612 * We are being explicitly informed that a new instance is starting,
613 * and this means that:
614 * - the absolute deadline of the entity has to be placed at
615 * current time + relative deadline;
616 * - the runtime of the entity has to be set to the maximum value.
618 * The capability of specifying such event is useful whenever a -deadline
619 * entity wants to (try to!) synchronize its behaviour with the scheduler's
620 * one, and to (try to!) reconcile itself with its own scheduling
623 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
625 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
626 struct rq *rq = rq_of_dl_rq(dl_rq);
628 WARN_ON(dl_se->dl_boosted);
629 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
632 * We are racing with the deadline timer. So, do nothing because
633 * the deadline timer handler will take care of properly recharging
634 * the runtime and postponing the deadline
636 if (dl_se->dl_throttled)
640 * We use the regular wall clock time to set deadlines in the
641 * future; in fact, we must consider execution overheads (time
642 * spent on hardirq context, etc.).
644 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
645 dl_se->runtime = dl_se->dl_runtime;
649 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
650 * possibility of a entity lasting more than what it declared, and thus
651 * exhausting its runtime.
653 * Here we are interested in making runtime overrun possible, but we do
654 * not want a entity which is misbehaving to affect the scheduling of all
656 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
657 * is used, in order to confine each entity within its own bandwidth.
659 * This function deals exactly with that, and ensures that when the runtime
660 * of a entity is replenished, its deadline is also postponed. That ensures
661 * the overrunning entity can't interfere with other entity in the system and
662 * can't make them miss their deadlines. Reasons why this kind of overruns
663 * could happen are, typically, a entity voluntarily trying to overcome its
664 * runtime, or it just underestimated it during sched_setattr().
666 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
667 struct sched_dl_entity *pi_se)
669 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
670 struct rq *rq = rq_of_dl_rq(dl_rq);
672 BUG_ON(pi_se->dl_runtime <= 0);
675 * This could be the case for a !-dl task that is boosted.
676 * Just go with full inherited parameters.
678 if (dl_se->dl_deadline == 0) {
679 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
680 dl_se->runtime = pi_se->dl_runtime;
683 if (dl_se->dl_yielded && dl_se->runtime > 0)
687 * We keep moving the deadline away until we get some
688 * available runtime for the entity. This ensures correct
689 * handling of situations where the runtime overrun is
692 while (dl_se->runtime <= 0) {
693 dl_se->deadline += pi_se->dl_period;
694 dl_se->runtime += pi_se->dl_runtime;
698 * At this point, the deadline really should be "in
699 * the future" with respect to rq->clock. If it's
700 * not, we are, for some reason, lagging too much!
701 * Anyway, after having warn userspace abut that,
702 * we still try to keep the things running by
703 * resetting the deadline and the budget of the
706 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
707 printk_deferred_once("sched: DL replenish lagged too much\n");
708 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
709 dl_se->runtime = pi_se->dl_runtime;
712 if (dl_se->dl_yielded)
713 dl_se->dl_yielded = 0;
714 if (dl_se->dl_throttled)
715 dl_se->dl_throttled = 0;
719 * Here we check if --at time t-- an entity (which is probably being
720 * [re]activated or, in general, enqueued) can use its remaining runtime
721 * and its current deadline _without_ exceeding the bandwidth it is
722 * assigned (function returns true if it can't). We are in fact applying
723 * one of the CBS rules: when a task wakes up, if the residual runtime
724 * over residual deadline fits within the allocated bandwidth, then we
725 * can keep the current (absolute) deadline and residual budget without
726 * disrupting the schedulability of the system. Otherwise, we should
727 * refill the runtime and set the deadline a period in the future,
728 * because keeping the current (absolute) deadline of the task would
729 * result in breaking guarantees promised to other tasks (refer to
730 * Documentation/scheduler/sched-deadline.txt for more informations).
732 * This function returns true if:
734 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
736 * IOW we can't recycle current parameters.
738 * Notice that the bandwidth check is done against the deadline. For
739 * task with deadline equal to period this is the same of using
740 * dl_period instead of dl_deadline in the equation above.
742 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
743 struct sched_dl_entity *pi_se, u64 t)
748 * left and right are the two sides of the equation above,
749 * after a bit of shuffling to use multiplications instead
752 * Note that none of the time values involved in the two
753 * multiplications are absolute: dl_deadline and dl_runtime
754 * are the relative deadline and the maximum runtime of each
755 * instance, runtime is the runtime left for the last instance
756 * and (deadline - t), since t is rq->clock, is the time left
757 * to the (absolute) deadline. Even if overflowing the u64 type
758 * is very unlikely to occur in both cases, here we scale down
759 * as we want to avoid that risk at all. Scaling down by 10
760 * means that we reduce granularity to 1us. We are fine with it,
761 * since this is only a true/false check and, anyway, thinking
762 * of anything below microseconds resolution is actually fiction
763 * (but still we want to give the user that illusion >;).
765 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
766 right = ((dl_se->deadline - t) >> DL_SCALE) *
767 (pi_se->dl_runtime >> DL_SCALE);
769 return dl_time_before(right, left);
773 * Revised wakeup rule [1]: For self-suspending tasks, rather then
774 * re-initializing task's runtime and deadline, the revised wakeup
775 * rule adjusts the task's runtime to avoid the task to overrun its
778 * Reasoning: a task may overrun the density if:
779 * runtime / (deadline - t) > dl_runtime / dl_deadline
781 * Therefore, runtime can be adjusted to:
782 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
784 * In such way that runtime will be equal to the maximum density
785 * the task can use without breaking any rule.
787 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
788 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
791 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
793 u64 laxity = dl_se->deadline - rq_clock(rq);
796 * If the task has deadline < period, and the deadline is in the past,
797 * it should already be throttled before this check.
799 * See update_dl_entity() comments for further details.
801 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
803 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
807 * Regarding the deadline, a task with implicit deadline has a relative
808 * deadline == relative period. A task with constrained deadline has a
809 * relative deadline <= relative period.
811 * We support constrained deadline tasks. However, there are some restrictions
812 * applied only for tasks which do not have an implicit deadline. See
813 * update_dl_entity() to know more about such restrictions.
