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