| 1 | /* |
| 2 | * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR |
| 3 | * policies) |
| 4 | */ |
| 5 | |
| 6 | #ifdef CONFIG_SMP |
| 7 | |
| 8 | static inline int rt_overloaded(struct rq *rq) |
| 9 | { |
| 10 | return atomic_read(&rq->rd->rto_count); |
| 11 | } |
| 12 | |
| 13 | static inline void rt_set_overload(struct rq *rq) |
| 14 | { |
| 15 | cpu_set(rq->cpu, rq->rd->rto_mask); |
| 16 | /* |
| 17 | * Make sure the mask is visible before we set |
| 18 | * the overload count. That is checked to determine |
| 19 | * if we should look at the mask. It would be a shame |
| 20 | * if we looked at the mask, but the mask was not |
| 21 | * updated yet. |
| 22 | */ |
| 23 | wmb(); |
| 24 | atomic_inc(&rq->rd->rto_count); |
| 25 | } |
| 26 | |
| 27 | static inline void rt_clear_overload(struct rq *rq) |
| 28 | { |
| 29 | /* the order here really doesn't matter */ |
| 30 | atomic_dec(&rq->rd->rto_count); |
| 31 | cpu_clear(rq->cpu, rq->rd->rto_mask); |
| 32 | } |
| 33 | |
| 34 | static void update_rt_migration(struct rq *rq) |
| 35 | { |
| 36 | if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) { |
| 37 | rt_set_overload(rq); |
| 38 | rq->rt.overloaded = 1; |
| 39 | } else { |
| 40 | rt_clear_overload(rq); |
| 41 | rq->rt.overloaded = 0; |
| 42 | } |
| 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 | #ifdef CONFIG_SMP |
| 154 | static int find_lowest_rq(struct task_struct *task); |
| 155 | |
| 156 | static int select_task_rq_rt(struct task_struct *p, int sync) |
| 157 | { |
| 158 | struct rq *rq = task_rq(p); |
| 159 | |
| 160 | /* |
| 161 | * If the current task is an RT task, then |
| 162 | * try to see if we can wake this RT task up on another |
| 163 | * runqueue. Otherwise simply start this RT task |
| 164 | * on its current runqueue. |
| 165 | * |
| 166 | * We want to avoid overloading runqueues. Even if |
| 167 | * the RT task is of higher priority than the current RT task. |
| 168 | * RT tasks behave differently than other tasks. If |
| 169 | * one gets preempted, we try to push it off to another queue. |
| 170 | * So trying to keep a preempting RT task on the same |
| 171 | * cache hot CPU will force the running RT task to |
| 172 | * a cold CPU. So we waste all the cache for the lower |
| 173 | * RT task in hopes of saving some of a RT task |
| 174 | * that is just being woken and probably will have |
| 175 | * cold cache anyway. |
| 176 | */ |
| 177 | if (unlikely(rt_task(rq->curr)) && |
| 178 | (p->nr_cpus_allowed > 1)) { |
| 179 | int cpu = find_lowest_rq(p); |
| 180 | |
| 181 | return (cpu == -1) ? task_cpu(p) : cpu; |
| 182 | } |
| 183 | |
| 184 | /* |
| 185 | * Otherwise, just let it ride on the affined RQ and the |
| 186 | * post-schedule router will push the preempted task away |
| 187 | */ |
| 188 | return task_cpu(p); |
| 189 | } |
| 190 | #endif /* CONFIG_SMP */ |
| 191 | |
| 192 | /* |
| 193 | * Preempt the current task with a newly woken task if needed: |
| 194 | */ |
| 195 | static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p) |
| 196 | { |
| 197 | if (p->prio < rq->curr->prio) |
| 198 | resched_task(rq->curr); |
| 199 | } |
| 200 | |
| 201 | static struct task_struct *pick_next_task_rt(struct rq *rq) |
| 202 | { |
| 203 | struct rt_prio_array *array = &rq->rt.active; |
| 204 | struct task_struct *next; |
