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         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 (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
  43                 if (!rq->rt.overloaded) {
  44                         rt_set_overload(rq);
  45                         rq->rt.overloaded = 1;
  46                 }
  47         } else if (rq->rt.overloaded) {
  48                 rt_clear_overload(rq);
  49                 rq->rt.overloaded = 0;
  50         }
  51 }
  52 #endif /* CONFIG_SMP */
  53
  54 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  55 {
  56         return container_of(rt_se, struct task_struct, rt);
  57 }
  58
  59 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
  60 {
  61         return !list_empty(&rt_se->run_list);
  62 }
  63
  64 #ifdef CONFIG_RT_GROUP_SCHED
  65
  66 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
  67 {
  68         if (!rt_rq->tg)
  69                 return RUNTIME_INF;
  70
  71         return rt_rq->rt_runtime;
  72 }
  73
  74 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
  75 {
  76         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
  77 }
  78
  79 #define for_each_leaf_rt_rq(rt_rq, rq) \
  80         list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
  81
  82 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
  83 {
  84         return rt_rq->rq;
  85 }
  86
  87 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
  88 {
  89         return rt_se->rt_rq;
  90 }
  91
  92 #define for_each_sched_rt_entity(rt_se) \
  93         for (; rt_se; rt_se = rt_se->parent)
  94
  95 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
  96 {
  97         return rt_se->my_q;
  98 }
  99
 100 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
 101 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
 102
 103 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 104 {
 105         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 106
 107         if (rt_se && !on_rt_rq(rt_se) && rt_rq->rt_nr_running) {
 108                 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 109
 110                 enqueue_rt_entity(rt_se);
 111                 if (rt_rq->highest_prio < curr->prio)
 112                         resched_task(curr);
 113         }
 114 }
 115
 116 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 117 {
 118         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 119
 120         if (rt_se && on_rt_rq(rt_se))
 121                 dequeue_rt_entity(rt_se);
 122 }
 123
 124 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 125 {
 126         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 127 }
 128
 129 static int rt_se_boosted(struct sched_rt_entity *rt_se)
 130 {
 131         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 132         struct task_struct *p;
 133
 134         if (rt_rq)
 135                 return !!rt_rq->rt_nr_boosted;
 136
 137         p = rt_task_of(rt_se);
 138         return p->prio != p->normal_prio;
 139 }
 140
 141 #ifdef CONFIG_SMP
 142 static inline cpumask_t sched_rt_period_mask(void)
 143 {
 144         return cpu_rq(smp_processor_id())->rd->span;
 145 }
 146 #else
 147 static inline cpumask_t sched_rt_period_mask(void)
 148 {
 149         return cpu_online_map;
 150 }
 151 #endif
 152
 153 static inline
 154 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 155 {
 156         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
 157 }
 158
 159 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 160 {
 161         return &rt_rq->tg->rt_bandwidth;
 162 }
 163
 164 #else /* !CONFIG_RT_GROUP_SCHED */
 165
 166 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 167 {
 168         return rt_rq->rt_runtime;
 169 }
 170
 171 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 172 {
 173         return ktime_to_ns(def_rt_bandwidth.rt_period);
 174 }
 175
 176 #define for_each_leaf_rt_rq(rt_rq, rq) \
 177         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 178
 179 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
 180 {
 181         return container_of(rt_rq, struct rq, rt);
 182 }
 183
 184 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
 185 {
 186         struct task_struct *p = rt_task_of(rt_se);
 187         struct rq *rq = task_rq(p);
 188
 189         return &rq->rt;
 190 }
 191
 192 #define for_each_sched_rt_entity(rt_se) \
 193         for (; rt_se; rt_se = NULL)
 194
 195 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 196 {
 197         return NULL;
 198 }
 199
 200 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 201 {
 202 }
 203
 204 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 205 {
 206 }
 207
 208 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 209 {
 210         return rt_rq->rt_throttled;
 211 }
 212
 213 static inline cpumask_t sched_rt_period_mask(void)
 214 {
 215         return cpu_online_map;
 216 }
 217
 218 static inline
 219 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 220 {
 221         return &cpu_rq(cpu)->rt;
 222 }
 223
 224 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 225 {
 226         return &def_rt_bandwidth;
 227 }
 228
 229 #endif /* CONFIG_RT_GROUP_SCHED */
 230
 231 #ifdef CONFIG_SMP
 232 static int do_balance_runtime(struct rt_rq *rt_rq)
 233 {
 234         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 235         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 236         int i, weight, more = 0;
 237         u64 rt_period;
 238
 239         weight = cpus_weight(rd->span);
 240
 241         spin_lock(&rt_b->rt_runtime_lock);
 242         rt_period = ktime_to_ns(rt_b->rt_period);
 243         for_each_cpu_mask_nr(i, rd->span) {
 244                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 245                 s64 diff;
 246
 247                 if (iter == rt_rq)
 248                         continue;
 249
 250                 spin_lock(&iter->rt_runtime_lock);
 251                 if (iter->rt_runtime == RUNTIME_INF)
 252                         goto next;
 253
 254                 diff = iter->rt_runtime - iter->rt_time;
 255                 if (diff > 0) {
 256                         diff = div_u64((u64)diff, weight);
 257                         if (rt_rq->rt_runtime + diff > rt_period)
 258                                 diff = rt_period - rt_rq->rt_runtime;
 259                         iter->rt_runtime -= diff;
 260                         rt_rq->rt_runtime += diff;
 261                         more = 1;
 262                         if (rt_rq->rt_runtime == rt_period) {
 263                                 spin_unlock(&iter->rt_runtime_lock);
 264                                 break;
 265                         }
 266                 }
 267 next:
 268                 spin_unlock(&iter->rt_runtime_lock);
