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         cpumask_set_cpu(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         cpumask_clear_cpu(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 < 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 const struct cpumask *sched_rt_period_mask(void)
 143 {
 144         return cpu_rq(smp_processor_id())->rd->span;
 145 }
 146 #else
 147 static inline const struct cpumask *sched_rt_period_mask(void)
 148 {
 149         return cpu_online_mask;
 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 const struct cpumask *sched_rt_period_mask(void)
 216 {
 217         return cpu_online_mask;
 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 = cpumask_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(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(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         const struct cpumask *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(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.curr;
 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                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 541                         spin_lock(&rt_rq->rt_runtime_lock);
 542                         rt_rq->rt_time += delta_exec;
 543                         if (sched_rt_runtime_exceeded(rt_rq))
 544                                 resched_task(curr);
 545                         spin_unlock(&rt_rq->rt_runtime_lock);
 546                 }
 547         }
 548 }
 549
 550 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 551
 552 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
 553
 554 static inline int next_prio(struct rq *rq)
 555 {
 556         struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
 557
 558         if (next && rt_prio(next->prio))
 559                 return next->prio;
 560         else
 561                 return MAX_RT_PRIO;
 562 }
 563 #endif
 564
 565 static inline
 566 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 567 {
 568         int prio = rt_se_prio(rt_se);
 569 #ifdef CONFIG_SMP
 570         struct rq *rq = rq_of_rt_rq(rt_rq);
 571 #endif
 572
 573         WARN_ON(!rt_prio(prio));
 574         rt_rq->rt_nr_running++;
 575 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 576         if (prio < rt_rq->highest_prio.curr) {
 577
 578                 /*
 579                  * If the new task is higher in priority than anything on the
 580                  * run-queue, we have a new high that must be published to
 581                  * the world.  We also know that the previous high becomes
 582                  * our next-highest.
 583                  */
 584                 rt_rq->highest_prio.next = rt_rq->highest_prio.curr;
 585                 rt_rq->highest_prio.curr = prio;
 586 #ifdef CONFIG_SMP
 587                 if (rq->online)
 588                         cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
 589 #endif
 590         } else if (prio == rt_rq->highest_prio.curr)
 591                 /*
 592                  * If the next task is equal in priority to the highest on
 593                  * the run-queue, then we implicitly know that the next highest
 594                  * task cannot be any lower than current
 595                  */
 596                 rt_rq->highest_prio.next = prio;
 597         else if (prio < rt_rq->highest_prio.next)
 598                 /*
 599                  * Otherwise, we need to recompute next-highest
 600                  */
 601                 rt_rq->highest_prio.next = next_prio(rq);
 602 #endif
 603 #ifdef CONFIG_SMP
 604         if (rt_se->nr_cpus_allowed > 1)
 605                 rq->rt.rt_nr_migratory++;
 606
 607         update_rt_migration(rq);
 608 #endif
 609 #ifdef CONFIG_RT_GROUP_SCHED
 610         if (rt_se_boosted(rt_se))
 611                 rt_rq->rt_nr_boosted++;
 612
 613         if (rt_rq->tg)
 614                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 615 #else
 616         start_rt_bandwidth(&def_rt_bandwidth);
 617 #endif
 618 }
 619
 620 static inline
 621 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 622 {
 623 #ifdef CONFIG_SMP
 624         struct rq *rq = rq_of_rt_rq(rt_rq);
 625         int highest_prio = rt_rq->highest_prio.curr;
 626 #endif
 627
 628         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 629         WARN_ON(!rt_rq->rt_nr_running);
 630         rt_rq->rt_nr_running--;
 631 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 632         if (rt_rq->rt_nr_running) {
 633                 int prio = rt_se_prio(rt_se);
 634
 635                 WARN_ON(prio < rt_rq->highest_prio.curr);
 636
 637                 /*
 638                  * This may have been our highest or next-highest priority
 639                  * task and therefore we may have some recomputation to do
 640                  */
 641                 if (prio == rt_rq->highest_prio.curr) {
 642                         struct rt_prio_array *array = &rt_rq->active;
 643
 644                         rt_rq->highest_prio.curr =
 645                                 sched_find_first_bit(array->bitmap);
 646                 }
 647
 648                 if (prio <= rt_rq->highest_prio.next)
 649                         rt_rq->highest_prio.next = next_prio(rq);
 650         } else
 651                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
 652 #endif
 653 #ifdef CONFIG_SMP
 654         if (rt_se->nr_cpus_allowed > 1)
 655                 rq->rt.rt_nr_migratory--;
 656
 657         if (rq->online && rt_rq->highest_prio.curr != highest_prio)
 658                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
 659
 660         update_rt_migration(rq);
 661 #endif /* CONFIG_SMP */
 662 #ifdef CONFIG_RT_GROUP_SCHED
 663         if (rt_se_boosted(rt_se))
 664                 rt_rq->rt_nr_boosted--;
 665
 666         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 667 #endif
 668 }
 669
 670 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
 671 {
 672         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 673         struct rt_prio_array *array = &rt_rq->active;
 674         struct rt_rq *group_rq = group_rt_rq(rt_se);
 675         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 676
 677         /*
 678          * Don't enqueue the group if its throttled, or when empty.
