1 /*
2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3 * policies)
4 */
5
6 #ifdef CONFIG_SMP
7
8 static inline int rt_overloaded(struct rq *rq)
9 {
10 return atomic_read(&rq->rd->rto_count);
11 }
12
13 static inline void rt_set_overload(struct rq *rq)
14 {
15 if (!rq->online)
16 return;
17
18 cpu_set(rq->cpu, rq->rd->rto_mask);
19 /*
20 * Make sure the mask is visible before we set
21 * the overload count. That is checked to determine
22 * if we should look at the mask. It would be a shame
23 * if we looked at the mask, but the mask was not
24 * updated yet.
25 */
26 wmb();
27 atomic_inc(&rq->rd->rto_count);
28 }
29
30 static inline void rt_clear_overload(struct rq *rq)
31 {
32 if (!rq->online)
33 return;
34
35 /* the order here really doesn't matter */
36 atomic_dec(&rq->rd->rto_count);
37 cpu_clear(rq->cpu, rq->rd->rto_mask);
38 }
39
40 static void update_rt_migration(struct rq *rq)
41 {
42 if (unlikely(num_online_cpus() == 1))
43 return;
44
45 if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
46 if (!rq->rt.overloaded) {
47 rt_set_overload(rq);
48 rq->rt.overloaded = 1;
49 }
50 } else if (rq->rt.overloaded) {
51 rt_clear_overload(rq);
52 rq->rt.overloaded = 0;
53 }
54 }
55 #endif /* CONFIG_SMP */
56
57 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
58 {
59 return container_of(rt_se, struct task_struct, rt);
60 }
61
62 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
63 {
64 return !list_empty(&rt_se->run_list);
65 }
66
67 #ifdef CONFIG_RT_GROUP_SCHED
68
69 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
70 {
71 if (!rt_rq->tg)
72 return RUNTIME_INF;
73
74 return rt_rq->tg->rt_runtime;
75 }
76
77 #define for_each_leaf_rt_rq(rt_rq, rq) \
78 list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
79
80 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
81 {
82 return rt_rq->rq;
83 }
84
85 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
86 {
87 return rt_se->rt_rq;
88 }
89
90 #define for_each_sched_rt_entity(rt_se) \
91 for (; rt_se; rt_se = rt_se->parent)
92
93 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
94 {
95 return rt_se->my_q;
96 }
97
98 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
99 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
100
101 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
102 {
103 struct sched_rt_entity *rt_se = rt_rq->rt_se;
104
105 if (rt_se && !on_rt_rq(rt_se) && rt_rq->rt_nr_running) {
106 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
107
108 enqueue_rt_entity(rt_se);
109 if (rt_rq->highest_prio < curr->prio)
110 resched_task(curr);
111 }
112 }
113
114 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
115 {
116 struct sched_rt_entity *rt_se = rt_rq->rt_se;
117
118 if (rt_se && on_rt_rq(rt_se))
119 dequeue_rt_entity(rt_se);
120 }
121
122 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
123 {
124 return 0;
125 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
126 }
127
128 static int rt_se_boosted(struct sched_rt_entity *rt_se)
129 {
130 struct rt_rq *rt_rq = group_rt_rq(rt_se);
131 struct task_struct *p;
132
133 if (rt_rq)
134 return !!rt_rq->rt_nr_boosted;
135
136 p = rt_task_of(rt_se);
137 return p->prio != p->normal_prio;
138 }
139
140 #else
141
142 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
143 {
144 if (sysctl_sched_rt_runtime == -1)
145 return RUNTIME_INF;
146
147 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
148 }
149
150 #define for_each_leaf_rt_rq(rt_rq, rq) \
151 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
152
153 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
154 {
155 return container_of(rt_rq, struct rq, rt);
156 }
157
158 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
159 {
160 struct task_struct *p = rt_task_of(rt_se);
161 struct rq *rq = task_rq(p);
162
163 return &rq->rt;
164 }
165
166 #define for_each_sched_rt_entity(rt_se) \
167 for (; rt_se; rt_se = NULL)
168
169 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
170 {
171 return NULL;
172 }
173
174 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
175 {
176 }
177
178 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
179 {
180 }
181
182 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
183 {
184 return 0;
185 return rt_rq->rt_throttled;
186 }
187 #endif
188
189 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
190 {
191 #ifdef CONFIG_RT_GROUP_SCHED
192 struct rt_rq *rt_rq = group_rt_rq(rt_se);
193
194 if (rt_rq)
195 return rt_rq->highest_prio;
196 #endif
197
198 return rt_task_of(rt_se)->prio;
199 }
200
201 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
202 {
203 u64 runtime = sched_rt_runtime(rt_rq);
204
205 return 0;
206 if (runtime == RUNTIME_INF)
207 return 0;
208
209 if (rt_rq->rt_throttled)
210 return rt_rq_throttled(rt_rq);
211
212 if (rt_rq->rt_time > runtime) {
213 struct rq *rq = rq_of_rt_rq(rt_rq);
214
215 rq->rt_throttled = 1;
216 rt_rq->rt_throttled = 1;
217
218 if (rt_rq_throttled(rt_rq)) {
219 sched_rt_rq_dequeue(rt_rq);
220 return 1;
221 }
222 }
223
224 return 0;
225 }
226
227 static void update_sched_rt_period(struct rq *rq)
228 {
229 struct rt_rq *rt_rq;
230 u64 period;
231
232 return;
233 while (rq->clock > rq->rt_period_expire) {
234 period = (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
235 rq->rt_period_expire += period;
236
237 for_each_leaf_rt_rq(rt_rq, rq) {
238 u64 runtime = sched_rt_runtime(rt_rq);
239
240 rt_rq->rt_time -= min(rt_rq->rt_time, runtime);
241 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
242 rt_rq->rt_throttled = 0;
243 sched_rt_rq_enqueue(rt_rq);
244 }
245 }
246
247 rq->rt_throttled = 0;
248 }
249 }
250
251 /*
252 * Update the current task's runtime statistics. Skip current tasks that
253 * are not in our scheduling class.
