1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* 3 * Scheduler internal types and methods: 4 */ 5 #ifndef _KERNEL_SCHED_SCHED_H 6 #define _KERNEL_SCHED_SCHED_H 7 8 #include <linux/sched/affinity.h> 9 #include <linux/sched/autogroup.h> 10 #include <linux/sched/cpufreq.h> 11 #include <linux/sched/deadline.h> 12 #include <linux/sched.h> 13 #include <linux/sched/loadavg.h> 14 #include <linux/sched/mm.h> 15 #include <linux/sched/rseq_api.h> 16 #include <linux/sched/signal.h> 17 #include <linux/sched/smt.h> 18 #include <linux/sched/stat.h> 19 #include <linux/sched/sysctl.h> 20 #include <linux/sched/task_flags.h> 21 #include <linux/sched/task.h> 22 #include <linux/sched/topology.h> 23 24 #include <linux/atomic.h> 25 #include <linux/bitmap.h> 26 #include <linux/bug.h> 27 #include <linux/capability.h> 28 #include <linux/cgroup_api.h> 29 #include <linux/cgroup.h> 30 #include <linux/context_tracking.h> 31 #include <linux/cpufreq.h> 32 #include <linux/cpumask_api.h> 33 #include <linux/ctype.h> 34 #include <linux/file.h> 35 #include <linux/fs_api.h> 36 #include <linux/hrtimer_api.h> 37 #include <linux/interrupt.h> 38 #include <linux/irq_work.h> 39 #include <linux/jiffies.h> 40 #include <linux/kref_api.h> 41 #include <linux/kthread.h> 42 #include <linux/ktime_api.h> 43 #include <linux/lockdep_api.h> 44 #include <linux/lockdep.h> 45 #include <linux/minmax.h> 46 #include <linux/mm.h> 47 #include <linux/module.h> 48 #include <linux/mutex_api.h> 49 #include <linux/plist.h> 50 #include <linux/poll.h> 51 #include <linux/proc_fs.h> 52 #include <linux/profile.h> 53 #include <linux/psi.h> 54 #include <linux/rcupdate.h> 55 #include <linux/seq_file.h> 56 #include <linux/seqlock.h> 57 #include <linux/softirq.h> 58 #include <linux/spinlock_api.h> 59 #include <linux/static_key.h> 60 #include <linux/stop_machine.h> 61 #include <linux/syscalls_api.h> 62 #include <linux/syscalls.h> 63 #include <linux/tick.h> 64 #include <linux/topology.h> 65 #include <linux/types.h> 66 #include <linux/u64_stats_sync_api.h> 67 #include <linux/uaccess.h> 68 #include <linux/wait_api.h> 69 #include <linux/wait_bit.h> 70 #include <linux/workqueue_api.h> 71 72 #include <trace/events/power.h> 73 #include <trace/events/sched.h> 74 75 #include "../workqueue_internal.h" 76 77 #ifdef CONFIG_PARAVIRT 78 # include <asm/paravirt.h> 79 # include <asm/paravirt_api_clock.h> 80 #endif 81 82 #include "cpupri.h" 83 #include "cpudeadline.h" 84 85 #ifdef CONFIG_SCHED_DEBUG 86 # define SCHED_WARN_ON(x) WARN_ONCE(x, #x) 87 #else 88 # define SCHED_WARN_ON(x) ({ (void)(x), 0; }) 89 #endif 90 91 struct rq; 92 struct cpuidle_state; 93 94 /* task_struct::on_rq states: */ 95 #define TASK_ON_RQ_QUEUED 1 96 #define TASK_ON_RQ_MIGRATING 2 97 98 extern __read_mostly int scheduler_running; 99 100 extern unsigned long calc_load_update; 101 extern atomic_long_t calc_load_tasks; 102 103 extern void calc_global_load_tick(struct rq *this_rq); 104 extern long calc_load_fold_active(struct rq *this_rq, long adjust); 105 106 extern void call_trace_sched_update_nr_running(struct rq *rq, int count); 107 108 extern int sysctl_sched_rt_period; 109 extern int sysctl_sched_rt_runtime; 110 extern int sched_rr_timeslice; 111 112 /* 113 * Helpers for converting nanosecond timing to jiffy resolution 114 */ 115 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) 116 117 /* 118 * Increase resolution of nice-level calculations for 64-bit architectures. 119 * The extra resolution improves shares distribution and load balancing of 120 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup 121 * hierarchies, especially on larger systems. This is not a user-visible change 122 * and does not change the user-interface for setting shares/weights. 123 * 124 * We increase resolution only if we have enough bits to allow this increased 125 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit 126 * are pretty high and the returns do not justify the increased costs. 127 * 128 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to 129 * increase coverage and consistency always enable it on 64-bit platforms. 130 */ 131 #ifdef CONFIG_64BIT 132 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT) 133 # define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT) 134 # define scale_load_down(w) \ 135 ({ \ 136 unsigned long __w = (w); \ 137 if (__w) \ 138 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \ 139 __w; \ 140 }) 141 #else 142 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT) 143 # define scale_load(w) (w) 144 # define scale_load_down(w) (w) 145 #endif 146 147 /* 148 * Task weight (visible to users) and its load (invisible to users) have 149 * independent resolution, but they should be well calibrated. We use 150 * scale_load() and scale_load_down(w) to convert between them. The 151 * following must be true: 152 * 153 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD 154 * 155 */ 156 #define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT) 157 158 /* 159 * Single value that decides SCHED_DEADLINE internal math precision. 160 * 10 -> just above 1us 161 * 9 -> just above 0.5us 162 */ 163 #define DL_SCALE 10 164 165 /* 166 * Single value that denotes runtime == period, ie unlimited time. 167 */ 168 #define RUNTIME_INF ((u64)~0ULL) 169 170 static inline int idle_policy(int policy) 171 { 172 return policy == SCHED_IDLE; 173 } 174 static inline int fair_policy(int policy) 175 { 176 return policy == SCHED_NORMAL || policy == SCHED_BATCH; 177 } 178 179 static inline int rt_policy(int policy) 180 { 181 return policy == SCHED_FIFO || policy == SCHED_RR; 182 } 183 184 static inline int dl_policy(int policy) 185 { 186 return policy == SCHED_DEADLINE; 187 } 188 static inline bool valid_policy(int policy) 189 { 190 return idle_policy(policy) || fair_policy(policy) || 191 rt_policy(policy) || dl_policy(policy); 192 } 193 194 static inline int task_has_idle_policy(struct task_struct *p) 195 { 196 return idle_policy(p->policy); 197 } 198 199 static inline int task_has_rt_policy(struct task_struct *p) 200 { 201 return rt_policy(p->policy); 202 } 203 204 static inline int task_has_dl_policy(struct task_struct *p) 205 { 206 return dl_policy(p->policy); 207 } 208 209 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT) 210 211 static inline void update_avg(u64 *avg, u64 sample) 212 { 213 s64 diff = sample - *avg; 214 *avg += diff / 8; 215 } 216 217 /* 218 * Shifting a value by an exponent greater *or equal* to the size of said value 219 * is UB; cap at size-1. 220 */ 221 #define shr_bound(val, shift) \ 222 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1)) 223 224 /* 225 * !! For sched_setattr_nocheck() (kernel) only !! 226 * 227 * This is actually gross. :( 228 * 229 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE 230 * tasks, but still be able to sleep. We need this on platforms that cannot 231 * atomically change clock frequency. Remove once fast switching will be 232 * available on such platforms. 233 * 234 * SUGOV stands for SchedUtil GOVernor. 235 */ 236 #define SCHED_FLAG_SUGOV 0x10000000 237 238 #define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV) 239 240 static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se) 241 { 242 #ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL 243 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV); 244 #else 245 return false; 246 #endif 247 } 248 249 /* 250 * Tells if entity @a should preempt entity @b. 251 */ 252 static inline bool dl_entity_preempt(const struct sched_dl_entity *a, 253 const struct sched_dl_entity *b) 254 { 255 return dl_entity_is_special(a) || 256 dl_time_before(a->deadline, b->deadline); 257 } 258 259 /* 260 * This is the priority-queue data structure of the RT scheduling class: 261 */ 262 struct rt_prio_array { 263 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ 264 struct list_head queue[MAX_RT_PRIO]; 265 }; 266 267 struct rt_bandwidth { 268 /* nests inside the rq lock: */ 269 raw_spinlock_t rt_runtime_lock; 270 ktime_t rt_period; 271 u64 rt_runtime; 272 struct hrtimer rt_period_timer; 273 unsigned int rt_period_active; 274 }; 275 276 static inline int dl_bandwidth_enabled(void) 277 { 278 return sysctl_sched_rt_runtime >= 0; 279 } 280 281 /* 282 * To keep the bandwidth of -deadline tasks under control 283 * we need some place where: 284 * - store the maximum -deadline bandwidth of each cpu; 285 * - cache the fraction of bandwidth that is currently allocated in 286 * each root domain; 287 * 288 * This is all done in the data structure below. It is similar to the 289 * one used for RT-throttling (rt_bandwidth), with the main difference 290 * that, since here we are only interested in admission control, we 291 * do not decrease any runtime while the group "executes", neither we 292 * need a timer to replenish it. 293 * 294 * With respect to SMP, bandwidth is given on a per root domain basis, 295 * meaning that: 296 * - bw (< 100%) is the deadline bandwidth of each CPU; 297 * - total_bw is the currently allocated bandwidth in each root domain; 298 */ 299 struct dl_bw { 300 raw_spinlock_t lock; 301 u64 bw; 302 u64 total_bw; 303 }; 304 305 extern void init_dl_bw(struct dl_bw *dl_b); 306 extern int sched_dl_global_validate(void); 307 extern void sched_dl_do_global(void); 308 extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr); 309 extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr); 310 extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr); 311 extern bool __checkparam_dl(const struct sched_attr *attr); 312 extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr); 313 extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial); 314 extern int dl_bw_check_overflow(int cpu); 315 316 /* 317 * SCHED_DEADLINE supports servers (nested scheduling) with the following 318 * interface: 319 * 320 * dl_se::rq -- runqueue we belong to. 321 * 322 * dl_se::server_has_tasks() -- used on bandwidth enforcement; we 'stop' the 323 * server when it runs out of tasks to run. 324 * 325 * dl_se::server_pick() -- nested pick_next_task(); we yield the period if this 326 * returns NULL. 327 * 328 * dl_server_update() -- called from update_curr_common(), propagates runtime 329 * to the server. 330 * 331 * dl_server_start() 332 * dl_server_stop() -- start/stop the server when it has (no) tasks. 333 * 334 * dl_server_init() -- initializes the server. 335 */ 336 extern void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec); 337 extern void dl_server_start(struct sched_dl_entity *dl_se); 338 extern void dl_server_stop(struct sched_dl_entity *dl_se); 339 extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq, 340 dl_server_has_tasks_f has_tasks, 341 dl_server_pick_f pick); 342 343 #ifdef CONFIG_CGROUP_SCHED 344 345 struct cfs_rq; 346 struct rt_rq; 347 348 extern struct list_head task_groups; 349 350 struct cfs_bandwidth { 351 #ifdef CONFIG_CFS_BANDWIDTH 352 raw_spinlock_t lock; 353 ktime_t period; 354 u64 quota; 355 u64 runtime; 356 u64 burst; 357 u64 runtime_snap; 358 s64 hierarchical_quota; 359 360 u8 idle; 361 u8 period_active; 362 u8 slack_started; 363 struct hrtimer period_timer; 364 struct hrtimer slack_timer; 365 struct list_head throttled_cfs_rq; 366 367 /* Statistics: */ 368 int nr_periods; 369 int nr_throttled; 370 int nr_burst; 371 u64 throttled_time; 372 u64 burst_time; 373 #endif 374 }; 375 376 /* Task group related information */ 377 struct task_group { 378 struct cgroup_subsys_state css; 379 380 #ifdef CONFIG_FAIR_GROUP_SCHED 381 /* schedulable entities of this group on each CPU */ 382 struct sched_entity **se; 383 /* runqueue "owned" by this group on each CPU */ 384 struct cfs_rq **cfs_rq; 385 unsigned long shares; 386 387 /* A positive value indicates that this is a SCHED_IDLE group. */ 388 int idle; 389 390 #ifdef CONFIG_SMP 391 /* 392 * load_avg can be heavily contended at clock tick time, so put 393 * it in its own cacheline separated from the fields above which 394 * will also be accessed at each tick. 395 */ 396 atomic_long_t load_avg ____cacheline_aligned; 397 #endif 398 #endif 399 400 #ifdef CONFIG_RT_GROUP_SCHED 401 struct sched_rt_entity **rt_se; 402 struct rt_rq **rt_rq; 403 404 struct rt_bandwidth rt_bandwidth; 405 #endif 406 407 struct rcu_head rcu; 408 struct list_head list; 409 410 struct task_group *parent; 411 struct list_head siblings; 412 struct list_head children; 413 414 #ifdef CONFIG_SCHED_AUTOGROUP 415 struct autogroup *autogroup; 416 #endif 417 418 struct cfs_bandwidth cfs_bandwidth; 419 420 #ifdef CONFIG_UCLAMP_TASK_GROUP 421 /* The two decimal precision [%] value requested from user-space */ 422 unsigned int uclamp_pct[UCLAMP_CNT]; 423 /* Clamp values requested for a task group */ 424 struct uclamp_se uclamp_req[UCLAMP_CNT]; 425 /* Effective clamp values used for a task group */ 426 struct uclamp_se uclamp[UCLAMP_CNT]; 427 #endif 428 429 }; 430 431 #ifdef CONFIG_FAIR_GROUP_SCHED 432 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD 433 434 /* 435 * A weight of 0 or 1 can cause arithmetics problems. 