| 205 | struct list_head *queue; |
| 206 | int idx; |
| 207 | |
| 208 | idx = sched_find_first_bit(array->bitmap); |
| 209 | if (idx >= MAX_RT_PRIO) |
| 210 | return NULL; |
| 211 | |
| 212 | queue = array->queue + idx; |
| 213 | next = list_entry(queue->next, struct task_struct, run_list); |
| 214 | |
| 215 | next->se.exec_start = rq->clock; |
| 216 | |
| 217 | return next; |
| 218 | } |
| 219 | |
| 220 | static void put_prev_task_rt(struct rq *rq, struct task_struct *p) |
| 221 | { |
| 222 | update_curr_rt(rq); |
| 223 | p->se.exec_start = 0; |
| 224 | } |
| 225 | |
| 226 | #ifdef CONFIG_SMP |
| 227 | /* Only try algorithms three times */ |
| 228 | #define RT_MAX_TRIES 3 |
| 229 | |
| 230 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest); |
| 231 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep); |
| 232 | |
| 233 | static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) |
| 234 | { |
| 235 | if (!task_running(rq, p) && |
| 236 | (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) && |
| 237 | (p->nr_cpus_allowed > 1)) |
| 238 | return 1; |
| 239 | return 0; |
| 240 | } |
| 241 | |
| 242 | /* Return the second highest RT task, NULL otherwise */ |
| 243 | static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu) |
| 244 | { |
| 245 | struct rt_prio_array *array = &rq->rt.active; |
| 246 | struct task_struct *next; |
| 247 | struct list_head *queue; |
| 248 | int idx; |
| 249 | |
| 250 | if (likely(rq->rt.rt_nr_running < 2)) |
| 251 | return NULL; |
| 252 | |
| 253 | idx = sched_find_first_bit(array->bitmap); |
| 254 | if (unlikely(idx >= MAX_RT_PRIO)) { |
| 255 | WARN_ON(1); /* rt_nr_running is bad */ |
| 256 | return NULL; |
| 257 | } |
| 258 | |
| 259 | queue = array->queue + idx; |
| 260 | BUG_ON(list_empty(queue)); |
| 261 | |
| 262 | next = list_entry(queue->next, struct task_struct, run_list); |
| 263 | if (unlikely(pick_rt_task(rq, next, cpu))) |
| 264 | goto out; |
| 265 | |
| 266 | if (queue->next->next != queue) { |
| 267 | /* same prio task */ |
| 268 | next = list_entry(queue->next->next, struct task_struct, |
| 269 | run_list); |
| 270 | if (pick_rt_task(rq, next, cpu)) |
| 271 | goto out; |
| 272 | } |
| 273 | |
| 274 | retry: |
| 275 | /* slower, but more flexible */ |
| 276 | idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1); |
| 277 | if (unlikely(idx >= MAX_RT_PRIO)) |
| 278 | return NULL; |
| 279 | |
| 280 | queue = array->queue + idx; |
| 281 | BUG_ON(list_empty(queue)); |
| 282 | |
| 283 | list_for_each_entry(next, queue, run_list) { |
| 284 | if (pick_rt_task(rq, next, cpu)) |
| 285 | goto out; |
| 286 | } |
| 287 | |
| 288 | goto retry; |
| 289 | |
| 290 | out: |
| 291 | return next; |
| 292 | } |
| 293 | |
| 294 | static DEFINE_PER_CPU(cpumask_t, local_cpu_mask); |
| 295 | |
| 296 | static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask) |
| 297 | { |
| 298 | int lowest_prio = -1; |
| 299 | int lowest_cpu = -1; |
| 300 | int count = 0; |
| 301 | int cpu; |
| 302 | |
| 303 | cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed); |
| 304 | |
| 305 | /* |
| 306 | * Scan each rq for the lowest prio. |
| 307 | */ |
| 308 | for_each_cpu_mask(cpu, *lowest_mask) { |
| 309 | struct rq *rq = cpu_rq(cpu); |
| 310 | |
| 311 | /* We look for lowest RT prio or non-rt CPU */ |
| 312 | if (rq->rt.highest_prio >= MAX_RT_PRIO) { |
