1 /* 2 * Deadline Scheduling Class (SCHED_DEADLINE) 3 * 4 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS). 5 * 6 * Tasks that periodically executes their instances for less than their 7 * runtime won't miss any of their deadlines. 8 * Tasks that are not periodic or sporadic or that tries to execute more 9 * than their reserved bandwidth will be slowed down (and may potentially 10 * miss some of their deadlines), and won't affect any other task. 11 * 12 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>, 13 * Juri Lelli <juri.lelli@gmail.com>, 14 * Michael Trimarchi <michael@amarulasolutions.com>, 15 * Fabio Checconi <fchecconi@gmail.com> 16 */ 17 #include "sched.h" 18 19 #include <linux/slab.h> 20 #include <uapi/linux/sched/types.h> 21 22 struct dl_bandwidth def_dl_bandwidth; 23 24 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se) 25 { 26 return container_of(dl_se, struct task_struct, dl); 27 } 28 29 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq) 30 { 31 return container_of(dl_rq, struct rq, dl); 32 } 33 34 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se) 35 { 36 struct task_struct *p = dl_task_of(dl_se); 37 struct rq *rq = task_rq(p); 38 39 return &rq->dl; 40 } 41 42 static inline int on_dl_rq(struct sched_dl_entity *dl_se) 43 { 44 return !RB_EMPTY_NODE(&dl_se->rb_node); 45 } 46 47 #ifdef CONFIG_SMP 48 static inline struct dl_bw *dl_bw_of(int i) 49 { 50 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), 51 "sched RCU must be held"); 52 return &cpu_rq(i)->rd->dl_bw; 53 } 54 55 static inline int dl_bw_cpus(int i) 56 { 57 struct root_domain *rd = cpu_rq(i)->rd; 58 int cpus = 0; 59 60 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), 61 "sched RCU must be held"); 62 for_each_cpu_and(i, rd->span, cpu_active_mask) 63 cpus++; 64 65 return cpus; 66 } 67 #else 68 static inline struct dl_bw *dl_bw_of(int i) 69 { 70 return &cpu_rq(i)->dl.dl_bw; 71 } 72 73 static inline int dl_bw_cpus(int i) 74 { 75 return 1; 76 } 77 #endif 78 79 static inline 80 void add_running_bw(u64 dl_bw, struct dl_rq *dl_rq) 81 { 82 u64 old = dl_rq->running_bw; 83 84 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); 85 dl_rq->running_bw += dl_bw; 86 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */ 87 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); 88 } 89 90 static inline 91 void sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq) 92 { 93 u64 old = dl_rq->running_bw; 94 95 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); 96 dl_rq->running_bw -= dl_bw; 97 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */ 98 if (dl_rq->running_bw > old) 99 dl_rq->running_bw = 0; 100 } 101 102 static inline 103 void add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) 104 { 105 u64 old = dl_rq->this_bw; 106 107 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); 108 dl_rq->this_bw += dl_bw; 109 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */ 110 } 111 112 static inline 113 void sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq) 114 { 115 u64 old = dl_rq->this_bw; 116 117 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock); 118 dl_rq->this_bw -= dl_bw; 119 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */ 120 if (dl_rq->this_bw > old) 121 dl_rq->this_bw = 0; 122 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw); 123 } 124 125 void dl_change_utilization(struct task_struct *p, u64 new_bw) 126 { 127 struct rq *rq; 128 129 if (task_on_rq_queued(p)) 130 return; 131 132 rq = task_rq(p); 133 if (p->dl.dl_non_contending) { 134 sub_running_bw(p->dl.dl_bw, &rq->dl); 135 p->dl.dl_non_contending = 0; 136 /* 137 * If the timer handler is currently running and the 138 * timer cannot be cancelled, inactive_task_timer() 139 * will see that dl_not_contending is not set, and 140 * will not touch the rq's active utilization, 141 * so we are still safe. 142 */ 143 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) 144 put_task_struct(p); 145 } 146 sub_rq_bw(p->dl.dl_bw, &rq->dl); 147 add_rq_bw(new_bw, &rq->dl); 148 } 149 150 /* 151 * The utilization of a task cannot be immediately removed from 152 * the rq active utilization (running_bw) when the task blocks. 153 * Instead, we have to wait for the so called "0-lag time". 154 * 155 * If a task blocks before the "0-lag time", a timer (the inactive 156 * timer) is armed, and running_bw is decreased when the timer 157 * fires. 158 * 159 * If the task wakes up again before the inactive timer fires, 160 * the timer is cancelled, whereas if the task wakes up after the 161 * inactive timer fired (and running_bw has been decreased) the 162 * task's utilization has to be added to running_bw again. 163 * A flag in the deadline scheduling entity (dl_non_contending) 164 * is used to avoid race conditions between the inactive timer handler 165 * and task wakeups. 166 * 167 * The following diagram shows how running_bw is updated. A task is 168 * "ACTIVE" when its utilization contributes to running_bw; an 169 * "ACTIVE contending" task is in the TASK_RUNNING state, while an 170 * "ACTIVE non contending" task is a blocked task for which the "0-lag time" 171 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag" 172 * time already passed, which does not contribute to running_bw anymore. 173 * +------------------+ 174 * wakeup | ACTIVE | 175 * +------------------>+ contending | 176 * | add_running_bw | | 177 * | +----+------+------+ 178 * | | ^ 179 * | dequeue | | 180 * +--------+-------+ | | 181 * | | t >= 0-lag | | wakeup 182 * | INACTIVE |<---------------+ | 183 * | | sub_running_bw | | 184 * +--------+-------+ | | 185 * ^ | | 186 * | t < 0-lag | | 187 * | | | 188 * | V | 189 * | +----+------+------+ 190 * | sub_running_bw | ACTIVE | 191 * +-------------------+ | 192 * inactive timer | non contending | 193 * fired +------------------+ 194 * 195 * The task_non_contending() function is invoked when a task 196 * blocks, and checks if the 0-lag time already passed or 197 * not (in the first case, it directly updates running_bw; 198 * in the second case, it arms the inactive timer). 199 * 200 * The task_contending() function is invoked when a task wakes 201 * up, and checks if the task is still in the "ACTIVE non contending" 202 * state or not (in the second case, it updates running_bw). 203 */ 204 static void task_non_contending(struct task_struct *p) 205 { 206 struct sched_dl_entity *dl_se = &p->dl; 207 struct hrtimer *timer = &dl_se->inactive_timer; 208 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 209 struct rq *rq = rq_of_dl_rq(dl_rq); 210 s64 zerolag_time; 211 212 /* 213 * If this is a non-deadline task that has been boosted, 214 * do nothing 215 */ 216 if (dl_se->dl_runtime == 0) 217 return; 218 219 WARN_ON(hrtimer_active(&dl_se->inactive_timer)); 220 WARN_ON(dl_se->dl_non_contending); 221 222 zerolag_time = dl_se->deadline - 223 div64_long((dl_se->runtime * dl_se->dl_period), 224 dl_se->dl_runtime); 225 226 /* 227 * Using relative times instead of the absolute "0-lag time" 228 * allows to simplify the code 229 */ 230 zerolag_time -= rq_clock(rq); 231 232 /* 233 * If the "0-lag time" already passed, decrease the active 234 * utilization now, instead of starting a timer 235 */ 236 if (zerolag_time < 0) { 237 if (dl_task(p)) 238 sub_running_bw(dl_se->dl_bw, dl_rq); 239 if (!dl_task(p) || p->state == TASK_DEAD) { 240 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 241 242 if (p->state == TASK_DEAD) 243 sub_rq_bw(p->dl.dl_bw, &rq->dl); 244 raw_spin_lock(&dl_b->lock); 245 __dl_clear(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); 246 __dl_clear_params(p); 247 raw_spin_unlock(&dl_b->lock); 248 } 249 250 return; 251 } 252 253 dl_se->dl_non_contending = 1; 254 get_task_struct(p); 255 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL); 256 } 257 258 static void task_contending(struct sched_dl_entity *dl_se, int flags) 259 { 260 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 261 262 /* 263 * If this is a non-deadline task that has been boosted, 264 * do nothing 265 */ 266 if (dl_se->dl_runtime == 0) 267 return; 268 269 if (flags & ENQUEUE_MIGRATED) 270 add_rq_bw(dl_se->dl_bw, dl_rq); 271 272 if (dl_se->dl_non_contending) { 273 dl_se->dl_non_contending = 0; 274 /* 275 * If the timer handler is currently running and the 276 * timer cannot be cancelled, inactive_task_timer() 277 * will see that dl_not_contending is not set, and 278 * will not touch the rq's active utilization, 279 * so we are still safe. 280 */ 281 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) 282 put_task_struct(dl_task_of(dl_se)); 283 } else { 284 /* 285 * Since "dl_non_contending" is not set, the 286 * task's utilization has already been removed from 287 * active utilization (either when the task blocked, 288 * when the "inactive timer" fired). 289 * So, add it back. 290 */ 291 add_running_bw(dl_se->dl_bw, dl_rq); 292 } 293 } 294 295 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq) 296 { 297 struct sched_dl_entity *dl_se = &p->dl; 298 299 return dl_rq->rb_leftmost == &dl_se->rb_node; 300 } 301 302 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime) 303 { 304 raw_spin_lock_init(&dl_b->dl_runtime_lock); 305 dl_b->dl_period = period; 306 dl_b->dl_runtime = runtime; 307 } 308 309 void init_dl_bw(struct dl_bw *dl_b) 310 { 311 raw_spin_lock_init(&dl_b->lock); 312 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); 313 if (global_rt_runtime() == RUNTIME_INF) 314 dl_b->bw = -1; 315 else 316 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); 317 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); 318 dl_b->total_bw = 0; 319 } 320 321 void init_dl_rq(struct dl_rq *dl_rq) 322 { 323 dl_rq->rb_root = RB_ROOT; 324 325 #ifdef CONFIG_SMP 326 /* zero means no -deadline tasks */ 327 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; 328 329 dl_rq->dl_nr_migratory = 0; 330 dl_rq->overloaded = 0; 331 dl_rq->pushable_dl_tasks_root = RB_ROOT; 332 #else 333 init_dl_bw(&dl_rq->dl_bw); 334 #endif 335 336 dl_rq->running_bw = 0; 337 dl_rq->this_bw = 0; 338 init_dl_rq_bw_ratio(dl_rq); 339 } 340 341 #ifdef CONFIG_SMP 342 343 static inline int dl_overloaded(struct rq *rq) 344 { 345 return atomic_read(&rq->rd->dlo_count); 346 } 347 348 static inline void dl_set_overload(struct rq *rq) 349 { 350 if (!rq->online) 351 return; 352 353 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); 354 /* 355 * Must be visible before the overload count is 356 * set (as in sched_rt.c). 