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>
21 struct dl_bandwidth def_dl_bandwidth
;
23 static inline struct task_struct
*dl_task_of(struct sched_dl_entity
*dl_se
)
25 return container_of(dl_se
, struct task_struct
, dl
);
28 static inline struct rq
*rq_of_dl_rq(struct dl_rq
*dl_rq
)
30 return container_of(dl_rq
, struct rq
, dl
);
33 static inline struct dl_rq
*dl_rq_of_se(struct sched_dl_entity
*dl_se
)
35 struct task_struct
*p
= dl_task_of(dl_se
);
36 struct rq
*rq
= task_rq(p
);
41 static inline int on_dl_rq(struct sched_dl_entity
*dl_se
)
43 return !RB_EMPTY_NODE(&dl_se
->rb_node
);
46 static inline int is_leftmost(struct task_struct
*p
, struct dl_rq
*dl_rq
)
48 struct sched_dl_entity
*dl_se
= &p
->dl
;
50 return dl_rq
->rb_leftmost
== &dl_se
->rb_node
;
53 void init_dl_bandwidth(struct dl_bandwidth
*dl_b
, u64 period
, u64 runtime
)
55 raw_spin_lock_init(&dl_b
->dl_runtime_lock
);
56 dl_b
->dl_period
= period
;
57 dl_b
->dl_runtime
= runtime
;
60 void init_dl_bw(struct dl_bw
*dl_b
)
62 raw_spin_lock_init(&dl_b
->lock
);
63 raw_spin_lock(&def_dl_bandwidth
.dl_runtime_lock
);
64 if (global_rt_runtime() == RUNTIME_INF
)
67 dl_b
->bw
= to_ratio(global_rt_period(), global_rt_runtime());
68 raw_spin_unlock(&def_dl_bandwidth
.dl_runtime_lock
);
72 void init_dl_rq(struct dl_rq
*dl_rq
)
74 dl_rq
->rb_root
= RB_ROOT
;
77 /* zero means no -deadline tasks */
78 dl_rq
->earliest_dl
.curr
= dl_rq
->earliest_dl
.next
= 0;
80 dl_rq
->dl_nr_migratory
= 0;
81 dl_rq
->overloaded
= 0;
82 dl_rq
->pushable_dl_tasks_root
= RB_ROOT
;
84 init_dl_bw(&dl_rq
->dl_bw
);
90 static inline int dl_overloaded(struct rq
*rq
)
92 return atomic_read(&rq
->rd
->dlo_count
);
95 static inline void dl_set_overload(struct rq
*rq
)
100 cpumask_set_cpu(rq
->cpu
, rq
->rd
->dlo_mask
);
102 * Must be visible before the overload count is
103 * set (as in sched_rt.c).
105 * Matched by the barrier in pull_dl_task().
108 atomic_inc(&rq
->rd
->dlo_count
);
111 static inline void dl_clear_overload(struct rq
*rq
)
116 atomic_dec(&rq
->rd
->dlo_count
);
117 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->dlo_mask
);
120 static void update_dl_migration(struct dl_rq
*dl_rq
)
122 if (dl_rq
->dl_nr_migratory
&& dl_rq
->dl_nr_running
> 1) {
123 if (!dl_rq
->overloaded
) {
124 dl_set_overload(rq_of_dl_rq(dl_rq
));
125 dl_rq
->overloaded
= 1;
127 } else if (dl_rq
->overloaded
) {
128 dl_clear_overload(rq_of_dl_rq(dl_rq
));
129 dl_rq
->overloaded
= 0;
133 static void inc_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
135 struct task_struct
*p
= dl_task_of(dl_se
);
137 if (p
->nr_cpus_allowed
> 1)
138 dl_rq
->dl_nr_migratory
++;
140 update_dl_migration(dl_rq
);
143 static void dec_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
145 struct task_struct
*p
= dl_task_of(dl_se
);
147 if (p
->nr_cpus_allowed
> 1)
148 dl_rq
->dl_nr_migratory
--;
150 update_dl_migration(dl_rq
);
154 * The list of pushable -deadline task is not a plist, like in
155 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
157 static void enqueue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
159 struct dl_rq
*dl_rq
= &rq
->dl
;
160 struct rb_node
**link
= &dl_rq
->pushable_dl_tasks_root
.rb_node
;
161 struct rb_node
*parent
= NULL
;
162 struct task_struct
*entry
;
165 BUG_ON(!RB_EMPTY_NODE(&p
->pushable_dl_tasks
));
169 entry
= rb_entry(parent
, struct task_struct
,
171 if (dl_entity_preempt(&p
->dl
, &entry
->dl
))
172 link
= &parent
->rb_left
;
174 link
= &parent
->rb_right
;
180 dl_rq
->pushable_dl_tasks_leftmost
= &p
->pushable_dl_tasks
;
181 dl_rq
->earliest_dl
.next
= p
->dl
.deadline
;
184 rb_link_node(&p
->pushable_dl_tasks
, parent
, link
);
185 rb_insert_color(&p
->pushable_dl_tasks
, &dl_rq
->pushable_dl_tasks_root
);
188 static void dequeue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
190 struct dl_rq
*dl_rq
= &rq
->dl
;
192 if (RB_EMPTY_NODE(&p
->pushable_dl_tasks
))
195 if (dl_rq
->pushable_dl_tasks_leftmost
== &p
->pushable_dl_tasks
) {
196 struct rb_node
*next_node
;
198 next_node
= rb_next(&p
->pushable_dl_tasks
);
199 dl_rq
->pushable_dl_tasks_leftmost
= next_node
;
201 dl_rq
->earliest_dl
.next
= rb_entry(next_node
,
202 struct task_struct
, pushable_dl_tasks
)->dl
.deadline
;
206 rb_erase(&p
->pushable_dl_tasks
, &dl_rq
->pushable_dl_tasks_root
);
207 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
210 static inline int has_pushable_dl_tasks(struct rq
*rq
)
212 return !RB_EMPTY_ROOT(&rq
->dl
.pushable_dl_tasks_root
);
215 static int push_dl_task(struct rq
*rq
);
217 static inline bool need_pull_dl_task(struct rq
*rq
, struct task_struct
*prev
)
219 return dl_task(prev
);
222 static DEFINE_PER_CPU(struct callback_head
, dl_push_head
);
223 static DEFINE_PER_CPU(struct callback_head
, dl_pull_head
);
225 static void push_dl_tasks(struct rq
*);
226 static void pull_dl_task(struct rq
*);
228 static inline void queue_push_tasks(struct rq
*rq
)
230 if (!has_pushable_dl_tasks(rq
))
233 queue_balance_callback(rq
, &per_cpu(dl_push_head
, rq
->cpu
), push_dl_tasks
);
236 static inline void queue_pull_task(struct rq
*rq
)
238 queue_balance_callback(rq
, &per_cpu(dl_pull_head
, rq
->cpu
), pull_dl_task
);
241 static struct rq
*find_lock_later_rq(struct task_struct
*task
, struct rq
*rq
);
243 static struct rq
*dl_task_offline_migration(struct rq
*rq
, struct task_struct
*p
)
245 struct rq
*later_rq
= NULL
;
246 bool fallback
= false;
248 later_rq
= find_lock_later_rq(p
, rq
);
254 * If we cannot preempt any rq, fall back to pick any
258 cpu
= cpumask_any_and(cpu_active_mask
, tsk_cpus_allowed(p
));
259 if (cpu
>= nr_cpu_ids
) {
261 * Fail to find any suitable cpu.
