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