1 /* SPDX-License-Identifier: GPL-2.0 */ 2 3 #include <linux/sched.h> 4 #include <linux/sched/autogroup.h> 5 #include <linux/sched/sysctl.h> 6 #include <linux/sched/topology.h> 7 #include <linux/sched/rt.h> 8 #include <linux/sched/deadline.h> 9 #include <linux/sched/clock.h> 10 #include <linux/sched/wake_q.h> 11 #include <linux/sched/signal.h> 12 #include <linux/sched/numa_balancing.h> 13 #include <linux/sched/mm.h> 14 #include <linux/sched/cpufreq.h> 15 #include <linux/sched/stat.h> 16 #include <linux/sched/nohz.h> 17 #include <linux/sched/debug.h> 18 #include <linux/sched/hotplug.h> 19 #include <linux/sched/task.h> 20 #include <linux/sched/task_stack.h> 21 #include <linux/sched/cputime.h> 22 #include <linux/sched/init.h> 23 24 #include <linux/u64_stats_sync.h> 25 #include <linux/kernel_stat.h> 26 #include <linux/binfmts.h> 27 #include <linux/mutex.h> 28 #include <linux/spinlock.h> 29 #include <linux/stop_machine.h> 30 #include <linux/irq_work.h> 31 #include <linux/tick.h> 32 #include <linux/slab.h> 33 34 #ifdef CONFIG_PARAVIRT 35 #include <asm/paravirt.h> 36 #endif 37 38 #include "cpupri.h" 39 #include "cpudeadline.h" 40 #include "cpuacct.h" 41 42 #ifdef CONFIG_SCHED_DEBUG 43 # define SCHED_WARN_ON(x) WARN_ONCE(x, #x) 44 #else 45 # define SCHED_WARN_ON(x) ({ (void)(x), 0; }) 46 #endif 47 48 struct rq; 49 struct cpuidle_state; 50 51 /* task_struct::on_rq states: */ 52 #define TASK_ON_RQ_QUEUED 1 53 #define TASK_ON_RQ_MIGRATING 2 54 55 extern __read_mostly int scheduler_running; 56 57 extern unsigned long calc_load_update; 58 extern atomic_long_t calc_load_tasks; 59 60 extern void calc_global_load_tick(struct rq *this_rq); 61 extern long calc_load_fold_active(struct rq *this_rq, long adjust); 62 63 #ifdef CONFIG_SMP 64 extern void cpu_load_update_active(struct rq *this_rq); 65 #else 66 static inline void cpu_load_update_active(struct rq *this_rq) { } 67 #endif 68 69 /* 70 * Helpers for converting nanosecond timing to jiffy resolution 71 */ 72 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) 73 74 /* 75 * Increase resolution of nice-level calculations for 64-bit architectures. 76 * The extra resolution improves shares distribution and load balancing of 77 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup 78 * hierarchies, especially on larger systems. This is not a user-visible change 79 * and does not change the user-interface for setting shares/weights. 80 * 81 * We increase resolution only if we have enough bits to allow this increased 82 * resolution (i.e. 64bit). The costs for increasing resolution when 32bit are 83 * pretty high and the returns do not justify the increased costs. 84 * 85 * Really only required when CONFIG_FAIR_GROUP_SCHED is also set, but to 86 * increase coverage and consistency always enable it on 64bit platforms. 87 */ 88 #ifdef CONFIG_64BIT 89 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT) 90 # define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT) 91 # define scale_load_down(w) ((w) >> SCHED_FIXEDPOINT_SHIFT) 92 #else 93 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT) 94 # define scale_load(w) (w) 95 # define scale_load_down(w) (w) 96 #endif 97 98 /* 99 * Task weight (visible to users) and its load (invisible to users) have 100 * independent resolution, but they should be well calibrated. We use 101 * scale_load() and scale_load_down(w) to convert between them. The 102 * following must be true: 103 * 104 * scale_load(sched_prio_to_weight[USER_PRIO(NICE_TO_PRIO(0))]) == NICE_0_LOAD 105 * 106 */ 107 #define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT) 108 109 /* 110 * Single value that decides SCHED_DEADLINE internal math precision. 111 * 10 -> just above 1us 112 * 9 -> just above 0.5us 113 */ 114 #define DL_SCALE (10) 115 116 /* 117 * These are the 'tuning knobs' of the scheduler: 118 */ 119 120 /* 121 * single value that denotes runtime == period, ie unlimited time. 122 */ 123 #define RUNTIME_INF ((u64)~0ULL) 124 125 static inline int idle_policy(int policy) 126 { 127 return policy == SCHED_IDLE; 128 } 129 static inline int fair_policy(int policy) 130 { 131 return policy == SCHED_NORMAL || policy == SCHED_BATCH; 132 } 133 134 static inline int rt_policy(int policy) 135 { 136 return policy == SCHED_FIFO || policy == SCHED_RR; 137 } 138 139 static inline int dl_policy(int policy) 140 { 141 return policy == SCHED_DEADLINE; 142 } 143 static inline bool valid_policy(int policy) 144 { 145 return idle_policy(policy) || fair_policy(policy) || 146 rt_policy(policy) || dl_policy(policy); 147 } 148 149 static inline int task_has_rt_policy(struct task_struct *p) 150 { 151 return rt_policy(p->policy); 152 } 153 154 static inline int task_has_dl_policy(struct task_struct *p) 155 { 156 return dl_policy(p->policy); 157 } 158 159 /* 160 * Tells if entity @a should preempt entity @b. 161 */ 162 static inline bool 163 dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b) 164 { 165 return dl_time_before(a->deadline, b->deadline); 166 } 167 168 /* 169 * This is the priority-queue data structure of the RT scheduling class: 170 */ 171 struct rt_prio_array { 172 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ 173 struct list_head queue[MAX_RT_PRIO]; 174 }; 175 176 struct rt_bandwidth { 177 /* nests inside the rq lock: */ 178 raw_spinlock_t rt_runtime_lock; 179 ktime_t rt_period; 180 u64 rt_runtime; 181 struct hrtimer rt_period_timer; 182 unsigned int rt_period_active; 183 }; 184 185 void __dl_clear_params(struct task_struct *p); 186 187 /* 188 * To keep the bandwidth of -deadline tasks and groups under control 189 * we need some place where: 190 * - store the maximum -deadline bandwidth of the system (the group); 191 * - cache the fraction of that bandwidth that is currently allocated. 192 * 193 * This is all done in the data structure below. It is similar to the 194 * one used for RT-throttling (rt_bandwidth), with the main difference 195 * that, since here we are only interested in admission control, we 196 * do not decrease any runtime while the group "executes", neither we 197 * need a timer to replenish it. 198 * 199 * With respect to SMP, the bandwidth is given on a per-CPU basis, 200 * meaning that: 201 * - dl_bw (< 100%) is the bandwidth of the system (group) on each CPU; 202 * - dl_total_bw array contains, in the i-eth element, the currently 203 * allocated bandwidth on the i-eth CPU. 204 * Moreover, groups consume bandwidth on each CPU, while tasks only 205 * consume bandwidth on the CPU they're running on. 206 * Finally, dl_total_bw_cpu is used to cache the index of dl_total_bw 207 * that will be shown the next time the proc or cgroup controls will 208 * be red. It on its turn can be changed by writing on its own 209 * control. 