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                 rt_rq->rt_throttled = 0;
 513                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 514                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 515         }
 516 }
 517
 518 static void disable_runtime(struct rq *rq)
 519 {
 520         unsigned long flags;
 521
 522         raw_spin_lock_irqsave(&rq->lock, flags);
 523         __disable_runtime(rq);
 524         raw_spin_unlock_irqrestore(&rq->lock, flags);
 525 }
 526
 527 static void __enable_runtime(struct rq *rq)
 528 {
 529         rt_rq_iter_t iter;
 530         struct rt_rq *rt_rq;
 531
 532         if (unlikely(!scheduler_running))
 533                 return;
 534
 535         /*
 536          * Reset each runqueue's bandwidth settings
 537          */
 538         for_each_rt_rq(rt_rq, iter, rq) {
 539                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 540
 541                 raw_spin_lock(&rt_b->rt_runtime_lock);
 542                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 543                 rt_rq->rt_runtime = rt_b->rt_runtime;
 544                 rt_rq->rt_time = 0;
 545                 rt_rq->rt_throttled = 0;
 546                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 547                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 548         }
 549 }
 550
 551 static void enable_runtime(struct rq *rq)
 552 {
 553         unsigned long flags;
 554
 555         raw_spin_lock_irqsave(&rq->lock, flags);
 556         __enable_runtime(rq);
 557         raw_spin_unlock_irqrestore(&rq->lock, flags);
 558 }
 559
 560 static int balance_runtime(struct rt_rq *rt_rq)
 561 {
 562         int more = 0;
 563
 564         if (!sched_feat(RT_RUNTIME_SHARE))
 565                 return more;
 566
 567         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 568                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 569                 more = do_balance_runtime(rt_rq);
 570                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 571         }
 572
 573         return more;
 574 }
 575 #else /* !CONFIG_SMP */
 576 static inline int balance_runtime(struct rt_rq *rt_rq)
 577 {
 578         return 0;
 579 }
 580 #endif /* CONFIG_SMP */
 581
 582 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 583 {
 584         int i, idle = 1;
 585         const struct cpumask *span;
 586
 587         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 588                 return 1;
 589
 590         span = sched_rt_period_mask();
 591 #ifdef CONFIG_RT_GROUP_SCHED
 592         /*
 593          * FIXME: isolated CPUs should really leave the root task group,
 594          * whether they are isolcpus or were isolated via cpusets, lest
 595          * the timer run on a CPU which does not service all runqueues,
 596          * potentially leaving other CPUs indefinitely throttled.  If
 597          * isolation is really required, the user will turn the throttle
 598          * off to kill the perturbations it causes anyway.  Meanwhile,
 599          * this maintains functionality for boot and/or troubleshooting.
 600          */
 601         if (rt_b == &root_task_group.rt_bandwidth)
 602                 span = cpu_online_mask;
 603 #endif
 604         for_each_cpu(i, span) {
 605                 int enqueue = 0;
 606                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 607                 struct rq *rq = rq_of_rt_rq(rt_rq);
 608
 609                 raw_spin_lock(&rq->lock);
 610                 if (rt_rq->rt_time) {
 611                         u64 runtime;
 612
 613                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 614                         if (rt_rq->rt_throttled)
 615                                 balance_runtime(rt_rq);
 616                         runtime = rt_rq->rt_runtime;
 617                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 618                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 619                                 rt_rq->rt_throttled = 0;
 620                                 enqueue = 1;
 621
 622                                 /*
 623                                  * Force a clock update if the CPU was idle,
 624                                  * lest wakeup -> unthrottle time accumulate.
 625                                  */
 626                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
 627                                         rq->skip_clock_update = -1;
 628                         }
 629                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 630                                 idle = 0;
 631                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 632                 } else if (rt_rq->rt_nr_running) {
 633                         idle = 0;
 634                         if (!rt_rq_throttled(rt_rq))
 635                                 enqueue = 1;
 636                 }
 637
 638                 if (enqueue)
 639                         sched_rt_rq_enqueue(rt_rq);
 640                 raw_spin_unlock(&rq->lock);
 641         }
 642
 643         return idle;
 644 }
 645
 646 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 647 {
 648 #ifdef CONFIG_RT_GROUP_SCHED
 649         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 650
 651         if (rt_rq)
 652                 return rt_rq->highest_prio.curr;
 653 #endif
 654
 655         return rt_task_of(rt_se)->prio;
 656 }
 657
 658 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 659 {
 660         u64 runtime = sched_rt_runtime(rt_rq);
 661
 662         if (rt_rq->rt_throttled)
 663                 return rt_rq_throttled(rt_rq);
 664
 665         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 666                 return 0;
 667
 668         balance_runtime(rt_rq);
 669         runtime = sched_rt_runtime(rt_rq);
 670         if (runtime == RUNTIME_INF)
 671                 return 0;
 672
 673         if (rt_rq->rt_time > runtime) {
 674                 rt_rq->rt_throttled = 1;
 675                 printk_once(KERN_WARNING "sched: RT throttling activated\n");
 676                 if (rt_rq_throttled(rt_rq)) {
 677                         sched_rt_rq_dequeue(rt_rq);
 678                         return 1;
 679                 }
 680         }
 681
 682         return 0;
 683 }
 684
 685 /*
 686  * Update the current task's runtime statistics. Skip current tasks that
 687  * are not in our scheduling class.
