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");
62 if (cpumask_subset(rd
->span
, cpu_active_mask
))
63 return cpumask_weight(rd
->span
);
67 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
73 static inline unsigned long __dl_bw_capacity(int i
)
75 struct root_domain
*rd
= cpu_rq(i
)->rd
;
76 unsigned long cap
= 0;
78 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
79 "sched RCU must be held");
81 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
82 cap
+= capacity_orig_of(i
);
88 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
89 * of the CPU the task is running on rather rd's \Sum CPU capacity.
91 static inline unsigned long dl_bw_capacity(int i
)
93 if (!static_branch_unlikely(&sched_asym_cpucapacity
) &&
94 capacity_orig_of(i
) == SCHED_CAPACITY_SCALE
) {
95 return dl_bw_cpus(i
) << SCHED_CAPACITY_SHIFT
;
97 return __dl_bw_capacity(i
);
101 static inline struct dl_bw
*dl_bw_of(int i
)
103 return &cpu_rq(i
)->dl
.dl_bw
;
106 static inline int dl_bw_cpus(int i
)
111 static inline unsigned long dl_bw_capacity(int i
)
113 return SCHED_CAPACITY_SCALE
;
118 void __add_running_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
120 u64 old
= dl_rq
->running_bw
;
122 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
123 dl_rq
->running_bw
+= dl_bw
;
124 SCHED_WARN_ON(dl_rq
->running_bw
< old
); /* overflow */
125 SCHED_WARN_ON(dl_rq
->running_bw
> dl_rq
->this_bw
);
126 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
127 cpufreq_update_util(rq_of_dl_rq(dl_rq
), 0);
131 void __sub_running_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
133 u64 old
= dl_rq
->running_bw
;
135 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
136 dl_rq
->running_bw
-= dl_bw
;
137 SCHED_WARN_ON(dl_rq
->running_bw
> old
); /* underflow */
138 if (dl_rq
->running_bw
> old
)
139 dl_rq
->running_bw
= 0;
140 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
141 cpufreq_update_util(rq_of_dl_rq(dl_rq
), 0);
145 void __add_rq_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
147 u64 old
= dl_rq
->this_bw
;
149 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
150 dl_rq
->this_bw
+= dl_bw
;
151 SCHED_WARN_ON(dl_rq
->this_bw
< old
); /* overflow */
155 void __sub_rq_bw(u64 dl_bw
, struct dl_rq
*dl_rq
)
157 u64 old
= dl_rq
->this_bw
;
159 lockdep_assert_held(&(rq_of_dl_rq(dl_rq
))->lock
);
160 dl_rq
->this_bw
-= dl_bw
;
161 SCHED_WARN_ON(dl_rq
->this_bw
> old
); /* underflow */
162 if (dl_rq
->this_bw
> old
)
164 SCHED_WARN_ON(dl_rq
->running_bw
> dl_rq
->this_bw
);
168 void add_rq_bw(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
170 if (!dl_entity_is_special(dl_se
))
171 __add_rq_bw(dl_se
->dl_bw
, dl_rq
);
175 void sub_rq_bw(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
177 if (!dl_entity_is_special(dl_se
))
178 __sub_rq_bw(dl_se
->dl_bw
, dl_rq
);
182 void add_running_bw(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
184 if (!dl_entity_is_special(dl_se
))
185 __add_running_bw(dl_se
->dl_bw
, dl_rq
);
189 void sub_running_bw(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
191 if (!dl_entity_is_special(dl_se
))
192 __sub_running_bw(dl_se
->dl_bw
, dl_rq
);
195 static void dl_change_utilization(struct task_struct
*p
, u64 new_bw
)
199 BUG_ON(p
->dl
.flags
& SCHED_FLAG_SUGOV
);
201 if (task_on_rq_queued(p
))
205 if (p
->dl
.dl_non_contending
) {
206 sub_running_bw(&p
->dl
, &rq
->dl
);
207 p
->dl
.dl_non_contending
= 0;
209 * If the timer handler is currently running and the
210 * timer cannot be cancelled, inactive_task_timer()
211 * will see that dl_not_contending is not set, and
212 * will not touch the rq's active utilization,
213 * so we are still safe.
215 if (hrtimer_try_to_cancel(&p
->dl
.inactive_timer
) == 1)
218 __sub_rq_bw(p
->dl
.dl_bw
, &rq
->dl
);
219 __add_rq_bw(new_bw
, &rq
->dl
);
223 * The utilization of a task cannot be immediately removed from
224 * the rq active utilization (running_bw) when the task blocks.
225 * Instead, we have to wait for the so called "0-lag time".
227 * If a task blocks before the "0-lag time", a timer (the inactive
228 * timer) is armed, and running_bw is decreased when the timer
231 * If the task wakes up again before the inactive timer fires,
232 * the timer is cancelled, whereas if the task wakes up after the
233 * inactive timer fired (and running_bw has been decreased) the
234 * task's utilization has to be added to running_bw again.
235 * A flag in the deadline scheduling entity (dl_non_contending)
236 * is used to avoid race conditions between the inactive timer handler
239 * The following diagram shows how running_bw is updated. A task is
240 * "ACTIVE" when its utilization contributes to running_bw; an
241 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
242 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
243 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
244 * time already passed, which does not contribute to running_bw anymore.
245 * +------------------+
247 * +------------------>+ contending |
248 * | add_running_bw | |
249 * | +----+------+------+
252 * +--------+-------+ | |
253 * | | t >= 0-lag | | wakeup
254 * | INACTIVE |<---------------+ |
255 * | | sub_running_bw | |
256 * +--------+-------+ | |
261 * | +----+------+------+
262 * | sub_running_bw | ACTIVE |
263 * +-------------------+ |
264 * inactive timer | non contending |
265 * fired +------------------+
267 * The task_non_contending() function is invoked when a task
268 * blocks, and checks if the 0-lag time already passed or
269 * not (in the first case, it directly updates running_bw;
270 * in the second case, it arms the inactive timer).
272 * The task_contending() function is invoked when a task wakes
273 * up, and checks if the task is still in the "ACTIVE non contending"
274 * state or not (in the second case, it updates running_bw).
276 static void task_non_contending(struct task_struct
*p
)
278 struct sched_dl_entity
*dl_se
= &p
->dl
;
279 struct hrtimer
*timer
= &dl_se
->inactive_timer
;
280 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
281 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
285 * If this is a non-deadline task that has been boosted,
288 if (dl_se
->dl_runtime
== 0)
291 if (dl_entity_is_special(dl_se
))
294 WARN_ON(dl_se
->dl_non_contending
);
296 zerolag_time
= dl_se
->deadline
-
297 div64_long((dl_se
->runtime
* dl_se
->dl_period
),
301 * Using relative times instead of the absolute "0-lag time"
302 * allows to simplify the code
304 zerolag_time
-= rq_clock(rq
);
307 * If the "0-lag time" already passed, decrease the active
308 * utilization now, instead of starting a timer
310 if ((zerolag_time
< 0) || hrtimer_active(&dl_se
->inactive_timer
)) {
312 sub_running_bw(dl_se
, dl_rq
);
313 if (!dl_task(p
) || p
->state
== TASK_DEAD
) {
314 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
316 if (p
->state
== TASK_DEAD
)
317 sub_rq_bw(&p
->dl
, &rq
->dl
);
318 raw_spin_lock(&dl_b
->lock
);
319 __dl_sub(dl_b
, p
->dl
.dl_bw
, dl_bw_cpus(task_cpu(p
)));
320 __dl_clear_params(p
);
321 raw_spin_unlock(&dl_b
->lock
);
327 dl_se
->dl_non_contending
= 1;
329 hrtimer_start(timer
, ns_to_ktime(zerolag_time
), HRTIMER_MODE_REL_HARD
);
332 static void task_contending(struct sched_dl_entity
*dl_se
, int flags
)
334 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
337 * If this is a non-deadline task that has been boosted,
340 if (dl_se
->dl_runtime
== 0)
343 if (flags
& ENQUEUE_MIGRATED
)
344 add_rq_bw(dl_se
, dl_rq
);
346 if (dl_se
->dl_non_contending
) {
347 dl_se
->dl_non_contending
= 0;
349 * If the timer handler is currently running and the
350 * timer cannot be cancelled, inactive_task_timer()
351 * will see that dl_not_contending is not set, and
352 * will not touch the rq's active utilization,
353 * so we are still safe.
355 if (hrtimer_try_to_cancel(&dl_se
->inactive_timer
) == 1)
356 put_task_struct(dl_task_of(dl_se
));
359 * Since "dl_non_contending" is not set, the
360 * task's utilization has already been removed from
361 * active utilization (either when the task blocked,
362 * when the "inactive timer" fired).
365 add_running_bw(dl_se
, dl_rq
);
369 static inline int is_leftmost(struct task_struct
*p
, struct dl_rq
*dl_rq
)
371 struct sched_dl_entity
*dl_se
= &p
->dl
;
373 return dl_rq
->root
.rb_leftmost
== &dl_se
->rb_node
;
376 static void init_dl_rq_bw_ratio(struct dl_rq
*dl_rq
);
378 void init_dl_bandwidth(struct dl_bandwidth
*dl_b
, u64 period
, u64 runtime
)
380 raw_spin_lock_init(&dl_b
->dl_runtime_lock
);
381 dl_b
->dl_period
= period
;
382 dl_b
->dl_runtime
= runtime
;
385 void init_dl_bw(struct dl_bw
*dl_b
)
387 raw_spin_lock_init(&dl_b
->lock
);
388 raw_spin_lock(&def_dl_bandwidth
.dl_runtime_lock
);
389 if (global_rt_runtime() == RUNTIME_INF
)
392 dl_b
->bw
= to_ratio(global_rt_period(), global_rt_runtime());
393 raw_spin_unlock(&def_dl_bandwidth
.dl_runtime_lock
);
397 void init_dl_rq(struct dl_rq
*dl_rq
)
399 dl_rq
->root
= RB_ROOT_CACHED
;
402 /* zero means no -deadline tasks */
403 dl_rq
->earliest_dl
.curr
= dl_rq
->earliest_dl
.next
= 0;
405 dl_rq
->dl_nr_migratory
= 0;
406 dl_rq
->overloaded
= 0;
407 dl_rq
->pushable_dl_tasks_root
= RB_ROOT_CACHED
;
409 init_dl_bw(&dl_rq
->dl_bw
);
412 dl_rq
->running_bw
= 0;
414 init_dl_rq_bw_ratio(dl_rq
);
419 static inline int dl_overloaded(struct rq
*rq
)
421 return atomic_read(&rq
->rd
->dlo_count
);
424 static inline void dl_set_overload(struct rq
*rq
)
429 cpumask_set_cpu(rq
->cpu
, rq
->rd
->dlo_mask
);
431 * Must be visible before the overload count is
432 * set (as in sched_rt.c).
