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/irq_work.h> 10 #include <linux/tick.h> 11 #include <linux/slab.h> 12 13 #include "cpupri.h" 14 #include "cpudeadline.h" 15 #include "cpuacct.h" 16 17 struct rq; 18 struct cpuidle_state; 19 20 /* task_struct::on_rq states: */ 21 #define TASK_ON_RQ_QUEUED 1 22 #define TASK_ON_RQ_MIGRATING 2 23 24 extern __read_mostly int scheduler_running; 25 26 extern unsigned long calc_load_update; 27 extern atomic_long_t calc_load_tasks; 28 29 extern void calc_global_load_tick(struct rq *this_rq); 30 extern long calc_load_fold_active(struct rq *this_rq); 31 32 #ifdef CONFIG_SMP 33 extern void update_cpu_load_active(struct rq *this_rq); 34 #else 35 static inline void update_cpu_load_active(struct rq *this_rq) { } 36 #endif 37 38 /* 39 * Helpers for converting nanosecond timing to jiffy resolution 40 */ 41 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) 42 43 /* 44 * Increase resolution of nice-level calculations for 64-bit architectures. 45 * The extra resolution improves shares distribution and load balancing of 46 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup 47 * hierarchies, especially on larger systems. This is not a user-visible change 48 * and does not change the user-interface for setting shares/weights. 49 * 50 * We increase resolution only if we have enough bits to allow this increased 51 * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution 52 * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the 53 * increased costs. 54 */ 55 #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load */ 56 # define SCHED_LOAD_RESOLUTION 10 57 # define scale_load(w) ((w) << SCHED_LOAD_RESOLUTION) 58 # define scale_load_down(w) ((w) >> SCHED_LOAD_RESOLUTION) 59 #else 60 # define SCHED_LOAD_RESOLUTION 0 61 # define scale_load(w) (w) 62 # define scale_load_down(w) (w) 63 #endif 64 65 #define SCHED_LOAD_SHIFT (10 + SCHED_LOAD_RESOLUTION) 66 #define SCHED_LOAD_SCALE (1L << SCHED_LOAD_SHIFT) 67 68 #define NICE_0_LOAD SCHED_LOAD_SCALE 69 #define NICE_0_SHIFT SCHED_LOAD_SHIFT 70 71 /* 72 * Single value that decides SCHED_DEADLINE internal math precision. 73 * 10 -> just above 1us 74 * 9 -> just above 0.5us 75 */ 76 #define DL_SCALE (10) 77 78 /* 79 * These are the 'tuning knobs' of the scheduler: 80 */ 81 82 /* 83 * single value that denotes runtime == period, ie unlimited time. 84 */ 85 #define RUNTIME_INF ((u64)~0ULL) 86 87 static inline int idle_policy(int policy) 88 { 89 return policy == SCHED_IDLE; 90 } 91 static inline int fair_policy(int policy) 92 { 93 return policy == SCHED_NORMAL || policy == SCHED_BATCH; 94 } 95 96 static inline int rt_policy(int policy) 97 { 98 return policy == SCHED_FIFO || policy == SCHED_RR; 99 } 100 101 static inline int dl_policy(int policy) 102 { 103 return policy == SCHED_DEADLINE; 104 } 105 static inline bool valid_policy(int policy) 106 { 107 return idle_policy(policy) || fair_policy(policy) || 108 rt_policy(policy) || dl_policy(policy); 109 } 110 111 static inline int task_has_rt_policy(struct task_struct *p) 112 { 113 return rt_policy(p->policy); 114 } 115 116 static inline int task_has_dl_policy(struct task_struct *p) 117 { 118 return dl_policy(p->policy); 119 } 120 121 /* 122 * Tells if entity @a should preempt entity @b. 123 */ 124 static inline bool 125 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b) 126 { 127 return dl_time_before(a->deadline, b->deadline); 128 } 129 130 /* 131 * This is the priority-queue data structure of the RT scheduling class: 132 */ 133 struct rt_prio_array { 134 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ 135 struct list_head queue[MAX_RT_PRIO]; 136 }; 137 138 struct rt_bandwidth { 139 /* nests inside the rq lock: */ 140 raw_spinlock_t rt_runtime_lock; 141 ktime_t rt_period; 142 u64 rt_runtime; 143 struct hrtimer rt_period_timer; 144 unsigned int rt_period_active; 145 }; 146 147 void __dl_clear_params(struct task_struct *p); 148 149 /* 150 * To keep the bandwidth of -deadline tasks and groups under control 151 * we need some place where: 152 * - store the maximum -deadline bandwidth of the system (the group); 153 * - cache the fraction of that bandwidth that is currently allocated. 154 * 155 * This is all done in the data structure below. It is similar to the 156 * one used for RT-throttling (rt_bandwidth), with the main difference 157 * that, since here we are only interested in admission control, we 158 * do not decrease any runtime while the group "executes", neither we 159 * need a timer to replenish it. 160 * 161 * With respect to SMP, the bandwidth is given on a per-CPU basis, 162 * meaning that: 163 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU; 164 * - dl_total_bw array contains, in the i-eth element, the currently 165 * allocated bandwidth on the i-eth CPU. 166 * Moreover, groups consume bandwidth on each CPU, while tasks only 167 * consume bandwidth on the CPU they're running on. 168 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw 169 * that will be shown the next time the proc or cgroup controls will 170 * be red. It on its turn can be changed by writing on its own 171 * control. 172 */ 173 struct dl_bandwidth { 174 raw_spinlock_t dl_runtime_lock; 175 u64 dl_runtime; 176 u64 dl_period; 177 }; 178 179 static inline int dl_bandwidth_enabled(void) 180 { 181 return sysctl_sched_rt_runtime >= 0; 182 } 183 184 extern struct dl_bw *dl_bw_of(int i); 185 186 struct dl_bw { 187 raw_spinlock_t lock; 188 u64 bw, total_bw; 189 }; 190 191 static inline 192 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 193 { 194 dl_b->total_bw -= tsk_bw; 195 } 196 197 static inline 198 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 199 { 200 dl_b->total_bw += tsk_bw; 201 } 202 203 static inline 204 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 205 { 206 return dl_b->bw != -1 && 207 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 208 } 209 210 extern struct mutex sched_domains_mutex; 211 212 #ifdef CONFIG_CGROUP_SCHED 213 214 #include <linux/cgroup.h> 215 216 struct cfs_rq; 217 struct rt_rq; 218 219 extern struct list_head task_groups; 220 221 struct cfs_bandwidth { 222 #ifdef CONFIG_CFS_BANDWIDTH 223 raw_spinlock_t lock; 224 ktime_t period; 225 u64 quota, runtime; 226 s64 hierarchical_quota; 227 u64 runtime_expires; 228 229 int idle, period_active; 230 struct hrtimer period_timer, slack_timer; 231 struct list_head throttled_cfs_rq; 232 233 /* statistics */ 234 int nr_periods, nr_throttled; 235 u64 throttled_time; 236 #endif 237 }; 238 239 /* task group related information */ 240 struct task_group { 241 struct cgroup_subsys_state css; 242 243 #ifdef CONFIG_FAIR_GROUP_SCHED 244 /* schedulable entities of this group on each cpu */ 245 struct sched_entity **se; 246 /* runqueue "owned" by this group on each cpu */ 247 struct cfs_rq **cfs_rq; 248 unsigned long shares; 249 250 #ifdef CONFIG_SMP 251 /* 252 * load_avg can be heavily contended at clock tick time, so put 253 * it in its own cacheline separated from the fields above which 254 * will also be accessed at each tick. 255 */ 256 atomic_long_t load_avg ____cacheline_aligned; 257 #endif 258 #endif 259 260 #ifdef CONFIG_RT_GROUP_SCHED 261 struct sched_rt_entity **rt_se; 262 struct rt_rq **rt_rq; 263 264 struct rt_bandwidth rt_bandwidth; 265 #endif 266 267 struct rcu_head rcu; 268 struct list_head list; 269 270 struct task_group *parent; 271 struct list_head siblings; 272 struct list_head children; 273 274 #ifdef CONFIG_SCHED_AUTOGROUP 275 struct autogroup *autogroup; 276 #endif 277 278 struct cfs_bandwidth cfs_bandwidth; 279 }; 280 281 #ifdef CONFIG_FAIR_GROUP_SCHED 282 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD 283 284 /* 285 * A weight of 0 or 1 can cause arithmetics problems. 286 * A weight of a cfs_rq is the sum of weights of which entities 287 * are queued on this cfs_rq, so a weight of a entity should not be 288 * too large, so as the shares value of a task group. 289 * (The default weight is 1024 - so there's no practical 290 * limitation from this.) 291 */ 292 #define MIN_SHARES (1UL << 1) 293 #define MAX_SHARES (1UL << 18) 294 #endif 295 296 typedef int (*tg_visitor)(struct task_group *, void *); 297 298 extern int walk_tg_tree_from(struct task_group *from, 299 tg_visitor down, tg_visitor up, void *data); 300 301 /* 302 * Iterate the full tree, calling @down when first entering a node and @up when 303 * leaving it for the final time. 304 * 305 * Caller must hold rcu_lock or sufficient equivalent. 306 */ 307 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) 308 { 309 return walk_tg_tree_from(&root_task_group, down, up, data); 310 } 311 312 extern int tg_nop(struct task_group *tg, void *data); 313 314 extern void free_fair_sched_group(struct task_group *tg); 315 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent); 316 extern void unregister_fair_sched_group(struct task_group *tg, int cpu); 317 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 318 struct sched_entity *se, int cpu, 319 struct sched_entity *parent); 320 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 321 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 322 323 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b); 324 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 325 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq); 326 327 extern void free_rt_sched_group(struct task_group *tg); 328 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent); 329 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 330 struct sched_rt_entity *rt_se, int cpu, 331 struct sched_rt_entity *parent); 332 333 extern struct task_group *sched_create_group(struct task_group *parent); 334 extern void sched_online_group(struct task_group *tg, 335 struct task_group *parent); 336 extern void sched_destroy_group(struct task_group *tg); 337 extern void sched_offline_group(struct task_group *tg); 338 339 extern void sched_move_task(struct task_struct *tsk); 340 341 #ifdef CONFIG_FAIR_GROUP_SCHED 342 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 343 344 #ifdef CONFIG_SMP 345 extern void set_task_rq_fair(struct sched_entity *se, 346 struct cfs_rq *prev, struct cfs_rq *next); 347 #else /* !CONFIG_SMP */ 348 static inline void set_task_rq_fair(struct sched_entity *se, 349 struct cfs_rq *prev, struct cfs_rq *next) { } 350 #endif /* CONFIG_SMP */ 351 #endif /* CONFIG_FAIR_GROUP_SCHED */ 352 353 #else /* CONFIG_CGROUP_SCHED */ 354 355 struct cfs_bandwidth { }; 356 357 #endif /* CONFIG_CGROUP_SCHED */ 358 359 /* CFS-related fields in a runqueue */ 360 struct cfs_rq { 361 struct load_weight load; 362 unsigned int nr_running, h_nr_running; 363 364 u64 exec_clock; 365 u64 min_vruntime; 366 #ifndef CONFIG_64BIT 367 u64 min_vruntime_copy; 368 #endif 369 370 struct rb_root tasks_timeline; 371 struct rb_node *rb_leftmost; 372 373 /* 374 * 'curr' points to currently running entity on this cfs_rq. 375 * It is set to NULL otherwise (i.e when none are currently running). 376 */ 377 struct sched_entity *curr, *next, *last, *skip; 378 379 #ifdef CONFIG_SCHED_DEBUG 380 unsigned int nr_spread_over; 381 #endif 382 383 #ifdef CONFIG_SMP 384 /* 385 * CFS load tracking 386 */ 387 struct sched_avg avg; 388 u64 runnable_load_sum; 389 unsigned long runnable_load_avg; 390 #ifdef CONFIG_FAIR_GROUP_SCHED 391 unsigned long tg_load_avg_contrib; 392 #endif 393 atomic_long_t removed_load_avg, removed_util_avg; 394 #ifndef CONFIG_64BIT 395 u64 load_last_update_time_copy; 396 #endif 397 398 #ifdef CONFIG_FAIR_GROUP_SCHED 399 /* 400 * h_load = weight * f(tg) 401 * 402 * Where f(tg) is the recursive weight fraction assigned to 403 * this group. 404 */ 405 unsigned long h_load; 406 u64 last_h_load_update; 407 struct sched_entity *h_load_next; 408 #endif /* CONFIG_FAIR_GROUP_SCHED */ 409 #endif /* CONFIG_SMP */ 410 411 #ifdef CONFIG_FAIR_GROUP_SCHED 412 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ 413 414 /* 415 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in 416 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities 417 * (like users, containers etc.) 418 * 419 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This 420 * list is used during load balance. 