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         if (!rq->online)
  16                 return;
  17
  18         cpu_set(rq->cpu, rq->rd->rto_mask);
  19         /*
  20          * Make sure the mask is visible before we set
  21          * the overload count. That is checked to determine
  22          * if we should look at the mask. It would be a shame
  23          * if we looked at the mask, but the mask was not
  24          * updated yet.
  25          */
  26         wmb();
  27         atomic_inc(&rq->rd->rto_count);
  28 }
  29
  30 static inline void rt_clear_overload(struct rq *rq)
  31 {
  32         if (!rq->online)
  33                 return;
  34
  35         /* the order here really doesn't matter */
  36         atomic_dec(&rq->rd->rto_count);
  37         cpu_clear(rq->cpu, rq->rd->rto_mask);
  38 }
  39
  40 static void update_rt_migration(struct rq *rq)
  41 {
  42         if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
  43                 if (!rq->rt.overloaded) {
  44                         rt_set_overload(rq);
  45                         rq->rt.overloaded = 1;
  46                 }
  47         } else if (rq->rt.overloaded) {
  48                 rt_clear_overload(rq);
  49                 rq->rt.overloaded = 0;
  50         }
  51 }
  52 #endif /* CONFIG_SMP */
  53
  54 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  55 {
  56         return container_of(rt_se, struct task_struct, rt);
  57 }
  58
  59 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
  60 {
  61         return !list_empty(&rt_se->run_list);
  62 }
  63
  64 #ifdef CONFIG_RT_GROUP_SCHED
  65
  66 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
  67 {
  68         if (!rt_rq->tg)
  69                 return RUNTIME_INF;
  70
  71         return rt_rq->rt_runtime;
  72 }
  73
  74 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
  75 {
  76         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
  77 }
  78
  79 #define for_each_leaf_rt_rq(rt_rq, rq) \
  80         list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
  81
  82 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
  83 {
  84         return rt_rq->rq;
  85 }
  86
  87 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
  88 {
  89         return rt_se->rt_rq;
  90 }
  91
  92 #define for_each_sched_rt_entity(rt_se) \
  93         for (; rt_se; rt_se = rt_se->parent)
  94
  95 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
  96 {
  97         return rt_se->my_q;
  98 }
  99
 100 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
 101 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
 102
 103 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 104 {
 105         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 106         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 107
 108         if (rt_rq->rt_nr_running) {
 109                 if (rt_se && !on_rt_rq(rt_se))
 110                         enqueue_rt_entity(rt_se);
 111                 if (rt_rq->highest_prio < curr->prio)
 112                         resched_task(curr);
 113         }
 114 }
 115
 116 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 117 {
 118         struct sched_rt_entity *rt_se = rt_rq->rt_se;
 119
 120         if (rt_se && on_rt_rq(rt_se))
 121                 dequeue_rt_entity(rt_se);
 122 }
 123
 124 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 125 {
 126         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 127 }
 128
 129 static int rt_se_boosted(struct sched_rt_entity *rt_se)
 130 {
 131         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 132         struct task_struct *p;
 133
 134         if (rt_rq)
 135                 return !!rt_rq->rt_nr_boosted;
 136
 137         p = rt_task_of(rt_se);
 138         return p->prio != p->normal_prio;
 139 }
 140
 141 #ifdef CONFIG_SMP
 142 static inline cpumask_t sched_rt_period_mask(void)
 143 {
 144         return cpu_rq(smp_processor_id())->rd->span;
 145 }
 146 #else
 147 static inline cpumask_t sched_rt_period_mask(void)
 148 {
 149         return cpu_online_map;
 150 }
 151 #endif
 152
 153 static inline
 154 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 155 {
 156         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
 157 }
 158
 159 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 160 {
 161         return &rt_rq->tg->rt_bandwidth;
 162 }
 163
 164 #else /* !CONFIG_RT_GROUP_SCHED */
 165
 166 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 167 {
 168         return rt_rq->rt_runtime;
 169 }
 170
 171 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 172 {
 173         return ktime_to_ns(def_rt_bandwidth.rt_period);
 174 }
 175
 176 #define for_each_leaf_rt_rq(rt_rq, rq) \
 177         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 178
 179 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
 180 {
 181         return container_of(rt_rq, struct rq, rt);
 182 }
 183
 184 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
 185 {
 186         struct task_struct *p = rt_task_of(rt_se);
 187         struct rq *rq = task_rq(p);
 188
 189         return &rq->rt;
 190 }
 191
 192 #define for_each_sched_rt_entity(rt_se) \
 193         for (; rt_se; rt_se = NULL)
 194
 195 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 196 {
 197         return NULL;
 198 }
 199
 200 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 201 {
 202         if (rt_rq->rt_nr_running)
 203                 resched_task(rq_of_rt_rq(rt_rq)->curr);
 204 }
 205
 206 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 207 {
 208 }
 209
 210 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 211 {
 212         return rt_rq->rt_throttled;
 213 }
 214
 215 static inline cpumask_t sched_rt_period_mask(void)
 216 {
 217         return cpu_online_map;
 218 }
 219
 220 static inline
 221 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 222 {
 223         return &cpu_rq(cpu)->rt;
 224 }
 225
 226 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 227 {
 228         return &def_rt_bandwidth;
 229 }
 230
 231 #endif /* CONFIG_RT_GROUP_SCHED */
 232
 233 #ifdef CONFIG_SMP
 234 /*
 235  * We ran out of runtime, see if we can borrow some from our neighbours.
