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         cpumask_set_cpu(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         cpumask_clear_cpu(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_rcu(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 task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 106         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 107
 108         if (rt_rq->rt_nr_running) {
 109                 if (rt_se && !on_rt_rq(rt_se))
 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 const struct cpumask *sched_rt_period_mask(void)
 143 {
 144         return cpu_rq(smp_processor_id())->rd->span;
 145 }
 146 #else
 147 static inline const struct cpumask *sched_rt_period_mask(void)
 148 {
 149         return cpu_online_mask;
 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 const struct cpumask *sched_rt_period_mask(void)
 216 {
 217         return cpu_online_mask;
 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 /*
 235  * We ran out of runtime, see if we can borrow some from our neighbours.
 236  */
 237 static int do_balance_runtime(struct rt_rq *rt_rq)
 238 {
 239         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 240         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 241         int i, weight, more = 0;
 242         u64 rt_period;
 243
 244         weight = cpumask_weight(rd->span);
 245
 246         spin_lock(&rt_b->rt_runtime_lock);
 247         rt_period = ktime_to_ns(rt_b->rt_period);
 248         for_each_cpu(i, rd->span) {
 249                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 250                 s64 diff;
 251
 252                 if (iter == rt_rq)
 253                         continue;
 254
 255                 spin_lock(&iter->rt_runtime_lock);
 256                 /*
 257                  * Either all rqs have inf runtime and there's nothing to steal
 258                  * or __disable_runtime() below sets a specific rq to inf to
 259                  * indicate its been disabled and disalow stealing.
 260                  */
 261                 if (iter->rt_runtime == RUNTIME_INF)
 262                         goto next;
 263
 264                 /*
 265                  * From runqueues with spare time, take 1/n part of their
 266                  * spare time, but no more than our period.
 267                  */
 268                 diff = iter->rt_runtime - iter->rt_time;
 269                 if (diff > 0) {
 270                         diff = div_u64((u64)diff, weight);
 271                         if (rt_rq->rt_runtime + diff > rt_period)
 272                                 diff = rt_period - rt_rq->rt_runtime;
 273                         iter->rt_runtime -= diff;
 274                         rt_rq->rt_runtime += diff;
 275                         more = 1;
 276                         if (rt_rq->rt_runtime == rt_period) {
 277                                 spin_unlock(&iter->rt_runtime_lock);
 278                                 break;
 279                         }
 280                 }
 281 next:
 282                 spin_unlock(&iter->rt_runtime_lock);
 283         }
 284         spin_unlock(&rt_b->rt_runtime_lock);
 285
 286         return more;
 287 }
 288
 289 /*
 290  * Ensure this RQ takes back all the runtime it lend to its neighbours.
 291  */
 292 static void __disable_runtime(struct rq *rq)
 293 {
 294         struct root_domain *rd = rq->rd;
 295         struct rt_rq *rt_rq;
 296
 297         if (unlikely(!scheduler_running))
 298                 return;
 299
 300         for_each_leaf_rt_rq(rt_rq, rq) {
 301                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 302                 s64 want;
 303                 int i;
 304
 305                 spin_lock(&rt_b->rt_runtime_lock);
 306                 spin_lock(&rt_rq->rt_runtime_lock);
 307                 /*
 308                  * Either we're all inf and nobody needs to borrow, or we're
 309                  * already disabled and thus have nothing to do, or we have
 310                  * exactly the right amount of runtime to take out.
 311                  */
 312                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 313                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 314                         goto balanced;
 315                 spin_unlock(&rt_rq->rt_runtime_lock);
 316
 317                 /*
 318                  * Calculate the difference between what we started out with
 319                  * and what we current have, that's the amount of runtime
 320                  * we lend and now have to reclaim.
 321                  */
 322                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 323
 324                 /*
 325                  * Greedy reclaim, take back as much as we can.
 326                  */
 327                 for_each_cpu(i, rd->span) {
 328                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 329                         s64 diff;
 330
 331                         /*
 332                          * Can't reclaim from ourselves or disabled runqueues.
 333                          */
 334                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 335                                 continue;
 336
 337                         spin_lock(&iter->rt_runtime_lock);
 338                         if (want > 0) {
 339                                 diff = min_t(s64, iter->rt_runtime, want);
 340                                 iter->rt_runtime -= diff;
 341                                 want -= diff;
 342                         } else {
 343                                 iter->rt_runtime -= want;
 344                                 want -= want;
 345                         }
 346                         spin_unlock(&iter->rt_runtime_lock);
 347
 348                         if (!want)
 349                                 break;
 350                 }
 351
 352                 spin_lock(&rt_rq->rt_runtime_lock);
 353                 /*
 354                  * We cannot be left wanting - that would mean some runtime
 355                  * leaked out of the system.
 356                  */
 357                 BUG_ON(want);
 358 balanced:
 359                 /*
 360                  * Disable all the borrow logic by pretending we have inf
 361                  * runtime - in which case borrowing doesn't make sense.