421 */ 422 int on_list; 423 struct list_head leaf_cfs_rq_list; 424 struct task_group *tg; /* group that "owns" this runqueue */ 425 426 #ifdef CONFIG_CFS_BANDWIDTH 427 int runtime_enabled; 428 u64 runtime_expires; 429 s64 runtime_remaining; 430 431 u64 throttled_clock, throttled_clock_task; 432 u64 throttled_clock_task_time; 433 int throttled, throttle_count; 434 struct list_head throttled_list; 435 #endif /* CONFIG_CFS_BANDWIDTH */ 436 #endif /* CONFIG_FAIR_GROUP_SCHED */ 437 }; 438 439 static inline int rt_bandwidth_enabled(void) 440 { 441 return sysctl_sched_rt_runtime >= 0; 442 } 443 444 /* RT IPI pull logic requires IRQ_WORK */ 445 #ifdef CONFIG_IRQ_WORK 446 # define HAVE_RT_PUSH_IPI 447 #endif 448 449 /* Real-Time classes' related field in a runqueue: */ 450 struct rt_rq { 451 struct rt_prio_array active; 452 unsigned int rt_nr_running; 453 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 454 struct { 455 int curr; /* highest queued rt task prio */ 456 #ifdef CONFIG_SMP 457 int next; /* next highest */ 458 #endif 459 } highest_prio; 460 #endif 461 #ifdef CONFIG_SMP 462 unsigned long rt_nr_migratory; 463 unsigned long rt_nr_total; 464 int overloaded; 465 struct plist_head pushable_tasks; 466 #ifdef HAVE_RT_PUSH_IPI 467 int push_flags; 468 int push_cpu; 469 struct irq_work push_work; 470 raw_spinlock_t push_lock; 471 #endif 472 #endif /* CONFIG_SMP */ 473 int rt_queued; 474 475 int rt_throttled; 476 u64 rt_time; 477 u64 rt_runtime; 478 /* Nests inside the rq lock: */ 479 raw_spinlock_t rt_runtime_lock; 480 481 #ifdef CONFIG_RT_GROUP_SCHED 482 unsigned long rt_nr_boosted; 483 484 struct rq *rq; 485 struct task_group *tg; 486 #endif 487 }; 488 489 /* Deadline class' related fields in a runqueue */ 490 struct dl_rq { 491 /* runqueue is an rbtree, ordered by deadline */ 492 struct rb_root rb_root; 493 struct rb_node *rb_leftmost; 494 495 unsigned long dl_nr_running; 496 497 #ifdef CONFIG_SMP 498 /* 499 * Deadline values of the currently executing and the 500 * earliest ready task on this rq. Caching these facilitates 501 * the decision wether or not a ready but not running task 502 * should migrate somewhere else. 503 */ 504 struct { 505 u64 curr; 506 u64 next; 507 } earliest_dl; 508 509 unsigned long dl_nr_migratory; 510 int overloaded; 511 512 /* 513 * Tasks on this rq that can be pushed away. They are kept in 514 * an rb-tree, ordered by tasks' deadlines, with caching 515 * of the leftmost (earliest deadline) element. 516 */ 517 struct rb_root pushable_dl_tasks_root; 518 struct rb_node *pushable_dl_tasks_leftmost; 519 #else 520 struct dl_bw dl_bw; 521 #endif 522 }; 523 524 #ifdef CONFIG_SMP 525 526 /* 527 * We add the notion of a root-domain which will be used to define per-domain 528 * variables. Each exclusive cpuset essentially defines an island domain by 529 * fully partitioning the member cpus from any other cpuset. Whenever a new 530 * exclusive cpuset is created, we also create and attach a new root-domain 531 * object. 532 * 533 */ 534 struct root_domain { 535 atomic_t refcount; 536 atomic_t rto_count; 537 struct rcu_head rcu; 538 cpumask_var_t span; 539 cpumask_var_t online; 540 541 /* Indicate more than one runnable task for any CPU */ 542 bool overload; 543 544 /* 545 * The bit corresponding to a CPU gets set here if such CPU has more 546 * than one runnable -deadline task (as it is below for RT tasks). 547 */ 548 cpumask_var_t dlo_mask; 549 atomic_t dlo_count; 550 struct dl_bw dl_bw; 551 struct cpudl cpudl; 552 553 /* 554 * The "RT overload" flag: it gets set if a CPU has more than 555 * one runnable RT task. 556 */ 557 cpumask_var_t rto_mask; 558 struct cpupri cpupri; 559 }; 560 561 extern struct root_domain def_root_domain; 562 563 #endif /* CONFIG_SMP */ 564 565 /* 566 * This is the main, per-CPU runqueue data structure. 567 * 568 * Locking rule: those places that want to lock multiple runqueues 569 * (such as the load balancing or the thread migration code), lock 570 * acquire operations must be ordered by ascending &runqueue. 571 */ 572 struct rq { 573 /* runqueue lock: */ 574 raw_spinlock_t lock; 575 576 /* 577 * nr_running and cpu_load should be in the same cacheline because 578 * remote CPUs use both these fields when doing load calculation. 579 */ 580 unsigned int nr_running; 581 #ifdef CONFIG_NUMA_BALANCING 582 unsigned int nr_numa_running; 583 unsigned int nr_preferred_running; 584 #endif 585 #define CPU_LOAD_IDX_MAX 5 586 unsigned long cpu_load[CPU_LOAD_IDX_MAX]; 587 unsigned long last_load_update_tick; 588 #ifdef CONFIG_NO_HZ_COMMON 589 u64 nohz_stamp; 590 unsigned long nohz_flags; 591 #endif 592 #ifdef CONFIG_NO_HZ_FULL 593 unsigned long last_sched_tick; 594 #endif 595 /* capture load from *all* tasks on this cpu: */ 596 struct load_weight load; 597 unsigned long nr_load_updates; 598 u64 nr_switches; 599 600 struct cfs_rq cfs; 601 struct rt_rq rt; 602 struct dl_rq dl; 603 604 #ifdef CONFIG_FAIR_GROUP_SCHED 605 /* list of leaf cfs_rq on this cpu: */ 606 struct list_head leaf_cfs_rq_list; 607 #endif /* CONFIG_FAIR_GROUP_SCHED */ 608 609 /* 610 * This is part of a global counter where only the total sum 611 * over all CPUs matters. A task can increase this counter on 612 * one CPU and if it got migrated afterwards it may decrease 613 * it on another CPU. Always updated under the runqueue lock: 614 */ 615 unsigned long nr_uninterruptible; 616 617 struct task_struct *curr, *idle, *stop; 618 unsigned long next_balance; 619 struct mm_struct *prev_mm; 620 621 unsigned int clock_skip_update; 622 u64 clock; 623 u64 clock_task; 624 625 atomic_t nr_iowait; 626 627 #ifdef CONFIG_SMP 628 struct root_domain *rd; 629 struct sched_domain *sd; 630 631 unsigned long cpu_capacity; 632 unsigned long cpu_capacity_orig; 633 634 struct callback_head *balance_callback; 635 636 unsigned char idle_balance; 637 /* For active balancing */ 638 int active_balance; 639 int push_cpu; 640 struct cpu_stop_work active_balance_work; 641 /* cpu of this runqueue: */ 642 int cpu; 643 int online; 644 645 struct list_head cfs_tasks; 646 647 u64 rt_avg; 648 u64 age_stamp; 649 u64 idle_stamp; 650 u64 avg_idle; 651 652 /* This is used to determine avg_idle's max value */ 653 u64 max_idle_balance_cost; 654 #endif 655 656 