262 * The task will never come back!
264 BUG_ON(dl_bandwidth_enabled());
267 * If admission control is disabled we
268 * try a little harder to let the task
271 cpu
= cpumask_any(cpu_active_mask
);
273 later_rq
= cpu_rq(cpu
);
274 double_lock_balance(rq
, later_rq
);
278 * By now the task is replenished and enqueued; migrate it.
280 deactivate_task(rq
, p
, 0);
281 set_task_cpu(p
, later_rq
->cpu
);
282 activate_task(later_rq
, p
, 0);
285 resched_curr(later_rq
);
287 double_unlock_balance(later_rq
, rq
);
295 void enqueue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
300 void dequeue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
305 void inc_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
310 void dec_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
314 static inline bool need_pull_dl_task(struct rq
*rq
, struct task_struct
*prev
)
319 static inline void pull_dl_task(struct rq
*rq
)
323 static inline void queue_push_tasks(struct rq
*rq
)
327 static inline void queue_pull_task(struct rq
*rq
)
330 #endif /* CONFIG_SMP */
332 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
333 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
334 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
,
338 * We are being explicitly informed that a new instance is starting,
339 * and this means that:
340 * - the absolute deadline of the entity has to be placed at
341 * current time + relative deadline;
342 * - the runtime of the entity has to be set to the maximum value.
344 * The capability of specifying such event is useful whenever a -deadline
345 * entity wants to (try to!) synchronize its behaviour with the scheduler's
346 * one, and to (try to!) reconcile itself with its own scheduling
349 static inline void setup_new_dl_entity(struct sched_dl_entity
*dl_se
,
350 struct sched_dl_entity
*pi_se
)
352 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
353 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
355 WARN_ON(!dl_se
->dl_new
|| dl_se
->dl_throttled
);
358 * We use the regular wall clock time to set deadlines in the
359 * future; in fact, we must consider execution overheads (time
360 * spent on hardirq context, etc.).
362 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
363 dl_se
->runtime
= pi_se
->dl_runtime
;
368 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
369 * possibility of a entity lasting more than what it declared, and thus
370 * exhausting its runtime.
372 * Here we are interested in making runtime overrun possible, but we do
373 * not want a entity which is misbehaving to affect the scheduling of all
375 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
376 * is used, in order to confine each entity within its own bandwidth.
378 * This function deals exactly with that, and ensures that when the runtime
379 * of a entity is replenished, its deadline is also postponed. That ensures
380 * the overrunning entity can't interfere with other entity in the system and
381 * can't make them miss their deadlines. Reasons why this kind of overruns
382 * could happen are, typically, a entity voluntarily trying to overcome its
383 * runtime, or it just underestimated it during sched_setattr().
385 static void replenish_dl_entity(struct sched_dl_entity
*dl_se
,
386 struct sched_dl_entity
*pi_se
)
388 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
389 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
391 BUG_ON(pi_se
->dl_runtime
<= 0);
394 * This could be the case for a !-dl task that is boosted.
395 * Just go with full inherited parameters.
397 if (dl_se
->dl_deadline
== 0) {
398 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
399 dl_se
->runtime
= pi_se
->dl_runtime
;
403 * We keep moving the deadline away until we get some
404 * available runtime for the entity. This ensures correct
405 * handling of situations where the runtime overrun is
408 while (dl_se
->runtime
<= 0) {
409 dl_se
->deadline
+= pi_se
->dl_period
;
410 dl_se
->runtime
+= pi_se
->dl_runtime
;
414 * At this point, the deadline really should be "in
415 * the future" with respect to rq->clock. If it's
416 * not, we are, for some reason, lagging too much!
417 * Anyway, after having warn userspace abut that,
418 * we still try to keep the things running by
419 * resetting the deadline and the budget of the
422 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
))) {
423 printk_deferred_once("sched: DL replenish lagged too much\n");
424 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
425 dl_se
->runtime
= pi_se
->dl_runtime
;
428 if (dl_se
->dl_yielded
)
429 dl_se
->dl_yielded
= 0;
430 if (dl_se
->dl_throttled
)
431 dl_se
->dl_throttled
= 0;
435 * Here we check if --at time t-- an entity (which is probably being
436 * [re]activated or, in general, enqueued) can use its remaining runtime
437 * and its current deadline _without_ exceeding the bandwidth it is
438 * assigned (function returns true if it can't). We are in fact applying
439 * one of the CBS rules: when a task wakes up, if the residual runtime
440 * over residual deadline fits within the allocated bandwidth, then we
441 * can keep the current (absolute) deadline and residual budget without
442 * disrupting the schedulability of the system. Otherwise, we should
443 * refill the runtime and set the deadline a period in the future,
444 * because keeping the current (absolute) deadline of the task would
445 * result in breaking guarantees promised to other tasks (refer to
446 * Documentation/scheduler/sched-deadline.txt for more informations).
448 * This function returns true if:
450 * runtime / (deadline - t) > dl_runtime / dl_period ,
452 * IOW we can't recycle current parameters.
454 * Notice that the bandwidth check is done against the period. For
455 * task with deadline equal to period this is the same of using
456 * dl_deadline instead of dl_period in the equation above.