436 * A weight of a cfs_rq is the sum of weights of which entities 437 * are queued on this cfs_rq, so a weight of a entity should not be 438 * too large, so as the shares value of a task group. 439 * (The default weight is 1024 - so there's no practical 440 * limitation from this.) 441 */ 442 #define MIN_SHARES (1UL << 1) 443 #define MAX_SHARES (1UL << 18) 444 #endif 445 446 typedef int (*tg_visitor)(struct task_group *, void *); 447 448 extern int walk_tg_tree_from(struct task_group *from, 449 tg_visitor down, tg_visitor up, void *data); 450 451 /* 452 * Iterate the full tree, calling @down when first entering a node and @up when 453 * leaving it for the final time. 454 * 455 * Caller must hold rcu_lock or sufficient equivalent. 456 */ 457 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data) 458 { 459 return walk_tg_tree_from(&root_task_group, down, up, data); 460 } 461 462 extern int tg_nop(struct task_group *tg, void *data); 463 464 #ifdef CONFIG_FAIR_GROUP_SCHED 465 extern void free_fair_sched_group(struct task_group *tg); 466 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent); 467 extern void online_fair_sched_group(struct task_group *tg); 468 extern void unregister_fair_sched_group(struct task_group *tg); 469 #else 470 static inline void free_fair_sched_group(struct task_group *tg) { } 471 static inline int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 472 { 473 return 1; 474 } 475 static inline void online_fair_sched_group(struct task_group *tg) { } 476 static inline void unregister_fair_sched_group(struct task_group *tg) { } 477 #endif 478 479 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 480 struct sched_entity *se, int cpu, 481 struct sched_entity *parent); 482 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent); 483 484 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b); 485 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b); 486 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq); 487 extern bool cfs_task_bw_constrained(struct task_struct *p); 488 489 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, 490 struct sched_rt_entity *rt_se, int cpu, 491 struct sched_rt_entity *parent); 492 extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us); 493 extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us); 494 extern long sched_group_rt_runtime(struct task_group *tg); 495 extern long sched_group_rt_period(struct task_group *tg); 496 extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk); 497 498 extern struct task_group *sched_create_group(struct task_group *parent); 499 extern void sched_online_group(struct task_group *tg, 500 struct task_group *parent); 501 extern void sched_destroy_group(struct task_group *tg); 502 extern void sched_release_group(struct task_group *tg); 503 504 extern void sched_move_task(struct task_struct *tsk); 505 506 #ifdef CONFIG_FAIR_GROUP_SCHED 507 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares); 508 509 extern int sched_group_set_idle(struct task_group *tg, long idle); 510 511 #ifdef CONFIG_SMP 512 extern void set_task_rq_fair(struct sched_entity *se, 513 struct cfs_rq *prev, struct cfs_rq *next); 514 #else /* !CONFIG_SMP */ 515 static inline void set_task_rq_fair(struct sched_entity *se, 516 struct cfs_rq *prev, struct cfs_rq *next) { } 517 #endif /* CONFIG_SMP */ 518 #endif /* CONFIG_FAIR_GROUP_SCHED */ 519 520 #else /* CONFIG_CGROUP_SCHED */ 521 522 struct cfs_bandwidth { }; 523 static inline bool cfs_task_bw_constrained(struct task_struct *p) { return false; } 524 525 #endif /* CONFIG_CGROUP_SCHED */ 526 527 extern void unregister_rt_sched_group(struct task_group *tg); 528 extern void free_rt_sched_group(struct task_group *tg); 529 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent); 530 531 /* 532 * u64_u32_load/u64_u32_store 533 * 534 * Use a copy of a u64 value to protect against data race. This is only 535 * applicable for 32-bits architectures. 536 */ 537 #ifdef CONFIG_64BIT 538 # define u64_u32_load_copy(var, copy) var 539 # define u64_u32_store_copy(var, copy, val) (var = val) 540 #else 541 # define u64_u32_load_copy(var, copy) \ 542 ({ \ 543 u64 __val, __val_copy; \ 544 do { \ 545 __val_copy = copy; \ 546 /* \ 547 * paired with u64_u32_store_copy(), ordering access \ 548 * to var and copy. \ 549 */ \ 550 smp_rmb(); \ 551 __val = var; \ 552 } while (__val != __val_copy); \ 553 __val; \ 554 }) 555 # define u64_u32_store_copy(var, copy, val) \ 556 do { \ 557 typeof(val) __val = (val); \ 558 var = __val; \ 559 /* \ 560 * paired with u64_u32_load_copy(), ordering access to var and \ 561 * copy. \ 562 */ \ 563 smp_wmb(); \ 564 copy = __val; \ 565 } while (0) 566 #endif 567 # define u64_u32_load(var) u64_u32_load_copy(var, var##_copy) 568 # define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val) 569 570 /* CFS-related fields in a runqueue */ 571 struct cfs_rq { 572 struct load_weight load; 573 unsigned int nr_running; 574 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */ 575 unsigned int idle_nr_running; /* SCHED_IDLE */ 576 unsigned int idle_h_nr_running; /* SCHED_IDLE */ 577 578 s64 avg_vruntime; 579 u64 avg_load; 580 581 u64 exec_clock; 582 u64 min_vruntime; 583 #ifdef CONFIG_SCHED_CORE 584 unsigned int forceidle_seq; 585 u64 min_vruntime_fi; 586 #endif 587 588 #ifndef CONFIG_64BIT 589 u64 min_vruntime_copy; 590 #endif 591 592 struct rb_root_cached tasks_timeline; 593 594 /* 595 * 'curr' points to currently running entity on this cfs_rq. 596 * It is set to NULL otherwise (i.e when none are currently running). 597 */ 598 struct sched_entity *curr; 599 struct sched_entity *next; 600 601 #ifdef CONFIG_SCHED_DEBUG 602 unsigned int nr_spread_over; 603 #endif 604 605 #ifdef CONFIG_SMP 606 /* 607 * CFS load tracking 608 */ 609 struct sched_avg avg; 610 #ifndef CONFIG_64BIT 611 u64 last_update_time_copy; 612 #endif 613 struct { 614 raw_spinlock_t lock ____cacheline_aligned; 615 int nr; 616 unsigned long load_avg; 617 unsigned long util_avg; 618 unsigned long runnable_avg; 619 } removed; 620 621 #ifdef CONFIG_FAIR_GROUP_SCHED 622 u64 last_update_tg_load_avg; 623 unsigned long tg_load_avg_contrib; 624 long propagate; 625 long prop_runnable_sum; 626 627 /* 628 * h_load = weight * f(tg) 629 * 630 * Where f(tg) is the recursive weight fraction assigned to 631 * this group. 632 */ 633 unsigned long h_load; 634 u64 last_h_load_update; 635 struct sched_entity *h_load_next; 636 #endif /* CONFIG_FAIR_GROUP_SCHED */ 637 #endif /* CONFIG_SMP */ 638 639 #ifdef CONFIG_FAIR_GROUP_SCHED 640 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */ 641 642 /* 643 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in 644 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities 645 * (like users, containers etc.) 646 * 647 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU. 648 * This list is used during load balance. 649 */ 650 int on_list; 651 struct list_head leaf_cfs_rq_list; 652 struct task_group *tg; /* group that "owns" this runqueue */ 653 654 /* Locally cached copy of our task_group's idle value */ 655 int idle; 656 657 #ifdef CONFIG_CFS_BANDWIDTH 658 int runtime_enabled; 659 s64 runtime_remaining; 660 661 u64 throttled_pelt_idle; 662 #ifndef CONFIG_64BIT 663 u64 throttled_pelt_idle_copy; 664 #endif 665 u64 throttled_clock; 666 u64 throttled_clock_pelt; 667 u64 throttled_clock_pelt_time; 668 u64 throttled_clock_self; 669 u64 throttled_clock_self_time; 670 int throttled; 671 int throttle_count; 672 struct list_head throttled_list; 673 struct list_head throttled_csd_list; 674 #endif /* CONFIG_CFS_BANDWIDTH */ 675 #endif /* CONFIG_FAIR_GROUP_SCHED */ 676 }; 677 678 static inline int rt_bandwidth_enabled(void) 679 { 680 return sysctl_sched_rt_runtime >= 0; 681 } 682 683 /* RT IPI pull logic requires IRQ_WORK */ 684 #if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP) 685 # define HAVE_RT_PUSH_IPI 686 #endif 687 688 /* Real-Time classes' related field in a runqueue: */ 689 struct rt_rq { 690 struct rt_prio_array active; 691 unsigned int rt_nr_running; 692 unsigned int rr_nr_running; 693 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED 694 struct { 695 int curr; /* highest queued rt task prio */ 696 #ifdef CONFIG_SMP 697 int next; /* next highest */ 698 #endif 699 } highest_prio; 700 #endif 701 #ifdef CONFIG_SMP 702 int overloaded; 703 struct plist_head pushable_tasks; 704 705 #endif /* CONFIG_SMP */ 706 int rt_queued; 707 708 int rt_throttled; 709 u64 rt_time; 710 u64 rt_runtime; 711 /* Nests inside the rq lock: */ 712 raw_spinlock_t rt_runtime_lock; 713 714 #ifdef CONFIG_RT_GROUP_SCHED 715 unsigned int rt_nr_boosted; 716 717 struct rq *rq; 718 struct task_group *tg; 719 #endif 720 }; 721 722 static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq) 723 { 724 return rt_rq->rt_queued && rt_rq->rt_nr_running; 725 } 726 727 /* Deadline class' related fields in a runqueue */ 728 struct dl_rq { 729 /* runqueue is an rbtree, ordered by deadline */ 730 struct rb_root_cached root; 731 732 unsigned int dl_nr_running; 733 734 #ifdef CONFIG_SMP 735 /* 736 * Deadline values of the currently executing and the 737 * earliest ready task on this rq. Caching these facilitates 738 * the decision whether or not a ready but not running task 739 * should migrate somewhere else. 740 */ 741 struct { 742 u64 curr; 743 u64 next; 744 } earliest_dl; 745 746 int overloaded; 747 748 /* 749 * Tasks on this rq that can be pushed away. They are kept in 750 * an rb-tree, ordered by tasks' deadlines, with caching 751 * of the leftmost (earliest deadline) element. 752 */ 753 struct rb_root_cached pushable_dl_tasks_root; 754 #else 755 struct dl_bw dl_bw; 756 #endif 757 /* 758 * "Active utilization" for this runqueue: increased when a 759 * task wakes up (becomes TASK_RUNNING) and decreased when a 760 * task blocks 761 */ 762 u64 running_bw; 763 764 /* 765 * Utilization of the tasks "assigned" to this runqueue (including 766 * the tasks that are in runqueue and the tasks that executed on this 767 * CPU and blocked). Increased when a task moves to this runqueue, and 768 * decreased when the task moves away (migrates, changes scheduling 769 * policy, or terminates). 770 * This is needed to compute the "inactive utilization" for the 771 * runqueue (inactive utilization = this_bw - running_bw). 772 */ 773 u64 this_bw; 774 u64 extra_bw; 775 776 /* 777 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM 778 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB). 779 */ 780 u64 max_bw; 781 782 /* 783 * Inverse of the fraction of CPU utilization that can be reclaimed 784 * by the GRUB algorithm. 785 */ 786 u64 bw_ratio; 787 }; 788 789 #ifdef CONFIG_FAIR_GROUP_SCHED 790 /* An entity is a task if it doesn't "own" a runqueue */ 791 #define entity_is_task(se) (!se->my_q) 792 793 static inline void se_update_runnable(struct sched_entity *se) 794 { 795 if (!entity_is_task(se)) 796 se->runnable_weight = se->my_q->h_nr_running; 797 } 798 799 static inline long se_runnable(struct sched_entity *se) 800 { 801 if (entity_is_task(se)) 802 return !!se->on_rq; 803 else 804 return se->runnable_weight; 805 } 806 807 #else 808 #define entity_is_task(se) 1 809 810 static inline void se_update_runnable(struct sched_entity *se) {} 811 812 static inline long se_runnable(struct sched_entity *se) 813 { 814 return !!se->on_rq; 815 } 816 #endif 817 818 #ifdef CONFIG_SMP 819 /* 820 * XXX we want to get rid of these helpers and use the full load resolution. 821 */ 822 static inline long se_weight(struct sched_entity *se) 823 { 824 return scale_load_down(se->load.weight); 825 } 826 827 828 static inline bool sched_asym_prefer(int a, int b) 829 { 830 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b); 831 } 832 833 struct perf_domain { 834 struct em_perf_domain *em_pd; 835 struct perf_domain *next; 836 struct rcu_head rcu; 837 }; 838 839 /* Scheduling group status flags */ 840 #define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */ 841 #define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */ 842 843 /* 844 * We add the notion of a root-domain which will be used to define per-domain 845 * variables. Each exclusive cpuset essentially defines an island domain by 846 * fully partitioning the member CPUs from any other cpuset. Whenever a new 847 * exclusive cpuset is created, we also create and attach a new root-domain 848 * object. 849 * 850 */ 851 struct root_domain { 852 atomic_t refcount; 853 atomic_t rto_count; 854 struct rcu_head rcu; 855 cpumask_var_t span; 856 cpumask_var_t online; 857 858 /* 859 * Indicate pullable load on at least one CPU, e.g: 860 * - More than one runnable task 861 * - Running task is misfit 862 */ 863 int overload; 864 865 /* Indicate one or more cpus over-utilized (tipping point) */ 866 int overutilized; 867 868 /* 869 * The bit corresponding to a CPU gets set here if such CPU has more 870 * than one runnable -deadline task (as it is below for RT tasks). 