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 void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 128 {
 129         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 130         plist_node_init(&p->pushable_tasks, p->prio);
 131         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
 132 }
 133
 134 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 135 {
 136         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 137 }
 138
 139 static inline int has_pushable_tasks(struct rq *rq)
 140 {
 141         return !plist_head_empty(&rq->rt.pushable_tasks);
 142 }
 143
 144 #else
 145
 146 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 147 {
 148 }
 149
 150 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 151 {
 152 }
 153
 154 static inline
 155 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 156 {
 157 }
 158
 159 static inline
 160 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 161 {
 162 }
 163
 164 #endif /* CONFIG_SMP */
 165
 166 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
 167 {
 168         return !list_empty(&rt_se->run_list);
 169 }
 170
 171 #ifdef CONFIG_RT_GROUP_SCHED
 172
 173 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 174 {
 175         if (!rt_rq->tg)
 176                 return RUNTIME_INF;
 177
 178         return rt_rq->rt_runtime;
 179 }
 180
 181 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 182 {
 183         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
 184 }
 185
 186 typedef struct task_group *rt_rq_iter_t;
 187
 188 static inline struct task_group *next_task_group(struct task_group *tg)
 189 {
 190         do {
 191                 tg = list_entry_rcu(tg->list.next,
 192                         typeof(struct task_group), list);
 193         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
 194
 195         if (&tg->list == &task_groups)
 196                 tg = NULL;
 197
 198         return tg;
 199 }
 200
 201 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
 202         for (iter = container_of(&task_groups, typeof(*iter), list);    \
 203                 (iter = next_task_group(iter)) &&                       \
 204                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
 205
 206 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
 207 {
 208         list_add_rcu(&rt_rq->leaf_rt_rq_list,
 209                         &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
 210 }
 211
 212 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
 213 {
 214         list_del_rcu(&rt_rq->leaf_rt_rq_list);
 215 }
 216
 217 #define for_each_leaf_rt_rq(rt_rq, rq) \
 218         list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
 219
 220 #define for_each_sched_rt_entity(rt_se) \
 221         for (; rt_se; rt_se = rt_se->parent)
 222
 223 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 224 {
 225         return rt_se->my_q;
 226 }
 227
 228 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
 229 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
 230
 231 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 232 {
 233         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 234         struct sched_rt_entity *rt_se;
 235
 236         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 237
 238         rt_se = rt_rq->tg->rt_se[cpu];
 239
 240         if (rt_rq->rt_nr_running) {
 241                 if (rt_se && !on_rt_rq(rt_se))
 242                         enqueue_rt_entity(rt_se, false);
 243                 if (rt_rq->highest_prio.curr < curr->prio)
 244                         resched_task(curr);
 245         }
 246 }
 247
 248 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 249 {
 250         struct sched_rt_entity *rt_se;
 251         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 252
 253         rt_se = rt_rq->tg->rt_se[cpu];
 254
 255         if (rt_se && on_rt_rq(rt_se))
 256                 dequeue_rt_entity(rt_se);
 257 }
 258
 259 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 260 {
 261         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 262 }
 263
 264 static int rt_se_boosted(struct sched_rt_entity *rt_se)
 265 {
 266         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 267         struct task_struct *p;
 268
 269         if (rt_rq)
 270                 return !!rt_rq->rt_nr_boosted;
 271
 272         p = rt_task_of(rt_se);
 273         return p->prio != p->normal_prio;
 274 }
 275
 276 #ifdef CONFIG_SMP
 277 static inline const struct cpumask *sched_rt_period_mask(void)
 278 {
 279         return cpu_rq(smp_processor_id())->rd->span;
 280 }
 281 #else
 282 static inline const struct cpumask *sched_rt_period_mask(void)
 283 {
 284         return cpu_online_mask;
 285 }
 286 #endif
 287
 288 static inline
 289 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 290 {
 291         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
 292 }
 293
 294 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 295 {
 296         return &rt_rq->tg->rt_bandwidth;
 297 }
 298
 299 #else /* !CONFIG_RT_GROUP_SCHED */
 300
 301 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 302 {
 303         return rt_rq->rt_runtime;
 304 }
 305
 306 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 307 {
 308         return ktime_to_ns(def_rt_bandwidth.rt_period);
 309 }
 310
 311 typedef struct rt_rq *rt_rq_iter_t;
 312
 313 #define for_each_rt_rq(rt_rq, iter, rq) \
 314         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 315
 316 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
 317 {
 318 }
 319
 320 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
 321 {
 322 }
 323
 324 #define for_each_leaf_rt_rq(rt_rq, rq) \
 325         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 326
 327 #define for_each_sched_rt_entity(rt_se) \
 328         for (; rt_se; rt_se = NULL)
 329
 330 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 331 {
 332         return NULL;
 333 }
 334
 335 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 336 {
 337         if (rt_rq->rt_nr_running)
 338                 resched_task(rq_of_rt_rq(rt_rq)->curr);
 339 }
 340
 341 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 342 {
 343 }
 344
 345 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 346 {
 347         return rt_rq->rt_throttled;
 348 }
 349
 350 static inline const struct cpumask *sched_rt_period_mask(void)
 351 {
 352         return cpu_online_mask;
 353 }
 354
 355 static inline
 356 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 357 {
 358         return &cpu_rq(cpu)->rt;
 359 }
 360
 361 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 362 {
 363         return &def_rt_bandwidth;
 364 }
 365
 366 #endif /* CONFIG_RT_GROUP_SCHED */
 367
 368 #ifdef CONFIG_SMP
 369 /*
 370  * We ran out of runtime, see if we can borrow some from our neighbours.
 371  */
 372 static int do_balance_runtime(struct rt_rq *rt_rq)
 373 {
 374         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 375         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 376         int i, weight, more = 0;
 377         u64 rt_period;
 378
 379         weight = cpumask_weight(rd->span);
 380
 381         raw_spin_lock(&rt_b->rt_runtime_lock);
 382         rt_period = ktime_to_ns(rt_b->rt_period);
 383         for_each_cpu(i, rd->span) {
 384                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 385                 s64 diff;
 386
 387                 if (iter == rt_rq)
 388                         continue;
 389
 390                 raw_spin_lock(&iter->rt_runtime_lock);
 391                 /*
 392                  * Either all rqs have inf runtime and there's nothing to steal
 393                  * or __disable_runtime() below sets a specific rq to inf to
 394                  * indicate its been disabled and disalow stealing.
