2 * Deadline Scheduling Class (SCHED_DEADLINE)
4 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6 * Tasks that periodically executes their instances for less than their
7 * runtime won't miss any of their deadlines.
8 * Tasks that are not periodic or sporadic or that tries to execute more
9 * than their reserved bandwidth will be slowed down (and may potentially
10 * miss some of their deadlines), and won't affect any other task.
12 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
13 * Juri Lelli <juri.lelli@gmail.com>,
14 * Michael Trimarchi <michael@amarulasolutions.com>,
15 * Fabio Checconi <fchecconi@gmail.com>
19 #include <linux/slab.h>
20 #include <uapi/linux/sched/types.h>
22 struct dl_bandwidth def_dl_bandwidth
;
24 static inline struct task_struct
*dl_task_of(struct sched_dl_entity
*dl_se
)
26 return container_of(dl_se
, struct task_struct
, dl
);
29 static inline struct rq
*rq_of_dl_rq(struct dl_rq
*dl_rq
)
31 return container_of(dl_rq
, struct rq
, dl
);
34 static inline struct dl_rq
*dl_rq_of_se(struct sched_dl_entity
*dl_se
)
36 struct task_struct
*p
= dl_task_of(dl_se
);
37 struct rq
*rq
= task_rq(p
);
42 static inline int on_dl_rq(struct sched_dl_entity
*dl_se
)
44 return !RB_EMPTY_NODE(&dl_se
->rb_node
);
48 static inline struct dl_bw
*dl_bw_of(int i
)
50 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
51 "sched RCU must be held");
52 return &cpu_rq(i
)->rd
->dl_bw
;
55 static inline int dl_bw_cpus(int i
)
57 struct root_domain
*rd
= cpu_rq(i
)->rd
;
60 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
61 "sched RCU must be held");
62 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
68 static inline struct dl_bw
*dl_bw_of(int i
)
70 return &cpu_rq(i
)->dl
.dl_bw
;
73 static inline int dl_bw_cpus(int i
)
80 void add_running_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
82 u64 old
= dl_rq
->running_bw
;
84 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
85 dl_rq
->running_bw
+= dl_bw
;
86 SCHED_WARN_ON(dl_rq
->running_bw
< old
); /* overflow */
87 SCHED_WARN_ON(dl_rq
->running_bw
> dl_rq
->this_bw
);
91 void sub_running_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
93 u64 old
= dl_rq
->running_bw
;
95 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
96 dl_rq
->running_bw
-= dl_bw
;
97 SCHED_WARN_ON(dl_rq
->running_bw
> old
); /* underflow */
98 if (dl_rq
->running_bw
> old
)
99 dl_rq
->running_bw
= 0;
103 void add_rq_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
105 u64 old
= dl_rq
->this_bw
;
107 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
108 dl_rq
->this_bw
+= dl_bw
;
109 SCHED_WARN_ON(dl_rq
->this_bw
< old
); /* overflow */
113 void sub_rq_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
115 u64 old
= dl_rq
->this_bw
;
117 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
118 dl_rq
->this_bw
-= dl_bw
;
119 SCHED_WARN_ON(dl_rq
->this_bw
> old
); /* underflow */
120 if (dl_rq
->this_bw
> old
)
122 SCHED_WARN_ON(dl_rq
->running_bw
> dl_rq
->this_bw
);
125 void dl_change_utilization(struct task_struct
*p
, u64 new_bw
)
129 if (task_on_rq_queued(p
))
133 if (p
->dl
.dl_non_contending
) {
134 sub_running_bw(p
->dl
.dl_bw
, &rq
->dl
);
135 p
->dl
.dl_non_contending
= 0;
137 * If the timer handler is currently running and the
138 * timer cannot be cancelled, inactive_task_timer()
139 * will see that dl_not_contending is not set, and
140 * will not touch the rq's active utilization,
141 * so we are still safe.
143 if (hrtimer_try_to_cancel(&p
->dl
.inactive_timer
) == 1)
146 sub_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
147 add_rq_bw(new_bw
, &rq
->dl
);
151 * The utilization of a task cannot be immediately removed from
152 * the rq active utilization (running_bw) when the task blocks.
153 * Instead, we have to wait for the so called "0-lag time".
155 * If a task blocks before the "0-lag time", a timer (the inactive
156 * timer) is armed, and running_bw is decreased when the timer
159 * If the task wakes up again before the inactive timer fires,
160 * the timer is cancelled, whereas if the task wakes up after the
161 * inactive timer fired (and running_bw has been decreased) the
162 * task's utilization has to be added to running_bw again.
163 * A flag in the deadline scheduling entity (dl_non_contending)
164 * is used to avoid race conditions between the inactive timer handler
167 * The following diagram shows how running_bw is updated. A task is
168 * "ACTIVE" when its utilization contributes to running_bw; an
169 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
170 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
171 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
172 * time already passed, which does not contribute to running_bw anymore.
173 * +------------------+
175 * +------------------>+ contending |
176 * | add_running_bw | |
177 * | +----+------+------+
180 * +--------+-------+ | |
181 * | | t >= 0-lag | | wakeup
182 * | INACTIVE |<---------------+ |
183 * | | sub_running_bw | |
184 * +--------+-------+ | |
189 * | +----+------+------+
190 * | sub_running_bw | ACTIVE |
191 * +-------------------+ |
192 * inactive timer | non contending |
193 * fired +------------------+
195 * The task_non_contending() function is invoked when a task
196 * blocks, and checks if the 0-lag time already passed or
197 * not (in the first case, it directly updates running_bw;
198 * in the second case, it arms the inactive timer).
200 * The task_contending() function is invoked when a task wakes
201 * up, and checks if the task is still in the "ACTIVE non contending"
202 * state or not (in the second case, it updates running_bw).
204 static void task_non_contending(struct task_struct
*p
)
206 struct sched_dl_entity
*dl_se
= &p
->dl
;
207 struct hrtimer
*timer
= &dl_se
->inactive_timer
;
208 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
209 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
213 * If this is a non-deadline task that has been boosted,
216 if (dl_se
->dl_runtime
== 0)
219 WARN_ON(hrtimer_active(&dl_se
->inactive_timer
));
220 WARN_ON(dl_se
->dl_non_contending
);
222 zerolag_time
= dl_se
->deadline
-
223 div64_long((dl_se
->runtime
* dl_se
->dl_period
),
227 * Using relative times instead of the absolute "0-lag time"
228 * allows to simplify the code
230 zerolag_time
-= rq_clock(rq
);
233 * If the "0-lag time" already passed, decrease the active
234 * utilization now, instead of starting a timer
236 if (zerolag_time
< 0) {
238 sub_running_bw(dl_se
->dl_bw
, dl_rq
);
239 if (!dl_task(p
) || p
->state
== TASK_DEAD
) {
240 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
242 if (p
->state
== TASK_DEAD
)
243 sub_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
244 raw_spin_lock(&dl_b
->lock
);
245 __dl_clear(dl_b
, p
->dl
.dl_bw
, dl_bw_cpus(task_cpu(p
)));
246 __dl_clear_params(p
);
247 raw_spin_unlock(&dl_b
->lock
);
253 dl_se
->dl_non_contending
= 1;
255 hrtimer_start(timer
, ns_to_ktime(zerolag_time
), HRTIMER_MODE_REL
);
258 static void task_contending(struct sched_dl_entity
*dl_se
, int flags
)
260 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
263 * If this is a non-deadline task that has been boosted,
266 if (dl_se
->dl_runtime
== 0)
269 if (flags
& ENQUEUE_MIGRATED
)
270 add_rq_bw(dl_se
->dl_bw
, dl_rq
);
272 if (dl_se
->dl_non_contending
) {
273 dl_se
->dl_non_contending
= 0;
275 * If the timer handler is currently running and the
276 * timer cannot be cancelled, inactive_task_timer()
277 * will see that dl_not_contending is not set, and
278 * will not touch the rq's active utilization,
279 * so we are still safe.
281 if (hrtimer_try_to_cancel(&dl_se
->inactive_timer
) == 1)
282 put_task_struct(dl_task_of(dl_se
));
285 * Since "dl_non_contending" is not set, the
286 * task's utilization has already been removed from
287 * active utilization (either when the task blocked,
288 * when the "inactive timer" fired).
291 add_running_bw(dl_se
->dl_bw
, dl_rq
);
295 static inline int is_leftmost(struct task_struct
*p
, struct dl_rq
*dl_rq
)
297 struct sched_dl_entity
*dl_se
= &p
->dl
;
299 return dl_rq
->rb_leftmost
== &dl_se
->rb_node
;
302 void init_dl_bandwidth(struct dl_bandwidth
*dl_b
, u64 period
, u64 runtime
)
304 raw_spin_lock_init(&dl_b
->dl_runtime_lock
);
305 dl_b
->dl_period
= period
;
306 dl_b
->dl_runtime
= runtime
;
309 void init_dl_bw(struct dl_bw
*dl_b
)
311 raw_spin_lock_init(&dl_b
->lock
);
312 raw_spin_lock(&def_dl_bandwidth
.dl_runtime_lock
);
313 if (global_rt_runtime() == RUNTIME_INF
)
316 dl_b
->bw
= to_ratio(global_rt_period(), global_rt_runtime());
317 raw_spin_unlock(&def_dl_bandwidth
.dl_runtime_lock
);
321 void init_dl_rq(struct dl_rq
*dl_rq
)
323 dl_rq
->rb_root
= RB_ROOT
;
326 /* zero means no -deadline tasks */
327 dl_rq
->earliest_dl
.curr
= dl_rq
->earliest_dl
.next
= 0;
329 dl_rq
->dl_nr_migratory
= 0;
330 dl_rq
->overloaded
= 0;
331 dl_rq
->pushable_dl_tasks_root
= RB_ROOT
;
333 init_dl_bw(&dl_rq
->dl_bw
);
336 dl_rq
->running_bw
= 0;
338 init_dl_rq_bw_ratio(dl_rq
);
343 static inline int dl_overloaded(struct rq
*rq
)
345 return atomic_read(&rq
->rd
->dlo_count
);
348 static inline void dl_set_overload(struct rq
*rq
)
353 cpumask_set_cpu(rq
->cpu
, rq
->rd
->dlo_mask
);
355 * Must be visible before the overload count is
356 * set (as in sched_rt.c).