 679          * The latter is a consequence of the former when a child group
 680          * get throttled and the current group doesn't have any other
 681          * active members.
 682          */
 683         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 684                 return;
 685
 686         list_add_tail(&rt_se->run_list, queue);
 687         __set_bit(rt_se_prio(rt_se), array->bitmap);
 688
 689         inc_rt_tasks(rt_se, rt_rq);
 690 }
 691
 692 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 693 {
 694         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 695         struct rt_prio_array *array = &rt_rq->active;
 696
 697         list_del_init(&rt_se->run_list);
 698         if (list_empty(array->queue + rt_se_prio(rt_se)))
 699                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 700
 701         dec_rt_tasks(rt_se, rt_rq);
 702 }
 703
 704 /*
 705  * Because the prio of an upper entry depends on the lower
 706  * entries, we must remove entries top - down.
 707  */
 708 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 709 {
 710         struct sched_rt_entity *back = NULL;
 711
 712         for_each_sched_rt_entity(rt_se) {
 713                 rt_se->back = back;
 714                 back = rt_se;
 715         }
 716
 717         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 718                 if (on_rt_rq(rt_se))
 719                         __dequeue_rt_entity(rt_se);
 720         }
 721 }
 722
 723 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
 724 {
 725         dequeue_rt_stack(rt_se);
 726         for_each_sched_rt_entity(rt_se)
 727                 __enqueue_rt_entity(rt_se);
 728 }
 729
 730 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 731 {
 732         dequeue_rt_stack(rt_se);
 733
 734         for_each_sched_rt_entity(rt_se) {
 735                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 736
 737                 if (rt_rq && rt_rq->rt_nr_running)
 738                         __enqueue_rt_entity(rt_se);
 739         }
 740 }
 741
 742 /*
 743  * Adding/removing a task to/from a priority array:
 744  */
 745 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 746 {
 747         struct sched_rt_entity *rt_se = &p->rt;
 748
 749         if (wakeup)
 750                 rt_se->timeout = 0;
 751
 752         enqueue_rt_entity(rt_se);
 753
 754         inc_cpu_load(rq, p->se.load.weight);
 755 }
 756
 757 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 758 {
 759         struct sched_rt_entity *rt_se = &p->rt;
 760
 761         update_curr_rt(rq);
 762         dequeue_rt_entity(rt_se);
 763
 764         dec_cpu_load(rq, p->se.load.weight);
 765 }
 766
 767 /*
 768  * Put task to the end of the run list without the overhead of dequeue
 769  * followed by enqueue.
