1 /* 2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR 3 * policies) 4 */ 5 6 #include "sched.h" 7 8 #include <linux/slab.h> 9 10 int sched_rr_timeslice = RR_TIMESLICE; 11 12 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); 13 14 struct rt_bandwidth def_rt_bandwidth; 15 16 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) 17 { 18 struct rt_bandwidth *rt_b = 19 container_of(timer, struct rt_bandwidth, rt_period_timer); 20 ktime_t now; 21 int overrun; 22 int idle = 0; 23 24 for (;;) { 25 now = hrtimer_cb_get_time(timer); 26 overrun = hrtimer_forward(timer, now, rt_b->rt_period); 27 28 if (!overrun) 29 break; 30 31 idle = do_sched_rt_period_timer(rt_b, overrun); 32 } 33 34 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 35 } 36 37 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) 38 { 39 rt_b->rt_period = ns_to_ktime(period); 40 rt_b->rt_runtime = runtime; 41 42 raw_spin_lock_init(&rt_b->rt_runtime_lock); 43 44 hrtimer_init(&rt_b->rt_period_timer, 45 CLOCK_MONOTONIC, HRTIMER_MODE_REL); 46 rt_b->rt_period_timer.function = sched_rt_period_timer; 47 } 48 49 static void start_rt_bandwidth(struct rt_bandwidth *rt_b) 50 { 51 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) 52 return; 53 54 if (hrtimer_active(&rt_b->rt_period_timer)) 55 return; 56 57 raw_spin_lock(&rt_b->rt_runtime_lock); 58 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period); 59 raw_spin_unlock(&rt_b->rt_runtime_lock); 60 } 61 62 void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) 63 { 64 struct rt_prio_array *array; 65 int i; 66 67 array = &rt_rq->active; 68 for (i = 0; i < MAX_RT_PRIO; i++) { 69 INIT_LIST_HEAD(array->queue + i); 70 __clear_bit(i, array->bitmap); 71 } 72 /* delimiter for bitsearch: */ 73 __set_bit(MAX_RT_PRIO, array->bitmap); 74 75 #if defined CONFIG_SMP 76 rt_rq->highest_prio.curr = MAX_RT_PRIO; 77 rt_rq->highest_prio.next = MAX_RT_PRIO; 78 rt_rq->rt_nr_migratory = 0; 79 rt_rq->overloaded = 0; 80 plist_head_init(&rt_rq->pushable_tasks); 81 #endif 82 /* We start is dequeued state, because no RT tasks are queued */ 83 rt_rq->rt_queued = 0; 84 85 rt_rq->rt_time = 0; 86 rt_rq->rt_throttled = 0; 87 rt_rq->rt_runtime = 0; 88 raw_spin_lock_init(&rt_rq->rt_runtime_lock); 89 } 90 91 #ifdef CONFIG_RT_GROUP_SCHED 92 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) 93 { 94 hrtimer_cancel(&rt_b->rt_period_timer); 95 } 96 97 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q) 98 99 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) 100 { 101 #ifdef CONFIG_SCHED_DEBUG 102 WARN_ON_ONCE(!rt_entity_is_task(rt_se)); 103 #endif 104 return container_of(rt_se, struct task_struct, rt); 105 } 106 107 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) 108 { 109 return rt_rq->rq; 110 } 111 112 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) 113 { 114 return rt_se->rt_rq; 115 } 116 117 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) 118 { 119 struct rt_rq *rt_rq = rt_se->rt_rq; 120 121 return rt_rq->rq; 122 } 123 124 void free_rt_sched_group(struct task_group *tg) 125 { 126 int i; 127 128 if (tg->rt_se) 129 destroy_rt_bandwidth(&tg->rt_bandwidth); 130 131 for_each_possible_cpu(i) { 132 if (tg->rt_rq) 133 kfree(tg->rt_rq[i]); 134 if (tg->rt_se) 135 kfree(tg->rt_se[i]); 136 } 137 138 kfree(tg->rt_rq); 139 kfree(tg->rt_se); 140 } 141 142 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 143 struct sched_rt_entity *rt_se, int cpu, 144 struct sched_rt_entity *parent) 145 { 146 struct rq *rq = cpu_rq(cpu); 147 148 rt_rq->highest_prio.curr = MAX_RT_PRIO; 149 rt_rq->rt_nr_boosted = 0; 150 rt_rq->rq = rq; 151 rt_rq->tg = tg; 152 153 tg->rt_rq[cpu] = rt_rq; 154 tg->rt_se[cpu] = rt_se; 155 156 if (!rt_se) 157 return; 158 159 if (!parent) 160 rt_se->rt_rq = &rq->rt; 161 else 162 rt_se->rt_rq = parent->my_q; 163 164 rt_se->my_q = rt_rq; 165 rt_se->parent = parent; 166 INIT_LIST_HEAD(&rt_se->run_list); 167 } 168 169 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) 170 { 171 struct rt_rq *rt_rq; 172 struct sched_rt_entity *rt_se; 173 int i; 174 175 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); 176 if (!tg->rt_rq) 177 goto err; 178 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); 179 if (!tg->rt_se) 180 goto err; 181 182 init_rt_bandwidth(&tg->rt_bandwidth, 183 ktime_to_ns(def_rt_bandwidth.rt_period), 0); 184 185 for_each_possible_cpu(i) { 186 rt_rq = kzalloc_node(sizeof(struct rt_rq), 187 GFP_KERNEL, cpu_to_node(i)); 188 if (!rt_rq) 189 goto err; 190 191 rt_se = kzalloc_node(sizeof(struct sched_rt_entity), 192 GFP_KERNEL, cpu_to_node(i)); 193 if (!rt_se) 194 goto err_free_rq; 195 196 init_rt_rq(rt_rq, cpu_rq(i)); 197 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; 198 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]); 199 } 200 201 return 1; 202 203 err_free_rq: 204 kfree(rt_rq); 205 err: 206 return 0; 207 } 208 209 #else /* CONFIG_RT_GROUP_SCHED */ 210 211 #define rt_entity_is_task(rt_se) (1) 212 213 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) 214 { 215 return container_of(rt_se, struct task_struct, rt); 216 } 217 218 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) 219 { 220 return container_of(rt_rq, struct rq, rt); 221 } 222 223 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) 224 { 225 struct task_struct *p = rt_task_of(rt_se); 226 227 return task_rq(p); 228 } 229 230 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) 231 { 232 struct rq *rq = rq_of_rt_se(rt_se); 233 234 return &rq->rt; 235 } 236 237 void free_rt_sched_group(struct task_group *tg) { } 238 239 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) 240 { 241 return 1; 242 } 243 #endif /* CONFIG_RT_GROUP_SCHED */ 244 245 #ifdef CONFIG_SMP 246 247 static int pull_rt_task(struct rq *this_rq); 248 249 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) 250 { 251 /* Try to pull RT tasks here if we lower this rq's prio */ 252 return rq->rt.highest_prio.curr > prev->prio; 253 } 254 255 static inline int rt_overloaded(struct rq *rq) 256 { 257 return atomic_read(&rq->rd->rto_count); 258 } 259 260 static inline void rt_set_overload(struct rq *rq) 261 { 262 if (!rq->online) 263 return; 264 265 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask); 266 /* 267 * Make sure the mask is visible before we set 268 * the overload count. That is checked to determine 269 * if we should look at the mask. It would be a shame 270 * if we looked at the mask, but the mask was not 271 * updated yet. 272 * 273 * Matched by the barrier in pull_rt_task(). 