1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * kernel/workqueue.c - generic async execution with shared worker pool 4 * 5 * Copyright (C) 2002 Ingo Molnar 6 * 7 * Derived from the taskqueue/keventd code by: 8 * David Woodhouse <dwmw2@infradead.org> 9 * Andrew Morton 10 * Kai Petzke <wpp@marie.physik.tu-berlin.de> 11 * Theodore Ts'o <tytso@mit.edu> 12 * 13 * Made to use alloc_percpu by Christoph Lameter. 14 * 15 * Copyright (C) 2010 SUSE Linux Products GmbH 16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 17 * 18 * This is the generic async execution mechanism. Work items as are 19 * executed in process context. The worker pool is shared and 20 * automatically managed. There are two worker pools for each CPU (one for 21 * normal work items and the other for high priority ones) and some extra 22 * pools for workqueues which are not bound to any specific CPU - the 23 * number of these backing pools is dynamic. 24 * 25 * Please read Documentation/core-api/workqueue.rst for details. 26 */ 27 28 #include <linux/export.h> 29 #include <linux/kernel.h> 30 #include <linux/sched.h> 31 #include <linux/init.h> 32 #include <linux/signal.h> 33 #include <linux/completion.h> 34 #include <linux/workqueue.h> 35 #include <linux/slab.h> 36 #include <linux/cpu.h> 37 #include <linux/notifier.h> 38 #include <linux/kthread.h> 39 #include <linux/hardirq.h> 40 #include <linux/mempolicy.h> 41 #include <linux/freezer.h> 42 #include <linux/debug_locks.h> 43 #include <linux/lockdep.h> 44 #include <linux/idr.h> 45 #include <linux/jhash.h> 46 #include <linux/hashtable.h> 47 #include <linux/rculist.h> 48 #include <linux/nodemask.h> 49 #include <linux/moduleparam.h> 50 #include <linux/uaccess.h> 51 #include <linux/sched/isolation.h> 52 #include <linux/sched/debug.h> 53 #include <linux/nmi.h> 54 #include <linux/kvm_para.h> 55 #include <linux/delay.h> 56 57 #include "workqueue_internal.h" 58 59 enum { 60 /* 61 * worker_pool flags 62 * 63 * A bound pool is either associated or disassociated with its CPU. 64 * While associated (!DISASSOCIATED), all workers are bound to the 65 * CPU and none has %WORKER_UNBOUND set and concurrency management 66 * is in effect. 67 * 68 * While DISASSOCIATED, the cpu may be offline and all workers have 69 * %WORKER_UNBOUND set and concurrency management disabled, and may 70 * be executing on any CPU. The pool behaves as an unbound one. 71 * 72 * Note that DISASSOCIATED should be flipped only while holding 73 * wq_pool_attach_mutex to avoid changing binding state while 74 * worker_attach_to_pool() is in progress. 75 */ 76 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 77 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 78 79 /* worker flags */ 80 WORKER_DIE = 1 << 1, /* die die die */ 81 WORKER_IDLE = 1 << 2, /* is idle */ 82 WORKER_PREP = 1 << 3, /* preparing to run works */ 83 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 84 WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 85 WORKER_REBOUND = 1 << 8, /* worker was rebound */ 86 87 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 88 WORKER_UNBOUND | WORKER_REBOUND, 89 90 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 91 92 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 93 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 94 95 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 96 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 97 98 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 99 /* call for help after 10ms 100 (min two ticks) */ 101 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 102 CREATE_COOLDOWN = HZ, /* time to breath after fail */ 103 104 /* 105 * Rescue workers are used only on emergencies and shared by 106 * all cpus. Give MIN_NICE. 107 */ 108 RESCUER_NICE_LEVEL = MIN_NICE, 109 HIGHPRI_NICE_LEVEL = MIN_NICE, 110 111 WQ_NAME_LEN = 32, 112 }; 113 114 /* 115 * Structure fields follow one of the following exclusion rules. 116 * 117 * I: Modifiable by initialization/destruction paths and read-only for 118 * everyone else. 119 * 120 * P: Preemption protected. Disabling preemption is enough and should 121 * only be modified and accessed from the local cpu. 122 * 123 * L: pool->lock protected. Access with pool->lock held. 124 * 125 * K: Only modified by worker while holding pool->lock. Can be safely read by 126 * self, while holding pool->lock or from IRQ context if %current is the 127 * kworker. 128 * 129 * S: Only modified by worker self. 130 * 131 * A: wq_pool_attach_mutex protected. 132 * 133 * PL: wq_pool_mutex protected. 134 * 135 * PR: wq_pool_mutex protected for writes. RCU protected for reads. 136 * 137 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 138 * 139 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 140 * RCU for reads. 141 * 142 * WQ: wq->mutex protected. 143 * 144 * WR: wq->mutex protected for writes. RCU protected for reads. 145 * 146 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read 147 * with READ_ONCE() without locking. 148 * 149 * MD: wq_mayday_lock protected. 150 * 151 * WD: Used internally by the watchdog. 152 */ 153 154 /* struct worker is defined in workqueue_internal.h */ 155 156 struct worker_pool { 157 raw_spinlock_t lock; /* the pool lock */ 158 int cpu; /* I: the associated cpu */ 159 int node; /* I: the associated node ID */ 160 int id; /* I: pool ID */ 161 unsigned int flags; /* L: flags */ 162 163 unsigned long watchdog_ts; /* L: watchdog timestamp */ 164 bool cpu_stall; /* WD: stalled cpu bound pool */ 165 166 /* 167 * The counter is incremented in a process context on the associated CPU 168 * w/ preemption disabled, and decremented or reset in the same context 169 * but w/ pool->lock held. The readers grab pool->lock and are 170 * guaranteed to see if the counter reached zero. 171 */ 172 int nr_running; 173 174 struct list_head worklist; /* L: list of pending works */ 175 176 int nr_workers; /* L: total number of workers */ 177 int nr_idle; /* L: currently idle workers */ 178 179 struct list_head idle_list; /* L: list of idle workers */ 180 struct timer_list idle_timer; /* L: worker idle timeout */ 181 struct work_struct idle_cull_work; /* L: worker idle cleanup */ 182 183 struct timer_list mayday_timer; /* L: SOS timer for workers */ 184 185 /* a workers is either on busy_hash or idle_list, or the manager */ 186 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 187 /* L: hash of busy workers */ 188 189 struct worker *manager; /* L: purely informational */ 190 struct list_head workers; /* A: attached workers */ 191 struct list_head dying_workers; /* A: workers about to die */ 192 struct completion *detach_completion; /* all workers detached */ 193 194 struct ida worker_ida; /* worker IDs for task name */ 195 196 struct workqueue_attrs *attrs; /* I: worker attributes */ 197 struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 198 int refcnt; /* PL: refcnt for unbound pools */ 199 200 /* 201 * Destruction of pool is RCU protected to allow dereferences 202 * from get_work_pool(). 203 */ 204 struct rcu_head rcu; 205 }; 206 207 /* 208 * Per-pool_workqueue statistics. These can be monitored using 209 * tools/workqueue/wq_monitor.py. 210 */ 211 enum pool_workqueue_stats { 212 PWQ_STAT_STARTED, /* work items started execution */ 213 PWQ_STAT_COMPLETED, /* work items completed execution */ 214 PWQ_STAT_CPU_TIME, /* total CPU time consumed */ 215 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ 216 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 217 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ 218 PWQ_STAT_MAYDAY, /* maydays to rescuer */ 219 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 220 221 PWQ_NR_STATS, 222 }; 223 224 /* 225 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 226 * of work_struct->data are used for flags and the remaining high bits 227 * point to the pwq; thus, pwqs need to be aligned at two's power of the 228 * number of flag bits. 229 */ 230 struct pool_workqueue { 231 struct worker_pool *pool; /* I: the associated pool */ 232 struct workqueue_struct *wq; /* I: the owning workqueue */ 233 int work_color; /* L: current color */ 234 int flush_color; /* L: flushing color */ 235 int refcnt; /* L: reference count */ 236 int nr_in_flight[WORK_NR_COLORS]; 237 /* L: nr of in_flight works */ 238 239 /* 240 * nr_active management and WORK_STRUCT_INACTIVE: 241 * 242 * When pwq->nr_active >= max_active, new work item is queued to 243 * pwq->inactive_works instead of pool->worklist and marked with 244 * WORK_STRUCT_INACTIVE. 245 * 246 * All work items marked with WORK_STRUCT_INACTIVE do not participate 247 * in pwq->nr_active and all work items in pwq->inactive_works are 248 * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 249 * work items are in pwq->inactive_works. Some of them are ready to 250 * run in pool->worklist or worker->scheduled. Those work itmes are 251 * only struct wq_barrier which is used for flush_work() and should 252 * not participate in pwq->nr_active. For non-barrier work item, it 253 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 254 */ 255 int nr_active; /* L: nr of active works */ 256 struct list_head inactive_works; /* L: inactive works */ 257 struct list_head pwqs_node; /* WR: node on wq->pwqs */ 258 struct list_head mayday_node; /* MD: node on wq->maydays */ 259 260 u64 stats[PWQ_NR_STATS]; 261 262 /* 263 * Release of unbound pwq is punted to a kthread_worker. See put_pwq() 264 * and pwq_release_workfn() for details. pool_workqueue itself is also 265 * RCU protected so that the first pwq can be determined without 266 * grabbing wq->mutex. 267 */ 268 struct kthread_work release_work; 269 struct rcu_head rcu; 270 } __aligned(1 << WORK_STRUCT_FLAG_BITS); 271 272 /* 273 * Structure used to wait for workqueue flush. 274 */ 275 struct wq_flusher { 276 struct list_head list; /* WQ: list of flushers */ 277 int flush_color; /* WQ: flush color waiting for */ 278 struct completion done; /* flush completion */ 279 }; 280 281 struct wq_device; 282 283 /* 284 * The externally visible workqueue. It relays the issued work items to 285 * the appropriate worker_pool through its pool_workqueues. 286 */ 287 struct workqueue_struct { 288 struct list_head pwqs; /* WR: all pwqs of this wq */ 289 struct list_head list; /* PR: list of all workqueues */ 290 291 struct mutex mutex; /* protects this wq */ 292 int work_color; /* WQ: current work color */ 293 int flush_color; /* WQ: current flush color */ 294 atomic_t nr_pwqs_to_flush; /* flush in progress */ 295 struct wq_flusher *first_flusher; /* WQ: first flusher */ 296 struct list_head flusher_queue; /* WQ: flush waiters */ 297 struct list_head flusher_overflow; /* WQ: flush overflow list */ 298 299 struct list_head maydays; /* MD: pwqs requesting rescue */ 300 struct worker *rescuer; /* MD: rescue worker */ 301 302 int nr_drainers; /* WQ: drain in progress */ 303 int max_active; /* WO: max active works */ 304 int saved_max_active; /* WQ: saved max_active */ 305 306 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 307 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 308 309 #ifdef CONFIG_SYSFS 310 struct wq_device *wq_dev; /* I: for sysfs interface */ 311 #endif 312 #ifdef CONFIG_LOCKDEP 313 char *lock_name; 314 struct lock_class_key key; 315 struct lockdep_map lockdep_map; 316 #endif 317 char name[WQ_NAME_LEN]; /* I: workqueue name */ 318 319 /* 320 * Destruction of workqueue_struct is RCU protected to allow walking 321 * the workqueues list without grabbing wq_pool_mutex. 322 * This is used to dump all workqueues from sysrq. 323 */ 324 struct rcu_head rcu; 325 326 /* hot fields used during command issue, aligned to cacheline */ 327 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 328 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */ 329 }; 330 331 static struct kmem_cache *pwq_cache; 332 333 /* 334 * Each pod type describes how CPUs should be grouped for unbound workqueues. 335 * See the comment above workqueue_attrs->affn_scope. 336 */ 337 struct wq_pod_type { 338 int nr_pods; /* number of pods */ 339 cpumask_var_t *pod_cpus; /* pod -> cpus */ 340 int *pod_node; /* pod -> node */ 341 int *cpu_pod; /* cpu -> pod */ 342 }; 343 344 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; 345 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE; 346 347 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { 348 [WQ_AFFN_DFL] = "default", 349 [WQ_AFFN_CPU] = "cpu", 350 [WQ_AFFN_SMT] = "smt", 351 [WQ_AFFN_CACHE] = "cache", 352 [WQ_AFFN_NUMA] = "numa", 353 [WQ_AFFN_SYSTEM] = "system", 354 }; 355 356 /* 357 * Per-cpu work items which run for longer than the following threshold are 358 * automatically considered CPU intensive and excluded from concurrency 359 * management to prevent them from noticeably delaying other per-cpu work items. 360 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 361 * The actual value is initialized in wq_cpu_intensive_thresh_init(). 362 */ 363 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 364 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 365 366 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 367 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 368 module_param_named(power_efficient, wq_power_efficient, bool, 0444); 369 370 static bool wq_online; /* can kworkers be created yet? */ 371 372 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ 373 static struct workqueue_attrs *wq_update_pod_attrs_buf; 374 375 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 376 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 377 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 378 /* wait for manager to go away */ 379 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 380 381 static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 382 static bool workqueue_freezing; /* PL: have wqs started freezing? */ 383 384 /* PL&A: allowable cpus for unbound wqs and work items */ 385 static cpumask_var_t wq_unbound_cpumask; 386 387 /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 388 static struct cpumask wq_cmdline_cpumask __initdata; 389 390 /* CPU where unbound work was last round robin scheduled from this CPU */ 391 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 392 393 /* 394 * Local execution of unbound work items is no longer guaranteed. The 395 * following always forces round-robin CPU selection on unbound work items 396 * to uncover usages which depend on it. 397 */ 398 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 399 static bool wq_debug_force_rr_cpu = true; 400 #else 401 static bool wq_debug_force_rr_cpu = false; 402 #endif 403 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 404 405 /* the per-cpu worker pools */ 406 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 407 408 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 409 410 /* PL: hash of all unbound pools keyed by pool->attrs */ 411 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 412 413 /* I: attributes used when instantiating standard unbound pools on demand */ 414 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 415 416 /* I: attributes used when instantiating ordered pools on demand */ 417 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 418 419 /* 420 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 421 * process context while holding a pool lock. Bounce to a dedicated kthread 422 * worker to avoid A-A deadlocks. 423 */ 424 static struct kthread_worker *pwq_release_worker; 425 426 struct workqueue_struct *system_wq __read_mostly; 427 EXPORT_SYMBOL(system_wq); 428 struct workqueue_struct *system_highpri_wq __read_mostly; 429 EXPORT_SYMBOL_GPL(system_highpri_wq); 430 struct workqueue_struct *system_long_wq __read_mostly; 431 EXPORT_SYMBOL_GPL(system_long_wq); 432 struct workqueue_struct *system_unbound_wq __read_mostly; 433 EXPORT_SYMBOL_GPL(system_unbound_wq); 434 struct workqueue_struct *system_freezable_wq __read_mostly; 435 EXPORT_SYMBOL_GPL(system_freezable_wq); 436 struct workqueue_struct *system_power_efficient_wq __read_mostly; 437 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 438 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 439 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 440 441 static int worker_thread(void *__worker); 442 static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 443 static void show_pwq(struct pool_workqueue *pwq); 444 static void show_one_worker_pool(struct worker_pool *pool); 445 446 #define CREATE_TRACE_POINTS 447 #include <trace/events/workqueue.h> 448 449 #define assert_rcu_or_pool_mutex() \ 450 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 451 !lockdep_is_held(&wq_pool_mutex), \ 452 "RCU or wq_pool_mutex should be held") 453 454 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 455 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 456 !lockdep_is_held(&wq->mutex) && \ 457 !lockdep_is_held(&wq_pool_mutex), \ 458 "RCU, wq->mutex or wq_pool_mutex should be held") 459 460 #define for_each_cpu_worker_pool(pool, cpu) \ 461 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 462 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 463 (pool)++) 464 465 /** 466 * for_each_pool - iterate through all worker_pools in the system 467 * @pool: iteration cursor 468 * @pi: integer used for iteration 469 * 470 * This must be called either with wq_pool_mutex held or RCU read 471 * locked. If the pool needs to be used beyond the locking in effect, the 472 * caller is responsible for guaranteeing that the pool stays online. 473 * 474 * The if/else clause exists only for the lockdep assertion and can be 475 * ignored. 476 */ 477 #define for_each_pool(pool, pi) \ 478 idr_for_each_entry(&worker_pool_idr, pool, pi) \ 479 if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 480 else 481 482 /** 483 * for_each_pool_worker - iterate through all workers of a worker_pool 484 * @worker: iteration cursor 485 * @pool: worker_pool to iterate workers of 486 * 487 * This must be called with wq_pool_attach_mutex. 488 * 489 * The if/else clause exists only for the lockdep assertion and can be 490 * ignored. 491 */ 492 #define for_each_pool_worker(worker, pool) \ 493 list_for_each_entry((worker), &(pool)->workers, node) \ 494 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 495 else 496 497 /** 498 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 499 * @pwq: iteration cursor 500 * @wq: the target workqueue 501 * 502 * This must be called either with wq->mutex held or RCU read locked. 503 * If the pwq needs to be used beyond the locking in effect, the caller is 504 * responsible for guaranteeing that the pwq stays online. 505 * 506 * The if/else clause exists only for the lockdep assertion and can be 507 * ignored. 508 */ 509 #define for_each_pwq(pwq, wq) \ 510 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 511 lockdep_is_held(&(wq->mutex))) 512 513 #ifdef CONFIG_DEBUG_OBJECTS_WORK 514 515 static const struct debug_obj_descr work_debug_descr; 516 517 static void *work_debug_hint(void *addr) 518 { 519 return ((struct work_struct *) addr)->func; 520 } 521 522 static bool work_is_static_object(void *addr) 523 { 524 struct work_struct *work = addr; 525 526 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 527 } 528 529 /* 530 * fixup_init is called when: 531 * - an active object is initialized 532 */ 533 static bool work_fixup_init(void *addr, enum debug_obj_state state) 534 { 535 struct work_struct *work = addr; 536 537 switch (state) { 538 case ODEBUG_STATE_ACTIVE: 539 cancel_work_sync(work); 540 debug_object_init(work, &work_debug_descr); 541 return true; 542 default: 543 return false; 544 } 545 } 546 547 /* 548 * fixup_free is called when: 549 * - an active object is freed 550 */ 551 static bool work_fixup_free(void *addr, enum debug_obj_state state) 552 { 553 struct work_struct *work = addr; 554 555 switch (state) { 556 case ODEBUG_STATE_ACTIVE: 557 cancel_work_sync(work); 558 debug_object_free(work, &work_debug_descr); 559 return true; 560 default: 561 return false; 562 } 563 } 564 565 static const struct debug_obj_descr work_debug_descr = { 566 .name = "work_struct", 567 .debug_hint = work_debug_hint, 568 .is_static_object = work_is_static_object, 569 .fixup_init = work_fixup_init, 570 .fixup_free = work_fixup_free, 571 }; 572 573 static inline void debug_work_activate(struct work_struct *work) 574 { 575 debug_object_activate(work, &work_debug_descr); 576 } 577 578 static inline void debug_work_deactivate(struct work_struct *work) 579 { 580 debug_object_deactivate(work, &work_debug_descr); 581 } 582 583 void __init_work(struct work_struct *work, int onstack) 584 { 585 if (onstack) 586 debug_object_init_on_stack(work, &work_debug_descr); 587 else 588 debug_object_init(work, &work_debug_descr); 589 } 590 EXPORT_SYMBOL_GPL(__init_work); 591 592 void destroy_work_on_stack(struct work_struct *work) 593 { 594 debug_object_free(work, &work_debug_descr); 595 } 596 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 597 598 void destroy_delayed_work_on_stack(struct delayed_work *work) 599 { 600 destroy_timer_on_stack(&work->timer); 601 debug_object_free(&work->work, &work_debug_descr); 602 } 603 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 604 605 #else 606 static inline void debug_work_activate(struct work_struct *work) { } 607 static inline void debug_work_deactivate(struct work_struct *work) { } 608 #endif 609 610 /** 611 * worker_pool_assign_id - allocate ID and assign it to @pool 612 * @pool: the pool pointer of interest 613 * 614 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 615 * successfully, -errno on failure. 616 */ 617 static int worker_pool_assign_id(struct worker_pool *pool) 618 { 619 int ret; 620 621 lockdep_assert_held(&wq_pool_mutex); 622 623 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 624 GFP_KERNEL); 625 if (ret >= 0) { 626 pool->id = ret; 627 return 0; 628 } 629 return ret; 630 } 631 632 static unsigned int work_color_to_flags(int color) 633 { 634 return color << WORK_STRUCT_COLOR_SHIFT; 635 } 636 637 static int get_work_color(unsigned long work_data) 638 { 639 return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 640 ((1 << WORK_STRUCT_COLOR_BITS) - 1); 641 } 642 643 static int work_next_color(int color) 644 { 645 return (color + 1) % WORK_NR_COLORS; 646 } 647 648 /* 649 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 650 * contain the pointer to the queued pwq. Once execution starts, the flag 651 * is cleared and the high bits contain OFFQ flags and pool ID. 652 * 653 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 654 * and clear_work_data() can be used to set the pwq, pool or clear 655 * work->data. These functions should only be called while the work is 656 * owned - ie. while the PENDING bit is set. 657 * 658 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 659 * corresponding to a work. Pool is available once the work has been 660 * queued anywhere after initialization until it is sync canceled. pwq is 661 * available only while the work item is queued. 