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)) {
37 if (!rq->rt.overloaded) {
39 rq->rt.overloaded = 1;
41 } else if (rq->rt.overloaded) {
42 rt_clear_overload(rq);
43 rq->rt.overloaded = 0;
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->rt.run_list, array->queue + p->prio);
115 __set_bit(p->prio, array->bitmap);
116 inc_cpu_load(rq, p->se.load.weight);
125 * Adding/removing a task to/from a priority array:
127 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
129 struct rt_prio_array *array = &rq->rt.active;
133 list_del(&p->rt.run_list);
134 if (list_empty(array->queue + p->prio))
135 __clear_bit(p->prio, array->bitmap);
136 dec_cpu_load(rq, p->se.load.weight);
142 * Put task to the end of the run list without the overhead of dequeue
143 * followed by enqueue.
145 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
147 struct rt_prio_array *array = &rq->rt.active;
149 list_move_tail(&p->rt.run_list, array->queue + p->prio);
153 yield_task_rt(struct rq *rq)
155 requeue_task_rt(rq, rq->curr);
159 static int find_lowest_rq(struct task_struct *task);
161 static int select_task_rq_rt(struct task_struct *p, int sync)
163 struct rq *rq = task_rq(p);
166 * If the current task is an RT task, then
167 * try to see if we can wake this RT task up on another
168 * runqueue. Otherwise simply start this RT task
169 * on its current runqueue.
171 * We want to avoid overloading runqueues. Even if
172 * the RT task is of higher priority than the current RT task.
173 * RT tasks behave differently than other tasks. If
174 * one gets preempted, we try to push it off to another queue.
175 * So trying to keep a preempting RT task on the same
176 * cache hot CPU will force the running RT task to
177 * a cold CPU. So we waste all the cache for the lower
178 * RT task in hopes of saving some of a RT task
179 * that is just being woken and probably will have
182 if (unlikely(rt_task(rq->curr)) &&
183 (p->nr_cpus_allowed > 1)) {
184 int cpu = find_lowest_rq(p);
186 return (cpu == -1) ? task_cpu(p) : cpu;
190 * Otherwise, just let it ride on the affined RQ and the
191 * post-schedule router will push the preempted task away
195 #endif /* CONFIG_SMP */
198 * Preempt the current task with a newly woken task if needed:
200 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
202 if (p->prio < rq->curr->prio)
203 resched_task(rq->curr);
206 static struct task_struct *pick_next_task_rt(struct rq *rq)
208 struct rt_prio_array *array = &rq->rt.active;
209 struct task_struct *next;
210 struct list_head *queue;
213 idx = sched_find_first_bit(array->bitmap);
214 if (idx >= MAX_RT_PRIO)
217 queue = array->queue + idx;
218 next = list_entry(queue->next, struct task_struct, rt.run_list);
220 next->se.exec_start = rq->clock;
225 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
228 p->se.exec_start = 0;
232 /* Only try algorithms three times */
233 #define RT_MAX_TRIES 3
235 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
236 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
238 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
240 if (!task_running(rq, p) &&
241 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
242 (p->nr_cpus_allowed > 1))
247 /* Return the second highest RT task, NULL otherwise */
248 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
250 struct rt_prio_array *array = &rq->rt.active;
251 struct task_struct *next;
252 struct list_head *queue;
255 if (likely(rq->rt.rt_nr_running < 2))
258 idx = sched_find_first_bit(array->bitmap);
259 if (unlikely(idx >= MAX_RT_PRIO)) {
260 WARN_ON(1); /* rt_nr_running is bad */
264 queue = array->queue + idx;
265 BUG_ON(list_empty(queue));
267 next = list_entry(queue->next, struct task_struct, rt.run_list);
268 if (unlikely(pick_rt_task(rq, next, cpu)))
271 if (queue->next->next != queue) {
273 next = list_entry(queue->next->next, struct task_struct,
275 if (pick_rt_task(rq, next, cpu))
280 /* slower, but more flexible */
281 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
282 if (unlikely(idx >= MAX_RT_PRIO))
285 queue = array->queue + idx;
286 BUG_ON(list_empty(queue));
288 list_for_each_entry(next, queue, rt.run_list) {
289 if (pick_rt_task(rq, next, cpu))
299 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
301 static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
303 int lowest_prio = -1;
308 cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed);
311 * Scan each rq for the lowest prio.
313 for_each_cpu_mask(cpu, *lowest_mask) {
314 struct rq *rq = cpu_rq(cpu);
316 /* We look for lowest RT prio or non-rt CPU */
317 if (rq->rt.highest_prio >= MAX_RT_PRIO) {
319 * if we already found a low RT queue
320 * and now we found this non-rt queue
321 * clear the mask and set our bit.