434 * Matched by the barrier in pull_dl_task().
437 atomic_inc(&rq
->rd
->dlo_count
);
440 static inline void dl_clear_overload(struct rq
*rq
)
445 atomic_dec(&rq
->rd
->dlo_count
);
446 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->dlo_mask
);
449 static void update_dl_migration(struct dl_rq
*dl_rq
)
451 if (dl_rq
->dl_nr_migratory
&& dl_rq
->dl_nr_running
> 1) {
452 if (!dl_rq
->overloaded
) {
453 dl_set_overload(rq_of_dl_rq(dl_rq
));
454 dl_rq
->overloaded
= 1;
456 } else if (dl_rq
->overloaded
) {
457 dl_clear_overload(rq_of_dl_rq(dl_rq
));
458 dl_rq
->overloaded
= 0;
462 static void inc_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
464 struct task_struct
*p
= dl_task_of(dl_se
);
466 if (p
->nr_cpus_allowed
> 1)
467 dl_rq
->dl_nr_migratory
++;
469 update_dl_migration(dl_rq
);
472 static void dec_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
474 struct task_struct
*p
= dl_task_of(dl_se
);
476 if (p
->nr_cpus_allowed
> 1)
477 dl_rq
->dl_nr_migratory
--;
479 update_dl_migration(dl_rq
);
483 * The list of pushable -deadline task is not a plist, like in
484 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
486 static void enqueue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
488 struct dl_rq
*dl_rq
= &rq
->dl
;
489 struct rb_node
**link
= &dl_rq
->pushable_dl_tasks_root
.rb_root
.rb_node
;
490 struct rb_node
*parent
= NULL
;
491 struct task_struct
*entry
;
492 bool leftmost
= true;
494 BUG_ON(!RB_EMPTY_NODE(&p
->pushable_dl_tasks
));
498 entry
= rb_entry(parent
, struct task_struct
,
500 if (dl_entity_preempt(&p
->dl
, &entry
->dl
))
501 link
= &parent
->rb_left
;
503 link
= &parent
->rb_right
;
509 dl_rq
->earliest_dl
.next
= p
->dl
.deadline
;
511 rb_link_node(&p
->pushable_dl_tasks
, parent
, link
);
512 rb_insert_color_cached(&p
->pushable_dl_tasks
,
513 &dl_rq
->pushable_dl_tasks_root
, leftmost
);
516 static void dequeue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
518 struct dl_rq
*dl_rq
= &rq
->dl
;
520 if (RB_EMPTY_NODE(&p
->pushable_dl_tasks
))
523 if (dl_rq
->pushable_dl_tasks_root
.rb_leftmost
== &p
->pushable_dl_tasks
) {
524 struct rb_node
*next_node
;
526 next_node
= rb_next(&p
->pushable_dl_tasks
);
528 dl_rq
->earliest_dl
.next
= rb_entry(next_node
,
529 struct task_struct
, pushable_dl_tasks
)->dl
.deadline
;
533 rb_erase_cached(&p
->pushable_dl_tasks
, &dl_rq
->pushable_dl_tasks_root
);
534 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
537 static inline int has_pushable_dl_tasks(struct rq
*rq
)
539 return !RB_EMPTY_ROOT(&rq
->dl
.pushable_dl_tasks_root
.rb_root
);
542 static int push_dl_task(struct rq
*rq
);
544 static inline bool need_pull_dl_task(struct rq
*rq
, struct task_struct
*prev
)
546 return dl_task(prev
);
549 static DEFINE_PER_CPU(struct callback_head
, dl_push_head
);
550 static DEFINE_PER_CPU(struct callback_head
, dl_pull_head
);
552 static void push_dl_tasks(struct rq
*);
553 static void pull_dl_task(struct rq
*);
555 static inline void deadline_queue_push_tasks(struct rq
*rq
)
557 if (!has_pushable_dl_tasks(rq
))
560 queue_balance_callback(rq
, &per_cpu(dl_push_head
, rq
->cpu
), push_dl_tasks
);
563 static inline void deadline_queue_pull_task(struct rq
*rq
)
565 queue_balance_callback(rq
, &per_cpu(dl_pull_head
, rq
->cpu
), pull_dl_task
);
568 static struct rq
*find_lock_later_rq(struct task_struct
*task
, struct rq
*rq
);
570 static struct rq
*dl_task_offline_migration(struct rq
*rq
, struct task_struct
*p
)
572 struct rq
*later_rq
= NULL
;
575 later_rq
= find_lock_later_rq(p
, rq
);
580 * If we cannot preempt any rq, fall back to pick any
583 cpu
= cpumask_any_and(cpu_active_mask
, p
->cpus_ptr
);
584 if (cpu
>= nr_cpu_ids
) {
586 * Failed to find any suitable CPU.
587 * The task will never come back!
589 BUG_ON(dl_bandwidth_enabled());
592 * If admission control is disabled we
593 * try a little harder to let the task
596 cpu
= cpumask_any(cpu_active_mask
);
598 later_rq
= cpu_rq(cpu
);
599 double_lock_balance(rq
, later_rq
);
602 if (p
->dl
.dl_non_contending
|| p
->dl
.dl_throttled
) {
604 * Inactive timer is armed (or callback is running, but
605 * waiting for us to release rq locks). In any case, when it
606 * will fire (or continue), it will see running_bw of this
607 * task migrated to later_rq (and correctly handle it).
609 sub_running_bw(&p
->dl
, &rq
->dl
);
610 sub_rq_bw(&p
->dl
, &rq
->dl
);
612 add_rq_bw(&p
->dl
, &later_rq
->dl
);
613 add_running_bw(&p
->dl
, &later_rq
->dl
);
615 sub_rq_bw(&p
->dl
, &rq
->dl
);
616 add_rq_bw(&p
->dl
, &later_rq
->dl
);
620 * And we finally need to fixup root_domain(s) bandwidth accounting,
621 * since p is still hanging out in the old (now moved to default) root
624 dl_b
= &rq
->rd
->dl_bw
;
625 raw_spin_lock(&dl_b
->lock
);
626 __dl_sub(dl_b
, p
->dl
.dl_bw
, cpumask_weight(rq
->rd
->span
));
627 raw_spin_unlock(&dl_b
->lock
);
629 dl_b
= &later_rq
->rd
->dl_bw
;
630 raw_spin_lock(&dl_b
->lock
);
631 __dl_add(dl_b
, p
->dl
.dl_bw
, cpumask_weight(later_rq
->rd
->span
));
632 raw_spin_unlock(&dl_b
->lock
);
634 set_task_cpu(p
, later_rq
->cpu
);
635 double_unlock_balance(later_rq
, rq
);
643 void enqueue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
648 void dequeue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
653 void inc_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
658 void dec_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
662 static inline bool need_pull_dl_task(struct rq
*rq
, struct task_struct
*prev
)
667 static inline void pull_dl_task(struct rq
*rq
)
671 static inline void deadline_queue_push_tasks(struct rq
*rq
)
675 static inline void deadline_queue_pull_task(struct rq
*rq
)
678 #endif /* CONFIG_SMP */
680 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
681 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
682 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
685 * We are being explicitly informed that a new instance is starting,
686 * and this means that:
687 * - the absolute deadline of the entity has to be placed at
688 * current time + relative deadline;
689 * - the runtime of the entity has to be set to the maximum value.
691 * The capability of specifying such event is useful whenever a -deadline
692 * entity wants to (try to!) synchronize its behaviour with the scheduler's
693 * one, and to (try to!) reconcile itself with its own scheduling
696 static inline void setup_new_dl_entity(struct sched_dl_entity
*dl_se
)
698 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
699 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
701 WARN_ON(dl_se
->dl_boosted
);
702 WARN_ON(dl_time_before(rq_clock(rq
), dl_se
->deadline
));
705 * We are racing with the deadline timer. So, do nothing because
706 * the deadline timer handler will take care of properly recharging
707 * the runtime and postponing the deadline
709 if (dl_se
->dl_throttled
)
713 * We use the regular wall clock time to set deadlines in the
714 * future; in fact, we must consider execution overheads (time
715 * spent on hardirq context, etc.).
717 dl_se
->deadline
= rq_clock(rq
) + dl_se
->dl_deadline
;
718 dl_se
->runtime
= dl_se
->dl_runtime
;
722 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
723 * possibility of a entity lasting more than what it declared, and thus
724 * exhausting its runtime.
726 * Here we are interested in making runtime overrun possible, but we do
727 * not want a entity which is misbehaving to affect the scheduling of all
729 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
730 * is used, in order to confine each entity within its own bandwidth.
732 * This function deals exactly with that, and ensures that when the runtime
733 * of a entity is replenished, its deadline is also postponed. That ensures
734 * the overrunning entity can't interfere with other entity in the system and
735 * can't make them miss their deadlines. Reasons why this kind of overruns
736 * could happen are, typically, a entity voluntarily trying to overcome its
737 * runtime, or it just underestimated it during sched_setattr().
739 static void replenish_dl_entity(struct sched_dl_entity
*dl_se
,
740 struct sched_dl_entity
*pi_se
)
742 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
743 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
745 BUG_ON(pi_se
->dl_runtime
<= 0);
748 * This could be the case for a !-dl task that is boosted.
749 * Just go with full inherited parameters.
751 if (dl_se
->dl_deadline
== 0) {
752 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
753 dl_se
->runtime
= pi_se
->dl_runtime
;
756 if (dl_se
->dl_yielded
&& dl_se
->runtime
> 0)
760 * We keep moving the deadline away until we get some
761 * available runtime for the entity. This ensures correct
762 * handling of situations where the runtime overrun is
765 while (dl_se
->runtime
<= 0) {
766 dl_se
->deadline
+= pi_se
->dl_period
;
767 dl_se
->runtime
+= pi_se
->dl_runtime
;
771 * At this point, the deadline really should be "in
772 * the future" with respect to rq->clock. If it's
773 * not, we are, for some reason, lagging too much!
