2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
8 static inline int rt_overloaded(struct rq
*rq
)
10 return atomic_read(&rq
->rd
->rto_count
);
13 static inline void rt_set_overload(struct rq
*rq
)
15 cpu_set(rq
->cpu
, rq
->rd
->rto_mask
);
17 * Make sure the mask is visible before we set
18 * the overload count. That is checked to determine
19 * if we should look at the mask. It would be a shame
20 * if we looked at the mask, but the mask was not
24 atomic_inc(&rq
->rd
->rto_count
);
27 static inline void rt_clear_overload(struct rq
*rq
)
29 /* the order here really doesn't matter */
30 atomic_dec(&rq
->rd
->rto_count
);
31 cpu_clear(rq
->cpu
, rq
->rd
->rto_mask
);
34 static void update_rt_migration(struct rq
*rq
)
36 if (rq
->rt
.rt_nr_migratory
&& (rq
->rt
.rt_nr_running
> 1)) {
37 if (!rq
->rt
.overloaded
) {
39 rq
->rt
.overloaded
= 1;
41 } else if (rq
->rt
.overloaded
) {
42 rt_clear_overload(rq
);
43 rq
->rt
.overloaded
= 0;
46 #endif /* CONFIG_SMP */
49 * Update the current task's runtime statistics. Skip current tasks that
50 * are not in our scheduling class.
52 static void update_curr_rt(struct rq
*rq
)
54 struct task_struct
*curr
= rq
->curr
;
57 if (!task_has_rt_policy(curr
))
60 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
61 if (unlikely((s64
)delta_exec
< 0))
64 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
66 curr
->se
.sum_exec_runtime
+= delta_exec
;
67 curr
->se
.exec_start
= rq
->clock
;
68 cpuacct_charge(curr
, delta_exec
);
71 static inline void inc_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
74 rq
->rt
.rt_nr_running
++;
76 if (p
->prio
< rq
->rt
.highest_prio
)
77 rq
->rt
.highest_prio
= p
->prio
;
78 if (p
->nr_cpus_allowed
> 1)
79 rq
->rt
.rt_nr_migratory
++;
81 update_rt_migration(rq
);
82 #endif /* CONFIG_SMP */
85 static inline void dec_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
88 WARN_ON(!rq
->rt
.rt_nr_running
);
89 rq
->rt
.rt_nr_running
--;
91 if (rq
->rt
.rt_nr_running
) {
92 struct rt_prio_array
*array
;
94 WARN_ON(p
->prio
< rq
->rt
.highest_prio
);
95 if (p
->prio
== rq
->rt
.highest_prio
) {
97 array
= &rq
->rt
.active
;
99 sched_find_first_bit(array
->bitmap
);
100 } /* otherwise leave rq->highest prio alone */
102 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
103 if (p
->nr_cpus_allowed
> 1)
104 rq
->rt
.rt_nr_migratory
--;
106 update_rt_migration(rq
);
107 #endif /* CONFIG_SMP */
110 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
112 struct rt_prio_array
*array
= &rq
->rt
.active
;
114 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
115 __set_bit(p
->prio
, array
->bitmap
);
116 inc_cpu_load(rq
, p
->se
.load
.weight
);
122 * Adding/removing a task to/from a priority array:
124 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
126 struct rt_prio_array
*array
= &rq
->rt
.active
;
130 list_del(&p
->run_list
);
131 if (list_empty(array
->queue
+ p
->prio
))
132 __clear_bit(p
->prio
, array
->bitmap
);
133 dec_cpu_load(rq
, p
->se
.load
.weight
);
139 * Put task to the end of the run list without the overhead of dequeue
140 * followed by enqueue.
142 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
144 struct rt_prio_array
*array
= &rq
->rt
.active
;
146 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
150 yield_task_rt(struct rq
*rq
)
152 requeue_task_rt(rq
, rq
->curr
);
156 static int find_lowest_rq(struct task_struct
*task
);
158 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
160 struct rq
*rq
= task_rq(p
);
163 * If the current task is an RT task, then
164 * try to see if we can wake this RT task up on another
165 * runqueue. Otherwise simply start this RT task
166 * on its current runqueue.
168 * We want to avoid overloading runqueues. Even if
169 * the RT task is of higher priority than the current RT task.
