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