815 * The dl_is_implicit() returns true if the task has an implicit deadline.
817 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
819 return dl_se->dl_deadline == dl_se->dl_period;
823 * When a deadline entity is placed in the runqueue, its runtime and deadline
824 * might need to be updated. This is done by a CBS wake up rule. There are two
825 * different rules: 1) the original CBS; and 2) the Revisited CBS.
827 * When the task is starting a new period, the Original CBS is used. In this
828 * case, the runtime is replenished and a new absolute deadline is set.
830 * When a task is queued before the begin of the next period, using the
831 * remaining runtime and deadline could make the entity to overflow, see
832 * dl_entity_overflow() to find more about runtime overflow. When such case
833 * is detected, the runtime and deadline need to be updated.
835 * If the task has an implicit deadline, i.e., deadline == period, the Original
836 * CBS is applied. the runtime is replenished and a new absolute deadline is
837 * set, as in the previous cases.
839 * However, the Original CBS does not work properly for tasks with
840 * deadline < period, which are said to have a constrained deadline. By
841 * applying the Original CBS, a constrained deadline task would be able to run
842 * runtime/deadline in a period. With deadline < period, the task would
843 * overrun the runtime/period allowed bandwidth, breaking the admission test.
845 * In order to prevent this misbehave, the Revisited CBS is used for
846 * constrained deadline tasks when a runtime overflow is detected. In the
847 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
848 * the remaining runtime of the task is reduced to avoid runtime overflow.
849 * Please refer to the comments update_dl_revised_wakeup() function to find
850 * more about the Revised CBS rule.
852 static void update_dl_entity(struct sched_dl_entity *dl_se,
853 struct sched_dl_entity *pi_se)
855 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
856 struct rq *rq = rq_of_dl_rq(dl_rq);
858 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
859 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
861 if (unlikely(!dl_is_implicit(dl_se) &&
862 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
863 !dl_se->dl_boosted)){
864 update_dl_revised_wakeup(dl_se, rq);
868 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
869 dl_se->runtime = pi_se->dl_runtime;
873 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
875 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
879 * If the entity depleted all its runtime, and if we want it to sleep
880 * while waiting for some new execution time to become available, we
881 * set the bandwidth replenishment timer to the replenishment instant
882 * and try to activate it.
884 * Notice that it is important for the caller to know if the timer
885 * actually started or not (i.e., the replenishment instant is in
886 * the future or in the past).
888 static int start_dl_timer(struct task_struct *p)
890 struct sched_dl_entity *dl_se = &p->dl;
891 struct hrtimer *timer = &dl_se->dl_timer;
892 struct rq *rq = task_rq(p);
896 lockdep_assert_held(&rq->lock);
899 * We want the timer to fire at the deadline, but considering
900 * that it is actually coming from rq->clock and not from
901 * hrtimer's time base reading.
903 act = ns_to_ktime(dl_next_period(dl_se));
904 now = hrtimer_cb_get_time(timer);
905 delta = ktime_to_ns(now) - rq_clock(rq);
906 act = ktime_add_ns(act, delta);
909 * If the expiry time already passed, e.g., because the value
910 * chosen as the deadline is too small, don't even try to
911 * start the timer in the past!
913 if (ktime_us_delta(act, now) < 0)
917 * !enqueued will guarantee another callback; even if one is already in
918 * progress. This ensures a balanced {get,put}_task_struct().
920 * The race against __run_timer() clearing the enqueued state is
921 * harmless because we're holding task_rq()->lock, therefore the timer
922 * expiring after we've done the check will wait on its task_rq_lock()
923 * and observe our state.
925 if (!hrtimer_is_queued(timer)) {
927 hrtimer_start(timer, act, HRTIMER_MODE_ABS);
934 * This is the bandwidth enforcement timer callback. If here, we know
935 * a task is not on its dl_rq, since the fact that the timer was running
936 * means the task is throttled and needs a runtime replenishment.
938 * However, what we actually do depends on the fact the task is active,
939 * (it is on its rq) or has been removed from there by a call to
940 * dequeue_task_dl(). In the former case we must issue the runtime
941 * replenishment and add the task back to the dl_rq; in the latter, we just
942 * do nothing but clearing dl_throttled, so that runtime and deadline
943 * updating (and the queueing back to dl_rq) will be done by the
944 * next call to enqueue_task_dl().
946 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
948 struct sched_dl_entity *dl_se = container_of(timer,
949 struct sched_dl_entity,
951 struct task_struct *p = dl_task_of(dl_se);
955 rq = task_rq_lock(p, &rf);
958 * The task might have changed its scheduling policy to something
959 * different than SCHED_DEADLINE (through switched_from_dl()).
965 * The task might have been boosted by someone else and might be in the
966 * boosting/deboosting path, its not throttled.
968 if (dl_se->dl_boosted)
972 * Spurious timer due to start_dl_timer() race; or we already received
973 * a replenishment from rt_mutex_setprio().
975 if (!dl_se->dl_throttled)
982 * If the throttle happened during sched-out; like:
989 * __dequeue_task_dl()
992 * We can be both throttled and !queued. Replenish the counter
993 * but do not enqueue -- wait for our wakeup to do that.
995 if (!task_on_rq_queued(p)) {
996 replenish_dl_entity(dl_se, dl_se);
1001 if (unlikely(!rq->online)) {
1003 * If the runqueue is no longer available, migrate the
1004 * task elsewhere. This necessarily changes rq.
1006 lockdep_unpin_lock(&rq->lock, rf.cookie);
1007 rq = dl_task_offline_migration(rq, p);
1008 rf.cookie = lockdep_pin_lock(&rq->lock);
1009 update_rq_clock(rq);
1012 * Now that the task has been migrated to the new RQ and we
1013 * have that locked, proceed as normal and enqueue the task
1019 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1020 if (dl_task(rq->curr))
1021 check_preempt_curr_dl(rq, p, 0);
1027 * Queueing this task back might have overloaded rq, check if we need
1028 * to kick someone away.
