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 #include <linux/highmem.h> 10 #include <linux/hrtimer_api.h> 11 #include <linux/ktime_api.h> 12 #include <linux/sched/signal.h> 13 #include <linux/syscalls_api.h> 14 #include <linux/debug_locks.h> 15 #include <linux/prefetch.h> 16 #include <linux/capability.h> 17 #include <linux/pgtable_api.h> 18 #include <linux/wait_bit.h> 19 #include <linux/jiffies.h> 20 #include <linux/spinlock_api.h> 21 #include <linux/cpumask_api.h> 22 #include <linux/lockdep_api.h> 23 #include <linux/hardirq.h> 24 #include <linux/softirq.h> 25 #include <linux/refcount_api.h> 26 #include <linux/topology.h> 27 #include <linux/sched/clock.h> 28 #include <linux/sched/cond_resched.h> 29 #include <linux/sched/cputime.h> 30 #include <linux/sched/debug.h> 31 #include <linux/sched/hotplug.h> 32 #include <linux/sched/init.h> 33 #include <linux/sched/isolation.h> 34 #include <linux/sched/loadavg.h> 35 #include <linux/sched/mm.h> 36 #include <linux/sched/nohz.h> 37 #include <linux/sched/rseq_api.h> 38 #include <linux/sched/rt.h> 39 40 #include <linux/blkdev.h> 41 #include <linux/context_tracking.h> 42 #include <linux/cpuset.h> 43 #include <linux/delayacct.h> 44 #include <linux/init_task.h> 45 #include <linux/interrupt.h> 46 #include <linux/ioprio.h> 47 #include <linux/kallsyms.h> 48 #include <linux/kcov.h> 49 #include <linux/kprobes.h> 50 #include <linux/llist_api.h> 51 #include <linux/mmu_context.h> 52 #include <linux/mmzone.h> 53 #include <linux/mutex_api.h> 54 #include <linux/nmi.h> 55 #include <linux/nospec.h> 56 #include <linux/perf_event_api.h> 57 #include <linux/profile.h> 58 #include <linux/psi.h> 59 #include <linux/rcuwait_api.h> 60 #include <linux/sched/wake_q.h> 61 #include <linux/scs.h> 62 #include <linux/slab.h> 63 #include <linux/syscalls.h> 64 #include <linux/vtime.h> 65 #include <linux/wait_api.h> 66 #include <linux/workqueue_api.h> 67 68 #ifdef CONFIG_PREEMPT_DYNAMIC 69 # ifdef CONFIG_GENERIC_ENTRY 70 # include <linux/entry-common.h> 71 # endif 72 #endif 73 74 #include <uapi/linux/sched/types.h> 75 76 #include <asm/switch_to.h> 77 #include <asm/tlb.h> 78 79 #define CREATE_TRACE_POINTS 80 #include <linux/sched/rseq_api.h> 81 #include <trace/events/sched.h> 82 #undef CREATE_TRACE_POINTS 83 84 #include "sched.h" 85 #include "stats.h" 86 #include "autogroup.h" 87 88 #include "autogroup.h" 89 #include "pelt.h" 90 #include "smp.h" 91 #include "stats.h" 92 93 #include "../workqueue_internal.h" 94 #include "../../io_uring/io-wq.h" 95 #include "../smpboot.h" 96 97 /* 98 * Export tracepoints that act as a bare tracehook (ie: have no trace event 99 * associated with them) to allow external modules to probe them. 100 */ 101 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp); 102 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp); 103 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp); 104 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp); 105 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp); 106 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp); 107 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp); 108 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp); 109 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp); 110 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp); 111 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp); 112 113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115 #ifdef CONFIG_SCHED_DEBUG 116 /* 117 * Debugging: various feature bits 118 * 119 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of 120 * sysctl_sched_features, defined in sched.h, to allow constants propagation 121 * at compile time and compiler optimization based on features default. 122 */ 123 #define SCHED_FEAT(name, enabled) \ 124 (1UL << __SCHED_FEAT_##name) * enabled | 125 const_debug unsigned int sysctl_sched_features = 126 #include "features.h" 127 0; 128 #undef SCHED_FEAT 129 130 /* 131 * Print a warning if need_resched is set for the given duration (if 132 * LATENCY_WARN is enabled). 133 * 134 * If sysctl_resched_latency_warn_once is set, only one warning will be shown 135 * per boot. 136 */ 137 __read_mostly int sysctl_resched_latency_warn_ms = 100; 138 __read_mostly int sysctl_resched_latency_warn_once = 1; 139 #endif /* CONFIG_SCHED_DEBUG */ 140 141 /* 142 * Number of tasks to iterate in a single balance run. 143 * Limited because this is done with IRQs disabled. 144 */ 145 #ifdef CONFIG_PREEMPT_RT 146 const_debug unsigned int sysctl_sched_nr_migrate = 8; 147 #else 148 const_debug unsigned int sysctl_sched_nr_migrate = 32; 149 #endif 150 151 __read_mostly int scheduler_running; 152 153 #ifdef CONFIG_SCHED_CORE 154 155 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled); 156 157 /* kernel prio, less is more */ 158 static inline int __task_prio(struct task_struct *p) 159 { 160 if (p->sched_class == &stop_sched_class) /* trumps deadline */ 161 return -2; 162 163 if (rt_prio(p->prio)) /* includes deadline */ 164 return p->prio; /* [-1, 99] */ 165 166 if (p->sched_class == &idle_sched_class) 167 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */ 168 169 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */ 170 } 171 172 /* 173 * l(a,b) 174 * le(a,b) := !l(b,a) 175 * g(a,b) := l(b,a) 176 * ge(a,b) := !l(a,b) 177 */ 178 179 /* real prio, less is less */ 180 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi) 181 { 182 183 int pa = __task_prio(a), pb = __task_prio(b); 184 185 if (-pa < -pb) 186 return true; 187 188 if (-pb < -pa) 189 return false; 190 191 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */ 192 return !dl_time_before(a->dl.deadline, b->dl.deadline); 193 194 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */ 195 return cfs_prio_less(a, b, in_fi); 196 197 return false; 198 } 199 200 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b) 201 { 202 if (a->core_cookie < b->core_cookie) 203 return true; 204 205 if (a->core_cookie > b->core_cookie) 206 return false; 207 208 /* flip prio, so high prio is leftmost */ 209 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count)) 210 return true; 211 212 return false; 213 } 214 215 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node) 216 217 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b) 218 { 219 return __sched_core_less(__node_2_sc(a), __node_2_sc(b)); 220 } 221 222 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node) 223 { 224 const struct task_struct *p = __node_2_sc(node); 225 unsigned long cookie = (unsigned long)key; 226 227 if (cookie < p->core_cookie) 228 return -1; 229 230 if (cookie > p->core_cookie) 231 return 1; 232 233 return 0; 234 } 235 236 void sched_core_enqueue(struct rq *rq, struct task_struct *p) 237 { 238 rq->core->core_task_seq++; 239 240 if (!p->core_cookie) 241 return; 242 243 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less); 244 } 245 246 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) 247 { 248 rq->core->core_task_seq++; 249 250 if (sched_core_enqueued(p)) { 251 rb_erase(&p->core_node, &rq->core_tree); 252 RB_CLEAR_NODE(&p->core_node); 253 } 254 255 /* 256 * Migrating the last task off the cpu, with the cpu in forced idle 257 * state. Reschedule to create an accounting edge for forced idle, 258 * and re-examine whether the core is still in forced idle state. 259 */ 260 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 && 261 rq->core->core_forceidle_count && rq->curr == rq->idle) 262 resched_curr(rq); 263 } 264 265 /* 266 * Find left-most (aka, highest priority) task matching @cookie. 267 */ 268 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie) 269 { 270 struct rb_node *node; 271 272 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp); 273 /* 274 * The idle task always matches any cookie! 275 */ 276 if (!node) 277 return idle_sched_class.pick_task(rq); 278 279 return __node_2_sc(node); 280 } 281 282 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie) 283 { 284 struct rb_node *node = &p->core_node; 285 286 node = rb_next(node); 287 if (!node) 288 return NULL; 289 290 p = container_of(node, struct task_struct, core_node); 291 if (p->core_cookie != cookie) 292 return NULL; 293 294 return p; 295 } 296 297 /* 298 * Magic required such that: 299 * 300 * raw_spin_rq_lock(rq); 301 * ... 302 * raw_spin_rq_unlock(rq); 303 * 304 * ends up locking and unlocking the _same_ lock, and all CPUs 305 * always agree on what rq has what lock. 306 * 307 * XXX entirely possible to selectively enable cores, don't bother for now. 308 */ 309 310 static DEFINE_MUTEX(sched_core_mutex); 311 static atomic_t sched_core_count; 312 static struct cpumask sched_core_mask; 313 314 static void sched_core_lock(int cpu, unsigned long *flags) 315 { 316 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 317 int t, i = 0; 318 319 local_irq_save(*flags); 320 for_each_cpu(t, smt_mask) 321 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++); 322 } 323 324 static void sched_core_unlock(int cpu, unsigned long *flags) 325 { 326 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 327 int t; 328 329 for_each_cpu(t, smt_mask) 330 raw_spin_unlock(&cpu_rq(t)->__lock); 331 local_irq_restore(*flags); 332 } 333 334 static void __sched_core_flip(bool enabled) 335 { 336 unsigned long flags; 337 int cpu, t; 338 339 cpus_read_lock(); 340 341 /* 342 * Toggle the online cores, one by one. 343 */ 344 cpumask_copy(&sched_core_mask, cpu_online_mask); 345 for_each_cpu(cpu, &sched_core_mask) { 346 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 347 348 sched_core_lock(cpu, &flags); 349 350 for_each_cpu(t, smt_mask) 351 cpu_rq(t)->core_enabled = enabled; 352 353 cpu_rq(cpu)->core->core_forceidle_start = 0; 354 355 sched_core_unlock(cpu, &flags); 356 357 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask); 358 } 359 360 /* 361 * Toggle the offline CPUs. 362 */ 363 cpumask_copy(&sched_core_mask, cpu_possible_mask); 364 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask); 365 366 for_each_cpu(cpu, &sched_core_mask) 367 cpu_rq(cpu)->core_enabled = enabled; 368 369 cpus_read_unlock(); 370 } 371 372 static void sched_core_assert_empty(void) 373 { 374 int cpu; 375 376 for_each_possible_cpu(cpu) 377 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree)); 378 } 379 380 static void __sched_core_enable(void) 381 { 382 static_branch_enable(&__sched_core_enabled); 383 /* 384 * Ensure all previous instances of raw_spin_rq_*lock() have finished 385 * and future ones will observe !sched_core_disabled(). 386 */ 387 synchronize_rcu(); 388 __sched_core_flip(true); 389 sched_core_assert_empty(); 390 } 391 392 static void __sched_core_disable(void) 393 { 394 sched_core_assert_empty(); 395 __sched_core_flip(false); 396 static_branch_disable(&__sched_core_enabled); 397 } 398 399 void sched_core_get(void) 400 { 401 if (atomic_inc_not_zero(&sched_core_count)) 402 return; 403 404 mutex_lock(&sched_core_mutex); 405 if (!atomic_read(&sched_core_count)) 406 __sched_core_enable(); 407 408 smp_mb__before_atomic(); 409 atomic_inc(&sched_core_count); 410 mutex_unlock(&sched_core_mutex); 411 } 412 413 static void __sched_core_put(struct work_struct *work) 414 { 415 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) { 416 __sched_core_disable(); 417 mutex_unlock(&sched_core_mutex); 418 } 419 } 420 421 void sched_core_put(void) 422 { 423 static DECLARE_WORK(_work, __sched_core_put); 424 425 /* 426 * "There can be only one" 427 * 428 * Either this is the last one, or we don't actually need to do any 429 * 'work'. If it is the last *again*, we rely on 430 * WORK_STRUCT_PENDING_BIT. 431 */ 432 if (!atomic_add_unless(&sched_core_count, -1, 1)) 433 schedule_work(&_work); 434 } 435 436 #else /* !CONFIG_SCHED_CORE */ 437 438 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { } 439 static inline void 440 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { } 441 442 #endif /* CONFIG_SCHED_CORE */ 443 444 /* 445 * Serialization rules: 446 * 447 * Lock order: 448 * 449 * p->pi_lock 450 * rq->lock 451 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls) 452 * 453 * rq1->lock 454 * rq2->lock where: rq1 < rq2 455 * 456 * Regular state: 457 * 458 * Normal scheduling state is serialized by rq->lock. __schedule() takes the 459 * local CPU's rq->lock, it optionally removes the task from the runqueue and 460 * always looks at the local rq data structures to find the most eligible task 461 * to run next. 462 * 463 * Task enqueue is also under rq->lock, possibly taken from another CPU. 464 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to 465 * the local CPU to avoid bouncing the runqueue state around [ see 466 * ttwu_queue_wakelist() ] 467 * 468 * Task wakeup, specifically wakeups that involve migration, are horribly 469 * complicated to avoid having to take two rq->locks. 470 * 471 * Special state: 472 * 473 * System-calls and anything external will use task_rq_lock() which acquires 474 * both p->pi_lock and rq->lock. As a consequence the state they change is 475 * stable while holding either lock: 476 * 477 * - sched_setaffinity()/ 478 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed 479 * - set_user_nice(): p->se.load, p->*prio 480 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio, 481 * p->se.load, p->rt_priority, 482 * p->dl.dl_{runtime, deadline, period, flags, bw, density} 483 * - sched_setnuma(): p->numa_preferred_nid 484 * - sched_move_task()/ 485 * cpu_cgroup_fork(): p->sched_task_group 486 * - uclamp_update_active() p->uclamp* 487 * 488 * p->state <- TASK_*: 489 * 490 * is changed locklessly using set_current_state(), __set_current_state() or 491 * set_special_state(), see their respective comments, or by 492 * try_to_wake_up(). This latter uses p->pi_lock to serialize against 493 * concurrent self. 494 * 495 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }: 496 * 497 * is set by activate_task() and cleared by deactivate_task(), under 498 * rq->lock. Non-zero indicates the task is runnable, the special 499 * ON_RQ_MIGRATING state is used for migration without holding both 500 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock(). 501 * 502 * p->on_cpu <- { 0, 1 }: 503 * 504 * is set by prepare_task() and cleared by finish_task() such that it will be 505 * set before p is scheduled-in and cleared after p is scheduled-out, both 506 * under rq->lock. Non-zero indicates the task is running on its CPU. 507 * 508 * [ The astute reader will observe that it is possible for two tasks on one 509 * CPU to have ->on_cpu = 1 at the same time. ] 510 * 511 * task_cpu(p): is changed by set_task_cpu(), the rules are: 512 * 513 * - Don't call set_task_cpu() on a blocked task: 514 * 515 * We don't care what CPU we're not running on, this simplifies hotplug, 516 * the CPU assignment of blocked tasks isn't required to be valid. 517 * 518 * - for try_to_wake_up(), called under p->pi_lock: 519 * 520 * This allows try_to_wake_up() to only take one rq->lock, see its comment. 521 * 522 * - for migration called under rq->lock: 523 * [ see task_on_rq_migrating() in task_rq_lock() ] 524 * 525 * o move_queued_task() 526 * o detach_task() 527 * 528 * - for migration called under double_rq_lock(): 529 * 530 * o __migrate_swap_task() 531 * o push_rt_task() / pull_rt_task() 532 * o push_dl_task() / pull_dl_task() 533 * o dl_task_offline_migration() 534 * 535 */ 536 537 void raw_spin_rq_lock_nested(struct rq *rq, int subclass) 538 { 539 raw_spinlock_t *lock; 540 541 /* Matches synchronize_rcu() in __sched_core_enable() */ 542 preempt_disable(); 543 if (sched_core_disabled()) { 544 raw_spin_lock_nested(&rq->__lock, subclass); 545 /* preempt_count *MUST* be > 1 */ 546 preempt_enable_no_resched(); 547 return; 548 } 549 550 for (;;) { 551 lock = __rq_lockp(rq); 552 raw_spin_lock_nested(lock, subclass); 553 if (likely(lock == __rq_lockp(rq))) { 554 /* preempt_count *MUST* be > 1 */ 555 preempt_enable_no_resched(); 556 return; 557 } 558 raw_spin_unlock(lock); 559 } 560 } 561 562 bool raw_spin_rq_trylock(struct rq *rq) 563 { 564 raw_spinlock_t *lock; 565 bool ret; 566 567 /* Matches synchronize_rcu() in __sched_core_enable() */ 568 preempt_disable(); 569 if (sched_core_disabled()) { 570 ret = raw_spin_trylock(&rq->__lock); 571 preempt_enable(); 572 return ret; 573 } 574 575 for (;;) { 576 lock = __rq_lockp(rq); 577 ret = raw_spin_trylock(lock); 578 if (!ret || (likely(lock == __rq_lockp(rq)))) { 579 preempt_enable(); 580 return ret; 581 } 582 raw_spin_unlock(lock); 583 } 584 } 585 586 void raw_spin_rq_unlock(struct rq *rq) 587 { 588 raw_spin_unlock(rq_lockp(rq)); 589 } 590 591 #ifdef CONFIG_SMP 592 /* 593 * double_rq_lock - safely lock two runqueues 594 */ 595 void double_rq_lock(struct rq *rq1, struct rq *rq2) 596 { 597 lockdep_assert_irqs_disabled(); 598 599 if (rq_order_less(rq2, rq1)) 600 swap(rq1, rq2); 601 602 raw_spin_rq_lock(rq1); 603 if (__rq_lockp(rq1) != __rq_lockp(rq2)) 604 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING); 605 606 double_rq_clock_clear_update(rq1, rq2); 607 } 608 #endif 609 610 /* 611 * __task_rq_lock - lock the rq @p resides on. 612 */ 613 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) 614 __acquires(rq->lock) 615 { 616 struct rq *rq; 617 618 lockdep_assert_held(&p->pi_lock); 619 620 for (;;) { 621 rq = task_rq(p); 622 raw_spin_rq_lock(rq); 623 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { 624 rq_pin_lock(rq, rf); 625 return rq; 626 } 627 raw_spin_rq_unlock(rq); 628 629 while (unlikely(task_on_rq_migrating(p))) 630 cpu_relax(); 631 } 632 } 633 634 /* 635 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 636 */ 637 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) 638 __acquires(p->pi_lock) 639 __acquires(rq->lock) 640 { 641 struct rq *rq; 642 643 for (;;) { 644 raw_spin_lock_irqsave(&p->pi_lock, rf->flags); 645 rq = task_rq(p); 646 raw_spin_rq_lock(rq); 647 /* 648 * move_queued_task() task_rq_lock() 649 * 650 * ACQUIRE (rq->lock) 651 * [S] ->on_rq = MIGRATING [L] rq = task_rq() 652 * WMB (__set_task_cpu()) ACQUIRE (rq->lock); 653 * [S] ->cpu = new_cpu [L] task_rq() 654 * [L] ->on_rq 655 * RELEASE (rq->lock) 656 * 657 * If we observe the old CPU in task_rq_lock(), the acquire of 658 * the old rq->lock will fully serialize against the stores. 659 * 660 * If we observe the new CPU in task_rq_lock(), the address 661 * dependency headed by '[L] rq = task_rq()' and the acquire 662 * will pair with the WMB to ensure we then also see migrating. 663 */ 664 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { 665 rq_pin_lock(rq, rf); 666 return rq; 667 } 668 raw_spin_rq_unlock(rq); 669 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); 670 671 while (unlikely(task_on_rq_migrating(p))) 672 cpu_relax(); 673 } 674 } 675 676 /* 677 * RQ-clock updating methods: 678 */ 679 680 static void update_rq_clock_task(struct rq *rq, s64 delta) 681 { 682 /* 683 * In theory, the compile should just see 0 here, and optimize out the call 684 * to sched_rt_avg_update. But I don't trust it... 685 */ 686 s64 __maybe_unused steal = 0, irq_delta = 0; 687 688 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 689 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 690 691 /* 692 * Since irq_time is only updated on {soft,}irq_exit, we might run into 693 * this case when a previous update_rq_clock() happened inside a 694 * {soft,}irq region. 695 * 696 * When this happens, we stop ->clock_task and only update the 697 * prev_irq_time stamp to account for the part that fit, so that a next 698 * update will consume the rest. This ensures ->clock_task is 699 * monotonic. 700 * 701 * It does however cause some slight miss-attribution of {soft,}irq 702 * time, a more accurate solution would be to update the irq_time using 703 * the current rq->clock timestamp, except that would require using 704 * atomic ops. 705 */ 706 if (irq_delta > delta) 707 irq_delta = delta; 708 709 rq->prev_irq_time += irq_delta; 710 delta -= irq_delta; 711 #endif 712 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 713 if (static_key_false((¶virt_steal_rq_enabled))) { 714 steal = paravirt_steal_clock(cpu_of(rq)); 715 steal -= rq->prev_steal_time_rq; 716 717 if (unlikely(steal > delta)) 718 steal = delta; 719 720 rq->prev_steal_time_rq += steal; 721 delta -= steal; 722 } 723 #endif 724 725 rq->clock_task += delta; 726 727 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 728 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 729 update_irq_load_avg(rq, irq_delta + steal); 730 #endif 731 update_rq_clock_pelt(rq, delta); 732 } 733 734 void update_rq_clock(struct rq *rq) 735 { 736 s64 delta; 737 738 lockdep_assert_rq_held(rq); 739 740 if (rq->clock_update_flags & RQCF_ACT_SKIP) 741 return; 742 743 #ifdef CONFIG_SCHED_DEBUG 744 if (sched_feat(WARN_DOUBLE_CLOCK)) 745 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED); 746 rq->clock_update_flags |= RQCF_UPDATED; 747 #endif 748 749 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 750 if (delta < 0) 751 return; 752 rq->clock += delta; 753 update_rq_clock_task(rq, delta); 754 } 755 756 #ifdef CONFIG_SCHED_HRTICK 757 /* 758 * Use HR-timers to deliver accurate preemption points. 759 */ 760 761 static void hrtick_clear(struct rq *rq) 762 { 763 if (hrtimer_active(&rq->hrtick_timer)) 764 hrtimer_cancel(&rq->hrtick_timer); 765 } 766 767 /* 768 * High-resolution timer tick. 769 * Runs from hardirq context with interrupts disabled. 770 */ 771 static enum hrtimer_restart hrtick(struct hrtimer *timer) 772 { 773 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 774 struct rq_flags rf; 775 776 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 777 778 rq_lock(rq, &rf); 779 update_rq_clock(rq); 780 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 781 rq_unlock(rq, &rf); 782 783 return HRTIMER_NORESTART; 784 } 785 786 #ifdef CONFIG_SMP 787 788 static void __hrtick_restart(struct rq *rq) 789 { 790 struct hrtimer *timer = &rq->hrtick_timer; 791 ktime_t time = rq->hrtick_time; 792 793 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD); 794 } 795 796 /* 797 * called from hardirq (IPI) context 798 */ 799 static void __hrtick_start(void *arg) 800 { 801 struct rq *rq = arg; 802 struct rq_flags rf; 803 804 rq_lock(rq, &rf); 805 __hrtick_restart(rq); 806 rq_unlock(rq, &rf); 807 } 808 809 /* 810 * Called to set the hrtick timer state. 811 * 812 * called with rq->lock held and irqs disabled 813 */ 814 void hrtick_start(struct rq *rq, u64 delay) 815 { 816 struct hrtimer *timer = &rq->hrtick_timer; 817 s64 delta; 818 819 /* 820 * Don't schedule slices shorter than 10000ns, that just 821 * doesn't make sense and can cause timer DoS. 822 */ 823 delta = max_t(s64, delay, 10000LL); 824 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta); 825 826 if (rq == this_rq()) 827 __hrtick_restart(rq); 828 else 829 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 830 } 831 832 #else 833 /* 834 * Called to set the hrtick timer state. 835 * 836 * called with rq->lock held and irqs disabled 837 */ 838 void hrtick_start(struct rq *rq, u64 delay) 839 { 840 /* 841 * Don't schedule slices shorter than 10000ns, that just 842 * doesn't make sense. Rely on vruntime for fairness. 843 */ 844 delay = max_t(u64, delay, 10000LL); 845 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), 846 HRTIMER_MODE_REL_PINNED_HARD); 847 } 848 849 #endif /* CONFIG_SMP */ 850 851 static void hrtick_rq_init(struct rq *rq) 852 { 853 #ifdef CONFIG_SMP 854 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq); 855 #endif 856 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); 857 rq->hrtick_timer.function = hrtick; 858 } 859 #else /* CONFIG_SCHED_HRTICK */ 860 static inline void hrtick_clear(struct rq *rq) 861 { 862 } 863 864 static inline void hrtick_rq_init(struct rq *rq) 865 { 866 } 867 #endif /* CONFIG_SCHED_HRTICK */ 868 869 /* 870 * cmpxchg based fetch_or, macro so it works for different integer types 871 */ 872 #define fetch_or(ptr, mask) \ 873 ({ \ 874 typeof(ptr) _ptr = (ptr); \ 875 typeof(mask) _mask = (mask); \ 876 typeof(*_ptr) _val = *_ptr; \ 877 \ 878 do { \ 879 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \ 880 _val; \ 881 }) 882 883 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 884 /* 885 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 886 * this avoids any races wrt polling state changes and thereby avoids 887 * spurious IPIs. 888 */ 889 static inline bool set_nr_and_not_polling(struct task_struct *p) 890 { 891 struct thread_info *ti = task_thread_info(p); 892 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 893 } 894 895 /* 896 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 897 * 898 * If this returns true, then the idle task promises to call 899 * sched_ttwu_pending() and reschedule soon. 900 */ 901 static bool set_nr_if_polling(struct task_struct *p) 902 { 903 struct thread_info *ti = task_thread_info(p); 904 typeof(ti->flags) val = READ_ONCE(ti->flags); 905 906 for (;;) { 907 if (!(val & _TIF_POLLING_NRFLAG)) 908 return false; 909 if (val & _TIF_NEED_RESCHED) 910 return true; 911 if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED)) 912 break; 913 } 914 return true; 915 } 916 917 #else 918 static inline bool set_nr_and_not_polling(struct task_struct *p) 919 { 920 set_tsk_need_resched(p); 921 return true; 922 } 923 924 #ifdef CONFIG_SMP 925 static inline bool set_nr_if_polling(struct task_struct *p) 926 { 927 return false; 928 } 929 #endif 930 #endif 931 932 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task) 933 { 934 struct wake_q_node *node = &task->wake_q; 935 936 /* 937 * Atomically grab the task, if ->wake_q is !nil already it means 938 * it's already queued (either by us or someone else) and will get the 939 * wakeup due to that. 940 * 941 * In order to ensure that a pending wakeup will observe our pending 942 * state, even in the failed case, an explicit smp_mb() must be used. 943 */ 944 smp_mb__before_atomic(); 945 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL))) 946 return false; 947 948 /* 949 * The head is context local, there can be no concurrency. 950 */ 951 *head->lastp = node; 952 head->lastp = &node->next; 953 return true; 954 } 955 956 /** 957 * wake_q_add() - queue a wakeup for 'later' waking. 958 * @head: the wake_q_head to add @task to 959 * @task: the task to queue for 'later' wakeup 960 * 961 * Queue a task for later wakeup, most likely by the wake_up_q() call in the 962 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come 963 * instantly. 964 * 965 * This function must be used as-if it were wake_up_process(); IOW the task 966 * must be ready to be woken at this location. 967 */ 968 void wake_q_add(struct wake_q_head *head, struct task_struct *task) 969 { 970 if (__wake_q_add(head, task)) 971 get_task_struct(task); 972 } 973 974 /** 975 * wake_q_add_safe() - safely queue a wakeup for 'later' waking. 976 * @head: the wake_q_head to add @task to 977 * @task: the task to queue for 'later' wakeup 978 * 979 * Queue a task for later wakeup, most likely by the wake_up_q() call in the 980 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come 981 * instantly. 982 * 983 * This function must be used as-if it were wake_up_process(); IOW the task 984 * must be ready to be woken at this location. 985 * 986 * This function is essentially a task-safe equivalent to wake_q_add(). Callers 987 * that already hold reference to @task can call the 'safe' version and trust 988 * wake_q to do the right thing depending whether or not the @task is already 989 * queued for wakeup. 990 */ 991 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task) 992 { 993 if (!__wake_q_add(head, task)) 994 put_task_struct(task); 995 } 996 997 void wake_up_q(struct wake_q_head *head) 998 { 999 struct wake_q_node *node = head->first; 1000 1001 while (node != WAKE_Q_TAIL) { 1002 struct task_struct *task; 1003 1004 task = container_of(node, struct task_struct, wake_q); 1005 /* Task can safely be re-inserted now: */ 1006 node = node->next; 1007 task->wake_q.next = NULL; 1008 1009 /* 1010 * wake_up_process() executes a full barrier, which pairs with 1011 * the queueing in wake_q_add() so as not to miss wakeups. 1012 */ 1013 wake_up_process(task); 1014 put_task_struct(task); 1015 } 1016 } 1017 1018 /* 1019 * resched_curr - mark rq's current task 'to be rescheduled now'. 1020 * 1021 * On UP this means the setting of the need_resched flag, on SMP it 1022 * might also involve a cross-CPU call to trigger the scheduler on 1023 * the target CPU. 1024 */ 1025 void resched_curr(struct rq *rq) 1026 { 1027 struct task_struct *curr = rq->curr; 1028 int cpu; 1029 1030 lockdep_assert_rq_held(rq); 1031 1032 if (test_tsk_need_resched(curr)) 1033 return; 1034 1035 cpu = cpu_of(rq); 1036 1037 if (cpu == smp_processor_id()) { 1038 set_tsk_need_resched(curr); 1039 set_preempt_need_resched(); 1040 return; 1041 } 1042 1043 if (set_nr_and_not_polling(curr)) 1044 smp_send_reschedule(cpu); 1045 else 1046 trace_sched_wake_idle_without_ipi(cpu); 1047 } 1048 1049 void resched_cpu(int cpu) 1050 { 1051 struct rq *rq = cpu_rq(cpu); 1052 unsigned long flags; 1053 1054 raw_spin_rq_lock_irqsave(rq, flags); 1055 if (cpu_online(cpu) || cpu == smp_processor_id()) 1056 resched_curr(rq); 1057 raw_spin_rq_unlock_irqrestore(rq, flags); 1058 } 1059 1060 #ifdef CONFIG_SMP 1061 #ifdef CONFIG_NO_HZ_COMMON 1062 /* 1063 * In the semi idle case, use the nearest busy CPU for migrating timers 1064 * from an idle CPU. This is good for power-savings. 1065 * 1066 * We don't do similar optimization for completely idle system, as 1067 * selecting an idle CPU will add more delays to the timers than intended 1068 * (as that CPU's timer base may not be uptodate wrt jiffies etc). 1069 */ 1070 int get_nohz_timer_target(void) 1071 { 1072 int i, cpu = smp_processor_id(), default_cpu = -1; 1073 struct sched_domain *sd; 1074 const struct cpumask *hk_mask; 1075 1076 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) { 1077 if (!idle_cpu(cpu)) 1078 return cpu; 1079 default_cpu = cpu; 1080 } 1081 1082 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER); 1083 1084 rcu_read_lock(); 1085 for_each_domain(cpu, sd) { 1086 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) { 1087 if (cpu == i) 1088 continue; 1089 1090 if (!idle_cpu(i)) { 1091 cpu = i; 1092 goto unlock; 1093 } 1094 } 1095 } 1096 1097 if (default_cpu == -1) 1098 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER); 1099 cpu = default_cpu; 1100 unlock: 1101 rcu_read_unlock(); 1102 return cpu; 1103 } 1104 1105 /* 1106 * When add_timer_on() enqueues a timer into the timer wheel of an 1107 * idle CPU then this timer might expire before the next timer event 1108 * which is scheduled to wake up that CPU. In case of a completely 1109 * idle system the next event might even be infinite time into the 1110 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 1111 * leaves the inner idle loop so the newly added timer is taken into 1112 * account when the CPU goes back to idle and evaluates the timer 1113 * wheel for the next timer event. 1114 */ 1115 static void wake_up_idle_cpu(int cpu) 1116 { 1117 struct rq *rq = cpu_rq(cpu); 1118 1119 if (cpu == smp_processor_id()) 1120 return; 1121 1122 if (set_nr_and_not_polling(rq->idle)) 1123 smp_send_reschedule(cpu); 1124 else 1125 trace_sched_wake_idle_without_ipi(cpu); 1126 } 1127 1128 static bool wake_up_full_nohz_cpu(int cpu) 1129 { 1130 /* 1131 * We just need the target to call irq_exit() and re-evaluate 1132 * the next tick. The nohz full kick at least implies that. 1133 * If needed we can still optimize that later with an 1134 * empty IRQ. 1135 */ 1136 if (cpu_is_offline(cpu)) 1137 return true; /* Don't try to wake offline CPUs. */ 1138 if (tick_nohz_full_cpu(cpu)) { 1139 if (cpu != smp_processor_id() || 1140 tick_nohz_tick_stopped()) 1141 tick_nohz_full_kick_cpu(cpu); 1142 return true; 1143 } 1144 1145 return false; 1146 } 1147 1148 /* 1149 * Wake up the specified CPU. If the CPU is going offline, it is the 1150 * caller's responsibility to deal with the lost wakeup, for example, 1151 * by hooking into the CPU_DEAD notifier like timers and hrtimers do. 1152 */ 1153 void wake_up_nohz_cpu(int cpu) 1154 { 1155 if (!wake_up_full_nohz_cpu(cpu)) 1156 wake_up_idle_cpu(cpu); 1157 } 1158 1159 static void nohz_csd_func(void *info) 1160 { 1161 struct rq *rq = info; 1162 int cpu = cpu_of(rq); 1163 unsigned int flags; 1164 1165 /* 1166 * Release the rq::nohz_csd. 1167 */ 1168 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu)); 1169 WARN_ON(!(flags & NOHZ_KICK_MASK)); 1170 1171 rq->idle_balance = idle_cpu(cpu); 1172 if (rq->idle_balance && !need_resched()) { 1173 rq->nohz_idle_balance = flags; 1174 raise_softirq_irqoff(SCHED_SOFTIRQ); 1175 } 1176 } 1177 1178 #endif /* CONFIG_NO_HZ_COMMON */ 1179 1180 #ifdef CONFIG_NO_HZ_FULL 1181 bool sched_can_stop_tick(struct rq *rq) 1182 { 1183 int fifo_nr_running; 1184 1185 /* Deadline tasks, even if single, need the tick */ 1186 if (rq->dl.dl_nr_running) 1187 return false; 1188 1189 /* 1190 * If there are more than one RR tasks, we need the tick to affect the 1191 * actual RR behaviour. 1192 */ 1193 if (rq->rt.rr_nr_running) { 1194 if (rq->rt.rr_nr_running == 1) 1195 return true; 1196 else 1197 return false; 1198 } 1199 1200 /* 1201 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no 1202 * forced preemption between FIFO tasks. 1203 */ 1204 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running; 1205 if (fifo_nr_running) 1206 return true; 1207 1208 /* 1209 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left; 1210 * if there's more than one we need the tick for involuntary 1211 * preemption. 1212 */ 1213 if (rq->nr_running > 1) 1214 return false; 1215 1216 return true; 1217 } 1218 #endif /* CONFIG_NO_HZ_FULL */ 1219 #endif /* CONFIG_SMP */ 1220 1221 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 1222 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 1223 /* 1224 * Iterate task_group tree rooted at *from, calling @down when first entering a 1225 * node and @up when leaving it for the final time. 1226 * 1227 * Caller must hold rcu_lock or sufficient equivalent. 1228 */ 1229 int walk_tg_tree_from(struct task_group *from, 1230 tg_visitor down, tg_visitor up, void *data) 1231 { 1232 struct task_group *parent, *child; 1233 int ret; 1234 1235 parent = from; 1236 1237 down: 1238 ret = (*down)(parent, data); 1239 if (ret) 1240 goto out; 1241 list_for_each_entry_rcu(child, &parent->children, siblings) { 1242 parent = child; 1243 goto down; 1244 1245 up: 1246 continue; 1247 } 1248 ret = (*up)(parent, data); 1249 if (ret || parent == from) 1250 goto out; 1251 1252 child = parent; 1253 parent = parent->parent; 1254 if (parent) 1255 goto up; 1256 out: 1257 return ret; 1258 } 1259 1260 int tg_nop(struct task_group *tg, void *data) 1261 { 1262 return 0; 1263 } 1264 #endif 1265 1266 static void set_load_weight(struct task_struct *p, bool update_load) 1267 { 1268 int prio = p->static_prio - MAX_RT_PRIO; 1269 struct load_weight *load = &p->se.load; 1270 1271 /* 1272 * SCHED_IDLE tasks get minimal weight: 1273 */ 1274 if (task_has_idle_policy(p)) { 1275 load->weight = scale_load(WEIGHT_IDLEPRIO); 1276 load->inv_weight = WMULT_IDLEPRIO; 1277 return; 1278 } 1279 1280 /* 1281 * SCHED_OTHER tasks have to update their load when changing their 1282 * weight 1283 */ 1284 if (update_load && p->sched_class == &fair_sched_class) { 1285 reweight_task(p, prio); 1286 } else { 1287 load->weight = scale_load(sched_prio_to_weight[prio]); 1288 load->inv_weight = sched_prio_to_wmult[prio]; 1289 } 1290 } 1291 1292 #ifdef CONFIG_UCLAMP_TASK 1293 /* 1294 * Serializes updates of utilization clamp values 1295 * 1296 * The (slow-path) user-space triggers utilization clamp value updates which 1297 * can require updates on (fast-path) scheduler's data structures used to 1298 * support enqueue/dequeue operations. 1299 * While the per-CPU rq lock protects fast-path update operations, user-space 1300 * requests are serialized using a mutex to reduce the risk of conflicting 1301 * updates or API abuses. 1302 */ 1303 static DEFINE_MUTEX(uclamp_mutex); 1304 1305 /* Max allowed minimum utilization */ 1306 static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE; 1307 1308 /* Max allowed maximum utilization */ 1309 static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE; 1310 1311 /* 1312 * By default RT tasks run at the maximum performance point/capacity of the 1313 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to 1314 * SCHED_CAPACITY_SCALE. 1315 * 1316 * This knob allows admins to change the default behavior when uclamp is being 1317 * used. In battery powered devices, particularly, running at the maximum 1318 * capacity and frequency will increase energy consumption and shorten the 1319 * battery life. 1320 * 1321 * This knob only affects RT tasks that their uclamp_se->user_defined == false. 1322 * 1323 * This knob will not override the system default sched_util_clamp_min defined 1324 * above. 1325 */ 1326 static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE; 1327 1328 /* All clamps are required to be less or equal than these values */ 1329 static struct uclamp_se uclamp_default[UCLAMP_CNT]; 1330 1331 /* 1332 * This static key is used to reduce the uclamp overhead in the fast path. It 1333 * primarily disables the call to uclamp_rq_{inc, dec}() in 1334 * enqueue/dequeue_task(). 1335 * 1336 * This allows users to continue to enable uclamp in their kernel config with 1337 * minimum uclamp overhead in the fast path. 1338 * 1339 * As soon as userspace modifies any of the uclamp knobs, the static key is 1340 * enabled, since we have an actual users that make use of uclamp 1341 * functionality. 1342 * 1343 * The knobs that would enable this static key are: 1344 * 1345 * * A task modifying its uclamp value with sched_setattr(). 1346 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs. 1347 * * An admin modifying the cgroup cpu.uclamp.{min, max} 1348 */ 1349 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used); 1350 1351 /* Integer rounded range for each bucket */ 1352 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS) 1353 1354 #define for_each_clamp_id(clamp_id) \ 1355 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++) 1356 1357 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value) 1358 { 1359 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1); 1360 } 1361 1362 static inline unsigned int uclamp_none(enum uclamp_id clamp_id) 1363 { 1364 if (clamp_id == UCLAMP_MIN) 1365 return 0; 1366 return SCHED_CAPACITY_SCALE; 1367 } 1368 1369 static inline void uclamp_se_set(struct uclamp_se *uc_se, 1370 unsigned int value, bool user_defined) 1371 { 1372 uc_se->value = value; 1373 uc_se->bucket_id = uclamp_bucket_id(value); 1374 uc_se->user_defined = user_defined; 1375 } 1376 1377 static inline unsigned int 1378 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id, 1379 unsigned int clamp_value) 1380 { 1381 /* 1382 * Avoid blocked utilization pushing up the frequency when we go 1383 * idle (which drops the max-clamp) by retaining the last known 1384 * max-clamp. 1385 */ 1386 if (clamp_id == UCLAMP_MAX) { 1387 rq->uclamp_flags |= UCLAMP_FLAG_IDLE; 1388 return clamp_value; 1389 } 1390 1391 return uclamp_none(UCLAMP_MIN); 1392 } 1393 1394 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id, 1395 unsigned int clamp_value) 1396 { 1397 /* Reset max-clamp retention only on idle exit */ 1398 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE)) 1399 return; 1400 1401 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value); 1402 } 1403 1404 static inline 1405 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id, 1406 unsigned int clamp_value) 1407 { 1408 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket; 1409 int bucket_id = UCLAMP_BUCKETS - 1; 1410 1411 /* 1412 * Since both min and max clamps are max aggregated, find the 1413 * top most bucket with tasks in. 1414 */ 1415 for ( ; bucket_id >= 0; bucket_id--) { 1416 if (!bucket[bucket_id].tasks) 1417 continue; 1418 return bucket[bucket_id].value; 1419 } 1420 1421 /* No tasks -- default clamp values */ 1422 return uclamp_idle_value(rq, clamp_id, clamp_value); 1423 } 1424 1425 static void __uclamp_update_util_min_rt_default(struct task_struct *p) 1426 { 1427 unsigned int default_util_min; 1428 struct uclamp_se *uc_se; 1429 1430 lockdep_assert_held(&p->pi_lock); 1431 1432 uc_se = &p->uclamp_req[UCLAMP_MIN]; 1433 1434 /* Only sync if user didn't override the default */ 1435 if (uc_se->user_defined) 1436 return; 1437 1438 default_util_min = sysctl_sched_uclamp_util_min_rt_default; 1439 uclamp_se_set(uc_se, default_util_min, false); 1440 } 1441 1442 static void uclamp_update_util_min_rt_default(struct task_struct *p) 1443 { 1444 struct rq_flags rf; 1445 struct rq *rq; 1446 1447 if (!rt_task(p)) 1448 return; 1449 1450 /* Protect updates to p->uclamp_* */ 1451 rq = task_rq_lock(p, &rf); 1452 __uclamp_update_util_min_rt_default(p); 1453 task_rq_unlock(rq, p, &rf); 1454 } 1455 1456 static inline struct uclamp_se 1457 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id) 1458 { 1459 /* Copy by value as we could modify it */ 1460 struct uclamp_se uc_req = p->uclamp_req[clamp_id]; 1461 #ifdef CONFIG_UCLAMP_TASK_GROUP 1462 unsigned int tg_min, tg_max, value; 1463 1464 /* 1465 * Tasks in autogroups or root task group will be 1466 * restricted by system defaults. 