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 (tsk_nr_cpus_allowed(p
) > 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 (tsk_nr_cpus_allowed(p
) > 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
;
247 later_rq
= find_lock_later_rq(p
, rq
);
252 * If we cannot preempt any rq, fall back to pick any
255 cpu
= cpumask_any_and(cpu_active_mask
, tsk_cpus_allowed(p
));
256 if (cpu
>= nr_cpu_ids
) {
258 * Fail to find any suitable cpu.
259 * The task will never come back!
261 BUG_ON(dl_bandwidth_enabled());
264 * If admission control is disabled we
265 * try a little harder to let the task
268 cpu
= cpumask_any(cpu_active_mask
);
270 later_rq
= cpu_rq(cpu
);
271 double_lock_balance(rq
, later_rq
);
274 set_task_cpu(p
, later_rq
->cpu
);
275 double_unlock_balance(later_rq
, rq
);
283 void enqueue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
288 void dequeue_pushable_dl_task(struct rq
*rq
, struct task_struct
*p
)
293 void inc_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
298 void dec_dl_migration(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
302 static inline bool need_pull_dl_task(struct rq
*rq
, struct task_struct
*prev
)
307 static inline void pull_dl_task(struct rq
*rq
)
311 static inline void queue_push_tasks(struct rq
*rq
)
315 static inline void queue_pull_task(struct rq
*rq
)
318 #endif /* CONFIG_SMP */
320 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
321 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
);
322 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
,
326 * We are being explicitly informed that a new instance is starting,
327 * and this means that:
328 * - the absolute deadline of the entity has to be placed at
329 * current time + relative deadline;
330 * - the runtime of the entity has to be set to the maximum value.
332 * The capability of specifying such event is useful whenever a -deadline
333 * entity wants to (try to!) synchronize its behaviour with the scheduler's
334 * one, and to (try to!) reconcile itself with its own scheduling
337 static inline void setup_new_dl_entity(struct sched_dl_entity
*dl_se
)
339 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
340 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
342 WARN_ON(dl_se
->dl_boosted
);
343 WARN_ON(dl_time_before(rq_clock(rq
), dl_se
->deadline
));
346 * We are racing with the deadline timer. So, do nothing because
347 * the deadline timer handler will take care of properly recharging
348 * the runtime and postponing the deadline
350 if (dl_se
->dl_throttled
)
354 * We use the regular wall clock time to set deadlines in the
355 * future; in fact, we must consider execution overheads (time
356 * spent on hardirq context, etc.).
358 dl_se
->deadline
= rq_clock(rq
) + dl_se
->dl_deadline
;
359 dl_se
->runtime
= dl_se
->dl_runtime
;
363 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
364 * possibility of a entity lasting more than what it declared, and thus
365 * exhausting its runtime.
367 * Here we are interested in making runtime overrun possible, but we do
368 * not want a entity which is misbehaving to affect the scheduling of all
370 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
371 * is used, in order to confine each entity within its own bandwidth.
373 * This function deals exactly with that, and ensures that when the runtime
374 * of a entity is replenished, its deadline is also postponed. That ensures
375 * the overrunning entity can't interfere with other entity in the system and
376 * can't make them miss their deadlines. Reasons why this kind of overruns
377 * could happen are, typically, a entity voluntarily trying to overcome its
378 * runtime, or it just underestimated it during sched_setattr().
380 static void replenish_dl_entity(struct sched_dl_entity
*dl_se
,
381 struct sched_dl_entity
*pi_se
)
383 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
384 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
386 BUG_ON(pi_se
->dl_runtime
<= 0);
389 * This could be the case for a !-dl task that is boosted.
390 * Just go with full inherited parameters.
392 if (dl_se
->dl_deadline
== 0) {
393 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
394 dl_se
->runtime
= pi_se
->dl_runtime
;
397 if (dl_se
->dl_yielded
&& dl_se
->runtime
> 0)
401 * We keep moving the deadline away until we get some
402 * available runtime for the entity. This ensures correct
403 * handling of situations where the runtime overrun is
406 while (dl_se
->runtime
<= 0) {
407 dl_se
->deadline
+= pi_se
->dl_period
;
408 dl_se
->runtime
+= pi_se
->dl_runtime
;
412 * At this point, the deadline really should be "in
413 * the future" with respect to rq->clock. If it's
414 * not, we are, for some reason, lagging too much!
415 * Anyway, after having warn userspace abut that,
416 * we still try to keep the things running by
417 * resetting the deadline and the budget of the
420 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
))) {
421 printk_deferred_once("sched: DL replenish lagged too much\n");
422 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
423 dl_se
->runtime
= pi_se
->dl_runtime
;
426 if (dl_se
->dl_yielded
)
427 dl_se
->dl_yielded
= 0;
428 if (dl_se
->dl_throttled
)
429 dl_se
->dl_throttled
= 0;
433 * Here we check if --at time t-- an entity (which is probably being
434 * [re]activated or, in general, enqueued) can use its remaining runtime
435 * and its current deadline _without_ exceeding the bandwidth it is
436 * assigned (function returns true if it can't). We are in fact applying
437 * one of the CBS rules: when a task wakes up, if the residual runtime
438 * over residual deadline fits within the allocated bandwidth, then we
439 * can keep the current (absolute) deadline and residual budget without
440 * disrupting the schedulability of the system. Otherwise, we should
441 * refill the runtime and set the deadline a period in the future,
442 * because keeping the current (absolute) deadline of the task would
443 * result in breaking guarantees promised to other tasks (refer to
444 * Documentation/scheduler/sched-deadline.txt for more informations).