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 657 u64 prev_irq_time; 658 #endif 659 #ifdef CONFIG_PARAVIRT 660 u64 prev_steal_time; 661 #endif 662 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 663 u64 prev_steal_time_rq; 664 #endif 665 666 /* calc_load related fields */ 667 unsigned long calc_load_update; 668 long calc_load_active; 669 670 #ifdef CONFIG_SCHED_HRTICK 671 #ifdef CONFIG_SMP 672 int hrtick_csd_pending; 673 struct call_single_data hrtick_csd; 674 #endif 675 struct hrtimer hrtick_timer; 676 #endif 677 678 #ifdef CONFIG_SCHEDSTATS 679 /* latency stats */ 680 struct sched_info rq_sched_info; 681 unsigned long long rq_cpu_time; 682 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 683 684 /* sys_sched_yield() stats */ 685 unsigned int yld_count; 686 687 /* schedule() stats */ 688 unsigned int sched_count; 689 unsigned int sched_goidle; 690 691 /* try_to_wake_up() stats */ 692 unsigned int ttwu_count; 693 unsigned int ttwu_local; 694 #endif 695 696 #ifdef CONFIG_SMP 697 struct llist_head wake_list; 698 #endif 699 700 #ifdef CONFIG_CPU_IDLE 701 /* Must be inspected within a rcu lock section */ 702 struct cpuidle_state *idle_state; 703 #endif 704 }; 705 706 static inline int cpu_of(struct rq *rq) 707 { 708 #ifdef CONFIG_SMP 709 return rq->cpu; 710 #else 711 return 0; 712 #endif 713 } 714 715 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 716 717 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 718 #define this_rq() this_cpu_ptr(&runqueues) 719 #define task_rq(p) cpu_rq(task_cpu(p)) 720 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 721 #define raw_rq() raw_cpu_ptr(&runqueues) 722 723 static inline u64 __rq_clock_broken(struct rq *rq) 724 { 725 return READ_ONCE(rq->clock); 726 } 727 728 static inline u64 rq_clock(struct rq *rq) 729 { 730 lockdep_assert_held(&rq->lock); 731 return rq->clock; 732 } 733 734 static inline u64 rq_clock_task(struct rq *rq) 735 { 736 lockdep_assert_held(&rq->lock); 737 return rq->clock_task; 738 } 739 740 #define RQCF_REQ_SKIP 0x01 741 #define RQCF_ACT_SKIP 0x02 742 743 static inline void rq_clock_skip_update(struct rq *rq, bool skip) 744 { 745 lockdep_assert_held(&rq->lock); 746 if (skip) 747 rq->clock_skip_update |= RQCF_REQ_SKIP; 748 else 749 rq->clock_skip_update &= ~RQCF_REQ_SKIP; 750 } 751 752 #ifdef CONFIG_NUMA 753 enum numa_topology_type { 754 NUMA_DIRECT, 755 NUMA_GLUELESS_MESH, 756 NUMA_BACKPLANE, 757 }; 758 extern enum numa_topology_type sched_numa_topology_type; 759 extern int sched_max_numa_distance; 760 extern bool find_numa_distance(int distance); 761 #endif 762 763 #ifdef CONFIG_NUMA_BALANCING 764 /* The regions in numa_faults array from task_struct */ 765 enum numa_faults_stats { 766 NUMA_MEM = 0, 767 NUMA_CPU, 768 NUMA_MEMBUF, 769 NUMA_CPUBUF 770 }; 771 extern void sched_setnuma(struct task_struct *p, int node); 772 extern int migrate_task_to(struct task_struct *p, int cpu); 773 extern int migrate_swap(struct task_struct *, struct task_struct *); 774 #endif /* CONFIG_NUMA_BALANCING */ 775 776 #ifdef CONFIG_SMP 777 778 static inline void 779 queue_balance_callback(struct rq *rq, 780 struct callback_head *head, 781 void (*func)(struct rq *rq)) 782 { 783 lockdep_assert_held(&rq->lock); 784 785 if (unlikely(head->next)) 786 return; 787 788 head->func = (void (*)(struct callback_head *))func; 789 head->next = rq->balance_callback; 790 rq->balance_callback = head; 791 } 792 793 extern void sched_ttwu_pending(void); 794 795 #define rcu_dereference_check_sched_domain(p) \ 796 rcu_dereference_check((p), \ 797 lockdep_is_held(&sched_domains_mutex)) 798 799 /* 800 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 801 * See detach_destroy_domains: synchronize_sched for details. 802 * 803 * The domain tree of any CPU may only be accessed from within 804 * preempt-disabled sections. 805 */ 806 #define for_each_domain(cpu, __sd) \ 807 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 808 __sd; __sd = __sd->parent) 809 810 #define for_each_lower_domain(sd) for (; sd; sd = sd->child) 811 812 /** 813 * highest_flag_domain - Return highest sched_domain containing flag. 814 * @cpu: The cpu whose highest level of sched domain is to 815 * be returned. 816 * @flag: The flag to check for the highest sched_domain 817 * for the given cpu. 818 * 819 * Returns the highest sched_domain of a cpu which contains the given flag. 820 */ 821 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 822 { 823 struct sched_domain *sd, *hsd = NULL; 824 825 for_each_domain(cpu, sd) { 826 if (!(sd->flags & flag)) 827 break; 828 hsd = sd; 829 } 830 831 return hsd; 832 } 833 834 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) 835 { 836 struct sched_domain *sd; 837 838 for_each_domain(cpu, sd) { 839 if (sd->flags & flag) 840 break; 841 } 842 843 return sd; 844 } 845 846 DECLARE_PER_CPU(struct sched_domain *, sd_llc); 847 DECLARE_PER_CPU(int, sd_llc_size); 848 DECLARE_PER_CPU(int, sd_llc_id); 849 DECLARE_PER_CPU(struct sched_domain *, sd_numa); 850 DECLARE_PER_CPU(struct sched_domain *, sd_busy); 851 DECLARE_PER_CPU(struct sched_domain *, sd_asym); 852 853 struct sched_group_capacity { 854 atomic_t ref; 855 /* 856 * CPU capacity of this group, SCHED_LOAD_SCALE being max capacity 857 * for a single CPU. 858 */ 859 unsigned int capacity; 860 unsigned long next_update; 861 int imbalance; /* XXX unrelated to capacity but shared group state */ 862 /* 863 * Number of busy cpus in this group. 864 */ 865 atomic_t nr_busy_cpus; 866 867 unsigned long cpumask[0]; /* iteration mask */ 868 }; 869 870 struct sched_group { 871 struct sched_group *next; /* Must be a circular list */ 872 atomic_t ref; 873 874 unsigned int group_weight; 875 struct sched_group_capacity *sgc; 876 877 /* 878 * The CPUs this group covers. 879 * 880 * NOTE: this field is variable length. (Allocated dynamically 881 * by attaching extra space to the end of the structure, 882 * depending on how many CPUs the kernel has booted up with) 883 */ 884 unsigned long cpumask[0]; 885 }; 886 887 static inline struct cpumask *sched_group_cpus(struct sched_group *sg) 888 { 889 return to_cpumask(sg->cpumask); 890 } 891 892 /* 893 * cpumask masking which cpus in the group are allowed to iterate up the domain 894 * tree. 895 */ 896 static inline struct cpumask *sched_group_mask(struct sched_group *sg) 897 { 898 return to_cpumask(sg->sgc->cpumask); 899 } 900 901 /** 902 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. 