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 cpu_set(rq
->cpu
, rt_overload_mask
);
21 * Make sure the mask is visible before we set
22 * the overload count. That is checked to determine
23 * if we should look at the mask. It would be a shame
24 * if we looked at the mask, but the mask was not
28 atomic_inc(&rto_count
);
30 static inline void rt_clear_overload(struct rq
*rq
)
32 /* the order here really doesn't matter */
33 atomic_dec(&rto_count
);
34 cpu_clear(rq
->cpu
, rt_overload_mask
);
37 static void update_rt_migration(struct rq
*rq
)
39 if (rq
->rt
.rt_nr_migratory
&& (rq
->rt
.rt_nr_running
> 1))
42 rt_clear_overload(rq
);
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 task will not preempt the RQ, try to find a better RQ
162 * before we even activate the task
164 if ((p
->prio
>= rq
->rt
.highest_prio
)
165 && (p
->nr_cpus_allowed
> 1)) {
166 int cpu
= find_lowest_rq(p
);
168 return (cpu
== -1) ? task_cpu(p
) : cpu
;
172 * Otherwise, just let it ride on the affined RQ and the
173 * post-schedule router will push the preempted task away
177 #endif /* CONFIG_SMP */
180 * Preempt the current task with a newly woken task if needed:
182 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
184 if (p
->prio
< rq
->curr
->prio
)
185 resched_task(rq
->curr
);
188 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
190 struct rt_prio_array
*array
= &rq
->rt
.active
;
191 struct task_struct
*next
;
192 struct list_head
*queue
;
195 idx
= sched_find_first_bit(array
->bitmap
);
196 if (idx
>= MAX_RT_PRIO
)
199 queue
= array
->queue
+ idx
;
200 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
202 next
->se
.exec_start
= rq
->clock
;
207 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
210 p
->se
.exec_start
= 0;
214 /* Only try algorithms three times */
215 #define RT_MAX_TRIES 3
217 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
218 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
220 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
222 if (!task_running(rq
, p
) &&
223 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
224 (p
->nr_cpus_allowed
> 1))
229 /* Return the second highest RT task, NULL otherwise */
230 static struct task_struct
*pick_next_highest_task_rt(struct rq
*rq
,
233 struct rt_prio_array
*array
= &rq
->rt
.active
;
234 struct task_struct
*next
;
235 struct list_head
*queue
;
238 assert_spin_locked(&rq
->lock
);
240 if (likely(rq
->rt
.rt_nr_running
< 2))
243 idx
= sched_find_first_bit(array
->bitmap
);
244 if (unlikely(idx
>= MAX_RT_PRIO
)) {
245 WARN_ON(1); /* rt_nr_running is bad */
249 queue
= array
->queue
+ idx
;
250 BUG_ON(list_empty(queue
));
252 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
253 if (unlikely(pick_rt_task(rq
, next
, cpu
)))
256 if (queue
->next
->next
!= queue
) {
258 next
= list_entry(queue
->next
->next
, struct task_struct
, run_list
);
259 if (pick_rt_task(rq
, next
, cpu
))
264 /* slower, but more flexible */
265 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
266 if (unlikely(idx
>= MAX_RT_PRIO
))
269 queue
= array
->queue
+ idx
;
270 BUG_ON(list_empty(queue
));
272 list_for_each_entry(next
, queue
, run_list
) {
273 if (pick_rt_task(rq
, next
, cpu
))
283 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
284 static DEFINE_PER_CPU(cpumask_t
, valid_cpu_mask
);
286 static int find_lowest_cpus(struct task_struct
*task
, cpumask_t
*lowest_mask
)
289 cpumask_t
*valid_mask
= &__get_cpu_var(valid_cpu_mask
);
290 int lowest_prio
= -1;
293 cpus_clear(*lowest_mask
);
294 cpus_and(*valid_mask
, cpu_online_map
, task
->cpus_allowed
);
297 * Scan each rq for the lowest prio.
