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_RT_GROUP_SCHED
   7
   8 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
   9
  10 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  11 {
  12 #ifdef CONFIG_SCHED_DEBUG
  13         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
  14 #endif
  15         return container_of(rt_se, struct task_struct, rt);
  16 }
  17
  18 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
  19 {
  20         return rt_rq->rq;
  21 }
  22
  23 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
  24 {
  25         return rt_se->rt_rq;
  26 }
  27
  28 #else /* CONFIG_RT_GROUP_SCHED */
  29
  30 #define rt_entity_is_task(rt_se) (1)
  31
  32 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  33 {
  34         return container_of(rt_se, struct task_struct, rt);
  35 }
  36
  37 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
  38 {
  39         return container_of(rt_rq, struct rq, rt);
  40 }
  41
  42 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
  43 {
  44         struct task_struct *p = rt_task_of(rt_se);
  45         struct rq *rq = task_rq(p);
  46
  47         return &rq->rt;
  48 }
  49
  50 #endif /* CONFIG_RT_GROUP_SCHED */
  51
  52 #ifdef CONFIG_SMP
  53
  54 static inline int rt_overloaded(struct rq *rq)
  55 {
  56         return atomic_read(&rq->rd->rto_count);
  57 }
  58
  59 static inline void rt_set_overload(struct rq *rq)
  60 {
  61         if (!rq->online)
  62                 return;
  63
  64         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
  65         /*
  66          * Make sure the mask is visible before we set
  67          * the overload count. That is checked to determine
  68          * if we should look at the mask. It would be a shame
  69          * if we looked at the mask, but the mask was not
  70          * updated yet.
  71          */
  72         wmb();
  73         atomic_inc(&rq->rd->rto_count);
  74 }
  75
  76 static inline void rt_clear_overload(struct rq *rq)
  77 {
  78         if (!rq->online)
  79                 return;
  80
  81         /* the order here really doesn't matter */
  82         atomic_dec(&rq->rd->rto_count);
  83         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
  84 }
  85
  86 static void update_rt_migration(struct rt_rq *rt_rq)
  87 {
  88         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
  89                 if (!rt_rq->overloaded) {
  90                         rt_set_overload(rq_of_rt_rq(rt_rq));
  91                         rt_rq->overloaded = 1;
  92                 }
  93         } else if (rt_rq->overloaded) {
  94                 rt_clear_overload(rq_of_rt_rq(rt_rq));
  95                 rt_rq->overloaded = 0;
  96         }
  97 }
  98
  99 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 100 {
 101         if (!rt_entity_is_task(rt_se))
 102                 return;
 103
 104         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
 105
 106         rt_rq->rt_nr_total++;
 107         if (rt_se->nr_cpus_allowed > 1)
 108                 rt_rq->rt_nr_migratory++;
 109
 110         update_rt_migration(rt_rq);
 111 }
 112
 113 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 114 {
 115         if (!rt_entity_is_task(rt_se))
 116                 return;
 117
 118         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
 119
 120         rt_rq->rt_nr_total--;
 121         if (rt_se->nr_cpus_allowed > 1)
 122                 rt_rq->rt_nr_migratory--;
 123
 124         update_rt_migration(rt_rq);
 125 }
 126
 127 static inline int has_pushable_tasks(struct rq *rq)
 128 {
 129         return !plist_head_empty(&rq->rt.pushable_tasks);
 130 }
 131
 132 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 133 {
 134         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 135         plist_node_init(&p->pushable_tasks, p->prio);
 136         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
 137
 138         /* Update the highest prio pushable task */
 139         if (p->prio < rq->rt.highest_prio.next)
 140                 rq->rt.highest_prio.next = p->prio;
 141 }
 142
 143 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 144 {
 145         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 146
 147         /* Update the new highest prio pushable task */
 148         if (has_pushable_tasks(rq)) {
 149                 p = plist_first_entry(&rq->rt.pushable_tasks,
 150                                       struct task_struct, pushable_tasks);
 151                 rq->rt.highest_prio.next = p->prio;
 152         } else
 153                 rq->rt.highest_prio.next = MAX_RT_PRIO;
 154 }
 155
 156 #else
 157
 158 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 159 {
 160 }
 161
 162 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 163 {
 164 }
 165
 166 static inline
 167 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 168 {
 169 }
 170
 171 static inline
 172 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 173 {
 174 }
 175
 176 #endif /* CONFIG_SMP */
 177
 178 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
 179 {
 180         return !list_empty(&rt_se->run_list);
 181 }
 182
 183 #ifdef CONFIG_RT_GROUP_SCHED
 184
 185 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 186 {
 187         if (!rt_rq->tg)
 188                 return RUNTIME_INF;
 189
 190         return rt_rq->rt_runtime;
 191 }
 192
 193 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 194 {
 195         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
 196 }
 197
 198 typedef struct task_group *rt_rq_iter_t;
 199
 200 static inline struct task_group *next_task_group(struct task_group *tg)
 201 {
 202         do {
 203                 tg = list_entry_rcu(tg->list.next,
 204                         typeof(struct task_group), list);
 205         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
 206
 207         if (&tg->list == &task_groups)
 208                 tg = NULL;
 209
 210         return tg;
 211 }
 212
 213 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
 214         for (iter = container_of(&task_groups, typeof(*iter), list);    \
 215                 (iter = next_task_group(iter)) &&                       \
 216                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
 217
 218 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
 219 {
 220         list_add_rcu(&rt_rq->leaf_rt_rq_list,
 221                         &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
 222 }
 223
 224 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
 225 {
 226         list_del_rcu(&rt_rq->leaf_rt_rq_list);
 227 }
 228
 229 #define for_each_leaf_rt_rq(rt_rq, rq) \
 230         list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
 231
 232 #define for_each_sched_rt_entity(rt_se) \
 233         for (; rt_se; rt_se = rt_se->parent)
 234
 235 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 236 {
 237         return rt_se->my_q;
 238 }
 239
 240 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
 241 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
 242
 243 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 244 {
 245         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 246         struct sched_rt_entity *rt_se;
 247
 248         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 249
 250         rt_se = rt_rq->tg->rt_se[cpu];
 251
 252         if (rt_rq->rt_nr_running) {
 253                 if (rt_se && !on_rt_rq(rt_se))
 254                         enqueue_rt_entity(rt_se, false);
 255                 if (rt_rq->highest_prio.curr < curr->prio)
 256                         resched_task(curr);
 257         }
 258 }
 259
 260 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 261 {
 262         struct sched_rt_entity *rt_se;
 263         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 264
 265         rt_se = rt_rq->tg->rt_se[cpu];
 266
 267         if (rt_se && on_rt_rq(rt_se))
 268                 dequeue_rt_entity(rt_se);
 269 }
 270
 271 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 272 {
