git.oblomov.eu Git - linux-2.6/blob - 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         curr->se.exec_start = rq->clock;
 487         cpuacct_charge(curr, delta_exec);
 488
 489         for_each_sched_rt_entity(rt_se) {
 490                 rt_rq = rt_rq_of_se(rt_se);
 491
 492                 spin_lock(&rt_rq->rt_runtime_lock);
 493                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 494                         rt_rq->rt_time += delta_exec;
 495                         if (sched_rt_runtime_exceeded(rt_rq))
 496                                 resched_task(curr);
 497                 }
 498                 spin_unlock(&rt_rq->rt_runtime_lock);
 499         }
 500 }
 501
 502 static inline
 503 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 504 {
 505         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 506         rt_rq->rt_nr_running++;
 507 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 508         if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
 509 #ifdef CONFIG_SMP
 510                 struct rq *rq = rq_of_rt_rq(rt_rq);
 511 #endif
 512
 513                 rt_rq->highest_prio = rt_se_prio(rt_se);
 514 #ifdef CONFIG_SMP
 515                 if (rq->online)
 516                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 517                                    rt_se_prio(rt_se));
 518 #endif
 519         }
 520 #endif
 521 #ifdef CONFIG_SMP
 522         if (rt_se->nr_cpus_allowed > 1) {
 523                 struct rq *rq = rq_of_rt_rq(rt_rq);
 524
 525                 rq->rt.rt_nr_migratory++;
 526         }
 527
 528         update_rt_migration(rq_of_rt_rq(rt_rq));
 529 #endif
 530 #ifdef CONFIG_RT_GROUP_SCHED
 531         if (rt_se_boosted(rt_se))
 532                 rt_rq->rt_nr_boosted++;
 533
 534         if (rt_rq->tg)
 535                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 536 #else
 537         start_rt_bandwidth(&def_rt_bandwidth);
 538 #endif
 539 }
 540
 541 static inline
 542 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 543 {
 544 #ifdef CONFIG_SMP
 545         int highest_prio = rt_rq->highest_prio;
 546 #endif
 547
 548         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 549         WARN_ON(!rt_rq->rt_nr_running);
 550         rt_rq->rt_nr_running--;
 551 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 552         if (rt_rq->rt_nr_running) {
 553                 struct rt_prio_array *array;
 554
 555                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
 556                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
 557                         /* recalculate */
 558                         array = &rt_rq->active;
 559                         rt_rq->highest_prio =
 560                                 sched_find_first_bit(array->bitmap);
 561                 } /* otherwise leave rq->highest prio alone */
 562         } else
 563                 rt_rq->highest_prio = MAX_RT_PRIO;
 564 #endif
 565 #ifdef CONFIG_SMP
 566         if (rt_se->nr_cpus_allowed > 1) {
 567                 struct rq *rq = rq_of_rt_rq(rt_rq);
 568                 rq->rt.rt_nr_migratory--;
 569         }
 570
 571         if (rt_rq->highest_prio != highest_prio) {
 572                 struct rq *rq = rq_of_rt_rq(rt_rq);
 573
 574                 if (rq->online)
 575                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 576                                    rt_rq->highest_prio);
 577         }
 578
 579         update_rt_migration(rq_of_rt_rq(rt_rq));
 580 #endif /* CONFIG_SMP */
 581 #ifdef CONFIG_RT_GROUP_SCHED
 582         if (rt_se_boosted(rt_se))
 583                 rt_rq->rt_nr_boosted--;
 584
 585         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 586 #endif
 587 }
 588
 589 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
 590 {
 591         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 592         struct rt_prio_array *array = &rt_rq->active;
 593         struct rt_rq *group_rq = group_rt_rq(rt_se);
 594         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 595
 596         /*
 597          * Don't enqueue the group if its throttled, or when empty.
 598          * The latter is a consequence of the former when a child group
 599          * get throttled and the current group doesn't have any other
 600          * active members.
