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 */
48 static int sched_rt_ratio_exceeded(struct rq *rq, struct rt_rq *rt_rq)
52 if (sysctl_sched_rt_ratio == SCHED_RT_FRAC)
55 if (rt_rq->rt_throttled)
58 period = (u64)sysctl_sched_rt_period * NSEC_PER_MSEC;
59 ratio = (period * sysctl_sched_rt_ratio) >> SCHED_RT_FRAC_SHIFT;
61 if (rt_rq->rt_time > ratio) {
62 rt_rq->rt_throttled = rq->clock + period - rt_rq->rt_time;
69 static void update_sched_rt_period(struct rq *rq)
71 while (rq->clock > rq->rt_period_expire) {
74 period = (u64)sysctl_sched_rt_period * NSEC_PER_MSEC;
75 ratio = (period * sysctl_sched_rt_ratio) >> SCHED_RT_FRAC_SHIFT;
77 rq->rt.rt_time -= min(rq->rt.rt_time, ratio);
78 rq->rt_period_expire += period;
82 * When the rt throttle is expired, let them rip.
83 * (XXX: use hrtick when available)
85 if (rq->rt.rt_throttled && rq->clock > rq->rt.rt_throttled) {
86 rq->rt.rt_throttled = 0;
87 if (!sched_rt_ratio_exceeded(rq, &rq->rt))
88 resched_task(rq->curr);
93 * Update the current task's runtime statistics. Skip current tasks that
94 * are not in our scheduling class.
96 static void update_curr_rt(struct rq *rq)
98 struct task_struct *curr = rq->curr;
101 if (!task_has_rt_policy(curr))
104 delta_exec = rq->clock - curr->se.exec_start;
105 if (unlikely((s64)delta_exec < 0))
108 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
110 curr->se.sum_exec_runtime += delta_exec;
111 curr->se.exec_start = rq->clock;
112 cpuacct_charge(curr, delta_exec);
114 rq->rt.rt_time += delta_exec;
115 update_sched_rt_period(rq);
116 if (sched_rt_ratio_exceeded(rq, &rq->rt))
120 static inline void inc_rt_tasks(struct task_struct *p, struct rq *rq)
122 WARN_ON(!rt_task(p));
123 rq->rt.rt_nr_running++;
125 if (p->prio < rq->rt.highest_prio)
126 rq->rt.highest_prio = p->prio;
127 if (p->nr_cpus_allowed > 1)
128 rq->rt.rt_nr_migratory++;
130 update_rt_migration(rq);
131 #endif /* CONFIG_SMP */
134 static inline void dec_rt_tasks(struct task_struct *p, struct rq *rq)
136 WARN_ON(!rt_task(p));
137 WARN_ON(!rq->rt.rt_nr_running);
138 rq->rt.rt_nr_running--;
140 if (rq->rt.rt_nr_running) {
141 struct rt_prio_array *array;
143 WARN_ON(p->prio < rq->rt.highest_prio);
144 if (p->prio == rq->rt.highest_prio) {
146 array = &rq->rt.active;
147 rq->rt.highest_prio =
148 sched_find_first_bit(array->bitmap);
149 } /* otherwise leave rq->highest prio alone */
151 rq->rt.highest_prio = MAX_RT_PRIO;
152 if (p->nr_cpus_allowed > 1)
153 rq->rt.rt_nr_migratory--;
155 update_rt_migration(rq);
156 #endif /* CONFIG_SMP */
159 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
161 struct rt_prio_array *array = &rq->rt.active;
163 list_add_tail(&p->rt.run_list, array->queue + p->prio);
164 __set_bit(p->prio, array->bitmap);
165 inc_cpu_load(rq, p->se.load.weight);
174 * Adding/removing a task to/from a priority array:
176 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
178 struct rt_prio_array *array = &rq->rt.active;
182 list_del(&p->rt.run_list);
183 if (list_empty(array->queue + p->prio))
184 __clear_bit(p->prio, array->bitmap);
185 dec_cpu_load(rq, p->se.load.weight);
191 * Put task to the end of the run list without the overhead of dequeue
192 * followed by enqueue.