 362                  */
 363                 rt_rq->rt_runtime = RUNTIME_INF;
 364                 spin_unlock(&rt_rq->rt_runtime_lock);
 365                 spin_unlock(&rt_b->rt_runtime_lock);
 366         }
 367 }
 368
 369 static void disable_runtime(struct rq *rq)
 370 {
 371         unsigned long flags;
 372
 373         spin_lock_irqsave(&rq->lock, flags);
 374         __disable_runtime(rq);
 375         spin_unlock_irqrestore(&rq->lock, flags);
 376 }
 377
 378 static void __enable_runtime(struct rq *rq)
 379 {
 380         struct rt_rq *rt_rq;
 381
 382         if (unlikely(!scheduler_running))
 383                 return;
 384
 385         /*
 386          * Reset each runqueue's bandwidth settings
 387          */
 388         for_each_leaf_rt_rq(rt_rq, rq) {
 389                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 390
 391                 spin_lock(&rt_b->rt_runtime_lock);
 392                 spin_lock(&rt_rq->rt_runtime_lock);
 393                 rt_rq->rt_runtime = rt_b->rt_runtime;
 394                 rt_rq->rt_time = 0;
 395                 rt_rq->rt_throttled = 0;
 396                 spin_unlock(&rt_rq->rt_runtime_lock);
 397                 spin_unlock(&rt_b->rt_runtime_lock);
 398         }
 399 }
 400
 401 static void enable_runtime(struct rq *rq)
 402 {
 403         unsigned long flags;
 404
 405         spin_lock_irqsave(&rq->lock, flags);
 406         __enable_runtime(rq);
 407         spin_unlock_irqrestore(&rq->lock, flags);
 408 }
 409
 410 static int balance_runtime(struct rt_rq *rt_rq)
 411 {
 412         int more = 0;
 413
 414         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 415                 spin_unlock(&rt_rq->rt_runtime_lock);
 416                 more = do_balance_runtime(rt_rq);
 417                 spin_lock(&rt_rq->rt_runtime_lock);
 418         }
 419
 420         return more;
 421 }
 422 #else /* !CONFIG_SMP */
 423 static inline int balance_runtime(struct rt_rq *rt_rq)
 424 {
 425         return 0;
 426 }
 427 #endif /* CONFIG_SMP */
 428
 429 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 430 {
 431         int i, idle = 1;
 432         const struct cpumask *span;
 433
 434         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 435                 return 1;
 436
 437         span = sched_rt_period_mask();
 438         for_each_cpu(i, span) {
 439                 int enqueue = 0;
 440                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 441                 struct rq *rq = rq_of_rt_rq(rt_rq);
 442
 443                 spin_lock(&rq->lock);
 444                 if (rt_rq->rt_time) {
 445                         u64 runtime;
 446
 447                         spin_lock(&rt_rq->rt_runtime_lock);
 448                         if (rt_rq->rt_throttled)
 449                                 balance_runtime(rt_rq);
 450                         runtime = rt_rq->rt_runtime;
 451                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 452                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 453                                 rt_rq->rt_throttled = 0;
 454                                 enqueue = 1;
 455                         }
 456                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 457                                 idle = 0;
 458                         spin_unlock(&rt_rq->rt_runtime_lock);
 459                 } else if (rt_rq->rt_nr_running)
 460                         idle = 0;
 461
 462                 if (enqueue)
 463                         sched_rt_rq_enqueue(rt_rq);
 464                 spin_unlock(&rq->lock);
 465         }
 466
 467         return idle;
 468 }
 469
 470 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 471 {
 472 #ifdef CONFIG_RT_GROUP_SCHED
 473         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 474
 475         if (rt_rq)
 476                 return rt_rq->highest_prio;
 477 #endif
 478
 479         return rt_task_of(rt_se)->prio;
 480 }
 481
 482 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 483 {
 484         u64 runtime = sched_rt_runtime(rt_rq);
 485
 486         if (rt_rq->rt_throttled)
 487                 return rt_rq_throttled(rt_rq);
 488
 489         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 490                 return 0;
 491
 492         balance_runtime(rt_rq);
 493         runtime = sched_rt_runtime(rt_rq);
 494         if (runtime == RUNTIME_INF)
 495                 return 0;
 496
 497         if (rt_rq->rt_time > runtime) {
 498                 rt_rq->rt_throttled = 1;
 499                 if (rt_rq_throttled(rt_rq)) {
 500                         sched_rt_rq_dequeue(rt_rq);
 501                         return 1;
 502                 }
 503         }
 504
 505         return 0;
 506 }
 507
 508 /*
 509  * Update the current task's runtime statistics. Skip current tasks that
 510  * are not in our scheduling class.