 236  */
 237 static int do_balance_runtime(struct rt_rq *rt_rq)
 238 {
 239         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 240         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 241         int i, weight, more = 0;
 242         u64 rt_period;
 243
 244         weight = cpus_weight(rd->span);
 245
 246         spin_lock(&rt_b->rt_runtime_lock);
 247         rt_period = ktime_to_ns(rt_b->rt_period);
 248         for_each_cpu_mask_nr(i, rd->span) {
 249                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 250                 s64 diff;
 251
 252                 if (iter == rt_rq)
 253                         continue;
 254
 255                 spin_lock(&iter->rt_runtime_lock);
 256                 /*
 257                  * Either all rqs have inf runtime and there's nothing to steal
 258                  * or __disable_runtime() below sets a specific rq to inf to
 259                  * indicate its been disabled and disalow stealing.
 260                  */
 261                 if (iter->rt_runtime == RUNTIME_INF)
 262                         goto next;
 263
 264                 /*
 265                  * From runqueues with spare time, take 1/n part of their
 266                  * spare time, but no more than our period.
 267                  */
 268                 diff = iter->rt_runtime - iter->rt_time;
 269                 if (diff > 0) {
 270                         diff = div_u64((u64)diff, weight);
 271                         if (rt_rq->rt_runtime + diff > rt_period)
 272                                 diff = rt_period - rt_rq->rt_runtime;
 273                         iter->rt_runtime -= diff;
 274                         rt_rq->rt_runtime += diff;
 275                         more = 1;
 276                         if (rt_rq->rt_runtime == rt_period) {
 277                                 spin_unlock(&iter->rt_runtime_lock);
 278                                 break;
 279                         }
 280                 }
 281 next:
 282                 spin_unlock(&iter->rt_runtime_lock);
 283         }
 284         spin_unlock(&rt_b->rt_runtime_lock);
 285
 286         return more;
 287 }
 288
 289 /*
 290  * Ensure this RQ takes back all the runtime it lend to its neighbours.
 291  */
 292 static void __disable_runtime(struct rq *rq)
 293 {
 294         struct root_domain *rd = rq->rd;
 295         struct rt_rq *rt_rq;
 296
 297         if (unlikely(!scheduler_running))
 298                 return;
 299
 300         for_each_leaf_rt_rq(rt_rq, rq) {
 301                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 302                 s64 want;
 303                 int i;
 304
 305                 spin_lock(&rt_b->rt_runtime_lock);
 306                 spin_lock(&rt_rq->rt_runtime_lock);
 307                 /*
 308                  * Either we're all inf and nobody needs to borrow, or we're
 309                  * already disabled and thus have nothing to do, or we have
 310                  * exactly the right amount of runtime to take out.
 311                  */
 312                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 313                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 314                         goto balanced;
 315                 spin_unlock(&rt_rq->rt_runtime_lock);
 316
 317                 /*
 318                  * Calculate the difference between what we started out with
 319                  * and what we current have, that's the amount of runtime
 320                  * we lend and now have to reclaim.
 321                  */
 322                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 323
 324                 /*
 325                  * Greedy reclaim, take back as much as we can.
 326                  */
 327                 for_each_cpu_mask(i, rd->span) {
 328                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 329                         s64 diff;
 330
 331                         /*
 332                          * Can't reclaim from ourselves or disabled runqueues.
 333                          */
 334                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 335                                 continue;
 336
 337                         spin_lock(&iter->rt_runtime_lock);
 338                         if (want > 0) {
 339                                 diff = min_t(s64, iter->rt_runtime, want);
 340                                 iter->rt_runtime -= diff;
 341                                 want -= diff;
 342                         } else {
 343                                 iter->rt_runtime -= want;
 344                                 want -= want;
 345                         }
 346                         spin_unlock(&iter->rt_runtime_lock);
 347
 348                         if (!want)
 349                                 break;
 350                 }
 351
 352                 spin_lock(&rt_rq->rt_runtime_lock);
 353                 /*
 354                  * We cannot be left wanting - that would mean some runtime
 355                  * leaked out of the system.
 356                  */
 357                 BUG_ON(want);
 358 balanced:
 359                 /*
 360                  * Disable all the borrow logic by pretending we have inf
 361                  * runtime - in which case borrowing doesn't make sense.
 362                  */
 363                 rt_rq->rt_runtime = RUNTIME_INF;
 364                 spin_unlock(&rt_rq->rt_runtime_lock);
 365                 spin_unlock(&rt_b->rt_runtime_lock);
 366         }
 367 }
 368
 369 static void disable_runtime(struct rq *rq)
 370 {
 371         unsigned long flags;
 372
 373         spin_lock_irqsave(&rq->lock, flags);
 374         __disable_runtime(rq);
 375         spin_unlock_irqrestore(&rq->lock, flags);
 376 }
 377
 378 static void __enable_runtime(struct rq *rq)
 379 {
 380         struct rt_rq *rt_rq;
 381
 382         if (unlikely(!scheduler_running))
 383                 return;
 384
 385         /*
 386          * Reset each runqueue's bandwidth settings
 387          */
 388         for_each_leaf_rt_rq(rt_rq, rq) {
 389                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 390
 391                 spin_lock(&rt_b->rt_runtime_lock);
 392                 spin_lock(&rt_rq->rt_runtime_lock);
 393                 rt_rq->rt_runtime = rt_b->rt_runtime;
 394                 rt_rq->rt_time = 0;
 395                 rt_rq->rt_throttled = 0;
 396                 spin_unlock(&rt_rq->rt_runtime_lock);
 397                 spin_unlock(&rt_b->rt_runtime_lock);
 398         }
 399 }
 400
 401 static void enable_runtime(struct rq *rq)
 402 {
 403         unsigned long flags;
 404
 405         spin_lock_irqsave(&rq->lock, flags);
 406         __enable_runtime(rq);
 407         spin_unlock_irqrestore(&rq->lock, flags);
 408 }
 409
 410 static int balance_runtime(struct rt_rq *rt_rq)
 411 {
 412         int more = 0;
 413
 414         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 415                 spin_unlock(&rt_rq->rt_runtime_lock);
 416                 more = do_balance_runtime(rt_rq);
 417                 spin_lock(&rt_rq->rt_runtime_lock);
 418         }
 419
 420         return more;
 421 }
 422 #else /* !CONFIG_SMP */
 423 static inline int balance_runtime(struct rt_rq *rt_rq)
 424 {
 425         return 0;
 426 }
 427 #endif /* CONFIG_SMP */
 428
 429 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 430 {
 431         int i, idle = 1;
 432         cpumask_t span;
 433
