sched: de-SCHED_OTHER-ize the RT path
[linux-2.6.git] / kernel / sched_rt.c
1 /*
2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3  * policies)
4  */
5
6 #ifdef CONFIG_SMP
7 static cpumask_t rt_overload_mask;
8 static atomic_t rto_count;
9 static inline int rt_overloaded(void)
10 {
11         return atomic_read(&rto_count);
12 }
13 static inline cpumask_t *rt_overload(void)
14 {
15         return &rt_overload_mask;
16 }
17 static inline void rt_set_overload(struct rq *rq)
18 {
19         cpu_set(rq->cpu, rt_overload_mask);
20         /*
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
25          * updated yet.
26          */
27         wmb();
28         atomic_inc(&rto_count);
29 }
30 static inline void rt_clear_overload(struct rq *rq)
31 {
32         /* the order here really doesn't matter */
33         atomic_dec(&rto_count);
34         cpu_clear(rq->cpu, rt_overload_mask);
35 }
36
37 static void update_rt_migration(struct rq *rq)
38 {
39         if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1))
40                 rt_set_overload(rq);
41         else
42                 rt_clear_overload(rq);
43 }
44 #endif /* CONFIG_SMP */
45
46 /*
47  * Update the current task's runtime statistics. Skip current tasks that
48  * are not in our scheduling class.
49  */
50 static void update_curr_rt(struct rq *rq)
51 {
52         struct task_struct *curr = rq->curr;
53         u64 delta_exec;
54
55         if (!task_has_rt_policy(curr))
56                 return;
57
58         delta_exec = rq->clock - curr->se.exec_start;
59         if (unlikely((s64)delta_exec < 0))
60                 delta_exec = 0;
61
62         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
63
64         curr->se.sum_exec_runtime += delta_exec;
65         curr->se.exec_start = rq->clock;
66         cpuacct_charge(curr, delta_exec);
67 }
68
69 static inline void inc_rt_tasks(struct task_struct *p, struct rq *rq)
70 {
71         WARN_ON(!rt_task(p));
72         rq->rt.rt_nr_running++;
73 #ifdef CONFIG_SMP
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++;
78
79         update_rt_migration(rq);
80 #endif /* CONFIG_SMP */
81 }
82
83 static inline void dec_rt_tasks(struct task_struct *p, struct rq *rq)
84 {
85         WARN_ON(!rt_task(p));
86         WARN_ON(!rq->rt.rt_nr_running);
87         rq->rt.rt_nr_running--;
88 #ifdef CONFIG_SMP
89         if (rq->rt.rt_nr_running) {
90                 struct rt_prio_array *array;
91
92                 WARN_ON(p->prio < rq->rt.highest_prio);
93                 if (p->prio == rq->rt.highest_prio) {
94                         /* recalculate */
95                         array = &rq->rt.active;
96                         rq->rt.highest_prio =
97                                 sched_find_first_bit(array->bitmap);
98                 } /* otherwise leave rq->highest prio alone */
99         } else
100                 rq->rt.highest_prio = MAX_RT_PRIO;
101         if (p->nr_cpus_allowed > 1)
102                 rq->rt.rt_nr_migratory--;
103
104         update_rt_migration(rq);
105 #endif /* CONFIG_SMP */
106 }
107
108 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
109 {
110         struct rt_prio_array *array = &rq->rt.active;
111
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);
115
116         inc_rt_tasks(p, rq);
117 }
118
119 /*
120  * Adding/removing a task to/from a priority array:
121  */
122 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
123 {
124         struct rt_prio_array *array = &rq->rt.active;
125
126         update_curr_rt(rq);
127
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);
132
133         dec_rt_tasks(p, rq);
134 }
135
136 /*
137  * Put task to the end of the run list without the overhead of dequeue
138  * followed by enqueue.