 511  */
 512 static void update_curr_rt(struct rq *rq)
 513 {
 514         struct task_struct *curr = rq->curr;
 515         struct sched_rt_entity *rt_se = &curr->rt;
 516         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 517         u64 delta_exec;
 518
 519         if (!task_has_rt_policy(curr))
 520                 return;
 521
 522         delta_exec = rq->clock - curr->se.exec_start;
 523         if (unlikely((s64)delta_exec < 0))
 524                 delta_exec = 0;
 525
 526         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
 527
 528         curr->se.sum_exec_runtime += delta_exec;
 529         account_group_exec_runtime(curr, delta_exec);
 530
 531         curr->se.exec_start = rq->clock;
 532         cpuacct_charge(curr, delta_exec);
 533
 534         if (!rt_bandwidth_enabled())
 535                 return;
 536
 537         for_each_sched_rt_entity(rt_se) {
 538                 rt_rq = rt_rq_of_se(rt_se);
 539
 540                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 541                         spin_lock(&rt_rq->rt_runtime_lock);
 542                         rt_rq->rt_time += delta_exec;
 543                         if (sched_rt_runtime_exceeded(rt_rq))
 544                                 resched_task(curr);
 545                         spin_unlock(&rt_rq->rt_runtime_lock);
 546                 }
 547         }
 548 }
 549
 550 static inline
 551 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 552 {
 553         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 554         rt_rq->rt_nr_running++;
 555 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 556         if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
 557 #ifdef CONFIG_SMP
 558                 struct rq *rq = rq_of_rt_rq(rt_rq);
 559 #endif
 560
 561                 rt_rq->highest_prio = rt_se_prio(rt_se);
 562 #ifdef CONFIG_SMP
 563                 if (rq->online)
 564                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 565                                    rt_se_prio(rt_se));
 566 #endif
 567         }
 568 #endif
 569 #ifdef CONFIG_SMP
 570         if (rt_se->nr_cpus_allowed > 1) {
 571                 struct rq *rq = rq_of_rt_rq(rt_rq);
 572
 573                 rq->rt.rt_nr_migratory++;
 574         }
 575
 576         update_rt_migration(rq_of_rt_rq(rt_rq));
 577 #endif
 578 #ifdef CONFIG_RT_GROUP_SCHED
 579         if (rt_se_boosted(rt_se))
 580                 rt_rq->rt_nr_boosted++;
 581
 582         if (rt_rq->tg)
 583                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 584 #else
 585         start_rt_bandwidth(&def_rt_bandwidth);
 586 #endif
 587 }
 588
 589 static inline
 590 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 591 {
 592 #ifdef CONFIG_SMP
 593         int highest_prio = rt_rq->highest_prio;
 594 #endif
 595
 596         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 597         WARN_ON(!rt_rq->rt_nr_running);
 598         rt_rq->rt_nr_running--;
 599 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 600         if (rt_rq->rt_nr_running) {
 601                 struct rt_prio_array *array;
 602
 603                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
 604                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
 605                         /* recalculate */
 606                         array = &rt_rq->active;
 607                         rt_rq->highest_prio =
 608                                 sched_find_first_bit(array->bitmap);
 609                 } /* otherwise leave rq->highest prio alone */
 610         } else
 611                 rt_rq->highest_prio = MAX_RT_PRIO;
 612 #endif
 613 #ifdef CONFIG_SMP
 614         if (rt_se->nr_cpus_allowed > 1) {
 615                 struct rq *rq = rq_of_rt_rq(rt_rq);
 616                 rq->rt.rt_nr_migratory--;
 617         }
 618
 619         if (rt_rq->highest_prio != highest_prio) {
 620                 struct rq *rq = rq_of_rt_rq(rt_rq);
 621
 622                 if (rq->online)
 623                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 624                                    rt_rq->highest_prio);
 625         }
 626
 627         update_rt_migration(rq_of_rt_rq(rt_rq));
 628 #endif /* CONFIG_SMP */
 629 #ifdef CONFIG_RT_GROUP_SCHED
 630         if (rt_se_boosted(rt_se))
 631                 rt_rq->rt_nr_boosted--;
 632
 633         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 634 #endif
 635 }
 636
 637 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
 638 {
 639         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 640         struct rt_prio_array *array = &rt_rq->active;
 641         struct rt_rq *group_rq = group_rt_rq(rt_se);
 642         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 643
 644         /*
 645          * Don't enqueue the group if its throttled, or when empty.
 646          * The latter is a consequence of the former when a child group
 647          * get throttled and the current group doesn't have any other
 648          * active members.
 649          */
 650         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 651                 return;
 652
 653         list_add_tail(&rt_se->run_list, queue);
 654         __set_bit(rt_se_prio(rt_se), array->bitmap);
 655
 656         inc_rt_tasks(rt_se, rt_rq);
 657 }
 658
 659 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 660 {
 661         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 662         struct rt_prio_array *array = &rt_rq->active;
 663
 664         list_del_init(&rt_se->run_list);
 665         if (list_empty(array->queue + rt_se_prio(rt_se)))
 666                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 667
 668         dec_rt_tasks(rt_se, rt_rq);
 669 }
 670
 671 /*
 672  * Because the prio of an upper entry depends on the lower
 673  * entries, we must remove entries top - down.
 674  */
 675 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 676 {
 677         struct sched_rt_entity *back = NULL;
 678
 679         for_each_sched_rt_entity(rt_se) {
 680                 rt_se->back = back;
 681                 back = rt_se;
 682         }
 683
 684         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 685                 if (on_rt_rq(rt_se))
 686                         __dequeue_rt_entity(rt_se);
 687         }
 688 }
 689
 690 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 691 {
 692         dequeue_rt_stack(rt_se);
 693         for_each_sched_rt_entity(rt_se)
 694                 __enqueue_rt_entity(rt_se);
 695 }
 696
 697 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 698 {
 699         dequeue_rt_stack(rt_se);
 700
 701         for_each_sched_rt_entity(rt_se) {
 702                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 703
 704                 if (rt_rq && rt_rq->rt_nr_running)
 705                         __enqueue_rt_entity(rt_se);
 706         }
 707 }
 708
 709 /*
 710  * Adding/removing a task to/from a priority array:
 711  */
 712 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 713 {
 714         struct sched_rt_entity *rt_se = &p->rt;
 715
 716         if (wakeup)
 717                 rt_se->timeout = 0;
 718
 719         enqueue_rt_entity(rt_se);
 720
 721         inc_cpu_load(rq, p->se.load.weight);
 722 }
 723
 724 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 725 {
 726         struct sched_rt_entity *rt_se = &p->rt;
 727
 728         update_curr_rt(rq);
 729         dequeue_rt_entity(rt_se);
 730
 731         dec_cpu_load(rq, p->se.load.weight);
 732 }
 733
 734 /*
 735  * Put task to the end of the run list without the overhead of dequeue
 736  * followed by enqueue.