 601          */
 602         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 603                 return;
 604
 605         list_add_tail(&rt_se->run_list, queue);
 606         __set_bit(rt_se_prio(rt_se), array->bitmap);
 607
 608         inc_rt_tasks(rt_se, rt_rq);
 609 }
 610
 611 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 612 {
 613         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 614         struct rt_prio_array *array = &rt_rq->active;
 615
 616         list_del_init(&rt_se->run_list);
 617         if (list_empty(array->queue + rt_se_prio(rt_se)))
 618                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 619
 620         dec_rt_tasks(rt_se, rt_rq);
 621 }
 622
 623 /*
 624  * Because the prio of an upper entry depends on the lower
 625  * entries, we must remove entries top - down.
 626  */
 627 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 628 {
 629         struct sched_rt_entity *back = NULL;
 630
 631         for_each_sched_rt_entity(rt_se) {
 632                 rt_se->back = back;
 633                 back = rt_se;
 634         }
 635
 636         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 637                 if (on_rt_rq(rt_se))
 638                         __dequeue_rt_entity(rt_se);
 639         }
 640 }
 641
 642 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 643 {
 644         dequeue_rt_stack(rt_se);
 645         for_each_sched_rt_entity(rt_se)
 646                 __enqueue_rt_entity(rt_se);
 647 }
 648
 649 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 650 {
 651         dequeue_rt_stack(rt_se);
 652
 653         for_each_sched_rt_entity(rt_se) {
 654                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 655
 656                 if (rt_rq && rt_rq->rt_nr_running)
 657                         __enqueue_rt_entity(rt_se);
 658         }
 659 }
 660
 661 /*
 662  * Adding/removing a task to/from a priority array:
 663  */
 664 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 665 {
 666         struct sched_rt_entity *rt_se = &p->rt;
 667
 668         if (wakeup)
 669                 rt_se->timeout = 0;
 670
 671         enqueue_rt_entity(rt_se);
 672
 673         inc_cpu_load(rq, p->se.load.weight);
 674 }
 675
 676 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 677 {
 678         struct sched_rt_entity *rt_se = &p->rt;
 679
 680         update_curr_rt(rq);
 681         dequeue_rt_entity(rt_se);
 682
 683         dec_cpu_load(rq, p->se.load.weight);
 684 }
 685
 686 /*
 687  * Put task to the end of the run list without the overhead of dequeue
 688  * followed by enqueue.
 689  */
 690 static void
 691 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 692 {
 693         if (on_rt_rq(rt_se)) {
 694                 struct rt_prio_array *array = &rt_rq->active;
 695                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 696
 697                 if (head)
 698                         list_move(&rt_se->run_list, queue);
 699                 else
 700                         list_move_tail(&rt_se->run_list, queue);
 701         }
 702 }
 703
 704 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 705 {
 706         struct sched_rt_entity *rt_se = &p->rt;
 707         struct rt_rq *rt_rq;
 708
 709         for_each_sched_rt_entity(rt_se) {
 710                 rt_rq = rt_rq_of_se(rt_se);
 711                 requeue_rt_entity(rt_rq, rt_se, head);
 712         }
 713 }
 714
 715 static void yield_task_rt(struct rq *rq)
 716 {
 717         requeue_task_rt(rq, rq->curr, 0);
 718 }
 719
 720 #ifdef CONFIG_SMP
 721 static int find_lowest_rq(struct task_struct *task);
 722
 723 static int select_task_rq_rt(struct task_struct *p, int sync)
 724 {
 725         struct rq *rq = task_rq(p);
 726
 727         /*
 728          * If the current task is an RT task, then
 729          * try to see if we can wake this RT task up on another
 730          * runqueue. Otherwise simply start this RT task
 731          * on its current runqueue.
 732          *
 733          * We want to avoid overloading runqueues. Even if
 734          * the RT task is of higher priority than the current RT task.
 735          * RT tasks behave differently than other tasks. If
 736          * one gets preempted, we try to push it off to another queue.
