1 /* 2 * kernel/workqueue.c - generic async execution with shared worker pool 3 * 4 * Copyright (C) 2002 Ingo Molnar 5 * 6 * Derived from the taskqueue/keventd code by: 7 * David Woodhouse <dwmw2@infradead.org> 8 * Andrew Morton 9 * Kai Petzke <wpp@marie.physik.tu-berlin.de> 10 * Theodore Ts'o <tytso@mit.edu> 11 * 12 * Made to use alloc_percpu by Christoph Lameter. 13 * 14 * Copyright (C) 2010 SUSE Linux Products GmbH 15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16 * 17 * This is the generic async execution mechanism. Work items as are 18 * executed in process context. The worker pool is shared and 19 * automatically managed. There are two worker pools for each CPU (one for 20 * normal work items and the other for high priority ones) and some extra 21 * pools for workqueues which are not bound to any specific CPU - the 22 * number of these backing pools is dynamic. 23 * 24 * Please read Documentation/core-api/workqueue.rst for details. 25 */ 26 27 #include <linux/export.h> 28 #include <linux/kernel.h> 29 #include <linux/sched.h> 30 #include <linux/init.h> 31 #include <linux/signal.h> 32 #include <linux/completion.h> 33 #include <linux/workqueue.h> 34 #include <linux/slab.h> 35 #include <linux/cpu.h> 36 #include <linux/notifier.h> 37 #include <linux/kthread.h> 38 #include <linux/hardirq.h> 39 #include <linux/mempolicy.h> 40 #include <linux/freezer.h> 41 #include <linux/debug_locks.h> 42 #include <linux/lockdep.h> 43 #include <linux/idr.h> 44 #include <linux/jhash.h> 45 #include <linux/hashtable.h> 46 #include <linux/rculist.h> 47 #include <linux/nodemask.h> 48 #include <linux/moduleparam.h> 49 #include <linux/uaccess.h> 50 #include <linux/sched/isolation.h> 51 #include <linux/nmi.h> 52 53 #include "workqueue_internal.h" 54 55 enum { 56 /* 57 * worker_pool flags 58 * 59 * A bound pool is either associated or disassociated with its CPU. 60 * While associated (!DISASSOCIATED), all workers are bound to the 61 * CPU and none has %WORKER_UNBOUND set and concurrency management 62 * is in effect. 63 * 64 * While DISASSOCIATED, the cpu may be offline and all workers have 65 * %WORKER_UNBOUND set and concurrency management disabled, and may 66 * be executing on any CPU. The pool behaves as an unbound one. 67 * 68 * Note that DISASSOCIATED should be flipped only while holding 69 * wq_pool_attach_mutex to avoid changing binding state while 70 * worker_attach_to_pool() is in progress. 71 */ 72 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 73 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 74 75 /* worker flags */ 76 WORKER_DIE = 1 << 1, /* die die die */ 77 WORKER_IDLE = 1 << 2, /* is idle */ 78 WORKER_PREP = 1 << 3, /* preparing to run works */ 79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 81 WORKER_REBOUND = 1 << 8, /* worker was rebound */ 82 83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 84 WORKER_UNBOUND | WORKER_REBOUND, 85 86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 87 88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 90 91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 93 94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 95 /* call for help after 10ms 96 (min two ticks) */ 97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 98 CREATE_COOLDOWN = HZ, /* time to breath after fail */ 99 100 /* 101 * Rescue workers are used only on emergencies and shared by 102 * all cpus. Give MIN_NICE. 103 */ 104 RESCUER_NICE_LEVEL = MIN_NICE, 105 HIGHPRI_NICE_LEVEL = MIN_NICE, 106 107 WQ_NAME_LEN = 24, 108 }; 109 110 /* 111 * Structure fields follow one of the following exclusion rules. 112 * 113 * I: Modifiable by initialization/destruction paths and read-only for 114 * everyone else. 115 * 116 * P: Preemption protected. Disabling preemption is enough and should 117 * only be modified and accessed from the local cpu. 118 * 119 * L: pool->lock protected. Access with pool->lock held. 120 * 121 * X: During normal operation, modification requires pool->lock and should 122 * be done only from local cpu. Either disabling preemption on local 123 * cpu or grabbing pool->lock is enough for read access. If 124 * POOL_DISASSOCIATED is set, it's identical to L. 125 * 126 * A: wq_pool_attach_mutex protected. 127 * 128 * PL: wq_pool_mutex protected. 129 * 130 * PR: wq_pool_mutex protected for writes. RCU protected for reads. 131 * 132 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 133 * 134 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 135 * RCU for reads. 136 * 137 * WQ: wq->mutex protected. 138 * 139 * WR: wq->mutex protected for writes. RCU protected for reads. 140 * 141 * MD: wq_mayday_lock protected. 142 */ 143 144 /* struct worker is defined in workqueue_internal.h */ 145 146 struct worker_pool { 147 spinlock_t lock; /* the pool lock */ 148 int cpu; /* I: the associated cpu */ 149 int node; /* I: the associated node ID */ 150 int id; /* I: pool ID */ 151 unsigned int flags; /* X: flags */ 152 153 unsigned long watchdog_ts; /* L: watchdog timestamp */ 154 155 struct list_head worklist; /* L: list of pending works */ 156 157 int nr_workers; /* L: total number of workers */ 158 int nr_idle; /* L: currently idle workers */ 159 160 struct list_head idle_list; /* X: list of idle workers */ 161 struct timer_list idle_timer; /* L: worker idle timeout */ 162 struct timer_list mayday_timer; /* L: SOS timer for workers */ 163 164 /* a workers is either on busy_hash or idle_list, or the manager */ 165 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 166 /* L: hash of busy workers */ 167 168 struct worker *manager; /* L: purely informational */ 169 struct list_head workers; /* A: attached workers */ 170 struct completion *detach_completion; /* all workers detached */ 171 172 struct ida worker_ida; /* worker IDs for task name */ 173 174 struct workqueue_attrs *attrs; /* I: worker attributes */ 175 struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 176 int refcnt; /* PL: refcnt for unbound pools */ 177 178 /* 179 * The current concurrency level. As it's likely to be accessed 180 * from other CPUs during try_to_wake_up(), put it in a separate 181 * cacheline. 182 */ 183 atomic_t nr_running ____cacheline_aligned_in_smp; 184 185 /* 186 * Destruction of pool is RCU protected to allow dereferences 187 * from get_work_pool(). 188 */ 189 struct rcu_head rcu; 190 } ____cacheline_aligned_in_smp; 191 192 /* 193 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 194 * of work_struct->data are used for flags and the remaining high bits 195 * point to the pwq; thus, pwqs need to be aligned at two's power of the 196 * number of flag bits. 197 */ 198 struct pool_workqueue { 199 struct worker_pool *pool; /* I: the associated pool */ 200 struct workqueue_struct *wq; /* I: the owning workqueue */ 201 int work_color; /* L: current color */ 202 int flush_color; /* L: flushing color */ 203 int refcnt; /* L: reference count */ 204 int nr_in_flight[WORK_NR_COLORS]; 205 /* L: nr of in_flight works */ 206 int nr_active; /* L: nr of active works */ 207 int max_active; /* L: max active works */ 208 struct list_head delayed_works; /* L: delayed works */ 209 struct list_head pwqs_node; /* WR: node on wq->pwqs */ 210 struct list_head mayday_node; /* MD: node on wq->maydays */ 211 212 /* 213 * Release of unbound pwq is punted to system_wq. See put_pwq() 214 * and pwq_unbound_release_workfn() for details. pool_workqueue 215 * itself is also RCU protected so that the first pwq can be 216 * determined without grabbing wq->mutex. 217 */ 218 struct work_struct unbound_release_work; 219 struct rcu_head rcu; 220 } __aligned(1 << WORK_STRUCT_FLAG_BITS); 221 222 /* 223 * Structure used to wait for workqueue flush. 224 */ 225 struct wq_flusher { 226 struct list_head list; /* WQ: list of flushers */ 227 int flush_color; /* WQ: flush color waiting for */ 228 struct completion done; /* flush completion */ 229 }; 230 231 struct wq_device; 232 233 /* 234 * The externally visible workqueue. It relays the issued work items to 235 * the appropriate worker_pool through its pool_workqueues. 236 */ 237 struct workqueue_struct { 238 struct list_head pwqs; /* WR: all pwqs of this wq */ 239 struct list_head list; /* PR: list of all workqueues */ 240 241 struct mutex mutex; /* protects this wq */ 242 int work_color; /* WQ: current work color */ 243 int flush_color; /* WQ: current flush color */ 244 atomic_t nr_pwqs_to_flush; /* flush in progress */ 245 struct wq_flusher *first_flusher; /* WQ: first flusher */ 246 struct list_head flusher_queue; /* WQ: flush waiters */ 247 struct list_head flusher_overflow; /* WQ: flush overflow list */ 248 249 struct list_head maydays; /* MD: pwqs requesting rescue */ 250 struct worker *rescuer; /* I: rescue worker */ 251 252 int nr_drainers; /* WQ: drain in progress */ 253 int saved_max_active; /* WQ: saved pwq max_active */ 254 255 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 256 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 257 258 #ifdef CONFIG_SYSFS 259 struct wq_device *wq_dev; /* I: for sysfs interface */ 260 #endif 261 #ifdef CONFIG_LOCKDEP 262 char *lock_name; 263 struct lock_class_key key; 264 struct lockdep_map lockdep_map; 265 #endif 266 char name[WQ_NAME_LEN]; /* I: workqueue name */ 267 268 /* 269 * Destruction of workqueue_struct is RCU protected to allow walking 270 * the workqueues list without grabbing wq_pool_mutex. 271 * This is used to dump all workqueues from sysrq. 272 */ 273 struct rcu_head rcu; 274 275 /* hot fields used during command issue, aligned to cacheline */ 276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 279 }; 280 281 static struct kmem_cache *pwq_cache; 282 283 static cpumask_var_t *wq_numa_possible_cpumask; 284 /* possible CPUs of each node */ 285 286 static bool wq_disable_numa; 287 module_param_named(disable_numa, wq_disable_numa, bool, 0444); 288 289 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 290 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 291 module_param_named(power_efficient, wq_power_efficient, bool, 0444); 292 293 static bool wq_online; /* can kworkers be created yet? */ 294 295 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 296 297 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 298 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 299 300 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 301 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 302 static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 303 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ 304 305 static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 306 static bool workqueue_freezing; /* PL: have wqs started freezing? */ 307 308 /* PL: allowable cpus for unbound wqs and work items */ 309 static cpumask_var_t wq_unbound_cpumask; 310 311 /* CPU where unbound work was last round robin scheduled from this CPU */ 312 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 313 314 /* 315 * Local execution of unbound work items is no longer guaranteed. The 316 * following always forces round-robin CPU selection on unbound work items 317 * to uncover usages which depend on it. 318 */ 319 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 320 static bool wq_debug_force_rr_cpu = true; 321 #else 322 static bool wq_debug_force_rr_cpu = false; 323 #endif 324 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 325 326 /* the per-cpu worker pools */ 327 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 328 329 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 330 331 /* PL: hash of all unbound pools keyed by pool->attrs */ 332 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 333 334 /* I: attributes used when instantiating standard unbound pools on demand */ 335 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 336 337 /* I: attributes used when instantiating ordered pools on demand */ 338 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 339 340 struct workqueue_struct *system_wq __read_mostly; 341 EXPORT_SYMBOL(system_wq); 342 struct workqueue_struct *system_highpri_wq __read_mostly; 343 EXPORT_SYMBOL_GPL(system_highpri_wq); 344 struct workqueue_struct *system_long_wq __read_mostly; 345 EXPORT_SYMBOL_GPL(system_long_wq); 346 struct workqueue_struct *system_unbound_wq __read_mostly; 347 EXPORT_SYMBOL_GPL(system_unbound_wq); 348 struct workqueue_struct *system_freezable_wq __read_mostly; 349 EXPORT_SYMBOL_GPL(system_freezable_wq); 350 struct workqueue_struct *system_power_efficient_wq __read_mostly; 351 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 352 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 353 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 354 355 static int worker_thread(void *__worker); 356 static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 357 358 #define CREATE_TRACE_POINTS 359 #include <trace/events/workqueue.h> 360 361 #define assert_rcu_or_pool_mutex() \ 362 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 363 !lockdep_is_held(&wq_pool_mutex), \ 364 "RCU or wq_pool_mutex should be held") 365 366 #define assert_rcu_or_wq_mutex(wq) \ 367 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 368 !lockdep_is_held(&wq->mutex), \ 369 "RCU or wq->mutex should be held") 370 371 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 372 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 373 !lockdep_is_held(&wq->mutex) && \ 374 !lockdep_is_held(&wq_pool_mutex), \ 375 "RCU, wq->mutex or wq_pool_mutex should be held") 376 377 #define for_each_cpu_worker_pool(pool, cpu) \ 378 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 379 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 380 (pool)++) 381 382 /** 383 * for_each_pool - iterate through all worker_pools in the system 384 * @pool: iteration cursor 385 * @pi: integer used for iteration 386 * 387 * This must be called either with wq_pool_mutex held or RCU read 388 * locked. If the pool needs to be used beyond the locking in effect, the 389 * caller is responsible for guaranteeing that the pool stays online. 390 * 391 * The if/else clause exists only for the lockdep assertion and can be 392 * ignored. 393 */ 394 #define for_each_pool(pool, pi) \ 395 idr_for_each_entry(&worker_pool_idr, pool, pi) \ 396 if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 397 else 398 399 /** 400 * for_each_pool_worker - iterate through all workers of a worker_pool 401 * @worker: iteration cursor 402 * @pool: worker_pool to iterate workers of 403 * 404 * This must be called with wq_pool_attach_mutex. 405 * 406 * The if/else clause exists only for the lockdep assertion and can be 407 * ignored. 408 */ 409 #define for_each_pool_worker(worker, pool) \ 410 list_for_each_entry((worker), &(pool)->workers, node) \ 411 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 412 else 413 414 /** 415 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 416 * @pwq: iteration cursor 417 * @wq: the target workqueue 418 * 419 * This must be called either with wq->mutex held or RCU read locked. 420 * If the pwq needs to be used beyond the locking in effect, the caller is 421 * responsible for guaranteeing that the pwq stays online. 422 * 423 * The if/else clause exists only for the lockdep assertion and can be 424 * ignored. 425 */ 426 #define for_each_pwq(pwq, wq) \ 427 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 428 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 429 else 430 431 #ifdef CONFIG_DEBUG_OBJECTS_WORK 432 433 static struct debug_obj_descr work_debug_descr; 434 435 static void *work_debug_hint(void *addr) 436 { 437 return ((struct work_struct *) addr)->func; 438 } 439 440 static bool work_is_static_object(void *addr) 441 { 442 struct work_struct *work = addr; 443 444 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 445 } 446 447 /* 448 * fixup_init is called when: 449 * - an active object is initialized 450 */ 451 static bool work_fixup_init(void *addr, enum debug_obj_state state) 452 { 453 struct work_struct *work = addr; 454 455 switch (state) { 456 case ODEBUG_STATE_ACTIVE: 457 cancel_work_sync(work); 458 debug_object_init(work, &work_debug_descr); 459 return true; 460 default: 461 return false; 462 } 463 } 464 465 /* 466 * fixup_free is called when: 467 * - an active object is freed 468 */ 469 static bool work_fixup_free(void *addr, enum debug_obj_state state) 470 { 471 struct work_struct *work = addr; 472 473 switch (state) { 474 case ODEBUG_STATE_ACTIVE: 475 cancel_work_sync(work); 476 debug_object_free(work, &work_debug_descr); 477 return true; 478 default: 479 return false; 480 } 481 } 482 483 static struct debug_obj_descr work_debug_descr = { 484 .name = "work_struct", 485 .debug_hint = work_debug_hint, 486 .is_static_object = work_is_static_object, 487 .fixup_init = work_fixup_init, 488 .fixup_free = work_fixup_free, 489 }; 490 491 static inline void debug_work_activate(struct work_struct *work) 492 { 493 debug_object_activate(work, &work_debug_descr); 494 } 495 496 static inline void debug_work_deactivate(struct work_struct *work) 497 { 498 debug_object_deactivate(work, &work_debug_descr); 499 } 500 501 void __init_work(struct work_struct *work, int onstack) 502 { 503 if (onstack) 504 debug_object_init_on_stack(work, &work_debug_descr); 505 else 506 debug_object_init(work, &work_debug_descr); 507 } 508 EXPORT_SYMBOL_GPL(__init_work); 509 510 void destroy_work_on_stack(struct work_struct *work) 511 { 512 debug_object_free(work, &work_debug_descr); 513 } 514 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 515 516 void destroy_delayed_work_on_stack(struct delayed_work *work) 517 { 518 destroy_timer_on_stack(&work->timer); 519 debug_object_free(&work->work, &work_debug_descr); 520 } 521 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 522 523 #else 524 static inline void debug_work_activate(struct work_struct *work) { } 525 static inline void debug_work_deactivate(struct work_struct *work) { } 526 #endif 527 528 /** 529 * worker_pool_assign_id - allocate ID and assing it to @pool 530 * @pool: the pool pointer of interest 531 * 532 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 533 * successfully, -errno on failure. 534 */ 535 static int worker_pool_assign_id(struct worker_pool *pool) 536 { 537 int ret; 538 539 lockdep_assert_held(&wq_pool_mutex); 540 541 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 542 GFP_KERNEL); 543 if (ret >= 0) { 544 pool->id = ret; 545 return 0; 546 } 547 return ret; 548 } 549 550 /** 551 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 552 * @wq: the target workqueue 553 * @node: the node ID 554 * 555 * This must be called with any of wq_pool_mutex, wq->mutex or RCU 556 * read locked. 557 * If the pwq needs to be used beyond the locking in effect, the caller is 558 * responsible for guaranteeing that the pwq stays online. 559 * 560 * Return: The unbound pool_workqueue for @node. 561 */ 562 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 563 int node) 564 { 565 assert_rcu_or_wq_mutex_or_pool_mutex(wq); 566 567 /* 568 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 569 * delayed item is pending. The plan is to keep CPU -> NODE 570 * mapping valid and stable across CPU on/offlines. Once that 571 * happens, this workaround can be removed. 572 */ 573 if (unlikely(node == NUMA_NO_NODE)) 574 return wq->dfl_pwq; 575 576 return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 577 } 578 579 static unsigned int work_color_to_flags(int color) 580 { 581 return color << WORK_STRUCT_COLOR_SHIFT; 582 } 583 584 static int get_work_color(struct work_struct *work) 585 { 586 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 587 ((1 << WORK_STRUCT_COLOR_BITS) - 1); 588 } 589 590 static int work_next_color(int color) 591 { 592 return (color + 1) % WORK_NR_COLORS; 593 } 594 595 /* 596 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 597 * contain the pointer to the queued pwq. Once execution starts, the flag 598 * is cleared and the high bits contain OFFQ flags and pool ID. 599 * 600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 601 * and clear_work_data() can be used to set the pwq, pool or clear 602 * work->data. These functions should only be called while the work is 603 * owned - ie. while the PENDING bit is set. 604 * 605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 606 * corresponding to a work. Pool is available once the work has been 607 * queued anywhere after initialization until it is sync canceled. pwq is 608 * available only while the work item is queued. 609 * 610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being 611 * canceled. While being canceled, a work item may have its PENDING set 612 * but stay off timer and worklist for arbitrarily long and nobody should 613 * try to steal the PENDING bit. 614 */ 615 static inline void set_work_data(struct work_struct *work, unsigned long data, 616 unsigned long flags) 617 { 618 WARN_ON_ONCE(!work_pending(work)); 619 atomic_long_set(&work->data, data | flags | work_static(work)); 620 } 621 622 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 623 unsigned long extra_flags) 624 { 625 set_work_data(work, (unsigned long)pwq, 626 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 627 } 628 629 static void set_work_pool_and_keep_pending(struct work_struct *work, 630 int pool_id) 631 { 632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 633 WORK_STRUCT_PENDING); 634 } 635 636 static void set_work_pool_and_clear_pending(struct work_struct *work, 637 int pool_id) 638 { 639 /* 640 * The following wmb is paired with the implied mb in 641 * test_and_set_bit(PENDING) and ensures all updates to @work made 642 * here are visible to and precede any updates by the next PENDING 643 * owner. 644 */ 645 smp_wmb(); 646 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 647 /* 648 * The following mb guarantees that previous clear of a PENDING bit 649 * will not be reordered with any speculative LOADS or STORES from 650 * work->current_func, which is executed afterwards. This possible 651 * reordering can lead to a missed execution on attempt to queue 652 * the same @work. E.g. consider this case: 653 * 654 * CPU#0 CPU#1 655 * ---------------------------- -------------------------------- 656 * 657 * 1 STORE event_indicated 658 * 2 queue_work_on() { 659 * 3 test_and_set_bit(PENDING) 660 * 4 } set_..._and_clear_pending() { 661 * 5 set_work_data() # clear bit 662 * 6 smp_mb() 663 * 7 work->current_func() { 664 * 8 LOAD event_indicated 665 * } 666 * 667 * Without an explicit full barrier speculative LOAD on line 8 can 668 * be executed before CPU#0 does STORE on line 1. If that happens, 669 * CPU#0 observes the PENDING bit is still set and new execution of 670 * a @work is not queued in a hope, that CPU#1 will eventually 671 * finish the queued @work. Meanwhile CPU#1 does not see 672 * event_indicated is set, because speculative LOAD was executed 673 * before actual STORE. 