2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
7 static cpumask_t rt_overload_mask
;
8 static atomic_t rto_count
;
9 static inline int rt_overloaded(void)
11 return atomic_read(&rto_count
);
13 static inline cpumask_t
*rt_overload(void)
15 return &rt_overload_mask
;
17 static inline void rt_set_overload(struct rq
*rq
)
19 rq
->rt
.overloaded
= 1;
20 cpu_set(rq
->cpu
, rt_overload_mask
);
22 * Make sure the mask is visible before we set
23 * the overload count. That is checked to determine
24 * if we should look at the mask. It would be a shame
25 * if we looked at the mask, but the mask was not
29 atomic_inc(&rto_count
);
31 static inline void rt_clear_overload(struct rq
*rq
)
33 /* the order here really doesn't matter */
34 atomic_dec(&rto_count
);
35 cpu_clear(rq
->cpu
, rt_overload_mask
);
36 rq
->rt
.overloaded
= 0;
39 static void update_rt_migration(struct rq
*rq
)
41 if (rq
->rt
.rt_nr_migratory
&& (rq
->rt
.rt_nr_running
> 1))
44 rt_clear_overload(rq
);
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 assert_spin_locked(&rq
->lock
);
254 if (likely(rq
->rt
.rt_nr_running
< 2))
257 idx
= sched_find_first_bit(array
->bitmap
);
258 if (unlikely(idx
>= MAX_RT_PRIO
)) {
259 WARN_ON(1); /* rt_nr_running is bad */
263 queue
= array
->queue
+ idx
;
264 BUG_ON(list_empty(queue
));
266 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
267 if (unlikely(pick_rt_task(rq
, next
, cpu
)))
270 if (queue
->next
->next
!= queue
) {
272 next
= list_entry(queue
->next
->next
, struct task_struct
,
274 if (pick_rt_task(rq
, next
, cpu
))
279 /* slower, but more flexible */
280 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
281 if (unlikely(idx
>= MAX_RT_PRIO
))
284 queue
= array
->queue
+ idx
;
285 BUG_ON(list_empty(queue
));
287 list_for_each_entry(next
, queue
, run_list
) {
288 if (pick_rt_task(rq
, next
, cpu
))
298 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
300 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
302 int lowest_prio
= -1;
307 cpus_and(*lowest_mask
, cpu_online_map
, task
->cpus_allowed
);
310 * Scan each rq for the lowest prio.
312 for_each_cpu_mask(cpu
, *lowest_mask
) {
313 struct rq
*rq
= cpu_rq(cpu
);
315 /* We look for lowest RT prio or non-rt CPU */
316 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
318 * if we already found a low RT queue
319 * and now we found this non-rt queue
320 * clear the mask and set our bit.
321 * Otherwise just return the queue as is
322 * and the count==1 will cause the algorithm
323 * to use the first bit found.
325 if (lowest_cpu
!= -1) {
326 cpus_clear(*lowest_mask
);
327 cpu_set(rq
->cpu
, *lowest_mask
);
332 /* no locking for now */
333 if ((rq
->rt
.highest_prio
> task
->prio
)
334 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
335 if (rq
->rt
.highest_prio
> lowest_prio
) {
336 /* new low - clear old data */
337 lowest_prio
= rq
->rt
.highest_prio
;
343 cpu_clear(cpu
, *lowest_mask
);
347 * Clear out all the set bits that represent
348 * runqueues that were of higher prio than
351 if (lowest_cpu
> 0) {
353 * Perhaps we could add another cpumask op to
354 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
355 * Then that could be optimized to use memset and such.
357 for_each_cpu_mask(cpu
, *lowest_mask
) {
358 if (cpu
>= lowest_cpu
)
360 cpu_clear(cpu
, *lowest_mask
);
367 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
371 /* "this_cpu" is cheaper to preempt than a remote processor */
372 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
375 first
= first_cpu(*mask
);
376 if (first
!= NR_CPUS
)
382 static int find_lowest_rq(struct task_struct
*task
)
384 struct sched_domain
*sd
;
385 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
386 int this_cpu
= smp_processor_id();
387 int cpu
= task_cpu(task
);
388 int count
= find_lowest_cpus(task
, lowest_mask
);
391 return -1; /* No targets found */
394 * There is no sense in performing an optimal search if only one
398 return first_cpu(*lowest_mask
);
401 * At this point we have built a mask of cpus representing the
402 * lowest priority tasks in the system. Now we want to elect
403 * the best one based on our affinity and topology.
405 * We prioritize the last cpu that the task executed on since
406 * it is most likely cache-hot in that location.
408 if (cpu_isset(cpu
, *lowest_mask
))
412 * Otherwise, we consult the sched_domains span maps to figure
413 * out which cpu is logically closest to our hot cache data.