 737          * So trying to keep a preempting RT task on the same
 738          * cache hot CPU will force the running RT task to
 739          * a cold CPU. So we waste all the cache for the lower
 740          * RT task in hopes of saving some of a RT task
 741          * that is just being woken and probably will have
 742          * cold cache anyway.
 743          */
 744         if (unlikely(rt_task(rq->curr)) &&
 745             (p->rt.nr_cpus_allowed > 1)) {
 746                 int cpu = find_lowest_rq(p);
 747
 748                 return (cpu == -1) ? task_cpu(p) : cpu;
 749         }
 750
 751         /*
 752          * Otherwise, just let it ride on the affined RQ and the
 753          * post-schedule router will push the preempted task away
 754          */
 755         return task_cpu(p);
 756 }
 757
 758 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
 759 {
 760         cpumask_t mask;
 761
 762         if (rq->curr->rt.nr_cpus_allowed == 1)
 763                 return;
 764
 765         if (p->rt.nr_cpus_allowed != 1
 766             && cpupri_find(&rq->rd->cpupri, p, &mask))
 767                 return;
 768
 769         if (!cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
 770                 return;
 771
 772         /*
 773          * There appears to be other cpus that can accept
 774          * current and none to run 'p', so lets reschedule
 775          * to try and push current away:
 776          */
 777         requeue_task_rt(rq, p, 1);
 778         resched_task(rq->curr);
 779 }
 780
 781 #endif /* CONFIG_SMP */
 782
 783 /*
 784  * Preempt the current task with a newly woken task if needed:
 785  */
 786 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
 787 {
 788         if (p->prio < rq->curr->prio) {
 789                 resched_task(rq->curr);
 790                 return;
 791         }
 792
 793 #ifdef CONFIG_SMP
 794         /*
 795          * If:
 796          *
 797          * - the newly woken task is of equal priority to the current task
 798          * - the newly woken task is non-migratable while current is migratable
 799          * - current will be preempted on the next reschedule
 800          *
 801          * we should check to see if current can readily move to a different
 802          * cpu.  If so, we will reschedule to allow the push logic to try
 803          * to move current somewhere else, making room for our non-migratable
 804          * task.
 805          */
 806         if (p->prio == rq->curr->prio && !need_resched())
 807                 check_preempt_equal_prio(rq, p);
 808 #endif
 809 }
 810
 811 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 812                                                    struct rt_rq *rt_rq)
 813 {
 814         struct rt_prio_array *array = &rt_rq->active;
 815         struct sched_rt_entity *next = NULL;
 816         struct list_head *queue;
 817         int idx;
 818
 819         idx = sched_find_first_bit(array->bitmap);
 820         BUG_ON(idx >= MAX_RT_PRIO);
 821
 822         queue = array->queue + idx;
 823         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 824
 825         return next;
 826 }
 827
 828 static struct task_struct *pick_next_task_rt(struct rq *rq)
 829 {
 830         struct sched_rt_entity *rt_se;
 831         struct task_struct *p;
 832         struct rt_rq *rt_rq;
 833
 834         rt_rq = &rq->rt;
 835
 836         if (unlikely(!rt_rq->rt_nr_running))
 837                 return NULL;
 838
 839         if (rt_rq_throttled(rt_rq))
 840                 return NULL;
 841
 842         do {
 843                 rt_se = pick_next_rt_entity(rq, rt_rq);
 844                 BUG_ON(!rt_se);
 845                 rt_rq = group_rt_rq(rt_se);
 846         } while (rt_rq);
 847
 848         p = rt_task_of(rt_se);
 849         p->se.exec_start = rq->clock;
 850         return p;
 851 }
 852
 853 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 854 {
 855         update_curr_rt(rq);
 856         p->se.exec_start = 0;
 857 }
 858
 859 #ifdef CONFIG_SMP
 860
 861 /* Only try algorithms three times */
 862 #define RT_MAX_TRIES 3
 863
 864 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
 865 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest);
 866
 867 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 868
 869 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 870 {
 871         if (!task_running(rq, p) &&
 872             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
 873             (p->rt.nr_cpus_allowed > 1))
 874                 return 1;
 875         return 0;
 876 }
 877
 878 /* Return the second highest RT task, NULL otherwise */
 879 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
 880 {
 881         struct task_struct *next = NULL;
 882         struct sched_rt_entity *rt_se;