170 * RT tasks behave differently than other tasks. If
171 * one gets preempted, we try to push it off to another queue.
172 * So trying to keep a preempting RT task on the same
173 * cache hot CPU will force the running RT task to
174 * a cold CPU. So we waste all the cache for the lower
175 * RT task in hopes of saving some of a RT task
176 * that is just being woken and probably will have
179 if (unlikely(rt_task(rq
->curr
)) &&
180 (p
->nr_cpus_allowed
> 1)) {
181 int cpu
= find_lowest_rq(p
);
183 return (cpu
== -1) ? task_cpu(p
) : cpu
;
187 * Otherwise, just let it ride on the affined RQ and the
188 * post-schedule router will push the preempted task away
192 #endif /* CONFIG_SMP */
195 * Preempt the current task with a newly woken task if needed:
197 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
199 if (p
->prio
< rq
->curr
->prio
)
200 resched_task(rq
->curr
);
203 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
205 struct rt_prio_array
*array
= &rq
->rt
.active
;
206 struct task_struct
*next
;
207 struct list_head
*queue
;
210 idx
= sched_find_first_bit(array
->bitmap
);
211 if (idx
>= MAX_RT_PRIO
)
214 queue
= array
->queue
+ idx
;
215 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
217 next
->se
.exec_start
= rq
->clock
;
222 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
225 p
->se
.exec_start
= 0;
229 /* Only try algorithms three times */
230 #define RT_MAX_TRIES 3
232 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
233 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
235 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
237 if (!task_running(rq
, p
) &&
238 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
239 (p
->nr_cpus_allowed
> 1))
244 /* Return the second highest RT task, NULL otherwise */
245 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
247 struct rt_prio_array
*array
= &rq
->rt
.active
;
248 struct task_struct
*next
;
249 struct list_head
*queue
;
252 if (likely(rq
->rt
.rt_nr_running
< 2))
255 idx
= sched_find_first_bit(array
->bitmap
);
256 if (unlikely(idx
>= MAX_RT_PRIO
)) {
257 WARN_ON(1); /* rt_nr_running is bad */
261 queue
= array
->queue
+ idx
;
262 BUG_ON(list_empty(queue
));
264 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
265 if (unlikely(pick_rt_task(rq
, next
, cpu
)))
268 if (queue
->next
->next
!= queue
) {
270 next
= list_entry(queue
->next
->next
, struct task_struct
,
272 if (pick_rt_task(rq
, next
, cpu
))
277 /* slower, but more flexible */
278 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
279 if (unlikely(idx
>= MAX_RT_PRIO
))
282 queue
= array
->queue
+ idx
;
283 BUG_ON(list_empty(queue
));
285 list_for_each_entry(next
, queue
, run_list
) {
286 if (pick_rt_task(rq
, next
, cpu
))
296 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
298 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
300 int lowest_prio
= -1;
305 cpus_and(*lowest_mask
, task_rq(task
)->rd
->online
, task
->cpus_allowed
);
308 * Scan each rq for the lowest prio.
310 for_each_cpu_mask(cpu
, *lowest_mask
) {
311 struct rq
*rq
= cpu_rq(cpu
);
313 /* We look for lowest RT prio or non-rt CPU */
314 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
316 * if we already found a low RT queue
317 * and now we found this non-rt queue
318 * clear the mask and set our bit.
319 * Otherwise just return the queue as is
320 * and the count==1 will cause the algorithm
321 * to use the first bit found.
323 if (lowest_cpu
!= -1) {
324 cpus_clear(*lowest_mask
);
325 cpu_set(rq
->cpu
, *lowest_mask
);
330 /* no locking for now */
331 if ((rq
->rt
.highest_prio
> task
->prio
)
332 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
333 if (rq
->rt
.highest_prio
> lowest_prio
) {
334 /* new low - clear old data */
335 lowest_prio
= rq
->rt
.highest_prio
;
341 cpu_clear(cpu
, *lowest_mask
);
345 * Clear out all the set bits that represent
346 * runqueues that were of higher prio than
349 if (lowest_cpu
> 0) {
351 * Perhaps we could add another cpumask op to
352 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
353 * Then that could be optimized to use memset and such.