 434         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 435                 return 1;
 436
 437         span = sched_rt_period_mask();
 438         for_each_cpu_mask(i, span) {
 439                 int enqueue = 0;
 440                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 441                 struct rq *rq = rq_of_rt_rq(rt_rq);
 442
 443                 spin_lock(&rq->lock);
 444                 if (rt_rq->rt_time) {
 445                         u64 runtime;
 446
 447                         spin_lock(&rt_rq->rt_runtime_lock);
 448                         if (rt_rq->rt_throttled)
 449                                 balance_runtime(rt_rq);
 450                         runtime = rt_rq->rt_runtime;
 451                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 452                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 453                                 rt_rq->rt_throttled = 0;
 454                                 enqueue = 1;
 455                         }
 456                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 457                                 idle = 0;
 458                         spin_unlock(&rt_rq->rt_runtime_lock);
 459                 } else if (rt_rq->rt_nr_running)
 460                         idle = 0;
 461
 462                 if (enqueue)
 463                         sched_rt_rq_enqueue(rt_rq);
 464                 spin_unlock(&rq->lock);
 465         }
 466
 467         return idle;
 468 }
 469
 470 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 471 {
 472 #ifdef CONFIG_RT_GROUP_SCHED
 473         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 474
 475         if (rt_rq)
 476                 return rt_rq->highest_prio;
 477 #endif
 478
 479         return rt_task_of(rt_se)->prio;
 480 }
 481
 482 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 483 {
 484         u64 runtime = sched_rt_runtime(rt_rq);
 485
 486         if (rt_rq->rt_throttled)
 487                 return rt_rq_throttled(rt_rq);
 488
 489         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 490                 return 0;
 491
 492         balance_runtime(rt_rq);
 493         runtime = sched_rt_runtime(rt_rq);
 494         if (runtime == RUNTIME_INF)
 495                 return 0;
 496
 497         if (rt_rq->rt_time > runtime) {
 498                 rt_rq->rt_throttled = 1;
 499                 if (rt_rq_throttled(rt_rq)) {
 500                         sched_rt_rq_dequeue(rt_rq);
 501                         return 1;
 502                 }
 503         }
 504
 505         return 0;
 506 }
 507
 508 /*
 509  * Update the current task's runtime statistics. Skip current tasks that
 510  * are not in our scheduling class.
 511  */
 512 static void update_curr_rt(struct rq *rq)
 513 {
 514         struct task_struct *curr = rq->curr;
 515         struct sched_rt_entity *rt_se = &curr->rt;
 516         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 517         u64 delta_exec;
 518
 519         if (!task_has_rt_policy(curr))
 520                 return;
 521
 522         delta_exec = rq->clock - curr->se.exec_start;
 523         if (unlikely((s64)delta_exec < 0))
 524                 delta_exec = 0;
 525
 526         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
 527
 528         curr->se.sum_exec_runtime += delta_exec;
 529         account_group_exec_runtime(curr, delta_exec);
 530
 531         curr->se.exec_start = rq->clock;
 532         cpuacct_charge(curr, delta_exec);
 533
 534         if (!rt_bandwidth_enabled())
 535                 return;
 536
 537         for_each_sched_rt_entity(rt_se) {
 538                 rt_rq = rt_rq_of_se(rt_se);
 539
 540                 spin_lock(&rt_rq->rt_runtime_lock);
 541                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 542                         rt_rq->rt_time += delta_exec;
 543                         if (sched_rt_runtime_exceeded(rt_rq))
 544                                 resched_task(curr);
 545                 }
 546                 spin_unlock(&rt_rq->rt_runtime_lock);
 547         }
 548 }
 549
 550 static inline
 551 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 552 {
 553         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 554         rt_rq->rt_nr_running++;
 555 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 556         if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
 557 #ifdef CONFIG_SMP
 558                 struct rq *rq = rq_of_rt_rq(rt_rq);
 559 #endif
 560
 561                 rt_rq->highest_prio = rt_se_prio(rt_se);
 562 #ifdef CONFIG_SMP
 563                 if (rq->online)
 564                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 565                                    rt_se_prio(rt_se));
 566 #endif
 567         }
 568 #endif
 569 #ifdef CONFIG_SMP
 570         if (rt_se->nr_cpus_allowed > 1) {
 571                 struct rq *rq = rq_of_rt_rq(rt_rq);
 572
 573                 rq->rt.rt_nr_migratory++;
 574         }
 575
 576         update_rt_migration(rq_of_rt_rq(rt_rq));
 577 #endif
 578 #ifdef CONFIG_RT_GROUP_SCHED
 579         if (rt_se_boosted(rt_se))
 580                 rt_rq->rt_nr_boosted++;
 581
 582         if (rt_rq->tg)
 583                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 584 #else
 585         start_rt_bandwidth(&def_rt_bandwidth);
 586 #endif
 587 }
 588
 589 static inline
 590 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 591 {
 592 #ifdef CONFIG_SMP
 593         int highest_prio = rt_rq->highest_prio;
 594 #endif
 595
 596         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 597         WARN_ON(!rt_rq->rt_nr_running);
 598         rt_rq->rt_nr_running--;
 599 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 600         if (rt_rq->rt_nr_running) {
 601                 struct rt_prio_array *array;
 602
 603                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
 604                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
 605                         /* recalculate */
 606                         array = &rt_rq->active;
 607                         rt_rq->highest_prio =
 608                                 sched_find_first_bit(array->bitmap);
 609                 } /* otherwise leave rq->highest prio alone */
 610         } else
 611                 rt_rq->highest_prio = MAX_RT_PRIO;
 612 #endif
 613 #ifdef CONFIG_SMP
 614         if (rt_se->nr_cpus_allowed > 1) {
 615                 struct rq *rq = rq_of_rt_rq(rt_rq);
 616                 rq->rt.rt_nr_migratory--;
 617         }
 618
 619         if (rt_rq->highest_prio != highest_prio) {
 620                 struct rq *rq = rq_of_rt_rq(rt_rq);
 621
 622                 if (rq->online)
 623                         cpupri_set(&rq->rd->cpupri, rq->cpu,
 624                                    rt_rq->highest_prio);
 625         }
 626
 627         update_rt_migration(rq_of_rt_rq(rt_rq));
 628 #endif /* CONFIG_SMP */
 629 #ifdef CONFIG_RT_GROUP_SCHED
 630         if (rt_se_boosted(rt_se))
 631                 rt_rq->rt_nr_boosted--;
 632
 633         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 634 #endif
 635 }
 636
 637 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
 638 {
 639         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 640         struct rt_prio_array *array = &rt_rq->active;
 641         struct rt_rq *group_rq = group_rt_rq(rt_se);
 642         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 643
 644         /*
 645          * Don't enqueue the group if its throttled, or when empty.