357 * 358 * Matched by the barrier in pull_dl_task(). 359 */ 360 smp_wmb(); 361 atomic_inc(&rq->rd->dlo_count); 362 } 363 364 static inline void dl_clear_overload(struct rq *rq) 365 { 366 if (!rq->online) 367 return; 368 369 atomic_dec(&rq->rd->dlo_count); 370 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); 371 } 372 373 static void update_dl_migration(struct dl_rq *dl_rq) 374 { 375 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { 376 if (!dl_rq->overloaded) { 377 dl_set_overload(rq_of_dl_rq(dl_rq)); 378 dl_rq->overloaded = 1; 379 } 380 } else if (dl_rq->overloaded) { 381 dl_clear_overload(rq_of_dl_rq(dl_rq)); 382 dl_rq->overloaded = 0; 383 } 384 } 385 386 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 387 { 388 struct task_struct *p = dl_task_of(dl_se); 389 390 if (p->nr_cpus_allowed > 1) 391 dl_rq->dl_nr_migratory++; 392 393 update_dl_migration(dl_rq); 394 } 395 396 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 397 { 398 struct task_struct *p = dl_task_of(dl_se); 399 400 if (p->nr_cpus_allowed > 1) 401 dl_rq->dl_nr_migratory--; 402 403 update_dl_migration(dl_rq); 404 } 405 406 /* 407 * The list of pushable -deadline task is not a plist, like in 408 * sched_rt.c, it is an rb-tree with tasks ordered by deadline. 409 */ 410 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) 411 { 412 struct dl_rq *dl_rq = &rq->dl; 413 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node; 414 struct rb_node *parent = NULL; 415 struct task_struct *entry; 416 int leftmost = 1; 417 418 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); 419 420 while (*link) { 421 parent = *link; 422 entry = rb_entry(parent, struct task_struct, 423 pushable_dl_tasks); 424 if (dl_entity_preempt(&p->dl, &entry->dl)) 425 link = &parent->rb_left; 426 else { 427 link = &parent->rb_right; 428 leftmost = 0; 429 } 430 } 431 432 if (leftmost) { 433 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks; 434 dl_rq->earliest_dl.next = p->dl.deadline; 435 } 436 437 rb_link_node(&p->pushable_dl_tasks, parent, link); 438 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); 439 } 440 441 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) 442 { 443 struct dl_rq *dl_rq = &rq->dl; 444 445 if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) 446 return; 447 448 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) { 449 struct rb_node *next_node; 450 451 next_node = rb_next(&p->pushable_dl_tasks); 452 dl_rq->pushable_dl_tasks_leftmost = next_node; 453 if (next_node) { 454 dl_rq->earliest_dl.next = rb_entry(next_node, 455 struct task_struct, pushable_dl_tasks)->dl.deadline; 456 } 457 } 458 459 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); 460 RB_CLEAR_NODE(&p->pushable_dl_tasks); 461 } 462 463 static inline int has_pushable_dl_tasks(struct rq *rq) 464 { 465 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root); 466 } 467 468 static int push_dl_task(struct rq *rq); 469 470 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) 471 { 472 return dl_task(prev); 473 } 474 475 static DEFINE_PER_CPU(struct callback_head, dl_push_head); 476 static DEFINE_PER_CPU(struct callback_head, dl_pull_head); 477 478 static void push_dl_tasks(struct rq *); 479 static void pull_dl_task(struct rq *); 480 481 static inline void queue_push_tasks(struct rq *rq) 482 { 483 if (!has_pushable_dl_tasks(rq)) 484 return; 485 486 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks); 487 } 488 489 static inline void queue_pull_task(struct rq *rq) 490 { 491 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task); 492 } 493 494 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq); 495 496 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p) 497 { 498 struct rq *later_rq = NULL; 499 500 later_rq = find_lock_later_rq(p, rq); 501 if (!later_rq) { 502 int cpu; 503 504 /* 505 * If we cannot preempt any rq, fall back to pick any 506 * online cpu. 507 */ 508 cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed); 509 if (cpu >= nr_cpu_ids) { 510 /* 511 * Fail to find any suitable cpu. 512 * The task will never come back! 513 */ 514 BUG_ON(dl_bandwidth_enabled()); 515 516 /* 517 * If admission control is disabled we 518 * try a little harder to let the task 519 * run. 520 */ 521 cpu = cpumask_any(cpu_active_mask); 522 } 523 later_rq = cpu_rq(cpu); 524 double_lock_balance(rq, later_rq); 525 } 526 527 set_task_cpu(p, later_rq->cpu); 528 double_unlock_balance(later_rq, rq); 529 530 return later_rq; 531 } 532 533 #else 534 535 static inline 536 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) 537 { 538 } 539 540 static inline 541 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) 542 { 543 } 544 545 static inline 546 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 547 { 548 } 549 550 static inline 551 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 552 { 553 } 554 555 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) 556 { 557 return false; 558 } 559 560 static inline void pull_dl_task(struct rq *rq) 561 { 562 } 563 564 static inline void queue_push_tasks(struct rq *rq) 565 { 566 } 567 568 static inline void queue_pull_task(struct rq *rq) 569 { 570 } 571 #endif /* CONFIG_SMP */ 572 573 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); 574 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); 575 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, 576 int flags); 577 578 /* 579 * We are being explicitly informed that a new instance is starting, 580 * and this means that: 581 * - the absolute deadline of the entity has to be placed at 582 * current time + relative deadline; 583 * - the runtime of the entity has to be set to the maximum value. 584 * 585 * The capability of specifying such event is useful whenever a -deadline 586 * entity wants to (try to!) synchronize its behaviour with the scheduler's 587 * one, and to (try to!) reconcile itself with its own scheduling 588 * parameters. 589 */ 590 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se) 591 { 592 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 593 struct rq *rq = rq_of_dl_rq(dl_rq); 594 595 WARN_ON(dl_se->dl_boosted); 596 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline)); 597 598 /* 599 * We are racing with the deadline timer. So, do nothing because 600 * the deadline timer handler will take care of properly recharging 601 * the runtime and postponing the deadline 602 */ 603 if (dl_se->dl_throttled) 604 return; 605 606 /* 607 * We use the regular wall clock time to set deadlines in the 608 * future; in fact, we must consider execution overheads (time 609 * spent on hardirq context, etc.). 610 */ 611 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline; 612 dl_se->runtime = dl_se->dl_runtime; 613 } 614 615 /* 616 * Pure Earliest Deadline First (EDF) scheduling does not deal with the 617 * possibility of a entity lasting more than what it declared, and thus 618 * exhausting its runtime. 619 * 620 * Here we are interested in making runtime overrun possible, but we do 621 * not want a entity which is misbehaving to affect the scheduling of all 622 * other entities. 623 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) 624 * is used, in order to confine each entity within its own bandwidth. 625 * 626 * This function deals exactly with that, and ensures that when the runtime 627 * of a entity is replenished, its deadline is also postponed. That ensures 628 * the overrunning entity can't interfere with other entity in the system and 629 * can't make them miss their deadlines. Reasons why this kind of overruns 630 * could happen are, typically, a entity voluntarily trying to overcome its 631 * runtime, or it just underestimated it during sched_setattr(). 632 */ 633 static void replenish_dl_entity(struct sched_dl_entity *dl_se, 634 struct sched_dl_entity *pi_se) 635 { 636 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 637 struct rq *rq = rq_of_dl_rq(dl_rq); 638 639 BUG_ON(pi_se->dl_runtime <= 0); 640 641 /* 642 * This could be the case for a !-dl task that is boosted. 643 * Just go with full inherited parameters. 644 */ 645 if (dl_se->dl_deadline == 0) { 646 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; 647 dl_se->runtime = pi_se->dl_runtime; 648 } 649 650 if (dl_se->dl_yielded && dl_se->runtime > 0) 651 dl_se->runtime = 0; 652 653 /* 654 * We keep moving the deadline away until we get some 655 * available runtime for the entity. This ensures correct 656 * handling of situations where the runtime overrun is 657 * arbitrary large. 658 */ 659 while (dl_se->runtime <= 0) { 660 dl_se->deadline += pi_se->dl_period; 661 dl_se->runtime += pi_se->dl_runtime; 662 } 663 664 /* 665 * At this point, the deadline really should be "in 666 * the future" with respect to rq->clock. If it's 667 * not, we are, for some reason, lagging too much! 668 * Anyway, after having warn userspace abut that, 669 * we still try to keep the things running by 670 * resetting the deadline and the budget of the 671 * entity. 672 */ 673 if (dl_time_before(dl_se->deadline, rq_clock(rq))) { 674 printk_deferred_once("sched: DL replenish lagged too much\n"); 675 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; 676 dl_se->runtime = pi_se->dl_runtime; 677 } 678 679 if (dl_se->dl_yielded) 680 dl_se->dl_yielded = 0; 681 if (dl_se->dl_throttled) 682 dl_se->dl_throttled = 0; 683 } 684 685 /* 686 * Here we check if --at time t-- an entity (which is probably being 687 * [re]activated or, in general, enqueued) can use its remaining runtime 688 * and its current deadline _without_ exceeding the bandwidth it is 689 * assigned (function returns true if it can't). We are in fact applying 690 * one of the CBS rules: when a task wakes up, if the residual runtime 691 * over residual deadline fits within the allocated bandwidth, then we 692 * can keep the current (absolute) deadline and residual budget without 693 * disrupting the schedulability of the system. Otherwise, we should 694 * refill the runtime and set the deadline a period in the future, 695 * because keeping the current (absolute) deadline of the task would 696 * result in breaking guarantees promised to other tasks (refer to 697 * Documentation/scheduler/sched-deadline.txt for more informations). 