662 * 663 * %WORK_OFFQ_CANCELING is used to mark a work item which is being 664 * canceled. While being canceled, a work item may have its PENDING set 665 * but stay off timer and worklist for arbitrarily long and nobody should 666 * try to steal the PENDING bit. 667 */ 668 static inline void set_work_data(struct work_struct *work, unsigned long data, 669 unsigned long flags) 670 { 671 WARN_ON_ONCE(!work_pending(work)); 672 atomic_long_set(&work->data, data | flags | work_static(work)); 673 } 674 675 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 676 unsigned long extra_flags) 677 { 678 set_work_data(work, (unsigned long)pwq, 679 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 680 } 681 682 static void set_work_pool_and_keep_pending(struct work_struct *work, 683 int pool_id) 684 { 685 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 686 WORK_STRUCT_PENDING); 687 } 688 689 static void set_work_pool_and_clear_pending(struct work_struct *work, 690 int pool_id) 691 { 692 /* 693 * The following wmb is paired with the implied mb in 694 * test_and_set_bit(PENDING) and ensures all updates to @work made 695 * here are visible to and precede any updates by the next PENDING 696 * owner. 697 */ 698 smp_wmb(); 699 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 700 /* 701 * The following mb guarantees that previous clear of a PENDING bit 702 * will not be reordered with any speculative LOADS or STORES from 703 * work->current_func, which is executed afterwards. This possible 704 * reordering can lead to a missed execution on attempt to queue 705 * the same @work. E.g. consider this case: 706 * 707 * CPU#0 CPU#1 708 * ---------------------------- -------------------------------- 709 * 710 * 1 STORE event_indicated 711 * 2 queue_work_on() { 712 * 3 test_and_set_bit(PENDING) 713 * 4 } set_..._and_clear_pending() { 714 * 5 set_work_data() # clear bit 715 * 6 smp_mb() 716 * 7 work->current_func() { 717 * 8 LOAD event_indicated 718 * } 719 * 720 * Without an explicit full barrier speculative LOAD on line 8 can 721 * be executed before CPU#0 does STORE on line 1. If that happens, 722 * CPU#0 observes the PENDING bit is still set and new execution of 723 * a @work is not queued in a hope, that CPU#1 will eventually 724 * finish the queued @work. Meanwhile CPU#1 does not see 725 * event_indicated is set, because speculative LOAD was executed 726 * before actual STORE. 727 */ 728 smp_mb(); 729 } 730 731 static void clear_work_data(struct work_struct *work) 732 { 733 smp_wmb(); /* see set_work_pool_and_clear_pending() */ 734 set_work_data(work, WORK_STRUCT_NO_POOL, 0); 735 } 736 737 static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 738 { 739 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); 740 } 741 742 static struct pool_workqueue *get_work_pwq(struct work_struct *work) 743 { 744 unsigned long data = atomic_long_read(&work->data); 745 746 if (data & WORK_STRUCT_PWQ) 747 return work_struct_pwq(data); 748 else 749 return NULL; 750 } 751 752 /** 753 * get_work_pool - return the worker_pool a given work was associated with 754 * @work: the work item of interest 755 * 756 * Pools are created and destroyed under wq_pool_mutex, and allows read 757 * access under RCU read lock. As such, this function should be 758 * called under wq_pool_mutex or inside of a rcu_read_lock() region. 759 * 760 * All fields of the returned pool are accessible as long as the above 761 * mentioned locking is in effect. If the returned pool needs to be used 762 * beyond the critical section, the caller is responsible for ensuring the 763 * returned pool is and stays online. 764 * 765 * Return: The worker_pool @work was last associated with. %NULL if none. 766 */ 767 static struct worker_pool *get_work_pool(struct work_struct *work) 768 { 769 unsigned long data = atomic_long_read(&work->data); 770 int pool_id; 771 772 assert_rcu_or_pool_mutex(); 773 774 if (data & WORK_STRUCT_PWQ) 775 return work_struct_pwq(data)->pool; 776 777 pool_id = data >> WORK_OFFQ_POOL_SHIFT; 778 if (pool_id == WORK_OFFQ_POOL_NONE) 779 return NULL; 780 781 return idr_find(&worker_pool_idr, pool_id); 782 } 783 784 /** 785 * get_work_pool_id - return the worker pool ID a given work is associated with 786 * @work: the work item of interest 787 * 788 * Return: The worker_pool ID @work was last associated with. 789 * %WORK_OFFQ_POOL_NONE if none. 790 */ 791 static int get_work_pool_id(struct work_struct *work) 792 { 793 unsigned long data = atomic_long_read(&work->data); 794 795 if (data & WORK_STRUCT_PWQ) 796 return work_struct_pwq(data)->pool->id; 797 798 return data >> WORK_OFFQ_POOL_SHIFT; 799 } 800 801 static void mark_work_canceling(struct work_struct *work) 802 { 803 unsigned long pool_id = get_work_pool_id(work); 804 805 pool_id <<= WORK_OFFQ_POOL_SHIFT; 806 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 807 } 808 809 static bool work_is_canceling(struct work_struct *work) 810 { 811 unsigned long data = atomic_long_read(&work->data); 812 813 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 814 } 815 816 /* 817 * Policy functions. These define the policies on how the global worker 818 * pools are managed. Unless noted otherwise, these functions assume that 819 * they're being called with pool->lock held. 820 */ 821 822 /* 823 * Need to wake up a worker? Called from anything but currently 824 * running workers. 825 * 826 * Note that, because unbound workers never contribute to nr_running, this 827 * function will always return %true for unbound pools as long as the 828 * worklist isn't empty. 829 */ 830 static bool need_more_worker(struct worker_pool *pool) 831 { 832 return !list_empty(&pool->worklist) && !pool->nr_running; 833 } 834 835 /* Can I start working? Called from busy but !running workers. */ 836 static bool may_start_working(struct worker_pool *pool) 837 { 838 return pool->nr_idle; 839 } 840 841 /* Do I need to keep working? Called from currently running workers. */ 842 static bool keep_working(struct worker_pool *pool) 843 { 844 return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 845 } 846 847 /* Do we need a new worker? Called from manager. */ 848 static bool need_to_create_worker(struct worker_pool *pool) 849 { 850 return need_more_worker(pool) && !may_start_working(pool); 851 } 852 853 /* Do we have too many workers and should some go away? */ 854 static bool too_many_workers(struct worker_pool *pool) 855 { 856 bool managing = pool->flags & POOL_MANAGER_ACTIVE; 857 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 858 int nr_busy = pool->nr_workers - nr_idle; 859 860 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 861 } 862 863 /** 864 * worker_set_flags - set worker flags and adjust nr_running accordingly 865 * @worker: self 866 * @flags: flags to set 867 * 868 * Set @flags in @worker->flags and adjust nr_running accordingly. 869 */ 870 static inline void worker_set_flags(struct worker *worker, unsigned int flags) 871 { 872 struct worker_pool *pool = worker->pool; 873 874 lockdep_assert_held(&pool->lock); 875 876 /* If transitioning into NOT_RUNNING, adjust nr_running. */ 877 if ((flags & WORKER_NOT_RUNNING) && 878 !(worker->flags & WORKER_NOT_RUNNING)) { 879 pool->nr_running--; 880 } 881 882 worker->flags |= flags; 883 } 884 885 /** 886 * worker_clr_flags - clear worker flags and adjust nr_running accordingly 887 * @worker: self 888 * @flags: flags to clear 889 * 890 * Clear @flags in @worker->flags and adjust nr_running accordingly. 891 */ 892 static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 893 { 894 struct worker_pool *pool = worker->pool; 895 unsigned int oflags = worker->flags; 896 897 lockdep_assert_held(&pool->lock); 898 899 worker->flags &= ~flags; 900 901 /* 902 * If transitioning out of NOT_RUNNING, increment nr_running. Note 903 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 904 * of multiple flags, not a single flag. 905 */ 906 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 907 if (!(worker->flags & WORKER_NOT_RUNNING)) 908 pool->nr_running++; 909 } 910 911 /* Return the first idle worker. Called with pool->lock held. */ 912 static struct worker *first_idle_worker(struct worker_pool *pool) 913 { 914 if (unlikely(list_empty(&pool->idle_list))) 915 return NULL; 916 917 return list_first_entry(&pool->idle_list, struct worker, entry); 918 } 919 920 /** 921 * worker_enter_idle - enter idle state 922 * @worker: worker which is entering idle state 923 * 924 * @worker is entering idle state. Update stats and idle timer if 925 * necessary. 926 * 927 * LOCKING: 928 * raw_spin_lock_irq(pool->lock). 929 */ 930 static void worker_enter_idle(struct worker *worker) 931 { 932 struct worker_pool *pool = worker->pool; 933 934 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 935 WARN_ON_ONCE(!list_empty(&worker->entry) && 936 (worker->hentry.next || worker->hentry.pprev))) 937 return; 938 939 /* can't use worker_set_flags(), also called from create_worker() */ 940 worker->flags |= WORKER_IDLE; 941 pool->nr_idle++; 942 worker->last_active = jiffies; 943 944 /* idle_list is LIFO */ 945 list_add(&worker->entry, &pool->idle_list); 946 947 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 948 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 949 950 /* Sanity check nr_running. */ 951 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 952 } 953 954 /** 955 * worker_leave_idle - leave idle state 956 * @worker: worker which is leaving idle state 957 * 958 * @worker is leaving idle state. Update stats. 959 * 960 * LOCKING: 961 * raw_spin_lock_irq(pool->lock). 962 */ 963 static void worker_leave_idle(struct worker *worker) 964 { 965 struct worker_pool *pool = worker->pool; 966 967 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 968 return; 969 worker_clr_flags(worker, WORKER_IDLE); 970 pool->nr_idle--; 971 list_del_init(&worker->entry); 972 } 973 974 /** 975 * find_worker_executing_work - find worker which is executing a work 976 * @pool: pool of interest 977 * @work: work to find worker for 978 * 979 * Find a worker which is executing @work on @pool by searching 980 * @pool->busy_hash which is keyed by the address of @work. For a worker 981 * to match, its current execution should match the address of @work and 982 * its work function. This is to avoid unwanted dependency between 983 * unrelated work executions through a work item being recycled while still 984 * being executed. 985 * 986 * This is a bit tricky. A work item may be freed once its execution 987 * starts and nothing prevents the freed area from being recycled for 988 * another work item. If the same work item address ends up being reused 989 * before the original execution finishes, workqueue will identify the 990 * recycled work item as currently executing and make it wait until the 991 * current execution finishes, introducing an unwanted dependency. 992 * 993 * This function checks the work item address and work function to avoid 994 * false positives. Note that this isn't complete as one may construct a 995 * work function which can introduce dependency onto itself through a 996 * recycled work item. Well, if somebody wants to shoot oneself in the 997 * foot that badly, there's only so much we can do, and if such deadlock 998 * actually occurs, it should be easy to locate the culprit work function. 999 * 1000 * CONTEXT: 1001 * raw_spin_lock_irq(pool->lock). 1002 * 1003 * Return: 1004 * Pointer to worker which is executing @work if found, %NULL 1005 * otherwise. 1006 */ 1007 static struct worker *find_worker_executing_work(struct worker_pool *pool, 1008 struct work_struct *work) 1009 { 1010 struct worker *worker; 1011 1012 hash_for_each_possible(pool->busy_hash, worker, hentry, 1013 (unsigned long)work) 1014 if (worker->current_work == work && 1015 worker->current_func == work->func) 1016 return worker; 1017 1018 return NULL; 1019 } 1020 1021 /** 1022 * move_linked_works - move linked works to a list 1023 * @work: start of series of works to be scheduled 1024 * @head: target list to append @work to 1025 * @nextp: out parameter for nested worklist walking 1026 * 1027 * Schedule linked works starting from @work to @head. Work series to be 1028 * scheduled starts at @work and includes any consecutive work with 1029 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on 1030 * @nextp. 1031 * 1032 * CONTEXT: 1033 * raw_spin_lock_irq(pool->lock). 1034 */ 1035 static void move_linked_works(struct work_struct *work, struct list_head *head, 1036 struct work_struct **nextp) 1037 { 1038 struct work_struct *n; 1039 1040 /* 1041 * Linked worklist will always end before the end of the list, 1042 * use NULL for list head. 1043 */ 1044 list_for_each_entry_safe_from(work, n, NULL, entry) { 1045 list_move_tail(&work->entry, head); 1046 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1047 break; 1048 } 1049 1050 /* 1051 * If we're already inside safe list traversal and have moved 1052 * multiple works to the scheduled queue, the next position 1053 * needs to be updated. 1054 */ 1055 if (nextp) 1056 *nextp = n; 1057 } 1058 1059 /** 1060 * assign_work - assign a work item and its linked work items to a worker 1061 * @work: work to assign 1062 * @worker: worker to assign to 1063 * @nextp: out parameter for nested worklist walking 1064 * 1065 * Assign @work and its linked work items to @worker. If @work is already being 1066 * executed by another worker in the same pool, it'll be punted there. 1067 * 1068 * If @nextp is not NULL, it's updated to point to the next work of the last 1069 * scheduled work. This allows assign_work() to be nested inside 1070 * list_for_each_entry_safe(). 1071 * 1072 * Returns %true if @work was successfully assigned to @worker. %false if @work 1073 * was punted to another worker already executing it. 1074 */ 1075 static bool assign_work(struct work_struct *work, struct worker *worker, 1076 struct work_struct **nextp) 1077 { 1078 struct worker_pool *pool = worker->pool; 1079 struct worker *collision; 1080 1081 lockdep_assert_held(&pool->lock); 1082 1083 /* 1084 * A single work shouldn't be executed concurrently by multiple workers. 1085 * __queue_work() ensures that @work doesn't jump to a different pool 1086 * while still running in the previous pool. Here, we should ensure that 1087 * @work is not executed concurrently by multiple workers from the same 1088 * pool. Check whether anyone is already processing the work. If so, 1089 * defer the work to the currently executing one. 1090 */ 1091 collision = find_worker_executing_work(pool, work); 1092 if (unlikely(collision)) { 1093 move_linked_works(work, &collision->scheduled, nextp); 1094 return false; 1095 } 1096 1097 move_linked_works(work, &worker->scheduled, nextp); 1098 return true; 1099 } 1100 1101 /** 1102 * kick_pool - wake up an idle worker if necessary 1103 * @pool: pool to kick 1104 * 1105 * @pool may have pending work items. Wake up worker if necessary. Returns 1106 * whether a worker was woken up. 1107 */ 1108 static bool kick_pool(struct worker_pool *pool) 1109 { 1110 struct worker *worker = first_idle_worker(pool); 1111 struct task_struct *p; 1112 1113 lockdep_assert_held(&pool->lock); 1114 1115 if (!need_more_worker(pool) || !worker) 1116 return false; 1117 1118 p = worker->task; 1119 1120 #ifdef CONFIG_SMP 1121 /* 1122 * Idle @worker is about to execute @work and waking up provides an 1123 * opportunity to migrate @worker at a lower cost by setting the task's 1124 * wake_cpu field. Let's see if we want to move @worker to improve 1125 * execution locality. 1126 * 1127 * We're waking the worker that went idle the latest and there's some 1128 * chance that @worker is marked idle but hasn't gone off CPU yet. If 1129 * so, setting the wake_cpu won't do anything. As this is a best-effort 1130 * optimization and the race window is narrow, let's leave as-is for 1131 * now. If this becomes pronounced, we can skip over workers which are 1132 * still on cpu when picking an idle worker. 1133 * 1134 * If @pool has non-strict affinity, @worker might have ended up outside 1135 * its affinity scope. Repatriate. 1136 */ 1137 if (!pool->attrs->affn_strict && 1138 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { 1139 struct work_struct *work = list_first_entry(&pool->worklist, 1140 struct work_struct, entry); 1141 p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask); 1142 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; 1143 } 1144 #endif 1145 wake_up_process(p); 1146 return true; 1147 } 1148 1149 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 1150 1151 /* 1152 * Concurrency-managed per-cpu work items that hog CPU for longer than 1153 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 1154 * which prevents them from stalling other concurrency-managed work items. If a 1155 * work function keeps triggering this mechanism, it's likely that the work item 1156 * should be using an unbound workqueue instead. 1157 * 1158 * wq_cpu_intensive_report() tracks work functions which trigger such conditions 1159 * and report them so that they can be examined and converted to use unbound 1160 * workqueues as appropriate. To avoid flooding the console, each violating work 1161 * function is tracked and reported with exponential backoff. 1162 */ 1163 #define WCI_MAX_ENTS 128 1164 1165 struct wci_ent { 1166 work_func_t func; 1167 atomic64_t cnt; 1168 struct hlist_node hash_node; 1169 }; 1170 1171 static struct wci_ent wci_ents[WCI_MAX_ENTS]; 1172 static int wci_nr_ents; 1173 static DEFINE_RAW_SPINLOCK(wci_lock); 1174 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 1175 1176 static struct wci_ent *wci_find_ent(work_func_t func) 1177 { 1178 struct wci_ent *ent; 1179 1180 hash_for_each_possible_rcu(wci_hash, ent, hash_node, 1181 (unsigned long)func) { 1182 if (ent->func == func) 1183 return ent; 1184 } 1185 return NULL; 1186 } 1187 1188 static void wq_cpu_intensive_report(work_func_t func) 1189 { 1190 struct wci_ent *ent; 1191 1192 restart: 1193 ent = wci_find_ent(func); 1194 if (ent) { 1195 u64 cnt; 1196 1197 /* 1198 * Start reporting from the fourth time and back off 1199 * exponentially. 1200 */ 1201 cnt = atomic64_inc_return_relaxed(&ent->cnt); 1202 if (cnt >= 4 && is_power_of_2(cnt)) 1203 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 1204 ent->func, wq_cpu_intensive_thresh_us, 1205 atomic64_read(&ent->cnt)); 1206 return; 1207 } 1208 1209 /* 1210 * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 1211 * is exhausted, something went really wrong and we probably made enough 1212 * noise already. 1213 */ 1214 if (wci_nr_ents >= WCI_MAX_ENTS) 1215 return; 1216 1217 raw_spin_lock(&wci_lock); 1218 1219 if (wci_nr_ents >= WCI_MAX_ENTS) { 1220 raw_spin_unlock(&wci_lock); 1221 return; 1222 } 1223 1224 if (wci_find_ent(func)) { 1225 raw_spin_unlock(&wci_lock); 1226 goto restart; 1227 } 1228 1229 ent = &wci_ents[wci_nr_ents++]; 1230 ent->func = func; 1231 atomic64_set(&ent->cnt, 1); 1232 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 1233 1234 raw_spin_unlock(&wci_lock); 1235 } 1236 1237 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1238 static void wq_cpu_intensive_report(work_func_t func) {} 1239 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1240 1241 /** 1242 * wq_worker_running - a worker is running again 1243 * @task: task waking up 1244 * 1245 * This function is called when a worker returns from schedule() 1246 */ 1247 void wq_worker_running(struct task_struct *task) 1248 { 1249 struct worker *worker = kthread_data(task); 1250 1251 if (!READ_ONCE(worker->sleeping)) 1252 return; 1253 1254 /* 1255 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 1256 * and the nr_running increment below, we may ruin the nr_running reset 1257 * and leave with an unexpected pool->nr_running == 1 on the newly unbound 1258 * pool. Protect against such race. 1259 */ 1260 preempt_disable(); 1261 if (!(worker->flags & WORKER_NOT_RUNNING)) 1262 worker->pool->nr_running++; 1263 preempt_enable(); 1264 1265 /* 1266 * CPU intensive auto-detection cares about how long a work item hogged 1267 * CPU without sleeping. Reset the starting timestamp on wakeup. 1268 */ 1269 worker->current_at = worker->task->se.sum_exec_runtime; 1270 1271 WRITE_ONCE(worker->sleeping, 0); 1272 } 1273 1274 /** 1275 * wq_worker_sleeping - a worker is going to sleep 1276 * @task: task going to sleep 1277 * 1278 * This function is called from schedule() when a busy worker is 1279 * going to sleep. 1280 */ 1281 void wq_worker_sleeping(struct task_struct *task) 1282 { 1283 struct worker *worker = kthread_data(task); 1284 struct worker_pool *pool; 1285 1286 /* 1287 * Rescuers, which may not have all the fields set up like normal 1288 * workers, also reach here, let's not access anything before 1289 * checking NOT_RUNNING. 1290 */ 1291 if (worker->flags & WORKER_NOT_RUNNING) 1292 return; 1293 1294 pool = worker->pool; 1295 1296 /* Return if preempted before wq_worker_running() was reached */ 1297 if (READ_ONCE(worker->sleeping)) 1298 return; 1299 1300 WRITE_ONCE(worker->sleeping, 1); 1301 raw_spin_lock_irq(&pool->lock); 1302 1303 /* 1304 * Recheck in case unbind_workers() preempted us. We don't 1305 * want to decrement nr_running after the worker is unbound 1306 * and nr_running has been reset. 1307 */ 1308 if (worker->flags & WORKER_NOT_RUNNING) { 1309 raw_spin_unlock_irq(&pool->lock); 1310 return; 1311 } 1312 1313 pool->nr_running--; 1314 if (kick_pool(pool)) 1315 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1316 1317 raw_spin_unlock_irq(&pool->lock); 1318 } 1319 1320 /** 1321 * wq_worker_tick - a scheduler tick occurred while a kworker is running 1322 * @task: task currently running 1323 * 1324 * Called from scheduler_tick(). We're in the IRQ context and the current 1325 * worker's fields which follow the 'K' locking rule can be accessed safely. 1326 */ 1327 void wq_worker_tick(struct task_struct *task) 1328 { 1329 struct worker *worker = kthread_data(task); 1330 struct pool_workqueue *pwq = worker->current_pwq; 1331 struct worker_pool *pool = worker->pool; 1332 1333 if (!pwq) 1334 return; 1335 1336 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 1337 1338 if (!wq_cpu_intensive_thresh_us) 1339 return; 1340 1341 /* 1342 * If the current worker is concurrency managed and hogged the CPU for 1343 * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1344 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1345 * 1346 * Set @worker->sleeping means that @worker is in the process of 1347 * switching out voluntarily and won't be contributing to 1348 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1349 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1350 * double decrements. The task is releasing the CPU anyway. Let's skip. 1351 * We probably want to make this prettier in the future. 