674 */ 675 smp_mb(); 676 } 677 678 static void clear_work_data(struct work_struct *work) 679 { 680 smp_wmb(); /* see set_work_pool_and_clear_pending() */ 681 set_work_data(work, WORK_STRUCT_NO_POOL, 0); 682 } 683 684 static struct pool_workqueue *get_work_pwq(struct work_struct *work) 685 { 686 unsigned long data = atomic_long_read(&work->data); 687 688 if (data & WORK_STRUCT_PWQ) 689 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 690 else 691 return NULL; 692 } 693 694 /** 695 * get_work_pool - return the worker_pool a given work was associated with 696 * @work: the work item of interest 697 * 698 * Pools are created and destroyed under wq_pool_mutex, and allows read 699 * access under RCU read lock. As such, this function should be 700 * called under wq_pool_mutex or inside of a rcu_read_lock() region. 701 * 702 * All fields of the returned pool are accessible as long as the above 703 * mentioned locking is in effect. If the returned pool needs to be used 704 * beyond the critical section, the caller is responsible for ensuring the 705 * returned pool is and stays online. 706 * 707 * Return: The worker_pool @work was last associated with. %NULL if none. 708 */ 709 static struct worker_pool *get_work_pool(struct work_struct *work) 710 { 711 unsigned long data = atomic_long_read(&work->data); 712 int pool_id; 713 714 assert_rcu_or_pool_mutex(); 715 716 if (data & WORK_STRUCT_PWQ) 717 return ((struct pool_workqueue *) 718 (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 719 720 pool_id = data >> WORK_OFFQ_POOL_SHIFT; 721 if (pool_id == WORK_OFFQ_POOL_NONE) 722 return NULL; 723 724 return idr_find(&worker_pool_idr, pool_id); 725 } 726 727 /** 728 * get_work_pool_id - return the worker pool ID a given work is associated with 729 * @work: the work item of interest 730 * 731 * Return: The worker_pool ID @work was last associated with. 732 * %WORK_OFFQ_POOL_NONE if none. 733 */ 734 static int get_work_pool_id(struct work_struct *work) 735 { 736 unsigned long data = atomic_long_read(&work->data); 737 738 if (data & WORK_STRUCT_PWQ) 739 return ((struct pool_workqueue *) 740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 741 742 return data >> WORK_OFFQ_POOL_SHIFT; 743 } 744 745 static void mark_work_canceling(struct work_struct *work) 746 { 747 unsigned long pool_id = get_work_pool_id(work); 748 749 pool_id <<= WORK_OFFQ_POOL_SHIFT; 750 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 751 } 752 753 static bool work_is_canceling(struct work_struct *work) 754 { 755 unsigned long data = atomic_long_read(&work->data); 756 757 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 758 } 759 760 /* 761 * Policy functions. These define the policies on how the global worker 762 * pools are managed. Unless noted otherwise, these functions assume that 763 * they're being called with pool->lock held. 764 */ 765 766 static bool __need_more_worker(struct worker_pool *pool) 767 { 768 return !atomic_read(&pool->nr_running); 769 } 770 771 /* 772 * Need to wake up a worker? Called from anything but currently 773 * running workers. 774 * 775 * Note that, because unbound workers never contribute to nr_running, this 776 * function will always return %true for unbound pools as long as the 777 * worklist isn't empty. 778 */ 779 static bool need_more_worker(struct worker_pool *pool) 780 { 781 return !list_empty(&pool->worklist) && __need_more_worker(pool); 782 } 783 784 /* Can I start working? Called from busy but !running workers. */ 785 static bool may_start_working(struct worker_pool *pool) 786 { 787 return pool->nr_idle; 788 } 789 790 /* Do I need to keep working? Called from currently running workers. */ 791 static bool keep_working(struct worker_pool *pool) 792 { 793 return !list_empty(&pool->worklist) && 794 atomic_read(&pool->nr_running) <= 1; 795 } 796 797 /* Do we need a new worker? Called from manager. */ 798 static bool need_to_create_worker(struct worker_pool *pool) 799 { 800 return need_more_worker(pool) && !may_start_working(pool); 801 } 802 803 /* Do we have too many workers and should some go away? */ 804 static bool too_many_workers(struct worker_pool *pool) 805 { 806 bool managing = pool->flags & POOL_MANAGER_ACTIVE; 807 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 808 int nr_busy = pool->nr_workers - nr_idle; 809 810 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 811 } 812 813 /* 814 * Wake up functions. 815 */ 816 817 /* Return the first idle worker. Safe with preemption disabled */ 818 static struct worker *first_idle_worker(struct worker_pool *pool) 819 { 820 if (unlikely(list_empty(&pool->idle_list))) 821 return NULL; 822 823 return list_first_entry(&pool->idle_list, struct worker, entry); 824 } 825 826 /** 827 * wake_up_worker - wake up an idle worker 828 * @pool: worker pool to wake worker from 829 * 830 * Wake up the first idle worker of @pool. 831 * 832 * CONTEXT: 833 * spin_lock_irq(pool->lock). 834 */ 835 static void wake_up_worker(struct worker_pool *pool) 836 { 837 struct worker *worker = first_idle_worker(pool); 838 839 if (likely(worker)) 840 wake_up_process(worker->task); 841 } 842 843 /** 844 * wq_worker_running - a worker is running again 845 * @task: task waking up 846 * 847 * This function is called when a worker returns from schedule() 848 */ 849 void wq_worker_running(struct task_struct *task) 850 { 851 struct worker *worker = kthread_data(task); 852 853 if (!worker->sleeping) 854 return; 855 if (!(worker->flags & WORKER_NOT_RUNNING)) 856 atomic_inc(&worker->pool->nr_running); 857 worker->sleeping = 0; 858 } 859 860 /** 861 * wq_worker_sleeping - a worker is going to sleep 862 * @task: task going to sleep 863 * 864 * This function is called from schedule() when a busy worker is 865 * going to sleep. 866 */ 867 void wq_worker_sleeping(struct task_struct *task) 868 { 869 struct worker *next, *worker = kthread_data(task); 870 struct worker_pool *pool; 871 872 /* 873 * Rescuers, which may not have all the fields set up like normal 874 * workers, also reach here, let's not access anything before 875 * checking NOT_RUNNING. 876 */ 877 if (worker->flags & WORKER_NOT_RUNNING) 878 return; 879 880 pool = worker->pool; 881 882 if (WARN_ON_ONCE(worker->sleeping)) 883 return; 884 885 worker->sleeping = 1; 886 spin_lock_irq(&pool->lock); 887 888 /* 889 * The counterpart of the following dec_and_test, implied mb, 890 * worklist not empty test sequence is in insert_work(). 891 * Please read comment there. 892 * 893 * NOT_RUNNING is clear. This means that we're bound to and 894 * running on the local cpu w/ rq lock held and preemption 895 * disabled, which in turn means that none else could be 896 * manipulating idle_list, so dereferencing idle_list without pool 897 * lock is safe. 898 */ 899 if (atomic_dec_and_test(&pool->nr_running) && 900 !list_empty(&pool->worklist)) { 901 next = first_idle_worker(pool); 902 if (next) 903 wake_up_process(next->task); 904 } 905 spin_unlock_irq(&pool->lock); 906 } 907 908 /** 909 * wq_worker_last_func - retrieve worker's last work function 910 * @task: Task to retrieve last work function of. 911 * 912 * Determine the last function a worker executed. This is called from 913 * the scheduler to get a worker's last known identity. 914 * 915 * CONTEXT: 916 * spin_lock_irq(rq->lock) 917 * 918 * This function is called during schedule() when a kworker is going 919 * to sleep. It's used by psi to identify aggregation workers during 920 * dequeuing, to allow periodic aggregation to shut-off when that 921 * worker is the last task in the system or cgroup to go to sleep. 922 * 923 * As this function doesn't involve any workqueue-related locking, it 924 * only returns stable values when called from inside the scheduler's 925 * queuing and dequeuing paths, when @task, which must be a kworker, 926 * is guaranteed to not be processing any works. 927 * 928 * Return: 929 * The last work function %current executed as a worker, NULL if it 930 * hasn't executed any work yet. 931 */ 932 work_func_t wq_worker_last_func(struct task_struct *task) 933 { 934 struct worker *worker = kthread_data(task); 935 936 return worker->last_func; 937 } 938 939 /** 940 * worker_set_flags - set worker flags and adjust nr_running accordingly 941 * @worker: self 942 * @flags: flags to set 943 * 944 * Set @flags in @worker->flags and adjust nr_running accordingly. 945 * 946 * CONTEXT: 947 * spin_lock_irq(pool->lock) 948 */ 949 static inline void worker_set_flags(struct worker *worker, unsigned int flags) 950 { 951 struct worker_pool *pool = worker->pool; 952 953 WARN_ON_ONCE(worker->task != current); 954 955 /* If transitioning into NOT_RUNNING, adjust nr_running. */ 956 if ((flags & WORKER_NOT_RUNNING) && 957 !(worker->flags & WORKER_NOT_RUNNING)) { 958 atomic_dec(&pool->nr_running); 959 } 960 961 worker->flags |= flags; 962 } 963 964 /** 965 * worker_clr_flags - clear worker flags and adjust nr_running accordingly 966 * @worker: self 967 * @flags: flags to clear 968 * 969 * Clear @flags in @worker->flags and adjust nr_running accordingly. 970 * 971 * CONTEXT: 972 * spin_lock_irq(pool->lock) 973 */ 974 static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 975 { 976 struct worker_pool *pool = worker->pool; 977 unsigned int oflags = worker->flags; 978 979 WARN_ON_ONCE(worker->task != current); 980 981 worker->flags &= ~flags; 982 983 /* 984 * If transitioning out of NOT_RUNNING, increment nr_running. Note 985 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 986 * of multiple flags, not a single flag. 987 */ 988 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 989 if (!(worker->flags & WORKER_NOT_RUNNING)) 990 atomic_inc(&pool->nr_running); 991 } 992 993 /** 994 * find_worker_executing_work - find worker which is executing a work 995 * @pool: pool of interest 996 * @work: work to find worker for 997 * 998 * Find a worker which is executing @work on @pool by searching 999 * @pool->busy_hash which is keyed by the address of @work. For a worker 1000 * to match, its current execution should match the address of @work and 1001 * its work function. This is to avoid unwanted dependency between 1002 * unrelated work executions through a work item being recycled while still 1003 * being executed. 1004 * 1005 * This is a bit tricky. A work item may be freed once its execution 1006 * starts and nothing prevents the freed area from being recycled for 1007 * another work item. If the same work item address ends up being reused 1008 * before the original execution finishes, workqueue will identify the 1009 * recycled work item as currently executing and make it wait until the 1010 * current execution finishes, introducing an unwanted dependency. 1011 * 1012 * This function checks the work item address and work function to avoid 1013 * false positives. Note that this isn't complete as one may construct a 1014 * work function which can introduce dependency onto itself through a 1015 * recycled work item. Well, if somebody wants to shoot oneself in the 1016 * foot that badly, there's only so much we can do, and if such deadlock 1017 * actually occurs, it should be easy to locate the culprit work function. 1018 * 1019 * CONTEXT: 1020 * spin_lock_irq(pool->lock). 1021 * 1022 * Return: 1023 * Pointer to worker which is executing @work if found, %NULL 1024 * otherwise. 1025 */ 1026 static struct worker *find_worker_executing_work(struct worker_pool *pool, 1027 struct work_struct *work) 1028 { 1029 struct worker *worker; 1030 1031 hash_for_each_possible(pool->busy_hash, worker, hentry, 1032 (unsigned long)work) 1033 if (worker->current_work == work && 1034 worker->current_func == work->func) 1035 return worker; 1036 1037 return NULL; 1038 } 1039 1040 /** 1041 * move_linked_works - move linked works to a list 1042 * @work: start of series of works to be scheduled 1043 * @head: target list to append @work to 1044 * @nextp: out parameter for nested worklist walking 1045 * 1046 * Schedule linked works starting from @work to @head. Work series to 1047 * be scheduled starts at @work and includes any consecutive work with 1048 * WORK_STRUCT_LINKED set in its predecessor. 1049 * 1050 * If @nextp is not NULL, it's updated to point to the next work of 1051 * the last scheduled work. This allows move_linked_works() to be 1052 * nested inside outer list_for_each_entry_safe(). 1053 * 1054 * CONTEXT: 1055 * spin_lock_irq(pool->lock). 1056 */ 1057 static void move_linked_works(struct work_struct *work, struct list_head *head, 1058 struct work_struct **nextp) 1059 { 1060 struct work_struct *n; 1061 1062 /* 1063 * Linked worklist will always end before the end of the list, 1064 * use NULL for list head. 1065 */ 1066 list_for_each_entry_safe_from(work, n, NULL, entry) { 1067 list_move_tail(&work->entry, head); 1068 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1069 break; 1070 } 1071 1072 /* 1073 * If we're already inside safe list traversal and have moved 1074 * multiple works to the scheduled queue, the next position 1075 * needs to be updated. 1076 */ 1077 if (nextp) 1078 *nextp = n; 1079 } 1080 1081 /** 1082 * get_pwq - get an extra reference on the specified pool_workqueue 1083 * @pwq: pool_workqueue to get 1084 * 1085 * Obtain an extra reference on @pwq. The caller should guarantee that 1086 * @pwq has positive refcnt and be holding the matching pool->lock. 1087 */ 1088 static void get_pwq(struct pool_workqueue *pwq) 1089 { 1090 lockdep_assert_held(&pwq->pool->lock); 1091 WARN_ON_ONCE(pwq->refcnt <= 0); 1092 pwq->refcnt++; 1093 } 1094 1095 /** 1096 * put_pwq - put a pool_workqueue reference 1097 * @pwq: pool_workqueue to put 1098 * 1099 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 1100 * destruction. The caller should be holding the matching pool->lock. 1101 */ 1102 static void put_pwq(struct pool_workqueue *pwq) 1103 { 1104 lockdep_assert_held(&pwq->pool->lock); 1105 if (likely(--pwq->refcnt)) 1106 return; 1107 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 1108 return; 1109 /* 1110 * @pwq can't be released under pool->lock, bounce to 1111 * pwq_unbound_release_workfn(). This never recurses on the same 1112 * pool->lock as this path is taken only for unbound workqueues and 1113 * the release work item is scheduled on a per-cpu workqueue. To 1114 * avoid lockdep warning, unbound pool->locks are given lockdep 1115 * subclass of 1 in get_unbound_pool(). 1116 */ 1117 schedule_work(&pwq->unbound_release_work); 1118 } 1119 1120 /** 1121 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1122 * @pwq: pool_workqueue to put (can be %NULL) 1123 * 1124 * put_pwq() with locking. This function also allows %NULL @pwq. 1125 */ 1126 static void put_pwq_unlocked(struct pool_workqueue *pwq) 1127 { 1128 if (pwq) { 1129 /* 1130 * As both pwqs and pools are RCU protected, the 1131 * following lock operations are safe. 1132 */ 1133 spin_lock_irq(&pwq->pool->lock); 1134 put_pwq(pwq); 1135 spin_unlock_irq(&pwq->pool->lock); 1136 } 1137 } 1138 1139 static void pwq_activate_delayed_work(struct work_struct *work) 1140 { 1141 struct pool_workqueue *pwq = get_work_pwq(work); 1142 1143 trace_workqueue_activate_work(work); 1144 if (list_empty(&pwq->pool->worklist)) 1145 pwq->pool->watchdog_ts = jiffies; 1146 move_linked_works(work, &pwq->pool->worklist, NULL); 1147 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1148 pwq->nr_active++; 1149 } 1150 1151 static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 1152 { 1153 struct work_struct *work = list_first_entry(&pwq->delayed_works, 1154 struct work_struct, entry); 1155 1156 pwq_activate_delayed_work(work); 1157 } 1158 1159 /** 1160 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1161 * @pwq: pwq of interest 1162 * @color: color of work which left the queue 1163 * 1164 * A work either has completed or is removed from pending queue, 1165 * decrement nr_in_flight of its pwq and handle workqueue flushing. 1166 * 1167 * CONTEXT: 1168 * spin_lock_irq(pool->lock). 1169 */ 1170 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1171 { 1172 /* uncolored work items don't participate in flushing or nr_active */ 1173 if (color == WORK_NO_COLOR) 1174 goto out_put; 1175 1176 pwq->nr_in_flight[color]--; 1177 1178 pwq->nr_active--; 1179 if (!list_empty(&pwq->delayed_works)) { 1180 /* one down, submit a delayed one */ 1181 if (pwq->nr_active < pwq->max_active) 1182 pwq_activate_first_delayed(pwq); 1183 } 1184 1185 /* is flush in progress and are we at the flushing tip? */ 1186 if (likely(pwq->flush_color != color)) 1187 goto out_put; 1188 1189 /* are there still in-flight works? */ 1190 if (pwq->nr_in_flight[color]) 1191 goto out_put; 1192 1193 /* this pwq is done, clear flush_color */ 1194 pwq->flush_color = -1; 1195 1196 /* 1197 * If this was the last pwq, wake up the first flusher. It 1198 * will handle the rest. 1199 */ 1200 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1201 complete(&pwq->wq->first_flusher->done); 1202 out_put: 1203 put_pwq(pwq); 1204 } 1205 1206 /** 1207 * try_to_grab_pending - steal work item from worklist and disable irq 1208 * @work: work item to steal 1209 * @is_dwork: @work is a delayed_work 1210 * @flags: place to store irq state 1211 * 1212 * Try to grab PENDING bit of @work. This function can handle @work in any 1213 * stable state - idle, on timer or on worklist. 1214 * 1215 * Return: 1216 * 1 if @work was pending and we successfully stole PENDING 1217 * 0 if @work was idle and we claimed PENDING 1218 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1219 * -ENOENT if someone else is canceling @work, this state may persist 1220 * for arbitrarily long 1221 * 1222 * Note: 1223 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1224 * interrupted while holding PENDING and @work off queue, irq must be 1225 * disabled on entry. This, combined with delayed_work->timer being 1226 * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1227 * 1228 * On successful return, >= 0, irq is disabled and the caller is 1229 * responsible for releasing it using local_irq_restore(*@flags). 1230 * 1231 * This function is safe to call from any context including IRQ handler. 1232 */ 1233 static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1234 unsigned long *flags) 1235 { 1236 struct worker_pool *pool; 1237 struct pool_workqueue *pwq; 1238 1239 local_irq_save(*flags); 1240 1241 /* try to steal the timer if it exists */ 1242 if (is_dwork) { 1243 struct delayed_work *dwork = to_delayed_work(work); 1244 1245 /* 1246 * dwork->timer is irqsafe. If del_timer() fails, it's 1247 * guaranteed that the timer is not queued anywhere and not 1248 * running on the local CPU. 1249 */ 1250 if (likely(del_timer(&dwork->timer))) 1251 return 1; 1252 } 1253 1254 /* try to claim PENDING the normal way */ 1255 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1256 return 0; 1257 1258 rcu_read_lock(); 1259 /* 1260 * The queueing is in progress, or it is already queued. Try to 1261 * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1262 */ 1263 pool = get_work_pool(work); 1264 if (!pool) 1265 goto fail; 1266 1267 spin_lock(&pool->lock); 1268 /* 1269 * work->data is guaranteed to point to pwq only while the work 1270 * item is queued on pwq->wq, and both updating work->data to point 1271 * to pwq on queueing and to pool on dequeueing are done under 1272 * pwq->pool->lock. This in turn guarantees that, if work->data 1273 * points to pwq which is associated with a locked pool, the work 1274 * item is currently queued on that pool. 1275 */ 1276 pwq = get_work_pwq(work); 1277 if (pwq && pwq->pool == pool) { 1278 debug_work_deactivate(work); 1279 1280 /* 1281 * A delayed work item cannot be grabbed directly because 1282 * it might have linked NO_COLOR work items which, if left 1283 * on the delayed_list, will confuse pwq->nr_active 1284 * management later on and cause stall. Make sure the work 1285 * item is activated before grabbing. 1286 */ 1287 if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1288 pwq_activate_delayed_work(work); 1289 1290 list_del_init(&work->entry); 1291 pwq_dec_nr_in_flight(pwq, get_work_color(work)); 1292 1293 /* work->data points to pwq iff queued, point to pool */ 1294 set_work_pool_and_keep_pending(work, pool->id); 1295 1296 spin_unlock(&pool->lock); 1297 rcu_read_unlock(); 1298 return 1; 1299 } 1300 spin_unlock(&pool->lock); 1301 fail: 1302 rcu_read_unlock(); 1303 local_irq_restore(*flags); 1304 if (work_is_canceling(work)) 1305 return -ENOENT; 1306 cpu_relax(); 1307 return -EAGAIN; 1308 } 1309 1310 /** 1311 * insert_work - insert a work into a pool 1312 * @pwq: pwq @work belongs to 1313 * @work: work to insert 1314 * @head: insertion point 1315 * @extra_flags: extra WORK_STRUCT_* flags to set 1316 * 1317 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1318 * work_struct flags. 1319 * 1320 * CONTEXT: 1321 * spin_lock_irq(pool->lock). 1322 */ 1323 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1324 struct list_head *head, unsigned int extra_flags) 1325 { 1326 struct worker_pool *pool = pwq->pool; 1327 1328 /* we own @work, set data and link */ 1329 set_work_pwq(work, pwq, extra_flags); 1330 list_add_tail(&work->entry, head); 1331 get_pwq(pwq); 1332 1333 /* 1334 * Ensure either wq_worker_sleeping() sees the above 1335 * list_add_tail() or we see zero nr_running to avoid workers lying 1336 * around lazily while there are works to be processed. 1337 */ 1338 smp_mb(); 1339 1340 if (__need_more_worker(pool)) 1341 wake_up_worker(pool); 1342 } 1343 1344 /* 1345 * Test whether @work is being queued from another work executing on the 1346 * same workqueue. 1347 */ 1348 static bool is_chained_work(struct workqueue_struct *wq) 1349 { 1350 struct worker *worker; 1351 1352 worker = current_wq_worker(); 1353 /* 1354 * Return %true iff I'm a worker executing a work item on @wq. If 1355 * I'm @worker, it's safe to dereference it without locking. 1356 */ 1357 return worker && worker->current_pwq->wq == wq; 1358 } 1359 1360 /* 1361 * When queueing an unbound work item to a wq, prefer local CPU if allowed 1362 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1363 * avoid perturbing sensitive tasks. 1364 */ 1365 static int wq_select_unbound_cpu(int cpu) 1366 { 1367 static bool printed_dbg_warning; 1368 int new_cpu; 1369 1370 if (likely(!wq_debug_force_rr_cpu)) { 1371 if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1372 return cpu; 1373 } else if (!printed_dbg_warning) { 1374 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1375 printed_dbg_warning = true; 1376 } 1377 1378 if (cpumask_empty(wq_unbound_cpumask)) 1379 return cpu; 1380 1381 new_cpu = __this_cpu_read(wq_rr_cpu_last); 1382 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1383 if (unlikely(new_cpu >= nr_cpu_ids)) { 1384 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1385 if (unlikely(new_cpu >= nr_cpu_ids)) 1386 return cpu; 1387 } 1388 __this_cpu_write(wq_rr_cpu_last, new_cpu); 1389 1390 return new_cpu; 1391 } 1392 1393 static void __queue_work(int cpu, struct workqueue_struct *wq, 1394 struct work_struct *work) 1395 { 1396 struct pool_workqueue *pwq; 1397 struct worker_pool *last_pool; 1398 struct list_head *worklist; 1399 unsigned int work_flags; 1400 unsigned int req_cpu = cpu; 1401 1402 /* 1403 * While a work item is PENDING && off queue, a task trying to 1404 * steal the PENDING will busy-loop waiting for it to either get 1405 * queued or lose PENDING. Grabbing PENDING and queueing should 1406 * happen with IRQ disabled. 