139  */
140 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
141 {
142         struct rt_prio_array *array = &rq->rt.active;
143
144         list_move_tail(&p->run_list, array->queue + p->prio);
145 }
146
147 static void
148 yield_task_rt(struct rq *rq)
149 {
150         requeue_task_rt(rq, rq->curr);
151 }
152
153 #ifdef CONFIG_SMP
154 static int select_task_rq_rt(struct task_struct *p, int sync)
155 {
156         return task_cpu(p);
157 }
158 #endif /* CONFIG_SMP */
159
160 /*
161  * Preempt the current task with a newly woken task if needed:
162  */
163 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
164 {
165         if (p->prio < rq->curr->prio)
166                 resched_task(rq->curr);
167 }
168
169 static struct task_struct *pick_next_task_rt(struct rq *rq)
170 {
171         struct rt_prio_array *array = &rq->rt.active;
172         struct task_struct *next;
173         struct list_head *queue;
174         int idx;
175
176         idx = sched_find_first_bit(array->bitmap);
177         if (idx >= MAX_RT_PRIO)
178                 return NULL;
179
180         queue = array->queue + idx;
181         next = list_entry(queue->next, struct task_struct, run_list);
182
183         next->se.exec_start = rq->clock;
184
185         return next;
186 }
187
188 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
189 {
190         update_curr_rt(rq);
191         p->se.exec_start = 0;
192 }
193
194 #ifdef CONFIG_SMP
195 /* Only try algorithms three times */
196 #define RT_MAX_TRIES 3
197
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);
200
201 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
202 {
203         if (!task_running(rq, p) &&
204             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
205             (p->nr_cpus_allowed > 1))
206                 return 1;
207         return 0;
208 }
209
210 /* Return the second highest RT task, NULL otherwise */
211 static struct task_struct *pick_next_highest_task_rt(struct rq *rq,
212                                                      int cpu)
213 {
214         struct rt_prio_array *array = &rq->rt.active;
215         struct task_struct *next;
216         struct list_head *queue;
217         int idx;
218
219         assert_spin_locked(&rq->lock);
220
221         if (likely(rq->rt.rt_nr_running < 2))
222                 return NULL;
223
224         idx = sched_find_first_bit(array->bitmap);
225         if (unlikely(idx >= MAX_RT_PRIO)) {
226                 WARN_ON(1); /* rt_nr_running is bad */
227                 return NULL;
228         }
229
230         queue = array->queue + idx;
231         BUG_ON(list_empty(queue));
232
233         next = list_entry(queue->next, struct task_struct, run_list);
234         if (unlikely(pick_rt_task(rq, next, cpu)))
235                 goto out;
236
237         if (queue->next->next != queue) {
238                 /* same prio task */
239                 next = list_entry(queue->next->next, struct task_struct, run_list);
240                 if (pick_rt_task(rq, next, cpu))
241                         goto out;
242         }
243
244  retry:
245         /* slower, but more flexible */
246         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
247         if (unlikely(idx >= MAX_RT_PRIO))
248                 return NULL;
249
250         queue = array->queue + idx;
251         BUG_ON(list_empty(queue));
252
253         list_for_each_entry(next, queue, run_list) {
254                 if (pick_rt_task(rq, next, cpu))
255                         goto out;
256         }
257
258         goto retry;
259
260  out:
261         return next;
262 }
263
264 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
265
266 /* Will lock the rq it finds */
267 static struct rq *find_lock_lowest_rq(struct task_struct *task,
268                                       struct rq *this_rq)
269 {
270         struct rq *lowest_rq = NULL;
271         int cpu;
272         int tries;
273         cpumask_t *cpu_mask = &__get_cpu_var(local_cpu_mask);
274
275         cpus_and(*cpu_mask, cpu_online_map, task->cpus_allowed);
276
277         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
278                 /*
279                  * Scan each rq for the lowest prio.
280                  */
281                 for_each_cpu_mask(cpu, *cpu_mask) {
282                         struct rq *rq = &per_cpu(runqueues, cpu);
283
284                         if (cpu == this_rq->cpu)
285                                 continue;
286
287                         /* We look for lowest RT prio or non-rt CPU */
288                         if (rq->rt.highest_prio >= MAX_RT_PRIO) {
289                                 lowest_rq = rq;
290                                 break;
291                         }
292
293                         /* no locking for now */
294                         if (rq->rt.highest_prio > task->prio &&
295                             (!lowest_rq || rq->rt.highest_prio > lowest_rq->rt.highest_prio)) {
296                                 lowest_rq = rq;
297                         }
298                 }
299
300                 if (!lowest_rq)
301                         break;
302
303                 /* if the prio of this runqueue changed, try again */
304                 if (double_lock_balance(this_rq, lowest_rq)) {
305                         /*
306                          * We had to unlock the run queue. In
307                          * the mean time, task could have
308                          * migrated already or had its affinity changed.