210 */ 211 struct dl_bandwidth { 212 raw_spinlock_t dl_runtime_lock; 213 u64 dl_runtime; 214 u64 dl_period; 215 }; 216 217 static inline int dl_bandwidth_enabled(void) 218 { 219 return sysctl_sched_rt_runtime >= 0; 220 } 221 222 struct dl_bw { 223 raw_spinlock_t lock; 224 u64 bw, total_bw; 225 }; 226 227 static inline void __dl_update(struct dl_bw *dl_b, s64 bw); 228 229 static inline 230 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw, int cpus) 231 { 232 dl_b->total_bw -= tsk_bw; 233 __dl_update(dl_b, (s32)tsk_bw / cpus); 234 } 235 236 static inline 237 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus) 238 { 239 dl_b->total_bw += tsk_bw; 240 __dl_update(dl_b, -((s32)tsk_bw / cpus)); 241 } 242 243 static inline 244 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 245 { 246 return dl_b->bw != -1 && 247 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 248 } 249 250 void dl_change_utilization(struct task_struct *p, u64 new_bw); 251 extern void init_dl_bw(struct dl_bw *dl_b); 252 extern int sched_dl_global_validate(void); 253 extern void sched_dl_do_global(void); 254 extern int sched_dl_overflow(struct task_struct *p, int policy, 255 const struct sched_attr *attr); 256 extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr); 257 extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr); 258 extern bool __checkparam_dl(const struct sched_attr *attr); 259 extern void __dl_clear_params(struct task_struct *p); 260 extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr); 261 extern int dl_task_can_attach(struct task_struct *p, 262 const struct cpumask *cs_cpus_allowed); 263 extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, 264 const struct cpumask *trial); 265 extern bool dl_cpu_busy(unsigned int cpu); 266 267 #ifdef CONFIG_CGROUP_SCHED 268 269 #include <linux/cgroup.h> 270 271 struct cfs_rq; 272 struct rt_rq; 273 274 extern struct list_head task_groups; 275 276 struct cfs_bandwidth { 277 #ifdef CONFIG_CFS_BANDWIDTH 278 raw_spinlock_t lock; 279 ktime_t period; 280 u64 quota, runtime; 281 s64 hierarchical_quota; 282 u64 runtime_expires; 283 284 int idle, period_active; 285 struct hrtimer period_timer, slack_timer; 286 struct list_head throttled_cfs_rq; 287 288 /* statistics */ 289 int nr_periods, nr_throttled; 290 u64 throttled_time; 291 #endif 292 }; 293 294 /* task group related information */ 295 struct task_group { 296 struct cgroup_subsys_state css; 297 298 #ifdef CONFIG_FAIR_GROUP_SCHED 299 /* schedulable entities of this group on each cpu */ 300 struct sched_entity **se; 301 /* runqueue "owned" by this group on each cpu */ 302 struct cfs_rq **cfs_rq; 303 unsigned long shares; 304 305 #ifdef CONFIG_SMP 306 /* 307 * load_avg can be heavily contended at clock tick time, so put 308 * it in its own cacheline separated from the fields above which 309 * will also be accessed at each tick. 310 */ 311 atomic_long_t load_avg ____cacheline_aligned; 312 #endif 313 #endif 314 315 #ifdef CONFIG_RT_GROUP_SCHED 316 struct sched_rt_entity **rt_se; 317 struct rt_rq **rt_rq; 318 319 struct rt_bandwidth rt_bandwidth; 320 #endif 321 322 struct rcu_head rcu; 323 struct list_head list; 324 325 struct task_group *parent; 326 struct list_head siblings; 327 struct list_head children; 328 329 #ifdef CONFIG_SCHED_AUTOGROUP 330 struct autogroup *autogroup; 331 #endif 332 333 struct cfs_bandwidth cfs_bandwidth; 334 }; 335 336 #ifdef CONFIG_FAIR_GROUP_SCHED 337 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD 338 339 /* 340 * A weight of 0 or 1 can cause arithmetics problems. 341 * A weight of a cfs_rq is the sum of weights of which entities 342 * are queued on this cfs_rq, so a weight of a entity should not be 343 * too large, so as the shares value of a task group. 344 * (The default weight is 1024 - so there's no practical 345 * limitation from this.) 346 */ 347 #define MIN_SHARES (1UL << 1) 348 #define MAX_SHARES (1UL << 18) 349 #endif 350 351 typedef int (*tg_visitor)(struct task_group *, void *); 352 353 extern int walk_tg_tree_from(struct task_group *from, 354 tg_visitor down, tg_visitor up, void *data); 355 356 /* 357 * Iterate the full tree, calling @down when first entering a node and @up when 358 * leaving it for the final time. 359 * 360 * Caller must hold rcu_lock or sufficient equivalent. 361 */ 362 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) 363 { 364 return walk_tg_tree_from(&root_task_group, down, up, data); 365 } 366 367 extern int tg_nop(struct task_group *tg, void *data); 368 369 extern void free_fair_sched_group(struct task_group *tg); 370 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent); 371 extern void online_fair_sched_group(struct task_group *tg); 372 extern void unregister_fair_sched_group(struct task_group *tg); 373 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 374 struct sched_entity *se, int cpu, 375 struct sched_entity *parent); 376 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 377 378 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b); 379 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 380 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq); 381 382 extern void free_rt_sched_group(struct task_group *tg); 383 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent); 384 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 385 struct sched_rt_entity *rt_se, int cpu, 386 struct sched_rt_entity *parent); 387 extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us); 388 extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us); 389 extern long sched_group_rt_runtime(struct task_group *tg); 390 extern long sched_group_rt_period(struct task_group *tg); 391 extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk); 392 393 extern struct task_group *sched_create_group(struct task_group *parent); 394 extern void sched_online_group(struct task_group *tg, 395 struct task_group *parent); 396 extern void sched_destroy_group(struct task_group *tg); 397 extern void sched_offline_group(struct task_group *tg); 398 399 extern void sched_move_task(struct task_struct *tsk); 400 401 #ifdef CONFIG_FAIR_GROUP_SCHED 402 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 403 404 #ifdef CONFIG_SMP 405 extern void set_task_rq_fair(struct sched_entity *se, 406 struct cfs_rq *prev, struct cfs_rq *next); 407 #else /* !CONFIG_SMP */ 408 static inline void set_task_rq_fair(struct sched_entity *se, 409 struct cfs_rq *prev, struct cfs_rq *next) { } 410 #endif /* CONFIG_SMP */ 411 #endif /* CONFIG_FAIR_GROUP_SCHED */ 412 413 #else /* CONFIG_CGROUP_SCHED */ 414 415 struct cfs_bandwidth { }; 416 417 #endif /* CONFIG_CGROUP_SCHED */ 418 419 /* CFS-related fields in a runqueue */ 420 struct cfs_rq { 421 struct load_weight load; 422 unsigned int nr_running, h_nr_running; 423 424 u64 exec_clock; 425 u64 min_vruntime; 426 #ifndef CONFIG_64BIT 427 u64 min_vruntime_copy; 428 #endif 429 430 struct rb_root_cached tasks_timeline; 431 432 /* 433 * 'curr' points to currently running entity on this cfs_rq. 434 * It is set to NULL otherwise (i.e when none are currently running). 435 */ 436 struct sched_entity *curr, *next, *last, *skip; 437 438 #ifdef CONFIG_SCHED_DEBUG 439 unsigned int nr_spread_over; 440 #endif 441 442 #ifdef CONFIG_SMP 443 /* 444 * CFS load tracking 445 */ 446 struct sched_avg avg; 447 u64 runnable_load_sum; 448 unsigned long runnable_load_avg; 449 #ifdef CONFIG_FAIR_GROUP_SCHED 450 unsigned long tg_load_avg_contrib; 451 unsigned long propagate_avg; 452 #endif 453 atomic_long_t removed_load_avg, removed_util_avg; 454 #ifndef CONFIG_64BIT 455 u64 load_last_update_time_copy; 456 #endif 457 458 #ifdef CONFIG_FAIR_GROUP_SCHED 459 /* 460 * h_load = weight * f(tg) 461 * 462 * Where f(tg) is the recursive weight fraction assigned to 463 * this group. 464 */ 465 unsigned long h_load; 466 u64 last_h_load_update; 467 struct sched_entity *h_load_next; 468 #endif /* CONFIG_FAIR_GROUP_SCHED */ 469 #endif /* CONFIG_SMP */ 470 471 #ifdef CONFIG_FAIR_GROUP_SCHED 472 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ 473 474 /* 475 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in 476 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities 477 * (like users, containers etc.) 478 * 479 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This 480 * list is used during load balance. 