299 for_each_cpu_mask(cpu
, *valid_mask
) {
300 struct rq
*rq
= cpu_rq(cpu
);
302 /* We look for lowest RT prio or non-rt CPU */
303 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
305 cpus_clear(*lowest_mask
);
306 cpu_set(rq
->cpu
, *lowest_mask
);
310 /* no locking for now */
311 if ((rq
->rt
.highest_prio
> task
->prio
)
312 && (rq
->rt
.highest_prio
>= lowest_prio
)) {
313 if (rq
->rt
.highest_prio
> lowest_prio
) {
314 /* new low - clear old data */
315 lowest_prio
= rq
->rt
.highest_prio
;
316 cpus_clear(*lowest_mask
);
318 cpu_set(rq
->cpu
, *lowest_mask
);
326 static inline int pick_optimal_cpu(int this_cpu
, cpumask_t
*mask
)
330 /* "this_cpu" is cheaper to preempt than a remote processor */
331 if ((this_cpu
!= -1) && cpu_isset(this_cpu
, *mask
))
334 first
= first_cpu(*mask
);
335 if (first
!= NR_CPUS
)
341 static int find_lowest_rq(struct task_struct
*task
)
343 struct sched_domain
*sd
;
344 cpumask_t
*lowest_mask
= &__get_cpu_var(local_cpu_mask
);
345 int this_cpu
= smp_processor_id();
346 int cpu
= task_cpu(task
);
348 if (!find_lowest_cpus(task
, lowest_mask
))
352 * At this point we have built a mask of cpus representing the
353 * lowest priority tasks in the system. Now we want to elect
354 * the best one based on our affinity and topology.
356 * We prioritize the last cpu that the task executed on since
357 * it is most likely cache-hot in that location.
359 if (cpu_isset(cpu
, *lowest_mask
))
363 * Otherwise, we consult the sched_domains span maps to figure
364 * out which cpu is logically closest to our hot cache data.
367 this_cpu
= -1; /* Skip this_cpu opt if the same */
369 for_each_domain(cpu
, sd
) {
370 if (sd
->flags
& SD_WAKE_AFFINE
) {
371 cpumask_t domain_mask
;
374 cpus_and(domain_mask
, sd
->span
, *lowest_mask
);
376 best_cpu
= pick_optimal_cpu(this_cpu
,
384 * And finally, if there were no matches within the domains
385 * just give the caller *something* to work with from the compatible
388 return pick_optimal_cpu(this_cpu
, lowest_mask
);
391 /* Will lock the rq it finds */
392 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
,
395 struct rq
*lowest_rq
= NULL
;
399 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
400 cpu
= find_lowest_rq(task
);
402 if ((cpu
== -1) || (cpu
== rq
->cpu
))
405 lowest_rq
= cpu_rq(cpu
);
407 /* if the prio of this runqueue changed, try again */
408 if (double_lock_balance(rq
, lowest_rq
)) {
410 * We had to unlock the run queue. In
411 * the mean time, task could have
412 * migrated already or had its affinity changed.
413 * Also make sure that it wasn't scheduled on its rq.
415 if (unlikely(task_rq(task
) != rq
||
416 !cpu_isset(lowest_rq
->cpu
, task
->cpus_allowed
) ||
417 task_running(rq
, task
) ||
419 spin_unlock(&lowest_rq
->lock
);
425 /* If this rq is still suitable use it. */
426 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
430 spin_unlock(&lowest_rq
->lock
);
438 * If the current CPU has more than one RT task, see if the non
439 * running task can migrate over to a CPU that is running a task
440 * of lesser priority.
442 static int push_rt_task(struct rq
*rq
)
444 struct task_struct
*next_task
;
445 struct rq
*lowest_rq
;
447 int paranoid
= RT_MAX_TRIES
;
449 assert_spin_locked(&rq
->lock
);
451 next_task
= pick_next_highest_task_rt(rq
, -1);
456 if (unlikely(next_task
== rq
->curr
)) {
462 * It's possible that the next_task slipped in of
463 * higher priority than current. If that's the case
464 * just reschedule current.
