1 2 #include <linux/sched.h> 3 #include <linux/mutex.h> 4 #include <linux/spinlock.h> 5 #include <linux/stop_machine.h> 6 7 #include "cpupri.h" 8 9 extern __read_mostly int scheduler_running; 10 11 /* 12 * Convert user-nice values [ -20 ... 0 ... 19 ] 13 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], 14 * and back. 15 */ 16 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) 17 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) 18 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) 19 20 /* 21 * 'User priority' is the nice value converted to something we 22 * can work with better when scaling various scheduler parameters, 23 * it's a [ 0 ... 39 ] range. 24 */ 25 #define USER_PRIO(p) ((p)-MAX_RT_PRIO) 26 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) 27 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) 28 29 /* 30 * Helpers for converting nanosecond timing to jiffy resolution 31 */ 32 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) 33 34 #define NICE_0_LOAD SCHED_LOAD_SCALE 35 #define NICE_0_SHIFT SCHED_LOAD_SHIFT 36 37 /* 38 * These are the 'tuning knobs' of the scheduler: 39 */ 40 41 /* 42 * single value that denotes runtime == period, ie unlimited time. 43 */ 44 #define RUNTIME_INF ((u64)~0ULL) 45 46 static inline int rt_policy(int policy) 47 { 48 if (policy == SCHED_FIFO || policy == SCHED_RR) 49 return 1; 50 return 0; 51 } 52 53 static inline int task_has_rt_policy(struct task_struct *p) 54 { 55 return rt_policy(p->policy); 56 } 57 58 /* 59 * This is the priority-queue data structure of the RT scheduling class: 60 */ 61 struct rt_prio_array { 62 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ 63 struct list_head queue[MAX_RT_PRIO]; 64 }; 65 66 struct rt_bandwidth { 67 /* nests inside the rq lock: */ 68 raw_spinlock_t rt_runtime_lock; 69 ktime_t rt_period; 70 u64 rt_runtime; 71 struct hrtimer rt_period_timer; 72 }; 73 74 extern struct mutex sched_domains_mutex; 75 76 #ifdef CONFIG_CGROUP_SCHED 77 78 #include <linux/cgroup.h> 79 80 struct cfs_rq; 81 struct rt_rq; 82 83 static LIST_HEAD(task_groups); 84 85 struct cfs_bandwidth { 86 #ifdef CONFIG_CFS_BANDWIDTH 87 raw_spinlock_t lock; 88 ktime_t period; 89 u64 quota, runtime; 90 s64 hierarchal_quota; 91 u64 runtime_expires; 92 93 int idle, timer_active; 94 struct hrtimer period_timer, slack_timer; 95 struct list_head throttled_cfs_rq; 96 97 /* statistics */ 98 int nr_periods, nr_throttled; 99 u64 throttled_time; 100 #endif 101 }; 102 103 /* task group related information */ 104 struct task_group { 105 struct cgroup_subsys_state css; 106 107 #ifdef CONFIG_FAIR_GROUP_SCHED 108 /* schedulable entities of this group on each cpu */ 109 struct sched_entity **se; 110 /* runqueue "owned" by this group on each cpu */ 111 struct cfs_rq **cfs_rq; 112 unsigned long shares; 113 114 atomic_t load_weight; 115 #endif 116 117 #ifdef CONFIG_RT_GROUP_SCHED 118 struct sched_rt_entity **rt_se; 119 struct rt_rq **rt_rq; 120 121 struct rt_bandwidth rt_bandwidth; 122 #endif 123 124 struct rcu_head rcu; 125 struct list_head list; 126 127 struct task_group *parent; 128 struct list_head siblings; 129 struct list_head children; 130 131 #ifdef CONFIG_SCHED_AUTOGROUP 132 struct autogroup *autogroup; 133 #endif 134 135 struct cfs_bandwidth cfs_bandwidth; 136 }; 137 138 #ifdef CONFIG_FAIR_GROUP_SCHED 139 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD 140 141 /* 142 * A weight of 0 or 1 can cause arithmetics problems. 143 * A weight of a cfs_rq is the sum of weights of which entities 144 * are queued on this cfs_rq, so a weight of a entity should not be 145 * too large, so as the shares value of a task group. 146 * (The default weight is 1024 - so there's no practical 147 * limitation from this.) 148 */ 149 #define MIN_SHARES (1UL << 1) 150 #define MAX_SHARES (1UL << 18) 151 #endif 152 153 /* Default task group. 154 * Every task in system belong to this group at bootup. 