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 extern struct 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 #ifdef CONFIG_SMP 378 unsigned long h_load_throttle; 379 #endif /* CONFIG_SMP */ 380 #endif /* CONFIG_FAIR_GROUP_SCHED */ 381 382 #ifdef CONFIG_RT_GROUP_SCHED 383 struct list_head leaf_rt_rq_list; 384 #endif 385 386 /* 387 * This is part of a global counter where only the total sum 388 * over all CPUs matters. A task can increase this counter on 389 * one CPU and if it got migrated afterwards it may decrease 390 * it on another CPU. Always updated under the runqueue lock: 391 */ 392 unsigned long nr_uninterruptible; 393 394 struct task_struct *curr, *idle, *stop; 395 unsigned long next_balance; 396 struct mm_struct *prev_mm; 397 398 u64 clock; 399 u64 clock_task; 400 401 atomic_t nr_iowait; 402 403 #ifdef CONFIG_SMP 404 struct root_domain *rd; 405 struct sched_domain *sd; 406 407 unsigned long cpu_power; 408 409 unsigned char idle_balance; 410 /* For active balancing */ 411 int post_schedule; 412 int active_balance; 413 int push_cpu; 414 struct cpu_stop_work active_balance_work; 415 /* cpu of this runqueue: */ 416 int cpu; 417 int online; 418 419 struct list_head cfs_tasks; 420 421 u64 rt_avg; 422 u64 age_stamp; 423 u64 idle_stamp; 424 u64 avg_idle; 425 #endif 426 427 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 428 u64 prev_irq_time; 429 #endif 430 #ifdef CONFIG_PARAVIRT 431 u64 prev_steal_time; 432 #endif 433 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 434 u64 prev_steal_time_rq; 435 #endif 436 437 /* calc_load related fields */ 438 unsigned long calc_load_update; 439 long calc_load_active; 440 441 #ifdef CONFIG_SCHED_HRTICK 442 #ifdef CONFIG_SMP 443 int hrtick_csd_pending; 444 struct call_single_data hrtick_csd; 445 #endif 446 struct hrtimer hrtick_timer; 447 #endif 448 449 #ifdef CONFIG_SCHEDSTATS 450 /* latency stats */ 451 struct sched_info rq_sched_info; 452 unsigned long long rq_cpu_time; 453 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 454 455 /* sys_sched_yield() stats */ 456 unsigned int yld_count; 457 458 /* schedule() stats */ 459 unsigned int sched_count; 460 unsigned int sched_goidle; 461 462 /* try_to_wake_up() stats */ 463 unsigned int ttwu_count; 464 unsigned int ttwu_local; 465 #endif 466 467 #ifdef CONFIG_SMP 468 struct llist_head wake_list; 469 #endif 470 }; 471 472 static inline int cpu_of(struct rq *rq) 473 { 474 #ifdef CONFIG_SMP 475 return rq->cpu; 476 #else 477 return 0; 478 #endif 479 } 480 481 DECLARE_PER_CPU(struct rq, runqueues); 482 483 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 484 #define this_rq() (&__get_cpu_var(runqueues)) 485 #define task_rq(p) cpu_rq(task_cpu(p)) 486 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 487 #define raw_rq() (&__raw_get_cpu_var(runqueues)) 488 489 #ifdef CONFIG_SMP 490 491 #define rcu_dereference_check_sched_domain(p) \ 492 rcu_dereference_check((p), \ 493 lockdep_is_held(&sched_domains_mutex)) 494 495 /* 496 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 497 * See detach_destroy_domains: synchronize_sched for details. 498 * 499 * The domain tree of any CPU may only be accessed from within 500 * preempt-disabled sections. 501 */ 502 #define for_each_domain(cpu, __sd) \ 503 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 504 __sd; __sd = __sd->parent) 505 506 #define for_each_lower_domain(sd) for (; sd; sd = sd->child) 507 508 /** 509 * highest_flag_domain - Return highest sched_domain containing flag. 510 * @cpu: The cpu whose highest level of sched domain is to 511 * be returned. 512 * @flag: The flag to check for the highest sched_domain 513 * for the given cpu. 514 * 515 * Returns the highest sched_domain of a cpu which contains the given flag. 516 */ 517 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 518 { 519 struct sched_domain *sd, *hsd = NULL; 520 521 for_each_domain(cpu, sd) { 522 if (!