481 */ 482 int on_list; 483 struct list_head leaf_cfs_rq_list; 484 struct task_group *tg; /* group that "owns" this runqueue */ 485 486 #ifdef CONFIG_CFS_BANDWIDTH 487 int runtime_enabled; 488 u64 runtime_expires; 489 s64 runtime_remaining; 490 491 u64 throttled_clock, throttled_clock_task; 492 u64 throttled_clock_task_time; 493 int throttled, throttle_count; 494 struct list_head throttled_list; 495 #endif /* CONFIG_CFS_BANDWIDTH */ 496 #endif /* CONFIG_FAIR_GROUP_SCHED */ 497 }; 498 499 static inline int rt_bandwidth_enabled(void) 500 { 501 return sysctl_sched_rt_runtime >= 0; 502 } 503 504 /* RT IPI pull logic requires IRQ_WORK */ 505 #ifdef CONFIG_IRQ_WORK 506 # define HAVE_RT_PUSH_IPI 507 #endif 508 509 /* Real-Time classes' related field in a runqueue: */ 510 struct rt_rq { 511 struct rt_prio_array active; 512 unsigned int rt_nr_running; 513 unsigned int rr_nr_running; 514 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 515 struct { 516 int curr; /* highest queued rt task prio */ 517 #ifdef CONFIG_SMP 518 int next; /* next highest */ 519 #endif 520 } highest_prio; 521 #endif 522 #ifdef CONFIG_SMP 523 unsigned long rt_nr_migratory; 524 unsigned long rt_nr_total; 525 int overloaded; 526 struct plist_head pushable_tasks; 527 #ifdef HAVE_RT_PUSH_IPI 528 int push_flags; 529 int push_cpu; 530 struct irq_work push_work; 531 raw_spinlock_t push_lock; 532 #endif 533 #endif /* CONFIG_SMP */ 534 int rt_queued; 535 536 int rt_throttled; 537 u64 rt_time; 538 u64 rt_runtime; 539 /* Nests inside the rq lock: */ 540 raw_spinlock_t rt_runtime_lock; 541 542 #ifdef CONFIG_RT_GROUP_SCHED 543 unsigned long rt_nr_boosted; 544 545 struct rq *rq; 546 struct task_group *tg; 547 #endif 548 }; 549 550 /* Deadline class' related fields in a runqueue */ 551 struct dl_rq { 552 /* runqueue is an rbtree, ordered by deadline */ 553 struct rb_root_cached root; 554 555 unsigned long dl_nr_running; 556 557 #ifdef CONFIG_SMP 558 /* 559 * Deadline values of the currently executing and the 560 * earliest ready task on this rq. Caching these facilitates 561 * the decision wether or not a ready but not running task 562 * should migrate somewhere else. 563 */ 564 struct { 565 u64 curr; 566 u64 next; 567 } earliest_dl; 568 569 unsigned long dl_nr_migratory; 570 int overloaded; 571 572 /* 573 * Tasks on this rq that can be pushed away. They are kept in 574 * an rb-tree, ordered by tasks' deadlines, with caching 575 * of the leftmost (earliest deadline) element. 576 */ 577 struct rb_root_cached pushable_dl_tasks_root; 578 #else 579 struct dl_bw dl_bw; 580 #endif 581 /* 582 * "Active utilization" for this runqueue: increased when a 583 * task wakes up (becomes TASK_RUNNING) and decreased when a 584 * task blocks 585 */ 586 u64 running_bw; 587 588 /* 589 * Utilization of the tasks "assigned" to this runqueue (including 590 * the tasks that are in runqueue and the tasks that executed on this 591 * CPU and blocked). Increased when a task moves to this runqueue, and 592 * decreased when the task moves away (migrates, changes scheduling 593 * policy, or terminates). 594 * This is needed to compute the "inactive utilization" for the 595 * runqueue (inactive utilization = this_bw - running_bw). 596 */ 597 u64 this_bw; 598 u64 extra_bw; 599 600 /* 601 * Inverse of the fraction of CPU utilization that can be reclaimed 602 * by the GRUB algorithm. 603 */ 604 u64 bw_ratio; 605 }; 606 607 #ifdef CONFIG_SMP 608 609 static inline bool sched_asym_prefer(int a, int b) 610 { 611 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b); 612 } 613 614 /* 615 * We add the notion of a root-domain which will be used to define per-domain 616 * variables. Each exclusive cpuset essentially defines an island domain by 617 * fully partitioning the member cpus from any other cpuset. Whenever a new 618 * exclusive cpuset is created, we also create and attach a new root-domain 619 * object. 620 * 621 */ 622 struct root_domain { 623 atomic_t refcount; 624 atomic_t rto_count; 625 struct rcu_head rcu; 626 cpumask_var_t span; 627 cpumask_var_t online; 628 629 /* Indicate more than one runnable task for any CPU */ 630 bool overload; 631 632 /* 633 * The bit corresponding to a CPU gets set here if such CPU has more 634 * than one runnable -deadline task (as it is below for RT tasks). 635 */ 636 cpumask_var_t dlo_mask; 637 atomic_t dlo_count; 638 struct dl_bw dl_bw; 639 struct cpudl cpudl; 640 641 /* 642 * The "RT overload" flag: it gets set if a CPU has more than 643 * one runnable RT task. 644 */ 645 cpumask_var_t rto_mask; 646 struct cpupri cpupri; 647 648 unsigned long max_cpu_capacity; 649 }; 650 651 extern struct root_domain def_root_domain; 652 extern struct mutex sched_domains_mutex; 653 654 extern void init_defrootdomain(void); 655 extern int sched_init_domains(const struct cpumask *cpu_map); 656 extern void rq_attach_root(struct rq *rq, struct root_domain *rd); 657 658 #endif /* CONFIG_SMP */ 659 660 /* 661 * This is the main, per-CPU runqueue data structure. 662 * 663 * Locking rule: those places that want to lock multiple runqueues 664 * (such as the load balancing or the thread migration code), lock 665 * acquire operations must be ordered by ascending &runqueue. 666 */ 667 struct rq { 668 /* runqueue lock: */ 669 raw_spinlock_t lock; 670 671 /* 672 * nr_running and cpu_load should be in the same cacheline because 673 * remote CPUs use both these fields when doing load calculation. 674 */ 675 unsigned int nr_running; 676 #ifdef CONFIG_NUMA_BALANCING 677 unsigned int nr_numa_running; 678 unsigned int nr_preferred_running; 679 #endif 680 #define CPU_LOAD_IDX_MAX 5 681 unsigned long cpu_load[CPU_LOAD_IDX_MAX]; 682 #ifdef CONFIG_NO_HZ_COMMON 683 #ifdef CONFIG_SMP 684 unsigned long last_load_update_tick; 685 #endif /* CONFIG_SMP */ 686 unsigned long nohz_flags; 687 #endif /* CONFIG_NO_HZ_COMMON */ 688 #ifdef CONFIG_NO_HZ_FULL 689 unsigned long last_sched_tick; 690 #endif 691 /* capture load from *all* tasks on this cpu: */ 692 struct load_weight load; 693 unsigned long nr_load_updates; 694 u64 nr_switches; 695 696 struct cfs_rq cfs; 697 struct rt_rq rt; 698 struct dl_rq dl; 699 700 #ifdef CONFIG_FAIR_GROUP_SCHED 701 /* list of leaf cfs_rq on this cpu: */ 702 struct list_head leaf_cfs_rq_list; 703 struct list_head *tmp_alone_branch; 704 #endif /* CONFIG_FAIR_GROUP_SCHED */ 705 706 /* 707 * This is part of a global counter where only the total sum 708 * over all CPUs matters. A task can increase this counter on 709 * one CPU and if it got migrated afterwards it may decrease 710 * it on another CPU. Always updated under the runqueue lock: 711 */ 712 unsigned long nr_uninterruptible; 713 714 struct task_struct *curr, *idle, *stop; 715 unsigned long next_balance; 716 struct mm_struct *prev_mm; 717 718 unsigned int clock_update_flags; 719 u64 clock; 720 u64 clock_task; 721 722 atomic_t nr_iowait; 723 724 #ifdef CONFIG_SMP 725 struct root_domain *rd; 726 struct sched_domain *sd; 727 728 unsigned long cpu_capacity; 729 unsigned long cpu_capacity_orig; 730 731 struct callback_head *balance_callback; 732 733 unsigned char idle_balance; 734 /* For active balancing */ 735 int active_balance; 736 int push_cpu; 737 struct cpu_stop_work active_balance_work; 738 /* cpu of this runqueue: */ 739 int cpu; 740 int online; 741 742 struct list_head cfs_tasks; 743 744 u64 rt_avg; 745 u64 age_stamp; 746 u64 idle_stamp; 747 u64 avg_idle; 748 749 /* This is used to determine avg_idle's max value */ 750 u64 max_idle_balance_cost; 751 #endif 752 753 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 754 u64 prev_irq_time; 755 #endif 756 #ifdef CONFIG_PARAVIRT 757 u64 prev_steal_time; 758 #endif 759 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 760 u64 prev_steal_time_rq; 761 #endif 762 763 /* calc_load related fields */ 764 unsigned long calc_load_update; 765 long calc_load_active; 766 767 #ifdef CONFIG_SCHED_HRTICK 768 #ifdef CONFIG_SMP 769 int hrtick_csd_pending; 770 call_single_data_t hrtick_csd; 771 #endif 772 struct hrtimer hrtick_timer; 773 #endif 774 775 #ifdef CONFIG_SCHEDSTATS 