2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
9 * The "RT overload" flag: it gets set if a CPU has more than
10 * one runnable RT task.
12 static cpumask_t rt_overload_mask
;
13 static atomic_t rto_count
;
15 static inline int rt_overloaded(void)
17 return atomic_read(&rto_count
);
20 static inline cpumask_t
*rt_overload(void)
22 return &rt_overload_mask
;
25 static inline void rt_set_overload(struct rq
*rq
)
27 rq
->rt
.overloaded
= 1;
28 cpu_set(rq
->cpu
, rt_overload_mask
);
30 * Make sure the mask is visible before we set
31 * the overload count. That is checked to determine
32 * if we should look at the mask. It would be a shame
33 * if we looked at the mask, but the mask was not
37 atomic_inc(&rto_count
);
40 static inline void rt_clear_overload(struct rq
*rq
)
42 /* the order here really doesn't matter */
43 atomic_dec(&rto_count
);
44 cpu_clear(rq
->cpu
, rt_overload_mask
);
45 rq
->rt
.overloaded
= 0;
48 static void update_rt_migration(struct rq
*rq
)
50 if (rq
->rt
.rt_nr_migratory
&& (rq
->rt
.rt_nr_running
> 1))
53 rt_clear_overload(rq
);
55 #endif /* CONFIG_SMP */
58 * Update the current task's runtime statistics. Skip current tasks that
59 * are not in our scheduling class.
61 static void update_curr_rt(struct rq
*rq
)
63 struct task_struct
*curr
= rq
->curr
;
66 if (!task_has_rt_policy(curr
))
69 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
70 if (unlikely((s64
)delta_exec
< 0))
73 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
75 curr
->se
.sum_exec_runtime
+= delta_exec
;
76 curr
->se
.exec_start
= rq
->clock
;
77 cpuacct_charge(curr
, delta_exec
);
80 static inline void inc_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
83 rq
->rt
.rt_nr_running
++;
85 if (p
->prio
< rq
->rt
.highest_prio
)
86 rq
->rt
.highest_prio
= p
->prio
;
87 if (p
->nr_cpus_allowed
> 1)
88 rq
->rt
.rt_nr_migratory
++;
90 update_rt_migration(rq
);
91 #endif /* CONFIG_SMP */
94 static inline void dec_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
97 WARN_ON(!rq
->rt
.rt_nr_running
);
98 rq
->rt
.rt_nr_running
--;
100 if (rq
->rt
.rt_nr_running
) {
101 struct rt_prio_array
*array
;
103 WARN_ON(p
->prio
< rq
->rt
.highest_prio
);
104 if (p
->prio
== rq
->rt
.highest_prio
) {
106 array
= &rq
->rt
.active
;
107 rq
->rt
.highest_prio
=
108 sched_find_first_bit(array
->bitmap
);
109 } /* otherwise leave rq->highest prio alone */
111 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
112 if (p
->nr_cpus_allowed
> 1)
113 rq
->rt
.rt_nr_migratory
--;
115 update_rt_migration(rq
);
116 #endif /* CONFIG_SMP */
119 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
121 struct rt_prio_array
*array
= &rq
->rt
.active
;
123 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
124 __set_bit(p
->prio
, array
->bitmap
);
125 inc_cpu_load(rq
, p
->se
.load
.weight
);
131 * Adding/removing a task to/from a priority array:
133 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
135 struct rt_prio_array
*array
= &rq
->rt
.active
;
139 list_del(&p
->run_list
);
140 if (list_empty(array
->queue
+ p
->prio
))
141 __clear_bit(p
->prio
, array
->bitmap
);
142 dec_cpu_load(rq
, p
->se
.load
.weight
);
148 * Put task to the end of the run list without the overhead of dequeue
149 * followed by enqueue.
151 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
153 struct rt_prio_array
*array
= &rq
->rt
.active
;
155 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
159 yield_task_rt(struct rq
*rq
)
161 requeue_task_rt(rq
, rq
->curr
);
165 static int find_lowest_rq(struct task_struct
*task
);
167 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
169 struct rq
*rq
= task_rq(p
);
172 * If the current task is an RT task, then
173 * try to see if we can wake this RT task up on another
174 * runqueue. Otherwise simply start this RT task
175 * on its current runqueue.