 737  */
 738 static void
 739 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 740 {
 741         if (on_rt_rq(rt_se)) {
 742                 struct rt_prio_array *array = &rt_rq->active;
 743                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 744
 745                 if (head)
 746                         list_move(&rt_se->run_list, queue);
 747                 else
 748                         list_move_tail(&rt_se->run_list, queue);
 749         }
 750 }
 751
 752 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 753 {
 754         struct sched_rt_entity *rt_se = &p->rt;
 755         struct rt_rq *rt_rq;
 756
 757         for_each_sched_rt_entity(rt_se) {
 758                 rt_rq = rt_rq_of_se(rt_se);
 759                 requeue_rt_entity(rt_rq, rt_se, head);
 760         }
 761 }
 762
 763 static void yield_task_rt(struct rq *rq)
 764 {
 765         requeue_task_rt(rq, rq->curr, 0);
 766 }
 767
 768 #ifdef CONFIG_SMP
 769 static int find_lowest_rq(struct task_struct *task);
 770
 771 static int select_task_rq_rt(struct task_struct *p, int sync)
 772 {
 773         struct rq *rq = task_rq(p);
 774
 775         /*
 776          * If the current task is an RT task, then
 777          * try to see if we can wake this RT task up on another
 778          * runqueue. Otherwise simply start this RT task
 779          * on its current runqueue.
 780          *
 781          * We want to avoid overloading runqueues. Even if
 782          * the RT task is of higher priority than the current RT task.
 783          * RT tasks behave differently than other tasks. If
 784          * one gets preempted, we try to push it off to another queue.
 785          * So trying to keep a preempting RT task on the same
 786          * cache hot CPU will force the running RT task to
 787          * a cold CPU. So we waste all the cache for the lower
 788          * RT task in hopes of saving some of a RT task
 789          * that is just being woken and probably will have
 790          * cold cache anyway.
 791          */
 792         if (unlikely(rt_task(rq->curr)) &&
 793             (p->rt.nr_cpus_allowed > 1)) {
 794                 int cpu = find_lowest_rq(p);
 795
 796                 return (cpu == -1) ? task_cpu(p) : cpu;
 797         }
 798
 799         /*
 800          * Otherwise, just let it ride on the affined RQ and the
 801          * post-schedule router will push the preempted task away
 802          */
 803         return task_cpu(p);
 804 }
 805
 806 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
 807 {
 808         cpumask_var_t mask;
 809
 810         if (rq->curr->rt.nr_cpus_allowed == 1)
 811                 return;
 812
 813         if (!alloc_cpumask_var(&mask, GFP_ATOMIC))
 814                 return;
 815
 816         if (p->rt.nr_cpus_allowed != 1
 817             && cpupri_find(&rq->rd->cpupri, p, mask))
 818                 goto free;
 819
 820         if (!cpupri_find(&rq->rd->cpupri, rq->curr, mask))
 821                 goto free;
 822
 823         /*
 824          * There appears to be other cpus that can accept
 825          * current and none to run 'p', so lets reschedule
 826          * to try and push current away:
 827          */
 828         requeue_task_rt(rq, p, 1);
 829         resched_task(rq->curr);
 830 free:
 831         free_cpumask_var(mask);
 832 }
 833
 834 #endif /* CONFIG_SMP */
 835
 836 /*
 837  * Preempt the current task with a newly woken task if needed:
 838  */
 839 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
 840 {
 841         if (p->prio < rq->curr->prio) {
 842                 resched_task(rq->curr);
 843                 return;
 844         }
 845
 846 #ifdef CONFIG_SMP
 847         /*
 848          * If:
 849          *
 850          * - the newly woken task is of equal priority to the current task
 851          * - the newly woken task is non-migratable while current is migratable
 852          * - current will be preempted on the next reschedule
 853          *
 854          * we should check to see if current can readily move to a different
 855          * cpu.  If so, we will reschedule to allow the push logic to try
 856          * to move current somewhere else, making room for our non-migratable
 857          * task.