 770  */
 771 static void
 772 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 773 {
 774         if (on_rt_rq(rt_se)) {
 775                 struct rt_prio_array *array = &rt_rq->active;
 776                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 777
 778                 if (head)
 779                         list_move(&rt_se->run_list, queue);
 780                 else
 781                         list_move_tail(&rt_se->run_list, queue);
 782         }
 783 }
 784
 785 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 786 {
 787         struct sched_rt_entity *rt_se = &p->rt;
 788         struct rt_rq *rt_rq;
 789
 790         for_each_sched_rt_entity(rt_se) {
 791                 rt_rq = rt_rq_of_se(rt_se);
 792                 requeue_rt_entity(rt_rq, rt_se, head);
 793         }
 794 }
 795
 796 static void yield_task_rt(struct rq *rq)
 797 {
 798         requeue_task_rt(rq, rq->curr, 0);
 799 }
 800
 801 #ifdef CONFIG_SMP
 802 static int find_lowest_rq(struct task_struct *task);
 803
 804 static int select_task_rq_rt(struct task_struct *p, int sync)
 805 {
 806         struct rq *rq = task_rq(p);
 807
 808         /*
 809          * If the current task is an RT task, then
 810          * try to see if we can wake this RT task up on another
 811          * runqueue. Otherwise simply start this RT task
 812          * on its current runqueue.
 813          *
 814          * We want to avoid overloading runqueues. Even if
 815          * the RT task is of higher priority than the current RT task.
 816          * RT tasks behave differently than other tasks. If
 817          * one gets preempted, we try to push it off to another queue.
 818          * So trying to keep a preempting RT task on the same
 819          * cache hot CPU will force the running RT task to
 820          * a cold CPU. So we waste all the cache for the lower
 821          * RT task in hopes of saving some of a RT task
 822          * that is just being woken and probably will have
 823          * cold cache anyway.
 824          */
 825         if (unlikely(rt_task(rq->curr)) &&
 826             (p->rt.nr_cpus_allowed > 1)) {
 827                 int cpu = find_lowest_rq(p);
 828
 829                 return (cpu == -1) ? task_cpu(p) : cpu;
 830         }
 831
 832         /*
 833          * Otherwise, just let it ride on the affined RQ and the
 834          * post-schedule router will push the preempted task away
 835          */
 836         return task_cpu(p);
 837 }
 838
 839 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
 840 {
 841         cpumask_var_t mask;
 842
 843         if (rq->curr->rt.nr_cpus_allowed == 1)
 844                 return;
 845
 846         if (!alloc_cpumask_var(&mask, GFP_ATOMIC))
 847                 return;
 848
 849         if (p->rt.nr_cpus_allowed != 1
 850             && cpupri_find(&rq->rd->cpupri, p, mask))
 851                 goto free;
 852
 853         if (!cpupri_find(&rq->rd->cpupri, rq->curr, mask))
 854                 goto free;
 855
 856         /*
 857          * There appears to be other cpus that can accept
 858          * current and none to run 'p', so lets reschedule
 859          * to try and push current away:
 860          */
 861         requeue_task_rt(rq, p, 1);
 862         resched_task(rq->curr);
 863 free:
 864         free_cpumask_var(mask);
 865 }
 866
 867 #endif /* CONFIG_SMP */
 868
 869 /*
 870  * Preempt the current task with a newly woken task if needed:
 871  */
 872 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
 873 {
 874         if (p->prio < rq->curr->prio) {
 875                 resched_task(rq->curr);
 876                 return;
 877         }
 878
 879 #ifdef CONFIG_SMP
 880         /*
 881          * If:
 882          *
 883          * - the newly woken task is of equal priority to the current task
 884          * - the newly woken task is non-migratable while current is migratable
 885          * - current will be preempted on the next reschedule
 886          *
 887          * we should check to see if current can readily move to a different
 888          * cpu.  If so, we will reschedule to allow the push logic to try
 889          * to move current somewhere else, making room for our non-migratable
 890          * task.