698 * 699 * This function returns true if: 700 * 701 * runtime / (deadline - t) > dl_runtime / dl_deadline , 702 * 703 * IOW we can't recycle current parameters. 704 * 705 * Notice that the bandwidth check is done against the deadline. For 706 * task with deadline equal to period this is the same of using 707 * dl_period instead of dl_deadline in the equation above. 708 */ 709 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, 710 struct sched_dl_entity *pi_se, u64 t) 711 { 712 u64 left, right; 713 714 /* 715 * left and right are the two sides of the equation above, 716 * after a bit of shuffling to use multiplications instead 717 * of divisions. 718 * 719 * Note that none of the time values involved in the two 720 * multiplications are absolute: dl_deadline and dl_runtime 721 * are the relative deadline and the maximum runtime of each 722 * instance, runtime is the runtime left for the last instance 723 * and (deadline - t), since t is rq->clock, is the time left 724 * to the (absolute) deadline. Even if overflowing the u64 type 725 * is very unlikely to occur in both cases, here we scale down 726 * as we want to avoid that risk at all. Scaling down by 10 727 * means that we reduce granularity to 1us. We are fine with it, 728 * since this is only a true/false check and, anyway, thinking 729 * of anything below microseconds resolution is actually fiction 730 * (but still we want to give the user that illusion >;). 731 */ 732 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); 733 right = ((dl_se->deadline - t) >> DL_SCALE) * 734 (pi_se->dl_runtime >> DL_SCALE); 735 736 return dl_time_before(right, left); 737 } 738 739 /* 740 * Revised wakeup rule [1]: For self-suspending tasks, rather then 741 * re-initializing task's runtime and deadline, the revised wakeup 742 * rule adjusts the task's runtime to avoid the task to overrun its 743 * density. 744 * 745 * Reasoning: a task may overrun the density if: 746 * runtime / (deadline - t) > dl_runtime / dl_deadline 747 * 748 * Therefore, runtime can be adjusted to: 749 * runtime = (dl_runtime / dl_deadline) * (deadline - t) 750 * 751 * In such way that runtime will be equal to the maximum density 752 * the task can use without breaking any rule. 753 * 754 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant 755 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24. 756 */ 757 static void 758 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq) 759 { 760 u64 laxity = dl_se->deadline - rq_clock(rq); 761 762 /* 763 * If the task has deadline < period, and the deadline is in the past, 764 * it should already be throttled before this check. 765 * 766 * See update_dl_entity() comments for further details. 767 */ 768 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq))); 769 770 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT; 771 } 772 773 /* 774 * Regarding the deadline, a task with implicit deadline has a relative 775 * deadline == relative period. A task with constrained deadline has a 776 * relative deadline <= relative period. 777 * 778 * We support constrained deadline tasks. However, there are some restrictions 779 * applied only for tasks which do not have an implicit deadline. See 780 * update_dl_entity() to know more about such restrictions. 781 * 782 * The dl_is_implicit() returns true if the task has an implicit deadline. 783 */ 784 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se) 785 { 786 return dl_se->dl_deadline == dl_se->dl_period; 787 } 788 789 /* 790 * When a deadline entity is placed in the runqueue, its runtime and deadline 791 * might need to be updated. This is done by a CBS wake up rule. There are two 792 * different rules: 1) the original CBS; and 2) the Revisited CBS. 793 * 794 * When the task is starting a new period, the Original CBS is used. In this 795 * case, the runtime is replenished and a new absolute deadline is set. 796 * 797 * When a task is queued before the begin of the next period, using the 798 * remaining runtime and deadline could make the entity to overflow, see 799 * dl_entity_overflow() to find more about runtime overflow. When such case 800 * is detected, the runtime and deadline need to be updated. 801 * 802 * If the task has an implicit deadline, i.e., deadline == period, the Original 803 * CBS is applied. the runtime is replenished and a new absolute deadline is 804 * set, as in the previous cases. 805 * 806 * However, the Original CBS does not work properly for tasks with 807 * deadline < period, which are said to have a constrained deadline. By 808 * applying the Original CBS, a constrained deadline task would be able to run 809 * runtime/deadline in a period. With deadline < period, the task would 810 * overrun the runtime/period allowed bandwidth, breaking the admission test. 811 * 812 * In order to prevent this misbehave, the Revisited CBS is used for 813 * constrained deadline tasks when a runtime overflow is detected. In the 814 * Revisited CBS, rather than replenishing & setting a new absolute deadline, 815 * the remaining runtime of the task is reduced to avoid runtime overflow. 816 * Please refer to the comments update_dl_revised_wakeup() function to find 817 * more about the Revised CBS rule. 818 */ 819 static void update_dl_entity(struct sched_dl_entity *dl_se, 820 struct sched_dl_entity *pi_se) 821 { 822 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 823 struct rq *rq = rq_of_dl_rq(dl_rq); 824 825 if (dl_time_before(dl_se->deadline, rq_clock(rq)) || 826 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { 827 828 if (unlikely(!dl_is_implicit(dl_se) && 829 !dl_time_before(dl_se->deadline, rq_clock(rq)) && 830 !dl_se->dl_boosted)){ 831 update_dl_revised_wakeup(dl_se, rq); 832 return; 833 } 834 835 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; 836 dl_se->runtime = pi_se->dl_runtime; 837 } 838 } 839 840 static inline u64 dl_next_period(struct sched_dl_entity *dl_se) 841 { 842 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period; 843 } 844 845 /* 846 * If the entity depleted all its runtime, and if we want it to sleep 847 * while waiting for some new execution time to become available, we 848 * set the bandwidth replenishment timer to the replenishment instant 849 * and try to activate it. 850 * 851 * Notice that it is important for the caller to know if the timer 852 * actually started or not (i.e., the replenishment instant is in 853 * the future or in the past). 854 */ 855 static int start_dl_timer(struct task_struct *p) 856 { 857 struct sched_dl_entity *dl_se = &p->dl; 858 struct hrtimer *timer = &dl_se->dl_timer; 859 struct rq *rq = task_rq(p); 860 ktime_t now, act; 861 s64 delta; 862 863 lockdep_assert_held(&rq->lock); 864 865 /* 866 * We want the timer to fire at the deadline, but considering 867 * that it is actually coming from rq->clock and not from 868 * hrtimer's time base reading. 869 */ 870 act = ns_to_ktime(dl_next_period(dl_se)); 871 now = hrtimer_cb_get_time(timer); 872 delta = ktime_to_ns(now) - rq_clock(rq); 873 act = ktime_add_ns(act, delta); 874 875 /* 876 * If the expiry time already passed, e.g., because the value 877 * chosen as the deadline is too small, don't even try to 878 * start the timer in the past! 879 */ 880 if (ktime_us_delta(act, now) < 0) 881 return 0; 882 883 /* 884 * !enqueued will guarantee another callback; even if one is already in 885 * progress. This ensures a balanced {get,put}_task_struct(). 886 * 887 * The race against __run_timer() clearing the enqueued state is 888 * harmless because we're holding task_rq()->lock, therefore the timer 889 * expiring after we've done the check will wait on its task_rq_lock() 890 * and observe our state. 891 */ 892 if (!hrtimer_is_queued(timer)) { 893 get_task_struct(p); 894 hrtimer_start(timer, act, HRTIMER_MODE_ABS); 895 } 896 897 return 1; 898 } 899 900 /* 901 * This is the bandwidth enforcement timer callback. If here, we know 902 * a task is not on its dl_rq, since the fact that the timer was running 903 * means the task is throttled and needs a runtime replenishment. 904 * 905 * However, what we actually do depends on the fact the task is active, 906 * (it is on its rq) or has been removed from there by a call to 907 * dequeue_task_dl(). In the former case we must issue the runtime 908 * replenishment and add the task back to the dl_rq; in the latter, we just 909 * do nothing but clearing dl_throttled, so that runtime and deadline 910 * updating (and the queueing back to dl_rq) will be done by the 911 * next call to enqueue_task_dl(). 912 */ 913 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) 914 { 915 struct sched_dl_entity *dl_se = container_of(timer, 916 struct sched_dl_entity, 917 dl_timer); 918 struct task_struct *p = dl_task_of(dl_se); 919 struct rq_flags rf; 920 struct rq *rq; 921 922 rq = task_rq_lock(p, &rf); 923 924 /* 925 * The task might have changed its scheduling policy to something 926 * different than SCHED_DEADLINE (through switched_from_dl()). 927 */ 928 if (!dl_task(p)) 929 goto unlock; 930 931 /* 932 * The task might have been boosted by someone else and might be in the 933 * boosting/deboosting path, its not throttled. 934 */ 935 if (dl_se->dl_boosted) 936 goto unlock; 937 938 /* 939 * Spurious timer due to start_dl_timer() race; or we already received 940 * a replenishment from rt_mutex_setprio(). 