 395                  */
 396                 if (iter->rt_runtime == RUNTIME_INF)
 397                         goto next;
 398
 399                 /*
 400                  * From runqueues with spare time, take 1/n part of their
 401                  * spare time, but no more than our period.
 402                  */
 403                 diff = iter->rt_runtime - iter->rt_time;
 404                 if (diff > 0) {
 405                         diff = div_u64((u64)diff, weight);
 406                         if (rt_rq->rt_runtime + diff > rt_period)
 407                                 diff = rt_period - rt_rq->rt_runtime;
 408                         iter->rt_runtime -= diff;
 409                         rt_rq->rt_runtime += diff;
 410                         more = 1;
 411                         if (rt_rq->rt_runtime == rt_period) {
 412                                 raw_spin_unlock(&iter->rt_runtime_lock);
 413                                 break;
 414                         }
 415                 }
 416 next:
 417                 raw_spin_unlock(&iter->rt_runtime_lock);
 418         }
 419         raw_spin_unlock(&rt_b->rt_runtime_lock);
 420
 421         return more;
 422 }
 423
 424 /*
 425  * Ensure this RQ takes back all the runtime it lend to its neighbours.
 426  */
 427 static void __disable_runtime(struct rq *rq)
 428 {
 429         struct root_domain *rd = rq->rd;
 430         rt_rq_iter_t iter;
 431         struct rt_rq *rt_rq;
 432
 433         if (unlikely(!scheduler_running))
 434                 return;
 435
 436         for_each_rt_rq(rt_rq, iter, rq) {
 437                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 438                 s64 want;
 439                 int i;
 440
 441                 raw_spin_lock(&rt_b->rt_runtime_lock);
 442                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 443                 /*
 444                  * Either we're all inf and nobody needs to borrow, or we're
 445                  * already disabled and thus have nothing to do, or we have
 446                  * exactly the right amount of runtime to take out.
 447                  */
 448                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 449                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 450                         goto balanced;
 451                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 452
 453                 /*
 454                  * Calculate the difference between what we started out with
 455                  * and what we current have, that's the amount of runtime
 456                  * we lend and now have to reclaim.
 457                  */
 458                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 459
 460                 /*
 461                  * Greedy reclaim, take back as much as we can.
 462                  */
 463                 for_each_cpu(i, rd->span) {
 464                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 465                         s64 diff;
 466
 467                         /*
 468                          * Can't reclaim from ourselves or disabled runqueues.
 469                          */
 470                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 471                                 continue;
 472
 473                         raw_spin_lock(&iter->rt_runtime_lock);
 474                         if (want > 0) {
 475                                 diff = min_t(s64, iter->rt_runtime, want);
 476                                 iter->rt_runtime -= diff;
 477                                 want -= diff;
 478                         } else {
 479                                 iter->rt_runtime -= want;
 480                                 want -= want;
 481                         }
 482                         raw_spin_unlock(&iter->rt_runtime_lock);
 483
 484                         if (!want)
 485                                 break;
 486                 }
 487
 488                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 489                 /*
 490                  * We cannot be left wanting - that would mean some runtime
 491                  * leaked out of the system.
 492                  */
 493                 BUG_ON(want);
 494 balanced:
 495                 /*
 496                  * Disable all the borrow logic by pretending we have inf
 497                  * runtime - in which case borrowing doesn't make sense.
 498                  */
 499                 rt_rq->rt_runtime = RUNTIME_INF;
 500                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 501                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 502         }
 503 }
 504
 505 static void disable_runtime(struct rq *rq)
 506 {
 507         unsigned long flags;
 508
 509         raw_spin_lock_irqsave(&rq->lock, flags);
 510         __disable_runtime(rq);
 511         raw_spin_unlock_irqrestore(&rq->lock, flags);
 512 }
 513
 514 static void __enable_runtime(struct rq *rq)
 515 {
 516         rt_rq_iter_t iter;
 517         struct rt_rq *rt_rq;
 518
 519         if (unlikely(!scheduler_running))
 520                 return;
 521
 522         /*
 523          * Reset each runqueue's bandwidth settings
 524          */
 525         for_each_rt_rq(rt_rq, iter, rq) {
 526                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 527
 528                 raw_spin_lock(&rt_b->rt_runtime_lock);
 529                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 530                 rt_rq->rt_runtime = rt_b->rt_runtime;
 531                 rt_rq->rt_time = 0;
 532                 rt_rq->rt_throttled = 0;
 533                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 534                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 535         }
 536 }
 537
 538 static void enable_runtime(struct rq *rq)
 539 {
 540         unsigned long flags;
 541
 542         raw_spin_lock_irqsave(&rq->lock, flags);
 543         __enable_runtime(rq);
 544         raw_spin_unlock_irqrestore(&rq->lock, flags);
 545 }
 546
 547 static int balance_runtime(struct rt_rq *rt_rq)
 548 {
 549         int more = 0;
 550
 551         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 552                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 553                 more = do_balance_runtime(rt_rq);
 554                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 555         }
 556
 557         return more;
 558 }
 559 #else /* !CONFIG_SMP */
 560 static inline int balance_runtime(struct rt_rq *rt_rq)
 561 {
 562         return 0;
 563 }
 564 #endif /* CONFIG_SMP */
 565
 566 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 567 {
 568         int i, idle = 1;
 569         const struct cpumask *span;
 570
 571         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 572                 return 1;
 573
 574         span = sched_rt_period_mask();
 575         for_each_cpu(i, span) {
 576                 int enqueue = 0;
 577                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 578                 struct rq *rq = rq_of_rt_rq(rt_rq);
 579
 580                 raw_spin_lock(&rq->lock);
 581                 if (rt_rq->rt_time) {
 582                         u64 runtime;
 583
 584                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 585                         if (rt_rq->rt_throttled)
 586                                 balance_runtime(rt_rq);
 587                         runtime = rt_rq->rt_runtime;
 588                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 589                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 590                                 rt_rq->rt_throttled = 0;
 591                                 enqueue = 1;
 592
 593                                 /*
 594                                  * Force a clock update if the CPU was idle,
 595                                  * lest wakeup -> unthrottle time accumulate.
