kernel/sched_rt.c

   1 /*
   2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
   3  * policies)
   4  */
   5
   6 #ifdef CONFIG_RT_GROUP_SCHED
   7
   8 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
   9
  10 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  11 {
  12 #ifdef CONFIG_SCHED_DEBUG
  13         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
  14 #endif
  15         return container_of(rt_se, struct task_struct, rt);
  16 }
  17
  18 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
  19 {
  20         return rt_rq->rq;
  21 }
  22
  23 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
  24 {
  25         return rt_se->rt_rq;
  26 }
  27
  28 #else /* CONFIG_RT_GROUP_SCHED */
  29
  30 #define rt_entity_is_task(rt_se) (1)
  31
  32 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
  33 {
  34         return container_of(rt_se, struct task_struct, rt);
  35 }
  36
  37 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
  38 {
  39         return container_of(rt_rq, struct rq, rt);
  40 }
  41
  42 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
  43 {
  44         struct task_struct *p = rt_task_of(rt_se);
  45         struct rq *rq = task_rq(p);
  46
  47         return &rq->rt;
  48 }
  49
  50 #endif /* CONFIG_RT_GROUP_SCHED */
  51
  52 #ifdef CONFIG_SMP
  53
  54 static inline int rt_overloaded(struct rq *rq)
  55 {
  56         return atomic_read(&rq->rd->rto_count);
  57 }
  58
  59 static inline void rt_set_overload(struct rq *rq)
  60 {
  61         if (!rq->online)
  62                 return;
  63
  64         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
  65         /*
  66          * Make sure the mask is visible before we set
  67          * the overload count. That is checked to determine
  68          * if we should look at the mask. It would be a shame
  69          * if we looked at the mask, but the mask was not
  70          * updated yet.
  71          */
  72         wmb();
  73         atomic_inc(&rq->rd->rto_count);
  74 }
  75
  76 static inline void rt_clear_overload(struct rq *rq)
  77 {
  78         if (!rq->online)
  79                 return;
  80
  81         /* the order here really doesn't matter */
  82         atomic_dec(&rq->rd->rto_count);
  83         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
  84 }
  85
  86 static void update_rt_migration(struct rt_rq *rt_rq)
  87 {
  88         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
  89                 if (!rt_rq->overloaded) {
  90                         rt_set_overload(rq_of_rt_rq(rt_rq));
  91                         rt_rq->overloaded = 1;
  92                 }
  93         } else if (rt_rq->overloaded) {
  94                 rt_clear_overload(rq_of_rt_rq(rt_rq));
  95                 rt_rq->overloaded = 0;
  96         }
  97 }
  98
  99 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 100 {
 101         if (!rt_entity_is_task(rt_se))
 102                 return;
 103
 104         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
 105
 106         rt_rq->rt_nr_total++;
 107         if (rt_se->nr_cpus_allowed > 1)
 108                 rt_rq->rt_nr_migratory++;
 109
 110         update_rt_migration(rt_rq);
 111 }
 112
 113 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 114 {
 115         if (!rt_entity_is_task(rt_se))
 116                 return;
 117
 118         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
 119
 120         rt_rq->rt_nr_total--;
 121         if (rt_se->nr_cpus_allowed > 1)
 122                 rt_rq->rt_nr_migratory--;
 123
 124         update_rt_migration(rt_rq);
 125 }
 126
 127 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 128 {
 129         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 130         plist_node_init(&p->pushable_tasks, p->prio);
 131         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
 132 }
 133
 134 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 135 {
 136         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
 137 }
 138
 139 static inline int has_pushable_tasks(struct rq *rq)
 140 {
 141         return !plist_head_empty(&rq->rt.pushable_tasks);
 142 }
 143
 144 #else
 145
 146 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
 147 {
 148 }
 149
 150 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
 151 {
 152 }
 153
 154 static inline
 155 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 156 {
 157 }
 158
 159 static inline
 160 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 161 {
 162 }
 163
 164 #endif /* CONFIG_SMP */
 165
 166 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
 167 {
 168         return !list_empty(&rt_se->run_list);
 169 }
 170
 171 #ifdef CONFIG_RT_GROUP_SCHED
 172
 173 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 174 {
 175         if (!rt_rq->tg)
 176                 return RUNTIME_INF;
 177
 178         return rt_rq->rt_runtime;
 179 }
 180
 181 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 182 {
 183         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
 184 }
 185
 186 typedef struct task_group *rt_rq_iter_t;
 187
 188 #define for_each_rt_rq(rt_rq, iter, rq) \
 189         for (iter = list_entry_rcu(task_groups.next, typeof(*iter), list); \
 190              (&iter->list != &task_groups) && \
 191              (rt_rq = iter->rt_rq[cpu_of(rq)]); \
 192              iter = list_entry_rcu(iter->list.next, typeof(*iter), list))
 193
 194 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
 195 {
 196         list_add_rcu(&rt_rq->leaf_rt_rq_list,
 197                         &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
 198 }
 199
 200 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
 201 {
 202         list_del_rcu(&rt_rq->leaf_rt_rq_list);
 203 }
 204
 205 #define for_each_leaf_rt_rq(rt_rq, rq) \
 206         list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
 207
 208 #define for_each_sched_rt_entity(rt_se) \
 209         for (; rt_se; rt_se = rt_se->parent)
 210
 211 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 212 {
 213         return rt_se->my_q;
 214 }
 215
 216 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
 217 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
 218
 219 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 220 {
 221         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
 222         struct sched_rt_entity *rt_se;
 223
 224         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 225
 226         rt_se = rt_rq->tg->rt_se[cpu];
 227
 228         if (rt_rq->rt_nr_running) {
 229                 if (rt_se && !on_rt_rq(rt_se))
 230                         enqueue_rt_entity(rt_se, false);
 231                 if (rt_rq->highest_prio.curr < curr->prio)
 232                         resched_task(curr);
 233         }
 234 }
 235
 236 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 237 {
 238         struct sched_rt_entity *rt_se;
 239         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
 240
 241         rt_se = rt_rq->tg->rt_se[cpu];
 242
 243         if (rt_se && on_rt_rq(rt_se))
 244                 dequeue_rt_entity(rt_se);
 245 }
 246
 247 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 248 {
 249         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
 250 }
 251
 252 static int rt_se_boosted(struct sched_rt_entity *rt_se)
 253 {
 254         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 255         struct task_struct *p;
 256
 257         if (rt_rq)
 258                 return !!rt_rq->rt_nr_boosted;
 259
 260         p = rt_task_of(rt_se);
 261         return p->prio != p->normal_prio;
 262 }
 263
 264 #ifdef CONFIG_SMP
 265 static inline const struct cpumask *sched_rt_period_mask(void)
 266 {