(sd->flags & flag)) 523 break; 524 hsd = sd; 525 } 526 527 return hsd; 528 } 529 530 DECLARE_PER_CPU(struct sched_domain *, sd_llc); 531 DECLARE_PER_CPU(int, sd_llc_id); 532 533 extern int group_balance_cpu(struct sched_group *sg); 534 535 #endif /* CONFIG_SMP */ 536 537 #include "stats.h" 538 #include "auto_group.h" 539 540 #ifdef CONFIG_CGROUP_SCHED 541 542 /* 543 * Return the group to which this tasks belongs. 544 * 545 * We cannot use task_subsys_state() and friends because the cgroup 546 * subsystem changes that value before the cgroup_subsys::attach() method 547 * is called, therefore we cannot pin it and might observe the wrong value. 548 * 549 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup 550 * core changes this before calling sched_move_task(). 551 * 552 * Instead we use a 'copy' which is updated from sched_move_task() while 553 * holding both task_struct::pi_lock and rq::lock. 554 */ 555 static inline struct task_group *task_group(struct task_struct *p) 556 { 557 return p->sched_task_group; 558 } 559 560 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 561 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 562 { 563 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 564 struct task_group *tg = task_group(p); 565 #endif 566 567 #ifdef CONFIG_FAIR_GROUP_SCHED 568 p->se.cfs_rq = tg->cfs_rq[cpu]; 569 p->se.parent = tg->se[cpu]; 570 #endif 571 572 #ifdef CONFIG_RT_GROUP_SCHED 573 p->rt.rt_rq = tg->rt_rq[cpu]; 574 p->rt.parent = tg->rt_se[cpu]; 575 #endif 576 } 577 578 #else /* CONFIG_CGROUP_SCHED */ 579 580 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 581 static inline struct task_group *task_group(struct task_struct *p) 582 { 583 return NULL; 584 } 585 586 #endif /* CONFIG_CGROUP_SCHED */ 587 588 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 589 { 590 set_task_rq(p, cpu); 591 #ifdef CONFIG_SMP 592 /* 593 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 594 * successfuly executed on another CPU. We must ensure that updates of 595 * per-task data have been completed by this moment. 596 */ 597 smp_wmb(); 598 task_thread_info(p)->cpu = cpu; 599 #endif 600 } 601 602 /* 603 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 604 */ 605 #ifdef CONFIG_SCHED_DEBUG 606 # include <linux/static_key.h> 607 # define const_debug __read_mostly 608 #else 609 # define const_debug const 610 #endif 611 612 extern const_debug unsigned int sysctl_sched_features; 613 614 #define SCHED_FEAT(name, enabled) \ 615 __SCHED_FEAT_##name , 616 617 enum { 618 #include "features.h" 619 __SCHED_FEAT_NR, 620 }; 621 622 #undef SCHED_FEAT 623 624 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL) 625 static __always_inline bool static_branch__true(struct static_key *key) 626 { 627 return static_key_true(key); /* Not out of line branch. */ 628 } 629 630 static __always_inline bool static_branch__false(struct static_key *key) 631 { 632 return static_key_false(key); /* Out of line branch. */ 633 } 634 635 #define SCHED_FEAT(name, enabled) \ 636 static __always_inline bool static_branch_##name(struct static_key *key) \ 637 { \ 638 return static_branch__##enabled(key); \ 639 } 640 641 #include "features.h" 642 643 #undef SCHED_FEAT 644 645 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 646 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 647 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */ 648 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 649 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */ 650 651 static inline u64 global_rt_period(void) 652 { 653 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 654 } 655 656 static inline u64 global_rt_runtime(void) 657 { 658 if (sysctl_sched_rt_runtime < 0) 659 return RUNTIME_INF; 660 661 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 662 } 663 664 665 666 static inline int task_current(struct rq *rq, struct task_struct *p) 667 { 668 return rq->curr == p; 669 } 670 671 static inline int task_running(struct rq *rq, struct task_struct *p) 672 { 673 #ifdef CONFIG_SMP 674 