1467 */ 1468 if (task_group_is_autogroup(task_group(p))) 1469 return uc_req; 1470 if (task_group(p) == &root_task_group) 1471 return uc_req; 1472 1473 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value; 1474 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value; 1475 value = uc_req.value; 1476 value = clamp(value, tg_min, tg_max); 1477 uclamp_se_set(&uc_req, value, false); 1478 #endif 1479 1480 return uc_req; 1481 } 1482 1483 /* 1484 * The effective clamp bucket index of a task depends on, by increasing 1485 * priority: 1486 * - the task specific clamp value, when explicitly requested from userspace 1487 * - the task group effective clamp value, for tasks not either in the root 1488 * group or in an autogroup 1489 * - the system default clamp value, defined by the sysadmin 1490 */ 1491 static inline struct uclamp_se 1492 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id) 1493 { 1494 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id); 1495 struct uclamp_se uc_max = uclamp_default[clamp_id]; 1496 1497 /* System default restrictions always apply */ 1498 if (unlikely(uc_req.value > uc_max.value)) 1499 return uc_max; 1500 1501 return uc_req; 1502 } 1503 1504 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id) 1505 { 1506 struct uclamp_se uc_eff; 1507 1508 /* Task currently refcounted: use back-annotated (effective) value */ 1509 if (p->uclamp[clamp_id].active) 1510 return (unsigned long)p->uclamp[clamp_id].value; 1511 1512 uc_eff = uclamp_eff_get(p, clamp_id); 1513 1514 return (unsigned long)uc_eff.value; 1515 } 1516 1517 /* 1518 * When a task is enqueued on a rq, the clamp bucket currently defined by the 1519 * task's uclamp::bucket_id is refcounted on that rq. This also immediately 1520 * updates the rq's clamp value if required. 1521 * 1522 * Tasks can have a task-specific value requested from user-space, track 1523 * within each bucket the maximum value for tasks refcounted in it. 1524 * This "local max aggregation" allows to track the exact "requested" value 1525 * for each bucket when all its RUNNABLE tasks require the same clamp. 1526 */ 1527 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p, 1528 enum uclamp_id clamp_id) 1529 { 1530 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id]; 1531 struct uclamp_se *uc_se = &p->uclamp[clamp_id]; 1532 struct uclamp_bucket *bucket; 1533 1534 lockdep_assert_rq_held(rq); 1535 1536 /* Update task effective clamp */ 1537 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id); 1538 1539 bucket = &uc_rq->bucket[uc_se->bucket_id]; 1540 bucket->tasks++; 1541 uc_se->active = true; 1542 1543 uclamp_idle_reset(rq, clamp_id, uc_se->value); 1544 1545 /* 1546 * Local max aggregation: rq buckets always track the max 1547 * "requested" clamp value of its RUNNABLE tasks. 1548 */ 1549 if (bucket->tasks == 1 || uc_se->value > bucket->value) 1550 bucket->value = uc_se->value; 1551 1552 if (uc_se->value > READ_ONCE(uc_rq->value)) 1553 WRITE_ONCE(uc_rq->value, uc_se->value); 1554 } 1555 1556 /* 1557 * When a task is dequeued from a rq, the clamp bucket refcounted by the task 1558 * is released. If this is the last task reference counting the rq's max 1559 * active clamp value, then the rq's clamp value is updated. 1560 * 1561 * Both refcounted tasks and rq's cached clamp values are expected to be 1562 * always valid. If it's detected they are not, as defensive programming, 1563 * enforce the expected state and warn. 1564 */ 1565 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p, 1566 enum uclamp_id clamp_id) 1567 { 1568 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id]; 1569 struct uclamp_se *uc_se = &p->uclamp[clamp_id]; 1570 struct uclamp_bucket *bucket; 1571 unsigned int bkt_clamp; 1572 unsigned int rq_clamp; 1573 1574 lockdep_assert_rq_held(rq); 1575 1576 /* 1577 * If sched_uclamp_used was enabled after task @p was enqueued, 1578 * we could end up with unbalanced call to uclamp_rq_dec_id(). 1579 * 1580 * In this case the uc_se->active flag should be false since no uclamp 1581 * accounting was performed at enqueue time and we can just return 1582 * here. 1583 * 1584 * Need to be careful of the following enqueue/dequeue ordering 1585 * problem too 1586 * 1587 * enqueue(taskA) 1588 * // sched_uclamp_used gets enabled 1589 * enqueue(taskB) 1590 * dequeue(taskA) 1591 * // Must not decrement bucket->tasks here 1592 * dequeue(taskB) 1593 * 1594 * where we could end up with stale data in uc_se and 1595 * bucket[uc_se->bucket_id]. 1596 * 1597 * The following check here eliminates the possibility of such race. 1598 */ 1599 if (unlikely(!uc_se->active)) 1600 return; 1601 1602 bucket = &uc_rq->bucket[uc_se->bucket_id]; 1603 1604 SCHED_WARN_ON(!bucket->tasks); 1605 if (likely(bucket->tasks)) 1606 bucket->tasks--; 1607 1608 uc_se->active = false; 1609 1610 /* 1611 * Keep "local max aggregation" simple and accept to (possibly) 1612 * overboost some RUNNABLE tasks in the same bucket. 1613 * The rq clamp bucket value is reset to its base value whenever 1614 * there are no more RUNNABLE tasks refcounting it. 1615 */ 1616 if (likely(bucket->tasks)) 1617 return; 1618 1619 rq_clamp = READ_ONCE(uc_rq->value); 1620 /* 1621 * Defensive programming: this should never happen. If it happens, 1622 * e.g. due to future modification, warn and fixup the expected value. 1623 */ 1624 SCHED_WARN_ON(bucket->value > rq_clamp); 1625 if (bucket->value >= rq_clamp) { 1626 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value); 1627 WRITE_ONCE(uc_rq->value, bkt_clamp); 1628 } 1629 } 1630 1631 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) 1632 { 1633 enum uclamp_id clamp_id; 1634 1635 /* 1636 * Avoid any overhead until uclamp is actually used by the userspace. 1637 * 1638 * The condition is constructed such that a NOP is generated when 1639 * sched_uclamp_used is disabled. 1640 */ 1641 if (!static_branch_unlikely(&sched_uclamp_used)) 1642 return; 1643 1644 if (unlikely(!p->sched_class->uclamp_enabled)) 1645 return; 1646 1647 for_each_clamp_id(clamp_id) 1648 uclamp_rq_inc_id(rq, p, clamp_id); 1649 1650 /* Reset clamp idle holding when there is one RUNNABLE task */ 1651 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE) 1652 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE; 1653 } 1654 1655 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) 1656 { 1657 enum uclamp_id clamp_id; 1658 1659 /* 1660 * Avoid any overhead until uclamp is actually used by the userspace. 1661 * 1662 * The condition is constructed such that a NOP is generated when 1663 * sched_uclamp_used is disabled. 1664 */ 1665 if (!static_branch_unlikely(&sched_uclamp_used)) 1666 return; 1667 1668 if (unlikely(!p->sched_class->uclamp_enabled)) 1669 return; 1670 1671 for_each_clamp_id(clamp_id) 1672 uclamp_rq_dec_id(rq, p, clamp_id); 1673 } 1674 1675 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p, 1676 enum uclamp_id clamp_id) 1677 { 1678 if (!p->uclamp[clamp_id].active) 1679 return; 1680 1681 uclamp_rq_dec_id(rq, p, clamp_id); 1682 uclamp_rq_inc_id(rq, p, clamp_id); 1683 1684 /* 1685 * Make sure to clear the idle flag if we've transiently reached 0 1686 * active tasks on rq. 1687 */ 1688 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE)) 1689 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE; 1690 } 1691 1692 static inline void 1693 uclamp_update_active(struct task_struct *p) 1694 { 1695 enum uclamp_id clamp_id; 1696 struct rq_flags rf; 1697 struct rq *rq; 1698 1699 /* 1700 * Lock the task and the rq where the task is (or was) queued. 1701 * 1702 * We might lock the (previous) rq of a !RUNNABLE task, but that's the 1703 * price to pay to safely serialize util_{min,max} updates with 1704 * enqueues, dequeues and migration operations. 1705 * This is the same locking schema used by __set_cpus_allowed_ptr(). 1706 */ 1707 rq = task_rq_lock(p, &rf); 1708 1709 /* 1710 * Setting the clamp bucket is serialized by task_rq_lock(). 1711 * If the task is not yet RUNNABLE and its task_struct is not 1712 * affecting a valid clamp bucket, the next time it's enqueued, 1713 * it will already see the updated clamp bucket value. 1714 */ 1715 for_each_clamp_id(clamp_id) 1716 uclamp_rq_reinc_id(rq, p, clamp_id); 1717 1718 task_rq_unlock(rq, p, &rf); 1719 } 1720 1721 #ifdef CONFIG_UCLAMP_TASK_GROUP 1722 static inline void 1723 uclamp_update_active_tasks(struct cgroup_subsys_state *css) 1724 { 1725 struct css_task_iter it; 1726 struct task_struct *p; 1727 1728 css_task_iter_start(css, 0, &it); 1729 while ((p = css_task_iter_next(&it))) 1730 uclamp_update_active(p); 1731 css_task_iter_end(&it); 1732 } 1733 1734 static void cpu_util_update_eff(struct cgroup_subsys_state *css); 1735 #endif 1736 1737 #ifdef CONFIG_SYSCTL 1738 #ifdef CONFIG_UCLAMP_TASK 1739 #ifdef CONFIG_UCLAMP_TASK_GROUP 1740 static void uclamp_update_root_tg(void) 1741 { 1742 struct task_group *tg = &root_task_group; 1743 1744 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN], 1745 sysctl_sched_uclamp_util_min, false); 1746 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX], 1747 sysctl_sched_uclamp_util_max, false); 1748 1749 rcu_read_lock(); 1750 cpu_util_update_eff(&root_task_group.css); 1751 rcu_read_unlock(); 1752 } 1753 #else 1754 static void uclamp_update_root_tg(void) { } 1755 #endif 1756 1757 static void uclamp_sync_util_min_rt_default(void) 1758 { 1759 struct task_struct *g, *p; 1760 1761 /* 1762 * copy_process() sysctl_uclamp 1763 * uclamp_min_rt = X; 1764 * write_lock(&tasklist_lock) read_lock(&tasklist_lock) 1765 * // link thread smp_mb__after_spinlock() 1766 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock); 1767 * sched_post_fork() for_each_process_thread() 1768 * __uclamp_sync_rt() __uclamp_sync_rt() 1769 * 1770 * Ensures that either sched_post_fork() will observe the new 1771 * uclamp_min_rt or for_each_process_thread() will observe the new 1772 * task. 1773 */ 1774 read_lock(&tasklist_lock); 1775 smp_mb__after_spinlock(); 1776 read_unlock(&tasklist_lock); 1777 1778 rcu_read_lock(); 1779 for_each_process_thread(g, p) 1780 uclamp_update_util_min_rt_default(p); 1781 rcu_read_unlock(); 1782 } 1783 1784 static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write, 1785 void *buffer, size_t *lenp, loff_t *ppos) 1786 { 1787 bool update_root_tg = false; 1788 int old_min, old_max, old_min_rt; 1789 int result; 1790 1791 mutex_lock(&uclamp_mutex); 1792 old_min = sysctl_sched_uclamp_util_min; 1793 old_max = sysctl_sched_uclamp_util_max; 1794 old_min_rt = sysctl_sched_uclamp_util_min_rt_default; 1795 1796 result = proc_dointvec(table, write, buffer, lenp, ppos); 1797 if (result) 1798 goto undo; 1799 if (!write) 1800 goto done; 1801 1802 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max || 1803 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE || 1804 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) { 1805 1806 result = -EINVAL; 1807 goto undo; 1808 } 1809 1810 if (old_min != sysctl_sched_uclamp_util_min) { 1811 uclamp_se_set(&uclamp_default[UCLAMP_MIN], 1812 sysctl_sched_uclamp_util_min, false); 1813 update_root_tg = true; 1814 } 1815 if (old_max != sysctl_sched_uclamp_util_max) { 1816 uclamp_se_set(&uclamp_default[UCLAMP_MAX], 1817 sysctl_sched_uclamp_util_max, false); 1818 update_root_tg = true; 1819 } 1820 1821 if (update_root_tg) { 1822 static_branch_enable(&sched_uclamp_used); 1823 uclamp_update_root_tg(); 1824 } 1825 1826 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) { 1827 static_branch_enable(&sched_uclamp_used); 1828 uclamp_sync_util_min_rt_default(); 1829 } 1830 1831 /* 1832 * We update all RUNNABLE tasks only when task groups are in use. 1833 * Otherwise, keep it simple and do just a lazy update at each next 1834 * task enqueue time. 1835 */ 1836 1837 goto done; 1838 1839 undo: 1840 sysctl_sched_uclamp_util_min = old_min; 1841 sysctl_sched_uclamp_util_max = old_max; 1842 sysctl_sched_uclamp_util_min_rt_default = old_min_rt; 1843 done: 1844 mutex_unlock(&uclamp_mutex); 1845 1846 return result; 1847 } 1848 #endif 1849 #endif 1850 1851 static int uclamp_validate(struct task_struct *p, 1852 const struct sched_attr *attr) 1853 { 1854 int util_min = p->uclamp_req[UCLAMP_MIN].value; 1855 int util_max = p->uclamp_req[UCLAMP_MAX].value; 1856 1857 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) { 1858 util_min = attr->sched_util_min; 1859 1860 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1) 1861 return -EINVAL; 1862 } 1863 1864 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) { 1865 util_max = attr->sched_util_max; 1866 1867 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1) 1868 return -EINVAL; 1869 } 1870 1871 if (util_min != -1 && util_max != -1 && util_min > util_max) 1872 return -EINVAL; 1873 1874 /* 1875 * We have valid uclamp attributes; make sure uclamp is enabled. 1876 * 1877 * We need to do that here, because enabling static branches is a 1878 * blocking operation which obviously cannot be done while holding 1879 * scheduler locks. 1880 */ 1881 static_branch_enable(&sched_uclamp_used); 1882 1883 return 0; 1884 } 1885 1886 static bool uclamp_reset(const struct sched_attr *attr, 1887 enum uclamp_id clamp_id, 1888 struct uclamp_se *uc_se) 1889 { 1890 /* Reset on sched class change for a non user-defined clamp value. */ 1891 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) && 1892 !uc_se->user_defined) 1893 return true; 1894 1895 /* Reset on sched_util_{min,max} == -1. */ 1896 if (clamp_id == UCLAMP_MIN && 1897 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN && 1898 attr->sched_util_min == -1) { 1899 return true; 1900 } 1901 1902 if (clamp_id == UCLAMP_MAX && 1903 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX && 1904 attr->sched_util_max == -1) { 1905 return true; 1906 } 1907 1908 return false; 1909 } 1910 1911 static void __setscheduler_uclamp(struct task_struct *p, 1912 const struct sched_attr *attr) 1913 { 1914 enum uclamp_id clamp_id; 1915 1916 for_each_clamp_id(clamp_id) { 1917 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id]; 1918 unsigned int value; 1919 1920 if (!uclamp_reset(attr, clamp_id, uc_se)) 1921 continue; 1922 1923 /* 1924 * RT by default have a 100% boost value that could be modified 1925 * at runtime. 1926 */ 1927 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN)) 1928 value = sysctl_sched_uclamp_util_min_rt_default; 1929 else 1930 value = uclamp_none(clamp_id); 1931 1932 uclamp_se_set(uc_se, value, false); 1933 1934 } 1935 1936 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP))) 1937 return; 1938 1939 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN && 1940 attr->sched_util_min != -1) { 1941 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN], 1942 attr->sched_util_min, true); 1943 } 1944 1945 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX && 1946 attr->sched_util_max != -1) { 1947 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX], 1948 attr->sched_util_max, true); 1949 } 1950 } 1951 1952 static void uclamp_fork(struct task_struct *p) 1953 { 1954 enum uclamp_id clamp_id; 1955 1956 /* 1957 * We don't need to hold task_rq_lock() when updating p->uclamp_* here 1958 * as the task is still at its early fork stages. 1959 */ 1960 for_each_clamp_id(clamp_id) 1961 p->uclamp[clamp_id].active = false; 1962 1963 if (likely(!p->sched_reset_on_fork)) 1964 return; 1965 1966 for_each_clamp_id(clamp_id) { 1967 uclamp_se_set(&p->uclamp_req[clamp_id], 1968 uclamp_none(clamp_id), false); 1969 } 1970 } 1971 1972 static void uclamp_post_fork(struct task_struct *p) 1973 { 1974 uclamp_update_util_min_rt_default(p); 1975 } 1976 1977 static void __init init_uclamp_rq(struct rq *rq) 1978 { 1979 enum uclamp_id clamp_id; 1980 struct uclamp_rq *uc_rq = rq->uclamp; 1981 1982 for_each_clamp_id(clamp_id) { 1983 uc_rq[clamp_id] = (struct uclamp_rq) { 1984 .value = uclamp_none(clamp_id) 1985 }; 1986 } 1987 1988 rq->uclamp_flags = UCLAMP_FLAG_IDLE; 1989 } 1990 1991 static void __init init_uclamp(void) 1992 { 1993 struct uclamp_se uc_max = {}; 1994 enum uclamp_id clamp_id; 1995 int cpu; 1996 1997 for_each_possible_cpu(cpu) 1998 init_uclamp_rq(cpu_rq(cpu)); 1999 2000 for_each_clamp_id(clamp_id) { 2001 uclamp_se_set(&init_task.uclamp_req[clamp_id], 2002 uclamp_none(clamp_id), false); 2003 } 2004 2005 /* System defaults allow max clamp values for both indexes */ 2006 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false); 2007 for_each_clamp_id(clamp_id) { 2008 uclamp_default[clamp_id] = uc_max; 2009 #ifdef CONFIG_UCLAMP_TASK_GROUP 2010 root_task_group.uclamp_req[clamp_id] = uc_max; 2011 root_task_group.uclamp[clamp_id] = uc_max; 2012 #endif 2013 } 2014 } 2015 2016 #else /* CONFIG_UCLAMP_TASK */ 2017 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { } 2018 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { } 2019 static inline int uclamp_validate(struct task_struct *p, 2020 const struct sched_attr *attr) 2021 { 2022 return -EOPNOTSUPP; 2023 } 2024 static void __setscheduler_uclamp(struct task_struct *p, 2025 const struct sched_attr *attr) { } 2026 static inline void uclamp_fork(struct task_struct *p) { } 2027 static inline void uclamp_post_fork(struct task_struct *p) { } 2028 static inline void init_uclamp(void) { } 2029 #endif /* CONFIG_UCLAMP_TASK */ 2030 2031 bool sched_task_on_rq(struct task_struct *p) 2032 { 2033 return task_on_rq_queued(p); 2034 } 2035 2036 unsigned long get_wchan(struct task_struct *p) 2037 { 2038 unsigned long ip = 0; 2039 unsigned int state; 2040 2041 if (!p || p == current) 2042 return 0; 2043 2044 /* Only get wchan if task is blocked and we can keep it that way. */ 2045 raw_spin_lock_irq(&p->pi_lock); 2046 state = READ_ONCE(p->__state); 2047 smp_rmb(); /* see try_to_wake_up() */ 2048 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq) 2049 ip = __get_wchan(p); 2050 raw_spin_unlock_irq(&p->pi_lock); 2051 2052 return ip; 2053 } 2054 2055 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 2056 { 2057 if (!(flags & ENQUEUE_NOCLOCK)) 2058 update_rq_clock(rq); 2059 2060 if (!(flags & ENQUEUE_RESTORE)) { 2061 sched_info_enqueue(rq, p); 2062 psi_enqueue(p, flags & ENQUEUE_WAKEUP); 2063 } 2064 2065 uclamp_rq_inc(rq, p); 2066 p->sched_class->enqueue_task(rq, p, flags); 2067 2068 if (sched_core_enabled(rq)) 2069 sched_core_enqueue(rq, p); 2070 } 2071 2072 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 2073 { 2074 if (sched_core_enabled(rq)) 2075 sched_core_dequeue(rq, p, flags); 2076 2077 if (!(flags & DEQUEUE_NOCLOCK)) 2078 update_rq_clock(rq); 2079 2080 if (!(flags & DEQUEUE_SAVE)) { 2081 sched_info_dequeue(rq, p); 2082 psi_dequeue(p, flags & DEQUEUE_SLEEP); 2083 } 2084 2085 uclamp_rq_dec(rq, p); 2086 p->sched_class->dequeue_task(rq, p, flags); 2087 } 2088 2089 void activate_task(struct rq *rq, struct task_struct *p, int flags) 2090 { 2091 enqueue_task(rq, p, flags); 2092 2093 p->on_rq = TASK_ON_RQ_QUEUED; 2094 } 2095 2096 void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 2097 { 2098 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING; 2099 2100 dequeue_task(rq, p, flags); 2101 } 2102 2103 static inline int __normal_prio(int policy, int rt_prio, int nice) 2104 { 2105 int prio; 2106 2107 if (dl_policy(policy)) 2108 prio = MAX_DL_PRIO - 1; 2109 else if (rt_policy(policy)) 2110 prio = MAX_RT_PRIO - 1 - rt_prio; 2111 else 2112 prio = NICE_TO_PRIO(nice); 2113 2114 return prio; 2115 } 2116 2117 /* 2118 * Calculate the expected normal priority: i.e. priority 2119 * without taking RT-inheritance into account. Might be 2120 * boosted by interactivity modifiers. Changes upon fork, 2121 * setprio syscalls, and whenever the interactivity 2122 * estimator recalculates. 2123 */ 2124 static inline int normal_prio(struct task_struct *p) 2125 { 2126 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio)); 2127 } 2128 2129 /* 2130 * Calculate the current priority, i.e. the priority 2131 * taken into account by the scheduler. This value might 2132 * be boosted by RT tasks, or might be boosted by 2133 * interactivity modifiers. Will be RT if the task got 2134 * RT-boosted. If not then it returns p->normal_prio. 2135 */ 2136 static int effective_prio(struct task_struct *p) 2137 { 2138 p->normal_prio = normal_prio(p); 2139 /* 2140 * If we are RT tasks or we were boosted to RT priority, 2141 * keep the priority unchanged. Otherwise, update priority 2142 * to the normal priority: 2143 */ 2144 if (!rt_prio(p->prio)) 2145 return p->normal_prio; 2146 return p->prio; 2147 } 2148 2149 /** 2150 * task_curr - is this task currently executing on a CPU? 2151 * @p: the task in question. 2152 * 2153 * Return: 1 if the task is currently executing. 0 otherwise. 2154 */ 2155 inline int task_curr(const struct task_struct *p) 2156 { 2157 return cpu_curr(task_cpu(p)) == p; 2158 } 2159 2160 /* 2161 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock, 2162 * use the balance_callback list if you want balancing. 2163 * 2164 * this means any call to check_class_changed() must be followed by a call to 2165 * balance_callback(). 2166 */ 2167 static inline void check_class_changed(struct rq *rq, struct task_struct *p, 2168 const struct sched_class *prev_class, 2169 int oldprio) 2170 { 2171 if (prev_class != p->sched_class) { 2172 if (prev_class->switched_from) 2173 prev_class->switched_from(rq, p); 2174 2175 p->sched_class->switched_to(rq, p); 2176 } else if (oldprio != p->prio || dl_task(p)) 2177 p->sched_class->prio_changed(rq, p, oldprio); 2178 } 2179 2180 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 2181 { 2182 if (p->sched_class == rq->curr->sched_class) 2183 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 2184 else if (sched_class_above(p->sched_class, rq->curr->sched_class)) 2185 resched_curr(rq); 2186 2187 /* 2188 * A queue event has occurred, and we're going to schedule. In 2189 * this case, we can save a useless back to back clock update. 2190 */ 2191 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) 2192 rq_clock_skip_update(rq); 2193 } 2194 2195 #ifdef CONFIG_SMP 2196 2197 static void 2198 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags); 2199 2200 static int __set_cpus_allowed_ptr(struct task_struct *p, 2201 const struct cpumask *new_mask, 2202 u32 flags); 2203 2204 static void migrate_disable_switch(struct rq *rq, struct task_struct *p) 2205 { 2206 if (likely(!p->migration_disabled)) 2207 return; 2208 2209 if (p->cpus_ptr != &p->cpus_mask) 2210 return; 2211 2212 /* 2213 * Violates locking rules! see comment in __do_set_cpus_allowed(). 2214 */ 2215 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE); 2216 } 2217 2218 void migrate_disable(void) 2219 { 2220 struct task_struct *p = current; 2221 2222 if (p->migration_disabled) { 2223 p->migration_disabled++; 2224 return; 2225 } 2226 2227 preempt_disable(); 2228 this_rq()->nr_pinned++; 2229 p->migration_disabled = 1; 2230 preempt_enable(); 2231 } 2232 EXPORT_SYMBOL_GPL(migrate_disable); 2233 2234 void migrate_enable(void) 2235 { 2236 struct task_struct *p = current; 2237 2238 if (p->migration_disabled > 1) { 2239 p->migration_disabled--; 2240 return; 2241 } 2242 2243 if (WARN_ON_ONCE(!p->migration_disabled)) 2244 return; 2245 2246 /* 2247 * Ensure stop_task runs either before or after this, and that 2248 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule(). 2249 */ 2250 preempt_disable(); 2251 if (p->cpus_ptr != &p->cpus_mask) 2252 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE); 2253 /* 2254 * Mustn't clear migration_disabled() until cpus_ptr points back at the 2255 * regular cpus_mask, otherwise things that race (eg. 2256 * select_fallback_rq) get confused. 2257 */ 2258 barrier(); 2259 p->migration_disabled = 0; 2260 this_rq()->nr_pinned--; 2261 preempt_enable(); 2262 } 2263 EXPORT_SYMBOL_GPL(migrate_enable); 2264 2265 static inline bool rq_has_pinned_tasks(struct rq *rq) 2266 { 2267 return rq->nr_pinned; 2268 } 2269 2270 /* 2271 * Per-CPU kthreads are allowed to run on !active && online CPUs, see 2272 * __set_cpus_allowed_ptr() and select_fallback_rq(). 2273 */ 2274 static inline bool is_cpu_allowed(struct task_struct *p, int cpu) 2275 { 2276 /* When not in the task's cpumask, no point in looking further. */ 2277 if (!cpumask_test_cpu(cpu, p->cpus_ptr)) 2278 return false; 2279 2280 /* migrate_disabled() must be allowed to finish. */ 2281 if (is_migration_disabled(p)) 2282 return cpu_online(cpu); 2283 2284 /* Non kernel threads are not allowed during either online or offline. */ 2285 if (!(p->flags & PF_KTHREAD)) 2286 return cpu_active(cpu) && task_cpu_possible(cpu, p); 2287 2288 /* KTHREAD_IS_PER_CPU is always allowed. */ 2289 if (kthread_is_per_cpu(p)) 2290 return cpu_online(cpu); 2291 2292 /* Regular kernel threads don't get to stay during offline. */ 2293 if (cpu_dying(cpu)) 2294 return false; 2295 2296 /* But are allowed during online. */ 2297 return cpu_online(cpu); 2298 } 2299 2300 /* 2301 * This is how migration works: 2302 * 2303 * 1) we invoke migration_cpu_stop() on the target CPU using 2304 * stop_one_cpu(). 2305 * 2) stopper starts to run (implicitly forcing the migrated thread 2306 * off the CPU) 2307 * 3) it checks whether the migrated task is still in the wrong runqueue. 2308 * 4) if it's in the wrong runqueue then the migration thread removes 2309 * it and puts it into the right queue. 2310 * 5) stopper completes and stop_one_cpu() returns and the migration 2311 * is done. 2312 */ 2313 2314 /* 2315 * move_queued_task - move a queued task to new rq. 2316 * 2317 * Returns (locked) new rq. Old rq's lock is released. 2318 */ 2319 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf, 2320 struct task_struct *p, int new_cpu) 2321 { 2322 lockdep_assert_rq_held(rq); 2323 2324 deactivate_task(rq, p, DEQUEUE_NOCLOCK); 2325 set_task_cpu(p, new_cpu); 2326 rq_unlock(rq, rf); 2327 2328 rq = cpu_rq(new_cpu); 2329 2330 rq_lock(rq, rf); 2331 BUG_ON(task_cpu(p) != new_cpu); 2332 activate_task(rq, p, 0); 2333 check_preempt_curr(rq, p, 0); 2334 2335 return rq; 2336 } 2337 2338 struct migration_arg { 2339 struct task_struct *task; 2340 int dest_cpu; 2341 struct set_affinity_pending *pending; 2342 }; 2343 2344 /* 2345 * @refs: number of wait_for_completion() 2346 * @stop_pending: is @stop_work in use 2347 */ 2348 struct set_affinity_pending { 2349 refcount_t refs; 2350 unsigned int stop_pending; 2351 struct completion done; 2352 struct cpu_stop_work stop_work; 2353 struct migration_arg arg; 2354 }; 2355 2356 /* 2357 * Move (not current) task off this CPU, onto the destination CPU. We're doing 2358 * this because either it can't run here any more (set_cpus_allowed() 2359 * away from this CPU, or CPU going down), or because we're 2360 * attempting to rebalance this task on exec (sched_exec). 2361 * 2362 * So we race with normal scheduler movements, but that's OK, as long 2363 * as the task is no longer on this CPU. 2364 */ 2365 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf, 2366 struct task_struct *p, int dest_cpu) 2367 { 2368 /* Affinity changed (again). */ 2369 if (!is_cpu_allowed(p, dest_cpu)) 2370 return rq; 2371 2372 update_rq_clock(rq); 2373 rq = move_queued_task(rq, rf, p, dest_cpu); 2374 2375 return rq; 2376 } 2377 2378 /* 2379 * migration_cpu_stop - this will be executed by a highprio stopper thread 2380 * and performs thread migration by bumping thread off CPU then 2381 * 'pushing' onto another runqueue. 2382 */ 2383 static int migration_cpu_stop(void *data) 2384 { 2385 struct migration_arg *arg = data; 2386 struct set_affinity_pending *pending = arg->pending; 2387 struct task_struct *p = arg->task; 2388 struct rq *rq = this_rq(); 2389 bool complete = false; 2390 struct rq_flags rf; 2391 2392 /* 2393 * The original target CPU might have gone down and we might 2394 * be on another CPU but it doesn't matter. 2395 */ 2396 local_irq_save(rf.flags); 2397 /* 2398 * We need to explicitly wake pending tasks before running 2399 * __migrate_task() such that we will not miss enforcing cpus_ptr 2400 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. 2401 */ 2402 flush_smp_call_function_queue(); 2403 2404 raw_spin_lock(&p->pi_lock); 2405 rq_lock(rq, &rf); 2406 2407 /* 2408 * If we were passed a pending, then ->stop_pending was set, thus 2409 * p->migration_pending must have remained stable. 2410 */ 2411 WARN_ON_ONCE(pending && pending != p->migration_pending); 2412 2413 /* 2414 * If task_rq(p) != rq, it cannot be migrated here, because we're 2415 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because 2416 * we're holding p->pi_lock. 2417 */ 2418 if (task_rq(p) == rq) { 2419 if (is_migration_disabled(p)) 2420 goto out; 2421 2422 if (pending) { 2423 p->migration_pending = NULL; 2424 complete = true; 2425 2426 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) 2427 goto out; 2428 } 2429 2430 if (task_on_rq_queued(p)) 2431 rq = __migrate_task(rq, &rf, p, arg->dest_cpu); 2432 else 2433 p->wake_cpu = arg->dest_cpu; 2434 2435 /* 2436 * XXX __migrate_task() can fail, at which point we might end 2437 * up running on a dodgy CPU, AFAICT this can only happen 2438 * during CPU hotplug, at which point we'll get pushed out 2439 * anyway, so it's probably not a big deal. 2440 */ 2441 2442 } else if (pending) { 2443 /* 2444 * This happens when we get migrated between migrate_enable()'s 2445 * preempt_enable() and scheduling the stopper task. At that 2446 * point we're a regular task again and not current anymore. 2447 * 2448 * A !PREEMPT kernel has a giant hole here, which makes it far 2449 * more likely. 2450 */ 2451 2452 /* 2453 * The task moved before the stopper got to run. We're holding 2454 * ->pi_lock, so the allowed mask is stable - if it got 2455 * somewhere allowed, we're done. 2456 */ 2457 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) { 2458 p->migration_pending = NULL; 2459 complete = true; 2460 goto out; 2461 } 2462 2463 /* 2464 * When migrate_enable() hits a rq mis-match we can't reliably 2465 * determine is_migration_disabled() and so have to chase after 2466 * it. 2467 */ 2468 WARN_ON_ONCE(!pending->stop_pending); 2469 task_rq_unlock(rq, p, &rf); 2470 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop, 2471 &pending->arg, &pending->stop_work); 2472 return 0; 2473 } 2474 out: 2475 if (pending) 2476 pending->stop_pending = false; 2477 task_rq_unlock(rq, p, &rf); 2478 2479 if (complete) 2480 complete_all(&pending->done); 2481 2482 return 0; 2483 } 2484 2485 int push_cpu_stop(void *arg) 2486 { 2487 struct rq *lowest_rq = NULL, *rq = this_rq(); 2488 struct task_struct *p = arg; 2489 2490 raw_spin_lock_irq(&p->pi_lock); 2491 raw_spin_rq_lock(rq); 2492 2493 if (task_rq(p) != rq) 2494 goto out_unlock; 2495 2496 if (is_migration_disabled(p)) { 2497 p->migration_flags |= MDF_PUSH; 2498 goto out_unlock; 2499 } 2500 2501 p->migration_flags &= ~MDF_PUSH; 2502 2503 if (p->sched_class->find_lock_rq) 2504 lowest_rq = p->sched_class->find_lock_rq(p, rq); 2505 2506 if (!lowest_rq) 2507 goto out_unlock; 2508 2509 // XXX validate p is still the highest prio task 2510 if (task_rq(p) == rq) { 2511 deactivate_task(rq, p, 0); 2512 set_task_cpu(p, lowest_rq->cpu); 2513 activate_task(lowest_rq, p, 0); 2514 resched_curr(lowest_rq); 2515 } 2516 2517 double_unlock_balance(rq, lowest_rq); 2518 2519 out_unlock: 2520 rq->push_busy = false; 2521 raw_spin_rq_unlock(rq); 2522 raw_spin_unlock_irq(&p->pi_lock); 2523 2524 put_task_struct(p); 2525 return 0; 2526 } 2527 2528 /* 2529 * sched_class::set_cpus_allowed must do the below, but is not required to 2530 * actually call this function. 2531 */ 2532 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags) 2533 { 2534 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) { 2535 p->cpus_ptr = new_mask; 2536 return; 2537 } 2538 2539 cpumask_copy(&p->cpus_mask, new_mask); 2540 p->nr_cpus_allowed = cpumask_weight(new_mask); 2541 } 2542 2543 static void 2544 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags) 2545 { 2546 struct rq *rq = task_rq(p); 2547 bool queued, running; 2548 2549 /* 2550 * This here violates the locking rules for affinity, since we're only 2551 * supposed to change these variables while holding both rq->lock and 2552 * p->pi_lock. 2553 * 2554 * HOWEVER, it magically works, because ttwu() is the only code that 2555 * accesses these variables under p->pi_lock and only does so after 2556 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule() 2557 * before finish_task(). 2558 * 2559 * XXX do further audits, this smells like something putrid. 2560 */ 2561 if (flags & SCA_MIGRATE_DISABLE) 2562 SCHED_WARN_ON(!p->on_cpu); 2563 else 2564 lockdep_assert_held(&p->pi_lock); 2565 2566 queued = task_on_rq_queued(p); 2567 running = task_current(rq, p); 2568 2569 if (queued) { 2570 /* 2571 * Because __kthread_bind() calls this on blocked tasks without 2572 * holding rq->lock. 2573 */ 2574 lockdep_assert_rq_held(rq); 2575 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK); 2576 } 2577 if (running) 2578 put_prev_task(rq, p); 2579 2580 p->sched_class->set_cpus_allowed(p, new_mask, flags); 2581 2582 if (queued) 2583 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); 2584 if (running) 2585 set_next_task(rq, p); 2586 } 2587 2588 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 2589 { 2590 __do_set_cpus_allowed(p, new_mask, 0); 2591 } 2592 2593 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src, 2594 int node) 2595 { 2596 if (!src->user_cpus_ptr) 2597 return 0; 2598 2599 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node); 2600 if (!dst->user_cpus_ptr) 2601 return -ENOMEM; 2602 2603 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr); 2604 return 0; 2605 } 2606 2607 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p) 2608 { 2609 struct cpumask *user_mask = NULL; 2610 2611 swap(p->user_cpus_ptr, user_mask); 2612 2613 return user_mask; 2614 } 2615 2616 void release_user_cpus_ptr(struct task_struct *p) 2617 { 2618 kfree(clear_user_cpus_ptr(p)); 2619 } 2620 2621 /* 2622 * This function is wildly self concurrent; here be dragons. 2623 * 2624 * 2625 * When given a valid mask, __set_cpus_allowed_ptr() must block until the 2626 * designated task is enqueued on an allowed CPU. If that task is currently 2627 * running, we have to kick it out using the CPU stopper. 2628 * 2629 * Migrate-Disable comes along and tramples all over our nice sandcastle. 2630 * Consider: 2631 * 2632 * Initial conditions: P0->cpus_mask = [0, 1] 2633 * 2634 * P0@CPU0 P1 2635 * 2636 * migrate_disable(); 2637 * <preempted> 2638 * set_cpus_allowed_ptr(P0, [1]); 2639 * 2640 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes 2641 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region). 2642 * This means we need the following scheme: 2643 * 2644 * P0@CPU0 P1 2645 * 2646 * migrate_disable(); 2647 * <preempted> 2648 * set_cpus_allowed_ptr(P0, [1]); 2649 * <blocks> 2650 * <resumes> 2651 * migrate_enable(); 2652 * __set_cpus_allowed_ptr(); 2653 * <wakes local stopper> 2654 * `--> <woken on migration completion> 2655 * 2656 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple 2657 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any 2658 * task p are serialized by p->pi_lock, which we can leverage: the one that 2659 * should come into effect at the end of the Migrate-Disable region is the last 2660 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask), 2661 * but we still need to properly signal those waiting tasks at the appropriate 2662 * moment. 2663 * 2664 * This is implemented using struct set_affinity_pending. The first 2665 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will 2666 * setup an instance of that struct and install it on the targeted task_struct. 2667 * Any and all further callers will reuse that instance. Those then wait for 2668 * a completion signaled at the tail of the CPU stopper callback (1), triggered 2669 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()). 2670 * 2671 * 2672 * (1) In the cases covered above. There is one more where the completion is 2673 * signaled within affine_move_task() itself: when a subsequent affinity request 2674 * occurs after the stopper bailed out due to the targeted task still being 2675 * Migrate-Disable. Consider: 2676 * 2677 * Initial conditions: P0->cpus_mask = [0, 1] 2678 * 2679 * CPU0 P1 P2 2680 * <P0> 2681 * migrate_disable(); 2682 * <preempted> 2683 * set_cpus_allowed_ptr(P0, [1]); 2684 * <blocks> 2685 * <migration/0> 2686 * migration_cpu_stop() 2687 * is_migration_disabled() 2688 * <bails> 2689 * set_cpus_allowed_ptr(P0, [0, 1]); 2690 * <signal completion> 2691 * <awakes> 2692 * 2693 * Note that the above is safe vs a concurrent migrate_enable(), as any 2694 * pending affinity completion is preceded by an uninstallation of 2695 * p->migration_pending done with p->pi_lock held. 2696 */ 2697 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf, 2698 int dest_cpu, unsigned int flags) 2699 { 2700 struct set_affinity_pending my_pending = { }, *pending = NULL; 2701 bool stop_pending, complete = false; 2702 2703 /* Can the task run on the task's current CPU? If so, we're done */ 2704 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) { 2705 struct task_struct *push_task = NULL; 2706 2707 if ((flags & SCA_MIGRATE_ENABLE) && 2708 (p->migration_flags & MDF_PUSH) && !rq->push_busy) { 2709 rq->push_busy = true; 2710 push_task = get_task_struct(p); 2711 } 2712 2713 /* 2714 * If there are pending waiters, but no pending stop_work, 2715 * then complete now. 2716 */ 2717 pending = p->migration_pending; 2718 if (pending && !pending->stop_pending) { 2719 p->migration_pending = NULL; 2720 complete = true; 2721 } 2722 2723 task_rq_unlock(rq, p, rf); 2724 2725 if (push_task) { 2726 stop_one_cpu_nowait(rq->cpu, push_cpu_stop, 2727 p, &rq->push_work); 2728 } 2729 2730 if (complete) 2731 complete_all(&pending->done); 2732 2733 return 0; 2734 } 2735 2736 if (!(flags & SCA_MIGRATE_ENABLE)) { 2737 /* serialized by p->pi_lock */ 2738 if (!p->migration_pending) { 2739 /* Install the request */ 2740 refcount_set(&my_pending.refs, 1); 2741 init_completion(&my_pending.done); 2742 my_pending.arg = (struct migration_arg) { 2743 .task = p, 2744 .dest_cpu = dest_cpu, 2745 .pending = &my_pending, 2746 }; 2747 2748 p->migration_pending = &my_pending; 2749 } else { 2750 pending = p->migration_pending; 2751 refcount_inc(&pending->refs); 2752 /* 2753 * Affinity has changed, but we've already installed a 2754 * pending. migration_cpu_stop() *must* see this, else 2755 * we risk a completion of the pending despite having a 2756 * task on a disallowed CPU. 2757 * 2758 * Serialized by p->pi_lock, so this is safe. 2759 */ 2760 pending->arg.dest_cpu = dest_cpu; 2761 } 2762 } 2763 pending = p->migration_pending; 2764 /* 2765 * - !MIGRATE_ENABLE: 2766 * we'll have installed a pending if there wasn't one already. 2767 * 2768 * - MIGRATE_ENABLE: 2769 * we're here because the current CPU isn't matching anymore, 2770 * the only way that can happen is because of a concurrent 2771 * set_cpus_allowed_ptr() call, which should then still be 2772 * pending completion. 2773 * 2774 * Either way, we really should have a @pending here. 2775 */ 2776 if (WARN_ON_ONCE(!pending)) { 2777 task_rq_unlock(rq, p, rf); 2778 return -EINVAL; 2779 } 2780 2781 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) { 2782 /* 2783 * MIGRATE_ENABLE gets here because 'p == current', but for 2784 * anything else we cannot do is_migration_disabled(), punt 2785 * and have the stopper function handle it all race-free. 2786 */ 2787 stop_pending = pending->stop_pending; 2788 if (!stop_pending) 2789 pending->stop_pending = true; 2790 2791 if (flags & SCA_MIGRATE_ENABLE) 2792 p->migration_flags &= ~MDF_PUSH; 2793 2794 task_rq_unlock(rq, p, rf); 2795 2796 if (!stop_pending) { 2797 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop, 2798 &pending->arg, &pending->stop_work); 2799 } 2800 2801 if (flags & SCA_MIGRATE_ENABLE) 2802 return 0; 2803 } else { 2804 2805 if (!is_migration_disabled(p)) { 2806 if (task_on_rq_queued(p)) 2807 rq = move_queued_task(rq, rf, p, dest_cpu); 2808 2809 if (!pending->stop_pending) { 2810 p->migration_pending = NULL; 2811 complete = true; 2812 } 2813 } 2814 task_rq_unlock(rq, p, rf); 2815 2816 if (complete) 2817 complete_all(&pending->done); 2818 } 2819 2820 wait_for_completion(&pending->done); 2821 2822 if (refcount_dec_and_test(&pending->refs)) 2823 wake_up_var(&pending->refs); /* No UaF, just an address */ 2824 2825 /* 2826 * Block the original owner of &pending until all subsequent callers 2827 * have seen the completion and decremented the refcount 2828 */ 2829 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs)); 2830 2831 /* ARGH */ 2832 WARN_ON_ONCE(my_pending.stop_pending); 2833 2834 return 0; 2835 } 2836 2837 /* 2838 * Called with both p->pi_lock and rq->lock held; drops both before returning. 