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