358 * Matched by the barrier in pull_dl_task().
361 atomic_inc(&rq
->rd
->dlo_count
);
364 static inline void dl_clear_overload(struct rq
*rq
)
369 atomic_dec(&rq
->rd
->dlo_count
);
370 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->dlo_mask
);
373 static void update_dl_migration(struct dl_rq
*dl_rq
)
375 if (dl_rq
->dl_nr_migratory
&& dl_rq
->dl_nr_running
> 1) {
376 if (!dl_rq
->overloaded
) {
377 dl_set_overload(rq_of_dl_rq(dl_rq
));
378 dl_rq
->overloaded
= 1;
380 } else if (dl_rq
->overloaded
) {
381 dl_clear_overload(rq_of_dl_rq(dl_rq
));
382 dl_rq
->overloaded
= 0;
386 static void inc_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
388 struct task_struct
*p
= dl_task_of(dl_se
);
390 if (p
->nr_cpus_allowed
> 1)
391 dl_rq
->dl_nr_migratory
++;
393 update_dl_migration(dl_rq
);
396 static void dec_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
398 struct task_struct
*p
= dl_task_of(dl_se
);
400 if (p
->nr_cpus_allowed
> 1)
401 dl_rq
->dl_nr_migratory
--;
403 update_dl_migration(dl_rq
);
407 * The list of pushable -deadline task is not a plist, like in
408 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
410 static void enqueue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
412 struct dl_rq
*dl_rq
= &rq
->dl
;
413 struct rb_node
**link
= &dl_rq
->pushable_dl_tasks_root
.rb_node
;
414 struct rb_node
*parent
= NULL
;
415 struct task_struct
*entry
;
418 BUG_ON(!RB_EMPTY_NODE(&p
->pushable_dl_tasks
));
422 entry
= rb_entry(parent
, struct task_struct
,
424 if (dl_entity_preempt(&p
->dl
, &entry
->dl
))
425 link
= &parent
->rb_left
;
427 link
= &parent
->rb_right
;
433 dl_rq
->pushable_dl_tasks_leftmost
= &p
->pushable_dl_tasks
;
434 dl_rq
->earliest_dl
.next
= p
->dl
.deadline
;
437 rb_link_node(&p
->pushable_dl_tasks
, parent
, link
);
438 rb_insert_color(&p
->pushable_dl_tasks
, &dl_rq
->pushable_dl_tasks_root
);
441 static void dequeue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
443 struct dl_rq
*dl_rq
= &rq
->dl
;
445 if (RB_EMPTY_NODE(&p
->pushable_dl_tasks
))
448 if (dl_rq
->pushable_dl_tasks_leftmost
== &p
->pushable_dl_tasks
) {
449 struct rb_node
*next_node
;
451 next_node
= rb_next(&p
->pushable_dl_tasks
);
452 dl_rq
->pushable_dl_tasks_leftmost
= next_node
;
454 dl_rq
->earliest_dl
.next
= rb_entry(next_node
,
455 struct task_struct
, pushable_dl_tasks
)->dl
.deadline
;
459 rb_erase(&p
->pushable_dl_tasks
, &dl_rq
->pushable_dl_tasks_root
);
460 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
463 static inline int has_pushable_dl_tasks(struct rq
*rq
)
465 return !RB_EMPTY_ROOT(&rq
->dl
.pushable_dl_tasks_root
);
468 static int push_dl_task(struct rq
*rq
);
470 static inline bool need_pull_dl_task(struct rq
*rq
, struct task_struct
*prev
)
472 return dl_task(prev
);
475 static DEFINE_PER_CPU(struct callback_head
, dl_push_head
);
476 static DEFINE_PER_CPU(struct callback_head
, dl_pull_head
);
478 static void push_dl_tasks(struct rq
*);
479 static void pull_dl_task(struct rq
*);
481 static inline void queue_push_tasks(struct rq
*rq
)
483 if (!has_pushable_dl_tasks(rq
))
486 queue_balance_callback(rq
, &per_cpu(dl_push_head
, rq
->cpu
), push_dl_tasks
);
489 static inline void queue_pull_task(struct rq
*rq
)
491 queue_balance_callback(rq
, &per_cpu(dl_pull_head
, rq
->cpu
), pull_dl_task
);
494 static struct rq
*find_lock_later_rq(struct task_struct
*task
, struct rq
*rq
);
496 static struct rq
*dl_task_offline_migration(struct rq
*rq
, struct task_struct
*p
)
498 struct rq
*later_rq
= NULL
;
500 later_rq
= find_lock_later_rq(p
, rq
);
505 * If we cannot preempt any rq, fall back to pick any
508 cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
509 if (cpu
>= nr_cpu_ids
) {
511 * Fail to find any suitable cpu.
512 * The task will never come back!
514 BUG_ON(dl_bandwidth_enabled());
517 * If admission control is disabled we
518 * try a little harder to let the task
521 cpu
= cpumask_any(cpu_active_mask
);
523 later_rq
= cpu_rq(cpu
);
524 double_lock_balance(rq
, later_rq
);
527 set_task_cpu(p
, later_rq
->cpu
);
528 double_unlock_balance(later_rq
, rq
);
536 void enqueue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
541 void dequeue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
546 void inc_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
551 void dec_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
555 static inline bool need_pull_dl_task(struct rq
*rq
, struct task_struct
*prev
)
560 static inline void pull_dl_task(struct rq
*rq
)
564 static inline void queue_push_tasks(struct rq
*rq
)
568 static inline void queue_pull_task(struct rq
*rq
)
571 #endif /* CONFIG_SMP */
573 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
574 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
575 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
,
579 * We are being explicitly informed that a new instance is starting,
580 * and this means that:
581 * - the absolute deadline of the entity has to be placed at
582 * current time + relative deadline;
583 * - the runtime of the entity has to be set to the maximum value.
585 * The capability of specifying such event is useful whenever a -deadline
586 * entity wants to (try to!) synchronize its behaviour with the scheduler's
587 * one, and to (try to!) reconcile itself with its own scheduling
590 static inline void setup_new_dl_entity(struct sched_dl_entity
*dl_se
)
592 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
593 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
595 WARN_ON(dl_se
->dl_boosted
);
596 WARN_ON(dl_time_before(rq_clock(rq
), dl_se
->deadline
));
599 * We are racing with the deadline timer. So, do nothing because
600 * the deadline timer handler will take care of properly recharging
601 * the runtime and postponing the deadline
603 if (dl_se
->dl_throttled
)
607 * We use the regular wall clock time to set deadlines in the
608 * future; in fact, we must consider execution overheads (time
609 * spent on hardirq context, etc.).
611 dl_se
->deadline
= rq_clock(rq
) + dl_se
->dl_deadline
;
612 dl_se
->runtime
= dl_se
->dl_runtime
;
616 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
617 * possibility of a entity lasting more than what it declared, and thus
618 * exhausting its runtime.
620 * Here we are interested in making runtime overrun possible, but we do
621 * not want a entity which is misbehaving to affect the scheduling of all
623 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
624 * is used, in order to confine each entity within its own bandwidth.
626 * This function deals exactly with that, and ensures that when the runtime
627 * of a entity is replenished, its deadline is also postponed. That ensures
628 * the overrunning entity can't interfere with other entity in the system and
629 * can't make them miss their deadlines. Reasons why this kind of overruns
630 * could happen are, typically, a entity voluntarily trying to overcome its
631 * runtime, or it just underestimated it during sched_setattr().
633 static void replenish_dl_entity(struct sched_dl_entity
*dl_se
,
634 struct sched_dl_entity
*pi_se
)
636 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
637 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
639 BUG_ON(pi_se
->dl_runtime
<= 0);
642 * This could be the case for a !-dl task that is boosted.
643 * Just go with full inherited parameters.
645 if (dl_se
->dl_deadline
== 0) {
646 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
647 dl_se
->runtime
= pi_se
->dl_runtime
;
650 if (dl_se
->dl_yielded
&& dl_se
->runtime
> 0)
654 * We keep moving the deadline away until we get some
655 * available runtime for the entity. This ensures correct
656 * handling of situations where the runtime overrun is
659 while (dl_se
->runtime
<= 0) {
660 dl_se
->deadline
+= pi_se
->dl_period
;
661 dl_se
->runtime
+= pi_se
->dl_runtime
;
665 * At this point, the deadline really should be "in
666 * the future" with respect to rq->clock. If it's
667 * not, we are, for some reason, lagging too much!
668 * Anyway, after having warn userspace abut that,
669 * we still try to keep the things running by
670 * resetting the deadline and the budget of the
673 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
))) {
674 printk_deferred_once("sched: DL replenish lagged too much\n");
675 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
676 dl_se
->runtime
= pi_se
->dl_runtime
;
679 if (dl_se
->dl_yielded
)
680 dl_se
->dl_yielded
= 0;
681 if (dl_se
->dl_throttled
)
682 dl_se
->dl_throttled
= 0;
686 * Here we check if --at time t-- an entity (which is probably being
687 * [re]activated or, in general, enqueued) can use its remaining runtime
688 * and its current deadline _without_ exceeding the bandwidth it is
689 * assigned (function returns true if it can't). We are in fact applying
690 * one of the CBS rules: when a task wakes up, if the residual runtime
691 * over residual deadline fits within the allocated bandwidth, then we
692 * can keep the current (absolute) deadline and residual budget without
693 * disrupting the schedulability of the system. Otherwise, we should
694 * refill the runtime and set the deadline a period in the future,
695 * because keeping the current (absolute) deadline of the task would
696 * result in breaking guarantees promised to other tasks (refer to
697 * Documentation/scheduler/sched-deadline.txt for more informations).