1407 */ 1408 lockdep_assert_irqs_disabled(); 1409 1410 debug_work_activate(work); 1411 1412 /* if draining, only works from the same workqueue are allowed */ 1413 if (unlikely(wq->flags & __WQ_DRAINING) && 1414 WARN_ON_ONCE(!is_chained_work(wq))) 1415 return; 1416 rcu_read_lock(); 1417 retry: 1418 if (req_cpu == WORK_CPU_UNBOUND) 1419 cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1420 1421 /* pwq which will be used unless @work is executing elsewhere */ 1422 if (!(wq->flags & WQ_UNBOUND)) 1423 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1424 else 1425 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1426 1427 /* 1428 * If @work was previously on a different pool, it might still be 1429 * running there, in which case the work needs to be queued on that 1430 * pool to guarantee non-reentrancy. 1431 */ 1432 last_pool = get_work_pool(work); 1433 if (last_pool && last_pool != pwq->pool) { 1434 struct worker *worker; 1435 1436 spin_lock(&last_pool->lock); 1437 1438 worker = find_worker_executing_work(last_pool, work); 1439 1440 if (worker && worker->current_pwq->wq == wq) { 1441 pwq = worker->current_pwq; 1442 } else { 1443 /* meh... not running there, queue here */ 1444 spin_unlock(&last_pool->lock); 1445 spin_lock(&pwq->pool->lock); 1446 } 1447 } else { 1448 spin_lock(&pwq->pool->lock); 1449 } 1450 1451 /* 1452 * pwq is determined and locked. For unbound pools, we could have 1453 * raced with pwq release and it could already be dead. If its 1454 * refcnt is zero, repeat pwq selection. Note that pwqs never die 1455 * without another pwq replacing it in the numa_pwq_tbl or while 1456 * work items are executing on it, so the retrying is guaranteed to 1457 * make forward-progress. 1458 */ 1459 if (unlikely(!pwq->refcnt)) { 1460 if (wq->flags & WQ_UNBOUND) { 1461 spin_unlock(&pwq->pool->lock); 1462 cpu_relax(); 1463 goto retry; 1464 } 1465 /* oops */ 1466 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 1467 wq->name, cpu); 1468 } 1469 1470 /* pwq determined, queue */ 1471 trace_workqueue_queue_work(req_cpu, pwq, work); 1472 1473 if (WARN_ON(!list_empty(&work->entry))) 1474 goto out; 1475 1476 pwq->nr_in_flight[pwq->work_color]++; 1477 work_flags = work_color_to_flags(pwq->work_color); 1478 1479 if (likely(pwq->nr_active < pwq->max_active)) { 1480 trace_workqueue_activate_work(work); 1481 pwq->nr_active++; 1482 worklist = &pwq->pool->worklist; 1483 if (list_empty(worklist)) 1484 pwq->pool->watchdog_ts = jiffies; 1485 } else { 1486 work_flags |= WORK_STRUCT_DELAYED; 1487 worklist = &pwq->delayed_works; 1488 } 1489 1490 insert_work(pwq, work, worklist, work_flags); 1491 1492 out: 1493 spin_unlock(&pwq->pool->lock); 1494 rcu_read_unlock(); 1495 } 1496 1497 /** 1498 * queue_work_on - queue work on specific cpu 1499 * @cpu: CPU number to execute work on 1500 * @wq: workqueue to use 1501 * @work: work to queue 1502 * 1503 * We queue the work to a specific CPU, the caller must ensure it 1504 * can't go away. 1505 * 1506 * Return: %false if @work was already on a queue, %true otherwise. 1507 */ 1508 bool queue_work_on(int cpu, struct workqueue_struct *wq, 1509 struct work_struct *work) 1510 { 1511 bool ret = false; 1512 unsigned long flags; 1513 1514 local_irq_save(flags); 1515 1516 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1517 __queue_work(cpu, wq, work); 1518 ret = true; 1519 } 1520 1521 local_irq_restore(flags); 1522 return ret; 1523 } 1524 EXPORT_SYMBOL(queue_work_on); 1525 1526 /** 1527 * workqueue_select_cpu_near - Select a CPU based on NUMA node 1528 * @node: NUMA node ID that we want to select a CPU from 1529 * 1530 * This function will attempt to find a "random" cpu available on a given 1531 * node. If there are no CPUs available on the given node it will return 1532 * WORK_CPU_UNBOUND indicating that we should just schedule to any 1533 * available CPU if we need to schedule this work. 1534 */ 1535 static int workqueue_select_cpu_near(int node) 1536 { 1537 int cpu; 1538 1539 /* No point in doing this if NUMA isn't enabled for workqueues */ 1540 if (!wq_numa_enabled) 1541 return WORK_CPU_UNBOUND; 1542 1543 /* Delay binding to CPU if node is not valid or online */ 1544 if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 1545 return WORK_CPU_UNBOUND; 1546 1547 /* Use local node/cpu if we are already there */ 1548 cpu = raw_smp_processor_id(); 1549 if (node == cpu_to_node(cpu)) 1550 return cpu; 1551 1552 /* Use "random" otherwise know as "first" online CPU of node */ 1553 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 1554 1555 /* If CPU is valid return that, otherwise just defer */ 1556 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 1557 } 1558 1559 /** 1560 * queue_work_node - queue work on a "random" cpu for a given NUMA node 1561 * @node: NUMA node that we are targeting the work for 1562 * @wq: workqueue to use 1563 * @work: work to queue 1564 * 1565 * We queue the work to a "random" CPU within a given NUMA node. The basic 1566 * idea here is to provide a way to somehow associate work with a given 1567 * NUMA node. 1568 * 1569 * This function will only make a best effort attempt at getting this onto 1570 * the right NUMA node. If no node is requested or the requested node is 1571 * offline then we just fall back to standard queue_work behavior. 1572 * 1573 * Currently the "random" CPU ends up being the first available CPU in the 1574 * intersection of cpu_online_mask and the cpumask of the node, unless we 1575 * are running on the node. In that case we just use the current CPU. 1576 * 1577 * Return: %false if @work was already on a queue, %true otherwise. 1578 */ 1579 bool queue_work_node(int node, struct workqueue_struct *wq, 1580 struct work_struct *work) 1581 { 1582 unsigned long flags; 1583 bool ret = false; 1584 1585 /* 1586 * This current implementation is specific to unbound workqueues. 1587 * Specifically we only return the first available CPU for a given 1588 * node instead of cycling through individual CPUs within the node. 1589 * 1590 * If this is used with a per-cpu workqueue then the logic in 1591 * workqueue_select_cpu_near would need to be updated to allow for 1592 * some round robin type logic. 1593 */ 1594 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 1595 1596 local_irq_save(flags); 1597 1598 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1599 int cpu = workqueue_select_cpu_near(node); 1600 1601 __queue_work(cpu, wq, work); 1602 ret = true; 1603 } 1604 1605 local_irq_restore(flags); 1606 return ret; 1607 } 1608 EXPORT_SYMBOL_GPL(queue_work_node); 1609 1610 void delayed_work_timer_fn(struct timer_list *t) 1611 { 1612 struct delayed_work *dwork = from_timer(dwork, t, timer); 1613 1614 /* should have been called from irqsafe timer with irq already off */ 1615 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 1616 } 1617 EXPORT_SYMBOL(delayed_work_timer_fn); 1618 1619 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 1620 struct delayed_work *dwork, unsigned long delay) 1621 { 1622 struct timer_list *timer = &dwork->timer; 1623 struct work_struct *work = &dwork->work; 1624 1625 WARN_ON_ONCE(!wq); 1626 WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1627 WARN_ON_ONCE(timer_pending(timer)); 1628 WARN_ON_ONCE(!list_empty(&work->entry)); 1629 1630 /* 1631 * If @delay is 0, queue @dwork->work immediately. This is for 1632 * both optimization and correctness. The earliest @timer can 1633 * expire is on the closest next tick and delayed_work users depend 1634 * on that there's no such delay when @delay is 0. 1635 */ 1636 if (!delay) { 1637 __queue_work(cpu, wq, &dwork->work); 1638 return; 1639 } 1640 1641 dwork->wq = wq; 1642 dwork->cpu = cpu; 1643 timer->expires = jiffies + delay; 1644 1645 if (unlikely(cpu != WORK_CPU_UNBOUND)) 1646 add_timer_on(timer, cpu); 1647 else 1648 add_timer(timer); 1649 } 1650 1651 /** 1652 * queue_delayed_work_on - queue work on specific CPU after delay 1653 * @cpu: CPU number to execute work on 1654 * @wq: workqueue to use 1655 * @dwork: work to queue 1656 * @delay: number of jiffies to wait before queueing 1657 * 1658 * Return: %false if @work was already on a queue, %true otherwise. If 1659 * @delay is zero and @dwork is idle, it will be scheduled for immediate 1660 * execution. 1661 */ 1662 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 1663 struct delayed_work *dwork, unsigned long delay) 1664 { 1665 struct work_struct *work = &dwork->work; 1666 bool ret = false; 1667 unsigned long flags; 1668 1669 /* read the comment in __queue_work() */ 1670 local_irq_save(flags); 1671 1672 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1673 __queue_delayed_work(cpu, wq, dwork, delay); 1674 ret = true; 1675 } 1676 1677 local_irq_restore(flags); 1678 return ret; 1679 } 1680 EXPORT_SYMBOL(queue_delayed_work_on); 1681 1682 /** 1683 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 1684 * @cpu: CPU number to execute work on 1685 * @wq: workqueue to use 1686 * @dwork: work to queue 1687 * @delay: number of jiffies to wait before queueing 1688 * 1689 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 1690 * modify @dwork's timer so that it expires after @delay. If @delay is 1691 * zero, @work is guaranteed to be scheduled immediately regardless of its 1692 * current state. 1693 * 1694 * Return: %false if @dwork was idle and queued, %true if @dwork was 1695 * pending and its timer was modified. 1696 * 1697 * This function is safe to call from any context including IRQ handler. 1698 * See try_to_grab_pending() for details. 1699 */ 1700 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 1701 struct delayed_work *dwork, unsigned long delay) 1702 { 1703 unsigned long flags; 1704 int ret; 1705 1706 do { 1707 ret = try_to_grab_pending(&dwork->work, true, &flags); 1708 } while (unlikely(ret == -EAGAIN)); 1709 1710 if (likely(ret >= 0)) { 1711 __queue_delayed_work(cpu, wq, dwork, delay); 1712 local_irq_restore(flags); 1713 } 1714 1715 /* -ENOENT from try_to_grab_pending() becomes %true */ 1716 return ret; 1717 } 1718 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 1719 1720 static void rcu_work_rcufn(struct rcu_head *rcu) 1721 { 1722 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 1723 1724 /* read the comment in __queue_work() */ 1725 local_irq_disable(); 1726 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 1727 local_irq_enable(); 1728 } 1729 1730 /** 1731 * queue_rcu_work - queue work after a RCU grace period 1732 * @wq: workqueue to use 1733 * @rwork: work to queue 1734 * 1735 * Return: %false if @rwork was already pending, %true otherwise. Note 1736 * that a full RCU grace period is guaranteed only after a %true return. 1737 * While @rwork is guaranteed to be executed after a %false return, the 1738 * execution may happen before a full RCU grace period has passed. 1739 */ 1740 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 1741 { 1742 struct work_struct *work = &rwork->work; 1743 1744 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1745 rwork->wq = wq; 1746 call_rcu(&rwork->rcu, rcu_work_rcufn); 1747 return true; 1748 } 1749 1750 return false; 1751 } 1752 EXPORT_SYMBOL(queue_rcu_work); 1753 1754 /** 1755 * worker_enter_idle - enter idle state 1756 * @worker: worker which is entering idle state 1757 * 1758 * @worker is entering idle state. Update stats and idle timer if 1759 * necessary. 1760 * 1761 * LOCKING: 1762 * spin_lock_irq(pool->lock). 1763 */ 1764 static void worker_enter_idle(struct worker *worker) 1765 { 1766 struct worker_pool *pool = worker->pool; 1767 1768 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 1769 WARN_ON_ONCE(!list_empty(&worker->entry) && 1770 (worker->hentry.next || worker->hentry.pprev))) 1771 return; 1772 1773 /* can't use worker_set_flags(), also called from create_worker() */ 1774 worker->flags |= WORKER_IDLE; 1775 pool->nr_idle++; 1776 worker->last_active = jiffies; 1777 1778 /* idle_list is LIFO */ 1779 list_add(&worker->entry, &pool->idle_list); 1780 1781 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1782 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1783 1784 /* 1785 * Sanity check nr_running. Because unbind_workers() releases 1786 * pool->lock between setting %WORKER_UNBOUND and zapping 1787 * nr_running, the warning may trigger spuriously. Check iff 1788 * unbind is not in progress. 1789 */ 1790 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1791 pool->nr_workers == pool->nr_idle && 1792 atomic_read(&pool->nr_running)); 1793 } 1794 1795 /** 1796 * worker_leave_idle - leave idle state 1797 * @worker: worker which is leaving idle state 1798 * 1799 * @worker is leaving idle state. Update stats. 1800 * 1801 * LOCKING: 1802 * spin_lock_irq(pool->lock). 1803 */ 1804 static void worker_leave_idle(struct worker *worker) 1805 { 1806 struct worker_pool *pool = worker->pool; 1807 1808 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 1809 return; 1810 worker_clr_flags(worker, WORKER_IDLE); 1811 pool->nr_idle--; 1812 list_del_init(&worker->entry); 1813 } 1814 1815 static struct worker *alloc_worker(int node) 1816 { 1817 struct worker *worker; 1818 1819 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1820 if (worker) { 1821 INIT_LIST_HEAD(&worker->entry); 1822 INIT_LIST_HEAD(&worker->scheduled); 1823 INIT_LIST_HEAD(&worker->node); 1824 /* on creation a worker is in !idle && prep state */ 1825 worker->flags = WORKER_PREP; 1826 } 1827 return worker; 1828 } 1829 1830 /** 1831 * worker_attach_to_pool() - attach a worker to a pool 1832 * @worker: worker to be attached 1833 * @pool: the target pool 1834 * 1835 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 1836 * cpu-binding of @worker are kept coordinated with the pool across 1837 * cpu-[un]hotplugs. 1838 */ 1839 static void worker_attach_to_pool(struct worker *worker, 1840 struct worker_pool *pool) 1841 { 1842 mutex_lock(&wq_pool_attach_mutex); 1843 1844 /* 1845 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1846 * online CPUs. It'll be re-applied when any of the CPUs come up. 1847 */ 1848 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 1849 1850 /* 1851 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 1852 * stable across this function. See the comments above the flag 1853 * definition for details. 1854 */ 1855 if (pool->flags & POOL_DISASSOCIATED) 1856 worker->flags |= WORKER_UNBOUND; 1857 1858 list_add_tail(&worker->node, &pool->workers); 1859 worker->pool = pool; 1860 1861 mutex_unlock(&wq_pool_attach_mutex); 1862 } 1863 1864 /** 1865 * worker_detach_from_pool() - detach a worker from its pool 1866 * @worker: worker which is attached to its pool 1867 * 1868 * Undo the attaching which had been done in worker_attach_to_pool(). The 1869 * caller worker shouldn't access to the pool after detached except it has 1870 * other reference to the pool. 1871 */ 1872 static void worker_detach_from_pool(struct worker *worker) 1873 { 1874 struct worker_pool *pool = worker->pool; 1875 struct completion *detach_completion = NULL; 1876 1877 mutex_lock(&wq_pool_attach_mutex); 1878 1879 list_del(&worker->node); 1880 worker->pool = NULL; 1881 1882 if (list_empty(&pool->workers)) 1883 detach_completion = pool->detach_completion; 1884 mutex_unlock(&wq_pool_attach_mutex); 1885 1886 /* clear leftover flags without pool->lock after it is detached */ 1887 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1888 1889 if (detach_completion) 1890 complete(detach_completion); 1891 } 1892 1893 /** 1894 * create_worker - create a new workqueue worker 1895 * @pool: pool the new worker will belong to 1896 * 1897 * Create and start a new worker which is attached to @pool. 1898 * 1899 * CONTEXT: 1900 * Might sleep. Does GFP_KERNEL allocations. 1901 * 1902 * Return: 1903 * Pointer to the newly created worker. 1904 */ 1905 static struct worker *create_worker(struct worker_pool *pool) 1906 { 1907 struct worker *worker = NULL; 1908 int id = -1; 1909 char id_buf[16]; 1910 1911 /* ID is needed to determine kthread name */ 1912 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1913 if (id < 0) 1914 goto fail; 1915 1916 worker = alloc_worker(pool->node); 1917 if (!worker) 1918 goto fail; 1919 1920 worker->id = id; 1921 1922 if (pool->cpu >= 0) 1923 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1924 pool->attrs->nice < 0 ? "H" : ""); 1925 else 1926 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1927 1928 worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1929 "kworker/%s", id_buf); 1930 if (IS_ERR(worker->task)) 1931 goto fail; 1932 1933 set_user_nice(worker->task, pool->attrs->nice); 1934 kthread_bind_mask(worker->task, pool->attrs->cpumask); 1935 1936 /* successful, attach the worker to the pool */ 1937 worker_attach_to_pool(worker, pool); 1938 1939 /* start the newly created worker */ 1940 spin_lock_irq(&pool->lock); 1941 worker->pool->nr_workers++; 1942 worker_enter_idle(worker); 1943 wake_up_process(worker->task); 1944 spin_unlock_irq(&pool->lock); 1945 1946 return worker; 1947 1948 fail: 1949 if (id >= 0) 1950 ida_simple_remove(&pool->worker_ida, id); 1951 kfree(worker); 1952 return NULL; 1953 } 1954 1955 /** 1956 * destroy_worker - destroy a workqueue worker 1957 * @worker: worker to be destroyed 1958 * 1959 * Destroy @worker and adjust @pool stats accordingly. The worker should 1960 * be idle. 1961 * 1962 * CONTEXT: 1963 * spin_lock_irq(pool->lock). 1964 */ 1965 static void destroy_worker(struct worker *worker) 1966 { 1967 struct worker_pool *pool = worker->pool; 1968 1969 lockdep_assert_held(&pool->lock); 1970 1971 /* sanity check frenzy */ 1972 if (WARN_ON(worker->current_work) || 1973 WARN_ON(!list_empty(&worker->scheduled)) || 1974 WARN_ON(!(worker->flags & WORKER_IDLE))) 1975 return; 1976 1977 pool->nr_workers--; 1978 pool->nr_idle--; 1979 1980 list_del_init(&worker->entry); 1981 worker->flags |= WORKER_DIE; 1982 wake_up_process(worker->task); 1983 } 1984 1985 static void idle_worker_timeout(struct timer_list *t) 1986 { 1987 struct worker_pool *pool = from_timer(pool, t, idle_timer); 1988 1989 spin_lock_irq(&pool->lock); 1990 1991 while (too_many_workers(pool)) { 1992 struct worker *worker; 1993 unsigned long expires; 1994 1995 /* idle_list is kept in LIFO order, check the last one */ 1996 worker = list_entry(pool->idle_list.prev, struct worker, entry); 1997 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1998 1999 if (time_before(jiffies, expires)) { 2000 mod_timer(&pool->idle_timer, expires); 2001 break; 2002 } 2003 2004 destroy_worker(worker); 2005 } 2006 2007 spin_unlock_irq(&pool->lock); 2008 } 2009 2010 static void send_mayday(struct work_struct *work) 2011 { 2012 struct pool_workqueue *pwq = get_work_pwq(work); 2013 struct workqueue_struct *wq = pwq->wq; 2014 2015 lockdep_assert_held(&wq_mayday_lock); 2016 2017 if (!wq->rescuer) 2018 return; 2019 2020 /* mayday mayday mayday */ 2021 if (list_empty(&pwq->mayday_node)) { 2022 /* 2023 * If @pwq is for an unbound wq, its base ref may be put at 2024 * any time due to an attribute change. Pin @pwq until the 2025 * rescuer is done with it. 2026 */ 2027 get_pwq(pwq); 2028 list_add_tail(&pwq->mayday_node, &wq->maydays); 2029 wake_up_process(wq->rescuer->task); 2030 } 2031 } 2032 2033 static void pool_mayday_timeout(struct timer_list *t) 2034 { 2035 struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2036 struct work_struct *work; 2037 2038 spin_lock_irq(&pool->lock); 2039 spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2040 2041 if (need_to_create_worker(pool)) { 2042 /* 2043 * We've been trying to create a new worker but 2044 * haven't been successful. We might be hitting an 2045 * allocation deadlock. Send distress signals to 2046 * rescuers. 2047 */ 2048 list_for_each_entry(work, &pool->worklist, entry) 2049 send_mayday(work); 2050 } 2051 2052 spin_unlock(&wq_mayday_lock); 2053 spin_unlock_irq(&pool->lock); 2054 2055 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2056 } 2057 2058 /** 2059 * maybe_create_worker - create a new worker if necessary 2060 * @pool: pool to create a new worker for 2061 * 2062 * Create a new worker for @pool if necessary. @pool is guaranteed to 2063 * have at least one idle worker on return from this function. If 2064 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 2065 * sent to all rescuers with works scheduled on @pool to resolve 2066 * possible allocation deadlock. 2067 * 2068 * On return, need_to_create_worker() is guaranteed to be %false and 2069 * may_start_working() %true. 2070 * 2071 * LOCKING: 2072 * spin_lock_irq(pool->lock) which may be released and regrabbed 2073 * multiple times. Does GFP_KERNEL allocations. Called only from 2074 * manager. 2075 */ 2076 static void maybe_create_worker(struct worker_pool *pool) 2077 __releases(&pool->lock) 2078 __acquires(&pool->lock) 2079 { 2080 restart: 2081 spin_unlock_irq(&pool->lock); 2082 2083 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 2084 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2085 2086 while (true) { 2087 if (create_worker(pool) || !need_to_create_worker(pool)) 2088 break; 2089 2090 schedule_timeout_interruptible(CREATE_COOLDOWN); 2091 2092 if (!need_to_create_worker(pool)) 2093 break; 2094 } 2095 2096 del_timer_sync(&pool->mayday_timer); 2097 spin_lock_irq(&pool->lock); 2098 /* 2099 * This is necessary even after a new worker was just successfully 2100 * created as @pool->lock was dropped and the new worker might have 2101 * already become busy. 2102 */ 2103 if (need_to_create_worker(pool)) 2104 goto restart; 2105 } 2106 2107 /** 2108 * manage_workers - manage worker pool 2109 * @worker: self 2110 * 2111 * Assume the manager role and manage the worker pool @worker belongs 2112 * to. At any given time, there can be only zero or one manager per 2113 * pool. The exclusion is handled automatically by this function. 2114 * 2115 * The caller can safely start processing works on false return. On 2116 * true return, it's guaranteed that need_to_create_worker() is false 2117 * and may_start_working() is true. 2118 * 2119 * CONTEXT: 2120 * spin_lock_irq(pool->lock) which may be released and regrabbed 2121 * multiple times. Does GFP_KERNEL allocations. 2122 * 2123 * Return: 2124 * %false if the pool doesn't need management and the caller can safely 2125 * start processing works, %true if management function was performed and 2126 * the conditions that the caller verified before calling the function may 2127 * no longer be true. 2128 */ 2129 static bool manage_workers(struct worker *worker) 2130 { 2131 struct worker_pool *pool = worker->pool; 2132 2133 if (pool->flags & POOL_MANAGER_ACTIVE) 2134 return false; 2135 2136 pool->flags |= POOL_MANAGER_ACTIVE; 2137 pool->manager = worker; 2138 2139 maybe_create_worker(pool); 2140 2141 pool->manager = NULL; 2142 pool->flags &= ~POOL_MANAGER_ACTIVE; 2143 wake_up(&wq_manager_wait); 2144 return true; 2145 } 2146 2147 /** 2148 * process_one_work - process single work 2149 * @worker: self 2150 * @work: work to process 2151 * 2152 * Process @work. This function contains all the logics necessary to 2153 * process a single work including synchronization against and 2154 * interaction with other workers on the same cpu, queueing and 2155 * flushing. As long as context requirement is met, any worker can 2156 * call this function to process a work. 2157 * 2158 * CONTEXT: 2159 * spin_lock_irq(pool->lock) which is released and regrabbed. 