1 2 #include <linux/sched.h> 3 #include <linux/sched/sysctl.h> 4 #include <linux/sched/rt.h> 5 #include <linux/sched/deadline.h> 6 #include <linux/mutex.h> 7 #include <linux/spinlock.h> 8 #include <linux/stop_machine.h> 9 #include <linux/tick.h> 10 #include <linux/slab.h> 11 12 #include "cpupri.h" 13 #include "cpudeadline.h" 14 #include "cpuacct.h" 15 16 struct rq; 17 18 extern __read_mostly int scheduler_running; 19 20 extern unsigned long calc_load_update; 21 extern atomic_long_t calc_load_tasks; 22 23 extern long calc_load_fold_active(struct rq *this_rq); 24 extern void update_cpu_load_active(struct rq *this_rq); 25 26 /* 27 * Convert user-nice values [ -20 ... 0 ... 19 ] 28 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], 29 * and back. 30 */ 31 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) 32 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) 33 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) 34 35 /* 36 * 'User priority' is the nice value converted to something we 37 * can work with better when scaling various scheduler parameters, 38 * it's a [ 0 ... 39 ] range. 39 */ 40 #define USER_PRIO(p) ((p)-MAX_RT_PRIO) 41 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) 42 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) 43 44 /* 45 * Helpers for converting nanosecond timing to jiffy resolution 46 */ 47 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) 48 49 /* 50 * Increase resolution of nice-level calculations for 64-bit architectures. 51 * The extra resolution improves shares distribution and load balancing of 52 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup 53 * hierarchies, especially on larger systems. This is not a user-visible change 54 * and does not change the user-interface for setting shares/weights. 55 * 56 * We increase resolution only if we have enough bits to allow this increased 57 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution 58 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the 59 * increased costs. 60 */ 61 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */ 62 # define SCHED_LOAD_RESOLUTION 10 63 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION) 64 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION) 65 #else 66 # define SCHED_LOAD_RESOLUTION 0 67 # define scale_load(w) (w) 68 # define scale_load_down(w) (w) 69 #endif 70 71 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION) 72 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT) 73 74 #define NICE_0_LOAD SCHED_LOAD_SCALE 75 #define NICE_0_SHIFT SCHED_LOAD_SHIFT 76 77 /* 78 * Single value that decides SCHED_DEADLINE internal math precision. 79 * 10 -> just above 1us 80 * 9 -> just above 0.5us 81 */ 82 #define DL_SCALE (10) 83 84 /* 85 * These are the 'tuning knobs' of the scheduler: 86 */ 87 88 /* 89 * single value that denotes runtime == period, ie unlimited time. 90 */ 91 #define RUNTIME_INF ((u64)~0ULL) 92 93 static inline int fair_policy(int policy) 94 { 95 return policy == SCHED_NORMAL || policy == SCHED_BATCH; 96 } 97 98 static inline int rt_policy(int policy) 99 { 100 return policy == SCHED_FIFO || policy == SCHED_RR; 101 } 102 103 static inline int dl_policy(int policy) 104 { 105 return policy == SCHED_DEADLINE; 106 } 107 108 static inline int task_has_rt_policy(struct task_struct *p) 109 { 110 return rt_policy(p->policy); 111 } 112 113 static inline int task_has_dl_policy(struct task_struct *p) 114 { 115 return dl_policy(p->policy); 116 } 117 118 static inline bool dl_time_before(u64 a, u64 b) 119 { 120 return (s64)(a - b) < 0; 121 } 122 123 /* 124 * Tells if entity @a should preempt entity @b. 125 */ 126 static inline bool 127 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b) 128 { 129 return dl_time_before(a->deadline, b->deadline); 130 } 131 132 /* 133 * This is the priority-queue data structure of the RT scheduling class: 134 */ 135 struct rt_prio_array { 136 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ 137 struct list_head queue[MAX_RT_PRIO]; 138 }; 139 140 struct rt_bandwidth { 141 /* nests inside the rq lock: */ 142 raw_spinlock_t rt_runtime_lock; 143 ktime_t rt_period; 144 u64 rt_runtime; 145 struct hrtimer rt_period_timer; 146 }; 147 /* 148 * To keep the bandwidth of -deadline tasks and groups under control 149 * we need some place where: 150 * - store the maximum -deadline bandwidth of the system (the group); 151 * - cache the fraction of that bandwidth that is currently allocated. 152 * 153 * This is all done in the data structure below. It is similar to the 154 * one used for RT-throttling (rt_bandwidth), with the main difference 155 * that, since here we are only interested in admission control, we 156 * do not decrease any runtime while the group "executes", neither we 157 * need a timer to replenish it. 158 * 159 * With respect to SMP, the bandwidth is given on a per-CPU basis, 160 * meaning that: 161 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU; 162 * - dl_total_bw array contains, in the i-eth element, the currently 163 * allocated bandwidth on the i-eth CPU. 164 * Moreover, groups consume bandwidth on each CPU, while tasks only 165 * consume bandwidth on the CPU they're running on. 166 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw 167 * that will be shown the next time the proc or cgroup controls will 168 * be red. It on its turn can be changed by writing on its own 169 * control. 