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 select_task_rq_rt(struct task_struct
*p
, int sync
)
158 #endif /* CONFIG_SMP */
161 * Preempt the current task with a newly woken task if needed:
163 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
)
165 if (p
->prio
< rq
->curr
->prio
)
166 resched_task(rq
->curr
);
169 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
171 struct rt_prio_array
*array
= &rq
->rt
.active
;
172 struct task_struct
*next
;
173 struct list_head
*queue
;
176 idx
= sched_find_first_bit(array
->bitmap
);
177 if (idx
>= MAX_RT_PRIO
)
180 queue
= array
->queue
+ idx
;
181 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
183 next
->se
.exec_start
= rq
->clock
;
188 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
191 p
->se
.exec_start
= 0;
195 /* Only try algorithms three times */
196 #define RT_MAX_TRIES 3
198 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
);
199 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
);
201 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
203 if (!task_running(rq
, p
) &&
204 (cpu
< 0 || cpu_isset(cpu
, p
->cpus_allowed
)) &&
205 (p
->nr_cpus_allowed
> 1))
210 /* Return the second highest RT task, NULL otherwise */
211 static struct task_struct
*pick_next_highest_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 assert_spin_locked(&rq
->lock
);
221 if (likely(rq
->rt
.rt_nr_running
< 2))
224 idx
= sched_find_first_bit(array
->bitmap
);
225 if (unlikely(idx
>= MAX_RT_PRIO
)) {
226 WARN_ON(1); /* rt_nr_running is bad */
230 queue
= array
->queue
+ idx
;
231 BUG_ON(list_empty(queue
));
233 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
234 if (unlikely(pick_rt_task(rq
, next
, cpu
)))
237 if (queue
->next
->next
!= queue
) {
239 next
= list_entry(queue
->next
->next
, struct task_struct
, run_list
);
240 if (pick_rt_task(rq
, next
, cpu
))
245 /* slower, but more flexible */
246 idx
= find_next_bit(array
->bitmap
, MAX_RT_PRIO
, idx
+1);
247 if (unlikely(idx
>= MAX_RT_PRIO
))
250 queue
= array
->queue
+ idx
;
251 BUG_ON(list_empty(queue
));
253 list_for_each_entry(next
, queue
, run_list
) {
254 if (pick_rt_task(rq
, next
, cpu
))
264 static DEFINE_PER_CPU(cpumask_t
, local_cpu_mask
);
266 static int find_lowest_rq(struct task_struct
*task
)
269 cpumask_t
*cpu_mask
= &__get_cpu_var(local_cpu_mask
);
270 struct rq
*lowest_rq
= NULL
;
272 cpus_and(*cpu_mask
, cpu_online_map
, task
->cpus_allowed
);
275 * Scan each rq for the lowest prio.
277 for_each_cpu_mask(cpu
, *cpu_mask
) {
278 struct rq
*rq
= cpu_rq(cpu
);
283 /* We look for lowest RT prio or non-rt CPU */
284 if (rq
->rt
.highest_prio
>= MAX_RT_PRIO
) {
289 /* no locking for now */
290 if (rq
->rt
.highest_prio
> task
->prio
&&
291 (!lowest_rq
|| rq
->rt
.highest_prio
> lowest_rq
->rt
.highest_prio
)) {
296 return lowest_rq
? lowest_rq
->cpu
: -1;
299 /* Will lock the rq it finds */
300 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
,
303 struct rq
*lowest_rq
= NULL
;
307 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
308 cpu
= find_lowest_rq(task
);
313 lowest_rq
= cpu_rq(cpu
);
315 /* if the prio of this runqueue changed, try again */
316 if (double_lock_balance(rq
, lowest_rq
)) {
318 * We had to unlock the run queue. In
319 * the mean time, task could have
320 * migrated already or had its affinity changed.
321 * Also make sure that it wasn't scheduled on its rq.
323 if (unlikely(task_rq(task
) != rq
||
324 !cpu_isset(lowest_rq
->cpu
, task
->cpus_allowed
) ||
325 task_running(rq
, task
) ||
327 spin_unlock(&lowest_rq
->lock
);
333 /* If this rq is still suitable use it. */
334 if (lowest_rq
->rt
.highest_prio
> task
->prio
)
338 spin_unlock(&lowest_rq
->lock
);
346 * If the current CPU has more than one RT task, see if the non
347 * running task can migrate over to a CPU that is running a task
348 * of lesser priority.
