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 void rt_set_overload(struct rq
*rq
)
22 rq
->rt
.overloaded
= 1;
23 cpu_set(rq
->cpu
, rt_overload_mask
);
25 * Make sure the mask is visible before we set
26 * the overload count. That is checked to determine
27 * if we should look at the mask. It would be a shame
28 * if we looked at the mask, but the mask was not
32 atomic_inc(&rto_count
);
35 static inline void rt_clear_overload(struct rq
*rq
)
37 /* the order here really doesn't matter */
38 atomic_dec(&rto_count
);
39 cpu_clear(rq
->cpu
, rt_overload_mask
);
40 rq
->rt
.overloaded
= 0;
43 static void update_rt_migration(struct rq
*rq
)
45 if (rq
->rt
.rt_nr_migratory
&& (rq
->rt
.rt_nr_running
> 1))
48 rt_clear_overload(rq
);
50 #endif /* CONFIG_SMP */
53 * Update the current task's runtime statistics. Skip current tasks that
54 * are not in our scheduling class.
56 static void update_curr_rt(struct rq
*rq
)
58 struct task_struct
*curr
= rq
->curr
;
61 if (!task_has_rt_policy(curr
))
64 delta_exec
= rq
->clock
- curr
->se
.exec_start
;
65 if (unlikely((s64
)delta_exec
< 0))
68 schedstat_set(curr
->se
.exec_max
, max(curr
->se
.exec_max
, delta_exec
));
70 curr
->se
.sum_exec_runtime
+= delta_exec
;
71 curr
->se
.exec_start
= rq
->clock
;
72 cpuacct_charge(curr
, delta_exec
);
75 static inline void inc_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
78 rq
->rt
.rt_nr_running
++;
80 if (p
->prio
< rq
->rt
.highest_prio
)
81 rq
->rt
.highest_prio
= p
->prio
;
82 if (p
->nr_cpus_allowed
> 1)
83 rq
->rt
.rt_nr_migratory
++;
85 update_rt_migration(rq
);
86 #endif /* CONFIG_SMP */
89 static inline void dec_rt_tasks(struct task_struct
*p
, struct rq
*rq
)
92 WARN_ON(!rq
->rt
.rt_nr_running
);
93 rq
->rt
.rt_nr_running
--;
95 if (rq
->rt
.rt_nr_running
) {
96 struct rt_prio_array
*array
;
98 WARN_ON(p
->prio
< rq
->rt
.highest_prio
);
99 if (p
->prio
== rq
->rt
.highest_prio
) {
101 array
= &rq
->rt
.active
;
102 rq
->rt
.highest_prio
=
103 sched_find_first_bit(array
->bitmap
);
104 } /* otherwise leave rq->highest prio alone */
106 rq
->rt
.highest_prio
= MAX_RT_PRIO
;
107 if (p
->nr_cpus_allowed
> 1)
108 rq
->rt
.rt_nr_migratory
--;
110 update_rt_migration(rq
);
111 #endif /* CONFIG_SMP */
114 static void enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
116 struct rt_prio_array
*array
= &rq
->rt
.active
;
118 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
119 __set_bit(p
->prio
, array
->bitmap
);
120 inc_cpu_load(rq
, p
->se
.load
.weight
);
126 * Adding/removing a task to/from a priority array:
128 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int sleep
)
130 struct rt_prio_array
*array
= &rq
->rt
.active
;
134 list_del(&p
->run_list
);
135 if (list_empty(array
->queue
+ p
->prio
))
136 __clear_bit(p
->prio
, array
->bitmap
);
137 dec_cpu_load(rq
, p
->se
.load
.weight
);
143 * Put task to the end of the run list without the overhead of dequeue
144 * followed by enqueue.
146 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
)
148 struct rt_prio_array
*array
= &rq
->rt
.active
;
150 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
154 yield_task_rt(struct rq
*rq
)
156 requeue_task_rt(rq
, rq
->curr
);
160 static int find_lowest_rq(struct task_struct
*task
);
162 static int select_task_rq_rt(struct task_struct
*p
, int sync
)
164 struct rq
*rq
= task_rq(p
);
167 * If the current task is an RT task, then
168 * try to see if we can wake this RT task up on another
169 * runqueue. Otherwise simply start this RT task
170 * on its current runqueue.