 267         return cpu_rq(smp_processor_id())->rd->span;
 268 }
 269 #else
 270 static inline const struct cpumask *sched_rt_period_mask(void)
 271 {
 272         return cpu_online_mask;
 273 }
 274 #endif
 275
 276 static inline
 277 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 278 {
 279         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
 280 }
 281
 282 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 283 {
 284         return &rt_rq->tg->rt_bandwidth;
 285 }
 286
 287 #else /* !CONFIG_RT_GROUP_SCHED */
 288
 289 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
 290 {
 291         return rt_rq->rt_runtime;
 292 }
 293
 294 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
 295 {
 296         return ktime_to_ns(def_rt_bandwidth.rt_period);
 297 }
 298
 299 typedef struct rt_rq *rt_rq_iter_t;
 300
 301 #define for_each_rt_rq(rt_rq, iter, rq) \
 302         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 303
 304 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
 305 {
 306 }
 307
 308 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
 309 {
 310 }
 311
 312 #define for_each_leaf_rt_rq(rt_rq, rq) \
 313         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
 314
 315 #define for_each_sched_rt_entity(rt_se) \
 316         for (; rt_se; rt_se = NULL)
 317
 318 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
 319 {
 320         return NULL;
 321 }
 322
 323 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
 324 {
 325         if (rt_rq->rt_nr_running)
 326                 resched_task(rq_of_rt_rq(rt_rq)->curr);
 327 }
 328
 329 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
 330 {
 331 }
 332
 333 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
 334 {
 335         return rt_rq->rt_throttled;
 336 }
 337
 338 static inline const struct cpumask *sched_rt_period_mask(void)
 339 {
 340         return cpu_online_mask;
 341 }
 342
 343 static inline
 344 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
 345 {
 346         return &cpu_rq(cpu)->rt;
 347 }
 348
 349 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
 350 {
 351         return &def_rt_bandwidth;
 352 }
 353
 354 #endif /* CONFIG_RT_GROUP_SCHED */
 355
 356 #ifdef CONFIG_SMP
 357 /*
 358  * We ran out of runtime, see if we can borrow some from our neighbours.
 359  */
 360 static int do_balance_runtime(struct rt_rq *rt_rq)
 361 {
 362         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 363         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
 364         int i, weight, more = 0;
 365         u64 rt_period;
 366
 367         weight = cpumask_weight(rd->span);
 368
 369         raw_spin_lock(&rt_b->rt_runtime_lock);
 370         rt_period = ktime_to_ns(rt_b->rt_period);
 371         for_each_cpu(i, rd->span) {
 372                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 373                 s64 diff;
 374
 375                 if (iter == rt_rq)
 376                         continue;
 377
 378                 raw_spin_lock(&iter->rt_runtime_lock);
 379                 /*
 380                  * Either all rqs have inf runtime and there's nothing to steal
 381                  * or __disable_runtime() below sets a specific rq to inf to
 382                  * indicate its been disabled and disalow stealing.
 383                  */
 384                 if (iter->rt_runtime == RUNTIME_INF)
 385                         goto next;
 386
 387                 /*
 388                  * From runqueues with spare time, take 1/n part of their
 389                  * spare time, but no more than our period.
 390                  */
 391                 diff = iter->rt_runtime - iter->rt_time;
 392                 if (diff > 0) {
 393                         diff = div_u64((u64)diff, weight);
 394                         if (rt_rq->rt_runtime + diff > rt_period)
 395                                 diff = rt_period - rt_rq->rt_runtime;
 396                         iter->rt_runtime -= diff;
 397                         rt_rq->rt_runtime += diff;
 398                         more = 1;
 399                         if (rt_rq->rt_runtime == rt_period) {
 400                                 raw_spin_unlock(&iter->rt_runtime_lock);
 401                                 break;
 402                         }
 403                 }
 404 next:
 405                 raw_spin_unlock(&iter->rt_runtime_lock);
 406         }
 407         raw_spin_unlock(&rt_b->rt_runtime_lock);
 408
 409         return more;
 410 }
 411
 412 /*
 413  * Ensure this RQ takes back all the runtime it lend to its neighbours.
 414  */
 415 static void __disable_runtime(struct rq *rq)
 416 {
 417         struct root_domain *rd = rq->rd;
 418         rt_rq_iter_t iter;
 419         struct rt_rq *rt_rq;
 420
 421         if (unlikely(!scheduler_running))
 422                 return;
 423
 424         for_each_rt_rq(rt_rq, iter, rq) {
 425                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 426                 s64 want;
 427                 int i;
 428
 429                 raw_spin_lock(&rt_b->rt_runtime_lock);
 430                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 431                 /*
 432                  * Either we're all inf and nobody needs to borrow, or we're
 433                  * already disabled and thus have nothing to do, or we have
 434                  * exactly the right amount of runtime to take out.
 435                  */
 436                 if (rt_rq->rt_runtime == RUNTIME_INF ||
 437                                 rt_rq->rt_runtime == rt_b->rt_runtime)
 438                         goto balanced;
 439                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 440
 441                 /*
 442                  * Calculate the difference between what we started out with
 443                  * and what we current have, that's the amount of runtime
 444                  * we lend and now have to reclaim.
 445                  */
 446                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
 447
 448                 /*
 449                  * Greedy reclaim, take back as much as we can.
 450                  */
 451                 for_each_cpu(i, rd->span) {
 452                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
 453                         s64 diff;
 454
 455                         /*
 456                          * Can't reclaim from ourselves or disabled runqueues.
 457                          */
 458                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
 459                                 continue;
 460
 461                         raw_spin_lock(&iter->rt_runtime_lock);
 462                         if (want > 0) {
 463                                 diff = min_t(s64, iter->rt_runtime, want);
 464                                 iter->rt_runtime -= diff;
 465                                 want -= diff;
 466                         } else {
 467                                 iter->rt_runtime -= want;
 468                                 want -= want;
 469                         }
 470                         raw_spin_unlock(&iter->rt_runtime_lock);
 471
 472                         if (!want)
 473                                 break;
 474                 }
 475
 476                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 477                 /*
 478                  * We cannot be left wanting - that would mean some runtime
 479                  * leaked out of the system.
 480                  */
 481                 BUG_ON(want);
 482 balanced:
 483                 /*
 484                  * Disable all the borrow logic by pretending we have inf
 485                  * runtime - in which case borrowing doesn't make sense.