903 * @group: The group whose first cpu is to be returned. 904 */ 905 static inline unsigned int group_first_cpu(struct sched_group *group) 906 { 907 return cpumask_first(sched_group_cpus(group)); 908 } 909 910 extern int group_balance_cpu(struct sched_group *sg); 911 912 #else 913 914 static inline void sched_ttwu_pending(void) { } 915 916 #endif /* CONFIG_SMP */ 917 918 #include "stats.h" 919 #include "auto_group.h" 920 921 #ifdef CONFIG_CGROUP_SCHED 922 923 /* 924 * Return the group to which this tasks belongs. 925 * 926 * We cannot use task_css() and friends because the cgroup subsystem 927 * changes that value before the cgroup_subsys::attach() method is called, 928 * therefore we cannot pin it and might observe the wrong value. 929 * 930 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup 931 * core changes this before calling sched_move_task(). 932 * 933 * Instead we use a 'copy' which is updated from sched_move_task() while 934 * holding both task_struct::pi_lock and rq::lock. 935 */ 936 static inline struct task_group *task_group(struct task_struct *p) 937 { 938 return p->sched_task_group; 939 } 940 941 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 942 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 943 { 944 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 945 struct task_group *tg = task_group(p); 946 #endif 947 948 #ifdef CONFIG_FAIR_GROUP_SCHED 949 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]); 950 p->se.cfs_rq = tg->cfs_rq[cpu]; 951 p->se.parent = tg->se[cpu]; 952 #endif 953 954 #ifdef CONFIG_RT_GROUP_SCHED 955 p->rt.rt_rq = tg->rt_rq[cpu]; 956 p->rt.parent = tg->rt_se[cpu]; 957 #endif 958 } 959 960 #else /* CONFIG_CGROUP_SCHED */ 961 962 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 963 static inline struct task_group *task_group(struct task_struct *p) 964 { 965 return NULL; 966 } 967 968 #endif /* CONFIG_CGROUP_SCHED */ 969 970 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 971 { 972 set_task_rq(p, cpu); 973 #ifdef CONFIG_SMP 974 /* 975 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 976 * successfuly executed on another CPU. We must ensure that updates of 977 * per-task data have been completed by this moment. 978 */ 979 smp_wmb(); 980 task_thread_info(p)->cpu = cpu; 981 p->wake_cpu = cpu; 982 #endif 983 } 984 985 /* 986 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 987 */ 988 #ifdef CONFIG_SCHED_DEBUG 989 # include <linux/static_key.h> 990 # define const_debug __read_mostly 991 #else 992 # define const_debug const 993 #endif 994 995 extern const_debug unsigned int sysctl_sched_features; 996 997 #define SCHED_FEAT(name, enabled) \ 998 __SCHED_FEAT_##name , 999 1000 enum { 1001 #include "features.h" 1002 __SCHED_FEAT_NR, 1003 }; 1004 1005 #undef SCHED_FEAT 1006 1007 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL) 1008 #define SCHED_FEAT(name, enabled) \ 1009 static __always_inline bool static_branch_##name(struct static_key *key) \ 1010 { \ 1011 return static_key_##enabled(key); \ 1012 } 1013 1014 #include "features.h" 1015 1016 #undef SCHED_FEAT 1017 1018 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 1019 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 1020 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */ 1021 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 1022 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */ 1023 1024 extern struct static_key_false sched_numa_balancing; 1025 1026 static inline u64 global_rt_period(void) 1027 { 1028 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 1029 } 1030 1031 static inline u64 global_rt_runtime(void) 1032 { 1033 if (sysctl_sched_rt_runtime < 0) 1034 return RUNTIME_INF; 1035 1036 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 1037 } 1038 1039 static inline int task_current(struct rq *rq, struct task_struct *p) 1040 { 1041 return rq->curr == p; 1042 } 1043 1044 static inline int task_running(struct rq *rq, struct task_struct *p) 1045 { 1046 #ifdef CONFIG_SMP 1047 return p->on_cpu; 1048 #else 1049 return task_current(rq, p); 1050 #endif 1051 } 1052 1053 static inline int task_on_rq_queued(struct task_struct *p) 1054 { 1055 return p->on_rq == TASK_ON_RQ_QUEUED; 1056 } 1057 1058 static inline int task_on_rq_migrating(struct task_struct *p) 1059 { 1060 return p->on_rq == TASK_ON_RQ_MIGRATING; 1061 } 1062 1063 #ifndef prepare_arch_switch 1064 # define prepare_arch_switch(next) do { } while (0) 1065 #endif 1066 #ifndef finish_arch_post_lock_switch 1067 # define finish_arch_post_lock_switch() do { } while (0) 1068 #endif 1069 1070 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 1071 { 1072 #ifdef CONFIG_SMP 1073 /* 1074 * We can optimise this out completely for !SMP, because the 1075 * SMP rebalancing from interrupt is the only thing that cares 1076 * here. 1077 */ 1078 next->on_cpu = 1; 1079 #endif 1080 } 1081 1082 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 1083 { 1084 #ifdef CONFIG_SMP 1085 /* 1086 * After ->on_cpu is cleared, the task can be moved to a different CPU. 1087 * We must ensure this doesn't happen until the switch is completely 1088 * finished. 1089 * 1090 * In particular, the load of prev->state in finish_task_switch() must 1091 * happen before this. 1092 * 1093 * Pairs with the smp_cond_acquire() in try_to_wake_up(). 