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 static cpumask_t rt_overload_mask;
   8 static atomic_t rto_count;
   9 static inline int rt_overloaded(void)
  10 {
  11         return atomic_read(&rto_count);
  12 }
  13 static inline cpumask_t *rt_overload(void)
  14 {
  15         return &rt_overload_mask;
  16 }
  17 static inline void rt_set_overload(struct rq *rq)
  18 {
  19         cpu_set(rq->cpu, rt_overload_mask);
  20         /*
  21          * Make sure the mask is visible before we set
  22          * the overload count. That is checked to determine
  23          * if we should look at the mask. It would be a shame
  24          * if we looked at the mask, but the mask was not
  25          * updated yet.
  26          */
  27         wmb();
  28         atomic_inc(&rto_count);
  29 }
  30 static inline void rt_clear_overload(struct rq *rq)
  31 {
  32         /* the order here really doesn't matter */
  33         atomic_dec(&rto_count);
  34         cpu_clear(rq->cpu, rt_overload_mask);
  35 }
  36
  37 static void update_rt_migration(struct rq *rq)
  38 {
  39         if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1))
  40                 rt_set_overload(rq);
  41         else
  42                 rt_clear_overload(rq);
  43 }
  44 #endif /* CONFIG_SMP */
  45
  46 /*
  47  * Update the current task's runtime statistics. Skip current tasks that
  48  * are not in our scheduling class.
  49  */
  50 static void update_curr_rt(struct rq *rq)
  51 {
  52         struct task_struct *curr = rq->curr;
  53         u64 delta_exec;
  54
  55         if (!task_has_rt_policy(curr))
  56                 return;
  57
  58         delta_exec = rq->clock - curr->se.exec_start;
  59         if (unlikely((s64)delta_exec < 0))
  60                 delta_exec = 0;
  61
  62         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
  63
  64         curr->se.sum_exec_runtime += delta_exec;
  65         curr->se.exec_start = rq->clock;
  66         cpuacct_charge(curr, delta_exec);
  67 }
  68
  69 static inline void inc_rt_tasks(struct task_struct *p, struct rq *rq)
  70 {
  71         WARN_ON(!rt_task(p));
  72         rq->rt.rt_nr_running++;
  73 #ifdef CONFIG_SMP
  74         if (p->prio < rq->rt.highest_prio)
  75                 rq->rt.highest_prio = p->prio;
  76         if (p->nr_cpus_allowed > 1)
  77                 rq->rt.rt_nr_migratory++;
  78
  79         update_rt_migration(rq);
  80 #endif /* CONFIG_SMP */
  81 }
  82
  83 static inline void dec_rt_tasks(struct task_struct *p, struct rq *rq)
  84 {
  85         WARN_ON(!rt_task(p));
  86         WARN_ON(!rq->rt.rt_nr_running);
  87         rq->rt.rt_nr_running--;
  88 #ifdef CONFIG_SMP
  89         if (rq->rt.rt_nr_running) {
  90                 struct rt_prio_array *array;
  91
  92                 WARN_ON(p->prio < rq->rt.highest_prio);
  93                 if (p->prio == rq->rt.highest_prio) {
  94                         /* recalculate */
  95                         array = &rq->rt.active;
  96                         rq->rt.highest_prio =
  97                                 sched_find_first_bit(array->bitmap);
  98                 } /* otherwise leave rq->highest prio alone */
  99         } else
 100                 rq->rt.highest_prio = MAX_RT_PRIO;
 101         if (p->nr_cpus_allowed > 1)
 102                 rq->rt.rt_nr_migratory--;
 103
 104         update_rt_migration(rq);
 105 #endif /* CONFIG_SMP */
 106 }
 107
 108 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
 109 {
 110         struct rt_prio_array *array = &rq->rt.active;
 111
 112         list_add_tail(&p->run_list, array->queue + p->prio);
 113         __set_bit(p->prio, array->bitmap);
 114         inc_cpu_load(rq, p->se.load.weight);
 115
 116         inc_rt_tasks(p, rq);
 117 }
 118
 119 /*
 120  * Adding/removing a task to/from a priority array:
 121  */
 122 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
 123 {
 124         struct rt_prio_array *array = &rq->rt.active;
 125
 126         update_curr_rt(rq);
 127
 128         list_del(&p->run_list);
 129         if (list_empty(array->queue + p->prio))
 130                 __clear_bit(p->prio, array->bitmap);
 131         dec_cpu_load(rq, p->se.load.weight);
 132
 133         dec_rt_tasks(p, rq);
 134 }
 135
 136 /*
 137  * Put task to the end of the run list without the overhead of dequeue
 138  * followed by enqueue.
