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)) {
38 rq
->rt
.overloaded
= 1;
40 rt_clear_overload(rq
);
41 rq
->rt
.overloaded
= 0;
44 #endif /* CONFIG_SMP */
47 * Update the current task's runtime statistics. Skip current tasks that
48 * are not in our scheduling class.
50 static void update_curr_rt(struct rq
*rq
)
52 struct task_struct
*curr
= rq
->curr
;
55 if (!task_has_rt_policy(curr
))
58 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
59 if (unlikely((s64
)delta_exec
< 0))
62 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
64 curr
->se
.sum_exec_runtime
+= delta_exec
;
65 curr
->se
.exec_start
= rq
->clock
;
66 cpuacct_charge(curr
, delta_exec
);
69 static inline void inc_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
72 rq
->rt
.rt_nr_running
++;
74 if (p
->prio
< rq
->rt
.highest_prio
)
75 rq
->rt
.highest_prio
= p
->prio
;
76 if (p
->nr_cpus_allowed
> 1)
77 rq
->rt
.rt_nr_migratory
++;
79 update_rt_migration(rq
);
80 #endif /* CONFIG_SMP */
83 static inline void dec_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
86 WARN_ON(!rq
->rt
.rt_nr_running
);
87 rq
->rt
.rt_nr_running
--;
89 if (rq
->rt
.rt_nr_running
) {
90 struct rt_prio_array
*array
;
92 WARN_ON(p
->prio
< rq
->rt
.highest_prio
);
93 if (p
->prio
== rq
->rt
.highest_prio
) {
95 array
= &rq
->rt
.active
;
97 sched_find_first_bit(array
->bitmap
);
98 } /* otherwise leave rq->highest prio alone */
100 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
101 if (p
->nr_cpus_allowed
> 1)
102 rq
->rt
.rt_nr_migratory
--;
104 update_rt_migration(rq
);
105 #endif /* CONFIG_SMP */
108 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
110 struct rt_prio_array
*array
= &rq
->rt
.active
;
112 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
113 __set_bit(p
->prio
, array
->bitmap
);
114 inc_cpu_load(rq
, p
->se
.load
.weight
);
120 * Adding/removing a task to/from a priority array:
122 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
124 struct rt_prio_array
*array
= &rq
->rt
.active
;
128 list_del(&p
->run_list
);
129 if (list_empty(array
->queue
+ p
->prio
))
130 __clear_bit(p
->prio
, array
->bitmap
);
131 dec_cpu_load(rq
, p
->se
.load
.weight
);
137 * Put task to the end of the run list without the overhead of dequeue
138 * followed by enqueue.
140 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
142 struct rt_prio_array
*array
= &rq
->rt
.active
;
144 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
148 yield_task_rt(struct rq
*rq
)
150 requeue_task_rt(rq
, rq
->curr
);
154 static int find_lowest_rq(struct task_struct
*task
);
156 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
158 struct rq
*rq
= task_rq(p
);
161 * If the current task is an RT task, then
162 * try to see if we can wake this RT task up on another
163 * runqueue. Otherwise simply start this RT task
164 * on its current runqueue.
166 * We want to avoid overloading runqueues. Even if
167 * the RT task is of higher priority than the current RT task.
168 * RT tasks behave differently than other tasks. If
169 * one gets preempted, we try to push it off to another queue.
170 * So trying to keep a preempting RT task on the same
171 * cache hot CPU will force the running RT task to
172 * a cold CPU. So we waste all the cache for the lower
173 * RT task in hopes of saving some of a RT task
174 * that is just being woken and probably will have
177 if (unlikely(rt_task(rq
->curr
)) &&
178 (p
->nr_cpus_allowed
> 1)) {
179 int cpu
= find_lowest_rq(p
);
181 return (cpu
== -1) ? task_cpu(p
) : cpu
;
185 * Otherwise, just let it ride on the affined RQ and the
186 * post-schedule router will push the preempted task away
190 #endif /* CONFIG_SMP */
193 * Preempt the current task with a newly woken task if needed:
195 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
197 if (p
->prio
< rq
->curr
->prio
)
198 resched_task(rq
->curr
);
201 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
203 struct rt_prio_array
*array
= &rq
->rt
.active
;
204 struct task_struct
*next
;
205 struct list_head
*queue
;
208 idx
= sched_find_first_bit(array
->bitmap
);
209 if (idx
>= MAX_RT_PRIO
)
212 queue
= array
->queue
+ idx
;
213 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
215 next
->se
.exec_start
= rq
->clock
;
220 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
223 p
->se
.exec_start
= 0;
227 /* Only try algorithms three times */
228 #define RT_MAX_TRIES 3
230 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
231 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
233 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
235 if (!task_running(rq
, p
) &&
236 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
237 (p
->nr_cpus_allowed
> 1))
242 /* Return the second highest RT task, NULL otherwise */
243 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
245 struct rt_prio_array
*array
= &rq
->rt
.active
;
246 struct task_struct
*next
;
247 struct list_head
*queue
;
250 if (likely(rq
->rt
.rt_nr_running
< 2))
253 idx
= sched_find_first_bit(array
->bitmap
);
254 if (unlikely(idx
>= MAX_RT_PRIO
)) {
255 WARN_ON(1); /* rt_nr_running is bad */
259 queue
= array
->queue
+ idx
;
260 BUG_ON(list_empty(queue
));
262 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
263 if (unlikely(pick_rt_task(rq
, next
, cpu
)))
266 if (queue
->next
->next
!= queue
) {
268 next
= list_entry(queue
->next
->next
, struct task_struct
,
270 if (pick_rt_task(rq
, next
, cpu
))
275 /* slower, but more flexible */
276 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
277 if (unlikely(idx
>= MAX_RT_PRIO
))
280 queue
= array
->queue
+ idx
;
281 BUG_ON(list_empty(queue
));
283 list_for_each_entry(next
, queue
, run_list
) {
284 if (pick_rt_task(rq
, next
, cpu
))
294 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
296 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
298 int lowest_prio
= -1;
303 cpus_and(*lowest_mask
, task_rq(task
)->rd
->online
, task
->cpus_allowed
);
306 * Scan each rq for the lowest prio.
