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 * Preempt the current task with a newly woken task if needed:
156 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
158 if (p->prio < rq->curr->prio)
159 resched_task(rq->curr);
162 static struct task_struct *pick_next_task_rt(struct rq *rq)
164 struct rt_prio_array *array = &rq->rt.active;
165 struct task_struct *next;
166 struct list_head *queue;
169 idx = sched_find_first_bit(array->bitmap);
170 if (idx >= MAX_RT_PRIO)
173 queue = array->queue + idx;
174 next = list_entry(queue->next, struct task_struct, run_list);
176 next->se.exec_start = rq->clock;
181 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
184 p->se.exec_start = 0;
188 /* Only try algorithms three times */
189 #define RT_MAX_TRIES 3
191 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
192 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
194 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
196 if (!task_running(rq, p) &&
197 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
198 (p->nr_cpus_allowed > 1))
203 /* Return the second highest RT task, NULL otherwise */
204 static struct task_struct *pick_next_highest_task_rt(struct rq *rq,
207 struct rt_prio_array *array = &rq->rt.active;
208 struct task_struct *next;
209 struct list_head *queue;
212 assert_spin_locked(&rq->lock);
214 if (likely(rq->rt.rt_nr_running < 2))
217 idx = sched_find_first_bit(array->bitmap);
218 if (unlikely(idx >= MAX_RT_PRIO)) {
219 WARN_ON(1); /* rt_nr_running is bad */
223 queue = array->queue + idx;
224 BUG_ON(list_empty(queue));
226 next = list_entry(queue->next, struct task_struct, run_list);
227 if (unlikely(pick_rt_task(rq, next, cpu)))
230 if (queue->next->next != queue) {
232 next = list_entry(queue->next->next, struct task_struct, run_list);
233 if (pick_rt_task(rq, next, cpu))
238 /* slower, but more flexible */
239 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
240 if (unlikely(idx >= MAX_RT_PRIO))
243 queue = array->queue + idx;
244 BUG_ON(list_empty(queue));
246 list_for_each_entry(next, queue, run_list) {
247 if (pick_rt_task(rq, next, cpu))
257 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
259 /* Will lock the rq it finds */
260 static struct rq *find_lock_lowest_rq(struct task_struct *task,
263 struct rq *lowest_rq = NULL;
266 cpumask_t *cpu_mask = &__get_cpu_var(local_cpu_mask);
268 cpus_and(*cpu_mask, cpu_online_map, task->cpus_allowed);
270 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
272 * Scan each rq for the lowest prio.
274 for_each_cpu_mask(cpu, *cpu_mask) {
275 struct rq *rq = &per_cpu(runqueues, cpu);
277 if (cpu == this_rq->cpu)
280 /* We look for lowest RT prio or non-rt CPU */
281 if (rq->rt.highest_prio >= MAX_RT_PRIO) {
286 /* no locking for now */
287 if (rq->rt.highest_prio > task->prio &&
288 (!lowest_rq || rq->rt.highest_prio > lowest_rq->rt.highest_prio)) {
296 /* if the prio of this runqueue changed, try again */
297 if (double_lock_balance(this_rq, lowest_rq)) {
299 * We had to unlock the run queue. In
300 * the mean time, task could have
301 * migrated already or had its affinity changed.
302 * Also make sure that it wasn't scheduled on its rq.
304 if (unlikely(task_rq(task) != this_rq ||
305 !cpu_isset(lowest_rq->cpu, task->cpus_allowed) ||
306 task_running(this_rq, task) ||
308 spin_unlock(&lowest_rq->lock);
314 /* If this rq is still suitable use it. */
315 if (lowest_rq->rt.highest_prio > task->prio)
319 spin_unlock(&lowest_rq->lock);
327 * If the current CPU has more than one RT task, see if the non
328 * running task can migrate over to a CPU that is running a task
329 * of lesser priority.
331 static int push_rt_task(struct rq *this_rq)
333 struct task_struct *next_task;
334 struct rq *lowest_rq;
336 int paranoid = RT_MAX_TRIES;
338 assert_spin_locked(&this_rq->lock);
340 next_task = pick_next_highest_task_rt(this_rq, -1);
345 if (unlikely(next_task == this_rq->curr)) {
351 * It's possible that the next_task slipped in of
352 * higher priority than current. If that's the case
353 * just reschedule current.