2160 */ 2161 static void process_one_work(struct worker *worker, struct work_struct *work) 2162 __releases(&pool->lock) 2163 __acquires(&pool->lock) 2164 { 2165 struct pool_workqueue *pwq = get_work_pwq(work); 2166 struct worker_pool *pool = worker->pool; 2167 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 2168 int work_color; 2169 struct worker *collision; 2170 #ifdef CONFIG_LOCKDEP 2171 /* 2172 * It is permissible to free the struct work_struct from 2173 * inside the function that is called from it, this we need to 2174 * take into account for lockdep too. To avoid bogus "held 2175 * lock freed" warnings as well as problems when looking into 2176 * work->lockdep_map, make a copy and use that here. 2177 */ 2178 struct lockdep_map lockdep_map; 2179 2180 lockdep_copy_map(&lockdep_map, &work->lockdep_map); 2181 #endif 2182 /* ensure we're on the correct CPU */ 2183 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2184 raw_smp_processor_id() != pool->cpu); 2185 2186 /* 2187 * A single work shouldn't be executed concurrently by 2188 * multiple workers on a single cpu. Check whether anyone is 2189 * already processing the work. If so, defer the work to the 2190 * currently executing one. 2191 */ 2192 collision = find_worker_executing_work(pool, work); 2193 if (unlikely(collision)) { 2194 move_linked_works(work, &collision->scheduled, NULL); 2195 return; 2196 } 2197 2198 /* claim and dequeue */ 2199 debug_work_deactivate(work); 2200 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2201 worker->current_work = work; 2202 worker->current_func = work->func; 2203 worker->current_pwq = pwq; 2204 work_color = get_work_color(work); 2205 2206 /* 2207 * Record wq name for cmdline and debug reporting, may get 2208 * overridden through set_worker_desc(). 2209 */ 2210 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 2211 2212 list_del_init(&work->entry); 2213 2214 /* 2215 * CPU intensive works don't participate in concurrency management. 2216 * They're the scheduler's responsibility. This takes @worker out 2217 * of concurrency management and the next code block will chain 2218 * execution of the pending work items. 2219 */ 2220 if (unlikely(cpu_intensive)) 2221 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2222 2223 /* 2224 * Wake up another worker if necessary. The condition is always 2225 * false for normal per-cpu workers since nr_running would always 2226 * be >= 1 at this point. This is used to chain execution of the 2227 * pending work items for WORKER_NOT_RUNNING workers such as the 2228 * UNBOUND and CPU_INTENSIVE ones. 2229 */ 2230 if (need_more_worker(pool)) 2231 wake_up_worker(pool); 2232 2233 /* 2234 * Record the last pool and clear PENDING which should be the last 2235 * update to @work. Also, do this inside @pool->lock so that 2236 * PENDING and queued state changes happen together while IRQ is 2237 * disabled. 2238 */ 2239 set_work_pool_and_clear_pending(work, pool->id); 2240 2241 spin_unlock_irq(&pool->lock); 2242 2243 lock_map_acquire(&pwq->wq->lockdep_map); 2244 lock_map_acquire(&lockdep_map); 2245 /* 2246 * Strictly speaking we should mark the invariant state without holding 2247 * any locks, that is, before these two lock_map_acquire()'s. 2248 * 2249 * However, that would result in: 2250 * 2251 * A(W1) 2252 * WFC(C) 2253 * A(W1) 2254 * C(C) 2255 * 2256 * Which would create W1->C->W1 dependencies, even though there is no 2257 * actual deadlock possible. There are two solutions, using a 2258 * read-recursive acquire on the work(queue) 'locks', but this will then 2259 * hit the lockdep limitation on recursive locks, or simply discard 2260 * these locks. 2261 * 2262 * AFAICT there is no possible deadlock scenario between the 2263 * flush_work() and complete() primitives (except for single-threaded 2264 * workqueues), so hiding them isn't a problem. 2265 */ 2266 lockdep_invariant_state(true); 2267 trace_workqueue_execute_start(work); 2268 worker->current_func(work); 2269 /* 2270 * While we must be careful to not use "work" after this, the trace 2271 * point will only record its address. 2272 */ 2273 trace_workqueue_execute_end(work); 2274 lock_map_release(&lockdep_map); 2275 lock_map_release(&pwq->wq->lockdep_map); 2276 2277 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2278 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2279 " last function: %ps\n", 2280 current->comm, preempt_count(), task_pid_nr(current), 2281 worker->current_func); 2282 debug_show_held_locks(current); 2283 dump_stack(); 2284 } 2285 2286 /* 2287 * The following prevents a kworker from hogging CPU on !PREEMPT 2288 * kernels, where a requeueing work item waiting for something to 2289 * happen could deadlock with stop_machine as such work item could 2290 * indefinitely requeue itself while all other CPUs are trapped in 2291 * stop_machine. At the same time, report a quiescent RCU state so 2292 * the same condition doesn't freeze RCU. 2293 */ 2294 cond_resched(); 2295 2296 spin_lock_irq(&pool->lock); 2297 2298 /* clear cpu intensive status */ 2299 if (unlikely(cpu_intensive)) 2300 worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2301 2302 /* tag the worker for identification in schedule() */ 2303 worker->last_func = worker->current_func; 2304 2305 /* we're done with it, release */ 2306 hash_del(&worker->hentry); 2307 worker->current_work = NULL; 2308 worker->current_func = NULL; 2309 worker->current_pwq = NULL; 2310 pwq_dec_nr_in_flight(pwq, work_color); 2311 } 2312 2313 /** 2314 * process_scheduled_works - process scheduled works 2315 * @worker: self 2316 * 2317 * Process all scheduled works. Please note that the scheduled list 2318 * may change while processing a work, so this function repeatedly 2319 * fetches a work from the top and executes it. 2320 * 2321 * CONTEXT: 2322 * spin_lock_irq(pool->lock) which may be released and regrabbed 2323 * multiple times. 2324 */ 2325 static void process_scheduled_works(struct worker *worker) 2326 { 2327 while (!list_empty(&worker->scheduled)) { 2328 struct work_struct *work = list_first_entry(&worker->scheduled, 2329 struct work_struct, entry); 2330 process_one_work(worker, work); 2331 } 2332 } 2333 2334 static void set_pf_worker(bool val) 2335 { 2336 mutex_lock(&wq_pool_attach_mutex); 2337 if (val) 2338 current->flags |= PF_WQ_WORKER; 2339 else 2340 current->flags &= ~PF_WQ_WORKER; 2341 mutex_unlock(&wq_pool_attach_mutex); 2342 } 2343 2344 /** 2345 * worker_thread - the worker thread function 2346 * @__worker: self 2347 * 2348 * The worker thread function. All workers belong to a worker_pool - 2349 * either a per-cpu one or dynamic unbound one. These workers process all 2350 * work items regardless of their specific target workqueue. The only 2351 * exception is work items which belong to workqueues with a rescuer which 2352 * will be explained in rescuer_thread(). 2353 * 2354 * Return: 0 2355 */ 2356 static int worker_thread(void *__worker) 2357 { 2358 struct worker *worker = __worker; 2359 struct worker_pool *pool = worker->pool; 2360 2361 /* tell the scheduler that this is a workqueue worker */ 2362 set_pf_worker(true); 2363 woke_up: 2364 spin_lock_irq(&pool->lock); 2365 2366 /* am I supposed to die? */ 2367 if (unlikely(worker->flags & WORKER_DIE)) { 2368 spin_unlock_irq(&pool->lock); 2369 WARN_ON_ONCE(!list_empty(&worker->entry)); 2370 set_pf_worker(false); 2371 2372 set_task_comm(worker->task, "kworker/dying"); 2373 ida_simple_remove(&pool->worker_ida, worker->id); 2374 worker_detach_from_pool(worker); 2375 kfree(worker); 2376 return 0; 2377 } 2378 2379 worker_leave_idle(worker); 2380 recheck: 2381 /* no more worker necessary? */ 2382 if (!need_more_worker(pool)) 2383 goto sleep; 2384 2385 /* do we need to manage? */ 2386 if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2387 goto recheck; 2388 2389 /* 2390 * ->scheduled list can only be filled while a worker is 2391 * preparing to process a work or actually processing it. 2392 * Make sure nobody diddled with it while I was sleeping. 2393 */ 2394 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2395 2396 /* 2397 * Finish PREP stage. We're guaranteed to have at least one idle 2398 * worker or that someone else has already assumed the manager 2399 * role. This is where @worker starts participating in concurrency 2400 * management if applicable and concurrency management is restored 2401 * after being rebound. See rebind_workers() for details. 2402 */ 2403 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2404 2405 do { 2406 struct work_struct *work = 2407 list_first_entry(&pool->worklist, 2408 struct work_struct, entry); 2409 2410 pool->watchdog_ts = jiffies; 2411 2412 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2413 /* optimization path, not strictly necessary */ 2414 process_one_work(worker, work); 2415 if (unlikely(!list_empty(&worker->scheduled))) 2416 process_scheduled_works(worker); 2417 } else { 2418 move_linked_works(work, &worker->scheduled, NULL); 2419 process_scheduled_works(worker); 2420 } 2421 } while (keep_working(pool)); 2422 2423 worker_set_flags(worker, WORKER_PREP); 2424 sleep: 2425 /* 2426 * pool->lock is held and there's no work to process and no need to 2427 * manage, sleep. Workers are woken up only while holding 2428 * pool->lock or from local cpu, so setting the current state 2429 * before releasing pool->lock is enough to prevent losing any 2430 * event. 2431 */ 2432 worker_enter_idle(worker); 2433 __set_current_state(TASK_IDLE); 2434 spin_unlock_irq(&pool->lock); 2435 schedule(); 2436 goto woke_up; 2437 } 2438 2439 /** 2440 * rescuer_thread - the rescuer thread function 2441 * @__rescuer: self 2442 * 2443 * Workqueue rescuer thread function. There's one rescuer for each 2444 * workqueue which has WQ_MEM_RECLAIM set. 2445 * 2446 * Regular work processing on a pool may block trying to create a new 2447 * worker which uses GFP_KERNEL allocation which has slight chance of 2448 * developing into deadlock if some works currently on the same queue 2449 * need to be processed to satisfy the GFP_KERNEL allocation. This is 2450 * the problem rescuer solves. 2451 * 2452 * When such condition is possible, the pool summons rescuers of all 2453 * workqueues which have works queued on the pool and let them process 2454 * those works so that forward progress can be guaranteed. 2455 * 2456 * This should happen rarely. 2457 * 2458 * Return: 0 2459 */ 2460 static int rescuer_thread(void *__rescuer) 2461 { 2462 struct worker *rescuer = __rescuer; 2463 struct workqueue_struct *wq = rescuer->rescue_wq; 2464 struct list_head *scheduled = &rescuer->scheduled; 2465 bool should_stop; 2466 2467 set_user_nice(current, RESCUER_NICE_LEVEL); 2468 2469 /* 2470 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2471 * doesn't participate in concurrency management. 2472 */ 2473 set_pf_worker(true); 2474 repeat: 2475 set_current_state(TASK_IDLE); 2476 2477 /* 2478 * By the time the rescuer is requested to stop, the workqueue 2479 * shouldn't have any work pending, but @wq->maydays may still have 2480 * pwq(s) queued. This can happen by non-rescuer workers consuming 2481 * all the work items before the rescuer got to them. Go through 2482 * @wq->maydays processing before acting on should_stop so that the 2483 * list is always empty on exit. 2484 */ 2485 should_stop = kthread_should_stop(); 2486 2487 /* see whether any pwq is asking for help */ 2488 spin_lock_irq(&wq_mayday_lock); 2489 2490 while (!list_empty(&wq->maydays)) { 2491 struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2492 struct pool_workqueue, mayday_node); 2493 struct worker_pool *pool = pwq->pool; 2494 struct work_struct *work, *n; 2495 bool first = true; 2496 2497 __set_current_state(TASK_RUNNING); 2498 list_del_init(&pwq->mayday_node); 2499 2500 spin_unlock_irq(&wq_mayday_lock); 2501 2502 worker_attach_to_pool(rescuer, pool); 2503 2504 spin_lock_irq(&pool->lock); 2505 2506 /* 2507 * Slurp in all works issued via this workqueue and 2508 * process'em. 2509 */ 2510 WARN_ON_ONCE(!list_empty(scheduled)); 2511 list_for_each_entry_safe(work, n, &pool->worklist, entry) { 2512 if (get_work_pwq(work) == pwq) { 2513 if (first) 2514 pool->watchdog_ts = jiffies; 2515 move_linked_works(work, scheduled, &n); 2516 } 2517 first = false; 2518 } 2519 2520 if (!list_empty(scheduled)) { 2521 process_scheduled_works(rescuer); 2522 2523 /* 2524 * The above execution of rescued work items could 2525 * have created more to rescue through 2526 * pwq_activate_first_delayed() or chained 2527 * queueing. Let's put @pwq back on mayday list so 2528 * that such back-to-back work items, which may be 2529 * being used to relieve memory pressure, don't 2530 * incur MAYDAY_INTERVAL delay inbetween. 2531 */ 2532 if (need_to_create_worker(pool)) { 2533 spin_lock(&wq_mayday_lock); 2534 get_pwq(pwq); 2535 list_move_tail(&pwq->mayday_node, &wq->maydays); 2536 spin_unlock(&wq_mayday_lock); 2537 } 2538 } 2539 2540 /* 2541 * Put the reference grabbed by send_mayday(). @pool won't 2542 * go away while we're still attached to it. 2543 */ 2544 put_pwq(pwq); 2545 2546 /* 2547 * Leave this pool. If need_more_worker() is %true, notify a 2548 * regular worker; otherwise, we end up with 0 concurrency 2549 * and stalling the execution. 2550 */ 2551 if (need_more_worker(pool)) 2552 wake_up_worker(pool); 2553 2554 spin_unlock_irq(&pool->lock); 2555 2556 worker_detach_from_pool(rescuer); 2557 2558 spin_lock_irq(&wq_mayday_lock); 2559 } 2560 2561 spin_unlock_irq(&wq_mayday_lock); 2562 2563 if (should_stop) { 2564 __set_current_state(TASK_RUNNING); 2565 set_pf_worker(false); 2566 return 0; 2567 } 2568 2569 /* rescuers should never participate in concurrency management */ 2570 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2571 schedule(); 2572 goto repeat; 2573 } 2574 2575 /** 2576 * check_flush_dependency - check for flush dependency sanity 2577 * @target_wq: workqueue being flushed 2578 * @target_work: work item being flushed (NULL for workqueue flushes) 2579 * 2580 * %current is trying to flush the whole @target_wq or @target_work on it. 2581 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2582 * reclaiming memory or running on a workqueue which doesn't have 2583 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2584 * a deadlock. 2585 */ 2586 static void check_flush_dependency(struct workqueue_struct *target_wq, 2587 struct work_struct *target_work) 2588 { 2589 work_func_t target_func = target_work ? target_work->func : NULL; 2590 struct worker *worker; 2591 2592 if (target_wq->flags & WQ_MEM_RECLAIM) 2593 return; 2594 2595 worker = current_wq_worker(); 2596 2597 WARN_ONCE(current->flags & PF_MEMALLOC, 2598 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2599 current->pid, current->comm, target_wq->name, target_func); 2600 WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 2601 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2602 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2603 worker->current_pwq->wq->name, worker->current_func, 2604 target_wq->name, target_func); 2605 } 2606 2607 struct wq_barrier { 2608 struct work_struct work; 2609 struct completion done; 2610 struct task_struct *task; /* purely informational */ 2611 }; 2612 2613 static void wq_barrier_func(struct work_struct *work) 2614 { 2615 struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2616 complete(&barr->done); 2617 } 2618 2619 /** 2620 * insert_wq_barrier - insert a barrier work 2621 * @pwq: pwq to insert barrier into 2622 * @barr: wq_barrier to insert 2623 * @target: target work to attach @barr to 2624 * @worker: worker currently executing @target, NULL if @target is not executing 2625 * 2626 * @barr is linked to @target such that @barr is completed only after 2627 * @target finishes execution. Please note that the ordering 2628 * guarantee is observed only with respect to @target and on the local 2629 * cpu. 2630 * 2631 * Currently, a queued barrier can't be canceled. This is because 2632 * try_to_grab_pending() can't determine whether the work to be 2633 * grabbed is at the head of the queue and thus can't clear LINKED 2634 * flag of the previous work while there must be a valid next work 2635 * after a work with LINKED flag set. 2636 * 2637 * Note that when @worker is non-NULL, @target may be modified 2638 * underneath us, so we can't reliably determine pwq from @target. 2639 * 2640 * CONTEXT: 2641 * spin_lock_irq(pool->lock). 2642 */ 2643 static void insert_wq_barrier(struct pool_workqueue *pwq, 2644 struct wq_barrier *barr, 2645 struct work_struct *target, struct worker *worker) 2646 { 2647 struct list_head *head; 2648 unsigned int linked = 0; 2649 2650 /* 2651 * debugobject calls are safe here even with pool->lock locked 2652 * as we know for sure that this will not trigger any of the 2653 * checks and call back into the fixup functions where we 2654 * might deadlock. 2655 */ 2656 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 2657 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2658 2659 init_completion_map(&barr->done, &target->lockdep_map); 2660 2661 barr->task = current; 2662 2663 /* 2664 * If @target is currently being executed, schedule the 2665 * barrier to the worker; otherwise, put it after @target. 2666 */ 2667 if (worker) 2668 head = worker->scheduled.next; 2669 else { 2670 unsigned long *bits = work_data_bits(target); 2671 2672 head = target->entry.next; 2673 /* there can already be other linked works, inherit and set */ 2674 linked = *bits & WORK_STRUCT_LINKED; 2675 __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2676 } 2677 2678 debug_work_activate(&barr->work); 2679 insert_work(pwq, &barr->work, head, 2680 work_color_to_flags(WORK_NO_COLOR) | linked); 2681 } 2682 2683 /** 2684 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 2685 * @wq: workqueue being flushed 2686 * @flush_color: new flush color, < 0 for no-op 2687 * @work_color: new work color, < 0 for no-op 2688 * 2689 * Prepare pwqs for workqueue flushing. 2690 * 2691 * If @flush_color is non-negative, flush_color on all pwqs should be 2692 * -1. If no pwq has in-flight commands at the specified color, all 2693 * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2694 * has in flight commands, its pwq->flush_color is set to 2695 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 2696 * wakeup logic is armed and %true is returned. 2697 * 2698 * The caller should have initialized @wq->first_flusher prior to 2699 * calling this function with non-negative @flush_color. If 2700 * @flush_color is negative, no flush color update is done and %false 2701 * is returned. 2702 * 2703 * If @work_color is non-negative, all pwqs should have the same 2704 * work_color which is previous to @work_color and all will be 2705 * advanced to @work_color. 2706 * 2707 * CONTEXT: 2708 * mutex_lock(wq->mutex). 2709 * 2710 * Return: 2711 * %true if @flush_color >= 0 and there's something to flush. %false 2712 * otherwise. 2713 */ 2714 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 2715 int flush_color, int work_color) 2716 { 2717 bool wait = false; 2718 struct pool_workqueue *pwq; 2719 2720 if (flush_color >= 0) { 2721 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2722 atomic_set(&wq->nr_pwqs_to_flush, 1); 2723 } 2724 2725 for_each_pwq(pwq, wq) { 2726 struct worker_pool *pool = pwq->pool; 2727 2728 spin_lock_irq(&pool->lock); 2729 2730 if (flush_color >= 0) { 2731 WARN_ON_ONCE(pwq->flush_color != -1); 2732 2733 if (pwq->nr_in_flight[flush_color]) { 2734 pwq->flush_color = flush_color; 2735 atomic_inc(&wq->nr_pwqs_to_flush); 2736 wait = true; 2737 } 2738 } 2739 2740 if (work_color >= 0) { 2741 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2742 pwq->work_color = work_color; 2743 } 2744 2745 spin_unlock_irq(&pool->lock); 2746 } 2747 2748 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 2749 complete(&wq->first_flusher->done); 2750 2751 return wait; 2752 } 2753 2754 /** 2755 * flush_workqueue - ensure that any scheduled work has run to completion. 2756 * @wq: workqueue to flush 2757 * 2758 * This function sleeps until all work items which were queued on entry 2759 * have finished execution, but it is not livelocked by new incoming ones. 2760 */ 2761 void flush_workqueue(struct workqueue_struct *wq) 2762 { 2763 struct wq_flusher this_flusher = { 2764 .list = LIST_HEAD_INIT(this_flusher.list), 2765 .flush_color = -1, 2766 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 2767 }; 2768 int next_color; 2769 2770 if (WARN_ON(!wq_online)) 2771 return; 2772 2773 lock_map_acquire(&wq->lockdep_map); 2774 lock_map_release(&wq->lockdep_map); 2775 2776 mutex_lock(&wq->mutex); 2777 2778 /* 2779 * Start-to-wait phase 2780 */ 2781 next_color = work_next_color(wq->work_color); 2782 2783 if (next_color != wq->flush_color) { 2784 /* 2785 * Color space is not full. The current work_color 2786 * becomes our flush_color and work_color is advanced 2787 * by one. 2788 */ 2789 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 2790 this_flusher.flush_color = wq->work_color; 2791 wq->work_color = next_color; 2792 2793 if (!wq->first_flusher) { 2794 /* no flush in progress, become the first flusher */ 2795 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 2796 2797 wq->first_flusher = &this_flusher; 2798 2799 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 2800 wq->work_color)) { 2801 /* nothing to flush, done */ 2802 wq->flush_color = next_color; 2803 wq->first_flusher = NULL; 2804 goto out_unlock; 2805 } 2806 } else { 2807 /* wait in queue */ 2808 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 2809 list_add_tail(&this_flusher.list, &wq->flusher_queue); 2810 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 2811 } 2812 } else { 2813 /* 2814 * Oops, color space is full, wait on overflow queue. 2815 * The next flush completion will assign us 2816 * flush_color and transfer to flusher_queue. 2817 */ 2818 list_add_tail(&this_flusher.list, &wq->flusher_overflow); 2819 } 2820 2821 check_flush_dependency(wq, NULL); 2822 2823 mutex_unlock(&wq->mutex); 2824 2825 wait_for_completion(&this_flusher.done); 2826 2827 /* 2828 * Wake-up-and-cascade phase 2829 * 2830 * First flushers are responsible for cascading flushes and 2831 * handling overflow. Non-first flushers can simply return. 