1352 */ 1353 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1354 worker->task->se.sum_exec_runtime - worker->current_at < 1355 wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1356 return; 1357 1358 raw_spin_lock(&pool->lock); 1359 1360 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 1361 wq_cpu_intensive_report(worker->current_func); 1362 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1363 1364 if (kick_pool(pool)) 1365 pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1366 1367 raw_spin_unlock(&pool->lock); 1368 } 1369 1370 /** 1371 * wq_worker_last_func - retrieve worker's last work function 1372 * @task: Task to retrieve last work function of. 1373 * 1374 * Determine the last function a worker executed. This is called from 1375 * the scheduler to get a worker's last known identity. 1376 * 1377 * CONTEXT: 1378 * raw_spin_lock_irq(rq->lock) 1379 * 1380 * This function is called during schedule() when a kworker is going 1381 * to sleep. It's used by psi to identify aggregation workers during 1382 * dequeuing, to allow periodic aggregation to shut-off when that 1383 * worker is the last task in the system or cgroup to go to sleep. 1384 * 1385 * As this function doesn't involve any workqueue-related locking, it 1386 * only returns stable values when called from inside the scheduler's 1387 * queuing and dequeuing paths, when @task, which must be a kworker, 1388 * is guaranteed to not be processing any works. 1389 * 1390 * Return: 1391 * The last work function %current executed as a worker, NULL if it 1392 * hasn't executed any work yet. 1393 */ 1394 work_func_t wq_worker_last_func(struct task_struct *task) 1395 { 1396 struct worker *worker = kthread_data(task); 1397 1398 return worker->last_func; 1399 } 1400 1401 /** 1402 * get_pwq - get an extra reference on the specified pool_workqueue 1403 * @pwq: pool_workqueue to get 1404 * 1405 * Obtain an extra reference on @pwq. The caller should guarantee that 1406 * @pwq has positive refcnt and be holding the matching pool->lock. 1407 */ 1408 static void get_pwq(struct pool_workqueue *pwq) 1409 { 1410 lockdep_assert_held(&pwq->pool->lock); 1411 WARN_ON_ONCE(pwq->refcnt <= 0); 1412 pwq->refcnt++; 1413 } 1414 1415 /** 1416 * put_pwq - put a pool_workqueue reference 1417 * @pwq: pool_workqueue to put 1418 * 1419 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 1420 * destruction. The caller should be holding the matching pool->lock. 1421 */ 1422 static void put_pwq(struct pool_workqueue *pwq) 1423 { 1424 lockdep_assert_held(&pwq->pool->lock); 1425 if (likely(--pwq->refcnt)) 1426 return; 1427 /* 1428 * @pwq can't be released under pool->lock, bounce to a dedicated 1429 * kthread_worker to avoid A-A deadlocks. 1430 */ 1431 kthread_queue_work(pwq_release_worker, &pwq->release_work); 1432 } 1433 1434 /** 1435 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1436 * @pwq: pool_workqueue to put (can be %NULL) 1437 * 1438 * put_pwq() with locking. This function also allows %NULL @pwq. 1439 */ 1440 static void put_pwq_unlocked(struct pool_workqueue *pwq) 1441 { 1442 if (pwq) { 1443 /* 1444 * As both pwqs and pools are RCU protected, the 1445 * following lock operations are safe. 1446 */ 1447 raw_spin_lock_irq(&pwq->pool->lock); 1448 put_pwq(pwq); 1449 raw_spin_unlock_irq(&pwq->pool->lock); 1450 } 1451 } 1452 1453 static bool pwq_is_empty(struct pool_workqueue *pwq) 1454 { 1455 return !pwq->nr_active && list_empty(&pwq->inactive_works); 1456 } 1457 1458 static void pwq_activate_inactive_work(struct work_struct *work) 1459 { 1460 struct pool_workqueue *pwq = get_work_pwq(work); 1461 1462 trace_workqueue_activate_work(work); 1463 if (list_empty(&pwq->pool->worklist)) 1464 pwq->pool->watchdog_ts = jiffies; 1465 move_linked_works(work, &pwq->pool->worklist, NULL); 1466 __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1467 pwq->nr_active++; 1468 } 1469 1470 static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 1471 { 1472 struct work_struct *work = list_first_entry(&pwq->inactive_works, 1473 struct work_struct, entry); 1474 1475 pwq_activate_inactive_work(work); 1476 } 1477 1478 /** 1479 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1480 * @pwq: pwq of interest 1481 * @work_data: work_data of work which left the queue 1482 * 1483 * A work either has completed or is removed from pending queue, 1484 * decrement nr_in_flight of its pwq and handle workqueue flushing. 1485 * 1486 * CONTEXT: 1487 * raw_spin_lock_irq(pool->lock). 1488 */ 1489 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1490 { 1491 int color = get_work_color(work_data); 1492 1493 if (!(work_data & WORK_STRUCT_INACTIVE)) { 1494 pwq->nr_active--; 1495 if (!list_empty(&pwq->inactive_works)) { 1496 /* one down, submit an inactive one */ 1497 if (pwq->nr_active < READ_ONCE(pwq->wq->max_active)) 1498 pwq_activate_first_inactive(pwq); 1499 } 1500 } 1501 1502 pwq->nr_in_flight[color]--; 1503 1504 /* is flush in progress and are we at the flushing tip? */ 1505 if (likely(pwq->flush_color != color)) 1506 goto out_put; 1507 1508 /* are there still in-flight works? */ 1509 if (pwq->nr_in_flight[color]) 1510 goto out_put; 1511 1512 /* this pwq is done, clear flush_color */ 1513 pwq->flush_color = -1; 1514 1515 /* 1516 * If this was the last pwq, wake up the first flusher. It 1517 * will handle the rest. 1518 */ 1519 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1520 complete(&pwq->wq->first_flusher->done); 1521 out_put: 1522 put_pwq(pwq); 1523 } 1524 1525 /** 1526 * try_to_grab_pending - steal work item from worklist and disable irq 1527 * @work: work item to steal 1528 * @is_dwork: @work is a delayed_work 1529 * @flags: place to store irq state 1530 * 1531 * Try to grab PENDING bit of @work. This function can handle @work in any 1532 * stable state - idle, on timer or on worklist. 1533 * 1534 * Return: 1535 * 1536 * ======== ================================================================ 1537 * 1 if @work was pending and we successfully stole PENDING 1538 * 0 if @work was idle and we claimed PENDING 1539 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1540 * -ENOENT if someone else is canceling @work, this state may persist 1541 * for arbitrarily long 1542 * ======== ================================================================ 1543 * 1544 * Note: 1545 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1546 * interrupted while holding PENDING and @work off queue, irq must be 1547 * disabled on entry. This, combined with delayed_work->timer being 1548 * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1549 * 1550 * On successful return, >= 0, irq is disabled and the caller is 1551 * responsible for releasing it using local_irq_restore(*@flags). 1552 * 1553 * This function is safe to call from any context including IRQ handler. 1554 */ 1555 static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1556 unsigned long *flags) 1557 { 1558 struct worker_pool *pool; 1559 struct pool_workqueue *pwq; 1560 1561 local_irq_save(*flags); 1562 1563 /* try to steal the timer if it exists */ 1564 if (is_dwork) { 1565 struct delayed_work *dwork = to_delayed_work(work); 1566 1567 /* 1568 * dwork->timer is irqsafe. If del_timer() fails, it's 1569 * guaranteed that the timer is not queued anywhere and not 1570 * running on the local CPU. 1571 */ 1572 if (likely(del_timer(&dwork->timer))) 1573 return 1; 1574 } 1575 1576 /* try to claim PENDING the normal way */ 1577 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1578 return 0; 1579 1580 rcu_read_lock(); 1581 /* 1582 * The queueing is in progress, or it is already queued. Try to 1583 * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1584 */ 1585 pool = get_work_pool(work); 1586 if (!pool) 1587 goto fail; 1588 1589 raw_spin_lock(&pool->lock); 1590 /* 1591 * work->data is guaranteed to point to pwq only while the work 1592 * item is queued on pwq->wq, and both updating work->data to point 1593 * to pwq on queueing and to pool on dequeueing are done under 1594 * pwq->pool->lock. This in turn guarantees that, if work->data 1595 * points to pwq which is associated with a locked pool, the work 1596 * item is currently queued on that pool. 1597 */ 1598 pwq = get_work_pwq(work); 1599 if (pwq && pwq->pool == pool) { 1600 debug_work_deactivate(work); 1601 1602 /* 1603 * A cancelable inactive work item must be in the 1604 * pwq->inactive_works since a queued barrier can't be 1605 * canceled (see the comments in insert_wq_barrier()). 1606 * 1607 * An inactive work item cannot be grabbed directly because 1608 * it might have linked barrier work items which, if left 1609 * on the inactive_works list, will confuse pwq->nr_active 1610 * management later on and cause stall. Make sure the work 1611 * item is activated before grabbing. 1612 */ 1613 if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1614 pwq_activate_inactive_work(work); 1615 1616 list_del_init(&work->entry); 1617 pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 1618 1619 /* work->data points to pwq iff queued, point to pool */ 1620 set_work_pool_and_keep_pending(work, pool->id); 1621 1622 raw_spin_unlock(&pool->lock); 1623 rcu_read_unlock(); 1624 return 1; 1625 } 1626 raw_spin_unlock(&pool->lock); 1627 fail: 1628 rcu_read_unlock(); 1629 local_irq_restore(*flags); 1630 if (work_is_canceling(work)) 1631 return -ENOENT; 1632 cpu_relax(); 1633 return -EAGAIN; 1634 } 1635 1636 /** 1637 * insert_work - insert a work into a pool 1638 * @pwq: pwq @work belongs to 1639 * @work: work to insert 1640 * @head: insertion point 1641 * @extra_flags: extra WORK_STRUCT_* flags to set 1642 * 1643 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1644 * work_struct flags. 1645 * 1646 * CONTEXT: 1647 * raw_spin_lock_irq(pool->lock). 1648 */ 1649 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1650 struct list_head *head, unsigned int extra_flags) 1651 { 1652 debug_work_activate(work); 1653 1654 /* record the work call stack in order to print it in KASAN reports */ 1655 kasan_record_aux_stack_noalloc(work); 1656 1657 /* we own @work, set data and link */ 1658 set_work_pwq(work, pwq, extra_flags); 1659 list_add_tail(&work->entry, head); 1660 get_pwq(pwq); 1661 } 1662 1663 /* 1664 * Test whether @work is being queued from another work executing on the 1665 * same workqueue. 1666 */ 1667 static bool is_chained_work(struct workqueue_struct *wq) 1668 { 1669 struct worker *worker; 1670 1671 worker = current_wq_worker(); 1672 /* 1673 * Return %true iff I'm a worker executing a work item on @wq. If 1674 * I'm @worker, it's safe to dereference it without locking. 1675 */ 1676 return worker && worker->current_pwq->wq == wq; 1677 } 1678 1679 /* 1680 * When queueing an unbound work item to a wq, prefer local CPU if allowed 1681 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1682 * avoid perturbing sensitive tasks. 1683 */ 1684 static int wq_select_unbound_cpu(int cpu) 1685 { 1686 int new_cpu; 1687 1688 if (likely(!wq_debug_force_rr_cpu)) { 1689 if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1690 return cpu; 1691 } else { 1692 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1693 } 1694 1695 new_cpu = __this_cpu_read(wq_rr_cpu_last); 1696 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1697 if (unlikely(new_cpu >= nr_cpu_ids)) { 1698 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1699 if (unlikely(new_cpu >= nr_cpu_ids)) 1700 return cpu; 1701 } 1702 __this_cpu_write(wq_rr_cpu_last, new_cpu); 1703 1704 return new_cpu; 1705 } 1706 1707 static void __queue_work(int cpu, struct workqueue_struct *wq, 1708 struct work_struct *work) 1709 { 1710 struct pool_workqueue *pwq; 1711 struct worker_pool *last_pool, *pool; 1712 unsigned int work_flags; 1713 unsigned int req_cpu = cpu; 1714 1715 /* 1716 * While a work item is PENDING && off queue, a task trying to 1717 * steal the PENDING will busy-loop waiting for it to either get 1718 * queued or lose PENDING. Grabbing PENDING and queueing should 1719 * happen with IRQ disabled. 1720 */ 1721 lockdep_assert_irqs_disabled(); 1722 1723 1724 /* 1725 * For a draining wq, only works from the same workqueue are 1726 * allowed. The __WQ_DESTROYING helps to spot the issue that 1727 * queues a new work item to a wq after destroy_workqueue(wq). 1728 */ 1729 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 1730 WARN_ON_ONCE(!is_chained_work(wq)))) 1731 return; 1732 rcu_read_lock(); 1733 retry: 1734 /* pwq which will be used unless @work is executing elsewhere */ 1735 if (req_cpu == WORK_CPU_UNBOUND) { 1736 if (wq->flags & WQ_UNBOUND) 1737 cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1738 else 1739 cpu = raw_smp_processor_id(); 1740 } 1741 1742 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 1743 pool = pwq->pool; 1744 1745 /* 1746 * If @work was previously on a different pool, it might still be 1747 * running there, in which case the work needs to be queued on that 1748 * pool to guarantee non-reentrancy. 1749 */ 1750 last_pool = get_work_pool(work); 1751 if (last_pool && last_pool != pool) { 1752 struct worker *worker; 1753 1754 raw_spin_lock(&last_pool->lock); 1755 1756 worker = find_worker_executing_work(last_pool, work); 1757 1758 if (worker && worker->current_pwq->wq == wq) { 1759 pwq = worker->current_pwq; 1760 pool = pwq->pool; 1761 WARN_ON_ONCE(pool != last_pool); 1762 } else { 1763 /* meh... not running there, queue here */ 1764 raw_spin_unlock(&last_pool->lock); 1765 raw_spin_lock(&pool->lock); 1766 } 1767 } else { 1768 raw_spin_lock(&pool->lock); 1769 } 1770 1771 /* 1772 * pwq is determined and locked. For unbound pools, we could have raced 1773 * with pwq release and it could already be dead. If its refcnt is zero, 1774 * repeat pwq selection. Note that unbound pwqs never die without 1775 * another pwq replacing it in cpu_pwq or while work items are executing 1776 * on it, so the retrying is guaranteed to make forward-progress. 1777 */ 1778 if (unlikely(!pwq->refcnt)) { 1779 if (wq->flags & WQ_UNBOUND) { 1780 raw_spin_unlock(&pool->lock); 1781 cpu_relax(); 1782 goto retry; 1783 } 1784 /* oops */ 1785 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 1786 wq->name, cpu); 1787 } 1788 1789 /* pwq determined, queue */ 1790 trace_workqueue_queue_work(req_cpu, pwq, work); 1791 1792 if (WARN_ON(!list_empty(&work->entry))) 1793 goto out; 1794 1795 pwq->nr_in_flight[pwq->work_color]++; 1796 work_flags = work_color_to_flags(pwq->work_color); 1797 1798 /* 1799 * Limit the number of concurrently active work items to max_active. 1800 * @work must also queue behind existing inactive work items to maintain 1801 * ordering when max_active changes. See wq_adjust_max_active(). 1802 */ 1803 if (list_empty(&pwq->inactive_works) && 1804 pwq->nr_active < READ_ONCE(pwq->wq->max_active)) { 1805 if (list_empty(&pool->worklist)) 1806 pool->watchdog_ts = jiffies; 1807 1808 trace_workqueue_activate_work(work); 1809 pwq->nr_active++; 1810 insert_work(pwq, work, &pool->worklist, work_flags); 1811 kick_pool(pool); 1812 } else { 1813 work_flags |= WORK_STRUCT_INACTIVE; 1814 insert_work(pwq, work, &pwq->inactive_works, work_flags); 1815 } 1816 1817 out: 1818 raw_spin_unlock(&pool->lock); 1819 rcu_read_unlock(); 1820 } 1821 1822 /** 1823 * queue_work_on - queue work on specific cpu 1824 * @cpu: CPU number to execute work on 1825 * @wq: workqueue to use 1826 * @work: work to queue 1827 * 1828 * We queue the work to a specific CPU, the caller must ensure it 1829 * can't go away. Callers that fail to ensure that the specified 1830 * CPU cannot go away will execute on a randomly chosen CPU. 1831 * But note well that callers specifying a CPU that never has been 1832 * online will get a splat. 1833 * 1834 * Return: %false if @work was already on a queue, %true otherwise. 1835 */ 1836 bool queue_work_on(int cpu, struct workqueue_struct *wq, 1837 struct work_struct *work) 1838 { 1839 bool ret = false; 1840 unsigned long flags; 1841 1842 local_irq_save(flags); 1843 1844 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1845 __queue_work(cpu, wq, work); 1846 ret = true; 1847 } 1848 1849 local_irq_restore(flags); 1850 return ret; 1851 } 1852 EXPORT_SYMBOL(queue_work_on); 1853 1854 /** 1855 * select_numa_node_cpu - Select a CPU based on NUMA node 1856 * @node: NUMA node ID that we want to select a CPU from 1857 * 1858 * This function will attempt to find a "random" cpu available on a given 1859 * node. If there are no CPUs available on the given node it will return 1860 * WORK_CPU_UNBOUND indicating that we should just schedule to any 1861 * available CPU if we need to schedule this work. 1862 */ 1863 static int select_numa_node_cpu(int node) 1864 { 1865 int cpu; 1866 1867 /* Delay binding to CPU if node is not valid or online */ 1868 if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 1869 return WORK_CPU_UNBOUND; 1870 1871 /* Use local node/cpu if we are already there */ 1872 cpu = raw_smp_processor_id(); 1873 if (node == cpu_to_node(cpu)) 1874 return cpu; 1875 1876 /* Use "random" otherwise know as "first" online CPU of node */ 1877 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 1878 1879 /* If CPU is valid return that, otherwise just defer */ 1880 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 1881 } 1882 1883 /** 1884 * queue_work_node - queue work on a "random" cpu for a given NUMA node 1885 * @node: NUMA node that we are targeting the work for 1886 * @wq: workqueue to use 1887 * @work: work to queue 1888 * 1889 * We queue the work to a "random" CPU within a given NUMA node. The basic 1890 * idea here is to provide a way to somehow associate work with a given 1891 * NUMA node. 1892 * 1893 * This function will only make a best effort attempt at getting this onto 1894 * the right NUMA node. If no node is requested or the requested node is 1895 * offline then we just fall back to standard queue_work behavior. 1896 * 1897 * Currently the "random" CPU ends up being the first available CPU in the 1898 * intersection of cpu_online_mask and the cpumask of the node, unless we 1899 * are running on the node. In that case we just use the current CPU. 1900 * 1901 * Return: %false if @work was already on a queue, %true otherwise. 1902 */ 1903 bool queue_work_node(int node, struct workqueue_struct *wq, 1904 struct work_struct *work) 1905 { 1906 unsigned long flags; 1907 bool ret = false; 1908 1909 /* 1910 * This current implementation is specific to unbound workqueues. 1911 * Specifically we only return the first available CPU for a given 1912 * node instead of cycling through individual CPUs within the node. 1913 * 1914 * If this is used with a per-cpu workqueue then the logic in 1915 * workqueue_select_cpu_near would need to be updated to allow for 1916 * some round robin type logic. 1917 */ 1918 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 1919 1920 local_irq_save(flags); 1921 1922 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1923 int cpu = select_numa_node_cpu(node); 1924 1925 __queue_work(cpu, wq, work); 1926 ret = true; 1927 } 1928 1929 local_irq_restore(flags); 1930 return ret; 1931 } 1932 EXPORT_SYMBOL_GPL(queue_work_node); 1933 1934 void delayed_work_timer_fn(struct timer_list *t) 1935 { 1936 struct delayed_work *dwork = from_timer(dwork, t, timer); 1937 1938 /* should have been called from irqsafe timer with irq already off */ 1939 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 1940 } 1941 EXPORT_SYMBOL(delayed_work_timer_fn); 1942 1943 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 1944 struct delayed_work *dwork, unsigned long delay) 1945 { 1946 struct timer_list *timer = &dwork->timer; 1947 struct work_struct *work = &dwork->work; 1948 1949 WARN_ON_ONCE(!wq); 1950 WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1951 WARN_ON_ONCE(timer_pending(timer)); 1952 WARN_ON_ONCE(!list_empty(&work->entry)); 1953 1954 /* 1955 * If @delay is 0, queue @dwork->work immediately. This is for 1956 * both optimization and correctness. The earliest @timer can 1957 * expire is on the closest next tick and delayed_work users depend 1958 * on that there's no such delay when @delay is 0. 1959 */ 1960 if (!delay) { 1961 __queue_work(cpu, wq, &dwork->work); 1962 return; 1963 } 1964 1965 dwork->wq = wq; 1966 dwork->cpu = cpu; 1967 timer->expires = jiffies + delay; 1968 1969 if (unlikely(cpu != WORK_CPU_UNBOUND)) 1970 add_timer_on(timer, cpu); 1971 else 1972 add_timer(timer); 1973 } 1974 1975 /** 1976 * queue_delayed_work_on - queue work on specific CPU after delay 1977 * @cpu: CPU number to execute work on 1978 * @wq: workqueue to use 1979 * @dwork: work to queue 1980 * @delay: number of jiffies to wait before queueing 1981 * 1982 * Return: %false if @work was already on a queue, %true otherwise. If 1983 * @delay is zero and @dwork is idle, it will be scheduled for immediate 1984 * execution. 1985 */ 1986 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 1987 struct delayed_work *dwork, unsigned long delay) 1988 { 1989 struct work_struct *work = &dwork->work; 1990 bool ret = false; 1991 unsigned long flags; 1992 1993 /* read the comment in __queue_work() */ 1994 local_irq_save(flags); 1995 1996 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1997 __queue_delayed_work(cpu, wq, dwork, delay); 1998 ret = true; 1999 } 2000 2001 local_irq_restore(flags); 2002 return ret; 2003 } 2004 EXPORT_SYMBOL(queue_delayed_work_on); 2005 2006 /** 2007 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 2008 * @cpu: CPU number to execute work on 2009 * @wq: workqueue to use 2010 * @dwork: work to queue 2011 * @delay: number of jiffies to wait before queueing 2012 * 2013 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 2014 * modify @dwork's timer so that it expires after @delay. If @delay is 2015 * zero, @work is guaranteed to be scheduled immediately regardless of its 2016 * current state. 2017 * 2018 * Return: %false if @dwork was idle and queued, %true if @dwork was 2019 * pending and its timer was modified. 2020 * 2021 * This function is safe to call from any context including IRQ handler. 2022 * See try_to_grab_pending() for details. 2023 */ 2024 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 2025 struct delayed_work *dwork, unsigned long delay) 2026 { 2027 unsigned long flags; 2028 int ret; 2029 2030 do { 2031 ret = try_to_grab_pending(&dwork->work, true, &flags); 2032 } while (unlikely(ret == -EAGAIN)); 2033 2034 if (likely(ret >= 0)) { 2035 __queue_delayed_work(cpu, wq, dwork, delay); 2036 local_irq_restore(flags); 2037 } 2038 2039 /* -ENOENT from try_to_grab_pending() becomes %true */ 2040 return ret; 2041 } 2042 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 2043 2044 static void rcu_work_rcufn(struct rcu_head *rcu) 2045 { 2046 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 2047 2048 /* read the comment in __queue_work() */ 2049 local_irq_disable(); 2050 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 2051 local_irq_enable(); 2052 } 2053 2054 /** 2055 * queue_rcu_work - queue work after a RCU grace period 2056 * @wq: workqueue to use 2057 * @rwork: work to queue 2058 * 2059 * Return: %false if @rwork was already pending, %true otherwise. Note 2060 * that a full RCU grace period is guaranteed only after a %true return. 