416 this_cpu
= -1; /* Skip this_cpu opt if the same */
418 for_each_domain(cpu
, sd
) {
419 if (sd
->flags
& SD_WAKE_AFFINE
) {
420 cpumask_t domain_mask
;
423 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
425 best_cpu
= pick_optimal_cpu(this_cpu
,
433 * And finally, if there were no matches within the domains
434 * just give the caller *something* to work with from the compatible
437 return pick_optimal_cpu(this_cpu
, lowest_mask
);
440 /* Will lock the rq it finds */
441 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
443 struct rq
*lowest_rq
= NULL
;
447 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
448 cpu
= find_lowest_rq(task
);
450 if ((cpu
== -1) || (cpu
== rq
->cpu
))
453 lowest_rq
= cpu_rq(cpu
);
455 /* if the prio of this runqueue changed, try again */
456 if (double_lock_balance(rq
, lowest_rq
)) {
458 * We had to unlock the run queue. In
459 * the mean time, task could have
460 * migrated already or had its affinity changed.
461 * Also make sure that it wasn't scheduled on its rq.
463 if (unlikely(task_rq(task
) != rq
||
464 !cpu_isset(lowest_rq
->cpu
,
465 task
->cpus_allowed
) ||
466 task_running(rq
, task
) ||
469 spin_unlock(&lowest_rq
->lock
);
475 /* If this rq is still suitable use it. */
476 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
480 spin_unlock(&lowest_rq
->lock
);
488 * If the current CPU has more than one RT task, see if the non
489 * running task can migrate over to a CPU that is running a task
490 * of lesser priority.
492 static int push_rt_task(struct rq
*rq
)
494 struct task_struct
*next_task
;
495 struct rq
*lowest_rq
;
497 int paranoid
= RT_MAX_TRIES
;
499 assert_spin_locked(&rq
->lock
);
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 assert_spin_locked(&lowest_rq
->lock
);
547 deactivate_task(rq
, next_task
, 0);
548 set_task_cpu(next_task
, lowest_rq
->cpu
);
549 activate_task(lowest_rq
, next_task
, 0);
551 resched_task(lowest_rq
->curr
);
553 spin_unlock(&lowest_rq
->lock
);
557 put_task_struct(next_task
);
563 * TODO: Currently we just use the second highest prio task on
564 * the queue, and stop when it can't migrate (or there's
565 * no more RT tasks). There may be a case where a lower
566 * priority RT task has a different affinity than the
567 * higher RT task. In this case the lower RT task could
568 * possibly be able to migrate where as the higher priority
569 * RT task could not. We currently ignore this issue.
570 * Enhancements are welcome!
572 static void push_rt_tasks(struct rq
*rq
)
574 /* push_rt_task will return true if it moved an RT */
575 while (push_rt_task(rq
))
579 static int pull_rt_task(struct rq
*this_rq
)
581 struct task_struct
*next
;
582 struct task_struct
*p
;
584 cpumask_t
*rto_cpumask
;
585 int this_cpu
= this_rq
->cpu
;
589 assert_spin_locked(&this_rq
->lock
);
592 * If cpusets are used, and we have overlapping
593 * run queue cpusets, then this algorithm may not catch all.
594 * This is just the price you pay on trying to keep
595 * dirtying caches down on large SMP machines.
597 if (likely(!rt_overloaded()))
600 next
= pick_next_task_rt(this_rq
);
602 rto_cpumask
= rt_overload();
604 for_each_cpu_mask(cpu
, *rto_cpumask
) {
608 src_rq
= cpu_rq(cpu
);
609 if (unlikely(src_rq
->rt
.rt_nr_running
<= 1)) {
611 * It is possible that overlapping cpusets
612 * will miss clearing a non overloaded runqueue.
615 if (double_lock_balance(this_rq
, src_rq
)) {
616 /* unlocked our runqueue lock */
617 struct task_struct
*old_next
= next
;
618 next
= pick_next_task_rt(this_rq
);
619 if (next
!= old_next
)
622 if (likely(src_rq
->rt
.rt_nr_running
<= 1))
624 * Small chance that this_rq->curr changed
625 * but it's really harmless here.
627 rt_clear_overload(this_rq
);
630 * Heh, the src_rq is now overloaded, since
631 * we already have the src_rq lock, go straight
632 * to pulling tasks from it.
635 spin_unlock(&src_rq
->lock
);
640 * We can potentially drop this_rq's lock in
641 * double_lock_balance, and another CPU could
642 * steal our next task - hence we must cause
643 * the caller to recalculate the next task
646 if (double_lock_balance(this_rq
, src_rq
)) {
647 struct task_struct
*old_next
= next
;
648 next
= pick_next_task_rt(this_rq
);
649 if (next
!= old_next
)
654 * Are there still pullable RT tasks?
656 if (src_rq
->rt
.rt_nr_running
<= 1) {
657 spin_unlock(&src_rq
->lock
);
662 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
665 * Do we have an RT task that preempts
666 * the to-be-scheduled task?
668 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
669 WARN_ON(p
== src_rq
->curr
);
670 WARN_ON(!p
->se
.on_rq
);
673 * There's a chance that p is higher in priority
674 * than what's currently running on its cpu.