| 313 | /* |
| 314 | * if we already found a low RT queue |
| 315 | * and now we found this non-rt queue |
| 316 | * clear the mask and set our bit. |
| 317 | * Otherwise just return the queue as is |
| 318 | * and the count==1 will cause the algorithm |
| 319 | * to use the first bit found. |
| 320 | */ |
| 321 | if (lowest_cpu != -1) { |
| 322 | cpus_clear(*lowest_mask); |
| 323 | cpu_set(rq->cpu, *lowest_mask); |
| 324 | } |
| 325 | return 1; |
| 326 | } |
| 327 | |
| 328 | /* no locking for now */ |
| 329 | if ((rq->rt.highest_prio > task->prio) |
| 330 | && (rq->rt.highest_prio >= lowest_prio)) { |
| 331 | if (rq->rt.highest_prio > lowest_prio) { |
| 332 | /* new low - clear old data */ |
| 333 | lowest_prio = rq->rt.highest_prio; |
| 334 | lowest_cpu = cpu; |
| 335 | count = 0; |
| 336 | } |
| 337 | count++; |
| 338 | } else |
| 339 | cpu_clear(cpu, *lowest_mask); |
| 340 | } |
| 341 | |
| 342 | /* |
| 343 | * Clear out all the set bits that represent |
| 344 | * runqueues that were of higher prio than |
| 345 | * the lowest_prio. |
| 346 | */ |
| 347 | if (lowest_cpu > 0) { |
| 348 | /* |
| 349 | * Perhaps we could add another cpumask op to |
| 350 | * zero out bits. Like cpu_zero_bits(cpumask, nrbits); |
| 351 | * Then that could be optimized to use memset and such. |
| 352 | */ |
| 353 | for_each_cpu_mask(cpu, *lowest_mask) { |
| 354 | if (cpu >= lowest_cpu) |
| 355 | break; |
| 356 | cpu_clear(cpu, *lowest_mask); |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | return count; |
| 361 | } |
| 362 | |
| 363 | static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask) |
| 364 | { |
| 365 | int first; |
| 366 | |
| 367 | /* "this_cpu" is cheaper to preempt than a remote processor */ |
| 368 | if ((this_cpu != -1) && cpu_isset(this_cpu, *mask)) |
| 369 | return this_cpu; |
| 370 | |
| 371 | first = first_cpu(*mask); |
| 372 | if (first != NR_CPUS) |
| 373 | return first; |
| 374 | |
| 375 | return -1; |
| 376 | } |
| 377 | |
| 378 | static int find_lowest_rq(struct task_struct *task) |
| 379 | { |
| 380 | struct sched_domain *sd; |
| 381 | cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask); |
| 382 | int this_cpu = smp_processor_id(); |
| 383 | int cpu = task_cpu(task); |
| 384 | int count = find_lowest_cpus(task, lowest_mask); |
| 385 | |
| 386 | if (!count) |
| 387 | return -1; /* No targets found */ |
| 388 | |
| 389 | /* |
| 390 | * There is no sense in performing an optimal search if only one |
| 391 | * target is found. |
| 392 | */ |
| 393 | if (count == 1) |
| 394 | return first_cpu(*lowest_mask); |
| 395 | |
| 396 | /* |
| 397 | * At this point we have built a mask of cpus representing the |
| 398 | * lowest priority tasks in the system. Now we want to elect |
| 399 | * the best one based on our affinity and topology. |
| 400 | * |
| 401 | * We prioritize the last cpu that the task executed on since |
| 402 | * it is most likely cache-hot in that location. |
| 403 | */ |
| 404 | if (cpu_isset(cpu, *lowest_mask)) |
| 405 | return cpu; |
| 406 | |
| 407 | /* |
| 408 | * Otherwise, we consult the sched_domains span maps to figure |
| 409 | * out which cpu is logically closest to our hot cache data. |
| 410 | */ |
| 411 | if (this_cpu == cpu) |
| 412 | this_cpu = -1; /* Skip this_cpu opt if the same */ |
