1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * kernel/sched/core.c 4 * 5 * Core kernel scheduler code and related syscalls 6 * 7 * Copyright (C) 1991-2002 Linus Torvalds 8 */ 9 #define CREATE_TRACE_POINTS 10 #include <trace/events/sched.h> 11 #undef CREATE_TRACE_POINTS 12 13 #include "sched.h" 14 15 #include <linux/nospec.h> 16 17 #include <linux/kcov.h> 18 #include <linux/scs.h> 19 20 #include <asm/switch_to.h> 21 #include <asm/tlb.h> 22 23 #include "../workqueue_internal.h" 24 #include "../../fs/io-wq.h" 25 #include "../smpboot.h" 26 27 #include "pelt.h" 28 #include "smp.h" 29 30 /* 31 * Export tracepoints that act as a bare tracehook (ie: have no trace event 32 * associated with them) to allow external modules to probe them. 33 */ 34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp); 35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp); 36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp); 37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp); 38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp); 39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp); 40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp); 41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp); 42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp); 43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp); 44 45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 46 47 #ifdef CONFIG_SCHED_DEBUG 48 /* 49 * Debugging: various feature bits 50 * 51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of 52 * sysctl_sched_features, defined in sched.h, to allow constants propagation 53 * at compile time and compiler optimization based on features default. 54 */ 55 #define SCHED_FEAT(name, enabled) \ 56 (1UL << __SCHED_FEAT_##name) * enabled | 57 const_debug unsigned int sysctl_sched_features = 58 #include "features.h" 59 0; 60 #undef SCHED_FEAT 61 62 /* 63 * Print a warning if need_resched is set for the given duration (if 64 * LATENCY_WARN is enabled). 65 * 66 * If sysctl_resched_latency_warn_once is set, only one warning will be shown 67 * per boot. 68 */ 69 __read_mostly int sysctl_resched_latency_warn_ms = 100; 70 __read_mostly int sysctl_resched_latency_warn_once = 1; 71 #endif /* CONFIG_SCHED_DEBUG */ 72 73 /* 74 * Number of tasks to iterate in a single balance run. 75 * Limited because this is done with IRQs disabled. 76 */ 77 const_debug unsigned int sysctl_sched_nr_migrate = 32; 78 79 /* 80 * period over which we measure -rt task CPU usage in us. 81 * default: 1s 82 */ 83 unsigned int sysctl_sched_rt_period = 1000000; 84 85 __read_mostly int scheduler_running; 86 87 #ifdef CONFIG_SCHED_CORE 88 89 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled); 90 91 /* kernel prio, less is more */ 92 static inline int __task_prio(struct task_struct *p) 93 { 94 if (p->sched_class == &stop_sched_class) /* trumps deadline */ 95 return -2; 96 97 if (rt_prio(p->prio)) /* includes deadline */ 98 return p->prio; /* [-1, 99] */ 99 100 if (p->sched_class == &idle_sched_class) 101 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */ 102 103 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */ 104 } 105 106 /* 107 * l(a,b) 108 * le(a,b) := !l(b,a) 109 * g(a,b) := l(b,a) 110 * ge(a,b) := !l(a,b) 111 */ 112 113 /* real prio, less is less */ 114 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi) 115 { 116 117 int pa = __task_prio(a), pb = __task_prio(b); 118 119 if (-pa < -pb) 120 return true; 121 122 if (-pb < -pa) 123 return false; 124 125 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */ 126 return !dl_time_before(a->dl.deadline, b->dl.deadline); 127 128 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */ 129 return cfs_prio_less(a, b, in_fi); 130 131 return false; 132 } 133 134 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b) 135 { 136 if (a->core_cookie < b->core_cookie) 137 return true; 138 139 if (a->core_cookie > b->core_cookie) 140 return false; 141 142 /* flip prio, so high prio is leftmost */ 143 if (prio_less(b, a, task_rq(a)->core->core_forceidle)) 144 return true; 145 146 return false; 147 } 148 149 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node) 150 151 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b) 152 { 153 return __sched_core_less(__node_2_sc(a), __node_2_sc(b)); 154 } 155 156 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node) 157 { 158 const struct task_struct *p = __node_2_sc(node); 159 unsigned long cookie = (unsigned long)key; 160 161 if (cookie < p->core_cookie) 162 return -1; 163 164 if (cookie > p->core_cookie) 165 return 1; 166 167 return 0; 168 } 169 170 void sched_core_enqueue(struct rq *rq, struct task_struct *p) 171 { 172 rq->core->core_task_seq++; 173 174 if (!p->core_cookie) 175 return; 176 177 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less); 178 } 179 180 void sched_core_dequeue(struct rq *rq, struct task_struct *p) 181 { 182 rq->core->core_task_seq++; 183 184 if (!sched_core_enqueued(p)) 185 return; 186 187 rb_erase(&p->core_node, &rq->core_tree); 188 RB_CLEAR_NODE(&p->core_node); 189 } 190 191 /* 192 * Find left-most (aka, highest priority) task matching @cookie. 193 */ 194 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie) 195 { 196 struct rb_node *node; 197 198 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp); 199 /* 200 * The idle task always matches any cookie! 201 */ 202 if (!node) 203 return idle_sched_class.pick_task(rq); 204 205 return __node_2_sc(node); 206 } 207 208 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie) 209 { 210 struct rb_node *node = &p->core_node; 211 212 node = rb_next(node); 213 if (!node) 214 return NULL; 215 216 p = container_of(node, struct task_struct, core_node); 217 if (p->core_cookie != cookie) 218 return NULL; 219 220 return p; 221 } 222 223 /* 224 * Magic required such that: 225 * 226 * raw_spin_rq_lock(rq); 227 * ... 228 * raw_spin_rq_unlock(rq); 229 * 230 * ends up locking and unlocking the _same_ lock, and all CPUs 231 * always agree on what rq has what lock. 232 * 233 * XXX entirely possible to selectively enable cores, don't bother for now. 234 */ 235 236 static DEFINE_MUTEX(sched_core_mutex); 237 static atomic_t sched_core_count; 238 static struct cpumask sched_core_mask; 239 240 static void sched_core_lock(int cpu, unsigned long *flags) 241 { 242 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 243 int t, i = 0; 244 245 local_irq_save(*flags); 246 for_each_cpu(t, smt_mask) 247 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++); 248 } 249 250 static void sched_core_unlock(int cpu, unsigned long *flags) 251 { 252 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 253 int t; 254 255 for_each_cpu(t, smt_mask) 256 raw_spin_unlock(&cpu_rq(t)->__lock); 257 local_irq_restore(*flags); 258 } 259 260 static void __sched_core_flip(bool enabled) 261 { 262 unsigned long flags; 263 int cpu, t; 264 265 cpus_read_lock(); 266 267 /* 268 * Toggle the online cores, one by one. 269 */ 270 cpumask_copy(&sched_core_mask, cpu_online_mask); 271 for_each_cpu(cpu, &sched_core_mask) { 272 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 273 274 sched_core_lock(cpu, &flags); 275 276 for_each_cpu(t, smt_mask) 277 cpu_rq(t)->core_enabled = enabled; 278 279 sched_core_unlock(cpu, &flags); 280 281 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask); 282 } 283 284 /* 285 * Toggle the offline CPUs. 286 */ 287 cpumask_copy(&sched_core_mask, cpu_possible_mask); 288 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask); 289 290 for_each_cpu(cpu, &sched_core_mask) 291 cpu_rq(cpu)->core_enabled = enabled; 292 293 cpus_read_unlock(); 294 } 295 296 static void sched_core_assert_empty(void) 297 { 298 int cpu; 299 300 for_each_possible_cpu(cpu) 301 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree)); 302 } 303 304 static void __sched_core_enable(void) 305 { 306 static_branch_enable(&__sched_core_enabled); 307 /* 308 * Ensure all previous instances of raw_spin_rq_*lock() have finished 309 * and future ones will observe !sched_core_disabled(). 310 */ 311 synchronize_rcu(); 312 __sched_core_flip(true); 313 sched_core_assert_empty(); 314 } 315 316 static void __sched_core_disable(void) 317 { 318 sched_core_assert_empty(); 319 __sched_core_flip(false); 320 static_branch_disable(&__sched_core_enabled); 321 } 322 323 void sched_core_get(void) 324 { 325 if (atomic_inc_not_zero(&sched_core_count)) 326 return; 327 328 mutex_lock(&sched_core_mutex); 329 if (!atomic_read(&sched_core_count)) 330 __sched_core_enable(); 331 332 smp_mb__before_atomic(); 333 atomic_inc(&sched_core_count); 334 mutex_unlock(&sched_core_mutex); 335 } 336 337 static void __sched_core_put(struct work_struct *work) 338 { 339 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) { 340 __sched_core_disable(); 341 mutex_unlock(&sched_core_mutex); 342 } 343 } 344 345 void sched_core_put(void) 346 { 347 static DECLARE_WORK(_work, __sched_core_put); 348 349 /* 350 * "There can be only one" 351 * 352 * Either this is the last one, or we don't actually need to do any 353 * 'work'. If it is the last *again*, we rely on 354 * WORK_STRUCT_PENDING_BIT. 355 */ 356 if (!atomic_add_unless(&sched_core_count, -1, 1)) 357 schedule_work(&_work); 358 } 359 360 #else /* !CONFIG_SCHED_CORE */ 361 362 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { } 363 static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { } 364 365 #endif /* CONFIG_SCHED_CORE */ 366 367 /* 368 * part of the period that we allow rt tasks to run in us. 369 * default: 0.95s 370 */ 371 int sysctl_sched_rt_runtime = 950000; 372 373 374 /* 375 * Serialization rules: 376 * 377 * Lock order: 378 * 379 * p->pi_lock 380 * rq->lock 381 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls) 382 * 383 * rq1->lock 384 * rq2->lock where: rq1 < rq2 385 * 386 * Regular state: 387 * 388 * Normal scheduling state is serialized by rq->lock. __schedule() takes the 389 * local CPU's rq->lock, it optionally removes the task from the runqueue and 390 * always looks at the local rq data structures to find the most eligible task 391 * to run next. 392 * 393 * Task enqueue is also under rq->lock, possibly taken from another CPU. 394 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to 395 * the local CPU to avoid bouncing the runqueue state around [ see 396 * ttwu_queue_wakelist() ] 397 * 398 * Task wakeup, specifically wakeups that involve migration, are horribly 399 * complicated to avoid having to take two rq->locks. 400 * 401 * Special state: 402 * 403 * System-calls and anything external will use task_rq_lock() which acquires 404 * both p->pi_lock and rq->lock. As a consequence the state they change is 405 * stable while holding either lock: 406 * 407 * - sched_setaffinity()/ 408 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed 409 * - set_user_nice(): p->se.load, p->*prio 410 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio, 411 * p->se.load, p->rt_priority, 412 * p->dl.dl_{runtime, deadline, period, flags, bw, density} 413 * - sched_setnuma(): p->numa_preferred_nid 414 * - sched_move_task()/ 415 * cpu_cgroup_fork(): p->sched_task_group 416 * - uclamp_update_active() p->uclamp* 417 * 418 * p->state <- TASK_*: 419 * 420 * is changed locklessly using set_current_state(), __set_current_state() or 421 * set_special_state(), see their respective comments, or by 422 * try_to_wake_up(). This latter uses p->pi_lock to serialize against 423 * concurrent self. 424 * 425 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }: 426 * 427 * is set by activate_task() and cleared by deactivate_task(), under 428 * rq->lock. Non-zero indicates the task is runnable, the special 429 * ON_RQ_MIGRATING state is used for migration without holding both 430 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock(). 431 * 432 * p->on_cpu <- { 0, 1 }: 433 * 434 * is set by prepare_task() and cleared by finish_task() such that it will be 435 * set before p is scheduled-in and cleared after p is scheduled-out, both 436 * under rq->lock. Non-zero indicates the task is running on its CPU. 437 * 438 * [ The astute reader will observe that it is possible for two tasks on one 439 * CPU to have ->on_cpu = 1 at the same time. ] 440 * 441 * task_cpu(p): is changed by set_task_cpu(), the rules are: 442 * 443 * - Don't call set_task_cpu() on a blocked task: 444 * 445 * We don't care what CPU we're not running on, this simplifies hotplug, 446 * the CPU assignment of blocked tasks isn't required to be valid. 447 * 448 * - for try_to_wake_up(), called under p->pi_lock: 449 * 450 * This allows try_to_wake_up() to only take one rq->lock, see its comment. 451 * 452 * - for migration called under rq->lock: 453 * [ see task_on_rq_migrating() in task_rq_lock() ] 454 * 455 * o move_queued_task() 456 * o detach_task() 457 * 458 * - for migration called under double_rq_lock(): 459 * 460 * o __migrate_swap_task() 461 * o push_rt_task() / pull_rt_task() 462 * o push_dl_task() / pull_dl_task() 463 * o dl_task_offline_migration() 464 * 465 */ 466 467 void raw_spin_rq_lock_nested(struct rq *rq, int subclass) 468 { 469 raw_spinlock_t *lock; 470 471 /* Matches synchronize_rcu() in __sched_core_enable() */ 472 preempt_disable(); 473 if (sched_core_disabled()) { 474 raw_spin_lock_nested(&rq->__lock, subclass); 475 /* preempt_count *MUST* be > 1 */ 476 preempt_enable_no_resched(); 477 return; 478 } 479 480 for (;;) { 481 lock = __rq_lockp(rq); 482 raw_spin_lock_nested(lock, subclass); 483 if (likely(lock == __rq_lockp(rq))) { 484 /* preempt_count *MUST* be > 1 */ 485 preempt_enable_no_resched(); 486 return; 487 } 488 raw_spin_unlock(lock); 489 } 490 } 491 492 bool raw_spin_rq_trylock(struct rq *rq) 493 { 494 raw_spinlock_t *lock; 495 bool ret; 496 497 /* Matches synchronize_rcu() in __sched_core_enable() */ 498 preempt_disable(); 499 if (sched_core_disabled()) { 500 ret = raw_spin_trylock(&rq->__lock); 501 preempt_enable(); 502 return ret; 503 } 504 505 for (;;) { 506 lock = __rq_lockp(rq); 507 ret = raw_spin_trylock(lock); 508 if (!ret || (likely(lock == __rq_lockp(rq)))) { 509 preempt_enable(); 510 return ret; 511 } 512 raw_spin_unlock(lock); 513 } 514 } 515 516 void raw_spin_rq_unlock(struct rq *rq) 517 { 518 raw_spin_unlock(rq_lockp(rq)); 519 } 520 521 #ifdef CONFIG_SMP 522 /* 523 * double_rq_lock - safely lock two runqueues 524 */ 525 void double_rq_lock(struct rq *rq1, struct rq *rq2) 526 { 527 lockdep_assert_irqs_disabled(); 528 529 if (rq_order_less(rq2, rq1)) 530 swap(rq1, rq2); 531 532 raw_spin_rq_lock(rq1); 533 if (__rq_lockp(rq1) == __rq_lockp(rq2)) 534 return; 535 536 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING); 537 } 538 #endif 539 540 /* 541 * __task_rq_lock - lock the rq @p resides on. 542 */ 543 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) 544 __acquires(rq->lock) 545 { 546 struct rq *rq; 547 548 lockdep_assert_held(&p->pi_lock); 549 550 for (;;) { 551 rq = task_rq(p); 552 raw_spin_rq_lock(rq); 553 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { 554 rq_pin_lock(rq, rf); 555 return rq; 556 } 557 raw_spin_rq_unlock(rq); 558 559 while (unlikely(task_on_rq_migrating(p))) 560 cpu_relax(); 561 } 562 } 563 564 /* 565 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 566 */ 567 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) 568 __acquires(p->pi_lock) 569 __acquires(rq->lock) 570 { 571 struct rq *rq; 572 573 for (;;) { 574 raw_spin_lock_irqsave(&p->pi_lock, rf->flags); 575 rq = task_rq(p); 576 raw_spin_rq_lock(rq); 577 /* 578 * move_queued_task() task_rq_lock() 579 * 580 * ACQUIRE (rq->lock) 581 * [S] ->on_rq = MIGRATING [L] rq = task_rq() 582 * WMB (__set_task_cpu()) ACQUIRE (rq->lock); 583 * [S] ->cpu = new_cpu [L] task_rq() 584 * [L] ->on_rq 585 * RELEASE (rq->lock) 586 * 587 * If we observe the old CPU in task_rq_lock(), the acquire of 588 * the old rq->lock will fully serialize against the stores. 589 * 590 * If we observe the new CPU in task_rq_lock(), the address 591 * dependency headed by '[L] rq = task_rq()' and the acquire 592 * will pair with the WMB to ensure we then also see migrating. 593 */ 594 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { 595 rq_pin_lock(rq, rf); 596 return rq; 597 } 598 raw_spin_rq_unlock(rq); 599 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); 600 601 while (unlikely(task_on_rq_migrating(p))) 602 cpu_relax(); 603 } 604 } 605 606 /* 607 * RQ-clock updating methods: 608 */ 609 610 static void update_rq_clock_task(struct rq *rq, s64 delta) 611 { 612 /* 613 * In theory, the compile should just see 0 here, and optimize out the call 614 * to sched_rt_avg_update. But I don't trust it... 615 */ 616 s64 __maybe_unused steal = 0, irq_delta = 0; 617 618 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 619 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 620 621 /* 622 * Since irq_time is only updated on {soft,}irq_exit, we might run into 623 * this case when a previous update_rq_clock() happened inside a 624 * {soft,}irq region. 625 * 626 * When this happens, we stop ->clock_task and only update the 627 * prev_irq_time stamp to account for the part that fit, so that a next 628 * update will consume the rest. This ensures ->clock_task is 629 * monotonic. 630 * 631 * It does however cause some slight miss-attribution of {soft,}irq 632 * time, a more accurate solution would be to update the irq_time using 633 * the current rq->clock timestamp, except that would require using 634 * atomic ops. 635 */ 636 if (irq_delta > delta) 637 irq_delta = delta; 638 639 rq->prev_irq_time += irq_delta; 640 delta -= irq_delta; 641 #endif 642 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 643 if (static_key_false((¶virt_steal_rq_enabled))) { 644 steal = paravirt_steal_clock(cpu_of(rq)); 645 steal -= rq->prev_steal_time_rq; 646 647 if (unlikely(steal > delta)) 648 steal = delta; 649 650 rq->prev_steal_time_rq += steal; 651 delta -= steal; 652 } 653 #endif 654 655 rq->clock_task += delta; 656 657 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 658 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 659 update_irq_load_avg(rq, irq_delta + steal); 660 #endif 661 update_rq_clock_pelt(rq, delta); 662 } 663 664 void update_rq_clock(struct rq *rq) 665 { 666 s64 delta; 667 668 lockdep_assert_rq_held(rq); 669 670 if (rq->clock_update_flags & RQCF_ACT_SKIP) 671 return; 672 673 #ifdef CONFIG_SCHED_DEBUG 674 if (sched_feat(WARN_DOUBLE_CLOCK)) 675 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED); 676 rq->clock_update_flags |= RQCF_UPDATED; 677 #endif 678 679 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 680 if (delta < 0) 681 return; 682 rq->clock += delta; 683 update_rq_clock_task(rq, delta); 684 } 685 686 #ifdef CONFIG_SCHED_HRTICK 687 /* 688 * Use HR-timers to deliver accurate preemption points. 689 */ 690 691 static void hrtick_clear(struct rq *rq) 692 { 693 if (hrtimer_active(&rq->hrtick_timer)) 694 hrtimer_cancel(&rq->hrtick_timer); 695 } 696 697 /* 698 * High-resolution timer tick. 699 * Runs from hardirq context with interrupts disabled. 700 */ 701 static enum hrtimer_restart hrtick(struct hrtimer *timer) 702 { 703 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 704 struct rq_flags rf; 705 706 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 707 708 rq_lock(rq, &rf); 709 update_rq_clock(rq); 710 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 711 rq_unlock(rq, &rf); 712 713 return HRTIMER_NORESTART; 714 } 715 716 #ifdef CONFIG_SMP 717 718 static void __hrtick_restart(struct rq *rq) 719 { 720 struct hrtimer *timer = &rq->hrtick_timer; 721 ktime_t time = rq->hrtick_time; 722 723 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD); 724 } 725 726 /* 727 * called from hardirq (IPI) context 728 */ 729 static void __hrtick_start(void *arg) 730 { 731 struct rq *rq = arg; 732 struct rq_flags rf; 733 734 rq_lock(rq, &rf); 735 __hrtick_restart(rq); 736 rq_unlock(rq, &rf); 737 } 738 739 /* 740 * Called to set the hrtick timer state. 741 * 742 * called with rq->lock held and irqs disabled 743 */ 744 void hrtick_start(struct rq *rq, u64 delay) 745 { 746 struct hrtimer *timer = &rq->hrtick_timer; 747 s64 delta; 748 749 /* 750 * Don't schedule slices shorter than 10000ns, that just 751 * doesn't make sense and can cause timer DoS. 752 */ 753 delta = max_t(s64, delay, 10000LL); 754 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta); 755 756 if (rq == this_rq()) 757 __hrtick_restart(rq); 758 else 759 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 760 } 761 762 #else 763 /* 764 * Called to set the hrtick timer state. 765 * 766 * called with rq->lock held and irqs disabled 767 */ 768 void hrtick_start(struct rq *rq, u64 delay) 769 { 770 /* 771 * Don't schedule slices shorter than 10000ns, that just 772 * doesn't make sense. Rely on vruntime for fairness. 773 */ 774 delay = max_t(u64, delay, 10000LL); 775 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), 776 HRTIMER_MODE_REL_PINNED_HARD); 777 } 778 779 #endif /* CONFIG_SMP */ 780 781 static void hrtick_rq_init(struct rq *rq) 782 { 783 #ifdef CONFIG_SMP 784 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq); 785 #endif 786 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); 787 rq->hrtick_timer.function = hrtick; 788 } 789 #else /* CONFIG_SCHED_HRTICK */ 790 static inline void hrtick_clear(struct rq *rq) 791 { 792 } 793 794 static inline void hrtick_rq_init(struct rq *rq) 795 { 796 } 797 #endif /* CONFIG_SCHED_HRTICK */ 798 799 /* 800 * cmpxchg based fetch_or, macro so it works for different integer types 801 */ 802 #define fetch_or(ptr, mask) \ 803 ({ \ 804 typeof(ptr) _ptr = (ptr); \ 805 typeof(mask) _mask = (mask); \ 806 typeof(*_ptr) _old, _val = *_ptr; \ 807 \ 808 for (;;) { \ 809 _old = cmpxchg(_ptr, _val, _val | _mask); \ 810 if (_old == _val) \ 811 break; \ 812 _val = _old; \ 813 } \ 814 _old; \ 815 }) 816 817 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 818 /* 819 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 820 * this avoids any races wrt polling state changes and thereby avoids 821 * spurious IPIs. 822 */ 823 static bool set_nr_and_not_polling(struct task_struct *p) 824 { 825 struct thread_info *ti = task_thread_info(p); 826 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 827 } 828 829 /* 830 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 831 * 832 * If this returns true, then the idle task promises to call 833 * sched_ttwu_pending() and reschedule soon. 834 */ 835 static bool set_nr_if_polling(struct task_struct *p) 836 { 837 struct thread_info *ti = task_thread_info(p); 838 typeof(ti->flags) old, val = READ_ONCE(ti->flags); 839 840 for (;;) { 841 if (!(val & _TIF_POLLING_NRFLAG)) 842 return false; 843 if (val & _TIF_NEED_RESCHED) 844 return true; 845 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 846 if (old == val) 847 break; 848 val = old; 849 } 850 return true; 851 } 852 853 #else 854 static bool set_nr_and_not_polling(struct task_struct *p) 855 { 856 set_tsk_need_resched(p); 857 return true; 858 } 859 860 #ifdef CONFIG_SMP 861 static bool set_nr_if_polling(struct task_struct *p) 862 { 863 return false; 864 } 865 #endif 866 #endif 867 868 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task) 869 { 870 struct wake_q_node *node = &task->wake_q; 871 872 /* 873 * Atomically grab the task, if ->wake_q is !nil already it means 874 * it's already queued (either by us or someone else) and will get the 875 * wakeup due to that. 876 * 877 * In order to ensure that a pending wakeup will observe our pending 878 * state, even in the failed case, an explicit smp_mb() must be used. 879 */ 880 smp_mb__before_atomic(); 881 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL))) 882 return false; 883 884 /* 885 * The head is context local, there can be no concurrency. 886 */ 887 *head->lastp = node; 888 head->lastp = &node->next; 889 return true; 890 } 891 892 /** 893 * wake_q_add() - queue a wakeup for 'later' waking. 894 * @head: the wake_q_head to add @task to 895 * @task: the task to queue for 'later' wakeup 896 * 897 * Queue a task for later wakeup, most likely by the wake_up_q() call in the 898 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come 899 * instantly. 900 * 901 * This function must be used as-if it were wake_up_process(); IOW the task 902 * must be ready to be woken at this location. 903 */ 904 void wake_q_add(struct wake_q_head *head, struct task_struct *task) 905 { 906 if (__wake_q_add(head, task)) 907 get_task_struct(task); 908 } 909 910 /** 911 * wake_q_add_safe() - safely queue a wakeup for 'later' waking. 912 * @head: the wake_q_head to add @task to 913 * @task: the task to queue for 'later' wakeup 914 * 915 * Queue a task for later wakeup, most likely by the wake_up_q() call in the 916 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come 917 * instantly. 918 * 919 * This function must be used as-if it were wake_up_process(); IOW the task 920 * must be ready to be woken at this location. 921 * 922 * This function is essentially a task-safe equivalent to wake_q_add(). Callers 923 * that already hold reference to @task can call the 'safe' version and trust 924 * wake_q to do the right thing depending whether or not the @task is already 925 * queued for wakeup. 926 */ 927 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task) 928 { 929 if (!__wake_q_add(head, task)) 930 put_task_struct(task); 931 } 932 933 void wake_up_q(struct wake_q_head *head) 934 { 935 struct wake_q_node *node = head->first; 936 937 while (node != WAKE_Q_TAIL) { 938 struct task_struct *task; 939 940 task = container_of(node, struct task_struct, wake_q); 941 /* Task can safely be re-inserted now: */ 942 node = node->next; 943 task->wake_q.next = NULL; 944 945 /* 946 * wake_up_process() executes a full barrier, which pairs with 947 * the queueing in wake_q_add() so as not to miss wakeups. 948 */ 949 wake_up_process(task); 950 put_task_struct(task); 951 } 952 } 953 954 /* 955 * resched_curr - mark rq's current task 'to be rescheduled now'. 956 * 957 * On UP this means the setting of the need_resched flag, on SMP it 958 * might also involve a cross-CPU call to trigger the scheduler on 959 * the target CPU. 960 */ 961 void resched_curr(struct rq *rq) 962 { 963 struct task_struct *curr = rq->curr; 964 int cpu; 965 966 lockdep_assert_rq_held(rq); 967 968 if (test_tsk_need_resched(curr)) 969 return; 970 971 cpu = cpu_of(rq); 972 973 if (cpu == smp_processor_id()) { 974 set_tsk_need_resched(curr); 975 set_preempt_need_resched(); 976 return; 977 } 978 979 if (set_nr_and_not_polling(curr)) 980 smp_send_reschedule(cpu); 981 else 982 trace_sched_wake_idle_without_ipi(cpu); 983 } 984 985 void resched_cpu(int cpu) 986 { 987 struct rq *rq = cpu_rq(cpu); 988 unsigned long flags; 989 990 raw_spin_rq_lock_irqsave(rq, flags); 991 if (cpu_online(cpu) || cpu == smp_processor_id()) 992 resched_curr(rq); 993 raw_spin_rq_unlock_irqrestore(rq, flags); 994 } 995 996 #ifdef CONFIG_SMP 997 #ifdef CONFIG_NO_HZ_COMMON 998 /* 999 * In the semi idle case, use the nearest busy CPU for migrating timers 1000 * from an idle CPU. This is good for power-savings. 1001 * 1002 * We don't do similar optimization for completely idle system, as 1003 * selecting an idle CPU will add more delays to the timers than intended 1004 * (as that CPU's timer base may not be uptodate wrt jiffies etc). 1005 */ 1006 int get_nohz_timer_target(void) 1007 { 1008 int i, cpu = smp_processor_id(), default_cpu = -1; 1009 struct sched_domain *sd; 1010 1011 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) { 1012 if (!idle_cpu(cpu)) 1013 return cpu; 1014 default_cpu = cpu; 1015 } 1016 1017 rcu_read_lock(); 1018 for_each_domain(cpu, sd) { 1019 for_each_cpu_and(i, sched_domain_span(sd), 1020 housekeeping_cpumask(HK_FLAG_TIMER)) { 1021 if (cpu == i) 1022 continue; 1023 1024 if (!idle_cpu(i)) { 1025 cpu = i; 1026 goto unlock; 1027 } 1028 } 1029 } 1030 1031 if (default_cpu == -1) 1032 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER); 1033 cpu = default_cpu; 1034 unlock: 1035 rcu_read_unlock(); 1036 return cpu; 1037 } 1038 1039 /* 1040 * When add_timer_on() enqueues a timer into the timer wheel of an 1041 * idle CPU then this timer might expire before the next timer event 1042 * which is scheduled to wake up that CPU. In case of a completely 1043 * idle system the next event might even be infinite time into the 1044 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 1045 * leaves the inner idle loop so the newly added timer is taken into 1046 * account when the CPU goes back to idle and evaluates the timer 1047 * wheel for the next timer event. 1048 */ 1049 static void wake_up_idle_cpu(int cpu) 1050 { 1051 struct rq *rq = cpu_rq(cpu); 1052 1053 if (cpu == smp_processor_id()) 1054 return; 1055 1056 if (set_nr_and_not_polling(rq->idle)) 1057 smp_send_reschedule(cpu); 1058 else 1059 trace_sched_wake_idle_without_ipi(cpu); 1060 } 1061 1062 static bool wake_up_full_nohz_cpu(int cpu) 1063 { 1064 /* 1065 * We just need the target to call irq_exit() and re-evaluate 1066 * the next tick. The nohz full kick at least implies that. 1067 * If needed we can still optimize that later with an 1068 * empty IRQ. 1069 */ 1070 if (cpu_is_offline(cpu)) 1071 return true; /* Don't try to wake offline CPUs. */ 1072 if (tick_nohz_full_cpu(cpu)) { 1073 if (cpu != smp_processor_id() || 1074 tick_nohz_tick_stopped()) 1075 tick_nohz_full_kick_cpu(cpu); 1076 return true; 1077 } 1078 1079 return false; 1080 } 1081 1082 /* 1083 * Wake up the specified CPU. If the CPU is going offline, it is the 1084 * caller's responsibility to deal with the lost wakeup, for example, 1085 * by hooking into the CPU_DEAD notifier like timers and hrtimers do. 1086 */ 1087 void wake_up_nohz_cpu(int cpu) 1088 { 1089 if (!wake_up_full_nohz_cpu(cpu)) 1090 wake_up_idle_cpu(cpu); 1091 } 1092 1093 static void nohz_csd_func(void *info) 1094 { 1095 struct rq *rq = info; 1096 int cpu = cpu_of(rq); 1097 unsigned int flags; 1098 1099 /* 1100 * Release the rq::nohz_csd. 1101 */ 1102 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu)); 1103 WARN_ON(!(flags & NOHZ_KICK_MASK)); 1104 1105 rq->idle_balance = idle_cpu(cpu); 1106 if (rq->idle_balance && !need_resched()) { 1107 rq->nohz_idle_balance = flags; 1108 raise_softirq_irqoff(SCHED_SOFTIRQ); 1109 } 1110 } 1111 1112 #endif /* CONFIG_NO_HZ_COMMON */ 1113 1114 #ifdef CONFIG_NO_HZ_FULL 1115 bool sched_can_stop_tick(struct rq *rq) 1116 { 1117 int fifo_nr_running; 1118 1119 /* Deadline tasks, even if single, need the tick */ 1120 if (rq->dl.dl_nr_running) 1121 return false; 1122 1123 /* 1124 * If there are more than one RR tasks, we need the tick to affect the 1125 * actual RR behaviour. 1126 */ 1127 if (rq->rt.rr_nr_running) { 1128 if (rq->rt.rr_nr_running == 1) 1129 return true; 1130 else 1131 return false; 1132 } 1133 1134 /* 1135 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no 1136 * forced preemption between FIFO tasks. 1137 */ 1138 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running; 1139 if (fifo_nr_running) 1140 return true; 1141 1142 /* 1143 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left; 1144 * if there's more than one we need the tick for involuntary 1145 * preemption. 1146 */ 1147 if (rq->nr_running > 1) 1148 return false; 1149 1150 return true; 1151 } 1152 #endif /* CONFIG_NO_HZ_FULL */ 1153 #endif /* CONFIG_SMP */ 1154 1155 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 1156 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 1157 /* 1158 * Iterate task_group tree rooted at *from, calling @down when first entering a 1159 * node and @up when leaving it for the final time. 1160 * 1161 * Caller must hold rcu_lock or sufficient equivalent. 1162 */ 1163 int walk_tg_tree_from(struct task_group *from, 1164 tg_visitor down, tg_visitor up, void *data) 1165 { 1166 struct task_group *parent, *child; 1167 int ret; 1168 1169 parent = from; 1170 1171 down: 1172 ret = (*down)(parent, data); 1173 if (ret) 1174 goto out; 1175 list_for_each_entry_rcu(child, &parent->children, siblings) { 1176 parent = child; 1177 goto down; 1178 1179 up: 1180 continue; 1181 } 1182 ret = (*up)(parent, data); 1183 if (ret || parent == from) 1184 goto out; 1185 1186 child = parent; 1187 parent = parent->parent; 1188 if (parent) 1189 goto up; 1190 out: 1191 return ret; 1192 } 1193 1194 int tg_nop(struct task_group *tg, void *data) 1195 { 1196 return 0; 1197 } 1198 #endif 1199 1200 static void set_load_weight(struct task_struct *p, bool update_load) 1201 { 1202 int prio = p->static_prio - MAX_RT_PRIO; 1203 struct load_weight *load = &p->se.load; 1204 1205 /* 1206 * SCHED_IDLE tasks get minimal weight: 1207 */ 1208 if (task_has_idle_policy(p)) { 1209 load->weight = scale_load(WEIGHT_IDLEPRIO); 1210 load->inv_weight = WMULT_IDLEPRIO; 1211 return; 1212 } 1213 1214 /* 1215 * SCHED_OTHER tasks have to update their load when changing their 1216 * weight 1217 */ 1218 if (update_load && p->sched_class == &fair_sched_class) { 1219 reweight_task(p, prio); 1220 } else { 1221 load->weight = scale_load(sched_prio_to_weight[prio]); 1222 load->inv_weight = sched_prio_to_wmult[prio]; 1223 } 1224 } 1225 1226 #ifdef CONFIG_UCLAMP_TASK 1227 /* 1228 * Serializes updates of utilization clamp values 1229 * 1230 * The (slow-path) user-space triggers utilization clamp value updates which 1231 * can require updates on (fast-path) scheduler's data structures used to 1232 * support enqueue/dequeue operations. 1233 * While the per-CPU rq lock protects fast-path update operations, user-space 1234 * requests are serialized using a mutex to reduce the risk of conflicting 1235 * updates or API abuses. 1236 */ 1237 static DEFINE_MUTEX(uclamp_mutex); 1238 1239 /* Max allowed minimum utilization */ 1240 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE; 1241 1242 /* Max allowed maximum utilization */ 1243 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE; 1244 1245 /* 1246 * By default RT tasks run at the maximum performance point/capacity of the 1247 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to 1248 * SCHED_CAPACITY_SCALE. 1249 * 1250 * This knob allows admins to change the default behavior when uclamp is being 1251 * used. In battery powered devices, particularly, running at the maximum 1252 * capacity and frequency will increase energy consumption and shorten the 1253 * battery life. 1254 * 1255 * This knob only affects RT tasks that their uclamp_se->user_defined == false. 1256 * 1257 * This knob will not override the system default sched_util_clamp_min defined 1258 * above. 1259 */ 1260 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE; 1261 1262 /* All clamps are required to be less or equal than these values */ 1263 static struct uclamp_se uclamp_default[UCLAMP_CNT]; 1264 1265 /* 1266 * This static key is used to reduce the uclamp overhead in the fast path. It 1267 * primarily disables the call to uclamp_rq_{inc, dec}() in 1268 * enqueue/dequeue_task(). 1269 * 1270 * This allows users to continue to enable uclamp in their kernel config with 1271 * minimum uclamp overhead in the fast path. 1272 * 1273 * As soon as userspace modifies any of the uclamp knobs, the static key is 1274 * enabled, since we have an actual users that make use of uclamp 1275 * functionality. 1276 * 1277 * The knobs that would enable this static key are: 1278 * 1279 * * A task modifying its uclamp value with sched_setattr(). 1280 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs. 1281 * * An admin modifying the cgroup cpu.uclamp.{min, max} 1282 */ 1283 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used); 1284 1285 /* Integer rounded range for each bucket */ 1286 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS) 1287 1288 #define for_each_clamp_id(clamp_id) \ 1289 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++) 1290 1291 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value) 1292 { 1293 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1); 1294 } 1295 1296 static inline unsigned int uclamp_none(enum uclamp_id clamp_id) 1297 { 1298 if (clamp_id == UCLAMP_MIN) 1299 return 0; 1300 return SCHED_CAPACITY_SCALE; 1301 } 1302 1303 static inline void uclamp_se_set(struct uclamp_se *uc_se, 1304 unsigned int value, bool user_defined) 1305 { 1306 uc_se->value = value; 1307 uc_se->bucket_id = uclamp_bucket_id(value); 1308 uc_se->user_defined = user_defined; 1309 } 1310 1311 static inline unsigned int 1312 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id, 1313 unsigned int clamp_value) 1314 { 1315 /* 1316 * Avoid blocked utilization pushing up the frequency when we go 1317 * idle (which drops the max-clamp) by retaining the last known 1318 * max-clamp. 