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