 139  */
 140 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
 141 {
 142         struct rt_prio_array *array = &rq->rt.active;
 143
 144         list_move_tail(&p->run_list, array->queue + p->prio);
 145 }
 146
 147 static void
 148 yield_task_rt(struct rq *rq)
 149 {
 150         requeue_task_rt(rq, rq->curr);
 151 }
 152
 153 /*
 154  * Preempt the current task with a newly woken task if needed:
 155  */
 156 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
 157 {
 158         if (p->prio < rq->curr->prio)
 159                 resched_task(rq->curr);
 160 }
 161
 162 static struct task_struct *pick_next_task_rt(struct rq *rq)
 163 {
 164         struct rt_prio_array *array = &rq->rt.active;
 165         struct task_struct *next;
 166         struct list_head *queue;
 167         int idx;
 168
 169         idx = sched_find_first_bit(array->bitmap);
 170         if (idx >= MAX_RT_PRIO)
 171                 return NULL;
 172
 173         queue = array->queue + idx;
 174         next = list_entry(queue->next, struct task_struct, run_list);
 175
 176         next->se.exec_start = rq->clock;
 177
 178         return next;
 179 }
 180
 181 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
 182 {
 183         update_curr_rt(rq);
 184         p->se.exec_start = 0;
 185 }
 186
 187 #ifdef CONFIG_SMP
 188 /* Only try algorithms three times */
 189 #define RT_MAX_TRIES 3
 190
 191 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
 192 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
 193
 194 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
 195 {
 196         if (!task_running(rq, p) &&
 197             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
 198             (p->nr_cpus_allowed > 1))
 199                 return 1;
 200         return 0;
 201 }
 202
 203 /* Return the second highest RT task, NULL otherwise */
 204 static struct task_struct *pick_next_highest_task_rt(struct rq *rq,
 205                                                      int cpu)
 206 {
 207         struct rt_prio_array *array = &rq->rt.active;
 208         struct task_struct *next;
 209         struct list_head *queue;
 210         int idx;
 211
 212         assert_spin_locked(&rq->lock);
 213
 214         if (likely(rq->rt.rt_nr_running < 2))
 215                 return NULL;
 216
 217         idx = sched_find_first_bit(array->bitmap);
 218         if (unlikely(idx >= MAX_RT_PRIO)) {
 219                 WARN_ON(1); /* rt_nr_running is bad */
 220                 return NULL;
 221         }
 222
 223         queue = array->queue + idx;
 224         BUG_ON(list_empty(queue));
 225
 226         next = list_entry(queue->next, struct task_struct, run_list);
 227         if (unlikely(pick_rt_task(rq, next, cpu)))
 228                 goto out;
 229
 230         if (queue->next->next != queue) {
 231                 /* same prio task */
 232                 next = list_entry(queue->next->next, struct task_struct, run_list);
 233                 if (pick_rt_task(rq, next, cpu))
 234                         goto out;
 235         }
 236
 237  retry:
 238         /* slower, but more flexible */
 239         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
 240         if (unlikely(idx >= MAX_RT_PRIO))
 241                 return NULL;
 242
 243         queue = array->queue + idx;
 244         BUG_ON(list_empty(queue));
 245
 246         list_for_each_entry(next, queue, run_list) {
 247                 if (pick_rt_task(rq, next, cpu))
 248                         goto out;
 249         }
 250
 251         goto retry;
 252
 253  out:
 254         return next;
 255 }
 256
 257 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
 258
 259 /* Will lock the rq it finds */
 260 static struct rq *find_lock_lowest_rq(struct task_struct *task,
 261                                       struct rq *this_rq)
 262 {
 263         struct rq *lowest_rq = NULL;
 264         int cpu;
 265         int tries;
 266         cpumask_t *cpu_mask = &__get_cpu_var(local_cpu_mask);
 267
 268         cpus_and(*cpu_mask, cpu_online_map, task->cpus_allowed);
 269
 270         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
 271                 /*
 272                  * Scan each rq for the lowest prio.
 273                  */
 274                 for_each_cpu_mask(cpu, *cpu_mask) {
 275                         struct rq *rq = &per_cpu(runqueues, cpu);
 276
 277                         if (cpu == this_rq->cpu)
 278                                 continue;
 279
 280                         /* We look for lowest RT prio or non-rt CPU */
 281                         if (rq->rt.highest_prio >= MAX_RT_PRIO) {
 282                                 lowest_rq = rq;
 283                                 break;
 284                         }
 285
 286                         /* no locking for now */
 287                         if (rq->rt.highest_prio > task->prio &&
 288                             (!lowest_rq || rq->rt.highest_prio > lowest_rq->rt.highest_prio)) {
 289                                 lowest_rq = rq;
 290                         }
 291                 }
 292
 293                 if (!lowest_rq)
 294                         break;
 295
 296                 /* if the prio of this runqueue changed, try again */
 297                 if (double_lock_balance(this_rq, lowest_rq)) {
 298                         /*
 299                          * We had to unlock the run queue. In
 300                          * the mean time, task could have
 301                          * migrated already or had its affinity changed.