308 for_each_cpu_mask(cpu
, *lowest_mask
) {
309 struct rq
*rq
= cpu_rq(cpu
);
311 /* We look for lowest RT prio or non-rt CPU */
312 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
314 * if we already found a low RT queue
315 * and now we found this non-rt queue
316 * clear the mask and set our bit.
317 * Otherwise just return the queue as is
318 * and the count==1 will cause the algorithm
319 * to use the first bit found.
321 if (lowest_cpu
!= -1) {
322 cpus_clear(*lowest_mask
);
323 cpu_set(rq
->cpu
, *lowest_mask
);
328 /* no locking for now */
329 if ((rq
->rt
.highest_prio
> task
->prio
)
330 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
331 if (rq
->rt
.highest_prio
> lowest_prio
) {
332 /* new low - clear old data */
333 lowest_prio
= rq
->rt
.highest_prio
;
339 cpu_clear(cpu
, *lowest_mask
);
343 * Clear out all the set bits that represent
344 * runqueues that were of higher prio than
347 if (lowest_cpu
> 0) {
349 * Perhaps we could add another cpumask op to
350 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
351 * Then that could be optimized to use memset and such.
353 for_each_cpu_mask(cpu
, *lowest_mask
) {
354 if (cpu
>= lowest_cpu
)
356 cpu_clear(cpu
, *lowest_mask
);
363 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
367 /* "this_cpu" is cheaper to preempt than a remote processor */
368 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
371 first
= first_cpu(*mask
);
372 if (first
!= NR_CPUS
)
378 static int find_lowest_rq(struct task_struct
*task
)
380 struct sched_domain
*sd
;
381 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
382 int this_cpu
= smp_processor_id();
383 int cpu
= task_cpu(task
);
384 int count
= find_lowest_cpus(task
, lowest_mask
);
387 return -1; /* No targets found */
390 * There is no sense in performing an optimal search if only one
394 return first_cpu(*lowest_mask
);
397 * At this point we have built a mask of cpus representing the
398 * lowest priority tasks in the system. Now we want to elect
399 * the best one based on our affinity and topology.
401 * We prioritize the last cpu that the task executed on since
402 * it is most likely cache-hot in that location.
404 if (cpu_isset(cpu
, *lowest_mask
))
408 * Otherwise, we consult the sched_domains span maps to figure
409 * out which cpu is logically closest to our hot cache data.
412 this_cpu
= -1; /* Skip this_cpu opt if the same */
414 for_each_domain(cpu
, sd
) {
415 if (sd
->flags
& SD_WAKE_AFFINE
) {
416 cpumask_t domain_mask
;
419 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
421 best_cpu
= pick_optimal_cpu(this_cpu
,
429 * And finally, if there were no matches within the domains
430 * just give the caller *something* to work with from the compatible
433 return pick_optimal_cpu(this_cpu
, lowest_mask
);
436 /* Will lock the rq it finds */
437 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
439 struct rq
*lowest_rq
= NULL
;
443 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
444 cpu
= find_lowest_rq(task
);
446 if ((cpu
== -1) || (cpu
== rq
->cpu
))
449 lowest_rq
= cpu_rq(cpu
);
451 /* if the prio of this runqueue changed, try again */
452 if (double_lock_balance(rq
, lowest_rq
)) {
454 * We had to unlock the run queue. In
455 * the mean time, task could have
456 * migrated already or had its affinity changed.
