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
  53 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
  54 {
  55         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
  56         plist_node_init(&p->pushable_tasks, p->prio);
  57         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
  58 }
  59
  60 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
  61 {
  62         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
  63 }
  64
  65 #else
  66
  67 #define enqueue_pushable_task(rq, p) do { } while (0)
  68 #define dequeue_pushable_task(rq, p) do { } while (0)
  69
  70 #endif /* CONFIG_SMP */
  71
  72 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  73 {
  74         return container_of(rt_se, struct task_struct, rt);
  75 }
  76
  77 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
  78 {
  79         return !list_empty(&rt_se->run_list);
  80 }
  81
  82 #ifdef CONFIG_RT_GROUP_SCHED
  83
  84 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
  85 {
  86         if (!rt_rq->tg)
  87                 return RUNTIME_INF;
  88
  89         return rt_rq->rt_runtime;
  90 }
  91
  92 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
  93 {
  94         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
  95 }
  96
  97 #define for_each_leaf_rt_rq(rt_rq, rq) \
  98         list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
  99
 100 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
 101 {
 102         return rt_rq->rq;
 103 }
 104
 105 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
 106 {
 107         return rt_se->rt_rq;
 108 }
 109
 110 #define for_each_sched_rt_entity(rt_se) \
 111         for (; rt_se; rt_se = rt_se->parent)
 112
 113 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 114 {
 115         return rt_se->my_q;
 116 }
 117
 118 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
 119 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
 120
 121 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 122 {
 123         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 124         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 125
 126         if (rt_rq->rt_nr_running) {
 127                 if (rt_se && !on_rt_rq(rt_se))
 128                         enqueue_rt_entity(rt_se);
 129                 if (rt_rq->highest_prio.curr < curr->prio)
 130                         resched_task(curr);
 131         }
 132 }
 133
 134 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 135 {
 136         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 137
 138         if (rt_se && on_rt_rq(rt_se))
 139                 dequeue_rt_entity(rt_se);
 140 }
 141
 142 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 143 {
 144         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 145 }
 146
 147 static int rt_se_boosted(struct sched_rt_entity *rt_se)
 148 {
 149         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 150         struct task_struct *p;
 151
 152         if (rt_rq)
 153                 return !!rt_rq->rt_nr_boosted;
 154
 155         p = rt_task_of(rt_se);
 156         return p->prio != p->normal_prio;
 157 }
 158
 159 #ifdef CONFIG_SMP
 160 static inline const struct cpumask *sched_rt_period_mask(void)
 161 {
 162         return cpu_rq(smp_processor_id())->rd->span;
 163 }
 164 #else
 165 static inline const struct cpumask *sched_rt_period_mask(void)
 166 {
 167         return cpu_online_mask;
 168 }
 169 #endif
 170
 171 static inline
 172 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 173 {
 174         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
 175 }
 176
 177 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 178 {
 179         return &rt_rq->tg->rt_bandwidth;
 180 }
 181
 182 #else /* !CONFIG_RT_GROUP_SCHED */
 183
 184 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 185 {
 186         return rt_rq->rt_runtime;
 187 }
 188
 189 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 190 {
 191         return ktime_to_ns(def_rt_bandwidth.rt_period);
 192 }
 193
 194 #define for_each_leaf_rt_rq(rt_rq, rq) \
 195         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 196
 197 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
 198 {
 199         return container_of(rt_rq, struct rq, rt);
 200 }
 201
 202 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
 203 {
 204         struct task_struct *p = rt_task_of(rt_se);
 205         struct rq *rq = task_rq(p);
 206
 207         return &rq->rt;
 208 }
 209
 210 #define for_each_sched_rt_entity(rt_se) \
 211         for (; rt_se; rt_se = NULL)
 212
 213 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 214 {
 215         return NULL;
 216 }
 217
 218 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 219 {
 220         if (rt_rq->rt_nr_running)
 221                 resched_task(rq_of_rt_rq(rt_rq)->curr);
 222 }
 223
 224 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 225 {
 226 }
 227
 228 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 229 {
 230         return rt_rq->rt_throttled;
 231 }
 232
 233 static inline const struct cpumask *sched_rt_period_mask(void)
 234 {
 235         return cpu_online_mask;
 236 }
 237
 238 static inline
 239 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 240 {
 241         return &cpu_rq(cpu)->rt;
 242 }
 243
 244 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 245 {
 246         return &def_rt_bandwidth;
 247 }
 248
 249 #endif /* CONFIG_RT_GROUP_SCHED */
 250
 251 #ifdef CONFIG_SMP
 252 /*
 253  * We ran out of runtime, see if we can borrow some from our neighbours.
 254  */
 255 static int do_balance_runtime(struct rt_rq *rt_rq)
 256 {
 257         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 258         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 259         int i, weight, more = 0;
 260         u64 rt_period;
 261
 262         weight = cpumask_weight(rd->span);
 263
 264         spin_lock(&rt_b->rt_runtime_lock);
 265         rt_period = ktime_to_ns(rt_b->rt_period);
 266         for_each_cpu(i, rd->span) {
 267                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 268                 s64 diff;
 269
 270                 if (iter == rt_rq)
 271                         continue;
 272
 273                 spin_lock(&iter->rt_runtime_lock);
 274                 /*
 275                  * Either all rqs have inf runtime and there's nothing to steal
 276                  * or __disable_runtime() below sets a specific rq to inf to
 277                  * indicate its been disabled and disalow stealing.
 278                  */
 279                 if (iter->rt_runtime == RUNTIME_INF)
 280                         goto next;
 281
 282                 /*
 283                  * From runqueues with spare time, take 1/n part of their
 284                  * spare time, but no more than our period.