155 */ 156 extern struct task_group root_task_group; 157 158 typedef int (*tg_visitor)(struct task_group *, void *); 159 160 extern int walk_tg_tree_from(struct task_group *from, 161 tg_visitor down, tg_visitor up, void *data); 162 163 /* 164 * Iterate the full tree, calling @down when first entering a node and @up when 165 * leaving it for the final time. 166 * 167 * Caller must hold rcu_lock or sufficient equivalent. 168 */ 169 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) 170 { 171 return walk_tg_tree_from(&root_task_group, down, up, data); 172 } 173 174 extern int tg_nop(struct task_group *tg, void *data); 175 176 extern void free_fair_sched_group(struct task_group *tg); 177 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent); 178 extern void unregister_fair_sched_group(struct task_group *tg, int cpu); 179 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 180 struct sched_entity *se, int cpu, 181 struct sched_entity *parent); 182 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 183 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 184 185 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b); 186 extern void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 187 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq); 188 189 extern void free_rt_sched_group(struct task_group *tg); 190 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent); 191 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 192 struct sched_rt_entity *rt_se, int cpu, 193 struct sched_rt_entity *parent); 194 195 #else /* CONFIG_CGROUP_SCHED */ 196 197 struct cfs_bandwidth { }; 198 199 #endif /* CONFIG_CGROUP_SCHED */ 200 201 /* CFS-related fields in a runqueue */ 202 struct cfs_rq { 203 struct load_weight load; 204 unsigned int nr_running, h_nr_running; 205 206 u64 exec_clock; 207 u64 min_vruntime; 208 #ifndef CONFIG_64BIT 209 u64 min_vruntime_copy; 210 #endif 211 212 struct rb_root tasks_timeline; 213 struct rb_node *rb_leftmost; 214 215 /* 216 * 'curr' points to currently running entity on this cfs_rq. 217 * It is set to NULL otherwise (i.e when none are currently running). 218 */ 219 struct sched_entity *curr, *next, *last, *skip; 220 221 #ifdef CONFIG_SCHED_DEBUG 222 unsigned int nr_spread_over; 223 #endif 224 225 #ifdef CONFIG_FAIR_GROUP_SCHED 226 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ 227 228 /* 229 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in 230 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities 231 * (like users, containers etc.) 232 * 233 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This 234 * list is used during load balance. 235 */ 236 int on_list; 237 struct list_head leaf_cfs_rq_list; 238 struct task_group *tg; /* group that "owns" this runqueue */ 239 240 #ifdef CONFIG_SMP 241 /* 242 * h_load = weight * f(tg) 243 * 244 * Where f(tg) is the recursive weight fraction assigned to 245 * this group. 246 */ 247 unsigned long h_load; 248 249 /* 250 * Maintaining per-cpu shares distribution for group scheduling 251 * 252 * load_stamp is the last time we updated the load average 253 * load_last is the last time we updated the load average and saw load 254 * load_unacc_exec_time is currently unaccounted execution time 255 */ 256 u64 load_avg; 257 u64 load_period; 258 u64 load_stamp, load_last, load_unacc_exec_time; 259 260 unsigned long load_contribution; 261 #endif /* CONFIG_SMP */ 262 #ifdef CONFIG_CFS_BANDWIDTH 263 int runtime_enabled; 264 u64 runtime_expires; 265 s64 runtime_remaining; 266 267 u64 throttled_timestamp; 268 int throttled, throttle_count; 269 struct list_head throttled_list; 270 #endif /* CONFIG_CFS_BANDWIDTH */ 271 #endif /* CONFIG_FAIR_GROUP_SCHED */ 272 }; 273 274 static inline int rt_bandwidth_enabled(void) 275 { 276 return sysctl_sched_rt_runtime >= 0; 277 } 278 279 /* Real-Time classes' related field in a runqueue: */ 280 struct rt_rq { 281 struct rt_prio_array active; 282 unsigned int rt_nr_running; 283 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 284 struct { 285 int curr; /* highest queued rt task prio */ 286 #ifdef CONFIG_SMP 287 int next; /* next highest */ 288 #endif 289 } highest_prio; 290 #endif 291 #ifdef CONFIG_SMP 292 unsigned long rt_nr_migratory; 293 unsigned long rt_nr_total; 294 int overloaded; 295 struct