 688  */
 689 static void update_curr_rt(struct rq *rq)
 690 {
 691         struct task_struct *curr = rq->curr;
 692         struct sched_rt_entity *rt_se = &curr->rt;
 693         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 694         u64 delta_exec;
 695
 696         if (curr->sched_class != &rt_sched_class)
 697                 return;
 698
 699         delta_exec = rq->clock_task - curr->se.exec_start;
 700         if (unlikely((s64)delta_exec < 0))
 701                 delta_exec = 0;
 702
 703         schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
 704
 705         curr->se.sum_exec_runtime += delta_exec;
 706         account_group_exec_runtime(curr, delta_exec);
 707
 708         curr->se.exec_start = rq->clock_task;
 709         cpuacct_charge(curr, delta_exec);
 710
 711         sched_rt_avg_update(rq, delta_exec);
 712
 713         if (!rt_bandwidth_enabled())
 714                 return;
 715
 716         for_each_sched_rt_entity(rt_se) {
 717                 rt_rq = rt_rq_of_se(rt_se);
 718
 719                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 720                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 721                         rt_rq->rt_time += delta_exec;
 722                         if (sched_rt_runtime_exceeded(rt_rq))
 723                                 resched_task(curr);
 724                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 725                 }
 726         }
 727 }
 728
 729 #if defined CONFIG_SMP
 730
 731 static void
 732 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 733 {
 734         struct rq *rq = rq_of_rt_rq(rt_rq);
 735
 736 #ifdef CONFIG_RT_GROUP_SCHED
 737         /*
 738          * Change rq's cpupri only if rt_rq is the top queue.
 739          */
 740         if (&rq->rt != rt_rq)
 741                 return;
 742 #endif
 743         if (rq->online && prio < prev_prio)
 744                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
 745 }
 746
 747 static void
 748 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 749 {
 750         struct rq *rq = rq_of_rt_rq(rt_rq);
 751
 752 #ifdef CONFIG_RT_GROUP_SCHED
 753         /*
 754          * Change rq's cpupri only if rt_rq is the top queue.
 755          */
 756         if (&rq->rt != rt_rq)
 757                 return;
 758 #endif
 759         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
 760                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
 761 }
 762
 763 #else /* CONFIG_SMP */
 764
 765 static inline
 766 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 767 static inline
 768 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 769
 770 #endif /* CONFIG_SMP */
 771
 772 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 773 static void
 774 inc_rt_prio(struct rt_rq *rt_rq, int prio)
 775 {
 776         int prev_prio = rt_rq->highest_prio.curr;
 777
 778         if (prio < prev_prio)
 779                 rt_rq->highest_prio.curr = prio;
 780
 781         inc_rt_prio_smp(rt_rq, prio, prev_prio);
 782 }
 783
 784 static void
 785 dec_rt_prio(struct rt_rq *rt_rq, int prio)
 786 {
 787         int prev_prio = rt_rq->highest_prio.curr;
 788
 789         if (rt_rq->rt_nr_running) {
 790
 791                 WARN_ON(prio < prev_prio);
 792
 793                 /*
 794                  * This may have been our highest task, and therefore
 795                  * we may have some recomputation to do
 796                  */
 797                 if (prio == prev_prio) {
 798                         struct rt_prio_array *array = &rt_rq->active;
 799
 800                         rt_rq->highest_prio.curr =
 801                                 sched_find_first_bit(array->bitmap);
 802                 }
 803
 804         } else
 805                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
 806
 807         dec_rt_prio_smp(rt_rq, prio, prev_prio);
 808 }
 809
 810 #else
 811
 812 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
 813 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
 814
 815 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
 816
 817 #ifdef CONFIG_RT_GROUP_SCHED
 818
 819 static void
 820 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 821 {
 822         if (rt_se_boosted(rt_se))
 823                 rt_rq->rt_nr_boosted++;
 824
 825         if (rt_rq->tg)
 826                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 827 }
 828
 829 static void
 830 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 831 {
 832         if (rt_se_boosted(rt_se))
 833                 rt_rq->rt_nr_boosted--;
 834
 835         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 836 }
 837
 838 #else /* CONFIG_RT_GROUP_SCHED */
 839
 840 static void
 841 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 842 {
 843         start_rt_bandwidth(&def_rt_bandwidth);
 844 }
 845
 846 static inline
 847 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
 848
 849 #endif /* CONFIG_RT_GROUP_SCHED */
 850
 851 static inline
 852 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 853 {
 854         int prio = rt_se_prio(rt_se);
 855
 856         WARN_ON(!rt_prio(prio));
 857         rt_rq->rt_nr_running++;
 858
 859         inc_rt_prio(rt_rq, prio);
 860         inc_rt_migration(rt_se, rt_rq);
 861         inc_rt_group(rt_se, rt_rq);
 862 }
 863
 864 static inline
 865 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 866 {
 867         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 868         WARN_ON(!rt_rq->rt_nr_running);
 869         rt_rq->rt_nr_running--;
 870
 871         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
 872         dec_rt_migration(rt_se, rt_rq);
 873         dec_rt_group(rt_se, rt_rq);
 874 }
 875
 876 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 877 {
 878         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 879         struct rt_prio_array *array = &rt_rq->active;
 880         struct rt_rq *group_rq = group_rt_rq(rt_se);
 881         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 882
 883         /*
 884          * Don't enqueue the group if its throttled, or when empty.
