1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* 3 * Scheduler internal types and methods: 4 */ 5 #include <linux/sched.h> 6 7 #include <linux/sched/autogroup.h> 8 #include <linux/sched/clock.h> 9 #include <linux/sched/coredump.h> 10 #include <linux/sched/cpufreq.h> 11 #include <linux/sched/cputime.h> 12 #include <linux/sched/deadline.h> 13 #include <linux/sched/debug.h> 14 #include <linux/sched/hotplug.h> 15 #include <linux/sched/idle.h> 16 #include <linux/sched/init.h> 17 #include <linux/sched/isolation.h> 18 #include <linux/sched/jobctl.h> 19 #include <linux/sched/loadavg.h> 20 #include <linux/sched/mm.h> 21 #include <linux/sched/nohz.h> 22 #include <linux/sched/numa_balancing.h> 23 #include <linux/sched/prio.h> 24 #include <linux/sched/rt.h> 25 #include <linux/sched/signal.h> 26 #include <linux/sched/smt.h> 27 #include <linux/sched/stat.h> 28 #include <linux/sched/sysctl.h> 29 #include <linux/sched/task.h> 30 #include <linux/sched/task_stack.h> 31 #include <linux/sched/topology.h> 32 #include <linux/sched/user.h> 33 #include <linux/sched/wake_q.h> 34 #include <linux/sched/xacct.h> 35 36 #include <uapi/linux/sched/types.h> 37 38 #include <linux/binfmts.h> 39 #include <linux/blkdev.h> 40 #include <linux/compat.h> 41 #include <linux/context_tracking.h> 42 #include <linux/cpufreq.h> 43 #include <linux/cpuidle.h> 44 #include <linux/cpuset.h> 45 #include <linux/ctype.h> 46 #include <linux/debugfs.h> 47 #include <linux/delayacct.h> 48 #include <linux/energy_model.h> 49 #include <linux/init_task.h> 50 #include <linux/kprobes.h> 51 #include <linux/kthread.h> 52 #include <linux/membarrier.h> 53 #include <linux/migrate.h> 54 #include <linux/mmu_context.h> 55 #include <linux/nmi.h> 56 #include <linux/proc_fs.h> 57 #include <linux/prefetch.h> 58 #include <linux/profile.h> 59 #include <linux/psi.h> 60 #include <linux/rcupdate_wait.h> 61 #include <linux/security.h> 62 #include <linux/stop_machine.h> 63 #include <linux/suspend.h> 64 #include <linux/swait.h> 65 #include <linux/syscalls.h> 66 #include <linux/task_work.h> 67 #include <linux/tsacct_kern.h> 68 69 #include <asm/tlb.h> 70 71 #ifdef CONFIG_PARAVIRT 72 # include <asm/paravirt.h> 73 #endif 74 75 #include "cpupri.h" 76 #include "cpudeadline.h" 77 78 #ifdef CONFIG_SCHED_DEBUG 79 # define SCHED_WARN_ON(x) WARN_ONCE(x, #x) 80 #else 81 # define SCHED_WARN_ON(x) ({ (void)(x), 0; }) 82 #endif 83 84 struct rq; 85 struct cpuidle_state; 86 87 /* task_struct::on_rq states: */ 88 #define TASK_ON_RQ_QUEUED 1 89 #define TASK_ON_RQ_MIGRATING 2 90 91 extern __read_mostly int scheduler_running; 92 93 extern unsigned long calc_load_update; 94 extern atomic_long_t calc_load_tasks; 95 96 extern void calc_global_load_tick(struct rq *this_rq); 97 extern long calc_load_fold_active(struct rq *this_rq, long adjust); 98 99 /* 100 * Helpers for converting nanosecond timing to jiffy resolution 101 */ 102 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) 103 104 /* 105 * Increase resolution of nice-level calculations for 64-bit architectures. 106 * The extra resolution improves shares distribution and load balancing of 107 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup 108 * hierarchies, especially on larger systems. This is not a user-visible change 109 * and does not change the user-interface for setting shares/weights. 110 * 111 * We increase resolution only if we have enough bits to allow this increased 112 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit 113 * are pretty high and the returns do not justify the increased costs. 114 * 115 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to 116 * increase coverage and consistency always enable it on 64-bit platforms. 117 */ 118 #ifdef CONFIG_64BIT 119 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT) 120 # define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT) 121 # define scale_load_down(w) \ 122 ({ \ 123 unsigned long __w = (w); \ 124 if (__w) \ 125 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \ 126 __w; \ 127 }) 128 #else 129 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT) 130 # define scale_load(w) (w) 131 # define scale_load_down(w) (w) 132 #endif 133 134 /* 135 * Task weight (visible to users) and its load (invisible to users) have 136 * independent resolution, but they should be well calibrated. We use 137 * scale_load() and scale_load_down(w) to convert between them. The 138 * following must be true: 139 * 140 * scale_load(sched_prio_to_weight[USER_PRIO(NICE_TO_PRIO(0))]) == NICE_0_LOAD 141 * 142 */ 143 #define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT) 144 145 /* 146 * Single value that decides SCHED_DEADLINE internal math precision. 147 * 10 -> just above 1us 148 * 9 -> just above 0.5us 149 */ 150 #define DL_SCALE 10 151 152 /* 153 * Single value that denotes runtime == period, ie unlimited time. 154 */ 155 #define RUNTIME_INF ((u64)~0ULL) 156 157 static inline int idle_policy(int policy) 158 { 159 return policy == SCHED_IDLE; 160 } 161 static inline int fair_policy(int policy) 162 { 163 return policy == SCHED_NORMAL || policy == SCHED_BATCH; 164 } 165 166 static inline int rt_policy(int policy) 167 { 168 return policy == SCHED_FIFO || policy == SCHED_RR; 169 } 170 171 static inline int dl_policy(int policy) 172 { 173 return policy == SCHED_DEADLINE; 174 } 175 static inline bool valid_policy(int policy) 176 { 177 return idle_policy(policy) || fair_policy(policy) || 178 rt_policy(policy) || dl_policy(policy); 179 } 180 181 static inline int task_has_idle_policy(struct task_struct *p) 182 { 183 return idle_policy(p->policy); 184 } 185 186 static inline int task_has_rt_policy(struct task_struct *p) 187 { 188 return rt_policy(p->policy); 189 } 190 191 static inline int task_has_dl_policy(struct task_struct *p) 192 { 193 return dl_policy(p->policy); 194 } 195 196 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT) 197 198 /* 199 * !! For sched_setattr_nocheck() (kernel) only !! 200 * 201 * This is actually gross. :( 202 * 203 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE 204 * tasks, but still be able to sleep. We need this on platforms that cannot 205 * atomically change clock frequency. Remove once fast switching will be 206 * available on such platforms. 207 * 208 * SUGOV stands for SchedUtil GOVernor. 209 */ 210 #define SCHED_FLAG_SUGOV 0x10000000 211 212 static inline bool dl_entity_is_special(struct sched_dl_entity *dl_se) 213 { 214 #ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL 215 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV); 216 #else 217 return false; 218 #endif 219 } 220 221 /* 222 * Tells if entity @a should preempt entity @b. 223 */ 224 static inline bool 225 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b) 226 { 227 return dl_entity_is_special(a) || 228 dl_time_before(a->deadline, b->deadline); 229 } 230 231 /* 232 * This is the priority-queue data structure of the RT scheduling class: 233 */ 234 struct rt_prio_array { 235 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ 236 struct list_head queue[MAX_RT_PRIO]; 237 }; 238 239 struct rt_bandwidth { 240 /* nests inside the rq lock: */ 241 raw_spinlock_t rt_runtime_lock; 242 ktime_t rt_period; 243 u64 rt_runtime; 244 struct hrtimer rt_period_timer; 245 unsigned int rt_period_active; 246 }; 247 248 void __dl_clear_params(struct task_struct *p); 249 250 /* 251 * To keep the bandwidth of -deadline tasks and groups under control 252 * we need some place where: 253 * - store the maximum -deadline bandwidth of the system (the group); 254 * - cache the fraction of that bandwidth that is currently allocated. 255 * 256 * This is all done in the data structure below. It is similar to the 257 * one used for RT-throttling (rt_bandwidth), with the main difference 258 * that, since here we are only interested in admission control, we 259 * do not decrease any runtime while the group "executes", neither we 260 * need a timer to replenish it. 261 * 262 * With respect to SMP, the bandwidth is given on a per-CPU basis, 263 * meaning that: 264 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU; 265 * - dl_total_bw array contains, in the i-eth element, the currently 266 * allocated bandwidth on the i-eth CPU. 267 * Moreover, groups consume bandwidth on each CPU, while tasks only 268 * consume bandwidth on the CPU they're running on. 269 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw 270 * that will be shown the next time the proc or cgroup controls will 271 * be red. It on its turn can be changed by writing on its own 272 * control. 