2832 */ 2833 if (wq->first_flusher != &this_flusher) 2834 return; 2835 2836 mutex_lock(&wq->mutex); 2837 2838 /* we might have raced, check again with mutex held */ 2839 if (wq->first_flusher != &this_flusher) 2840 goto out_unlock; 2841 2842 wq->first_flusher = NULL; 2843 2844 WARN_ON_ONCE(!list_empty(&this_flusher.list)); 2845 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 2846 2847 while (true) { 2848 struct wq_flusher *next, *tmp; 2849 2850 /* complete all the flushers sharing the current flush color */ 2851 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 2852 if (next->flush_color != wq->flush_color) 2853 break; 2854 list_del_init(&next->list); 2855 complete(&next->done); 2856 } 2857 2858 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 2859 wq->flush_color != work_next_color(wq->work_color)); 2860 2861 /* this flush_color is finished, advance by one */ 2862 wq->flush_color = work_next_color(wq->flush_color); 2863 2864 /* one color has been freed, handle overflow queue */ 2865 if (!list_empty(&wq->flusher_overflow)) { 2866 /* 2867 * Assign the same color to all overflowed 2868 * flushers, advance work_color and append to 2869 * flusher_queue. This is the start-to-wait 2870 * phase for these overflowed flushers. 2871 */ 2872 list_for_each_entry(tmp, &wq->flusher_overflow, list) 2873 tmp->flush_color = wq->work_color; 2874 2875 wq->work_color = work_next_color(wq->work_color); 2876 2877 list_splice_tail_init(&wq->flusher_overflow, 2878 &wq->flusher_queue); 2879 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 2880 } 2881 2882 if (list_empty(&wq->flusher_queue)) { 2883 WARN_ON_ONCE(wq->flush_color != wq->work_color); 2884 break; 2885 } 2886 2887 /* 2888 * Need to flush more colors. Make the next flusher 2889 * the new first flusher and arm pwqs. 2890 */ 2891 WARN_ON_ONCE(wq->flush_color == wq->work_color); 2892 WARN_ON_ONCE(wq->flush_color != next->flush_color); 2893 2894 list_del_init(&next->list); 2895 wq->first_flusher = next; 2896 2897 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 2898 break; 2899 2900 /* 2901 * Meh... this color is already done, clear first 2902 * flusher and repeat cascading. 2903 */ 2904 wq->first_flusher = NULL; 2905 } 2906 2907 out_unlock: 2908 mutex_unlock(&wq->mutex); 2909 } 2910 EXPORT_SYMBOL(flush_workqueue); 2911 2912 /** 2913 * drain_workqueue - drain a workqueue 2914 * @wq: workqueue to drain 2915 * 2916 * Wait until the workqueue becomes empty. While draining is in progress, 2917 * only chain queueing is allowed. IOW, only currently pending or running 2918 * work items on @wq can queue further work items on it. @wq is flushed 2919 * repeatedly until it becomes empty. The number of flushing is determined 2920 * by the depth of chaining and should be relatively short. Whine if it 2921 * takes too long. 2922 */ 2923 void drain_workqueue(struct workqueue_struct *wq) 2924 { 2925 unsigned int flush_cnt = 0; 2926 struct pool_workqueue *pwq; 2927 2928 /* 2929 * __queue_work() needs to test whether there are drainers, is much 2930 * hotter than drain_workqueue() and already looks at @wq->flags. 2931 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 2932 */ 2933 mutex_lock(&wq->mutex); 2934 if (!wq->nr_drainers++) 2935 wq->flags |= __WQ_DRAINING; 2936 mutex_unlock(&wq->mutex); 2937 reflush: 2938 flush_workqueue(wq); 2939 2940 mutex_lock(&wq->mutex); 2941 2942 for_each_pwq(pwq, wq) { 2943 bool drained; 2944 2945 spin_lock_irq(&pwq->pool->lock); 2946 drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2947 spin_unlock_irq(&pwq->pool->lock); 2948 2949 if (drained) 2950 continue; 2951 2952 if (++flush_cnt == 10 || 2953 (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2954 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 2955 wq->name, flush_cnt); 2956 2957 mutex_unlock(&wq->mutex); 2958 goto reflush; 2959 } 2960 2961 if (!--wq->nr_drainers) 2962 wq->flags &= ~__WQ_DRAINING; 2963 mutex_unlock(&wq->mutex); 2964 } 2965 EXPORT_SYMBOL_GPL(drain_workqueue); 2966 2967 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2968 bool from_cancel) 2969 { 2970 struct worker *worker = NULL; 2971 struct worker_pool *pool; 2972 struct pool_workqueue *pwq; 2973 2974 might_sleep(); 2975 2976 rcu_read_lock(); 2977 pool = get_work_pool(work); 2978 if (!pool) { 2979 rcu_read_unlock(); 2980 return false; 2981 } 2982 2983 spin_lock_irq(&pool->lock); 2984 /* see the comment in try_to_grab_pending() with the same code */ 2985 pwq = get_work_pwq(work); 2986 if (pwq) { 2987 if (unlikely(pwq->pool != pool)) 2988 goto already_gone; 2989 } else { 2990 worker = find_worker_executing_work(pool, work); 2991 if (!worker) 2992 goto already_gone; 2993 pwq = worker->current_pwq; 2994 } 2995 2996 check_flush_dependency(pwq->wq, work); 2997 2998 insert_wq_barrier(pwq, barr, work, worker); 2999 spin_unlock_irq(&pool->lock); 3000 3001 /* 3002 * Force a lock recursion deadlock when using flush_work() inside a 3003 * single-threaded or rescuer equipped workqueue. 3004 * 3005 * For single threaded workqueues the deadlock happens when the work 3006 * is after the work issuing the flush_work(). For rescuer equipped 3007 * workqueues the deadlock happens when the rescuer stalls, blocking 3008 * forward progress. 3009 */ 3010 if (!from_cancel && 3011 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3012 lock_map_acquire(&pwq->wq->lockdep_map); 3013 lock_map_release(&pwq->wq->lockdep_map); 3014 } 3015 rcu_read_unlock(); 3016 return true; 3017 already_gone: 3018 spin_unlock_irq(&pool->lock); 3019 rcu_read_unlock(); 3020 return false; 3021 } 3022 3023 static bool __flush_work(struct work_struct *work, bool from_cancel) 3024 { 3025 struct wq_barrier barr; 3026 3027 if (WARN_ON(!wq_online)) 3028 return false; 3029 3030 if (WARN_ON(!work->func)) 3031 return false; 3032 3033 if (!from_cancel) { 3034 lock_map_acquire(&work->lockdep_map); 3035 lock_map_release(&work->lockdep_map); 3036 } 3037 3038 if (start_flush_work(work, &barr, from_cancel)) { 3039 wait_for_completion(&barr.done); 3040 destroy_work_on_stack(&barr.work); 3041 return true; 3042 } else { 3043 return false; 3044 } 3045 } 3046 3047 /** 3048 * flush_work - wait for a work to finish executing the last queueing instance 3049 * @work: the work to flush 3050 * 3051 * Wait until @work has finished execution. @work is guaranteed to be idle 3052 * on return if it hasn't been requeued since flush started. 3053 * 3054 * Return: 3055 * %true if flush_work() waited for the work to finish execution, 3056 * %false if it was already idle. 3057 */ 3058 bool flush_work(struct work_struct *work) 3059 { 3060 return __flush_work(work, false); 3061 } 3062 EXPORT_SYMBOL_GPL(flush_work); 3063 3064 struct cwt_wait { 3065 wait_queue_entry_t wait; 3066 struct work_struct *work; 3067 }; 3068 3069 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 3070 { 3071 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 3072 3073 if (cwait->work != key) 3074 return 0; 3075 return autoremove_wake_function(wait, mode, sync, key); 3076 } 3077 3078 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3079 { 3080 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3081 unsigned long flags; 3082 int ret; 3083 3084 do { 3085 ret = try_to_grab_pending(work, is_dwork, &flags); 3086 /* 3087 * If someone else is already canceling, wait for it to 3088 * finish. flush_work() doesn't work for PREEMPT_NONE 3089 * because we may get scheduled between @work's completion 3090 * and the other canceling task resuming and clearing 3091 * CANCELING - flush_work() will return false immediately 3092 * as @work is no longer busy, try_to_grab_pending() will 3093 * return -ENOENT as @work is still being canceled and the 3094 * other canceling task won't be able to clear CANCELING as 3095 * we're hogging the CPU. 3096 * 3097 * Let's wait for completion using a waitqueue. As this 3098 * may lead to the thundering herd problem, use a custom 3099 * wake function which matches @work along with exclusive 3100 * wait and wakeup. 3101 */ 3102 if (unlikely(ret == -ENOENT)) { 3103 struct cwt_wait cwait; 3104 3105 init_wait(&cwait.wait); 3106 cwait.wait.func = cwt_wakefn; 3107 cwait.work = work; 3108 3109 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 3110 TASK_UNINTERRUPTIBLE); 3111 if (work_is_canceling(work)) 3112 schedule(); 3113 finish_wait(&cancel_waitq, &cwait.wait); 3114 } 3115 } while (unlikely(ret < 0)); 3116 3117 /* tell other tasks trying to grab @work to back off */ 3118 mark_work_canceling(work); 3119 local_irq_restore(flags); 3120 3121 /* 3122 * This allows canceling during early boot. We know that @work 3123 * isn't executing. 3124 */ 3125 if (wq_online) 3126 __flush_work(work, true); 3127 3128 clear_work_data(work); 3129 3130 /* 3131 * Paired with prepare_to_wait() above so that either 3132 * waitqueue_active() is visible here or !work_is_canceling() is 3133 * visible there. 3134 */ 3135 smp_mb(); 3136 if (waitqueue_active(&cancel_waitq)) 3137 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 3138 3139 return ret; 3140 } 3141 3142 /** 3143 * cancel_work_sync - cancel a work and wait for it to finish 3144 * @work: the work to cancel 3145 * 3146 * Cancel @work and wait for its execution to finish. This function 3147 * can be used even if the work re-queues itself or migrates to 3148 * another workqueue. On return from this function, @work is 3149 * guaranteed to be not pending or executing on any CPU. 3150 * 3151 * cancel_work_sync(&delayed_work->work) must not be used for 3152 * delayed_work's. Use cancel_delayed_work_sync() instead. 3153 * 3154 * The caller must ensure that the workqueue on which @work was last 3155 * queued can't be destroyed before this function returns. 3156 * 3157 * Return: 3158 * %true if @work was pending, %false otherwise. 3159 */ 3160 bool cancel_work_sync(struct work_struct *work) 3161 { 3162 return __cancel_work_timer(work, false); 3163 } 3164 EXPORT_SYMBOL_GPL(cancel_work_sync); 3165 3166 /** 3167 * flush_delayed_work - wait for a dwork to finish executing the last queueing 3168 * @dwork: the delayed work to flush 3169 * 3170 * Delayed timer is cancelled and the pending work is queued for 3171 * immediate execution. Like flush_work(), this function only 3172 * considers the last queueing instance of @dwork. 3173 * 3174 * Return: 3175 * %true if flush_work() waited for the work to finish execution, 3176 * %false if it was already idle. 3177 */ 3178 bool flush_delayed_work(struct delayed_work *dwork) 3179 { 3180 local_irq_disable(); 3181 if (del_timer_sync(&dwork->timer)) 3182 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 3183 local_irq_enable(); 3184 return flush_work(&dwork->work); 3185 } 3186 EXPORT_SYMBOL(flush_delayed_work); 3187 3188 /** 3189 * flush_rcu_work - wait for a rwork to finish executing the last queueing 3190 * @rwork: the rcu work to flush 3191 * 3192 * Return: 3193 * %true if flush_rcu_work() waited for the work to finish execution, 3194 * %false if it was already idle. 3195 */ 3196 bool flush_rcu_work(struct rcu_work *rwork) 3197 { 3198 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 3199 rcu_barrier(); 3200 flush_work(&rwork->work); 3201 return true; 3202 } else { 3203 return flush_work(&rwork->work); 3204 } 3205 } 3206 EXPORT_SYMBOL(flush_rcu_work); 3207 3208 static bool __cancel_work(struct work_struct *work, bool is_dwork) 3209 { 3210 unsigned long flags; 3211 int ret; 3212 3213 do { 3214 ret = try_to_grab_pending(work, is_dwork, &flags); 3215 } while (unlikely(ret == -EAGAIN)); 3216 3217 if (unlikely(ret < 0)) 3218 return false; 3219 3220 set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3221 local_irq_restore(flags); 3222 return ret; 3223 } 3224 3225 /** 3226 * cancel_delayed_work - cancel a delayed work 3227 * @dwork: delayed_work to cancel 3228 * 3229 * Kill off a pending delayed_work. 3230 * 3231 * Return: %true if @dwork was pending and canceled; %false if it wasn't 3232 * pending. 3233 * 3234 * Note: 3235 * The work callback function may still be running on return, unless 3236 * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3237 * use cancel_delayed_work_sync() to wait on it. 3238 * 3239 * This function is safe to call from any context including IRQ handler. 3240 */ 3241 bool cancel_delayed_work(struct delayed_work *dwork) 3242 { 3243 return __cancel_work(&dwork->work, true); 3244 } 3245 EXPORT_SYMBOL(cancel_delayed_work); 3246 3247 /** 3248 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3249 * @dwork: the delayed work cancel 3250 * 3251 * This is cancel_work_sync() for delayed works. 3252 * 3253 * Return: 3254 * %true if @dwork was pending, %false otherwise. 3255 */ 3256 bool cancel_delayed_work_sync(struct delayed_work *dwork) 3257 { 3258 return __cancel_work_timer(&dwork->work, true); 3259 } 3260 EXPORT_SYMBOL(cancel_delayed_work_sync); 3261 3262 /** 3263 * schedule_on_each_cpu - execute a function synchronously on each online CPU 3264 * @func: the function to call 3265 * 3266 * schedule_on_each_cpu() executes @func on each online CPU using the 3267 * system workqueue and blocks until all CPUs have completed. 3268 * schedule_on_each_cpu() is very slow. 3269 * 3270 * Return: 3271 * 0 on success, -errno on failure. 3272 */ 3273 int schedule_on_each_cpu(work_func_t func) 3274 { 3275 int cpu; 3276 struct work_struct __percpu *works; 3277 3278 works = alloc_percpu(struct work_struct); 3279 if (!works) 3280 return -ENOMEM; 3281 3282 get_online_cpus(); 3283 3284 for_each_online_cpu(cpu) { 3285 struct work_struct *work = per_cpu_ptr(works, cpu); 3286 3287 INIT_WORK(work, func); 3288 schedule_work_on(cpu, work); 3289 } 3290 3291 for_each_online_cpu(cpu) 3292 flush_work(per_cpu_ptr(works, cpu)); 3293 3294 put_online_cpus(); 3295 free_percpu(works); 3296 return 0; 3297 } 3298 3299 /** 3300 * execute_in_process_context - reliably execute the routine with user context 3301 * @fn: the function to execute 3302 * @ew: guaranteed storage for the execute work structure (must 3303 * be available when the work executes) 3304 * 3305 * Executes the function immediately if process context is available, 3306 * otherwise schedules the function for delayed execution. 3307 * 3308 * Return: 0 - function was executed 3309 * 1 - function was scheduled for execution 3310 */ 3311 int execute_in_process_context(work_func_t fn, struct execute_work *ew) 3312 { 3313 if (!in_interrupt()) { 3314 fn(&ew->work); 3315 return 0; 3316 } 3317 3318 INIT_WORK(&ew->work, fn); 3319 schedule_work(&ew->work); 3320 3321 return 1; 3322 } 3323 EXPORT_SYMBOL_GPL(execute_in_process_context); 3324 3325 /** 3326 * free_workqueue_attrs - free a workqueue_attrs 3327 * @attrs: workqueue_attrs to free 3328 * 3329 * Undo alloc_workqueue_attrs(). 3330 */ 3331 void free_workqueue_attrs(struct workqueue_attrs *attrs) 3332 { 3333 if (attrs) { 3334 free_cpumask_var(attrs->cpumask); 3335 kfree(attrs); 3336 } 3337 } 3338 3339 /** 3340 * alloc_workqueue_attrs - allocate a workqueue_attrs 3341 * @gfp_mask: allocation mask to use 3342 * 3343 * Allocate a new workqueue_attrs, initialize with default settings and 3344 * return it. 3345 * 3346 * Return: The allocated new workqueue_attr on success. %NULL on failure. 3347 */ 3348 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 3349 { 3350 struct workqueue_attrs *attrs; 3351 3352 attrs = kzalloc(sizeof(*attrs), gfp_mask); 3353 if (!attrs) 3354 goto fail; 3355 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 3356 goto fail; 3357 3358 cpumask_copy(attrs->cpumask, cpu_possible_mask); 3359 return attrs; 3360 fail: 3361 free_workqueue_attrs(attrs); 3362 return NULL; 3363 } 3364 3365 static void copy_workqueue_attrs(struct workqueue_attrs *to, 3366 const struct workqueue_attrs *from) 3367 { 3368 to->nice = from->nice; 3369 cpumask_copy(to->cpumask, from->cpumask); 3370 /* 3371 * Unlike hash and equality test, this function doesn't ignore 3372 * ->no_numa as it is used for both pool and wq attrs. Instead, 3373 * get_unbound_pool() explicitly clears ->no_numa after copying. 3374 */ 3375 to->no_numa = from->no_numa; 3376 } 3377 3378 /* hash value of the content of @attr */ 3379 static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 3380 { 3381 u32 hash = 0; 3382 3383 hash = jhash_1word(attrs->nice, hash); 3384 hash = jhash(cpumask_bits(attrs->cpumask), 3385 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 3386 return hash; 3387 } 3388 3389 /* content equality test */ 3390 static bool wqattrs_equal(const struct workqueue_attrs *a, 3391 const struct workqueue_attrs *b) 3392 { 3393 if (a->nice != b->nice) 3394 return false; 3395 if (!cpumask_equal(a->cpumask, b->cpumask)) 3396 return false; 3397 return true; 3398 } 3399 3400 /** 3401 * init_worker_pool - initialize a newly zalloc'd worker_pool 3402 * @pool: worker_pool to initialize 3403 * 3404 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3405 * 3406 * Return: 0 on success, -errno on failure. Even on failure, all fields 3407 * inside @pool proper are initialized and put_unbound_pool() can be called 3408 * on @pool safely to release it. 3409 */ 3410 static int init_worker_pool(struct worker_pool *pool) 3411 { 3412 spin_lock_init(&pool->lock); 3413 pool->id = -1; 3414 pool->cpu = -1; 3415 pool->node = NUMA_NO_NODE; 3416 pool->flags |= POOL_DISASSOCIATED; 3417 pool->watchdog_ts = jiffies; 3418 INIT_LIST_HEAD(&pool->worklist); 3419 INIT_LIST_HEAD(&pool->idle_list); 3420 hash_init(pool->busy_hash); 3421 3422 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 3423 3424 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 3425 3426 INIT_LIST_HEAD(&pool->workers); 3427 3428 ida_init(&pool->worker_ida); 3429 INIT_HLIST_NODE(&pool->hash_node); 3430 pool->refcnt = 1; 3431 3432 /* shouldn't fail above this point */ 3433 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 3434 if (!pool->attrs) 3435 return -ENOMEM; 3436 return 0; 3437 } 3438 3439 #ifdef CONFIG_LOCKDEP 3440 static void wq_init_lockdep(struct workqueue_struct *wq) 3441 { 3442 char *lock_name; 3443 3444 lockdep_register_key(&wq->key); 3445 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3446 if (!lock_name) 3447 lock_name = wq->name; 3448 3449 wq->lock_name = lock_name; 3450 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3451 } 3452 3453 static void wq_unregister_lockdep(struct workqueue_struct *wq) 3454 { 3455 lockdep_unregister_key(&wq->key); 3456 } 3457 3458 static void wq_free_lockdep(struct workqueue_struct *wq) 3459 { 3460 if (wq->lock_name != wq->name) 3461 kfree(wq->lock_name); 3462 } 3463 #else 3464 static void wq_init_lockdep(struct workqueue_struct *wq) 3465 { 3466 } 3467 3468 static void wq_unregister_lockdep(struct workqueue_struct *wq) 3469 { 3470 } 3471 3472 static void wq_free_lockdep(struct workqueue_struct *wq) 3473 { 3474 } 3475 #endif 3476 3477 static void rcu_free_wq(struct rcu_head *rcu) 3478 { 3479 struct workqueue_struct *wq = 3480 container_of(rcu, struct workqueue_struct, rcu); 3481 3482 wq_free_lockdep(wq); 3483 3484 if (!(wq->flags & WQ_UNBOUND)) 3485 free_percpu(wq->cpu_pwqs); 3486 else 3487 free_workqueue_attrs(wq->unbound_attrs); 3488 3489 kfree(wq->rescuer); 3490 kfree(wq); 3491 } 3492 3493 static void rcu_free_pool(struct rcu_head *rcu) 3494 { 3495 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 3496 3497 ida_destroy(&pool->worker_ida); 3498 free_workqueue_attrs(pool->attrs); 3499 kfree(pool); 3500 } 3501 3502 /** 3503 * put_unbound_pool - put a worker_pool 3504 * @pool: worker_pool to put 3505 * 3506 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3507 * safe manner. get_unbound_pool() calls this function on its failure path 3508 * and this function should be able to release pools which went through, 3509 * successfully or not, init_worker_pool(). 3510 * 3511 * Should be called with wq_pool_mutex held. 3512 */ 3513 static void put_unbound_pool(struct worker_pool *pool) 3514 { 3515 DECLARE_COMPLETION_ONSTACK(detach_completion); 3516 struct worker *worker; 3517 3518 lockdep_assert_held(&wq_pool_mutex); 3519 3520 if (--pool->refcnt) 3521 return; 3522 3523 /* sanity checks */ 3524 if (WARN_ON(!(pool->cpu < 0)) || 3525 WARN_ON(!list_empty(&pool->worklist))) 3526 return; 3527 3528 /* release id and unhash */ 3529 if (pool->id >= 0) 3530 idr_remove(&worker_pool_idr, pool->id); 3531 hash_del(&pool->hash_node); 3532 3533 /* 3534 * Become the manager and destroy all workers. This prevents 3535 * @pool's workers from blocking on attach_mutex. We're the last 3536 * manager and @pool gets freed with the flag set. 3537 */ 3538 spin_lock_irq(&pool->lock); 3539 wait_event_lock_irq(wq_manager_wait, 3540 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); 3541 pool->flags |= POOL_MANAGER_ACTIVE; 3542 3543 while ((worker = first_idle_worker(pool))) 3544 destroy_worker(worker); 3545 WARN_ON(pool->nr_workers || pool->nr_idle); 3546 spin_unlock_irq(&pool->lock); 3547 3548 mutex_lock(&wq_pool_attach_mutex); 3549 if (!list_empty(&pool->workers)) 3550 pool->detach_completion = &detach_completion; 3551 mutex_unlock(&wq_pool_attach_mutex); 3552 3553 if (pool->detach_completion) 3554 wait_for_completion(pool->detach_completion); 3555 3556 /* shut down the timers */ 3557 del_timer_sync(&pool->idle_timer); 3558 del_timer_sync(&pool->mayday_timer); 3559 3560 /* RCU protected to allow dereferences from get_work_pool() */ 3561 call_rcu(&pool->rcu, rcu_free_pool); 3562 } 3563 3564 /** 3565 * get_unbound_pool - get a worker_pool with the specified attributes 3566 * @attrs: the attributes of the worker_pool to get 3567 * 3568 * Obtain a worker_pool which has the same attributes as @attrs, bump the 3569 * reference count and return it. If there already is a matching 3570 * worker_pool, it will be used; otherwise, this function attempts to 3571 * create a new one. 3572 * 3573 * Should be called with wq_pool_mutex held. 3574 * 3575 * Return: On success, a worker_pool with the same attributes as @attrs. 3576 * On failure, %NULL. 3577 */ 3578 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 3579 { 3580 u32 hash = wqattrs_hash(attrs); 3581 struct worker_pool *pool; 3582 int node; 3583 int target_node = NUMA_NO_NODE; 3584 3585 lockdep_assert_held(&wq_pool_mutex); 3586 3587 /* do we already have a matching pool? */ 3588 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 3589 if (wqattrs_equal(pool->attrs, attrs)) { 3590 pool->refcnt++; 3591 return pool; 3592 } 3593 } 3594 3595 /* if cpumask is contained inside a NUMA node, we belong to that node */ 3596 if (wq_numa_enabled) { 3597 for_each_node(node) { 3598 if (cpumask_subset(attrs->cpumask, 3599 wq_numa_possible_cpumask[node])) { 3600 target_node = node; 3601 break; 3602 } 3603 } 3604 } 3605 3606 /* nope, create a new one */ 3607 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 3608 if (!pool || init_worker_pool(pool) < 0) 3609 goto fail; 3610 3611 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 3612 copy_workqueue_attrs(pool->attrs, attrs); 3613 pool->node = target_node; 3614 3615 /* 3616 * no_numa isn't a worker_pool attribute, always clear it. See 3617 * 'struct workqueue_attrs' comments for detail. 3618 */ 3619 pool->attrs->no_numa = false; 3620 3621 if (worker_pool_assign_id(pool) < 0) 3622 goto fail; 3623 3624 /* create and start the initial worker */ 3625 if (wq_online && !create_worker(pool)) 3626 goto fail; 3627 3628 /* install */ 3629 hash_add(unbound_pool_hash, &pool->hash_node, hash); 3630 3631 return pool; 3632 fail: 3633 if (pool) 3634 put_unbound_pool(pool); 3635 return NULL; 3636 } 3637 3638 static void rcu_free_pwq(struct rcu_head *rcu) 3639 { 3640 kmem_cache_free(pwq_cache, 3641 container_of(rcu, struct pool_workqueue, rcu)); 3642 } 3643 3644 /* 3645 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 3646 * and needs to be destroyed. 