 596                                  */
 597                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
 598                                         rq->skip_clock_update = -1;
 599                         }
 600                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 601                                 idle = 0;
 602                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 603                 } else if (rt_rq->rt_nr_running) {
 604                         idle = 0;
 605                         if (!rt_rq_throttled(rt_rq))
 606                                 enqueue = 1;
 607                 }
 608
 609                 if (enqueue)
 610                         sched_rt_rq_enqueue(rt_rq);
 611                 raw_spin_unlock(&rq->lock);
 612         }
 613
 614         return idle;
 615 }
 616
 617 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 618 {
 619 #ifdef CONFIG_RT_GROUP_SCHED
 620         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 621
 622         if (rt_rq)
 623                 return rt_rq->highest_prio.curr;
 624 #endif
 625
 626         return rt_task_of(rt_se)->prio;
 627 }
 628
 629 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 630 {
 631         u64 runtime = sched_rt_runtime(rt_rq);
 632
 633         if (rt_rq->rt_throttled)
 634                 return rt_rq_throttled(rt_rq);
 635
 636         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 637                 return 0;
 638
 639         balance_runtime(rt_rq);
 640         runtime = sched_rt_runtime(rt_rq);
 641         if (runtime == RUNTIME_INF)
 642                 return 0;
 643
 644         if (rt_rq->rt_time > runtime) {
 645                 rt_rq->rt_throttled = 1;
 646                 if (rt_rq_throttled(rt_rq)) {
 647                         sched_rt_rq_dequeue(rt_rq);
 648                         return 1;
 649                 }
 650         }
 651
 652         return 0;
 653 }
 654
 655 /*
 656  * Update the current task's runtime statistics. Skip current tasks that
 657  * are not in our scheduling class.
 658  */
 659 static void update_curr_rt(struct rq *rq)
 660 {
 661         struct task_struct *curr = rq->curr;
 662         struct sched_rt_entity *rt_se = &curr->rt;
 663         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 664         u64 delta_exec;
 665
 666         if (curr->sched_class != &rt_sched_class)
 667                 return;
 668
 669         delta_exec = rq->clock_task - curr->se.exec_start;
 670         if (unlikely((s64)delta_exec < 0))
 671                 delta_exec = 0;
 672
 673         schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
 674
 675         curr->se.sum_exec_runtime += delta_exec;
 676         account_group_exec_runtime(curr, delta_exec);
 677
 678         curr->se.exec_start = rq->clock_task;
 679         cpuacct_charge(curr, delta_exec);
 680
 681         sched_rt_avg_update(rq, delta_exec);
 682
 683         if (!rt_bandwidth_enabled())
 684                 return;
 685
 686         for_each_sched_rt_entity(rt_se) {
 687                 rt_rq = rt_rq_of_se(rt_se);
 688
 689                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 690                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 691                         rt_rq->rt_time += delta_exec;
 692                         if (sched_rt_runtime_exceeded(rt_rq))
 693                                 resched_task(curr);
 694                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 695                 }
 696         }
 697 }
 698
 699 #if defined CONFIG_SMP
 700
 701 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
 702
 703 static inline int next_prio(struct rq *rq)
 704 {
 705         struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
 706
 707         if (next)
 708                 return next->prio;
 709         else
 710                 return MAX_RT_PRIO;
 711 }
 712
 713 static void
 714 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 715 {
 716         struct rq *rq = rq_of_rt_rq(rt_rq);
 717
 718         if (prio < prev_prio) {
 719
 720                 /*
 721                  * If the new task is higher in priority than anything on the
 722                  * run-queue, we know that the previous high becomes our
 723                  * next-highest.
 724                  */
 725                 rt_rq->highest_prio.next = prev_prio;
 726
 727                 if (rq->online)
 728                         cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
 729
 730         } else if (prio == rt_rq->highest_prio.curr)
 731                 /*
 732                  * If the next task is equal in priority to the highest on
 733                  * the run-queue, then we implicitly know that the next highest
 734                  * task cannot be any lower than current
 735                  */
 736                 rt_rq->highest_prio.next = prio;
 737         else if (prio < rt_rq->highest_prio.next)
 738                 /*
 739                  * Otherwise, we need to recompute next-highest
 740                  */
 741                 rt_rq->highest_prio.next = next_prio(rq);
 742 }
 743
 744 static void
 745 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 746 {
 747         struct rq *rq = rq_of_rt_rq(rt_rq);
 748
 749         if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
 750                 rt_rq->highest_prio.next = next_prio(rq);
 751
 752         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
 753                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
 754 }
 755
 756 #else /* CONFIG_SMP */
 757
 758 static inline
 759 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 760 static inline
 761 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 762
 763 #endif /* CONFIG_SMP */
 764
 765 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 766 static void
 767 inc_rt_prio(struct rt_rq *rt_rq, int prio)
 768 {
 769         int prev_prio = rt_rq->highest_prio.curr;
 770
 771         if (prio < prev_prio)
 772                 rt_rq->highest_prio.curr = prio;
 773
 774         inc_rt_prio_smp(rt_rq, prio, prev_prio);
 775 }
 776
 777 static void
 778 dec_rt_prio(struct rt_rq *rt_rq, int prio)
 779 {
 780         int prev_prio = rt_rq->highest_prio.curr;
 781
 782         if (rt_rq->rt_nr_running) {
 783
 784                 WARN_ON(prio < prev_prio);
 785
 786                 /*
 787                  * This may have been our highest task, and therefore
 788                  * we may have some recomputation to do
 789                  */
 790                 if (prio == prev_prio) {
 791                         struct rt_prio_array *array = &rt_rq->active;
 792
 793                         rt_rq->highest_prio.curr =
 794                                 sched_find_first_bit(array->bitmap);
 795                 }
 796
 797         } else
 798                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
 799
 800         dec_rt_prio_smp(rt_rq, prio, prev_prio);
 801 }
 802
 803 #else
 804
 805 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
 806 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
 807
 808 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
 809
 810 #ifdef CONFIG_RT_GROUP_SCHED
 811
 812 static void
 813 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 814 {
 815         if (rt_se_boosted(rt_se))
 816                 rt_rq->rt_nr_boosted++;
 817
 818         if (rt_rq->tg)
 819                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 820 }
 821
 822 static void
 823 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 824 {
 825         if (rt_se_boosted(rt_se))
 826                 rt_rq->rt_nr_boosted--;
 827
 828         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 829 }
 830
 831 #else /* CONFIG_RT_GROUP_SCHED */
 832
 833 static void
 834 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 835 {
 836         start_rt_bandwidth(&def_rt_bandwidth);
 837 }
 838
 839 static inline
 840 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
 841
 842 #endif /* CONFIG_RT_GROUP_SCHED */
 843
 844 static inline
 845 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 846 {
 847         int prio = rt_se_prio(rt_se);
 848
 849         WARN_ON(!rt_prio(prio));
 850         rt_rq->rt_nr_running++;
 851
 852         inc_rt_prio(rt_rq, prio);
 853         inc_rt_migration(rt_se, rt_rq);
 854         inc_rt_group(rt_se, rt_rq);
 855 }
 856
 857 static inline
 858 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 859 {
 860         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 861         WARN_ON(!rt_rq->rt_nr_running);
 862         rt_rq->rt_nr_running--;
 863
 864         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
 865         dec_rt_migration(rt_se, rt_rq);
 866         dec_rt_group(rt_se, rt_rq);
 867 }
 868
 869 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 870 {
 871         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 872         struct rt_prio_array *array = &rt_rq->active;
 873         struct rt_rq *group_rq = group_rt_rq(rt_se);
 874         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 875
 876         /*
 877          * Don't enqueue the group if its throttled, or when empty.