273 */ 274 struct dl_bandwidth { 275 raw_spinlock_t dl_runtime_lock; 276 u64 dl_runtime; 277 u64 dl_period; 278 }; 279 280 static inline int dl_bandwidth_enabled(void) 281 { 282 return sysctl_sched_rt_runtime >= 0; 283 } 284 285 struct dl_bw { 286 raw_spinlock_t lock; 287 u64 bw; 288 u64 total_bw; 289 }; 290 291 static inline void __dl_update(struct dl_bw *dl_b, s64 bw); 292 293 static inline 294 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus) 295 { 296 dl_b->total_bw -= tsk_bw; 297 __dl_update(dl_b, (s32)tsk_bw / cpus); 298 } 299 300 static inline 301 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus) 302 { 303 dl_b->total_bw += tsk_bw; 304 __dl_update(dl_b, -((s32)tsk_bw / cpus)); 305 } 306 307 static inline 308 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 309 { 310 return dl_b->bw != -1 && 311 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 312 } 313 314 extern void init_dl_bw(struct dl_bw *dl_b); 315 extern int sched_dl_global_validate(void); 316 extern void sched_dl_do_global(void); 317 extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr); 318 extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr); 319 extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr); 320 extern bool __checkparam_dl(const struct sched_attr *attr); 321 extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr); 322 extern int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed); 323 extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial); 324 extern bool dl_cpu_busy(unsigned int cpu); 325 326 #ifdef CONFIG_CGROUP_SCHED 327 328 #include <linux/cgroup.h> 329 #include <linux/psi.h> 330 331 struct cfs_rq; 332 struct rt_rq; 333 334 extern struct list_head task_groups; 335 336 struct cfs_bandwidth { 337 #ifdef CONFIG_CFS_BANDWIDTH 338 raw_spinlock_t lock; 339 ktime_t period; 340 u64 quota; 341 u64 runtime; 342 s64 hierarchical_quota; 343 344 u8 idle; 345 u8 period_active; 346 u8 distribute_running; 347 u8 slack_started; 348 struct hrtimer period_timer; 349 struct hrtimer slack_timer; 350 struct list_head throttled_cfs_rq; 351 352 /* Statistics: */ 353 int nr_periods; 354 int nr_throttled; 355 u64 throttled_time; 356 #endif 357 }; 358 359 /* Task group related information */ 360 struct task_group { 361 struct cgroup_subsys_state css; 362 363 #ifdef CONFIG_FAIR_GROUP_SCHED 364 /* schedulable entities of this group on each CPU */ 365 struct sched_entity **se; 366 /* runqueue "owned" by this group on each CPU */ 367 struct cfs_rq **cfs_rq; 368 unsigned long shares; 369 370 #ifdef CONFIG_SMP 371 /* 372 * load_avg can be heavily contended at clock tick time, so put 373 * it in its own cacheline separated from the fields above which 374 * will also be accessed at each tick. 375 */ 376 atomic_long_t load_avg ____cacheline_aligned; 377 #endif 378 #endif 379 380 #ifdef CONFIG_RT_GROUP_SCHED 381 struct sched_rt_entity **rt_se; 382 struct rt_rq **rt_rq; 383 384 struct rt_bandwidth rt_bandwidth; 385 #endif 386 387 struct rcu_head rcu; 388 struct list_head list; 389 390 struct task_group *parent; 391 struct list_head siblings; 392 struct list_head children; 393 394 #ifdef CONFIG_SCHED_AUTOGROUP 395 struct autogroup *autogroup; 396 #endif 397 398 struct cfs_bandwidth cfs_bandwidth; 399 400 #ifdef CONFIG_UCLAMP_TASK_GROUP 401 /* The two decimal precision [%] value requested from user-space */ 402 unsigned int uclamp_pct[UCLAMP_CNT]; 403 /* Clamp values requested for a task group */ 404 struct uclamp_se uclamp_req[UCLAMP_CNT]; 405 /* Effective clamp values used for a task group */ 406 struct uclamp_se uclamp[UCLAMP_CNT]; 407 #endif 408 409 }; 410 411 #ifdef CONFIG_FAIR_GROUP_SCHED 412 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD 413 414 /* 415 * A weight of 0 or 1 can cause arithmetics problems. 416 * A weight of a cfs_rq is the sum of weights of which entities 417 * are queued on this cfs_rq, so a weight of a entity should not be 418 * too large, so as the shares value of a task group. 419 * (The default weight is 1024 - so there's no practical 420 * limitation from this.) 421 */ 422 #define MIN_SHARES (1UL << 1) 423 #define MAX_SHARES (1UL << 18) 424 #endif 425 426 typedef int (*tg_visitor)(struct task_group *, void *); 427 428 extern int walk_tg_tree_from(struct task_group *from, 429 tg_visitor down, tg_visitor up, void *data); 430 431 /* 432 * Iterate the full tree, calling @down when first entering a node and @up when 433 * leaving it for the final time. 434 * 435 * Caller must hold rcu_lock or sufficient equivalent. 436 */ 437 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) 438 { 439 return walk_tg_tree_from(&root_task_group, down, up, data); 440 } 441 442 extern int tg_nop(struct task_group *tg, void *data); 443 444 extern void free_fair_sched_group(struct task_group *tg); 445 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent); 446 extern void online_fair_sched_group(struct task_group *tg); 447 extern void unregister_fair_sched_group(struct task_group *tg); 448 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 449 struct sched_entity *se, int cpu, 450 struct sched_entity *parent); 451 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 452 453 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b); 454 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 455 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq); 456 457 extern void free_rt_sched_group(struct task_group *tg); 458 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent); 459 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 460 struct sched_rt_entity *rt_se, int cpu, 461 struct sched_rt_entity *parent); 462 extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us); 463 extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us); 464 extern long sched_group_rt_runtime(struct task_group *tg); 465 extern long sched_group_rt_period(struct task_group *tg); 466 extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk); 467 468 extern struct task_group *sched_create_group(struct task_group *parent); 469 extern void sched_online_group(struct task_group *tg, 470 struct task_group *parent); 471 extern void sched_destroy_group(struct task_group *tg); 472 extern void sched_offline_group(struct task_group *tg); 473 474 extern void sched_move_task(struct task_struct *tsk); 475 476 #ifdef CONFIG_FAIR_GROUP_SCHED 477 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 478 479 #ifdef CONFIG_SMP 480 extern void set_task_rq_fair(struct sched_entity *se, 481 struct cfs_rq *prev, struct cfs_rq *next); 482 #else /* !CONFIG_SMP */ 483 static inline void set_task_rq_fair(struct sched_entity *se, 484 struct cfs_rq *prev, struct cfs_rq *next) { } 485 #endif /* CONFIG_SMP */ 486 #endif /* CONFIG_FAIR_GROUP_SCHED */ 487 488 #else /* CONFIG_CGROUP_SCHED */ 489 490 struct cfs_bandwidth { }; 491 492 #endif /* CONFIG_CGROUP_SCHED */ 493 494 /* CFS-related fields in a runqueue */ 495 struct cfs_rq { 496 struct load_weight load; 497 unsigned int nr_running; 498 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */ 499 unsigned int idle_h_nr_running; /* SCHED_IDLE */ 500 501 u64 exec_clock; 502 u64 min_vruntime; 503 #ifndef CONFIG_64BIT 504 u64 min_vruntime_copy; 505 #endif 506 507 struct rb_root_cached tasks_timeline; 508 509 /* 510 * 'curr' points to currently running entity on this cfs_rq. 511 * It is set to NULL otherwise (i.e when none are currently running). 512 */ 513 struct sched_entity *curr; 514 struct sched_entity *next; 515 struct sched_entity *last; 516 struct sched_entity *skip; 517 518 #ifdef CONFIG_SCHED_DEBUG 519 unsigned int nr_spread_over; 520 #endif 521 522 #ifdef CONFIG_SMP 523 /* 524 * CFS load tracking 525 */ 526 struct sched_avg avg; 527 #ifndef CONFIG_64BIT 528 u64 load_last_update_time_copy; 529 #endif 530 struct { 531 raw_spinlock_t lock ____cacheline_aligned; 532 int nr; 533 unsigned long load_avg; 534 unsigned long util_avg; 535 unsigned long runnable_avg; 536 } removed; 537 538 #ifdef CONFIG_FAIR_GROUP_SCHED 539 unsigned long tg_load_avg_contrib; 540 long propagate; 541 long prop_runnable_sum; 542 543 /* 544 * h_load = weight * f(tg) 545 * 546 * Where f(tg) is the recursive weight fraction assigned to 547 * this group. 548 */ 549 unsigned long h_load; 550 u64 last_h_load_update; 551 struct sched_entity *h_load_next; 552 #endif /* CONFIG_FAIR_GROUP_SCHED */ 553 #endif /* CONFIG_SMP */ 554 555 #ifdef CONFIG_FAIR_GROUP_SCHED 556 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */ 557 558 /* 559 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in 560 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities 561 * (like users, containers etc.) 562 * 563 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU. 564 * This list is used during load balance. 565 */ 566 int on_list; 567 struct list_head leaf_cfs_rq_list; 568 struct task_group *tg; /* group that "owns" this runqueue */ 569 570 #ifdef CONFIG_CFS_BANDWIDTH 571 int runtime_enabled; 572 s64 runtime_remaining; 573 574 u64 throttled_clock; 575 u64 throttled_clock_task; 576 u64 throttled_clock_task_time; 577 int throttled; 578 int throttle_count; 579 struct list_head throttled_list; 580 #endif /* CONFIG_CFS_BANDWIDTH */ 581 #endif /* CONFIG_FAIR_GROUP_SCHED */ 582 }; 583 584 static inline int rt_bandwidth_enabled(void) 585 { 586 return sysctl_sched_rt_runtime >= 0; 587 } 588 589 /* RT IPI pull logic requires IRQ_WORK */ 590 #if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP) 591 # define HAVE_RT_PUSH_IPI 592 #endif 593 594 /* Real-Time classes' related field in a runqueue: */ 595 struct rt_rq { 596 struct rt_prio_array active; 597 unsigned int rt_nr_running; 598 unsigned int rr_nr_running; 599 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 600 struct { 601 int curr; /* highest queued rt task prio */ 602 #ifdef CONFIG_SMP 603 int next; /* next highest */ 604 #endif 605 } highest_prio; 606 #endif 607 #ifdef CONFIG_SMP 608 unsigned long rt_nr_migratory; 609 unsigned long rt_nr_total; 610 int overloaded; 611 struct plist_head pushable_tasks; 612 613 #endif /* CONFIG_SMP */ 614 int rt_queued; 615 616 int rt_throttled; 617 u64 rt_time; 618 u64 rt_runtime; 619 /* Nests inside the rq lock: */ 620 raw_spinlock_t rt_runtime_lock; 621 622 #ifdef CONFIG_RT_GROUP_SCHED 623 unsigned long rt_nr_boosted; 624 625 struct rq *rq; 626 struct task_group *tg; 627 #endif 628 }; 629 630 static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq) 631 { 632 return rt_rq->rt_queued && rt_rq->rt_nr_running; 633 } 634 635 /* Deadline class' related fields in a runqueue */ 636 struct dl_rq { 637 /* runqueue is an rbtree, ordered by deadline */ 638 struct rb_root_cached root; 639 640 unsigned long dl_nr_running; 641 642 #ifdef CONFIG_SMP 643 /* 644 * Deadline values of the currently executing and the 645 * earliest ready task on this rq. Caching these facilitates 646 * the decision whether or not a ready but not running task 647 * should migrate somewhere else. 648 */ 649 struct { 650 u64 curr; 651 u64 next; 652 } earliest_dl; 653 654 unsigned long dl_nr_migratory; 655 int overloaded; 656 657 /* 658 * Tasks on this rq that can be pushed away. They are kept in 659 * an rb-tree, ordered by tasks' deadlines, with caching 660 * of the leftmost (earliest deadline) element. 