776 /* latency stats */ 777 struct sched_info rq_sched_info; 778 unsigned long long rq_cpu_time; 779 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 780 781 /* sys_sched_yield() stats */ 782 unsigned int yld_count; 783 784 /* schedule() stats */ 785 unsigned int sched_count; 786 unsigned int sched_goidle; 787 788 /* try_to_wake_up() stats */ 789 unsigned int ttwu_count; 790 unsigned int ttwu_local; 791 #endif 792 793 #ifdef CONFIG_SMP 794 struct llist_head wake_list; 795 #endif 796 797 #ifdef CONFIG_CPU_IDLE 798 /* Must be inspected within a rcu lock section */ 799 struct cpuidle_state *idle_state; 800 #endif 801 }; 802 803 static inline int cpu_of(struct rq *rq) 804 { 805 #ifdef CONFIG_SMP 806 return rq->cpu; 807 #else 808 return 0; 809 #endif 810 } 811 812 813 #ifdef CONFIG_SCHED_SMT 814 815 extern struct static_key_false sched_smt_present; 816 817 extern void __update_idle_core(struct rq *rq); 818 819 static inline void update_idle_core(struct rq *rq) 820 { 821 if (static_branch_unlikely(&sched_smt_present)) 822 __update_idle_core(rq); 823 } 824 825 #else 826 static inline void update_idle_core(struct rq *rq) { } 827 #endif 828 829 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 830 831 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 832 #define this_rq() this_cpu_ptr(&runqueues) 833 #define task_rq(p) cpu_rq(task_cpu(p)) 834 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 835 #define raw_rq() raw_cpu_ptr(&runqueues) 836 837 static inline u64 __rq_clock_broken(struct rq *rq) 838 { 839 return READ_ONCE(rq->clock); 840 } 841 842 /* 843 * rq::clock_update_flags bits 844 * 845 * %RQCF_REQ_SKIP - will request skipping of clock update on the next 846 * call to __schedule(). This is an optimisation to avoid 847 * neighbouring rq clock updates. 848 * 849 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is 850 * in effect and calls to update_rq_clock() are being ignored. 851 * 852 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been 853 * made to update_rq_clock() since the last time rq::lock was pinned. 854 * 855 * If inside of __schedule(), clock_update_flags will have been 856 * shifted left (a left shift is a cheap operation for the fast path 857 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use, 858 * 859 * if (rq-clock_update_flags >= RQCF_UPDATED) 860 * 861 * to check if %RQCF_UPADTED is set. It'll never be shifted more than 862 * one position though, because the next rq_unpin_lock() will shift it 863 * back. 864 */ 865 #define RQCF_REQ_SKIP 0x01 866 #define RQCF_ACT_SKIP 0x02 867 #define RQCF_UPDATED 0x04 868 869 static inline void assert_clock_updated(struct rq *rq) 870 { 871 /* 872 * The only reason for not seeing a clock update since the 873 * last rq_pin_lock() is if we're currently skipping updates. 874 */ 875 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP); 876 } 877 878 static inline u64 rq_clock(struct rq *rq) 879 { 880 lockdep_assert_held(&rq->lock); 881 assert_clock_updated(rq); 882 883 return rq->clock; 884 } 885 886 static inline u64 rq_clock_task(struct rq *rq) 887 { 888 lockdep_assert_held(&rq->lock); 889 assert_clock_updated(rq); 890 891 return rq->clock_task; 892 } 893 894 static inline void rq_clock_skip_update(struct rq *rq, bool skip) 895 { 896 lockdep_assert_held(&rq->lock); 897 if (skip) 898 rq->clock_update_flags |= RQCF_REQ_SKIP; 899 else 900 rq->clock_update_flags &= ~RQCF_REQ_SKIP; 901 } 902 903 struct rq_flags { 904 unsigned long flags; 905 struct pin_cookie cookie; 906 #ifdef CONFIG_SCHED_DEBUG 907 /* 908 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the 909 * current pin context is stashed here in case it needs to be 910 * restored in rq_repin_lock(). 911 */ 912 unsigned int clock_update_flags; 913 #endif 914 }; 915 916 static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf) 917 { 918 rf->cookie = lockdep_pin_lock(&rq->lock); 919 920 #ifdef CONFIG_SCHED_DEBUG 921 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP); 922 rf->clock_update_flags = 0; 923 #endif 924 } 925 926 static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf) 927 { 928 #ifdef CONFIG_SCHED_DEBUG 929 if (rq->clock_update_flags > RQCF_ACT_SKIP) 930 rf->clock_update_flags = RQCF_UPDATED; 931 #endif 932 933 lockdep_unpin_lock(&rq->lock, rf->cookie); 934 } 935 936 static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf) 937 { 938 lockdep_repin_lock(&rq->lock, rf->cookie); 939 940 #ifdef CONFIG_SCHED_DEBUG 941 /* 942 * Restore the value we stashed in @rf for this pin context. 943 */ 944 rq->clock_update_flags |= rf->clock_update_flags; 945 #endif 946 } 947 948 #ifdef CONFIG_NUMA 949 enum numa_topology_type { 950 NUMA_DIRECT, 951 NUMA_GLUELESS_MESH, 952 NUMA_BACKPLANE, 953 }; 954 extern enum numa_topology_type sched_numa_topology_type; 955 extern int sched_max_numa_distance; 956 extern bool find_numa_distance(int distance); 957 #endif 958 959 #ifdef CONFIG_NUMA 960 extern void sched_init_numa(void); 961 extern void sched_domains_numa_masks_set(unsigned int cpu); 962 extern void sched_domains_numa_masks_clear(unsigned int cpu); 963 #else 964 static inline void sched_init_numa(void) { } 965 static inline void sched_domains_numa_masks_set(unsigned int cpu) { } 966 static inline void sched_domains_numa_masks_clear(unsigned int cpu) { } 967 #endif 968 969 #ifdef CONFIG_NUMA_BALANCING 970 /* The regions in numa_faults array from task_struct */ 971 enum numa_faults_stats { 972 NUMA_MEM = 0, 973 NUMA_CPU, 974 NUMA_MEMBUF, 975 NUMA_CPUBUF 976 }; 977 extern void sched_setnuma(struct task_struct *p, int node); 978 extern int migrate_task_to(struct task_struct *p, int cpu); 979 extern int migrate_swap(struct task_struct *, struct task_struct *); 980 #endif /* CONFIG_NUMA_BALANCING */ 981 982 #ifdef CONFIG_SMP 983 984 static inline void 985 queue_balance_callback(struct rq *rq, 986 struct callback_head *head, 987 void (*func)(struct rq *rq)) 988 { 989 lockdep_assert_held(&rq->lock); 990 991 if (unlikely(head->next)) 992 return; 993 994 head->func = (void (*)(struct callback_head *))func; 995 head->next = rq->balance_callback; 996 rq->balance_callback = head; 997 } 998 999 extern void sched_ttwu_pending(void); 1000 1001 #define rcu_dereference_check_sched_domain(p) \ 1002 rcu_dereference_check((p), \ 1003 lockdep_is_held(&sched_domains_mutex)) 1004 1005 /* 1006 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 1007 * See detach_destroy_domains: synchronize_sched for details. 1008 * 1009 * The domain tree of any CPU may only be accessed from within 1010 * preempt-disabled sections. 1011 */ 1012 #define for_each_domain(cpu, __sd) \ 1013 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 1014 __sd; __sd = __sd->parent) 1015 1016 #define for_each_lower_domain(sd) for (; sd; sd = sd->child) 1017 1018 /** 1019 * highest_flag_domain - Return highest sched_domain containing flag. 1020 * @cpu: The cpu whose highest level of sched domain is to 1021 * be returned. 1022 * @flag: The flag to check for the highest sched_domain 1023 * for the given cpu. 1024 * 1025 * Returns the highest sched_domain of a cpu which contains the given flag. 1026 */ 1027 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 1028 { 1029 struct sched_domain *sd, *hsd = NULL; 1030 1031 for_each_domain(cpu, sd) { 1032 if (!