457 * Also make sure that it wasn't scheduled on its rq.
459 if (unlikely(task_rq(task
) != rq
||
460 !cpu_isset(lowest_rq
->cpu
,
461 task
->cpus_allowed
) ||
462 task_running(rq
, task
) ||
465 spin_unlock(&lowest_rq
->lock
);
471 /* If this rq is still suitable use it. */
472 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
476 spin_unlock(&lowest_rq
->lock
);
484 * If the current CPU has more than one RT task, see if the non
485 * running task can migrate over to a CPU that is running a task
486 * of lesser priority.
488 static int push_rt_task(struct rq
*rq
)
490 struct task_struct
*next_task
;
491 struct rq
*lowest_rq
;
493 int paranoid
= RT_MAX_TRIES
;
495 if (!rq
->rt
.overloaded
)
498 next_task
= pick_next_highest_task_rt(rq
, -1);
503 if (unlikely(next_task
== rq
->curr
)) {
509 * It's possible that the next_task slipped in of
510 * higher priority than current. If that's the case
511 * just reschedule current.
513 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
514 resched_task(rq
->curr
);
518 /* We might release rq lock */
519 get_task_struct(next_task
);
521 /* find_lock_lowest_rq locks the rq if found */
522 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
524 struct task_struct
*task
;
526 * find lock_lowest_rq releases rq->lock
527 * so it is possible that next_task has changed.
528 * If it has, then try again.
530 task
= pick_next_highest_task_rt(rq
, -1);
531 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
532 put_task_struct(next_task
);
539 deactivate_task(rq
, next_task
, 0);
540 set_task_cpu(next_task
, lowest_rq
->cpu
);
541 activate_task(lowest_rq
, next_task
, 0);
543 resched_task(lowest_rq
->curr
);
545 spin_unlock(&lowest_rq
->lock
);
549 put_task_struct(next_task
);
555 * TODO: Currently we just use the second highest prio task on
556 * the queue, and stop when it can't migrate (or there's
557 * no more RT tasks). There may be a case where a lower
558 * priority RT task has a different affinity than the
559 * higher RT task. In this case the lower RT task could
560 * possibly be able to migrate where as the higher priority
561 * RT task could not. We currently ignore this issue.
562 * Enhancements are welcome!
564 static void push_rt_tasks(struct rq
*rq
)
566 /* push_rt_task will return true if it moved an RT */
567 while (push_rt_task(rq
))
571 static int pull_rt_task(struct rq
*this_rq
)
573 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
574 struct task_struct
*p
, *next
;
577 if (likely(!rt_overloaded(this_rq
)))
580 next
= pick_next_task_rt(this_rq
);
582 for_each_cpu_mask(cpu
, this_rq
->rd
->rto_mask
) {
586 src_rq
= cpu_rq(cpu
);
587 if (unlikely(src_rq
->rt
.rt_nr_running
<= 1)) {
589 * It is possible that overlapping cpusets
590 * will miss clearing a non overloaded runqueue.
593 if (double_lock_balance(this_rq
, src_rq
)) {
594 /* unlocked our runqueue lock */
595 struct task_struct
*old_next
= next
;
597 next
= pick_next_task_rt(this_rq
);
598 if (next
!= old_next
)
601 if (likely(src_rq
->rt
.rt_nr_running
<= 1)) {
603 * Small chance that this_rq->curr changed
604 * but it's really harmless here.
606 rt_clear_overload(this_rq
);
609 * Heh, the src_rq is now overloaded, since
610 * we already have the src_rq lock, go straight
611 * to pulling tasks from it.
615 spin_unlock(&src_rq
->lock
);
620 * We can potentially drop this_rq's lock in
621 * double_lock_balance, and another CPU could
622 * steal our next task - hence we must cause
623 * the caller to recalculate the next task
626 if (double_lock_balance(this_rq
, src_rq
)) {
627 struct task_struct
*old_next
= next
;
629 next
= pick_next_task_rt(this_rq
);
630 if (next
!= old_next
)
635 * Are there still pullable RT tasks?
637 if (src_rq
->rt
.rt_nr_running
<= 1) {
638 spin_unlock(&src_rq
->lock
);
643 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
646 * Do we have an RT task that preempts
647 * the to-be-scheduled task?
649 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
650 WARN_ON(p
== src_rq
->curr
);
651 WARN_ON(!p
->se
.on_rq
);
654 * There's a chance that p is higher in priority
655 * than what's currently running on its cpu.
656 * This is just that p is wakeing up and hasn't
657 * had a chance to schedule. We only pull
658 * p if it is lower in priority than the
659 * current task on the run queue or
660 * this_rq next task is lower in prio than
661 * the current task on that rq.