699 * This function returns true if:
701 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
703 * IOW we can't recycle current parameters.
705 * Notice that the bandwidth check is done against the deadline. For
706 * task with deadline equal to period this is the same of using
707 * dl_period instead of dl_deadline in the equation above.
709 static bool dl_entity_overflow(struct sched_dl_entity
*dl_se
,
710 struct sched_dl_entity
*pi_se
, u64 t
)
715 * left and right are the two sides of the equation above,
716 * after a bit of shuffling to use multiplications instead
719 * Note that none of the time values involved in the two
720 * multiplications are absolute: dl_deadline and dl_runtime
721 * are the relative deadline and the maximum runtime of each
722 * instance, runtime is the runtime left for the last instance
723 * and (deadline - t), since t is rq->clock, is the time left
724 * to the (absolute) deadline. Even if overflowing the u64 type
725 * is very unlikely to occur in both cases, here we scale down
726 * as we want to avoid that risk at all. Scaling down by 10
727 * means that we reduce granularity to 1us. We are fine with it,
728 * since this is only a true/false check and, anyway, thinking
729 * of anything below microseconds resolution is actually fiction
730 * (but still we want to give the user that illusion >;).
732 left
= (pi_se
->dl_deadline
>> DL_SCALE
) * (dl_se
->runtime
>> DL_SCALE
);
733 right
= ((dl_se
->deadline
- t
) >> DL_SCALE
) *
734 (pi_se
->dl_runtime
>> DL_SCALE
);
736 return dl_time_before(right
, left
);
740 * Revised wakeup rule [1]: For self-suspending tasks, rather then
741 * re-initializing task's runtime and deadline, the revised wakeup
742 * rule adjusts the task's runtime to avoid the task to overrun its
745 * Reasoning: a task may overrun the density if:
746 * runtime / (deadline - t) > dl_runtime / dl_deadline
748 * Therefore, runtime can be adjusted to:
749 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
751 * In such way that runtime will be equal to the maximum density
752 * the task can use without breaking any rule.
754 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
755 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
758 update_dl_revised_wakeup(struct sched_dl_entity
*dl_se
, struct rq
*rq
)
760 u64 laxity
= dl_se
->deadline
- rq_clock(rq
);
763 * If the task has deadline < period, and the deadline is in the past,
764 * it should already be throttled before this check.
766 * See update_dl_entity() comments for further details.
768 WARN_ON(dl_time_before(dl_se
->deadline
, rq_clock(rq
)));
770 dl_se
->runtime
= (dl_se
->dl_density
* laxity
) >> BW_SHIFT
;
774 * Regarding the deadline, a task with implicit deadline has a relative
775 * deadline == relative period. A task with constrained deadline has a
776 * relative deadline <= relative period.
778 * We support constrained deadline tasks. However, there are some restrictions
779 * applied only for tasks which do not have an implicit deadline. See
780 * update_dl_entity() to know more about such restrictions.
782 * The dl_is_implicit() returns true if the task has an implicit deadline.
784 static inline bool dl_is_implicit(struct sched_dl_entity
*dl_se
)
786 return dl_se
->dl_deadline
== dl_se
->dl_period
;
790 * When a deadline entity is placed in the runqueue, its runtime and deadline
791 * might need to be updated. This is done by a CBS wake up rule. There are two
792 * different rules: 1) the original CBS; and 2) the Revisited CBS.
794 * When the task is starting a new period, the Original CBS is used. In this
795 * case, the runtime is replenished and a new absolute deadline is set.
797 * When a task is queued before the begin of the next period, using the
798 * remaining runtime and deadline could make the entity to overflow, see
799 * dl_entity_overflow() to find more about runtime overflow. When such case
800 * is detected, the runtime and deadline need to be updated.
802 * If the task has an implicit deadline, i.e., deadline == period, the Original
803 * CBS is applied. the runtime is replenished and a new absolute deadline is
804 * set, as in the previous cases.
806 * However, the Original CBS does not work properly for tasks with
807 * deadline < period, which are said to have a constrained deadline. By
808 * applying the Original CBS, a constrained deadline task would be able to run
809 * runtime/deadline in a period. With deadline < period, the task would
810 * overrun the runtime/period allowed bandwidth, breaking the admission test.
812 * In order to prevent this misbehave, the Revisited CBS is used for
813 * constrained deadline tasks when a runtime overflow is detected. In the
814 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
815 * the remaining runtime of the task is reduced to avoid runtime overflow.
816 * Please refer to the comments update_dl_revised_wakeup() function to find
817 * more about the Revised CBS rule.
819 static void update_dl_entity(struct sched_dl_entity
*dl_se
,
820 struct sched_dl_entity
*pi_se
)
822 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
823 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
825 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
)) ||
826 dl_entity_overflow(dl_se
, pi_se
, rq_clock(rq
))) {
828 if (unlikely(!dl_is_implicit(dl_se
) &&
829 !dl_time_before(dl_se
->deadline
, rq_clock(rq
)) &&
830 !dl_se
->dl_boosted
)){
831 update_dl_revised_wakeup(dl_se
, rq
);
835 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
836 dl_se
->runtime
= pi_se
->dl_runtime
;
840 static inline u64
dl_next_period(struct sched_dl_entity
*dl_se
)
842 return dl_se
->deadline
- dl_se
->dl_deadline
+ dl_se
->dl_period
;
846 * If the entity depleted all its runtime, and if we want it to sleep
847 * while waiting for some new execution time to become available, we
848 * set the bandwidth replenishment timer to the replenishment instant
849 * and try to activate it.
851 * Notice that it is important for the caller to know if the timer
852 * actually started or not (i.e., the replenishment instant is in
853 * the future or in the past).
855 static int start_dl_timer(struct task_struct
*p
)
857 struct sched_dl_entity
*dl_se
= &p
->dl
;
858 struct hrtimer
*timer
= &dl_se
->dl_timer
;
859 struct rq
*rq
= task_rq(p
);
863 lockdep_assert_held(&rq
->lock
);
866 * We want the timer to fire at the deadline, but considering
867 * that it is actually coming from rq->clock and not from
868 * hrtimer's time base reading.
870 act
= ns_to_ktime(dl_next_period(dl_se
));
871 now
= hrtimer_cb_get_time(timer
);
872 delta
= ktime_to_ns(now
) - rq_clock(rq
);
873 act
= ktime_add_ns(act
, delta
);
876 * If the expiry time already passed, e.g., because the value
877 * chosen as the deadline is too small, don't even try to
878 * start the timer in the past!
880 if (ktime_us_delta(act
, now
) < 0)
884 * !enqueued will guarantee another callback; even if one is already in
885 * progress. This ensures a balanced {get,put}_task_struct().
887 * The race against __run_timer() clearing the enqueued state is
888 * harmless because we're holding task_rq()->lock, therefore the timer
889 * expiring after we've done the check will wait on its task_rq_lock()
890 * and observe our state.
892 if (!hrtimer_is_queued(timer
)) {
894 hrtimer_start(timer
, act
, HRTIMER_MODE_ABS
);
901 * This is the bandwidth enforcement timer callback. If here, we know
902 * a task is not on its dl_rq, since the fact that the timer was running
903 * means the task is throttled and needs a runtime replenishment.
905 * However, what we actually do depends on the fact the task is active,
906 * (it is on its rq) or has been removed from there by a call to
907 * dequeue_task_dl(). In the former case we must issue the runtime
908 * replenishment and add the task back to the dl_rq; in the latter, we just
909 * do nothing but clearing dl_throttled, so that runtime and deadline
910 * updating (and the queueing back to dl_rq) will be done by the
911 * next call to enqueue_task_dl().
913 static enum hrtimer_restart
dl_task_timer(struct hrtimer
*timer
)
915 struct sched_dl_entity
*dl_se
= container_of(timer
,
916 struct sched_dl_entity
,
918 struct task_struct
*p
= dl_task_of(dl_se
);
922 rq
= task_rq_lock(p
, &rf
);
925 * The task might have changed its scheduling policy to something
926 * different than SCHED_DEADLINE (through switched_from_dl()).
932 * The task might have been boosted by someone else and might be in the
933 * boosting/deboosting path, its not throttled.
935 if (dl_se
->dl_boosted
)
939 * Spurious timer due to start_dl_timer() race; or we already received
940 * a replenishment from rt_mutex_setprio().
942 if (!dl_se
->dl_throttled
)
949 * If the throttle happened during sched-out; like:
956 * __dequeue_task_dl()
959 * We can be both throttled and !queued. Replenish the counter
960 * but do not enqueue -- wait for our wakeup to do that.
962 if (!task_on_rq_queued(p
)) {
963 replenish_dl_entity(dl_se
, dl_se
);
968 if (unlikely(!rq
->online
)) {
970 * If the runqueue is no longer available, migrate the
971 * task elsewhere. This necessarily changes rq.
973 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
974 rq
= dl_task_offline_migration(rq
, p
);
975 rf
.cookie
= lockdep_pin_lock(&rq
->lock
);
979 * Now that the task has been migrated to the new RQ and we
980 * have that locked, proceed as normal and enqueue the task
986 enqueue_task_dl(rq
, p
, ENQUEUE_REPLENISH
);
987 if (dl_task(rq
->curr
))
988 check_preempt_curr_dl(rq
, p
, 0);
994 * Queueing this task back might have overloaded rq, check if we need
995 * to kick someone away.