return p->on_cpu; 675 #else 676 return task_current(rq, p); 677 #endif 678 } 679 680 681 #ifndef prepare_arch_switch 682 # define prepare_arch_switch(next) do { } while (0) 683 #endif 684 #ifndef finish_arch_switch 685 # define finish_arch_switch(prev) do { } while (0) 686 #endif 687 #ifndef finish_arch_post_lock_switch 688 # define finish_arch_post_lock_switch() do { } while (0) 689 #endif 690 691 #ifndef __ARCH_WANT_UNLOCKED_CTXSW 692 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 693 { 694 #ifdef CONFIG_SMP 695 /* 696 * We can optimise this out completely for !SMP, because the 697 * SMP rebalancing from interrupt is the only thing that cares 698 * here. 699 */ 700 next->on_cpu = 1; 701 #endif 702 } 703 704 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 705 { 706 #ifdef CONFIG_SMP 707 /* 708 * After ->on_cpu is cleared, the task can be moved to a different CPU. 709 * We must ensure this doesn't happen until the switch is completely 710 * finished. 711 */ 712 smp_wmb(); 713 prev->on_cpu = 0; 714 #endif 715 #ifdef CONFIG_DEBUG_SPINLOCK 716 /* this is a valid case when another task releases the spinlock */ 717 rq->lock.owner = current; 718 #endif 719 /* 720 * If we are tracking spinlock dependencies then we have to 721 * fix up the runqueue lock - which gets 'carried over' from 722 * prev into current: 723 */ 724 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); 725 726 raw_spin_unlock_irq(&rq->lock); 727 } 728 729 #else /* __ARCH_WANT_UNLOCKED_CTXSW */ 730 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 731 { 732 #ifdef CONFIG_SMP 733 /* 734 * We can optimise this out completely for !SMP, because the 735 * SMP rebalancing from interrupt is the only thing that cares 736 * here. 737 */ 738 next->on_cpu = 1; 739 #endif 740 raw_spin_unlock(&rq->lock); 741 } 742 743 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 744 { 745 #ifdef CONFIG_SMP 746 /* 747 * After ->on_cpu is cleared, the task can be moved to a different CPU. 748 * We must ensure this doesn't happen until the switch is completely 749 * finished. 750 */ 751 smp_wmb(); 752 prev->on_cpu = 0; 753 #endif 754 local_irq_enable(); 755 } 756 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ 757 758 759 static inline void update_load_add(struct load_weight *lw, unsigned long inc) 760 { 761 lw->weight += inc; 762 lw->inv_weight = 0; 763 } 764 765 static inline void update_load_sub(struct load_weight *lw, unsigned long dec) 766 { 767 lw->weight -= dec; 768 lw->inv_weight = 0; 769 } 770 771 static inline void update_load_set(struct load_weight *lw, unsigned long w) 772 { 773 lw->weight = w; 774 lw->inv_weight = 0; 775 } 776 777 /* 778 * To aid in avoiding the subversion of "niceness" due to uneven distribution 779 * of tasks with abnormal "nice" values across CPUs the contribution that 780 * each task makes to its run queue's load is weighted according to its 781 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 782 * scaled version of the new time slice allocation that they receive on time 783 * slice expiry etc. 784 */ 785 786 #define WEIGHT_IDLEPRIO 3 787 #define WMULT_IDLEPRIO 1431655765 788 789 /* 790 * Nice levels are multiplicative, with a gentle 10% change for every 791 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to 792 * nice 1, it will get ~10% less CPU time than another CPU-bound task 793 * that remained on nice 0. 794 * 795 * The "10% effect" is relative and cumulative: from _any_ nice level, 796 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level 797 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. 798 * If a task goes up by ~10% and another task goes down by ~10% then 799 * the relative distance between them is ~25%.) 800 */ 801 static const int prio_to_weight[40] = { 802 /* -20 */ 88761, 71755, 56483, 46273, 36291, 803 /* -15 */ 29154, 23254, 18705, 14949, 11916, 804 /* -10 */ 9548, 7620, 6100, 4904, 3906, 805 /* -5 */ 3121, 2501, 1991, 1586, 1277, 806 /* 0 */ 1024, 820, 655, 526, 423, 807 /* 5 */ 335, 272, 215, 172, 137, 808 /* 10 */ 110, 87, 70, 56, 45, 809 /* 15 */ 36, 29, 23, 18, 15, 810 }; 811 812 /* 813 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. 