663 if (p
->prio
< src_rq
->curr
->prio
||
664 (next
&& next
->prio
< src_rq
->curr
->prio
))
669 deactivate_task(src_rq
, p
, 0);
670 set_task_cpu(p
, this_cpu
);
671 activate_task(this_rq
, p
, 0);
673 * We continue with the search, just in
674 * case there's an even higher prio task
675 * in another runqueue. (low likelyhood
678 * Update next so that we won't pick a task
679 * on another cpu with a priority lower (or equal)
680 * than the one we just picked.
686 spin_unlock(&src_rq
->lock
);
692 static void pre_schedule_rt(struct rq
*rq
, struct task_struct
*prev
)
694 /* Try to pull RT tasks here if we lower this rq's prio */
695 if (unlikely(rt_task(prev
)) && rq
->rt
.highest_prio
> prev
->prio
)
699 static void post_schedule_rt(struct rq
*rq
)
702 * If we have more than one rt_task queued, then
703 * see if we can push the other rt_tasks off to other CPUS.
704 * Note we may release the rq lock, and since
705 * the lock was owned by prev, we need to release it
706 * first via finish_lock_switch and then reaquire it here.
708 if (unlikely(rq
->rt
.overloaded
)) {
709 spin_lock_irq(&rq
->lock
);
711 spin_unlock_irq(&rq
->lock
);
716 static void task_wake_up_rt(struct rq
*rq
, struct task_struct
*p
)
718 if (!task_running(rq
, p
) &&
719 (p
->prio
>= rq
->rt
.highest_prio
) &&
725 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
726 unsigned long max_load_move
,
727 struct sched_domain
*sd
, enum cpu_idle_type idle
,
728 int *all_pinned
, int *this_best_prio
)
730 /* don't touch RT tasks */
735 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
736 struct sched_domain
*sd
, enum cpu_idle_type idle
)
738 /* don't touch RT tasks */
742 static void set_cpus_allowed_rt(struct task_struct
*p
, cpumask_t
*new_mask
)
744 int weight
= cpus_weight(*new_mask
);
749 * Update the migration status of the RQ if we have an RT task
750 * which is running AND changing its weight value.
752 if (p
->se
.on_rq
&& (weight
!= p
->nr_cpus_allowed
)) {
753 struct rq
*rq
= task_rq(p
);
755 if ((p
->nr_cpus_allowed
<= 1) && (weight
> 1)) {
756 rq
->rt
.rt_nr_migratory
++;
757 } else if ((p
->nr_cpus_allowed
> 1) && (weight
<= 1)) {
758 BUG_ON(!rq
->rt
.rt_nr_migratory
);
759 rq
->rt
.rt_nr_migratory
--;
762 update_rt_migration(rq
);
765 p
->cpus_allowed
= *new_mask
;
766 p
->nr_cpus_allowed
= weight
;
769 /* Assumes rq->lock is held */
770 static void join_domain_rt(struct rq
*rq
)
772 if (rq
->rt
.overloaded
)
776 /* Assumes rq->lock is held */
777 static void leave_domain_rt(struct rq
*rq
)
779 if (rq
->rt
.overloaded
)
780 rt_clear_overload(rq
);
782 #endif /* CONFIG_SMP */
784 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
)
789 * RR tasks need a special form of timeslice management.
790 * FIFO tasks have no timeslices.
792 if (p
->policy
!= SCHED_RR
)
798 p
->time_slice
= DEF_TIMESLICE
;
801 * Requeue to the end of queue if we are not the only element
804 if (p
->run_list
.prev
!= p
->run_list
.next
) {
805 requeue_task_rt(rq
, p
);
806 set_tsk_need_resched(p
);
810 static void set_curr_task_rt(struct rq
*rq
)
812 struct task_struct
*p
= rq
->curr
;
814 p
->se
.exec_start
= rq
->clock
;
817 const struct sched_class rt_sched_class
= {
818 .next
= &fair_sched_class
,
819 .enqueue_task
= enqueue_task_rt
,
820 .dequeue_task
= dequeue_task_rt
,
821 .yield_task
= yield_task_rt
,
823 .select_task_rq
= select_task_rq_rt
,
824 #endif /* CONFIG_SMP */
826 .check_preempt_curr
= check_preempt_curr_rt
,
828 .pick_next_task
= pick_next_task_rt
,
829 .put_prev_task
= put_prev_task_rt
,
832 .load_balance
= load_balance_rt
,
833 .move_one_task
= move_one_task_rt
,
834 .set_cpus_allowed
= set_cpus_allowed_rt
,
835 .join_domain
= join_domain_rt
,
836 .leave_domain
= leave_domain_rt
,
837 .pre_schedule
= pre_schedule_rt
,
838 .post_schedule
= post_schedule_rt
,
839 .task_wake_up
= task_wake_up_rt
,
842 .set_curr_task
= set_curr_task_rt
,
843 .task_tick
= task_tick_rt
,