 858          */
 859         if (p->prio == rq->curr->prio && !need_resched())
 860                 check_preempt_equal_prio(rq, p);
 861 #endif
 862 }
 863
 864 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 865                                                    struct rt_rq *rt_rq)
 866 {
 867         struct rt_prio_array *array = &rt_rq->active;
 868         struct sched_rt_entity *next = NULL;
 869         struct list_head *queue;
 870         int idx;
 871
 872         idx = sched_find_first_bit(array->bitmap);
 873         BUG_ON(idx >= MAX_RT_PRIO);
 874
 875         queue = array->queue + idx;
 876         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 877
 878         return next;
 879 }
 880
 881 static struct task_struct *pick_next_task_rt(struct rq *rq)
 882 {
 883         struct sched_rt_entity *rt_se;
 884         struct task_struct *p;
 885         struct rt_rq *rt_rq;
 886
 887         rt_rq = &rq->rt;
 888
 889         if (unlikely(!rt_rq->rt_nr_running))
 890                 return NULL;
 891
 892         if (rt_rq_throttled(rt_rq))
 893                 return NULL;
 894
 895         do {
 896                 rt_se = pick_next_rt_entity(rq, rt_rq);
 897                 BUG_ON(!rt_se);
 898                 rt_rq = group_rt_rq(rt_se);
 899         } while (rt_rq);
 900
 901         p = rt_task_of(rt_se);
 902         p->se.exec_start = rq->clock;
 903         return p;
 904 }
 905
 906 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 907 {
 908         update_curr_rt(rq);
 909         p->se.exec_start = 0;
 910 }
 911
 912 #ifdef CONFIG_SMP
 913
 914 /* Only try algorithms three times */
 915 #define RT_MAX_TRIES 3
 916
 917 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 918
 919 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 920 {
 921         if (!task_running(rq, p) &&
 922             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
 923             (p->rt.nr_cpus_allowed > 1))
 924                 return 1;
 925         return 0;
 926 }
 927
 928 /* Return the second highest RT task, NULL otherwise */
 929 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
 930 {
 931         struct task_struct *next = NULL;
 932         struct sched_rt_entity *rt_se;
 933         struct rt_prio_array *array;
 934         struct rt_rq *rt_rq;
 935         int idx;
 936
 937         for_each_leaf_rt_rq(rt_rq, rq) {
 938                 array = &rt_rq->active;
 939                 idx = sched_find_first_bit(array->bitmap);
 940  next_idx:
 941                 if (idx >= MAX_RT_PRIO)
 942                         continue;
 943                 if (next && next->prio < idx)
 944                         continue;
 945                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
 946                         struct task_struct *p = rt_task_of(rt_se);
 947                         if (pick_rt_task(rq, p, cpu)) {
 948                                 next = p;
 949                                 break;
 950                         }
 951                 }
 952                 if (!next) {
 953                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 954                         goto next_idx;
 955                 }
 956         }
 957
 958         return next;
 959 }
 960
 961 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
 962
 963 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
 964 {
 965         int first;
 966
 967         /* "this_cpu" is cheaper to preempt than a remote processor */
 968         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
 969                 return this_cpu;
 970
 971         first = first_cpu(*mask);
 972         if (first != NR_CPUS)
 973                 return first;
 974
 975         return -1;
 976 }
 977
 978 static int find_lowest_rq(struct task_struct *task)
 979 {
 980         struct sched_domain *sd;
 981         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
 982         int this_cpu = smp_processor_id();
 983         int cpu      = task_cpu(task);
 984
 985         if (task->rt.nr_cpus_allowed == 1)
 986                 return -1; /* No other targets possible */
 987
 988         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
 989                 return -1; /* No targets found */
 990
 991         /*
 992          * Only consider CPUs that are usable for migration.
 993          * I guess we might want to change cpupri_find() to ignore those
 994          * in the first place.
 995          */
 996         cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
 997
 998         /*
 999          * At this point we have built a mask of cpus representing the
1000          * lowest priority tasks in the system.  Now we want to elect
1001          * the best one based on our affinity and topology.
1002          *
1003          * We prioritize the last cpu that the task executed on since
1004          * it is most likely cache-hot in that location.
1005          */
1006         if (cpumask_test_cpu(cpu, lowest_mask))
1007                 return cpu;
1008
1009         /*
1010          * Otherwise, we consult the sched_domains span maps to figure
1011          * out which cpu is logically closest to our hot cache data.
1012          */
1013         if (this_cpu == cpu)
1014                 this_cpu = -1; /* Skip this_cpu opt if the same */
1015
1016         for_each_domain(cpu, sd) {
1017                 if (sd->flags & SD_WAKE_AFFINE) {
1018                         cpumask_t domain_mask;
1019                         int       best_cpu;
1020
1021                         cpumask_and(&domain_mask, sched_domain_span(sd),
1022                                     lowest_mask);
1023
1024                         best_cpu = pick_optimal_cpu(this_cpu,
1025                                                     &domain_mask);
1026                         if (best_cpu != -1)
1027                                 return best_cpu;
1028                 }
1029         }
1030
1031         /*
1032          * And finally, if there were no matches within the domains
1033          * just give the caller *something* to work with from the compatible
1034          * locations.
1035          */
1036         return pick_optimal_cpu(this_cpu, lowest_mask);
1037 }
1038
1039 /* Will lock the rq it finds */
1040 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1041 {
1042         struct rq *lowest_rq = NULL;
1043         int tries;
1044         int cpu;
1045
1046         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1047                 cpu = find_lowest_rq(task);
1048
1049                 if ((cpu == -1) || (cpu == rq->cpu))
1050                         break;
1051
1052                 lowest_rq = cpu_rq(cpu);
1053
1054                 /* if the prio of this runqueue changed, try again */
1055                 if (double_lock_balance(rq, lowest_rq)) {
1056                         /*
1057                          * We had to unlock the run queue. In
1058                          * the mean time, task could have
1059                          * migrated already or had its affinity changed.