1319 */ 1320 if (clamp_id == UCLAMP_MAX) { 1321 rq->uclamp_flags |= UCLAMP_FLAG_IDLE; 1322 return clamp_value; 1323 } 1324 1325 return uclamp_none(UCLAMP_MIN); 1326 } 1327 1328 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id, 1329 unsigned int clamp_value) 1330 { 1331 /* Reset max-clamp retention only on idle exit */ 1332 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE)) 1333 return; 1334 1335 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value); 1336 } 1337 1338 static inline 1339 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id, 1340 unsigned int clamp_value) 1341 { 1342 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket; 1343 int bucket_id = UCLAMP_BUCKETS - 1; 1344 1345 /* 1346 * Since both min and max clamps are max aggregated, find the 1347 * top most bucket with tasks in. 1348 */ 1349 for ( ; bucket_id >= 0; bucket_id--) { 1350 if (!bucket[bucket_id].tasks) 1351 continue; 1352 return bucket[bucket_id].value; 1353 } 1354 1355 /* No tasks -- default clamp values */ 1356 return uclamp_idle_value(rq, clamp_id, clamp_value); 1357 } 1358 1359 static void __uclamp_update_util_min_rt_default(struct task_struct *p) 1360 { 1361 unsigned int default_util_min; 1362 struct uclamp_se *uc_se; 1363 1364 lockdep_assert_held(&p->pi_lock); 1365 1366 uc_se = &p->uclamp_req[UCLAMP_MIN]; 1367 1368 /* Only sync if user didn't override the default */ 1369 if (uc_se->user_defined) 1370 return; 1371 1372 default_util_min = sysctl_sched_uclamp_util_min_rt_default; 1373 uclamp_se_set(uc_se, default_util_min, false); 1374 } 1375 1376 static void uclamp_update_util_min_rt_default(struct task_struct *p) 1377 { 1378 struct rq_flags rf; 1379 struct rq *rq; 1380 1381 if (!rt_task(p)) 1382 return; 1383 1384 /* Protect updates to p->uclamp_* */ 1385 rq = task_rq_lock(p, &rf); 1386 __uclamp_update_util_min_rt_default(p); 1387 task_rq_unlock(rq, p, &rf); 1388 } 1389 1390 static void uclamp_sync_util_min_rt_default(void) 1391 { 1392 struct task_struct *g, *p; 1393 1394 /* 1395 * copy_process() sysctl_uclamp 1396 * uclamp_min_rt = X; 1397 * write_lock(&tasklist_lock) read_lock(&tasklist_lock) 1398 * // link thread smp_mb__after_spinlock() 1399 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock); 1400 * sched_post_fork() for_each_process_thread() 1401 * __uclamp_sync_rt() __uclamp_sync_rt() 1402 * 1403 * Ensures that either sched_post_fork() will observe the new 1404 * uclamp_min_rt or for_each_process_thread() will observe the new 1405 * task. 1406 */ 1407 read_lock(&tasklist_lock); 1408 smp_mb__after_spinlock(); 1409 read_unlock(&tasklist_lock); 1410 1411 rcu_read_lock(); 1412 for_each_process_thread(g, p) 1413 uclamp_update_util_min_rt_default(p); 1414 rcu_read_unlock(); 1415 } 1416 1417 static inline struct uclamp_se 1418 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id) 1419 { 1420 /* Copy by value as we could modify it */ 1421 struct uclamp_se uc_req = p->uclamp_req[clamp_id]; 1422 #ifdef CONFIG_UCLAMP_TASK_GROUP 1423 unsigned int tg_min, tg_max, value; 1424 1425 /* 1426 * Tasks in autogroups or root task group will be 1427 * restricted by system defaults. 1428 */ 1429 if (task_group_is_autogroup(task_group(p))) 1430 return uc_req; 1431 if (task_group(p) == &root_task_group) 1432 return uc_req; 1433 1434 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value; 1435 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value; 1436 value = uc_req.value; 1437 value = clamp(value, tg_min, tg_max); 1438 uclamp_se_set(&uc_req, value, false); 1439 #endif 1440 1441 return uc_req; 1442 } 1443 1444 /* 1445 * The effective clamp bucket index of a task depends on, by increasing 1446 * priority: 1447 * - the task specific clamp value, when explicitly requested from userspace 1448 * - the task group effective clamp value, for tasks not either in the root 1449 * group or in an autogroup 1450 * - the system default clamp value, defined by the sysadmin 1451 */ 1452 static inline struct uclamp_se 1453 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id) 1454 { 1455 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id); 1456 struct uclamp_se uc_max = uclamp_default[clamp_id]; 1457 1458 /* System default restrictions always apply */ 1459 if (unlikely(uc_req.value > uc_max.value)) 1460 return uc_max; 1461 1462 return uc_req; 1463 } 1464 1465 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id) 1466 { 1467 struct uclamp_se uc_eff; 1468 1469 /* Task currently refcounted: use back-annotated (effective) value */ 1470 if (p->uclamp[clamp_id].active) 1471 return (unsigned long)p->uclamp[clamp_id].value; 1472 1473 uc_eff = uclamp_eff_get(p, clamp_id); 1474 1475 return (unsigned long)uc_eff.value; 1476 } 1477 1478 /* 1479 * When a task is enqueued on a rq, the clamp bucket currently defined by the 1480 * task's uclamp::bucket_id is refcounted on that rq. This also immediately 1481 * updates the rq's clamp value if required. 1482 * 1483 * Tasks can have a task-specific value requested from user-space, track 1484 * within each bucket the maximum value for tasks refcounted in it. 1485 * This "local max aggregation" allows to track the exact "requested" value 1486 * for each bucket when all its RUNNABLE tasks require the same clamp. 1487 */ 1488 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p, 1489 enum uclamp_id clamp_id) 1490 { 1491 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id]; 1492 struct uclamp_se *uc_se = &p->uclamp[clamp_id]; 1493 struct uclamp_bucket *bucket; 1494 1495 lockdep_assert_rq_held(rq); 1496 1497 /* Update task effective clamp */ 1498 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id); 1499 1500 bucket = &uc_rq->bucket[uc_se->bucket_id]; 1501 bucket->tasks++; 1502 uc_se->active = true; 1503 1504 uclamp_idle_reset(rq, clamp_id, uc_se->value); 1505 1506 /* 1507 * Local max aggregation: rq buckets always track the max 1508 * "requested" clamp value of its RUNNABLE tasks. 1509 */ 1510 if (bucket->tasks == 1 || uc_se->value > bucket->value) 1511 bucket->value = uc_se->value; 1512 1513 if (uc_se->value > READ_ONCE(uc_rq->value)) 1514 WRITE_ONCE(uc_rq->value, uc_se->value); 1515 } 1516 1517 /* 1518 * When a task is dequeued from a rq, the clamp bucket refcounted by the task 1519 * is released. If this is the last task reference counting the rq's max 1520 * active clamp value, then the rq's clamp value is updated. 1521 * 1522 * Both refcounted tasks and rq's cached clamp values are expected to be 1523 * always valid. If it's detected they are not, as defensive programming, 1524 * enforce the expected state and warn. 1525 */ 1526 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p, 1527 enum uclamp_id clamp_id) 1528 { 1529 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id]; 1530 struct uclamp_se *uc_se = &p->uclamp[clamp_id]; 1531 struct uclamp_bucket *bucket; 1532 unsigned int bkt_clamp; 1533 unsigned int rq_clamp; 1534 1535 lockdep_assert_rq_held(rq); 1536 1537 /* 1538 * If sched_uclamp_used was enabled after task @p was enqueued, 1539 * we could end up with unbalanced call to uclamp_rq_dec_id(). 1540 * 1541 * In this case the uc_se->active flag should be false since no uclamp 1542 * accounting was performed at enqueue time and we can just return 1543 * here. 1544 * 1545 * Need to be careful of the following enqueue/dequeue ordering 1546 * problem too 1547 * 1548 * enqueue(taskA) 1549 * // sched_uclamp_used gets enabled 1550 * enqueue(taskB) 1551 * dequeue(taskA) 1552 * // Must not decrement bucket->tasks here 1553 * dequeue(taskB) 1554 * 1555 * where we could end up with stale data in uc_se and 1556 * bucket[uc_se->bucket_id]. 1557 * 1558 * The following check here eliminates the possibility of such race. 1559 */ 1560 if (unlikely(!uc_se->active)) 1561 return; 1562 1563 bucket = &uc_rq->bucket[uc_se->bucket_id]; 1564 1565 SCHED_WARN_ON(!bucket->tasks); 1566 if (likely(bucket->tasks)) 1567 bucket->tasks--; 1568 1569 uc_se->active = false; 1570 1571 /* 1572 * Keep "local max aggregation" simple and accept to (possibly) 1573 * overboost some RUNNABLE tasks in the same bucket. 1574 * The rq clamp bucket value is reset to its base value whenever 1575 * there are no more RUNNABLE tasks refcounting it. 1576 */ 1577 if (likely(bucket->tasks)) 1578 return; 1579 1580 rq_clamp = READ_ONCE(uc_rq->value); 1581 /* 1582 * Defensive programming: this should never happen. If it happens, 1583 * e.g. due to future modification, warn and fixup the expected value. 1584 */ 1585 SCHED_WARN_ON(bucket->value > rq_clamp); 1586 if (bucket->value >= rq_clamp) { 1587 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value); 1588 WRITE_ONCE(uc_rq->value, bkt_clamp); 1589 } 1590 } 1591 1592 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) 1593 { 1594 enum uclamp_id clamp_id; 1595 1596 /* 1597 * Avoid any overhead until uclamp is actually used by the userspace. 1598 * 1599 * The condition is constructed such that a NOP is generated when 1600 * sched_uclamp_used is disabled. 1601 */ 1602 if (!static_branch_unlikely(&sched_uclamp_used)) 1603 return; 1604 1605 if (unlikely(!p->sched_class->uclamp_enabled)) 1606 return; 1607 1608 for_each_clamp_id(clamp_id) 1609 uclamp_rq_inc_id(rq, p, clamp_id); 1610 1611 /* Reset clamp idle holding when there is one RUNNABLE task */ 1612 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE) 1613 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE; 1614 } 1615 1616 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) 1617 { 1618 enum uclamp_id clamp_id; 1619 1620 /* 1621 * Avoid any overhead until uclamp is actually used by the userspace. 1622 * 1623 * The condition is constructed such that a NOP is generated when 1624 * sched_uclamp_used is disabled. 1625 */ 1626 if (!static_branch_unlikely(&sched_uclamp_used)) 1627 return; 1628 1629 if (unlikely(!p->sched_class->uclamp_enabled)) 1630 return; 1631 1632 for_each_clamp_id(clamp_id) 1633 uclamp_rq_dec_id(rq, p, clamp_id); 1634 } 1635 1636 static inline void 1637 uclamp_update_active(struct task_struct *p) 1638 { 1639 enum uclamp_id clamp_id; 1640 struct rq_flags rf; 1641 struct rq *rq; 1642 1643 /* 1644 * Lock the task and the rq where the task is (or was) queued. 1645 * 1646 * We might lock the (previous) rq of a !RUNNABLE task, but that's the 1647 * price to pay to safely serialize util_{min,max} updates with 1648 * enqueues, dequeues and migration operations. 1649 * This is the same locking schema used by __set_cpus_allowed_ptr(). 1650 */ 1651 rq = task_rq_lock(p, &rf); 1652 1653 /* 1654 * Setting the clamp bucket is serialized by task_rq_lock(). 1655 * If the task is not yet RUNNABLE and its task_struct is not 1656 * affecting a valid clamp bucket, the next time it's enqueued, 1657 * it will already see the updated clamp bucket value. 1658 */ 1659 for_each_clamp_id(clamp_id) { 1660 if (p->uclamp[clamp_id].active) { 1661 uclamp_rq_dec_id(rq, p, clamp_id); 1662 uclamp_rq_inc_id(rq, p, clamp_id); 1663 } 1664 } 1665 1666 task_rq_unlock(rq, p, &rf); 1667 } 1668 1669 #ifdef CONFIG_UCLAMP_TASK_GROUP 1670 static inline void 1671 uclamp_update_active_tasks(struct cgroup_subsys_state *css) 1672 { 1673 struct css_task_iter it; 1674 struct task_struct *p; 1675 1676 css_task_iter_start(css, 0, &it); 1677 while ((p = css_task_iter_next(&it))) 1678 uclamp_update_active(p); 1679 css_task_iter_end(&it); 1680 } 1681 1682 static void cpu_util_update_eff(struct cgroup_subsys_state *css); 1683 static void uclamp_update_root_tg(void) 1684 { 1685 struct task_group *tg = &root_task_group; 1686 1687 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN], 1688 sysctl_sched_uclamp_util_min, false); 1689 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX], 1690 sysctl_sched_uclamp_util_max, false); 1691 1692 rcu_read_lock(); 1693 cpu_util_update_eff(&root_task_group.css); 1694 rcu_read_unlock(); 1695 } 1696 #else 1697 static void uclamp_update_root_tg(void) { } 1698 #endif 1699 1700 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write, 1701 void *buffer, size_t *lenp, loff_t *ppos) 1702 { 1703 bool update_root_tg = false; 1704 int old_min, old_max, old_min_rt; 1705 int result; 1706 1707 mutex_lock(&uclamp_mutex); 1708 old_min = sysctl_sched_uclamp_util_min; 1709 old_max = sysctl_sched_uclamp_util_max; 1710 old_min_rt = sysctl_sched_uclamp_util_min_rt_default; 1711 1712 result = proc_dointvec(table, write, buffer, lenp, ppos); 1713 if (result) 1714 goto undo; 1715 if (!write) 1716 goto done; 1717 1718 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max || 1719 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE || 1720 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) { 1721 1722 result = -EINVAL; 1723 goto undo; 1724 } 1725 1726 if (old_min != sysctl_sched_uclamp_util_min) { 1727 uclamp_se_set(&uclamp_default[UCLAMP_MIN], 1728 sysctl_sched_uclamp_util_min, false); 1729 update_root_tg = true; 1730 } 1731 if (old_max != sysctl_sched_uclamp_util_max) { 1732 uclamp_se_set(&uclamp_default[UCLAMP_MAX], 1733 sysctl_sched_uclamp_util_max, false); 1734 update_root_tg = true; 1735 } 1736 1737 if (update_root_tg) { 1738 static_branch_enable(&sched_uclamp_used); 1739 uclamp_update_root_tg(); 1740 } 1741 1742 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) { 1743 static_branch_enable(&sched_uclamp_used); 1744 uclamp_sync_util_min_rt_default(); 1745 } 1746 1747 /* 1748 * We update all RUNNABLE tasks only when task groups are in use. 1749 * Otherwise, keep it simple and do just a lazy update at each next 1750 * task enqueue time. 1751 */ 1752 1753 goto done; 1754 1755 undo: 1756 sysctl_sched_uclamp_util_min = old_min; 1757 sysctl_sched_uclamp_util_max = old_max; 1758 sysctl_sched_uclamp_util_min_rt_default = old_min_rt; 1759 done: 1760 mutex_unlock(&uclamp_mutex); 1761 1762 return result; 1763 } 1764 1765 static int uclamp_validate(struct task_struct *p, 1766 const struct sched_attr *attr) 1767 { 1768 int util_min = p->uclamp_req[UCLAMP_MIN].value; 1769 int util_max = p->uclamp_req[UCLAMP_MAX].value; 1770 1771 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) { 1772 util_min = attr->sched_util_min; 1773 1774 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1) 1775 return -EINVAL; 1776 } 1777 1778 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) { 1779 util_max = attr->sched_util_max; 1780 1781 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1) 1782 return -EINVAL; 1783 } 1784 1785 if (util_min != -1 && util_max != -1 && util_min > util_max) 1786 return -EINVAL; 1787 1788 /* 1789 * We have valid uclamp attributes; make sure uclamp is enabled. 1790 * 1791 * We need to do that here, because enabling static branches is a 1792 * blocking operation which obviously cannot be done while holding 1793 * scheduler locks. 1794 */ 1795 static_branch_enable(&sched_uclamp_used); 1796 1797 return 0; 1798 } 1799 1800 static bool uclamp_reset(const struct sched_attr *attr, 1801 enum uclamp_id clamp_id, 1802 struct uclamp_se *uc_se) 1803 { 1804 /* Reset on sched class change for a non user-defined clamp value. */ 1805 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) && 1806 !uc_se->user_defined) 1807 return true; 1808 1809 /* Reset on sched_util_{min,max} == -1. */ 1810 if (clamp_id == UCLAMP_MIN && 1811 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN && 1812 attr->sched_util_min == -1) { 1813 return true; 1814 } 1815 1816 if (clamp_id == UCLAMP_MAX && 1817 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX && 1818 attr->sched_util_max == -1) { 1819 return true; 1820 } 1821 1822 return false; 1823 } 1824 1825 static void __setscheduler_uclamp(struct task_struct *p, 1826 const struct sched_attr *attr) 1827 { 1828 enum uclamp_id clamp_id; 1829 1830 for_each_clamp_id(clamp_id) { 1831 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id]; 1832 unsigned int value; 1833 1834 if (!uclamp_reset(attr, clamp_id, uc_se)) 1835 continue; 1836 1837 /* 1838 * RT by default have a 100% boost value that could be modified 1839 * at runtime. 1840 */ 1841 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN)) 1842 value = sysctl_sched_uclamp_util_min_rt_default; 1843 else 1844 value = uclamp_none(clamp_id); 1845 1846 uclamp_se_set(uc_se, value, false); 1847 1848 } 1849 1850 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP))) 1851 return; 1852 1853 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN && 1854 attr->sched_util_min != -1) { 1855 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN], 1856 attr->sched_util_min, true); 1857 } 1858 1859 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX && 1860 attr->sched_util_max != -1) { 1861 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX], 1862 attr->sched_util_max, true); 1863 } 1864 } 1865 1866 static void uclamp_fork(struct task_struct *p) 1867 { 1868 enum uclamp_id clamp_id; 1869 1870 /* 1871 * We don't need to hold task_rq_lock() when updating p->uclamp_* here 1872 * as the task is still at its early fork stages. 1873 */ 1874 for_each_clamp_id(clamp_id) 1875 p->uclamp[clamp_id].active = false; 1876 1877 if (likely(!p->sched_reset_on_fork)) 1878 return; 1879 1880 for_each_clamp_id(clamp_id) { 1881 uclamp_se_set(&p->uclamp_req[clamp_id], 1882 uclamp_none(clamp_id), false); 1883 } 1884 } 1885 1886 static void uclamp_post_fork(struct task_struct *p) 1887 { 1888 uclamp_update_util_min_rt_default(p); 1889 } 1890 1891 static void __init init_uclamp_rq(struct rq *rq) 1892 { 1893 enum uclamp_id clamp_id; 1894 struct uclamp_rq *uc_rq = rq->uclamp; 1895 1896 for_each_clamp_id(clamp_id) { 1897 uc_rq[clamp_id] = (struct uclamp_rq) { 1898 .value = uclamp_none(clamp_id) 1899 }; 1900 } 1901 1902 rq->uclamp_flags = 0; 1903 } 1904 1905 static void __init init_uclamp(void) 1906 { 1907 struct uclamp_se uc_max = {}; 1908 enum uclamp_id clamp_id; 1909 int cpu; 1910 1911 for_each_possible_cpu(cpu) 1912 init_uclamp_rq(cpu_rq(cpu)); 1913 1914 for_each_clamp_id(clamp_id) { 1915 uclamp_se_set(&init_task.uclamp_req[clamp_id], 1916 uclamp_none(clamp_id), false); 1917 } 1918 1919 /* System defaults allow max clamp values for both indexes */ 1920 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false); 1921 for_each_clamp_id(clamp_id) { 1922 uclamp_default[clamp_id] = uc_max; 1923 #ifdef CONFIG_UCLAMP_TASK_GROUP 1924 root_task_group.uclamp_req[clamp_id] = uc_max; 1925 root_task_group.uclamp[clamp_id] = uc_max; 1926 #endif 1927 } 1928 } 1929 1930 #else /* CONFIG_UCLAMP_TASK */ 1931 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { } 1932 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { } 1933 static inline int uclamp_validate(struct task_struct *p, 1934 const struct sched_attr *attr) 1935 { 1936 return -EOPNOTSUPP; 1937 } 1938 static void __setscheduler_uclamp(struct task_struct *p, 1939 const struct sched_attr *attr) { } 1940 static inline void uclamp_fork(struct task_struct *p) { } 1941 static inline void uclamp_post_fork(struct task_struct *p) { } 1942 static inline void init_uclamp(void) { } 1943 #endif /* CONFIG_UCLAMP_TASK */ 1944 1945 bool sched_task_on_rq(struct task_struct *p) 1946 { 1947 return task_on_rq_queued(p); 1948 } 1949 1950 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 1951 { 1952 if (!(flags & ENQUEUE_NOCLOCK)) 1953 update_rq_clock(rq); 1954 1955 if (!(flags & ENQUEUE_RESTORE)) { 1956 sched_info_enqueue(rq, p); 1957 psi_enqueue(p, flags & ENQUEUE_WAKEUP); 1958 } 1959 1960 uclamp_rq_inc(rq, p); 1961 p->sched_class->enqueue_task(rq, p, flags); 1962 1963 if (sched_core_enabled(rq)) 1964 sched_core_enqueue(rq, p); 1965 } 1966 1967 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 1968 { 1969 if (sched_core_enabled(rq)) 1970 sched_core_dequeue(rq, p); 1971 1972 if (!(flags & DEQUEUE_NOCLOCK)) 1973 update_rq_clock(rq); 1974 1975 if (!(flags & DEQUEUE_SAVE)) { 1976 sched_info_dequeue(rq, p); 1977 psi_dequeue(p, flags & DEQUEUE_SLEEP); 1978 } 1979 1980 uclamp_rq_dec(rq, p); 1981 p->sched_class->dequeue_task(rq, p, flags); 1982 } 1983 1984 void activate_task(struct rq *rq, struct task_struct *p, int flags) 1985 { 1986 enqueue_task(rq, p, flags); 1987 1988 p->on_rq = TASK_ON_RQ_QUEUED; 1989 } 1990 1991 void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 1992 { 1993 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING; 1994 1995 dequeue_task(rq, p, flags); 1996 } 1997 1998 static inline int __normal_prio(int policy, int rt_prio, int nice) 1999 { 2000 int prio; 2001 2002 if (dl_policy(policy)) 2003 prio = MAX_DL_PRIO - 1; 2004 else if (rt_policy(policy)) 2005 prio = MAX_RT_PRIO - 1 - rt_prio; 2006 else 2007 prio = NICE_TO_PRIO(nice); 2008 2009 return prio; 2010 } 2011 2012 /* 2013 * Calculate the expected normal priority: i.e. priority 2014 * without taking RT-inheritance into account. Might be 2015 * boosted by interactivity modifiers. Changes upon fork, 2016 * setprio syscalls, and whenever the interactivity 2017 * estimator recalculates. 2018 */ 2019 static inline int normal_prio(struct task_struct *p) 2020 { 2021 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio)); 2022 } 2023 2024 /* 2025 * Calculate the current priority, i.e. the priority 2026 * taken into account by the scheduler. This value might 2027 * be boosted by RT tasks, or might be boosted by 2028 * interactivity modifiers. Will be RT if the task got 2029 * RT-boosted. If not then it returns p->normal_prio. 2030 */ 2031 static int effective_prio(struct task_struct *p) 2032 { 2033 p->normal_prio = normal_prio(p); 2034 /* 2035 * If we are RT tasks or we were boosted to RT priority, 2036 * keep the priority unchanged. Otherwise, update priority 2037 * to the normal priority: 2038 */ 2039 if (!rt_prio(p->prio)) 2040 return p->normal_prio; 2041 return p->prio; 2042 } 2043 2044 /** 2045 * task_curr - is this task currently executing on a CPU? 2046 * @p: the task in question. 2047 * 2048 * Return: 1 if the task is currently executing. 0 otherwise. 2049 */ 2050 inline int task_curr(const struct task_struct *p) 2051 { 2052 return cpu_curr(task_cpu(p)) == p; 2053 } 2054 2055 /* 2056 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock, 2057 * use the balance_callback list if you want balancing. 2058 * 2059 * this means any call to check_class_changed() must be followed by a call to 2060 * balance_callback(). 2061 */ 2062 static inline void check_class_changed(struct rq *rq, struct task_struct *p, 2063 const struct sched_class *prev_class, 2064 int oldprio) 2065 { 2066 if (prev_class != p->sched_class) { 2067 if (prev_class->switched_from) 2068 prev_class->switched_from(rq, p); 2069 2070 p->sched_class->switched_to(rq, p); 2071 } else if (oldprio != p->prio || dl_task(p)) 2072 p->sched_class->prio_changed(rq, p, oldprio); 2073 } 2074 2075 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 2076 { 2077 if (p->sched_class == rq->curr->sched_class) 2078 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 2079 else if (p->sched_class > rq->curr->sched_class) 2080 resched_curr(rq); 2081 2082 /* 2083 * A queue event has occurred, and we're going to schedule. In 2084 * this case, we can save a useless back to back clock update. 2085 */ 2086 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) 2087 rq_clock_skip_update(rq); 2088 } 2089 2090 #ifdef CONFIG_SMP 2091 2092 static void 2093 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags); 2094 2095 static int __set_cpus_allowed_ptr(struct task_struct *p, 2096 const struct cpumask *new_mask, 2097 u32 flags); 2098 2099 static void migrate_disable_switch(struct rq *rq, struct task_struct *p) 2100 { 2101 if (likely(!p->migration_disabled)) 2102 return; 2103 2104 if (p->cpus_ptr != &p->cpus_mask) 2105 return; 2106 2107 /* 2108 * Violates locking rules! see comment in __do_set_cpus_allowed(). 2109 */ 2110 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE); 2111 } 2112 2113 void migrate_disable(void) 2114 { 2115 struct task_struct *p = current; 2116 2117 if (p->migration_disabled) { 2118 p->migration_disabled++; 2119 return; 2120 } 2121 2122 preempt_disable(); 2123 this_rq()->nr_pinned++; 2124 p->migration_disabled = 1; 2125 preempt_enable(); 2126 } 2127 EXPORT_SYMBOL_GPL(migrate_disable); 2128 2129 void migrate_enable(void) 2130 { 2131 struct task_struct *p = current; 2132 2133 if (p->migration_disabled > 1) { 2134 p->migration_disabled--; 2135 return; 2136 } 2137 2138 /* 2139 * Ensure stop_task runs either before or after this, and that 2140 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule(). 2141 */ 2142 preempt_disable(); 2143 if (p->cpus_ptr != &p->cpus_mask) 2144 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE); 2145 /* 2146 * Mustn't clear migration_disabled() until cpus_ptr points back at the 2147 * regular cpus_mask, otherwise things that race (eg. 2148 * select_fallback_rq) get confused. 2149 */ 2150 barrier(); 2151 p->migration_disabled = 0; 2152 this_rq()->nr_pinned--; 2153 preempt_enable(); 2154 } 2155 EXPORT_SYMBOL_GPL(migrate_enable); 2156 2157 static inline bool rq_has_pinned_tasks(struct rq *rq) 2158 { 2159 return rq->nr_pinned; 2160 } 2161 2162 /* 2163 * Per-CPU kthreads are allowed to run on !active && online CPUs, see 2164 * __set_cpus_allowed_ptr() and select_fallback_rq(). 2165 */ 2166 static inline bool is_cpu_allowed(struct task_struct *p, int cpu) 2167 { 2168 /* When not in the task's cpumask, no point in looking further. */ 2169 if (!cpumask_test_cpu(cpu, p->cpus_ptr)) 2170 return false; 2171 2172 /* migrate_disabled() must be allowed to finish. */ 2173 if (is_migration_disabled(p)) 2174 return cpu_online(cpu); 2175 2176 /* Non kernel threads are not allowed during either online or offline. */ 2177 if (!(p->flags & PF_KTHREAD)) 2178 return cpu_active(cpu); 2179 2180 /* KTHREAD_IS_PER_CPU is always allowed. */ 2181 if (kthread_is_per_cpu(p)) 2182 return cpu_online(cpu); 2183 2184 /* Regular kernel threads don't get to stay during offline. */ 2185 if (cpu_dying(cpu)) 2186 return false; 2187 2188 /* But are allowed during online. */ 2189 return cpu_online(cpu); 2190 } 2191 2192 /* 2193 * This is how migration works: 2194 * 2195 * 1) we invoke migration_cpu_stop() on the target CPU using 2196 * stop_one_cpu(). 2197 * 2) stopper starts to run (implicitly forcing the migrated thread 2198 * off the CPU) 2199 * 3) it checks whether the migrated task is still in the wrong runqueue. 2200 * 4) if it's in the wrong runqueue then the migration thread removes 2201 * it and puts it into the right queue. 2202 * 5) stopper completes and stop_one_cpu() returns and the migration 2203 * is done. 2204 */ 2205 2206 /* 2207 * move_queued_task - move a queued task to new rq. 2208 * 2209 * Returns (locked) new rq. Old rq's lock is released. 2210 */ 2211 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf, 2212 struct task_struct *p, int new_cpu) 2213 { 2214 lockdep_assert_rq_held(rq); 2215 2216 deactivate_task(rq, p, DEQUEUE_NOCLOCK); 2217 set_task_cpu(p, new_cpu); 2218 rq_unlock(rq, rf); 2219 2220 rq = cpu_rq(new_cpu); 2221 2222 rq_lock(rq, rf); 2223 BUG_ON(task_cpu(p) != new_cpu); 2224 activate_task(rq, p, 0); 2225 check_preempt_curr(rq, p, 0); 2226 2227 return rq; 2228 } 2229 2230 struct migration_arg { 2231 struct task_struct *task; 2232 int dest_cpu; 2233 struct set_affinity_pending *pending; 2234 }; 2235 2236 /* 2237 * @refs: number of wait_for_completion() 2238 * @stop_pending: is @stop_work in use 2239 */ 2240 struct set_affinity_pending { 2241 refcount_t refs; 2242 unsigned int stop_pending; 2243 struct completion done; 2244 struct cpu_stop_work stop_work; 2245 struct migration_arg arg; 2246 }; 2247 2248 /* 2249 * Move (not current) task off this CPU, onto the destination CPU. We're doing 2250 * this because either it can't run here any more (set_cpus_allowed() 2251 * away from this CPU, or CPU going down), or because we're 2252 * attempting to rebalance this task on exec (sched_exec). 2253 * 2254 * So we race with normal scheduler movements, but that's OK, as long 2255 * as the task is no longer on this CPU. 2256 */ 2257 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf, 2258 struct task_struct *p, int dest_cpu) 2259 { 2260 /* Affinity changed (again). */ 2261 if (!is_cpu_allowed(p, dest_cpu)) 2262 return rq; 2263 2264 update_rq_clock(rq); 2265 rq = move_queued_task(rq, rf, p, dest_cpu); 2266 2267 return rq; 2268 } 2269 2270 /* 2271 * migration_cpu_stop - this will be executed by a highprio stopper thread 2272 * and performs thread migration by bumping thread off CPU then 2273 * 'pushing' onto another runqueue. 2274 */ 2275 static int migration_cpu_stop(void *data) 2276 { 2277 struct migration_arg *arg = data; 2278 struct set_affinity_pending *pending = arg->pending; 2279 struct task_struct *p = arg->task; 2280 struct rq *rq = this_rq(); 2281 bool complete = false; 2282 struct rq_flags rf; 2283 2284 /* 2285 * The original target CPU might have gone down and we might 2286 * be on another CPU but it doesn't matter. 2287 */ 2288 local_irq_save(rf.flags); 2289 /* 2290 * We need to explicitly wake pending tasks before running 2291 * __migrate_task() such that we will not miss enforcing cpus_ptr 2292 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. 2293 */ 2294 flush_smp_call_function_from_idle(); 2295 2296 raw_spin_lock(&p->pi_lock); 2297 rq_lock(rq, &rf); 2298 2299 /* 2300 * If we were passed a pending, then ->stop_pending was set, thus 2301 * p->migration_pending must have remained stable. 2302 */ 2303 WARN_ON_ONCE(pending && pending != p->migration_pending); 2304 2305 /* 2306 * If task_rq(p) != rq, it cannot be migrated here, because we're 2307 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because 2308 * we're holding p->pi_lock. 2309 */ 2310 if (task_rq(p) == rq) { 2311 if (is_migration_disabled(p)) 2312 goto out; 2313 2314 if (pending) { 2315 p->migration_pending = NULL; 2316 complete = true; 2317 2318 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) 2319 goto out; 2320 } 2321 2322 if (task_on_rq_queued(p)) 2323 rq = __migrate_task(rq, &rf, p, arg->dest_cpu); 2324 else 2325 p->wake_cpu = arg->dest_cpu; 2326 2327 /* 2328 * XXX __migrate_task() can fail, at which point we might end 2329 * up running on a dodgy CPU, AFAICT this can only happen 2330 * during CPU hotplug, at which point we'll get pushed out 2331 * anyway, so it's probably not a big deal. 2332 */ 2333 2334 } else if (pending) { 2335 /* 2336 * This happens when we get migrated between migrate_enable()'s 2337 * preempt_enable() and scheduling the stopper task. At that 2338 * point we're a regular task again and not current anymore. 2339 * 2340 * A !PREEMPT kernel has a giant hole here, which makes it far 2341 * more likely. 2342 */ 2343 2344 /* 2345 * The task moved before the stopper got to run. We're holding 2346 * ->pi_lock, so the allowed mask is stable - if it got 2347 * somewhere allowed, we're done. 2348 */ 2349 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) { 2350 p->migration_pending = NULL; 2351 complete = true; 2352 goto out; 2353 } 2354 2355 /* 2356 * When migrate_enable() hits a rq mis-match we can't reliably 2357 * determine is_migration_disabled() and so have to chase after 2358 * it. 2359 */ 2360 WARN_ON_ONCE(!pending->stop_pending); 2361 task_rq_unlock(rq, p, &rf); 2362 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop, 2363 &pending->arg, &pending->stop_work); 2364 return 0; 2365 } 2366 out: 2367 if (pending) 2368 pending->stop_pending = false; 2369 task_rq_unlock(rq, p, &rf); 2370 2371 if (complete) 2372 complete_all(&pending->done); 2373 2374 return 0; 2375 } 2376 2377 int push_cpu_stop(void *arg) 2378 { 2379 struct rq *lowest_rq = NULL, *rq = this_rq(); 2380 struct task_struct *p = arg; 2381 2382 raw_spin_lock_irq(&p->pi_lock); 2383 raw_spin_rq_lock(rq); 2384 2385 if (task_rq(p) != rq) 2386 goto out_unlock; 2387 2388 if (is_migration_disabled(p)) { 2389 p->migration_flags |= MDF_PUSH; 2390 goto out_unlock; 2391 } 2392 2393 p->migration_flags &= ~MDF_PUSH; 2394 2395 if (p->sched_class->find_lock_rq) 2396 lowest_rq = p->sched_class->find_lock_rq(p, rq); 2397 2398 if (!lowest_rq) 2399 goto out_unlock; 2400 2401 // XXX validate p is still the highest prio task 2402 if (task_rq(p) == rq) { 2403 deactivate_task(rq, p, 0); 2404 set_task_cpu(p, lowest_rq->cpu); 2405 activate_task(lowest_rq, p, 0); 2406 resched_curr(lowest_rq); 2407 } 2408 2409 double_unlock_balance(rq, lowest_rq); 2410 2411 out_unlock: 2412 rq->push_busy = false; 2413 raw_spin_rq_unlock(rq); 2414 raw_spin_unlock_irq(&p->pi_lock); 2415 2416 put_task_struct(p); 2417 return 0; 2418 } 2419 2420 /* 2421 * sched_class::set_cpus_allowed must do the below, but is not required to 2422 * actually call this function. 2423 */ 2424 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags) 2425 { 2426 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) { 2427 p->cpus_ptr = new_mask; 2428 return; 2429 } 2430 2431 cpumask_copy(&p->cpus_mask, new_mask); 2432 p->nr_cpus_allowed = cpumask_weight(new_mask); 2433 } 2434 2435 static void 2436 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags) 2437 { 2438 struct rq *rq = task_rq(p); 2439 bool queued, running; 2440 2441 /* 2442 * This here violates the locking rules for affinity, since we're only 2443 * supposed to change these variables while holding both rq->lock and 2444 * p->pi_lock. 2445 * 2446 * HOWEVER, it magically works, because ttwu() is the only code that 2447 * accesses these variables under p->pi_lock and only does so after 2448 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule() 2449 * before finish_task(). 2450 * 2451 * XXX do further audits, this smells like something putrid. 2452 */ 2453 if (flags & SCA_MIGRATE_DISABLE) 2454 SCHED_WARN_ON(!p->on_cpu); 2455 else 2456 lockdep_assert_held(&p->pi_lock); 2457 2458 queued = task_on_rq_queued(p); 2459 running = task_current(rq, p); 2460 2461 if (queued) { 2462 /* 2463 * Because __kthread_bind() calls this on blocked tasks without 2464 * holding rq->lock. 2465 */ 2466 lockdep_assert_rq_held(rq); 2467 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK); 2468 } 2469 if (running) 2470 put_prev_task(rq, p); 2471 2472 p->sched_class->set_cpus_allowed(p, new_mask, flags); 2473 2474 if (queued) 2475 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); 2476 if (running) 2477 set_next_task(rq, p); 2478 } 2479 2480 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 2481 { 2482 __do_set_cpus_allowed(p, new_mask, 0); 2483 } 2484 2485 /* 2486 * This function is wildly self concurrent; here be dragons. 2487 * 2488 * 2489 * When given a valid mask, __set_cpus_allowed_ptr() must block until the 2490 * designated task is enqueued on an allowed CPU. If that task is currently 2491 * running, we have to kick it out using the CPU stopper. 2492 * 2493 * Migrate-Disable comes along and tramples all over our nice sandcastle. 2494 * Consider: 2495 * 2496 * Initial conditions: P0->cpus_mask = [0, 1] 2497 * 2498 * P0@CPU0 P1 2499 * 2500 * migrate_disable(); 2501 * <preempted> 2502 * set_cpus_allowed_ptr(P0, [1]); 2503 * 2504 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes 2505 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region). 2506 * This means we need the following scheme: 2507 * 2508 * P0@CPU0 P1 2509 * 2510 * migrate_disable(); 2511 * <preempted> 2512 * set_cpus_allowed_ptr(P0, [1]); 2513 * <blocks> 2514 * <resumes> 2515 * migrate_enable(); 2516 * __set_cpus_allowed_ptr(); 2517 * <wakes local stopper> 2518 * `--> <woken on migration completion> 2519 * 2520 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple 2521 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any 2522 * task p are serialized by p->pi_lock, which we can leverage: the one that 2523 * should come into effect at the end of the Migrate-Disable region is the last 2524 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask), 2525 * but we still need to properly signal those waiting tasks at the appropriate 2526 * moment. 2527 * 2528 * This is implemented using struct set_affinity_pending. The first 2529 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will 2530 * setup an instance of that struct and install it on the targeted task_struct. 2531 * Any and all further callers will reuse that instance. Those then wait for 2532 * a completion signaled at the tail of the CPU stopper callback (1), triggered 2533 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()). 2534 * 2535 * 2536 * (1) In the cases covered above. There is one more where the completion is 2537 * signaled within affine_move_task() itself: when a subsequent affinity request 2538 * occurs after the stopper bailed out due to the targeted task still being 2539 * Migrate-Disable. Consider: 2540 * 2541 * Initial conditions: P0->cpus_mask = [0, 1] 2542 * 2543 * CPU0 P1 P2 2544 * <P0> 2545 * migrate_disable(); 2546 * <preempted> 2547 * set_cpus_allowed_ptr(P0, [1]); 2548 * <blocks> 2549 * <migration/0> 2550 * migration_cpu_stop() 2551 * is_migration_disabled() 2552 * <bails> 2553 * set_cpus_allowed_ptr(P0, [0, 1]); 2554 * <signal completion> 2555 * <awakes> 2556 * 2557 * Note that the above is safe vs a concurrent migrate_enable(), as any 2558 * pending affinity completion is preceded by an uninstallation of 2559 * p->migration_pending done with p->pi_lock held. 2560 */ 2561 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf, 2562 int dest_cpu, unsigned int flags) 2563 { 2564 struct set_affinity_pending my_pending = { }, *pending = NULL; 2565 bool stop_pending, complete = false; 2566 2567 /* Can the task run on the task's current CPU? If so, we're done */ 2568 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) { 2569 struct task_struct *push_task = NULL; 2570 2571 if ((flags & SCA_MIGRATE_ENABLE) && 2572 (p->migration_flags & MDF_PUSH) && !rq->push_busy) { 2573 rq->push_busy = true; 2574 push_task = get_task_struct(p); 2575 } 2576 2577 /* 2578 * If there are pending waiters, but no pending stop_work, 2579 * then complete now. 2580 */ 2581 pending = p->migration_pending; 2582 if (pending && !pending->stop_pending) { 2583 p->migration_pending = NULL; 2584 complete = true; 2585 } 2586 2587 task_rq_unlock(rq, p, rf); 2588 2589 if (push_task) { 2590 stop_one_cpu_nowait(rq->cpu, push_cpu_stop, 2591 p, &rq->push_work); 2592 } 2593 2594 if (complete) 2595 complete_all(&pending->done); 2596 2597 return 0; 2598 } 2599 2600 if (!(flags & SCA_MIGRATE_ENABLE)) { 2601 /* serialized by p->pi_lock */ 2602 if (!p->migration_pending) { 2603 /* Install the request */ 2604 refcount_set(&my_pending.refs, 1); 2605 init_completion(&my_pending.done); 2606 my_pending.arg = (struct migration_arg) { 2607 .task = p, 2608 .dest_cpu = dest_cpu, 2609 .pending = &my_pending, 2610 }; 2611 2612 p->migration_pending = &my_pending; 2613 } else { 2614 pending = p->migration_pending; 2615 refcount_inc(&pending->refs); 2616 /* 2617 * Affinity has changed, but we've already installed a 2618 * pending. migration_cpu_stop() *must* see this, else 2619 * we risk a completion of the pending despite having a 2620 * task on a disallowed CPU. 2621 * 2622 * Serialized by p->pi_lock, so this is safe. 2623 */ 2624 pending->arg.dest_cpu = dest_cpu; 2625 } 2626 } 2627 pending = p->migration_pending; 2628 /* 2629 * - !MIGRATE_ENABLE: 2630 * we'll have installed a pending if there wasn't one already. 