 273         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 274 }
 275
 276 static int rt_se_boosted(struct sched_rt_entity *rt_se)
 277 {
 278         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 279         struct task_struct *p;
 280
 281         if (rt_rq)
 282                 return !!rt_rq->rt_nr_boosted;
 283
 284         p = rt_task_of(rt_se);
 285         return p->prio != p->normal_prio;
 286 }
 287
 288 #ifdef CONFIG_SMP
 289 static inline const struct cpumask *sched_rt_period_mask(void)
 290 {
 291         return cpu_rq(smp_processor_id())->rd->span;
 292 }
 293 #else
 294 static inline const struct cpumask *sched_rt_period_mask(void)
 295 {
 296         return cpu_online_mask;
 297 }
 298 #endif
 299
 300 static inline
 301 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 302 {
 303         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
 304 }
 305
 306 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 307 {
 308         return &rt_rq->tg->rt_bandwidth;
 309 }
 310
 311 #else /* !CONFIG_RT_GROUP_SCHED */
 312
 313 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 314 {
 315         return rt_rq->rt_runtime;
 316 }
 317
 318 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 319 {
 320         return ktime_to_ns(def_rt_bandwidth.rt_period);
 321 }
 322
 323 typedef struct rt_rq *rt_rq_iter_t;
 324
 325 #define for_each_rt_rq(rt_rq, iter, rq) \
 326         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 327
 328 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
 329 {
 330 }
 331
 332 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
 333 {
 334 }
 335
 336 #define for_each_leaf_rt_rq(rt_rq, rq) \
 337         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 338
 339 #define for_each_sched_rt_entity(rt_se) \
 340         for (; rt_se; rt_se = NULL)
 341
 342 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 343 {
 344         return NULL;
 345 }
 346
 347 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 348 {
 349         if (rt_rq->rt_nr_running)
 350                 resched_task(rq_of_rt_rq(rt_rq)->curr);
 351 }
 352
 353 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 354 {
 355 }
 356
 357 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 358 {
 359         return rt_rq->rt_throttled;
 360 }
 361
 362 static inline const struct cpumask *sched_rt_period_mask(void)
 363 {
 364         return cpu_online_mask;
 365 }
 366
 367 static inline
 368 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 369 {
 370         return &cpu_rq(cpu)->rt;
 371 }
 372
 373 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 374 {
 375         return &def_rt_bandwidth;
 376 }
 377
 378 #endif /* CONFIG_RT_GROUP_SCHED */
 379
 380 #ifdef CONFIG_SMP
 381 /*
 382  * We ran out of runtime, see if we can borrow some from our neighbours.
 383  */
 384 static int do_balance_runtime(struct rt_rq *rt_rq)
 385 {
 386         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 387         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 388         int i, weight, more = 0;
 389         u64 rt_period;
 390
 391         weight = cpumask_weight(rd->span);
 392
 393         raw_spin_lock(&rt_b->rt_runtime_lock);
 394         rt_period = ktime_to_ns(rt_b->rt_period);
 395         for_each_cpu(i, rd->span) {
 396                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 397                 s64 diff;
 398
 399                 if (iter == rt_rq)
 400                         continue;
 401
 402                 raw_spin_lock(&iter->rt_runtime_lock);
 403                 /*
 404                  * Either all rqs have inf runtime and there's nothing to steal
 405                  * or __disable_runtime() below sets a specific rq to inf to
 406                  * indicate its been disabled and disalow stealing.
 407                  */
 408                 if (iter->rt_runtime == RUNTIME_INF)
 409                         goto next;
 410
 411                 /*
 412                  * From runqueues with spare time, take 1/n part of their
 413                  * spare time, but no more than our period.
 414                  */
 415                 diff = iter->rt_runtime - iter->rt_time;
 416                 if (diff > 0) {
 417                         diff = div_u64((u64)diff, weight);
 418                         if (rt_rq->rt_runtime + diff > rt_period)
 419                                 diff = rt_period - rt_rq->rt_runtime;
 420                         iter->rt_runtime -= diff;
 421                         rt_rq->rt_runtime += diff;
 422                         more = 1;
 423                         if (rt_rq->rt_runtime == rt_period) {
 424                                 raw_spin_unlock(&iter->rt_runtime_lock);
 425                                 break;
 426                         }
 427                 }
 428 next:
 429                 raw_spin_unlock(&iter->rt_runtime_lock);
 430         }
 431         raw_spin_unlock(&rt_b->rt_runtime_lock);
 432
 433         return more;
 434 }
 435
 436 /*
 437  * Ensure this RQ takes back all the runtime it lend to its neighbours.
 438  */
 439 static void __disable_runtime(struct rq *rq)
 440 {
 441         struct root_domain *rd = rq->rd;
 442         rt_rq_iter_t iter;
 443         struct rt_rq *rt_rq;
 444
 445         if (unlikely(!scheduler_running))
 446                 return;
 447
 448         for_each_rt_rq(rt_rq, iter, rq) {
 449                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 450                 s64 want;
 451                 int i;
 452
 453                 raw_spin_lock(&rt_b->rt_runtime_lock);
 454                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 455                 /*
 456                  * Either we're all inf and nobody needs to borrow, or we're
 457                  * already disabled and thus have nothing to do, or we have
 458                  * exactly the right amount of runtime to take out.
 459                  */
 460                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 461                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 462                         goto balanced;
 463                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 464
 465                 /*
 466                  * Calculate the difference between what we started out with
 467                  * and what we current have, that's the amount of runtime
 468                  * we lend and now have to reclaim.
 469                  */
 470                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 471
 472                 /*
 473                  * Greedy reclaim, take back as much as we can.
 474                  */
 475                 for_each_cpu(i, rd->span) {
 476                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 477                         s64 diff;
 478
 479                         /*
 480                          * Can't reclaim from ourselves or disabled runqueues.
 481                          */
 482                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 483                                 continue;
 484
 485                         raw_spin_lock(&iter->rt_runtime_lock);
 486                         if (want > 0) {
 487                                 diff = min_t(s64, iter->rt_runtime, want);
 488                                 iter->rt_runtime -= diff;
 489                                 want -= diff;
 490                         } else {
 491                                 iter->rt_runtime -= want;
 492                                 want -= want;
 493                         }
 494                         raw_spin_unlock(&iter->rt_runtime_lock);
 495
 496                         if (!want)
 497                                 break;
 498                 }
 499
 500                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 501                 /*
 502                  * We cannot be left wanting - that would mean some runtime
 503                  * leaked out of the system.