2061 * While @rwork is guaranteed to be executed after a %false return, the 2062 * execution may happen before a full RCU grace period has passed. 2063 */ 2064 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 2065 { 2066 struct work_struct *work = &rwork->work; 2067 2068 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 2069 rwork->wq = wq; 2070 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 2071 return true; 2072 } 2073 2074 return false; 2075 } 2076 EXPORT_SYMBOL(queue_rcu_work); 2077 2078 static struct worker *alloc_worker(int node) 2079 { 2080 struct worker *worker; 2081 2082 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2083 if (worker) { 2084 INIT_LIST_HEAD(&worker->entry); 2085 INIT_LIST_HEAD(&worker->scheduled); 2086 INIT_LIST_HEAD(&worker->node); 2087 /* on creation a worker is in !idle && prep state */ 2088 worker->flags = WORKER_PREP; 2089 } 2090 return worker; 2091 } 2092 2093 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 2094 { 2095 if (pool->cpu < 0 && pool->attrs->affn_strict) 2096 return pool->attrs->__pod_cpumask; 2097 else 2098 return pool->attrs->cpumask; 2099 } 2100 2101 /** 2102 * worker_attach_to_pool() - attach a worker to a pool 2103 * @worker: worker to be attached 2104 * @pool: the target pool 2105 * 2106 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 2107 * cpu-binding of @worker are kept coordinated with the pool across 2108 * cpu-[un]hotplugs. 2109 */ 2110 static void worker_attach_to_pool(struct worker *worker, 2111 struct worker_pool *pool) 2112 { 2113 mutex_lock(&wq_pool_attach_mutex); 2114 2115 /* 2116 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 2117 * stable across this function. See the comments above the flag 2118 * definition for details. 2119 */ 2120 if (pool->flags & POOL_DISASSOCIATED) 2121 worker->flags |= WORKER_UNBOUND; 2122 else 2123 kthread_set_per_cpu(worker->task, pool->cpu); 2124 2125 if (worker->rescue_wq) 2126 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2127 2128 list_add_tail(&worker->node, &pool->workers); 2129 worker->pool = pool; 2130 2131 mutex_unlock(&wq_pool_attach_mutex); 2132 } 2133 2134 /** 2135 * worker_detach_from_pool() - detach a worker from its pool 2136 * @worker: worker which is attached to its pool 2137 * 2138 * Undo the attaching which had been done in worker_attach_to_pool(). The 2139 * caller worker shouldn't access to the pool after detached except it has 2140 * other reference to the pool. 2141 */ 2142 static void worker_detach_from_pool(struct worker *worker) 2143 { 2144 struct worker_pool *pool = worker->pool; 2145 struct completion *detach_completion = NULL; 2146 2147 mutex_lock(&wq_pool_attach_mutex); 2148 2149 kthread_set_per_cpu(worker->task, -1); 2150 list_del(&worker->node); 2151 worker->pool = NULL; 2152 2153 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 2154 detach_completion = pool->detach_completion; 2155 mutex_unlock(&wq_pool_attach_mutex); 2156 2157 /* clear leftover flags without pool->lock after it is detached */ 2158 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2159 2160 if (detach_completion) 2161 complete(detach_completion); 2162 } 2163 2164 /** 2165 * create_worker - create a new workqueue worker 2166 * @pool: pool the new worker will belong to 2167 * 2168 * Create and start a new worker which is attached to @pool. 2169 * 2170 * CONTEXT: 2171 * Might sleep. Does GFP_KERNEL allocations. 2172 * 2173 * Return: 2174 * Pointer to the newly created worker. 2175 */ 2176 static struct worker *create_worker(struct worker_pool *pool) 2177 { 2178 struct worker *worker; 2179 int id; 2180 char id_buf[23]; 2181 2182 /* ID is needed to determine kthread name */ 2183 id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 2184 if (id < 0) { 2185 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 2186 ERR_PTR(id)); 2187 return NULL; 2188 } 2189 2190 worker = alloc_worker(pool->node); 2191 if (!worker) { 2192 pr_err_once("workqueue: Failed to allocate a worker\n"); 2193 goto fail; 2194 } 2195 2196 worker->id = id; 2197 2198 if (pool->cpu >= 0) 2199 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2200 pool->attrs->nice < 0 ? "H" : ""); 2201 else 2202 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2203 2204 worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2205 "kworker/%s", id_buf); 2206 if (IS_ERR(worker->task)) { 2207 if (PTR_ERR(worker->task) == -EINTR) { 2208 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 2209 id_buf); 2210 } else { 2211 pr_err_once("workqueue: Failed to create a worker thread: %pe", 2212 worker->task); 2213 } 2214 goto fail; 2215 } 2216 2217 set_user_nice(worker->task, pool->attrs->nice); 2218 kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 2219 2220 /* successful, attach the worker to the pool */ 2221 worker_attach_to_pool(worker, pool); 2222 2223 /* start the newly created worker */ 2224 raw_spin_lock_irq(&pool->lock); 2225 2226 worker->pool->nr_workers++; 2227 worker_enter_idle(worker); 2228 kick_pool(pool); 2229 2230 /* 2231 * @worker is waiting on a completion in kthread() and will trigger hung 2232 * check if not woken up soon. As kick_pool() might not have waken it 2233 * up, wake it up explicitly once more. 2234 */ 2235 wake_up_process(worker->task); 2236 2237 raw_spin_unlock_irq(&pool->lock); 2238 2239 return worker; 2240 2241 fail: 2242 ida_free(&pool->worker_ida, id); 2243 kfree(worker); 2244 return NULL; 2245 } 2246 2247 static void unbind_worker(struct worker *worker) 2248 { 2249 lockdep_assert_held(&wq_pool_attach_mutex); 2250 2251 kthread_set_per_cpu(worker->task, -1); 2252 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2253 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2254 else 2255 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2256 } 2257 2258 static void wake_dying_workers(struct list_head *cull_list) 2259 { 2260 struct worker *worker, *tmp; 2261 2262 list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2263 list_del_init(&worker->entry); 2264 unbind_worker(worker); 2265 /* 2266 * If the worker was somehow already running, then it had to be 2267 * in pool->idle_list when set_worker_dying() happened or we 2268 * wouldn't have gotten here. 2269 * 2270 * Thus, the worker must either have observed the WORKER_DIE 2271 * flag, or have set its state to TASK_IDLE. Either way, the 2272 * below will be observed by the worker and is safe to do 2273 * outside of pool->lock. 2274 */ 2275 wake_up_process(worker->task); 2276 } 2277 } 2278 2279 /** 2280 * set_worker_dying - Tag a worker for destruction 2281 * @worker: worker to be destroyed 2282 * @list: transfer worker away from its pool->idle_list and into list 2283 * 2284 * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2285 * should be idle. 2286 * 2287 * CONTEXT: 2288 * raw_spin_lock_irq(pool->lock). 2289 */ 2290 static void set_worker_dying(struct worker *worker, struct list_head *list) 2291 { 2292 struct worker_pool *pool = worker->pool; 2293 2294 lockdep_assert_held(&pool->lock); 2295 lockdep_assert_held(&wq_pool_attach_mutex); 2296 2297 /* sanity check frenzy */ 2298 if (WARN_ON(worker->current_work) || 2299 WARN_ON(!list_empty(&worker->scheduled)) || 2300 WARN_ON(!(worker->flags & WORKER_IDLE))) 2301 return; 2302 2303 pool->nr_workers--; 2304 pool->nr_idle--; 2305 2306 worker->flags |= WORKER_DIE; 2307 2308 list_move(&worker->entry, list); 2309 list_move(&worker->node, &pool->dying_workers); 2310 } 2311 2312 /** 2313 * idle_worker_timeout - check if some idle workers can now be deleted. 2314 * @t: The pool's idle_timer that just expired 2315 * 2316 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 2317 * worker_leave_idle(), as a worker flicking between idle and active while its 2318 * pool is at the too_many_workers() tipping point would cause too much timer 2319 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 2320 * it expire and re-evaluate things from there. 2321 */ 2322 static void idle_worker_timeout(struct timer_list *t) 2323 { 2324 struct worker_pool *pool = from_timer(pool, t, idle_timer); 2325 bool do_cull = false; 2326 2327 if (work_pending(&pool->idle_cull_work)) 2328 return; 2329 2330 raw_spin_lock_irq(&pool->lock); 2331 2332 if (too_many_workers(pool)) { 2333 struct worker *worker; 2334 unsigned long expires; 2335 2336 /* idle_list is kept in LIFO order, check the last one */ 2337 worker = list_entry(pool->idle_list.prev, struct worker, entry); 2338 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2339 do_cull = !time_before(jiffies, expires); 2340 2341 if (!do_cull) 2342 mod_timer(&pool->idle_timer, expires); 2343 } 2344 raw_spin_unlock_irq(&pool->lock); 2345 2346 if (do_cull) 2347 queue_work(system_unbound_wq, &pool->idle_cull_work); 2348 } 2349 2350 /** 2351 * idle_cull_fn - cull workers that have been idle for too long. 2352 * @work: the pool's work for handling these idle workers 2353 * 2354 * This goes through a pool's idle workers and gets rid of those that have been 2355 * idle for at least IDLE_WORKER_TIMEOUT seconds. 2356 * 2357 * We don't want to disturb isolated CPUs because of a pcpu kworker being 2358 * culled, so this also resets worker affinity. This requires a sleepable 2359 * context, hence the split between timer callback and work item. 2360 */ 2361 static void idle_cull_fn(struct work_struct *work) 2362 { 2363 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2364 LIST_HEAD(cull_list); 2365 2366 /* 2367 * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2368 * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2369 * path. This is required as a previously-preempted worker could run after 2370 * set_worker_dying() has happened but before wake_dying_workers() did. 2371 */ 2372 mutex_lock(&wq_pool_attach_mutex); 2373 raw_spin_lock_irq(&pool->lock); 2374 2375 while (too_many_workers(pool)) { 2376 struct worker *worker; 2377 unsigned long expires; 2378 2379 worker = list_entry(pool->idle_list.prev, struct worker, entry); 2380 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2381 2382 if (time_before(jiffies, expires)) { 2383 mod_timer(&pool->idle_timer, expires); 2384 break; 2385 } 2386 2387 set_worker_dying(worker, &cull_list); 2388 } 2389 2390 raw_spin_unlock_irq(&pool->lock); 2391 wake_dying_workers(&cull_list); 2392 mutex_unlock(&wq_pool_attach_mutex); 2393 } 2394 2395 static void send_mayday(struct work_struct *work) 2396 { 2397 struct pool_workqueue *pwq = get_work_pwq(work); 2398 struct workqueue_struct *wq = pwq->wq; 2399 2400 lockdep_assert_held(&wq_mayday_lock); 2401 2402 if (!wq->rescuer) 2403 return; 2404 2405 /* mayday mayday mayday */ 2406 if (list_empty(&pwq->mayday_node)) { 2407 /* 2408 * If @pwq is for an unbound wq, its base ref may be put at 2409 * any time due to an attribute change. Pin @pwq until the 2410 * rescuer is done with it. 2411 */ 2412 get_pwq(pwq); 2413 list_add_tail(&pwq->mayday_node, &wq->maydays); 2414 wake_up_process(wq->rescuer->task); 2415 pwq->stats[PWQ_STAT_MAYDAY]++; 2416 } 2417 } 2418 2419 static void pool_mayday_timeout(struct timer_list *t) 2420 { 2421 struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2422 struct work_struct *work; 2423 2424 raw_spin_lock_irq(&pool->lock); 2425 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2426 2427 if (need_to_create_worker(pool)) { 2428 /* 2429 * We've been trying to create a new worker but 2430 * haven't been successful. We might be hitting an 2431 * allocation deadlock. Send distress signals to 2432 * rescuers. 2433 */ 2434 list_for_each_entry(work, &pool->worklist, entry) 2435 send_mayday(work); 2436 } 2437 2438 raw_spin_unlock(&wq_mayday_lock); 2439 raw_spin_unlock_irq(&pool->lock); 2440 2441 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2442 } 2443 2444 /** 2445 * maybe_create_worker - create a new worker if necessary 2446 * @pool: pool to create a new worker for 2447 * 2448 * Create a new worker for @pool if necessary. @pool is guaranteed to 2449 * have at least one idle worker on return from this function. If 2450 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 2451 * sent to all rescuers with works scheduled on @pool to resolve 2452 * possible allocation deadlock. 2453 * 2454 * On return, need_to_create_worker() is guaranteed to be %false and 2455 * may_start_working() %true. 2456 * 2457 * LOCKING: 2458 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2459 * multiple times. Does GFP_KERNEL allocations. Called only from 2460 * manager. 2461 */ 2462 static void maybe_create_worker(struct worker_pool *pool) 2463 __releases(&pool->lock) 2464 __acquires(&pool->lock) 2465 { 2466 restart: 2467 raw_spin_unlock_irq(&pool->lock); 2468 2469 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 2470 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2471 2472 while (true) { 2473 if (create_worker(pool) || !need_to_create_worker(pool)) 2474 break; 2475 2476 schedule_timeout_interruptible(CREATE_COOLDOWN); 2477 2478 if (!need_to_create_worker(pool)) 2479 break; 2480 } 2481 2482 del_timer_sync(&pool->mayday_timer); 2483 raw_spin_lock_irq(&pool->lock); 2484 /* 2485 * This is necessary even after a new worker was just successfully 2486 * created as @pool->lock was dropped and the new worker might have 2487 * already become busy. 2488 */ 2489 if (need_to_create_worker(pool)) 2490 goto restart; 2491 } 2492 2493 /** 2494 * manage_workers - manage worker pool 2495 * @worker: self 2496 * 2497 * Assume the manager role and manage the worker pool @worker belongs 2498 * to. At any given time, there can be only zero or one manager per 2499 * pool. The exclusion is handled automatically by this function. 2500 * 2501 * The caller can safely start processing works on false return. On 2502 * true return, it's guaranteed that need_to_create_worker() is false 2503 * and may_start_working() is true. 2504 * 2505 * CONTEXT: 2506 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2507 * multiple times. Does GFP_KERNEL allocations. 2508 * 2509 * Return: 2510 * %false if the pool doesn't need management and the caller can safely 2511 * start processing works, %true if management function was performed and 2512 * the conditions that the caller verified before calling the function may 2513 * no longer be true. 2514 */ 2515 static bool manage_workers(struct worker *worker) 2516 { 2517 struct worker_pool *pool = worker->pool; 2518 2519 if (pool->flags & POOL_MANAGER_ACTIVE) 2520 return false; 2521 2522 pool->flags |= POOL_MANAGER_ACTIVE; 2523 pool->manager = worker; 2524 2525 maybe_create_worker(pool); 2526 2527 pool->manager = NULL; 2528 pool->flags &= ~POOL_MANAGER_ACTIVE; 2529 rcuwait_wake_up(&manager_wait); 2530 return true; 2531 } 2532 2533 /** 2534 * process_one_work - process single work 2535 * @worker: self 2536 * @work: work to process 2537 * 2538 * Process @work. This function contains all the logics necessary to 2539 * process a single work including synchronization against and 2540 * interaction with other workers on the same cpu, queueing and 2541 * flushing. As long as context requirement is met, any worker can 2542 * call this function to process a work. 2543 * 2544 * CONTEXT: 2545 * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2546 */ 2547 static void process_one_work(struct worker *worker, struct work_struct *work) 2548 __releases(&pool->lock) 2549 __acquires(&pool->lock) 2550 { 2551 struct pool_workqueue *pwq = get_work_pwq(work); 2552 struct worker_pool *pool = worker->pool; 2553 unsigned long work_data; 2554 #ifdef CONFIG_LOCKDEP 2555 /* 2556 * It is permissible to free the struct work_struct from 2557 * inside the function that is called from it, this we need to 2558 * take into account for lockdep too. To avoid bogus "held 2559 * lock freed" warnings as well as problems when looking into 2560 * work->lockdep_map, make a copy and use that here. 2561 */ 2562 struct lockdep_map lockdep_map; 2563 2564 lockdep_copy_map(&lockdep_map, &work->lockdep_map); 2565 #endif 2566 /* ensure we're on the correct CPU */ 2567 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2568 raw_smp_processor_id() != pool->cpu); 2569 2570 /* claim and dequeue */ 2571 debug_work_deactivate(work); 2572 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2573 worker->current_work = work; 2574 worker->current_func = work->func; 2575 worker->current_pwq = pwq; 2576 worker->current_at = worker->task->se.sum_exec_runtime; 2577 work_data = *work_data_bits(work); 2578 worker->current_color = get_work_color(work_data); 2579 2580 /* 2581 * Record wq name for cmdline and debug reporting, may get 2582 * overridden through set_worker_desc(). 2583 */ 2584 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 2585 2586 list_del_init(&work->entry); 2587 2588 /* 2589 * CPU intensive works don't participate in concurrency management. 2590 * They're the scheduler's responsibility. This takes @worker out 2591 * of concurrency management and the next code block will chain 2592 * execution of the pending work items. 2593 */ 2594 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2595 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2596 2597 /* 2598 * Kick @pool if necessary. It's always noop for per-cpu worker pools 2599 * since nr_running would always be >= 1 at this point. This is used to 2600 * chain execution of the pending work items for WORKER_NOT_RUNNING 2601 * workers such as the UNBOUND and CPU_INTENSIVE ones. 2602 */ 2603 kick_pool(pool); 2604 2605 /* 2606 * Record the last pool and clear PENDING which should be the last 2607 * update to @work. Also, do this inside @pool->lock so that 2608 * PENDING and queued state changes happen together while IRQ is 2609 * disabled. 2610 */ 2611 set_work_pool_and_clear_pending(work, pool->id); 2612 2613 pwq->stats[PWQ_STAT_STARTED]++; 2614 raw_spin_unlock_irq(&pool->lock); 2615 2616 lock_map_acquire(&pwq->wq->lockdep_map); 2617 lock_map_acquire(&lockdep_map); 2618 /* 2619 * Strictly speaking we should mark the invariant state without holding 2620 * any locks, that is, before these two lock_map_acquire()'s. 2621 * 2622 * However, that would result in: 2623 * 2624 * A(W1) 2625 * WFC(C) 2626 * A(W1) 2627 * C(C) 2628 * 2629 * Which would create W1->C->W1 dependencies, even though there is no 2630 * actual deadlock possible. There are two solutions, using a 2631 * read-recursive acquire on the work(queue) 'locks', but this will then 2632 * hit the lockdep limitation on recursive locks, or simply discard 2633 * these locks. 2634 * 2635 * AFAICT there is no possible deadlock scenario between the 2636 * flush_work() and complete() primitives (except for single-threaded 2637 * workqueues), so hiding them isn't a problem. 2638 */ 2639 lockdep_invariant_state(true); 2640 trace_workqueue_execute_start(work); 2641 worker->current_func(work); 2642 /* 2643 * While we must be careful to not use "work" after this, the trace 2644 * point will only record its address. 2645 */ 2646 trace_workqueue_execute_end(work, worker->current_func); 2647 pwq->stats[PWQ_STAT_COMPLETED]++; 2648 lock_map_release(&lockdep_map); 2649 lock_map_release(&pwq->wq->lockdep_map); 2650 2651 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2652 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2653 " last function: %ps\n", 2654 current->comm, preempt_count(), task_pid_nr(current), 2655 worker->current_func); 2656 debug_show_held_locks(current); 2657 dump_stack(); 2658 } 2659 2660 /* 2661 * The following prevents a kworker from hogging CPU on !PREEMPTION 2662 * kernels, where a requeueing work item waiting for something to 2663 * happen could deadlock with stop_machine as such work item could 2664 * indefinitely requeue itself while all other CPUs are trapped in 2665 * stop_machine. At the same time, report a quiescent RCU state so 2666 * the same condition doesn't freeze RCU. 2667 */ 2668 cond_resched(); 2669 2670 raw_spin_lock_irq(&pool->lock); 2671 2672 /* 2673 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2674 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2675 * wq_cpu_intensive_thresh_us. Clear it. 2676 */ 2677 worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2678 2679 /* tag the worker for identification in schedule() */ 2680 worker->last_func = worker->current_func; 2681 2682 /* we're done with it, release */ 2683 hash_del(&worker->hentry); 2684 worker->current_work = NULL; 2685 worker->current_func = NULL; 2686 worker->current_pwq = NULL; 2687 worker->current_color = INT_MAX; 2688 pwq_dec_nr_in_flight(pwq, work_data); 2689 } 2690 2691 /** 2692 * process_scheduled_works - process scheduled works 2693 * @worker: self 2694 * 2695 * Process all scheduled works. Please note that the scheduled list 2696 * may change while processing a work, so this function repeatedly 2697 * fetches a work from the top and executes it. 2698 * 2699 * CONTEXT: 2700 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2701 * multiple times. 2702 */ 2703 static void process_scheduled_works(struct worker *worker) 2704 { 2705 struct work_struct *work; 2706 bool first = true; 2707 2708 while ((work = list_first_entry_or_null(&worker->scheduled, 2709 struct work_struct, entry))) { 2710 if (first) { 2711 worker->pool->watchdog_ts = jiffies; 2712 first = false; 2713 } 2714 process_one_work(worker, work); 2715 } 2716 } 2717 2718 static void set_pf_worker(bool val) 2719 { 2720 mutex_lock(&wq_pool_attach_mutex); 2721 if (val) 2722 current->flags |= PF_WQ_WORKER; 2723 else 2724 current->flags &= ~PF_WQ_WORKER; 2725 mutex_unlock(&wq_pool_attach_mutex); 2726 } 2727 2728 /** 2729 * worker_thread - the worker thread function 2730 * @__worker: self 2731 * 2732 * The worker thread function. All workers belong to a worker_pool - 2733 * either a per-cpu one or dynamic unbound one. These workers process all 2734 * work items regardless of their specific target workqueue. The only 2735 * exception is work items which belong to workqueues with a rescuer which 2736 * will be explained in rescuer_thread(). 