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_SMP
   7
   8 static inline int rt_overloaded(struct rq *rq)
   9 {
  10         return atomic_read(&rq->rd->rto_count);
  11 }
  12
  13 static inline void rt_set_overload(struct rq *rq)
  14 {
  15         cpu_set(rq->cpu, rq->rd->rto_mask);
  16         /*
  17          * Make sure the mask is visible before we set
  18          * the overload count. That is checked to determine
  19          * if we should look at the mask. It would be a shame
  20          * if we looked at the mask, but the mask was not
  21          * updated yet.
  22          */
  23         wmb();
  24         atomic_inc(&rq->rd->rto_count);
  25 }
  26
  27 static inline void rt_clear_overload(struct rq *rq)
  28 {
  29         /* the order here really doesn't matter */
  30         atomic_dec(&rq->rd->rto_count);
  31         cpu_clear(rq->cpu, rq->rd->rto_mask);
  32 }
  33
  34 static void update_rt_migration(struct rq *rq)
  35 {
  36         if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
  37                 if (!rq->rt.overloaded) {
  38                         rt_set_overload(rq);
  39                         rq->rt.overloaded = 1;
  40                 }
  41         } else if (rq->rt.overloaded) {
  42                 rt_clear_overload(rq);
  43                 rq->rt.overloaded = 0;
  44         }
  45 }
  46 #endif /* CONFIG_SMP */
  47
  48 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  49 {
  50         return container_of(rt_se, struct task_struct, rt);
  51 }
  52
  53 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
  54 {
  55         return !list_empty(&rt_se->run_list);
  56 }
  57
  58 #ifdef CONFIG_RT_GROUP_SCHED
  59
  60 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
  61 {
  62         if (!rt_rq->tg)
  63                 return RUNTIME_INF;
  64
  65         return rt_rq->tg->rt_runtime;
  66 }
  67
  68 #define for_each_leaf_rt_rq(rt_rq, rq) \
  69         list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
  70
  71 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
  72 {
  73         return rt_rq->rq;
  74 }
  75
  76 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
  77 {
  78         return rt_se->rt_rq;
  79 }
  80
  81 #define for_each_sched_rt_entity(rt_se) \
  82         for (; rt_se; rt_se = rt_se->parent)
  83
  84 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
  85 {
  86         return rt_se->my_q;
  87 }
  88
  89 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
  90 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
  91
  92 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
  93 {
  94         struct sched_rt_entity *rt_se = rt_rq->rt_se;
  95
  96         if (rt_se && !on_rt_rq(rt_se) && rt_rq->rt_nr_running) {
  97                 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
  98
  99                 enqueue_rt_entity(rt_se);
 100                 if (rt_rq->highest_prio < curr->prio)
 101                         resched_task(curr);
 102         }
 103 }
 104
 105 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 106 {
 107         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 108
 109         if (rt_se && on_rt_rq(rt_se))
 110                 dequeue_rt_entity(rt_se);
 111 }
 112
 113 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 114 {
 115         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 116 }
 117
 118 static int rt_se_boosted(struct sched_rt_entity *rt_se)
 119 {
 120         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 121         struct task_struct *p;
 122
 123         if (rt_rq)
 124                 return !!rt_rq->rt_nr_boosted;
 125
 126         p = rt_task_of(rt_se);
 127         return p->prio != p->normal_prio;
 128 }
 129
 130 #else
 131
 132 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 133 {
 134         if (sysctl_sched_rt_runtime == -1)
 135                 return RUNTIME_INF;
 136
 137         return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
 138 }
 139
 140 #define for_each_leaf_rt_rq(rt_rq, rq) \
 141         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 142
 143 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
 144 {
 145         return container_of(rt_rq, struct rq, rt);
 146 }
 147
 148 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
 149 {
 150         struct task_struct *p = rt_task_of(rt_se);
 151         struct rq *rq = task_rq(p);
 152
 153         return &rq->rt;
 154 }
 155
 156 #define for_each_sched_rt_entity(rt_se) \
 157         for (; rt_se; rt_se = NULL)
 158
 159 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 160 {
 161         return NULL;
 162 }
 163
 164 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 165 {
 166 }
 167
 168 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 169 {
 170 }
 171
 172 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 173 {
 174         return rt_rq->rt_throttled;
 175 }
 176 #endif
 177
 178 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 179 {
 180 #ifdef CONFIG_RT_GROUP_SCHED
 181         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 182
 183         if (rt_rq)
 184                 return rt_rq->highest_prio;
 185 #endif
 186
 187         return rt_task_of(rt_se)->prio;
 188 }
 189
 190 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 191 {
 192         u64 runtime = sched_rt_runtime(rt_rq);
 193
 194         if (runtime == RUNTIME_INF)
 195                 return 0;
 196
 197         if (rt_rq->rt_throttled)
 198                 return rt_rq_throttled(rt_rq);