774 * Anyway, after having warn userspace abut that,
775 * we still try to keep the things running by
776 * resetting the deadline and the budget of the
779 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
))) {
780 printk_deferred_once("sched: DL replenish lagged too much\n");
781 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
782 dl_se
->runtime
= pi_se
->dl_runtime
;
785 if (dl_se
->dl_yielded
)
786 dl_se
->dl_yielded
= 0;
787 if (dl_se
->dl_throttled
)
788 dl_se
->dl_throttled
= 0;
792 * Here we check if --at time t-- an entity (which is probably being
793 * [re]activated or, in general, enqueued) can use its remaining runtime
794 * and its current deadline _without_ exceeding the bandwidth it is
795 * assigned (function returns true if it can't). We are in fact applying
796 * one of the CBS rules: when a task wakes up, if the residual runtime
797 * over residual deadline fits within the allocated bandwidth, then we
798 * can keep the current (absolute) deadline and residual budget without
799 * disrupting the schedulability of the system. Otherwise, we should
800 * refill the runtime and set the deadline a period in the future,
801 * because keeping the current (absolute) deadline of the task would
802 * result in breaking guarantees promised to other tasks (refer to
803 * Documentation/scheduler/sched-deadline.rst for more information).
805 * This function returns true if:
807 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
809 * IOW we can't recycle current parameters.
811 * Notice that the bandwidth check is done against the deadline. For
812 * task with deadline equal to period this is the same of using
813 * dl_period instead of dl_deadline in the equation above.
815 static bool dl_entity_overflow(struct sched_dl_entity
*dl_se
,
816 struct sched_dl_entity
*pi_se
, u64 t
)
821 * left and right are the two sides of the equation above,
822 * after a bit of shuffling to use multiplications instead
825 * Note that none of the time values involved in the two
826 * multiplications are absolute: dl_deadline and dl_runtime
827 * are the relative deadline and the maximum runtime of each
828 * instance, runtime is the runtime left for the last instance
829 * and (deadline - t), since t is rq->clock, is the time left
830 * to the (absolute) deadline. Even if overflowing the u64 type
831 * is very unlikely to occur in both cases, here we scale down
832 * as we want to avoid that risk at all. Scaling down by 10
833 * means that we reduce granularity to 1us. We are fine with it,
834 * since this is only a true/false check and, anyway, thinking
835 * of anything below microseconds resolution is actually fiction
836 * (but still we want to give the user that illusion >;).
838 left
= (pi_se
->dl_deadline
>> DL_SCALE
) * (dl_se
->runtime
>> DL_SCALE
);
839 right
= ((dl_se
->deadline
- t
) >> DL_SCALE
) *
840 (pi_se
->dl_runtime
>> DL_SCALE
);
842 return dl_time_before(right
, left
);
846 * Revised wakeup rule [1]: For self-suspending tasks, rather then
847 * re-initializing task's runtime and deadline, the revised wakeup
848 * rule adjusts the task's runtime to avoid the task to overrun its
851 * Reasoning: a task may overrun the density if:
852 * runtime / (deadline - t) > dl_runtime / dl_deadline
854 * Therefore, runtime can be adjusted to:
855 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
857 * In such way that runtime will be equal to the maximum density
858 * the task can use without breaking any rule.
860 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
861 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
864 update_dl_revised_wakeup(struct sched_dl_entity
*dl_se
, struct rq
*rq
)
866 u64 laxity
= dl_se
->deadline
- rq_clock(rq
);
869 * If the task has deadline < period, and the deadline is in the past,
870 * it should already be throttled before this check.
872 * See update_dl_entity() comments for further details.
874 WARN_ON(dl_time_before(dl_se
->deadline
, rq_clock(rq
)));
876 dl_se
->runtime
= (dl_se
->dl_density
* laxity
) >> BW_SHIFT
;
880 * Regarding the deadline, a task with implicit deadline has a relative
881 * deadline == relative period. A task with constrained deadline has a
882 * relative deadline <= relative period.
884 * We support constrained deadline tasks. However, there are some restrictions
885 * applied only for tasks which do not have an implicit deadline. See
886 * update_dl_entity() to know more about such restrictions.
888 * The dl_is_implicit() returns true if the task has an implicit deadline.
890 static inline bool dl_is_implicit(struct sched_dl_entity
*dl_se
)
892 return dl_se
->dl_deadline
== dl_se
->dl_period
;
896 * When a deadline entity is placed in the runqueue, its runtime and deadline
897 * might need to be updated. This is done by a CBS wake up rule. There are two
898 * different rules: 1) the original CBS; and 2) the Revisited CBS.
900 * When the task is starting a new period, the Original CBS is used. In this
901 * case, the runtime is replenished and a new absolute deadline is set.
903 * When a task is queued before the begin of the next period, using the
904 * remaining runtime and deadline could make the entity to overflow, see
905 * dl_entity_overflow() to find more about runtime overflow. When such case
906 * is detected, the runtime and deadline need to be updated.
908 * If the task has an implicit deadline, i.e., deadline == period, the Original
909 * CBS is applied. the runtime is replenished and a new absolute deadline is
910 * set, as in the previous cases.
912 * However, the Original CBS does not work properly for tasks with
913 * deadline < period, which are said to have a constrained deadline. By
914 * applying the Original CBS, a constrained deadline task would be able to run
915 * runtime/deadline in a period. With deadline < period, the task would
916 * overrun the runtime/period allowed bandwidth, breaking the admission test.
918 * In order to prevent this misbehave, the Revisited CBS is used for
919 * constrained deadline tasks when a runtime overflow is detected. In the
920 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
921 * the remaining runtime of the task is reduced to avoid runtime overflow.
922 * Please refer to the comments update_dl_revised_wakeup() function to find
923 * more about the Revised CBS rule.
925 static void update_dl_entity(struct sched_dl_entity
*dl_se
,
926 struct sched_dl_entity
*pi_se
)
928 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
929 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
931 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
)) ||
932 dl_entity_overflow(dl_se
, pi_se
, rq_clock(rq
))) {
934 if (unlikely(!dl_is_implicit(dl_se
) &&
935 !dl_time_before(dl_se
->deadline
, rq_clock(rq
)) &&
936 !dl_se
->dl_boosted
)){
937 update_dl_revised_wakeup(dl_se
, rq
);
941 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
942 dl_se
->runtime
= pi_se
->dl_runtime
;
946 static inline u64
dl_next_period(struct sched_dl_entity
*dl_se
)
948 return dl_se
->deadline
- dl_se
->dl_deadline
+ dl_se
->dl_period
;
952 * If the entity depleted all its runtime, and if we want it to sleep
953 * while waiting for some new execution time to become available, we
954 * set the bandwidth replenishment timer to the replenishment instant
955 * and try to activate it.
957 * Notice that it is important for the caller to know if the timer
958 * actually started or not (i.e., the replenishment instant is in
959 * the future or in the past).
961 static int start_dl_timer(struct task_struct
*p
)
963 struct sched_dl_entity
*dl_se
= &p
->dl
;
964 struct hrtimer
*timer
= &dl_se
->dl_timer
;
965 struct rq
*rq
= task_rq(p
);
969 lockdep_assert_held(&rq
->lock
);
972 * We want the timer to fire at the deadline, but considering
973 * that it is actually coming from rq->clock and not from
974 * hrtimer's time base reading.
976 act
= ns_to_ktime(dl_next_period(dl_se
));
977 now
= hrtimer_cb_get_time(timer
);
978 delta
= ktime_to_ns(now
) - rq_clock(rq
);
979 act
= ktime_add_ns(act
, delta
);
982 * If the expiry time already passed, e.g., because the value
983 * chosen as the deadline is too small, don't even try to
984 * start the timer in the past!
986 if (ktime_us_delta(act
, now
) < 0)
990 * !enqueued will guarantee another callback; even if one is already in
991 * progress. This ensures a balanced {get,put}_task_struct().
993 * The race against __run_timer() clearing the enqueued state is
994 * harmless because we're holding task_rq()->lock, therefore the timer
995 * expiring after we've done the check will wait on its task_rq_lock()
996 * and observe our state.
998 if (!hrtimer_is_queued(timer
)) {
1000 hrtimer_start(timer
, act
, HRTIMER_MODE_ABS_HARD
);
1007 * This is the bandwidth enforcement timer callback. If here, we know
1008 * a task is not on its dl_rq, since the fact that the timer was running
1009 * means the task is throttled and needs a runtime replenishment.
1011 * However, what we actually do depends on the fact the task is active,
1012 * (it is on its rq) or has been removed from there by a call to
1013 * dequeue_task_dl(). In the former case we must issue the runtime
1014 * replenishment and add the task back to the dl_rq; in the latter, we just
1015 * do nothing but clearing dl_throttled, so that runtime and deadline
1016 * updating (and the queueing back to dl_rq) will be done by the
1017 * next call to enqueue_task_dl().
1019 static enum hrtimer_restart
dl_task_timer(struct hrtimer
*timer
)
1021 struct sched_dl_entity
*dl_se
= container_of(timer
,
1022 struct sched_dl_entity
,
1024 struct task_struct
*p
= dl_task_of(dl_se
);
1028 rq
= task_rq_lock(p
, &rf
);
1031 * The task might have changed its scheduling policy to something
1032 * different than SCHED_DEADLINE (through switched_from_dl()).
1038 * The task might have been boosted by someone else and might be in the
1039 * boosting/deboosting path, its not throttled.
1041 if (dl_se
->dl_boosted
)
1045 * Spurious timer due to start_dl_timer() race; or we already received
1046 * a replenishment from rt_mutex_setprio().
1048 if (!dl_se
->dl_throttled
)
1052 update_rq_clock(rq
);
1055 * If the throttle happened during sched-out; like:
1062 * __dequeue_task_dl()
1065 * We can be both throttled and !queued. Replenish the counter
1066 * but do not enqueue -- wait for our wakeup to do that.
1068 if (!task_on_rq_queued(p
)) {
1069 replenish_dl_entity(dl_se
, dl_se
);
1074 if (unlikely(!rq
->online
)) {
1076 * If the runqueue is no longer available, migrate the
1077 * task elsewhere. This necessarily changes rq.
1079 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
1080 rq
= dl_task_offline_migration(rq
, p
);
1081 rf
.cookie
= lockdep_pin_lock(&rq
->lock
);
1082 update_rq_clock(rq
);
1085 * Now that the task has been migrated to the new RQ and we
1086 * have that locked, proceed as normal and enqueue the task
1092 enqueue_task_dl(rq
, p
, ENQUEUE_REPLENISH
);
1093 if (dl_task(rq
->curr
))
1094 check_preempt_curr_dl(rq
, p
, 0);
1100 * Queueing this task back might have overloaded rq, check if we need
1101 * to kick someone away.
1103 if (has_pushable_dl_tasks(rq
)) {
1105 * Nothing relies on rq->lock after this, so its safe to drop
1108 rq_unpin_lock(rq
, &rf
);
1110 rq_repin_lock(rq
, &rf
);
1115 task_rq_unlock(rq
, p
, &rf
);
1118 * This can free the task_struct, including this hrtimer, do not touch
1119 * anything related to that after this.