 285                  */
 286                 diff = iter->rt_runtime - iter->rt_time;
 287                 if (diff > 0) {
 288                         diff = div_u64((u64)diff, weight);
 289                         if (rt_rq->rt_runtime + diff > rt_period)
 290                                 diff = rt_period - rt_rq->rt_runtime;
 291                         iter->rt_runtime -= diff;
 292                         rt_rq->rt_runtime += diff;
 293                         more = 1;
 294                         if (rt_rq->rt_runtime == rt_period) {
 295                                 spin_unlock(&iter->rt_runtime_lock);
 296                                 break;
 297                         }
 298                 }
 299 next:
 300                 spin_unlock(&iter->rt_runtime_lock);
 301         }
 302         spin_unlock(&rt_b->rt_runtime_lock);
 303
 304         return more;
 305 }
 306
 307 /*
 308  * Ensure this RQ takes back all the runtime it lend to its neighbours.
 309  */
 310 static void __disable_runtime(struct rq *rq)
 311 {
 312         struct root_domain *rd = rq->rd;
 313         struct rt_rq *rt_rq;
 314
 315         if (unlikely(!scheduler_running))
 316                 return;
 317
 318         for_each_leaf_rt_rq(rt_rq, rq) {
 319                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 320                 s64 want;
 321                 int i;
 322
 323                 spin_lock(&rt_b->rt_runtime_lock);
 324                 spin_lock(&rt_rq->rt_runtime_lock);
 325                 /*
 326                  * Either we're all inf and nobody needs to borrow, or we're
 327                  * already disabled and thus have nothing to do, or we have
 328                  * exactly the right amount of runtime to take out.
 329                  */
 330                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 331                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 332                         goto balanced;
 333                 spin_unlock(&rt_rq->rt_runtime_lock);
 334
 335                 /*
 336                  * Calculate the difference between what we started out with
 337                  * and what we current have, that's the amount of runtime
 338                  * we lend and now have to reclaim.
 339                  */
 340                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 341
 342                 /*
 343                  * Greedy reclaim, take back as much as we can.
 344                  */
 345                 for_each_cpu(i, rd->span) {
 346                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 347                         s64 diff;
 348
 349                         /*
 350                          * Can't reclaim from ourselves or disabled runqueues.
 351                          */
 352                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 353                                 continue;
 354
 355                         spin_lock(&iter->rt_runtime_lock);
 356                         if (want > 0) {
 357                                 diff = min_t(s64, iter->rt_runtime, want);
 358                                 iter->rt_runtime -= diff;
 359                                 want -= diff;
 360                         } else {
 361                                 iter->rt_runtime -= want;
 362                                 want -= want;
 363                         }
 364                         spin_unlock(&iter->rt_runtime_lock);
 365
 366                         if (!want)
 367                                 break;
 368                 }
 369
 370                 spin_lock(&rt_rq->rt_runtime_lock);
 371                 /*
 372                  * We cannot be left wanting - that would mean some runtime
 373                  * leaked out of the system.
 374                  */
 375                 BUG_ON(want);
 376 balanced:
 377                 /*
 378                  * Disable all the borrow logic by pretending we have inf
 379                  * runtime - in which case borrowing doesn't make sense.
 380                  */
 381                 rt_rq->rt_runtime = RUNTIME_INF;
 382                 spin_unlock(&rt_rq->rt_runtime_lock);
 383                 spin_unlock(&rt_b->rt_runtime_lock);
 384         }
 385 }
 386
 387 static void disable_runtime(struct rq *rq)
 388 {
 389         unsigned long flags;
 390
 391         spin_lock_irqsave(&rq->lock, flags);
 392         __disable_runtime(rq);
 393         spin_unlock_irqrestore(&rq->lock, flags);
 394 }
 395
 396 static void __enable_runtime(struct rq *rq)
 397 {
 398         struct rt_rq *rt_rq;
 399
 400         if (unlikely(!scheduler_running))
 401                 return;
 402
 403         /*
 404          * Reset each runqueue's bandwidth settings
 405          */
 406         for_each_leaf_rt_rq(rt_rq, rq) {
 407                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 408
 409                 spin_lock(&rt_b->rt_runtime_lock);
 410                 spin_lock(&rt_rq->rt_runtime_lock);
 411                 rt_rq->rt_runtime = rt_b->rt_runtime;
 412                 rt_rq->rt_time = 0;
 413                 rt_rq->rt_throttled = 0;
 414                 spin_unlock(&rt_rq->rt_runtime_lock);
 415                 spin_unlock(&rt_b->rt_runtime_lock);
 416         }
 417 }
 418
 419 static void enable_runtime(struct rq *rq)
 420 {
 421         unsigned long flags;
 422
 423         spin_lock_irqsave(&rq->lock, flags);
 424         __enable_runtime(rq);
 425         spin_unlock_irqrestore(&rq->lock, flags);
 426 }
 427
 428 static int balance_runtime(struct rt_rq *rt_rq)
 429 {
 430         int more = 0;
 431
 432         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 433                 spin_unlock(&rt_rq->rt_runtime_lock);
 434                 more = do_balance_runtime(rt_rq);
 435                 spin_lock(&rt_rq->rt_runtime_lock);
 436         }
 437
 438         return more;
 439 }
 440 #else /* !CONFIG_SMP */
 441 static inline int balance_runtime(struct rt_rq *rt_rq)
 442 {
 443         return 0;
 444 }
 445 #endif /* CONFIG_SMP */
 446
 447 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 448 {
 449         int i, idle = 1;
 450         const struct cpumask *span;
 451
 452         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 453                 return 1;
 454
 455         span = sched_rt_period_mask();
 456         for_each_cpu(i, span) {
 457                 int enqueue = 0;
 458                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 459                 struct rq *rq = rq_of_rt_rq(rt_rq);
 460
 461                 spin_lock(&rq->lock);
 462                 if (rt_rq->rt_time) {
 463                         u64 runtime;
 464
 465                         spin_lock(&rt_rq->rt_runtime_lock);
 466                         if (rt_rq->rt_throttled)
 467                                 balance_runtime(rt_rq);
 468                         runtime = rt_rq->rt_runtime;
 469                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 470                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 471                                 rt_rq->rt_throttled = 0;
 472                                 enqueue = 1;
 473                         }
 474                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 475                                 idle = 0;
 476                         spin_unlock(&rt_rq->rt_runtime_lock);
 477                 } else if (rt_rq->rt_nr_running)
 478                         idle = 0;
 479
 480                 if (enqueue)
 481                         sched_rt_rq_enqueue(rt_rq);
 482                 spin_unlock(&rq->lock);
 483         }
 484
 485         return idle;
 486 }
 487
 488 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 489 {
 490 #ifdef CONFIG_RT_GROUP_SCHED
 491         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 492
 493         if (rt_rq)
 494                 return rt_rq->highest_prio.curr;
 495 #endif
 496
 497         return rt_task_of(rt_se)->prio;
 498 }
 499
 500 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 501 {
 502         u64 runtime = sched_rt_runtime(rt_rq);
 503
 504         if (rt_rq->rt_throttled)
 505                 return rt_rq_throttled(rt_rq);
 506
 507         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 508                 return 0;
 509
 510         balance_runtime(rt_rq);
 511         runtime = sched_rt_runtime(rt_rq);
 512         if (runtime == RUNTIME_INF)
 513                 return 0;
 514
 515         if (rt_rq->rt_time > runtime) {
 516                 rt_rq->rt_throttled = 1;
 517                 if (rt_rq_throttled(rt_rq)) {
 518                         sched_rt_rq_dequeue(rt_rq);
 519                         return 1;
 520                 }
 521         }
 522
 523         return 0;
 524 }
 525
 526 /*
 527  * Update the current task's runtime statistics. Skip current tasks that
 528  * are not in our scheduling class.