 486                  */
 487                 rt_rq->rt_runtime = RUNTIME_INF;
 488                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 489                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 490         }
 491 }
 492
 493 static void disable_runtime(struct rq *rq)
 494 {
 495         unsigned long flags;
 496
 497         raw_spin_lock_irqsave(&rq->lock, flags);
 498         __disable_runtime(rq);
 499         raw_spin_unlock_irqrestore(&rq->lock, flags);
 500 }
 501
 502 static void __enable_runtime(struct rq *rq)
 503 {
 504         rt_rq_iter_t iter;
 505         struct rt_rq *rt_rq;
 506
 507         if (unlikely(!scheduler_running))
 508                 return;
 509
 510         /*
 511          * Reset each runqueue's bandwidth settings
 512          */
 513         for_each_rt_rq(rt_rq, iter, rq) {
 514                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
 515
 516                 raw_spin_lock(&rt_b->rt_runtime_lock);
 517                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 518                 rt_rq->rt_runtime = rt_b->rt_runtime;
 519                 rt_rq->rt_time = 0;
 520                 rt_rq->rt_throttled = 0;
 521                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 522                 raw_spin_unlock(&rt_b->rt_runtime_lock);
 523         }
 524 }
 525
 526 static void enable_runtime(struct rq *rq)
 527 {
 528         unsigned long flags;
 529
 530         raw_spin_lock_irqsave(&rq->lock, flags);
 531         __enable_runtime(rq);
 532         raw_spin_unlock_irqrestore(&rq->lock, flags);
 533 }
 534
 535 static int balance_runtime(struct rt_rq *rt_rq)
 536 {
 537         int more = 0;
 538
 539         if (rt_rq->rt_time > rt_rq->rt_runtime) {
 540                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
 541                 more = do_balance_runtime(rt_rq);
 542                 raw_spin_lock(&rt_rq->rt_runtime_lock);
 543         }
 544
 545         return more;
 546 }
 547 #else /* !CONFIG_SMP */
 548 static inline int balance_runtime(struct rt_rq *rt_rq)
 549 {
 550         return 0;
 551 }
 552 #endif /* CONFIG_SMP */
 553
 554 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 555 {
 556         int i, idle = 1;
 557         const struct cpumask *span;
 558
 559         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
 560                 return 1;
 561
 562         span = sched_rt_period_mask();
 563         for_each_cpu(i, span) {
 564                 int enqueue = 0;
 565                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
 566                 struct rq *rq = rq_of_rt_rq(rt_rq);
 567
 568                 raw_spin_lock(&rq->lock);
 569                 if (rt_rq->rt_time) {
 570                         u64 runtime;
 571
 572                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 573                         if (rt_rq->rt_throttled)
 574                                 balance_runtime(rt_rq);
 575                         runtime = rt_rq->rt_runtime;
 576                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
 577                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
 578                                 rt_rq->rt_throttled = 0;
 579                                 enqueue = 1;
 580
 581                                 /*
 582                                  * Force a clock update if the CPU was idle,
 583                                  * lest wakeup -> unthrottle time accumulate.
 584                                  */
 585                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
 586                                         rq->skip_clock_update = -1;
 587                         }
 588                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
 589                                 idle = 0;
 590                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 591                 } else if (rt_rq->rt_nr_running) {
 592                         idle = 0;
 593                         if (!rt_rq_throttled(rt_rq))
 594                                 enqueue = 1;
 595                 }
 596
 597                 if (enqueue)
 598                         sched_rt_rq_enqueue(rt_rq);
 599                 raw_spin_unlock(&rq->lock);
 600         }
 601
 602         return idle;
 603 }
 604
 605 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
 606 {
 607 #ifdef CONFIG_RT_GROUP_SCHED
 608         struct rt_rq *rt_rq = group_rt_rq(rt_se);
 609
 610         if (rt_rq)
 611                 return rt_rq->highest_prio.curr;
 612 #endif
 613
 614         return rt_task_of(rt_se)->prio;
 615 }
 616
 617 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
 618 {
 619         u64 runtime = sched_rt_runtime(rt_rq);
 620
 621         if (rt_rq->rt_throttled)
 622                 return rt_rq_throttled(rt_rq);
 623
 624         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
 625                 return 0;
 626
 627         balance_runtime(rt_rq);
 628         runtime = sched_rt_runtime(rt_rq);
 629         if (runtime == RUNTIME_INF)
 630                 return 0;
 631
 632         if (rt_rq->rt_time > runtime) {
 633                 rt_rq->rt_throttled = 1;
 634                 if (rt_rq_throttled(rt_rq)) {
 635                         sched_rt_rq_dequeue(rt_rq);
 636                         return 1;
 637                 }
 638         }
 639
 640         return 0;
 641 }
 642
 643 /*
 644  * Update the current task's runtime statistics. Skip current tasks that
 645  * are not in our scheduling class.
 646  */
 647 static void update_curr_rt(struct rq *rq)
 648 {
 649         struct task_struct *curr = rq->curr;
 650         struct sched_rt_entity *rt_se = &curr->rt;
 651         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 652         u64 delta_exec;
 653
 654         if (curr->sched_class != &rt_sched_class)
 655                 return;
 656
 657         delta_exec = rq->clock_task - curr->se.exec_start;
 658         if (unlikely((s64)delta_exec < 0))
 659                 delta_exec = 0;
 660
 661         schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
 662
 663         curr->se.sum_exec_runtime += delta_exec;
 664         account_group_exec_runtime(curr, delta_exec);
 665
 666         curr->se.exec_start = rq->clock_task;
 667         cpuacct_charge(curr, delta_exec);
 668
 669         sched_rt_avg_update(rq, delta_exec);
 670
 671         if (!rt_bandwidth_enabled())
 672                 return;
 673
 674         for_each_sched_rt_entity(rt_se) {
 675                 rt_rq = rt_rq_of_se(rt_se);
 676
 677                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
 678                         raw_spin_lock(&rt_rq->rt_runtime_lock);
 679                         rt_rq->rt_time += delta_exec;
 680                         if (sched_rt_runtime_exceeded(rt_rq))
 681                                 resched_task(curr);
 682                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
 683                 }
 684         }
 685 }
 686
 687 #if defined CONFIG_SMP
 688
 689 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
 690
 691 static inline int next_prio(struct rq *rq)
 692 {
 693         struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
 694
 695         if (next && rt_prio(next->prio))
 696                 return next->prio;
 697         else
 698                 return MAX_RT_PRIO;
 699 }
 700
 701 static void
 702 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 703 {
 704         struct rq *rq = rq_of_rt_rq(rt_rq);
 705
 706         if (prio < prev_prio) {
 707
 708                 /*
 709                  * If the new task is higher in priority than anything on the
 710                  * run-queue, we know that the previous high becomes our
 711                  * next-highest.