1123 return HRTIMER_NORESTART
;
1126 void init_dl_task_timer(struct sched_dl_entity
*dl_se
)
1128 struct hrtimer
*timer
= &dl_se
->dl_timer
;
1130 hrtimer_init(timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
1131 timer
->function
= dl_task_timer
;
1135 * During the activation, CBS checks if it can reuse the current task's
1136 * runtime and period. If the deadline of the task is in the past, CBS
1137 * cannot use the runtime, and so it replenishes the task. This rule
1138 * works fine for implicit deadline tasks (deadline == period), and the
1139 * CBS was designed for implicit deadline tasks. However, a task with
1140 * constrained deadline (deadline < period) might be awakened after the
1141 * deadline, but before the next period. In this case, replenishing the
1142 * task would allow it to run for runtime / deadline. As in this case
1143 * deadline < period, CBS enables a task to run for more than the
1144 * runtime / period. In a very loaded system, this can cause a domino
1145 * effect, making other tasks miss their deadlines.
1147 * To avoid this problem, in the activation of a constrained deadline
1148 * task after the deadline but before the next period, throttle the
1149 * task and set the replenishing timer to the begin of the next period,
1150 * unless it is boosted.
1152 static inline void dl_check_constrained_dl(struct sched_dl_entity
*dl_se
)
1154 struct task_struct
*p
= dl_task_of(dl_se
);
1155 struct rq
*rq
= rq_of_dl_rq(dl_rq_of_se(dl_se
));
1157 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
)) &&
1158 dl_time_before(rq_clock(rq
), dl_next_period(dl_se
))) {
1159 if (unlikely(dl_se
->dl_boosted
|| !start_dl_timer(p
)))
1161 dl_se
->dl_throttled
= 1;
1162 if (dl_se
->runtime
> 0)
1168 int dl_runtime_exceeded(struct sched_dl_entity
*dl_se
)
1170 return (dl_se
->runtime
<= 0);
1173 extern bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
);
1176 * This function implements the GRUB accounting rule:
1177 * according to the GRUB reclaiming algorithm, the runtime is
1178 * not decreased as "dq = -dt", but as
1179 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1180 * where u is the utilization of the task, Umax is the maximum reclaimable
1181 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1182 * as the difference between the "total runqueue utilization" and the
1183 * runqueue active utilization, and Uextra is the (per runqueue) extra
1184 * reclaimable utilization.
1185 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1186 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1188 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1189 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1190 * Since delta is a 64 bit variable, to have an overflow its value
1191 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1192 * So, overflow is not an issue here.
1194 static u64
grub_reclaim(u64 delta
, struct rq
*rq
, struct sched_dl_entity
*dl_se
)
1196 u64 u_inact
= rq
->dl
.this_bw
- rq
->dl
.running_bw
; /* Utot - Uact */
1198 u64 u_act_min
= (dl_se
->dl_bw
* rq
->dl
.bw_ratio
) >> RATIO_SHIFT
;
1201 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1202 * we compare u_inact + rq->dl.extra_bw with
1203 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1204 * u_inact + rq->dl.extra_bw can be larger than
1205 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1206 * leading to wrong results)
1208 if (u_inact
+ rq
->dl
.extra_bw
> BW_UNIT
- u_act_min
)
1211 u_act
= BW_UNIT
- u_inact
- rq
->dl
.extra_bw
;
1213 return (delta
* u_act
) >> BW_SHIFT
;
1217 * Update the current task's runtime statistics (provided it is still
1218 * a -deadline task and has not been removed from the dl_rq).
1220 static void update_curr_dl(struct rq
*rq
)
1222 struct task_struct
*curr
= rq
->curr
;
1223 struct sched_dl_entity
*dl_se
= &curr
->dl
;
1224 u64 delta_exec
, scaled_delta_exec
;
1225 int cpu
= cpu_of(rq
);
1228 if (!dl_task(curr
) || !on_dl_rq(dl_se
))
1232 * Consumed budget is computed considering the time as
1233 * observed by schedulable tasks (excluding time spent
1234 * in hardirq context, etc.). Deadlines are instead
1235 * computed using hard walltime. This seems to be the more
1236 * natural solution, but the full ramifications of this
1237 * approach need further study.
1239 now
= rq_clock_task(rq
);
1240 delta_exec
= now
- curr
->se
.exec_start
;
1241 if (unlikely((s64
)delta_exec
<= 0)) {
1242 if (unlikely(dl_se
->dl_yielded
))
1247 schedstat_set(curr
->se
.statistics
.exec_max
,
1248 max(curr
->se
.statistics
.exec_max
, delta_exec
));
1250 curr
->se
.sum_exec_runtime
+= delta_exec
;
1251 account_group_exec_runtime(curr
, delta_exec
);
1253 curr
->se
.exec_start
= now
;
1254 cgroup_account_cputime(curr
, delta_exec
);
1256 if (dl_entity_is_special(dl_se
))
1260 * For tasks that participate in GRUB, we implement GRUB-PA: the
1261 * spare reclaimed bandwidth is used to clock down frequency.
1263 * For the others, we still need to scale reservation parameters
1264 * according to current frequency and CPU maximum capacity.
1266 if (unlikely(dl_se
->flags
& SCHED_FLAG_RECLAIM
)) {
1267 scaled_delta_exec
= grub_reclaim(delta_exec
,
1271 unsigned long scale_freq
= arch_scale_freq_capacity(cpu
);
1272 unsigned long scale_cpu
= arch_scale_cpu_capacity(cpu
);
1274 scaled_delta_exec
= cap_scale(delta_exec
, scale_freq
);
1275 scaled_delta_exec
= cap_scale(scaled_delta_exec
, scale_cpu
);
1278 dl_se
->runtime
-= scaled_delta_exec
;
1281 if (dl_runtime_exceeded(dl_se
) || dl_se
->dl_yielded
) {
1282 dl_se
->dl_throttled
= 1;
1284 /* If requested, inform the user about runtime overruns. */
1285 if (dl_runtime_exceeded(dl_se
) &&
1286 (dl_se
->flags
& SCHED_FLAG_DL_OVERRUN
))
1287 dl_se
->dl_overrun
= 1;
1289 __dequeue_task_dl(rq
, curr
, 0);
1290 if (unlikely(dl_se
->dl_boosted
|| !start_dl_timer(curr
)))
1291 enqueue_task_dl(rq
, curr
, ENQUEUE_REPLENISH
);
1293 if (!is_leftmost(curr
, &rq
->dl
))
1298 * Because -- for now -- we share the rt bandwidth, we need to
1299 * account our runtime there too, otherwise actual rt tasks
1300 * would be able to exceed the shared quota.
1302 * Account to the root rt group for now.
1304 * The solution we're working towards is having the RT groups scheduled
1305 * using deadline servers -- however there's a few nasties to figure
1306 * out before that can happen.
1308 if (rt_bandwidth_enabled()) {
1309 struct rt_rq
*rt_rq
= &rq
->rt
;
1311 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
1313 * We'll let actual RT tasks worry about the overflow here, we
1314 * have our own CBS to keep us inline; only account when RT
1315 * bandwidth is relevant.
1317 if (sched_rt_bandwidth_account(rt_rq
))
1318 rt_rq
->rt_time
+= delta_exec
;
1319 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
1323 static enum hrtimer_restart
inactive_task_timer(struct hrtimer
*timer
)
1325 struct sched_dl_entity
*dl_se
= container_of(timer
,
1326 struct sched_dl_entity
,
1328 struct task_struct
*p
= dl_task_of(dl_se
);
1332 rq
= task_rq_lock(p
, &rf
);
1335 update_rq_clock(rq
);
1337 if (!dl_task(p
) || p
->state
== TASK_DEAD
) {
1338 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
1340 if (p
->state
== TASK_DEAD
&& dl_se
->dl_non_contending
) {
1341 sub_running_bw(&p
->dl
, dl_rq_of_se(&p
->dl
));
1342 sub_rq_bw(&p
->dl
, dl_rq_of_se(&p
->dl
));
1343 dl_se
->dl_non_contending
= 0;
1346 raw_spin_lock(&dl_b
->lock
);
1347 __dl_sub(dl_b
, p
->dl
.dl_bw
, dl_bw_cpus(task_cpu(p
)));
1348 raw_spin_unlock(&dl_b
->lock
);
1349 __dl_clear_params(p
);
1353 if (dl_se
->dl_non_contending
== 0)
1356 sub_running_bw(dl_se
, &rq
->dl
);
1357 dl_se
->dl_non_contending
= 0;
1359 task_rq_unlock(rq
, p
, &rf
);
1362 return HRTIMER_NORESTART
;
1365 void init_dl_inactive_task_timer(struct sched_dl_entity
*dl_se
)
1367 struct hrtimer
*timer
= &dl_se
->inactive_timer
;
1369 hrtimer_init(timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
1370 timer
->function
= inactive_task_timer
;
1375 static void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
1377 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
1379 if (dl_rq
->earliest_dl
.curr
== 0 ||
1380 dl_time_before(deadline
, dl_rq
->earliest_dl
.curr
)) {
1381 dl_rq
->earliest_dl
.curr
= deadline
;
1382 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, deadline
);
1386 static void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
1388 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
1391 * Since we may have removed our earliest (and/or next earliest)
1392 * task we must recompute them.