941 */ 942 if (!dl_se->dl_throttled) 943 goto unlock; 944 945 sched_clock_tick(); 946 update_rq_clock(rq); 947 948 /* 949 * If the throttle happened during sched-out; like: 950 * 951 * schedule() 952 * deactivate_task() 953 * dequeue_task_dl() 954 * update_curr_dl() 955 * start_dl_timer() 956 * __dequeue_task_dl() 957 * prev->on_rq = 0; 958 * 959 * We can be both throttled and !queued. Replenish the counter 960 * but do not enqueue -- wait for our wakeup to do that. 961 */ 962 if (!task_on_rq_queued(p)) { 963 replenish_dl_entity(dl_se, dl_se); 964 goto unlock; 965 } 966 967 #ifdef CONFIG_SMP 968 if (unlikely(!rq->online)) { 969 /* 970 * If the runqueue is no longer available, migrate the 971 * task elsewhere. This necessarily changes rq. 972 */ 973 lockdep_unpin_lock(&rq->lock, rf.cookie); 974 rq = dl_task_offline_migration(rq, p); 975 rf.cookie = lockdep_pin_lock(&rq->lock); 976 update_rq_clock(rq); 977 978 /* 979 * Now that the task has been migrated to the new RQ and we 980 * have that locked, proceed as normal and enqueue the task 981 * there. 982 */ 983 } 984 #endif 985 986 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); 987 if (dl_task(rq->curr)) 988 check_preempt_curr_dl(rq, p, 0); 989 else 990 resched_curr(rq); 991 992 #ifdef CONFIG_SMP 993 /* 994 * Queueing this task back might have overloaded rq, check if we need 995 * to kick someone away. 996 */ 997 if (has_pushable_dl_tasks(rq)) { 998 /* 999 * Nothing relies on rq->lock after this, so its safe to drop 1000 * rq->lock. 1001 */ 1002 rq_unpin_lock(rq, &rf); 1003 push_dl_task(rq); 1004 rq_repin_lock(rq, &rf); 1005 } 1006 #endif 1007 1008 unlock: 1009 task_rq_unlock(rq, p, &rf); 1010 1011 /* 1012 * This can free the task_struct, including this hrtimer, do not touch 1013 * anything related to that after this. 1014 */ 1015 put_task_struct(p); 1016 1017 return HRTIMER_NORESTART; 1018 } 1019 1020 void init_dl_task_timer(struct sched_dl_entity *dl_se) 1021 { 1022 struct hrtimer *timer = &dl_se->dl_timer; 1023 1024 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1025 timer->function = dl_task_timer; 1026 } 1027 1028 /* 1029 * During the activation, CBS checks if it can reuse the current task's 1030 * runtime and period. If the deadline of the task is in the past, CBS 1031 * cannot use the runtime, and so it replenishes the task. This rule 1032 * works fine for implicit deadline tasks (deadline == period), and the 1033 * CBS was designed for implicit deadline tasks. However, a task with 1034 * constrained deadline (deadine < period) might be awakened after the 1035 * deadline, but before the next period. In this case, replenishing the 1036 * task would allow it to run for runtime / deadline. As in this case 1037 * deadline < period, CBS enables a task to run for more than the 1038 * runtime / period. In a very loaded system, this can cause a domino 1039 * effect, making other tasks miss their deadlines. 1040 * 1041 * To avoid this problem, in the activation of a constrained deadline 1042 * task after the deadline but before the next period, throttle the 1043 * task and set the replenishing timer to the begin of the next period, 1044 * unless it is boosted. 1045 */ 1046 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se) 1047 { 1048 struct task_struct *p = dl_task_of(dl_se); 1049 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se)); 1050 1051 if (dl_time_before(dl_se->deadline, rq_clock(rq)) && 1052 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) { 1053 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p))) 1054 return; 1055 dl_se->dl_throttled = 1; 1056 if (dl_se->runtime > 0) 1057 dl_se->runtime = 0; 1058 } 1059 } 1060 1061 static 1062 int dl_runtime_exceeded(struct sched_dl_entity *dl_se) 1063 { 1064 return (dl_se->runtime <= 0); 1065 } 1066 1067 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); 1068 1069 /* 1070 * This function implements the GRUB accounting rule: 1071 * according to the GRUB reclaiming algorithm, the runtime is 1072 * not decreased as "dq = -dt", but as 1073 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt", 1074 * where u is the utilization of the task, Umax is the maximum reclaimable 1075 * utilization, Uinact is the (per-runqueue) inactive utilization, computed 1076 * as the difference between the "total runqueue utilization" and the 1077 * runqueue active utilization, and Uextra is the (per runqueue) extra 1078 * reclaimable utilization. 1079 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations 1080 * multiplied by 2^BW_SHIFT, the result has to be shifted right by 1081 * BW_SHIFT. 1082 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT, 1083 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT. 1084 * Since delta is a 64 bit variable, to have an overflow its value 1085 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds. 1086 * So, overflow is not an issue here. 1087 */ 1088 u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se) 1089 { 1090 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */ 1091 u64 u_act; 1092 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT; 1093 1094 /* 1095 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)}, 1096 * we compare u_inact + rq->dl.extra_bw with 1097 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because 1098 * u_inact + rq->dl.extra_bw can be larger than 1099 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative 1100 * leading to wrong results) 1101 */ 1102 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min) 1103 u_act = u_act_min; 1104 else 1105 u_act = BW_UNIT - u_inact - rq->dl.extra_bw; 1106 1107 return (delta * u_act) >> BW_SHIFT; 1108 } 1109 1110 /* 1111 * Update the current task's runtime statistics (provided it is still 1112 * a -deadline task and has not been removed from the dl_rq). 1113 */ 1114 static void update_curr_dl(struct rq *rq) 1115 { 1116 struct task_struct *curr = rq->curr; 1117 struct sched_dl_entity *dl_se = &curr->dl; 1118 u64 delta_exec; 1119 1120 if (!dl_task(curr) || !on_dl_rq(dl_se)) 1121 return; 1122 1123 /* 1124 * Consumed budget is computed considering the time as 1125 * observed by schedulable tasks (excluding time spent 1126 * in hardirq context, etc.). Deadlines are instead 1127 * computed using hard walltime. This seems to be the more 1128 * natural solution, but the full ramifications of this 1129 * approach need further study. 1130 */ 1131 delta_exec = rq_clock_task(rq) - curr->se.exec_start; 1132 if (unlikely((s64)delta_exec <= 0)) { 1133 if (unlikely(dl_se->dl_yielded)) 1134 goto throttle; 1135 return; 1136 } 1137 1138 /* kick cpufreq (see the comment in kernel/sched/sched.h). */ 1139 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_DL); 1140 1141 schedstat_set(curr->se.statistics.exec_max, 1142 max(curr->se.statistics.exec_max, delta_exec)); 1143 1144 curr->se.sum_exec_runtime += delta_exec; 1145 account_group_exec_runtime(curr, delta_exec); 1146 1147 curr->se.exec_start = rq_clock_task(rq); 1148 cpuacct_charge(curr, delta_exec); 1149 1150 sched_rt_avg_update(rq, delta_exec); 1151 1152 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) 1153 delta_exec = grub_reclaim(delta_exec, rq, &curr->dl); 1154 dl_se->runtime -= delta_exec; 1155 1156 throttle: 1157 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) { 1158 dl_se->dl_throttled = 1; 1159 __dequeue_task_dl(rq, curr, 0); 1160 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr))) 1161 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); 1162 1163 if (!is_leftmost(curr, &rq->dl)) 1164 resched_curr(rq); 1165 } 1166 1167 /* 1168 * Because -- for now -- we share the rt bandwidth, we need to 1169 * account our runtime there too, otherwise actual rt tasks 1170 * would be able to exceed the shared quota. 1171 * 1172 * Account to the root rt group for now. 1173 * 1174 * The solution we're working towards is having the RT groups scheduled 1175 * using deadline servers -- however there's a few nasties to figure 1176 * out before that can happen. 1177 */ 1178 if (rt_bandwidth_enabled()) { 1179 struct rt_rq *rt_rq = &rq->rt; 1180 1181 raw_spin_lock(&rt_rq->rt_runtime_lock); 1182 /* 1183 * We'll let actual RT tasks worry about the overflow here, we 1184 * have our own CBS to keep us inline; only account when RT 1185 * bandwidth is relevant. 1186 */ 1187 if (sched_rt_bandwidth_account(rt_rq)) 1188 rt_rq->rt_time += delta_exec; 1189 raw_spin_unlock(&rt_rq->rt_runtime_lock); 1190 } 1191 } 1192 1193 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer) 1194 { 1195 struct sched_dl_entity *dl_se = container_of(timer, 1196 struct sched_dl_entity, 1197 inactive_timer); 1198 struct task_struct *p = dl_task_of(dl_se); 1199 struct rq_flags rf; 1200 struct rq *rq; 1201 1202 rq = task_rq_lock(p, &rf); 1203 1204 if (!dl_task(p) || p->state == TASK_DEAD) { 1205 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1206 1207 if (p->state == TASK_DEAD && dl_se->dl_non_contending) { 1208 sub_running_bw(p->dl.dl_bw, dl_rq_of_se(&p->dl)); 1209 sub_rq_bw(p->dl.dl_bw, dl_rq_of_se(&p->dl)); 1210 dl_se->dl_non_contending = 0; 1211 } 1212 1213 raw_spin_lock(&dl_b->lock); 1214 __dl_clear(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); 1215 raw_spin_unlock(&dl_b->lock); 1216 __dl_clear_params(p); 1217 1218 goto unlock; 1219 } 1220 if (dl_se->dl_non_contending == 0) 1221 goto unlock; 1222 1223 sched_clock_tick(); 1224 update_rq_clock(rq); 1225 1226 sub_running_bw(dl_se->dl_bw, &rq->dl); 1227 dl_se->dl_non_contending = 0; 1228 unlock: 1229 task_rq_unlock(rq, p, &rf); 1230 put_task_struct(p); 1231 1232 return HRTIMER_NORESTART; 1233 } 1234 1235 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se) 1236 { 1237 struct hrtimer *timer = &dl_se->inactive_timer; 1238 1239 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1240 timer->function = inactive_task_timer; 1241 } 1242 1243 #ifdef CONFIG_SMP 1244 1245 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) 1246 { 1247 struct rq *rq = rq_of_dl_rq(dl_rq); 1248 1249 if (dl_rq->earliest_dl.curr == 0 || 1250 dl_time_before(deadline, dl_rq->earliest_dl.curr)) { 1251 dl_rq->earliest_dl.curr = deadline; 1252 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline); 1253 } 1254 } 1255 1256 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) 1257 { 1258 struct rq *rq = rq_of_dl_rq(dl_rq); 1259 1260 /* 1261 * Since we may have removed our earliest (and/or next earliest) 1262 * task we must recompute them. 1263 */ 1264 if (!dl_rq->dl_nr_running) { 1265 dl_rq->earliest_dl.curr = 0; 1266 dl_rq->earliest_dl.next = 0; 1267 cpudl_clear(&rq->rd->cpudl, rq->cpu); 1268 } else { 1269 struct rb_node *leftmost = dl_rq->rb_leftmost; 1270 struct sched_dl_entity *entry; 1271 1272 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); 1273 dl_rq->earliest_dl.curr = entry->deadline; 1274 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline); 1275 } 1276 } 1277 1278 #else 1279 1280 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} 1281 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} 1282 1283 #endif /* CONFIG_SMP */ 1284 1285 static inline 1286 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 1287 { 1288 int prio = dl_task_of(dl_se)->prio; 1289 u64 deadline = dl_se->deadline; 1290 1291 WARN_ON(!dl_prio(prio)); 1292 dl_rq->dl_nr_running++; 1293 add_nr_running(rq_of_dl_rq(dl_rq), 1); 1294 1295 inc_dl_deadline(dl_rq, deadline); 1296 inc_dl_migration(dl_se, dl_rq); 1297 } 1298 1299 static inline 1300 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 1301 { 1302 int prio = dl_task_of(dl_se)->prio; 1303 1304 WARN_ON(!dl_prio(prio)); 1305 WARN_ON(!dl_rq->dl_nr_running); 1306 dl_rq->dl_nr_running--; 1307 sub_nr_running(rq_of_dl_rq(dl_rq), 1); 1308 1309 dec_dl_deadline(dl_rq, dl_se->deadline); 1310 dec_dl_migration(dl_se, dl_rq); 1311 } 1312 1313 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) 1314 { 1315 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 1316 struct rb_node **link = &dl_rq->rb_root.rb_node; 1317 struct rb_node *parent = NULL; 1318 struct sched_dl_entity *entry; 1319 int leftmost = 1; 1320 1321 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); 1322 1323 while (*link) { 1324 parent = *link; 1325 entry = rb_entry(parent, struct sched_dl_entity, rb_node); 1326 if (dl_time_before(dl_se->deadline, entry->deadline)) 1327 link = &parent->rb_left; 1328 else { 1329 link = &parent->rb_right; 1330 leftmost = 0; 1331 } 1332 } 1333 1334 if (leftmost) 1335 dl_rq->rb_leftmost = &dl_se->rb_node; 1336 1337 rb_link_node(&dl_se->rb_node, parent, link); 1338 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root); 1339 1340 inc_dl_tasks(dl_se, dl_rq); 1341 } 1342 1343 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) 1344 { 1345 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 1346 1347 if (RB_EMPTY_NODE(&dl_se->rb_node)) 1348 return; 1349 1350 if (dl_rq->rb_leftmost == &dl_se->rb_node) { 1351 struct rb_node *next_node; 1352 1353 next_node = rb_next(&dl_se->rb_node); 1354 dl_rq->rb_leftmost = next_node; 1355 } 1356 1357 rb_erase(&dl_se->rb_node, &dl_rq->rb_root); 1358 RB_CLEAR_NODE(&dl_se->rb_node); 1359 1360 dec_dl_tasks(dl_se, dl_rq); 1361 } 1362 1363 static void 1364 enqueue_dl_entity(struct sched_dl_entity *dl_se, 1365 struct sched_dl_entity *pi_se, int flags) 1366 { 1367 BUG_ON(on_dl_rq(dl_se)); 1368 1369 /* 1370 * If this is a wakeup or a new instance, the scheduling 1371 * parameters of the task might need updating. Otherwise, 1372 * we want a replenishment of its runtime. 1373 */ 1374 if (flags & ENQUEUE_WAKEUP) { 1375 task_contending(dl_se, flags); 1376 update_dl_entity(dl_se, pi_se); 1377 } else if (flags & ENQUEUE_REPLENISH) { 1378 replenish_dl_entity(dl_se, pi_se); 1379 } 1380 1381 __enqueue_dl_entity(dl_se); 1382 } 1383 1384 static void dequeue_dl_entity(struct sched_dl_entity *dl_se) 1385 { 1386 __dequeue_dl_entity(dl_se); 1387 } 1388 1389 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) 1390 { 1391 struct task_struct *pi_task = rt_mutex_get_top_task(p); 1392 struct sched_dl_entity *pi_se = &p->dl; 1393 1394 /* 1395 * Use the scheduling parameters of the top pi-waiter task if: 1396 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND 1397 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is 1398 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting 1399 * boosted due to a SCHED_DEADLINE pi-waiter). 1400 * Otherwise we keep our runtime and deadline. 1401 */ 1402 if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) { 1403 pi_se = &pi_task->dl; 1404 } else if (!dl_prio(p->normal_prio)) { 1405 /* 1406 * Special case in which we have a !SCHED_DEADLINE task 1407 * that is going to be deboosted, but exceeds its 1408 * runtime while doing so. No point in replenishing 1409 * it, as it's going to return back to its original 1410 * scheduling class after this. 1411 */ 1412 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); 1413 return; 1414 } 1415 1416 /* 1417 * Check if a constrained deadline task was activated 1418 * after the deadline but before the next period. 1419 * If that is the case, the task will be throttled and 1420 * the replenishment timer will be set to the next period. 1421 */ 1422 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl)) 1423 dl_check_constrained_dl(&p->dl); 1424 1425 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) { 1426 add_rq_bw(p->dl.dl_bw, &rq->dl); 1427 add_running_bw(p->dl.dl_bw, &rq->dl); 1428 } 1429 1430 /* 1431 * If p is throttled, we do not enqueue it. In fact, if it exhausted 1432 * its budget it needs a replenishment and, since it now is on 1433 * its rq, the bandwidth timer callback (which clearly has not 1434 * run yet) will take care of this. 1435 * However, the active utilization does not depend on the fact 1436 * that the task is on the runqueue or not (but depends on the 1437 * task's state - in GRUB parlance, "inactive" vs "active contending"). 1438 * In other words, even if a task is throttled its utilization must 1439 * be counted in the active utilization; hence, we need to call 1440 * add_running_bw(). 1441 */ 1442 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) { 1443 if (flags & ENQUEUE_WAKEUP) 1444 task_contending(&p->dl, flags); 1445 1446 return; 1447 } 1448 1449 enqueue_dl_entity(&p->dl, pi_se, flags); 1450 1451 if (!task_current(rq, p) && p->nr_cpus_allowed > 1) 1452 enqueue_pushable_dl_task(rq, p); 1453 } 1454 1455 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) 1456 { 1457 dequeue_dl_entity(&p->dl); 1458 dequeue_pushable_dl_task(rq, p); 1459 } 1460 1461 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) 1462 { 1463 update_curr_dl(rq); 1464 __dequeue_task_dl(rq, p, flags); 1465 1466 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) { 1467 sub_running_bw(p->dl.dl_bw, &rq->dl); 1468 sub_rq_bw(p->dl.dl_bw, &rq->dl); 1469 } 1470 1471 /* 1472 * This check allows to start the inactive timer (or to immediately 1473 * decrease the active utilization, if needed) in two cases: 1474 * when the task blocks and when it is terminating 1475 * (p->state == TASK_DEAD). We can handle the two cases in the same 1476 * way, because from GRUB's point of view the same thing is happening 1477 * (the task moves from "active contending" to "active non contending" 1478 * or "inactive") 1479 */ 1480 if (flags & DEQUEUE_SLEEP) 1481 task_non_contending(p); 1482 } 1483 1484 /* 1485 * Yield task semantic for -deadline tasks is: 1486 * 1487 * get off from the CPU until our next instance, with 1488 * a new runtime. This is of little use now, since we 1489 * don't have a bandwidth reclaiming mechanism. Anyway, 1490 * bandwidth reclaiming is planned for the future, and 1491 * yield_task_dl will indicate that some spare budget 1492 * is available for other task instances to use it. 1493 */ 1494 static void yield_task_dl(struct rq *rq) 1495 { 1496 /* 1497 * We make the task go to sleep until its current deadline by 1498 * forcing its runtime to zero. This way, update_curr_dl() stops 1499 * it and the bandwidth timer will wake it up and will give it 1500 * new scheduling parameters (thanks to dl_yielded=1). 1501 */ 1502 rq->curr->dl.dl_yielded = 1; 1503 1504 update_rq_clock(rq); 1505 update_curr_dl(rq); 1506 /* 1507 * Tell update_rq_clock() that we've just updated, 1508 * so we don't do microscopic update in schedule() 1509 * and double the fastpath cost. 1510 */ 1511 rq_clock_skip_update(rq, true); 1512 } 1513 1514 #ifdef CONFIG_SMP 1515 1516 static int find_later_rq(struct task_struct *task); 1517 1518 static int 1519 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) 1520 { 1521 struct task_struct *curr; 1522 struct rq *rq; 1523 1524 if (sd_flag != SD_BALANCE_WAKE) 1525 goto out; 1526 1527 rq = cpu_rq(cpu); 1528 1529 rcu_read_lock(); 1530 curr = READ_ONCE(rq->curr); /* unlocked access */ 1531 1532 /* 1533 * If we are dealing with a -deadline task, we must 1534 * decide where to wake it up. 1535 * If it has a later deadline and the current task 1536 * on this rq can't move (provided the waking task 1537 * can!) we prefer to send it somewhere else. On the 1538 * other hand, if it has a shorter deadline, we 1539 * try to make it stay here, it might be important. 1540 */ 1541 if (unlikely(dl_task(curr)) && 1542 (curr->nr_cpus_allowed < 2 || 1543 !dl_entity_preempt(&p->dl, &curr->dl)) && 1544 (p->nr_cpus_allowed > 1)) { 1545 int target = find_later_rq(p); 1546 1547 if (target != -1 && 1548 (dl_time_before(p->dl.deadline, 1549 cpu_rq(target)->dl.earliest_dl.curr) || 1550 (cpu_rq(target)->dl.dl_nr_running == 0))) 1551 cpu = target; 1552 } 1553 rcu_read_unlock(); 1554 1555 out: 1556 return cpu; 1557 } 1558 1559 static void migrate_task_rq_dl(struct task_struct *p) 1560 { 1561 struct rq *rq; 1562 1563 if (p->state != TASK_WAKING) 1564 return; 1565 1566 rq = task_rq(p); 1567 /* 1568 * Since p->state == TASK_WAKING, set_task_cpu() has been called 1569 * from try_to_wake_up(). Hence, p->pi_lock is locked, but 1570 * rq->lock is not... So, lock it 1571 */ 1572 raw_spin_lock(&rq->lock); 1573 if (p->dl.dl_non_contending) { 1574 sub_running_bw(p->dl.dl_bw, &rq->dl); 1575 p->dl.dl_non_contending = 0; 1576 /* 1577 * If the timer handler is currently running and the 1578 * timer cannot be cancelled, inactive_task_timer() 1579 * will see that dl_not_contending is not set, and 1580 * will not touch the rq's active utilization, 1581 * so we are still safe. 1582 */ 1583 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) 1584 put_task_struct(p); 1585 } 1586 sub_rq_bw(p->dl.dl_bw, &rq->dl); 1587 raw_spin_unlock(&rq->lock); 1588 } 1589 1590 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) 1591 { 1592 /* 1593 * Current can't be migrated, useless to reschedule, 1594 * let's hope p can move out. 1595 */ 1596 if (rq->curr->nr_cpus_allowed == 1 || 1597 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1) 1598 return; 1599 1600 /* 1601 * p is migratable, so let's not schedule it and 1602 * see if it is pushed or pulled somewhere else. 1603 */ 1604 if (p->nr_cpus_allowed != 1 && 1605 cpudl_find(&rq->rd->cpudl, p, NULL) != -1) 1606 return; 1607 1608 resched_curr(rq); 1609 } 1610 1611 #endif /* CONFIG_SMP */ 1612 1613 /* 1614 * Only called when both the current and waking task are -deadline 1615 * tasks. 