194 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
196 struct rt_prio_array *array = &rq->rt.active;
198 list_move_tail(&p->rt.run_list, array->queue + p->prio);
202 yield_task_rt(struct rq *rq)
204 requeue_task_rt(rq, rq->curr);
208 static int find_lowest_rq(struct task_struct *task);
210 static int select_task_rq_rt(struct task_struct *p, int sync)
212 struct rq *rq = task_rq(p);
215 * If the current task is an RT task, then
216 * try to see if we can wake this RT task up on another
217 * runqueue. Otherwise simply start this RT task
218 * on its current runqueue.
220 * We want to avoid overloading runqueues. Even if
221 * the RT task is of higher priority than the current RT task.
222 * RT tasks behave differently than other tasks. If
223 * one gets preempted, we try to push it off to another queue.
224 * So trying to keep a preempting RT task on the same
225 * cache hot CPU will force the running RT task to
226 * a cold CPU. So we waste all the cache for the lower
227 * RT task in hopes of saving some of a RT task
228 * that is just being woken and probably will have
231 if (unlikely(rt_task(rq->curr)) &&
232 (p->nr_cpus_allowed > 1)) {
233 int cpu = find_lowest_rq(p);
235 return (cpu == -1) ? task_cpu(p) : cpu;
239 * Otherwise, just let it ride on the affined RQ and the
240 * post-schedule router will push the preempted task away
244 #endif /* CONFIG_SMP */
247 * Preempt the current task with a newly woken task if needed:
249 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
251 if (p->prio < rq->curr->prio)
252 resched_task(rq->curr);
255 static struct task_struct *pick_next_task_rt(struct rq *rq)
257 struct rt_prio_array *array = &rq->rt.active;
258 struct task_struct *next;
259 struct list_head *queue;
260 struct rt_rq *rt_rq = &rq->rt;
263 if (sched_rt_ratio_exceeded(rq, rt_rq))
266 idx = sched_find_first_bit(array->bitmap);
267 if (idx >= MAX_RT_PRIO)
270 queue = array->queue + idx;
271 next = list_entry(queue->next, struct task_struct, rt.run_list);
273 next->se.exec_start = rq->clock;
278 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
281 p->se.exec_start = 0;
285 /* Only try algorithms three times */
286 #define RT_MAX_TRIES 3
288 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
289 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
291 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
293 if (!task_running(rq, p) &&
294 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
295 (p->nr_cpus_allowed > 1))
300 /* Return the second highest RT task, NULL otherwise */
301 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
303 struct rt_prio_array *array = &rq->rt.active;
304 struct task_struct *next;
305 struct list_head *queue;
308 if (likely(rq->rt.rt_nr_running < 2))
311 idx = sched_find_first_bit(array->bitmap);
312 if (unlikely(idx >= MAX_RT_PRIO)) {
313 WARN_ON(1); /* rt_nr_running is bad */
317 queue = array->queue + idx;
318 BUG_ON(list_empty(queue));
320 next = list_entry(queue->next, struct task_struct, rt.run_list);
321 if (unlikely(pick_rt_task(rq, next, cpu)))
324 if (queue->next->next != queue) {
326 next = list_entry(queue->next->next, struct task_struct,
328 if (pick_rt_task(rq, next, cpu))
333 /* slower, but more flexible */
334 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
335 if (unlikely(idx >= MAX_RT_PRIO))
338 queue = array->queue + idx;
339 BUG_ON(list_empty(queue));
341 list_for_each_entry(next, queue, rt.run_list) {
342 if (pick_rt_task(rq, next, cpu))
352 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
354 static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
356 int lowest_prio = -1;
361 cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed);
364 * Scan each rq for the lowest prio.
366 for_each_cpu_mask(cpu, *lowest_mask) {
367 struct rq *rq = cpu_rq(cpu);
369 /* We look for lowest RT prio or non-rt CPU */
370 if (rq->rt.highest_prio >= MAX_RT_PRIO) {
372 * if we already found a low RT queue
373 * and now we found this non-rt queue
374 * clear the mask and set our bit.
375 * Otherwise just return the queue as is
376 * and the count==1 will cause the algorithm
377 * to use the first bit found.
379 if (lowest_cpu != -1) {
380 cpus_clear(*lowest_mask);
381 cpu_set(rq->cpu, *lowest_mask);
386 /* no locking for now */
387 if ((rq->rt.highest_prio > task->prio)
388 && (rq->rt.highest_prio >= lowest_prio)) {
389 if (rq->rt.highest_prio > lowest_prio) {
390 /* new low - clear old data */
391 lowest_prio = rq->rt.highest_prio;
397 cpu_clear(cpu, *lowest_mask);
401 * Clear out all the set bits that represent
402 * runqueues that were of higher prio than
405 if (lowest_cpu > 0) {
407 * Perhaps we could add another cpumask op to
408 * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
409 * Then that could be optimized to use memset and such.