 646          * The latter is a consequence of the former when a child group
 647          * get throttled and the current group doesn't have any other
 648          * active members.
 649          */
 650         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 651                 return;
 652
 653         list_add_tail(&rt_se->run_list, queue);
 654         __set_bit(rt_se_prio(rt_se), array->bitmap);
 655
 656         inc_rt_tasks(rt_se, rt_rq);
 657 }
 658
 659 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 660 {
 661         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 662         struct rt_prio_array *array = &rt_rq->active;
 663
 664         list_del_init(&rt_se->run_list);
 665         if (list_empty(array->queue + rt_se_prio(rt_se)))
 666                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 667
 668         dec_rt_tasks(rt_se, rt_rq);
 669 }
 670
 671 /*
 672  * Because the prio of an upper entry depends on the lower
 673  * entries, we must remove entries top - down.
 674  */
 675 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 676 {
 677         struct sched_rt_entity *back = NULL;
 678
 679         for_each_sched_rt_entity(rt_se) {
 680                 rt_se->back = back;
 681                 back = rt_se;
 682         }
 683
 684         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 685                 if (on_rt_rq(rt_se))
 686                         __dequeue_rt_entity(rt_se);
 687         }
 688 }
 689
 690 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 691 {
 692         dequeue_rt_stack(rt_se);
 693         for_each_sched_rt_entity(rt_se)
 694                 __enqueue_rt_entity(rt_se);
 695 }
 696
 697 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 698 {
 699         dequeue_rt_stack(rt_se);
 700
 701         for_each_sched_rt_entity(rt_se) {
 702                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 703
 704                 if (rt_rq && rt_rq->rt_nr_running)
 705                         __enqueue_rt_entity(rt_se);
 706         }
 707 }
 708
 709 /*
 710  * Adding/removing a task to/from a priority array:
 711  */
 712 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 713 {
 714         struct sched_rt_entity *rt_se = &p->rt;
 715
 716         if (wakeup)
 717                 rt_se->timeout = 0;
 718
 719         enqueue_rt_entity(rt_se);
 720
 721         inc_cpu_load(rq, p->se.load.weight);
 722 }
 723
 724 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 725 {
 726         struct sched_rt_entity *rt_se = &p->rt;
 727
 728         update_curr_rt(rq);
 729         dequeue_rt_entity(rt_se);
 730
 731         dec_cpu_load(rq, p->se.load.weight);
 732 }
 733
 734 /*
 735  * Put task to the end of the run list without the overhead of dequeue
 736  * followed by enqueue.
 737  */
 738 static void
 739 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 740 {
 741         if (on_rt_rq(rt_se)) {
 742                 struct rt_prio_array *array = &rt_rq->active;
 743                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 744
 745                 if (head)
 746                         list_move(&rt_se->run_list, queue);
 747                 else
 748                         list_move_tail(&rt_se->run_list, queue);
 749         }
 750 }
 751
 752 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 753 {
 754         struct sched_rt_entity *rt_se = &p->rt;
 755         struct rt_rq *rt_rq;
 756
 757         for_each_sched_rt_entity(rt_se) {
 758                 rt_rq = rt_rq_of_se(rt_se);
 759                 requeue_rt_entity(rt_rq, rt_se, head);
 760         }
 761 }
 762
 763 static void yield_task_rt(struct rq *rq)
 764 {
 765         requeue_task_rt(rq, rq->curr, 0);
 766 }
 767
 768 #ifdef CONFIG_SMP
 769 static int find_lowest_rq(struct task_struct *task);
 770
 771 static int select_task_rq_rt(struct task_struct *p, int sync)
 772 {
 773         struct rq *rq = task_rq(p);
 774
 775         /*
 776          * If the current task is an RT task, then
 777          * try to see if we can wake this RT task up on another
 778          * runqueue. Otherwise simply start this RT task
 779          * on its current runqueue.
 780          *
 781          * We want to avoid overloading runqueues. Even if
 782          * the RT task is of higher priority than the current RT task.
 783          * RT tasks behave differently than other tasks. If
 784          * one gets preempted, we try to push it off to another queue.
 785          * So trying to keep a preempting RT task on the same
 786          * cache hot CPU will force the running RT task to
 787          * a cold CPU. So we waste all the cache for the lower
 788          * RT task in hopes of saving some of a RT task
 789          * that is just being woken and probably will have
 790          * cold cache anyway.