322 * Otherwise just return the queue as is
323 * and the count==1 will cause the algorithm
324 * to use the first bit found.
326 if (lowest_cpu != -1) {
327 cpus_clear(*lowest_mask);
328 cpu_set(rq->cpu, *lowest_mask);
333 /* no locking for now */
334 if ((rq->rt.highest_prio > task->prio)
335 && (rq->rt.highest_prio >= lowest_prio)) {
336 if (rq->rt.highest_prio > lowest_prio) {
337 /* new low - clear old data */
338 lowest_prio = rq->rt.highest_prio;
344 cpu_clear(cpu, *lowest_mask);
348 * Clear out all the set bits that represent
349 * runqueues that were of higher prio than
352 if (lowest_cpu > 0) {
354 * Perhaps we could add another cpumask op to
355 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
356 * Then that could be optimized to use memset and such.
358 for_each_cpu_mask(cpu, *lowest_mask) {
359 if (cpu >= lowest_cpu)
361 cpu_clear(cpu, *lowest_mask);
368 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
372 /* "this_cpu" is cheaper to preempt than a remote processor */
373 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
376 first = first_cpu(*mask);
377 if (first != NR_CPUS)
383 static int find_lowest_rq(struct task_struct *task)
385 struct sched_domain *sd;
386 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
387 int this_cpu = smp_processor_id();
388 int cpu = task_cpu(task);
389 int count = find_lowest_cpus(task, lowest_mask);
392 return -1; /* No targets found */
395 * There is no sense in performing an optimal search if only one
399 return first_cpu(*lowest_mask);
402 * At this point we have built a mask of cpus representing the
403 * lowest priority tasks in the system. Now we want to elect
404 * the best one based on our affinity and topology.
406 * We prioritize the last cpu that the task executed on since
407 * it is most likely cache-hot in that location.
409 if (cpu_isset(cpu, *lowest_mask))
413 * Otherwise, we consult the sched_domains span maps to figure
414 * out which cpu is logically closest to our hot cache data.
417 this_cpu = -1; /* Skip this_cpu opt if the same */
419 for_each_domain(cpu, sd) {
420 if (sd->flags & SD_WAKE_AFFINE) {
421 cpumask_t domain_mask;
424 cpus_and(domain_mask, sd->span, *lowest_mask);
426 best_cpu = pick_optimal_cpu(this_cpu,
434 * And finally, if there were no matches within the domains
435 * just give the caller *something* to work with from the compatible
438 return pick_optimal_cpu(this_cpu, lowest_mask);
441 /* Will lock the rq it finds */
442 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
444 struct rq *lowest_rq = NULL;
448 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
449 cpu = find_lowest_rq(task);
451 if ((cpu == -1) || (cpu == rq->cpu))
454 lowest_rq = cpu_rq(cpu);
456 /* if the prio of this runqueue changed, try again */
457 if (double_lock_balance(rq, lowest_rq)) {
459 * We had to unlock the run queue. In
460 * the mean time, task could have
461 * migrated already or had its affinity changed.
462 * Also make sure that it wasn't scheduled on its rq.
464 if (unlikely(task_rq(task) != rq ||
465 !cpu_isset(lowest_rq->cpu,
466 task->cpus_allowed) ||
467 task_running(rq, task) ||
470 spin_unlock(&lowest_rq->lock);
476 /* If this rq is still suitable use it. */
477 if (lowest_rq->rt.highest_prio > task->prio)
481 spin_unlock(&lowest_rq->lock);
489 * If the current CPU has more than one RT task, see if the non
490 * running task can migrate over to a CPU that is running a task
491 * of lesser priority.
493 static int push_rt_task(struct rq *rq)
495 struct task_struct *next_task;
496 struct rq *lowest_rq;
498 int paranoid = RT_MAX_TRIES;
500 if (!rq->rt.overloaded)
503 next_task = pick_next_highest_task_rt(rq, -1);
508 if (unlikely(next_task == rq->curr)) {
514 * It's possible that the next_task slipped in of
515 * higher priority than current. If that's the case
516 * just reschedule current.
518 if (unlikely(next_task->prio < rq->curr->prio)) {
519 resched_task(rq->curr);
523 /* We might release rq lock */
524 get_task_struct(next_task);
526 /* find_lock_lowest_rq locks the rq if found */
527 lowest_rq = find_lock_lowest_rq(next_task, rq);
529 struct task_struct *task;
531 * find lock_lowest_rq releases rq->lock
532 * so it is possible that next_task has changed.