458 static bool dl_entity_overflow(struct sched_dl_entity
*dl_se
,
459 struct sched_dl_entity
*pi_se
, u64 t
)
464 * left and right are the two sides of the equation above,
465 * after a bit of shuffling to use multiplications instead
468 * Note that none of the time values involved in the two
469 * multiplications are absolute: dl_deadline and dl_runtime
470 * are the relative deadline and the maximum runtime of each
471 * instance, runtime is the runtime left for the last instance
472 * and (deadline - t), since t is rq->clock, is the time left
473 * to the (absolute) deadline. Even if overflowing the u64 type
474 * is very unlikely to occur in both cases, here we scale down
475 * as we want to avoid that risk at all. Scaling down by 10
476 * means that we reduce granularity to 1us. We are fine with it,
477 * since this is only a true/false check and, anyway, thinking
478 * of anything below microseconds resolution is actually fiction
479 * (but still we want to give the user that illusion >;).
481 left
= (pi_se
->dl_period
>> DL_SCALE
) * (dl_se
->runtime
>> DL_SCALE
);
482 right
= ((dl_se
->deadline
- t
) >> DL_SCALE
) *
483 (pi_se
->dl_runtime
>> DL_SCALE
);
485 return dl_time_before(right
, left
);
489 * When a -deadline entity is queued back on the runqueue, its runtime and
490 * deadline might need updating.
492 * The policy here is that we update the deadline of the entity only if:
493 * - the current deadline is in the past,
494 * - using the remaining runtime with the current deadline would make
495 * the entity exceed its bandwidth.
497 static void update_dl_entity(struct sched_dl_entity
*dl_se
,
498 struct sched_dl_entity
*pi_se
)
500 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
501 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
504 * The arrival of a new instance needs special treatment, i.e.,
505 * the actual scheduling parameters have to be "renewed".
508 setup_new_dl_entity(dl_se
, pi_se
);
512 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
)) ||
513 dl_entity_overflow(dl_se
, pi_se
, rq_clock(rq
))) {
514 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
515 dl_se
->runtime
= pi_se
->dl_runtime
;
520 * If the entity depleted all its runtime, and if we want it to sleep
521 * while waiting for some new execution time to become available, we
522 * set the bandwidth enforcement timer to the replenishment instant
523 * and try to activate it.
525 * Notice that it is important for the caller to know if the timer
526 * actually started or not (i.e., the replenishment instant is in
527 * the future or in the past).
529 static int start_dl_timer(struct task_struct
*p
)
531 struct sched_dl_entity
*dl_se
= &p
->dl
;
532 struct hrtimer
*timer
= &dl_se
->dl_timer
;
533 struct rq
*rq
= task_rq(p
);
537 lockdep_assert_held(&rq
->lock
);
540 * We want the timer to fire at the deadline, but considering
541 * that it is actually coming from rq->clock and not from
542 * hrtimer's time base reading.
544 act
= ns_to_ktime(dl_se
->deadline
);
545 now
= hrtimer_cb_get_time(timer
);
546 delta
= ktime_to_ns(now
) - rq_clock(rq
);
547 act
= ktime_add_ns(act
, delta
);
550 * If the expiry time already passed, e.g., because the value
551 * chosen as the deadline is too small, don't even try to
552 * start the timer in the past!
554 if (ktime_us_delta(act
, now
) < 0)
558 * !enqueued will guarantee another callback; even if one is already in
559 * progress. This ensures a balanced {get,put}_task_struct().
561 * The race against __run_timer() clearing the enqueued state is
562 * harmless because we're holding task_rq()->lock, therefore the timer
563 * expiring after we've done the check will wait on its task_rq_lock()
564 * and observe our state.
566 if (!hrtimer_is_queued(timer
)) {
568 hrtimer_start(timer
, act
, HRTIMER_MODE_ABS
);
575 * This is the bandwidth enforcement timer callback. If here, we know
576 * a task is not on its dl_rq, since the fact that the timer was running
577 * means the task is throttled and needs a runtime replenishment.
579 * However, what we actually do depends on the fact the task is active,
580 * (it is on its rq) or has been removed from there by a call to
581 * dequeue_task_dl(). In the former case we must issue the runtime
582 * replenishment and add the task back to the dl_rq; in the latter, we just
583 * do nothing but clearing dl_throttled, so that runtime and deadline
584 * updating (and the queueing back to dl_rq) will be done by the
585 * next call to enqueue_task_dl().
587 static enum hrtimer_restart
dl_task_timer(struct hrtimer
*timer
)
589 struct sched_dl_entity
*dl_se
= container_of(timer
,
590 struct sched_dl_entity
,
592 struct task_struct
*p
= dl_task_of(dl_se
);
596 rq
= task_rq_lock(p
, &flags
);
599 * The task might have changed its scheduling policy to something
600 * different than SCHED_DEADLINE (through switched_fromd_dl()).
603 __dl_clear_params(p
);
608 * This is possible if switched_from_dl() raced against a running
609 * callback that took the above !dl_task() path and we've since then
610 * switched back into SCHED_DEADLINE.
612 * There's nothing to do except drop our task reference.
618 * The task might have been boosted by someone else and might be in the
619 * boosting/deboosting path, its not throttled.
621 if (dl_se
->dl_boosted
)
625 * Spurious timer due to start_dl_timer() race; or we already received
626 * a replenishment from rt_mutex_setprio().
628 if (!dl_se
->dl_throttled
)
635 * If the throttle happened during sched-out; like:
642 * __dequeue_task_dl()
645 * We can be both throttled and !queued. Replenish the counter
646 * but do not enqueue -- wait for our wakeup to do that.
648 if (!task_on_rq_queued(p
)) {
649 replenish_dl_entity(dl_se
, dl_se
);
653 enqueue_task_dl(rq
, p
, ENQUEUE_REPLENISH
);
654 if (dl_task(rq
->curr
))
655 check_preempt_curr_dl(rq
, p
, 0);
661 * Perform balancing operations here; after the replenishments. We
662 * cannot drop rq->lock before this, otherwise the assertion in
663 * start_dl_timer() about not missing updates is not true.
665 * If we find that the rq the task was on is no longer available, we
666 * need to select a new rq.
668 * XXX figure out if select_task_rq_dl() deals with offline cpus.
670 if (unlikely(!rq
->online
))
671 rq
= dl_task_offline_migration(rq
, p
);
674 * Queueing this task back might have overloaded rq, check if we need
675 * to kick someone away.