661 */ 662 struct rb_root_cached pushable_dl_tasks_root; 663 #else 664 struct dl_bw dl_bw; 665 #endif 666 /* 667 * "Active utilization" for this runqueue: increased when a 668 * task wakes up (becomes TASK_RUNNING) and decreased when a 669 * task blocks 670 */ 671 u64 running_bw; 672 673 /* 674 * Utilization of the tasks "assigned" to this runqueue (including 675 * the tasks that are in runqueue and the tasks that executed on this 676 * CPU and blocked). Increased when a task moves to this runqueue, and 677 * decreased when the task moves away (migrates, changes scheduling 678 * policy, or terminates). 679 * This is needed to compute the "inactive utilization" for the 680 * runqueue (inactive utilization = this_bw - running_bw). 681 */ 682 u64 this_bw; 683 u64 extra_bw; 684 685 /* 686 * Inverse of the fraction of CPU utilization that can be reclaimed 687 * by the GRUB algorithm. 688 */ 689 u64 bw_ratio; 690 }; 691 692 #ifdef CONFIG_FAIR_GROUP_SCHED 693 /* An entity is a task if it doesn't "own" a runqueue */ 694 #define entity_is_task(se) (!se->my_q) 695 696 static inline void se_update_runnable(struct sched_entity *se) 697 { 698 if (!entity_is_task(se)) 699 se->runnable_weight = se->my_q->h_nr_running; 700 } 701 702 static inline long se_runnable(struct sched_entity *se) 703 { 704 if (entity_is_task(se)) 705 return !!se->on_rq; 706 else 707 return se->runnable_weight; 708 } 709 710 #else 711 #define entity_is_task(se) 1 712 713 static inline void se_update_runnable(struct sched_entity *se) {} 714 715 static inline long se_runnable(struct sched_entity *se) 716 { 717 return !!se->on_rq; 718 } 719 #endif 720 721 #ifdef CONFIG_SMP 722 /* 723 * XXX we want to get rid of these helpers and use the full load resolution. 724 */ 725 static inline long se_weight(struct sched_entity *se) 726 { 727 return scale_load_down(se->load.weight); 728 } 729 730 731 static inline bool sched_asym_prefer(int a, int b) 732 { 733 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b); 734 } 735 736 struct perf_domain { 737 struct em_perf_domain *em_pd; 738 struct perf_domain *next; 739 struct rcu_head rcu; 740 }; 741 742 /* Scheduling group status flags */ 743 #define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */ 744 #define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */ 745 746 /* 747 * We add the notion of a root-domain which will be used to define per-domain 748 * variables. Each exclusive cpuset essentially defines an island domain by 749 * fully partitioning the member CPUs from any other cpuset. Whenever a new 750 * exclusive cpuset is created, we also create and attach a new root-domain 751 * object. 752 * 753 */ 754 struct root_domain { 755 atomic_t refcount; 756 atomic_t rto_count; 757 struct rcu_head rcu; 758 cpumask_var_t span; 759 cpumask_var_t online; 760 761 /* 762 * Indicate pullable load on at least one CPU, e.g: 763 * - More than one runnable task 764 * - Running task is misfit 765 */ 766 int overload; 767 768 /* Indicate one or more cpus over-utilized (tipping point) */ 769 int overutilized; 770 771 /* 772 * The bit corresponding to a CPU gets set here if such CPU has more 773 * than one runnable -deadline task (as it is below for RT tasks). 774 */ 775 cpumask_var_t dlo_mask; 776 atomic_t dlo_count; 777 struct dl_bw dl_bw; 778 struct cpudl cpudl; 779 780 #ifdef HAVE_RT_PUSH_IPI 781 /* 782 * For IPI pull requests, loop across the rto_mask. 783 */ 784 struct irq_work rto_push_work; 785 raw_spinlock_t rto_lock; 786 /* These are only updated and read within rto_lock */ 787 int rto_loop; 788 int rto_cpu; 789 /* These atomics are updated outside of a lock */ 790 atomic_t rto_loop_next; 791 atomic_t rto_loop_start; 792 #endif 793 /* 794 * The "RT overload" flag: it gets set if a CPU has more than 795 * one runnable RT task. 796 */ 797 cpumask_var_t rto_mask; 798 struct cpupri cpupri; 799 800 unsigned long max_cpu_capacity; 801 802 /* 803 * NULL-terminated list of performance domains intersecting with the 804 * CPUs of the rd. Protected by RCU. 805 */ 806 struct perf_domain __rcu *pd; 807 }; 808 809 extern void init_defrootdomain(void); 810 extern int sched_init_domains(const struct cpumask *cpu_map); 811 extern void rq_attach_root(struct rq *rq, struct root_domain *rd); 812 extern void sched_get_rd(struct root_domain *rd); 813 extern void sched_put_rd(struct root_domain *rd); 814 815 #ifdef HAVE_RT_PUSH_IPI 816 extern void rto_push_irq_work_func(struct irq_work *work); 817 #endif 818 #endif /* CONFIG_SMP */ 819 820 #ifdef CONFIG_UCLAMP_TASK 821 /* 822 * struct uclamp_bucket - Utilization clamp bucket 823 * @value: utilization clamp value for tasks on this clamp bucket 824 * @tasks: number of RUNNABLE tasks on this clamp bucket 825 * 826 * Keep track of how many tasks are RUNNABLE for a given utilization 827 * clamp value. 828 */ 829 struct uclamp_bucket { 830 unsigned long value : bits_per(SCHED_CAPACITY_SCALE); 831 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE); 832 }; 833 834 /* 835 * struct uclamp_rq - rq's utilization clamp 836 * @value: currently active clamp values for a rq 837 * @bucket: utilization clamp buckets affecting a rq 838 * 839 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values. 840 * A clamp value is affecting a rq when there is at least one task RUNNABLE 841 * (or actually running) with that value. 842 * 843 * There are up to UCLAMP_CNT possible different clamp values, currently there 844 * are only two: minimum utilization and maximum utilization. 845 * 846 * All utilization clamping values are MAX aggregated, since: 847 * - for util_min: we want to run the CPU at least at the max of the minimum 848 * utilization required by its currently RUNNABLE tasks. 849 * - for util_max: we want to allow the CPU to run up to the max of the 850 * maximum utilization allowed by its currently RUNNABLE tasks. 851 * 852 * Since on each system we expect only a limited number of different 853 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track 854 * the metrics required to compute all the per-rq utilization clamp values. 855 */ 856 struct uclamp_rq { 857 unsigned int value; 858 struct uclamp_bucket bucket[UCLAMP_BUCKETS]; 859 }; 860 #endif /* CONFIG_UCLAMP_TASK */ 861 862 /* 863 * This is the main, per-CPU runqueue data structure. 864 * 865 * Locking rule: those places that want to lock multiple runqueues 866 * (such as the load balancing or the thread migration code), lock 867 * acquire operations must be ordered by ascending &runqueue. 868 */ 869 struct rq { 870 /* runqueue lock: */ 871 raw_spinlock_t lock; 872 873 /* 874 * nr_running and cpu_load should be in the same cacheline because 875 * remote CPUs use both these fields when doing load calculation. 876 */ 877 unsigned int nr_running; 878 #ifdef CONFIG_NUMA_BALANCING 879 unsigned int nr_numa_running; 880 unsigned int nr_preferred_running; 881 unsigned int numa_migrate_on; 882 #endif 883 #ifdef CONFIG_NO_HZ_COMMON 884 #ifdef CONFIG_SMP 885 unsigned long last_load_update_tick; 886 unsigned long last_blocked_load_update_tick; 887 unsigned int has_blocked_load; 888 #endif /* CONFIG_SMP */ 889 unsigned int nohz_tick_stopped; 890 atomic_t nohz_flags; 891 #endif /* CONFIG_NO_HZ_COMMON */ 892 893 unsigned long nr_load_updates; 894 u64 nr_switches; 895 896 #ifdef CONFIG_UCLAMP_TASK 897 /* Utilization clamp values based on CPU's RUNNABLE tasks */ 898 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned; 899 unsigned int uclamp_flags; 900 #define UCLAMP_FLAG_IDLE 0x01 901 #endif 902 903 struct cfs_rq cfs; 904 struct rt_rq rt; 905 struct dl_rq dl; 906 907 #ifdef CONFIG_FAIR_GROUP_SCHED 908 /* list of leaf cfs_rq on this CPU: */ 909 struct list_head leaf_cfs_rq_list; 910 struct list_head *tmp_alone_branch; 911 #endif /* CONFIG_FAIR_GROUP_SCHED */ 912 913 /* 914 * This is part of a global counter where only the total sum 915 * over all CPUs matters. A task can increase this counter on 916 * one CPU and if it got migrated afterwards it may decrease 917 * it on another CPU. Always updated under the runqueue lock: 918 */ 919 unsigned long nr_uninterruptible; 920 921 struct task_struct __rcu *curr; 922 struct task_struct *idle; 923 struct task_struct *stop; 924 unsigned long next_balance; 925 struct mm_struct *prev_mm; 926 927 unsigned int clock_update_flags; 928 u64 clock; 929 /* Ensure that all clocks are in the same cache line */ 930 u64 clock_task ____cacheline_aligned; 931 u64 clock_pelt; 932 unsigned long lost_idle_time; 933 934 atomic_t nr_iowait; 935 936 #ifdef CONFIG_MEMBARRIER 937 int membarrier_state; 938 #endif 939 940 #ifdef CONFIG_SMP 941 struct root_domain *rd; 942 struct sched_domain __rcu *sd; 943 944 unsigned long cpu_capacity; 945 unsigned long cpu_capacity_orig; 946 947 struct callback_head *balance_callback; 948 949 unsigned char idle_balance; 950 951 unsigned long misfit_task_load; 952 953 /* For active balancing */ 954 int active_balance; 955 int push_cpu; 956 struct cpu_stop_work active_balance_work; 957 958 /* CPU of this runqueue: */ 959 int cpu; 960 int online; 961 962 struct list_head cfs_tasks; 963 964 struct sched_avg avg_rt; 965 struct sched_avg avg_dl; 966 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 967 struct sched_avg avg_irq; 968 #endif 969 #ifdef CONFIG_SCHED_THERMAL_PRESSURE 970 struct sched_avg avg_thermal; 971 #endif 972 u64 idle_stamp; 973 u64 avg_idle; 974 975 /* This is used to determine avg_idle's max value */ 976 u64 max_idle_balance_cost; 977 #endif 978 979 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 980 u64 prev_irq_time; 981 #endif 982 #ifdef CONFIG_PARAVIRT 983 u64 prev_steal_time; 984 #endif 985 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 986 u64 prev_steal_time_rq; 987 #endif 988 989 /* calc_load related fields */ 990 unsigned long calc_load_update; 991 long calc_load_active; 992 993 #ifdef CONFIG_SCHED_HRTICK 994 #ifdef CONFIG_SMP 995 call_single_data_t hrtick_csd; 996 #endif 997 struct hrtimer hrtick_timer; 998 #endif 999 1000 #ifdef CONFIG_SCHEDSTATS 1001 /* latency stats */ 1002 struct sched_info rq_sched_info; 1003 unsigned long long rq_cpu_time; 1004 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 1005 1006 /* sys_sched_yield() stats */ 1007 unsigned int yld_count; 1008 1009 /* schedule() stats */ 1010 unsigned int sched_count; 1011 unsigned int sched_goidle; 1012 1013 /* try_to_wake_up() stats */ 1014 unsigned int ttwu_count; 1015 unsigned int ttwu_local; 1016 #endif 1017 1018 #ifdef CONFIG_SMP 1019 struct llist_head wake_list; 1020 #endif 1021 1022 #ifdef CONFIG_CPU_IDLE 1023 /* Must be inspected within a rcu lock section */ 1024 struct cpuidle_state *idle_state; 1025 #endif 1026 }; 1027 1028 #ifdef CONFIG_FAIR_GROUP_SCHED 1029 1030 /* CPU runqueue to which this cfs_rq is attached */ 1031 static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 1032 { 1033 return cfs_rq->rq; 1034 } 1035 1036 #else 1037 1038 static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 1039 { 1040 return container_of(cfs_rq, struct rq, cfs); 1041 } 1042 #endif 1043 1044 static inline int cpu_of(struct rq *rq) 1045 { 1046 #ifdef CONFIG_SMP 1047 return rq->cpu; 1048 #else 1049 return 0; 1050 #endif 1051 } 1052 1053 1054 #ifdef CONFIG_SCHED_SMT 1055 extern void __update_idle_core(struct rq *rq); 1056 1057 static inline void update_idle_core(struct rq *rq) 1058 { 1059 if (static_branch_unlikely(&sched_smt_present)) 1060 __update_idle_core(rq); 1061 } 1062 1063 #else 1064 static inline void update_idle_core(struct rq *rq) { } 1065 #endif 1066 1067 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 1068 1069 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 1070 #define this_rq() this_cpu_ptr(&runqueues) 1071 #define task_rq(p) cpu_rq(task_cpu(p)) 1072 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 1073 #define raw_rq() raw_cpu_ptr(&runqueues) 1074 1075 extern void update_rq_clock(struct rq *rq); 1076 1077 static inline u64 __rq_clock_broken(struct rq *rq) 1078 { 1079 return READ_ONCE(rq->clock); 1080 } 1081 1082 /* 1083 * rq::clock_update_flags bits 1084 * 1085 * %RQCF_REQ_SKIP - will request skipping of clock update on the next 1086 * call to __schedule(). This is an optimisation to avoid 1087 * neighbouring rq clock updates. 1088 * 1089 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is 1090 * in effect and calls to update_rq_clock() are being ignored. 1091 * 1092 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been 1093 * made to update_rq_clock() since the last time rq::lock was pinned. 1094 * 1095 * If inside of __schedule(), clock_update_flags will have been 1096 * shifted left (a left shift is a cheap operation for the fast path 1097 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use, 1098 * 1099 * if (rq-clock_update_flags >= RQCF_UPDATED) 1100 * 1101 * to check if %RQCF_UPADTED is set. It'll never be shifted more than 1102 * one position though, because the next rq_unpin_lock() will shift it 1103 * back. 1104 */ 1105 #define RQCF_REQ_SKIP 0x01 1106 #define RQCF_ACT_SKIP 0x02 1107 #define RQCF_UPDATED 0x04 1108 1109 static inline void assert_clock_updated(struct rq *rq) 1110 { 1111 /* 1112 * The only reason for not seeing a clock update since the 1113 * last rq_pin_lock() is if we're currently skipping updates. 1114 */ 1115 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP); 1116 } 1117 1118 static inline u64 rq_clock(struct rq *rq) 1119 { 1120 lockdep_assert_held(&rq->lock); 1121 assert_clock_updated(rq); 1122 1123 return rq->clock; 1124 } 1125 1126 static inline u64 rq_clock_task(struct rq *rq) 1127 { 1128 lockdep_assert_held(&rq->lock); 1129 assert_clock_updated(rq); 1130 1131 return rq->clock_task; 1132 } 1133 1134 /** 1135 * By default the decay is the default pelt decay period. 1136 * The decay shift can change the decay period in 1137 * multiples of 32. 1138 * Decay shift Decay period(ms) 1139 * 0 32 1140 * 1 64 1141 * 2 128 1142 * 3 256 1143 * 4 512 1144 */ 1145 extern int sched_thermal_decay_shift; 1146 1147 static inline u64 rq_clock_thermal(struct rq *rq) 1148 { 1149 return rq_clock_task(rq) >> sched_thermal_decay_shift; 1150 } 1151 1152 static inline void rq_clock_skip_update(struct rq *rq) 1153 { 1154 lockdep_assert_held(&rq->lock); 1155 rq->clock_update_flags |= RQCF_REQ_SKIP; 1156 } 1157 1158 /* 1159 * See rt task throttling, which is the only time a skip 1160 * request is cancelled. 1161 */ 1162 static inline void rq_clock_cancel_skipupdate(struct rq *rq) 1163 { 1164 lockdep_assert_held(&rq->lock); 1165 rq->clock_update_flags &= ~RQCF_REQ_SKIP; 1166 } 1167 1168 struct rq_flags { 1169 unsigned long flags; 1170 struct pin_cookie cookie; 1171 #ifdef CONFIG_SCHED_DEBUG 1172 /* 1173 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the 1174 * current pin context is stashed here in case it needs to be 1175 * restored in rq_repin_lock(). 1176 */ 1177 unsigned int clock_update_flags; 1178 #endif 1179 }; 1180 1181 static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf) 1182 { 1183 rf->cookie = lockdep_pin_lock(&rq->lock); 1184 1185 #ifdef CONFIG_SCHED_DEBUG 1186 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP); 1187 rf->clock_update_flags = 0; 1188 #endif 1189 } 1190 1191 static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf) 1192 { 1193 #ifdef CONFIG_SCHED_DEBUG 1194 if (rq->clock_update_flags > RQCF_ACT_SKIP) 1195 rf->clock_update_flags = RQCF_UPDATED; 1196 #endif 1197 1198 lockdep_unpin_lock(&rq->lock, rf->cookie); 1199 } 1200 1201 static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf) 1202 { 1203 lockdep_repin_lock(&rq->lock, rf->cookie); 1204 1205 #ifdef CONFIG_SCHED_DEBUG 1206 /* 1207 * Restore the value we stashed in @rf for this pin context. 1208 */ 1209 rq->clock_update_flags |= rf->clock_update_flags; 1210 #endif 1211 } 1212 1213 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1214 __acquires(rq->lock); 1215 1216 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1217 __acquires(p->pi_lock) 1218 __acquires(rq->lock); 1219 1220 static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf) 1221 __releases(rq->lock) 1222 { 1223 rq_unpin_lock(rq, rf); 1224 raw_spin_unlock(&rq->lock); 1225 } 1226 1227 static inline void 1228 task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf) 1229 __releases(rq->lock) 1230 __releases(p->pi_lock) 1231 { 1232 rq_unpin_lock(rq, rf); 1233 raw_spin_unlock(&rq->lock); 1234 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); 1235 } 1236 1237 static inline void 1238 rq_lock_irqsave(struct rq *rq, struct rq_flags *rf) 1239 __acquires(rq->lock) 1240 { 1241 raw_spin_lock_irqsave(&rq->lock, rf->flags); 1242 rq_pin_lock(rq, rf); 1243 } 1244 1245 static inline void 1246 rq_lock_irq(struct rq *rq, struct rq_flags *rf) 1247 __acquires(rq->lock) 1248 { 1249 raw_spin_lock_irq(&rq->lock); 1250 rq_pin_lock(rq, rf); 1251 } 1252 1253 static inline void 1254 rq_lock(struct rq *rq, struct rq_flags *rf) 1255 __acquires(rq->lock) 1256 { 1257 raw_spin_lock(&rq->lock); 1258 rq_pin_lock(rq, rf); 1259 } 1260 1261 static inline void 1262 rq_relock(struct rq *rq, struct rq_flags *rf) 1263 __acquires(rq->lock) 1264 { 1265 raw_spin_lock(&rq->lock); 1266 rq_repin_lock(rq, rf); 1267 } 1268 1269 static inline void 1270 rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf) 1271 __releases(rq->lock) 1272 { 1273 rq_unpin_lock(rq, rf); 1274 raw_spin_unlock_irqrestore(&rq->lock, rf->flags); 1275 } 1276 1277 static inline void 1278 rq_unlock_irq(struct rq *rq, struct rq_flags *rf) 1279 __releases(rq->lock) 1280 { 1281 rq_unpin_lock(rq, rf); 1282 raw_spin_unlock_irq(&rq->lock); 1283 } 1284 1285 static inline void 1286 rq_unlock(struct rq *rq, struct rq_flags *rf) 1287 __releases(rq->lock) 1288 { 1289 rq_unpin_lock(rq, rf); 1290 raw_spin_unlock(&rq->lock); 1291 } 1292 1293 static inline struct