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         if (rt_rq->rt_nr_running)
 203                 resched_task(rq_of_rt_rq(rt_rq)->curr);
 204 }
 205
 206 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 207 {
 208 }
 209
 210 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 211 {
 212         return rt_rq->rt_throttled;
 213 }
 214
 215 static inline cpumask_t sched_rt_period_mask(void)
 216 {
 217         return cpu_online_map;
 218 }
 219
 220 static inline
 221 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 222 {
 223         return &cpu_rq(cpu)->rt;
 224 }
 225
 226 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 227 {
 228         return &def_rt_bandwidth;
 229 }
 230
 231 #endif /* CONFIG_RT_GROUP_SCHED */
 232
 233 #ifdef CONFIG_SMP
 234 static int do_balance_runtime(struct rt_rq *rt_rq)
 235 {
 236         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 237         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 238         int i, weight, more = 0;
 239         u64 rt_period;
 240
 241         weight = cpus_weight(rd->span);
 242
 243         spin_lock(&rt_b->rt_runtime_lock);
 244         rt_period = ktime_to_ns(rt_b->rt_period);
 245         for_each_cpu_mask_nr(i, rd->span) {
 246                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 247                 s64 diff;
 248
 249                 if (iter == rt_rq)
 250                         continue;
 251
 252                 spin_lock(&iter->rt_runtime_lock);
 253                 if (iter->rt_runtime == RUNTIME_INF)
 254                         goto next;
 255
 256                 diff = iter->rt_runtime - iter->rt_time;
 257                 if (diff > 0) {
 258                         diff = div_u64((u64)diff, weight);
 259                         if (rt_rq->rt_runtime + diff > rt_period)
 260                                 diff = rt_period - rt_rq->rt_runtime;
 261                         iter->rt_runtime -= diff;
 262                         rt_rq->rt_runtime += diff;
 263                         more = 1;
 264                         if (rt_rq->rt_runtime == rt_period) {
 265                                 spin_unlock(&iter->rt_runtime_lock);
 266                                 break;
 267                         }
 268                 }
 269 next:
 270                 spin_unlock(&iter->rt_runtime_lock);
 271         }
 272         spin_unlock(&rt_b->rt_runtime_lock);
 273
 274         return more;
 275 }
 276
 277 static void __disable_runtime(struct rq *rq)
 278 {
 279         struct root_domain *rd = rq->rd;
 280         struct rt_rq *rt_rq;
 281
 282         if (unlikely(!scheduler_running))
 283                 return;
 284
 285         for_each_leaf_rt_rq(rt_rq, rq) {
 286                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 287                 s64 want;
 288                 int i;
 289
 290                 spin_lock(&rt_b->rt_runtime_lock);
 291                 spin_lock(&rt_rq->rt_runtime_lock);
 292                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 293                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 294                         goto balanced;
 295                 spin_unlock(&rt_rq->rt_runtime_lock);
 296
 297                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 298
 299                 for_each_cpu_mask(i, rd->span) {
 300                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 301                         s64 diff;
 302
 303                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 304                                 continue;
 305
 306                         spin_lock(&iter->rt_runtime_lock);
 307                         if (want > 0) {
 308                                 diff = min_t(s64, iter->rt_runtime, want);
 309                                 iter->rt_runtime -= diff;
 310                                 want -= diff;
 311                         } else {
 312                                 iter->rt_runtime -= want;
 313                                 want -= want;
 314                         }
 315                         spin_unlock(&iter->rt_runtime_lock);
 316
 317                         if (!want)
 318                                 break;
 319                 }
 320
 321                 spin_lock(&rt_rq->rt_runtime_lock);
 322                 BUG_ON(want);
 323 balanced:
 324                 rt_rq->rt_runtime = RUNTIME_INF;
 325                 spin_unlock(&rt_rq->rt_runtime_lock);
 326                 spin_unlock(&rt_b->rt_runtime_lock);
 327         }
 328 }
 329
 330 static void disable_runtime(struct rq *rq)
 331 {
 332         unsigned long flags;
 333
 334         spin_lock_irqsave(&rq->lock, flags);
 335         __disable_runtime(rq);
 336         spin_unlock_irqrestore(&rq->lock, flags);
 337 }
 338
 339 static void __enable_runtime(struct rq *rq)
 340 {
 341         struct rt_rq *rt_rq;
 342
 343         if (unlikely(!scheduler_running))
 344                 return;
 345
 346         for_each_leaf_rt_rq(rt_rq, rq) {
 347                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 348
 349                 spin_lock(&rt_b->rt_runtime_lock);
 350                 spin_lock(&rt_rq->rt_runtime_lock);
 351                 rt_rq->rt_runtime = rt_b->rt_runtime;
 352                 rt_rq->rt_time = 0;
 353                 spin_unlock(&rt_rq->rt_runtime_lock);
 354                 spin_unlock(&rt_b->rt_runtime_lock);
 355         }
 356 }
 357
 358 static void enable_runtime(struct rq *rq)
 359 {
 360         unsigned long flags;