(sd->flags & flag)) 1033 break; 1034 hsd = sd; 1035 } 1036 1037 return hsd; 1038 } 1039 1040 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) 1041 { 1042 struct sched_domain *sd; 1043 1044 for_each_domain(cpu, sd) { 1045 if (sd->flags & flag) 1046 break; 1047 } 1048 1049 return sd; 1050 } 1051 1052 DECLARE_PER_CPU(struct sched_domain *, sd_llc); 1053 DECLARE_PER_CPU(int, sd_llc_size); 1054 DECLARE_PER_CPU(int, sd_llc_id); 1055 DECLARE_PER_CPU(struct sched_domain_shared *, sd_llc_shared); 1056 DECLARE_PER_CPU(struct sched_domain *, sd_numa); 1057 DECLARE_PER_CPU(struct sched_domain *, sd_asym); 1058 1059 struct sched_group_capacity { 1060 atomic_t ref; 1061 /* 1062 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity 1063 * for a single CPU. 1064 */ 1065 unsigned long capacity; 1066 unsigned long min_capacity; /* Min per-CPU capacity in group */ 1067 unsigned long next_update; 1068 int imbalance; /* XXX unrelated to capacity but shared group state */ 1069 1070 #ifdef CONFIG_SCHED_DEBUG 1071 int id; 1072 #endif 1073 1074 unsigned long cpumask[0]; /* balance mask */ 1075 }; 1076 1077 struct sched_group { 1078 struct sched_group *next; /* Must be a circular list */ 1079 atomic_t ref; 1080 1081 unsigned int group_weight; 1082 struct sched_group_capacity *sgc; 1083 int asym_prefer_cpu; /* cpu of highest priority in group */ 1084 1085 /* 1086 * The CPUs this group covers. 1087 * 1088 * NOTE: this field is variable length. (Allocated dynamically 1089 * by attaching extra space to the end of the structure, 1090 * depending on how many CPUs the kernel has booted up with) 1091 */ 1092 unsigned long cpumask[0]; 1093 }; 1094 1095 static inline struct cpumask *sched_group_span(struct sched_group *sg) 1096 { 1097 return to_cpumask(sg->cpumask); 1098 } 1099 1100 /* 1101 * See build_balance_mask(). 1102 */ 1103 static inline struct cpumask *group_balance_mask(struct sched_group *sg) 1104 { 1105 return to_cpumask(sg->sgc->cpumask); 1106 } 1107 1108 /** 1109 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group. 1110 * @group: The group whose first cpu is to be returned. 1111 */ 1112 static inline unsigned int group_first_cpu(struct sched_group *group) 1113 { 1114 return cpumask_first(sched_group_span(group)); 1115 } 1116 1117 extern int group_balance_cpu(struct sched_group *sg); 1118 1119 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 1120 void register_sched_domain_sysctl(void); 1121 void dirty_sched_domain_sysctl(int cpu); 1122 void unregister_sched_domain_sysctl(void); 1123 #else 1124 static inline void register_sched_domain_sysctl(void) 1125 { 1126 } 1127 static inline void dirty_sched_domain_sysctl(int cpu) 1128 { 1129 } 1130 static inline void unregister_sched_domain_sysctl(void) 1131 { 1132 } 1133 #endif 1134 1135 #else 1136 1137 static inline void sched_ttwu_pending(void) { } 1138 1139 #endif /* CONFIG_SMP */ 1140 1141 #include "stats.h" 1142 #include "autogroup.h" 1143 1144 #ifdef CONFIG_CGROUP_SCHED 1145 1146 /* 1147 * Return the group to which this tasks belongs. 1148 * 1149 * We cannot use task_css() and friends because the cgroup subsystem 1150 * changes that value before the cgroup_subsys::attach() method is called, 1151 * therefore we cannot pin it and might observe the wrong value. 1152 * 1153 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup 1154 * core changes this before calling sched_move_task(). 1155 * 1156 * Instead we use a 'copy' which is updated from sched_move_task() while 1157 * holding both task_struct::pi_lock and rq::lock. 1158 */ 1159 static inline struct task_group *task_group(struct task_struct *p) 1160 { 1161 return p->sched_task_group; 1162 } 1163 1164 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 1165 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 1166 { 1167 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 1168 struct task_group *tg = task_group(p); 1169 #endif 1170 1171 #ifdef CONFIG_FAIR_GROUP_SCHED 1172 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]); 1173 p->se.cfs_rq = tg->cfs_rq[cpu]; 1174 p->se.parent = tg->se[cpu]; 1175 #endif 1176 1177 #ifdef CONFIG_RT_GROUP_SCHED 1178 p->rt.rt_rq = tg->rt_rq[cpu]; 1179 p->rt.parent = tg->rt_se[cpu]; 1180 #endif 1181 } 1182 1183 #else /* CONFIG_CGROUP_SCHED */ 1184 1185 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 1186 static inline struct task_group *task_group(struct task_struct *p) 1187 { 1188 return NULL; 1189 } 1190 1191 #endif /* CONFIG_CGROUP_SCHED */ 1192 1193 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 1194 { 1195 set_task_rq(p, cpu); 1196 #ifdef CONFIG_SMP 1197 /* 1198 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 1199 * successfuly executed on another CPU. We must ensure that updates of 1200 * per-task data have been completed by this moment. 1201 */ 1202 smp_wmb(); 1203 #ifdef CONFIG_THREAD_INFO_IN_TASK 1204 p->cpu = cpu; 1205 #else 1206 task_thread_info(p)->cpu = cpu; 1207 #endif 1208 p->wake_cpu = cpu; 1209 #endif 1210 } 1211 1212 /* 1213 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 1214 */ 1215 #ifdef CONFIG_SCHED_DEBUG 1216 # include <linux/static_key.h> 1217 # define const_debug __read_mostly 1218 #else 1219 # define const_debug const 1220 #endif 1221 1222 extern const_debug unsigned int sysctl_sched_features; 1223 1224 #define SCHED_FEAT(name, enabled) \ 1225 __SCHED_FEAT_##name , 1226 1227 enum { 1228 #include "features.h" 1229 __SCHED_FEAT_NR, 1230 }; 1231 1232 #undef SCHED_FEAT 1233 1234 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL) 1235 #define SCHED_FEAT(name, enabled) \ 1236 static __always_inline bool static_branch_##name(struct static_key *key) \ 1237 { \ 1238 return static_key_##enabled(key); \ 1239 } 1240 1241 #include "features.h" 1242 1243 #undef SCHED_FEAT 1244 1245 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 1246 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 1247 #else /* !(SCHED_DEBUG && HAVE_JUMP_LABEL) */ 1248 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 1249 #endif /* SCHED_DEBUG && HAVE_JUMP_LABEL */ 1250 1251 extern struct static_key_false sched_numa_balancing; 1252 extern struct static_key_false sched_schedstats; 1253 1254 static inline u64 global_rt_period(void) 1255 { 1256 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 1257 } 1258 1259 static inline u64 global_rt_runtime(void) 1260 { 1261 if (sysctl_sched_rt_runtime < 0) 1262 return RUNTIME_INF; 1263 1264 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 1265 } 1266 1267 static inline int task_current(struct rq *rq, struct task_struct *p) 1268 { 1269 return rq->curr == p; 1270 } 1271 1272 static inline int task_running(struct rq *rq, struct task_struct *p) 1273 { 1274 #ifdef CONFIG_SMP 1275 return p->on_cpu; 1276 #else 1277 return task_current(rq, p); 1278 #endif 1279 } 1280 1281 static inline int task_on_rq_queued(struct task_struct *p) 1282 { 1283 return p->on_rq == TASK_ON_RQ_QUEUED; 1284 } 1285 1286 static inline int task_on_rq_migrating(struct task_struct *p) 1287 { 1288 return p->on_rq == TASK_ON_RQ_MIGRATING; 1289 } 1290 1291 #ifndef prepare_arch_switch 1292 # define prepare_arch_switch(next) do { } while (0) 1293 #endif 1294 #ifndef finish_arch_post_lock_switch 1295 # define finish_arch_post_lock_switch() do { } while (0) 1296 #endif 1297 1298 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) 1299 { 1300 #ifdef CONFIG_SMP 1301 /* 1302 * We can optimise this out completely for !SMP, because the 1303 * SMP rebalancing from interrupt is the only thing that cares 1304 * here. 1305 */ 1306 next->on_cpu = 1; 1307 #endif 1308 } 1309 1310 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) 1311 { 1312 #ifdef CONFIG_SMP 1313 /* 1314 * After ->on_cpu is cleared, the task can be moved to a different CPU. 1315 * We must ensure this doesn't happen until the switch is completely 1316 * finished. 1317 * 1318 * In particular, the load of prev->state in finish_task_switch() must 1319 * happen before this. 1320 * 1321 * Pairs with the smp_cond_load_acquire() in try_to_wake_up(). 