2737 * 2738 * Return: 0 2739 */ 2740 static int worker_thread(void *__worker) 2741 { 2742 struct worker *worker = __worker; 2743 struct worker_pool *pool = worker->pool; 2744 2745 /* tell the scheduler that this is a workqueue worker */ 2746 set_pf_worker(true); 2747 woke_up: 2748 raw_spin_lock_irq(&pool->lock); 2749 2750 /* am I supposed to die? */ 2751 if (unlikely(worker->flags & WORKER_DIE)) { 2752 raw_spin_unlock_irq(&pool->lock); 2753 set_pf_worker(false); 2754 2755 set_task_comm(worker->task, "kworker/dying"); 2756 ida_free(&pool->worker_ida, worker->id); 2757 worker_detach_from_pool(worker); 2758 WARN_ON_ONCE(!list_empty(&worker->entry)); 2759 kfree(worker); 2760 return 0; 2761 } 2762 2763 worker_leave_idle(worker); 2764 recheck: 2765 /* no more worker necessary? */ 2766 if (!need_more_worker(pool)) 2767 goto sleep; 2768 2769 /* do we need to manage? */ 2770 if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2771 goto recheck; 2772 2773 /* 2774 * ->scheduled list can only be filled while a worker is 2775 * preparing to process a work or actually processing it. 2776 * Make sure nobody diddled with it while I was sleeping. 2777 */ 2778 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2779 2780 /* 2781 * Finish PREP stage. We're guaranteed to have at least one idle 2782 * worker or that someone else has already assumed the manager 2783 * role. This is where @worker starts participating in concurrency 2784 * management if applicable and concurrency management is restored 2785 * after being rebound. See rebind_workers() for details. 2786 */ 2787 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2788 2789 do { 2790 struct work_struct *work = 2791 list_first_entry(&pool->worklist, 2792 struct work_struct, entry); 2793 2794 if (assign_work(work, worker, NULL)) 2795 process_scheduled_works(worker); 2796 } while (keep_working(pool)); 2797 2798 worker_set_flags(worker, WORKER_PREP); 2799 sleep: 2800 /* 2801 * pool->lock is held and there's no work to process and no need to 2802 * manage, sleep. Workers are woken up only while holding 2803 * pool->lock or from local cpu, so setting the current state 2804 * before releasing pool->lock is enough to prevent losing any 2805 * event. 2806 */ 2807 worker_enter_idle(worker); 2808 __set_current_state(TASK_IDLE); 2809 raw_spin_unlock_irq(&pool->lock); 2810 schedule(); 2811 goto woke_up; 2812 } 2813 2814 /** 2815 * rescuer_thread - the rescuer thread function 2816 * @__rescuer: self 2817 * 2818 * Workqueue rescuer thread function. There's one rescuer for each 2819 * workqueue which has WQ_MEM_RECLAIM set. 2820 * 2821 * Regular work processing on a pool may block trying to create a new 2822 * worker which uses GFP_KERNEL allocation which has slight chance of 2823 * developing into deadlock if some works currently on the same queue 2824 * need to be processed to satisfy the GFP_KERNEL allocation. This is 2825 * the problem rescuer solves. 2826 * 2827 * When such condition is possible, the pool summons rescuers of all 2828 * workqueues which have works queued on the pool and let them process 2829 * those works so that forward progress can be guaranteed. 2830 * 2831 * This should happen rarely. 2832 * 2833 * Return: 0 2834 */ 2835 static int rescuer_thread(void *__rescuer) 2836 { 2837 struct worker *rescuer = __rescuer; 2838 struct workqueue_struct *wq = rescuer->rescue_wq; 2839 bool should_stop; 2840 2841 set_user_nice(current, RESCUER_NICE_LEVEL); 2842 2843 /* 2844 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2845 * doesn't participate in concurrency management. 2846 */ 2847 set_pf_worker(true); 2848 repeat: 2849 set_current_state(TASK_IDLE); 2850 2851 /* 2852 * By the time the rescuer is requested to stop, the workqueue 2853 * shouldn't have any work pending, but @wq->maydays may still have 2854 * pwq(s) queued. This can happen by non-rescuer workers consuming 2855 * all the work items before the rescuer got to them. Go through 2856 * @wq->maydays processing before acting on should_stop so that the 2857 * list is always empty on exit. 2858 */ 2859 should_stop = kthread_should_stop(); 2860 2861 /* see whether any pwq is asking for help */ 2862 raw_spin_lock_irq(&wq_mayday_lock); 2863 2864 while (!list_empty(&wq->maydays)) { 2865 struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2866 struct pool_workqueue, mayday_node); 2867 struct worker_pool *pool = pwq->pool; 2868 struct work_struct *work, *n; 2869 2870 __set_current_state(TASK_RUNNING); 2871 list_del_init(&pwq->mayday_node); 2872 2873 raw_spin_unlock_irq(&wq_mayday_lock); 2874 2875 worker_attach_to_pool(rescuer, pool); 2876 2877 raw_spin_lock_irq(&pool->lock); 2878 2879 /* 2880 * Slurp in all works issued via this workqueue and 2881 * process'em. 2882 */ 2883 WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2884 list_for_each_entry_safe(work, n, &pool->worklist, entry) { 2885 if (get_work_pwq(work) == pwq && 2886 assign_work(work, rescuer, &n)) 2887 pwq->stats[PWQ_STAT_RESCUED]++; 2888 } 2889 2890 if (!list_empty(&rescuer->scheduled)) { 2891 process_scheduled_works(rescuer); 2892 2893 /* 2894 * The above execution of rescued work items could 2895 * have created more to rescue through 2896 * pwq_activate_first_inactive() or chained 2897 * queueing. Let's put @pwq back on mayday list so 2898 * that such back-to-back work items, which may be 2899 * being used to relieve memory pressure, don't 2900 * incur MAYDAY_INTERVAL delay inbetween. 2901 */ 2902 if (pwq->nr_active && need_to_create_worker(pool)) { 2903 raw_spin_lock(&wq_mayday_lock); 2904 /* 2905 * Queue iff we aren't racing destruction 2906 * and somebody else hasn't queued it already. 2907 */ 2908 if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2909 get_pwq(pwq); 2910 list_add_tail(&pwq->mayday_node, &wq->maydays); 2911 } 2912 raw_spin_unlock(&wq_mayday_lock); 2913 } 2914 } 2915 2916 /* 2917 * Put the reference grabbed by send_mayday(). @pool won't 2918 * go away while we're still attached to it. 2919 */ 2920 put_pwq(pwq); 2921 2922 /* 2923 * Leave this pool. Notify regular workers; otherwise, we end up 2924 * with 0 concurrency and stalling the execution. 2925 */ 2926 kick_pool(pool); 2927 2928 raw_spin_unlock_irq(&pool->lock); 2929 2930 worker_detach_from_pool(rescuer); 2931 2932 raw_spin_lock_irq(&wq_mayday_lock); 2933 } 2934 2935 raw_spin_unlock_irq(&wq_mayday_lock); 2936 2937 if (should_stop) { 2938 __set_current_state(TASK_RUNNING); 2939 set_pf_worker(false); 2940 return 0; 2941 } 2942 2943 /* rescuers should never participate in concurrency management */ 2944 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2945 schedule(); 2946 goto repeat; 2947 } 2948 2949 /** 2950 * check_flush_dependency - check for flush dependency sanity 2951 * @target_wq: workqueue being flushed 2952 * @target_work: work item being flushed (NULL for workqueue flushes) 2953 * 2954 * %current is trying to flush the whole @target_wq or @target_work on it. 2955 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2956 * reclaiming memory or running on a workqueue which doesn't have 2957 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2958 * a deadlock. 2959 */ 2960 static void check_flush_dependency(struct workqueue_struct *target_wq, 2961 struct work_struct *target_work) 2962 { 2963 work_func_t target_func = target_work ? target_work->func : NULL; 2964 struct worker *worker; 2965 2966 if (target_wq->flags & WQ_MEM_RECLAIM) 2967 return; 2968 2969 worker = current_wq_worker(); 2970 2971 WARN_ONCE(current->flags & PF_MEMALLOC, 2972 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2973 current->pid, current->comm, target_wq->name, target_func); 2974 WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 2975 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2976 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2977 worker->current_pwq->wq->name, worker->current_func, 2978 target_wq->name, target_func); 2979 } 2980 2981 struct wq_barrier { 2982 struct work_struct work; 2983 struct completion done; 2984 struct task_struct *task; /* purely informational */ 2985 }; 2986 2987 static void wq_barrier_func(struct work_struct *work) 2988 { 2989 struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2990 complete(&barr->done); 2991 } 2992 2993 /** 2994 * insert_wq_barrier - insert a barrier work 2995 * @pwq: pwq to insert barrier into 2996 * @barr: wq_barrier to insert 2997 * @target: target work to attach @barr to 2998 * @worker: worker currently executing @target, NULL if @target is not executing 2999 * 3000 * @barr is linked to @target such that @barr is completed only after 3001 * @target finishes execution. Please note that the ordering 3002 * guarantee is observed only with respect to @target and on the local 3003 * cpu. 3004 * 3005 * Currently, a queued barrier can't be canceled. This is because 3006 * try_to_grab_pending() can't determine whether the work to be 3007 * grabbed is at the head of the queue and thus can't clear LINKED 3008 * flag of the previous work while there must be a valid next work 3009 * after a work with LINKED flag set. 3010 * 3011 * Note that when @worker is non-NULL, @target may be modified 3012 * underneath us, so we can't reliably determine pwq from @target. 3013 * 3014 * CONTEXT: 3015 * raw_spin_lock_irq(pool->lock). 3016 */ 3017 static void insert_wq_barrier(struct pool_workqueue *pwq, 3018 struct wq_barrier *barr, 3019 struct work_struct *target, struct worker *worker) 3020 { 3021 unsigned int work_flags = 0; 3022 unsigned int work_color; 3023 struct list_head *head; 3024 3025 /* 3026 * debugobject calls are safe here even with pool->lock locked 3027 * as we know for sure that this will not trigger any of the 3028 * checks and call back into the fixup functions where we 3029 * might deadlock. 3030 */ 3031 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 3032 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 3033 3034 init_completion_map(&barr->done, &target->lockdep_map); 3035 3036 barr->task = current; 3037 3038 /* The barrier work item does not participate in pwq->nr_active. */ 3039 work_flags |= WORK_STRUCT_INACTIVE; 3040 3041 /* 3042 * If @target is currently being executed, schedule the 3043 * barrier to the worker; otherwise, put it after @target. 3044 */ 3045 if (worker) { 3046 head = worker->scheduled.next; 3047 work_color = worker->current_color; 3048 } else { 3049 unsigned long *bits = work_data_bits(target); 3050 3051 head = target->entry.next; 3052 /* there can already be other linked works, inherit and set */ 3053 work_flags |= *bits & WORK_STRUCT_LINKED; 3054 work_color = get_work_color(*bits); 3055 __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3056 } 3057 3058 pwq->nr_in_flight[work_color]++; 3059 work_flags |= work_color_to_flags(work_color); 3060 3061 insert_work(pwq, &barr->work, head, work_flags); 3062 } 3063 3064 /** 3065 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 3066 * @wq: workqueue being flushed 3067 * @flush_color: new flush color, < 0 for no-op 3068 * @work_color: new work color, < 0 for no-op 3069 * 3070 * Prepare pwqs for workqueue flushing. 3071 * 3072 * If @flush_color is non-negative, flush_color on all pwqs should be 3073 * -1. If no pwq has in-flight commands at the specified color, all 3074 * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3075 * has in flight commands, its pwq->flush_color is set to 3076 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 3077 * wakeup logic is armed and %true is returned. 3078 * 3079 * The caller should have initialized @wq->first_flusher prior to 3080 * calling this function with non-negative @flush_color. If 3081 * @flush_color is negative, no flush color update is done and %false 3082 * is returned. 3083 * 3084 * If @work_color is non-negative, all pwqs should have the same 3085 * work_color which is previous to @work_color and all will be 3086 * advanced to @work_color. 3087 * 3088 * CONTEXT: 3089 * mutex_lock(wq->mutex). 3090 * 3091 * Return: 3092 * %true if @flush_color >= 0 and there's something to flush. %false 3093 * otherwise. 3094 */ 3095 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 3096 int flush_color, int work_color) 3097 { 3098 bool wait = false; 3099 struct pool_workqueue *pwq; 3100 3101 if (flush_color >= 0) { 3102 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3103 atomic_set(&wq->nr_pwqs_to_flush, 1); 3104 } 3105 3106 for_each_pwq(pwq, wq) { 3107 struct worker_pool *pool = pwq->pool; 3108 3109 raw_spin_lock_irq(&pool->lock); 3110 3111 if (flush_color >= 0) { 3112 WARN_ON_ONCE(pwq->flush_color != -1); 3113 3114 if (pwq->nr_in_flight[flush_color]) { 3115 pwq->flush_color = flush_color; 3116 atomic_inc(&wq->nr_pwqs_to_flush); 3117 wait = true; 3118 } 3119 } 3120 3121 if (work_color >= 0) { 3122 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3123 pwq->work_color = work_color; 3124 } 3125 3126 raw_spin_unlock_irq(&pool->lock); 3127 } 3128 3129 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 3130 complete(&wq->first_flusher->done); 3131 3132 return wait; 3133 } 3134 3135 /** 3136 * __flush_workqueue - ensure that any scheduled work has run to completion. 3137 * @wq: workqueue to flush 3138 * 3139 * This function sleeps until all work items which were queued on entry 3140 * have finished execution, but it is not livelocked by new incoming ones. 3141 */ 3142 void __flush_workqueue(struct workqueue_struct *wq) 3143 { 3144 struct wq_flusher this_flusher = { 3145 .list = LIST_HEAD_INIT(this_flusher.list), 3146 .flush_color = -1, 3147 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 3148 }; 3149 int next_color; 3150 3151 if (WARN_ON(!wq_online)) 3152 return; 3153 3154 lock_map_acquire(&wq->lockdep_map); 3155 lock_map_release(&wq->lockdep_map); 3156 3157 mutex_lock(&wq->mutex); 3158 3159 /* 3160 * Start-to-wait phase 3161 */ 3162 next_color = work_next_color(wq->work_color); 3163 3164 if (next_color != wq->flush_color) { 3165 /* 3166 * Color space is not full. The current work_color 3167 * becomes our flush_color and work_color is advanced 3168 * by one. 3169 */ 3170 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 3171 this_flusher.flush_color = wq->work_color; 3172 wq->work_color = next_color; 3173 3174 if (!wq->first_flusher) { 3175 /* no flush in progress, become the first flusher */ 3176 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 3177 3178 wq->first_flusher = &this_flusher; 3179 3180 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 3181 wq->work_color)) { 3182 /* nothing to flush, done */ 3183 wq->flush_color = next_color; 3184 wq->first_flusher = NULL; 3185 goto out_unlock; 3186 } 3187 } else { 3188 /* wait in queue */ 3189 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 3190 list_add_tail(&this_flusher.list, &wq->flusher_queue); 3191 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 3192 } 3193 } else { 3194 /* 3195 * Oops, color space is full, wait on overflow queue. 3196 * The next flush completion will assign us 3197 * flush_color and transfer to flusher_queue. 3198 */ 3199 list_add_tail(&this_flusher.list, &wq->flusher_overflow); 3200 } 3201 3202 check_flush_dependency(wq, NULL); 3203 3204 mutex_unlock(&wq->mutex); 3205 3206 wait_for_completion(&this_flusher.done); 3207 3208 /* 3209 * Wake-up-and-cascade phase 3210 * 3211 * First flushers are responsible for cascading flushes and 3212 * handling overflow. Non-first flushers can simply return. 3213 */ 3214 if (READ_ONCE(wq->first_flusher) != &this_flusher) 3215 return; 3216 3217 mutex_lock(&wq->mutex); 3218 3219 /* we might have raced, check again with mutex held */ 3220 if (wq->first_flusher != &this_flusher) 3221 goto out_unlock; 3222 3223 WRITE_ONCE(wq->first_flusher, NULL); 3224 3225 WARN_ON_ONCE(!list_empty(&this_flusher.list)); 3226 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 3227 3228 while (true) { 3229 struct wq_flusher *next, *tmp; 3230 3231 /* complete all the flushers sharing the current flush color */ 3232 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 3233 if (next->flush_color != wq->flush_color) 3234 break; 3235 list_del_init(&next->list); 3236 complete(&next->done); 3237 } 3238 3239 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 3240 wq->flush_color != work_next_color(wq->work_color)); 3241 3242 /* this flush_color is finished, advance by one */ 3243 wq->flush_color = work_next_color(wq->flush_color); 3244 3245 /* one color has been freed, handle overflow queue */ 3246 if (!list_empty(&wq->flusher_overflow)) { 3247 /* 3248 * Assign the same color to all overflowed 3249 * flushers, advance work_color and append to 3250 * flusher_queue. This is the start-to-wait 3251 * phase for these overflowed flushers. 3252 */ 3253 list_for_each_entry(tmp, &wq->flusher_overflow, list) 3254 tmp->flush_color = wq->work_color; 3255 3256 wq->work_color = work_next_color(wq->work_color); 3257 3258 list_splice_tail_init(&wq->flusher_overflow, 3259 &wq->flusher_queue); 3260 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 3261 } 3262 3263 if (list_empty(&wq->flusher_queue)) { 3264 WARN_ON_ONCE(wq->flush_color != wq->work_color); 3265 break; 3266 } 3267 3268 /* 3269 * Need to flush more colors. Make the next flusher 3270 * the new first flusher and arm pwqs. 3271 */ 3272 WARN_ON_ONCE(wq->flush_color == wq->work_color); 3273 WARN_ON_ONCE(wq->flush_color != next->flush_color); 3274 3275 list_del_init(&next->list); 3276 wq->first_flusher = next; 3277 3278 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 3279 break; 3280 3281 /* 3282 * Meh... this color is already done, clear first 3283 * flusher and repeat cascading. 3284 */ 3285 wq->first_flusher = NULL; 3286 } 3287 3288 out_unlock: 3289 mutex_unlock(&wq->mutex); 3290 } 3291 EXPORT_SYMBOL(__flush_workqueue); 3292 3293 /** 3294 * drain_workqueue - drain a workqueue 3295 * @wq: workqueue to drain 3296 * 3297 * Wait until the workqueue becomes empty. While draining is in progress, 3298 * only chain queueing is allowed. IOW, only currently pending or running 3299 * work items on @wq can queue further work items on it. @wq is flushed 3300 * repeatedly until it becomes empty. The number of flushing is determined 3301 * by the depth of chaining and should be relatively short. Whine if it 3302 * takes too long. 3303 */ 3304 void drain_workqueue(struct workqueue_struct *wq) 3305 { 3306 unsigned int flush_cnt = 0; 3307 struct pool_workqueue *pwq; 3308 3309 /* 3310 * __queue_work() needs to test whether there are drainers, is much 3311 * hotter than drain_workqueue() and already looks at @wq->flags. 3312 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 3313 */ 3314 mutex_lock(&wq->mutex); 3315 if (!wq->nr_drainers++) 3316 wq->flags |= __WQ_DRAINING; 3317 mutex_unlock(&wq->mutex); 3318 reflush: 3319 __flush_workqueue(wq); 3320 3321 mutex_lock(&wq->mutex); 3322 3323 for_each_pwq(pwq, wq) { 3324 bool drained; 3325 3326 raw_spin_lock_irq(&pwq->pool->lock); 3327 drained = pwq_is_empty(pwq); 3328 raw_spin_unlock_irq(&pwq->pool->lock); 3329 3330 if (drained) 3331 continue; 3332 3333 if (++flush_cnt == 10 || 3334 (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3335 pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3336 wq->name, __func__, flush_cnt); 3337 3338 mutex_unlock(&wq->mutex); 3339 goto reflush; 3340 } 3341 3342 if (!--wq->nr_drainers) 3343 wq->flags &= ~__WQ_DRAINING; 3344 mutex_unlock(&wq->mutex); 3345 } 3346 EXPORT_SYMBOL_GPL(drain_workqueue); 3347 3348 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3349 bool from_cancel) 3350 { 3351 struct worker *worker = NULL; 3352 struct worker_pool *pool; 3353 struct pool_workqueue *pwq; 3354 3355 might_sleep(); 3356 3357 rcu_read_lock(); 3358 pool = get_work_pool(work); 3359 if (!pool) { 3360 rcu_read_unlock(); 3361 return false; 3362 } 3363 3364 raw_spin_lock_irq(&pool->lock); 3365 /* see the comment in try_to_grab_pending() with the same code */ 3366 pwq = get_work_pwq(work); 3367 if (pwq) { 3368 if (unlikely(pwq->pool != pool)) 3369 goto already_gone; 3370 } else { 3371 worker = find_worker_executing_work(pool, work); 3372 if (!worker) 3373 goto already_gone; 3374 pwq = worker->current_pwq; 3375 } 3376 3377 check_flush_dependency(pwq->wq, work); 3378 3379 insert_wq_barrier(pwq, barr, work, worker); 3380 raw_spin_unlock_irq(&pool->lock); 3381 3382 /* 3383 * Force a lock recursion deadlock when using flush_work() inside a 3384 * single-threaded or rescuer equipped workqueue. 3385 * 3386 * For single threaded workqueues the deadlock happens when the work 3387 * is after the work issuing the flush_work(). For rescuer equipped 3388 * workqueues the deadlock happens when the rescuer stalls, blocking 3389 * forward progress. 3390 */ 3391 if (!from_cancel && 3392 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3393 lock_map_acquire(&pwq->wq->lockdep_map); 3394 lock_map_release(&pwq->wq->lockdep_map); 3395 } 3396 rcu_read_unlock(); 3397 return true; 3398 already_gone: 3399 raw_spin_unlock_irq(&pool->lock); 3400 rcu_read_unlock(); 3401 return false; 3402 } 3403 3404 static bool __flush_work(struct work_struct *work, bool from_cancel) 3405 { 3406 struct wq_barrier barr; 3407 3408 if (WARN_ON(!wq_online)) 3409 return false; 3410 3411 if (WARN_ON(!work->func)) 3412 return false; 3413 3414 lock_map_acquire(&work->lockdep_map); 3415 lock_map_release(&work->lockdep_map); 3416 3417 if (start_flush_work(work, &barr, from_cancel)) { 3418 wait_for_completion(&barr.done); 3419 destroy_work_on_stack(&barr.work); 3420 return true; 3421 } else { 3422 return false; 3423 } 3424 } 3425 3426 /** 3427 * flush_work - wait for a work to finish executing the last queueing instance 3428 * @work: the work to flush 3429 * 3430 * Wait until @work has finished execution. @work is guaranteed to be idle 3431 * on return if it hasn't been requeued since flush started. 3432 * 3433 * Return: 3434 * %true if flush_work() waited for the work to finish execution, 3435 * %false if it was already idle. 3436 */ 3437 bool flush_work(struct work_struct *work) 3438 { 3439 return __flush_work(work, false); 3440 } 3441 EXPORT_SYMBOL_GPL(flush_work); 3442 3443 struct cwt_wait { 3444 wait_queue_entry_t wait; 3445 struct work_struct *work; 3446 }; 3447 3448 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 3449 { 3450 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 3451 3452 if (cwait->work != key) 3453 return 0; 3454 return autoremove_wake_function(wait, mode, sync, key); 3455 } 3456 3457 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3458 { 3459 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3460 unsigned long flags; 3461 int ret; 3462 3463 do { 3464 ret = try_to_grab_pending(work, is_dwork, &flags); 3465 /* 3466 * If someone else is already canceling, wait for it to 3467 * finish. flush_work() doesn't work for PREEMPT_NONE 3468 * because we may get scheduled between @work's completion 3469 * and the other canceling task resuming and clearing 3470 * CANCELING - flush_work() will return false immediately 3471 * as @work is no longer busy, try_to_grab_pending() will 3472 * return -ENOENT as @work is still being canceled and the 3473 * other canceling task won't be able to clear CANCELING as 3474 * we're hogging the CPU. 3475 * 3476 * Let's wait for completion using a waitqueue. As this 3477 * may lead to the thundering herd problem, use a custom 3478 * wake function which matches @work along with exclusive 3479 * wait and wakeup. 