997 if (has_pushable_dl_tasks(rq
)) {
999 * Nothing relies on rq->lock after this, so its safe to drop
1002 rq_unpin_lock(rq
, &rf
);
1004 rq_repin_lock(rq
, &rf
);
1009 task_rq_unlock(rq
, p
, &rf
);
1012 * This can free the task_struct, including this hrtimer, do not touch
1013 * anything related to that after this.
1017 return HRTIMER_NORESTART
;
1020 void init_dl_task_timer(struct sched_dl_entity
*dl_se
)
1022 struct hrtimer
*timer
= &dl_se
->dl_timer
;
1024 hrtimer_init(timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1025 timer
->function
= dl_task_timer
;
1029 * During the activation, CBS checks if it can reuse the current task's
1030 * runtime and period. If the deadline of the task is in the past, CBS
1031 * cannot use the runtime, and so it replenishes the task. This rule
1032 * works fine for implicit deadline tasks (deadline == period), and the
1033 * CBS was designed for implicit deadline tasks. However, a task with
1034 * constrained deadline (deadine < period) might be awakened after the
1035 * deadline, but before the next period. In this case, replenishing the
1036 * task would allow it to run for runtime / deadline. As in this case
1037 * deadline < period, CBS enables a task to run for more than the
1038 * runtime / period. In a very loaded system, this can cause a domino
1039 * effect, making other tasks miss their deadlines.
1041 * To avoid this problem, in the activation of a constrained deadline
1042 * task after the deadline but before the next period, throttle the
1043 * task and set the replenishing timer to the begin of the next period,
1044 * unless it is boosted.
1046 static inline void dl_check_constrained_dl(struct sched_dl_entity
*dl_se
)
1048 struct task_struct
*p
= dl_task_of(dl_se
);
1049 struct rq
*rq
= rq_of_dl_rq(dl_rq_of_se(dl_se
));
1051 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
)) &&
1052 dl_time_before(rq_clock(rq
), dl_next_period(dl_se
))) {
1053 if (unlikely(dl_se
->dl_boosted
|| !start_dl_timer(p
)))
1055 dl_se
->dl_throttled
= 1;
1056 if (dl_se
->runtime
> 0)
1062 int dl_runtime_exceeded(struct sched_dl_entity
*dl_se
)
1064 return (dl_se
->runtime
<= 0);
1067 extern bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
);
1070 * This function implements the GRUB accounting rule:
1071 * according to the GRUB reclaiming algorithm, the runtime is
1072 * not decreased as "dq = -dt", but as
1073 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1074 * where u is the utilization of the task, Umax is the maximum reclaimable
1075 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1076 * as the difference between the "total runqueue utilization" and the
1077 * runqueue active utilization, and Uextra is the (per runqueue) extra
1078 * reclaimable utilization.
1079 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1080 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1082 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1083 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1084 * Since delta is a 64 bit variable, to have an overflow its value
1085 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1086 * So, overflow is not an issue here.
1088 u64
grub_reclaim(u64 delta
, struct rq
*rq
, struct sched_dl_entity
*dl_se
)
1090 u64 u_inact
= rq
->dl
.this_bw
- rq
->dl
.running_bw
; /* Utot - Uact */
1092 u64 u_act_min
= (dl_se
->dl_bw
* rq
->dl
.bw_ratio
) >> RATIO_SHIFT
;
1095 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1096 * we compare u_inact + rq->dl.extra_bw with
1097 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1098 * u_inact + rq->dl.extra_bw can be larger than
1099 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1100 * leading to wrong results)
1102 if (u_inact
+ rq
->dl
.extra_bw
> BW_UNIT
- u_act_min
)
1105 u_act
= BW_UNIT
- u_inact
- rq
->dl
.extra_bw
;
1107 return (delta
* u_act
) >> BW_SHIFT
;
1111 * Update the current task's runtime statistics (provided it is still
1112 * a -deadline task and has not been removed from the dl_rq).
1114 static void update_curr_dl(struct rq
*rq
)
1116 struct task_struct
*curr
= rq
->curr
;
1117 struct sched_dl_entity
*dl_se
= &curr
->dl
;
1120 if (!dl_task(curr
) || !on_dl_rq(dl_se
))
1124 * Consumed budget is computed considering the time as
1125 * observed by schedulable tasks (excluding time spent
1126 * in hardirq context, etc.). Deadlines are instead
1127 * computed using hard walltime. This seems to be the more
1128 * natural solution, but the full ramifications of this
1129 * approach need further study.
1131 delta_exec
= rq_clock_task(rq
) - curr
->se
.exec_start
;
1132 if (unlikely((s64
)delta_exec
<= 0)) {
1133 if (unlikely(dl_se
->dl_yielded
))
1138 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
1139 cpufreq_update_this_cpu(rq
, SCHED_CPUFREQ_DL
);
1141 schedstat_set(curr
->se
.statistics
.exec_max
,
1142 max(curr
->se
.statistics
.exec_max
, delta_exec
));
1144 curr
->se
.sum_exec_runtime
+= delta_exec
;
1145 account_group_exec_runtime(curr
, delta_exec
);
1147 curr
->se
.exec_start
= rq_clock_task(rq
);
1148 cpuacct_charge(curr
, delta_exec
);
1150 sched_rt_avg_update(rq
, delta_exec
);
1152 if (unlikely(dl_se
->flags
& SCHED_FLAG_RECLAIM
))
1153 delta_exec
= grub_reclaim(delta_exec
, rq
, &curr
->dl
);
1154 dl_se
->runtime
-= delta_exec
;
1157 if (dl_runtime_exceeded(dl_se
) || dl_se
->dl_yielded
) {
1158 dl_se
->dl_throttled
= 1;
1159 __dequeue_task_dl(rq
, curr
, 0);
1160 if (unlikely(dl_se
->dl_boosted
|| !start_dl_timer(curr
)))
1161 enqueue_task_dl(rq
, curr
, ENQUEUE_REPLENISH
);
1163 if (!is_leftmost(curr
, &rq
->dl
))
1168 * Because -- for now -- we share the rt bandwidth, we need to
1169 * account our runtime there too, otherwise actual rt tasks
1170 * would be able to exceed the shared quota.
1172 * Account to the root rt group for now.
1174 * The solution we're working towards is having the RT groups scheduled
1175 * using deadline servers -- however there's a few nasties to figure
1176 * out before that can happen.
1178 if (rt_bandwidth_enabled()) {
1179 struct rt_rq
*rt_rq
= &rq
->rt
;
1181 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
1183 * We'll let actual RT tasks worry about the overflow here, we
1184 * have our own CBS to keep us inline; only account when RT
1185 * bandwidth is relevant.
1187 if (sched_rt_bandwidth_account(rt_rq
))
1188 rt_rq
->rt_time
+= delta_exec
;
1189 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
1193 static enum hrtimer_restart
inactive_task_timer(struct hrtimer
*timer
)
1195 struct sched_dl_entity
*dl_se
= container_of(timer
,
1196 struct sched_dl_entity
,
1198 struct task_struct
*p
= dl_task_of(dl_se
);
1202 rq
= task_rq_lock(p
, &rf
);
1204 if (!dl_task(p
) || p
->state
== TASK_DEAD
) {
1205 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
1207 if (p
->state
== TASK_DEAD
&& dl_se
->dl_non_contending
) {
1208 sub_running_bw(p
->dl
.dl_bw
, dl_rq_of_se(&p
->dl
));
1209 sub_rq_bw(p
->dl
.dl_bw
, dl_rq_of_se(&p
->dl
));
1210 dl_se
->dl_non_contending
= 0;
1213 raw_spin_lock(&dl_b
->lock
);
1214 __dl_clear(dl_b
, p
->dl
.dl_bw
, dl_bw_cpus(task_cpu(p
)));
1215 raw_spin_unlock(&dl_b
->lock
);
1216 __dl_clear_params(p
);
1220 if (dl_se
->dl_non_contending
== 0)
1224 update_rq_clock(rq
);
1226 sub_running_bw(dl_se
->dl_bw
, &rq
->dl
);
1227 dl_se
->dl_non_contending
= 0;
1229 task_rq_unlock(rq
, p
, &rf
);
1232 return HRTIMER_NORESTART
;
1235 void init_dl_inactive_task_timer(struct sched_dl_entity
*dl_se
)
1237 struct hrtimer
*timer
= &dl_se
->inactive_timer
;
1239 hrtimer_init(timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1240 timer
->function
= inactive_task_timer
;
1245 static void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
1247 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
1249 if (dl_rq
->earliest_dl
.curr
== 0 ||
1250 dl_time_before(deadline
, dl_rq
->earliest_dl
.curr
)) {
1251 dl_rq
->earliest_dl
.curr
= deadline
;
1252 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, deadline
);
1256 static void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
1258 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
1261 * Since we may have removed our earliest (and/or next earliest)
1262 * task we must recompute them.