 199
 200         if (rt_rq->rt_time > runtime) {
 201                 struct rq *rq = rq_of_rt_rq(rt_rq);
 202
 203                 rq->rt_throttled = 1;
 204                 rt_rq->rt_throttled = 1;
 205
 206                 if (rt_rq_throttled(rt_rq)) {
 207                         sched_rt_rq_dequeue(rt_rq);
 208                         return 1;
 209                 }
 210         }
 211
 212         return 0;
 213 }
 214
 215 static void update_sched_rt_period(struct rq *rq)
 216 {
 217         struct rt_rq *rt_rq;
 218         u64 period;
 219
 220         while (rq->clock > rq->rt_period_expire) {
 221                 period = (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
 222                 rq->rt_period_expire += period;
 223
 224                 for_each_leaf_rt_rq(rt_rq, rq) {
 225                         u64 runtime = sched_rt_runtime(rt_rq);
 226
 227                         rt_rq->rt_time -= min(rt_rq->rt_time, runtime);
 228                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 229                                 rt_rq->rt_throttled = 0;
 230                                 sched_rt_rq_enqueue(rt_rq);
 231                         }
 232                 }
 233
 234                 rq->rt_throttled = 0;
 235         }
 236 }
 237
 238 /*
 239  * Update the current task's runtime statistics. Skip current tasks that
 240  * are not in our scheduling class.
 241  */
 242 static void update_curr_rt(struct rq *rq)
 243 {
 244         struct task_struct *curr = rq->curr;
 245         struct sched_rt_entity *rt_se = &curr->rt;
 246         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 247         u64 delta_exec;
 248
 249         if (!task_has_rt_policy(curr))
 250                 return;
 251
 252         delta_exec = rq->clock - curr->se.exec_start;
 253         if (unlikely((s64)delta_exec < 0))
 254                 delta_exec = 0;
 255
 256         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
 257
 258         curr->se.sum_exec_runtime += delta_exec;
 259         curr->se.exec_start = rq->clock;
 260         cpuacct_charge(curr, delta_exec);
 261
 262         rt_rq->rt_time += delta_exec;
 263         if (sched_rt_runtime_exceeded(rt_rq))
 264                 resched_task(curr);
 265 }
 266
 267 static inline
 268 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 269 {
 270         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 271         rt_rq->rt_nr_running++;
 272 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 273         if (rt_se_prio(rt_se) < rt_rq->highest_prio)
 274                 rt_rq->highest_prio = rt_se_prio(rt_se);
 275 #endif
 276 #ifdef CONFIG_SMP
 277         if (rt_se->nr_cpus_allowed > 1) {
 278                 struct rq *rq = rq_of_rt_rq(rt_rq);
 279                 rq->rt.rt_nr_migratory++;
 280         }
 281
 282         update_rt_migration(rq_of_rt_rq(rt_rq));
 283 #endif
 284 #ifdef CONFIG_RT_GROUP_SCHED
 285         if (rt_se_boosted(rt_se))
 286                 rt_rq->rt_nr_boosted++;
 287 #endif
 288 }
 289
 290 static inline
 291 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 292 {
 293         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 294         WARN_ON(!rt_rq->rt_nr_running);
 295         rt_rq->rt_nr_running--;
 296 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 297         if (rt_rq->rt_nr_running) {
 298                 struct rt_prio_array *array;
 299
 300                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
 301                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
 302                         /* recalculate */
 303                         array = &rt_rq->active;
 304                         rt_rq->highest_prio =
 305                                 sched_find_first_bit(array->bitmap);
 306                 } /* otherwise leave rq->highest prio alone */
 307         } else
 308                 rt_rq->highest_prio = MAX_RT_PRIO;
 309 #endif
 310 #ifdef CONFIG_SMP
 311         if (rt_se->nr_cpus_allowed > 1) {
 312                 struct rq *rq = rq_of_rt_rq(rt_rq);
 313                 rq->rt.rt_nr_migratory--;
 314         }
 315
 316         update_rt_migration(rq_of_rt_rq(rt_rq));
 317 #endif /* CONFIG_SMP */
 318 #ifdef CONFIG_RT_GROUP_SCHED
 319         if (rt_se_boosted(rt_se))
 320                 rt_rq->rt_nr_boosted--;
 321
 322         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 323 #endif
 324 }
 325
 326 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 327 {
 328         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 329         struct rt_prio_array *array = &rt_rq->active;
 330         struct rt_rq *group_rq = group_rt_rq(rt_se);
 331
 332         if (group_rq && rt_rq_throttled(group_rq))
 333                 return;
 334
 335         list_add_tail(&rt_se->run_list, array->queue + rt_se_prio(rt_se));
 336         __set_bit(rt_se_prio(rt_se), array->bitmap);
 337
 338         inc_rt_tasks(rt_se, rt_rq);
 339 }
 340
 341 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 342 {
 343         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 344         struct rt_prio_array *array = &rt_rq->active;
 345
 346         list_del_init(&rt_se->run_list);
 347         if (list_empty(array->queue + rt_se_prio(rt_se)))
 348                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 349
 350         dec_rt_tasks(rt_se, rt_rq);
 351 }
 352
 353 /*
 354  * Because the prio of an upper entry depends on the lower
 355  * entries, we must remove entries top - down.