 891          */
 892         if (p->prio == rq->curr->prio && !need_resched())
 893                 check_preempt_equal_prio(rq, p);
 894 #endif
 895 }
 896
 897 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
 898                                                    struct rt_rq *rt_rq)
 899 {
 900         struct rt_prio_array *array = &rt_rq->active;
 901         struct sched_rt_entity *next = NULL;
 902         struct list_head *queue;
 903         int idx;
 904
 905         idx = sched_find_first_bit(array->bitmap);
 906         BUG_ON(idx >= MAX_RT_PRIO);
 907
 908         queue = array->queue + idx;
 909         next = list_entry(queue->next, struct sched_rt_entity, run_list);
 910
 911         return next;
 912 }
 913
 914 static struct task_struct *pick_next_task_rt(struct rq *rq)
 915 {
 916         struct sched_rt_entity *rt_se;
 917         struct task_struct *p;
 918         struct rt_rq *rt_rq;
 919
 920         rt_rq = &rq->rt;
 921
 922         if (unlikely(!rt_rq->rt_nr_running))
 923                 return NULL;
 924
 925         if (rt_rq_throttled(rt_rq))
 926                 return NULL;
 927
 928         do {
 929                 rt_se = pick_next_rt_entity(rq, rt_rq);
 930                 BUG_ON(!rt_se);
 931                 rt_rq = group_rt_rq(rt_se);
 932         } while (rt_rq);
 933
 934         p = rt_task_of(rt_se);
 935         p->se.exec_start = rq->clock;
 936         return p;
 937 }
 938
 939 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 940 {
 941         update_curr_rt(rq);
 942         p->se.exec_start = 0;
 943 }
 944
 945 #ifdef CONFIG_SMP
 946
 947 /* Only try algorithms three times */
 948 #define RT_MAX_TRIES 3
 949
 950 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 951
 952 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 953 {
 954         if (!task_running(rq, p) &&
 955             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
 956             (p->rt.nr_cpus_allowed > 1))
 957                 return 1;
 958         return 0;
 959 }
 960
 961 /* Return the second highest RT task, NULL otherwise */
 962 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
 963 {
 964         struct task_struct *next = NULL;
 965         struct sched_rt_entity *rt_se;
 966         struct rt_prio_array *array;
 967         struct rt_rq *rt_rq;
 968         int idx;
 969
 970         for_each_leaf_rt_rq(rt_rq, rq) {
 971                 array = &rt_rq->active;
 972                 idx = sched_find_first_bit(array->bitmap);
 973  next_idx:
 974                 if (idx >= MAX_RT_PRIO)
 975                         continue;
 976                 if (next && next->prio < idx)
 977                         continue;
 978                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
 979                         struct task_struct *p = rt_task_of(rt_se);
 980                         if (pick_rt_task(rq, p, cpu)) {
 981                                 next = p;
 982                                 break;
 983                         }
 984                 }
 985                 if (!next) {
 986                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 987                         goto next_idx;
 988                 }
 989         }
 990
 991         return next;
 992 }
 993
 994 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
 995
 996 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
 997 {
 998         int first;
 999
1000         /* "this_cpu" is cheaper to preempt than a remote processor */
1001         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
1002                 return this_cpu;
1003
1004         first = first_cpu(*mask);
1005         if (first != NR_CPUS)
1006                 return first;
1007
1008         return -1;
1009 }
1010
1011 static int find_lowest_rq(struct task_struct *task)
1012 {
1013         struct sched_domain *sd;
1014         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1015         int this_cpu = smp_processor_id();
1016         int cpu      = task_cpu(task);
1017
1018         if (task->rt.nr_cpus_allowed == 1)
1019                 return -1; /* No other targets possible */
1020
1021         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1022                 return -1; /* No targets found */
1023
1024         /*
1025          * Only consider CPUs that are usable for migration.
1026          * I guess we might want to change cpupri_find() to ignore those
1027          * in the first place.
1028          */
1029         cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
1030
1031         /*
1032          * At this point we have built a mask of cpus representing the
1033          * lowest priority tasks in the system.  Now we want to elect
1034          * the best one based on our affinity and topology.
1035          *
1036          * We prioritize the last cpu that the task executed on since
1037          * it is most likely cache-hot in that location.
1038          */
1039         if (cpumask_test_cpu(cpu, lowest_mask))
1040                 return cpu;
1041
1042         /*
1043          * Otherwise, we consult the sched_domains span maps to figure
1044          * out which cpu is logically closest to our hot cache data.
1045          */
1046         if (this_cpu == cpu)
1047                 this_cpu = -1; /* Skip this_cpu opt if the same */
1048
1049         for_each_domain(cpu, sd) {
1050                 if (sd->flags & SD_WAKE_AFFINE) {
1051                         cpumask_t domain_mask;
1052                         int       best_cpu;
1053
1054                         cpumask_and(&domain_mask, sched_domain_span(sd),
1055                                     lowest_mask);
1056
1057                         best_cpu = pick_optimal_cpu(this_cpu,
1058                                                     &domain_mask);
1059                         if (best_cpu != -1)
1060                                 return best_cpu;
1061                 }
1062         }
1063
1064         /*
1065          * And finally, if there were no matches within the domains
1066          * just give the caller *something* to work with from the compatible
1067          * locations.