 885          * The latter is a consequence of the former when a child group
 886          * get throttled and the current group doesn't have any other
 887          * active members.
 888          */
 889         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 890                 return;
 891
 892         if (!rt_rq->rt_nr_running)
 893                 list_add_leaf_rt_rq(rt_rq);
 894
 895         if (head)
 896                 list_add(&rt_se->run_list, queue);
 897         else
 898                 list_add_tail(&rt_se->run_list, queue);
 899         __set_bit(rt_se_prio(rt_se), array->bitmap);
 900
 901         inc_rt_tasks(rt_se, rt_rq);
 902 }
 903
 904 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 905 {
 906         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 907         struct rt_prio_array *array = &rt_rq->active;
 908
 909         list_del_init(&rt_se->run_list);
 910         if (list_empty(array->queue + rt_se_prio(rt_se)))
 911                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 912
 913         dec_rt_tasks(rt_se, rt_rq);
 914         if (!rt_rq->rt_nr_running)
 915                 list_del_leaf_rt_rq(rt_rq);
 916 }
 917
 918 /*
 919  * Because the prio of an upper entry depends on the lower
 920  * entries, we must remove entries top - down.
 921  */
 922 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 923 {
 924         struct sched_rt_entity *back = NULL;
 925
 926         for_each_sched_rt_entity(rt_se) {
 927                 rt_se->back = back;
 928                 back = rt_se;
 929         }
 930
 931         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 932                 if (on_rt_rq(rt_se))
 933                         __dequeue_rt_entity(rt_se);
 934         }
 935 }
 936
 937 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 938 {
 939         dequeue_rt_stack(rt_se);
 940         for_each_sched_rt_entity(rt_se)
 941                 __enqueue_rt_entity(rt_se, head);
 942 }
 943
 944 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 945 {
 946         dequeue_rt_stack(rt_se);
 947
 948         for_each_sched_rt_entity(rt_se) {
 949                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 950
 951                 if (rt_rq && rt_rq->rt_nr_running)
 952                         __enqueue_rt_entity(rt_se, false);
 953         }
 954 }
 955
 956 /*
 957  * Adding/removing a task to/from a priority array:
 958  */
 959 static void
 960 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 961 {
 962         struct sched_rt_entity *rt_se = &p->rt;
 963
 964         if (flags & ENQUEUE_WAKEUP)
 965                 rt_se->timeout = 0;
 966
 967         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
 968
 969         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
 970                 enqueue_pushable_task(rq, p);
 971
 972         inc_nr_running(rq);
 973 }
 974
 975 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 976 {
 977         struct sched_rt_entity *rt_se = &p->rt;
 978
 979         update_curr_rt(rq);
 980         dequeue_rt_entity(rt_se);
 981
 982         dequeue_pushable_task(rq, p);
 983
 984         dec_nr_running(rq);
 985 }
 986
 987 /*
 988  * Put task to the end of the run list without the overhead of dequeue
 989  * followed by enqueue.
 990  */
 991 static void
 992 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 993 {
 994         if (on_rt_rq(rt_se)) {
 995                 struct rt_prio_array *array = &rt_rq->active;
 996                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 997
 998                 if (head)
 999                         list_move(&rt_se->run_list, queue);
1000                 else
1001                         list_move_tail(&rt_se->run_list, queue);
1002         }
1003 }
1004
1005 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1006 {
1007         struct sched_rt_entity *rt_se = &p->rt;
1008         struct rt_rq *rt_rq;
1009
1010         for_each_sched_rt_entity(rt_se) {
1011                 rt_rq = rt_rq_of_se(rt_se);
1012                 requeue_rt_entity(rt_rq, rt_se, head);
1013         }
1014 }
1015
1016 static void yield_task_rt(struct rq *rq)
1017 {
1018         requeue_task_rt(rq, rq->curr, 0);
1019 }
1020
1021 #ifdef CONFIG_SMP
1022 static int find_lowest_rq(struct task_struct *task);
1023
1024 static int
1025 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
1026 {
1027         struct task_struct *curr;
1028         struct rq *rq;
1029         int cpu;
1030
1031         cpu = task_cpu(p);
1032
1033         /* For anything but wake ups, just return the task_cpu */
1034         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1035                 goto out;
1036
1037         rq = cpu_rq(cpu);
1038
1039         rcu_read_lock();
1040         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1041
1042         /*
1043          * If the current task on @p's runqueue is an RT task, then
1044          * try to see if we can wake this RT task up on another
1045          * runqueue. Otherwise simply start this RT task
1046          * on its current runqueue.
1047          *
1048          * We want to avoid overloading runqueues. If the woken
1049          * task is a higher priority, then it will stay on this CPU
1050          * and the lower prio task should be moved to another CPU.
1051          * Even though this will probably make the lower prio task
1052          * lose its cache, we do not want to bounce a higher task
1053          * around just because it gave up its CPU, perhaps for a
1054          * lock?