 302                          * Also make sure that it wasn't scheduled on its rq.
 303                          */
 304                         if (unlikely(task_rq(task) != this_rq ||
 305                                      !cpu_isset(lowest_rq->cpu, task->cpus_allowed) ||
 306                                      task_running(this_rq, task) ||
 307                                      !task->se.on_rq)) {
 308                                 spin_unlock(&lowest_rq->lock);
 309                                 lowest_rq = NULL;
 310                                 break;
 311                         }
 312                 }
 313
 314                 /* If this rq is still suitable use it. */
 315                 if (lowest_rq->rt.highest_prio > task->prio)
 316                         break;
 317
 318                 /* try again */
 319                 spin_unlock(&lowest_rq->lock);
 320                 lowest_rq = NULL;
 321         }
 322
 323         return lowest_rq;
 324 }
 325
 326 /*
 327  * If the current CPU has more than one RT task, see if the non
 328  * running task can migrate over to a CPU that is running a task
 329  * of lesser priority.
 330  */
 331 static int push_rt_task(struct rq *this_rq)
 332 {
 333         struct task_struct *next_task;
 334         struct rq *lowest_rq;
 335         int ret = 0;
 336         int paranoid = RT_MAX_TRIES;
 337
 338         assert_spin_locked(&this_rq->lock);
 339
 340         next_task = pick_next_highest_task_rt(this_rq, -1);
 341         if (!next_task)
 342                 return 0;
 343
 344  retry:
 345         if (unlikely(next_task == this_rq->curr)) {
 346                 WARN_ON(1);
 347                 return 0;
 348         }
 349
 350         /*
 351          * It's possible that the next_task slipped in of
 352          * higher priority than current. If that's the case
 353          * just reschedule current.
 354          */
 355         if (unlikely(next_task->prio < this_rq->curr->prio)) {
 356                 resched_task(this_rq->curr);
 357                 return 0;
 358         }
 359
 360         /* We might release this_rq lock */
 361         get_task_struct(next_task);
 362
 363         /* find_lock_lowest_rq locks the rq if found */
 364         lowest_rq = find_lock_lowest_rq(next_task, this_rq);
 365         if (!lowest_rq) {
 366                 struct task_struct *task;
 367                 /*
 368                  * find lock_lowest_rq releases this_rq->lock
 369                  * so it is possible that next_task has changed.
 370                  * If it has, then try again.
 371                  */
 372                 task = pick_next_highest_task_rt(this_rq, -1);
 373                 if (unlikely(task != next_task) && task && paranoid--) {
 374                         put_task_struct(next_task);
 375                         next_task = task;
 376                         goto retry;
 377                 }
 378                 goto out;
 379         }
 380
 381         assert_spin_locked(&lowest_rq->lock);
 382
 383         deactivate_task(this_rq, next_task, 0);
 384         set_task_cpu(next_task, lowest_rq->cpu);
 385         activate_task(lowest_rq, next_task, 0);
 386
 387         resched_task(lowest_rq->curr);
 388
 389         spin_unlock(&lowest_rq->lock);
 390
 391         ret = 1;
 392 out:
 393         put_task_struct(next_task);
 394
 395         return ret;
 396 }
 397
 398 /*
 399  * TODO: Currently we just use the second highest prio task on
 400  *       the queue, and stop when it can't migrate (or there's
 401  *       no more RT tasks).  There may be a case where a lower
 402  *       priority RT task has a different affinity than the
 403  *       higher RT task. In this case the lower RT task could
 404  *       possibly be able to migrate where as the higher priority
 405  *       RT task could not.  We currently ignore this issue.
 406  *       Enhancements are welcome!
 407  */
 408 static void push_rt_tasks(struct rq *rq)
 409 {
 410         /* push_rt_task will return true if it moved an RT */
 411         while (push_rt_task(rq))
 412                 ;
 413 }
 414
 415 static int pull_rt_task(struct rq *this_rq)
 416 {
 417         struct task_struct *next;
 418         struct task_struct *p;
 419         struct rq *src_rq;
 420         cpumask_t *rto_cpumask;
 421         int this_cpu = this_rq->cpu;
 422         int cpu;
 423         int ret = 0;
 424
 425         assert_spin_locked(&this_rq->lock);
 426
 427         /*
 428          * If cpusets are used, and we have overlapping
 429          * run queue cpusets, then this algorithm may not catch all.