446 * This function returns true if:
448 * runtime / (deadline - t) > dl_runtime / dl_period ,
450 * IOW we can't recycle current parameters.
452 * Notice that the bandwidth check is done against the period. For
453 * task with deadline equal to period this is the same of using
454 * dl_deadline instead of dl_period in the equation above.
456 static bool dl_entity_overflow(struct sched_dl_entity
*dl_se
,
457 struct sched_dl_entity
*pi_se
, u64 t
)
462 * left and right are the two sides of the equation above,
463 * after a bit of shuffling to use multiplications instead
466 * Note that none of the time values involved in the two
467 * multiplications are absolute: dl_deadline and dl_runtime
468 * are the relative deadline and the maximum runtime of each
469 * instance, runtime is the runtime left for the last instance
470 * and (deadline - t), since t is rq->clock, is the time left
471 * to the (absolute) deadline. Even if overflowing the u64 type
472 * is very unlikely to occur in both cases, here we scale down
473 * as we want to avoid that risk at all. Scaling down by 10
474 * means that we reduce granularity to 1us. We are fine with it,
475 * since this is only a true/false check and, anyway, thinking
476 * of anything below microseconds resolution is actually fiction
477 * (but still we want to give the user that illusion >;).
479 left
= (pi_se
->dl_period
>> DL_SCALE
) * (dl_se
->runtime
>> DL_SCALE
);
480 right
= ((dl_se
->deadline
- t
) >> DL_SCALE
) *
481 (pi_se
->dl_runtime
>> DL_SCALE
);
483 return dl_time_before(right
, left
);
487 * When a -deadline entity is queued back on the runqueue, its runtime and
488 * deadline might need updating.
490 * The policy here is that we update the deadline of the entity only if:
491 * - the current deadline is in the past,
492 * - using the remaining runtime with the current deadline would make
493 * the entity exceed its bandwidth.
495 static void update_dl_entity(struct sched_dl_entity
*dl_se
,
496 struct sched_dl_entity
*pi_se
)
498 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
499 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
501 if (dl_time_before(dl_se
->deadline
, rq_clock(rq
)) ||
502 dl_entity_overflow(dl_se
, pi_se
, rq_clock(rq
))) {
503 dl_se
->deadline
= rq_clock(rq
) + pi_se
->dl_deadline
;
504 dl_se
->runtime
= pi_se
->dl_runtime
;
509 * If the entity depleted all its runtime, and if we want it to sleep
510 * while waiting for some new execution time to become available, we
511 * set the bandwidth enforcement timer to the replenishment instant
512 * and try to activate it.
514 * Notice that it is important for the caller to know if the timer
515 * actually started or not (i.e., the replenishment instant is in
516 * the future or in the past).
518 static int start_dl_timer(struct task_struct
*p
)
520 struct sched_dl_entity
*dl_se
= &p
->dl
;
521 struct hrtimer
*timer
= &dl_se
->dl_timer
;
522 struct rq
*rq
= task_rq(p
);
526 lockdep_assert_held(&rq
->lock
);
529 * We want the timer to fire at the deadline, but considering
530 * that it is actually coming from rq->clock and not from
531 * hrtimer's time base reading.
533 act
= ns_to_ktime(dl_se
->deadline
);
534 now
= hrtimer_cb_get_time(timer
);
535 delta
= ktime_to_ns(now
) - rq_clock(rq
);
536 act
= ktime_add_ns(act
, delta
);
539 * If the expiry time already passed, e.g., because the value
540 * chosen as the deadline is too small, don't even try to
541 * start the timer in the past!
543 if (ktime_us_delta(act
, now
) < 0)
547 * !enqueued will guarantee another callback; even if one is already in
548 * progress. This ensures a balanced {get,put}_task_struct().
550 * The race against __run_timer() clearing the enqueued state is
551 * harmless because we're holding task_rq()->lock, therefore the timer
552 * expiring after we've done the check will wait on its task_rq_lock()
553 * and observe our state.
555 if (!hrtimer_is_queued(timer
)) {
557 hrtimer_start(timer
, act
, HRTIMER_MODE_ABS
);
564 * This is the bandwidth enforcement timer callback. If here, we know
565 * a task is not on its dl_rq, since the fact that the timer was running
566 * means the task is throttled and needs a runtime replenishment.
568 * However, what we actually do depends on the fact the task is active,
569 * (it is on its rq) or has been removed from there by a call to
570 * dequeue_task_dl(). In the former case we must issue the runtime
571 * replenishment and add the task back to the dl_rq; in the latter, we just
572 * do nothing but clearing dl_throttled, so that runtime and deadline
573 * updating (and the queueing back to dl_rq) will be done by the
574 * next call to enqueue_task_dl().
576 static enum hrtimer_restart
dl_task_timer(struct hrtimer
*timer
)
578 struct sched_dl_entity
*dl_se
= container_of(timer
,
579 struct sched_dl_entity
,
581 struct task_struct
*p
= dl_task_of(dl_se
);
585 rq
= task_rq_lock(p
, &rf
);
588 * The task might have changed its scheduling policy to something
589 * different than SCHED_DEADLINE (through switched_fromd_dl()).
592 __dl_clear_params(p
);
597 * The task might have been boosted by someone else and might be in the
598 * boosting/deboosting path, its not throttled.
600 if (dl_se
->dl_boosted
)
604 * Spurious timer due to start_dl_timer() race; or we already received
605 * a replenishment from rt_mutex_setprio().