 883         struct rt_prio_array *array;
 884         struct rt_rq *rt_rq;
 885         int idx;
 886
 887         for_each_leaf_rt_rq(rt_rq, rq) {
 888                 array = &rt_rq->active;
 889                 idx = sched_find_first_bit(array->bitmap);
 890  next_idx:
 891                 if (idx >= MAX_RT_PRIO)
 892                         continue;
 893                 if (next && next->prio < idx)
 894                         continue;
 895                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
 896                         struct task_struct *p = rt_task_of(rt_se);
 897                         if (pick_rt_task(rq, p, cpu)) {
 898                                 next = p;
 899                                 break;
 900                         }
 901                 }
 902                 if (!next) {
 903                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 904                         goto next_idx;
 905                 }
 906         }
 907
 908         return next;
 909 }
 910
 911 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
 912
 913 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
 914 {
 915         int first;
 916
 917         /* "this_cpu" is cheaper to preempt than a remote processor */
 918         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
 919                 return this_cpu;
 920
 921         first = first_cpu(*mask);
 922         if (first != NR_CPUS)
 923                 return first;
 924
 925         return -1;
 926 }
 927
 928 static int find_lowest_rq(struct task_struct *task)
 929 {
 930         struct sched_domain *sd;
 931         cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
 932         int this_cpu = smp_processor_id();
 933         int cpu      = task_cpu(task);
 934
 935         if (task->rt.nr_cpus_allowed == 1)
 936                 return -1; /* No other targets possible */
 937
 938         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
 939                 return -1; /* No targets found */
 940
 941         /*
 942          * Only consider CPUs that are usable for migration.
 943          * I guess we might want to change cpupri_find() to ignore those
 944          * in the first place.
 945          */
 946         cpus_and(*lowest_mask, *lowest_mask, cpu_active_map);
 947
 948         /*
 949          * At this point we have built a mask of cpus representing the
 950          * lowest priority tasks in the system.  Now we want to elect
 951          * the best one based on our affinity and topology.
 952          *
 953          * We prioritize the last cpu that the task executed on since
 954          * it is most likely cache-hot in that location.
 955          */
 956         if (cpu_isset(cpu, *lowest_mask))
 957                 return cpu;
 958
 959         /*
 960          * Otherwise, we consult the sched_domains span maps to figure
 961          * out which cpu is logically closest to our hot cache data.
 962          */
 963         if (this_cpu == cpu)
 964                 this_cpu = -1; /* Skip this_cpu opt if the same */
 965
 966         for_each_domain(cpu, sd) {
 967                 if (sd->flags & SD_WAKE_AFFINE) {
 968                         cpumask_t domain_mask;
 969                         int       best_cpu;
 970
 971                         cpus_and(domain_mask, sd->span, *lowest_mask);
 972
 973                         best_cpu = pick_optimal_cpu(this_cpu,
 974                                                     &domain_mask);
 975                         if (best_cpu != -1)
 976                                 return best_cpu;
 977                 }
 978         }
 979
 980         /*
 981          * And finally, if there were no matches within the domains
 982          * just give the caller *something* to work with from the compatible
 983          * locations.
 984          */
 985         return pick_optimal_cpu(this_cpu, lowest_mask);
 986 }
 987
 988 /* Will lock the rq it finds */
 989 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
 990 {
 991         struct rq *lowest_rq = NULL;
 992         int tries;
 993         int cpu;
 994
 995         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
 996                 cpu = find_lowest_rq(task);
 997
 998                 if ((cpu == -1) || (cpu == rq->cpu))
 999                         break;
1000
1001                 lowest_rq = cpu_rq(cpu);
1002
1003                 /* if the prio of this runqueue changed, try again */
1004                 if (double_lock_balance(rq, lowest_rq)) {
1005                         /*
1006                          * We had to unlock the run queue. In
1007                          * the mean time, task could have
1008                          * migrated already or had its affinity changed.