355 if (unlikely(next_task->prio < this_rq->curr->prio)) {
356 resched_task(this_rq->curr);
360 /* We might release this_rq lock */
361 get_task_struct(next_task);
363 /* find_lock_lowest_rq locks the rq if found */
364 lowest_rq = find_lock_lowest_rq(next_task, this_rq);
366 struct task_struct *task;
368 * find lock_lowest_rq releases this_rq->lock
369 * so it is possible that next_task has changed.
370 * If it has, then try again.
372 task = pick_next_highest_task_rt(this_rq, -1);
373 if (unlikely(task != next_task) && task && paranoid--) {
374 put_task_struct(next_task);
381 assert_spin_locked(&lowest_rq->lock);
383 deactivate_task(this_rq, next_task, 0);
384 set_task_cpu(next_task, lowest_rq->cpu);
385 activate_task(lowest_rq, next_task, 0);
387 resched_task(lowest_rq->curr);
389 spin_unlock(&lowest_rq->lock);
393 put_task_struct(next_task);
399 * TODO: Currently we just use the second highest prio task on
400 * the queue, and stop when it can't migrate (or there's
401 * no more RT tasks). There may be a case where a lower
402 * priority RT task has a different affinity than the
403 * higher RT task. In this case the lower RT task could
404 * possibly be able to migrate where as the higher priority
405 * RT task could not. We currently ignore this issue.
406 * Enhancements are welcome!
408 static void push_rt_tasks(struct rq *rq)
410 /* push_rt_task will return true if it moved an RT */
411 while (push_rt_task(rq))
415 static int pull_rt_task(struct rq *this_rq)
417 struct task_struct *next;
418 struct task_struct *p;
420 cpumask_t *rto_cpumask;
421 int this_cpu = this_rq->cpu;
425 assert_spin_locked(&this_rq->lock);
428 * If cpusets are used, and we have overlapping
429 * run queue cpusets, then this algorithm may not catch all.
430 * This is just the price you pay on trying to keep
431 * dirtying caches down on large SMP machines.
433 if (likely(!rt_overloaded()))
436 next = pick_next_task_rt(this_rq);
438 rto_cpumask = rt_overload();
440 for_each_cpu_mask(cpu, *rto_cpumask) {
444 src_rq = cpu_rq(cpu);
445 if (unlikely(src_rq->rt.rt_nr_running <= 1)) {
447 * It is possible that overlapping cpusets
448 * will miss clearing a non overloaded runqueue.
451 if (double_lock_balance(this_rq, src_rq)) {
452 /* unlocked our runqueue lock */
453 struct task_struct *old_next = next;
454 next = pick_next_task_rt(this_rq);
455 if (next != old_next)
458 if (likely(src_rq->rt.rt_nr_running <= 1))
460 * Small chance that this_rq->curr changed
461 * but it's really harmless here.
463 rt_clear_overload(this_rq);
466 * Heh, the src_rq is now overloaded, since
467 * we already have the src_rq lock, go straight
468 * to pulling tasks from it.
471 spin_unlock(&src_rq->lock);
476 * We can potentially drop this_rq's lock in
477 * double_lock_balance, and another CPU could
478 * steal our next task - hence we must cause
479 * the caller to recalculate the next task
482 if (double_lock_balance(this_rq, src_rq)) {
483 struct task_struct *old_next = next;
484 next = pick_next_task_rt(this_rq);
485 if (next != old_next)
490 * Are there still pullable RT tasks?
492 if (src_rq->rt.rt_nr_running <= 1) {
493 spin_unlock(&src_rq->lock);
498 p = pick_next_highest_task_rt(src_rq, this_cpu);
501 * Do we have an RT task that preempts
502 * the to-be-scheduled task?
504 if (p && (!next || (p->prio < next->prio))) {
505 WARN_ON(p == src_rq->curr);
506 WARN_ON(!p->se.on_rq);
509 * There's a chance that p is higher in priority
510 * than what's currently running on its cpu.
511 * This is just that p is wakeing up and hasn't
512 * had a chance to schedule. We only pull
513 * p if it is lower in priority than the
514 * current task on the run queue or
515 * this_rq next task is lower in prio than
516 * the current task on that rq.