1094 */ 1095 smp_store_release(&prev->on_cpu, 0); 1096 #endif 1097 #ifdef CONFIG_DEBUG_SPINLOCK 1098 /* this is a valid case when another task releases the spinlock */ 1099 rq->lock.owner = current; 1100 #endif 1101 /* 1102 * If we are tracking spinlock dependencies then we have to 1103 * fix up the runqueue lock - which gets 'carried over' from 1104 * prev into current: 1105 */ 1106 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); 1107 1108 raw_spin_unlock_irq(&rq->lock); 1109 } 1110 1111 /* 1112 * wake flags 1113 */ 1114 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */ 1115 #define WF_FORK 0x02 /* child wakeup after fork */ 1116 #define WF_MIGRATED 0x4 /* internal use, task got migrated */ 1117 1118 /* 1119 * To aid in avoiding the subversion of "niceness" due to uneven distribution 1120 * of tasks with abnormal "nice" values across CPUs the contribution that 1121 * each task makes to its run queue's load is weighted according to its 1122 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 1123 * scaled version of the new time slice allocation that they receive on time 1124 * slice expiry etc. 1125 */ 1126 1127 #define WEIGHT_IDLEPRIO 3 1128 #define WMULT_IDLEPRIO 1431655765 1129 1130 extern const int sched_prio_to_weight[40]; 1131 extern const u32 sched_prio_to_wmult[40]; 1132 1133 #define ENQUEUE_WAKEUP 0x01 1134 #define ENQUEUE_HEAD 0x02 1135 #ifdef CONFIG_SMP 1136 #define ENQUEUE_WAKING 0x04 /* sched_class::task_waking was called */ 1137 #else 1138 #define ENQUEUE_WAKING 0x00 1139 #endif 1140 #define ENQUEUE_REPLENISH 0x08 1141 #define ENQUEUE_RESTORE 0x10 1142 1143 #define DEQUEUE_SLEEP 0x01 1144 #define DEQUEUE_SAVE 0x02 1145 1146 #define RETRY_TASK ((void *)-1UL) 1147 1148 struct sched_class { 1149 const struct sched_class *next; 1150 1151 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags); 1152 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags); 1153 void (*yield_task) (struct rq *rq); 1154 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt); 1155 1156 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags); 1157 1158 /* 1159 * It is the responsibility of the pick_next_task() method that will 1160 * return the next task to call put_prev_task() on the @prev task or 1161 * something equivalent. 1162 * 1163 * May return RETRY_TASK when it finds a higher prio class has runnable 1164 * tasks. 1165 */ 1166 struct task_struct * (*pick_next_task) (struct rq *rq, 1167 struct task_struct *prev); 1168 void (*put_prev_task) (struct rq *rq, struct task_struct *p); 1169 1170 #ifdef CONFIG_SMP 1171 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags); 1172 void (*migrate_task_rq)(struct task_struct *p); 1173 1174 void (*task_waking) (struct task_struct *task); 1175 void (*task_woken) (struct rq *this_rq, struct task_struct *task); 1176 1177 void (*set_cpus_allowed)(struct task_struct *p, 1178 const struct cpumask *newmask); 1179 1180 void (*rq_online)(struct rq *rq); 1181 void (*rq_offline)(struct rq *rq); 1182 #endif 1183 1184 void (*set_curr_task) (struct rq *rq); 1185 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued); 1186 void (*task_fork) (struct task_struct *p); 1187 void (*task_dead) (struct task_struct *p); 1188 1189 /* 1190 * The switched_from() call is allowed to drop rq->lock, therefore we 1191 * cannot assume the switched_from/switched_to pair is serliazed by 1192 * rq->lock. They are however serialized by p->pi_lock. 1193 */ 1194 void (*switched_from) (struct rq *this_rq, struct task_struct *task); 1195 void (*switched_to) (struct rq *this_rq, struct task_struct *task); 1196 void (*prio_changed) (struct rq *this_rq, struct task_struct *task, 1197 int oldprio); 1198 1199 unsigned int (*get_rr_interval) (struct rq *rq, 1200 struct task_struct *task); 1201 1202 void (*update_curr) (struct rq *rq); 1203 1204 #ifdef CONFIG_FAIR_GROUP_SCHED 1205 void (*task_move_group) (struct task_struct *p); 1206 #endif 1207 }; 1208 1209 static inline void put_prev_task(struct rq *rq, struct task_struct *prev) 1210 { 1211 prev->sched_class->put_prev_task(rq, prev); 1212 } 1213 1214 #define sched_class_highest (&stop_sched_class) 1215 #define for_each_class(class) \ 1216 for (class = sched_class_highest; class; class = class->next) 1217 1218 extern const struct sched_class stop_sched_class; 1219 extern const struct sched_class dl_sched_class; 1220 extern const struct sched_class rt_sched_class; 1221 extern const struct sched_class fair_sched_class; 1222 extern const struct sched_class idle_sched_class; 1223 1224 1225 #ifdef CONFIG_SMP 1226 1227 extern void update_group_capacity(struct sched_domain *sd, int cpu); 1228 1229 extern void trigger_load_balance(struct rq *rq); 1230 1231 extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask); 1232 1233 #endif 1234 1235 #ifdef CONFIG_CPU_IDLE 1236 static inline void idle_set_state(struct rq *rq, 1237 struct cpuidle_state *idle_state) 1238 { 1239 rq->idle_state = idle_state; 1240 } 1241 1242 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1243 { 1244 WARN_ON(!rcu_read_lock_held()); 1245 return rq->idle_state; 1246 } 1247 #else 1248 static inline void idle_set_state(struct rq *rq, 1249 struct cpuidle_state *idle_state) 1250 { 1251 } 1252 1253 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1254 { 1255 return NULL; 1256 } 1257 #endif 1258 1259 extern void sysrq_sched_debug_show(void); 1260 extern void sched_init_granularity(void); 1261 extern void update_max_interval(void); 1262 1263 extern void init_sched_dl_class(void); 1264 extern void init_sched_rt_class(void); 1265 extern void init_sched_fair_class(void); 1266 1267 extern void resched_curr(struct rq *rq); 1268 extern void resched_cpu(int cpu); 1269 1270 extern struct rt_bandwidth def_rt_bandwidth; 1271 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 1272 1273 extern struct dl_bandwidth def_dl_bandwidth; 1274 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime); 1275 extern void init_dl_task_timer(struct sched_dl_entity *dl_se); 1276 1277 unsigned long to_ratio(u64 period, u64 runtime); 1278 1279 extern void init_entity_runnable_average(struct sched_entity *se); 1280 1281 static inline void add_nr_running(struct rq *rq, unsigned count) 1282 { 1283 unsigned prev_nr = rq->nr_running; 1284 1285 rq->nr_running = prev_nr + count; 1286 1287 if (prev_nr < 2 && rq->nr_running >= 2) { 1288 #ifdef CONFIG_SMP 1289 if (!rq->rd->overload) 1290 rq->rd->overload = true; 1291 #endif 1292 1293 #ifdef CONFIG_NO_HZ_FULL 1294 if (tick_nohz_full_cpu(rq->cpu)) { 1295 /* 1296 * Tick is needed if more than one task runs on a CPU. 1297 * Send the target an IPI to kick it out of nohz mode. 1298 * 1299 * We assume that IPI implies full memory barrier and the 1300 * new value of rq->nr_running is visible on reception 1301 * from the target. 