plist_head pushable_tasks; 296 #endif 297 int rt_throttled; 298 u64 rt_time; 299 u64 rt_runtime; 300 /* Nests inside the rq lock: */ 301 raw_spinlock_t rt_runtime_lock; 302 303 #ifdef CONFIG_RT_GROUP_SCHED 304 unsigned long rt_nr_boosted; 305 306 struct rq *rq; 307 struct list_head leaf_rt_rq_list; 308 struct task_group *tg; 309 #endif 310 }; 311 312 #ifdef CONFIG_SMP 313 314 /* 315 * We add the notion of a root-domain which will be used to define per-domain 316 * variables. Each exclusive cpuset essentially defines an island domain by 317 * fully partitioning the member cpus from any other cpuset. Whenever a new 318 * exclusive cpuset is created, we also create and attach a new root-domain 319 * object. 320 * 321 */ 322 struct root_domain { 323 atomic_t refcount; 324 atomic_t rto_count; 325 struct rcu_head rcu; 326 cpumask_var_t span; 327 cpumask_var_t online; 328 329 /* 330 * The "RT overload" flag: it gets set if a CPU has more than 331 * one runnable RT task. 332 */ 333 cpumask_var_t rto_mask; 334 struct cpupri cpupri; 335 }; 336 337 extern struct root_domain def_root_domain; 338 339 #endif /* CONFIG_SMP */ 340 341 /* 342 * This is the main, per-CPU runqueue data structure. 343 * 344 * Locking rule: those places that want to lock multiple runqueues 345 * (such as the load balancing or the thread migration code), lock 346 * acquire operations must be ordered by ascending &runqueue. 347 */ 348 struct rq { 349 /* runqueue lock: */ 350 raw_spinlock_t lock; 351 352 /* 353 * nr_running and cpu_load should be in the same cacheline because 354 * remote CPUs use both these fields when doing load calculation. 355 */ 356 unsigned int nr_running; 357 #define CPU_LOAD_IDX_MAX 5 358 unsigned long cpu_load[CPU_LOAD_IDX_MAX]; 359 unsigned long last_load_update_tick; 360 #ifdef CONFIG_NO_HZ 361 u64 nohz_stamp; 362 unsigned long nohz_flags; 363 #endif 364 int skip_clock_update; 365 366 /* capture load from *all* tasks on this cpu: */ 367 struct load_weight load; 368 unsigned long nr_load_updates; 369 u64 nr_switches; 370 371 struct cfs_rq cfs; 372 struct rt_rq rt; 373 374 #ifdef CONFIG_FAIR_GROUP_SCHED 375 /* list of leaf cfs_rq on this cpu: */ 376 struct list_head leaf_cfs_rq_list; 377 #endif 378 #ifdef CONFIG_RT_GROUP_SCHED 379 struct list_head leaf_rt_rq_list; 380 #endif 381 382 /* 383 * This is part of a global counter where only the total sum 384 * over all CPUs matters. A task can increase this counter on 385 * one CPU and if it got migrated afterwards it may decrease 386 * it on another CPU. Always updated under the runqueue lock: 387 */ 388 unsigned long nr_uninterruptible; 389 390 struct task_struct *curr, *idle, *stop; 391 unsigned long next_balance; 392 struct mm_struct *prev_mm; 393 394 u64 clock; 395 u64 clock_task; 396 397 atomic_t nr_iowait; 398 399 #ifdef CONFIG_SMP 400 struct root_domain *rd; 401 struct sched_domain *sd; 402 403 unsigned long cpu_power; 404 405 unsigned char idle_balance; 406 /* For active balancing */ 407 int post_schedule; 408 int active_balance; 409 int push_cpu; 410 struct cpu_stop_work active_balance_work; 411 /* cpu of this runqueue: */ 412 int cpu; 413 int online; 414 415 struct list_head cfs_tasks; 416 417 u64 rt_avg; 418 u64 age_stamp; 419 u64 idle_stamp; 420 u64 avg_idle; 421 #endif 422 423 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 424 u64 prev_irq_time; 425 #endif 426 #ifdef CONFIG_PARAVIRT 427 u64 prev_steal_time; 428 #endif 429 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 430 u64 prev_steal_time_rq; 431 #endif 432 433 /* calc_load related fields */ 434 unsigned long calc_load_update; 435 long calc_load_active; 436 437 #ifdef CONFIG_SCHED_HRTICK 438 #ifdef CONFIG_SMP 439 int hrtick_csd_pending; 440 struct call_single_data hrtick_csd; 441 #endif 442 struct hrtimer hrtick_timer; 443 #endif 444 445 #ifdef CONFIG_SCHEDSTATS 446 /* latency stats */ 447 struct sched_info rq_sched_info; 448 unsigned long long rq_cpu_time; 449 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 450 451 /* sys_sched_yield() stats */ 452 unsigned int yld_count; 453 454 /* schedule() stats */ 