2839 */ 2840 static int __set_cpus_allowed_ptr_locked(struct task_struct *p, 2841 const struct cpumask *new_mask, 2842 u32 flags, 2843 struct rq *rq, 2844 struct rq_flags *rf) 2845 __releases(rq->lock) 2846 __releases(p->pi_lock) 2847 { 2848 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p); 2849 const struct cpumask *cpu_valid_mask = cpu_active_mask; 2850 bool kthread = p->flags & PF_KTHREAD; 2851 struct cpumask *user_mask = NULL; 2852 unsigned int dest_cpu; 2853 int ret = 0; 2854 2855 update_rq_clock(rq); 2856 2857 if (kthread || is_migration_disabled(p)) { 2858 /* 2859 * Kernel threads are allowed on online && !active CPUs, 2860 * however, during cpu-hot-unplug, even these might get pushed 2861 * away if not KTHREAD_IS_PER_CPU. 2862 * 2863 * Specifically, migration_disabled() tasks must not fail the 2864 * cpumask_any_and_distribute() pick below, esp. so on 2865 * SCA_MIGRATE_ENABLE, otherwise we'll not call 2866 * set_cpus_allowed_common() and actually reset p->cpus_ptr. 2867 */ 2868 cpu_valid_mask = cpu_online_mask; 2869 } 2870 2871 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) { 2872 ret = -EINVAL; 2873 goto out; 2874 } 2875 2876 /* 2877 * Must re-check here, to close a race against __kthread_bind(), 2878 * sched_setaffinity() is not guaranteed to observe the flag. 2879 */ 2880 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) { 2881 ret = -EINVAL; 2882 goto out; 2883 } 2884 2885 if (!(flags & SCA_MIGRATE_ENABLE)) { 2886 if (cpumask_equal(&p->cpus_mask, new_mask)) 2887 goto out; 2888 2889 if (WARN_ON_ONCE(p == current && 2890 is_migration_disabled(p) && 2891 !cpumask_test_cpu(task_cpu(p), new_mask))) { 2892 ret = -EBUSY; 2893 goto out; 2894 } 2895 } 2896 2897 /* 2898 * Picking a ~random cpu helps in cases where we are changing affinity 2899 * for groups of tasks (ie. cpuset), so that load balancing is not 2900 * immediately required to distribute the tasks within their new mask. 2901 */ 2902 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask); 2903 if (dest_cpu >= nr_cpu_ids) { 2904 ret = -EINVAL; 2905 goto out; 2906 } 2907 2908 __do_set_cpus_allowed(p, new_mask, flags); 2909 2910 if (flags & SCA_USER) 2911 user_mask = clear_user_cpus_ptr(p); 2912 2913 ret = affine_move_task(rq, p, rf, dest_cpu, flags); 2914 2915 kfree(user_mask); 2916 2917 return ret; 2918 2919 out: 2920 task_rq_unlock(rq, p, rf); 2921 2922 return ret; 2923 } 2924 2925 /* 2926 * Change a given task's CPU affinity. Migrate the thread to a 2927 * proper CPU and schedule it away if the CPU it's executing on 2928 * is removed from the allowed bitmask. 2929 * 2930 * NOTE: the caller must have a valid reference to the task, the 2931 * task must not exit() & deallocate itself prematurely. The 2932 * call is not atomic; no spinlocks may be held. 2933 */ 2934 static int __set_cpus_allowed_ptr(struct task_struct *p, 2935 const struct cpumask *new_mask, u32 flags) 2936 { 2937 struct rq_flags rf; 2938 struct rq *rq; 2939 2940 rq = task_rq_lock(p, &rf); 2941 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf); 2942 } 2943 2944 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 2945 { 2946 return __set_cpus_allowed_ptr(p, new_mask, 0); 2947 } 2948 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 2949 2950 /* 2951 * Change a given task's CPU affinity to the intersection of its current 2952 * affinity mask and @subset_mask, writing the resulting mask to @new_mask 2953 * and pointing @p->user_cpus_ptr to a copy of the old mask. 2954 * If the resulting mask is empty, leave the affinity unchanged and return 2955 * -EINVAL. 2956 */ 2957 static int restrict_cpus_allowed_ptr(struct task_struct *p, 2958 struct cpumask *new_mask, 2959 const struct cpumask *subset_mask) 2960 { 2961 struct cpumask *user_mask = NULL; 2962 struct rq_flags rf; 2963 struct rq *rq; 2964 int err; 2965 2966 if (!p->user_cpus_ptr) { 2967 user_mask = kmalloc(cpumask_size(), GFP_KERNEL); 2968 if (!user_mask) 2969 return -ENOMEM; 2970 } 2971 2972 rq = task_rq_lock(p, &rf); 2973 2974 /* 2975 * Forcefully restricting the affinity of a deadline task is 2976 * likely to cause problems, so fail and noisily override the 2977 * mask entirely. 2978 */ 2979 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { 2980 err = -EPERM; 2981 goto err_unlock; 2982 } 2983 2984 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) { 2985 err = -EINVAL; 2986 goto err_unlock; 2987 } 2988 2989 /* 2990 * We're about to butcher the task affinity, so keep track of what 2991 * the user asked for in case we're able to restore it later on. 2992 */ 2993 if (user_mask) { 2994 cpumask_copy(user_mask, p->cpus_ptr); 2995 p->user_cpus_ptr = user_mask; 2996 } 2997 2998 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf); 2999 3000 err_unlock: 3001 task_rq_unlock(rq, p, &rf); 3002 kfree(user_mask); 3003 return err; 3004 } 3005 3006 /* 3007 * Restrict the CPU affinity of task @p so that it is a subset of 3008 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the 3009 * old affinity mask. If the resulting mask is empty, we warn and walk 3010 * up the cpuset hierarchy until we find a suitable mask. 3011 */ 3012 void force_compatible_cpus_allowed_ptr(struct task_struct *p) 3013 { 3014 cpumask_var_t new_mask; 3015 const struct cpumask *override_mask = task_cpu_possible_mask(p); 3016 3017 alloc_cpumask_var(&new_mask, GFP_KERNEL); 3018 3019 /* 3020 * __migrate_task() can fail silently in the face of concurrent 3021 * offlining of the chosen destination CPU, so take the hotplug 3022 * lock to ensure that the migration succeeds. 3023 */ 3024 cpus_read_lock(); 3025 if (!cpumask_available(new_mask)) 3026 goto out_set_mask; 3027 3028 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask)) 3029 goto out_free_mask; 3030 3031 /* 3032 * We failed to find a valid subset of the affinity mask for the 3033 * task, so override it based on its cpuset hierarchy. 3034 */ 3035 cpuset_cpus_allowed(p, new_mask); 3036 override_mask = new_mask; 3037 3038 out_set_mask: 3039 if (printk_ratelimit()) { 3040 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n", 3041 task_pid_nr(p), p->comm, 3042 cpumask_pr_args(override_mask)); 3043 } 3044 3045 WARN_ON(set_cpus_allowed_ptr(p, override_mask)); 3046 out_free_mask: 3047 cpus_read_unlock(); 3048 free_cpumask_var(new_mask); 3049 } 3050 3051 static int 3052 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask); 3053 3054 /* 3055 * Restore the affinity of a task @p which was previously restricted by a 3056 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free) 3057 * @p->user_cpus_ptr. 3058 * 3059 * It is the caller's responsibility to serialise this with any calls to 3060 * force_compatible_cpus_allowed_ptr(@p). 3061 */ 3062 void relax_compatible_cpus_allowed_ptr(struct task_struct *p) 3063 { 3064 struct cpumask *user_mask = p->user_cpus_ptr; 3065 unsigned long flags; 3066 3067 /* 3068 * Try to restore the old affinity mask. If this fails, then 3069 * we free the mask explicitly to avoid it being inherited across 3070 * a subsequent fork(). 3071 */ 3072 if (!user_mask || !__sched_setaffinity(p, user_mask)) 3073 return; 3074 3075 raw_spin_lock_irqsave(&p->pi_lock, flags); 3076 user_mask = clear_user_cpus_ptr(p); 3077 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3078 3079 kfree(user_mask); 3080 } 3081 3082 void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 3083 { 3084 #ifdef CONFIG_SCHED_DEBUG 3085 unsigned int state = READ_ONCE(p->__state); 3086 3087 /* 3088 * We should never call set_task_cpu() on a blocked task, 3089 * ttwu() will sort out the placement. 3090 */ 3091 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq); 3092 3093 /* 3094 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING, 3095 * because schedstat_wait_{start,end} rebase migrating task's wait_start 3096 * time relying on p->on_rq. 3097 */ 3098 WARN_ON_ONCE(state == TASK_RUNNING && 3099 p->sched_class == &fair_sched_class && 3100 (p->on_rq && !task_on_rq_migrating(p))); 3101 3102 #ifdef CONFIG_LOCKDEP 3103 /* 3104 * The caller should hold either p->pi_lock or rq->lock, when changing 3105 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 3106 * 3107 * sched_move_task() holds both and thus holding either pins the cgroup, 3108 * see task_group(). 3109 * 3110 * Furthermore, all task_rq users should acquire both locks, see 3111 * task_rq_lock(). 3112 */ 3113 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 3114 lockdep_is_held(__rq_lockp(task_rq(p))))); 3115 #endif 3116 /* 3117 * Clearly, migrating tasks to offline CPUs is a fairly daft thing. 3118 */ 3119 WARN_ON_ONCE(!cpu_online(new_cpu)); 3120 3121 WARN_ON_ONCE(is_migration_disabled(p)); 3122 #endif 3123 3124 trace_sched_migrate_task(p, new_cpu); 3125 3126 if (task_cpu(p) != new_cpu) { 3127 if (p->sched_class->migrate_task_rq) 3128 p->sched_class->migrate_task_rq(p, new_cpu); 3129 p->se.nr_migrations++; 3130 rseq_migrate(p); 3131 perf_event_task_migrate(p); 3132 } 3133 3134 __set_task_cpu(p, new_cpu); 3135 } 3136 3137 #ifdef CONFIG_NUMA_BALANCING 3138 static void __migrate_swap_task(struct task_struct *p, int cpu) 3139 { 3140 if (task_on_rq_queued(p)) { 3141 struct rq *src_rq, *dst_rq; 3142 struct rq_flags srf, drf; 3143 3144 src_rq = task_rq(p); 3145 dst_rq = cpu_rq(cpu); 3146 3147 rq_pin_lock(src_rq, &srf); 3148 rq_pin_lock(dst_rq, &drf); 3149 3150 deactivate_task(src_rq, p, 0); 3151 set_task_cpu(p, cpu); 3152 activate_task(dst_rq, p, 0); 3153 check_preempt_curr(dst_rq, p, 0); 3154 3155 rq_unpin_lock(dst_rq, &drf); 3156 rq_unpin_lock(src_rq, &srf); 3157 3158 } else { 3159 /* 3160 * Task isn't running anymore; make it appear like we migrated 3161 * it before it went to sleep. This means on wakeup we make the 3162 * previous CPU our target instead of where it really is. 3163 */ 3164 p->wake_cpu = cpu; 3165 } 3166 } 3167 3168 struct migration_swap_arg { 3169 struct task_struct *src_task, *dst_task; 3170 int src_cpu, dst_cpu; 3171 }; 3172 3173 static int migrate_swap_stop(void *data) 3174 { 3175 struct migration_swap_arg *arg = data; 3176 struct rq *src_rq, *dst_rq; 3177 int ret = -EAGAIN; 3178 3179 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu)) 3180 return -EAGAIN; 3181 3182 src_rq = cpu_rq(arg->src_cpu); 3183 dst_rq = cpu_rq(arg->dst_cpu); 3184 3185 double_raw_lock(&arg->src_task->pi_lock, 3186 &arg->dst_task->pi_lock); 3187 double_rq_lock(src_rq, dst_rq); 3188 3189 if (task_cpu(arg->dst_task) != arg->dst_cpu) 3190 goto unlock; 3191 3192 if (task_cpu(arg->src_task) != arg->src_cpu) 3193 goto unlock; 3194 3195 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr)) 3196 goto unlock; 3197 3198 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr)) 3199 goto unlock; 3200 3201 __migrate_swap_task(arg->src_task, arg->dst_cpu); 3202 __migrate_swap_task(arg->dst_task, arg->src_cpu); 3203 3204 ret = 0; 3205 3206 unlock: 3207 double_rq_unlock(src_rq, dst_rq); 3208 raw_spin_unlock(&arg->dst_task->pi_lock); 3209 raw_spin_unlock(&arg->src_task->pi_lock); 3210 3211 return ret; 3212 } 3213 3214 /* 3215 * Cross migrate two tasks 3216 */ 3217 int migrate_swap(struct task_struct *cur, struct task_struct *p, 3218 int target_cpu, int curr_cpu) 3219 { 3220 struct migration_swap_arg arg; 3221 int ret = -EINVAL; 3222 3223 arg = (struct migration_swap_arg){ 3224 .src_task = cur, 3225 .src_cpu = curr_cpu, 3226 .dst_task = p, 3227 .dst_cpu = target_cpu, 3228 }; 3229 3230 if (arg.src_cpu == arg.dst_cpu) 3231 goto out; 3232 3233 /* 3234 * These three tests are all lockless; this is OK since all of them 3235 * will be re-checked with proper locks held further down the line. 3236 */ 3237 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 3238 goto out; 3239 3240 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr)) 3241 goto out; 3242 3243 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr)) 3244 goto out; 3245 3246 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 3247 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 3248 3249 out: 3250 return ret; 3251 } 3252 #endif /* CONFIG_NUMA_BALANCING */ 3253 3254 /* 3255 * wait_task_inactive - wait for a thread to unschedule. 3256 * 3257 * If @match_state is nonzero, it's the @p->state value just checked and 3258 * not expected to change. If it changes, i.e. @p might have woken up, 3259 * then return zero. When we succeed in waiting for @p to be off its CPU, 3260 * we return a positive number (its total switch count). If a second call 3261 * a short while later returns the same number, the caller can be sure that 3262 * @p has remained unscheduled the whole time. 3263 * 3264 * The caller must ensure that the task *will* unschedule sometime soon, 3265 * else this function might spin for a *long* time. This function can't 3266 * be called with interrupts off, or it may introduce deadlock with 3267 * smp_call_function() if an IPI is sent by the same process we are 3268 * waiting to become inactive. 3269 */ 3270 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state) 3271 { 3272 int running, queued; 3273 struct rq_flags rf; 3274 unsigned long ncsw; 3275 struct rq *rq; 3276 3277 for (;;) { 3278 /* 3279 * We do the initial early heuristics without holding 3280 * any task-queue locks at all. We'll only try to get 3281 * the runqueue lock when things look like they will 3282 * work out! 3283 */ 3284 rq = task_rq(p); 3285 3286 /* 3287 * If the task is actively running on another CPU 3288 * still, just relax and busy-wait without holding 3289 * any locks. 3290 * 3291 * NOTE! Since we don't hold any locks, it's not 3292 * even sure that "rq" stays as the right runqueue! 3293 * But we don't care, since "task_running()" will 3294 * return false if the runqueue has changed and p 3295 * is actually now running somewhere else! 3296 */ 3297 while (task_running(rq, p)) { 3298 if (match_state && unlikely(READ_ONCE(p->__state) != match_state)) 3299 return 0; 3300 cpu_relax(); 3301 } 3302 3303 /* 3304 * Ok, time to look more closely! We need the rq 3305 * lock now, to be *sure*. If we're wrong, we'll 3306 * just go back and repeat. 3307 */ 3308 rq = task_rq_lock(p, &rf); 3309 trace_sched_wait_task(p); 3310 running = task_running(rq, p); 3311 queued = task_on_rq_queued(p); 3312 ncsw = 0; 3313 if (!match_state || READ_ONCE(p->__state) == match_state) 3314 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 3315 task_rq_unlock(rq, p, &rf); 3316 3317 /* 3318 * If it changed from the expected state, bail out now. 3319 */ 3320 if (unlikely(!ncsw)) 3321 break; 3322 3323 /* 3324 * Was it really running after all now that we 3325 * checked with the proper locks actually held? 3326 * 3327 * Oops. Go back and try again.. 3328 */ 3329 if (unlikely(running)) { 3330 cpu_relax(); 3331 continue; 3332 } 3333 3334 /* 3335 * It's not enough that it's not actively running, 3336 * it must be off the runqueue _entirely_, and not 3337 * preempted! 3338 * 3339 * So if it was still runnable (but just not actively 3340 * running right now), it's preempted, and we should 3341 * yield - it could be a while. 3342 */ 3343 if (unlikely(queued)) { 3344 ktime_t to = NSEC_PER_SEC / HZ; 3345 3346 set_current_state(TASK_UNINTERRUPTIBLE); 3347 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD); 3348 continue; 3349 } 3350 3351 /* 3352 * Ahh, all good. It wasn't running, and it wasn't 3353 * runnable, which means that it will never become 3354 * running in the future either. We're all done! 3355 */ 3356 break; 3357 } 3358 3359 return ncsw; 3360 } 3361 3362 /*** 3363 * kick_process - kick a running thread to enter/exit the kernel 3364 * @p: the to-be-kicked thread 3365 * 3366 * Cause a process which is running on another CPU to enter 3367 * kernel-mode, without any delay. (to get signals handled.) 3368 * 3369 * NOTE: this function doesn't have to take the runqueue lock, 3370 * because all it wants to ensure is that the remote task enters 3371 * the kernel. If the IPI races and the task has been migrated 3372 * to another CPU then no harm is done and the purpose has been 3373 * achieved as well. 3374 */ 3375 void kick_process(struct task_struct *p) 3376 { 3377 int cpu; 3378 3379 preempt_disable(); 3380 cpu = task_cpu(p); 3381 if ((cpu != smp_processor_id()) && task_curr(p)) 3382 smp_send_reschedule(cpu); 3383 preempt_enable(); 3384 } 3385 EXPORT_SYMBOL_GPL(kick_process); 3386 3387 /* 3388 * ->cpus_ptr is protected by both rq->lock and p->pi_lock 3389 * 3390 * A few notes on cpu_active vs cpu_online: 3391 * 3392 * - cpu_active must be a subset of cpu_online 3393 * 3394 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU, 3395 * see __set_cpus_allowed_ptr(). At this point the newly online 3396 * CPU isn't yet part of the sched domains, and balancing will not 3397 * see it. 3398 * 3399 * - on CPU-down we clear cpu_active() to mask the sched domains and 3400 * avoid the load balancer to place new tasks on the to be removed 3401 * CPU. Existing tasks will remain running there and will be taken 3402 * off. 3403 * 3404 * This means that fallback selection must not select !active CPUs. 3405 * And can assume that any active CPU must be online. Conversely 3406 * select_task_rq() below may allow selection of !active CPUs in order 3407 * to satisfy the above rules. 3408 */ 3409 static int select_fallback_rq(int cpu, struct task_struct *p) 3410 { 3411 int nid = cpu_to_node(cpu); 3412 const struct cpumask *nodemask = NULL; 3413 enum { cpuset, possible, fail } state = cpuset; 3414 int dest_cpu; 3415 3416 /* 3417 * If the node that the CPU is on has been offlined, cpu_to_node() 3418 * will return -1. There is no CPU on the node, and we should 3419 * select the CPU on the other node. 3420 */ 3421 if (nid != -1) { 3422 nodemask = cpumask_of_node(nid); 3423 3424 /* Look for allowed, online CPU in same node. */ 3425 for_each_cpu(dest_cpu, nodemask) { 3426 if (is_cpu_allowed(p, dest_cpu)) 3427 return dest_cpu; 3428 } 3429 } 3430 3431 for (;;) { 3432 /* Any allowed, online CPU? */ 3433 for_each_cpu(dest_cpu, p->cpus_ptr) { 3434 if (!is_cpu_allowed(p, dest_cpu)) 3435 continue; 3436 3437 goto out; 3438 } 3439 3440 /* No more Mr. Nice Guy. */ 3441 switch (state) { 3442 case cpuset: 3443 if (cpuset_cpus_allowed_fallback(p)) { 3444 state = possible; 3445 break; 3446 } 3447 fallthrough; 3448 case possible: 3449 /* 3450 * XXX When called from select_task_rq() we only 3451 * hold p->pi_lock and again violate locking order. 3452 * 3453 * More yuck to audit. 3454 */ 3455 do_set_cpus_allowed(p, task_cpu_possible_mask(p)); 3456 state = fail; 3457 break; 3458 case fail: 3459 BUG(); 3460 break; 3461 } 3462 } 3463 3464 out: 3465 if (state != cpuset) { 3466 /* 3467 * Don't tell them about moving exiting tasks or 3468 * kernel threads (both mm NULL), since they never 3469 * leave kernel. 3470 */ 3471 if (p->mm && printk_ratelimit()) { 3472 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 3473 task_pid_nr(p), p->comm, cpu); 3474 } 3475 } 3476 3477 return dest_cpu; 3478 } 3479 3480 /* 3481 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable. 3482 */ 3483 static inline 3484 int select_task_rq(struct task_struct *p, int cpu, int wake_flags) 3485 { 3486 lockdep_assert_held(&p->pi_lock); 3487 3488 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p)) 3489 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags); 3490 else 3491 cpu = cpumask_any(p->cpus_ptr); 3492 3493 /* 3494 * In order not to call set_task_cpu() on a blocking task we need 3495 * to rely on ttwu() to place the task on a valid ->cpus_ptr 3496 * CPU. 3497 * 3498 * Since this is common to all placement strategies, this lives here. 3499 * 3500 * [ this allows ->select_task() to simply return task_cpu(p) and 3501 * not worry about this generic constraint ] 3502 */ 3503 if (unlikely(!is_cpu_allowed(p, cpu))) 3504 cpu = select_fallback_rq(task_cpu(p), p); 3505 3506 return cpu; 3507 } 3508 3509 void sched_set_stop_task(int cpu, struct task_struct *stop) 3510 { 3511 static struct lock_class_key stop_pi_lock; 3512 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 3513 struct task_struct *old_stop = cpu_rq(cpu)->stop; 3514 3515 if (stop) { 3516 /* 3517 * Make it appear like a SCHED_FIFO task, its something 3518 * userspace knows about and won't get confused about. 3519 * 3520 * Also, it will make PI more or less work without too 3521 * much confusion -- but then, stop work should not 3522 * rely on PI working anyway. 3523 */ 3524 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 3525 3526 stop->sched_class = &stop_sched_class; 3527 3528 /* 3529 * The PI code calls rt_mutex_setprio() with ->pi_lock held to 3530 * adjust the effective priority of a task. As a result, 3531 * rt_mutex_setprio() can trigger (RT) balancing operations, 3532 * which can then trigger wakeups of the stop thread to push 3533 * around the current task. 3534 * 3535 * The stop task itself will never be part of the PI-chain, it 3536 * never blocks, therefore that ->pi_lock recursion is safe. 3537 * Tell lockdep about this by placing the stop->pi_lock in its 3538 * own class. 3539 */ 3540 lockdep_set_class(&stop->pi_lock, &stop_pi_lock); 3541 } 3542 3543 cpu_rq(cpu)->stop = stop; 3544 3545 if (old_stop) { 3546 /* 3547 * Reset it back to a normal scheduling class so that 3548 * it can die in pieces. 3549 */ 3550 old_stop->sched_class = &rt_sched_class; 3551 } 3552 } 3553 3554 #else /* CONFIG_SMP */ 3555 3556 static inline int __set_cpus_allowed_ptr(struct task_struct *p, 3557 const struct cpumask *new_mask, 3558 u32 flags) 3559 { 3560 return set_cpus_allowed_ptr(p, new_mask); 3561 } 3562 3563 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { } 3564 3565 static inline bool rq_has_pinned_tasks(struct rq *rq) 3566 { 3567 return false; 3568 } 3569 3570 #endif /* !CONFIG_SMP */ 3571 3572 static void 3573 ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 3574 { 3575 struct rq *rq; 3576 3577 if (!schedstat_enabled()) 3578 return; 3579 3580 rq = this_rq(); 3581 3582 #ifdef CONFIG_SMP 3583 if (cpu == rq->cpu) { 3584 __schedstat_inc(rq->ttwu_local); 3585 __schedstat_inc(p->stats.nr_wakeups_local); 3586 } else { 3587 struct sched_domain *sd; 3588 3589 __schedstat_inc(p->stats.nr_wakeups_remote); 3590 rcu_read_lock(); 3591 for_each_domain(rq->cpu, sd) { 3592 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 3593 __schedstat_inc(sd->ttwu_wake_remote); 3594 break; 3595 } 3596 } 3597 rcu_read_unlock(); 3598 } 3599 3600 if (wake_flags & WF_MIGRATED) 3601 __schedstat_inc(p->stats.nr_wakeups_migrate); 3602 #endif /* CONFIG_SMP */ 3603 3604 __schedstat_inc(rq->ttwu_count); 3605 __schedstat_inc(p->stats.nr_wakeups); 3606 3607 if (wake_flags & WF_SYNC) 3608 __schedstat_inc(p->stats.nr_wakeups_sync); 3609 } 3610 3611 /* 3612 * Mark the task runnable and perform wakeup-preemption. 3613 */ 3614 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags, 3615 struct rq_flags *rf) 3616 { 3617 check_preempt_curr(rq, p, wake_flags); 3618 WRITE_ONCE(p->__state, TASK_RUNNING); 3619 trace_sched_wakeup(p); 3620 3621 #ifdef CONFIG_SMP 3622 if (p->sched_class->task_woken) { 3623 /* 3624 * Our task @p is fully woken up and running; so it's safe to 3625 * drop the rq->lock, hereafter rq is only used for statistics. 3626 */ 3627 rq_unpin_lock(rq, rf); 3628 p->sched_class->task_woken(rq, p); 3629 rq_repin_lock(rq, rf); 3630 } 3631 3632 if (rq->idle_stamp) { 3633 u64 delta = rq_clock(rq) - rq->idle_stamp; 3634 u64 max = 2*rq->max_idle_balance_cost; 3635 3636 update_avg(&rq->avg_idle, delta); 3637 3638 if (rq->avg_idle > max) 3639 rq->avg_idle = max; 3640 3641 rq->wake_stamp = jiffies; 3642 rq->wake_avg_idle = rq->avg_idle / 2; 3643 3644 rq->idle_stamp = 0; 3645 } 3646 #endif 3647 } 3648 3649 static void 3650 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags, 3651 struct rq_flags *rf) 3652 { 3653 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK; 3654 3655 lockdep_assert_rq_held(rq); 3656 3657 if (p->sched_contributes_to_load) 3658 rq->nr_uninterruptible--; 3659 3660 #ifdef CONFIG_SMP 3661 if (wake_flags & WF_MIGRATED) 3662 en_flags |= ENQUEUE_MIGRATED; 3663 else 3664 #endif 3665 if (p->in_iowait) { 3666 delayacct_blkio_end(p); 3667 atomic_dec(&task_rq(p)->nr_iowait); 3668 } 3669 3670 activate_task(rq, p, en_flags); 3671 ttwu_do_wakeup(rq, p, wake_flags, rf); 3672 } 3673 3674 /* 3675 * Consider @p being inside a wait loop: 3676 * 3677 * for (;;) { 3678 * set_current_state(TASK_UNINTERRUPTIBLE); 3679 * 3680 * if (CONDITION) 3681 * break; 3682 * 3683 * schedule(); 3684 * } 3685 * __set_current_state(TASK_RUNNING); 3686 * 3687 * between set_current_state() and schedule(). In this case @p is still 3688 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in 3689 * an atomic manner. 3690 * 3691 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq 3692 * then schedule() must still happen and p->state can be changed to 3693 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we 3694 * need to do a full wakeup with enqueue. 3695 * 3696 * Returns: %true when the wakeup is done, 3697 * %false otherwise. 3698 */ 3699 static int ttwu_runnable(struct task_struct *p, int wake_flags) 3700 { 3701 struct rq_flags rf; 3702 struct rq *rq; 3703 int ret = 0; 3704 3705 rq = __task_rq_lock(p, &rf); 3706 if (task_on_rq_queued(p)) { 3707 /* check_preempt_curr() may use rq clock */ 3708 update_rq_clock(rq); 3709 ttwu_do_wakeup(rq, p, wake_flags, &rf); 3710 ret = 1; 3711 } 3712 __task_rq_unlock(rq, &rf); 3713 3714 return ret; 3715 } 3716 3717 #ifdef CONFIG_SMP 3718 void sched_ttwu_pending(void *arg) 3719 { 3720 struct llist_node *llist = arg; 3721 struct rq *rq = this_rq(); 3722 struct task_struct *p, *t; 3723 struct rq_flags rf; 3724 3725 if (!llist) 3726 return; 3727 3728 /* 3729 * rq::ttwu_pending racy indication of out-standing wakeups. 3730 * Races such that false-negatives are possible, since they 3731 * are shorter lived that false-positives would be. 3732 */ 3733 WRITE_ONCE(rq->ttwu_pending, 0); 3734 3735 rq_lock_irqsave(rq, &rf); 3736 update_rq_clock(rq); 3737 3738 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) { 3739 if (WARN_ON_ONCE(p->on_cpu)) 3740 smp_cond_load_acquire(&p->on_cpu, !VAL); 3741 3742 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq))) 3743 set_task_cpu(p, cpu_of(rq)); 3744 3745 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf); 3746 } 3747 3748 rq_unlock_irqrestore(rq, &rf); 3749 } 3750 3751 void send_call_function_single_ipi(int cpu) 3752 { 3753 struct rq *rq = cpu_rq(cpu); 3754 3755 if (!set_nr_if_polling(rq->idle)) 3756 arch_send_call_function_single_ipi(cpu); 3757 else 3758 trace_sched_wake_idle_without_ipi(cpu); 3759 } 3760 3761 /* 3762 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if 3763 * necessary. The wakee CPU on receipt of the IPI will queue the task 3764 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost 3765 * of the wakeup instead of the waker. 3766 */ 3767 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) 3768 { 3769 struct rq *rq = cpu_rq(cpu); 3770 3771 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED); 3772 3773 WRITE_ONCE(rq->ttwu_pending, 1); 3774 __smp_call_single_queue(cpu, &p->wake_entry.llist); 3775 } 3776 3777 void wake_up_if_idle(int cpu) 3778 { 3779 struct rq *rq = cpu_rq(cpu); 3780 struct rq_flags rf; 3781 3782 rcu_read_lock(); 3783 3784 if (!is_idle_task(rcu_dereference(rq->curr))) 3785 goto out; 3786 3787 rq_lock_irqsave(rq, &rf); 3788 if (is_idle_task(rq->curr)) 3789 resched_curr(rq); 3790 /* Else CPU is not idle, do nothing here: */ 3791 rq_unlock_irqrestore(rq, &rf); 3792 3793 out: 3794 rcu_read_unlock(); 3795 } 3796 3797 bool cpus_share_cache(int this_cpu, int that_cpu) 3798 { 3799 if (this_cpu == that_cpu) 3800 return true; 3801 3802 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 3803 } 3804 3805 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu) 3806 { 3807 /* 3808 * Do not complicate things with the async wake_list while the CPU is 3809 * in hotplug state. 3810 */ 3811 if (!cpu_active(cpu)) 3812 return false; 3813 3814 /* Ensure the task will still be allowed to run on the CPU. */ 3815 if (!cpumask_test_cpu(cpu, p->cpus_ptr)) 3816 return false; 3817 3818 /* 3819 * If the CPU does not share cache, then queue the task on the 3820 * remote rqs wakelist to avoid accessing remote data. 3821 */ 3822 if (!cpus_share_cache(smp_processor_id(), cpu)) 3823 return true; 3824 3825 if (cpu == smp_processor_id()) 3826 return false; 3827 3828 /* 3829 * If the wakee cpu is idle, or the task is descheduling and the 3830 * only running task on the CPU, then use the wakelist to offload 3831 * the task activation to the idle (or soon-to-be-idle) CPU as 3832 * the current CPU is likely busy. nr_running is checked to 3833 * avoid unnecessary task stacking. 3834 * 3835 * Note that we can only get here with (wakee) p->on_rq=0, 3836 * p->on_cpu can be whatever, we've done the dequeue, so 3837 * the wakee has been accounted out of ->nr_running. 3838 */ 3839 if (!cpu_rq(cpu)->nr_running) 3840 return true; 3841 3842 return false; 3843 } 3844 3845 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) 3846 { 3847 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) { 3848 sched_clock_cpu(cpu); /* Sync clocks across CPUs */ 3849 __ttwu_queue_wakelist(p, cpu, wake_flags); 3850 return true; 3851 } 3852 3853 return false; 3854 } 3855 3856 #else /* !CONFIG_SMP */ 3857 3858 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags) 3859 { 3860 return false; 3861 } 3862 3863 #endif /* CONFIG_SMP */ 3864 3865 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags) 3866 { 3867 struct rq *rq = cpu_rq(cpu); 3868 struct rq_flags rf; 3869 3870 if (ttwu_queue_wakelist(p, cpu, wake_flags)) 3871 return; 3872 3873 rq_lock(rq, &rf); 3874 update_rq_clock(rq); 3875 ttwu_do_activate(rq, p, wake_flags, &rf); 3876 rq_unlock(rq, &rf); 3877 } 3878 3879 /* 3880 * Invoked from try_to_wake_up() to check whether the task can be woken up. 3881 * 3882 * The caller holds p::pi_lock if p != current or has preemption 3883 * disabled when p == current. 3884 * 3885 * The rules of PREEMPT_RT saved_state: 3886 * 3887 * The related locking code always holds p::pi_lock when updating 3888 * p::saved_state, which means the code is fully serialized in both cases. 3889 * 3890 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other 3891 * bits set. This allows to distinguish all wakeup scenarios. 3892 */ 3893 static __always_inline 3894 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success) 3895 { 3896 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) { 3897 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) && 3898 state != TASK_RTLOCK_WAIT); 3899 } 3900 3901 if (READ_ONCE(p->__state) & state) { 3902 *success = 1; 3903 return true; 3904 } 3905 3906 #ifdef CONFIG_PREEMPT_RT 3907 /* 3908 * Saved state preserves the task state across blocking on 3909 * an RT lock. If the state matches, set p::saved_state to 3910 * TASK_RUNNING, but do not wake the task because it waits 3911 * for a lock wakeup. Also indicate success because from 3912 * the regular waker's point of view this has succeeded. 3913 * 3914 * After acquiring the lock the task will restore p::__state 3915 * from p::saved_state which ensures that the regular 3916 * wakeup is not lost. The restore will also set 3917 * p::saved_state to TASK_RUNNING so any further tests will 3918 * not result in false positives vs. @success 3919 */ 3920 if (p->saved_state & state) { 3921 p->saved_state = TASK_RUNNING; 3922 *success = 1; 3923 } 3924 #endif 3925 return false; 3926 } 3927 3928 /* 3929 * Notes on Program-Order guarantees on SMP systems. 3930 * 3931 * MIGRATION 3932 * 3933 * The basic program-order guarantee on SMP systems is that when a task [t] 3934 * migrates, all its activity on its old CPU [c0] happens-before any subsequent 3935 * execution on its new CPU [c1]. 3936 * 3937 * For migration (of runnable tasks) this is provided by the following means: 3938 * 3939 * A) UNLOCK of the rq(c0)->lock scheduling out task t 3940 * B) migration for t is required to synchronize *both* rq(c0)->lock and 3941 * rq(c1)->lock (if not at the same time, then in that order). 3942 * C) LOCK of the rq(c1)->lock scheduling in task 3943 * 3944 * Release/acquire chaining guarantees that B happens after A and C after B. 3945 * Note: the CPU doing B need not be c0 or c1 3946 * 3947 * Example: 3948 * 3949 * CPU0 CPU1 CPU2 3950 * 3951 * LOCK rq(0)->lock 3952 * sched-out X 3953 * sched-in Y 3954 * UNLOCK rq(0)->lock 3955 * 3956 * LOCK rq(0)->lock // orders against CPU0 3957 * dequeue X 3958 * UNLOCK rq(0)->lock 3959 * 3960 * LOCK rq(1)->lock 3961 * enqueue X 3962 * UNLOCK rq(1)->lock 3963 * 3964 * LOCK rq(1)->lock // orders against CPU2 3965 * sched-out Z 3966 * sched-in X 3967 * UNLOCK rq(1)->lock 3968 * 3969 * 3970 * BLOCKING -- aka. SLEEP + WAKEUP 3971 * 3972 * For blocking we (obviously) need to provide the same guarantee as for 3973 * migration. However the means are completely different as there is no lock 3974 * chain to provide order. Instead we do: 3975 * 3976 * 1) smp_store_release(X->on_cpu, 0) -- finish_task() 3977 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up() 3978 * 3979 * Example: 3980 * 3981 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule) 3982 * 3983 * LOCK rq(0)->lock LOCK X->pi_lock 3984 * dequeue X 3985 * sched-out X 3986 * smp_store_release(X->on_cpu, 0); 3987 * 3988 * smp_cond_load_acquire(&X->on_cpu, !VAL); 3989 * X->state = WAKING 3990 * set_task_cpu(X,2) 3991 * 3992 * LOCK rq(2)->lock 3993 * enqueue X 3994 * X->state = RUNNING 3995 * UNLOCK rq(2)->lock 3996 * 3997 * LOCK rq(2)->lock // orders against CPU1 3998 * sched-out Z 3999 * sched-in X 4000 * UNLOCK rq(2)->lock 4001 * 4002 * UNLOCK X->pi_lock 4003 * UNLOCK rq(0)->lock 4004 * 4005 * 4006 * However, for wakeups there is a second guarantee we must provide, namely we 4007 * must ensure that CONDITION=1 done by the caller can not be reordered with 4008 * accesses to the task state; see try_to_wake_up() and set_current_state(). 4009 */ 4010 4011 /** 4012 * try_to_wake_up - wake up a thread 4013 * @p: the thread to be awakened 4014 * @state: the mask of task states that can be woken 4015 * @wake_flags: wake modifier flags (WF_*) 4016 * 4017 * Conceptually does: 4018 * 4019 * If (@state & @p->state) @p->state = TASK_RUNNING. 4020 * 4021 * If the task was not queued/runnable, also place it back on a runqueue. 4022 * 4023 * This function is atomic against schedule() which would dequeue the task. 4024 * 4025 * It issues a full memory barrier before accessing @p->state, see the comment 4026 * with set_current_state(). 4027 * 4028 * Uses p->pi_lock to serialize against concurrent wake-ups. 4029 * 4030 * Relies on p->pi_lock stabilizing: 4031 * - p->sched_class 4032 * - p->cpus_ptr 4033 * - p->sched_task_group 4034 * in order to do migration, see its use of select_task_rq()/set_task_cpu(). 4035 * 4036 * Tries really hard to only take one task_rq(p)->lock for performance. 4037 * Takes rq->lock in: 4038 * - ttwu_runnable() -- old rq, unavoidable, see comment there; 4039 * - ttwu_queue() -- new rq, for enqueue of the task; 4040 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us. 4041 * 4042 * As a consequence we race really badly with just about everything. See the 4043 * many memory barriers and their comments for details. 4044 * 4045 * Return: %true if @p->state changes (an actual wakeup was done), 4046 * %false otherwise. 4047 */ 4048 static int 4049 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 4050 { 4051 unsigned long flags; 4052 int cpu, success = 0; 4053 4054 preempt_disable(); 4055 if (p == current) { 4056 /* 4057 * We're waking current, this means 'p->on_rq' and 'task_cpu(p) 4058 * == smp_processor_id()'. Together this means we can special 4059 * case the whole 'p->on_rq && ttwu_runnable()' case below 4060 * without taking any locks. 4061 * 4062 * In particular: 4063 * - we rely on Program-Order guarantees for all the ordering, 4064 * - we're serialized against set_special_state() by virtue of 4065 * it disabling IRQs (this allows not taking ->pi_lock). 4066 */ 4067 if (!ttwu_state_match(p, state, &success)) 4068 goto out; 4069 4070 trace_sched_waking(p); 4071 WRITE_ONCE(p->__state, TASK_RUNNING); 4072 trace_sched_wakeup(p); 4073 goto out; 4074 } 4075 4076 /* 4077 * If we are going to wake up a thread waiting for CONDITION we 4078 * need to ensure that CONDITION=1 done by the caller can not be 4079 * reordered with p->state check below. This pairs with smp_store_mb() 4080 * in set_current_state() that the waiting thread does. 4081 */ 4082 raw_spin_lock_irqsave(&p->pi_lock, flags); 4083 smp_mb__after_spinlock(); 4084 if (!ttwu_state_match(p, state, &success)) 4085 goto unlock; 4086 4087 trace_sched_waking(p); 4088 4089 /* 4090 * Ensure we load p->on_rq _after_ p->state, otherwise it would 4091 * be possible to, falsely, observe p->on_rq == 0 and get stuck 4092 * in smp_cond_load_acquire() below. 4093 * 4094 * sched_ttwu_pending() try_to_wake_up() 4095 * STORE p->on_rq = 1 LOAD p->state 4096 * UNLOCK rq->lock 4097 * 4098 * __schedule() (switch to task 'p') 4099 * LOCK rq->lock smp_rmb(); 4100 * smp_mb__after_spinlock(); 4101 * UNLOCK rq->lock 4102 * 4103 * [task p] 4104 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq 4105 * 4106 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in 4107 * __schedule(). See the comment for smp_mb__after_spinlock(). 4108 * 4109 * A similar smb_rmb() lives in try_invoke_on_locked_down_task(). 4110 */ 4111 smp_rmb(); 4112 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags)) 4113 goto unlock; 4114 4115 #ifdef CONFIG_SMP 4116 /* 4117 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be 4118 * possible to, falsely, observe p->on_cpu == 0. 4119 * 4120 * One must be running (->on_cpu == 1) in order to remove oneself 4121 * from the runqueue. 4122 * 4123 * __schedule() (switch to task 'p') try_to_wake_up() 4124 * STORE p->on_cpu = 1 LOAD p->on_rq 4125 * UNLOCK rq->lock 4126 * 4127 * __schedule() (put 'p' to sleep) 4128 * LOCK rq->lock smp_rmb(); 4129 * smp_mb__after_spinlock(); 4130 * STORE p->on_rq = 0 LOAD p->on_cpu 4131 * 4132 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in 4133 * __schedule(). See the comment for smp_mb__after_spinlock(). 4134 * 4135 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure 4136 * schedule()'s deactivate_task() has 'happened' and p will no longer 4137 * care about it's own p->state. See the comment in __schedule(). 4138 */ 4139 smp_acquire__after_ctrl_dep(); 4140 4141 /* 4142 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq 4143 * == 0), which means we need to do an enqueue, change p->state to 4144 * TASK_WAKING such that we can unlock p->pi_lock before doing the 4145 * enqueue, such as ttwu_queue_wakelist(). 4146 */ 4147 WRITE_ONCE(p->__state, TASK_WAKING); 4148 4149 /* 4150 * If the owning (remote) CPU is still in the middle of schedule() with 4151 * this task as prev, considering queueing p on the remote CPUs wake_list 4152 * which potentially sends an IPI instead of spinning on p->on_cpu to 4153 * let the waker make forward progress. This is safe because IRQs are 4154 * disabled and the IPI will deliver after on_cpu is cleared. 4155 * 4156 * Ensure we load task_cpu(p) after p->on_cpu: 4157 * 4158 * set_task_cpu(p, cpu); 4159 * STORE p->cpu = @cpu 4160 * __schedule() (switch to task 'p') 4161 * LOCK rq->lock 4162 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu) 4163 * STORE p->on_cpu = 1 LOAD p->cpu 4164 * 4165 * to ensure we observe the correct CPU on which the task is currently 4166 * scheduling. 4167 */ 4168 if (smp_load_acquire(&p->on_cpu) && 4169 ttwu_queue_wakelist(p, task_cpu(p), wake_flags)) 4170 goto unlock; 4171 4172 /* 4173 * If the owning (remote) CPU is still in the middle of schedule() with 4174 * this task as prev, wait until it's done referencing the task. 4175 * 4176 * Pairs with the smp_store_release() in finish_task(). 4177 * 4178 * This ensures that tasks getting woken will be fully ordered against 4179 * their previous state and preserve Program Order. 4180 */ 4181 smp_cond_load_acquire(&p->on_cpu, !VAL); 4182 4183 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU); 4184 if (task_cpu(p) != cpu) { 4185 if (p->in_iowait) { 4186 delayacct_blkio_end(p); 4187 atomic_dec(&task_rq(p)->nr_iowait); 4188 } 4189 4190 wake_flags |= WF_MIGRATED; 4191 psi_ttwu_dequeue(p); 4192 set_task_cpu(p, cpu); 4193 } 4194 #else 4195 cpu = task_cpu(p); 4196 #endif /* CONFIG_SMP */ 4197 4198 ttwu_queue(p, cpu, wake_flags); 4199 unlock: 4200 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4201 out: 4202 if (success) 4203 ttwu_stat(p, task_cpu(p), wake_flags); 4204 preempt_enable(); 4205 4206 return success; 4207 } 4208 4209 /** 4210 * task_call_func - Invoke a function on task in fixed state 4211 * @p: Process for which the function is to be invoked, can be @current. 4212 * @func: Function to invoke. 4213 * @arg: Argument to function. 