 712                  */
 713                 rt_rq->highest_prio.next = prev_prio;
 714
 715                 if (rq->online)
 716                         cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
 717
 718         } else if (prio == rt_rq->highest_prio.curr)
 719                 /*
 720                  * If the next task is equal in priority to the highest on
 721                  * the run-queue, then we implicitly know that the next highest
 722                  * task cannot be any lower than current
 723                  */
 724                 rt_rq->highest_prio.next = prio;
 725         else if (prio < rt_rq->highest_prio.next)
 726                 /*
 727                  * Otherwise, we need to recompute next-highest
 728                  */
 729                 rt_rq->highest_prio.next = next_prio(rq);
 730 }
 731
 732 static void
 733 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
 734 {
 735         struct rq *rq = rq_of_rt_rq(rt_rq);
 736
 737         if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
 738                 rt_rq->highest_prio.next = next_prio(rq);
 739
 740         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
 741                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
 742 }
 743
 744 #else /* CONFIG_SMP */
 745
 746 static inline
 747 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 748 static inline
 749 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
 750
 751 #endif /* CONFIG_SMP */
 752
 753 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
 754 static void
 755 inc_rt_prio(struct rt_rq *rt_rq, int prio)
 756 {
 757         int prev_prio = rt_rq->highest_prio.curr;
 758
 759         if (prio < prev_prio)
 760                 rt_rq->highest_prio.curr = prio;
 761
 762         inc_rt_prio_smp(rt_rq, prio, prev_prio);
 763 }
 764
 765 static void
 766 dec_rt_prio(struct rt_rq *rt_rq, int prio)
 767 {
 768         int prev_prio = rt_rq->highest_prio.curr;
 769
 770         if (rt_rq->rt_nr_running) {
 771
 772                 WARN_ON(prio < prev_prio);
 773
 774                 /*
 775                  * This may have been our highest task, and therefore
 776                  * we may have some recomputation to do
 777                  */
 778                 if (prio == prev_prio) {
 779                         struct rt_prio_array *array = &rt_rq->active;
 780
 781                         rt_rq->highest_prio.curr =
 782                                 sched_find_first_bit(array->bitmap);
 783                 }
 784
 785         } else
 786                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
 787
 788         dec_rt_prio_smp(rt_rq, prio, prev_prio);
 789 }
 790
 791 #else
 792
 793 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
 794 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
 795
 796 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
 797
 798 #ifdef CONFIG_RT_GROUP_SCHED
 799
 800 static void
 801 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 802 {
 803         if (rt_se_boosted(rt_se))
 804                 rt_rq->rt_nr_boosted++;
 805
 806         if (rt_rq->tg)
 807                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
 808 }
 809
 810 static void
 811 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 812 {
 813         if (rt_se_boosted(rt_se))
 814                 rt_rq->rt_nr_boosted--;
 815
 816         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
 817 }
 818
 819 #else /* CONFIG_RT_GROUP_SCHED */
 820
 821 static void
 822 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 823 {
 824         start_rt_bandwidth(&def_rt_bandwidth);
 825 }
 826
 827 static inline
 828 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
 829
 830 #endif /* CONFIG_RT_GROUP_SCHED */
 831
 832 static inline
 833 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 834 {
 835         int prio = rt_se_prio(rt_se);
 836
 837         WARN_ON(!rt_prio(prio));
 838         rt_rq->rt_nr_running++;
 839
 840         inc_rt_prio(rt_rq, prio);
 841         inc_rt_migration(rt_se, rt_rq);
 842         inc_rt_group(rt_se, rt_rq);
 843 }
 844
 845 static inline
 846 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
 847 {
 848         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
 849         WARN_ON(!rt_rq->rt_nr_running);
 850         rt_rq->rt_nr_running--;
 851
 852         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
 853         dec_rt_migration(rt_se, rt_rq);
 854         dec_rt_group(rt_se, rt_rq);
 855 }
 856
 857 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 858 {
 859         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 860         struct rt_prio_array *array = &rt_rq->active;
 861         struct rt_rq *group_rq = group_rt_rq(rt_se);
 862         struct list_head *queue = array->queue + rt_se_prio(rt_se);
 863
 864         /*
 865          * Don't enqueue the group if its throttled, or when empty.
 866          * The latter is a consequence of the former when a child group
 867          * get throttled and the current group doesn't have any other
 868          * active members.
 869          */
 870         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
 871                 return;
 872
 873         if (!rt_rq->rt_nr_running)
 874                 list_add_leaf_rt_rq(rt_rq);
 875
 876         if (head)
 877                 list_add(&rt_se->run_list, queue);
 878         else
 879                 list_add_tail(&rt_se->run_list, queue);
 880         __set_bit(rt_se_prio(rt_se), array->bitmap);
 881
 882         inc_rt_tasks(rt_se, rt_rq);
 883 }
 884
 885 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
 886 {
 887         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
 888         struct rt_prio_array *array = &rt_rq->active;
 889
 890         list_del_init(&rt_se->run_list);
 891         if (list_empty(array->queue + rt_se_prio(rt_se)))
 892                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
 893
 894         dec_rt_tasks(rt_se, rt_rq);
 895         if (!rt_rq->rt_nr_running)
 896                 list_del_leaf_rt_rq(rt_rq);
 897 }
 898
 899 /*
 900  * Because the prio of an upper entry depends on the lower
 901  * entries, we must remove entries top - down.
 902  */
 903 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
 904 {
 905         struct sched_rt_entity *back = NULL;
 906
 907         for_each_sched_rt_entity(rt_se) {
 908                 rt_se->back = back;
 909                 back = rt_se;
 910         }
 911
 912         for (rt_se = back; rt_se; rt_se = rt_se->back) {
 913                 if (on_rt_rq(rt_se))
 914                         __dequeue_rt_entity(rt_se);
 915         }
 916 }
 917
 918 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
 919 {
 920         dequeue_rt_stack(rt_se);
 921         for_each_sched_rt_entity(rt_se)
 922                 __enqueue_rt_entity(rt_se, head);
 923 }
 924
 925 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
 926 {
 927         dequeue_rt_stack(rt_se);
 928
 929         for_each_sched_rt_entity(rt_se) {
 930                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
 931
 932                 if (rt_rq && rt_rq->rt_nr_running)
 933                         __enqueue_rt_entity(rt_se, false);
 934         }
 935 }
 936
 937 /*
 938  * Adding/removing a task to/from a priority array:
 939  */
 940 static void
 941 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 942 {
 943         struct sched_rt_entity *rt_se = &p->rt;
 944
 945         if (flags & ENQUEUE_WAKEUP)
 946                 rt_se->timeout = 0;
 947
 948         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
 949
 950         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
 951                 enqueue_pushable_task(rq, p);
 952 }
 953
 954 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
 955 {
 956         struct sched_rt_entity *rt_se = &p->rt;
 957
 958         update_curr_rt(rq);
 959         dequeue_rt_entity(rt_se);
 960
 961         dequeue_pushable_task(rq, p);
 962 }
 963
 964 /*
 965  * Put task to the end of the run list without the overhead of dequeue
 966  * followed by enqueue.