355 for_each_cpu_mask(cpu
, *lowest_mask
) {
356 if (cpu
>= lowest_cpu
)
358 cpu_clear(cpu
, *lowest_mask
);
365 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
369 /* "this_cpu" is cheaper to preempt than a remote processor */
370 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
373 first
= first_cpu(*mask
);
374 if (first
!= NR_CPUS
)
380 static int find_lowest_rq(struct task_struct
*task
)
382 struct sched_domain
*sd
;
383 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
384 int this_cpu
= smp_processor_id();
385 int cpu
= task_cpu(task
);
386 int count
= find_lowest_cpus(task
, lowest_mask
);
389 return -1; /* No targets found */
392 * There is no sense in performing an optimal search if only one
396 return first_cpu(*lowest_mask
);
399 * At this point we have built a mask of cpus representing the
400 * lowest priority tasks in the system. Now we want to elect
401 * the best one based on our affinity and topology.
403 * We prioritize the last cpu that the task executed on since
404 * it is most likely cache-hot in that location.
406 if (cpu_isset(cpu
, *lowest_mask
))
410 * Otherwise, we consult the sched_domains span maps to figure
411 * out which cpu is logically closest to our hot cache data.
414 this_cpu
= -1; /* Skip this_cpu opt if the same */
416 for_each_domain(cpu
, sd
) {
417 if (sd
->flags
& SD_WAKE_AFFINE
) {
418 cpumask_t domain_mask
;
421 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
423 best_cpu
= pick_optimal_cpu(this_cpu
,
431 * And finally, if there were no matches within the domains
432 * just give the caller *something* to work with from the compatible
435 return pick_optimal_cpu(this_cpu
, lowest_mask
);
438 /* Will lock the rq it finds */
439 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
441 struct rq
*lowest_rq
= NULL
;
445 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
446 cpu
= find_lowest_rq(task
);
448 if ((cpu
== -1) || (cpu
== rq
->cpu
))
451 lowest_rq
= cpu_rq(cpu
);
453 /* if the prio of this runqueue changed, try again */
454 if (double_lock_balance(rq
, lowest_rq
)) {
456 * We had to unlock the run queue. In
457 * the mean time, task could have
458 * migrated already or had its affinity changed.
459 * Also make sure that it wasn't scheduled on its rq.
461 if (unlikely(task_rq(task
) != rq
||
462 !cpu_isset(lowest_rq
->cpu
,
463 task
->cpus_allowed
) ||
464 task_running(rq
, task
) ||
467 spin_unlock(&lowest_rq
->lock
);
473 /* If this rq is still suitable use it. */
474 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
478 spin_unlock(&lowest_rq
->lock
);
486 * If the current CPU has more than one RT task, see if the non
487 * running task can migrate over to a CPU that is running a task
488 * of lesser priority.
490 static int push_rt_task(struct rq
*rq
)
492 struct task_struct
*next_task
;
493 struct rq
*lowest_rq
;
495 int paranoid
= RT_MAX_TRIES
;
497 if (!rq
->rt
.overloaded
)
500 next_task
= pick_next_highest_task_rt(rq
, -1);
505 if (unlikely(next_task
== rq
->curr
)) {
511 * It's possible that the next_task slipped in of
512 * higher priority than current. If that's the case
513 * just reschedule current.
515 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
516 resched_task(rq
->curr
);
520 /* We might release rq lock */
521 get_task_struct(next_task
);
523 /* find_lock_lowest_rq locks the rq if found */
524 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
526 struct task_struct
*task
;
528 * find lock_lowest_rq releases rq->lock
529 * so it is possible that next_task has changed.
530 * If it has, then try again.
532 task
= pick_next_highest_task_rt(rq
, -1);
533 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
534 put_task_struct(next_task
);
541 deactivate_task(rq
, next_task
, 0);
542 set_task_cpu(next_task
, lowest_rq
->cpu
);
543 activate_task(lowest_rq
, next_task
, 0);
545 resched_task(lowest_rq
->curr
);
547 spin_unlock(&lowest_rq
->lock
);
551 put_task_struct(next_task
);
557 * TODO: Currently we just use the second highest prio task on
558 * the queue, and stop when it can't migrate (or there's
559 * no more RT tasks). There may be a case where a lower
560 * priority RT task has a different affinity than the
561 * higher RT task. In this case the lower RT task could
562 * possibly be able to migrate where as the higher priority
563 * RT task could not. We currently ignore this issue.