| 413 | |
| 414 | for_each_domain(cpu, sd) { |
| 415 | if (sd->flags & SD_WAKE_AFFINE) { |
| 416 | cpumask_t domain_mask; |
| 417 | int best_cpu; |
| 418 | |
| 419 | cpus_and(domain_mask, sd->span, *lowest_mask); |
| 420 | |
| 421 | best_cpu = pick_optimal_cpu(this_cpu, |
| 422 | &domain_mask); |
| 423 | if (best_cpu != -1) |
| 424 | return best_cpu; |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | /* |
| 429 | * And finally, if there were no matches within the domains |
| 430 | * just give the caller *something* to work with from the compatible |
| 431 | * locations. |
| 432 | */ |
| 433 | return pick_optimal_cpu(this_cpu, lowest_mask); |
| 434 | } |
| 435 | |
| 436 | /* Will lock the rq it finds */ |
| 437 | static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) |
| 438 | { |
| 439 | struct rq *lowest_rq = NULL; |
| 440 | int tries; |
| 441 | int cpu; |
| 442 | |
| 443 | for (tries = 0; tries < RT_MAX_TRIES; tries++) { |
| 444 | cpu = find_lowest_rq(task); |
| 445 | |
| 446 | if ((cpu == -1) || (cpu == rq->cpu)) |
| 447 | break; |
| 448 | |
| 449 | lowest_rq = cpu_rq(cpu); |
| 450 | |
| 451 | /* if the prio of this runqueue changed, try again */ |
| 452 | if (double_lock_balance(rq, lowest_rq)) { |
| 453 | /* |
| 454 | * We had to unlock the run queue. In |
| 455 | * the mean time, task could have |
| 456 | * migrated already or had its affinity changed. |
| 457 | * Also make sure that it wasn't scheduled on its rq. |
| 458 | */ |
| 459 | if (unlikely(task_rq(task) != rq || |
| 460 | !cpu_isset(lowest_rq->cpu, |
| 461 | task->cpus_allowed) || |
| 462 | task_running(rq, task) || |
| 463 | !task->se.on_rq)) { |
| 464 | |
| 465 | spin_unlock(&lowest_rq->lock); |
| 466 | lowest_rq = NULL; |
| 467 | break; |
| 468 | } |
| 469 | } |
| 470 | |
| 471 | /* If this rq is still suitable use it. */ |
| 472 | if (lowest_rq->rt.highest_prio > task->prio) |
| 473 | break; |
| 474 | |
| 475 | /* try again */ |
| 476 | spin_unlock(&lowest_rq->lock); |
| 477 | lowest_rq = NULL; |
| 478 | } |
| 479 | |
| 480 | return lowest_rq; |
| 481 | } |
| 482 | |
| 483 | /* |
| 484 | * If the current CPU has more than one RT task, see if the non |
| 485 | * running task can migrate over to a CPU that is running a task |
| 486 | * of lesser priority. |
| 487 | */ |
| 488 | static int push_rt_task(struct rq *rq) |
| 489 | { |
| 490 | struct task_struct *next_task; |
| 491 | struct rq *lowest_rq; |
| 492 | int ret = 0; |
| 493 | int paranoid = RT_MAX_TRIES; |
| 494 | |
| 495 | if (!rq->rt.overloaded) |
| 496 | return 0; |
| 497 | |
| 498 | next_task = pick_next_highest_task_rt(rq, -1); |
| 499 | if (!next_task) |
| 500 | return 0; |
| 501 | |
| 502 | retry: |
| 503 | if (unlikely(next_task == rq->curr)) { |
| 504 | WARN_ON(1); |
| 505 | return 0; |
| 506 | } |
| 507 | |
| 508 | /* |
| 509 | * It's possible that the next_task slipped in of |
| 510 | * higher priority than current. If that's the case |
| 511 | * just reschedule current. |
| 512 | */ |
| 513 | if (unlikely(next_task->prio < rq->curr->prio)) { |
| 514 | resched_task(rq->curr); |
| 515 | return 0; |
| 516 | } |
| 517 | |
| 518 | /* We might release rq lock */ |
| 519 | get_task_struct(next_task); |
| 520 | |
| 521 | /* find_lock_lowest_rq locks the rq if found */ |
| 522 | lowest_rq = find_lock_lowest_rq(next_task, rq); |