1055          *
1056          * For equal prio tasks, we just let the scheduler sort it out.
1057          *
1058          * Otherwise, just let it ride on the affined RQ and the
1059          * post-schedule router will push the preempted task away
1060          *
1061          * This test is optimistic, if we get it wrong the load-balancer
1062          * will have to sort it out.
1063          */
1064         if (curr && unlikely(rt_task(curr)) &&
1065             (curr->rt.nr_cpus_allowed < 2 ||
1066              curr->prio <= p->prio) &&
1067             (p->rt.nr_cpus_allowed > 1)) {
1068                 int target = find_lowest_rq(p);
1069
1070                 /*
1071                  * Don't bother moving it if the destination CPU is
1072                  * not running a lower priority task.
1073                  */
1074                 if (target != -1 &&
1075                     p->prio < cpu_rq(target)->rt.highest_prio.curr)
1076                         cpu = target;
1077         }
1078         rcu_read_unlock();
1079
1080 out:
1081         return cpu;
1082 }
1083
1084 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1085 {
1086         if (rq->curr->rt.nr_cpus_allowed == 1)
1087                 return;
1088
1089         if (p->rt.nr_cpus_allowed != 1
1090             && cpupri_find(&rq->rd->cpupri, p, NULL))
1091                 return;
1092
1093         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1094                 return;
1095
1096         /*
1097          * There appears to be other cpus that can accept
1098          * current and none to run 'p', so lets reschedule
1099          * to try and push current away:
1100          */
1101         requeue_task_rt(rq, p, 1);
1102         resched_task(rq->curr);
1103 }
1104
1105 #endif /* CONFIG_SMP */
1106
1107 /*
1108  * Preempt the current task with a newly woken task if needed:
1109  */
1110 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1111 {
1112         if (p->prio < rq->curr->prio) {
1113                 resched_task(rq->curr);
1114                 return;
1115         }
1116
1117 #ifdef CONFIG_SMP
1118         /*
1119          * If:
1120          *
1121          * - the newly woken task is of equal priority to the current task
1122          * - the newly woken task is non-migratable while current is migratable
1123          * - current will be preempted on the next reschedule
1124          *
1125          * we should check to see if current can readily move to a different
1126          * cpu.  If so, we will reschedule to allow the push logic to try
1127          * to move current somewhere else, making room for our non-migratable
1128          * task.
1129          */
1130         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1131                 check_preempt_equal_prio(rq, p);
1132 #endif
1133 }
1134
1135 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1136                                                    struct rt_rq *rt_rq)
1137 {
1138         struct rt_prio_array *array = &rt_rq->active;
1139         struct sched_rt_entity *next = NULL;
1140         struct list_head *queue;
1141         int idx;
1142
1143         idx = sched_find_first_bit(array->bitmap);
1144         BUG_ON(idx >= MAX_RT_PRIO);
1145
1146         queue = array->queue + idx;
1147         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1148
1149         return next;
1150 }
1151
1152 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1153 {
1154         struct sched_rt_entity *rt_se;
1155         struct task_struct *p;
1156         struct rt_rq *rt_rq;
1157
1158         rt_rq = &rq->rt;
1159
1160         if (!rt_rq->rt_nr_running)
1161                 return NULL;
1162
1163         if (rt_rq_throttled(rt_rq))
1164                 return NULL;
1165
1166         do {
1167                 rt_se = pick_next_rt_entity(rq, rt_rq);
1168                 BUG_ON(!rt_se);
1169                 rt_rq = group_rt_rq(rt_se);
1170         } while (rt_rq);
1171
1172         p = rt_task_of(rt_se);
1173         p->se.exec_start = rq->clock_task;
1174
1175         return p;
1176 }
1177
1178 static struct task_struct *pick_next_task_rt(struct rq *rq)
1179 {
1180         struct task_struct *p = _pick_next_task_rt(rq);
1181
1182         /* The running task is never eligible for pushing */
1183         if (p)
1184                 dequeue_pushable_task(rq, p);
1185
1186 #ifdef CONFIG_SMP
1187         /*
1188          * We detect this state here so that we can avoid taking the RQ
1189          * lock again later if there is no need to push
1190          */
1191         rq->post_schedule = has_pushable_tasks(rq);
1192 #endif
1193
1194         return p;
1195 }
1196
1197 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1198 {
1199         update_curr_rt(rq);
1200
1201         /*
1202          * The previous task needs to be made eligible for pushing
1203          * if it is still active
1204          */
1205         if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1206                 enqueue_pushable_task(rq, p);
1207 }
1208
1209 #ifdef CONFIG_SMP
1210
1211 /* Only try algorithms three times */
1212 #define RT_MAX_TRIES 3
1213
1214 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1215
1216 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1217 {
1218         if (!task_running(rq, p) &&
1219             (cpu < 0 || cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) &&
1220             (p->rt.nr_cpus_allowed > 1))
1221                 return 1;
1222         return 0;
1223 }
1224
1225 /* Return the second highest RT task, NULL otherwise */
1226 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1227 {
1228         struct task_struct *next = NULL;
1229         struct sched_rt_entity *rt_se;
1230         struct rt_prio_array *array;
1231         struct rt_rq *rt_rq;
1232         int idx;
1233
1234         for_each_leaf_rt_rq(rt_rq, rq) {
1235                 array = &rt_rq->active;
1236                 idx = sched_find_first_bit(array->bitmap);
1237 next_idx:
1238                 if (idx >= MAX_RT_PRIO)
1239                         continue;
1240                 if (next && next->prio < idx)
1241                         continue;
1242                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1243                         struct task_struct *p;
1244
1245                         if (!rt_entity_is_task(rt_se))
1246                                 continue;
1247
1248                         p = rt_task_of(rt_se);
1249                         if (pick_rt_task(rq, p, cpu)) {