350 static int push_rt_task(struct rq
*rq
)
352 struct task_struct
*next_task
;
353 struct rq
*lowest_rq
;
355 int paranoid
= RT_MAX_TRIES
;
357 assert_spin_locked(&rq
->lock
);
359 next_task
= pick_next_highest_task_rt(rq
, -1);
364 if (unlikely(next_task
== rq
->curr
)) {
370 * It's possible that the next_task slipped in of
371 * higher priority than current. If that's the case
372 * just reschedule current.
374 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
375 resched_task(rq
->curr
);
379 /* We might release rq lock */
380 get_task_struct(next_task
);
382 /* find_lock_lowest_rq locks the rq if found */
383 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
385 struct task_struct
*task
;
387 * find lock_lowest_rq releases rq->lock
388 * so it is possible that next_task has changed.
389 * If it has, then try again.
391 task
= pick_next_highest_task_rt(rq
, -1);
392 if (unlikely(task
!= next_task
) && task
&& paranoid
--) {
393 put_task_struct(next_task
);
400 assert_spin_locked(&lowest_rq
->lock
);
402 deactivate_task(rq
, next_task
, 0);
403 set_task_cpu(next_task
, lowest_rq
->cpu
);
404 activate_task(lowest_rq
, next_task
, 0);
406 resched_task(lowest_rq
->curr
);
408 spin_unlock(&lowest_rq
->lock
);
412 put_task_struct(next_task
);
418 * TODO: Currently we just use the second highest prio task on
419 * the queue, and stop when it can't migrate (or there's
420 * no more RT tasks). There may be a case where a lower
421 * priority RT task has a different affinity than the
422 * higher RT task. In this case the lower RT task could
423 * possibly be able to migrate where as the higher priority
424 * RT task could not. We currently ignore this issue.
425 * Enhancements are welcome!
427 static void push_rt_tasks(struct rq
*rq
)
429 /* push_rt_task will return true if it moved an RT */
430 while (push_rt_task(rq
))
434 static int pull_rt_task(struct rq
*this_rq
)
436 struct task_struct
*next
;
437 struct task_struct
*p
;
439 cpumask_t
*rto_cpumask
;
440 int this_cpu
= this_rq
->cpu
;
444 assert_spin_locked(&this_rq
->lock
);
447 * If cpusets are used, and we have overlapping
448 * run queue cpusets, then this algorithm may not catch all.
449 * This is just the price you pay on trying to keep
450 * dirtying caches down on large SMP machines.
452 if (likely(!rt_overloaded()))
455 next
= pick_next_task_rt(this_rq
);
457 rto_cpumask
= rt_overload();
459 for_each_cpu_mask(cpu
, *rto_cpumask
) {
463 src_rq
= cpu_rq(cpu
);
464 if (unlikely(src_rq
->rt
.rt_nr_running
<= 1)) {
466 * It is possible that overlapping cpusets
467 * will miss clearing a non overloaded runqueue.
470 if (double_lock_balance(this_rq
, src_rq
)) {
471 /* unlocked our runqueue lock */
472 struct task_struct
*old_next
= next
;
473 next
= pick_next_task_rt(this_rq
);
474 if (next
!= old_next
)
477 if (likely(src_rq
->rt
.rt_nr_running
<= 1))
479 * Small chance that this_rq->curr changed
480 * but it's really harmless here.
482 rt_clear_overload(this_rq
);
485 * Heh, the src_rq is now overloaded, since
486 * we already have the src_rq lock, go straight
487 * to pulling tasks from it.
490 spin_unlock(&src_rq
->lock
);
495 * We can potentially drop this_rq's lock in
496 * double_lock_balance, and another CPU could
497 * steal our next task - hence we must cause
498 * the caller to recalculate the next task
501 if (double_lock_balance(this_rq
, src_rq
)) {
502 struct task_struct
*old_next
= next
;
503 next
= pick_next_task_rt(this_rq
);
504 if (next
!= old_next
)
509 * Are there still pullable RT tasks?
511 if (src_rq
->rt
.rt_nr_running
<= 1) {
512 spin_unlock(&src_rq
->lock
);
517 p
= pick_next_highest_task_rt(src_rq
, this_cpu
);
520 * Do we have an RT task that preempts
521 * the to-be-scheduled task?
523 if (p
&& (!next
|| (p
->prio
< next
->prio
))) {
524 WARN_ON(p
== src_rq
->curr
);
525 WARN_ON(!p
->se
.on_rq
);
528 * There's a chance that p is higher in priority
529 * than what's currently running on its cpu.
530 * This is just that p is wakeing up and hasn't
531 * had a chance to schedule. We only pull
532 * p if it is lower in priority than the
533 * current task on the run queue or
534 * this_rq next task is lower in prio than
535 * the current task on that rq.