 878          * The latter is a consequence of the former when a child group
 879          * get throttled and the current group doesn't have any other
 880          * active members.
 881          */
 882         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 883                 return;
 884
 885         if (!rt_rq->rt_nr_running)
 886                 list_add_leaf_rt_rq(rt_rq);
 887
 888         if (head)
 889                 list_add(&rt_se->run_list, queue);
 890         else
 891                 list_add_tail(&rt_se->run_list, queue);
 892         __set_bit(rt_se_prio(rt_se), array->bitmap);
 893
 894         inc_rt_tasks(rt_se, rt_rq);
 895 }
 896
 897 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 898 {
 899         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 900         struct rt_prio_array *array = &rt_rq->active;
 901
 902         list_del_init(&rt_se->run_list);
 903         if (list_empty(array->queue + rt_se_prio(rt_se)))
 904                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 905
 906         dec_rt_tasks(rt_se, rt_rq);
 907         if (!rt_rq->rt_nr_running)
 908                 list_del_leaf_rt_rq(rt_rq);
 909 }
 910
 911 /*
 912  * Because the prio of an upper entry depends on the lower
 913  * entries, we must remove entries top - down.
 914  */
 915 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 916 {
 917         struct sched_rt_entity *back = NULL;
 918
 919         for_each_sched_rt_entity(rt_se) {
 920                 rt_se->back = back;
 921                 back = rt_se;
 922         }
 923
 924         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 925                 if (on_rt_rq(rt_se))
 926                         __dequeue_rt_entity(rt_se);
 927         }
 928 }
 929
 930 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 931 {
 932         dequeue_rt_stack(rt_se);
 933         for_each_sched_rt_entity(rt_se)
 934                 __enqueue_rt_entity(rt_se, head);
 935 }
 936
 937 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 938 {
 939         dequeue_rt_stack(rt_se);
 940
 941         for_each_sched_rt_entity(rt_se) {
 942                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 943
 944                 if (rt_rq && rt_rq->rt_nr_running)
 945                         __enqueue_rt_entity(rt_se, false);
 946         }
 947 }
 948
 949 /*
 950  * Adding/removing a task to/from a priority array:
 951  */
 952 static void
 953 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 954 {
 955         struct sched_rt_entity *rt_se = &p->rt;
 956
 957         if (flags & ENQUEUE_WAKEUP)
 958                 rt_se->timeout = 0;
 959
 960         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
 961
 962         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
 963                 enqueue_pushable_task(rq, p);
 964 }
 965
 966 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 967 {
 968         struct sched_rt_entity *rt_se = &p->rt;
 969
 970         update_curr_rt(rq);
 971         dequeue_rt_entity(rt_se);
 972
 973         dequeue_pushable_task(rq, p);
 974 }
 975
 976 /*
 977  * Put task to the end of the run list without the overhead of dequeue
 978  * followed by enqueue.
 979  */
 980 static void
 981 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 982 {
 983         if (on_rt_rq(rt_se)) {
 984                 struct rt_prio_array *array = &rt_rq->active;
 985                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 986
 987                 if (head)
 988                         list_move(&rt_se->run_list, queue);
 989                 else
 990                         list_move_tail(&rt_se->run_list, queue);
 991         }
 992 }
 993
 994 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 995 {
 996         struct sched_rt_entity *rt_se = &p->rt;
 997         struct rt_rq *rt_rq;
 998
 999         for_each_sched_rt_entity(rt_se) {
1000                 rt_rq = rt_rq_of_se(rt_se);
1001                 requeue_rt_entity(rt_rq, rt_se, head);
1002         }
1003 }
1004
1005 static void yield_task_rt(struct rq *rq)
1006 {
1007         requeue_task_rt(rq, rq->curr, 0);
1008 }
1009
1010 #ifdef CONFIG_SMP
1011 static int find_lowest_rq(struct task_struct *task);
1012
1013 static int
1014 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
1015 {
1016         struct task_struct *curr;
1017         struct rq *rq;
1018         int cpu;
1019
1020         cpu = task_cpu(p);
1021
1022         /* For anything but wake ups, just return the task_cpu */
1023         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1024                 goto out;
1025
1026         rq = cpu_rq(cpu);
1027
1028         rcu_read_lock();
1029         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1030
1031         /*
1032          * If the current task on @p's runqueue is an RT task, then
1033          * try to see if we can wake this RT task up on another
1034          * runqueue. Otherwise simply start this RT task
1035          * on its current runqueue.
1036          *
1037          * We want to avoid overloading runqueues. If the woken
1038          * task is a higher priority, then it will stay on this CPU
1039          * and the lower prio task should be moved to another CPU.
1040          * Even though this will probably make the lower prio task
1041          * lose its cache, we do not want to bounce a higher task
1042          * around just because it gave up its CPU, perhaps for a
1043          * lock?