274 */ 275 smp_wmb(); 276 atomic_inc(&rq->rd->rto_count); 277 } 278 279 static inline void rt_clear_overload(struct rq *rq) 280 { 281 if (!rq->online) 282 return; 283 284 /* the order here really doesn't matter */ 285 atomic_dec(&rq->rd->rto_count); 286 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask); 287 } 288 289 static void update_rt_migration(struct rt_rq *rt_rq) 290 { 291 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) { 292 if (!rt_rq->overloaded) { 293 rt_set_overload(rq_of_rt_rq(rt_rq)); 294 rt_rq->overloaded = 1; 295 } 296 } else if (rt_rq->overloaded) { 297 rt_clear_overload(rq_of_rt_rq(rt_rq)); 298 rt_rq->overloaded = 0; 299 } 300 } 301 302 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 303 { 304 struct task_struct *p; 305 306 if (!rt_entity_is_task(rt_se)) 307 return; 308 309 p = rt_task_of(rt_se); 310 rt_rq = &rq_of_rt_rq(rt_rq)->rt; 311 312 rt_rq->rt_nr_total++; 313 if (p->nr_cpus_allowed > 1) 314 rt_rq->rt_nr_migratory++; 315 316 update_rt_migration(rt_rq); 317 } 318 319 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 320 { 321 struct task_struct *p; 322 323 if (!rt_entity_is_task(rt_se)) 324 return; 325 326 p = rt_task_of(rt_se); 327 rt_rq = &rq_of_rt_rq(rt_rq)->rt; 328 329 rt_rq->rt_nr_total--; 330 if (p->nr_cpus_allowed > 1) 331 rt_rq->rt_nr_migratory--; 332 333 update_rt_migration(rt_rq); 334 } 335 336 static inline int has_pushable_tasks(struct rq *rq) 337 { 338 return !plist_head_empty(&rq->rt.pushable_tasks); 339 } 340 341 static inline void set_post_schedule(struct rq *rq) 342 { 343 /* 344 * We detect this state here so that we can avoid taking the RQ 345 * lock again later if there is no need to push 346 */ 347 rq->post_schedule = has_pushable_tasks(rq); 348 } 349 350 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p) 351 { 352 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); 353 plist_node_init(&p->pushable_tasks, p->prio); 354 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks); 355 356 /* Update the highest prio pushable task */ 357 if (p->prio < rq->rt.highest_prio.next) 358 rq->rt.highest_prio.next = p->prio; 359 } 360 361 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p) 362 { 363 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); 364 365 /* Update the new highest prio pushable task */ 366 if (has_pushable_tasks(rq)) { 367 p = plist_first_entry(&rq->rt.pushable_tasks, 368 struct task_struct, pushable_tasks); 369 rq->rt.highest_prio.next = p->prio; 370 } else 371 rq->rt.highest_prio.next = MAX_RT_PRIO; 372 } 373 374 #else 375 376 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p) 377 { 378 } 379 380 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p) 381 { 382 } 383 384 static inline 385 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 386 { 387 } 388 389 static inline 390 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 391 { 392 } 393 394 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) 395 { 396 return false; 397 } 398 399 static inline int pull_rt_task(struct rq *this_rq) 400 { 401 return 0; 402 } 403 404 static inline void set_post_schedule(struct rq *rq) 405 { 406 } 407 #endif /* CONFIG_SMP */ 408 409 static void enqueue_top_rt_rq(struct rt_rq *rt_rq); 410 static void dequeue_top_rt_rq(struct rt_rq *rt_rq); 411 412 static inline int on_rt_rq(struct sched_rt_entity *rt_se) 413 { 414 return !list_empty(&rt_se->run_list); 415 } 416 417 #ifdef CONFIG_RT_GROUP_SCHED 418 419 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) 420 { 421 if (!rt_rq->tg) 422 return RUNTIME_INF; 423 424 return rt_rq->rt_runtime; 425 } 426 427 static inline u64 sched_rt_period(struct rt_rq *rt_rq) 428 { 429 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); 430 } 431 432 typedef struct task_group *rt_rq_iter_t; 433 434 static inline struct task_group *next_task_group(struct task_group *tg) 435 { 436 do { 437 tg = list_entry_rcu(tg->list.next, 438 typeof(struct task_group), list); 439 } while (&tg->list != &task_groups && task_group_is_autogroup(tg)); 440 441 if (&tg->list == &task_groups) 442 tg = NULL; 443 444 return tg; 445 } 446 447 #define for_each_rt_rq(rt_rq, iter, rq) \ 448 for (iter = container_of(&task_groups, typeof(*iter), list); \ 449 (iter = next_task_group(iter)) && \ 450 (rt_rq = iter->rt_rq[cpu_of(rq)]);) 451 452 #define for_each_sched_rt_entity(rt_se) \ 453 for (; rt_se; rt_se = rt_se->parent) 454 455 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) 456 { 457 return rt_se->my_q; 458 } 459 460 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head); 461 static void dequeue_rt_entity(struct sched_rt_entity *rt_se); 462 463 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) 464 { 465 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; 466 struct rq *rq = rq_of_rt_rq(rt_rq); 467 struct sched_rt_entity *rt_se; 468 469 int cpu = cpu_of(rq); 470 471 rt_se = rt_rq->tg->rt_se[cpu]; 472 473 if (rt_rq->rt_nr_running) { 474 if (!rt_se) 475 enqueue_top_rt_rq(rt_rq); 476 else if (!on_rt_rq(rt_se)) 477 enqueue_rt_entity(rt_se, false); 478 479 if (rt_rq->highest_prio.curr < curr->prio) 480 resched_curr(rq); 481 } 482 } 483 484 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) 485 { 486 struct sched_rt_entity *rt_se; 487 int cpu = cpu_of(rq_of_rt_rq(rt_rq)); 488 489 rt_se = rt_rq->tg->rt_se[cpu]; 490 491 if (!rt_se) 492 dequeue_top_rt_rq(rt_rq); 493 else if (on_rt_rq(rt_se)) 494 dequeue_rt_entity(rt_se); 495 } 496 497 static inline int rt_rq_throttled(struct rt_rq *rt_rq) 498 { 499 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; 500 } 501 502 static int rt_se_boosted(struct sched_rt_entity *rt_se) 503 { 504 struct rt_rq *rt_rq = group_rt_rq(rt_se); 505 struct task_struct *p; 506 507 if (rt_rq) 508 return !!rt_rq->rt_nr_boosted; 509 510 p = rt_task_of(rt_se); 511 return p->prio != p->normal_prio; 512 } 513 514 #ifdef CONFIG_SMP 515 static inline const struct cpumask *sched_rt_period_mask(void) 516 { 517 return this_rq()->rd->span; 518 } 519 #else 520 static inline const struct cpumask *sched_rt_period_mask(void) 521 { 522 return cpu_online_mask; 523 } 524 #endif 525 526 static inline 527 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) 528 { 529 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; 530 } 531 532 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) 533 { 534 return &rt_rq->tg->rt_bandwidth; 535 } 536 537 #else /* !CONFIG_RT_GROUP_SCHED */ 538 539 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) 540 { 541 return rt_rq->rt_runtime; 542 } 543 544 static inline u64 sched_rt_period(struct rt_rq *rt_rq) 545 { 546 return ktime_to_ns(def_rt_bandwidth.rt_period); 547 } 548 549 typedef struct rt_rq *rt_rq_iter_t; 550 551 #define for_each_rt_rq(rt_rq, iter, rq) \ 552 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL) 553 554 #define for_each_sched_rt_entity(rt_se) \ 555 for (; rt_se; rt_se = NULL) 556 557 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) 558 { 559 return NULL; 560 } 561 562 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) 563 { 564 struct rq *rq = rq_of_rt_rq(rt_rq); 565 566 if (!rt_rq->rt_nr_running) 567 return; 568 569 enqueue_top_rt_rq(rt_rq); 570 resched_curr(rq); 571 } 572 573 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) 574 { 575 dequeue_top_rt_rq(rt_rq); 576 } 577 578 static inline int rt_rq_throttled(struct rt_rq *rt_rq) 579 { 580 return rt_rq->rt_throttled; 581 } 582 583 static inline const struct cpumask *sched_rt_period_mask(void) 584 { 585 return cpu_online_mask; 586 } 587 588 static inline 589 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) 590 { 591 return &cpu_rq(cpu)->rt; 592 } 593 594 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) 595 { 596 return &def_rt_bandwidth; 597 } 598 599 #endif /* CONFIG_RT_GROUP_SCHED */ 600 601 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq) 602 { 603 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 604 605 return (hrtimer_active(&rt_b->rt_period_timer) || 606 rt_rq->rt_time < rt_b->rt_runtime); 607 } 608 609 #ifdef CONFIG_SMP 610 /* 611 * We ran out of runtime, see if we can borrow some from our neighbours. 