 504                  */
 505                 BUG_ON(want);
 506 balanced:
 507                 /*
 508                  * Disable all the borrow logic by pretending we have inf
 509                  * runtime - in which case borrowing doesn't make sense.
 510                  */
 511                 rt_rq->rt_runtime = RUNTIME_INF;
 512                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 513                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 514         }
 515 }
 516
 517 static void disable_runtime(struct rq *rq)
 518 {
 519         unsigned long flags;
 520
 521         raw_spin_lock_irqsave(&rq->lock, flags);
 522         __disable_runtime(rq);
 523         raw_spin_unlock_irqrestore(&rq->lock, flags);
 524 }
 525
 526 static void __enable_runtime(struct rq *rq)
 527 {
 528         rt_rq_iter_t iter;
 529         struct rt_rq *rt_rq;
 530
 531         if (unlikely(!scheduler_running))
 532                 return;
 533
 534         /*
 535          * Reset each runqueue's bandwidth settings
 536          */
 537         for_each_rt_rq(rt_rq, iter, rq) {
 538                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 539
 540                 raw_spin_lock(&rt_b->rt_runtime_lock);
 541                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 542                 rt_rq->rt_runtime = rt_b->rt_runtime;
 543                 rt_rq->rt_time = 0;
 544                 rt_rq->rt_throttled = 0;
 545                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 546                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 547         }
 548 }
 549
 550 static void enable_runtime(struct rq *rq)
 551 {
 552         unsigned long flags;
 553
 554         raw_spin_lock_irqsave(&rq->lock, flags);
 555         __enable_runtime(rq);
 556         raw_spin_unlock_irqrestore(&rq->lock, flags);
 557 }
 558
 559 static int balance_runtime(struct rt_rq *rt_rq)
 560 {
 561         int more = 0;
 562
 563         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 564                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 565                 more = do_balance_runtime(rt_rq);
 566                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 567         }
 568
 569         return more;
 570 }
 571 #else /* !CONFIG_SMP */
 572 static inline int balance_runtime(struct rt_rq *rt_rq)
 573 {
 574         return 0;
 575 }
 576 #endif /* CONFIG_SMP */
 577
 578 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 579 {
 580         int i, idle = 1;
 581         const struct cpumask *span;
 582
 583         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 584                 return 1;
 585
 586         span = sched_rt_period_mask();
 587         for_each_cpu(i, span) {
 588                 int enqueue = 0;
 589                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 590                 struct rq *rq = rq_of_rt_rq(rt_rq);
 591
 592                 raw_spin_lock(&rq->lock);
 593                 if (rt_rq->rt_time) {
 594                         u64 runtime;
 595
 596                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 597                         if (rt_rq->rt_throttled)
 598                                 balance_runtime(rt_rq);
 599                         runtime = rt_rq->rt_runtime;
 600                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 601                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 602                                 rt_rq->rt_throttled = 0;
 603                                 enqueue = 1;
 604
 605                                 /*
 606                                  * Force a clock update if the CPU was idle,
 607                                  * lest wakeup -> unthrottle time accumulate.
 608                                  */
 609                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
 610                                         rq->skip_clock_update = -1;
 611                         }
 612                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 613                                 idle = 0;
 614                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 615                 } else if (rt_rq->rt_nr_running) {
 616                         idle = 0;
 617                         if (!rt_rq_throttled(rt_rq))
 618                                 enqueue = 1;
 619                 }
 620
 621                 if (enqueue)
 622                         sched_rt_rq_enqueue(rt_rq);
 623                 raw_spin_unlock(&rq->lock);
 624         }
 625
 626         return idle;
 627 }
 628
 629 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 630 {
 631 #ifdef CONFIG_RT_GROUP_SCHED
 632         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 633
 634         if (rt_rq)
 635                 return rt_rq->highest_prio.curr;
 636 #endif
 637
 638         return rt_task_of(rt_se)->prio;
 639 }
 640
 641 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 642 {
 643         u64 runtime = sched_rt_runtime(rt_rq);
 644
 645         if (rt_rq->rt_throttled)
 646                 return rt_rq_throttled(rt_rq);
 647
 648         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 649                 return 0;
 650
 651         balance_runtime(rt_rq);
 652         runtime = sched_rt_runtime(rt_rq);
 653         if (runtime == RUNTIME_INF)
 654                 return 0;
 655
 656         if (rt_rq->rt_time > runtime) {
 657                 rt_rq->rt_throttled = 1;
 658                 if (rt_rq_throttled(rt_rq)) {
 659                         sched_rt_rq_dequeue(rt_rq);
 660                         return 1;
 661                 }
 662         }
 663
 664         return 0;
 665 }
 666
 667 /*
 668  * Update the current task's runtime statistics. Skip current tasks that
 669  * are not in our scheduling class.