3647 */ 3648 static void pwq_unbound_release_workfn(struct work_struct *work) 3649 { 3650 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 3651 unbound_release_work); 3652 struct workqueue_struct *wq = pwq->wq; 3653 struct worker_pool *pool = pwq->pool; 3654 bool is_last; 3655 3656 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 3657 return; 3658 3659 mutex_lock(&wq->mutex); 3660 list_del_rcu(&pwq->pwqs_node); 3661 is_last = list_empty(&wq->pwqs); 3662 mutex_unlock(&wq->mutex); 3663 3664 mutex_lock(&wq_pool_mutex); 3665 put_unbound_pool(pool); 3666 mutex_unlock(&wq_pool_mutex); 3667 3668 call_rcu(&pwq->rcu, rcu_free_pwq); 3669 3670 /* 3671 * If we're the last pwq going away, @wq is already dead and no one 3672 * is gonna access it anymore. Schedule RCU free. 3673 */ 3674 if (is_last) { 3675 wq_unregister_lockdep(wq); 3676 call_rcu(&wq->rcu, rcu_free_wq); 3677 } 3678 } 3679 3680 /** 3681 * pwq_adjust_max_active - update a pwq's max_active to the current setting 3682 * @pwq: target pool_workqueue 3683 * 3684 * If @pwq isn't freezing, set @pwq->max_active to the associated 3685 * workqueue's saved_max_active and activate delayed work items 3686 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 3687 */ 3688 static void pwq_adjust_max_active(struct pool_workqueue *pwq) 3689 { 3690 struct workqueue_struct *wq = pwq->wq; 3691 bool freezable = wq->flags & WQ_FREEZABLE; 3692 unsigned long flags; 3693 3694 /* for @wq->saved_max_active */ 3695 lockdep_assert_held(&wq->mutex); 3696 3697 /* fast exit for non-freezable wqs */ 3698 if (!freezable && pwq->max_active == wq->saved_max_active) 3699 return; 3700 3701 /* this function can be called during early boot w/ irq disabled */ 3702 spin_lock_irqsave(&pwq->pool->lock, flags); 3703 3704 /* 3705 * During [un]freezing, the caller is responsible for ensuring that 3706 * this function is called at least once after @workqueue_freezing 3707 * is updated and visible. 3708 */ 3709 if (!freezable || !workqueue_freezing) { 3710 pwq->max_active = wq->saved_max_active; 3711 3712 while (!list_empty(&pwq->delayed_works) && 3713 pwq->nr_active < pwq->max_active) 3714 pwq_activate_first_delayed(pwq); 3715 3716 /* 3717 * Need to kick a worker after thawed or an unbound wq's 3718 * max_active is bumped. It's a slow path. Do it always. 3719 */ 3720 wake_up_worker(pwq->pool); 3721 } else { 3722 pwq->max_active = 0; 3723 } 3724 3725 spin_unlock_irqrestore(&pwq->pool->lock, flags); 3726 } 3727 3728 /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3729 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3730 struct worker_pool *pool) 3731 { 3732 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3733 3734 memset(pwq, 0, sizeof(*pwq)); 3735 3736 pwq->pool = pool; 3737 pwq->wq = wq; 3738 pwq->flush_color = -1; 3739 pwq->refcnt = 1; 3740 INIT_LIST_HEAD(&pwq->delayed_works); 3741 INIT_LIST_HEAD(&pwq->pwqs_node); 3742 INIT_LIST_HEAD(&pwq->mayday_node); 3743 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3744 } 3745 3746 /* sync @pwq with the current state of its associated wq and link it */ 3747 static void link_pwq(struct pool_workqueue *pwq) 3748 { 3749 struct workqueue_struct *wq = pwq->wq; 3750 3751 lockdep_assert_held(&wq->mutex); 3752 3753 /* may be called multiple times, ignore if already linked */ 3754 if (!list_empty(&pwq->pwqs_node)) 3755 return; 3756 3757 /* set the matching work_color */ 3758 pwq->work_color = wq->work_color; 3759 3760 /* sync max_active to the current setting */ 3761 pwq_adjust_max_active(pwq); 3762 3763 /* link in @pwq */ 3764 list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3765 } 3766 3767 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3768 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3769 const struct workqueue_attrs *attrs) 3770 { 3771 struct worker_pool *pool; 3772 struct pool_workqueue *pwq; 3773 3774 lockdep_assert_held(&wq_pool_mutex); 3775 3776 pool = get_unbound_pool(attrs); 3777 if (!pool) 3778 return NULL; 3779 3780 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3781 if (!pwq) { 3782 put_unbound_pool(pool); 3783 return NULL; 3784 } 3785 3786 init_pwq(pwq, wq, pool); 3787 return pwq; 3788 } 3789 3790 /** 3791 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3792 * @attrs: the wq_attrs of the default pwq of the target workqueue 3793 * @node: the target NUMA node 3794 * @cpu_going_down: if >= 0, the CPU to consider as offline 3795 * @cpumask: outarg, the resulting cpumask 3796 * 3797 * Calculate the cpumask a workqueue with @attrs should use on @node. If 3798 * @cpu_going_down is >= 0, that cpu is considered offline during 3799 * calculation. The result is stored in @cpumask. 3800 * 3801 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 3802 * enabled and @node has online CPUs requested by @attrs, the returned 3803 * cpumask is the intersection of the possible CPUs of @node and 3804 * @attrs->cpumask. 3805 * 3806 * The caller is responsible for ensuring that the cpumask of @node stays 3807 * stable. 3808 * 3809 * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3810 * %false if equal. 3811 */ 3812 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 3813 int cpu_going_down, cpumask_t *cpumask) 3814 { 3815 if (!wq_numa_enabled || attrs->no_numa) 3816 goto use_dfl; 3817 3818 /* does @node have any online CPUs @attrs wants? */ 3819 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 3820 if (cpu_going_down >= 0) 3821 cpumask_clear_cpu(cpu_going_down, cpumask); 3822 3823 if (cpumask_empty(cpumask)) 3824 goto use_dfl; 3825 3826 /* yeap, return possible CPUs in @node that @attrs wants */ 3827 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 3828 3829 if (cpumask_empty(cpumask)) { 3830 pr_warn_once("WARNING: workqueue cpumask: online intersect > " 3831 "possible intersect\n"); 3832 return false; 3833 } 3834 3835 return !cpumask_equal(cpumask, attrs->cpumask); 3836 3837 use_dfl: 3838 cpumask_copy(cpumask, attrs->cpumask); 3839 return false; 3840 } 3841 3842 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 3843 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 3844 int node, 3845 struct pool_workqueue *pwq) 3846 { 3847 struct pool_workqueue *old_pwq; 3848 3849 lockdep_assert_held(&wq_pool_mutex); 3850 lockdep_assert_held(&wq->mutex); 3851 3852 /* link_pwq() can handle duplicate calls */ 3853 link_pwq(pwq); 3854 3855 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 3856 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 3857 return old_pwq; 3858 } 3859 3860 /* context to store the prepared attrs & pwqs before applying */ 3861 struct apply_wqattrs_ctx { 3862 struct workqueue_struct *wq; /* target workqueue */ 3863 struct workqueue_attrs *attrs; /* attrs to apply */ 3864 struct list_head list; /* queued for batching commit */ 3865 struct pool_workqueue *dfl_pwq; 3866 struct pool_workqueue *pwq_tbl[]; 3867 }; 3868 3869 /* free the resources after success or abort */ 3870 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 3871 { 3872 if (ctx) { 3873 int node; 3874 3875 for_each_node(node) 3876 put_pwq_unlocked(ctx->pwq_tbl[node]); 3877 put_pwq_unlocked(ctx->dfl_pwq); 3878 3879 free_workqueue_attrs(ctx->attrs); 3880 3881 kfree(ctx); 3882 } 3883 } 3884 3885 /* allocate the attrs and pwqs for later installation */ 3886 static struct apply_wqattrs_ctx * 3887 apply_wqattrs_prepare(struct workqueue_struct *wq, 3888 const struct workqueue_attrs *attrs) 3889 { 3890 struct apply_wqattrs_ctx *ctx; 3891 struct workqueue_attrs *new_attrs, *tmp_attrs; 3892 int node; 3893 3894 lockdep_assert_held(&wq_pool_mutex); 3895 3896 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 3897 3898 new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 3899 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 3900 if (!ctx || !new_attrs || !tmp_attrs) 3901 goto out_free; 3902 3903 /* 3904 * Calculate the attrs of the default pwq. 3905 * If the user configured cpumask doesn't overlap with the 3906 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3907 */ 3908 copy_workqueue_attrs(new_attrs, attrs); 3909 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3910 if (unlikely(cpumask_empty(new_attrs->cpumask))) 3911 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 3912 3913 /* 3914 * We may create multiple pwqs with differing cpumasks. Make a 3915 * copy of @new_attrs which will be modified and used to obtain 3916 * pools. 3917 */ 3918 copy_workqueue_attrs(tmp_attrs, new_attrs); 3919 3920 /* 3921 * If something goes wrong during CPU up/down, we'll fall back to 3922 * the default pwq covering whole @attrs->cpumask. Always create 3923 * it even if we don't use it immediately. 3924 */ 3925 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 3926 if (!ctx->dfl_pwq) 3927 goto out_free; 3928 3929 for_each_node(node) { 3930 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 3931 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 3932 if (!ctx->pwq_tbl[node]) 3933 goto out_free; 3934 } else { 3935 ctx->dfl_pwq->refcnt++; 3936 ctx->pwq_tbl[node] = ctx->dfl_pwq; 3937 } 3938 } 3939 3940 /* save the user configured attrs and sanitize it. */ 3941 copy_workqueue_attrs(new_attrs, attrs); 3942 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 3943 ctx->attrs = new_attrs; 3944 3945 ctx->wq = wq; 3946 free_workqueue_attrs(tmp_attrs); 3947 return ctx; 3948 3949 out_free: 3950 free_workqueue_attrs(tmp_attrs); 3951 free_workqueue_attrs(new_attrs); 3952 apply_wqattrs_cleanup(ctx); 3953 return NULL; 3954 } 3955 3956 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 3957 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 3958 { 3959 int node; 3960 3961 /* all pwqs have been created successfully, let's install'em */ 3962 mutex_lock(&ctx->wq->mutex); 3963 3964 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 3965 3966 /* save the previous pwq and install the new one */ 3967 for_each_node(node) 3968 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 3969 ctx->pwq_tbl[node]); 3970 3971 /* @dfl_pwq might not have been used, ensure it's linked */ 3972 link_pwq(ctx->dfl_pwq); 3973 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 3974 3975 mutex_unlock(&ctx->wq->mutex); 3976 } 3977 3978 static void apply_wqattrs_lock(void) 3979 { 3980 /* CPUs should stay stable across pwq creations and installations */ 3981 get_online_cpus(); 3982 mutex_lock(&wq_pool_mutex); 3983 } 3984 3985 static void apply_wqattrs_unlock(void) 3986 { 3987 mutex_unlock(&wq_pool_mutex); 3988 put_online_cpus(); 3989 } 3990 3991 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3992 const struct workqueue_attrs *attrs) 3993 { 3994 struct apply_wqattrs_ctx *ctx; 3995 3996 /* only unbound workqueues can change attributes */ 3997 if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3998 return -EINVAL; 3999 4000 /* creating multiple pwqs breaks ordering guarantee */ 4001 if (!list_empty(&wq->pwqs)) { 4002 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4003 return -EINVAL; 4004 4005 wq->flags &= ~__WQ_ORDERED; 4006 } 4007 4008 ctx = apply_wqattrs_prepare(wq, attrs); 4009 if (!ctx) 4010 return -ENOMEM; 4011 4012 /* the ctx has been prepared successfully, let's commit it */ 4013 apply_wqattrs_commit(ctx); 4014 apply_wqattrs_cleanup(ctx); 4015 4016 return 0; 4017 } 4018 4019 /** 4020 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 4021 * @wq: the target workqueue 4022 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 4023 * 4024 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 4025 * machines, this function maps a separate pwq to each NUMA node with 4026 * possibles CPUs in @attrs->cpumask so that work items are affine to the 4027 * NUMA node it was issued on. Older pwqs are released as in-flight work 4028 * items finish. Note that a work item which repeatedly requeues itself 4029 * back-to-back will stay on its current pwq. 4030 * 4031 * Performs GFP_KERNEL allocations. 4032 * 4033 * Return: 0 on success and -errno on failure. 4034 */ 4035 int apply_workqueue_attrs(struct workqueue_struct *wq, 4036 const struct workqueue_attrs *attrs) 4037 { 4038 int ret; 4039 4040 apply_wqattrs_lock(); 4041 ret = apply_workqueue_attrs_locked(wq, attrs); 4042 apply_wqattrs_unlock(); 4043 4044 return ret; 4045 } 4046 EXPORT_SYMBOL_GPL(apply_workqueue_attrs); 4047 4048 /** 4049 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 4050 * @wq: the target workqueue 4051 * @cpu: the CPU coming up or going down 4052 * @online: whether @cpu is coming up or going down 4053 * 4054 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 4055 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 4056 * @wq accordingly. 4057 * 4058 * If NUMA affinity can't be adjusted due to memory allocation failure, it 4059 * falls back to @wq->dfl_pwq which may not be optimal but is always 4060 * correct. 4061 * 4062 * Note that when the last allowed CPU of a NUMA node goes offline for a 4063 * workqueue with a cpumask spanning multiple nodes, the workers which were 4064 * already executing the work items for the workqueue will lose their CPU 4065 * affinity and may execute on any CPU. This is similar to how per-cpu 4066 * workqueues behave on CPU_DOWN. If a workqueue user wants strict 4067 * affinity, it's the user's responsibility to flush the work item from 4068 * CPU_DOWN_PREPARE. 4069 */ 4070 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 4071 bool online) 4072 { 4073 int node = cpu_to_node(cpu); 4074 int cpu_off = online ? -1 : cpu; 4075 struct pool_workqueue *old_pwq = NULL, *pwq; 4076 struct workqueue_attrs *target_attrs; 4077 cpumask_t *cpumask; 4078 4079 lockdep_assert_held(&wq_pool_mutex); 4080 4081 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4082 wq->unbound_attrs->no_numa) 4083 return; 4084 4085 /* 4086 * We don't wanna alloc/free wq_attrs for each wq for each CPU. 4087 * Let's use a preallocated one. The following buf is protected by 4088 * CPU hotplug exclusion. 4089 */ 4090 target_attrs = wq_update_unbound_numa_attrs_buf; 4091 cpumask = target_attrs->cpumask; 4092 4093 copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 4094 pwq = unbound_pwq_by_node(wq, node); 4095 4096 /* 4097 * Let's determine what needs to be done. If the target cpumask is 4098 * different from the default pwq's, we need to compare it to @pwq's 4099 * and create a new one if they don't match. If the target cpumask 4100 * equals the default pwq's, the default pwq should be used. 4101 */ 4102 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 4103 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4104 return; 4105 } else { 4106 goto use_dfl_pwq; 4107 } 4108 4109 /* create a new pwq */ 4110 pwq = alloc_unbound_pwq(wq, target_attrs); 4111 if (!pwq) { 4112 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 4113 wq->name); 4114 goto use_dfl_pwq; 4115 } 4116 4117 /* Install the new pwq. */ 4118 mutex_lock(&wq->mutex); 4119 old_pwq = numa_pwq_tbl_install(wq, node, pwq); 4120 goto out_unlock; 4121 4122 use_dfl_pwq: 4123 mutex_lock(&wq->mutex); 4124 spin_lock_irq(&wq->dfl_pwq->pool->lock); 4125 get_pwq(wq->dfl_pwq); 4126 spin_unlock_irq(&wq->dfl_pwq->pool->lock); 4127 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 4128 out_unlock: 4129 mutex_unlock(&wq->mutex); 4130 put_pwq_unlocked(old_pwq); 4131 } 4132 4133 static int alloc_and_link_pwqs(struct workqueue_struct *wq) 4134 { 4135 bool highpri = wq->flags & WQ_HIGHPRI; 4136 int cpu, ret; 4137 4138 if (!(wq->flags & WQ_UNBOUND)) { 4139 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4140 if (!wq->cpu_pwqs) 4141 return -ENOMEM; 4142 4143 for_each_possible_cpu(cpu) { 4144 struct pool_workqueue *pwq = 4145 per_cpu_ptr(wq->cpu_pwqs, cpu); 4146 struct worker_pool *cpu_pools = 4147 per_cpu(cpu_worker_pools, cpu); 4148 4149 init_pwq(pwq, wq, &cpu_pools[highpri]); 4150 4151 mutex_lock(&wq->mutex); 4152 link_pwq(pwq); 4153 mutex_unlock(&wq->mutex); 4154 } 4155 return 0; 4156 } else if (wq->flags & __WQ_ORDERED) { 4157 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 4158 /* there should only be single pwq for ordering guarantee */ 4159 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 4160 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 4161 "ordering guarantee broken for workqueue %s\n", wq->name); 4162 return ret; 4163 } else { 4164 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 4165 } 4166 } 4167 4168 static int wq_clamp_max_active(int max_active, unsigned int flags, 4169 const char *name) 4170 { 4171 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4172 4173 if (max_active < 1 || max_active > lim) 4174 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4175 max_active, name, 1, lim); 4176 4177 return clamp_val(max_active, 1, lim); 4178 } 4179 4180 /* 4181 * Workqueues which may be used during memory reclaim should have a rescuer 4182 * to guarantee forward progress. 4183 */ 4184 static int init_rescuer(struct workqueue_struct *wq) 4185 { 4186 struct worker *rescuer; 4187 int ret; 4188 4189 if (!(wq->flags & WQ_MEM_RECLAIM)) 4190 return 0; 4191 4192 rescuer = alloc_worker(NUMA_NO_NODE); 4193 if (!rescuer) 4194 return -ENOMEM; 4195 4196 rescuer->rescue_wq = wq; 4197 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4198 ret = PTR_ERR_OR_ZERO(rescuer->task); 4199 if (ret) { 4200 kfree(rescuer); 4201 return ret; 4202 } 4203 4204 wq->rescuer = rescuer; 4205 kthread_bind_mask(rescuer->task, cpu_possible_mask); 4206 wake_up_process(rescuer->task); 4207 4208 return 0; 4209 } 4210 4211 __printf(1, 4) 4212 struct workqueue_struct *alloc_workqueue(const char *fmt, 4213 unsigned int flags, 4214 int max_active, ...) 4215 { 4216 size_t tbl_size = 0; 4217 va_list args; 4218 struct workqueue_struct *wq; 4219 struct pool_workqueue *pwq; 4220 4221 /* 4222 * Unbound && max_active == 1 used to imply ordered, which is no 4223 * longer the case on NUMA machines due to per-node pools. While 4224 * alloc_ordered_workqueue() is the right way to create an ordered 4225 * workqueue, keep the previous behavior to avoid subtle breakages 4226 * on NUMA. 4227 */ 4228 if ((flags & WQ_UNBOUND) && max_active == 1) 4229 flags |= __WQ_ORDERED; 4230 4231 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4232 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4233 flags |= WQ_UNBOUND; 4234 4235 /* allocate wq and format name */ 4236 if (flags & WQ_UNBOUND) 4237 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4238 4239 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4240 if (!wq) 4241 return NULL; 4242 4243 if (flags & WQ_UNBOUND) { 4244 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 4245 if (!wq->unbound_attrs) 4246 goto err_free_wq; 4247 } 4248 4249 va_start(args, max_active); 4250 vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4251 va_end(args); 4252 4253 max_active = max_active ?: WQ_DFL_ACTIVE; 4254 max_active = wq_clamp_max_active(max_active, flags, wq->name); 4255 4256 /* init wq */ 4257 wq->flags = flags; 4258 wq->saved_max_active = max_active; 4259 mutex_init(&wq->mutex); 4260 atomic_set(&wq->nr_pwqs_to_flush, 0); 4261 INIT_LIST_HEAD(&wq->pwqs); 4262 INIT_LIST_HEAD(&wq->flusher_queue); 4263 INIT_LIST_HEAD(&wq->flusher_overflow); 4264 INIT_LIST_HEAD(&wq->maydays); 4265 4266 wq_init_lockdep(wq); 4267 INIT_LIST_HEAD(&wq->list); 4268 4269 if (alloc_and_link_pwqs(wq) < 0) 4270 goto err_unreg_lockdep; 4271 4272 if (wq_online && init_rescuer(wq) < 0) 4273 goto err_destroy; 4274 4275 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4276 goto err_destroy; 4277 4278 /* 4279 * wq_pool_mutex protects global freeze state and workqueues list. 4280 * Grab it, adjust max_active and add the new @wq to workqueues 4281 * list. 4282 */ 4283 mutex_lock(&wq_pool_mutex); 4284 4285 mutex_lock(&wq->mutex); 4286 for_each_pwq(pwq, wq) 4287 pwq_adjust_max_active(pwq); 4288 mutex_unlock(&wq->mutex); 4289 4290 list_add_tail_rcu(&wq->list, &workqueues); 4291 4292 mutex_unlock(&wq_pool_mutex); 4293 4294 return wq; 4295 4296 err_unreg_lockdep: 4297 wq_unregister_lockdep(wq); 4298 wq_free_lockdep(wq); 4299 err_free_wq: 4300 free_workqueue_attrs(wq->unbound_attrs); 4301 kfree(wq); 4302 return NULL; 4303 err_destroy: 4304 destroy_workqueue(wq); 4305 return NULL; 4306 } 4307 EXPORT_SYMBOL_GPL(alloc_workqueue); 4308 4309 /** 4310 * destroy_workqueue - safely terminate a workqueue 4311 * @wq: target workqueue 4312 * 4313 * Safely destroy a workqueue. All work currently pending will be done first. 4314 */ 4315 void destroy_workqueue(struct workqueue_struct *wq) 4316 { 4317 struct pool_workqueue *pwq; 4318 int node; 4319 4320 /* drain it before proceeding with destruction */ 4321 drain_workqueue(wq); 4322 4323 /* sanity checks */ 4324 mutex_lock(&wq->mutex); 4325 for_each_pwq(pwq, wq) { 4326 int i; 4327 4328 for (i = 0; i < WORK_NR_COLORS; i++) { 4329 if (WARN_ON(pwq->nr_in_flight[i])) { 4330 mutex_unlock(&wq->mutex); 4331 show_workqueue_state(); 4332 return; 4333 } 4334 } 4335 4336 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 4337 WARN_ON(pwq->nr_active) || 4338 WARN_ON(!list_empty(&pwq->delayed_works))) { 4339 mutex_unlock(&wq->mutex); 4340 show_workqueue_state(); 4341 return; 4342 } 4343 } 4344 mutex_unlock(&wq->mutex); 4345 4346 /* 4347 * wq list is used to freeze wq, remove from list after 4348 * flushing is complete in case freeze races us. 4349 */ 4350 mutex_lock(&wq_pool_mutex); 4351 list_del_rcu(&wq->list); 4352 mutex_unlock(&wq_pool_mutex); 4353 4354 workqueue_sysfs_unregister(wq); 4355 4356 if (wq->rescuer) 4357 kthread_stop(wq->rescuer->task); 4358 4359 if (!(wq->flags & WQ_UNBOUND)) { 4360 wq_unregister_lockdep(wq); 4361 /* 4362 * The base ref is never dropped on per-cpu pwqs. Directly 4363 * schedule RCU free. 4364 */ 4365 call_rcu(&wq->rcu, rcu_free_wq); 4366 } else { 4367 /* 4368 * We're the sole accessor of @wq at this point. Directly 4369 * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 4370 * @wq will be freed when the last pwq is released. 4371 */ 4372 for_each_node(node) { 4373 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 4374 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 4375 put_pwq_unlocked(pwq); 4376 } 4377 4378 /* 4379 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 4380 * put. Don't access it afterwards. 4381 */ 4382 pwq = wq->dfl_pwq; 4383 wq->dfl_pwq = NULL; 4384 put_pwq_unlocked(pwq); 4385 } 4386 } 4387 EXPORT_SYMBOL_GPL(destroy_workqueue); 4388 4389 /** 4390 * workqueue_set_max_active - adjust max_active of a workqueue 4391 * @wq: target workqueue 4392 * @max_active: new max_active value. 4393 * 4394 * Set max_active of @wq to @max_active. 4395 * 4396 * CONTEXT: 4397 * Don't call from IRQ context. 