170 */ 171 struct dl_bandwidth { 172 raw_spinlock_t dl_runtime_lock; 173 u64 dl_runtime; 174 u64 dl_period; 175 }; 176 177 static inline int dl_bandwidth_enabled(void) 178 { 179 return sysctl_sched_rt_runtime >= 0; 180 } 181 182 extern struct dl_bw *dl_bw_of(int i); 183 184 struct dl_bw { 185 raw_spinlock_t lock; 186 u64 bw, total_bw; 187 }; 188 189 extern struct mutex sched_domains_mutex; 190 191 #ifdef CONFIG_CGROUP_SCHED 192 193 #include <linux/cgroup.h> 194 195 struct cfs_rq; 196 struct rt_rq; 197 198 extern struct list_head task_groups; 199 200 struct cfs_bandwidth { 201 #ifdef CONFIG_CFS_BANDWIDTH 202 raw_spinlock_t lock; 203 ktime_t period; 204 u64 quota, runtime; 205 s64 hierarchal_quota; 206 u64 runtime_expires; 207 208 int idle, timer_active; 209 struct hrtimer period_timer, slack_timer; 210 struct list_head throttled_cfs_rq; 211 212 /* statistics */ 213 int nr_periods, nr_throttled; 214 u64 throttled_time; 215 #endif 216 }; 217 218 /* task group related information */ 219 struct task_group { 220 struct cgroup_subsys_state css; 221 222 #ifdef CONFIG_FAIR_GROUP_SCHED 223 /* schedulable entities of this group on each cpu */ 224 struct sched_entity **se; 225 /* runqueue "owned" by this group on each cpu */ 226 struct cfs_rq **cfs_rq; 227 unsigned long shares; 228 229 #ifdef CONFIG_SMP 230 atomic_long_t load_avg; 231 atomic_t runnable_avg; 232 #endif 233 #endif 234 235 #ifdef CONFIG_RT_GROUP_SCHED 236 struct sched_rt_entity **rt_se; 237 struct rt_rq **rt_rq; 238 239 struct rt_bandwidth rt_bandwidth; 240 #endif 241 242 struct rcu_head rcu; 243 struct list_head list; 244 245 struct task_group *parent; 246 struct list_head siblings; 247 struct list_head children; 248 249 #ifdef CONFIG_SCHED_AUTOGROUP 250 struct autogroup *autogroup; 251 #endif 252 253 struct cfs_bandwidth cfs_bandwidth; 254 }; 255 256 #ifdef CONFIG_FAIR_GROUP_SCHED 257 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD 258 259 /* 260 * A weight of 0 or 1 can cause arithmetics problems. 261 * A weight of a cfs_rq is the sum of weights of which entities 262 * are queued on this cfs_rq, so a weight of a entity should not be 263 * too large, so as the shares value of a task group. 264 * (The default weight is 1024 - so there's no practical 265 * limitation from this.) 266 */ 267 #define MIN_SHARES (1UL << 1) 268 #define MAX_SHARES (1UL << 18) 269 #endif 270 271 typedef int (*tg_visitor)(struct task_group *, void *); 272 273 extern int walk_tg_tree_from(struct task_group *from, 274 tg_visitor down, tg_visitor up, void *data); 275 276 /* 277 * Iterate the full tree, calling @down when first entering a node and @up when 278 * leaving it for the final time. 279 * 280 * Caller must hold rcu_lock or sufficient equivalent. 281 */ 282 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) 283 { 284 return walk_tg_tree_from(&root_task_group, down, up, data); 285 } 286 287 extern int tg_nop(struct task_group *tg, void *data); 288 289 extern void free_fair_sched_group(struct task_group *tg); 290 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent); 291 extern void unregister_fair_sched_group(struct task_group *tg, int cpu); 292 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 293 struct sched_entity *se, int cpu, 294 struct sched_entity *parent); 295 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 296 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 297 298 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b); 299 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 300 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq); 301 302 extern void free_rt_sched_group(struct task_group *tg); 303 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent); 304 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 305 struct sched_rt_entity *rt_se, int cpu, 306 struct sched_rt_entity *parent); 307 308 extern struct task_group *sched_create_group(struct task_group *parent); 309 extern void sched_online_group(struct task_group *tg, 310 struct task_group *parent); 311 extern void sched_destroy_group(struct task_group *tg); 312 extern void sched_offline_group(struct task_group *tg); 313 314 extern void sched_move_task(struct task_struct *tsk); 315 316 #ifdef CONFIG_FAIR_GROUP_SCHED 317 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 318 #endif 319 320 #else /* CONFIG_CGROUP_SCHED */ 321 322 struct cfs_bandwidth { }; 323 324 #endif /* CONFIG_CGROUP_SCHED */ 325 326 /* CFS-related fields in a runqueue */ 327 struct cfs_rq { 328 struct load_weight load; 329 unsigned int nr_running, h_nr_running; 330 331 u64 exec_clock; 332 u64 min_vruntime; 333 #ifndef CONFIG_64BIT 334 u64 min_vruntime_copy; 335 #endif 336 337 struct rb_root tasks_timeline; 338 struct rb_node *rb_leftmost; 339 340 /* 341 * 'curr' points to currently running entity on this cfs_rq. 342 * It is set to NULL otherwise (i.e when none are currently running). 343 */ 344 struct sched_entity *curr, *next, *last, *skip; 345 346 #ifdef CONFIG_SCHED_DEBUG 347 unsigned int nr_spread_over; 348 #endif 349 350 #ifdef CONFIG_SMP 351 /* 352 * CFS Load tracking 353 * Under CFS, load is tracked on a per-entity basis and aggregated up. 354 * This allows for the description of both thread and group usage (in 355 * the FAIR_GROUP_SCHED case). 356 */ 357 unsigned long runnable_load_avg, blocked_load_avg; 358 atomic64_t decay_counter; 359 u64 last_decay; 360 atomic_long_t removed_load; 361 362 #ifdef CONFIG_FAIR_GROUP_SCHED 363 /* Required to track per-cpu representation of a task_group */ 364 u32 tg_runnable_contrib; 365 unsigned long tg_load_contrib; 366 367 /* 368 * h_load = weight * f(tg) 369 * 370 * Where f(tg) is the recursive weight fraction assigned to 371 * this group. 