1 2 #include <linux/sched.h> 3 #include <linux/sched/sysctl.h> 4 #include <linux/sched/rt.h> 5 #include <linux/u64_stats_sync.h> 6 #include <linux/sched/deadline.h> 7 #include <linux/binfmts.h> 8 #include <linux/mutex.h> 9 #include <linux/spinlock.h> 10 #include <linux/stop_machine.h> 11 #include <linux/irq_work.h> 12 #include <linux/tick.h> 13 #include <linux/slab.h> 14 15 #include "cpupri.h" 16 #include "cpudeadline.h" 17 #include "cpuacct.h" 18 19 #ifdef CONFIG_SCHED_DEBUG 20 #define SCHED_WARN_ON(x) WARN_ONCE(x, #x) 21 #else 22 #define SCHED_WARN_ON(x) ((void)(x)) 23 #endif 24 25 struct rq; 26 struct cpuidle_state; 27 28 /* task_struct::on_rq states: */ 29 #define TASK_ON_RQ_QUEUED 1 30 #define TASK_ON_RQ_MIGRATING 2 31 32 extern __read_mostly int scheduler_running; 33 34 extern unsigned long calc_load_update; 35 extern atomic_long_t calc_load_tasks; 36 37 extern void calc_global_load_tick(struct rq *this_rq); 38 extern long calc_load_fold_active(struct rq *this_rq, long adjust); 39 40 #ifdef CONFIG_SMP 41 extern void cpu_load_update_active(struct rq *this_rq); 42 #else 43 static inline void cpu_load_update_active(struct rq *this_rq) { } 44 #endif 45 46 /* 47 * Helpers for converting nanosecond timing to jiffy resolution 48 */ 49 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) 50 51 /* 52 * Increase resolution of nice-level calculations for 64-bit architectures. 53 * The extra resolution improves shares distribution and load balancing of 54 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup 55 * hierarchies, especially on larger systems. This is not a user-visible change 56 * and does not change the user-interface for setting shares/weights. 57 * 58 * We increase resolution only if we have enough bits to allow this increased 59 * resolution (i.e. 64bit). The costs for increasing resolution when 32bit are 60 * pretty high and the returns do not justify the increased costs. 61 * 62 * Really only required when CONFIG_FAIR_GROUP_SCHED is also set, but to 63 * increase coverage and consistency always enable it on 64bit platforms. 64 */ 65 #ifdef CONFIG_64BIT 66 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT) 67 # define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT) 68 # define scale_load_down(w) ((w) >> SCHED_FIXEDPOINT_SHIFT) 69 #else 70 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT) 71 # define scale_load(w) (w) 72 # define scale_load_down(w) (w) 73 #endif 74 75 /* 76 * Task weight (visible to users) and its load (invisible to users) have 77 * independent resolution, but they should be well calibrated. We use 78 * scale_load() and scale_load_down(w) to convert between them. The 79 * following must be true: 80 * 81 * scale_load(sched_prio_to_weight[USER_PRIO(NICE_TO_PRIO(0))]) == NICE_0_LOAD 82 * 83 */ 84 #define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT) 85 86 /* 87 * Single value that decides SCHED_DEADLINE internal math precision. 88 * 10 -> just above 1us 89 * 9 -> just above 0.5us 90 */ 91 #define DL_SCALE (10) 92 93 /* 94 * These are the 'tuning knobs' of the scheduler: 95 */ 96 97 /* 98 * single value that denotes runtime == period, ie unlimited time. 99 */ 100 #define RUNTIME_INF ((u64)~0ULL) 101 102 static inline int idle_policy(int policy) 103 { 104 return policy == SCHED_IDLE; 105 } 106 static inline int fair_policy(int policy) 107 { 108 return policy == SCHED_NORMAL || policy == SCHED_BATCH; 109 } 110 111 static inline int rt_policy(int policy) 112 { 113 return policy == SCHED_FIFO || policy == SCHED_RR; 114 } 115 116 static inline int dl_policy(int policy) 117 { 118 return policy == SCHED_DEADLINE; 119 } 120 static inline bool valid_policy(int policy) 121 { 122 return idle_policy(policy) || fair_policy(policy) || 123 rt_policy(policy) || dl_policy(policy); 124 } 125 126 static inline int task_has_rt_policy(struct task_struct *p) 127 { 128 return rt_policy(p->policy); 129 } 130 131 static inline int task_has_dl_policy(struct task_struct *p) 132 { 133 return dl_policy(p->policy); 134 } 135 136 /* 137 * Tells if entity @a should preempt entity @b. 138 */ 139 static inline bool 140 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b) 141 { 142 return dl_time_before(a->deadline, b->deadline); 143 } 144 145 /* 146 * This is the priority-queue data structure of the RT scheduling class: 147 */ 148 struct rt_prio_array { 149 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ 150 struct list_head queue[MAX_RT_PRIO]; 151 }; 152 153 struct rt_bandwidth { 154 /* nests inside the rq lock: */ 155 raw_spinlock_t rt_runtime_lock; 156 ktime_t rt_period; 157 u64 rt_runtime; 158 struct hrtimer rt_period_timer; 159 unsigned int rt_period_active; 160 }; 161 162 void __dl_clear_params(struct task_struct *p); 163 164 /* 165 * To keep the bandwidth of -deadline tasks and groups under control 166 * we need some place where: 167 * - store the maximum -deadline bandwidth of the system (the group); 168 * - cache the fraction of that bandwidth that is currently allocated. 169 * 170 * This is all done in the data structure below. It is similar to the 171 * one used for RT-throttling (rt_bandwidth), with the main difference 172 * that, since here we are only interested in admission control, we 173 * do not decrease any runtime while the group "executes", neither we 174 * need a timer to replenish it. 175 * 176 * With respect to SMP, the bandwidth is given on a per-CPU basis, 177 * meaning that: 178 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU; 179 * - dl_total_bw array contains, in the i-eth element, the currently 180 * allocated bandwidth on the i-eth CPU. 181 * Moreover, groups consume bandwidth on each CPU, while tasks only 182 * consume bandwidth on the CPU they're running on. 183 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw 184 * that will be shown the next time the proc or cgroup controls will 185 * be red. It on its turn can be changed by writing on its own 186 * control. 