309                          * Also make sure that it wasn't scheduled on its rq.
310                          */
311                         if (unlikely(task_rq(task) != this_rq ||
312                                      !cpu_isset(lowest_rq->cpu, task->cpus_allowed) ||
313                                      task_running(this_rq, task) ||
314                                      !task->se.on_rq)) {
315                                 spin_unlock(&lowest_rq->lock);
316                                 lowest_rq = NULL;
317                                 break;
318                         }
319                 }
320
321                 /* If this rq is still suitable use it. */
322                 if (lowest_rq->rt.highest_prio > task->prio)
323                         break;
324
325                 /* try again */
326                 spin_unlock(&lowest_rq->lock);
327                 lowest_rq = NULL;
328         }
329
330         return lowest_rq;
331 }
332
333 /*
334  * If the current CPU has more than one RT task, see if the non
335  * running task can migrate over to a CPU that is running a task
336  * of lesser priority.
337  */
338 static int push_rt_task(struct rq *rq)
339 {
340         struct task_struct *next_task;
341         struct rq *lowest_rq;
342         int ret = 0;
343         int paranoid = RT_MAX_TRIES;
344
345         assert_spin_locked(&rq->lock);
346
347         next_task = pick_next_highest_task_rt(rq, -1);
348         if (!next_task)
349                 return 0;
350
351  retry:
352         if (unlikely(next_task == rq->curr)) {
353                 WARN_ON(1);
354                 return 0;
355         }
356
357         /*
358          * It's possible that the next_task slipped in of
359          * higher priority than current. If that's the case
360          * just reschedule current.
361          */
362         if (unlikely(next_task->prio < rq->curr->prio)) {
363                 resched_task(rq->curr);
364                 return 0;
365         }
366
367         /* We might release rq lock */
368         get_task_struct(next_task);
369
370         /* find_lock_lowest_rq locks the rq if found */
371         lowest_rq = find_lock_lowest_rq(next_task, rq);
372         if (!lowest_rq) {
373                 struct task_struct *task;
374                 /*
375                  * find lock_lowest_rq releases rq->lock
376                  * so it is possible that next_task has changed.
377                  * If it has, then try again.
378                  */
379                 task = pick_next_highest_task_rt(rq, -1);
380                 if (unlikely(task != next_task) && task && paranoid--) {
381                         put_task_struct(next_task);
382                         next_task = task;
383                         goto retry;
384                 }
385                 goto out;
386         }
387
388         assert_spin_locked(&lowest_rq->lock);
389
390         deactivate_task(rq, next_task, 0);
391         set_task_cpu(next_task, lowest_rq->cpu);
392         activate_task(lowest_rq, next_task, 0);
393
394         resched_task(lowest_rq->curr);
395
396         spin_unlock(&lowest_rq->lock);
397
398         ret = 1;
399 out:
400         put_task_struct(next_task);
401
402         return ret;
403 }
404
405 /*
406  * TODO: Currently we just use the second highest prio task on
407  *       the queue, and stop when it can't migrate (or there's
408  *       no more RT tasks).  There may be a case where a lower
409  *       priority RT task has a different affinity than the
410  *       higher RT task. In this case the lower RT task could
411  *       possibly be able to migrate where as the higher priority
412  *       RT task could not.  We currently ignore this issue.
413  *       Enhancements are welcome!
414  */
415 static void push_rt_tasks(struct rq *rq)
416 {
417         /* push_rt_task will return true if it moved an RT */
418         while (push_rt_task(rq))
419                 ;
420 }
421
422 static int pull_rt_task(struct rq *this_rq)
423 {
424         struct task_struct *next;
425         struct task_struct *p;
426         struct rq *src_rq;
427         cpumask_t *rto_cpumask;
428         int this_cpu = this_rq->cpu;
429         int cpu;
430         int ret = 0;
431
432         assert_spin_locked(&this_rq->lock);
433
434         /*
435          * If cpusets are used, and we have overlapping
436          * run queue cpusets, then this algorithm may not catch all.