 529  */
 530 static void update_curr_rt(struct rq *rq)
 531 {
 532         struct task_struct *curr = rq->curr;
 533         struct sched_rt_entity *rt_se = &curr->rt;
 534         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 535         u64 delta_exec;
 536
 537         if (!task_has_rt_policy(curr))
 538                 return;
 539
 540         delta_exec = rq->clock - curr->se.exec_start;
 541         if (unlikely((s64)delta_exec < 0))
 542                 delta_exec = 0;
 543
 544         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
 545
 546         curr->se.sum_exec_runtime += delta_exec;
 547         account_group_exec_runtime(curr, delta_exec);
 548
 549         curr->se.exec_start = rq->clock;
 550         cpuacct_charge(curr, delta_exec);
 551
 552         if (!rt_bandwidth_enabled())
 553                 return;
 554
 555         for_each_sched_rt_entity(rt_se) {
 556                 rt_rq = rt_rq_of_se(rt_se);
 557
 558                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 559                         spin_lock(&rt_rq->rt_runtime_lock);
 560                         rt_rq->rt_time += delta_exec;
 561                         if (sched_rt_runtime_exceeded(rt_rq))
 562                                 resched_task(curr);
 563                         spin_unlock(&rt_rq->rt_runtime_lock);
 564                 }
 565         }
 566 }
 567
 568 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 569
 570 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
 571
 572 static inline int next_prio(struct rq *rq)
 573 {
 574         struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
 575
 576         if (next && rt_prio(next->prio))
 577                 return next->prio;
 578         else
 579                 return MAX_RT_PRIO;
 580 }
 581 #endif
 582
 583 static inline
 584 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 585 {
 586         int prio = rt_se_prio(rt_se);
 587 #ifdef CONFIG_SMP
 588         struct rq *rq = rq_of_rt_rq(rt_rq);
 589 #endif
 590
 591         WARN_ON(!rt_prio(prio));
 592         rt_rq->rt_nr_running++;
 593 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 594         if (prio < rt_rq->highest_prio.curr) {
 595
 596                 /*
 597                  * If the new task is higher in priority than anything on the
 598                  * run-queue, we have a new high that must be published to
 599                  * the world.  We also know that the previous high becomes
 600                  * our next-highest.
 601                  */
 602                 rt_rq->highest_prio.next = rt_rq->highest_prio.curr;
 603                 rt_rq->highest_prio.curr = prio;
 604 #ifdef CONFIG_SMP
 605                 if (rq->online)
 606                         cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
 607 #endif
 608         } else if (prio == rt_rq->highest_prio.curr)
 609                 /*
 610                  * If the next task is equal in priority to the highest on
 611                  * the run-queue, then we implicitly know that the next highest
 612                  * task cannot be any lower than current
 613                  */
 614                 rt_rq->highest_prio.next = prio;
 615         else if (prio < rt_rq->highest_prio.next)
 616                 /*
 617                  * Otherwise, we need to recompute next-highest
 618                  */
 619                 rt_rq->highest_prio.next = next_prio(rq);
 620 #endif
 621 #ifdef CONFIG_SMP
 622         if (rt_se->nr_cpus_allowed > 1)
 623                 rq->rt.rt_nr_migratory++;
 624
 625         update_rt_migration(rq);
 626 #endif
 627 #ifdef CONFIG_RT_GROUP_SCHED
 628         if (rt_se_boosted(rt_se))
 629                 rt_rq->rt_nr_boosted++;
 630
 631         if (rt_rq->tg)
 632                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 633 #else
 634         start_rt_bandwidth(&def_rt_bandwidth);
 635 #endif
 636 }
 637
 638 static inline
 639 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 640 {
 641 #ifdef CONFIG_SMP
 642         struct rq *rq = rq_of_rt_rq(rt_rq);
 643         int highest_prio = rt_rq->highest_prio.curr;
 644 #endif
 645
 646         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 647         WARN_ON(!rt_rq->rt_nr_running);
 648         rt_rq->rt_nr_running--;
 649 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 650         if (rt_rq->rt_nr_running) {
 651                 int prio = rt_se_prio(rt_se);
 652
 653                 WARN_ON(prio < rt_rq->highest_prio.curr);
 654
 655                 /*
 656                  * This may have been our highest or next-highest priority
 657                  * task and therefore we may have some recomputation to do
 658                  */
 659                 if (prio == rt_rq->highest_prio.curr) {
 660                         struct rt_prio_array *array = &rt_rq->active;
 661
 662                         rt_rq->highest_prio.curr =
 663                                 sched_find_first_bit(array->bitmap);
 664                 }
 665
 666                 if (prio <= rt_rq->highest_prio.next)
 667                         rt_rq->highest_prio.next = next_prio(rq);
 668         } else
 669                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
 670 #endif
 671 #ifdef CONFIG_SMP
 672         if (rt_se->nr_cpus_allowed > 1)
 673                 rq->rt.rt_nr_migratory--;
 674
 675         if (rq->online && rt_rq->highest_prio.curr != highest_prio)
 676                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
 677
 678         update_rt_migration(rq);
 679 #endif /* CONFIG_SMP */
 680 #ifdef CONFIG_RT_GROUP_SCHED
 681         if (rt_se_boosted(rt_se))
 682                 rt_rq->rt_nr_boosted--;
 683
 684         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 685 #endif
 686 }
 687
 688 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
 689 {
 690         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 691         struct rt_prio_array *array = &rt_rq->active;
 692         struct rt_rq *group_rq = group_rt_rq(rt_se);
 693         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 694
 695         /*
 696          * Don't enqueue the group if its throttled, or when empty.