612 */ 613 static int do_balance_runtime(struct rt_rq *rt_rq) 614 { 615 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 616 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd; 617 int i, weight, more = 0; 618 u64 rt_period; 619 620 weight = cpumask_weight(rd->span); 621 622 raw_spin_lock(&rt_b->rt_runtime_lock); 623 rt_period = ktime_to_ns(rt_b->rt_period); 624 for_each_cpu(i, rd->span) { 625 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); 626 s64 diff; 627 628 if (iter == rt_rq) 629 continue; 630 631 raw_spin_lock(&iter->rt_runtime_lock); 632 /* 633 * Either all rqs have inf runtime and there's nothing to steal 634 * or __disable_runtime() below sets a specific rq to inf to 635 * indicate its been disabled and disalow stealing. 636 */ 637 if (iter->rt_runtime == RUNTIME_INF) 638 goto next; 639 640 /* 641 * From runqueues with spare time, take 1/n part of their 642 * spare time, but no more than our period. 643 */ 644 diff = iter->rt_runtime - iter->rt_time; 645 if (diff > 0) { 646 diff = div_u64((u64)diff, weight); 647 if (rt_rq->rt_runtime + diff > rt_period) 648 diff = rt_period - rt_rq->rt_runtime; 649 iter->rt_runtime -= diff; 650 rt_rq->rt_runtime += diff; 651 more = 1; 652 if (rt_rq->rt_runtime == rt_period) { 653 raw_spin_unlock(&iter->rt_runtime_lock); 654 break; 655 } 656 } 657 next: 658 raw_spin_unlock(&iter->rt_runtime_lock); 659 } 660 raw_spin_unlock(&rt_b->rt_runtime_lock); 661 662 return more; 663 } 664 665 /* 666 * Ensure this RQ takes back all the runtime it lend to its neighbours. 667 */ 668 static void __disable_runtime(struct rq *rq) 669 { 670 struct root_domain *rd = rq->rd; 671 rt_rq_iter_t iter; 672 struct rt_rq *rt_rq; 673 674 if (unlikely(!scheduler_running)) 675 return; 676 677 for_each_rt_rq(rt_rq, iter, rq) { 678 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 679 s64 want; 680 int i; 681 682 raw_spin_lock(&rt_b->rt_runtime_lock); 683 raw_spin_lock(&rt_rq->rt_runtime_lock); 684 /* 685 * Either we're all inf and nobody needs to borrow, or we're 686 * already disabled and thus have nothing to do, or we have 687 * exactly the right amount of runtime to take out. 688 */ 689 if (rt_rq->rt_runtime == RUNTIME_INF || 690 rt_rq->rt_runtime == rt_b->rt_runtime) 691 goto balanced; 692 raw_spin_unlock(&rt_rq->rt_runtime_lock); 693 694 /* 695 * Calculate the difference between what we started out with 696 * and what we current have, that's the amount of runtime 697 * we lend and now have to reclaim. 698 */ 699 want = rt_b->rt_runtime - rt_rq->rt_runtime; 700 701 /* 702 * Greedy reclaim, take back as much as we can. 703 */ 704 for_each_cpu(i, rd->span) { 705 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); 706 s64 diff; 707 708 /* 709 * Can't reclaim from ourselves or disabled runqueues. 710 */ 711 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF) 712 continue; 713 714 raw_spin_lock(&iter->rt_runtime_lock); 715 if (want > 0) { 716 diff = min_t(s64, iter->rt_runtime, want); 717 iter->rt_runtime -= diff; 718 want -= diff; 719 } else { 720 iter->rt_runtime -= want; 721 want -= want; 722 } 723 raw_spin_unlock(&iter->rt_runtime_lock); 724 725 if (!want) 726 break; 727 } 728 729 raw_spin_lock(&rt_rq->rt_runtime_lock); 730 /* 731 * We cannot be left wanting - that would mean some runtime 732 * leaked out of the system. 733 */ 734 BUG_ON(want); 735 balanced: 736 /* 737 * Disable all the borrow logic by pretending we have inf 738 * runtime - in which case borrowing doesn't make sense. 739 */ 740 rt_rq->rt_runtime = RUNTIME_INF; 741 rt_rq->rt_throttled = 0; 742 raw_spin_unlock(&rt_rq->rt_runtime_lock); 743 raw_spin_unlock(&rt_b->rt_runtime_lock); 744 745 /* Make rt_rq available for pick_next_task() */ 746 sched_rt_rq_enqueue(rt_rq); 747 } 748 } 749 750 static void __enable_runtime(struct rq *rq) 751 { 752 rt_rq_iter_t iter; 753 struct rt_rq *rt_rq; 754 755 if (unlikely(!scheduler_running)) 756 return; 757 758 /* 759 * Reset each runqueue's bandwidth settings 760 */ 761 for_each_rt_rq(rt_rq, iter, rq) { 762 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 763 764 raw_spin_lock(&rt_b->rt_runtime_lock); 765 raw_spin_lock(&rt_rq->rt_runtime_lock); 766 rt_rq->rt_runtime = rt_b->rt_runtime; 767 rt_rq->rt_time = 0; 768 rt_rq->rt_throttled = 0; 769 raw_spin_unlock(&rt_rq->rt_runtime_lock); 770 raw_spin_unlock(&rt_b->rt_runtime_lock); 771 } 772 } 773 774 static int balance_runtime(struct rt_rq *rt_rq) 775 { 776 int more = 0; 777 778 if (!sched_feat(RT_RUNTIME_SHARE)) 779 return more; 780 781 if (rt_rq->rt_time > rt_rq->rt_runtime) { 782 raw_spin_unlock(&rt_rq->rt_runtime_lock); 783 more = do_balance_runtime(rt_rq); 784 raw_spin_lock(&rt_rq->rt_runtime_lock); 785 } 786 787 return more; 788 } 789 #else /* !CONFIG_SMP */ 790 static inline int balance_runtime(struct rt_rq *rt_rq) 791 { 792 return 0; 793 } 794 #endif /* CONFIG_SMP */ 795 796 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) 797 { 798 int i, idle = 1, throttled = 0; 799 const struct cpumask *span; 800 801 span = sched_rt_period_mask(); 802 #ifdef CONFIG_RT_GROUP_SCHED 803 /* 804 * FIXME: isolated CPUs should really leave the root task group, 805 * whether they are isolcpus or were isolated via cpusets, lest 806 * the timer run on a CPU which does not service all runqueues, 807 * potentially leaving other CPUs indefinitely throttled. If 808 * isolation is really required, the user will turn the throttle 809 * off to kill the perturbations it causes anyway. Meanwhile, 810 * this maintains functionality for boot and/or troubleshooting. 811 */ 812 if (rt_b == &root_task_group.rt_bandwidth) 813 span = cpu_online_mask; 814 #endif 815 for_each_cpu(i, span) { 816 int enqueue = 0; 817 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); 818 struct rq *rq = rq_of_rt_rq(rt_rq); 819 820 raw_spin_lock(&rq->lock); 821 if (rt_rq->rt_time) { 822 u64 runtime; 823 824 raw_spin_lock(&rt_rq->rt_runtime_lock); 825 if (rt_rq->rt_throttled) 826 balance_runtime(rt_rq); 827 runtime = rt_rq->rt_runtime; 828 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); 829 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { 830 rt_rq->rt_throttled = 0; 831 enqueue = 1; 832 833 /* 834 * Force a clock update if the CPU was idle, 835 * lest wakeup -> unthrottle time accumulate. 