 356  *
 357  * XXX: O(1/2 h^2) because we can only walk up, not down the chain.
 358  *      doesn't matter much for now, as h=2 for GROUP_SCHED.
 359  */
 360 static void dequeue_rt_stack(struct task_struct *p)
 361 {
 362         struct sched_rt_entity *rt_se, *top_se;
 363
 364         /*
 365          * dequeue all, top - down.
 366          */
 367         do {
 368                 rt_se = &p->rt;
 369                 top_se = NULL;
 370                 for_each_sched_rt_entity(rt_se) {
 371                         if (on_rt_rq(rt_se))
 372                                 top_se = rt_se;
 373                 }
 374                 if (top_se)
 375                         dequeue_rt_entity(top_se);
 376         } while (top_se);
 377 }
 378
 379 /*
 380  * Adding/removing a task to/from a priority array:
 381  */
 382 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 383 {
 384         struct sched_rt_entity *rt_se = &p->rt;
 385
 386         if (wakeup)
 387                 rt_se->timeout = 0;
 388
 389         dequeue_rt_stack(p);
 390
 391         /*
 392          * enqueue everybody, bottom - up.
 393          */
 394         for_each_sched_rt_entity(rt_se)
 395                 enqueue_rt_entity(rt_se);
 396
 397         inc_cpu_load(rq, p->se.load.weight);
 398 }
 399
 400 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 401 {
 402         struct sched_rt_entity *rt_se = &p->rt;
 403         struct rt_rq *rt_rq;
 404
 405         update_curr_rt(rq);
 406
 407         dequeue_rt_stack(p);
 408
 409         /*
 410          * re-enqueue all non-empty rt_rq entities.
 411          */
 412         for_each_sched_rt_entity(rt_se) {
 413                 rt_rq = group_rt_rq(rt_se);
 414                 if (rt_rq && rt_rq->rt_nr_running)
 415                         enqueue_rt_entity(rt_se);
 416         }
 417
 418         dec_cpu_load(rq, p->se.load.weight);
 419 }
 420
 421 /*
 422  * Put task to the end of the run list without the overhead of dequeue
 423  * followed by enqueue.
 424  */
 425 static
 426 void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
 427 {
 428         struct rt_prio_array *array = &rt_rq->active;
 429
 430         list_move_tail(&rt_se->run_list, array->queue + rt_se_prio(rt_se));
 431 }
 432
 433 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
 434 {
 435         struct sched_rt_entity *rt_se = &p->rt;
 436         struct rt_rq *rt_rq;
 437
 438         for_each_sched_rt_entity(rt_se) {
 439                 rt_rq = rt_rq_of_se(rt_se);
 440                 requeue_rt_entity(rt_rq, rt_se);
 441         }
 442 }
 443
 444 static void yield_task_rt(struct rq *rq)
 445 {
 446         requeue_task_rt(rq, rq->curr);
 447 }
 448
 449 #ifdef CONFIG_SMP
 450 static int find_lowest_rq(struct task_struct *task);
 451
 452 static int select_task_rq_rt(struct task_struct *p, int sync)
 453 {
 454         struct rq *rq = task_rq(p);
 455
 456         /*
 457          * If the current task is an RT task, then
 458          * try to see if we can wake this RT task up on another
 459          * runqueue. Otherwise simply start this RT task
 460          * on its current runqueue.
 461          *
 462          * We want to avoid overloading runqueues. Even if
 463          * the RT task is of higher priority than the current RT task.
 464          * RT tasks behave differently than other tasks. If
 465          * one gets preempted, we try to push it off to another queue.
 466          * So trying to keep a preempting RT task on the same
 467          * cache hot CPU will force the running RT task to
 468          * a cold CPU. So we waste all the cache for the lower
 469          * RT task in hopes of saving some of a RT task
 470          * that is just being woken and probably will have
 471          * cold cache anyway.