1616 */ 1617 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, 1618 int flags) 1619 { 1620 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { 1621 resched_curr(rq); 1622 return; 1623 } 1624 1625 #ifdef CONFIG_SMP 1626 /* 1627 * In the unlikely case current and p have the same deadline 1628 * let us try to decide what's the best thing to do... 1629 */ 1630 if ((p->dl.deadline == rq->curr->dl.deadline) && 1631 !test_tsk_need_resched(rq->curr)) 1632 check_preempt_equal_dl(rq, p); 1633 #endif /* CONFIG_SMP */ 1634 } 1635 1636 #ifdef CONFIG_SCHED_HRTICK 1637 static void start_hrtick_dl(struct rq *rq, struct task_struct *p) 1638 { 1639 hrtick_start(rq, p->dl.runtime); 1640 } 1641 #else /* !CONFIG_SCHED_HRTICK */ 1642 static void start_hrtick_dl(struct rq *rq, struct task_struct *p) 1643 { 1644 } 1645 #endif 1646 1647 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, 1648 struct dl_rq *dl_rq) 1649 { 1650 struct rb_node *left = dl_rq->rb_leftmost; 1651 1652 if (!left) 1653 return NULL; 1654 1655 return rb_entry(left, struct sched_dl_entity, rb_node); 1656 } 1657 1658 struct task_struct * 1659 pick_next_task_dl(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 1660 { 1661 struct sched_dl_entity *dl_se; 1662 struct task_struct *p; 1663 struct dl_rq *dl_rq; 1664 1665 dl_rq = &rq->dl; 1666 1667 if (need_pull_dl_task(rq, prev)) { 1668 /* 1669 * This is OK, because current is on_cpu, which avoids it being 1670 * picked for load-balance and preemption/IRQs are still 1671 * disabled avoiding further scheduler activity on it and we're 1672 * being very careful to re-start the picking loop. 1673 */ 1674 rq_unpin_lock(rq, rf); 1675 pull_dl_task(rq); 1676 rq_repin_lock(rq, rf); 1677 /* 1678 * pull_dl_task() can drop (and re-acquire) rq->lock; this 1679 * means a stop task can slip in, in which case we need to 1680 * re-start task selection. 1681 */ 1682 if (rq->stop && task_on_rq_queued(rq->stop)) 1683 return RETRY_TASK; 1684 } 1685 1686 /* 1687 * When prev is DL, we may throttle it in put_prev_task(). 1688 * So, we update time before we check for dl_nr_running. 1689 */ 1690 if (prev->sched_class == &dl_sched_class) 1691 update_curr_dl(rq); 1692 1693 if (unlikely(!dl_rq->dl_nr_running)) 1694 return NULL; 1695 1696 put_prev_task(rq, prev); 1697 1698 dl_se = pick_next_dl_entity(rq, dl_rq); 1699 BUG_ON(!dl_se); 1700 1701 p = dl_task_of(dl_se); 1702 p->se.exec_start = rq_clock_task(rq); 1703 1704 /* Running task will never be pushed. */ 1705 dequeue_pushable_dl_task(rq, p); 1706 1707 if (hrtick_enabled(rq)) 1708 start_hrtick_dl(rq, p); 1709 1710 queue_push_tasks(rq); 1711 1712 return p; 1713 } 1714 1715 static void put_prev_task_dl(struct rq *rq, struct task_struct *p) 1716 { 1717 update_curr_dl(rq); 1718 1719 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) 1720 enqueue_pushable_dl_task(rq, p); 1721 } 1722 1723 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) 1724 { 1725 update_curr_dl(rq); 1726 1727 /* 1728 * Even when we have runtime, update_curr_dl() might have resulted in us 1729 * not being the leftmost task anymore. In that case NEED_RESCHED will 1730 * be set and schedule() will start a new hrtick for the next task. 1731 */ 1732 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 && 1733 is_leftmost(p, &rq->dl)) 1734 start_hrtick_dl(rq, p); 1735 } 1736 1737 static void task_fork_dl(struct task_struct *p) 1738 { 1739 /* 1740 * SCHED_DEADLINE tasks cannot fork and this is achieved through 1741 * sched_fork() 1742 */ 1743 } 1744 1745 static void set_curr_task_dl(struct rq *rq) 1746 { 1747 struct task_struct *p = rq->curr; 1748 1749 p->se.exec_start = rq_clock_task(rq); 1750 1751 /* You can't push away the running task */ 1752 dequeue_pushable_dl_task(rq, p); 1753 } 1754 1755 #ifdef CONFIG_SMP 1756 1757 /* Only try algorithms three times */ 1758 #define DL_MAX_TRIES 3 1759 1760 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) 1761 { 1762 if (!task_running(rq, p) && 1763 cpumask_test_cpu(cpu, &p->cpus_allowed)) 1764 return 1; 1765 return 0; 1766 } 1767 1768 /* 1769 * Return the earliest pushable rq's task, which is suitable to be executed 1770 * on the CPU, NULL otherwise: 1771 */ 1772 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu) 1773 { 1774 struct rb_node *next_node = rq->dl.pushable_dl_tasks_leftmost; 1775 struct task_struct *p = NULL; 1776 1777 if (!has_pushable_dl_tasks(rq)) 1778 return NULL; 1779 1780 next_node: 1781 if (next_node) { 1782 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks); 1783 1784 if (pick_dl_task(rq, p, cpu)) 1785 return p; 1786 1787 next_node = rb_next(next_node); 1788 goto next_node; 1789 } 1790 1791 return NULL; 1792 } 1793 1794 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); 1795 1796 static int find_later_rq(struct task_struct *task) 1797 { 1798 struct sched_domain *sd; 1799 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); 1800 int this_cpu = smp_processor_id(); 1801 int best_cpu, cpu = task_cpu(task); 1802 1803 /* Make sure the mask is initialized first */ 1804 if (unlikely(!later_mask)) 1805 return -1; 1806 1807 if (task->nr_cpus_allowed == 1) 1808 return -1; 1809 1810 /* 1811 * We have to consider system topology and task affinity 1812 * first, then we can look for a suitable cpu. 1813 */ 1814 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl, 1815 task, later_mask); 1816 if (best_cpu == -1) 1817 return -1; 1818 1819 /* 1820 * If we are here, some target has been found, 1821 * the most suitable of which is cached in best_cpu. 1822 * This is, among the runqueues where the current tasks 1823 * have later deadlines than the task's one, the rq 1824 * with the latest possible one. 1825 * 1826 * Now we check how well this matches with task's 1827 * affinity and system topology. 1828 * 1829 * The last cpu where the task run is our first 1830 * guess, since it is most likely cache-hot there. 1831 */ 1832 if (cpumask_test_cpu(cpu, later_mask)) 1833 return cpu; 1834 /* 1835 * Check if this_cpu is to be skipped (i.e., it is 1836 * not in the mask) or not. 1837 */ 1838 if (!cpumask_test_cpu(this_cpu, later_mask)) 1839 this_cpu = -1; 1840 1841 rcu_read_lock(); 1842 for_each_domain(cpu, sd) { 1843 if (sd->flags & SD_WAKE_AFFINE) { 1844 1845 /* 1846 * If possible, preempting this_cpu is 1847 * cheaper than migrating. 1848 */ 1849 if (this_cpu != -1 && 1850 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { 1851 rcu_read_unlock(); 1852 return this_cpu; 1853 } 1854 1855 /* 1856 * Last chance: if best_cpu is valid and is 1857 * in the mask, that becomes our choice. 1858 */ 1859 if (best_cpu < nr_cpu_ids && 1860 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) { 1861 rcu_read_unlock(); 1862 return best_cpu; 1863 } 1864 } 1865 } 1866 rcu_read_unlock(); 1867 1868 /* 1869 * At this point, all our guesses failed, we just return 1870 * 'something', and let the caller sort the things out. 1871 */ 1872 if (this_cpu != -1) 1873 return this_cpu; 1874 1875 cpu = cpumask_any(later_mask); 1876 if (cpu < nr_cpu_ids) 1877 return cpu; 1878 1879 return -1; 1880 } 1881 1882 /* Locks the rq it finds */ 1883 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) 1884 { 1885 struct rq *later_rq = NULL; 1886 int tries; 1887 int cpu; 1888 1889 for (tries = 0; tries < DL_MAX_TRIES; tries++) { 1890 cpu = find_later_rq(task); 1891 1892 if ((cpu == -1) || (cpu == rq->cpu)) 1893 break; 1894 1895 later_rq = cpu_rq(cpu); 1896 1897 if (later_rq->dl.dl_nr_running && 1898 !dl_time_before(task->dl.deadline, 1899 later_rq->dl.earliest_dl.curr)) { 1900 /* 1901 * Target rq has tasks of equal or earlier deadline, 1902 * retrying does not release any lock and is unlikely 1903 * to yield a different result. 1904 */ 1905 later_rq = NULL; 1906 break; 1907 } 1908 1909 /* Retry if something changed. */ 1910 if (double_lock_balance(rq, later_rq)) { 1911 if (unlikely(task_rq(task) != rq || 1912 !cpumask_test_cpu(later_rq->cpu, &task->cpus_allowed) || 1913 task_running(rq, task) || 1914 !dl_task(task) || 1915 !task_on_rq_queued(task))) { 1916 double_unlock_balance(rq, later_rq); 1917 later_rq = NULL; 1918 break; 1919 } 1920 } 1921 1922 /* 1923 * If the rq we found has no -deadline task, or 1924 * its earliest one has a later deadline than our 1925 * task, the rq is a good one. 1926 */ 1927 if (!later_rq->dl.dl_nr_running || 1928 dl_time_before(task->dl.deadline, 1929 later_rq->dl.earliest_dl.curr)) 1930 break; 1931 1932 /* Otherwise we try again. */ 1933 double_unlock_balance(rq, later_rq); 1934 later_rq = NULL; 1935 } 1936 1937 return later_rq; 1938 } 1939 1940 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) 1941 { 1942 struct task_struct *p; 1943 1944 if (!has_pushable_dl_tasks(rq)) 1945 return NULL; 1946 1947 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost, 1948 struct task_struct, pushable_dl_tasks); 1949 1950 BUG_ON(rq->cpu != task_cpu(p)); 1951 BUG_ON(task_current(rq, p)); 1952 BUG_ON(p->nr_cpus_allowed <= 1); 1953 1954 BUG_ON(!task_on_rq_queued(p)); 1955 BUG_ON(!dl_task(p)); 1956 1957 return p; 1958 } 1959 1960 /* 1961 * See if the non running -deadline tasks on this rq 1962 * can be sent to some other CPU where they can preempt 1963 * and start executing. 