 967  */
 968 static void
 969 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
 970 {
 971         if (on_rt_rq(rt_se)) {
 972                 struct rt_prio_array *array = &rt_rq->active;
 973                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
 974
 975                 if (head)
 976                         list_move(&rt_se->run_list, queue);
 977                 else
 978                         list_move_tail(&rt_se->run_list, queue);
 979         }
 980 }
 981
 982 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
 983 {
 984         struct sched_rt_entity *rt_se = &p->rt;
 985         struct rt_rq *rt_rq;
 986
 987         for_each_sched_rt_entity(rt_se) {
 988                 rt_rq = rt_rq_of_se(rt_se);
 989                 requeue_rt_entity(rt_rq, rt_se, head);
 990         }
 991 }
 992
 993 static void yield_task_rt(struct rq *rq)
 994 {
 995         requeue_task_rt(rq, rq->curr, 0);
 996 }
 997
 998 #ifdef CONFIG_SMP
 999 static int find_lowest_rq(struct task_struct *task);
1000
1001 static int
1002 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
1003 {
1004         struct task_struct *curr;
1005         struct rq *rq;
1006         int cpu;
1007
1008         if (sd_flag != SD_BALANCE_WAKE)
1009                 return smp_processor_id();
1010
1011         cpu = task_cpu(p);
1012         rq = cpu_rq(cpu);
1013
1014         rcu_read_lock();
1015         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1016
1017         /*
1018          * If the current task on @p's runqueue is an RT task, then
1019          * try to see if we can wake this RT task up on another
1020          * runqueue. Otherwise simply start this RT task
1021          * on its current runqueue.
1022          *
1023          * We want to avoid overloading runqueues. If the woken
1024          * task is a higher priority, then it will stay on this CPU
1025          * and the lower prio task should be moved to another CPU.
1026          * Even though this will probably make the lower prio task
1027          * lose its cache, we do not want to bounce a higher task
1028          * around just because it gave up its CPU, perhaps for a
1029          * lock?
1030          *
1031          * For equal prio tasks, we just let the scheduler sort it out.
1032          *
1033          * Otherwise, just let it ride on the affined RQ and the
1034          * post-schedule router will push the preempted task away
1035          *
1036          * This test is optimistic, if we get it wrong the load-balancer
1037          * will have to sort it out.
1038          */
1039         if (curr && unlikely(rt_task(curr)) &&
1040             (curr->rt.nr_cpus_allowed < 2 ||
1041              curr->prio < p->prio) &&
1042             (p->rt.nr_cpus_allowed > 1)) {
1043                 int target = find_lowest_rq(p);
1044
1045                 if (target != -1)
1046                         cpu = target;
1047         }
1048         rcu_read_unlock();
1049
1050         return cpu;
1051 }
1052
1053 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1054 {
1055         if (rq->curr->rt.nr_cpus_allowed == 1)
1056                 return;
1057
1058         if (p->rt.nr_cpus_allowed != 1
1059             && cpupri_find(&rq->rd->cpupri, p, NULL))
1060                 return;
1061
1062         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1063                 return;
1064
1065         /*
1066          * There appears to be other cpus that can accept
1067          * current and none to run 'p', so lets reschedule
1068          * to try and push current away:
1069          */
1070         requeue_task_rt(rq, p, 1);
1071         resched_task(rq->curr);
1072 }
1073
1074 #endif /* CONFIG_SMP */
1075
1076 /*
1077  * Preempt the current task with a newly woken task if needed:
1078  */
1079 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1080 {
1081         if (p->prio < rq->curr->prio) {
1082                 resched_task(rq->curr);
1083                 return;
1084         }
1085
1086 #ifdef CONFIG_SMP
1087         /*
1088          * If:
1089          *
1090          * - the newly woken task is of equal priority to the current task
1091          * - the newly woken task is non-migratable while current is migratable
1092          * - current will be preempted on the next reschedule
1093          *
1094          * we should check to see if current can readily move to a different
1095          * cpu.  If so, we will reschedule to allow the push logic to try
1096          * to move current somewhere else, making room for our non-migratable
1097          * task.
1098          */
1099         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1100                 check_preempt_equal_prio(rq, p);
1101 #endif
1102 }
1103
1104 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1105                                                    struct rt_rq *rt_rq)
1106 {
1107         struct rt_prio_array *array = &rt_rq->active;
1108         struct sched_rt_entity *next = NULL;
1109         struct list_head *queue;
1110         int idx;
1111
1112         idx = sched_find_first_bit(array->bitmap);
1113         BUG_ON(idx >= MAX_RT_PRIO);
1114
1115         queue = array->queue + idx;
1116         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1117
1118         return next;
1119 }
1120
1121 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1122 {
1123         struct sched_rt_entity *rt_se;
1124         struct task_struct *p;
1125         struct rt_rq *rt_rq;
1126
1127         rt_rq = &rq->rt;
1128
1129         if (unlikely(!rt_rq->rt_nr_running))
1130                 return NULL;
1131
1132         if (rt_rq_throttled(rt_rq))
1133                 return NULL;
1134
1135         do {
1136                 rt_se = pick_next_rt_entity(rq, rt_rq);
1137                 BUG_ON(!rt_se);
1138                 rt_rq = group_rt_rq(rt_se);
1139         } while (rt_rq);
1140
1141         p = rt_task_of(rt_se);
1142         p->se.exec_start = rq->clock_task;
1143
1144         return p;
1145 }
1146
1147 static struct task_struct *pick_next_task_rt(struct rq *rq)
1148 {
1149         struct task_struct *p = _pick_next_task_rt(rq);
1150
1151         /* The running task is never eligible for pushing */
1152         if (p)
1153                 dequeue_pushable_task(rq, p);
1154
1155 #ifdef CONFIG_SMP
1156         /*
1157          * We detect this state here so that we can avoid taking the RQ
1158          * lock again later if there is no need to push
1159          */
1160         rq->post_schedule = has_pushable_tasks(rq);
1161 #endif
1162
1163         return p;
1164 }
1165
1166 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1167 {
1168         update_curr_rt(rq);
1169         p->se.exec_start = 0;
1170
1171         /*
1172          * The previous task needs to be made eligible for pushing
1173          * if it is still active
1174          */
1175         if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1176                 enqueue_pushable_task(rq, p);
1177 }
1178
1179 #ifdef CONFIG_SMP
1180
1181 /* Only try algorithms three times */
1182 #define RT_MAX_TRIES 3
1183
1184 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1185
1186 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1187 {
1188         if (!task_running(rq, p) &&
1189             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1190             (p->rt.nr_cpus_allowed > 1))
1191                 return 1;
1192         return 0;
1193 }
1194
1195 /* Return the second highest RT task, NULL otherwise */
1196 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1197 {
1198         struct task_struct *next = NULL;
1199         struct sched_rt_entity *rt_se;
1200         struct rt_prio_array *array;
1201         struct rt_rq *rt_rq;
1202         int idx;
1203
1204         for_each_leaf_rt_rq(rt_rq, rq) {
1205                 array = &rt_rq->active;
1206                 idx = sched_find_first_bit(array->bitmap);
1207 next_idx:
1208                 if (idx >= MAX_RT_PRIO)
1209                         continue;
1210                 if (next && next->prio < idx)
1211                         continue;
1212                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1213                         struct task_struct *p;
1214
1215                         if (!rt_entity_is_task(rt_se))
1216                                 continue;
1217
1218                         p = rt_task_of(rt_se);
1219                         if (pick_rt_task(rq, p, cpu)) {
1220                                 next = p;
1221                                 break;
1222                         }
1223                 }
1224                 if (!next) {
1225                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1226                         goto next_idx;
1227                 }
1228         }
1229
1230         return next;
1231 }
1232
1233 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1234
1235 static int find_lowest_rq(struct task_struct *task)
1236 {
1237         struct sched_domain *sd;
1238         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1239         int this_cpu = smp_processor_id();
1240         int cpu      = task_cpu(task);
1241
1242         /* Make sure the mask is initialized first */
1243         if (unlikely(!lowest_mask))
1244                 return -1;
1245
1246         if (task->rt.nr_cpus_allowed == 1)
1247                 return -1; /* No other targets possible */
1248
1249         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1250                 return -1; /* No targets found */
1251
1252         /*
1253          * At this point we have built a mask of cpus representing the
1254          * lowest priority tasks in the system.  Now we want to elect
1255          * the best one based on our affinity and topology.