 697          * The latter is a consequence of the former when a child group
 698          * get throttled and the current group doesn't have any other
 699          * active members.
 700          */
 701         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 702                 return;
 703
 704         list_add_tail(&rt_se->run_list, queue);
 705         __set_bit(rt_se_prio(rt_se), array->bitmap);
 706
 707         inc_rt_tasks(rt_se, rt_rq);
 708 }
 709
 710 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 711 {
 712         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 713         struct rt_prio_array *array = &rt_rq->active;
 714
 715         list_del_init(&rt_se->run_list);
 716         if (list_empty(array->queue + rt_se_prio(rt_se)))
 717                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 718
 719         dec_rt_tasks(rt_se, rt_rq);
 720 }
 721
 722 /*
 723  * Because the prio of an upper entry depends on the lower
 724  * entries, we must remove entries top - down.
 725  */
 726 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 727 {
 728         struct sched_rt_entity *back = NULL;
 729
 730         for_each_sched_rt_entity(rt_se) {
 731                 rt_se->back = back;
 732                 back = rt_se;
 733         }
 734
 735         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 736                 if (on_rt_rq(rt_se))
 737                         __dequeue_rt_entity(rt_se);
 738         }
 739 }
 740
 741 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 742 {
 743         dequeue_rt_stack(rt_se);
 744         for_each_sched_rt_entity(rt_se)
 745                 __enqueue_rt_entity(rt_se);
 746 }
 747
 748 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 749 {
 750         dequeue_rt_stack(rt_se);
 751
 752         for_each_sched_rt_entity(rt_se) {
 753                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 754
 755                 if (rt_rq && rt_rq->rt_nr_running)
 756                         __enqueue_rt_entity(rt_se);
 757         }
 758 }
 759
 760 /*
 761  * Adding/removing a task to/from a priority array:
 762  */
 763 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 764 {
 765         struct sched_rt_entity *rt_se = &p->rt;
 766
 767         if (wakeup)
 768                 rt_se->timeout = 0;
 769
 770         enqueue_rt_entity(rt_se);
 771
 772         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
 773                 enqueue_pushable_task(rq, p);
 774
 775         inc_cpu_load(rq, p->se.load.weight);
 776 }
 777
 778 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 779 {
 780         struct sched_rt_entity *rt_se = &p->rt;
 781
 782         update_curr_rt(rq);
 783         dequeue_rt_entity(rt_se);
 784
 785         dequeue_pushable_task(rq, p);
 786
 787         dec_cpu_load(rq, p->se.load.weight);
 788 }
 789
 790 /*
 791  * Put task to the end of the run list without the overhead of dequeue
 792  * followed by enqueue.
 793  */
 794 static void
 795 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 796 {
 797         if (on_rt_rq(rt_se)) {
 798                 struct rt_prio_array *array = &rt_rq->active;
 799                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 800
 801                 if (head)
 802                         list_move(&rt_se->run_list, queue);
 803                 else
 804                         list_move_tail(&rt_se->run_list, queue);
 805         }
 806 }
 807
 808 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 809 {
 810         struct sched_rt_entity *rt_se = &p->rt;
 811         struct rt_rq *rt_rq;
 812
 813         for_each_sched_rt_entity(rt_se) {
 814                 rt_rq = rt_rq_of_se(rt_se);
 815                 requeue_rt_entity(rt_rq, rt_se, head);
 816         }
 817 }
 818
 819 static void yield_task_rt(struct rq *rq)
 820 {
 821         requeue_task_rt(rq, rq->curr, 0);
 822 }
 823
 824 #ifdef CONFIG_SMP
 825 static int find_lowest_rq(struct task_struct *task);
 826
 827 static int select_task_rq_rt(struct task_struct *p, int sync)
 828 {
 829         struct rq *rq = task_rq(p);
 830
 831         /*
 832          * If the current task is an RT task, then
 833          * try to see if we can wake this RT task up on another
 834          * runqueue. Otherwise simply start this RT task
 835          * on its current runqueue.
 836          *
 837          * We want to avoid overloading runqueues. Even if
 838          * the RT task is of higher priority than the current RT task.
 839          * RT tasks behave differently than other tasks. If
 840          * one gets preempted, we try to push it off to another queue.
 841          * So trying to keep a preempting RT task on the same
 842          * cache hot CPU will force the running RT task to
 843          * a cold CPU. So we waste all the cache for the lower
 844          * RT task in hopes of saving some of a RT task
 845          * that is just being woken and probably will have
 846          * cold cache anyway.