 670  */
 671 static void update_curr_rt(struct rq *rq)
 672 {
 673         struct task_struct *curr = rq->curr;
 674         struct sched_rt_entity *rt_se = &curr->rt;
 675         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 676         u64 delta_exec;
 677
 678         if (curr->sched_class != &rt_sched_class)
 679                 return;
 680
 681         delta_exec = rq->clock_task - curr->se.exec_start;
 682         if (unlikely((s64)delta_exec < 0))
 683                 delta_exec = 0;
 684
 685         schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
 686
 687         curr->se.sum_exec_runtime += delta_exec;
 688         account_group_exec_runtime(curr, delta_exec);
 689
 690         curr->se.exec_start = rq->clock_task;
 691         cpuacct_charge(curr, delta_exec);
 692
 693         sched_rt_avg_update(rq, delta_exec);
 694
 695         if (!rt_bandwidth_enabled())
 696                 return;
 697
 698         for_each_sched_rt_entity(rt_se) {
 699                 rt_rq = rt_rq_of_se(rt_se);
 700
 701                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 702                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 703                         rt_rq->rt_time += delta_exec;
 704                         if (sched_rt_runtime_exceeded(rt_rq))
 705                                 resched_task(curr);
 706                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 707                 }
 708         }
 709 }
 710
 711 #if defined CONFIG_SMP
 712
 713 static void
 714 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 715 {
 716         struct rq *rq = rq_of_rt_rq(rt_rq);
 717
 718         if (rq->online && prio < prev_prio)
 719                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
 720 }
 721
 722 static void
 723 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 724 {
 725         struct rq *rq = rq_of_rt_rq(rt_rq);
 726
 727         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
 728                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
 729 }
 730
 731 #else /* CONFIG_SMP */
 732
 733 static inline
 734 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 735 static inline
 736 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 737
 738 #endif /* CONFIG_SMP */
 739
 740 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 741 static void
 742 inc_rt_prio(struct rt_rq *rt_rq, int prio)
 743 {
 744         int prev_prio = rt_rq->highest_prio.curr;
 745
 746         if (prio < prev_prio)
 747                 rt_rq->highest_prio.curr = prio;
 748
 749         inc_rt_prio_smp(rt_rq, prio, prev_prio);
 750 }
 751
 752 static void
 753 dec_rt_prio(struct rt_rq *rt_rq, int prio)
 754 {
 755         int prev_prio = rt_rq->highest_prio.curr;
 756
 757         if (rt_rq->rt_nr_running) {
 758
 759                 WARN_ON(prio < prev_prio);
 760
 761                 /*
 762                  * This may have been our highest task, and therefore
 763                  * we may have some recomputation to do
 764                  */
 765                 if (prio == prev_prio) {
 766                         struct rt_prio_array *array = &rt_rq->active;
 767
 768                         rt_rq->highest_prio.curr =
 769                                 sched_find_first_bit(array->bitmap);
 770                 }
 771
 772         } else
 773                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
 774
 775         dec_rt_prio_smp(rt_rq, prio, prev_prio);
 776 }
 777
 778 #else
 779
 780 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
 781 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
 782
 783 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
 784
 785 #ifdef CONFIG_RT_GROUP_SCHED
 786
 787 static void
 788 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 789 {
 790         if (rt_se_boosted(rt_se))
 791                 rt_rq->rt_nr_boosted++;
 792
 793         if (rt_rq->tg)
 794                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 795 }
 796
 797 static void
 798 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 799 {
 800         if (rt_se_boosted(rt_se))
 801                 rt_rq->rt_nr_boosted--;
 802
 803         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 804 }
 805
 806 #else /* CONFIG_RT_GROUP_SCHED */
 807
 808 static void
 809 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 810 {
 811         start_rt_bandwidth(&def_rt_bandwidth);
 812 }
 813
 814 static inline
 815 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
 816
 817 #endif /* CONFIG_RT_GROUP_SCHED */
 818
 819 static inline
 820 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 821 {
 822         int prio = rt_se_prio(rt_se);
 823
 824         WARN_ON(!rt_prio(prio));
 825         rt_rq->rt_nr_running++;
 826
 827         inc_rt_prio(rt_rq, prio);
 828         inc_rt_migration(rt_se, rt_rq);
 829         inc_rt_group(rt_se, rt_rq);
 830 }
 831
 832 static inline
 833 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 834 {
 835         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 836         WARN_ON(!rt_rq->rt_nr_running);
 837         rt_rq->rt_nr_running--;
 838
 839         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
 840         dec_rt_migration(rt_se, rt_rq);
 841         dec_rt_group(rt_se, rt_rq);
 842 }
 843
 844 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 845 {
 846         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 847         struct rt_prio_array *array = &rt_rq->active;
 848         struct rt_rq *group_rq = group_rt_rq(rt_se);
 849         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 850
 851         /*
 852          * Don't enqueue the group if its throttled, or when empty.
 853          * The latter is a consequence of the former when a child group
 854          * get throttled and the current group doesn't have any other
 855          * active members.
 856          */
 857         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 858                 return;
 859
 860         if (!rt_rq->rt_nr_running)
 861                 list_add_leaf_rt_rq(rt_rq);
 862
 863         if (head)
 864                 list_add(&rt_se->run_list, queue);
 865         else
 866                 list_add_tail(&rt_se->run_list, queue);
 867         __set_bit(rt_se_prio(rt_se), array->bitmap);
 868
 869         inc_rt_tasks(rt_se, rt_rq);
 870 }
 871
 872 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 873 {
 874         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 875         struct rt_prio_array *array = &rt_rq->active;
 876
 877         list_del_init(&rt_se->run_list);
 878         if (list_empty(array->queue + rt_se_prio(rt_se)))
 879                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 880
 881         dec_rt_tasks(rt_se, rt_rq);
 882         if (!rt_rq->rt_nr_running)
 883                 list_del_leaf_rt_rq(rt_rq);
 884 }
 885
 886 /*
 887  * Because the prio of an upper entry depends on the lower
 888  * entries, we must remove entries top - down.
 889  */
 890 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 891 {
 892         struct sched_rt_entity *back = NULL;
 893
 894         for_each_sched_rt_entity(rt_se) {
 895                 rt_se->back = back;
 896                 back = rt_se;
 897         }
 898
 899         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 900                 if (on_rt_rq(rt_se))
 901                         __dequeue_rt_entity(rt_se);
 902         }
 903 }
 904
 905 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 906 {
 907         dequeue_rt_stack(rt_se);
 908         for_each_sched_rt_entity(rt_se)
 909                 __enqueue_rt_entity(rt_se, head);
 910 }
 911
 912 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 913 {
 914         dequeue_rt_stack(rt_se);
 915
 916         for_each_sched_rt_entity(rt_se) {
 917                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 918
 919                 if (rt_rq && rt_rq->rt_nr_running)
 920                         __enqueue_rt_entity(rt_se, false);
 921         }
 922 }
 923
 924 /*
 925  * Adding/removing a task to/from a priority array:
 926  */
 927 static void
 928 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 929 {
 930         struct sched_rt_entity *rt_se = &p->rt;
 931
 932         if (flags & ENQUEUE_WAKEUP)
 933                 rt_se->timeout = 0;
 934
 935         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
 936
 937         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
 938                 enqueue_pushable_task(rq, p);
 939
 940         inc_nr_running(rq);
 941 }
 942
 943 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 944 {
 945         struct sched_rt_entity *rt_se = &p->rt;
 946
 947         update_curr_rt(rq);
 948         dequeue_rt_entity(rt_se);
 949
 950         dequeue_pushable_task(rq, p);
 951
 952         dec_nr_running(rq);
 953 }
 954
 955 /*
 956  * Put task to the end of the run list without the overhead of dequeue
 957  * followed by enqueue.