3480 */ 3481 if (unlikely(ret == -ENOENT)) { 3482 struct cwt_wait cwait; 3483 3484 init_wait(&cwait.wait); 3485 cwait.wait.func = cwt_wakefn; 3486 cwait.work = work; 3487 3488 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 3489 TASK_UNINTERRUPTIBLE); 3490 if (work_is_canceling(work)) 3491 schedule(); 3492 finish_wait(&cancel_waitq, &cwait.wait); 3493 } 3494 } while (unlikely(ret < 0)); 3495 3496 /* tell other tasks trying to grab @work to back off */ 3497 mark_work_canceling(work); 3498 local_irq_restore(flags); 3499 3500 /* 3501 * This allows canceling during early boot. We know that @work 3502 * isn't executing. 3503 */ 3504 if (wq_online) 3505 __flush_work(work, true); 3506 3507 clear_work_data(work); 3508 3509 /* 3510 * Paired with prepare_to_wait() above so that either 3511 * waitqueue_active() is visible here or !work_is_canceling() is 3512 * visible there. 3513 */ 3514 smp_mb(); 3515 if (waitqueue_active(&cancel_waitq)) 3516 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 3517 3518 return ret; 3519 } 3520 3521 /** 3522 * cancel_work_sync - cancel a work and wait for it to finish 3523 * @work: the work to cancel 3524 * 3525 * Cancel @work and wait for its execution to finish. This function 3526 * can be used even if the work re-queues itself or migrates to 3527 * another workqueue. On return from this function, @work is 3528 * guaranteed to be not pending or executing on any CPU. 3529 * 3530 * cancel_work_sync(&delayed_work->work) must not be used for 3531 * delayed_work's. Use cancel_delayed_work_sync() instead. 3532 * 3533 * The caller must ensure that the workqueue on which @work was last 3534 * queued can't be destroyed before this function returns. 3535 * 3536 * Return: 3537 * %true if @work was pending, %false otherwise. 3538 */ 3539 bool cancel_work_sync(struct work_struct *work) 3540 { 3541 return __cancel_work_timer(work, false); 3542 } 3543 EXPORT_SYMBOL_GPL(cancel_work_sync); 3544 3545 /** 3546 * flush_delayed_work - wait for a dwork to finish executing the last queueing 3547 * @dwork: the delayed work to flush 3548 * 3549 * Delayed timer is cancelled and the pending work is queued for 3550 * immediate execution. Like flush_work(), this function only 3551 * considers the last queueing instance of @dwork. 3552 * 3553 * Return: 3554 * %true if flush_work() waited for the work to finish execution, 3555 * %false if it was already idle. 3556 */ 3557 bool flush_delayed_work(struct delayed_work *dwork) 3558 { 3559 local_irq_disable(); 3560 if (del_timer_sync(&dwork->timer)) 3561 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 3562 local_irq_enable(); 3563 return flush_work(&dwork->work); 3564 } 3565 EXPORT_SYMBOL(flush_delayed_work); 3566 3567 /** 3568 * flush_rcu_work - wait for a rwork to finish executing the last queueing 3569 * @rwork: the rcu work to flush 3570 * 3571 * Return: 3572 * %true if flush_rcu_work() waited for the work to finish execution, 3573 * %false if it was already idle. 3574 */ 3575 bool flush_rcu_work(struct rcu_work *rwork) 3576 { 3577 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 3578 rcu_barrier(); 3579 flush_work(&rwork->work); 3580 return true; 3581 } else { 3582 return flush_work(&rwork->work); 3583 } 3584 } 3585 EXPORT_SYMBOL(flush_rcu_work); 3586 3587 static bool __cancel_work(struct work_struct *work, bool is_dwork) 3588 { 3589 unsigned long flags; 3590 int ret; 3591 3592 do { 3593 ret = try_to_grab_pending(work, is_dwork, &flags); 3594 } while (unlikely(ret == -EAGAIN)); 3595 3596 if (unlikely(ret < 0)) 3597 return false; 3598 3599 set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3600 local_irq_restore(flags); 3601 return ret; 3602 } 3603 3604 /* 3605 * See cancel_delayed_work() 3606 */ 3607 bool cancel_work(struct work_struct *work) 3608 { 3609 return __cancel_work(work, false); 3610 } 3611 EXPORT_SYMBOL(cancel_work); 3612 3613 /** 3614 * cancel_delayed_work - cancel a delayed work 3615 * @dwork: delayed_work to cancel 3616 * 3617 * Kill off a pending delayed_work. 3618 * 3619 * Return: %true if @dwork was pending and canceled; %false if it wasn't 3620 * pending. 3621 * 3622 * Note: 3623 * The work callback function may still be running on return, unless 3624 * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3625 * use cancel_delayed_work_sync() to wait on it. 3626 * 3627 * This function is safe to call from any context including IRQ handler. 3628 */ 3629 bool cancel_delayed_work(struct delayed_work *dwork) 3630 { 3631 return __cancel_work(&dwork->work, true); 3632 } 3633 EXPORT_SYMBOL(cancel_delayed_work); 3634 3635 /** 3636 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3637 * @dwork: the delayed work cancel 3638 * 3639 * This is cancel_work_sync() for delayed works. 3640 * 3641 * Return: 3642 * %true if @dwork was pending, %false otherwise. 3643 */ 3644 bool cancel_delayed_work_sync(struct delayed_work *dwork) 3645 { 3646 return __cancel_work_timer(&dwork->work, true); 3647 } 3648 EXPORT_SYMBOL(cancel_delayed_work_sync); 3649 3650 /** 3651 * schedule_on_each_cpu - execute a function synchronously on each online CPU 3652 * @func: the function to call 3653 * 3654 * schedule_on_each_cpu() executes @func on each online CPU using the 3655 * system workqueue and blocks until all CPUs have completed. 3656 * schedule_on_each_cpu() is very slow. 3657 * 3658 * Return: 3659 * 0 on success, -errno on failure. 3660 */ 3661 int schedule_on_each_cpu(work_func_t func) 3662 { 3663 int cpu; 3664 struct work_struct __percpu *works; 3665 3666 works = alloc_percpu(struct work_struct); 3667 if (!works) 3668 return -ENOMEM; 3669 3670 cpus_read_lock(); 3671 3672 for_each_online_cpu(cpu) { 3673 struct work_struct *work = per_cpu_ptr(works, cpu); 3674 3675 INIT_WORK(work, func); 3676 schedule_work_on(cpu, work); 3677 } 3678 3679 for_each_online_cpu(cpu) 3680 flush_work(per_cpu_ptr(works, cpu)); 3681 3682 cpus_read_unlock(); 3683 free_percpu(works); 3684 return 0; 3685 } 3686 3687 /** 3688 * execute_in_process_context - reliably execute the routine with user context 3689 * @fn: the function to execute 3690 * @ew: guaranteed storage for the execute work structure (must 3691 * be available when the work executes) 3692 * 3693 * Executes the function immediately if process context is available, 3694 * otherwise schedules the function for delayed execution. 3695 * 3696 * Return: 0 - function was executed 3697 * 1 - function was scheduled for execution 3698 */ 3699 int execute_in_process_context(work_func_t fn, struct execute_work *ew) 3700 { 3701 if (!in_interrupt()) { 3702 fn(&ew->work); 3703 return 0; 3704 } 3705 3706 INIT_WORK(&ew->work, fn); 3707 schedule_work(&ew->work); 3708 3709 return 1; 3710 } 3711 EXPORT_SYMBOL_GPL(execute_in_process_context); 3712 3713 /** 3714 * free_workqueue_attrs - free a workqueue_attrs 3715 * @attrs: workqueue_attrs to free 3716 * 3717 * Undo alloc_workqueue_attrs(). 3718 */ 3719 void free_workqueue_attrs(struct workqueue_attrs *attrs) 3720 { 3721 if (attrs) { 3722 free_cpumask_var(attrs->cpumask); 3723 free_cpumask_var(attrs->__pod_cpumask); 3724 kfree(attrs); 3725 } 3726 } 3727 3728 /** 3729 * alloc_workqueue_attrs - allocate a workqueue_attrs 3730 * 3731 * Allocate a new workqueue_attrs, initialize with default settings and 3732 * return it. 3733 * 3734 * Return: The allocated new workqueue_attr on success. %NULL on failure. 3735 */ 3736 struct workqueue_attrs *alloc_workqueue_attrs(void) 3737 { 3738 struct workqueue_attrs *attrs; 3739 3740 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 3741 if (!attrs) 3742 goto fail; 3743 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 3744 goto fail; 3745 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 3746 goto fail; 3747 3748 cpumask_copy(attrs->cpumask, cpu_possible_mask); 3749 attrs->affn_scope = WQ_AFFN_DFL; 3750 return attrs; 3751 fail: 3752 free_workqueue_attrs(attrs); 3753 return NULL; 3754 } 3755 3756 static void copy_workqueue_attrs(struct workqueue_attrs *to, 3757 const struct workqueue_attrs *from) 3758 { 3759 to->nice = from->nice; 3760 cpumask_copy(to->cpumask, from->cpumask); 3761 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 3762 to->affn_strict = from->affn_strict; 3763 3764 /* 3765 * Unlike hash and equality test, copying shouldn't ignore wq-only 3766 * fields as copying is used for both pool and wq attrs. Instead, 3767 * get_unbound_pool() explicitly clears the fields. 3768 */ 3769 to->affn_scope = from->affn_scope; 3770 to->ordered = from->ordered; 3771 } 3772 3773 /* 3774 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 3775 * comments in 'struct workqueue_attrs' definition. 3776 */ 3777 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 3778 { 3779 attrs->affn_scope = WQ_AFFN_NR_TYPES; 3780 attrs->ordered = false; 3781 } 3782 3783 /* hash value of the content of @attr */ 3784 static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 3785 { 3786 u32 hash = 0; 3787 3788 hash = jhash_1word(attrs->nice, hash); 3789 hash = jhash(cpumask_bits(attrs->cpumask), 3790 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 3791 hash = jhash(cpumask_bits(attrs->__pod_cpumask), 3792 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 3793 hash = jhash_1word(attrs->affn_strict, hash); 3794 return hash; 3795 } 3796 3797 /* content equality test */ 3798 static bool wqattrs_equal(const struct workqueue_attrs *a, 3799 const struct workqueue_attrs *b) 3800 { 3801 if (a->nice != b->nice) 3802 return false; 3803 if (!cpumask_equal(a->cpumask, b->cpumask)) 3804 return false; 3805 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 3806 return false; 3807 if (a->affn_strict != b->affn_strict) 3808 return false; 3809 return true; 3810 } 3811 3812 /* Update @attrs with actually available CPUs */ 3813 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 3814 const cpumask_t *unbound_cpumask) 3815 { 3816 /* 3817 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 3818 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 3819 * @unbound_cpumask. 3820 */ 3821 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 3822 if (unlikely(cpumask_empty(attrs->cpumask))) 3823 cpumask_copy(attrs->cpumask, unbound_cpumask); 3824 } 3825 3826 /* find wq_pod_type to use for @attrs */ 3827 static const struct wq_pod_type * 3828 wqattrs_pod_type(const struct workqueue_attrs *attrs) 3829 { 3830 enum wq_affn_scope scope; 3831 struct wq_pod_type *pt; 3832 3833 /* to synchronize access to wq_affn_dfl */ 3834 lockdep_assert_held(&wq_pool_mutex); 3835 3836 if (attrs->affn_scope == WQ_AFFN_DFL) 3837 scope = wq_affn_dfl; 3838 else 3839 scope = attrs->affn_scope; 3840 3841 pt = &wq_pod_types[scope]; 3842 3843 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 3844 likely(pt->nr_pods)) 3845 return pt; 3846 3847 /* 3848 * Before workqueue_init_topology(), only SYSTEM is available which is 3849 * initialized in workqueue_init_early(). 3850 */ 3851 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 3852 BUG_ON(!pt->nr_pods); 3853 return pt; 3854 } 3855 3856 /** 3857 * init_worker_pool - initialize a newly zalloc'd worker_pool 3858 * @pool: worker_pool to initialize 3859 * 3860 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3861 * 3862 * Return: 0 on success, -errno on failure. Even on failure, all fields 3863 * inside @pool proper are initialized and put_unbound_pool() can be called 3864 * on @pool safely to release it. 3865 */ 3866 static int init_worker_pool(struct worker_pool *pool) 3867 { 3868 raw_spin_lock_init(&pool->lock); 3869 pool->id = -1; 3870 pool->cpu = -1; 3871 pool->node = NUMA_NO_NODE; 3872 pool->flags |= POOL_DISASSOCIATED; 3873 pool->watchdog_ts = jiffies; 3874 INIT_LIST_HEAD(&pool->worklist); 3875 INIT_LIST_HEAD(&pool->idle_list); 3876 hash_init(pool->busy_hash); 3877 3878 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 3879 INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 3880 3881 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 3882 3883 INIT_LIST_HEAD(&pool->workers); 3884 INIT_LIST_HEAD(&pool->dying_workers); 3885 3886 ida_init(&pool->worker_ida); 3887 INIT_HLIST_NODE(&pool->hash_node); 3888 pool->refcnt = 1; 3889 3890 /* shouldn't fail above this point */ 3891 pool->attrs = alloc_workqueue_attrs(); 3892 if (!pool->attrs) 3893 return -ENOMEM; 3894 3895 wqattrs_clear_for_pool(pool->attrs); 3896 3897 return 0; 3898 } 3899 3900 #ifdef CONFIG_LOCKDEP 3901 static void wq_init_lockdep(struct workqueue_struct *wq) 3902 { 3903 char *lock_name; 3904 3905 lockdep_register_key(&wq->key); 3906 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3907 if (!lock_name) 3908 lock_name = wq->name; 3909 3910 wq->lock_name = lock_name; 3911 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3912 } 3913 3914 static void wq_unregister_lockdep(struct workqueue_struct *wq) 3915 { 3916 lockdep_unregister_key(&wq->key); 3917 } 3918 3919 static void wq_free_lockdep(struct workqueue_struct *wq) 3920 { 3921 if (wq->lock_name != wq->name) 3922 kfree(wq->lock_name); 3923 } 3924 #else 3925 static void wq_init_lockdep(struct workqueue_struct *wq) 3926 { 3927 } 3928 3929 static void wq_unregister_lockdep(struct workqueue_struct *wq) 3930 { 3931 } 3932 3933 static void wq_free_lockdep(struct workqueue_struct *wq) 3934 { 3935 } 3936 #endif 3937 3938 static void rcu_free_wq(struct rcu_head *rcu) 3939 { 3940 struct workqueue_struct *wq = 3941 container_of(rcu, struct workqueue_struct, rcu); 3942 3943 wq_free_lockdep(wq); 3944 free_percpu(wq->cpu_pwq); 3945 free_workqueue_attrs(wq->unbound_attrs); 3946 kfree(wq); 3947 } 3948 3949 static void rcu_free_pool(struct rcu_head *rcu) 3950 { 3951 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 3952 3953 ida_destroy(&pool->worker_ida); 3954 free_workqueue_attrs(pool->attrs); 3955 kfree(pool); 3956 } 3957 3958 /** 3959 * put_unbound_pool - put a worker_pool 3960 * @pool: worker_pool to put 3961 * 3962 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3963 * safe manner. get_unbound_pool() calls this function on its failure path 3964 * and this function should be able to release pools which went through, 3965 * successfully or not, init_worker_pool(). 3966 * 3967 * Should be called with wq_pool_mutex held. 3968 */ 3969 static void put_unbound_pool(struct worker_pool *pool) 3970 { 3971 DECLARE_COMPLETION_ONSTACK(detach_completion); 3972 struct worker *worker; 3973 LIST_HEAD(cull_list); 3974 3975 lockdep_assert_held(&wq_pool_mutex); 3976 3977 if (--pool->refcnt) 3978 return; 3979 3980 /* sanity checks */ 3981 if (WARN_ON(!(pool->cpu < 0)) || 3982 WARN_ON(!list_empty(&pool->worklist))) 3983 return; 3984 3985 /* release id and unhash */ 3986 if (pool->id >= 0) 3987 idr_remove(&worker_pool_idr, pool->id); 3988 hash_del(&pool->hash_node); 3989 3990 /* 3991 * Become the manager and destroy all workers. This prevents 3992 * @pool's workers from blocking on attach_mutex. We're the last 3993 * manager and @pool gets freed with the flag set. 3994 * 3995 * Having a concurrent manager is quite unlikely to happen as we can 3996 * only get here with 3997 * pwq->refcnt == pool->refcnt == 0 3998 * which implies no work queued to the pool, which implies no worker can 3999 * become the manager. However a worker could have taken the role of 4000 * manager before the refcnts dropped to 0, since maybe_create_worker() 4001 * drops pool->lock 4002 */ 4003 while (true) { 4004 rcuwait_wait_event(&manager_wait, 4005 !(pool->flags & POOL_MANAGER_ACTIVE), 4006 TASK_UNINTERRUPTIBLE); 4007 4008 mutex_lock(&wq_pool_attach_mutex); 4009 raw_spin_lock_irq(&pool->lock); 4010 if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 4011 pool->flags |= POOL_MANAGER_ACTIVE; 4012 break; 4013 } 4014 raw_spin_unlock_irq(&pool->lock); 4015 mutex_unlock(&wq_pool_attach_mutex); 4016 } 4017 4018 while ((worker = first_idle_worker(pool))) 4019 set_worker_dying(worker, &cull_list); 4020 WARN_ON(pool->nr_workers || pool->nr_idle); 4021 raw_spin_unlock_irq(&pool->lock); 4022 4023 wake_dying_workers(&cull_list); 4024 4025 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 4026 pool->detach_completion = &detach_completion; 4027 mutex_unlock(&wq_pool_attach_mutex); 4028 4029 if (pool->detach_completion) 4030 wait_for_completion(pool->detach_completion); 4031 4032 /* shut down the timers */ 4033 del_timer_sync(&pool->idle_timer); 4034 cancel_work_sync(&pool->idle_cull_work); 4035 del_timer_sync(&pool->mayday_timer); 4036 4037 /* RCU protected to allow dereferences from get_work_pool() */ 4038 call_rcu(&pool->rcu, rcu_free_pool); 4039 } 4040 4041 /** 4042 * get_unbound_pool - get a worker_pool with the specified attributes 4043 * @attrs: the attributes of the worker_pool to get 4044 * 4045 * Obtain a worker_pool which has the same attributes as @attrs, bump the 4046 * reference count and return it. If there already is a matching 4047 * worker_pool, it will be used; otherwise, this function attempts to 4048 * create a new one. 4049 * 4050 * Should be called with wq_pool_mutex held. 4051 * 4052 * Return: On success, a worker_pool with the same attributes as @attrs. 4053 * On failure, %NULL. 4054 */ 4055 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 4056 { 4057 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 4058 u32 hash = wqattrs_hash(attrs); 4059 struct worker_pool *pool; 4060 int pod, node = NUMA_NO_NODE; 4061 4062 lockdep_assert_held(&wq_pool_mutex); 4063 4064 /* do we already have a matching pool? */ 4065 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 4066 if (wqattrs_equal(pool->attrs, attrs)) { 4067 pool->refcnt++; 4068 return pool; 4069 } 4070 } 4071 4072 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 4073 for (pod = 0; pod < pt->nr_pods; pod++) { 4074 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 4075 node = pt->pod_node[pod]; 4076 break; 4077 } 4078 } 4079 4080 /* nope, create a new one */ 4081 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 4082 if (!pool || init_worker_pool(pool) < 0) 4083 goto fail; 4084 4085 pool->node = node; 4086 copy_workqueue_attrs(pool->attrs, attrs); 4087 wqattrs_clear_for_pool(pool->attrs); 4088 4089 if (worker_pool_assign_id(pool) < 0) 4090 goto fail; 4091 4092 /* create and start the initial worker */ 4093 if (wq_online && !create_worker(pool)) 4094 goto fail; 4095 4096 /* install */ 4097 hash_add(unbound_pool_hash, &pool->hash_node, hash); 4098 4099 return pool; 4100 fail: 4101 if (pool) 4102 put_unbound_pool(pool); 4103 return NULL; 4104 } 4105 4106 static void rcu_free_pwq(struct rcu_head *rcu) 4107 { 4108 kmem_cache_free(pwq_cache, 4109 container_of(rcu, struct pool_workqueue, rcu)); 4110 } 4111 4112 /* 4113 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 4114 * refcnt and needs to be destroyed. 4115 */ 4116 static void pwq_release_workfn(struct kthread_work *work) 4117 { 4118 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 4119 release_work); 4120 struct workqueue_struct *wq = pwq->wq; 4121 struct worker_pool *pool = pwq->pool; 4122 bool is_last = false; 4123 4124 /* 4125 * When @pwq is not linked, it doesn't hold any reference to the 4126 * @wq, and @wq is invalid to access. 4127 */ 4128 if (!list_empty(&pwq->pwqs_node)) { 4129 mutex_lock(&wq->mutex); 4130 list_del_rcu(&pwq->pwqs_node); 4131 is_last = list_empty(&wq->pwqs); 4132 mutex_unlock(&wq->mutex); 4133 } 4134 4135 if (wq->flags & WQ_UNBOUND) { 4136 mutex_lock(&wq_pool_mutex); 4137 put_unbound_pool(pool); 4138 mutex_unlock(&wq_pool_mutex); 4139 } 4140 4141 call_rcu(&pwq->rcu, rcu_free_pwq); 4142 4143 /* 4144 * If we're the last pwq going away, @wq is already dead and no one 4145 * is gonna access it anymore. Schedule RCU free. 4146 */ 4147 if (is_last) { 4148 wq_unregister_lockdep(wq); 4149 call_rcu(&wq->rcu, rcu_free_wq); 4150 } 4151 } 4152 4153 /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4154 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4155 struct worker_pool *pool) 4156 { 4157 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4158 4159 memset(pwq, 0, sizeof(*pwq)); 4160 4161 pwq->pool = pool; 4162 pwq->wq = wq; 4163 pwq->flush_color = -1; 4164 pwq->refcnt = 1; 4165 INIT_LIST_HEAD(&pwq->inactive_works); 4166 INIT_LIST_HEAD(&pwq->pwqs_node); 4167 INIT_LIST_HEAD(&pwq->mayday_node); 4168 kthread_init_work(&pwq->release_work, pwq_release_workfn); 4169 } 4170 4171 /* sync @pwq with the current state of its associated wq and link it */ 4172 static void link_pwq(struct pool_workqueue *pwq) 4173 { 4174 struct workqueue_struct *wq = pwq->wq; 4175 4176 lockdep_assert_held(&wq->mutex); 4177 4178 /* may be called multiple times, ignore if already linked */ 4179 if (!list_empty(&pwq->pwqs_node)) 4180 return; 4181 4182 /* set the matching work_color */ 4183 pwq->work_color = wq->work_color; 4184 4185 /* link in @pwq */ 4186 list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4187 } 4188 4189 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4190 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4191 const struct workqueue_attrs *attrs) 4192 { 4193 struct worker_pool *pool; 4194 struct pool_workqueue *pwq; 4195 4196 lockdep_assert_held(&wq_pool_mutex); 4197 4198 pool = get_unbound_pool(attrs); 4199 if (!pool) 4200 return NULL; 4201 4202 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4203 if (!pwq) { 4204 put_unbound_pool(pool); 4205 return NULL; 4206 } 4207 4208 init_pwq(pwq, wq, pool); 4209 return pwq; 4210 } 4211 4212 /** 4213 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 4214 * @attrs: the wq_attrs of the default pwq of the target workqueue 4215 * @cpu: the target CPU 4216 * @cpu_going_down: if >= 0, the CPU to consider as offline 4217 * 4218 * Calculate the cpumask a workqueue with @attrs should use on @pod. If 4219 * @cpu_going_down is >= 0, that cpu is considered offline during calculation. 4220 * The result is stored in @attrs->__pod_cpumask. 4221 * 4222 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 4223 * and @pod has online CPUs requested by @attrs, the returned cpumask is the 4224 * intersection of the possible CPUs of @pod and @attrs->cpumask. 4225 * 4226 * The caller is responsible for ensuring that the cpumask of @pod stays stable. 