607 if (!dl_se
->dl_throttled
)
614 * If the throttle happened during sched-out; like:
621 * __dequeue_task_dl()
624 * We can be both throttled and !queued. Replenish the counter
625 * but do not enqueue -- wait for our wakeup to do that.
627 if (!task_on_rq_queued(p
)) {
628 replenish_dl_entity(dl_se
, dl_se
);
633 if (unlikely(!rq
->online
)) {
635 * If the runqueue is no longer available, migrate the
636 * task elsewhere. This necessarily changes rq.
638 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
639 rq
= dl_task_offline_migration(rq
, p
);
640 rf
.cookie
= lockdep_pin_lock(&rq
->lock
);
643 * Now that the task has been migrated to the new RQ and we
644 * have that locked, proceed as normal and enqueue the task
650 enqueue_task_dl(rq
, p
, ENQUEUE_REPLENISH
);
651 if (dl_task(rq
->curr
))
652 check_preempt_curr_dl(rq
, p
, 0);
658 * Queueing this task back might have overloaded rq, check if we need
659 * to kick someone away.
661 if (has_pushable_dl_tasks(rq
)) {
663 * Nothing relies on rq->lock after this, so its safe to drop
666 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
668 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
673 task_rq_unlock(rq
, p
, &rf
);
676 * This can free the task_struct, including this hrtimer, do not touch
677 * anything related to that after this.
681 return HRTIMER_NORESTART
;
684 void init_dl_task_timer(struct sched_dl_entity
*dl_se
)
686 struct hrtimer
*timer
= &dl_se
->dl_timer
;
688 hrtimer_init(timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
689 timer
->function
= dl_task_timer
;
693 int dl_runtime_exceeded(struct sched_dl_entity
*dl_se
)
695 return (dl_se
->runtime
<= 0);
698 extern bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
);
701 * Update the current task's runtime statistics (provided it is still
702 * a -deadline task and has not been removed from the dl_rq).
704 static void update_curr_dl(struct rq
*rq
)
706 struct task_struct
*curr
= rq
->curr
;
707 struct sched_dl_entity
*dl_se
= &curr
->dl
;
710 if (!dl_task(curr
) || !on_dl_rq(dl_se
))
714 * Consumed budget is computed considering the time as
715 * observed by schedulable tasks (excluding time spent
716 * in hardirq context, etc.). Deadlines are instead
717 * computed using hard walltime. This seems to be the more
718 * natural solution, but the full ramifications of this
719 * approach need further study.
721 delta_exec
= rq_clock_task(rq
) - curr
->se
.exec_start
;
722 if (unlikely((s64
)delta_exec
<= 0)) {
723 if (unlikely(dl_se
->dl_yielded
))
728 /* kick cpufreq (see the comment in linux/cpufreq.h). */
729 if (cpu_of(rq
) == smp_processor_id())
730 cpufreq_trigger_update(rq_clock(rq
));
732 schedstat_set(curr
->se
.statistics
.exec_max
,
733 max(curr
->se
.statistics
.exec_max
, delta_exec
));
735 curr
->se
.sum_exec_runtime
+= delta_exec
;
736 account_group_exec_runtime(curr
, delta_exec
);
738 curr
->se
.exec_start
= rq_clock_task(rq
);
739 cpuacct_charge(curr
, delta_exec
);
741 sched_rt_avg_update(rq
, delta_exec
);
743 dl_se
->runtime
-= delta_exec
;
746 if (dl_runtime_exceeded(dl_se
) || dl_se
->dl_yielded
) {
747 dl_se
->dl_throttled
= 1;
748 __dequeue_task_dl(rq
, curr
, 0);
749 if (unlikely(dl_se
->dl_boosted
|| !start_dl_timer(curr
)))
750 enqueue_task_dl(rq
, curr
, ENQUEUE_REPLENISH
);
752 if (!is_leftmost(curr
, &rq
->dl
))
757 * Because -- for now -- we share the rt bandwidth, we need to
758 * account our runtime there too, otherwise actual rt tasks
759 * would be able to exceed the shared quota.
761 * Account to the root rt group for now.
763 * The solution we're working towards is having the RT groups scheduled
764 * using deadline servers -- however there's a few nasties to figure
765 * out before that can happen.
767 if (rt_bandwidth_enabled()) {
768 struct rt_rq
*rt_rq
= &rq
->rt
;
770 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
772 * We'll let actual RT tasks worry about the overflow here, we
773 * have our own CBS to keep us inline; only account when RT
774 * bandwidth is relevant.
776 if (sched_rt_bandwidth_account(rt_rq
))
777 rt_rq
->rt_time
+= delta_exec
;
778 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
784 static void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
786 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
788 if (dl_rq
->earliest_dl
.curr
== 0 ||
789 dl_time_before(deadline
, dl_rq
->earliest_dl
.curr
)) {
790 dl_rq
->earliest_dl
.curr
= deadline
;
791 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, deadline
);
795 static void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
)
797 struct rq
*rq
= rq_of_dl_rq(dl_rq
);
800 * Since we may have removed our earliest (and/or next earliest)
801 * task we must recompute them.