 958  */
 959 static void
 960 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 961 {
 962         if (on_rt_rq(rt_se)) {
 963                 struct rt_prio_array *array = &rt_rq->active;
 964                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 965
 966                 if (head)
 967                         list_move(&rt_se->run_list, queue);
 968                 else
 969                         list_move_tail(&rt_se->run_list, queue);
 970         }
 971 }
 972
 973 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 974 {
 975         struct sched_rt_entity *rt_se = &p->rt;
 976         struct rt_rq *rt_rq;
 977
 978         for_each_sched_rt_entity(rt_se) {
 979                 rt_rq = rt_rq_of_se(rt_se);
 980                 requeue_rt_entity(rt_rq, rt_se, head);
 981         }
 982 }
 983
 984 static void yield_task_rt(struct rq *rq)
 985 {
 986         requeue_task_rt(rq, rq->curr, 0);
 987 }
 988
 989 #ifdef CONFIG_SMP
 990 static int find_lowest_rq(struct task_struct *task);
 991
 992 static int
 993 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
 994 {
 995         struct task_struct *curr;
 996         struct rq *rq;
 997         int cpu;
 998
 999         cpu = task_cpu(p);
1000
1001         /* For anything but wake ups, just return the task_cpu */
1002         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1003                 goto out;
1004
1005         rq = cpu_rq(cpu);
1006
1007         rcu_read_lock();
1008         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1009
1010         /*
1011          * If the current task on @p's runqueue is an RT task, then
1012          * try to see if we can wake this RT task up on another
1013          * runqueue. Otherwise simply start this RT task
1014          * on its current runqueue.
1015          *
1016          * We want to avoid overloading runqueues. If the woken
1017          * task is a higher priority, then it will stay on this CPU
1018          * and the lower prio task should be moved to another CPU.
1019          * Even though this will probably make the lower prio task
1020          * lose its cache, we do not want to bounce a higher task
1021          * around just because it gave up its CPU, perhaps for a
1022          * lock?
1023          *
1024          * For equal prio tasks, we just let the scheduler sort it out.
1025          *
1026          * Otherwise, just let it ride on the affined RQ and the
1027          * post-schedule router will push the preempted task away
1028          *
1029          * This test is optimistic, if we get it wrong the load-balancer
1030          * will have to sort it out.
1031          */
1032         if (curr && unlikely(rt_task(curr)) &&
1033             (curr->rt.nr_cpus_allowed < 2 ||
1034              curr->prio < p->prio) &&
1035             (p->rt.nr_cpus_allowed > 1)) {
1036                 int target = find_lowest_rq(p);
1037
1038                 if (target != -1)
1039                         cpu = target;
1040         }
1041         rcu_read_unlock();
1042
1043 out:
1044         return cpu;
1045 }
1046
1047 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1048 {
1049         if (rq->curr->rt.nr_cpus_allowed == 1)
1050                 return;
1051
1052         if (p->rt.nr_cpus_allowed != 1
1053             && cpupri_find(&rq->rd->cpupri, p, NULL))
1054                 return;
1055
1056         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1057                 return;
1058
1059         /*
1060          * There appears to be other cpus that can accept
1061          * current and none to run 'p', so lets reschedule
1062          * to try and push current away:
1063          */
1064         requeue_task_rt(rq, p, 1);
1065         resched_task(rq->curr);
1066 }
1067
1068 #endif /* CONFIG_SMP */
1069
1070 /*
1071  * Preempt the current task with a newly woken task if needed:
1072  */
1073 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1074 {
1075         if (p->prio < rq->curr->prio) {
1076                 resched_task(rq->curr);
1077                 return;
1078         }
1079
1080 #ifdef CONFIG_SMP
1081         /*
1082          * If:
1083          *
1084          * - the newly woken task is of equal priority to the current task
1085          * - the newly woken task is non-migratable while current is migratable
1086          * - current will be preempted on the next reschedule
1087          *
1088          * we should check to see if current can readily move to a different
1089          * cpu.  If so, we will reschedule to allow the push logic to try
1090          * to move current somewhere else, making room for our non-migratable
1091          * task.
1092          */
1093         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1094                 check_preempt_equal_prio(rq, p);
1095 #endif
1096 }
1097
1098 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1099                                                    struct rt_rq *rt_rq)
1100 {
1101         struct rt_prio_array *array = &rt_rq->active;
1102         struct sched_rt_entity *next = NULL;
1103         struct list_head *queue;
1104         int idx;
1105
1106         idx = sched_find_first_bit(array->bitmap);
1107         BUG_ON(idx >= MAX_RT_PRIO);
1108
1109         queue = array->queue + idx;
1110         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1111
1112         return next;
1113 }
1114
1115 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1116 {
1117         struct sched_rt_entity *rt_se;
1118         struct task_struct *p;
1119         struct rt_rq *rt_rq;
1120
1121         rt_rq = &rq->rt;
1122
1123         if (!rt_rq->rt_nr_running)
1124                 return NULL;
1125
1126         if (rt_rq_throttled(rt_rq))
1127                 return NULL;
1128
1129         do {
1130                 rt_se = pick_next_rt_entity(rq, rt_rq);
1131                 BUG_ON(!rt_se);
1132                 rt_rq = group_rt_rq(rt_se);
1133         } while (rt_rq);
1134
1135         p = rt_task_of(rt_se);
1136         p->se.exec_start = rq->clock_task;
1137
1138         return p;
1139 }
1140
1141 static struct task_struct *pick_next_task_rt(struct rq *rq)
1142 {
1143         struct task_struct *p = _pick_next_task_rt(rq);
1144
1145         /* The running task is never eligible for pushing */
1146         if (p)
1147                 dequeue_pushable_task(rq, p);
1148
1149 #ifdef CONFIG_SMP
1150         /*
1151          * We detect this state here so that we can avoid taking the RQ
1152          * lock again later if there is no need to push
1153          */
1154         rq->post_schedule = has_pushable_tasks(rq);
1155 #endif
1156
1157         return p;
1158 }
1159
1160 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1161 {
1162         update_curr_rt(rq);
1163
1164         /*
1165          * The previous task needs to be made eligible for pushing
1166          * if it is still active
1167          */
1168         if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1169                 enqueue_pushable_task(rq, p);
1170 }
1171
1172 #ifdef CONFIG_SMP
1173
1174 /* Only try algorithms three times */
1175 #define RT_MAX_TRIES 3
1176
1177 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1178
1179 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1180 {
1181         if (!task_running(rq, p) &&
1182             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1183             (p->rt.nr_cpus_allowed > 1))
1184                 return 1;
1185         return 0;
1186 }
1187
1188 /* Return the second highest RT task, NULL otherwise */
1189 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1190 {
1191         struct task_struct *next = NULL;
1192         struct sched_rt_entity *rt_se;
1193         struct rt_prio_array *array;
1194         struct rt_rq *rt_rq;
1195         int idx;
1196
1197         for_each_leaf_rt_rq(rt_rq, rq) {
1198                 array = &rt_rq->active;
1199                 idx = sched_find_first_bit(array->bitmap);
1200 next_idx:
1201                 if (idx >= MAX_RT_PRIO)
1202                         continue;
1203                 if (next && next->prio < idx)
1204                         continue;
1205                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1206                         struct task_struct *p;
1207
1208                         if (!rt_entity_is_task(rt_se))
1209                                 continue;
1210
1211                         p = rt_task_of(rt_se);
1212                         if (pick_rt_task(rq, p, cpu)) {
1213                                 next = p;
1214                                 break;
1215                         }
1216                 }
1217                 if (!next) {
1218                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1219                         goto next_idx;
1220                 }
1221         }
1222
1223         return next;
1224 }
1225
1226 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1227
1228 static int find_lowest_rq(struct task_struct *task)
1229 {
1230         struct sched_domain *sd;
1231         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1232         int this_cpu = smp_processor_id();
1233         int cpu      = task_cpu(task);
1234
1235         /* Make sure the mask is initialized first */
1236         if (unlikely(!lowest_mask))
1237                 return -1;
1238
1239         if (task->rt.nr_cpus_allowed == 1)
1240                 return -1; /* No other targets possible */
1241
1242         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1243                 return -1; /* No targets found */
1244
1245         /*
1246          * At this point we have built a mask of cpus representing the
1247          * lowest priority tasks in the system.  Now we want to elect
1248          * the best one based on our affinity and topology.
