1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR 4 * policies) 5 */ 6 #include "sched.h" 7 8 int sched_rr_timeslice = RR_TIMESLICE; 9 int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE; 10 11 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); 12 13 struct rt_bandwidth def_rt_bandwidth; 14 15 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) 16 { 17 struct rt_bandwidth *rt_b = 18 container_of(timer, struct rt_bandwidth, rt_period_timer); 19 int idle = 0; 20 int overrun; 21 22 raw_spin_lock(&rt_b->rt_runtime_lock); 23 for (;;) { 24 overrun = hrtimer_forward_now(timer, rt_b->rt_period); 25 if (!overrun) 26 break; 27 28 raw_spin_unlock(&rt_b->rt_runtime_lock); 29 idle = do_sched_rt_period_timer(rt_b, overrun); 30 raw_spin_lock(&rt_b->rt_runtime_lock); 31 } 32 if (idle) 33 rt_b->rt_period_active = 0; 34 raw_spin_unlock(&rt_b->rt_runtime_lock); 35 36 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 37 } 38 39 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) 40 { 41 rt_b->rt_period = ns_to_ktime(period); 42 rt_b->rt_runtime = runtime; 43 44 raw_spin_lock_init(&rt_b->rt_runtime_lock); 45 46 hrtimer_init(&rt_b->rt_period_timer, 47 CLOCK_MONOTONIC, HRTIMER_MODE_REL); 48 rt_b->rt_period_timer.function = sched_rt_period_timer; 49 } 50 51 static void start_rt_bandwidth(struct rt_bandwidth *rt_b) 52 { 53 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF) 54 return; 55 56 raw_spin_lock(&rt_b->rt_runtime_lock); 57 if (!rt_b->rt_period_active) { 58 rt_b->rt_period_active = 1; 59 /* 60 * SCHED_DEADLINE updates the bandwidth, as a run away 61 * RT task with a DL task could hog a CPU. But DL does 62 * not reset the period. If a deadline task was running 63 * without an RT task running, it can cause RT tasks to 64 * throttle when they start up. Kick the timer right away 65 * to update the period. 66 */ 67 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0)); 68 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED); 69 } 70 raw_spin_unlock(&rt_b->rt_runtime_lock); 71 } 72 73 void init_rt_rq(struct rt_rq *rt_rq) 74 { 75 struct rt_prio_array *array; 76 int i; 77 78 array = &rt_rq->active; 79 for (i = 0; i < MAX_RT_PRIO; i++) { 80 INIT_LIST_HEAD(array->queue + i); 81 __clear_bit(i, array->bitmap); 82 } 83 /* delimiter for bitsearch: */ 84 __set_bit(MAX_RT_PRIO, array->bitmap); 85 86 #if defined CONFIG_SMP 87 rt_rq->highest_prio.curr = MAX_RT_PRIO; 88 rt_rq->highest_prio.next = MAX_RT_PRIO; 89 rt_rq->rt_nr_migratory = 0; 90 rt_rq->overloaded = 0; 91 plist_head_init(&rt_rq->pushable_tasks); 92 #endif /* CONFIG_SMP */ 93 /* We start is dequeued state, because no RT tasks are queued */ 94 rt_rq->rt_queued = 0; 95 96 rt_rq->rt_time = 0; 97 rt_rq->rt_throttled = 0; 98 rt_rq->rt_runtime = 0; 99 raw_spin_lock_init(&rt_rq->rt_runtime_lock); 100 } 101 102 #ifdef CONFIG_RT_GROUP_SCHED 103 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) 104 { 105 hrtimer_cancel(&rt_b->rt_period_timer); 106 } 107 108 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q) 109 110 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) 111 { 112 #ifdef CONFIG_SCHED_DEBUG 113 WARN_ON_ONCE(!rt_entity_is_task(rt_se)); 114 #endif 115 return container_of(rt_se, struct task_struct, rt); 116 } 117 118 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) 119 { 120 return rt_rq->rq; 121 } 122 123 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) 124 { 125 return rt_se->rt_rq; 126 } 127 128 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) 129 { 130 struct rt_rq *rt_rq = rt_se->rt_rq; 131 132 return rt_rq->rq; 133 } 134 135 void free_rt_sched_group(struct task_group *tg) 136 { 137 int i; 138 139 if (tg->rt_se) 140 destroy_rt_bandwidth(&tg->rt_bandwidth); 141 142 for_each_possible_cpu(i) { 143 if (tg->rt_rq) 144 kfree(tg->rt_rq[i]); 145 if (tg->rt_se) 146 kfree(tg->rt_se[i]); 147 } 148 149 kfree(tg->rt_rq); 150 kfree(tg->rt_se); 151 } 152 153 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 154 struct sched_rt_entity *rt_se, int cpu, 155 struct sched_rt_entity *parent) 156 { 157 struct rq *rq = cpu_rq(cpu); 158 159 rt_rq->highest_prio.curr = MAX_RT_PRIO; 160 rt_rq->rt_nr_boosted = 0; 161 rt_rq->rq = rq; 162 rt_rq->tg = tg; 163 164 tg->rt_rq[cpu] = rt_rq; 165 tg->rt_se[cpu] = rt_se; 166 167 if (!rt_se) 168 return; 169 170 if (!parent) 171 rt_se->rt_rq = &rq->rt; 172 else 173 rt_se->rt_rq = parent->my_q; 174 175 rt_se->my_q = rt_rq; 176 rt_se->parent = parent; 177 INIT_LIST_HEAD(&rt_se->run_list); 178 } 179 180 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) 181 { 182 struct rt_rq *rt_rq; 183 struct sched_rt_entity *rt_se; 184 int i; 185 186 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); 187 if (!tg->rt_rq) 188 goto err; 189 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); 190 if (!tg->rt_se) 191 goto err; 192 193 init_rt_bandwidth(&tg->rt_bandwidth, 194 ktime_to_ns(def_rt_bandwidth.rt_period), 0); 195 196 for_each_possible_cpu(i) { 197 rt_rq = kzalloc_node(sizeof(struct rt_rq), 198 GFP_KERNEL, cpu_to_node(i)); 199 if (!rt_rq) 200 goto err; 201 202 rt_se = kzalloc_node(sizeof(struct sched_rt_entity), 203 GFP_KERNEL, cpu_to_node(i)); 204 if (!rt_se) 205 goto err_free_rq; 206 207 init_rt_rq(rt_rq); 208 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; 209 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]); 210 } 211 212 return 1; 213 214 err_free_rq: 215 kfree(rt_rq); 216 err: 217 return 0; 218 } 219 220 #else /* CONFIG_RT_GROUP_SCHED */ 221 222 #define rt_entity_is_task(rt_se) (1) 223 224 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se) 225 { 226 return container_of(rt_se, struct task_struct, rt); 227 } 228 229 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq) 230 { 231 return container_of(rt_rq, struct rq, rt); 232 } 233 234 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se) 235 { 236 struct task_struct *p = rt_task_of(rt_se); 237 238 return task_rq(p); 239 } 240 241 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se) 242 { 243 struct rq *rq = rq_of_rt_se(rt_se); 244 245 return &rq->rt; 246 } 247 248 void free_rt_sched_group(struct task_group *tg) { } 249 250 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) 251 { 252 return 1; 253 } 254 #endif /* CONFIG_RT_GROUP_SCHED */ 255 256 #ifdef CONFIG_SMP 257 258 static void pull_rt_task(struct rq *this_rq); 259 260 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) 261 { 262 /* Try to pull RT tasks here if we lower this rq's prio */ 263 return rq->rt.highest_prio.curr > prev->prio; 264 } 265 266 static inline int rt_overloaded(struct rq *rq) 267 { 268 return atomic_read(&rq->rd->rto_count); 269 } 270 271 static inline void rt_set_overload(struct rq *rq) 272 { 273 if (!rq->online) 274 return; 275 276 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask); 277 /* 278 * Make sure the mask is visible before we set 279 * the overload count. That is checked to determine 280 * if we should look at the mask. It would be a shame 281 * if we looked at the mask, but the mask was not 282 * updated yet. 283 * 284 * Matched by the barrier in pull_rt_task(). 285 */ 286 smp_wmb(); 287 atomic_inc(&rq->rd->rto_count); 288 } 289 290 static inline void rt_clear_overload(struct rq *rq) 291 { 292 if (!rq->online) 293 return; 294 295 /* the order here really doesn't matter */ 296 atomic_dec(&rq->rd->rto_count); 297 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask); 298 } 299 300 static void update_rt_migration(struct rt_rq *rt_rq) 301 { 302 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) { 303 if (!rt_rq->overloaded) { 304 rt_set_overload(rq_of_rt_rq(rt_rq)); 305 rt_rq->overloaded = 1; 306 } 307 } else if (rt_rq->overloaded) { 308 rt_clear_overload(rq_of_rt_rq(rt_rq)); 309 rt_rq->overloaded = 0; 310 } 311 } 312 313 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 314 { 315 struct task_struct *p; 316 317 if (!rt_entity_is_task(rt_se)) 318 return; 319 320 p = rt_task_of(rt_se); 321 rt_rq = &rq_of_rt_rq(rt_rq)->rt; 322 323 rt_rq->rt_nr_total++; 324 if (p->nr_cpus_allowed > 1) 325 rt_rq->rt_nr_migratory++; 326 327 update_rt_migration(rt_rq); 328 } 329 330 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 331 { 332 struct task_struct *p; 333 334 if (!rt_entity_is_task(rt_se)) 335 return; 336 337 p = rt_task_of(rt_se); 338 rt_rq = &rq_of_rt_rq(rt_rq)->rt; 339 340 rt_rq->rt_nr_total--; 341 if (p->nr_cpus_allowed > 1) 342 rt_rq->rt_nr_migratory--; 343 344 update_rt_migration(rt_rq); 345 } 346 347 static inline int has_pushable_tasks(struct rq *rq) 348 { 349 return !plist_head_empty(&rq->rt.pushable_tasks); 350 } 351 352 static DEFINE_PER_CPU(struct callback_head, rt_push_head); 353 static DEFINE_PER_CPU(struct callback_head, rt_pull_head); 354 355 static void push_rt_tasks(struct rq *); 356 static void pull_rt_task(struct rq *); 357 358 static inline void rt_queue_push_tasks(struct rq *rq) 359 { 360 if (!has_pushable_tasks(rq)) 361 return; 362 363 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks); 364 } 365 366 static inline