1322 */ 1323 smp_store_release(&prev->on_cpu, 0); 1324 #endif 1325 #ifdef CONFIG_DEBUG_SPINLOCK 1326 /* this is a valid case when another task releases the spinlock */ 1327 rq->lock.owner = current; 1328 #endif 1329 /* 1330 * If we are tracking spinlock dependencies then we have to 1331 * fix up the runqueue lock - which gets 'carried over' from 1332 * prev into current: 1333 */ 1334 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); 1335 1336 raw_spin_unlock_irq(&rq->lock); 1337 } 1338 1339 /* 1340 * wake flags 1341 */ 1342 #define WF_SYNC 0x01 /* waker goes to sleep after wakeup */ 1343 #define WF_FORK 0x02 /* child wakeup after fork */ 1344 #define WF_MIGRATED 0x4 /* internal use, task got migrated */ 1345 1346 /* 1347 * To aid in avoiding the subversion of "niceness" due to uneven distribution 1348 * of tasks with abnormal "nice" values across CPUs the contribution that 1349 * each task makes to its run queue's load is weighted according to its 1350 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 1351 * scaled version of the new time slice allocation that they receive on time 1352 * slice expiry etc. 1353 */ 1354 1355 #define WEIGHT_IDLEPRIO 3 1356 #define WMULT_IDLEPRIO 1431655765 1357 1358 extern const int sched_prio_to_weight[40]; 1359 extern const u32 sched_prio_to_wmult[40]; 1360 1361 /* 1362 * {de,en}queue flags: 1363 * 1364 * DEQUEUE_SLEEP - task is no longer runnable 1365 * ENQUEUE_WAKEUP - task just became runnable 1366 * 1367 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks 1368 * are in a known state which allows modification. Such pairs 1369 * should preserve as much state as possible. 1370 * 1371 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location 1372 * in the runqueue. 1373 * 1374 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified) 1375 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline) 1376 * ENQUEUE_MIGRATED - the task was migrated during wakeup 1377 * 1378 */ 1379 1380 #define DEQUEUE_SLEEP 0x01 1381 #define DEQUEUE_SAVE 0x02 /* matches ENQUEUE_RESTORE */ 1382 #define DEQUEUE_MOVE 0x04 /* matches ENQUEUE_MOVE */ 1383 #define DEQUEUE_NOCLOCK 0x08 /* matches ENQUEUE_NOCLOCK */ 1384 1385 #define ENQUEUE_WAKEUP 0x01 1386 #define ENQUEUE_RESTORE 0x02 1387 #define ENQUEUE_MOVE 0x04 1388 #define ENQUEUE_NOCLOCK 0x08 1389 1390 #define ENQUEUE_HEAD 0x10 1391 #define ENQUEUE_REPLENISH 0x20 1392 #ifdef CONFIG_SMP 1393 #define ENQUEUE_MIGRATED 0x40 1394 #else 1395 #define ENQUEUE_MIGRATED 0x00 1396 #endif 1397 1398 #define RETRY_TASK ((void *)-1UL) 1399 1400 struct sched_class { 1401 const struct sched_class *next; 1402 1403 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags); 1404 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags); 1405 void (*yield_task) (struct rq *rq); 1406 bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt); 1407 1408 void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags); 1409 1410 /* 1411 * It is the responsibility of the pick_next_task() method that will 1412 * return the next task to call put_prev_task() on the @prev task or 1413 * something equivalent. 1414 * 1415 * May return RETRY_TASK when it finds a higher prio class has runnable 1416 * tasks. 1417 */ 1418 struct task_struct * (*pick_next_task) (struct rq *rq, 1419 struct task_struct *prev, 1420 struct rq_flags *rf); 1421 void (*put_prev_task) (struct rq *rq, struct task_struct *p); 1422 1423 #ifdef CONFIG_SMP 1424 int (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags); 1425 void (*migrate_task_rq)(struct task_struct *p); 1426 1427 void (*task_woken) (struct rq *this_rq, struct task_struct *task); 1428 1429 void (*set_cpus_allowed)(struct task_struct *p, 1430 const struct cpumask *newmask); 1431 1432 void (*rq_online)(struct rq *rq); 1433 void (*rq_offline)(struct rq *rq); 1434 #endif 1435 1436 void (*set_curr_task) (struct rq *rq); 1437 void (*task_tick) (struct rq *rq, struct task_struct *p, int queued); 1438 void (*task_fork) (struct task_struct *p); 1439 void (*task_dead) (struct task_struct *p); 1440 1441 /* 1442 * The switched_from() call is allowed to drop rq->lock, therefore we 1443 * cannot assume the switched_from/switched_to pair is serliazed by 1444 * rq->lock. They are however serialized by p->pi_lock. 1445 */ 1446 void (*switched_from) (struct rq *this_rq, struct task_struct *task); 1447 void (*switched_to) (struct rq *this_rq, struct task_struct *task); 1448 void (*prio_changed) (struct rq *this_rq, struct task_struct *task, 1449 int oldprio); 1450 1451 unsigned int (*get_rr_interval) (struct rq *rq, 1452 struct task_struct *task); 1453 1454 void (*update_curr) (struct rq *rq); 1455 1456 #define TASK_SET_GROUP 0 1457 #define TASK_MOVE_GROUP 1 1458 1459 #ifdef CONFIG_FAIR_GROUP_SCHED 1460 void (*task_change_group) (struct task_struct *p, int type); 1461 #endif 1462 }; 1463 1464 static inline void put_prev_task(struct rq *rq, struct task_struct *prev) 1465 { 1466 prev->sched_class->put_prev_task(rq, prev); 1467 } 1468 1469 static inline void set_curr_task(struct rq *rq, struct task_struct *curr) 1470 { 1471 curr->sched_class->set_curr_task(rq); 1472 } 1473 1474 #ifdef CONFIG_SMP 1475 #define sched_class_highest (&stop_sched_class) 1476 #else 1477 #define sched_class_highest (&dl_sched_class) 1478 #endif 1479 #define for_each_class(class) \ 1480 for (class = sched_class_highest; class; class = class->next) 1481 1482 extern const struct sched_class stop_sched_class; 1483 extern const struct sched_class dl_sched_class; 1484 extern const struct sched_class rt_sched_class; 1485 extern const struct sched_class fair_sched_class; 1486 extern const struct sched_class idle_sched_class; 1487 1488 1489 #ifdef CONFIG_SMP 1490 1491 extern void update_group_capacity(struct sched_domain *sd, int cpu); 1492 1493 extern void trigger_load_balance(struct rq *rq); 1494 1495 extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask); 1496 1497 #endif 1498 1499 #ifdef CONFIG_CPU_IDLE 1500 static inline void idle_set_state(struct rq *rq, 1501 struct cpuidle_state *idle_state) 1502 { 1503 rq->idle_state = idle_state; 1504 } 1505 1506 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1507 { 1508 SCHED_WARN_ON(!rcu_read_lock_held()); 1509 return rq->idle_state; 1510 } 1511 #else 1512 static inline void idle_set_state(struct rq *rq, 1513 struct cpuidle_state *idle_state) 1514 { 1515 } 1516 1517 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 1518 { 1519 return NULL; 1520 } 1521 #endif 1522 1523 extern void schedule_idle(void); 1524 1525 extern void sysrq_sched_debug_show(void); 1526 extern void sched_init_granularity(void); 1527 extern void update_max_interval(void); 1528 1529 extern void init_sched_dl_class(void); 1530 extern void init_sched_rt_class(void); 1531 extern void init_sched_fair_class(void); 1532 1533 extern void resched_curr(struct rq *rq); 1534 extern void resched_cpu(int cpu); 1535 1536 extern struct rt_bandwidth def_rt_bandwidth; 1537 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 1538 1539 extern struct dl_bandwidth def_dl_bandwidth; 1540 extern void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime); 1541 extern void init_dl_task_timer(struct sched_dl_entity *dl_se); 1542 extern void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se); 1543 extern void init_dl_rq_bw_ratio(struct dl_rq *dl_rq); 1544 1545 #define BW_SHIFT 20 1546 #define BW_UNIT (1 << BW_SHIFT) 1547 #define RATIO_SHIFT 8 1548 unsigned long to_ratio(u64 period, u64 runtime); 1549 1550 extern void init_entity_runnable_average(struct sched_entity *se); 1551 extern void post_init_entity_util_avg(struct sched_entity *se); 1552 1553 #ifdef CONFIG_NO_HZ_FULL 1554 extern bool sched_can_stop_tick(struct rq *rq); 1555 1556 /* 1557 * Tick may be needed by tasks in the runqueue depending on their policy and 1558 * requirements. If tick is needed, lets send the target an IPI to kick it out of 1559 * nohz mode if necessary. 