1249          *
1250          * We prioritize the last cpu that the task executed on since
1251          * it is most likely cache-hot in that location.
1252          */
1253         if (cpumask_test_cpu(cpu, lowest_mask))
1254                 return cpu;
1255
1256         /*
1257          * Otherwise, we consult the sched_domains span maps to figure
1258          * out which cpu is logically closest to our hot cache data.
1259          */
1260         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1261                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1262
1263         rcu_read_lock();
1264         for_each_domain(cpu, sd) {
1265                 if (sd->flags & SD_WAKE_AFFINE) {
1266                         int best_cpu;
1267
1268                         /*
1269                          * "this_cpu" is cheaper to preempt than a
1270                          * remote processor.
1271                          */
1272                         if (this_cpu != -1 &&
1273                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1274                                 rcu_read_unlock();
1275                                 return this_cpu;
1276                         }
1277
1278                         best_cpu = cpumask_first_and(lowest_mask,
1279                                                      sched_domain_span(sd));
1280                         if (best_cpu < nr_cpu_ids) {
1281                                 rcu_read_unlock();
1282                                 return best_cpu;
1283                         }
1284                 }
1285         }
1286         rcu_read_unlock();
1287
1288         /*
1289          * And finally, if there were no matches within the domains
1290          * just give the caller *something* to work with from the compatible
1291          * locations.
1292          */
1293         if (this_cpu != -1)
1294                 return this_cpu;
1295
1296         cpu = cpumask_any(lowest_mask);
1297         if (cpu < nr_cpu_ids)
1298                 return cpu;
1299         return -1;
1300 }
1301
1302 /* Will lock the rq it finds */
1303 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1304 {
1305         struct rq *lowest_rq = NULL;
1306         int tries;
1307         int cpu;
1308
1309         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1310                 cpu = find_lowest_rq(task);
1311
1312                 if ((cpu == -1) || (cpu == rq->cpu))
1313                         break;
1314
1315                 lowest_rq = cpu_rq(cpu);
1316
1317                 /* if the prio of this runqueue changed, try again */
1318                 if (double_lock_balance(rq, lowest_rq)) {
1319                         /*
1320                          * We had to unlock the run queue. In
1321                          * the mean time, task could have
1322                          * migrated already or had its affinity changed.
1323                          * Also make sure that it wasn't scheduled on its rq.
1324                          */
1325                         if (unlikely(task_rq(task) != rq ||
1326                                      !cpumask_test_cpu(lowest_rq->cpu,
1327                                                        &task->cpus_allowed) ||
1328                                      task_running(rq, task) ||
1329                                      !task->on_rq)) {
1330
1331                                 raw_spin_unlock(&lowest_rq->lock);
1332                                 lowest_rq = NULL;
1333                                 break;
1334                         }
1335                 }
1336
1337                 /* If this rq is still suitable use it. */
1338                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1339                         break;
1340
1341                 /* try again */
1342                 double_unlock_balance(rq, lowest_rq);
1343                 lowest_rq = NULL;
1344         }
1345
1346         return lowest_rq;
1347 }
1348
1349 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1350 {
1351         struct task_struct *p;
1352
1353         if (!has_pushable_tasks(rq))
1354                 return NULL;
1355
1356         p = plist_first_entry(&rq->rt.pushable_tasks,
1357                               struct task_struct, pushable_tasks);
1358
1359         BUG_ON(rq->cpu != task_cpu(p));
1360         BUG_ON(task_current(rq, p));
1361         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1362
1363         BUG_ON(!p->on_rq);
1364         BUG_ON(!rt_task(p));
1365
1366         return p;
1367 }
1368
1369 /*
1370  * If the current CPU has more than one RT task, see if the non
1371  * running task can migrate over to a CPU that is running a task
1372  * of lesser priority.
1373  */
1374 static int push_rt_task(struct rq *rq)
1375 {
1376         struct task_struct *next_task;
1377         struct rq *lowest_rq;
1378         int ret = 0;
1379
1380         if (!rq->rt.overloaded)
1381                 return 0;
1382
1383         next_task = pick_next_pushable_task(rq);
1384         if (!next_task)
1385                 return 0;
1386
1387 retry:
1388         if (unlikely(next_task == rq->curr)) {
1389                 WARN_ON(1);
1390                 return 0;
1391         }
1392
1393         /*
1394          * It's possible that the next_task slipped in of
1395          * higher priority than current. If that's the case
1396          * just reschedule current.
1397          */
1398         if (unlikely(next_task->prio < rq->curr->prio)) {
1399                 resched_task(rq->curr);
1400                 return 0;
1401         }
1402
1403         /* We might release rq lock */
1404         get_task_struct(next_task);
1405
1406         /* find_lock_lowest_rq locks the rq if found */
1407         lowest_rq = find_lock_lowest_rq(next_task, rq);
1408         if (!lowest_rq) {
1409                 struct task_struct *task;
1410                 /*
1411                  * find_lock_lowest_rq releases rq->lock
1412                  * so it is possible that next_task has migrated.
1413                  *
1414                  * We need to make sure that the task is still on the same
1415                  * run-queue and is also still the next task eligible for
1416                  * pushing.
1417                  */
1418                 task = pick_next_pushable_task(rq);
1419                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1420                         /*
1421                          * The task hasn't migrated, and is still the next
1422                          * eligible task, but we failed to find a run-queue
1423                          * to push it to.  Do not retry in this case, since
1424                          * other cpus will pull from us when ready.