void rt_queue_pull_task(struct rq *rq) 367 { 368 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task); 369 } 370 371 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p) 372 { 373 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); 374 plist_node_init(&p->pushable_tasks, p->prio); 375 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks); 376 377 /* Update the highest prio pushable task */ 378 if (p->prio < rq->rt.highest_prio.next) 379 rq->rt.highest_prio.next = p->prio; 380 } 381 382 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p) 383 { 384 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks); 385 386 /* Update the new highest prio pushable task */ 387 if (has_pushable_tasks(rq)) { 388 p = plist_first_entry(&rq->rt.pushable_tasks, 389 struct task_struct, pushable_tasks); 390 rq->rt.highest_prio.next = p->prio; 391 } else 392 rq->rt.highest_prio.next = MAX_RT_PRIO; 393 } 394 395 #else 396 397 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p) 398 { 399 } 400 401 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p) 402 { 403 } 404 405 static inline 406 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 407 { 408 } 409 410 static inline 411 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 412 { 413 } 414 415 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev) 416 { 417 return false; 418 } 419 420 static inline void pull_rt_task(struct rq *this_rq) 421 { 422 } 423 424 static inline void rt_queue_push_tasks(struct rq *rq) 425 { 426 } 427 #endif /* CONFIG_SMP */ 428 429 static void enqueue_top_rt_rq(struct rt_rq *rt_rq); 430 static void dequeue_top_rt_rq(struct rt_rq *rt_rq); 431 432 static inline int on_rt_rq(struct sched_rt_entity *rt_se) 433 { 434 return rt_se->on_rq; 435 } 436 437 #ifdef CONFIG_RT_GROUP_SCHED 438 439 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) 440 { 441 if (!rt_rq->tg) 442 return RUNTIME_INF; 443 444 return rt_rq->rt_runtime; 445 } 446 447 static inline u64 sched_rt_period(struct rt_rq *rt_rq) 448 { 449 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period); 450 } 451 452 typedef struct task_group *rt_rq_iter_t; 453 454 static inline struct task_group *next_task_group(struct task_group *tg) 455 { 456 do { 457 tg = list_entry_rcu(tg->list.next, 458 typeof(struct task_group), list); 459 } while (&tg->list != &task_groups && task_group_is_autogroup(tg)); 460 461 if (&tg->list == &task_groups) 462 tg = NULL; 463 464 return tg; 465 } 466 467 #define for_each_rt_rq(rt_rq, iter, rq) \ 468 for (iter = container_of(&task_groups, typeof(*iter), list); \ 469 (iter = next_task_group(iter)) && \ 470 (rt_rq = iter->rt_rq[cpu_of(rq)]);) 471 472 #define for_each_sched_rt_entity(rt_se) \ 473 for (; rt_se; rt_se = rt_se->parent) 474 475 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) 476 { 477 return rt_se->my_q; 478 } 479 480 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); 481 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags); 482 483 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq) 484 { 485 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr; 486 struct rq *rq = rq_of_rt_rq(rt_rq); 487 struct sched_rt_entity *rt_se; 488 489 int cpu = cpu_of(rq); 490 491 rt_se = rt_rq->tg->rt_se[cpu]; 492 493 if (rt_rq->rt_nr_running) { 494 if (!rt_se) 495 enqueue_top_rt_rq(rt_rq); 496 else if (!on_rt_rq(rt_se)) 497 enqueue_rt_entity(rt_se, 0); 498 499 if (rt_rq->highest_prio.curr < curr->prio) 500 resched_curr(rq); 501 } 502 } 503 504 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq) 505 { 506 struct sched_rt_entity *rt_se; 507 int cpu = cpu_of(rq_of_rt_rq(rt_rq)); 508 509 rt_se = rt_rq->tg->rt_se[cpu]; 510 511 if (!rt_se) 512 dequeue_top_rt_rq(rt_rq); 513 else if (on_rt_rq(rt_se)) 514 dequeue_rt_entity(rt_se, 0); 515 } 516 517 static inline int rt_rq_throttled(struct rt_rq *rt_rq) 518 { 519 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted; 520 } 521 522 static int rt_se_boosted(struct sched_rt_entity *rt_se) 523 { 524 struct rt_rq *rt_rq = group_rt_rq(rt_se); 525 struct task_struct *p; 526 527 if (rt_rq) 528 return !!rt_rq->rt_nr_boosted; 529 530 p = rt_task_of(rt_se); 531 return p->prio != p->normal_prio; 532 } 533 534 #ifdef CONFIG_SMP 535 static inline const struct cpumask *sched_rt_period_mask(void) 536 { 537 return this_rq()->rd->span; 538 } 539 #else 540 static inline const struct cpumask *sched_rt_period_mask(void) 541 { 542 return cpu_online_mask; 543 } 544 #endif 545 546 static inline 547 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) 548 { 549 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu]; 550 } 551 552 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) 553 { 554 return &rt_rq->tg->rt_bandwidth; 555 } 556 557 #else /* !CONFIG_RT_GROUP_SCHED */ 558 559 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq) 560 { 561 return rt_rq->rt_runtime; 562 } 563 564 static inline u64 sched_rt_period(struct rt_rq *rt_rq) 565 { 566 return ktime_to_ns(def_rt_bandwidth.rt_period); 567 } 568 569 typedef struct rt_rq *rt_rq_iter_t; 570 571 #define for_each_rt_rq(rt_rq, iter, rq) \ 572 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL) 573 574 #define for_each_sched_rt_entity(rt_se) \ 575 for (; rt_se; rt_se = NULL) 576 577 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se) 578 { 579 return NULL; 580 } 581 582 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq) 583 { 584 struct rq *rq = rq_of_rt_rq(rt_rq); 585 586 if (!rt_rq->rt_nr_running) 587 return; 588 589 enqueue_top_rt_rq(rt_rq); 590 resched_curr(rq); 591 } 592 593 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq) 594 { 595 dequeue_top_rt_rq(rt_rq); 596 } 597 598 static inline int rt_rq_throttled(struct rt_rq *rt_rq) 599 { 600 return rt_rq->rt_throttled; 601 } 602 603 static inline const struct cpumask *sched_rt_period_mask(void) 604 { 605 return cpu_online_mask; 606 } 607 608 static inline 609 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu) 610 { 611 return &cpu_rq(cpu)->rt; 612 } 613 614 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq) 615 { 616 return &def_rt_bandwidth; 617 } 618 619 #endif /* CONFIG_RT_GROUP_SCHED */ 620 621 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq) 622 { 623 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 624 625 return (hrtimer_active(&rt_b->rt_period_timer) || 626 rt_rq->rt_time < rt_b->rt_runtime); 627 } 628 629 #ifdef CONFIG_SMP 630 /* 631 * We ran out of runtime, see if we can borrow some from our neighbours. 632 */ 633 static void do_balance_runtime(struct rt_rq *rt_rq) 634 { 635 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 636 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd; 637 int i, weight; 638 u64 rt_period; 639 640 weight = cpumask_weight(rd->span); 641 642 raw_spin_lock(&rt_b->rt_runtime_lock); 643 rt_period = ktime_to_ns(rt_b->rt_period); 644 for_each_cpu(i, rd->span) { 645 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); 646 s64 diff; 647 648 if (iter == rt_rq) 649 continue; 650 651 raw_spin_lock(&iter->rt_runtime_lock); 652 /* 653 * Either all rqs have inf runtime and there's nothing to steal 654 * or __disable_runtime() below sets a specific rq to inf to 655 * indicate its been disabled and disalow stealing. 656 */ 657 if (iter->rt_runtime == RUNTIME_INF) 658 goto next; 659 660 /* 661 * From runqueues with spare time, take 1/n part of their 662 * spare time, but no more than our period. 663 */ 664 diff = iter->rt_runtime - iter->rt_time; 665 if (diff > 0) { 666 diff = div_u64((u64)diff, weight); 667 if (rt_rq->rt_runtime + diff > rt_period) 668 diff = rt_period - rt_rq->rt_runtime; 669 iter->rt_runtime -= diff; 670 rt_rq->rt_runtime += diff; 671 if (rt_rq->rt_runtime == rt_period) { 672 raw_spin_unlock(&iter->rt_runtime_lock); 673 break; 674 } 675 } 676 next: 677 raw_spin_unlock(&iter->rt_runtime_lock); 678 } 679 raw_spin_unlock(&rt_b->rt_runtime_lock); 680 } 681 682 /* 683 * Ensure this RQ takes back all the runtime it lend to its neighbours. 684 */ 685 static void __disable_runtime(struct rq *rq) 686 { 687 struct root_domain *rd = rq->rd; 688 rt_rq_iter_t iter; 689 struct rt_rq *rt_rq; 690 691 if (unlikely(!scheduler_running)) 692 return; 693 694 for_each_rt_rq(rt_rq, iter, rq) { 695 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 696 s64 want; 697 int i; 698 699 raw_spin_lock(&rt_b->rt_runtime_lock); 700 raw_spin_lock(&rt_rq->rt_runtime_lock); 701 /* 702 * Either we're all inf and nobody needs to borrow, or we're 703 * already disabled and thus have nothing to do, or we have 704 * exactly the right amount of runtime to take out. 705 */ 706 if (rt_rq->rt_runtime == RUNTIME_INF || 707 rt_rq->rt_runtime == rt_b->rt_runtime) 708 goto balanced; 709 raw_spin_unlock(&rt_rq->rt_runtime_lock); 710 711 /* 712 * Calculate the difference between what we started out with 713 * and what we current have, that's the amount of runtime 714 * we lend and now have to reclaim. 