 472          */
 473         if (unlikely(rt_task(rq->curr)) &&
 474             (p->rt.nr_cpus_allowed > 1)) {
 475                 int cpu = find_lowest_rq(p);
 476
 477                 return (cpu == -1) ? task_cpu(p) : cpu;
 478         }
 479
 480         /*
 481          * Otherwise, just let it ride on the affined RQ and the
 482          * post-schedule router will push the preempted task away
 483          */
 484         return task_cpu(p);
 485 }
 486 #endif /* CONFIG_SMP */
 487
 488 /*
 489  * Preempt the current task with a newly woken task if needed:
 490  */
 491 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
 492 {
 493         if (p->prio < rq->curr->prio)
 494                 resched_task(rq->curr);
 495 }
 496
 497 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 498                                                    struct rt_rq *rt_rq)
 499 {
 500         struct rt_prio_array *array = &rt_rq->active;
 501         struct sched_rt_entity *next = NULL;
 502         struct list_head *queue;
 503         int idx;
 504
 505         idx = sched_find_first_bit(array->bitmap);
 506         BUG_ON(idx >= MAX_RT_PRIO);
 507
 508         queue = array->queue + idx;
 509         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 510
 511         return next;
 512 }
 513
 514 static struct task_struct *pick_next_task_rt(struct rq *rq)
 515 {
 516         struct sched_rt_entity *rt_se;
 517         struct task_struct *p;
 518         struct rt_rq *rt_rq;
 519
 520         rt_rq = &rq->rt;
 521
 522         if (unlikely(!rt_rq->rt_nr_running))
 523                 return NULL;
 524
 525         if (rt_rq_throttled(rt_rq))
 526                 return NULL;
 527
 528         do {
 529                 rt_se = pick_next_rt_entity(rq, rt_rq);
 530                 BUG_ON(!rt_se);
 531                 rt_rq = group_rt_rq(rt_se);
 532         } while (rt_rq);
 533
 534         p = rt_task_of(rt_se);
 535         p->se.exec_start = rq->clock;
 536         return p;
 537 }
 538
 539 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 540 {
 541         update_curr_rt(rq);
 542         p->se.exec_start = 0;
 543 }
 544
 545 #ifdef CONFIG_SMP
 546
 547 /* Only try algorithms three times */
 548 #define RT_MAX_TRIES 3
 549
 550 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
 551 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 552
 553 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 554 {
 555         if (!task_running(rq, p) &&
 556             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
 557             (p->rt.nr_cpus_allowed > 1))
 558                 return 1;
 559         return 0;
 560 }
 561
 562 /* Return the second highest RT task, NULL otherwise */
 563 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
 564 {
 565         struct task_struct *next = NULL;
 566         struct sched_rt_entity *rt_se;
 567         struct rt_prio_array *array;
 568         struct rt_rq *rt_rq;
 569         int idx;
 570
 571         for_each_leaf_rt_rq(rt_rq, rq) {
 572                 array = &rt_rq->active;
 573                 idx = sched_find_first_bit(array->bitmap);
 574  next_idx:
 575                 if (idx >= MAX_RT_PRIO)
 576                         continue;
 577                 if (next && next->prio < idx)
 578                         continue;
 579                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
 580                         struct task_struct *p = rt_task_of(rt_se);
 581                         if (pick_rt_task(rq, p, cpu)) {
 582                                 next = p;
 583                                 break;
 584                         }
 585                 }
 586                 if (!next) {
 587                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 588                         goto next_idx;
 589                 }
 590         }
 591
 592         return next;
 593 }
 594
 595 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
 596
 597 static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
 598 {
 599         int       lowest_prio = -1;
 600         int       lowest_cpu  = -1;
 601         int       count       = 0;
 602         int       cpu;
 603
 604         cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed);
 605
 606         /*
 607          * Scan each rq for the lowest prio.
 608          */
 609         for_each_cpu_mask(cpu, *lowest_mask) {
 610                 struct rq *rq = cpu_rq(cpu);
 611
 612                 /* We look for lowest RT prio or non-rt CPU */
 613                 if (rq->rt.highest_prio >= MAX_RT_PRIO) {
 614                         /*
 615                          * if we already found a low RT queue
 616                          * and now we found this non-rt queue
 617                          * clear the mask and set our bit.
 618                          * Otherwise just return the queue as is
 619                          * and the count==1 will cause the algorithm
 620                          * to use the first bit found.