2631 * 2632 * - MIGRATE_ENABLE: 2633 * we're here because the current CPU isn't matching anymore, 2634 * the only way that can happen is because of a concurrent 2635 * set_cpus_allowed_ptr() call, which should then still be 2636 * pending completion. 2637 * 2638 * Either way, we really should have a @pending here. 2639 */ 2640 if (WARN_ON_ONCE(!pending)) { 2641 task_rq_unlock(rq, p, rf); 2642 return -EINVAL; 2643 } 2644 2645 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) { 2646 /* 2647 * MIGRATE_ENABLE gets here because 'p == current', but for 2648 * anything else we cannot do is_migration_disabled(), punt 2649 * and have the stopper function handle it all race-free. 2650 */ 2651 stop_pending = pending->stop_pending; 2652 if (!stop_pending) 2653 pending->stop_pending = true; 2654 2655 if (flags & SCA_MIGRATE_ENABLE) 2656 p->migration_flags &= ~MDF_PUSH; 2657 2658 task_rq_unlock(rq, p, rf); 2659 2660 if (!stop_pending) { 2661 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop, 2662 &pending->arg, &pending->stop_work); 2663 } 2664 2665 if (flags & SCA_MIGRATE_ENABLE) 2666 return 0; 2667 } else { 2668 2669 if (!is_migration_disabled(p)) { 2670 if (task_on_rq_queued(p)) 2671 rq = move_queued_task(rq, rf, p, dest_cpu); 2672 2673 if (!pending->stop_pending) { 2674 p->migration_pending = NULL; 2675 complete = true; 2676 } 2677 } 2678 task_rq_unlock(rq, p, rf); 2679 2680 if (complete) 2681 complete_all(&pending->done); 2682 } 2683 2684 wait_for_completion(&pending->done); 2685 2686 if (refcount_dec_and_test(&pending->refs)) 2687 wake_up_var(&pending->refs); /* No UaF, just an address */ 2688 2689 /* 2690 * Block the original owner of &pending until all subsequent callers 2691 * have seen the completion and decremented the refcount 2692 */ 2693 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs)); 2694 2695 /* ARGH */ 2696 WARN_ON_ONCE(my_pending.stop_pending); 2697 2698 return 0; 2699 } 2700 2701 /* 2702 * Change a given task's CPU affinity. Migrate the thread to a 2703 * proper CPU and schedule it away if the CPU it's executing on 2704 * is removed from the allowed bitmask. 2705 * 2706 * NOTE: the caller must have a valid reference to the task, the 2707 * task must not exit() & deallocate itself prematurely. The 2708 * call is not atomic; no spinlocks may be held. 2709 */ 2710 static int __set_cpus_allowed_ptr(struct task_struct *p, 2711 const struct cpumask *new_mask, 2712 u32 flags) 2713 { 2714 const struct cpumask *cpu_valid_mask = cpu_active_mask; 2715 unsigned int dest_cpu; 2716 struct rq_flags rf; 2717 struct rq *rq; 2718 int ret = 0; 2719 2720 rq = task_rq_lock(p, &rf); 2721 update_rq_clock(rq); 2722 2723 if (p->flags & PF_KTHREAD || is_migration_disabled(p)) { 2724 /* 2725 * Kernel threads are allowed on online && !active CPUs, 2726 * however, during cpu-hot-unplug, even these might get pushed 2727 * away if not KTHREAD_IS_PER_CPU. 2728 * 2729 * Specifically, migration_disabled() tasks must not fail the 2730 * cpumask_any_and_distribute() pick below, esp. so on 2731 * SCA_MIGRATE_ENABLE, otherwise we'll not call 2732 * set_cpus_allowed_common() and actually reset p->cpus_ptr. 2733 */ 2734 cpu_valid_mask = cpu_online_mask; 2735 } 2736 2737 /* 2738 * Must re-check here, to close a race against __kthread_bind(), 2739 * sched_setaffinity() is not guaranteed to observe the flag. 2740 */ 2741 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) { 2742 ret = -EINVAL; 2743 goto out; 2744 } 2745 2746 if (!(flags & SCA_MIGRATE_ENABLE)) { 2747 if (cpumask_equal(&p->cpus_mask, new_mask)) 2748 goto out; 2749 2750 if (WARN_ON_ONCE(p == current && 2751 is_migration_disabled(p) && 2752 !cpumask_test_cpu(task_cpu(p), new_mask))) { 2753 ret = -EBUSY; 2754 goto out; 2755 } 2756 } 2757 2758 /* 2759 * Picking a ~random cpu helps in cases where we are changing affinity 2760 * for groups of tasks (ie. cpuset), so that load balancing is not 2761 * immediately required to distribute the tasks within their new mask. 2762 */ 2763 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask); 2764 if (dest_cpu >= nr_cpu_ids) { 2765 ret = -EINVAL; 2766 goto out; 2767 } 2768 2769 __do_set_cpus_allowed(p, new_mask, flags); 2770 2771 return affine_move_task(rq, p, &rf, dest_cpu, flags); 2772 2773 out: 2774 task_rq_unlock(rq, p, &rf); 2775 2776 return ret; 2777 } 2778 2779 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 2780 { 2781 return __set_cpus_allowed_ptr(p, new_mask, 0); 2782 } 2783 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 2784 2785 void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 2786 { 2787 #ifdef CONFIG_SCHED_DEBUG 2788 unsigned int state = READ_ONCE(p->__state); 2789 2790 /* 2791 * We should never call set_task_cpu() on a blocked task, 2792 * ttwu() will sort out the placement. 2793 */ 2794 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq); 2795 2796 /* 2797 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING, 2798 * because schedstat_wait_{start,end} rebase migrating task's wait_start 2799 * time relying on p->on_rq. 2800 */ 2801 WARN_ON_ONCE(state == TASK_RUNNING && 2802 p->sched_class == &fair_sched_class && 2803 (p->on_rq && !task_on_rq_migrating(p))); 2804 2805 #ifdef CONFIG_LOCKDEP 2806 /* 2807 * The caller should hold either p->pi_lock or rq->lock, when changing 2808 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 2809 * 2810 * sched_move_task() holds both and thus holding either pins the cgroup, 2811 * see task_group(). 2812 * 2813 * Furthermore, all task_rq users should acquire both locks, see 2814 * task_rq_lock(). 2815 */ 2816 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 2817 lockdep_is_held(__rq_lockp(task_rq(p))))); 2818 #endif 2819 /* 2820 * Clearly, migrating tasks to offline CPUs is a fairly daft thing. 2821 */ 2822 WARN_ON_ONCE(!cpu_online(new_cpu)); 2823 2824 WARN_ON_ONCE(is_migration_disabled(p)); 2825 #endif 2826 2827 trace_sched_migrate_task(p, new_cpu); 2828 2829 if (task_cpu(p) != new_cpu) { 2830 if (p->sched_class->migrate_task_rq) 2831 p->sched_class->migrate_task_rq(p, new_cpu); 2832 p->se.nr_migrations++; 2833 rseq_migrate(p); 2834 perf_event_task_migrate(p); 2835 } 2836 2837 __set_task_cpu(p, new_cpu); 2838 } 2839 2840 #ifdef CONFIG_NUMA_BALANCING 2841 static void __migrate_swap_task(struct task_struct *p, int cpu) 2842 { 2843 if (task_on_rq_queued(p)) { 2844 struct rq *src_rq, *dst_rq; 2845 struct rq_flags srf, drf; 2846 2847 src_rq = task_rq(p); 2848 dst_rq = cpu_rq(cpu); 2849 2850 rq_pin_lock(src_rq, &srf); 2851 rq_pin_lock(dst_rq, &drf); 2852 2853 deactivate_task(src_rq, p, 0); 2854 set_task_cpu(p, cpu); 2855 activate_task(dst_rq, p, 0); 2856 check_preempt_curr(dst_rq, p, 0); 2857 2858 rq_unpin_lock(dst_rq, &drf); 2859 rq_unpin_lock(src_rq, &srf); 2860 2861 } else { 2862 /* 2863 * Task isn't running anymore; make it appear like we migrated 2864 * it before it went to sleep. This means on wakeup we make the 2865 * previous CPU our target instead of where it really is. 2866 */ 2867 p->wake_cpu = cpu; 2868 } 2869 } 2870 2871 struct migration_swap_arg { 2872 struct task_struct *src_task, *dst_task; 2873 int src_cpu, dst_cpu; 2874 }; 2875 2876 static int migrate_swap_stop(void *data) 2877 { 2878 struct migration_swap_arg *arg = data; 2879 struct rq *src_rq, *dst_rq; 2880 int ret = -EAGAIN; 2881 2882 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu)) 2883 return -EAGAIN; 2884 2885 src_rq = cpu_rq(arg->src_cpu); 2886 dst_rq = cpu_rq(arg->dst_cpu); 2887 2888 double_raw_lock(&arg->src_task->pi_lock, 2889 &arg->dst_task->pi_lock); 2890 double_rq_lock(src_rq, dst_rq); 2891 2892 if (task_cpu(arg->dst_task) != arg->dst_cpu) 2893 goto unlock; 2894 2895 if (task_cpu(arg->src_task) != arg->src_cpu) 2896 goto unlock; 2897 2898 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr)) 2899 goto unlock; 2900 2901 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr)) 2902 goto unlock; 2903 2904 __migrate_swap_task(arg->src_task, arg->dst_cpu); 2905 __migrate_swap_task(arg->dst_task, arg->src_cpu); 2906 2907 ret = 0; 2908 2909 unlock: 2910 double_rq_unlock(src_rq, dst_rq); 2911 raw_spin_unlock(&arg->dst_task->pi_lock); 2912 raw_spin_unlock(&arg->src_task->pi_lock); 2913 2914 return ret; 2915 } 2916 2917 /* 2918 * Cross migrate two tasks 2919 */ 2920 int migrate_swap(struct task_struct *cur, struct task_struct *p, 2921 int target_cpu, int curr_cpu) 2922 { 2923 struct migration_swap_arg arg; 2924 int ret = -EINVAL; 2925 2926 arg = (struct migration_swap_arg){ 2927 .src_task = cur, 2928 .src_cpu = curr_cpu, 2929 .dst_task = p, 2930 .dst_cpu = target_cpu, 2931 }; 2932 2933 if (arg.src_cpu == arg.dst_cpu) 2934 goto out; 2935 2936 /* 2937 * These three tests are all lockless; this is OK since all of them 2938 * will be re-checked with proper locks held further down the line. 2939 */ 2940 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 2941 goto out; 2942 2943 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr)) 2944 goto out; 2945 2946 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr)) 2947 goto out; 2948 2949 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 2950 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 2951 2952 out: 2953 return ret; 2954 } 2955 #endif /* CONFIG_NUMA_BALANCING */ 2956 2957 /* 2958 * wait_task_inactive - wait for a thread to unschedule. 2959 * 2960 * If @match_state is nonzero, it's the @p->state value just checked and 2961 * not expected to change. If it changes, i.e. @p might have woken up, 2962 * then return zero. When we succeed in waiting for @p to be off its CPU, 2963 * we return a positive number (its total switch count). If a second call 2964 * a short while later returns the same number, the caller can be sure that 2965 * @p has remained unscheduled the whole time. 2966 * 2967 * The caller must ensure that the task *will* unschedule sometime soon, 2968 * else this function might spin for a *long* time. This function can't 2969 * be called with interrupts off, or it may introduce deadlock with 2970 * smp_call_function() if an IPI is sent by the same process we are 2971 * waiting to become inactive. 2972 */ 2973 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state) 2974 { 2975 int running, queued; 2976 struct rq_flags rf; 2977 unsigned long ncsw; 2978 struct rq *rq; 2979 2980 for (;;) { 2981 /* 2982 * We do the initial early heuristics without holding 2983 * any task-queue locks at all. We'll only try to get 2984 * the runqueue lock when things look like they will 2985 * work out! 2986 */ 2987 rq = task_rq(p); 2988 2989 /* 2990 * If the task is actively running on another CPU 2991 * still, just relax and busy-wait without holding 2992 * any locks. 2993 * 2994 * NOTE! Since we don't hold any locks, it's not 2995 * even sure that "rq" stays as the right runqueue! 2996 * But we don't care, since "task_running()" will 2997 * return false if the runqueue has changed and p 2998 * is actually now running somewhere else! 2999 */ 3000 while (task_running(rq, p)) { 3001 if (match_state && unlikely(READ_ONCE(p->__state) != match_state)) 3002 return 0; 3003 cpu_relax(); 3004 } 3005 3006 /* 3007 * Ok, time to look more closely! We need the rq 3008 * lock now, to be *sure*. If we're wrong, we'll 3009 * just go back and repeat. 3010 */ 3011 rq = task_rq_lock(p, &rf); 3012 trace_sched_wait_task(p); 3013 running = task_running(rq, p); 3014 queued = task_on_rq_queued(p); 3015 ncsw = 0; 3016 if (!match_state || READ_ONCE(p->__state) == match_state) 3017 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 3018 task_rq_unlock(rq, p, &rf); 3019 3020 /* 3021 * If it changed from the expected state, bail out now. 3022 */ 3023 if (unlikely(!ncsw)) 3024 break; 3025 3026 /* 3027 * Was it really running after all now that we 3028 * checked with the proper locks actually held? 3029 * 3030 * Oops. Go back and try again.. 3031 */ 3032 if (unlikely(running)) { 3033 cpu_relax(); 3034 continue; 3035 } 3036 3037 /* 3038 * It's not enough that it's not actively running, 3039 * it must be off the runqueue _entirely_, and not 3040 * preempted! 3041 * 3042 * So if it was still runnable (but just not actively 3043 * running right now), it's preempted, and we should 3044 * yield - it could be a while. 3045 */ 3046 if (unlikely(queued)) { 3047 ktime_t to = NSEC_PER_SEC / HZ; 3048 3049 set_current_state(TASK_UNINTERRUPTIBLE); 3050 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 3051 continue; 3052 } 3053 3054 /* 3055 * Ahh, all good. It wasn't running, and it wasn't 3056 * runnable, which means that it will never become 3057 * running in the future either. We're all done! 3058 */ 3059 break; 3060 } 3061 3062 return ncsw; 3063 } 3064 3065 /*** 3066 * kick_process - kick a running thread to enter/exit the kernel 3067 * @p: the to-be-kicked thread 3068 * 3069 * Cause a process which is running on another CPU to enter 3070 * kernel-mode, without any delay. (to get signals handled.) 3071 * 3072 * NOTE: this function doesn't have to take the runqueue lock, 3073 * because all it wants to ensure is that the remote task enters 3074 * the kernel. If the IPI races and the task has been migrated 3075 * to another CPU then no harm is done and the purpose has been 3076 * achieved as well. 3077 */ 3078 void kick_process(struct task_struct *p) 3079 { 3080 int cpu; 3081 3082 preempt_disable(); 3083 cpu = task_cpu(p); 3084 if ((cpu != smp_processor_id()) && task_curr(p)) 3085 smp_send_reschedule(cpu); 3086 preempt_enable(); 3087 } 3088 EXPORT_SYMBOL_GPL(kick_process); 3089 3090 /* 3091 * ->cpus_ptr is protected by both rq->lock and p->pi_lock 3092 * 3093 * A few notes on cpu_active vs cpu_online: 3094 * 3095 * - cpu_active must be a subset of cpu_online 3096 * 3097 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU, 3098 * see __set_cpus_allowed_ptr(). At this point the newly online 3099 * CPU isn't yet part of the sched domains, and balancing will not 3100 * see it. 3101 * 3102 * - on CPU-down we clear cpu_active() to mask the sched domains and 3103 * avoid the load balancer to place new tasks on the to be removed 3104 * CPU. Existing tasks will remain running there and will be taken 3105 * off. 3106 * 3107 * This means that fallback selection must not select !active CPUs. 3108 * And can assume that any active CPU must be online. Conversely 3109 * select_task_rq() below may allow selection of !active CPUs in order 3110 * to satisfy the above rules. 3111 */ 3112 static int select_fallback_rq(int cpu, struct task_struct *p) 3113 { 3114 int nid = cpu_to_node(cpu); 3115 const struct cpumask *nodemask = NULL; 3116 enum { cpuset, possible, fail } state = cpuset; 3117 int dest_cpu; 3118 3119 /* 3120 * If the node that the CPU is on has been offlined, cpu_to_node() 3121 * will return -1. There is no CPU on the node, and we should 3122 * select the CPU on the other node. 3123 */ 3124 if (nid != -1) { 3125 nodemask = cpumask_of_node(nid); 3126 3127 /* Look for allowed, online CPU in same node. */ 3128 for_each_cpu(dest_cpu, nodemask) { 3129 if (!cpu_active(dest_cpu)) 3130 continue; 3131 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr)) 3132 return dest_cpu; 3133 } 3134 } 3135 3136 for (;;) { 3137 /* Any allowed, online CPU? */ 3138 for_each_cpu(dest_cpu, p->cpus_ptr) { 3139 if (!is_cpu_allowed(p, dest_cpu)) 3140 continue; 3141 3142 goto out; 3143 } 3144 3145 /* No more Mr. Nice Guy. */ 3146 switch (state) { 3147 case cpuset: 3148 if (IS_ENABLED(CONFIG_CPUSETS)) { 3149 cpuset_cpus_allowed_fallback(p); 3150 state = possible; 3151 break; 3152 } 3153 fallthrough; 3154 case possible: 3155 /* 3156 * XXX When called from select_task_rq() we only 3157 * hold p->pi_lock and again violate locking order. 3158 * 3159 * More yuck to audit. 3160 */ 3161 do_set_cpus_allowed(p, cpu_possible_mask); 3162 state = fail; 3163 break; 3164 3165 case fail: 3166 BUG(); 3167 break; 3168 } 3169 } 3170 3171 out: 3172 if (state != cpuset) { 3173 /* 3174 * Don't tell them about moving exiting tasks or 3175 * kernel threads (both mm NULL), since they never 3176 * leave kernel. 3177 */ 3178 if (p->mm && printk_ratelimit()) { 3179 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 3180 task_pid_nr(p), p->comm, cpu); 3181 } 3182 } 3183 3184 return dest_cpu; 3185 } 3186 3187 /* 3188 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable. 3189 */ 3190 static inline 3191 int select_task_rq(struct task_struct *p, int cpu, int wake_flags) 3192 { 3193 lockdep_assert_held(&p->pi_lock); 3194 3195 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p)) 3196 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags); 3197 else 3198 cpu = cpumask_any(p->cpus_ptr); 3199 3200 /* 3201 * In order not to call set_task_cpu() on a blocking task we need 3202 * to rely on ttwu() to place the task on a valid ->cpus_ptr 3203 * CPU. 3204 * 3205 * Since this is common to all placement strategies, this lives here. 3206 * 3207 * [ this allows ->select_task() to simply return task_cpu(p) and 3208 * not worry about this generic constraint ] 3209 */ 3210 if (unlikely(!is_cpu_allowed(p, cpu))) 3211 cpu = select_fallback_rq(task_cpu(p), p); 3212 3213 return cpu; 3214 } 3215 3216 void sched_set_stop_task(int cpu, struct task_struct *stop) 3217 { 3218 static struct lock_class_key stop_pi_lock; 3219 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 3220 struct task_struct *old_stop = cpu_rq(cpu)->stop; 3221 3222 if (stop) { 3223 /* 3224 * Make it appear like a SCHED_FIFO task, its something 3225 * userspace knows about and won't get confused about. 3226 * 3227 * Also, it will make PI more or less work without too 3228 * much confusion -- but then, stop work should not 3229 * rely on PI working anyway. 3230 */ 3231 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 3232 3233 stop->sched_class = &stop_sched_class; 3234 3235 /* 3236 * The PI code calls rt_mutex_setprio() with ->pi_lock held to 3237 * adjust the effective priority of a task. As a result, 3238 * rt_mutex_setprio() can trigger (RT) balancing operations, 3239 * which can then trigger wakeups of the stop thread to push 3240 * around the current task. 3241 * 3242 * The stop task itself will never be part of the PI-chain, it 3243 * never blocks, therefore that ->pi_lock recursion is safe. 3244 * Tell lockdep about this by placing the stop->pi_lock in its 3245 * own class. 3246 */ 3247 lockdep_set_class(&stop->pi_lock, &stop_pi_lock); 3248 } 3249 3250 cpu_rq(cpu)->stop = stop; 3251 3252 if (old_stop) { 3253 /* 3254 * Reset it back to a normal scheduling class so that 3255 * it can die in pieces. 3256 */ 3257 old_stop->sched_class = &rt_sched_class; 3258 } 3259 } 3260 3261 #else /* CONFIG_SMP */ 3262 3263 static inline int __set_cpus_allowed_ptr(struct task_struct *p, 3264 const struct cpumask *new_mask, 3265 u32 flags) 3266 { 3267 return set_cpus_allowed_ptr(p, new_mask); 3268 } 3269 3270 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { } 3271 3272 static inline bool rq_has_pinned_tasks(struct rq *rq) 3273 { 3274 return false; 3275 } 3276 3277 #endif /* !CONFIG_SMP */ 3278 3279 static void 3280 ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 3281 { 3282 struct rq *rq; 3283 3284 if (!schedstat_enabled()) 3285 return; 3286 3287 rq = this_rq(); 3288 3289 #ifdef CONFIG_SMP 3290 if (cpu == rq->cpu) { 3291 __schedstat_inc(rq->ttwu_local); 3292 __schedstat_inc(p->se.statistics.nr_wakeups_local); 3293 } else { 3294 struct sched_domain *sd; 3295 3296 __schedstat_inc(p->se.statistics.nr_wakeups_remote); 3297 rcu_read_lock(); 3298 for_each_domain(rq->cpu, sd) { 3299 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 3300 __schedstat_inc(sd->ttwu_wake_remote); 3301 break; 3302 } 3303 } 3304 rcu_read_unlock(); 3305 } 3306 3307 if (wake_flags & WF_MIGRATED) 3308 __schedstat_inc(p->se.statistics.nr_wakeups_migrate); 3309 #endif /* CONFIG_SMP */ 3310 3311 __schedstat_inc(rq->ttwu_count); 3312 __schedstat_inc(p->se.statistics.nr_wakeups); 3313 3314 if (wake_flags & WF_SYNC) 3315 __schedstat_inc(p->se.statistics.nr_wakeups_sync); 3316 } 3317 3318 /* 3319 * Mark the task runnable and perform wakeup-preemption. 3320 */ 3321 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags, 3322 struct rq_flags *rf) 3323 { 3324 check_preempt_curr(rq, p, wake_flags); 3325 WRITE_ONCE(p->__state, TASK_RUNNING); 3326 trace_sched_wakeup(p); 3327 3328 #ifdef CONFIG_SMP 3329 if (p->sched_class->task_woken) { 3330 /* 3331 * Our task @p is fully woken up and running; so it's safe to 3332 * drop the rq->lock, hereafter rq is only used for statistics. 3333 */ 3334 rq_unpin_lock(rq, rf); 3335 p->sched_class->task_woken(rq, p); 3336 rq_repin_lock(rq, rf); 3337 } 3338 3339 if (rq->idle_stamp) { 3340 u64 delta = rq_clock(rq) - rq->idle_stamp; 3341 u64 max = 2*rq->max_idle_balance_cost; 3342 3343 update_avg(&rq->avg_idle, delta); 3344 3345 if (rq->avg_idle > max) 3346 rq->avg_idle = max; 3347 3348 rq->wake_stamp = jiffies; 3349 rq->wake_avg_idle = rq->avg_idle / 2; 3350 3351 rq->idle_stamp = 0; 3352 } 3353 #endif 3354 } 3355 3356 static void 3357 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags, 3358 struct rq_flags *rf) 3359 { 3360 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK; 3361 3362 lockdep_assert_rq_held(rq); 3363 3364 if (p->sched_contributes_to_load) 3365 rq->nr_uninterruptible--; 3366 3367 #ifdef CONFIG_SMP 3368 if (wake_flags & WF_MIGRATED) 3369 en_flags |= ENQUEUE_MIGRATED; 3370 else 3371 #endif 3372 if (p->in_iowait) { 3373 delayacct_blkio_end(p); 3374 atomic_dec(&task_rq(p)->nr_iowait); 3375 } 3376 3377 activate_task(rq, p, en_flags); 3378 ttwu_do_wakeup(rq, p, wake_flags, rf); 3379 } 3380 3381 /* 3382 * Consider @p being inside a wait loop: 3383 * 3384 * for (;;) { 3385 * set_current_state(TASK_UNINTERRUPTIBLE); 3386 * 3387 * if (CONDITION) 3388 * break; 3389 * 3390 * schedule(); 3391 * } 3392 * __set_current_state(TASK_RUNNING); 3393 * 3394 * between set_current_state() and schedule(). In this case @p is still 3395 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in 3396 * an atomic manner. 3397 * 3398 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq 3399 * then schedule() must still happen and p->state can be changed to 3400 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we 3401 * need to do a full wakeup with enqueue. 3402 * 3403 * Returns: %true when the wakeup is done, 3404 * %false otherwise. 3405 */ 3406 static int ttwu_runnable(struct task_struct *p, int wake_flags) 3407 { 3408 struct rq_flags rf; 3409 struct rq *rq; 3410 int ret = 0; 3411 3412 rq = __task_rq_lock(p, &rf); 3413 if (task_on_rq_queued(p)) { 3414 /* check_preempt_curr() may use rq clock */ 3415 update_rq_clock(rq); 3416 ttwu_do_wakeup(rq, p, wake_flags, &rf); 3417 ret = 1; 3418 } 3419 __task_rq_unlock(rq, &rf); 3420 3421 return ret; 3422 } 3423 3424 #ifdef CONFIG_SMP 3425 void sched_ttwu_pending(void *arg) 3426 { 3427 struct llist_node *llist = arg; 3428 struct rq *rq = this_rq(); 3429 struct task_struct *p, *t; 3430 struct rq_flags rf; 3431 3432 if (!llist) 3433 return; 3434 3435 /* 3436 * rq::ttwu_pending racy indication of out-standing wakeups. 3437 * Races such that false-negatives are possible, since they 3438 * are shorter lived that false-positives would be. 3439 */ 3440 WRITE_ONCE(rq->ttwu_pending, 0); 3441 3442 rq_lock_irqsave(rq, &rf); 3443 update_rq_clock(rq); 3444 3445 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) { 3446 if (WARN_ON_ONCE(p->on_cpu)) 3447 smp_cond_load_acquire(&p->on_cpu, !VAL); 3448 3449 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq))) 3450 set_task_cpu(p, cpu_of(rq)); 3451 3452 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf); 3453 } 3454 3455 rq_unlock_irqrestore(rq, &rf); 3456 } 3457 3458 void send_call_function_single_ipi(int cpu) 3459 { 3460 struct rq *rq = cpu_rq(cpu); 3461 3462 if (!set_nr_if_polling(rq->idle)) 3463 arch_send_call_function_single_ipi(cpu); 3464 else 3465 trace_sched_wake_idle_without_ipi(cpu); 3466 } 3467 3468 /* 3469 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if 3470 * necessary. The wakee CPU on receipt of the IPI will queue the task 3471 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost 3472 * of the wakeup instead of the waker. 3473 */ 3474 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) 3475 { 3476 struct rq *rq = cpu_rq(cpu); 3477 3478 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED); 3479 3480 WRITE_ONCE(rq->ttwu_pending, 1); 3481 __smp_call_single_queue(cpu, &p->wake_entry.llist); 3482 } 3483 3484 void wake_up_if_idle(int cpu) 3485 { 3486 struct rq *rq = cpu_rq(cpu); 3487 struct rq_flags rf; 3488 3489 rcu_read_lock(); 3490 3491 if (!is_idle_task(rcu_dereference(rq->curr))) 3492 goto out; 3493 3494 if (set_nr_if_polling(rq->idle)) { 3495 trace_sched_wake_idle_without_ipi(cpu); 3496 } else { 3497 rq_lock_irqsave(rq, &rf); 3498 if (is_idle_task(rq->curr)) 3499 smp_send_reschedule(cpu); 3500 /* Else CPU is not idle, do nothing here: */ 3501 rq_unlock_irqrestore(rq, &rf); 3502 } 3503 3504 out: 3505 rcu_read_unlock(); 3506 } 3507 3508 bool cpus_share_cache(int this_cpu, int that_cpu) 3509 { 3510 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 3511 } 3512 3513 static inline bool ttwu_queue_cond(int cpu, int wake_flags) 3514 { 3515 /* 3516 * Do not complicate things with the async wake_list while the CPU is 3517 * in hotplug state. 3518 */ 3519 if (!cpu_active(cpu)) 3520 return false; 3521 3522 /* 3523 * If the CPU does not share cache, then queue the task on the 3524 * remote rqs wakelist to avoid accessing remote data. 3525 */ 3526 if (!cpus_share_cache(smp_processor_id(), cpu)) 3527 return true; 3528 3529 /* 3530 * If the task is descheduling and the only running task on the 3531 * CPU then use the wakelist to offload the task activation to 3532 * the soon-to-be-idle CPU as the current CPU is likely busy. 3533 * nr_running is checked to avoid unnecessary task stacking. 3534 */ 3535 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1) 3536 return true; 3537 3538 return false; 3539 } 3540 3541 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) 3542 { 3543 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) { 3544 if (WARN_ON_ONCE(cpu == smp_processor_id())) 3545 return false; 3546 3547 sched_clock_cpu(cpu); /* Sync clocks across CPUs */ 3548 __ttwu_queue_wakelist(p, cpu, wake_flags); 3549 return true; 3550 } 3551 3552 return false; 3553 } 3554 3555 #else /* !CONFIG_SMP */ 3556 3557 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) 3558 { 3559 return false; 3560 } 3561 3562 #endif /* CONFIG_SMP */ 3563 3564 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags) 3565 { 3566 struct rq *rq = cpu_rq(cpu); 3567 struct rq_flags rf; 3568 3569 if (ttwu_queue_wakelist(p, cpu, wake_flags)) 3570 return; 3571 3572 rq_lock(rq, &rf); 3573 update_rq_clock(rq); 3574 ttwu_do_activate(rq, p, wake_flags, &rf); 3575 rq_unlock(rq, &rf); 3576 } 3577 3578 /* 3579 * Notes on Program-Order guarantees on SMP systems. 3580 * 3581 * MIGRATION 3582 * 3583 * The basic program-order guarantee on SMP systems is that when a task [t] 3584 * migrates, all its activity on its old CPU [c0] happens-before any subsequent 3585 * execution on its new CPU [c1]. 3586 * 3587 * For migration (of runnable tasks) this is provided by the following means: 3588 * 3589 * A) UNLOCK of the rq(c0)->lock scheduling out task t 3590 * B) migration for t is required to synchronize *both* rq(c0)->lock and 3591 * rq(c1)->lock (if not at the same time, then in that order). 3592 * C) LOCK of the rq(c1)->lock scheduling in task 3593 * 3594 * Release/acquire chaining guarantees that B happens after A and C after B. 3595 * Note: the CPU doing B need not be c0 or c1 3596 * 3597 * Example: 3598 * 3599 * CPU0 CPU1 CPU2 3600 * 3601 * LOCK rq(0)->lock 3602 * sched-out X 3603 * sched-in Y 3604 * UNLOCK rq(0)->lock 3605 * 3606 * LOCK rq(0)->lock // orders against CPU0 3607 * dequeue X 3608 * UNLOCK rq(0)->lock 3609 * 3610 * LOCK rq(1)->lock 3611 * enqueue X 3612 * UNLOCK rq(1)->lock 3613 * 3614 * LOCK rq(1)->lock // orders against CPU2 3615 * sched-out Z 3616 * sched-in X 3617 * UNLOCK rq(1)->lock 3618 * 3619 * 3620 * BLOCKING -- aka. SLEEP + WAKEUP 3621 * 3622 * For blocking we (obviously) need to provide the same guarantee as for 3623 * migration. However the means are completely different as there is no lock 3624 * chain to provide order. Instead we do: 3625 * 3626 * 1) smp_store_release(X->on_cpu, 0) -- finish_task() 3627 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up() 3628 * 3629 * Example: 3630 * 3631 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule) 3632 * 3633 * LOCK rq(0)->lock LOCK X->pi_lock 3634 * dequeue X 3635 * sched-out X 3636 * smp_store_release(X->on_cpu, 0); 3637 * 3638 * smp_cond_load_acquire(&X->on_cpu, !VAL); 3639 * X->state = WAKING 3640 * set_task_cpu(X,2) 3641 * 3642 * LOCK rq(2)->lock 3643 * enqueue X 3644 * X->state = RUNNING 3645 * UNLOCK rq(2)->lock 3646 * 3647 * LOCK rq(2)->lock // orders against CPU1 3648 * sched-out Z 3649 * sched-in X 3650 * UNLOCK rq(2)->lock 3651 * 3652 * UNLOCK X->pi_lock 3653 * UNLOCK rq(0)->lock 3654 * 3655 * 3656 * However, for wakeups there is a second guarantee we must provide, namely we 3657 * must ensure that CONDITION=1 done by the caller can not be reordered with 3658 * accesses to the task state; see try_to_wake_up() and set_current_state(). 3659 */ 3660 3661 /** 3662 * try_to_wake_up - wake up a thread 3663 * @p: the thread to be awakened 3664 * @state: the mask of task states that can be woken 3665 * @wake_flags: wake modifier flags (WF_*) 3666 * 3667 * Conceptually does: 3668 * 3669 * If (@state & @p->state) @p->state = TASK_RUNNING. 3670 * 3671 * If the task was not queued/runnable, also place it back on a runqueue. 3672 * 3673 * This function is atomic against schedule() which would dequeue the task. 3674 * 3675 * It issues a full memory barrier before accessing @p->state, see the comment 3676 * with set_current_state(). 3677 * 3678 * Uses p->pi_lock to serialize against concurrent wake-ups. 3679 * 3680 * Relies on p->pi_lock stabilizing: 3681 * - p->sched_class 3682 * - p->cpus_ptr 3683 * - p->sched_task_group 3684 * in order to do migration, see its use of select_task_rq()/set_task_cpu(). 3685 * 3686 * Tries really hard to only take one task_rq(p)->lock for performance. 3687 * Takes rq->lock in: 3688 * - ttwu_runnable() -- old rq, unavoidable, see comment there; 3689 * - ttwu_queue() -- new rq, for enqueue of the task; 3690 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us. 3691 * 3692 * As a consequence we race really badly with just about everything. See the 3693 * many memory barriers and their comments for details. 3694 * 3695 * Return: %true if @p->state changes (an actual wakeup was done), 3696 * %false otherwise. 3697 */ 3698 static int 3699 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 3700 { 3701 unsigned long flags; 3702 int cpu, success = 0; 3703 3704 preempt_disable(); 3705 if (p == current) { 3706 /* 3707 * We're waking current, this means 'p->on_rq' and 'task_cpu(p) 3708 * == smp_processor_id()'. Together this means we can special 3709 * case the whole 'p->on_rq && ttwu_runnable()' case below 3710 * without taking any locks. 3711 * 3712 * In particular: 3713 * - we rely on Program-Order guarantees for all the ordering, 3714 * - we're serialized against set_special_state() by virtue of 3715 * it disabling IRQs (this allows not taking ->pi_lock). 3716 */ 3717 if (!(READ_ONCE(p->__state) & state)) 3718 goto out; 3719 3720 success = 1; 3721 trace_sched_waking(p); 3722 WRITE_ONCE(p->__state, TASK_RUNNING); 3723 trace_sched_wakeup(p); 3724 goto out; 3725 } 3726 3727 /* 3728 * If we are going to wake up a thread waiting for CONDITION we 3729 * need to ensure that CONDITION=1 done by the caller can not be 3730 * reordered with p->state check below. This pairs with smp_store_mb() 3731 * in set_current_state() that the waiting thread does. 3732 */ 3733 raw_spin_lock_irqsave(&p->pi_lock, flags); 3734 smp_mb__after_spinlock(); 3735 if (!(READ_ONCE(p->__state) & state)) 3736 goto unlock; 3737 3738 trace_sched_waking(p); 3739 3740 /* We're going to change ->state: */ 3741 success = 1; 3742 3743 /* 3744 * Ensure we load p->on_rq _after_ p->state, otherwise it would 3745 * be possible to, falsely, observe p->on_rq == 0 and get stuck 3746 * in smp_cond_load_acquire() below. 3747 * 3748 * sched_ttwu_pending() try_to_wake_up() 3749 * STORE p->on_rq = 1 LOAD p->state 3750 * UNLOCK rq->lock 3751 * 3752 * __schedule() (switch to task 'p') 3753 * LOCK rq->lock smp_rmb(); 3754 * smp_mb__after_spinlock(); 3755 * UNLOCK rq->lock 3756 * 3757 * [task p] 3758 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq 3759 * 3760 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in 3761 * __schedule(). See the comment for smp_mb__after_spinlock(). 3762 * 3763 * A similar smb_rmb() lives in try_invoke_on_locked_down_task(). 3764 */ 3765 smp_rmb(); 3766 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags)) 3767 goto unlock; 3768 3769 #ifdef CONFIG_SMP 3770 /* 3771 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be 3772 * possible to, falsely, observe p->on_cpu == 0. 3773 * 3774 * One must be running (->on_cpu == 1) in order to remove oneself 3775 * from the runqueue. 3776 * 3777 * __schedule() (switch to task 'p') try_to_wake_up() 3778 * STORE p->on_cpu = 1 LOAD p->on_rq 3779 * UNLOCK rq->lock 3780 * 3781 * __schedule() (put 'p' to sleep) 3782 * LOCK rq->lock smp_rmb(); 3783 * smp_mb__after_spinlock(); 3784 * STORE p->on_rq = 0 LOAD p->on_cpu 3785 * 3786 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in 3787 * __schedule(). See the comment for smp_mb__after_spinlock(). 3788 * 3789 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure 3790 * schedule()'s deactivate_task() has 'happened' and p will no longer 3791 * care about it's own p->state. See the comment in __schedule(). 3792 */ 3793 smp_acquire__after_ctrl_dep(); 3794 3795 /* 3796 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq 3797 * == 0), which means we need to do an enqueue, change p->state to 3798 * TASK_WAKING such that we can unlock p->pi_lock before doing the 3799 * enqueue, such as ttwu_queue_wakelist(). 3800 */ 3801 WRITE_ONCE(p->__state, TASK_WAKING); 3802 3803 /* 3804 * If the owning (remote) CPU is still in the middle of schedule() with 3805 * this task as prev, considering queueing p on the remote CPUs wake_list 3806 * which potentially sends an IPI instead of spinning on p->on_cpu to 3807 * let the waker make forward progress. This is safe because IRQs are 3808 * disabled and the IPI will deliver after on_cpu is cleared. 3809 * 3810 * Ensure we load task_cpu(p) after p->on_cpu: 3811 * 3812 * set_task_cpu(p, cpu); 3813 * STORE p->cpu = @cpu 3814 * __schedule() (switch to task 'p') 3815 * LOCK rq->lock 3816 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu) 3817 * STORE p->on_cpu = 1 LOAD p->cpu 3818 * 3819 * to ensure we observe the correct CPU on which the task is currently 3820 * scheduling. 3821 */ 3822 if (smp_load_acquire(&p->on_cpu) && 3823 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU)) 3824 goto unlock; 3825 3826 /* 3827 * If the owning (remote) CPU is still in the middle of schedule() with 3828 * this task as prev, wait until it's done referencing the task. 3829 * 3830 * Pairs with the smp_store_release() in finish_task(). 3831 * 3832 * This ensures that tasks getting woken will be fully ordered against 3833 * their previous state and preserve Program Order. 3834 */ 3835 smp_cond_load_acquire(&p->on_cpu, !VAL); 3836 3837 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU); 3838 if (task_cpu(p) != cpu) { 3839 if (p->in_iowait) { 3840 delayacct_blkio_end(p); 3841 atomic_dec(&task_rq(p)->nr_iowait); 3842 } 3843 3844 wake_flags |= WF_MIGRATED; 3845 psi_ttwu_dequeue(p); 3846 set_task_cpu(p, cpu); 3847 } 3848 #else 3849 cpu = task_cpu(p); 3850 #endif /* CONFIG_SMP */ 3851 3852 ttwu_queue(p, cpu, wake_flags); 3853 unlock: 3854 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3855 out: 3856 if (success) 3857 ttwu_stat(p, task_cpu(p), wake_flags); 3858 preempt_enable(); 3859 3860 return success; 3861 } 3862 3863 /** 3864 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state 3865 * @p: Process for which the function is to be invoked, can be @current. 3866 * @func: Function to invoke. 3867 * @arg: Argument to function. 3868 * 3869 * If the specified task can be quickly locked into a definite state 3870 * (either sleeping or on a given runqueue), arrange to keep it in that 3871 * state while invoking @func(@arg). This function can use ->on_rq and 3872 * task_curr() to work out what the state is, if required. Given that 3873 * @func can be invoked with a runqueue lock held, it had better be quite 3874 * lightweight. 3875 * 3876 * Returns: 3877 * @false if the task slipped out from under the locks. 3878 * @true if the task was locked onto a runqueue or is sleeping. 3879 * However, @func can override this by returning @false. 3880 */ 3881 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg) 3882 { 3883 struct rq_flags rf; 3884 bool ret = false; 3885 struct rq *rq; 3886 3887 raw_spin_lock_irqsave(&p->pi_lock, rf.flags); 3888 if (p->on_rq) { 3889 rq = __task_rq_lock(p, &rf); 3890 if (task_rq(p) == rq) 3891 ret = func(p, arg); 3892 rq_unlock(rq, &rf); 3893 } else { 3894 switch (READ_ONCE(p->__state)) { 3895 case TASK_RUNNING: 3896 case TASK_WAKING: 3897 break; 3898 default: 3899 smp_rmb(); // See smp_rmb() comment in try_to_wake_up(). 