1560 */ 1561 static inline void sched_update_tick_dependency(struct rq *rq) 1562 { 1563 int cpu; 1564 1565 if (!tick_nohz_full_enabled()) 1566 return; 1567 1568 cpu = cpu_of(rq); 1569 1570 if (!tick_nohz_full_cpu(cpu)) 1571 return; 1572 1573 if (sched_can_stop_tick(rq)) 1574 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED); 1575 else 1576 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED); 1577 } 1578 #else 1579 static inline void sched_update_tick_dependency(struct rq *rq) { } 1580 #endif 1581 1582 static inline void add_nr_running(struct rq *rq, unsigned count) 1583 { 1584 unsigned prev_nr = rq->nr_running; 1585 1586 rq->nr_running = prev_nr + count; 1587 1588 if (prev_nr < 2 && rq->nr_running >= 2) { 1589 #ifdef CONFIG_SMP 1590 if (!rq->rd->overload) 1591 rq->rd->overload = true; 1592 #endif 1593 } 1594 1595 sched_update_tick_dependency(rq); 1596 } 1597 1598 static inline void sub_nr_running(struct rq *rq, unsigned count) 1599 { 1600 rq->nr_running -= count; 1601 /* Check if we still need preemption */ 1602 sched_update_tick_dependency(rq); 1603 } 1604 1605 static inline void rq_last_tick_reset(struct rq *rq) 1606 { 1607 #ifdef CONFIG_NO_HZ_FULL 1608 rq->last_sched_tick = jiffies; 1609 #endif 1610 } 1611 1612 extern void update_rq_clock(struct rq *rq); 1613 1614 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 1615 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 1616 1617 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags); 1618 1619 extern const_debug unsigned int sysctl_sched_time_avg; 1620 extern const_debug unsigned int sysctl_sched_nr_migrate; 1621 extern const_debug unsigned int sysctl_sched_migration_cost; 1622 1623 static inline u64 sched_avg_period(void) 1624 { 1625 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2; 1626 } 1627 1628 #ifdef CONFIG_SCHED_HRTICK 1629 1630 /* 1631 * Use hrtick when: 1632 * - enabled by features 1633 * - hrtimer is actually high res 1634 */ 1635 static inline int hrtick_enabled(struct rq *rq) 1636 { 1637 if (!sched_feat(HRTICK)) 1638 return 0; 1639 if (!cpu_active(cpu_of(rq))) 1640 return 0; 1641 return hrtimer_is_hres_active(&rq->hrtick_timer); 1642 } 1643 1644 void hrtick_start(struct rq *rq, u64 delay); 1645 1646 #else 1647 1648 static inline int hrtick_enabled(struct rq *rq) 1649 { 1650 return 0; 1651 } 1652 1653 #endif /* CONFIG_SCHED_HRTICK */ 1654 1655 #ifdef CONFIG_SMP 1656 extern void sched_avg_update(struct rq *rq); 1657 1658 #ifndef arch_scale_freq_capacity 1659 static __always_inline 1660 unsigned long arch_scale_freq_capacity(struct sched_domain *sd, int cpu) 1661 { 1662 return SCHED_CAPACITY_SCALE; 1663 } 1664 #endif 1665 1666 #ifndef arch_scale_cpu_capacity 1667 static __always_inline 1668 unsigned long arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) 1669 { 1670 if (sd && (sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) 1671 return sd->smt_gain / sd->span_weight; 1672 1673 return SCHED_CAPACITY_SCALE; 1674 } 1675 #endif 1676 1677 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) 1678 { 1679 rq->rt_avg += rt_delta * arch_scale_freq_capacity(NULL, cpu_of(rq)); 1680 sched_avg_update(rq); 1681 } 1682 #else 1683 static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta) { } 1684 static inline void sched_avg_update(struct rq *rq) { } 1685 #endif 1686 1687 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1688 __acquires(rq->lock); 1689 1690 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1691 __acquires(p->pi_lock) 1692 __acquires(rq->lock); 1693 1694 static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf) 1695 __releases(rq->lock) 1696 { 1697 rq_unpin_lock(rq, rf); 1698 raw_spin_unlock(&rq->lock); 1699 } 1700 1701 static inline void 1702 task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf) 1703 __releases(rq->lock) 1704 __releases(p->pi_lock) 1705 { 1706 rq_unpin_lock(rq, rf); 1707 raw_spin_unlock(&rq->lock); 1708 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); 1709 } 1710 1711 static inline void 1712 rq_lock_irqsave(struct rq *rq, struct rq_flags *rf) 1713 __acquires(rq->lock) 1714 { 1715 raw_spin_lock_irqsave(&rq->lock, rf->flags); 1716 rq_pin_lock(rq, rf); 1717 } 1718 1719 static inline void 1720 rq_lock_irq(struct rq *rq, struct rq_flags *rf) 1721 __acquires(rq->lock) 1722 { 1723 raw_spin_lock_irq(&rq->lock); 1724 rq_pin_lock(rq, rf); 1725 } 1726 1727 static inline void 1728 rq_lock(struct rq *rq, struct rq_flags *rf) 1729 __acquires(rq->lock) 1730 { 1731 raw_spin_lock(&rq->lock); 1732 rq_pin_lock(rq, rf); 1733 } 1734 1735 static inline void 1736 rq_relock(struct rq *rq, struct rq_flags *rf) 1737 __acquires(rq->lock) 1738 { 1739 raw_spin_lock(&rq->lock); 1740 rq_repin_lock(rq, rf); 1741 } 1742 1743 static inline void 1744 rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf) 1745 __releases(rq->lock) 1746 { 1747 rq_unpin_lock(rq, rf); 1748 raw_spin_unlock_irqrestore(&rq->lock, rf->flags); 1749 } 1750 1751 static inline void 1752 rq_unlock_irq(struct rq *rq, struct rq_flags *rf) 1753 __releases(rq->lock) 1754 { 1755 rq_unpin_lock(rq, rf); 1756 raw_spin_unlock_irq(&rq->lock); 1757 } 1758 1759 static inline void 1760 rq_unlock(struct rq *rq, struct rq_flags *rf) 1761 __releases(rq->lock) 1762 { 1763 rq_unpin_lock(rq, rf); 1764 raw_spin_unlock(&rq->lock); 1765 } 1766 1767 #ifdef CONFIG_SMP 1768 #ifdef CONFIG_PREEMPT 1769 1770 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2); 1771 1772 /* 1773 * fair double_lock_balance: Safely acquires both rq->locks in a fair 1774 * way at the expense of forcing extra atomic operations in all 1775 * invocations. This assures that the double_lock is acquired using the 1776 * same underlying policy as the spinlock_t on this architecture, which 1777 * reduces latency compared to the unfair variant below. However, it 1778 * also adds more overhead and therefore may reduce throughput. 1779 */ 1780 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1781 __releases(this_rq->lock) 1782 __acquires(busiest->lock) 1783 __acquires(this_rq->lock) 1784 { 1785 raw_spin_unlock(&this_rq->lock); 1786 double_rq_lock(this_rq, busiest); 1787 1788 return 1; 1789 } 1790 1791 #else 1792 /* 1793 * Unfair double_lock_balance: Optimizes throughput at the expense of 1794 * latency by eliminating extra atomic operations when the locks are 1795 * already in proper order on entry. This favors lower cpu-ids and will 1796 * grant the double lock to lower cpus over higher ids under contention, 1797 * regardless of entry order into the function. 1798 */ 1799 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 1800 __releases(this_rq->lock) 1801 __acquires(busiest->lock) 1802 __acquires(this_rq->lock) 1803 { 1804 int ret = 0; 1805 1806 if (unlikely(!raw_spin_trylock(&busiest->lock))) { 1807 if (busiest < this_rq) { 1808 raw_spin_unlock(&this_rq->lock); 1809 raw_spin_lock(&busiest->lock); 1810 raw_spin_lock_nested(&this_rq->lock, 1811 SINGLE_DEPTH_NESTING); 1812 ret = 1; 1813 } else 1814 raw_spin_lock_nested(&busiest->lock, 1815 SINGLE_DEPTH_NESTING); 1816 } 1817 return ret; 1818 } 1819 1820 #endif /* CONFIG_PREEMPT */ 1821 1822 /* 1823 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 1824 */ 1825 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 1826 { 1827 if (unlikely(!irqs_disabled())) { 1828 /* printk() doesn't work good under rq->lock */ 1829 raw_spin_unlock(&this_rq->lock); 1830 BUG_ON(1); 1831 } 1832 1833 return _double_lock_balance(this_rq, busiest); 1834 } 1835 1836 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 1837 __releases(busiest->lock) 1838 { 1839 raw_spin_unlock(&busiest->lock); 1840 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_); 1841 } 1842 1843 static inline void double_lock(spinlock_t *l1, spinlock_t *l2) 1844 { 1845 if (l1 > l2) 1846 swap(l1, l2); 1847 1848 spin_lock(l1); 1849 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1850 } 1851 1852 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2) 1853 { 1854 if (l1 > l2) 1855 swap(l1, l2); 1856 1857 spin_lock_irq(l1); 1858 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1859 } 1860 1861 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2) 1862 { 1863 if (l1 > l2) 1864 swap(l1, l2); 1865 1866 raw_spin_lock(l1); 1867 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 1868 } 1869 1870 /* 1871 * double_rq_lock - safely lock two runqueues 1872 * 1873 * Note this does not disable interrupts like task_rq_lock, 1874 * you need to do so manually before calling. 