1044          *
1045          * For equal prio tasks, we just let the scheduler sort it out.
1046          *
1047          * Otherwise, just let it ride on the affined RQ and the
1048          * post-schedule router will push the preempted task away
1049          *
1050          * This test is optimistic, if we get it wrong the load-balancer
1051          * will have to sort it out.
1052          */
1053         if (curr && unlikely(rt_task(curr)) &&
1054             (curr->rt.nr_cpus_allowed < 2 ||
1055              curr->prio < p->prio) &&
1056             (p->rt.nr_cpus_allowed > 1)) {
1057                 int target = find_lowest_rq(p);
1058
1059                 if (target != -1)
1060                         cpu = target;
1061         }
1062         rcu_read_unlock();
1063
1064 out:
1065         return cpu;
1066 }
1067
1068 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1069 {
1070         if (rq->curr->rt.nr_cpus_allowed == 1)
1071                 return;
1072
1073         if (p->rt.nr_cpus_allowed != 1
1074             && cpupri_find(&rq->rd->cpupri, p, NULL))
1075                 return;
1076
1077         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1078                 return;
1079
1080         /*
1081          * There appears to be other cpus that can accept
1082          * current and none to run 'p', so lets reschedule
1083          * to try and push current away:
1084          */
1085         requeue_task_rt(rq, p, 1);
1086         resched_task(rq->curr);
1087 }
1088
1089 #endif /* CONFIG_SMP */
1090
1091 /*
1092  * Preempt the current task with a newly woken task if needed:
1093  */
1094 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1095 {
1096         if (p->prio < rq->curr->prio) {
1097                 resched_task(rq->curr);
1098                 return;
1099         }
1100
1101 #ifdef CONFIG_SMP
1102         /*
1103          * If:
1104          *
1105          * - the newly woken task is of equal priority to the current task
1106          * - the newly woken task is non-migratable while current is migratable
1107          * - current will be preempted on the next reschedule
1108          *
1109          * we should check to see if current can readily move to a different
1110          * cpu.  If so, we will reschedule to allow the push logic to try
1111          * to move current somewhere else, making room for our non-migratable
1112          * task.
1113          */
1114         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1115                 check_preempt_equal_prio(rq, p);
1116 #endif
1117 }
1118
1119 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1120                                                    struct rt_rq *rt_rq)
1121 {
1122         struct rt_prio_array *array = &rt_rq->active;
1123         struct sched_rt_entity *next = NULL;
1124         struct list_head *queue;
1125         int idx;
1126
1127         idx = sched_find_first_bit(array->bitmap);
1128         BUG_ON(idx >= MAX_RT_PRIO);
1129
1130         queue = array->queue + idx;
1131         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1132
1133         return next;
1134 }
1135
1136 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1137 {
1138         struct sched_rt_entity *rt_se;
1139         struct task_struct *p;
1140         struct rt_rq *rt_rq;
1141
1142         rt_rq = &rq->rt;
1143
1144         if (!rt_rq->rt_nr_running)
1145                 return NULL;
1146
1147         if (rt_rq_throttled(rt_rq))
1148                 return NULL;
1149
1150         do {
1151                 rt_se = pick_next_rt_entity(rq, rt_rq);
1152                 BUG_ON(!rt_se);
1153                 rt_rq = group_rt_rq(rt_se);
1154         } while (rt_rq);
1155
1156         p = rt_task_of(rt_se);
1157         p->se.exec_start = rq->clock_task;
1158
1159         return p;
1160 }
1161
1162 static struct task_struct *pick_next_task_rt(struct rq *rq)
1163 {
1164         struct task_struct *p = _pick_next_task_rt(rq);
1165
1166         /* The running task is never eligible for pushing */
1167         if (p)
1168                 dequeue_pushable_task(rq, p);
1169
1170 #ifdef CONFIG_SMP
1171         /*
1172          * We detect this state here so that we can avoid taking the RQ
1173          * lock again later if there is no need to push
1174          */
1175         rq->post_schedule = has_pushable_tasks(rq);
1176 #endif
1177
1178         return p;
1179 }
1180
1181 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1182 {
1183         update_curr_rt(rq);
1184
1185         /*
1186          * The previous task needs to be made eligible for pushing
1187          * if it is still active
1188          */
1189         if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1190                 enqueue_pushable_task(rq, p);
1191 }
1192
1193 #ifdef CONFIG_SMP
1194
1195 /* Only try algorithms three times */
1196 #define RT_MAX_TRIES 3
1197
1198 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1199
1200 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1201 {
1202         if (!task_running(rq, p) &&
1203             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1204             (p->rt.nr_cpus_allowed > 1))
1205                 return 1;
1206         return 0;
1207 }
1208
1209 /* Return the second highest RT task, NULL otherwise */
1210 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1211 {
1212         struct task_struct *next = NULL;
1213         struct sched_rt_entity *rt_se;
1214         struct rt_prio_array *array;
1215         struct rt_rq *rt_rq;
1216         int idx;
1217
1218         for_each_leaf_rt_rq(rt_rq, rq) {
1219                 array = &rt_rq->active;
1220                 idx = sched_find_first_bit(array->bitmap);
1221 next_idx:
1222                 if (idx >= MAX_RT_PRIO)
1223                         continue;
1224                 if (next && next->prio < idx)
1225                         continue;
1226                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1227                         struct task_struct *p;
1228
1229                         if (!rt_entity_is_task(rt_se))
1230                                 continue;
1231
1232                         p = rt_task_of(rt_se);
1233                         if (pick_rt_task(rq, p, cpu)) {
1234                                 next = p;
1235                                 break;
1236                         }
1237                 }
1238                 if (!next) {
1239                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1240                         goto next_idx;
1241                 }
1242         }
1243
1244         return next;
1245 }
1246
1247 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1248
1249 static int find_lowest_rq(struct task_struct *task)
1250 {
1251         struct sched_domain *sd;
1252         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1253         int this_cpu = smp_processor_id();
1254         int cpu      = task_cpu(task);
1255
1256         /* Make sure the mask is initialized first */
1257         if (unlikely(!lowest_mask))
1258                 return -1;
1259
1260         if (task->rt.nr_cpus_allowed == 1)
1261                 return -1; /* No other targets possible */
1262
1263         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1264                 return -1; /* No targets found */
1265
1266         /*
1267          * At this point we have built a mask of cpus representing the
1268          * lowest priority tasks in the system.  Now we want to elect
1269          * the best one based on our affinity and topology.