3900 if (!p->on_rq) 3901 ret = func(p, arg); 3902 } 3903 } 3904 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags); 3905 return ret; 3906 } 3907 3908 /** 3909 * wake_up_process - Wake up a specific process 3910 * @p: The process to be woken up. 3911 * 3912 * Attempt to wake up the nominated process and move it to the set of runnable 3913 * processes. 3914 * 3915 * Return: 1 if the process was woken up, 0 if it was already running. 3916 * 3917 * This function executes a full memory barrier before accessing the task state. 3918 */ 3919 int wake_up_process(struct task_struct *p) 3920 { 3921 return try_to_wake_up(p, TASK_NORMAL, 0); 3922 } 3923 EXPORT_SYMBOL(wake_up_process); 3924 3925 int wake_up_state(struct task_struct *p, unsigned int state) 3926 { 3927 return try_to_wake_up(p, state, 0); 3928 } 3929 3930 /* 3931 * Perform scheduler related setup for a newly forked process p. 3932 * p is forked by current. 3933 * 3934 * __sched_fork() is basic setup used by init_idle() too: 3935 */ 3936 static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 3937 { 3938 p->on_rq = 0; 3939 3940 p->se.on_rq = 0; 3941 p->se.exec_start = 0; 3942 p->se.sum_exec_runtime = 0; 3943 p->se.prev_sum_exec_runtime = 0; 3944 p->se.nr_migrations = 0; 3945 p->se.vruntime = 0; 3946 INIT_LIST_HEAD(&p->se.group_node); 3947 3948 #ifdef CONFIG_FAIR_GROUP_SCHED 3949 p->se.cfs_rq = NULL; 3950 #endif 3951 3952 #ifdef CONFIG_SCHEDSTATS 3953 /* Even if schedstat is disabled, there should not be garbage */ 3954 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 3955 #endif 3956 3957 RB_CLEAR_NODE(&p->dl.rb_node); 3958 init_dl_task_timer(&p->dl); 3959 init_dl_inactive_task_timer(&p->dl); 3960 __dl_clear_params(p); 3961 3962 INIT_LIST_HEAD(&p->rt.run_list); 3963 p->rt.timeout = 0; 3964 p->rt.time_slice = sched_rr_timeslice; 3965 p->rt.on_rq = 0; 3966 p->rt.on_list = 0; 3967 3968 #ifdef CONFIG_PREEMPT_NOTIFIERS 3969 INIT_HLIST_HEAD(&p->preempt_notifiers); 3970 #endif 3971 3972 #ifdef CONFIG_COMPACTION 3973 p->capture_control = NULL; 3974 #endif 3975 init_numa_balancing(clone_flags, p); 3976 #ifdef CONFIG_SMP 3977 p->wake_entry.u_flags = CSD_TYPE_TTWU; 3978 p->migration_pending = NULL; 3979 #endif 3980 } 3981 3982 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing); 3983 3984 #ifdef CONFIG_NUMA_BALANCING 3985 3986 void set_numabalancing_state(bool enabled) 3987 { 3988 if (enabled) 3989 static_branch_enable(&sched_numa_balancing); 3990 else 3991 static_branch_disable(&sched_numa_balancing); 3992 } 3993 3994 #ifdef CONFIG_PROC_SYSCTL 3995 int sysctl_numa_balancing(struct ctl_table *table, int write, 3996 void *buffer, size_t *lenp, loff_t *ppos) 3997 { 3998 struct ctl_table t; 3999 int err; 4000 int state = static_branch_likely(&sched_numa_balancing); 4001 4002 if (write && !capable(CAP_SYS_ADMIN)) 4003 return -EPERM; 4004 4005 t = *table; 4006 t.data = &state; 4007 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 4008 if (err < 0) 4009 return err; 4010 if (write) 4011 set_numabalancing_state(state); 4012 return err; 4013 } 4014 #endif 4015 #endif 4016 4017 #ifdef CONFIG_SCHEDSTATS 4018 4019 DEFINE_STATIC_KEY_FALSE(sched_schedstats); 4020 4021 static void set_schedstats(bool enabled) 4022 { 4023 if (enabled) 4024 static_branch_enable(&sched_schedstats); 4025 else 4026 static_branch_disable(&sched_schedstats); 4027 } 4028 4029 void force_schedstat_enabled(void) 4030 { 4031 if (!schedstat_enabled()) { 4032 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n"); 4033 static_branch_enable(&sched_schedstats); 4034 } 4035 } 4036 4037 static int __init setup_schedstats(char *str) 4038 { 4039 int ret = 0; 4040 if (!str) 4041 goto out; 4042 4043 if (!strcmp(str, "enable")) { 4044 set_schedstats(true); 4045 ret = 1; 4046 } else if (!strcmp(str, "disable")) { 4047 set_schedstats(false); 4048 ret = 1; 4049 } 4050 out: 4051 if (!ret) 4052 pr_warn("Unable to parse schedstats=\n"); 4053 4054 return ret; 4055 } 4056 __setup("schedstats=", setup_schedstats); 4057 4058 #ifdef CONFIG_PROC_SYSCTL 4059 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer, 4060 size_t *lenp, loff_t *ppos) 4061 { 4062 struct ctl_table t; 4063 int err; 4064 int state = static_branch_likely(&sched_schedstats); 4065 4066 if (write && !capable(CAP_SYS_ADMIN)) 4067 return -EPERM; 4068 4069 t = *table; 4070 t.data = &state; 4071 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 4072 if (err < 0) 4073 return err; 4074 if (write) 4075 set_schedstats(state); 4076 return err; 4077 } 4078 #endif /* CONFIG_PROC_SYSCTL */ 4079 #endif /* CONFIG_SCHEDSTATS */ 4080 4081 /* 4082 * fork()/clone()-time setup: 4083 */ 4084 int sched_fork(unsigned long clone_flags, struct task_struct *p) 4085 { 4086 unsigned long flags; 4087 4088 __sched_fork(clone_flags, p); 4089 /* 4090 * We mark the process as NEW here. This guarantees that 4091 * nobody will actually run it, and a signal or other external 4092 * event cannot wake it up and insert it on the runqueue either. 4093 */ 4094 p->__state = TASK_NEW; 4095 4096 /* 4097 * Make sure we do not leak PI boosting priority to the child. 4098 */ 4099 p->prio = current->normal_prio; 4100 4101 uclamp_fork(p); 4102 4103 /* 4104 * Revert to default priority/policy on fork if requested. 4105 */ 4106 if (unlikely(p->sched_reset_on_fork)) { 4107 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 4108 p->policy = SCHED_NORMAL; 4109 p->static_prio = NICE_TO_PRIO(0); 4110 p->rt_priority = 0; 4111 } else if (PRIO_TO_NICE(p->static_prio) < 0) 4112 p->static_prio = NICE_TO_PRIO(0); 4113 4114 p->prio = p->normal_prio = p->static_prio; 4115 set_load_weight(p, false); 4116 4117 /* 4118 * We don't need the reset flag anymore after the fork. It has 4119 * fulfilled its duty: 4120 */ 4121 p->sched_reset_on_fork = 0; 4122 } 4123 4124 if (dl_prio(p->prio)) 4125 return -EAGAIN; 4126 else if (rt_prio(p->prio)) 4127 p->sched_class = &rt_sched_class; 4128 else 4129 p->sched_class = &fair_sched_class; 4130 4131 init_entity_runnable_average(&p->se); 4132 4133 /* 4134 * The child is not yet in the pid-hash so no cgroup attach races, 4135 * and the cgroup is pinned to this child due to cgroup_fork() 4136 * is ran before sched_fork(). 4137 * 4138 * Silence PROVE_RCU. 4139 */ 4140 raw_spin_lock_irqsave(&p->pi_lock, flags); 4141 rseq_migrate(p); 4142 /* 4143 * We're setting the CPU for the first time, we don't migrate, 4144 * so use __set_task_cpu(). 4145 */ 4146 __set_task_cpu(p, smp_processor_id()); 4147 if (p->sched_class->task_fork) 4148 p->sched_class->task_fork(p); 4149 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4150 4151 #ifdef CONFIG_SCHED_INFO 4152 if (likely(sched_info_on())) 4153 memset(&p->sched_info, 0, sizeof(p->sched_info)); 4154 #endif 4155 #if defined(CONFIG_SMP) 4156 p->on_cpu = 0; 4157 #endif 4158 init_task_preempt_count(p); 4159 #ifdef CONFIG_SMP 4160 plist_node_init(&p->pushable_tasks, MAX_PRIO); 4161 RB_CLEAR_NODE(&p->pushable_dl_tasks); 4162 #endif 4163 return 0; 4164 } 4165 4166 void sched_post_fork(struct task_struct *p) 4167 { 4168 uclamp_post_fork(p); 4169 } 4170 4171 unsigned long to_ratio(u64 period, u64 runtime) 4172 { 4173 if (runtime == RUNTIME_INF) 4174 return BW_UNIT; 4175 4176 /* 4177 * Doing this here saves a lot of checks in all 4178 * the calling paths, and returning zero seems 4179 * safe for them anyway. 4180 */ 4181 if (period == 0) 4182 return 0; 4183 4184 return div64_u64(runtime << BW_SHIFT, period); 4185 } 4186 4187 /* 4188 * wake_up_new_task - wake up a newly created task for the first time. 4189 * 4190 * This function will do some initial scheduler statistics housekeeping 4191 * that must be done for every newly created context, then puts the task 4192 * on the runqueue and wakes it. 4193 */ 4194 void wake_up_new_task(struct task_struct *p) 4195 { 4196 struct rq_flags rf; 4197 struct rq *rq; 4198 4199 raw_spin_lock_irqsave(&p->pi_lock, rf.flags); 4200 WRITE_ONCE(p->__state, TASK_RUNNING); 4201 #ifdef CONFIG_SMP 4202 /* 4203 * Fork balancing, do it here and not earlier because: 4204 * - cpus_ptr can change in the fork path 4205 * - any previously selected CPU might disappear through hotplug 4206 * 4207 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq, 4208 * as we're not fully set-up yet. 4209 */ 4210 p->recent_used_cpu = task_cpu(p); 4211 rseq_migrate(p); 4212 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK)); 4213 #endif 4214 rq = __task_rq_lock(p, &rf); 4215 update_rq_clock(rq); 4216 post_init_entity_util_avg(p); 4217 4218 activate_task(rq, p, ENQUEUE_NOCLOCK); 4219 trace_sched_wakeup_new(p); 4220 check_preempt_curr(rq, p, WF_FORK); 4221 #ifdef CONFIG_SMP 4222 if (p->sched_class->task_woken) { 4223 /* 4224 * Nothing relies on rq->lock after this, so it's fine to 4225 * drop it. 4226 */ 4227 rq_unpin_lock(rq, &rf); 4228 p->sched_class->task_woken(rq, p); 4229 rq_repin_lock(rq, &rf); 4230 } 4231 #endif 4232 task_rq_unlock(rq, p, &rf); 4233 } 4234 4235 #ifdef CONFIG_PREEMPT_NOTIFIERS 4236 4237 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key); 4238 4239 void preempt_notifier_inc(void) 4240 { 4241 static_branch_inc(&preempt_notifier_key); 4242 } 4243 EXPORT_SYMBOL_GPL(preempt_notifier_inc); 4244 4245 void preempt_notifier_dec(void) 4246 { 4247 static_branch_dec(&preempt_notifier_key); 4248 } 4249 EXPORT_SYMBOL_GPL(preempt_notifier_dec); 4250 4251 /** 4252 * preempt_notifier_register - tell me when current is being preempted & rescheduled 4253 * @notifier: notifier struct to register 4254 */ 4255 void preempt_notifier_register(struct preempt_notifier *notifier) 4256 { 4257 if (!static_branch_unlikely(&preempt_notifier_key)) 4258 WARN(1, "registering preempt_notifier while notifiers disabled\n"); 4259 4260 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 4261 } 4262 EXPORT_SYMBOL_GPL(preempt_notifier_register); 4263 4264 /** 4265 * preempt_notifier_unregister - no longer interested in preemption notifications 4266 * @notifier: notifier struct to unregister 4267 * 4268 * This is *not* safe to call from within a preemption notifier. 4269 */ 4270 void preempt_notifier_unregister(struct preempt_notifier *notifier) 4271 { 4272 hlist_del(¬ifier->link); 4273 } 4274 EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 4275 4276 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr) 4277 { 4278 struct preempt_notifier *notifier; 4279 4280 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 4281 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 4282 } 4283 4284 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) 4285 { 4286 if (static_branch_unlikely(&preempt_notifier_key)) 4287 __fire_sched_in_preempt_notifiers(curr); 4288 } 4289 4290 static void 4291 __fire_sched_out_preempt_notifiers(struct task_struct *curr, 4292 struct task_struct *next) 4293 { 4294 struct preempt_notifier *notifier; 4295 4296 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 4297 notifier->ops->sched_out(notifier, next); 4298 } 4299 4300 static __always_inline void 4301 fire_sched_out_preempt_notifiers(struct task_struct *curr, 4302 struct task_struct *next) 4303 { 4304 if (static_branch_unlikely(&preempt_notifier_key)) 4305 __fire_sched_out_preempt_notifiers(curr, next); 4306 } 4307 4308 #else /* !CONFIG_PREEMPT_NOTIFIERS */ 4309 4310 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) 4311 { 4312 } 4313 4314 static inline void 4315 fire_sched_out_preempt_notifiers(struct task_struct *curr, 4316 struct task_struct *next) 4317 { 4318 } 4319 4320 #endif /* CONFIG_PREEMPT_NOTIFIERS */ 4321 4322 static inline void prepare_task(struct task_struct *next) 4323 { 4324 #ifdef CONFIG_SMP 4325 /* 4326 * Claim the task as running, we do this before switching to it 4327 * such that any running task will have this set. 4328 * 4329 * See the ttwu() WF_ON_CPU case and its ordering comment. 4330 */ 4331 WRITE_ONCE(next->on_cpu, 1); 4332 #endif 4333 } 4334 4335 static inline void finish_task(struct task_struct *prev) 4336 { 4337 #ifdef CONFIG_SMP 4338 /* 4339 * This must be the very last reference to @prev from this CPU. After 4340 * p->on_cpu is cleared, the task can be moved to a different CPU. We 4341 * must ensure this doesn't happen until the switch is completely 4342 * finished. 4343 * 4344 * In particular, the load of prev->state in finish_task_switch() must 4345 * happen before this. 4346 * 4347 * Pairs with the smp_cond_load_acquire() in try_to_wake_up(). 4348 */ 4349 smp_store_release(&prev->on_cpu, 0); 4350 #endif 4351 } 4352 4353 #ifdef CONFIG_SMP 4354 4355 static void do_balance_callbacks(struct rq *rq, struct callback_head *head) 4356 { 4357 void (*func)(struct rq *rq); 4358 struct callback_head *next; 4359 4360 lockdep_assert_rq_held(rq); 4361 4362 while (head) { 4363 func = (void (*)(struct rq *))head->func; 4364 next = head->next; 4365 head->next = NULL; 4366 head = next; 4367 4368 func(rq); 4369 } 4370 } 4371 4372 static void balance_push(struct rq *rq); 4373 4374 struct callback_head balance_push_callback = { 4375 .next = NULL, 4376 .func = (void (*)(struct callback_head *))balance_push, 4377 }; 4378 4379 static inline struct callback_head *splice_balance_callbacks(struct rq *rq) 4380 { 4381 struct callback_head *head = rq->balance_callback; 4382 4383 lockdep_assert_rq_held(rq); 4384 if (head) 4385 rq->balance_callback = NULL; 4386 4387 return head; 4388 } 4389 4390 static void __balance_callbacks(struct rq *rq) 4391 { 4392 do_balance_callbacks(rq, splice_balance_callbacks(rq)); 4393 } 4394 4395 static inline void balance_callbacks(struct rq *rq, struct callback_head *head) 4396 { 4397 unsigned long flags; 4398 4399 if (unlikely(head)) { 4400 raw_spin_rq_lock_irqsave(rq, flags); 4401 do_balance_callbacks(rq, head); 4402 raw_spin_rq_unlock_irqrestore(rq, flags); 4403 } 4404 } 4405 4406 #else 4407 4408 static inline void __balance_callbacks(struct rq *rq) 4409 { 4410 } 4411 4412 static inline struct callback_head *splice_balance_callbacks(struct rq *rq) 4413 { 4414 return NULL; 4415 } 4416 4417 static inline void balance_callbacks(struct rq *rq, struct callback_head *head) 4418 { 4419 } 4420 4421 #endif 4422 4423 static inline void 4424 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf) 4425 { 4426 /* 4427 * Since the runqueue lock will be released by the next 4428 * task (which is an invalid locking op but in the case 4429 * of the scheduler it's an obvious special-case), so we 4430 * do an early lockdep release here: 4431 */ 4432 rq_unpin_lock(rq, rf); 4433 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_); 4434 #ifdef CONFIG_DEBUG_SPINLOCK 4435 /* this is a valid case when another task releases the spinlock */ 4436 rq_lockp(rq)->owner = next; 4437 #endif 4438 } 4439 4440 static inline void finish_lock_switch(struct rq *rq) 4441 { 4442 /* 4443 * If we are tracking spinlock dependencies then we have to 4444 * fix up the runqueue lock - which gets 'carried over' from 4445 * prev into current: 4446 */ 4447 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_); 4448 __balance_callbacks(rq); 4449 raw_spin_rq_unlock_irq(rq); 4450 } 4451 4452 /* 4453 * NOP if the arch has not defined these: 4454 */ 4455 4456 #ifndef prepare_arch_switch 4457 # define prepare_arch_switch(next) do { } while (0) 4458 #endif 4459 4460 #ifndef finish_arch_post_lock_switch 4461 # define finish_arch_post_lock_switch() do { } while (0) 4462 #endif 4463 4464 static inline void kmap_local_sched_out(void) 4465 { 4466 #ifdef CONFIG_KMAP_LOCAL 4467 if (unlikely(current->kmap_ctrl.idx)) 4468 __kmap_local_sched_out(); 4469 #endif 4470 } 4471 4472 static inline void kmap_local_sched_in(void) 4473 { 4474 #ifdef CONFIG_KMAP_LOCAL 4475 if (unlikely(current->kmap_ctrl.idx)) 4476 __kmap_local_sched_in(); 4477 #endif 4478 } 4479 4480 /** 4481 * prepare_task_switch - prepare to switch tasks 4482 * @rq: the runqueue preparing to switch 4483 * @prev: the current task that is being switched out 4484 * @next: the task we are going to switch to. 4485 * 4486 * This is called with the rq lock held and interrupts off. It must 4487 * be paired with a subsequent finish_task_switch after the context 4488 * switch. 4489 * 4490 * prepare_task_switch sets up locking and calls architecture specific 4491 * hooks. 4492 */ 4493 static inline void 4494 prepare_task_switch(struct rq *rq, struct task_struct *prev, 4495 struct task_struct *next) 4496 { 4497 kcov_prepare_switch(prev); 4498 sched_info_switch(rq, prev, next); 4499 perf_event_task_sched_out(prev, next); 4500 rseq_preempt(prev); 4501 fire_sched_out_preempt_notifiers(prev, next); 4502 kmap_local_sched_out(); 4503 prepare_task(next); 4504 prepare_arch_switch(next); 4505 } 4506 4507 /** 4508 * finish_task_switch - clean up after a task-switch 4509 * @prev: the thread we just switched away from. 4510 * 4511 * finish_task_switch must be called after the context switch, paired 4512 * with a prepare_task_switch call before the context switch. 4513 * finish_task_switch will reconcile locking set up by prepare_task_switch, 4514 * and do any other architecture-specific cleanup actions. 4515 * 4516 * Note that we may have delayed dropping an mm in context_switch(). If 4517 * so, we finish that here outside of the runqueue lock. (Doing it 4518 * with the lock held can cause deadlocks; see schedule() for 4519 * details.) 4520 * 4521 * The context switch have flipped the stack from under us and restored the 4522 * local variables which were saved when this task called schedule() in the 4523 * past. prev == current is still correct but we need to recalculate this_rq 4524 * because prev may have moved to another CPU. 4525 */ 4526 static struct rq *finish_task_switch(struct task_struct *prev) 4527 __releases(rq->lock) 4528 { 4529 struct rq *rq = this_rq(); 4530 struct mm_struct *mm = rq->prev_mm; 4531 long prev_state; 4532 4533 /* 4534 * The previous task will have left us with a preempt_count of 2 4535 * because it left us after: 4536 * 4537 * schedule() 4538 * preempt_disable(); // 1 4539 * __schedule() 4540 * raw_spin_lock_irq(&rq->lock) // 2 4541 * 4542 * Also, see FORK_PREEMPT_COUNT. 4543 */ 4544 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET, 4545 "corrupted preempt_count: %s/%d/0x%x\n", 4546 current->comm, current->pid, preempt_count())) 4547 preempt_count_set(FORK_PREEMPT_COUNT); 4548 4549 rq->prev_mm = NULL; 4550 4551 /* 4552 * A task struct has one reference for the use as "current". 4553 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 4554 * schedule one last time. The schedule call will never return, and 4555 * the scheduled task must drop that reference. 4556 * 4557 * We must observe prev->state before clearing prev->on_cpu (in 4558 * finish_task), otherwise a concurrent wakeup can get prev 4559 * running on another CPU and we could rave with its RUNNING -> DEAD 4560 * transition, resulting in a double drop. 4561 */ 4562 prev_state = READ_ONCE(prev->__state); 4563 vtime_task_switch(prev); 4564 perf_event_task_sched_in(prev, current); 4565 finish_task(prev); 4566 tick_nohz_task_switch(); 4567 finish_lock_switch(rq); 4568 finish_arch_post_lock_switch(); 4569 kcov_finish_switch(current); 4570 /* 4571 * kmap_local_sched_out() is invoked with rq::lock held and 4572 * interrupts disabled. There is no requirement for that, but the 4573 * sched out code does not have an interrupt enabled section. 4574 * Restoring the maps on sched in does not require interrupts being 4575 * disabled either. 4576 */ 4577 kmap_local_sched_in(); 4578 4579 fire_sched_in_preempt_notifiers(current); 4580 /* 4581 * When switching through a kernel thread, the loop in 4582 * membarrier_{private,global}_expedited() may have observed that 4583 * kernel thread and not issued an IPI. It is therefore possible to 4584 * schedule between user->kernel->user threads without passing though 4585 * switch_mm(). Membarrier requires a barrier after storing to 4586 * rq->curr, before returning to userspace, so provide them here: 4587 * 4588 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly 4589 * provided by mmdrop(), 4590 * - a sync_core for SYNC_CORE. 4591 */ 4592 if (mm) { 4593 membarrier_mm_sync_core_before_usermode(mm); 4594 mmdrop(mm); 4595 } 4596 if (unlikely(prev_state == TASK_DEAD)) { 4597 if (prev->sched_class->task_dead) 4598 prev->sched_class->task_dead(prev); 4599 4600 /* 4601 * Remove function-return probe instances associated with this 4602 * task and put them back on the free list. 4603 */ 4604 kprobe_flush_task(prev); 4605 4606 /* Task is done with its stack. */ 4607 put_task_stack(prev); 4608 4609 put_task_struct_rcu_user(prev); 4610 } 4611 4612 return rq; 4613 } 4614 4615 /** 4616 * schedule_tail - first thing a freshly forked thread must call. 4617 * @prev: the thread we just switched away from. 4618 */ 4619 asmlinkage __visible void schedule_tail(struct task_struct *prev) 4620 __releases(rq->lock) 4621 { 4622 /* 4623 * New tasks start with FORK_PREEMPT_COUNT, see there and 4624 * finish_task_switch() for details. 4625 * 4626 * finish_task_switch() will drop rq->lock() and lower preempt_count 4627 * and the preempt_enable() will end up enabling preemption (on 4628 * PREEMPT_COUNT kernels). 4629 */ 4630 4631 finish_task_switch(prev); 4632 preempt_enable(); 4633 4634 if (current->set_child_tid) 4635 put_user(task_pid_vnr(current), current->set_child_tid); 4636 4637 calculate_sigpending(); 4638 } 4639 4640 /* 4641 * context_switch - switch to the new MM and the new thread's register state. 4642 */ 4643 static __always_inline struct rq * 4644 context_switch(struct rq *rq, struct task_struct *prev, 4645 struct task_struct *next, struct rq_flags *rf) 4646 { 4647 prepare_task_switch(rq, prev, next); 4648 4649 /* 4650 * For paravirt, this is coupled with an exit in switch_to to 4651 * combine the page table reload and the switch backend into 4652 * one hypercall. 4653 */ 4654 arch_start_context_switch(prev); 4655 4656 /* 4657 * kernel -> kernel lazy + transfer active 4658 * user -> kernel lazy + mmgrab() active 4659 * 4660 * kernel -> user switch + mmdrop() active 4661 * user -> user switch 4662 */ 4663 if (!next->mm) { // to kernel 4664 enter_lazy_tlb(prev->active_mm, next); 4665 4666 next->active_mm = prev->active_mm; 4667 if (prev->mm) // from user 4668 mmgrab(prev->active_mm); 4669 else 4670 prev->active_mm = NULL; 4671 } else { // to user 4672 membarrier_switch_mm(rq, prev->active_mm, next->mm); 4673 /* 4674 * sys_membarrier() requires an smp_mb() between setting 4675 * rq->curr / membarrier_switch_mm() and returning to userspace. 4676 * 4677 * The below provides this either through switch_mm(), or in 4678 * case 'prev->active_mm == next->mm' through 4679 * finish_task_switch()'s mmdrop(). 4680 */ 4681 switch_mm_irqs_off(prev->active_mm, next->mm, next); 4682 4683 if (!prev->mm) { // from kernel 4684 /* will mmdrop() in finish_task_switch(). */ 4685 rq->prev_mm = prev->active_mm; 4686 prev->active_mm = NULL; 4687 } 4688 } 4689 4690 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP); 4691 4692 prepare_lock_switch(rq, next, rf); 4693 4694 /* Here we just switch the register state and the stack. */ 4695 switch_to(prev, next, prev); 4696 barrier(); 4697 4698 return finish_task_switch(prev); 4699 } 4700 4701 /* 4702 * nr_running and nr_context_switches: 4703 * 4704 * externally visible scheduler statistics: current number of runnable 4705 * threads, total number of context switches performed since bootup. 4706 */ 4707 unsigned int nr_running(void) 4708 { 4709 unsigned int i, sum = 0; 4710 4711 for_each_online_cpu(i) 4712 sum += cpu_rq(i)->nr_running; 4713 4714 return sum; 4715 } 4716 4717 /* 4718 * Check if only the current task is running on the CPU. 4719 * 4720 * Caution: this function does not check that the caller has disabled 4721 * preemption, thus the result might have a time-of-check-to-time-of-use 4722 * race. The caller is responsible to use it correctly, for example: 4723 * 4724 * - from a non-preemptible section (of course) 4725 * 4726 * - from a thread that is bound to a single CPU 4727 * 4728 * - in a loop with very short iterations (e.g. a polling loop) 4729 */ 4730 bool single_task_running(void) 4731 { 4732 return raw_rq()->nr_running == 1; 4733 } 4734 EXPORT_SYMBOL(single_task_running); 4735 4736 unsigned long long nr_context_switches(void) 4737 { 4738 int i; 4739 unsigned long long sum = 0; 4740 4741 for_each_possible_cpu(i) 4742 sum += cpu_rq(i)->nr_switches; 4743 4744 return sum; 4745 } 4746 4747 /* 4748 * Consumers of these two interfaces, like for example the cpuidle menu 4749 * governor, are using nonsensical data. Preferring shallow idle state selection 4750 * for a CPU that has IO-wait which might not even end up running the task when 4751 * it does become runnable. 4752 */ 4753 4754 unsigned int nr_iowait_cpu(int cpu) 4755 { 4756 return atomic_read(&cpu_rq(cpu)->nr_iowait); 4757 } 4758 4759 /* 4760 * IO-wait accounting, and how it's mostly bollocks (on SMP). 4761 * 4762 * The idea behind IO-wait account is to account the idle time that we could 4763 * have spend running if it were not for IO. That is, if we were to improve the 4764 * storage performance, we'd have a proportional reduction in IO-wait time. 4765 * 4766 * This all works nicely on UP, where, when a task blocks on IO, we account 4767 * idle time as IO-wait, because if the storage were faster, it could've been 4768 * running and we'd not be idle. 4769 * 4770 * This has been extended to SMP, by doing the same for each CPU. This however 4771 * is broken. 4772 * 4773 * Imagine for instance the case where two tasks block on one CPU, only the one 4774 * CPU will have IO-wait accounted, while the other has regular idle. Even 4775 * though, if the storage were faster, both could've ran at the same time, 4776 * utilising both CPUs. 4777 * 4778 * This means, that when looking globally, the current IO-wait accounting on 4779 * SMP is a lower bound, by reason of under accounting. 4780 * 4781 * Worse, since the numbers are provided per CPU, they are sometimes 4782 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly 4783 * associated with any one particular CPU, it can wake to another CPU than it 4784 * blocked on. This means the per CPU IO-wait number is meaningless. 4785 * 4786 * Task CPU affinities can make all that even more 'interesting'. 4787 */ 4788 4789 unsigned int nr_iowait(void) 4790 { 4791 unsigned int i, sum = 0; 4792 4793 for_each_possible_cpu(i) 4794 sum += nr_iowait_cpu(i); 4795 4796 return sum; 4797 } 4798 4799 #ifdef CONFIG_SMP 4800 4801 /* 4802 * sched_exec - execve() is a valuable balancing opportunity, because at 4803 * this point the task has the smallest effective memory and cache footprint. 4804 */ 4805 void sched_exec(void) 4806 { 4807 struct task_struct *p = current; 4808 unsigned long flags; 4809 int dest_cpu; 4810 4811 raw_spin_lock_irqsave(&p->pi_lock, flags); 4812 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC); 4813 if (dest_cpu == smp_processor_id()) 4814 goto unlock; 4815 4816 if (likely(cpu_active(dest_cpu))) { 4817 struct migration_arg arg = { p, dest_cpu }; 4818 4819 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4820 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 4821 return; 4822 } 4823 unlock: 4824 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4825 } 4826 4827 #endif 4828 4829 DEFINE_PER_CPU(struct kernel_stat, kstat); 4830 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 4831 4832 EXPORT_PER_CPU_SYMBOL(kstat); 4833 EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 4834 4835 /* 4836 * The function fair_sched_class.update_curr accesses the struct curr 4837 * and its field curr->exec_start; when called from task_sched_runtime(), 4838 * we observe a high rate of cache misses in practice. 4839 * Prefetching this data results in improved performance. 4840 */ 4841 static inline void prefetch_curr_exec_start(struct task_struct *p) 4842 { 4843 #ifdef CONFIG_FAIR_GROUP_SCHED 4844 struct sched_entity *curr = (&p->se)->cfs_rq->curr; 4845 #else 4846 struct sched_entity *curr = (&task_rq(p)->cfs)->curr; 4847 #endif 4848 prefetch(curr); 4849 prefetch(&curr->exec_start); 4850 } 4851 4852 /* 4853 * Return accounted runtime for the task. 4854 * In case the task is currently running, return the runtime plus current's 4855 * pending runtime that have not been accounted yet. 4856 */ 4857 unsigned long long task_sched_runtime(struct task_struct *p) 4858 { 4859 struct rq_flags rf; 4860 struct rq *rq; 4861 u64 ns; 4862 4863 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 4864 /* 4865 * 64-bit doesn't need locks to atomically read a 64-bit value. 4866 * So we have a optimization chance when the task's delta_exec is 0. 4867 * Reading ->on_cpu is racy, but this is ok. 4868 * 4869 * If we race with it leaving CPU, we'll take a lock. So we're correct. 4870 * If we race with it entering CPU, unaccounted time is 0. This is 4871 * indistinguishable from the read occurring a few cycles earlier. 4872 * If we see ->on_cpu without ->on_rq, the task is leaving, and has 4873 * been accounted, so we're correct here as well. 4874 */ 4875 if (!p->on_cpu || !task_on_rq_queued(p)) 4876 return p->se.sum_exec_runtime; 4877 #endif 4878 4879 rq = task_rq_lock(p, &rf); 4880 /* 4881 * Must be ->curr _and_ ->on_rq. If dequeued, we would 4882 * project cycles that may never be accounted to this 4883 * thread, breaking clock_gettime(). 4884 */ 4885 if (task_current(rq, p) && task_on_rq_queued(p)) { 4886 prefetch_curr_exec_start(p); 4887 update_rq_clock(rq); 4888 p->sched_class->update_curr(rq); 4889 } 4890 ns = p->se.sum_exec_runtime; 4891 task_rq_unlock(rq, p, &rf); 4892 4893 return ns; 4894 } 4895 4896 #ifdef CONFIG_SCHED_DEBUG 4897 static u64 cpu_resched_latency(struct rq *rq) 4898 { 4899 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms); 4900 u64 resched_latency, now = rq_clock(rq); 4901 static bool warned_once; 4902 4903 if (sysctl_resched_latency_warn_once && warned_once) 4904 return 0; 4905 4906 if (!need_resched() || !latency_warn_ms) 4907 return 0; 4908 4909 if (system_state == SYSTEM_BOOTING) 4910 return 0; 4911 4912 if (!rq->last_seen_need_resched_ns) { 4913 rq->last_seen_need_resched_ns = now; 4914 rq->ticks_without_resched = 0; 4915 return 0; 4916 } 4917 4918 rq->ticks_without_resched++; 4919 resched_latency = now - rq->last_seen_need_resched_ns; 4920 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC) 4921 return 0; 4922 4923 warned_once = true; 4924 4925 return resched_latency; 4926 } 4927 4928 static int __init setup_resched_latency_warn_ms(char *str) 4929 { 4930 long val; 4931 4932 if ((kstrtol(str, 0, &val))) { 4933 pr_warn("Unable to set resched_latency_warn_ms\n"); 4934 return 1; 4935 } 4936 4937 sysctl_resched_latency_warn_ms = val; 4938 return 1; 4939 } 4940 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms); 4941 #else 4942 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; } 4943 #endif /* CONFIG_SCHED_DEBUG */ 4944 4945 /* 4946 * This function gets called by the timer code, with HZ frequency. 4947 * We call it with interrupts disabled. 4948 */ 4949 void scheduler_tick(void) 4950 { 4951 int cpu = smp_processor_id(); 4952 struct rq *rq = cpu_rq(cpu); 4953 struct task_struct *curr = rq->curr; 4954 struct rq_flags rf; 4955 unsigned long thermal_pressure; 4956 u64 resched_latency; 4957 4958 arch_scale_freq_tick(); 4959 sched_clock_tick(); 4960 4961 rq_lock(rq, &rf); 4962 4963 update_rq_clock(rq); 4964 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); 4965 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure); 4966 curr->sched_class->task_tick(rq, curr, 0); 4967 if (sched_feat(LATENCY_WARN)) 4968 resched_latency = cpu_resched_latency(rq); 4969 calc_global_load_tick(rq); 4970 4971 rq_unlock(rq, &rf); 4972 4973 if (sched_feat(LATENCY_WARN) && resched_latency) 4974 resched_latency_warn(cpu, resched_latency); 4975 4976 perf_event_task_tick(); 4977 4978 #ifdef CONFIG_SMP 4979 rq->idle_balance = idle_cpu(cpu); 4980 trigger_load_balance(rq); 4981 #endif 4982 } 4983 4984 #ifdef CONFIG_NO_HZ_FULL 4985 4986 struct tick_work { 4987 int cpu; 4988 atomic_t state; 4989 struct delayed_work work; 4990 }; 4991 /* Values for ->state, see diagram below. */ 4992 #define TICK_SCHED_REMOTE_OFFLINE 0 4993 #define TICK_SCHED_REMOTE_OFFLINING 1 4994 #define TICK_SCHED_REMOTE_RUNNING 2 4995 4996 /* 4997 * State diagram for ->state: 4998 * 4999 * 5000 * TICK_SCHED_REMOTE_OFFLINE 5001 * | ^ 5002 * | | 5003 * | | sched_tick_remote() 5004 * | | 5005 * | | 5006 * +--TICK_SCHED_REMOTE_OFFLINING 5007 * | ^ 5008 * | | 5009 * sched_tick_start() | | sched_tick_stop() 5010 * | | 5011 * V | 5012 * TICK_SCHED_REMOTE_RUNNING 5013 * 5014 * 5015 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote() 5016 * and sched_tick_start() are happy to leave the state in RUNNING. 5017 */ 5018 5019 static struct tick_work __percpu *tick_work_cpu; 5020 5021 static void sched_tick_remote(struct work_struct *work) 5022 { 5023 struct delayed_work *dwork = to_delayed_work(work); 5024 struct tick_work *twork = container_of(dwork, struct tick_work, work); 5025 int cpu = twork->cpu; 5026 struct rq *rq = cpu_rq(cpu); 5027 struct task_struct *curr; 5028 struct rq_flags rf; 5029 u64 delta; 5030 int os; 5031 5032 /* 5033 * Handle the tick only if it appears the remote CPU is running in full 5034 * dynticks mode. The check is racy by nature, but missing a tick or 5035 * having one too much is no big deal because the scheduler tick updates 5036 * statistics and checks timeslices in a time-independent way, regardless 5037 * of when exactly it is running. 5038 */ 5039 if (!tick_nohz_tick_stopped_cpu(cpu)) 5040 goto out_requeue; 5041 5042 rq_lock_irq(rq, &rf); 5043 curr = rq->curr; 5044 if (cpu_is_offline(cpu)) 5045 goto out_unlock; 5046 5047 update_rq_clock(rq); 5048 5049 if (!is_idle_task(curr)) { 5050 /* 5051 * Make sure the next tick runs within a reasonable 5052 * amount of time. 5053 */ 5054 delta = rq_clock_task(rq) - curr->se.exec_start; 5055 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3); 5056 } 5057 curr->sched_class->task_tick(rq, curr, 0); 5058 5059 calc_load_nohz_remote(rq); 5060 out_unlock: 5061 rq_unlock_irq(rq, &rf); 5062 out_requeue: 5063 5064 /* 5065 * Run the remote tick once per second (1Hz). This arbitrary 5066 * frequency is large enough to avoid overload but short enough 5067 * to keep scheduler internal stats reasonably up to date. But 5068 * first update state to reflect hotplug activity if required. 5069 */ 5070 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING); 5071 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE); 5072 if (os == TICK_SCHED_REMOTE_RUNNING) 5073 queue_delayed_work(system_unbound_wq, dwork, HZ); 5074 } 5075 5076 static void sched_tick_start(int cpu) 5077 { 5078 int os; 5079 struct tick_work *twork; 5080 5081 if (housekeeping_cpu(cpu, HK_FLAG_TICK)) 5082 return; 5083 5084 WARN_ON_ONCE(!tick_work_cpu); 5085 5086 twork = per_cpu_ptr(tick_work_cpu, cpu); 5087 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING); 5088 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING); 5089 if (os == TICK_SCHED_REMOTE_OFFLINE) { 5090 twork->cpu = cpu; 5091 INIT_DELAYED_WORK(&twork->work, sched_tick_remote); 5092 queue_delayed_work(system_unbound_wq, &twork->work, HZ); 5093 } 5094 } 5095 5096 #ifdef CONFIG_HOTPLUG_CPU 5097 static void sched_tick_stop(int cpu) 5098 { 5099 struct tick_work *twork; 5100 int os; 5101 5102 if (housekeeping_cpu(cpu, HK_FLAG_TICK)) 5103 return; 5104 5105 WARN_ON_ONCE(!tick_work_cpu); 5106 5107 twork = per_cpu_ptr(tick_work_cpu, cpu); 5108 /* There cannot be competing actions, but don't rely on stop-machine. */ 5109 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING); 5110 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING); 5111 /* Don't cancel, as this would mess up the state machine. */ 5112 } 5113 #endif /* CONFIG_HOTPLUG_CPU */ 5114 5115 int __init sched_tick_offload_init(void) 5116 { 5117 tick_work_cpu = alloc_percpu(struct tick_work); 5118 BUG_ON(!tick_work_cpu); 5119 return 0; 5120 } 5121 5122 #else /* !CONFIG_NO_HZ_FULL */ 5123 static inline void sched_tick_start(int cpu) { } 5124 static inline void sched_tick_stop(int cpu) { } 5125 #endif 5126 5127 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \ 5128 defined(CONFIG_TRACE_PREEMPT_TOGGLE)) 5129 /* 5130 * If the value passed in is equal to the current preempt count 5131 * then we just disabled preemption. Start timing the latency. 5132 */ 5133 static inline void preempt_latency_start(int val) 5134 { 5135 if (preempt_count() == val) { 5136 unsigned long ip = get_lock_parent_ip(); 5137 #ifdef CONFIG_DEBUG_PREEMPT 5138 current->preempt_disable_ip = ip; 5139 #endif 5140 trace_preempt_off(CALLER_ADDR0, ip); 5141 } 5142 } 5143 5144 void preempt_count_add(int val) 5145 { 5146 #ifdef CONFIG_DEBUG_PREEMPT 5147 /* 5148 * Underflow? 5149 */ 5150 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 5151 return; 5152 #endif 5153 __preempt_count_add(val); 5154 #ifdef CONFIG_DEBUG_PREEMPT 5155 /* 5156 * Spinlock count overflowing soon? 5157 */ 5158 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 5159 PREEMPT_MASK - 10); 5160 #endif 5161 preempt_latency_start(val); 5162 } 5163 EXPORT_SYMBOL(preempt_count_add); 5164 NOKPROBE_SYMBOL(preempt_count_add); 5165 5166 /* 5167 * If the value passed in equals to the current preempt count 5168 * then we just enabled preemption. Stop timing the latency. 5169 */ 5170 static inline void preempt_latency_stop(int val) 5171 { 5172 if (preempt_count() == val) 5173 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip()); 5174 } 5175 5176 void preempt_count_sub(int val) 5177 { 5178 #ifdef CONFIG_DEBUG_PREEMPT 5179 /* 5180 * Underflow? 5181 */ 5182 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 5183 return; 5184 /* 5185 * Is the spinlock portion underflowing? 5186 */ 5187 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 5188 !