677 if (has_pushable_dl_tasks(rq
)) {
679 * Nothing relies on rq->lock after this, so its safe to drop
682 lockdep_unpin_lock(&rq
->lock
);
684 lockdep_pin_lock(&rq
->lock
);
689 task_rq_unlock(rq
, p
, &flags
);
692 * This can free the task_struct, including this hrtimer, do not touch
693 * anything related to that after this.
697 return HRTIMER_NORESTART
;
700 void init_dl_task_timer(struct sched_dl_entity
*dl_se
)
702 struct hrtimer
*timer
= &dl_se
->dl_timer
;
704 hrtimer_init(timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
705 timer
->function
= dl_task_timer
;
709 int dl_runtime_exceeded(struct sched_dl_entity
*dl_se
)
711 return (dl_se
->runtime
<= 0);
714 extern bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
);
717 * Update the current task's runtime statistics (provided it is still
718 * a -deadline task and has not been removed from the dl_rq).
720 static void update_curr_dl(struct rq
*rq
)
722 struct task_struct
*curr
= rq
->curr
;
723 struct sched_dl_entity
*dl_se
= &curr
->dl
;
726 if (!dl_task(curr
) || !on_dl_rq(dl_se
))
729 /* Kick cpufreq (see the comment in linux/cpufreq.h). */
730 if (cpu_of(rq
) == smp_processor_id())
731 cpufreq_trigger_update(rq_clock(rq
));
734 * Consumed budget is computed considering the time as
735 * observed by schedulable tasks (excluding time spent
736 * in hardirq context, etc.). Deadlines are instead
737 * computed using hard walltime. This seems to be the more
738 * natural solution, but the full ramifications of this
739 * approach need further study.
741 delta_exec
= rq_clock_task(rq
) - curr
->se
.exec_start
;
742 if (unlikely((s64
)delta_exec
<= 0))
745 schedstat_set(curr
->se
.statistics
.exec_max
,
746 max(curr
->se
.statistics
.exec_max
, delta_exec
));
748 curr
->se
.sum_exec_runtime
+= delta_exec
;
749 account_group_exec_runtime(curr
, delta_exec
);
751 curr
->se
.exec_start
= rq_clock_task(rq
);
752 cpuacct_charge(curr
, delta_exec
);
754 sched_rt_avg_update(rq
, delta_exec
);
756 dl_se
->runtime
-= dl_se
->dl_yielded
? 0 : delta_exec
;
757 if (dl_runtime_exceeded(dl_se
)) {
758 dl_se
->dl_throttled
= 1;
759 __dequeue_task_dl(rq
, curr
, 0);
760 if (unlikely(dl_se
->dl_boosted
|| !start_dl_timer(curr
)))
761 enqueue_task_dl(rq
, curr
, ENQUEUE_REPLENISH
);
763 if (!is_leftmost(curr
, &rq
->dl
))
768 * Because -- for now -- we share the rt bandwidth, we need to
769 * account our runtime there too, otherwise actual rt tasks
770 * would be able to exceed the shared quota.
772 * Account to the root rt group for now.
774 * The solution we're working towards is having the RT groups scheduled
775 * using deadline servers -- however there's a few nasties to figure
776 * out before that can happen.
778 if (rt_bandwidth_enabled()) {
779 struct rt_rq
*rt_rq
= &rq
->rt
;
781 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
783 * We'll let actual RT tasks worry about the overflow here, we
784 * have our own CBS to keep us inline; only account when RT
785 * bandwidth is relevant.
787 if (sched_rt_bandwidth_account(rt_rq
))
788 rt_rq
->rt_time
+= delta_exec
;
789 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
795 static void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
797 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
799 if (dl_rq
->earliest_dl
.curr
== 0 ||
800 dl_time_before(deadline
, dl_rq
->earliest_dl
.curr
)) {
801 dl_rq
->earliest_dl
.curr
= deadline
;
802 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, deadline
, 1);
806 static void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
808 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
811 * Since we may have removed our earliest (and/or next earliest)
812 * task we must recompute them.
814 if (!dl_rq
->dl_nr_running
) {
815 dl_rq
->earliest_dl
.curr
= 0;
816 dl_rq
->earliest_dl
.next
= 0;
817 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, 0, 0);
819 struct rb_node
*leftmost
= dl_rq
->rb_leftmost
;
820 struct sched_dl_entity
*entry
;
822 entry
= rb_entry(leftmost
, struct sched_dl_entity
, rb_node
);
823 dl_rq
->earliest_dl
.curr
= entry
->deadline
;
824 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, entry
->deadline
, 1);
830 static inline void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
831 static inline void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
833 #endif /* CONFIG_SMP */
836 void inc_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
838 int prio
= dl_task_of(dl_se
)->prio
;
839 u64 deadline
= dl_se
->deadline
;
841 WARN_ON(!dl_prio(prio
));
842 dl_rq
->dl_nr_running
++;
843 add_nr_running(rq_of_dl_rq(dl_rq
), 1);
845 inc_dl_deadline(dl_rq
, deadline
);
846 inc_dl_migration(dl_se
, dl_rq
);
850 void dec_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
852 int prio
= dl_task_of(dl_se
)->prio
;
854 WARN_ON(!dl_prio(prio
));
855 WARN_ON(!dl_rq
->dl_nr_running
);
856 dl_rq
->dl_nr_running
--;
857 sub_nr_running(rq_of_dl_rq(dl_rq
), 1);
859 dec_dl_deadline(dl_rq
, dl_se
->deadline
);
860 dec_dl_migration(dl_se
, dl_rq
);
863 static void __enqueue_dl_entity(struct sched_dl_entity
*dl_se
)
865 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
866 struct rb_node
**link
= &dl_rq
->rb_root
.rb_node
;
867 struct rb_node
*parent
= NULL
;
868 struct sched_dl_entity
*entry
;
871 BUG_ON(!RB_EMPTY_NODE(&dl_se
->rb_node
));
875 entry
= rb_entry(parent
, struct sched_dl_entity
, rb_node
);
876 if (dl_time_before(dl_se
->deadline
, entry
->deadline
))
877 link
= &parent
->rb_left
;
879 link
= &parent
->rb_right
;
885 dl_rq
->rb_leftmost
= &dl_se
->rb_node
;
887 rb_link_node(&dl_se
->rb_node
, parent
, link
);
888 rb_insert_color(&dl_se
->rb_node
, &dl_rq
->rb_root
);
890 inc_dl_tasks(dl_se
, dl_rq
);
893 static void __dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
895 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
897 if (RB_EMPTY_NODE(&dl_se
->rb_node
))
900 if (dl_rq
->rb_leftmost
== &dl_se
->rb_node
) {
901 struct rb_node
*next_node
;
903 next_node
= rb_next(&dl_se
->rb_node
);
904 dl_rq
->rb_leftmost
= next_node
;
907 rb_erase(&dl_se
->rb_node
, &dl_rq
->rb_root
);
908 RB_CLEAR_NODE(&dl_se
->rb_node
);
910 dec_dl_tasks(dl_se
, dl_rq
);
914 enqueue_dl_entity(struct sched_dl_entity
*dl_se
,
915 struct sched_dl_entity
*pi_se
, int flags
)
917 BUG_ON(on_dl_rq(dl_se
));
920 * If this is a wakeup or a new instance, the scheduling
921 * parameters of the task might need updating. Otherwise,
922 * we want a replenishment of its runtime.