254 */
255 static void update_curr_rt(struct rq *rq)
256 {
257 struct task_struct *curr = rq->curr;
258 struct sched_rt_entity *rt_se = &curr->rt;
259 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
260 u64 delta_exec;
261
262 if (!task_has_rt_policy(curr))
263 return;
264
265 delta_exec = rq->clock - curr->se.exec_start;
266 if (unlikely((s64)delta_exec < 0))
267 delta_exec = 0;
268
269 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
270
271 curr->se.sum_exec_runtime += delta_exec;
272 curr->se.exec_start = rq->clock;
273 cpuacct_charge(curr, delta_exec);
274
275 // rt_rq->rt_time += delta_exec;
276 if (sched_rt_runtime_exceeded(rt_rq))
277 resched_task(curr);
278 }
279
280 static inline
281 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
282 {
283 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
284 rt_rq->rt_nr_running++;
285 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
286 if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
287 struct rq *rq = rq_of_rt_rq(rt_rq);
288 rt_rq->highest_prio = rt_se_prio(rt_se);
289
290 if (rq->online)
291 cpupri_set(&rq->rd->cpupri, rq->cpu,
292 rt_se_prio(rt_se));
293 }
294 #endif
295 #ifdef CONFIG_SMP
296 if (rt_se->nr_cpus_allowed > 1) {
297 struct rq *rq = rq_of_rt_rq(rt_rq);
298 rq->rt.rt_nr_migratory++;
299 }
300
301 update_rt_migration(rq_of_rt_rq(rt_rq));
302 #endif
303 #ifdef CONFIG_RT_GROUP_SCHED
304 if (rt_se_boosted(rt_se))
305 rt_rq->rt_nr_boosted++;
306 #endif
307 }
308
309 static inline
310 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
311 {
312 #ifdef CONFIG_SMP
313 int highest_prio = rt_rq->highest_prio;
314 #endif
315
316 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
317 WARN_ON(!rt_rq->rt_nr_running);
318 rt_rq->rt_nr_running--;
319 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
320 if (rt_rq->rt_nr_running) {
321 struct rt_prio_array *array;
322
323 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
324 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
325 /* recalculate */
326 array = &rt_rq->active;
327 rt_rq->highest_prio =
328 sched_find_first_bit(array->bitmap);
329 } /* otherwise leave rq->highest prio alone */
330 } else
331 rt_rq->highest_prio = MAX_RT_PRIO;
332 #endif
333 #ifdef CONFIG_SMP
334 if (rt_se->nr_cpus_allowed > 1) {
335 struct rq *rq = rq_of_rt_rq(rt_rq);
336 BUG_ON(!rq->rt.rt_nr_migratory);
337 rq->rt.rt_nr_migratory--;
338 }
339
340 if (rt_rq->highest_prio != highest_prio) {
341 struct rq *rq = rq_of_rt_rq(rt_rq);
342
343 if (rq->online)
344 cpupri_set(&rq->rd->cpupri, rq->cpu,
345 rt_rq->highest_prio);
346 }
347
348 update_rt_migration(rq_of_rt_rq(rt_rq));
349 #endif /* CONFIG_SMP */
350 #ifdef CONFIG_RT_GROUP_SCHED
351 if (rt_se_boosted(rt_se))
352 rt_rq->rt_nr_boosted--;
353
354 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
355 #endif
356 }
357
358 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
359 {
360 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
361 struct rt_prio_array *array = &rt_rq->active;
362 struct rt_rq *group_rq = group_rt_rq(rt_se);
363
364 if (group_rq && rt_rq_throttled(group_rq))
365 return;
366
367 if (rt_se->nr_cpus_allowed == 1)
368 list_add_tail(&rt_se->run_list,
369 array->xqueue + rt_se_prio(rt_se));
370 else
371 list_add_tail(&rt_se->run_list,
372 array->squeue + rt_se_prio(rt_se));
373
374 __set_bit(rt_se_prio(rt_se), array->bitmap);
375
376 inc_rt_tasks(rt_se, rt_rq);
377 }
378
379 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
380 {
381 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
382 struct rt_prio_array *array = &rt_rq->active;
383
384 list_del_init(&rt_se->run_list);
385 if (list_empty(array->squeue + rt_se_prio(rt_se))
386 && list_empty(array->xqueue + rt_se_prio(rt_se)))
387 __clear_bit(rt_se_prio(rt_se), array->bitmap);
388
389 dec_rt_tasks(rt_se, rt_rq);
390 }
391
392 /*
393 * Because the prio of an upper entry depends on the lower
394 * entries, we must remove entries top - down.