 269         }
 270         spin_unlock(&rt_b->rt_runtime_lock);
 271
 272         return more;
 273 }
 274
 275 static void __disable_runtime(struct rq *rq)
 276 {
 277         struct root_domain *rd = rq->rd;
 278         struct rt_rq *rt_rq;
 279
 280         if (unlikely(!scheduler_running))
 281                 return;
 282
 283         for_each_leaf_rt_rq(rt_rq, rq) {
 284                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 285                 s64 want;
 286                 int i;
 287
 288                 spin_lock(&rt_b->rt_runtime_lock);
 289                 spin_lock(&rt_rq->rt_runtime_lock);
 290                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 291                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 292                         goto balanced;
 293                 spin_unlock(&rt_rq->rt_runtime_lock);
 294
 295                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 296
 297                 for_each_cpu_mask(i, rd->span) {
 298                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 299                         s64 diff;
 300
 301                         if (iter == rt_rq)
 302                                 continue;
 303
 304                         spin_lock(&iter->rt_runtime_lock);
 305                         if (want > 0) {
 306                                 diff = min_t(s64, iter->rt_runtime, want);
 307                                 iter->rt_runtime -= diff;
 308                                 want -= diff;
 309                         } else {
 310                                 iter->rt_runtime -= want;
 311                                 want -= want;
 312                         }
 313                         spin_unlock(&iter->rt_runtime_lock);
 314
 315                         if (!want)
 316                                 break;
 317                 }
 318
 319                 spin_lock(&rt_rq->rt_runtime_lock);
 320                 BUG_ON(want);
 321 balanced:
 322                 rt_rq->rt_runtime = RUNTIME_INF;
 323                 spin_unlock(&rt_rq->rt_runtime_lock);
 324                 spin_unlock(&rt_b->rt_runtime_lock);
 325         }
 326 }
 327
 328 static void disable_runtime(struct rq *rq)
 329 {
 330         unsigned long flags;
 331
 332         spin_lock_irqsave(&rq->lock, flags);
 333         __disable_runtime(rq);
 334         spin_unlock_irqrestore(&rq->lock, flags);
 335 }
 336
 337 static void __enable_runtime(struct rq *rq)
 338 {
 339         struct rt_rq *rt_rq;
 340
 341         if (unlikely(!scheduler_running))
 342                 return;
 343
 344         for_each_leaf_rt_rq(rt_rq, rq) {
 345                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 346
 347                 spin_lock(&rt_b->rt_runtime_lock);
 348                 spin_lock(&rt_rq->rt_runtime_lock);
 349                 rt_rq->rt_runtime = rt_b->rt_runtime;
 350                 rt_rq->rt_time = 0;
 351                 spin_unlock(&rt_rq->rt_runtime_lock);
 352                 spin_unlock(&rt_b->rt_runtime_lock);
 353         }
 354 }
 355
 356 static void enable_runtime(struct rq *rq)
 357 {
 358         unsigned long flags;
 359
 360         spin_lock_irqsave(&rq->lock, flags);
 361         __enable_runtime(rq);
 362         spin_unlock_irqrestore(&rq->lock, flags);
 363 }
 364
 365 static int balance_runtime(struct rt_rq *rt_rq)
 366 {
 367         int more = 0;
 368
 369         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 370                 spin_unlock(&rt_rq->rt_runtime_lock);
 371                 more = do_balance_runtime(rt_rq);
 372                 spin_lock(&rt_rq->rt_runtime_lock);
 373         }
 374
 375         return more;
 376 }
 377 #else /* !CONFIG_SMP */
 378 static inline int balance_runtime(struct rt_rq *rt_rq)
 379 {
 380         return 0;
 381 }
 382 #endif /* CONFIG_SMP */
 383
 384 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 385 {
 386         int i, idle = 1;
 387         cpumask_t span;
 388
 389         if (rt_b->rt_runtime == RUNTIME_INF)
 390                 return 1;
 391
 392         span = sched_rt_period_mask();
 393         for_each_cpu_mask(i, span) {
 394                 int enqueue = 0;
 395                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 396                 struct rq *rq = rq_of_rt_rq(rt_rq);
 397
 398                 spin_lock(&rq->lock);
 399                 if (rt_rq->rt_time) {
 400                         u64 runtime;
 401
 402                         spin_lock(&rt_rq->rt_runtime_lock);
 403                         if (rt_rq->rt_throttled)
 404                                 balance_runtime(rt_rq);
 405                         runtime = rt_rq->rt_runtime;
 406                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 407                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 408                                 rt_rq->rt_throttled = 0;
 409                                 enqueue = 1;
 410                         }
 411                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 412                                 idle = 0;
 413                         spin_unlock(&rt_rq->rt_runtime_lock);
 414                 } else if (rt_rq->rt_nr_running)
 415                         idle = 0;
 416
 417                 if (enqueue)
 418                         sched_rt_rq_enqueue(rt_rq);
 419                 spin_unlock(&rq->lock);
 420         }
 421
 422         return idle;
 423 }
 424
 425 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 426 {
 427 #ifdef CONFIG_RT_GROUP_SCHED
 428         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 429
 430         if (rt_rq)
 431                 return rt_rq->highest_prio;
 432 #endif
 433
 434         return rt_task_of(rt_se)->prio;
 435 }
 436
 437 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 438 {
 439         u64 runtime = sched_rt_runtime(rt_rq);
 440
 441         if (runtime == RUNTIME_INF)
 442                 return 0;
 443
 444         if (rt_rq->rt_throttled)
 445                 return rt_rq_throttled(rt_rq);
 446
 447         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 448                 return 0;
 449
 450         balance_runtime(rt_rq);
 451         runtime = sched_rt_runtime(rt_rq);
 452         if (runtime == RUNTIME_INF)
 453                 return 0;
 454
 455         if (rt_rq->rt_time > runtime) {
 456                 rt_rq->rt_throttled = 1;
 457                 if (rt_rq_throttled(rt_rq)) {
 458                         sched_rt_rq_dequeue(rt_rq);
 459                         return 1;
 460                 }
 461         }
 462
 463         return 0;
 464 }
 465
 466 /*
 467  * Update the current task's runtime statistics. Skip current tasks that
 468  * are not in our scheduling class.