rq * 1294 this_rq_lock_irq(struct rq_flags *rf) 1295 __acquires(rq->lock) 1296 { 1297 struct rq *rq; 1298 1299 local_irq_disable(); 1300 rq = this_rq(); 1301 rq_lock(rq, rf); 1302 return rq; 1303 } 1304 1305 #ifdef CONFIG_NUMA 1306 enum numa_topology_type { 1307 NUMA_DIRECT, 1308 NUMA_GLUELESS_MESH, 1309 NUMA_BACKPLANE, 1310 }; 1311 extern enum numa_topology_type sched_numa_topology_type; 1312 extern int sched_max_numa_distance; 1313 extern bool find_numa_distance(int distance); 1314 extern void sched_init_numa(void); 1315 extern void sched_domains_numa_masks_set(unsigned int cpu); 1316 extern void sched_domains_numa_masks_clear(unsigned int cpu); 1317 extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu); 1318 #else 1319 static inline void sched_init_numa(void) { } 1320 static inline void sched_domains_numa_masks_set(unsigned int cpu) { } 1321 static inline void sched_domains_numa_masks_clear(unsigned int cpu) { } 1322 static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu) 1323 { 1324 return nr_cpu_ids; 1325 } 1326 #endif 1327 1328 #ifdef CONFIG_NUMA_BALANCING 1329 /* The regions in numa_faults array from task_struct */ 1330 enum numa_faults_stats { 1331 NUMA_MEM = 0, 1332 NUMA_CPU, 1333 NUMA_MEMBUF, 1334 NUMA_CPUBUF 1335 }; 1336 extern void sched_setnuma(struct task_struct *p, int node); 1337 extern int migrate_task_to(struct task_struct *p, int cpu); 1338 extern int migrate_swap(struct task_struct *p, struct task_struct *t, 1339 int cpu, int scpu); 1340 extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p); 1341 #else 1342 static inline void 1343 init_numa_balancing(unsigned long clone_flags, struct task_struct *p) 1344 { 1345 } 1346 #endif /* CONFIG_NUMA_BALANCING */ 1347 1348 #ifdef CONFIG_SMP 1349 1350 static inline void 1351 queue_balance_callback(struct rq *rq, 1352 struct callback_head *head, 1353 void (*func)(struct rq *rq)) 1354 { 1355 lockdep_assert_held(&rq->lock); 1356 1357 if (unlikely(head->next)) 1358 return; 1359 1360 head->func = (void (*)(struct callback_head *))func; 1361 head->next = rq->balance_callback; 1362 rq->balance_callback = head; 1363 } 1364 1365 extern void sched_ttwu_pending(void); 1366 1367 #define rcu_dereference_check_sched_domain(p) \ 1368 rcu_dereference_check((p), \ 1369 lockdep_is_held(&sched_domains_mutex)) 1370 1371 /* 1372 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 1373 * See destroy_sched_domains: call_rcu for details. 1374 * 1375 * The domain tree of any CPU may only be accessed from within 1376 * preempt-disabled sections. 1377 */ 1378 #define for_each_domain(cpu, __sd) \ 1379 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 1380 __sd; __sd = __sd->parent) 1381 1382 /** 1383 * highest_flag_domain - Return highest sched_domain containing flag. 1384 * @cpu: The CPU whose highest level of sched domain is to 1385 * be returned. 1386 * @flag: The flag to check for the highest sched_domain 1387 * for the given CPU. 1388 * 1389 * Returns the highest sched_domain of a CPU which contains the given flag. 1390 */ 1391 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 1392 { 1393 struct sched_domain *sd, *hsd = NULL; 1394 1395 for_each_domain(cpu, sd) { 1396 if (!(sd->flags & flag)) 1397 break; 1398 hsd = sd; 1399 } 1400 1401 return hsd; 1402 } 1403 1404 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) 1405 { 1406 struct sched_domain *sd; 1407 1408 for_each_domain(cpu, sd) { 1409 if (sd->flags & flag) 1410 break; 1411 } 1412 1413 return sd; 1414 } 1415 1416 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc); 1417 DECLARE_PER_CPU(int, sd_llc_size); 1418 DECLARE_PER_CPU(int, sd_llc_id); 1419 DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared); 1420 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa); 1421 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing); 1422 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity); 1423 extern struct static_key_false sched_asym_cpucapacity; 1424 1425 struct sched_group_capacity { 1426 atomic_t ref; 1427 /* 1428 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity 1429 * for a single CPU. 1430 */ 1431 unsigned long capacity; 1432 unsigned long min_capacity; /* Min per-CPU capacity in group */ 1433 unsigned long max_capacity; /* Max per-CPU capacity in group */ 1434 unsigned long next_update; 1435 int imbalance; /* XXX unrelated to capacity but shared group state */ 1436 1437 #ifdef CONFIG_SCHED_DEBUG 1438 int id; 1439 #endif 1440 1441 unsigned long cpumask[0]; /* Balance mask */ 1442 }; 1443 1444 struct sched_group { 1445 struct sched_group *next; /* Must be a circular list */ 1446 atomic_t ref; 1447 1448 unsigned int group_weight; 1449 struct sched_group_capacity *sgc; 1450 int asym_prefer_cpu; /* CPU of highest priority in group */ 1451 1452 /* 1453 * The CPUs this group covers. 1454 * 1455 * NOTE: this field is variable length. (Allocated dynamically 1456 * by attaching extra space to the end of the structure, 1457 * depending on how many CPUs the kernel has booted up with) 1458 */ 1459 unsigned long cpumask[0]; 1460 }; 1461 1462 static inline struct cpumask *sched_group_span(struct sched_group *sg) 1463 { 1464 return to_cpumask(sg->cpumask); 1465 } 1466 1467 /* 1468 * See build_balance_mask(). 1469 */ 1470 static inline struct cpumask *group_balance_mask(struct sched_group *sg) 1471 { 1472 return to_cpumask(sg->sgc->cpumask); 1473 } 1474 1475 /** 1476 * group_first_cpu - Returns the first CPU in the cpumask of a sched_group. 1477 * @group: The group whose first CPU is to be returned. 1478 */ 1479 static inline unsigned int group_first_cpu(struct sched_group *group) 1480 { 1481 return cpumask_first(sched_group_span(group)); 1482 } 1483 1484 extern int group_balance_cpu(struct sched_group *sg); 1485 1486 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 1487 void register_sched_domain_sysctl(void); 1488 void dirty_sched_domain_sysctl(int cpu); 1489 void unregister_sched_domain_sysctl(void); 1490 #else 1491 static inline void register_sched_domain_sysctl(void) 1492 { 1493 } 1494 static inline void dirty_sched_domain_sysctl(int cpu) 1495 { 1496 } 1497 static inline void unregister_sched_domain_sysctl(void) 1498 { 1499 } 1500 #endif 1501 1502 extern int newidle_balance(struct rq *this_rq, struct rq_flags *rf); 1503 1504 #else 1505 1506 static inline void sched_ttwu_pending(void) { } 1507 1508 static inline int newidle_balance(struct rq *this_rq, struct rq_flags *rf) { return 0; } 1509 1510 #endif /* CONFIG_SMP */ 1511 1512 #include "stats.h" 1513 #include "autogroup.h" 1514 1515 #ifdef CONFIG_CGROUP_SCHED 1516 1517 /* 1518 * Return the group to which this tasks belongs. 1519 * 1520 * We cannot use task_css() and friends because the cgroup subsystem 1521 * changes that value before the cgroup_subsys::attach() method is called, 1522 * therefore we cannot pin it and might observe the wrong value. 1523 * 1524 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup 1525 * core changes this before calling sched_move_task(). 1526 * 1527 * Instead we use a 'copy' which is updated from sched_move_task() while 1528 * holding both task_struct::pi_lock and rq::lock. 1529 */ 1530 static inline struct task_group *task_group(struct task_struct *p) 1531 { 1532 return p->sched_task_group; 1533 } 1534 1535 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 1536 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 1537 { 1538 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 1539 struct task_group *tg = task_group(p); 1540 #endif 1541 1542 #ifdef CONFIG_FAIR_GROUP_SCHED 1543 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]); 1544 p->se.cfs_rq = tg->cfs_rq[cpu]; 1545 p->se.parent = tg->se[cpu]; 1546 #endif 1547 1548 #ifdef CONFIG_RT_GROUP_SCHED 1549 p->rt.rt_rq = tg->rt_rq[cpu]; 1550 p->rt.parent = tg->rt_se[cpu]; 1551 #endif 1552 } 1553 1554 #else /* CONFIG_CGROUP_SCHED */ 1555 1556 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 1557 static inline struct task_group *task_group(struct task_struct *p) 1558 { 1559 return NULL; 1560 } 1561 1562 #endif /* CONFIG_CGROUP_SCHED */ 1563 1564 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 1565 { 1566 set_task_rq(p, cpu); 1567 #ifdef CONFIG_SMP 1568 /* 1569 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 1570 * successfully executed on another CPU. We must ensure that updates of 1571 * per-task data have been completed by this moment. 1572 */ 1573 smp_wmb(); 1574 #ifdef CONFIG_THREAD_INFO_IN_TASK 1575 WRITE_ONCE(p->cpu, cpu); 1576 #else 1577 WRITE_ONCE(task_thread_info(p)->cpu, cpu); 1578 #endif 1579 p->wake_cpu = cpu; 1580 #endif 1581 } 1582 1583 /* 1584 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 1585 */ 1586 #ifdef CONFIG_SCHED_DEBUG 1587 # include <linux/static_key.h> 1588 # define const_debug __read_mostly 1589 #else 1590 # define const_debug const 1591 #endif 1592 1593 #define SCHED_FEAT(name, enabled) \ 1594 __SCHED_FEAT_##name , 1595 1596 enum { 1597 #include "features.h" 1598 __SCHED_FEAT_NR, 1599 }; 1600 1601 #undef SCHED_FEAT 1602 1603 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL) 1604 1605 /* 1606 * To support run-time toggling of sched features, all the translation units 1607 * (but core.c) reference the sysctl_sched_features defined in core.c. 1608 */ 1609 extern const_debug unsigned int sysctl_sched_features; 1610 1611 #define SCHED_FEAT(name, enabled) \ 1612 static __always_inline bool static_branch_##name(struct static_key *key) \ 1613 { \ 1614 return static_key_##enabled(key); \ 1615 } 1616 1617 #include "features.h" 1618 #undef SCHED_FEAT 1619 1620 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 1621 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 1622 1623 #else /* !