1068          */
1069         return pick_optimal_cpu(this_cpu, lowest_mask);
1070 }
1071
1072 /* Will lock the rq it finds */
1073 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1074 {
1075         struct rq *lowest_rq = NULL;
1076         int tries;
1077         int cpu;
1078
1079         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1080                 cpu = find_lowest_rq(task);
1081
1082                 if ((cpu == -1) || (cpu == rq->cpu))
1083                         break;
1084
1085                 lowest_rq = cpu_rq(cpu);
1086
1087                 /* if the prio of this runqueue changed, try again */
1088                 if (double_lock_balance(rq, lowest_rq)) {
1089                         /*
1090                          * We had to unlock the run queue. In
1091                          * the mean time, task could have
1092                          * migrated already or had its affinity changed.
1093                          * Also make sure that it wasn't scheduled on its rq.
1094                          */
1095                         if (unlikely(task_rq(task) != rq ||
1096                                      !cpumask_test_cpu(lowest_rq->cpu,
1097                                                        &task->cpus_allowed) ||
1098                                      task_running(rq, task) ||
1099                                      !task->se.on_rq)) {
1100
1101                                 spin_unlock(&lowest_rq->lock);
1102                                 lowest_rq = NULL;
1103                                 break;
1104                         }
1105                 }
1106
1107                 /* If this rq is still suitable use it. */
1108                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1109                         break;
1110
1111                 /* try again */
1112                 double_unlock_balance(rq, lowest_rq);
1113                 lowest_rq = NULL;
1114         }
1115
1116         return lowest_rq;
1117 }
1118
1119 /*
1120  * If the current CPU has more than one RT task, see if the non
1121  * running task can migrate over to a CPU that is running a task
1122  * of lesser priority.
1123  */
1124 static int push_rt_task(struct rq *rq)
1125 {
1126         struct task_struct *next_task;
1127         struct rq *lowest_rq;
1128         int ret = 0;
1129         int paranoid = RT_MAX_TRIES;
1130
1131         if (!rq->rt.overloaded)
1132                 return 0;
1133
1134         next_task = pick_next_highest_task_rt(rq, -1);
1135         if (!next_task)
1136                 return 0;
1137
1138  retry:
1139         if (unlikely(next_task == rq->curr)) {
1140                 WARN_ON(1);
1141                 return 0;
1142         }
1143
1144         /*
1145          * It's possible that the next_task slipped in of
1146          * higher priority than current. If that's the case
1147          * just reschedule current.
1148          */
1149         if (unlikely(next_task->prio < rq->curr->prio)) {
1150                 resched_task(rq->curr);
1151                 return 0;
1152         }
1153
1154         /* We might release rq lock */
1155         get_task_struct(next_task);
1156
1157         /* find_lock_lowest_rq locks the rq if found */
1158         lowest_rq = find_lock_lowest_rq(next_task, rq);
1159         if (!lowest_rq) {
1160                 struct task_struct *task;
1161                 /*
1162                  * find lock_lowest_rq releases rq->lock
1163                  * so it is possible that next_task has changed.
1164                  * If it has, then try again.
1165                  */
1166                 task = pick_next_highest_task_rt(rq, -1);
1167                 if (unlikely(task != next_task) && task && paranoid--) {
1168                         put_task_struct(next_task);
1169                         next_task = task;
1170                         goto retry;
1171                 }
1172                 goto out;
1173         }
1174
1175         deactivate_task(rq, next_task, 0);
1176         set_task_cpu(next_task, lowest_rq->cpu);
1177         activate_task(lowest_rq, next_task, 0);
1178
1179         resched_task(lowest_rq->curr);
1180
1181         double_unlock_balance(rq, lowest_rq);
1182
1183         ret = 1;
1184 out:
1185         put_task_struct(next_task);
1186
1187         return ret;
1188 }
1189
1190 /*
1191  * TODO: Currently we just use the second highest prio task on
1192  *       the queue, and stop when it can't migrate (or there's
1193  *       no more RT tasks).  There may be a case where a lower
1194  *       priority RT task has a different affinity than the
1195  *       higher RT task. In this case the lower RT task could
1196  *       possibly be able to migrate where as the higher priority
1197  *       RT task could not.  We currently ignore this issue.