172 * We want to avoid overloading runqueues. Even if
173 * the RT task is of higher priority than the current RT task.
174 * RT tasks behave differently than other tasks. If
175 * one gets preempted, we try to push it off to another queue.
176 * So trying to keep a preempting RT task on the same
177 * cache hot CPU will force the running RT task to
178 * a cold CPU. So we waste all the cache for the lower
179 * RT task in hopes of saving some of a RT task
180 * that is just being woken and probably will have
183 if (unlikely(rt_task(rq
->curr
)) &&
184 (p
->nr_cpus_allowed
> 1)) {
185 int cpu
= find_lowest_rq(p
);
187 return (cpu
== -1) ? task_cpu(p
) : cpu
;
191 * Otherwise, just let it ride on the affined RQ and the
192 * post-schedule router will push the preempted task away
196 #endif /* CONFIG_SMP */
199 * Preempt the current task with a newly woken task if needed:
201 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
203 if (p
->prio
< rq
->curr
->prio
)
204 resched_task(rq
->curr
);
207 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
209 struct rt_prio_array
*array
= &rq
->rt
.active
;
210 struct task_struct
*next
;
211 struct list_head
*queue
;
214 idx
= sched_find_first_bit(array
->bitmap
);
215 if (idx
>= MAX_RT_PRIO
)
218 queue
= array
->queue
+ idx
;
219 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
221 next
->se
.exec_start
= rq
->clock
;
226 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
229 p
->se
.exec_start
= 0;
233 /* Only try algorithms three times */
234 #define RT_MAX_TRIES 3
236 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
237 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
239 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
241 if (!task_running(rq
, p
) &&
242 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
243 (p
->nr_cpus_allowed
> 1))
248 /* Return the second highest RT task, NULL otherwise */
249 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
, int cpu
)
251 struct rt_prio_array
*array
= &rq
->rt
.active
;
252 struct task_struct
*next
;
253 struct list_head
*queue
;
256 if (likely(rq
->rt
.rt_nr_running
< 2))
259 idx
= sched_find_first_bit(array
->bitmap
);
260 if (unlikely(idx
>= MAX_RT_PRIO
)) {
261 WARN_ON(1); /* rt_nr_running is bad */
265 queue
= array
->queue
+ idx
;
266 BUG_ON(list_empty(queue
));
268 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
269 if (unlikely(pick_rt_task(rq
, next
, cpu
)))
272 if (queue
->next
->next
!= queue
) {
274 next
= list_entry(queue
->next
->next
, struct task_struct
,
276 if (pick_rt_task(rq
, next
, cpu
))
281 /* slower, but more flexible */
282 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
283 if (unlikely(idx
>= MAX_RT_PRIO
))
286 queue
= array
->queue
+ idx
;
287 BUG_ON(list_empty(queue
));
289 list_for_each_entry(next
, queue
, run_list
) {
290 if (pick_rt_task(rq
, next
, cpu
))
300 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
302 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
304 int lowest_prio
= -1;
309 cpus_and(*lowest_mask
, cpu_online_map
, task
->cpus_allowed
);
312 * Scan each rq for the lowest prio.
314 for_each_cpu_mask(cpu
, *lowest_mask
) {
315 struct rq
*rq
= cpu_rq(cpu
);
317 /* We look for lowest RT prio or non-rt CPU */
318 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
320 * if we already found a low RT queue
321 * and now we found this non-rt queue
322 * clear the mask and set our bit.
323 * Otherwise just return the queue as is
324 * and the count==1 will cause the algorithm
325 * to use the first bit found.