1030 if (has_pushable_dl_tasks(rq)) {
1032 * Nothing relies on rq->lock after this, so its safe to drop
1035 rq_unpin_lock(rq, &rf);
1037 rq_repin_lock(rq, &rf);
1042 task_rq_unlock(rq, p, &rf);
1045 * This can free the task_struct, including this hrtimer, do not touch
1046 * anything related to that after this.
1050 return HRTIMER_NORESTART;
1053 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1055 struct hrtimer *timer = &dl_se->dl_timer;
1057 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1058 timer->function = dl_task_timer;
1062 * During the activation, CBS checks if it can reuse the current task's
1063 * runtime and period. If the deadline of the task is in the past, CBS
1064 * cannot use the runtime, and so it replenishes the task. This rule
1065 * works fine for implicit deadline tasks (deadline == period), and the
1066 * CBS was designed for implicit deadline tasks. However, a task with
1067 * constrained deadline (deadine < period) might be awakened after the
1068 * deadline, but before the next period. In this case, replenishing the
1069 * task would allow it to run for runtime / deadline. As in this case
1070 * deadline < period, CBS enables a task to run for more than the
1071 * runtime / period. In a very loaded system, this can cause a domino
1072 * effect, making other tasks miss their deadlines.
1074 * To avoid this problem, in the activation of a constrained deadline
1075 * task after the deadline but before the next period, throttle the
1076 * task and set the replenishing timer to the begin of the next period,
1077 * unless it is boosted.
1079 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1081 struct task_struct *p = dl_task_of(dl_se);
1082 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1084 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1085 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1086 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1088 dl_se->dl_throttled = 1;
1089 if (dl_se->runtime > 0)
1095 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1097 return (dl_se->runtime <= 0);
1100 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1103 * This function implements the GRUB accounting rule:
1104 * according to the GRUB reclaiming algorithm, the runtime is
1105 * not decreased as "dq = -dt", but as
1106 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1107 * where u is the utilization of the task, Umax is the maximum reclaimable
1108 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1109 * as the difference between the "total runqueue utilization" and the
1110 * runqueue active utilization, and Uextra is the (per runqueue) extra
1111 * reclaimable utilization.
1112 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1113 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1115 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1116 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1117 * Since delta is a 64 bit variable, to have an overflow its value
1118 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1119 * So, overflow is not an issue here.
1121 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1123 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1125 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1128 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1129 * we compare u_inact + rq->dl.extra_bw with
1130 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1131 * u_inact + rq->dl.extra_bw can be larger than
1132 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1133 * leading to wrong results)
1135 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1138 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1140 return (delta * u_act) >> BW_SHIFT;
1144 * Update the current task's runtime statistics (provided it is still
1145 * a -deadline task and has not been removed from the dl_rq).
1147 static void update_curr_dl(struct rq *rq)
1149 struct task_struct *curr = rq->curr;
1150 struct sched_dl_entity *dl_se = &curr->dl;
1151 u64 delta_exec, scaled_delta_exec;
1152 int cpu = cpu_of(rq);
1155 if (!dl_task(curr) || !on_dl_rq(dl_se))
1159 * Consumed budget is computed considering the time as
1160 * observed by schedulable tasks (excluding time spent
1161 * in hardirq context, etc.). Deadlines are instead
1162 * computed using hard walltime. This seems to be the more
1163 * natural solution, but the full ramifications of this
1164 * approach need further study.
1166 now = rq_clock_task(rq);
1167 delta_exec = now - curr->se.exec_start;
1168 if (unlikely((s64)delta_exec <= 0)) {
1169 if (unlikely(dl_se->dl_yielded))
1174 schedstat_set(curr->se.statistics.exec_max,
1175 max(curr->se.statistics.exec_max, delta_exec));
1177 curr->se.sum_exec_runtime += delta_exec;
1178 account_group_exec_runtime(curr, delta_exec);
1180 curr->se.exec_start = now;
1181 cgroup_account_cputime(curr, delta_exec);
1183 if (dl_entity_is_special(dl_se))
1187 * For tasks that participate in GRUB, we implement GRUB-PA: the
1188 * spare reclaimed bandwidth is used to clock down frequency.
1190 * For the others, we still need to scale reservation parameters
1191 * according to current frequency and CPU maximum capacity.
1193 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1194 scaled_delta_exec = grub_reclaim(delta_exec,
1198 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1199 unsigned long scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
1201 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1202 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1205 dl_se->runtime -= scaled_delta_exec;
1208 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1209 dl_se->dl_throttled = 1;
1211 /* If requested, inform the user about runtime overruns. */
1212 if (dl_runtime_exceeded(dl_se) &&
1213 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1214 dl_se->dl_overrun = 1;
1216 __dequeue_task_dl(rq, curr, 0);
1217 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
1218 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1220 if (!is_leftmost(curr, &rq->dl))
1225 * Because -- for now -- we share the rt bandwidth, we need to
1226 * account our runtime there too, otherwise actual rt tasks
1227 * would be able to exceed the shared quota.
1229 * Account to the root rt group for now.
1231 * The solution we're working towards is having the RT groups scheduled
1232 * using deadline servers -- however there's a few nasties to figure
1233 * out before that can happen.
1235 if (rt_bandwidth_enabled()) {
1236 struct rt_rq *rt_rq = &rq->rt;
1238 raw_spin_lock(&rt_rq->rt_runtime_lock);
1240 * We'll let actual RT tasks worry about the overflow here, we
1241 * have our own CBS to keep us inline; only account when RT
1242 * bandwidth is relevant.
