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 = rq_of_rt_rq(rt_rq)->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 (!sched_feat(RT_RUNTIME_SHARE))
 564                 return more;
 565
 566         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 567                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 568                 more = do_balance_runtime(rt_rq);
 569                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 570         }
 571
 572         return more;
 573 }
 574 #else /* !CONFIG_SMP */
 575 static inline int balance_runtime(struct rt_rq *rt_rq)
 576 {
 577         return 0;
 578 }
 579 #endif /* CONFIG_SMP */
 580
 581 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 582 {
 583         int i, idle = 1;
 584         const struct cpumask *span;
 585
 586         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 587                 return 1;
 588
 589         span = sched_rt_period_mask();
 590         for_each_cpu(i, span) {
 591                 int enqueue = 0;
 592                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 593                 struct rq *rq = rq_of_rt_rq(rt_rq);
 594
 595                 raw_spin_lock(&rq->lock);
 596                 if (rt_rq->rt_time) {
 597                         u64 runtime;
 598
 599                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 600                         if (rt_rq->rt_throttled)
 601                                 balance_runtime(rt_rq);
 602                         runtime = rt_rq->rt_runtime;
 603                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 604                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 605                                 rt_rq->rt_throttled = 0;
 606                                 enqueue = 1;
 607
 608                                 /*
 609                                  * Force a clock update if the CPU was idle,
 610                                  * lest wakeup -> unthrottle time accumulate.
 611                                  */
 612                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
 613                                         rq->skip_clock_update = -1;
 614                         }
 615                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 616                                 idle = 0;
 617                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 618                 } else if (rt_rq->rt_nr_running) {
 619                         idle = 0;
 620                         if (!rt_rq_throttled(rt_rq))
 621                                 enqueue = 1;
 622                 }
 623
 624                 if (enqueue)
 625                         sched_rt_rq_enqueue(rt_rq);
 626                 raw_spin_unlock(&rq->lock);
 627         }
 628
 629         return idle;
 630 }
 631
 632 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 633 {
 634 #ifdef CONFIG_RT_GROUP_SCHED
 635         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 636
 637         if (rt_rq)
 638                 return rt_rq->highest_prio.curr;
 639 #endif
 640
 641         return rt_task_of(rt_se)->prio;
 642 }
 643
 644 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 645 {
 646         u64 runtime = sched_rt_runtime(rt_rq);
 647
 648         if (rt_rq->rt_throttled)
 649                 return rt_rq_throttled(rt_rq);
 650
 651         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 652                 return 0;
 653
 654         balance_runtime(rt_rq);
 655         runtime = sched_rt_runtime(rt_rq);
 656         if (runtime == RUNTIME_INF)
 657                 return 0;
 658
 659         if (rt_rq->rt_time > runtime) {
 660                 rt_rq->rt_throttled = 1;
 661                 printk_once(KERN_WARNING "sched: RT throttling activated\n");
 662                 if (rt_rq_throttled(rt_rq)) {
 663                         sched_rt_rq_dequeue(rt_rq);
 664                         return 1;
 665                 }
 666         }
 667
 668         return 0;
 669 }
 670
 671 /*
 672  * Update the current task's runtime statistics. Skip current tasks that
 673  * are not in our scheduling class.
 674  */
 675 static void update_curr_rt(struct rq *rq)
 676 {
 677         struct task_struct *curr = rq->curr;
 678         struct sched_rt_entity *rt_se = &curr->rt;
 679         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 680         u64 delta_exec;
 681
 682         if (curr->sched_class != &rt_sched_class)
 683                 return;
 684
 685         delta_exec = rq->clock_task - curr->se.exec_start;
 686         if (unlikely((s64)delta_exec < 0))
 687                 delta_exec = 0;
 688
 689         schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
 690
 691         curr->se.sum_exec_runtime += delta_exec;
 692         account_group_exec_runtime(curr, delta_exec);
 693
 694         curr->se.exec_start = rq->clock_task;
 695         cpuacct_charge(curr, delta_exec);
 696
 697         sched_rt_avg_update(rq, delta_exec);
 698
 699         if (!rt_bandwidth_enabled())
 700                 return;
 701
 702         for_each_sched_rt_entity(rt_se) {
 703                 rt_rq = rt_rq_of_se(rt_se);
 704
 705                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 706                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 707                         rt_rq->rt_time += delta_exec;
 708                         if (sched_rt_runtime_exceeded(rt_rq))
 709                                 resched_task(curr);
 710                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 711                 }
 712         }
 713 }
 714
 715 #if defined CONFIG_SMP
 716
 717 static void
 718 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 719 {
 720         struct rq *rq = rq_of_rt_rq(rt_rq);
 721
 722 #ifdef CONFIG_RT_GROUP_SCHED
 723         /*
 724          * Change rq's cpupri only if rt_rq is the top queue.