4227 */ 4228 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, 4229 int cpu_going_down) 4230 { 4231 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 4232 int pod = pt->cpu_pod[cpu]; 4233 4234 /* does @pod have any online CPUs @attrs wants? */ 4235 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 4236 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); 4237 if (cpu_going_down >= 0) 4238 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); 4239 4240 if (cpumask_empty(attrs->__pod_cpumask)) { 4241 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 4242 return; 4243 } 4244 4245 /* yeap, return possible CPUs in @pod that @attrs wants */ 4246 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); 4247 4248 if (cpumask_empty(attrs->__pod_cpumask)) 4249 pr_warn_once("WARNING: workqueue cpumask: online intersect > " 4250 "possible intersect\n"); 4251 } 4252 4253 /* install @pwq into @wq's cpu_pwq and return the old pwq */ 4254 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 4255 int cpu, struct pool_workqueue *pwq) 4256 { 4257 struct pool_workqueue *old_pwq; 4258 4259 lockdep_assert_held(&wq_pool_mutex); 4260 lockdep_assert_held(&wq->mutex); 4261 4262 /* link_pwq() can handle duplicate calls */ 4263 link_pwq(pwq); 4264 4265 old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4266 rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq); 4267 return old_pwq; 4268 } 4269 4270 /* context to store the prepared attrs & pwqs before applying */ 4271 struct apply_wqattrs_ctx { 4272 struct workqueue_struct *wq; /* target workqueue */ 4273 struct workqueue_attrs *attrs; /* attrs to apply */ 4274 struct list_head list; /* queued for batching commit */ 4275 struct pool_workqueue *dfl_pwq; 4276 struct pool_workqueue *pwq_tbl[]; 4277 }; 4278 4279 /* free the resources after success or abort */ 4280 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 4281 { 4282 if (ctx) { 4283 int cpu; 4284 4285 for_each_possible_cpu(cpu) 4286 put_pwq_unlocked(ctx->pwq_tbl[cpu]); 4287 put_pwq_unlocked(ctx->dfl_pwq); 4288 4289 free_workqueue_attrs(ctx->attrs); 4290 4291 kfree(ctx); 4292 } 4293 } 4294 4295 /* allocate the attrs and pwqs for later installation */ 4296 static struct apply_wqattrs_ctx * 4297 apply_wqattrs_prepare(struct workqueue_struct *wq, 4298 const struct workqueue_attrs *attrs, 4299 const cpumask_var_t unbound_cpumask) 4300 { 4301 struct apply_wqattrs_ctx *ctx; 4302 struct workqueue_attrs *new_attrs; 4303 int cpu; 4304 4305 lockdep_assert_held(&wq_pool_mutex); 4306 4307 if (WARN_ON(attrs->affn_scope < 0 || 4308 attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 4309 return ERR_PTR(-EINVAL); 4310 4311 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 4312 4313 new_attrs = alloc_workqueue_attrs(); 4314 if (!ctx || !new_attrs) 4315 goto out_free; 4316 4317 /* 4318 * If something goes wrong during CPU up/down, we'll fall back to 4319 * the default pwq covering whole @attrs->cpumask. Always create 4320 * it even if we don't use it immediately. 4321 */ 4322 copy_workqueue_attrs(new_attrs, attrs); 4323 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 4324 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 4325 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 4326 if (!ctx->dfl_pwq) 4327 goto out_free; 4328 4329 for_each_possible_cpu(cpu) { 4330 if (new_attrs->ordered) { 4331 ctx->dfl_pwq->refcnt++; 4332 ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 4333 } else { 4334 wq_calc_pod_cpumask(new_attrs, cpu, -1); 4335 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 4336 if (!ctx->pwq_tbl[cpu]) 4337 goto out_free; 4338 } 4339 } 4340 4341 /* save the user configured attrs and sanitize it. */ 4342 copy_workqueue_attrs(new_attrs, attrs); 4343 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 4344 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 4345 ctx->attrs = new_attrs; 4346 4347 ctx->wq = wq; 4348 return ctx; 4349 4350 out_free: 4351 free_workqueue_attrs(new_attrs); 4352 apply_wqattrs_cleanup(ctx); 4353 return ERR_PTR(-ENOMEM); 4354 } 4355 4356 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 4357 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 4358 { 4359 int cpu; 4360 4361 /* all pwqs have been created successfully, let's install'em */ 4362 mutex_lock(&ctx->wq->mutex); 4363 4364 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 4365 4366 /* save the previous pwq and install the new one */ 4367 for_each_possible_cpu(cpu) 4368 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 4369 ctx->pwq_tbl[cpu]); 4370 4371 /* @dfl_pwq might not have been used, ensure it's linked */ 4372 link_pwq(ctx->dfl_pwq); 4373 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 4374 4375 mutex_unlock(&ctx->wq->mutex); 4376 } 4377 4378 static void apply_wqattrs_lock(void) 4379 { 4380 /* CPUs should stay stable across pwq creations and installations */ 4381 cpus_read_lock(); 4382 mutex_lock(&wq_pool_mutex); 4383 } 4384 4385 static void apply_wqattrs_unlock(void) 4386 { 4387 mutex_unlock(&wq_pool_mutex); 4388 cpus_read_unlock(); 4389 } 4390 4391 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4392 const struct workqueue_attrs *attrs) 4393 { 4394 struct apply_wqattrs_ctx *ctx; 4395 4396 /* only unbound workqueues can change attributes */ 4397 if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4398 return -EINVAL; 4399 4400 /* creating multiple pwqs breaks ordering guarantee */ 4401 if (!list_empty(&wq->pwqs)) { 4402 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4403 return -EINVAL; 4404 4405 wq->flags &= ~__WQ_ORDERED; 4406 } 4407 4408 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 4409 if (IS_ERR(ctx)) 4410 return PTR_ERR(ctx); 4411 4412 /* the ctx has been prepared successfully, let's commit it */ 4413 apply_wqattrs_commit(ctx); 4414 apply_wqattrs_cleanup(ctx); 4415 4416 return 0; 4417 } 4418 4419 /** 4420 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 4421 * @wq: the target workqueue 4422 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 4423 * 4424 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 4425 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 4426 * work items are affine to the pod it was issued on. Older pwqs are released as 4427 * in-flight work items finish. Note that a work item which repeatedly requeues 4428 * itself back-to-back will stay on its current pwq. 4429 * 4430 * Performs GFP_KERNEL allocations. 4431 * 4432 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4433 * 4434 * Return: 0 on success and -errno on failure. 4435 */ 4436 int apply_workqueue_attrs(struct workqueue_struct *wq, 4437 const struct workqueue_attrs *attrs) 4438 { 4439 int ret; 4440 4441 lockdep_assert_cpus_held(); 4442 4443 mutex_lock(&wq_pool_mutex); 4444 ret = apply_workqueue_attrs_locked(wq, attrs); 4445 mutex_unlock(&wq_pool_mutex); 4446 4447 return ret; 4448 } 4449 4450 /** 4451 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug 4452 * @wq: the target workqueue 4453 * @cpu: the CPU to update pool association for 4454 * @hotplug_cpu: the CPU coming up or going down 4455 * @online: whether @cpu is coming up or going down 4456 * 4457 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 4458 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of 4459 * @wq accordingly. 4460 * 4461 * 4462 * If pod affinity can't be adjusted due to memory allocation failure, it falls 4463 * back to @wq->dfl_pwq which may not be optimal but is always correct. 4464 * 4465 * Note that when the last allowed CPU of a pod goes offline for a workqueue 4466 * with a cpumask spanning multiple pods, the workers which were already 4467 * executing the work items for the workqueue will lose their CPU affinity and 4468 * may execute on any CPU. This is similar to how per-cpu workqueues behave on 4469 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 4470 * responsibility to flush the work item from CPU_DOWN_PREPARE. 4471 */ 4472 static void wq_update_pod(struct workqueue_struct *wq, int cpu, 4473 int hotplug_cpu, bool online) 4474 { 4475 int off_cpu = online ? -1 : hotplug_cpu; 4476 struct pool_workqueue *old_pwq = NULL, *pwq; 4477 struct workqueue_attrs *target_attrs; 4478 4479 lockdep_assert_held(&wq_pool_mutex); 4480 4481 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 4482 return; 4483 4484 /* 4485 * We don't wanna alloc/free wq_attrs for each wq for each CPU. 4486 * Let's use a preallocated one. The following buf is protected by 4487 * CPU hotplug exclusion. 4488 */ 4489 target_attrs = wq_update_pod_attrs_buf; 4490 4491 copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 4492 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 4493 4494 /* nothing to do if the target cpumask matches the current pwq */ 4495 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); 4496 pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu), 4497 lockdep_is_held(&wq_pool_mutex)); 4498 if (wqattrs_equal(target_attrs, pwq->pool->attrs)) 4499 return; 4500 4501 /* create a new pwq */ 4502 pwq = alloc_unbound_pwq(wq, target_attrs); 4503 if (!pwq) { 4504 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 4505 wq->name); 4506 goto use_dfl_pwq; 4507 } 4508 4509 /* Install the new pwq. */ 4510 mutex_lock(&wq->mutex); 4511 old_pwq = install_unbound_pwq(wq, cpu, pwq); 4512 goto out_unlock; 4513 4514 use_dfl_pwq: 4515 mutex_lock(&wq->mutex); 4516 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 4517 get_pwq(wq->dfl_pwq); 4518 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 4519 old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq); 4520 out_unlock: 4521 mutex_unlock(&wq->mutex); 4522 put_pwq_unlocked(old_pwq); 4523 } 4524 4525 static int alloc_and_link_pwqs(struct workqueue_struct *wq) 4526 { 4527 bool highpri = wq->flags & WQ_HIGHPRI; 4528 int cpu, ret; 4529 4530 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 4531 if (!wq->cpu_pwq) 4532 goto enomem; 4533 4534 if (!(wq->flags & WQ_UNBOUND)) { 4535 for_each_possible_cpu(cpu) { 4536 struct pool_workqueue **pwq_p = 4537 per_cpu_ptr(wq->cpu_pwq, cpu); 4538 struct worker_pool *pool = 4539 &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); 4540 4541 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 4542 pool->node); 4543 if (!*pwq_p) 4544 goto enomem; 4545 4546 init_pwq(*pwq_p, wq, pool); 4547 4548 mutex_lock(&wq->mutex); 4549 link_pwq(*pwq_p); 4550 mutex_unlock(&wq->mutex); 4551 } 4552 return 0; 4553 } 4554 4555 cpus_read_lock(); 4556 if (wq->flags & __WQ_ORDERED) { 4557 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 4558 /* there should only be single pwq for ordering guarantee */ 4559 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 4560 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 4561 "ordering guarantee broken for workqueue %s\n", wq->name); 4562 } else { 4563 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 4564 } 4565 cpus_read_unlock(); 4566 4567 /* for unbound pwq, flush the pwq_release_worker ensures that the 4568 * pwq_release_workfn() completes before calling kfree(wq). 4569 */ 4570 if (ret) 4571 kthread_flush_worker(pwq_release_worker); 4572 4573 return ret; 4574 4575 enomem: 4576 if (wq->cpu_pwq) { 4577 for_each_possible_cpu(cpu) { 4578 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 4579 4580 if (pwq) 4581 kmem_cache_free(pwq_cache, pwq); 4582 } 4583 free_percpu(wq->cpu_pwq); 4584 wq->cpu_pwq = NULL; 4585 } 4586 return -ENOMEM; 4587 } 4588 4589 static int wq_clamp_max_active(int max_active, unsigned int flags, 4590 const char *name) 4591 { 4592 if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 4593 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4594 max_active, name, 1, WQ_MAX_ACTIVE); 4595 4596 return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 4597 } 4598 4599 /* 4600 * Workqueues which may be used during memory reclaim should have a rescuer 4601 * to guarantee forward progress. 4602 */ 4603 static int init_rescuer(struct workqueue_struct *wq) 4604 { 4605 struct worker *rescuer; 4606 int ret; 4607 4608 if (!(wq->flags & WQ_MEM_RECLAIM)) 4609 return 0; 4610 4611 rescuer = alloc_worker(NUMA_NO_NODE); 4612 if (!rescuer) { 4613 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 4614 wq->name); 4615 return -ENOMEM; 4616 } 4617 4618 rescuer->rescue_wq = wq; 4619 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name); 4620 if (IS_ERR(rescuer->task)) { 4621 ret = PTR_ERR(rescuer->task); 4622 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 4623 wq->name, ERR_PTR(ret)); 4624 kfree(rescuer); 4625 return ret; 4626 } 4627 4628 wq->rescuer = rescuer; 4629 kthread_bind_mask(rescuer->task, cpu_possible_mask); 4630 wake_up_process(rescuer->task); 4631 4632 return 0; 4633 } 4634 4635 /** 4636 * wq_adjust_max_active - update a wq's max_active to the current setting 4637 * @wq: target workqueue 4638 * 4639 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and 4640 * activate inactive work items accordingly. If @wq is freezing, clear 4641 * @wq->max_active to zero. 4642 */ 4643 static void wq_adjust_max_active(struct workqueue_struct *wq) 4644 { 4645 struct pool_workqueue *pwq; 4646 4647 lockdep_assert_held(&wq->mutex); 4648 4649 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { 4650 WRITE_ONCE(wq->max_active, 0); 4651 return; 4652 } 4653 4654 if (wq->max_active == wq->saved_max_active) 4655 return; 4656 4657 /* 4658 * Update @wq->max_active and then kick inactive work items if more 4659 * active work items are allowed. This doesn't break work item ordering 4660 * because new work items are always queued behind existing inactive 4661 * work items if there are any. 4662 */ 4663 WRITE_ONCE(wq->max_active, wq->saved_max_active); 4664 4665 for_each_pwq(pwq, wq) { 4666 unsigned long flags; 4667 4668 /* this function can be called during early boot w/ irq disabled */ 4669 raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4670 4671 while (!list_empty(&pwq->inactive_works) && 4672 pwq->nr_active < wq->max_active) 4673 pwq_activate_first_inactive(pwq); 4674 4675 kick_pool(pwq->pool); 4676 4677 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 4678 } 4679 } 4680 4681 __printf(1, 4) 4682 struct workqueue_struct *alloc_workqueue(const char *fmt, 4683 unsigned int flags, 4684 int max_active, ...) 4685 { 4686 va_list args; 4687 struct workqueue_struct *wq; 4688 int len; 4689 4690 /* 4691 * Unbound && max_active == 1 used to imply ordered, which is no longer 4692 * the case on many machines due to per-pod pools. While 4693 * alloc_ordered_workqueue() is the right way to create an ordered 4694 * workqueue, keep the previous behavior to avoid subtle breakages. 4695 */ 4696 if ((flags & WQ_UNBOUND) && max_active == 1) 4697 flags |= __WQ_ORDERED; 4698 4699 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4700 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4701 flags |= WQ_UNBOUND; 4702 4703 /* allocate wq and format name */ 4704 wq = kzalloc(sizeof(*wq), GFP_KERNEL); 4705 if (!wq) 4706 return NULL; 4707 4708 if (flags & WQ_UNBOUND) { 4709 wq->unbound_attrs = alloc_workqueue_attrs(); 4710 if (!wq->unbound_attrs) 4711 goto err_free_wq; 4712 } 4713 4714 va_start(args, max_active); 4715 len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4716 va_end(args); 4717 4718 if (len >= WQ_NAME_LEN) 4719 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", wq->name); 4720 4721 max_active = max_active ?: WQ_DFL_ACTIVE; 4722 max_active = wq_clamp_max_active(max_active, flags, wq->name); 4723 4724 /* init wq */ 4725 wq->flags = flags; 4726 wq->max_active = max_active; 4727 wq->saved_max_active = max_active; 4728 mutex_init(&wq->mutex); 4729 atomic_set(&wq->nr_pwqs_to_flush, 0); 4730 INIT_LIST_HEAD(&wq->pwqs); 4731 INIT_LIST_HEAD(&wq->flusher_queue); 4732 INIT_LIST_HEAD(&wq->flusher_overflow); 4733 INIT_LIST_HEAD(&wq->maydays); 4734 4735 wq_init_lockdep(wq); 4736 INIT_LIST_HEAD(&wq->list); 4737 4738 if (alloc_and_link_pwqs(wq) < 0) 4739 goto err_unreg_lockdep; 4740 4741 if (wq_online && init_rescuer(wq) < 0) 4742 goto err_destroy; 4743 4744 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4745 goto err_destroy; 4746 4747 /* 4748 * wq_pool_mutex protects global freeze state and workqueues list. 4749 * Grab it, adjust max_active and add the new @wq to workqueues 4750 * list. 4751 */ 4752 mutex_lock(&wq_pool_mutex); 4753 4754 mutex_lock(&wq->mutex); 4755 wq_adjust_max_active(wq); 4756 mutex_unlock(&wq->mutex); 4757 4758 list_add_tail_rcu(&wq->list, &workqueues); 4759 4760 mutex_unlock(&wq_pool_mutex); 4761 4762 return wq; 4763 4764 err_unreg_lockdep: 4765 wq_unregister_lockdep(wq); 4766 wq_free_lockdep(wq); 4767 err_free_wq: 4768 free_workqueue_attrs(wq->unbound_attrs); 4769 kfree(wq); 4770 return NULL; 4771 err_destroy: 4772 destroy_workqueue(wq); 4773 return NULL; 4774 } 4775 EXPORT_SYMBOL_GPL(alloc_workqueue); 4776 4777 static bool pwq_busy(struct pool_workqueue *pwq) 4778 { 4779 int i; 4780 4781 for (i = 0; i < WORK_NR_COLORS; i++) 4782 if (pwq->nr_in_flight[i]) 4783 return true; 4784 4785 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4786 return true; 4787 if (!pwq_is_empty(pwq)) 4788 return true; 4789 4790 return false; 4791 } 4792 4793 /** 4794 * destroy_workqueue - safely terminate a workqueue 4795 * @wq: target workqueue 4796 * 4797 * Safely destroy a workqueue. All work currently pending will be done first. 4798 */ 4799 void destroy_workqueue(struct workqueue_struct *wq) 4800 { 4801 struct pool_workqueue *pwq; 4802 int cpu; 4803 4804 /* 4805 * Remove it from sysfs first so that sanity check failure doesn't 4806 * lead to sysfs name conflicts. 4807 */ 4808 workqueue_sysfs_unregister(wq); 4809 4810 /* mark the workqueue destruction is in progress */ 4811 mutex_lock(&wq->mutex); 4812 wq->flags |= __WQ_DESTROYING; 4813 mutex_unlock(&wq->mutex); 4814 4815 /* drain it before proceeding with destruction */ 4816 drain_workqueue(wq); 4817 4818 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4819 if (wq->rescuer) { 4820 struct worker *rescuer = wq->rescuer; 4821 4822 /* this prevents new queueing */ 4823 raw_spin_lock_irq(&wq_mayday_lock); 4824 wq->rescuer = NULL; 4825 raw_spin_unlock_irq(&wq_mayday_lock); 4826 4827 /* rescuer will empty maydays list before exiting */ 4828 kthread_stop(rescuer->task); 4829 kfree(rescuer); 4830 } 4831 4832 /* 4833 * Sanity checks - grab all the locks so that we wait for all 4834 * in-flight operations which may do put_pwq(). 4835 */ 4836 mutex_lock(&wq_pool_mutex); 4837 mutex_lock(&wq->mutex); 4838 for_each_pwq(pwq, wq) { 4839 raw_spin_lock_irq(&pwq->pool->lock); 4840 if (WARN_ON(pwq_busy(pwq))) { 4841 pr_warn("%s: %s has the following busy pwq\n", 4842 __func__, wq->name); 4843 show_pwq(pwq); 4844 raw_spin_unlock_irq(&pwq->pool->lock); 4845 mutex_unlock(&wq->mutex); 4846 mutex_unlock(&wq_pool_mutex); 4847 show_one_workqueue(wq); 4848 return; 4849 } 4850 raw_spin_unlock_irq(&pwq->pool->lock); 4851 } 4852 mutex_unlock(&wq->mutex); 4853 4854 /* 4855 * wq list is used to freeze wq, remove from list after 4856 * flushing is complete in case freeze races us. 4857 */ 4858 list_del_rcu(&wq->list); 4859 mutex_unlock(&wq_pool_mutex); 4860 4861 /* 4862 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 4863 * to put the base refs. @wq will be auto-destroyed from the last 4864 * pwq_put. RCU read lock prevents @wq from going away from under us. 4865 */ 4866 rcu_read_lock(); 4867 4868 for_each_possible_cpu(cpu) { 4869 pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4870 RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL); 4871 put_pwq_unlocked(pwq); 4872 } 4873 4874 put_pwq_unlocked(wq->dfl_pwq); 4875 wq->dfl_pwq = NULL; 4876 4877 rcu_read_unlock(); 4878 } 4879 EXPORT_SYMBOL_GPL(destroy_workqueue); 4880 4881 /** 4882 * workqueue_set_max_active - adjust max_active of a workqueue 4883 * @wq: target workqueue 4884 * @max_active: new max_active value. 4885 * 4886 * Set max_active of @wq to @max_active. 4887 * 4888 * CONTEXT: 4889 * Don't call from IRQ context. 4890 */ 4891 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4892 { 4893 /* disallow meddling with max_active for ordered workqueues */ 4894 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4895 return; 4896 4897 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4898 4899 mutex_lock(&wq->mutex); 4900 4901 wq->flags &= ~__WQ_ORDERED; 4902 wq->saved_max_active = max_active; 4903 wq_adjust_max_active(wq); 4904 4905 mutex_unlock(&wq->mutex); 4906 } 4907 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4908 4909 /** 4910 * current_work - retrieve %current task's work struct 4911 * 4912 * Determine if %current task is a workqueue worker and what it's working on. 4913 * Useful to find out the context that the %current task is running in. 4914 * 4915 * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 4916 */ 4917 struct work_struct *current_work(void) 4918 { 4919 struct worker *worker = current_wq_worker(); 4920 4921 return worker ? worker->current_work : NULL; 4922 } 4923 EXPORT_SYMBOL(current_work); 4924 4925 /** 4926 * current_is_workqueue_rescuer - is %current workqueue rescuer? 4927 * 4928 * Determine whether %current is a workqueue rescuer. Can be used from 4929 * work functions to determine whether it's being run off the rescuer task. 4930 * 4931 * Return: %true if %current is a workqueue rescuer. %false otherwise. 4932 */ 4933 bool current_is_workqueue_rescuer(void) 4934 { 4935 struct worker *worker = current_wq_worker(); 4936 4937 return worker && worker->rescue_wq; 4938 } 4939 4940 /** 4941 * workqueue_congested - test whether a workqueue is congested 4942 * @cpu: CPU in question 4943 * @wq: target workqueue 4944 * 4945 * Test whether @wq's cpu workqueue for @cpu is congested. There is 4946 * no synchronization around this function and the test result is 4947 * unreliable and only useful as advisory hints or for debugging. 4948 * 4949 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4950 * 4951 * With the exception of ordered workqueues, all workqueues have per-cpu 4952 * pool_workqueues, each with its own congested state. A workqueue being 4953 * congested on one CPU doesn't mean that the workqueue is contested on any 4954 * other CPUs. 4955 * 4956 * Return: 4957 * %true if congested, %false otherwise. 4958 */ 4959 bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4960 { 4961 struct pool_workqueue *pwq; 4962 bool ret; 4963 4964 rcu_read_lock(); 4965 preempt_disable(); 4966 4967 if (cpu == WORK_CPU_UNBOUND) 4968 cpu = smp_processor_id(); 4969 4970 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 4971 ret = !list_empty(&pwq->inactive_works); 4972 4973 preempt_enable(); 4974 rcu_read_unlock(); 4975 4976 return ret; 4977 } 4978 EXPORT_SYMBOL_GPL(workqueue_congested); 4979 4980 /** 4981 * work_busy - test whether a work is currently pending or running 4982 * @work: the work to be tested 4983 * 4984 * Test whether @work is currently pending or running. There is no 4985 * synchronization around this function and the test result is 4986 * unreliable and only useful as advisory hints or for debugging. 