871 */ 872 cpumask_var_t dlo_mask; 873 atomic_t dlo_count; 874 struct dl_bw dl_bw; 875 struct cpudl cpudl; 876 877 /* 878 * Indicate whether a root_domain's dl_bw has been checked or 879 * updated. It's monotonously increasing value. 880 * 881 * Also, some corner cases, like 'wrap around' is dangerous, but given 882 * that u64 is 'big enough'. So that shouldn't be a concern. 883 */ 884 u64 visit_gen; 885 886 #ifdef HAVE_RT_PUSH_IPI 887 /* 888 * For IPI pull requests, loop across the rto_mask. 889 */ 890 struct irq_work rto_push_work; 891 raw_spinlock_t rto_lock; 892 /* These are only updated and read within rto_lock */ 893 int rto_loop; 894 int rto_cpu; 895 /* These atomics are updated outside of a lock */ 896 atomic_t rto_loop_next; 897 atomic_t rto_loop_start; 898 #endif 899 /* 900 * The "RT overload" flag: it gets set if a CPU has more than 901 * one runnable RT task. 902 */ 903 cpumask_var_t rto_mask; 904 struct cpupri cpupri; 905 906 unsigned long max_cpu_capacity; 907 908 /* 909 * NULL-terminated list of performance domains intersecting with the 910 * CPUs of the rd. Protected by RCU. 911 */ 912 struct perf_domain __rcu *pd; 913 }; 914 915 extern void init_defrootdomain(void); 916 extern int sched_init_domains(const struct cpumask *cpu_map); 917 extern void rq_attach_root(struct rq *rq, struct root_domain *rd); 918 extern void sched_get_rd(struct root_domain *rd); 919 extern void sched_put_rd(struct root_domain *rd); 920 921 #ifdef HAVE_RT_PUSH_IPI 922 extern void rto_push_irq_work_func(struct irq_work *work); 923 #endif 924 #endif /* CONFIG_SMP */ 925 926 #ifdef CONFIG_UCLAMP_TASK 927 /* 928 * struct uclamp_bucket - Utilization clamp bucket 929 * @value: utilization clamp value for tasks on this clamp bucket 930 * @tasks: number of RUNNABLE tasks on this clamp bucket 931 * 932 * Keep track of how many tasks are RUNNABLE for a given utilization 933 * clamp value. 934 */ 935 struct uclamp_bucket { 936 unsigned long value : bits_per(SCHED_CAPACITY_SCALE); 937 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE); 938 }; 939 940 /* 941 * struct uclamp_rq - rq's utilization clamp 942 * @value: currently active clamp values for a rq 943 * @bucket: utilization clamp buckets affecting a rq 944 * 945 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values. 946 * A clamp value is affecting a rq when there is at least one task RUNNABLE 947 * (or actually running) with that value. 948 * 949 * There are up to UCLAMP_CNT possible different clamp values, currently there 950 * are only two: minimum utilization and maximum utilization. 951 * 952 * All utilization clamping values are MAX aggregated, since: 953 * - for util_min: we want to run the CPU at least at the max of the minimum 954 * utilization required by its currently RUNNABLE tasks. 955 * - for util_max: we want to allow the CPU to run up to the max of the 956 * maximum utilization allowed by its currently RUNNABLE tasks. 957 * 958 * Since on each system we expect only a limited number of different 959 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track 960 * the metrics required to compute all the per-rq utilization clamp values. 961 */ 962 struct uclamp_rq { 963 unsigned int value; 964 struct uclamp_bucket bucket[UCLAMP_BUCKETS]; 965 }; 966 967 DECLARE_STATIC_KEY_FALSE(sched_uclamp_used); 968 #endif /* CONFIG_UCLAMP_TASK */ 969 970 struct rq; 971 struct balance_callback { 972 struct balance_callback *next; 973 void (*func)(struct rq *rq); 974 }; 975 976 /* 977 * This is the main, per-CPU runqueue data structure. 978 * 979 * Locking rule: those places that want to lock multiple runqueues 980 * (such as the load balancing or the thread migration code), lock 981 * acquire operations must be ordered by ascending &runqueue. 982 */ 983 struct rq { 984 /* runqueue lock: */ 985 raw_spinlock_t __lock; 986 987 unsigned int nr_running; 988 #ifdef CONFIG_NUMA_BALANCING 989 unsigned int nr_numa_running; 990 unsigned int nr_preferred_running; 991 unsigned int numa_migrate_on; 992 #endif 993 #ifdef CONFIG_NO_HZ_COMMON 994 #ifdef CONFIG_SMP 995 unsigned long last_blocked_load_update_tick; 996 unsigned int has_blocked_load; 997 call_single_data_t nohz_csd; 998 #endif /* CONFIG_SMP */ 999 unsigned int nohz_tick_stopped; 1000 atomic_t nohz_flags; 1001 #endif /* CONFIG_NO_HZ_COMMON */ 1002 1003 #ifdef CONFIG_SMP 1004 unsigned int ttwu_pending; 1005 #endif 1006 u64 nr_switches; 1007 1008 #ifdef CONFIG_UCLAMP_TASK 1009 /* Utilization clamp values based on CPU's RUNNABLE tasks */ 1010 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned; 1011 unsigned int uclamp_flags; 1012 #define UCLAMP_FLAG_IDLE 0x01 1013 #endif 1014 1015 struct cfs_rq cfs; 1016 struct rt_rq rt; 1017 struct dl_rq dl; 1018 1019 #ifdef CONFIG_FAIR_GROUP_SCHED 1020 /* list of leaf cfs_rq on this CPU: */ 1021 struct list_head leaf_cfs_rq_list; 1022 struct list_head *tmp_alone_branch; 1023 #endif /* CONFIG_FAIR_GROUP_SCHED */ 1024 1025 /* 1026 * This is part of a global counter where only the total sum 1027 * over all CPUs matters. A task can increase this counter on 1028 * one CPU and if it got migrated afterwards it may decrease 1029 * it on another CPU. Always updated under the runqueue lock: 1030 */ 1031 unsigned int nr_uninterruptible; 1032 1033 struct task_struct __rcu *curr; 1034 struct task_struct *idle; 1035 struct task_struct *stop; 1036 unsigned long next_balance; 1037 struct mm_struct *prev_mm; 1038 1039 unsigned int clock_update_flags; 1040 u64 clock; 1041 /* Ensure that all clocks are in the same cache line */ 1042 u64 clock_task ____cacheline_aligned; 1043 u64 clock_pelt; 1044 unsigned long lost_idle_time; 1045 u64 clock_pelt_idle; 1046 u64 clock_idle; 1047 #ifndef CONFIG_64BIT 1048 u64 clock_pelt_idle_copy; 1049 u64 clock_idle_copy; 1050 #endif 1051 1052 atomic_t nr_iowait; 1053 1054 #ifdef CONFIG_SCHED_DEBUG 1055 u64 last_seen_need_resched_ns; 1056 int ticks_without_resched; 1057 #endif 1058 1059 #ifdef CONFIG_MEMBARRIER 1060 int membarrier_state; 1061 #endif 1062 1063 #ifdef CONFIG_SMP 1064 struct root_domain *rd; 1065 struct sched_domain __rcu *sd; 1066 1067 unsigned long cpu_capacity; 1068 1069 struct balance_callback *balance_callback; 1070 1071 unsigned char nohz_idle_balance; 1072 unsigned char idle_balance; 1073 1074 unsigned long misfit_task_load; 1075 1076 /* For active balancing */ 1077 int active_balance; 1078 int push_cpu; 1079 struct cpu_stop_work active_balance_work; 1080 1081 /* CPU of this runqueue: */ 1082 int cpu; 1083 int online; 1084 1085 struct list_head cfs_tasks; 1086 1087 struct sched_avg avg_rt; 1088 struct sched_avg avg_dl; 1089 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 1090 struct sched_avg avg_irq; 1091 #endif 1092 #ifdef CONFIG_SCHED_THERMAL_PRESSURE 1093 struct sched_avg avg_thermal; 1094 #endif 1095 u64 idle_stamp; 1096 u64 avg_idle; 1097 1098 /* This is used to determine avg_idle's max value */ 1099 u64 max_idle_balance_cost; 1100 1101 #ifdef CONFIG_HOTPLUG_CPU 1102 struct rcuwait hotplug_wait; 1103 #endif 1104 #endif /* CONFIG_SMP */ 1105 1106 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 1107 u64 prev_irq_time; 1108 #endif 1109 #ifdef CONFIG_PARAVIRT 1110 u64 prev_steal_time; 1111 #endif 1112 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 1113 u64 prev_steal_time_rq; 1114 #endif 1115 1116 /* calc_load related fields */ 1117 unsigned long calc_load_update; 1118 long calc_load_active; 1119 1120 #ifdef CONFIG_SCHED_HRTICK 1121 #ifdef CONFIG_SMP 1122 call_single_data_t hrtick_csd; 1123 #endif 1124 struct hrtimer hrtick_timer; 1125 ktime_t hrtick_time; 1126 #endif 1127 1128 #ifdef CONFIG_SCHEDSTATS 1129 /* latency stats */ 1130 struct sched_info rq_sched_info; 1131 unsigned long long rq_cpu_time; 1132 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */ 1133 1134 /* sys_sched_yield() stats */ 1135 unsigned int yld_count; 1136 1137 /* schedule() stats */ 1138 unsigned int sched_count; 1139 unsigned int sched_goidle; 1140 1141 /* try_to_wake_up() stats */ 1142 unsigned int ttwu_count; 1143 unsigned int ttwu_local; 1144 #endif 1145 1146 #ifdef CONFIG_CPU_IDLE 1147 /* Must be inspected within a rcu lock section */ 1148 struct cpuidle_state *idle_state; 1149 #endif 1150 1151 #ifdef CONFIG_SMP 1152 unsigned int nr_pinned; 1153 #endif 1154 unsigned int push_busy; 1155 struct cpu_stop_work push_work; 1156 1157 #ifdef CONFIG_SCHED_CORE 1158 /* per rq */ 1159 struct rq *core; 1160 struct task_struct *core_pick; 1161 unsigned int core_enabled; 1162 unsigned int core_sched_seq; 1163 struct rb_root core_tree; 1164 1165 /* shared state -- careful with sched_core_cpu_deactivate() */ 1166 unsigned int core_task_seq; 1167 unsigned int core_pick_seq; 1168 unsigned long core_cookie; 1169 unsigned int core_forceidle_count; 1170 unsigned int core_forceidle_seq; 1171 unsigned int core_forceidle_occupation; 1172 u64 core_forceidle_start; 1173 #endif 1174 1175 /* Scratch cpumask to be temporarily used under rq_lock */ 1176 cpumask_var_t scratch_mask; 1177 1178 #if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP) 1179 call_single_data_t cfsb_csd; 1180 struct list_head cfsb_csd_list; 1181 #endif 1182 }; 1183 1184 #ifdef CONFIG_FAIR_GROUP_SCHED 1185 1186 /* CPU runqueue to which this cfs_rq is attached */ 1187 static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 1188 { 1189 return cfs_rq->rq; 1190 } 1191 1192 #else 1193 1194 static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 1195 { 1196 return container_of(cfs_rq, struct rq, cfs); 1197 } 1198 #endif 1199 1200 static inline int cpu_of(struct rq *rq) 1201 { 1202 #ifdef CONFIG_SMP 1203 return rq->cpu; 1204 #else 1205 return 0; 1206 #endif 1207 } 1208 1209 #define MDF_PUSH 0x01 1210 1211 static inline bool is_migration_disabled(struct task_struct *p) 1212 { 1213 #ifdef CONFIG_SMP 1214 return p->migration_disabled; 1215 #else 1216 return false; 1217 #endif 1218 } 1219 1220 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 1221 1222 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) 1223 #define this_rq() this_cpu_ptr(&runqueues) 1224 #define task_rq(p) cpu_rq(task_cpu(p)) 1225 #define cpu_curr(cpu) (cpu_rq(cpu)->curr) 1226 #define raw_rq() raw_cpu_ptr(&runqueues) 1227 1228 struct sched_group; 1229 #ifdef CONFIG_SCHED_CORE 1230 static inline struct cpumask *sched_group_span(struct sched_group *sg); 1231 1232 DECLARE_STATIC_KEY_FALSE(__sched_core_enabled); 1233 1234 static inline bool sched_core_enabled(struct rq *rq) 1235 { 1236 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled; 1237 } 1238 1239 static inline bool sched_core_disabled(void) 1240 { 1241 return !static_branch_unlikely(&__sched_core_enabled); 1242 } 1243 1244 /* 1245 * Be careful with this function; not for general use. The return value isn't 1246 * stable unless you actually hold a relevant rq->__lock. 1247 */ 1248 static inline raw_spinlock_t *rq_lockp(struct rq *rq) 1249 { 1250 if (sched_core_enabled(rq)) 1251 return &rq->core->__lock; 1252 1253 return &rq->__lock; 1254 } 1255 1256 static inline raw_spinlock_t *__rq_lockp(struct rq *rq) 1257 { 1258 if (rq->core_enabled) 1259 return &rq->core->__lock; 1260 1261 return &rq->__lock; 1262 } 1263 1264 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b, 1265 bool fi); 1266 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi); 1267 1268 /* 1269 * Helpers to check if the CPU's core cookie matches with the task's cookie 1270 * when core scheduling is enabled. 1271 * A special case is that the task's cookie always matches with CPU's core 1272 * cookie if the CPU is in an idle core. 1273 */ 1274 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p) 1275 { 1276 /* Ignore cookie match if core scheduler is not enabled on the CPU. */ 1277 if (!sched_core_enabled(rq)) 1278 return true; 1279 1280 return rq->core->core_cookie == p->core_cookie; 1281 } 1282 1283 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p) 1284 { 1285 bool idle_core = true; 1286 int cpu; 1287 1288 /* Ignore cookie match if core scheduler is not enabled on the CPU. */ 1289 if (!sched_core_enabled(rq)) 1290 return true; 1291 1292 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) { 1293 if (!available_idle_cpu(cpu)) { 1294 idle_core = false; 1295 break; 1296 } 1297 } 1298 1299 /* 1300 * A CPU in an idle core is always the best choice for tasks with 1301 * cookies. 