1198  *       Enhancements are welcome!
1199  */
1200 static void push_rt_tasks(struct rq *rq)
1201 {
1202         /* push_rt_task will return true if it moved an RT */
1203         while (push_rt_task(rq))
1204                 ;
1205 }
1206
1207 static int pull_rt_task(struct rq *this_rq)
1208 {
1209         int this_cpu = this_rq->cpu, ret = 0, cpu;
1210         struct task_struct *p, *next;
1211         struct rq *src_rq;
1212
1213         if (likely(!rt_overloaded(this_rq)))
1214                 return 0;
1215
1216         next = pick_next_task_rt(this_rq);
1217
1218         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1219                 if (this_cpu == cpu)
1220                         continue;
1221
1222                 src_rq = cpu_rq(cpu);
1223                 /*
1224                  * We can potentially drop this_rq's lock in
1225                  * double_lock_balance, and another CPU could
1226                  * steal our next task - hence we must cause
1227                  * the caller to recalculate the next task
1228                  * in that case:
1229                  */
1230                 if (double_lock_balance(this_rq, src_rq)) {
1231                         struct task_struct *old_next = next;
1232
1233                         next = pick_next_task_rt(this_rq);
1234                         if (next != old_next)
1235                                 ret = 1;
1236                 }
1237
1238                 /*
1239                  * Are there still pullable RT tasks?
1240                  */
1241                 if (src_rq->rt.rt_nr_running <= 1)
1242                         goto skip;
1243
1244                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1245
1246                 /*
1247                  * Do we have an RT task that preempts
1248                  * the to-be-scheduled task?
1249                  */
1250                 if (p && (!next || (p->prio < next->prio))) {
1251                         WARN_ON(p == src_rq->curr);
1252                         WARN_ON(!p->se.on_rq);
1253
1254                         /*
1255                          * There's a chance that p is higher in priority
1256                          * than what's currently running on its cpu.
1257                          * This is just that p is wakeing up and hasn't
1258                          * had a chance to schedule. We only pull
1259                          * p if it is lower in priority than the
1260                          * current task on the run queue or
1261                          * this_rq next task is lower in prio than
1262                          * the current task on that rq.
1263                          */
1264                         if (p->prio < src_rq->curr->prio ||
1265                             (next && next->prio < src_rq->curr->prio))
1266                                 goto skip;
1267
1268                         ret = 1;
1269
1270                         deactivate_task(src_rq, p, 0);
1271                         set_task_cpu(p, this_cpu);
1272                         activate_task(this_rq, p, 0);
1273                         /*
1274                          * We continue with the search, just in
1275                          * case there's an even higher prio task
1276                          * in another runqueue. (low likelyhood
1277                          * but possible)
1278                          *
1279                          * Update next so that we won't pick a task
1280                          * on another cpu with a priority lower (or equal)
1281                          * than the one we just picked.
1282                          */
1283                         next = p;
1284
1285                 }
1286  skip:
1287                 double_unlock_balance(this_rq, src_rq);
1288         }
1289
1290         return ret;
1291 }
1292
1293 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1294 {
1295         /* Try to pull RT tasks here if we lower this rq's prio */
1296         if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1297                 pull_rt_task(rq);
1298 }
1299
1300 static void post_schedule_rt(struct rq *rq)
1301 {
1302         /*
1303          * If we have more than one rt_task queued, then
1304          * see if we can push the other rt_tasks off to other CPUS.
1305          * Note we may release the rq lock, and since
1306          * the lock was owned by prev, we need to release it
1307          * first via finish_lock_switch and then reaquire it here.