564 * Enhancements are welcome!
566 static void push_rt_tasks(struct rq
*rq
)
568 /* push_rt_task will return true if it moved an RT */
569 while (push_rt_task(rq
))
573 static int pull_rt_task(struct rq
*this_rq
)
575 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
576 struct task_struct
*p
, *next
;
579 if (likely(!rt_overloaded(this_rq
)))
582 next
= pick_next_task_rt(this_rq
);
584 for_each_cpu_mask(cpu
, this_rq
->rd
->rto_mask
) {
588 src_rq
= cpu_rq(cpu
);
590 * We can potentially drop this_rq's lock in
591 * double_lock_balance, and another CPU could
592 * steal our next task - hence we must cause
593 * the caller to recalculate the next task
596 if (double_lock_balance(this_rq
, src_rq
)) {
597 struct task_struct
*old_next
= next
;
599 next
= pick_next_task_rt(this_rq
);
600 if (next
!= old_next
)
605 * Are there still pullable RT tasks?
607 if (src_rq
->rt
.rt_nr_running
<= 1) {
608 spin_unlock(&src_rq
->lock
);
612 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
615 * Do we have an RT task that preempts
616 * the to-be-scheduled task?
618 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
619 WARN_ON(p
== src_rq
->curr
);
620 WARN_ON(!p
->se
.on_rq
);
623 * There's a chance that p is higher in priority
624 * than what's currently running on its cpu.
625 * This is just that p is wakeing up and hasn't
626 * had a chance to schedule. We only pull
627 * p if it is lower in priority than the
628 * current task on the run queue or
629 * this_rq next task is lower in prio than
630 * the current task on that rq.
632 if (p
->prio
< src_rq
->curr
->prio
||
633 (next
&& next
->prio
< src_rq
->curr
->prio
))
638 deactivate_task(src_rq
, p
, 0);
639 set_task_cpu(p
, this_cpu
);
640 activate_task(this_rq
, p
, 0);
642 * We continue with the search, just in
643 * case there's an even higher prio task
644 * in another runqueue. (low likelyhood
647 * Update next so that we won't pick a task
648 * on another cpu with a priority lower (or equal)
649 * than the one we just picked.
655 spin_unlock(&src_rq
->lock
);
661 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
663 /* Try to pull RT tasks here if we lower this rq's prio */
664 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
668 static void post_schedule_rt(struct rq
*rq
)
671 * If we have more than one rt_task queued, then
672 * see if we can push the other rt_tasks off to other CPUS.
673 * Note we may release the rq lock, and since
674 * the lock was owned by prev, we need to release it
675 * first via finish_lock_switch and then reaquire it here.
677 if (unlikely(rq
->rt
.overloaded
)) {
678 spin_lock_irq(&rq
->lock
);
680 spin_unlock_irq(&rq
->lock
);
685 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
687 if (!task_running(rq
, p
) &&
688 (p
->prio
>= rq
->rt
.highest_prio
) &&
694 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
695 unsigned long max_load_move
,
696 struct sched_domain
*sd
, enum cpu_idle_type idle
,
697 int *all_pinned
, int *this_best_prio
)
699 /* don't touch RT tasks */
704 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
705 struct sched_domain
*sd
, enum cpu_idle_type idle
)
707 /* don't touch RT tasks */
711 static void set_cpus_allowed_rt(struct task_struct
*p
, cpumask_t
*new_mask
)
713 int weight
= cpus_weight(*new_mask
);
718 * Update the migration status of the RQ if we have an RT task
719 * which is running AND changing its weight value.