836 */ 837 if (rt_rq->rt_nr_running && rq->curr == rq->idle) 838 rq->skip_clock_update = -1; 839 } 840 if (rt_rq->rt_time || rt_rq->rt_nr_running) 841 idle = 0; 842 raw_spin_unlock(&rt_rq->rt_runtime_lock); 843 } else if (rt_rq->rt_nr_running) { 844 idle = 0; 845 if (!rt_rq_throttled(rt_rq)) 846 enqueue = 1; 847 } 848 if (rt_rq->rt_throttled) 849 throttled = 1; 850 851 if (enqueue) 852 sched_rt_rq_enqueue(rt_rq); 853 raw_spin_unlock(&rq->lock); 854 } 855 856 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)) 857 return 1; 858 859 return idle; 860 } 861 862 static inline int rt_se_prio(struct sched_rt_entity *rt_se) 863 { 864 #ifdef CONFIG_RT_GROUP_SCHED 865 struct rt_rq *rt_rq = group_rt_rq(rt_se); 866 867 if (rt_rq) 868 return rt_rq->highest_prio.curr; 869 #endif 870 871 return rt_task_of(rt_se)->prio; 872 } 873 874 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) 875 { 876 u64 runtime = sched_rt_runtime(rt_rq); 877 878 if (rt_rq->rt_throttled) 879 return rt_rq_throttled(rt_rq); 880 881 if (runtime >= sched_rt_period(rt_rq)) 882 return 0; 883 884 balance_runtime(rt_rq); 885 runtime = sched_rt_runtime(rt_rq); 886 if (runtime == RUNTIME_INF) 887 return 0; 888 889 if (rt_rq->rt_time > runtime) { 890 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 891 892 /* 893 * Don't actually throttle groups that have no runtime assigned 894 * but accrue some time due to boosting. 895 */ 896 if (likely(rt_b->rt_runtime)) { 897 rt_rq->rt_throttled = 1; 898 printk_deferred_once("sched: RT throttling activated\n"); 899 } else { 900 /* 901 * In case we did anyway, make it go away, 902 * replenishment is a joke, since it will replenish us 903 * with exactly 0 ns. 904 */ 905 rt_rq->rt_time = 0; 906 } 907 908 if (rt_rq_throttled(rt_rq)) { 909 sched_rt_rq_dequeue(rt_rq); 910 return 1; 911 } 912 } 913 914 return 0; 915 } 916 917 /* 918 * Update the current task's runtime statistics. Skip current tasks that 919 * are not in our scheduling class. 920 */ 921 static void update_curr_rt(struct rq *rq) 922 { 923 struct task_struct *curr = rq->curr; 924 struct sched_rt_entity *rt_se = &curr->rt; 925 u64 delta_exec; 926 927 if (curr->sched_class != &rt_sched_class) 928 return; 929 930 delta_exec = rq_clock_task(rq) - curr->se.exec_start; 931 if (unlikely((s64)delta_exec <= 0)) 932 return; 933 934 schedstat_set(curr->se.statistics.exec_max, 935 max(curr->se.statistics.exec_max, delta_exec)); 936 937 curr->se.sum_exec_runtime += delta_exec; 938 account_group_exec_runtime(curr, delta_exec); 939 940 curr->se.exec_start = rq_clock_task(rq); 941 cpuacct_charge(curr, delta_exec); 942 943 sched_rt_avg_update(rq, delta_exec); 944 945 if (!rt_bandwidth_enabled()) 946 return; 947 948 for_each_sched_rt_entity(rt_se) { 949 struct rt_rq *rt_rq = rt_rq_of_se(rt_se); 950 951 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) { 952 raw_spin_lock(&rt_rq->rt_runtime_lock); 953 rt_rq->rt_time += delta_exec; 954 if (sched_rt_runtime_exceeded(rt_rq)) 955 resched_curr(rq); 956 raw_spin_unlock(&rt_rq->rt_runtime_lock); 957 } 958 } 959 } 960 961 static void 962 dequeue_top_rt_rq(struct rt_rq *rt_rq) 963 { 964 struct rq *rq = rq_of_rt_rq(rt_rq); 965 966 BUG_ON(&rq->rt != rt_rq); 967 968 if (!rt_rq->rt_queued) 969 return; 970 971 BUG_ON(!rq->nr_running); 972 973 sub_nr_running(rq, rt_rq->rt_nr_running); 974 rt_rq->rt_queued = 0; 975 } 976 977 static void 978 enqueue_top_rt_rq(struct rt_rq *rt_rq) 979 { 980 struct rq *rq = rq_of_rt_rq(rt_rq); 981 982 BUG_ON(&rq->rt != rt_rq); 983 984 if (rt_rq->rt_queued) 985 return; 986 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running) 987 return; 988 989 add_nr_running(rq, rt_rq->rt_nr_running); 990 rt_rq->rt_queued = 1; 991 } 992 993 #if defined CONFIG_SMP 994 995 static void 996 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) 997 { 998 struct rq *rq = rq_of_rt_rq(rt_rq); 999 1000 #ifdef CONFIG_RT_GROUP_SCHED 1001 /* 1002 * Change rq's cpupri only if rt_rq is the top queue. 1003 */ 1004 if (&rq->rt != rt_rq) 1005 return; 1006 #endif 1007 if (rq->online && prio < prev_prio) 1008 cpupri_set(&rq->rd->cpupri, rq->cpu, prio); 1009 } 1010 1011 static void 1012 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) 1013 { 1014 struct rq *rq = rq_of_rt_rq(rt_rq); 1015 1016 #ifdef CONFIG_RT_GROUP_SCHED 1017 /* 1018 * Change rq's cpupri only if rt_rq is the top queue. 1019 */ 1020 if (&rq->rt != rt_rq) 1021 return; 1022 #endif 1023 if (rq->online && rt_rq->highest_prio.curr != prev_prio) 1024 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr); 1025 } 1026 1027 #else /* CONFIG_SMP */ 1028 1029 static inline 1030 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} 1031 static inline 1032 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} 1033 1034 #endif /* CONFIG_SMP */ 1035 1036 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 1037 static void 1038 inc_rt_prio(struct rt_rq *rt_rq, int prio) 1039 { 1040 int prev_prio = rt_rq->highest_prio.curr; 1041 1042 if (prio < prev_prio) 1043 rt_rq->highest_prio.curr = prio; 1044 1045 inc_rt_prio_smp(rt_rq, prio, prev_prio); 1046 } 1047 1048 static void 1049 dec_rt_prio(struct rt_rq *rt_rq, int prio) 1050 { 1051 int prev_prio = rt_rq->highest_prio.curr; 1052 1053 if (rt_rq->rt_nr_running) { 1054 1055 WARN_ON(prio < prev_prio); 1056 1057 /* 1058 * This may have been our highest task, and therefore 1059 * we may have some recomputation to do 1060 */ 1061 if (prio == prev_prio) { 1062 struct rt_prio_array *array = &rt_rq->active; 1063 1064 rt_rq->highest_prio.curr = 1065 sched_find_first_bit(array->bitmap); 1066 } 1067 1068 } else 1069 rt_rq->highest_prio.curr = MAX_RT_PRIO; 1070 1071 dec_rt_prio_smp(rt_rq, prio, prev_prio); 1072 } 1073 1074 #else 1075 1076 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {} 1077 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {} 1078 1079 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */ 1080 1081 #ifdef CONFIG_RT_GROUP_SCHED 1082 1083 static void 1084 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1085 { 1086 if (rt_se_boosted(rt_se)) 1087 rt_rq->rt_nr_boosted++; 1088 1089 if (rt_rq->tg) 1090 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); 1091 } 1092 1093 static void 1094 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1095 { 1096 if (rt_se_boosted(rt_se)) 1097 rt_rq->rt_nr_boosted--; 1098 1099 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); 1100 } 1101 1102 #else /* CONFIG_RT_GROUP_SCHED */ 1103 1104 static void 1105 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1106 { 1107 start_rt_bandwidth(&def_rt_bandwidth); 1108 } 1109 1110 static inline 1111 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} 1112 1113 #endif /* CONFIG_RT_GROUP_SCHED */ 1114 1115 static inline 1116 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se) 1117 { 1118 struct rt_rq *group_rq = group_rt_rq(rt_se); 1119 1120 if (group_rq) 1121 return group_rq->rt_nr_running; 1122 else 1123 return 1; 1124 } 1125 1126 static inline 1127 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1128 { 1129 int prio = rt_se_prio(rt_se); 1130 1131 WARN_ON(!rt_prio(prio)); 1132 rt_rq->rt_nr_running += rt_se_nr_running(rt_se); 1133 1134 inc_rt_prio(rt_rq, prio); 1135 inc_rt_migration(rt_se, rt_rq); 1136 inc_rt_group(rt_se, rt_rq); 1137 } 1138 1139 static inline 1140 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1141 { 1142 WARN_ON(!rt_prio(rt_se_prio(rt_se))); 1143 WARN_ON(!rt_rq->rt_nr_running); 1144 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se); 1145 1146 dec_rt_prio(rt_rq, rt_se_prio(rt_se)); 1147 dec_rt_migration(rt_se, rt_rq); 1148 dec_rt_group(rt_se, rt_rq); 1149 } 1150 1151 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head) 1152 { 1153 struct rt_rq *rt_rq = rt_rq_of_se(rt_se); 1154 struct rt_prio_array *array = &rt_rq->active; 1155 struct rt_rq *group_rq = group_rt_rq(rt_se); 1156 struct list_head *queue = array->queue + rt_se_prio(rt_se); 1157 1158 /* 1159 * Don't enqueue the group if its throttled, or when empty. 1160 * The latter is a consequence of the former when a child group 1161 * get throttled and the current group doesn't have any other 1162 * active members. 1163 */ 1164 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) 1165 return; 1166 1167 if (head) 1168 list_add(&rt_se->run_list, queue); 1169 else 1170 list_add_tail(&rt_se->run_list, queue); 1171 __set_bit(rt_se_prio(rt_se), array->bitmap); 1172 1173 inc_rt_tasks(rt_se, rt_rq); 1174 } 1175 1176 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se) 1177 { 1178 struct rt_rq *rt_rq = rt_rq_of_se(rt_se); 1179 struct rt_prio_array *array = &rt_rq->active; 1180 1181 list_del_init(&rt_se->run_list); 1182 if (list_empty(array->queue + rt_se_prio(rt_se))) 1183 __clear_bit(rt_se_prio(rt_se), array->bitmap); 1184 1185 dec_rt_tasks(rt_se, rt_rq); 1186 } 1187 1188 /* 1189 * Because the prio of an upper entry depends on the lower 1190 * entries, we must remove entries top - down. 1191 */ 1192 static void dequeue_rt_stack(struct sched_rt_entity *rt_se) 1193 { 1194 struct sched_rt_entity *back = NULL; 1195 1196 for_each_sched_rt_entity(rt_se) { 1197 rt_se->back = back; 1198 back = rt_se; 1199 } 1200 1201 dequeue_top_rt_rq(rt_rq_of_se(back)); 1202 1203 for (rt_se = back; rt_se; rt_se = rt_se->back) { 1204 if (on_rt_rq(rt_se)) 1205 __dequeue_rt_entity(rt_se); 1206 } 1207 } 1208 1209 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head) 1210 { 1211 struct rq *rq = rq_of_rt_se(rt_se); 1212 1213 dequeue_rt_stack(rt_se); 1214 for_each_sched_rt_entity(rt_se) 1215 __enqueue_rt_entity(rt_se, head); 1216 enqueue_top_rt_rq(&rq->rt); 1217 } 1218 1219 static void dequeue_rt_entity(struct sched_rt_entity *rt_se) 1220 { 1221 struct rq *rq = rq_of_rt_se(rt_se); 1222 1223 dequeue_rt_stack(rt_se); 1224 1225 for_each_sched_rt_entity(rt_se) { 1226 struct rt_rq *rt_rq = group_rt_rq(rt_se); 1227 1228 if (rt_rq && rt_rq->rt_nr_running) 1229 __enqueue_rt_entity(rt_se, false); 1230 } 1231 enqueue_top_rt_rq(&rq->rt); 1232 } 1233 1234 /* 1235 * Adding/removing a task to/from a priority array: 1236 */ 1237 static void 1238 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags) 1239 { 1240 struct sched_rt_entity *rt_se = &p->rt; 1241 1242 if (flags & ENQUEUE_WAKEUP) 1243 rt_se->timeout = 0; 1244 1245 enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD); 1246 1247 if (!task_current(rq, p) && p->nr_cpus_allowed > 1) 1248 enqueue_pushable_task(rq, p); 1249 } 1250 1251 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags) 1252 { 1253 struct sched_rt_entity *rt_se = &p->rt; 1254 1255 update_curr_rt(rq); 1256 dequeue_rt_entity(rt_se); 1257 1258 dequeue_pushable_task(rq, p); 1259 } 1260 1261 /* 1262 * Put task to the head or the end of the run list without the overhead of 1263 * dequeue followed by enqueue. 1264 */ 1265 static void 1266 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head) 1267 { 1268 if (on_rt_rq(rt_se)) { 1269 struct rt_prio_array *array = &rt_rq->active; 1270 struct list_head *queue = array->queue + rt_se_prio(rt_se); 1271 1272 if (head) 1273 list_move(&rt_se->run_list, queue); 1274 else 1275 list_move_tail(&rt_se->run_list, queue); 1276 } 1277 } 1278 1279 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head) 1280 { 1281 struct sched_rt_entity *rt_se = &p->rt; 1282 struct rt_rq *rt_rq; 1283 1284 for_each_sched_rt_entity(rt_se) { 1285 rt_rq = rt_rq_of_se(rt_se); 1286 requeue_rt_entity(rt_rq, rt_se, head); 1287 } 1288 } 1289 1290 static void yield_task_rt(struct rq *rq) 1291 { 1292 requeue_task_rt(rq, rq->curr, 0); 1293 } 1294 1295 #ifdef CONFIG_SMP 1296 static int find_lowest_rq(struct task_struct *task); 1297 1298 static int 1299 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags) 1300 { 1301 struct task_struct *curr; 1302 struct rq *rq; 1303 1304 if (p->nr_cpus_allowed == 1) 1305 goto out; 1306 1307 /* For anything but wake ups, just return the task_cpu */ 1308 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK) 1309 goto out; 1310 1311 rq = cpu_rq(cpu); 1312 1313 rcu_read_lock(); 1314 curr = ACCESS_ONCE(rq->curr); /* unlocked access */ 1315 1316 /* 1317 * If the current task on @p's runqueue is an RT task, then 1318 * try to see if we can wake this RT task up on another 1319 * runqueue. Otherwise simply start this RT task 1320 * on its current runqueue. 1321 * 1322 * We want to avoid overloading runqueues. If the woken 1323 * task is a higher priority, then it will stay on this CPU 1324 * and the lower prio task should be moved to another CPU. 1325 * Even though this will probably make the lower prio task 1326 * lose its cache, we do not want to bounce a higher task 1327 * around just because it gave up its CPU, perhaps for a 1328 * lock? 1329 * 1330 * For equal prio tasks, we just let the scheduler sort it out. 1331 * 1332 * Otherwise, just let it ride on the affined RQ and the 1333 * post-schedule router will push the preempted task away 1334 * 1335 * This test is optimistic, if we get it wrong the load-balancer 1336 * will have to sort it out. 1337 */ 1338 if (curr && unlikely(rt_task(curr)) && 1339 (curr->nr_cpus_allowed < 2 || 1340 curr->prio <= p->prio)) { 1341 int target = find_lowest_rq(p); 1342 1343 if (target != -1) 1344 cpu = target; 1345 } 1346 rcu_read_unlock(); 1347 1348 out: 1349 return cpu; 1350 } 1351 1352 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p) 1353 { 1354 if (rq->curr->nr_cpus_allowed == 1) 1355 return; 1356 1357 if (p->nr_cpus_allowed != 1 1358 && cpupri_find(&rq->rd->cpupri, p, NULL)) 1359 return; 1360 1361 if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL)) 1362 return; 1363 1364 /* 