1302 */ 1303 return idle_core || rq->core->core_cookie == p->core_cookie; 1304 } 1305 1306 static inline bool sched_group_cookie_match(struct rq *rq, 1307 struct task_struct *p, 1308 struct sched_group *group) 1309 { 1310 int cpu; 1311 1312 /* Ignore cookie match if core scheduler is not enabled on the CPU. */ 1313 if (!sched_core_enabled(rq)) 1314 return true; 1315 1316 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) { 1317 if (sched_core_cookie_match(cpu_rq(cpu), p)) 1318 return true; 1319 } 1320 return false; 1321 } 1322 1323 static inline bool sched_core_enqueued(struct task_struct *p) 1324 { 1325 return !RB_EMPTY_NODE(&p->core_node); 1326 } 1327 1328 extern void sched_core_enqueue(struct rq *rq, struct task_struct *p); 1329 extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags); 1330 1331 extern void sched_core_get(void); 1332 extern void sched_core_put(void); 1333 1334 #else /* !CONFIG_SCHED_CORE */ 1335 1336 static inline bool sched_core_enabled(struct rq *rq) 1337 { 1338 return false; 1339 } 1340 1341 static inline bool sched_core_disabled(void) 1342 { 1343 return true; 1344 } 1345 1346 static inline raw_spinlock_t *rq_lockp(struct rq *rq) 1347 { 1348 return &rq->__lock; 1349 } 1350 1351 static inline raw_spinlock_t *__rq_lockp(struct rq *rq) 1352 { 1353 return &rq->__lock; 1354 } 1355 1356 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p) 1357 { 1358 return true; 1359 } 1360 1361 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p) 1362 { 1363 return true; 1364 } 1365 1366 static inline bool sched_group_cookie_match(struct rq *rq, 1367 struct task_struct *p, 1368 struct sched_group *group) 1369 { 1370 return true; 1371 } 1372 #endif /* CONFIG_SCHED_CORE */ 1373 1374 static inline void lockdep_assert_rq_held(struct rq *rq) 1375 { 1376 lockdep_assert_held(__rq_lockp(rq)); 1377 } 1378 1379 extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass); 1380 extern bool raw_spin_rq_trylock(struct rq *rq); 1381 extern void raw_spin_rq_unlock(struct rq *rq); 1382 1383 static inline void raw_spin_rq_lock(struct rq *rq) 1384 { 1385 raw_spin_rq_lock_nested(rq, 0); 1386 } 1387 1388 static inline void raw_spin_rq_lock_irq(struct rq *rq) 1389 { 1390 local_irq_disable(); 1391 raw_spin_rq_lock(rq); 1392 } 1393 1394 static inline void raw_spin_rq_unlock_irq(struct rq *rq) 1395 { 1396 raw_spin_rq_unlock(rq); 1397 local_irq_enable(); 1398 } 1399 1400 static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq) 1401 { 1402 unsigned long flags; 1403 local_irq_save(flags); 1404 raw_spin_rq_lock(rq); 1405 return flags; 1406 } 1407 1408 static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags) 1409 { 1410 raw_spin_rq_unlock(rq); 1411 local_irq_restore(flags); 1412 } 1413 1414 #define raw_spin_rq_lock_irqsave(rq, flags) \ 1415 do { \ 1416 flags = _raw_spin_rq_lock_irqsave(rq); \ 1417 } while (0) 1418 1419 #ifdef CONFIG_SCHED_SMT 1420 extern void __update_idle_core(struct rq *rq); 1421 1422 static inline void update_idle_core(struct rq *rq) 1423 { 1424 if (static_branch_unlikely(&sched_smt_present)) 1425 __update_idle_core(rq); 1426 } 1427 1428 #else 1429 static inline void update_idle_core(struct rq *rq) { } 1430 #endif 1431 1432 #ifdef CONFIG_FAIR_GROUP_SCHED 1433 static inline struct task_struct *task_of(struct sched_entity *se) 1434 { 1435 SCHED_WARN_ON(!entity_is_task(se)); 1436 return container_of(se, struct task_struct, se); 1437 } 1438 1439 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 1440 { 1441 return p->se.cfs_rq; 1442 } 1443 1444 /* runqueue on which this entity is (to be) queued */ 1445 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se) 1446 { 1447 return se->cfs_rq; 1448 } 1449 1450 /* runqueue "owned" by this group */ 1451 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 1452 { 1453 return grp->my_q; 1454 } 1455 1456 #else 1457 1458 #define task_of(_se) container_of(_se, struct task_struct, se) 1459 1460 static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p) 1461 { 1462 return &task_rq(p)->cfs; 1463 } 1464 1465 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se) 1466 { 1467 const struct task_struct *p = task_of(se); 1468 struct rq *rq = task_rq(p); 1469 1470 return &rq->cfs; 1471 } 1472 1473 /* runqueue "owned" by this group */ 1474 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 1475 { 1476 return NULL; 1477 } 1478 #endif 1479 1480 extern void update_rq_clock(struct rq *rq); 1481 1482 /* 1483 * rq::clock_update_flags bits 1484 * 1485 * %RQCF_REQ_SKIP - will request skipping of clock update on the next 1486 * call to __schedule(). This is an optimisation to avoid 1487 * neighbouring rq clock updates. 1488 * 1489 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is 1490 * in effect and calls to update_rq_clock() are being ignored. 1491 * 1492 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been 1493 * made to update_rq_clock() since the last time rq::lock was pinned. 1494 * 1495 * If inside of __schedule(), clock_update_flags will have been 1496 * shifted left (a left shift is a cheap operation for the fast path 1497 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use, 1498 * 1499 * if (rq-clock_update_flags >= RQCF_UPDATED) 1500 * 1501 * to check if %RQCF_UPDATED is set. It'll never be shifted more than 1502 * one position though, because the next rq_unpin_lock() will shift it 1503 * back. 1504 */ 1505 #define RQCF_REQ_SKIP 0x01 1506 #define RQCF_ACT_SKIP 0x02 1507 #define RQCF_UPDATED 0x04 1508 1509 static inline void assert_clock_updated(struct rq *rq) 1510 { 1511 /* 1512 * The only reason for not seeing a clock update since the 1513 * last rq_pin_lock() is if we're currently skipping updates. 1514 */ 1515 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP); 1516 } 1517 1518 static inline u64 rq_clock(struct rq *rq) 1519 { 1520 lockdep_assert_rq_held(rq); 1521 assert_clock_updated(rq); 1522 1523 return rq->clock; 1524 } 1525 1526 static inline u64 rq_clock_task(struct rq *rq) 1527 { 1528 lockdep_assert_rq_held(rq); 1529 assert_clock_updated(rq); 1530 1531 return rq->clock_task; 1532 } 1533 1534 /** 1535 * By default the decay is the default pelt decay period. 1536 * The decay shift can change the decay period in 1537 * multiples of 32. 1538 * Decay shift Decay period(ms) 1539 * 0 32 1540 * 1 64 1541 * 2 128 1542 * 3 256 1543 * 4 512 1544 */ 1545 extern int sched_thermal_decay_shift; 1546 1547 static inline u64 rq_clock_thermal(struct rq *rq) 1548 { 1549 return rq_clock_task(rq) >> sched_thermal_decay_shift; 1550 } 1551 1552 static inline void rq_clock_skip_update(struct rq *rq) 1553 { 1554 lockdep_assert_rq_held(rq); 1555 rq->clock_update_flags |= RQCF_REQ_SKIP; 1556 } 1557 1558 /* 1559 * See rt task throttling, which is the only time a skip 1560 * request is canceled. 1561 */ 1562 static inline void rq_clock_cancel_skipupdate(struct rq *rq) 1563 { 1564 lockdep_assert_rq_held(rq); 1565 rq->clock_update_flags &= ~RQCF_REQ_SKIP; 1566 } 1567 1568 /* 1569 * During cpu offlining and rq wide unthrottling, we can trigger 1570 * an update_rq_clock() for several cfs and rt runqueues (Typically 1571 * when using list_for_each_entry_*) 1572 * rq_clock_start_loop_update() can be called after updating the clock 1573 * once and before iterating over the list to prevent multiple update. 1574 * After the iterative traversal, we need to call rq_clock_stop_loop_update() 1575 * to clear RQCF_ACT_SKIP of rq->clock_update_flags. 1576 */ 1577 static inline void rq_clock_start_loop_update(struct rq *rq) 1578 { 1579 lockdep_assert_rq_held(rq); 1580 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP); 1581 rq->clock_update_flags |= RQCF_ACT_SKIP; 1582 } 1583 1584 static inline void rq_clock_stop_loop_update(struct rq *rq) 1585 { 1586 lockdep_assert_rq_held(rq); 1587 rq->clock_update_flags &= ~RQCF_ACT_SKIP; 1588 } 1589 1590 struct rq_flags { 1591 unsigned long flags; 1592 struct pin_cookie cookie; 1593 #ifdef CONFIG_SCHED_DEBUG 1594 /* 1595 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the 1596 * current pin context is stashed here in case it needs to be 1597 * restored in rq_repin_lock(). 1598 */ 1599 unsigned int clock_update_flags; 1600 #endif 1601 }; 1602 1603 extern struct balance_callback balance_push_callback; 1604 1605 /* 1606 * Lockdep annotation that avoids accidental unlocks; it's like a 1607 * sticky/continuous lockdep_assert_held(). 1608 * 1609 * This avoids code that has access to 'struct rq *rq' (basically everything in 1610 * the scheduler) from accidentally unlocking the rq if they do not also have a 1611 * copy of the (on-stack) 'struct rq_flags rf'. 1612 * 1613 * Also see Documentation/locking/lockdep-design.rst. 1614 */ 1615 static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf) 1616 { 1617 rf->cookie = lockdep_pin_lock(__rq_lockp(rq)); 1618 1619 #ifdef CONFIG_SCHED_DEBUG 1620 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP); 1621 rf->clock_update_flags = 0; 1622 #ifdef CONFIG_SMP 1623 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback); 1624 #endif 1625 #endif 1626 } 1627 1628 static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf) 1629 { 1630 #ifdef CONFIG_SCHED_DEBUG 1631 if (rq->clock_update_flags > RQCF_ACT_SKIP) 1632 rf->clock_update_flags = RQCF_UPDATED; 1633 #endif 1634 1635 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie); 1636 } 1637 1638 static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf) 1639 { 1640 lockdep_repin_lock(__rq_lockp(rq), rf->cookie); 1641 1642 #ifdef CONFIG_SCHED_DEBUG 1643 /* 1644 * Restore the value we stashed in @rf for this pin context. 1645 */ 1646 rq->clock_update_flags |= rf->clock_update_flags; 1647 #endif 1648 } 1649 1650 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1651 __acquires(rq->lock); 1652 1653 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf) 1654 __acquires(p->pi_lock) 1655 __acquires(rq->lock); 1656 1657 static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf) 1658 __releases(rq->lock) 1659 { 1660 rq_unpin_lock(rq, rf); 1661 raw_spin_rq_unlock(rq); 1662 } 1663 1664 static inline void 1665 task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf) 1666 __releases(rq->lock) 1667 __releases(p->pi_lock) 1668 { 1669 rq_unpin_lock(rq, rf); 1670 raw_spin_rq_unlock(rq); 1671 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags); 1672 } 1673 1674 DEFINE_LOCK_GUARD_1(task_rq_lock, struct task_struct, 1675 _T->rq = task_rq_lock(_T->lock, &_T->rf), 1676 task_rq_unlock(_T->rq, _T->lock, &_T->rf), 1677 struct rq *rq; struct rq_flags rf) 1678 1679 static inline void 1680 rq_lock_irqsave(struct rq *rq, struct rq_flags *rf) 1681 __acquires(rq->lock) 1682 { 1683 raw_spin_rq_lock_irqsave(rq, rf->flags); 1684 rq_pin_lock(rq, rf); 1685 } 1686 1687 static inline void 1688 rq_lock_irq(struct rq *rq, struct rq_flags *rf) 1689 __acquires(rq->lock) 1690 { 1691 raw_spin_rq_lock_irq(rq); 1692 rq_pin_lock(rq, rf); 1693 } 1694 1695 static inline void 1696 rq_lock(struct rq *rq, struct rq_flags *rf) 1697 __acquires(rq->lock) 1698 { 1699 raw_spin_rq_lock(rq); 1700 rq_pin_lock(rq, rf); 1701 } 1702 1703 static inline void 1704 rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf) 1705 __releases(rq->lock) 1706 { 1707 rq_unpin_lock(rq, rf); 1708 raw_spin_rq_unlock_irqrestore(rq, rf->flags); 1709 } 1710 1711 static inline void 1712 rq_unlock_irq(struct rq *rq, struct rq_flags *rf) 1713 __releases(rq->lock) 1714 { 1715 rq_unpin_lock(rq, rf); 1716 raw_spin_rq_unlock_irq(rq); 1717 } 1718 1719 static inline void 1720 rq_unlock(struct rq *rq, struct rq_flags *rf) 1721 __releases(rq->lock) 1722 { 1723 rq_unpin_lock(rq, rf); 1724 raw_spin_rq_unlock(rq); 1725 } 1726 1727 DEFINE_LOCK_GUARD_1(rq_lock, struct rq, 1728 rq_lock(_T->lock, &_T->rf), 1729 rq_unlock(_T->lock, &_T->rf), 1730 struct rq_flags rf) 1731 1732 DEFINE_LOCK_GUARD_1(rq_lock_irq, struct rq, 1733 rq_lock_irq(_T->lock, &_T->rf), 1734 rq_unlock_irq(_T->lock, &_T->rf), 1735 struct rq_flags rf) 1736 1737 DEFINE_LOCK_GUARD_1(rq_lock_irqsave, struct rq, 1738 rq_lock_irqsave(_T->lock, &_T->rf), 1739 rq_unlock_irqrestore(_T->lock, &_T->rf), 1740 struct rq_flags rf) 1741 1742 static inline struct rq * 1743 this_rq_lock_irq(struct rq_flags *rf) 1744 __acquires(rq->lock) 1745 { 1746 struct rq *rq; 1747 1748 local_irq_disable(); 1749 rq = this_rq(); 1750 rq_lock(rq, rf); 1751 return rq; 1752 } 1753 1754 #ifdef CONFIG_NUMA 1755 enum numa_topology_type { 1756 NUMA_DIRECT, 1757 NUMA_GLUELESS_MESH, 1758 NUMA_BACKPLANE, 1759 }; 1760 extern enum numa_topology_type sched_numa_topology_type; 1761 extern int sched_max_numa_distance; 1762 extern bool find_numa_distance(int distance); 1763 extern void sched_init_numa(int offline_node); 1764 extern void sched_update_numa(int cpu, bool online); 1765 extern void sched_domains_numa_masks_set(unsigned int cpu); 1766 extern void sched_domains_numa_masks_clear(unsigned int cpu); 1767 extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu); 1768 #else 1769 static inline void sched_init_numa(int offline_node) { } 1770 static inline void sched_update_numa(int cpu, bool online) { } 1771 static inline void sched_domains_numa_masks_set(unsigned int cpu) { } 1772 static inline void sched_domains_numa_masks_clear(unsigned int cpu) { } 1773 static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu) 1774 { 1775 return nr_cpu_ids; 1776 } 1777 #endif 1778 1779 #ifdef CONFIG_NUMA_BALANCING 1780 /* The regions in numa_faults array from task_struct */ 1781 enum numa_faults_stats { 1782 NUMA_MEM = 0, 1783 NUMA_CPU, 1784 NUMA_MEMBUF, 1785 NUMA_CPUBUF 1786 }; 1787 extern void sched_setnuma(struct task_struct *p, int node); 1788 extern int migrate_task_to(struct task_struct *p, int cpu); 1789 extern int migrate_swap(struct task_struct *p, struct task_struct *t, 1790 int cpu, int scpu); 1791 extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p); 1792 #else 1793 static inline void 1794 init_numa_balancing(unsigned long clone_flags, struct task_struct *p) 1795 { 1796 } 1797 #endif /* CONFIG_NUMA_BALANCING */ 1798 1799 #ifdef CONFIG_SMP 1800 1801 static inline void 1802 queue_balance_callback(struct rq *rq, 1803 struct balance_callback *head, 1804 void (*func)(struct rq *rq)) 1805 { 1806 lockdep_assert_rq_held(rq); 1807 1808 /* 1809 * Don't (re)queue an already queued item; nor queue anything when 1810 * balance_push() is active, see the comment with 1811 * balance_push_callback. 