327 if (lowest_cpu
!= -1) {
328 cpus_clear(*lowest_mask
);
329 cpu_set(rq
->cpu
, *lowest_mask
);
334 /* no locking for now */
335 if ((rq
->rt
.highest_prio
> task
->prio
)
336 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
337 if (rq
->rt
.highest_prio
> lowest_prio
) {
338 /* new low - clear old data */
339 lowest_prio
= rq
->rt
.highest_prio
;
345 cpu_clear(cpu
, *lowest_mask
);
349 * Clear out all the set bits that represent
350 * runqueues that were of higher prio than
353 if (lowest_cpu
> 0) {
355 * Perhaps we could add another cpumask op to
356 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
357 * Then that could be optimized to use memset and such.
359 for_each_cpu_mask(cpu
, *lowest_mask
) {
360 if (cpu
>= lowest_cpu
)
362 cpu_clear(cpu
, *lowest_mask
);
369 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
373 /* "this_cpu" is cheaper to preempt than a remote processor */
374 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
377 first
= first_cpu(*mask
);
378 if (first
!= NR_CPUS
)
384 static int find_lowest_rq(struct task_struct
*task
)
386 struct sched_domain
*sd
;
387 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
388 int this_cpu
= smp_processor_id();
389 int cpu
= task_cpu(task
);
390 int count
= find_lowest_cpus(task
, lowest_mask
);
393 return -1; /* No targets found */
396 * There is no sense in performing an optimal search if only one
400 return first_cpu(*lowest_mask
);
403 * At this point we have built a mask of cpus representing the
404 * lowest priority tasks in the system. Now we want to elect
405 * the best one based on our affinity and topology.
407 * We prioritize the last cpu that the task executed on since
408 * it is most likely cache-hot in that location.
410 if (cpu_isset(cpu
, *lowest_mask
))
414 * Otherwise, we consult the sched_domains span maps to figure
415 * out which cpu is logically closest to our hot cache data.
418 this_cpu
= -1; /* Skip this_cpu opt if the same */
420 for_each_domain(cpu
, sd
) {
421 if (sd
->flags
& SD_WAKE_AFFINE
) {
422 cpumask_t domain_mask
;
425 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
427 best_cpu
= pick_optimal_cpu(this_cpu
,
435 * And finally, if there were no matches within the domains
436 * just give the caller *something* to work with from the compatible
439 return pick_optimal_cpu(this_cpu
, lowest_mask
);
442 /* Will lock the rq it finds */
443 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
445 struct rq
*lowest_rq
= NULL
;
449 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
450 cpu
= find_lowest_rq(task
);
452 if ((cpu
== -1) || (cpu
== rq
->cpu
))
455 lowest_rq
= cpu_rq(cpu
);
457 /* if the prio of this runqueue changed, try again */
458 if (double_lock_balance(rq
, lowest_rq
)) {
460 * We had to unlock the run queue. In
461 * the mean time, task could have
462 * migrated already or had its affinity changed.
463 * Also make sure that it wasn't scheduled on its rq.
465 if (unlikely(task_rq(task
) != rq
||
466 !cpu_isset(lowest_rq
->cpu
,
467 task
->cpus_allowed
) ||
468 task_running(rq
, task
) ||
471 spin_unlock(&lowest_rq
->lock
);
477 /* If this rq is still suitable use it. */
478 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
482 spin_unlock(&lowest_rq
->lock
);
490 * If the current CPU has more than one RT task, see if the non
491 * running task can migrate over to a CPU that is running a task
492 * of lesser priority.
494 static int push_rt_task(struct rq
*rq
)
496 struct task_struct
*next_task
;
497 struct rq
*lowest_rq
;
499 int paranoid
= RT_MAX_TRIES
;
501 if (!rq
->rt
.overloaded
)
504 next_task
= pick_next_highest_task_rt(rq
, -1);
509 if (unlikely(next_task
== rq
->curr
)) {
515 * It's possible that the next_task slipped in of
516 * higher priority than current. If that's the case
517 * just reschedule current.
519 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
520 resched_task(rq
->curr
);
524 /* We might release rq lock */
525 get_task_struct(next_task
);
527 /* find_lock_lowest_rq locks the rq if found */
528 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
530 struct task_struct
*task
;
532 * find lock_lowest_rq releases rq->lock
533 * so it is possible that next_task has changed.
534 * If it has, then try again.