 469  */
 470 static void update_curr_rt(struct rq *rq)
 471 {
 472         struct task_struct *curr = rq->curr;
 473         struct sched_rt_entity *rt_se = &curr->rt;
 474         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 475         u64 delta_exec;
 476
 477         if (!task_has_rt_policy(curr))
 478                 return;
 479
 480         delta_exec = rq->clock - curr->se.exec_start;
 481         if (unlikely((s64)delta_exec < 0))
 482                 delta_exec = 0;
 483
 484         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
 485
 486         curr->se.sum_exec_runtime += delta_exec;
 487         curr->se.exec_start = rq->clock;
 488         cpuacct_charge(curr, delta_exec);
 489
 490         for_each_sched_rt_entity(rt_se) {
 491                 rt_rq = rt_rq_of_se(rt_se);
 492
 493                 spin_lock(&rt_rq->rt_runtime_lock);
 494                 rt_rq->rt_time += delta_exec;
 495                 if (sched_rt_runtime_exceeded(rt_rq))
 496                         resched_task(curr);
 497                 spin_unlock(&rt_rq->rt_runtime_lock);
 498         }
 499 }
 500
 501 static inline
 502 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 503 {
 504         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 505         rt_rq->rt_nr_running++;
 506 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 507         if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
 508 #ifdef CONFIG_SMP
 509                 struct rq *rq = rq_of_rt_rq(rt_rq);
 510 #endif
 511
 512                 rt_rq->highest_prio = rt_se_prio(rt_se);
 513 #ifdef CONFIG_SMP
 514                 if (rq->online)
 515                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 516                                    rt_se_prio(rt_se));
 517 #endif
 518         }
 519 #endif
 520 #ifdef CONFIG_SMP
 521         if (rt_se->nr_cpus_allowed > 1) {
 522                 struct rq *rq = rq_of_rt_rq(rt_rq);
 523
 524                 rq->rt.rt_nr_migratory++;
 525         }
 526
 527         update_rt_migration(rq_of_rt_rq(rt_rq));
 528 #endif
 529 #ifdef CONFIG_RT_GROUP_SCHED
 530         if (rt_se_boosted(rt_se))
 531                 rt_rq->rt_nr_boosted++;
 532
 533         if (rt_rq->tg)
 534                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 535 #else
 536         start_rt_bandwidth(&def_rt_bandwidth);
 537 #endif
 538 }
 539
 540 static inline
 541 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 542 {
 543 #ifdef CONFIG_SMP
 544         int highest_prio = rt_rq->highest_prio;
 545 #endif
 546
 547         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 548         WARN_ON(!rt_rq->rt_nr_running);
 549         rt_rq->rt_nr_running--;
 550 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 551         if (rt_rq->rt_nr_running) {
 552                 struct rt_prio_array *array;
 553
 554                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
 555                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
 556                         /* recalculate */
 557                         array = &rt_rq->active;
 558                         rt_rq->highest_prio =
 559                                 sched_find_first_bit(array->bitmap);
 560                 } /* otherwise leave rq->highest prio alone */
 561         } else
 562                 rt_rq->highest_prio = MAX_RT_PRIO;
 563 #endif
 564 #ifdef CONFIG_SMP
 565         if (rt_se->nr_cpus_allowed > 1) {
 566                 struct rq *rq = rq_of_rt_rq(rt_rq);
 567                 rq->rt.rt_nr_migratory--;
 568         }
 569
 570         if (rt_rq->highest_prio != highest_prio) {
 571                 struct rq *rq = rq_of_rt_rq(rt_rq);
 572
 573                 if (rq->online)
 574                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 575                                    rt_rq->highest_prio);
 576         }
 577
 578         update_rt_migration(rq_of_rt_rq(rt_rq));
 579 #endif /* CONFIG_SMP */
 580 #ifdef CONFIG_RT_GROUP_SCHED
 581         if (rt_se_boosted(rt_se))
 582                 rt_rq->rt_nr_boosted--;
 583
 584         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 585 #endif
 586 }
 587
 588 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
 589 {
 590         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 591         struct rt_prio_array *array = &rt_rq->active;
 592         struct rt_rq *group_rq = group_rt_rq(rt_se);
 593         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 594
 595         /*
 596          * Don't enqueue the group if its throttled, or when empty.
 597          * The latter is a consequence of the former when a child group
 598          * get throttled and the current group doesn't have any other
 599          * active members.