1964 */ 1965 static int push_dl_task(struct rq *rq) 1966 { 1967 struct task_struct *next_task; 1968 struct rq *later_rq; 1969 int ret = 0; 1970 1971 if (!rq->dl.overloaded) 1972 return 0; 1973 1974 next_task = pick_next_pushable_dl_task(rq); 1975 if (!next_task) 1976 return 0; 1977 1978 retry: 1979 if (unlikely(next_task == rq->curr)) { 1980 WARN_ON(1); 1981 return 0; 1982 } 1983 1984 /* 1985 * If next_task preempts rq->curr, and rq->curr 1986 * can move away, it makes sense to just reschedule 1987 * without going further in pushing next_task. 1988 */ 1989 if (dl_task(rq->curr) && 1990 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && 1991 rq->curr->nr_cpus_allowed > 1) { 1992 resched_curr(rq); 1993 return 0; 1994 } 1995 1996 /* We might release rq lock */ 1997 get_task_struct(next_task); 1998 1999 /* Will lock the rq it'll find */ 2000 later_rq = find_lock_later_rq(next_task, rq); 2001 if (!later_rq) { 2002 struct task_struct *task; 2003 2004 /* 2005 * We must check all this again, since 2006 * find_lock_later_rq releases rq->lock and it is 2007 * then possible that next_task has migrated. 2008 */ 2009 task = pick_next_pushable_dl_task(rq); 2010 if (task == next_task) { 2011 /* 2012 * The task is still there. We don't try 2013 * again, some other cpu will pull it when ready. 2014 */ 2015 goto out; 2016 } 2017 2018 if (!task) 2019 /* No more tasks */ 2020 goto out; 2021 2022 put_task_struct(next_task); 2023 next_task = task; 2024 goto retry; 2025 } 2026 2027 deactivate_task(rq, next_task, 0); 2028 sub_running_bw(next_task->dl.dl_bw, &rq->dl); 2029 sub_rq_bw(next_task->dl.dl_bw, &rq->dl); 2030 set_task_cpu(next_task, later_rq->cpu); 2031 add_rq_bw(next_task->dl.dl_bw, &later_rq->dl); 2032 add_running_bw(next_task->dl.dl_bw, &later_rq->dl); 2033 activate_task(later_rq, next_task, 0); 2034 ret = 1; 2035 2036 resched_curr(later_rq); 2037 2038 double_unlock_balance(rq, later_rq); 2039 2040 out: 2041 put_task_struct(next_task); 2042 2043 return ret; 2044 } 2045 2046 static void push_dl_tasks(struct rq *rq) 2047 { 2048 /* push_dl_task() will return true if it moved a -deadline task */ 2049 while (push_dl_task(rq)) 2050 ; 2051 } 2052 2053 static void pull_dl_task(struct rq *this_rq) 2054 { 2055 int this_cpu = this_rq->cpu, cpu; 2056 struct task_struct *p; 2057 bool resched = false; 2058 struct rq *src_rq; 2059 u64 dmin = LONG_MAX; 2060 2061 if (likely(!dl_overloaded(this_rq))) 2062 return; 2063 2064 /* 2065 * Match the barrier from dl_set_overloaded; this guarantees that if we 2066 * see overloaded we must also see the dlo_mask bit. 2067 */ 2068 smp_rmb(); 2069 2070 for_each_cpu(cpu, this_rq->rd->dlo_mask) { 2071 if (this_cpu == cpu) 2072 continue; 2073 2074 src_rq = cpu_rq(cpu); 2075 2076 /* 2077 * It looks racy, abd it is! However, as in sched_rt.c, 2078 * we are fine with this. 2079 */ 2080 if (this_rq->dl.dl_nr_running && 2081 dl_time_before(this_rq->dl.earliest_dl.curr, 2082 src_rq->dl.earliest_dl.next)) 2083 continue; 2084 2085 /* Might drop this_rq->lock */ 2086 double_lock_balance(this_rq, src_rq); 2087 2088 /* 2089 * If there are no more pullable tasks on the 2090 * rq, we're done with it. 2091 */ 2092 if (src_rq->dl.dl_nr_running <= 1) 2093 goto skip; 2094 2095 p = pick_earliest_pushable_dl_task(src_rq, this_cpu); 2096 2097 /* 2098 * We found a task to be pulled if: 2099 * - it preempts our current (if there's one), 2100 * - it will preempt the last one we pulled (if any). 2101 */ 2102 if (p && dl_time_before(p->dl.deadline, dmin) && 2103 (!this_rq->dl.dl_nr_running || 2104 dl_time_before(p->dl.deadline, 2105 this_rq->dl.earliest_dl.curr))) { 2106 WARN_ON(p == src_rq->curr); 2107 WARN_ON(!task_on_rq_queued(p)); 2108 2109 /* 2110 * Then we pull iff p has actually an earlier 2111 * deadline than the current task of its runqueue. 2112 */ 2113 if (dl_time_before(p->dl.deadline, 2114 src_rq->curr->dl.deadline)) 2115 goto skip; 2116 2117 resched = true; 2118 2119 deactivate_task(src_rq, p, 0); 2120 sub_running_bw(p->dl.dl_bw, &src_rq->dl); 2121 sub_rq_bw(p->dl.dl_bw, &src_rq->dl); 2122 set_task_cpu(p, this_cpu); 2123 add_rq_bw(p->dl.dl_bw, &this_rq->dl); 2124 add_running_bw(p->dl.dl_bw, &this_rq->dl); 2125 activate_task(this_rq, p, 0); 2126 dmin = p->dl.deadline; 2127 2128 /* Is there any other task even earlier? */ 2129 } 2130 skip: 2131 double_unlock_balance(this_rq, src_rq); 2132 } 2133 2134 if (resched) 2135 resched_curr(this_rq); 2136 } 2137 2138 /* 2139 * Since the task is not running and a reschedule is not going to happen 2140 * anytime soon on its runqueue, we try pushing it away now. 2141 */ 2142 static void task_woken_dl(struct rq *rq, struct task_struct *p) 2143 { 2144 if (!task_running(rq, p) && 2145 !test_tsk_need_resched(rq->curr) && 2146 p->nr_cpus_allowed > 1 && 2147 dl_task(rq->curr) && 2148 (rq->curr->nr_cpus_allowed < 2 || 2149 !dl_entity_preempt(&p->dl, &rq->curr->dl))) { 2150 push_dl_tasks(rq); 2151 } 2152 } 2153 2154 static void set_cpus_allowed_dl(struct task_struct *p, 2155 const struct cpumask *new_mask) 2156 { 2157 struct root_domain *src_rd; 2158 struct rq *rq; 2159 2160 BUG_ON(!dl_task(p)); 2161 2162 rq = task_rq(p); 2163 src_rd = rq->rd; 2164 /* 2165 * Migrating a SCHED_DEADLINE task between exclusive 2166 * cpusets (different root_domains) entails a bandwidth 2167 * update. We already made space for us in the destination 2168 * domain (see cpuset_can_attach()). 2169 */ 2170 if (!cpumask_intersects(src_rd->span, new_mask)) { 2171 struct dl_bw *src_dl_b; 2172 2173 src_dl_b = dl_bw_of(cpu_of(rq)); 2174 /* 2175 * We now free resources of the root_domain we are migrating 2176 * off. In the worst case, sched_setattr() may temporary fail 2177 * until we complete the update. 2178 */ 2179 raw_spin_lock(&src_dl_b->lock); 2180 __dl_clear(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p))); 2181 raw_spin_unlock(&src_dl_b->lock); 2182 } 2183 2184 set_cpus_allowed_common(p, new_mask); 2185 } 2186 2187 /* Assumes rq->lock is held */ 2188 static void rq_online_dl(struct rq *rq) 2189 { 2190 if (rq->dl.overloaded) 2191 dl_set_overload(rq); 2192 2193 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu); 2194 if (rq->dl.dl_nr_running > 0) 2195 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr); 2196 } 2197 2198 /* Assumes rq->lock is held */ 2199 static void rq_offline_dl(struct rq *rq) 2200 { 2201 if (rq->dl.overloaded) 2202 dl_clear_overload(rq); 2203 2204 cpudl_clear(&rq->rd->cpudl, rq->cpu); 2205 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu); 2206 } 2207 2208 void __init init_sched_dl_class(void) 2209 { 2210 unsigned int i; 2211 2212 for_each_possible_cpu(i) 2213 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), 2214 GFP_KERNEL, cpu_to_node(i)); 2215 } 2216 2217 #endif /* CONFIG_SMP */ 2218 2219 static void switched_from_dl(struct rq *rq, struct task_struct *p) 2220 { 2221 /* 2222 * task_non_contending() can start the "inactive timer" (if the 0-lag 2223 * time is in the future). If the task switches back to dl before 2224 * the "inactive timer" fires, it can continue to consume its current 2225 * runtime using its current deadline. If it stays outside of 2226 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer() 2227 * will reset the task parameters. 2228 */ 2229 if (task_on_rq_queued(p) && p->dl.dl_runtime) 2230 task_non_contending(p); 2231 2232 if (!task_on_rq_queued(p)) 2233 sub_rq_bw(p->dl.dl_bw, &rq->dl); 2234 2235 /* 2236 * We cannot use inactive_task_timer() to invoke sub_running_bw() 2237 * at the 0-lag time, because the task could have been migrated 2238 * while SCHED_OTHER in the meanwhile. 2239 */ 2240 if (p->dl.dl_non_contending) 2241 p->dl.dl_non_contending = 0; 2242 2243 /* 2244 * Since this might be the only -deadline task on the rq, 2245 * this is the right place to try to pull some other one 2246 * from an overloaded cpu, if any. 2247 */ 2248 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) 2249 return; 2250 2251 queue_pull_task(rq); 2252 } 2253 2254 /* 2255 * When switching to -deadline, we may overload the rq, then 2256 * we try to push someone off, if possible. 2257 */ 2258 static void switched_to_dl(struct rq *rq, struct task_struct *p) 2259 { 2260 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1) 2261 put_task_struct(p); 2262 2263 /* If p is not queued we will update its parameters at next wakeup. */ 2264 if (!task_on_rq_queued(p)) { 2265 add_rq_bw(p->dl.dl_bw, &rq->dl); 2266 2267 return; 2268 } 2269 /* 2270 * If p is boosted we already updated its params in 2271 * rt_mutex_setprio()->enqueue_task(..., ENQUEUE_REPLENISH), 2272 * p's deadline being now already after rq_clock(rq). 2273 */ 2274 if (dl_time_before(p->dl.deadline, rq_clock(rq))) 2275 setup_new_dl_entity(&p->dl); 2276 2277 if (rq->curr != p) { 2278 #ifdef CONFIG_SMP 2279 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded) 2280 queue_push_tasks(rq); 2281 #endif 2282 if (dl_task(rq->curr)) 2283 check_preempt_curr_dl(rq, p, 0); 2284 else 2285 resched_curr(rq); 2286 } 2287 } 2288 2289 /* 2290 * If the scheduling parameters of a -deadline task changed, 2291 * a push or pull operation might be needed. 2292 */ 2293 static void prio_changed_dl(struct rq *rq, struct task_struct *p, 2294 int oldprio) 2295 { 2296 if (task_on_rq_queued(p) || rq->curr == p) { 2297 #ifdef CONFIG_SMP 2298 /* 2299 * This might be too much, but unfortunately 2300 * we don't have the old deadline value, and 2301 * we can't argue if the task is increasing 2302 * or lowering its prio, so... 2303 */ 2304 if (!rq->dl.overloaded) 2305 queue_pull_task(rq); 2306 2307 /* 2308 * If we now have a earlier deadline task than p, 2309 * then reschedule, provided p is still on this 2310 * runqueue. 2311 */ 2312 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline)) 2313 resched_curr(rq); 2314 #else 2315 /* 2316 * Again, we don't know if p has a earlier 2317 * or later deadline, so let's blindly set a 2318 * (maybe not needed) rescheduling point. 2319 */ 2320 resched_curr(rq); 2321 #endif /* CONFIG_SMP */ 2322 } 2323 } 2324 2325 const struct sched_class dl_sched_class = { 2326 .next = &rt_sched_class, 2327 .enqueue_task = enqueue_task_dl, 2328 .dequeue_task = dequeue_task_dl, 2329 .yield_task = yield_task_dl, 2330 2331 .check_preempt_curr = check_preempt_curr_dl, 2332 2333 .pick_next_task = pick_next_task_dl, 2334 .put_prev_task = put_prev_task_dl, 2335 2336 #ifdef CONFIG_SMP 2337 .select_task_rq = select_task_rq_dl, 2338 .migrate_task_rq = migrate_task_rq_dl, 2339 .set_cpus_allowed = set_cpus_allowed_dl, 2340 .rq_online = rq_online_dl, 2341 .rq_offline = rq_offline_dl, 2342 .task_woken = task_woken_dl, 2343 #endif 2344 2345 .set_curr_task = set_curr_task_dl, 2346 .task_tick = task_tick_dl, 2347 .task_fork = task_fork_dl, 2348 2349 .prio_changed = prio_changed_dl, 2350 .switched_from = switched_from_dl, 2351 .switched_to = switched_to_dl, 2352 2353 .update_curr = update_curr_dl, 2354 }; 2355 2356 int sched_dl_global_validate(void) 2357 { 2358 u64 runtime = global_rt_runtime(); 2359 u64 period = global_rt_period(); 2360 u64 new_bw = to_ratio(period, runtime); 2361 struct dl_bw *dl_b; 2362 int cpu, ret = 0; 2363 unsigned long flags; 2364 2365 /* 2366 * Here we want to check the bandwidth not being set to some 2367 * value smaller than the currently allocated bandwidth in 2368 * any of the root_domains. 