 361
 362         spin_lock_irqsave(&rq->lock, flags);
 363         __enable_runtime(rq);
 364         spin_unlock_irqrestore(&rq->lock, flags);
 365 }
 366
 367 static int balance_runtime(struct rt_rq *rt_rq)
 368 {
 369         int more = 0;
 370
 371         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 372                 spin_unlock(&rt_rq->rt_runtime_lock);
 373                 more = do_balance_runtime(rt_rq);
 374                 spin_lock(&rt_rq->rt_runtime_lock);
 375         }
 376
 377         return more;
 378 }
 379 #else /* !CONFIG_SMP */
 380 static inline int balance_runtime(struct rt_rq *rt_rq)
 381 {
 382         return 0;
 383 }
 384 #endif /* CONFIG_SMP */
 385
 386 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 387 {
 388         int i, idle = 1;
 389         cpumask_t span;
 390
 391         if (rt_b->rt_runtime == RUNTIME_INF)
 392                 return 1;
 393
 394         span = sched_rt_period_mask();
 395         for_each_cpu_mask(i, span) {
 396                 int enqueue = 0;
 397                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 398                 struct rq *rq = rq_of_rt_rq(rt_rq);
 399
 400                 spin_lock(&rq->lock);
 401                 if (rt_rq->rt_time) {
 402                         u64 runtime;
 403
 404                         spin_lock(&rt_rq->rt_runtime_lock);
 405                         if (rt_rq->rt_throttled)
 406                                 balance_runtime(rt_rq);
 407                         runtime = rt_rq->rt_runtime;
 408                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 409                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 410                                 rt_rq->rt_throttled = 0;
 411                                 enqueue = 1;
 412                         }
 413                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 414                                 idle = 0;
 415                         spin_unlock(&rt_rq->rt_runtime_lock);
 416                 } else if (rt_rq->rt_nr_running)
 417                         idle = 0;
 418
 419                 if (enqueue)
 420                         sched_rt_rq_enqueue(rt_rq);
 421                 spin_unlock(&rq->lock);
 422         }
 423
 424         return idle;
 425 }
 426
 427 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 428 {
 429 #ifdef CONFIG_RT_GROUP_SCHED
 430         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 431
 432         if (rt_rq)
 433                 return rt_rq->highest_prio;
 434 #endif
 435
 436         return rt_task_of(rt_se)->prio;
 437 }
 438
 439 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 440 {
 441         u64 runtime = sched_rt_runtime(rt_rq);
 442
 443         if (rt_rq->rt_throttled)
 444                 return rt_rq_throttled(rt_rq);
 445
 446         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 447                 return 0;
 448
 449         balance_runtime(rt_rq);
 450         runtime = sched_rt_runtime(rt_rq);
 451         if (runtime == RUNTIME_INF)
 452                 return 0;
 453
 454         if (rt_rq->rt_time > runtime) {
 455                 rt_rq->rt_throttled = 1;
 456                 if (rt_rq_throttled(rt_rq)) {
 457                         sched_rt_rq_dequeue(rt_rq);
 458                         return 1;
 459                 }
 460         }
 461
 462         return 0;
 463 }
 464
 465 /*
 466  * Update the current task's runtime statistics. Skip current tasks that
 467  * are not in our scheduling class.
 468  */
 469 static void update_curr_rt(struct rq *rq)
 470 {
 471         struct task_struct *curr = rq->curr;
 472         struct sched_rt_entity *rt_se = &curr->rt;
 473         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 474         u64 delta_exec;
 475
 476         if (!task_has_rt_policy(curr))
 477                 return;
 478
 479         delta_exec = rq->clock - curr->se.exec_start;
 480         if (unlikely((s64)delta_exec < 0))
 481                 delta_exec = 0;
 482
 483         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
 484
 485         curr->se.sum_exec_runtime += delta_exec;
 486         account_group_exec_runtime(curr, delta_exec);
 487
 488         curr->se.exec_start = rq->clock;
 489         cpuacct_charge(curr, delta_exec);
 490
 491         for_each_sched_rt_entity(rt_se) {
 492                 rt_rq = rt_rq_of_se(rt_se);
 493
 494                 spin_lock(&rt_rq->rt_runtime_lock);
 495                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 496                         rt_rq->rt_time += delta_exec;
 497                         if (sched_rt_runtime_exceeded(rt_rq))
 498                                 resched_task(curr);
 499                 }
 500                 spin_unlock(&rt_rq->rt_runtime_lock);
 501         }
 502 }
 503
 504 static inline
 505 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 506 {
 507         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 508         rt_rq->rt_nr_running++;
 509 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 510         if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
 511 #ifdef CONFIG_SMP
 512                 struct rq *rq = rq_of_rt_rq(rt_rq);
 513 #endif
 514
 515                 rt_rq->highest_prio = rt_se_prio(rt_se);
 516 #ifdef CONFIG_SMP
 517                 if (rq->online)
 518                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 519                                    rt_se_prio(rt_se));
 520 #endif
 521         }
 522 #endif
 523 #ifdef CONFIG_SMP