187 */ 188 struct dl_bandwidth { 189 raw_spinlock_t dl_runtime_lock; 190 u64 dl_runtime; 191 u64 dl_period; 192 }; 193 194 static inline int dl_bandwidth_enabled(void) 195 { 196 return sysctl_sched_rt_runtime >= 0; 197 } 198 199 extern struct dl_bw *dl_bw_of(int i); 200 201 struct dl_bw { 202 raw_spinlock_t lock; 203 u64 bw, total_bw; 204 }; 205 206 static inline 207 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 208 { 209 dl_b->total_bw -= tsk_bw; 210 } 211 212 static inline 213 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 214 { 215 dl_b->total_bw += tsk_bw; 216 } 217 218 static inline 219 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 220 { 221 return dl_b->bw != -1 && 222 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 223 } 224 225 extern struct mutex sched_domains_mutex; 226 227 #ifdef CONFIG_CGROUP_SCHED 228 229 #include <linux/cgroup.h> 230 231 struct cfs_rq; 232 struct rt_rq; 233 234 extern struct list_head task_groups; 235 236 struct cfs_bandwidth { 237 #ifdef CONFIG_CFS_BANDWIDTH 238 raw_spinlock_t lock; 239 ktime_t period; 240 u64 quota, runtime; 241 s64 hierarchical_quota; 242 u64 runtime_expires; 243 244 int idle, period_active; 245 struct hrtimer period_timer, slack_timer; 246 struct list_head throttled_cfs_rq; 247 248 /* statistics */ 249 int nr_periods, nr_throttled; 250 u64 throttled_time; 251 #endif 252 }; 253 254 /* task group related information */ 255 struct task_group { 256 struct cgroup_subsys_state css; 257 258 #ifdef CONFIG_FAIR_GROUP_SCHED 259 /* schedulable entities of this group on each cpu */ 260 struct sched_entity **se; 261 /* runqueue "owned" by this group on each cpu */ 262 struct cfs_rq **cfs_rq; 263 unsigned long shares; 264 265 #ifdef CONFIG_SMP 266 /* 267 * load_avg can be heavily contended at clock tick time, so put 268 * it in its own cacheline separated from the fields above which 269 * will also be accessed at each tick. 270 */ 271 atomic_long_t load_avg ____cacheline_aligned; 272 #endif 273 #endif 274 275 #ifdef CONFIG_RT_GROUP_SCHED 276 struct sched_rt_entity **rt_se; 277 struct rt_rq **rt_rq; 278 279 struct rt_bandwidth rt_bandwidth; 280 #endif 281 282 struct rcu_head rcu; 283 struct list_head list; 284 285 struct task_group *parent; 286 struct list_head siblings; 287 struct list_head children; 288 289 #ifdef CONFIG_SCHED_AUTOGROUP 290 struct autogroup *autogroup; 291 #endif 292 293 struct cfs_bandwidth cfs_bandwidth; 294 }; 295 296 #ifdef CONFIG_FAIR_GROUP_SCHED 297 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD 298 299 /* 300 * A weight of 0 or 1 can cause arithmetics problems. 301 * A weight of a cfs_rq is the sum of weights of which entities 302 * are queued on this cfs_rq, so a weight of a entity should not be 303 * too large, so as the shares value of a task group. 304 * (The default weight is 1024 - so there's no practical 305 * limitation from this.) 306 */ 307 #define MIN_SHARES (1UL << 1) 308 #define MAX_SHARES (1UL << 18) 309 #endif 310 311 typedef int (*tg_visitor)(struct task_group *, void *); 312 313 extern int walk_tg_tree_from(struct task_group *from, 314 tg_visitor down, tg_visitor up, void *data); 315 316 /* 317 * Iterate the full tree, calling @down when first entering a node and @up when 318 * leaving it for the final time. 319 * 320 * Caller must hold rcu_lock or sufficient equivalent. 321 */ 322 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) 323 { 324 return walk_tg_tree_from(&root_task_group, down, up, data); 325 } 326 327 extern int tg_nop(struct task_group *tg, void *data); 328 329 extern void free_fair_sched_group(struct task_group *tg); 330 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent); 331 extern void online_fair_sched_group(struct task_group *tg); 332 extern void unregister_fair_sched_group(struct task_group *tg); 333 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 334 struct sched_entity *se, int cpu, 335 struct sched_entity *parent); 336 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 337 338 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b); 339 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 340 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq); 341 342 extern void free_rt_sched_group(struct task_group *tg); 343 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent); 344 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 345 struct sched_rt_entity *rt_se, int cpu, 346 struct sched_rt_entity *parent); 347 348 extern struct task_group *sched_create_group(struct task_group *parent); 349 extern void sched_online_group(struct task_group *tg, 350 struct task_group *parent); 351 extern void sched_destroy_group(struct task_group *tg); 352 extern void sched_offline_group(struct task_group *tg); 353 354 extern void sched_move_task(struct task_struct *tsk); 355 356 #ifdef CONFIG_FAIR_GROUP_SCHED 357 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 358 359 #ifdef CONFIG_SMP 360 extern void set_task_rq_fair(struct sched_entity *se, 361 struct cfs_rq *prev, struct cfs_rq *next); 362 #else /* !CONFIG_SMP */ 363 static inline void set_task_rq_fair(struct sched_entity *se, 364 struct cfs_rq *prev, struct cfs_rq *next) { } 365 #endif /* CONFIG_SMP */ 366 #endif /* CONFIG_FAIR_GROUP_SCHED */ 367 368 #else /* CONFIG_CGROUP_SCHED */ 369 370 struct cfs_bandwidth { }; 371 372 #endif /* CONFIG_CGROUP_SCHED */ 373 374 /* CFS-related fields in a runqueue */ 375 struct cfs_rq { 376 struct load_weight load; 377 unsigned int nr_running, h_nr_running; 378 379 u64 exec_clock; 380 u64 min_vruntime; 381 #ifndef CONFIG_64BIT 382 u64 min_vruntime_copy; 383 #endif 384 385 struct rb_root tasks_timeline; 386 struct rb_node *rb_leftmost; 387 388 /* 389 * 'curr' points to currently running entity on this cfs_rq. 390 * It is set to NULL otherwise (i.e when none are currently running). 391 */ 392 struct sched_entity *curr, *next, *last, *skip; 393 394 #ifdef CONFIG_SCHED_DEBUG 395 unsigned int nr_spread_over; 396 #endif 397 398 #ifdef CONFIG_SMP 399 /* 400 * CFS load tracking 401 */ 402 struct sched_avg avg; 403 u64 runnable_load_sum; 404 unsigned long runnable_load_avg; 405 #ifdef CONFIG_FAIR_GROUP_SCHED 406 unsigned long tg_load_avg_contrib; 407 #endif 408 atomic_long_t removed_load_avg, removed_util_avg; 409 #ifndef CONFIG_64BIT 410 u64 load_last_update_time_copy; 411 #endif 412 413 #ifdef CONFIG_FAIR_GROUP_SCHED 414 /* 415 * h_load = weight * f(tg) 416 * 417 * Where f(tg) is the recursive weight fraction assigned to 418 * this group. 419 */ 420 unsigned long h_load; 421 u64 last_h_load_update; 422 struct sched_entity *h_load_next; 423 #endif /* CONFIG_FAIR_GROUP_SCHED */ 424 #endif /* CONFIG_SMP */ 425 426 #ifdef CONFIG_FAIR_GROUP_SCHED 427 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ 428 429 /* 430 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in 431 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities 432 * (like users, containers etc.) 433 * 434 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This 435 * list is used during load balance. 