437          * This is just the price you pay on trying to keep
438          * dirtying caches down on large SMP machines.
439          */
440         if (likely(!rt_overloaded()))
441                 return 0;
442
443         next = pick_next_task_rt(this_rq);
444
445         rto_cpumask = rt_overload();
446
447         for_each_cpu_mask(cpu, *rto_cpumask) {
448                 if (this_cpu == cpu)
449                         continue;
450
451                 src_rq = cpu_rq(cpu);
452                 if (unlikely(src_rq->rt.rt_nr_running <= 1)) {
453                         /*
454                          * It is possible that overlapping cpusets
455                          * will miss clearing a non overloaded runqueue.
456                          * Clear it now.
457                          */
458                         if (double_lock_balance(this_rq, src_rq)) {
459                                 /* unlocked our runqueue lock */
460                                 struct task_struct *old_next = next;
461                                 next = pick_next_task_rt(this_rq);
462                                 if (next != old_next)
463                                         ret = 1;
464                         }
465                         if (likely(src_rq->rt.rt_nr_running <= 1))
466                                 /*
467                                  * Small chance that this_rq->curr changed
468                                  * but it's really harmless here.
469                                  */
470                                 rt_clear_overload(this_rq);
471                         else
472                                 /*
473                                  * Heh, the src_rq is now overloaded, since
474                                  * we already have the src_rq lock, go straight
475                                  * to pulling tasks from it.
476                                  */
477                                 goto try_pulling;
478                         spin_unlock(&src_rq->lock);
479                         continue;
480                 }
481
482                 /*
483                  * We can potentially drop this_rq's lock in
484                  * double_lock_balance, and another CPU could
485                  * steal our next task - hence we must cause
486                  * the caller to recalculate the next task
487                  * in that case:
488                  */
489                 if (double_lock_balance(this_rq, src_rq)) {
490                         struct task_struct *old_next = next;
491                         next = pick_next_task_rt(this_rq);
492                         if (next != old_next)
493                                 ret = 1;
494                 }
495
496                 /*
497                  * Are there still pullable RT tasks?
498                  */
499                 if (src_rq->rt.rt_nr_running <= 1) {
500                         spin_unlock(&src_rq->lock);
501                         continue;
502                 }
503
504  try_pulling:
505                 p = pick_next_highest_task_rt(src_rq, this_cpu);
506
507                 /*
508                  * Do we have an RT task that preempts
509                  * the to-be-scheduled task?
510                  */
511                 if (p && (!next || (p->prio < next->prio))) {
512                         WARN_ON(p == src_rq->curr);
513                         WARN_ON(!p->se.on_rq);
514
515                         /*
516                          * There's a chance that p is higher in priority
517                          * than what's currently running on its cpu.
518                          * This is just that p is wakeing up and hasn't
519                          * had a chance to schedule. We only pull
520                          * p if it is lower in priority than the
521                          * current task on the run queue or
522                          * this_rq next task is lower in prio than
523                          * the current task on that rq.
524                          */
525                         if (p->prio < src_rq->curr->prio ||
526                             (next && next->prio < src_rq->curr->prio))
527                                 goto bail;
528
529                         ret = 1;
530
531                         deactivate_task(src_rq, p, 0);
532                         set_task_cpu(p, this_cpu);
533                         activate_task(this_rq, p, 0);
534                         /*
535                          * We continue with the search, just in
536                          * case there's an even higher prio task
537                          * in another runqueue. (low likelyhood
538                          * but possible)
539                          */
540
541                         /*
542                          * Update next so that we won't pick a task
543                          * on another cpu with a priority lower (or equal)
544                          * than the one we just picked.
545                          */
546                         next = p;
547
548                 }
549  bail:
550                 spin_unlock(&src_rq->lock);
551         }
552
553         return ret;
554 }
555
556 static void schedule_balance_rt(struct rq *rq,
557                                 struct task_struct *prev)
558 {
559         /* Try to pull RT tasks here if we lower this rq's prio */
560         if (unlikely(rt_task(prev)) &&
561             rq->rt.highest_prio > prev->prio)
562                 pull_rt_task(rq);
563 }
564
565 static void schedule_tail_balance_rt(struct rq *rq)
566 {
567         /*
568          * If we have more than one rt_task queued, then
569          * see if we can push the other rt_tasks off to other CPUS.