 847          */
 848         if (unlikely(rt_task(rq->curr)) &&
 849             (p->rt.nr_cpus_allowed > 1)) {
 850                 int cpu = find_lowest_rq(p);
 851
 852                 return (cpu == -1) ? task_cpu(p) : cpu;
 853         }
 854
 855         /*
 856          * Otherwise, just let it ride on the affined RQ and the
 857          * post-schedule router will push the preempted task away
 858          */
 859         return task_cpu(p);
 860 }
 861
 862 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
 863 {
 864         cpumask_var_t mask;
 865
 866         if (rq->curr->rt.nr_cpus_allowed == 1)
 867                 return;
 868
 869         if (!alloc_cpumask_var(&mask, GFP_ATOMIC))
 870                 return;
 871
 872         if (p->rt.nr_cpus_allowed != 1
 873             && cpupri_find(&rq->rd->cpupri, p, mask))
 874                 goto free;
 875
 876         if (!cpupri_find(&rq->rd->cpupri, rq->curr, mask))
 877                 goto free;
 878
 879         /*
 880          * There appears to be other cpus that can accept
 881          * current and none to run 'p', so lets reschedule
 882          * to try and push current away:
 883          */
 884         requeue_task_rt(rq, p, 1);
 885         resched_task(rq->curr);
 886 free:
 887         free_cpumask_var(mask);
 888 }
 889
 890 #endif /* CONFIG_SMP */
 891
 892 /*
 893  * Preempt the current task with a newly woken task if needed:
 894  */
 895 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
 896 {
 897         if (p->prio < rq->curr->prio) {
 898                 resched_task(rq->curr);
 899                 return;
 900         }
 901
 902 #ifdef CONFIG_SMP
 903         /*
 904          * If:
 905          *
 906          * - the newly woken task is of equal priority to the current task
 907          * - the newly woken task is non-migratable while current is migratable
 908          * - current will be preempted on the next reschedule
 909          *
 910          * we should check to see if current can readily move to a different
 911          * cpu.  If so, we will reschedule to allow the push logic to try
 912          * to move current somewhere else, making room for our non-migratable
 913          * task.
 914          */
 915         if (p->prio == rq->curr->prio && !need_resched())
 916                 check_preempt_equal_prio(rq, p);
 917 #endif
 918 }
 919
 920 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 921                                                    struct rt_rq *rt_rq)
 922 {
 923         struct rt_prio_array *array = &rt_rq->active;
 924         struct sched_rt_entity *next = NULL;
 925         struct list_head *queue;
 926         int idx;
 927
 928         idx = sched_find_first_bit(array->bitmap);
 929         BUG_ON(idx >= MAX_RT_PRIO);
 930
 931         queue = array->queue + idx;
 932         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 933
 934         return next;
 935 }
 936
 937 static struct task_struct *_pick_next_task_rt(struct rq *rq)
 938 {
 939         struct sched_rt_entity *rt_se;
 940         struct task_struct *p;
 941         struct rt_rq *rt_rq;
 942
 943         rt_rq = &rq->rt;
 944
 945         if (unlikely(!rt_rq->rt_nr_running))
 946                 return NULL;
 947
 948         if (rt_rq_throttled(rt_rq))
 949                 return NULL;
 950
 951         do {
 952                 rt_se = pick_next_rt_entity(rq, rt_rq);
 953                 BUG_ON(!rt_se);
 954                 rt_rq = group_rt_rq(rt_se);
 955         } while (rt_rq);
 956
 957         p = rt_task_of(rt_se);
 958         p->se.exec_start = rq->clock;
 959
 960         return p;
 961 }
 962
 963 static struct task_struct *pick_next_task_rt(struct rq *rq)
 964 {
 965         struct task_struct *p = _pick_next_task_rt(rq);
 966
 967         /* The running task is never eligible for pushing */
 968         if (p)
 969                 dequeue_pushable_task(rq, p);
 970
 971         return p;
 972 }
 973
 974 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 975 {
 976         update_curr_rt(rq);
 977         p->se.exec_start = 0;
 978
 979         /*
 980          * The previous task needs to be made eligible for pushing
 981          * if it is still active
 982          */
 983         if (p->se.on_rq && p->rt.nr_cpus_allowed > 1)
 984                 enqueue_pushable_task(rq, p);
 985 }
 986
 987 #ifdef CONFIG_SMP
 988
 989 /* Only try algorithms three times */
 990 #define RT_MAX_TRIES 3
 991
 992 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 993
 994 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 995 {
 996         if (!task_running(rq, p) &&
 997             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
 998             (p->rt.nr_cpus_allowed > 1))
 999                 return 1;
1000         return 0;
1001 }
1002
1003 /* Return the second highest RT task, NULL otherwise */
1004 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1005 {
1006         struct task_struct *next = NULL;
1007         struct sched_rt_entity *rt_se;
1008         struct rt_prio_array *array;
1009         struct rt_rq *rt_rq;
1010         int idx;
1011
1012         for_each_leaf_rt_rq(rt_rq, rq) {
1013                 array = &rt_rq->active;
1014                 idx = sched_find_first_bit(array->bitmap);
1015  next_idx:
1016                 if (idx >= MAX_RT_PRIO)
1017                         continue;
1018                 if (next && next->prio < idx)
1019                         continue;
1020                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1021                         struct task_struct *p = rt_task_of(rt_se);
1022                         if (pick_rt_task(rq, p, cpu)) {
1023                                 next = p;
1024                                 break;
1025                         }
1026                 }
1027                 if (!next) {
1028                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1029                         goto next_idx;
1030                 }
1031         }
1032
1033         return next;
1034 }
1035
1036 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1037
1038 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
1039 {
1040         int first;
1041
1042         /* "this_cpu" is cheaper to preempt than a remote processor */
1043         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
1044                 return this_cpu;
1045
1046         first = first_cpu(*mask);
1047         if (first != NR_CPUS)
1048                 return first;
1049
1050         return -1;
1051 }
1052
1053 static int find_lowest_rq(struct task_struct *task)
1054 {
1055         struct sched_domain *sd;
1056         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1057         int this_cpu = smp_processor_id();
1058         int cpu      = task_cpu(task);
1059
1060         if (task->rt.nr_cpus_allowed == 1)
1061                 return -1; /* No other targets possible */
1062
1063         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1064                 return -1; /* No targets found */
1065
1066         /*
1067          * Only consider CPUs that are usable for migration.
1068          * I guess we might want to change cpupri_find() to ignore those
1069          * in the first place.
1070          */
1071         cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
1072
1073         /*
1074          * At this point we have built a mask of cpus representing the
1075          * lowest priority tasks in the system.  Now we want to elect
1076          * the best one based on our affinity and topology.
1077          *
1078          * We prioritize the last cpu that the task executed on since
1079          * it is most likely cache-hot in that location.
