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