715 */ 716 want = rt_b->rt_runtime - rt_rq->rt_runtime; 717 718 /* 719 * Greedy reclaim, take back as much as we can. 720 */ 721 for_each_cpu(i, rd->span) { 722 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i); 723 s64 diff; 724 725 /* 726 * Can't reclaim from ourselves or disabled runqueues. 727 */ 728 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF) 729 continue; 730 731 raw_spin_lock(&iter->rt_runtime_lock); 732 if (want > 0) { 733 diff = min_t(s64, iter->rt_runtime, want); 734 iter->rt_runtime -= diff; 735 want -= diff; 736 } else { 737 iter->rt_runtime -= want; 738 want -= want; 739 } 740 raw_spin_unlock(&iter->rt_runtime_lock); 741 742 if (!want) 743 break; 744 } 745 746 raw_spin_lock(&rt_rq->rt_runtime_lock); 747 /* 748 * We cannot be left wanting - that would mean some runtime 749 * leaked out of the system. 750 */ 751 BUG_ON(want); 752 balanced: 753 /* 754 * Disable all the borrow logic by pretending we have inf 755 * runtime - in which case borrowing doesn't make sense. 756 */ 757 rt_rq->rt_runtime = RUNTIME_INF; 758 rt_rq->rt_throttled = 0; 759 raw_spin_unlock(&rt_rq->rt_runtime_lock); 760 raw_spin_unlock(&rt_b->rt_runtime_lock); 761 762 /* Make rt_rq available for pick_next_task() */ 763 sched_rt_rq_enqueue(rt_rq); 764 } 765 } 766 767 static void __enable_runtime(struct rq *rq) 768 { 769 rt_rq_iter_t iter; 770 struct rt_rq *rt_rq; 771 772 if (unlikely(!scheduler_running)) 773 return; 774 775 /* 776 * Reset each runqueue's bandwidth settings 777 */ 778 for_each_rt_rq(rt_rq, iter, rq) { 779 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 780 781 raw_spin_lock(&rt_b->rt_runtime_lock); 782 raw_spin_lock(&rt_rq->rt_runtime_lock); 783 rt_rq->rt_runtime = rt_b->rt_runtime; 784 rt_rq->rt_time = 0; 785 rt_rq->rt_throttled = 0; 786 raw_spin_unlock(&rt_rq->rt_runtime_lock); 787 raw_spin_unlock(&rt_b->rt_runtime_lock); 788 } 789 } 790 791 static void balance_runtime(struct rt_rq *rt_rq) 792 { 793 if (!sched_feat(RT_RUNTIME_SHARE)) 794 return; 795 796 if (rt_rq->rt_time > rt_rq->rt_runtime) { 797 raw_spin_unlock(&rt_rq->rt_runtime_lock); 798 do_balance_runtime(rt_rq); 799 raw_spin_lock(&rt_rq->rt_runtime_lock); 800 } 801 } 802 #else /* !CONFIG_SMP */ 803 static inline void balance_runtime(struct rt_rq *rt_rq) {} 804 #endif /* CONFIG_SMP */ 805 806 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun) 807 { 808 int i, idle = 1, throttled = 0; 809 const struct cpumask *span; 810 811 span = sched_rt_period_mask(); 812 #ifdef CONFIG_RT_GROUP_SCHED 813 /* 814 * FIXME: isolated CPUs should really leave the root task group, 815 * whether they are isolcpus or were isolated via cpusets, lest 816 * the timer run on a CPU which does not service all runqueues, 817 * potentially leaving other CPUs indefinitely throttled. If 818 * isolation is really required, the user will turn the throttle 819 * off to kill the perturbations it causes anyway. Meanwhile, 820 * this maintains functionality for boot and/or troubleshooting. 821 */ 822 if (rt_b == &root_task_group.rt_bandwidth) 823 span = cpu_online_mask; 824 #endif 825 for_each_cpu(i, span) { 826 int enqueue = 0; 827 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i); 828 struct rq *rq = rq_of_rt_rq(rt_rq); 829 int skip; 830 831 /* 832 * When span == cpu_online_mask, taking each rq->lock 833 * can be time-consuming. Try to avoid it when possible. 834 */ 835 raw_spin_lock(&rt_rq->rt_runtime_lock); 836 skip = !rt_rq->rt_time && !rt_rq->rt_nr_running; 837 raw_spin_unlock(&rt_rq->rt_runtime_lock); 838 if (skip) 839 continue; 840 841 raw_spin_lock(&rq->lock); 842 update_rq_clock(rq); 843 844 if (rt_rq->rt_time) { 845 u64 runtime; 846 847 raw_spin_lock(&rt_rq->rt_runtime_lock); 848 if (rt_rq->rt_throttled) 849 balance_runtime(rt_rq); 850 runtime = rt_rq->rt_runtime; 851 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime); 852 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) { 853 rt_rq->rt_throttled = 0; 854 enqueue = 1; 855 856 /* 857 * When we're idle and a woken (rt) task is 858 * throttled check_preempt_curr() will set 859 * skip_update and the time between the wakeup 860 * and this unthrottle will get accounted as 861 * 'runtime'. 862 */ 863 if (rt_rq->rt_nr_running && rq->curr == rq->idle) 864 rq_clock_cancel_skipupdate(rq); 865 } 866 if (rt_rq->rt_time || rt_rq->rt_nr_running) 867 idle = 0; 868 raw_spin_unlock(&rt_rq->rt_runtime_lock); 869 } else if (rt_rq->rt_nr_running) { 870 idle = 0; 871 if (!rt_rq_throttled(rt_rq)) 872 enqueue = 1; 873 } 874 if (rt_rq->rt_throttled) 875 throttled = 1; 876 877 if (enqueue) 878 sched_rt_rq_enqueue(rt_rq); 879 raw_spin_unlock(&rq->lock); 880 } 881 882 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)) 883 return 1; 884 885 return idle; 886 } 887 888 static inline int rt_se_prio(struct sched_rt_entity *rt_se) 889 { 890 #ifdef CONFIG_RT_GROUP_SCHED 891 struct rt_rq *rt_rq = group_rt_rq(rt_se); 892 893 if (rt_rq) 894 return rt_rq->highest_prio.curr; 895 #endif 896 897 return rt_task_of(rt_se)->prio; 898 } 899 900 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq) 901 { 902 u64 runtime = sched_rt_runtime(rt_rq); 903 904 if (rt_rq->rt_throttled) 905 return rt_rq_throttled(rt_rq); 906 907 if (runtime >= sched_rt_period(rt_rq)) 908 return 0; 909 910 balance_runtime(rt_rq); 911 runtime = sched_rt_runtime(rt_rq); 912 if (runtime == RUNTIME_INF) 913 return 0; 914 915 if (rt_rq->rt_time > runtime) { 916 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq); 917 918 /* 919 * Don't actually throttle groups that have no runtime assigned 920 * but accrue some time due to boosting. 921 */ 922 if (likely(rt_b->rt_runtime)) { 923 rt_rq->rt_throttled = 1; 924 printk_deferred_once("sched: RT throttling activated\n"); 925 } else { 926 /* 927 * In case we did anyway, make it go away, 928 * replenishment is a joke, since it will replenish us 929 * with exactly 0 ns. 930 */ 931 rt_rq->rt_time = 0; 932 } 933 934 if (rt_rq_throttled(rt_rq)) { 935 sched_rt_rq_dequeue(rt_rq); 936 return 1; 937 } 938 } 939 940 return 0; 941 } 942 943 /* 944 * Update the current task's runtime statistics. Skip current tasks that 945 * are not in our scheduling class. 946 */ 947 static void update_curr_rt(struct rq *rq) 948 { 949 struct task_struct *curr = rq->curr; 950 struct sched_rt_entity *rt_se = &curr->rt; 951 u64 delta_exec; 952 u64 now; 953 954 if (curr->sched_class != &rt_sched_class) 955 return; 956 957 now = rq_clock_task(rq); 958 delta_exec = now - curr->se.exec_start; 959 if (unlikely((s64)delta_exec <= 0)) 960 return; 961 962 schedstat_set(curr->se.statistics.exec_max, 963 max(curr->se.statistics.exec_max, delta_exec)); 964 965 curr->se.sum_exec_runtime += delta_exec; 966 account_group_exec_runtime(curr, delta_exec); 967 968 curr->se.exec_start = now; 969 cgroup_account_cputime(curr, delta_exec); 970 971 sched_rt_avg_update(rq, delta_exec); 972 973 if (!rt_bandwidth_enabled()) 974 return; 975 976 for_each_sched_rt_entity(rt_se) { 977 struct rt_rq *rt_rq = rt_rq_of_se(rt_se); 978 979 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) { 980 raw_spin_lock(&rt_rq->rt_runtime_lock); 981 rt_rq->rt_time += delta_exec; 982 if (sched_rt_runtime_exceeded(rt_rq)) 983 resched_curr(rq); 984 raw_spin_unlock(&rt_rq->rt_runtime_lock); 985 } 986 } 987 } 988 989 static void 990 dequeue_top_rt_rq(struct rt_rq *rt_rq) 991 { 992 struct rq *rq = rq_of_rt_rq(rt_rq); 993 994 BUG_ON(&rq->rt != rt_rq); 995 996 if (!rt_rq->rt_queued) 997 return; 998 999 BUG_ON(!rq->nr_running); 1000 1001 sub_nr_running(rq, rt_rq->rt_nr_running); 1002 rt_rq->rt_queued = 0; 1003 1004 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */ 1005 cpufreq_update_util(rq, 0); 1006 } 1007 1008 static void 1009 enqueue_top_rt_rq(struct rt_rq *rt_rq) 1010 { 1011 struct rq *rq = rq_of_rt_rq(rt_rq); 1012 1013 BUG_ON(&rq->rt != rt_rq); 1014 1015 if (rt_rq->rt_queued) 1016 return; 1017 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running) 1018 return; 1019 1020 add_nr_running(rq, rt_rq->rt_nr_running); 1021 rt_rq->rt_queued = 1; 1022 1023 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */ 1024 cpufreq_update_util(rq, 0); 1025 } 1026 1027 #if defined CONFIG_SMP 1028 1029 static void 1030 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) 1031 { 1032 struct rq *rq = rq_of_rt_rq(rt_rq); 1033 1034 #ifdef CONFIG_RT_GROUP_SCHED 1035 /* 1036 * Change rq's cpupri only if rt_rq is the top queue. 1037 */ 1038 if (&rq->rt != rt_rq) 1039 return; 1040 #endif 1041 if (rq->online && prio < prev_prio) 1042 cpupri_set(&rq->rd->cpupri, rq->cpu, prio); 1043 } 1044 1045 static void 1046 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) 1047 { 1048 struct rq *rq = rq_of_rt_rq(rt_rq); 1049 1050 #ifdef CONFIG_RT_GROUP_SCHED 1051 /* 1052 * Change rq's cpupri only if rt_rq is the top queue. 