| 523 | if (!lowest_rq) { |
| 524 | struct task_struct *task; |
| 525 | /* |
| 526 | * find lock_lowest_rq releases rq->lock |
| 527 | * so it is possible that next_task has changed. |
| 528 | * If it has, then try again. |
| 529 | */ |
| 530 | task = pick_next_highest_task_rt(rq, -1); |
| 531 | if (unlikely(task != next_task) && task && paranoid--) { |
| 532 | put_task_struct(next_task); |
| 533 | next_task = task; |
| 534 | goto retry; |
| 535 | } |
| 536 | goto out; |
| 537 | } |
| 538 | |
| 539 | deactivate_task(rq, next_task, 0); |
| 540 | set_task_cpu(next_task, lowest_rq->cpu); |
| 541 | activate_task(lowest_rq, next_task, 0); |
| 542 | |
| 543 | resched_task(lowest_rq->curr); |
| 544 | |
| 545 | spin_unlock(&lowest_rq->lock); |
| 546 | |
| 547 | ret = 1; |
| 548 | out: |
| 549 | put_task_struct(next_task); |
| 550 | |
| 551 | return ret; |
| 552 | } |
| 553 | |
| 554 | /* |
| 555 | * TODO: Currently we just use the second highest prio task on |
| 556 | * the queue, and stop when it can't migrate (or there's |
| 557 | * no more RT tasks). There may be a case where a lower |
| 558 | * priority RT task has a different affinity than the |
| 559 | * higher RT task. In this case the lower RT task could |
| 560 | * possibly be able to migrate where as the higher priority |
| 561 | * RT task could not. We currently ignore this issue. |
| 562 | * Enhancements are welcome! |
| 563 | */ |
| 564 | static void push_rt_tasks(struct rq *rq) |
| 565 | { |
| 566 | /* push_rt_task will return true if it moved an RT */ |
| 567 | while (push_rt_task(rq)) |
| 568 | ; |
| 569 | } |
| 570 | |
| 571 | static int pull_rt_task(struct rq *this_rq) |
| 572 | { |
| 573 | int this_cpu = this_rq->cpu, ret = 0, cpu; |
| 574 | struct task_struct *p, *next; |
| 575 | struct rq *src_rq; |
| 576 | |
| 577 | if (likely(!rt_overloaded(this_rq))) |
| 578 | return 0; |
| 579 | |
| 580 | next = pick_next_task_rt(this_rq); |
| 581 | |
| 582 | for_each_cpu_mask(cpu, this_rq->rd->rto_mask) { |
| 583 | if (this_cpu == cpu) |
| 584 | continue; |
| 585 | |
| 586 | src_rq = cpu_rq(cpu); |
| 587 | if (unlikely(src_rq->rt.rt_nr_running <= 1)) { |
| 588 | /* |
| 589 | * It is possible that overlapping cpusets |
| 590 | * will miss clearing a non overloaded runqueue. |
| 591 | * Clear it now. |
| 592 | */ |
| 593 | if (double_lock_balance(this_rq, src_rq)) { |
| 594 | /* unlocked our runqueue lock */ |
| 595 | struct task_struct *old_next = next; |
| 596 | |
| 597 | next = pick_next_task_rt(this_rq); |
| 598 | if (next != old_next) |
| 599 | ret = 1; |
| 600 | } |
| 601 | if (likely(src_rq->rt.rt_nr_running <= 1)) { |
| 602 | /* |
| 603 | * Small chance that this_rq->curr changed |
| 604 | * but it's really harmless here. |
| 605 | */ |
| 606 | rt_clear_overload(this_rq); |
| 607 | } else { |
| 608 | /* |
| 609 | * Heh, the src_rq is now overloaded, since |
| 610 | * we already have the src_rq lock, go straight |
| 611 | * to pulling tasks from it. |
| 612 | */ |
| 613 | goto try_pulling; |
| 614 | } |
| 615 | spin_unlock(&src_rq->lock); |
| 616 | continue; |
| 617 | } |
| 618 | |
| 619 | /* |
| 620 | * We can potentially drop this_rq's lock in |
| 621 | * double_lock_balance, and another CPU could |
| 622 | * steal our next task - hence we must cause |
| 623 | * the caller to recalculate the next task |