1009                          * Also make sure that it wasn't scheduled on its rq.
1010                          */
1011                         if (unlikely(task_rq(task) != rq ||
1012                                      !cpu_isset(lowest_rq->cpu,
1013                                                 task->cpus_allowed) ||
1014                                      task_running(rq, task) ||
1015                                      !task->se.on_rq)) {
1016
1017                                 spin_unlock(&lowest_rq->lock);
1018                                 lowest_rq = NULL;
1019                                 break;
1020                         }
1021                 }
1022
1023                 /* If this rq is still suitable use it. */
1024                 if (lowest_rq->rt.highest_prio > task->prio)
1025                         break;
1026
1027                 /* try again */
1028                 double_unlock_balance(rq, lowest_rq);
1029                 lowest_rq = NULL;
1030         }
1031
1032         return lowest_rq;
1033 }
1034
1035 /*
1036  * If the current CPU has more than one RT task, see if the non
1037  * running task can migrate over to a CPU that is running a task
1038  * of lesser priority.
1039  */
1040 static int push_rt_task(struct rq *rq)
1041 {
1042         struct task_struct *next_task;
1043         struct rq *lowest_rq;
1044         int ret = 0;
1045         int paranoid = RT_MAX_TRIES;
1046
1047         if (!rq->rt.overloaded)
1048                 return 0;
1049
1050         next_task = pick_next_highest_task_rt(rq, -1);
1051         if (!next_task)
1052                 return 0;
1053
1054  retry:
1055         if (unlikely(next_task == rq->curr)) {
1056                 WARN_ON(1);
1057                 return 0;
1058         }
1059
1060         /*
1061          * It's possible that the next_task slipped in of
1062          * higher priority than current. If that's the case
1063          * just reschedule current.
1064          */
1065         if (unlikely(next_task->prio < rq->curr->prio)) {
1066                 resched_task(rq->curr);
1067                 return 0;
1068         }
1069
1070         /* We might release rq lock */
1071         get_task_struct(next_task);
1072
1073         /* find_lock_lowest_rq locks the rq if found */
1074         lowest_rq = find_lock_lowest_rq(next_task, rq);
1075         if (!lowest_rq) {
1076                 struct task_struct *task;
1077                 /*
1078                  * find lock_lowest_rq releases rq->lock
1079                  * so it is possible that next_task has changed.
1080                  * If it has, then try again.
1081                  */
1082                 task = pick_next_highest_task_rt(rq, -1);
1083                 if (unlikely(task != next_task) && task && paranoid--) {
1084                         put_task_struct(next_task);
1085                         next_task = task;
1086                         goto retry;
1087                 }
1088                 goto out;
1089         }
1090
1091         deactivate_task(rq, next_task, 0);
1092         set_task_cpu(next_task, lowest_rq->cpu);
1093         activate_task(lowest_rq, next_task, 0);
1094
1095         resched_task(lowest_rq->curr);
1096
1097         double_unlock_balance(rq, lowest_rq);
1098
1099         ret = 1;
1100 out:
1101         put_task_struct(next_task);
1102
1103         return ret;
1104 }
1105
1106 /*
1107  * TODO: Currently we just use the second highest prio task on
1108  *       the queue, and stop when it can't migrate (or there's
1109  *       no more RT tasks).  There may be a case where a lower
1110  *       priority RT task has a different affinity than the
1111  *       higher RT task. In this case the lower RT task could
1112  *       possibly be able to migrate where as the higher priority
1113  *       RT task could not.  We currently ignore this issue.
1114  *       Enhancements are welcome!