 600          */
 601         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 602                 return;
 603
 604         list_add_tail(&rt_se->run_list, queue);
 605         __set_bit(rt_se_prio(rt_se), array->bitmap);
 606
 607         inc_rt_tasks(rt_se, rt_rq);
 608 }
 609
 610 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 611 {
 612         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 613         struct rt_prio_array *array = &rt_rq->active;
 614
 615         list_del_init(&rt_se->run_list);
 616         if (list_empty(array->queue + rt_se_prio(rt_se)))
 617                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 618
 619         dec_rt_tasks(rt_se, rt_rq);
 620 }
 621
 622 /*
 623  * Because the prio of an upper entry depends on the lower
 624  * entries, we must remove entries top - down.
 625  */
 626 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 627 {
 628         struct sched_rt_entity *back = NULL;
 629
 630         for_each_sched_rt_entity(rt_se) {
 631                 rt_se->back = back;
 632                 back = rt_se;
 633         }
 634
 635         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 636                 if (on_rt_rq(rt_se))
 637                         __dequeue_rt_entity(rt_se);
 638         }
 639 }
 640
 641 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 642 {
 643         dequeue_rt_stack(rt_se);
 644         for_each_sched_rt_entity(rt_se)
 645                 __enqueue_rt_entity(rt_se);
 646 }
 647
 648 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 649 {
 650         dequeue_rt_stack(rt_se);
 651
 652         for_each_sched_rt_entity(rt_se) {
 653                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 654
 655                 if (rt_rq && rt_rq->rt_nr_running)
 656                         __enqueue_rt_entity(rt_se);
 657         }
 658 }
 659
 660 /*
 661  * Adding/removing a task to/from a priority array:
 662  */
 663 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 664 {
 665         struct sched_rt_entity *rt_se = &p->rt;
 666
 667         if (wakeup)
 668                 rt_se->timeout = 0;
 669
 670         enqueue_rt_entity(rt_se);
 671
 672         inc_cpu_load(rq, p->se.load.weight);
 673 }
 674
 675 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 676 {
 677         struct sched_rt_entity *rt_se = &p->rt;
 678
 679         update_curr_rt(rq);
 680         dequeue_rt_entity(rt_se);
 681
 682         dec_cpu_load(rq, p->se.load.weight);
 683 }
 684
 685 /*
 686  * Put task to the end of the run list without the overhead of dequeue
 687  * followed by enqueue.
 688  */
 689 static void
 690 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 691 {
 692         if (on_rt_rq(rt_se)) {
 693                 struct rt_prio_array *array = &rt_rq->active;
 694                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 695
 696                 if (head)
 697                         list_move(&rt_se->run_list, queue);
 698                 else
 699                         list_move_tail(&rt_se->run_list, queue);
 700         }
 701 }
 702
 703 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 704 {
 705         struct sched_rt_entity *rt_se = &p->rt;
 706         struct rt_rq *rt_rq;
 707
 708         for_each_sched_rt_entity(rt_se) {
 709                 rt_rq = rt_rq_of_se(rt_se);
 710                 requeue_rt_entity(rt_rq, rt_se, head);
 711         }
 712 }
 713
 714 static void yield_task_rt(struct rq *rq)
 715 {
 716         requeue_task_rt(rq, rq->curr, 0);
 717 }
 718
 719 #ifdef CONFIG_SMP
 720 static int find_lowest_rq(struct task_struct *task);
 721
 722 static int select_task_rq_rt(struct task_struct *p, int sync)
 723 {
 724         struct rq *rq = task_rq(p);
 725
 726         /*
 727          * If the current task is an RT task, then
 728          * try to see if we can wake this RT task up on another
 729          * runqueue. Otherwise simply start this RT task
 730          * on its current runqueue.
 731          *
 732          * We want to avoid overloading runqueues. Even if
 733          * the RT task is of higher priority than the current RT task.
 734          * RT tasks behave differently than other tasks. If
 735          * one gets preempted, we try to push it off to another queue.
 736          * So trying to keep a preempting RT task on the same
 737          * cache hot CPU will force the running RT task to
 738          * a cold CPU. So we waste all the cache for the lower
 739          * RT task in hopes of saving some of a RT task
 740          * that is just being woken and probably will have
 741          * cold cache anyway.
 742          */
 743         if (unlikely(rt_task(rq->curr)) &&
 744             (p->rt.nr_cpus_allowed > 1)) {
 745                 int cpu = find_lowest_rq(p);
 746
 747                 return (cpu == -1) ? task_cpu(p) : cpu;
 748         }
 749
 750         /*
 751          * Otherwise, just let it ride on the affined RQ and the
 752          * post-schedule router will push the preempted task away
 753          */
 754         return task_cpu(p);
 755 }
 756
 757 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
 758 {
 759         cpumask_t mask;
 760
 761         if (rq->curr->rt.nr_cpus_allowed == 1)
 762                 return;
 763
 764         if (p->rt.nr_cpus_allowed != 1
 765             && cpupri_find(&rq->rd->cpupri, p, &mask))
 766                 return;
 767
 768         if (!cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
 769                 return;
 770
 771         /*
 772          * There appears to be other cpus that can accept
 773          * current and none to run 'p', so lets reschedule
 774          * to try and push current away:
 775          */
 776         requeue_task_rt(rq, p, 1);
 777         resched_task(rq->curr);
 778 }
 779
 780 #endif /* CONFIG_SMP */
 781
 782 /*
 783  * Preempt the current task with a newly woken task if needed:
 784  */
 785 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
 786 {
 787         if (p->prio < rq->curr->prio) {
 788                 resched_task(rq->curr);
 789                 return;
 790         }
 791
 792 #ifdef CONFIG_SMP
 793         /*
 794          * If:
 795          *
 796          * - the newly woken task is of equal priority to the current task
 797          * - the newly woken task is non-migratable while current is migratable
 798          * - current will be preempted on the next reschedule
 799          *
 800          * we should check to see if current can readily move to a different
 801          * cpu.  If so, we will reschedule to allow the push logic to try
 802          * to move current somewhere else, making room for our non-migratable
 803          * task.