436 */ 437 int on_list; 438 struct list_head leaf_cfs_rq_list; 439 struct task_group *tg; /* group that "owns" this runqueue */ 440 441 #ifdef CONFIG_CFS_BANDWIDTH 442 int runtime_enabled; 443 u64 runtime_expires; 444 s64 runtime_remaining; 445 446 u64 throttled_clock, throttled_clock_task; 447 u64 throttled_clock_task_time; 448 int throttled, throttle_count; 449 struct list_head throttled_list; 450 #endif /* CONFIG_CFS_BANDWIDTH */ 451 #endif /* CONFIG_FAIR_GROUP_SCHED */ 452 }; 453 454 static inline int rt_bandwidth_enabled(void) 455 { 456 return sysctl_sched_rt_runtime >= 0; 457 } 458 459 /* RT IPI pull logic requires IRQ_WORK */ 460 #ifdef CONFIG_IRQ_WORK 461 # define HAVE_RT_PUSH_IPI 462 #endif 463 464 /* Real-Time classes' related field in a runqueue: */ 465 struct rt_rq { 466 struct rt_prio_array active; 467 unsigned int rt_nr_running; 468 unsigned int rr_nr_running; 469 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 470 struct { 471 int curr; /* highest queued rt task prio */ 472 #ifdef CONFIG_SMP 473 int next; /* next highest */ 474 #endif 475 } highest_prio; 476 #endif 477 #ifdef CONFIG_SMP 478 unsigned long rt_nr_migratory; 479 unsigned long rt_nr_total; 480 int overloaded; 481 struct plist_head pushable_tasks; 482 #ifdef HAVE_RT_PUSH_IPI 483 int push_flags; 484 int push_cpu; 485 struct irq_work push_work; 486 raw_spinlock_t push_lock; 487 #endif 488 #endif /* CONFIG_SMP */ 489 int rt_queued; 490 491 int rt_throttled; 492 u64 rt_time; 493 u64 rt_runtime; 494 /* Nests inside the rq lock: */ 495 raw_spinlock_t rt_runtime_lock; 496 497 #ifdef CONFIG_RT_GROUP_SCHED 498 unsigned long rt_nr_boosted; 499 500 struct rq *rq; 501 struct task_group *tg; 502 #endif 503 }; 504 505 /* Deadline class' related fields in a runqueue */ 506 struct dl_rq { 507 /* runqueue is an rbtree, ordered by deadline */ 508 struct rb_root rb_root; 509 struct rb_node *rb_leftmost; 510 511 unsigned long dl_nr_running; 512 513 #ifdef CONFIG_SMP 514 /* 515 * Deadline values of the currently executing and the 516 * earliest ready task on this rq. Caching these facilitates 517 * the decision wether or not a ready but not running task 518 * should migrate somewhere else. 519 */ 520 struct { 521 u64 curr; 522 u64 next; 523 } earliest_dl; 524 525 unsigned long dl_nr_migratory; 526 int overloaded; 527 528 /* 529 * Tasks on this rq that can be pushed away. They are kept in 530 * an rb-tree, ordered by tasks' deadlines, with caching 531 * of the leftmost (earliest deadline) element. 532 */ 533 struct rb_root pushable_dl_tasks_root; 534 struct rb_node *pushable_dl_tasks_leftmost; 535 #else 536 struct dl_bw dl_bw; 537 #endif 538 }; 539 540 #ifdef CONFIG_SMP 541 542 /* 543 * We add the notion of a root-domain which will be used to define per-domain 544 * variables. Each exclusive cpuset essentially defines an island domain by 545 * fully partitioning the member cpus from any other cpuset. Whenever a new 546 * exclusive cpuset is created, we also create and attach a new root-domain 547 * object. 548 * 549 */ 550 struct root_domain { 551 atomic_t refcount; 552 atomic_t rto_count; 553 struct rcu_head rcu; 554 cpumask_var_t span; 555 cpumask_var_t online; 556 557 /* Indicate more than one runnable task for any CPU */ 558 bool overload; 559 560 /* 561 * The bit corresponding to a CPU gets set here if such CPU has more 562 * than one runnable -deadline task (as it is below for RT tasks). 563 */ 564 cpumask_var_t dlo_mask; 565 atomic_t dlo_count; 566 struct dl_bw dl_bw; 567 struct cpudl cpudl; 568 569 /* 570 * The "RT overload" flag: it gets set if a CPU has more than 571 * one runnable RT task. 572 */ 573 cpumask_var_t rto_mask; 574 struct cpupri cpupri; 575 576 unsigned long max_cpu_capacity; 577 }; 578 579 extern struct root_domain def_root_domain; 580 581 #endif /* CONFIG_SMP */ 582 583 /* 584 * This is the main, per-CPU runqueue data structure. 585 * 586 * Locking rule: those places that want to lock multiple runqueues 587 * (such as the load balancing or the thread migration code), lock 588 * acquire operations must be ordered by ascending &runqueue. 589 */ 590 struct rq { 591 /* runqueue lock: */ 592 raw_spinlock_t lock; 593 594 /* 595 * nr_running and cpu_load should be in the same cacheline because 596 * remote CPUs use both these fields when doing load calculation. 597 */ 598 unsigned int nr_running; 599 #ifdef CONFIG_NUMA_BALANCING 600 unsigned int nr_numa_running; 601 unsigned int nr_preferred_running; 602 #endif 603 #define CPU_LOAD_IDX_MAX 5 604 unsigned long cpu_load[CPU_LOAD_IDX_MAX]; 605 #ifdef CONFIG_NO_HZ_COMMON 606 #ifdef CONFIG_SMP 607 unsigned long last_load_update_tick; 608 #endif /* CONFIG_SMP */ 609 unsigned long nohz_flags; 610 #endif /* CONFIG_NO_HZ_COMMON */ 611 #ifdef CONFIG_NO_HZ_FULL 612 unsigned long last_sched_tick; 613 #endif 614 /* capture load from *all* tasks on this cpu: */ 615 struct load_weight load; 616 unsigned long nr_load_updates; 617 u64 nr_switches; 618 619 struct cfs_rq cfs; 620 struct rt_rq rt; 621 struct dl_rq dl; 622 623 #ifdef CONFIG_FAIR_GROUP_SCHED 624 /* list of leaf cfs_rq on this cpu: */ 625 struct list_head leaf_cfs_rq_list; 626 #endif /* CONFIG_FAIR_GROUP_SCHED */ 627 628 /* 629 * This is part of a global counter where only the total sum 630 * over all CPUs matters. A task can increase this counter on 631 * one CPU and if it got migrated afterwards it may decrease 632 * it on another CPU. Always updated under the runqueue lock: 633 */ 634 unsigned long nr_uninterruptible; 635 636 struct task_struct *curr, *idle, *stop; 637 unsigned long next_balance; 638 struct mm_struct *prev_mm; 639 640 unsigned int clock_skip_update; 641 u64 clock; 642 u64 clock_task; 643 644 atomic_t nr_iowait; 645 646 #ifdef CONFIG_SMP 647 struct root_domain *rd; 648 struct sched_domain *sd; 649 650 unsigned long cpu_capacity; 651 unsigned long cpu_capacity_orig; 652 653 struct callback_head *balance_callback; 654 655 unsigned char idle_balance; 656 /* For active balancing */ 657 int active_balance; 658 int push_cpu; 659 struct cpu_stop_work active_balance_work; 660 /* cpu of this runqueue: */ 661 int cpu; 662 int online; 663 664 struct list_head cfs_tasks; 665 666 u64 rt_avg; 667 u64 age_stamp; 668 u64 idle_stamp; 669 u64 avg_idle; 670 671 /* This is used to determine avg_idle's max value */ 672 u64 max_idle_balance_cost; 673 #endif 674 675 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 676 u64 prev_irq_time; 677 #endif 678 #ifdef CONFIG_PARAVIRT 679 u64 prev_steal_time; 680 #endif 681 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 682 u64 prev_steal_time_rq; 683 #endif 684 685 /* calc_load related fields */ 686 unsigned long calc_load_update; 687 long calc_load_active; 688 689 #ifdef CONFIG_SCHED_HRTICK 690 #ifdef CONFIG_SMP 691 int