(SCHED_DEBUG && CONFIG_JUMP_LABEL) */ 1624 1625 /* 1626 * Each translation unit has its own copy of sysctl_sched_features to allow 1627 * constants propagation at compile time and compiler optimization based on 1628 * features default. 1629 */ 1630 #define SCHED_FEAT(name, enabled) \ 1631 (1UL << __SCHED_FEAT_##name) * enabled | 1632 static const_debug __maybe_unused unsigned int sysctl_sched_features = 1633 #include "features.h" 1634 0; 1635 #undef SCHED_FEAT 1636 1637 #define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 1638 1639 #endif /* SCHED_DEBUG && CONFIG_JUMP_LABEL */ 1640 1641 extern struct static_key_false sched_numa_balancing; 1642 extern struct static_key_false sched_schedstats; 1643 1644 static inline u64 global_rt_period(void) 1645 { 1646 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 1647 } 1648 1649 static inline u64 global_rt_runtime(void) 1650 { 1651 if (sysctl_sched_rt_runtime < 0) 1652 return RUNTIME_INF; 1653 1654 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 1655 } 1656 1657 static inline int task_current(struct rq *rq, struct task_struct *p) 1658 { 1659 return rq->curr == p; 1660 } 1661 1662 static inline int task_running(struct rq *rq, struct task_struct *p) 1663 { 1664 #ifdef CONFIG_SMP 1665 return p->on_cpu; 1666 #else 1667 return task_current(rq, p); 1668 #endif 1669 } 1670 1671 static inline int task_on_rq_queued(struct task_struct *p) 1672 { 1673 return p->on_rq == TASK_ON_RQ_QUEUED; 1674 } 1675 1676 static inline int task_on_rq_migrating(struct task_struct *p) 1677 { 1678 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING; 1679 } 1680 1681 /* 1682 * wake flags 1683 */ 1684 #define WF_SYNC 0x01 /* Waker goes to sleep after wakeup */ 1685 #define WF_FORK 0x02 /* Child wakeup after fork */ 1686 #define WF_MIGRATED 0x4 /* Internal use, task got migrated */ 1687 1688 /* 1689 * To aid in avoiding the subversion of "niceness" due to uneven distribution 1690 * of tasks with abnormal "nice" values across CPUs the contribution that 1691 * each task makes to its run queue's load is weighted according to its 1692 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 1693 * scaled version of the new time slice allocation that they receive on time 1694 * slice expiry etc. 1695 */ 1696 1697 #define WEIGHT_IDLEPRIO 3 1698 #define WMULT_IDLEPRIO 1431655765 1699 1700 extern const int sched_prio_to_weight[40]; 1701 extern const u32 sched_prio_to_wmult[40]; 1702 1703 /* 1704 * {de,en}queue flags: 1705 * 1706 * DEQUEUE_SLEEP - task is no longer runnable 1707 * ENQUEUE_WAKEUP - task just became runnable 1708 * 1709 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks 1710 * are in a known state which allows modification. Such pairs 1711 * should preserve as much state as possible. 1712 * 1713 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location 1714 * in the runqueue. 1715 * 1716 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified) 1717 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline) 1718 * ENQUEUE_MIGRATED - the task was migrated during wakeup 1719 * 1720 */ 1721 1722 #define DEQUEUE_SLEEP 0x01 1723 #define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */ 1724 #define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */ 1725 #define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */ 1726 1727 #define ENQUEUE_WAKEUP 0x01 1728 #define ENQUEUE_RESTORE 0x02 1729 #define ENQUEUE_MOVE 0x04 1730 #define ENQUEUE_NOCLOCK 0x08 1731 1732 #define ENQUEUE_HEAD 0x10 1733 #define ENQUEUE_REPLENISH 0x20 1734 #ifdef CONFIG_SMP 1735 #define ENQUEUE_MIGRATED 0x40 1736 #else 1737 #define ENQUEUE_MIGRATED 0x00 1738 #endif 1739 1740 #define RETRY_TASK ((void *)-1UL) 1741 1742 struct sched_class { 1743 const struct sched_class *next; 1744 1745 #ifdef CONFIG_UCLAMP_TASK 1746 int uclamp_enabled; 1747 #endif 1748 1749 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags); 1750 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags); 1751 void (*yield_task) (struct rq *rq); 1752 bool (*yield_to_task)(struct rq *rq, struct task_struct *p, bool preempt); 1753 1754 void (*check_preempt_curr)(struct rq *rq, struct task_struct *p, int flags); 1755 1756 struct task_struct *(*pick_next_task)(struct rq *rq); 1757 1758 void (*put_prev_task)(struct rq *rq, struct task_struct *p); 1759 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first); 1760 1761 #ifdef CONFIG_SMP 1762 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf); 1763 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags); 1764 void (*migrate_task_rq)(struct task_struct *p, int new_cpu); 1765 1766 void (*task_woken)(struct rq *this_rq, struct task_struct *task); 1767 1768 void (*set_cpus_allowed)(struct task_struct *p, 1769 const struct cpumask *newmask); 1770 1771 void (*rq_online)(struct rq *rq); 1772 void (*rq_offline)(struct rq *rq); 1773 #endif 1774 1775 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued); 1776 void (*task_fork)(struct task_struct *p); 1777 void (*task_dead)(struct task_struct *p); 1778 1779 /* 1780 * The switched_from() call is allowed to drop rq->lock, therefore we 1781 * cannot assume the switched_from/switched_to pair is serliazed by 1782 * rq->lock. They are however serialized by p->pi_lock. 1783 */ 1784 void (*switched_from)(struct rq *this_rq, struct task_struct *task); 1785 void (*switched_to) (struct rq *this_rq, struct task_struct *task); 1786 void (*prio_changed) (struct rq *this_rq, struct task_struct *task, 1787 int oldprio); 1788 1789 unsigned int (*get_rr_interval)(struct rq *rq, 1790 struct task_struct *task); 1791 1792 void (*update_curr)(struct rq *rq); 1793 1794 #define TASK_SET_GROUP 0 1795 #define TASK_MOVE_GROUP 1 1796 1797 #ifdef CONFIG_FAIR_GROUP_SCHED 1798 void (*task_change_group)(struct task_struct *p, int type); 1799 #endif 1800 }; 1801 1802 static inline void put_prev_task(struct rq *rq, struct task_struct *prev) 1803 { 1804 WARN_ON_ONCE(rq->curr != prev); 1805 prev->sched_class->put_prev_task(rq, prev); 1806 } 1807 1808 static inline void set_next_task(struct rq *rq, struct task_struct *next) 1809 { 1810 WARN_ON_ONCE(rq->curr != next); 1811 next->sched_class->set_next_task(rq, next, false); 1812 } 1813 1814 #ifdef CONFIG_SMP 1815 #define sched_class_highest (&stop_sched_class) 1816 #else 1817 #define sched_class_highest (&dl_sched_class) 1818 #endif 1819 1820 #define for_class_range(class, _from, _to) \ 1821 for (class = (_from); class != (_to); class = class->next) 1822 1823 #define for_each_class(class) \ 1824 for_class_range(class, sched_class_highest, NULL) 1825 1826 extern const struct sched_class stop_sched_class; 1827 extern const struct sched_class dl_sched_class; 1828 extern const struct sched_class rt_sched_class; 1829 extern const struct sched_class fair_sched_class; 1830 extern const struct sched_class idle_sched_class; 1831 1832 static inline bool sched_stop_runnable(struct rq *rq) 1833 { 1834 return rq->stop && task_on_rq_queued(rq->stop); 1835 } 1836 1837 static inline bool sched_dl_runnable(struct rq *rq) 1838 { 1839 return rq->dl.dl_nr_running > 0; 1840 } 1841 1842 static inline bool sched_rt_runnable(struct rq *rq) 1843 { 1844 return rq->rt.rt_queued > 0; 1845 } 1846 1847 static inline bool sched_fair_runnable(struct rq *rq) 1848 { 1849 return rq->cfs.nr_running > 0; 1850 } 1851 1852 extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf); 1853 extern struct task_struct *pick_next_task_idle(struct rq *rq); 1854 1855 #ifdef CONFIG_SMP 1856 1857 extern void update_group_capacity(struct sched_domain *sd, int cpu); 1858 1859 extern void trigger_load_balance(struct rq *rq); 1860 1861 extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask); 1862 1863 #endif 1864 1865 #ifdef CONFIG_CPU_IDLE 1866 static inline void idle_set_state(struct rq *rq, 1867 struct cpuidle_state *idle_state) 1868 { 1869 rq->idle_state = idle_state; 1870 } 1871 1872 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1873 { 1874 SCHED_WARN_ON(!rcu_read_lock_held()); 1875 1876 return rq->idle_state; 1877 } 1878 #else 1879 static inline void idle_set_state(struct rq *rq, 1880 struct cpuidle_state *idle_state) 1881 { 1882 } 1883 1884 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1885 { 1886 return NULL; 1887 } 1888 #endif 1889 1890 extern void schedule_idle(void); 1891 1892 extern void sysrq_sched_debug_show(void); 1893 extern void sched_init_granularity(void); 1894 extern void update_max_interval(void); 1895 1896 extern void init_sched_dl_class(void); 1897 extern void init_sched_rt_class(void); 1898 extern void init_sched_fair_class(void); 1899 1900 extern void reweight_task(struct task_struct *p, int prio); 1901 1902 extern void resched_curr(struct rq *rq); 1903 extern void resched_cpu(int cpu); 1904 1905 extern struct rt_bandwidth def_rt_bandwidth; 1906 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 1907 1908 extern struct dl_bandwidth def_dl_bandwidth; 1909 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime); 1910 extern void init_dl_task_timer(struct sched_dl_entity *dl_se); 1911 extern void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se); 1912 1913 #define BW_SHIFT 20 1914 #define BW_UNIT (1 << BW_SHIFT) 1915 #define RATIO_SHIFT 8 1916 unsigned long to_ratio(u64 period, u64 runtime); 1917 1918 extern void init_entity_runnable_average(struct sched_entity *se); 1919 extern void post_init_entity_util_avg(struct task_struct *p); 1920 1921 #ifdef CONFIG_NO_HZ_FULL 1922 extern bool sched_can_stop_tick(struct rq *rq); 1923 extern int __init sched_tick_offload_init(void); 1924 1925 /* 1926 * Tick may be needed by tasks in the runqueue depending on their policy and 1927 * requirements. If tick is needed, lets send the target an IPI to kick it out of 1928 * nohz mode if necessary. 