395 *
396 * XXX: O(1/2 h^2) because we can only walk up, not down the chain.
397 * doesn't matter much for now, as h=2 for GROUP_SCHED.
398 */
399 static void dequeue_rt_stack(struct task_struct *p)
400 {
401 struct sched_rt_entity *rt_se, *top_se;
402
403 /*
404 * dequeue all, top - down.
405 */
406 do {
407 rt_se = &p->rt;
408 top_se = NULL;
409 for_each_sched_rt_entity(rt_se) {
410 if (on_rt_rq(rt_se))
411 top_se = rt_se;
412 }
413 if (top_se)
414 dequeue_rt_entity(top_se);
415 } while (top_se);
416 }
417
418 static inline void incr_rt_nr_uninterruptible(struct task_struct *p,
419 struct rq *rq)
420 {
421 rq->rt.rt_nr_uninterruptible++;
422 }
423
424 static inline void decr_rt_nr_uninterruptible(struct task_struct *p,
425 struct rq *rq)
426 {
427 rq->rt.rt_nr_uninterruptible--;
428 }
429
430 unsigned long rt_nr_running(void)
431 {
432 unsigned long i, sum = 0;
433
434 for_each_online_cpu(i)
435 sum += cpu_rq(i)->rt.rt_nr_running;
436
437 return sum;
438 }
439
440 unsigned long rt_nr_running_cpu(int cpu)
441 {
442 return cpu_rq(cpu)->rt.rt_nr_running;
443 }
444
445 unsigned long rt_nr_uninterruptible(void)
446 {
447 unsigned long i, sum = 0;
448
449 for_each_online_cpu(i)
450 sum += cpu_rq(i)->rt.rt_nr_uninterruptible;
451
452 /*
453 * Since we read the counters lockless, it might be slightly
454 * inaccurate. Do not allow it to go below zero though:
455 */
456 if (unlikely((long)sum < 0))
457 sum = 0;
458
459 return sum;
460 }
461
462 unsigned long rt_nr_uninterruptible_cpu(int cpu)
463 {
464 return cpu_rq(cpu)->rt.rt_nr_uninterruptible;
465 }
466
467 /*
468 * Adding/removing a task to/from a priority array:
469 */
470 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
471 {
472 struct sched_rt_entity *rt_se = &p->rt;
473
474 if (wakeup)
475 rt_se->timeout = 0;
476
477 dequeue_rt_stack(p);
478
479 /*
480 * enqueue everybody, bottom - up.
481 */
482 for_each_sched_rt_entity(rt_se)
483 enqueue_rt_entity(rt_se);
484
485 if (p->state == TASK_UNINTERRUPTIBLE)
486 decr_rt_nr_uninterruptible(p, rq);
487 }
488
489 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
490 {
491 struct sched_rt_entity *rt_se = &p->rt;
492 struct rt_rq *rt_rq;
493
494 update_curr_rt(rq);
495
496 if (p->state == TASK_UNINTERRUPTIBLE)
497 incr_rt_nr_uninterruptible(p, rq);
498
499 dequeue_rt_stack(p);
500
501 /*
502 * re-enqueue all non-empty rt_rq entities.
503 */
504 for_each_sched_rt_entity(rt_se) {
505 rt_rq = group_rt_rq(rt_se);
506 if (rt_rq && rt_rq->rt_nr_running)
507 enqueue_rt_entity(rt_se);
508 }
509 }
510
511 /*
512 * Put task to the end of the run list without the overhead of dequeue
513 * followed by enqueue.