177 * We want to avoid overloading runqueues. Even if
178 * the RT task is of higher priority than the current RT task.
179 * RT tasks behave differently than other tasks. If
180 * one gets preempted, we try to push it off to another queue.
181 * So trying to keep a preempting RT task on the same
182 * cache hot CPU will force the running RT task to
183 * a cold CPU. So we waste all the cache for the lower
184 * RT task in hopes of saving some of a RT task
185 * that is just being woken and probably will have
188 if (unlikely(rt_task(rq
->curr
)) &&
189 (p
->nr_cpus_allowed
> 1)) {
190 int cpu
= find_lowest_rq(p
);
192 return (cpu
== -1) ? task_cpu(p
) : cpu
;
196 * Otherwise, just let it ride on the affined RQ and the
197 * post-schedule router will push the preempted task away
201 #endif /* CONFIG_SMP */
204 * Preempt the current task with a newly woken task if needed:
206 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
208 if (p
->prio
< rq
->curr
->prio
)
209 resched_task(rq
->curr
);
212 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
214 struct rt_prio_array
*array
= &rq
->rt
.active
;
215 struct task_struct
*next
;
216 struct list_head
*queue
;
219 idx
= sched_find_first_bit(array
->bitmap
);
220 if (idx
>= MAX_RT_PRIO
)
223 queue
= array
->queue
+ idx
;
224 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
226 next
->se
.exec_start
= rq
->clock
;
231 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
234 p
->se
.exec_start
= 0;
238 /* Only try algorithms three times */
239 #define RT_MAX_TRIES 3
241 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
242 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
244 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
246 if (!task_running(rq
, p
) &&
247 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
248 (p
->nr_cpus_allowed
> 1))
253 /* Return the second highest RT task, NULL otherwise */
254 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
256 struct rt_prio_array
*array
= &rq
->rt
.active
;
257 struct task_struct
*next
;
258 struct list_head
*queue
;
261 assert_spin_locked(&rq
->lock
);
263 if (likely(rq
->rt
.rt_nr_running
< 2))
266 idx
= sched_find_first_bit(array
->bitmap
);
267 if (unlikely(idx
>= MAX_RT_PRIO
)) {
268 WARN_ON(1); /* rt_nr_running is bad */
272 queue
= array
->queue
+ idx
;
273 BUG_ON(list_empty(queue
));
275 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
276 if (unlikely(pick_rt_task(rq
, next
, cpu
)))
279 if (queue
->next
->next
!= queue
) {
281 next
= list_entry(queue
->next
->next
, struct task_struct
,
283 if (pick_rt_task(rq
, next
, cpu
))
288 /* slower, but more flexible */
289 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
290 if (unlikely(idx
>= MAX_RT_PRIO
))
293 queue
= array
->queue
+ idx
;
294 BUG_ON(list_empty(queue
));
296 list_for_each_entry(next
, queue
, run_list
) {
297 if (pick_rt_task(rq
, next
, cpu
))
307 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
309 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
311 int lowest_prio
= -1;
316 cpus_and(*lowest_mask
, cpu_online_map
, task
->cpus_allowed
);
319 * Scan each rq for the lowest prio.
321 for_each_cpu_mask(cpu
, *lowest_mask
) {
322 struct rq
*rq
= cpu_rq(cpu
);
324 /* We look for lowest RT prio or non-rt CPU */
325 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
327 * if we already found a low RT queue
328 * and now we found this non-rt queue
329 * clear the mask and set our bit.
330 * Otherwise just return the queue as is
331 * and the count==1 will cause the algorithm
332 * to use the first bit found.
334 if (lowest_cpu
!= -1) {
335 cpus_clear(*lowest_mask
);
336 cpu_set(rq
->cpu
, *lowest_mask
);
341 /* no locking for now */
342 if ((rq
->rt
.highest_prio
> task
->prio
)
343 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
344 if (rq
->rt
.highest_prio
> lowest_prio
) {
345 /* new low - clear old data */
346 lowest_prio
= rq
->rt
.highest_prio
;
352 cpu_clear(cpu
, *lowest_mask
);
356 * Clear out all the set bits that represent
357 * runqueues that were of higher prio than
360 if (lowest_cpu
> 0) {
362 * Perhaps we could add another cpumask op to
363 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
364 * Then that could be optimized to use memset and such.