 804          */
 805         if (p->prio == rq->curr->prio && !need_resched())
 806                 check_preempt_equal_prio(rq, p);
 807 #endif
 808 }
 809
 810 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 811                                                    struct rt_rq *rt_rq)
 812 {
 813         struct rt_prio_array *array = &rt_rq->active;
 814         struct sched_rt_entity *next = NULL;
 815         struct list_head *queue;
 816         int idx;
 817
 818         idx = sched_find_first_bit(array->bitmap);
 819         BUG_ON(idx >= MAX_RT_PRIO);
 820
 821         queue = array->queue + idx;
 822         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 823
 824         return next;
 825 }
 826
 827 static struct task_struct *pick_next_task_rt(struct rq *rq)
 828 {
 829         struct sched_rt_entity *rt_se;
 830         struct task_struct *p;
 831         struct rt_rq *rt_rq;
 832
 833         rt_rq = &rq->rt;
 834
 835         if (unlikely(!rt_rq->rt_nr_running))
 836                 return NULL;
 837
 838         if (rt_rq_throttled(rt_rq))
 839                 return NULL;
 840
 841         do {
 842                 rt_se = pick_next_rt_entity(rq, rt_rq);
 843                 BUG_ON(!rt_se);
 844                 rt_rq = group_rt_rq(rt_se);
 845         } while (rt_rq);
 846
 847         p = rt_task_of(rt_se);
 848         p->se.exec_start = rq->clock;
 849         return p;
 850 }
 851
 852 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 853 {
 854         update_curr_rt(rq);
 855         p->se.exec_start = 0;
 856 }
 857
 858 #ifdef CONFIG_SMP
 859
 860 /* Only try algorithms three times */
 861 #define RT_MAX_TRIES 3
 862
 863 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
 864 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 865
 866 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 867 {
 868         if (!task_running(rq, p) &&
 869             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
 870             (p->rt.nr_cpus_allowed > 1))
 871                 return 1;
 872         return 0;
 873 }
 874
 875 /* Return the second highest RT task, NULL otherwise */
 876 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
 877 {
 878         struct task_struct *next = NULL;
 879         struct sched_rt_entity *rt_se;
 880         struct rt_prio_array *array;
 881         struct rt_rq *rt_rq;
 882         int idx;
 883
 884         for_each_leaf_rt_rq(rt_rq, rq) {
 885                 array = &rt_rq->active;
 886                 idx = sched_find_first_bit(array->bitmap);
 887  next_idx:
 888                 if (idx >= MAX_RT_PRIO)
 889                         continue;
 890                 if (next && next->prio < idx)
 891                         continue;
 892                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
 893                         struct task_struct *p = rt_task_of(rt_se);
 894                         if (pick_rt_task(rq, p, cpu)) {
 895                                 next = p;
 896                                 break;
 897                         }
 898                 }
 899                 if (!next) {
 900                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 901                         goto next_idx;
 902                 }
 903         }
 904
 905         return next;
 906 }
 907
 908 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
 909
 910 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
 911 {
 912         int first;
 913
 914         /* "this_cpu" is cheaper to preempt than a remote processor */
 915         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
 916                 return this_cpu;
 917
 918         first = first_cpu(*mask);
 919         if (first != NR_CPUS)
 920                 return first;
 921
 922         return -1;
 923 }
 924
 925 static int find_lowest_rq(struct task_struct *task)
 926 {
 927         struct sched_domain *sd;
 928         cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
 929         int this_cpu = smp_processor_id();
 930         int cpu      = task_cpu(task);
 931
 932         if (task->rt.nr_cpus_allowed == 1)
 933                 return -1; /* No other targets possible */
 934
 935         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
 936                 return -1; /* No targets found */
 937
 938         /*
 939          * Only consider CPUs that are usable for migration.
 940          * I guess we might want to change cpupri_find() to ignore those
 941          * in the first place.
 942          */
 943         cpus_and(*lowest_mask, *lowest_mask, cpu_active_map);
 944
 945         /*
 946          * At this point we have built a mask of cpus representing the
 947          * lowest priority tasks in the system.  Now we want to elect
 948          * the best one based on our affinity and topology.
 949          *
 950          * We prioritize the last cpu that the task executed on since
 951          * it is most likely cache-hot in that location.
 952          */
 953         if (cpu_isset(cpu, *lowest_mask))
 954                 return cpu;
 955
 956         /*
 957          * Otherwise, we consult the sched_domains span maps to figure
 958          * out which cpu is logically closest to our hot cache data.
 959          */
 960         if (this_cpu == cpu)
 961                 this_cpu = -1; /* Skip this_cpu opt if the same */
 962
 963         for_each_domain(cpu, sd) {
 964                 if (sd->flags & SD_WAKE_AFFINE) {
 965                         cpumask_t domain_mask;
 966                         int       best_cpu;
 967
 968                         cpus_and(domain_mask, sd->span, *lowest_mask);
 969
 970                         best_cpu = pick_optimal_cpu(this_cpu,
 971                                                     &domain_mask);
 972                         if (best_cpu != -1)
 973                                 return best_cpu;
 974                 }
 975         }
 976
 977         /*
 978          * And finally, if there were no matches within the domains
 979          * just give the caller *something* to work with from the compatible
 980          * locations.