533 * If it has, then try again.
535 task = pick_next_highest_task_rt(rq, -1);
536 if (unlikely(task != next_task) && task && paranoid--) {
537 put_task_struct(next_task);
544 deactivate_task(rq, next_task, 0);
545 set_task_cpu(next_task, lowest_rq->cpu);
546 activate_task(lowest_rq, next_task, 0);
548 resched_task(lowest_rq->curr);
550 spin_unlock(&lowest_rq->lock);
554 put_task_struct(next_task);
560 * TODO: Currently we just use the second highest prio task on
561 * the queue, and stop when it can't migrate (or there's
562 * no more RT tasks). There may be a case where a lower
563 * priority RT task has a different affinity than the
564 * higher RT task. In this case the lower RT task could
565 * possibly be able to migrate where as the higher priority
566 * RT task could not. We currently ignore this issue.
567 * Enhancements are welcome!
569 static void push_rt_tasks(struct rq *rq)
571 /* push_rt_task will return true if it moved an RT */
572 while (push_rt_task(rq))
576 static int pull_rt_task(struct rq *this_rq)
578 int this_cpu = this_rq->cpu, ret = 0, cpu;
579 struct task_struct *p, *next;
582 if (likely(!rt_overloaded(this_rq)))
585 next = pick_next_task_rt(this_rq);
587 for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
591 src_rq = cpu_rq(cpu);
593 * We can potentially drop this_rq's lock in
594 * double_lock_balance, and another CPU could
595 * steal our next task - hence we must cause
596 * the caller to recalculate the next task
599 if (double_lock_balance(this_rq, src_rq)) {
600 struct task_struct *old_next = next;
602 next = pick_next_task_rt(this_rq);
603 if (next != old_next)
608 * Are there still pullable RT tasks?
610 if (src_rq->rt.rt_nr_running <= 1) {
611 spin_unlock(&src_rq->lock);
615 p = pick_next_highest_task_rt(src_rq, this_cpu);
618 * Do we have an RT task that preempts
619 * the to-be-scheduled task?
621 if (p && (!next || (p->prio < next->prio))) {
622 WARN_ON(p == src_rq->curr);
623 WARN_ON(!p->se.on_rq);
626 * There's a chance that p is higher in priority
627 * than what's currently running on its cpu.
628 * This is just that p is wakeing up and hasn't
629 * had a chance to schedule. We only pull
630 * p if it is lower in priority than the
631 * current task on the run queue or
632 * this_rq next task is lower in prio than
633 * the current task on that rq.
635 if (p->prio < src_rq->curr->prio ||
636 (next && next->prio < src_rq->curr->prio))
641 deactivate_task(src_rq, p, 0);
642 set_task_cpu(p, this_cpu);
643 activate_task(this_rq, p, 0);
645 * We continue with the search, just in
646 * case there's an even higher prio task
647 * in another runqueue. (low likelyhood
650 * Update next so that we won't pick a task
651 * on another cpu with a priority lower (or equal)
652 * than the one we just picked.
658 spin_unlock(&src_rq->lock);
664 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
666 /* Try to pull RT tasks here if we lower this rq's prio */
667 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
671 static void post_schedule_rt(struct rq *rq)
674 * If we have more than one rt_task queued, then
675 * see if we can push the other rt_tasks off to other CPUS.
676 * Note we may release the rq lock, and since
677 * the lock was owned by prev, we need to release it
678 * first via finish_lock_switch and then reaquire it here.
680 if (unlikely(rq->rt.overloaded)) {
681 spin_lock_irq(&rq->lock);
683 spin_unlock_irq(&rq->lock);
688 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
690 if (!task_running(rq, p) &&
691 (p->prio >= rq->rt.highest_prio) &&
697 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
698 unsigned long max_load_move,
699 struct sched_domain *sd, enum cpu_idle_type idle,
700 int *all_pinned, int *this_best_prio)
702 /* don't touch RT tasks */
707 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
708 struct sched_domain *sd, enum cpu_idle_type idle)
710 /* don't touch RT tasks */
714 static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
716 int weight = cpus_weight(*new_mask);
721 * Update the migration status of the RQ if we have an RT task
722 * which is running AND changing its weight value.