675 * This is just that p is wakeing up and hasn't
676 * had a chance to schedule. We only pull
677 * p if it is lower in priority than the
678 * current task on the run queue or
679 * this_rq next task is lower in prio than
680 * the current task on that rq.
682 if (p
->prio
< src_rq
->curr
->prio
||
683 (next
&& next
->prio
< src_rq
->curr
->prio
))
688 deactivate_task(src_rq
, p
, 0);
689 set_task_cpu(p
, this_cpu
);
690 activate_task(this_rq
, p
, 0);
692 * We continue with the search, just in
693 * case there's an even higher prio task
694 * in another runqueue. (low likelyhood
699 * Update next so that we won't pick a task
700 * on another cpu with a priority lower (or equal)
701 * than the one we just picked.
707 spin_unlock(&src_rq
->lock
);
713 static void schedule_balance_rt(struct rq
*rq
,
714 struct task_struct
*prev
)
716 /* Try to pull RT tasks here if we lower this rq's prio */
717 if (unlikely(rt_task(prev
)) &&
718 rq
->rt
.highest_prio
> prev
->prio
)
722 static void schedule_tail_balance_rt(struct rq
*rq
)
725 * If we have more than one rt_task queued, then
726 * see if we can push the other rt_tasks off to other CPUS.
727 * Note we may release the rq lock, and since
728 * the lock was owned by prev, we need to release it
729 * first via finish_lock_switch and then reaquire it here.
731 if (unlikely(rq
->rt
.overloaded
)) {
732 spin_lock_irq(&rq
->lock
);
734 spin_unlock_irq(&rq
->lock
);
739 static void wakeup_balance_rt(struct rq
*rq
, struct task_struct
*p
)
741 if (unlikely(rt_task(p
)) &&
742 !task_running(rq
, p
) &&
743 (p
->prio
>= rq
->rt
.highest_prio
) &&
749 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
750 unsigned long max_load_move
,
751 struct sched_domain
*sd
, enum cpu_idle_type idle
,
752 int *all_pinned
, int *this_best_prio
)
754 /* don't touch RT tasks */
759 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
760 struct sched_domain
*sd
, enum cpu_idle_type idle
)
762 /* don't touch RT tasks */
765 static void set_cpus_allowed_rt(struct task_struct
*p
, cpumask_t
*new_mask
)
767 int weight
= cpus_weight(*new_mask
);
772 * Update the migration status of the RQ if we have an RT task
773 * which is running AND changing its weight value.
775 if (p
->se
.on_rq
&& (weight
!= p
->nr_cpus_allowed
)) {
776 struct rq
*rq
= task_rq(p
);
778 if ((p
->nr_cpus_allowed
<= 1) && (weight
> 1))
779 rq
->rt
.rt_nr_migratory
++;
780 else if((p
->nr_cpus_allowed
> 1) && (weight
<= 1)) {
781 BUG_ON(!rq
->rt
.rt_nr_migratory
);
782 rq
->rt
.rt_nr_migratory
--;
785 update_rt_migration(rq
);
788 p
->cpus_allowed
= *new_mask
;
789 p
->nr_cpus_allowed
= weight
;
791 #else /* CONFIG_SMP */
792 # define schedule_tail_balance_rt(rq) do { } while (0)
793 # define schedule_balance_rt(rq, prev) do { } while (0)
794 # define wakeup_balance_rt(rq, p) do { } while (0)
795 #endif /* CONFIG_SMP */
797 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
)
802 * RR tasks need a special form of timeslice management.
803 * FIFO tasks have no timeslices.
805 if (p
->policy
!= SCHED_RR
)
811 p
->time_slice
= DEF_TIMESLICE
;
814 * Requeue to the end of queue if we are not the only element
817 if (p
->run_list
.prev
!= p
->run_list
.next
) {
818 requeue_task_rt(rq
, p
);
819 set_tsk_need_resched(p
);
823 static void set_curr_task_rt(struct rq
*rq
)
825 struct task_struct
*p
= rq
->curr
;
827 p
->se
.exec_start
= rq
->clock
;
830 const struct sched_class rt_sched_class
= {
831 .next
= &fair_sched_class
,
832 .enqueue_task
= enqueue_task_rt
,
833 .dequeue_task
= dequeue_task_rt
,
834 .yield_task
= yield_task_rt
,
836 .select_task_rq
= select_task_rq_rt
,
837 #endif /* CONFIG_SMP */
839 .check_preempt_curr
= check_preempt_curr_rt
,
841 .pick_next_task
= pick_next_task_rt
,
842 .put_prev_task
= put_prev_task_rt
,
845 .load_balance
= load_balance_rt
,
846 .move_one_task
= move_one_task_rt
,
847 .set_cpus_allowed
= set_cpus_allowed_rt
,
850 .set_curr_task
= set_curr_task_rt
,
851 .task_tick
= task_tick_rt
,