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