570          * Note we may release the rq lock, and since
571          * the lock was owned by prev, we need to release it
572          * first via finish_lock_switch and then reaquire it here.
573          */
574         if (unlikely(rq->rt.rt_nr_running > 1)) {
575                 spin_lock_irq(&rq->lock);
576                 push_rt_tasks(rq);
577                 spin_unlock_irq(&rq->lock);
578         }
579 }
580
581
582 static void wakeup_balance_rt(struct rq *rq, struct task_struct *p)
583 {
584         if (unlikely(rt_task(p)) &&
585             !task_running(rq, p) &&
586             (p->prio >= rq->curr->prio))
587                 push_rt_tasks(rq);
588 }
589
590 static unsigned long
591 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
592                 unsigned long max_load_move,
593                 struct sched_domain *sd, enum cpu_idle_type idle,
594                 int *all_pinned, int *this_best_prio)
595 {
596         /* don't touch RT tasks */
597         return 0;
598 }
599
600 static int
601 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
602                  struct sched_domain *sd, enum cpu_idle_type idle)
603 {
604         /* don't touch RT tasks */
605         return 0;
606 }
607 static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
608 {
609         int weight = cpus_weight(*new_mask);
610
611         BUG_ON(!rt_task(p));
612
613         /*
614          * Update the migration status of the RQ if we have an RT task
615          * which is running AND changing its weight value.
616          */
617         if (p->se.on_rq && (weight != p->nr_cpus_allowed)) {
618                 struct rq *rq = task_rq(p);
619
620                 if ((p->nr_cpus_allowed <= 1) && (weight > 1))
621                         rq->rt.rt_nr_migratory++;
622                 else if((p->nr_cpus_allowed > 1) && (weight <= 1)) {
623                         BUG_ON(!rq->rt.rt_nr_migratory);
624                         rq->rt.rt_nr_migratory--;
625                 }
626
627                 update_rt_migration(rq);
628         }
629
630         p->cpus_allowed    = *new_mask;
631         p->nr_cpus_allowed = weight;
632 }
633 #else /* CONFIG_SMP */
634 # define schedule_tail_balance_rt(rq)   do { } while (0)
635 # define schedule_balance_rt(rq, prev)  do { } while (0)
636 # define wakeup_balance_rt(rq, p)       do { } while (0)
637 #endif /* CONFIG_SMP */
638
639 static void task_tick_rt(struct rq *rq, struct task_struct *p)
640 {
641         update_curr_rt(rq);
642
643         /*
644          * RR tasks need a special form of timeslice management.
645          * FIFO tasks have no timeslices.
646          */
647         if (p->policy != SCHED_RR)
648                 return;
649
650         if (--p->time_slice)
651                 return;
652
653         p->time_slice = DEF_TIMESLICE;
654
655         /*
656          * Requeue to the end of queue if we are not the only element
657          * on the queue:
658          */
659         if (p->run_list.prev != p->run_list.next) {
660                 requeue_task_rt(rq, p);
661                 set_tsk_need_resched(p);
662         }
663 }
664
665 static void set_curr_task_rt(struct rq *rq)
666 {
667         struct task_struct *p = rq->curr;
668
669         p->se.exec_start = rq->clock;
670 }
671
672 const struct sched_class rt_sched_class = {
673         .next                   = &fair_sched_class,
674         .enqueue_task           = enqueue_task_rt,
675         .dequeue_task           = dequeue_task_rt,
676         .yield_task             = yield_task_rt,
677 #ifdef CONFIG_SMP
678         .select_task_rq         = select_task_rq_rt,
679 #endif /* CONFIG_SMP */
680
681         .check_preempt_curr     = check_preempt_curr_rt,
682
683         .pick_next_task         = pick_next_task_rt,
684         .put_prev_task          = put_prev_task_rt,
685
686 #ifdef CONFIG_SMP
687         .load_balance           = load_balance_rt,
688         .move_one_task          = move_one_task_rt,
689         .set_cpus_allowed       = set_cpus_allowed_rt,
690 #endif
691
692         .set_curr_task          = set_curr_task_rt,
693         .task_tick              = task_tick_rt,
694 };