803 if (!dl_rq
->dl_nr_running
) {
804 dl_rq
->earliest_dl
.curr
= 0;
805 dl_rq
->earliest_dl
.next
= 0;
806 cpudl_clear(&rq
->rd
->cpudl
, rq
->cpu
);
808 struct rb_node
*leftmost
= dl_rq
->rb_leftmost
;
809 struct sched_dl_entity
*entry
;
811 entry
= rb_entry(leftmost
, struct sched_dl_entity
, rb_node
);
812 dl_rq
->earliest_dl
.curr
= entry
->deadline
;
813 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, entry
->deadline
);
819 static inline void inc_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
820 static inline void dec_dl_deadline(struct dl_rq
*dl_rq
, u64 deadline
) {}
822 #endif /* CONFIG_SMP */
825 void inc_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
827 int prio
= dl_task_of(dl_se
)->prio
;
828 u64 deadline
= dl_se
->deadline
;
830 WARN_ON(!dl_prio(prio
));
831 dl_rq
->dl_nr_running
++;
832 add_nr_running(rq_of_dl_rq(dl_rq
), 1);
834 inc_dl_deadline(dl_rq
, deadline
);
835 inc_dl_migration(dl_se
, dl_rq
);
839 void dec_dl_tasks(struct sched_dl_entity
*dl_se
, struct dl_rq
*dl_rq
)
841 int prio
= dl_task_of(dl_se
)->prio
;
843 WARN_ON(!dl_prio(prio
));
844 WARN_ON(!dl_rq
->dl_nr_running
);
845 dl_rq
->dl_nr_running
--;
846 sub_nr_running(rq_of_dl_rq(dl_rq
), 1);
848 dec_dl_deadline(dl_rq
, dl_se
->deadline
);
849 dec_dl_migration(dl_se
, dl_rq
);
852 static void __enqueue_dl_entity(struct sched_dl_entity
*dl_se
)
854 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
855 struct rb_node
**link
= &dl_rq
->rb_root
.rb_node
;
856 struct rb_node
*parent
= NULL
;
857 struct sched_dl_entity
*entry
;
860 BUG_ON(!RB_EMPTY_NODE(&dl_se
->rb_node
));
864 entry
= rb_entry(parent
, struct sched_dl_entity
, rb_node
);
865 if (dl_time_before(dl_se
->deadline
, entry
->deadline
))
866 link
= &parent
->rb_left
;
868 link
= &parent
->rb_right
;
874 dl_rq
->rb_leftmost
= &dl_se
->rb_node
;
876 rb_link_node(&dl_se
->rb_node
, parent
, link
);
877 rb_insert_color(&dl_se
->rb_node
, &dl_rq
->rb_root
);
879 inc_dl_tasks(dl_se
, dl_rq
);
882 static void __dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
884 struct dl_rq
*dl_rq
= dl_rq_of_se(dl_se
);
886 if (RB_EMPTY_NODE(&dl_se
->rb_node
))
889 if (dl_rq
->rb_leftmost
== &dl_se
->rb_node
) {
890 struct rb_node
*next_node
;
892 next_node
= rb_next(&dl_se
->rb_node
);
893 dl_rq
->rb_leftmost
= next_node
;
896 rb_erase(&dl_se
->rb_node
, &dl_rq
->rb_root
);
897 RB_CLEAR_NODE(&dl_se
->rb_node
);
899 dec_dl_tasks(dl_se
, dl_rq
);
903 enqueue_dl_entity(struct sched_dl_entity
*dl_se
,
904 struct sched_dl_entity
*pi_se
, int flags
)
906 BUG_ON(on_dl_rq(dl_se
));
909 * If this is a wakeup or a new instance, the scheduling
910 * parameters of the task might need updating. Otherwise,
911 * we want a replenishment of its runtime.
913 if (flags
& ENQUEUE_WAKEUP
)
914 update_dl_entity(dl_se
, pi_se
);
915 else if (flags
& ENQUEUE_REPLENISH
)
916 replenish_dl_entity(dl_se
, pi_se
);
918 __enqueue_dl_entity(dl_se
);
921 static void dequeue_dl_entity(struct sched_dl_entity
*dl_se
)
923 __dequeue_dl_entity(dl_se
);
926 static void enqueue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
928 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
929 struct sched_dl_entity
*pi_se
= &p
->dl
;
932 * Use the scheduling parameters of the top pi-waiter
933 * task if we have one and its (absolute) deadline is
934 * smaller than our one... OTW we keep our runtime and
937 if (pi_task
&& p
->dl
.dl_boosted
&& dl_prio(pi_task
->normal_prio
)) {
938 pi_se
= &pi_task
->dl
;
939 } else if (!dl_prio(p
->normal_prio
)) {
941 * Special case in which we have a !SCHED_DEADLINE task
942 * that is going to be deboosted, but exceedes its
943 * runtime while doing so. No point in replenishing
944 * it, as it's going to return back to its original
945 * scheduling class after this.
947 BUG_ON(!p
->dl
.dl_boosted
|| flags
!= ENQUEUE_REPLENISH
);
952 * If p is throttled, we do nothing. In fact, if it exhausted
953 * its budget it needs a replenishment and, since it now is on
954 * its rq, the bandwidth timer callback (which clearly has not
955 * run yet) will take care of this.
957 if (p
->dl
.dl_throttled
&& !(flags
& ENQUEUE_REPLENISH
))
960 enqueue_dl_entity(&p
->dl
, pi_se
, flags
);
962 if (!task_current(rq
, p
) && tsk_nr_cpus_allowed(p
) > 1)
963 enqueue_pushable_dl_task(rq
, p
);
966 static void __dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
968 dequeue_dl_entity(&p
->dl
);
969 dequeue_pushable_dl_task(rq
, p
);
972 static void dequeue_task_dl(struct rq
*rq
, struct task_struct
*p
, int flags
)
975 __dequeue_task_dl(rq
, p
, flags
);
979 * Yield task semantic for -deadline tasks is:
981 * get off from the CPU until our next instance, with
982 * a new runtime. This is of little use now, since we
983 * don't have a bandwidth reclaiming mechanism. Anyway,
984 * bandwidth reclaiming is planned for the future, and
985 * yield_task_dl will indicate that some spare budget
986 * is available for other task instances to use it.