hrtick_csd_pending; 692 struct call_single_data hrtick_csd; 693 #endif 694 struct hrtimer hrtick_timer; 695 #endif 696 697 #ifdef CONFIG_SCHEDSTATS 698 /* latency stats */ 699 struct sched_info rq_sched_info; 700 unsigned long long rq_cpu_time; 701 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 702 703 /* sys_sched_yield() stats */ 704 unsigned int yld_count; 705 706 /* schedule() stats */ 707 unsigned int sched_count; 708 unsigned int sched_goidle; 709 710 /* try_to_wake_up() stats */ 711 unsigned int ttwu_count; 712 unsigned int ttwu_local; 713 #endif 714 715 #ifdef CONFIG_SMP 716 struct llist_head wake_list; 717 #endif 718 719 #ifdef CONFIG_CPU_IDLE 720 /* Must be inspected within a rcu lock section */ 721 struct cpuidle_state *idle_state; 722 #endif 723 }; 724 725 static inline int cpu_of(struct rq *rq) 726 { 727 #ifdef CONFIG_SMP 728 return rq->cpu; 729 #else 730 return 0; 731 #endif 732 } 733 734 735 #ifdef CONFIG_SCHED_SMT 736 737 extern struct static_key_false sched_smt_present; 738 739 extern void __update_idle_core(struct rq *rq); 740 741 static inline void update_idle_core(struct rq *rq) 742 { 743 if (static_branch_unlikely(&sched_smt_present)) 744 __update_idle_core(rq); 745 } 746 747 #else 748 static inline void update_idle_core(struct rq *rq) { } 749 #endif 750 751 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 752 753 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 754 #define this_rq() this_cpu_ptr(&runqueues) 755 #define task_rq(p) cpu_rq(task_cpu(p)) 756 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 757 #define raw_rq() raw_cpu_ptr(&runqueues) 758 759 static inline u64 __rq_clock_broken(struct rq *rq) 760 { 761 return READ_ONCE(rq->clock); 762 } 763 764 static inline u64 rq_clock(struct rq *rq) 765 { 766 lockdep_assert_held(&rq->lock); 767 return rq->clock; 768 } 769 770 static inline u64 rq_clock_task(struct rq *rq) 771 { 772 lockdep_assert_held(&rq->lock); 773 return rq->clock_task; 774 } 775 776 #define RQCF_REQ_SKIP 0x01 777 #define RQCF_ACT_SKIP 0x02 778 779 static inline void rq_clock_skip_update(struct rq *rq, bool skip) 780 { 781 lockdep_assert_held(&rq->lock); 782 if (skip) 783 rq->clock_skip_update |= RQCF_REQ_SKIP; 784 else 785 rq->clock_skip_update &= ~RQCF_REQ_SKIP; 786 } 787 788 #ifdef CONFIG_NUMA 789 enum numa_topology_type { 790 NUMA_DIRECT, 791 NUMA_GLUELESS_MESH, 792 NUMA_BACKPLANE, 793 }; 794 extern enum numa_topology_type sched_numa_topology_type; 795 extern int sched_max_numa_distance; 796 extern bool find_numa_distance(int distance); 797 #endif 798 799 #ifdef CONFIG_NUMA_BALANCING 800 /* The regions in numa_faults array from task_struct */ 801 enum numa_faults_stats { 802 NUMA_MEM = 0, 803 NUMA_CPU, 804 NUMA_MEMBUF, 805 NUMA_CPUBUF 806 }; 807 extern void sched_setnuma(struct task_struct *p, int node); 808 extern int migrate_task_to(struct task_struct *p, int cpu); 809 extern int migrate_swap(struct task_struct *, struct task_struct *); 810 #endif /* CONFIG_NUMA_BALANCING */ 811 812 #ifdef CONFIG_SMP 813 814 static inline void 815 queue_balance_callback(struct rq *rq, 816 struct callback_head *head, 817 void (*func)(struct rq *rq)) 818 { 819 lockdep_assert_held(&rq->lock); 820 821 if (unlikely(head->next)) 822 return; 823 824 head->func = (void (*)(struct callback_head *))func; 825 head->next = rq->balance_callback; 826 rq->balance_callback = head; 827 } 828 829 extern void sched_ttwu_pending(void); 830 831 #define rcu_dereference_check_sched_domain(p) \ 832 rcu_dereference_check((p), \ 833 lockdep_is_held(&sched_domains_mutex)) 834 835 /* 836 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 837 * See detach_destroy_domains: synchronize_sched for details. 838 * 839 * The domain tree of any CPU may only be accessed from within 840 * preempt-disabled sections. 841 */ 842 #define for_each_domain(cpu, __sd) \ 843 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 844 __sd; __sd = __sd->parent) 845 846 #define for_each_lower_domain(sd) for (; sd; sd = sd->child) 847 848 /** 849 * highest_flag_domain - Return highest sched_domain containing flag. 850 * @cpu: The cpu whose highest level of sched domain is to 851 * be returned. 852 * @flag: The flag to check for the highest sched_domain 853 * for the given cpu. 854 * 855 * Returns the highest sched_domain of a cpu which contains the given flag. 856 */ 857 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 858 { 859 struct sched_domain *sd, *hsd = NULL; 860 861 for_each_domain(cpu, sd) { 862 if (!(sd->flags & flag)) 863 break; 864 hsd = sd; 865 } 866 867 return hsd; 868 } 869 870 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) 871 { 872 struct sched_domain *sd; 873 874 for_each_domain(cpu, sd) { 875 if (sd->flags & flag) 876 break; 877 } 878 879 return sd; 880 } 881 882 DECLARE_PER_CPU(struct sched_domain *, sd_llc); 883 DECLARE_PER_CPU(int, sd_llc_size); 884 DECLARE_PER_CPU(int, sd_llc_id); 885 DECLARE_PER_CPU(struct sched_domain_shared *, sd_llc_shared); 886 DECLARE_PER_CPU(struct sched_domain *, sd_numa); 887 DECLARE_PER_CPU(struct sched_domain *, sd_asym); 888 889 struct sched_group_capacity { 890 atomic_t ref; 891 /* 892 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity 893 * for a single CPU. 894 */ 895 unsigned int capacity; 896 unsigned long next_update; 897 int imbalance; /* XXX unrelated to capacity but shared group state */ 898 899 unsigned long cpumask[0]; /* iteration mask */ 900 }; 901 902 struct sched_group { 903 struct sched_group *next; /* Must be a circular list */ 904 atomic_t ref; 905 906 unsigned int group_weight; 907 struct sched_group_capacity *sgc; 908 909 /* 910 * The CPUs this group covers. 911 * 912 * NOTE: this field is variable length. (Allocated dynamically 913 * by attaching extra space to the end of the structure, 914 * depending on how many CPUs the kernel has booted up with) 915 */ 916 unsigned long cpumask[0]; 917 }; 918 919 static inline struct cpumask *sched_group_cpus(struct sched_group *sg) 920 { 921 return to_cpumask(sg->cpumask); 922 } 923 924 /* 925 * cpumask masking which cpus in the group are allowed to iterate up the domain 926 * tree. 927 */ 928 static inline struct cpumask *sched_group_mask(struct sched_group *sg) 929 { 930 return to_cpumask(sg->sgc->cpumask); 931 } 932 933 /** 934 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. 935 * @group: The group whose first cpu is to be returned. 