1929 */ 1930 static inline void sched_update_tick_dependency(struct rq *rq) 1931 { 1932 int cpu; 1933 1934 if (!tick_nohz_full_enabled()) 1935 return; 1936 1937 cpu = cpu_of(rq); 1938 1939 if (!tick_nohz_full_cpu(cpu)) 1940 return; 1941 1942 if (sched_can_stop_tick(rq)) 1943 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED); 1944 else 1945 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED); 1946 } 1947 #else 1948 static inline int sched_tick_offload_init(void) { return 0; } 1949 static inline void sched_update_tick_dependency(struct rq *rq) { } 1950 #endif 1951 1952 static inline void add_nr_running(struct rq *rq, unsigned count) 1953 { 1954 unsigned prev_nr = rq->nr_running; 1955 1956 rq->nr_running = prev_nr + count; 1957 1958 #ifdef CONFIG_SMP 1959 if (prev_nr < 2 && rq->nr_running >= 2) { 1960 if (!READ_ONCE(rq->rd->overload)) 1961 WRITE_ONCE(rq->rd->overload, 1); 1962 } 1963 #endif 1964 1965 sched_update_tick_dependency(rq); 1966 } 1967 1968 static inline void sub_nr_running(struct rq *rq, unsigned count) 1969 { 1970 rq->nr_running -= count; 1971 /* Check if we still need preemption */ 1972 sched_update_tick_dependency(rq); 1973 } 1974 1975 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 1976 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 1977 1978 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); 1979 1980 extern const_debug unsigned int sysctl_sched_nr_migrate; 1981 extern const_debug unsigned int sysctl_sched_migration_cost; 1982 1983 #ifdef CONFIG_SCHED_HRTICK 1984 1985 /* 1986 * Use hrtick when: 1987 * - enabled by features 1988 * - hrtimer is actually high res 1989 */ 1990 static inline int hrtick_enabled(struct rq *rq) 1991 { 1992 if (!sched_feat(HRTICK)) 1993 return 0; 1994 if (!cpu_active(cpu_of(rq))) 1995 return 0; 1996 return hrtimer_is_hres_active(&rq->hrtick_timer); 1997 } 1998 1999 void hrtick_start(struct rq *rq, u64 delay); 2000 2001 #else 2002 2003 static inline int hrtick_enabled(struct rq *rq) 2004 { 2005 return 0; 2006 } 2007 2008 #endif /* CONFIG_SCHED_HRTICK */ 2009 2010 #ifndef arch_scale_freq_tick 2011 static __always_inline 2012 void arch_scale_freq_tick(void) 2013 { 2014 } 2015 #endif 2016 2017 #ifndef arch_scale_freq_capacity 2018 static __always_inline 2019 unsigned long arch_scale_freq_capacity(int cpu) 2020 { 2021 return SCHED_CAPACITY_SCALE; 2022 } 2023 #endif 2024 2025 #ifdef CONFIG_SMP 2026 #ifdef CONFIG_PREEMPTION 2027 2028 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2); 2029 2030 /* 2031 * fair double_lock_balance: Safely acquires both rq->locks in a fair 2032 * way at the expense of forcing extra atomic operations in all 2033 * invocations. This assures that the double_lock is acquired using the 2034 * same underlying policy as the spinlock_t on this architecture, which 2035 * reduces latency compared to the unfair variant below. However, it 2036 * also adds more overhead and therefore may reduce throughput. 2037 */ 2038 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 2039 __releases(this_rq->lock) 2040 __acquires(busiest->lock) 2041 __acquires(this_rq->lock) 2042 { 2043 raw_spin_unlock(&this_rq->lock); 2044 double_rq_lock(this_rq, busiest); 2045 2046 return 1; 2047 } 2048 2049 #else 2050 /* 2051 * Unfair double_lock_balance: Optimizes throughput at the expense of 2052 * latency by eliminating extra atomic operations when the locks are 2053 * already in proper order on entry. This favors lower CPU-ids and will 2054 * grant the double lock to lower CPUs over higher ids under contention, 2055 * regardless of entry order into the function. 2056 */ 2057 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 2058 __releases(this_rq->lock) 2059 __acquires(busiest->lock) 2060 __acquires(this_rq->lock) 2061 { 2062 int ret = 0; 2063 2064 if (unlikely(!raw_spin_trylock(&busiest->lock))) { 2065 if (busiest < this_rq) { 2066 raw_spin_unlock(&this_rq->lock); 2067 raw_spin_lock(&busiest->lock); 2068 raw_spin_lock_nested(&this_rq->lock, 2069 SINGLE_DEPTH_NESTING); 2070 ret = 1; 2071 } else 2072 raw_spin_lock_nested(&busiest->lock, 2073 SINGLE_DEPTH_NESTING); 2074 } 2075 return ret; 2076 } 2077 2078 #endif /* CONFIG_PREEMPTION */ 2079 2080 /* 2081 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 2082 */ 2083 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 2084 { 2085 if (unlikely(!irqs_disabled())) { 2086 /* printk() doesn't work well under rq->lock */ 2087 raw_spin_unlock(&this_rq->lock); 2088 BUG_ON(1); 2089 } 2090 2091 return _double_lock_balance(this_rq, busiest); 2092 } 2093 2094 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 2095 __releases(busiest->lock) 2096 { 2097 raw_spin_unlock(&busiest->lock); 2098 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 2099 } 2100 2101 static inline void double_lock(spinlock_t *l1, spinlock_t *l2) 2102 { 2103 if (l1 > l2) 2104 swap(l1, l2); 2105 2106 spin_lock(l1); 2107 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 2108 } 2109 2110 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2) 2111 { 2112 if (l1 > l2) 2113 swap(l1, l2); 2114 2115 spin_lock_irq(l1); 2116 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 2117 } 2118 2119 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2) 2120 { 2121 if (l1 > l2) 2122 swap(l1, l2); 2123 2124 raw_spin_lock(l1); 2125 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 2126 } 2127 2128 /* 2129 * double_rq_lock - safely lock two runqueues 2130 * 2131 * Note this does not disable interrupts like task_rq_lock, 2132 * you need to do so manually before calling. 2133 */ 2134 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 2135 __acquires(rq1->lock) 2136 __acquires(rq2->lock) 2137 { 2138 BUG_ON(!irqs_disabled()); 2139 if (rq1 == rq2) { 2140 raw_spin_lock(&rq1->lock); 2141 __acquire(rq2->lock); /* Fake it out ;) */ 2142 } else { 2143 if (rq1 < rq2) { 2144 raw_spin_lock(&rq1->lock); 2145 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); 2146 } else { 2147 raw_spin_lock(&rq2->lock); 2148 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); 2149 } 2150 } 2151 } 2152 2153 /* 2154 * double_rq_unlock - safely unlock two runqueues 2155 * 2156 * Note this does not restore interrupts like task_rq_unlock, 2157 * you need to do so manually after calling. 2158 */ 2159 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 2160 __releases(rq1->lock) 2161 __releases(rq2->lock) 2162 { 2163 raw_spin_unlock(&rq1->lock); 2164 if (rq1 != rq2) 2165 raw_spin_unlock(&rq2->lock); 2166 else 2167 __release(rq2->lock); 2168 } 2169 2170 extern void set_rq_online (struct rq *rq); 2171 extern void set_rq_offline(struct rq *rq); 2172 extern bool sched_smp_initialized; 2173 2174 #else /* CONFIG_SMP */ 2175 2176 /* 2177 * double_rq_lock - safely lock two runqueues 2178 * 2179 * Note this does not disable interrupts like task_rq_lock, 2180 * you need to do so manually before calling. 2181 */ 2182 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 2183 __acquires(rq1->lock) 2184 __acquires(rq2->lock) 2185 { 2186 BUG_ON(!irqs_disabled()); 2187 BUG_ON(rq1 != rq2); 2188 raw_spin_lock(&rq1->lock); 2189 __acquire(rq2->lock); /* Fake it out ;) */ 2190 } 2191 2192 /* 2193 * double_rq_unlock - safely unlock two runqueues 2194 * 2195 * Note this does not restore interrupts like task_rq_unlock, 2196 * you need to do so manually after calling. 