411 for_each_cpu_mask(cpu, *lowest_mask) {
412 if (cpu >= lowest_cpu)
414 cpu_clear(cpu, *lowest_mask);
421 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
425 /* "this_cpu" is cheaper to preempt than a remote processor */
426 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
429 first = first_cpu(*mask);
430 if (first != NR_CPUS)
436 static int find_lowest_rq(struct task_struct *task)
438 struct sched_domain *sd;
439 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
440 int this_cpu = smp_processor_id();
441 int cpu = task_cpu(task);
442 int count = find_lowest_cpus(task, lowest_mask);
445 return -1; /* No targets found */
448 * There is no sense in performing an optimal search if only one
452 return first_cpu(*lowest_mask);
455 * At this point we have built a mask of cpus representing the
456 * lowest priority tasks in the system. Now we want to elect
457 * the best one based on our affinity and topology.
459 * We prioritize the last cpu that the task executed on since
460 * it is most likely cache-hot in that location.
462 if (cpu_isset(cpu, *lowest_mask))
466 * Otherwise, we consult the sched_domains span maps to figure
467 * out which cpu is logically closest to our hot cache data.
470 this_cpu = -1; /* Skip this_cpu opt if the same */
472 for_each_domain(cpu, sd) {
473 if (sd->flags & SD_WAKE_AFFINE) {
474 cpumask_t domain_mask;
477 cpus_and(domain_mask, sd->span, *lowest_mask);
479 best_cpu = pick_optimal_cpu(this_cpu,
487 * And finally, if there were no matches within the domains
488 * just give the caller *something* to work with from the compatible
491 return pick_optimal_cpu(this_cpu, lowest_mask);
494 /* Will lock the rq it finds */
495 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
497 struct rq *lowest_rq = NULL;
501 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
502 cpu = find_lowest_rq(task);
504 if ((cpu == -1) || (cpu == rq->cpu))
507 lowest_rq = cpu_rq(cpu);
509 /* if the prio of this runqueue changed, try again */
510 if (double_lock_balance(rq, lowest_rq)) {
512 * We had to unlock the run queue. In
513 * the mean time, task could have
514 * migrated already or had its affinity changed.
515 * Also make sure that it wasn't scheduled on its rq.
517 if (unlikely(task_rq(task) != rq ||
518 !cpu_isset(lowest_rq->cpu,
519 task->cpus_allowed) ||
520 task_running(rq, task) ||
523 spin_unlock(&lowest_rq->lock);
529 /* If this rq is still suitable use it. */
530 if (lowest_rq->rt.highest_prio > task->prio)
534 spin_unlock(&lowest_rq->lock);
542 * If the current CPU has more than one RT task, see if the non
543 * running task can migrate over to a CPU that is running a task
544 * of lesser priority.
546 static int push_rt_task(struct rq *rq)
548 struct task_struct *next_task;
549 struct rq *lowest_rq;
551 int paranoid = RT_MAX_TRIES;
553 if (!rq->rt.overloaded)
556 next_task = pick_next_highest_task_rt(rq, -1);
561 if (unlikely(next_task == rq->curr)) {
567 * It's possible that the next_task slipped in of
568 * higher priority than current. If that's the case
569 * just reschedule current.
571 if (unlikely(next_task->prio < rq->curr->prio)) {
572 resched_task(rq->curr);
576 /* We might release rq lock */
577 get_task_struct(next_task);
579 /* find_lock_lowest_rq locks the rq if found */
580 lowest_rq = find_lock_lowest_rq(next_task, rq);
582 struct task_struct *task;
584 * find lock_lowest_rq releases rq->lock
585 * so it is possible that next_task has changed.
586 * If it has, then try again.