4214 * 4215 * Fix the task in it's current state by avoiding wakeups and or rq operations 4216 * and call @func(@arg) on it. This function can use ->on_rq and task_curr() 4217 * to work out what the state is, if required. Given that @func can be invoked 4218 * with a runqueue lock held, it had better be quite lightweight. 4219 * 4220 * Returns: 4221 * Whatever @func returns 4222 */ 4223 int task_call_func(struct task_struct *p, task_call_f func, void *arg) 4224 { 4225 struct rq *rq = NULL; 4226 unsigned int state; 4227 struct rq_flags rf; 4228 int ret; 4229 4230 raw_spin_lock_irqsave(&p->pi_lock, rf.flags); 4231 4232 state = READ_ONCE(p->__state); 4233 4234 /* 4235 * Ensure we load p->on_rq after p->__state, otherwise it would be 4236 * possible to, falsely, observe p->on_rq == 0. 4237 * 4238 * See try_to_wake_up() for a longer comment. 4239 */ 4240 smp_rmb(); 4241 4242 /* 4243 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when 4244 * the task is blocked. Make sure to check @state since ttwu() can drop 4245 * locks at the end, see ttwu_queue_wakelist(). 4246 */ 4247 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq) 4248 rq = __task_rq_lock(p, &rf); 4249 4250 /* 4251 * At this point the task is pinned; either: 4252 * - blocked and we're holding off wakeups (pi->lock) 4253 * - woken, and we're holding off enqueue (rq->lock) 4254 * - queued, and we're holding off schedule (rq->lock) 4255 * - running, and we're holding off de-schedule (rq->lock) 4256 * 4257 * The called function (@func) can use: task_curr(), p->on_rq and 4258 * p->__state to differentiate between these states. 4259 */ 4260 ret = func(p, arg); 4261 4262 if (rq) 4263 rq_unlock(rq, &rf); 4264 4265 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags); 4266 return ret; 4267 } 4268 4269 /** 4270 * cpu_curr_snapshot - Return a snapshot of the currently running task 4271 * @cpu: The CPU on which to snapshot the task. 4272 * 4273 * Returns the task_struct pointer of the task "currently" running on 4274 * the specified CPU. If the same task is running on that CPU throughout, 4275 * the return value will be a pointer to that task's task_struct structure. 4276 * If the CPU did any context switches even vaguely concurrently with the 4277 * execution of this function, the return value will be a pointer to the 4278 * task_struct structure of a randomly chosen task that was running on 4279 * that CPU somewhere around the time that this function was executing. 4280 * 4281 * If the specified CPU was offline, the return value is whatever it 4282 * is, perhaps a pointer to the task_struct structure of that CPU's idle 4283 * task, but there is no guarantee. Callers wishing a useful return 4284 * value must take some action to ensure that the specified CPU remains 4285 * online throughout. 4286 * 4287 * This function executes full memory barriers before and after fetching 4288 * the pointer, which permits the caller to confine this function's fetch 4289 * with respect to the caller's accesses to other shared variables. 4290 */ 4291 struct task_struct *cpu_curr_snapshot(int cpu) 4292 { 4293 struct task_struct *t; 4294 4295 smp_mb(); /* Pairing determined by caller's synchronization design. */ 4296 t = rcu_dereference(cpu_curr(cpu)); 4297 smp_mb(); /* Pairing determined by caller's synchronization design. */ 4298 return t; 4299 } 4300 4301 /** 4302 * wake_up_process - Wake up a specific process 4303 * @p: The process to be woken up. 4304 * 4305 * Attempt to wake up the nominated process and move it to the set of runnable 4306 * processes. 4307 * 4308 * Return: 1 if the process was woken up, 0 if it was already running. 4309 * 4310 * This function executes a full memory barrier before accessing the task state. 4311 */ 4312 int wake_up_process(struct task_struct *p) 4313 { 4314 return try_to_wake_up(p, TASK_NORMAL, 0); 4315 } 4316 EXPORT_SYMBOL(wake_up_process); 4317 4318 int wake_up_state(struct task_struct *p, unsigned int state) 4319 { 4320 return try_to_wake_up(p, state, 0); 4321 } 4322 4323 /* 4324 * Perform scheduler related setup for a newly forked process p. 4325 * p is forked by current. 4326 * 4327 * __sched_fork() is basic setup used by init_idle() too: 4328 */ 4329 static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 4330 { 4331 p->on_rq = 0; 4332 4333 p->se.on_rq = 0; 4334 p->se.exec_start = 0; 4335 p->se.sum_exec_runtime = 0; 4336 p->se.prev_sum_exec_runtime = 0; 4337 p->se.nr_migrations = 0; 4338 p->se.vruntime = 0; 4339 INIT_LIST_HEAD(&p->se.group_node); 4340 4341 #ifdef CONFIG_FAIR_GROUP_SCHED 4342 p->se.cfs_rq = NULL; 4343 #endif 4344 4345 #ifdef CONFIG_SCHEDSTATS 4346 /* Even if schedstat is disabled, there should not be garbage */ 4347 memset(&p->stats, 0, sizeof(p->stats)); 4348 #endif 4349 4350 RB_CLEAR_NODE(&p->dl.rb_node); 4351 init_dl_task_timer(&p->dl); 4352 init_dl_inactive_task_timer(&p->dl); 4353 __dl_clear_params(p); 4354 4355 INIT_LIST_HEAD(&p->rt.run_list); 4356 p->rt.timeout = 0; 4357 p->rt.time_slice = sched_rr_timeslice; 4358 p->rt.on_rq = 0; 4359 p->rt.on_list = 0; 4360 4361 #ifdef CONFIG_PREEMPT_NOTIFIERS 4362 INIT_HLIST_HEAD(&p->preempt_notifiers); 4363 #endif 4364 4365 #ifdef CONFIG_COMPACTION 4366 p->capture_control = NULL; 4367 #endif 4368 init_numa_balancing(clone_flags, p); 4369 #ifdef CONFIG_SMP 4370 p->wake_entry.u_flags = CSD_TYPE_TTWU; 4371 p->migration_pending = NULL; 4372 #endif 4373 } 4374 4375 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing); 4376 4377 #ifdef CONFIG_NUMA_BALANCING 4378 4379 int sysctl_numa_balancing_mode; 4380 4381 static void __set_numabalancing_state(bool enabled) 4382 { 4383 if (enabled) 4384 static_branch_enable(&sched_numa_balancing); 4385 else 4386 static_branch_disable(&sched_numa_balancing); 4387 } 4388 4389 void set_numabalancing_state(bool enabled) 4390 { 4391 if (enabled) 4392 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL; 4393 else 4394 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED; 4395 __set_numabalancing_state(enabled); 4396 } 4397 4398 #ifdef CONFIG_PROC_SYSCTL 4399 static void reset_memory_tiering(void) 4400 { 4401 struct pglist_data *pgdat; 4402 4403 for_each_online_pgdat(pgdat) { 4404 pgdat->nbp_threshold = 0; 4405 pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE); 4406 pgdat->nbp_th_start = jiffies_to_msecs(jiffies); 4407 } 4408 } 4409 4410 int sysctl_numa_balancing(struct ctl_table *table, int write, 4411 void *buffer, size_t *lenp, loff_t *ppos) 4412 { 4413 struct ctl_table t; 4414 int err; 4415 int state = sysctl_numa_balancing_mode; 4416 4417 if (write && !capable(CAP_SYS_ADMIN)) 4418 return -EPERM; 4419 4420 t = *table; 4421 t.data = &state; 4422 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 4423 if (err < 0) 4424 return err; 4425 if (write) { 4426 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) && 4427 (state & NUMA_BALANCING_MEMORY_TIERING)) 4428 reset_memory_tiering(); 4429 sysctl_numa_balancing_mode = state; 4430 __set_numabalancing_state(state); 4431 } 4432 return err; 4433 } 4434 #endif 4435 #endif 4436 4437 #ifdef CONFIG_SCHEDSTATS 4438 4439 DEFINE_STATIC_KEY_FALSE(sched_schedstats); 4440 4441 static void set_schedstats(bool enabled) 4442 { 4443 if (enabled) 4444 static_branch_enable(&sched_schedstats); 4445 else 4446 static_branch_disable(&sched_schedstats); 4447 } 4448 4449 void force_schedstat_enabled(void) 4450 { 4451 if (!schedstat_enabled()) { 4452 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n"); 4453 static_branch_enable(&sched_schedstats); 4454 } 4455 } 4456 4457 static int __init setup_schedstats(char *str) 4458 { 4459 int ret = 0; 4460 if (!str) 4461 goto out; 4462 4463 if (!strcmp(str, "enable")) { 4464 set_schedstats(true); 4465 ret = 1; 4466 } else if (!strcmp(str, "disable")) { 4467 set_schedstats(false); 4468 ret = 1; 4469 } 4470 out: 4471 if (!ret) 4472 pr_warn("Unable to parse schedstats=\n"); 4473 4474 return ret; 4475 } 4476 __setup("schedstats=", setup_schedstats); 4477 4478 #ifdef CONFIG_PROC_SYSCTL 4479 static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer, 4480 size_t *lenp, loff_t *ppos) 4481 { 4482 struct ctl_table t; 4483 int err; 4484 int state = static_branch_likely(&sched_schedstats); 4485 4486 if (write && !capable(CAP_SYS_ADMIN)) 4487 return -EPERM; 4488 4489 t = *table; 4490 t.data = &state; 4491 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 4492 if (err < 0) 4493 return err; 4494 if (write) 4495 set_schedstats(state); 4496 return err; 4497 } 4498 #endif /* CONFIG_PROC_SYSCTL */ 4499 #endif /* CONFIG_SCHEDSTATS */ 4500 4501 #ifdef CONFIG_SYSCTL 4502 static struct ctl_table sched_core_sysctls[] = { 4503 #ifdef CONFIG_SCHEDSTATS 4504 { 4505 .procname = "sched_schedstats", 4506 .data = NULL, 4507 .maxlen = sizeof(unsigned int), 4508 .mode = 0644, 4509 .proc_handler = sysctl_schedstats, 4510 .extra1 = SYSCTL_ZERO, 4511 .extra2 = SYSCTL_ONE, 4512 }, 4513 #endif /* CONFIG_SCHEDSTATS */ 4514 #ifdef CONFIG_UCLAMP_TASK 4515 { 4516 .procname = "sched_util_clamp_min", 4517 .data = &sysctl_sched_uclamp_util_min, 4518 .maxlen = sizeof(unsigned int), 4519 .mode = 0644, 4520 .proc_handler = sysctl_sched_uclamp_handler, 4521 }, 4522 { 4523 .procname = "sched_util_clamp_max", 4524 .data = &sysctl_sched_uclamp_util_max, 4525 .maxlen = sizeof(unsigned int), 4526 .mode = 0644, 4527 .proc_handler = sysctl_sched_uclamp_handler, 4528 }, 4529 { 4530 .procname = "sched_util_clamp_min_rt_default", 4531 .data = &sysctl_sched_uclamp_util_min_rt_default, 4532 .maxlen = sizeof(unsigned int), 4533 .mode = 0644, 4534 .proc_handler = sysctl_sched_uclamp_handler, 4535 }, 4536 #endif /* CONFIG_UCLAMP_TASK */ 4537 {} 4538 }; 4539 static int __init sched_core_sysctl_init(void) 4540 { 4541 register_sysctl_init("kernel", sched_core_sysctls); 4542 return 0; 4543 } 4544 late_initcall(sched_core_sysctl_init); 4545 #endif /* CONFIG_SYSCTL */ 4546 4547 /* 4548 * fork()/clone()-time setup: 4549 */ 4550 int sched_fork(unsigned long clone_flags, struct task_struct *p) 4551 { 4552 __sched_fork(clone_flags, p); 4553 /* 4554 * We mark the process as NEW here. This guarantees that 4555 * nobody will actually run it, and a signal or other external 4556 * event cannot wake it up and insert it on the runqueue either. 4557 */ 4558 p->__state = TASK_NEW; 4559 4560 /* 4561 * Make sure we do not leak PI boosting priority to the child. 4562 */ 4563 p->prio = current->normal_prio; 4564 4565 uclamp_fork(p); 4566 4567 /* 4568 * Revert to default priority/policy on fork if requested. 4569 */ 4570 if (unlikely(p->sched_reset_on_fork)) { 4571 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 4572 p->policy = SCHED_NORMAL; 4573 p->static_prio = NICE_TO_PRIO(0); 4574 p->rt_priority = 0; 4575 } else if (PRIO_TO_NICE(p->static_prio) < 0) 4576 p->static_prio = NICE_TO_PRIO(0); 4577 4578 p->prio = p->normal_prio = p->static_prio; 4579 set_load_weight(p, false); 4580 4581 /* 4582 * We don't need the reset flag anymore after the fork. It has 4583 * fulfilled its duty: 4584 */ 4585 p->sched_reset_on_fork = 0; 4586 } 4587 4588 if (dl_prio(p->prio)) 4589 return -EAGAIN; 4590 else if (rt_prio(p->prio)) 4591 p->sched_class = &rt_sched_class; 4592 else 4593 p->sched_class = &fair_sched_class; 4594 4595 init_entity_runnable_average(&p->se); 4596 4597 4598 #ifdef CONFIG_SCHED_INFO 4599 if (likely(sched_info_on())) 4600 memset(&p->sched_info, 0, sizeof(p->sched_info)); 4601 #endif 4602 #if defined(CONFIG_SMP) 4603 p->on_cpu = 0; 4604 #endif 4605 init_task_preempt_count(p); 4606 #ifdef CONFIG_SMP 4607 plist_node_init(&p->pushable_tasks, MAX_PRIO); 4608 RB_CLEAR_NODE(&p->pushable_dl_tasks); 4609 #endif 4610 return 0; 4611 } 4612 4613 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs) 4614 { 4615 unsigned long flags; 4616 4617 /* 4618 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly 4619 * required yet, but lockdep gets upset if rules are violated. 4620 */ 4621 raw_spin_lock_irqsave(&p->pi_lock, flags); 4622 #ifdef CONFIG_CGROUP_SCHED 4623 if (1) { 4624 struct task_group *tg; 4625 tg = container_of(kargs->cset->subsys[cpu_cgrp_id], 4626 struct task_group, css); 4627 tg = autogroup_task_group(p, tg); 4628 p->sched_task_group = tg; 4629 } 4630 #endif 4631 rseq_migrate(p); 4632 /* 4633 * We're setting the CPU for the first time, we don't migrate, 4634 * so use __set_task_cpu(). 4635 */ 4636 __set_task_cpu(p, smp_processor_id()); 4637 if (p->sched_class->task_fork) 4638 p->sched_class->task_fork(p); 4639 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4640 } 4641 4642 void sched_post_fork(struct task_struct *p) 4643 { 4644 uclamp_post_fork(p); 4645 } 4646 4647 unsigned long to_ratio(u64 period, u64 runtime) 4648 { 4649 if (runtime == RUNTIME_INF) 4650 return BW_UNIT; 4651 4652 /* 4653 * Doing this here saves a lot of checks in all 4654 * the calling paths, and returning zero seems 4655 * safe for them anyway. 4656 */ 4657 if (period == 0) 4658 return 0; 4659 4660 return div64_u64(runtime << BW_SHIFT, period); 4661 } 4662 4663 /* 4664 * wake_up_new_task - wake up a newly created task for the first time. 4665 * 4666 * This function will do some initial scheduler statistics housekeeping 4667 * that must be done for every newly created context, then puts the task 4668 * on the runqueue and wakes it. 4669 */ 4670 void wake_up_new_task(struct task_struct *p) 4671 { 4672 struct rq_flags rf; 4673 struct rq *rq; 4674 4675 raw_spin_lock_irqsave(&p->pi_lock, rf.flags); 4676 WRITE_ONCE(p->__state, TASK_RUNNING); 4677 #ifdef CONFIG_SMP 4678 /* 4679 * Fork balancing, do it here and not earlier because: 4680 * - cpus_ptr can change in the fork path 4681 * - any previously selected CPU might disappear through hotplug 4682 * 4683 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq, 4684 * as we're not fully set-up yet. 4685 */ 4686 p->recent_used_cpu = task_cpu(p); 4687 rseq_migrate(p); 4688 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK)); 4689 #endif 4690 rq = __task_rq_lock(p, &rf); 4691 update_rq_clock(rq); 4692 post_init_entity_util_avg(p); 4693 4694 activate_task(rq, p, ENQUEUE_NOCLOCK); 4695 trace_sched_wakeup_new(p); 4696 check_preempt_curr(rq, p, WF_FORK); 4697 #ifdef CONFIG_SMP 4698 if (p->sched_class->task_woken) { 4699 /* 4700 * Nothing relies on rq->lock after this, so it's fine to 4701 * drop it. 4702 */ 4703 rq_unpin_lock(rq, &rf); 4704 p->sched_class->task_woken(rq, p); 4705 rq_repin_lock(rq, &rf); 4706 } 4707 #endif 4708 task_rq_unlock(rq, p, &rf); 4709 } 4710 4711 #ifdef CONFIG_PREEMPT_NOTIFIERS 4712 4713 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key); 4714 4715 void preempt_notifier_inc(void) 4716 { 4717 static_branch_inc(&preempt_notifier_key); 4718 } 4719 EXPORT_SYMBOL_GPL(preempt_notifier_inc); 4720 4721 void preempt_notifier_dec(void) 4722 { 4723 static_branch_dec(&preempt_notifier_key); 4724 } 4725 EXPORT_SYMBOL_GPL(preempt_notifier_dec); 4726 4727 /** 4728 * preempt_notifier_register - tell me when current is being preempted & rescheduled 4729 * @notifier: notifier struct to register 4730 */ 4731 void preempt_notifier_register(struct preempt_notifier *notifier) 4732 { 4733 if (!static_branch_unlikely(&preempt_notifier_key)) 4734 WARN(1, "registering preempt_notifier while notifiers disabled\n"); 4735 4736 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 4737 } 4738 EXPORT_SYMBOL_GPL(preempt_notifier_register); 4739 4740 /** 4741 * preempt_notifier_unregister - no longer interested in preemption notifications 4742 * @notifier: notifier struct to unregister 4743 * 4744 * This is *not* safe to call from within a preemption notifier. 4745 */ 4746 void preempt_notifier_unregister(struct preempt_notifier *notifier) 4747 { 4748 hlist_del(¬ifier->link); 4749 } 4750 EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 4751 4752 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr) 4753 { 4754 struct preempt_notifier *notifier; 4755 4756 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 4757 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 4758 } 4759 4760 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) 4761 { 4762 if (static_branch_unlikely(&preempt_notifier_key)) 4763 __fire_sched_in_preempt_notifiers(curr); 4764 } 4765 4766 static void 4767 __fire_sched_out_preempt_notifiers(struct task_struct *curr, 4768 struct task_struct *next) 4769 { 4770 struct preempt_notifier *notifier; 4771 4772 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 4773 notifier->ops->sched_out(notifier, next); 4774 } 4775 4776 static __always_inline void 4777 fire_sched_out_preempt_notifiers(struct task_struct *curr, 4778 struct task_struct *next) 4779 { 4780 if (static_branch_unlikely(&preempt_notifier_key)) 4781 __fire_sched_out_preempt_notifiers(curr, next); 4782 } 4783 4784 #else /* !CONFIG_PREEMPT_NOTIFIERS */ 4785 4786 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr) 4787 { 4788 } 4789 4790 static inline void 4791 fire_sched_out_preempt_notifiers(struct task_struct *curr, 4792 struct task_struct *next) 4793 { 4794 } 4795 4796 #endif /* CONFIG_PREEMPT_NOTIFIERS */ 4797 4798 static inline void prepare_task(struct task_struct *next) 4799 { 4800 #ifdef CONFIG_SMP 4801 /* 4802 * Claim the task as running, we do this before switching to it 4803 * such that any running task will have this set. 4804 * 4805 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and 4806 * its ordering comment. 4807 */ 4808 WRITE_ONCE(next->on_cpu, 1); 4809 #endif 4810 } 4811 4812 static inline void finish_task(struct task_struct *prev) 4813 { 4814 #ifdef CONFIG_SMP 4815 /* 4816 * This must be the very last reference to @prev from this CPU. After 4817 * p->on_cpu is cleared, the task can be moved to a different CPU. We 4818 * must ensure this doesn't happen until the switch is completely 4819 * finished. 4820 * 4821 * In particular, the load of prev->state in finish_task_switch() must 4822 * happen before this. 4823 * 4824 * Pairs with the smp_cond_load_acquire() in try_to_wake_up(). 4825 */ 4826 smp_store_release(&prev->on_cpu, 0); 4827 #endif 4828 } 4829 4830 #ifdef CONFIG_SMP 4831 4832 static void do_balance_callbacks(struct rq *rq, struct callback_head *head) 4833 { 4834 void (*func)(struct rq *rq); 4835 struct callback_head *next; 4836 4837 lockdep_assert_rq_held(rq); 4838 4839 while (head) { 4840 func = (void (*)(struct rq *))head->func; 4841 next = head->next; 4842 head->next = NULL; 4843 head = next; 4844 4845 func(rq); 4846 } 4847 } 4848 4849 static void balance_push(struct rq *rq); 4850 4851 /* 4852 * balance_push_callback is a right abuse of the callback interface and plays 4853 * by significantly different rules. 4854 * 4855 * Where the normal balance_callback's purpose is to be ran in the same context 4856 * that queued it (only later, when it's safe to drop rq->lock again), 4857 * balance_push_callback is specifically targeted at __schedule(). 4858 * 4859 * This abuse is tolerated because it places all the unlikely/odd cases behind 4860 * a single test, namely: rq->balance_callback == NULL. 4861 */ 4862 struct callback_head balance_push_callback = { 4863 .next = NULL, 4864 .func = (void (*)(struct callback_head *))balance_push, 4865 }; 4866 4867 static inline struct callback_head * 4868 __splice_balance_callbacks(struct rq *rq, bool split) 4869 { 4870 struct callback_head *head = rq->balance_callback; 4871 4872 if (likely(!head)) 4873 return NULL; 4874 4875 lockdep_assert_rq_held(rq); 4876 /* 4877 * Must not take balance_push_callback off the list when 4878 * splice_balance_callbacks() and balance_callbacks() are not 4879 * in the same rq->lock section. 4880 * 4881 * In that case it would be possible for __schedule() to interleave 4882 * and observe the list empty. 4883 */ 4884 if (split && head == &balance_push_callback) 4885 head = NULL; 4886 else 4887 rq->balance_callback = NULL; 4888 4889 return head; 4890 } 4891 4892 static inline struct callback_head *splice_balance_callbacks(struct rq *rq) 4893 { 4894 return __splice_balance_callbacks(rq, true); 4895 } 4896 4897 static void __balance_callbacks(struct rq *rq) 4898 { 4899 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false)); 4900 } 4901 4902 static inline void balance_callbacks(struct rq *rq, struct callback_head *head) 4903 { 4904 unsigned long flags; 4905 4906 if (unlikely(head)) { 4907 raw_spin_rq_lock_irqsave(rq, flags); 4908 do_balance_callbacks(rq, head); 4909 raw_spin_rq_unlock_irqrestore(rq, flags); 4910 } 4911 } 4912 4913 #else 4914 4915 static inline void __balance_callbacks(struct rq *rq) 4916 { 4917 } 4918 4919 static inline struct callback_head *splice_balance_callbacks(struct rq *rq) 4920 { 4921 return NULL; 4922 } 4923 4924 static inline void balance_callbacks(struct rq *rq, struct callback_head *head) 4925 { 4926 } 4927 4928 #endif 4929 4930 static inline void 4931 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf) 4932 { 4933 /* 4934 * Since the runqueue lock will be released by the next 4935 * task (which is an invalid locking op but in the case 4936 * of the scheduler it's an obvious special-case), so we 4937 * do an early lockdep release here: 4938 */ 4939 rq_unpin_lock(rq, rf); 4940 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_); 4941 #ifdef CONFIG_DEBUG_SPINLOCK 4942 /* this is a valid case when another task releases the spinlock */ 4943 rq_lockp(rq)->owner = next; 4944 #endif 4945 } 4946 4947 static inline void finish_lock_switch(struct rq *rq) 4948 { 4949 /* 4950 * If we are tracking spinlock dependencies then we have to 4951 * fix up the runqueue lock - which gets 'carried over' from 4952 * prev into current: 4953 */ 4954 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_); 4955 __balance_callbacks(rq); 4956 raw_spin_rq_unlock_irq(rq); 4957 } 4958 4959 /* 4960 * NOP if the arch has not defined these: 4961 */ 4962 4963 #ifndef prepare_arch_switch 4964 # define prepare_arch_switch(next) do { } while (0) 4965 #endif 4966 4967 #ifndef finish_arch_post_lock_switch 4968 # define finish_arch_post_lock_switch() do { } while (0) 4969 #endif 4970 4971 static inline void kmap_local_sched_out(void) 4972 { 4973 #ifdef CONFIG_KMAP_LOCAL 4974 if (unlikely(current->kmap_ctrl.idx)) 4975 __kmap_local_sched_out(); 4976 #endif 4977 } 4978 4979 static inline void kmap_local_sched_in(void) 4980 { 4981 #ifdef CONFIG_KMAP_LOCAL 4982 if (unlikely(current->kmap_ctrl.idx)) 4983 __kmap_local_sched_in(); 4984 #endif 4985 } 4986 4987 /** 4988 * prepare_task_switch - prepare to switch tasks 4989 * @rq: the runqueue preparing to switch 4990 * @prev: the current task that is being switched out 4991 * @next: the task we are going to switch to. 4992 * 4993 * This is called with the rq lock held and interrupts off. It must 4994 * be paired with a subsequent finish_task_switch after the context 4995 * switch. 4996 * 4997 * prepare_task_switch sets up locking and calls architecture specific 4998 * hooks. 4999 */ 5000 static inline void 5001 prepare_task_switch(struct rq *rq, struct task_struct *prev, 5002 struct task_struct *next) 5003 { 5004 kcov_prepare_switch(prev); 5005 sched_info_switch(rq, prev, next); 5006 perf_event_task_sched_out(prev, next); 5007 rseq_preempt(prev); 5008 fire_sched_out_preempt_notifiers(prev, next); 5009 kmap_local_sched_out(); 5010 prepare_task(next); 5011 prepare_arch_switch(next); 5012 } 5013 5014 /** 5015 * finish_task_switch - clean up after a task-switch 5016 * @prev: the thread we just switched away from. 5017 * 5018 * finish_task_switch must be called after the context switch, paired 5019 * with a prepare_task_switch call before the context switch. 5020 * finish_task_switch will reconcile locking set up by prepare_task_switch, 5021 * and do any other architecture-specific cleanup actions. 5022 * 5023 * Note that we may have delayed dropping an mm in context_switch(). If 5024 * so, we finish that here outside of the runqueue lock. (Doing it 5025 * with the lock held can cause deadlocks; see schedule() for 5026 * details.) 5027 * 5028 * The context switch have flipped the stack from under us and restored the 5029 * local variables which were saved when this task called schedule() in the 5030 * past. prev == current is still correct but we need to recalculate this_rq 5031 * because prev may have moved to another CPU. 5032 */ 5033 static struct rq *finish_task_switch(struct task_struct *prev) 5034 __releases(rq->lock) 5035 { 5036 struct rq *rq = this_rq(); 5037 struct mm_struct *mm = rq->prev_mm; 5038 unsigned int prev_state; 5039 5040 /* 5041 * The previous task will have left us with a preempt_count of 2 5042 * because it left us after: 5043 * 5044 * schedule() 5045 * preempt_disable(); // 1 5046 * __schedule() 5047 * raw_spin_lock_irq(&rq->lock) // 2 5048 * 5049 * Also, see FORK_PREEMPT_COUNT. 5050 */ 5051 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET, 5052 "corrupted preempt_count: %s/%d/0x%x\n", 5053 current->comm, current->pid, preempt_count())) 5054 preempt_count_set(FORK_PREEMPT_COUNT); 5055 5056 rq->prev_mm = NULL; 5057 5058 /* 5059 * A task struct has one reference for the use as "current". 5060 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 5061 * schedule one last time. The schedule call will never return, and 5062 * the scheduled task must drop that reference. 5063 * 5064 * We must observe prev->state before clearing prev->on_cpu (in 5065 * finish_task), otherwise a concurrent wakeup can get prev 5066 * running on another CPU and we could rave with its RUNNING -> DEAD 5067 * transition, resulting in a double drop. 5068 */ 5069 prev_state = READ_ONCE(prev->__state); 5070 vtime_task_switch(prev); 5071 perf_event_task_sched_in(prev, current); 5072 finish_task(prev); 5073 tick_nohz_task_switch(); 5074 finish_lock_switch(rq); 5075 finish_arch_post_lock_switch(); 5076 kcov_finish_switch(current); 5077 /* 5078 * kmap_local_sched_out() is invoked with rq::lock held and 5079 * interrupts disabled. There is no requirement for that, but the 5080 * sched out code does not have an interrupt enabled section. 5081 * Restoring the maps on sched in does not require interrupts being 5082 * disabled either. 5083 */ 5084 kmap_local_sched_in(); 5085 5086 fire_sched_in_preempt_notifiers(current); 5087 /* 5088 * When switching through a kernel thread, the loop in 5089 * membarrier_{private,global}_expedited() may have observed that 5090 * kernel thread and not issued an IPI. It is therefore possible to 5091 * schedule between user->kernel->user threads without passing though 5092 * switch_mm(). Membarrier requires a barrier after storing to 5093 * rq->curr, before returning to userspace, so provide them here: 5094 * 5095 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly 5096 * provided by mmdrop(), 5097 * - a sync_core for SYNC_CORE. 5098 */ 5099 if (mm) { 5100 membarrier_mm_sync_core_before_usermode(mm); 5101 mmdrop_sched(mm); 5102 } 5103 if (unlikely(prev_state == TASK_DEAD)) { 5104 if (prev->sched_class->task_dead) 5105 prev->sched_class->task_dead(prev); 5106 5107 /* Task is done with its stack. */ 5108 put_task_stack(prev); 5109 5110 put_task_struct_rcu_user(prev); 5111 } 5112 5113 return rq; 5114 } 5115 5116 /** 5117 * schedule_tail - first thing a freshly forked thread must call. 5118 * @prev: the thread we just switched away from. 5119 */ 5120 asmlinkage __visible void schedule_tail(struct task_struct *prev) 5121 __releases(rq->lock) 5122 { 5123 /* 5124 * New tasks start with FORK_PREEMPT_COUNT, see there and 5125 * finish_task_switch() for details. 5126 * 5127 * finish_task_switch() will drop rq->lock() and lower preempt_count 5128 * and the preempt_enable() will end up enabling preemption (on 5129 * PREEMPT_COUNT kernels). 5130 */ 5131 5132 finish_task_switch(prev); 5133 preempt_enable(); 5134 5135 if (current->set_child_tid) 5136 put_user(task_pid_vnr(current), current->set_child_tid); 5137 5138 calculate_sigpending(); 5139 } 5140 5141 /* 5142 * context_switch - switch to the new MM and the new thread's register state. 5143 */ 5144 static __always_inline struct rq * 5145 context_switch(struct rq *rq, struct task_struct *prev, 5146 struct task_struct *next, struct rq_flags *rf) 5147 { 5148 prepare_task_switch(rq, prev, next); 5149 5150 /* 5151 * For paravirt, this is coupled with an exit in switch_to to 5152 * combine the page table reload and the switch backend into 5153 * one hypercall. 5154 */ 5155 arch_start_context_switch(prev); 5156 5157 /* 5158 * kernel -> kernel lazy + transfer active 5159 * user -> kernel lazy + mmgrab() active 5160 * 5161 * kernel -> user switch + mmdrop() active 5162 * user -> user switch 5163 */ 5164 if (!next->mm) { // to kernel 5165 enter_lazy_tlb(prev->active_mm, next); 5166 5167 next->active_mm = prev->active_mm; 5168 if (prev->mm) // from user 5169 mmgrab(prev->active_mm); 5170 else 5171 prev->active_mm = NULL; 5172 } else { // to user 5173 membarrier_switch_mm(rq, prev->active_mm, next->mm); 5174 /* 5175 * sys_membarrier() requires an smp_mb() between setting 5176 * rq->curr / membarrier_switch_mm() and returning to userspace. 5177 * 5178 * The below provides this either through switch_mm(), or in 5179 * case 'prev->active_mm == next->mm' through 5180 * finish_task_switch()'s mmdrop(). 5181 */ 5182 switch_mm_irqs_off(prev->active_mm, next->mm, next); 5183 lru_gen_use_mm(next->mm); 5184 5185 if (!prev->mm) { // from kernel 5186 /* will mmdrop() in finish_task_switch(). */ 5187 rq->prev_mm = prev->active_mm; 5188 prev->active_mm = NULL; 5189 } 5190 } 5191 5192 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP); 5193 5194 prepare_lock_switch(rq, next, rf); 5195 5196 /* Here we just switch the register state and the stack. */ 5197 switch_to(prev, next, prev); 5198 barrier(); 5199 5200 return finish_task_switch(prev); 5201 } 5202 5203 /* 5204 * nr_running and nr_context_switches: 5205 * 5206 * externally visible scheduler statistics: current number of runnable 5207 * threads, total number of context switches performed since bootup. 5208 */ 5209 unsigned int nr_running(void) 5210 { 5211 unsigned int i, sum = 0; 5212 5213 for_each_online_cpu(i) 5214 sum += cpu_rq(i)->nr_running; 5215 5216 return sum; 5217 } 5218 5219 /* 5220 * Check if only the current task is running on the CPU. 5221 * 5222 * Caution: this function does not check that the caller has disabled 5223 * preemption, thus the result might have a time-of-check-to-time-of-use 5224 * race. The caller is responsible to use it correctly, for example: 5225 * 5226 * - from a non-preemptible section (of course) 5227 * 5228 * - from a thread that is bound to a single CPU 5229 * 5230 * - in a loop with very short iterations (e.g. a polling loop) 5231 */ 5232 bool single_task_running(void) 5233 { 5234 return raw_rq()->nr_running == 1; 5235 } 5236 EXPORT_SYMBOL(single_task_running); 5237 5238 unsigned long long nr_context_switches(void) 5239 { 5240 int i; 5241 unsigned long long sum = 0; 5242 5243 for_each_possible_cpu(i) 5244 sum += cpu_rq(i)->nr_switches; 5245 5246 return sum; 5247 } 5248 5249 /* 5250 * Consumers of these two interfaces, like for example the cpuidle menu 5251 * governor, are using nonsensical data. Preferring shallow idle state selection 5252 * for a CPU that has IO-wait which might not even end up running the task when 5253 * it does become runnable. 5254 */ 5255 5256 unsigned int nr_iowait_cpu(int cpu) 5257 { 5258 return atomic_read(&cpu_rq(cpu)->nr_iowait); 5259 } 5260 5261 /* 5262 * IO-wait accounting, and how it's mostly bollocks (on SMP). 5263 * 5264 * The idea behind IO-wait account is to account the idle time that we could 5265 * have spend running if it were not for IO. That is, if we were to improve the 5266 * storage performance, we'd have a proportional reduction in IO-wait time. 5267 * 5268 * This all works nicely on UP, where, when a task blocks on IO, we account 5269 * idle time as IO-wait, because if the storage were faster, it could've been 5270 * running and we'd not be idle. 5271 * 5272 * This has been extended to SMP, by doing the same for each CPU. This however 5273 * is broken. 5274 * 5275 * Imagine for instance the case where two tasks block on one CPU, only the one 5276 * CPU will have IO-wait accounted, while the other has regular idle. Even 5277 * though, if the storage were faster, both could've ran at the same time, 5278 * utilising both CPUs. 5279 * 5280 * This means, that when looking globally, the current IO-wait accounting on 5281 * SMP is a lower bound, by reason of under accounting. 5282 * 5283 * Worse, since the numbers are provided per CPU, they are sometimes 5284 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly 5285 * associated with any one particular CPU, it can wake to another CPU than it 5286 * blocked on. This means the per CPU IO-wait number is meaningless. 5287 * 5288 * Task CPU affinities can make all that even more 'interesting'. 5289 */ 5290 5291 unsigned int nr_iowait(void) 5292 { 5293 unsigned int i, sum = 0; 5294 5295 for_each_possible_cpu(i) 5296 sum += nr_iowait_cpu(i); 5297 5298 return sum; 5299 } 5300 5301 #ifdef CONFIG_SMP 5302 5303 /* 5304 * sched_exec - execve() is a valuable balancing opportunity, because at 5305 * this point the task has the smallest effective memory and cache footprint. 5306 */ 5307 void sched_exec(void) 5308 { 5309 struct task_struct *p = current; 5310 unsigned long flags; 5311 int dest_cpu; 5312 5313 raw_spin_lock_irqsave(&p->pi_lock, flags); 5314 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC); 5315 if (dest_cpu == smp_processor_id()) 5316 goto unlock; 5317 5318 if (likely(cpu_active(dest_cpu))) { 5319 struct migration_arg arg = { p, dest_cpu }; 5320 5321 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 5322 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 5323 return; 5324 } 5325 unlock: 5326 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 5327 } 5328 5329 #endif 5330 5331 DEFINE_PER_CPU(struct kernel_stat, kstat); 5332 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 5333 5334 EXPORT_PER_CPU_SYMBOL(kstat); 5335 EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 5336 5337 /* 5338 * The function fair_sched_class.update_curr accesses the struct curr 5339 * and its field curr->exec_start; when called from task_sched_runtime(), 5340 * we observe a high rate of cache misses in practice. 5341 * Prefetching this data results in improved performance. 5342 */ 5343 static inline void prefetch_curr_exec_start(struct task_struct *p) 5344 { 5345 #ifdef CONFIG_FAIR_GROUP_SCHED 5346 struct sched_entity *curr = (&p->se)->cfs_rq->curr; 5347 #else 5348 struct sched_entity *curr = (&task_rq(p)->cfs)->curr; 5349 #endif 5350 prefetch(curr); 5351 prefetch(&curr->exec_start); 5352 } 5353 5354 /* 5355 * Return accounted runtime for the task. 5356 * In case the task is currently running, return the runtime plus current's 5357 * pending runtime that have not been accounted yet. 5358 */ 5359 unsigned long long task_sched_runtime(struct task_struct *p) 5360 { 5361 struct rq_flags rf; 5362 struct rq *rq; 5363 u64 ns; 5364 5365 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 5366 /* 5367 * 64-bit doesn't need locks to atomically read a 64-bit value. 5368 * So we have a optimization chance when the task's delta_exec is 0. 5369 * Reading ->on_cpu is racy, but this is ok. 5370 * 5371 * If we race with it leaving CPU, we'll take a lock. So we're correct. 5372 * If we race with it entering CPU, unaccounted time is 0. This is 5373 * indistinguishable from the read occurring a few cycles earlier. 5374 * If we see ->on_cpu without ->on_rq, the task is leaving, and has 5375 * been accounted, so we're correct here as well. 5376 */ 5377 if (!p->on_cpu || !task_on_rq_queued(p)) 5378 return p->se.sum_exec_runtime; 5379 #endif 5380 5381 rq = task_rq_lock(p, &rf); 5382 /* 5383 * Must be ->curr _and_ ->on_rq. If dequeued, we would 5384 * project cycles that may never be accounted to this 5385 * thread, breaking clock_gettime(). 5386 */ 5387 if (task_current(rq, p) && task_on_rq_queued(p)) { 5388 prefetch_curr_exec_start(p); 5389 update_rq_clock(rq); 5390 p->sched_class->update_curr(rq); 5391 } 5392 ns = p->se.sum_exec_runtime; 5393 task_rq_unlock(rq, p, &rf); 5394 5395 return ns; 5396 } 5397 5398 #ifdef CONFIG_SCHED_DEBUG 5399 static u64 cpu_resched_latency(struct rq *rq) 5400 { 5401 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms); 5402 u64 resched_latency, now = rq_clock(rq); 5403 static bool warned_once; 5404 5405 if (sysctl_resched_latency_warn_once && warned_once) 5406 return 0; 5407 5408 if (!need_resched() || !latency_warn_ms) 5409 return 0; 5410 5411 if (system_state == SYSTEM_BOOTING) 5412 return 0; 5413 5414 if (!rq->last_seen_need_resched_ns) { 5415 rq->last_seen_need_resched_ns = now; 5416 rq->ticks_without_resched = 0; 5417 return 0; 5418 } 5419 5420 rq->ticks_without_resched++; 5421 resched_latency = now - rq->last_seen_need_resched_ns; 5422 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC) 5423 return 0; 5424 5425 warned_once = true; 5426 5427 return resched_latency; 5428 } 5429 5430 static int __init setup_resched_latency_warn_ms(char *str) 5431 { 5432 long val; 5433 5434 if ((kstrtol(str, 0, &val))) { 5435 pr_warn("Unable to set resched_latency_warn_ms\n"); 5436 return 1; 5437 } 5438 5439 sysctl_resched_latency_warn_ms = val; 5440 return 1; 5441 } 5442 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms); 5443 #else 5444 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; } 5445 #endif /* CONFIG_SCHED_DEBUG */ 5446 5447 /* 5448 * This function gets called by the timer code, with HZ frequency. 5449 * We call it with interrupts disabled. 5450 */ 5451 void scheduler_tick(void) 5452 { 5453 int cpu = smp_processor_id(); 5454 struct rq *rq = cpu_rq(cpu); 5455 struct task_struct *curr = rq->curr; 5456 struct rq_flags rf; 5457 unsigned long thermal_pressure; 5458 u64 resched_latency; 5459 5460 arch_scale_freq_tick(); 5461 sched_clock_tick(); 5462 5463 rq_lock(rq, &rf); 5464 5465 update_rq_clock(rq); 5466 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); 5467 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure); 5468 curr->sched_class->task_tick(rq, curr, 0); 5469 if (sched_feat(LATENCY_WARN)) 5470 resched_latency = cpu_resched_latency(rq); 5471 calc_global_load_tick(rq); 5472 sched_core_tick(rq); 5473 5474 rq_unlock(rq, &rf); 5475 5476 if (sched_feat(LATENCY_WARN) && resched_latency) 5477 resched_latency_warn(cpu, resched_latency); 5478 5479 perf_event_task_tick(); 5480 5481 #ifdef CONFIG_SMP 5482 rq->idle_balance = idle_cpu(cpu); 5483 trigger_load_balance(rq); 5484 #endif 5485 } 5486 5487 #ifdef CONFIG_NO_HZ_FULL 5488 5489 struct tick_work { 5490 int cpu; 5491 atomic_t state; 5492 struct delayed_work work; 5493 }; 5494 /* Values for ->state, see diagram below. */ 5495 #define TICK_SCHED_REMOTE_OFFLINE 0 5496 #define TICK_SCHED_REMOTE_OFFLINING 1 5497 #define TICK_SCHED_REMOTE_RUNNING 2 5498 5499 /* 5500 * State diagram for ->state: 5501 * 5502 * 5503 * TICK_SCHED_REMOTE_OFFLINE 5504 * | ^ 5505 * | | 5506 * | | sched_tick_remote() 5507 * | | 5508 * | | 5509 * +--TICK_SCHED_REMOTE_OFFLINING 5510 * | ^ 5511 * | | 5512 * sched_tick_start() | | sched_tick_stop() 5513 * | | 5514 * V | 5515 * TICK_SCHED_REMOTE_RUNNING 5516 * 5517 * 5518 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote() 5519 * and sched_tick_start() are happy to leave the state in RUNNING. 5520 */ 5521 5522 static struct tick_work __percpu *tick_work_cpu; 5523 5524 static void sched_tick_remote(struct work_struct *work) 5525 { 5526 struct delayed_work *dwork = to_delayed_work(work); 5527 struct tick_work *twork = container_of(dwork, struct tick_work, work); 5528 int cpu = twork->cpu; 5529 struct rq *rq = cpu_rq(cpu); 5530 struct task_struct *curr; 5531 struct rq_flags rf; 5532 u64 delta; 5533 int os; 5534 5535 /* 5536 * Handle the tick only if it appears the remote CPU is running in full 5537 * dynticks mode. The check is racy by nature, but missing a tick or 5538 * having one too much is no big deal because the scheduler tick updates 5539 * statistics and checks timeslices in a time-independent way, regardless 5540 * of when exactly it is running. 5541 */ 5542 if (!tick_nohz_tick_stopped_cpu(cpu)) 5543 goto out_requeue; 5544 5545 rq_lock_irq(rq, &rf); 5546 curr = rq->curr; 5547 if (cpu_is_offline(cpu)) 5548 goto out_unlock; 5549 5550 update_rq_clock(rq); 5551 5552 if (!is_idle_task(curr)) { 5553 /* 5554 * Make sure the next tick runs within a reasonable 5555 * amount of time. 5556 */ 5557 delta = rq_clock_task(rq) - curr->se.exec_start; 5558 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3); 5559 } 5560 curr->sched_class->task_tick(rq, curr, 0); 5561 5562 calc_load_nohz_remote(rq); 5563 out_unlock: 5564 rq_unlock_irq(rq, &rf); 5565 out_requeue: 5566 5567 /* 5568 * Run the remote tick once per second (1Hz). This arbitrary 5569 * frequency is large enough to avoid overload but short enough 5570 * to keep scheduler internal stats reasonably up to date. But 5571 * first update state to reflect hotplug activity if required. 