936 */ 937 static inline unsigned int group_first_cpu(struct sched_group *group) 938 { 939 return cpumask_first(sched_group_cpus(group)); 940 } 941 942 extern int group_balance_cpu(struct sched_group *sg); 943 944 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 945 void register_sched_domain_sysctl(void); 946 void unregister_sched_domain_sysctl(void); 947 #else 948 static inline void register_sched_domain_sysctl(void) 949 { 950 } 951 static inline void unregister_sched_domain_sysctl(void) 952 { 953 } 954 #endif 955 956 #else 957 958 static inline void sched_ttwu_pending(void) { } 959 960 #endif /* CONFIG_SMP */ 961 962 #include "stats.h" 963 #include "auto_group.h" 964 965 #ifdef CONFIG_CGROUP_SCHED 966 967 /* 968 * Return the group to which this tasks belongs. 969 * 970 * We cannot use task_css() and friends because the cgroup subsystem 971 * changes that value before the cgroup_subsys::attach() method is called, 972 * therefore we cannot pin it and might observe the wrong value. 973 * 974 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup 975 * core changes this before calling sched_move_task(). 976 * 977 * Instead we use a 'copy' which is updated from sched_move_task() while 978 * holding both task_struct::pi_lock and rq::lock. 979 */ 980 static inline struct task_group *task_group(struct task_struct *p) 981 { 982 return p->sched_task_group; 983 } 984 985 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 986 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 987 { 988 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 989 struct task_group *tg = task_group(p); 990 #endif 991 992 #ifdef CONFIG_FAIR_GROUP_SCHED 993 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]); 994 p->se.cfs_rq = tg->cfs_rq[cpu]; 995 p->se.parent = tg->se[cpu]; 996 #endif 997 998 #ifdef CONFIG_RT_GROUP_SCHED 999 p->rt.rt_rq = tg->rt_rq[cpu]; 1000 p->rt.parent = tg->rt_se[cpu]; 1001 #endif 1002 } 1003 1004 #else /* CONFIG_CGROUP_SCHED */ 1005 1006 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 1007 static inline struct task_group *task_group(struct task_struct *p) 1008 { 1009 return NULL; 1010 } 1011 1012 #endif /* CONFIG_CGROUP_SCHED */ 1013 1014 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 1015 { 1016 set_task_rq(p, cpu); 1017 #ifdef CONFIG_SMP 1018 /* 1019 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 1020 * successfuly executed on another CPU. We must ensure that updates of 1021 * per-task data have been completed by this moment. 1022 */ 1023 smp_wmb(); 1024 #ifdef CONFIG_THREAD_INFO_IN_TASK 1025 p->cpu = cpu; 1026 #else 1027 task_thread_info(p)->cpu = cpu; 1028 #endif 1029 p->wake_cpu = cpu; 1030 #endif 1031 } 1032 1033 /* 1034 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 1035 */ 1036 #ifdef CONFIG_SCHED_DEBUG 1037 # include <linux/static_key.h> 1038 # define const_debug __read_mostly 1039 #else 1040 # define const_debug const 1041 #endif 1042 1043 extern const_debug unsigned int sysctl_sched_features; 1044 1045 #define SCHED_FEAT(name, enabled) \ 1046 __SCHED_FEAT_##name , 1047 1048 enum { 1049 #include "features.h" 1050 __SCHED_FEAT_NR, 1051 }; 1052 1053 #undef SCHED_FEAT 1054 1055 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL) 1056 #define SCHED_FEAT(name, enabled) \ 1057 static __always_inline bool static_branch_##name(struct static_key *key) \ 1058 { \ 1059 return static_key_##enabled(key); \ 1060 } 1061 1062 #include "features.h" 1063 1064 #undef SCHED_FEAT 1065 1066 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 1067 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 1068 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */ 1069 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 1070 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */ 1071 1072 extern struct static_key_false sched_numa_balancing; 1073 extern struct static_key_false sched_schedstats; 1074 1075 static inline u64 global_rt_period(void) 1076 { 1077 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 1078 } 1079 1080 static inline u64 global_rt_runtime(void) 1081 { 1082 if (sysctl_sched_rt_runtime < 0) 1083 return RUNTIME_INF; 1084 1085 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 1086 } 1087 1088 static inline int task_current(struct rq *rq, struct task_struct *p) 1089 { 1090 return rq->curr == p; 1091 } 1092 1093 static inline int task_running(struct rq *rq, struct task_struct *p) 1094 { 1095 #ifdef CONFIG_SMP 1096 return p->on_cpu; 1097 #else 1098 return task_current(rq, p); 1099 #endif 1100 } 1101 1102 static inline int task_on_rq_queued(struct task_struct *p) 1103 { 1104 return p->on_rq == TASK_ON_RQ_QUEUED; 1105 } 1106 1107 static inline int task_on_rq_migrating(struct task_struct *p) 1108 { 1109 return p->on_rq == TASK_ON_RQ_MIGRATING; 1110 } 1111 1112 #ifndef prepare_arch_switch 1113 # define prepare_arch_switch(next) do { } while (0) 1114 #endif 1115 #ifndef finish_arch_post_lock_switch 1116 # define finish_arch_post_lock_switch() do { } while (0) 1117 #endif 1118 1119 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 1120 { 1121 #ifdef CONFIG_SMP 1122 /* 1123 * We can optimise this out completely for !SMP, because the 1124 * SMP rebalancing from interrupt is the only thing that cares 1125 * here. 1126 */ 1127 next->on_cpu = 1; 1128 #endif 1129 } 1130 1131 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 1132 { 1133 #ifdef CONFIG_SMP 1134 /* 1135 * After ->on_cpu is cleared, the task can be moved to a different CPU. 1136 * We must ensure this doesn't happen until the switch is completely 1137 * finished. 1138 * 1139 * In particular, the load of prev->state in finish_task_switch() must 1140 * happen before this. 1141 * 1142 * Pairs with the smp_cond_load_acquire() in try_to_wake_up(). 1143 */ 1144 smp_store_release(&prev->on_cpu, 0); 1145 #endif 1146 #ifdef CONFIG_DEBUG_SPINLOCK 1147 /* this is a valid case when another task releases the spinlock */ 1148 rq->lock.owner = current; 1149 #endif 1150 /* 1151 * If we are tracking spinlock dependencies then we have to 1152 * fix up the runqueue lock - which gets 'carried over' from 1153 * prev into current: 1154 */ 1155 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); 1156 1157 raw_spin_unlock_irq(&rq->lock); 1158 } 1159 1160 /* 1161 * wake flags 1162 */ 1163 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */ 1164 #define WF_FORK 0x02 /* child wakeup after fork */ 1165 #define WF_MIGRATED 0x4 /* internal use, task got migrated */ 1166 1167 /* 1168 * To aid in avoiding the subversion of "niceness" due to uneven distribution 1169 * of tasks with abnormal "nice" values across CPUs the contribution that 1170 * each task makes to its run queue's load is weighted according to its 1171 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 1172 * scaled version of the new time slice allocation that they receive on time 1173 * slice expiry etc. 1174 */ 1175 1176 #define WEIGHT_IDLEPRIO 3 1177 #define WMULT_IDLEPRIO 1431655765 1178 1179 extern const int sched_prio_to_weight[40]; 1180 extern const u32 sched_prio_to_wmult[40]; 1181 1182 /* 1183 * {de,en}queue flags: 1184 * 1185 * DEQUEUE_SLEEP - task is no longer runnable 1186 * ENQUEUE_WAKEUP - task just became runnable 1187 * 1188 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks 1189 * are in a known state which allows modification. Such pairs 1190 * should preserve as much state as possible. 