455 unsigned int sched_count; 456 unsigned int sched_goidle; 457 458 /* try_to_wake_up() stats */ 459 unsigned int ttwu_count; 460 unsigned int ttwu_local; 461 #endif 462 463 #ifdef CONFIG_SMP 464 struct llist_head wake_list; 465 #endif 466 }; 467 468 static inline int cpu_of(struct rq *rq) 469 { 470 #ifdef CONFIG_SMP 471 return rq->cpu; 472 #else 473 return 0; 474 #endif 475 } 476 477 DECLARE_PER_CPU(struct rq, runqueues); 478 479 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 480 #define this_rq() (&__get_cpu_var(runqueues)) 481 #define task_rq(p) cpu_rq(task_cpu(p)) 482 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 483 #define raw_rq() (&__raw_get_cpu_var(runqueues)) 484 485 #ifdef CONFIG_SMP 486 487 #define rcu_dereference_check_sched_domain(p) \ 488 rcu_dereference_check((p), \ 489 lockdep_is_held(&sched_domains_mutex)) 490 491 /* 492 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 493 * See detach_destroy_domains: synchronize_sched for details. 494 * 495 * The domain tree of any CPU may only be accessed from within 496 * preempt-disabled sections. 497 */ 498 #define for_each_domain(cpu, __sd) \ 499 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 500 __sd; __sd = __sd->parent) 501 502 #define for_each_lower_domain(sd) for (; sd; sd = sd->child) 503 504 /** 505 * highest_flag_domain - Return highest sched_domain containing flag. 506 * @cpu: The cpu whose highest level of sched domain is to 507 * be returned. 508 * @flag: The flag to check for the highest sched_domain 509 * for the given cpu. 510 * 511 * Returns the highest sched_domain of a cpu which contains the given flag. 512 */ 513 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 514 { 515 struct sched_domain *sd, *hsd = NULL; 516 517 for_each_domain(cpu, sd) { 518 if (!(sd->flags & flag)) 519 break; 520 hsd = sd; 521 } 522 523 return hsd; 524 } 525 526 DECLARE_PER_CPU(struct sched_domain *, sd_llc); 527 DECLARE_PER_CPU(int, sd_llc_id); 528 529 extern int group_balance_cpu(struct sched_group *sg); 530 531 #endif /* CONFIG_SMP */ 532 533 #include "stats.h" 534 #include "auto_group.h" 535 536 #ifdef CONFIG_CGROUP_SCHED 537 538 /* 539 * Return the group to which this tasks belongs. 540 * 541 * We use task_subsys_state_check() and extend the RCU verification with 542 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each 543 * task it moves into the cgroup. Therefore by holding either of those locks, 544 * we pin the task to the current cgroup. 545 */ 546 static inline struct task_group *task_group(struct task_struct *p) 547 { 548 struct task_group *tg; 549 struct cgroup_subsys_state *css; 550 551 css = task_subsys_state_check(p, cpu_cgroup_subsys_id, 552 lockdep_is_held(&p->pi_lock) || 553 lockdep_is_held(&task_rq(p)->lock)); 554 tg = container_of(css, struct task_group, css); 555 556 return autogroup_task_group(p, tg); 557 } 558 559 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 560 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 561 { 562 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 563 struct task_group *tg = task_group(p); 564 #endif 565 566 #ifdef CONFIG_FAIR_GROUP_SCHED 567 p->se.cfs_rq = tg->cfs_rq[cpu]; 568 p->se.parent = tg->se[cpu]; 569 #endif 570 571 #ifdef CONFIG_RT_GROUP_SCHED 572 p->rt.rt_rq = tg->rt_rq[cpu]; 573 p->rt.parent = tg->rt_se[cpu]; 574 #endif 575 } 576 577 #else /* CONFIG_CGROUP_SCHED */ 578 579 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 580 static inline struct task_group *task_group(struct task_struct *p) 581 { 582 return NULL; 583 } 584 585 #endif /* CONFIG_CGROUP_SCHED */ 586 587 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 588 { 589 set_task_rq(p, cpu); 590 #ifdef CONFIG_SMP 591 /* 592 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 593 * successfuly executed on another CPU. We must ensure that updates of 594 * per-task data have been completed by this moment. 