1425                          */
1426                         goto out;
1427                 }
1428
1429                 if (!task)
1430                         /* No more tasks, just exit */
1431                         goto out;
1432
1433                 /*
1434                  * Something has shifted, try again.
1435                  */
1436                 put_task_struct(next_task);
1437                 next_task = task;
1438                 goto retry;
1439         }
1440
1441         deactivate_task(rq, next_task, 0);
1442         set_task_cpu(next_task, lowest_rq->cpu);
1443         activate_task(lowest_rq, next_task, 0);
1444         ret = 1;
1445
1446         resched_task(lowest_rq->curr);
1447
1448         double_unlock_balance(rq, lowest_rq);
1449
1450 out:
1451         put_task_struct(next_task);
1452
1453         return ret;
1454 }
1455
1456 static void push_rt_tasks(struct rq *rq)
1457 {
1458         /* push_rt_task will return true if it moved an RT */
1459         while (push_rt_task(rq))
1460                 ;
1461 }
1462
1463 static int pull_rt_task(struct rq *this_rq)
1464 {
1465         int this_cpu = this_rq->cpu, ret = 0, cpu;
1466         struct task_struct *p;
1467         struct rq *src_rq;
1468
1469         if (likely(!rt_overloaded(this_rq)))
1470                 return 0;
1471
1472         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1473                 if (this_cpu == cpu)
1474                         continue;
1475
1476                 src_rq = cpu_rq(cpu);
1477
1478                 /*
1479                  * Don't bother taking the src_rq->lock if the next highest
1480                  * task is known to be lower-priority than our current task.
1481                  * This may look racy, but if this value is about to go
1482                  * logically higher, the src_rq will push this task away.
1483                  * And if its going logically lower, we do not care
1484                  */
1485                 if (src_rq->rt.highest_prio.next >=
1486                     this_rq->rt.highest_prio.curr)
1487                         continue;
1488
1489                 /*
1490                  * We can potentially drop this_rq's lock in
1491                  * double_lock_balance, and another CPU could
1492                  * alter this_rq
1493                  */
1494                 double_lock_balance(this_rq, src_rq);
1495
1496                 /*
1497                  * Are there still pullable RT tasks?
1498                  */
1499                 if (src_rq->rt.rt_nr_running <= 1)
1500                         goto skip;
1501
1502                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1503
1504                 /*
1505                  * Do we have an RT task that preempts
1506                  * the to-be-scheduled task?
1507                  */
1508                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1509                         WARN_ON(p == src_rq->curr);
1510                         WARN_ON(!p->on_rq);
1511
1512                         /*
1513                          * There's a chance that p is higher in priority
1514                          * than what's currently running on its cpu.
1515                          * This is just that p is wakeing up and hasn't
1516                          * had a chance to schedule. We only pull
1517                          * p if it is lower in priority than the
1518                          * current task on the run queue
1519                          */
1520                         if (p->prio < src_rq->curr->prio)
1521                                 goto skip;
1522
1523                         ret = 1;
1524
1525                         deactivate_task(src_rq, p, 0);
1526                         set_task_cpu(p, this_cpu);
1527                         activate_task(this_rq, p, 0);
1528                         /*
1529                          * We continue with the search, just in
1530                          * case there's an even higher prio task
1531                          * in another runqueue. (low likelihood
1532                          * but possible)
1533                          */
1534                 }
1535 skip:
1536                 double_unlock_balance(this_rq, src_rq);
1537         }
1538
1539         return ret;
1540 }
1541
1542 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1543 {
1544         /* Try to pull RT tasks here if we lower this rq's prio */
1545         if (rq->rt.highest_prio.curr > prev->prio)
1546                 pull_rt_task(rq);
1547 }
1548
1549 static void post_schedule_rt(struct rq *rq)
1550 {
1551         push_rt_tasks(rq);
1552 }
1553
1554 /*
1555  * If we are not running and we are not going to reschedule soon, we should
1556  * try to push tasks away now
1557  */
1558 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1559 {
1560         if (!task_running(rq, p) &&
1561             !test_tsk_need_resched(rq->curr) &&
1562             has_pushable_tasks(rq) &&
1563             p->rt.nr_cpus_allowed > 1 &&
1564             rt_task(rq->curr) &&
1565             (rq->curr->rt.nr_cpus_allowed < 2 ||
1566              rq->curr->prio < p->prio))
1567                 push_rt_tasks(rq);
1568 }
1569
1570 static void set_cpus_allowed_rt(struct task_struct *p,
1571                                 const struct cpumask *new_mask)
1572 {
1573         int weight = cpumask_weight(new_mask);
1574
1575         BUG_ON(!rt_task(p));
1576
1577         /*
1578          * Update the migration status of the RQ if we have an RT task
1579          * which is running AND changing its weight value.
1580          */
1581         if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1582                 struct rq *rq = task_rq(p);
1583
1584                 if (!task_current(rq, p)) {
1585                         /*
1586                          * Make sure we dequeue this task from the pushable list
1587                          * before going further.  It will either remain off of
1588                          * the list because we are no longer pushable, or it
1589                          * will be requeued.
1590                          */
1591                         if (p->rt.nr_cpus_allowed > 1)
1592                                 dequeue_pushable_task(rq, p);
1593
1594                         /*
1595                          * Requeue if our weight is changing and still > 1
1596                          */
1597                         if (weight > 1)
1598                                 enqueue_pushable_task(rq, p);
1599
1600                 }
1601
1602                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1603                         rq->rt.rt_nr_migratory++;
1604                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1605                         BUG_ON(!rq->rt.rt_nr_migratory);
1606                         rq->rt.rt_nr_migratory--;
1607                 }
1608
1609                 update_rt_migration(&rq->rt);
1610         }
1611
1612         cpumask_copy(&p->cpus_allowed, new_mask);
1613         p->rt.nr_cpus_allowed = weight;
1614 }
1615
1616 /* Assumes rq->lock is held */
1617 static void rq_online_rt(struct rq *rq)
1618 {
1619         if (rq->rt.overloaded)
1620                 rt_set_overload(rq);
1621
1622         __enable_runtime(rq);
1623
1624         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1625 }
1626
1627 /* Assumes rq->lock is held */
1628 static void rq_offline_rt(struct rq *rq)
1629 {
1630         if (rq->rt.overloaded)
1631                 rt_clear_overload(rq);
1632
1633         __disable_runtime(rq);
1634
1635         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1636 }
1637
1638 /*
1639  * When switch from the rt queue, we bring ourselves to a position
1640  * that we might want to pull RT tasks from other runqueues.