1191 * 1192 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location 1193 * in the runqueue. 1194 * 1195 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified) 1196 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline) 1197 * ENQUEUE_MIGRATED - the task was migrated during wakeup 1198 * 1199 */ 1200 1201 #define DEQUEUE_SLEEP 0x01 1202 #define DEQUEUE_SAVE 0x02 /* matches ENQUEUE_RESTORE */ 1203 #define DEQUEUE_MOVE 0x04 /* matches ENQUEUE_MOVE */ 1204 1205 #define ENQUEUE_WAKEUP 0x01 1206 #define ENQUEUE_RESTORE 0x02 1207 #define ENQUEUE_MOVE 0x04 1208 1209 #define ENQUEUE_HEAD 0x08 1210 #define ENQUEUE_REPLENISH 0x10 1211 #ifdef CONFIG_SMP 1212 #define ENQUEUE_MIGRATED 0x20 1213 #else 1214 #define ENQUEUE_MIGRATED 0x00 1215 #endif 1216 1217 #define RETRY_TASK ((void *)-1UL) 1218 1219 struct sched_class { 1220 const struct sched_class *next; 1221 1222 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags); 1223 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags); 1224 void (*yield_task) (struct rq *rq); 1225 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt); 1226 1227 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags); 1228 1229 /* 1230 * It is the responsibility of the pick_next_task() method that will 1231 * return the next task to call put_prev_task() on the @prev task or 1232 * something equivalent. 1233 * 1234 * May return RETRY_TASK when it finds a higher prio class has runnable 1235 * tasks. 1236 */ 1237 struct task_struct * (*pick_next_task) (struct rq *rq, 1238 struct task_struct *prev, 1239 struct pin_cookie cookie); 1240 void (*put_prev_task) (struct rq *rq, struct task_struct *p); 1241 1242 #ifdef CONFIG_SMP 1243 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags); 1244 void (*migrate_task_rq)(struct task_struct *p); 1245 1246 void (*task_woken) (struct rq *this_rq, struct task_struct *task); 1247 1248 void (*set_cpus_allowed)(struct task_struct *p, 1249 const struct cpumask *newmask); 1250 1251 void (*rq_online)(struct rq *rq); 1252 void (*rq_offline)(struct rq *rq); 1253 #endif 1254 1255 void (*set_curr_task) (struct rq *rq); 1256 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued); 1257 void (*task_fork) (struct task_struct *p); 1258 void (*task_dead) (struct task_struct *p); 1259 1260 /* 1261 * The switched_from() call is allowed to drop rq->lock, therefore we 1262 * cannot assume the switched_from/switched_to pair is serliazed by 1263 * rq->lock. They are however serialized by p->pi_lock. 1264 */ 1265 void (*switched_from) (struct rq *this_rq, struct task_struct *task); 1266 void (*switched_to) (struct rq *this_rq, struct task_struct *task); 1267 void (*prio_changed) (struct rq *this_rq, struct task_struct *task, 1268 int oldprio); 1269 1270 unsigned int (*get_rr_interval) (struct rq *rq, 1271 struct task_struct *task); 1272 1273 void (*update_curr) (struct rq *rq); 1274 1275 #define TASK_SET_GROUP 0 1276 #define TASK_MOVE_GROUP 1 1277 1278 #ifdef CONFIG_FAIR_GROUP_SCHED 1279 void (*task_change_group) (struct task_struct *p, int type); 1280 #endif 1281 }; 1282 1283 static inline void put_prev_task(struct rq *rq, struct task_struct *prev) 1284 { 1285 prev->sched_class->put_prev_task(rq, prev); 1286 } 1287 1288 static inline void set_curr_task(struct rq *rq, struct task_struct *curr) 1289 { 1290 curr->sched_class->set_curr_task(rq); 1291 } 1292 1293 #define sched_class_highest (&stop_sched_class) 1294 #define for_each_class(class) \ 1295 for (class = sched_class_highest; class; class = class->next) 1296 1297 extern const struct sched_class stop_sched_class; 1298 extern const struct sched_class dl_sched_class; 1299 extern const struct sched_class rt_sched_class; 1300 extern const struct sched_class fair_sched_class; 1301 extern const struct sched_class idle_sched_class; 1302 1303 1304 #ifdef CONFIG_SMP 1305 1306 extern void update_group_capacity(struct sched_domain *sd, int cpu); 1307 1308 extern void trigger_load_balance(struct rq *rq); 1309 1310 extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask); 1311 1312 #endif 1313 1314 #ifdef CONFIG_CPU_IDLE 1315 static inline void idle_set_state(struct rq *rq, 1316 struct cpuidle_state *idle_state) 1317 { 1318 rq->idle_state = idle_state; 1319 } 1320 1321 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1322 { 1323 SCHED_WARN_ON(!rcu_read_lock_held()); 1324 return rq->idle_state; 1325 } 1326 #else 1327 static inline void idle_set_state(struct rq *rq, 1328 struct cpuidle_state *idle_state) 1329 { 1330 } 1331 1332 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1333 { 1334 return NULL; 1335 } 1336 #endif 1337 1338 extern void sysrq_sched_debug_show(void); 1339 extern void sched_init_granularity(void); 1340 extern void update_max_interval(void); 1341 1342 extern void init_sched_dl_class(void); 1343 extern void init_sched_rt_class(void); 1344 extern void init_sched_fair_class(void); 1345 1346 extern void resched_curr(struct rq *rq); 1347 extern void resched_cpu(int cpu); 1348 1349 extern struct rt_bandwidth def_rt_bandwidth; 1350 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 1351 1352 extern struct dl_bandwidth def_dl_bandwidth; 1353 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime); 1354 extern void init_dl_task_timer(struct sched_dl_entity *dl_se); 1355 1356 unsigned long to_ratio(u64 period, u64 runtime); 1357 1358 extern void init_entity_runnable_average(struct sched_entity *se); 1359 extern void post_init_entity_util_avg(struct sched_entity *se); 1360 1361 #ifdef CONFIG_NO_HZ_FULL 1362 extern bool sched_can_stop_tick(struct rq *rq); 1363 1364 /* 1365 * Tick may be needed by tasks in the runqueue depending on their policy and 1366 * requirements. If tick is needed, lets send the target an IPI to kick it out of 1367 * nohz mode if necessary. 