 981          */
 982         return pick_optimal_cpu(this_cpu, lowest_mask);
 983 }
 984
 985 /* Will lock the rq it finds */
 986 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
 987 {
 988         struct rq *lowest_rq = NULL;
 989         int tries;
 990         int cpu;
 991
 992         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
 993                 cpu = find_lowest_rq(task);
 994
 995                 if ((cpu == -1) || (cpu == rq->cpu))
 996                         break;
 997
 998                 lowest_rq = cpu_rq(cpu);
 999
1000                 /* if the prio of this runqueue changed, try again */
1001                 if (double_lock_balance(rq, lowest_rq)) {
1002                         /*
1003                          * We had to unlock the run queue. In
1004                          * the mean time, task could have
1005                          * migrated already or had its affinity changed.
1006                          * Also make sure that it wasn't scheduled on its rq.
1007                          */
1008                         if (unlikely(task_rq(task) != rq ||
1009                                      !cpu_isset(lowest_rq->cpu,
1010                                                 task->cpus_allowed) ||
1011                                      task_running(rq, task) ||
1012                                      !task->se.on_rq)) {
1013
1014                                 spin_unlock(&lowest_rq->lock);
1015                                 lowest_rq = NULL;
1016                                 break;
1017                         }
1018                 }
1019
1020                 /* If this rq is still suitable use it. */
1021                 if (lowest_rq->rt.highest_prio > task->prio)
1022                         break;
1023
1024                 /* try again */
1025                 spin_unlock(&lowest_rq->lock);
1026                 lowest_rq = NULL;
1027         }
1028
1029         return lowest_rq;
1030 }
1031
1032 /*
1033  * If the current CPU has more than one RT task, see if the non
1034  * running task can migrate over to a CPU that is running a task
1035  * of lesser priority.
1036  */
1037 static int push_rt_task(struct rq *rq)
1038 {
1039         struct task_struct *next_task;
1040         struct rq *lowest_rq;
1041         int ret = 0;
1042         int paranoid = RT_MAX_TRIES;
1043
1044         if (!rq->rt.overloaded)
1045                 return 0;
1046
1047         next_task = pick_next_highest_task_rt(rq, -1);
1048         if (!next_task)
1049                 return 0;
1050
1051  retry:
1052         if (unlikely(next_task == rq->curr)) {
1053                 WARN_ON(1);
1054                 return 0;
1055         }
1056
1057         /*
1058          * It's possible that the next_task slipped in of
1059          * higher priority than current. If that's the case
1060          * just reschedule current.
1061          */
1062         if (unlikely(next_task->prio < rq->curr->prio)) {
1063                 resched_task(rq->curr);
1064                 return 0;
1065         }
1066
1067         /* We might release rq lock */
1068         get_task_struct(next_task);
1069
1070         /* find_lock_lowest_rq locks the rq if found */
1071         lowest_rq = find_lock_lowest_rq(next_task, rq);
1072         if (!lowest_rq) {
1073                 struct task_struct *task;
1074                 /*
1075                  * find lock_lowest_rq releases rq->lock
1076                  * so it is possible that next_task has changed.
1077                  * If it has, then try again.
1078                  */
1079                 task = pick_next_highest_task_rt(rq, -1);
1080                 if (unlikely(task != next_task) && task && paranoid--) {
1081                         put_task_struct(next_task);
1082                         next_task = task;
1083                         goto retry;
1084                 }
1085                 goto out;
1086         }
1087
1088         deactivate_task(rq, next_task, 0);
1089         set_task_cpu(next_task, lowest_rq->cpu);
1090         activate_task(lowest_rq, next_task, 0);
1091
1092         resched_task(lowest_rq->curr);
1093
1094         spin_unlock(&lowest_rq->lock);
1095
1096         ret = 1;
1097 out:
1098         put_task_struct(next_task);
1099
1100         return ret;
1101 }
1102
1103 /*
1104  * TODO: Currently we just use the second highest prio task on
1105  *       the queue, and stop when it can't migrate (or there's
1106  *       no more RT tasks).  There may be a case where a lower
1107  *       priority RT task has a different affinity than the
1108  *       higher RT task. In this case the lower RT task could
1109  *       possibly be able to migrate where as the higher priority
1110  *       RT task could not.  We currently ignore this issue.
1111  *       Enhancements are welcome!
1112  */
1113 static void push_rt_tasks(struct rq *rq)
1114 {
1115         /* push_rt_task will return true if it moved an RT */
1116         while (push_rt_task(rq))
1117                 ;
1118 }
1119
1120 static int pull_rt_task(struct rq *this_rq)
1121 {
1122         int this_cpu = this_rq->cpu, ret = 0, cpu;
1123         struct task_struct *p, *next;
1124         struct rq *src_rq;
1125
1126         if (likely(!rt_overloaded(this_rq)))
1127                 return 0;
1128
1129         next = pick_next_task_rt(this_rq);
1130
1131         for_each_cpu_mask_nr(cpu, this_rq->rd->rto_mask) {
1132                 if (this_cpu == cpu)
1133                         continue;
1134
1135                 src_rq = cpu_rq(cpu);
1136                 /*
1137                  * We can potentially drop this_rq's lock in
1138                  * double_lock_balance, and another CPU could
1139                  * steal our next task - hence we must cause
1140                  * the caller to recalculate the next task
1141                  * in that case:
1142                  */
1143                 if (double_lock_balance(this_rq, src_rq)) {
1144                         struct task_struct *old_next = next;
1145
1146                         next = pick_next_task_rt(this_rq);
1147                         if (next != old_next)
1148                                 ret = 1;
1149                 }
1150
1151                 /*
1152                  * Are there still pullable RT tasks?