595 */ 596 smp_wmb(); 597 task_thread_info(p)->cpu = cpu; 598 #endif 599 } 600 601 /* 602 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 603 */ 604 #ifdef CONFIG_SCHED_DEBUG 605 # include <linux/static_key.h> 606 # define const_debug __read_mostly 607 #else 608 # define const_debug const 609 #endif 610 611 extern const_debug unsigned int sysctl_sched_features; 612 613 #define SCHED_FEAT(name, enabled) \ 614 __SCHED_FEAT_##name , 615 616 enum { 617 #include "features.h" 618 __SCHED_FEAT_NR, 619 }; 620 621 #undef SCHED_FEAT 622 623 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL) 624 static __always_inline bool static_branch__true(struct static_key *key) 625 { 626 return static_key_true(key); /* Not out of line branch. */ 627 } 628 629 static __always_inline bool static_branch__false(struct static_key *key) 630 { 631 return static_key_false(key); /* Out of line branch. */ 632 } 633 634 #define SCHED_FEAT(name, enabled) \ 635 static __always_inline bool static_branch_##name(struct static_key *key) \ 636 { \ 637 return static_branch__##enabled(key); \ 638 } 639 640 #include "features.h" 641 642 #undef SCHED_FEAT 643 644 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 645 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 646 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */ 647 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 648 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */ 649 650 static inline u64 global_rt_period(void) 651 { 652 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 653 } 654 655 static inline u64 global_rt_runtime(void) 656 { 657 if (sysctl_sched_rt_runtime < 0) 658 return RUNTIME_INF; 659 660 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 661 } 662 663 664 665 static inline int task_current(struct rq *rq, struct task_struct *p) 666 { 667 return rq->curr == p; 668 } 669 670 static inline int task_running(struct rq *rq, struct task_struct *p) 671 { 672 #ifdef CONFIG_SMP 673 return p->on_cpu; 674 #else 675 return task_current(rq, p); 676 #endif 677 } 678 679 680 #ifndef prepare_arch_switch 681 # define prepare_arch_switch(next) do { } while (0) 682 #endif 683 #ifndef finish_arch_switch 684 # define finish_arch_switch(prev) do { } while (0) 685 #endif 686 #ifndef finish_arch_post_lock_switch 687 # define finish_arch_post_lock_switch() do { } while (0) 688 #endif 689 690 #ifndef __ARCH_WANT_UNLOCKED_CTXSW 691 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 692 { 693 #ifdef CONFIG_SMP 694 /* 695 * We can optimise this out completely for !SMP, because the 696 * SMP rebalancing from interrupt is the only thing that cares 697 * here. 698 */ 699 next->on_cpu = 1; 700 #endif 701 } 702 703 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 704 { 705 #ifdef CONFIG_SMP 706 /* 707 * After ->on_cpu is cleared, the task can be moved to a different CPU. 708 * We must ensure this doesn't happen until the switch is completely 709 * finished. 710 */ 711 smp_wmb(); 712 prev->on_cpu = 0; 713 #endif 714 #ifdef CONFIG_DEBUG_SPINLOCK 715 /* this is a valid case when another task releases the spinlock */ 716 rq->lock.owner = current; 717 #endif 718 /* 719 * If we are tracking spinlock dependencies then we have to 720 * fix up the runqueue lock - which gets 'carried over' from 721 * prev into current: 722 */ 723 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); 724 725 raw_spin_unlock_irq(&rq->lock); 726 } 727 728 #else /* __ARCH_WANT_UNLOCKED_CTXSW */ 729 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 730 { 731 #ifdef CONFIG_SMP 732 /* 733 * We can optimise this out completely for !SMP, because the 734 * SMP rebalancing from interrupt is the only thing that cares 735 * here. 736 */ 737 next->on_cpu = 1; 738 #endif 739 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 740 raw_spin_unlock_irq(&rq->lock); 741 #else 742 raw_spin_unlock(&rq->lock); 743 #endif 744 } 745 746 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 747 { 748 #ifdef CONFIG_SMP 749 /* 750 * After ->on_cpu is cleared, the task can be moved to a different CPU. 751 * We must ensure this doesn't happen until the switch is completely 752 * finished. 