514 *
515 * Note: We always enqueue the task to the shared-queue, regardless of its
516 * previous position w.r.t. exclusive vs shared. This is so that exclusive RR
517 * tasks fairly round-robin with all tasks on the runqueue, not just other
518 * exclusive tasks.
519 */
520 static
521 void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
522 {
523 struct rt_prio_array *array = &rt_rq->active;
524
525 list_del_init(&rt_se->run_list);
526 list_add_tail(&rt_se->run_list, array->squeue + rt_se_prio(rt_se));
527 }
528
529 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
530 {
531 struct sched_rt_entity *rt_se = &p->rt;
532 struct rt_rq *rt_rq;
533
534 for_each_sched_rt_entity(rt_se) {
535 rt_rq = rt_rq_of_se(rt_se);
536 requeue_rt_entity(rt_rq, rt_se);
537 }
538 }
539
540 static void yield_task_rt(struct rq *rq)
541 {
542 requeue_task_rt(rq, rq->curr);
543 }
544
545 #ifdef CONFIG_SMP
546 static int find_lowest_rq(struct task_struct *task);
547
548 static int select_task_rq_rt(struct task_struct *p, int sync)
549 {
550 struct rq *rq = task_rq(p);
551
552 /*
553 * If the current task is an RT task, then
554 * try to see if we can wake this RT task up on another
555 * runqueue. Otherwise simply start this RT task
556 * on its current runqueue.
557 *
558 * We want to avoid overloading runqueues. Even if
559 * the RT task is of higher priority than the current RT task.
560 * RT tasks behave differently than other tasks. If
561 * one gets preempted, we try to push it off to another queue.
562 * So trying to keep a preempting RT task on the same
563 * cache hot CPU will force the running RT task to
564 * a cold CPU. So we waste all the cache for the lower
565 * RT task in hopes of saving some of a RT task
566 * that is just being woken and probably will have
567 * cold cache anyway.
568 */
569 if (unlikely(rt_task(rq->curr)) &&
570 (p->rt.nr_cpus_allowed > 1)) {
571 int cpu = find_lowest_rq(p);
572
573 return (cpu == -1) ? task_cpu(p) : cpu;
574 }
575
576 /*
577 * Otherwise, just let it ride on the affined RQ and the
578 * post-schedule router will push the preempted task away
579 */
580 return task_cpu(p);
581 }
582 #endif /* CONFIG_SMP */
583
584 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
585 struct rt_rq *rt_rq);
586
587 /*
588 * Preempt the current task with a newly woken task if needed:
589 */
590 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
591 {
592 if (p->prio < rq->curr->prio) {
593 resched_task(rq->curr);
594 return;
595 }
596
597 #ifdef CONFIG_SMP
598 /*
599 * If:
600 *
601 * - the newly woken task is of equal priority to the current task
602 * - the newly woken task is non-migratable while current is migratable
603 * - current will be preempted on the next reschedule
604 *
605 * we should check to see if current can readily move to a different
606 * cpu. If so, we will reschedule to allow the push logic to try
607 * to move current somewhere else, making room for our non-migratable
608 * task.
609 */
610 if((p->prio == rq->curr->prio)
611 && p->rt.nr_cpus_allowed == 1
612 && rq->curr->rt.nr_cpus_allowed != 1
613 && pick_next_rt_entity(rq, &rq->rt) != &rq->curr->rt) {
614 cpumask_t mask;
615
616 if (cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
617 /*
618 * There appears to be other cpus that can accept
619 * current, so lets reschedule to try and push it away
620 */
621 resched_task(rq->curr);
622 }
623 #endif
624 }
625
626 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
627 struct rt_rq *rt_rq)
628 {
629 struct rt_prio_array *array = &rt_rq->active;
630 struct sched_rt_entity *next = NULL;
631 struct list_head *queue;
632 int idx;
633
634 idx = sched_find_first_bit(array->bitmap);
635 BUG_ON(idx >= MAX_RT_PRIO);
636
637 queue = array->xqueue + idx;
638 if (!list_empty(queue))
639 next = list_entry(queue->next, struct sched_rt_entity,
640 run_list);
641 else {
642 queue = array->squeue + idx;
643 next = list_entry(queue->next, struct sched_rt_entity,
644 run_list);
645 }
646
647 return next;
648 }
649
650 static struct task_struct *pick_next_task_rt(struct rq *rq)
651 {
652 struct sched_rt_entity *rt_se;
653 struct task_struct *p;
654 struct rt_rq *rt_rq;
655
656 rt_rq = &rq->rt;
657
658 if (unlikely(!rt_rq->rt_nr_running))
659 return NULL;
660
661 if (rt_rq_throttled(rt_rq))
662 return NULL;
663
664 do {
665 rt_se = pick_next_rt_entity(rq, rt_rq);
666 BUG_ON(!rt_se);
667 rt_rq = group_rt_rq(rt_se);
668 } while (rt_rq);
669
670 p = rt_task_of(rt_se);
671 p->se.exec_start = rq->clock;
672 return p;
673 }
674
675 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