1394 if (!dl_rq
->dl_nr_running
) {
1395 dl_rq
->earliest_dl
.curr
= 0;
1396 dl_rq
->earliest_dl
.next
= 0;
1397 cpudl_clear(&rq
->rd
->cpudl
, rq
->cpu
);
1399 struct rb_node
*leftmost
= dl_rq
->root
.rb_leftmost
;
1400 struct sched_dl_entity
*entry
;
1402 entry
= rb_entry(leftmost
, struct sched_dl_entity
, rb_node
);
1403 dl_rq
->earliest_dl
.curr
= entry
->deadline
;
1404 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, entry
->deadline
);
1410 static inline void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
1411 static inline void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
1413 #endif /* CONFIG_SMP */
1416 void inc_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
1418 int prio
= dl_task_of(dl_se
)->prio
;
1419 u64 deadline
= dl_se
->deadline
;
1421 WARN_ON(!dl_prio(prio
));
1422 dl_rq
->dl_nr_running
++;
1423 add_nr_running(rq_of_dl_rq(dl_rq
), 1);
1425 inc_dl_deadline(dl_rq
, deadline
);
1426 inc_dl_migration(dl_se
, dl_rq
);
1430 void dec_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
1432 int prio
= dl_task_of(dl_se
)->prio
;
1434 WARN_ON(!dl_prio(prio
));
1435 WARN_ON(!dl_rq
->dl_nr_running
);
1436 dl_rq
->dl_nr_running
--;
1437 sub_nr_running(rq_of_dl_rq(dl_rq
), 1);
1439 dec_dl_deadline(dl_rq
, dl_se
->deadline
);
1440 dec_dl_migration(dl_se
, dl_rq
);
1443 static void __enqueue_dl_entity(struct sched_dl_entity
*dl_se
)
1445 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
1446 struct rb_node
**link
= &dl_rq
->root
.rb_root
.rb_node
;
1447 struct rb_node
*parent
= NULL
;
1448 struct sched_dl_entity
*entry
;
1451 BUG_ON(!RB_EMPTY_NODE(&dl_se
->rb_node
));
1455 entry
= rb_entry(parent
, struct sched_dl_entity
, rb_node
);
1456 if (dl_time_before(dl_se
->deadline
, entry
->deadline
))
1457 link
= &parent
->rb_left
;
1459 link
= &parent
->rb_right
;
1464 rb_link_node(&dl_se
->rb_node
, parent
, link
);
1465 rb_insert_color_cached(&dl_se
->rb_node
, &dl_rq
->root
, leftmost
);
1467 inc_dl_tasks(dl_se
, dl_rq
);
1470 static void __dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
1472 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
1474 if (RB_EMPTY_NODE(&dl_se
->rb_node
))
1477 rb_erase_cached(&dl_se
->rb_node
, &dl_rq
->root
);
1478 RB_CLEAR_NODE(&dl_se
->rb_node
);
1480 dec_dl_tasks(dl_se
, dl_rq
);
1484 enqueue_dl_entity(struct sched_dl_entity
*dl_se
,
1485 struct sched_dl_entity
*pi_se
, int flags
)
1487 BUG_ON(on_dl_rq(dl_se
));
1490 * If this is a wakeup or a new instance, the scheduling
1491 * parameters of the task might need updating. Otherwise,
1492 * we want a replenishment of its runtime.
1494 if (flags
& ENQUEUE_WAKEUP
) {
1495 task_contending(dl_se
, flags
);
1496 update_dl_entity(dl_se
, pi_se
);
1497 } else if (flags
& ENQUEUE_REPLENISH
) {
1498 replenish_dl_entity(dl_se
, pi_se
);
1499 } else if ((flags
& ENQUEUE_RESTORE
) &&
1500 dl_time_before(dl_se
->deadline
,
1501 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se
))))) {
1502 setup_new_dl_entity(dl_se
);
1505 __enqueue_dl_entity(dl_se
);
1508 static void dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
1510 __dequeue_dl_entity(dl_se
);
1513 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
1515 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
1516 struct sched_dl_entity
*pi_se
= &p
->dl
;
1519 * Use the scheduling parameters of the top pi-waiter task if:
1520 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
1521 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
1522 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
1523 * boosted due to a SCHED_DEADLINE pi-waiter).
1524 * Otherwise we keep our runtime and deadline.
1526 if (pi_task
&& dl_prio(pi_task
->normal_prio
) && p
->dl
.dl_boosted
) {
1527 pi_se
= &pi_task
->dl
;
1529 * Because of delays in the detection of the overrun of a
1530 * thread's runtime, it might be the case that a thread
1531 * goes to sleep in a rt mutex with negative runtime. As
1532 * a consequence, the thread will be throttled.
1534 * While waiting for the mutex, this thread can also be
1535 * boosted via PI, resulting in a thread that is throttled
1536 * and boosted at the same time.
1538 * In this case, the boost overrides the throttle.
1540 if (p
->dl
.dl_throttled
) {
1542 * The replenish timer needs to be canceled. No
1543 * problem if it fires concurrently: boosted threads
1544 * are ignored in dl_task_timer().
1546 hrtimer_try_to_cancel(&p
->dl
.dl_timer
);
1547 p
->dl
.dl_throttled
= 0;
1549 } else if (!dl_prio(p
->normal_prio
)) {
1551 * Special case in which we have a !SCHED_DEADLINE task that is going
1552 * to be deboosted, but exceeds its runtime while doing so. No point in
1553 * replenishing it, as it's going to return back to its original
1554 * scheduling class after this. If it has been throttled, we need to
1555 * clear the flag, otherwise the task may wake up as throttled after
1556 * being boosted again with no means to replenish the runtime and clear
1559 p
->dl
.dl_throttled
= 0;
1560 BUG_ON(!p
->dl
.dl_boosted
|| flags
!= ENQUEUE_REPLENISH
);
1565 * Check if a constrained deadline task was activated
1566 * after the deadline but before the next period.
1567 * If that is the case, the task will be throttled and
1568 * the replenishment timer will be set to the next period.
1570 if (!p
->dl
.dl_throttled
&& !dl_is_implicit(&p
->dl
))
1571 dl_check_constrained_dl(&p
->dl
);
1573 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
|| flags
& ENQUEUE_RESTORE
) {
1574 add_rq_bw(&p
->dl
, &rq
->dl
);
1575 add_running_bw(&p
->dl
, &rq
->dl
);
1579 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1580 * its budget it needs a replenishment and, since it now is on
1581 * its rq, the bandwidth timer callback (which clearly has not
1582 * run yet) will take care of this.
1583 * However, the active utilization does not depend on the fact
1584 * that the task is on the runqueue or not (but depends on the
1585 * task's state - in GRUB parlance, "inactive" vs "active contending").
1586 * In other words, even if a task is throttled its utilization must
1587 * be counted in the active utilization; hence, we need to call
1590 if (p
->dl
.dl_throttled
&& !(flags
& ENQUEUE_REPLENISH
)) {
1591 if (flags
& ENQUEUE_WAKEUP
)
1592 task_contending(&p
->dl
, flags
);
1597 enqueue_dl_entity(&p
->dl
, pi_se
, flags
);
1599 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
1600 enqueue_pushable_dl_task(rq
, p
);
1603 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
1605 dequeue_dl_entity(&p
->dl
);
1606 dequeue_pushable_dl_task(rq
, p
);
1609 static void dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
1612 __dequeue_task_dl(rq
, p
, flags
);
1614 if (p
->on_rq
== TASK_ON_RQ_MIGRATING
|| flags
& DEQUEUE_SAVE
) {
1615 sub_running_bw(&p
->dl
, &rq
->dl
);
1616 sub_rq_bw(&p
->dl
, &rq
->dl
);
1620 * This check allows to start the inactive timer (or to immediately
1621 * decrease the active utilization, if needed) in two cases:
1622 * when the task blocks and when it is terminating
1623 * (p->state == TASK_DEAD). We can handle the two cases in the same
1624 * way, because from GRUB's point of view the same thing is happening
1625 * (the task moves from "active contending" to "active non contending"
1628 if (flags
& DEQUEUE_SLEEP
)
1629 task_non_contending(p
);
1633 * Yield task semantic for -deadline tasks is:
1635 * get off from the CPU until our next instance, with
1636 * a new runtime. This is of little use now, since we
1637 * don't have a bandwidth reclaiming mechanism. Anyway,
1638 * bandwidth reclaiming is planned for the future, and
1639 * yield_task_dl will indicate that some spare budget
1640 * is available for other task instances to use it.
1642 static void yield_task_dl(struct rq
*rq
)
1645 * We make the task go to sleep until its current deadline by
1646 * forcing its runtime to zero. This way, update_curr_dl() stops
1647 * it and the bandwidth timer will wake it up and will give it
1648 * new scheduling parameters (thanks to dl_yielded=1).
1650 rq
->curr
->dl
.dl_yielded
= 1;
1652 update_rq_clock(rq
);
1655 * Tell update_rq_clock() that we've just updated,
1656 * so we don't do microscopic update in schedule()
1657 * and double the fastpath cost.
1659 rq_clock_skip_update(rq
);
1664 static int find_later_rq(struct task_struct
*task
);
1667 select_task_rq_dl(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
)
1669 struct task_struct
*curr
;
1673 if (sd_flag
!= SD_BALANCE_WAKE
)
1679 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1682 * If we are dealing with a -deadline task, we must
1683 * decide where to wake it up.
1684 * If it has a later deadline and the current task
1685 * on this rq can't move (provided the waking task
1686 * can!) we prefer to send it somewhere else. On the
1687 * other hand, if it has a shorter deadline, we
1688 * try to make it stay here, it might be important.
1690 select_rq
= unlikely(dl_task(curr
)) &&
1691 (curr
->nr_cpus_allowed
< 2 ||
1692 !dl_entity_preempt(&p
->dl
, &curr
->dl
)) &&
1693 p
->nr_cpus_allowed
> 1;
1696 * Take the capacity of the CPU into account to
1697 * ensure it fits the requirement of the task.
1699 if (static_branch_unlikely(&sched_asym_cpucapacity
))
1700 select_rq
|= !dl_task_fits_capacity(p
, cpu
);
1703 int target
= find_later_rq(p
);
1706 (dl_time_before(p
->dl
.deadline
,
1707 cpu_rq(target
)->dl
.earliest_dl
.curr
) ||
1708 (cpu_rq(target
)->dl
.dl_nr_running
== 0)))
1717 static void migrate_task_rq_dl(struct task_struct
*p
, int new_cpu __maybe_unused
)
1721 if (p
->state
!= TASK_WAKING
)
1726 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1727 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1728 * rq->lock is not... So, lock it
1730 raw_spin_lock(&rq
->lock
);
1731 if (p
->dl
.dl_non_contending
) {
1732 sub_running_bw(&p
->dl
, &rq
->dl
);
1733 p
->dl
.dl_non_contending
= 0;
1735 * If the timer handler is currently running and the
1736 * timer cannot be cancelled, inactive_task_timer()
1737 * will see that dl_not_contending is not set, and
1738 * will not touch the rq's active utilization,
1739 * so we are still safe.
1741 if (hrtimer_try_to_cancel(&p
->dl
.inactive_timer
) == 1)
1744 sub_rq_bw(&p
->dl
, &rq
->dl
);
1745 raw_spin_unlock(&rq
->lock
);
1748 static void check_preempt_equal_dl(struct rq
*rq
, struct task_struct
*p
)
1751 * Current can't be migrated, useless to reschedule,
1752 * let's hope p can move out.
1754 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1755 !cpudl_find(&rq
->rd
->cpudl
, rq
->curr
, NULL
))
1759 * p is migratable, so let's not schedule it and
1760 * see if it is pushed or pulled somewhere else.