372 */ 373 unsigned long h_load; 374 u64 last_h_load_update; 375 struct sched_entity *h_load_next; 376 #endif /* CONFIG_FAIR_GROUP_SCHED */ 377 #endif /* CONFIG_SMP */ 378 379 #ifdef CONFIG_FAIR_GROUP_SCHED 380 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ 381 382 /* 383 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in 384 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities 385 * (like users, containers etc.) 386 * 387 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This 388 * list is used during load balance. 389 */ 390 int on_list; 391 struct list_head leaf_cfs_rq_list; 392 struct task_group *tg; /* group that "owns" this runqueue */ 393 394 #ifdef CONFIG_CFS_BANDWIDTH 395 int runtime_enabled; 396 u64 runtime_expires; 397 s64 runtime_remaining; 398 399 u64 throttled_clock, throttled_clock_task; 400 u64 throttled_clock_task_time; 401 int throttled, throttle_count; 402 struct list_head throttled_list; 403 #endif /* CONFIG_CFS_BANDWIDTH */ 404 #endif /* CONFIG_FAIR_GROUP_SCHED */ 405 }; 406 407 static inline int rt_bandwidth_enabled(void) 408 { 409 return sysctl_sched_rt_runtime >= 0; 410 } 411 412 /* Real-Time classes' related field in a runqueue: */ 413 struct rt_rq { 414 struct rt_prio_array active; 415 unsigned int rt_nr_running; 416 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 417 struct { 418 int curr; /* highest queued rt task prio */ 419 #ifdef CONFIG_SMP 420 int next; /* next highest */ 421 #endif 422 } highest_prio; 423 #endif 424 #ifdef CONFIG_SMP 425 unsigned long rt_nr_migratory; 426 unsigned long rt_nr_total; 427 int overloaded; 428 struct plist_head pushable_tasks; 429 #endif 430 int rt_throttled; 431 u64 rt_time; 432 u64 rt_runtime; 433 /* Nests inside the rq lock: */ 434 raw_spinlock_t rt_runtime_lock; 435 436 #ifdef CONFIG_RT_GROUP_SCHED 437 unsigned long rt_nr_boosted; 438 439 struct rq *rq; 440 struct task_group *tg; 441 #endif 442 }; 443 444 /* Deadline class' related fields in a runqueue */ 445 struct dl_rq { 446 /* runqueue is an rbtree, ordered by deadline */ 447 struct rb_root rb_root; 448 struct rb_node *rb_leftmost; 449 450 unsigned long dl_nr_running; 451 452 #ifdef CONFIG_SMP 453 /* 454 * Deadline values of the currently executing and the 455 * earliest ready task on this rq. Caching these facilitates 456 * the decision wether or not a ready but not running task 457 * should migrate somewhere else. 458 */ 459 struct { 460 u64 curr; 461 u64 next; 462 } earliest_dl; 463 464 unsigned long dl_nr_migratory; 465 unsigned long dl_nr_total; 466 int overloaded; 467 468 /* 469 * Tasks on this rq that can be pushed away. They are kept in 470 * an rb-tree, ordered by tasks' deadlines, with caching 471 * of the leftmost (earliest deadline) element. 472 */ 473 struct rb_root pushable_dl_tasks_root; 474 struct rb_node *pushable_dl_tasks_leftmost; 475 #else 476 struct dl_bw dl_bw; 477 #endif 478 }; 479 480 #ifdef CONFIG_SMP 481 482 /* 483 * We add the notion of a root-domain which will be used to define per-domain 484 * variables. Each exclusive cpuset essentially defines an island domain by 485 * fully partitioning the member cpus from any other cpuset. Whenever a new 486 * exclusive cpuset is created, we also create and attach a new root-domain 487 * object. 488 * 489 */ 490 struct root_domain { 491 atomic_t refcount; 492 atomic_t rto_count; 493 struct rcu_head rcu; 494 cpumask_var_t span; 495 cpumask_var_t online; 496 497 /* 498 * The bit corresponding to a CPU gets set here if such CPU has more 499 * than one runnable -deadline task (as it is below for RT tasks). 500 */ 501 cpumask_var_t dlo_mask; 502 atomic_t dlo_count; 503 struct dl_bw dl_bw; 504 struct cpudl cpudl; 505 506 /* 507 * The "RT overload" flag: it gets set if a CPU has more than 508 * one runnable RT task. 509 */ 510 cpumask_var_t rto_mask; 511 struct cpupri cpupri; 512 }; 513 514 extern struct root_domain def_root_domain; 515 516 #endif /* CONFIG_SMP */ 517 518 /* 519 * This is the main, per-CPU runqueue data structure. 520 * 521 * Locking rule: those places that want to lock multiple runqueues 522 * (such as the load balancing or the thread migration code), lock 523 * acquire operations must be ordered by ascending &runqueue. 524 */ 525 struct rq { 526 /* runqueue lock: */ 527 raw_spinlock_t lock; 528 529 /* 530 * nr_running and cpu_load should be in the same cacheline because 531 * remote CPUs use both these fields when doing load calculation. 532 */ 533 unsigned int nr_running; 534 #ifdef CONFIG_NUMA_BALANCING 535 unsigned int nr_numa_running; 536 unsigned int nr_preferred_running; 537 #endif 538 #define CPU_LOAD_IDX_MAX 5 539 unsigned long cpu_load[CPU_LOAD_IDX_MAX]; 540 unsigned long last_load_update_tick; 541 #ifdef CONFIG_NO_HZ_COMMON 542 u64 nohz_stamp; 543 unsigned long nohz_flags; 544 #endif 545 #ifdef CONFIG_NO_HZ_FULL 546 unsigned long last_sched_tick; 547 #endif 548 int skip_clock_update; 549 550 /* capture load from *all* tasks on this cpu: */ 551 struct load_weight load; 552 unsigned long nr_load_updates; 553 u64 nr_switches; 554 555 struct cfs_rq cfs; 556 struct rt_rq rt; 557 struct dl_rq dl; 558 559 #ifdef CONFIG_FAIR_GROUP_SCHED 560 /* list of leaf cfs_rq on this cpu: */ 561 struct list_head leaf_cfs_rq_list; 562 #endif /* CONFIG_FAIR_GROUP_SCHED */ 563 564 #ifdef CONFIG_RT_GROUP_SCHED 565 struct list_head leaf_rt_rq_list; 566 #endif 567 568 /* 569 * This is part of a global counter where only the total sum 570 * over all CPUs matters. A task can increase this counter on 571 * one CPU and if it got migrated afterwards it may decrease 572 * it on another CPU. Always updated under the runqueue lock: 573 */ 574 unsigned long nr_uninterruptible; 575 576 struct