366 for_each_cpu_mask(cpu
, *lowest_mask
) {
367 if (cpu
>= lowest_cpu
)
369 cpu_clear(cpu
, *lowest_mask
);
376 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
380 /* "this_cpu" is cheaper to preempt than a remote processor */
381 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
384 first
= first_cpu(*mask
);
385 if (first
!= NR_CPUS
)
391 static int find_lowest_rq(struct task_struct
*task
)
393 struct sched_domain
*sd
;
394 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
395 int this_cpu
= smp_processor_id();
396 int cpu
= task_cpu(task
);
397 int count
= find_lowest_cpus(task
, lowest_mask
);
400 return -1; /* No targets found */
403 * There is no sense in performing an optimal search if only one
407 return first_cpu(*lowest_mask
);
410 * At this point we have built a mask of cpus representing the
411 * lowest priority tasks in the system. Now we want to elect
412 * the best one based on our affinity and topology.
414 * We prioritize the last cpu that the task executed on since
415 * it is most likely cache-hot in that location.
417 if (cpu_isset(cpu
, *lowest_mask
))
421 * Otherwise, we consult the sched_domains span maps to figure
422 * out which cpu is logically closest to our hot cache data.
425 this_cpu
= -1; /* Skip this_cpu opt if the same */
427 for_each_domain(cpu
, sd
) {
428 if (sd
->flags
& SD_WAKE_AFFINE
) {
429 cpumask_t domain_mask
;
432 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
434 best_cpu
= pick_optimal_cpu(this_cpu
,
442 * And finally, if there were no matches within the domains
443 * just give the caller *something* to work with from the compatible
446 return pick_optimal_cpu(this_cpu
, lowest_mask
);
449 /* Will lock the rq it finds */
450 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
452 struct rq
*lowest_rq
= NULL
;
456 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
457 cpu
= find_lowest_rq(task
);
459 if ((cpu
== -1) || (cpu
== rq
->cpu
))
462 lowest_rq
= cpu_rq(cpu
);
464 /* if the prio of this runqueue changed, try again */
465 if (double_lock_balance(rq
, lowest_rq
)) {
467 * We had to unlock the run queue. In
468 * the mean time, task could have
469 * migrated already or had its affinity changed.
470 * Also make sure that it wasn't scheduled on its rq.
472 if (unlikely(task_rq(task
) != rq
||
473 !cpu_isset(lowest_rq
->cpu
,
474 task
->cpus_allowed
) ||
475 task_running(rq
, task
) ||
478 spin_unlock(&lowest_rq
->lock
);
484 /* If this rq is still suitable use it. */
485 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
489 spin_unlock(&lowest_rq
->lock
);
497 * If the current CPU has more than one RT task, see if the non
498 * running task can migrate over to a CPU that is running a task
499 * of lesser priority.
501 static int push_rt_task(struct rq
*rq
)
503 struct task_struct
*next_task
;
504 struct rq
*lowest_rq
;
506 int paranoid
= RT_MAX_TRIES
;
508 assert_spin_locked(&rq
->lock
);
510 if (!rq
->rt
.overloaded
)
513 next_task
= pick_next_highest_task_rt(rq
, -1);
518 if (unlikely(next_task
== rq
->curr
)) {
524 * It's possible that the next_task slipped in of
525 * higher priority than current. If that's the case
526 * just reschedule current.
528 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
529 resched_task(rq
->curr
);
533 /* We might release rq lock */
534 get_task_struct(next_task
);
536 /* find_lock_lowest_rq locks the rq if found */
537 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
539 struct task_struct
*task
;
541 * find lock_lowest_rq releases rq->lock
542 * so it is possible that next_task has changed.
543 * If it has, then try again.