1308          */
1309         if (unlikely(rq->rt.overloaded)) {
1310                 spin_lock_irq(&rq->lock);
1311                 push_rt_tasks(rq);
1312                 spin_unlock_irq(&rq->lock);
1313         }
1314 }
1315
1316 /*
1317  * If we are not running and we are not going to reschedule soon, we should
1318  * try to push tasks away now
1319  */
1320 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1321 {
1322         if (!task_running(rq, p) &&
1323             !test_tsk_need_resched(rq->curr) &&
1324             rq->rt.overloaded)
1325                 push_rt_tasks(rq);
1326 }
1327
1328 static unsigned long
1329 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1330                 unsigned long max_load_move,
1331                 struct sched_domain *sd, enum cpu_idle_type idle,
1332                 int *all_pinned, int *this_best_prio)
1333 {
1334         /* don't touch RT tasks */
1335         return 0;
1336 }
1337
1338 static int
1339 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1340                  struct sched_domain *sd, enum cpu_idle_type idle)
1341 {
1342         /* don't touch RT tasks */
1343         return 0;
1344 }
1345
1346 static void set_cpus_allowed_rt(struct task_struct *p,
1347                                 const struct cpumask *new_mask)
1348 {
1349         int weight = cpumask_weight(new_mask);
1350
1351         BUG_ON(!rt_task(p));
1352
1353         /*
1354          * Update the migration status of the RQ if we have an RT task
1355          * which is running AND changing its weight value.
1356          */
1357         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1358                 struct rq *rq = task_rq(p);
1359
1360                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1361                         rq->rt.rt_nr_migratory++;
1362                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1363                         BUG_ON(!rq->rt.rt_nr_migratory);
1364                         rq->rt.rt_nr_migratory--;
1365                 }
1366
1367                 update_rt_migration(rq);
1368         }
1369
1370         cpumask_copy(&p->cpus_allowed, new_mask);
1371         p->rt.nr_cpus_allowed = weight;
1372 }
1373
1374 /* Assumes rq->lock is held */
1375 static void rq_online_rt(struct rq *rq)
1376 {
1377         if (rq->rt.overloaded)
1378                 rt_set_overload(rq);
1379
1380         __enable_runtime(rq);
1381
1382         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1383 }
1384
1385 /* Assumes rq->lock is held */
1386 static void rq_offline_rt(struct rq *rq)
1387 {
1388         if (rq->rt.overloaded)
1389                 rt_clear_overload(rq);
1390
1391         __disable_runtime(rq);
1392
1393         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1394 }
1395
1396 /*
1397  * When switch from the rt queue, we bring ourselves to a position
1398  * that we might want to pull RT tasks from other runqueues.
1399  */
1400 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1401                            int running)
1402 {
1403         /*
1404          * If there are other RT tasks then we will reschedule
1405          * and the scheduling of the other RT tasks will handle
1406          * the balancing. But if we are the last RT task
1407          * we may need to handle the pulling of RT tasks
1408          * now.
1409          */
1410         if (!rq->rt.rt_nr_running)
1411                 pull_rt_task(rq);
1412 }
1413
1414 static inline void init_sched_rt_class(void)
1415 {
1416         unsigned int i;
1417
1418         for_each_possible_cpu(i)
1419                 alloc_cpumask_var(&per_cpu(local_cpu_mask, i), GFP_KERNEL);
1420 }
1421 #endif /* CONFIG_SMP */
1422
1423 /*
1424  * When switching a task to RT, we may overload the runqueue
1425  * with RT tasks. In this case we try to push them off to
1426  * other runqueues.
1427  */
1428 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1429                            int running)
1430 {
1431         int check_resched = 1;
1432
1433         /*
1434          * If we are already running, then there's nothing
1435          * that needs to be done. But if we are not running
1436          * we may need to preempt the current running task.
1437          * If that current running task is also an RT task
1438          * then see if we can move to another run queue.
1439          */
1440         if (!running) {
1441 #ifdef CONFIG_SMP
1442                 if (rq->rt.overloaded && push_rt_task(rq) &&
1443                     /* Don't resched if we changed runqueues */
1444                     rq != task_rq(p))
1445                         check_resched = 0;
1446 #endif /* CONFIG_SMP */
1447                 if (check_resched && p->prio < rq->curr->prio)
1448                         resched_task(rq->curr);
1449         }
1450 }
1451
1452 /*
1453  * Priority of the task has changed. This may cause
1454  * us to initiate a push or pull.