 725          */
 726         if (&rq->rt != rt_rq)
 727                 return;
 728 #endif
 729         if (rq->online && prio < prev_prio)
 730                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
 731 }
 732
 733 static void
 734 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 735 {
 736         struct rq *rq = rq_of_rt_rq(rt_rq);
 737
 738 #ifdef CONFIG_RT_GROUP_SCHED
 739         /*
 740          * Change rq's cpupri only if rt_rq is the top queue.
 741          */
 742         if (&rq->rt != rt_rq)
 743                 return;
 744 #endif
 745         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
 746                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
 747 }
 748
 749 #else /* CONFIG_SMP */
 750
 751 static inline
 752 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 753 static inline
 754 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 755
 756 #endif /* CONFIG_SMP */
 757
 758 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 759 static void
 760 inc_rt_prio(struct rt_rq *rt_rq, int prio)
 761 {
 762         int prev_prio = rt_rq->highest_prio.curr;
 763
 764         if (prio < prev_prio)
 765                 rt_rq->highest_prio.curr = prio;
 766
 767         inc_rt_prio_smp(rt_rq, prio, prev_prio);
 768 }
 769
 770 static void
 771 dec_rt_prio(struct rt_rq *rt_rq, int prio)
 772 {
 773         int prev_prio = rt_rq->highest_prio.curr;
 774
 775         if (rt_rq->rt_nr_running) {
 776
 777                 WARN_ON(prio < prev_prio);
 778
 779                 /*
 780                  * This may have been our highest task, and therefore
 781                  * we may have some recomputation to do
 782                  */
 783                 if (prio == prev_prio) {
 784                         struct rt_prio_array *array = &rt_rq->active;
 785
 786                         rt_rq->highest_prio.curr =
 787                                 sched_find_first_bit(array->bitmap);
 788                 }
 789
 790         } else
 791                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
 792
 793         dec_rt_prio_smp(rt_rq, prio, prev_prio);
 794 }
 795
 796 #else
 797
 798 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
 799 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
 800
 801 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
 802
 803 #ifdef CONFIG_RT_GROUP_SCHED
 804
 805 static void
 806 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 807 {
 808         if (rt_se_boosted(rt_se))
 809                 rt_rq->rt_nr_boosted++;
 810
 811         if (rt_rq->tg)
 812                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 813 }
 814
 815 static void
 816 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 817 {
 818         if (rt_se_boosted(rt_se))
 819                 rt_rq->rt_nr_boosted--;
 820
 821         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 822 }
 823
 824 #else /* CONFIG_RT_GROUP_SCHED */
 825
 826 static void
 827 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 828 {
 829         start_rt_bandwidth(&def_rt_bandwidth);
 830 }
 831
 832 static inline
 833 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
 834
 835 #endif /* CONFIG_RT_GROUP_SCHED */
 836
 837 static inline
 838 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 839 {
 840         int prio = rt_se_prio(rt_se);
 841
 842         WARN_ON(!rt_prio(prio));
 843         rt_rq->rt_nr_running++;
 844
 845         inc_rt_prio(rt_rq, prio);
 846         inc_rt_migration(rt_se, rt_rq);
 847         inc_rt_group(rt_se, rt_rq);
 848 }
 849
 850 static inline
 851 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 852 {
 853         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 854         WARN_ON(!rt_rq->rt_nr_running);
 855         rt_rq->rt_nr_running--;
 856
 857         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
 858         dec_rt_migration(rt_se, rt_rq);
 859         dec_rt_group(rt_se, rt_rq);
 860 }
 861
 862 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 863 {
 864         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 865         struct rt_prio_array *array = &rt_rq->active;
 866         struct rt_rq *group_rq = group_rt_rq(rt_se);
 867         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 868
 869         /*
 870          * Don't enqueue the group if its throttled, or when empty.
 871          * The latter is a consequence of the former when a child group
 872          * get throttled and the current group doesn't have any other
 873          * active members.
 874          */
 875         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 876                 return;
 877
 878         if (!rt_rq->rt_nr_running)
 879                 list_add_leaf_rt_rq(rt_rq);
 880
 881         if (head)
 882                 list_add(&rt_se->run_list, queue);
 883         else
 884                 list_add_tail(&rt_se->run_list, queue);
 885         __set_bit(rt_se_prio(rt_se), array->bitmap);
 886
 887         inc_rt_tasks(rt_se, rt_rq);
 888 }
 889
 890 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 891 {
 892         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 893         struct rt_prio_array *array = &rt_rq->active;
 894
 895         list_del_init(&rt_se->run_list);
 896         if (list_empty(array->queue + rt_se_prio(rt_se)))
 897                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 898
 899         dec_rt_tasks(rt_se, rt_rq);
 900         if (!rt_rq->rt_nr_running)
 901                 list_del_leaf_rt_rq(rt_rq);
 902 }
 903
 904 /*
 905  * Because the prio of an upper entry depends on the lower
 906  * entries, we must remove entries top - down.