1762 if (p
->nr_cpus_allowed
!= 1 &&
1763 cpudl_find(&rq
->rd
->cpudl
, p
, NULL
))
1769 static int balance_dl(struct rq
*rq
, struct task_struct
*p
, struct rq_flags
*rf
)
1771 if (!on_dl_rq(&p
->dl
) && need_pull_dl_task(rq
, p
)) {
1773 * This is OK, because current is on_cpu, which avoids it being
1774 * picked for load-balance and preemption/IRQs are still
1775 * disabled avoiding further scheduler activity on it and we've
1776 * not yet started the picking loop.
1778 rq_unpin_lock(rq
, rf
);
1780 rq_repin_lock(rq
, rf
);
1783 return sched_stop_runnable(rq
) || sched_dl_runnable(rq
);
1785 #endif /* CONFIG_SMP */
1788 * Only called when both the current and waking task are -deadline
1791 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
,
1794 if (dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
)) {
1801 * In the unlikely case current and p have the same deadline
1802 * let us try to decide what's the best thing to do...
1804 if ((p
->dl
.deadline
== rq
->curr
->dl
.deadline
) &&
1805 !test_tsk_need_resched(rq
->curr
))
1806 check_preempt_equal_dl(rq
, p
);
1807 #endif /* CONFIG_SMP */
1810 #ifdef CONFIG_SCHED_HRTICK
1811 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1813 hrtick_start(rq
, p
->dl
.runtime
);
1815 #else /* !CONFIG_SCHED_HRTICK */
1816 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1821 static void set_next_task_dl(struct rq
*rq
, struct task_struct
*p
, bool first
)
1823 p
->se
.exec_start
= rq_clock_task(rq
);
1825 /* You can't push away the running task */
1826 dequeue_pushable_dl_task(rq
, p
);
1831 if (hrtick_enabled(rq
))
1832 start_hrtick_dl(rq
, p
);
1834 if (rq
->curr
->sched_class
!= &dl_sched_class
)
1835 update_dl_rq_load_avg(rq_clock_pelt(rq
), rq
, 0);
1837 deadline_queue_push_tasks(rq
);
1840 static struct sched_dl_entity
*pick_next_dl_entity(struct rq
*rq
,
1841 struct dl_rq
*dl_rq
)
1843 struct rb_node
*left
= rb_first_cached(&dl_rq
->root
);
1848 return rb_entry(left
, struct sched_dl_entity
, rb_node
);
1851 static struct task_struct
*pick_next_task_dl(struct rq
*rq
)
1853 struct sched_dl_entity
*dl_se
;
1854 struct dl_rq
*dl_rq
= &rq
->dl
;
1855 struct task_struct
*p
;
1857 if (!sched_dl_runnable(rq
))
1860 dl_se
= pick_next_dl_entity(rq
, dl_rq
);
1862 p
= dl_task_of(dl_se
);
1863 set_next_task_dl(rq
, p
, true);
1867 static void put_prev_task_dl(struct rq
*rq
, struct task_struct
*p
)
1871 update_dl_rq_load_avg(rq_clock_pelt(rq
), rq
, 1);
1872 if (on_dl_rq(&p
->dl
) && p
->nr_cpus_allowed
> 1)
1873 enqueue_pushable_dl_task(rq
, p
);
1877 * scheduler tick hitting a task of our scheduling class.
1879 * NOTE: This function can be called remotely by the tick offload that
1880 * goes along full dynticks. Therefore no local assumption can be made
1881 * and everything must be accessed through the @rq and @curr passed in
1884 static void task_tick_dl(struct rq
*rq
, struct task_struct
*p
, int queued
)
1888 update_dl_rq_load_avg(rq_clock_pelt(rq
), rq
, 1);
1890 * Even when we have runtime, update_curr_dl() might have resulted in us
1891 * not being the leftmost task anymore. In that case NEED_RESCHED will
1892 * be set and schedule() will start a new hrtick for the next task.
1894 if (hrtick_enabled(rq
) && queued
&& p
->dl
.runtime
> 0 &&
1895 is_leftmost(p
, &rq
->dl
))
1896 start_hrtick_dl(rq
, p
);
1899 static void task_fork_dl(struct task_struct
*p
)
1902 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1909 /* Only try algorithms three times */
1910 #define DL_MAX_TRIES 3
1912 static int pick_dl_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1914 if (!task_running(rq
, p
) &&
1915 cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1921 * Return the earliest pushable rq's task, which is suitable to be executed
1922 * on the CPU, NULL otherwise:
1924 static struct task_struct
*pick_earliest_pushable_dl_task(struct rq
*rq
, int cpu
)
1926 struct rb_node
*next_node
= rq
->dl
.pushable_dl_tasks_root
.rb_leftmost
;
1927 struct task_struct
*p
= NULL
;
1929 if (!has_pushable_dl_tasks(rq
))
1934 p
= rb_entry(next_node
, struct task_struct
, pushable_dl_tasks
);
1936 if (pick_dl_task(rq
, p
, cpu
))
1939 next_node
= rb_next(next_node
);
1946 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask_dl
);
1948 static int find_later_rq(struct task_struct
*task
)
1950 struct sched_domain
*sd
;
1951 struct cpumask
*later_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask_dl
);
1952 int this_cpu
= smp_processor_id();
1953 int cpu
= task_cpu(task
);
1955 /* Make sure the mask is initialized first */
1956 if (unlikely(!later_mask
))
1959 if (task
->nr_cpus_allowed
== 1)
1963 * We have to consider system topology and task affinity
1964 * first, then we can look for a suitable CPU.
1966 if (!cpudl_find(&task_rq(task
)->rd
->cpudl
, task
, later_mask
))
1970 * If we are here, some targets have been found, including
1971 * the most suitable which is, among the runqueues where the
1972 * current tasks have later deadlines than the task's one, the
1973 * rq with the latest possible one.
1975 * Now we check how well this matches with task's
1976 * affinity and system topology.
1978 * The last CPU where the task run is our first
1979 * guess, since it is most likely cache-hot there.
1981 if (cpumask_test_cpu(cpu
, later_mask
))
1984 * Check if this_cpu is to be skipped (i.e., it is
1985 * not in the mask) or not.
1987 if (!cpumask_test_cpu(this_cpu
, later_mask
))
1991 for_each_domain(cpu
, sd
) {
1992 if (sd
->flags
& SD_WAKE_AFFINE
) {
1996 * If possible, preempting this_cpu is
1997 * cheaper than migrating.
1999 if (this_cpu
!= -1 &&
2000 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
2005 best_cpu
= cpumask_first_and(later_mask
,
2006 sched_domain_span(sd
));
2008 * Last chance: if a CPU being in both later_mask
2009 * and current sd span is valid, that becomes our
2010 * choice. Of course, the latest possible CPU is
2011 * already under consideration through later_mask.
2013 if (best_cpu
< nr_cpu_ids
) {
2022 * At this point, all our guesses failed, we just return
2023 * 'something', and let the caller sort the things out.
2028 cpu
= cpumask_any(later_mask
);
2029 if (cpu
< nr_cpu_ids
)
2035 /* Locks the rq it finds */
2036 static struct rq
*find_lock_later_rq(struct task_struct
*task
, struct rq
*rq
)
2038 struct rq
*later_rq
= NULL
;
2042 for (tries
= 0; tries
< DL_MAX_TRIES
; tries
++) {
2043 cpu
= find_later_rq(task
);
2045 if ((cpu
== -1) || (cpu
== rq
->cpu
))
2048 later_rq
= cpu_rq(cpu
);
2050 if (later_rq
->dl
.dl_nr_running
&&
2051 !dl_time_before(task
->dl
.deadline
,
2052 later_rq
->dl
.earliest_dl
.curr
)) {
2054 * Target rq has tasks of equal or earlier deadline,
2055 * retrying does not release any lock and is unlikely
2056 * to yield a different result.
2062 /* Retry if something changed. */
2063 if (double_lock_balance(rq
, later_rq
)) {
2064 if (unlikely(task_rq(task
) != rq
||
2065 !cpumask_test_cpu(later_rq
->cpu
, task
->cpus_ptr
) ||
2066 task_running(rq
, task
) ||
2068 !task_on_rq_queued(task
))) {
2069 double_unlock_balance(rq
, later_rq
);
2076 * If the rq we found has no -deadline task, or
2077 * its earliest one has a later deadline than our
2078 * task, the rq is a good one.
2080 if (!later_rq
->dl
.dl_nr_running
||
2081 dl_time_before(task
->dl
.deadline
,
2082 later_rq
->dl
.earliest_dl
.curr
))
2085 /* Otherwise we try again. */
2086 double_unlock_balance(rq
, later_rq
);
2093 static struct task_struct
*pick_next_pushable_dl_task(struct rq
*rq
)
2095 struct task_struct
*p
;
2097 if (!has_pushable_dl_tasks(rq
))
2100 p
= rb_entry(rq
->dl
.pushable_dl_tasks_root
.rb_leftmost
,
2101 struct task_struct
, pushable_dl_tasks
);
2103 BUG_ON(rq
->cpu
!= task_cpu(p
));
2104 BUG_ON(task_current(rq
, p
));
2105 BUG_ON(p
->nr_cpus_allowed
<= 1);
2107 BUG_ON(!task_on_rq_queued(p
));
2108 BUG_ON(!dl_task(p
));
2114 * See if the non running -deadline tasks on this rq
2115 * can be sent to some other CPU where they can preempt
2116 * and start executing.
2118 static int push_dl_task(struct rq
*rq
)
2120 struct task_struct
*next_task
;
2121 struct rq
*later_rq
;
2124 if (!rq
->dl
.overloaded
)
2127 next_task
= pick_next_pushable_dl_task(rq
);
2132 if (WARN_ON(next_task
== rq
->curr
))
2136 * If next_task preempts rq->curr, and rq->curr
2137 * can move away, it makes sense to just reschedule
2138 * without going further in pushing next_task.
2140 if (dl_task(rq
->curr
) &&
2141 dl_time_before(next_task
->dl
.deadline
, rq
->curr
->dl
.deadline
) &&
2142 rq
->curr
->nr_cpus_allowed
> 1) {
2147 /* We might release rq lock */
2148 get_task_struct(next_task
);
2150 /* Will lock the rq it'll find */
2151 later_rq
= find_lock_later_rq(next_task
, rq
);
2153 struct task_struct
*task
;
2156 * We must check all this again, since
2157 * find_lock_later_rq releases rq->lock and it is
2158 * then possible that next_task has migrated.
2160 task
= pick_next_pushable_dl_task(rq
);
2161 if (task
== next_task
) {
2163 * The task is still there. We don't try
2164 * again, some other CPU will pull it when ready.