466 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
467 resched_task(rq
->curr
);
471 /* We might release rq lock */
472 get_task_struct(next_task
);
474 /* find_lock_lowest_rq locks the rq if found */
475 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
477 struct task_struct
*task
;
479 * find lock_lowest_rq releases rq->lock
480 * so it is possible that next_task has changed.
481 * If it has, then try again.
483 task
= pick_next_highest_task_rt(rq
, -1);
484 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
485 put_task_struct(next_task
);
492 assert_spin_locked(&lowest_rq
->lock
);
494 deactivate_task(rq
, next_task
, 0);
495 set_task_cpu(next_task
, lowest_rq
->cpu
);
496 activate_task(lowest_rq
, next_task
, 0);
498 resched_task(lowest_rq
->curr
);
500 spin_unlock(&lowest_rq
->lock
);
504 put_task_struct(next_task
);
510 * TODO: Currently we just use the second highest prio task on
511 * the queue, and stop when it can't migrate (or there's
512 * no more RT tasks). There may be a case where a lower
513 * priority RT task has a different affinity than the
514 * higher RT task. In this case the lower RT task could
515 * possibly be able to migrate where as the higher priority
516 * RT task could not. We currently ignore this issue.
517 * Enhancements are welcome!
519 static void push_rt_tasks(struct rq
*rq
)
521 /* push_rt_task will return true if it moved an RT */
522 while (push_rt_task(rq
))
526 static int pull_rt_task(struct rq
*this_rq
)
528 struct task_struct
*next
;
529 struct task_struct
*p
;
531 cpumask_t
*rto_cpumask
;
532 int this_cpu
= this_rq
->cpu
;
536 assert_spin_locked(&this_rq
->lock
);
539 * If cpusets are used, and we have overlapping
540 * run queue cpusets, then this algorithm may not catch all.
541 * This is just the price you pay on trying to keep
542 * dirtying caches down on large SMP machines.
544 if (likely(!rt_overloaded()))
547 next
= pick_next_task_rt(this_rq
);
549 rto_cpumask
= rt_overload();
551 for_each_cpu_mask(cpu
, *rto_cpumask
) {
555 src_rq
= cpu_rq(cpu
);
556 if (unlikely(src_rq
->rt
.rt_nr_running
<= 1)) {
558 * It is possible that overlapping cpusets
559 * will miss clearing a non overloaded runqueue.
562 if (double_lock_balance(this_rq
, src_rq
)) {
563 /* unlocked our runqueue lock */
564 struct task_struct
*old_next
= next
;
565 next
= pick_next_task_rt(this_rq
);
566 if (next
!= old_next
)
569 if (likely(src_rq
->rt
.rt_nr_running
<= 1))
571 * Small chance that this_rq->curr changed
572 * but it's really harmless here.
574 rt_clear_overload(this_rq
);
577 * Heh, the src_rq is now overloaded, since
578 * we already have the src_rq lock, go straight
579 * to pulling tasks from it.
582 spin_unlock(&src_rq
->lock
);
587 * We can potentially drop this_rq's lock in
588 * double_lock_balance, and another CPU could
589 * steal our next task - hence we must cause
590 * the caller to recalculate the next task
593 if (double_lock_balance(this_rq
, src_rq
)) {
594 struct task_struct
*old_next
= next
;
595 next
= pick_next_task_rt(this_rq
);
596 if (next
!= old_next
)
601 * Are there still pullable RT tasks?
603 if (src_rq
->rt
.rt_nr_running
<= 1) {
604 spin_unlock(&src_rq
->lock
);
609 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
612 * Do we have an RT task that preempts
613 * the to-be-scheduled task?
615 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
616 WARN_ON(p
== src_rq
->curr
);
617 WARN_ON(!p
->se
.on_rq
);
620 * There's a chance that p is higher in priority
621 * than what's currently running on its cpu.