988 static void yield_task_dl(struct rq
*rq
)
991 * We make the task go to sleep until its current deadline by
992 * forcing its runtime to zero. This way, update_curr_dl() stops
993 * it and the bandwidth timer will wake it up and will give it
994 * new scheduling parameters (thanks to dl_yielded=1).
996 rq
->curr
->dl
.dl_yielded
= 1;
1001 * Tell update_rq_clock() that we've just updated,
1002 * so we don't do microscopic update in schedule()
1003 * and double the fastpath cost.
1005 rq_clock_skip_update(rq
, true);
1010 static int find_later_rq(struct task_struct
*task
);
1013 select_task_rq_dl(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
)
1015 struct task_struct
*curr
;
1018 if (sd_flag
!= SD_BALANCE_WAKE
)
1024 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1027 * If we are dealing with a -deadline task, we must
1028 * decide where to wake it up.
1029 * If it has a later deadline and the current task
1030 * on this rq can't move (provided the waking task
1031 * can!) we prefer to send it somewhere else. On the
1032 * other hand, if it has a shorter deadline, we
1033 * try to make it stay here, it might be important.
1035 if (unlikely(dl_task(curr
)) &&
1036 (tsk_nr_cpus_allowed(curr
) < 2 ||
1037 !dl_entity_preempt(&p
->dl
, &curr
->dl
)) &&
1038 (tsk_nr_cpus_allowed(p
) > 1)) {
1039 int target
= find_later_rq(p
);
1042 (dl_time_before(p
->dl
.deadline
,
1043 cpu_rq(target
)->dl
.earliest_dl
.curr
) ||
1044 (cpu_rq(target
)->dl
.dl_nr_running
== 0)))
1053 static void check_preempt_equal_dl(struct rq
*rq
, struct task_struct
*p
)
1056 * Current can't be migrated, useless to reschedule,
1057 * let's hope p can move out.
1059 if (tsk_nr_cpus_allowed(rq
->curr
) == 1 ||
1060 cpudl_find(&rq
->rd
->cpudl
, rq
->curr
, NULL
) == -1)
1064 * p is migratable, so let's not schedule it and
1065 * see if it is pushed or pulled somewhere else.
1067 if (tsk_nr_cpus_allowed(p
) != 1 &&
1068 cpudl_find(&rq
->rd
->cpudl
, p
, NULL
) != -1)
1074 #endif /* CONFIG_SMP */
1077 * Only called when both the current and waking task are -deadline
1080 static void check_preempt_curr_dl(struct rq
*rq
, struct task_struct
*p
,
1083 if (dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
)) {
1090 * In the unlikely case current and p have the same deadline
1091 * let us try to decide what's the best thing to do...
1093 if ((p
->dl
.deadline
== rq
->curr
->dl
.deadline
) &&
1094 !test_tsk_need_resched(rq
->curr
))
1095 check_preempt_equal_dl(rq
, p
);
1096 #endif /* CONFIG_SMP */
1099 #ifdef CONFIG_SCHED_HRTICK
1100 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1102 hrtick_start(rq
, p
->dl
.runtime
);
1104 #else /* !CONFIG_SCHED_HRTICK */
1105 static void start_hrtick_dl(struct rq
*rq
, struct task_struct
*p
)
1110 static struct sched_dl_entity
*pick_next_dl_entity(struct rq
*rq
,
1111 struct dl_rq
*dl_rq
)
1113 struct rb_node
*left
= dl_rq
->rb_leftmost
;
1118 return rb_entry(left
, struct sched_dl_entity
, rb_node
);
1121 struct task_struct
*
1122 pick_next_task_dl(struct rq
*rq
, struct task_struct
*prev
, struct pin_cookie cookie
)
1124 struct sched_dl_entity
*dl_se
;
1125 struct task_struct
*p
;
1126 struct dl_rq
*dl_rq
;
1130 if (need_pull_dl_task(rq
, prev
)) {
1132 * This is OK, because current is on_cpu, which avoids it being
1133 * picked for load-balance and preemption/IRQs are still
1134 * disabled avoiding further scheduler activity on it and we're
1135 * being very careful to re-start the picking loop.
1137 lockdep_unpin_lock(&rq
->lock
, cookie
);
1139 lockdep_repin_lock(&rq
->lock
, cookie
);
1141 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1142 * means a stop task can slip in, in which case we need to
1143 * re-start task selection.
1145 if (rq
->stop
&& task_on_rq_queued(rq
->stop
))
1150 * When prev is DL, we may throttle it in put_prev_task().
1151 * So, we update time before we check for dl_nr_running.
1153 if (prev
->sched_class
== &dl_sched_class
)
1156 if (unlikely(!dl_rq
->dl_nr_running
))
1159 put_prev_task(rq
, prev
);
1161 dl_se
= pick_next_dl_entity(rq
, dl_rq
);
1164 p
= dl_task_of(dl_se
);
1165 p
->se
.exec_start
= rq_clock_task(rq
);
1167 /* Running task will never be pushed. */
1168 dequeue_pushable_dl_task(rq
, p
);
1170 if (hrtick_enabled(rq
))
1171 start_hrtick_dl(rq
, p
);
1173 queue_push_tasks(rq
);
1178 static void put_prev_task_dl(struct rq
*rq
, struct task_struct
*p
)
1182 if (on_dl_rq(&p
->dl
) && tsk_nr_cpus_allowed(p
) > 1)
1183 enqueue_pushable_dl_task(rq
, p
);
1186 static void task_tick_dl(struct rq
*rq
, struct task_struct
*p
, int queued
)
1191 * Even when we have runtime, update_curr_dl() might have resulted in us
1192 * not being the leftmost task anymore. In that case NEED_RESCHED will
1193 * be set and schedule() will start a new hrtick for the next task.