676 {
677 update_curr_rt(rq);
678 p->se.exec_start = 0;
679 }
680
681 #ifdef CONFIG_SMP
682
683 /* Only try algorithms three times */
684 #define RT_MAX_TRIES 3
685
686 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
687 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
688
689 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
690 {
691 if (!task_running(rq, p) &&
692 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
693 (p->rt.nr_cpus_allowed > 1))
694 return 1;
695 return 0;
696 }
697
698 /* Return the second highest RT task, NULL otherwise */
699 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
700 {
701 struct task_struct *next = NULL;
702 struct sched_rt_entity *rt_se;
703 struct rt_prio_array *array;
704 struct rt_rq *rt_rq;
705 int idx;
706
707 for_each_leaf_rt_rq(rt_rq, rq) {
708 array = &rt_rq->active;
709 idx = sched_find_first_bit(array->bitmap);
710 next_idx:
711 if (idx >= MAX_RT_PRIO)
712 continue;
713 if (next && next->prio < idx)
714 continue;
715 list_for_each_entry(rt_se, array->squeue + idx, run_list) {
716 struct task_struct *p = rt_task_of(rt_se);
717 if (pick_rt_task(rq, p, cpu)) {
718 next = p;
719 break;
720 }
721 }
722 if (!next) {
723 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
724 goto next_idx;
725 }
726 }
727
728 return next;
729 }
730
731 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
732
733 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
734 {
735 int first;
736
737 /* "this_cpu" is cheaper to preempt than a remote processor */
738 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
739 return this_cpu;
740
741 first = first_cpu(*mask);
742 if (first != NR_CPUS)
743 return first;
744
745 return -1;
746 }
747
748 static int find_lowest_rq(struct task_struct *task)
749 {
750 struct sched_domain *sd;
751 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
752 int this_cpu = smp_processor_id();
753 int cpu = task_cpu(task);
754
755 if (task->rt.nr_cpus_allowed == 1)
756 return -1; /* No other targets possible */
757
758 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
759 return -1; /* No targets found */
760
761 /*
762 * At this point we have built a mask of cpus representing the
763 * lowest priority tasks in the system. Now we want to elect
764 * the best one based on our affinity and topology.
765 *
766 * We prioritize the last cpu that the task executed on since
767 * it is most likely cache-hot in that location.
768 */
769 if (cpu_isset(cpu, *lowest_mask))
770 return cpu;
771
772 /*
773 * Otherwise, we consult the sched_domains span maps to figure
774 * out which cpu is logically closest to our hot cache data.
775 */
776 if (this_cpu == cpu)
777 this_cpu = -1; /* Skip this_cpu opt if the same */
778
779 for_each_domain(cpu, sd) {
780 if (sd->flags & SD_WAKE_AFFINE) {
781 cpumask_t domain_mask;
782 int best_cpu;
783
784 cpus_and(domain_mask, sd->span, *lowest_mask);
785
786 best_cpu = pick_optimal_cpu(this_cpu,
787 &domain_mask);
788 if (best_cpu != -1)
789 return best_cpu;
790 }
791 }
792
793 /*
794 * And finally, if there were no matches within the domains
795 * just give the caller *something* to work with from the compatible
796 * locations.
797 */
798 return pick_optimal_cpu(this_cpu, lowest_mask);
799 }
800
801 /* Will lock the rq it finds */
802 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
803 {
804 struct rq *lowest_rq = NULL;
805 int tries;
806 int cpu;
807
808 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
809 cpu = find_lowest_rq(task);
810
811 if ((cpu == -1) || (cpu == rq->cpu))
812 break;
813
814 lowest_rq = cpu_rq(cpu);
815
816 /* if the prio of this runqueue changed, try again */
817 if (double_lock_balance(rq, lowest_rq)) {
818 /*
819 * We had to unlock the run queue. In
820 * the mean time, task could have
821 * migrated already or had its affinity changed.
822 * Also make sure that it wasn't scheduled on its rq.
823 */
824 if (unlikely(task_rq(task) != rq ||
825 !cpu_isset(lowest_rq->cpu,
826 task->cpus_allowed) ||
827 task_running(rq, task) ||
828 !task->se.on_rq)) {
829
830 spin_unlock(&lowest_rq->lock);
831 lowest_rq = NULL;
832 break;
833 }
834 }
835
836 /* If this rq is still suitable use it. */
837 if (lowest_rq->rt.highest_prio > task->prio)
838 break;
839
840 /* try again */
841 spin_unlock(&lowest_rq->lock);
842 lowest_rq = NULL;
843 }
844
845 return lowest_rq;
846 }
847
848 /*
849 * If the current CPU has more than one RT task, see if the non
850 * running task can migrate over to a CPU that is running a task
851 * of lesser priority.