1060                          * Also make sure that it wasn't scheduled on its rq.
1061                          */
1062                         if (unlikely(task_rq(task) != rq ||
1063                                      !cpumask_test_cpu(lowest_rq->cpu,
1064                                                        &task->cpus_allowed) ||
1065                                      task_running(rq, task) ||
1066                                      !task->se.on_rq)) {
1067
1068                                 spin_unlock(&lowest_rq->lock);
1069                                 lowest_rq = NULL;
1070                                 break;
1071                         }
1072                 }
1073
1074                 /* If this rq is still suitable use it. */
1075                 if (lowest_rq->rt.highest_prio > task->prio)
1076                         break;
1077
1078                 /* try again */
1079                 double_unlock_balance(rq, lowest_rq);
1080                 lowest_rq = NULL;
1081         }
1082
1083         return lowest_rq;
1084 }
1085
1086 /*
1087  * If the current CPU has more than one RT task, see if the non
1088  * running task can migrate over to a CPU that is running a task
1089  * of lesser priority.
1090  */
1091 static int push_rt_task(struct rq *rq)
1092 {
1093         struct task_struct *next_task;
1094         struct rq *lowest_rq;
1095         int ret = 0;
1096         int paranoid = RT_MAX_TRIES;
1097
1098         if (!rq->rt.overloaded)
1099                 return 0;
1100
1101         next_task = pick_next_highest_task_rt(rq, -1);
1102         if (!next_task)
1103                 return 0;
1104
1105  retry:
1106         if (unlikely(next_task == rq->curr)) {
1107                 WARN_ON(1);
1108                 return 0;
1109         }
1110
1111         /*
1112          * It's possible that the next_task slipped in of
1113          * higher priority than current. If that's the case
1114          * just reschedule current.
1115          */
1116         if (unlikely(next_task->prio < rq->curr->prio)) {
1117                 resched_task(rq->curr);
1118                 return 0;
1119         }
1120
1121         /* We might release rq lock */
1122         get_task_struct(next_task);
1123
1124         /* find_lock_lowest_rq locks the rq if found */
1125         lowest_rq = find_lock_lowest_rq(next_task, rq);
1126         if (!lowest_rq) {
1127                 struct task_struct *task;
1128                 /*
1129                  * find lock_lowest_rq releases rq->lock
1130                  * so it is possible that next_task has changed.
1131                  * If it has, then try again.
1132                  */
1133                 task = pick_next_highest_task_rt(rq, -1);
1134                 if (unlikely(task != next_task) && task && paranoid--) {
1135                         put_task_struct(next_task);
1136                         next_task = task;
1137                         goto retry;
1138                 }
1139                 goto out;
1140         }
1141
1142         deactivate_task(rq, next_task, 0);
1143         set_task_cpu(next_task, lowest_rq->cpu);
1144         activate_task(lowest_rq, next_task, 0);
1145
1146         resched_task(lowest_rq->curr);
1147
1148         double_unlock_balance(rq, lowest_rq);
1149
1150         ret = 1;
1151 out:
1152         put_task_struct(next_task);
1153
1154         return ret;
1155 }
1156
1157 /*
1158  * TODO: Currently we just use the second highest prio task on
1159  *       the queue, and stop when it can't migrate (or there's
1160  *       no more RT tasks).  There may be a case where a lower
1161  *       priority RT task has a different affinity than the
1162  *       higher RT task. In this case the lower RT task could
1163  *       possibly be able to migrate where as the higher priority
1164  *       RT task could not.  We currently ignore this issue.
1165  *       Enhancements are welcome!
1166  */
1167 static void push_rt_tasks(struct rq *rq)
1168 {
1169         /* push_rt_task will return true if it moved an RT */
1170         while (push_rt_task(rq))
1171                 ;
1172 }
1173
1174 static int pull_rt_task(struct rq *this_rq)
1175 {
1176         int this_cpu = this_rq->cpu, ret = 0, cpu;
1177         struct task_struct *p, *next;
1178         struct rq *src_rq;
1179
1180         if (likely(!rt_overloaded(this_rq)))
1181                 return 0;
1182
1183         next = pick_next_task_rt(this_rq);
1184
1185         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1186                 if (this_cpu == cpu)
1187                         continue;
1188
1189                 src_rq = cpu_rq(cpu);
1190                 /*
1191                  * We can potentially drop this_rq's lock in
1192                  * double_lock_balance, and another CPU could
1193                  * steal our next task - hence we must cause
1194                  * the caller to recalculate the next task
1195                  * in that case:
1196                  */
1197                 if (double_lock_balance(this_rq, src_rq)) {
1198                         struct task_struct *old_next = next;
1199
1200                         next = pick_next_task_rt(this_rq);
1201                         if (next != old_next)
1202                                 ret = 1;
1203                 }
1204
1205                 /*
1206                  * Are there still pullable RT tasks?
1207                  */
1208                 if (src_rq->rt.rt_nr_running <= 1)
1209                         goto skip;
1210
1211                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1212
1213                 /*
1214                  * Do we have an RT task that preempts
1215                  * the to-be-scheduled task?