2369 * 2370 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 2371 * cycling on root_domains... Discussion on different/better 2372 * solutions is welcome! 2373 */ 2374 for_each_possible_cpu(cpu) { 2375 rcu_read_lock_sched(); 2376 dl_b = dl_bw_of(cpu); 2377 2378 raw_spin_lock_irqsave(&dl_b->lock, flags); 2379 if (new_bw < dl_b->total_bw) 2380 ret = -EBUSY; 2381 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 2382 2383 rcu_read_unlock_sched(); 2384 2385 if (ret) 2386 break; 2387 } 2388 2389 return ret; 2390 } 2391 2392 void init_dl_rq_bw_ratio(struct dl_rq *dl_rq) 2393 { 2394 if (global_rt_runtime() == RUNTIME_INF) { 2395 dl_rq->bw_ratio = 1 << RATIO_SHIFT; 2396 dl_rq->extra_bw = 1 << BW_SHIFT; 2397 } else { 2398 dl_rq->bw_ratio = to_ratio(global_rt_runtime(), 2399 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT); 2400 dl_rq->extra_bw = to_ratio(global_rt_period(), 2401 global_rt_runtime()); 2402 } 2403 } 2404 2405 void sched_dl_do_global(void) 2406 { 2407 u64 new_bw = -1; 2408 struct dl_bw *dl_b; 2409 int cpu; 2410 unsigned long flags; 2411 2412 def_dl_bandwidth.dl_period = global_rt_period(); 2413 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 2414 2415 if (global_rt_runtime() != RUNTIME_INF) 2416 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 2417 2418 /* 2419 * FIXME: As above... 2420 */ 2421 for_each_possible_cpu(cpu) { 2422 rcu_read_lock_sched(); 2423 dl_b = dl_bw_of(cpu); 2424 2425 raw_spin_lock_irqsave(&dl_b->lock, flags); 2426 dl_b->bw = new_bw; 2427 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 2428 2429 rcu_read_unlock_sched(); 2430 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl); 2431 } 2432 } 2433 2434 /* 2435 * We must be sure that accepting a new task (or allowing changing the 2436 * parameters of an existing one) is consistent with the bandwidth 2437 * constraints. If yes, this function also accordingly updates the currently 2438 * allocated bandwidth to reflect the new situation. 2439 * 2440 * This function is called while holding p's rq->lock. 2441 */ 2442 int sched_dl_overflow(struct task_struct *p, int policy, 2443 const struct sched_attr *attr) 2444 { 2445 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2446 u64 period = attr->sched_period ?: attr->sched_deadline; 2447 u64 runtime = attr->sched_runtime; 2448 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2449 int cpus, err = -1; 2450 2451 /* !deadline task may carry old deadline bandwidth */ 2452 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p)) 2453 return 0; 2454 2455 /* 2456 * Either if a task, enters, leave, or stays -deadline but changes 2457 * its parameters, we may need to update accordingly the total 2458 * allocated bandwidth of the container. 2459 */ 2460 raw_spin_lock(&dl_b->lock); 2461 cpus = dl_bw_cpus(task_cpu(p)); 2462 if (dl_policy(policy) && !task_has_dl_policy(p) && 2463 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2464 if (hrtimer_active(&p->dl.inactive_timer)) 2465 __dl_clear(dl_b, p->dl.dl_bw, cpus); 2466 __dl_add(dl_b, new_bw, cpus); 2467 err = 0; 2468 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2469 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2470 /* 2471 * XXX this is slightly incorrect: when the task 2472 * utilization decreases, we should delay the total 2473 * utilization change until the task's 0-lag point. 2474 * But this would require to set the task's "inactive 2475 * timer" when the task is not inactive. 2476 */ 2477 __dl_clear(dl_b, p->dl.dl_bw, cpus); 2478 __dl_add(dl_b, new_bw, cpus); 2479 dl_change_utilization(p, new_bw); 2480 err = 0; 2481 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2482 /* 2483 * Do not decrease the total deadline utilization here, 2484 * switched_from_dl() will take care to do it at the correct 2485 * (0-lag) time. 2486 */ 2487 err = 0; 2488 } 2489 raw_spin_unlock(&dl_b->lock); 2490 2491 return err; 2492 } 2493 2494 /* 2495 * This function initializes the sched_dl_entity of a newly becoming 2496 * SCHED_DEADLINE task. 2497 * 2498 * Only the static values are considered here, the actual runtime and the 2499 * absolute deadline will be properly calculated when the task is enqueued 2500 * for the first time with its new policy. 2501 */ 2502 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr) 2503 { 2504 struct sched_dl_entity *dl_se = &p->dl; 2505 2506 dl_se->dl_runtime = attr->sched_runtime; 2507 dl_se->dl_deadline = attr->sched_deadline; 2508 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 2509 dl_se->flags = attr->sched_flags; 2510 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 2511 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime); 2512 } 2513 2514 void __getparam_dl(struct task_struct *p, struct sched_attr *attr) 2515 { 2516 struct sched_dl_entity *dl_se = &p->dl; 2517 2518 attr->sched_priority = p->rt_priority; 2519 attr->sched_runtime = dl_se->dl_runtime; 2520 attr->sched_deadline = dl_se->dl_deadline; 2521 attr->sched_period = dl_se->dl_period; 2522 attr->sched_flags = dl_se->flags; 2523 } 2524 2525 /* 2526 * This function validates the new parameters of a -deadline task. 2527 * We ask for the deadline not being zero, and greater or equal 2528 * than the runtime, as well as the period of being zero or 2529 * greater than deadline. Furthermore, we have to be sure that 2530 * user parameters are above the internal resolution of 1us (we 2531 * check sched_runtime only since it is always the smaller one) and 2532 * below 2^63 ns (we have to check both sched_deadline and 2533 * sched_period, as the latter can be zero). 2534 */ 2535 bool __checkparam_dl(const struct sched_attr *attr) 2536 { 2537 /* deadline != 0 */ 2538 if (attr->sched_deadline == 0) 2539 return false; 2540 2541 /* 2542 * Since we truncate DL_SCALE bits, make sure we're at least 2543 * that big. 2544 */ 2545 if (attr->sched_runtime < (1ULL << DL_SCALE)) 2546 return false; 2547 2548 /* 2549 * Since we use the MSB for wrap-around and sign issues, make 2550 * sure it's not set (mind that period can be equal to zero). 2551 */ 2552 if (attr->sched_deadline & (1ULL << 63) || 2553 attr->sched_period & (1ULL << 63)) 2554 return false; 2555 2556 /* runtime <= deadline <= period (if period != 0) */ 2557 if ((attr->sched_period != 0 && 2558 attr->sched_period < attr->sched_deadline) || 2559 attr->sched_deadline < attr->sched_runtime) 2560 return false; 2561 2562 return true; 2563 } 2564 2565 /* 2566 * This function clears the sched_dl_entity static params. 2567 */ 2568 void __dl_clear_params(struct task_struct *p) 2569 { 2570 struct sched_dl_entity *dl_se = &p->dl; 2571 2572 dl_se->dl_runtime = 0; 2573 dl_se->dl_deadline = 0; 2574 dl_se->dl_period = 0; 2575 dl_se->flags = 0; 2576 dl_se->dl_bw = 0; 2577 dl_se->dl_density = 0; 2578 2579 dl_se->dl_throttled = 0; 2580 dl_se->dl_yielded = 0; 2581 dl_se->dl_non_contending = 0; 2582 } 2583 2584 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr) 2585 { 2586 struct sched_dl_entity *dl_se = &p->dl; 2587 2588 if (dl_se->dl_runtime != attr->sched_runtime || 2589 dl_se->dl_deadline != attr->sched_deadline || 2590 dl_se->dl_period != attr->sched_period || 2591 dl_se->flags != attr->sched_flags) 2592 return true; 2593 2594 return false; 2595 } 2596 2597 #ifdef CONFIG_SMP 2598 int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed) 2599 { 2600 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask, 2601 cs_cpus_allowed); 2602 struct dl_bw *dl_b; 2603 bool overflow; 2604 int cpus, ret; 2605 unsigned long flags; 2606 2607 rcu_read_lock_sched(); 2608 dl_b = dl_bw_of(dest_cpu); 2609 raw_spin_lock_irqsave(&dl_b->lock, flags); 2610 cpus = dl_bw_cpus(dest_cpu); 2611 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw); 2612 if (overflow) 2613 ret = -EBUSY; 2614 else { 2615 /* 2616 * We reserve space for this task in the destination 2617 * root_domain, as we can't fail after this point. 2618 * We will free resources in the source root_domain 2619 * later on (see set_cpus_allowed_dl()). 2620 */ 2621 __dl_add(dl_b, p->dl.dl_bw, cpus); 2622 ret = 0; 2623 } 2624 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 2625 rcu_read_unlock_sched(); 2626 return ret; 2627 } 2628 2629 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, 2630 const struct cpumask *trial) 2631 { 2632 int ret = 1, trial_cpus; 2633 struct dl_bw *cur_dl_b; 2634 unsigned long flags; 2635 2636 rcu_read_lock_sched(); 2637 cur_dl_b = dl_bw_of(cpumask_any(cur)); 2638 trial_cpus = cpumask_weight(trial); 2639 2640 raw_spin_lock_irqsave(&cur_dl_b->lock, flags); 2641 if (cur_dl_b->bw != -1 && 2642 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) 2643 ret = 0; 2644 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); 2645 rcu_read_unlock_sched(); 2646 return ret; 2647 } 2648 2649 bool dl_cpu_busy(unsigned int cpu) 2650 { 2651 unsigned long flags; 2652 struct dl_bw *dl_b; 2653 bool overflow; 2654 int cpus; 2655 2656 rcu_read_lock_sched(); 2657 dl_b = dl_bw_of(cpu); 2658 raw_spin_lock_irqsave(&dl_b->lock, flags); 2659 cpus = dl_bw_cpus(cpu); 2660 overflow = __dl_overflow(dl_b, cpus, 0, 0); 2661 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 2662 rcu_read_unlock_sched(); 2663 return overflow; 2664 } 2665 #endif 2666 2667 #ifdef CONFIG_SCHED_DEBUG 2668 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); 2669 2670 void print_dl_stats(struct seq_file *m, int cpu) 2671 { 2672 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); 2673 } 2674 #endif /* CONFIG_SCHED_DEBUG */ 2675