1115  */
1116 static void push_rt_tasks(struct rq *rq)
1117 {
1118         /* push_rt_task will return true if it moved an RT */
1119         while (push_rt_task(rq))
1120                 ;
1121 }
1122
1123 static int pull_rt_task(struct rq *this_rq)
1124 {
1125         int this_cpu = this_rq->cpu, ret = 0, cpu;
1126         struct task_struct *p, *next;
1127         struct rq *src_rq;
1128
1129         if (likely(!rt_overloaded(this_rq)))
1130                 return 0;
1131
1132         next = pick_next_task_rt(this_rq);
1133
1134         for_each_cpu_mask_nr(cpu, this_rq->rd->rto_mask) {
1135                 if (this_cpu == cpu)
1136                         continue;
1137
1138                 src_rq = cpu_rq(cpu);
1139                 /*
1140                  * We can potentially drop this_rq's lock in
1141                  * double_lock_balance, and another CPU could
1142                  * steal our next task - hence we must cause
1143                  * the caller to recalculate the next task
1144                  * in that case:
1145                  */
1146                 if (double_lock_balance(this_rq, src_rq)) {
1147                         struct task_struct *old_next = next;
1148
1149                         next = pick_next_task_rt(this_rq);
1150                         if (next != old_next)
1151                                 ret = 1;
1152                 }
1153
1154                 /*
1155                  * Are there still pullable RT tasks?
1156                  */
1157                 if (src_rq->rt.rt_nr_running <= 1)
1158                         goto skip;
1159
1160                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1161
1162                 /*
1163                  * Do we have an RT task that preempts
1164                  * the to-be-scheduled task?
1165                  */
1166                 if (p && (!next || (p->prio < next->prio))) {
1167                         WARN_ON(p == src_rq->curr);
1168                         WARN_ON(!p->se.on_rq);
1169
1170                         /*
1171                          * There's a chance that p is higher in priority
1172                          * than what's currently running on its cpu.
1173                          * This is just that p is wakeing up and hasn't
1174                          * had a chance to schedule. We only pull
1175                          * p if it is lower in priority than the
1176                          * current task on the run queue or
1177                          * this_rq next task is lower in prio than
1178                          * the current task on that rq.
1179                          */
1180                         if (p->prio < src_rq->curr->prio ||
1181                             (next && next->prio < src_rq->curr->prio))
1182                                 goto skip;
1183
1184                         ret = 1;
1185
1186                         deactivate_task(src_rq, p, 0);
1187                         set_task_cpu(p, this_cpu);
1188                         activate_task(this_rq, p, 0);
1189                         /*
1190                          * We continue with the search, just in
1191                          * case there's an even higher prio task
1192                          * in another runqueue. (low likelyhood
1193                          * but possible)
1194                          *
1195                          * Update next so that we won't pick a task
1196                          * on another cpu with a priority lower (or equal)
1197                          * than the one we just picked.
1198                          */
1199                         next = p;
1200
1201                 }
1202  skip:
1203                 double_unlock_balance(this_rq, src_rq);
1204         }
1205
1206         return ret;
1207 }
1208
1209 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1210 {
1211         /* Try to pull RT tasks here if we lower this rq's prio */
1212         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1213                 pull_rt_task(rq);
1214 }
1215
1216 static void post_schedule_rt(struct rq *rq)
1217 {
1218         /*
1219          * If we have more than one rt_task queued, then
1220          * see if we can push the other rt_tasks off to other CPUS.
1221          * Note we may release the rq lock, and since
1222          * the lock was owned by prev, we need to release it
1223          * first via finish_lock_switch and then reaquire it here.