1812 */ 1813 if (unlikely(head->next || rq->balance_callback == &balance_push_callback)) 1814 return; 1815 1816 head->func = func; 1817 head->next = rq->balance_callback; 1818 rq->balance_callback = head; 1819 } 1820 1821 #define rcu_dereference_check_sched_domain(p) \ 1822 rcu_dereference_check((p), \ 1823 lockdep_is_held(&sched_domains_mutex)) 1824 1825 /* 1826 * The domain tree (rq->sd) is protected by RCU's quiescent state transition. 1827 * See destroy_sched_domains: call_rcu for details. 1828 * 1829 * The domain tree of any CPU may only be accessed from within 1830 * preempt-disabled sections. 1831 */ 1832 #define for_each_domain(cpu, __sd) \ 1833 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \ 1834 __sd; __sd = __sd->parent) 1835 1836 /* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */ 1837 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) | 1838 static const unsigned int SD_SHARED_CHILD_MASK = 1839 #include <linux/sched/sd_flags.h> 1840 0; 1841 #undef SD_FLAG 1842 1843 /** 1844 * highest_flag_domain - Return highest sched_domain containing flag. 1845 * @cpu: The CPU whose highest level of sched domain is to 1846 * be returned. 1847 * @flag: The flag to check for the highest sched_domain 1848 * for the given CPU. 1849 * 1850 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has 1851 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag. 1852 */ 1853 static inline struct sched_domain *highest_flag_domain(int cpu, int flag) 1854 { 1855 struct sched_domain *sd, *hsd = NULL; 1856 1857 for_each_domain(cpu, sd) { 1858 if (sd->flags & flag) { 1859 hsd = sd; 1860 continue; 1861 } 1862 1863 /* 1864 * Stop the search if @flag is known to be shared at lower 1865 * levels. It will not be found further up. 1866 */ 1867 if (flag & SD_SHARED_CHILD_MASK) 1868 break; 1869 } 1870 1871 return hsd; 1872 } 1873 1874 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) 1875 { 1876 struct sched_domain *sd; 1877 1878 for_each_domain(cpu, sd) { 1879 if (sd->flags & flag) 1880 break; 1881 } 1882 1883 return sd; 1884 } 1885 1886 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc); 1887 DECLARE_PER_CPU(int, sd_llc_size); 1888 DECLARE_PER_CPU(int, sd_llc_id); 1889 DECLARE_PER_CPU(int, sd_share_id); 1890 DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared); 1891 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa); 1892 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing); 1893 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity); 1894 extern struct static_key_false sched_asym_cpucapacity; 1895 extern struct static_key_false sched_cluster_active; 1896 1897 static __always_inline bool sched_asym_cpucap_active(void) 1898 { 1899 return static_branch_unlikely(&sched_asym_cpucapacity); 1900 } 1901 1902 struct sched_group_capacity { 1903 atomic_t ref; 1904 /* 1905 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity 1906 * for a single CPU. 1907 */ 1908 unsigned long capacity; 1909 unsigned long min_capacity; /* Min per-CPU capacity in group */ 1910 unsigned long max_capacity; /* Max per-CPU capacity in group */ 1911 unsigned long next_update; 1912 int imbalance; /* XXX unrelated to capacity but shared group state */ 1913 1914 #ifdef CONFIG_SCHED_DEBUG 1915 int id; 1916 #endif 1917 1918 unsigned long cpumask[]; /* Balance mask */ 1919 }; 1920 1921 struct sched_group { 1922 struct sched_group *next; /* Must be a circular list */ 1923 atomic_t ref; 1924 1925 unsigned int group_weight; 1926 unsigned int cores; 1927 struct sched_group_capacity *sgc; 1928 int asym_prefer_cpu; /* CPU of highest priority in group */ 1929 int flags; 1930 1931 /* 1932 * The CPUs this group covers. 1933 * 1934 * NOTE: this field is variable length. (Allocated dynamically 1935 * by attaching extra space to the end of the structure, 1936 * depending on how many CPUs the kernel has booted up with) 1937 */ 1938 unsigned long cpumask[]; 1939 }; 1940 1941 static inline struct cpumask *sched_group_span(struct sched_group *sg) 1942 { 1943 return to_cpumask(sg->cpumask); 1944 } 1945 1946 /* 1947 * See build_balance_mask(). 1948 */ 1949 static inline struct cpumask *group_balance_mask(struct sched_group *sg) 1950 { 1951 return to_cpumask(sg->sgc->cpumask); 1952 } 1953 1954 extern int group_balance_cpu(struct sched_group *sg); 1955 1956 #ifdef CONFIG_SCHED_DEBUG 1957 void update_sched_domain_debugfs(void); 1958 void dirty_sched_domain_sysctl(int cpu); 1959 #else 1960 static inline void update_sched_domain_debugfs(void) 1961 { 1962 } 1963 static inline void dirty_sched_domain_sysctl(int cpu) 1964 { 1965 } 1966 #endif 1967 1968 extern int sched_update_scaling(void); 1969 1970 static inline const struct cpumask *task_user_cpus(struct task_struct *p) 1971 { 1972 if (!p->user_cpus_ptr) 1973 return cpu_possible_mask; /* &init_task.cpus_mask */ 1974 return p->user_cpus_ptr; 1975 } 1976 #endif /* CONFIG_SMP */ 1977 1978 #include "stats.h" 1979 1980 #if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS) 1981 1982 extern void __sched_core_account_forceidle(struct rq *rq); 1983 1984 static inline void sched_core_account_forceidle(struct rq *rq) 1985 { 1986 if (schedstat_enabled()) 1987 __sched_core_account_forceidle(rq); 1988 } 1989 1990 extern void __sched_core_tick(struct rq *rq); 1991 1992 static inline void sched_core_tick(struct rq *rq) 1993 { 1994 if (sched_core_enabled(rq) && schedstat_enabled()) 1995 __sched_core_tick(rq); 1996 } 1997 1998 #else 1999 2000 static inline void sched_core_account_forceidle(struct rq *rq) {} 2001 2002 static inline void sched_core_tick(struct rq *rq) {} 2003 2004 #endif /* CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS */ 2005 2006 #ifdef CONFIG_CGROUP_SCHED 2007 2008 /* 2009 * Return the group to which this tasks belongs. 2010 * 2011 * We cannot use task_css() and friends because the cgroup subsystem 2012 * changes that value before the cgroup_subsys::attach() method is called, 2013 * therefore we cannot pin it and might observe the wrong value. 2014 * 2015 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup 2016 * core changes this before calling sched_move_task(). 2017 * 2018 * Instead we use a 'copy' which is updated from sched_move_task() while 2019 * holding both task_struct::pi_lock and rq::lock. 2020 */ 2021 static inline struct task_group *task_group(struct task_struct *p) 2022 { 2023 return p->sched_task_group; 2024 } 2025 2026 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ 2027 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) 2028 { 2029 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED) 2030 struct task_group *tg = task_group(p); 2031 #endif 2032 2033 #ifdef CONFIG_FAIR_GROUP_SCHED 2034 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]); 2035 p->se.cfs_rq = tg->cfs_rq[cpu]; 2036 p->se.parent = tg->se[cpu]; 2037 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0; 2038 #endif 2039 2040 #ifdef CONFIG_RT_GROUP_SCHED 2041 p->rt.rt_rq = tg->rt_rq[cpu]; 2042 p->rt.parent = tg->rt_se[cpu]; 2043 #endif 2044 } 2045 2046 #else /* CONFIG_CGROUP_SCHED */ 2047 2048 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } 2049 static inline struct task_group *task_group(struct task_struct *p) 2050 { 2051 return NULL; 2052 } 2053 2054 #endif /* CONFIG_CGROUP_SCHED */ 2055 2056 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) 2057 { 2058 set_task_rq(p, cpu); 2059 #ifdef CONFIG_SMP 2060 /* 2061 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be 2062 * successfully executed on another CPU. We must ensure that updates of 2063 * per-task data have been completed by this moment. 2064 */ 2065 smp_wmb(); 2066 WRITE_ONCE(task_thread_info(p)->cpu, cpu); 2067 p->wake_cpu = cpu; 2068 #endif 2069 } 2070 2071 /* 2072 * Tunables that become constants when CONFIG_SCHED_DEBUG is off: 2073 */ 2074 #ifdef CONFIG_SCHED_DEBUG 2075 # define const_debug __read_mostly 2076 #else 2077 # define const_debug const 2078 #endif 2079 2080 #define SCHED_FEAT(name, enabled) \ 2081 __SCHED_FEAT_##name , 2082 2083 enum { 2084 #include "features.h" 2085 __SCHED_FEAT_NR, 2086 }; 2087 2088 #undef SCHED_FEAT 2089 2090 #ifdef CONFIG_SCHED_DEBUG 2091 2092 /* 2093 * To support run-time toggling of sched features, all the translation units 2094 * (but core.c) reference the sysctl_sched_features defined in core.c. 2095 */ 2096 extern const_debug unsigned int sysctl_sched_features; 2097 2098 #ifdef CONFIG_JUMP_LABEL 2099 #define SCHED_FEAT(name, enabled) \ 2100 static __always_inline bool static_branch_##name(struct static_key *key) \ 2101 { \ 2102 return static_key_##enabled(key); \ 2103 } 2104 2105 #include "features.h" 2106 #undef SCHED_FEAT 2107 2108 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR]; 2109 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x])) 2110 2111 #else /* !CONFIG_JUMP_LABEL */ 2112 2113 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 2114 2115 #endif /* CONFIG_JUMP_LABEL */ 2116 2117 #else /* !SCHED_DEBUG */ 2118 2119 /* 2120 * Each translation unit has its own copy of sysctl_sched_features to allow 2121 * constants propagation at compile time and compiler optimization based on 2122 * features default. 2123 */ 2124 #define SCHED_FEAT(name, enabled) \ 2125 (1UL << __SCHED_FEAT_##name) * enabled | 2126 static const_debug __maybe_unused unsigned int sysctl_sched_features = 2127 #include "features.h" 2128 0; 2129 #undef SCHED_FEAT 2130 2131 #define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) 2132 2133 #endif /* SCHED_DEBUG */ 2134 2135 extern struct static_key_false sched_numa_balancing; 2136 extern struct static_key_false sched_schedstats; 2137 2138 static inline u64 global_rt_period(void) 2139 { 2140 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; 2141 } 2142 2143 static inline u64 global_rt_runtime(void) 2144 { 2145 if (sysctl_sched_rt_runtime < 0) 2146 return RUNTIME_INF; 2147 2148 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; 2149 } 2150 2151 static inline int task_current(struct rq *rq, struct task_struct *p) 2152 { 2153 return rq->curr == p; 2154 } 2155 2156 static inline int task_on_cpu(struct rq *rq, struct task_struct *p) 2157 { 2158 #ifdef CONFIG_SMP 2159 return p->on_cpu; 2160 #else 2161 return task_current(rq, p); 2162 #endif 2163 } 2164 2165 static inline int task_on_rq_queued(struct task_struct *p) 2166 { 2167 return p->on_rq == TASK_ON_RQ_QUEUED; 2168 } 2169 2170 static inline int task_on_rq_migrating(struct task_struct *p) 2171 { 2172 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING; 2173 } 2174 2175 /* Wake flags. The first three directly map to some SD flag value */ 2176 #define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */ 2177 #define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */ 2178 #define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */ 2179 2180 #define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */ 2181 #define WF_MIGRATED 0x20 /* Internal use, task got migrated */ 2182 #define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */ 2183 2184 #ifdef CONFIG_SMP 2185 static_assert(WF_EXEC == SD_BALANCE_EXEC); 2186 static_assert(WF_FORK == SD_BALANCE_FORK); 2187 static_assert(WF_TTWU == SD_BALANCE_WAKE); 2188 #endif 2189 2190 /* 2191 * To aid in avoiding the subversion of "niceness" due to uneven distribution 2192 * of tasks with abnormal "nice" values across CPUs the contribution that 2193 * each task makes to its run queue's load is weighted according to its 2194 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a 2195 * scaled version of the new time slice allocation that they receive on time 2196 * slice expiry etc. 2197 */ 2198 2199 #define WEIGHT_IDLEPRIO 3 2200 #define WMULT_IDLEPRIO 1431655765 2201 2202 extern const int sched_prio_to_weight[40]; 2203 extern const u32 sched_prio_to_wmult[40]; 2204 2205 /* 2206 * {de,en}queue flags: 2207 * 2208 * DEQUEUE_SLEEP - task is no longer runnable 2209 * ENQUEUE_WAKEUP - task just became runnable 2210 * 2211 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks 2212 * are in a known state which allows modification. Such pairs 2213 * should preserve as much state as possible. 