4398 */ 4399 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4400 { 4401 struct pool_workqueue *pwq; 4402 4403 /* disallow meddling with max_active for ordered workqueues */ 4404 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4405 return; 4406 4407 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4408 4409 mutex_lock(&wq->mutex); 4410 4411 wq->flags &= ~__WQ_ORDERED; 4412 wq->saved_max_active = max_active; 4413 4414 for_each_pwq(pwq, wq) 4415 pwq_adjust_max_active(pwq); 4416 4417 mutex_unlock(&wq->mutex); 4418 } 4419 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4420 4421 /** 4422 * current_work - retrieve %current task's work struct 4423 * 4424 * Determine if %current task is a workqueue worker and what it's working on. 4425 * Useful to find out the context that the %current task is running in. 4426 * 4427 * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 4428 */ 4429 struct work_struct *current_work(void) 4430 { 4431 struct worker *worker = current_wq_worker(); 4432 4433 return worker ? worker->current_work : NULL; 4434 } 4435 EXPORT_SYMBOL(current_work); 4436 4437 /** 4438 * current_is_workqueue_rescuer - is %current workqueue rescuer? 4439 * 4440 * Determine whether %current is a workqueue rescuer. Can be used from 4441 * work functions to determine whether it's being run off the rescuer task. 4442 * 4443 * Return: %true if %current is a workqueue rescuer. %false otherwise. 4444 */ 4445 bool current_is_workqueue_rescuer(void) 4446 { 4447 struct worker *worker = current_wq_worker(); 4448 4449 return worker && worker->rescue_wq; 4450 } 4451 4452 /** 4453 * workqueue_congested - test whether a workqueue is congested 4454 * @cpu: CPU in question 4455 * @wq: target workqueue 4456 * 4457 * Test whether @wq's cpu workqueue for @cpu is congested. There is 4458 * no synchronization around this function and the test result is 4459 * unreliable and only useful as advisory hints or for debugging. 4460 * 4461 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4462 * Note that both per-cpu and unbound workqueues may be associated with 4463 * multiple pool_workqueues which have separate congested states. A 4464 * workqueue being congested on one CPU doesn't mean the workqueue is also 4465 * contested on other CPUs / NUMA nodes. 4466 * 4467 * Return: 4468 * %true if congested, %false otherwise. 4469 */ 4470 bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4471 { 4472 struct pool_workqueue *pwq; 4473 bool ret; 4474 4475 rcu_read_lock(); 4476 preempt_disable(); 4477 4478 if (cpu == WORK_CPU_UNBOUND) 4479 cpu = smp_processor_id(); 4480 4481 if (!(wq->flags & WQ_UNBOUND)) 4482 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 4483 else 4484 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4485 4486 ret = !list_empty(&pwq->delayed_works); 4487 preempt_enable(); 4488 rcu_read_unlock(); 4489 4490 return ret; 4491 } 4492 EXPORT_SYMBOL_GPL(workqueue_congested); 4493 4494 /** 4495 * work_busy - test whether a work is currently pending or running 4496 * @work: the work to be tested 4497 * 4498 * Test whether @work is currently pending or running. There is no 4499 * synchronization around this function and the test result is 4500 * unreliable and only useful as advisory hints or for debugging. 4501 * 4502 * Return: 4503 * OR'd bitmask of WORK_BUSY_* bits. 4504 */ 4505 unsigned int work_busy(struct work_struct *work) 4506 { 4507 struct worker_pool *pool; 4508 unsigned long flags; 4509 unsigned int ret = 0; 4510 4511 if (work_pending(work)) 4512 ret |= WORK_BUSY_PENDING; 4513 4514 rcu_read_lock(); 4515 pool = get_work_pool(work); 4516 if (pool) { 4517 spin_lock_irqsave(&pool->lock, flags); 4518 if (find_worker_executing_work(pool, work)) 4519 ret |= WORK_BUSY_RUNNING; 4520 spin_unlock_irqrestore(&pool->lock, flags); 4521 } 4522 rcu_read_unlock(); 4523 4524 return ret; 4525 } 4526 EXPORT_SYMBOL_GPL(work_busy); 4527 4528 /** 4529 * set_worker_desc - set description for the current work item 4530 * @fmt: printf-style format string 4531 * @...: arguments for the format string 4532 * 4533 * This function can be called by a running work function to describe what 4534 * the work item is about. If the worker task gets dumped, this 4535 * information will be printed out together to help debugging. The 4536 * description can be at most WORKER_DESC_LEN including the trailing '\0'. 4537 */ 4538 void set_worker_desc(const char *fmt, ...) 4539 { 4540 struct worker *worker = current_wq_worker(); 4541 va_list args; 4542 4543 if (worker) { 4544 va_start(args, fmt); 4545 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 4546 va_end(args); 4547 } 4548 } 4549 EXPORT_SYMBOL_GPL(set_worker_desc); 4550 4551 /** 4552 * print_worker_info - print out worker information and description 4553 * @log_lvl: the log level to use when printing 4554 * @task: target task 4555 * 4556 * If @task is a worker and currently executing a work item, print out the 4557 * name of the workqueue being serviced and worker description set with 4558 * set_worker_desc() by the currently executing work item. 4559 * 4560 * This function can be safely called on any task as long as the 4561 * task_struct itself is accessible. While safe, this function isn't 4562 * synchronized and may print out mixups or garbages of limited length. 4563 */ 4564 void print_worker_info(const char *log_lvl, struct task_struct *task) 4565 { 4566 work_func_t *fn = NULL; 4567 char name[WQ_NAME_LEN] = { }; 4568 char desc[WORKER_DESC_LEN] = { }; 4569 struct pool_workqueue *pwq = NULL; 4570 struct workqueue_struct *wq = NULL; 4571 struct worker *worker; 4572 4573 if (!(task->flags & PF_WQ_WORKER)) 4574 return; 4575 4576 /* 4577 * This function is called without any synchronization and @task 4578 * could be in any state. Be careful with dereferences. 4579 */ 4580 worker = kthread_probe_data(task); 4581 4582 /* 4583 * Carefully copy the associated workqueue's workfn, name and desc. 4584 * Keep the original last '\0' in case the original is garbage. 4585 */ 4586 probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 4587 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 4588 probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 4589 probe_kernel_read(name, wq->name, sizeof(name) - 1); 4590 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 4591 4592 if (fn || name[0] || desc[0]) { 4593 printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 4594 if (strcmp(name, desc)) 4595 pr_cont(" (%s)", desc); 4596 pr_cont("\n"); 4597 } 4598 } 4599 4600 static void pr_cont_pool_info(struct worker_pool *pool) 4601 { 4602 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 4603 if (pool->node != NUMA_NO_NODE) 4604 pr_cont(" node=%d", pool->node); 4605 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 4606 } 4607 4608 static void pr_cont_work(bool comma, struct work_struct *work) 4609 { 4610 if (work->func == wq_barrier_func) { 4611 struct wq_barrier *barr; 4612 4613 barr = container_of(work, struct wq_barrier, work); 4614 4615 pr_cont("%s BAR(%d)", comma ? "," : "", 4616 task_pid_nr(barr->task)); 4617 } else { 4618 pr_cont("%s %ps", comma ? "," : "", work->func); 4619 } 4620 } 4621 4622 static void show_pwq(struct pool_workqueue *pwq) 4623 { 4624 struct worker_pool *pool = pwq->pool; 4625 struct work_struct *work; 4626 struct worker *worker; 4627 bool has_in_flight = false, has_pending = false; 4628 int bkt; 4629 4630 pr_info(" pwq %d:", pool->id); 4631 pr_cont_pool_info(pool); 4632 4633 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 4634 !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 4635 4636 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 4637 if (worker->current_pwq == pwq) { 4638 has_in_flight = true; 4639 break; 4640 } 4641 } 4642 if (has_in_flight) { 4643 bool comma = false; 4644 4645 pr_info(" in-flight:"); 4646 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 4647 if (worker->current_pwq != pwq) 4648 continue; 4649 4650 pr_cont("%s %d%s:%ps", comma ? "," : "", 4651 task_pid_nr(worker->task), 4652 worker == pwq->wq->rescuer ? "(RESCUER)" : "", 4653 worker->current_func); 4654 list_for_each_entry(work, &worker->scheduled, entry) 4655 pr_cont_work(false, work); 4656 comma = true; 4657 } 4658 pr_cont("\n"); 4659 } 4660 4661 list_for_each_entry(work, &pool->worklist, entry) { 4662 if (get_work_pwq(work) == pwq) { 4663 has_pending = true; 4664 break; 4665 } 4666 } 4667 if (has_pending) { 4668 bool comma = false; 4669 4670 pr_info(" pending:"); 4671 list_for_each_entry(work, &pool->worklist, entry) { 4672 if (get_work_pwq(work) != pwq) 4673 continue; 4674 4675 pr_cont_work(comma, work); 4676 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 4677 } 4678 pr_cont("\n"); 4679 } 4680 4681 if (!list_empty(&pwq->delayed_works)) { 4682 bool comma = false; 4683 4684 pr_info(" delayed:"); 4685 list_for_each_entry(work, &pwq->delayed_works, entry) { 4686 pr_cont_work(comma, work); 4687 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 4688 } 4689 pr_cont("\n"); 4690 } 4691 } 4692 4693 /** 4694 * show_workqueue_state - dump workqueue state 4695 * 4696 * Called from a sysrq handler or try_to_freeze_tasks() and prints out 4697 * all busy workqueues and pools. 4698 */ 4699 void show_workqueue_state(void) 4700 { 4701 struct workqueue_struct *wq; 4702 struct worker_pool *pool; 4703 unsigned long flags; 4704 int pi; 4705 4706 rcu_read_lock(); 4707 4708 pr_info("Showing busy workqueues and worker pools:\n"); 4709 4710 list_for_each_entry_rcu(wq, &workqueues, list) { 4711 struct pool_workqueue *pwq; 4712 bool idle = true; 4713 4714 for_each_pwq(pwq, wq) { 4715 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 4716 idle = false; 4717 break; 4718 } 4719 } 4720 if (idle) 4721 continue; 4722 4723 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 4724 4725 for_each_pwq(pwq, wq) { 4726 spin_lock_irqsave(&pwq->pool->lock, flags); 4727 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 4728 show_pwq(pwq); 4729 spin_unlock_irqrestore(&pwq->pool->lock, flags); 4730 /* 4731 * We could be printing a lot from atomic context, e.g. 4732 * sysrq-t -> show_workqueue_state(). Avoid triggering 4733 * hard lockup. 4734 */ 4735 touch_nmi_watchdog(); 4736 } 4737 } 4738 4739 for_each_pool(pool, pi) { 4740 struct worker *worker; 4741 bool first = true; 4742 4743 spin_lock_irqsave(&pool->lock, flags); 4744 if (pool->nr_workers == pool->nr_idle) 4745 goto next_pool; 4746 4747 pr_info("pool %d:", pool->id); 4748 pr_cont_pool_info(pool); 4749 pr_cont(" hung=%us workers=%d", 4750 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 4751 pool->nr_workers); 4752 if (pool->manager) 4753 pr_cont(" manager: %d", 4754 task_pid_nr(pool->manager->task)); 4755 list_for_each_entry(worker, &pool->idle_list, entry) { 4756 pr_cont(" %s%d", first ? "idle: " : "", 4757 task_pid_nr(worker->task)); 4758 first = false; 4759 } 4760 pr_cont("\n"); 4761 next_pool: 4762 spin_unlock_irqrestore(&pool->lock, flags); 4763 /* 4764 * We could be printing a lot from atomic context, e.g. 4765 * sysrq-t -> show_workqueue_state(). Avoid triggering 4766 * hard lockup. 4767 */ 4768 touch_nmi_watchdog(); 4769 } 4770 4771 rcu_read_unlock(); 4772 } 4773 4774 /* used to show worker information through /proc/PID/{comm,stat,status} */ 4775 void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 4776 { 4777 int off; 4778 4779 /* always show the actual comm */ 4780 off = strscpy(buf, task->comm, size); 4781 if (off < 0) 4782 return; 4783 4784 /* stabilize PF_WQ_WORKER and worker pool association */ 4785 mutex_lock(&wq_pool_attach_mutex); 4786 4787 if (task->flags & PF_WQ_WORKER) { 4788 struct worker *worker = kthread_data(task); 4789 struct worker_pool *pool = worker->pool; 4790 4791 if (pool) { 4792 spin_lock_irq(&pool->lock); 4793 /* 4794 * ->desc tracks information (wq name or 4795 * set_worker_desc()) for the latest execution. If 4796 * current, prepend '+', otherwise '-'. 4797 */ 4798 if (worker->desc[0] != '\0') { 4799 if (worker->current_work) 4800 scnprintf(buf + off, size - off, "+%s", 4801 worker->desc); 4802 else 4803 scnprintf(buf + off, size - off, "-%s", 4804 worker->desc); 4805 } 4806 spin_unlock_irq(&pool->lock); 4807 } 4808 } 4809 4810 mutex_unlock(&wq_pool_attach_mutex); 4811 } 4812 4813 #ifdef CONFIG_SMP 4814 4815 /* 4816 * CPU hotplug. 4817 * 4818 * There are two challenges in supporting CPU hotplug. Firstly, there 4819 * are a lot of assumptions on strong associations among work, pwq and 4820 * pool which make migrating pending and scheduled works very 4821 * difficult to implement without impacting hot paths. Secondly, 4822 * worker pools serve mix of short, long and very long running works making 4823 * blocked draining impractical. 4824 * 4825 * This is solved by allowing the pools to be disassociated from the CPU 4826 * running as an unbound one and allowing it to be reattached later if the 4827 * cpu comes back online. 4828 */ 4829 4830 static void unbind_workers(int cpu) 4831 { 4832 struct worker_pool *pool; 4833 struct worker *worker; 4834 4835 for_each_cpu_worker_pool(pool, cpu) { 4836 mutex_lock(&wq_pool_attach_mutex); 4837 spin_lock_irq(&pool->lock); 4838 4839 /* 4840 * We've blocked all attach/detach operations. Make all workers 4841 * unbound and set DISASSOCIATED. Before this, all workers 4842 * except for the ones which are still executing works from 4843 * before the last CPU down must be on the cpu. After 4844 * this, they may become diasporas. 4845 */ 4846 for_each_pool_worker(worker, pool) 4847 worker->flags |= WORKER_UNBOUND; 4848 4849 pool->flags |= POOL_DISASSOCIATED; 4850 4851 spin_unlock_irq(&pool->lock); 4852 mutex_unlock(&wq_pool_attach_mutex); 4853 4854 /* 4855 * Call schedule() so that we cross rq->lock and thus can 4856 * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4857 * This is necessary as scheduler callbacks may be invoked 4858 * from other cpus. 4859 */ 4860 schedule(); 4861 4862 /* 4863 * Sched callbacks are disabled now. Zap nr_running. 4864 * After this, nr_running stays zero and need_more_worker() 4865 * and keep_working() are always true as long as the 4866 * worklist is not empty. This pool now behaves as an 4867 * unbound (in terms of concurrency management) pool which 4868 * are served by workers tied to the pool. 4869 */ 4870 atomic_set(&pool->nr_running, 0); 4871 4872 /* 4873 * With concurrency management just turned off, a busy 4874 * worker blocking could lead to lengthy stalls. Kick off 4875 * unbound chain execution of currently pending work items. 4876 */ 4877 spin_lock_irq(&pool->lock); 4878 wake_up_worker(pool); 4879 spin_unlock_irq(&pool->lock); 4880 } 4881 } 4882 4883 /** 4884 * rebind_workers - rebind all workers of a pool to the associated CPU 4885 * @pool: pool of interest 4886 * 4887 * @pool->cpu is coming online. Rebind all workers to the CPU. 4888 */ 4889 static void rebind_workers(struct worker_pool *pool) 4890 { 4891 struct worker *worker; 4892 4893 lockdep_assert_held(&wq_pool_attach_mutex); 4894 4895 /* 4896 * Restore CPU affinity of all workers. As all idle workers should 4897 * be on the run-queue of the associated CPU before any local 4898 * wake-ups for concurrency management happen, restore CPU affinity 4899 * of all workers first and then clear UNBOUND. As we're called 4900 * from CPU_ONLINE, the following shouldn't fail. 4901 */ 4902 for_each_pool_worker(worker, pool) 4903 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4904 pool->attrs->cpumask) < 0); 4905 4906 spin_lock_irq(&pool->lock); 4907 4908 pool->flags &= ~POOL_DISASSOCIATED; 4909 4910 for_each_pool_worker(worker, pool) { 4911 unsigned int worker_flags = worker->flags; 4912 4913 /* 4914 * A bound idle worker should actually be on the runqueue 4915 * of the associated CPU for local wake-ups targeting it to 4916 * work. Kick all idle workers so that they migrate to the 4917 * associated CPU. Doing this in the same loop as 4918 * replacing UNBOUND with REBOUND is safe as no worker will 4919 * be bound before @pool->lock is released. 4920 */ 4921 if (worker_flags & WORKER_IDLE) 4922 wake_up_process(worker->task); 4923 4924 /* 4925 * We want to clear UNBOUND but can't directly call 4926 * worker_clr_flags() or adjust nr_running. Atomically 4927 * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4928 * @worker will clear REBOUND using worker_clr_flags() when 4929 * it initiates the next execution cycle thus restoring 4930 * concurrency management. Note that when or whether 4931 * @worker clears REBOUND doesn't affect correctness. 4932 * 4933 * WRITE_ONCE() is necessary because @worker->flags may be 4934 * tested without holding any lock in 4935 * wq_worker_running(). Without it, NOT_RUNNING test may 4936 * fail incorrectly leading to premature concurrency 4937 * management operations. 4938 */ 4939 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4940 worker_flags |= WORKER_REBOUND; 4941 worker_flags &= ~WORKER_UNBOUND; 4942 WRITE_ONCE(worker->flags, worker_flags); 4943 } 4944 4945 spin_unlock_irq(&pool->lock); 4946 } 4947 4948 /** 4949 * restore_unbound_workers_cpumask - restore cpumask of unbound workers 4950 * @pool: unbound pool of interest 4951 * @cpu: the CPU which is coming up 4952 * 4953 * An unbound pool may end up with a cpumask which doesn't have any online 4954 * CPUs. When a worker of such pool get scheduled, the scheduler resets 4955 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 4956 * online CPU before, cpus_allowed of all its workers should be restored. 4957 */ 4958 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 4959 { 4960 static cpumask_t cpumask; 4961 struct worker *worker; 4962 4963 lockdep_assert_held(&wq_pool_attach_mutex); 4964 4965 /* is @cpu allowed for @pool? */ 4966 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 4967 return; 4968 4969 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 4970 4971 /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4972 for_each_pool_worker(worker, pool) 4973 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 4974 } 4975 4976 int workqueue_prepare_cpu(unsigned int cpu) 4977 { 4978 struct worker_pool *pool; 4979 4980 for_each_cpu_worker_pool(pool, cpu) { 4981 if (pool->nr_workers) 4982 continue; 4983 if (!create_worker(pool)) 4984 return -ENOMEM; 4985 } 4986 return 0; 4987 } 4988 4989 int workqueue_online_cpu(unsigned int cpu) 4990 { 4991 struct worker_pool *pool; 4992 struct workqueue_struct *wq; 4993 int pi; 4994 4995 mutex_lock(&wq_pool_mutex); 4996 4997 for_each_pool(pool, pi) { 4998 mutex_lock(&wq_pool_attach_mutex); 4999 5000 if (pool->cpu == cpu) 5001 rebind_workers(pool); 5002 else if (pool->cpu < 0) 5003 restore_unbound_workers_cpumask(pool, cpu); 5004 5005 mutex_unlock(&wq_pool_attach_mutex); 5006 } 5007 5008 /* update NUMA affinity of unbound workqueues */ 5009 list_for_each_entry(wq, &workqueues, list) 5010 wq_update_unbound_numa(wq, cpu, true); 5011 5012 mutex_unlock(&wq_pool_mutex); 5013 return 0; 5014 } 5015 5016 int workqueue_offline_cpu(unsigned int cpu) 5017 { 5018 struct workqueue_struct *wq; 5019 5020 /* unbinding per-cpu workers should happen on the local CPU */ 5021 if (WARN_ON(cpu != smp_processor_id())) 5022 return -1; 5023 5024 unbind_workers(cpu); 5025 5026 /* update NUMA affinity of unbound workqueues */ 5027 mutex_lock(&wq_pool_mutex); 5028 list_for_each_entry(wq, &workqueues, list) 5029 wq_update_unbound_numa(wq, cpu, false); 5030 mutex_unlock(&wq_pool_mutex); 5031 5032 return 0; 5033 } 5034 5035 struct work_for_cpu { 5036 struct work_struct work; 5037 long (*fn)(void *); 5038 void *arg; 5039 long ret; 5040 }; 5041 5042 static void work_for_cpu_fn(struct work_struct *work) 5043 { 5044 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5045 5046 wfc->ret = wfc->fn(wfc->arg); 5047 } 5048 5049 /** 5050 * work_on_cpu - run a function in thread context on a particular cpu 5051 * @cpu: the cpu to run on 5052 * @fn: the function to run 5053 * @arg: the function arg 5054 * 5055 * It is up to the caller to ensure that the cpu doesn't go offline. 5056 * The caller must not hold any locks which would prevent @fn from completing. 5057 * 5058 * Return: The value @fn returns. 5059 */ 5060 long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 5061 { 5062 struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 5063 5064 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5065 schedule_work_on(cpu, &wfc.work); 5066 flush_work(&wfc.work); 5067 destroy_work_on_stack(&wfc.work); 5068 return wfc.ret; 5069 } 5070 EXPORT_SYMBOL_GPL(work_on_cpu); 5071 5072 /** 5073 * work_on_cpu_safe - run a function in thread context on a particular cpu 5074 * @cpu: the cpu to run on 5075 * @fn: the function to run 5076 * @arg: the function argument 5077 * 5078 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 5079 * any locks which would prevent @fn from completing. 5080 * 5081 * Return: The value @fn returns. 5082 */ 5083 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 5084 { 5085 long ret = -ENODEV; 5086 5087 get_online_cpus(); 5088 if (cpu_online(cpu)) 5089 ret = work_on_cpu(cpu, fn, arg); 5090 put_online_cpus(); 5091 return ret; 5092 } 5093 EXPORT_SYMBOL_GPL(work_on_cpu_safe); 5094 #endif /* CONFIG_SMP */ 5095 5096 #ifdef CONFIG_FREEZER 5097 5098 /** 5099 * freeze_workqueues_begin - begin freezing workqueues 5100 * 5101 * Start freezing workqueues. After this function returns, all freezable 5102 * workqueues will queue new works to their delayed_works list instead of 5103 * pool->worklist. 5104 * 5105 * CONTEXT: 5106 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5107 */ 5108 void freeze_workqueues_begin(void) 5109 { 5110 struct workqueue_struct *wq; 5111 struct pool_workqueue *pwq; 5112 5113 mutex_lock(&wq_pool_mutex); 5114 5115 WARN_ON_ONCE(workqueue_freezing); 5116 workqueue_freezing = true; 5117 5118 list_for_each_entry(wq, &workqueues, list) { 5119 mutex_lock(&wq->mutex); 5120 for_each_pwq(pwq, wq) 5121 pwq_adjust_max_active(pwq); 5122 mutex_unlock(&wq->mutex); 5123 } 5124 5125 mutex_unlock(&wq_pool_mutex); 5126 } 5127 5128 /** 5129 * freeze_workqueues_busy - are freezable workqueues still busy? 5130 * 5131 * Check whether freezing is complete. This function must be called 5132 * between freeze_workqueues_begin() and thaw_workqueues(). 5133 * 5134 * CONTEXT: 5135 * Grabs and releases wq_pool_mutex. 5136 * 5137 * Return: 5138 * %true if some freezable workqueues are still busy. %false if freezing 5139 * is complete. 5140 */ 5141 bool freeze_workqueues_busy(void) 5142 { 5143 bool busy = false; 5144 struct workqueue_struct *wq; 5145 struct pool_workqueue *pwq; 5146 5147 mutex_lock(&wq_pool_mutex); 5148 5149 WARN_ON_ONCE(!workqueue_freezing); 5150 5151 list_for_each_entry(wq, &workqueues, list) { 5152 if (!(wq->flags & WQ_FREEZABLE)) 5153 continue; 5154 /* 5155 * nr_active is monotonically decreasing. It's safe 5156 * to peek without lock. 5157 */ 5158 rcu_read_lock(); 5159 for_each_pwq(pwq, wq) { 5160 WARN_ON_ONCE(pwq->nr_active < 0); 5161 if (pwq->nr_active) { 5162 busy = true; 5163 rcu_read_unlock(); 5164 goto out_unlock; 5165 } 5166 } 5167 rcu_read_unlock(); 5168 } 5169 out_unlock: 5170 mutex_unlock(&wq_pool_mutex); 5171 return busy; 5172 } 5173 5174 /** 5175 * thaw_workqueues - thaw workqueues 5176 * 5177 * Thaw workqueues. Normal queueing is restored and all collected 5178 * frozen works are transferred to their respective pool worklists. 