1195 if (hrtick_enabled(rq
) && queued
&& p
->dl
.runtime
> 0 &&
1196 is_leftmost(p
, &rq
->dl
))
1197 start_hrtick_dl(rq
, p
);
1200 static void task_fork_dl(struct task_struct
*p
)
1203 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1208 static void task_dead_dl(struct task_struct
*p
)
1210 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
1213 * Since we are TASK_DEAD we won't slip out of the domain!
1215 raw_spin_lock_irq(&dl_b
->lock
);
1216 /* XXX we should retain the bw until 0-lag */
1217 dl_b
->total_bw
-= p
->dl
.dl_bw
;
1218 raw_spin_unlock_irq(&dl_b
->lock
);
1221 static void set_curr_task_dl(struct rq
*rq
)
1223 struct task_struct
*p
= rq
->curr
;
1225 p
->se
.exec_start
= rq_clock_task(rq
);
1227 /* You can't push away the running task */
1228 dequeue_pushable_dl_task(rq
, p
);
1233 /* Only try algorithms three times */
1234 #define DL_MAX_TRIES 3
1236 static int pick_dl_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1238 if (!task_running(rq
, p
) &&
1239 cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
1245 * Return the earliest pushable rq's task, which is suitable to be executed
1246 * on the CPU, NULL otherwise:
1248 static struct task_struct
*pick_earliest_pushable_dl_task(struct rq
*rq
, int cpu
)
1250 struct rb_node
*next_node
= rq
->dl
.pushable_dl_tasks_leftmost
;
1251 struct task_struct
*p
= NULL
;
1253 if (!has_pushable_dl_tasks(rq
))
1258 p
= rb_entry(next_node
, struct task_struct
, pushable_dl_tasks
);
1260 if (pick_dl_task(rq
, p
, cpu
))
1263 next_node
= rb_next(next_node
);
1270 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask_dl
);
1272 static int find_later_rq(struct task_struct
*task
)
1274 struct sched_domain
*sd
;
1275 struct cpumask
*later_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask_dl
);
1276 int this_cpu
= smp_processor_id();
1277 int best_cpu
, cpu
= task_cpu(task
);
1279 /* Make sure the mask is initialized first */
1280 if (unlikely(!later_mask
))
1283 if (tsk_nr_cpus_allowed(task
) == 1)
1287 * We have to consider system topology and task affinity
1288 * first, then we can look for a suitable cpu.
1290 best_cpu
= cpudl_find(&task_rq(task
)->rd
->cpudl
,
1296 * If we are here, some target has been found,
1297 * the most suitable of which is cached in best_cpu.
1298 * This is, among the runqueues where the current tasks
1299 * have later deadlines than the task's one, the rq
1300 * with the latest possible one.
1302 * Now we check how well this matches with task's
1303 * affinity and system topology.
1305 * The last cpu where the task run is our first
1306 * guess, since it is most likely cache-hot there.
1308 if (cpumask_test_cpu(cpu
, later_mask
))
1311 * Check if this_cpu is to be skipped (i.e., it is
1312 * not in the mask) or not.
1314 if (!cpumask_test_cpu(this_cpu
, later_mask
))
1318 for_each_domain(cpu
, sd
) {
1319 if (sd
->flags
& SD_WAKE_AFFINE
) {
1322 * If possible, preempting this_cpu is
1323 * cheaper than migrating.
1325 if (this_cpu
!= -1 &&
1326 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1332 * Last chance: if best_cpu is valid and is
1333 * in the mask, that becomes our choice.
1335 if (best_cpu
< nr_cpu_ids
&&
1336 cpumask_test_cpu(best_cpu
, sched_domain_span(sd
))) {
1345 * At this point, all our guesses failed, we just return
1346 * 'something', and let the caller sort the things out.
1351 cpu
= cpumask_any(later_mask
);
1352 if (cpu
< nr_cpu_ids
)
1358 /* Locks the rq it finds */
1359 static struct rq
*find_lock_later_rq(struct task_struct
*task
, struct rq
*rq
)
1361 struct rq
*later_rq
= NULL
;
1365 for (tries
= 0; tries
< DL_MAX_TRIES
; tries
++) {
1366 cpu
= find_later_rq(task
);
1368 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1371 later_rq
= cpu_rq(cpu
);
1373 if (later_rq
->dl
.dl_nr_running
&&
1374 !dl_time_before(task
->dl
.deadline
,
1375 later_rq
->dl
.earliest_dl
.curr
)) {
1377 * Target rq has tasks of equal or earlier deadline,
1378 * retrying does not release any lock and is unlikely
1379 * to yield a different result.
1385 /* Retry if something changed. */
1386 if (double_lock_balance(rq
, later_rq
)) {
1387 if (unlikely(task_rq(task
) != rq
||
1388 !cpumask_test_cpu(later_rq
->cpu
,
1389 tsk_cpus_allowed(task
)) ||
1390 task_running(rq
, task
) ||
1392 !task_on_rq_queued(task
))) {
1393 double_unlock_balance(rq
, later_rq
);
1400 * If the rq we found has no -deadline task, or
1401 * its earliest one has a later deadline than our
1402 * task, the rq is a good one.