1365 * There appears to be other cpus that can accept 1366 * current and none to run 'p', so lets reschedule 1367 * to try and push current away: 1368 */ 1369 requeue_task_rt(rq, p, 1); 1370 resched_curr(rq); 1371 } 1372 1373 #endif /* CONFIG_SMP */ 1374 1375 /* 1376 * Preempt the current task with a newly woken task if needed: 1377 */ 1378 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags) 1379 { 1380 if (p->prio < rq->curr->prio) { 1381 resched_curr(rq); 1382 return; 1383 } 1384 1385 #ifdef CONFIG_SMP 1386 /* 1387 * If: 1388 * 1389 * - the newly woken task is of equal priority to the current task 1390 * - the newly woken task is non-migratable while current is migratable 1391 * - current will be preempted on the next reschedule 1392 * 1393 * we should check to see if current can readily move to a different 1394 * cpu. If so, we will reschedule to allow the push logic to try 1395 * to move current somewhere else, making room for our non-migratable 1396 * task. 1397 */ 1398 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr)) 1399 check_preempt_equal_prio(rq, p); 1400 #endif 1401 } 1402 1403 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, 1404 struct rt_rq *rt_rq) 1405 { 1406 struct rt_prio_array *array = &rt_rq->active; 1407 struct sched_rt_entity *next = NULL; 1408 struct list_head *queue; 1409 int idx; 1410 1411 idx = sched_find_first_bit(array->bitmap); 1412 BUG_ON(idx >= MAX_RT_PRIO); 1413 1414 queue = array->queue + idx; 1415 next = list_entry(queue->next, struct sched_rt_entity, run_list); 1416 1417 return next; 1418 } 1419 1420 static struct task_struct *_pick_next_task_rt(struct rq *rq) 1421 { 1422 struct sched_rt_entity *rt_se; 1423 struct task_struct *p; 1424 struct rt_rq *rt_rq = &rq->rt; 1425 1426 do { 1427 rt_se = pick_next_rt_entity(rq, rt_rq); 1428 BUG_ON(!rt_se); 1429 rt_rq = group_rt_rq(rt_se); 1430 } while (rt_rq); 1431 1432 p = rt_task_of(rt_se); 1433 p->se.exec_start = rq_clock_task(rq); 1434 1435 return p; 1436 } 1437 1438 static struct task_struct * 1439 pick_next_task_rt(struct rq *rq, struct task_struct *prev) 1440 { 1441 struct task_struct *p; 1442 struct rt_rq *rt_rq = &rq->rt; 1443 1444 if (need_pull_rt_task(rq, prev)) { 1445 pull_rt_task(rq); 1446 /* 1447 * pull_rt_task() can drop (and re-acquire) rq->lock; this 1448 * means a dl or stop task can slip in, in which case we need 1449 * to re-start task selection. 1450 */ 1451 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) || 1452 rq->dl.dl_nr_running)) 1453 return RETRY_TASK; 1454 } 1455 1456 /* 1457 * We may dequeue prev's rt_rq in put_prev_task(). 1458 * So, we update time before rt_nr_running check. 1459 */ 1460 if (prev->sched_class == &rt_sched_class) 1461 update_curr_rt(rq); 1462 1463 if (!rt_rq->rt_queued) 1464 return NULL; 1465 1466 put_prev_task(rq, prev); 1467 1468 p = _pick_next_task_rt(rq); 1469 1470 /* The running task is never eligible for pushing */ 1471 dequeue_pushable_task(rq, p); 1472 1473 set_post_schedule(rq); 1474 1475 return p; 1476 } 1477 1478 static void put_prev_task_rt(struct rq *rq, struct task_struct *p) 1479 { 1480 update_curr_rt(rq); 1481 1482 /* 1483 * The previous task needs to be made eligible for pushing 1484 * if it is still active 1485 */ 1486 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1) 1487 enqueue_pushable_task(rq, p); 1488 } 1489 1490 #ifdef CONFIG_SMP 1491 1492 /* Only try algorithms three times */ 1493 #define RT_MAX_TRIES 3 1494 1495 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) 1496 { 1497 if (!task_running(rq, p) && 1498 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) 1499 return 1; 1500 return 0; 1501 } 1502 1503 /* 1504 * Return the highest pushable rq's task, which is suitable to be executed 1505 * on the cpu, NULL otherwise 1506 */ 1507 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu) 1508 { 1509 struct plist_head *head = &rq->rt.pushable_tasks; 1510 struct task_struct *p; 1511 1512 if (!has_pushable_tasks(rq)) 1513 return NULL; 1514 1515 plist_for_each_entry(p, head, pushable_tasks) { 1516 if (pick_rt_task(rq, p, cpu)) 1517 return p; 1518 } 1519 1520 return NULL; 1521 } 1522 1523 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask); 1524 1525 static int find_lowest_rq(struct task_struct *task) 1526 { 1527 struct sched_domain *sd; 1528 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask); 1529 int this_cpu = smp_processor_id(); 1530 int cpu = task_cpu(task); 1531 1532 /* Make sure the mask is initialized first */ 1533 if (unlikely(!lowest_mask)) 1534 return -1; 1535 1536 if (task->nr_cpus_allowed == 1) 1537 return -1; /* No other targets possible */ 1538 1539 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask)) 1540 return -1; /* No targets found */ 1541 1542 /* 1543 * At this point we have built a mask of cpus representing the 1544 * lowest priority tasks in the system. Now we want to elect 1545 * the best one based on our affinity and topology. 1546 * 1547 * We prioritize the last cpu that the task executed on since 1548 * it is most likely cache-hot in that location. 1549 */ 1550 if (cpumask_test_cpu(cpu, lowest_mask)) 1551 return cpu; 1552 1553 /* 1554 * Otherwise, we consult the sched_domains span maps to figure 1555 * out which cpu is logically closest to our hot cache data. 1556 */ 1557 if (!cpumask_test_cpu(this_cpu, lowest_mask)) 1558 this_cpu = -1; /* Skip this_cpu opt if not among lowest */ 1559 1560 rcu_read_lock(); 1561 for_each_domain(cpu, sd) { 1562 if (sd->flags & SD_WAKE_AFFINE) { 1563 int best_cpu; 1564 1565 /* 1566 * "this_cpu" is cheaper to preempt than a 1567 * remote processor. 1568 */ 1569 if (this_cpu != -1 && 1570 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { 1571 rcu_read_unlock(); 1572 return this_cpu; 1573 } 1574 1575 best_cpu = cpumask_first_and(lowest_mask, 1576 sched_domain_span(sd)); 1577 if (best_cpu < nr_cpu_ids) { 1578 rcu_read_unlock(); 1579 return best_cpu; 1580 } 1581 } 1582 } 1583 rcu_read_unlock(); 1584 1585 /* 1586 * And finally, if there were no matches within the domains 1587 * just give the caller *something* to work with from the compatible 1588 * locations. 1589 */ 1590 if (this_cpu != -1) 1591 return this_cpu; 1592 1593 cpu = cpumask_any(lowest_mask); 1594 if (cpu < nr_cpu_ids) 1595 return cpu; 1596 return -1; 1597 } 1598 1599 /* Will lock the rq it finds */ 1600 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) 1601 { 1602 struct rq *lowest_rq = NULL; 1603 int tries; 1604 int cpu; 1605 1606 for (tries = 0; tries < RT_MAX_TRIES; tries++) { 1607 cpu = find_lowest_rq(task); 1608 1609 if ((cpu == -1) || (cpu == rq->cpu)) 1610 break; 1611 1612 lowest_rq = cpu_rq(cpu); 1613 1614 /* if the prio of this runqueue changed, try again */ 1615 if (double_lock_balance(rq, lowest_rq)) { 1616 /* 1617 * We had to unlock the run queue. In 1618 * the mean time, task could have 1619 * migrated already or had its affinity changed. 