588 task = pick_next_highest_task_rt(rq, -1);
589 if (unlikely(task != next_task) && task && paranoid--) {
590 put_task_struct(next_task);
597 deactivate_task(rq, next_task, 0);
598 set_task_cpu(next_task, lowest_rq->cpu);
599 activate_task(lowest_rq, next_task, 0);
601 resched_task(lowest_rq->curr);
603 spin_unlock(&lowest_rq->lock);
607 put_task_struct(next_task);
613 * TODO: Currently we just use the second highest prio task on
614 * the queue, and stop when it can't migrate (or there's
615 * no more RT tasks). There may be a case where a lower
616 * priority RT task has a different affinity than the
617 * higher RT task. In this case the lower RT task could
618 * possibly be able to migrate where as the higher priority
619 * RT task could not. We currently ignore this issue.
620 * Enhancements are welcome!
622 static void push_rt_tasks(struct rq *rq)
624 /* push_rt_task will return true if it moved an RT */
625 while (push_rt_task(rq))
629 static int pull_rt_task(struct rq *this_rq)
631 int this_cpu = this_rq->cpu, ret = 0, cpu;
632 struct task_struct *p, *next;
635 if (likely(!rt_overloaded(this_rq)))
638 next = pick_next_task_rt(this_rq);
640 for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
644 src_rq = cpu_rq(cpu);
646 * We can potentially drop this_rq's lock in
647 * double_lock_balance, and another CPU could
648 * steal our next task - hence we must cause
649 * the caller to recalculate the next task
652 if (double_lock_balance(this_rq, src_rq)) {
653 struct task_struct *old_next = next;
655 next = pick_next_task_rt(this_rq);
656 if (next != old_next)
661 * Are there still pullable RT tasks?
663 if (src_rq->rt.rt_nr_running <= 1) {
664 spin_unlock(&src_rq->lock);
668 p = pick_next_highest_task_rt(src_rq, this_cpu);
671 * Do we have an RT task that preempts
672 * the to-be-scheduled task?
674 if (p && (!next || (p->prio < next->prio))) {
675 WARN_ON(p == src_rq->curr);
676 WARN_ON(!p->se.on_rq);
679 * There's a chance that p is higher in priority
680 * than what's currently running on its cpu.
681 * This is just that p is wakeing up and hasn't
682 * had a chance to schedule. We only pull
683 * p if it is lower in priority than the
684 * current task on the run queue or
685 * this_rq next task is lower in prio than
686 * the current task on that rq.
688 if (p->prio < src_rq->curr->prio ||
689 (next && next->prio < src_rq->curr->prio))
694 deactivate_task(src_rq, p, 0);
695 set_task_cpu(p, this_cpu);
696 activate_task(this_rq, p, 0);
698 * We continue with the search, just in
699 * case there's an even higher prio task
700 * in another runqueue. (low likelyhood
703 * Update next so that we won't pick a task
704 * on another cpu with a priority lower (or equal)
705 * than the one we just picked.
711 spin_unlock(&src_rq->lock);
717 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
719 /* Try to pull RT tasks here if we lower this rq's prio */
720 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
724 static void post_schedule_rt(struct rq *rq)
727 * If we have more than one rt_task queued, then
728 * see if we can push the other rt_tasks off to other CPUS.
729 * Note we may release the rq lock, and since
730 * the lock was owned by prev, we need to release it
731 * first via finish_lock_switch and then reaquire it here.
733 if (unlikely(rq->rt.overloaded)) {
734 spin_lock_irq(&rq->lock);
736 spin_unlock_irq(&rq->lock);
741 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
743 if (!task_running(rq, p) &&
744 (p->prio >= rq->rt.highest_prio) &&
750 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
751 unsigned long max_load_move,
752 struct sched_domain *sd, enum cpu_idle_type idle,
753 int *all_pinned, int *this_best_prio)
755 /* don't touch RT tasks */
760 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
761 struct sched_domain *sd, enum cpu_idle_type idle)
763 /* don't touch RT tasks */
767 static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
769 int weight = cpus_weight(*new_mask);
774 * Update the migration status of the RQ if we have an RT task
775 * which is running AND changing its weight value.
777 if (p->se.on_rq && (weight != p->nr_cpus_allowed)) {
778 struct rq *rq = task_rq(p);
780 if ((p->nr_cpus_allowed <= 1) && (weight > 1)) {
781 rq->rt.rt_nr_migratory++;
782 } else if ((p->nr_cpus_allowed > 1) && (weight <= 1)) {
783 BUG_ON(!rq->rt.rt_nr_migratory);
784 rq->rt.rt_nr_migratory--;
787 update_rt_migration(rq);
790 p->cpus_allowed = *new_mask;
791 p->nr_cpus_allowed = weight;
794 /* Assumes rq->lock is held */
795 static void join_domain_rt(struct rq *rq)
797 if (rq->rt.overloaded)
801 /* Assumes rq->lock is held */
802 static void leave_domain_rt(struct rq *rq)
804 if (rq->rt.overloaded)
805 rt_clear_overload(rq);
809 * When switch from the rt queue, we bring ourselves to a position
810 * that we might want to pull RT tasks from other runqueues.