924 if (dl_se
->dl_new
|| flags
& ENQUEUE_WAKEUP
)
925 update_dl_entity(dl_se
, pi_se
);
926 else if (flags
& ENQUEUE_REPLENISH
)
927 replenish_dl_entity(dl_se
, pi_se
);
929 __enqueue_dl_entity(dl_se
);
932 static void dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
934 __dequeue_dl_entity(dl_se
);
937 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
939 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
940 struct sched_dl_entity
*pi_se
= &p
->dl
;
943 * Use the scheduling parameters of the top pi-waiter
944 * task if we have one and its (absolute) deadline is
945 * smaller than our one... OTW we keep our runtime and
948 if (pi_task
&& p
->dl
.dl_boosted
&& dl_prio(pi_task
->normal_prio
)) {
949 pi_se
= &pi_task
->dl
;
950 } else if (!dl_prio(p
->normal_prio
)) {
952 * Special case in which we have a !SCHED_DEADLINE task
953 * that is going to be deboosted, but exceedes its
954 * runtime while doing so. No point in replenishing
955 * it, as it's going to return back to its original
956 * scheduling class after this.
958 BUG_ON(!p
->dl
.dl_boosted
|| flags
!= ENQUEUE_REPLENISH
);
963 * If p is throttled, we do nothing. In fact, if it exhausted
964 * its budget it needs a replenishment and, since it now is on
965 * its rq, the bandwidth timer callback (which clearly has not
966 * run yet) will take care of this.
968 if (p
->dl
.dl_throttled
&& !(flags
& ENQUEUE_REPLENISH
))
971 enqueue_dl_entity(&p
->dl
, pi_se
, flags
);
973 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
974 enqueue_pushable_dl_task(rq
, p
);
977 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
979 dequeue_dl_entity(&p
->dl
);
980 dequeue_pushable_dl_task(rq
, p
);
983 static void dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
986 __dequeue_task_dl(rq
, p
, flags
);
990 * Yield task semantic for -deadline tasks is:
992 * get off from the CPU until our next instance, with
993 * a new runtime. This is of little use now, since we
994 * don't have a bandwidth reclaiming mechanism. Anyway,
995 * bandwidth reclaiming is planned for the future, and
996 * yield_task_dl will indicate that some spare budget
997 * is available for other task instances to use it.
999 static void yield_task_dl(struct rq
*rq
)
1001 struct task_struct
*p
= rq
->curr
;
1004 * We make the task go to sleep until its current deadline by
1005 * forcing its runtime to zero. This way, update_curr_dl() stops
1006 * it and the bandwidth timer will wake it up and will give it
1007 * new scheduling parameters (thanks to dl_yielded=1).
1009 if (p
->dl
.runtime
> 0) {
1010 rq
->curr
->dl
.dl_yielded
= 1;
1013 update_rq_clock(rq
);
1016 * Tell update_rq_clock() that we've just updated,
1017 * so we don't do microscopic update in schedule()
1018 * and double the fastpath cost.
1020 rq_clock_skip_update(rq
, true);
1025 static int find_later_rq(struct task_struct
*task
);
1028 select_task_rq_dl(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
)
1030 struct task_struct
*curr
;
1033 if (sd_flag
!= SD_BALANCE_WAKE
)
1039 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1042 * If we are dealing with a -deadline task, we must
1043 * decide where to wake it up.
1044 * If it has a later deadline and the current task
1045 * on this rq can't move (provided the waking task
1046 * can!) we prefer to send it somewhere else. On the
1047 * other hand, if it has a shorter deadline, we
1048 * try to make it stay here, it might be important.
1050 if (unlikely(dl_task(curr
)) &&
1051 (curr
->nr_cpus_allowed
< 2 ||
1052 !dl_entity_preempt(&p
->dl
, &curr
->dl
)) &&
1053 (p
->nr_cpus_allowed
> 1)) {
1054 int target
= find_later_rq(p
);
1057 (dl_time_before(p
->dl
.deadline
,
1058 cpu_rq(target
)->dl
.earliest_dl
.curr
) ||
1059 (cpu_rq(target
)->dl
.dl_nr_running
== 0)))
1068 static void check_preempt_equal_dl(struct rq
*rq
, struct task_struct
*p
)
1071 * Current can't be migrated, useless to reschedule,
1072 * let's hope p can move out.
1074 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1075 cpudl_find(&rq
->rd
->cpudl
, rq
->curr
, NULL
) == -1)
1079 * p is migratable, so let's not schedule it and
1080 * see if it is pushed or pulled somewhere else.
1082 if (p
->nr_cpus_allowed
!= 1 &&
1083 cpudl_find(&rq
->rd
->cpudl
, p
, NULL
) != -1)
1089 #endif /* CONFIG_SMP */
1092 * Only called when both the current and waking task are -deadline
1095 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
,
1098 if (dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
)) {
1105 * In the unlikely case current and p have the same deadline
1106 * let us try to decide what's the best thing to do...