1404 if (!later_rq
->dl
.dl_nr_running
||
1405 dl_time_before(task
->dl
.deadline
,
1406 later_rq
->dl
.earliest_dl
.curr
))
1409 /* Otherwise we try again. */
1410 double_unlock_balance(rq
, later_rq
);
1417 static struct task_struct
*pick_next_pushable_dl_task(struct rq
*rq
)
1419 struct task_struct
*p
;
1421 if (!has_pushable_dl_tasks(rq
))
1424 p
= rb_entry(rq
->dl
.pushable_dl_tasks_leftmost
,
1425 struct task_struct
, pushable_dl_tasks
);
1427 BUG_ON(rq
->cpu
!= task_cpu(p
));
1428 BUG_ON(task_current(rq
, p
));
1429 BUG_ON(tsk_nr_cpus_allowed(p
) <= 1);
1431 BUG_ON(!task_on_rq_queued(p
));
1432 BUG_ON(!dl_task(p
));
1438 * See if the non running -deadline tasks on this rq
1439 * can be sent to some other CPU where they can preempt
1440 * and start executing.
1442 static int push_dl_task(struct rq
*rq
)
1444 struct task_struct
*next_task
;
1445 struct rq
*later_rq
;
1448 if (!rq
->dl
.overloaded
)
1451 next_task
= pick_next_pushable_dl_task(rq
);
1456 if (unlikely(next_task
== rq
->curr
)) {
1462 * If next_task preempts rq->curr, and rq->curr
1463 * can move away, it makes sense to just reschedule
1464 * without going further in pushing next_task.
1466 if (dl_task(rq
->curr
) &&
1467 dl_time_before(next_task
->dl
.deadline
, rq
->curr
->dl
.deadline
) &&
1468 tsk_nr_cpus_allowed(rq
->curr
) > 1) {
1473 /* We might release rq lock */
1474 get_task_struct(next_task
);
1476 /* Will lock the rq it'll find */
1477 later_rq
= find_lock_later_rq(next_task
, rq
);
1479 struct task_struct
*task
;
1482 * We must check all this again, since
1483 * find_lock_later_rq releases rq->lock and it is
1484 * then possible that next_task has migrated.
1486 task
= pick_next_pushable_dl_task(rq
);
1487 if (task_cpu(next_task
) == rq
->cpu
&& task
== next_task
) {
1489 * The task is still there. We don't try
1490 * again, some other cpu will pull it when ready.
1499 put_task_struct(next_task
);
1504 deactivate_task(rq
, next_task
, 0);
1505 set_task_cpu(next_task
, later_rq
->cpu
);
1506 activate_task(later_rq
, next_task
, 0);
1509 resched_curr(later_rq
);
1511 double_unlock_balance(rq
, later_rq
);
1514 put_task_struct(next_task
);
1519 static void push_dl_tasks(struct rq
*rq
)
1521 /* push_dl_task() will return true if it moved a -deadline task */
1522 while (push_dl_task(rq
))
1526 static void pull_dl_task(struct rq
*this_rq
)
1528 int this_cpu
= this_rq
->cpu
, cpu
;
1529 struct task_struct
*p
;
1530 bool resched
= false;
1532 u64 dmin
= LONG_MAX
;
1534 if (likely(!dl_overloaded(this_rq
)))
1538 * Match the barrier from dl_set_overloaded; this guarantees that if we
1539 * see overloaded we must also see the dlo_mask bit.
1543 for_each_cpu(cpu
, this_rq
->rd
->dlo_mask
) {
1544 if (this_cpu
== cpu
)
1547 src_rq
= cpu_rq(cpu
);
1550 * It looks racy, abd it is! However, as in sched_rt.c,
1551 * we are fine with this.
1553 if (this_rq
->dl
.dl_nr_running
&&
1554 dl_time_before(this_rq
->dl
.earliest_dl
.curr
,
1555 src_rq
->dl
.earliest_dl
.next
))
1558 /* Might drop this_rq->lock */
1559 double_lock_balance(this_rq
, src_rq
);
1562 * If there are no more pullable tasks on the
1563 * rq, we're done with it.
1565 if (src_rq
->dl
.dl_nr_running
<= 1)
1568 p
= pick_earliest_pushable_dl_task(src_rq
, this_cpu
);
1571 * We found a task to be pulled if:
1572 * - it preempts our current (if there's one),
1573 * - it will preempt the last one we pulled (if any).
1575 if (p
&& dl_time_before(p
->dl
.deadline
, dmin
) &&
1576 (!this_rq
->dl
.dl_nr_running
||
1577 dl_time_before(p
->dl
.deadline
,
1578 this_rq
->dl
.earliest_dl
.curr
))) {
1579 WARN_ON(p
== src_rq
->curr
);
1580 WARN_ON(!task_on_rq_queued(p
));
1583 * Then we pull iff p has actually an earlier
1584 * deadline than the current task of its runqueue.
1586 if (dl_time_before(p
->dl
.deadline
,
1587 src_rq
->curr
->dl
.deadline
))
1592 deactivate_task(src_rq
, p
, 0);
1593 set_task_cpu(p
, this_cpu
);
1594 activate_task(this_rq
, p
, 0);
1595 dmin
= p
->dl
.deadline
;
1597 /* Is there any other task even earlier? */
1600 double_unlock_balance(this_rq
, src_rq
);
1604 resched_curr(this_rq
);
1608 * Since the task is not running and a reschedule is not going to happen
1609 * anytime soon on its runqueue, we try pushing it away now.