 907  */
 908 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 909 {
 910         struct sched_rt_entity *back = NULL;
 911
 912         for_each_sched_rt_entity(rt_se) {
 913                 rt_se->back = back;
 914                 back = rt_se;
 915         }
 916
 917         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 918                 if (on_rt_rq(rt_se))
 919                         __dequeue_rt_entity(rt_se);
 920         }
 921 }
 922
 923 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 924 {
 925         dequeue_rt_stack(rt_se);
 926         for_each_sched_rt_entity(rt_se)
 927                 __enqueue_rt_entity(rt_se, head);
 928 }
 929
 930 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 931 {
 932         dequeue_rt_stack(rt_se);
 933
 934         for_each_sched_rt_entity(rt_se) {
 935                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 936
 937                 if (rt_rq && rt_rq->rt_nr_running)
 938                         __enqueue_rt_entity(rt_se, false);
 939         }
 940 }
 941
 942 /*
 943  * Adding/removing a task to/from a priority array:
 944  */
 945 static void
 946 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 947 {
 948         struct sched_rt_entity *rt_se = &p->rt;
 949
 950         if (flags & ENQUEUE_WAKEUP)
 951                 rt_se->timeout = 0;
 952
 953         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
 954
 955         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
 956                 enqueue_pushable_task(rq, p);
 957
 958         inc_nr_running(rq);
 959 }
 960
 961 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 962 {
 963         struct sched_rt_entity *rt_se = &p->rt;
 964
 965         update_curr_rt(rq);
 966         dequeue_rt_entity(rt_se);
 967
 968         dequeue_pushable_task(rq, p);
 969
 970         dec_nr_running(rq);
 971 }
 972
 973 /*
 974  * Put task to the end of the run list without the overhead of dequeue
 975  * followed by enqueue.
 976  */
 977 static void
 978 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 979 {
 980         if (on_rt_rq(rt_se)) {
 981                 struct rt_prio_array *array = &rt_rq->active;
 982                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 983
 984                 if (head)
 985                         list_move(&rt_se->run_list, queue);
 986                 else
 987                         list_move_tail(&rt_se->run_list, queue);
 988         }
 989 }
 990
 991 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 992 {
 993         struct sched_rt_entity *rt_se = &p->rt;
 994         struct rt_rq *rt_rq;
 995
 996         for_each_sched_rt_entity(rt_se) {
 997                 rt_rq = rt_rq_of_se(rt_se);
 998                 requeue_rt_entity(rt_rq, rt_se, head);
 999         }
1000 }
1001
1002 static void yield_task_rt(struct rq *rq)
1003 {
1004         requeue_task_rt(rq, rq->curr, 0);
1005 }
1006
1007 #ifdef CONFIG_SMP
1008 static int find_lowest_rq(struct task_struct *task);
1009
1010 static int
1011 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
1012 {
1013         struct task_struct *curr;
1014         struct rq *rq;
1015         int cpu;
1016
1017         cpu = task_cpu(p);
1018
1019         /* For anything but wake ups, just return the task_cpu */
1020         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1021                 goto out;
1022
1023         rq = cpu_rq(cpu);
1024
1025         rcu_read_lock();
1026         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1027
1028         /*
1029          * If the current task on @p's runqueue is an RT task, then
1030          * try to see if we can wake this RT task up on another
1031          * runqueue. Otherwise simply start this RT task
1032          * on its current runqueue.
1033          *
1034          * We want to avoid overloading runqueues. If the woken
1035          * task is a higher priority, then it will stay on this CPU
1036          * and the lower prio task should be moved to another CPU.
1037          * Even though this will probably make the lower prio task
1038          * lose its cache, we do not want to bounce a higher task
1039          * around just because it gave up its CPU, perhaps for a
1040          * lock?
1041          *
1042          * For equal prio tasks, we just let the scheduler sort it out.
1043          *
1044          * Otherwise, just let it ride on the affined RQ and the
1045          * post-schedule router will push the preempted task away
1046          *
1047          * This test is optimistic, if we get it wrong the load-balancer
1048          * will have to sort it out.