536 task
= pick_next_highest_task_rt(rq
, -1);
537 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
538 put_task_struct(next_task
);
545 deactivate_task(rq
, next_task
, 0);
546 set_task_cpu(next_task
, lowest_rq
->cpu
);
547 activate_task(lowest_rq
, next_task
, 0);
549 resched_task(lowest_rq
->curr
);
551 spin_unlock(&lowest_rq
->lock
);
555 put_task_struct(next_task
);
561 * TODO: Currently we just use the second highest prio task on
562 * the queue, and stop when it can't migrate (or there's
563 * no more RT tasks). There may be a case where a lower
564 * priority RT task has a different affinity than the
565 * higher RT task. In this case the lower RT task could
566 * possibly be able to migrate where as the higher priority
567 * RT task could not. We currently ignore this issue.
568 * Enhancements are welcome!
570 static void push_rt_tasks(struct rq
*rq
)
572 /* push_rt_task will return true if it moved an RT */
573 while (push_rt_task(rq
))
577 static int pull_rt_task(struct rq
*this_rq
)
579 int this_cpu
= this_rq
->cpu
, ret
= 0, cpu
;
580 struct task_struct
*p
, *next
;
584 * If cpusets are used, and we have overlapping
585 * run queue cpusets, then this algorithm may not catch all.
586 * This is just the price you pay on trying to keep
587 * dirtying caches down on large SMP machines.
589 if (likely(!rt_overloaded()))
592 next
= pick_next_task_rt(this_rq
);
594 for_each_cpu_mask(cpu
, rt_overload_mask
) {
598 src_rq
= cpu_rq(cpu
);
599 if (unlikely(src_rq
->rt
.rt_nr_running
<= 1)) {
601 * It is possible that overlapping cpusets
602 * will miss clearing a non overloaded runqueue.
605 if (double_lock_balance(this_rq
, src_rq
)) {
606 /* unlocked our runqueue lock */
607 struct task_struct
*old_next
= next
;
609 next
= pick_next_task_rt(this_rq
);
610 if (next
!= old_next
)
613 if (likely(src_rq
->rt
.rt_nr_running
<= 1)) {
615 * Small chance that this_rq->curr changed
616 * but it's really harmless here.
618 rt_clear_overload(this_rq
);
621 * Heh, the src_rq is now overloaded, since
622 * we already have the src_rq lock, go straight
623 * to pulling tasks from it.
627 spin_unlock(&src_rq
->lock
);
632 * We can potentially drop this_rq's lock in
633 * double_lock_balance, and another CPU could
634 * steal our next task - hence we must cause
635 * the caller to recalculate the next task
638 if (double_lock_balance(this_rq
, src_rq
)) {
639 struct task_struct
*old_next
= next
;
641 next
= pick_next_task_rt(this_rq
);
642 if (next
!= old_next
)
647 * Are there still pullable RT tasks?
649 if (src_rq
->rt
.rt_nr_running
<= 1) {
650 spin_unlock(&src_rq
->lock
);
655 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
658 * Do we have an RT task that preempts
659 * the to-be-scheduled task?
661 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
662 WARN_ON(p
== src_rq
->curr
);
663 WARN_ON(!p
->se
.on_rq
);
666 * There's a chance that p is higher in priority
667 * than what's currently running on its cpu.
668 * This is just that p is wakeing up and hasn't
669 * had a chance to schedule. We only pull
670 * p if it is lower in priority than the
671 * current task on the run queue or
672 * this_rq next task is lower in prio than
673 * the current task on that rq.
675 if (p
->prio
< src_rq
->curr
->prio
||
676 (next
&& next
->prio
< src_rq
->curr
->prio
))
681 deactivate_task(src_rq
, p
, 0);
682 set_task_cpu(p
, this_cpu
);
683 activate_task(this_rq
, p
, 0);
685 * We continue with the search, just in
686 * case there's an even higher prio task
687 * in another runqueue. (low likelyhood
690 * Update next so that we won't pick a task
691 * on another cpu with a priority lower (or equal)
692 * than the one we just picked.