task_struct *curr, *idle, *stop; 577 unsigned long next_balance; 578 struct mm_struct *prev_mm; 579 580 u64 clock; 581 u64 clock_task; 582 583 atomic_t nr_iowait; 584 585 #ifdef CONFIG_SMP 586 struct root_domain *rd; 587 struct sched_domain *sd; 588 589 unsigned long cpu_power; 590 591 unsigned char idle_balance; 592 /* For active balancing */ 593 int post_schedule; 594 int active_balance; 595 int push_cpu; 596 struct cpu_stop_work active_balance_work; 597 /* cpu of this runqueue: */ 598 int cpu; 599 int online; 600 601 struct list_head cfs_tasks; 602 603 u64 rt_avg; 604 u64 age_stamp; 605 u64 idle_stamp; 606 u64 avg_idle; 607 608 /* This is used to determine avg_idle's max value */ 609 u64 max_idle_balance_cost; 610 #endif 611 612 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 613 u64 prev_irq_time; 614 #endif 615 #ifdef CONFIG_PARAVIRT 616 u64 prev_steal_time; 617 #endif 618 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 619 u64 prev_steal_time_rq; 620 #endif 621 622 /* calc_load related fields */ 623 unsigned long calc_load_update; 624 long calc_load_active; 625 626 #ifdef CONFIG_SCHED_HRTICK 627 #ifdef CONFIG_SMP 628 int hrtick_csd_pending; 629 struct call_single_data hrtick_csd; 630 #endif 631 struct hrtimer hrtick_timer; 632 #endif 633 634 #ifdef CONFIG_SCHEDSTATS 635 /* latency stats */ 636 struct sched_info rq_sched_info; 637 unsigned long long rq_cpu_time; 638 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 639 640 /* sys_sched_yield() stats */ 641 unsigned int yld_count; 642 643 /* schedule() stats */ 644 unsigned int sched_count; 645 unsigned int sched_goidle; 646 647 /* try_to_wake_up() stats */ 648 unsigned int ttwu_count; 649 unsigned int ttwu_local; 650 #endif 651 652 #ifdef CONFIG_SMP 653 struct llist_head wake_list; 654 #endif 655 656 struct sched_avg avg; 657 }; 658 659 static inline int cpu_of(struct rq *rq) 660 { 661 #ifdef CONFIG_SMP 662 return rq->cpu; 663 #else 664 return 0; 665 #endif 666 } 667 668 DECLARE_PER_CPU(struct rq, runqueues); 669 670 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 671 #define this_rq() (&__get_cpu_var(runqueues)) 672 #define task_rq(p) cpu_rq(task_cpu(p)) 673 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 674 #define raw_rq() (&__raw_get_cpu_var(runqueues)) 675 676 static inline u64 rq_clock(struct rq *rq) 677 { 678 return rq->clock; 679 } 680 681 static inline u64 rq_clock_task(struct rq *rq) 682 { 683 return rq->clock_task; 684 } 685 686 #ifdef CONFIG_NUMA_BALANCING 687 extern void sched_setnuma(struct task_struct *p, int node); 688 extern int migrate_task_to(struct task_struct *p, int cpu); 689 extern int migrate_swap(struct task_struct *, struct task_struct *); 690 #endif /* CONFIG_NUMA_BALANCING */ 691 692 #ifdef CONFIG_SMP 693 694 #define rcu_dereference_check_sched_domain(p) \ 695 rcu_dereference_check((p), \ 696 lockdep_is_held(&sched_domains_mutex)) 697 698 /* 699 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 700 * See detach_destroy_domains: synchronize_sched for details. 701 * 702 * The domain tree of any CPU may only be accessed from within 703 * preempt-disabled sections. 704 */ 705 #define for_each_domain(cpu, __sd) \ 706 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 707 __sd; __sd = __sd->parent) 708 709 #define for_each_lower_domain(sd) for (; sd; sd = sd->child) 710 711 /** 712 * highest_flag_domain - Return highest sched_domain containing flag. 713 * @cpu: The cpu whose highest level of sched domain is to 714 * be returned. 715 * @flag: The flag to check for the highest sched_domain 716 * for the given cpu. 717 * 718 * Returns the highest sched_domain of a cpu which contains the given flag. 719 */ 720 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 721 { 722 struct sched_domain *sd, *hsd = NULL; 723 724 for_each_domain(cpu, sd) { 725 if (!(sd->flags & flag)) 726 break; 727 hsd = sd; 728 } 729 730 return hsd; 731 } 732 733 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) 734 { 735 struct sched_domain *sd; 736 737 for_each_domain(cpu, sd) { 738 if (sd->flags & flag) 739 break; 740 } 741 742 return sd; 743 } 744 745 DECLARE_PER_CPU(struct sched_domain *, sd_llc); 746 DECLARE_PER_CPU(int, sd_llc_size); 747 DECLARE_PER_CPU(int, sd_llc_id); 748 DECLARE_PER_CPU(struct sched_domain *, sd_numa); 749 DECLARE_PER_CPU(struct sched_domain *, sd_busy); 750 DECLARE_PER_CPU(struct sched_domain *, sd_asym); 751 752 struct sched_group_power { 753 atomic_t ref; 754 /* 755 * CPU power of this group, SCHED_LOAD_SCALE being max power for a 756 * single CPU. 757 */ 758 unsigned int power, power_orig; 759 unsigned long next_update; 760 int imbalance; /* XXX unrelated to power but shared group state */ 761 /* 762 * Number of busy cpus in this group. 763 */ 764 atomic_t nr_busy_cpus; 765 766 unsigned long cpumask[0]; /* iteration mask */ 767 }; 768 769 struct sched_group { 770 struct sched_group *next; /* Must be a circular list */ 771 atomic_t ref; 772 773 unsigned int group_weight; 774 struct sched_group_power *sgp; 775 776 /* 777 * The CPUs this group covers. 778 * 779 * NOTE: this field is variable length. (Allocated dynamically 780 * by attaching extra space to the end of the structure, 781 * depending on how many CPUs the kernel has booted up with) 782 */ 783 unsigned long cpumask[0]; 784 }; 785 786 static inline struct cpumask *sched_group_cpus(struct sched_group *sg) 787 { 788 return to_cpumask(sg->cpumask); 789 } 790 791 /* 792 * cpumask masking which cpus in the group are allowed to iterate up the domain 793 * tree. 794 */ 795 static inline struct cpumask *sched_group_mask(struct sched_group *sg) 796 { 797 return to_cpumask(sg->sgp->cpumask); 798 } 799 800 /** 801 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. 