1049          */
1050         if (curr && unlikely(rt_task(curr)) &&
1051             (curr->rt.nr_cpus_allowed < 2 ||
1052              curr->prio <= p->prio) &&
1053             (p->rt.nr_cpus_allowed > 1)) {
1054                 int target = find_lowest_rq(p);
1055
1056                 if (target != -1)
1057                         cpu = target;
1058         }
1059         rcu_read_unlock();
1060
1061 out:
1062         return cpu;
1063 }
1064
1065 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1066 {
1067         if (rq->curr->rt.nr_cpus_allowed == 1)
1068                 return;
1069
1070         if (p->rt.nr_cpus_allowed != 1
1071             && cpupri_find(&rq->rd->cpupri, p, NULL))
1072                 return;
1073
1074         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1075                 return;
1076
1077         /*
1078          * There appears to be other cpus that can accept
1079          * current and none to run 'p', so lets reschedule
1080          * to try and push current away:
1081          */
1082         requeue_task_rt(rq, p, 1);
1083         resched_task(rq->curr);
1084 }
1085
1086 #endif /* CONFIG_SMP */
1087
1088 /*
1089  * Preempt the current task with a newly woken task if needed:
1090  */
1091 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1092 {
1093         if (p->prio < rq->curr->prio) {
1094                 resched_task(rq->curr);
1095                 return;
1096         }
1097
1098 #ifdef CONFIG_SMP
1099         /*
1100          * If:
1101          *
1102          * - the newly woken task is of equal priority to the current task
1103          * - the newly woken task is non-migratable while current is migratable
1104          * - current will be preempted on the next reschedule
1105          *
1106          * we should check to see if current can readily move to a different
1107          * cpu.  If so, we will reschedule to allow the push logic to try
1108          * to move current somewhere else, making room for our non-migratable
1109          * task.
1110          */
1111         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1112                 check_preempt_equal_prio(rq, p);
1113 #endif
1114 }
1115
1116 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1117                                                    struct rt_rq *rt_rq)
1118 {
1119         struct rt_prio_array *array = &rt_rq->active;
1120         struct sched_rt_entity *next = NULL;
1121         struct list_head *queue;
1122         int idx;
1123
1124         idx = sched_find_first_bit(array->bitmap);
1125         BUG_ON(idx >= MAX_RT_PRIO);
1126
1127         queue = array->queue + idx;
1128         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1129
1130         return next;
1131 }
1132
1133 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1134 {
1135         struct sched_rt_entity *rt_se;
1136         struct task_struct *p;
1137         struct rt_rq *rt_rq;
1138
1139         rt_rq = &rq->rt;
1140
1141         if (!rt_rq->rt_nr_running)
1142                 return NULL;
1143
1144         if (rt_rq_throttled(rt_rq))
1145                 return NULL;
1146
1147         do {
1148                 rt_se = pick_next_rt_entity(rq, rt_rq);
1149                 BUG_ON(!rt_se);
1150                 rt_rq = group_rt_rq(rt_se);
1151         } while (rt_rq);
1152
1153         p = rt_task_of(rt_se);
1154         p->se.exec_start = rq->clock_task;
1155
1156         return p;
1157 }
1158
1159 static struct task_struct *pick_next_task_rt(struct rq *rq)
1160 {
1161         struct task_struct *p = _pick_next_task_rt(rq);
1162
1163         /* The running task is never eligible for pushing */
1164         if (p)
1165                 dequeue_pushable_task(rq, p);
1166
1167 #ifdef CONFIG_SMP
1168         /*
1169          * We detect this state here so that we can avoid taking the RQ
1170          * lock again later if there is no need to push
1171          */
1172         rq->post_schedule = has_pushable_tasks(rq);
1173 #endif
1174
1175         return p;
1176 }
1177
1178 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1179 {
1180         update_curr_rt(rq);
1181
1182         /*
1183          * The previous task needs to be made eligible for pushing
1184          * if it is still active
1185          */
1186         if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1187                 enqueue_pushable_task(rq, p);
1188 }
1189
1190 #ifdef CONFIG_SMP
1191
1192 /* Only try algorithms three times */
1193 #define RT_MAX_TRIES 3
1194
1195 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1196
1197 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1198 {
1199         if (!task_running(rq, p) &&
1200             (cpu < 0 || cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) &&
1201             (p->rt.nr_cpus_allowed > 1))
1202                 return 1;
1203         return 0;
1204 }
1205
1206 /* Return the second highest RT task, NULL otherwise */
1207 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1208 {
1209         struct task_struct *next = NULL;
1210         struct sched_rt_entity *rt_se;
1211         struct rt_prio_array *array;
1212         struct rt_rq *rt_rq;
1213         int idx;
1214
1215         for_each_leaf_rt_rq(rt_rq, rq) {
1216                 array = &rt_rq->active;
1217                 idx = sched_find_first_bit(array->bitmap);
1218 next_idx:
1219                 if (idx >= MAX_RT_PRIO)
1220                         continue;
1221                 if (next && next->prio < idx)
1222                         continue;
1223                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1224                         struct task_struct *p;
1225
1226                         if (!rt_entity_is_task(rt_se))
1227                                 continue;
1228
1229                         p = rt_task_of(rt_se);
1230                         if (pick_rt_task(rq, p, cpu)) {
1231                                 next = p;
1232                                 break;
1233                         }
1234                 }
1235                 if (!next) {
1236                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1237                         goto next_idx;
1238                 }
1239         }
1240
1241         return next;
1242 }
1243
1244 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1245
1246 static int find_lowest_rq(struct task_struct *task)
1247 {
1248         struct sched_domain *sd;
1249         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1250         int this_cpu = smp_processor_id();
1251         int cpu      = task_cpu(task);
1252
1253         /* Make sure the mask is initialized first */
1254         if (unlikely(!lowest_mask))
1255                 return -1;
1256
1257         if (task->rt.nr_cpus_allowed == 1)
1258                 return -1; /* No other targets possible */
1259
1260         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1261                 return -1; /* No targets found */
1262
1263         /*
1264          * At this point we have built a mask of cpus representing the
1265          * lowest priority tasks in the system.  Now we want to elect
1266          * the best one based on our affinity and topology.