1302 */ 1303 tick_nohz_full_kick_cpu(rq->cpu); 1304 } 1305 #endif 1306 } 1307 } 1308 1309 static inline void sub_nr_running(struct rq *rq, unsigned count) 1310 { 1311 rq->nr_running -= count; 1312 } 1313 1314 static inline void rq_last_tick_reset(struct rq *rq) 1315 { 1316 #ifdef CONFIG_NO_HZ_FULL 1317 rq->last_sched_tick = jiffies; 1318 #endif 1319 } 1320 1321 extern void update_rq_clock(struct rq *rq); 1322 1323 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 1324 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 1325 1326 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); 1327 1328 extern const_debug unsigned int sysctl_sched_time_avg; 1329 extern const_debug unsigned int sysctl_sched_nr_migrate; 1330 extern const_debug unsigned int sysctl_sched_migration_cost; 1331 1332 static inline u64 sched_avg_period(void) 1333 { 1334 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; 1335 } 1336 1337 #ifdef CONFIG_SCHED_HRTICK 1338 1339 /* 1340 * Use hrtick when: 1341 * - enabled by features 1342 * - hrtimer is actually high res 1343 */ 1344 static inline int hrtick_enabled(struct rq *rq) 1345 { 1346 if (!sched_feat(HRTICK)) 1347 return 0; 1348 if (!cpu_active(cpu_of(rq))) 1349 return 0; 1350 return hrtimer_is_hres_active(&rq->hrtick_timer); 1351 } 1352 1353 void hrtick_start(struct rq *rq, u64 delay); 1354 1355 #else 1356 1357 static inline int hrtick_enabled(struct rq *rq) 1358 { 1359 return 0; 1360 } 1361 1362 #endif /* CONFIG_SCHED_HRTICK */ 1363 1364 #ifdef CONFIG_SMP 1365 extern void sched_avg_update(struct rq *rq); 1366 1367 #ifndef arch_scale_freq_capacity 1368 static __always_inline 1369 unsigned long arch_scale_freq_capacity(struct sched_domain *sd, int cpu) 1370 { 1371 return SCHED_CAPACITY_SCALE; 1372 } 1373 #endif 1374 1375 #ifndef arch_scale_cpu_capacity 1376 static __always_inline 1377 unsigned long arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) 1378 { 1379 if (sd && (sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) 1380 return sd->smt_gain / sd->span_weight; 1381 1382 return SCHED_CAPACITY_SCALE; 1383 } 1384 #endif 1385 1386 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) 1387 { 1388 rq->rt_avg += rt_delta * arch_scale_freq_capacity(NULL, cpu_of(rq)); 1389 sched_avg_update(rq); 1390 } 1391 #else 1392 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { } 1393 static inline void sched_avg_update(struct rq *rq) { } 1394 #endif 1395 1396 /* 1397 * __task_rq_lock - lock the rq @p resides on. 1398 */ 1399 static inline struct rq *__task_rq_lock(struct task_struct *p) 1400 __acquires(rq->lock) 1401 { 1402 struct rq *rq; 1403 1404 lockdep_assert_held(&p->pi_lock); 1405 1406 for (;;) { 1407 rq = task_rq(p); 1408 raw_spin_lock(&rq->lock); 1409 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { 1410 lockdep_pin_lock(&rq->lock); 1411 return rq; 1412 } 1413 raw_spin_unlock(&rq->lock); 1414 1415 while (unlikely(task_on_rq_migrating(p))) 1416 cpu_relax(); 1417 } 1418 } 1419 1420 /* 1421 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 1422 */ 1423 static inline struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 1424 __acquires(p->pi_lock) 1425 __acquires(rq->lock) 1426 { 1427 struct rq *rq; 1428 1429 for (;;) { 1430 raw_spin_lock_irqsave(&p->pi_lock, *flags); 1431 rq = task_rq(p); 1432 raw_spin_lock(&rq->lock); 1433 /* 1434 * move_queued_task() task_rq_lock() 1435 * 1436 * ACQUIRE (rq->lock) 1437 * [S] ->on_rq = MIGRATING [L] rq = task_rq() 1438 * WMB (__set_task_cpu()) ACQUIRE (rq->lock); 1439 * [S] ->cpu = new_cpu [L] task_rq() 1440 * [L] ->on_rq 1441 * RELEASE (rq->lock) 1442 * 1443 * If we observe the old cpu in task_rq_lock, the acquire of 1444 * the old rq->lock will fully serialize against the stores. 1445 * 1446 * If we observe the new cpu in task_rq_lock, the acquire will 1447 * pair with the WMB to ensure we must then also see migrating. 1448 */ 1449 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) { 1450 lockdep_pin_lock(&rq->lock); 1451 return rq; 1452 } 1453 raw_spin_unlock(&rq->lock); 1454 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 1455 1456 while (unlikely(task_on_rq_migrating(p))) 1457 cpu_relax(); 1458 } 1459 } 1460 1461 static inline void __task_rq_unlock(struct rq *rq) 1462 __releases(rq->lock) 1463 { 1464 lockdep_unpin_lock(&rq->lock); 1465 raw_spin_unlock(&rq->lock); 1466 } 1467 1468 static inline void 1469 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 1470 __releases(rq->lock) 1471 __releases(p->pi_lock) 1472 { 1473 lockdep_unpin_lock(&rq->lock); 1474 raw_spin_unlock(&rq->lock); 1475 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 1476 } 1477 1478 #ifdef CONFIG_SMP 1479 #ifdef CONFIG_PREEMPT 1480 1481 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2); 1482 1483 /* 1484 * fair double_lock_balance: Safely acquires both rq->locks in a fair 1485 * way at the expense of forcing extra atomic operations in all 1486 * invocations. This assures that the double_lock is acquired using the 1487 * same underlying policy as the spinlock_t on this architecture, which 1488 * reduces latency compared to the unfair variant below. However, it 1489 * also adds more overhead and therefore may reduce throughput. 1490 */ 1491 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1492 __releases(this_rq->lock) 1493 __acquires(busiest->lock) 1494 __acquires(this_rq->lock) 1495 { 1496 raw_spin_unlock(&this_rq->lock); 1497 double_rq_lock(this_rq, busiest); 1498 1499 return 1; 1500 } 1501 1502 #else 1503 /* 1504 * Unfair double_lock_balance: Optimizes throughput at the expense of 1505 * latency by eliminating extra atomic operations when the locks are 1506 * already in proper order on entry. This favors lower cpu-ids and will 1507 * grant the double lock to lower cpus over higher ids under contention, 1508 * regardless of entry order into the function. 