1244 if (sched_rt_bandwidth_account(rt_rq))
1245 rt_rq->rt_time += delta_exec;
1246 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1250 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1252 struct sched_dl_entity *dl_se = container_of(timer,
1253 struct sched_dl_entity,
1255 struct task_struct *p = dl_task_of(dl_se);
1259 rq = task_rq_lock(p, &rf);
1262 update_rq_clock(rq);
1264 if (!dl_task(p) || p->state == TASK_DEAD) {
1265 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1267 if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1268 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1269 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1270 dl_se->dl_non_contending = 0;
1273 raw_spin_lock(&dl_b->lock);
1274 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1275 raw_spin_unlock(&dl_b->lock);
1276 __dl_clear_params(p);
1280 if (dl_se->dl_non_contending == 0)
1283 sub_running_bw(dl_se, &rq->dl);
1284 dl_se->dl_non_contending = 0;
1286 task_rq_unlock(rq, p, &rf);
1289 return HRTIMER_NORESTART;
1292 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1294 struct hrtimer *timer = &dl_se->inactive_timer;
1296 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1297 timer->function = inactive_task_timer;
1302 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1304 struct rq *rq = rq_of_dl_rq(dl_rq);
1306 if (dl_rq->earliest_dl.curr == 0 ||
1307 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1308 dl_rq->earliest_dl.curr = deadline;
1309 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1313 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1315 struct rq *rq = rq_of_dl_rq(dl_rq);
1318 * Since we may have removed our earliest (and/or next earliest)
1319 * task we must recompute them.
1321 if (!dl_rq->dl_nr_running) {
1322 dl_rq->earliest_dl.curr = 0;
1323 dl_rq->earliest_dl.next = 0;
1324 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1326 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1327 struct sched_dl_entity *entry;
1329 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1330 dl_rq->earliest_dl.curr = entry->deadline;
1331 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1337 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1338 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1340 #endif /* CONFIG_SMP */
1343 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1345 int prio = dl_task_of(dl_se)->prio;
1346 u64 deadline = dl_se->deadline;
1348 WARN_ON(!dl_prio(prio));
1349 dl_rq->dl_nr_running++;
1350 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1352 inc_dl_deadline(dl_rq, deadline);
1353 inc_dl_migration(dl_se, dl_rq);
1357 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1359 int prio = dl_task_of(dl_se)->prio;
1361 WARN_ON(!dl_prio(prio));
1362 WARN_ON(!dl_rq->dl_nr_running);
1363 dl_rq->dl_nr_running--;
1364 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1366 dec_dl_deadline(dl_rq, dl_se->deadline);
1367 dec_dl_migration(dl_se, dl_rq);
1370 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1372 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1373 struct rb_node **link = &dl_rq->root.rb_root.rb_node;
1374 struct rb_node *parent = NULL;
1375 struct sched_dl_entity *entry;
1378 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1382 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
1383 if (dl_time_before(dl_se->deadline, entry->deadline))
1384 link = &parent->rb_left;
1386 link = &parent->rb_right;
1391 rb_link_node(&dl_se->rb_node, parent, link);
1392 rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost);
1394 inc_dl_tasks(dl_se, dl_rq);
1397 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1399 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1401 if (RB_EMPTY_NODE(&dl_se->rb_node))
1404 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1405 RB_CLEAR_NODE(&dl_se->rb_node);
1407 dec_dl_tasks(dl_se, dl_rq);
1411 enqueue_dl_entity(struct sched_dl_entity *dl_se,
1412 struct sched_dl_entity *pi_se, int flags)
1414 BUG_ON(on_dl_rq(dl_se));
1417 * If this is a wakeup or a new instance, the scheduling
1418 * parameters of the task might need updating. Otherwise,
1419 * we want a replenishment of its runtime.
1421 if (flags & ENQUEUE_WAKEUP) {
1422 task_contending(dl_se, flags);
1423 update_dl_entity(dl_se, pi_se);
1424 } else if (flags & ENQUEUE_REPLENISH) {
1425 replenish_dl_entity(dl_se, pi_se);
1426 } else if ((flags & ENQUEUE_RESTORE) &&
1427 dl_time_before(dl_se->deadline,
1428 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1429 setup_new_dl_entity(dl_se);
1432 __enqueue_dl_entity(dl_se);
1435 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1437 __dequeue_dl_entity(dl_se);
1440 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1442 struct task_struct *pi_task = rt_mutex_get_top_task(p);
1443 struct sched_dl_entity *pi_se = &p->dl;
1446 * Use the scheduling parameters of the top pi-waiter task if:
1447 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1448 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1449 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1450 * boosted due to a SCHED_DEADLINE pi-waiter).
1451 * Otherwise we keep our runtime and deadline.
1453 if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) {
1454 pi_se = &pi_task->dl;
1455 } else if (!dl_prio(p->normal_prio)) {
1457 * Special case in which we have a !SCHED_DEADLINE task
1458 * that is going to be deboosted, but exceeds its
1459 * runtime while doing so. No point in replenishing
1460 * it, as it's going to return back to its original
1461 * scheduling class after this.
1463 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
1468 * Check if a constrained deadline task was activated
1469 * after the deadline but before the next period.
1470 * If that is the case, the task will be throttled and
1471 * the replenishment timer will be set to the next period.
1473 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1474 dl_check_constrained_dl(&p->dl);
1476 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1477 add_rq_bw(&p->dl, &rq->dl);
1478 add_running_bw(&p->dl, &rq->dl);
1482 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1483 * its budget it needs a replenishment and, since it now is on
1484 * its rq, the bandwidth timer callback (which clearly has not
1485 * run yet) will take care of this.
1486 * However, the active utilization does not depend on the fact
1487 * that the task is on the runqueue or not (but depends on the
1488 * task's state - in GRUB parlance, "inactive" vs "active contending").
1489 * In other words, even if a task is throttled its utilization must
1490 * be counted in the active utilization; hence, we need to call
1493 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1494 if (flags & ENQUEUE_WAKEUP)
1495 task_contending(&p->dl, flags);
1500 enqueue_dl_entity(&p->dl, pi_se, flags);
1502 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1503 enqueue_pushable_dl_task(rq, p);
1506 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1508 dequeue_dl_entity(&p->dl);
1509 dequeue_pushable_dl_task(rq, p);
1512 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1515 __dequeue_task_dl(rq, p, flags);
1517 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1518 sub_running_bw(&p->dl, &rq->dl);
1519 sub_rq_bw(&p->dl, &rq->dl);
1523 * This check allows to start the inactive timer (or to immediately
1524 * decrease the active utilization, if needed) in two cases:
1525 * when the task blocks and when it is terminating
1526 * (p->state == TASK_DEAD). We can handle the two cases in the same
1527 * way, because from GRUB's point of view the same thing is happening
1528 * (the task moves from "active contending" to "active non contending"
1531 if (flags & DEQUEUE_SLEEP)
1532 task_non_contending(p);
1536 * Yield task semantic for -deadline tasks is:
1538 * get off from the CPU until our next instance, with
1539 * a new runtime. This is of little use now, since we
1540 * don't have a bandwidth reclaiming mechanism. Anyway,
1541 * bandwidth reclaiming is planned for the future, and
1542 * yield_task_dl will indicate that some spare budget
1543 * is available for other task instances to use it.