1108 if ((p
->dl
.deadline
== rq
->curr
->dl
.deadline
) &&
1109 !test_tsk_need_resched(rq
->curr
))
1110 check_preempt_equal_dl(rq
, p
);
1111 #endif /* CONFIG_SMP */
1114 #ifdef CONFIG_SCHED_HRTICK
1115 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1117 hrtick_start(rq
, p
->dl
.runtime
);
1119 #else /* !CONFIG_SCHED_HRTICK */
1120 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1125 static struct sched_dl_entity
*pick_next_dl_entity(struct rq
*rq
,
1126 struct dl_rq
*dl_rq
)
1128 struct rb_node
*left
= dl_rq
->rb_leftmost
;
1133 return rb_entry(left
, struct sched_dl_entity
, rb_node
);
1136 struct task_struct
*pick_next_task_dl(struct rq
*rq
, struct task_struct
*prev
)
1138 struct sched_dl_entity
*dl_se
;
1139 struct task_struct
*p
;
1140 struct dl_rq
*dl_rq
;
1144 if (need_pull_dl_task(rq
, prev
)) {
1146 * This is OK, because current is on_cpu, which avoids it being
1147 * picked for load-balance and preemption/IRQs are still
1148 * disabled avoiding further scheduler activity on it and we're
1149 * being very careful to re-start the picking loop.
1151 lockdep_unpin_lock(&rq
->lock
);
1153 lockdep_pin_lock(&rq
->lock
);
1155 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1156 * means a stop task can slip in, in which case we need to
1157 * re-start task selection.
1159 if (rq
->stop
&& task_on_rq_queued(rq
->stop
))
1164 * When prev is DL, we may throttle it in put_prev_task().
1165 * So, we update time before we check for dl_nr_running.
1167 if (prev
->sched_class
== &dl_sched_class
)
1170 if (unlikely(!dl_rq
->dl_nr_running
))
1173 put_prev_task(rq
, prev
);
1175 dl_se
= pick_next_dl_entity(rq
, dl_rq
);
1178 p
= dl_task_of(dl_se
);
1179 p
->se
.exec_start
= rq_clock_task(rq
);
1181 /* Running task will never be pushed. */
1182 dequeue_pushable_dl_task(rq
, p
);
1184 if (hrtick_enabled(rq
))
1185 start_hrtick_dl(rq
, p
);
1187 queue_push_tasks(rq
);
1192 static void put_prev_task_dl(struct rq
*rq
, struct task_struct
*p
)
1196 if (on_dl_rq(&p
->dl
) && p
->nr_cpus_allowed
> 1)
1197 enqueue_pushable_dl_task(rq
, p
);
1200 static void task_tick_dl(struct rq
*rq
, struct task_struct
*p
, int queued
)
1205 * Even when we have runtime, update_curr_dl() might have resulted in us
1206 * not being the leftmost task anymore. In that case NEED_RESCHED will
1207 * be set and schedule() will start a new hrtick for the next task.
1209 if (hrtick_enabled(rq
) && queued
&& p
->dl
.runtime
> 0 &&
1210 is_leftmost(p
, &rq
->dl
))
1211 start_hrtick_dl(rq
, p
);
1214 static void task_fork_dl(struct task_struct
*p
)
1217 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1222 static void task_dead_dl(struct task_struct
*p
)
1224 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
1227 * Since we are TASK_DEAD we won't slip out of the domain!
1229 raw_spin_lock_irq(&dl_b
->lock
);
1230 /* XXX we should retain the bw until 0-lag */
1231 dl_b
->total_bw
-= p
->dl
.dl_bw
;
1232 raw_spin_unlock_irq(&dl_b
->lock
);
1235 static void set_curr_task_dl(struct rq
*rq
)
1237 struct task_struct
*p
= rq
->curr
;
1239 p
->se
.exec_start
= rq_clock_task(rq
);
1241 /* You can't push away the running task */
1242 dequeue_pushable_dl_task(rq
, p
);
1247 /* Only try algorithms three times */
1248 #define DL_MAX_TRIES 3
1250 static int pick_dl_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1252 if (!task_running(rq
, p
) &&
1253 cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
1259 * Return the earliest pushable rq's task, which is suitable to be executed
1260 * on the CPU, NULL otherwise:
1262 static struct task_struct
*pick_earliest_pushable_dl_task(struct rq
*rq
, int cpu
)
1264 struct rb_node
*next_node
= rq
->dl
.pushable_dl_tasks_leftmost
;
1265 struct task_struct
*p
= NULL
;
1267 if (!has_pushable_dl_tasks(rq
))
1272 p
= rb_entry(next_node
, struct task_struct
, pushable_dl_tasks
);
1274 if (pick_dl_task(rq
, p
, cpu
))
1277 next_node
= rb_next(next_node
);
1284 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask_dl
);
1286 static int find_later_rq(struct task_struct
*task
)
1288 struct sched_domain
*sd
;
1289 struct cpumask
*later_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask_dl
);
1290 int this_cpu
= smp_processor_id();
1291 int best_cpu
, cpu
= task_cpu(task
);
1293 /* Make sure the mask is initialized first */
1294 if (unlikely(!later_mask
))
1297 if (task
->nr_cpus_allowed
== 1)
1301 * We have to consider system topology and task affinity
1302 * first, then we can look for a suitable cpu.
1304 best_cpu
= cpudl_find(&task_rq(task
)->rd
->cpudl
,
1310 * If we are here, some target has been found,
1311 * the most suitable of which is cached in best_cpu.
1312 * This is, among the runqueues where the current tasks
1313 * have later deadlines than the task's one, the rq
1314 * with the latest possible one.
1316 * Now we check how well this matches with task's
1317 * affinity and system topology.
1319 * The last cpu where the task run is our first
1320 * guess, since it is most likely cache-hot there.
1322 if (cpumask_test_cpu(cpu
, later_mask
))
1325 * Check if this_cpu is to be skipped (i.e., it is
1326 * not in the mask) or not.
1328 if (!cpumask_test_cpu(this_cpu
, later_mask
))
1332 for_each_domain(cpu
, sd
) {
1333 if (sd
->flags
& SD_WAKE_AFFINE
) {
1336 * If possible, preempting this_cpu is
1337 * cheaper than migrating.
1339 if (this_cpu
!= -1 &&
1340 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1346 * Last chance: if best_cpu is valid and is
1347 * in the mask, that becomes our choice.
1349 if (best_cpu
< nr_cpu_ids
&&
1350 cpumask_test_cpu(best_cpu
, sched_domain_span(sd
))) {
1359 * At this point, all our guesses failed, we just return
1360 * 'something', and let the caller sort the things out.