1270          *
1271          * We prioritize the last cpu that the task executed on since
1272          * it is most likely cache-hot in that location.
1273          */
1274         if (cpumask_test_cpu(cpu, lowest_mask))
1275                 return cpu;
1276
1277         /*
1278          * Otherwise, we consult the sched_domains span maps to figure
1279          * out which cpu is logically closest to our hot cache data.
1280          */
1281         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1282                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1283
1284         rcu_read_lock();
1285         for_each_domain(cpu, sd) {
1286                 if (sd->flags & SD_WAKE_AFFINE) {
1287                         int best_cpu;
1288
1289                         /*
1290                          * "this_cpu" is cheaper to preempt than a
1291                          * remote processor.
1292                          */
1293                         if (this_cpu != -1 &&
1294                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1295                                 rcu_read_unlock();
1296                                 return this_cpu;
1297                         }
1298
1299                         best_cpu = cpumask_first_and(lowest_mask,
1300                                                      sched_domain_span(sd));
1301                         if (best_cpu < nr_cpu_ids) {
1302                                 rcu_read_unlock();
1303                                 return best_cpu;
1304                         }
1305                 }
1306         }
1307         rcu_read_unlock();
1308
1309         /*
1310          * And finally, if there were no matches within the domains
1311          * just give the caller *something* to work with from the compatible
1312          * locations.
1313          */
1314         if (this_cpu != -1)
1315                 return this_cpu;
1316
1317         cpu = cpumask_any(lowest_mask);
1318         if (cpu < nr_cpu_ids)
1319                 return cpu;
1320         return -1;
1321 }
1322
1323 /* Will lock the rq it finds */
1324 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1325 {
1326         struct rq *lowest_rq = NULL;
1327         int tries;
1328         int cpu;
1329
1330         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1331                 cpu = find_lowest_rq(task);
1332
1333                 if ((cpu == -1) || (cpu == rq->cpu))
1334                         break;
1335
1336                 lowest_rq = cpu_rq(cpu);
1337
1338                 /* if the prio of this runqueue changed, try again */
1339                 if (double_lock_balance(rq, lowest_rq)) {
1340                         /*
1341                          * We had to unlock the run queue. In
1342                          * the mean time, task could have
1343                          * migrated already or had its affinity changed.
1344                          * Also make sure that it wasn't scheduled on its rq.
1345                          */
1346                         if (unlikely(task_rq(task) != rq ||
1347                                      !cpumask_test_cpu(lowest_rq->cpu,
1348                                                        &task->cpus_allowed) ||
1349                                      task_running(rq, task) ||
1350                                      !task->on_rq)) {
1351
1352                                 raw_spin_unlock(&lowest_rq->lock);
1353                                 lowest_rq = NULL;
1354                                 break;
1355                         }
1356                 }
1357
1358                 /* If this rq is still suitable use it. */
1359                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1360                         break;
1361
1362                 /* try again */
1363                 double_unlock_balance(rq, lowest_rq);
1364                 lowest_rq = NULL;
1365         }
1366
1367         return lowest_rq;
1368 }
1369
1370 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1371 {
1372         struct task_struct *p;
1373
1374         if (!has_pushable_tasks(rq))
1375                 return NULL;
1376
1377         p = plist_first_entry(&rq->rt.pushable_tasks,
1378                               struct task_struct, pushable_tasks);
1379
1380         BUG_ON(rq->cpu != task_cpu(p));
1381         BUG_ON(task_current(rq, p));
1382         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1383
1384         BUG_ON(!p->on_rq);
1385         BUG_ON(!rt_task(p));
1386
1387         return p;
1388 }
1389
1390 /*
1391  * If the current CPU has more than one RT task, see if the non
1392  * running task can migrate over to a CPU that is running a task
1393  * of lesser priority.
1394  */
1395 static int push_rt_task(struct rq *rq)
1396 {
1397         struct task_struct *next_task;
1398         struct rq *lowest_rq;
1399         int ret = 0;
1400
1401         if (!rq->rt.overloaded)
1402                 return 0;
1403
1404         next_task = pick_next_pushable_task(rq);
1405         if (!next_task)
1406                 return 0;
1407
1408 retry:
1409         if (unlikely(next_task == rq->curr)) {
1410                 WARN_ON(1);
1411                 return 0;
1412         }
1413
1414         /*
1415          * It's possible that the next_task slipped in of
1416          * higher priority than current. If that's the case
1417          * just reschedule current.
1418          */
1419         if (unlikely(next_task->prio < rq->curr->prio)) {
1420                 resched_task(rq->curr);
1421                 return 0;
1422         }
1423
1424         /* We might release rq lock */
1425         get_task_struct(next_task);
1426
1427         /* find_lock_lowest_rq locks the rq if found */
1428         lowest_rq = find_lock_lowest_rq(next_task, rq);
1429         if (!lowest_rq) {
1430                 struct task_struct *task;
1431                 /*
1432                  * find_lock_lowest_rq releases rq->lock
1433                  * so it is possible that next_task has migrated.
1434                  *
1435                  * We need to make sure that the task is still on the same
1436                  * run-queue and is also still the next task eligible for
1437                  * pushing.
1438                  */
1439                 task = pick_next_pushable_task(rq);
1440                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1441                         /*
1442                          * The task hasn't migrated, and is still the next
1443                          * eligible task, but we failed to find a run-queue
1444                          * to push it to.  Do not retry in this case, since
1445                          * other cpus will pull from us when ready.
1446                          */
1447                         goto out;
1448                 }
1449
1450                 if (!task)
1451                         /* No more tasks, just exit */
1452                         goto out;
1453
1454                 /*
1455                  * Something has shifted, try again.
