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