1256          *
1257          * We prioritize the last cpu that the task executed on since
1258          * it is most likely cache-hot in that location.
1259          */
1260         if (cpumask_test_cpu(cpu, lowest_mask))
1261                 return cpu;
1262
1263         /*
1264          * Otherwise, we consult the sched_domains span maps to figure
1265          * out which cpu is logically closest to our hot cache data.
1266          */
1267         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1268                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1269
1270         rcu_read_lock();
1271         for_each_domain(cpu, sd) {
1272                 if (sd->flags & SD_WAKE_AFFINE) {
1273                         int best_cpu;
1274
1275                         /*
1276                          * "this_cpu" is cheaper to preempt than a
1277                          * remote processor.
1278                          */
1279                         if (this_cpu != -1 &&
1280                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1281                                 rcu_read_unlock();
1282                                 return this_cpu;
1283                         }
1284
1285                         best_cpu = cpumask_first_and(lowest_mask,
1286                                                      sched_domain_span(sd));
1287                         if (best_cpu < nr_cpu_ids) {
1288                                 rcu_read_unlock();
1289                                 return best_cpu;
1290                         }
1291                 }
1292         }
1293         rcu_read_unlock();
1294
1295         /*
1296          * And finally, if there were no matches within the domains
1297          * just give the caller *something* to work with from the compatible
1298          * locations.
1299          */
1300         if (this_cpu != -1)
1301                 return this_cpu;
1302
1303         cpu = cpumask_any(lowest_mask);
1304         if (cpu < nr_cpu_ids)
1305                 return cpu;
1306         return -1;
1307 }
1308
1309 /* Will lock the rq it finds */
1310 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1311 {
1312         struct rq *lowest_rq = NULL;
1313         int tries;
1314         int cpu;
1315
1316         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1317                 cpu = find_lowest_rq(task);
1318
1319                 if ((cpu == -1) || (cpu == rq->cpu))
1320                         break;
1321
1322                 lowest_rq = cpu_rq(cpu);
1323
1324                 /* if the prio of this runqueue changed, try again */
1325                 if (double_lock_balance(rq, lowest_rq)) {
1326                         /*
1327                          * We had to unlock the run queue. In
1328                          * the mean time, task could have
1329                          * migrated already or had its affinity changed.
1330                          * Also make sure that it wasn't scheduled on its rq.
1331                          */
1332                         if (unlikely(task_rq(task) != rq ||
1333                                      !cpumask_test_cpu(lowest_rq->cpu,
1334                                                        &task->cpus_allowed) ||
1335                                      task_running(rq, task) ||
1336                                      !task->on_rq)) {
1337
1338                                 raw_spin_unlock(&lowest_rq->lock);
1339                                 lowest_rq = NULL;
1340                                 break;
1341                         }
1342                 }
1343
1344                 /* If this rq is still suitable use it. */
1345                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1346                         break;
1347
1348                 /* try again */
1349                 double_unlock_balance(rq, lowest_rq);
1350                 lowest_rq = NULL;
1351         }
1352
1353         return lowest_rq;
1354 }
1355
1356 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1357 {
1358         struct task_struct *p;
1359
1360         if (!has_pushable_tasks(rq))
1361                 return NULL;
1362
1363         p = plist_first_entry(&rq->rt.pushable_tasks,
1364                               struct task_struct, pushable_tasks);
1365
1366         BUG_ON(rq->cpu != task_cpu(p));
1367         BUG_ON(task_current(rq, p));
1368         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1369
1370         BUG_ON(!p->on_rq);
1371         BUG_ON(!rt_task(p));
1372
1373         return p;
1374 }
1375
1376 /*
1377  * If the current CPU has more than one RT task, see if the non
1378  * running task can migrate over to a CPU that is running a task
1379  * of lesser priority.
1380  */
1381 static int push_rt_task(struct rq *rq)
1382 {
1383         struct task_struct *next_task;
1384         struct rq *lowest_rq;
1385
1386         if (!rq->rt.overloaded)
1387                 return 0;
1388
1389         next_task = pick_next_pushable_task(rq);
1390         if (!next_task)
1391                 return 0;
1392
1393 retry:
1394         if (unlikely(next_task == rq->curr)) {
1395                 WARN_ON(1);
1396                 return 0;
1397         }
1398
1399         /*
1400          * It's possible that the next_task slipped in of
1401          * higher priority than current. If that's the case
1402          * just reschedule current.
1403          */
1404         if (unlikely(next_task->prio < rq->curr->prio)) {
1405                 resched_task(rq->curr);
1406                 return 0;
1407         }
1408
1409         /* We might release rq lock */
1410         get_task_struct(next_task);
1411
1412         /* find_lock_lowest_rq locks the rq if found */
1413         lowest_rq = find_lock_lowest_rq(next_task, rq);
1414         if (!lowest_rq) {
1415                 struct task_struct *task;
1416                 /*
1417                  * find lock_lowest_rq releases rq->lock
1418                  * so it is possible that next_task has migrated.