1080          */
1081         if (cpumask_test_cpu(cpu, lowest_mask))
1082                 return cpu;
1083
1084         /*
1085          * Otherwise, we consult the sched_domains span maps to figure
1086          * out which cpu is logically closest to our hot cache data.
1087          */
1088         if (this_cpu == cpu)
1089                 this_cpu = -1; /* Skip this_cpu opt if the same */
1090
1091         for_each_domain(cpu, sd) {
1092                 if (sd->flags & SD_WAKE_AFFINE) {
1093                         cpumask_t domain_mask;
1094                         int       best_cpu;
1095
1096                         cpumask_and(&domain_mask, sched_domain_span(sd),
1097                                     lowest_mask);
1098
1099                         best_cpu = pick_optimal_cpu(this_cpu,
1100                                                     &domain_mask);
1101                         if (best_cpu != -1)
1102                                 return best_cpu;
1103                 }
1104         }
1105
1106         /*
1107          * And finally, if there were no matches within the domains
1108          * just give the caller *something* to work with from the compatible
1109          * locations.
1110          */
1111         return pick_optimal_cpu(this_cpu, lowest_mask);
1112 }
1113
1114 /* Will lock the rq it finds */
1115 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1116 {
1117         struct rq *lowest_rq = NULL;
1118         int tries;
1119         int cpu;
1120
1121         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1122                 cpu = find_lowest_rq(task);
1123
1124                 if ((cpu == -1) || (cpu == rq->cpu))
1125                         break;
1126
1127                 lowest_rq = cpu_rq(cpu);
1128
1129                 /* if the prio of this runqueue changed, try again */
1130                 if (double_lock_balance(rq, lowest_rq)) {
1131                         /*
1132                          * We had to unlock the run queue. In
1133                          * the mean time, task could have
1134                          * migrated already or had its affinity changed.
1135                          * Also make sure that it wasn't scheduled on its rq.
1136                          */
1137                         if (unlikely(task_rq(task) != rq ||
1138                                      !cpumask_test_cpu(lowest_rq->cpu,
1139                                                        &task->cpus_allowed) ||
1140                                      task_running(rq, task) ||
1141                                      !task->se.on_rq)) {
1142
1143                                 spin_unlock(&lowest_rq->lock);
1144                                 lowest_rq = NULL;
1145                                 break;
1146                         }
1147                 }
1148
1149                 /* If this rq is still suitable use it. */
1150                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1151                         break;
1152
1153                 /* try again */
1154                 double_unlock_balance(rq, lowest_rq);
1155                 lowest_rq = NULL;
1156         }
1157
1158         return lowest_rq;
1159 }
1160
1161 static inline int has_pushable_tasks(struct rq *rq)
1162 {
1163         return !plist_head_empty(&rq->rt.pushable_tasks);
1164 }
1165
1166 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1167 {
1168         struct task_struct *p;
1169
1170         if (!has_pushable_tasks(rq))
1171                 return NULL;
1172
1173         p = plist_first_entry(&rq->rt.pushable_tasks,
1174                               struct task_struct, pushable_tasks);
1175
1176         BUG_ON(rq->cpu != task_cpu(p));
1177         BUG_ON(task_current(rq, p));
1178         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1179
1180         BUG_ON(!p->se.on_rq);
1181         BUG_ON(!rt_task(p));
1182
1183         return p;
1184 }
1185
1186 /*
1187  * If the current CPU has more than one RT task, see if the non
1188  * running task can migrate over to a CPU that is running a task
1189  * of lesser priority.
1190  */
1191 static int push_rt_task(struct rq *rq)
1192 {
1193         struct task_struct *next_task;
1194         struct rq *lowest_rq;
1195         int paranoid = RT_MAX_TRIES;
1196
1197         if (!rq->rt.overloaded)
1198                 return 0;
1199
1200         next_task = pick_next_pushable_task(rq);
1201         if (!next_task)
1202                 return 0;
1203
1204  retry:
1205         if (unlikely(next_task == rq->curr)) {
1206                 WARN_ON(1);
1207                 return 0;
1208         }
1209
1210         /*
1211          * It's possible that the next_task slipped in of
1212          * higher priority than current. If that's the case
1213          * just reschedule current.
1214          */
1215         if (unlikely(next_task->prio < rq->curr->prio)) {
1216                 resched_task(rq->curr);
1217                 return 0;
1218         }
1219
1220         /* We might release rq lock */
1221         get_task_struct(next_task);
1222
1223         /* find_lock_lowest_rq locks the rq if found */
1224         lowest_rq = find_lock_lowest_rq(next_task, rq);
1225         if (!lowest_rq) {
1226                 struct task_struct *task;
1227                 /*
1228                  * find lock_lowest_rq releases rq->lock
1229                  * so it is possible that next_task has changed.
1230                  * If it has, then try again.
1231                  */
1232                 task = pick_next_pushable_task(rq);
1233                 if (unlikely(task != next_task) && task && paranoid--) {
1234                         put_task_struct(next_task);
1235                         next_task = task;
1236                         goto retry;
1237                 }
1238
1239                 /*
1240                  * Once we have failed to push this task, we will not
1241                  * try again, since the other cpus will pull from us
1242                  * when they are ready
1243                  */
1244                 dequeue_pushable_task(rq, next_task);
1245                 goto out;
1246         }
1247
1248         deactivate_task(rq, next_task, 0);
1249         set_task_cpu(next_task, lowest_rq->cpu);
1250         activate_task(lowest_rq, next_task, 0);
1251
1252         resched_task(lowest_rq->curr);
1253
1254         double_unlock_balance(rq, lowest_rq);
1255
1256 out:
1257         put_task_struct(next_task);
1258
1259         return 1;
1260 }
1261
1262 static void push_rt_tasks(struct rq *rq)
1263 {
1264         /* push_rt_task will return true if it moved an RT */
1265         while (push_rt_task(rq))
1266                 ;
1267 }
1268
1269 static int pull_rt_task(struct rq *this_rq)
1270 {
1271         int this_cpu = this_rq->cpu, ret = 0, cpu;
1272         struct task_struct *p;
1273         struct rq *src_rq;
1274
1275         if (likely(!rt_overloaded(this_rq)))
1276                 return 0;
1277
1278         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1279                 if (this_cpu == cpu)
1280                         continue;
1281
1282                 src_rq = cpu_rq(cpu);
1283
1284                 /*
1285                  * Don't bother taking the src_rq->lock if the next highest
1286                  * task is known to be lower-priority than our current task.