1267          *
1268          * We prioritize the last cpu that the task executed on since
1269          * it is most likely cache-hot in that location.
1270          */
1271         if (cpumask_test_cpu(cpu, lowest_mask))
1272                 return cpu;
1273
1274         /*
1275          * Otherwise, we consult the sched_domains span maps to figure
1276          * out which cpu is logically closest to our hot cache data.
1277          */
1278         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1279                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1280
1281         rcu_read_lock();
1282         for_each_domain(cpu, sd) {
1283                 if (sd->flags & SD_WAKE_AFFINE) {
1284                         int best_cpu;
1285
1286                         /*
1287                          * "this_cpu" is cheaper to preempt than a
1288                          * remote processor.
1289                          */
1290                         if (this_cpu != -1 &&
1291                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1292                                 rcu_read_unlock();
1293                                 return this_cpu;
1294                         }
1295
1296                         best_cpu = cpumask_first_and(lowest_mask,
1297                                                      sched_domain_span(sd));
1298                         if (best_cpu < nr_cpu_ids) {
1299                                 rcu_read_unlock();
1300                                 return best_cpu;
1301                         }
1302                 }
1303         }
1304         rcu_read_unlock();
1305
1306         /*
1307          * And finally, if there were no matches within the domains
1308          * just give the caller *something* to work with from the compatible
1309          * locations.
1310          */
1311         if (this_cpu != -1)
1312                 return this_cpu;
1313
1314         cpu = cpumask_any(lowest_mask);
1315         if (cpu < nr_cpu_ids)
1316                 return cpu;
1317         return -1;
1318 }
1319
1320 /* Will lock the rq it finds */
1321 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1322 {
1323         struct rq *lowest_rq = NULL;
1324         int tries;
1325         int cpu;
1326
1327         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1328                 cpu = find_lowest_rq(task);
1329
1330                 if ((cpu == -1) || (cpu == rq->cpu))
1331                         break;
1332
1333                 lowest_rq = cpu_rq(cpu);
1334
1335                 /* if the prio of this runqueue changed, try again */
1336                 if (double_lock_balance(rq, lowest_rq)) {
1337                         /*
1338                          * We had to unlock the run queue. In
1339                          * the mean time, task could have
1340                          * migrated already or had its affinity changed.
1341                          * Also make sure that it wasn't scheduled on its rq.
1342                          */
1343                         if (unlikely(task_rq(task) != rq ||
1344                                      !cpumask_test_cpu(lowest_rq->cpu,
1345                                                        tsk_cpus_allowed(task)) ||
1346                                      task_running(rq, task) ||
1347                                      !task->on_rq)) {
1348
1349                                 raw_spin_unlock(&lowest_rq->lock);
1350                                 lowest_rq = NULL;
1351                                 break;
1352                         }
1353                 }
1354
1355                 /* If this rq is still suitable use it. */
1356                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1357                         break;
1358
1359                 /* try again */
1360                 double_unlock_balance(rq, lowest_rq);
1361                 lowest_rq = NULL;
1362         }
1363
1364         return lowest_rq;
1365 }
1366
1367 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1368 {
1369         struct task_struct *p;
1370
1371         if (!has_pushable_tasks(rq))
1372                 return NULL;
1373
1374         p = plist_first_entry(&rq->rt.pushable_tasks,
1375                               struct task_struct, pushable_tasks);
1376
1377         BUG_ON(rq->cpu != task_cpu(p));
1378         BUG_ON(task_current(rq, p));
1379         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1380
1381         BUG_ON(!p->on_rq);
1382         BUG_ON(!rt_task(p));
1383
1384         return p;
1385 }
1386
1387 /*
1388  * If the current CPU has more than one RT task, see if the non
1389  * running task can migrate over to a CPU that is running a task
1390  * of lesser priority.
1391  */
1392 static int push_rt_task(struct rq *rq)
1393 {
1394         struct task_struct *next_task;
1395         struct rq *lowest_rq;
1396         int ret = 0;
1397
1398         if (!rq->rt.overloaded)
1399                 return 0;
1400
1401         next_task = pick_next_pushable_task(rq);
1402         if (!next_task)
1403                 return 0;
1404
1405 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1406        if (unlikely(task_running(rq, next_task)))
1407                return 0;
1408 #endif
1409
1410 retry:
1411         if (unlikely(next_task == rq->curr)) {
1412                 WARN_ON(1);
1413                 return 0;
1414         }
1415
1416         /*
1417          * It's possible that the next_task slipped in of
1418          * higher priority than current. If that's the case
1419          * just reschedule current.