| 624 | * in that case: |
| 625 | */ |
| 626 | if (double_lock_balance(this_rq, src_rq)) { |
| 627 | struct task_struct *old_next = next; |
| 628 | |
| 629 | next = pick_next_task_rt(this_rq); |
| 630 | if (next != old_next) |
| 631 | ret = 1; |
| 632 | } |
| 633 | |
| 634 | /* |
| 635 | * Are there still pullable RT tasks? |
| 636 | */ |
| 637 | if (src_rq->rt.rt_nr_running <= 1) { |
| 638 | spin_unlock(&src_rq->lock); |
| 639 | continue; |
| 640 | } |
| 641 | |
| 642 | try_pulling: |
| 643 | p = pick_next_highest_task_rt(src_rq, this_cpu); |
| 644 | |
| 645 | /* |
| 646 | * Do we have an RT task that preempts |
| 647 | * the to-be-scheduled task? |
| 648 | */ |
| 649 | if (p && (!next || (p->prio < next->prio))) { |
| 650 | WARN_ON(p == src_rq->curr); |
| 651 | WARN_ON(!p->se.on_rq); |
| 652 | |
| 653 | /* |
| 654 | * There's a chance that p is higher in priority |
| 655 | * than what's currently running on its cpu. |
| 656 | * This is just that p is wakeing up and hasn't |
| 657 | * had a chance to schedule. We only pull |
| 658 | * p if it is lower in priority than the |
| 659 | * current task on the run queue or |
| 660 | * this_rq next task is lower in prio than |
| 661 | * the current task on that rq. |
| 662 | */ |
| 663 | if (p->prio < src_rq->curr->prio || |
| 664 | (next && next->prio < src_rq->curr->prio)) |
| 665 | goto out; |
| 666 | |
| 667 | ret = 1; |
| 668 | |
| 669 | deactivate_task(src_rq, p, 0); |
| 670 | set_task_cpu(p, this_cpu); |
| 671 | activate_task(this_rq, p, 0); |
| 672 | /* |
| 673 | * We continue with the search, just in |
| 674 | * case there's an even higher prio task |
| 675 | * in another runqueue. (low likelyhood |
| 676 | * but possible) |
| 677 | * |
| 678 | * Update next so that we won't pick a task |
| 679 | * on another cpu with a priority lower (or equal) |
| 680 | * than the one we just picked. |
| 681 | */ |
| 682 | next = p; |
| 683 | |
| 684 | } |
| 685 | out: |
| 686 | spin_unlock(&src_rq->lock); |
| 687 | } |
| 688 | |
| 689 | return ret; |
| 690 | } |
| 691 | |
| 692 | static void pre_schedule_rt(struct rq *rq, struct task_struct *prev) |
| 693 | { |
| 694 | /* Try to pull RT tasks here if we lower this rq's prio */ |
| 695 | if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio) |
| 696 | pull_rt_task(rq); |
| 697 | } |
| 698 | |
| 699 | static void post_schedule_rt(struct rq *rq) |
| 700 | { |
| 701 | /* |
| 702 | * If we have more than one rt_task queued, then |
| 703 | * see if we can push the other rt_tasks off to other CPUS. |
| 704 | * Note we may release the rq lock, and since |
| 705 | * the lock was owned by prev, we need to release it |
| 706 | * first via finish_lock_switch and then reaquire it here. |
| 707 | */ |
| 708 | if (unlikely(rq->rt.overloaded)) { |
| 709 | spin_lock_irq(&rq->lock); |
| 710 | push_rt_tasks(rq); |
| 711 | spin_unlock_irq(&rq->lock); |
| 712 | } |
| 713 | } |
| 714 | |
| 715 | |
| 716 | static void task_wake_up_rt(struct rq *rq, struct task_struct *p) |
| 717 | { |
| 718 | if (!task_running(rq, p) && |
| 719 | (p->prio >= rq->rt.highest_prio) && |
| 720 | rq->rt.overloaded) |
| 721 | push_rt_tasks(rq); |
| 722 | } |
| 723 | |
| 724 | static unsigned long |
| 725 | load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 726 | unsigned long max_load_move, |