537 if (p
->prio
< src_rq
->curr
->prio
||
538 (next
&& next
->prio
< src_rq
->curr
->prio
))
543 deactivate_task(src_rq
, p
, 0);
544 set_task_cpu(p
, this_cpu
);
545 activate_task(this_rq
, p
, 0);
547 * We continue with the search, just in
548 * case there's an even higher prio task
549 * in another runqueue. (low likelyhood
554 * Update next so that we won't pick a task
555 * on another cpu with a priority lower (or equal)
556 * than the one we just picked.
562 spin_unlock(&src_rq
->lock
);
568 static void schedule_balance_rt(struct rq
*rq
,
569 struct task_struct
*prev
)
571 /* Try to pull RT tasks here if we lower this rq's prio */
572 if (unlikely(rt_task(prev
)) &&
573 rq
->rt
.highest_prio
> prev
->prio
)
577 static void schedule_tail_balance_rt(struct rq
*rq
)
580 * If we have more than one rt_task queued, then
581 * see if we can push the other rt_tasks off to other CPUS.
582 * Note we may release the rq lock, and since
583 * the lock was owned by prev, we need to release it
584 * first via finish_lock_switch and then reaquire it here.
586 if (unlikely(rq
->rt
.rt_nr_running
> 1)) {
587 spin_lock_irq(&rq
->lock
);
589 spin_unlock_irq(&rq
->lock
);
594 static void wakeup_balance_rt(struct rq
*rq
, struct task_struct
*p
)
596 if (unlikely(rt_task(p
)) &&
597 !task_running(rq
, p
) &&
598 (p
->prio
>= rq
->curr
->prio
))
603 load_balance_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
604 unsigned long max_load_move
,
605 struct sched_domain
*sd
, enum cpu_idle_type idle
,
606 int *all_pinned
, int *this_best_prio
)
608 /* don't touch RT tasks */
613 move_one_task_rt(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
614 struct sched_domain
*sd
, enum cpu_idle_type idle
)
616 /* don't touch RT tasks */
619 static void set_cpus_allowed_rt(struct task_struct
*p
, cpumask_t
*new_mask
)
621 int weight
= cpus_weight(*new_mask
);
626 * Update the migration status of the RQ if we have an RT task
627 * which is running AND changing its weight value.
629 if (p
->se
.on_rq
&& (weight
!= p
->nr_cpus_allowed
)) {
630 struct rq
*rq
= task_rq(p
);
632 if ((p
->nr_cpus_allowed
<= 1) && (weight
> 1))
633 rq
->rt
.rt_nr_migratory
++;
634 else if((p
->nr_cpus_allowed
> 1) && (weight
<= 1)) {
635 BUG_ON(!rq
->rt
.rt_nr_migratory
);
636 rq
->rt
.rt_nr_migratory
--;
639 update_rt_migration(rq
);
642 p
->cpus_allowed
= *new_mask
;
643 p
->nr_cpus_allowed
= weight
;
645 #else /* CONFIG_SMP */
646 # define schedule_tail_balance_rt(rq) do { } while (0)
647 # define schedule_balance_rt(rq, prev) do { } while (0)
648 # define wakeup_balance_rt(rq, p) do { } while (0)
649 #endif /* CONFIG_SMP */
651 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
)
656 * RR tasks need a special form of timeslice management.
657 * FIFO tasks have no timeslices.
659 if (p
->policy
!= SCHED_RR
)
665 p
->time_slice
= DEF_TIMESLICE
;
668 * Requeue to the end of queue if we are not the only element
671 if (p
->run_list
.prev
!= p
->run_list
.next
) {
672 requeue_task_rt(rq
, p
);
673 set_tsk_need_resched(p
);
677 static void set_curr_task_rt(struct rq
*rq
)
679 struct task_struct
*p
= rq
->curr
;
681 p
->se
.exec_start
= rq
->clock
;
684 const struct sched_class rt_sched_class
= {
685 .next
= &fair_sched_class
,
686 .enqueue_task
= enqueue_task_rt
,
687 .dequeue_task
= dequeue_task_rt
,
688 .yield_task
= yield_task_rt
,
690 .select_task_rq
= select_task_rq_rt
,
691 #endif /* CONFIG_SMP */
693 .check_preempt_curr
= check_preempt_curr_rt
,
695 .pick_next_task
= pick_next_task_rt
,
696 .put_prev_task
= put_prev_task_rt
,
699 .load_balance
= load_balance_rt
,
700 .move_one_task
= move_one_task_rt
,
701 .set_cpus_allowed
= set_cpus_allowed_rt
,
704 .set_curr_task
= set_curr_task_rt
,
705 .task_tick
= task_tick_rt
,