 524         if (rt_se->nr_cpus_allowed > 1) {
 525                 struct rq *rq = rq_of_rt_rq(rt_rq);
 526
 527                 rq->rt.rt_nr_migratory++;
 528         }
 529
 530         update_rt_migration(rq_of_rt_rq(rt_rq));
 531 #endif
 532 #ifdef CONFIG_RT_GROUP_SCHED
 533         if (rt_se_boosted(rt_se))
 534                 rt_rq->rt_nr_boosted++;
 535
 536         if (rt_rq->tg)
 537                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 538 #else
 539         start_rt_bandwidth(&def_rt_bandwidth);
 540 #endif
 541 }
 542
 543 static inline
 544 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 545 {
 546 #ifdef CONFIG_SMP
 547         int highest_prio = rt_rq->highest_prio;
 548 #endif
 549
 550         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 551         WARN_ON(!rt_rq->rt_nr_running);
 552         rt_rq->rt_nr_running--;
 553 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 554         if (rt_rq->rt_nr_running) {
 555                 struct rt_prio_array *array;
 556
 557                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
 558                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
 559                         /* recalculate */
 560                         array = &rt_rq->active;
 561                         rt_rq->highest_prio =
 562                                 sched_find_first_bit(array->bitmap);
 563                 } /* otherwise leave rq->highest prio alone */
 564         } else
 565                 rt_rq->highest_prio = MAX_RT_PRIO;
 566 #endif
 567 #ifdef CONFIG_SMP
 568         if (rt_se->nr_cpus_allowed > 1) {
 569                 struct rq *rq = rq_of_rt_rq(rt_rq);
 570                 rq->rt.rt_nr_migratory--;
 571         }
 572
 573         if (rt_rq->highest_prio != highest_prio) {
 574                 struct rq *rq = rq_of_rt_rq(rt_rq);
 575
 576                 if (rq->online)
 577                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 578                                    rt_rq->highest_prio);
 579         }
 580
 581         update_rt_migration(rq_of_rt_rq(rt_rq));
 582 #endif /* CONFIG_SMP */
 583 #ifdef CONFIG_RT_GROUP_SCHED
 584         if (rt_se_boosted(rt_se))
 585                 rt_rq->rt_nr_boosted--;
 586
 587         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 588 #endif
 589 }
 590
 591 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
 592 {
 593         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 594         struct rt_prio_array *array = &rt_rq->active;
 595         struct rt_rq *group_rq = group_rt_rq(rt_se);
 596         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 597
 598         /*
 599          * Don't enqueue the group if its throttled, or when empty.
 600          * The latter is a consequence of the former when a child group
 601          * get throttled and the current group doesn't have any other
 602          * active members.
 603          */
 604         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 605                 return;
 606
 607         list_add_tail(&rt_se->run_list, queue);
 608         __set_bit(rt_se_prio(rt_se), array->bitmap);
 609
 610         inc_rt_tasks(rt_se, rt_rq);
 611 }
 612
 613 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 614 {
 615         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 616         struct rt_prio_array *array = &rt_rq->active;
 617
 618         list_del_init(&rt_se->run_list);
 619         if (list_empty(array->queue + rt_se_prio(rt_se)))
 620                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 621
 622         dec_rt_tasks(rt_se, rt_rq);
 623 }
 624
 625 /*
 626  * Because the prio of an upper entry depends on the lower
 627  * entries, we must remove entries top - down.
 628  */
 629 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 630 {
 631         struct sched_rt_entity *back = NULL;
 632
 633         for_each_sched_rt_entity(rt_se) {
 634                 rt_se->back = back;
 635                 back = rt_se;
 636         }
 637
 638         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 639                 if (on_rt_rq(rt_se))
 640                         __dequeue_rt_entity(rt_se);
 641         }
 642 }
 643
 644 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 645 {
 646         dequeue_rt_stack(rt_se);
 647         for_each_sched_rt_entity(rt_se)
 648                 __enqueue_rt_entity(rt_se);
 649 }
 650
 651 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 652 {
 653         dequeue_rt_stack(rt_se);
 654
 655         for_each_sched_rt_entity(rt_se) {
 656                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 657
 658                 if (rt_rq && rt_rq->rt_nr_running)
 659                         __enqueue_rt_entity(rt_se);
 660         }
 661 }
 662
 663 /*
 664  * Adding/removing a task to/from a priority array:
 665  */
 666 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 667 {
 668         struct sched_rt_entity *rt_se = &p->rt;
 669
 670         if (wakeup)
 671                 rt_se->timeout = 0;
 672
 673         enqueue_rt_entity(rt_se);
 674
 675         inc_cpu_load(rq, p->se.load.weight);
 676 }
 677
 678 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 679 {
 680         struct sched_rt_entity *rt_se = &p->rt;
 681
 682         update_curr_rt(rq);
 683         dequeue_rt_entity(rt_se);
 684
 685         dec_cpu_load(rq, p->se.load.weight);
 686 }
 687
 688 /*
 689  * Put task to the end of the run list without the overhead of dequeue
 690  * followed by enqueue.