| 727 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 728 | int *all_pinned, int *this_best_prio) |
| 729 | { |
| 730 | /* don't touch RT tasks */ |
| 731 | return 0; |
| 732 | } |
| 733 | |
| 734 | static int |
| 735 | move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 736 | struct sched_domain *sd, enum cpu_idle_type idle) |
| 737 | { |
| 738 | /* don't touch RT tasks */ |
| 739 | return 0; |
| 740 | } |
| 741 | |
| 742 | static void set_cpus_allowed_rt(struct task_struct *p, cpumask_t *new_mask) |
| 743 | { |
| 744 | int weight = cpus_weight(*new_mask); |
| 745 | |
| 746 | BUG_ON(!rt_task(p)); |
| 747 | |
| 748 | /* |
| 749 | * Update the migration status of the RQ if we have an RT task |
| 750 | * which is running AND changing its weight value. |
| 751 | */ |
| 752 | if (p->se.on_rq && (weight != p->nr_cpus_allowed)) { |
| 753 | struct rq *rq = task_rq(p); |
| 754 | |
| 755 | if ((p->nr_cpus_allowed <= 1) && (weight > 1)) { |
| 756 | rq->rt.rt_nr_migratory++; |
| 757 | } else if ((p->nr_cpus_allowed > 1) && (weight <= 1)) { |
| 758 | BUG_ON(!rq->rt.rt_nr_migratory); |
| 759 | rq->rt.rt_nr_migratory--; |
| 760 | } |
| 761 | |
| 762 | update_rt_migration(rq); |
| 763 | } |
| 764 | |
| 765 | p->cpus_allowed = *new_mask; |
| 766 | p->nr_cpus_allowed = weight; |
| 767 | } |
| 768 | |
| 769 | /* Assumes rq->lock is held */ |
| 770 | static void join_domain_rt(struct rq *rq) |
| 771 | { |
| 772 | if (rq->rt.overloaded) |
| 773 | rt_set_overload(rq); |
| 774 | } |
| 775 | |
| 776 | /* Assumes rq->lock is held */ |
| 777 | static void leave_domain_rt(struct rq *rq) |
| 778 | { |
| 779 | if (rq->rt.overloaded) |
| 780 | rt_clear_overload(rq); |
| 781 | } |
| 782 | |
| 783 | /* |
| 784 | * When switch from the rt queue, we bring ourselves to a position |
| 785 | * that we might want to pull RT tasks from other runqueues. |
| 786 | */ |
| 787 | static void switched_from_rt(struct rq *rq, struct task_struct *p, |
| 788 | int running) |
| 789 | { |
| 790 | /* |
| 791 | * If there are other RT tasks then we will reschedule |
| 792 | * and the scheduling of the other RT tasks will handle |
| 793 | * the balancing. But if we are the last RT task |
| 794 | * we may need to handle the pulling of RT tasks |
| 795 | * now. |
| 796 | */ |
| 797 | if (!rq->rt.rt_nr_running) |
| 798 | pull_rt_task(rq); |
| 799 | } |
| 800 | #endif /* CONFIG_SMP */ |
| 801 | |
| 802 | /* |
| 803 | * When switching a task to RT, we may overload the runqueue |
| 804 | * with RT tasks. In this case we try to push them off to |
| 805 | * other runqueues. |
| 806 | */ |
| 807 | static void switched_to_rt(struct rq *rq, struct task_struct *p, |
| 808 | int running) |
| 809 | { |
| 810 | int check_resched = 1; |
| 811 | |
| 812 | /* |
| 813 | * If we are already running, then there's nothing |
| 814 | * that needs to be done. But if we are not running |
| 815 | * we may need to preempt the current running task. |
| 816 | * If that current running task is also an RT task |
| 817 | * then see if we can move to another run queue. |
| 818 | */ |
| 819 | if (!running) { |
| 820 | #ifdef CONFIG_SMP |
| 821 | if (rq->rt.overloaded && push_rt_task(rq) && |
| 822 | /* Don't resched if we changed runqueues */ |
| 823 | rq != task_rq(p)) |
| 824 | check_resched = 0; |
| 825 | #endif /* CONFIG_SMP */ |