1368 */ 1369 static inline void sched_update_tick_dependency(struct rq *rq) 1370 { 1371 int cpu; 1372 1373 if (!tick_nohz_full_enabled()) 1374 return; 1375 1376 cpu = cpu_of(rq); 1377 1378 if (!tick_nohz_full_cpu(cpu)) 1379 return; 1380 1381 if (sched_can_stop_tick(rq)) 1382 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED); 1383 else 1384 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED); 1385 } 1386 #else 1387 static inline void sched_update_tick_dependency(struct rq *rq) { } 1388 #endif 1389 1390 static inline void add_nr_running(struct rq *rq, unsigned count) 1391 { 1392 unsigned prev_nr = rq->nr_running; 1393 1394 rq->nr_running = prev_nr + count; 1395 1396 if (prev_nr < 2 && rq->nr_running >= 2) { 1397 #ifdef CONFIG_SMP 1398 if (!rq->rd->overload) 1399 rq->rd->overload = true; 1400 #endif 1401 } 1402 1403 sched_update_tick_dependency(rq); 1404 } 1405 1406 static inline void sub_nr_running(struct rq *rq, unsigned count) 1407 { 1408 rq->nr_running -= count; 1409 /* Check if we still need preemption */ 1410 sched_update_tick_dependency(rq); 1411 } 1412 1413 static inline void rq_last_tick_reset(struct rq *rq) 1414 { 1415 #ifdef CONFIG_NO_HZ_FULL 1416 rq->last_sched_tick = jiffies; 1417 #endif 1418 } 1419 1420 extern void update_rq_clock(struct rq *rq); 1421 1422 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 1423 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 1424 1425 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); 1426 1427 extern const_debug unsigned int sysctl_sched_time_avg; 1428 extern const_debug unsigned int sysctl_sched_nr_migrate; 1429 extern const_debug unsigned int sysctl_sched_migration_cost; 1430 1431 static inline u64 sched_avg_period(void) 1432 { 1433 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; 1434 } 1435 1436 #ifdef CONFIG_SCHED_HRTICK 1437 1438 /* 1439 * Use hrtick when: 1440 * - enabled by features 1441 * - hrtimer is actually high res 1442 */ 1443 static inline int hrtick_enabled(struct rq *rq) 1444 { 1445 if (!sched_feat(HRTICK)) 1446 return 0; 1447 if (!cpu_active(cpu_of(rq))) 1448 return 0; 1449 return hrtimer_is_hres_active(&rq->hrtick_timer); 1450 } 1451 1452 void hrtick_start(struct rq *rq, u64 delay); 1453 1454 #else 1455 1456 static inline int hrtick_enabled(struct rq *rq) 1457 { 1458 return 0; 1459 } 1460 1461 #endif /* CONFIG_SCHED_HRTICK */ 1462 1463 #ifdef CONFIG_SMP 1464 extern void sched_avg_update(struct rq *rq); 1465 1466 #ifndef arch_scale_freq_capacity 1467 static __always_inline 1468 unsigned long arch_scale_freq_capacity(struct sched_domain *sd, int cpu) 1469 { 1470 return SCHED_CAPACITY_SCALE; 1471 } 1472 #endif 1473 1474 #ifndef arch_scale_cpu_capacity 1475 static __always_inline 1476 unsigned long arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) 1477 { 1478 if (sd && (sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) 1479 return sd->smt_gain / sd->span_weight; 1480 1481 return SCHED_CAPACITY_SCALE; 1482 } 1483 #endif 1484 1485 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) 1486 { 1487 rq->rt_avg += rt_delta * arch_scale_freq_capacity(NULL, cpu_of(rq)); 1488 sched_avg_update(rq); 1489 } 1490 #else 1491 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { } 1492 static inline void sched_avg_update(struct rq *rq) { } 1493 #endif 1494 1495 struct rq_flags { 1496 unsigned long flags; 1497 struct pin_cookie cookie; 1498 }; 1499 1500 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1501 __acquires(rq->lock); 1502 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1503 __acquires(p->pi_lock) 1504 __acquires(rq->lock); 1505 1506 static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf) 1507 __releases(rq->lock) 1508 { 1509 lockdep_unpin_lock(&rq->lock, rf->cookie); 1510 raw_spin_unlock(&rq->lock); 1511 } 1512 1513 static inline void 1514 task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf) 1515 __releases(rq->lock) 1516 __releases(p->pi_lock) 1517 { 1518 lockdep_unpin_lock(&rq->lock, rf->cookie); 1519 raw_spin_unlock(&rq->lock); 1520 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); 1521 } 1522 1523 #ifdef CONFIG_SMP 1524 #ifdef CONFIG_PREEMPT 1525 1526 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2); 1527 1528 /* 1529 * fair double_lock_balance: Safely acquires both rq->locks in a fair 1530 * way at the expense of forcing extra atomic operations in all 1531 * invocations. This assures that the double_lock is acquired using the 1532 * same underlying policy as the spinlock_t on this architecture, which 1533 * reduces latency compared to the unfair variant below. However, it 1534 * also adds more overhead and therefore may reduce throughput. 1535 */ 1536 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1537 __releases(this_rq->lock) 1538 __acquires(busiest->lock) 1539 __acquires(this_rq->lock) 1540 { 1541 raw_spin_unlock(&this_rq->lock); 1542 double_rq_lock(this_rq, busiest); 1543 1544 return 1; 1545 } 1546 1547 #else 1548 /* 1549 * Unfair double_lock_balance: Optimizes throughput at the expense of 1550 * latency by eliminating extra atomic operations when the locks are 1551 * already in proper order on entry. This favors lower cpu-ids and will 1552 * grant the double lock to lower cpus over higher ids under contention, 1553 * regardless of entry order into the function. 1554 */ 1555 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1556 __releases(this_rq->lock) 1557 __acquires(busiest->lock) 1558 __acquires(this_rq->lock) 1559 { 1560 int ret = 0; 1561 1562 if (unlikely(!raw_spin_trylock(&busiest->lock))) { 1563 if (busiest < this_rq) { 1564 raw_spin_unlock(&this_rq->lock); 1565 raw_spin_lock(&busiest->lock); 1566 raw_spin_lock_nested(&this_rq->lock, 1567 SINGLE_DEPTH_NESTING); 1568 ret = 1; 1569 } else 1570 raw_spin_lock_nested(&busiest->lock, 1571 SINGLE_DEPTH_NESTING); 1572 } 1573 return ret; 1574 } 1575 1576 #endif /* CONFIG_PREEMPT */ 1577 1578 /* 1579 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 1580 */ 1581 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 1582 { 1583 if (unlikely(!irqs_disabled())) { 1584 /* printk() doesn't work good under rq->lock */ 1585 raw_spin_unlock(&this_rq->lock); 1586 BUG_ON(1); 1587 } 1588 1589 return _double_lock_balance(this_rq, busiest); 1590 } 1591 1592 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 1593 __releases(busiest->lock) 1594 { 1595 raw_spin_unlock(&busiest->lock); 1596 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1597 } 1598 1599 static inline void double_lock(spinlock_t *l1, spinlock_t *l2) 1600 { 1601 if (l1 > l2) 1602 swap(l1, l2); 1603 1604 spin_lock(l1); 1605 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1606 } 1607 1608 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2) 1609 { 1610 if (l1 > l2) 1611 swap(l1, l2); 1612 1613 spin_lock_irq(l1); 1614 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1615 } 1616 1617 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2) 1618 { 1619 if (l1 > l2) 1620 swap(l1, l2); 1621 1622 raw_spin_lock(l1); 1623 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1624 } 1625 1626 /* 1627 * double_rq_lock - safely lock two runqueues 1628 * 1629 * Note this does not disable interrupts like task_rq_lock, 1630 * you need to do so manually before calling. 