5572 */ 5573 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING); 5574 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE); 5575 if (os == TICK_SCHED_REMOTE_RUNNING) 5576 queue_delayed_work(system_unbound_wq, dwork, HZ); 5577 } 5578 5579 static void sched_tick_start(int cpu) 5580 { 5581 int os; 5582 struct tick_work *twork; 5583 5584 if (housekeeping_cpu(cpu, HK_TYPE_TICK)) 5585 return; 5586 5587 WARN_ON_ONCE(!tick_work_cpu); 5588 5589 twork = per_cpu_ptr(tick_work_cpu, cpu); 5590 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING); 5591 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING); 5592 if (os == TICK_SCHED_REMOTE_OFFLINE) { 5593 twork->cpu = cpu; 5594 INIT_DELAYED_WORK(&twork->work, sched_tick_remote); 5595 queue_delayed_work(system_unbound_wq, &twork->work, HZ); 5596 } 5597 } 5598 5599 #ifdef CONFIG_HOTPLUG_CPU 5600 static void sched_tick_stop(int cpu) 5601 { 5602 struct tick_work *twork; 5603 int os; 5604 5605 if (housekeeping_cpu(cpu, HK_TYPE_TICK)) 5606 return; 5607 5608 WARN_ON_ONCE(!tick_work_cpu); 5609 5610 twork = per_cpu_ptr(tick_work_cpu, cpu); 5611 /* There cannot be competing actions, but don't rely on stop-machine. */ 5612 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING); 5613 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING); 5614 /* Don't cancel, as this would mess up the state machine. */ 5615 } 5616 #endif /* CONFIG_HOTPLUG_CPU */ 5617 5618 int __init sched_tick_offload_init(void) 5619 { 5620 tick_work_cpu = alloc_percpu(struct tick_work); 5621 BUG_ON(!tick_work_cpu); 5622 return 0; 5623 } 5624 5625 #else /* !CONFIG_NO_HZ_FULL */ 5626 static inline void sched_tick_start(int cpu) { } 5627 static inline void sched_tick_stop(int cpu) { } 5628 #endif 5629 5630 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \ 5631 defined(CONFIG_TRACE_PREEMPT_TOGGLE)) 5632 /* 5633 * If the value passed in is equal to the current preempt count 5634 * then we just disabled preemption. Start timing the latency. 5635 */ 5636 static inline void preempt_latency_start(int val) 5637 { 5638 if (preempt_count() == val) { 5639 unsigned long ip = get_lock_parent_ip(); 5640 #ifdef CONFIG_DEBUG_PREEMPT 5641 current->preempt_disable_ip = ip; 5642 #endif 5643 trace_preempt_off(CALLER_ADDR0, ip); 5644 } 5645 } 5646 5647 void preempt_count_add(int val) 5648 { 5649 #ifdef CONFIG_DEBUG_PREEMPT 5650 /* 5651 * Underflow? 5652 */ 5653 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 5654 return; 5655 #endif 5656 __preempt_count_add(val); 5657 #ifdef CONFIG_DEBUG_PREEMPT 5658 /* 5659 * Spinlock count overflowing soon? 5660 */ 5661 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 5662 PREEMPT_MASK - 10); 5663 #endif 5664 preempt_latency_start(val); 5665 } 5666 EXPORT_SYMBOL(preempt_count_add); 5667 NOKPROBE_SYMBOL(preempt_count_add); 5668 5669 /* 5670 * If the value passed in equals to the current preempt count 5671 * then we just enabled preemption. Stop timing the latency. 5672 */ 5673 static inline void preempt_latency_stop(int val) 5674 { 5675 if (preempt_count() == val) 5676 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip()); 5677 } 5678 5679 void preempt_count_sub(int val) 5680 { 5681 #ifdef CONFIG_DEBUG_PREEMPT 5682 /* 5683 * Underflow? 5684 */ 5685 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 5686 return; 5687 /* 5688 * Is the spinlock portion underflowing? 5689 */ 5690 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 5691 !(preempt_count() & PREEMPT_MASK))) 5692 return; 5693 #endif 5694 5695 preempt_latency_stop(val); 5696 __preempt_count_sub(val); 5697 } 5698 EXPORT_SYMBOL(preempt_count_sub); 5699 NOKPROBE_SYMBOL(preempt_count_sub); 5700 5701 #else 5702 static inline void preempt_latency_start(int val) { } 5703 static inline void preempt_latency_stop(int val) { } 5704 #endif 5705 5706 static inline unsigned long get_preempt_disable_ip(struct task_struct *p) 5707 { 5708 #ifdef CONFIG_DEBUG_PREEMPT 5709 return p->preempt_disable_ip; 5710 #else 5711 return 0; 5712 #endif 5713 } 5714 5715 /* 5716 * Print scheduling while atomic bug: 5717 */ 5718 static noinline void __schedule_bug(struct task_struct *prev) 5719 { 5720 /* Save this before calling printk(), since that will clobber it */ 5721 unsigned long preempt_disable_ip = get_preempt_disable_ip(current); 5722 5723 if (oops_in_progress) 5724 return; 5725 5726 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 5727 prev->comm, prev->pid, preempt_count()); 5728 5729 debug_show_held_locks(prev); 5730 print_modules(); 5731 if (irqs_disabled()) 5732 print_irqtrace_events(prev); 5733 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT) 5734 && in_atomic_preempt_off()) { 5735 pr_err("Preemption disabled at:"); 5736 print_ip_sym(KERN_ERR, preempt_disable_ip); 5737 } 5738 if (panic_on_warn) 5739 panic("scheduling while atomic\n"); 5740 5741 dump_stack(); 5742 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 5743 } 5744 5745 /* 5746 * Various schedule()-time debugging checks and statistics: 5747 */ 5748 static inline void schedule_debug(struct task_struct *prev, bool preempt) 5749 { 5750 #ifdef CONFIG_SCHED_STACK_END_CHECK 5751 if (task_stack_end_corrupted(prev)) 5752 panic("corrupted stack end detected inside scheduler\n"); 5753 5754 if (task_scs_end_corrupted(prev)) 5755 panic("corrupted shadow stack detected inside scheduler\n"); 5756 #endif 5757 5758 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP 5759 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) { 5760 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n", 5761 prev->comm, prev->pid, prev->non_block_count); 5762 dump_stack(); 5763 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 5764 } 5765 #endif 5766 5767 if (unlikely(in_atomic_preempt_off())) { 5768 __schedule_bug(prev); 5769 preempt_count_set(PREEMPT_DISABLED); 5770 } 5771 rcu_sleep_check(); 5772 SCHED_WARN_ON(ct_state() == CONTEXT_USER); 5773 5774 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 5775 5776 schedstat_inc(this_rq()->sched_count); 5777 } 5778 5779 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev, 5780 struct rq_flags *rf) 5781 { 5782 #ifdef CONFIG_SMP 5783 const struct sched_class *class; 5784 /* 5785 * We must do the balancing pass before put_prev_task(), such 5786 * that when we release the rq->lock the task is in the same 5787 * state as before we took rq->lock. 5788 * 5789 * We can terminate the balance pass as soon as we know there is 5790 * a runnable task of @class priority or higher. 5791 */ 5792 for_class_range(class, prev->sched_class, &idle_sched_class) { 5793 if (class->balance(rq, prev, rf)) 5794 break; 5795 } 5796 #endif 5797 5798 put_prev_task(rq, prev); 5799 } 5800 5801 /* 5802 * Pick up the highest-prio task: 5803 */ 5804 static inline struct task_struct * 5805 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 5806 { 5807 const struct sched_class *class; 5808 struct task_struct *p; 5809 5810 /* 5811 * Optimization: we know that if all tasks are in the fair class we can 5812 * call that function directly, but only if the @prev task wasn't of a 5813 * higher scheduling class, because otherwise those lose the 5814 * opportunity to pull in more work from other CPUs. 5815 */ 5816 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) && 5817 rq->nr_running == rq->cfs.h_nr_running)) { 5818 5819 p = pick_next_task_fair(rq, prev, rf); 5820 if (unlikely(p == RETRY_TASK)) 5821 goto restart; 5822 5823 /* Assume the next prioritized class is idle_sched_class */ 5824 if (!p) { 5825 put_prev_task(rq, prev); 5826 p = pick_next_task_idle(rq); 5827 } 5828 5829 return p; 5830 } 5831 5832 restart: 5833 put_prev_task_balance(rq, prev, rf); 5834 5835 for_each_class(class) { 5836 p = class->pick_next_task(rq); 5837 if (p) 5838 return p; 5839 } 5840 5841 BUG(); /* The idle class should always have a runnable task. */ 5842 } 5843 5844 #ifdef CONFIG_SCHED_CORE 5845 static inline bool is_task_rq_idle(struct task_struct *t) 5846 { 5847 return (task_rq(t)->idle == t); 5848 } 5849 5850 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie) 5851 { 5852 return is_task_rq_idle(a) || (a->core_cookie == cookie); 5853 } 5854 5855 static inline bool cookie_match(struct task_struct *a, struct task_struct *b) 5856 { 5857 if (is_task_rq_idle(a) || is_task_rq_idle(b)) 5858 return true; 5859 5860 return a->core_cookie == b->core_cookie; 5861 } 5862 5863 static inline struct task_struct *pick_task(struct rq *rq) 5864 { 5865 const struct sched_class *class; 5866 struct task_struct *p; 5867 5868 for_each_class(class) { 5869 p = class->pick_task(rq); 5870 if (p) 5871 return p; 5872 } 5873 5874 BUG(); /* The idle class should always have a runnable task. */ 5875 } 5876 5877 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi); 5878 5879 static void queue_core_balance(struct rq *rq); 5880 5881 static struct task_struct * 5882 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 5883 { 5884 struct task_struct *next, *p, *max = NULL; 5885 const struct cpumask *smt_mask; 5886 bool fi_before = false; 5887 bool core_clock_updated = (rq == rq->core); 5888 unsigned long cookie; 5889 int i, cpu, occ = 0; 5890 struct rq *rq_i; 5891 bool need_sync; 5892 5893 if (!sched_core_enabled(rq)) 5894 return __pick_next_task(rq, prev, rf); 5895 5896 cpu = cpu_of(rq); 5897 5898 /* Stopper task is switching into idle, no need core-wide selection. */ 5899 if (cpu_is_offline(cpu)) { 5900 /* 5901 * Reset core_pick so that we don't enter the fastpath when 5902 * coming online. core_pick would already be migrated to 5903 * another cpu during offline. 5904 */ 5905 rq->core_pick = NULL; 5906 return __pick_next_task(rq, prev, rf); 5907 } 5908 5909 /* 5910 * If there were no {en,de}queues since we picked (IOW, the task 5911 * pointers are all still valid), and we haven't scheduled the last 5912 * pick yet, do so now. 5913 * 5914 * rq->core_pick can be NULL if no selection was made for a CPU because 5915 * it was either offline or went offline during a sibling's core-wide 5916 * selection. In this case, do a core-wide selection. 5917 */ 5918 if (rq->core->core_pick_seq == rq->core->core_task_seq && 5919 rq->core->core_pick_seq != rq->core_sched_seq && 5920 rq->core_pick) { 5921 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq); 5922 5923 next = rq->core_pick; 5924 if (next != prev) { 5925 put_prev_task(rq, prev); 5926 set_next_task(rq, next); 5927 } 5928 5929 rq->core_pick = NULL; 5930 goto out; 5931 } 5932 5933 put_prev_task_balance(rq, prev, rf); 5934 5935 smt_mask = cpu_smt_mask(cpu); 5936 need_sync = !!rq->core->core_cookie; 5937 5938 /* reset state */ 5939 rq->core->core_cookie = 0UL; 5940 if (rq->core->core_forceidle_count) { 5941 if (!core_clock_updated) { 5942 update_rq_clock(rq->core); 5943 core_clock_updated = true; 5944 } 5945 sched_core_account_forceidle(rq); 5946 /* reset after accounting force idle */ 5947 rq->core->core_forceidle_start = 0; 5948 rq->core->core_forceidle_count = 0; 5949 rq->core->core_forceidle_occupation = 0; 5950 need_sync = true; 5951 fi_before = true; 5952 } 5953 5954 /* 5955 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq 5956 * 5957 * @task_seq guards the task state ({en,de}queues) 5958 * @pick_seq is the @task_seq we did a selection on 5959 * @sched_seq is the @pick_seq we scheduled 5960 * 5961 * However, preemptions can cause multiple picks on the same task set. 5962 * 'Fix' this by also increasing @task_seq for every pick. 5963 */ 5964 rq->core->core_task_seq++; 5965 5966 /* 5967 * Optimize for common case where this CPU has no cookies 5968 * and there are no cookied tasks running on siblings. 5969 */ 5970 if (!need_sync) { 5971 next = pick_task(rq); 5972 if (!next->core_cookie) { 5973 rq->core_pick = NULL; 5974 /* 5975 * For robustness, update the min_vruntime_fi for 5976 * unconstrained picks as well. 5977 */ 5978 WARN_ON_ONCE(fi_before); 5979 task_vruntime_update(rq, next, false); 5980 goto out_set_next; 5981 } 5982 } 5983 5984 /* 5985 * For each thread: do the regular task pick and find the max prio task 5986 * amongst them. 5987 * 5988 * Tie-break prio towards the current CPU 5989 */ 5990 for_each_cpu_wrap(i, smt_mask, cpu) { 5991 rq_i = cpu_rq(i); 5992 5993 /* 5994 * Current cpu always has its clock updated on entrance to 5995 * pick_next_task(). If the current cpu is not the core, 5996 * the core may also have been updated above. 5997 */ 5998 if (i != cpu && (rq_i != rq->core || !core_clock_updated)) 5999 update_rq_clock(rq_i); 6000 6001 p = rq_i->core_pick = pick_task(rq_i); 6002 if (!max || prio_less(max, p, fi_before)) 6003 max = p; 6004 } 6005 6006 cookie = rq->core->core_cookie = max->core_cookie; 6007 6008 /* 6009 * For each thread: try and find a runnable task that matches @max or 6010 * force idle. 6011 */ 6012 for_each_cpu(i, smt_mask) { 6013 rq_i = cpu_rq(i); 6014 p = rq_i->core_pick; 6015 6016 if (!cookie_equals(p, cookie)) { 6017 p = NULL; 6018 if (cookie) 6019 p = sched_core_find(rq_i, cookie); 6020 if (!p) 6021 p = idle_sched_class.pick_task(rq_i); 6022 } 6023 6024 rq_i->core_pick = p; 6025 6026 if (p == rq_i->idle) { 6027 if (rq_i->nr_running) { 6028 rq->core->core_forceidle_count++; 6029 if (!fi_before) 6030 rq->core->core_forceidle_seq++; 6031 } 6032 } else { 6033 occ++; 6034 } 6035 } 6036 6037 if (schedstat_enabled() && rq->core->core_forceidle_count) { 6038 rq->core->core_forceidle_start = rq_clock(rq->core); 6039 rq->core->core_forceidle_occupation = occ; 6040 } 6041 6042 rq->core->core_pick_seq = rq->core->core_task_seq; 6043 next = rq->core_pick; 6044 rq->core_sched_seq = rq->core->core_pick_seq; 6045 6046 /* Something should have been selected for current CPU */ 6047 WARN_ON_ONCE(!next); 6048 6049 /* 6050 * Reschedule siblings 6051 * 6052 * NOTE: L1TF -- at this point we're no longer running the old task and 6053 * sending an IPI (below) ensures the sibling will no longer be running 6054 * their task. This ensures there is no inter-sibling overlap between 6055 * non-matching user state. 6056 */ 6057 for_each_cpu(i, smt_mask) { 6058 rq_i = cpu_rq(i); 6059 6060 /* 6061 * An online sibling might have gone offline before a task 6062 * could be picked for it, or it might be offline but later 6063 * happen to come online, but its too late and nothing was 6064 * picked for it. That's Ok - it will pick tasks for itself, 6065 * so ignore it. 6066 */ 6067 if (!rq_i->core_pick) 6068 continue; 6069 6070 /* 6071 * Update for new !FI->FI transitions, or if continuing to be in !FI: 6072 * fi_before fi update? 6073 * 0 0 1 6074 * 0 1 1 6075 * 1 0 1 6076 * 1 1 0 6077 */ 6078 if (!(fi_before && rq->core->core_forceidle_count)) 6079 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count); 6080 6081 rq_i->core_pick->core_occupation = occ; 6082 6083 if (i == cpu) { 6084 rq_i->core_pick = NULL; 6085 continue; 6086 } 6087 6088 /* Did we break L1TF mitigation requirements? */ 6089 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick)); 6090 6091 if (rq_i->curr == rq_i->core_pick) { 6092 rq_i->core_pick = NULL; 6093 continue; 6094 } 6095 6096 resched_curr(rq_i); 6097 } 6098 6099 out_set_next: 6100 set_next_task(rq, next); 6101 out: 6102 if (rq->core->core_forceidle_count && next == rq->idle) 6103 queue_core_balance(rq); 6104 6105 return next; 6106 } 6107 6108 static bool try_steal_cookie(int this, int that) 6109 { 6110 struct rq *dst = cpu_rq(this), *src = cpu_rq(that); 6111 struct task_struct *p; 6112 unsigned long cookie; 6113 bool success = false; 6114 6115 local_irq_disable(); 6116 double_rq_lock(dst, src); 6117 6118 cookie = dst->core->core_cookie; 6119 if (!cookie) 6120 goto unlock; 6121 6122 if (dst->curr != dst->idle) 6123 goto unlock; 6124 6125 p = sched_core_find(src, cookie); 6126 if (p == src->idle) 6127 goto unlock; 6128 6129 do { 6130 if (p == src->core_pick || p == src->curr) 6131 goto next; 6132 6133 if (!is_cpu_allowed(p, this)) 6134 goto next; 6135 6136 if (p->core_occupation > dst->idle->core_occupation) 6137 goto next; 6138 6139 deactivate_task(src, p, 0); 6140 set_task_cpu(p, this); 6141 activate_task(dst, p, 0); 6142 6143 resched_curr(dst); 6144 6145 success = true; 6146 break; 6147 6148 next: 6149 p = sched_core_next(p, cookie); 6150 } while (p); 6151 6152 unlock: 6153 double_rq_unlock(dst, src); 6154 local_irq_enable(); 6155 6156 return success; 6157 } 6158 6159 static bool steal_cookie_task(int cpu, struct sched_domain *sd) 6160 { 6161 int i; 6162 6163 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) { 6164 if (i == cpu) 6165 continue; 6166 6167 if (need_resched()) 6168 break; 6169 6170 if (try_steal_cookie(cpu, i)) 6171 return true; 6172 } 6173 6174 return false; 6175 } 6176 6177 static void sched_core_balance(struct rq *rq) 6178 { 6179 struct sched_domain *sd; 6180 int cpu = cpu_of(rq); 6181 6182 preempt_disable(); 6183 rcu_read_lock(); 6184 raw_spin_rq_unlock_irq(rq); 6185 for_each_domain(cpu, sd) { 6186 if (need_resched()) 6187 break; 6188 6189 if (steal_cookie_task(cpu, sd)) 6190 break; 6191 } 6192 raw_spin_rq_lock_irq(rq); 6193 rcu_read_unlock(); 6194 preempt_enable(); 6195 } 6196 6197 static DEFINE_PER_CPU(struct callback_head, core_balance_head); 6198 6199 static void queue_core_balance(struct rq *rq) 6200 { 6201 if (!sched_core_enabled(rq)) 6202 return; 6203 6204 if (!rq->core->core_cookie) 6205 return; 6206 6207 if (!rq->nr_running) /* not forced idle */ 6208 return; 6209 6210 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance); 6211 } 6212 6213 static void sched_core_cpu_starting(unsigned int cpu) 6214 { 6215 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 6216 struct rq *rq = cpu_rq(cpu), *core_rq = NULL; 6217 unsigned long flags; 6218 int t; 6219 6220 sched_core_lock(cpu, &flags); 6221 6222 WARN_ON_ONCE(rq->core != rq); 6223 6224 /* if we're the first, we'll be our own leader */ 6225 if (cpumask_weight(smt_mask) == 1) 6226 goto unlock; 6227 6228 /* find the leader */ 6229 for_each_cpu(t, smt_mask) { 6230 if (t == cpu) 6231 continue; 6232 rq = cpu_rq(t); 6233 if (rq->core == rq) { 6234 core_rq = rq; 6235 break; 6236 } 6237 } 6238 6239 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */ 6240 goto unlock; 6241 6242 /* install and validate core_rq */ 6243 for_each_cpu(t, smt_mask) { 6244 rq = cpu_rq(t); 6245 6246 if (t == cpu) 6247 rq->core = core_rq; 6248 6249 WARN_ON_ONCE(rq->core != core_rq); 6250 } 6251 6252 unlock: 6253 sched_core_unlock(cpu, &flags); 6254 } 6255 6256 static void sched_core_cpu_deactivate(unsigned int cpu) 6257 { 6258 const struct cpumask *smt_mask = cpu_smt_mask(cpu); 6259 struct rq *rq = cpu_rq(cpu), *core_rq = NULL; 6260 unsigned long flags; 6261 int t; 6262 6263 sched_core_lock(cpu, &flags); 6264 6265 /* if we're the last man standing, nothing to do */ 6266 if (cpumask_weight(smt_mask) == 1) { 6267 WARN_ON_ONCE(rq->core != rq); 6268 goto unlock; 6269 } 6270 6271 /* if we're not the leader, nothing to do */ 6272 if (rq->core != rq) 6273 goto unlock; 6274 6275 /* find a new leader */ 6276 for_each_cpu(t, smt_mask) { 6277 if (t == cpu) 6278 continue; 6279 core_rq = cpu_rq(t); 6280 break; 6281 } 6282 6283 if (WARN_ON_ONCE(!core_rq)) /* impossible */ 6284 goto unlock; 6285 6286 /* copy the shared state to the new leader */ 6287 core_rq->core_task_seq = rq->core_task_seq; 6288 core_rq->core_pick_seq = rq->core_pick_seq; 6289 core_rq->core_cookie = rq->core_cookie; 6290 core_rq->core_forceidle_count = rq->core_forceidle_count; 6291 core_rq->core_forceidle_seq = rq->core_forceidle_seq; 6292 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation; 6293 6294 /* 6295 * Accounting edge for forced idle is handled in pick_next_task(). 6296 * Don't need another one here, since the hotplug thread shouldn't 6297 * have a cookie. 6298 */ 6299 core_rq->core_forceidle_start = 0; 6300 6301 /* install new leader */ 6302 for_each_cpu(t, smt_mask) { 6303 rq = cpu_rq(t); 6304 rq->core = core_rq; 6305 } 6306 6307 unlock: 6308 sched_core_unlock(cpu, &flags); 6309 } 6310 6311 static inline void sched_core_cpu_dying(unsigned int cpu) 6312 { 6313 struct rq *rq = cpu_rq(cpu); 6314 6315 if (rq->core != rq) 6316 rq->core = rq; 6317 } 6318 6319 #else /* !CONFIG_SCHED_CORE */ 6320 6321 static inline void sched_core_cpu_starting(unsigned int cpu) {} 6322 static inline void sched_core_cpu_deactivate(unsigned int cpu) {} 6323 static inline void sched_core_cpu_dying(unsigned int cpu) {} 6324 6325 static struct task_struct * 6326 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 6327 { 6328 return __pick_next_task(rq, prev, rf); 6329 } 6330 6331 #endif /* CONFIG_SCHED_CORE */ 6332 6333 /* 6334 * Constants for the sched_mode argument of __schedule(). 6335 * 6336 * The mode argument allows RT enabled kernels to differentiate a 6337 * preemption from blocking on an 'sleeping' spin/rwlock. Note that 6338 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to 6339 * optimize the AND operation out and just check for zero. 6340 */ 6341 #define SM_NONE 0x0 6342 #define SM_PREEMPT 0x1 6343 #define SM_RTLOCK_WAIT 0x2 6344 6345 #ifndef CONFIG_PREEMPT_RT 6346 # define SM_MASK_PREEMPT (~0U) 6347 #else 6348 # define SM_MASK_PREEMPT SM_PREEMPT 6349 #endif 6350 6351 /* 6352 * __schedule() is the main scheduler function. 6353 * 6354 * The main means of driving the scheduler and thus entering this function are: 6355 * 6356 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 6357 * 6358 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 6359 * paths. For example, see arch/x86/entry_64.S. 6360 * 6361 * To drive preemption between tasks, the scheduler sets the flag in timer 6362 * interrupt handler scheduler_tick(). 6363 * 6364 * 3. Wakeups don't really cause entry into schedule(). They add a 6365 * task to the run-queue and that's it. 6366 * 6367 * Now, if the new task added to the run-queue preempts the current 6368 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 6369 * called on the nearest possible occasion: 6370 * 6371 * - If the kernel is preemptible (CONFIG_PREEMPTION=y): 6372 * 6373 * - in syscall or exception context, at the next outmost 6374 * preempt_enable(). (this might be as soon as the wake_up()'s 6375 * spin_unlock()!) 6376 * 6377 * - in IRQ context, return from interrupt-handler to 6378 * preemptible context 6379 * 6380 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set) 6381 * then at the next: 6382 * 6383 * - cond_resched() call 6384 * - explicit schedule() call 6385 * - return from syscall or exception to user-space 6386 * - return from interrupt-handler to user-space 6387 * 6388 * WARNING: must be called with preemption disabled! 6389 */ 6390 static void __sched notrace __schedule(unsigned int sched_mode) 6391 { 6392 struct task_struct *prev, *next; 6393 unsigned long *switch_count; 6394 unsigned long prev_state; 6395 struct rq_flags rf; 6396 struct rq *rq; 6397 int cpu; 6398 6399 cpu = smp_processor_id(); 6400 rq = cpu_rq(cpu); 6401 prev = rq->curr; 6402 6403 schedule_debug(prev, !!sched_mode); 6404 6405 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL)) 6406 hrtick_clear(rq); 6407 6408 local_irq_disable(); 6409 rcu_note_context_switch(!!sched_mode); 6410 6411 /* 6412 * Make sure that signal_pending_state()->signal_pending() below 6413 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 6414 * done by the caller to avoid the race with signal_wake_up(): 6415 * 6416 * __set_current_state(@state) signal_wake_up() 6417 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING) 6418 * wake_up_state(p, state) 6419 * LOCK rq->lock LOCK p->pi_state 6420 * smp_mb__after_spinlock() smp_mb__after_spinlock() 6421 * if (signal_pending_state()) if (p->state & @state) 6422 * 6423 * Also, the membarrier system call requires a full memory barrier 6424 * after coming from user-space, before storing to rq->curr. 6425 */ 6426 rq_lock(rq, &rf); 6427 smp_mb__after_spinlock(); 6428 6429 /* Promote REQ to ACT */ 6430 rq->clock_update_flags <<= 1; 6431 update_rq_clock(rq); 6432 6433 switch_count = &prev->nivcsw; 6434 6435 /* 6436 * We must load prev->state once (task_struct::state is volatile), such 6437 * that we form a control dependency vs deactivate_task() below. 6438 */ 6439 prev_state = READ_ONCE(prev->__state); 6440 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) { 6441 if (signal_pending_state(prev_state, prev)) { 6442 WRITE_ONCE(prev->__state, TASK_RUNNING); 6443 } else { 6444 prev->sched_contributes_to_load = 6445 (prev_state & TASK_UNINTERRUPTIBLE) && 6446 !(prev_state & TASK_NOLOAD) && 6447 !(prev->flags & PF_FROZEN); 6448 6449 if (prev->sched_contributes_to_load) 6450 rq->nr_uninterruptible++; 6451 6452 /* 6453 * __schedule() ttwu() 6454 * prev_state = prev->state; if (p->on_rq && ...) 6455 * if (prev_state) goto out; 6456 * p->on_rq = 0; smp_acquire__after_ctrl_dep(); 6457 * p->state = TASK_WAKING 6458 * 6459 * Where __schedule() and ttwu() have matching control dependencies. 6460 * 6461 * After this, schedule() must not care about p->state any more. 6462 */ 6463 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK); 6464 6465 if (prev->in_iowait) { 6466 atomic_inc(&rq->nr_iowait); 6467 delayacct_blkio_start(); 6468 } 6469 } 6470 switch_count = &prev->nvcsw; 6471 } 6472 6473 next = pick_next_task(rq, prev, &rf); 6474 clear_tsk_need_resched(prev); 6475 clear_preempt_need_resched(); 6476 #ifdef CONFIG_SCHED_DEBUG 6477 rq->last_seen_need_resched_ns = 0; 6478 #endif 6479 6480 if (likely(prev != next)) { 6481 rq->nr_switches++; 6482 /* 6483 * RCU users of rcu_dereference(rq->curr) may not see 6484 * changes to task_struct made by pick_next_task(). 6485 */ 6486 RCU_INIT_POINTER(rq->curr, next); 6487 /* 6488 * The membarrier system call requires each architecture 6489 * to have a full memory barrier after updating 6490 * rq->curr, before returning to user-space. 6491 * 6492 * Here are the schemes providing that barrier on the 6493 * various architectures: 6494 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC. 6495 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC. 6496 * - finish_lock_switch() for weakly-ordered 6497 * architectures where spin_unlock is a full barrier, 6498 * - switch_to() for arm64 (weakly-ordered, spin_unlock 6499 * is a RELEASE barrier), 6500 */ 6501 ++*switch_count; 6502 6503 migrate_disable_switch(rq, prev); 6504 psi_sched_switch(prev, next, !task_on_rq_queued(prev)); 6505 6506 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state); 6507 6508 /* Also unlocks the rq: */ 6509 rq = context_switch(rq, prev, next, &rf); 6510 } else { 6511 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP); 6512 6513 rq_unpin_lock(rq, &rf); 6514 __balance_callbacks(rq); 6515 raw_spin_rq_unlock_irq(rq); 6516 } 6517 } 6518 6519 void __noreturn do_task_dead(void) 6520 { 6521 /* Causes final put_task_struct in finish_task_switch(): */ 6522 set_special_state(TASK_DEAD); 6523 6524 /* Tell freezer to ignore us: */ 6525 current->flags |= PF_NOFREEZE; 6526 6527 __schedule(SM_NONE); 6528 BUG(); 6529 6530 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */ 6531 for (;;) 6532 cpu_relax(); 6533 } 6534 6535 static inline void sched_submit_work(struct task_struct *tsk) 6536 { 6537 unsigned int task_flags; 6538 6539 if (task_is_running(tsk)) 6540 return; 6541 6542 task_flags = tsk->flags; 6543 /* 6544 * If a worker goes to sleep, notify and ask workqueue whether it 6545 * wants to wake up a task to maintain concurrency. 6546 */ 6547 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) { 6548 if (task_flags & PF_WQ_WORKER) 6549 wq_worker_sleeping(tsk); 6550 else 6551 io_wq_worker_sleeping(tsk); 6552 } 6553 6554 /* 6555 * spinlock and rwlock must not flush block requests. This will 6556 * deadlock if the callback attempts to acquire a lock which is 6557 * already acquired. 6558 */ 6559 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT); 6560 6561 /* 6562 * If we are going to sleep and we have plugged IO queued, 6563 * make sure to submit it to avoid deadlocks. 6564 */ 6565 blk_flush_plug(tsk->plug, true); 6566 } 6567 6568 static void sched_update_worker(struct task_struct *tsk) 6569 { 6570 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) { 6571 if (tsk->flags & PF_WQ_WORKER) 6572 wq_worker_running(tsk); 6573 else 6574 io_wq_worker_running(tsk); 6575 } 6576 } 6577 6578 asmlinkage __visible void __sched schedule(void) 6579 { 6580 struct task_struct *tsk = current; 6581 6582 sched_submit_work(tsk); 6583 do { 6584 preempt_disable(); 6585 __schedule(SM_NONE); 6586 sched_preempt_enable_no_resched(); 6587 } while (need_resched()); 6588 sched_update_worker(tsk); 6589 } 6590 EXPORT_SYMBOL(schedule); 6591 6592 /* 6593 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted 6594 * state (have scheduled out non-voluntarily) by making sure that all 6595 * tasks have either left the run queue or have gone into user space. 6596 * As idle tasks do not do either, they must not ever be preempted 6597 * (schedule out non-voluntarily). 6598 * 6599 * schedule_idle() is similar to schedule_preempt_disable() except that it 6600 * never enables preemption because it does not call sched_submit_work(). 6601 */ 6602 void __sched schedule_idle(void) 6603 { 6604 /* 6605 * As this skips calling sched_submit_work(), which the idle task does 6606 * regardless because that function is a nop when the task is in a 6607 * TASK_RUNNING state, make sure this isn't used someplace that the 6608 * current task can be in any other state. Note, idle is always in the 6609 * TASK_RUNNING state. 6610 */ 6611 WARN_ON_ONCE(current->__state); 6612 do { 6613 __schedule(SM_NONE); 6614 } while (need_resched()); 6615 } 6616 6617 #if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK) 6618 asmlinkage __visible void __sched schedule_user(void) 6619 { 6620 /* 6621 * If we come here after a random call to set_need_resched(), 6622 * or we have been woken up remotely but the IPI has not yet arrived, 6623 * we haven't yet exited the RCU idle mode. Do it here manually until 6624 * we find a better solution. 6625 * 6626 * NB: There are buggy callers of this function. Ideally we 6627 * should warn if prev_state != CONTEXT_USER, but that will trigger 6628 * too frequently to make sense yet. 6629 */ 6630 enum ctx_state prev_state = exception_enter(); 6631 schedule(); 6632 exception_exit(prev_state); 6633 } 6634 #endif 6635 6636 /** 6637 * schedule_preempt_disabled - called with preemption disabled 6638 * 6639 * Returns with preemption disabled. Note: preempt_count must be 1 6640 */ 6641 void __sched schedule_preempt_disabled(void) 6642 { 6643 sched_preempt_enable_no_resched(); 6644 schedule(); 6645 preempt_disable(); 6646 } 6647 6648 #ifdef CONFIG_PREEMPT_RT 6649 void __sched notrace schedule_rtlock(void) 6650 { 6651 do { 6652 preempt_disable(); 6653 __schedule(SM_RTLOCK_WAIT); 6654 sched_preempt_enable_no_resched(); 6655 } while (need_resched()); 6656 } 6657 NOKPROBE_SYMBOL(schedule_rtlock); 6658 #endif 6659 6660 static void __sched notrace preempt_schedule_common(void) 6661 { 6662 do { 6663 /* 6664 * Because the function tracer can trace preempt_count_sub() 6665 * and it also uses preempt_enable/disable_notrace(), if 6666 * NEED_RESCHED is set, the preempt_enable_notrace() called 6667 * by the function tracer will call this function again and 6668 * cause infinite recursion. 6669 * 6670 * Preemption must be disabled here before the function 6671 * tracer can trace. Break up preempt_disable() into two 6672 * calls. One to disable preemption without fear of being 6673 * traced. The other to still record the preemption latency, 6674 * which can also be traced by the function tracer. 6675 */ 6676 preempt_disable_notrace(); 6677 preempt_latency_start(1); 6678 __schedule(SM_PREEMPT); 6679 preempt_latency_stop(1); 6680 preempt_enable_no_resched_notrace(); 6681 6682 /* 6683 * Check again in case we missed a preemption opportunity 6684 * between schedule and now. 6685 */ 6686 } while (need_resched()); 6687 } 6688 6689 #ifdef CONFIG_PREEMPTION 6690 /* 6691 * This is the entry point to schedule() from in-kernel preemption 6692 * off of preempt_enable. 6693 */ 6694 asmlinkage __visible void __sched notrace preempt_schedule(void) 6695 { 6696 /* 6697 * If there is a non-zero preempt_count or interrupts are disabled, 6698 * we do not want to preempt the current task. Just return.. 6699 */ 6700 if (likely(!preemptible())) 6701 return; 6702 preempt_schedule_common(); 6703 } 6704 NOKPROBE_SYMBOL(preempt_schedule); 6705 EXPORT_SYMBOL(preempt_schedule); 6706 6707 #ifdef CONFIG_PREEMPT_DYNAMIC 6708 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL) 6709 #ifndef preempt_schedule_dynamic_enabled 6710 #define preempt_schedule_dynamic_enabled preempt_schedule 6711 #define preempt_schedule_dynamic_disabled NULL 6712 #endif 6713 DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled); 6714 EXPORT_STATIC_CALL_TRAMP(preempt_schedule); 6715 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY) 6716 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule); 6717 void __sched notrace dynamic_preempt_schedule(void) 6718 { 6719 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule)) 6720 return; 6721 preempt_schedule(); 6722 } 6723 NOKPROBE_SYMBOL(dynamic_preempt_schedule); 6724 EXPORT_SYMBOL(dynamic_preempt_schedule); 6725 #endif 6726 #endif 6727 6728 /** 6729 * preempt_schedule_notrace - preempt_schedule called by tracing 6730 * 6731 * The tracing infrastructure uses preempt_enable_notrace to prevent 6732 * recursion and tracing preempt enabling caused by the tracing 6733 * infrastructure itself. But as tracing can happen in areas coming 6734 * from userspace or just about to enter userspace, a preempt enable 6735 * can occur before user_exit() is called. This will cause the scheduler 6736 * to be called when the system is still in usermode. 6737 * 6738 * To prevent this, the preempt_enable_notrace will use this function 6739 * instead of preempt_schedule() to exit user context if needed before 6740 * calling the scheduler. 6741 */ 6742 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void) 6743 { 6744 enum ctx_state prev_ctx; 6745 6746 if (likely(!preemptible())) 6747 return; 6748 6749 do { 6750 /* 6751 * Because the function tracer can trace preempt_count_sub() 6752 * and it also uses preempt_enable/disable_notrace(), if 6753 * NEED_RESCHED is set, the preempt_enable_notrace() called 6754 * by the function tracer will call this function again and 6755 * cause infinite recursion. 6756 * 6757 * Preemption must be disabled here before the function 6758 * tracer can trace. Break up preempt_disable() into two 6759 * calls. One to disable preemption without fear of being 6760 * traced. The other to still record the preemption latency, 6761 * which can also be traced by the function tracer. 6762 */ 6763 preempt_disable_notrace(); 6764 preempt_latency_start(1); 6765 /* 6766 * Needs preempt disabled in case user_exit() is traced 6767 * and the tracer calls preempt_enable_notrace() causing 6768 * an infinite recursion. 6769 */ 6770 prev_ctx = exception_enter(); 6771 __schedule(SM_PREEMPT); 6772 exception_exit(prev_ctx); 6773 6774 preempt_latency_stop(1); 6775 preempt_enable_no_resched_notrace(); 6776 } while (need_resched()); 6777 } 6778 EXPORT_SYMBOL_GPL(preempt_schedule_notrace); 6779 6780 #ifdef CONFIG_PREEMPT_DYNAMIC 6781 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL) 6782 #ifndef preempt_schedule_notrace_dynamic_enabled 6783 #define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace 6784 #define preempt_schedule_notrace_dynamic_disabled NULL 6785 #endif 6786 DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled); 6787 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace); 6788 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY) 6789 static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace); 6790 void __sched notrace dynamic_preempt_schedule_notrace(void) 6791 { 6792 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace)) 6793 return; 6794 preempt_schedule_notrace(); 6795 } 6796 NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace); 6797 EXPORT_SYMBOL(dynamic_preempt_schedule_notrace); 6798 #endif 6799 #endif 6800 6801 #endif /* CONFIG_PREEMPTION */ 6802 6803 /* 6804 * This is the entry point to schedule() from kernel preemption 6805 * off of irq context. 6806 * Note, that this is called and return with irqs disabled. This will 6807 * protect us against recursive calling from irq. 6808 */ 6809 asmlinkage __visible void __sched preempt_schedule_irq(void) 6810 { 6811 enum ctx_state prev_state; 6812 6813 /* Catch callers which need to be fixed */ 6814 BUG_ON(preempt_count() || !irqs_disabled()); 6815 6816 prev_state = exception_enter(); 6817 6818 do { 6819 preempt_disable(); 6820 local_irq_enable(); 6821 __schedule(SM_PREEMPT); 6822 local_irq_disable(); 6823 sched_preempt_enable_no_resched(); 6824 } while (need_resched()); 6825 6826 exception_exit(prev_state); 6827 } 6828 6829 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags, 6830 void *key) 6831 { 6832 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC); 6833 return try_to_wake_up(curr->private, mode, wake_flags); 6834 } 6835 EXPORT_SYMBOL(default_wake_function); 6836 6837 static void __setscheduler_prio(struct task_struct *p, int prio) 6838 { 6839 if (dl_prio(prio)) 6840 p->sched_class = &dl_sched_class; 6841 else if (rt_prio(prio)) 6842 p->sched_class = &rt_sched_class; 6843 else 6844 p->sched_class = &fair_sched_class; 6845 6846 p->prio = prio; 6847 } 6848 6849 #ifdef CONFIG_RT_MUTEXES 6850 6851 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio) 6852 { 6853 if (pi_task) 6854 prio = min(prio, pi_task->prio); 6855 6856 return prio; 6857 } 6858 6859 static inline int rt_effective_prio(struct task_struct *p, int prio) 6860 { 6861 struct task_struct *pi_task = rt_mutex_get_top_task(p); 6862 6863 return __rt_effective_prio(pi_task, prio); 6864 } 6865 6866 /* 6867 * rt_mutex_setprio - set the current priority of a task 6868 * @p: task to boost 6869 * @pi_task: donor task 6870 * 6871 * This function changes the 'effective' priority of a task. It does 6872 * not touch ->normal_prio like __setscheduler(). 6873 * 6874 * Used by the rt_mutex code to implement priority inheritance 6875 * logic. Call site only calls if the priority of the task changed. 6876 */ 6877 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task) 6878 { 6879 int prio, oldprio, queued, running, queue_flag = 6880 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; 6881 const struct sched_class *prev_class; 6882 struct rq_flags rf; 6883 struct rq *rq; 6884 6885 /* XXX used to be waiter->prio, not waiter->task->prio */ 6886 prio = __rt_effective_prio(pi_task, p->normal_prio); 6887 6888 /* 6889 * If nothing changed; bail early. 6890 */ 6891 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio)) 6892 return; 6893 6894 rq = __task_rq_lock(p, &rf); 6895 update_rq_clock(rq); 6896 /* 6897 * Set under pi_lock && rq->lock, such that the value can be used under 6898 * either lock. 6899 * 6900 * Note that there is loads of tricky to make this pointer cache work 6901 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to 6902 * ensure a task is de-boosted (pi_task is set to NULL) before the 6903 * task is allowed to run again (and can exit). This ensures the pointer 6904 * points to a blocked task -- which guarantees the task is present. 6905 */ 6906 p->pi_top_task = pi_task; 6907 6908 /* 6909 * For FIFO/RR we only need to set prio, if that matches we're done. 6910 */ 6911 if (prio == p->prio && !dl_prio(prio)) 6912 goto out_unlock; 6913 6914 /* 6915 * Idle task boosting is a nono in general. There is one 6916 * exception, when PREEMPT_RT and NOHZ is active: 6917 * 6918 * The idle task calls get_next_timer_interrupt() and holds 6919 * the timer wheel base->lock on the CPU and another CPU wants 6920 * to access the timer (probably to cancel it). We can safely 6921 * ignore the boosting request, as the idle CPU runs this code 6922 * with interrupts disabled and will complete the lock 6923 * protected section without being interrupted. So there is no 6924 * real need to boost. 