1153                  */
1154                 if (src_rq->rt.rt_nr_running <= 1)
1155                         goto skip;
1156
1157                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1158
1159                 /*
1160                  * Do we have an RT task that preempts
1161                  * the to-be-scheduled task?
1162                  */
1163                 if (p && (!next || (p->prio < next->prio))) {
1164                         WARN_ON(p == src_rq->curr);
1165                         WARN_ON(!p->se.on_rq);
1166
1167                         /*
1168                          * There's a chance that p is higher in priority
1169                          * than what's currently running on its cpu.
1170                          * This is just that p is wakeing up and hasn't
1171                          * had a chance to schedule. We only pull
1172                          * p if it is lower in priority than the
1173                          * current task on the run queue or
1174                          * this_rq next task is lower in prio than
1175                          * the current task on that rq.
1176                          */
1177                         if (p->prio < src_rq->curr->prio ||
1178                             (next && next->prio < src_rq->curr->prio))
1179                                 goto skip;
1180
1181                         ret = 1;
1182
1183                         deactivate_task(src_rq, p, 0);
1184                         set_task_cpu(p, this_cpu);
1185                         activate_task(this_rq, p, 0);
1186                         /*
1187                          * We continue with the search, just in
1188                          * case there's an even higher prio task
1189                          * in another runqueue. (low likelyhood
1190                          * but possible)
1191                          *
1192                          * Update next so that we won't pick a task
1193                          * on another cpu with a priority lower (or equal)
1194                          * than the one we just picked.
1195                          */
1196                         next = p;
1197
1198                 }
1199  skip:
1200                 spin_unlock(&src_rq->lock);
1201         }
1202
1203         return ret;
1204 }
1205
1206 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1207 {
1208         /* Try to pull RT tasks here if we lower this rq's prio */
1209         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1210                 pull_rt_task(rq);
1211 }
1212
1213 static void post_schedule_rt(struct rq *rq)
1214 {
1215         /*
1216          * If we have more than one rt_task queued, then
1217          * see if we can push the other rt_tasks off to other CPUS.
1218          * Note we may release the rq lock, and since
1219          * the lock was owned by prev, we need to release it
1220          * first via finish_lock_switch and then reaquire it here.
1221          */
1222         if (unlikely(rq->rt.overloaded)) {
1223                 spin_lock_irq(&rq->lock);
1224                 push_rt_tasks(rq);
1225                 spin_unlock_irq(&rq->lock);
1226         }
1227 }
1228
1229 /*
1230  * If we are not running and we are not going to reschedule soon, we should
1231  * try to push tasks away now
1232  */
1233 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1234 {
1235         if (!task_running(rq, p) &&
1236             !test_tsk_need_resched(rq->curr) &&
1237             rq->rt.overloaded)
1238                 push_rt_tasks(rq);
1239 }
1240
1241 static unsigned long
1242 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1243                 unsigned long max_load_move,
1244                 struct sched_domain *sd, enum cpu_idle_type idle,
1245                 int *all_pinned, int *this_best_prio)
1246 {
1247         /* don't touch RT tasks */
1248         return 0;
1249 }
1250
1251 static int
1252 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1253                  struct sched_domain *sd, enum cpu_idle_type idle)
1254 {
1255         /* don't touch RT tasks */
1256         return 0;
1257 }
1258
1259 static void set_cpus_allowed_rt(struct task_struct *p,
1260                                 const cpumask_t *new_mask)
1261 {
1262         int weight = cpus_weight(*new_mask);
1263
1264         BUG_ON(!rt_task(p));
1265
1266         /*
1267          * Update the migration status of the RQ if we have an RT task
1268          * which is running AND changing its weight value.
1269          */
1270         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1271                 struct rq *rq = task_rq(p);
1272
1273                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1274                         rq->rt.rt_nr_migratory++;
1275                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1276                         BUG_ON(!rq->rt.rt_nr_migratory);
1277                         rq->rt.rt_nr_migratory--;
1278                 }
1279
1280                 update_rt_migration(rq);
1281         }
1282
1283         p->cpus_allowed    = *new_mask;
1284         p->rt.nr_cpus_allowed = weight;
1285 }
1286
1287 /* Assumes rq->lock is held */
1288 static void rq_online_rt(struct rq *rq)
1289 {
1290         if (rq->rt.overloaded)
1291                 rt_set_overload(rq);
1292
1293         __enable_runtime(rq);
1294
1295         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1296 }
1297
1298 /* Assumes rq->lock is held */
1299 static void rq_offline_rt(struct rq *rq)
1300 {
1301         if (rq->rt.overloaded)
1302                 rt_clear_overload(rq);
1303
1304         __disable_runtime(rq);
1305
1306         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1307 }
1308
1309 /*
1310  * When switch from the rt queue, we bring ourselves to a position
1311  * that we might want to pull RT tasks from other runqueues.
1312  */
1313 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1314                            int running)
1315 {
1316         /*
1317          * If there are other RT tasks then we will reschedule
1318          * and the scheduling of the other RT tasks will handle
1319          * the balancing. But if we are the last RT task
1320          * we may need to handle the pulling of RT tasks
1321          * now.