545 task
= pick_next_highest_task_rt(rq
, -1);
546 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
547 put_task_struct(next_task
);
554 assert_spin_locked(&lowest_rq
->lock
);
556 deactivate_task(rq
, next_task
, 0);
557 set_task_cpu(next_task
, lowest_rq
->cpu
);
558 activate_task(lowest_rq
, next_task
, 0);
560 resched_task(lowest_rq
->curr
);
562 spin_unlock(&lowest_rq
->lock
);
566 put_task_struct(next_task
);
572 * TODO: Currently we just use the second highest prio task on
573 * the queue, and stop when it can't migrate (or there's
574 * no more RT tasks). There may be a case where a lower
575 * priority RT task has a different affinity than the
576 * higher RT task. In this case the lower RT task could
577 * possibly be able to migrate where as the higher priority
578 * RT task could not. We currently ignore this issue.
579 * Enhancements are welcome!
581 static void push_rt_tasks(struct rq
*rq
)
583 /* push_rt_task will return true if it moved an RT */
584 while (push_rt_task(rq
))
588 static int pull_rt_task(struct rq
*this_rq
)
590 struct task_struct
*next
;
591 struct task_struct
*p
;
593 cpumask_t
*rto_cpumask
;
594 int this_cpu
= this_rq
->cpu
;
598 assert_spin_locked(&this_rq
->lock
);
601 * If cpusets are used, and we have overlapping
602 * run queue cpusets, then this algorithm may not catch all.
603 * This is just the price you pay on trying to keep
604 * dirtying caches down on large SMP machines.
606 if (likely(!rt_overloaded()))
609 next
= pick_next_task_rt(this_rq
);
611 rto_cpumask
= rt_overload();
613 for_each_cpu_mask(cpu
, *rto_cpumask
) {
617 src_rq
= cpu_rq(cpu
);
618 if (unlikely(src_rq
->rt
.rt_nr_running
<= 1)) {
620 * It is possible that overlapping cpusets
621 * will miss clearing a non overloaded runqueue.
624 if (double_lock_balance(this_rq
, src_rq
)) {
625 /* unlocked our runqueue lock */
626 struct task_struct
*old_next
= next
;
627 next
= pick_next_task_rt(this_rq
);
628 if (next
!= old_next
)
631 if (likely(src_rq
->rt
.rt_nr_running
<= 1))
633 * Small chance that this_rq->curr changed
634 * but it's really harmless here.
636 rt_clear_overload(this_rq
);
639 * Heh, the src_rq is now overloaded, since
640 * we already have the src_rq lock, go straight
641 * to pulling tasks from it.
644 spin_unlock(&src_rq
->lock
);
649 * We can potentially drop this_rq's lock in
650 * double_lock_balance, and another CPU could
651 * steal our next task - hence we must cause
652 * the caller to recalculate the next task
655 if (double_lock_balance(this_rq
, src_rq
)) {
656 struct task_struct
*old_next
= next
;
657 next
= pick_next_task_rt(this_rq
);
658 if (next
!= old_next
)
663 * Are there still pullable RT tasks?
665 if (src_rq
->rt
.rt_nr_running
<= 1) {
666 spin_unlock(&src_rq
->lock
);
671 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
674 * Do we have an RT task that preempts
675 * the to-be-scheduled task?
677 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
678 WARN_ON(p
== src_rq
->curr
);
679 WARN_ON(!p
->se
.on_rq
);
682 * There's a chance that p is higher in priority
683 * than what's currently running on its cpu.
684 * This is just that p is wakeing up and hasn't
685 * had a chance to schedule. We only pull
686 * p if it is lower in priority than the
687 * current task on the run queue or
688 * this_rq next task is lower in prio than
689 * the current task on that rq.
691 if (p
->prio
< src_rq
->curr
->prio
||
692 (next
&& next
->prio
< src_rq
->curr
->prio
))
697 deactivate_task(src_rq
, p
, 0);
698 set_task_cpu(p
, this_cpu
);
699 activate_task(this_rq
, p
, 0);
701 * We continue with the search, just in
702 * case there's an even higher prio task
703 * in another runqueue. (low likelyhood
708 * Update next so that we won't pick a task
709 * on another cpu with a priority lower (or equal)
710 * than the one we just picked.