5179 * 5180 * CONTEXT: 5181 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5182 */ 5183 void thaw_workqueues(void) 5184 { 5185 struct workqueue_struct *wq; 5186 struct pool_workqueue *pwq; 5187 5188 mutex_lock(&wq_pool_mutex); 5189 5190 if (!workqueue_freezing) 5191 goto out_unlock; 5192 5193 workqueue_freezing = false; 5194 5195 /* restore max_active and repopulate worklist */ 5196 list_for_each_entry(wq, &workqueues, list) { 5197 mutex_lock(&wq->mutex); 5198 for_each_pwq(pwq, wq) 5199 pwq_adjust_max_active(pwq); 5200 mutex_unlock(&wq->mutex); 5201 } 5202 5203 out_unlock: 5204 mutex_unlock(&wq_pool_mutex); 5205 } 5206 #endif /* CONFIG_FREEZER */ 5207 5208 static int workqueue_apply_unbound_cpumask(void) 5209 { 5210 LIST_HEAD(ctxs); 5211 int ret = 0; 5212 struct workqueue_struct *wq; 5213 struct apply_wqattrs_ctx *ctx, *n; 5214 5215 lockdep_assert_held(&wq_pool_mutex); 5216 5217 list_for_each_entry(wq, &workqueues, list) { 5218 if (!(wq->flags & WQ_UNBOUND)) 5219 continue; 5220 /* creating multiple pwqs breaks ordering guarantee */ 5221 if (wq->flags & __WQ_ORDERED) 5222 continue; 5223 5224 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5225 if (!ctx) { 5226 ret = -ENOMEM; 5227 break; 5228 } 5229 5230 list_add_tail(&ctx->list, &ctxs); 5231 } 5232 5233 list_for_each_entry_safe(ctx, n, &ctxs, list) { 5234 if (!ret) 5235 apply_wqattrs_commit(ctx); 5236 apply_wqattrs_cleanup(ctx); 5237 } 5238 5239 return ret; 5240 } 5241 5242 /** 5243 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5244 * @cpumask: the cpumask to set 5245 * 5246 * The low-level workqueues cpumask is a global cpumask that limits 5247 * the affinity of all unbound workqueues. This function check the @cpumask 5248 * and apply it to all unbound workqueues and updates all pwqs of them. 5249 * 5250 * Retun: 0 - Success 5251 * -EINVAL - Invalid @cpumask 5252 * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5253 */ 5254 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5255 { 5256 int ret = -EINVAL; 5257 cpumask_var_t saved_cpumask; 5258 5259 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 5260 return -ENOMEM; 5261 5262 /* 5263 * Not excluding isolated cpus on purpose. 5264 * If the user wishes to include them, we allow that. 5265 */ 5266 cpumask_and(cpumask, cpumask, cpu_possible_mask); 5267 if (!cpumask_empty(cpumask)) { 5268 apply_wqattrs_lock(); 5269 5270 /* save the old wq_unbound_cpumask. */ 5271 cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5272 5273 /* update wq_unbound_cpumask at first and apply it to wqs. */ 5274 cpumask_copy(wq_unbound_cpumask, cpumask); 5275 ret = workqueue_apply_unbound_cpumask(); 5276 5277 /* restore the wq_unbound_cpumask when failed. */ 5278 if (ret < 0) 5279 cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5280 5281 apply_wqattrs_unlock(); 5282 } 5283 5284 free_cpumask_var(saved_cpumask); 5285 return ret; 5286 } 5287 5288 #ifdef CONFIG_SYSFS 5289 /* 5290 * Workqueues with WQ_SYSFS flag set is visible to userland via 5291 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 5292 * following attributes. 5293 * 5294 * per_cpu RO bool : whether the workqueue is per-cpu or unbound 5295 * max_active RW int : maximum number of in-flight work items 5296 * 5297 * Unbound workqueues have the following extra attributes. 5298 * 5299 * pool_ids RO int : the associated pool IDs for each node 5300 * nice RW int : nice value of the workers 5301 * cpumask RW mask : bitmask of allowed CPUs for the workers 5302 * numa RW bool : whether enable NUMA affinity 5303 */ 5304 struct wq_device { 5305 struct workqueue_struct *wq; 5306 struct device dev; 5307 }; 5308 5309 static struct workqueue_struct *dev_to_wq(struct device *dev) 5310 { 5311 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 5312 5313 return wq_dev->wq; 5314 } 5315 5316 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 5317 char *buf) 5318 { 5319 struct workqueue_struct *wq = dev_to_wq(dev); 5320 5321 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 5322 } 5323 static DEVICE_ATTR_RO(per_cpu); 5324 5325 static ssize_t max_active_show(struct device *dev, 5326 struct device_attribute *attr, char *buf) 5327 { 5328 struct workqueue_struct *wq = dev_to_wq(dev); 5329 5330 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 5331 } 5332 5333 static ssize_t max_active_store(struct device *dev, 5334 struct device_attribute *attr, const char *buf, 5335 size_t count) 5336 { 5337 struct workqueue_struct *wq = dev_to_wq(dev); 5338 int val; 5339 5340 if (sscanf(buf, "%d", &val) != 1 || val <= 0) 5341 return -EINVAL; 5342 5343 workqueue_set_max_active(wq, val); 5344 return count; 5345 } 5346 static DEVICE_ATTR_RW(max_active); 5347 5348 static struct attribute *wq_sysfs_attrs[] = { 5349 &dev_attr_per_cpu.attr, 5350 &dev_attr_max_active.attr, 5351 NULL, 5352 }; 5353 ATTRIBUTE_GROUPS(wq_sysfs); 5354 5355 static ssize_t wq_pool_ids_show(struct device *dev, 5356 struct device_attribute *attr, char *buf) 5357 { 5358 struct workqueue_struct *wq = dev_to_wq(dev); 5359 const char *delim = ""; 5360 int node, written = 0; 5361 5362 get_online_cpus(); 5363 rcu_read_lock(); 5364 for_each_node(node) { 5365 written += scnprintf(buf + written, PAGE_SIZE - written, 5366 "%s%d:%d", delim, node, 5367 unbound_pwq_by_node(wq, node)->pool->id); 5368 delim = " "; 5369 } 5370 written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 5371 rcu_read_unlock(); 5372 put_online_cpus(); 5373 5374 return written; 5375 } 5376 5377 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 5378 char *buf) 5379 { 5380 struct workqueue_struct *wq = dev_to_wq(dev); 5381 int written; 5382 5383 mutex_lock(&wq->mutex); 5384 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 5385 mutex_unlock(&wq->mutex); 5386 5387 return written; 5388 } 5389 5390 /* prepare workqueue_attrs for sysfs store operations */ 5391 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 5392 { 5393 struct workqueue_attrs *attrs; 5394 5395 lockdep_assert_held(&wq_pool_mutex); 5396 5397 attrs = alloc_workqueue_attrs(GFP_KERNEL); 5398 if (!attrs) 5399 return NULL; 5400 5401 copy_workqueue_attrs(attrs, wq->unbound_attrs); 5402 return attrs; 5403 } 5404 5405 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 5406 const char *buf, size_t count) 5407 { 5408 struct workqueue_struct *wq = dev_to_wq(dev); 5409 struct workqueue_attrs *attrs; 5410 int ret = -ENOMEM; 5411 5412 apply_wqattrs_lock(); 5413 5414 attrs = wq_sysfs_prep_attrs(wq); 5415 if (!attrs) 5416 goto out_unlock; 5417 5418 if (sscanf(buf, "%d", &attrs->nice) == 1 && 5419 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5420 ret = apply_workqueue_attrs_locked(wq, attrs); 5421 else 5422 ret = -EINVAL; 5423 5424 out_unlock: 5425 apply_wqattrs_unlock(); 5426 free_workqueue_attrs(attrs); 5427 return ret ?: count; 5428 } 5429 5430 static ssize_t wq_cpumask_show(struct device *dev, 5431 struct device_attribute *attr, char *buf) 5432 { 5433 struct workqueue_struct *wq = dev_to_wq(dev); 5434 int written; 5435 5436 mutex_lock(&wq->mutex); 5437 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5438 cpumask_pr_args(wq->unbound_attrs->cpumask)); 5439 mutex_unlock(&wq->mutex); 5440 return written; 5441 } 5442 5443 static ssize_t wq_cpumask_store(struct device *dev, 5444 struct device_attribute *attr, 5445 const char *buf, size_t count) 5446 { 5447 struct workqueue_struct *wq = dev_to_wq(dev); 5448 struct workqueue_attrs *attrs; 5449 int ret = -ENOMEM; 5450 5451 apply_wqattrs_lock(); 5452 5453 attrs = wq_sysfs_prep_attrs(wq); 5454 if (!attrs) 5455 goto out_unlock; 5456 5457 ret = cpumask_parse(buf, attrs->cpumask); 5458 if (!ret) 5459 ret = apply_workqueue_attrs_locked(wq, attrs); 5460 5461 out_unlock: 5462 apply_wqattrs_unlock(); 5463 free_workqueue_attrs(attrs); 5464 return ret ?: count; 5465 } 5466 5467 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 5468 char *buf) 5469 { 5470 struct workqueue_struct *wq = dev_to_wq(dev); 5471 int written; 5472 5473 mutex_lock(&wq->mutex); 5474 written = scnprintf(buf, PAGE_SIZE, "%d\n", 5475 !wq->unbound_attrs->no_numa); 5476 mutex_unlock(&wq->mutex); 5477 5478 return written; 5479 } 5480 5481 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 5482 const char *buf, size_t count) 5483 { 5484 struct workqueue_struct *wq = dev_to_wq(dev); 5485 struct workqueue_attrs *attrs; 5486 int v, ret = -ENOMEM; 5487 5488 apply_wqattrs_lock(); 5489 5490 attrs = wq_sysfs_prep_attrs(wq); 5491 if (!attrs) 5492 goto out_unlock; 5493 5494 ret = -EINVAL; 5495 if (sscanf(buf, "%d", &v) == 1) { 5496 attrs->no_numa = !v; 5497 ret = apply_workqueue_attrs_locked(wq, attrs); 5498 } 5499 5500 out_unlock: 5501 apply_wqattrs_unlock(); 5502 free_workqueue_attrs(attrs); 5503 return ret ?: count; 5504 } 5505 5506 static struct device_attribute wq_sysfs_unbound_attrs[] = { 5507 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 5508 __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 5509 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 5510 __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 5511 __ATTR_NULL, 5512 }; 5513 5514 static struct bus_type wq_subsys = { 5515 .name = "workqueue", 5516 .dev_groups = wq_sysfs_groups, 5517 }; 5518 5519 static ssize_t wq_unbound_cpumask_show(struct device *dev, 5520 struct device_attribute *attr, char *buf) 5521 { 5522 int written; 5523 5524 mutex_lock(&wq_pool_mutex); 5525 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5526 cpumask_pr_args(wq_unbound_cpumask)); 5527 mutex_unlock(&wq_pool_mutex); 5528 5529 return written; 5530 } 5531 5532 static ssize_t wq_unbound_cpumask_store(struct device *dev, 5533 struct device_attribute *attr, const char *buf, size_t count) 5534 { 5535 cpumask_var_t cpumask; 5536 int ret; 5537 5538 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5539 return -ENOMEM; 5540 5541 ret = cpumask_parse(buf, cpumask); 5542 if (!ret) 5543 ret = workqueue_set_unbound_cpumask(cpumask); 5544 5545 free_cpumask_var(cpumask); 5546 return ret ? ret : count; 5547 } 5548 5549 static struct device_attribute wq_sysfs_cpumask_attr = 5550 __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5551 wq_unbound_cpumask_store); 5552 5553 static int __init wq_sysfs_init(void) 5554 { 5555 int err; 5556 5557 err = subsys_virtual_register(&wq_subsys, NULL); 5558 if (err) 5559 return err; 5560 5561 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 5562 } 5563 core_initcall(wq_sysfs_init); 5564 5565 static void wq_device_release(struct device *dev) 5566 { 5567 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 5568 5569 kfree(wq_dev); 5570 } 5571 5572 /** 5573 * workqueue_sysfs_register - make a workqueue visible in sysfs 5574 * @wq: the workqueue to register 5575 * 5576 * Expose @wq in sysfs under /sys/bus/workqueue/devices. 5577 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 5578 * which is the preferred method. 5579 * 5580 * Workqueue user should use this function directly iff it wants to apply 5581 * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 5582 * apply_workqueue_attrs() may race against userland updating the 5583 * attributes. 5584 * 5585 * Return: 0 on success, -errno on failure. 5586 */ 5587 int workqueue_sysfs_register(struct workqueue_struct *wq) 5588 { 5589 struct wq_device *wq_dev; 5590 int ret; 5591 5592 /* 5593 * Adjusting max_active or creating new pwqs by applying 5594 * attributes breaks ordering guarantee. Disallow exposing ordered 5595 * workqueues. 5596 */ 5597 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 5598 return -EINVAL; 5599 5600 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 5601 if (!wq_dev) 5602 return -ENOMEM; 5603 5604 wq_dev->wq = wq; 5605 wq_dev->dev.bus = &wq_subsys; 5606 wq_dev->dev.release = wq_device_release; 5607 dev_set_name(&wq_dev->dev, "%s", wq->name); 5608 5609 /* 5610 * unbound_attrs are created separately. Suppress uevent until 5611 * everything is ready. 5612 */ 5613 dev_set_uevent_suppress(&wq_dev->dev, true); 5614 5615 ret = device_register(&wq_dev->dev); 5616 if (ret) { 5617 put_device(&wq_dev->dev); 5618 wq->wq_dev = NULL; 5619 return ret; 5620 } 5621 5622 if (wq->flags & WQ_UNBOUND) { 5623 struct device_attribute *attr; 5624 5625 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 5626 ret = device_create_file(&wq_dev->dev, attr); 5627 if (ret) { 5628 device_unregister(&wq_dev->dev); 5629 wq->wq_dev = NULL; 5630 return ret; 5631 } 5632 } 5633 } 5634 5635 dev_set_uevent_suppress(&wq_dev->dev, false); 5636 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 5637 return 0; 5638 } 5639 5640 /** 5641 * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 5642 * @wq: the workqueue to unregister 5643 * 5644 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 5645 */ 5646 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 5647 { 5648 struct wq_device *wq_dev = wq->wq_dev; 5649 5650 if (!wq->wq_dev) 5651 return; 5652 5653 wq->wq_dev = NULL; 5654 device_unregister(&wq_dev->dev); 5655 } 5656 #else /* CONFIG_SYSFS */ 5657 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 5658 #endif /* CONFIG_SYSFS */ 5659 5660 /* 5661 * Workqueue watchdog. 5662 * 5663 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 5664 * flush dependency, a concurrency managed work item which stays RUNNING 5665 * indefinitely. Workqueue stalls can be very difficult to debug as the 5666 * usual warning mechanisms don't trigger and internal workqueue state is 5667 * largely opaque. 5668 * 5669 * Workqueue watchdog monitors all worker pools periodically and dumps 5670 * state if some pools failed to make forward progress for a while where 5671 * forward progress is defined as the first item on ->worklist changing. 5672 * 5673 * This mechanism is controlled through the kernel parameter 5674 * "workqueue.watchdog_thresh" which can be updated at runtime through the 5675 * corresponding sysfs parameter file. 5676 */ 5677 #ifdef CONFIG_WQ_WATCHDOG 5678 5679 static unsigned long wq_watchdog_thresh = 30; 5680 static struct timer_list wq_watchdog_timer; 5681 5682 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 5683 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 5684 5685 static void wq_watchdog_reset_touched(void) 5686 { 5687 int cpu; 5688 5689 wq_watchdog_touched = jiffies; 5690 for_each_possible_cpu(cpu) 5691 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 5692 } 5693 5694 static void wq_watchdog_timer_fn(struct timer_list *unused) 5695 { 5696 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 5697 bool lockup_detected = false; 5698 struct worker_pool *pool; 5699 int pi; 5700 5701 if (!thresh) 5702 return; 5703 5704 rcu_read_lock(); 5705 5706 for_each_pool(pool, pi) { 5707 unsigned long pool_ts, touched, ts; 5708 5709 if (list_empty(&pool->worklist)) 5710 continue; 5711 5712 /* get the latest of pool and touched timestamps */ 5713 pool_ts = READ_ONCE(pool->watchdog_ts); 5714 touched = READ_ONCE(wq_watchdog_touched); 5715 5716 if (time_after(pool_ts, touched)) 5717 ts = pool_ts; 5718 else 5719 ts = touched; 5720 5721 if (pool->cpu >= 0) { 5722 unsigned long cpu_touched = 5723 READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 5724 pool->cpu)); 5725 if (time_after(cpu_touched, ts)) 5726 ts = cpu_touched; 5727 } 5728 5729 /* did we stall? */ 5730 if (time_after(jiffies, ts + thresh)) { 5731 lockup_detected = true; 5732 pr_emerg("BUG: workqueue lockup - pool"); 5733 pr_cont_pool_info(pool); 5734 pr_cont(" stuck for %us!\n", 5735 jiffies_to_msecs(jiffies - pool_ts) / 1000); 5736 } 5737 } 5738 5739 rcu_read_unlock(); 5740 5741 if (lockup_detected) 5742 show_workqueue_state(); 5743 5744 wq_watchdog_reset_touched(); 5745 mod_timer(&wq_watchdog_timer, jiffies + thresh); 5746 } 5747 5748 notrace void wq_watchdog_touch(int cpu) 5749 { 5750 if (cpu >= 0) 5751 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 5752 else 5753 wq_watchdog_touched = jiffies; 5754 } 5755 5756 static void wq_watchdog_set_thresh(unsigned long thresh) 5757 { 5758 wq_watchdog_thresh = 0; 5759 del_timer_sync(&wq_watchdog_timer); 5760 5761 if (thresh) { 5762 wq_watchdog_thresh = thresh; 5763 wq_watchdog_reset_touched(); 5764 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 5765 } 5766 } 5767 5768 static int wq_watchdog_param_set_thresh(const char *val, 5769 const struct kernel_param *kp) 5770 { 5771 unsigned long thresh; 5772 int ret; 5773 5774 ret = kstrtoul(val, 0, &thresh); 5775 if (ret) 5776 return ret; 5777 5778 if (system_wq) 5779 wq_watchdog_set_thresh(thresh); 5780 else 5781 wq_watchdog_thresh = thresh; 5782 5783 return 0; 5784 } 5785 5786 static const struct kernel_param_ops wq_watchdog_thresh_ops = { 5787 .set = wq_watchdog_param_set_thresh, 5788 .get = param_get_ulong, 5789 }; 5790 5791 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 5792 0644); 5793 5794 static void wq_watchdog_init(void) 5795 { 5796 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 5797 wq_watchdog_set_thresh(wq_watchdog_thresh); 5798 } 5799 5800 #else /* CONFIG_WQ_WATCHDOG */ 5801 5802 static inline void wq_watchdog_init(void) { } 5803 5804 #endif /* CONFIG_WQ_WATCHDOG */ 5805 5806 static void __init wq_numa_init(void) 5807 { 5808 cpumask_var_t *tbl; 5809 int node, cpu; 5810 5811 if (num_possible_nodes() <= 1) 5812 return; 5813 5814 if (wq_disable_numa) { 5815 pr_info("workqueue: NUMA affinity support disabled\n"); 5816 return; 5817 } 5818 5819 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 5820 BUG_ON(!wq_update_unbound_numa_attrs_buf); 5821 5822 /* 5823 * We want masks of possible CPUs of each node which isn't readily 5824 * available. Build one from cpu_to_node() which should have been 5825 * fully initialized by now. 5826 */ 5827 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5828 BUG_ON(!tbl); 5829 5830 for_each_node(node) 5831 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 5832 node_online(node) ? node : NUMA_NO_NODE)); 5833 5834 for_each_possible_cpu(cpu) { 5835 node = cpu_to_node(cpu); 5836 if (WARN_ON(node == NUMA_NO_NODE)) { 5837 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5838 /* happens iff arch is bonkers, let's just proceed */ 5839 return; 5840 } 5841 cpumask_set_cpu(cpu, tbl[node]); 5842 } 5843 5844 wq_numa_possible_cpumask = tbl; 5845 wq_numa_enabled = true; 5846 } 5847 5848 /** 5849 * workqueue_init_early - early init for workqueue subsystem 5850 * 5851 * This is the first half of two-staged workqueue subsystem initialization 5852 * and invoked as soon as the bare basics - memory allocation, cpumasks and 5853 * idr are up. It sets up all the data structures and system workqueues 5854 * and allows early boot code to create workqueues and queue/cancel work 5855 * items. Actual work item execution starts only after kthreads can be 5856 * created and scheduled right before early initcalls. 5857 */ 5858 int __init workqueue_init_early(void) 5859 { 5860 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 5861 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 5862 int i, cpu; 5863 5864 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5865 5866 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5867 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 5868 5869 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5870 5871 /* initialize CPU pools */ 5872 for_each_possible_cpu(cpu) { 5873 struct worker_pool *pool; 5874 5875 i = 0; 5876 for_each_cpu_worker_pool(pool, cpu) { 5877 BUG_ON(init_worker_pool(pool)); 5878 pool->cpu = cpu; 5879 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 5880 pool->attrs->nice = std_nice[i++]; 5881 pool->node = cpu_to_node(cpu); 5882 5883 /* alloc pool ID */ 5884 mutex_lock(&wq_pool_mutex); 5885 BUG_ON(worker_pool_assign_id(pool)); 5886 mutex_unlock(&wq_pool_mutex); 5887 } 5888 } 5889 5890 /* create default unbound and ordered wq attrs */ 5891 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 5892 struct workqueue_attrs *attrs; 5893 5894 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 5895 attrs->nice = std_nice[i]; 5896 unbound_std_wq_attrs[i] = attrs; 5897 5898 /* 5899 * An ordered wq should have only one pwq as ordering is 5900 * guaranteed by max_active which is enforced by pwqs. 5901 * Turn off NUMA so that dfl_pwq is used for all nodes. 5902 */ 5903 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 5904 attrs->nice = std_nice[i]; 5905 attrs->no_numa = true; 5906 ordered_wq_attrs[i] = attrs; 5907 } 5908 5909 system_wq = alloc_workqueue("events", 0, 0); 5910 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5911 system_long_wq = alloc_workqueue("events_long", 0, 0); 5912 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5913 WQ_UNBOUND_MAX_ACTIVE); 5914 system_freezable_wq = alloc_workqueue("events_freezable", 5915 WQ_FREEZABLE, 0); 5916 system_power_efficient_wq = alloc_workqueue("events_power_efficient", 5917 WQ_POWER_EFFICIENT, 0); 5918 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 5919 WQ_FREEZABLE | WQ_POWER_EFFICIENT, 5920 0); 5921 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 5922 !system_unbound_wq || !system_freezable_wq || 5923 !system_power_efficient_wq || 5924 !system_freezable_power_efficient_wq); 5925 5926 return 0; 5927 } 5928 5929 /** 5930 * workqueue_init - bring workqueue subsystem fully online 5931 * 5932 * This is the latter half of two-staged workqueue subsystem initialization 5933 * and invoked as soon as kthreads can be created and scheduled. 5934 * Workqueues have been created and work items queued on them, but there 5935 * are no kworkers executing the work items yet. Populate the worker pools 5936 * with the initial workers and enable future kworker creations. 5937 */ 5938 int __init workqueue_init(void) 5939 { 5940 struct workqueue_struct *wq; 5941 struct worker_pool *pool; 5942 int cpu, bkt; 5943 5944 /* 5945 * It'd be simpler to initialize NUMA in workqueue_init_early() but 5946 * CPU to node mapping may not be available that early on some 5947 * archs such as power and arm64. As per-cpu pools created 5948 * previously could be missing node hint and unbound pools NUMA 5949 * affinity, fix them up. 5950 * 5951 * Also, while iterating workqueues, create rescuers if requested. 5952 */ 5953 wq_numa_init(); 5954 5955 mutex_lock(&wq_pool_mutex); 5956 5957 for_each_possible_cpu(cpu) { 5958 for_each_cpu_worker_pool(pool, cpu) { 5959 pool->node = cpu_to_node(cpu); 5960 } 5961 } 5962 5963 list_for_each_entry(wq, &workqueues, list) { 5964 wq_update_unbound_numa(wq, smp_processor_id(), true); 5965 WARN(init_rescuer(wq), 5966 "workqueue: failed to create early rescuer for %s", 5967 wq->name); 5968 } 5969 5970 mutex_unlock(&wq_pool_mutex); 5971 5972 /* create the initial workers */ 5973 for_each_online_cpu(cpu) { 5974 for_each_cpu_worker_pool(pool, cpu) { 5975 pool->flags &= ~POOL_DISASSOCIATED; 5976 BUG_ON(!create_worker(pool)); 5977 } 5978 } 5979 5980 hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 5981 BUG_ON(!create_worker(pool)); 5982 5983 wq_online = true; 5984 wq_watchdog_init(); 5985 5986 return 0; 5987 } 5988