753 */ 754 smp_wmb(); 755 prev->on_cpu = 0; 756 #endif 757 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW 758 local_irq_enable(); 759 #endif 760 } 761 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ 762 763 764 static inline void update_load_add(struct load_weight *lw, unsigned long inc) 765 { 766 lw->weight += inc; 767 lw->inv_weight = 0; 768 } 769 770 static inline void update_load_sub(struct load_weight *lw, unsigned long dec) 771 { 772 lw->weight -= dec; 773 lw->inv_weight = 0; 774 } 775 776 static inline void update_load_set(struct load_weight *lw, unsigned long w) 777 { 778 lw->weight = w; 779 lw->inv_weight = 0; 780 } 781 782 /* 783 * To aid in avoiding the subversion of "niceness" due to uneven distribution 784 * of tasks with abnormal "nice" values across CPUs the contribution that 785 * each task makes to its run queue's load is weighted according to its 786 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 787 * scaled version of the new time slice allocation that they receive on time 788 * slice expiry etc. 789 */ 790 791 #define WEIGHT_IDLEPRIO 3 792 #define WMULT_IDLEPRIO 1431655765 793 794 /* 795 * Nice levels are multiplicative, with a gentle 10% change for every 796 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to 797 * nice 1, it will get ~10% less CPU time than another CPU-bound task 798 * that remained on nice 0. 799 * 800 * The "10% effect" is relative and cumulative: from _any_ nice level, 801 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level 802 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. 803 * If a task goes up by ~10% and another task goes down by ~10% then 804 * the relative distance between them is ~25%.) 805 */ 806 static const int prio_to_weight[40] = { 807 /* -20 */ 88761, 71755, 56483, 46273, 36291, 808 /* -15 */ 29154, 23254, 18705, 14949, 11916, 809 /* -10 */ 9548, 7620, 6100, 4904, 3906, 810 /* -5 */ 3121, 2501, 1991, 1586, 1277, 811 /* 0 */ 1024, 820, 655, 526, 423, 812 /* 5 */ 335, 272, 215, 172, 137, 813 /* 10 */ 110, 87, 70, 56, 45, 814 /* 15 */ 36, 29, 23, 18, 15, 815 }; 816 817 /* 818 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. 819 * 820 * In cases where the weight does not change often, we can use the 821 * precalculated inverse to speed up arithmetics by turning divisions 822 * into multiplications: 823 */ 824 static const u32 prio_to_wmult[40] = { 825 /* -20 */ 48388, 59856, 76040, 92818, 118348, 826 /* -15 */ 147320, 184698, 229616, 287308, 360437, 827 /* -10 */ 449829, 563644, 704093, 875809, 1099582, 828 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, 829 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, 830 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, 831 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, 832 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, 833 }; 834 835 /* Time spent by the tasks of the cpu accounting group executing in ... */ 836 enum cpuacct_stat_index { 837 CPUACCT_STAT_USER, /* ... user mode */ 838 CPUACCT_STAT_SYSTEM, /* ... kernel mode */ 839 840 CPUACCT_STAT_NSTATS, 841 }; 842 843 844 #define sched_class_highest (&stop_sched_class) 845 #define for_each_class(class) \ 846 for (class = sched_class_highest; class; class = class->next) 847 848 extern const struct sched_class stop_sched_class; 849 extern const struct sched_class rt_sched_class; 850 extern const struct sched_class fair_sched_class; 851 extern const struct sched_class idle_sched_class; 852 853 854 #ifdef CONFIG_SMP 855 856 extern void trigger_load_balance(struct rq *rq, int cpu); 857 extern void idle_balance(int this_cpu, struct rq *this_rq); 858 859 #else /* CONFIG_SMP */ 860 861 static inline void idle_balance(int cpu, struct rq *rq) 862 { 863 } 864 865 #endif 866 867 extern void sysrq_sched_debug_show(void); 868 extern void sched_init_granularity(void); 869 extern void update_max_interval(void); 870 extern void update_group_power(struct sched_domain *sd, int cpu); 871 extern int update_runtime(struct notifier_block *nfb, unsigned long action, void *hcpu); 872 extern void init_sched_rt_class(void); 873 extern void init_sched_fair_class(void); 874 875 extern void resched_task(struct task_struct *p); 876 extern void resched_cpu(int cpu); 877 