518 if (p->prio < src_rq->curr->prio ||
519 (next && next->prio < src_rq->curr->prio))
524 deactivate_task(src_rq, p, 0);
525 set_task_cpu(p, this_cpu);
526 activate_task(this_rq, p, 0);
528 * We continue with the search, just in
529 * case there's an even higher prio task
530 * in another runqueue. (low likelyhood
535 * Update next so that we won't pick a task
536 * on another cpu with a priority lower (or equal)
537 * than the one we just picked.
543 spin_unlock(&src_rq->lock);
549 static void schedule_balance_rt(struct rq *rq,
550 struct task_struct *prev)
552 /* Try to pull RT tasks here if we lower this rq's prio */
553 if (unlikely(rt_task(prev)) &&
554 rq->rt.highest_prio > prev->prio)
558 static void schedule_tail_balance_rt(struct rq *rq)
561 * If we have more than one rt_task queued, then
562 * see if we can push the other rt_tasks off to other CPUS.
563 * Note we may release the rq lock, and since
564 * the lock was owned by prev, we need to release it
565 * first via finish_lock_switch and then reaquire it here.
567 if (unlikely(rq->rt.rt_nr_running > 1)) {
568 spin_lock_irq(&rq->lock);
570 spin_unlock_irq(&rq->lock);
575 static void wakeup_balance_rt(struct rq *rq, struct task_struct *p)
577 if (unlikely(rt_task(p)) &&
578 !task_running(rq, p) &&
579 (p->prio >= rq->curr->prio))
584 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
585 unsigned long max_load_move,
586 struct sched_domain *sd, enum cpu_idle_type idle,
587 int *all_pinned, int *this_best_prio)
589 /* don't touch RT tasks */
594 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
595 struct sched_domain *sd, enum cpu_idle_type idle)
597 /* don't touch RT tasks */
600 static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
602 int weight = cpus_weight(*new_mask);
607 * Update the migration status of the RQ if we have an RT task
608 * which is running AND changing its weight value.
610 if (p->se.on_rq && (weight != p->nr_cpus_allowed)) {
611 struct rq *rq = task_rq(p);
613 if ((p->nr_cpus_allowed <= 1) && (weight > 1))
614 rq->rt.rt_nr_migratory++;
615 else if((p->nr_cpus_allowed > 1) && (weight <= 1)) {
616 BUG_ON(!rq->rt.rt_nr_migratory);
617 rq->rt.rt_nr_migratory--;
620 update_rt_migration(rq);
623 p->cpus_allowed = *new_mask;
624 p->nr_cpus_allowed = weight;
626 #else /* CONFIG_SMP */
627 # define schedule_tail_balance_rt(rq) do { } while (0)
628 # define schedule_balance_rt(rq, prev) do { } while (0)
629 # define wakeup_balance_rt(rq, p) do { } while (0)
630 #endif /* CONFIG_SMP */
632 static void task_tick_rt(struct rq *rq, struct task_struct *p)
637 * RR tasks need a special form of timeslice management.
638 * FIFO tasks have no timeslices.
640 if (p->policy != SCHED_RR)
646 p->time_slice = DEF_TIMESLICE;
649 * Requeue to the end of queue if we are not the only element
652 if (p->run_list.prev != p->run_list.next) {
653 requeue_task_rt(rq, p);
654 set_tsk_need_resched(p);
658 static void set_curr_task_rt(struct rq *rq)
660 struct task_struct *p = rq->curr;
662 p->se.exec_start = rq->clock;
665 const struct sched_class rt_sched_class = {
666 .next = &fair_sched_class,
667 .enqueue_task = enqueue_task_rt,
668 .dequeue_task = dequeue_task_rt,
669 .yield_task = yield_task_rt,
671 .check_preempt_curr = check_preempt_curr_rt,
673 .pick_next_task = pick_next_task_rt,
674 .put_prev_task = put_prev_task_rt,
677 .load_balance = load_balance_rt,
678 .move_one_task = move_one_task_rt,
679 .set_cpus_allowed = set_cpus_allowed_rt,
682 .set_curr_task = set_curr_task_rt,
683 .task_tick = task_tick_rt,