1264 if (!dl_rq
->dl_nr_running
) {
1265 dl_rq
->earliest_dl
.curr
= 0;
1266 dl_rq
->earliest_dl
.next
= 0;
1267 cpudl_clear(&rq
->rd
->cpudl
, rq
->cpu
);
1269 struct rb_node
*leftmost
= dl_rq
->rb_leftmost
;
1270 struct sched_dl_entity
*entry
;
1272 entry
= rb_entry(leftmost
, struct sched_dl_entity
, rb_node
);
1273 dl_rq
->earliest_dl
.curr
= entry
->deadline
;
1274 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, entry
->deadline
);
1280 static inline void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
1281 static inline void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
1283 #endif /* CONFIG_SMP */
1286 void inc_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
1288 int prio
= dl_task_of(dl_se
)->prio
;
1289 u64 deadline
= dl_se
->deadline
;
1291 WARN_ON(!dl_prio(prio
));
1292 dl_rq
->dl_nr_running
++;
1293 add_nr_running(rq_of_dl_rq(dl_rq
), 1);
1295 inc_dl_deadline(dl_rq
, deadline
);
1296 inc_dl_migration(dl_se
, dl_rq
);
1300 void dec_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
1302 int prio
= dl_task_of(dl_se
)->prio
;
1304 WARN_ON(!dl_prio(prio
));
1305 WARN_ON(!dl_rq
->dl_nr_running
);
1306 dl_rq
->dl_nr_running
--;
1307 sub_nr_running(rq_of_dl_rq(dl_rq
), 1);
1309 dec_dl_deadline(dl_rq
, dl_se
->deadline
);
1310 dec_dl_migration(dl_se
, dl_rq
);
1313 static void __enqueue_dl_entity(struct sched_dl_entity
*dl_se
)
1315 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
1316 struct rb_node
**link
= &dl_rq
->rb_root
.rb_node
;
1317 struct rb_node
*parent
= NULL
;
1318 struct sched_dl_entity
*entry
;
1321 BUG_ON(!RB_EMPTY_NODE(&dl_se
->rb_node
));
1325 entry
= rb_entry(parent
, struct sched_dl_entity
, rb_node
);
1326 if (dl_time_before(dl_se
->deadline
, entry
->deadline
))
1327 link
= &parent
->rb_left
;
1329 link
= &parent
->rb_right
;
1335 dl_rq
->rb_leftmost
= &dl_se
->rb_node
;
1337 rb_link_node(&dl_se
->rb_node
, parent
, link
);
1338 rb_insert_color(&dl_se
->rb_node
, &dl_rq
->rb_root
);
1340 inc_dl_tasks(dl_se
, dl_rq
);
1343 static void __dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
1345 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
1347 if (RB_EMPTY_NODE(&dl_se
->rb_node
))
1350 if (dl_rq
->rb_leftmost
== &dl_se
->rb_node
) {
1351 struct rb_node
*next_node
;
1353 next_node
= rb_next(&dl_se
->rb_node
);
1354 dl_rq
->rb_leftmost
= next_node
;
1357 rb_erase(&dl_se
->rb_node
, &dl_rq
->rb_root
);
1358 RB_CLEAR_NODE(&dl_se
->rb_node
);
1360 dec_dl_tasks(dl_se
, dl_rq
);
1364 enqueue_dl_entity(struct sched_dl_entity
*dl_se
,
1365 struct sched_dl_entity
*pi_se
, int flags
)
1367 BUG_ON(on_dl_rq(dl_se
));
1370 * If this is a wakeup or a new instance, the scheduling
1371 * parameters of the task might need updating. Otherwise,
1372 * we want a replenishment of its runtime.
1374 if (flags
& ENQUEUE_WAKEUP
) {
1375 task_contending(dl_se
, flags
);
1376 update_dl_entity(dl_se
, pi_se
);
1377 } else if (flags
& ENQUEUE_REPLENISH
) {
1378 replenish_dl_entity(dl_se
, pi_se
);
1381 __enqueue_dl_entity(dl_se
);
1384 static void dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
1386 __dequeue_dl_entity(dl_se
);
1389 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
1391 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
1392 struct sched_dl_entity
*pi_se
= &p
->dl
;
1395 * Use the scheduling parameters of the top pi-waiter task if:
1396 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1397 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1398 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1399 * boosted due to a SCHED_DEADLINE pi-waiter).
1400 * Otherwise we keep our runtime and deadline.
1402 if (pi_task
&& dl_prio(pi_task
->normal_prio
) && p
->dl
.dl_boosted
) {
1403 pi_se
= &pi_task
->dl
;
1404 } else if (!dl_prio(p
->normal_prio
)) {
1406 * Special case in which we have a !SCHED_DEADLINE task
1407 * that is going to be deboosted, but exceeds its
1408 * runtime while doing so. No point in replenishing
1409 * it, as it's going to return back to its original
1410 * scheduling class after this.
1412 BUG_ON(!p
->dl
.dl_boosted
|| flags
!= ENQUEUE_REPLENISH
);
1417 * Check if a constrained deadline task was activated
1418 * after the deadline but before the next period.
1419 * If that is the case, the task will be throttled and
1420 * the replenishment timer will be set to the next period.
1422 if (!p
->dl
.dl_throttled
&& !dl_is_implicit(&p
->dl
))
1423 dl_check_constrained_dl(&p
->dl
);
1425 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
|| flags
& ENQUEUE_RESTORE
) {
1426 add_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
1427 add_running_bw(p
->dl
.dl_bw
, &rq
->dl
);
1431 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1432 * its budget it needs a replenishment and, since it now is on
1433 * its rq, the bandwidth timer callback (which clearly has not
1434 * run yet) will take care of this.
1435 * However, the active utilization does not depend on the fact
1436 * that the task is on the runqueue or not (but depends on the
1437 * task's state - in GRUB parlance, "inactive" vs "active contending").
1438 * In other words, even if a task is throttled its utilization must
1439 * be counted in the active utilization; hence, we need to call
1442 if (p
->dl
.dl_throttled
&& !(flags
& ENQUEUE_REPLENISH
)) {
1443 if (flags
& ENQUEUE_WAKEUP
)
1444 task_contending(&p
->dl
, flags
);
1449 enqueue_dl_entity(&p
->dl
, pi_se
, flags
);
1451 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
1452 enqueue_pushable_dl_task(rq
, p
);
1455 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
1457 dequeue_dl_entity(&p
->dl
);
1458 dequeue_pushable_dl_task(rq
, p
);
1461 static void dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
1464 __dequeue_task_dl(rq
, p
, flags
);
1466 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
|| flags
& DEQUEUE_SAVE
) {
1467 sub_running_bw(p
->dl
.dl_bw
, &rq
->dl
);
1468 sub_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
1472 * This check allows to start the inactive timer (or to immediately
1473 * decrease the active utilization, if needed) in two cases:
1474 * when the task blocks and when it is terminating
1475 * (p->state == TASK_DEAD). We can handle the two cases in the same
1476 * way, because from GRUB's point of view the same thing is happening
1477 * (the task moves from "active contending" to "active non contending"
1480 if (flags
& DEQUEUE_SLEEP
)
1481 task_non_contending(p
);
1485 * Yield task semantic for -deadline tasks is:
1487 * get off from the CPU until our next instance, with
1488 * a new runtime. This is of little use now, since we
1489 * don't have a bandwidth reclaiming mechanism. Anyway,
1490 * bandwidth reclaiming is planned for the future, and
1491 * yield_task_dl will indicate that some spare budget
1492 * is available for other task instances to use it.
1494 static void yield_task_dl(struct rq
*rq
)
1497 * We make the task go to sleep until its current deadline by
1498 * forcing its runtime to zero. This way, update_curr_dl() stops
1499 * it and the bandwidth timer will wake it up and will give it
1500 * new scheduling parameters (thanks to dl_yielded=1).
1502 rq
->curr
->dl
.dl_yielded
= 1;
1504 update_rq_clock(rq
);
1507 * Tell update_rq_clock() that we've just updated,
1508 * so we don't do microscopic update in schedule()
1509 * and double the fastpath cost.
1511 rq_clock_skip_update(rq
, true);
1516 static int find_later_rq(struct task_struct
*task
);
1519 select_task_rq_dl(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
)
1521 struct task_struct
*curr
;
1524 if (sd_flag
!= SD_BALANCE_WAKE
)
1530 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1533 * If we are dealing with a -deadline task, we must
1534 * decide where to wake it up.
1535 * If it has a later deadline and the current task
1536 * on this rq can't move (provided the waking task
1537 * can!) we prefer to send it somewhere else. On the
1538 * other hand, if it has a shorter deadline, we
1539 * try to make it stay here, it might be important.
1541 if (unlikely(dl_task(curr
)) &&
1542 (curr
->nr_cpus_allowed
< 2 ||
1543 !dl_entity_preempt(&p
->dl
, &curr
->dl
)) &&
1544 (p
->nr_cpus_allowed
> 1)) {
1545 int target
= find_later_rq(p
);
1548 (dl_time_before(p
->dl
.deadline
,
1549 cpu_rq(target
)->dl
.earliest_dl
.curr
) ||
1550 (cpu_rq(target
)->dl
.dl_nr_running
== 0)))
1559 static void migrate_task_rq_dl(struct task_struct
*p
)
1563 if (p
->state
!= TASK_WAKING
)
1568 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1569 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1570 * rq->lock is not... So, lock it
1572 raw_spin_lock(&rq
->lock
);
1573 if (p
->dl
.dl_non_contending
) {
1574 sub_running_bw(p
->dl
.dl_bw
, &rq
->dl
);
1575 p
->dl
.dl_non_contending
= 0;
1577 * If the timer handler is currently running and the
1578 * timer cannot be cancelled, inactive_task_timer()
1579 * will see that dl_not_contending is not set, and
1580 * will not touch the rq's active utilization,
1581 * so we are still safe.
1583 if (hrtimer_try_to_cancel(&p
->dl
.inactive_timer
) == 1)
1586 sub_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
1587 raw_spin_unlock(&rq
->lock
);
1590 static void check_preempt_equal_dl(struct rq
*rq
, struct task_struct
*p
)
1593 * Current can't be migrated, useless to reschedule,
1594 * let's hope p can move out.
1596 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1597 cpudl_find(&rq
->rd
->cpudl
, rq
->curr
, NULL
) == -1)
1601 * p is migratable, so let's not schedule it and
1602 * see if it is pushed or pulled somewhere else.
1604 if (p
->nr_cpus_allowed
!= 1 &&
1605 cpudl_find(&rq
->rd
->cpudl
, p
, NULL
) != -1)
1611 #endif /* CONFIG_SMP */
1614 * Only called when both the current and waking task are -deadline
1617 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
,
1620 if (dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
)) {
1627 * In the unlikely case current and p have the same deadline
1628 * let us try to decide what's the best thing to do...