721 if (p
->se
.on_rq
&& (weight
!= p
->nr_cpus_allowed
)) {
722 struct rq
*rq
= task_rq(p
);
724 if ((p
->nr_cpus_allowed
<= 1) && (weight
> 1)) {
725 rq
->rt
.rt_nr_migratory
++;
726 } else if ((p
->nr_cpus_allowed
> 1) && (weight
<= 1)) {
727 BUG_ON(!rq
->rt
.rt_nr_migratory
);
728 rq
->rt
.rt_nr_migratory
--;
731 update_rt_migration(rq
);
734 p
->cpus_allowed
= *new_mask
;
735 p
->nr_cpus_allowed
= weight
;
738 /* Assumes rq->lock is held */
739 static void join_domain_rt(struct rq
*rq
)
741 if (rq
->rt
.overloaded
)
745 /* Assumes rq->lock is held */
746 static void leave_domain_rt(struct rq
*rq
)
748 if (rq
->rt
.overloaded
)
749 rt_clear_overload(rq
);
753 * When switch from the rt queue, we bring ourselves to a position
754 * that we might want to pull RT tasks from other runqueues.
756 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
,
760 * If there are other RT tasks then we will reschedule
761 * and the scheduling of the other RT tasks will handle
762 * the balancing. But if we are the last RT task
763 * we may need to handle the pulling of RT tasks
766 if (!rq
->rt
.rt_nr_running
)
769 #endif /* CONFIG_SMP */
772 * When switching a task to RT, we may overload the runqueue
773 * with RT tasks. In this case we try to push them off to
776 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
,
779 int check_resched
= 1;
782 * If we are already running, then there's nothing
783 * that needs to be done. But if we are not running
784 * we may need to preempt the current running task.
785 * If that current running task is also an RT task
786 * then see if we can move to another run queue.
790 if (rq
->rt
.overloaded
&& push_rt_task(rq
) &&
791 /* Don't resched if we changed runqueues */
794 #endif /* CONFIG_SMP */
795 if (check_resched
&& p
->prio
< rq
->curr
->prio
)
796 resched_task(rq
->curr
);
801 * Priority of the task has changed. This may cause
802 * us to initiate a push or pull.
804 static void prio_changed_rt(struct rq
*rq
, struct task_struct
*p
,
805 int oldprio
, int running
)
810 * If our priority decreases while running, we
811 * may need to pull tasks to this runqueue.
813 if (oldprio
< p
->prio
)
816 * If there's a higher priority task waiting to run
819 if (p
->prio
> rq
->rt
.highest_prio
)
822 /* For UP simply resched on drop of prio */
823 if (oldprio
< p
->prio
)
825 #endif /* CONFIG_SMP */
828 * This task is not running, but if it is
829 * greater than the current running task
832 if (p
->prio
< rq
->curr
->prio
)
833 resched_task(rq
->curr
);
838 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
)
843 * RR tasks need a special form of timeslice management.
844 * FIFO tasks have no timeslices.
846 if (p
->policy
!= SCHED_RR
)
852 p
->time_slice
= DEF_TIMESLICE
;
855 * Requeue to the end of queue if we are not the only element
858 if (p
->run_list
.prev
!= p
->run_list
.next
) {
859 requeue_task_rt(rq
, p
);
860 set_tsk_need_resched(p
);
864 static void set_curr_task_rt(struct rq
*rq
)
866 struct task_struct
*p
= rq
->curr
;
868 p
->se
.exec_start
= rq
->clock
;
871 const struct sched_class rt_sched_class
= {
872 .next
= &fair_sched_class
,
873 .enqueue_task
= enqueue_task_rt
,
874 .dequeue_task
= dequeue_task_rt
,
875 .yield_task
= yield_task_rt
,
877 .select_task_rq
= select_task_rq_rt
,
878 #endif /* CONFIG_SMP */
880 .check_preempt_curr
= check_preempt_curr_rt
,
882 .pick_next_task
= pick_next_task_rt
,
883 .put_prev_task
= put_prev_task_rt
,
886 .load_balance
= load_balance_rt
,
887 .move_one_task
= move_one_task_rt
,
888 .set_cpus_allowed
= set_cpus_allowed_rt
,
889 .join_domain
= join_domain_rt
,
890 .leave_domain
= leave_domain_rt
,
891 .pre_schedule
= pre_schedule_rt
,
892 .post_schedule
= post_schedule_rt
,
893 .task_wake_up
= task_wake_up_rt
,
894 .switched_from
= switched_from_rt
,
897 .set_curr_task
= set_curr_task_rt
,
898 .task_tick
= task_tick_rt
,
900 .prio_changed
= prio_changed_rt
,
901 .switched_to
= switched_to_rt
,