2214 * 2215 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location 2216 * in the runqueue. 2217 * 2218 * NOCLOCK - skip the update_rq_clock() (avoids double updates) 2219 * 2220 * MIGRATION - p->on_rq == TASK_ON_RQ_MIGRATING (used for DEADLINE) 2221 * 2222 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified) 2223 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline) 2224 * ENQUEUE_MIGRATED - the task was migrated during wakeup 2225 * 2226 */ 2227 2228 #define DEQUEUE_SLEEP 0x01 2229 #define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */ 2230 #define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */ 2231 #define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */ 2232 #define DEQUEUE_MIGRATING 0x100 /* Matches ENQUEUE_MIGRATING */ 2233 2234 #define ENQUEUE_WAKEUP 0x01 2235 #define ENQUEUE_RESTORE 0x02 2236 #define ENQUEUE_MOVE 0x04 2237 #define ENQUEUE_NOCLOCK 0x08 2238 2239 #define ENQUEUE_HEAD 0x10 2240 #define ENQUEUE_REPLENISH 0x20 2241 #ifdef CONFIG_SMP 2242 #define ENQUEUE_MIGRATED 0x40 2243 #else 2244 #define ENQUEUE_MIGRATED 0x00 2245 #endif 2246 #define ENQUEUE_INITIAL 0x80 2247 #define ENQUEUE_MIGRATING 0x100 2248 2249 #define RETRY_TASK ((void *)-1UL) 2250 2251 struct affinity_context { 2252 const struct cpumask *new_mask; 2253 struct cpumask *user_mask; 2254 unsigned int flags; 2255 }; 2256 2257 extern s64 update_curr_common(struct rq *rq); 2258 2259 struct sched_class { 2260 2261 #ifdef CONFIG_UCLAMP_TASK 2262 int uclamp_enabled; 2263 #endif 2264 2265 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags); 2266 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags); 2267 void (*yield_task) (struct rq *rq); 2268 bool (*yield_to_task)(struct rq *rq, struct task_struct *p); 2269 2270 void (*wakeup_preempt)(struct rq *rq, struct task_struct *p, int flags); 2271 2272 struct task_struct *(*pick_next_task)(struct rq *rq); 2273 2274 void (*put_prev_task)(struct rq *rq, struct task_struct *p); 2275 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first); 2276 2277 #ifdef CONFIG_SMP 2278 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf); 2279 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags); 2280 2281 struct task_struct * (*pick_task)(struct rq *rq); 2282 2283 void (*migrate_task_rq)(struct task_struct *p, int new_cpu); 2284 2285 void (*task_woken)(struct rq *this_rq, struct task_struct *task); 2286 2287 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx); 2288 2289 void (*rq_online)(struct rq *rq); 2290 void (*rq_offline)(struct rq *rq); 2291 2292 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq); 2293 #endif 2294 2295 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued); 2296 void (*task_fork)(struct task_struct *p); 2297 void (*task_dead)(struct task_struct *p); 2298 2299 /* 2300 * The switched_from() call is allowed to drop rq->lock, therefore we 2301 * cannot assume the switched_from/switched_to pair is serialized by 2302 * rq->lock. They are however serialized by p->pi_lock. 2303 */ 2304 void (*switched_from)(struct rq *this_rq, struct task_struct *task); 2305 void (*switched_to) (struct rq *this_rq, struct task_struct *task); 2306 void (*prio_changed) (struct rq *this_rq, struct task_struct *task, 2307 int oldprio); 2308 2309 unsigned int (*get_rr_interval)(struct rq *rq, 2310 struct task_struct *task); 2311 2312 void (*update_curr)(struct rq *rq); 2313 2314 #ifdef CONFIG_FAIR_GROUP_SCHED 2315 void (*task_change_group)(struct task_struct *p); 2316 #endif 2317 2318 #ifdef CONFIG_SCHED_CORE 2319 int (*task_is_throttled)(struct task_struct *p, int cpu); 2320 #endif 2321 }; 2322 2323 static inline void put_prev_task(struct rq *rq, struct task_struct *prev) 2324 { 2325 WARN_ON_ONCE(rq->curr != prev); 2326 prev->sched_class->put_prev_task(rq, prev); 2327 } 2328 2329 static inline void set_next_task(struct rq *rq, struct task_struct *next) 2330 { 2331 next->sched_class->set_next_task(rq, next, false); 2332 } 2333 2334 2335 /* 2336 * Helper to define a sched_class instance; each one is placed in a separate 2337 * section which is ordered by the linker script: 2338 * 2339 * include/asm-generic/vmlinux.lds.h 2340 * 2341 * *CAREFUL* they are laid out in *REVERSE* order!!! 2342 * 2343 * Also enforce alignment on the instance, not the type, to guarantee layout. 2344 */ 2345 #define DEFINE_SCHED_CLASS(name) \ 2346 const struct sched_class name##_sched_class \ 2347 __aligned(__alignof__(struct sched_class)) \ 2348 __section("__" #name "_sched_class") 2349 2350 /* Defined in include/asm-generic/vmlinux.lds.h */ 2351 extern struct sched_class __sched_class_highest[]; 2352 extern struct sched_class __sched_class_lowest[]; 2353 2354 #define for_class_range(class, _from, _to) \ 2355 for (class = (_from); class < (_to); class++) 2356 2357 #define for_each_class(class) \ 2358 for_class_range(class, __sched_class_highest, __sched_class_lowest) 2359 2360 #define sched_class_above(_a, _b) ((_a) < (_b)) 2361 2362 extern const struct sched_class stop_sched_class; 2363 extern const struct sched_class dl_sched_class; 2364 extern const struct sched_class rt_sched_class; 2365 extern const struct sched_class fair_sched_class; 2366 extern const struct sched_class idle_sched_class; 2367 2368 static inline bool sched_stop_runnable(struct rq *rq) 2369 { 2370 return rq->stop && task_on_rq_queued(rq->stop); 2371 } 2372 2373 static inline bool sched_dl_runnable(struct rq *rq) 2374 { 2375 return rq->dl.dl_nr_running > 0; 2376 } 2377 2378 static inline bool sched_rt_runnable(struct rq *rq) 2379 { 2380 return rq->rt.rt_queued > 0; 2381 } 2382 2383 static inline bool sched_fair_runnable(struct rq *rq) 2384 { 2385 return rq->cfs.nr_running > 0; 2386 } 2387 2388 extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf); 2389 extern struct task_struct *pick_next_task_idle(struct rq *rq); 2390 2391 #define SCA_CHECK 0x01 2392 #define SCA_MIGRATE_DISABLE 0x02 2393 #define SCA_MIGRATE_ENABLE 0x04 2394 #define SCA_USER 0x08 2395 2396 #ifdef CONFIG_SMP 2397 2398 extern void update_group_capacity(struct sched_domain *sd, int cpu); 2399 2400 extern void trigger_load_balance(struct rq *rq); 2401 2402 extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx); 2403 2404 static inline struct task_struct *get_push_task(struct rq *rq) 2405 { 2406 struct task_struct *p = rq->curr; 2407 2408 lockdep_assert_rq_held(rq); 2409 2410 if (rq->push_busy) 2411 return NULL; 2412 2413 if (p->nr_cpus_allowed == 1) 2414 return NULL; 2415 2416 if (p->migration_disabled) 2417 return NULL; 2418 2419 rq->push_busy = true; 2420 return get_task_struct(p); 2421 } 2422 2423 extern int push_cpu_stop(void *arg); 2424 2425 #endif 2426 2427 #ifdef CONFIG_CPU_IDLE 2428 static inline void idle_set_state(struct rq *rq, 2429 struct cpuidle_state *idle_state) 2430 { 2431 rq->idle_state = idle_state; 2432 } 2433 2434 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 2435 { 2436 SCHED_WARN_ON(!rcu_read_lock_held()); 2437 2438 return rq->idle_state; 2439 } 2440 #else 2441 static inline void idle_set_state(struct rq *rq, 2442 struct cpuidle_state *idle_state) 2443 { 2444 } 2445 2446 static inline struct cpuidle_state *idle_get_state(struct rq *rq) 2447 { 2448 return NULL; 2449 } 2450 #endif 2451 2452 extern void schedule_idle(void); 2453 asmlinkage void schedule_user(void); 2454 2455 extern void sysrq_sched_debug_show(void); 2456 extern void sched_init_granularity(void); 2457 extern void update_max_interval(void); 2458 2459 extern void init_sched_dl_class(void); 2460 extern void init_sched_rt_class(void); 2461 extern void init_sched_fair_class(void); 2462 2463 extern void reweight_task(struct task_struct *p, int prio); 2464 2465 extern void resched_curr(struct rq *rq); 2466 extern void resched_cpu(int cpu); 2467 2468 extern struct rt_bandwidth def_rt_bandwidth; 2469 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime); 2470 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); 2471 2472 extern void init_dl_entity(struct sched_dl_entity *dl_se); 2473 2474 #define BW_SHIFT 20 2475 #define BW_UNIT (1 << BW_SHIFT) 2476 #define RATIO_SHIFT 8 2477 #define MAX_BW_BITS (64 - BW_SHIFT) 2478 #define MAX_BW ((1ULL << MAX_BW_BITS) - 1) 2479 unsigned long to_ratio(u64 period, u64 runtime); 2480 2481 extern void init_entity_runnable_average(struct sched_entity *se); 2482 extern void post_init_entity_util_avg(struct task_struct *p); 2483 2484 #ifdef CONFIG_NO_HZ_FULL 2485 extern bool sched_can_stop_tick(struct rq *rq); 2486 extern int __init sched_tick_offload_init(void); 2487 2488 /* 2489 * Tick may be needed by tasks in the runqueue depending on their policy and 2490 * requirements. If tick is needed, lets send the target an IPI to kick it out of 2491 * nohz mode if necessary. 2492 */ 2493 static inline void sched_update_tick_dependency(struct rq *rq) 2494 { 2495 int cpu = cpu_of(rq); 2496 2497 if (!tick_nohz_full_cpu(cpu)) 2498 return; 2499 2500 if (sched_can_stop_tick(rq)) 2501 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED); 2502 else 2503 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED); 2504 } 2505 #else 2506 static inline int sched_tick_offload_init(void) { return 0; } 2507 static inline void sched_update_tick_dependency(struct rq *rq) { } 2508 #endif 2509 2510 static inline void add_nr_running(struct rq *rq, unsigned count) 2511 { 2512 unsigned prev_nr = rq->nr_running; 2513 2514 rq->nr_running = prev_nr + count; 2515 if (trace_sched_update_nr_running_tp_enabled()) { 2516 call_trace_sched_update_nr_running(rq, count); 2517 } 2518 2519 #ifdef CONFIG_SMP 2520 if (prev_nr < 2 && rq->nr_running >= 2) { 2521 if (!READ_ONCE(rq->rd->overload)) 2522 WRITE_ONCE(rq->rd->overload, 1); 2523 } 2524 #endif 2525 2526 sched_update_tick_dependency(rq); 2527 } 2528 2529 static inline void sub_nr_running(struct rq *rq, unsigned count) 2530 { 2531 rq->nr_running -= count; 2532 if (trace_sched_update_nr_running_tp_enabled()) { 2533 call_trace_sched_update_nr_running(rq, -count); 2534 } 2535 2536 /* Check if we still need preemption */ 2537 sched_update_tick_dependency(rq); 2538 } 2539 2540 extern void activate_task(struct rq *rq, struct task_struct *p, int flags); 2541 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags); 2542 2543 extern void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags); 2544 2545 #ifdef CONFIG_PREEMPT_RT 2546 #define SCHED_NR_MIGRATE_BREAK 8 2547 #else 2548 #define SCHED_NR_MIGRATE_BREAK 32 2549 #endif 2550 2551 extern const_debug unsigned int sysctl_sched_nr_migrate; 2552 extern const_debug unsigned int sysctl_sched_migration_cost; 2553 2554 extern unsigned int sysctl_sched_base_slice; 2555 2556 #ifdef CONFIG_SCHED_DEBUG 2557 extern int sysctl_resched_latency_warn_ms; 2558 extern int sysctl_resched_latency_warn_once; 2559 2560 extern unsigned int sysctl_sched_tunable_scaling; 2561 2562 extern unsigned int sysctl_numa_balancing_scan_delay; 2563 extern unsigned int sysctl_numa_balancing_scan_period_min; 2564 extern unsigned int sysctl_numa_balancing_scan_period_max; 2565 extern unsigned int sysctl_numa_balancing_scan_size; 2566 extern unsigned int sysctl_numa_balancing_hot_threshold; 2567 #endif 2568 2569 #ifdef CONFIG_SCHED_HRTICK 2570 2571 /* 2572 * Use hrtick when: 2573 * - enabled by features 2574 * - hrtimer is actually high res 2575 */ 2576 static inline int hrtick_enabled(struct rq *rq) 2577 { 2578 if (!cpu_active(cpu_of(rq))) 2579 return 0; 2580 return hrtimer_is_hres_active(&rq->hrtick_timer); 2581 } 2582 2583 static inline int hrtick_enabled_fair(struct rq *rq) 2584 { 2585 if (!sched_feat(HRTICK)) 2586 return 0; 2587 return hrtick_enabled(rq); 2588 } 2589 2590 static inline int hrtick_enabled_dl(struct rq *rq) 2591 { 2592 if (!sched_feat(HRTICK_DL)) 2593 return 0; 2594 return hrtick_enabled(rq); 2595 } 2596 2597 void hrtick_start(struct rq *rq, u64 delay); 2598 2599 #else 2600 2601 static inline int hrtick_enabled_fair(struct rq *rq) 2602 { 2603 return 0; 2604 } 2605 2606 static inline int hrtick_enabled_dl(struct rq *rq) 2607 { 2608 return 0; 2609 } 2610 2611 static inline int hrtick_enabled(struct rq *rq) 2612 { 2613 return 0; 2614 } 2615 2616 #endif /* CONFIG_SCHED_HRTICK */ 2617 2618 #ifndef arch_scale_freq_tick 2619 static __always_inline 2620 void arch_scale_freq_tick(void) 2621 { 2622 } 2623 #endif 2624 2625 #ifndef arch_scale_freq_capacity 2626 /** 2627 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU. 2628 * @cpu: the CPU in question. 