878 extern struct rt_bandwidth def_rt_bandwidth; 879 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 880 881 extern void update_idle_cpu_load(struct rq *this_rq); 882 883 #ifdef CONFIG_CGROUP_CPUACCT 884 #include <linux/cgroup.h> 885 /* track cpu usage of a group of tasks and its child groups */ 886 struct cpuacct { 887 struct cgroup_subsys_state css; 888 /* cpuusage holds pointer to a u64-type object on every cpu */ 889 u64 __percpu *cpuusage; 890 struct kernel_cpustat __percpu *cpustat; 891 }; 892 893 /* return cpu accounting group corresponding to this container */ 894 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) 895 { 896 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), 897 struct cpuacct, css); 898 } 899 900 /* return cpu accounting group to which this task belongs */ 901 static inline struct cpuacct *task_ca(struct task_struct *tsk) 902 { 903 return container_of(task_subsys_state(tsk, cpuacct_subsys_id), 904 struct cpuacct, css); 905 } 906 907 static inline struct cpuacct *parent_ca(struct cpuacct *ca) 908 { 909 if (!ca || !ca->css.cgroup->parent) 910 return NULL; 911 return cgroup_ca(ca->css.cgroup->parent); 912 } 913 914 extern void cpuacct_charge(struct task_struct *tsk, u64 cputime); 915 #else 916 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} 917 #endif 918 919 static inline void inc_nr_running(struct rq *rq) 920 { 921 rq->nr_running++; 922 } 923 924 static inline void dec_nr_running(struct rq *rq) 925 { 926 rq->nr_running--; 927 } 928 929 extern void update_rq_clock(struct rq *rq); 930 931 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 932 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 933 934 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); 935 936 extern const_debug unsigned int sysctl_sched_time_avg; 937 extern const_debug unsigned int sysctl_sched_nr_migrate; 938 extern const_debug unsigned int sysctl_sched_migration_cost; 939 940 static inline u64 sched_avg_period(void) 941 { 942 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; 943 } 944 945 void calc_load_account_idle(struct rq *this_rq); 946 947 #ifdef CONFIG_SCHED_HRTICK 948 949 /* 950 * Use hrtick when: 951 * - enabled by features 952 * - hrtimer is actually high res 953 */ 954 static inline int hrtick_enabled(struct rq *rq) 955 { 956 if (!sched_feat(HRTICK)) 957 return 0; 958 if (!cpu_active(cpu_of(rq))) 959 return 0; 960 return hrtimer_is_hres_active(&rq->hrtick_timer); 961 } 962 963 void hrtick_start(struct rq *rq, u64 delay); 964 965 #else 966 967 static inline int hrtick_enabled(struct rq *rq) 968 { 969 return 0; 970 } 971 972 #endif /* CONFIG_SCHED_HRTICK */ 973 974 #ifdef CONFIG_SMP 975 extern void sched_avg_update(struct rq *rq); 976 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) 977 { 978 rq->rt_avg += rt_delta; 979 sched_avg_update(rq); 980 } 981 #else 982 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { } 983 static inline void sched_avg_update(struct rq *rq) { } 984 #endif 985 986 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period); 987 988 #ifdef CONFIG_SMP 989 #ifdef CONFIG_PREEMPT 990 991 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2); 992 993 /* 994 * fair double_lock_balance: Safely acquires both rq->locks in a fair 995 * way at the expense of forcing extra atomic operations in all 996 * invocations. This assures that the double_lock is acquired using the 997 * same underlying policy as the spinlock_t on this architecture, which 998 * reduces latency compared to the unfair variant below. However, it 999 * also adds more overhead and therefore may reduce throughput. 1000 */ 1001 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1002 __releases(this_rq->lock) 1003 __acquires(busiest->lock) 1004 __acquires(this_rq->lock) 1005 { 1006 raw_spin_unlock(&this_rq->lock); 1007 double_rq_lock(this_rq, busiest); 1008 1009 return 1; 1010 } 1011 1012 #else 1013 /* 1014 * Unfair double_lock_balance: Optimizes throughput at the expense of 1015 * latency by eliminating extra atomic operations when the locks are 1016 * already in proper order on entry. This favors lower cpu-ids and will 1017 * grant the double lock to lower cpus over higher ids under contention, 1018 * regardless of entry order into the function. 