 621                          */
 622                         if (lowest_cpu != -1) {
 623                                 cpus_clear(*lowest_mask);
 624                                 cpu_set(rq->cpu, *lowest_mask);
 625                         }
 626                         return 1;
 627                 }
 628
 629                 /* no locking for now */
 630                 if ((rq->rt.highest_prio > task->prio)
 631                     && (rq->rt.highest_prio >= lowest_prio)) {
 632                         if (rq->rt.highest_prio > lowest_prio) {
 633                                 /* new low - clear old data */
 634                                 lowest_prio = rq->rt.highest_prio;
 635                                 lowest_cpu = cpu;
 636                                 count = 0;
 637                         }
 638                         count++;
 639                 } else
 640                         cpu_clear(cpu, *lowest_mask);
 641         }
 642
 643         /*
 644          * Clear out all the set bits that represent
 645          * runqueues that were of higher prio than
 646          * the lowest_prio.
 647          */
 648         if (lowest_cpu > 0) {
 649                 /*
 650                  * Perhaps we could add another cpumask op to
 651                  * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
 652                  * Then that could be optimized to use memset and such.
 653                  */
 654                 for_each_cpu_mask(cpu, *lowest_mask) {
 655                         if (cpu >= lowest_cpu)
 656                                 break;
 657                         cpu_clear(cpu, *lowest_mask);
 658                 }
 659         }
 660
 661         return count;
 662 }
 663
 664 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
 665 {
 666         int first;
 667
 668         /* "this_cpu" is cheaper to preempt than a remote processor */
 669         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
 670                 return this_cpu;
 671
 672         first = first_cpu(*mask);
 673         if (first != NR_CPUS)
 674                 return first;
 675
 676         return -1;
 677 }
 678
 679 static int find_lowest_rq(struct task_struct *task)
 680 {
 681         struct sched_domain *sd;
 682         cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
 683         int this_cpu = smp_processor_id();
 684         int cpu      = task_cpu(task);
 685         int count    = find_lowest_cpus(task, lowest_mask);
 686
 687         if (!count)
 688                 return -1; /* No targets found */
 689
 690         /*
 691          * There is no sense in performing an optimal search if only one
 692          * target is found.
 693          */
 694         if (count == 1)
 695                 return first_cpu(*lowest_mask);
 696
 697         /*
 698          * At this point we have built a mask of cpus representing the
 699          * lowest priority tasks in the system.  Now we want to elect
 700          * the best one based on our affinity and topology.
 701          *
 702          * We prioritize the last cpu that the task executed on since
 703          * it is most likely cache-hot in that location.
 704          */
 705         if (cpu_isset(cpu, *lowest_mask))
 706                 return cpu;
 707
 708         /*
 709          * Otherwise, we consult the sched_domains span maps to figure
 710          * out which cpu is logically closest to our hot cache data.
 711          */
 712         if (this_cpu == cpu)
 713                 this_cpu = -1; /* Skip this_cpu opt if the same */
 714
 715         for_each_domain(cpu, sd) {
 716                 if (sd->flags & SD_WAKE_AFFINE) {
 717                         cpumask_t domain_mask;
 718                         int       best_cpu;
 719
 720                         cpus_and(domain_mask, sd->span, *lowest_mask);
 721
 722                         best_cpu = pick_optimal_cpu(this_cpu,
 723                                                     &domain_mask);
 724                         if (best_cpu != -1)
 725                                 return best_cpu;
 726                 }
 727         }
 728
 729         /*
 730          * And finally, if there were no matches within the domains
 731          * just give the caller *something* to work with from the compatible
 732          * locations.
 733          */
 734         return pick_optimal_cpu(this_cpu, lowest_mask);
 735 }
 736
 737 /* Will lock the rq it finds */
 738 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
 739 {
 740         struct rq *lowest_rq = NULL;
 741         int tries;
 742         int cpu;
 743
 744         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
 745                 cpu = find_lowest_rq(task);
 746
 747                 if ((cpu == -1) || (cpu == rq->cpu))
 748                         break;
 749
 750                 lowest_rq = cpu_rq(cpu);
 751
 752                 /* if the prio of this runqueue changed, try again */
 753                 if (double_lock_balance(rq, lowest_rq)) {
 754                         /*
 755                          * We had to unlock the run queue. In
 756                          * the mean time, task could have
 757                          * migrated already or had its affinity changed.
 758                          * Also make sure that it wasn't scheduled on its rq.
 759                          */
 760                         if (unlikely(task_rq(task) != rq ||
 761                                      !cpu_isset(lowest_rq->cpu,
 762                                                 task->cpus_allowed) ||
 763                                      task_running(rq, task) ||
 764                                      !task->se.on_rq)) {
 765
 766                                 spin_unlock(&lowest_rq->lock);
 767                                 lowest_rq = NULL;
 768                                 break;
 769                         }
 770                 }
 771
 772                 /* If this rq is still suitable use it. */
 773                 if (lowest_rq->rt.highest_prio > task->prio)
 774                         break;
 775
 776                 /* try again */
 777                 spin_unlock(&lowest_rq->lock);
 778                 lowest_rq = NULL;
 779         }
 780
 781         return lowest_rq;
 782 }
 783
 784 /*
 785  * If the current CPU has more than one RT task, see if the non
 786  * running task can migrate over to a CPU that is running a task
 787  * of lesser priority.