1630 if ((p
->dl
.deadline
== rq
->curr
->dl
.deadline
) &&
1631 !test_tsk_need_resched(rq
->curr
))
1632 check_preempt_equal_dl(rq
, p
);
1633 #endif /* CONFIG_SMP */
1636 #ifdef CONFIG_SCHED_HRTICK
1637 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1639 hrtick_start(rq
, p
->dl
.runtime
);
1641 #else /* !CONFIG_SCHED_HRTICK */
1642 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1647 static struct sched_dl_entity
*pick_next_dl_entity(struct rq
*rq
,
1648 struct dl_rq
*dl_rq
)
1650 struct rb_node
*left
= dl_rq
->rb_leftmost
;
1655 return rb_entry(left
, struct sched_dl_entity
, rb_node
);
1658 struct task_struct
*
1659 pick_next_task_dl(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
1661 struct sched_dl_entity
*dl_se
;
1662 struct task_struct
*p
;
1663 struct dl_rq
*dl_rq
;
1667 if (need_pull_dl_task(rq
, prev
)) {
1669 * This is OK, because current is on_cpu, which avoids it being
1670 * picked for load-balance and preemption/IRQs are still
1671 * disabled avoiding further scheduler activity on it and we're
1672 * being very careful to re-start the picking loop.
1674 rq_unpin_lock(rq
, rf
);
1676 rq_repin_lock(rq
, rf
);
1678 * pull_dl_task() can drop (and re-acquire) rq->lock; this
1679 * means a stop task can slip in, in which case we need to
1680 * re-start task selection.
1682 if (rq
->stop
&& task_on_rq_queued(rq
->stop
))
1687 * When prev is DL, we may throttle it in put_prev_task().
1688 * So, we update time before we check for dl_nr_running.
1690 if (prev
->sched_class
== &dl_sched_class
)
1693 if (unlikely(!dl_rq
->dl_nr_running
))
1696 put_prev_task(rq
, prev
);
1698 dl_se
= pick_next_dl_entity(rq
, dl_rq
);
1701 p
= dl_task_of(dl_se
);
1702 p
->se
.exec_start
= rq_clock_task(rq
);
1704 /* Running task will never be pushed. */
1705 dequeue_pushable_dl_task(rq
, p
);
1707 if (hrtick_enabled(rq
))
1708 start_hrtick_dl(rq
, p
);
1710 queue_push_tasks(rq
);
1715 static void put_prev_task_dl(struct rq
*rq
, struct task_struct
*p
)
1719 if (on_dl_rq(&p
->dl
) && p
->nr_cpus_allowed
> 1)
1720 enqueue_pushable_dl_task(rq
, p
);
1723 static void task_tick_dl(struct rq
*rq
, struct task_struct
*p
, int queued
)
1728 * Even when we have runtime, update_curr_dl() might have resulted in us
1729 * not being the leftmost task anymore. In that case NEED_RESCHED will
1730 * be set and schedule() will start a new hrtick for the next task.
1732 if (hrtick_enabled(rq
) && queued
&& p
->dl
.runtime
> 0 &&
1733 is_leftmost(p
, &rq
->dl
))
1734 start_hrtick_dl(rq
, p
);
1737 static void task_fork_dl(struct task_struct
*p
)
1740 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1745 static void set_curr_task_dl(struct rq
*rq
)
1747 struct task_struct
*p
= rq
->curr
;
1749 p
->se
.exec_start
= rq_clock_task(rq
);
1751 /* You can't push away the running task */
1752 dequeue_pushable_dl_task(rq
, p
);
1757 /* Only try algorithms three times */
1758 #define DL_MAX_TRIES 3
1760 static int pick_dl_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1762 if (!task_running(rq
, p
) &&
1763 cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1769 * Return the earliest pushable rq's task, which is suitable to be executed
1770 * on the CPU, NULL otherwise:
1772 static struct task_struct
*pick_earliest_pushable_dl_task(struct rq
*rq
, int cpu
)
1774 struct rb_node
*next_node
= rq
->dl
.pushable_dl_tasks_leftmost
;
1775 struct task_struct
*p
= NULL
;
1777 if (!has_pushable_dl_tasks(rq
))
1782 p
= rb_entry(next_node
, struct task_struct
, pushable_dl_tasks
);
1784 if (pick_dl_task(rq
, p
, cpu
))
1787 next_node
= rb_next(next_node
);
1794 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask_dl
);
1796 static int find_later_rq(struct task_struct
*task
)
1798 struct sched_domain
*sd
;
1799 struct cpumask
*later_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask_dl
);
1800 int this_cpu
= smp_processor_id();
1801 int best_cpu
, cpu
= task_cpu(task
);
1803 /* Make sure the mask is initialized first */
1804 if (unlikely(!later_mask
))
1807 if (task
->nr_cpus_allowed
== 1)
1811 * We have to consider system topology and task affinity
1812 * first, then we can look for a suitable cpu.
1814 best_cpu
= cpudl_find(&task_rq(task
)->rd
->cpudl
,
1820 * If we are here, some target has been found,
1821 * the most suitable of which is cached in best_cpu.
1822 * This is, among the runqueues where the current tasks
1823 * have later deadlines than the task's one, the rq
1824 * with the latest possible one.
1826 * Now we check how well this matches with task's
1827 * affinity and system topology.
1829 * The last cpu where the task run is our first
1830 * guess, since it is most likely cache-hot there.
1832 if (cpumask_test_cpu(cpu
, later_mask
))
1835 * Check if this_cpu is to be skipped (i.e., it is
1836 * not in the mask) or not.
1838 if (!cpumask_test_cpu(this_cpu
, later_mask
))
1842 for_each_domain(cpu
, sd
) {
1843 if (sd
->flags
& SD_WAKE_AFFINE
) {
1846 * If possible, preempting this_cpu is
1847 * cheaper than migrating.
1849 if (this_cpu
!= -1 &&
1850 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1856 * Last chance: if best_cpu is valid and is
1857 * in the mask, that becomes our choice.
1859 if (best_cpu
< nr_cpu_ids
&&
1860 cpumask_test_cpu(best_cpu
, sched_domain_span(sd
))) {
1869 * At this point, all our guesses failed, we just return
1870 * 'something', and let the caller sort the things out.
1875 cpu
= cpumask_any(later_mask
);
1876 if (cpu
< nr_cpu_ids
)
1882 /* Locks the rq it finds */
1883 static struct rq
*find_lock_later_rq(struct task_struct
*task
, struct rq
*rq
)
1885 struct rq
*later_rq
= NULL
;
1889 for (tries
= 0; tries
< DL_MAX_TRIES
; tries
++) {
1890 cpu
= find_later_rq(task
);
1892 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1895 later_rq
= cpu_rq(cpu
);
1897 if (later_rq
->dl
.dl_nr_running
&&
1898 !dl_time_before(task
->dl
.deadline
,
1899 later_rq
->dl
.earliest_dl
.curr
)) {
1901 * Target rq has tasks of equal or earlier deadline,
1902 * retrying does not release any lock and is unlikely
1903 * to yield a different result.
1909 /* Retry if something changed. */
1910 if (double_lock_balance(rq
, later_rq
)) {
1911 if (unlikely(task_rq(task
) != rq
||
1912 !cpumask_test_cpu(later_rq
->cpu
, &task
->cpus_allowed
) ||
1913 task_running(rq
, task
) ||
1915 !task_on_rq_queued(task
))) {
1916 double_unlock_balance(rq
, later_rq
);
1923 * If the rq we found has no -deadline task, or
1924 * its earliest one has a later deadline than our
1925 * task, the rq is a good one.
1927 if (!later_rq
->dl
.dl_nr_running
||
1928 dl_time_before(task
->dl
.deadline
,
1929 later_rq
->dl
.earliest_dl
.curr
))
1932 /* Otherwise we try again. */
1933 double_unlock_balance(rq
, later_rq
);
1940 static struct task_struct
*pick_next_pushable_dl_task(struct rq
*rq
)
1942 struct task_struct
*p
;
1944 if (!has_pushable_dl_tasks(rq
))
1947 p
= rb_entry(rq
->dl
.pushable_dl_tasks_leftmost
,
1948 struct task_struct
, pushable_dl_tasks
);
1950 BUG_ON(rq
->cpu
!= task_cpu(p
));
1951 BUG_ON(task_current(rq
, p
));
1952 BUG_ON(p
->nr_cpus_allowed
<= 1);
1954 BUG_ON(!task_on_rq_queued(p
));
1955 BUG_ON(!dl_task(p
));
1961 * See if the non running -deadline tasks on this rq
1962 * can be sent to some other CPU where they can preempt
1963 * and start executing.
1965 static int push_dl_task(struct rq
*rq
)
1967 struct task_struct
*next_task
;
1968 struct rq
*later_rq
;
1971 if (!rq
->dl
.overloaded
)
1974 next_task
= pick_next_pushable_dl_task(rq
);
1979 if (unlikely(next_task
== rq
->curr
)) {
1985 * If next_task preempts rq->curr, and rq->curr
1986 * can move away, it makes sense to just reschedule
1987 * without going further in pushing next_task.
1989 if (dl_task(rq
->curr
) &&
1990 dl_time_before(next_task
->dl
.deadline
, rq
->curr
->dl
.deadline
) &&
1991 rq
->curr
->nr_cpus_allowed
> 1) {
1996 /* We might release rq lock */
1997 get_task_struct(next_task
);
1999 /* Will lock the rq it'll find */
2000 later_rq
= find_lock_later_rq(next_task
, rq
);
2002 struct task_struct
*task
;
2005 * We must check all this again, since
2006 * find_lock_later_rq releases rq->lock and it is
2007 * then possible that next_task has migrated.