6925 */ 6926 if (unlikely(p == rq->idle)) { 6927 WARN_ON(p != rq->curr); 6928 WARN_ON(p->pi_blocked_on); 6929 goto out_unlock; 6930 } 6931 6932 trace_sched_pi_setprio(p, pi_task); 6933 oldprio = p->prio; 6934 6935 if (oldprio == prio) 6936 queue_flag &= ~DEQUEUE_MOVE; 6937 6938 prev_class = p->sched_class; 6939 queued = task_on_rq_queued(p); 6940 running = task_current(rq, p); 6941 if (queued) 6942 dequeue_task(rq, p, queue_flag); 6943 if (running) 6944 put_prev_task(rq, p); 6945 6946 /* 6947 * Boosting condition are: 6948 * 1. -rt task is running and holds mutex A 6949 * --> -dl task blocks on mutex A 6950 * 6951 * 2. -dl task is running and holds mutex A 6952 * --> -dl task blocks on mutex A and could preempt the 6953 * running task 6954 */ 6955 if (dl_prio(prio)) { 6956 if (!dl_prio(p->normal_prio) || 6957 (pi_task && dl_prio(pi_task->prio) && 6958 dl_entity_preempt(&pi_task->dl, &p->dl))) { 6959 p->dl.pi_se = pi_task->dl.pi_se; 6960 queue_flag |= ENQUEUE_REPLENISH; 6961 } else { 6962 p->dl.pi_se = &p->dl; 6963 } 6964 } else if (rt_prio(prio)) { 6965 if (dl_prio(oldprio)) 6966 p->dl.pi_se = &p->dl; 6967 if (oldprio < prio) 6968 queue_flag |= ENQUEUE_HEAD; 6969 } else { 6970 if (dl_prio(oldprio)) 6971 p->dl.pi_se = &p->dl; 6972 if (rt_prio(oldprio)) 6973 p->rt.timeout = 0; 6974 } 6975 6976 __setscheduler_prio(p, prio); 6977 6978 if (queued) 6979 enqueue_task(rq, p, queue_flag); 6980 if (running) 6981 set_next_task(rq, p); 6982 6983 check_class_changed(rq, p, prev_class, oldprio); 6984 out_unlock: 6985 /* Avoid rq from going away on us: */ 6986 preempt_disable(); 6987 6988 rq_unpin_lock(rq, &rf); 6989 __balance_callbacks(rq); 6990 raw_spin_rq_unlock(rq); 6991 6992 preempt_enable(); 6993 } 6994 #else 6995 static inline int rt_effective_prio(struct task_struct *p, int prio) 6996 { 6997 return prio; 6998 } 6999 #endif 7000 7001 void set_user_nice(struct task_struct *p, long nice) 7002 { 7003 bool queued, running; 7004 int old_prio; 7005 struct rq_flags rf; 7006 struct rq *rq; 7007 7008 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 7009 return; 7010 /* 7011 * We have to be careful, if called from sys_setpriority(), 7012 * the task might be in the middle of scheduling on another CPU. 7013 */ 7014 rq = task_rq_lock(p, &rf); 7015 update_rq_clock(rq); 7016 7017 /* 7018 * The RT priorities are set via sched_setscheduler(), but we still 7019 * allow the 'normal' nice value to be set - but as expected 7020 * it won't have any effect on scheduling until the task is 7021 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 7022 */ 7023 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 7024 p->static_prio = NICE_TO_PRIO(nice); 7025 goto out_unlock; 7026 } 7027 queued = task_on_rq_queued(p); 7028 running = task_current(rq, p); 7029 if (queued) 7030 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK); 7031 if (running) 7032 put_prev_task(rq, p); 7033 7034 p->static_prio = NICE_TO_PRIO(nice); 7035 set_load_weight(p, true); 7036 old_prio = p->prio; 7037 p->prio = effective_prio(p); 7038 7039 if (queued) 7040 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); 7041 if (running) 7042 set_next_task(rq, p); 7043 7044 /* 7045 * If the task increased its priority or is running and 7046 * lowered its priority, then reschedule its CPU: 7047 */ 7048 p->sched_class->prio_changed(rq, p, old_prio); 7049 7050 out_unlock: 7051 task_rq_unlock(rq, p, &rf); 7052 } 7053 EXPORT_SYMBOL(set_user_nice); 7054 7055 /* 7056 * is_nice_reduction - check if nice value is an actual reduction 7057 * 7058 * Similar to can_nice() but does not perform a capability check. 7059 * 7060 * @p: task 7061 * @nice: nice value 7062 */ 7063 static bool is_nice_reduction(const struct task_struct *p, const int nice) 7064 { 7065 /* Convert nice value [19,-20] to rlimit style value [1,40]: */ 7066 int nice_rlim = nice_to_rlimit(nice); 7067 7068 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE)); 7069 } 7070 7071 /* 7072 * can_nice - check if a task can reduce its nice value 7073 * @p: task 7074 * @nice: nice value 7075 */ 7076 int can_nice(const struct task_struct *p, const int nice) 7077 { 7078 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE); 7079 } 7080 7081 #ifdef __ARCH_WANT_SYS_NICE 7082 7083 /* 7084 * sys_nice - change the priority of the current process. 7085 * @increment: priority increment 7086 * 7087 * sys_setpriority is a more generic, but much slower function that 7088 * does similar things. 7089 */ 7090 SYSCALL_DEFINE1(nice, int, increment) 7091 { 7092 long nice, retval; 7093 7094 /* 7095 * Setpriority might change our priority at the same moment. 7096 * We don't have to worry. Conceptually one call occurs first 7097 * and we have a single winner. 7098 */ 7099 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 7100 nice = task_nice(current) + increment; 7101 7102 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 7103 if (increment < 0 && !can_nice(current, nice)) 7104 return -EPERM; 7105 7106 retval = security_task_setnice(current, nice); 7107 if (retval) 7108 return retval; 7109 7110 set_user_nice(current, nice); 7111 return 0; 7112 } 7113 7114 #endif 7115 7116 /** 7117 * task_prio - return the priority value of a given task. 7118 * @p: the task in question. 7119 * 7120 * Return: The priority value as seen by users in /proc. 7121 * 7122 * sched policy return value kernel prio user prio/nice 7123 * 7124 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19] 7125 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99] 7126 * deadline -101 -1 0 7127 */ 7128 int task_prio(const struct task_struct *p) 7129 { 7130 return p->prio - MAX_RT_PRIO; 7131 } 7132 7133 /** 7134 * idle_cpu - is a given CPU idle currently? 7135 * @cpu: the processor in question. 7136 * 7137 * Return: 1 if the CPU is currently idle. 0 otherwise. 7138 */ 7139 int idle_cpu(int cpu) 7140 { 7141 struct rq *rq = cpu_rq(cpu); 7142 7143 if (rq->curr != rq->idle) 7144 return 0; 7145 7146 if (rq->nr_running) 7147 return 0; 7148 7149 #ifdef CONFIG_SMP 7150 if (rq->ttwu_pending) 7151 return 0; 7152 #endif 7153 7154 return 1; 7155 } 7156 7157 /** 7158 * available_idle_cpu - is a given CPU idle for enqueuing work. 7159 * @cpu: the CPU in question. 7160 * 7161 * Return: 1 if the CPU is currently idle. 0 otherwise. 7162 */ 7163 int available_idle_cpu(int cpu) 7164 { 7165 if (!idle_cpu(cpu)) 7166 return 0; 7167 7168 if (vcpu_is_preempted(cpu)) 7169 return 0; 7170 7171 return 1; 7172 } 7173 7174 /** 7175 * idle_task - return the idle task for a given CPU. 7176 * @cpu: the processor in question. 7177 * 7178 * Return: The idle task for the CPU @cpu. 7179 */ 7180 struct task_struct *idle_task(int cpu) 7181 { 7182 return cpu_rq(cpu)->idle; 7183 } 7184 7185 #ifdef CONFIG_SMP 7186 /* 7187 * This function computes an effective utilization for the given CPU, to be 7188 * used for frequency selection given the linear relation: f = u * f_max. 7189 * 7190 * The scheduler tracks the following metrics: 7191 * 7192 * cpu_util_{cfs,rt,dl,irq}() 7193 * cpu_bw_dl() 7194 * 7195 * Where the cfs,rt and dl util numbers are tracked with the same metric and 7196 * synchronized windows and are thus directly comparable. 7197 * 7198 * The cfs,rt,dl utilization are the running times measured with rq->clock_task 7199 * which excludes things like IRQ and steal-time. These latter are then accrued 7200 * in the irq utilization. 7201 * 7202 * The DL bandwidth number otoh is not a measured metric but a value computed 7203 * based on the task model parameters and gives the minimal utilization 7204 * required to meet deadlines. 7205 */ 7206 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs, 7207 enum cpu_util_type type, 7208 struct task_struct *p) 7209 { 7210 unsigned long dl_util, util, irq, max; 7211 struct rq *rq = cpu_rq(cpu); 7212 7213 max = arch_scale_cpu_capacity(cpu); 7214 7215 if (!uclamp_is_used() && 7216 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) { 7217 return max; 7218 } 7219 7220 /* 7221 * Early check to see if IRQ/steal time saturates the CPU, can be 7222 * because of inaccuracies in how we track these -- see 7223 * update_irq_load_avg(). 7224 */ 7225 irq = cpu_util_irq(rq); 7226 if (unlikely(irq >= max)) 7227 return max; 7228 7229 /* 7230 * Because the time spend on RT/DL tasks is visible as 'lost' time to 7231 * CFS tasks and we use the same metric to track the effective 7232 * utilization (PELT windows are synchronized) we can directly add them 7233 * to obtain the CPU's actual utilization. 7234 * 7235 * CFS and RT utilization can be boosted or capped, depending on 7236 * utilization clamp constraints requested by currently RUNNABLE 7237 * tasks. 7238 * When there are no CFS RUNNABLE tasks, clamps are released and 7239 * frequency will be gracefully reduced with the utilization decay. 7240 */ 7241 util = util_cfs + cpu_util_rt(rq); 7242 if (type == FREQUENCY_UTIL) 7243 util = uclamp_rq_util_with(rq, util, p); 7244 7245 dl_util = cpu_util_dl(rq); 7246 7247 /* 7248 * For frequency selection we do not make cpu_util_dl() a permanent part 7249 * of this sum because we want to use cpu_bw_dl() later on, but we need 7250 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such 7251 * that we select f_max when there is no idle time. 7252 * 7253 * NOTE: numerical errors or stop class might cause us to not quite hit 7254 * saturation when we should -- something for later. 7255 */ 7256 if (util + dl_util >= max) 7257 return max; 7258 7259 /* 7260 * OTOH, for energy computation we need the estimated running time, so 7261 * include util_dl and ignore dl_bw. 7262 */ 7263 if (type == ENERGY_UTIL) 7264 util += dl_util; 7265 7266 /* 7267 * There is still idle time; further improve the number by using the 7268 * irq metric. Because IRQ/steal time is hidden from the task clock we 7269 * need to scale the task numbers: 7270 * 7271 * max - irq 7272 * U' = irq + --------- * U 7273 * max 7274 */ 7275 util = scale_irq_capacity(util, irq, max); 7276 util += irq; 7277 7278 /* 7279 * Bandwidth required by DEADLINE must always be granted while, for 7280 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism 7281 * to gracefully reduce the frequency when no tasks show up for longer 7282 * periods of time. 7283 * 7284 * Ideally we would like to set bw_dl as min/guaranteed freq and util + 7285 * bw_dl as requested freq. However, cpufreq is not yet ready for such 7286 * an interface. So, we only do the latter for now. 7287 */ 7288 if (type == FREQUENCY_UTIL) 7289 util += cpu_bw_dl(rq); 7290 7291 return min(max, util); 7292 } 7293 7294 unsigned long sched_cpu_util(int cpu) 7295 { 7296 return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL); 7297 } 7298 #endif /* CONFIG_SMP */ 7299 7300 /** 7301 * find_process_by_pid - find a process with a matching PID value. 7302 * @pid: the pid in question. 7303 * 7304 * The task of @pid, if found. %NULL otherwise. 7305 */ 7306 static struct task_struct *find_process_by_pid(pid_t pid) 7307 { 7308 return pid ? find_task_by_vpid(pid) : current; 7309 } 7310 7311 /* 7312 * sched_setparam() passes in -1 for its policy, to let the functions 7313 * it calls know not to change it. 7314 */ 7315 #define SETPARAM_POLICY -1 7316 7317 static void __setscheduler_params(struct task_struct *p, 7318 const struct sched_attr *attr) 7319 { 7320 int policy = attr->sched_policy; 7321 7322 if (policy == SETPARAM_POLICY) 7323 policy = p->policy; 7324 7325 p->policy = policy; 7326 7327 if (dl_policy(policy)) 7328 __setparam_dl(p, attr); 7329 else if (fair_policy(policy)) 7330 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 7331 7332 /* 7333 * __sched_setscheduler() ensures attr->sched_priority == 0 when 7334 * !rt_policy. Always setting this ensures that things like 7335 * getparam()/getattr() don't report silly values for !rt tasks. 7336 */ 7337 p->rt_priority = attr->sched_priority; 7338 p->normal_prio = normal_prio(p); 7339 set_load_weight(p, true); 7340 } 7341 7342 /* 7343 * Check the target process has a UID that matches the current process's: 7344 */ 7345 static bool check_same_owner(struct task_struct *p) 7346 { 7347 const struct cred *cred = current_cred(), *pcred; 7348 bool match; 7349 7350 rcu_read_lock(); 7351 pcred = __task_cred(p); 7352 match = (uid_eq(cred->euid, pcred->euid) || 7353 uid_eq(cred->euid, pcred->uid)); 7354 rcu_read_unlock(); 7355 return match; 7356 } 7357 7358 /* 7359 * Allow unprivileged RT tasks to decrease priority. 7360 * Only issue a capable test if needed and only once to avoid an audit 7361 * event on permitted non-privileged operations: 7362 */ 7363 static int user_check_sched_setscheduler(struct task_struct *p, 7364 const struct sched_attr *attr, 7365 int policy, int reset_on_fork) 7366 { 7367 if (fair_policy(policy)) { 7368 if (attr->sched_nice < task_nice(p) && 7369 !is_nice_reduction(p, attr->sched_nice)) 7370 goto req_priv; 7371 } 7372 7373 if (rt_policy(policy)) { 7374 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO); 7375 7376 /* Can't set/change the rt policy: */ 7377 if (policy != p->policy && !rlim_rtprio) 7378 goto req_priv; 7379 7380 /* Can't increase priority: */ 7381 if (attr->sched_priority > p->rt_priority && 7382 attr->sched_priority > rlim_rtprio) 7383 goto req_priv; 7384 } 7385 7386 /* 7387 * Can't set/change SCHED_DEADLINE policy at all for now 7388 * (safest behavior); in the future we would like to allow 7389 * unprivileged DL tasks to increase their relative deadline 7390 * or reduce their runtime (both ways reducing utilization) 7391 */ 7392 if (dl_policy(policy)) 7393 goto req_priv; 7394 7395 /* 7396 * Treat SCHED_IDLE as nice 20. Only allow a switch to 7397 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 7398 */ 7399 if (task_has_idle_policy(p) && !idle_policy(policy)) { 7400 if (!is_nice_reduction(p, task_nice(p))) 7401 goto req_priv; 7402 } 7403 7404 /* Can't change other user's priorities: */ 7405 if (!check_same_owner(p)) 7406 goto req_priv; 7407 7408 /* Normal users shall not reset the sched_reset_on_fork flag: */ 7409 if (p->sched_reset_on_fork && !reset_on_fork) 7410 goto req_priv; 7411 7412 return 0; 7413 7414 req_priv: 7415 if (!capable(CAP_SYS_NICE)) 7416 return -EPERM; 7417 7418 return 0; 7419 } 7420 7421 static int __sched_setscheduler(struct task_struct *p, 7422 const struct sched_attr *attr, 7423 bool user, bool pi) 7424 { 7425 int oldpolicy = -1, policy = attr->sched_policy; 7426 int retval, oldprio, newprio, queued, running; 7427 const struct sched_class *prev_class; 7428 struct callback_head *head; 7429 struct rq_flags rf; 7430 int reset_on_fork; 7431 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; 7432 struct rq *rq; 7433 7434 /* The pi code expects interrupts enabled */ 7435 BUG_ON(pi && in_interrupt()); 7436 recheck: 7437 /* Double check policy once rq lock held: */ 7438 if (policy < 0) { 7439 reset_on_fork = p->sched_reset_on_fork; 7440 policy = oldpolicy = p->policy; 7441 } else { 7442 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 7443 7444 if (!valid_policy(policy)) 7445 return -EINVAL; 7446 } 7447 7448 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV)) 7449 return -EINVAL; 7450 7451 /* 7452 * Valid priorities for SCHED_FIFO and SCHED_RR are 7453 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL, 7454 * SCHED_BATCH and SCHED_IDLE is 0. 7455 */ 7456 if (attr->sched_priority > MAX_RT_PRIO-1) 7457 return -EINVAL; 7458 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 7459 (rt_policy(policy) != (attr->sched_priority != 0))) 7460 return -EINVAL; 7461 7462 if (user) { 7463 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork); 7464 if (retval) 7465 return retval; 7466 7467 if (attr->sched_flags & SCHED_FLAG_SUGOV) 7468 return -EINVAL; 7469 7470 retval = security_task_setscheduler(p); 7471 if (retval) 7472 return retval; 7473 } 7474 7475 /* Update task specific "requested" clamps */ 7476 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) { 7477 retval = uclamp_validate(p, attr); 7478 if (retval) 7479 return retval; 7480 } 7481 7482 if (pi) 7483 cpuset_read_lock(); 7484 7485 /* 7486 * Make sure no PI-waiters arrive (or leave) while we are 7487 * changing the priority of the task: 7488 * 7489 * To be able to change p->policy safely, the appropriate 7490 * runqueue lock must be held. 7491 */ 7492 rq = task_rq_lock(p, &rf); 7493 update_rq_clock(rq); 7494 7495 /* 7496 * Changing the policy of the stop threads its a very bad idea: 7497 */ 7498 if (p == rq->stop) { 7499 retval = -EINVAL; 7500 goto unlock; 7501 } 7502 7503 /* 7504 * If not changing anything there's no need to proceed further, 7505 * but store a possible modification of reset_on_fork. 7506 */ 7507 if (unlikely(policy == p->policy)) { 7508 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 7509 goto change; 7510 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 7511 goto change; 7512 if (dl_policy(policy) && dl_param_changed(p, attr)) 7513 goto change; 7514 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) 7515 goto change; 7516 7517 p->sched_reset_on_fork = reset_on_fork; 7518 retval = 0; 7519 goto unlock; 7520 } 7521 change: 7522 7523 if (user) { 7524 #ifdef CONFIG_RT_GROUP_SCHED 7525 /* 7526 * Do not allow realtime tasks into groups that have no runtime 7527 * assigned. 7528 */ 7529 if (rt_bandwidth_enabled() && rt_policy(policy) && 7530 task_group(p)->rt_bandwidth.rt_runtime == 0 && 7531 !task_group_is_autogroup(task_group(p))) { 7532 retval = -EPERM; 7533 goto unlock; 7534 } 7535 #endif 7536 #ifdef CONFIG_SMP 7537 if (dl_bandwidth_enabled() && dl_policy(policy) && 7538 !(attr->sched_flags & SCHED_FLAG_SUGOV)) { 7539 cpumask_t *span = rq->rd->span; 7540 7541 /* 7542 * Don't allow tasks with an affinity mask smaller than 7543 * the entire root_domain to become SCHED_DEADLINE. We 7544 * will also fail if there's no bandwidth available. 7545 */ 7546 if (!cpumask_subset(span, p->cpus_ptr) || 7547 rq->rd->dl_bw.bw == 0) { 7548 retval = -EPERM; 7549 goto unlock; 7550 } 7551 } 7552 #endif 7553 } 7554 7555 /* Re-check policy now with rq lock held: */ 7556 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 7557 policy = oldpolicy = -1; 7558 task_rq_unlock(rq, p, &rf); 7559 if (pi) 7560 cpuset_read_unlock(); 7561 goto recheck; 7562 } 7563 7564 /* 7565 * If setscheduling to SCHED_DEADLINE (or changing the parameters 7566 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 7567 * is available. 7568 */ 7569 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) { 7570 retval = -EBUSY; 7571 goto unlock; 7572 } 7573 7574 p->sched_reset_on_fork = reset_on_fork; 7575 oldprio = p->prio; 7576 7577 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice); 7578 if (pi) { 7579 /* 7580 * Take priority boosted tasks into account. If the new 7581 * effective priority is unchanged, we just store the new 7582 * normal parameters and do not touch the scheduler class and 7583 * the runqueue. This will be done when the task deboost 7584 * itself. 7585 */ 7586 newprio = rt_effective_prio(p, newprio); 7587 if (newprio == oldprio) 7588 queue_flags &= ~DEQUEUE_MOVE; 7589 } 7590 7591 queued = task_on_rq_queued(p); 7592 running = task_current(rq, p); 7593 if (queued) 7594 dequeue_task(rq, p, queue_flags); 7595 if (running) 7596 put_prev_task(rq, p); 7597 7598 prev_class = p->sched_class; 7599 7600 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) { 7601 __setscheduler_params(p, attr); 7602 __setscheduler_prio(p, newprio); 7603 } 7604 __setscheduler_uclamp(p, attr); 7605 7606 if (queued) { 7607 /* 7608 * We enqueue to tail when the priority of a task is 7609 * increased (user space view). 7610 */ 7611 if (oldprio < p->prio) 7612 queue_flags |= ENQUEUE_HEAD; 7613 7614 enqueue_task(rq, p, queue_flags); 7615 } 7616 if (running) 7617 set_next_task(rq, p); 7618 7619 check_class_changed(rq, p, prev_class, oldprio); 7620 7621 /* Avoid rq from going away on us: */ 7622 preempt_disable(); 7623 head = splice_balance_callbacks(rq); 7624 task_rq_unlock(rq, p, &rf); 7625 7626 if (pi) { 7627 cpuset_read_unlock(); 7628 rt_mutex_adjust_pi(p); 7629 } 7630 7631 /* Run balance callbacks after we've adjusted the PI chain: */ 7632 balance_callbacks(rq, head); 7633 preempt_enable(); 7634 7635 return 0; 7636 7637 unlock: 7638 task_rq_unlock(rq, p, &rf); 7639 if (pi) 7640 cpuset_read_unlock(); 7641 return retval; 7642 } 7643 7644 static int _sched_setscheduler(struct task_struct *p, int policy, 7645 const struct sched_param *param, bool check) 7646 { 7647 struct sched_attr attr = { 7648 .sched_policy = policy, 7649 .sched_priority = param->sched_priority, 7650 .sched_nice = PRIO_TO_NICE(p->static_prio), 7651 }; 7652 7653 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ 7654 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { 7655 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 7656 policy &= ~SCHED_RESET_ON_FORK; 7657 attr.sched_policy = policy; 7658 } 7659 7660 return __sched_setscheduler(p, &attr, check, true); 7661 } 7662 /** 7663 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 7664 * @p: the task in question. 7665 * @policy: new policy. 7666 * @param: structure containing the new RT priority. 7667 * 7668 * Use sched_set_fifo(), read its comment. 7669 * 7670 * Return: 0 on success. An error code otherwise. 7671 * 7672 * NOTE that the task may be already dead. 7673 */ 7674 int sched_setscheduler(struct task_struct *p, int policy, 7675 const struct sched_param *param) 7676 { 7677 return _sched_setscheduler(p, policy, param, true); 7678 } 7679 7680 int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 7681 { 7682 return __sched_setscheduler(p, attr, true, true); 7683 } 7684 7685 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr) 7686 { 7687 return __sched_setscheduler(p, attr, false, true); 7688 } 7689 EXPORT_SYMBOL_GPL(sched_setattr_nocheck); 7690 7691 /** 7692 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 7693 * @p: the task in question. 7694 * @policy: new policy. 7695 * @param: structure containing the new RT priority. 7696 * 7697 * Just like sched_setscheduler, only don't bother checking if the 7698 * current context has permission. For example, this is needed in 7699 * stop_machine(): we create temporary high priority worker threads, 7700 * but our caller might not have that capability. 7701 * 7702 * Return: 0 on success. An error code otherwise. 7703 */ 7704 int sched_setscheduler_nocheck(struct task_struct *p, int policy, 7705 const struct sched_param *param) 7706 { 7707 return _sched_setscheduler(p, policy, param, false); 7708 } 7709 7710 /* 7711 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally 7712 * incapable of resource management, which is the one thing an OS really should 7713 * be doing. 7714 * 7715 * This is of course the reason it is limited to privileged users only. 7716 * 7717 * Worse still; it is fundamentally impossible to compose static priority 7718 * workloads. You cannot take two correctly working static prio workloads 7719 * and smash them together and still expect them to work. 7720 * 7721 * For this reason 'all' FIFO tasks the kernel creates are basically at: 7722 * 7723 * MAX_RT_PRIO / 2 7724 * 7725 * The administrator _MUST_ configure the system, the kernel simply doesn't 7726 * know enough information to make a sensible choice. 7727 */ 7728 void sched_set_fifo(struct task_struct *p) 7729 { 7730 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 }; 7731 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0); 7732 } 7733 EXPORT_SYMBOL_GPL(sched_set_fifo); 7734 7735 /* 7736 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL. 7737 */ 7738 void sched_set_fifo_low(struct task_struct *p) 7739 { 7740 struct sched_param sp = { .sched_priority = 1 }; 7741 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0); 7742 } 7743 EXPORT_SYMBOL_GPL(sched_set_fifo_low); 7744 7745 void sched_set_normal(struct task_struct *p, int nice) 7746 { 7747 struct sched_attr attr = { 7748 .sched_policy = SCHED_NORMAL, 7749 .sched_nice = nice, 7750 }; 7751 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0); 7752 } 7753 EXPORT_SYMBOL_GPL(sched_set_normal); 7754 7755 static int 7756 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 7757 { 7758 struct sched_param lparam; 7759 struct task_struct *p; 7760 int retval; 7761 7762 if (!param || pid < 0) 7763 return -EINVAL; 7764 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 7765 return -EFAULT; 7766 7767 rcu_read_lock(); 7768 retval = -ESRCH; 7769 p = find_process_by_pid(pid); 7770 if (likely(p)) 7771 get_task_struct(p); 7772 rcu_read_unlock(); 7773 7774 if (likely(p)) { 7775 retval = sched_setscheduler(p, policy, &lparam); 7776 put_task_struct(p); 7777 } 7778 7779 return retval; 7780 } 7781 7782 /* 7783 * Mimics kernel/events/core.c perf_copy_attr(). 7784 */ 7785 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr) 7786 { 7787 u32 size; 7788 int ret; 7789 7790 /* Zero the full structure, so that a short copy will be nice: */ 7791 memset(attr, 0, sizeof(*attr)); 7792 7793 ret = get_user(size, &uattr->size); 7794 if (ret) 7795 return ret; 7796 7797 /* ABI compatibility quirk: */ 7798 if (!size) 7799 size = SCHED_ATTR_SIZE_VER0; 7800 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE) 7801 goto err_size; 7802 7803 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size); 7804 if (ret) { 7805 if (ret == -E2BIG) 7806 goto err_size; 7807 return ret; 7808 } 7809 7810 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) && 7811 size < SCHED_ATTR_SIZE_VER1) 7812 return -EINVAL; 7813 7814 /* 7815 * XXX: Do we want to be lenient like existing syscalls; or do we want 7816 * to be strict and return an error on out-of-bounds values? 7817 */ 7818 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 7819 7820 return 0; 7821 7822 err_size: 7823 put_user(sizeof(*attr), &uattr->size); 7824 return -E2BIG; 7825 } 7826 7827 static void get_params(struct task_struct *p, struct sched_attr *attr) 7828 { 7829 if (task_has_dl_policy(p)) 7830 __getparam_dl(p, attr); 7831 else if (task_has_rt_policy(p)) 7832 attr->sched_priority = p->rt_priority; 7833 else 7834 attr->sched_nice = task_nice(p); 7835 } 7836 7837 /** 7838 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 7839 * @pid: the pid in question. 7840 * @policy: new policy. 7841 * @param: structure containing the new RT priority. 7842 * 7843 * Return: 0 on success. An error code otherwise. 7844 */ 7845 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param) 7846 { 7847 if (policy < 0) 7848 return -EINVAL; 7849 7850 return do_sched_setscheduler(pid, policy, param); 7851 } 7852 7853 /** 7854 * sys_sched_setparam - set/change the RT priority of a thread 7855 * @pid: the pid in question. 7856 * @param: structure containing the new RT priority. 7857 * 7858 * Return: 0 on success. An error code otherwise. 7859 */ 7860 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 7861 { 7862 return do_sched_setscheduler(pid, SETPARAM_POLICY, param); 7863 } 7864 7865 /** 7866 * sys_sched_setattr - same as above, but with extended sched_attr 7867 * @pid: the pid in question. 7868 * @uattr: structure containing the extended parameters. 7869 * @flags: for future extension. 7870 */ 7871 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 7872 unsigned int, flags) 7873 { 7874 struct sched_attr attr; 7875 struct task_struct *p; 7876 int retval; 7877 7878 if (!uattr || pid < 0 || flags) 7879 return -EINVAL; 7880 7881 retval = sched_copy_attr(uattr, &attr); 7882 if (retval) 7883 return retval; 7884 7885 if ((int)attr.sched_policy < 0) 7886 return -EINVAL; 7887 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY) 7888 attr.sched_policy = SETPARAM_POLICY; 7889 7890 rcu_read_lock(); 7891 retval = -ESRCH; 7892 p = find_process_by_pid(pid); 7893 if (likely(p)) 7894 get_task_struct(p); 7895 rcu_read_unlock(); 7896 7897 if (likely(p)) { 7898 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS) 7899 get_params(p, &attr); 7900 retval = sched_setattr(p, &attr); 7901 put_task_struct(p); 7902 } 7903 7904 return retval; 7905 } 7906 7907 /** 7908 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 7909 * @pid: the pid in question. 7910 * 7911 * Return: On success, the policy of the thread. Otherwise, a negative error 7912 * code. 7913 */ 7914 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 7915 { 7916 struct task_struct *p; 7917 int retval; 7918 7919 if (pid < 0) 7920 return -EINVAL; 7921 7922 retval = -ESRCH; 7923 rcu_read_lock(); 7924 p = find_process_by_pid(pid); 7925 if (p) { 7926 retval = security_task_getscheduler(p); 7927 if (!retval) 7928 retval = p->policy 7929 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 7930 } 7931 rcu_read_unlock(); 7932 return retval; 7933 } 7934 7935 /** 7936 * sys_sched_getparam - get the RT priority of a thread 7937 * @pid: the pid in question. 7938 * @param: structure containing the RT priority. 7939 * 7940 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 7941 * code. 7942 */ 7943 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 7944 { 7945 struct sched_param lp = { .sched_priority = 0 }; 7946 struct task_struct *p; 7947 int retval; 7948 7949 if (!param || pid < 0) 7950 return -EINVAL; 7951 7952 rcu_read_lock(); 7953 p = find_process_by_pid(pid); 7954 retval = -ESRCH; 7955 if (!p) 7956 goto out_unlock; 7957 7958 retval = security_task_getscheduler(p); 7959 if (retval) 7960 goto out_unlock; 7961 7962 if (task_has_rt_policy(p)) 7963 lp.sched_priority = p->rt_priority; 7964 rcu_read_unlock(); 7965 7966 /* 7967 * This one might sleep, we cannot do it with a spinlock held ... 7968 */ 7969 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 7970 7971 return retval; 7972 7973 out_unlock: 7974 rcu_read_unlock(); 7975 return retval; 7976 } 7977 7978 /* 7979 * Copy the kernel size attribute structure (which might be larger 7980 * than what user-space knows about) to user-space. 7981 * 7982 * Note that all cases are valid: user-space buffer can be larger or 7983 * smaller than the kernel-space buffer. The usual case is that both 7984 * have the same size. 7985 */ 7986 static int 7987 sched_attr_copy_to_user(struct sched_attr __user *uattr, 7988 struct sched_attr *kattr, 7989 unsigned int usize) 7990 { 7991 unsigned int ksize = sizeof(*kattr); 7992 7993 if (!access_ok(uattr, usize)) 7994 return -EFAULT; 7995 7996 /* 7997 * sched_getattr() ABI forwards and backwards compatibility: 7998 * 7999 * If usize == ksize then we just copy everything to user-space and all is good. 8000 * 8001 * If usize < ksize then we only copy as much as user-space has space for, 8002 * this keeps ABI compatibility as well. We skip the rest. 8003 * 8004 * If usize > ksize then user-space is using a newer version of the ABI, 8005 * which part the kernel doesn't know about. Just ignore it - tooling can 8006 * detect the kernel's knowledge of attributes from the attr->size value 8007 * which is set to ksize in this case. 8008 */ 8009 kattr->size = min(usize, ksize); 8010 8011 if (copy_to_user(uattr, kattr, kattr->size)) 8012 return -EFAULT; 8013 8014 return 0; 8015 } 8016 8017 /** 8018 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 8019 * @pid: the pid in question. 8020 * @uattr: structure containing the extended parameters. 8021 * @usize: sizeof(attr) for fwd/bwd comp. 8022 * @flags: for future extension. 8023 */ 8024 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 8025 unsigned int, usize, unsigned int, flags) 8026 { 8027 struct sched_attr kattr = { }; 8028 struct task_struct *p; 8029 int retval; 8030 8031 if (!uattr || pid < 0 || usize > PAGE_SIZE || 8032 usize < SCHED_ATTR_SIZE_VER0 || flags) 8033 return -EINVAL; 8034 8035 rcu_read_lock(); 8036 p = find_process_by_pid(pid); 8037 retval = -ESRCH; 8038 if (!p) 8039 goto out_unlock; 8040 8041 retval = security_task_getscheduler(p); 8042 if (retval) 8043 goto out_unlock; 8044 8045 kattr.sched_policy = p->policy; 8046 if (p->sched_reset_on_fork) 8047 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 8048 get_params(p, &kattr); 8049 kattr.sched_flags &= SCHED_FLAG_ALL; 8050 8051 #ifdef CONFIG_UCLAMP_TASK 8052 /* 8053 * This could race with another potential updater, but this is fine 8054 * because it'll correctly read the old or the new value. We don't need 8055 * to guarantee who wins the race as long as it doesn't return garbage. 8056 */ 8057 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value; 8058 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value; 8059 #endif 8060 8061 rcu_read_unlock(); 8062 8063 return sched_attr_copy_to_user(uattr, &kattr, usize); 8064 8065 out_unlock: 8066 rcu_read_unlock(); 8067 return retval; 8068 } 8069 8070 #ifdef CONFIG_SMP 8071 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask) 8072 { 8073 int ret = 0; 8074 8075 /* 8076 * If the task isn't a deadline task or admission control is 8077 * disabled then we don't care about affinity changes. 8078 */ 8079 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled()) 8080 return 0; 8081 8082 /* 8083 * Since bandwidth control happens on root_domain basis, 8084 * if admission test is enabled, we only admit -deadline 8085 * tasks allowed to run on all the CPUs in the task's 8086 * root_domain. 8087 */ 8088 rcu_read_lock(); 8089 if (!cpumask_subset(task_rq(p)->rd->span, mask)) 8090 ret = -EBUSY; 8091 rcu_read_unlock(); 8092 return ret; 8093 } 8094 #endif 8095 8096 static int 8097 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask) 8098 { 8099 int retval; 8100 cpumask_var_t cpus_allowed, new_mask; 8101 8102 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) 8103 return -ENOMEM; 8104 8105 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 8106 retval = -ENOMEM; 8107 goto out_free_cpus_allowed; 8108 } 8109 8110 cpuset_cpus_allowed(p, cpus_allowed); 8111 cpumask_and(new_mask, mask, cpus_allowed); 8112 8113 retval = dl_task_check_affinity(p, new_mask); 8114 if (retval) 8115 goto out_free_new_mask; 8116 again: 8117 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER); 8118 if (retval) 8119 goto out_free_new_mask; 8120 8121 cpuset_cpus_allowed(p, cpus_allowed); 8122 if (!cpumask_subset(new_mask, cpus_allowed)) { 8123 /* 8124 * We must have raced with a concurrent cpuset update. 8125 * Just reset the cpumask to the cpuset's cpus_allowed. 8126 */ 8127 cpumask_copy(new_mask, cpus_allowed); 8128 goto again; 8129 } 8130 8131 out_free_new_mask: 8132 free_cpumask_var(new_mask); 8133 out_free_cpus_allowed: 8134 free_cpumask_var(cpus_allowed); 8135 return retval; 8136 } 8137 8138 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 8139 { 8140 struct task_struct *p; 8141 int retval; 8142 8143 rcu_read_lock(); 8144 8145 p = find_process_by_pid(pid); 8146 if (!p) { 8147 rcu_read_unlock(); 8148 return -ESRCH; 8149 } 8150 8151 /* Prevent p going away */ 8152 get_task_struct(p); 8153 rcu_read_unlock(); 8154 8155 if (p->flags & PF_NO_SETAFFINITY) { 8156 retval = -EINVAL; 8157 goto out_put_task; 8158 } 8159 8160 if (!check_same_owner(p)) { 8161 rcu_read_lock(); 8162 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 8163 rcu_read_unlock(); 8164 retval = -EPERM; 8165 goto out_put_task; 8166 } 8167 rcu_read_unlock(); 8168 } 8169 8170 retval = security_task_setscheduler(p); 8171 if (retval) 8172 goto out_put_task; 8173 8174 retval = __sched_setaffinity(p, in_mask); 8175 out_put_task: 8176 put_task_struct(p); 8177 return retval; 8178 } 8179 8180 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 8181 struct cpumask *new_mask) 8182 { 8183 if (len < cpumask_size()) 8184 cpumask_clear(new_mask); 8185 else if (len > cpumask_size()) 8186 len = cpumask_size(); 8187 8188 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 8189 } 8190 8191 /** 8192 * sys_sched_setaffinity - set the CPU affinity of a process 8193 * @pid: pid of the process 8194 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 8195 * @user_mask_ptr: user-space pointer to the new CPU mask 8196 * 8197 * Return: 0 on success. An error code otherwise. 8198 */ 8199 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 8200 unsigned long __user *, user_mask_ptr) 8201 { 8202 cpumask_var_t new_mask; 8203 int retval; 8204 8205 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 8206 return -ENOMEM; 8207 8208 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 8209 if (retval == 0) 8210 retval = sched_setaffinity(pid, new_mask); 8211 free_cpumask_var(new_mask); 8212 return retval; 8213 } 8214 8215 long sched_getaffinity(pid_t pid, struct cpumask *mask) 8216 { 8217 struct task_struct *p; 8218 unsigned long flags; 8219 int retval; 8220 8221 rcu_read_lock(); 8222 8223 retval = -ESRCH; 8224 p = find_process_by_pid(pid); 8225 if (!p) 8226 goto out_unlock; 8227 8228 retval = security_task_getscheduler(p); 8229 if (retval) 8230 goto out_unlock; 8231 8232 raw_spin_lock_irqsave(&p->pi_lock, flags); 8233 cpumask_and(mask, &p->cpus_mask, cpu_active_mask); 8234 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 8235 8236 out_unlock: 8237 rcu_read_unlock(); 8238 8239 return retval; 8240 } 8241 8242 /** 8243 * sys_sched_getaffinity - get the CPU affinity of a process 8244 * @pid: pid of the process 8245 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 8246 * @user_mask_ptr: user-space pointer to hold the current CPU mask 8247 * 8248 * Return: size of CPU mask copied to user_mask_ptr on success. An 8249 * error code otherwise. 8250 */ 8251 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 8252 unsigned long __user *, user_mask_ptr) 8253 { 8254 int ret; 8255 cpumask_var_t mask; 8256 8257 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 8258 return -EINVAL; 8259 if (len & (sizeof(unsigned long)-1)) 8260 return -EINVAL; 8261 8262 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 8263 return -ENOMEM; 8264 8265 ret = sched_getaffinity(pid, mask); 8266 if (ret == 0) { 8267 unsigned int retlen = min(len, cpumask_size()); 8268 8269 if (copy_to_user(user_mask_ptr, mask, retlen)) 8270 ret = -EFAULT; 8271 else 8272 ret = retlen; 8273 } 8274 free_cpumask_var(mask); 8275 8276 return ret; 8277 } 8278 8279 static void do_sched_yield(void) 8280 { 8281 struct rq_flags rf; 8282 struct rq *rq; 8283 8284 rq = this_rq_lock_irq(&rf); 8285 8286 schedstat_inc(rq->yld_count); 8287 current->sched_class->yield_task(rq); 8288 8289 preempt_disable(); 8290 rq_unlock_irq(rq, &rf); 8291 sched_preempt_enable_no_resched(); 8292 8293 schedule(); 8294 } 8295 8296 /** 8297 * sys_sched_yield - yield the current processor to other threads. 8298 * 8299 * This function yields the current CPU to other tasks. If there are no 8300 * other threads running on this CPU then this function will return. 8301 * 8302 * Return: 0. 8303 */ 8304 SYSCALL_DEFINE0(sched_yield) 8305 { 8306 do_sched_yield(); 8307 return 0; 8308 } 8309 8310 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC) 8311 int __sched __cond_resched(void) 8312 { 8313 if (should_resched(0)) { 8314 preempt_schedule_common(); 8315 return 1; 8316 } 8317 /* 8318 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick 8319 * whether the current CPU is in an RCU read-side critical section, 8320 * so the tick can report quiescent states even for CPUs looping 8321 * in kernel context. In contrast, in non-preemptible kernels, 8322 * RCU readers leave no in-memory hints, which means that CPU-bound 8323 * processes executing in kernel context might never report an 8324 * RCU quiescent state. Therefore, the following code causes 8325 * cond_resched() to report a quiescent state, but only when RCU 8326 * is in urgent need of one. 8327 */ 8328 #ifndef CONFIG_PREEMPT_RCU 8329 rcu_all_qs(); 8330 #endif 8331 return 0; 8332 } 8333 EXPORT_SYMBOL(__cond_resched); 8334 #endif 8335 8336 #ifdef CONFIG_PREEMPT_DYNAMIC 8337 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL) 8338 #define cond_resched_dynamic_enabled __cond_resched 8339 #define cond_resched_dynamic_disabled ((void *)&__static_call_return0) 8340 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched); 8341 EXPORT_STATIC_CALL_TRAMP(cond_resched); 8342 8343 #define might_resched_dynamic_enabled __cond_resched 8344 #define might_resched_dynamic_disabled ((void *)&__static_call_return0) 8345 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched); 8346 EXPORT_STATIC_CALL_TRAMP(might_resched); 8347 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY) 8348 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched); 8349 int __sched dynamic_cond_resched(void) 8350 { 8351 if (!static_branch_unlikely(&sk_dynamic_cond_resched)) 8352 return 0; 8353 return __cond_resched(); 8354 } 8355 EXPORT_SYMBOL(dynamic_cond_resched); 8356 8357 static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched); 8358 int __sched dynamic_might_resched(void) 8359 { 8360 if (!static_branch_unlikely(&sk_dynamic_might_resched)) 8361 return 0; 8362 return __cond_resched(); 8363 } 8364 EXPORT_SYMBOL(dynamic_might_resched); 8365 #endif 8366 #endif 8367 8368 /* 8369 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 8370 * call schedule, and on return reacquire the lock. 8371 * 8372 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level 8373 * operations here to prevent schedule() from being called twice (once via 8374 * spin_unlock(), once by hand). 8375 */ 8376 int __cond_resched_lock(spinlock_t *lock) 8377 { 8378 int resched = should_resched(PREEMPT_LOCK_OFFSET); 8379 int ret = 0; 8380 8381 lockdep_assert_held(lock); 8382 8383 if (spin_needbreak(lock) || resched) { 8384 spin_unlock(lock); 8385 if (!_cond_resched()) 8386 cpu_relax(); 8387 ret = 1; 8388 spin_lock(lock); 8389 } 8390 return ret; 8391 } 8392 EXPORT_SYMBOL(__cond_resched_lock); 8393 8394 int __cond_resched_rwlock_read(rwlock_t *lock) 8395 { 8396 int resched = should_resched(PREEMPT_LOCK_OFFSET); 8397 int ret = 0; 8398 8399 lockdep_assert_held_read(lock); 8400 8401 if (rwlock_needbreak(lock) || resched) { 8402 read_unlock(lock); 8403 if (!_cond_resched()) 8404 cpu_relax(); 8405 ret = 1; 8406 read_lock(lock); 8407 } 8408 return ret; 8409 } 8410 EXPORT_SYMBOL(__cond_resched_rwlock_read); 8411 8412 int __cond_resched_rwlock_write(rwlock_t *lock) 8413 { 8414 int resched = should_resched(PREEMPT_LOCK_OFFSET); 8415 int ret = 0; 8416 8417 lockdep_assert_held_write(lock); 8418 8419 if (rwlock_needbreak(lock) || resched) { 8420 write_unlock(lock); 8421 if (!