802 * @group: The group whose first cpu is to be returned. 803 */ 804 static inline unsigned int group_first_cpu(struct sched_group *group) 805 { 806 return cpumask_first(sched_group_cpus(group)); 807 } 808 809 extern int group_balance_cpu(struct sched_group *sg); 810 811 #endif /* CONFIG_SMP */ 812 813 #include "stats.h" 814 #include "auto_group.h" 815 816 #ifdef CONFIG_CGROUP_SCHED 817 818 /* 819 * Return the group to which this tasks belongs. 820 * 821 * We cannot use task_css() and friends because the cgroup subsystem 822 * changes that value before the cgroup_subsys::attach() method is called, 823 * therefore we cannot pin it and might observe the wrong value. 824 * 825 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup 826 * core changes this before calling sched_move_task(). 827 * 828 * Instead we use a 'copy' which is updated from sched_move_task() while 829 * holding both task_struct::pi_lock and rq::lock. 830 */ 831 static inline struct task_group *task_group(struct task_struct *p) 832 { 833 return p->sched_task_group; 834 } 835 836 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 837 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 838 { 839 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 840 struct task_group *tg = task_group(p); 841 #endif 842 843 #ifdef CONFIG_FAIR_GROUP_SCHED 844 p->se.cfs_rq = tg->cfs_rq[cpu]; 845 p->se.parent = tg->se[cpu]; 846 #endif 847 848 #ifdef CONFIG_RT_GROUP_SCHED 849 p->rt.rt_rq = tg->rt_rq[cpu]; 850 p->rt.parent = tg->rt_se[cpu]; 851 #endif 852 } 853 854 #else /* CONFIG_CGROUP_SCHED */ 855 856 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 857 static inline struct task_group *task_group(struct task_struct *p) 858 { 859 return NULL; 860 } 861 862 #endif /* CONFIG_CGROUP_SCHED */ 863 864 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 865 { 866 set_task_rq(p, cpu); 867 #ifdef CONFIG_SMP 868 /* 869 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 870 * successfuly executed on another CPU. We must ensure that updates of 871 * per-task data have been completed by this moment. 872 */ 873 smp_wmb(); 874 task_thread_info(p)->cpu = cpu; 875 p->wake_cpu = cpu; 876 #endif 877 } 878 879 /* 880 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 881 */ 882 #ifdef CONFIG_SCHED_DEBUG 883 # include <linux/static_key.h> 884 # define const_debug __read_mostly 885 #else 886 # define const_debug const 887 #endif 888 889 extern const_debug unsigned int sysctl_sched_features; 890 891 #define SCHED_FEAT(name, enabled) \ 892 __SCHED_FEAT_##name , 893 894 enum { 895 #include "features.h" 896 __SCHED_FEAT_NR, 897 }; 898 899 #undef SCHED_FEAT 900 901 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL) 902 static __always_inline bool static_branch__true(struct static_key *key) 903 { 904 return static_key_true(key); /* Not out of line branch. */ 905 } 906 907 static __always_inline bool static_branch__false(struct static_key *key) 908 { 909 return static_key_false(key); /* Out of line branch. */ 910 } 911 912 #define SCHED_FEAT(name, enabled) \ 913 static __always_inline bool static_branch_##name(struct static_key *key) \ 914 { \ 915 return static_branch__##enabled(key); \ 916 } 917 918 #include "features.h" 919 920 #undef SCHED_FEAT 921 922 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 923 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 924 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */ 925 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 926 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */ 927 928 #ifdef CONFIG_NUMA_BALANCING 929 #define sched_feat_numa(x) sched_feat(x) 930 #ifdef CONFIG_SCHED_DEBUG 931 #define numabalancing_enabled sched_feat_numa(NUMA) 932 #else 933 extern bool numabalancing_enabled; 934 #endif /* CONFIG_SCHED_DEBUG */ 935 #else 936 #define sched_feat_numa(x) (0) 937 #define numabalancing_enabled (0) 938 #endif /* CONFIG_NUMA_BALANCING */ 939 940 static inline u64 global_rt_period(void) 941 { 942 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 943 } 944 945 static inline u64 global_rt_runtime(void) 946 { 947 if (sysctl_sched_rt_runtime < 0) 948 return RUNTIME_INF; 949 950 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 951 } 952 953 static inline int task_current(struct rq *rq, struct task_struct *p) 954 { 955 return rq->curr == p; 956 } 957 958 static inline int task_running(struct rq *rq, struct task_struct *p) 959 { 960 #ifdef CONFIG_SMP 961 return p->on_cpu; 962 #else 963 return task_current(rq, p); 964 #endif 965 } 966 967 968 #ifndef prepare_arch_switch 969 # define prepare_arch_switch(next) do { } while (0) 970 #endif 971 #ifndef finish_arch_switch 972 # define finish_arch_switch(prev) do { } while (0) 973 #endif 974 #ifndef finish_arch_post_lock_switch 975 # define finish_arch_post_lock_switch() do { } while (0) 976 #endif 977 978 #ifndef __ARCH_WANT_UNLOCKED_CTXSW 979 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 980 { 981 #ifdef CONFIG_SMP 982 /* 983 * We can optimise this out completely for !SMP, because the 984 * SMP rebalancing from interrupt is the only thing that cares 985 * here. 986 */ 987 next->on_cpu = 1; 988 #endif 989 } 990 991 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 992 { 993 #ifdef CONFIG_SMP 994 /* 995 * After ->on_cpu is cleared, the task can be moved to a different CPU. 996 * We must ensure this doesn't happen until the switch is completely 997 * finished. 