(preempt_count() & PREEMPT_MASK))) 5189 return; 5190 #endif 5191 5192 preempt_latency_stop(val); 5193 __preempt_count_sub(val); 5194 } 5195 EXPORT_SYMBOL(preempt_count_sub); 5196 NOKPROBE_SYMBOL(preempt_count_sub); 5197 5198 #else 5199 static inline void preempt_latency_start(int val) { } 5200 static inline void preempt_latency_stop(int val) { } 5201 #endif 5202 5203 static inline unsigned long get_preempt_disable_ip(struct task_struct *p) 5204 { 5205 #ifdef CONFIG_DEBUG_PREEMPT 5206 return p->preempt_disable_ip; 5207 #else 5208 return 0; 5209 #endif 5210 } 5211 5212 /* 5213 * Print scheduling while atomic bug: 5214 */ 5215 static noinline void __schedule_bug(struct task_struct *prev) 5216 { 5217 /* Save this before calling printk(), since that will clobber it */ 5218 unsigned long preempt_disable_ip = get_preempt_disable_ip(current); 5219 5220 if (oops_in_progress) 5221 return; 5222 5223 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 5224 prev->comm, prev->pid, preempt_count()); 5225 5226 debug_show_held_locks(prev); 5227 print_modules(); 5228 if (irqs_disabled()) 5229 print_irqtrace_events(prev); 5230 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT) 5231 && in_atomic_preempt_off()) { 5232 pr_err("Preemption disabled at:"); 5233 print_ip_sym(KERN_ERR, preempt_disable_ip); 5234 } 5235 if (panic_on_warn) 5236 panic("scheduling while atomic\n"); 5237 5238 dump_stack(); 5239 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 5240 } 5241 5242 /* 5243 * Various schedule()-time debugging checks and statistics: 5244 */ 5245 static inline void schedule_debug(struct task_struct *prev, bool preempt) 5246 { 5247 #ifdef CONFIG_SCHED_STACK_END_CHECK 5248 if (task_stack_end_corrupted(prev)) 5249 panic("corrupted stack end detected inside scheduler\n"); 5250 5251 if (task_scs_end_corrupted(prev)) 5252 panic("corrupted shadow stack detected inside scheduler\n"); 5253 #endif 5254 5255 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP 5256 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) { 5257 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n", 5258 prev->comm, prev->pid, prev->non_block_count); 5259 dump_stack(); 5260 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 5261 } 5262 #endif 5263 5264 if (unlikely(in_atomic_preempt_off())) { 5265 __schedule_bug(prev); 5266 preempt_count_set(PREEMPT_DISABLED); 5267 } 5268 rcu_sleep_check(); 5269 SCHED_WARN_ON(ct_state() == CONTEXT_USER); 5270 5271 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 5272 5273 schedstat_inc(this_rq()->sched_count); 5274 } 5275 5276 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev, 5277 struct rq_flags *rf) 5278 { 5279 #ifdef CONFIG_SMP 5280 const struct sched_class *class; 5281 /* 5282 * We must do the balancing pass before put_prev_task(), such 5283 * that when we release the rq->lock the task is in the same 5284 * state as before we took rq->lock. 5285 * 5286 * We can terminate the balance pass as soon as we know there is 5287 * a runnable task of @class priority or higher. 5288 */ 5289 for_class_range(class, prev->sched_class, &idle_sched_class) { 5290 if (class->balance(rq, prev, rf)) 5291 break; 5292 } 5293 #endif 5294 5295 put_prev_task(rq, prev); 5296 } 5297 5298 /* 5299 * Pick up the highest-prio task: 5300 */ 5301 static inline struct task_struct * 5302 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 5303 { 5304 const struct sched_class *class; 5305 struct task_struct *p; 5306 5307 /* 5308 * Optimization: we know that if all tasks are in the fair class we can 5309 * call that function directly, but only if the @prev task wasn't of a 5310 * higher scheduling class, because otherwise those lose the 5311 * opportunity to pull in more work from other CPUs. 5312 */ 5313 if (likely(prev->sched_class <= &fair_sched_class && 5314 rq->nr_running == rq->cfs.h_nr_running)) { 5315 5316 p = pick_next_task_fair(rq, prev, rf); 5317 if (unlikely(p == RETRY_TASK)) 5318 goto restart; 5319 5320 /* Assume the next prioritized class is idle_sched_class */ 5321 if (!p) { 5322 put_prev_task(rq, prev); 5323 p = pick_next_task_idle(rq); 5324 } 5325 5326 return p; 5327 } 5328 5329 restart: 5330 put_prev_task_balance(rq, prev, rf); 5331 5332 for_each_class(class) { 5333 p = class->pick_next_task(rq); 5334 if (p) 5335 return p; 5336 } 5337 5338 /* The idle class should always have a runnable task: */ 5339 BUG(); 5340 } 5341 5342 #ifdef CONFIG_SCHED_CORE 5343 static inline bool is_task_rq_idle(struct task_struct *t) 5344 { 5345 return (task_rq(t)->idle == t); 5346 } 5347 5348 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie) 5349 { 5350 return is_task_rq_idle(a) || (a->core_cookie == cookie); 5351 } 5352 5353 static inline bool cookie_match(struct task_struct *a, struct task_struct *b) 5354 { 5355 if (is_task_rq_idle(a) || is_task_rq_idle(b)) 5356 return true; 5357 5358 return a->core_cookie == b->core_cookie; 5359 } 5360 5361 // XXX fairness/fwd progress conditions 5362 /* 5363 * Returns 5364 * - NULL if there is no runnable task for this class. 5365 * - the highest priority task for this runqueue if it matches 5366 * rq->core->core_cookie or its priority is greater than max. 5367 * - Else returns idle_task. 5368 */ 5369 static struct task_struct * 5370 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi) 5371 { 5372 struct task_struct *class_pick, *cookie_pick; 5373 unsigned long cookie = rq->core->core_cookie; 5374 5375 class_pick = class->pick_task(rq); 5376 if (!class_pick) 5377 return NULL; 5378 5379 if (!cookie) { 5380 /* 5381 * If class_pick is tagged, return it only if it has 5382 * higher priority than max. 5383 */ 5384 if (max && class_pick->core_cookie && 5385 prio_less(class_pick, max, in_fi)) 5386 return idle_sched_class.pick_task(rq); 5387 5388 return class_pick; 5389 } 5390 5391 /* 5392 * If class_pick is idle or matches cookie, return early. 5393 */ 5394 if (cookie_equals(class_pick, cookie)) 5395 return class_pick; 5396 5397 cookie_pick = sched_core_find(rq, cookie); 5398 5399 /* 5400 * If class > max && class > cookie, it is the highest priority task on 5401 * the core (so far) and it must be selected, otherwise we must go with 5402 * the cookie pick in order to satisfy the constraint. 5403 */ 5404 if (prio_less(cookie_pick, class_pick, in_fi) && 5405 (!max || prio_less(max, class_pick, in_fi))) 5406 return class_pick; 5407 5408 return cookie_pick; 5409 } 5410 5411 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi); 5412 5413 static struct task_struct * 5414 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 5415 { 5416 struct task_struct *next, *max = NULL; 5417 const struct sched_class *class; 5418 const struct cpumask *smt_mask; 5419 bool fi_before = false; 5420 int i, j, cpu, occ = 0; 5421 bool need_sync; 5422 5423 if (!sched_core_enabled(rq)) 5424 return __pick_next_task(rq, prev, rf); 5425 5426 cpu = cpu_of(rq); 5427 5428 /* Stopper task is switching into idle, no need core-wide selection. */ 5429 if (cpu_is_offline(cpu)) { 5430 /* 5431 * Reset core_pick so that we don't enter the fastpath when 5432 * coming online. core_pick would already be migrated to 5433 * another cpu during offline. 5434 */ 5435 rq->core_pick = NULL; 5436 return __pick_next_task(rq, prev, rf); 5437 } 5438 5439 /* 5440 * If there were no {en,de}queues since we picked (IOW, the task 5441 * pointers are all still valid), and we haven't scheduled the last 5442 * pick yet, do so now. 5443 * 5444 * rq->core_pick can be NULL if no selection was made for a CPU because 5445 * it was either offline or went offline during a sibling's core-wide 5446 * selection. In this case, do a core-wide selection. 5447 */ 5448 if (rq->core->core_pick_seq == rq->core->core_task_seq && 5449 rq->core->core_pick_seq != rq->core_sched_seq && 5450 rq->core_pick) { 5451 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq); 5452 5453 next = rq->core_pick; 5454 if (next != prev) { 5455 put_prev_task(rq, prev); 5456 set_next_task(rq, next); 5457 } 5458 5459 rq->core_pick = NULL; 5460 return next; 5461 } 5462 5463 put_prev_task_balance(rq, prev, rf); 5464 5465 smt_mask = cpu_smt_mask(cpu); 5466 need_sync = !!rq->core->core_cookie; 5467 5468 /* reset state */ 5469 rq->core->core_cookie = 0UL; 5470 if (rq->core->core_forceidle) { 5471 need_sync = true; 5472 fi_before = true; 5473 rq->core->core_forceidle = false; 5474 } 5475 5476 /* 5477 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq 5478 * 5479 * @task_seq guards the task state ({en,de}queues) 5480 * @pick_seq is the @task_seq we did a selection on 5481 * @sched_seq is the @pick_seq we scheduled 5482 * 5483 * However, preemptions can cause multiple picks on the same task set. 5484 * 'Fix' this by also increasing @task_seq for every pick. 5485 */ 5486 rq->core->core_task_seq++; 5487 5488 /* 5489 * Optimize for common case where this CPU has no cookies 5490 * and there are no cookied tasks running on siblings. 5491 */ 5492 if (!need_sync) { 5493 for_each_class(class) { 5494 next = class->pick_task(rq); 5495 if (next) 5496 break; 5497 } 5498 5499 if (!next->core_cookie) { 5500 rq->core_pick = NULL; 5501 /* 5502 * For robustness, update the min_vruntime_fi for 5503 * unconstrained picks as well. 5504 */ 5505 WARN_ON_ONCE(fi_before); 5506 task_vruntime_update(rq, next, false); 5507 goto done; 5508 } 5509 } 5510 5511 for_each_cpu(i, smt_mask) { 5512 struct rq *rq_i = cpu_rq(i); 5513 5514 rq_i->core_pick = NULL; 5515 5516 if (i != cpu) 5517 update_rq_clock(rq_i); 5518 } 5519 5520 /* 5521 * Try and select tasks for each sibling in descending sched_class 5522 * order. 5523 */ 5524 for_each_class(class) { 5525 again: 5526 for_each_cpu_wrap(i, smt_mask, cpu) { 5527 struct rq *rq_i = cpu_rq(i); 5528 struct task_struct *p; 5529 5530 if (rq_i->core_pick) 5531 continue; 5532 5533 /* 5534 * If this sibling doesn't yet have a suitable task to 5535 * run; ask for the most eligible task, given the 5536 * highest priority task already selected for this 5537 * core. 5538 */ 5539 p = pick_task(rq_i, class, max, fi_before); 5540 if (!p) 5541 continue; 5542 5543 if (!is_task_rq_idle(p)) 5544 occ++; 5545 5546 rq_i->core_pick = p; 5547 if (rq_i->idle == p && rq_i->nr_running) { 5548 rq->core->core_forceidle = true; 5549 if (!fi_before) 5550 rq->core->core_forceidle_seq++; 5551 } 5552 5553 /* 5554 * If this new candidate is of higher priority than the 5555 * previous; and they're incompatible; we need to wipe 5556 * the slate and start over. pick_task makes sure that 5557 * p's priority is more than max if it doesn't match 5558 * max's cookie. 5559 * 5560 * NOTE: this is a linear max-filter and is thus bounded 5561 * in execution time. 5562 */ 5563 if (!max || !cookie_match(max, p)) { 5564 struct task_struct *old_max = max; 5565 5566 rq->core->core_cookie = p->core_cookie; 5567 max = p; 5568 5569 if (old_max) { 5570 rq->core->core_forceidle = false; 5571 for_each_cpu(j, smt_mask) { 5572 if (j == i) 5573 continue; 5574 5575 cpu_rq(j)->core_pick = NULL; 5576 } 5577 occ = 1; 5578 goto again; 5579 } 5580 } 5581 } 5582 } 5583 5584 rq->core->core_pick_seq = rq->core->core_task_seq; 5585 next = rq->core_pick; 5586 rq->core_sched_seq = rq->core->core_pick_seq; 5587 5588 /* Something should have been selected for current CPU */ 5589 WARN_ON_ONCE(!next); 5590 5591 /* 5592 * Reschedule siblings 5593 * 5594 * NOTE: L1TF -- at this point we're no longer running the old task and 5595 * sending an IPI (below) ensures the sibling will no longer be running 5596 * their task. This ensures there is no inter-sibling overlap between 5597 * non-matching user state. 5598 */ 5599 for_each_cpu(i, smt_mask) { 5600 struct rq *rq_i = cpu_rq(i); 5601 5602 /* 5603 * An online sibling might have gone offline before a task 5604 * could be picked for it, or it might be offline but later 5605 * happen to come online, but its too late and nothing was 5606 * picked for it. That's Ok - it will pick tasks for itself, 5607 * so ignore it. 5608 */ 5609 if (!rq_i->core_pick) 5610 continue; 5611 5612 /* 5613 * Update for new !FI->FI transitions, or if continuing to be in !FI: 5614 * fi_before fi update? 5615 * 0 0 1 5616 * 0 1 1 5617 * 1 0 1 5618 * 1 1 0 5619 */ 5620 if (!(fi_before && rq->core->core_forceidle)) 5621 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle); 5622 5623 rq_i->core_pick->core_occupation = occ; 5624 5625 if (i == cpu) { 5626 rq_i->core_pick = NULL; 5627 continue; 5628 } 5629 5630 /* Did we break L1TF mitigation requirements? */ 5631 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick)); 5632 5633 if (rq_i->curr == rq_i->core_pick) { 5634 rq_i->core_pick = NULL; 5635 continue; 5636 } 5637 5638 resched_curr(rq_i); 5639 } 5640 5641 done: 5642 set_next_task(rq, next); 5643 return next; 5644 } 5645 5646 static bool try_steal_cookie(int this, int that) 5647 { 5648 struct rq *dst = cpu_rq(this), *src = cpu_rq(that); 5649 struct task_struct *p; 5650 unsigned long cookie; 5651 bool success = false; 5652 5653 local_irq_disable(); 5654 double_rq_lock(dst, src); 5655 5656 cookie = dst->core->core_cookie; 5657 if (!cookie) 5658 goto unlock; 5659 5660 if (dst->curr != dst->idle) 5661 goto unlock; 5662 5663 p = sched_core_find(src, cookie); 5664 if (p == src->idle) 5665 goto unlock; 5666 5667 do { 5668 if (p == src->core_pick || p == src->curr) 5669 goto next; 5670 5671 if (!cpumask_test_cpu(this, &p->cpus_mask)) 5672 goto next; 5673 5674 if (p->core_occupation > dst->idle->core_occupation) 5675 goto next; 5676 5677 p->on_rq = TASK_ON_RQ_MIGRATING; 5678 deactivate_task(src, p, 0); 5679 set_task_cpu(p, this); 5680 activate_task(dst, p, 0); 5681 p->on_rq = TASK_ON_RQ_QUEUED; 5682 5683 resched_curr(dst); 5684 5685 success = true; 5686 break; 5687 5688 next: 5689 p = sched_core_next(p, cookie); 5690 } while (p); 5691 5692 unlock: 5693 double_rq_unlock(dst, src); 5694 local_irq_enable(); 5695 5696 return success; 5697 } 5698 5699 static bool steal_cookie_task(int cpu, struct sched_domain *sd) 5700 { 5701 int i; 5702 5703 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) { 5704 if (i == cpu) 5705 continue; 5706 5707 if (need_resched()) 5708 break; 5709 5710 if (try_steal_cookie(cpu, i)) 5711 return true; 5712 } 5713 5714 return false; 5715 } 5716 5717 static void sched_core_balance(struct rq *rq) 5718 { 5719 struct sched_domain *sd; 5720 int cpu = cpu_of(rq); 5721 5722 preempt_disable(); 5723 rcu_read_lock(); 5724 raw_spin_rq_unlock_irq(rq); 5725 for_each_domain(cpu, sd) { 5726 if (need_resched()) 5727 break; 5728 5729 if (steal_cookie_task(cpu, sd)) 5730 break; 5731 } 5732 raw_spin_rq_lock_irq(rq); 5733 rcu_read_unlock(); 5734 preempt_enable(); 5735 } 5736 5737 static DEFINE_PER_CPU(struct callback_head, core_balance_head); 5738 5739 void queue_core_balance(struct rq *rq) 5740 { 5741 if (!sched_core_enabled(rq)) 5742 return; 5743 5744 if (!rq->core->core_cookie) 5745 return; 5746 5747 if (!rq->nr_running) /* not forced idle */ 5748 return; 5749 5750 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance); 5751 } 5752 5753 static void sched_core_cpu_starting(unsigned int cpu) 5754 { 5755 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 5756 struct rq *rq = cpu_rq(cpu), *core_rq = NULL; 5757 unsigned long flags; 5758 int t; 5759 5760 sched_core_lock(cpu, &flags); 5761 5762 WARN_ON_ONCE(rq->core != rq); 5763 5764 /* if we're the first, we'll be our own leader */ 5765 if (cpumask_weight(smt_mask) == 1) 5766 goto unlock; 5767 5768 /* find the leader */ 5769 for_each_cpu(t, smt_mask) { 5770 if (t == cpu) 5771 continue; 5772 rq = cpu_rq(t); 5773 if (rq->core == rq) { 5774 core_rq = rq; 5775 break; 5776 } 5777 } 5778 5779 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */ 5780 goto unlock; 5781 5782 /* install and validate core_rq */ 5783 for_each_cpu(t, smt_mask) { 5784 rq = cpu_rq(t); 5785 5786 if (t == cpu) 5787 rq->core = core_rq; 5788 5789 WARN_ON_ONCE(rq->core != core_rq); 5790 } 5791 5792 unlock: 5793 sched_core_unlock(cpu, &flags); 5794 } 5795 5796 static void sched_core_cpu_deactivate(unsigned int cpu) 5797 { 5798 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 5799 struct rq *rq = cpu_rq(cpu), *core_rq = NULL; 5800 unsigned long flags; 5801 int t; 5802 5803 sched_core_lock(cpu, &flags); 5804 5805 /* if we're the last man standing, nothing to do */ 5806 if (cpumask_weight(smt_mask) == 1) { 5807 WARN_ON_ONCE(rq->core != rq); 5808 goto unlock; 5809 } 5810 5811 /* if we're not the leader, nothing to do */ 5812 if (rq->core != rq) 5813 goto unlock; 5814 5815 /* find a new leader */ 5816 for_each_cpu(t, smt_mask) { 5817 if (t == cpu) 5818 continue; 5819 core_rq = cpu_rq(t); 5820 break; 5821 } 5822 5823 if (WARN_ON_ONCE(!core_rq)) /* impossible */ 5824 goto unlock; 5825 5826 /* copy the shared state to the new leader */ 5827 core_rq->core_task_seq = rq->core_task_seq; 5828 core_rq->core_pick_seq = rq->core_pick_seq; 5829 core_rq->core_cookie = rq->core_cookie; 5830 core_rq->core_forceidle = rq->core_forceidle; 5831 core_rq->core_forceidle_seq = rq->core_forceidle_seq; 5832 5833 /* install new leader */ 5834 for_each_cpu(t, smt_mask) { 5835 rq = cpu_rq(t); 5836 rq->core = core_rq; 5837 } 5838 5839 unlock: 5840 sched_core_unlock(cpu, &flags); 5841 } 5842 5843 static inline void sched_core_cpu_dying(unsigned int cpu) 5844 { 5845 struct rq *rq = cpu_rq(cpu); 5846 5847 if (rq->core != rq) 5848 rq->core = rq; 5849 } 5850 5851 #else /* !CONFIG_SCHED_CORE */ 5852 5853 static inline void sched_core_cpu_starting(unsigned int cpu) {} 5854 static inline void sched_core_cpu_deactivate(unsigned int cpu) {} 5855 static inline void sched_core_cpu_dying(unsigned int cpu) {} 5856 5857 static struct task_struct * 5858 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 5859 { 5860 return __pick_next_task(rq, prev, rf); 5861 } 5862 5863 #endif /* CONFIG_SCHED_CORE */ 5864 5865 /* 5866 * __schedule() is the main scheduler function. 5867 * 5868 * The main means of driving the scheduler and thus entering this function are: 5869 * 5870 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 5871 * 5872 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 5873 * paths. For example, see arch/x86/entry_64.S. 5874 * 5875 * To drive preemption between tasks, the scheduler sets the flag in timer 5876 * interrupt handler scheduler_tick(). 5877 * 5878 * 3. Wakeups don't really cause entry into schedule(). They add a 5879 * task to the run-queue and that's it. 5880 * 5881 * Now, if the new task added to the run-queue preempts the current 5882 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 5883 * called on the nearest possible occasion: 5884 * 5885 * - If the kernel is preemptible (CONFIG_PREEMPTION=y): 5886 * 5887 * - in syscall or exception context, at the next outmost 5888 * preempt_enable(). (this might be as soon as the wake_up()'s 5889 * spin_unlock()!) 5890 * 5891 * - in IRQ context, return from interrupt-handler to 5892 * preemptible context 5893 * 5894 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set) 5895 * then at the next: 5896 * 5897 * - cond_resched() call 5898 * - explicit schedule() call 5899 * - return from syscall or exception to user-space 5900 * - return from interrupt-handler to user-space 5901 * 5902 * WARNING: must be called with preemption disabled! 5903 */ 5904 static void __sched notrace __schedule(bool preempt) 5905 { 5906 struct task_struct *prev, *next; 5907 unsigned long *switch_count; 5908 unsigned long prev_state; 5909 struct rq_flags rf; 5910 struct rq *rq; 5911 int cpu; 5912 5913 cpu = smp_processor_id(); 5914 rq = cpu_rq(cpu); 5915 prev = rq->curr; 5916 5917 schedule_debug(prev, preempt); 5918 5919 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL)) 5920 hrtick_clear(rq); 5921 5922 local_irq_disable(); 5923 rcu_note_context_switch(preempt); 5924 5925 /* 5926 * Make sure that signal_pending_state()->signal_pending() below 5927 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 5928 * done by the caller to avoid the race with signal_wake_up(): 5929 * 5930 * __set_current_state(@state) signal_wake_up() 5931 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING) 5932 * wake_up_state(p, state) 5933 * LOCK rq->lock LOCK p->pi_state 5934 * smp_mb__after_spinlock() smp_mb__after_spinlock() 5935 * if (signal_pending_state()) if (p->state & @state) 5936 * 5937 * Also, the membarrier system call requires a full memory barrier 5938 * after coming from user-space, before storing to rq->curr. 5939 */ 5940 rq_lock(rq, &rf); 5941 smp_mb__after_spinlock(); 5942 5943 /* Promote REQ to ACT */ 5944 rq->clock_update_flags <<= 1; 5945 update_rq_clock(rq); 5946 5947 switch_count = &prev->nivcsw; 5948 5949 /* 5950 * We must load prev->state once (task_struct::state is volatile), such 5951 * that: 5952 * 5953 * - we form a control dependency vs deactivate_task() below. 5954 * - ptrace_{,un}freeze_traced() can change ->state underneath us. 5955 */ 5956 prev_state = READ_ONCE(prev->__state); 5957 if (!preempt && prev_state) { 5958 if (signal_pending_state(prev_state, prev)) { 5959 WRITE_ONCE(prev->__state, TASK_RUNNING); 5960 } else { 5961 prev->sched_contributes_to_load = 5962 (prev_state & TASK_UNINTERRUPTIBLE) && 5963 !(prev_state & TASK_NOLOAD) && 5964 !(prev->flags & PF_FROZEN); 5965 5966 if (prev->sched_contributes_to_load) 5967 rq->nr_uninterruptible++; 5968 5969 /* 5970 * __schedule() ttwu() 5971 * prev_state = prev->state; if (p->on_rq && ...) 5972 * if (prev_state) goto out; 5973 * p->on_rq = 0; smp_acquire__after_ctrl_dep(); 5974 * p->state = TASK_WAKING 5975 * 5976 * Where __schedule() and ttwu() have matching control dependencies. 5977 * 5978 * After this, schedule() must not care about p->state any more. 5979 */ 5980 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK); 5981 5982 if (prev->in_iowait) { 5983 atomic_inc(&rq->nr_iowait); 5984 delayacct_blkio_start(); 5985 } 5986 } 5987 switch_count = &prev->nvcsw; 5988 } 5989 5990 next = pick_next_task(rq, prev, &rf); 5991 clear_tsk_need_resched(prev); 5992 clear_preempt_need_resched(); 5993 #ifdef CONFIG_SCHED_DEBUG 5994 rq->last_seen_need_resched_ns = 0; 5995 #endif 5996 5997 if (likely(prev != next)) { 5998 rq->nr_switches++; 5999 /* 6000 * RCU users of rcu_dereference(rq->curr) may not see 6001 * changes to task_struct made by pick_next_task(). 6002 */ 6003 RCU_INIT_POINTER(rq->curr, next); 6004 /* 6005 * The membarrier system call requires each architecture 6006 * to have a full memory barrier after updating 6007 * rq->curr, before returning to user-space. 6008 * 6009 * Here are the schemes providing that barrier on the 6010 * various architectures: 6011 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC. 6012 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC. 6013 * - finish_lock_switch() for weakly-ordered 6014 * architectures where spin_unlock is a full barrier, 6015 * - switch_to() for arm64 (weakly-ordered, spin_unlock 6016 * is a RELEASE barrier), 6017 */ 6018 ++*switch_count; 6019 6020 migrate_disable_switch(rq, prev); 6021 psi_sched_switch(prev, next, !task_on_rq_queued(prev)); 6022 6023 trace_sched_switch(preempt, prev, next); 6024 6025 /* Also unlocks the rq: */ 6026 rq = context_switch(rq, prev, next, &rf); 6027 } else { 6028 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP); 6029 6030 rq_unpin_lock(rq, &rf); 6031 __balance_callbacks(rq); 6032 raw_spin_rq_unlock_irq(rq); 6033 } 6034 } 6035 6036 void __noreturn do_task_dead(void) 6037 { 6038 /* Causes final put_task_struct in finish_task_switch(): */ 6039 set_special_state(TASK_DEAD); 6040 6041 /* Tell freezer to ignore us: */ 6042 current->flags |= PF_NOFREEZE; 6043 6044 __schedule(false); 6045 BUG(); 6046 6047 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */ 6048 for (;;) 6049 cpu_relax(); 6050 } 6051 6052 static inline void sched_submit_work(struct task_struct *tsk) 6053 { 6054 unsigned int task_flags; 6055 6056 if (task_is_running(tsk)) 6057 return; 6058 6059 task_flags = tsk->flags; 6060 /* 6061 * If a worker went to sleep, notify and ask workqueue whether 6062 * it wants to wake up a task to maintain concurrency. 6063 * As this function is called inside the schedule() context, 6064 * we disable preemption to avoid it calling schedule() again 6065 * in the possible wakeup of a kworker and because wq_worker_sleeping() 6066 * requires it. 6067 */ 6068 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) { 6069 preempt_disable(); 6070 if (task_flags & PF_WQ_WORKER) 6071 wq_worker_sleeping(tsk); 6072 else 6073 io_wq_worker_sleeping(tsk); 6074 preempt_enable_no_resched(); 6075 } 6076 6077 if (tsk_is_pi_blocked(tsk)) 6078 return; 6079 6080 /* 6081 * If we are going to sleep and we have plugged IO queued, 6082 * make sure to submit it to avoid deadlocks. 6083 */ 6084 if (blk_needs_flush_plug(tsk)) 6085 blk_schedule_flush_plug(tsk); 6086 } 6087 6088 static void sched_update_worker(struct task_struct *tsk) 6089 { 6090 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) { 6091 if (tsk->flags & PF_WQ_WORKER) 6092 wq_worker_running(tsk); 6093 else 6094 io_wq_worker_running(tsk); 6095 } 6096 } 6097 6098 asmlinkage __visible void __sched schedule(void) 6099 { 6100 struct task_struct *tsk = current; 6101 6102 sched_submit_work(tsk); 6103 do { 6104 preempt_disable(); 6105 __schedule(false); 6106 sched_preempt_enable_no_resched(); 6107 } while (need_resched()); 6108 sched_update_worker(tsk); 6109 } 6110 EXPORT_SYMBOL(schedule); 6111 6112 /* 6113 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted 6114 * state (have scheduled out non-voluntarily) by making sure that all 6115 * tasks have either left the run queue or have gone into user space. 6116 * As idle tasks do not do either, they must not ever be preempted 6117 * (schedule out non-voluntarily). 6118 * 6119 * schedule_idle() is similar to schedule_preempt_disable() except that it 6120 * never enables preemption because it does not call sched_submit_work(). 6121 */ 6122 void __sched schedule_idle(void) 6123 { 6124 /* 6125 * As this skips calling sched_submit_work(), which the idle task does 6126 * regardless because that function is a nop when the task is in a 6127 * TASK_RUNNING state, make sure this isn't used someplace that the 6128 * current task can be in any other state. Note, idle is always in the 6129 * TASK_RUNNING state. 6130 */ 6131 WARN_ON_ONCE(current->__state); 6132 do { 6133 __schedule(false); 6134 } while (need_resched()); 6135 } 6136 6137 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK) 6138 asmlinkage __visible void __sched schedule_user(void) 6139 { 6140 /* 6141 * If we come here after a random call to set_need_resched(), 6142 * or we have been woken up remotely but the IPI has not yet arrived, 6143 * we haven't yet exited the RCU idle mode. Do it here manually until 6144 * we find a better solution. 6145 * 6146 * NB: There are buggy callers of this function. Ideally we 6147 * should warn if prev_state != CONTEXT_USER, but that will trigger 6148 * too frequently to make sense yet. 6149 */ 6150 enum ctx_state prev_state = exception_enter(); 6151 schedule(); 6152 exception_exit(prev_state); 6153 } 6154 #endif 6155 6156 /** 6157 * schedule_preempt_disabled - called with preemption disabled 6158 * 6159 * Returns with preemption disabled. Note: preempt_count must be 1 6160 */ 6161 void __sched schedule_preempt_disabled(void) 6162 { 6163 sched_preempt_enable_no_resched(); 6164 schedule(); 6165 preempt_disable(); 6166 } 6167 6168 static void __sched notrace preempt_schedule_common(void) 6169 { 6170 do { 6171 /* 6172 * Because the function tracer can trace preempt_count_sub() 6173 * and it also uses preempt_enable/disable_notrace(), if 6174 * NEED_RESCHED is set, the preempt_enable_notrace() called 6175 * by the function tracer will call this function again and 6176 * cause infinite recursion. 6177 * 6178 * Preemption must be disabled here before the function 6179 * tracer can trace. Break up preempt_disable() into two 6180 * calls. One to disable preemption without fear of being 6181 * traced. The other to still record the preemption latency, 6182 * which can also be traced by the function tracer. 6183 */ 6184 preempt_disable_notrace(); 6185 preempt_latency_start(1); 6186 __schedule(true); 6187 preempt_latency_stop(1); 6188 preempt_enable_no_resched_notrace(); 6189 6190 /* 6191 * Check again in case we missed a preemption opportunity 6192 * between schedule and now. 6193 */ 6194 } while (need_resched()); 6195 } 6196 6197 #ifdef CONFIG_PREEMPTION 6198 /* 6199 * This is the entry point to schedule() from in-kernel preemption 6200 * off of preempt_enable. 6201 */ 6202 asmlinkage __visible void __sched notrace preempt_schedule(void) 6203 { 6204 /* 6205 * If there is a non-zero preempt_count or interrupts are disabled, 6206 * we do not want to preempt the current task. Just return.. 6207 */ 6208 if (likely(!preemptible())) 6209 return; 6210 6211 preempt_schedule_common(); 6212 } 6213 NOKPROBE_SYMBOL(preempt_schedule); 6214 EXPORT_SYMBOL(preempt_schedule); 6215 6216 #ifdef CONFIG_PREEMPT_DYNAMIC 6217 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func); 6218 EXPORT_STATIC_CALL_TRAMP(preempt_schedule); 6219 #endif 6220 6221 6222 /** 6223 * preempt_schedule_notrace - preempt_schedule called by tracing 6224 * 6225 * The tracing infrastructure uses preempt_enable_notrace to prevent 6226 * recursion and tracing preempt enabling caused by the tracing 6227 * infrastructure itself. But as tracing can happen in areas coming 6228 * from userspace or just about to enter userspace, a preempt enable 6229 * can occur before user_exit() is called. This will cause the scheduler 6230 * to be called when the system is still in usermode. 6231 * 6232 * To prevent this, the preempt_enable_notrace will use this function 6233 * instead of preempt_schedule() to exit user context if needed before 6234 * calling the scheduler. 6235 */ 6236 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void) 6237 { 6238 enum ctx_state prev_ctx; 6239 6240 if (likely(!preemptible())) 6241 return; 6242 6243 do { 6244 /* 6245 * Because the function tracer can trace preempt_count_sub() 6246 * and it also uses preempt_enable/disable_notrace(), if 6247 * NEED_RESCHED is set, the preempt_enable_notrace() called 6248 * by the function tracer will call this function again and 6249 * cause infinite recursion. 6250 * 6251 * Preemption must be disabled here before the function 6252 * tracer can trace. Break up preempt_disable() into two 6253 * calls. One to disable preemption without fear of being 6254 * traced. The other to still record the preemption latency, 6255 * which can also be traced by the function tracer. 6256 */ 6257 preempt_disable_notrace(); 6258 preempt_latency_start(1); 6259 /* 6260 * Needs preempt disabled in case user_exit() is traced 6261 * and the tracer calls preempt_enable_notrace() causing 6262 * an infinite recursion. 6263 */ 6264 prev_ctx = exception_enter(); 6265 __schedule(true); 6266 exception_exit(prev_ctx); 6267 6268 preempt_latency_stop(1); 6269 preempt_enable_no_resched_notrace(); 6270 } while (need_resched()); 6271 } 6272 EXPORT_SYMBOL_GPL(preempt_schedule_notrace); 6273 6274 #ifdef CONFIG_PREEMPT_DYNAMIC 6275 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func); 6276 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace); 6277 #endif 6278 6279 #endif /* CONFIG_PREEMPTION */ 6280 6281 #ifdef CONFIG_PREEMPT_DYNAMIC 6282 6283 #include <linux/entry-common.h> 6284 6285 /* 6286 * SC:cond_resched 6287 * SC:might_resched 6288 * SC:preempt_schedule 6289 * SC:preempt_schedule_notrace 6290 * SC:irqentry_exit_cond_resched 6291 * 6292 * 6293 * NONE: 6294 * cond_resched <- __cond_resched 6295 * might_resched <- RET0 6296 * preempt_schedule <- NOP 6297 * preempt_schedule_notrace <- NOP 6298 * irqentry_exit_cond_resched <- NOP 6299 * 6300 * VOLUNTARY: 6301 * cond_resched <- __cond_resched 6302 * might_resched <- __cond_resched 6303 * preempt_schedule <- NOP 6304 * preempt_schedule_notrace <- NOP 6305 * irqentry_exit_cond_resched <- NOP 6306 * 6307 * FULL: 6308 * cond_resched <- RET0 6309 * might_resched <- RET0 6310 * preempt_schedule <- preempt_schedule 6311 * preempt_schedule_notrace <- preempt_schedule_notrace 6312 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched 6313 */ 6314 6315 enum { 6316 preempt_dynamic_none = 0, 6317 preempt_dynamic_voluntary, 6318 preempt_dynamic_full, 6319 }; 6320 6321 int preempt_dynamic_mode = preempt_dynamic_full; 6322 6323 int sched_dynamic_mode(const char *str) 6324 { 6325 if (!strcmp(str, "none")) 6326 return preempt_dynamic_none; 6327 6328 if (!strcmp(str, "voluntary")) 6329 return preempt_dynamic_voluntary; 6330 6331 if (!strcmp(str, "full")) 6332 return preempt_dynamic_full; 6333 6334 return -EINVAL; 6335 } 6336 6337 void sched_dynamic_update(int mode) 6338 { 6339 /* 6340 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in 6341 * the ZERO state, which is invalid. 6342 */ 6343 static_call_update(cond_resched, __cond_resched); 6344 static_call_update(might_resched, __cond_resched); 6345 static_call_update(preempt_schedule, __preempt_schedule_func); 6346 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func); 6347 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched); 6348 6349 switch (mode) { 6350 case preempt_dynamic_none: 6351 static_call_update(cond_resched, __cond_resched); 6352 static_call_update(might_resched, (void *)&__static_call_return0); 6353 static_call_update(preempt_schedule, NULL); 6354 static_call_update(preempt_schedule_notrace, NULL); 6355 static_call_update(irqentry_exit_cond_resched, NULL); 6356 pr_info("Dynamic Preempt: none\n"); 6357 break; 6358 6359 case preempt_dynamic_voluntary: 6360 static_call_update(cond_resched, __cond_resched); 6361 static_call_update(might_resched, __cond_resched); 6362 static_call_update(preempt_schedule, NULL); 6363 static_call_update(preempt_schedule_notrace, NULL); 6364 static_call_update(irqentry_exit_cond_resched, NULL); 6365 pr_info("Dynamic Preempt: voluntary\n"); 6366 break; 6367 6368 case preempt_dynamic_full: 6369 static_call_update(cond_resched, (void *)&__static_call_return0); 6370 static_call_update(might_resched, (void *)&__static_call_return0); 6371 static_call_update(preempt_schedule, __preempt_schedule_func); 6372 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func); 6373 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched); 6374 pr_info("Dynamic Preempt: full\n"); 6375 break; 6376 } 6377 6378 preempt_dynamic_mode = mode; 6379 } 6380 6381 static int __init setup_preempt_mode(char *str) 6382 { 6383 int mode = sched_dynamic_mode(str); 6384 if (mode < 0) { 6385 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str); 6386 return 1; 6387 } 6388 6389 sched_dynamic_update(mode); 6390 return 0; 6391 } 6392 __setup("preempt=", setup_preempt_mode); 6393 6394 #endif /* CONFIG_PREEMPT_DYNAMIC */ 6395 6396 /* 6397 * This is the entry point to schedule() from kernel preemption 6398 * off of irq context. 6399 * Note, that this is called and return with irqs disabled. This will 6400 * protect us against recursive calling from irq. 6401 */ 6402 asmlinkage __visible void __sched preempt_schedule_irq(void) 6403 { 6404 enum ctx_state prev_state; 6405 6406 /* Catch callers which need to be fixed */ 6407 BUG_ON(preempt_count() || !irqs_disabled()); 6408 6409 prev_state = exception_enter(); 6410 6411 do { 6412 preempt_disable(); 6413 local_irq_enable(); 6414 __schedule(true); 6415 local_irq_disable(); 6416 sched_preempt_enable_no_resched(); 6417 } while (need_resched()); 6418 6419 exception_exit(prev_state); 6420 } 6421 6422 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags, 6423 void *key) 6424 { 6425 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC); 6426 return try_to_wake_up(curr->private, mode, wake_flags); 6427 } 6428 EXPORT_SYMBOL(default_wake_function); 6429 6430 static void __setscheduler_prio(struct task_struct *p, int prio) 6431 { 6432 if (dl_prio(prio)) 6433 p->sched_class = &dl_sched_class; 6434 else if (rt_prio(prio)) 6435 p->sched_class = &rt_sched_class; 6436 else 6437 p->sched_class = &fair_sched_class; 6438 6439 p->prio = prio; 6440 } 6441 6442 #ifdef CONFIG_RT_MUTEXES 6443 6444 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio) 6445 { 6446 if (pi_task) 6447 prio = min(prio, pi_task->prio); 6448 6449 return prio; 6450 } 6451 6452 static inline int rt_effective_prio(struct task_struct *p, int prio) 6453 { 6454 struct task_struct *pi_task = rt_mutex_get_top_task(p); 6455 6456 return __rt_effective_prio(pi_task, prio); 6457 } 6458 6459 /* 6460 * rt_mutex_setprio - set the current priority of a task 6461 * @p: task to boost 6462 * @pi_task: donor task 6463 * 6464 * This function changes the 'effective' priority of a task. It does 6465 * not touch ->normal_prio like __setscheduler(). 6466 * 6467 * Used by the rt_mutex code to implement priority inheritance 6468 * logic. Call site only calls if the priority of the task changed. 6469 */ 6470 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task) 6471 { 6472 int prio, oldprio, queued, running, queue_flag = 6473 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; 6474 const struct sched_class *prev_class; 6475 struct rq_flags rf; 6476 struct rq *rq; 6477 6478 /* XXX used to be waiter->prio, not waiter->task->prio */ 6479 prio = __rt_effective_prio(pi_task, p->normal_prio); 6480 6481 /* 6482 * If nothing changed; bail early. 6483 */ 6484 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio)) 6485 return; 6486 6487 rq = __task_rq_lock(p, &rf); 6488 update_rq_clock(rq); 6489 /* 6490 * Set under pi_lock && rq->lock, such that the value can be used under 6491 * either lock. 6492 * 6493 * Note that there is loads of tricky to make this pointer cache work 6494 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to 6495 * ensure a task is de-boosted (pi_task is set to NULL) before the 6496 * task is allowed to run again (and can exit). This ensures the pointer 6497 * points to a blocked task -- which guarantees the task is present. 