1419                  *
1420                  * We need to make sure that the task is still on the same
1421                  * run-queue and is also still the next task eligible for
1422                  * pushing.
1423                  */
1424                 task = pick_next_pushable_task(rq);
1425                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1426                         /*
1427                          * If we get here, the task hasn't moved at all, but
1428                          * it has failed to push.  We will not try again,
1429                          * since the other cpus will pull from us when they
1430                          * are ready.
1431                          */
1432                         dequeue_pushable_task(rq, next_task);
1433                         goto out;
1434                 }
1435
1436                 if (!task)
1437                         /* No more tasks, just exit */
1438                         goto out;
1439
1440                 /*
1441                  * Something has shifted, try again.
1442                  */
1443                 put_task_struct(next_task);
1444                 next_task = task;
1445                 goto retry;
1446         }
1447
1448         deactivate_task(rq, next_task, 0);
1449         set_task_cpu(next_task, lowest_rq->cpu);
1450         activate_task(lowest_rq, next_task, 0);
1451
1452         resched_task(lowest_rq->curr);
1453
1454         double_unlock_balance(rq, lowest_rq);
1455
1456 out:
1457         put_task_struct(next_task);
1458
1459         return 1;
1460 }
1461
1462 static void push_rt_tasks(struct rq *rq)
1463 {
1464         /* push_rt_task will return true if it moved an RT */
1465         while (push_rt_task(rq))
1466                 ;
1467 }
1468
1469 static int pull_rt_task(struct rq *this_rq)
1470 {
1471         int this_cpu = this_rq->cpu, ret = 0, cpu;
1472         struct task_struct *p;
1473         struct rq *src_rq;
1474
1475         if (likely(!rt_overloaded(this_rq)))
1476                 return 0;
1477
1478         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1479                 if (this_cpu == cpu)
1480                         continue;
1481
1482                 src_rq = cpu_rq(cpu);
1483
1484                 /*
1485                  * Don't bother taking the src_rq->lock if the next highest
1486                  * task is known to be lower-priority than our current task.
1487                  * This may look racy, but if this value is about to go
1488                  * logically higher, the src_rq will push this task away.
1489                  * And if its going logically lower, we do not care
1490                  */
1491                 if (src_rq->rt.highest_prio.next >=
1492                     this_rq->rt.highest_prio.curr)
1493                         continue;
1494
1495                 /*
1496                  * We can potentially drop this_rq's lock in
1497                  * double_lock_balance, and another CPU could
1498                  * alter this_rq
1499                  */
1500                 double_lock_balance(this_rq, src_rq);
1501
1502                 /*
1503                  * Are there still pullable RT tasks?
1504                  */
1505                 if (src_rq->rt.rt_nr_running <= 1)
1506                         goto skip;
1507
1508                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1509
1510                 /*
1511                  * Do we have an RT task that preempts
1512                  * the to-be-scheduled task?
1513                  */
1514                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1515                         WARN_ON(p == src_rq->curr);
1516                         WARN_ON(!p->on_rq);
1517
1518                         /*
1519                          * There's a chance that p is higher in priority
1520                          * than what's currently running on its cpu.
1521                          * This is just that p is wakeing up and hasn't
1522                          * had a chance to schedule. We only pull
1523                          * p if it is lower in priority than the
1524                          * current task on the run queue
1525                          */
1526                         if (p->prio < src_rq->curr->prio)
1527                                 goto skip;
1528
1529                         ret = 1;
1530
1531                         deactivate_task(src_rq, p, 0);
1532                         set_task_cpu(p, this_cpu);
1533                         activate_task(this_rq, p, 0);
1534                         /*
1535                          * We continue with the search, just in
1536                          * case there's an even higher prio task
1537                          * in another runqueue. (low likelihood
1538                          * but possible)
1539                          */
1540                 }
1541 skip:
1542                 double_unlock_balance(this_rq, src_rq);
1543         }
1544
1545         return ret;
1546 }
1547
1548 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1549 {
1550         /* Try to pull RT tasks here if we lower this rq's prio */
1551         if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1552                 pull_rt_task(rq);
1553 }
1554
1555 static void post_schedule_rt(struct rq *rq)
1556 {
1557         push_rt_tasks(rq);
1558 }
1559
1560 /*
1561  * If we are not running and we are not going to reschedule soon, we should
1562  * try to push tasks away now
1563  */
1564 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1565 {
1566         if (!task_running(rq, p) &&
1567             !test_tsk_need_resched(rq->curr) &&
1568             has_pushable_tasks(rq) &&
1569             p->rt.nr_cpus_allowed > 1 &&
1570             rt_task(rq->curr) &&
1571             (rq->curr->rt.nr_cpus_allowed < 2 ||
1572              rq->curr->prio < p->prio))
1573                 push_rt_tasks(rq);
1574 }
1575
1576 static void set_cpus_allowed_rt(struct task_struct *p,
1577                                 const struct cpumask *new_mask)
1578 {
1579         int weight = cpumask_weight(new_mask);
1580
1581         BUG_ON(!rt_task(p));
1582
1583         /*
1584          * Update the migration status of the RQ if we have an RT task
1585          * which is running AND changing its weight value.
1586          */
1587         if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1588                 struct rq *rq = task_rq(p);
1589
1590                 if (!task_current(rq, p)) {
1591                         /*
1592                          * Make sure we dequeue this task from the pushable list
1593                          * before going further.  It will either remain off of
1594                          * the list because we are no longer pushable, or it
1595                          * will be requeued.
1596                          */
1597                         if (p->rt.nr_cpus_allowed > 1)
1598                                 dequeue_pushable_task(rq, p);
1599
1600                         /*
1601                          * Requeue if our weight is changing and still > 1
1602                          */
1603                         if (weight > 1)
1604                                 enqueue_pushable_task(rq, p);
1605
1606                 }
1607
1608                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1609                         rq->rt.rt_nr_migratory++;
1610                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1611                         BUG_ON(!rq->rt.rt_nr_migratory);
1612                         rq->rt.rt_nr_migratory--;
1613                 }
1614
1615                 update_rt_migration(&rq->rt);
1616         }
1617
1618         cpumask_copy(&p->cpus_allowed, new_mask);
1619         p->rt.nr_cpus_allowed = weight;
1620 }
1621
1622 /* Assumes rq->lock is held */
1623 static void rq_online_rt(struct rq *rq)
1624 {
1625         if (rq->rt.overloaded)
1626                 rt_set_overload(rq);
1627
1628         __enable_runtime(rq);
1629
1630         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1631 }
1632
1633 /* Assumes rq->lock is held */
1634 static void rq_offline_rt(struct rq *rq)
1635 {
1636         if (rq->rt.overloaded)
1637                 rt_clear_overload(rq);
1638
1639         __disable_runtime(rq);
1640
1641         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1642 }
1643
1644 /*
1645  * When switch from the rt queue, we bring ourselves to a position
1646  * that we might want to pull RT tasks from other runqueues.