998 */ 999 smp_wmb(); 1000 prev->on_cpu = 0; 1001 #endif 1002 #ifdef CONFIG_DEBUG_SPINLOCK 1003 /* this is a valid case when another task releases the spinlock */ 1004 rq->lock.owner = current; 1005 #endif 1006 /* 1007 * If we are tracking spinlock dependencies then we have to 1008 * fix up the runqueue lock - which gets 'carried over' from 1009 * prev into current: 1010 */ 1011 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); 1012 1013 raw_spin_unlock_irq(&rq->lock); 1014 } 1015 1016 #else /* __ARCH_WANT_UNLOCKED_CTXSW */ 1017 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 1018 { 1019 #ifdef CONFIG_SMP 1020 /* 1021 * We can optimise this out completely for !SMP, because the 1022 * SMP rebalancing from interrupt is the only thing that cares 1023 * here. 1024 */ 1025 next->on_cpu = 1; 1026 #endif 1027 raw_spin_unlock(&rq->lock); 1028 } 1029 1030 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 1031 { 1032 #ifdef CONFIG_SMP 1033 /* 1034 * After ->on_cpu is cleared, the task can be moved to a different CPU. 1035 * We must ensure this doesn't happen until the switch is completely 1036 * finished. 1037 */ 1038 smp_wmb(); 1039 prev->on_cpu = 0; 1040 #endif 1041 local_irq_enable(); 1042 } 1043 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ 1044 1045 /* 1046 * wake flags 1047 */ 1048 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */ 1049 #define WF_FORK 0x02 /* child wakeup after fork */ 1050 #define WF_MIGRATED 0x4 /* internal use, task got migrated */ 1051 1052 /* 1053 * To aid in avoiding the subversion of "niceness" due to uneven distribution 1054 * of tasks with abnormal "nice" values across CPUs the contribution that 1055 * each task makes to its run queue's load is weighted according to its 1056 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 1057 * scaled version of the new time slice allocation that they receive on time 1058 * slice expiry etc. 1059 */ 1060 1061 #define WEIGHT_IDLEPRIO 3 1062 #define WMULT_IDLEPRIO 1431655765 1063 1064 /* 1065 * Nice levels are multiplicative, with a gentle 10% change for every 1066 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to 1067 * nice 1, it will get ~10% less CPU time than another CPU-bound task 1068 * that remained on nice 0. 1069 * 1070 * The "10% effect" is relative and cumulative: from _any_ nice level, 1071 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level 1072 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. 1073 * If a task goes up by ~10% and another task goes down by ~10% then 1074 * the relative distance between them is ~25%.) 1075 */ 1076 static const int prio_to_weight[40] = { 1077 /* -20 */ 88761, 71755, 56483, 46273, 36291, 1078 /* -15 */ 29154, 23254, 18705, 14949, 11916, 1079 /* -10 */ 9548, 7620, 6100, 4904, 3906, 1080 /* -5 */ 3121, 2501, 1991, 1586, 1277, 1081 /* 0 */ 1024, 820, 655, 526, 423, 1082 /* 5 */ 335, 272, 215, 172, 137, 1083 /* 10 */ 110, 87, 70, 56, 45, 1084 /* 15 */ 36, 29, 23, 18, 15, 1085 }; 1086 1087 /* 1088 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. 1089 * 1090 * In cases where the weight does not change often, we can use the 1091 * precalculated inverse to speed up arithmetics by turning divisions 1092 * into multiplications: 1093 */ 1094 static const u32 prio_to_wmult[40] = { 1095 /* -20 */ 48388, 59856, 76040, 92818, 118348, 1096 /* -15 */ 147320, 184698, 229616, 287308, 360437, 1097 /* -10 */ 449829, 563644, 704093, 875809, 1099582, 1098 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, 1099 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, 1100 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, 1101 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, 1102 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, 1103 }; 1104 1105 #define ENQUEUE_WAKEUP 1 1106 #define ENQUEUE_HEAD 2 1107 #ifdef CONFIG_SMP 1108 #define ENQUEUE_WAKING 4 /* sched_class::task_waking was called */ 1109 #else 1110 #define ENQUEUE_WAKING 0 1111 #endif 1112 #define ENQUEUE_REPLENISH 8 1113 1114 #define DEQUEUE_SLEEP 1 1115 1116 struct sched_class { 1117 const struct sched_class *next; 1118 1119 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags); 1120 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags); 1121 void (*yield_task) (struct rq *rq); 1122 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt); 1123 1124 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags); 1125 1126 struct task_struct * (*pick_next_task) (struct rq *rq); 1127 void (*put_prev_task) (struct rq *rq, struct task_struct *p); 1128 1129 #ifdef CONFIG_SMP 1130 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags); 1131 void (*migrate_task_rq)(struct task_struct *p, int next_cpu); 1132 1133 void (*pre_schedule) (struct rq *this_rq, struct task_struct *task); 1134 void (*post_schedule) (struct rq *this_rq); 1135 void (*task_waking) (struct task_struct *task); 1136 void (*task_woken) (struct rq *this_rq, struct task_struct *task); 1137 1138 void (*set_cpus_allowed)(struct task_struct *p, 1139 const struct cpumask *newmask); 1140 1141 void (*rq_online)(struct rq *rq); 1142 void (*rq_offline)(struct rq *rq); 1143 #endif 1144 1145 void (*set_curr_task) (struct rq *rq); 1146 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued); 1147 void (*task_fork) (struct task_struct *p); 1148 void (*task_dead) (struct task_struct *p); 1149 1150 void (*switched_from) (struct rq *this_rq, struct task_struct *task); 1151 void (*switched_to) (struct rq *this_rq, struct task_struct *task); 1152 void (*prio_changed) (struct rq *this_rq, struct task_struct *task, 1153 int oldprio); 1154 1155 unsigned int (*get_rr_interval) (struct