1365 cpu
= cpumask_any(later_mask
);
1366 if (cpu
< nr_cpu_ids
)
1372 /* Locks the rq it finds */
1373 static struct rq
*find_lock_later_rq(struct task_struct
*task
, struct rq
*rq
)
1375 struct rq
*later_rq
= NULL
;
1379 for (tries
= 0; tries
< DL_MAX_TRIES
; tries
++) {
1380 cpu
= find_later_rq(task
);
1382 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1385 later_rq
= cpu_rq(cpu
);
1387 if (later_rq
->dl
.dl_nr_running
&&
1388 !dl_time_before(task
->dl
.deadline
,
1389 later_rq
->dl
.earliest_dl
.curr
)) {
1391 * Target rq has tasks of equal or earlier deadline,
1392 * retrying does not release any lock and is unlikely
1393 * to yield a different result.
1399 /* Retry if something changed. */
1400 if (double_lock_balance(rq
, later_rq
)) {
1401 if (unlikely(task_rq(task
) != rq
||
1402 !cpumask_test_cpu(later_rq
->cpu
,
1403 &task
->cpus_allowed
) ||
1404 task_running(rq
, task
) ||
1405 !task_on_rq_queued(task
))) {
1406 double_unlock_balance(rq
, later_rq
);
1413 * If the rq we found has no -deadline task, or
1414 * its earliest one has a later deadline than our
1415 * task, the rq is a good one.
1417 if (!later_rq
->dl
.dl_nr_running
||
1418 dl_time_before(task
->dl
.deadline
,
1419 later_rq
->dl
.earliest_dl
.curr
))
1422 /* Otherwise we try again. */
1423 double_unlock_balance(rq
, later_rq
);
1430 static struct task_struct
*pick_next_pushable_dl_task(struct rq
*rq
)
1432 struct task_struct
*p
;
1434 if (!has_pushable_dl_tasks(rq
))
1437 p
= rb_entry(rq
->dl
.pushable_dl_tasks_leftmost
,
1438 struct task_struct
, pushable_dl_tasks
);
1440 BUG_ON(rq
->cpu
!= task_cpu(p
));
1441 BUG_ON(task_current(rq
, p
));
1442 BUG_ON(p
->nr_cpus_allowed
<= 1);
1444 BUG_ON(!task_on_rq_queued(p
));
1445 BUG_ON(!dl_task(p
));
1451 * See if the non running -deadline tasks on this rq
1452 * can be sent to some other CPU where they can preempt
1453 * and start executing.
1455 static int push_dl_task(struct rq
*rq
)
1457 struct task_struct
*next_task
;
1458 struct rq
*later_rq
;
1461 if (!rq
->dl
.overloaded
)
1464 next_task
= pick_next_pushable_dl_task(rq
);
1469 if (unlikely(next_task
== rq
->curr
)) {
1475 * If next_task preempts rq->curr, and rq->curr
1476 * can move away, it makes sense to just reschedule
1477 * without going further in pushing next_task.
1479 if (dl_task(rq
->curr
) &&
1480 dl_time_before(next_task
->dl
.deadline
, rq
->curr
->dl
.deadline
) &&
1481 rq
->curr
->nr_cpus_allowed
> 1) {
1486 /* We might release rq lock */
1487 get_task_struct(next_task
);
1489 /* Will lock the rq it'll find */
1490 later_rq
= find_lock_later_rq(next_task
, rq
);
1492 struct task_struct
*task
;
1495 * We must check all this again, since
1496 * find_lock_later_rq releases rq->lock and it is
1497 * then possible that next_task has migrated.
1499 task
= pick_next_pushable_dl_task(rq
);
1500 if (task_cpu(next_task
) == rq
->cpu
&& task
== next_task
) {
1502 * The task is still there. We don't try
1503 * again, some other cpu will pull it when ready.
1512 put_task_struct(next_task
);
1517 deactivate_task(rq
, next_task
, 0);
1518 set_task_cpu(next_task
, later_rq
->cpu
);
1519 activate_task(later_rq
, next_task
, 0);
1522 resched_curr(later_rq
);
1524 double_unlock_balance(rq
, later_rq
);
1527 put_task_struct(next_task
);
1532 static void push_dl_tasks(struct rq
*rq
)
1534 /* push_dl_task() will return true if it moved a -deadline task */
1535 while (push_dl_task(rq
))
1539 static void pull_dl_task(struct rq
*this_rq
)
1541 int this_cpu
= this_rq
->cpu
, cpu
;
1542 struct task_struct
*p
;
1543 bool resched
= false;
1545 u64 dmin
= LONG_MAX
;
1547 if (likely(!dl_overloaded(this_rq
)))
1551 * Match the barrier from dl_set_overloaded; this guarantees that if we
1552 * see overloaded we must also see the dlo_mask bit.
1556 for_each_cpu(cpu
, this_rq
->rd
->dlo_mask
) {
1557 if (this_cpu
== cpu
)
1560 src_rq
= cpu_rq(cpu
);
1563 * It looks racy, abd it is! However, as in sched_rt.c,
1564 * we are fine with this.
1566 if (this_rq
->dl
.dl_nr_running
&&
1567 dl_time_before(this_rq
->dl
.earliest_dl
.curr
,
1568 src_rq
->dl
.earliest_dl
.next
))
1571 /* Might drop this_rq->lock */
1572 double_lock_balance(this_rq
, src_rq
);
1575 * If there are no more pullable tasks on the
1576 * rq, we're done with it.
1578 if (src_rq
->dl
.dl_nr_running
<= 1)
1581 p
= pick_earliest_pushable_dl_task(src_rq
, this_cpu
);
1584 * We found a task to be pulled if:
1585 * - it preempts our current (if there's one),
1586 * - it will preempt the last one we pulled (if any).
1588 if (p
&& dl_time_before(p
->dl
.deadline
, dmin
) &&
1589 (!this_rq
->dl
.dl_nr_running
||
1590 dl_time_before(p
->dl
.deadline
,
1591 this_rq
->dl
.earliest_dl
.curr
))) {
1592 WARN_ON(p
== src_rq
->curr
);
1593 WARN_ON(!task_on_rq_queued(p
));
1596 * Then we pull iff p has actually an earlier
1597 * deadline than the current task of its runqueue.