1420          */
1421         if (unlikely(next_task->prio < rq->curr->prio)) {
1422                 resched_task(rq->curr);
1423                 return 0;
1424         }
1425
1426         /* We might release rq lock */
1427         get_task_struct(next_task);
1428
1429         /* find_lock_lowest_rq locks the rq if found */
1430         lowest_rq = find_lock_lowest_rq(next_task, rq);
1431         if (!lowest_rq) {
1432                 struct task_struct *task;
1433                 /*
1434                  * find_lock_lowest_rq releases rq->lock
1435                  * so it is possible that next_task has migrated.
1436                  *
1437                  * We need to make sure that the task is still on the same
1438                  * run-queue and is also still the next task eligible for
1439                  * pushing.
1440                  */
1441                 task = pick_next_pushable_task(rq);
1442                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1443                         /*
1444                          * The task hasn't migrated, and is still the next
1445                          * eligible task, but we failed to find a run-queue
1446                          * to push it to.  Do not retry in this case, since
1447                          * other cpus will pull from us when ready.
1448                          */
1449                         goto out;
1450                 }
1451
1452                 if (!task)
1453                         /* No more tasks, just exit */
1454                         goto out;
1455
1456                 /*
1457                  * Something has shifted, try again.
1458                  */
1459                 put_task_struct(next_task);
1460                 next_task = task;
1461                 goto retry;
1462         }
1463
1464         deactivate_task(rq, next_task, 0);
1465         set_task_cpu(next_task, lowest_rq->cpu);
1466         activate_task(lowest_rq, next_task, 0);
1467         ret = 1;
1468
1469         resched_task(lowest_rq->curr);
1470
1471         double_unlock_balance(rq, lowest_rq);
1472
1473 out:
1474         put_task_struct(next_task);
1475
1476         return ret;
1477 }
1478
1479 static void push_rt_tasks(struct rq *rq)
1480 {
1481         /* push_rt_task will return true if it moved an RT */
1482         while (push_rt_task(rq))
1483                 ;
1484 }
1485
1486 static int pull_rt_task(struct rq *this_rq)
1487 {
1488         int this_cpu = this_rq->cpu, ret = 0, cpu;
1489         struct task_struct *p;
1490         struct rq *src_rq;
1491
1492         if (likely(!rt_overloaded(this_rq)))
1493                 return 0;
1494
1495         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1496                 if (this_cpu == cpu)
1497                         continue;
1498
1499                 src_rq = cpu_rq(cpu);
1500
1501                 /*
1502                  * Don't bother taking the src_rq->lock if the next highest
1503                  * task is known to be lower-priority than our current task.
1504                  * This may look racy, but if this value is about to go
1505                  * logically higher, the src_rq will push this task away.
1506                  * And if its going logically lower, we do not care
1507                  */
1508                 if (src_rq->rt.highest_prio.next >=
1509                     this_rq->rt.highest_prio.curr)
1510                         continue;
1511
1512                 /*
1513                  * We can potentially drop this_rq's lock in
1514                  * double_lock_balance, and another CPU could
1515                  * alter this_rq
1516                  */
1517                 double_lock_balance(this_rq, src_rq);
1518
1519                 /*
1520                  * Are there still pullable RT tasks?
1521                  */
1522                 if (src_rq->rt.rt_nr_running <= 1)
1523                         goto skip;
1524
1525                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1526
1527                 /*
1528                  * Do we have an RT task that preempts
1529                  * the to-be-scheduled task?
1530                  */
1531                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1532                         WARN_ON(p == src_rq->curr);
1533                         WARN_ON(!p->on_rq);
1534
1535                         /*
1536                          * There's a chance that p is higher in priority
1537                          * than what's currently running on its cpu.