1053 */ 1054 if (&rq->rt != rt_rq) 1055 return; 1056 #endif 1057 if (rq->online && rt_rq->highest_prio.curr != prev_prio) 1058 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr); 1059 } 1060 1061 #else /* CONFIG_SMP */ 1062 1063 static inline 1064 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} 1065 static inline 1066 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {} 1067 1068 #endif /* CONFIG_SMP */ 1069 1070 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 1071 static void 1072 inc_rt_prio(struct rt_rq *rt_rq, int prio) 1073 { 1074 int prev_prio = rt_rq->highest_prio.curr; 1075 1076 if (prio < prev_prio) 1077 rt_rq->highest_prio.curr = prio; 1078 1079 inc_rt_prio_smp(rt_rq, prio, prev_prio); 1080 } 1081 1082 static void 1083 dec_rt_prio(struct rt_rq *rt_rq, int prio) 1084 { 1085 int prev_prio = rt_rq->highest_prio.curr; 1086 1087 if (rt_rq->rt_nr_running) { 1088 1089 WARN_ON(prio < prev_prio); 1090 1091 /* 1092 * This may have been our highest task, and therefore 1093 * we may have some recomputation to do 1094 */ 1095 if (prio == prev_prio) { 1096 struct rt_prio_array *array = &rt_rq->active; 1097 1098 rt_rq->highest_prio.curr = 1099 sched_find_first_bit(array->bitmap); 1100 } 1101 1102 } else 1103 rt_rq->highest_prio.curr = MAX_RT_PRIO; 1104 1105 dec_rt_prio_smp(rt_rq, prio, prev_prio); 1106 } 1107 1108 #else 1109 1110 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {} 1111 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {} 1112 1113 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */ 1114 1115 #ifdef CONFIG_RT_GROUP_SCHED 1116 1117 static void 1118 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1119 { 1120 if (rt_se_boosted(rt_se)) 1121 rt_rq->rt_nr_boosted++; 1122 1123 if (rt_rq->tg) 1124 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth); 1125 } 1126 1127 static void 1128 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1129 { 1130 if (rt_se_boosted(rt_se)) 1131 rt_rq->rt_nr_boosted--; 1132 1133 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted); 1134 } 1135 1136 #else /* CONFIG_RT_GROUP_SCHED */ 1137 1138 static void 1139 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1140 { 1141 start_rt_bandwidth(&def_rt_bandwidth); 1142 } 1143 1144 static inline 1145 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {} 1146 1147 #endif /* CONFIG_RT_GROUP_SCHED */ 1148 1149 static inline 1150 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se) 1151 { 1152 struct rt_rq *group_rq = group_rt_rq(rt_se); 1153 1154 if (group_rq) 1155 return group_rq->rt_nr_running; 1156 else 1157 return 1; 1158 } 1159 1160 static inline 1161 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se) 1162 { 1163 struct rt_rq *group_rq = group_rt_rq(rt_se); 1164 struct task_struct *tsk; 1165 1166 if (group_rq) 1167 return group_rq->rr_nr_running; 1168 1169 tsk = rt_task_of(rt_se); 1170 1171 return (tsk->policy == SCHED_RR) ? 1 : 0; 1172 } 1173 1174 static inline 1175 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1176 { 1177 int prio = rt_se_prio(rt_se); 1178 1179 WARN_ON(!rt_prio(prio)); 1180 rt_rq->rt_nr_running += rt_se_nr_running(rt_se); 1181 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se); 1182 1183 inc_rt_prio(rt_rq, prio); 1184 inc_rt_migration(rt_se, rt_rq); 1185 inc_rt_group(rt_se, rt_rq); 1186 } 1187 1188 static inline 1189 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) 1190 { 1191 WARN_ON(!rt_prio(rt_se_prio(rt_se))); 1192 WARN_ON(!rt_rq->rt_nr_running); 1193 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se); 1194 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se); 1195 1196 dec_rt_prio(rt_rq, rt_se_prio(rt_se)); 1197 dec_rt_migration(rt_se, rt_rq); 1198 dec_rt_group(rt_se, rt_rq); 1199 } 1200 1201 /* 1202 * Change rt_se->run_list location unless SAVE && !MOVE 1203 * 1204 * assumes ENQUEUE/DEQUEUE flags match 1205 */ 1206 static inline bool move_entity(unsigned int flags) 1207 { 1208 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE) 1209 return false; 1210 1211 return true; 1212 } 1213 1214 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array) 1215 { 1216 list_del_init(&rt_se->run_list); 1217 1218 if (list_empty(array->queue + rt_se_prio(rt_se))) 1219 __clear_bit(rt_se_prio(rt_se), array->bitmap); 1220 1221 rt_se->on_list = 0; 1222 } 1223 1224 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) 1225 { 1226 struct rt_rq *rt_rq = rt_rq_of_se(rt_se); 1227 struct rt_prio_array *array = &rt_rq->active; 1228 struct rt_rq *group_rq = group_rt_rq(rt_se); 1229 struct list_head *queue = array->queue + rt_se_prio(rt_se); 1230 1231 /* 1232 * Don't enqueue the group if its throttled, or when empty. 1233 * The latter is a consequence of the former when a child group 1234 * get throttled and the current group doesn't have any other 1235 * active members. 1236 */ 1237 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) { 1238 if (rt_se->on_list) 1239 __delist_rt_entity(rt_se, array); 1240 return; 1241 } 1242 1243 if (move_entity(flags)) { 1244 WARN_ON_ONCE(rt_se->on_list); 1245 if (flags & ENQUEUE_HEAD) 1246 list_add(&rt_se->run_list, queue); 1247 else 1248 list_add_tail(&rt_se->run_list, queue); 1249 1250 __set_bit(rt_se_prio(rt_se), array->bitmap); 1251 rt_se->on_list = 1; 1252 } 1253 rt_se->on_rq = 1; 1254 1255 inc_rt_tasks(rt_se, rt_rq); 1256 } 1257 1258 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) 1259 { 1260 struct rt_rq *rt_rq = rt_rq_of_se(rt_se); 1261 struct rt_prio_array *array = &rt_rq->active; 1262 1263 if (move_entity(flags)) { 1264 WARN_ON_ONCE(!rt_se->on_list); 1265 __delist_rt_entity(rt_se, array); 1266 } 1267 rt_se->on_rq = 0; 1268 1269 dec_rt_tasks(rt_se, rt_rq); 1270 } 1271 1272 /* 1273 * Because the prio of an upper entry depends on the lower 1274 * entries, we must remove entries top - down. 1275 */ 1276 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags) 1277 { 1278 struct sched_rt_entity *back = NULL; 1279 1280 for_each_sched_rt_entity(rt_se) { 1281 rt_se->back = back; 1282 back = rt_se; 1283 } 1284 1285 dequeue_top_rt_rq(rt_rq_of_se(back)); 1286 1287 for (rt_se = back; rt_se; rt_se = rt_se->back) { 1288 if (on_rt_rq(rt_se)) 1289 __dequeue_rt_entity(rt_se, flags); 1290 } 1291 } 1292 1293 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) 1294 { 1295 struct rq *rq = rq_of_rt_se(rt_se); 1296 1297 dequeue_rt_stack(rt_se, flags); 1298 for_each_sched_rt_entity(rt_se) 1299 __enqueue_rt_entity(rt_se, flags); 1300 enqueue_top_rt_rq(&rq->rt); 1301 } 1302 1303 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags) 1304 { 1305 struct rq *rq = rq_of_rt_se(rt_se); 1306 1307 dequeue_rt_stack(rt_se, flags); 1308 1309 for_each_sched_rt_entity(rt_se) { 1310 struct rt_rq *rt_rq = group_rt_rq(rt_se); 1311 1312 if (rt_rq && rt_rq->rt_nr_running) 1313 __enqueue_rt_entity(rt_se, flags); 1314 } 1315 enqueue_top_rt_rq(&rq->rt); 1316 } 1317 1318 /* 1319 * Adding/removing a task to/from a priority array: 1320 */ 1321 static void 1322 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags) 1323 { 1324 struct sched_rt_entity *rt_se = &p->rt; 1325 1326 if (flags & ENQUEUE_WAKEUP) 1327 rt_se->timeout = 0; 1328 1329 enqueue_rt_entity(rt_se, flags); 1330 1331 if (!task_current(rq, p) && p->nr_cpus_allowed > 1) 1332 enqueue_pushable_task(rq, p); 1333 } 1334 1335 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags) 1336 { 1337 struct sched_rt_entity *rt_se = &p->rt; 1338 1339 update_curr_rt(rq); 1340 dequeue_rt_entity(rt_se, flags); 1341 1342 dequeue_pushable_task(rq, p); 1343 } 1344 1345 /* 1346 * Put task to the head or the end of the run list without the overhead of 1347 * dequeue followed by enqueue. 1348 */ 1349 static void 1350 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head) 1351 { 1352 if (on_rt_rq(rt_se)) { 1353 struct rt_prio_array *array = &rt_rq->active; 1354 struct list_head *queue = array->queue + rt_se_prio(rt_se); 1355 1356 if (head) 1357 list_move(&rt_se->run_list, queue); 1358 else 1359 list_move_tail(&rt_se->run_list, queue); 1360 } 1361 } 1362 1363 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head) 1364 { 1365 struct sched_rt_entity *rt_se = &p->rt; 1366 struct rt_rq *rt_rq; 1367 1368 for_each_sched_rt_entity(rt_se) { 1369 rt_rq = rt_rq_of_se(rt_se); 1370 requeue_rt_entity(rt_rq, rt_se, head); 1371 } 1372 } 1373 1374 static void yield_task_rt(struct rq *rq) 1375 { 1376 requeue_task_rt(rq, rq->curr, 0); 1377 } 1378 1379 #ifdef CONFIG_SMP 1380 static int find_lowest_rq(struct task_struct *task); 1381 1382 static int 1383 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags) 1384 { 1385 struct task_struct *curr; 1386 struct rq *rq; 1387 1388 /* For anything but wake ups, just return the task_cpu */ 1389 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK) 1390 goto out; 1391 1392 rq = cpu_rq(cpu); 1393 1394 rcu_read_lock(); 1395 curr = READ_ONCE(rq->curr); /* unlocked access */ 1396 1397 /* 1398 * If the current task on @p's runqueue is an RT task, then 1399 * try to see if we can wake this RT task up on another 1400 * runqueue. Otherwise simply start this RT task 1401 * on its current runqueue. 