814 * 815 * In cases where the weight does not change often, we can use the 816 * precalculated inverse to speed up arithmetics by turning divisions 817 * into multiplications: 818 */ 819 static const u32 prio_to_wmult[40] = { 820 /* -20 */ 48388, 59856, 76040, 92818, 118348, 821 /* -15 */ 147320, 184698, 229616, 287308, 360437, 822 /* -10 */ 449829, 563644, 704093, 875809, 1099582, 823 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, 824 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, 825 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, 826 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, 827 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, 828 }; 829 830 /* Time spent by the tasks of the cpu accounting group executing in ... */ 831 enum cpuacct_stat_index { 832 CPUACCT_STAT_USER, /* ... user mode */ 833 CPUACCT_STAT_SYSTEM, /* ... kernel mode */ 834 835 CPUACCT_STAT_NSTATS, 836 }; 837 838 839 #define sched_class_highest (&stop_sched_class) 840 #define for_each_class(class) \ 841 for (class = sched_class_highest; class; class = class->next) 842 843 extern const struct sched_class stop_sched_class; 844 extern const struct sched_class rt_sched_class; 845 extern const struct sched_class fair_sched_class; 846 extern const struct sched_class idle_sched_class; 847 848 849 #ifdef CONFIG_SMP 850 851 extern void trigger_load_balance(struct rq *rq, int cpu); 852 extern void idle_balance(int this_cpu, struct rq *this_rq); 853 854 #else /* CONFIG_SMP */ 855 856 static inline void idle_balance(int cpu, struct rq *rq) 857 { 858 } 859 860 #endif 861 862 extern void sysrq_sched_debug_show(void); 863 extern void sched_init_granularity(void); 864 extern void update_max_interval(void); 865 extern void update_group_power(struct sched_domain *sd, int cpu); 866 extern int update_runtime(struct notifier_block *nfb, unsigned long action, void *hcpu); 867 extern void init_sched_rt_class(void); 868 extern void init_sched_fair_class(void); 869 870 extern void resched_task(struct task_struct *p); 871 extern void resched_cpu(int cpu); 872 873 extern struct rt_bandwidth def_rt_bandwidth; 874 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 875 876 extern void update_idle_cpu_load(struct rq *this_rq); 877 878 #ifdef CONFIG_CGROUP_CPUACCT 879 #include <linux/cgroup.h> 880 /* track cpu usage of a group of tasks and its child groups */ 881 struct cpuacct { 882 struct cgroup_subsys_state css; 883 /* cpuusage holds pointer to a u64-type object on every cpu */ 884 u64 __percpu *cpuusage; 885 struct kernel_cpustat __percpu *cpustat; 886 }; 887 888 extern struct cgroup_subsys cpuacct_subsys; 889 extern struct cpuacct root_cpuacct; 890 891 /* return cpu accounting group corresponding to this container */ 892 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) 893 { 894 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), 895 struct cpuacct, css); 896 } 897 898 /* return cpu accounting group to which this task belongs */ 899 static inline struct cpuacct *task_ca(struct task_struct *tsk) 900 { 901 return container_of(task_subsys_state(tsk, cpuacct_subsys_id), 902 struct cpuacct, css); 903 } 904 905 static inline struct cpuacct *parent_ca(struct cpuacct *ca) 906 { 907 if (!ca || !ca->css.cgroup->parent) 908 return NULL; 909 return cgroup_ca(ca->css.cgroup->parent); 910 } 911 912 extern void cpuacct_charge(struct task_struct *tsk, u64 cputime); 913 #else 914 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} 915 #endif 916 917 #ifdef CONFIG_PARAVIRT 918 static inline u64 steal_ticks(u64 steal) 919 { 920 if (unlikely(steal > NSEC_PER_SEC)) 921 return div_u64(steal, TICK_NSEC); 922 923 return __iter_div_u64_rem(steal, TICK_NSEC, &steal); 924 } 925 #endif 926 927 static inline void inc_nr_running(struct rq *rq) 928 { 929 