1287                  * This may look racy, but if this value is about to go
1288                  * logically higher, the src_rq will push this task away.
1289                  * And if its going logically lower, we do not care
1290                  */
1291                 if (src_rq->rt.highest_prio.next >=
1292                     this_rq->rt.highest_prio.curr)
1293                         continue;
1294
1295                 /*
1296                  * We can potentially drop this_rq's lock in
1297                  * double_lock_balance, and another CPU could
1298                  * alter this_rq
1299                  */
1300                 double_lock_balance(this_rq, src_rq);
1301
1302                 /*
1303                  * Are there still pullable RT tasks?
1304                  */
1305                 if (src_rq->rt.rt_nr_running <= 1)
1306                         goto skip;
1307
1308                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1309
1310                 /*
1311                  * Do we have an RT task that preempts
1312                  * the to-be-scheduled task?
1313                  */
1314                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1315                         WARN_ON(p == src_rq->curr);
1316                         WARN_ON(!p->se.on_rq);
1317
1318                         /*
1319                          * There's a chance that p is higher in priority
1320                          * than what's currently running on its cpu.
1321                          * This is just that p is wakeing up and hasn't
1322                          * had a chance to schedule. We only pull
1323                          * p if it is lower in priority than the
1324                          * current task on the run queue
1325                          */
1326                         if (p->prio < src_rq->curr->prio)
1327                                 goto skip;
1328
1329                         ret = 1;
1330
1331                         deactivate_task(src_rq, p, 0);
1332                         set_task_cpu(p, this_cpu);
1333                         activate_task(this_rq, p, 0);
1334                         /*
1335                          * We continue with the search, just in
1336                          * case there's an even higher prio task
1337                          * in another runqueue. (low likelyhood
1338                          * but possible)
1339                          */
1340                 }
1341  skip:
1342                 double_unlock_balance(this_rq, src_rq);
1343         }
1344
1345         return ret;
1346 }
1347
1348 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1349 {
1350         /* Try to pull RT tasks here if we lower this rq's prio */
1351         if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1352                 pull_rt_task(rq);
1353 }
1354
1355 /*
1356  * assumes rq->lock is held
1357  */
1358 static int needs_post_schedule_rt(struct rq *rq)
1359 {
1360         return has_pushable_tasks(rq);
1361 }
1362
1363 static void post_schedule_rt(struct rq *rq)
1364 {
1365         /*
1366          * This is only called if needs_post_schedule_rt() indicates that
1367          * we need to push tasks away
1368          */
1369         spin_lock_irq(&rq->lock);
1370         push_rt_tasks(rq);
1371         spin_unlock_irq(&rq->lock);
1372 }
1373
1374 /*
1375  * If we are not running and we are not going to reschedule soon, we should
1376  * try to push tasks away now
1377  */
1378 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1379 {
1380         if (!task_running(rq, p) &&
1381             !test_tsk_need_resched(rq->curr) &&
1382             has_pushable_tasks(rq) &&
1383             p->rt.nr_cpus_allowed > 1)
1384                 push_rt_tasks(rq);
1385 }
1386
1387 static unsigned long
1388 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1389                 unsigned long max_load_move,
1390                 struct sched_domain *sd, enum cpu_idle_type idle,
1391                 int *all_pinned, int *this_best_prio)
1392 {
1393         /* don't touch RT tasks */
1394         return 0;
1395 }
1396
1397 static int
1398 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1399                  struct sched_domain *sd, enum cpu_idle_type idle)
1400 {
1401         /* don't touch RT tasks */
1402         return 0;
1403 }
1404
1405 static void set_cpus_allowed_rt(struct task_struct *p,
1406                                 const struct cpumask *new_mask)
1407 {
1408         int weight = cpumask_weight(new_mask);
1409
1410         BUG_ON(!rt_task(p));
1411
1412         /*
1413          * Update the migration status of the RQ if we have an RT task
1414          * which is running AND changing its weight value.
1415          */
1416         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1417                 struct rq *rq = task_rq(p);
1418
1419                 if (!task_current(rq, p)) {
1420                         /*
1421                          * Make sure we dequeue this task from the pushable list
1422                          * before going further.  It will either remain off of
1423                          * the list because we are no longer pushable, or it
1424                          * will be requeued.
1425                          */
1426                         if (p->rt.nr_cpus_allowed > 1)
1427                                 dequeue_pushable_task(rq, p);
1428
1429                         /*
1430                          * Requeue if our weight is changing and still > 1
1431                          */
1432                         if (weight > 1)
1433                                 enqueue_pushable_task(rq, p);
1434
1435                 }
1436
1437                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1438                         rq->rt.rt_nr_migratory++;
1439                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1440                         BUG_ON(!rq->rt.rt_nr_migratory);
1441                         rq->rt.rt_nr_migratory--;
1442                 }
1443
1444                 update_rt_migration(rq);
1445         }
1446
1447         cpumask_copy(&p->cpus_allowed, new_mask);
1448         p->rt.nr_cpus_allowed = weight;
1449 }
1450
1451 /* Assumes rq->lock is held */
1452 static void rq_online_rt(struct rq *rq)
1453 {
1454         if (rq->rt.overloaded)
1455                 rt_set_overload(rq);
1456
1457         __enable_runtime(rq);
1458
1459         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1460 }
1461
1462 /* Assumes rq->lock is held */
1463 static void rq_offline_rt(struct rq *rq)
1464 {
1465         if (rq->rt.overloaded)
1466                 rt_clear_overload(rq);
1467
1468         __disable_runtime(rq);
1469
1470         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1471 }
1472
1473 /*
1474  * When switch from the rt queue, we bring ourselves to a position
1475  * that we might want to pull RT tasks from other runqueues.