 691  */
 692 static void
 693 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 694 {
 695         if (on_rt_rq(rt_se)) {
 696                 struct rt_prio_array *array = &rt_rq->active;
 697                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 698
 699                 if (head)
 700                         list_move(&rt_se->run_list, queue);
 701                 else
 702                         list_move_tail(&rt_se->run_list, queue);
 703         }
 704 }
 705
 706 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 707 {
 708         struct sched_rt_entity *rt_se = &p->rt;
 709         struct rt_rq *rt_rq;
 710
 711         for_each_sched_rt_entity(rt_se) {
 712                 rt_rq = rt_rq_of_se(rt_se);
 713                 requeue_rt_entity(rt_rq, rt_se, head);
 714         }
 715 }
 716
 717 static void yield_task_rt(struct rq *rq)
 718 {
 719         requeue_task_rt(rq, rq->curr, 0);
 720 }
 721
 722 #ifdef CONFIG_SMP
 723 static int find_lowest_rq(struct task_struct *task);
 724
 725 static int select_task_rq_rt(struct task_struct *p, int sync)
 726 {
 727         struct rq *rq = task_rq(p);
 728
 729         /*
 730          * If the current task is an RT task, then
 731          * try to see if we can wake this RT task up on another
 732          * runqueue. Otherwise simply start this RT task
 733          * on its current runqueue.
 734          *
 735          * We want to avoid overloading runqueues. Even if
 736          * the RT task is of higher priority than the current RT task.
 737          * RT tasks behave differently than other tasks. If
 738          * one gets preempted, we try to push it off to another queue.
 739          * So trying to keep a preempting RT task on the same
 740          * cache hot CPU will force the running RT task to
 741          * a cold CPU. So we waste all the cache for the lower
 742          * RT task in hopes of saving some of a RT task
 743          * that is just being woken and probably will have
 744          * cold cache anyway.
 745          */
 746         if (unlikely(rt_task(rq->curr)) &&
 747             (p->rt.nr_cpus_allowed > 1)) {
 748                 int cpu = find_lowest_rq(p);
 749
 750                 return (cpu == -1) ? task_cpu(p) : cpu;
 751         }
 752
 753         /*
 754          * Otherwise, just let it ride on the affined RQ and the
 755          * post-schedule router will push the preempted task away
 756          */
 757         return task_cpu(p);
 758 }
 759
 760 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
 761 {
 762         cpumask_t mask;
 763
 764         if (rq->curr->rt.nr_cpus_allowed == 1)
 765                 return;
 766
 767         if (p->rt.nr_cpus_allowed != 1
 768             && cpupri_find(&rq->rd->cpupri, p, &mask))
 769                 return;
 770
 771         if (!cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
 772                 return;
 773
 774         /*
 775          * There appears to be other cpus that can accept
 776          * current and none to run 'p', so lets reschedule
 777          * to try and push current away:
 778          */
 779         requeue_task_rt(rq, p, 1);
 780         resched_task(rq->curr);
 781 }
 782
 783 #endif /* CONFIG_SMP */
 784
 785 /*
 786  * Preempt the current task with a newly woken task if needed:
 787  */
 788 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
 789 {
 790         if (p->prio < rq->curr->prio) {
 791                 resched_task(rq->curr);
 792                 return;
 793         }
 794
 795 #ifdef CONFIG_SMP
 796         /*
 797          * If:
 798          *
 799          * - the newly woken task is of equal priority to the current task
 800          * - the newly woken task is non-migratable while current is migratable
 801          * - current will be preempted on the next reschedule
 802          *
 803          * we should check to see if current can readily move to a different
 804          * cpu.  If so, we will reschedule to allow the push logic to try
 805          * to move current somewhere else, making room for our non-migratable
 806          * task.