1875 */ 1876 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1877 __acquires(rq1->lock) 1878 __acquires(rq2->lock) 1879 { 1880 BUG_ON(!irqs_disabled()); 1881 if (rq1 == rq2) { 1882 raw_spin_lock(&rq1->lock); 1883 __acquire(rq2->lock); /* Fake it out ;) */ 1884 } else { 1885 if (rq1 < rq2) { 1886 raw_spin_lock(&rq1->lock); 1887 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING); 1888 } else { 1889 raw_spin_lock(&rq2->lock); 1890 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING); 1891 } 1892 } 1893 } 1894 1895 /* 1896 * double_rq_unlock - safely unlock two runqueues 1897 * 1898 * Note this does not restore interrupts like task_rq_unlock, 1899 * you need to do so manually after calling. 1900 */ 1901 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1902 __releases(rq1->lock) 1903 __releases(rq2->lock) 1904 { 1905 raw_spin_unlock(&rq1->lock); 1906 if (rq1 != rq2) 1907 raw_spin_unlock(&rq2->lock); 1908 else 1909 __release(rq2->lock); 1910 } 1911 1912 extern void set_rq_online (struct rq *rq); 1913 extern void set_rq_offline(struct rq *rq); 1914 extern bool sched_smp_initialized; 1915 1916 #else /* CONFIG_SMP */ 1917 1918 /* 1919 * double_rq_lock - safely lock two runqueues 1920 * 1921 * Note this does not disable interrupts like task_rq_lock, 1922 * you need to do so manually before calling. 1923 */ 1924 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 1925 __acquires(rq1->lock) 1926 __acquires(rq2->lock) 1927 { 1928 BUG_ON(!irqs_disabled()); 1929 BUG_ON(rq1 != rq2); 1930 raw_spin_lock(&rq1->lock); 1931 __acquire(rq2->lock); /* Fake it out ;) */ 1932 } 1933 1934 /* 1935 * double_rq_unlock - safely unlock two runqueues 1936 * 1937 * Note this does not restore interrupts like task_rq_unlock, 1938 * you need to do so manually after calling. 1939 */ 1940 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 1941 __releases(rq1->lock) 1942 __releases(rq2->lock) 1943 { 1944 BUG_ON(rq1 != rq2); 1945 raw_spin_unlock(&rq1->lock); 1946 __release(rq2->lock); 1947 } 1948 1949 #endif 1950 1951 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 1952 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 1953 1954 #ifdef CONFIG_SCHED_DEBUG 1955 extern bool sched_debug_enabled; 1956 1957 extern void print_cfs_stats(struct seq_file *m, int cpu); 1958 extern void print_rt_stats(struct seq_file *m, int cpu); 1959 extern void print_dl_stats(struct seq_file *m, int cpu); 1960 extern void 1961 print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq); 1962 #ifdef CONFIG_NUMA_BALANCING 1963 extern void 1964 show_numa_stats(struct task_struct *p, struct seq_file *m); 1965 extern void 1966 print_numa_stats(struct seq_file *m, int node, unsigned long tsf, 1967 unsigned long tpf, unsigned long gsf, unsigned long gpf); 1968 #endif /* CONFIG_NUMA_BALANCING */ 1969 #endif /* CONFIG_SCHED_DEBUG */ 1970 1971 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 1972 extern void init_rt_rq(struct rt_rq *rt_rq); 1973 extern void init_dl_rq(struct dl_rq *dl_rq); 1974 1975 extern void cfs_bandwidth_usage_inc(void); 1976 extern void cfs_bandwidth_usage_dec(void); 1977 1978 #ifdef CONFIG_NO_HZ_COMMON 1979 enum rq_nohz_flag_bits { 1980 NOHZ_TICK_STOPPED, 1981 NOHZ_BALANCE_KICK, 1982 }; 1983 1984 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 1985 1986 extern void nohz_balance_exit_idle(unsigned int cpu); 1987 #else 1988 static inline void nohz_balance_exit_idle(unsigned int cpu) { } 1989 #endif 1990 1991 1992 #ifdef CONFIG_SMP 1993 static inline 1994 void __dl_update(struct dl_bw *dl_b, s64 bw) 1995 { 1996 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw); 1997 int i; 1998 1999 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(), 2000 "sched RCU must be held"); 2001 for_each_cpu_and(i, rd->span, cpu_active_mask) { 2002 struct rq *rq = cpu_rq(i); 2003 2004 rq->dl.extra_bw += bw; 2005 } 2006 } 2007 #else 2008 static inline 2009 void __dl_update(struct dl_bw *dl_b, s64 bw) 2010 { 2011 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw); 2012 2013 dl->extra_bw += bw; 2014 } 2015 #endif 2016 2017 2018 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 2019 struct irqtime { 2020 u64 total; 2021 u64 tick_delta; 2022 u64 irq_start_time; 2023 struct u64_stats_sync sync; 2024 }; 2025 2026 DECLARE_PER_CPU(struct irqtime, cpu_irqtime); 2027 2028 /* 2029 * Returns the irqtime minus the softirq time computed by ksoftirqd. 2030 * Otherwise ksoftirqd's sum_exec_runtime is substracted its own runtime 2031 * and never move forward. 2032 */ 2033 static inline u64 irq_time_read(int cpu) 2034 { 2035 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu); 2036 unsigned int seq; 2037 u64 total; 2038 2039 do { 2040 seq = __u64_stats_fetch_begin(&irqtime->sync); 2041 total = irqtime->total; 2042 } while (__u64_stats_fetch_retry(&irqtime->sync, seq)); 2043 2044 return total; 2045 } 2046 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 2047 2048 #ifdef CONFIG_CPU_FREQ 2049 DECLARE_PER_CPU(struct update_util_data *, cpufreq_update_util_data); 2050 2051 /** 2052 * cpufreq_update_util - Take a note about CPU utilization changes. 2053 * @rq: Runqueue to carry out the update for. 2054 * @flags: Update reason flags. 2055 * 2056 * This function is called by the scheduler on the CPU whose utilization is 2057 * being updated. 2058 * 2059 * It can only be called from RCU-sched read-side critical sections. 2060 * 2061 * The way cpufreq is currently arranged requires it to evaluate the CPU 2062 * performance state (frequency/voltage) on a regular basis to prevent it from 2063 * being stuck in a completely inadequate performance level for too long. 2064 * That is not guaranteed to happen if the updates are only triggered from CFS, 2065 * though, because they may not be coming in if RT or deadline tasks are active 2066 * all the time (or there are RT and DL tasks only). 2067 * 2068 * As a workaround for that issue, this function is called by the RT and DL 2069 * sched classes to trigger extra cpufreq updates to prevent it from stalling, 2070 * but that really is a band-aid. Going forward it should be replaced with 2071 * solutions targeted more specifically at RT and DL tasks. 2072 */ 2073 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) 2074 { 2075 struct update_util_data *data; 2076 2077 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data, 2078 cpu_of(rq))); 2079 if (data) 2080 data->func(data, rq_clock(rq), flags); 2081 } 2082 #else 2083 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {} 2084 #endif /* CONFIG_CPU_FREQ */ 2085 2086 #ifdef arch_scale_freq_capacity 2087 #ifndef arch_scale_freq_invariant 2088 #define arch_scale_freq_invariant() (true) 2089 #endif 2090 #else /* arch_scale_freq_capacity */ 2091 #define arch_scale_freq_invariant() (false) 2092 #endif 2093