1647  */
1648 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1649 {
1650         /*
1651          * If there are other RT tasks then we will reschedule
1652          * and the scheduling of the other RT tasks will handle
1653          * the balancing. But if we are the last RT task
1654          * we may need to handle the pulling of RT tasks
1655          * now.
1656          */
1657         if (p->on_rq && !rq->rt.rt_nr_running)
1658                 pull_rt_task(rq);
1659 }
1660
1661 static inline void init_sched_rt_class(void)
1662 {
1663         unsigned int i;
1664
1665         for_each_possible_cpu(i)
1666                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1667                                         GFP_KERNEL, cpu_to_node(i));
1668 }
1669 #endif /* CONFIG_SMP */
1670
1671 /*
1672  * When switching a task to RT, we may overload the runqueue
1673  * with RT tasks. In this case we try to push them off to
1674  * other runqueues.
1675  */
1676 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1677 {
1678         int check_resched = 1;
1679
1680         /*
1681          * If we are already running, then there's nothing
1682          * that needs to be done. But if we are not running
1683          * we may need to preempt the current running task.
1684          * If that current running task is also an RT task
1685          * then see if we can move to another run queue.
1686          */
1687         if (p->on_rq && rq->curr != p) {
1688 #ifdef CONFIG_SMP
1689                 if (rq->rt.overloaded && push_rt_task(rq) &&
1690                     /* Don't resched if we changed runqueues */
1691                     rq != task_rq(p))
1692                         check_resched = 0;
1693 #endif /* CONFIG_SMP */
1694                 if (check_resched && p->prio < rq->curr->prio)
1695                         resched_task(rq->curr);
1696         }
1697 }
1698
1699 /*
1700  * Priority of the task has changed. This may cause
1701  * us to initiate a push or pull.
1702  */
1703 static void
1704 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1705 {
1706         if (!p->on_rq)
1707                 return;
1708
1709         if (rq->curr == p) {
1710 #ifdef CONFIG_SMP
1711                 /*
1712                  * If our priority decreases while running, we
1713                  * may need to pull tasks to this runqueue.
1714                  */
1715                 if (oldprio < p->prio)
1716                         pull_rt_task(rq);
1717                 /*
1718                  * If there's a higher priority task waiting to run
1719                  * then reschedule. Note, the above pull_rt_task
1720                  * can release the rq lock and p could migrate.
1721                  * Only reschedule if p is still on the same runqueue.
1722                  */
1723                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1724                         resched_task(p);
1725 #else
1726                 /* For UP simply resched on drop of prio */
1727                 if (oldprio < p->prio)
1728                         resched_task(p);
1729 #endif /* CONFIG_SMP */
1730         } else {
1731                 /*
1732                  * This task is not running, but if it is
1733                  * greater than the current running task
1734                  * then reschedule.
1735                  */
1736                 if (p->prio < rq->curr->prio)
1737                         resched_task(rq->curr);
1738         }
1739 }
1740
1741 static void watchdog(struct rq *rq, struct task_struct *p)
1742 {
1743         unsigned long soft, hard;
1744
1745         /* max may change after cur was read, this will be fixed next tick */
1746         soft = task_rlimit(p, RLIMIT_RTTIME);
1747         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1748
1749         if (soft != RLIM_INFINITY) {
1750                 unsigned long next;
1751
1752                 p->rt.timeout++;
1753                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1754                 if (p->rt.timeout > next)
1755                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1756         }
1757 }
1758
1759 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1760 {
1761         update_curr_rt(rq);
1762
1763         watchdog(rq, p);
1764
1765         /*
1766          * RR tasks need a special form of timeslice management.
1767          * FIFO tasks have no timeslices.
1768          */
1769         if (p->policy != SCHED_RR)
1770                 return;
1771
1772         if (--p->rt.time_slice)
1773                 return;
1774
1775         p->rt.time_slice = DEF_TIMESLICE;
1776
1777         /*
1778          * Requeue to the end of queue if we are not the only element
1779          * on the queue:
1780          */
1781         if (p->rt.run_list.prev != p->rt.run_list.next) {
1782                 requeue_task_rt(rq, p, 0);
1783                 set_tsk_need_resched(p);
1784         }
1785 }
1786
1787 static void set_curr_task_rt(struct rq *rq)
1788 {
1789         struct task_struct *p = rq->curr;
1790
1791         p->se.exec_start = rq->clock_task;
1792
1793         /* The running task is never eligible for pushing */
1794         dequeue_pushable_task(rq, p);
1795 }
1796
1797 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1798 {
1799         /*
1800          * Time slice is 0 for SCHED_FIFO tasks
1801          */
1802         if (task->policy == SCHED_RR)
1803                 return DEF_TIMESLICE;
1804         else
1805                 return 0;
1806 }
1807
1808 static const struct sched_class rt_sched_class = {
1809         .next                   = &fair_sched_class,
1810         .enqueue_task           = enqueue_task_rt,
1811         .dequeue_task           = dequeue_task_rt,
1812         .yield_task             = yield_task_rt,
1813
1814         .check_preempt_curr     = check_preempt_curr_rt,
1815
1816         .pick_next_task         = pick_next_task_rt,
1817         .put_prev_task          = put_prev_task_rt,
1818
1819 #ifdef CONFIG_SMP
1820         .select_task_rq         = select_task_rq_rt,
1821
1822         .set_cpus_allowed       = set_cpus_allowed_rt,
1823         .rq_online              = rq_online_rt,
1824         .rq_offline             = rq_offline_rt,
1825         .pre_schedule           = pre_schedule_rt,
1826         .post_schedule          = post_schedule_rt,
1827         .task_woken             = task_woken_rt,
1828         .switched_from          = switched_from_rt,
1829 #endif
1830
1831         .set_curr_task          = set_curr_task_rt,
1832         .task_tick              = task_tick_rt,
1833
1834         .get_rr_interval        = get_rr_interval_rt,
1835
1836         .prio_changed           = prio_changed_rt,
1837         .switched_to            = switched_to_rt,
1838 };
1839
1840 #ifdef CONFIG_SCHED_DEBUG
1841 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1842
1843 static void print_rt_stats(struct seq_file *m, int cpu)
1844 {
1845         rt_rq_iter_t iter;
1846         struct rt_rq *rt_rq;
1847
1848         rcu_read_lock();
1849         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
1850                 print_rt_rq(m, cpu, rt_rq);
1851         rcu_read_unlock();
1852 }
1853 #endif /* CONFIG_SCHED_DEBUG */
1854