716 spin_unlock(&src_rq
->lock
);
722 static void schedule_balance_rt(struct rq
*rq
,
723 struct task_struct
*prev
)
725 /* Try to pull RT tasks here if we lower this rq's prio */
726 if (unlikely(rt_task(prev
)) &&
727 rq
->rt
.highest_prio
> prev
->prio
)
731 static void schedule_tail_balance_rt(struct rq
*rq
)
734 * If we have more than one rt_task queued, then
735 * see if we can push the other rt_tasks off to other CPUS.
736 * Note we may release the rq lock, and since
737 * the lock was owned by prev, we need to release it
738 * first via finish_lock_switch and then reaquire it here.
740 if (unlikely(rq
->rt
.overloaded
)) {
741 spin_lock_irq(&rq
->lock
);
743 spin_unlock_irq(&rq
->lock
);
748 static void wakeup_balance_rt(struct rq
*rq
, struct task_struct
*p
)
750 if (unlikely(rt_task(p
)) &&
751 !task_running(rq
, p
) &&
752 (p
->prio
>= rq
->rt
.highest_prio
) &&
758 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
759 unsigned long max_load_move
,
760 struct sched_domain
*sd
, enum cpu_idle_type idle
,
761 int *all_pinned
, int *this_best_prio
)
763 /* don't touch RT tasks */
768 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
769 struct sched_domain
*sd
, enum cpu_idle_type idle
)
771 /* don't touch RT tasks */
775 static void set_cpus_allowed_rt(struct task_struct
*p
, cpumask_t
*new_mask
)
777 int weight
= cpus_weight(*new_mask
);
782 * Update the migration status of the RQ if we have an RT task
783 * which is running AND changing its weight value.
785 if (p
->se
.on_rq
&& (weight
!= p
->nr_cpus_allowed
)) {
786 struct rq
*rq
= task_rq(p
);
788 if ((p
->nr_cpus_allowed
<= 1) && (weight
> 1)) {
789 rq
->rt
.rt_nr_migratory
++;
790 } else if ((p
->nr_cpus_allowed
> 1) && (weight
<= 1)) {
791 BUG_ON(!rq
->rt
.rt_nr_migratory
);
792 rq
->rt
.rt_nr_migratory
--;
795 update_rt_migration(rq
);
798 p
->cpus_allowed
= *new_mask
;
799 p
->nr_cpus_allowed
= weight
;
802 #else /* CONFIG_SMP */
803 # define schedule_tail_balance_rt(rq) do { } while (0)
804 # define schedule_balance_rt(rq, prev) do { } while (0)
805 # define wakeup_balance_rt(rq, p) do { } while (0)
806 #endif /* CONFIG_SMP */
808 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
)
813 * RR tasks need a special form of timeslice management.
814 * FIFO tasks have no timeslices.
816 if (p
->policy
!= SCHED_RR
)
822 p
->time_slice
= DEF_TIMESLICE
;
825 * Requeue to the end of queue if we are not the only element
828 if (p
->run_list
.prev
!= p
->run_list
.next
) {
829 requeue_task_rt(rq
, p
);
830 set_tsk_need_resched(p
);
834 static void set_curr_task_rt(struct rq
*rq
)
836 struct task_struct
*p
= rq
->curr
;
838 p
->se
.exec_start
= rq
->clock
;
841 const struct sched_class rt_sched_class
= {
842 .next
= &fair_sched_class
,
843 .enqueue_task
= enqueue_task_rt
,
844 .dequeue_task
= dequeue_task_rt
,
845 .yield_task
= yield_task_rt
,
847 .select_task_rq
= select_task_rq_rt
,
848 #endif /* CONFIG_SMP */
850 .check_preempt_curr
= check_preempt_curr_rt
,
852 .pick_next_task
= pick_next_task_rt
,
853 .put_prev_task
= put_prev_task_rt
,
856 .load_balance
= load_balance_rt
,
857 .move_one_task
= move_one_task_rt
,
858 .set_cpus_allowed
= set_cpus_allowed_rt
,
861 .set_curr_task
= set_curr_task_rt
,
862 .task_tick
= task_tick_rt
,