1019 */ 1020 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1021 __releases(this_rq->lock) 1022 __acquires(busiest->lock) 1023 __acquires(this_rq->lock) 1024 { 1025 int ret = 0; 1026 1027 if (unlikely(!raw_spin_trylock(&busiest->lock))) { 1028 if (busiest < this_rq) { 1029 raw_spin_unlock(&this_rq->lock); 1030 raw_spin_lock(&busiest->lock); 1031 raw_spin_lock_nested(&this_rq->lock, 1032 SINGLE_DEPTH_NESTING); 1033 ret = 1; 1034 } else 1035 raw_spin_lock_nested(&busiest->lock, 1036 SINGLE_DEPTH_NESTING); 1037 } 1038 return ret; 1039 } 1040 1041 #endif /* CONFIG_PREEMPT */ 1042 1043 /* 1044 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 1045 */ 1046 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 1047 { 1048 if (unlikely(!irqs_disabled())) { 1049 /* printk() doesn't work good under rq->lock */ 1050 raw_spin_unlock(&this_rq->lock); 1051 BUG_ON(1); 1052 } 1053 1054 return _double_lock_balance(this_rq, busiest); 1055 } 1056 1057 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 1058 __releases(busiest->lock) 1059 { 1060 raw_spin_unlock(&busiest->lock); 1061 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1062 } 1063 1064 /* 1065 * double_rq_lock - safely lock two runqueues 1066 * 1067 * Note this does not disable interrupts like task_rq_lock, 1068 * you need to do so manually before calling. 1069 */ 1070 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1071 __acquires(rq1->lock) 1072 __acquires(rq2->lock) 1073 { 1074 BUG_ON(!irqs_disabled()); 1075 if (rq1 == rq2) { 1076 raw_spin_lock(&rq1->lock); 1077 __acquire(rq2->lock); /* Fake it out ;) */ 1078 } else { 1079 if (rq1 < rq2) { 1080 raw_spin_lock(&rq1->lock); 1081 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); 1082 } else { 1083 raw_spin_lock(&rq2->lock); 1084 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); 1085 } 1086 } 1087 } 1088 1089 /* 1090 * double_rq_unlock - safely unlock two runqueues 1091 * 1092 * Note this does not restore interrupts like task_rq_unlock, 1093 * you need to do so manually after calling. 1094 */ 1095 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1096 __releases(rq1->lock) 1097 __releases(rq2->lock) 1098 { 1099 raw_spin_unlock(&rq1->lock); 1100 if (rq1 != rq2) 1101 raw_spin_unlock(&rq2->lock); 1102 else 1103 __release(rq2->lock); 1104 } 1105 1106 #else /* CONFIG_SMP */ 1107 1108 /* 1109 * double_rq_lock - safely lock two runqueues 1110 * 1111 * Note this does not disable interrupts like task_rq_lock, 1112 * you need to do so manually before calling. 1113 */ 1114 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1115 __acquires(rq1->lock) 1116 __acquires(rq2->lock) 1117 { 1118 BUG_ON(!irqs_disabled()); 1119 BUG_ON(rq1 != rq2); 1120 raw_spin_lock(&rq1->lock); 1121 __acquire(rq2->lock); /* Fake it out ;) */ 1122 } 1123 1124 /* 1125 * double_rq_unlock - safely unlock two runqueues 1126 * 1127 * Note this does not restore interrupts like task_rq_unlock, 1128 * you need to do so manually after calling. 1129 */ 1130 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1131 __releases(rq1->lock) 1132 __releases(rq2->lock) 1133 { 1134 BUG_ON(rq1 != rq2); 1135 raw_spin_unlock(&rq1->lock); 1136 __release(rq2->lock); 1137 } 1138 1139 #endif 1140 1141 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 1142 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 1143 extern void print_cfs_stats(struct seq_file *m, int cpu); 1144 extern void print_rt_stats(struct seq_file *m, int cpu); 1145 1146 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 1147 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq); 1148 extern void unthrottle_offline_cfs_rqs(struct rq *rq); 1149 1150 extern void account_cfs_bandwidth_used(int enabled, int was_enabled); 1151 1152 #ifdef CONFIG_NO_HZ 1153 enum rq_nohz_flag_bits { 1154 NOHZ_TICK_STOPPED, 1155 NOHZ_BALANCE_KICK, 1156 NOHZ_IDLE, 1157 }; 1158 1159 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 1160 #endif 1161