 791          */
 792         if (unlikely(rt_task(rq->curr)) &&
 793             (p->rt.nr_cpus_allowed > 1)) {
 794                 int cpu = find_lowest_rq(p);
 795
 796                 return (cpu == -1) ? task_cpu(p) : cpu;
 797         }
 798
 799         /*
 800          * Otherwise, just let it ride on the affined RQ and the
 801          * post-schedule router will push the preempted task away
 802          */
 803         return task_cpu(p);
 804 }
 805
 806 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
 807 {
 808         cpumask_t mask;
 809
 810         if (rq->curr->rt.nr_cpus_allowed == 1)
 811                 return;
 812
 813         if (p->rt.nr_cpus_allowed != 1
 814             && cpupri_find(&rq->rd->cpupri, p, &mask))
 815                 return;
 816
 817         if (!cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
 818                 return;
 819
 820         /*
 821          * There appears to be other cpus that can accept
 822          * current and none to run 'p', so lets reschedule
 823          * to try and push current away:
 824          */
 825         requeue_task_rt(rq, p, 1);
 826         resched_task(rq->curr);
 827 }
 828
 829 #endif /* CONFIG_SMP */
 830
 831 /*
 832  * Preempt the current task with a newly woken task if needed:
 833  */
 834 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
 835 {
 836         if (p->prio < rq->curr->prio) {
 837                 resched_task(rq->curr);
 838                 return;
 839         }
 840
 841 #ifdef CONFIG_SMP
 842         /*
 843          * If:
 844          *
 845          * - the newly woken task is of equal priority to the current task
 846          * - the newly woken task is non-migratable while current is migratable
 847          * - current will be preempted on the next reschedule
 848          *
 849          * we should check to see if current can readily move to a different
 850          * cpu.  If so, we will reschedule to allow the push logic to try
 851          * to move current somewhere else, making room for our non-migratable
 852          * task.
 853          */
 854         if (p->prio == rq->curr->prio && !need_resched())
 855                 check_preempt_equal_prio(rq, p);
 856 #endif
 857 }
 858
 859 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 860                                                    struct rt_rq *rt_rq)
 861 {
 862         struct rt_prio_array *array = &rt_rq->active;
 863         struct sched_rt_entity *next = NULL;
 864         struct list_head *queue;
 865         int idx;
 866
 867         idx = sched_find_first_bit(array->bitmap);
 868         BUG_ON(idx >= MAX_RT_PRIO);
 869
 870         queue = array->queue + idx;
 871         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 872
 873         return next;
 874 }
 875
 876 static struct task_struct *pick_next_task_rt(struct rq *rq)
 877 {
 878         struct sched_rt_entity *rt_se;
 879         struct task_struct *p;
 880         struct rt_rq *rt_rq;
 881
 882         rt_rq = &rq->rt;
 883
 884         if (unlikely(!rt_rq->rt_nr_running))
 885                 return NULL;
 886
 887         if (rt_rq_throttled(rt_rq))
 888                 return NULL;
 889
 890         do {
 891                 rt_se = pick_next_rt_entity(rq, rt_rq);
 892                 BUG_ON(!rt_se);
 893                 rt_rq = group_rt_rq(rt_se);
 894         } while (rt_rq);
 895
 896         p = rt_task_of(rt_se);
 897         p->se.exec_start = rq->clock;
 898         return p;
 899 }
 900
 901 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 902 {
 903         update_curr_rt(rq);
 904         p->se.exec_start = 0;
 905 }
 906
 907 #ifdef CONFIG_SMP
 908
 909 /* Only try algorithms three times */
 910 #define RT_MAX_TRIES 3
 911
 912 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
 913 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest);
 914
 915 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 916
 917 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 918 {
 919         if (!task_running(rq, p) &&
 920             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
 921             (p->rt.nr_cpus_allowed > 1))
 922                 return 1;
 923         return 0;
 924 }
 925
 926 /* Return the second highest RT task, NULL otherwise */
 927 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
 928 {
 929         struct task_struct *next = NULL;
 930         struct sched_rt_entity *rt_se;
 931         struct rt_prio_array *array;
 932         struct rt_rq *rt_rq;
 933         int idx;
 934
 935         for_each_leaf_rt_rq(rt_rq, rq) {
 936                 array = &rt_rq->active;
 937                 idx = sched_find_first_bit(array->bitmap);
 938  next_idx:
 939                 if (idx >= MAX_RT_PRIO)
 940                         continue;
 941                 if (next && next->prio < idx)
 942                         continue;
 943                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
 944                         struct task_struct *p = rt_task_of(rt_se);
 945                         if (pick_rt_task(rq, p, cpu)) {
 946                                 next = p;
 947                                 break;
 948                         }
 949                 }
 950                 if (!next) {
 951                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 952                         goto next_idx;
 953                 }
 954         }
 955
 956         return next;
 957 }
 958
 959 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
 960
 961 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
 962 {
 963         int first;
 964
 965         /* "this_cpu" is cheaper to preempt than a remote processor */
 966         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
 967                 return this_cpu;
 968
 969         first = first_cpu(*mask);
 970         if (first != NR_CPUS)
 971                 return first;
 972
 973         return -1;
 974 }
 975
 976 static int find_lowest_rq(struct task_struct *task)
 977 {
 978         struct sched_domain *sd;
 979         cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
 980         int this_cpu = smp_processor_id();
 981         int cpu      = task_cpu(task);
 982
 983         if (task->rt.nr_cpus_allowed == 1)
 984                 return -1; /* No other targets possible */
 985
 986         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
 987                 return -1; /* No targets found */
 988
 989         /*
 990          * Only consider CPUs that are usable for migration.
 991          * I guess we might want to change cpupri_find() to ignore those
 992          * in the first place.
 993          */
 994         cpus_and(*lowest_mask, *lowest_mask, cpu_active_map);
 995
 996         /*
 997          * At this point we have built a mask of cpus representing the
 998          * lowest priority tasks in the system.  Now we want to elect
 999          * the best one based on our affinity and topology.