1456                  */
1457                 put_task_struct(next_task);
1458                 next_task = task;
1459                 goto retry;
1460         }
1461
1462         deactivate_task(rq, next_task, 0);
1463         set_task_cpu(next_task, lowest_rq->cpu);
1464         activate_task(lowest_rq, next_task, 0);
1465         ret = 1;
1466
1467         resched_task(lowest_rq->curr);
1468
1469         double_unlock_balance(rq, lowest_rq);
1470
1471 out:
1472         put_task_struct(next_task);
1473
1474         return ret;
1475 }
1476
1477 static void push_rt_tasks(struct rq *rq)
1478 {
1479         /* push_rt_task will return true if it moved an RT */
1480         while (push_rt_task(rq))
1481                 ;
1482 }
1483
1484 static int pull_rt_task(struct rq *this_rq)
1485 {
1486         int this_cpu = this_rq->cpu, ret = 0, cpu;
1487         struct task_struct *p;
1488         struct rq *src_rq;
1489
1490         if (likely(!rt_overloaded(this_rq)))
1491                 return 0;
1492
1493         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1494                 if (this_cpu == cpu)
1495                         continue;
1496
1497                 src_rq = cpu_rq(cpu);
1498
1499                 /*
1500                  * Don't bother taking the src_rq->lock if the next highest
1501                  * task is known to be lower-priority than our current task.
1502                  * This may look racy, but if this value is about to go
1503                  * logically higher, the src_rq will push this task away.
1504                  * And if its going logically lower, we do not care
1505                  */
1506                 if (src_rq->rt.highest_prio.next >=
1507                     this_rq->rt.highest_prio.curr)
1508                         continue;
1509
1510                 /*
1511                  * We can potentially drop this_rq's lock in
1512                  * double_lock_balance, and another CPU could
1513                  * alter this_rq
1514                  */
1515                 double_lock_balance(this_rq, src_rq);
1516
1517                 /*
1518                  * Are there still pullable RT tasks?
1519                  */
1520                 if (src_rq->rt.rt_nr_running <= 1)
1521                         goto skip;
1522
1523                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1524
1525                 /*
1526                  * Do we have an RT task that preempts
1527                  * the to-be-scheduled task?
1528                  */
1529                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1530                         WARN_ON(p == src_rq->curr);
1531                         WARN_ON(!p->on_rq);
1532
1533                         /*
1534                          * There's a chance that p is higher in priority
1535                          * than what's currently running on its cpu.
1536                          * This is just that p is wakeing up and hasn't
1537                          * had a chance to schedule. We only pull
1538                          * p if it is lower in priority than the
1539                          * current task on the run queue
1540                          */
1541                         if (p->prio < src_rq->curr->prio)
1542                                 goto skip;
1543
1544                         ret = 1;
1545
1546                         deactivate_task(src_rq, p, 0);
1547                         set_task_cpu(p, this_cpu);
1548                         activate_task(this_rq, p, 0);
1549                         /*
1550                          * We continue with the search, just in
1551                          * case there's an even higher prio task
1552                          * in another runqueue. (low likelihood
1553                          * but possible)
1554                          */
1555                 }
1556 skip:
1557                 double_unlock_balance(this_rq, src_rq);
1558         }
1559
1560         return ret;
1561 }
1562
1563 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1564 {
1565         /* Try to pull RT tasks here if we lower this rq's prio */
1566         if (rq->rt.highest_prio.curr > prev->prio)
1567                 pull_rt_task(rq);
1568 }
1569
1570 static void post_schedule_rt(struct rq *rq)
1571 {
1572         push_rt_tasks(rq);
1573 }
1574
1575 /*
1576  * If we are not running and we are not going to reschedule soon, we should
1577  * try to push tasks away now
1578  */
1579 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1580 {
1581         if (!task_running(rq, p) &&
1582             !test_tsk_need_resched(rq->curr) &&
1583             has_pushable_tasks(rq) &&
1584             p->rt.nr_cpus_allowed > 1 &&
1585             rt_task(rq->curr) &&
1586             (rq->curr->rt.nr_cpus_allowed < 2 ||
1587              rq->curr->prio < p->prio))
1588                 push_rt_tasks(rq);
1589 }
1590
1591 static void set_cpus_allowed_rt(struct task_struct *p,
1592                                 const struct cpumask *new_mask)
1593 {
1594         int weight = cpumask_weight(new_mask);
1595
1596         BUG_ON(!rt_task(p));
1597
1598         /*
1599          * Update the migration status of the RQ if we have an RT task
1600          * which is running AND changing its weight value.
1601          */
1602         if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1603                 struct rq *rq = task_rq(p);
1604
1605                 if (!task_current(rq, p)) {
1606                         /*
1607                          * Make sure we dequeue this task from the pushable list
1608                          * before going further.  It will either remain off of
1609                          * the list because we are no longer pushable, or it
1610                          * will be requeued.
1611                          */
1612                         if (p->rt.nr_cpus_allowed > 1)
1613                                 dequeue_pushable_task(rq, p);
1614
1615                         /*
1616                          * Requeue if our weight is changing and still > 1
1617                          */
1618                         if (weight > 1)
1619                                 enqueue_pushable_task(rq, p);
1620
1621                 }
1622
1623                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1624                         rq->rt.rt_nr_migratory++;
1625                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1626                         BUG_ON(!rq->rt.rt_nr_migratory);
1627                         rq->rt.rt_nr_migratory--;
1628                 }
1629
1630                 update_rt_migration(&rq->rt);
1631         }
1632
1633         cpumask_copy(&p->cpus_allowed, new_mask);
1634         p->rt.nr_cpus_allowed = weight;
1635 }
1636
1637 /* Assumes rq->lock is held */
1638 static void rq_online_rt(struct rq *rq)
1639 {
1640         if (rq->rt.overloaded)
1641                 rt_set_overload(rq);
1642
1643         __enable_runtime(rq);
1644
1645         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1646 }
1647
1648 /* Assumes rq->lock is held */
1649 static void rq_offline_rt(struct rq *rq)
1650 {
1651         if (rq->rt.overloaded)
1652                 rt_clear_overload(rq);
1653
1654         __disable_runtime(rq);
1655
1656         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1657 }
1658
1659 /*
1660  * When switch from the rt queue, we bring ourselves to a position
1661  * that we might want to pull RT tasks from other runqueues.
1662  */
1663 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1664 {
1665         /*
1666          * If there are other RT tasks then we will reschedule
1667          * and the scheduling of the other RT tasks will handle
1668          * the balancing. But if we are the last RT task
1669          * we may need to handle the pulling of RT tasks
1670          * now.