1455  */
1456 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1457                             int oldprio, int running)
1458 {
1459         if (running) {
1460 #ifdef CONFIG_SMP
1461                 /*
1462                  * If our priority decreases while running, we
1463                  * may need to pull tasks to this runqueue.
1464                  */
1465                 if (oldprio < p->prio)
1466                         pull_rt_task(rq);
1467                 /*
1468                  * If there's a higher priority task waiting to run
1469                  * then reschedule. Note, the above pull_rt_task
1470                  * can release the rq lock and p could migrate.
1471                  * Only reschedule if p is still on the same runqueue.
1472                  */
1473                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1474                         resched_task(p);
1475 #else
1476                 /* For UP simply resched on drop of prio */
1477                 if (oldprio < p->prio)
1478                         resched_task(p);
1479 #endif /* CONFIG_SMP */
1480         } else {
1481                 /*
1482                  * This task is not running, but if it is
1483                  * greater than the current running task
1484                  * then reschedule.
1485                  */
1486                 if (p->prio < rq->curr->prio)
1487                         resched_task(rq->curr);
1488         }
1489 }
1490
1491 static void watchdog(struct rq *rq, struct task_struct *p)
1492 {
1493         unsigned long soft, hard;
1494
1495         if (!p->signal)
1496                 return;
1497
1498         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1499         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1500
1501         if (soft != RLIM_INFINITY) {
1502                 unsigned long next;
1503
1504                 p->rt.timeout++;
1505                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1506                 if (p->rt.timeout > next)
1507                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1508         }
1509 }
1510
1511 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1512 {
1513         update_curr_rt(rq);
1514
1515         watchdog(rq, p);
1516
1517         /*
1518          * RR tasks need a special form of timeslice management.
1519          * FIFO tasks have no timeslices.
1520          */
1521         if (p->policy != SCHED_RR)
1522                 return;
1523
1524         if (--p->rt.time_slice)
1525                 return;
1526
1527         p->rt.time_slice = DEF_TIMESLICE;
1528
1529         /*
1530          * Requeue to the end of queue if we are not the only element
1531          * on the queue:
1532          */
1533         if (p->rt.run_list.prev != p->rt.run_list.next) {
1534                 requeue_task_rt(rq, p, 0);
1535                 set_tsk_need_resched(p);
1536         }
1537 }
1538
1539 static void set_curr_task_rt(struct rq *rq)
1540 {
1541         struct task_struct *p = rq->curr;
1542
1543         p->se.exec_start = rq->clock;
1544 }
1545
1546 static const struct sched_class rt_sched_class = {
1547         .next                   = &fair_sched_class,
1548         .enqueue_task           = enqueue_task_rt,
1549         .dequeue_task           = dequeue_task_rt,
1550         .yield_task             = yield_task_rt,
1551
1552         .check_preempt_curr     = check_preempt_curr_rt,
1553
1554         .pick_next_task         = pick_next_task_rt,
1555         .put_prev_task          = put_prev_task_rt,
1556
1557 #ifdef CONFIG_SMP
1558         .select_task_rq         = select_task_rq_rt,
1559
1560         .load_balance           = load_balance_rt,
1561         .move_one_task          = move_one_task_rt,
1562         .set_cpus_allowed       = set_cpus_allowed_rt,
1563         .rq_online              = rq_online_rt,
1564         .rq_offline             = rq_offline_rt,
1565         .pre_schedule           = pre_schedule_rt,
1566         .post_schedule          = post_schedule_rt,
1567         .task_wake_up           = task_wake_up_rt,
1568         .switched_from          = switched_from_rt,
1569 #endif
1570
1571         .set_curr_task          = set_curr_task_rt,
1572         .task_tick              = task_tick_rt,
1573
1574         .prio_changed           = prio_changed_rt,
1575         .switched_to            = switched_to_rt,
1576 };
1577
1578 #ifdef CONFIG_SCHED_DEBUG
1579 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1580
1581 static void print_rt_stats(struct seq_file *m, int cpu)
1582 {
1583         struct rt_rq *rt_rq;
1584
1585         rcu_read_lock();
1586         for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1587                 print_rt_rq(m, cpu, rt_rq);
1588         rcu_read_unlock();
1589 }
1590 #endif /* CONFIG_SCHED_DEBUG */
1591