1402 * 1403 * We want to avoid overloading runqueues. If the woken 1404 * task is a higher priority, then it will stay on this CPU 1405 * and the lower prio task should be moved to another CPU. 1406 * Even though this will probably make the lower prio task 1407 * lose its cache, we do not want to bounce a higher task 1408 * around just because it gave up its CPU, perhaps for a 1409 * lock? 1410 * 1411 * For equal prio tasks, we just let the scheduler sort it out. 1412 * 1413 * Otherwise, just let it ride on the affined RQ and the 1414 * post-schedule router will push the preempted task away 1415 * 1416 * This test is optimistic, if we get it wrong the load-balancer 1417 * will have to sort it out. 1418 */ 1419 if (curr && unlikely(rt_task(curr)) && 1420 (curr->nr_cpus_allowed < 2 || 1421 curr->prio <= p->prio)) { 1422 int target = find_lowest_rq(p); 1423 1424 /* 1425 * Don't bother moving it if the destination CPU is 1426 * not running a lower priority task. 1427 */ 1428 if (target != -1 && 1429 p->prio < cpu_rq(target)->rt.highest_prio.curr) 1430 cpu = target; 1431 } 1432 rcu_read_unlock(); 1433 1434 out: 1435 return cpu; 1436 } 1437 1438 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p) 1439 { 1440 /* 1441 * Current can't be migrated, useless to reschedule, 1442 * let's hope p can move out. 1443 */ 1444 if (rq->curr->nr_cpus_allowed == 1 || 1445 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL)) 1446 return; 1447 1448 /* 1449 * p is migratable, so let's not schedule it and 1450 * see if it is pushed or pulled somewhere else. 1451 */ 1452 if (p->nr_cpus_allowed != 1 1453 && cpupri_find(&rq->rd->cpupri, p, NULL)) 1454 return; 1455 1456 /* 1457 * There appear to be other CPUs that can accept 1458 * the current task but none can run 'p', so lets reschedule 1459 * to try and push the current task away: 1460 */ 1461 requeue_task_rt(rq, p, 1); 1462 resched_curr(rq); 1463 } 1464 1465 #endif /* CONFIG_SMP */ 1466 1467 /* 1468 * Preempt the current task with a newly woken task if needed: 1469 */ 1470 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags) 1471 { 1472 if (p->prio < rq->curr->prio) { 1473 resched_curr(rq); 1474 return; 1475 } 1476 1477 #ifdef CONFIG_SMP 1478 /* 1479 * If: 1480 * 1481 * - the newly woken task is of equal priority to the current task 1482 * - the newly woken task is non-migratable while current is migratable 1483 * - current will be preempted on the next reschedule 1484 * 1485 * we should check to see if current can readily move to a different 1486 * cpu. If so, we will reschedule to allow the push logic to try 1487 * to move current somewhere else, making room for our non-migratable 1488 * task. 1489 */ 1490 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr)) 1491 check_preempt_equal_prio(rq, p); 1492 #endif 1493 } 1494 1495 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq, 1496 struct rt_rq *rt_rq) 1497 { 1498 struct rt_prio_array *array = &rt_rq->active; 1499 struct sched_rt_entity *next = NULL; 1500 struct list_head *queue; 1501 int idx; 1502 1503 idx = sched_find_first_bit(array->bitmap); 1504 BUG_ON(idx >= MAX_RT_PRIO); 1505 1506 queue = array->queue + idx; 1507 next = list_entry(queue->next, struct sched_rt_entity, run_list); 1508 1509 return next; 1510 } 1511 1512 static struct task_struct *_pick_next_task_rt(struct rq *rq) 1513 { 1514 struct sched_rt_entity *rt_se; 1515 struct task_struct *p; 1516 struct rt_rq *rt_rq = &rq->rt; 1517 1518 do { 1519 rt_se = pick_next_rt_entity(rq, rt_rq); 1520 BUG_ON(!rt_se); 1521 rt_rq = group_rt_rq(rt_se); 1522 } while (rt_rq); 1523 1524 p = rt_task_of(rt_se); 1525 p->se.exec_start = rq_clock_task(rq); 1526 1527 return p; 1528 } 1529 1530 static struct task_struct * 1531 pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 1532 { 1533 struct task_struct *p; 1534 struct rt_rq *rt_rq = &rq->rt; 1535 1536 if (need_pull_rt_task(rq, prev)) { 1537 /* 1538 * This is OK, because current is on_cpu, which avoids it being 1539 * picked for load-balance and preemption/IRQs are still 1540 * disabled avoiding further scheduler activity on it and we're 1541 * being very careful to re-start the picking loop. 1542 */ 1543 rq_unpin_lock(rq, rf); 1544 pull_rt_task(rq); 1545 rq_repin_lock(rq, rf); 1546 /* 1547 * pull_rt_task() can drop (and re-acquire) rq->lock; this 1548 * means a dl or stop task can slip in, in which case we need 1549 * to re-start task selection. 1550 */ 1551 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) || 1552 rq->dl.dl_nr_running)) 1553 return RETRY_TASK; 1554 } 1555 1556 /* 1557 * We may dequeue prev's rt_rq in put_prev_task(). 1558 * So, we update time before rt_nr_running check. 1559 */ 1560 if (prev->sched_class == &rt_sched_class) 1561 update_curr_rt(rq); 1562 1563 if (!rt_rq->rt_queued) 1564 return NULL; 1565 1566 put_prev_task(rq, prev); 1567 1568 p = _pick_next_task_rt(rq); 1569 1570 /* The running task is never eligible for pushing */ 1571 dequeue_pushable_task(rq, p); 1572 1573 rt_queue_push_tasks(rq); 1574 1575 return p; 1576 } 1577 1578 static void put_prev_task_rt(struct rq *rq, struct task_struct *p) 1579 { 1580 update_curr_rt(rq); 1581 1582 /* 1583 * The previous task needs to be made eligible for pushing 1584 * if it is still active 1585 */ 1586 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1) 1587 enqueue_pushable_task(rq, p); 1588 } 1589 1590 #ifdef CONFIG_SMP 1591 1592 /* Only try algorithms three times */ 1593 #define RT_MAX_TRIES 3 1594 1595 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu) 1596 { 1597 if (!task_running(rq, p) && 1598 cpumask_test_cpu(cpu, &p->cpus_allowed)) 1599 return 1; 1600 1601 return 0; 1602 } 1603 1604 /* 1605 * Return the highest pushable rq's task, which is suitable to be executed 1606 * on the CPU, NULL otherwise 1607 */ 1608 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu) 1609 { 1610 struct plist_head *head = &rq->rt.pushable_tasks; 1611 struct task_struct *p; 1612 1613 if (!has_pushable_tasks(rq)) 1614 return NULL; 1615 1616 plist_for_each_entry(p, head, pushable_tasks) { 1617 if (pick_rt_task(rq, p, cpu)) 1618 return p; 1619 } 1620 1621 return NULL; 1622 } 1623 1624 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask); 1625 1626 static int find_lowest_rq(struct task_struct *task) 1627 { 1628 struct sched_domain *sd; 1629 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask); 1630 int this_cpu = smp_processor_id(); 1631 int cpu = task_cpu(task); 1632 1633 /* Make sure the mask is initialized first */ 1634 if (unlikely(!lowest_mask)) 1635 return -1; 1636 1637 if (task->nr_cpus_allowed == 1) 1638 return -1; /* No other targets possible */ 1639 1640 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask)) 1641 return -1; /* No targets found */ 1642 1643 /* 1644 * At this point we have built a mask of CPUs representing the 1645 * lowest priority tasks in the system. Now we want to elect 1646 * the best one based on our affinity and topology. 1647 * 1648 * We prioritize the last CPU that the task executed on since 1649 * it is most likely cache-hot in that location. 1650 */ 1651 if (cpumask_test_cpu(cpu, lowest_mask)) 1652 return cpu; 1653 1654 /* 1655 * Otherwise, we consult the sched_domains span maps to figure 1656 * out which CPU is logically closest to our hot cache data. 1657 */ 1658 if (!cpumask_test_cpu(this_cpu, lowest_mask)) 1659 this_cpu = -1; /* Skip this_cpu opt if not among lowest */ 1660 1661 rcu_read_lock(); 1662 for_each_domain(cpu, sd) { 1663 if (sd->flags & SD_WAKE_AFFINE) { 1664 int best_cpu; 1665 1666 /* 1667 * "this_cpu" is cheaper to preempt than a 1668 * remote processor. 1669 */ 1670 if (this_cpu != -1 && 1671 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { 1672 rcu_read_unlock(); 1673 return this_cpu; 1674 } 1675 1676 best_cpu = cpumask_first_and(lowest_mask, 1677 sched_domain_span(sd)); 1678 if (best_cpu < nr_cpu_ids) { 1679 rcu_read_unlock(); 1680 return best_cpu; 1681 } 1682 } 1683 } 1684 rcu_read_unlock(); 1685 1686 /* 1687 * And finally, if there were no matches within the domains 1688 * just give the caller *something* to work with from the compatible 1689 * locations. 1690 */ 1691 if (this_cpu != -1) 1692 return this_cpu; 1693 1694 cpu = cpumask_any(lowest_mask); 1695 if (cpu < nr_cpu_ids) 1696 return cpu; 1697 1698 return -1; 1699 } 1700 1701 /* Will lock the rq it finds */ 1702 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) 1703 { 1704 struct rq *lowest_rq = NULL; 1705 int tries; 1706 int cpu; 1707 1708 for (tries = 0; tries < RT_MAX_TRIES; tries++) { 1709 cpu = find_lowest_rq(task); 1710 1711 if ((cpu == -1) || (cpu == rq->cpu)) 1712 break; 1713 1714 lowest_rq = cpu_rq(cpu); 1715 1716 if (lowest_rq->rt.highest_prio.curr <= task->prio) { 1717 /* 1718 * Target rq has tasks of equal or higher priority, 1719 * retrying does not release any lock and is unlikely 1720 * to yield a different result. 1721 */ 1722 lowest_rq = NULL; 1723 break; 1724 } 1725 1726 /* if the prio of this runqueue changed, try again */ 1727 if (double_lock_balance(rq, lowest_rq)) { 1728 /* 1729 * We had to unlock the run queue. In 1730 * the mean time, task could have 1731 * migrated already or had its affinity changed. 