1620 * Also make sure that it wasn't scheduled on its rq. 1621 */ 1622 if (unlikely(task_rq(task) != rq || 1623 !cpumask_test_cpu(lowest_rq->cpu, 1624 tsk_cpus_allowed(task)) || 1625 task_running(rq, task) || 1626 !task_on_rq_queued(task))) { 1627 1628 double_unlock_balance(rq, lowest_rq); 1629 lowest_rq = NULL; 1630 break; 1631 } 1632 } 1633 1634 /* If this rq is still suitable use it. */ 1635 if (lowest_rq->rt.highest_prio.curr > task->prio) 1636 break; 1637 1638 /* try again */ 1639 double_unlock_balance(rq, lowest_rq); 1640 lowest_rq = NULL; 1641 } 1642 1643 return lowest_rq; 1644 } 1645 1646 static struct task_struct *pick_next_pushable_task(struct rq *rq) 1647 { 1648 struct task_struct *p; 1649 1650 if (!has_pushable_tasks(rq)) 1651 return NULL; 1652 1653 p = plist_first_entry(&rq->rt.pushable_tasks, 1654 struct task_struct, pushable_tasks); 1655 1656 BUG_ON(rq->cpu != task_cpu(p)); 1657 BUG_ON(task_current(rq, p)); 1658 BUG_ON(p->nr_cpus_allowed <= 1); 1659 1660 BUG_ON(!task_on_rq_queued(p)); 1661 BUG_ON(!rt_task(p)); 1662 1663 return p; 1664 } 1665 1666 /* 1667 * If the current CPU has more than one RT task, see if the non 1668 * running task can migrate over to a CPU that is running a task 1669 * of lesser priority. 1670 */ 1671 static int push_rt_task(struct rq *rq) 1672 { 1673 struct task_struct *next_task; 1674 struct rq *lowest_rq; 1675 int ret = 0; 1676 1677 if (!rq->rt.overloaded) 1678 return 0; 1679 1680 next_task = pick_next_pushable_task(rq); 1681 if (!next_task) 1682 return 0; 1683 1684 retry: 1685 if (unlikely(next_task == rq->curr)) { 1686 WARN_ON(1); 1687 return 0; 1688 } 1689 1690 /* 1691 * It's possible that the next_task slipped in of 1692 * higher priority than current. If that's the case 1693 * just reschedule current. 1694 */ 1695 if (unlikely(next_task->prio < rq->curr->prio)) { 1696 resched_curr(rq); 1697 return 0; 1698 } 1699 1700 /* We might release rq lock */ 1701 get_task_struct(next_task); 1702 1703 /* find_lock_lowest_rq locks the rq if found */ 1704 lowest_rq = find_lock_lowest_rq(next_task, rq); 1705 if (!lowest_rq) { 1706 struct task_struct *task; 1707 /* 1708 * find_lock_lowest_rq releases rq->lock 1709 * so it is possible that next_task has migrated. 1710 * 1711 * We need to make sure that the task is still on the same 1712 * run-queue and is also still the next task eligible for 1713 * pushing. 1714 */ 1715 task = pick_next_pushable_task(rq); 1716 if (task_cpu(next_task) == rq->cpu && task == next_task) { 1717 /* 1718 * The task hasn't migrated, and is still the next 1719 * eligible task, but we failed to find a run-queue 1720 * to push it to. Do not retry in this case, since 1721 * other cpus will pull from us when ready. 1722 */ 1723 goto out; 1724 } 1725 1726 if (!task) 1727 /* No more tasks, just exit */ 1728 goto out; 1729 1730 /* 1731 * Something has shifted, try again. 1732 */ 1733 put_task_struct(next_task); 1734 next_task = task; 1735 goto retry; 1736 } 1737 1738 deactivate_task(rq, next_task, 0); 1739 set_task_cpu(next_task, lowest_rq->cpu); 1740 activate_task(lowest_rq, next_task, 0); 1741 ret = 1; 1742 1743 resched_curr(lowest_rq); 1744 1745 double_unlock_balance(rq, lowest_rq); 1746 1747 out: 1748 put_task_struct(next_task); 1749 1750 return ret; 1751 } 1752 1753 static void push_rt_tasks(struct rq *rq) 1754 { 1755 /* push_rt_task will return true if it moved an RT */ 1756 while (push_rt_task(rq)) 1757 ; 1758 } 1759 1760 static int pull_rt_task(struct rq *this_rq) 1761 { 1762 int this_cpu = this_rq->cpu, ret = 0, cpu; 1763 struct task_struct *p; 1764 struct rq *src_rq; 1765 1766 if (likely(!rt_overloaded(this_rq))) 1767 return 0; 1768 1769 /* 1770 * Match the barrier from rt_set_overloaded; this guarantees that if we 1771 * see overloaded we must also see the rto_mask bit. 1772 */ 1773 smp_rmb(); 1774 1775 for_each_cpu(cpu, this_rq->rd->rto_mask) { 1776 if (this_cpu == cpu) 1777 continue; 1778 1779 src_rq = cpu_rq(cpu); 1780 1781 /* 1782 * Don't bother taking the src_rq->lock if the next highest 1783 * task is known to be lower-priority than our current task. 1784 * This may look racy, but if this value is about to go 1785 * logically higher, the src_rq will push this task away. 1786 * And if its going logically lower, we do not care 1787 */ 1788 if (src_rq->rt.highest_prio.next >= 1789 this_rq->rt.highest_prio.curr) 1790 continue; 1791 1792 /* 1793 * We can potentially drop this_rq's lock in 1794 * double_lock_balance, and another CPU could 1795 * alter this_rq 1796 */ 1797 double_lock_balance(this_rq, src_rq); 1798 1799 /* 1800 * We can pull only a task, which is pushable 1801 * on its rq, and no others. 1802 */ 1803 p = pick_highest_pushable_task(src_rq, this_cpu); 1804 1805 /* 1806 * Do we have an RT task that preempts 1807 * the to-be-scheduled task? 1808 */ 1809 if (p && (p->prio < this_rq->rt.highest_prio.curr)) { 1810 WARN_ON(p == src_rq->curr); 1811 WARN_ON(!task_on_rq_queued(p)); 1812 1813 /* 1814 * There's a chance that p is higher in priority 1815 * than what's currently running on its cpu. 1816 * This is just that p is wakeing up and hasn't 1817 * had a chance to schedule. We only pull 1818 * p if it is lower in priority than the 1819 * current task on the run queue 1820 */ 1821 if (p->prio < src_rq->curr->prio) 1822 goto skip; 1823 1824 ret = 1; 1825 1826 deactivate_task(src_rq, p, 0); 1827 set_task_cpu(p, this_cpu); 1828 activate_task(this_rq, p, 0); 1829 /* 1830 * We continue with the search, just in 1831 * case there's an even higher prio task 1832 * in another runqueue. (low likelihood 1833 * but possible) 1834 */ 1835 } 1836 skip: 1837 double_unlock_balance(this_rq, src_rq); 1838 } 1839 1840 return ret; 1841 } 1842 1843 static void post_schedule_rt(struct rq *rq) 1844 { 1845 push_rt_tasks(rq); 1846 } 1847 1848 /* 1849 * If we are not running and we are not going to reschedule soon, we should 1850 * try to push tasks away now 1851 */ 1852 static void task_woken_rt(struct rq *rq, struct task_struct *p) 1853 { 1854 if (!task_running(rq, p) && 1855 !test_tsk_need_resched(rq->curr) && 1856 has_pushable_tasks(rq) && 1857 p->nr_cpus_allowed > 1 && 1858 (dl_task(rq->curr) || rt_task(rq->curr)) && 1859 (rq->curr->nr_cpus_allowed < 2 || 1860 rq->curr->prio <= p->prio)) 1861 push_rt_tasks(rq); 1862 } 1863 1864 static void set_cpus_allowed_rt(struct task_struct *p, 1865 const struct cpumask *new_mask) 1866 { 1867 struct rq *rq; 1868 int weight; 1869 1870 BUG_ON(!rt_task(p)); 1871 1872 if (!task_on_rq_queued(p)) 1873 return; 1874 1875 weight = cpumask_weight(new_mask); 1876 1877 /* 1878 * Only update if the process changes its state from whether it 1879 * can migrate or not. 1880 */ 1881 if ((p->nr_cpus_allowed > 1) == (weight > 1)) 1882 return; 1883 1884 rq = task_rq(p); 1885 1886 /* 1887 * The process used to be able to migrate OR it can now migrate 1888 */ 1889 if (weight <= 1) { 1890 if (!task_current(rq, p)) 1891 dequeue_pushable_task(rq, p); 1892 BUG_ON(!rq->rt.rt_nr_migratory); 1893 rq->rt.rt_nr_migratory--; 1894 } else { 1895 if (!task_current(rq, p)) 1896 enqueue_pushable_task(rq, p); 1897 rq->rt.rt_nr_migratory++; 1898 } 1899 1900 update_rt_migration(&rq->rt); 1901 } 1902 1903 /* Assumes rq->lock is held */ 1904 static void rq_online_rt(struct rq *rq) 1905 { 1906 if (rq->rt.overloaded) 1907 rt_set_overload(rq); 1908 1909 __enable_runtime(rq); 1910 1911 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr); 1912 } 1913 1914 /* Assumes rq->lock is held */ 1915 static void rq_offline_rt(struct rq *rq) 1916 { 1917 if (rq->rt.overloaded) 1918 rt_clear_overload(rq); 1919 1920 __disable_runtime(rq); 1921 1922 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID); 1923 } 1924 1925 /* 1926 * When switch from the rt queue, we bring ourselves to a position 1927 * that we might want to pull RT tasks from other runqueues. 