2629 * 2630 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e. 2631 * 2632 * f_curr 2633 * ------ * SCHED_CAPACITY_SCALE 2634 * f_max 2635 */ 2636 static __always_inline 2637 unsigned long arch_scale_freq_capacity(int cpu) 2638 { 2639 return SCHED_CAPACITY_SCALE; 2640 } 2641 #endif 2642 2643 #ifdef CONFIG_SCHED_DEBUG 2644 /* 2645 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to 2646 * acquire rq lock instead of rq_lock(). So at the end of these two functions 2647 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of 2648 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning. 2649 */ 2650 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) 2651 { 2652 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP); 2653 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */ 2654 #ifdef CONFIG_SMP 2655 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP); 2656 #endif 2657 } 2658 #else 2659 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {} 2660 #endif 2661 2662 #define DEFINE_LOCK_GUARD_2(name, type, _lock, _unlock, ...) \ 2663 __DEFINE_UNLOCK_GUARD(name, type, _unlock, type *lock2; __VA_ARGS__) \ 2664 static inline class_##name##_t class_##name##_constructor(type *lock, type *lock2) \ 2665 { class_##name##_t _t = { .lock = lock, .lock2 = lock2 }, *_T = &_t; \ 2666 _lock; return _t; } 2667 2668 #ifdef CONFIG_SMP 2669 2670 static inline bool rq_order_less(struct rq *rq1, struct rq *rq2) 2671 { 2672 #ifdef CONFIG_SCHED_CORE 2673 /* 2674 * In order to not have {0,2},{1,3} turn into into an AB-BA, 2675 * order by core-id first and cpu-id second. 2676 * 2677 * Notably: 2678 * 2679 * double_rq_lock(0,3); will take core-0, core-1 lock 2680 * double_rq_lock(1,2); will take core-1, core-0 lock 2681 * 2682 * when only cpu-id is considered. 2683 */ 2684 if (rq1->core->cpu < rq2->core->cpu) 2685 return true; 2686 if (rq1->core->cpu > rq2->core->cpu) 2687 return false; 2688 2689 /* 2690 * __sched_core_flip() relies on SMT having cpu-id lock order. 2691 */ 2692 #endif 2693 return rq1->cpu < rq2->cpu; 2694 } 2695 2696 extern void double_rq_lock(struct rq *rq1, struct rq *rq2); 2697 2698 #ifdef CONFIG_PREEMPTION 2699 2700 /* 2701 * fair double_lock_balance: Safely acquires both rq->locks in a fair 2702 * way at the expense of forcing extra atomic operations in all 2703 * invocations. This assures that the double_lock is acquired using the 2704 * same underlying policy as the spinlock_t on this architecture, which 2705 * reduces latency compared to the unfair variant below. However, it 2706 * also adds more overhead and therefore may reduce throughput. 2707 */ 2708 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 2709 __releases(this_rq->lock) 2710 __acquires(busiest->lock) 2711 __acquires(this_rq->lock) 2712 { 2713 raw_spin_rq_unlock(this_rq); 2714 double_rq_lock(this_rq, busiest); 2715 2716 return 1; 2717 } 2718 2719 #else 2720 /* 2721 * Unfair double_lock_balance: Optimizes throughput at the expense of 2722 * latency by eliminating extra atomic operations when the locks are 2723 * already in proper order on entry. This favors lower CPU-ids and will 2724 * grant the double lock to lower CPUs over higher ids under contention, 2725 * regardless of entry order into the function. 2726 */ 2727 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) 2728 __releases(this_rq->lock) 2729 __acquires(busiest->lock) 2730 __acquires(this_rq->lock) 2731 { 2732 if (__rq_lockp(this_rq) == __rq_lockp(busiest) || 2733 likely(raw_spin_rq_trylock(busiest))) { 2734 double_rq_clock_clear_update(this_rq, busiest); 2735 return 0; 2736 } 2737 2738 if (rq_order_less(this_rq, busiest)) { 2739 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING); 2740 double_rq_clock_clear_update(this_rq, busiest); 2741 return 0; 2742 } 2743 2744 raw_spin_rq_unlock(this_rq); 2745 double_rq_lock(this_rq, busiest); 2746 2747 return 1; 2748 } 2749 2750 #endif /* CONFIG_PREEMPTION */ 2751 2752 /* 2753 * double_lock_balance - lock the busiest runqueue, this_rq is locked already. 2754 */ 2755 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest) 2756 { 2757 lockdep_assert_irqs_disabled(); 2758 2759 return _double_lock_balance(this_rq, busiest); 2760 } 2761 2762 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest) 2763 __releases(busiest->lock) 2764 { 2765 if (__rq_lockp(this_rq) != __rq_lockp(busiest)) 2766 raw_spin_rq_unlock(busiest); 2767 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_); 2768 } 2769 2770 static inline void double_lock(spinlock_t *l1, spinlock_t *l2) 2771 { 2772 if (l1 > l2) 2773 swap(l1, l2); 2774 2775 spin_lock(l1); 2776 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 2777 } 2778 2779 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2) 2780 { 2781 if (l1 > l2) 2782 swap(l1, l2); 2783 2784 spin_lock_irq(l1); 2785 spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 2786 } 2787 2788 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2) 2789 { 2790 if (l1 > l2) 2791 swap(l1, l2); 2792 2793 raw_spin_lock(l1); 2794 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING); 2795 } 2796 2797 static inline void double_raw_unlock(raw_spinlock_t *l1, raw_spinlock_t *l2) 2798 { 2799 raw_spin_unlock(l1); 2800 raw_spin_unlock(l2); 2801 } 2802 2803 DEFINE_LOCK_GUARD_2(double_raw_spinlock, raw_spinlock_t, 2804 double_raw_lock(_T->lock, _T->lock2), 2805 double_raw_unlock(_T->lock, _T->lock2)) 2806 2807 /* 2808 * double_rq_unlock - safely unlock two runqueues 2809 * 2810 * Note this does not restore interrupts like task_rq_unlock, 2811 * you need to do so manually after calling. 2812 */ 2813 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 2814 __releases(rq1->lock) 2815 __releases(rq2->lock) 2816 { 2817 if (__rq_lockp(rq1) != __rq_lockp(rq2)) 2818 raw_spin_rq_unlock(rq2); 2819 else 2820 __release(rq2->lock); 2821 raw_spin_rq_unlock(rq1); 2822 } 2823 2824 extern void set_rq_online (struct rq *rq); 2825 extern void set_rq_offline(struct rq *rq); 2826 extern bool sched_smp_initialized; 2827 2828 #else /* CONFIG_SMP */ 2829 2830 /* 2831 * double_rq_lock - safely lock two runqueues 2832 * 2833 * Note this does not disable interrupts like task_rq_lock, 2834 * you need to do so manually before calling. 2835 */ 2836 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) 2837 __acquires(rq1->lock) 2838 __acquires(rq2->lock) 2839 { 2840 WARN_ON_ONCE(!irqs_disabled()); 2841 WARN_ON_ONCE(rq1 != rq2); 2842 raw_spin_rq_lock(rq1); 2843 __acquire(rq2->lock); /* Fake it out ;) */ 2844 double_rq_clock_clear_update(rq1, rq2); 2845 } 2846 2847 /* 2848 * double_rq_unlock - safely unlock two runqueues 2849 * 2850 * Note this does not restore interrupts like task_rq_unlock, 2851 * you need to do so manually after calling. 2852 */ 2853 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2) 2854 __releases(rq1->lock) 2855 __releases(rq2->lock) 2856 { 2857 WARN_ON_ONCE(rq1 != rq2); 2858 raw_spin_rq_unlock(rq1); 2859 __release(rq2->lock); 2860 } 2861 2862 #endif 2863 2864 DEFINE_LOCK_GUARD_2(double_rq_lock, struct rq, 2865 double_rq_lock(_T->lock, _T->lock2), 2866 double_rq_unlock(_T->lock, _T->lock2)) 2867 2868 extern struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq); 2869 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq); 2870 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq); 2871 2872 #ifdef CONFIG_SCHED_DEBUG 2873 extern bool sched_debug_verbose; 2874 2875 extern void print_cfs_stats(struct seq_file *m, int cpu); 2876 extern void print_rt_stats(struct seq_file *m, int cpu); 2877 extern void print_dl_stats(struct seq_file *m, int cpu); 2878 extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq); 2879 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq); 2880 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); 2881 2882 extern void resched_latency_warn(int cpu, u64 latency); 2883 #ifdef CONFIG_NUMA_BALANCING 2884 extern void 2885 show_numa_stats(struct task_struct *p, struct seq_file *m); 2886 extern void 2887 print_numa_stats(struct seq_file *m, int node, unsigned long tsf, 2888 unsigned long tpf, unsigned long gsf, unsigned long gpf); 2889 #endif /* CONFIG_NUMA_BALANCING */ 2890 #else 2891 static inline void resched_latency_warn(int cpu, u64 latency) {} 2892 #endif /* CONFIG_SCHED_DEBUG */ 2893 2894 extern void init_cfs_rq(struct cfs_rq *cfs_rq); 2895 extern void init_rt_rq(struct rt_rq *rt_rq); 2896 extern void init_dl_rq(struct dl_rq *dl_rq); 2897 2898 extern void cfs_bandwidth_usage_inc(void); 2899 extern void cfs_bandwidth_usage_dec(void); 2900 2901 #ifdef CONFIG_NO_HZ_COMMON 2902 #define NOHZ_BALANCE_KICK_BIT 0 2903 #define NOHZ_STATS_KICK_BIT 1 2904 #define NOHZ_NEWILB_KICK_BIT 2 2905 #define NOHZ_NEXT_KICK_BIT 3 2906 2907 /* Run rebalance_domains() */ 2908 #define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT) 2909 /* Update blocked load */ 2910 #define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT) 2911 /* Update blocked load when entering idle */ 2912 #define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT) 2913 /* Update nohz.next_balance */ 2914 #define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT) 2915 2916 #define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK) 2917 2918 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags) 2919 2920 extern void nohz_balance_exit_idle(struct rq *rq); 2921 #else 2922 static inline void nohz_balance_exit_idle(struct rq *rq) { } 2923 #endif 2924 2925 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON) 2926 extern void nohz_run_idle_balance(int cpu); 2927 #else 2928 static inline void nohz_run_idle_balance(int cpu) { } 2929 #endif 2930 2931 #ifdef CONFIG_IRQ_TIME_ACCOUNTING 2932 struct irqtime { 2933 u64 total; 2934 u64 tick_delta; 2935 u64 irq_start_time; 2936 struct u64_stats_sync sync; 2937 }; 2938 2939 DECLARE_PER_CPU(struct irqtime, cpu_irqtime); 2940 2941 /* 2942 * Returns the irqtime minus the softirq time computed by ksoftirqd. 2943 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime 2944 * and never move forward. 2945 */ 2946 static inline u64 irq_time_read(int cpu) 2947 { 2948 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu); 2949 unsigned int seq; 2950 u64 total; 2951 2952 do { 2953 seq = __u64_stats_fetch_begin(&irqtime->sync); 2954 total = irqtime->total; 2955 } while (__u64_stats_fetch_retry(&irqtime->sync, seq)); 2956 2957 return total; 2958 } 2959 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 2960 2961 #ifdef CONFIG_CPU_FREQ 2962 DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data); 2963 2964 /** 2965 * cpufreq_update_util - Take a note about CPU utilization changes. 2966 * @rq: Runqueue to carry out the update for. 2967 * @flags: Update reason flags. 2968 * 2969 * This function is called by the scheduler on the CPU whose utilization is 2970 * being updated. 2971 * 2972 * It can only be called from RCU-sched read-side critical sections. 2973 * 2974 * The way cpufreq is currently arranged requires it to evaluate the CPU 2975 * performance state (frequency/voltage) on a regular basis to prevent it from 2976 * being stuck in a completely inadequate performance level for too long. 2977 * That is not guaranteed to happen if the updates are only triggered from CFS 2978 * and DL, though, because they may not be coming in if only RT tasks are 2979 * active all the time (or there are RT tasks only). 2980 * 2981 * As a workaround for that issue, this function is called periodically by the 2982 * RT sched class to trigger extra cpufreq updates to prevent it from stalling, 2983 * but that really is a band-aid. Going forward it should be replaced with 2984 * solutions targeted more specifically at RT tasks. 