rq->nr_running++; 930 } 931 932 static inline void dec_nr_running(struct rq *rq) 933 { 934 rq->nr_running--; 935 } 936 937 extern void update_rq_clock(struct rq *rq); 938 939 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 940 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 941 942 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); 943 944 extern const_debug unsigned int sysctl_sched_time_avg; 945 extern const_debug unsigned int sysctl_sched_nr_migrate; 946 extern const_debug unsigned int sysctl_sched_migration_cost; 947 948 static inline u64 sched_avg_period(void) 949 { 950 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; 951 } 952 953 #ifdef CONFIG_SCHED_HRTICK 954 955 /* 956 * Use hrtick when: 957 * - enabled by features 958 * - hrtimer is actually high res 959 */ 960 static inline int hrtick_enabled(struct rq *rq) 961 { 962 if (!sched_feat(HRTICK)) 963 return 0; 964 if (!cpu_active(cpu_of(rq))) 965 return 0; 966 return hrtimer_is_hres_active(&rq->hrtick_timer); 967 } 968 969 void hrtick_start(struct rq *rq, u64 delay); 970 971 #else 972 973 static inline int hrtick_enabled(struct rq *rq) 974 { 975 return 0; 976 } 977 978 #endif /* CONFIG_SCHED_HRTICK */ 979 980 #ifdef CONFIG_SMP 981 extern void sched_avg_update(struct rq *rq); 982 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) 983 { 984 rq->rt_avg += rt_delta; 985 sched_avg_update(rq); 986 } 987 #else 988 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { } 989 static inline void sched_avg_update(struct rq *rq) { } 990 #endif 991 992 extern void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period); 993 994 #ifdef CONFIG_SMP 995 #ifdef CONFIG_PREEMPT 996 997 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2); 998 999 /* 1000 * fair double_lock_balance: Safely acquires both rq->locks in a fair 1001 * way at the expense of forcing extra atomic operations in all 1002 * invocations. This assures that the double_lock is acquired using the 1003 * same underlying policy as the spinlock_t on this architecture, which 1004 * reduces latency compared to the unfair variant below. However, it 1005 * also adds more overhead and therefore may reduce throughput. 1006 */ 1007 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1008 __releases(this_rq->lock) 1009 __acquires(busiest->lock) 1010 __acquires(this_rq->lock) 1011 { 1012 raw_spin_unlock(&this_rq->lock); 1013 double_rq_lock(this_rq, busiest); 1014 1015 return 1; 1016 } 1017 1018 #else 1019 /* 1020 * Unfair double_lock_balance: Optimizes throughput at the expense of 1021 * latency by eliminating extra atomic operations when the locks are 1022 * already in proper order on entry. This favors lower cpu-ids and will 1023 * grant the double lock to lower cpus over higher ids under contention, 1024 * regardless of entry order into the function. 1025 */ 1026 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1027 __releases(this_rq->lock) 1028 __acquires(busiest->lock) 1029 __acquires(this_rq->lock) 1030 { 1031 int ret = 0; 1032 1033 if (unlikely(!raw_spin_trylock(&busiest->lock))) { 1034 if (busiest < this_rq) { 1035 raw_spin_unlock(&this_rq->lock); 1036 raw_spin_lock(&busiest->lock); 1037 raw_spin_lock_nested(&this_rq->lock, 1038 SINGLE_DEPTH_NESTING); 1039 ret = 1; 1040 } else 1041 raw_spin_lock_nested(&busiest->lock, 1042 SINGLE_DEPTH_NESTING); 1043 } 1044 return ret; 1045 } 1046 1047 #endif /* CONFIG_PREEMPT */ 1048 1049 /* 1050 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 1051 */ 1052 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 1053 { 1054 if (unlikely(!irqs_disabled())) { 1055 /* printk() doesn't work good under rq->lock */ 1056 raw_spin_unlock(&this_rq->lock); 1057 BUG_ON(1); 1058 } 1059 1060 return _double_lock_balance(this_rq, busiest); 1061 } 1062 1063 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 1064 __releases(busiest->lock) 1065 { 1066 raw_spin_unlock(&busiest->lock); 1067 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1068 } 1069 1070 /* 1071 * double_rq_lock - safely lock two runqueues 1072 * 1073 * Note this does not disable interrupts like task_rq_lock, 1074 * you need to do so manually before calling. 