812 static void switched_from_rt(struct rq *rq, struct task_struct *p,
816 * If there are other RT tasks then we will reschedule
817 * and the scheduling of the other RT tasks will handle
818 * the balancing. But if we are the last RT task
819 * we may need to handle the pulling of RT tasks
822 if (!rq->rt.rt_nr_running)
825 #endif /* CONFIG_SMP */
828 * When switching a task to RT, we may overload the runqueue
829 * with RT tasks. In this case we try to push them off to
832 static void switched_to_rt(struct rq *rq, struct task_struct *p,
835 int check_resched = 1;
838 * If we are already running, then there's nothing
839 * that needs to be done. But if we are not running
840 * we may need to preempt the current running task.
841 * If that current running task is also an RT task
842 * then see if we can move to another run queue.
846 if (rq->rt.overloaded && push_rt_task(rq) &&
847 /* Don't resched if we changed runqueues */
850 #endif /* CONFIG_SMP */
851 if (check_resched && p->prio < rq->curr->prio)
852 resched_task(rq->curr);
857 * Priority of the task has changed. This may cause
858 * us to initiate a push or pull.
860 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
861 int oldprio, int running)
866 * If our priority decreases while running, we
867 * may need to pull tasks to this runqueue.
869 if (oldprio < p->prio)
872 * If there's a higher priority task waiting to run
875 if (p->prio > rq->rt.highest_prio)
878 /* For UP simply resched on drop of prio */
879 if (oldprio < p->prio)
881 #endif /* CONFIG_SMP */
884 * This task is not running, but if it is
885 * greater than the current running task
888 if (p->prio < rq->curr->prio)
889 resched_task(rq->curr);
893 static void watchdog(struct rq *rq, struct task_struct *p)
895 unsigned long soft, hard;
900 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
901 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
903 if (soft != RLIM_INFINITY) {
907 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
908 if (next > p->rt.timeout) {
909 u64 next_time = p->se.sum_exec_runtime;
911 next_time += next * (NSEC_PER_SEC/HZ);
912 if (p->it_sched_expires > next_time)
913 p->it_sched_expires = next_time;
915 p->it_sched_expires = p->se.sum_exec_runtime;
919 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
926 * RR tasks need a special form of timeslice management.
927 * FIFO tasks have no timeslices.
929 if (p->policy != SCHED_RR)
932 if (--p->rt.time_slice)
935 p->rt.time_slice = DEF_TIMESLICE;
938 * Requeue to the end of queue if we are not the only element
941 if (p->rt.run_list.prev != p->rt.run_list.next) {
942 requeue_task_rt(rq, p);
943 set_tsk_need_resched(p);
947 static void set_curr_task_rt(struct rq *rq)
949 struct task_struct *p = rq->curr;
951 p->se.exec_start = rq->clock;
954 const struct sched_class rt_sched_class = {
955 .next = &fair_sched_class,
956 .enqueue_task = enqueue_task_rt,
957 .dequeue_task = dequeue_task_rt,
958 .yield_task = yield_task_rt,
960 .select_task_rq = select_task_rq_rt,
961 #endif /* CONFIG_SMP */
963 .check_preempt_curr = check_preempt_curr_rt,
965 .pick_next_task = pick_next_task_rt,
966 .put_prev_task = put_prev_task_rt,
969 .load_balance = load_balance_rt,
970 .move_one_task = move_one_task_rt,
971 .set_cpus_allowed = set_cpus_allowed_rt,
972 .join_domain = join_domain_rt,
973 .leave_domain = leave_domain_rt,
974 .pre_schedule = pre_schedule_rt,
975 .post_schedule = post_schedule_rt,
976 .task_wake_up = task_wake_up_rt,
977 .switched_from = switched_from_rt,
980 .set_curr_task = set_curr_task_rt,
981 .task_tick = task_tick_rt,
983 .prio_changed = prio_changed_rt,
984 .switched_to = switched_to_rt,