rq *rq, 1156 struct task_struct *task); 1157 1158 #ifdef CONFIG_FAIR_GROUP_SCHED 1159 void (*task_move_group) (struct task_struct *p, int on_rq); 1160 #endif 1161 }; 1162 1163 #define sched_class_highest (&stop_sched_class) 1164 #define for_each_class(class) \ 1165 for (class = sched_class_highest; class; class = class->next) 1166 1167 extern const struct sched_class stop_sched_class; 1168 extern const struct sched_class dl_sched_class; 1169 extern const struct sched_class rt_sched_class; 1170 extern const struct sched_class fair_sched_class; 1171 extern const struct sched_class idle_sched_class; 1172 1173 1174 #ifdef CONFIG_SMP 1175 1176 extern void update_group_power(struct sched_domain *sd, int cpu); 1177 1178 extern void trigger_load_balance(struct rq *rq); 1179 extern void idle_balance(int this_cpu, struct rq *this_rq); 1180 1181 extern void idle_enter_fair(struct rq *this_rq); 1182 extern void idle_exit_fair(struct rq *this_rq); 1183 1184 #else /* CONFIG_SMP */ 1185 1186 static inline void idle_balance(int cpu, struct rq *rq) 1187 { 1188 } 1189 1190 #endif 1191 1192 extern void sysrq_sched_debug_show(void); 1193 extern void sched_init_granularity(void); 1194 extern void update_max_interval(void); 1195 1196 extern void init_sched_dl_class(void); 1197 extern void init_sched_rt_class(void); 1198 extern void init_sched_fair_class(void); 1199 extern void init_sched_dl_class(void); 1200 1201 extern void resched_task(struct task_struct *p); 1202 extern void resched_cpu(int cpu); 1203 1204 extern struct rt_bandwidth def_rt_bandwidth; 1205 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 1206 1207 extern struct dl_bandwidth def_dl_bandwidth; 1208 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime); 1209 extern void init_dl_task_timer(struct sched_dl_entity *dl_se); 1210 1211 unsigned long to_ratio(u64 period, u64 runtime); 1212 1213 extern void update_idle_cpu_load(struct rq *this_rq); 1214 1215 extern void init_task_runnable_average(struct task_struct *p); 1216 1217 #ifdef CONFIG_PARAVIRT 1218 static inline u64 steal_ticks(u64 steal) 1219 { 1220 if (unlikely(steal > NSEC_PER_SEC)) 1221 return div_u64(steal, TICK_NSEC); 1222 1223 return __iter_div_u64_rem(steal, TICK_NSEC, &steal); 1224 } 1225 #endif 1226 1227 static inline void inc_nr_running(struct rq *rq) 1228 { 1229 rq->nr_running++; 1230 1231 #ifdef CONFIG_NO_HZ_FULL 1232 if (rq->nr_running == 2) { 1233 if (tick_nohz_full_cpu(rq->cpu)) { 1234 /* Order rq->nr_running write against the IPI */ 1235 smp_wmb(); 1236 smp_send_reschedule(rq->cpu); 1237 } 1238 } 1239 #endif 1240 } 1241 1242 static inline void dec_nr_running(struct rq *rq) 1243 { 1244 rq->nr_running--; 1245 } 1246 1247 static inline void rq_last_tick_reset(struct rq *rq) 1248 { 1249 #ifdef CONFIG_NO_HZ_FULL 1250 rq->last_sched_tick = jiffies; 1251 #endif 1252 } 1253 1254 extern void update_rq_clock(struct rq *rq); 1255 1256 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 1257 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 1258 1259 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); 1260 1261 extern const_debug unsigned int sysctl_sched_time_avg; 1262 extern const_debug unsigned int sysctl_sched_nr_migrate; 1263 extern const_debug unsigned int sysctl_sched_migration_cost; 1264 1265 static inline u64 sched_avg_period(void) 1266 { 1267 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; 1268 } 1269 1270 #ifdef CONFIG_SCHED_HRTICK 1271 1272 /* 1273 * Use hrtick when: 1274 * - enabled by features 1275 * - hrtimer is actually high res 1276 */ 1277 static inline int hrtick_enabled(struct rq *rq) 1278 { 1279 if (!sched_feat(HRTICK)) 1280 return 0; 1281 if (!cpu_active(cpu_of(rq))) 1282 return 0; 1283 return hrtimer_is_hres_active(&rq->hrtick_timer); 1284 } 1285 1286 void hrtick_start(struct rq *rq, u64 delay); 1287 1288 #else 1289 1290 static inline int hrtick_enabled(struct rq *rq) 1291 { 1292 return 0; 1293 } 1294 1295 #endif /* CONFIG_SCHED_HRTICK */ 1296 1297 #ifdef CONFIG_SMP 1298 extern void sched_avg_update(struct rq *rq); 1299 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) 1300 { 1301 rq->rt_avg += rt_delta; 1302 sched_avg_update(rq); 1303 } 1304 #else 1305 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { } 1306 static inline void sched_avg_update(struct rq *rq) { } 1307 #endif 1308 1309 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period); 1310 1311 #ifdef CONFIG_SMP 1312 #ifdef CONFIG_PREEMPT 1313 1314 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2); 1315 1316 /* 1317 * fair double_lock_balance: Safely acquires both rq->locks in a fair 1318 * way at the expense of forcing extra atomic operations in all 1319 * invocations. This assures that the double_lock is acquired using the 1320 * same underlying policy as the spinlock_t on this architecture, which 1321 * reduces latency compared to the unfair variant below. However, it 1322 * also adds more overhead and therefore may reduce throughput. 1323 */ 1324 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1325 __releases(this_rq->lock) 1326 __acquires(busiest->lock) 1327 __acquires(this_rq->lock) 1328 { 1329 raw_spin_unlock(&this_rq->lock); 1330 double_rq_lock(this_rq, busiest); 1331 1332 return 1; 1333 } 1334 1335 #else 1336 /* 1337 * Unfair double_lock_balance: Optimizes throughput at the expense of 1338 * latency by eliminating extra atomic operations when the locks are 1339 * already in proper order on entry. This favors lower cpu-ids and will 1340 * grant the double lock to lower cpus over higher ids under contention, 1341 * regardless of entry order into the function. 