698 spin_unlock(&src_rq
->lock
);
704 static void schedule_balance_rt(struct rq
*rq
,
705 struct task_struct
*prev
)
707 /* Try to pull RT tasks here if we lower this rq's prio */
708 if (unlikely(rt_task(prev
)) &&
709 rq
->rt
.highest_prio
> prev
->prio
)
713 static void schedule_tail_balance_rt(struct rq
*rq
)
716 * If we have more than one rt_task queued, then
717 * see if we can push the other rt_tasks off to other CPUS.
718 * Note we may release the rq lock, and since
719 * the lock was owned by prev, we need to release it
720 * first via finish_lock_switch and then reaquire it here.
722 if (unlikely(rq
->rt
.overloaded
)) {
723 spin_lock_irq(&rq
->lock
);
725 spin_unlock_irq(&rq
->lock
);
730 static void wakeup_balance_rt(struct rq
*rq
, struct task_struct
*p
)
732 if (unlikely(rt_task(p
)) &&
733 !task_running(rq
, p
) &&
734 (p
->prio
>= rq
->rt
.highest_prio
) &&
740 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
741 unsigned long max_load_move
,
742 struct sched_domain
*sd
, enum cpu_idle_type idle
,
743 int *all_pinned
, int *this_best_prio
)
745 /* don't touch RT tasks */
750 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
751 struct sched_domain
*sd
, enum cpu_idle_type idle
)
753 /* don't touch RT tasks */
757 static void set_cpus_allowed_rt(struct task_struct
*p
, cpumask_t
*new_mask
)
759 int weight
= cpus_weight(*new_mask
);
764 * Update the migration status of the RQ if we have an RT task
765 * which is running AND changing its weight value.
767 if (p
->se
.on_rq
&& (weight
!= p
->nr_cpus_allowed
)) {
768 struct rq
*rq
= task_rq(p
);
770 if ((p
->nr_cpus_allowed
<= 1) && (weight
> 1)) {
771 rq
->rt
.rt_nr_migratory
++;
772 } else if ((p
->nr_cpus_allowed
> 1) && (weight
<= 1)) {
773 BUG_ON(!rq
->rt
.rt_nr_migratory
);
774 rq
->rt
.rt_nr_migratory
--;
777 update_rt_migration(rq
);
780 p
->cpus_allowed
= *new_mask
;
781 p
->nr_cpus_allowed
= weight
;
784 #else /* CONFIG_SMP */
785 # define schedule_tail_balance_rt(rq) do { } while (0)
786 # define schedule_balance_rt(rq, prev) do { } while (0)
787 # define wakeup_balance_rt(rq, p) do { } while (0)
788 #endif /* CONFIG_SMP */
790 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
)
795 * RR tasks need a special form of timeslice management.
796 * FIFO tasks have no timeslices.
798 if (p
->policy
!= SCHED_RR
)
804 p
->time_slice
= DEF_TIMESLICE
;
807 * Requeue to the end of queue if we are not the only element
810 if (p
->run_list
.prev
!= p
->run_list
.next
) {
811 requeue_task_rt(rq
, p
);
812 set_tsk_need_resched(p
);
816 static void set_curr_task_rt(struct rq
*rq
)
818 struct task_struct
*p
= rq
->curr
;
820 p
->se
.exec_start
= rq
->clock
;
823 const struct sched_class rt_sched_class
= {
824 .next
= &fair_sched_class
,
825 .enqueue_task
= enqueue_task_rt
,
826 .dequeue_task
= dequeue_task_rt
,
827 .yield_task
= yield_task_rt
,
829 .select_task_rq
= select_task_rq_rt
,
830 #endif /* CONFIG_SMP */
832 .check_preempt_curr
= check_preempt_curr_rt
,
834 .pick_next_task
= pick_next_task_rt
,
835 .put_prev_task
= put_prev_task_rt
,
838 .load_balance
= load_balance_rt
,
839 .move_one_task
= move_one_task_rt
,
840 .set_cpus_allowed
= set_cpus_allowed_rt
,
843 .set_curr_task
= set_curr_task_rt
,
844 .task_tick
= task_tick_rt
,