6498 */ 6499 p->pi_top_task = pi_task; 6500 6501 /* 6502 * For FIFO/RR we only need to set prio, if that matches we're done. 6503 */ 6504 if (prio == p->prio && !dl_prio(prio)) 6505 goto out_unlock; 6506 6507 /* 6508 * Idle task boosting is a nono in general. There is one 6509 * exception, when PREEMPT_RT and NOHZ is active: 6510 * 6511 * The idle task calls get_next_timer_interrupt() and holds 6512 * the timer wheel base->lock on the CPU and another CPU wants 6513 * to access the timer (probably to cancel it). We can safely 6514 * ignore the boosting request, as the idle CPU runs this code 6515 * with interrupts disabled and will complete the lock 6516 * protected section without being interrupted. So there is no 6517 * real need to boost. 6518 */ 6519 if (unlikely(p == rq->idle)) { 6520 WARN_ON(p != rq->curr); 6521 WARN_ON(p->pi_blocked_on); 6522 goto out_unlock; 6523 } 6524 6525 trace_sched_pi_setprio(p, pi_task); 6526 oldprio = p->prio; 6527 6528 if (oldprio == prio) 6529 queue_flag &= ~DEQUEUE_MOVE; 6530 6531 prev_class = p->sched_class; 6532 queued = task_on_rq_queued(p); 6533 running = task_current(rq, p); 6534 if (queued) 6535 dequeue_task(rq, p, queue_flag); 6536 if (running) 6537 put_prev_task(rq, p); 6538 6539 /* 6540 * Boosting condition are: 6541 * 1. -rt task is running and holds mutex A 6542 * --> -dl task blocks on mutex A 6543 * 6544 * 2. -dl task is running and holds mutex A 6545 * --> -dl task blocks on mutex A and could preempt the 6546 * running task 6547 */ 6548 if (dl_prio(prio)) { 6549 if (!dl_prio(p->normal_prio) || 6550 (pi_task && dl_prio(pi_task->prio) && 6551 dl_entity_preempt(&pi_task->dl, &p->dl))) { 6552 p->dl.pi_se = pi_task->dl.pi_se; 6553 queue_flag |= ENQUEUE_REPLENISH; 6554 } else { 6555 p->dl.pi_se = &p->dl; 6556 } 6557 } else if (rt_prio(prio)) { 6558 if (dl_prio(oldprio)) 6559 p->dl.pi_se = &p->dl; 6560 if (oldprio < prio) 6561 queue_flag |= ENQUEUE_HEAD; 6562 } else { 6563 if (dl_prio(oldprio)) 6564 p->dl.pi_se = &p->dl; 6565 if (rt_prio(oldprio)) 6566 p->rt.timeout = 0; 6567 } 6568 6569 __setscheduler_prio(p, prio); 6570 6571 if (queued) 6572 enqueue_task(rq, p, queue_flag); 6573 if (running) 6574 set_next_task(rq, p); 6575 6576 check_class_changed(rq, p, prev_class, oldprio); 6577 out_unlock: 6578 /* Avoid rq from going away on us: */ 6579 preempt_disable(); 6580 6581 rq_unpin_lock(rq, &rf); 6582 __balance_callbacks(rq); 6583 raw_spin_rq_unlock(rq); 6584 6585 preempt_enable(); 6586 } 6587 #else 6588 static inline int rt_effective_prio(struct task_struct *p, int prio) 6589 { 6590 return prio; 6591 } 6592 #endif 6593 6594 void set_user_nice(struct task_struct *p, long nice) 6595 { 6596 bool queued, running; 6597 int old_prio; 6598 struct rq_flags rf; 6599 struct rq *rq; 6600 6601 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 6602 return; 6603 /* 6604 * We have to be careful, if called from sys_setpriority(), 6605 * the task might be in the middle of scheduling on another CPU. 6606 */ 6607 rq = task_rq_lock(p, &rf); 6608 update_rq_clock(rq); 6609 6610 /* 6611 * The RT priorities are set via sched_setscheduler(), but we still 6612 * allow the 'normal' nice value to be set - but as expected 6613 * it won't have any effect on scheduling until the task is 6614 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 6615 */ 6616 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 6617 p->static_prio = NICE_TO_PRIO(nice); 6618 goto out_unlock; 6619 } 6620 queued = task_on_rq_queued(p); 6621 running = task_current(rq, p); 6622 if (queued) 6623 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK); 6624 if (running) 6625 put_prev_task(rq, p); 6626 6627 p->static_prio = NICE_TO_PRIO(nice); 6628 set_load_weight(p, true); 6629 old_prio = p->prio; 6630 p->prio = effective_prio(p); 6631 6632 if (queued) 6633 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); 6634 if (running) 6635 set_next_task(rq, p); 6636 6637 /* 6638 * If the task increased its priority or is running and 6639 * lowered its priority, then reschedule its CPU: 6640 */ 6641 p->sched_class->prio_changed(rq, p, old_prio); 6642 6643 out_unlock: 6644 task_rq_unlock(rq, p, &rf); 6645 } 6646 EXPORT_SYMBOL(set_user_nice); 6647 6648 /* 6649 * can_nice - check if a task can reduce its nice value 6650 * @p: task 6651 * @nice: nice value 6652 */ 6653 int can_nice(const struct task_struct *p, const int nice) 6654 { 6655 /* Convert nice value [19,-20] to rlimit style value [1,40]: */ 6656 int nice_rlim = nice_to_rlimit(nice); 6657 6658 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 6659 capable(CAP_SYS_NICE)); 6660 } 6661 6662 #ifdef __ARCH_WANT_SYS_NICE 6663 6664 /* 6665 * sys_nice - change the priority of the current process. 6666 * @increment: priority increment 6667 * 6668 * sys_setpriority is a more generic, but much slower function that 6669 * does similar things. 6670 */ 6671 SYSCALL_DEFINE1(nice, int, increment) 6672 { 6673 long nice, retval; 6674 6675 /* 6676 * Setpriority might change our priority at the same moment. 6677 * We don't have to worry. Conceptually one call occurs first 6678 * and we have a single winner. 6679 */ 6680 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 6681 nice = task_nice(current) + increment; 6682 6683 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 6684 if (increment < 0 && !can_nice(current, nice)) 6685 return -EPERM; 6686 6687 retval = security_task_setnice(current, nice); 6688 if (retval) 6689 return retval; 6690 6691 set_user_nice(current, nice); 6692 return 0; 6693 } 6694 6695 #endif 6696 6697 /** 6698 * task_prio - return the priority value of a given task. 6699 * @p: the task in question. 6700 * 6701 * Return: The priority value as seen by users in /proc. 6702 * 6703 * sched policy return value kernel prio user prio/nice 6704 * 6705 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19] 6706 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99] 6707 * deadline -101 -1 0 6708 */ 6709 int task_prio(const struct task_struct *p) 6710 { 6711 return p->prio - MAX_RT_PRIO; 6712 } 6713 6714 /** 6715 * idle_cpu - is a given CPU idle currently? 6716 * @cpu: the processor in question. 6717 * 6718 * Return: 1 if the CPU is currently idle. 0 otherwise. 6719 */ 6720 int idle_cpu(int cpu) 6721 { 6722 struct rq *rq = cpu_rq(cpu); 6723 6724 if (rq->curr != rq->idle) 6725 return 0; 6726 6727 if (rq->nr_running) 6728 return 0; 6729 6730 #ifdef CONFIG_SMP 6731 if (rq->ttwu_pending) 6732 return 0; 6733 #endif 6734 6735 return 1; 6736 } 6737 6738 /** 6739 * available_idle_cpu - is a given CPU idle for enqueuing work. 6740 * @cpu: the CPU in question. 6741 * 6742 * Return: 1 if the CPU is currently idle. 0 otherwise. 6743 */ 6744 int available_idle_cpu(int cpu) 6745 { 6746 if (!idle_cpu(cpu)) 6747 return 0; 6748 6749 if (vcpu_is_preempted(cpu)) 6750 return 0; 6751 6752 return 1; 6753 } 6754 6755 /** 6756 * idle_task - return the idle task for a given CPU. 6757 * @cpu: the processor in question. 6758 * 6759 * Return: The idle task for the CPU @cpu. 6760 */ 6761 struct task_struct *idle_task(int cpu) 6762 { 6763 return cpu_rq(cpu)->idle; 6764 } 6765 6766 #ifdef CONFIG_SMP 6767 /* 6768 * This function computes an effective utilization for the given CPU, to be 6769 * used for frequency selection given the linear relation: f = u * f_max. 6770 * 6771 * The scheduler tracks the following metrics: 6772 * 6773 * cpu_util_{cfs,rt,dl,irq}() 6774 * cpu_bw_dl() 6775 * 6776 * Where the cfs,rt and dl util numbers are tracked with the same metric and 6777 * synchronized windows and are thus directly comparable. 6778 * 6779 * The cfs,rt,dl utilization are the running times measured with rq->clock_task 6780 * which excludes things like IRQ and steal-time. These latter are then accrued 6781 * in the irq utilization. 6782 * 6783 * The DL bandwidth number otoh is not a measured metric but a value computed 6784 * based on the task model parameters and gives the minimal utilization 6785 * required to meet deadlines. 6786 */ 6787 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs, 6788 unsigned long max, enum cpu_util_type type, 6789 struct task_struct *p) 6790 { 6791 unsigned long dl_util, util, irq; 6792 struct rq *rq = cpu_rq(cpu); 6793 6794 if (!uclamp_is_used() && 6795 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) { 6796 return max; 6797 } 6798 6799 /* 6800 * Early check to see if IRQ/steal time saturates the CPU, can be 6801 * because of inaccuracies in how we track these -- see 6802 * update_irq_load_avg(). 6803 */ 6804 irq = cpu_util_irq(rq); 6805 if (unlikely(irq >= max)) 6806 return max; 6807 6808 /* 6809 * Because the time spend on RT/DL tasks is visible as 'lost' time to 6810 * CFS tasks and we use the same metric to track the effective 6811 * utilization (PELT windows are synchronized) we can directly add them 6812 * to obtain the CPU's actual utilization. 6813 * 6814 * CFS and RT utilization can be boosted or capped, depending on 6815 * utilization clamp constraints requested by currently RUNNABLE 6816 * tasks. 6817 * When there are no CFS RUNNABLE tasks, clamps are released and 6818 * frequency will be gracefully reduced with the utilization decay. 6819 */ 6820 util = util_cfs + cpu_util_rt(rq); 6821 if (type == FREQUENCY_UTIL) 6822 util = uclamp_rq_util_with(rq, util, p); 6823 6824 dl_util = cpu_util_dl(rq); 6825 6826 /* 6827 * For frequency selection we do not make cpu_util_dl() a permanent part 6828 * of this sum because we want to use cpu_bw_dl() later on, but we need 6829 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such 6830 * that we select f_max when there is no idle time. 6831 * 6832 * NOTE: numerical errors or stop class might cause us to not quite hit 6833 * saturation when we should -- something for later. 6834 */ 6835 if (util + dl_util >= max) 6836 return max; 6837 6838 /* 6839 * OTOH, for energy computation we need the estimated running time, so 6840 * include util_dl and ignore dl_bw. 6841 */ 6842 if (type == ENERGY_UTIL) 6843 util += dl_util; 6844 6845 /* 6846 * There is still idle time; further improve the number by using the 6847 * irq metric. Because IRQ/steal time is hidden from the task clock we 6848 * need to scale the task numbers: 6849 * 6850 * max - irq 6851 * U' = irq + --------- * U 6852 * max 6853 */ 6854 util = scale_irq_capacity(util, irq, max); 6855 util += irq; 6856 6857 /* 6858 * Bandwidth required by DEADLINE must always be granted while, for 6859 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism 6860 * to gracefully reduce the frequency when no tasks show up for longer 6861 * periods of time. 6862 * 6863 * Ideally we would like to set bw_dl as min/guaranteed freq and util + 6864 * bw_dl as requested freq. However, cpufreq is not yet ready for such 6865 * an interface. So, we only do the latter for now. 6866 */ 6867 if (type == FREQUENCY_UTIL) 6868 util += cpu_bw_dl(rq); 6869 6870 return min(max, util); 6871 } 6872 6873 unsigned long sched_cpu_util(int cpu, unsigned long max) 6874 { 6875 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max, 6876 ENERGY_UTIL, NULL); 6877 } 6878 #endif /* CONFIG_SMP */ 6879 6880 /** 6881 * find_process_by_pid - find a process with a matching PID value. 6882 * @pid: the pid in question. 6883 * 6884 * The task of @pid, if found. %NULL otherwise. 6885 */ 6886 static struct task_struct *find_process_by_pid(pid_t pid) 6887 { 6888 return pid ? find_task_by_vpid(pid) : current; 6889 } 6890 6891 /* 6892 * sched_setparam() passes in -1 for its policy, to let the functions 6893 * it calls know not to change it. 6894 */ 6895 #define SETPARAM_POLICY -1 6896 6897 static void __setscheduler_params(struct task_struct *p, 6898 const struct sched_attr *attr) 6899 { 6900 int policy = attr->sched_policy; 6901 6902 if (policy == SETPARAM_POLICY) 6903 policy = p->policy; 6904 6905 p->policy = policy; 6906 6907 if (dl_policy(policy)) 6908 __setparam_dl(p, attr); 6909 else if (fair_policy(policy)) 6910 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 6911 6912 /* 6913 * __sched_setscheduler() ensures attr->sched_priority == 0 when 6914 * !rt_policy. Always setting this ensures that things like 6915 * getparam()/getattr() don't report silly values for !rt tasks. 6916 */ 6917 p->rt_priority = attr->sched_priority; 6918 p->normal_prio = normal_prio(p); 6919 set_load_weight(p, true); 6920 } 6921 6922 /* 6923 * Check the target process has a UID that matches the current process's: 6924 */ 6925 static bool check_same_owner(struct task_struct *p) 6926 { 6927 const struct cred *cred = current_cred(), *pcred; 6928 bool match; 6929 6930 rcu_read_lock(); 6931 pcred = __task_cred(p); 6932 match = (uid_eq(cred->euid, pcred->euid) || 6933 uid_eq(cred->euid, pcred->uid)); 6934 rcu_read_unlock(); 6935 return match; 6936 } 6937 6938 static int __sched_setscheduler(struct task_struct *p, 6939 const struct sched_attr *attr, 6940 bool user, bool pi) 6941 { 6942 int oldpolicy = -1, policy = attr->sched_policy; 6943 int retval, oldprio, newprio, queued, running; 6944 const struct sched_class *prev_class; 6945 struct callback_head *head; 6946 struct rq_flags rf; 6947 int reset_on_fork; 6948 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; 6949 struct rq *rq; 6950 6951 /* The pi code expects interrupts enabled */ 6952 BUG_ON(pi && in_interrupt()); 6953 recheck: 6954 /* Double check policy once rq lock held: */ 6955 if (policy < 0) { 6956 reset_on_fork = p->sched_reset_on_fork; 6957 policy = oldpolicy = p->policy; 6958 } else { 6959 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 6960 6961 if (!valid_policy(policy)) 6962 return -EINVAL; 6963 } 6964 6965 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV)) 6966 return -EINVAL; 6967 6968 /* 6969 * Valid priorities for SCHED_FIFO and SCHED_RR are 6970 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL, 6971 * SCHED_BATCH and SCHED_IDLE is 0. 6972 */ 6973 if (attr->sched_priority > MAX_RT_PRIO-1) 6974 return -EINVAL; 6975 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 6976 (rt_policy(policy) != (attr->sched_priority != 0))) 6977 return -EINVAL; 6978 6979 /* 6980 * Allow unprivileged RT tasks to decrease priority: 6981 */ 6982 if (user && !capable(CAP_SYS_NICE)) { 6983 if (fair_policy(policy)) { 6984 if (attr->sched_nice < task_nice(p) && 6985 !can_nice(p, attr->sched_nice)) 6986 return -EPERM; 6987 } 6988 6989 if (rt_policy(policy)) { 6990 unsigned long rlim_rtprio = 6991 task_rlimit(p, RLIMIT_RTPRIO); 6992 6993 /* Can't set/change the rt policy: */ 6994 if (policy != p->policy && !rlim_rtprio) 6995 return -EPERM; 6996 6997 /* Can't increase priority: */ 6998 if (attr->sched_priority > p->rt_priority && 6999 attr->sched_priority > rlim_rtprio) 7000 return -EPERM; 7001 } 7002 7003 /* 7004 * Can't set/change SCHED_DEADLINE policy at all for now 7005 * (safest behavior); in the future we would like to allow 7006 * unprivileged DL tasks to increase their relative deadline 7007 * or reduce their runtime (both ways reducing utilization) 7008 */ 7009 if (dl_policy(policy)) 7010 return -EPERM; 7011 7012 /* 7013 * Treat SCHED_IDLE as nice 20. Only allow a switch to 7014 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 7015 */ 7016 if (task_has_idle_policy(p) && !idle_policy(policy)) { 7017 if (!can_nice(p, task_nice(p))) 7018 return -EPERM; 7019 } 7020 7021 /* Can't change other user's priorities: */ 7022 if (!check_same_owner(p)) 7023 return -EPERM; 7024 7025 /* Normal users shall not reset the sched_reset_on_fork flag: */ 7026 if (p->sched_reset_on_fork && !reset_on_fork) 7027 return -EPERM; 7028 } 7029 7030 if (user) { 7031 if (attr->sched_flags & SCHED_FLAG_SUGOV) 7032 return -EINVAL; 7033 7034 retval = security_task_setscheduler(p); 7035 if (retval) 7036 return retval; 7037 } 7038 7039 /* Update task specific "requested" clamps */ 7040 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) { 7041 retval = uclamp_validate(p, attr); 7042 if (retval) 7043 return retval; 7044 } 7045 7046 if (pi) 7047 cpuset_read_lock(); 7048 7049 /* 7050 * Make sure no PI-waiters arrive (or leave) while we are 7051 * changing the priority of the task: 7052 * 7053 * To be able to change p->policy safely, the appropriate 7054 * runqueue lock must be held. 7055 */ 7056 rq = task_rq_lock(p, &rf); 7057 update_rq_clock(rq); 7058 7059 /* 7060 * Changing the policy of the stop threads its a very bad idea: 7061 */ 7062 if (p == rq->stop) { 7063 retval = -EINVAL; 7064 goto unlock; 7065 } 7066 7067 /* 7068 * If not changing anything there's no need to proceed further, 7069 * but store a possible modification of reset_on_fork. 7070 */ 7071 if (unlikely(policy == p->policy)) { 7072 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 7073 goto change; 7074 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 7075 goto change; 7076 if (dl_policy(policy) && dl_param_changed(p, attr)) 7077 goto change; 7078 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) 7079 goto change; 7080 7081 p->sched_reset_on_fork = reset_on_fork; 7082 retval = 0; 7083 goto unlock; 7084 } 7085 change: 7086 7087 if (user) { 7088 #ifdef CONFIG_RT_GROUP_SCHED 7089 /* 7090 * Do not allow realtime tasks into groups that have no runtime 7091 * assigned. 7092 */ 7093 if (rt_bandwidth_enabled() && rt_policy(policy) && 7094 task_group(p)->rt_bandwidth.rt_runtime == 0 && 7095 !task_group_is_autogroup(task_group(p))) { 7096 retval = -EPERM; 7097 goto unlock; 7098 } 7099 #endif 7100 #ifdef CONFIG_SMP 7101 if (dl_bandwidth_enabled() && dl_policy(policy) && 7102 !(attr->sched_flags & SCHED_FLAG_SUGOV)) { 7103 cpumask_t *span = rq->rd->span; 7104 7105 /* 7106 * Don't allow tasks with an affinity mask smaller than 7107 * the entire root_domain to become SCHED_DEADLINE. We 7108 * will also fail if there's no bandwidth available. 7109 */ 7110 if (!cpumask_subset(span, p->cpus_ptr) || 7111 rq->rd->dl_bw.bw == 0) { 7112 retval = -EPERM; 7113 goto unlock; 7114 } 7115 } 7116 #endif 7117 } 7118 7119 /* Re-check policy now with rq lock held: */ 7120 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 7121 policy = oldpolicy = -1; 7122 task_rq_unlock(rq, p, &rf); 7123 if (pi) 7124 cpuset_read_unlock(); 7125 goto recheck; 7126 } 7127 7128 /* 7129 * If setscheduling to SCHED_DEADLINE (or changing the parameters 7130 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 7131 * is available. 7132 */ 7133 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) { 7134 retval = -EBUSY; 7135 goto unlock; 7136 } 7137 7138 p->sched_reset_on_fork = reset_on_fork; 7139 oldprio = p->prio; 7140 7141 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice); 7142 if (pi) { 7143 /* 7144 * Take priority boosted tasks into account. If the new 7145 * effective priority is unchanged, we just store the new 7146 * normal parameters and do not touch the scheduler class and 7147 * the runqueue. This will be done when the task deboost 7148 * itself. 7149 */ 7150 newprio = rt_effective_prio(p, newprio); 7151 if (newprio == oldprio) 7152 queue_flags &= ~DEQUEUE_MOVE; 7153 } 7154 7155 queued = task_on_rq_queued(p); 7156 running = task_current(rq, p); 7157 if (queued) 7158 dequeue_task(rq, p, queue_flags); 7159 if (running) 7160 put_prev_task(rq, p); 7161 7162 prev_class = p->sched_class; 7163 7164 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) { 7165 __setscheduler_params(p, attr); 7166 __setscheduler_prio(p, newprio); 7167 } 7168 __setscheduler_uclamp(p, attr); 7169 7170 if (queued) { 7171 /* 7172 * We enqueue to tail when the priority of a task is 7173 * increased (user space view). 7174 */ 7175 if (oldprio < p->prio) 7176 queue_flags |= ENQUEUE_HEAD; 7177 7178 enqueue_task(rq, p, queue_flags); 7179 } 7180 if (running) 7181 set_next_task(rq, p); 7182 7183 check_class_changed(rq, p, prev_class, oldprio); 7184 7185 /* Avoid rq from going away on us: */ 7186 preempt_disable(); 7187 head = splice_balance_callbacks(rq); 7188 task_rq_unlock(rq, p, &rf); 7189 7190 if (pi) { 7191 cpuset_read_unlock(); 7192 rt_mutex_adjust_pi(p); 7193 } 7194 7195 /* Run balance callbacks after we've adjusted the PI chain: */ 7196 balance_callbacks(rq, head); 7197 preempt_enable(); 7198 7199 return 0; 7200 7201 unlock: 7202 task_rq_unlock(rq, p, &rf); 7203 if (pi) 7204 cpuset_read_unlock(); 7205 return retval; 7206 } 7207 7208 static int _sched_setscheduler(struct task_struct *p, int policy, 7209 const struct sched_param *param, bool check) 7210 { 7211 struct sched_attr attr = { 7212 .sched_policy = policy, 7213 .sched_priority = param->sched_priority, 7214 .sched_nice = PRIO_TO_NICE(p->static_prio), 7215 }; 7216 7217 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ 7218 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { 7219 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 7220 policy &= ~SCHED_RESET_ON_FORK; 7221 attr.sched_policy = policy; 7222 } 7223 7224 return __sched_setscheduler(p, &attr, check, true); 7225 } 7226 /** 7227 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 7228 * @p: the task in question. 7229 * @policy: new policy. 7230 * @param: structure containing the new RT priority. 7231 * 7232 * Use sched_set_fifo(), read its comment. 7233 * 7234 * Return: 0 on success. An error code otherwise. 7235 * 7236 * NOTE that the task may be already dead. 7237 */ 7238 int sched_setscheduler(struct task_struct *p, int policy, 7239 const struct sched_param *param) 7240 { 7241 return _sched_setscheduler(p, policy, param, true); 7242 } 7243 7244 int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 7245 { 7246 return __sched_setscheduler(p, attr, true, true); 7247 } 7248 7249 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr) 7250 { 7251 return __sched_setscheduler(p, attr, false, true); 7252 } 7253 EXPORT_SYMBOL_GPL(sched_setattr_nocheck); 7254 7255 /** 7256 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 7257 * @p: the task in question. 7258 * @policy: new policy. 7259 * @param: structure containing the new RT priority. 7260 * 7261 * Just like sched_setscheduler, only don't bother checking if the 7262 * current context has permission. For example, this is needed in 7263 * stop_machine(): we create temporary high priority worker threads, 7264 * but our caller might not have that capability. 7265 * 7266 * Return: 0 on success. An error code otherwise. 7267 */ 7268 int sched_setscheduler_nocheck(struct task_struct *p, int policy, 7269 const struct sched_param *param) 7270 { 7271 return _sched_setscheduler(p, policy, param, false); 7272 } 7273 7274 /* 7275 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally 7276 * incapable of resource management, which is the one thing an OS really should 7277 * be doing. 7278 * 7279 * This is of course the reason it is limited to privileged users only. 7280 * 7281 * Worse still; it is fundamentally impossible to compose static priority 7282 * workloads. You cannot take two correctly working static prio workloads 7283 * and smash them together and still expect them to work. 7284 * 7285 * For this reason 'all' FIFO tasks the kernel creates are basically at: 7286 * 7287 * MAX_RT_PRIO / 2 7288 * 7289 * The administrator _MUST_ configure the system, the kernel simply doesn't 7290 * know enough information to make a sensible choice. 7291 */ 7292 void sched_set_fifo(struct task_struct *p) 7293 { 7294 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 }; 7295 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0); 7296 } 7297 EXPORT_SYMBOL_GPL(sched_set_fifo); 7298 7299 /* 7300 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL. 7301 */ 7302 void sched_set_fifo_low(struct task_struct *p) 7303 { 7304 struct sched_param sp = { .sched_priority = 1 }; 7305 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0); 7306 } 7307 EXPORT_SYMBOL_GPL(sched_set_fifo_low); 7308 7309 void sched_set_normal(struct task_struct *p, int nice) 7310 { 7311 struct sched_attr attr = { 7312 .sched_policy = SCHED_NORMAL, 7313 .sched_nice = nice, 7314 }; 7315 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0); 7316 } 7317 EXPORT_SYMBOL_GPL(sched_set_normal); 7318 7319 static int 7320 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 7321 { 7322 struct sched_param lparam; 7323 struct task_struct *p; 7324 int retval; 7325 7326 if (!param || pid < 0) 7327 return -EINVAL; 7328 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 7329 return -EFAULT; 7330 7331 rcu_read_lock(); 7332 retval = -ESRCH; 7333 p = find_process_by_pid(pid); 7334 if (likely(p)) 7335 get_task_struct(p); 7336 rcu_read_unlock(); 7337 7338 if (likely(p)) { 7339 retval = sched_setscheduler(p, policy, &lparam); 7340 put_task_struct(p); 7341 } 7342 7343 return retval; 7344 } 7345 7346 /* 7347 * Mimics kernel/events/core.c perf_copy_attr(). 7348 */ 7349 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr) 7350 { 7351 u32 size; 7352 int ret; 7353 7354 /* Zero the full structure, so that a short copy will be nice: */ 7355 memset(attr, 0, sizeof(*attr)); 7356 7357 ret = get_user(size, &uattr->size); 7358 if (ret) 7359 return ret; 7360 7361 /* ABI compatibility quirk: */ 7362 if (!size) 7363 size = SCHED_ATTR_SIZE_VER0; 7364 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE) 7365 goto err_size; 7366 7367 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size); 7368 if (ret) { 7369 if (ret == -E2BIG) 7370 goto err_size; 7371 return ret; 7372 } 7373 7374 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) && 7375 size < SCHED_ATTR_SIZE_VER1) 7376 return -EINVAL; 7377 7378 /* 7379 * XXX: Do we want to be lenient like existing syscalls; or do we want 7380 * to be strict and return an error on out-of-bounds values? 7381 */ 7382 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 7383 7384 return 0; 7385 7386 err_size: 7387 put_user(sizeof(*attr), &uattr->size); 7388 return -E2BIG; 7389 } 7390 7391 /** 7392 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 7393 * @pid: the pid in question. 7394 * @policy: new policy. 7395 * @param: structure containing the new RT priority. 7396 * 7397 * Return: 0 on success. An error code otherwise. 7398 */ 7399 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param) 7400 { 7401 if (policy < 0) 7402 return -EINVAL; 7403 7404 return do_sched_setscheduler(pid, policy, param); 7405 } 7406 7407 /** 7408 * sys_sched_setparam - set/change the RT priority of a thread 7409 * @pid: the pid in question. 7410 * @param: structure containing the new RT priority. 7411 * 7412 * Return: 0 on success. An error code otherwise. 7413 */ 7414 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 7415 { 7416 return do_sched_setscheduler(pid, SETPARAM_POLICY, param); 7417 } 7418 7419 /** 7420 * sys_sched_setattr - same as above, but with extended sched_attr 7421 * @pid: the pid in question. 7422 * @uattr: structure containing the extended parameters. 7423 * @flags: for future extension. 7424 */ 7425 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 7426 unsigned int, flags) 7427 { 7428 struct sched_attr attr; 7429 struct task_struct *p; 7430 int retval; 7431 7432 if (!uattr || pid < 0 || flags) 7433 return -EINVAL; 7434 7435 retval = sched_copy_attr(uattr, &attr); 7436 if (retval) 7437 return retval; 7438 7439 if ((int)attr.sched_policy < 0) 7440 return -EINVAL; 7441 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY) 7442 attr.sched_policy = SETPARAM_POLICY; 7443 7444 rcu_read_lock(); 7445 retval = -ESRCH; 7446 p = find_process_by_pid(pid); 7447 if (likely(p)) 7448 get_task_struct(p); 7449 rcu_read_unlock(); 7450 7451 if (likely(p)) { 7452 retval = sched_setattr(p, &attr); 7453 put_task_struct(p); 7454 } 7455 7456 return retval; 7457 } 7458 7459 /** 7460 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 7461 * @pid: the pid in question. 7462 * 7463 * Return: On success, the policy of the thread. Otherwise, a negative error 7464 * code. 7465 */ 7466 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 7467 { 7468 struct task_struct *p; 7469 int retval; 7470 7471 if (pid < 0) 7472 return -EINVAL; 7473 7474 retval = -ESRCH; 7475 rcu_read_lock(); 7476 p = find_process_by_pid(pid); 7477 if (p) { 7478 retval = security_task_getscheduler(p); 7479 if (!retval) 7480 retval = p->policy 7481 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 7482 } 7483 rcu_read_unlock(); 7484 return retval; 7485 } 7486 7487 /** 7488 * sys_sched_getparam - get the RT priority of a thread 7489 * @pid: the pid in question. 7490 * @param: structure containing the RT priority. 7491 * 7492 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 7493 * code. 7494 */ 7495 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 7496 { 7497 struct sched_param lp = { .sched_priority = 0 }; 7498 struct task_struct *p; 7499 int retval; 7500 7501 if (!param || pid < 0) 7502 return -EINVAL; 7503 7504 rcu_read_lock(); 7505 p = find_process_by_pid(pid); 7506 retval = -ESRCH; 7507 if (!p) 7508 goto out_unlock; 7509 7510 retval = security_task_getscheduler(p); 7511 if (retval) 7512 goto out_unlock; 7513 7514 if (task_has_rt_policy(p)) 7515 lp.sched_priority = p->rt_priority; 7516 rcu_read_unlock(); 7517 7518 /* 7519 * This one might sleep, we cannot do it with a spinlock held ... 7520 */ 7521 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 7522 7523 return retval; 7524 7525 out_unlock: 7526 rcu_read_unlock(); 7527 return retval; 7528 } 7529 7530 /* 7531 * Copy the kernel size attribute structure (which might be larger 7532 * than what user-space knows about) to user-space. 7533 * 7534 * Note that all cases are valid: user-space buffer can be larger or 7535 * smaller than the kernel-space buffer. The usual case is that both 7536 * have the same size. 7537 */ 7538 static int 7539 sched_attr_copy_to_user(struct sched_attr __user *uattr, 7540 struct sched_attr *kattr, 7541 unsigned int usize) 7542 { 7543 unsigned int ksize = sizeof(*kattr); 7544 7545 if (!access_ok(uattr, usize)) 7546 return -EFAULT; 7547 7548 /* 7549 * sched_getattr() ABI forwards and backwards compatibility: 7550 * 7551 * If usize == ksize then we just copy everything to user-space and all is good. 7552 * 7553 * If usize < ksize then we only copy as much as user-space has space for, 7554 * this keeps ABI compatibility as well. We skip the rest. 7555 * 7556 * If usize > ksize then user-space is using a newer version of the ABI, 7557 * which part the kernel doesn't know about. Just ignore it - tooling can 7558 * detect the kernel's knowledge of attributes from the attr->size value 7559 * which is set to ksize in this case. 7560 */ 7561 kattr->size = min(usize, ksize); 7562 7563 if (copy_to_user(uattr, kattr, kattr->size)) 7564 return -EFAULT; 7565 7566 return 0; 7567 } 7568 7569 /** 7570 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 7571 * @pid: the pid in question. 7572 * @uattr: structure containing the extended parameters. 7573 * @usize: sizeof(attr) for fwd/bwd comp. 7574 * @flags: for future extension. 7575 */ 7576 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 7577 unsigned int, usize, unsigned int, flags) 7578 { 7579 struct sched_attr kattr = { }; 7580 struct task_struct *p; 7581 int retval; 7582 7583 if (!uattr || pid < 0 || usize > PAGE_SIZE || 7584 usize < SCHED_ATTR_SIZE_VER0 || flags) 7585 return -EINVAL; 7586 7587 rcu_read_lock(); 7588 p = find_process_by_pid(pid); 7589 retval = -ESRCH; 7590 if (!p) 7591 goto out_unlock; 7592 7593 retval = security_task_getscheduler(p); 7594 if (retval) 7595 goto out_unlock; 7596 7597 kattr.sched_policy = p->policy; 7598 if (p->sched_reset_on_fork) 7599 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 7600 if (task_has_dl_policy(p)) 7601 __getparam_dl(p, &kattr); 7602 else if (task_has_rt_policy(p)) 7603 kattr.sched_priority = p->rt_priority; 7604 else 7605 kattr.sched_nice = task_nice(p); 7606 7607 #ifdef CONFIG_UCLAMP_TASK 7608 /* 7609 * This could race with another potential updater, but this is fine 7610 * because it'll correctly read the old or the new value. We don't need 7611 * to guarantee who wins the race as long as it doesn't return garbage. 7612 */ 7613 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value; 7614 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value; 7615 #endif 7616 7617 rcu_read_unlock(); 7618 7619 return sched_attr_copy_to_user(uattr, &kattr, usize); 7620 7621 out_unlock: 7622 rcu_read_unlock(); 7623 return retval; 7624 } 7625 7626 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 7627 { 7628 cpumask_var_t cpus_allowed, new_mask; 7629 struct task_struct *p; 7630 int retval; 7631 7632 rcu_read_lock(); 7633 7634 p = find_process_by_pid(pid); 7635 if (!p) { 7636 rcu_read_unlock(); 7637 return -ESRCH; 7638 } 7639 7640 /* Prevent p going away */ 7641 get_task_struct(p); 7642 rcu_read_unlock(); 7643 7644 if (p->flags & PF_NO_SETAFFINITY) { 7645 retval = -EINVAL; 7646 goto out_put_task; 7647 } 7648 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 7649 retval = -ENOMEM; 7650 goto out_put_task; 7651 } 7652 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 7653 retval = -ENOMEM; 7654 goto out_free_cpus_allowed; 7655 } 7656 retval = -EPERM; 7657 if (!check_same_owner(p)) { 7658 rcu_read_lock(); 7659 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 7660 rcu_read_unlock(); 7661 goto out_free_new_mask; 7662 } 7663 rcu_read_unlock(); 7664 } 7665 7666 retval = security_task_setscheduler(p); 7667 if (retval) 7668 goto out_free_new_mask; 7669 7670 7671 cpuset_cpus_allowed(p, cpus_allowed); 7672 cpumask_and(new_mask, in_mask, cpus_allowed); 7673 7674 /* 7675 * Since bandwidth control happens on root_domain basis, 7676 * if admission test is enabled, we only admit -deadline 7677 * tasks allowed to run on all the CPUs in the task's 7678 * root_domain. 7679 */ 7680 #ifdef CONFIG_SMP 7681 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { 7682 rcu_read_lock(); 7683 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { 7684 retval = -EBUSY; 7685 rcu_read_unlock(); 7686 goto out_free_new_mask; 7687 } 7688 rcu_read_unlock(); 7689 } 7690 #endif 7691 again: 7692 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK); 7693 7694 if (!retval) { 7695 cpuset_cpus_allowed(p, cpus_allowed); 7696 if (!cpumask_subset(new_mask, cpus_allowed)) { 7697 /* 7698 * We must have raced with a concurrent cpuset 7699 * update. Just reset the cpus_allowed to the 7700 * cpuset's cpus_allowed 7701 */ 7702 cpumask_copy(new_mask, cpus_allowed); 7703 goto again; 7704 } 7705 } 7706 out_free_new_mask: 7707 free_cpumask_var(new_mask); 7708 out_free_cpus_allowed: 7709 free_cpumask_var(cpus_allowed); 7710 out_put_task: 7711 put_task_struct(p); 7712 return retval; 7713 } 7714 7715 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 7716 struct cpumask *new_mask) 7717 { 7718 if (len < cpumask_size()) 7719 cpumask_clear(new_mask); 7720 else if (len > cpumask_size()) 7721 len = cpumask_size(); 7722 7723 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 7724 } 7725 7726 /** 7727 * sys_sched_setaffinity - set the CPU affinity of a process 7728 * @pid: pid of the process 7729 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 7730 * @user_mask_ptr: user-space pointer to the new CPU mask 7731 * 7732 * Return: 0 on success. An error code otherwise. 7733 */ 7734 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 7735 unsigned long __user *, user_mask_ptr) 7736 { 7737 cpumask_var_t new_mask; 7738 int retval; 7739 7740 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 7741 return -ENOMEM; 7742 7743 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 7744 if (retval == 0) 7745 retval = sched_setaffinity(pid, new_mask); 7746 free_cpumask_var(new_mask); 7747 return retval; 7748 } 7749 7750 long sched_getaffinity(pid_t pid, struct cpumask *mask) 7751 { 7752 struct task_struct *p; 7753 unsigned long flags; 7754 int retval; 7755 7756 rcu_read_lock(); 7757 7758 retval = -ESRCH; 7759 p = find_process_by_pid(pid); 7760 if (!p) 7761 goto out_unlock; 7762 7763 retval = security_task_getscheduler(p); 7764 if (retval) 7765 goto out_unlock; 7766 7767 raw_spin_lock_irqsave(&p->pi_lock, flags); 7768 cpumask_and(mask, &p->cpus_mask, cpu_active_mask); 7769 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 7770 7771 out_unlock: 7772 rcu_read_unlock(); 7773 7774 return retval; 7775 } 7776 7777 /** 7778 * sys_sched_getaffinity - get the CPU affinity of a process 7779 * @pid: pid of the process 7780 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 7781 * @user_mask_ptr: user-space pointer to hold the current CPU mask 7782 * 7783 * Return: size of CPU mask copied to user_mask_ptr on success. An 7784 * error code otherwise. 7785 */ 7786 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 7787 unsigned long __user *, user_mask_ptr) 7788 { 7789 int ret; 7790 cpumask_var_t mask; 7791 7792 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 7793 return -EINVAL; 7794 if (len & (sizeof(unsigned long)-1)) 7795 return -EINVAL; 7796 7797 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 7798 return -ENOMEM; 7799 7800 ret = sched_getaffinity(pid, mask); 7801 if (ret == 0) { 7802 unsigned int retlen = min(len, cpumask_size()); 7803 7804 if (copy_to_user(user_mask_ptr, mask, retlen)) 7805 ret = -EFAULT; 7806 else 7807 ret = retlen; 7808 } 7809 free_cpumask_var(mask); 7810 7811 return ret; 7812 } 7813 7814 static void do_sched_yield(void) 7815 { 7816 struct rq_flags rf; 7817 struct rq *rq; 7818 7819 rq = this_rq_lock_irq(&rf); 7820 7821 schedstat_inc(rq->yld_count); 7822 current->sched_class->yield_task(rq); 7823 7824 preempt_disable(); 7825 rq_unlock_irq(rq, &rf); 7826 sched_preempt_enable_no_resched(); 7827 7828 schedule(); 7829 } 7830 7831 /** 7832 * sys_sched_yield - yield the current processor to other threads. 7833 * 7834 * This function yields the current CPU to other tasks. If there are no 7835 * other threads running on this CPU then this function will return. 7836 * 7837 * Return: 0. 