 807          */
 808         if (p->prio == rq->curr->prio && !need_resched())
 809                 check_preempt_equal_prio(rq, p);
 810 #endif
 811 }
 812
 813 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 814                                                    struct rt_rq *rt_rq)
 815 {
 816         struct rt_prio_array *array = &rt_rq->active;
 817         struct sched_rt_entity *next = NULL;
 818         struct list_head *queue;
 819         int idx;
 820
 821         idx = sched_find_first_bit(array->bitmap);
 822         BUG_ON(idx >= MAX_RT_PRIO);
 823
 824         queue = array->queue + idx;
 825         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 826
 827         return next;
 828 }
 829
 830 static struct task_struct *pick_next_task_rt(struct rq *rq)
 831 {
 832         struct sched_rt_entity *rt_se;
 833         struct task_struct *p;
 834         struct rt_rq *rt_rq;
 835
 836         rt_rq = &rq->rt;
 837
 838         if (unlikely(!rt_rq->rt_nr_running))
 839                 return NULL;
 840
 841         if (rt_rq_throttled(rt_rq))
 842                 return NULL;
 843
 844         do {
 845                 rt_se = pick_next_rt_entity(rq, rt_rq);
 846                 BUG_ON(!rt_se);
 847                 rt_rq = group_rt_rq(rt_se);
 848         } while (rt_rq);
 849
 850         p = rt_task_of(rt_se);
 851         p->se.exec_start = rq->clock;
 852         return p;
 853 }
 854
 855 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 856 {
 857         update_curr_rt(rq);
 858         p->se.exec_start = 0;
 859 }
 860
 861 #ifdef CONFIG_SMP
 862
 863 /* Only try algorithms three times */
 864 #define RT_MAX_TRIES 3
 865
 866 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
 867 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest);
 868
 869 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 870
 871 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 872 {
 873         if (!task_running(rq, p) &&
 874             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
 875             (p->rt.nr_cpus_allowed > 1))
 876                 return 1;
 877         return 0;
 878 }
 879
 880 /* Return the second highest RT task, NULL otherwise */
 881 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
 882 {
 883         struct task_struct *next = NULL;
 884         struct sched_rt_entity *rt_se;
 885         struct rt_prio_array *array;
 886         struct rt_rq *rt_rq;
 887         int idx;
 888
 889         for_each_leaf_rt_rq(rt_rq, rq) {
 890                 array = &rt_rq->active;
 891                 idx = sched_find_first_bit(array->bitmap);
 892  next_idx:
 893                 if (idx >= MAX_RT_PRIO)
 894                         continue;
 895                 if (next && next->prio < idx)
 896                         continue;
 897                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
 898                         struct task_struct *p = rt_task_of(rt_se);
 899                         if (pick_rt_task(rq, p, cpu)) {
 900                                 next = p;
 901                                 break;
 902                         }
 903                 }
 904                 if (!next) {
 905                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 906                         goto next_idx;
 907                 }
 908         }
 909
 910         return next;
 911 }
 912
 913 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
 914
 915 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
 916 {
 917         int first;
 918
 919         /* "this_cpu" is cheaper to preempt than a remote processor */
 920         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
 921                 return this_cpu;
 922
 923         first = first_cpu(*mask);
 924         if (first != NR_CPUS)
 925                 return first;
 926
 927         return -1;
 928 }
 929
 930 static int find_lowest_rq(struct task_struct *task)
 931 {
 932         struct sched_domain *sd;
 933         cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
 934         int this_cpu = smp_processor_id();
 935         int cpu      = task_cpu(task);
 936
 937         if (task->rt.nr_cpus_allowed == 1)
 938                 return -1; /* No other targets possible */
 939
 940         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
 941                 return -1; /* No targets found */
 942
 943         /*
 944          * Only consider CPUs that are usable for migration.
 945          * I guess we might want to change cpupri_find() to ignore those
 946          * in the first place.
 947          */
 948         cpus_and(*lowest_mask, *lowest_mask, cpu_active_map);
 949
 950         /*
 951          * At this point we have built a mask of cpus representing the
 952          * lowest priority tasks in the system.  Now we want to elect
 953          * the best one based on our affinity and topology.
 954          *
 955          * We prioritize the last cpu that the task executed on since
 956          * it is most likely cache-hot in that location.
 957          */
 958         if (cpu_isset(cpu, *lowest_mask))
 959                 return cpu;
 960
 961         /*
 962          * Otherwise, we consult the sched_domains span maps to figure
 963          * out which cpu is logically closest to our hot cache data.
 964          */
 965         if (this_cpu == cpu)
 966                 this_cpu = -1; /* Skip this_cpu opt if the same */
 967
 968         for_each_domain(cpu, sd) {
 969                 if (sd->flags & SD_WAKE_AFFINE) {
 970                         cpumask_t domain_mask;
 971                         int       best_cpu;
 972
 973                         cpus_and(domain_mask, sd->span, *lowest_mask);
 974
 975                         best_cpu = pick_optimal_cpu(this_cpu,
 976                                                     &domain_mask);
 977                         if (best_cpu != -1)
 978                                 return best_cpu;
 979                 }
 980         }
 981
 982         /*
 983          * And finally, if there were no matches within the domains
 984          * just give the caller *something* to work with from the compatible
 985          * locations.
 986          */
 987         return pick_optimal_cpu(this_cpu, lowest_mask);
 988 }
 989
 990 /* Will lock the rq it finds */
 991 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
 992 {
 993         struct rq *lowest_rq = NULL;
 994         int tries;
 995         int cpu;
 996
 997         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
 998                 cpu = find_lowest_rq(task);
 999
1000                 if ((cpu == -1) || (cpu == rq->cpu))
1001                         break;
1002
1003                 lowest_rq = cpu_rq(cpu);
1004
1005                 /* if the prio of this runqueue changed, try again */
1006                 if (double_lock_balance(rq, lowest_rq)) {
1007                         /*
1008                          * We had to unlock the run queue. In
1009                          * the mean time, task could have
1010                          * migrated already or had its affinity changed.
1011                          * Also make sure that it wasn't scheduled on its rq.