1732 * Also make sure that it wasn't scheduled on its rq. 1733 */ 1734 if (unlikely(task_rq(task) != rq || 1735 !cpumask_test_cpu(lowest_rq->cpu, &task->cpus_allowed) || 1736 task_running(rq, task) || 1737 !rt_task(task) || 1738 !task_on_rq_queued(task))) { 1739 1740 double_unlock_balance(rq, lowest_rq); 1741 lowest_rq = NULL; 1742 break; 1743 } 1744 } 1745 1746 /* If this rq is still suitable use it. */ 1747 if (lowest_rq->rt.highest_prio.curr > task->prio) 1748 break; 1749 1750 /* try again */ 1751 double_unlock_balance(rq, lowest_rq); 1752 lowest_rq = NULL; 1753 } 1754 1755 return lowest_rq; 1756 } 1757 1758 static struct task_struct *pick_next_pushable_task(struct rq *rq) 1759 { 1760 struct task_struct *p; 1761 1762 if (!has_pushable_tasks(rq)) 1763 return NULL; 1764 1765 p = plist_first_entry(&rq->rt.pushable_tasks, 1766 struct task_struct, pushable_tasks); 1767 1768 BUG_ON(rq->cpu != task_cpu(p)); 1769 BUG_ON(task_current(rq, p)); 1770 BUG_ON(p->nr_cpus_allowed <= 1); 1771 1772 BUG_ON(!task_on_rq_queued(p)); 1773 BUG_ON(!rt_task(p)); 1774 1775 return p; 1776 } 1777 1778 /* 1779 * If the current CPU has more than one RT task, see if the non 1780 * running task can migrate over to a CPU that is running a task 1781 * of lesser priority. 1782 */ 1783 static int push_rt_task(struct rq *rq) 1784 { 1785 struct task_struct *next_task; 1786 struct rq *lowest_rq; 1787 int ret = 0; 1788 1789 if (!rq->rt.overloaded) 1790 return 0; 1791 1792 next_task = pick_next_pushable_task(rq); 1793 if (!next_task) 1794 return 0; 1795 1796 retry: 1797 if (unlikely(next_task == rq->curr)) { 1798 WARN_ON(1); 1799 return 0; 1800 } 1801 1802 /* 1803 * It's possible that the next_task slipped in of 1804 * higher priority than current. If that's the case 1805 * just reschedule current. 1806 */ 1807 if (unlikely(next_task->prio < rq->curr->prio)) { 1808 resched_curr(rq); 1809 return 0; 1810 } 1811 1812 /* We might release rq lock */ 1813 get_task_struct(next_task); 1814 1815 /* find_lock_lowest_rq locks the rq if found */ 1816 lowest_rq = find_lock_lowest_rq(next_task, rq); 1817 if (!lowest_rq) { 1818 struct task_struct *task; 1819 /* 1820 * find_lock_lowest_rq releases rq->lock 1821 * so it is possible that next_task has migrated. 1822 * 1823 * We need to make sure that the task is still on the same 1824 * run-queue and is also still the next task eligible for 1825 * pushing. 1826 */ 1827 task = pick_next_pushable_task(rq); 1828 if (task == next_task) { 1829 /* 1830 * The task hasn't migrated, and is still the next 1831 * eligible task, but we failed to find a run-queue 1832 * to push it to. Do not retry in this case, since 1833 * other CPUs will pull from us when ready. 1834 */ 1835 goto out; 1836 } 1837 1838 if (!task) 1839 /* No more tasks, just exit */ 1840 goto out; 1841 1842 /* 1843 * Something has shifted, try again. 1844 */ 1845 put_task_struct(next_task); 1846 next_task = task; 1847 goto retry; 1848 } 1849 1850 deactivate_task(rq, next_task, 0); 1851 set_task_cpu(next_task, lowest_rq->cpu); 1852 activate_task(lowest_rq, next_task, 0); 1853 ret = 1; 1854 1855 resched_curr(lowest_rq); 1856 1857 double_unlock_balance(rq, lowest_rq); 1858 1859 out: 1860 put_task_struct(next_task); 1861 1862 return ret; 1863 } 1864 1865 static void push_rt_tasks(struct rq *rq) 1866 { 1867 /* push_rt_task will return true if it moved an RT */ 1868 while (push_rt_task(rq)) 1869 ; 1870 } 1871 1872 #ifdef HAVE_RT_PUSH_IPI 1873 1874 /* 1875 * When a high priority task schedules out from a CPU and a lower priority 1876 * task is scheduled in, a check is made to see if there's any RT tasks 1877 * on other CPUs that are waiting to run because a higher priority RT task 1878 * is currently running on its CPU. In this case, the CPU with multiple RT 1879 * tasks queued on it (overloaded) needs to be notified that a CPU has opened 1880 * up that may be able to run one of its non-running queued RT tasks. 1881 * 1882 * All CPUs with overloaded RT tasks need to be notified as there is currently 1883 * no way to know which of these CPUs have the highest priority task waiting 1884 * to run. Instead of trying to take a spinlock on each of these CPUs, 1885 * which has shown to cause large latency when done on machines with many 1886 * CPUs, sending an IPI to the CPUs to have them push off the overloaded 1887 * RT tasks waiting to run. 1888 * 1889 * Just sending an IPI to each of the CPUs is also an issue, as on large 1890 * count CPU machines, this can cause an IPI storm on a CPU, especially 1891 * if its the only CPU with multiple RT tasks queued, and a large number 1892 * of CPUs scheduling a lower priority task at the same time. 1893 * 1894 * Each root domain has its own irq work function that can iterate over 1895 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT 1896 * tassk must be checked if there's one or many CPUs that are lowering 1897 * their priority, there's a single irq work iterator that will try to 1898 * push off RT tasks that are waiting to run. 1899 * 1900 * When a CPU schedules a lower priority task, it will kick off the 1901 * irq work iterator that will jump to each CPU with overloaded RT tasks. 1902 * As it only takes the first CPU that schedules a lower priority task 1903 * to start the process, the rto_start variable is incremented and if 1904 * the atomic result is one, then that CPU will try to take the rto_lock. 1905 * This prevents high contention on the lock as the process handles all 1906 * CPUs scheduling lower priority tasks. 1907 * 1908 * All CPUs that are scheduling a lower priority task will increment the 1909 * rt_loop_next variable. This will make sure that the irq work iterator 1910 * checks all RT overloaded CPUs whenever a CPU schedules a new lower 1911 * priority task, even if the iterator is in the middle of a scan. Incrementing 1912 * the rt_loop_next will cause the iterator to perform another scan. 1913 * 1914 */ 1915 static int rto_next_cpu(struct root_domain *rd) 1916 { 1917 int next; 1918 int cpu; 1919 1920 /* 1921 * When starting the IPI RT pushing, the rto_cpu is set to -1, 1922 * rt_next_cpu() will simply return the first CPU found in 1923 * the rto_mask. 1924 * 1925 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it 1926 * will return the next CPU found in the rto_mask. 1927 * 1928 * If there are no more CPUs left in the rto_mask, then a check is made 1929 * against rto_loop and rto_loop_next. rto_loop is only updated with 1930 * the rto_lock held, but any CPU may increment the rto_loop_next 1931 * without any locking. 1932 */ 1933 for (;;) { 1934 1935 /* When rto_cpu is -1 this acts like cpumask_first() */ 1936 cpu = cpumask_next(rd->rto_cpu, rd->rto_mask); 1937 1938 rd->rto_cpu = cpu; 1939 1940 if (cpu < nr_cpu_ids) 1941 return cpu; 1942 1943 rd->rto_cpu = -1; 1944 1945 /* 1946 * ACQUIRE ensures we see the @rto_mask changes 1947 * made prior to the @next value observed. 1948 * 1949 * Matches WMB in rt_set_overload(). 1950 */ 1951 next = atomic_read_acquire(&rd->rto_loop_next); 1952 1953 if (rd->rto_loop == next) 1954 break; 1955 1956 rd->rto_loop = next; 1957 } 1958 1959 return -1; 1960 } 1961 1962 static inline bool rto_start_trylock(atomic_t *v) 1963 { 1964 return !atomic_cmpxchg_acquire(v, 0, 1); 1965 } 1966 1967 static inline void rto_start_unlock(atomic_t *v) 1968 { 1969 atomic_set_release(v, 0); 1970 } 1971 1972 static void tell_cpu_to_push(struct rq *rq) 1973 { 1974 int cpu = -1; 1975 1976 /* Keep the loop going if the IPI is currently active */ 1977 atomic_inc(&rq->rd->rto_loop_next); 1978 1979 /* Only one CPU can initiate a loop at a time */ 1980 if (!rto_start_trylock(&rq->rd->rto_loop_start)) 1981 return; 1982 1983 raw_spin_lock(&rq->rd->rto_lock); 1984 1985 /* 1986 * The rto_cpu is updated under the lock, if it has a valid CPU 1987 * then the IPI is still running and will continue due to the 1988 * update to loop_next, and nothing needs to be done here. 1989 * Otherwise it is finishing up and an ipi needs to be sent. 1990 */ 1991 if (rq->rd->rto_cpu < 0) 1992 cpu = rto_next_cpu(rq->rd); 1993 1994 raw_spin_unlock(&rq->rd->rto_lock); 1995 1996 rto_start_unlock(&rq->rd->rto_loop_start); 1997 1998 if (cpu >= 0) { 1999 /* Make sure the rd does not get freed while pushing */ 2000 sched_get_rd(rq->rd); 2001 irq_work_queue_on(&rq->rd->rto_push_work, cpu); 2002 } 2003 } 2004 2005 /* Called from hardirq context */ 2006 void rto_push_irq_work_func(struct irq_work *work) 2007 { 2008 struct root_domain *rd = 2009 container_of(work, struct root_domain, rto_push_work); 2010 struct rq *rq; 2011 int cpu; 2012 2013 rq = this_rq(); 2014 2015 /* 2016 * We do not need to grab the lock to check for has_pushable_tasks. 2017 * When it gets updated, a check is made if a push is possible. 