1224          */
1225         if (unlikely(rq->rt.overloaded)) {
1226                 spin_lock_irq(&rq->lock);
1227                 push_rt_tasks(rq);
1228                 spin_unlock_irq(&rq->lock);
1229         }
1230 }
1231
1232 /*
1233  * If we are not running and we are not going to reschedule soon, we should
1234  * try to push tasks away now
1235  */
1236 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1237 {
1238         if (!task_running(rq, p) &&
1239             !test_tsk_need_resched(rq->curr) &&
1240             rq->rt.overloaded)
1241                 push_rt_tasks(rq);
1242 }
1243
1244 static unsigned long
1245 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1246                 unsigned long max_load_move,
1247                 struct sched_domain *sd, enum cpu_idle_type idle,
1248                 int *all_pinned, int *this_best_prio)
1249 {
1250         /* don't touch RT tasks */
1251         return 0;
1252 }
1253
1254 static int
1255 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1256                  struct sched_domain *sd, enum cpu_idle_type idle)
1257 {
1258         /* don't touch RT tasks */
1259         return 0;
1260 }
1261
1262 static void set_cpus_allowed_rt(struct task_struct *p,
1263                                 const cpumask_t *new_mask)
1264 {
1265         int weight = cpus_weight(*new_mask);
1266
1267         BUG_ON(!rt_task(p));
1268
1269         /*
1270          * Update the migration status of the RQ if we have an RT task
1271          * which is running AND changing its weight value.
1272          */
1273         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1274                 struct rq *rq = task_rq(p);
1275
1276                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1277                         rq->rt.rt_nr_migratory++;
1278                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1279                         BUG_ON(!rq->rt.rt_nr_migratory);
1280                         rq->rt.rt_nr_migratory--;
1281                 }
1282
1283                 update_rt_migration(rq);
1284         }
1285
1286         p->cpus_allowed    = *new_mask;
1287         p->rt.nr_cpus_allowed = weight;
1288 }
1289
1290 /* Assumes rq->lock is held */
1291 static void rq_online_rt(struct rq *rq)
1292 {
1293         if (rq->rt.overloaded)
1294                 rt_set_overload(rq);
1295
1296         __enable_runtime(rq);
1297
1298         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1299 }
1300
1301 /* Assumes rq->lock is held */
1302 static void rq_offline_rt(struct rq *rq)
1303 {
1304         if (rq->rt.overloaded)
1305                 rt_clear_overload(rq);
1306
1307         __disable_runtime(rq);
1308
1309         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1310 }
1311
1312 /*
1313  * When switch from the rt queue, we bring ourselves to a position
1314  * that we might want to pull RT tasks from other runqueues.
1315  */
1316 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1317                            int running)
1318 {
1319         /*
1320          * If there are other RT tasks then we will reschedule
1321          * and the scheduling of the other RT tasks will handle
1322          * the balancing. But if we are the last RT task
1323          * we may need to handle the pulling of RT tasks
1324          * now.
1325          */
1326         if (!rq->rt.rt_nr_running)
1327                 pull_rt_task(rq);
1328 }
1329 #endif /* CONFIG_SMP */
1330
1331 /*
1332  * When switching a task to RT, we may overload the runqueue
1333  * with RT tasks. In this case we try to push them off to
1334  * other runqueues.
1335  */
1336 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1337                            int running)
1338 {
1339         int check_resched = 1;
1340
1341         /*
1342          * If we are already running, then there's nothing
1343          * that needs to be done. But if we are not running
1344          * we may need to preempt the current running task.
1345          * If that current running task is also an RT task
1346          * then see if we can move to another run queue.
1347          */
1348         if (!running) {
1349 #ifdef CONFIG_SMP
1350                 if (rq->rt.overloaded && push_rt_task(rq) &&
1351                     /* Don't resched if we changed runqueues */
1352                     rq != task_rq(p))
1353                         check_resched = 0;
1354 #endif /* CONFIG_SMP */
1355                 if (check_resched && p->prio < rq->curr->prio)
1356                         resched_task(rq->curr);
1357         }
1358 }
1359
1360 /*
1361  * Priority of the task has changed. This may cause
1362  * us to initiate a push or pull.
1363  */
1364 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1365                             int oldprio, int running)
1366 {
1367         if (running) {
1368 #ifdef CONFIG_SMP
1369                 /*
1370                  * If our priority decreases while running, we
1371                  * may need to pull tasks to this runqueue.
1372                  */
1373                 if (oldprio < p->prio)
1374                         pull_rt_task(rq);
1375                 /*
1376                  * If there's a higher priority task waiting to run
1377                  * then reschedule. Note, the above pull_rt_task
1378                  * can release the rq lock and p could migrate.