1012                          */
1013                         if (unlikely(task_rq(task) != rq ||
1014                                      !cpu_isset(lowest_rq->cpu,
1015                                                 task->cpus_allowed) ||
1016                                      task_running(rq, task) ||
1017                                      !task->se.on_rq)) {
1018
1019                                 spin_unlock(&lowest_rq->lock);
1020                                 lowest_rq = NULL;
1021                                 break;
1022                         }
1023                 }
1024
1025                 /* If this rq is still suitable use it. */
1026                 if (lowest_rq->rt.highest_prio > task->prio)
1027                         break;
1028
1029                 /* try again */
1030                 double_unlock_balance(rq, lowest_rq);
1031                 lowest_rq = NULL;
1032         }
1033
1034         return lowest_rq;
1035 }
1036
1037 /*
1038  * If the current CPU has more than one RT task, see if the non
1039  * running task can migrate over to a CPU that is running a task
1040  * of lesser priority.
1041  */
1042 static int push_rt_task(struct rq *rq)
1043 {
1044         struct task_struct *next_task;
1045         struct rq *lowest_rq;
1046         int ret = 0;
1047         int paranoid = RT_MAX_TRIES;
1048
1049         if (!rq->rt.overloaded)
1050                 return 0;
1051
1052         next_task = pick_next_highest_task_rt(rq, -1);
1053         if (!next_task)
1054                 return 0;
1055
1056  retry:
1057         if (unlikely(next_task == rq->curr)) {
1058                 WARN_ON(1);
1059                 return 0;
1060         }
1061
1062         /*
1063          * It's possible that the next_task slipped in of
1064          * higher priority than current. If that's the case
1065          * just reschedule current.
1066          */
1067         if (unlikely(next_task->prio < rq->curr->prio)) {
1068                 resched_task(rq->curr);
1069                 return 0;
1070         }
1071
1072         /* We might release rq lock */
1073         get_task_struct(next_task);
1074
1075         /* find_lock_lowest_rq locks the rq if found */
1076         lowest_rq = find_lock_lowest_rq(next_task, rq);
1077         if (!lowest_rq) {
1078                 struct task_struct *task;
1079                 /*
1080                  * find lock_lowest_rq releases rq->lock
1081                  * so it is possible that next_task has changed.
1082                  * If it has, then try again.
1083                  */
1084                 task = pick_next_highest_task_rt(rq, -1);
1085                 if (unlikely(task != next_task) && task && paranoid--) {
1086                         put_task_struct(next_task);
1087                         next_task = task;
1088                         goto retry;
1089                 }
1090                 goto out;
1091         }
1092
1093         deactivate_task(rq, next_task, 0);
1094         set_task_cpu(next_task, lowest_rq->cpu);
1095         activate_task(lowest_rq, next_task, 0);
1096
1097         resched_task(lowest_rq->curr);
1098
1099         double_unlock_balance(rq, lowest_rq);
1100
1101         ret = 1;
1102 out:
1103         put_task_struct(next_task);
1104
1105         return ret;
1106 }
1107
1108 /*
1109  * TODO: Currently we just use the second highest prio task on
1110  *       the queue, and stop when it can't migrate (or there's
1111  *       no more RT tasks).  There may be a case where a lower
1112  *       priority RT task has a different affinity than the
1113  *       higher RT task. In this case the lower RT task could
1114  *       possibly be able to migrate where as the higher priority
1115  *       RT task could not.  We currently ignore this issue.
1116  *       Enhancements are welcome!
1117  */
1118 static void push_rt_tasks(struct rq *rq)
1119 {
1120         /* push_rt_task will return true if it moved an RT */
1121         while (push_rt_task(rq))
1122                 ;
1123 }
1124
1125 static int pull_rt_task(struct rq *this_rq)
1126 {
1127         int this_cpu = this_rq->cpu, ret = 0, cpu;
1128         struct task_struct *p, *next;
1129         struct rq *src_rq;
1130
1131         if (likely(!rt_overloaded(this_rq)))
1132                 return 0;
1133
1134         next = pick_next_task_rt(this_rq);
1135
1136         for_each_cpu_mask_nr(cpu, this_rq->rd->rto_mask) {
1137                 if (this_cpu == cpu)
1138                         continue;
1139
1140                 src_rq = cpu_rq(cpu);
1141                 /*
1142                  * We can potentially drop this_rq's lock in
1143                  * double_lock_balance, and another CPU could
1144                  * steal our next task - hence we must cause
1145                  * the caller to recalculate the next task
1146                  * in that case:
1147                  */
1148                 if (double_lock_balance(this_rq, src_rq)) {
1149                         struct task_struct *old_next = next;
1150
1151                         next = pick_next_task_rt(this_rq);
1152                         if (next != old_next)
1153                                 ret = 1;
1154                 }
1155
1156                 /*
1157                  * Are there still pullable RT tasks?
1158                  */
1159                 if (src_rq->rt.rt_nr_running <= 1)
1160                         goto skip;
1161
1162                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1163
1164                 /*
1165                  * Do we have an RT task that preempts
1166                  * the to-be-scheduled task?
1167                  */
1168                 if (p && (!next || (p->prio < next->prio))) {
1169                         WARN_ON(p == src_rq->curr);
1170                         WARN_ON(!p->se.on_rq);
1171
1172                         /*
1173                          * There's a chance that p is higher in priority
1174                          * than what's currently running on its cpu.