 430          * This is just the price you pay on trying to keep
 431          * dirtying caches down on large SMP machines.
 432          */
 433         if (likely(!rt_overloaded()))
 434                 return 0;
 435
 436         next = pick_next_task_rt(this_rq);
 437
 438         rto_cpumask = rt_overload();
 439
 440         for_each_cpu_mask(cpu, *rto_cpumask) {
 441                 if (this_cpu == cpu)
 442                         continue;
 443
 444                 src_rq = cpu_rq(cpu);
 445                 if (unlikely(src_rq->rt.rt_nr_running <= 1)) {
 446                         /*
 447                          * It is possible that overlapping cpusets
 448                          * will miss clearing a non overloaded runqueue.
 449                          * Clear it now.
 450                          */
 451                         if (double_lock_balance(this_rq, src_rq)) {
 452                                 /* unlocked our runqueue lock */
 453                                 struct task_struct *old_next = next;
 454                                 next = pick_next_task_rt(this_rq);
 455                                 if (next != old_next)
 456                                         ret = 1;
 457                         }
 458                         if (likely(src_rq->rt.rt_nr_running <= 1))
 459                                 /*
 460                                  * Small chance that this_rq->curr changed
 461                                  * but it's really harmless here.
 462                                  */
 463                                 rt_clear_overload(this_rq);
 464                         else
 465                                 /*
 466                                  * Heh, the src_rq is now overloaded, since
 467                                  * we already have the src_rq lock, go straight
 468                                  * to pulling tasks from it.
 469                                  */
 470                                 goto try_pulling;
 471                         spin_unlock(&src_rq->lock);
 472                         continue;
 473                 }
 474
 475                 /*
 476                  * We can potentially drop this_rq's lock in
 477                  * double_lock_balance, and another CPU could
 478                  * steal our next task - hence we must cause
 479                  * the caller to recalculate the next task
 480                  * in that case:
 481                  */
 482                 if (double_lock_balance(this_rq, src_rq)) {
 483                         struct task_struct *old_next = next;
 484                         next = pick_next_task_rt(this_rq);
 485                         if (next != old_next)
 486                                 ret = 1;
 487                 }
 488
 489                 /*
 490                  * Are there still pullable RT tasks?
 491                  */
 492                 if (src_rq->rt.rt_nr_running <= 1) {
 493                         spin_unlock(&src_rq->lock);
 494                         continue;
 495                 }
 496
 497  try_pulling:
 498                 p = pick_next_highest_task_rt(src_rq, this_cpu);
 499
 500                 /*
 501                  * Do we have an RT task that preempts
 502                  * the to-be-scheduled task?
 503                  */
 504                 if (p && (!next || (p->prio < next->prio))) {
 505                         WARN_ON(p == src_rq->curr);
 506                         WARN_ON(!p->se.on_rq);
 507
 508                         /*
 509                          * There's a chance that p is higher in priority
 510                          * than what's currently running on its cpu.
 511                          * This is just that p is wakeing up and hasn't
 512                          * had a chance to schedule. We only pull
 513                          * p if it is lower in priority than the
 514                          * current task on the run queue or
 515                          * this_rq next task is lower in prio than
 516                          * the current task on that rq.
 517                          */
 518                         if (p->prio < src_rq->curr->prio ||
 519                             (next && next->prio < src_rq->curr->prio))
 520                                 goto bail;
 521
 522                         ret = 1;
 523
 524                         deactivate_task(src_rq, p, 0);
 525                         set_task_cpu(p, this_cpu);
 526                         activate_task(this_rq, p, 0);
 527                         /*
 528                          * We continue with the search, just in
 529                          * case there's an even higher prio task
 530                          * in another runqueue. (low likelyhood
 531                          * but possible)
 532                          */
 533
 534                         /*
 535                          * Update next so that we won't pick a task
 536                          * on another cpu with a priority lower (or equal)
 537                          * than the one we just picked.
 538                          */
 539                         next = p;
 540
 541                 }
 542  bail:
 543                 spin_unlock(&src_rq->lock);
 544         }
 545
 546         return ret;
 547 }
 548
 549 static void schedule_balance_rt(struct rq *rq,
 550                                 struct task_struct *prev)
 551 {
 552         /* Try to pull RT tasks here if we lower this rq's prio */
 553         if (unlikely(rt_task(prev)) &&
 554             rq->rt.highest_prio > prev->prio)
 555                 pull_rt_task(rq);
 556 }
 557
 558 static void schedule_tail_balance_rt(struct rq *rq)
 559 {
 560         /*
 561          * If we have more than one rt_task queued, then
 562          * see if we can push the other rt_tasks off to other CPUS.