1000          *
1001          * We prioritize the last cpu that the task executed on since
1002          * it is most likely cache-hot in that location.
1003          */
1004         if (cpu_isset(cpu, *lowest_mask))
1005                 return cpu;
1006
1007         /*
1008          * Otherwise, we consult the sched_domains span maps to figure
1009          * out which cpu is logically closest to our hot cache data.
1010          */
1011         if (this_cpu == cpu)
1012                 this_cpu = -1; /* Skip this_cpu opt if the same */
1013
1014         for_each_domain(cpu, sd) {
1015                 if (sd->flags & SD_WAKE_AFFINE) {
1016                         cpumask_t domain_mask;
1017                         int       best_cpu;
1018
1019                         cpus_and(domain_mask, sd->span, *lowest_mask);
1020
1021                         best_cpu = pick_optimal_cpu(this_cpu,
1022                                                     &domain_mask);
1023                         if (best_cpu != -1)
1024                                 return best_cpu;
1025                 }
1026         }
1027
1028         /*
1029          * And finally, if there were no matches within the domains
1030          * just give the caller *something* to work with from the compatible
1031          * locations.
1032          */
1033         return pick_optimal_cpu(this_cpu, lowest_mask);
1034 }
1035
1036 /* Will lock the rq it finds */
1037 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1038 {
1039         struct rq *lowest_rq = NULL;
1040         int tries;
1041         int cpu;
1042
1043         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1044                 cpu = find_lowest_rq(task);
1045
1046                 if ((cpu == -1) || (cpu == rq->cpu))
1047                         break;
1048
1049                 lowest_rq = cpu_rq(cpu);
1050
1051                 /* if the prio of this runqueue changed, try again */
1052                 if (double_lock_balance(rq, lowest_rq)) {
1053                         /*
1054                          * We had to unlock the run queue. In
1055                          * the mean time, task could have
1056                          * migrated already or had its affinity changed.
1057                          * Also make sure that it wasn't scheduled on its rq.
1058                          */
1059                         if (unlikely(task_rq(task) != rq ||
1060                                      !cpu_isset(lowest_rq->cpu,
1061                                                 task->cpus_allowed) ||
1062                                      task_running(rq, task) ||
1063                                      !task->se.on_rq)) {
1064
1065                                 spin_unlock(&lowest_rq->lock);
1066                                 lowest_rq = NULL;
1067                                 break;
1068                         }
1069                 }
1070
1071                 /* If this rq is still suitable use it. */
1072                 if (lowest_rq->rt.highest_prio > task->prio)
1073                         break;
1074
1075                 /* try again */
1076                 double_unlock_balance(rq, lowest_rq);
1077                 lowest_rq = NULL;
1078         }
1079
1080         return lowest_rq;
1081 }
1082
1083 /*
1084  * If the current CPU has more than one RT task, see if the non
1085  * running task can migrate over to a CPU that is running a task
1086  * of lesser priority.
1087  */
1088 static int push_rt_task(struct rq *rq)
1089 {
1090         struct task_struct *next_task;
1091         struct rq *lowest_rq;
1092         int ret = 0;
1093         int paranoid = RT_MAX_TRIES;
1094
1095         if (!rq->rt.overloaded)
1096                 return 0;
1097
1098         next_task = pick_next_highest_task_rt(rq, -1);
1099         if (!next_task)
1100                 return 0;
1101
1102  retry:
1103         if (unlikely(next_task == rq->curr)) {
1104                 WARN_ON(1);
1105                 return 0;
1106         }
1107
1108         /*
1109          * It's possible that the next_task slipped in of
1110          * higher priority than current. If that's the case
1111          * just reschedule current.
1112          */
1113         if (unlikely(next_task->prio < rq->curr->prio)) {
1114                 resched_task(rq->curr);
1115                 return 0;
1116         }
1117
1118         /* We might release rq lock */
1119         get_task_struct(next_task);
1120
1121         /* find_lock_lowest_rq locks the rq if found */
1122         lowest_rq = find_lock_lowest_rq(next_task, rq);
1123         if (!lowest_rq) {
1124                 struct task_struct *task;
1125                 /*
1126                  * find lock_lowest_rq releases rq->lock
1127                  * so it is possible that next_task has changed.
1128                  * If it has, then try again.
1129                  */
1130                 task = pick_next_highest_task_rt(rq, -1);
1131                 if (unlikely(task != next_task) && task && paranoid--) {
1132                         put_task_struct(next_task);
1133                         next_task = task;
1134                         goto retry;
1135                 }
1136                 goto out;
1137         }
1138
1139         deactivate_task(rq, next_task, 0);
1140         set_task_cpu(next_task, lowest_rq->cpu);
1141         activate_task(lowest_rq, next_task, 0);
1142
1143         resched_task(lowest_rq->curr);
1144
1145         double_unlock_balance(rq, lowest_rq);
1146
1147         ret = 1;
1148 out:
1149         put_task_struct(next_task);
1150
1151         return ret;
1152 }
1153
1154 /*
1155  * TODO: Currently we just use the second highest prio task on
1156  *       the queue, and stop when it can't migrate (or there's
1157  *       no more RT tasks).  There may be a case where a lower
1158  *       priority RT task has a different affinity than the
1159  *       higher RT task. In this case the lower RT task could
1160  *       possibly be able to migrate where as the higher priority
1161  *       RT task could not.  We currently ignore this issue.
1162  *       Enhancements are welcome!