1379                  * Only reschedule if p is still on the same runqueue.
1380                  */
1381                 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1382                         resched_task(p);
1383 #else
1384                 /* For UP simply resched on drop of prio */
1385                 if (oldprio < p->prio)
1386                         resched_task(p);
1387 #endif /* CONFIG_SMP */
1388         } else {
1389                 /*
1390                  * This task is not running, but if it is
1391                  * greater than the current running task
1392                  * then reschedule.
1393                  */
1394                 if (p->prio < rq->curr->prio)
1395                         resched_task(rq->curr);
1396         }
1397 }
1398
1399 static void watchdog(struct rq *rq, struct task_struct *p)
1400 {
1401         unsigned long soft, hard;
1402
1403         if (!p->signal)
1404                 return;
1405
1406         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1407         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1408
1409         if (soft != RLIM_INFINITY) {
1410                 unsigned long next;
1411
1412                 p->rt.timeout++;
1413                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1414                 if (p->rt.timeout > next)
1415                         p->it_sched_expires = p->se.sum_exec_runtime;
1416         }
1417 }
1418
1419 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1420 {
1421         update_curr_rt(rq);
1422
1423         watchdog(rq, p);
1424
1425         /*
1426          * RR tasks need a special form of timeslice management.
1427          * FIFO tasks have no timeslices.
1428          */
1429         if (p->policy != SCHED_RR)
1430                 return;
1431
1432         if (--p->rt.time_slice)
1433                 return;
1434
1435         p->rt.time_slice = DEF_TIMESLICE;
1436
1437         /*
1438          * Requeue to the end of queue if we are not the only element
1439          * on the queue:
1440          */
1441         if (p->rt.run_list.prev != p->rt.run_list.next) {
1442                 requeue_task_rt(rq, p, 0);
1443                 set_tsk_need_resched(p);
1444         }
1445 }
1446
1447 static void set_curr_task_rt(struct rq *rq)
1448 {
1449         struct task_struct *p = rq->curr;
1450
1451         p->se.exec_start = rq->clock;
1452 }
1453
1454 static const struct sched_class rt_sched_class = {
1455         .next                   = &fair_sched_class,
1456         .enqueue_task           = enqueue_task_rt,
1457         .dequeue_task           = dequeue_task_rt,
1458         .yield_task             = yield_task_rt,
1459 #ifdef CONFIG_SMP
1460         .select_task_rq         = select_task_rq_rt,
1461 #endif /* CONFIG_SMP */
1462
1463         .check_preempt_curr     = check_preempt_curr_rt,
1464
1465         .pick_next_task         = pick_next_task_rt,
1466         .put_prev_task          = put_prev_task_rt,
1467
1468 #ifdef CONFIG_SMP
1469         .load_balance           = load_balance_rt,
1470         .move_one_task          = move_one_task_rt,
1471         .set_cpus_allowed       = set_cpus_allowed_rt,
1472         .rq_online              = rq_online_rt,
1473         .rq_offline             = rq_offline_rt,
1474         .pre_schedule           = pre_schedule_rt,
1475         .post_schedule          = post_schedule_rt,
1476         .task_wake_up           = task_wake_up_rt,
1477         .switched_from          = switched_from_rt,
1478 #endif
1479
1480         .set_curr_task          = set_curr_task_rt,
1481         .task_tick              = task_tick_rt,
1482
1483         .prio_changed           = prio_changed_rt,
1484         .switched_to            = switched_to_rt,
1485 };
1486
1487 #ifdef CONFIG_SCHED_DEBUG
1488 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1489
1490 static void print_rt_stats(struct seq_file *m, int cpu)
1491 {
1492         struct rt_rq *rt_rq;
1493
1494         rcu_read_lock();
1495         for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1496                 print_rt_rq(m, cpu, rt_rq);
1497         rcu_read_unlock();
1498 }
1499 #endif /* CONFIG_SCHED_DEBUG */