1538                          * This is just that p is wakeing up and hasn't
1539                          * had a chance to schedule. We only pull
1540                          * p if it is lower in priority than the
1541                          * current task on the run queue
1542                          */
1543                         if (p->prio < src_rq->curr->prio)
1544                                 goto skip;
1545
1546                         ret = 1;
1547
1548                         deactivate_task(src_rq, p, 0);
1549                         set_task_cpu(p, this_cpu);
1550                         activate_task(this_rq, p, 0);
1551                         /*
1552                          * We continue with the search, just in
1553                          * case there's an even higher prio task
1554                          * in another runqueue. (low likelihood
1555                          * but possible)
1556                          */
1557                 }
1558 skip:
1559                 double_unlock_balance(this_rq, src_rq);
1560         }
1561
1562         return ret;
1563 }
1564
1565 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1566 {
1567         /* Try to pull RT tasks here if we lower this rq's prio */
1568         if (rq->rt.highest_prio.curr > prev->prio)
1569                 pull_rt_task(rq);
1570 }
1571
1572 static void post_schedule_rt(struct rq *rq)
1573 {
1574         push_rt_tasks(rq);
1575 }
1576
1577 /*
1578  * If we are not running and we are not going to reschedule soon, we should
1579  * try to push tasks away now
1580  */
1581 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1582 {
1583         if (!task_running(rq, p) &&
1584             !test_tsk_need_resched(rq->curr) &&
1585             has_pushable_tasks(rq) &&
1586             p->rt.nr_cpus_allowed > 1 &&
1587             rt_task(rq->curr) &&
1588             (rq->curr->rt.nr_cpus_allowed < 2 ||
1589              rq->curr->prio <= p->prio))
1590                 push_rt_tasks(rq);
1591 }
1592
1593 static void set_cpus_allowed_rt(struct task_struct *p,
1594                                 const struct cpumask *new_mask)
1595 {
1596         int weight = cpumask_weight(new_mask);
1597
1598         BUG_ON(!rt_task(p));
1599
1600         /*
1601          * Update the migration status of the RQ if we have an RT task
1602          * which is running AND changing its weight value.
1603          */
1604         if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1605                 struct rq *rq = task_rq(p);
1606
1607                 if (!task_current(rq, p)) {
1608                         /*
1609                          * Make sure we dequeue this task from the pushable list
1610                          * before going further.  It will either remain off of
1611                          * the list because we are no longer pushable, or it
1612                          * will be requeued.
1613                          */
1614                         if (p->rt.nr_cpus_allowed > 1)
1615                                 dequeue_pushable_task(rq, p);
1616
1617                         /*
1618                          * Requeue if our weight is changing and still > 1
1619                          */
1620                         if (weight > 1)
1621                                 enqueue_pushable_task(rq, p);
1622
1623                 }
1624
1625                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1626                         rq->rt.rt_nr_migratory++;
1627                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1628                         BUG_ON(!rq->rt.rt_nr_migratory);
1629                         rq->rt.rt_nr_migratory--;
1630                 }
1631
1632                 update_rt_migration(&rq->rt);
1633         }
1634 }
1635
1636 /* Assumes rq->lock is held */
1637 static void rq_online_rt(struct rq *rq)
1638 {
1639         if (rq->rt.overloaded)
1640                 rt_set_overload(rq);
1641
1642         __enable_runtime(rq);
1643
1644         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1645 }
1646
1647 /* Assumes rq->lock is held */
1648 static void rq_offline_rt(struct rq *rq)
1649 {
1650         if (rq->rt.overloaded)
1651                 rt_clear_overload(rq);
1652
1653         __disable_runtime(rq);
1654
1655         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1656 }
1657
1658 /*
1659  * When switch from the rt queue, we bring ourselves to a position
1660  * that we might want to pull RT tasks from other runqueues.
1661  */
1662 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1663 {
1664         /*
1665          * If there are other RT tasks then we will reschedule
1666          * and the scheduling of the other RT tasks will handle
1667          * the balancing. But if we are the last RT task
1668          * we may need to handle the pulling of RT tasks
1669          * now.
1670          */
1671         if (p->on_rq && !rq->rt.rt_nr_running)
1672                 pull_rt_task(rq);
1673 }
1674
1675 static inline void init_sched_rt_class(void)
1676 {
1677         unsigned int i;
1678
1679         for_each_possible_cpu(i)
1680                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1681                                         GFP_KERNEL, cpu_to_node(i));
1682 }
1683 #endif /* CONFIG_SMP */
1684
1685 /*
1686  * When switching a task to RT, we may overload the runqueue
1687  * with RT tasks. In this case we try to push them off to
1688  * other runqueues.
1689  */
1690 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1691 {
1692         int check_resched = 1;
1693
1694         /*
1695          * If we are already running, then there's nothing
1696          * that needs to be done. But if we are not running
1697          * we may need to preempt the current running task.
1698          * If that current running task is also an RT task
1699          * then see if we can move to another run queue.
1700          */
1701         if (p->on_rq && rq->curr != p) {
1702 #ifdef CONFIG_SMP
1703                 if (rq->rt.overloaded && push_rt_task(rq) &&
1704                     /* Don't resched if we changed runqueues */
1705                     rq != task_rq(p))
1706                         check_resched = 0;
1707 #endif /* CONFIG_SMP */
1708                 if (check_resched && p->prio < rq->curr->prio)
1709                         resched_task(rq->curr);
1710         }
1711 }
1712
1713 /*
1714  * Priority of the task has changed. This may cause
1715  * us to initiate a push or pull.