1250                                 next = p;
1251                                 break;
1252                         }
1253                 }
1254                 if (!next) {
1255                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1256                         goto next_idx;
1257                 }
1258         }
1259
1260         return next;
1261 }
1262
1263 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1264
1265 static int find_lowest_rq(struct task_struct *task)
1266 {
1267         struct sched_domain *sd;
1268         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1269         int this_cpu = smp_processor_id();
1270         int cpu      = task_cpu(task);
1271
1272         /* Make sure the mask is initialized first */
1273         if (unlikely(!lowest_mask))
1274                 return -1;
1275
1276         if (task->rt.nr_cpus_allowed == 1)
1277                 return -1; /* No other targets possible */
1278
1279         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1280                 return -1; /* No targets found */
1281
1282         /*
1283          * At this point we have built a mask of cpus representing the
1284          * lowest priority tasks in the system.  Now we want to elect
1285          * the best one based on our affinity and topology.
1286          *
1287          * We prioritize the last cpu that the task executed on since
1288          * it is most likely cache-hot in that location.
1289          */
1290         if (cpumask_test_cpu(cpu, lowest_mask))
1291                 return cpu;
1292
1293         /*
1294          * Otherwise, we consult the sched_domains span maps to figure
1295          * out which cpu is logically closest to our hot cache data.
1296          */
1297         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1298                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1299
1300         rcu_read_lock();
1301         for_each_domain(cpu, sd) {
1302                 if (sd->flags & SD_WAKE_AFFINE) {
1303                         int best_cpu;
1304
1305                         /*
1306                          * "this_cpu" is cheaper to preempt than a
1307                          * remote processor.
1308                          */
1309                         if (this_cpu != -1 &&
1310                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1311                                 rcu_read_unlock();
1312                                 return this_cpu;
1313                         }
1314
1315                         best_cpu = cpumask_first_and(lowest_mask,
1316                                                      sched_domain_span(sd));
1317                         if (best_cpu < nr_cpu_ids) {
1318                                 rcu_read_unlock();
1319                                 return best_cpu;
1320                         }
1321                 }
1322         }
1323         rcu_read_unlock();
1324
1325         /*
1326          * And finally, if there were no matches within the domains
1327          * just give the caller *something* to work with from the compatible
1328          * locations.
1329          */
1330         if (this_cpu != -1)
1331                 return this_cpu;
1332
1333         cpu = cpumask_any(lowest_mask);
1334         if (cpu < nr_cpu_ids)
1335                 return cpu;
1336         return -1;
1337 }
1338
1339 /* Will lock the rq it finds */
1340 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1341 {
1342         struct rq *lowest_rq = NULL;
1343         int tries;
1344         int cpu;
1345
1346         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1347                 cpu = find_lowest_rq(task);
1348
1349                 if ((cpu == -1) || (cpu == rq->cpu))
1350                         break;
1351
1352                 lowest_rq = cpu_rq(cpu);
1353
1354                 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1355                         /*
1356                          * Target rq has tasks of equal or higher priority,
1357                          * retrying does not release any lock and is unlikely
1358                          * to yield a different result.
1359                          */
1360                         lowest_rq = NULL;
1361                         break;
1362                 }
1363
1364                 /* if the prio of this runqueue changed, try again */
1365                 if (double_lock_balance(rq, lowest_rq)) {
1366                         /*
1367                          * We had to unlock the run queue. In
1368                          * the mean time, task could have
1369                          * migrated already or had its affinity changed.
1370                          * Also make sure that it wasn't scheduled on its rq.
1371                          */
1372                         if (unlikely(task_rq(task) != rq ||
1373                                      !cpumask_test_cpu(lowest_rq->cpu,
1374                                                        tsk_cpus_allowed(task)) ||
1375                                      task_running(rq, task) ||
1376                                      !task->on_rq)) {
1377
1378                                 raw_spin_unlock(&lowest_rq->lock);
1379                                 lowest_rq = NULL;
1380                                 break;
1381                         }
1382                 }
1383
1384                 /* If this rq is still suitable use it. */
1385                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1386                         break;
1387
1388                 /* try again */
1389                 double_unlock_balance(rq, lowest_rq);
1390                 lowest_rq = NULL;
1391         }
1392
1393         return lowest_rq;
1394 }
1395
1396 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1397 {
1398         struct task_struct *p;
1399
1400         if (!has_pushable_tasks(rq))
1401                 return NULL;
1402
1403         p = plist_first_entry(&rq->rt.pushable_tasks,
1404                               struct task_struct, pushable_tasks);
1405
1406         BUG_ON(rq->cpu != task_cpu(p));
1407         BUG_ON(task_current(rq, p));
1408         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1409
1410         BUG_ON(!p->on_rq);
1411         BUG_ON(!rt_task(p));
1412
1413         return p;
1414 }
1415
1416 /*
1417  * If the current CPU has more than one RT task, see if the non
1418  * running task can migrate over to a CPU that is running a task
1419  * of lesser priority.