1175                          * This is just that p is wakeing up and hasn't
1176                          * had a chance to schedule. We only pull
1177                          * p if it is lower in priority than the
1178                          * current task on the run queue or
1179                          * this_rq next task is lower in prio than
1180                          * the current task on that rq.
1181                          */
1182                         if (p->prio < src_rq->curr->prio ||
1183                             (next && next->prio < src_rq->curr->prio))
1184                                 goto skip;
1185
1186                         ret = 1;
1187
1188                         deactivate_task(src_rq, p, 0);
1189                         set_task_cpu(p, this_cpu);
1190                         activate_task(this_rq, p, 0);
1191                         /*
1192                          * We continue with the search, just in
1193                          * case there's an even higher prio task
1194                          * in another runqueue. (low likelyhood
1195                          * but possible)
1196                          *
1197                          * Update next so that we won't pick a task
1198                          * on another cpu with a priority lower (or equal)
1199                          * than the one we just picked.
1200                          */
1201                         next = p;
1202
1203                 }
1204  skip:
1205                 double_unlock_balance(this_rq, src_rq);
1206         }
1207
1208         return ret;
1209 }
1210
1211 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1212 {
1213         /* Try to pull RT tasks here if we lower this rq's prio */
1214         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1215                 pull_rt_task(rq);
1216 }
1217
1218 static void post_schedule_rt(struct rq *rq)
1219 {
1220         /*
1221          * If we have more than one rt_task queued, then
1222          * see if we can push the other rt_tasks off to other CPUS.
1223          * Note we may release the rq lock, and since
1224          * the lock was owned by prev, we need to release it
1225          * first via finish_lock_switch and then reaquire it here.
1226          */
1227         if (unlikely(rq->rt.overloaded)) {
1228                 spin_lock_irq(&rq->lock);
1229                 push_rt_tasks(rq);
1230                 spin_unlock_irq(&rq->lock);
1231         }
1232 }
1233
1234 /*
1235  * If we are not running and we are not going to reschedule soon, we should
1236  * try to push tasks away now
1237  */
1238 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1239 {
1240         if (!task_running(rq, p) &&
1241             !test_tsk_need_resched(rq->curr) &&
1242             rq->rt.overloaded)
1243                 push_rt_tasks(rq);
1244 }
1245
1246 static unsigned long
1247 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1248                 unsigned long max_load_move,
1249                 struct sched_domain *sd, enum cpu_idle_type idle,
1250                 int *all_pinned, int *this_best_prio)
1251 {
1252         /* don't touch RT tasks */
1253         return 0;
1254 }
1255
1256 static int
1257 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1258                  struct sched_domain *sd, enum cpu_idle_type idle)
1259 {
1260         /* don't touch RT tasks */
1261         return 0;
1262 }
1263
1264 static void set_cpus_allowed_rt(struct task_struct *p,
1265                                 const cpumask_t *new_mask)
1266 {
1267         int weight = cpus_weight(*new_mask);
1268
1269         BUG_ON(!rt_task(p));
1270
1271         /*
1272          * Update the migration status of the RQ if we have an RT task
1273          * which is running AND changing its weight value.
1274          */
1275         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1276                 struct rq *rq = task_rq(p);
1277
1278                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1279                         rq->rt.rt_nr_migratory++;
1280                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1281                         BUG_ON(!rq->rt.rt_nr_migratory);
1282                         rq->rt.rt_nr_migratory--;
1283                 }
1284
1285                 update_rt_migration(rq);
1286         }
1287
1288         p->cpus_allowed    = *new_mask;
1289         p->rt.nr_cpus_allowed = weight;
1290 }
1291
1292 /* Assumes rq->lock is held */
1293 static void rq_online_rt(struct rq *rq)
1294 {
1295         if (rq->rt.overloaded)
1296                 rt_set_overload(rq);
1297
1298         __enable_runtime(rq);
1299
1300         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1301 }
1302
1303 /* Assumes rq->lock is held */
1304 static void rq_offline_rt(struct rq *rq)
1305 {
1306         if (rq->rt.overloaded)
1307                 rt_clear_overload(rq);
1308
1309         __disable_runtime(rq);
1310
1311         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1312 }
1313
1314 /*
1315  * When switch from the rt queue, we bring ourselves to a position
1316  * that we might want to pull RT tasks from other runqueues.
1317  */
1318 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1319                            int running)
1320 {
1321         /*
1322          * If there are other RT tasks then we will reschedule
1323          * and the scheduling of the other RT tasks will handle
1324          * the balancing. But if we are the last RT task
1325          * we may need to handle the pulling of RT tasks
1326          * now.
1327          */
1328         if (!rq->rt.rt_nr_running)
1329                 pull_rt_task(rq);
1330 }
1331 #endif /* CONFIG_SMP */
1332
1333 /*
1334  * When switching a task to RT, we may overload the runqueue
1335  * with RT tasks. In this case we try to push them off to
1336  * other runqueues.
1337  */
1338 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1339                            int running)
1340 {
1341         int check_resched = 1;
1342
1343         /*
1344          * If we are already running, then there's nothing
1345          * that needs to be done. But if we are not running
1346          * we may need to preempt the current running task.