1716  */
1717 static void
1718 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1719 {
1720         if (!p->on_rq)
1721                 return;
1722
1723         if (rq->curr == p) {
1724 #ifdef CONFIG_SMP
1725                 /*
1726                  * If our priority decreases while running, we
1727                  * may need to pull tasks to this runqueue.
1728                  */
1729                 if (oldprio < p->prio)
1730                         pull_rt_task(rq);
1731                 /*
1732                  * If there's a higher priority task waiting to run
1733                  * then reschedule. Note, the above pull_rt_task
1734                  * can release the rq lock and p could migrate.
1735                  * Only reschedule if p is still on the same runqueue.
1736                  */
1737                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1738                         resched_task(p);
1739 #else
1740                 /* For UP simply resched on drop of prio */
1741                 if (oldprio < p->prio)
1742                         resched_task(p);
1743 #endif /* CONFIG_SMP */
1744         } else {
1745                 /*
1746                  * This task is not running, but if it is
1747                  * greater than the current running task
1748                  * then reschedule.
1749                  */
1750                 if (p->prio < rq->curr->prio)
1751                         resched_task(rq->curr);
1752         }
1753 }
1754
1755 static void watchdog(struct rq *rq, struct task_struct *p)
1756 {
1757         unsigned long soft, hard;
1758
1759         /* max may change after cur was read, this will be fixed next tick */
1760         soft = task_rlimit(p, RLIMIT_RTTIME);
1761         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1762
1763         if (soft != RLIM_INFINITY) {
1764                 unsigned long next;
1765
1766                 p->rt.timeout++;
1767                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1768                 if (p->rt.timeout > next)
1769                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1770         }
1771 }
1772
1773 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1774 {
1775         struct sched_rt_entity *rt_se = &p->rt;
1776
1777         update_curr_rt(rq);
1778
1779         watchdog(rq, p);
1780
1781         /*
1782          * RR tasks need a special form of timeslice management.
1783          * FIFO tasks have no timeslices.
1784          */
1785         if (p->policy != SCHED_RR)
1786                 return;
1787
1788         if (--p->rt.time_slice)
1789                 return;
1790
1791         p->rt.time_slice = DEF_TIMESLICE;
1792
1793         /*
1794          * Requeue to the end of queue if we (and all of our ancestors) are the
1795          * only element on the queue
1796          */
1797         for_each_sched_rt_entity(rt_se) {
1798                 if (rt_se->run_list.prev != rt_se->run_list.next) {
1799                         requeue_task_rt(rq, p, 0);
1800                         set_tsk_need_resched(p);
1801                         return;
1802                 }
1803         }
1804 }
1805
1806 static void set_curr_task_rt(struct rq *rq)
1807 {
1808         struct task_struct *p = rq->curr;
1809
1810         p->se.exec_start = rq->clock_task;
1811
1812         /* The running task is never eligible for pushing */
1813         dequeue_pushable_task(rq, p);
1814 }
1815
1816 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1817 {
1818         /*
1819          * Time slice is 0 for SCHED_FIFO tasks
1820          */
1821         if (task->policy == SCHED_RR)
1822                 return DEF_TIMESLICE;
1823         else
1824                 return 0;
1825 }
1826
1827 static const struct sched_class rt_sched_class = {
1828         .next                   = &fair_sched_class,
1829         .enqueue_task           = enqueue_task_rt,
1830         .dequeue_task           = dequeue_task_rt,
1831         .yield_task             = yield_task_rt,
1832
1833         .check_preempt_curr     = check_preempt_curr_rt,
1834
1835         .pick_next_task         = pick_next_task_rt,
1836         .put_prev_task          = put_prev_task_rt,
1837
1838 #ifdef CONFIG_SMP
1839         .select_task_rq         = select_task_rq_rt,
1840
1841         .set_cpus_allowed       = set_cpus_allowed_rt,
1842         .rq_online              = rq_online_rt,
1843         .rq_offline             = rq_offline_rt,
1844         .pre_schedule           = pre_schedule_rt,
1845         .post_schedule          = post_schedule_rt,
1846         .task_woken             = task_woken_rt,
1847         .switched_from          = switched_from_rt,
1848 #endif
1849
1850         .set_curr_task          = set_curr_task_rt,
1851         .task_tick              = task_tick_rt,
1852
1853         .get_rr_interval        = get_rr_interval_rt,
1854
1855         .prio_changed           = prio_changed_rt,
1856         .switched_to            = switched_to_rt,
1857 };
1858
1859 #ifdef CONFIG_SCHED_DEBUG
1860 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1861
1862 static void print_rt_stats(struct seq_file *m, int cpu)
1863 {
1864         rt_rq_iter_t iter;
1865         struct rt_rq *rt_rq;
1866
1867         rcu_read_lock();
1868         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
1869                 print_rt_rq(m, cpu, rt_rq);
1870         rcu_read_unlock();
1871 }
1872 #endif /* CONFIG_SCHED_DEBUG */