1216                  */
1217                 if (p && (!next || (p->prio < next->prio))) {
1218                         WARN_ON(p == src_rq->curr);
1219                         WARN_ON(!p->se.on_rq);
1220
1221                         /*
1222                          * There's a chance that p is higher in priority
1223                          * than what's currently running on its cpu.
1224                          * This is just that p is wakeing up and hasn't
1225                          * had a chance to schedule. We only pull
1226                          * p if it is lower in priority than the
1227                          * current task on the run queue or
1228                          * this_rq next task is lower in prio than
1229                          * the current task on that rq.
1230                          */
1231                         if (p->prio < src_rq->curr->prio ||
1232                             (next && next->prio < src_rq->curr->prio))
1233                                 goto skip;
1234
1235                         ret = 1;
1236
1237                         deactivate_task(src_rq, p, 0);
1238                         set_task_cpu(p, this_cpu);
1239                         activate_task(this_rq, p, 0);
1240                         /*
1241                          * We continue with the search, just in
1242                          * case there's an even higher prio task
1243                          * in another runqueue. (low likelyhood
1244                          * but possible)
1245                          *
1246                          * Update next so that we won't pick a task
1247                          * on another cpu with a priority lower (or equal)
1248                          * than the one we just picked.
1249                          */
1250                         next = p;
1251
1252                 }
1253  skip:
1254                 double_unlock_balance(this_rq, src_rq);
1255         }
1256
1257         return ret;
1258 }
1259
1260 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1261 {
1262         /* Try to pull RT tasks here if we lower this rq's prio */
1263         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1264                 pull_rt_task(rq);
1265 }
1266
1267 static void post_schedule_rt(struct rq *rq)
1268 {
1269         /*
1270          * If we have more than one rt_task queued, then
1271          * see if we can push the other rt_tasks off to other CPUS.
1272          * Note we may release the rq lock, and since
1273          * the lock was owned by prev, we need to release it
1274          * first via finish_lock_switch and then reaquire it here.
1275          */
1276         if (unlikely(rq->rt.overloaded)) {
1277                 spin_lock_irq(&rq->lock);
1278                 push_rt_tasks(rq);
1279                 spin_unlock_irq(&rq->lock);
1280         }
1281 }
1282
1283 /*
1284  * If we are not running and we are not going to reschedule soon, we should
1285  * try to push tasks away now
1286  */
1287 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1288 {
1289         if (!task_running(rq, p) &&
1290             !test_tsk_need_resched(rq->curr) &&
1291             rq->rt.overloaded)
1292                 push_rt_tasks(rq);
1293 }
1294
1295 static unsigned long
1296 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1297                 unsigned long max_load_move,
1298                 struct sched_domain *sd, enum cpu_idle_type idle,
1299                 int *all_pinned, int *this_best_prio)
1300 {
1301         /* don't touch RT tasks */
1302         return 0;
1303 }
1304
1305 static int
1306 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1307                  struct sched_domain *sd, enum cpu_idle_type idle)
1308 {
1309         /* don't touch RT tasks */
1310         return 0;
1311 }
1312
1313 static void set_cpus_allowed_rt(struct task_struct *p,
1314                                 const struct cpumask *new_mask)
1315 {
1316         int weight = cpumask_weight(new_mask);
1317
1318         BUG_ON(!rt_task(p));
1319
1320         /*
1321          * Update the migration status of the RQ if we have an RT task
1322          * which is running AND changing its weight value.
1323          */
1324         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1325                 struct rq *rq = task_rq(p);
1326
1327                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1328                         rq->rt.rt_nr_migratory++;
1329                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1330                         BUG_ON(!rq->rt.rt_nr_migratory);
1331                         rq->rt.rt_nr_migratory--;
1332                 }
1333
1334                 update_rt_migration(rq);
1335         }
1336
1337         cpumask_copy(&p->cpus_allowed, new_mask);
1338         p->rt.nr_cpus_allowed = weight;
1339 }
1340
1341 /* Assumes rq->lock is held */
1342 static void rq_online_rt(struct rq *rq)
1343 {
1344         if (rq->rt.overloaded)
1345                 rt_set_overload(rq);
1346
1347         __enable_runtime(rq);
1348
1349         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1350 }
1351
1352 /* Assumes rq->lock is held */
1353 static void rq_offline_rt(struct rq *rq)
1354 {
1355         if (rq->rt.overloaded)
1356                 rt_clear_overload(rq);
1357
1358         __disable_runtime(rq);
1359
1360         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1361 }
1362
1363 /*
1364  * When switch from the rt queue, we bring ourselves to a position
1365  * that we might want to pull RT tasks from other runqueues.
1366  */
1367 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1368                            int running)
1369 {
1370         /*
1371          * If there are other RT tasks then we will reschedule
1372          * and the scheduling of the other RT tasks will handle
1373          * the balancing. But if we are the last RT task
1374          * we may need to handle the pulling of RT tasks
1375          * now.
1376          */
1377         if (!rq->rt.rt_nr_running)
1378                 pull_rt_task(rq);
1379 }
1380
1381 static inline void init_sched_rt_class(void)
1382 {
1383         unsigned int i;
1384
1385         for_each_possible_cpu(i)
1386                 alloc_cpumask_var(&per_cpu(local_cpu_mask, i), GFP_KERNEL);
1387 }
1388 #endif /* CONFIG_SMP */
1389
1390 /*
1391  * When switching a task to RT, we may overload the runqueue
1392  * with RT tasks. In this case we try to push them off to
1393  * other runqueues.