| 826 | if (check_resched && p->prio < rq->curr->prio) |
| 827 | resched_task(rq->curr); |
| 828 | } |
| 829 | } |
| 830 | |
| 831 | /* |
| 832 | * Priority of the task has changed. This may cause |
| 833 | * us to initiate a push or pull. |
| 834 | */ |
| 835 | static void prio_changed_rt(struct rq *rq, struct task_struct *p, |
| 836 | int oldprio, int running) |
| 837 | { |
| 838 | if (running) { |
| 839 | #ifdef CONFIG_SMP |
| 840 | /* |
| 841 | * If our priority decreases while running, we |
| 842 | * may need to pull tasks to this runqueue. |
| 843 | */ |
| 844 | if (oldprio < p->prio) |
| 845 | pull_rt_task(rq); |
| 846 | /* |
| 847 | * If there's a higher priority task waiting to run |
| 848 | * then reschedule. |
| 849 | */ |
| 850 | if (p->prio > rq->rt.highest_prio) |
| 851 | resched_task(p); |
| 852 | #else |
| 853 | /* For UP simply resched on drop of prio */ |
| 854 | if (oldprio < p->prio) |
| 855 | resched_task(p); |
| 856 | #endif /* CONFIG_SMP */ |
| 857 | } else { |
| 858 | /* |
| 859 | * This task is not running, but if it is |
| 860 | * greater than the current running task |
| 861 | * then reschedule. |
| 862 | */ |
| 863 | if (p->prio < rq->curr->prio) |
| 864 | resched_task(rq->curr); |
| 865 | } |
| 866 | } |
| 867 | |
| 868 | |
| 869 | static void task_tick_rt(struct rq *rq, struct task_struct *p) |
| 870 | { |
| 871 | update_curr_rt(rq); |
| 872 | |
| 873 | /* |
| 874 | * RR tasks need a special form of timeslice management. |
| 875 | * FIFO tasks have no timeslices. |
| 876 | */ |
| 877 | if (p->policy != SCHED_RR) |
| 878 | return; |
| 879 | |
| 880 | if (--p->time_slice) |
| 881 | return; |
| 882 | |
| 883 | p->time_slice = DEF_TIMESLICE; |
| 884 | |
| 885 | /* |
| 886 | * Requeue to the end of queue if we are not the only element |
| 887 | * on the queue: |
| 888 | */ |
| 889 | if (p->run_list.prev != p->run_list.next) { |
| 890 | requeue_task_rt(rq, p); |
| 891 | set_tsk_need_resched(p); |
| 892 | } |
| 893 | } |
| 894 | |
| 895 | static void set_curr_task_rt(struct rq *rq) |
| 896 | { |
| 897 | struct task_struct *p = rq->curr; |
| 898 | |
| 899 | p->se.exec_start = rq->clock; |
| 900 | } |
| 901 | |
| 902 | const struct sched_class rt_sched_class = { |
| 903 | .next = &fair_sched_class, |
| 904 | .enqueue_task = enqueue_task_rt, |
| 905 | .dequeue_task = dequeue_task_rt, |
| 906 | .yield_task = yield_task_rt, |
| 907 | #ifdef CONFIG_SMP |
| 908 | .select_task_rq = select_task_rq_rt, |
| 909 | #endif /* CONFIG_SMP */ |
| 910 | |
| 911 | .check_preempt_curr = check_preempt_curr_rt, |
| 912 | |
| 913 | .pick_next_task = pick_next_task_rt, |
| 914 | .put_prev_task = put_prev_task_rt, |
| 915 | |
| 916 | #ifdef CONFIG_SMP |
| 917 | .load_balance = load_balance_rt, |
| 918 | .move_one_task = move_one_task_rt, |
| 919 | .set_cpus_allowed = set_cpus_allowed_rt, |
| 920 | .join_domain = join_domain_rt, |
| 921 | .leave_domain = leave_domain_rt, |
| 922 | .pre_schedule = pre_schedule_rt, |
| 923 | .post_schedule = post_schedule_rt, |
| 924 | .task_wake_up = task_wake_up_rt, |
| 925 | .switched_from = switched_from_rt, |
| 926 | #endif |
| 927 | |
| 928 | .set_curr_task = set_curr_task_rt, |
| 929 | .task_tick = task_tick_rt, |
| 930 | |
| 931 | .prio_changed = prio_changed_rt, |
| 932 | .switched_to = switched_to_rt, |
| 933 | }; |