1641  */
1642 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1643 {
1644         /*
1645          * If there are other RT tasks then we will reschedule
1646          * and the scheduling of the other RT tasks will handle
1647          * the balancing. But if we are the last RT task
1648          * we may need to handle the pulling of RT tasks
1649          * now.
1650          */
1651         if (p->on_rq && !rq->rt.rt_nr_running)
1652                 pull_rt_task(rq);
1653 }
1654
1655 static inline void init_sched_rt_class(void)
1656 {
1657         unsigned int i;
1658
1659         for_each_possible_cpu(i)
1660                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1661                                         GFP_KERNEL, cpu_to_node(i));
1662 }
1663 #endif /* CONFIG_SMP */
1664
1665 /*
1666  * When switching a task to RT, we may overload the runqueue
1667  * with RT tasks. In this case we try to push them off to
1668  * other runqueues.
1669  */
1670 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1671 {
1672         int check_resched = 1;
1673
1674         /*
1675          * If we are already running, then there's nothing
1676          * that needs to be done. But if we are not running
1677          * we may need to preempt the current running task.
1678          * If that current running task is also an RT task
1679          * then see if we can move to another run queue.
1680          */
1681         if (p->on_rq && rq->curr != p) {
1682 #ifdef CONFIG_SMP
1683                 if (rq->rt.overloaded && push_rt_task(rq) &&
1684                     /* Don't resched if we changed runqueues */
1685                     rq != task_rq(p))
1686                         check_resched = 0;
1687 #endif /* CONFIG_SMP */
1688                 if (check_resched && p->prio < rq->curr->prio)
1689                         resched_task(rq->curr);
1690         }
1691 }
1692
1693 /*
1694  * Priority of the task has changed. This may cause
1695  * us to initiate a push or pull.
1696  */
1697 static void
1698 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1699 {
1700         if (!p->on_rq)
1701                 return;
1702
1703         if (rq->curr == p) {
1704 #ifdef CONFIG_SMP
1705                 /*
1706                  * If our priority decreases while running, we
1707                  * may need to pull tasks to this runqueue.
1708                  */
1709                 if (oldprio < p->prio)
1710                         pull_rt_task(rq);
1711                 /*
1712                  * If there's a higher priority task waiting to run
1713                  * then reschedule. Note, the above pull_rt_task
1714                  * can release the rq lock and p could migrate.
1715                  * Only reschedule if p is still on the same runqueue.
1716                  */
1717                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1718                         resched_task(p);
1719 #else
1720                 /* For UP simply resched on drop of prio */
1721                 if (oldprio < p->prio)
1722                         resched_task(p);
1723 #endif /* CONFIG_SMP */
1724         } else {
1725                 /*
1726                  * This task is not running, but if it is
1727                  * greater than the current running task
1728                  * then reschedule.
1729                  */
1730                 if (p->prio < rq->curr->prio)
1731                         resched_task(rq->curr);
1732         }
1733 }
1734
1735 static void watchdog(struct rq *rq, struct task_struct *p)
1736 {
1737         unsigned long soft, hard;
1738
1739         /* max may change after cur was read, this will be fixed next tick */
1740         soft = task_rlimit(p, RLIMIT_RTTIME);
1741         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1742
1743         if (soft != RLIM_INFINITY) {
1744                 unsigned long next;
1745
1746                 p->rt.timeout++;
1747                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1748                 if (p->rt.timeout > next)
1749                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1750         }
1751 }
1752
1753 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1754 {
1755         update_curr_rt(rq);
1756
1757         watchdog(rq, p);
1758
1759         /*
1760          * RR tasks need a special form of timeslice management.
1761          * FIFO tasks have no timeslices.
1762          */
1763         if (p->policy != SCHED_RR)
1764                 return;
1765
1766         if (--p->rt.time_slice)
1767                 return;
1768
1769         p->rt.time_slice = DEF_TIMESLICE;
1770
1771         /*
1772          * Requeue to the end of queue if we are not the only element
1773          * on the queue:
1774          */
1775         if (p->rt.run_list.prev != p->rt.run_list.next) {
1776                 requeue_task_rt(rq, p, 0);
1777                 set_tsk_need_resched(p);
1778         }
1779 }
1780
1781 static void set_curr_task_rt(struct rq *rq)
1782 {
1783         struct task_struct *p = rq->curr;
1784
1785         p->se.exec_start = rq->clock_task;
1786
1787         /* The running task is never eligible for pushing */
1788         dequeue_pushable_task(rq, p);
1789 }
1790
1791 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1792 {
1793         /*
1794          * Time slice is 0 for SCHED_FIFO tasks
1795          */
1796         if (task->policy == SCHED_RR)
1797                 return DEF_TIMESLICE;
1798         else
1799                 return 0;
1800 }
1801
1802 static const struct sched_class rt_sched_class = {
1803         .next                   = &fair_sched_class,
1804         .enqueue_task           = enqueue_task_rt,
1805         .dequeue_task           = dequeue_task_rt,
1806         .yield_task             = yield_task_rt,
1807
1808         .check_preempt_curr     = check_preempt_curr_rt,
1809
1810         .pick_next_task         = pick_next_task_rt,
1811         .put_prev_task          = put_prev_task_rt,
1812
1813 #ifdef CONFIG_SMP
1814         .select_task_rq         = select_task_rq_rt,
1815
1816         .set_cpus_allowed       = set_cpus_allowed_rt,
1817         .rq_online              = rq_online_rt,
1818         .rq_offline             = rq_offline_rt,
1819         .pre_schedule           = pre_schedule_rt,
1820         .post_schedule          = post_schedule_rt,
1821         .task_woken             = task_woken_rt,
1822         .switched_from          = switched_from_rt,
1823 #endif
1824
1825         .set_curr_task          = set_curr_task_rt,
1826         .task_tick              = task_tick_rt,
1827
1828         .get_rr_interval        = get_rr_interval_rt,
1829
1830         .prio_changed           = prio_changed_rt,
1831         .switched_to            = switched_to_rt,
1832 };
1833
1834 #ifdef CONFIG_SCHED_DEBUG
1835 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1836
1837 static void print_rt_stats(struct seq_file *m, int cpu)
1838 {
1839         rt_rq_iter_t iter;
1840         struct rt_rq *rt_rq;
1841
1842         rcu_read_lock();
1843         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
1844                 print_rt_rq(m, cpu, rt_rq);
1845         rcu_read_unlock();
1846 }
1847 #endif /* CONFIG_SCHED_DEBUG */