1322          */
1323         if (!rq->rt.rt_nr_running)
1324                 pull_rt_task(rq);
1325 }
1326 #endif /* CONFIG_SMP */
1327
1328 /*
1329  * When switching a task to RT, we may overload the runqueue
1330  * with RT tasks. In this case we try to push them off to
1331  * other runqueues.
1332  */
1333 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1334                            int running)
1335 {
1336         int check_resched = 1;
1337
1338         /*
1339          * If we are already running, then there's nothing
1340          * that needs to be done. But if we are not running
1341          * we may need to preempt the current running task.
1342          * If that current running task is also an RT task
1343          * then see if we can move to another run queue.
1344          */
1345         if (!running) {
1346 #ifdef CONFIG_SMP
1347                 if (rq->rt.overloaded && push_rt_task(rq) &&
1348                     /* Don't resched if we changed runqueues */
1349                     rq != task_rq(p))
1350                         check_resched = 0;
1351 #endif /* CONFIG_SMP */
1352                 if (check_resched && p->prio < rq->curr->prio)
1353                         resched_task(rq->curr);
1354         }
1355 }
1356
1357 /*
1358  * Priority of the task has changed. This may cause
1359  * us to initiate a push or pull.
1360  */
1361 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1362                             int oldprio, int running)
1363 {
1364         if (running) {
1365 #ifdef CONFIG_SMP
1366                 /*
1367                  * If our priority decreases while running, we
1368                  * may need to pull tasks to this runqueue.
1369                  */
1370                 if (oldprio < p->prio)
1371                         pull_rt_task(rq);
1372                 /*
1373                  * If there's a higher priority task waiting to run
1374                  * then reschedule. Note, the above pull_rt_task
1375                  * can release the rq lock and p could migrate.
1376                  * Only reschedule if p is still on the same runqueue.
1377                  */
1378                 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1379                         resched_task(p);
1380 #else
1381                 /* For UP simply resched on drop of prio */
1382                 if (oldprio < p->prio)
1383                         resched_task(p);
1384 #endif /* CONFIG_SMP */
1385         } else {
1386                 /*
1387                  * This task is not running, but if it is
1388                  * greater than the current running task
1389                  * then reschedule.
1390                  */
1391                 if (p->prio < rq->curr->prio)
1392                         resched_task(rq->curr);
1393         }
1394 }
1395
1396 static void watchdog(struct rq *rq, struct task_struct *p)
1397 {
1398         unsigned long soft, hard;
1399
1400         if (!p->signal)
1401                 return;
1402
1403         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1404         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1405
1406         if (soft != RLIM_INFINITY) {
1407                 unsigned long next;
1408
1409                 p->rt.timeout++;
1410                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1411                 if (p->rt.timeout > next)
1412                         p->it_sched_expires = p->se.sum_exec_runtime;
1413         }
1414 }
1415
1416 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1417 {
1418         update_curr_rt(rq);
1419
1420         watchdog(rq, p);
1421
1422         /*
1423          * RR tasks need a special form of timeslice management.
1424          * FIFO tasks have no timeslices.
1425          */
1426         if (p->policy != SCHED_RR)
1427                 return;
1428
1429         if (--p->rt.time_slice)
1430                 return;
1431
1432         p->rt.time_slice = DEF_TIMESLICE;
1433
1434         /*
1435          * Requeue to the end of queue if we are not the only element
1436          * on the queue:
1437          */
1438         if (p->rt.run_list.prev != p->rt.run_list.next) {
1439                 requeue_task_rt(rq, p, 0);
1440                 set_tsk_need_resched(p);
1441         }
1442 }
1443
1444 static void set_curr_task_rt(struct rq *rq)
1445 {
1446         struct task_struct *p = rq->curr;
1447
1448         p->se.exec_start = rq->clock;
1449 }
1450
1451 static const struct sched_class rt_sched_class = {
1452         .next                   = &fair_sched_class,
1453         .enqueue_task           = enqueue_task_rt,
1454         .dequeue_task           = dequeue_task_rt,
1455         .yield_task             = yield_task_rt,
1456 #ifdef CONFIG_SMP
1457         .select_task_rq         = select_task_rq_rt,
1458 #endif /* CONFIG_SMP */
1459
1460         .check_preempt_curr     = check_preempt_curr_rt,
1461
1462         .pick_next_task         = pick_next_task_rt,
1463         .put_prev_task          = put_prev_task_rt,
1464
1465 #ifdef CONFIG_SMP
1466         .load_balance           = load_balance_rt,
1467         .move_one_task          = move_one_task_rt,
1468         .set_cpus_allowed       = set_cpus_allowed_rt,
1469         .rq_online              = rq_online_rt,
1470         .rq_offline             = rq_offline_rt,
1471         .pre_schedule           = pre_schedule_rt,
1472         .post_schedule          = post_schedule_rt,
1473         .task_wake_up           = task_wake_up_rt,
1474         .switched_from          = switched_from_rt,
1475 #endif
1476
1477         .set_curr_task          = set_curr_task_rt,
1478         .task_tick              = task_tick_rt,
1479
1480         .prio_changed           = prio_changed_rt,
1481         .switched_to            = switched_to_rt,
1482 };
1483
1484 #ifdef CONFIG_SCHED_DEBUG
1485 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1486
1487 static void print_rt_stats(struct seq_file *m, int cpu)
1488 {
1489         struct rt_rq *rt_rq;
1490
1491         rcu_read_lock();
1492         for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1493                 print_rt_rq(m, cpu, rt_rq);
1494         rcu_read_unlock();
1495 }
1496 #endif /* CONFIG_SCHED_DEBUG */