1928 */ 1929 static void switched_from_rt(struct rq *rq, struct task_struct *p) 1930 { 1931 /* 1932 * If there are other RT tasks then we will reschedule 1933 * and the scheduling of the other RT tasks will handle 1934 * the balancing. But if we are the last RT task 1935 * we may need to handle the pulling of RT tasks 1936 * now. 1937 */ 1938 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running) 1939 return; 1940 1941 if (pull_rt_task(rq)) 1942 resched_curr(rq); 1943 } 1944 1945 void __init init_sched_rt_class(void) 1946 { 1947 unsigned int i; 1948 1949 for_each_possible_cpu(i) { 1950 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i), 1951 GFP_KERNEL, cpu_to_node(i)); 1952 } 1953 } 1954 #endif /* CONFIG_SMP */ 1955 1956 /* 1957 * When switching a task to RT, we may overload the runqueue 1958 * with RT tasks. In this case we try to push them off to 1959 * other runqueues. 1960 */ 1961 static void switched_to_rt(struct rq *rq, struct task_struct *p) 1962 { 1963 int check_resched = 1; 1964 1965 /* 1966 * If we are already running, then there's nothing 1967 * that needs to be done. But if we are not running 1968 * we may need to preempt the current running task. 1969 * If that current running task is also an RT task 1970 * then see if we can move to another run queue. 1971 */ 1972 if (task_on_rq_queued(p) && rq->curr != p) { 1973 #ifdef CONFIG_SMP 1974 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded && 1975 /* Don't resched if we changed runqueues */ 1976 push_rt_task(rq) && rq != task_rq(p)) 1977 check_resched = 0; 1978 #endif /* CONFIG_SMP */ 1979 if (check_resched && p->prio < rq->curr->prio) 1980 resched_curr(rq); 1981 } 1982 } 1983 1984 /* 1985 * Priority of the task has changed. This may cause 1986 * us to initiate a push or pull. 1987 */ 1988 static void 1989 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio) 1990 { 1991 if (!task_on_rq_queued(p)) 1992 return; 1993 1994 if (rq->curr == p) { 1995 #ifdef CONFIG_SMP 1996 /* 1997 * If our priority decreases while running, we 1998 * may need to pull tasks to this runqueue. 1999 */ 2000 if (oldprio < p->prio) 2001 pull_rt_task(rq); 2002 /* 2003 * If there's a higher priority task waiting to run 2004 * then reschedule. Note, the above pull_rt_task 2005 * can release the rq lock and p could migrate. 2006 * Only reschedule if p is still on the same runqueue. 2007 */ 2008 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p) 2009 resched_curr(rq); 2010 #else 2011 /* For UP simply resched on drop of prio */ 2012 if (oldprio < p->prio) 2013 resched_curr(rq); 2014 #endif /* CONFIG_SMP */ 2015 } else { 2016 /* 2017 * This task is not running, but if it is 2018 * greater than the current running task 2019 * then reschedule. 2020 */ 2021 if (p->prio < rq->curr->prio) 2022 resched_curr(rq); 2023 } 2024 } 2025 2026 static void watchdog(struct rq *rq, struct task_struct *p) 2027 { 2028 unsigned long soft, hard; 2029 2030 /* max may change after cur was read, this will be fixed next tick */ 2031 soft = task_rlimit(p, RLIMIT_RTTIME); 2032 hard = task_rlimit_max(p, RLIMIT_RTTIME); 2033 2034 if (soft != RLIM_INFINITY) { 2035 unsigned long next; 2036 2037 if (p->rt.watchdog_stamp != jiffies) { 2038 p->rt.timeout++; 2039 p->rt.watchdog_stamp = jiffies; 2040 } 2041 2042 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); 2043 if (p->rt.timeout > next) 2044 p->cputime_expires.sched_exp = p->se.sum_exec_runtime; 2045 } 2046 } 2047 2048 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) 2049 { 2050 struct sched_rt_entity *rt_se = &p->rt; 2051 2052 update_curr_rt(rq); 2053 2054 watchdog(rq, p); 2055 2056 /* 2057 * RR tasks need a special form of timeslice management. 2058 * FIFO tasks have no timeslices. 2059 */ 2060 if (p->policy != SCHED_RR) 2061 return; 2062 2063 if (--p->rt.time_slice) 2064 return; 2065 2066 p->rt.time_slice = sched_rr_timeslice; 2067 2068 /* 2069 * Requeue to the end of queue if we (and all of our ancestors) are not 2070 * the only element on the queue 2071 */ 2072 for_each_sched_rt_entity(rt_se) { 2073 if (rt_se->run_list.prev != rt_se->run_list.next) { 2074 requeue_task_rt(rq, p, 0); 2075 resched_curr(rq); 2076 return; 2077 } 2078 } 2079 } 2080 2081 static void set_curr_task_rt(struct rq *rq) 2082 { 2083 struct task_struct *p = rq->curr; 2084 2085 p->se.exec_start = rq_clock_task(rq); 2086 2087 /* The running task is never eligible for pushing */ 2088 dequeue_pushable_task(rq, p); 2089 } 2090 2091 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task) 2092 { 2093 /* 2094 * Time slice is 0 for SCHED_FIFO tasks 2095 */ 2096 if (task->policy == SCHED_RR) 2097 return sched_rr_timeslice; 2098 else 2099 return 0; 2100 } 2101 2102 const struct sched_class rt_sched_class = { 2103 .next = &fair_sched_class, 2104 .enqueue_task = enqueue_task_rt, 2105 .dequeue_task = dequeue_task_rt, 2106 .yield_task = yield_task_rt, 2107 2108 .check_preempt_curr = check_preempt_curr_rt, 2109 2110 .pick_next_task = pick_next_task_rt, 2111 .put_prev_task = put_prev_task_rt, 2112 2113 #ifdef CONFIG_SMP 2114 .select_task_rq = select_task_rq_rt, 2115 2116 .set_cpus_allowed = set_cpus_allowed_rt, 2117 .rq_online = rq_online_rt, 2118 .rq_offline = rq_offline_rt, 2119 .post_schedule = post_schedule_rt, 2120 .task_woken = task_woken_rt, 2121 .switched_from = switched_from_rt, 2122 #endif 2123 2124 .set_curr_task = set_curr_task_rt, 2125 .task_tick = task_tick_rt, 2126 2127 .get_rr_interval = get_rr_interval_rt, 2128 2129 .prio_changed = prio_changed_rt, 2130 .switched_to = switched_to_rt, 2131 }; 2132 2133 #ifdef CONFIG_SCHED_DEBUG 2134 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq); 2135 2136 void print_rt_stats(struct seq_file *m, int cpu) 2137 { 2138 rt_rq_iter_t iter; 2139 struct rt_rq *rt_rq; 2140 2141 rcu_read_lock(); 2142 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu)) 2143 print_rt_rq(m, cpu, rt_rq); 2144 rcu_read_unlock(); 2145 } 2146 #endif /* CONFIG_SCHED_DEBUG */ 2147