 788  */
 789 static int push_rt_task(struct rq *rq)
 790 {
 791         struct task_struct *next_task;
 792         struct rq *lowest_rq;
 793         int ret = 0;
 794         int paranoid = RT_MAX_TRIES;
 795
 796         if (!rq->rt.overloaded)
 797                 return 0;
 798
 799         next_task = pick_next_highest_task_rt(rq, -1);
 800         if (!next_task)
 801                 return 0;
 802
 803  retry:
 804         if (unlikely(next_task == rq->curr)) {
 805                 WARN_ON(1);
 806                 return 0;
 807         }
 808
 809         /*
 810          * It's possible that the next_task slipped in of
 811          * higher priority than current. If that's the case
 812          * just reschedule current.
 813          */
 814         if (unlikely(next_task->prio < rq->curr->prio)) {
 815                 resched_task(rq->curr);
 816                 return 0;
 817         }
 818
 819         /* We might release rq lock */
 820         get_task_struct(next_task);
 821
 822         /* find_lock_lowest_rq locks the rq if found */
 823         lowest_rq = find_lock_lowest_rq(next_task, rq);
 824         if (!lowest_rq) {
 825                 struct task_struct *task;
 826                 /*
 827                  * find lock_lowest_rq releases rq->lock
 828                  * so it is possible that next_task has changed.
 829                  * If it has, then try again.
 830                  */
 831                 task = pick_next_highest_task_rt(rq, -1);
 832                 if (unlikely(task != next_task) && task && paranoid--) {
 833                         put_task_struct(next_task);
 834                         next_task = task;
 835                         goto retry;
 836                 }
 837                 goto out;
 838         }
 839
 840         deactivate_task(rq, next_task, 0);
 841         set_task_cpu(next_task, lowest_rq->cpu);
 842         activate_task(lowest_rq, next_task, 0);
 843
 844         resched_task(lowest_rq->curr);
 845
 846         spin_unlock(&lowest_rq->lock);
 847
 848         ret = 1;
 849 out:
 850         put_task_struct(next_task);
 851
 852         return ret;
 853 }
 854
 855 /*
 856  * TODO: Currently we just use the second highest prio task on
 857  *       the queue, and stop when it can't migrate (or there's
 858  *       no more RT tasks).  There may be a case where a lower
 859  *       priority RT task has a different affinity than the
 860  *       higher RT task. In this case the lower RT task could
 861  *       possibly be able to migrate where as the higher priority
 862  *       RT task could not.  We currently ignore this issue.
 863  *       Enhancements are welcome!
 864  */
 865 static void push_rt_tasks(struct rq *rq)
 866 {
 867         /* push_rt_task will return true if it moved an RT */
 868         while (push_rt_task(rq))
 869                 ;
 870 }
 871
 872 static int pull_rt_task(struct rq *this_rq)
 873 {
 874         int this_cpu = this_rq->cpu, ret = 0, cpu;
 875         struct task_struct *p, *next;
 876         struct rq *src_rq;
 877
 878         if (likely(!rt_overloaded(this_rq)))
 879                 return 0;
 880
 881         next = pick_next_task_rt(this_rq);
 882
 883         for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
 884                 if (this_cpu == cpu)
 885                         continue;
 886
 887                 src_rq = cpu_rq(cpu);
 888                 /*
 889                  * We can potentially drop this_rq's lock in
 890                  * double_lock_balance, and another CPU could
 891                  * steal our next task - hence we must cause
 892                  * the caller to recalculate the next task
 893                  * in that case:
 894                  */
 895                 if (double_lock_balance(this_rq, src_rq)) {
 896                         struct task_struct *old_next = next;
 897
 898                         next = pick_next_task_rt(this_rq);
 899                         if (next != old_next)
 900                                 ret = 1;
 901                 }
 902
 903                 /*
 904                  * Are there still pullable RT tasks?
 905                  */
 906                 if (src_rq->rt.rt_nr_running <= 1)
 907                         goto skip;
 908
 909                 p = pick_next_highest_task_rt(src_rq, this_cpu);
 910
 911                 /*
 912                  * Do we have an RT task that preempts
 913                  * the to-be-scheduled task?
 914                  */
 915                 if (p && (!next || (p->prio < next->prio))) {
 916                         WARN_ON(p == src_rq->curr);
 917                         WARN_ON(!p->se.on_rq);
 918
 919                         /*
 920                          * There's a chance that p is higher in priority
 921                          * than what's currently running on its cpu.
