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