1163  */
1164 static void push_rt_tasks(struct rq *rq)
1165 {
1166         /* push_rt_task will return true if it moved an RT */
1167         while (push_rt_task(rq))
1168                 ;
1169 }
1170
1171 static int pull_rt_task(struct rq *this_rq)
1172 {
1173         int this_cpu = this_rq->cpu, ret = 0, cpu;
1174         struct task_struct *p, *next;
1175         struct rq *src_rq;
1176
1177         if (likely(!rt_overloaded(this_rq)))
1178                 return 0;
1179
1180         next = pick_next_task_rt(this_rq);
1181
1182         for_each_cpu_mask_nr(cpu, this_rq->rd->rto_mask) {
1183                 if (this_cpu == cpu)
1184                         continue;
1185
1186                 src_rq = cpu_rq(cpu);
1187                 /*
1188                  * We can potentially drop this_rq's lock in
1189                  * double_lock_balance, and another CPU could
1190                  * steal our next task - hence we must cause
1191                  * the caller to recalculate the next task
1192                  * in that case:
1193                  */
1194                 if (double_lock_balance(this_rq, src_rq)) {
1195                         struct task_struct *old_next = next;
1196
1197                         next = pick_next_task_rt(this_rq);
1198                         if (next != old_next)
1199                                 ret = 1;
1200                 }
1201
1202                 /*
1203                  * Are there still pullable RT tasks?
1204                  */
1205                 if (src_rq->rt.rt_nr_running <= 1)
1206                         goto skip;
1207
1208                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1209
1210                 /*
1211                  * Do we have an RT task that preempts
1212                  * the to-be-scheduled task?
1213                  */
1214                 if (p && (!next || (p->prio < next->prio))) {
1215                         WARN_ON(p == src_rq->curr);
1216                         WARN_ON(!p->se.on_rq);
1217
1218                         /*
1219                          * There's a chance that p is higher in priority
1220                          * than what's currently running on its cpu.
1221                          * This is just that p is wakeing up and hasn't
1222                          * had a chance to schedule. We only pull
1223                          * p if it is lower in priority than the
1224                          * current task on the run queue or
1225                          * this_rq next task is lower in prio than
1226                          * the current task on that rq.
1227                          */
1228                         if (p->prio < src_rq->curr->prio ||
1229                             (next && next->prio < src_rq->curr->prio))
1230                                 goto skip;
1231
1232                         ret = 1;
1233
1234                         deactivate_task(src_rq, p, 0);
1235                         set_task_cpu(p, this_cpu);
1236                         activate_task(this_rq, p, 0);
1237                         /*
1238                          * We continue with the search, just in
1239                          * case there's an even higher prio task
1240                          * in another runqueue. (low likelyhood
1241                          * but possible)
1242                          *
1243                          * Update next so that we won't pick a task
1244                          * on another cpu with a priority lower (or equal)
1245                          * than the one we just picked.
1246                          */
1247                         next = p;
1248
1249                 }
1250  skip:
1251                 double_unlock_balance(this_rq, src_rq);
1252         }
1253
1254         return ret;
1255 }
1256
1257 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1258 {
1259         /* Try to pull RT tasks here if we lower this rq's prio */
1260         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1261                 pull_rt_task(rq);
1262 }
1263
1264 static void post_schedule_rt(struct rq *rq)
1265 {
1266         /*
1267          * If we have more than one rt_task queued, then
1268          * see if we can push the other rt_tasks off to other CPUS.
1269          * Note we may release the rq lock, and since
1270          * the lock was owned by prev, we need to release it
1271          * first via finish_lock_switch and then reaquire it here.
1272          */
1273         if (unlikely(rq->rt.overloaded)) {
1274                 spin_lock_irq(&rq->lock);
1275                 push_rt_tasks(rq);
1276                 spin_unlock_irq(&rq->lock);
1277         }
1278 }
1279
1280 /*
1281  * If we are not running and we are not going to reschedule soon, we should
1282  * try to push tasks away now
1283  */
1284 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1285 {
1286         if (!task_running(rq, p) &&
1287             !test_tsk_need_resched(rq->curr) &&
1288             rq->rt.overloaded)
1289                 push_rt_tasks(rq);
1290 }
1291
1292 static unsigned long
1293 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1294                 unsigned long max_load_move,
1295                 struct sched_domain *sd, enum cpu_idle_type idle,
1296                 int *all_pinned, int *this_best_prio)
1297 {
1298         /* don't touch RT tasks */
1299         return 0;
1300 }
1301
1302 static int
1303 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1304                  struct sched_domain *sd, enum cpu_idle_type idle)
1305 {
1306         /* don't touch RT tasks */
1307         return 0;
1308 }
1309
1310 static void set_cpus_allowed_rt(struct task_struct *p,
1311                                 const cpumask_t *new_mask)
1312 {
1313         int weight = cpus_weight(*new_mask);
1314
1315         BUG_ON(!rt_task(p));
1316
1317         /*
1318          * Update the migration status of the RQ if we have an RT task
1319          * which is running AND changing its weight value.
1320          */
1321         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1322                 struct rq *rq = task_rq(p);
1323
1324                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1325                         rq->rt.rt_nr_migratory++;
1326                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1327                         BUG_ON(!rq->rt.rt_nr_migratory);
1328                         rq->rt.rt_nr_migratory--;
1329                 }
1330
1331                 update_rt_migration(rq);
1332         }
1333
1334         p->cpus_allowed    = *new_mask;
1335         p->rt.nr_cpus_allowed = weight;
1336 }
1337
1338 /* Assumes rq->lock is held */
1339 static void rq_online_rt(struct rq *rq)
1340 {
1341         if (rq->rt.overloaded)
1342                 rt_set_overload(rq);
1343
1344         __enable_runtime(rq);
1345
1346         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1347 }
1348
1349 /* Assumes rq->lock is held */
1350 static void rq_offline_rt(struct rq *rq)
1351 {
1352         if (rq->rt.overloaded)
1353                 rt_clear_overload(rq);
1354
1355         __disable_runtime(rq);
1356
1357         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1358 }
1359
1360 /*
1361  * When switch from the rt queue, we bring ourselves to a position
1362  * that we might want to pull RT tasks from other runqueues.