 922                          * This is just that p is wakeing up and hasn't
 923                          * had a chance to schedule. We only pull
 924                          * p if it is lower in priority than the
 925                          * current task on the run queue or
 926                          * this_rq next task is lower in prio than
 927                          * the current task on that rq.
 928                          */
 929                         if (p->prio < src_rq->curr->prio ||
 930                             (next && next->prio < src_rq->curr->prio))
 931                                 goto skip;
 932
 933                         ret = 1;
 934
 935                         deactivate_task(src_rq, p, 0);
 936                         set_task_cpu(p, this_cpu);
 937                         activate_task(this_rq, p, 0);
 938                         /*
 939                          * We continue with the search, just in
 940                          * case there's an even higher prio task
 941                          * in another runqueue. (low likelyhood
 942                          * but possible)
 943                          *
 944                          * Update next so that we won't pick a task
 945                          * on another cpu with a priority lower (or equal)
 946                          * than the one we just picked.
 947                          */
 948                         next = p;
 949
 950                 }
 951  skip:
 952                 spin_unlock(&src_rq->lock);
 953         }
 954
 955         return ret;
 956 }
 957
 958 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
 959 {
 960         /* Try to pull RT tasks here if we lower this rq's prio */
 961         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
 962                 pull_rt_task(rq);
 963 }
 964
 965 static void post_schedule_rt(struct rq *rq)
 966 {
 967         /*
 968          * If we have more than one rt_task queued, then
 969          * see if we can push the other rt_tasks off to other CPUS.
 970          * Note we may release the rq lock, and since
 971          * the lock was owned by prev, we need to release it
 972          * first via finish_lock_switch and then reaquire it here.
 973          */
 974         if (unlikely(rq->rt.overloaded)) {
 975                 spin_lock_irq(&rq->lock);
 976                 push_rt_tasks(rq);
 977                 spin_unlock_irq(&rq->lock);
 978         }
 979 }
 980
 981
 982 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
 983 {
 984         if (!task_running(rq, p) &&
 985             (p->prio >= rq->rt.highest_prio) &&
 986             rq->rt.overloaded)
 987                 push_rt_tasks(rq);
 988 }
 989
 990 static unsigned long
 991 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
 992                 unsigned long max_load_move,
 993                 struct sched_domain *sd, enum cpu_idle_type idle,
 994                 int *all_pinned, int *this_best_prio)
 995 {
 996         /* don't touch RT tasks */
 997         return 0;
 998 }
 999
1000 static int
1001 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1002                  struct sched_domain *sd, enum cpu_idle_type idle)
1003 {
1004         /* don't touch RT tasks */
1005         return 0;
1006 }
1007
1008 static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
1009 {
1010         int weight = cpus_weight(*new_mask);
1011
1012         BUG_ON(!rt_task(p));
1013
1014         /*
1015          * Update the migration status of the RQ if we have an RT task
1016          * which is running AND changing its weight value.
1017          */
1018         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1019                 struct rq *rq = task_rq(p);
1020
1021                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1022                         rq->rt.rt_nr_migratory++;
1023                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1024                         BUG_ON(!rq->rt.rt_nr_migratory);
1025                         rq->rt.rt_nr_migratory--;
1026                 }
1027
1028                 update_rt_migration(rq);
1029         }
1030
1031         p->cpus_allowed    = *new_mask;
1032         p->rt.nr_cpus_allowed = weight;
1033 }
1034
1035 /* Assumes rq->lock is held */
1036 static void join_domain_rt(struct rq *rq)
1037 {
1038         if (rq->rt.overloaded)
1039                 rt_set_overload(rq);
1040 }
1041
1042 /* Assumes rq->lock is held */
1043 static void leave_domain_rt(struct rq *rq)
1044 {
1045         if (rq->rt.overloaded)
1046                 rt_clear_overload(rq);
1047 }
1048
1049 /*
1050  * When switch from the rt queue, we bring ourselves to a position
1051  * that we might want to pull RT tasks from other runqueues.
1052  */
1053 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1054                            int running)
1055 {
1056         /*
1057          * If there are other RT tasks then we will reschedule
1058          * and the scheduling of the other RT tasks will handle
1059          * the balancing. But if we are the last RT task
1060          * we may need to handle the pulling of RT tasks
1061          * now.
1062          */
1063         if (!rq->rt.rt_nr_running)
1064                 pull_rt_task(rq);
1065 }
1066 #endif /* CONFIG_SMP */
1067
1068 /*
1069  * When switching a task to RT, we may overload the runqueue
1070  * with RT tasks. In this case we try to push them off to
1071  * other runqueues.