622 * This is just that p is wakeing up and hasn't
623 * had a chance to schedule. We only pull
624 * p if it is lower in priority than the
625 * current task on the run queue or
626 * this_rq next task is lower in prio than
627 * the current task on that rq.
629 if (p
->prio
< src_rq
->curr
->prio
||
630 (next
&& next
->prio
< src_rq
->curr
->prio
))
635 deactivate_task(src_rq
, p
, 0);
636 set_task_cpu(p
, this_cpu
);
637 activate_task(this_rq
, p
, 0);
639 * We continue with the search, just in
640 * case there's an even higher prio task
641 * in another runqueue. (low likelyhood
646 * Update next so that we won't pick a task
647 * on another cpu with a priority lower (or equal)
648 * than the one we just picked.
654 spin_unlock(&src_rq
->lock
);
660 static void schedule_balance_rt(struct rq
*rq
,
661 struct task_struct
*prev
)
663 /* Try to pull RT tasks here if we lower this rq's prio */
664 if (unlikely(rt_task(prev
)) &&
665 rq
->rt
.highest_prio
> prev
->prio
)
669 static void schedule_tail_balance_rt(struct rq
*rq
)
672 * If we have more than one rt_task queued, then
673 * see if we can push the other rt_tasks off to other CPUS.
674 * Note we may release the rq lock, and since
675 * the lock was owned by prev, we need to release it
676 * first via finish_lock_switch and then reaquire it here.
678 if (unlikely(rq
->rt
.rt_nr_running
> 1)) {
679 spin_lock_irq(&rq
->lock
);
681 spin_unlock_irq(&rq
->lock
);
686 static void wakeup_balance_rt(struct rq
*rq
, struct task_struct
*p
)
688 if (unlikely(rt_task(p
)) &&
689 !task_running(rq
, p
) &&
690 (p
->prio
>= rq
->curr
->prio
))
695 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
696 unsigned long max_load_move
,
697 struct sched_domain
*sd
, enum cpu_idle_type idle
,
698 int *all_pinned
, int *this_best_prio
)
700 /* don't touch RT tasks */
705 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
706 struct sched_domain
*sd
, enum cpu_idle_type idle
)
708 /* 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
;
737 #else /* CONFIG_SMP */
738 # define schedule_tail_balance_rt(rq) do { } while (0)
739 # define schedule_balance_rt(rq, prev) do { } while (0)
740 # define wakeup_balance_rt(rq, p) do { } while (0)
741 #endif /* CONFIG_SMP */
743 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
)
748 * RR tasks need a special form of timeslice management.
749 * FIFO tasks have no timeslices.
751 if (p
->policy
!= SCHED_RR
)
757 p
->time_slice
= DEF_TIMESLICE
;
760 * Requeue to the end of queue if we are not the only element
763 if (p
->run_list
.prev
!= p
->run_list
.next
) {
764 requeue_task_rt(rq
, p
);
765 set_tsk_need_resched(p
);
769 static void set_curr_task_rt(struct rq
*rq
)
771 struct task_struct
*p
= rq
->curr
;
773 p
->se
.exec_start
= rq
->clock
;
776 const struct sched_class rt_sched_class
= {
777 .next
= &fair_sched_class
,
778 .enqueue_task
= enqueue_task_rt
,
779 .dequeue_task
= dequeue_task_rt
,
780 .yield_task
= yield_task_rt
,
782 .select_task_rq
= select_task_rq_rt
,
783 #endif /* CONFIG_SMP */
785 .check_preempt_curr
= check_preempt_curr_rt
,
787 .pick_next_task
= pick_next_task_rt
,
788 .put_prev_task
= put_prev_task_rt
,
791 .load_balance
= load_balance_rt
,
792 .move_one_task
= move_one_task_rt
,
793 .set_cpus_allowed
= set_cpus_allowed_rt
,
796 .set_curr_task
= set_curr_task_rt
,
797 .task_tick
= task_tick_rt
,