2985 */ 2986 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) 2987 { 2988 struct update_util_data *data; 2989 2990 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data, 2991 cpu_of(rq))); 2992 if (data) 2993 data->func(data, rq_clock(rq), flags); 2994 } 2995 #else 2996 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {} 2997 #endif /* CONFIG_CPU_FREQ */ 2998 2999 #ifdef arch_scale_freq_capacity 3000 # ifndef arch_scale_freq_invariant 3001 # define arch_scale_freq_invariant() true 3002 # endif 3003 #else 3004 # define arch_scale_freq_invariant() false 3005 #endif 3006 3007 #ifdef CONFIG_SMP 3008 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs, 3009 unsigned long *min, 3010 unsigned long *max); 3011 3012 unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual, 3013 unsigned long min, 3014 unsigned long max); 3015 3016 3017 /* 3018 * Verify the fitness of task @p to run on @cpu taking into account the 3019 * CPU original capacity and the runtime/deadline ratio of the task. 3020 * 3021 * The function will return true if the original capacity of @cpu is 3022 * greater than or equal to task's deadline density right shifted by 3023 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise. 3024 */ 3025 static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu) 3026 { 3027 unsigned long cap = arch_scale_cpu_capacity(cpu); 3028 3029 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT); 3030 } 3031 3032 static inline unsigned long cpu_bw_dl(struct rq *rq) 3033 { 3034 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT; 3035 } 3036 3037 static inline unsigned long cpu_util_dl(struct rq *rq) 3038 { 3039 return READ_ONCE(rq->avg_dl.util_avg); 3040 } 3041 3042 3043 extern unsigned long cpu_util_cfs(int cpu); 3044 extern unsigned long cpu_util_cfs_boost(int cpu); 3045 3046 static inline unsigned long cpu_util_rt(struct rq *rq) 3047 { 3048 return READ_ONCE(rq->avg_rt.util_avg); 3049 } 3050 #endif 3051 3052 #ifdef CONFIG_UCLAMP_TASK 3053 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id); 3054 3055 static inline unsigned long uclamp_rq_get(struct rq *rq, 3056 enum uclamp_id clamp_id) 3057 { 3058 return READ_ONCE(rq->uclamp[clamp_id].value); 3059 } 3060 3061 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id, 3062 unsigned int value) 3063 { 3064 WRITE_ONCE(rq->uclamp[clamp_id].value, value); 3065 } 3066 3067 static inline bool uclamp_rq_is_idle(struct rq *rq) 3068 { 3069 return rq->uclamp_flags & UCLAMP_FLAG_IDLE; 3070 } 3071 3072 /* Is the rq being capped/throttled by uclamp_max? */ 3073 static inline bool uclamp_rq_is_capped(struct rq *rq) 3074 { 3075 unsigned long rq_util; 3076 unsigned long max_util; 3077 3078 if (!static_branch_likely(&sched_uclamp_used)) 3079 return false; 3080 3081 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq); 3082 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value); 3083 3084 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util; 3085 } 3086 3087 /* 3088 * When uclamp is compiled in, the aggregation at rq level is 'turned off' 3089 * by default in the fast path and only gets turned on once userspace performs 3090 * an operation that requires it. 3091 * 3092 * Returns true if userspace opted-in to use uclamp and aggregation at rq level 3093 * hence is active. 3094 */ 3095 static inline bool uclamp_is_used(void) 3096 { 3097 return static_branch_likely(&sched_uclamp_used); 3098 } 3099 #else /* CONFIG_UCLAMP_TASK */ 3100 static inline unsigned long uclamp_eff_value(struct task_struct *p, 3101 enum uclamp_id clamp_id) 3102 { 3103 if (clamp_id == UCLAMP_MIN) 3104 return 0; 3105 3106 return SCHED_CAPACITY_SCALE; 3107 } 3108 3109 static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; } 3110 3111 static inline bool uclamp_is_used(void) 3112 { 3113 return false; 3114 } 3115 3116 static inline unsigned long uclamp_rq_get(struct rq *rq, 3117 enum uclamp_id clamp_id) 3118 { 3119 if (clamp_id == UCLAMP_MIN) 3120 return 0; 3121 3122 return SCHED_CAPACITY_SCALE; 3123 } 3124 3125 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id, 3126 unsigned int value) 3127 { 3128 } 3129 3130 static inline bool uclamp_rq_is_idle(struct rq *rq) 3131 { 3132 return false; 3133 } 3134 #endif /* CONFIG_UCLAMP_TASK */ 3135 3136 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 3137 static inline unsigned long cpu_util_irq(struct rq *rq) 3138 { 3139 return READ_ONCE(rq->avg_irq.util_avg); 3140 } 3141 3142 static inline 3143 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max) 3144 { 3145 util *= (max - irq); 3146 util /= max; 3147 3148 return util; 3149 3150 } 3151 #else 3152 static inline unsigned long cpu_util_irq(struct rq *rq) 3153 { 3154 return 0; 3155 } 3156 3157 static inline 3158 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max) 3159 { 3160 return util; 3161 } 3162 #endif 3163 3164 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL) 3165 3166 #define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus))) 3167 3168 DECLARE_STATIC_KEY_FALSE(sched_energy_present); 3169 3170 static inline bool sched_energy_enabled(void) 3171 { 3172 return static_branch_unlikely(&sched_energy_present); 3173 } 3174 3175 extern struct cpufreq_governor schedutil_gov; 3176 3177 #else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */ 3178 3179 #define perf_domain_span(pd) NULL 3180 static inline bool sched_energy_enabled(void) { return false; } 3181 3182 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */ 3183 3184 #ifdef CONFIG_MEMBARRIER 3185 /* 3186 * The scheduler provides memory barriers required by membarrier between: 3187 * - prior user-space memory accesses and store to rq->membarrier_state, 3188 * - store to rq->membarrier_state and following user-space memory accesses. 3189 * In the same way it provides those guarantees around store to rq->curr. 3190 */ 3191 static inline void membarrier_switch_mm(struct rq *rq, 3192 struct mm_struct *prev_mm, 3193 struct mm_struct *next_mm) 3194 { 3195 int membarrier_state; 3196 3197 if (prev_mm == next_mm) 3198 return; 3199 3200 membarrier_state = atomic_read(&next_mm->membarrier_state); 3201 if (READ_ONCE(rq->membarrier_state) == membarrier_state) 3202 return; 3203 3204 WRITE_ONCE(rq->membarrier_state, membarrier_state); 3205 } 3206 #else 3207 static inline void membarrier_switch_mm(struct rq *rq, 3208 struct mm_struct *prev_mm, 3209 struct mm_struct *next_mm) 3210 { 3211 } 3212 #endif 3213 3214 #ifdef CONFIG_SMP 3215 static inline bool is_per_cpu_kthread(struct task_struct *p) 3216 { 3217 if (!(p->flags & PF_KTHREAD)) 3218 return false; 3219 3220 if (p->nr_cpus_allowed != 1) 3221 return false; 3222 3223 return true; 3224 } 3225 #endif 3226 3227 extern void swake_up_all_locked(struct swait_queue_head *q); 3228 extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait); 3229 3230 extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags); 3231 3232 #ifdef CONFIG_PREEMPT_DYNAMIC 3233 extern int preempt_dynamic_mode; 3234 extern int sched_dynamic_mode(const char *str); 3235 extern void sched_dynamic_update(int mode); 3236 #endif 3237 3238 #ifdef CONFIG_SCHED_MM_CID 3239 3240 #define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */ 3241 #define MM_CID_SCAN_DELAY 100 /* 100ms */ 3242 3243 extern raw_spinlock_t cid_lock; 3244 extern int use_cid_lock; 3245 3246 extern void sched_mm_cid_migrate_from(struct task_struct *t); 3247 extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t); 3248 extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr); 3249 extern void init_sched_mm_cid(struct task_struct *t); 3250 3251 static inline void __mm_cid_put(struct mm_struct *mm, int cid) 3252 { 3253 if (cid < 0) 3254 return; 3255 cpumask_clear_cpu(cid, mm_cidmask(mm)); 3256 } 3257 3258 /* 3259 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to 3260 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to 3261 * be held to transition to other states. 3262 * 3263 * State transitions synchronized with cmpxchg or try_cmpxchg need to be 3264 * consistent across cpus, which prevents use of this_cpu_cmpxchg. 3265 */ 3266 static inline void mm_cid_put_lazy(struct task_struct *t) 3267 { 3268 struct mm_struct *mm = t->mm; 3269 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid; 3270 int cid; 3271 3272 lockdep_assert_irqs_disabled(); 3273 cid = __this_cpu_read(pcpu_cid->cid); 3274 if (!mm_cid_is_lazy_put(cid) || 3275 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET)) 3276 return; 3277 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid)); 3278 } 3279 3280 static inline int mm_cid_pcpu_unset(struct mm_struct *mm) 3281 { 3282 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid; 3283 int cid, res; 3284 3285 lockdep_assert_irqs_disabled(); 3286 cid = __this_cpu_read(pcpu_cid->cid); 3287 for (;;) { 3288 if (mm_cid_is_unset(cid)) 3289 return MM_CID_UNSET; 3290 /* 3291 * Attempt transition from valid or lazy-put to unset. 3292 */ 3293 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET); 3294 if (res == cid) 3295 break; 3296 cid = res; 3297 } 3298 return cid; 3299 } 3300 3301 static inline void mm_cid_put(struct mm_struct *mm) 3302 { 3303 int cid; 3304 3305 lockdep_assert_irqs_disabled(); 3306 cid = mm_cid_pcpu_unset(mm); 3307 if (cid == MM_CID_UNSET) 3308 return; 3309 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid)); 3310 } 3311 3312 static inline int __mm_cid_try_get(struct mm_struct *mm) 3313 { 3314 struct cpumask *cpumask; 3315 int cid; 3316 3317 cpumask = mm_cidmask(mm); 3318 /* 3319 * Retry finding first zero bit if the mask is temporarily 3320 * filled. This only happens during concurrent remote-clear 3321 * which owns a cid without holding a rq lock. 3322 */ 3323 for (;;) { 3324 cid = cpumask_first_zero(cpumask); 3325 if (cid < nr_cpu_ids) 3326 break; 3327 cpu_relax(); 3328 } 3329 if (cpumask_test_and_set_cpu(cid, cpumask)) 3330 return -1; 3331 return cid; 3332 } 3333 3334 /* 3335 * Save a snapshot of the current runqueue time of this cpu 3336 * with the per-cpu cid value, allowing to estimate how recently it was used. 3337 */ 3338 static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm) 3339 { 3340 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq)); 3341 3342 lockdep_assert_rq_held(rq); 3343 WRITE_ONCE(pcpu_cid->time, rq->clock); 3344 } 3345 3346 static inline int __mm_cid_get(struct rq *rq, struct mm_struct *mm) 3347 { 3348 int cid; 3349 3350 /* 3351 * All allocations (even those using the cid_lock) are lock-free. If 3352 * use_cid_lock is set, hold the cid_lock to perform cid allocation to 3353 * guarantee forward progress. 3354 */ 3355 if (!READ_ONCE(use_cid_lock)) { 3356 cid = __mm_cid_try_get(mm); 3357 if (cid >= 0) 3358 goto end; 3359 raw_spin_lock(&cid_lock); 3360 } else { 3361 raw_spin_lock(&cid_lock); 3362 cid = __mm_cid_try_get(mm); 3363 if (cid >= 0) 3364 goto unlock; 3365 } 3366 3367 /* 3368 * cid concurrently allocated. Retry while forcing following 3369 * allocations to use the cid_lock to ensure forward progress. 3370 */ 3371 WRITE_ONCE(use_cid_lock, 1); 3372 /* 3373 * Set use_cid_lock before allocation. Only care about program order 3374 * because this is only required for forward progress. 3375 */ 3376 barrier(); 3377 /* 3378 * Retry until it succeeds. It is guaranteed to eventually succeed once 3379 * all newcoming allocations observe the use_cid_lock flag set. 3380 */ 3381 do { 3382 cid = __mm_cid_try_get(mm); 3383 cpu_relax(); 3384 } while (cid < 0); 3385 /* 3386 * Allocate before clearing use_cid_lock. Only care about 3387 * program order because this is for forward progress. 3388 */ 3389 barrier(); 3390 WRITE_ONCE(use_cid_lock, 0); 3391 unlock: 3392 raw_spin_unlock(&cid_lock); 3393 end: 3394 mm_cid_snapshot_time(rq, mm); 3395 return cid; 3396 } 3397 3398 static inline int mm_cid_get(struct rq *rq, struct mm_struct *mm) 3399 { 3400 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid; 3401 struct cpumask *cpumask; 3402 int cid; 3403 3404 lockdep_assert_rq_held(rq); 3405 cpumask = mm_cidmask(mm); 3406 cid = __this_cpu_read(pcpu_cid->cid); 3407 if (mm_cid_is_valid(cid)) { 3408 mm_cid_snapshot_time(rq, mm); 3409 return cid; 3410 } 3411 if (mm_cid_is_lazy_put(cid)) { 3412 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET)) 3413 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid)); 3414 } 3415 cid = __mm_cid_get(rq, mm); 3416 __this_cpu_write(pcpu_cid->cid, cid); 3417 return cid; 3418 } 3419 3420 static inline void switch_mm_cid(struct rq *rq, 3421 struct task_struct *prev, 3422 struct task_struct *next) 3423 { 3424 /* 3425 * Provide a memory barrier between rq->curr store and load of 3426 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition. 3427 * 3428 * Should be adapted if context_switch() is modified. 3429 */ 3430 if (!next->mm) { // to kernel 3431 /* 3432 * user -> kernel transition does not guarantee a barrier, but 3433 * we can use the fact that it performs an atomic operation in 3434 * mmgrab(). 3435 */ 3436 if (prev->mm) // from user 3437 smp_mb__after_mmgrab(); 3438 /* 3439 * kernel -> kernel transition does not change rq->curr->mm 3440 * state. It stays NULL. 3441 */ 3442 } else { // to user 3443 /* 3444 * kernel -> user transition does not provide a barrier 3445 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu]. 3446 * Provide it here. 3447 */ 3448 if (!prev->mm) // from kernel 3449 smp_mb(); 3450 /* 3451 * user -> user transition guarantees a memory barrier through 3452 * switch_mm() when current->mm changes. If current->mm is 3453 * unchanged, no barrier is needed. 3454 */ 3455 } 3456 if (prev->mm_cid_active) { 3457 mm_cid_snapshot_time(rq, prev->mm); 3458 mm_cid_put_lazy(prev); 3459 prev->mm_cid = -1; 3460 } 3461 if (next->mm_cid_active) 3462 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next->mm); 3463 } 3464 3465 #else 3466 static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { } 3467 static inline void sched_mm_cid_migrate_from(struct task_struct *t) { } 3468 static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { } 3469 static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { } 3470 static inline void init_sched_mm_cid(struct task_struct *t) { } 3471 #endif 3472 3473 extern u64 avg_vruntime(struct cfs_rq *cfs_rq); 3474 extern int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se); 3475 3476 #endif /* _KERNEL_SCHED_SCHED_H */ 3477