1611 static void task_woken_dl(struct rq
*rq
, struct task_struct
*p
)
1613 if (!task_running(rq
, p
) &&
1614 !test_tsk_need_resched(rq
->curr
) &&
1615 tsk_nr_cpus_allowed(p
) > 1 &&
1616 dl_task(rq
->curr
) &&
1617 (tsk_nr_cpus_allowed(rq
->curr
) < 2 ||
1618 !dl_entity_preempt(&p
->dl
, &rq
->curr
->dl
))) {
1623 static void set_cpus_allowed_dl(struct task_struct
*p
,
1624 const struct cpumask
*new_mask
)
1626 struct root_domain
*src_rd
;
1629 BUG_ON(!dl_task(p
));
1634 * Migrating a SCHED_DEADLINE task between exclusive
1635 * cpusets (different root_domains) entails a bandwidth
1636 * update. We already made space for us in the destination
1637 * domain (see cpuset_can_attach()).
1639 if (!cpumask_intersects(src_rd
->span
, new_mask
)) {
1640 struct dl_bw
*src_dl_b
;
1642 src_dl_b
= dl_bw_of(cpu_of(rq
));
1644 * We now free resources of the root_domain we are migrating
1645 * off. In the worst case, sched_setattr() may temporary fail
1646 * until we complete the update.
1648 raw_spin_lock(&src_dl_b
->lock
);
1649 __dl_clear(src_dl_b
, p
->dl
.dl_bw
);
1650 raw_spin_unlock(&src_dl_b
->lock
);
1653 set_cpus_allowed_common(p
, new_mask
);
1656 /* Assumes rq->lock is held */
1657 static void rq_online_dl(struct rq
*rq
)
1659 if (rq
->dl
.overloaded
)
1660 dl_set_overload(rq
);
1662 cpudl_set_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
1663 if (rq
->dl
.dl_nr_running
> 0)
1664 cpudl_set(&rq
->rd
->cpudl
, rq
->cpu
, rq
->dl
.earliest_dl
.curr
);
1667 /* Assumes rq->lock is held */
1668 static void rq_offline_dl(struct rq
*rq
)
1670 if (rq
->dl
.overloaded
)
1671 dl_clear_overload(rq
);
1673 cpudl_clear(&rq
->rd
->cpudl
, rq
->cpu
);
1674 cpudl_clear_freecpu(&rq
->rd
->cpudl
, rq
->cpu
);
1677 void __init
init_sched_dl_class(void)
1681 for_each_possible_cpu(i
)
1682 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl
, i
),
1683 GFP_KERNEL
, cpu_to_node(i
));
1686 #endif /* CONFIG_SMP */
1688 static void switched_from_dl(struct rq
*rq
, struct task_struct
*p
)
1691 * Start the deadline timer; if we switch back to dl before this we'll
1692 * continue consuming our current CBS slice. If we stay outside of
1693 * SCHED_DEADLINE until the deadline passes, the timer will reset the
1696 if (!start_dl_timer(p
))
1697 __dl_clear_params(p
);
1700 * Since this might be the only -deadline task on the rq,
1701 * this is the right place to try to pull some other one
1702 * from an overloaded cpu, if any.
1704 if (!task_on_rq_queued(p
) || rq
->dl
.dl_nr_running
)
1707 queue_pull_task(rq
);
1711 * When switching to -deadline, we may overload the rq, then
1712 * we try to push someone off, if possible.
1714 static void switched_to_dl(struct rq
*rq
, struct task_struct
*p
)
1717 /* If p is not queued we will update its parameters at next wakeup. */
1718 if (!task_on_rq_queued(p
))
1722 * If p is boosted we already updated its params in
1723 * rt_mutex_setprio()->enqueue_task(..., ENQUEUE_REPLENISH),
1724 * p's deadline being now already after rq_clock(rq).
1726 if (dl_time_before(p
->dl
.deadline
, rq_clock(rq
)))
1727 setup_new_dl_entity(&p
->dl
);
1729 if (rq
->curr
!= p
) {
1731 if (tsk_nr_cpus_allowed(p
) > 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 */
1778 const struct sched_class dl_sched_class
= {
1779 .next
= &rt_sched_class
,
1780 .enqueue_task
= enqueue_task_dl
,
1781 .dequeue_task
= dequeue_task_dl
,
1782 .yield_task
= yield_task_dl
,
1784 .check_preempt_curr
= check_preempt_curr_dl
,
1786 .pick_next_task
= pick_next_task_dl
,
1787 .put_prev_task
= put_prev_task_dl
,
1790 .select_task_rq
= select_task_rq_dl
,
1791 .set_cpus_allowed
= set_cpus_allowed_dl
,
1792 .rq_online
= rq_online_dl
,
1793 .rq_offline
= rq_offline_dl
,
1794 .task_woken
= task_woken_dl
,
1797 .set_curr_task
= set_curr_task_dl
,
1798 .task_tick
= task_tick_dl
,
1799 .task_fork
= task_fork_dl
,
1800 .task_dead
= task_dead_dl
,
1802 .prio_changed
= prio_changed_dl
,
1803 .switched_from
= switched_from_dl
,
1804 .switched_to
= switched_to_dl
,
1806 .update_curr
= update_curr_dl
,
1809 #ifdef CONFIG_SCHED_DEBUG
1810 extern void print_dl_rq(struct seq_file
*m
, int cpu
, struct dl_rq
*dl_rq
);
1812 void print_dl_stats(struct seq_file
*m
, int cpu
)
1814 print_dl_rq(m
, cpu
, &cpu_rq(cpu
)->dl
);
1816 #endif /* CONFIG_SCHED_DEBUG */