1671          */
1672         if (p->on_rq && !rq->rt.rt_nr_running)
1673                 pull_rt_task(rq);
1674 }
1675
1676 static inline void init_sched_rt_class(void)
1677 {
1678         unsigned int i;
1679
1680         for_each_possible_cpu(i)
1681                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1682                                         GFP_KERNEL, cpu_to_node(i));
1683 }
1684 #endif /* CONFIG_SMP */
1685
1686 /*
1687  * When switching a task to RT, we may overload the runqueue
1688  * with RT tasks. In this case we try to push them off to
1689  * other runqueues.
1690  */
1691 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1692 {
1693         int check_resched = 1;
1694
1695         /*
1696          * If we are already running, then there's nothing
1697          * that needs to be done. But if we are not running
1698          * we may need to preempt the current running task.
1699          * If that current running task is also an RT task
1700          * then see if we can move to another run queue.
1701          */
1702         if (p->on_rq && rq->curr != p) {
1703 #ifdef CONFIG_SMP
1704                 if (rq->rt.overloaded && push_rt_task(rq) &&
1705                     /* Don't resched if we changed runqueues */
1706                     rq != task_rq(p))
1707                         check_resched = 0;
1708 #endif /* CONFIG_SMP */
1709                 if (check_resched && p->prio < rq->curr->prio)
1710                         resched_task(rq->curr);
1711         }
1712 }
1713
1714 /*
1715  * Priority of the task has changed. This may cause
1716  * us to initiate a push or pull.
1717  */
1718 static void
1719 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1720 {
1721         if (!p->on_rq)
1722                 return;
1723
1724         if (rq->curr == p) {
1725 #ifdef CONFIG_SMP
1726                 /*
1727                  * If our priority decreases while running, we
1728                  * may need to pull tasks to this runqueue.
1729                  */
1730                 if (oldprio < p->prio)
1731                         pull_rt_task(rq);
1732                 /*
1733                  * If there's a higher priority task waiting to run
1734                  * then reschedule. Note, the above pull_rt_task
1735                  * can release the rq lock and p could migrate.
1736                  * Only reschedule if p is still on the same runqueue.
1737                  */
1738                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1739                         resched_task(p);
1740 #else
1741                 /* For UP simply resched on drop of prio */
1742                 if (oldprio < p->prio)
1743                         resched_task(p);
1744 #endif /* CONFIG_SMP */
1745         } else {
1746                 /*
1747                  * This task is not running, but if it is
1748                  * greater than the current running task
1749                  * then reschedule.
1750                  */
1751                 if (p->prio < rq->curr->prio)
1752                         resched_task(rq->curr);
1753         }
1754 }
1755
1756 static void watchdog(struct rq *rq, struct task_struct *p)
1757 {
1758         unsigned long soft, hard;
1759
1760         /* max may change after cur was read, this will be fixed next tick */
1761         soft = task_rlimit(p, RLIMIT_RTTIME);
1762         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1763
1764         if (soft != RLIM_INFINITY) {
1765                 unsigned long next;
1766
1767                 p->rt.timeout++;
1768                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1769                 if (p->rt.timeout > next)
1770                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1771         }
1772 }
1773
1774 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1775 {
1776         update_curr_rt(rq);
1777
1778         watchdog(rq, p);
1779
1780         /*
1781          * RR tasks need a special form of timeslice management.
1782          * FIFO tasks have no timeslices.
1783          */
1784         if (p->policy != SCHED_RR)
1785                 return;
1786
1787         if (--p->rt.time_slice)
1788                 return;
1789
1790         p->rt.time_slice = DEF_TIMESLICE;
1791
1792         /*
1793          * Requeue to the end of queue if we are not the only element
1794          * on the queue:
1795          */
1796         if (p->rt.run_list.prev != p->rt.run_list.next) {
1797                 requeue_task_rt(rq, p, 0);
1798                 set_tsk_need_resched(p);
1799         }
1800 }
1801
1802 static void set_curr_task_rt(struct rq *rq)
1803 {
1804         struct task_struct *p = rq->curr;
1805
1806         p->se.exec_start = rq->clock_task;
1807
1808         /* The running task is never eligible for pushing */
1809         dequeue_pushable_task(rq, p);
1810 }
1811
1812 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1813 {
1814         /*
1815          * Time slice is 0 for SCHED_FIFO tasks
1816          */
1817         if (task->policy == SCHED_RR)
1818                 return DEF_TIMESLICE;
1819         else
1820                 return 0;
1821 }
1822
1823 static const struct sched_class rt_sched_class = {
1824         .next                   = &fair_sched_class,
1825         .enqueue_task           = enqueue_task_rt,
1826         .dequeue_task           = dequeue_task_rt,
1827         .yield_task             = yield_task_rt,
1828
1829         .check_preempt_curr     = check_preempt_curr_rt,
1830
1831         .pick_next_task         = pick_next_task_rt,
1832         .put_prev_task          = put_prev_task_rt,
1833
1834 #ifdef CONFIG_SMP
1835         .select_task_rq         = select_task_rq_rt,
1836
1837         .set_cpus_allowed       = set_cpus_allowed_rt,
1838         .rq_online              = rq_online_rt,
1839         .rq_offline             = rq_offline_rt,
1840         .pre_schedule           = pre_schedule_rt,
1841         .post_schedule          = post_schedule_rt,
1842         .task_woken             = task_woken_rt,
1843         .switched_from          = switched_from_rt,
1844 #endif
1845
1846         .set_curr_task          = set_curr_task_rt,
1847         .task_tick              = task_tick_rt,
1848
1849         .get_rr_interval        = get_rr_interval_rt,
1850
1851         .prio_changed           = prio_changed_rt,
1852         .switched_to            = switched_to_rt,
1853 };
1854
1855 #ifdef CONFIG_SCHED_DEBUG
1856 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1857
1858 static void print_rt_stats(struct seq_file *m, int cpu)
1859 {
1860         rt_rq_iter_t iter;
1861         struct rt_rq *rt_rq;
1862
1863         rcu_read_lock();
1864         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
1865                 print_rt_rq(m, cpu, rt_rq);
1866         rcu_read_unlock();
1867 }
1868 #endif /* CONFIG_SCHED_DEBUG */
1869