2173 put_task_struct(next_task
);
2178 deactivate_task(rq
, next_task
, 0);
2179 set_task_cpu(next_task
, later_rq
->cpu
);
2182 * Update the later_rq clock here, because the clock is used
2183 * by the cpufreq_update_util() inside __add_running_bw().
2185 update_rq_clock(later_rq
);
2186 activate_task(later_rq
, next_task
, ENQUEUE_NOCLOCK
);
2189 resched_curr(later_rq
);
2191 double_unlock_balance(rq
, later_rq
);
2194 put_task_struct(next_task
);
2199 static void push_dl_tasks(struct rq
*rq
)
2201 /* push_dl_task() will return true if it moved a -deadline task */
2202 while (push_dl_task(rq
))
2206 static void pull_dl_task(struct rq
*this_rq
)
2208 int this_cpu
= this_rq
->cpu
, cpu
;
2209 struct task_struct
*p
;
2210 bool resched
= false;
2212 u64 dmin
= LONG_MAX
;
2214 if (likely(!dl_overloaded(this_rq
)))
2218 * Match the barrier from dl_set_overloaded; this guarantees that if we
2219 * see overloaded we must also see the dlo_mask bit.
2223 for_each_cpu(cpu
, this_rq
->rd
->dlo_mask
) {
2224 if (this_cpu
== cpu
)
2227 src_rq
= cpu_rq(cpu
);
2230 * It looks racy, abd it is! However, as in sched_rt.c,
2231 * we are fine with this.
2233 if (this_rq
->dl
.dl_nr_running
&&
2234 dl_time_before(this_rq
->dl
.earliest_dl
.curr
,
2235 src_rq
->dl
.earliest_dl
.next
))
2238 /* Might drop this_rq->lock */
2239 double_lock_balance(this_rq
, src_rq
);
2242 * If there are no more pullable tasks on the
2243 * rq, we're done with it.
2245 if (src_rq
->dl
.dl_nr_running
<= 1)
2248 p
= pick_earliest_pushable_dl_task(src_rq
, this_cpu
);
2251 * We found a task to be pulled if:
2252 * - it preempts our current (if there's one),
2253 * - it will preempt the last one we pulled (if any).
2255 if (p
&& dl_time_before(p
->dl
.deadline
, dmin
) &&
2256 (!this_rq
->dl
.dl_nr_running
||
2257 dl_time_before(p
->dl
.deadline
,
2258 this_rq
->dl
.earliest_dl
.curr
))) {
2259 WARN_ON(p
== src_rq
->curr
);
2260 WARN_ON(!task_on_rq_queued(p
));
2263 * Then we pull iff p has actually an earlier
2264 * deadline than the current task of its runqueue.
2266 if (dl_time_before(p
->dl
.deadline
,
2267 src_rq
->curr
->dl
.deadline
))
2272 deactivate_task(src_rq
, p
, 0);
2273 set_task_cpu(p
, this_cpu
);
2274 activate_task(this_rq
, p
, 0);
2275 dmin
= p
->dl
.deadline
;
2277 /* Is there any other task even earlier? */
2280 double_unlock_balance(this_rq
, src_rq
);
2284 resched_curr(this_rq
);
2288 * Since the task is not running and a reschedule is not going to happen
2289 * anytime soon on its runqueue, we try pushing it away now.
2291 static void task_woken_dl(struct rq
*rq
, struct task_struct
*p
)
2293 if (!task_running(rq
, p
) &&
2294 !test_tsk_need_resched(rq
->curr
) &&
2295 p
->nr_cpus_allowed
> 1 &&
2296 dl_task(rq
->curr
) &&
2297 (rq
->curr
->nr_cpus_allowed
< 2 ||
2298 !dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
))) {
2303 static void set_cpus_allowed_dl(struct task_struct
*p
,
2304 const struct cpumask
*new_mask
)
2306 struct root_domain
*src_rd
;
2309 BUG_ON(!dl_task(p
));
2314 * Migrating a SCHED_DEADLINE task between exclusive
2315 * cpusets (different root_domains) entails a bandwidth
2316 * update. We already made space for us in the destination
2317 * domain (see cpuset_can_attach()).
2319 if (!cpumask_intersects(src_rd
->span
, new_mask
)) {
2320 struct dl_bw
*src_dl_b
;
2322 src_dl_b
= dl_bw_of(cpu_of(rq
));
2324 * We now free resources of the root_domain we are migrating
2325 * off. In the worst case, sched_setattr() may temporary fail
2326 * until we complete the update.
2328 raw_spin_lock(&src_dl_b
->lock
);
2329 __dl_sub(src_dl_b
, p
->dl
.dl_bw
, dl_bw_cpus(task_cpu(p
)));
2330 raw_spin_unlock(&src_dl_b
->lock
);
2333 set_cpus_allowed_common(p
, new_mask
);
2336 /* Assumes rq->lock is held */
2337 static void rq_online_dl(struct rq
*rq
)
2339 if (rq
->dl
.overloaded
)
2340 dl_set_overload(rq
);
2342 cpudl_set_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
2343 if (rq
->dl
.dl_nr_running
> 0)
2344 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, rq
->dl
.earliest_dl
.curr
);
2347 /* Assumes rq->lock is held */
2348 static void rq_offline_dl(struct rq
*rq
)
2350 if (rq
->dl
.overloaded
)
2351 dl_clear_overload(rq
);
2353 cpudl_clear(&rq
->rd
->cpudl
, rq
->cpu
);
2354 cpudl_clear_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
2357 void __init
init_sched_dl_class(void)
2361 for_each_possible_cpu(i
)
2362 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl
, i
),
2363 GFP_KERNEL
, cpu_to_node(i
));
2366 void dl_add_task_root_domain(struct task_struct
*p
)
2372 rq
= task_rq_lock(p
, &rf
);
2376 dl_b
= &rq
->rd
->dl_bw
;
2377 raw_spin_lock(&dl_b
->lock
);
2379 __dl_add(dl_b
, p
->dl
.dl_bw
, cpumask_weight(rq
->rd
->span
));
2381 raw_spin_unlock(&dl_b
->lock
);
2384 task_rq_unlock(rq
, p
, &rf
);
2387 void dl_clear_root_domain(struct root_domain
*rd
)
2389 unsigned long flags
;
2391 raw_spin_lock_irqsave(&rd
->dl_bw
.lock
, flags
);
2392 rd
->dl_bw
.total_bw
= 0;
2393 raw_spin_unlock_irqrestore(&rd
->dl_bw
.lock
, flags
);
2396 #endif /* CONFIG_SMP */
2398 static void switched_from_dl(struct rq
*rq
, struct task_struct
*p
)
2401 * task_non_contending() can start the "inactive timer" (if the 0-lag
2402 * time is in the future). If the task switches back to dl before
2403 * the "inactive timer" fires, it can continue to consume its current
2404 * runtime using its current deadline. If it stays outside of
2405 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2406 * will reset the task parameters.
2408 if (task_on_rq_queued(p
) && p
->dl
.dl_runtime
)
2409 task_non_contending(p
);
2411 if (!task_on_rq_queued(p
)) {
2413 * Inactive timer is armed. However, p is leaving DEADLINE and
2414 * might migrate away from this rq while continuing to run on
2415 * some other class. We need to remove its contribution from
2416 * this rq running_bw now, or sub_rq_bw (below) will complain.
2418 if (p
->dl
.dl_non_contending
)
2419 sub_running_bw(&p
->dl
, &rq
->dl
);
2420 sub_rq_bw(&p
->dl
, &rq
->dl
);
2424 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2425 * at the 0-lag time, because the task could have been migrated
2426 * while SCHED_OTHER in the meanwhile.
2428 if (p
->dl
.dl_non_contending
)
2429 p
->dl
.dl_non_contending
= 0;
2432 * Since this might be the only -deadline task on the rq,
2433 * this is the right place to try to pull some other one
2434 * from an overloaded CPU, if any.
2436 if (!task_on_rq_queued(p
) || rq
->dl
.dl_nr_running
)
2439 deadline_queue_pull_task(rq
);
2443 * When switching to -deadline, we may overload the rq, then
2444 * we try to push someone off, if possible.
2446 static void switched_to_dl(struct rq
*rq
, struct task_struct
*p
)
2448 if (hrtimer_try_to_cancel(&p
->dl
.inactive_timer
) == 1)
2451 /* If p is not queued we will update its parameters at next wakeup. */
2452 if (!task_on_rq_queued(p
)) {
2453 add_rq_bw(&p
->dl
, &rq
->dl
);
2458 if (rq
->curr
!= p
) {
2460 if (p
->nr_cpus_allowed
> 1 && rq
->dl
.overloaded
)
2461 deadline_queue_push_tasks(rq
);
2463 if (dl_task(rq
->curr
))
2464 check_preempt_curr_dl(rq
, p
, 0);
2471 * If the scheduling parameters of a -deadline task changed,
2472 * a push or pull operation might be needed.
2474 static void prio_changed_dl(struct rq
*rq
, struct task_struct
*p
,
2477 if (task_on_rq_queued(p
) || rq
->curr
== p
) {
2480 * This might be too much, but unfortunately
2481 * we don't have the old deadline value, and
2482 * we can't argue if the task is increasing
2483 * or lowering its prio, so...
2485 if (!rq
->dl
.overloaded
)
2486 deadline_queue_pull_task(rq
);
2489 * If we now have a earlier deadline task than p,
2490 * then reschedule, provided p is still on this
2493 if (dl_time_before(rq
->dl
.earliest_dl
.curr
, p
->dl
.deadline
))
2497 * Again, we don't know if p has a earlier
2498 * or later deadline, so let's blindly set a
2499 * (maybe not needed) rescheduling point.
2502 #endif /* CONFIG_SMP */
2506 const struct sched_class dl_sched_class
2507 __attribute__((section("__dl_sched_class"))) = {
2508 .enqueue_task
= enqueue_task_dl
,
2509 .dequeue_task
= dequeue_task_dl
,
2510 .yield_task
= yield_task_dl
,
2512 .check_preempt_curr
= check_preempt_curr_dl
,
2514 .pick_next_task
= pick_next_task_dl
,
2515 .put_prev_task
= put_prev_task_dl
,
2516 .set_next_task
= set_next_task_dl
,
2519 .balance
= balance_dl
,
2520 .select_task_rq
= select_task_rq_dl
,
2521 .migrate_task_rq
= migrate_task_rq_dl
,
2522 .set_cpus_allowed
= set_cpus_allowed_dl
,
2523 .rq_online
= rq_online_dl
,
2524 .rq_offline
= rq_offline_dl
,
2525 .task_woken
= task_woken_dl
,
2528 .task_tick
= task_tick_dl
,
2529 .task_fork
= task_fork_dl
,
2531 .prio_changed
= prio_changed_dl
,
2532 .switched_from
= switched_from_dl
,
2533 .switched_to
= switched_to_dl
,
2535 .update_curr
= update_curr_dl
,
2538 int sched_dl_global_validate(void)
2540 u64 runtime
= global_rt_runtime();
2541 u64 period
= global_rt_period();
2542 u64 new_bw
= to_ratio(period
, runtime
);
2545 unsigned long flags
;
2548 * Here we want to check the bandwidth not being set to some
2549 * value smaller than the currently allocated bandwidth in
2550 * any of the root_domains.