852 */
853 static int push_rt_task(struct rq *rq)
854 {
855 struct task_struct *next_task;
856 struct rq *lowest_rq;
857 int ret = 0;
858 int paranoid = RT_MAX_TRIES;
859
860 if (!rq->rt.overloaded)
861 return 0;
862
863 next_task = pick_next_highest_task_rt(rq, -1);
864 if (!next_task)
865 return 0;
866
867 retry:
868 if (unlikely(next_task == rq->curr)) {
869 WARN_ON(1);
870 return 0;
871 }
872
873 /*
874 * It's possible that the next_task slipped in of
875 * higher priority than current. If that's the case
876 * just reschedule current.
877 */
878 if (unlikely(next_task->prio < rq->curr->prio)) {
879 resched_task(rq->curr);
880 return 0;
881 }
882
883 /* We might release rq lock */
884 get_task_struct(next_task);
885
886 /* find_lock_lowest_rq locks the rq if found */
887 lowest_rq = find_lock_lowest_rq(next_task, rq);
888 if (!lowest_rq) {
889 struct task_struct *task;
890 /*
891 * find lock_lowest_rq releases rq->lock
892 * so it is possible that next_task has changed.
893 * If it has, then try again.
894 */
895 task = pick_next_highest_task_rt(rq, -1);
896 if (unlikely(task != next_task) && task && paranoid--) {
897 put_task_struct(next_task);
898 next_task = task;
899 goto retry;
900 }
901 goto out;
902 }
903
904 deactivate_task(rq, next_task, 0);
905 set_task_cpu(next_task, lowest_rq->cpu);
906 activate_task(lowest_rq, next_task, 0);
907
908 resched_task(lowest_rq->curr);
909
910 schedstat_inc(rq, rto_pushed);
911
912 spin_unlock(&lowest_rq->lock);
913
914 ret = 1;
915 out:
916 put_task_struct(next_task);
917
918 return ret;
919 }
920
921 /*
922 * TODO: Currently we just use the second highest prio task on
923 * the queue, and stop when it can't migrate (or there's
924 * no more RT tasks). There may be a case where a lower
925 * priority RT task has a different affinity than the
926 * higher RT task. In this case the lower RT task could
927 * possibly be able to migrate where as the higher priority
928 * RT task could not. We currently ignore this issue.
929 * Enhancements are welcome!
930 */
931 static void push_rt_tasks(struct rq *rq)
932 {
933 /* push_rt_task will return true if it moved an RT */
934 while (push_rt_task(rq))
935 ;
936 }
937
938 static int pull_rt_task(struct rq *this_rq)
939 {
940 int this_cpu = this_rq->cpu, ret = 0, cpu;
941 struct task_struct *p, *next;
942 struct rq *src_rq;
943
944 if (likely(!rt_overloaded(this_rq)))
945 return 0;
946
947 next = pick_next_task_rt(this_rq);
948
949 for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
950 if (this_cpu == cpu)
951 continue;
952
953 src_rq = cpu_rq(cpu);
954 /*
955 * We can potentially drop this_rq's lock in
956 * double_lock_balance, and another CPU could
957 * steal our next task - hence we must cause
958 * the caller to recalculate the next task
959 * in that case:
960 */
961 if (double_lock_balance(this_rq, src_rq)) {
962 struct task_struct *old_next = next;
963
964 next = pick_next_task_rt(this_rq);
965 if (next != old_next)
966 ret = 1;
967 }
968
969 /*
970 * Are there still pullable RT tasks?
971 */
972 if (src_rq->rt.rt_nr_running <= 1)
973 goto skip;
974
975 p = pick_next_highest_task_rt(src_rq, this_cpu);
976
977 /*
978 * Do we have an RT task that preempts
979 * the to-be-scheduled task?
980 */
981 if (p && (!next || (p->prio < next->prio))) {
982 WARN_ON(p == src_rq->curr);
983 WARN_ON(!p->se.on_rq);
984
985 /*
986 * There's a chance that p is higher in priority
987 * than what's currently running on its cpu.
988 * This is just that p is wakeing up and hasn't
989 * had a chance to schedule. We only pull
990 * p if it is lower in priority than the
991 * current task on the run queue or
992 * this_rq next task is lower in prio than
993 * the current task on that rq.