7838 */ 7839 SYSCALL_DEFINE0(sched_yield) 7840 { 7841 do_sched_yield(); 7842 return 0; 7843 } 7844 7845 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC) 7846 int __sched __cond_resched(void) 7847 { 7848 if (should_resched(0)) { 7849 preempt_schedule_common(); 7850 return 1; 7851 } 7852 #ifndef CONFIG_PREEMPT_RCU 7853 rcu_all_qs(); 7854 #endif 7855 return 0; 7856 } 7857 EXPORT_SYMBOL(__cond_resched); 7858 #endif 7859 7860 #ifdef CONFIG_PREEMPT_DYNAMIC 7861 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched); 7862 EXPORT_STATIC_CALL_TRAMP(cond_resched); 7863 7864 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched); 7865 EXPORT_STATIC_CALL_TRAMP(might_resched); 7866 #endif 7867 7868 /* 7869 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 7870 * call schedule, and on return reacquire the lock. 7871 * 7872 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level 7873 * operations here to prevent schedule() from being called twice (once via 7874 * spin_unlock(), once by hand). 7875 */ 7876 int __cond_resched_lock(spinlock_t *lock) 7877 { 7878 int resched = should_resched(PREEMPT_LOCK_OFFSET); 7879 int ret = 0; 7880 7881 lockdep_assert_held(lock); 7882 7883 if (spin_needbreak(lock) || resched) { 7884 spin_unlock(lock); 7885 if (resched) 7886 preempt_schedule_common(); 7887 else 7888 cpu_relax(); 7889 ret = 1; 7890 spin_lock(lock); 7891 } 7892 return ret; 7893 } 7894 EXPORT_SYMBOL(__cond_resched_lock); 7895 7896 int __cond_resched_rwlock_read(rwlock_t *lock) 7897 { 7898 int resched = should_resched(PREEMPT_LOCK_OFFSET); 7899 int ret = 0; 7900 7901 lockdep_assert_held_read(lock); 7902 7903 if (rwlock_needbreak(lock) || resched) { 7904 read_unlock(lock); 7905 if (resched) 7906 preempt_schedule_common(); 7907 else 7908 cpu_relax(); 7909 ret = 1; 7910 read_lock(lock); 7911 } 7912 return ret; 7913 } 7914 EXPORT_SYMBOL(__cond_resched_rwlock_read); 7915 7916 int __cond_resched_rwlock_write(rwlock_t *lock) 7917 { 7918 int resched = should_resched(PREEMPT_LOCK_OFFSET); 7919 int ret = 0; 7920 7921 lockdep_assert_held_write(lock); 7922 7923 if (rwlock_needbreak(lock) || resched) { 7924 write_unlock(lock); 7925 if (resched) 7926 preempt_schedule_common(); 7927 else 7928 cpu_relax(); 7929 ret = 1; 7930 write_lock(lock); 7931 } 7932 return ret; 7933 } 7934 EXPORT_SYMBOL(__cond_resched_rwlock_write); 7935 7936 /** 7937 * yield - yield the current processor to other threads. 7938 * 7939 * Do not ever use this function, there's a 99% chance you're doing it wrong. 7940 * 7941 * The scheduler is at all times free to pick the calling task as the most 7942 * eligible task to run, if removing the yield() call from your code breaks 7943 * it, it's already broken. 7944 * 7945 * Typical broken usage is: 7946 * 7947 * while (!event) 7948 * yield(); 7949 * 7950 * where one assumes that yield() will let 'the other' process run that will 7951 * make event true. If the current task is a SCHED_FIFO task that will never 7952 * happen. Never use yield() as a progress guarantee!! 7953 * 7954 * If you want to use yield() to wait for something, use wait_event(). 7955 * If you want to use yield() to be 'nice' for others, use cond_resched(). 7956 * If you still want to use yield(), do not! 7957 */ 7958 void __sched yield(void) 7959 { 7960 set_current_state(TASK_RUNNING); 7961 do_sched_yield(); 7962 } 7963 EXPORT_SYMBOL(yield); 7964 7965 /** 7966 * yield_to - yield the current processor to another thread in 7967 * your thread group, or accelerate that thread toward the 7968 * processor it's on. 7969 * @p: target task 7970 * @preempt: whether task preemption is allowed or not 7971 * 7972 * It's the caller's job to ensure that the target task struct 7973 * can't go away on us before we can do any checks. 7974 * 7975 * Return: 7976 * true (>0) if we indeed boosted the target task. 7977 * false (0) if we failed to boost the target. 7978 * -ESRCH if there's no task to yield to. 7979 */ 7980 int __sched yield_to(struct task_struct *p, bool preempt) 7981 { 7982 struct task_struct *curr = current; 7983 struct rq *rq, *p_rq; 7984 unsigned long flags; 7985 int yielded = 0; 7986 7987 local_irq_save(flags); 7988 rq = this_rq(); 7989 7990 again: 7991 p_rq = task_rq(p); 7992 /* 7993 * If we're the only runnable task on the rq and target rq also 7994 * has only one task, there's absolutely no point in yielding. 7995 */ 7996 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 7997 yielded = -ESRCH; 7998 goto out_irq; 7999 } 8000 8001 double_rq_lock(rq, p_rq); 8002 if (task_rq(p) != p_rq) { 8003 double_rq_unlock(rq, p_rq); 8004 goto again; 8005 } 8006 8007 if (!curr->sched_class->yield_to_task) 8008 goto out_unlock; 8009 8010 if (curr->sched_class != p->sched_class) 8011 goto out_unlock; 8012 8013 if (task_running(p_rq, p) || !task_is_running(p)) 8014 goto out_unlock; 8015 8016 yielded = curr->sched_class->yield_to_task(rq, p); 8017 if (yielded) { 8018 schedstat_inc(rq->yld_count); 8019 /* 8020 * Make p's CPU reschedule; pick_next_entity takes care of 8021 * fairness. 8022 */ 8023 if (preempt && rq != p_rq) 8024 resched_curr(p_rq); 8025 } 8026 8027 out_unlock: 8028 double_rq_unlock(rq, p_rq); 8029 out_irq: 8030 local_irq_restore(flags); 8031 8032 if (yielded > 0) 8033 schedule(); 8034 8035 return yielded; 8036 } 8037 EXPORT_SYMBOL_GPL(yield_to); 8038 8039 int io_schedule_prepare(void) 8040 { 8041 int old_iowait = current->in_iowait; 8042 8043 current->in_iowait = 1; 8044 blk_schedule_flush_plug(current); 8045 8046 return old_iowait; 8047 } 8048 8049 void io_schedule_finish(int token) 8050 { 8051 current->in_iowait = token; 8052 } 8053 8054 /* 8055 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 8056 * that process accounting knows that this is a task in IO wait state. 8057 */ 8058 long __sched io_schedule_timeout(long timeout) 8059 { 8060 int token; 8061 long ret; 8062 8063 token = io_schedule_prepare(); 8064 ret = schedule_timeout(timeout); 8065 io_schedule_finish(token); 8066 8067 return ret; 8068 } 8069 EXPORT_SYMBOL(io_schedule_timeout); 8070 8071 void __sched io_schedule(void) 8072 { 8073 int token; 8074 8075 token = io_schedule_prepare(); 8076 schedule(); 8077 io_schedule_finish(token); 8078 } 8079 EXPORT_SYMBOL(io_schedule); 8080 8081 /** 8082 * sys_sched_get_priority_max - return maximum RT priority. 8083 * @policy: scheduling class. 8084 * 8085 * Return: On success, this syscall returns the maximum 8086 * rt_priority that can be used by a given scheduling class. 8087 * On failure, a negative error code is returned. 8088 */ 8089 SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 8090 { 8091 int ret = -EINVAL; 8092 8093 switch (policy) { 8094 case SCHED_FIFO: 8095 case SCHED_RR: 8096 ret = MAX_RT_PRIO-1; 8097 break; 8098 case SCHED_DEADLINE: 8099 case SCHED_NORMAL: 8100 case SCHED_BATCH: 8101 case SCHED_IDLE: 8102 ret = 0; 8103 break; 8104 } 8105 return ret; 8106 } 8107 8108 /** 8109 * sys_sched_get_priority_min - return minimum RT priority. 8110 * @policy: scheduling class. 8111 * 8112 * Return: On success, this syscall returns the minimum 8113 * rt_priority that can be used by a given scheduling class. 8114 * On failure, a negative error code is returned. 8115 */ 8116 SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 8117 { 8118 int ret = -EINVAL; 8119 8120 switch (policy) { 8121 case SCHED_FIFO: 8122 case SCHED_RR: 8123 ret = 1; 8124 break; 8125 case SCHED_DEADLINE: 8126 case SCHED_NORMAL: 8127 case SCHED_BATCH: 8128 case SCHED_IDLE: 8129 ret = 0; 8130 } 8131 return ret; 8132 } 8133 8134 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t) 8135 { 8136 struct task_struct *p; 8137 unsigned int time_slice; 8138 struct rq_flags rf; 8139 struct rq *rq; 8140 int retval; 8141 8142 if (pid < 0) 8143 return -EINVAL; 8144 8145 retval = -ESRCH; 8146 rcu_read_lock(); 8147 p = find_process_by_pid(pid); 8148 if (!p) 8149 goto out_unlock; 8150 8151 retval = security_task_getscheduler(p); 8152 if (retval) 8153 goto out_unlock; 8154 8155 rq = task_rq_lock(p, &rf); 8156 time_slice = 0; 8157 if (p->sched_class->get_rr_interval) 8158 time_slice = p->sched_class->get_rr_interval(rq, p); 8159 task_rq_unlock(rq, p, &rf); 8160 8161 rcu_read_unlock(); 8162 jiffies_to_timespec64(time_slice, t); 8163 return 0; 8164 8165 out_unlock: 8166 rcu_read_unlock(); 8167 return retval; 8168 } 8169 8170 /** 8171 * sys_sched_rr_get_interval - return the default timeslice of a process. 8172 * @pid: pid of the process. 8173 * @interval: userspace pointer to the timeslice value. 8174 * 8175 * this syscall writes the default timeslice value of a given process 8176 * into the user-space timespec buffer. A value of '0' means infinity. 8177 * 8178 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 8179 * an error code. 8180 */ 8181 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 8182 struct __kernel_timespec __user *, interval) 8183 { 8184 struct timespec64 t; 8185 int retval = sched_rr_get_interval(pid, &t); 8186 8187 if (retval == 0) 8188 retval = put_timespec64(&t, interval); 8189 8190 return retval; 8191 } 8192 8193 #ifdef CONFIG_COMPAT_32BIT_TIME 8194 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid, 8195 struct old_timespec32 __user *, interval) 8196 { 8197 struct timespec64 t; 8198 int retval = sched_rr_get_interval(pid, &t); 8199 8200 if (retval == 0) 8201 retval = put_old_timespec32(&t, interval); 8202 return retval; 8203 } 8204 #endif 8205 8206 void sched_show_task(struct task_struct *p) 8207 { 8208 unsigned long free = 0; 8209 int ppid; 8210 8211 if (!try_get_task_stack(p)) 8212 return; 8213 8214 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p)); 8215 8216 if (task_is_running(p)) 8217 pr_cont(" running task "); 8218 #ifdef CONFIG_DEBUG_STACK_USAGE 8219 free = stack_not_used(p); 8220 #endif 8221 ppid = 0; 8222 rcu_read_lock(); 8223 if (pid_alive(p)) 8224 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 8225 rcu_read_unlock(); 8226 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n", 8227 free, task_pid_nr(p), ppid, 8228 (unsigned long)task_thread_info(p)->flags); 8229 8230 print_worker_info(KERN_INFO, p); 8231 print_stop_info(KERN_INFO, p); 8232 show_stack(p, NULL, KERN_INFO); 8233 put_task_stack(p); 8234 } 8235 EXPORT_SYMBOL_GPL(sched_show_task); 8236 8237 static inline bool 8238 state_filter_match(unsigned long state_filter, struct task_struct *p) 8239 { 8240 unsigned int state = READ_ONCE(p->__state); 8241 8242 /* no filter, everything matches */ 8243 if (!state_filter) 8244 return true; 8245 8246 /* filter, but doesn't match */ 8247 if (!(state & state_filter)) 8248 return false; 8249 8250 /* 8251 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows 8252 * TASK_KILLABLE). 8253 */ 8254 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE) 8255 return false; 8256 8257 return true; 8258 } 8259 8260 8261 void show_state_filter(unsigned int state_filter) 8262 { 8263 struct task_struct *g, *p; 8264 8265 rcu_read_lock(); 8266 for_each_process_thread(g, p) { 8267 /* 8268 * reset the NMI-timeout, listing all files on a slow 8269 * console might take a lot of time: 8270 * Also, reset softlockup watchdogs on all CPUs, because 8271 * another CPU might be blocked waiting for us to process 8272 * an IPI. 8273 */ 8274 touch_nmi_watchdog(); 8275 touch_all_softlockup_watchdogs(); 8276 if (state_filter_match(state_filter, p)) 8277 sched_show_task(p); 8278 } 8279 8280 #ifdef CONFIG_SCHED_DEBUG 8281 if (!state_filter) 8282 sysrq_sched_debug_show(); 8283 #endif 8284 rcu_read_unlock(); 8285 /* 8286 * Only show locks if all tasks are dumped: 8287 */ 8288 if (!state_filter) 8289 debug_show_all_locks(); 8290 } 8291 8292 /** 8293 * init_idle - set up an idle thread for a given CPU 8294 * @idle: task in question 8295 * @cpu: CPU the idle task belongs to 8296 * 8297 * NOTE: this function does not set the idle thread's NEED_RESCHED 8298 * flag, to make booting more robust. 8299 */ 8300 void __init init_idle(struct task_struct *idle, int cpu) 8301 { 8302 struct rq *rq = cpu_rq(cpu); 8303 unsigned long flags; 8304 8305 __sched_fork(0, idle); 8306 8307 /* 8308 * The idle task doesn't need the kthread struct to function, but it 8309 * is dressed up as a per-CPU kthread and thus needs to play the part 8310 * if we want to avoid special-casing it in code that deals with per-CPU 8311 * kthreads. 8312 */ 8313 set_kthread_struct(idle); 8314 8315 raw_spin_lock_irqsave(&idle->pi_lock, flags); 8316 raw_spin_rq_lock(rq); 8317 8318 idle->__state = TASK_RUNNING; 8319 idle->se.exec_start = sched_clock(); 8320 /* 8321 * PF_KTHREAD should already be set at this point; regardless, make it 8322 * look like a proper per-CPU kthread. 8323 */ 8324 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY; 8325 kthread_set_per_cpu(idle, cpu); 8326 8327 scs_task_reset(idle); 8328 kasan_unpoison_task_stack(idle); 8329 8330 #ifdef CONFIG_SMP 8331 /* 8332 * It's possible that init_idle() gets called multiple times on a task, 8333 * in that case do_set_cpus_allowed() will not do the right thing. 8334 * 8335 * And since this is boot we can forgo the serialization. 8336 */ 8337 set_cpus_allowed_common(idle, cpumask_of(cpu), 0); 8338 #endif 8339 /* 8340 * We're having a chicken and egg problem, even though we are 8341 * holding rq->lock, the CPU isn't yet set to this CPU so the 8342 * lockdep check in task_group() will fail. 8343 * 8344 * Similar case to sched_fork(). / Alternatively we could 8345 * use task_rq_lock() here and obtain the other rq->lock. 8346 * 8347 * Silence PROVE_RCU 8348 */ 8349 rcu_read_lock(); 8350 __set_task_cpu(idle, cpu); 8351 rcu_read_unlock(); 8352 8353 rq->idle = idle; 8354 rcu_assign_pointer(rq->curr, idle); 8355 idle->on_rq = TASK_ON_RQ_QUEUED; 8356 #ifdef CONFIG_SMP 8357 idle->on_cpu = 1; 8358 #endif 8359 raw_spin_rq_unlock(rq); 8360 raw_spin_unlock_irqrestore(&idle->pi_lock, flags); 8361 8362 /* Set the preempt count _outside_ the spinlocks! */ 8363 init_idle_preempt_count(idle, cpu); 8364 8365 /* 8366 * The idle tasks have their own, simple scheduling class: 8367 */ 8368 idle->sched_class = &idle_sched_class; 8369 ftrace_graph_init_idle_task(idle, cpu); 8370 vtime_init_idle(idle, cpu); 8371 #ifdef CONFIG_SMP 8372 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 8373 #endif 8374 } 8375 8376 #ifdef CONFIG_SMP 8377 8378 int cpuset_cpumask_can_shrink(const struct cpumask *cur, 8379 const struct cpumask *trial) 8380 { 8381 int ret = 1; 8382 8383 if (!cpumask_weight(cur)) 8384 return ret; 8385 8386 ret = dl_cpuset_cpumask_can_shrink(cur, trial); 8387 8388 return ret; 8389 } 8390 8391 int task_can_attach(struct task_struct *p, 8392 const struct cpumask *cs_cpus_allowed) 8393 { 8394 int ret = 0; 8395 8396 /* 8397 * Kthreads which disallow setaffinity shouldn't be moved 8398 * to a new cpuset; we don't want to change their CPU 8399 * affinity and isolating such threads by their set of 8400 * allowed nodes is unnecessary. Thus, cpusets are not 8401 * applicable for such threads. This prevents checking for 8402 * success of set_cpus_allowed_ptr() on all attached tasks 8403 * before cpus_mask may be changed. 8404 */ 8405 if (p->flags & PF_NO_SETAFFINITY) { 8406 ret = -EINVAL; 8407 goto out; 8408 } 8409 8410 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span, 8411 cs_cpus_allowed)) 8412 ret = dl_task_can_attach(p, cs_cpus_allowed); 8413 8414 out: 8415 return ret; 8416 } 8417 8418 bool sched_smp_initialized __read_mostly; 8419 8420 #ifdef CONFIG_NUMA_BALANCING 8421 /* Migrate current task p to target_cpu */ 8422 int migrate_task_to(struct task_struct *p, int target_cpu) 8423 { 8424 struct migration_arg arg = { p, target_cpu }; 8425 int curr_cpu = task_cpu(p); 8426 8427 if (curr_cpu == target_cpu) 8428 return 0; 8429 8430 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr)) 8431 return -EINVAL; 8432 8433 /* TODO: This is not properly updating schedstats */ 8434 8435 trace_sched_move_numa(p, curr_cpu, target_cpu); 8436 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 8437 } 8438 8439 /* 8440 * Requeue a task on a given node and accurately track the number of NUMA 8441 * tasks on the runqueues 8442 */ 8443 void sched_setnuma(struct task_struct *p, int nid) 8444 { 8445 bool queued, running; 8446 struct rq_flags rf; 8447 struct rq *rq; 8448 8449 rq = task_rq_lock(p, &rf); 8450 queued = task_on_rq_queued(p); 8451 running = task_current(rq, p); 8452 8453 if (queued) 8454 dequeue_task(rq, p, DEQUEUE_SAVE); 8455 if (running) 8456 put_prev_task(rq, p); 8457 8458 p->numa_preferred_nid = nid; 8459 8460 if (queued) 8461 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); 8462 if (running) 8463 set_next_task(rq, p); 8464 task_rq_unlock(rq, p, &rf); 8465 } 8466 #endif /* CONFIG_NUMA_BALANCING */ 8467 8468 #ifdef CONFIG_HOTPLUG_CPU 8469 /* 8470 * Ensure that the idle task is using init_mm right before its CPU goes 8471 * offline. 8472 */ 8473 void idle_task_exit(void) 8474 { 8475 struct mm_struct *mm = current->active_mm; 8476 8477 BUG_ON(cpu_online(smp_processor_id())); 8478 BUG_ON(current != this_rq()->idle); 8479 8480 if (mm != &init_mm) { 8481 switch_mm(mm, &init_mm, current); 8482 finish_arch_post_lock_switch(); 8483 } 8484 8485 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */ 8486 } 8487 8488 static int __balance_push_cpu_stop(void *arg) 8489 { 8490 struct task_struct *p = arg; 8491 struct rq *rq = this_rq(); 8492 struct rq_flags rf; 8493 int cpu; 8494 8495 raw_spin_lock_irq(&p->pi_lock); 8496 rq_lock(rq, &rf); 8497 8498 update_rq_clock(rq); 8499 8500 if (task_rq(p) == rq && task_on_rq_queued(p)) { 8501 cpu = select_fallback_rq(rq->cpu, p); 8502 rq = __migrate_task(rq, &rf, p, cpu); 8503 } 8504 8505 rq_unlock(rq, &rf); 8506 raw_spin_unlock_irq(&p->pi_lock); 8507 8508 put_task_struct(p); 8509 8510 return 0; 8511 } 8512 8513 static DEFINE_PER_CPU(struct cpu_stop_work, push_work); 8514 8515 /* 8516 * Ensure we only run per-cpu kthreads once the CPU goes !active. 8517 * 8518 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only 8519 * effective when the hotplug motion is down. 8520 */ 8521 static void balance_push(struct rq *rq) 8522 { 8523 struct task_struct *push_task = rq->curr; 8524 8525 lockdep_assert_rq_held(rq); 8526 SCHED_WARN_ON(rq->cpu != smp_processor_id()); 8527 8528 /* 8529 * Ensure the thing is persistent until balance_push_set(.on = false); 8530 */ 8531 rq->balance_callback = &balance_push_callback; 8532 8533 /* 8534 * Only active while going offline. 8535 */ 8536 if (!cpu_dying(rq->cpu)) 8537 return; 8538 8539 /* 8540 * Both the cpu-hotplug and stop task are in this case and are 8541 * required to complete the hotplug process. 8542 */ 8543 if (kthread_is_per_cpu(push_task) || 8544 is_migration_disabled(push_task)) { 8545 8546 /* 8547 * If this is the idle task on the outgoing CPU try to wake 8548 * up the hotplug control thread which might wait for the 8549 * last task to vanish. The rcuwait_active() check is 8550 * accurate here because the waiter is pinned on this CPU 8551 * and can't obviously be running in parallel. 8552 * 8553 * On RT kernels this also has to check whether there are 8554 * pinned and scheduled out tasks on the runqueue. They 8555 * need to leave the migrate disabled section first. 8556 */ 8557 if (!rq->nr_running && !rq_has_pinned_tasks(rq) && 8558 rcuwait_active(&rq->hotplug_wait)) { 8559 raw_spin_rq_unlock(rq); 8560 rcuwait_wake_up(&rq->hotplug_wait); 8561 raw_spin_rq_lock(rq); 8562 } 8563 return; 8564 } 8565 8566 get_task_struct(push_task); 8567 /* 8568 * Temporarily drop rq->lock such that we can wake-up the stop task. 8569 * Both preemption and IRQs are still disabled. 8570 */ 8571 raw_spin_rq_unlock(rq); 8572 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task, 8573 this_cpu_ptr(&push_work)); 8574 /* 8575 * At this point need_resched() is true and we'll take the loop in 8576 * schedule(). The next pick is obviously going to be the stop task 8577 * which kthread_is_per_cpu() and will push this task away. 8578 */ 8579 raw_spin_rq_lock(rq); 8580 } 8581 8582 static void balance_push_set(int cpu, bool on) 8583 { 8584 struct rq *rq = cpu_rq(cpu); 8585 struct rq_flags rf; 8586 8587 rq_lock_irqsave(rq, &rf); 8588 if (on) { 8589 WARN_ON_ONCE(rq->balance_callback); 8590 rq->balance_callback = &balance_push_callback; 8591 } else if (rq->balance_callback == &balance_push_callback) { 8592 rq->balance_callback = NULL; 8593 } 8594 rq_unlock_irqrestore(rq, &rf); 8595 } 8596 8597 /* 8598 * Invoked from a CPUs hotplug control thread after the CPU has been marked 8599 * inactive. All tasks which are not per CPU kernel threads are either 8600 * pushed off this CPU now via balance_push() or placed on a different CPU 8601 * during wakeup. Wait until the CPU is quiescent. 8602 */ 8603 static void balance_hotplug_wait(void) 8604 { 8605 struct rq *rq = this_rq(); 8606 8607 rcuwait_wait_event(&rq->hotplug_wait, 8608 rq->nr_running == 1 && !rq_has_pinned_tasks(rq), 8609 TASK_UNINTERRUPTIBLE); 8610 } 8611 8612 #else 8613 8614 static inline void balance_push(struct rq *rq) 8615 { 8616 } 8617 8618 static inline void balance_push_set(int cpu, bool on) 8619 { 8620 } 8621 8622 static inline void balance_hotplug_wait(void) 8623 { 8624 } 8625 8626 #endif /* CONFIG_HOTPLUG_CPU */ 8627 8628 void set_rq_online(struct rq *rq) 8629 { 8630 if (!rq->online) { 8631 const struct sched_class *class; 8632 8633 cpumask_set_cpu(rq->cpu, rq->rd->online); 8634 rq->online = 1; 8635 8636 for_each_class(class) { 8637 if (class->rq_online) 8638 class->rq_online(rq); 8639 } 8640 } 8641 } 8642 8643 void set_rq_offline(struct rq *rq) 8644 { 8645 if (rq->online) { 8646 const struct sched_class *class; 8647 8648 for_each_class(class) { 8649 if (class->rq_offline) 8650 class->rq_offline(rq); 8651 } 8652 8653 cpumask_clear_cpu(rq->cpu, rq->rd->online); 8654 rq->online = 0; 8655 } 8656 } 8657 8658 /* 8659 * used to mark begin/end of suspend/resume: 8660 */ 8661 static int num_cpus_frozen; 8662 8663 /* 8664 * Update cpusets according to cpu_active mask. If cpusets are 8665 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 8666 * around partition_sched_domains(). 8667 * 8668 * If we come here as part of a suspend/resume, don't touch cpusets because we 8669 * want to restore it back to its original state upon resume anyway. 8670 */ 8671 static void cpuset_cpu_active(void) 8672 { 8673 if (cpuhp_tasks_frozen) { 8674 /* 8675 * num_cpus_frozen tracks how many CPUs are involved in suspend 8676 * resume sequence. As long as this is not the last online 8677 * operation in the resume sequence, just build a single sched 8678 * domain, ignoring cpusets. 8679 */ 8680 partition_sched_domains(1, NULL, NULL); 8681 if (--num_cpus_frozen) 8682 return; 8683 /* 8684 * This is the last CPU online operation. So fall through and 8685 * restore the original sched domains by considering the 8686 * cpuset configurations. 8687 */ 8688 cpuset_force_rebuild(); 8689 } 8690 cpuset_update_active_cpus(); 8691 } 8692 8693 static int cpuset_cpu_inactive(unsigned int cpu) 8694 { 8695 if (!cpuhp_tasks_frozen) { 8696 if (dl_cpu_busy(cpu)) 8697 return -EBUSY; 8698 cpuset_update_active_cpus(); 8699 } else { 8700 num_cpus_frozen++; 8701 partition_sched_domains(1, NULL, NULL); 8702 } 8703 return 0; 8704 } 8705 8706 int sched_cpu_activate(unsigned int cpu) 8707 { 8708 struct rq *rq = cpu_rq(cpu); 8709 struct rq_flags rf; 8710 8711 /* 8712 * Clear the balance_push callback and prepare to schedule 8713 * regular tasks. 8714 */ 8715 balance_push_set(cpu, false); 8716 8717 #ifdef CONFIG_SCHED_SMT 8718 /* 8719 * When going up, increment the number of cores with SMT present. 8720 */ 8721 if (cpumask_weight(cpu_smt_mask(cpu)) == 2) 8722 static_branch_inc_cpuslocked(&sched_smt_present); 8723 #endif 8724 set_cpu_active(cpu, true); 8725 8726 if (sched_smp_initialized) { 8727 sched_domains_numa_masks_set(cpu); 8728 cpuset_cpu_active(); 8729 } 8730 8731 /* 8732 * Put the rq online, if not already. This happens: 8733 * 8734 * 1) In the early boot process, because we build the real domains 8735 * after all CPUs have been brought up. 8736 * 8737 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the 8738 * domains. 8739 */ 8740 rq_lock_irqsave(rq, &rf); 8741 if (rq->rd) { 8742 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 8743 set_rq_online(rq); 8744 } 8745 rq_unlock_irqrestore(rq, &rf); 8746 8747 return 0; 8748 } 8749 8750 int sched_cpu_deactivate(unsigned int cpu) 8751 { 8752 struct rq *rq = cpu_rq(cpu); 8753 struct rq_flags rf; 8754 int ret; 8755 8756 /* 8757 * Remove CPU from nohz.idle_cpus_mask to prevent participating in 8758 * load balancing when not active 8759 */ 8760 nohz_balance_exit_idle(rq); 8761 8762 set_cpu_active(cpu, false); 8763 8764 /* 8765 * From this point forward, this CPU will refuse to run any task that 8766 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively 8767 * push those tasks away until this gets cleared, see 8768 * sched_cpu_dying(). 8769 */ 8770 balance_push_set(cpu, true); 8771 8772 /* 8773 * We've cleared cpu_active_mask / set balance_push, wait for all 8774 * preempt-disabled and RCU users of this state to go away such that 8775 * all new such users will observe it. 8776 * 8777 * Specifically, we rely on ttwu to no longer target this CPU, see 8778 * ttwu_queue_cond() and is_cpu_allowed(). 8779 * 8780 * Do sync before park smpboot threads to take care the rcu boost case. 8781 */ 8782 synchronize_rcu(); 8783 8784 rq_lock_irqsave(rq, &rf); 8785 if (rq->rd) { 8786 update_rq_clock(rq); 8787 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 8788 set_rq_offline(rq); 8789 } 8790 rq_unlock_irqrestore(rq, &rf); 8791 8792 #ifdef CONFIG_SCHED_SMT 8793 /* 8794 * When going down, decrement the number of cores with SMT present. 8795 */ 8796 if (cpumask_weight(cpu_smt_mask(cpu)) == 2) 8797 static_branch_dec_cpuslocked(&sched_smt_present); 8798 8799 sched_core_cpu_deactivate(cpu); 8800 #endif 8801 8802 if (!sched_smp_initialized) 8803 return 0; 8804 8805 ret = cpuset_cpu_inactive(cpu); 8806 if (ret) { 8807 balance_push_set(cpu, false); 8808 set_cpu_active(cpu, true); 8809 return ret; 8810 } 8811 sched_domains_numa_masks_clear(cpu); 8812 return 0; 8813 } 8814 8815 static void sched_rq_cpu_starting(unsigned int cpu) 8816 { 8817 struct rq *rq = cpu_rq(cpu); 8818 8819 rq->calc_load_update = calc_load_update; 8820 update_max_interval(); 8821 } 8822 8823 int sched_cpu_starting(unsigned int cpu) 8824 { 8825 sched_core_cpu_starting(cpu); 8826 sched_rq_cpu_starting(cpu); 8827 sched_tick_start(cpu); 8828 return 0; 8829 } 8830 8831 #ifdef CONFIG_HOTPLUG_CPU 8832 8833 /* 8834 * Invoked immediately before the stopper thread is invoked to bring the 8835 * CPU down completely. At this point all per CPU kthreads except the 8836 * hotplug thread (current) and the stopper thread (inactive) have been 8837 * either parked or have been unbound from the outgoing CPU. Ensure that 8838 * any of those which might be on the way out are gone. 8839 * 8840 * If after this point a bound task is being woken on this CPU then the 8841 * responsible hotplug callback has failed to do it's job. 8842 * sched_cpu_dying() will catch it with the appropriate fireworks. 8843 */ 8844 int sched_cpu_wait_empty(unsigned int cpu) 8845 { 8846 balance_hotplug_wait(); 8847 return 0; 8848 } 8849 8850 /* 8851 * Since this CPU is going 'away' for a while, fold any nr_active delta we 8852 * might have. Called from the CPU stopper task after ensuring that the 8853 * stopper is the last running task on the CPU, so nr_active count is 8854 * stable. We need to take the teardown thread which is calling this into 8855 * account, so we hand in adjust = 1 to the load calculation. 8856 * 8857 * Also see the comment "Global load-average calculations". 8858 */ 8859 static void calc_load_migrate(struct rq *rq) 8860 { 8861 long delta = calc_load_fold_active(rq, 1); 8862 8863 if (delta) 8864 atomic_long_add(delta, &calc_load_tasks); 8865 } 8866 8867 static void dump_rq_tasks(struct rq *rq, const char *loglvl) 8868 { 8869 struct task_struct *g, *p; 8870 int cpu = cpu_of(rq); 8871 8872 lockdep_assert_rq_held(rq); 8873 8874 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running); 8875 for_each_process_thread(g, p) { 8876 if (task_cpu(p) != cpu) 8877 continue; 8878 8879 if (!task_on_rq_queued(p)) 8880 continue; 8881 8882 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm); 8883 } 8884 } 8885 8886 int sched_cpu_dying(unsigned int cpu) 8887 { 8888 struct rq *rq = cpu_rq(cpu); 8889 struct rq_flags rf; 8890 8891 /* Handle pending wakeups and then migrate everything off */ 8892 sched_tick_stop(cpu); 8893 8894 rq_lock_irqsave(rq, &rf); 8895 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) { 8896 WARN(true, "Dying CPU not properly vacated!"); 8897 dump_rq_tasks(rq, KERN_WARNING); 8898 } 8899 rq_unlock_irqrestore(rq, &rf); 8900 8901 calc_load_migrate(rq); 8902 update_max_interval(); 8903 hrtick_clear(rq); 8904 sched_core_cpu_dying(cpu); 8905 return 0; 8906 } 8907 #endif 8908 8909 void __init sched_init_smp(void) 8910 { 8911 sched_init_numa(); 8912 8913 /* 8914 * There's no userspace yet to cause hotplug operations; hence all the 8915 * CPU masks are stable and all blatant races in the below code cannot 8916 * happen. 8917 */ 8918 mutex_lock(&sched_domains_mutex); 8919 sched_init_domains(cpu_active_mask); 8920 mutex_unlock(&sched_domains_mutex); 8921 8922 /* Move init over to a non-isolated CPU */ 8923 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0) 8924 BUG(); 8925 current->flags &= ~PF_NO_SETAFFINITY; 8926 sched_init_granularity(); 8927 8928 init_sched_rt_class(); 8929 init_sched_dl_class(); 8930 8931 sched_smp_initialized = true; 8932 } 8933 8934 static int __init migration_init(void) 8935 { 8936 sched_cpu_starting(smp_processor_id()); 8937 return 0; 8938 } 8939 early_initcall(migration_init); 8940 8941 #else 8942 void __init sched_init_smp(void) 8943 { 8944 sched_init_granularity(); 8945 } 8946 #endif /* CONFIG_SMP */ 8947 8948 int in_sched_functions(unsigned long addr) 8949 { 8950 return in_lock_functions(addr) || 8951 (addr >= (unsigned long)__sched_text_start 8952 && addr < (unsigned long)__sched_text_end); 8953 } 8954 8955 #ifdef CONFIG_CGROUP_SCHED 8956 /* 8957 * Default task group. 8958 * Every task in system belongs to this group at bootup. 8959 */ 8960 struct task_group root_task_group; 8961 LIST_HEAD(task_groups); 8962 8963 /* Cacheline aligned slab cache for task_group */ 8964 static struct kmem_cache *task_group_cache __read_mostly; 8965 #endif 8966 8967 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 8968 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask); 8969 8970 void __init sched_init(void) 8971 { 8972 unsigned long ptr = 0; 8973 int i; 8974 8975 /* Make sure the linker didn't screw up */ 8976 BUG_ON(&idle_sched_class + 1 != &fair_sched_class || 8977 &fair_sched_class + 1 != &rt_sched_class || 8978 &rt_sched_class + 1 != &dl_sched_class); 8979 #ifdef CONFIG_SMP 8980 BUG_ON(&dl_sched_class + 1 != &stop_sched_class); 8981 #endif 8982 8983 wait_bit_init(); 8984 8985 #ifdef CONFIG_FAIR_GROUP_SCHED 8986 ptr += 2 * nr_cpu_ids * sizeof(void **); 8987 #endif 8988 #ifdef CONFIG_RT_GROUP_SCHED 8989 ptr += 2 * nr_cpu_ids * sizeof(void **); 8990 #endif 8991 if (ptr) { 8992 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT); 8993 8994 #ifdef CONFIG_FAIR_GROUP_SCHED 8995 root_task_group.se = (struct sched_entity **)ptr; 8996 ptr += nr_cpu_ids * sizeof(void **); 8997 8998 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 8999 ptr += nr_cpu_ids * sizeof(void **); 9000 9001 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 9002 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 9003 #endif /* CONFIG_FAIR_GROUP_SCHED */ 9004 #ifdef CONFIG_RT_GROUP_SCHED 9005 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 9006 ptr += nr_cpu_ids * sizeof(void **); 9007 9008 root_task_group.rt_rq = (struct rt_rq **)ptr; 9009 ptr += nr_cpu_ids * sizeof(void **); 9010 9011 #endif /* CONFIG_RT_GROUP_SCHED */ 9012 } 9013 #ifdef CONFIG_CPUMASK_OFFSTACK 9014 for_each_possible_cpu(i) { 9015 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node( 9016 cpumask_size(), GFP_KERNEL, cpu_to_node(i)); 9017 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node( 9018 cpumask_size(), GFP_KERNEL, cpu_to_node(i)); 9019 } 9020 #endif /* CONFIG_CPUMASK_OFFSTACK */ 9021 9022 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime()); 9023 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime()); 9024 9025 #ifdef CONFIG_SMP 9026 init_defrootdomain(); 9027 #endif 9028 9029 #ifdef CONFIG_RT_GROUP_SCHED 9030 init_rt_bandwidth(&root_task_group.rt_bandwidth, 9031 global_rt_period(), global_rt_runtime()); 9032 #endif /* CONFIG_RT_GROUP_SCHED */ 9033 9034 #ifdef CONFIG_CGROUP_SCHED 9035 task_group_cache = KMEM_CACHE(task_group, 0); 9036 9037 list_add(&root_task_group.list, &task_groups); 9038 INIT_LIST_HEAD(&root_task_group.children); 9039 INIT_LIST_HEAD(&root_task_group.siblings); 9040 autogroup_init(&init_task); 9041 #endif /* CONFIG_CGROUP_SCHED */ 9042 9043 for_each_possible_cpu(i) { 9044 struct rq *rq; 9045 9046 rq = cpu_rq(i); 9047 raw_spin_lock_init(&rq->__lock); 9048 rq->nr_running = 0; 9049 rq->calc_load_active = 0; 9050 rq->calc_load_update = jiffies + LOAD_FREQ; 9051 init_cfs_rq(&rq->cfs); 9052 init_rt_rq(&rq->rt); 9053 init_dl_rq(&rq->dl); 9054 #ifdef CONFIG_FAIR_GROUP_SCHED 9055 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 9056 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; 9057 /* 9058 * How much CPU bandwidth does root_task_group get? 9059 * 9060 * In case of task-groups formed thr' the cgroup filesystem, it 9061 * gets 100% of the CPU resources in the system. This overall 9062 * system CPU resource is divided among the tasks of 9063 * root_task_group and its child task-groups in a fair manner, 9064 * based on each entity's (task or task-group's) weight 9065 * (se->load.weight). 9066 * 9067 * In other words, if root_task_group has 10 tasks of weight 9068 * 1024) and two child groups A0 and A1 (of weight 1024 each), 9069 * then A0's share of the CPU resource is: 9070 * 9071 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 9072 * 9073 * We achieve this by letting root_task_group's tasks sit 9074 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 9075 */ 9076 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 9077 #endif /* CONFIG_FAIR_GROUP_SCHED */ 9078 9079 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 9080 #ifdef CONFIG_RT_GROUP_SCHED 9081 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 9082 #endif 9083 #ifdef CONFIG_SMP 9084 rq->sd = NULL; 9085 rq->rd = NULL; 9086 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE; 9087 rq->balance_callback = &balance_push_callback; 9088 rq->active_balance = 0; 9089 rq->next_balance = jiffies; 9090 rq->push_cpu = 0; 9091 rq->cpu = i; 9092 rq->online = 0; 9093 rq->idle_stamp = 0; 9094 rq->avg_idle = 2*sysctl_sched_migration_cost; 9095 rq->wake_stamp = jiffies; 9096 rq->wake_avg_idle = rq->avg_idle; 9097 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 9098 9099 INIT_LIST_HEAD(&rq->cfs_tasks); 9100 9101 rq_attach_root(rq, &def_root_domain); 9102 #ifdef CONFIG_NO_HZ_COMMON 9103 rq->last_blocked_load_update_tick = jiffies; 9104 atomic_set(&rq->nohz_flags, 0); 9105 9106 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq); 9107 #endif 9108 #ifdef CONFIG_HOTPLUG_CPU 9109 rcuwait_init(&rq->hotplug_wait); 9110 #endif 9111 #endif /* CONFIG_SMP */ 9112 hrtick_rq_init(rq); 9113 atomic_set(&rq->nr_iowait, 0); 9114 9115 #ifdef CONFIG_SCHED_CORE 9116 rq->core = rq; 9117 rq->core_pick = NULL; 9118 rq->core_enabled = 0; 9119 rq->core_tree = RB_ROOT; 9120 rq->core_forceidle = false; 9121 9122 rq->core_cookie = 0UL; 9123 #endif 9124 } 9125 9126 set_load_weight(&init_task, false); 9127 9128 /* 9129 * The boot idle thread does lazy MMU switching as well: 9130 */ 9131 mmgrab(&init_mm); 9132 enter_lazy_tlb(&init_mm, current); 9133 9134 /* 9135 * Make us the idle thread. Technically, schedule() should not be 9136 * called from this thread, however somewhere below it might be, 9137 * but because we are the idle thread, we just pick up running again 9138 * when this runqueue becomes "idle". 9139 */ 9140 init_idle(current, smp_processor_id()); 9141 9142 calc_load_update = jiffies + LOAD_FREQ; 9143 9144 #ifdef CONFIG_SMP 9145 idle_thread_set_boot_cpu(); 9146 balance_push_set(smp_processor_id(), false); 9147 #endif 9148 init_sched_fair_class(); 9149 9150 psi_init(); 9151 9152 init_uclamp(); 9153 9154 scheduler_running = 1; 9155 } 9156 9157 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP 9158 static inline int preempt_count_equals(int preempt_offset) 9159 { 9160 int nested = preempt_count() + rcu_preempt_depth(); 9161 9162 return (nested == preempt_offset); 9163 } 9164 9165 void __might_sleep(const char *file, int line, int preempt_offset) 9166 { 9167 unsigned int state = get_current_state(); 9168 /* 9169 * Blocking primitives will set (and therefore destroy) current->state, 9170 * since we will exit with TASK_RUNNING make sure we enter with it, 9171 * otherwise we will destroy state. 