_cond_resched()) 8422 cpu_relax(); 8423 ret = 1; 8424 write_lock(lock); 8425 } 8426 return ret; 8427 } 8428 EXPORT_SYMBOL(__cond_resched_rwlock_write); 8429 8430 #ifdef CONFIG_PREEMPT_DYNAMIC 8431 8432 #ifdef CONFIG_GENERIC_ENTRY 8433 #include <linux/entry-common.h> 8434 #endif 8435 8436 /* 8437 * SC:cond_resched 8438 * SC:might_resched 8439 * SC:preempt_schedule 8440 * SC:preempt_schedule_notrace 8441 * SC:irqentry_exit_cond_resched 8442 * 8443 * 8444 * NONE: 8445 * cond_resched <- __cond_resched 8446 * might_resched <- RET0 8447 * preempt_schedule <- NOP 8448 * preempt_schedule_notrace <- NOP 8449 * irqentry_exit_cond_resched <- NOP 8450 * 8451 * VOLUNTARY: 8452 * cond_resched <- __cond_resched 8453 * might_resched <- __cond_resched 8454 * preempt_schedule <- NOP 8455 * preempt_schedule_notrace <- NOP 8456 * irqentry_exit_cond_resched <- NOP 8457 * 8458 * FULL: 8459 * cond_resched <- RET0 8460 * might_resched <- RET0 8461 * preempt_schedule <- preempt_schedule 8462 * preempt_schedule_notrace <- preempt_schedule_notrace 8463 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched 8464 */ 8465 8466 enum { 8467 preempt_dynamic_undefined = -1, 8468 preempt_dynamic_none, 8469 preempt_dynamic_voluntary, 8470 preempt_dynamic_full, 8471 }; 8472 8473 int preempt_dynamic_mode = preempt_dynamic_undefined; 8474 8475 int sched_dynamic_mode(const char *str) 8476 { 8477 if (!strcmp(str, "none")) 8478 return preempt_dynamic_none; 8479 8480 if (!strcmp(str, "voluntary")) 8481 return preempt_dynamic_voluntary; 8482 8483 if (!strcmp(str, "full")) 8484 return preempt_dynamic_full; 8485 8486 return -EINVAL; 8487 } 8488 8489 #if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL) 8490 #define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled) 8491 #define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled) 8492 #elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY) 8493 #define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key) 8494 #define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key) 8495 #else 8496 #error "Unsupported PREEMPT_DYNAMIC mechanism" 8497 #endif 8498 8499 void sched_dynamic_update(int mode) 8500 { 8501 /* 8502 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in 8503 * the ZERO state, which is invalid. 8504 */ 8505 preempt_dynamic_enable(cond_resched); 8506 preempt_dynamic_enable(might_resched); 8507 preempt_dynamic_enable(preempt_schedule); 8508 preempt_dynamic_enable(preempt_schedule_notrace); 8509 preempt_dynamic_enable(irqentry_exit_cond_resched); 8510 8511 switch (mode) { 8512 case preempt_dynamic_none: 8513 preempt_dynamic_enable(cond_resched); 8514 preempt_dynamic_disable(might_resched); 8515 preempt_dynamic_disable(preempt_schedule); 8516 preempt_dynamic_disable(preempt_schedule_notrace); 8517 preempt_dynamic_disable(irqentry_exit_cond_resched); 8518 pr_info("Dynamic Preempt: none\n"); 8519 break; 8520 8521 case preempt_dynamic_voluntary: 8522 preempt_dynamic_enable(cond_resched); 8523 preempt_dynamic_enable(might_resched); 8524 preempt_dynamic_disable(preempt_schedule); 8525 preempt_dynamic_disable(preempt_schedule_notrace); 8526 preempt_dynamic_disable(irqentry_exit_cond_resched); 8527 pr_info("Dynamic Preempt: voluntary\n"); 8528 break; 8529 8530 case preempt_dynamic_full: 8531 preempt_dynamic_disable(cond_resched); 8532 preempt_dynamic_disable(might_resched); 8533 preempt_dynamic_enable(preempt_schedule); 8534 preempt_dynamic_enable(preempt_schedule_notrace); 8535 preempt_dynamic_enable(irqentry_exit_cond_resched); 8536 pr_info("Dynamic Preempt: full\n"); 8537 break; 8538 } 8539 8540 preempt_dynamic_mode = mode; 8541 } 8542 8543 static int __init setup_preempt_mode(char *str) 8544 { 8545 int mode = sched_dynamic_mode(str); 8546 if (mode < 0) { 8547 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str); 8548 return 0; 8549 } 8550 8551 sched_dynamic_update(mode); 8552 return 1; 8553 } 8554 __setup("preempt=", setup_preempt_mode); 8555 8556 static void __init preempt_dynamic_init(void) 8557 { 8558 if (preempt_dynamic_mode == preempt_dynamic_undefined) { 8559 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) { 8560 sched_dynamic_update(preempt_dynamic_none); 8561 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) { 8562 sched_dynamic_update(preempt_dynamic_voluntary); 8563 } else { 8564 /* Default static call setting, nothing to do */ 8565 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT)); 8566 preempt_dynamic_mode = preempt_dynamic_full; 8567 pr_info("Dynamic Preempt: full\n"); 8568 } 8569 } 8570 } 8571 8572 #define PREEMPT_MODEL_ACCESSOR(mode) \ 8573 bool preempt_model_##mode(void) \ 8574 { \ 8575 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \ 8576 return preempt_dynamic_mode == preempt_dynamic_##mode; \ 8577 } \ 8578 EXPORT_SYMBOL_GPL(preempt_model_##mode) 8579 8580 PREEMPT_MODEL_ACCESSOR(none); 8581 PREEMPT_MODEL_ACCESSOR(voluntary); 8582 PREEMPT_MODEL_ACCESSOR(full); 8583 8584 #else /* !CONFIG_PREEMPT_DYNAMIC */ 8585 8586 static inline void preempt_dynamic_init(void) { } 8587 8588 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */ 8589 8590 /** 8591 * yield - yield the current processor to other threads. 8592 * 8593 * Do not ever use this function, there's a 99% chance you're doing it wrong. 8594 * 8595 * The scheduler is at all times free to pick the calling task as the most 8596 * eligible task to run, if removing the yield() call from your code breaks 8597 * it, it's already broken. 8598 * 8599 * Typical broken usage is: 8600 * 8601 * while (!event) 8602 * yield(); 8603 * 8604 * where one assumes that yield() will let 'the other' process run that will 8605 * make event true. If the current task is a SCHED_FIFO task that will never 8606 * happen. Never use yield() as a progress guarantee!! 8607 * 8608 * If you want to use yield() to wait for something, use wait_event(). 8609 * If you want to use yield() to be 'nice' for others, use cond_resched(). 8610 * If you still want to use yield(), do not! 8611 */ 8612 void __sched yield(void) 8613 { 8614 set_current_state(TASK_RUNNING); 8615 do_sched_yield(); 8616 } 8617 EXPORT_SYMBOL(yield); 8618 8619 /** 8620 * yield_to - yield the current processor to another thread in 8621 * your thread group, or accelerate that thread toward the 8622 * processor it's on. 8623 * @p: target task 8624 * @preempt: whether task preemption is allowed or not 8625 * 8626 * It's the caller's job to ensure that the target task struct 8627 * can't go away on us before we can do any checks. 8628 * 8629 * Return: 8630 * true (>0) if we indeed boosted the target task. 8631 * false (0) if we failed to boost the target. 8632 * -ESRCH if there's no task to yield to. 8633 */ 8634 int __sched yield_to(struct task_struct *p, bool preempt) 8635 { 8636 struct task_struct *curr = current; 8637 struct rq *rq, *p_rq; 8638 unsigned long flags; 8639 int yielded = 0; 8640 8641 local_irq_save(flags); 8642 rq = this_rq(); 8643 8644 again: 8645 p_rq = task_rq(p); 8646 /* 8647 * If we're the only runnable task on the rq and target rq also 8648 * has only one task, there's absolutely no point in yielding. 8649 */ 8650 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 8651 yielded = -ESRCH; 8652 goto out_irq; 8653 } 8654 8655 double_rq_lock(rq, p_rq); 8656 if (task_rq(p) != p_rq) { 8657 double_rq_unlock(rq, p_rq); 8658 goto again; 8659 } 8660 8661 if (!curr->sched_class->yield_to_task) 8662 goto out_unlock; 8663 8664 if (curr->sched_class != p->sched_class) 8665 goto out_unlock; 8666 8667 if (task_running(p_rq, p) || !task_is_running(p)) 8668 goto out_unlock; 8669 8670 yielded = curr->sched_class->yield_to_task(rq, p); 8671 if (yielded) { 8672 schedstat_inc(rq->yld_count); 8673 /* 8674 * Make p's CPU reschedule; pick_next_entity takes care of 8675 * fairness. 8676 */ 8677 if (preempt && rq != p_rq) 8678 resched_curr(p_rq); 8679 } 8680 8681 out_unlock: 8682 double_rq_unlock(rq, p_rq); 8683 out_irq: 8684 local_irq_restore(flags); 8685 8686 if (yielded > 0) 8687 schedule(); 8688 8689 return yielded; 8690 } 8691 EXPORT_SYMBOL_GPL(yield_to); 8692 8693 int io_schedule_prepare(void) 8694 { 8695 int old_iowait = current->in_iowait; 8696 8697 current->in_iowait = 1; 8698 blk_flush_plug(current->plug, true); 8699 return old_iowait; 8700 } 8701 8702 void io_schedule_finish(int token) 8703 { 8704 current->in_iowait = token; 8705 } 8706 8707 /* 8708 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 8709 * that process accounting knows that this is a task in IO wait state. 8710 */ 8711 long __sched io_schedule_timeout(long timeout) 8712 { 8713 int token; 8714 long ret; 8715 8716 token = io_schedule_prepare(); 8717 ret = schedule_timeout(timeout); 8718 io_schedule_finish(token); 8719 8720 return ret; 8721 } 8722 EXPORT_SYMBOL(io_schedule_timeout); 8723 8724 void __sched io_schedule(void) 8725 { 8726 int token; 8727 8728 token = io_schedule_prepare(); 8729 schedule(); 8730 io_schedule_finish(token); 8731 } 8732 EXPORT_SYMBOL(io_schedule); 8733 8734 /** 8735 * sys_sched_get_priority_max - return maximum RT priority. 8736 * @policy: scheduling class. 8737 * 8738 * Return: On success, this syscall returns the maximum 8739 * rt_priority that can be used by a given scheduling class. 8740 * On failure, a negative error code is returned. 8741 */ 8742 SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 8743 { 8744 int ret = -EINVAL; 8745 8746 switch (policy) { 8747 case SCHED_FIFO: 8748 case SCHED_RR: 8749 ret = MAX_RT_PRIO-1; 8750 break; 8751 case SCHED_DEADLINE: 8752 case SCHED_NORMAL: 8753 case SCHED_BATCH: 8754 case SCHED_IDLE: 8755 ret = 0; 8756 break; 8757 } 8758 return ret; 8759 } 8760 8761 /** 8762 * sys_sched_get_priority_min - return minimum RT priority. 8763 * @policy: scheduling class. 8764 * 8765 * Return: On success, this syscall returns the minimum 8766 * rt_priority that can be used by a given scheduling class. 8767 * On failure, a negative error code is returned. 8768 */ 8769 SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 8770 { 8771 int ret = -EINVAL; 8772 8773 switch (policy) { 8774 case SCHED_FIFO: 8775 case SCHED_RR: 8776 ret = 1; 8777 break; 8778 case SCHED_DEADLINE: 8779 case SCHED_NORMAL: 8780 case SCHED_BATCH: 8781 case SCHED_IDLE: 8782 ret = 0; 8783 } 8784 return ret; 8785 } 8786 8787 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t) 8788 { 8789 struct task_struct *p; 8790 unsigned int time_slice; 8791 struct rq_flags rf; 8792 struct rq *rq; 8793 int retval; 8794 8795 if (pid < 0) 8796 return -EINVAL; 8797 8798 retval = -ESRCH; 8799 rcu_read_lock(); 8800 p = find_process_by_pid(pid); 8801 if (!p) 8802 goto out_unlock; 8803 8804 retval = security_task_getscheduler(p); 8805 if (retval) 8806 goto out_unlock; 8807 8808 rq = task_rq_lock(p, &rf); 8809 time_slice = 0; 8810 if (p->sched_class->get_rr_interval) 8811 time_slice = p->sched_class->get_rr_interval(rq, p); 8812 task_rq_unlock(rq, p, &rf); 8813 8814 rcu_read_unlock(); 8815 jiffies_to_timespec64(time_slice, t); 8816 return 0; 8817 8818 out_unlock: 8819 rcu_read_unlock(); 8820 return retval; 8821 } 8822 8823 /** 8824 * sys_sched_rr_get_interval - return the default timeslice of a process. 8825 * @pid: pid of the process. 8826 * @interval: userspace pointer to the timeslice value. 8827 * 8828 * this syscall writes the default timeslice value of a given process 8829 * into the user-space timespec buffer. A value of '0' means infinity. 8830 * 8831 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 8832 * an error code. 8833 */ 8834 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 8835 struct __kernel_timespec __user *, interval) 8836 { 8837 struct timespec64 t; 8838 int retval = sched_rr_get_interval(pid, &t); 8839 8840 if (retval == 0) 8841 retval = put_timespec64(&t, interval); 8842 8843 return retval; 8844 } 8845 8846 #ifdef CONFIG_COMPAT_32BIT_TIME 8847 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid, 8848 struct old_timespec32 __user *, interval) 8849 { 8850 struct timespec64 t; 8851 int retval = sched_rr_get_interval(pid, &t); 8852 8853 if (retval == 0) 8854 retval = put_old_timespec32(&t, interval); 8855 return retval; 8856 } 8857 #endif 8858 8859 void sched_show_task(struct task_struct *p) 8860 { 8861 unsigned long free = 0; 8862 int ppid; 8863 8864 if (!try_get_task_stack(p)) 8865 return; 8866 8867 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p)); 8868 8869 if (task_is_running(p)) 8870 pr_cont(" running task "); 8871 #ifdef CONFIG_DEBUG_STACK_USAGE 8872 free = stack_not_used(p); 8873 #endif 8874 ppid = 0; 8875 rcu_read_lock(); 8876 if (pid_alive(p)) 8877 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 8878 rcu_read_unlock(); 8879 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n", 8880 free, task_pid_nr(p), ppid, 8881 read_task_thread_flags(p)); 8882 8883 print_worker_info(KERN_INFO, p); 8884 print_stop_info(KERN_INFO, p); 8885 show_stack(p, NULL, KERN_INFO); 8886 put_task_stack(p); 8887 } 8888 EXPORT_SYMBOL_GPL(sched_show_task); 8889 8890 static inline bool 8891 state_filter_match(unsigned long state_filter, struct task_struct *p) 8892 { 8893 unsigned int state = READ_ONCE(p->__state); 8894 8895 /* no filter, everything matches */ 8896 if (!state_filter) 8897 return true; 8898 8899 /* filter, but doesn't match */ 8900 if (!(state & state_filter)) 8901 return false; 8902 8903 /* 8904 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows 8905 * TASK_KILLABLE). 8906 */ 8907 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE) 8908 return false; 8909 8910 return true; 8911 } 8912 8913 8914 void show_state_filter(unsigned int state_filter) 8915 { 8916 struct task_struct *g, *p; 8917 8918 rcu_read_lock(); 8919 for_each_process_thread(g, p) { 8920 /* 8921 * reset the NMI-timeout, listing all files on a slow 8922 * console might take a lot of time: 8923 * Also, reset softlockup watchdogs on all CPUs, because 8924 * another CPU might be blocked waiting for us to process 8925 * an IPI. 8926 */ 8927 touch_nmi_watchdog(); 8928 touch_all_softlockup_watchdogs(); 8929 if (state_filter_match(state_filter, p)) 8930 sched_show_task(p); 8931 } 8932 8933 #ifdef CONFIG_SCHED_DEBUG 8934 if (!state_filter) 8935 sysrq_sched_debug_show(); 8936 #endif 8937 rcu_read_unlock(); 8938 /* 8939 * Only show locks if all tasks are dumped: 8940 */ 8941 if (!state_filter) 8942 debug_show_all_locks(); 8943 } 8944 8945 /** 8946 * init_idle - set up an idle thread for a given CPU 8947 * @idle: task in question 8948 * @cpu: CPU the idle task belongs to 8949 * 8950 * NOTE: this function does not set the idle thread's NEED_RESCHED 8951 * flag, to make booting more robust. 8952 */ 8953 void __init init_idle(struct task_struct *idle, int cpu) 8954 { 8955 struct rq *rq = cpu_rq(cpu); 8956 unsigned long flags; 8957 8958 __sched_fork(0, idle); 8959 8960 raw_spin_lock_irqsave(&idle->pi_lock, flags); 8961 raw_spin_rq_lock(rq); 8962 8963 idle->__state = TASK_RUNNING; 8964 idle->se.exec_start = sched_clock(); 8965 /* 8966 * PF_KTHREAD should already be set at this point; regardless, make it 8967 * look like a proper per-CPU kthread. 8968 */ 8969 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY; 8970 kthread_set_per_cpu(idle, cpu); 8971 8972 #ifdef CONFIG_SMP 8973 /* 8974 * It's possible that init_idle() gets called multiple times on a task, 8975 * in that case do_set_cpus_allowed() will not do the right thing. 8976 * 8977 * And since this is boot we can forgo the serialization. 8978 */ 8979 set_cpus_allowed_common(idle, cpumask_of(cpu), 0); 8980 #endif 8981 /* 8982 * We're having a chicken and egg problem, even though we are 8983 * holding rq->lock, the CPU isn't yet set to this CPU so the 8984 * lockdep check in task_group() will fail. 8985 * 8986 * Similar case to sched_fork(). / Alternatively we could 8987 * use task_rq_lock() here and obtain the other rq->lock. 8988 * 8989 * Silence PROVE_RCU 8990 */ 8991 rcu_read_lock(); 8992 __set_task_cpu(idle, cpu); 8993 rcu_read_unlock(); 8994 8995 rq->idle = idle; 8996 rcu_assign_pointer(rq->curr, idle); 8997 idle->on_rq = TASK_ON_RQ_QUEUED; 8998 #ifdef CONFIG_SMP 8999 idle->on_cpu = 1; 9000 #endif 9001 raw_spin_rq_unlock(rq); 9002 raw_spin_unlock_irqrestore(&idle->pi_lock, flags); 9003 9004 /* Set the preempt count _outside_ the spinlocks! */ 9005 init_idle_preempt_count(idle, cpu); 9006 9007 /* 9008 * The idle tasks have their own, simple scheduling class: 9009 */ 9010 idle->sched_class = &idle_sched_class; 9011 ftrace_graph_init_idle_task(idle, cpu); 9012 vtime_init_idle(idle, cpu); 9013 #ifdef CONFIG_SMP 9014 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 9015 #endif 9016 } 9017 9018 #ifdef CONFIG_SMP 9019 9020 int cpuset_cpumask_can_shrink(const struct cpumask *cur, 9021 const struct cpumask *trial) 9022 { 9023 int ret = 1; 9024 9025 if (cpumask_empty(cur)) 9026 return ret; 9027 9028 ret = dl_cpuset_cpumask_can_shrink(cur, trial); 9029 9030 return ret; 9031 } 9032 9033 int task_can_attach(struct task_struct *p, 9034 const struct cpumask *cs_effective_cpus) 9035 { 9036 int ret = 0; 9037 9038 /* 9039 * Kthreads which disallow setaffinity shouldn't be moved 9040 * to a new cpuset; we don't want to change their CPU 9041 * affinity and isolating such threads by their set of 9042 * allowed nodes is unnecessary. Thus, cpusets are not 9043 * applicable for such threads. This prevents checking for 9044 * success of set_cpus_allowed_ptr() on all attached tasks 9045 * before cpus_mask may be changed. 9046 */ 9047 if (p->flags & PF_NO_SETAFFINITY) { 9048 ret = -EINVAL; 9049 goto out; 9050 } 9051 9052 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span, 9053 cs_effective_cpus)) { 9054 int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus); 9055 9056 if (unlikely(cpu >= nr_cpu_ids)) 9057 return -EINVAL; 9058 ret = dl_cpu_busy(cpu, p); 9059 } 9060 9061 out: 9062 return ret; 9063 } 9064 9065 bool sched_smp_initialized __read_mostly; 9066 9067 #ifdef CONFIG_NUMA_BALANCING 9068 /* Migrate current task p to target_cpu */ 9069 int migrate_task_to(struct task_struct *p, int target_cpu) 9070 { 9071 struct migration_arg arg = { p, target_cpu }; 9072 int curr_cpu = task_cpu(p); 9073 9074 if (curr_cpu == target_cpu) 9075 return 0; 9076 9077 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr)) 9078 return -EINVAL; 9079 9080 /* TODO: This is not properly updating schedstats */ 9081 9082 trace_sched_move_numa(p, curr_cpu, target_cpu); 9083 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 9084 } 9085 9086 /* 9087 * Requeue a task on a given node and accurately track the number of NUMA 9088 * tasks on the runqueues 9089 */ 9090 void sched_setnuma(struct task_struct *p, int nid) 9091 { 9092 bool queued, running; 9093 struct rq_flags rf; 9094 struct rq *rq; 9095 9096 rq = task_rq_lock(p, &rf); 9097 queued = task_on_rq_queued(p); 9098 running = task_current(rq, p); 9099 9100 if (queued) 9101 dequeue_task(rq, p, DEQUEUE_SAVE); 9102 if (running) 9103 put_prev_task(rq, p); 9104 9105 p->numa_preferred_nid = nid; 9106 9107 if (queued) 9108 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK); 9109 if (running) 9110 set_next_task(rq, p); 9111 task_rq_unlock(rq, p, &rf); 9112 } 9113 #endif /* CONFIG_NUMA_BALANCING */ 9114 9115 #ifdef CONFIG_HOTPLUG_CPU 9116 /* 9117 * Ensure that the idle task is using init_mm right before its CPU goes 9118 * offline. 9119 */ 9120 void idle_task_exit(void) 9121 { 9122 struct mm_struct *mm = current->active_mm; 9123 9124 BUG_ON(cpu_online(smp_processor_id())); 9125 BUG_ON(current != this_rq()->idle); 9126 9127 if (mm != &init_mm) { 9128 switch_mm(mm, &init_mm, current); 9129 finish_arch_post_lock_switch(); 9130 } 9131 9132 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */ 9133 } 9134 9135 static int __balance_push_cpu_stop(void *arg) 9136 { 9137 struct task_struct *p = arg; 9138 struct rq *rq = this_rq(); 9139 struct rq_flags rf; 9140 int cpu; 9141 9142 raw_spin_lock_irq(&p->pi_lock); 9143 rq_lock(rq, &rf); 9144 9145 update_rq_clock(rq); 9146 9147 if (task_rq(p) == rq && task_on_rq_queued(p)) { 9148 cpu = select_fallback_rq(rq->cpu, p); 9149 rq = __migrate_task(rq, &rf, p, cpu); 9150 } 9151 9152 rq_unlock(rq, &rf); 9153 raw_spin_unlock_irq(&p->pi_lock); 9154 9155 put_task_struct(p); 9156 9157 return 0; 9158 } 9159 9160 static DEFINE_PER_CPU(struct cpu_stop_work, push_work); 9161 9162 /* 9163 * Ensure we only run per-cpu kthreads once the CPU goes !active. 9164 * 9165 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only 9166 * effective when the hotplug motion is down. 9167 */ 9168 static void balance_push(struct rq *rq) 9169 { 9170 struct task_struct *push_task = rq->curr; 9171 9172 lockdep_assert_rq_held(rq); 9173 9174 /* 9175 * Ensure the thing is persistent until balance_push_set(.on = false); 9176 */ 9177 rq->balance_callback = &balance_push_callback; 9178 9179 /* 9180 * Only active while going offline and when invoked on the outgoing 9181 * CPU. 9182 */ 9183 if (!cpu_dying(rq->cpu) || rq != this_rq()) 9184 return; 9185 9186 /* 9187 * Both the cpu-hotplug and stop task are in this case and are 9188 * required to complete the hotplug process. 9189 */ 9190 if (kthread_is_per_cpu(push_task) || 9191 is_migration_disabled(push_task)) { 9192 9193 /* 9194 * If this is the idle task on the outgoing CPU try to wake 9195 * up the hotplug control thread which might wait for the 9196 * last task to vanish. The rcuwait_active() check is 9197 * accurate here because the waiter is pinned on this CPU 9198 * and can't obviously be running in parallel. 9199 * 9200 * On RT kernels this also has to check whether there are 9201 * pinned and scheduled out tasks on the runqueue. They 9202 * need to leave the migrate disabled section first. 9203 */ 9204 if (!rq->nr_running && !rq_has_pinned_tasks(rq) && 9205 rcuwait_active(&rq->hotplug_wait)) { 9206 raw_spin_rq_unlock(rq); 9207 rcuwait_wake_up(&rq->hotplug_wait); 9208 raw_spin_rq_lock(rq); 9209 } 9210 return; 9211 } 9212 9213 get_task_struct(push_task); 9214 /* 9215 * Temporarily drop rq->lock such that we can wake-up the stop task. 9216 * Both preemption and IRQs are still disabled. 9217 */ 9218 raw_spin_rq_unlock(rq); 9219 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task, 9220 this_cpu_ptr(&push_work)); 9221 /* 9222 * At this point need_resched() is true and we'll take the loop in 9223 * schedule(). The next pick is obviously going to be the stop task 9224 * which kthread_is_per_cpu() and will push this task away. 9225 */ 9226 raw_spin_rq_lock(rq); 9227 } 9228 9229 static void balance_push_set(int cpu, bool on) 9230 { 9231 struct rq *rq = cpu_rq(cpu); 9232 struct rq_flags rf; 9233 9234 rq_lock_irqsave(rq, &rf); 9235 if (on) { 9236 WARN_ON_ONCE(rq->balance_callback); 9237 rq->balance_callback = &balance_push_callback; 9238 } else if (rq->balance_callback == &balance_push_callback) { 9239 rq->balance_callback = NULL; 9240 } 9241 rq_unlock_irqrestore(rq, &rf); 9242 } 9243 9244 /* 9245 * Invoked from a CPUs hotplug control thread after the CPU has been marked 9246 * inactive. All tasks which are not per CPU kernel threads are either 9247 * pushed off this CPU now via balance_push() or placed on a different CPU 9248 * during wakeup. Wait until the CPU is quiescent. 9249 */ 9250 static void balance_hotplug_wait(void) 9251 { 9252 struct rq *rq = this_rq(); 9253 9254 rcuwait_wait_event(&rq->hotplug_wait, 9255 rq->nr_running == 1 && !rq_has_pinned_tasks(rq), 9256 TASK_UNINTERRUPTIBLE); 9257 } 9258 9259 #else 9260 9261 static inline void balance_push(struct rq *rq) 9262 { 9263 } 9264 9265 static inline void balance_push_set(int cpu, bool on) 9266 { 9267 } 9268 9269 static inline void balance_hotplug_wait(void) 9270 { 9271 } 9272 9273 #endif /* CONFIG_HOTPLUG_CPU */ 9274 9275 void set_rq_online(struct rq *rq) 9276 { 9277 if (!rq->online) { 9278 const struct sched_class *class; 9279 9280 cpumask_set_cpu(rq->cpu, rq->rd->online); 9281 rq->online = 1; 9282 9283 for_each_class(class) { 9284 if (class->rq_online) 9285 class->rq_online(rq); 9286 } 9287 } 9288 } 9289 9290 void set_rq_offline(struct rq *rq) 9291 { 9292 if (rq->online) { 9293 const struct sched_class *class; 9294 9295 for_each_class(class) { 9296 if (class->rq_offline) 9297 class->rq_offline(rq); 9298 } 9299 9300 cpumask_clear_cpu(rq->cpu, rq->rd->online); 9301 rq->online = 0; 9302 } 9303 } 9304 9305 /* 9306 * used to mark begin/end of suspend/resume: 9307 */ 9308 static int num_cpus_frozen; 9309 9310 /* 9311 * Update cpusets according to cpu_active mask. If cpusets are 9312 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 9313 * around partition_sched_domains(). 9314 * 9315 * If we come here as part of a suspend/resume, don't touch cpusets because we 9316 * want to restore it back to its original state upon resume anyway. 9317 */ 9318 static void cpuset_cpu_active(void) 9319 { 9320 if (cpuhp_tasks_frozen) { 9321 /* 9322 * num_cpus_frozen tracks how many CPUs are involved in suspend 9323 * resume sequence. As long as this is not the last online 9324 * operation in the resume sequence, just build a single sched 9325 * domain, ignoring cpusets. 9326 */ 9327 partition_sched_domains(1, NULL, NULL); 9328 if (--num_cpus_frozen) 9329 return; 9330 /* 9331 * This is the last CPU online operation. So fall through and 9332 * restore the original sched domains by considering the 9333 * cpuset configurations. 9334 */ 9335 cpuset_force_rebuild(); 9336 } 9337 cpuset_update_active_cpus(); 9338 } 9339 9340 static int cpuset_cpu_inactive(unsigned int cpu) 9341 { 9342 if (!cpuhp_tasks_frozen) { 9343 int ret = dl_cpu_busy(cpu, NULL); 9344 9345 if (ret) 9346 return ret; 9347 cpuset_update_active_cpus(); 9348 } else { 9349 num_cpus_frozen++; 9350 partition_sched_domains(1, NULL, NULL); 9351 } 9352 return 0; 9353 } 9354 9355 int sched_cpu_activate(unsigned int cpu) 9356 { 9357 struct rq *rq = cpu_rq(cpu); 9358 struct rq_flags rf; 9359 9360 /* 9361 * Clear the balance_push callback and prepare to schedule 9362 * regular tasks. 9363 */ 9364 balance_push_set(cpu, false); 9365 9366 #ifdef CONFIG_SCHED_SMT 9367 /* 9368 * When going up, increment the number of cores with SMT present. 9369 */ 9370 if (cpumask_weight(cpu_smt_mask(cpu)) == 2) 9371 static_branch_inc_cpuslocked(&sched_smt_present); 9372 #endif 9373 set_cpu_active(cpu, true); 9374 9375 if (sched_smp_initialized) { 9376 sched_update_numa(cpu, true); 9377 sched_domains_numa_masks_set(cpu); 9378 cpuset_cpu_active(); 9379 } 9380 9381 /* 9382 * Put the rq online, if not already. This happens: 9383 * 9384 * 1) In the early boot process, because we build the real domains 9385 * after all CPUs have been brought up. 9386 * 9387 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the 9388 * domains. 9389 */ 9390 rq_lock_irqsave(rq, &rf); 9391 if (rq->rd) { 9392 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 9393 set_rq_online(rq); 9394 } 9395 rq_unlock_irqrestore(rq, &rf); 9396 9397 return 0; 9398 } 9399 9400 int sched_cpu_deactivate(unsigned int cpu) 9401 { 9402 struct rq *rq = cpu_rq(cpu); 9403 struct rq_flags rf; 9404 int ret; 9405 9406 /* 9407 * Remove CPU from nohz.idle_cpus_mask to prevent participating in 9408 * load balancing when not active 9409 */ 9410 nohz_balance_exit_idle(rq); 9411 9412 set_cpu_active(cpu, false); 9413 9414 /* 9415 * From this point forward, this CPU will refuse to run any task that 9416 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively 9417 * push those tasks away until this gets cleared, see 9418 * sched_cpu_dying(). 9419 */ 9420 balance_push_set(cpu, true); 9421 9422 /* 9423 * We've cleared cpu_active_mask / set balance_push, wait for all 9424 * preempt-disabled and RCU users of this state to go away such that 9425 * all new such users will observe it. 9426 * 9427 * Specifically, we rely on ttwu to no longer target this CPU, see 9428 * ttwu_queue_cond() and is_cpu_allowed(). 9429 * 9430 * Do sync before park smpboot threads to take care the rcu boost case. 9431 */ 9432 synchronize_rcu(); 9433 9434 rq_lock_irqsave(rq, &rf); 9435 if (rq->rd) { 9436 update_rq_clock(rq); 9437 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 9438 set_rq_offline(rq); 9439 } 9440 rq_unlock_irqrestore(rq, &rf); 9441 9442 #ifdef CONFIG_SCHED_SMT 9443 /* 9444 * When going down, decrement the number of cores with SMT present. 9445 */ 9446 if (cpumask_weight(cpu_smt_mask(cpu)) == 2) 9447 static_branch_dec_cpuslocked(&sched_smt_present); 9448 9449 sched_core_cpu_deactivate(cpu); 9450 #endif 9451 9452 if (!sched_smp_initialized) 9453 return 0; 9454 9455 sched_update_numa(cpu, false); 9456 ret = cpuset_cpu_inactive(cpu); 9457 if (ret) { 9458 balance_push_set(cpu, false); 9459 set_cpu_active(cpu, true); 9460 sched_update_numa(cpu, true); 9461 return ret; 9462 } 9463 sched_domains_numa_masks_clear(cpu); 9464 return 0; 9465 } 9466 9467 static void sched_rq_cpu_starting(unsigned int cpu) 9468 { 9469 struct rq *rq = cpu_rq(cpu); 9470 9471 rq->calc_load_update = calc_load_update; 9472 update_max_interval(); 9473 } 9474 9475 int sched_cpu_starting(unsigned int cpu) 9476 { 9477 sched_core_cpu_starting(cpu); 9478 sched_rq_cpu_starting(cpu); 9479 sched_tick_start(cpu); 9480 return 0; 9481 } 9482 9483 #ifdef CONFIG_HOTPLUG_CPU 9484 9485 /* 9486 * Invoked immediately before the stopper thread is invoked to bring the 9487 * CPU down completely. At this point all per CPU kthreads except the 9488 * hotplug thread (current) and the stopper thread (inactive) have been 9489 * either parked or have been unbound from the outgoing CPU. Ensure that 9490 * any of those which might be on the way out are gone. 9491 * 9492 * If after this point a bound task is being woken on this CPU then the 9493 * responsible hotplug callback has failed to do it's job. 9494 * sched_cpu_dying() will catch it with the appropriate fireworks. 9495 */ 9496 int sched_cpu_wait_empty(unsigned int cpu) 9497 { 9498 balance_hotplug_wait(); 9499 return 0; 9500 } 9501 9502 /* 9503 * Since this CPU is going 'away' for a while, fold any nr_active delta we 9504 * might have. Called from the CPU stopper task after ensuring that the 9505 * stopper is the last running task on the CPU, so nr_active count is 9506 * stable. We need to take the teardown thread which is calling this into 9507 * account, so we hand in adjust = 1 to the load calculation. 9508 * 9509 * Also see the comment "Global load-average calculations". 9510 */ 9511 static void calc_load_migrate(struct rq *rq) 9512 { 9513 long delta = calc_load_fold_active(rq, 1); 9514 9515 if (delta) 9516 atomic_long_add(delta, &calc_load_tasks); 9517 } 9518 9519 static void dump_rq_tasks(struct rq *rq, const char *loglvl) 9520 { 9521 struct task_struct *g, *p; 9522 int cpu = cpu_of(rq); 9523 9524 lockdep_assert_rq_held(rq); 9525 9526 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running); 9527 for_each_process_thread(g, p) { 9528 if (task_cpu(p) != cpu) 9529 continue; 9530 9531 if (!task_on_rq_queued(p)) 9532 continue; 9533 9534 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm); 9535 } 9536 } 9537 9538 int sched_cpu_dying(unsigned int cpu) 9539 { 9540 struct rq *rq = cpu_rq(cpu); 9541 struct rq_flags rf; 9542 9543 /* Handle pending wakeups and then migrate everything off */ 9544 sched_tick_stop(cpu); 9545 9546 rq_lock_irqsave(rq, &rf); 9547 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) { 9548 WARN(true, "Dying CPU not properly vacated!"); 9549 dump_rq_tasks(rq, KERN_WARNING); 9550 } 9551 rq_unlock_irqrestore(rq, &rf); 9552 9553 calc_load_migrate(rq); 9554 update_max_interval(); 9555 hrtick_clear(rq); 9556 sched_core_cpu_dying(cpu); 9557 return 0; 9558 } 9559 #endif 9560 9561 void __init sched_init_smp(void) 9562 { 9563 sched_init_numa(NUMA_NO_NODE); 9564 9565 /* 9566 * There's no userspace yet to cause hotplug operations; hence all the 9567 * CPU masks are stable and all blatant races in the below code cannot 9568 * happen. 9569 */ 9570 mutex_lock(&sched_domains_mutex); 9571 sched_init_domains(cpu_active_mask); 9572 mutex_unlock(&sched_domains_mutex); 9573 9574 /* Move init over to a non-isolated CPU */ 9575 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0) 9576 BUG(); 9577 current->flags &= ~PF_NO_SETAFFINITY; 9578 sched_init_granularity(); 9579 9580 init_sched_rt_class(); 9581 init_sched_dl_class(); 9582 9583 sched_smp_initialized = true; 9584 } 9585 9586 static int __init migration_init(void) 9587 { 9588 sched_cpu_starting(smp_processor_id()); 9589 return 0; 9590 } 9591 early_initcall(migration_init); 9592 9593 #else 9594 void __init sched_init_smp(void) 9595 { 9596 sched_init_granularity(); 9597 } 9598 #endif /* CONFIG_SMP */ 9599 9600 int in_sched_functions(unsigned long addr) 9601 { 9602 return in_lock_functions(addr) || 9603 (addr >= (unsigned long)__sched_text_start 9604 && addr < (unsigned long)__sched_text_end); 9605 } 9606 9607 #ifdef CONFIG_CGROUP_SCHED 9608 /* 9609 * Default task group. 9610 * Every task in system belongs to this group at bootup. 9611 */ 9612 struct task_group root_task_group; 9613 LIST_HEAD(task_groups); 9614 9615 /* Cacheline aligned slab cache for task_group */ 9616 static struct kmem_cache *task_group_cache __read_mostly; 9617 #endif 9618 9619 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 9620 DECLARE_PER_CPU(cpumask_var_t, select_rq_mask); 9621 9622 void __init sched_init(void) 9623 { 9624 unsigned long ptr = 0; 9625 int i; 9626 9627 /* Make sure the linker didn't screw up */ 9628 BUG_ON(&idle_sched_class != &fair_sched_class + 1 || 9629 &fair_sched_class != &rt_sched_class + 1 || 9630 &rt_sched_class != &dl_sched_class + 1); 9631 #ifdef CONFIG_SMP 9632 BUG_ON(&dl_sched_class != &stop_sched_class + 1); 9633 #endif 9634 9635 wait_bit_init(); 9636 9637 #ifdef CONFIG_FAIR_GROUP_SCHED 9638 ptr += 2 * nr_cpu_ids * sizeof(void **); 9639 #endif 9640 #ifdef CONFIG_RT_GROUP_SCHED 9641 ptr += 2 * nr_cpu_ids * sizeof(void **); 9642 #endif 9643 if (ptr) { 9644 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT); 9645 9646 #ifdef CONFIG_FAIR_GROUP_SCHED 9647 root_task_group.se = (struct sched_entity **)ptr; 9648 ptr += nr_cpu_ids * sizeof(void **); 9649 9650 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 9651 ptr += nr_cpu_ids * sizeof(void **); 9652 9653 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 9654 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 9655 #endif /* CONFIG_FAIR_GROUP_SCHED */ 9656 #ifdef CONFIG_RT_GROUP_SCHED 9657 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 9658 ptr += nr_cpu_ids * sizeof(void **); 9659 9660 root_task_group.rt_rq = (struct rt_rq **)ptr; 9661 ptr += nr_cpu_ids * sizeof(void **); 9662 9663 #endif /* CONFIG_RT_GROUP_SCHED */ 9664 } 9665 #ifdef CONFIG_CPUMASK_OFFSTACK 9666 for_each_possible_cpu(i) { 9667 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node( 9668 cpumask_size(), GFP_KERNEL, cpu_to_node(i)); 9669 per_cpu(select_rq_mask, i) = (cpumask_var_t)kzalloc_node( 9670 cpumask_size(), GFP_KERNEL, cpu_to_node(i)); 9671 } 9672 #endif /* CONFIG_CPUMASK_OFFSTACK */ 9673 9674 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime()); 9675 9676 #ifdef CONFIG_SMP 9677 init_defrootdomain(); 9678 #endif 9679 9680 #ifdef CONFIG_RT_GROUP_SCHED 9681 init_rt_bandwidth(&root_task_group.rt_bandwidth, 9682 global_rt_period(), global_rt_runtime()); 9683 #endif /* CONFIG_RT_GROUP_SCHED */ 9684 9685 #ifdef CONFIG_CGROUP_SCHED 9686 task_group_cache = KMEM_CACHE(task_group, 0); 9687 9688 list_add(&root_task_group.list, &task_groups); 9689 INIT_LIST_HEAD(&root_task_group.children); 9690 INIT_LIST_HEAD(&root_task_group.siblings); 9691 autogroup_init(&init_task); 9692 #endif /* CONFIG_CGROUP_SCHED */ 9693 9694 for_each_possible_cpu(i) { 9695 struct rq *rq; 9696 9697 rq = cpu_rq(i); 9698 raw_spin_lock_init(&rq->__lock); 9699 rq->nr_running = 0; 9700 rq->calc_load_active = 0; 9701 rq->calc_load_update = jiffies + LOAD_FREQ; 9702 init_cfs_rq(&rq->cfs); 9703 init_rt_rq(&rq->rt); 9704 init_dl_rq(&rq->dl); 9705 #ifdef CONFIG_FAIR_GROUP_SCHED 9706 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 9707 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; 9708 /* 9709 * How much CPU bandwidth does root_task_group get? 9710 * 9711 * In case of task-groups formed thr' the cgroup filesystem, it 9712 * gets 100% of the CPU resources in the system. This overall 9713 * system CPU resource is divided among the tasks of 9714 * root_task_group and its child task-groups in a fair manner, 9715 * based on each entity's (task or task-group's) weight 9716 * (se->load.weight). 9717 * 9718 * In other words, if root_task_group has 10 tasks of weight 9719 * 1024) and two child groups A0 and A1 (of weight 1024 each), 9720 * then A0's share of the CPU resource is: 9721 * 9722 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 9723 * 9724 * We achieve this by letting root_task_group's tasks sit 9725 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 9726 */ 9727 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 9728 #endif /* CONFIG_FAIR_GROUP_SCHED */ 9729 9730 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 9731 #ifdef CONFIG_RT_GROUP_SCHED 9732 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 9733 #endif 9734 #ifdef CONFIG_SMP 9735 rq->sd = NULL; 9736 rq->rd = NULL; 9737 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE; 9738 rq->balance_callback = &balance_push_callback; 9739 rq->active_balance = 0; 9740 rq->next_balance = jiffies; 9741 rq->push_cpu = 0; 9742 rq->cpu = i; 9743 rq->online = 0; 9744 rq->idle_stamp = 0; 9745 rq->avg_idle = 2*sysctl_sched_migration_cost; 9746 rq->wake_stamp = jiffies; 9747 rq->wake_avg_idle = rq->avg_idle; 9748 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 9749 9750 INIT_LIST_HEAD(&rq->cfs_tasks); 9751 9752 rq_attach_root(rq, &def_root_domain); 9753 #ifdef CONFIG_NO_HZ_COMMON 9754 rq->last_blocked_load_update_tick = jiffies; 9755 atomic_set(&rq->nohz_flags, 0); 9756 9757 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq); 9758 #endif 9759 #ifdef CONFIG_HOTPLUG_CPU 9760 rcuwait_init(&rq->hotplug_wait); 9761 #endif 9762 #endif /* CONFIG_SMP */ 9763 hrtick_rq_init(rq); 9764 atomic_set(&rq->nr_iowait, 0); 9765 9766 #ifdef CONFIG_SCHED_CORE 9767 rq->core = rq; 9768 rq->core_pick = NULL; 9769 rq->core_enabled = 0; 9770 rq->core_tree = RB_ROOT; 9771 rq->core_forceidle_count = 0; 9772 rq->core_forceidle_occupation = 0; 9773 rq->core_forceidle_start = 0; 9774 9775 rq->core_cookie = 0UL; 9776 #endif 9777 } 9778 9779 set_load_weight(&init_task, false); 9780 9781 /* 9782 * The boot idle thread does lazy MMU switching as well: 9783 */ 9784 mmgrab(&init_mm); 9785 enter_lazy_tlb(&init_mm, current); 9786 9787 /* 9788 * The idle task doesn't need the kthread struct to function, but it 9789 * is dressed up as a per-CPU kthread and thus needs to play the part 9790 * if we want to avoid special-casing it in code that deals with per-CPU 9791 * kthreads. 9792 */ 9793 WARN_ON(!set_kthread_struct(current)); 9794 9795 /* 9796 * Make us the idle thread. Technically, schedule() should not be 9797 * called from this thread, however somewhere below it might be, 9798 * but because we are the idle thread, we just pick up running again 9799 * when this runqueue becomes "idle". 9800 */ 9801 init_idle(current, smp_processor_id()); 9802 9803 calc_load_update = jiffies + LOAD_FREQ; 9804 9805 #ifdef CONFIG_SMP 9806 idle_thread_set_boot_cpu(); 9807 balance_push_set(smp_processor_id(), false); 9808 #endif 9809 init_sched_fair_class(); 9810 9811 psi_init(); 9812 9813 init_uclamp(); 9814 9815 preempt_dynamic_init(); 9816 9817 scheduler_running = 1; 9818 } 9819 9820 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP 9821 9822 void __might_sleep(const char *file, int line) 9823 { 9824 unsigned int state = get_current_state(); 9825 /* 9826 * Blocking primitives will set (and therefore destroy) current->state, 9827 * since we will exit with TASK_RUNNING make sure we enter with it, 9828 * otherwise we will destroy state. 