994 */
995 if (p->prio < src_rq->curr->prio ||
996 (next && next->prio < src_rq->curr->prio))
997 goto skip;
998
999 ret = 1;
1000
1001 deactivate_task(src_rq, p, 0);
1002 set_task_cpu(p, this_cpu);
1003 activate_task(this_rq, p, 0);
1004 /*
1005 * We continue with the search, just in
1006 * case there's an even higher prio task
1007 * in another runqueue. (low likelyhood
1008 * but possible)
1009 *
1010 * Update next so that we won't pick a task
1011 * on another cpu with a priority lower (or equal)
1012 * than the one we just picked.
1013 */
1014 next = p;
1015
1016 schedstat_inc(src_rq, rto_pulled);
1017 }
1018 skip:
1019 spin_unlock(&src_rq->lock);
1020 }
1021
1022 return ret;
1023 }
1024
1025 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1026 {
1027 /* Try to pull RT tasks here if we lower this rq's prio */
1028 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio) {
1029 pull_rt_task(rq);
1030 schedstat_inc(rq, rto_schedule);
1031 }
1032 }
1033
1034 static void post_schedule_rt(struct rq *rq)
1035 {
1036 /*
1037 * If we have more than one rt_task queued, then
1038 * see if we can push the other rt_tasks off to other CPUS.
1039 * Note we may release the rq lock, and since
1040 * the lock was owned by prev, we need to release it
1041 * first via finish_lock_switch and then reaquire it here.
1042 */
1043 if (unlikely(rq->rt.overloaded)) {
1044 spin_lock(&rq->lock);
1045 push_rt_tasks(rq);
1046 schedstat_inc(rq, rto_schedule_tail);
1047 spin_unlock(&rq->lock);
1048 }
1049 }
1050
1051
1052 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1053 {
1054 if (!task_running(rq, p) &&
1055 !test_tsk_need_resched(rq->curr) &&
1056 rq->rt.overloaded) {
1057 push_rt_tasks(rq);
1058 schedstat_inc(rq, rto_wakeup);
1059 }
1060 }
1061
1062 static unsigned long
1063 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1064 unsigned long max_load_move,
1065 struct sched_domain *sd, enum cpu_idle_type idle,
1066 int *all_pinned, int *this_best_prio)
1067 {
1068 /* don't touch RT tasks */
1069 return 0;
1070 }
1071
1072 static int
1073 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1074 struct sched_domain *sd, enum cpu_idle_type idle)
1075 {
1076 /* don't touch RT tasks */
1077 return 0;
1078 }
1079
1080 static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
1081 {
1082 int weight = cpus_weight(*new_mask);
1083
1084 BUG_ON(!rt_task(p));
1085
1086 /*
1087 * Update the migration status of the RQ if we have an RT task
1088 * which is running AND changing its weight value.
1089 */
1090 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1091 struct rq *rq = task_rq(p);
1092
1093 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1094 rq->rt.rt_nr_migratory++;
1095 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1096 BUG_ON(!rq->rt.rt_nr_migratory);
1097 rq->rt.rt_nr_migratory--;
1098 }
1099
1100 update_rt_migration(rq);
1101
1102 if (unlikely(weight == 1 || p->rt.nr_cpus_allowed == 1))
1103 /*
1104 * If either the new or old weight is a "1", we need
1105 * to requeue to properly move between shared and
1106 * exclusive queues.
1107 */
1108 requeue_task_rt(rq, p);
1109 }
1110
1111 p->cpus_allowed = *new_mask;
1112 p->rt.nr_cpus_allowed = weight;
1113 }
1114
1115 /* Assumes rq->lock is held */
1116 static void rq_online_rt(struct rq *rq)
1117 {
1118 if (rq->rt.overloaded)
1119 rt_set_overload(rq);
1120
1121 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1122 }
1123
1124 /* Assumes rq->lock is held */
1125 static void rq_offline_rt(struct rq *rq)
1126 {
1127 if (rq->rt.overloaded)
1128 rt_clear_overload(rq);
1129
1130 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1131 }
1132
1133 /*
1134 * When switch from the rt queue, we bring ourselves to a position
1135 * that we might want to pull RT tasks from other runqueues.
1136 */
1137 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1138 int running)
1139 {
1140 /*
1141 * If there are other RT tasks then we will reschedule
1142 * and the scheduling of the other RT tasks will handle
1143 * the balancing. But if we are the last RT task
1144 * we may need to handle the pulling of RT tasks
1145 * now.
1146 */
1147 if (!rq->rt.rt_nr_running)
1148 pull_rt_task(rq);
1149 }
1150 #endif /* CONFIG_SMP */
1151
1152 /*
1153 * When switching a task to RT, we may overload the runqueue
1154 * with RT tasks. In this case we try to push them off to
1155 * other runqueues.