2552 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
2553 * cycling on root_domains... Discussion on different/better
2554 * solutions is welcome!
2556 for_each_possible_cpu(cpu
) {
2557 rcu_read_lock_sched();
2558 dl_b
= dl_bw_of(cpu
);
2560 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2561 if (new_bw
< dl_b
->total_bw
)
2563 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2565 rcu_read_unlock_sched();
2574 static void init_dl_rq_bw_ratio(struct dl_rq
*dl_rq
)
2576 if (global_rt_runtime() == RUNTIME_INF
) {
2577 dl_rq
->bw_ratio
= 1 << RATIO_SHIFT
;
2578 dl_rq
->extra_bw
= 1 << BW_SHIFT
;
2580 dl_rq
->bw_ratio
= to_ratio(global_rt_runtime(),
2581 global_rt_period()) >> (BW_SHIFT
- RATIO_SHIFT
);
2582 dl_rq
->extra_bw
= to_ratio(global_rt_period(),
2583 global_rt_runtime());
2587 void sched_dl_do_global(void)
2592 unsigned long flags
;
2594 def_dl_bandwidth
.dl_period
= global_rt_period();
2595 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
2597 if (global_rt_runtime() != RUNTIME_INF
)
2598 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
2601 * FIXME: As above...
2603 for_each_possible_cpu(cpu
) {
2604 rcu_read_lock_sched();
2605 dl_b
= dl_bw_of(cpu
);
2607 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2609 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2611 rcu_read_unlock_sched();
2612 init_dl_rq_bw_ratio(&cpu_rq(cpu
)->dl
);
2617 * We must be sure that accepting a new task (or allowing changing the
2618 * parameters of an existing one) is consistent with the bandwidth
2619 * constraints. If yes, this function also accordingly updates the currently
2620 * allocated bandwidth to reflect the new situation.
2622 * This function is called while holding p's rq->lock.
2624 int sched_dl_overflow(struct task_struct
*p
, int policy
,
2625 const struct sched_attr
*attr
)
2627 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2628 u64 runtime
= attr
->sched_runtime
;
2629 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2630 int cpus
, err
= -1, cpu
= task_cpu(p
);
2631 struct dl_bw
*dl_b
= dl_bw_of(cpu
);
2634 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
2637 /* !deadline task may carry old deadline bandwidth */
2638 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2642 * Either if a task, enters, leave, or stays -deadline but changes
2643 * its parameters, we may need to update accordingly the total
2644 * allocated bandwidth of the container.
2646 raw_spin_lock(&dl_b
->lock
);
2647 cpus
= dl_bw_cpus(cpu
);
2648 cap
= dl_bw_capacity(cpu
);
2650 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2651 !__dl_overflow(dl_b
, cap
, 0, new_bw
)) {
2652 if (hrtimer_active(&p
->dl
.inactive_timer
))
2653 __dl_sub(dl_b
, p
->dl
.dl_bw
, cpus
);
2654 __dl_add(dl_b
, new_bw
, cpus
);
2656 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2657 !__dl_overflow(dl_b
, cap
, p
->dl
.dl_bw
, new_bw
)) {
2659 * XXX this is slightly incorrect: when the task
2660 * utilization decreases, we should delay the total
2661 * utilization change until the task's 0-lag point.
2662 * But this would require to set the task's "inactive
2663 * timer" when the task is not inactive.
2665 __dl_sub(dl_b
, p
->dl
.dl_bw
, cpus
);
2666 __dl_add(dl_b
, new_bw
, cpus
);
2667 dl_change_utilization(p
, new_bw
);
2669 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2671 * Do not decrease the total deadline utilization here,
2672 * switched_from_dl() will take care to do it at the correct
2677 raw_spin_unlock(&dl_b
->lock
);
2683 * This function initializes the sched_dl_entity of a newly becoming
2684 * SCHED_DEADLINE task.
2686 * Only the static values are considered here, the actual runtime and the
2687 * absolute deadline will be properly calculated when the task is enqueued
2688 * for the first time with its new policy.
2690 void __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
2692 struct sched_dl_entity
*dl_se
= &p
->dl
;
2694 dl_se
->dl_runtime
= attr
->sched_runtime
;
2695 dl_se
->dl_deadline
= attr
->sched_deadline
;
2696 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
2697 dl_se
->flags
= attr
->sched_flags
;
2698 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
2699 dl_se
->dl_density
= to_ratio(dl_se
->dl_deadline
, dl_se
->dl_runtime
);
2702 void __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
2704 struct sched_dl_entity
*dl_se
= &p
->dl
;
2706 attr
->sched_priority
= p
->rt_priority
;
2707 attr
->sched_runtime
= dl_se
->dl_runtime
;
2708 attr
->sched_deadline
= dl_se
->dl_deadline
;
2709 attr
->sched_period
= dl_se
->dl_period
;
2710 attr
->sched_flags
= dl_se
->flags
;
2714 * Default limits for DL period; on the top end we guard against small util
2715 * tasks still getting rediculous long effective runtimes, on the bottom end we
2716 * guard against timer DoS.
2718 unsigned int sysctl_sched_dl_period_max
= 1 << 22; /* ~4 seconds */
2719 unsigned int sysctl_sched_dl_period_min
= 100; /* 100 us */
2722 * This function validates the new parameters of a -deadline task.
2723 * We ask for the deadline not being zero, and greater or equal
2724 * than the runtime, as well as the period of being zero or
2725 * greater than deadline. Furthermore, we have to be sure that
2726 * user parameters are above the internal resolution of 1us (we
2727 * check sched_runtime only since it is always the smaller one) and
2728 * below 2^63 ns (we have to check both sched_deadline and
2729 * sched_period, as the latter can be zero).
2731 bool __checkparam_dl(const struct sched_attr
*attr
)
2733 u64 period
, max
, min
;
2735 /* special dl tasks don't actually use any parameter */
2736 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
2740 if (attr
->sched_deadline
== 0)
2744 * Since we truncate DL_SCALE bits, make sure we're at least
2747 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
2751 * Since we use the MSB for wrap-around and sign issues, make
2752 * sure it's not set (mind that period can be equal to zero).
2754 if (attr
->sched_deadline
& (1ULL << 63) ||
2755 attr
->sched_period
& (1ULL << 63))
2758 period
= attr
->sched_period
;
2760 period
= attr
->sched_deadline
;
2762 /* runtime <= deadline <= period (if period != 0) */
2763 if (period
< attr
->sched_deadline
||
2764 attr
->sched_deadline
< attr
->sched_runtime
)
2767 max
= (u64
)READ_ONCE(sysctl_sched_dl_period_max
) * NSEC_PER_USEC
;
2768 min
= (u64
)READ_ONCE(sysctl_sched_dl_period_min
) * NSEC_PER_USEC
;
2770 if (period
< min
|| period
> max
)
2777 * This function clears the sched_dl_entity static params.
2779 void __dl_clear_params(struct task_struct
*p
)
2781 struct sched_dl_entity
*dl_se
= &p
->dl
;
2783 dl_se
->dl_runtime
= 0;
2784 dl_se
->dl_deadline
= 0;
2785 dl_se
->dl_period
= 0;
2788 dl_se
->dl_density
= 0;
2790 dl_se
->dl_boosted
= 0;
2791 dl_se
->dl_throttled
= 0;
2792 dl_se
->dl_yielded
= 0;
2793 dl_se
->dl_non_contending
= 0;
2794 dl_se
->dl_overrun
= 0;
2797 bool dl_param_changed(struct task_struct
*p
, const struct sched_attr
*attr
)
2799 struct sched_dl_entity
*dl_se
= &p
->dl
;
2801 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
2802 dl_se
->dl_deadline
!= attr
->sched_deadline
||
2803 dl_se
->dl_period
!= attr
->sched_period
||
2804 dl_se
->flags
!= attr
->sched_flags
)
2811 int dl_task_can_attach(struct task_struct
*p
, const struct cpumask
*cs_cpus_allowed
)
2813 unsigned long flags
, cap
;
2814 unsigned int dest_cpu
;
2819 dest_cpu
= cpumask_any_and(cpu_active_mask
, cs_cpus_allowed
);
2821 rcu_read_lock_sched();
2822 dl_b
= dl_bw_of(dest_cpu
);
2823 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2824 cap
= dl_bw_capacity(dest_cpu
);
2825 overflow
= __dl_overflow(dl_b
, cap
, 0, p
->dl
.dl_bw
);
2830 * We reserve space for this task in the destination
2831 * root_domain, as we can't fail after this point.
2832 * We will free resources in the source root_domain
2833 * later on (see set_cpus_allowed_dl()).
2835 int cpus
= dl_bw_cpus(dest_cpu
);
2837 __dl_add(dl_b
, p
->dl
.dl_bw
, cpus
);
2840 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2841 rcu_read_unlock_sched();
2846 int dl_cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
2847 const struct cpumask
*trial
)
2849 int ret
= 1, trial_cpus
;
2850 struct dl_bw
*cur_dl_b
;
2851 unsigned long flags
;
2853 rcu_read_lock_sched();
2854 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
2855 trial_cpus
= cpumask_weight(trial
);
2857 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
2858 if (cur_dl_b
->bw
!= -1 &&
2859 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
2861 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
2862 rcu_read_unlock_sched();
2867 bool dl_cpu_busy(unsigned int cpu
)
2869 unsigned long flags
, cap
;
2873 rcu_read_lock_sched();
2874 dl_b
= dl_bw_of(cpu
);
2875 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
2876 cap
= dl_bw_capacity(cpu
);
2877 overflow
= __dl_overflow(dl_b
, cap
, 0, 0);
2878 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
2879 rcu_read_unlock_sched();
2885 #ifdef CONFIG_SCHED_DEBUG
2886 void print_dl_stats(struct seq_file
*m
, int cpu
)
2888 print_dl_rq(m
, cpu
, &cpu_rq(cpu
)->dl
);
2890 #endif /* CONFIG_SCHED_DEBUG */