9172 */ 9173 WARN_ONCE(state != TASK_RUNNING && current->task_state_change, 9174 "do not call blocking ops when !TASK_RUNNING; " 9175 "state=%x set at [<%p>] %pS\n", state, 9176 (void *)current->task_state_change, 9177 (void *)current->task_state_change); 9178 9179 ___might_sleep(file, line, preempt_offset); 9180 } 9181 EXPORT_SYMBOL(__might_sleep); 9182 9183 void ___might_sleep(const char *file, int line, int preempt_offset) 9184 { 9185 /* Ratelimiting timestamp: */ 9186 static unsigned long prev_jiffy; 9187 9188 unsigned long preempt_disable_ip; 9189 9190 /* WARN_ON_ONCE() by default, no rate limit required: */ 9191 rcu_sleep_check(); 9192 9193 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 9194 !is_idle_task(current) && !current->non_block_count) || 9195 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING || 9196 oops_in_progress) 9197 return; 9198 9199 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 9200 return; 9201 prev_jiffy = jiffies; 9202 9203 /* Save this before calling printk(), since that will clobber it: */ 9204 preempt_disable_ip = get_preempt_disable_ip(current); 9205 9206 printk(KERN_ERR 9207 "BUG: sleeping function called from invalid context at %s:%d\n", 9208 file, line); 9209 printk(KERN_ERR 9210 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n", 9211 in_atomic(), irqs_disabled(), current->non_block_count, 9212 current->pid, current->comm); 9213 9214 if (task_stack_end_corrupted(current)) 9215 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n"); 9216 9217 debug_show_held_locks(current); 9218 if (irqs_disabled()) 9219 print_irqtrace_events(current); 9220 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT) 9221 && !preempt_count_equals(preempt_offset)) { 9222 pr_err("Preemption disabled at:"); 9223 print_ip_sym(KERN_ERR, preempt_disable_ip); 9224 } 9225 dump_stack(); 9226 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 9227 } 9228 EXPORT_SYMBOL(___might_sleep); 9229 9230 void __cant_sleep(const char *file, int line, int preempt_offset) 9231 { 9232 static unsigned long prev_jiffy; 9233 9234 if (irqs_disabled()) 9235 return; 9236 9237 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT)) 9238 return; 9239 9240 if (preempt_count() > preempt_offset) 9241 return; 9242 9243 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 9244 return; 9245 prev_jiffy = jiffies; 9246 9247 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line); 9248 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 9249 in_atomic(), irqs_disabled(), 9250 current->pid, current->comm); 9251 9252 debug_show_held_locks(current); 9253 dump_stack(); 9254 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 9255 } 9256 EXPORT_SYMBOL_GPL(__cant_sleep); 9257 9258 #ifdef CONFIG_SMP 9259 void __cant_migrate(const char *file, int line) 9260 { 9261 static unsigned long prev_jiffy; 9262 9263 if (irqs_disabled()) 9264 return; 9265 9266 if (is_migration_disabled(current)) 9267 return; 9268 9269 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT)) 9270 return; 9271 9272 if (preempt_count() > 0) 9273 return; 9274 9275 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 9276 return; 9277 prev_jiffy = jiffies; 9278 9279 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line); 9280 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n", 9281 in_atomic(), irqs_disabled(), is_migration_disabled(current), 9282 current->pid, current->comm); 9283 9284 debug_show_held_locks(current); 9285 dump_stack(); 9286 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 9287 } 9288 EXPORT_SYMBOL_GPL(__cant_migrate); 9289 #endif 9290 #endif 9291 9292 #ifdef CONFIG_MAGIC_SYSRQ 9293 void normalize_rt_tasks(void) 9294 { 9295 struct task_struct *g, *p; 9296 struct sched_attr attr = { 9297 .sched_policy = SCHED_NORMAL, 9298 }; 9299 9300 read_lock(&tasklist_lock); 9301 for_each_process_thread(g, p) { 9302 /* 9303 * Only normalize user tasks: 9304 */ 9305 if (p->flags & PF_KTHREAD) 9306 continue; 9307 9308 p->se.exec_start = 0; 9309 schedstat_set(p->se.statistics.wait_start, 0); 9310 schedstat_set(p->se.statistics.sleep_start, 0); 9311 schedstat_set(p->se.statistics.block_start, 0); 9312 9313 if (!dl_task(p) && !rt_task(p)) { 9314 /* 9315 * Renice negative nice level userspace 9316 * tasks back to 0: 9317 */ 9318 if (task_nice(p) < 0) 9319 set_user_nice(p, 0); 9320 continue; 9321 } 9322 9323 __sched_setscheduler(p, &attr, false, false); 9324 } 9325 read_unlock(&tasklist_lock); 9326 } 9327 9328 #endif /* CONFIG_MAGIC_SYSRQ */ 9329 9330 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 9331 /* 9332 * These functions are only useful for the IA64 MCA handling, or kdb. 9333 * 9334 * They can only be called when the whole system has been 9335 * stopped - every CPU needs to be quiescent, and no scheduling 9336 * activity can take place. Using them for anything else would 9337 * be a serious bug, and as a result, they aren't even visible 9338 * under any other configuration. 9339 */ 9340 9341 /** 9342 * curr_task - return the current task for a given CPU. 9343 * @cpu: the processor in question. 9344 * 9345 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 9346 * 9347 * Return: The current task for @cpu. 9348 */ 9349 struct task_struct *curr_task(int cpu) 9350 { 9351 return cpu_curr(cpu); 9352 } 9353 9354 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 9355 9356 #ifdef CONFIG_IA64 9357 /** 9358 * ia64_set_curr_task - set the current task for a given CPU. 9359 * @cpu: the processor in question. 9360 * @p: the task pointer to set. 9361 * 9362 * Description: This function must only be used when non-maskable interrupts 9363 * are serviced on a separate stack. It allows the architecture to switch the 9364 * notion of the current task on a CPU in a non-blocking manner. This function 9365 * must be called with all CPU's synchronized, and interrupts disabled, the 9366 * and caller must save the original value of the current task (see 9367 * curr_task() above) and restore that value before reenabling interrupts and 9368 * re-starting the system. 9369 * 9370 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 9371 */ 9372 void ia64_set_curr_task(int cpu, struct task_struct *p) 9373 { 9374 cpu_curr(cpu) = p; 9375 } 9376 9377 #endif 9378 9379 #ifdef CONFIG_CGROUP_SCHED 9380 /* task_group_lock serializes the addition/removal of task groups */ 9381 static DEFINE_SPINLOCK(task_group_lock); 9382 9383 static inline void alloc_uclamp_sched_group(struct task_group *tg, 9384 struct task_group *parent) 9385 { 9386 #ifdef CONFIG_UCLAMP_TASK_GROUP 9387 enum uclamp_id clamp_id; 9388 9389 for_each_clamp_id(clamp_id) { 9390 uclamp_se_set(&tg->uclamp_req[clamp_id], 9391 uclamp_none(clamp_id), false); 9392 tg->uclamp[clamp_id] = parent->uclamp[clamp_id]; 9393 } 9394 #endif 9395 } 9396 9397 static void sched_free_group(struct task_group *tg) 9398 { 9399 free_fair_sched_group(tg); 9400 free_rt_sched_group(tg); 9401 autogroup_free(tg); 9402 kmem_cache_free(task_group_cache, tg); 9403 } 9404 9405 /* allocate runqueue etc for a new task group */ 9406 struct task_group *sched_create_group(struct task_group *parent) 9407 { 9408 struct task_group *tg; 9409 9410 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO); 9411 if (!tg) 9412 return ERR_PTR(-ENOMEM); 9413 9414 if (!alloc_fair_sched_group(tg, parent)) 9415 goto err; 9416 9417 if (!alloc_rt_sched_group(tg, parent)) 9418 goto err; 9419 9420 alloc_uclamp_sched_group(tg, parent); 9421 9422 return tg; 9423 9424 err: 9425 sched_free_group(tg); 9426 return ERR_PTR(-ENOMEM); 9427 } 9428 9429 void sched_online_group(struct task_group *tg, struct task_group *parent) 9430 { 9431 unsigned long flags; 9432 9433 spin_lock_irqsave(&task_group_lock, flags); 9434 list_add_rcu(&tg->list, &task_groups); 9435 9436 /* Root should already exist: */ 9437 WARN_ON(!parent); 9438 9439 tg->parent = parent; 9440 INIT_LIST_HEAD(&tg->children); 9441 list_add_rcu(&tg->siblings, &parent->children); 9442 spin_unlock_irqrestore(&task_group_lock, flags); 9443 9444 online_fair_sched_group(tg); 9445 } 9446 9447 /* rcu callback to free various structures associated with a task group */ 9448 static void sched_free_group_rcu(struct rcu_head *rhp) 9449 { 9450 /* Now it should be safe to free those cfs_rqs: */ 9451 sched_free_group(container_of(rhp, struct task_group, rcu)); 9452 } 9453 9454 void sched_destroy_group(struct task_group *tg) 9455 { 9456 /* Wait for possible concurrent references to cfs_rqs complete: */ 9457 call_rcu(&tg->rcu, sched_free_group_rcu); 9458 } 9459 9460 void sched_offline_group(struct task_group *tg) 9461 { 9462 unsigned long flags; 9463 9464 /* End participation in shares distribution: */ 9465 unregister_fair_sched_group(tg); 9466 9467 spin_lock_irqsave(&task_group_lock, flags); 9468 list_del_rcu(&tg->list); 9469 list_del_rcu(&tg->siblings); 9470 spin_unlock_irqrestore(&task_group_lock, flags); 9471 } 9472 9473 static void sched_change_group(struct task_struct *tsk, int type) 9474 { 9475 struct task_group *tg; 9476 9477 /* 9478 * All callers are synchronized by task_rq_lock(); we do not use RCU 9479 * which is pointless here. Thus, we pass "true" to task_css_check() 9480 * to prevent lockdep warnings. 9481 */ 9482 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), 9483 struct task_group, css); 9484 tg = autogroup_task_group(tsk, tg); 9485 tsk->sched_task_group = tg; 9486 9487 #ifdef CONFIG_FAIR_GROUP_SCHED 9488 if (tsk->sched_class->task_change_group) 9489 tsk->sched_class->task_change_group(tsk, type); 9490 else 9491 #endif 9492 set_task_rq(tsk, task_cpu(tsk)); 9493 } 9494 9495 /* 9496 * Change task's runqueue when it moves between groups. 9497 * 9498 * The caller of this function should have put the task in its new group by 9499 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect 9500 * its new group. 9501 */ 9502 void sched_move_task(struct task_struct *tsk) 9503 { 9504 int queued, running, queue_flags = 9505 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; 9506 struct rq_flags rf; 9507 struct rq *rq; 9508 9509 rq = task_rq_lock(tsk, &rf); 9510 update_rq_clock(rq); 9511 9512 running = task_current(rq, tsk); 9513 queued = task_on_rq_queued(tsk); 9514 9515 if (queued) 9516 dequeue_task(rq, tsk, queue_flags); 9517 if (running) 9518 put_prev_task(rq, tsk); 9519 9520 sched_change_group(tsk, TASK_MOVE_GROUP); 9521 9522 if (queued) 9523 enqueue_task(rq, tsk, queue_flags); 9524 if (running) { 9525 set_next_task(rq, tsk); 9526 /* 9527 * After changing group, the running task may have joined a 9528 * throttled one but it's still the running task. Trigger a 9529 * resched to make sure that task can still run. 9530 */ 9531 resched_curr(rq); 9532 } 9533 9534 task_rq_unlock(rq, tsk, &rf); 9535 } 9536 9537 static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 9538 { 9539 return css ? container_of(css, struct task_group, css) : NULL; 9540 } 9541 9542 static struct cgroup_subsys_state * 9543 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 9544 { 9545 struct task_group *parent = css_tg(parent_css); 9546 struct task_group *tg; 9547 9548 if (!parent) { 9549 /* This is early initialization for the top cgroup */ 9550 return &root_task_group.css; 9551 } 9552 9553 tg = sched_create_group(parent); 9554 if (IS_ERR(tg)) 9555 return ERR_PTR(-ENOMEM); 9556 9557 return &tg->css; 9558 } 9559 9560 /* Expose task group only after completing cgroup initialization */ 9561 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 9562 { 9563 struct task_group *tg = css_tg(css); 9564 struct task_group *parent = css_tg(css->parent); 9565 9566 if (parent) 9567 sched_online_group(tg, parent); 9568 9569 #ifdef CONFIG_UCLAMP_TASK_GROUP 9570 /* Propagate the effective uclamp value for the new group */ 9571 mutex_lock(&uclamp_mutex); 9572 rcu_read_lock(); 9573 cpu_util_update_eff(css); 9574 rcu_read_unlock(); 9575 mutex_unlock(&uclamp_mutex); 9576 #endif 9577 9578 return 0; 9579 } 9580 9581 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css) 9582 { 9583 struct task_group *tg = css_tg(css); 9584 9585 sched_offline_group(tg); 9586 } 9587 9588 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 9589 { 9590 struct task_group *tg = css_tg(css); 9591 9592 /* 9593 * Relies on the RCU grace period between css_released() and this. 9594 */ 9595 sched_free_group(tg); 9596 } 9597 9598 /* 9599 * This is called before wake_up_new_task(), therefore we really only 9600 * have to set its group bits, all the other stuff does not apply. 9601 */ 9602 static void cpu_cgroup_fork(struct task_struct *task) 9603 { 9604 struct rq_flags rf; 9605 struct rq *rq; 9606 9607 rq = task_rq_lock(task, &rf); 9608 9609 update_rq_clock(rq); 9610 sched_change_group(task, TASK_SET_GROUP); 9611 9612 task_rq_unlock(rq, task, &rf); 9613 } 9614 9615 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset) 9616 { 9617 struct task_struct *task; 9618 struct cgroup_subsys_state *css; 9619 int ret = 0; 9620 9621 cgroup_taskset_for_each(task, css, tset) { 9622 #ifdef CONFIG_RT_GROUP_SCHED 9623 if (!sched_rt_can_attach(css_tg(css), task)) 9624 return -EINVAL; 9625 #endif 9626 /* 9627 * Serialize against wake_up_new_task() such that if it's 9628 * running, we're sure to observe its full state. 9629 */ 9630 raw_spin_lock_irq(&task->pi_lock); 9631 /* 9632 * Avoid calling sched_move_task() before wake_up_new_task() 9633 * has happened. This would lead to problems with PELT, due to 9634 * move wanting to detach+attach while we're not attached yet. 9635 */ 9636 if (READ_ONCE(task->__state) == TASK_NEW) 9637 ret = -EINVAL; 9638 raw_spin_unlock_irq(&task->pi_lock); 9639 9640 if (ret) 9641 break; 9642 } 9643 return ret; 9644 } 9645 9646 static void cpu_cgroup_attach(struct cgroup_taskset *tset) 9647 { 9648 struct task_struct *task; 9649 struct cgroup_subsys_state *css; 9650 9651 cgroup_taskset_for_each(task, css, tset) 9652 sched_move_task(task); 9653 } 9654 9655 #ifdef CONFIG_UCLAMP_TASK_GROUP 9656 static void cpu_util_update_eff(struct cgroup_subsys_state *css) 9657 { 9658 struct cgroup_subsys_state *top_css = css; 9659 struct uclamp_se *uc_parent = NULL; 9660 struct uclamp_se *uc_se = NULL; 9661 unsigned int eff[UCLAMP_CNT]; 9662 enum uclamp_id clamp_id; 9663 unsigned int clamps; 9664 9665 lockdep_assert_held(&uclamp_mutex); 9666 SCHED_WARN_ON(!rcu_read_lock_held()); 9667 9668 css_for_each_descendant_pre(css, top_css) { 9669 uc_parent = css_tg(css)->parent 9670 ? css_tg(css)->parent->uclamp : NULL; 9671 9672 for_each_clamp_id(clamp_id) { 9673 /* Assume effective clamps matches requested clamps */ 9674 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value; 9675 /* Cap effective clamps with parent's effective clamps */ 9676 if (uc_parent && 9677 eff[clamp_id] > uc_parent[clamp_id].value) { 9678 eff[clamp_id] = uc_parent[clamp_id].value; 9679 } 9680 } 9681 /* Ensure protection is always capped by limit */ 9682 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]); 9683 9684 /* Propagate most restrictive effective clamps */ 9685 clamps = 0x0; 9686 uc_se = css_tg(css)->uclamp; 9687 for_each_clamp_id(clamp_id) { 9688 if (eff[clamp_id] == uc_se[clamp_id].value) 9689 continue; 9690 uc_se[clamp_id].value = eff[clamp_id]; 9691 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]); 9692 clamps |= (0x1 << clamp_id); 9693 } 9694 if (!clamps) { 9695 css = css_rightmost_descendant(css); 9696 continue; 9697 } 9698 9699 /* Immediately update descendants RUNNABLE tasks */ 9700 uclamp_update_active_tasks(css); 9701 } 9702 } 9703 9704 /* 9705 * Integer 10^N with a given N exponent by casting to integer the literal "1eN" 9706 * C expression. Since there is no way to convert a macro argument (N) into a 9707 * character constant, use two levels of macros. 9708 */ 9709 #define _POW10(exp) ((unsigned int)1e##exp) 9710 #define POW10(exp) _POW10(exp) 9711 9712 struct uclamp_request { 9713 #define UCLAMP_PERCENT_SHIFT 2 9714 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT)) 9715 s64 percent; 9716 u64 util; 9717 int ret; 9718 }; 9719 9720 static inline struct uclamp_request 9721 capacity_from_percent(char *buf) 9722 { 9723 struct uclamp_request req = { 9724 .percent = UCLAMP_PERCENT_SCALE, 9725 .util = SCHED_CAPACITY_SCALE, 9726 .ret = 0, 9727 }; 9728 9729 buf = strim(buf); 9730 if (strcmp(buf, "max")) { 9731 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT, 9732 &req.percent); 9733 if (req.ret) 9734 return req; 9735 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) { 9736 req.ret = -ERANGE; 9737 return req; 9738 } 9739 9740 req.util = req.percent << SCHED_CAPACITY_SHIFT; 9741 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE); 9742 } 9743 9744 return req; 9745 } 9746 9747 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf, 9748 size_t nbytes, loff_t off, 9749 enum uclamp_id clamp_id) 9750 { 9751 struct uclamp_request req; 9752 struct task_group *tg; 9753 9754 req = capacity_from_percent(buf); 9755 if (req.ret) 9756 return req.ret; 9757 9758 static_branch_enable(&sched_uclamp_used); 9759 9760 mutex_lock(&uclamp_mutex); 9761 rcu_read_lock(); 9762 9763 tg = css_tg(of_css(of)); 9764 if (tg->uclamp_req[clamp_id].value != req.util) 9765 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false); 9766 9767 /* 9768 * Because of not recoverable conversion rounding we keep track of the 9769 * exact requested value 9770 */ 9771 tg->uclamp_pct[clamp_id] = req.percent; 9772 9773 /* Update effective clamps to track the most restrictive value */ 9774 cpu_util_update_eff(of_css(of)); 9775 9776 rcu_read_unlock(); 9777 mutex_unlock(&uclamp_mutex); 9778 9779 return nbytes; 9780 } 9781 9782 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of, 9783 char *buf, size_t nbytes, 9784 loff_t off) 9785 { 9786 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN); 9787 } 9788 9789 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of, 9790 char *buf, size_t nbytes, 9791 loff_t off) 9792 { 9793 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX); 9794 } 9795 9796 static inline void cpu_uclamp_print(struct seq_file *sf, 9797 enum uclamp_id clamp_id) 9798 { 9799 struct task_group *tg; 9800 u64 util_clamp; 9801 u64 percent; 9802 u32 rem; 9803 9804 rcu_read_lock(); 9805 tg = css_tg(seq_css(sf)); 9806 util_clamp = tg->uclamp_req[clamp_id].value; 9807 rcu_read_unlock(); 9808 9809 if (util_clamp == SCHED_CAPACITY_SCALE) { 9810 seq_puts(sf, "max\n"); 9811 return; 9812 } 9813 9814 percent = tg->uclamp_pct[clamp_id]; 9815 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem); 9816 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem); 9817 } 9818 9819 static int cpu_uclamp_min_show(struct seq_file *sf, void *v) 9820 { 9821 cpu_uclamp_print(sf, UCLAMP_MIN); 9822 return 0; 9823 } 9824 9825 static int cpu_uclamp_max_show(struct seq_file *sf, void *v) 9826 { 9827 cpu_uclamp_print(sf, UCLAMP_MAX); 9828 return 0; 9829 } 9830 #endif /* CONFIG_UCLAMP_TASK_GROUP */ 9831 9832 #ifdef CONFIG_FAIR_GROUP_SCHED 9833 static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 9834 struct cftype *cftype, u64 shareval) 9835 { 9836 if (shareval > scale_load_down(ULONG_MAX)) 9837 shareval = MAX_SHARES; 9838 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 9839 } 9840 9841 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 9842 struct cftype *cft) 9843 { 9844 struct task_group *tg = css_tg(css); 9845 9846 return (u64) scale_load_down(tg->shares); 9847 } 9848 9849 #ifdef CONFIG_CFS_BANDWIDTH 9850 static DEFINE_MUTEX(cfs_constraints_mutex); 9851 9852 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 9853 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 9854 /* More than 203 days if BW_SHIFT equals 20. */ 9855 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC; 9856 9857 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 9858 9859 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota, 9860 u64 burst) 9861 { 9862 int i, ret = 0, runtime_enabled, runtime_was_enabled; 9863 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 9864 9865 if (tg == &root_task_group) 9866 return -EINVAL; 9867 9868 /* 9869 * Ensure we have at some amount of bandwidth every period. This is 9870 * to prevent reaching a state of large arrears when throttled via 9871 * entity_tick() resulting in prolonged exit starvation. 9872 */ 9873 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 9874 return -EINVAL; 9875 9876 /* 9877 * Likewise, bound things on the other side by preventing insane quota 9878 * periods. This also allows us to normalize in computing quota 9879 * feasibility. 9880 */ 9881 if (period > max_cfs_quota_period) 9882 return -EINVAL; 9883 9884 /* 9885 * Bound quota to defend quota against overflow during bandwidth shift. 9886 */ 9887 if (quota != RUNTIME_INF && quota > max_cfs_runtime) 9888 return -EINVAL; 9889 9890 if (quota != RUNTIME_INF && (burst > quota || 9891 burst + quota > max_cfs_runtime)) 9892 return -EINVAL; 9893 9894 /* 9895 * Prevent race between setting of cfs_rq->runtime_enabled and 9896 * unthrottle_offline_cfs_rqs(). 9897 */ 9898 get_online_cpus(); 9899 mutex_lock(&cfs_constraints_mutex); 9900 ret = __cfs_schedulable(tg, period, quota); 9901 if (ret) 9902 goto out_unlock; 9903 9904 runtime_enabled = quota != RUNTIME_INF; 9905 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 9906 /* 9907 * If we need to toggle cfs_bandwidth_used, off->on must occur 9908 * before making related changes, and on->off must occur afterwards 9909 */ 9910 if (runtime_enabled && !runtime_was_enabled) 9911 cfs_bandwidth_usage_inc(); 9912 raw_spin_lock_irq(&cfs_b->lock); 9913 cfs_b->period = ns_to_ktime(period); 9914 cfs_b->quota = quota; 9915 cfs_b->burst = burst; 9916 9917 __refill_cfs_bandwidth_runtime(cfs_b); 9918 9919 /* Restart the period timer (if active) to handle new period expiry: */ 9920 if (runtime_enabled) 9921 start_cfs_bandwidth(cfs_b); 9922 9923 raw_spin_unlock_irq(&cfs_b->lock); 9924 9925 for_each_online_cpu(i) { 9926 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 9927 struct rq *rq = cfs_rq->rq; 9928 struct rq_flags rf; 9929 9930 rq_lock_irq(rq, &rf); 9931 cfs_rq->runtime_enabled = runtime_enabled; 9932 cfs_rq->runtime_remaining = 0; 9933 9934 if (cfs_rq->throttled) 9935 unthrottle_cfs_rq(cfs_rq); 9936 rq_unlock_irq(rq, &rf); 9937 } 9938 if (runtime_was_enabled && !runtime_enabled) 9939 cfs_bandwidth_usage_dec(); 9940 out_unlock: 9941 mutex_unlock(&cfs_constraints_mutex); 9942 put_online_cpus(); 9943 9944 return ret; 9945 } 9946 9947 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 9948 { 9949 u64 quota, period, burst; 9950 9951 period = ktime_to_ns(tg->cfs_bandwidth.period); 9952 burst = tg->cfs_bandwidth.burst; 9953 if (cfs_quota_us < 0) 9954 quota = RUNTIME_INF; 9955 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC) 9956 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 9957 else 9958 return -EINVAL; 9959 9960 return tg_set_cfs_bandwidth(tg, period, quota, burst); 9961 } 9962 9963 static long tg_get_cfs_quota(struct task_group *tg) 9964 { 9965 u64 quota_us; 9966 9967 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 9968 return -1; 9969 9970 quota_us = tg->cfs_bandwidth.quota; 9971 do_div(quota_us, NSEC_PER_USEC); 9972 9973 return quota_us; 9974 } 9975 9976 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 9977 { 9978 u64 quota, period, burst; 9979 9980 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC) 9981 return -EINVAL; 9982 9983 period = (u64)cfs_period_us * NSEC_PER_USEC; 9984 quota = tg->cfs_bandwidth.quota; 9985 burst = tg->cfs_bandwidth.burst; 9986 9987 return tg_set_cfs_bandwidth(tg, period, quota, burst); 9988 } 9989 9990 static long tg_get_cfs_period(struct task_group *tg) 9991 { 9992 u64 cfs_period_us; 9993 9994 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 9995 do_div(cfs_period_us, NSEC_PER_USEC); 9996 9997 return cfs_period_us; 9998 } 9999 10000 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us) 10001 { 10002 u64 quota, period, burst; 10003 10004 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC) 10005 return -EINVAL; 10006 10007 burst = (u64)cfs_burst_us * NSEC_PER_USEC; 10008 period = ktime_to_ns(tg->cfs_bandwidth.period); 10009 quota = tg->cfs_bandwidth.quota; 10010 10011 return tg_set_cfs_bandwidth(tg, period, quota, burst); 10012 } 10013 10014 static long tg_get_cfs_burst(struct task_group *tg) 10015 { 10016 u64 burst_us; 10017 10018 burst_us = tg->cfs_bandwidth.burst; 10019 do_div(burst_us, NSEC_PER_USEC); 10020 10021 return burst_us; 10022 } 10023 10024 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 10025 struct cftype *cft) 10026 { 10027 return tg_get_cfs_quota(css_tg(css)); 10028 } 10029 10030 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 10031 struct cftype *cftype, s64 cfs_quota_us) 10032 { 10033 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 10034 } 10035 10036 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 10037 struct cftype *cft) 10038 { 10039 return tg_get_cfs_period(css_tg(css)); 10040 } 10041 10042 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 10043 struct cftype *cftype, u64 cfs_period_us) 10044 { 10045 return tg_set_cfs_period(css_tg(css), cfs_period_us); 10046 } 10047 10048 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css, 10049 struct cftype *cft) 10050 { 10051 return tg_get_cfs_burst(css_tg(css)); 10052 } 10053 10054 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css, 10055 struct cftype *cftype, u64 cfs_burst_us) 10056 { 10057 return tg_set_cfs_burst(css_tg(css), cfs_burst_us); 10058 } 10059 10060 struct cfs_schedulable_data { 10061 struct task_group *tg; 10062 u64 period, quota; 10063 }; 10064 10065 /* 10066 * normalize group quota/period to be quota/max_period 10067 * note: units are usecs 10068 */ 10069 static u64 normalize_cfs_quota(struct task_group *tg, 10070 struct cfs_schedulable_data *d) 10071 { 10072 u64 quota, period; 10073 10074 if (tg == d->tg) { 10075 period = d->period; 10076 quota = d->quota; 10077 } else { 10078 period = tg_get_cfs_period(tg); 10079 quota = tg_get_cfs_quota(tg); 10080 } 10081 10082 /* note: these should typically be equivalent */ 10083 if (quota == RUNTIME_INF || quota == -1) 10084 return RUNTIME_INF; 10085 10086 return to_ratio(period, quota); 10087 } 10088 10089 static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 10090 { 10091 struct cfs_schedulable_data *d = data; 10092 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 10093 s64 quota = 0, parent_quota = -1; 10094 10095 if (!tg->parent) { 10096 quota = RUNTIME_INF; 10097 } else { 10098 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 10099 10100 quota = normalize_cfs_quota(tg, d); 10101 parent_quota = parent_b->hierarchical_quota; 10102 10103 /* 10104 * Ensure max(child_quota) <= parent_quota. On cgroup2, 10105 * always take the min. On cgroup1, only inherit when no 10106 * limit is set: 10107 */ 10108 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) { 10109 quota = min(quota, parent_quota); 10110 } else { 10111 if (quota == RUNTIME_INF) 10112 quota = parent_quota; 10113 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 10114 return -EINVAL; 10115 } 10116 } 10117 cfs_b->hierarchical_quota = quota; 10118 10119 return 0; 10120 } 10121 10122 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 10123 { 10124 int ret; 10125 struct cfs_schedulable_data data = { 10126 .tg = tg, 10127 .period = period, 10128 .quota = quota, 10129 }; 10130 10131 if (quota != RUNTIME_INF) { 10132 do_div(data.period, NSEC_PER_USEC); 10133 do_div(data.quota, NSEC_PER_USEC); 10134 } 10135 10136 rcu_read_lock(); 10137 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 10138 rcu_read_unlock(); 10139 10140 return ret; 10141 } 10142 10143 static int cpu_cfs_stat_show(struct seq_file *sf, void *v) 10144 { 10145 struct task_group *tg = css_tg(seq_css(sf)); 10146 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 10147 10148 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 10149 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 10150 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 10151 10152 if (schedstat_enabled() && tg != &root_task_group) { 10153 u64 ws = 0; 10154 int i; 10155 10156 for_each_possible_cpu(i) 10157 ws += schedstat_val(tg->se[i]->statistics.wait_sum); 10158 10159 seq_printf(sf, "wait_sum %llu\n", ws); 10160 } 10161 10162 return 0; 10163 } 10164 #endif /* CONFIG_CFS_BANDWIDTH */ 10165 #endif /* CONFIG_FAIR_GROUP_SCHED */ 10166 10167 #ifdef CONFIG_RT_GROUP_SCHED 10168 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 10169 struct cftype *cft, s64 val) 10170 { 10171 return sched_group_set_rt_runtime(css_tg(css), val); 10172 } 10173 10174 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 10175 struct cftype *cft) 10176 { 10177 return sched_group_rt_runtime(css_tg(css)); 10178 } 10179 10180 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 10181 struct cftype *cftype, u64 rt_period_us) 10182 { 10183 return sched_group_set_rt_period(css_tg(css), rt_period_us); 10184 } 10185 10186 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 10187 struct cftype *cft) 10188 { 10189 return sched_group_rt_period(css_tg(css)); 10190 } 10191 #endif /* CONFIG_RT_GROUP_SCHED */ 10192 10193 static struct cftype cpu_legacy_files[] = { 10194 #ifdef CONFIG_FAIR_GROUP_SCHED 10195 { 10196 .name = "shares", 10197 .read_u64 = cpu_shares_read_u64, 10198 .write_u64 = cpu_shares_write_u64, 10199 }, 10200 #endif 10201 #ifdef CONFIG_CFS_BANDWIDTH 10202 { 10203 .name = "cfs_quota_us", 10204 .read_s64 = cpu_cfs_quota_read_s64, 10205 .write_s64 = cpu_cfs_quota_write_s64, 10206 }, 10207 { 10208 .name = "cfs_period_us", 10209 .read_u64 = cpu_cfs_period_read_u64, 10210 .write_u64 = cpu_cfs_period_write_u64, 10211 }, 10212 { 10213 .name = "cfs_burst_us", 10214 .read_u64 = cpu_cfs_burst_read_u64, 10215 .write_u64 = cpu_cfs_burst_write_u64, 10216 }, 10217 { 10218 .name = "stat", 10219 .seq_show = cpu_cfs_stat_show, 10220 }, 10221 #endif 10222 #ifdef CONFIG_RT_GROUP_SCHED 10223 { 10224 .name = "rt_runtime_us", 10225 .read_s64 = cpu_rt_runtime_read, 10226 .write_s64 = cpu_rt_runtime_write, 10227 }, 10228 { 10229 .name = "rt_period_us", 10230 .read_u64 = cpu_rt_period_read_uint, 10231 .write_u64 = cpu_rt_period_write_uint, 10232 }, 10233 #endif 10234 #ifdef CONFIG_UCLAMP_TASK_GROUP 10235 { 10236 .name = "uclamp.min", 10237 .flags = CFTYPE_NOT_ON_ROOT, 10238 .seq_show = cpu_uclamp_min_show, 10239 .write = cpu_uclamp_min_write, 10240 }, 10241 { 10242 .name = "uclamp.max", 10243 .flags = CFTYPE_NOT_ON_ROOT, 10244 .seq_show = cpu_uclamp_max_show, 10245 .write = cpu_uclamp_max_write, 10246 }, 10247 #endif 10248 { } /* Terminate */ 10249 }; 10250 10251 static int cpu_extra_stat_show(struct seq_file *sf, 10252 struct cgroup_subsys_state *css) 10253 { 10254 #ifdef CONFIG_CFS_BANDWIDTH 10255 { 10256 struct task_group *tg = css_tg(css); 10257 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 10258 u64 throttled_usec; 10259 10260 throttled_usec = cfs_b->throttled_time; 10261 do_div(throttled_usec, NSEC_PER_USEC); 10262 10263 seq_printf(sf, "nr_periods %d\n" 10264 "nr_throttled %d\n" 10265 "throttled_usec %llu\n", 10266 cfs_b->nr_periods, cfs_b->nr_throttled, 10267 throttled_usec); 10268 } 10269 #endif 10270 return 0; 10271 } 10272 10273 #ifdef CONFIG_FAIR_GROUP_SCHED 10274 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css, 10275 struct cftype *cft) 10276 { 10277 struct task_group *tg = css_tg(css); 10278 u64 weight = scale_load_down(tg->shares); 10279 10280 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024); 10281 } 10282 10283 static int cpu_weight_write_u64(struct cgroup_subsys_state *css, 10284 struct cftype *cft, u64 weight) 10285 { 10286 /* 10287 * cgroup weight knobs should use the common MIN, DFL and MAX 10288 * values which are 1, 100 and 10000 respectively. While it loses 10289 * a bit of range on both ends, it maps pretty well onto the shares 10290 * value used by scheduler and the round-trip conversions preserve 10291 * the original value over the entire range. 10292 */ 10293 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX) 10294 return -ERANGE; 10295 10296 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL); 10297 10298 return sched_group_set_shares(css_tg(css), scale_load(weight)); 10299 } 10300 10301 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css, 10302 struct cftype *cft) 10303 { 10304 unsigned long weight = scale_load_down(css_tg(css)->shares); 10305 int last_delta = INT_MAX; 10306 int prio, delta; 10307 10308 /* find the closest nice value to the current weight */ 10309 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) { 10310 delta = abs(sched_prio_to_weight[prio] - weight); 10311 if (delta >= last_delta) 10312 break; 10313 last_delta = delta; 10314 } 10315 10316 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO); 10317 } 10318 10319 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css, 10320 struct cftype *cft, s64 nice) 10321 { 10322 unsigned long weight; 10323 int idx; 10324 10325 if (nice < MIN_NICE || nice > MAX_NICE) 10326 return -ERANGE; 10327 10328 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO; 10329 idx = array_index_nospec(idx, 40); 10330 weight = sched_prio_to_weight[idx]; 10331 10332 return sched_group_set_shares(css_tg(css), scale_load(weight)); 10333 } 10334 #endif 10335 10336 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf, 10337 long period, long quota) 10338 { 10339 if (quota < 0) 10340 seq_puts(sf, "max"); 10341 else 10342 seq_printf(sf, "%ld", quota); 10343 10344 seq_printf(sf, " %ld\n", period); 10345 } 10346 10347 /* caller should put the current value in *@periodp before calling */ 10348 static int __maybe_unused cpu_period_quota_parse(char *buf, 10349 u64 *periodp, u64 *quotap) 10350 { 10351 char tok[21]; /* U64_MAX */ 10352 10353 if (sscanf(buf, "%20s %llu", tok, periodp) < 1) 10354 return -EINVAL; 10355 10356 *periodp *= NSEC_PER_USEC; 10357 10358 if (sscanf(tok, "%llu", quotap)) 10359 *quotap *= NSEC_PER_USEC; 10360 else if (!strcmp(tok, "max")) 10361 *quotap = RUNTIME_INF; 10362 else 10363 return -EINVAL; 10364 10365 return 0; 10366 } 10367 10368 #ifdef CONFIG_CFS_BANDWIDTH 10369 static int cpu_max_show(struct seq_file *sf, void *v) 10370 { 10371 struct task_group *tg = css_tg(seq_css(sf)); 10372 10373 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg)); 10374 return 0; 10375 } 10376 10377 static ssize_t cpu_max_write(struct kernfs_open_file *of, 10378 char *buf, size_t nbytes, loff_t off) 10379 { 10380 struct task_group *tg = css_tg(of_css(of)); 10381 u64 period = tg_get_cfs_period(tg); 10382 u64 burst = tg_get_cfs_burst(tg); 10383 u64 quota; 10384 int ret; 10385 10386 ret = cpu_period_quota_parse(buf, &period, "a); 10387 if (!ret) 10388 ret = tg_set_cfs_bandwidth(tg, period, quota, burst); 10389 return ret ?: nbytes; 10390 } 10391 #endif 10392 10393 static struct cftype cpu_files[] = { 10394 #ifdef CONFIG_FAIR_GROUP_SCHED 10395 { 10396 .name = "weight", 10397 .flags = CFTYPE_NOT_ON_ROOT, 10398 .read_u64 = cpu_weight_read_u64, 10399 .write_u64 = cpu_weight_write_u64, 10400 }, 10401 { 10402 .name = "weight.nice", 10403 .flags = CFTYPE_NOT_ON_ROOT, 10404 .read_s64 = cpu_weight_nice_read_s64, 10405 .write_s64 = cpu_weight_nice_write_s64, 10406 }, 10407 #endif 10408 #ifdef CONFIG_CFS_BANDWIDTH 10409 { 10410 .name = "max", 10411 .flags = CFTYPE_NOT_ON_ROOT, 10412 .seq_show = cpu_max_show, 10413 .write = cpu_max_write, 10414 }, 10415 { 10416 .name = "max.burst", 10417 .flags = CFTYPE_NOT_ON_ROOT, 10418 .read_u64 = cpu_cfs_burst_read_u64, 10419 .write_u64 = cpu_cfs_burst_write_u64, 10420 }, 10421 #endif 10422 #ifdef CONFIG_UCLAMP_TASK_GROUP 10423 { 10424 .name = "uclamp.min", 10425 .flags = CFTYPE_NOT_ON_ROOT, 10426 .seq_show = cpu_uclamp_min_show, 10427 .write = cpu_uclamp_min_write, 10428 }, 10429 { 10430 .name = "uclamp.max", 10431 .flags = CFTYPE_NOT_ON_ROOT, 10432 .seq_show = cpu_uclamp_max_show, 10433 .write = cpu_uclamp_max_write, 10434 }, 10435 #endif 10436 { } /* terminate */ 10437 }; 10438 10439 struct cgroup_subsys cpu_cgrp_subsys = { 10440 .css_alloc = cpu_cgroup_css_alloc, 10441 .css_online = cpu_cgroup_css_online, 10442 .css_released = cpu_cgroup_css_released, 10443 .css_free = cpu_cgroup_css_free, 10444 .css_extra_stat_show = cpu_extra_stat_show, 10445 .fork = cpu_cgroup_fork, 10446 .can_attach = cpu_cgroup_can_attach, 10447 .attach = cpu_cgroup_attach, 10448 .legacy_cftypes = cpu_legacy_files, 10449 .dfl_cftypes = cpu_files, 10450 .early_init = true, 10451 .threaded = true, 10452 }; 10453 10454 #endif /* CONFIG_CGROUP_SCHED */ 10455 10456 void dump_cpu_task(int cpu) 10457 { 10458 pr_info("Task dump for CPU %d:\n", cpu); 10459 sched_show_task(cpu_curr(cpu)); 10460 } 10461 10462 /* 10463 * Nice levels are multiplicative, with a gentle 10% change for every 10464 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to 10465 * nice 1, it will get ~10% less CPU time than another CPU-bound task 10466 * that remained on nice 0. 10467 * 10468 * The "10% effect" is relative and cumulative: from _any_ nice level, 10469 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level 10470 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. 10471 * If a task goes up by ~10% and another task goes down by ~10% then 10472 * the relative distance between them is ~25%.) 10473 */ 10474 const int sched_prio_to_weight[40] = { 10475 /* -20 */ 88761, 71755, 56483, 46273, 36291, 10476 /* -15 */ 29154, 23254, 18705, 14949, 11916, 10477 /* -10 */ 9548, 7620, 6100, 4904, 3906, 10478 /* -5 */ 3121, 2501, 1991, 1586, 1277, 10479 /* 0 */ 1024, 820, 655, 526, 423, 10480 /* 5 */ 335, 272, 215, 172, 137, 10481 /* 10 */ 110, 87, 70, 56, 45, 10482 /* 15 */ 36, 29, 23, 18, 15, 10483 }; 10484 10485 /* 10486 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated. 10487 * 10488 * In cases where the weight does not change often, we can use the 10489 * precalculated inverse to speed up arithmetics by turning divisions 10490 * into multiplications: 10491 */ 10492 const u32 sched_prio_to_wmult[40] = { 10493 /* -20 */ 48388, 59856, 76040, 92818, 118348, 10494 /* -15 */ 147320, 184698, 229616, 287308, 360437, 10495 /* -10 */ 449829, 563644, 704093, 875809, 1099582, 10496 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, 10497 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, 10498 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, 10499 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, 10500 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, 10501 }; 10502 10503 void call_trace_sched_update_nr_running(struct rq *rq, int count) 10504 { 10505 trace_sched_update_nr_running_tp(rq, count); 10506 } 10507