1545 static void yield_task_dl(struct rq *rq)
1548 * We make the task go to sleep until its current deadline by
1549 * forcing its runtime to zero. This way, update_curr_dl() stops
1550 * it and the bandwidth timer will wake it up and will give it
1551 * new scheduling parameters (thanks to dl_yielded=1).
1553 rq->curr->dl.dl_yielded = 1;
1555 update_rq_clock(rq);
1558 * Tell update_rq_clock() that we've just updated,
1559 * so we don't do microscopic update in schedule()
1560 * and double the fastpath cost.
1562 rq_clock_skip_update(rq);
1567 static int find_later_rq(struct task_struct *task);
1570 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
1572 struct task_struct *curr;
1575 if (sd_flag != SD_BALANCE_WAKE)
1581 curr = READ_ONCE(rq->curr); /* unlocked access */
1584 * If we are dealing with a -deadline task, we must
1585 * decide where to wake it up.
1586 * If it has a later deadline and the current task
1587 * on this rq can't move (provided the waking task
1588 * can!) we prefer to send it somewhere else. On the
1589 * other hand, if it has a shorter deadline, we
1590 * try to make it stay here, it might be important.
1592 if (unlikely(dl_task(curr)) &&
1593 (curr->nr_cpus_allowed < 2 ||
1594 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1595 (p->nr_cpus_allowed > 1)) {
1596 int target = find_later_rq(p);
1599 (dl_time_before(p->dl.deadline,
1600 cpu_rq(target)->dl.earliest_dl.curr) ||
1601 (cpu_rq(target)->dl.dl_nr_running == 0)))
1610 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1614 if (p->state != TASK_WAKING)
1619 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1620 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1621 * rq->lock is not... So, lock it
1623 raw_spin_lock(&rq->lock);
1624 if (p->dl.dl_non_contending) {
1625 sub_running_bw(&p->dl, &rq->dl);
1626 p->dl.dl_non_contending = 0;
1628 * If the timer handler is currently running and the
1629 * timer cannot be cancelled, inactive_task_timer()
1630 * will see that dl_not_contending is not set, and
1631 * will not touch the rq's active utilization,
1632 * so we are still safe.
1634 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1637 sub_rq_bw(&p->dl, &rq->dl);
1638 raw_spin_unlock(&rq->lock);
1641 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1644 * Current can't be migrated, useless to reschedule,
1645 * let's hope p can move out.
1647 if (rq->curr->nr_cpus_allowed == 1 ||
1648 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1652 * p is migratable, so let's not schedule it and
1653 * see if it is pushed or pulled somewhere else.
1655 if (p->nr_cpus_allowed != 1 &&
1656 cpudl_find(&rq->rd->cpudl, p, NULL))
1662 #endif /* CONFIG_SMP */
1665 * Only called when both the current and waking task are -deadline
1668 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1671 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1678 * In the unlikely case current and p have the same deadline
1679 * let us try to decide what's the best thing to do...
1681 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1682 !test_tsk_need_resched(rq->curr))
1683 check_preempt_equal_dl(rq, p);
1684 #endif /* CONFIG_SMP */
1687 #ifdef CONFIG_SCHED_HRTICK
1688 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1690 hrtick_start(rq, p->dl.runtime);
1692 #else /* !CONFIG_SCHED_HRTICK */
1693 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1698 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1699 struct dl_rq *dl_rq)
1701 struct rb_node *left = rb_first_cached(&dl_rq->root);
1706 return rb_entry(left, struct sched_dl_entity, rb_node);
1709 static struct task_struct *
1710 pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1712 struct sched_dl_entity *dl_se;
1713 struct task_struct *p;
1714 struct dl_rq *dl_rq;
1718 if (need_pull_dl_task(rq, prev)) {
1720 * This is OK, because current is on_cpu, which avoids it being
1721 * picked for load-balance and preemption/IRQs are still
1722 * disabled avoiding further scheduler activity on it and we're
1723 * being very careful to re-start the picking loop.
1725 rq_unpin_lock(rq, rf);
1727 rq_repin_lock(rq, rf);
1729 * pull_dl_task() can drop (and re-acquire) rq->lock; this
1730 * means a stop task can slip in, in which case we need to
1731 * re-start task selection.
1733 if (rq->stop && task_on_rq_queued(rq->stop))
1738 * When prev is DL, we may throttle it in put_prev_task().
1739 * So, we update time before we check for dl_nr_running.
1741 if (prev->sched_class == &dl_sched_class)
1744 if (unlikely(!dl_rq->dl_nr_running))
1747 put_prev_task(rq, prev);
1749 dl_se = pick_next_dl_entity(rq, dl_rq);
1752 p = dl_task_of(dl_se);
1753 p->se.exec_start = rq_clock_task(rq);
1755 /* Running task will never be pushed. */
1756 dequeue_pushable_dl_task(rq, p);
1758 if (hrtick_enabled(rq))
1759 start_hrtick_dl(rq, p);
1761 deadline_queue_push_tasks(rq);
1763 if (rq->curr->sched_class != &dl_sched_class)
1764 update_dl_rq_load_avg(rq_clock_task(rq), rq, 0);
1769 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1773 update_dl_rq_load_avg(rq_clock_task(rq), rq, 1);
1774 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1775 enqueue_pushable_dl_task(rq, p);
1779 * scheduler tick hitting a task of our scheduling class.
1781 * NOTE: This function can be called remotely by the tick offload that
1782 * goes along full dynticks. Therefore no local assumption can be made
1783 * and everything must be accessed through the @rq and @curr passed in
1786 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1790 update_dl_rq_load_avg(rq_clock_task(rq), rq, 1);
1792 * Even when we have runtime, update_curr_dl() might have resulted in us
1793 * not being the leftmost task anymore. In that case NEED_RESCHED will
1794 * be set and schedule() will start a new hrtick for the next task.