1342 */ 1343 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1344 __releases(this_rq->lock) 1345 __acquires(busiest->lock) 1346 __acquires(this_rq->lock) 1347 { 1348 int ret = 0; 1349 1350 if (unlikely(!raw_spin_trylock(&busiest->lock))) { 1351 if (busiest < this_rq) { 1352 raw_spin_unlock(&this_rq->lock); 1353 raw_spin_lock(&busiest->lock); 1354 raw_spin_lock_nested(&this_rq->lock, 1355 SINGLE_DEPTH_NESTING); 1356 ret = 1; 1357 } else 1358 raw_spin_lock_nested(&busiest->lock, 1359 SINGLE_DEPTH_NESTING); 1360 } 1361 return ret; 1362 } 1363 1364 #endif /* CONFIG_PREEMPT */ 1365 1366 /* 1367 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 1368 */ 1369 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 1370 { 1371 if (unlikely(!irqs_disabled())) { 1372 /* printk() doesn't work good under rq->lock */ 1373 raw_spin_unlock(&this_rq->lock); 1374 BUG_ON(1); 1375 } 1376 1377 return _double_lock_balance(this_rq, busiest); 1378 } 1379 1380 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 1381 __releases(busiest->lock) 1382 { 1383 raw_spin_unlock(&busiest->lock); 1384 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1385 } 1386 1387 static inline void double_lock(spinlock_t *l1, spinlock_t *l2) 1388 { 1389 if (l1 > l2) 1390 swap(l1, l2); 1391 1392 spin_lock(l1); 1393 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1394 } 1395 1396 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2) 1397 { 1398 if (l1 > l2) 1399 swap(l1, l2); 1400 1401 raw_spin_lock(l1); 1402 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1403 } 1404 1405 /* 1406 * double_rq_lock - safely lock two runqueues 1407 * 1408 * Note this does not disable interrupts like task_rq_lock, 1409 * you need to do so manually before calling. 1410 */ 1411 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1412 __acquires(rq1->lock) 1413 __acquires(rq2->lock) 1414 { 1415 BUG_ON(!irqs_disabled()); 1416 if (rq1 == rq2) { 1417 raw_spin_lock(&rq1->lock); 1418 __acquire(rq2->lock); /* Fake it out ;) */ 1419 } else { 1420 if (rq1 < rq2) { 1421 raw_spin_lock(&rq1->lock); 1422 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); 1423 } else { 1424 raw_spin_lock(&rq2->lock); 1425 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); 1426 } 1427 } 1428 } 1429 1430 /* 1431 * double_rq_unlock - safely unlock two runqueues 1432 * 1433 * Note this does not restore interrupts like task_rq_unlock, 1434 * you need to do so manually after calling. 1435 */ 1436 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1437 __releases(rq1->lock) 1438 __releases(rq2->lock) 1439 { 1440 raw_spin_unlock(&rq1->lock); 1441 if (rq1 != rq2) 1442 raw_spin_unlock(&rq2->lock); 1443 else 1444 __release(rq2->lock); 1445 } 1446 1447 #else /* CONFIG_SMP */ 1448 1449 /* 1450 * double_rq_lock - safely lock two runqueues 1451 * 1452 * Note this does not disable interrupts like task_rq_lock, 1453 * you need to do so manually before calling. 1454 */ 1455 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1456 __acquires(rq1->lock) 1457 __acquires(rq2->lock) 1458 { 1459 BUG_ON(!irqs_disabled()); 1460 BUG_ON(rq1 != rq2); 1461 raw_spin_lock(&rq1->lock); 1462 __acquire(rq2->lock); /* Fake it out ;) */ 1463 } 1464 1465 /* 1466 * double_rq_unlock - safely unlock two runqueues 1467 * 1468 * Note this does not restore interrupts like task_rq_unlock, 1469 * you need to do so manually after calling. 1470 */ 1471 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1472 __releases(rq1->lock) 1473 __releases(rq2->lock) 1474 { 1475 BUG_ON(rq1 != rq2); 1476 raw_spin_unlock(&rq1->lock); 1477 __release(rq2->lock); 1478 } 1479 1480 #endif 1481 1482 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 1483 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 1484 extern void print_cfs_stats(struct seq_file *m, int cpu); 1485 extern void print_rt_stats(struct seq_file *m, int cpu); 1486 1487 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 1488 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq); 1489 extern void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq); 1490 1491 extern void cfs_bandwidth_usage_inc(void); 1492 extern void cfs_bandwidth_usage_dec(void); 1493 1494 #ifdef CONFIG_NO_HZ_COMMON 1495 enum rq_nohz_flag_bits { 1496 NOHZ_TICK_STOPPED, 1497 NOHZ_BALANCE_KICK, 1498 }; 1499 1500 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 1501 #endif 1502 1503 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1504 1505 DECLARE_PER_CPU(u64, cpu_hardirq_time); 1506 DECLARE_PER_CPU(u64, cpu_softirq_time); 1507 1508 #ifndef CONFIG_64BIT 1509 DECLARE_PER_CPU(seqcount_t, irq_time_seq); 1510 1511 static inline void irq_time_write_begin(void) 1512 { 1513 __this_cpu_inc(irq_time_seq.sequence); 1514 smp_wmb(); 1515 } 1516 1517 static inline void irq_time_write_end(void) 1518 { 1519 smp_wmb(); 1520 __this_cpu_inc(irq_time_seq.sequence); 1521 } 1522 1523 static inline u64 irq_time_read(int cpu) 1524 { 1525 u64 irq_time; 1526 unsigned seq; 1527 1528 do { 1529 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 1530 irq_time = per_cpu(cpu_softirq_time, cpu) + 1531 per_cpu(cpu_hardirq_time, cpu); 1532 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 1533 1534 return irq_time; 1535 } 1536 #else /* CONFIG_64BIT */ 1537 static inline void irq_time_write_begin(void) 1538 { 1539 } 1540 1541 static inline void irq_time_write_end(void) 1542 { 1543 } 1544 1545 static inline u64 irq_time_read(int cpu) 1546 { 1547 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 1548 } 1549 #endif /* CONFIG_64BIT */ 1550 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 1551