1075 */ 1076 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1077 __acquires(rq1->lock) 1078 __acquires(rq2->lock) 1079 { 1080 BUG_ON(!irqs_disabled()); 1081 if (rq1 == rq2) { 1082 raw_spin_lock(&rq1->lock); 1083 __acquire(rq2->lock); /* Fake it out ;) */ 1084 } else { 1085 if (rq1 < rq2) { 1086 raw_spin_lock(&rq1->lock); 1087 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); 1088 } else { 1089 raw_spin_lock(&rq2->lock); 1090 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); 1091 } 1092 } 1093 } 1094 1095 /* 1096 * double_rq_unlock - safely unlock two runqueues 1097 * 1098 * Note this does not restore interrupts like task_rq_unlock, 1099 * you need to do so manually after calling. 1100 */ 1101 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1102 __releases(rq1->lock) 1103 __releases(rq2->lock) 1104 { 1105 raw_spin_unlock(&rq1->lock); 1106 if (rq1 != rq2) 1107 raw_spin_unlock(&rq2->lock); 1108 else 1109 __release(rq2->lock); 1110 } 1111 1112 #else /* CONFIG_SMP */ 1113 1114 /* 1115 * double_rq_lock - safely lock two runqueues 1116 * 1117 * Note this does not disable interrupts like task_rq_lock, 1118 * you need to do so manually before calling. 1119 */ 1120 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1121 __acquires(rq1->lock) 1122 __acquires(rq2->lock) 1123 { 1124 BUG_ON(!irqs_disabled()); 1125 BUG_ON(rq1 != rq2); 1126 raw_spin_lock(&rq1->lock); 1127 __acquire(rq2->lock); /* Fake it out ;) */ 1128 } 1129 1130 /* 1131 * double_rq_unlock - safely unlock two runqueues 1132 * 1133 * Note this does not restore interrupts like task_rq_unlock, 1134 * you need to do so manually after calling. 1135 */ 1136 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1137 __releases(rq1->lock) 1138 __releases(rq2->lock) 1139 { 1140 BUG_ON(rq1 != rq2); 1141 raw_spin_unlock(&rq1->lock); 1142 __release(rq2->lock); 1143 } 1144 1145 #endif 1146 1147 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 1148 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 1149 extern void print_cfs_stats(struct seq_file *m, int cpu); 1150 extern void print_rt_stats(struct seq_file *m, int cpu); 1151 1152 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 1153 extern void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq); 1154 1155 extern void account_cfs_bandwidth_used(int enabled, int was_enabled); 1156 1157 #ifdef CONFIG_NO_HZ 1158 enum rq_nohz_flag_bits { 1159 NOHZ_TICK_STOPPED, 1160 NOHZ_BALANCE_KICK, 1161 NOHZ_IDLE, 1162 }; 1163 1164 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 1165 #endif 1166 1167 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1168 1169 DECLARE_PER_CPU(u64, cpu_hardirq_time); 1170 DECLARE_PER_CPU(u64, cpu_softirq_time); 1171 1172 #ifndef CONFIG_64BIT 1173 DECLARE_PER_CPU(seqcount_t, irq_time_seq); 1174 1175 static inline void irq_time_write_begin(void) 1176 { 1177 __this_cpu_inc(irq_time_seq.sequence); 1178 smp_wmb(); 1179 } 1180 1181 static inline void irq_time_write_end(void) 1182 { 1183 smp_wmb(); 1184 __this_cpu_inc(irq_time_seq.sequence); 1185 } 1186 1187 static inline u64 irq_time_read(int cpu) 1188 { 1189 u64 irq_time; 1190 unsigned seq; 1191 1192 do { 1193 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 1194 irq_time = per_cpu(cpu_softirq_time, cpu) + 1195 per_cpu(cpu_hardirq_time, cpu); 1196 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 1197 1198 return irq_time; 1199 } 1200 #else /* CONFIG_64BIT */ 1201 static inline void irq_time_write_begin(void) 1202 { 1203 } 1204 1205 static inline void irq_time_write_end(void) 1206 { 1207 } 1208 1209 static inline u64 irq_time_read(int cpu) 1210 { 1211 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 1212 } 1213 #endif /* CONFIG_64BIT */ 1214 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 1215 1216