724 if (p->se.on_rq && (weight != p->nr_cpus_allowed)) {
725 struct rq *rq = task_rq(p);
727 if ((p->nr_cpus_allowed <= 1) && (weight > 1)) {
728 rq->rt.rt_nr_migratory++;
729 } else if ((p->nr_cpus_allowed > 1) && (weight <= 1)) {
730 BUG_ON(!rq->rt.rt_nr_migratory);
731 rq->rt.rt_nr_migratory--;
734 update_rt_migration(rq);
737 p->cpus_allowed = *new_mask;
738 p->nr_cpus_allowed = weight;
741 /* Assumes rq->lock is held */
742 static void join_domain_rt(struct rq *rq)
744 if (rq->rt.overloaded)
748 /* Assumes rq->lock is held */
749 static void leave_domain_rt(struct rq *rq)
751 if (rq->rt.overloaded)
752 rt_clear_overload(rq);
756 * When switch from the rt queue, we bring ourselves to a position
757 * that we might want to pull RT tasks from other runqueues.
759 static void switched_from_rt(struct rq *rq, struct task_struct *p,
763 * If there are other RT tasks then we will reschedule
764 * and the scheduling of the other RT tasks will handle
765 * the balancing. But if we are the last RT task
766 * we may need to handle the pulling of RT tasks
769 if (!rq->rt.rt_nr_running)
772 #endif /* CONFIG_SMP */
775 * When switching a task to RT, we may overload the runqueue
776 * with RT tasks. In this case we try to push them off to
779 static void switched_to_rt(struct rq *rq, struct task_struct *p,
782 int check_resched = 1;
785 * If we are already running, then there's nothing
786 * that needs to be done. But if we are not running
787 * we may need to preempt the current running task.
788 * If that current running task is also an RT task
789 * then see if we can move to another run queue.
793 if (rq->rt.overloaded && push_rt_task(rq) &&
794 /* Don't resched if we changed runqueues */
797 #endif /* CONFIG_SMP */
798 if (check_resched && p->prio < rq->curr->prio)
799 resched_task(rq->curr);
804 * Priority of the task has changed. This may cause
805 * us to initiate a push or pull.
807 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
808 int oldprio, int running)
813 * If our priority decreases while running, we
814 * may need to pull tasks to this runqueue.
816 if (oldprio < p->prio)
819 * If there's a higher priority task waiting to run
822 if (p->prio > rq->rt.highest_prio)
825 /* For UP simply resched on drop of prio */
826 if (oldprio < p->prio)
828 #endif /* CONFIG_SMP */
831 * This task is not running, but if it is
832 * greater than the current running task
835 if (p->prio < rq->curr->prio)
836 resched_task(rq->curr);
840 static void watchdog(struct rq *rq, struct task_struct *p)
842 unsigned long soft, hard;
847 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
848 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
850 if (soft != RLIM_INFINITY) {
854 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
855 if (next > p->rt.timeout) {
856 u64 next_time = p->se.sum_exec_runtime;
858 next_time += next * (NSEC_PER_SEC/HZ);
859 if (p->it_sched_expires > next_time)
860 p->it_sched_expires = next_time;
862 p->it_sched_expires = p->se.sum_exec_runtime;
866 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
873 * RR tasks need a special form of timeslice management.
874 * FIFO tasks have no timeslices.
876 if (p->policy != SCHED_RR)
879 if (--p->rt.time_slice)
882 p->rt.time_slice = DEF_TIMESLICE;
885 * Requeue to the end of queue if we are not the only element
888 if (p->rt.run_list.prev != p->rt.run_list.next) {
889 requeue_task_rt(rq, p);
890 set_tsk_need_resched(p);
894 static void set_curr_task_rt(struct rq *rq)
896 struct task_struct *p = rq->curr;
898 p->se.exec_start = rq->clock;
901 const struct sched_class rt_sched_class = {
902 .next = &fair_sched_class,
903 .enqueue_task = enqueue_task_rt,
904 .dequeue_task = dequeue_task_rt,
905 .yield_task = yield_task_rt,
907 .select_task_rq = select_task_rq_rt,
908 #endif /* CONFIG_SMP */
910 .check_preempt_curr = check_preempt_curr_rt,
912 .pick_next_task = pick_next_task_rt,
913 .put_prev_task = put_prev_task_rt,
916 .load_balance = load_balance_rt,
917 .move_one_task = move_one_task_rt,
918 .set_cpus_allowed = set_cpus_allowed_rt,
919 .join_domain = join_domain_rt,
920 .leave_domain = leave_domain_rt,
921 .pre_schedule = pre_schedule_rt,
922 .post_schedule = post_schedule_rt,
923 .task_wake_up = task_wake_up_rt,
924 .switched_from = switched_from_rt,
927 .set_curr_task = set_curr_task_rt,
928 .task_tick = task_tick_rt,
930 .prio_changed = prio_changed_rt,
931 .switched_to = switched_to_rt,