1420  */
1421 static int push_rt_task(struct rq *rq)
1422 {
1423         struct task_struct *next_task;
1424         struct rq *lowest_rq;
1425         int ret = 0;
1426
1427         if (!rq->rt.overloaded)
1428                 return 0;
1429
1430         next_task = pick_next_pushable_task(rq);
1431         if (!next_task)
1432                 return 0;
1433
1434 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1435        if (unlikely(task_running(rq, next_task)))
1436                return 0;
1437 #endif
1438
1439 retry:
1440         if (unlikely(next_task == rq->curr)) {
1441                 WARN_ON(1);
1442                 return 0;
1443         }
1444
1445         /*
1446          * It's possible that the next_task slipped in of
1447          * higher priority than current. If that's the case
1448          * just reschedule current.
1449          */
1450         if (unlikely(next_task->prio < rq->curr->prio)) {
1451                 resched_task(rq->curr);
1452                 return 0;
1453         }
1454
1455         /* We might release rq lock */
1456         get_task_struct(next_task);
1457
1458         /* find_lock_lowest_rq locks the rq if found */
1459         lowest_rq = find_lock_lowest_rq(next_task, rq);
1460         if (!lowest_rq) {
1461                 struct task_struct *task;
1462                 /*
1463                  * find_lock_lowest_rq releases rq->lock
1464                  * so it is possible that next_task has migrated.
1465                  *
1466                  * We need to make sure that the task is still on the same
1467                  * run-queue and is also still the next task eligible for
1468                  * pushing.
1469                  */
1470                 task = pick_next_pushable_task(rq);
1471                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1472                         /*
1473                          * The task hasn't migrated, and is still the next
1474                          * eligible task, but we failed to find a run-queue
1475                          * to push it to.  Do not retry in this case, since
1476                          * other cpus will pull from us when ready.
1477                          */
1478                         goto out;
1479                 }
1480
1481                 if (!task)
1482                         /* No more tasks, just exit */
1483                         goto out;
1484
1485                 /*
1486                  * Something has shifted, try again.
1487                  */
1488                 put_task_struct(next_task);
1489                 next_task = task;
1490                 goto retry;
1491         }
1492
1493         deactivate_task(rq, next_task, 0);
1494         set_task_cpu(next_task, lowest_rq->cpu);
1495         activate_task(lowest_rq, next_task, 0);
1496         ret = 1;
1497
1498         resched_task(lowest_rq->curr);
1499
1500         double_unlock_balance(rq, lowest_rq);
1501
1502 out:
1503         put_task_struct(next_task);
1504
1505         return ret;
1506 }
1507
1508 static void push_rt_tasks(struct rq *rq)
1509 {
1510         /* push_rt_task will return true if it moved an RT */
1511         while (push_rt_task(rq))
1512                 ;
1513 }
1514
1515 static int pull_rt_task(struct rq *this_rq)
1516 {
1517         int this_cpu = this_rq->cpu, ret = 0, cpu;
1518         struct task_struct *p;
1519         struct rq *src_rq;
1520
1521         if (likely(!rt_overloaded(this_rq)))
1522                 return 0;
1523
1524         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1525                 if (this_cpu == cpu)
1526                         continue;
1527
1528                 src_rq = cpu_rq(cpu);
1529
1530                 /*
1531                  * Don't bother taking the src_rq->lock if the next highest
1532                  * task is known to be lower-priority than our current task.
1533                  * This may look racy, but if this value is about to go
1534                  * logically higher, the src_rq will push this task away.
1535                  * And if its going logically lower, we do not care
1536                  */
1537                 if (src_rq->rt.highest_prio.next >=
1538                     this_rq->rt.highest_prio.curr)
1539                         continue;
1540
1541                 /*
1542                  * We can potentially drop this_rq's lock in
1543                  * double_lock_balance, and another CPU could
1544                  * alter this_rq
1545                  */
1546                 double_lock_balance(this_rq, src_rq);
1547
1548                 /*
1549                  * Are there still pullable RT tasks?
1550                  */
1551                 if (src_rq->rt.rt_nr_running <= 1)
1552                         goto skip;
1553
1554                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1555
1556                 /*
1557                  * Do we have an RT task that preempts
1558                  * the to-be-scheduled task?
1559                  */
1560                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1561                         WARN_ON(p == src_rq->curr);
1562                         WARN_ON(!p->on_rq);
1563
1564                         /*
1565                          * There's a chance that p is higher in priority
1566                          * than what's currently running on its cpu.