2009 task
= pick_next_pushable_dl_task(rq
);
2010 if (task
== next_task
) {
2012 * The task is still there. We don't try
2013 * again, some other cpu will pull it when ready.
2022 put_task_struct(next_task
);
2027 deactivate_task(rq
, next_task
, 0);
2028 sub_running_bw(next_task
->dl
.dl_bw
, &rq
->dl
);
2029 sub_rq_bw(next_task
->dl
.dl_bw
, &rq
->dl
);
2030 set_task_cpu(next_task
, later_rq
->cpu
);
2031 add_rq_bw(next_task
->dl
.dl_bw
, &later_rq
->dl
);
2032 add_running_bw(next_task
->dl
.dl_bw
, &later_rq
->dl
);
2033 activate_task(later_rq
, next_task
, 0);
2036 resched_curr(later_rq
);
2038 double_unlock_balance(rq
, later_rq
);
2041 put_task_struct(next_task
);
2046 static void push_dl_tasks(struct rq
*rq
)
2048 /* push_dl_task() will return true if it moved a -deadline task */
2049 while (push_dl_task(rq
))
2053 static void pull_dl_task(struct rq
*this_rq
)
2055 int this_cpu
= this_rq
->cpu
, cpu
;
2056 struct task_struct
*p
;
2057 bool resched
= false;
2059 u64 dmin
= LONG_MAX
;
2061 if (likely(!dl_overloaded(this_rq
)))
2065 * Match the barrier from dl_set_overloaded; this guarantees that if we
2066 * see overloaded we must also see the dlo_mask bit.
2070 for_each_cpu(cpu
, this_rq
->rd
->dlo_mask
) {
2071 if (this_cpu
== cpu
)
2074 src_rq
= cpu_rq(cpu
);
2077 * It looks racy, abd it is! However, as in sched_rt.c,
2078 * we are fine with this.
2080 if (this_rq
->dl
.dl_nr_running
&&
2081 dl_time_before(this_rq
->dl
.earliest_dl
.curr
,
2082 src_rq
->dl
.earliest_dl
.next
))
2085 /* Might drop this_rq->lock */
2086 double_lock_balance(this_rq
, src_rq
);
2089 * If there are no more pullable tasks on the
2090 * rq, we're done with it.
2092 if (src_rq
->dl
.dl_nr_running
<= 1)
2095 p
= pick_earliest_pushable_dl_task(src_rq
, this_cpu
);
2098 * We found a task to be pulled if:
2099 * - it preempts our current (if there's one),
2100 * - it will preempt the last one we pulled (if any).
2102 if (p
&& dl_time_before(p
->dl
.deadline
, dmin
) &&
2103 (!this_rq
->dl
.dl_nr_running
||
2104 dl_time_before(p
->dl
.deadline
,
2105 this_rq
->dl
.earliest_dl
.curr
))) {
2106 WARN_ON(p
== src_rq
->curr
);
2107 WARN_ON(!task_on_rq_queued(p
));
2110 * Then we pull iff p has actually an earlier
2111 * deadline than the current task of its runqueue.
2113 if (dl_time_before(p
->dl
.deadline
,
2114 src_rq
->curr
->dl
.deadline
))
2119 deactivate_task(src_rq
, p
, 0);
2120 sub_running_bw(p
->dl
.dl_bw
, &src_rq
->dl
);
2121 sub_rq_bw(p
->dl
.dl_bw
, &src_rq
->dl
);
2122 set_task_cpu(p
, this_cpu
);
2123 add_rq_bw(p
->dl
.dl_bw
, &this_rq
->dl
);
2124 add_running_bw(p
->dl
.dl_bw
, &this_rq
->dl
);
2125 activate_task(this_rq
, p
, 0);
2126 dmin
= p
->dl
.deadline
;
2128 /* Is there any other task even earlier? */
2131 double_unlock_balance(this_rq
, src_rq
);
2135 resched_curr(this_rq
);
2139 * Since the task is not running and a reschedule is not going to happen
2140 * anytime soon on its runqueue, we try pushing it away now.
2142 static void task_woken_dl(struct rq
*rq
, struct task_struct
*p
)
2144 if (!task_running(rq
, p
) &&
2145 !test_tsk_need_resched(rq
->curr
) &&
2146 p
->nr_cpus_allowed
> 1 &&
2147 dl_task(rq
->curr
) &&
2148 (rq
->curr
->nr_cpus_allowed
< 2 ||
2149 !dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
))) {
2154 static void set_cpus_allowed_dl(struct task_struct
*p
,
2155 const struct cpumask
*new_mask
)
2157 struct root_domain
*src_rd
;
2160 BUG_ON(!dl_task(p
));
2165 * Migrating a SCHED_DEADLINE task between exclusive
2166 * cpusets (different root_domains) entails a bandwidth
2167 * update. We already made space for us in the destination
2168 * domain (see cpuset_can_attach()).
2170 if (!cpumask_intersects(src_rd
->span
, new_mask
)) {
2171 struct dl_bw
*src_dl_b
;
2173 src_dl_b
= dl_bw_of(cpu_of(rq
));
2175 * We now free resources of the root_domain we are migrating
2176 * off. In the worst case, sched_setattr() may temporary fail
2177 * until we complete the update.
2179 raw_spin_lock(&src_dl_b
->lock
);
2180 __dl_clear(src_dl_b
, p
->dl
.dl_bw
, dl_bw_cpus(task_cpu(p
)));
2181 raw_spin_unlock(&src_dl_b
->lock
);
2184 set_cpus_allowed_common(p
, new_mask
);
2187 /* Assumes rq->lock is held */
2188 static void rq_online_dl(struct rq
*rq
)
2190 if (rq
->dl
.overloaded
)
2191 dl_set_overload(rq
);
2193 cpudl_set_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
2194 if (rq
->dl
.dl_nr_running
> 0)
2195 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, rq
->dl
.earliest_dl
.curr
);
2198 /* Assumes rq->lock is held */
2199 static void rq_offline_dl(struct rq
*rq
)
2201 if (rq
->dl
.overloaded
)
2202 dl_clear_overload(rq
);
2204 cpudl_clear(&rq
->rd
->cpudl
, rq
->cpu
);
2205 cpudl_clear_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
2208 void __init
init_sched_dl_class(void)
2212 for_each_possible_cpu(i
)
2213 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl
, i
),
2214 GFP_KERNEL
, cpu_to_node(i
));
2217 #endif /* CONFIG_SMP */
2219 static void switched_from_dl(struct rq
*rq
, struct task_struct
*p
)
2222 * task_non_contending() can start the "inactive timer" (if the 0-lag
2223 * time is in the future). If the task switches back to dl before
2224 * the "inactive timer" fires, it can continue to consume its current
2225 * runtime using its current deadline. If it stays outside of
2226 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2227 * will reset the task parameters.
2229 if (task_on_rq_queued(p
) && p
->dl
.dl_runtime
)
2230 task_non_contending(p
);
2232 if (!task_on_rq_queued(p
))
2233 sub_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
2236 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2237 * at the 0-lag time, because the task could have been migrated
2238 * while SCHED_OTHER in the meanwhile.
2240 if (p
->dl
.dl_non_contending
)
2241 p
->dl
.dl_non_contending
= 0;
2244 * Since this might be the only -deadline task on the rq,
2245 * this is the right place to try to pull some other one
2246 * from an overloaded cpu, if any.
2248 if (!task_on_rq_queued(p
) || rq
->dl
.dl_nr_running
)
2251 queue_pull_task(rq
);
2255 * When switching to -deadline, we may overload the rq, then
2256 * we try to push someone off, if possible.
2258 static void switched_to_dl(struct rq
*rq
, struct task_struct
*p
)
2260 if (hrtimer_try_to_cancel(&p
->dl
.inactive_timer
) == 1)
2263 /* If p is not queued we will update its parameters at next wakeup. */
2264 if (!task_on_rq_queued(p
)) {
2265 add_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
2270 * If p is boosted we already updated its params in
2271 * rt_mutex_setprio()->enqueue_task(..., ENQUEUE_REPLENISH),
2272 * p's deadline being now already after rq_clock(rq).
2274 if (dl_time_before(p
->dl
.deadline
, rq_clock(rq
)))
2275 setup_new_dl_entity(&p
->dl
);
2277 if (rq
->curr
!= p
) {
2279 if (p
->nr_cpus_allowed
> 1 && rq
->dl
.overloaded
)
2280 queue_push_tasks(rq
);
2282 if (dl_task(rq
->curr
))
2283 check_preempt_curr_dl(rq
, p
, 0);
2290 * If the scheduling parameters of a -deadline task changed,
2291 * a push or pull operation might be needed.
2293 static void prio_changed_dl(struct rq
*rq
, struct task_struct
*p
,
2296 if (task_on_rq_queued(p
) || rq
->curr
== p
) {
2299 * This might be too much, but unfortunately
2300 * we don't have the old deadline value, and
2301 * we can't argue if the task is increasing
2302 * or lowering its prio, so...
2304 if (!rq
->dl
.overloaded
)
2305 queue_pull_task(rq
);
2308 * If we now have a earlier deadline task than p,
2309 * then reschedule, provided p is still on this
2312 if (dl_time_before(rq
->dl
.earliest_dl
.curr
, p
->dl
.deadline
))
2316 * Again, we don't know if p has a earlier
2317 * or later deadline, so let's blindly set a
2318 * (maybe not needed) rescheduling point.