1796 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
1797 is_leftmost(p, &rq->dl))
1798 start_hrtick_dl(rq, p);
1801 static void task_fork_dl(struct task_struct *p)
1804 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1809 static void set_curr_task_dl(struct rq *rq)
1811 struct task_struct *p = rq->curr;
1813 p->se.exec_start = rq_clock_task(rq);
1815 /* You can't push away the running task */
1816 dequeue_pushable_dl_task(rq, p);
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 (unlikely(next_task == rq->curr)) {
2050 * If next_task preempts rq->curr, and rq->curr
2051 * can move away, it makes sense to just reschedule
2052 * without going further in pushing next_task.
2054 if (dl_task(rq->curr) &&
2055 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2056 rq->curr->nr_cpus_allowed > 1) {
2061 /* We might release rq lock */
2062 get_task_struct(next_task);
2064 /* Will lock the rq it'll find */
2065 later_rq = find_lock_later_rq(next_task, rq);
2067 struct task_struct *task;
2070 * We must check all this again, since
2071 * find_lock_later_rq releases rq->lock and it is
2072 * then possible that next_task has migrated.
2074 task = pick_next_pushable_dl_task(rq);
2075 if (task == next_task) {
2077 * The task is still there. We don't try
2078 * again, some other CPU will pull it when ready.
2087 put_task_struct(next_task);
2092 deactivate_task(rq, next_task, 0);
2093 sub_running_bw(&next_task->dl, &rq->dl);
2094 sub_rq_bw(&next_task->dl, &rq->dl);
2095 set_task_cpu(next_task, later_rq->cpu);
2096 add_rq_bw(&next_task->dl, &later_rq->dl);
2099 * Update the later_rq clock here, because the clock is used
2100 * by the cpufreq_update_util() inside __add_running_bw().
2102 update_rq_clock(later_rq);
2103 add_running_bw(&next_task->dl, &later_rq->dl);
2104 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
2107 resched_curr(later_rq);
2109 double_unlock_balance(rq, later_rq);
2112 put_task_struct(next_task);
2117 static void push_dl_tasks(struct rq *rq)
2119 /* push_dl_task() will return true if it moved a -deadline task */
2120 while (push_dl_task(rq))
2124 static void pull_dl_task(struct rq *this_rq)
2126 int this_cpu = this_rq->cpu, cpu;
2127 struct task_struct *p;
2128 bool resched = false;
2130 u64 dmin = LONG_MAX;
2132 if (likely(!dl_overloaded(this_rq)))
2136 * Match the barrier from dl_set_overloaded; this guarantees that if we
2137 * see overloaded we must also see the dlo_mask bit.
2141 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2142 if (this_cpu == cpu)
2145 src_rq = cpu_rq(cpu);
2148 * It looks racy, abd it is! However, as in sched_rt.c,
2149 * we are fine with this.
2151 if (this_rq->dl.dl_nr_running &&
2152 dl_time_before(this_rq->dl.earliest_dl.curr,
2153 src_rq->dl.earliest_dl.next))
2156 /* Might drop this_rq->lock */
2157 double_lock_balance(this_rq, src_rq);
2160 * If there are no more pullable tasks on the
2161 * rq, we're done with it.
2163 if (src_rq->dl.dl_nr_running <= 1)
2166 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2169 * We found a task to be pulled if:
2170 * - it preempts our current (if there's one),
2171 * - it will preempt the last one we pulled (if any).
2173 if (p && dl_time_before(p->dl.deadline, dmin) &&
2174 (!this_rq->dl.dl_nr_running ||
2175 dl_time_before(p->dl.deadline,
2176 this_rq->dl.earliest_dl.curr))) {
2177 WARN_ON(p == src_rq->curr);
2178 WARN_ON(!task_on_rq_queued(p));
2181 * Then we pull iff p has actually an earlier
2182 * deadline than the current task of its runqueue.
2184 if (dl_time_before(p->dl.deadline,
2185 src_rq->curr->dl.deadline))
2190 deactivate_task(src_rq, p, 0);
2191 sub_running_bw(&p->dl, &src_rq->dl);
2192 sub_rq_bw(&p->dl, &src_rq->dl);
2193 set_task_cpu(p, this_cpu);
2194 add_rq_bw(&p->dl, &this_rq->dl);
2195 add_running_bw(&p->dl, &this_rq->dl);
2196 activate_task(this_rq, p, 0);
2197 dmin = p->dl.deadline;
2199 /* Is there any other task even earlier? */
2202 double_unlock_balance(this_rq, src_rq);
2206 resched_curr(this_rq);
2210 * Since the task is not running and a reschedule is not going to happen
2211 * anytime soon on its runqueue, we try pushing it away now.
2213 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2215 if (!task_running(rq, p) &&
2216 !test_tsk_need_resched(rq->curr) &&
2217 p->nr_cpus_allowed > 1 &&
2218 dl_task(rq->curr) &&
2219 (rq->curr->nr_cpus_allowed < 2 ||
2220 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2225 static void set_cpus_allowed_dl(struct task_struct *p,
2226 const struct cpumask *new_mask)
2228 struct root_domain *src_rd;
2231 BUG_ON(!dl_task(p));
2236 * Migrating a SCHED_DEADLINE task between exclusive
2237 * cpusets (different root_domains) entails a bandwidth
2238 * update. We already made space for us in the destination
2239 * domain (see cpuset_can_attach()).
2241 if (!cpumask_intersects(src_rd->span, new_mask)) {
2242 struct dl_bw *src_dl_b;
2244 src_dl_b = dl_bw_of(cpu_of(rq));
2246 * We now free resources of the root_domain we are migrating
2247 * off. In the worst case, sched_setattr() may temporary fail
2248 * until we complete the update.