1509 */ 1510 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1511 __releases(this_rq->lock) 1512 __acquires(busiest->lock) 1513 __acquires(this_rq->lock) 1514 { 1515 int ret = 0; 1516 1517 if (unlikely(!raw_spin_trylock(&busiest->lock))) { 1518 if (busiest < this_rq) { 1519 raw_spin_unlock(&this_rq->lock); 1520 raw_spin_lock(&busiest->lock); 1521 raw_spin_lock_nested(&this_rq->lock, 1522 SINGLE_DEPTH_NESTING); 1523 ret = 1; 1524 } else 1525 raw_spin_lock_nested(&busiest->lock, 1526 SINGLE_DEPTH_NESTING); 1527 } 1528 return ret; 1529 } 1530 1531 #endif /* CONFIG_PREEMPT */ 1532 1533 /* 1534 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 1535 */ 1536 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 1537 { 1538 if (unlikely(!irqs_disabled())) { 1539 /* printk() doesn't work good under rq->lock */ 1540 raw_spin_unlock(&this_rq->lock); 1541 BUG_ON(1); 1542 } 1543 1544 return _double_lock_balance(this_rq, busiest); 1545 } 1546 1547 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 1548 __releases(busiest->lock) 1549 { 1550 raw_spin_unlock(&busiest->lock); 1551 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1552 } 1553 1554 static inline void double_lock(spinlock_t *l1, spinlock_t *l2) 1555 { 1556 if (l1 > l2) 1557 swap(l1, l2); 1558 1559 spin_lock(l1); 1560 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1561 } 1562 1563 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2) 1564 { 1565 if (l1 > l2) 1566 swap(l1, l2); 1567 1568 spin_lock_irq(l1); 1569 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1570 } 1571 1572 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2) 1573 { 1574 if (l1 > l2) 1575 swap(l1, l2); 1576 1577 raw_spin_lock(l1); 1578 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1579 } 1580 1581 /* 1582 * double_rq_lock - safely lock two runqueues 1583 * 1584 * Note this does not disable interrupts like task_rq_lock, 1585 * you need to do so manually before calling. 1586 */ 1587 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1588 __acquires(rq1->lock) 1589 __acquires(rq2->lock) 1590 { 1591 BUG_ON(!irqs_disabled()); 1592 if (rq1 == rq2) { 1593 raw_spin_lock(&rq1->lock); 1594 __acquire(rq2->lock); /* Fake it out ;) */ 1595 } else { 1596 if (rq1 < rq2) { 1597 raw_spin_lock(&rq1->lock); 1598 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); 1599 } else { 1600 raw_spin_lock(&rq2->lock); 1601 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); 1602 } 1603 } 1604 } 1605 1606 /* 1607 * double_rq_unlock - safely unlock two runqueues 1608 * 1609 * Note this does not restore interrupts like task_rq_unlock, 1610 * you need to do so manually after calling. 1611 */ 1612 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1613 __releases(rq1->lock) 1614 __releases(rq2->lock) 1615 { 1616 raw_spin_unlock(&rq1->lock); 1617 if (rq1 != rq2) 1618 raw_spin_unlock(&rq2->lock); 1619 else 1620 __release(rq2->lock); 1621 } 1622 1623 #else /* CONFIG_SMP */ 1624 1625 /* 1626 * double_rq_lock - safely lock two runqueues 1627 * 1628 * Note this does not disable interrupts like task_rq_lock, 1629 * you need to do so manually before calling. 1630 */ 1631 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1632 __acquires(rq1->lock) 1633 __acquires(rq2->lock) 1634 { 1635 BUG_ON(!irqs_disabled()); 1636 BUG_ON(rq1 != rq2); 1637 raw_spin_lock(&rq1->lock); 1638 __acquire(rq2->lock); /* Fake it out ;) */ 1639 } 1640 1641 /* 1642 * double_rq_unlock - safely unlock two runqueues 1643 * 1644 * Note this does not restore interrupts like task_rq_unlock, 1645 * you need to do so manually after calling. 1646 */ 1647 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1648 __releases(rq1->lock) 1649 __releases(rq2->lock) 1650 { 1651 BUG_ON(rq1 != rq2); 1652 raw_spin_unlock(&rq1->lock); 1653 __release(rq2->lock); 1654 } 1655 1656 #endif 1657 1658 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 1659 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 1660 1661 #ifdef CONFIG_SCHED_DEBUG 1662 extern void print_cfs_stats(struct seq_file *m, int cpu); 1663 extern void print_rt_stats(struct seq_file *m, int cpu); 1664 extern void print_dl_stats(struct seq_file *m, int cpu); 1665 extern void 1666 print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq); 1667 1668 #ifdef CONFIG_NUMA_BALANCING 1669 extern void 1670 show_numa_stats(struct task_struct *p, struct seq_file *m); 1671 extern void 1672 print_numa_stats(struct seq_file *m, int node, unsigned long tsf, 1673 unsigned long tpf, unsigned long gsf, unsigned long gpf); 1674 #endif /* CONFIG_NUMA_BALANCING */ 1675 #endif /* CONFIG_SCHED_DEBUG */ 1676 1677 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 1678 extern void init_rt_rq(struct rt_rq *rt_rq); 1679 extern void init_dl_rq(struct dl_rq *dl_rq); 1680 1681 extern void cfs_bandwidth_usage_inc(void); 1682 extern void cfs_bandwidth_usage_dec(void); 1683 1684 #ifdef CONFIG_NO_HZ_COMMON 1685 enum rq_nohz_flag_bits { 1686 NOHZ_TICK_STOPPED, 1687 NOHZ_BALANCE_KICK, 1688 }; 1689 1690 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 1691 #endif 1692 1693 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1694 1695 DECLARE_PER_CPU(u64, cpu_hardirq_time); 1696 DECLARE_PER_CPU(u64, cpu_softirq_time); 1697 1698 #ifndef CONFIG_64BIT 1699 DECLARE_PER_CPU(seqcount_t, irq_time_seq); 1700 1701 static inline void irq_time_write_begin(void) 1702 { 1703 __this_cpu_inc(irq_time_seq.sequence); 1704 smp_wmb(); 1705 } 1706 1707 static inline void irq_time_write_end(void) 1708 { 1709 smp_wmb(); 1710 __this_cpu_inc(irq_time_seq.sequence); 1711 } 1712 1713 static inline u64 irq_time_read(int cpu) 1714 { 1715 u64 irq_time; 1716 unsigned seq; 1717 1718 do { 1719 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 1720 irq_time = per_cpu(cpu_softirq_time, cpu) + 1721 per_cpu(cpu_hardirq_time, cpu); 1722 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 1723 1724 return irq_time; 1725 } 1726 #else /* CONFIG_64BIT */ 1727 static inline void irq_time_write_begin(void) 1728 { 1729 } 1730 1731 static inline void irq_time_write_end(void) 1732 { 1733 } 1734 1735 static inline u64 irq_time_read(int cpu) 1736 { 1737 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 1738 } 1739 #endif /* CONFIG_64BIT */ 1740 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 1741