2018 */ 2019 if (has_pushable_tasks(rq)) { 2020 raw_spin_lock(&rq->lock); 2021 push_rt_tasks(rq); 2022 raw_spin_unlock(&rq->lock); 2023 } 2024 2025 raw_spin_lock(&rd->rto_lock); 2026 2027 /* Pass the IPI to the next rt overloaded queue */ 2028 cpu = rto_next_cpu(rd); 2029 2030 raw_spin_unlock(&rd->rto_lock); 2031 2032 if (cpu < 0) { 2033 sched_put_rd(rd); 2034 return; 2035 } 2036 2037 /* Try the next RT overloaded CPU */ 2038 irq_work_queue_on(&rd->rto_push_work, cpu); 2039 } 2040 #endif /* HAVE_RT_PUSH_IPI */ 2041 2042 static void pull_rt_task(struct rq *this_rq) 2043 { 2044 int this_cpu = this_rq->cpu, cpu; 2045 bool resched = false; 2046 struct task_struct *p; 2047 struct rq *src_rq; 2048 int rt_overload_count = rt_overloaded(this_rq); 2049 2050 if (likely(!rt_overload_count)) 2051 return; 2052 2053 /* 2054 * Match the barrier from rt_set_overloaded; this guarantees that if we 2055 * see overloaded we must also see the rto_mask bit. 2056 */ 2057 smp_rmb(); 2058 2059 /* If we are the only overloaded CPU do nothing */ 2060 if (rt_overload_count == 1 && 2061 cpumask_test_cpu(this_rq->cpu, this_rq->rd->rto_mask)) 2062 return; 2063 2064 #ifdef HAVE_RT_PUSH_IPI 2065 if (sched_feat(RT_PUSH_IPI)) { 2066 tell_cpu_to_push(this_rq); 2067 return; 2068 } 2069 #endif 2070 2071 for_each_cpu(cpu, this_rq->rd->rto_mask) { 2072 if (this_cpu == cpu) 2073 continue; 2074 2075 src_rq = cpu_rq(cpu); 2076 2077 /* 2078 * Don't bother taking the src_rq->lock if the next highest 2079 * task is known to be lower-priority than our current task. 2080 * This may look racy, but if this value is about to go 2081 * logically higher, the src_rq will push this task away. 2082 * And if its going logically lower, we do not care 2083 */ 2084 if (src_rq->rt.highest_prio.next >= 2085 this_rq->rt.highest_prio.curr) 2086 continue; 2087 2088 /* 2089 * We can potentially drop this_rq's lock in 2090 * double_lock_balance, and another CPU could 2091 * alter this_rq 2092 */ 2093 double_lock_balance(this_rq, src_rq); 2094 2095 /* 2096 * We can pull only a task, which is pushable 2097 * on its rq, and no others. 2098 */ 2099 p = pick_highest_pushable_task(src_rq, this_cpu); 2100 2101 /* 2102 * Do we have an RT task that preempts 2103 * the to-be-scheduled task? 2104 */ 2105 if (p && (p->prio < this_rq->rt.highest_prio.curr)) { 2106 WARN_ON(p == src_rq->curr); 2107 WARN_ON(!task_on_rq_queued(p)); 2108 2109 /* 2110 * There's a chance that p is higher in priority 2111 * than what's currently running on its CPU. 2112 * This is just that p is wakeing up and hasn't 2113 * had a chance to schedule. We only pull 2114 * p if it is lower in priority than the 2115 * current task on the run queue 2116 */ 2117 if (p->prio < src_rq->curr->prio) 2118 goto skip; 2119 2120 resched = true; 2121 2122 deactivate_task(src_rq, p, 0); 2123 set_task_cpu(p, this_cpu); 2124 activate_task(this_rq, p, 0); 2125 /* 2126 * We continue with the search, just in 2127 * case there's an even higher prio task 2128 * in another runqueue. (low likelihood 2129 * but possible) 2130 */ 2131 } 2132 skip: 2133 double_unlock_balance(this_rq, src_rq); 2134 } 2135 2136 if (resched) 2137 resched_curr(this_rq); 2138 } 2139 2140 /* 2141 * If we are not running and we are not going to reschedule soon, we should 2142 * try to push tasks away now 2143 */ 2144 static void task_woken_rt(struct rq *rq, struct task_struct *p) 2145 { 2146 if (!task_running(rq, p) && 2147 !test_tsk_need_resched(rq->curr) && 2148 p->nr_cpus_allowed > 1 && 2149 (dl_task(rq->curr) || rt_task(rq->curr)) && 2150 (rq->curr->nr_cpus_allowed < 2 || 2151 rq->curr->prio <= p->prio)) 2152 push_rt_tasks(rq); 2153 } 2154 2155 /* Assumes rq->lock is held */ 2156 static void rq_online_rt(struct rq *rq) 2157 { 2158 if (rq->rt.overloaded) 2159 rt_set_overload(rq); 2160 2161 __enable_runtime(rq); 2162 2163 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr); 2164 } 2165 2166 /* Assumes rq->lock is held */ 2167 static void rq_offline_rt(struct rq *rq) 2168 { 2169 if (rq->rt.overloaded) 2170 rt_clear_overload(rq); 2171 2172 __disable_runtime(rq); 2173 2174 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID); 2175 } 2176 2177 /* 2178 * When switch from the rt queue, we bring ourselves to a position 2179 * that we might want to pull RT tasks from other runqueues. 2180 */ 2181 static void switched_from_rt(struct rq *rq, struct task_struct *p) 2182 { 2183 /* 2184 * If there are other RT tasks then we will reschedule 2185 * and the scheduling of the other RT tasks will handle 2186 * the balancing. But if we are the last RT task 2187 * we may need to handle the pulling of RT tasks 2188 * now. 2189 */ 2190 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running) 2191 return; 2192 2193 rt_queue_pull_task(rq); 2194 } 2195 2196 void __init init_sched_rt_class(void) 2197 { 2198 unsigned int i; 2199 2200 for_each_possible_cpu(i) { 2201 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i), 2202 GFP_KERNEL, cpu_to_node(i)); 2203 } 2204 } 2205 #endif /* CONFIG_SMP */ 2206 2207 /* 2208 * When switching a task to RT, we may overload the runqueue 2209 * with RT tasks. In this case we try to push them off to 2210 * other runqueues. 2211 */ 2212 static void switched_to_rt(struct rq *rq, struct task_struct *p) 2213 { 2214 /* 2215 * If we are already running, then there's nothing 2216 * that needs to be done. But if we are not running 2217 * we may need to preempt the current running task. 2218 * If that current running task is also an RT task 2219 * then see if we can move to another run queue. 2220 */ 2221 if (task_on_rq_queued(p) && rq->curr != p) { 2222 #ifdef CONFIG_SMP 2223 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded) 2224 rt_queue_push_tasks(rq); 2225 #endif /* CONFIG_SMP */ 2226 if (p->prio < rq->curr->prio && cpu_online(cpu_of(rq))) 2227 resched_curr(rq); 2228 } 2229 } 2230 2231 /* 2232 * Priority of the task has changed. This may cause 2233 * us to initiate a push or pull. 2234 */ 2235 static void 2236 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio) 2237 { 2238 if (!task_on_rq_queued(p)) 2239 return; 2240 2241 if (rq->curr == p) { 2242 #ifdef CONFIG_SMP 2243 /* 2244 * If our priority decreases while running, we 2245 * may need to pull tasks to this runqueue. 2246 */ 2247 if (oldprio < p->prio) 2248 rt_queue_pull_task(rq); 2249 2250 /* 2251 * If there's a higher priority task waiting to run 2252 * then reschedule. 2253 */ 2254 if (p->prio > rq->rt.highest_prio.curr) 2255 resched_curr(rq); 2256 #else 2257 /* For UP simply resched on drop of prio */ 2258 if (oldprio < p->prio) 2259 resched_curr(rq); 2260 #endif /* CONFIG_SMP */ 2261 } else { 2262 /* 2263 * This task is not running, but if it is 2264 * greater than the current running task 2265 * then reschedule. 2266 */ 2267 if (p->prio < rq->curr->prio) 2268 resched_curr(rq); 2269 } 2270 } 2271 2272 #ifdef CONFIG_POSIX_TIMERS 2273 static void watchdog(struct rq *rq, struct task_struct *p) 2274 { 2275 unsigned long soft, hard; 2276 2277 /* max may change after cur was read, this will be fixed next tick */ 2278 soft = task_rlimit(p, RLIMIT_RTTIME); 2279 hard = task_rlimit_max(p, RLIMIT_RTTIME); 2280 2281 if (soft != RLIM_INFINITY) { 2282 unsigned long next; 2283 2284 if (p->rt.watchdog_stamp != jiffies) { 2285 p->rt.timeout++; 2286 p->rt.watchdog_stamp = jiffies; 2287 } 2288 2289 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ); 2290 if (p->rt.timeout > next) 2291 p->cputime_expires.sched_exp = p->se.sum_exec_runtime; 2292 } 2293 } 2294 #else 2295 static inline void watchdog(struct rq *rq, struct task_struct *p) { } 2296 #endif 2297 2298 /* 2299 * scheduler tick hitting a task of our scheduling class. 2300 * 2301 * NOTE: This function can be called remotely by the tick offload that 2302 * goes along full dynticks. Therefore no local assumption can be made 2303 * and everything must be accessed through the @rq and @curr passed in 2304 * parameters. 2305 */ 2306 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued) 2307 { 2308 struct sched_rt_entity *rt_se = &p->rt; 2309 2310 update_curr_rt(rq); 2311 2312 watchdog(rq, p); 2313 2314 /* 2315 * RR tasks need a special form of timeslice management. 2316 * FIFO tasks have no timeslices. 2317 */ 2318 if (p->policy != SCHED_RR) 2319 return; 2320 2321 if (--p->rt.time_slice) 2322 return; 2323 2324 p->rt.time_slice = sched_rr_timeslice; 2325 2326 /* 2327 * Requeue to the end of queue if we (and all of our ancestors) are not 2328 * the only element on the queue 2329 */ 2330 for_each_sched_rt_entity(rt_se) { 2331 if (rt_se->run_list.prev != rt_se->run_list.next) { 2332 requeue_task_rt(rq, p, 0); 2333 resched_curr(rq); 2334 return; 2335 } 2336 } 2337 } 2338 2339 static void set_curr_task_rt(struct rq *rq) 2340 { 2341 struct task_struct *p = rq->curr; 2342 2343 p->se.exec_start = rq_clock_task(rq); 2344 2345 /* The running task is never eligible for pushing */ 2346 dequeue_pushable_task(rq, p); 2347 } 2348 2349 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task) 2350 { 2351 /* 2352 * Time slice is 0 for SCHED_FIFO tasks 2353 */ 2354 if (task->policy == SCHED_RR) 2355 return sched_rr_timeslice; 2356 else 2357 return 0; 2358 } 2359 2360 const struct sched_class rt_sched_class = { 2361 .next = &fair_sched_class, 2362 .enqueue_task = enqueue_task_rt, 2363 .dequeue_task = dequeue_task_rt, 2364 .yield_task = yield_task_rt, 2365 2366 .check_preempt_curr = check_preempt_curr_rt, 2367 2368 .pick_next_task = pick_next_task_rt, 2369 .put_prev_task = put_prev_task_rt, 2370 2371 #ifdef CONFIG_SMP 2372 .select_task_rq = select_task_rq_rt, 2373 2374 .set_cpus_allowed = set_cpus_allowed_common, 2375 .rq_online = rq_online_rt, 2376 .rq_offline = rq_offline_rt, 2377 .task_woken = task_woken_rt, 2378 .switched_from = switched_from_rt, 2379 #endif 2380 2381 .set_curr_task = set_curr_task_rt, 2382 .task_tick = task_tick_rt, 2383 2384 .get_rr_interval = get_rr_interval_rt, 2385 2386 .prio_changed = prio_changed_rt, 2387 .switched_to = switched_to_rt, 2388 2389 .update_curr = update_curr_rt, 2390 }; 2391 2392 #ifdef CONFIG_RT_GROUP_SCHED 2393 /* 2394 * Ensure that the real time constraints are schedulable. 