1599 if (dl_time_before(p
->dl
.deadline
,
1600 src_rq
->curr
->dl
.deadline
))
1605 deactivate_task(src_rq
, p
, 0);
1606 set_task_cpu(p
, this_cpu
);
1607 activate_task(this_rq
, p
, 0);
1608 dmin
= p
->dl
.deadline
;
1610 /* Is there any other task even earlier? */
1613 double_unlock_balance(this_rq
, src_rq
);
1617 resched_curr(this_rq
);
1621 * Since the task is not running and a reschedule is not going to happen
1622 * anytime soon on its runqueue, we try pushing it away now.
1624 static void task_woken_dl(struct rq
*rq
, struct task_struct
*p
)
1626 if (!task_running(rq
, p
) &&
1627 !test_tsk_need_resched(rq
->curr
) &&
1628 p
->nr_cpus_allowed
> 1 &&
1629 dl_task(rq
->curr
) &&
1630 (rq
->curr
->nr_cpus_allowed
< 2 ||
1631 !dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
))) {
1636 static void set_cpus_allowed_dl(struct task_struct
*p
,
1637 const struct cpumask
*new_mask
)
1639 struct root_domain
*src_rd
;
1642 BUG_ON(!dl_task(p
));
1647 * Migrating a SCHED_DEADLINE task between exclusive
1648 * cpusets (different root_domains) entails a bandwidth
1649 * update. We already made space for us in the destination
1650 * domain (see cpuset_can_attach()).
1652 if (!cpumask_intersects(src_rd
->span
, new_mask
)) {
1653 struct dl_bw
*src_dl_b
;
1655 src_dl_b
= dl_bw_of(cpu_of(rq
));
1657 * We now free resources of the root_domain we are migrating
1658 * off. In the worst case, sched_setattr() may temporary fail
1659 * until we complete the update.
1661 raw_spin_lock(&src_dl_b
->lock
);
1662 __dl_clear(src_dl_b
, p
->dl
.dl_bw
);
1663 raw_spin_unlock(&src_dl_b
->lock
);
1666 set_cpus_allowed_common(p
, new_mask
);
1669 /* Assumes rq->lock is held */
1670 static void rq_online_dl(struct rq
*rq
)
1672 if (rq
->dl
.overloaded
)
1673 dl_set_overload(rq
);
1675 cpudl_set_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
1676 if (rq
->dl
.dl_nr_running
> 0)
1677 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, rq
->dl
.earliest_dl
.curr
, 1);
1680 /* Assumes rq->lock is held */
1681 static void rq_offline_dl(struct rq
*rq
)
1683 if (rq
->dl
.overloaded
)
1684 dl_clear_overload(rq
);
1686 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, 0, 0);
1687 cpudl_clear_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
1690 void __init
init_sched_dl_class(void)
1694 for_each_possible_cpu(i
)
1695 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl
, i
),
1696 GFP_KERNEL
, cpu_to_node(i
));
1699 #endif /* CONFIG_SMP */
1701 static void switched_from_dl(struct rq
*rq
, struct task_struct
*p
)
1704 * Start the deadline timer; if we switch back to dl before this we'll
1705 * continue consuming our current CBS slice. If we stay outside of
1706 * SCHED_DEADLINE until the deadline passes, the timer will reset the
1709 if (!start_dl_timer(p
))
1710 __dl_clear_params(p
);
1713 * Since this might be the only -deadline task on the rq,
1714 * this is the right place to try to pull some other one
1715 * from an overloaded cpu, if any.
1717 if (!task_on_rq_queued(p
) || rq
->dl
.dl_nr_running
)
1720 queue_pull_task(rq
);
1724 * When switching to -deadline, we may overload the rq, then
1725 * we try to push someone off, if possible.
1727 static void switched_to_dl(struct rq
*rq
, struct task_struct
*p
)
1729 if (task_on_rq_queued(p
) && rq
->curr
!= p
) {
1731 if (p
->nr_cpus_allowed
> 1 && rq
->dl
.overloaded
)
1732 queue_push_tasks(rq
);
1734 if (dl_task(rq
->curr
))
1735 check_preempt_curr_dl(rq
, p
, 0);
1743 * If the scheduling parameters of a -deadline task changed,
1744 * a push or pull operation might be needed.
1746 static void prio_changed_dl(struct rq
*rq
, struct task_struct
*p
,
1749 if (task_on_rq_queued(p
) || rq
->curr
== p
) {
1752 * This might be too much, but unfortunately
1753 * we don't have the old deadline value, and
1754 * we can't argue if the task is increasing
1755 * or lowering its prio, so...
1757 if (!rq
->dl
.overloaded
)
1758 queue_pull_task(rq
);
1761 * If we now have a earlier deadline task than p,
1762 * then reschedule, provided p is still on this
1765 if (dl_time_before(rq
->dl
.earliest_dl
.curr
, p
->dl
.deadline
))
1769 * Again, we don't know if p has a earlier
1770 * or later deadline, so let's blindly set a
1771 * (maybe not needed) rescheduling point.
1774 #endif /* CONFIG_SMP */
1776 switched_to_dl(rq
, p
);
1779 const struct sched_class dl_sched_class
= {
1780 .next
= &rt_sched_class
,
1781 .enqueue_task
= enqueue_task_dl
,
1782 .dequeue_task
= dequeue_task_dl
,
1783 .yield_task
= yield_task_dl
,
1785 .check_preempt_curr
= check_preempt_curr_dl
,
1787 .pick_next_task
= pick_next_task_dl
,
1788 .put_prev_task
= put_prev_task_dl
,
1791 .select_task_rq
= select_task_rq_dl
,
1792 .set_cpus_allowed
= set_cpus_allowed_dl
,
1793 .rq_online
= rq_online_dl
,
1794 .rq_offline
= rq_offline_dl
,
1795 .task_woken
= task_woken_dl
,
1798 .set_curr_task
= set_curr_task_dl
,
1799 .task_tick
= task_tick_dl
,
1800 .task_fork
= task_fork_dl
,
1801 .task_dead
= task_dead_dl
,
1803 .prio_changed
= prio_changed_dl
,
1804 .switched_from
= switched_from_dl
,
1805 .switched_to
= switched_to_dl
,
1807 .update_curr
= update_curr_dl
,
1810 #ifdef CONFIG_SCHED_DEBUG
1811 extern void print_dl_rq(struct seq_file
*m
, int cpu
, struct dl_rq
*dl_rq
);
1813 void print_dl_stats(struct seq_file
*m
, int cpu
)
1815 print_dl_rq(m
, cpu
, &cpu_rq(cpu
)->dl
);
1817 #endif /* CONFIG_SCHED_DEBUG */