1394  */
1395 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1396                            int running)
1397 {
1398         int check_resched = 1;
1399
1400         /*
1401          * If we are already running, then there's nothing
1402          * that needs to be done. But if we are not running
1403          * we may need to preempt the current running task.
1404          * If that current running task is also an RT task
1405          * then see if we can move to another run queue.
1406          */
1407         if (!running) {
1408 #ifdef CONFIG_SMP
1409                 if (rq->rt.overloaded && push_rt_task(rq) &&
1410                     /* Don't resched if we changed runqueues */
1411                     rq != task_rq(p))
1412                         check_resched = 0;
1413 #endif /* CONFIG_SMP */
1414                 if (check_resched && p->prio < rq->curr->prio)
1415                         resched_task(rq->curr);
1416         }
1417 }
1418
1419 /*
1420  * Priority of the task has changed. This may cause
1421  * us to initiate a push or pull.
1422  */
1423 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1424                             int oldprio, int running)
1425 {
1426         if (running) {
1427 #ifdef CONFIG_SMP
1428                 /*
1429                  * If our priority decreases while running, we
1430                  * may need to pull tasks to this runqueue.
1431                  */
1432                 if (oldprio < p->prio)
1433                         pull_rt_task(rq);
1434                 /*
1435                  * If there's a higher priority task waiting to run
1436                  * then reschedule. Note, the above pull_rt_task
1437                  * can release the rq lock and p could migrate.
1438                  * Only reschedule if p is still on the same runqueue.
1439                  */
1440                 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1441                         resched_task(p);
1442 #else
1443                 /* For UP simply resched on drop of prio */
1444                 if (oldprio < p->prio)
1445                         resched_task(p);
1446 #endif /* CONFIG_SMP */
1447         } else {
1448                 /*
1449                  * This task is not running, but if it is
1450                  * greater than the current running task
1451                  * then reschedule.
1452                  */
1453                 if (p->prio < rq->curr->prio)
1454                         resched_task(rq->curr);
1455         }
1456 }
1457
1458 static void watchdog(struct rq *rq, struct task_struct *p)
1459 {
1460         unsigned long soft, hard;
1461
1462         if (!p->signal)
1463                 return;
1464
1465         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1466         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1467
1468         if (soft != RLIM_INFINITY) {
1469                 unsigned long next;
1470
1471                 p->rt.timeout++;
1472                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1473                 if (p->rt.timeout > next)
1474                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1475         }
1476 }
1477
1478 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1479 {
1480         update_curr_rt(rq);
1481
1482         watchdog(rq, p);
1483
1484         /*
1485          * RR tasks need a special form of timeslice management.
1486          * FIFO tasks have no timeslices.
1487          */
1488         if (p->policy != SCHED_RR)
1489                 return;
1490
1491         if (--p->rt.time_slice)
1492                 return;
1493
1494         p->rt.time_slice = DEF_TIMESLICE;
1495
1496         /*
1497          * Requeue to the end of queue if we are not the only element
1498          * on the queue:
1499          */
1500         if (p->rt.run_list.prev != p->rt.run_list.next) {
1501                 requeue_task_rt(rq, p, 0);
1502                 set_tsk_need_resched(p);
1503         }
1504 }
1505
1506 static void set_curr_task_rt(struct rq *rq)
1507 {
1508         struct task_struct *p = rq->curr;
1509
1510         p->se.exec_start = rq->clock;
1511 }
1512
1513 static const struct sched_class rt_sched_class = {
1514         .next                   = &fair_sched_class,
1515         .enqueue_task           = enqueue_task_rt,
1516         .dequeue_task           = dequeue_task_rt,
1517         .yield_task             = yield_task_rt,
1518
1519         .check_preempt_curr     = check_preempt_curr_rt,
1520
1521         .pick_next_task         = pick_next_task_rt,
1522         .put_prev_task          = put_prev_task_rt,
1523
1524 #ifdef CONFIG_SMP
1525         .select_task_rq         = select_task_rq_rt,
1526
1527         .load_balance           = load_balance_rt,
1528         .move_one_task          = move_one_task_rt,
1529         .set_cpus_allowed       = set_cpus_allowed_rt,
1530         .rq_online              = rq_online_rt,
1531         .rq_offline             = rq_offline_rt,
1532         .pre_schedule           = pre_schedule_rt,
1533         .post_schedule          = post_schedule_rt,
1534         .task_wake_up           = task_wake_up_rt,
1535         .switched_from          = switched_from_rt,
1536 #endif
1537
1538         .set_curr_task          = set_curr_task_rt,
1539         .task_tick              = task_tick_rt,
1540
1541         .prio_changed           = prio_changed_rt,
1542         .switched_to            = switched_to_rt,
1543 };
1544
1545 #ifdef CONFIG_SCHED_DEBUG
1546 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1547
1548 static void print_rt_stats(struct seq_file *m, int cpu)
1549 {
1550         struct rt_rq *rt_rq;
1551
1552         rcu_read_lock();
1553         for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1554                 print_rt_rq(m, cpu, rt_rq);
1555         rcu_read_unlock();
1556 }
1557 #endif /* CONFIG_SCHED_DEBUG */
1558