2197 */ 2198 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 2199 __releases(rq1->lock) 2200 __releases(rq2->lock) 2201 { 2202 BUG_ON(rq1 != rq2); 2203 raw_spin_unlock(&rq1->lock); 2204 __release(rq2->lock); 2205 } 2206 2207 #endif 2208 2209 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 2210 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 2211 2212 #ifdef CONFIG_SCHED_DEBUG 2213 extern bool sched_debug_enabled; 2214 2215 extern void print_cfs_stats(struct seq_file *m, int cpu); 2216 extern void print_rt_stats(struct seq_file *m, int cpu); 2217 extern void print_dl_stats(struct seq_file *m, int cpu); 2218 extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq); 2219 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq); 2220 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); 2221 #ifdef CONFIG_NUMA_BALANCING 2222 extern void 2223 show_numa_stats(struct task_struct *p, struct seq_file *m); 2224 extern void 2225 print_numa_stats(struct seq_file *m, int node, unsigned long tsf, 2226 unsigned long tpf, unsigned long gsf, unsigned long gpf); 2227 #endif /* CONFIG_NUMA_BALANCING */ 2228 #endif /* CONFIG_SCHED_DEBUG */ 2229 2230 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 2231 extern void init_rt_rq(struct rt_rq *rt_rq); 2232 extern void init_dl_rq(struct dl_rq *dl_rq); 2233 2234 extern void cfs_bandwidth_usage_inc(void); 2235 extern void cfs_bandwidth_usage_dec(void); 2236 2237 #ifdef CONFIG_NO_HZ_COMMON 2238 #define NOHZ_BALANCE_KICK_BIT 0 2239 #define NOHZ_STATS_KICK_BIT 1 2240 2241 #define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT) 2242 #define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT) 2243 2244 #define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK) 2245 2246 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 2247 2248 extern void nohz_balance_exit_idle(struct rq *rq); 2249 #else 2250 static inline void nohz_balance_exit_idle(struct rq *rq) { } 2251 #endif 2252 2253 2254 #ifdef CONFIG_SMP 2255 static inline 2256 void __dl_update(struct dl_bw *dl_b, s64 bw) 2257 { 2258 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw); 2259 int i; 2260 2261 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), 2262 "sched RCU must be held"); 2263 for_each_cpu_and(i, rd->span, cpu_active_mask) { 2264 struct rq *rq = cpu_rq(i); 2265 2266 rq->dl.extra_bw += bw; 2267 } 2268 } 2269 #else 2270 static inline 2271 void __dl_update(struct dl_bw *dl_b, s64 bw) 2272 { 2273 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw); 2274 2275 dl->extra_bw += bw; 2276 } 2277 #endif 2278 2279 2280 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 2281 struct irqtime { 2282 u64 total; 2283 u64 tick_delta; 2284 u64 irq_start_time; 2285 struct u64_stats_sync sync; 2286 }; 2287 2288 DECLARE_PER_CPU(struct irqtime, cpu_irqtime); 2289 2290 /* 2291 * Returns the irqtime minus the softirq time computed by ksoftirqd. 2292 * Otherwise ksoftirqd's sum_exec_runtime is substracted its own runtime 2293 * and never move forward. 2294 */ 2295 static inline u64 irq_time_read(int cpu) 2296 { 2297 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu); 2298 unsigned int seq; 2299 u64 total; 2300 2301 do { 2302 seq = __u64_stats_fetch_begin(&irqtime->sync); 2303 total = irqtime->total; 2304 } while (__u64_stats_fetch_retry(&irqtime->sync, seq)); 2305 2306 return total; 2307 } 2308 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 2309 2310 #ifdef CONFIG_CPU_FREQ 2311 DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data); 2312 2313 /** 2314 * cpufreq_update_util - Take a note about CPU utilization changes. 2315 * @rq: Runqueue to carry out the update for. 2316 * @flags: Update reason flags. 2317 * 2318 * This function is called by the scheduler on the CPU whose utilization is 2319 * being updated. 2320 * 2321 * It can only be called from RCU-sched read-side critical sections. 2322 * 2323 * The way cpufreq is currently arranged requires it to evaluate the CPU 2324 * performance state (frequency/voltage) on a regular basis to prevent it from 2325 * being stuck in a completely inadequate performance level for too long. 2326 * That is not guaranteed to happen if the updates are only triggered from CFS 2327 * and DL, though, because they may not be coming in if only RT tasks are 2328 * active all the time (or there are RT tasks only). 2329 * 2330 * As a workaround for that issue, this function is called periodically by the 2331 * RT sched class to trigger extra cpufreq updates to prevent it from stalling, 2332 * but that really is a band-aid. Going forward it should be replaced with 2333 * solutions targeted more specifically at RT tasks. 2334 */ 2335 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) 2336 { 2337 struct update_util_data *data; 2338 2339 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data, 2340 cpu_of(rq))); 2341 if (data) 2342 data->func(data, rq_clock(rq), flags); 2343 } 2344 #else 2345 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {} 2346 #endif /* CONFIG_CPU_FREQ */ 2347 2348 #ifdef CONFIG_UCLAMP_TASK 2349 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id); 2350 2351 static __always_inline 2352 unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util, 2353 struct task_struct *p) 2354 { 2355 unsigned long min_util = READ_ONCE(rq->uclamp[UCLAMP_MIN].value); 2356 unsigned long max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value); 2357 2358 if (p) { 2359 min_util = max(min_util, uclamp_eff_value(p, UCLAMP_MIN)); 2360 max_util = max(max_util, uclamp_eff_value(p, UCLAMP_MAX)); 2361 } 2362 2363 /* 2364 * Since CPU's {min,max}_util clamps are MAX aggregated considering 2365 * RUNNABLE tasks with _different_ clamps, we can end up with an 2366 * inversion. Fix it now when the clamps are applied. 2367 */ 2368 if (unlikely(min_util >= max_util)) 2369 return min_util; 2370 2371 return clamp(util, min_util, max_util); 2372 } 2373 #else /* CONFIG_UCLAMP_TASK */ 2374 static inline 2375 unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util, 2376 struct task_struct *p) 2377 { 2378 return util; 2379 } 2380 #endif /* CONFIG_UCLAMP_TASK */ 2381 2382 #ifdef arch_scale_freq_capacity 2383 # ifndef arch_scale_freq_invariant 2384 # define arch_scale_freq_invariant() true 2385 # endif 2386 #else 2387 # define arch_scale_freq_invariant() false 2388 #endif 2389 2390 #ifdef CONFIG_SMP 2391 static inline unsigned long capacity_orig_of(int cpu) 2392 { 2393 return cpu_rq(cpu)->cpu_capacity_orig; 2394 } 2395 #endif 2396 2397 /** 2398 * enum schedutil_type - CPU utilization type 2399 * @FREQUENCY_UTIL: Utilization used to select frequency 2400 * @ENERGY_UTIL: Utilization used during energy calculation 2401 * 2402 * The utilization signals of all scheduling classes (CFS/RT/DL) and IRQ time 2403 * need to be aggregated differently depending on the usage made of them. This 2404 * enum is used within schedutil_freq_util() to differentiate the types of 2405 * utilization expected by the callers, and adjust the aggregation accordingly. 2406 */ 2407 enum schedutil_type { 2408 FREQUENCY_UTIL, 2409 ENERGY_UTIL, 2410 }; 2411 2412 #ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL 2413 2414 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs, 2415 unsigned long max, enum schedutil_type type, 2416 struct task_struct *p); 2417 2418 static inline unsigned long cpu_bw_dl(struct rq *rq) 2419 { 2420 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT; 2421 } 2422 2423 static inline unsigned long cpu_util_dl(struct rq *rq) 2424 { 2425 return READ_ONCE(rq->avg_dl.util_avg); 2426 } 2427 2428 static inline unsigned long cpu_util_cfs(struct rq *rq) 2429 { 2430 unsigned long util = READ_ONCE(rq->cfs.avg.util_avg); 2431 2432 if (sched_feat(UTIL_EST)) { 2433 util = max_t(unsigned long, util, 2434 READ_ONCE(rq->cfs.avg.util_est.enqueued)); 2435 } 2436 2437 return util; 2438 } 2439 2440 static inline unsigned long cpu_util_rt(struct rq *rq) 2441 { 2442 return READ_ONCE(rq->avg_rt.util_avg); 2443 } 2444 #else /* CONFIG_CPU_FREQ_GOV_SCHEDUTIL */ 2445 static inline unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs, 2446 unsigned long max, enum schedutil_type type, 2447 struct task_struct *p) 2448 { 2449 return 0; 2450 } 2451 #endif /* CONFIG_CPU_FREQ_GOV_SCHEDUTIL */ 2452 2453 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 2454 static inline unsigned long cpu_util_irq(struct rq *rq) 2455 { 2456 return rq->avg_irq.util_avg; 2457 } 2458 2459 static inline 2460 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max) 2461 { 2462 util *= (max - irq); 2463 util /= max; 2464 2465 return util; 2466 2467 } 2468 #else 2469 static inline unsigned long cpu_util_irq(struct rq *rq) 2470 { 2471 return 0; 2472 } 2473 2474 static inline 2475 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max) 2476 { 2477 return util; 2478 } 2479 #endif 2480 2481 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL) 2482 2483 #define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus))) 2484 2485 DECLARE_STATIC_KEY_FALSE(sched_energy_present); 2486 2487 static inline bool sched_energy_enabled(void) 2488 { 2489 return static_branch_unlikely(&sched_energy_present); 2490 } 2491 2492 #else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */ 2493 2494 #define perf_domain_span(pd) NULL 2495 static inline bool sched_energy_enabled(void) { return false; } 2496 2497 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */ 2498 2499 #ifdef CONFIG_MEMBARRIER 2500 /* 2501 * The scheduler provides memory barriers required by membarrier between: 2502 * - prior user-space memory accesses and store to rq->membarrier_state, 2503 * - store to rq->membarrier_state and following user-space memory accesses. 2504 * In the same way it provides those guarantees around store to rq->curr. 2505 */ 2506 static inline void membarrier_switch_mm(struct rq *rq, 2507 struct mm_struct *prev_mm, 2508 struct mm_struct *next_mm) 2509 { 2510 int membarrier_state; 2511 2512 if (prev_mm == next_mm) 2513 return; 2514 2515 membarrier_state = atomic_read(&next_mm->membarrier_state); 2516 if (READ_ONCE(rq->membarrier_state) == membarrier_state) 2517 return; 2518 2519 WRITE_ONCE(rq->membarrier_state, membarrier_state); 2520 } 2521 #else 2522 static inline void membarrier_switch_mm(struct rq *rq, 2523 struct mm_struct *prev_mm, 2524 struct mm_struct *next_mm) 2525 { 2526 } 2527 #endif 2528 2529 #ifdef CONFIG_SMP 2530 static inline bool is_per_cpu_kthread(struct task_struct *p) 2531 { 2532 if (!(p->flags & PF_KTHREAD)) 2533 return false; 2534 2535 if (p->nr_cpus_allowed != 1) 2536 return false; 2537 2538 return true; 2539 } 2540 #endif 2541 2542 void swake_up_all_locked(struct swait_queue_head *q); 2543 void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait); 2544