2250 raw_spin_lock(&src_dl_b->lock);
2251 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2252 raw_spin_unlock(&src_dl_b->lock);
2255 set_cpus_allowed_common(p, new_mask);
2258 /* Assumes rq->lock is held */
2259 static void rq_online_dl(struct rq *rq)
2261 if (rq->dl.overloaded)
2262 dl_set_overload(rq);
2264 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2265 if (rq->dl.dl_nr_running > 0)
2266 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2269 /* Assumes rq->lock is held */
2270 static void rq_offline_dl(struct rq *rq)
2272 if (rq->dl.overloaded)
2273 dl_clear_overload(rq);
2275 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2276 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2279 void __init init_sched_dl_class(void)
2283 for_each_possible_cpu(i)
2284 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2285 GFP_KERNEL, cpu_to_node(i));
2288 #endif /* CONFIG_SMP */
2290 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2293 * task_non_contending() can start the "inactive timer" (if the 0-lag
2294 * time is in the future). If the task switches back to dl before
2295 * the "inactive timer" fires, it can continue to consume its current
2296 * runtime using its current deadline. If it stays outside of
2297 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2298 * will reset the task parameters.
2300 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2301 task_non_contending(p);
2303 if (!task_on_rq_queued(p)) {
2305 * Inactive timer is armed. However, p is leaving DEADLINE and
2306 * might migrate away from this rq while continuing to run on
2307 * some other class. We need to remove its contribution from
2308 * this rq running_bw now, or sub_rq_bw (below) will complain.
2310 if (p->dl.dl_non_contending)
2311 sub_running_bw(&p->dl, &rq->dl);
2312 sub_rq_bw(&p->dl, &rq->dl);
2316 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2317 * at the 0-lag time, because the task could have been migrated
2318 * while SCHED_OTHER in the meanwhile.
2320 if (p->dl.dl_non_contending)
2321 p->dl.dl_non_contending = 0;
2324 * Since this might be the only -deadline task on the rq,
2325 * this is the right place to try to pull some other one
2326 * from an overloaded CPU, if any.
2328 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2331 deadline_queue_pull_task(rq);
2335 * When switching to -deadline, we may overload the rq, then
2336 * we try to push someone off, if possible.
2338 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2340 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2343 /* If p is not queued we will update its parameters at next wakeup. */
2344 if (!task_on_rq_queued(p)) {
2345 add_rq_bw(&p->dl, &rq->dl);
2350 if (rq->curr != p) {
2352 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2353 deadline_queue_push_tasks(rq);
2355 if (dl_task(rq->curr))
2356 check_preempt_curr_dl(rq, p, 0);
2363 * If the scheduling parameters of a -deadline task changed,
2364 * a push or pull operation might be needed.
2366 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2369 if (task_on_rq_queued(p) || rq->curr == p) {
2372 * This might be too much, but unfortunately
2373 * we don't have the old deadline value, and
2374 * we can't argue if the task is increasing
2375 * or lowering its prio, so...
2377 if (!rq->dl.overloaded)
2378 deadline_queue_pull_task(rq);
2381 * If we now have a earlier deadline task than p,
2382 * then reschedule, provided p is still on this
2385 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2389 * Again, we don't know if p has a earlier
2390 * or later deadline, so let's blindly set a
2391 * (maybe not needed) rescheduling point.
2394 #endif /* CONFIG_SMP */
2398 const struct sched_class dl_sched_class = {
2399 .next = &rt_sched_class,
2400 .enqueue_task = enqueue_task_dl,
2401 .dequeue_task = dequeue_task_dl,
2402 .yield_task = yield_task_dl,
2404 .check_preempt_curr = check_preempt_curr_dl,
2406 .pick_next_task = pick_next_task_dl,
2407 .put_prev_task = put_prev_task_dl,
2410 .select_task_rq = select_task_rq_dl,
2411 .migrate_task_rq = migrate_task_rq_dl,
2412 .set_cpus_allowed = set_cpus_allowed_dl,
2413 .rq_online = rq_online_dl,
2414 .rq_offline = rq_offline_dl,
2415 .task_woken = task_woken_dl,
2418 .set_curr_task = set_curr_task_dl,
2419 .task_tick = task_tick_dl,
2420 .task_fork = task_fork_dl,
2422 .prio_changed = prio_changed_dl,
2423 .switched_from = switched_from_dl,
2424 .switched_to = switched_to_dl,
2426 .update_curr = update_curr_dl,
2429 int sched_dl_global_validate(void)
2431 u64 runtime = global_rt_runtime();
2432 u64 period = global_rt_period();
2433 u64 new_bw = to_ratio(period, runtime);
2436 unsigned long flags;
2439 * Here we want to check the bandwidth not being set to some
2440 * value smaller than the currently allocated bandwidth in
2441 * any of the root_domains.
2443 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2444 * cycling on root_domains... Discussion on different/better
2445 * solutions is welcome!
2447 for_each_possible_cpu(cpu) {
2448 rcu_read_lock_sched();
2449 dl_b = dl_bw_of(cpu);
2451 raw_spin_lock_irqsave(&dl_b->lock, flags);
2452 if (new_bw < dl_b->total_bw)
2454 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2456 rcu_read_unlock_sched();
2465 void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2467 if (global_rt_runtime() == RUNTIME_INF) {
2468 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2469 dl_rq->extra_bw = 1 << BW_SHIFT;
2471 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2472 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2473 dl_rq->extra_bw = to_ratio(global_rt_period(),
2474 global_rt_runtime());
2478 void sched_dl_do_global(void)
2483 unsigned long flags;
2485 def_dl_bandwidth.dl_period = global_rt_period();
2486 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2488 if (global_rt_runtime() != RUNTIME_INF)
2489 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2492 * FIXME: As above...
2494 for_each_possible_cpu(cpu) {
2495 rcu_read_lock_sched();
2496 dl_b = dl_bw_of(cpu);
2498 raw_spin_lock_irqsave(&dl_b->lock, flags);
2500 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2502 rcu_read_unlock_sched();
2503 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2508 * We must be sure that accepting a new task (or allowing changing the
2509 * parameters of an existing one) is consistent with the bandwidth
2510 * constraints. If yes, this function also accordingly updates the currently
2511 * allocated bandwidth to reflect the new situation.
2513 * This function is called while holding p's rq->lock.