2321 #endif /* CONFIG_SMP */
2325 const struct sched_class dl_sched_class
= {
2326 .next
= &rt_sched_class
,
2327 .enqueue_task
= enqueue_task_dl
,
2328 .dequeue_task
= dequeue_task_dl
,
2329 .yield_task
= yield_task_dl
,
2331 .check_preempt_curr
= check_preempt_curr_dl
,
2333 .pick_next_task
= pick_next_task_dl
,
2334 .put_prev_task
= put_prev_task_dl
,
2337 .select_task_rq
= select_task_rq_dl
,
2338 .migrate_task_rq
= migrate_task_rq_dl
,
2339 .set_cpus_allowed
= set_cpus_allowed_dl
,
2340 .rq_online
= rq_online_dl
,
2341 .rq_offline
= rq_offline_dl
,
2342 .task_woken
= task_woken_dl
,
2345 .set_curr_task
= set_curr_task_dl
,
2346 .task_tick
= task_tick_dl
,
2347 .task_fork
= task_fork_dl
,
2349 .prio_changed
= prio_changed_dl
,
2350 .switched_from
= switched_from_dl
,
2351 .switched_to
= switched_to_dl
,
2353 .update_curr
= update_curr_dl
,
2356 int sched_dl_global_validate(void)
2358 u64 runtime
= global_rt_runtime();
2359 u64 period
= global_rt_period();
2360 u64 new_bw
= to_ratio(period
, runtime
);
2363 unsigned long flags
;
2366 * Here we want to check the bandwidth not being set to some
2367 * value smaller than the currently allocated bandwidth in
2368 * any of the root_domains.
2370 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2371 * cycling on root_domains... Discussion on different/better
2372 * solutions is welcome!
2374 for_each_possible_cpu(cpu
) {
2375 rcu_read_lock_sched();
2376 dl_b
= dl_bw_of(cpu
);
2378 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2379 if (new_bw
< dl_b
->total_bw
)
2381 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2383 rcu_read_unlock_sched();
2392 void init_dl_rq_bw_ratio(struct dl_rq
*dl_rq
)
2394 if (global_rt_runtime() == RUNTIME_INF
) {
2395 dl_rq
->bw_ratio
= 1 << RATIO_SHIFT
;
2396 dl_rq
->extra_bw
= 1 << BW_SHIFT
;
2398 dl_rq
->bw_ratio
= to_ratio(global_rt_runtime(),
2399 global_rt_period()) >> (BW_SHIFT
- RATIO_SHIFT
);
2400 dl_rq
->extra_bw
= to_ratio(global_rt_period(),
2401 global_rt_runtime());
2405 void sched_dl_do_global(void)
2410 unsigned long flags
;
2412 def_dl_bandwidth
.dl_period
= global_rt_period();
2413 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
2415 if (global_rt_runtime() != RUNTIME_INF
)
2416 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
2419 * FIXME: As above...
2421 for_each_possible_cpu(cpu
) {
2422 rcu_read_lock_sched();
2423 dl_b
= dl_bw_of(cpu
);
2425 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2427 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2429 rcu_read_unlock_sched();
2430 init_dl_rq_bw_ratio(&cpu_rq(cpu
)->dl
);
2435 * We must be sure that accepting a new task (or allowing changing the
2436 * parameters of an existing one) is consistent with the bandwidth
2437 * constraints. If yes, this function also accordingly updates the currently
2438 * allocated bandwidth to reflect the new situation.
2440 * This function is called while holding p's rq->lock.
2442 int sched_dl_overflow(struct task_struct
*p
, int policy
,
2443 const struct sched_attr
*attr
)
2445 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2446 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2447 u64 runtime
= attr
->sched_runtime
;
2448 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2451 /* !deadline task may carry old deadline bandwidth */
2452 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2456 * Either if a task, enters, leave, or stays -deadline but changes
2457 * its parameters, we may need to update accordingly the total
2458 * allocated bandwidth of the container.
2460 raw_spin_lock(&dl_b
->lock
);
2461 cpus
= dl_bw_cpus(task_cpu(p
));
2462 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2463 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2464 if (hrtimer_active(&p
->dl
.inactive_timer
))
2465 __dl_clear(dl_b
, p
->dl
.dl_bw
, cpus
);
2466 __dl_add(dl_b
, new_bw
, cpus
);
2468 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2469 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2471 * XXX this is slightly incorrect: when the task
2472 * utilization decreases, we should delay the total
2473 * utilization change until the task's 0-lag point.
2474 * But this would require to set the task's "inactive
2475 * timer" when the task is not inactive.
2477 __dl_clear(dl_b
, p
->dl
.dl_bw
, cpus
);
2478 __dl_add(dl_b
, new_bw
, cpus
);
2479 dl_change_utilization(p
, new_bw
);
2481 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2483 * Do not decrease the total deadline utilization here,
2484 * switched_from_dl() will take care to do it at the correct
2489 raw_spin_unlock(&dl_b
->lock
);
2495 * This function initializes the sched_dl_entity of a newly becoming
2496 * SCHED_DEADLINE task.
2498 * Only the static values are considered here, the actual runtime and the
2499 * absolute deadline will be properly calculated when the task is enqueued
2500 * for the first time with its new policy.
2502 void __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
2504 struct sched_dl_entity
*dl_se
= &p
->dl
;
2506 dl_se
->dl_runtime
= attr
->sched_runtime
;
2507 dl_se
->dl_deadline
= attr
->sched_deadline
;
2508 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
2509 dl_se
->flags
= attr
->sched_flags
;
2510 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
2511 dl_se
->dl_density
= to_ratio(dl_se
->dl_deadline
, dl_se
->dl_runtime
);
2514 void __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
2516 struct sched_dl_entity
*dl_se
= &p
->dl
;
2518 attr
->sched_priority
= p
->rt_priority
;
2519 attr
->sched_runtime
= dl_se
->dl_runtime
;
2520 attr
->sched_deadline
= dl_se
->dl_deadline
;
2521 attr
->sched_period
= dl_se
->dl_period
;
2522 attr
->sched_flags
= dl_se
->flags
;
2526 * This function validates the new parameters of a -deadline task.
2527 * We ask for the deadline not being zero, and greater or equal
2528 * than the runtime, as well as the period of being zero or
2529 * greater than deadline. Furthermore, we have to be sure that
2530 * user parameters are above the internal resolution of 1us (we
2531 * check sched_runtime only since it is always the smaller one) and
2532 * below 2^63 ns (we have to check both sched_deadline and
2533 * sched_period, as the latter can be zero).
2535 bool __checkparam_dl(const struct sched_attr
*attr
)
2538 if (attr
->sched_deadline
== 0)
2542 * Since we truncate DL_SCALE bits, make sure we're at least
2545 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
2549 * Since we use the MSB for wrap-around and sign issues, make
2550 * sure it's not set (mind that period can be equal to zero).
2552 if (attr
->sched_deadline
& (1ULL << 63) ||
2553 attr
->sched_period
& (1ULL << 63))
2556 /* runtime <= deadline <= period (if period != 0) */
2557 if ((attr
->sched_period
!= 0 &&
2558 attr
->sched_period
< attr
->sched_deadline
) ||
2559 attr
->sched_deadline
< attr
->sched_runtime
)
2566 * This function clears the sched_dl_entity static params.
2568 void __dl_clear_params(struct task_struct
*p
)
2570 struct sched_dl_entity
*dl_se
= &p
->dl
;
2572 dl_se
->dl_runtime
= 0;
2573 dl_se
->dl_deadline
= 0;
2574 dl_se
->dl_period
= 0;
2577 dl_se
->dl_density
= 0;
2579 dl_se
->dl_throttled
= 0;
2580 dl_se
->dl_yielded
= 0;
2581 dl_se
->dl_non_contending
= 0;
2584 bool dl_param_changed(struct task_struct
*p
, const struct sched_attr
*attr
)
2586 struct sched_dl_entity
*dl_se
= &p
->dl
;
2588 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
2589 dl_se
->dl_deadline
!= attr
->sched_deadline
||
2590 dl_se
->dl_period
!= attr
->sched_period
||
2591 dl_se
->flags
!= attr
->sched_flags
)
2598 int dl_task_can_attach(struct task_struct
*p
, const struct cpumask
*cs_cpus_allowed
)
2600 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
2605 unsigned long flags
;
2607 rcu_read_lock_sched();
2608 dl_b
= dl_bw_of(dest_cpu
);
2609 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2610 cpus
= dl_bw_cpus(dest_cpu
);
2611 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
2616 * We reserve space for this task in the destination
2617 * root_domain, as we can't fail after this point.
2618 * We will free resources in the source root_domain
2619 * later on (see set_cpus_allowed_dl()).
2621 __dl_add(dl_b
, p
->dl
.dl_bw
, cpus
);
2624 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2625 rcu_read_unlock_sched();
2629 int dl_cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
2630 const struct cpumask
*trial
)
2632 int ret
= 1, trial_cpus
;
2633 struct dl_bw
*cur_dl_b
;
2634 unsigned long flags
;
2636 rcu_read_lock_sched();
2637 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
2638 trial_cpus
= cpumask_weight(trial
);
2640 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
2641 if (cur_dl_b
->bw
!= -1 &&
2642 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
2644 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
2645 rcu_read_unlock_sched();
2649 bool dl_cpu_busy(unsigned int cpu
)
2651 unsigned long flags
;
2656 rcu_read_lock_sched();
2657 dl_b
= dl_bw_of(cpu
);
2658 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2659 cpus
= dl_bw_cpus(cpu
);
2660 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
2661 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2662 rcu_read_unlock_sched();
2667 #ifdef CONFIG_SCHED_DEBUG
2668 extern void print_dl_rq(struct seq_file
*m
, int cpu
, struct dl_rq
*dl_rq
);
2670 void print_dl_stats(struct seq_file
*m
, int cpu
)
2672 print_dl_rq(m
, cpu
, &cpu_rq(cpu
)->dl
);
2674 #endif /* CONFIG_SCHED_DEBUG */