1567                          * This is just that p is wakeing up and hasn't
1568                          * had a chance to schedule. We only pull
1569                          * p if it is lower in priority than the
1570                          * current task on the run queue
1571                          */
1572                         if (p->prio < src_rq->curr->prio)
1573                                 goto skip;
1574
1575                         ret = 1;
1576
1577                         deactivate_task(src_rq, p, 0);
1578                         set_task_cpu(p, this_cpu);
1579                         activate_task(this_rq, p, 0);
1580                         /*
1581                          * We continue with the search, just in
1582                          * case there's an even higher prio task
1583                          * in another runqueue. (low likelihood
1584                          * but possible)
1585                          */
1586                 }
1587 skip:
1588                 double_unlock_balance(this_rq, src_rq);
1589         }
1590
1591         return ret;
1592 }
1593
1594 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1595 {
1596         /* Try to pull RT tasks here if we lower this rq's prio */
1597         if (rq->rt.highest_prio.curr > prev->prio)
1598                 pull_rt_task(rq);
1599 }
1600
1601 static void post_schedule_rt(struct rq *rq)
1602 {
1603         push_rt_tasks(rq);
1604 }
1605
1606 /*
1607  * If we are not running and we are not going to reschedule soon, we should
1608  * try to push tasks away now
1609  */
1610 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1611 {
1612         if (!task_running(rq, p) &&
1613             !test_tsk_need_resched(rq->curr) &&
1614             has_pushable_tasks(rq) &&
1615             p->rt.nr_cpus_allowed > 1 &&
1616             rt_task(rq->curr) &&
1617             (rq->curr->rt.nr_cpus_allowed < 2 ||
1618              rq->curr->prio <= p->prio))
1619                 push_rt_tasks(rq);
1620 }
1621
1622 static void set_cpus_allowed_rt(struct task_struct *p,
1623                                 const struct cpumask *new_mask)
1624 {
1625         int weight = cpumask_weight(new_mask);
1626
1627         BUG_ON(!rt_task(p));
1628
1629         /*
1630          * Update the migration status of the RQ if we have an RT task
1631          * which is running AND changing its weight value.
1632          */
1633         if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1634                 struct rq *rq = task_rq(p);
1635
1636                 if (!task_current(rq, p)) {
1637                         /*
1638                          * Make sure we dequeue this task from the pushable list
1639                          * before going further.  It will either remain off of
1640                          * the list because we are no longer pushable, or it
1641                          * will be requeued.
1642                          */
1643                         if (p->rt.nr_cpus_allowed > 1)
1644                                 dequeue_pushable_task(rq, p);
1645
1646                         /*
1647                          * Requeue if our weight is changing and still > 1
1648                          */
1649                         if (weight > 1)
1650                                 enqueue_pushable_task(rq, p);
1651
1652                 }
1653
1654                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1655                         rq->rt.rt_nr_migratory++;
1656                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1657                         BUG_ON(!rq->rt.rt_nr_migratory);
1658                         rq->rt.rt_nr_migratory--;
1659                 }
1660
1661                 update_rt_migration(&rq->rt);
1662         }
1663 }
1664
1665 /* Assumes rq->lock is held */
1666 static void rq_online_rt(struct rq *rq)
1667 {
1668         if (rq->rt.overloaded)
1669                 rt_set_overload(rq);
1670
1671         __enable_runtime(rq);
1672
1673         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1674 }
1675
1676 /* Assumes rq->lock is held */
1677 static void rq_offline_rt(struct rq *rq)
1678 {
1679         if (rq->rt.overloaded)
1680                 rt_clear_overload(rq);
1681
1682         __disable_runtime(rq);
1683
1684         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1685 }
1686
1687 /*
1688  * When switch from the rt queue, we bring ourselves to a position
1689  * that we might want to pull RT tasks from other runqueues.
1690  */
1691 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1692 {
1693         /*
1694          * If there are other RT tasks then we will reschedule
1695          * and the scheduling of the other RT tasks will handle
1696          * the balancing. But if we are the last RT task
1697          * we may need to handle the pulling of RT tasks
1698          * now.
1699          */
1700         if (p->on_rq && !rq->rt.rt_nr_running)
1701                 pull_rt_task(rq);
1702 }
1703
1704 static inline void init_sched_rt_class(void)
1705 {
1706         unsigned int i;
1707
1708         for_each_possible_cpu(i)
1709                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1710                                         GFP_KERNEL, cpu_to_node(i));
1711 }
1712 #endif /* CONFIG_SMP */
1713
1714 /*
1715  * When switching a task to RT, we may overload the runqueue
1716  * with RT tasks. In this case we try to push them off to
1717  * other runqueues.
1718  */
1719 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1720 {
1721         int check_resched = 1;
1722
1723         /*
1724          * If we are already running, then there's nothing
1725          * that needs to be done. But if we are not running
1726          * we may need to preempt the current running task.
1727          * If that current running task is also an RT task
1728          * then see if we can move to another run queue.
1729          */
1730         if (p->on_rq && rq->curr != p) {
1731 #ifdef CONFIG_SMP
1732                 if (rq->rt.overloaded && push_rt_task(rq) &&
1733                     /* Don't resched if we changed runqueues */
1734                     rq != task_rq(p))
1735                         check_resched = 0;
1736 #endif /* CONFIG_SMP */
1737                 if (check_resched && p->prio < rq->curr->prio)
1738                         resched_task(rq->curr);
1739         }
1740 }
1741
1742 /*
1743  * Priority of the task has changed. This may cause
1744  * us to initiate a push or pull.