2515 int sched_dl_overflow(struct task_struct *p, int policy,
2516 const struct sched_attr *attr)
2518 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2519 u64 period = attr->sched_period ?: attr->sched_deadline;
2520 u64 runtime = attr->sched_runtime;
2521 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2524 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2527 /* !deadline task may carry old deadline bandwidth */
2528 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2532 * Either if a task, enters, leave, or stays -deadline but changes
2533 * its parameters, we may need to update accordingly the total
2534 * allocated bandwidth of the container.
2536 raw_spin_lock(&dl_b->lock);
2537 cpus = dl_bw_cpus(task_cpu(p));
2538 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2539 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2540 if (hrtimer_active(&p->dl.inactive_timer))
2541 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2542 __dl_add(dl_b, new_bw, cpus);
2544 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2545 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2547 * XXX this is slightly incorrect: when the task
2548 * utilization decreases, we should delay the total
2549 * utilization change until the task's 0-lag point.
2550 * But this would require to set the task's "inactive
2551 * timer" when the task is not inactive.
2553 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2554 __dl_add(dl_b, new_bw, cpus);
2555 dl_change_utilization(p, new_bw);
2557 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2559 * Do not decrease the total deadline utilization here,
2560 * switched_from_dl() will take care to do it at the correct
2565 raw_spin_unlock(&dl_b->lock);
2571 * This function initializes the sched_dl_entity of a newly becoming
2572 * SCHED_DEADLINE task.
2574 * Only the static values are considered here, the actual runtime and the
2575 * absolute deadline will be properly calculated when the task is enqueued
2576 * for the first time with its new policy.
2578 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2580 struct sched_dl_entity *dl_se = &p->dl;
2582 dl_se->dl_runtime = attr->sched_runtime;
2583 dl_se->dl_deadline = attr->sched_deadline;
2584 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2585 dl_se->flags = attr->sched_flags;
2586 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2587 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2590 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2592 struct sched_dl_entity *dl_se = &p->dl;
2594 attr->sched_priority = p->rt_priority;
2595 attr->sched_runtime = dl_se->dl_runtime;
2596 attr->sched_deadline = dl_se->dl_deadline;
2597 attr->sched_period = dl_se->dl_period;
2598 attr->sched_flags = dl_se->flags;
2602 * This function validates the new parameters of a -deadline task.
2603 * We ask for the deadline not being zero, and greater or equal
2604 * than the runtime, as well as the period of being zero or
2605 * greater than deadline. Furthermore, we have to be sure that
2606 * user parameters are above the internal resolution of 1us (we
2607 * check sched_runtime only since it is always the smaller one) and
2608 * below 2^63 ns (we have to check both sched_deadline and
2609 * sched_period, as the latter can be zero).
2611 bool __checkparam_dl(const struct sched_attr *attr)
2613 /* special dl tasks don't actually use any parameter */
2614 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2618 if (attr->sched_deadline == 0)
2622 * Since we truncate DL_SCALE bits, make sure we're at least
2625 if (attr->sched_runtime < (1ULL << DL_SCALE))
2629 * Since we use the MSB for wrap-around and sign issues, make
2630 * sure it's not set (mind that period can be equal to zero).
2632 if (attr->sched_deadline & (1ULL << 63) ||
2633 attr->sched_period & (1ULL << 63))
2636 /* runtime <= deadline <= period (if period != 0) */
2637 if ((attr->sched_period != 0 &&
2638 attr->sched_period < attr->sched_deadline) ||
2639 attr->sched_deadline < attr->sched_runtime)
2646 * This function clears the sched_dl_entity static params.
2648 void __dl_clear_params(struct task_struct *p)
2650 struct sched_dl_entity *dl_se = &p->dl;
2652 dl_se->dl_runtime = 0;
2653 dl_se->dl_deadline = 0;
2654 dl_se->dl_period = 0;
2657 dl_se->dl_density = 0;
2659 dl_se->dl_throttled = 0;
2660 dl_se->dl_yielded = 0;
2661 dl_se->dl_non_contending = 0;
2662 dl_se->dl_overrun = 0;
2665 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2667 struct sched_dl_entity *dl_se = &p->dl;
2669 if (dl_se->dl_runtime != attr->sched_runtime ||
2670 dl_se->dl_deadline != attr->sched_deadline ||
2671 dl_se->dl_period != attr->sched_period ||
2672 dl_se->flags != attr->sched_flags)
2679 int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2681 unsigned int dest_cpu;
2685 unsigned long flags;
2687 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2689 rcu_read_lock_sched();
2690 dl_b = dl_bw_of(dest_cpu);
2691 raw_spin_lock_irqsave(&dl_b->lock, flags);
2692 cpus = dl_bw_cpus(dest_cpu);
2693 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
2698 * We reserve space for this task in the destination
2699 * root_domain, as we can't fail after this point.
2700 * We will free resources in the source root_domain
2701 * later on (see set_cpus_allowed_dl()).
2703 __dl_add(dl_b, p->dl.dl_bw, cpus);
2706 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2707 rcu_read_unlock_sched();
2712 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2713 const struct cpumask *trial)
2715 int ret = 1, trial_cpus;
2716 struct dl_bw *cur_dl_b;
2717 unsigned long flags;
2719 rcu_read_lock_sched();
2720 cur_dl_b = dl_bw_of(cpumask_any(cur));
2721 trial_cpus = cpumask_weight(trial);
2723 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2724 if (cur_dl_b->bw != -1 &&
2725 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2727 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2728 rcu_read_unlock_sched();
2733 bool dl_cpu_busy(unsigned int cpu)
2735 unsigned long flags;
2740 rcu_read_lock_sched();
2741 dl_b = dl_bw_of(cpu);
2742 raw_spin_lock_irqsave(&dl_b->lock, flags);
2743 cpus = dl_bw_cpus(cpu);
2744 overflow = __dl_overflow(dl_b, cpus, 0, 0);
2745 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2746 rcu_read_unlock_sched();
2752 #ifdef CONFIG_SCHED_DEBUG
2753 void print_dl_stats(struct seq_file *m, int cpu)
2755 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2757 #endif /* CONFIG_SCHED_DEBUG */