 563          * Note we may release the rq lock, and since
 564          * the lock was owned by prev, we need to release it
 565          * first via finish_lock_switch and then reaquire it here.
 566          */
 567         if (unlikely(rq->rt.rt_nr_running > 1)) {
 568                 spin_lock_irq(&rq->lock);
 569                 push_rt_tasks(rq);
 570                 spin_unlock_irq(&rq->lock);
 571         }
 572 }
 573
 574
 575 static void wakeup_balance_rt(struct rq *rq, struct task_struct *p)
 576 {
 577         if (unlikely(rt_task(p)) &&
 578             !task_running(rq, p) &&
 579             (p->prio >= rq->curr->prio))
 580                 push_rt_tasks(rq);
 581 }
 582
 583 static unsigned long
 584 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
 585                 unsigned long max_load_move,
 586                 struct sched_domain *sd, enum cpu_idle_type idle,
 587                 int *all_pinned, int *this_best_prio)
 588 {
 589         /* don't touch RT tasks */
 590         return 0;
 591 }
 592
 593 static int
 594 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
 595                  struct sched_domain *sd, enum cpu_idle_type idle)
 596 {
 597         /* don't touch RT tasks */
 598         return 0;
 599 }
 600 static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask)
 601 {
 602         int weight = cpus_weight(*new_mask);
 603
 604         BUG_ON(!rt_task(p));
 605
 606         /*
 607          * Update the migration status of the RQ if we have an RT task
 608          * which is running AND changing its weight value.
 609          */
 610         if (p->se.on_rq && (weight != p->nr_cpus_allowed)) {
 611                 struct rq *rq = task_rq(p);
 612
 613                 if ((p->nr_cpus_allowed <= 1) && (weight > 1))
 614                         rq->rt.rt_nr_migratory++;
 615                 else if((p->nr_cpus_allowed > 1) && (weight <= 1)) {
 616                         BUG_ON(!rq->rt.rt_nr_migratory);
 617                         rq->rt.rt_nr_migratory--;
 618                 }
 619
 620                 update_rt_migration(rq);
 621         }
 622
 623         p->cpus_allowed    = *new_mask;
 624         p->nr_cpus_allowed = weight;
 625 }
 626 #else /* CONFIG_SMP */
 627 # define schedule_tail_balance_rt(rq)   do { } while (0)
 628 # define schedule_balance_rt(rq, prev)  do { } while (0)
 629 # define wakeup_balance_rt(rq, p)       do { } while (0)
 630 #endif /* CONFIG_SMP */
 631
 632 static void task_tick_rt(struct rq *rq, struct task_struct *p)
 633 {
 634         update_curr_rt(rq);
 635
 636         /*
 637          * RR tasks need a special form of timeslice management.
 638          * FIFO tasks have no timeslices.
 639          */
 640         if (p->policy != SCHED_RR)
 641                 return;
 642
 643         if (--p->time_slice)
 644                 return;
 645
 646         p->time_slice = DEF_TIMESLICE;
 647
 648         /*
 649          * Requeue to the end of queue if we are not the only element
 650          * on the queue:
 651          */
 652         if (p->run_list.prev != p->run_list.next) {
 653                 requeue_task_rt(rq, p);
 654                 set_tsk_need_resched(p);
 655         }
 656 }
 657
 658 static void set_curr_task_rt(struct rq *rq)
 659 {
 660         struct task_struct *p = rq->curr;
 661
 662         p->se.exec_start = rq->clock;
 663 }
 664
 665 const struct sched_class rt_sched_class = {
 666         .next                   = &fair_sched_class,
 667         .enqueue_task           = enqueue_task_rt,
 668         .dequeue_task           = dequeue_task_rt,
 669         .yield_task             = yield_task_rt,
 670
 671         .check_preempt_curr     = check_preempt_curr_rt,
 672
 673         .pick_next_task         = pick_next_task_rt,
 674         .put_prev_task          = put_prev_task_rt,
 675
 676 #ifdef CONFIG_SMP
 677         .load_balance           = load_balance_rt,
 678         .move_one_task          = move_one_task_rt,
 679         .set_cpus_allowed       = set_cpus_allowed_rt,
 680 #endif
 681
 682         .set_curr_task          = set_curr_task_rt,
 683         .task_tick              = task_tick_rt,
 684 };