1745  */
1746 static void
1747 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1748 {
1749         if (!p->on_rq)
1750                 return;
1751
1752         if (rq->curr == p) {
1753 #ifdef CONFIG_SMP
1754                 /*
1755                  * If our priority decreases while running, we
1756                  * may need to pull tasks to this runqueue.
1757                  */
1758                 if (oldprio < p->prio)
1759                         pull_rt_task(rq);
1760                 /*
1761                  * If there's a higher priority task waiting to run
1762                  * then reschedule. Note, the above pull_rt_task
1763                  * can release the rq lock and p could migrate.
1764                  * Only reschedule if p is still on the same runqueue.
1765                  */
1766                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1767                         resched_task(p);
1768 #else
1769                 /* For UP simply resched on drop of prio */
1770                 if (oldprio < p->prio)
1771                         resched_task(p);
1772 #endif /* CONFIG_SMP */
1773         } else {
1774                 /*
1775                  * This task is not running, but if it is
1776                  * greater than the current running task
1777                  * then reschedule.
1778                  */
1779                 if (p->prio < rq->curr->prio)
1780                         resched_task(rq->curr);
1781         }
1782 }
1783
1784 static void watchdog(struct rq *rq, struct task_struct *p)
1785 {
1786         unsigned long soft, hard;
1787
1788         /* max may change after cur was read, this will be fixed next tick */
1789         soft = task_rlimit(p, RLIMIT_RTTIME);
1790         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1791
1792         if (soft != RLIM_INFINITY) {
1793                 unsigned long next;
1794
1795                 if (p->rt.watchdog_stamp != jiffies) {
1796                         p->rt.timeout++;
1797                         p->rt.watchdog_stamp = jiffies;
1798                 }
1799
1800                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1801                 if (p->rt.timeout > next)
1802                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1803         }
1804 }
1805
1806 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1807 {
1808         struct sched_rt_entity *rt_se = &p->rt;
1809
1810         update_curr_rt(rq);
1811
1812         watchdog(rq, p);
1813
1814         /*
1815          * RR tasks need a special form of timeslice management.
1816          * FIFO tasks have no timeslices.
1817          */
1818         if (p->policy != SCHED_RR)
1819                 return;
1820
1821         if (--p->rt.time_slice)
1822                 return;
1823
1824         p->rt.time_slice = DEF_TIMESLICE;
1825
1826         /*
1827          * Requeue to the end of queue if we (and all of our ancestors) are the
1828          * only element on the queue
1829          */
1830         for_each_sched_rt_entity(rt_se) {
1831                 if (rt_se->run_list.prev != rt_se->run_list.next) {
1832                         requeue_task_rt(rq, p, 0);
1833                         set_tsk_need_resched(p);
1834                         return;
1835                 }
1836         }
1837 }
1838
1839 static void set_curr_task_rt(struct rq *rq)
1840 {
1841         struct task_struct *p = rq->curr;
1842
1843         p->se.exec_start = rq->clock_task;
1844
1845         /* The running task is never eligible for pushing */
1846         dequeue_pushable_task(rq, p);
1847 }
1848
1849 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1850 {
1851         /*
1852          * Time slice is 0 for SCHED_FIFO tasks
1853          */
1854         if (task->policy == SCHED_RR)
1855                 return DEF_TIMESLICE;
1856         else
1857                 return 0;
1858 }
1859
1860 static const struct sched_class rt_sched_class = {
1861         .next                   = &fair_sched_class,
1862         .enqueue_task           = enqueue_task_rt,
1863         .dequeue_task           = dequeue_task_rt,
1864         .yield_task             = yield_task_rt,
1865
1866         .check_preempt_curr     = check_preempt_curr_rt,
1867
1868         .pick_next_task         = pick_next_task_rt,
1869         .put_prev_task          = put_prev_task_rt,
1870
1871 #ifdef CONFIG_SMP
1872         .select_task_rq         = select_task_rq_rt,
1873
1874         .set_cpus_allowed       = set_cpus_allowed_rt,
1875         .rq_online              = rq_online_rt,
1876         .rq_offline             = rq_offline_rt,
1877         .pre_schedule           = pre_schedule_rt,
1878         .post_schedule          = post_schedule_rt,
1879         .task_woken             = task_woken_rt,
1880         .switched_from          = switched_from_rt,
1881 #endif
1882
1883         .set_curr_task          = set_curr_task_rt,
1884         .task_tick              = task_tick_rt,
1885
1886         .get_rr_interval        = get_rr_interval_rt,
1887
1888         .prio_changed           = prio_changed_rt,
1889         .switched_to            = switched_to_rt,
1890 };
1891
1892 #ifdef CONFIG_SCHED_DEBUG
1893 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1894
1895 static void print_rt_stats(struct seq_file *m, int cpu)
1896 {
1897         rt_rq_iter_t iter;
1898         struct rt_rq *rt_rq;
1899
1900         rcu_read_lock();
1901         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
1902                 print_rt_rq(m, cpu, rt_rq);
1903         rcu_read_unlock();
1904 }
1905 #endif /* CONFIG_SCHED_DEBUG */