1631 */ 1632 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1633 __acquires(rq1->lock) 1634 __acquires(rq2->lock) 1635 { 1636 BUG_ON(!irqs_disabled()); 1637 if (rq1 == rq2) { 1638 raw_spin_lock(&rq1->lock); 1639 __acquire(rq2->lock); /* Fake it out ;) */ 1640 } else { 1641 if (rq1 < rq2) { 1642 raw_spin_lock(&rq1->lock); 1643 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); 1644 } else { 1645 raw_spin_lock(&rq2->lock); 1646 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); 1647 } 1648 } 1649 } 1650 1651 /* 1652 * double_rq_unlock - safely unlock two runqueues 1653 * 1654 * Note this does not restore interrupts like task_rq_unlock, 1655 * you need to do so manually after calling. 1656 */ 1657 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1658 __releases(rq1->lock) 1659 __releases(rq2->lock) 1660 { 1661 raw_spin_unlock(&rq1->lock); 1662 if (rq1 != rq2) 1663 raw_spin_unlock(&rq2->lock); 1664 else 1665 __release(rq2->lock); 1666 } 1667 1668 #else /* CONFIG_SMP */ 1669 1670 /* 1671 * double_rq_lock - safely lock two runqueues 1672 * 1673 * Note this does not disable interrupts like task_rq_lock, 1674 * you need to do so manually before calling. 1675 */ 1676 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1677 __acquires(rq1->lock) 1678 __acquires(rq2->lock) 1679 { 1680 BUG_ON(!irqs_disabled()); 1681 BUG_ON(rq1 != rq2); 1682 raw_spin_lock(&rq1->lock); 1683 __acquire(rq2->lock); /* Fake it out ;) */ 1684 } 1685 1686 /* 1687 * double_rq_unlock - safely unlock two runqueues 1688 * 1689 * Note this does not restore interrupts like task_rq_unlock, 1690 * you need to do so manually after calling. 1691 */ 1692 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1693 __releases(rq1->lock) 1694 __releases(rq2->lock) 1695 { 1696 BUG_ON(rq1 != rq2); 1697 raw_spin_unlock(&rq1->lock); 1698 __release(rq2->lock); 1699 } 1700 1701 #endif 1702 1703 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 1704 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 1705 1706 #ifdef CONFIG_SCHED_DEBUG 1707 extern void print_cfs_stats(struct seq_file *m, int cpu); 1708 extern void print_rt_stats(struct seq_file *m, int cpu); 1709 extern void print_dl_stats(struct seq_file *m, int cpu); 1710 extern void 1711 print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq); 1712 1713 #ifdef CONFIG_NUMA_BALANCING 1714 extern void 1715 show_numa_stats(struct task_struct *p, struct seq_file *m); 1716 extern void 1717 print_numa_stats(struct seq_file *m, int node, unsigned long tsf, 1718 unsigned long tpf, unsigned long gsf, unsigned long gpf); 1719 #endif /* CONFIG_NUMA_BALANCING */ 1720 #endif /* CONFIG_SCHED_DEBUG */ 1721 1722 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 1723 extern void init_rt_rq(struct rt_rq *rt_rq); 1724 extern void init_dl_rq(struct dl_rq *dl_rq); 1725 1726 extern void cfs_bandwidth_usage_inc(void); 1727 extern void cfs_bandwidth_usage_dec(void); 1728 1729 #ifdef CONFIG_NO_HZ_COMMON 1730 enum rq_nohz_flag_bits { 1731 NOHZ_TICK_STOPPED, 1732 NOHZ_BALANCE_KICK, 1733 }; 1734 1735 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 1736 1737 extern void nohz_balance_exit_idle(unsigned int cpu); 1738 #else 1739 static inline void nohz_balance_exit_idle(unsigned int cpu) { } 1740 #endif 1741 1742 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1743 struct irqtime { 1744 u64 hardirq_time; 1745 u64 softirq_time; 1746 u64 irq_start_time; 1747 struct u64_stats_sync sync; 1748 }; 1749 1750 DECLARE_PER_CPU(struct irqtime, cpu_irqtime); 1751 1752 static inline u64 irq_time_read(int cpu) 1753 { 1754 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu); 1755 unsigned int seq; 1756 u64 total; 1757 1758 do { 1759 seq = __u64_stats_fetch_begin(&irqtime->sync); 1760 total = irqtime->softirq_time + irqtime->hardirq_time; 1761 } while (__u64_stats_fetch_retry(&irqtime->sync, seq)); 1762 1763 return total; 1764 } 1765 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 1766 1767 #ifdef CONFIG_CPU_FREQ 1768 DECLARE_PER_CPU(struct update_util_data *, cpufreq_update_util_data); 1769 1770 /** 1771 * cpufreq_update_util - Take a note about CPU utilization changes. 1772 * @rq: Runqueue to carry out the update for. 1773 * @flags: Update reason flags. 1774 * 1775 * This function is called by the scheduler on the CPU whose utilization is 1776 * being updated. 1777 * 1778 * It can only be called from RCU-sched read-side critical sections. 1779 * 1780 * The way cpufreq is currently arranged requires it to evaluate the CPU 1781 * performance state (frequency/voltage) on a regular basis to prevent it from 1782 * being stuck in a completely inadequate performance level for too long. 1783 * That is not guaranteed to happen if the updates are only triggered from CFS, 1784 * though, because they may not be coming in if RT or deadline tasks are active 1785 * all the time (or there are RT and DL tasks only). 1786 * 1787 * As a workaround for that issue, this function is called by the RT and DL 1788 * sched classes to trigger extra cpufreq updates to prevent it from stalling, 1789 * but that really is a band-aid. Going forward it should be replaced with 1790 * solutions targeted more specifically at RT and DL tasks. 1791 */ 1792 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) 1793 { 1794 struct update_util_data *data; 1795 1796 data = rcu_dereference_sched(*this_cpu_ptr(&cpufreq_update_util_data)); 1797 if (data) 1798 data->func(data, rq_clock(rq), flags); 1799 } 1800 1801 static inline void cpufreq_update_this_cpu(struct rq *rq, unsigned int flags) 1802 { 1803 if (cpu_of(rq) == smp_processor_id()) 1804 cpufreq_update_util(rq, flags); 1805 } 1806 #else 1807 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {} 1808 static inline void cpufreq_update_this_cpu(struct rq *rq, unsigned int flags) {} 1809 #endif /* CONFIG_CPU_FREQ */ 1810 1811 #ifdef arch_scale_freq_capacity 1812 #ifndef arch_scale_freq_invariant 1813 #define arch_scale_freq_invariant() (true) 1814 #endif 1815 #else /* arch_scale_freq_capacity */ 1816 #define arch_scale_freq_invariant() (false) 1817 #endif 1818