9829 */ 9830 WARN_ONCE(state != TASK_RUNNING && current->task_state_change, 9831 "do not call blocking ops when !TASK_RUNNING; " 9832 "state=%x set at [<%p>] %pS\n", state, 9833 (void *)current->task_state_change, 9834 (void *)current->task_state_change); 9835 9836 __might_resched(file, line, 0); 9837 } 9838 EXPORT_SYMBOL(__might_sleep); 9839 9840 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip) 9841 { 9842 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT)) 9843 return; 9844 9845 if (preempt_count() == preempt_offset) 9846 return; 9847 9848 pr_err("Preemption disabled at:"); 9849 print_ip_sym(KERN_ERR, ip); 9850 } 9851 9852 static inline bool resched_offsets_ok(unsigned int offsets) 9853 { 9854 unsigned int nested = preempt_count(); 9855 9856 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT; 9857 9858 return nested == offsets; 9859 } 9860 9861 void __might_resched(const char *file, int line, unsigned int offsets) 9862 { 9863 /* Ratelimiting timestamp: */ 9864 static unsigned long prev_jiffy; 9865 9866 unsigned long preempt_disable_ip; 9867 9868 /* WARN_ON_ONCE() by default, no rate limit required: */ 9869 rcu_sleep_check(); 9870 9871 if ((resched_offsets_ok(offsets) && !irqs_disabled() && 9872 !is_idle_task(current) && !current->non_block_count) || 9873 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING || 9874 oops_in_progress) 9875 return; 9876 9877 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 9878 return; 9879 prev_jiffy = jiffies; 9880 9881 /* Save this before calling printk(), since that will clobber it: */ 9882 preempt_disable_ip = get_preempt_disable_ip(current); 9883 9884 pr_err("BUG: sleeping function called from invalid context at %s:%d\n", 9885 file, line); 9886 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n", 9887 in_atomic(), irqs_disabled(), current->non_block_count, 9888 current->pid, current->comm); 9889 pr_err("preempt_count: %x, expected: %x\n", preempt_count(), 9890 offsets & MIGHT_RESCHED_PREEMPT_MASK); 9891 9892 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) { 9893 pr_err("RCU nest depth: %d, expected: %u\n", 9894 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT); 9895 } 9896 9897 if (task_stack_end_corrupted(current)) 9898 pr_emerg("Thread overran stack, or stack corrupted\n"); 9899 9900 debug_show_held_locks(current); 9901 if (irqs_disabled()) 9902 print_irqtrace_events(current); 9903 9904 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK, 9905 preempt_disable_ip); 9906 9907 dump_stack(); 9908 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 9909 } 9910 EXPORT_SYMBOL(__might_resched); 9911 9912 void __cant_sleep(const char *file, int line, int preempt_offset) 9913 { 9914 static unsigned long prev_jiffy; 9915 9916 if (irqs_disabled()) 9917 return; 9918 9919 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT)) 9920 return; 9921 9922 if (preempt_count() > preempt_offset) 9923 return; 9924 9925 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 9926 return; 9927 prev_jiffy = jiffies; 9928 9929 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line); 9930 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 9931 in_atomic(), irqs_disabled(), 9932 current->pid, current->comm); 9933 9934 debug_show_held_locks(current); 9935 dump_stack(); 9936 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 9937 } 9938 EXPORT_SYMBOL_GPL(__cant_sleep); 9939 9940 #ifdef CONFIG_SMP 9941 void __cant_migrate(const char *file, int line) 9942 { 9943 static unsigned long prev_jiffy; 9944 9945 if (irqs_disabled()) 9946 return; 9947 9948 if (is_migration_disabled(current)) 9949 return; 9950 9951 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT)) 9952 return; 9953 9954 if (preempt_count() > 0) 9955 return; 9956 9957 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 9958 return; 9959 prev_jiffy = jiffies; 9960 9961 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line); 9962 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n", 9963 in_atomic(), irqs_disabled(), is_migration_disabled(current), 9964 current->pid, current->comm); 9965 9966 debug_show_held_locks(current); 9967 dump_stack(); 9968 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 9969 } 9970 EXPORT_SYMBOL_GPL(__cant_migrate); 9971 #endif 9972 #endif 9973 9974 #ifdef CONFIG_MAGIC_SYSRQ 9975 void normalize_rt_tasks(void) 9976 { 9977 struct task_struct *g, *p; 9978 struct sched_attr attr = { 9979 .sched_policy = SCHED_NORMAL, 9980 }; 9981 9982 read_lock(&tasklist_lock); 9983 for_each_process_thread(g, p) { 9984 /* 9985 * Only normalize user tasks: 9986 */ 9987 if (p->flags & PF_KTHREAD) 9988 continue; 9989 9990 p->se.exec_start = 0; 9991 schedstat_set(p->stats.wait_start, 0); 9992 schedstat_set(p->stats.sleep_start, 0); 9993 schedstat_set(p->stats.block_start, 0); 9994 9995 if (!dl_task(p) && !rt_task(p)) { 9996 /* 9997 * Renice negative nice level userspace 9998 * tasks back to 0: 9999 */ 10000 if (task_nice(p) < 0) 10001 set_user_nice(p, 0); 10002 continue; 10003 } 10004 10005 __sched_setscheduler(p, &attr, false, false); 10006 } 10007 read_unlock(&tasklist_lock); 10008 } 10009 10010 #endif /* CONFIG_MAGIC_SYSRQ */ 10011 10012 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 10013 /* 10014 * These functions are only useful for the IA64 MCA handling, or kdb. 10015 * 10016 * They can only be called when the whole system has been 10017 * stopped - every CPU needs to be quiescent, and no scheduling 10018 * activity can take place. Using them for anything else would 10019 * be a serious bug, and as a result, they aren't even visible 10020 * under any other configuration. 10021 */ 10022 10023 /** 10024 * curr_task - return the current task for a given CPU. 10025 * @cpu: the processor in question. 10026 * 10027 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 10028 * 10029 * Return: The current task for @cpu. 10030 */ 10031 struct task_struct *curr_task(int cpu) 10032 { 10033 return cpu_curr(cpu); 10034 } 10035 10036 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 10037 10038 #ifdef CONFIG_IA64 10039 /** 10040 * ia64_set_curr_task - set the current task for a given CPU. 10041 * @cpu: the processor in question. 10042 * @p: the task pointer to set. 10043 * 10044 * Description: This function must only be used when non-maskable interrupts 10045 * are serviced on a separate stack. It allows the architecture to switch the 10046 * notion of the current task on a CPU in a non-blocking manner. This function 10047 * must be called with all CPU's synchronized, and interrupts disabled, the 10048 * and caller must save the original value of the current task (see 10049 * curr_task() above) and restore that value before reenabling interrupts and 10050 * re-starting the system. 10051 * 10052 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 10053 */ 10054 void ia64_set_curr_task(int cpu, struct task_struct *p) 10055 { 10056 cpu_curr(cpu) = p; 10057 } 10058 10059 #endif 10060 10061 #ifdef CONFIG_CGROUP_SCHED 10062 /* task_group_lock serializes the addition/removal of task groups */ 10063 static DEFINE_SPINLOCK(task_group_lock); 10064 10065 static inline void alloc_uclamp_sched_group(struct task_group *tg, 10066 struct task_group *parent) 10067 { 10068 #ifdef CONFIG_UCLAMP_TASK_GROUP 10069 enum uclamp_id clamp_id; 10070 10071 for_each_clamp_id(clamp_id) { 10072 uclamp_se_set(&tg->uclamp_req[clamp_id], 10073 uclamp_none(clamp_id), false); 10074 tg->uclamp[clamp_id] = parent->uclamp[clamp_id]; 10075 } 10076 #endif 10077 } 10078 10079 static void sched_free_group(struct task_group *tg) 10080 { 10081 free_fair_sched_group(tg); 10082 free_rt_sched_group(tg); 10083 autogroup_free(tg); 10084 kmem_cache_free(task_group_cache, tg); 10085 } 10086 10087 static void sched_free_group_rcu(struct rcu_head *rcu) 10088 { 10089 sched_free_group(container_of(rcu, struct task_group, rcu)); 10090 } 10091 10092 static void sched_unregister_group(struct task_group *tg) 10093 { 10094 unregister_fair_sched_group(tg); 10095 unregister_rt_sched_group(tg); 10096 /* 10097 * We have to wait for yet another RCU grace period to expire, as 10098 * print_cfs_stats() might run concurrently. 10099 */ 10100 call_rcu(&tg->rcu, sched_free_group_rcu); 10101 } 10102 10103 /* allocate runqueue etc for a new task group */ 10104 struct task_group *sched_create_group(struct task_group *parent) 10105 { 10106 struct task_group *tg; 10107 10108 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO); 10109 if (!tg) 10110 return ERR_PTR(-ENOMEM); 10111 10112 if (!alloc_fair_sched_group(tg, parent)) 10113 goto err; 10114 10115 if (!alloc_rt_sched_group(tg, parent)) 10116 goto err; 10117 10118 alloc_uclamp_sched_group(tg, parent); 10119 10120 return tg; 10121 10122 err: 10123 sched_free_group(tg); 10124 return ERR_PTR(-ENOMEM); 10125 } 10126 10127 void sched_online_group(struct task_group *tg, struct task_group *parent) 10128 { 10129 unsigned long flags; 10130 10131 spin_lock_irqsave(&task_group_lock, flags); 10132 list_add_rcu(&tg->list, &task_groups); 10133 10134 /* Root should already exist: */ 10135 WARN_ON(!parent); 10136 10137 tg->parent = parent; 10138 INIT_LIST_HEAD(&tg->children); 10139 list_add_rcu(&tg->siblings, &parent->children); 10140 spin_unlock_irqrestore(&task_group_lock, flags); 10141 10142 online_fair_sched_group(tg); 10143 } 10144 10145 /* rcu callback to free various structures associated with a task group */ 10146 static void sched_unregister_group_rcu(struct rcu_head *rhp) 10147 { 10148 /* Now it should be safe to free those cfs_rqs: */ 10149 sched_unregister_group(container_of(rhp, struct task_group, rcu)); 10150 } 10151 10152 void sched_destroy_group(struct task_group *tg) 10153 { 10154 /* Wait for possible concurrent references to cfs_rqs complete: */ 10155 call_rcu(&tg->rcu, sched_unregister_group_rcu); 10156 } 10157 10158 void sched_release_group(struct task_group *tg) 10159 { 10160 unsigned long flags; 10161 10162 /* 10163 * Unlink first, to avoid walk_tg_tree_from() from finding us (via 10164 * sched_cfs_period_timer()). 10165 * 10166 * For this to be effective, we have to wait for all pending users of 10167 * this task group to leave their RCU critical section to ensure no new 10168 * user will see our dying task group any more. Specifically ensure 10169 * that tg_unthrottle_up() won't add decayed cfs_rq's to it. 10170 * 10171 * We therefore defer calling unregister_fair_sched_group() to 10172 * sched_unregister_group() which is guarantied to get called only after the 10173 * current RCU grace period has expired. 10174 */ 10175 spin_lock_irqsave(&task_group_lock, flags); 10176 list_del_rcu(&tg->list); 10177 list_del_rcu(&tg->siblings); 10178 spin_unlock_irqrestore(&task_group_lock, flags); 10179 } 10180 10181 static void sched_change_group(struct task_struct *tsk, int type) 10182 { 10183 struct task_group *tg; 10184 10185 /* 10186 * All callers are synchronized by task_rq_lock(); we do not use RCU 10187 * which is pointless here. Thus, we pass "true" to task_css_check() 10188 * to prevent lockdep warnings. 10189 */ 10190 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), 10191 struct task_group, css); 10192 tg = autogroup_task_group(tsk, tg); 10193 tsk->sched_task_group = tg; 10194 10195 #ifdef CONFIG_FAIR_GROUP_SCHED 10196 if (tsk->sched_class->task_change_group) 10197 tsk->sched_class->task_change_group(tsk, type); 10198 else 10199 #endif 10200 set_task_rq(tsk, task_cpu(tsk)); 10201 } 10202 10203 /* 10204 * Change task's runqueue when it moves between groups. 10205 * 10206 * The caller of this function should have put the task in its new group by 10207 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect 10208 * its new group. 10209 */ 10210 void sched_move_task(struct task_struct *tsk) 10211 { 10212 int queued, running, queue_flags = 10213 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK; 10214 struct rq_flags rf; 10215 struct rq *rq; 10216 10217 rq = task_rq_lock(tsk, &rf); 10218 update_rq_clock(rq); 10219 10220 running = task_current(rq, tsk); 10221 queued = task_on_rq_queued(tsk); 10222 10223 if (queued) 10224 dequeue_task(rq, tsk, queue_flags); 10225 if (running) 10226 put_prev_task(rq, tsk); 10227 10228 sched_change_group(tsk, TASK_MOVE_GROUP); 10229 10230 if (queued) 10231 enqueue_task(rq, tsk, queue_flags); 10232 if (running) { 10233 set_next_task(rq, tsk); 10234 /* 10235 * After changing group, the running task may have joined a 10236 * throttled one but it's still the running task. Trigger a 10237 * resched to make sure that task can still run. 10238 */ 10239 resched_curr(rq); 10240 } 10241 10242 task_rq_unlock(rq, tsk, &rf); 10243 } 10244 10245 static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 10246 { 10247 return css ? container_of(css, struct task_group, css) : NULL; 10248 } 10249 10250 static struct cgroup_subsys_state * 10251 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 10252 { 10253 struct task_group *parent = css_tg(parent_css); 10254 struct task_group *tg; 10255 10256 if (!parent) { 10257 /* This is early initialization for the top cgroup */ 10258 return &root_task_group.css; 10259 } 10260 10261 tg = sched_create_group(parent); 10262 if (IS_ERR(tg)) 10263 return ERR_PTR(-ENOMEM); 10264 10265 return &tg->css; 10266 } 10267 10268 /* Expose task group only after completing cgroup initialization */ 10269 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 10270 { 10271 struct task_group *tg = css_tg(css); 10272 struct task_group *parent = css_tg(css->parent); 10273 10274 if (parent) 10275 sched_online_group(tg, parent); 10276 10277 #ifdef CONFIG_UCLAMP_TASK_GROUP 10278 /* Propagate the effective uclamp value for the new group */ 10279 mutex_lock(&uclamp_mutex); 10280 rcu_read_lock(); 10281 cpu_util_update_eff(css); 10282 rcu_read_unlock(); 10283 mutex_unlock(&uclamp_mutex); 10284 #endif 10285 10286 return 0; 10287 } 10288 10289 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css) 10290 { 10291 struct task_group *tg = css_tg(css); 10292 10293 sched_release_group(tg); 10294 } 10295 10296 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 10297 { 10298 struct task_group *tg = css_tg(css); 10299 10300 /* 10301 * Relies on the RCU grace period between css_released() and this. 10302 */ 10303 sched_unregister_group(tg); 10304 } 10305 10306 /* 10307 * This is called before wake_up_new_task(), therefore we really only 10308 * have to set its group bits, all the other stuff does not apply. 10309 */ 10310 static void cpu_cgroup_fork(struct task_struct *task) 10311 { 10312 struct rq_flags rf; 10313 struct rq *rq; 10314 10315 rq = task_rq_lock(task, &rf); 10316 10317 update_rq_clock(rq); 10318 sched_change_group(task, TASK_SET_GROUP); 10319 10320 task_rq_unlock(rq, task, &rf); 10321 } 10322 10323 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset) 10324 { 10325 struct task_struct *task; 10326 struct cgroup_subsys_state *css; 10327 int ret = 0; 10328 10329 cgroup_taskset_for_each(task, css, tset) { 10330 #ifdef CONFIG_RT_GROUP_SCHED 10331 if (!sched_rt_can_attach(css_tg(css), task)) 10332 return -EINVAL; 10333 #endif 10334 /* 10335 * Serialize against wake_up_new_task() such that if it's 10336 * running, we're sure to observe its full state. 10337 */ 10338 raw_spin_lock_irq(&task->pi_lock); 10339 /* 10340 * Avoid calling sched_move_task() before wake_up_new_task() 10341 * has happened. This would lead to problems with PELT, due to 10342 * move wanting to detach+attach while we're not attached yet. 10343 */ 10344 if (READ_ONCE(task->__state) == TASK_NEW) 10345 ret = -EINVAL; 10346 raw_spin_unlock_irq(&task->pi_lock); 10347 10348 if (ret) 10349 break; 10350 } 10351 return ret; 10352 } 10353 10354 static void cpu_cgroup_attach(struct cgroup_taskset *tset) 10355 { 10356 struct task_struct *task; 10357 struct cgroup_subsys_state *css; 10358 10359 cgroup_taskset_for_each(task, css, tset) 10360 sched_move_task(task); 10361 } 10362 10363 #ifdef CONFIG_UCLAMP_TASK_GROUP 10364 static void cpu_util_update_eff(struct cgroup_subsys_state *css) 10365 { 10366 struct cgroup_subsys_state *top_css = css; 10367 struct uclamp_se *uc_parent = NULL; 10368 struct uclamp_se *uc_se = NULL; 10369 unsigned int eff[UCLAMP_CNT]; 10370 enum uclamp_id clamp_id; 10371 unsigned int clamps; 10372 10373 lockdep_assert_held(&uclamp_mutex); 10374 SCHED_WARN_ON(!rcu_read_lock_held()); 10375 10376 css_for_each_descendant_pre(css, top_css) { 10377 uc_parent = css_tg(css)->parent 10378 ? css_tg(css)->parent->uclamp : NULL; 10379 10380 for_each_clamp_id(clamp_id) { 10381 /* Assume effective clamps matches requested clamps */ 10382 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value; 10383 /* Cap effective clamps with parent's effective clamps */ 10384 if (uc_parent && 10385 eff[clamp_id] > uc_parent[clamp_id].value) { 10386 eff[clamp_id] = uc_parent[clamp_id].value; 10387 } 10388 } 10389 /* Ensure protection is always capped by limit */ 10390 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]); 10391 10392 /* Propagate most restrictive effective clamps */ 10393 clamps = 0x0; 10394 uc_se = css_tg(css)->uclamp; 10395 for_each_clamp_id(clamp_id) { 10396 if (eff[clamp_id] == uc_se[clamp_id].value) 10397 continue; 10398 uc_se[clamp_id].value = eff[clamp_id]; 10399 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]); 10400 clamps |= (0x1 << clamp_id); 10401 } 10402 if (!clamps) { 10403 css = css_rightmost_descendant(css); 10404 continue; 10405 } 10406 10407 /* Immediately update descendants RUNNABLE tasks */ 10408 uclamp_update_active_tasks(css); 10409 } 10410 } 10411 10412 /* 10413 * Integer 10^N with a given N exponent by casting to integer the literal "1eN" 10414 * C expression. Since there is no way to convert a macro argument (N) into a 10415 * character constant, use two levels of macros. 10416 */ 10417 #define _POW10(exp) ((unsigned int)1e##exp) 10418 #define POW10(exp) _POW10(exp) 10419 10420 struct uclamp_request { 10421 #define UCLAMP_PERCENT_SHIFT 2 10422 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT)) 10423 s64 percent; 10424 u64 util; 10425 int ret; 10426 }; 10427 10428 static inline struct uclamp_request 10429 capacity_from_percent(char *buf) 10430 { 10431 struct uclamp_request req = { 10432 .percent = UCLAMP_PERCENT_SCALE, 10433 .util = SCHED_CAPACITY_SCALE, 10434 .ret = 0, 10435 }; 10436 10437 buf = strim(buf); 10438 if (strcmp(buf, "max")) { 10439 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT, 10440 &req.percent); 10441 if (req.ret) 10442 return req; 10443 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) { 10444 req.ret = -ERANGE; 10445 return req; 10446 } 10447 10448 req.util = req.percent << SCHED_CAPACITY_SHIFT; 10449 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE); 10450 } 10451 10452 return req; 10453 } 10454 10455 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf, 10456 size_t nbytes, loff_t off, 10457 enum uclamp_id clamp_id) 10458 { 10459 struct uclamp_request req; 10460 struct task_group *tg; 10461 10462 req = capacity_from_percent(buf); 10463 if (req.ret) 10464 return req.ret; 10465 10466 static_branch_enable(&sched_uclamp_used); 10467 10468 mutex_lock(&uclamp_mutex); 10469 rcu_read_lock(); 10470 10471 tg = css_tg(of_css(of)); 10472 if (tg->uclamp_req[clamp_id].value != req.util) 10473 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false); 10474 10475 /* 10476 * Because of not recoverable conversion rounding we keep track of the 10477 * exact requested value 10478 */ 10479 tg->uclamp_pct[clamp_id] = req.percent; 10480 10481 /* Update effective clamps to track the most restrictive value */ 10482 cpu_util_update_eff(of_css(of)); 10483 10484 rcu_read_unlock(); 10485 mutex_unlock(&uclamp_mutex); 10486 10487 return nbytes; 10488 } 10489 10490 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of, 10491 char *buf, size_t nbytes, 10492 loff_t off) 10493 { 10494 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN); 10495 } 10496 10497 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of, 10498 char *buf, size_t nbytes, 10499 loff_t off) 10500 { 10501 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX); 10502 } 10503 10504 static inline void cpu_uclamp_print(struct seq_file *sf, 10505 enum uclamp_id clamp_id) 10506 { 10507 struct task_group *tg; 10508 u64 util_clamp; 10509 u64 percent; 10510 u32 rem; 10511 10512 rcu_read_lock(); 10513 tg = css_tg(seq_css(sf)); 10514 util_clamp = tg->uclamp_req[clamp_id].value; 10515 rcu_read_unlock(); 10516 10517 if (util_clamp == SCHED_CAPACITY_SCALE) { 10518 seq_puts(sf, "max\n"); 10519 return; 10520 } 10521 10522 percent = tg->uclamp_pct[clamp_id]; 10523 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem); 10524 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem); 10525 } 10526 10527 static int cpu_uclamp_min_show(struct seq_file *sf, void *v) 10528 { 10529 cpu_uclamp_print(sf, UCLAMP_MIN); 10530 return 0; 10531 } 10532 10533 static int cpu_uclamp_max_show(struct seq_file *sf, void *v) 10534 { 10535 cpu_uclamp_print(sf, UCLAMP_MAX); 10536 return 0; 10537 } 10538 #endif /* CONFIG_UCLAMP_TASK_GROUP */ 10539 10540 #ifdef CONFIG_FAIR_GROUP_SCHED 10541 static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 10542 struct cftype *cftype, u64 shareval) 10543 { 10544 if (shareval > scale_load_down(ULONG_MAX)) 10545 shareval = MAX_SHARES; 10546 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 10547 } 10548 10549 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 10550 struct cftype *cft) 10551 { 10552 struct task_group *tg = css_tg(css); 10553 10554 return (u64) scale_load_down(tg->shares); 10555 } 10556 10557 #ifdef CONFIG_CFS_BANDWIDTH 10558 static DEFINE_MUTEX(cfs_constraints_mutex); 10559 10560 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 10561 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 10562 /* More than 203 days if BW_SHIFT equals 20. */ 10563 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC; 10564 10565 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 10566 10567 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota, 10568 u64 burst) 10569 { 10570 int i, ret = 0, runtime_enabled, runtime_was_enabled; 10571 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 10572 10573 if (tg == &root_task_group) 10574 return -EINVAL; 10575 10576 /* 10577 * Ensure we have at some amount of bandwidth every period. This is 10578 * to prevent reaching a state of large arrears when throttled via 10579 * entity_tick() resulting in prolonged exit starvation. 10580 */ 10581 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 10582 return -EINVAL; 10583 10584 /* 10585 * Likewise, bound things on the other side by preventing insane quota 10586 * periods. This also allows us to normalize in computing quota 10587 * feasibility. 10588 */ 10589 if (period > max_cfs_quota_period) 10590 return -EINVAL; 10591 10592 /* 10593 * Bound quota to defend quota against overflow during bandwidth shift. 10594 */ 10595 if (quota != RUNTIME_INF && quota > max_cfs_runtime) 10596 return -EINVAL; 10597 10598 if (quota != RUNTIME_INF && (burst > quota || 10599 burst + quota > max_cfs_runtime)) 10600 return -EINVAL; 10601 10602 /* 10603 * Prevent race between setting of cfs_rq->runtime_enabled and 10604 * unthrottle_offline_cfs_rqs(). 10605 */ 10606 cpus_read_lock(); 10607 mutex_lock(&cfs_constraints_mutex); 10608 ret = __cfs_schedulable(tg, period, quota); 10609 if (ret) 10610 goto out_unlock; 10611 10612 runtime_enabled = quota != RUNTIME_INF; 10613 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 10614 /* 10615 * If we need to toggle cfs_bandwidth_used, off->on must occur 10616 * before making related changes, and on->off must occur afterwards 10617 */ 10618 if (runtime_enabled && !runtime_was_enabled) 10619 cfs_bandwidth_usage_inc(); 10620 raw_spin_lock_irq(&cfs_b->lock); 10621 cfs_b->period = ns_to_ktime(period); 10622 cfs_b->quota = quota; 10623 cfs_b->burst = burst; 10624 10625 __refill_cfs_bandwidth_runtime(cfs_b); 10626 10627 /* Restart the period timer (if active) to handle new period expiry: */ 10628 if (runtime_enabled) 10629 start_cfs_bandwidth(cfs_b); 10630 10631 raw_spin_unlock_irq(&cfs_b->lock); 10632 10633 for_each_online_cpu(i) { 10634 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 10635 struct rq *rq = cfs_rq->rq; 10636 struct rq_flags rf; 10637 10638 rq_lock_irq(rq, &rf); 10639 cfs_rq->runtime_enabled = runtime_enabled; 10640 cfs_rq->runtime_remaining = 0; 10641 10642 if (cfs_rq->throttled) 10643 unthrottle_cfs_rq(cfs_rq); 10644 rq_unlock_irq(rq, &rf); 10645 } 10646 if (runtime_was_enabled && !runtime_enabled) 10647 cfs_bandwidth_usage_dec(); 10648 out_unlock: 10649 mutex_unlock(&cfs_constraints_mutex); 10650 cpus_read_unlock(); 10651 10652 return ret; 10653 } 10654 10655 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 10656 { 10657 u64 quota, period, burst; 10658 10659 period = ktime_to_ns(tg->cfs_bandwidth.period); 10660 burst = tg->cfs_bandwidth.burst; 10661 if (cfs_quota_us < 0) 10662 quota = RUNTIME_INF; 10663 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC) 10664 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 10665 else 10666 return -EINVAL; 10667 10668 return tg_set_cfs_bandwidth(tg, period, quota, burst); 10669 } 10670 10671 static long tg_get_cfs_quota(struct task_group *tg) 10672 { 10673 u64 quota_us; 10674 10675 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 10676 return -1; 10677 10678 quota_us = tg->cfs_bandwidth.quota; 10679 do_div(quota_us, NSEC_PER_USEC); 10680 10681 return quota_us; 10682 } 10683 10684 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 10685 { 10686 u64 quota, period, burst; 10687 10688 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC) 10689 return -EINVAL; 10690 10691 period = (u64)cfs_period_us * NSEC_PER_USEC; 10692 quota = tg->cfs_bandwidth.quota; 10693 burst = tg->cfs_bandwidth.burst; 10694 10695 return tg_set_cfs_bandwidth(tg, period, quota, burst); 10696 } 10697 10698 static long tg_get_cfs_period(struct task_group *tg) 10699 { 10700 u64 cfs_period_us; 10701 10702 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 10703 do_div(cfs_period_us, NSEC_PER_USEC); 10704 10705 return cfs_period_us; 10706 } 10707 10708 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us) 10709 { 10710 u64 quota, period, burst; 10711 10712 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC) 10713 return -EINVAL; 10714 10715 burst = (u64)cfs_burst_us * NSEC_PER_USEC; 10716 period = ktime_to_ns(tg->cfs_bandwidth.period); 10717 quota = tg->cfs_bandwidth.quota; 10718 10719 return tg_set_cfs_bandwidth(tg, period, quota, burst); 10720 } 10721 10722 static long tg_get_cfs_burst(struct task_group *tg) 10723 { 10724 u64 burst_us; 10725 10726 burst_us = tg->cfs_bandwidth.burst; 10727 do_div(burst_us, NSEC_PER_USEC); 10728 10729 return burst_us; 10730 } 10731 10732 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 10733 struct cftype *cft) 10734 { 10735 return tg_get_cfs_quota(css_tg(css)); 10736 } 10737 10738 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 10739 struct cftype *cftype, s64 cfs_quota_us) 10740 { 10741 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 10742 } 10743 10744 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 10745 struct cftype *cft) 10746 { 10747 return tg_get_cfs_period(css_tg(css)); 10748 } 10749 10750 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 10751 struct cftype *cftype, u64 cfs_period_us) 10752 { 10753 return tg_set_cfs_period(css_tg(css), cfs_period_us); 10754 } 10755 10756 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css, 10757 struct cftype *cft) 10758 { 10759 return tg_get_cfs_burst(css_tg(css)); 10760 } 10761 10762 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css, 10763 struct cftype *cftype, u64 cfs_burst_us) 10764 { 10765 return tg_set_cfs_burst(css_tg(css), cfs_burst_us); 10766 } 10767 10768 struct cfs_schedulable_data { 10769 struct task_group *tg; 10770 u64 period, quota; 10771 }; 10772 10773 /* 10774 * normalize group quota/period to be quota/max_period 10775 * note: units are usecs 10776 */ 10777 static u64 normalize_cfs_quota(struct task_group *tg, 10778 struct cfs_schedulable_data *d) 10779 { 10780 u64 quota, period; 10781 10782 if (tg == d->tg) { 10783 period = d->period; 10784 quota = d->quota; 10785 } else { 10786 period = tg_get_cfs_period(tg); 10787 quota = tg_get_cfs_quota(tg); 10788 } 10789 10790 /* note: these should typically be equivalent */ 10791 if (quota == RUNTIME_INF || quota == -1) 10792 return RUNTIME_INF; 10793 10794 return to_ratio(period, quota); 10795 } 10796 10797 static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 10798 { 10799 struct cfs_schedulable_data *d = data; 10800 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 10801 s64 quota = 0, parent_quota = -1; 10802 10803 if (!tg->parent) { 10804 quota = RUNTIME_INF; 10805 } else { 10806 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 10807 10808 quota = normalize_cfs_quota(tg, d); 10809 parent_quota = parent_b->hierarchical_quota; 10810 10811 /* 10812 * Ensure max(child_quota) <= parent_quota. On cgroup2, 10813 * always take the min. On cgroup1, only inherit when no 10814 * limit is set: 10815 */ 10816 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) { 10817 quota = min(quota, parent_quota); 10818 } else { 10819 if (quota == RUNTIME_INF) 10820 quota = parent_quota; 10821 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 10822 return -EINVAL; 10823 } 10824 } 10825 cfs_b->hierarchical_quota = quota; 10826 10827 return 0; 10828 } 10829 10830 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 10831 { 10832 int ret; 10833 struct cfs_schedulable_data data = { 10834 .tg = tg, 10835 .period = period, 10836 .quota = quota, 10837 }; 10838 10839 if (quota != RUNTIME_INF) { 10840 do_div(data.period, NSEC_PER_USEC); 10841 do_div(data.quota, NSEC_PER_USEC); 10842 } 10843 10844 rcu_read_lock(); 10845 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 10846 rcu_read_unlock(); 10847 10848 return ret; 10849 } 10850 10851 static int cpu_cfs_stat_show(struct seq_file *sf, void *v) 10852 { 10853 struct task_group *tg = css_tg(seq_css(sf)); 10854 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 10855 10856 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 10857 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 10858 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 10859 10860 if (schedstat_enabled() && tg != &root_task_group) { 10861 struct sched_statistics *stats; 10862 u64 ws = 0; 10863 int i; 10864 10865 for_each_possible_cpu(i) { 10866 stats = __schedstats_from_se(tg->se[i]); 10867 ws += schedstat_val(stats->wait_sum); 10868 } 10869 10870 seq_printf(sf, "wait_sum %llu\n", ws); 10871 } 10872 10873 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst); 10874 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time); 10875 10876 return 0; 10877 } 10878 #endif /* CONFIG_CFS_BANDWIDTH */ 10879 #endif /* CONFIG_FAIR_GROUP_SCHED */ 10880 10881 #ifdef CONFIG_RT_GROUP_SCHED 10882 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 10883 struct cftype *cft, s64 val) 10884 { 10885 return sched_group_set_rt_runtime(css_tg(css), val); 10886 } 10887 10888 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 10889 struct cftype *cft) 10890 { 10891 return sched_group_rt_runtime(css_tg(css)); 10892 } 10893 10894 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 10895 struct cftype *cftype, u64 rt_period_us) 10896 { 10897 return sched_group_set_rt_period(css_tg(css), rt_period_us); 10898 } 10899 10900 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 10901 struct cftype *cft) 10902 { 10903 return sched_group_rt_period(css_tg(css)); 10904 } 10905 #endif /* CONFIG_RT_GROUP_SCHED */ 10906 10907 #ifdef CONFIG_FAIR_GROUP_SCHED 10908 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css, 10909 struct cftype *cft) 10910 { 10911 return css_tg(css)->idle; 10912 } 10913 10914 static int cpu_idle_write_s64(struct cgroup_subsys_state *css, 10915 struct cftype *cft, s64 idle) 10916 { 10917 return sched_group_set_idle(css_tg(css), idle); 10918 } 10919 #endif 10920 10921 static struct cftype cpu_legacy_files[] = { 10922 #ifdef CONFIG_FAIR_GROUP_SCHED 10923 { 10924 .name = "shares", 10925 .read_u64 = cpu_shares_read_u64, 10926 .write_u64 = cpu_shares_write_u64, 10927 }, 10928 { 10929 .name = "idle", 10930 .read_s64 = cpu_idle_read_s64, 10931 .write_s64 = cpu_idle_write_s64, 10932 }, 10933 #endif 10934 #ifdef CONFIG_CFS_BANDWIDTH 10935 { 10936 .name = "cfs_quota_us", 10937 .read_s64 = cpu_cfs_quota_read_s64, 10938 .write_s64 = cpu_cfs_quota_write_s64, 10939 }, 10940 { 10941 .name = "cfs_period_us", 10942 .read_u64 = cpu_cfs_period_read_u64, 10943 .write_u64 = cpu_cfs_period_write_u64, 10944 }, 10945 { 10946 .name = "cfs_burst_us", 10947 .read_u64 = cpu_cfs_burst_read_u64, 10948 .write_u64 = cpu_cfs_burst_write_u64, 10949 }, 10950 { 10951 .name = "stat", 10952 .seq_show = cpu_cfs_stat_show, 10953 }, 10954 #endif 10955 #ifdef CONFIG_RT_GROUP_SCHED 10956 { 10957 .name = "rt_runtime_us", 10958 .read_s64 = cpu_rt_runtime_read, 10959 .write_s64 = cpu_rt_runtime_write, 10960 }, 10961 { 10962 .name = "rt_period_us", 10963 .read_u64 = cpu_rt_period_read_uint, 10964 .write_u64 = cpu_rt_period_write_uint, 10965 }, 10966 #endif 10967 #ifdef CONFIG_UCLAMP_TASK_GROUP 10968 { 10969 .name = "uclamp.min", 10970 .flags = CFTYPE_NOT_ON_ROOT, 10971 .seq_show = cpu_uclamp_min_show, 10972 .write = cpu_uclamp_min_write, 10973 }, 10974 { 10975 .name = "uclamp.max", 10976 .flags = CFTYPE_NOT_ON_ROOT, 10977 .seq_show = cpu_uclamp_max_show, 10978 .write = cpu_uclamp_max_write, 10979 }, 10980 #endif 10981 { } /* Terminate */ 10982 }; 10983 10984 static int cpu_extra_stat_show(struct seq_file *sf, 10985 struct cgroup_subsys_state *css) 10986 { 10987 #ifdef CONFIG_CFS_BANDWIDTH 10988 { 10989 struct task_group *tg = css_tg(css); 10990 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 10991 u64 throttled_usec, burst_usec; 10992 10993 throttled_usec = cfs_b->throttled_time; 10994 do_div(throttled_usec, NSEC_PER_USEC); 10995 burst_usec = cfs_b->burst_time; 10996 do_div(burst_usec, NSEC_PER_USEC); 10997 10998 seq_printf(sf, "nr_periods %d\n" 10999 "nr_throttled %d\n" 11000 "throttled_usec %llu\n" 11001 "nr_bursts %d\n" 11002 "burst_usec %llu\n", 11003 cfs_b->nr_periods, cfs_b->nr_throttled, 11004 throttled_usec, cfs_b->nr_burst, burst_usec); 11005 } 11006 #endif 11007 return 0; 11008 } 11009 11010 #ifdef CONFIG_FAIR_GROUP_SCHED 11011 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css, 11012 struct cftype *cft) 11013 { 11014 struct task_group *tg = css_tg(css); 11015 u64 weight = scale_load_down(tg->shares); 11016 11017 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024); 11018 } 11019 11020 static int cpu_weight_write_u64(struct cgroup_subsys_state *css, 11021 struct cftype *cft, u64 weight) 11022 { 11023 /* 11024 * cgroup weight knobs should use the common MIN, DFL and MAX 11025 * values which are 1, 100 and 10000 respectively. While it loses 11026 * a bit of range on both ends, it maps pretty well onto the shares 11027 * value used by scheduler and the round-trip conversions preserve 11028 * the original value over the entire range. 11029 */ 11030 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX) 11031 return -ERANGE; 11032 11033 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL); 11034 11035 return sched_group_set_shares(css_tg(css), scale_load(weight)); 11036 } 11037 11038 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css, 11039 struct cftype *cft) 11040 { 11041 unsigned long weight = scale_load_down(css_tg(css)->shares); 11042 int last_delta = INT_MAX; 11043 int prio, delta; 11044 11045 /* find the closest nice value to the current weight */ 11046 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) { 11047 delta = abs(sched_prio_to_weight[prio] - weight); 11048 if (delta >= last_delta) 11049 break; 11050 last_delta = delta; 11051 } 11052 11053 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO); 11054 } 11055 11056 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css, 11057 struct cftype *cft, s64 nice) 11058 { 11059 unsigned long weight; 11060 int idx; 11061 11062 if (nice < MIN_NICE || nice > MAX_NICE) 11063 return -ERANGE; 11064 11065 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO; 11066 idx = array_index_nospec(idx, 40); 11067 weight = sched_prio_to_weight[idx]; 11068 11069 return sched_group_set_shares(css_tg(css), scale_load(weight)); 11070 } 11071 #endif 11072 11073 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf, 11074 long period, long quota) 11075 { 11076 if (quota < 0) 11077 seq_puts(sf, "max"); 11078 else 11079 seq_printf(sf, "%ld", quota); 11080 11081 seq_printf(sf, " %ld\n", period); 11082 } 11083 11084 /* caller should put the current value in *@periodp before calling */ 11085 static int __maybe_unused cpu_period_quota_parse(char *buf, 11086 u64 *periodp, u64 *quotap) 11087 { 11088 char tok[21]; /* U64_MAX */ 11089 11090 if (sscanf(buf, "%20s %llu", tok, periodp) < 1) 11091 return -EINVAL; 11092 11093 *periodp *= NSEC_PER_USEC; 11094 11095 if (sscanf(tok, "%llu", quotap)) 11096 *quotap *= NSEC_PER_USEC; 11097 else if (!strcmp(tok, "max")) 11098 *quotap = RUNTIME_INF; 11099 else 11100 return -EINVAL; 11101 11102 return 0; 11103 } 11104 11105 #ifdef CONFIG_CFS_BANDWIDTH 11106 static int cpu_max_show(struct seq_file *sf, void *v) 11107 { 11108 struct task_group *tg = css_tg(seq_css(sf)); 11109 11110 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg)); 11111 return 0; 11112 } 11113 11114 static ssize_t cpu_max_write(struct kernfs_open_file *of, 11115 char *buf, size_t nbytes, loff_t off) 11116 { 11117 struct task_group *tg = css_tg(of_css(of)); 11118 u64 period = tg_get_cfs_period(tg); 11119 u64 burst = tg_get_cfs_burst(tg); 11120 u64 quota; 11121 int ret; 11122 11123 ret = cpu_period_quota_parse(buf, &period, "a); 11124 if (!ret) 11125 ret = tg_set_cfs_bandwidth(tg, period, quota, burst); 11126 return ret ?: nbytes; 11127 } 11128 #endif 11129 11130 static struct cftype cpu_files[] = { 11131 #ifdef CONFIG_FAIR_GROUP_SCHED 11132 { 11133 .name = "weight", 11134 .flags = CFTYPE_NOT_ON_ROOT, 11135 .read_u64 = cpu_weight_read_u64, 11136 .write_u64 = cpu_weight_write_u64, 11137 }, 11138 { 11139 .name = "weight.nice", 11140 .flags = CFTYPE_NOT_ON_ROOT, 11141 .read_s64 = cpu_weight_nice_read_s64, 11142 .write_s64 = cpu_weight_nice_write_s64, 11143 }, 11144 { 11145 .name = "idle", 11146 .flags = CFTYPE_NOT_ON_ROOT, 11147 .read_s64 = cpu_idle_read_s64, 11148 .write_s64 = cpu_idle_write_s64, 11149 }, 11150 #endif 11151 #ifdef CONFIG_CFS_BANDWIDTH 11152 { 11153 .name = "max", 11154 .flags = CFTYPE_NOT_ON_ROOT, 11155 .seq_show = cpu_max_show, 11156 .write = cpu_max_write, 11157 }, 11158 { 11159 .name = "max.burst", 11160 .flags = CFTYPE_NOT_ON_ROOT, 11161 .read_u64 = cpu_cfs_burst_read_u64, 11162 .write_u64 = cpu_cfs_burst_write_u64, 11163 }, 11164 #endif 11165 #ifdef CONFIG_UCLAMP_TASK_GROUP 11166 { 11167 .name = "uclamp.min", 11168 .flags = CFTYPE_NOT_ON_ROOT, 11169 .seq_show = cpu_uclamp_min_show, 11170 .write = cpu_uclamp_min_write, 11171 }, 11172 { 11173 .name = "uclamp.max", 11174 .flags = CFTYPE_NOT_ON_ROOT, 11175 .seq_show = cpu_uclamp_max_show, 11176 .write = cpu_uclamp_max_write, 11177 }, 11178 #endif 11179 { } /* terminate */ 11180 }; 11181 11182 struct cgroup_subsys cpu_cgrp_subsys = { 11183 .css_alloc = cpu_cgroup_css_alloc, 11184 .css_online = cpu_cgroup_css_online, 11185 .css_released = cpu_cgroup_css_released, 11186 .css_free = cpu_cgroup_css_free, 11187 .css_extra_stat_show = cpu_extra_stat_show, 11188 .fork = cpu_cgroup_fork, 11189 .can_attach = cpu_cgroup_can_attach, 11190 .attach = cpu_cgroup_attach, 11191 .legacy_cftypes = cpu_legacy_files, 11192 .dfl_cftypes = cpu_files, 11193 .early_init = true, 11194 .threaded = true, 11195 }; 11196 11197 #endif /* CONFIG_CGROUP_SCHED */ 11198 11199 void dump_cpu_task(int cpu) 11200 { 11201 pr_info("Task dump for CPU %d:\n", cpu); 11202 sched_show_task(cpu_curr(cpu)); 11203 } 11204 11205 /* 11206 * Nice levels are multiplicative, with a gentle 10% change for every 11207 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to 11208 * nice 1, it will get ~10% less CPU time than another CPU-bound task 11209 * that remained on nice 0. 11210 * 11211 * The "10% effect" is relative and cumulative: from _any_ nice level, 11212 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level 11213 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. 11214 * If a task goes up by ~10% and another task goes down by ~10% then 11215 * the relative distance between them is ~25%.) 11216 */ 11217 const int sched_prio_to_weight[40] = { 11218 /* -20 */ 88761, 71755, 56483, 46273, 36291, 11219 /* -15 */ 29154, 23254, 18705, 14949, 11916, 11220 /* -10 */ 9548, 7620, 6100, 4904, 3906, 11221 /* -5 */ 3121, 2501, 1991, 1586, 1277, 11222 /* 0 */ 1024, 820, 655, 526, 423, 11223 /* 5 */ 335, 272, 215, 172, 137, 11224 /* 10 */ 110, 87, 70, 56, 45, 11225 /* 15 */ 36, 29, 23, 18, 15, 11226 }; 11227 11228 /* 11229 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated. 11230 * 11231 * In cases where the weight does not change often, we can use the 11232 * precalculated inverse to speed up arithmetics by turning divisions 11233 * into multiplications: 11234 */ 11235 const u32 sched_prio_to_wmult[40] = { 11236 /* -20 */ 48388, 59856, 76040, 92818, 118348, 11237 /* -15 */ 147320, 184698, 229616, 287308, 360437, 11238 /* -10 */ 449829, 563644, 704093, 875809, 1099582, 11239 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, 11240 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, 11241 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, 11242 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, 11243 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, 11244 }; 11245 11246 void call_trace_sched_update_nr_running(struct rq *rq, int count) 11247 { 11248 trace_sched_update_nr_running_tp(rq, count); 11249 } 11250