1347          * If that current running task is also an RT task
1348          * then see if we can move to another run queue.
1349          */
1350         if (!running) {
1351 #ifdef CONFIG_SMP
1352                 if (rq->rt.overloaded && push_rt_task(rq) &&
1353                     /* Don't resched if we changed runqueues */
1354                     rq != task_rq(p))
1355                         check_resched = 0;
1356 #endif /* CONFIG_SMP */
1357                 if (check_resched && p->prio < rq->curr->prio)
1358                         resched_task(rq->curr);
1359         }
1360 }
1361
1362 /*
1363  * Priority of the task has changed. This may cause
1364  * us to initiate a push or pull.
1365  */
1366 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1367                             int oldprio, int running)
1368 {
1369         if (running) {
1370 #ifdef CONFIG_SMP
1371                 /*
1372                  * If our priority decreases while running, we
1373                  * may need to pull tasks to this runqueue.
1374                  */
1375                 if (oldprio < p->prio)
1376                         pull_rt_task(rq);
1377                 /*
1378                  * If there's a higher priority task waiting to run
1379                  * then reschedule. Note, the above pull_rt_task
1380                  * can release the rq lock and p could migrate.
1381                  * Only reschedule if p is still on the same runqueue.
1382                  */
1383                 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1384                         resched_task(p);
1385 #else
1386                 /* For UP simply resched on drop of prio */
1387                 if (oldprio < p->prio)
1388                         resched_task(p);
1389 #endif /* CONFIG_SMP */
1390         } else {
1391                 /*
1392                  * This task is not running, but if it is
1393                  * greater than the current running task
1394                  * then reschedule.
1395                  */
1396                 if (p->prio < rq->curr->prio)
1397                         resched_task(rq->curr);
1398         }
1399 }
1400
1401 static void watchdog(struct rq *rq, struct task_struct *p)
1402 {
1403         unsigned long soft, hard;
1404
1405         if (!p->signal)
1406                 return;
1407
1408         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1409         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1410
1411         if (soft != RLIM_INFINITY) {
1412                 unsigned long next;
1413
1414                 p->rt.timeout++;
1415                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1416                 if (p->rt.timeout > next)
1417                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1418         }
1419 }
1420
1421 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1422 {
1423         update_curr_rt(rq);
1424
1425         watchdog(rq, p);
1426
1427         /*
1428          * RR tasks need a special form of timeslice management.
1429          * FIFO tasks have no timeslices.
1430          */
1431         if (p->policy != SCHED_RR)
1432                 return;
1433
1434         if (--p->rt.time_slice)
1435                 return;
1436
1437         p->rt.time_slice = DEF_TIMESLICE;
1438
1439         /*
1440          * Requeue to the end of queue if we are not the only element
1441          * on the queue:
1442          */
1443         if (p->rt.run_list.prev != p->rt.run_list.next) {
1444                 requeue_task_rt(rq, p, 0);
1445                 set_tsk_need_resched(p);
1446         }
1447 }
1448
1449 static void set_curr_task_rt(struct rq *rq)
1450 {
1451         struct task_struct *p = rq->curr;
1452
1453         p->se.exec_start = rq->clock;
1454 }
1455
1456 static const struct sched_class rt_sched_class = {
1457         .next                   = &fair_sched_class,
1458         .enqueue_task           = enqueue_task_rt,
1459         .dequeue_task           = dequeue_task_rt,
1460         .yield_task             = yield_task_rt,
1461 #ifdef CONFIG_SMP
1462         .select_task_rq         = select_task_rq_rt,
1463 #endif /* CONFIG_SMP */
1464
1465         .check_preempt_curr     = check_preempt_curr_rt,
1466
1467         .pick_next_task         = pick_next_task_rt,
1468         .put_prev_task          = put_prev_task_rt,
1469
1470 #ifdef CONFIG_SMP
1471         .load_balance           = load_balance_rt,
1472         .move_one_task          = move_one_task_rt,
1473         .set_cpus_allowed       = set_cpus_allowed_rt,
1474         .rq_online              = rq_online_rt,
1475         .rq_offline             = rq_offline_rt,
1476         .pre_schedule           = pre_schedule_rt,
1477         .post_schedule          = post_schedule_rt,
1478         .task_wake_up           = task_wake_up_rt,
1479         .switched_from          = switched_from_rt,
1480 #endif
1481
1482         .set_curr_task          = set_curr_task_rt,
1483         .task_tick              = task_tick_rt,
1484
1485         .prio_changed           = prio_changed_rt,
1486         .switched_to            = switched_to_rt,
1487 };
1488
1489 #ifdef CONFIG_SCHED_DEBUG
1490 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1491
1492 static void print_rt_stats(struct seq_file *m, int cpu)
1493 {
1494         struct rt_rq *rt_rq;
1495
1496         rcu_read_lock();
1497         for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1498                 print_rt_rq(m, cpu, rt_rq);
1499         rcu_read_unlock();
1500 }
1501 #endif /* CONFIG_SCHED_DEBUG */