1476  */
1477 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1478                            int running)
1479 {
1480         /*
1481          * If there are other RT tasks then we will reschedule
1482          * and the scheduling of the other RT tasks will handle
1483          * the balancing. But if we are the last RT task
1484          * we may need to handle the pulling of RT tasks
1485          * now.
1486          */
1487         if (!rq->rt.rt_nr_running)
1488                 pull_rt_task(rq);
1489 }
1490
1491 static inline void init_sched_rt_class(void)
1492 {
1493         unsigned int i;
1494
1495         for_each_possible_cpu(i)
1496                 alloc_cpumask_var(&per_cpu(local_cpu_mask, i), GFP_KERNEL);
1497 }
1498 #endif /* CONFIG_SMP */
1499
1500 /*
1501  * When switching a task to RT, we may overload the runqueue
1502  * with RT tasks. In this case we try to push them off to
1503  * other runqueues.
1504  */
1505 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1506                            int running)
1507 {
1508         int check_resched = 1;
1509
1510         /*
1511          * If we are already running, then there's nothing
1512          * that needs to be done. But if we are not running
1513          * we may need to preempt the current running task.
1514          * If that current running task is also an RT task
1515          * then see if we can move to another run queue.
1516          */
1517         if (!running) {
1518 #ifdef CONFIG_SMP
1519                 if (rq->rt.overloaded && push_rt_task(rq) &&
1520                     /* Don't resched if we changed runqueues */
1521                     rq != task_rq(p))
1522                         check_resched = 0;
1523 #endif /* CONFIG_SMP */
1524                 if (check_resched && p->prio < rq->curr->prio)
1525                         resched_task(rq->curr);
1526         }
1527 }
1528
1529 /*
1530  * Priority of the task has changed. This may cause
1531  * us to initiate a push or pull.
1532  */
1533 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1534                             int oldprio, int running)
1535 {
1536         if (running) {
1537 #ifdef CONFIG_SMP
1538                 /*
1539                  * If our priority decreases while running, we
1540                  * may need to pull tasks to this runqueue.
1541                  */
1542                 if (oldprio < p->prio)
1543                         pull_rt_task(rq);
1544                 /*
1545                  * If there's a higher priority task waiting to run
1546                  * then reschedule. Note, the above pull_rt_task
1547                  * can release the rq lock and p could migrate.
1548                  * Only reschedule if p is still on the same runqueue.
1549                  */
1550                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1551                         resched_task(p);
1552 #else
1553                 /* For UP simply resched on drop of prio */
1554                 if (oldprio < p->prio)
1555                         resched_task(p);
1556 #endif /* CONFIG_SMP */
1557         } else {
1558                 /*
1559                  * This task is not running, but if it is
1560                  * greater than the current running task
1561                  * then reschedule.
1562                  */
1563                 if (p->prio < rq->curr->prio)
1564                         resched_task(rq->curr);
1565         }
1566 }
1567
1568 static void watchdog(struct rq *rq, struct task_struct *p)
1569 {
1570         unsigned long soft, hard;
1571
1572         if (!p->signal)
1573                 return;
1574
1575         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1576         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1577
1578         if (soft != RLIM_INFINITY) {
1579                 unsigned long next;
1580
1581                 p->rt.timeout++;
1582                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1583                 if (p->rt.timeout > next)
1584                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1585         }
1586 }
1587
1588 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1589 {
1590         update_curr_rt(rq);
1591
1592         watchdog(rq, p);
1593
1594         /*
1595          * RR tasks need a special form of timeslice management.
1596          * FIFO tasks have no timeslices.
1597          */
1598         if (p->policy != SCHED_RR)
1599                 return;
1600
1601         if (--p->rt.time_slice)
1602                 return;
1603
1604         p->rt.time_slice = DEF_TIMESLICE;
1605
1606         /*
1607          * Requeue to the end of queue if we are not the only element
1608          * on the queue:
1609          */
1610         if (p->rt.run_list.prev != p->rt.run_list.next) {
1611                 requeue_task_rt(rq, p, 0);
1612                 set_tsk_need_resched(p);
1613         }
1614 }
1615
1616 static void set_curr_task_rt(struct rq *rq)
1617 {
1618         struct task_struct *p = rq->curr;
1619
1620         p->se.exec_start = rq->clock;
1621
1622         /* The running task is never eligible for pushing */
1623         dequeue_pushable_task(rq, p);
1624 }
1625
1626 static const struct sched_class rt_sched_class = {
1627         .next                   = &fair_sched_class,
1628         .enqueue_task           = enqueue_task_rt,
1629         .dequeue_task           = dequeue_task_rt,
1630         .yield_task             = yield_task_rt,
1631
1632         .check_preempt_curr     = check_preempt_curr_rt,
1633
1634         .pick_next_task         = pick_next_task_rt,
1635         .put_prev_task          = put_prev_task_rt,
1636
1637 #ifdef CONFIG_SMP
1638         .select_task_rq         = select_task_rq_rt,
1639
1640         .load_balance           = load_balance_rt,
1641         .move_one_task          = move_one_task_rt,
1642         .set_cpus_allowed       = set_cpus_allowed_rt,
1643         .rq_online              = rq_online_rt,
1644         .rq_offline             = rq_offline_rt,
1645         .pre_schedule           = pre_schedule_rt,
1646         .needs_post_schedule    = needs_post_schedule_rt,
1647         .post_schedule          = post_schedule_rt,
1648         .task_wake_up           = task_wake_up_rt,
1649         .switched_from          = switched_from_rt,
1650 #endif
1651
1652         .set_curr_task          = set_curr_task_rt,
1653         .task_tick              = task_tick_rt,
1654
1655         .prio_changed           = prio_changed_rt,
1656         .switched_to            = switched_to_rt,
1657 };
1658
1659 #ifdef CONFIG_SCHED_DEBUG
1660 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1661
1662 static void print_rt_stats(struct seq_file *m, int cpu)
1663 {
1664         struct rt_rq *rt_rq;
1665
1666         rcu_read_lock();
1667         for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1668                 print_rt_rq(m, cpu, rt_rq);
1669         rcu_read_unlock();
1670 }
1671 #endif /* CONFIG_SCHED_DEBUG */
1672