2395 */ 2396 static DEFINE_MUTEX(rt_constraints_mutex); 2397 2398 /* Must be called with tasklist_lock held */ 2399 static inline int tg_has_rt_tasks(struct task_group *tg) 2400 { 2401 struct task_struct *g, *p; 2402 2403 /* 2404 * Autogroups do not have RT tasks; see autogroup_create(). 2405 */ 2406 if (task_group_is_autogroup(tg)) 2407 return 0; 2408 2409 for_each_process_thread(g, p) { 2410 if (rt_task(p) && task_group(p) == tg) 2411 return 1; 2412 } 2413 2414 return 0; 2415 } 2416 2417 struct rt_schedulable_data { 2418 struct task_group *tg; 2419 u64 rt_period; 2420 u64 rt_runtime; 2421 }; 2422 2423 static int tg_rt_schedulable(struct task_group *tg, void *data) 2424 { 2425 struct rt_schedulable_data *d = data; 2426 struct task_group *child; 2427 unsigned long total, sum = 0; 2428 u64 period, runtime; 2429 2430 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 2431 runtime = tg->rt_bandwidth.rt_runtime; 2432 2433 if (tg == d->tg) { 2434 period = d->rt_period; 2435 runtime = d->rt_runtime; 2436 } 2437 2438 /* 2439 * Cannot have more runtime than the period. 2440 */ 2441 if (runtime > period && runtime != RUNTIME_INF) 2442 return -EINVAL; 2443 2444 /* 2445 * Ensure we don't starve existing RT tasks. 2446 */ 2447 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 2448 return -EBUSY; 2449 2450 total = to_ratio(period, runtime); 2451 2452 /* 2453 * Nobody can have more than the global setting allows. 2454 */ 2455 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 2456 return -EINVAL; 2457 2458 /* 2459 * The sum of our children's runtime should not exceed our own. 2460 */ 2461 list_for_each_entry_rcu(child, &tg->children, siblings) { 2462 period = ktime_to_ns(child->rt_bandwidth.rt_period); 2463 runtime = child->rt_bandwidth.rt_runtime; 2464 2465 if (child == d->tg) { 2466 period = d->rt_period; 2467 runtime = d->rt_runtime; 2468 } 2469 2470 sum += to_ratio(period, runtime); 2471 } 2472 2473 if (sum > total) 2474 return -EINVAL; 2475 2476 return 0; 2477 } 2478 2479 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 2480 { 2481 int ret; 2482 2483 struct rt_schedulable_data data = { 2484 .tg = tg, 2485 .rt_period = period, 2486 .rt_runtime = runtime, 2487 }; 2488 2489 rcu_read_lock(); 2490 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 2491 rcu_read_unlock(); 2492 2493 return ret; 2494 } 2495 2496 static int tg_set_rt_bandwidth(struct task_group *tg, 2497 u64 rt_period, u64 rt_runtime) 2498 { 2499 int i, err = 0; 2500 2501 /* 2502 * Disallowing the root group RT runtime is BAD, it would disallow the 2503 * kernel creating (and or operating) RT threads. 2504 */ 2505 if (tg == &root_task_group && rt_runtime == 0) 2506 return -EINVAL; 2507 2508 /* No period doesn't make any sense. */ 2509 if (rt_period == 0) 2510 return -EINVAL; 2511 2512 mutex_lock(&rt_constraints_mutex); 2513 read_lock(&tasklist_lock); 2514 err = __rt_schedulable(tg, rt_period, rt_runtime); 2515 if (err) 2516 goto unlock; 2517 2518 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 2519 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 2520 tg->rt_bandwidth.rt_runtime = rt_runtime; 2521 2522 for_each_possible_cpu(i) { 2523 struct rt_rq *rt_rq = tg->rt_rq[i]; 2524 2525 raw_spin_lock(&rt_rq->rt_runtime_lock); 2526 rt_rq->rt_runtime = rt_runtime; 2527 raw_spin_unlock(&rt_rq->rt_runtime_lock); 2528 } 2529 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 2530 unlock: 2531 read_unlock(&tasklist_lock); 2532 mutex_unlock(&rt_constraints_mutex); 2533 2534 return err; 2535 } 2536 2537 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 2538 { 2539 u64 rt_runtime, rt_period; 2540 2541 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 2542 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 2543 if (rt_runtime_us < 0) 2544 rt_runtime = RUNTIME_INF; 2545 2546 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 2547 } 2548 2549 long sched_group_rt_runtime(struct task_group *tg) 2550 { 2551 u64 rt_runtime_us; 2552 2553 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 2554 return -1; 2555 2556 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 2557 do_div(rt_runtime_us, NSEC_PER_USEC); 2558 return rt_runtime_us; 2559 } 2560 2561 int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us) 2562 { 2563 u64 rt_runtime, rt_period; 2564 2565 rt_period = rt_period_us * NSEC_PER_USEC; 2566 rt_runtime = tg->rt_bandwidth.rt_runtime; 2567 2568 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 2569 } 2570 2571 long sched_group_rt_period(struct task_group *tg) 2572 { 2573 u64 rt_period_us; 2574 2575 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 2576 do_div(rt_period_us, NSEC_PER_USEC); 2577 return rt_period_us; 2578 } 2579 2580 static int sched_rt_global_constraints(void) 2581 { 2582 int ret = 0; 2583 2584 mutex_lock(&rt_constraints_mutex); 2585 read_lock(&tasklist_lock); 2586 ret = __rt_schedulable(NULL, 0, 0); 2587 read_unlock(&tasklist_lock); 2588 mutex_unlock(&rt_constraints_mutex); 2589 2590 return ret; 2591 } 2592 2593 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 2594 { 2595 /* Don't accept realtime tasks when there is no way for them to run */ 2596 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 2597 return 0; 2598 2599 return 1; 2600 } 2601 2602 #else /* !CONFIG_RT_GROUP_SCHED */ 2603 static int sched_rt_global_constraints(void) 2604 { 2605 unsigned long flags; 2606 int i; 2607 2608 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 2609 for_each_possible_cpu(i) { 2610 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 2611 2612 raw_spin_lock(&rt_rq->rt_runtime_lock); 2613 rt_rq->rt_runtime = global_rt_runtime(); 2614 raw_spin_unlock(&rt_rq->rt_runtime_lock); 2615 } 2616 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 2617 2618 return 0; 2619 } 2620 #endif /* CONFIG_RT_GROUP_SCHED */ 2621 2622 static int sched_rt_global_validate(void) 2623 { 2624 if (sysctl_sched_rt_period <= 0) 2625 return -EINVAL; 2626 2627 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 2628 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 2629 return -EINVAL; 2630 2631 return 0; 2632 } 2633 2634 static void sched_rt_do_global(void) 2635 { 2636 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 2637 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 2638 } 2639 2640 int sched_rt_handler(struct ctl_table *table, int write, 2641 void __user *buffer, size_t *lenp, 2642 loff_t *ppos) 2643 { 2644 int old_period, old_runtime; 2645 static DEFINE_MUTEX(mutex); 2646 int ret; 2647 2648 mutex_lock(&mutex); 2649 old_period = sysctl_sched_rt_period; 2650 old_runtime = sysctl_sched_rt_runtime; 2651 2652 ret = proc_dointvec(table, write, buffer, lenp, ppos); 2653 2654 if (!ret && write) { 2655 ret = sched_rt_global_validate(); 2656 if (ret) 2657 goto undo; 2658 2659 ret = sched_dl_global_validate(); 2660 if (ret) 2661 goto undo; 2662 2663 ret = sched_rt_global_constraints(); 2664 if (ret) 2665 goto undo; 2666 2667 sched_rt_do_global(); 2668 sched_dl_do_global(); 2669 } 2670 if (0) { 2671 undo: 2672 sysctl_sched_rt_period = old_period; 2673 sysctl_sched_rt_runtime = old_runtime; 2674 } 2675 mutex_unlock(&mutex); 2676 2677 return ret; 2678 } 2679 2680 int sched_rr_handler(struct ctl_table *table, int write, 2681 void __user *buffer, size_t *lenp, 2682 loff_t *ppos) 2683 { 2684 int ret; 2685 static DEFINE_MUTEX(mutex); 2686 2687 mutex_lock(&mutex); 2688 ret = proc_dointvec(table, write, buffer, lenp, ppos); 2689 /* 2690 * Make sure that internally we keep jiffies. 2691 * Also, writing zero resets the timeslice to default: 2692 */ 2693 if (!ret && write) { 2694 sched_rr_timeslice = 2695 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE : 2696 msecs_to_jiffies(sysctl_sched_rr_timeslice); 2697 } 2698 mutex_unlock(&mutex); 2699 2700 return ret; 2701 } 2702 2703 #ifdef CONFIG_SCHED_DEBUG 2704 void print_rt_stats(struct seq_file *m, int cpu) 2705 { 2706 rt_rq_iter_t iter; 2707 struct rt_rq *rt_rq; 2708 2709 rcu_read_lock(); 2710 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu)) 2711 print_rt_rq(m, cpu, rt_rq); 2712 rcu_read_unlock(); 2713 } 2714 #endif /* CONFIG_SCHED_DEBUG */ 2715