1156 */
1157 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1158 int running)
1159 {
1160 int check_resched = 1;
1161
1162 /*
1163 * If we are already running, then there's nothing
1164 * that needs to be done. But if we are not running
1165 * we may need to preempt the current running task.
1166 * If that current running task is also an RT task
1167 * then see if we can move to another run queue.
1168 */
1169 if (!running) {
1170 #ifdef CONFIG_SMP
1171 if (rq->rt.overloaded && push_rt_task(rq) &&
1172 /* Don't resched if we changed runqueues */
1173 rq != task_rq(p))
1174 check_resched = 0;
1175 #endif /* CONFIG_SMP */
1176 if (check_resched && p->prio < rq->curr->prio)
1177 resched_task(rq->curr);
1178 }
1179 }
1180
1181 /*
1182 * Priority of the task has changed. This may cause
1183 * us to initiate a push or pull.
1184 */
1185 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1186 int oldprio, int running)
1187 {
1188 if (running) {
1189 #ifdef CONFIG_SMP
1190 /*
1191 * If our priority decreases while running, we
1192 * may need to pull tasks to this runqueue.
1193 */
1194 if (oldprio < p->prio)
1195 pull_rt_task(rq);
1196 /*
1197 * If there's a higher priority task waiting to run
1198 * then reschedule. Note, the above pull_rt_task
1199 * can release the rq lock and p could migrate.
1200 * Only reschedule if p is still on the same runqueue.
1201 */
1202 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1203 resched_task(p);
1204 #else
1205 /* For UP simply resched on drop of prio */
1206 if (oldprio < p->prio)
1207 resched_task(p);
1208 #endif /* CONFIG_SMP */
1209 } else {
1210 /*
1211 * This task is not running, but if it is
1212 * greater than the current running task
1213 * then reschedule.
1214 */
1215 if (p->prio < rq->curr->prio)
1216 resched_task(rq->curr);
1217 }
1218 }
1219
1220 static void watchdog(struct rq *rq, struct task_struct *p)
1221 {
1222 unsigned long soft, hard;
1223
1224 if (!p->signal)
1225 return;
1226
1227 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1228 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1229
1230 if (soft != RLIM_INFINITY) {
1231 unsigned long next;
1232
1233 p->rt.timeout++;
1234 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1235 if (p->rt.timeout > next)
1236 p->it_sched_expires = p->se.sum_exec_runtime;
1237 }
1238 }
1239
1240 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1241 {
1242 update_curr_rt(rq);
1243
1244 watchdog(rq, p);
1245
1246 /*
1247 * RR tasks need a special form of timeslice management.
1248 * FIFO tasks have no timeslices.
1249 */
1250 if (p->policy != SCHED_RR)
1251 return;
1252
1253 if (--p->rt.time_slice)
1254 return;
1255
1256 p->rt.time_slice = DEF_TIMESLICE;
1257
1258 /*
1259 * Requeue to the end of queue if we are not the only element
1260 * on the queue:
1261 */
1262 if (p->rt.run_list.prev != p->rt.run_list.next) {
1263 requeue_task_rt(rq, p);
1264 set_tsk_need_resched(p);
1265 }
1266 }
1267
1268 static void set_curr_task_rt(struct rq *rq)
1269 {
1270 struct task_struct *p = rq->curr;
1271
1272 p->se.exec_start = rq->clock;
1273 }
1274
1275 const struct sched_class rt_sched_class = {
1276 .next = &fair_sched_class,
1277 .enqueue_task = enqueue_task_rt,
1278 .dequeue_task = dequeue_task_rt,
1279 .yield_task = yield_task_rt,
1280 #ifdef CONFIG_SMP
1281 .select_task_rq = select_task_rq_rt,
1282 #endif /* CONFIG_SMP */
1283
1284 .check_preempt_curr = check_preempt_curr_rt,
1285
1286 .pick_next_task = pick_next_task_rt,
1287 .put_prev_task = put_prev_task_rt,
1288
1289 #ifdef CONFIG_SMP
1290 .load_balance = load_balance_rt,
1291 .move_one_task = move_one_task_rt,
1292 .set_cpus_allowed = set_cpus_allowed_rt,
1293 .rq_online = rq_online_rt,
1294 .rq_offline = rq_offline_rt,
1295 .pre_schedule = pre_schedule_rt,
1296 .post_schedule = post_schedule_rt,
1297 .task_wake_up = task_wake_up_rt,
1298 .switched_from = switched_from_rt,
1299 #endif
1300
1301 .set_curr_task = set_curr_task_rt,
1302 .task_tick = task_tick_rt,
1303
1304 .prio_changed = prio_changed_rt,
1305 .switched_to = switched_to_rt,
1306 };
1307
|
This page was automatically generated by the
LXR engine.
|