1072  */
1073 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1074                            int running)
1075 {
1076         int check_resched = 1;
1077
1078         /*
1079          * If we are already running, then there's nothing
1080          * that needs to be done. But if we are not running
1081          * we may need to preempt the current running task.
1082          * If that current running task is also an RT task
1083          * then see if we can move to another run queue.
1084          */
1085         if (!running) {
1086 #ifdef CONFIG_SMP
1087                 if (rq->rt.overloaded && push_rt_task(rq) &&
1088                     /* Don't resched if we changed runqueues */
1089                     rq != task_rq(p))
1090                         check_resched = 0;
1091 #endif /* CONFIG_SMP */
1092                 if (check_resched && p->prio < rq->curr->prio)
1093                         resched_task(rq->curr);
1094         }
1095 }
1096
1097 /*
1098  * Priority of the task has changed. This may cause
1099  * us to initiate a push or pull.
1100  */
1101 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1102                             int oldprio, int running)
1103 {
1104         if (running) {
1105 #ifdef CONFIG_SMP
1106                 /*
1107                  * If our priority decreases while running, we
1108                  * may need to pull tasks to this runqueue.
1109                  */
1110                 if (oldprio < p->prio)
1111                         pull_rt_task(rq);
1112                 /*
1113                  * If there's a higher priority task waiting to run
1114                  * then reschedule.
1115                  */
1116                 if (p->prio > rq->rt.highest_prio)
1117                         resched_task(p);
1118 #else
1119                 /* For UP simply resched on drop of prio */
1120                 if (oldprio < p->prio)
1121                         resched_task(p);
1122 #endif /* CONFIG_SMP */
1123         } else {
1124                 /*
1125                  * This task is not running, but if it is
1126                  * greater than the current running task
1127                  * then reschedule.
1128                  */
1129                 if (p->prio < rq->curr->prio)
1130                         resched_task(rq->curr);
1131         }
1132 }
1133
1134 static void watchdog(struct rq *rq, struct task_struct *p)
1135 {
1136         unsigned long soft, hard;
1137
1138         if (!p->signal)
1139                 return;
1140
1141         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1142         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1143
1144         if (soft != RLIM_INFINITY) {
1145                 unsigned long next;
1146
1147                 p->rt.timeout++;
1148                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1149                 if (p->rt.timeout > next)
1150                         p->it_sched_expires = p->se.sum_exec_runtime;
1151         }
1152 }
1153
1154 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1155 {
1156         update_curr_rt(rq);
1157
1158         watchdog(rq, p);
1159
1160         /*
1161          * RR tasks need a special form of timeslice management.
1162          * FIFO tasks have no timeslices.
1163          */
1164         if (p->policy != SCHED_RR)
1165                 return;
1166
1167         if (--p->rt.time_slice)
1168                 return;
1169
1170         p->rt.time_slice = DEF_TIMESLICE;
1171
1172         /*
1173          * Requeue to the end of queue if we are not the only element
1174          * on the queue:
1175          */
1176         if (p->rt.run_list.prev != p->rt.run_list.next) {
1177                 requeue_task_rt(rq, p);
1178                 set_tsk_need_resched(p);
1179         }
1180 }
1181
1182 static void set_curr_task_rt(struct rq *rq)
1183 {
1184         struct task_struct *p = rq->curr;
1185
1186         p->se.exec_start = rq->clock;
1187 }
1188
1189 const struct sched_class rt_sched_class = {
1190         .next                   = &fair_sched_class,
1191         .enqueue_task           = enqueue_task_rt,
1192         .dequeue_task           = dequeue_task_rt,
1193         .yield_task             = yield_task_rt,
1194 #ifdef CONFIG_SMP
1195         .select_task_rq         = select_task_rq_rt,
1196 #endif /* CONFIG_SMP */
1197
1198         .check_preempt_curr     = check_preempt_curr_rt,
1199
1200         .pick_next_task         = pick_next_task_rt,
1201         .put_prev_task          = put_prev_task_rt,
1202
1203 #ifdef CONFIG_SMP
1204         .load_balance           = load_balance_rt,
1205         .move_one_task          = move_one_task_rt,
1206         .set_cpus_allowed       = set_cpus_allowed_rt,
1207         .join_domain            = join_domain_rt,
1208         .leave_domain           = leave_domain_rt,
1209         .pre_schedule           = pre_schedule_rt,
1210         .post_schedule          = post_schedule_rt,
1211         .task_wake_up           = task_wake_up_rt,
1212         .switched_from          = switched_from_rt,
1213 #endif
1214
1215         .set_curr_task          = set_curr_task_rt,
1216         .task_tick              = task_tick_rt,
1217
1218         .prio_changed           = prio_changed_rt,
1219         .switched_to            = switched_to_rt,
1220 };