1363  */
1364 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1365                            int running)
1366 {
1367         /*
1368          * If there are other RT tasks then we will reschedule
1369          * and the scheduling of the other RT tasks will handle
1370          * the balancing. But if we are the last RT task
1371          * we may need to handle the pulling of RT tasks
1372          * now.
1373          */
1374         if (!rq->rt.rt_nr_running)
1375                 pull_rt_task(rq);
1376 }
1377 #endif /* CONFIG_SMP */
1378
1379 /*
1380  * When switching a task to RT, we may overload the runqueue
1381  * with RT tasks. In this case we try to push them off to
1382  * other runqueues.
1383  */
1384 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1385                            int running)
1386 {
1387         int check_resched = 1;
1388
1389         /*
1390          * If we are already running, then there's nothing
1391          * that needs to be done. But if we are not running
1392          * we may need to preempt the current running task.
1393          * If that current running task is also an RT task
1394          * then see if we can move to another run queue.
1395          */
1396         if (!running) {
1397 #ifdef CONFIG_SMP
1398                 if (rq->rt.overloaded && push_rt_task(rq) &&
1399                     /* Don't resched if we changed runqueues */
1400                     rq != task_rq(p))
1401                         check_resched = 0;
1402 #endif /* CONFIG_SMP */
1403                 if (check_resched && p->prio < rq->curr->prio)
1404                         resched_task(rq->curr);
1405         }
1406 }
1407
1408 /*
1409  * Priority of the task has changed. This may cause
1410  * us to initiate a push or pull.
1411  */
1412 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1413                             int oldprio, int running)
1414 {
1415         if (running) {
1416 #ifdef CONFIG_SMP
1417                 /*
1418                  * If our priority decreases while running, we
1419                  * may need to pull tasks to this runqueue.
1420                  */
1421                 if (oldprio < p->prio)
1422                         pull_rt_task(rq);
1423                 /*
1424                  * If there's a higher priority task waiting to run
1425                  * then reschedule. Note, the above pull_rt_task
1426                  * can release the rq lock and p could migrate.
1427                  * Only reschedule if p is still on the same runqueue.
1428                  */
1429                 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1430                         resched_task(p);
1431 #else
1432                 /* For UP simply resched on drop of prio */
1433                 if (oldprio < p->prio)
1434                         resched_task(p);
1435 #endif /* CONFIG_SMP */
1436         } else {
1437                 /*
1438                  * This task is not running, but if it is
1439                  * greater than the current running task
1440                  * then reschedule.
1441                  */
1442                 if (p->prio < rq->curr->prio)
1443                         resched_task(rq->curr);
1444         }
1445 }
1446
1447 static void watchdog(struct rq *rq, struct task_struct *p)
1448 {
1449         unsigned long soft, hard;
1450
1451         if (!p->signal)
1452                 return;
1453
1454         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1455         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1456
1457         if (soft != RLIM_INFINITY) {
1458                 unsigned long next;
1459
1460                 p->rt.timeout++;
1461                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1462                 if (p->rt.timeout > next)
1463                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1464         }
1465 }
1466
1467 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1468 {
1469         update_curr_rt(rq);
1470
1471         watchdog(rq, p);
1472
1473         /*
1474          * RR tasks need a special form of timeslice management.
1475          * FIFO tasks have no timeslices.
1476          */
1477         if (p->policy != SCHED_RR)
1478                 return;
1479
1480         if (--p->rt.time_slice)
1481                 return;
1482
1483         p->rt.time_slice = DEF_TIMESLICE;
1484
1485         /*
1486          * Requeue to the end of queue if we are not the only element
1487          * on the queue:
1488          */
1489         if (p->rt.run_list.prev != p->rt.run_list.next) {
1490                 requeue_task_rt(rq, p, 0);
1491                 set_tsk_need_resched(p);
1492         }
1493 }
1494
1495 static void set_curr_task_rt(struct rq *rq)
1496 {
1497         struct task_struct *p = rq->curr;
1498
1499         p->se.exec_start = rq->clock;
1500 }
1501
1502 static const struct sched_class rt_sched_class = {
1503         .next                   = &fair_sched_class,
1504         .enqueue_task           = enqueue_task_rt,
1505         .dequeue_task           = dequeue_task_rt,
1506         .yield_task             = yield_task_rt,
1507
1508         .check_preempt_curr     = check_preempt_curr_rt,
1509
1510         .pick_next_task         = pick_next_task_rt,
1511         .put_prev_task          = put_prev_task_rt,
1512
1513 #ifdef CONFIG_SMP
1514         .select_task_rq         = select_task_rq_rt,
1515
1516         .load_balance           = load_balance_rt,
1517         .move_one_task          = move_one_task_rt,
1518         .set_cpus_allowed       = set_cpus_allowed_rt,
1519         .rq_online              = rq_online_rt,
1520         .rq_offline             = rq_offline_rt,
1521         .pre_schedule           = pre_schedule_rt,
1522         .post_schedule          = post_schedule_rt,
1523         .task_wake_up           = task_wake_up_rt,
1524         .switched_from          = switched_from_rt,
1525 #endif
1526
1527         .set_curr_task          = set_curr_task_rt,
1528         .task_tick              = task_tick_rt,
1529
1530         .prio_changed           = prio_changed_rt,
1531         .switched_to            = switched_to_rt,
1532 };
1533
1534 #ifdef CONFIG_SCHED_DEBUG
1535 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1536
1537 static void print_rt_stats(struct seq_file *m, int cpu)
1538 {
1539         struct rt_rq *rt_rq;
1540
1541         rcu_read_lock();
1542         for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1543                 print_rt_rq(m, cpu, rt_rq);
1544         rcu_read_unlock();
1545 }
1546 #endif /* CONFIG_SCHED_DEBUG */