4987 * 4988 * Return: 4989 * OR'd bitmask of WORK_BUSY_* bits. 4990 */ 4991 unsigned int work_busy(struct work_struct *work) 4992 { 4993 struct worker_pool *pool; 4994 unsigned long flags; 4995 unsigned int ret = 0; 4996 4997 if (work_pending(work)) 4998 ret |= WORK_BUSY_PENDING; 4999 5000 rcu_read_lock(); 5001 pool = get_work_pool(work); 5002 if (pool) { 5003 raw_spin_lock_irqsave(&pool->lock, flags); 5004 if (find_worker_executing_work(pool, work)) 5005 ret |= WORK_BUSY_RUNNING; 5006 raw_spin_unlock_irqrestore(&pool->lock, flags); 5007 } 5008 rcu_read_unlock(); 5009 5010 return ret; 5011 } 5012 EXPORT_SYMBOL_GPL(work_busy); 5013 5014 /** 5015 * set_worker_desc - set description for the current work item 5016 * @fmt: printf-style format string 5017 * @...: arguments for the format string 5018 * 5019 * This function can be called by a running work function to describe what 5020 * the work item is about. If the worker task gets dumped, this 5021 * information will be printed out together to help debugging. The 5022 * description can be at most WORKER_DESC_LEN including the trailing '\0'. 5023 */ 5024 void set_worker_desc(const char *fmt, ...) 5025 { 5026 struct worker *worker = current_wq_worker(); 5027 va_list args; 5028 5029 if (worker) { 5030 va_start(args, fmt); 5031 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 5032 va_end(args); 5033 } 5034 } 5035 EXPORT_SYMBOL_GPL(set_worker_desc); 5036 5037 /** 5038 * print_worker_info - print out worker information and description 5039 * @log_lvl: the log level to use when printing 5040 * @task: target task 5041 * 5042 * If @task is a worker and currently executing a work item, print out the 5043 * name of the workqueue being serviced and worker description set with 5044 * set_worker_desc() by the currently executing work item. 5045 * 5046 * This function can be safely called on any task as long as the 5047 * task_struct itself is accessible. While safe, this function isn't 5048 * synchronized and may print out mixups or garbages of limited length. 5049 */ 5050 void print_worker_info(const char *log_lvl, struct task_struct *task) 5051 { 5052 work_func_t *fn = NULL; 5053 char name[WQ_NAME_LEN] = { }; 5054 char desc[WORKER_DESC_LEN] = { }; 5055 struct pool_workqueue *pwq = NULL; 5056 struct workqueue_struct *wq = NULL; 5057 struct worker *worker; 5058 5059 if (!(task->flags & PF_WQ_WORKER)) 5060 return; 5061 5062 /* 5063 * This function is called without any synchronization and @task 5064 * could be in any state. Be careful with dereferences. 5065 */ 5066 worker = kthread_probe_data(task); 5067 5068 /* 5069 * Carefully copy the associated workqueue's workfn, name and desc. 5070 * Keep the original last '\0' in case the original is garbage. 5071 */ 5072 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5073 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5074 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5075 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5076 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 5077 5078 if (fn || name[0] || desc[0]) { 5079 printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 5080 if (strcmp(name, desc)) 5081 pr_cont(" (%s)", desc); 5082 pr_cont("\n"); 5083 } 5084 } 5085 5086 static void pr_cont_pool_info(struct worker_pool *pool) 5087 { 5088 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 5089 if (pool->node != NUMA_NO_NODE) 5090 pr_cont(" node=%d", pool->node); 5091 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 5092 } 5093 5094 struct pr_cont_work_struct { 5095 bool comma; 5096 work_func_t func; 5097 long ctr; 5098 }; 5099 5100 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5101 { 5102 if (!pcwsp->ctr) 5103 goto out_record; 5104 if (func == pcwsp->func) { 5105 pcwsp->ctr++; 5106 return; 5107 } 5108 if (pcwsp->ctr == 1) 5109 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5110 else 5111 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5112 pcwsp->ctr = 0; 5113 out_record: 5114 if ((long)func == -1L) 5115 return; 5116 pcwsp->comma = comma; 5117 pcwsp->func = func; 5118 pcwsp->ctr = 1; 5119 } 5120 5121 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 5122 { 5123 if (work->func == wq_barrier_func) { 5124 struct wq_barrier *barr; 5125 5126 barr = container_of(work, struct wq_barrier, work); 5127 5128 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5129 pr_cont("%s BAR(%d)", comma ? "," : "", 5130 task_pid_nr(barr->task)); 5131 } else { 5132 if (!comma) 5133 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5134 pr_cont_work_flush(comma, work->func, pcwsp); 5135 } 5136 } 5137 5138 static void show_pwq(struct pool_workqueue *pwq) 5139 { 5140 struct pr_cont_work_struct pcws = { .ctr = 0, }; 5141 struct worker_pool *pool = pwq->pool; 5142 struct work_struct *work; 5143 struct worker *worker; 5144 bool has_in_flight = false, has_pending = false; 5145 int bkt; 5146 5147 pr_info(" pwq %d:", pool->id); 5148 pr_cont_pool_info(pool); 5149 5150 pr_cont(" active=%d refcnt=%d%s\n", 5151 pwq->nr_active, pwq->refcnt, 5152 !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 5153 5154 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 5155 if (worker->current_pwq == pwq) { 5156 has_in_flight = true; 5157 break; 5158 } 5159 } 5160 if (has_in_flight) { 5161 bool comma = false; 5162 5163 pr_info(" in-flight:"); 5164 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 5165 if (worker->current_pwq != pwq) 5166 continue; 5167 5168 pr_cont("%s %d%s:%ps", comma ? "," : "", 5169 task_pid_nr(worker->task), 5170 worker->rescue_wq ? "(RESCUER)" : "", 5171 worker->current_func); 5172 list_for_each_entry(work, &worker->scheduled, entry) 5173 pr_cont_work(false, work, &pcws); 5174 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 5175 comma = true; 5176 } 5177 pr_cont("\n"); 5178 } 5179 5180 list_for_each_entry(work, &pool->worklist, entry) { 5181 if (get_work_pwq(work) == pwq) { 5182 has_pending = true; 5183 break; 5184 } 5185 } 5186 if (has_pending) { 5187 bool comma = false; 5188 5189 pr_info(" pending:"); 5190 list_for_each_entry(work, &pool->worklist, entry) { 5191 if (get_work_pwq(work) != pwq) 5192 continue; 5193 5194 pr_cont_work(comma, work, &pcws); 5195 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 5196 } 5197 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 5198 pr_cont("\n"); 5199 } 5200 5201 if (!list_empty(&pwq->inactive_works)) { 5202 bool comma = false; 5203 5204 pr_info(" inactive:"); 5205 list_for_each_entry(work, &pwq->inactive_works, entry) { 5206 pr_cont_work(comma, work, &pcws); 5207 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 5208 } 5209 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 5210 pr_cont("\n"); 5211 } 5212 } 5213 5214 /** 5215 * show_one_workqueue - dump state of specified workqueue 5216 * @wq: workqueue whose state will be printed 5217 */ 5218 void show_one_workqueue(struct workqueue_struct *wq) 5219 { 5220 struct pool_workqueue *pwq; 5221 bool idle = true; 5222 unsigned long flags; 5223 5224 for_each_pwq(pwq, wq) { 5225 if (!pwq_is_empty(pwq)) { 5226 idle = false; 5227 break; 5228 } 5229 } 5230 if (idle) /* Nothing to print for idle workqueue */ 5231 return; 5232 5233 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 5234 5235 for_each_pwq(pwq, wq) { 5236 raw_spin_lock_irqsave(&pwq->pool->lock, flags); 5237 if (!pwq_is_empty(pwq)) { 5238 /* 5239 * Defer printing to avoid deadlocks in console 5240 * drivers that queue work while holding locks 5241 * also taken in their write paths. 5242 */ 5243 printk_deferred_enter(); 5244 show_pwq(pwq); 5245 printk_deferred_exit(); 5246 } 5247 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 5248 /* 5249 * We could be printing a lot from atomic context, e.g. 5250 * sysrq-t -> show_all_workqueues(). Avoid triggering 5251 * hard lockup. 5252 */ 5253 touch_nmi_watchdog(); 5254 } 5255 5256 } 5257 5258 /** 5259 * show_one_worker_pool - dump state of specified worker pool 5260 * @pool: worker pool whose state will be printed 5261 */ 5262 static void show_one_worker_pool(struct worker_pool *pool) 5263 { 5264 struct worker *worker; 5265 bool first = true; 5266 unsigned long flags; 5267 unsigned long hung = 0; 5268 5269 raw_spin_lock_irqsave(&pool->lock, flags); 5270 if (pool->nr_workers == pool->nr_idle) 5271 goto next_pool; 5272 5273 /* How long the first pending work is waiting for a worker. */ 5274 if (!list_empty(&pool->worklist)) 5275 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5276 5277 /* 5278 * Defer printing to avoid deadlocks in console drivers that 5279 * queue work while holding locks also taken in their write 5280 * paths. 5281 */ 5282 printk_deferred_enter(); 5283 pr_info("pool %d:", pool->id); 5284 pr_cont_pool_info(pool); 5285 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 5286 if (pool->manager) 5287 pr_cont(" manager: %d", 5288 task_pid_nr(pool->manager->task)); 5289 list_for_each_entry(worker, &pool->idle_list, entry) { 5290 pr_cont(" %s%d", first ? "idle: " : "", 5291 task_pid_nr(worker->task)); 5292 first = false; 5293 } 5294 pr_cont("\n"); 5295 printk_deferred_exit(); 5296 next_pool: 5297 raw_spin_unlock_irqrestore(&pool->lock, flags); 5298 /* 5299 * We could be printing a lot from atomic context, e.g. 5300 * sysrq-t -> show_all_workqueues(). Avoid triggering 5301 * hard lockup. 5302 */ 5303 touch_nmi_watchdog(); 5304 5305 } 5306 5307 /** 5308 * show_all_workqueues - dump workqueue state 5309 * 5310 * Called from a sysrq handler and prints out all busy workqueues and pools. 5311 */ 5312 void show_all_workqueues(void) 5313 { 5314 struct workqueue_struct *wq; 5315 struct worker_pool *pool; 5316 int pi; 5317 5318 rcu_read_lock(); 5319 5320 pr_info("Showing busy workqueues and worker pools:\n"); 5321 5322 list_for_each_entry_rcu(wq, &workqueues, list) 5323 show_one_workqueue(wq); 5324 5325 for_each_pool(pool, pi) 5326 show_one_worker_pool(pool); 5327 5328 rcu_read_unlock(); 5329 } 5330 5331 /** 5332 * show_freezable_workqueues - dump freezable workqueue state 5333 * 5334 * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5335 * still busy. 5336 */ 5337 void show_freezable_workqueues(void) 5338 { 5339 struct workqueue_struct *wq; 5340 5341 rcu_read_lock(); 5342 5343 pr_info("Showing freezable workqueues that are still busy:\n"); 5344 5345 list_for_each_entry_rcu(wq, &workqueues, list) { 5346 if (!(wq->flags & WQ_FREEZABLE)) 5347 continue; 5348 show_one_workqueue(wq); 5349 } 5350 5351 rcu_read_unlock(); 5352 } 5353 5354 /* used to show worker information through /proc/PID/{comm,stat,status} */ 5355 void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 5356 { 5357 int off; 5358 5359 /* always show the actual comm */ 5360 off = strscpy(buf, task->comm, size); 5361 if (off < 0) 5362 return; 5363 5364 /* stabilize PF_WQ_WORKER and worker pool association */ 5365 mutex_lock(&wq_pool_attach_mutex); 5366 5367 if (task->flags & PF_WQ_WORKER) { 5368 struct worker *worker = kthread_data(task); 5369 struct worker_pool *pool = worker->pool; 5370 5371 if (pool) { 5372 raw_spin_lock_irq(&pool->lock); 5373 /* 5374 * ->desc tracks information (wq name or 5375 * set_worker_desc()) for the latest execution. If 5376 * current, prepend '+', otherwise '-'. 5377 */ 5378 if (worker->desc[0] != '\0') { 5379 if (worker->current_work) 5380 scnprintf(buf + off, size - off, "+%s", 5381 worker->desc); 5382 else 5383 scnprintf(buf + off, size - off, "-%s", 5384 worker->desc); 5385 } 5386 raw_spin_unlock_irq(&pool->lock); 5387 } 5388 } 5389 5390 mutex_unlock(&wq_pool_attach_mutex); 5391 } 5392 5393 #ifdef CONFIG_SMP 5394 5395 /* 5396 * CPU hotplug. 5397 * 5398 * There are two challenges in supporting CPU hotplug. Firstly, there 5399 * are a lot of assumptions on strong associations among work, pwq and 5400 * pool which make migrating pending and scheduled works very 5401 * difficult to implement without impacting hot paths. Secondly, 5402 * worker pools serve mix of short, long and very long running works making 5403 * blocked draining impractical. 5404 * 5405 * This is solved by allowing the pools to be disassociated from the CPU 5406 * running as an unbound one and allowing it to be reattached later if the 5407 * cpu comes back online. 5408 */ 5409 5410 static void unbind_workers(int cpu) 5411 { 5412 struct worker_pool *pool; 5413 struct worker *worker; 5414 5415 for_each_cpu_worker_pool(pool, cpu) { 5416 mutex_lock(&wq_pool_attach_mutex); 5417 raw_spin_lock_irq(&pool->lock); 5418 5419 /* 5420 * We've blocked all attach/detach operations. Make all workers 5421 * unbound and set DISASSOCIATED. Before this, all workers 5422 * must be on the cpu. After this, they may become diasporas. 5423 * And the preemption disabled section in their sched callbacks 5424 * are guaranteed to see WORKER_UNBOUND since the code here 5425 * is on the same cpu. 5426 */ 5427 for_each_pool_worker(worker, pool) 5428 worker->flags |= WORKER_UNBOUND; 5429 5430 pool->flags |= POOL_DISASSOCIATED; 5431 5432 /* 5433 * The handling of nr_running in sched callbacks are disabled 5434 * now. Zap nr_running. After this, nr_running stays zero and 5435 * need_more_worker() and keep_working() are always true as 5436 * long as the worklist is not empty. This pool now behaves as 5437 * an unbound (in terms of concurrency management) pool which 5438 * are served by workers tied to the pool. 5439 */ 5440 pool->nr_running = 0; 5441 5442 /* 5443 * With concurrency management just turned off, a busy 5444 * worker blocking could lead to lengthy stalls. Kick off 5445 * unbound chain execution of currently pending work items. 5446 */ 5447 kick_pool(pool); 5448 5449 raw_spin_unlock_irq(&pool->lock); 5450 5451 for_each_pool_worker(worker, pool) 5452 unbind_worker(worker); 5453 5454 mutex_unlock(&wq_pool_attach_mutex); 5455 } 5456 } 5457 5458 /** 5459 * rebind_workers - rebind all workers of a pool to the associated CPU 5460 * @pool: pool of interest 5461 * 5462 * @pool->cpu is coming online. Rebind all workers to the CPU. 5463 */ 5464 static void rebind_workers(struct worker_pool *pool) 5465 { 5466 struct worker *worker; 5467 5468 lockdep_assert_held(&wq_pool_attach_mutex); 5469 5470 /* 5471 * Restore CPU affinity of all workers. As all idle workers should 5472 * be on the run-queue of the associated CPU before any local 5473 * wake-ups for concurrency management happen, restore CPU affinity 5474 * of all workers first and then clear UNBOUND. As we're called 5475 * from CPU_ONLINE, the following shouldn't fail. 5476 */ 5477 for_each_pool_worker(worker, pool) { 5478 kthread_set_per_cpu(worker->task, pool->cpu); 5479 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5480 pool_allowed_cpus(pool)) < 0); 5481 } 5482 5483 raw_spin_lock_irq(&pool->lock); 5484 5485 pool->flags &= ~POOL_DISASSOCIATED; 5486 5487 for_each_pool_worker(worker, pool) { 5488 unsigned int worker_flags = worker->flags; 5489 5490 /* 5491 * We want to clear UNBOUND but can't directly call 5492 * worker_clr_flags() or adjust nr_running. Atomically 5493 * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5494 * @worker will clear REBOUND using worker_clr_flags() when 5495 * it initiates the next execution cycle thus restoring 5496 * concurrency management. Note that when or whether 5497 * @worker clears REBOUND doesn't affect correctness. 5498 * 5499 * WRITE_ONCE() is necessary because @worker->flags may be 5500 * tested without holding any lock in 5501 * wq_worker_running(). Without it, NOT_RUNNING test may 5502 * fail incorrectly leading to premature concurrency 5503 * management operations. 5504 */ 5505 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5506 worker_flags |= WORKER_REBOUND; 5507 worker_flags &= ~WORKER_UNBOUND; 5508 WRITE_ONCE(worker->flags, worker_flags); 5509 } 5510 5511 raw_spin_unlock_irq(&pool->lock); 5512 } 5513 5514 /** 5515 * restore_unbound_workers_cpumask - restore cpumask of unbound workers 5516 * @pool: unbound pool of interest 5517 * @cpu: the CPU which is coming up 5518 * 5519 * An unbound pool may end up with a cpumask which doesn't have any online 5520 * CPUs. When a worker of such pool get scheduled, the scheduler resets 5521 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 5522 * online CPU before, cpus_allowed of all its workers should be restored. 5523 */ 5524 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 5525 { 5526 static cpumask_t cpumask; 5527 struct worker *worker; 5528 5529 lockdep_assert_held(&wq_pool_attach_mutex); 5530 5531 /* is @cpu allowed for @pool? */ 5532 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 5533 return; 5534 5535 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 5536 5537 /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5538 for_each_pool_worker(worker, pool) 5539 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 5540 } 5541 5542 int workqueue_prepare_cpu(unsigned int cpu) 5543 { 5544 struct worker_pool *pool; 5545 5546 for_each_cpu_worker_pool(pool, cpu) { 5547 if (pool->nr_workers) 5548 continue; 5549 if (!create_worker(pool)) 5550 return -ENOMEM; 5551 } 5552 return 0; 5553 } 5554 5555 int workqueue_online_cpu(unsigned int cpu) 5556 { 5557 struct worker_pool *pool; 5558 struct workqueue_struct *wq; 5559 int pi; 5560 5561 mutex_lock(&wq_pool_mutex); 5562 5563 for_each_pool(pool, pi) { 5564 mutex_lock(&wq_pool_attach_mutex); 5565 5566 if (pool->cpu == cpu) 5567 rebind_workers(pool); 5568 else if (pool->cpu < 0) 5569 restore_unbound_workers_cpumask(pool, cpu); 5570 5571 mutex_unlock(&wq_pool_attach_mutex); 5572 } 5573 5574 /* update pod affinity of unbound workqueues */ 5575 list_for_each_entry(wq, &workqueues, list) { 5576 struct workqueue_attrs *attrs = wq->unbound_attrs; 5577 5578 if (attrs) { 5579 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5580 int tcpu; 5581 5582 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5583 wq_update_pod(wq, tcpu, cpu, true); 5584 } 5585 } 5586 5587 mutex_unlock(&wq_pool_mutex); 5588 return 0; 5589 } 5590 5591 int workqueue_offline_cpu(unsigned int cpu) 5592 { 5593 struct workqueue_struct *wq; 5594 5595 /* unbinding per-cpu workers should happen on the local CPU */ 5596 if (WARN_ON(cpu != smp_processor_id())) 5597 return -1; 5598 5599 unbind_workers(cpu); 5600 5601 /* update pod affinity of unbound workqueues */ 5602 mutex_lock(&wq_pool_mutex); 5603 list_for_each_entry(wq, &workqueues, list) { 5604 struct workqueue_attrs *attrs = wq->unbound_attrs; 5605 5606 if (attrs) { 5607 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5608 int tcpu; 5609 5610 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5611 wq_update_pod(wq, tcpu, cpu, false); 5612 } 5613 } 5614 mutex_unlock(&wq_pool_mutex); 5615 5616 return 0; 5617 } 5618 5619 struct work_for_cpu { 5620 struct work_struct work; 5621 long (*fn)(void *); 5622 void *arg; 5623 long ret; 5624 }; 5625 5626 static void work_for_cpu_fn(struct work_struct *work) 5627 { 5628 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5629 5630 wfc->ret = wfc->fn(wfc->arg); 5631 } 5632 5633 /** 5634 * work_on_cpu_key - run a function in thread context on a particular cpu 5635 * @cpu: the cpu to run on 5636 * @fn: the function to run 5637 * @arg: the function arg 5638 * @key: The lock class key for lock debugging purposes 5639 * 5640 * It is up to the caller to ensure that the cpu doesn't go offline. 5641 * The caller must not hold any locks which would prevent @fn from completing. 5642 * 5643 * Return: The value @fn returns. 5644 */ 5645 long work_on_cpu_key(int cpu, long (*fn)(void *), 5646 void *arg, struct lock_class_key *key) 5647 { 5648 struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 5649 5650 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); 5651 schedule_work_on(cpu, &wfc.work); 5652 flush_work(&wfc.work); 5653 destroy_work_on_stack(&wfc.work); 5654 return wfc.ret; 5655 } 5656 EXPORT_SYMBOL_GPL(work_on_cpu_key); 5657 5658 /** 5659 * work_on_cpu_safe_key - run a function in thread context on a particular cpu 5660 * @cpu: the cpu to run on 5661 * @fn: the function to run 5662 * @arg: the function argument 5663 * @key: The lock class key for lock debugging purposes 5664 * 5665 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 5666 * any locks which would prevent @fn from completing. 5667 * 5668 * Return: The value @fn returns. 5669 */ 5670 long work_on_cpu_safe_key(int cpu, long (*fn)(void *), 5671 void *arg, struct lock_class_key *key) 5672 { 5673 long ret = -ENODEV; 5674 5675 cpus_read_lock(); 5676 if (cpu_online(cpu)) 5677 ret = work_on_cpu_key(cpu, fn, arg, key); 5678 cpus_read_unlock(); 5679 return ret; 5680 } 5681 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); 5682 #endif /* CONFIG_SMP */ 5683 5684 #ifdef CONFIG_FREEZER 5685 5686 /** 5687 * freeze_workqueues_begin - begin freezing workqueues 5688 * 5689 * Start freezing workqueues. After this function returns, all freezable 5690 * workqueues will queue new works to their inactive_works list instead of 5691 * pool->worklist. 5692 * 5693 * CONTEXT: 5694 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5695 */ 5696 void freeze_workqueues_begin(void) 5697 { 5698 struct workqueue_struct *wq; 5699 5700 mutex_lock(&wq_pool_mutex); 5701 5702 WARN_ON_ONCE(workqueue_freezing); 5703 workqueue_freezing = true; 5704 5705 list_for_each_entry(wq, &workqueues, list) { 5706 mutex_lock(&wq->mutex); 5707 wq_adjust_max_active(wq); 5708 mutex_unlock(&wq->mutex); 5709 } 5710 5711 mutex_unlock(&wq_pool_mutex); 5712 } 5713 5714 /** 5715 * freeze_workqueues_busy - are freezable workqueues still busy? 5716 * 5717 * Check whether freezing is complete. This function must be called 5718 * between freeze_workqueues_begin() and thaw_workqueues(). 5719 * 5720 * CONTEXT: 5721 * Grabs and releases wq_pool_mutex. 5722 * 5723 * Return: 5724 * %true if some freezable workqueues are still busy. %false if freezing 5725 * is complete. 5726 */ 5727 bool freeze_workqueues_busy(void) 5728 { 5729 bool busy = false; 5730 struct workqueue_struct *wq; 5731 struct pool_workqueue *pwq; 5732 5733 mutex_lock(&wq_pool_mutex); 5734 5735 WARN_ON_ONCE(!workqueue_freezing); 5736 5737 list_for_each_entry(wq, &workqueues, list) { 5738 if (!(wq->flags & WQ_FREEZABLE)) 5739 continue; 5740 /* 5741 * nr_active is monotonically decreasing. It's safe 5742 * to peek without lock. 5743 */ 5744 rcu_read_lock(); 5745 for_each_pwq(pwq, wq) { 5746 WARN_ON_ONCE(pwq->nr_active < 0); 5747 if (pwq->nr_active) { 5748 busy = true; 5749 rcu_read_unlock(); 5750 goto out_unlock; 5751 } 5752 } 5753 rcu_read_unlock(); 5754 } 5755 out_unlock: 5756 mutex_unlock(&wq_pool_mutex); 5757 return busy; 5758 } 5759 5760 /** 5761 * thaw_workqueues - thaw workqueues 5762 * 5763 * Thaw workqueues. Normal queueing is restored and all collected 5764 * frozen works are transferred to their respective pool worklists. 5765 * 5766 * CONTEXT: 5767 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5768 */ 5769 void thaw_workqueues(void) 5770 { 5771 struct workqueue_struct *wq; 5772 5773 mutex_lock(&wq_pool_mutex); 5774 5775 if (!workqueue_freezing) 5776 goto out_unlock; 5777 5778 workqueue_freezing = false; 5779 5780 /* restore max_active and repopulate worklist */ 5781 list_for_each_entry(wq, &workqueues, list) { 5782 mutex_lock(&wq->mutex); 5783 wq_adjust_max_active(wq); 5784 mutex_unlock(&wq->mutex); 5785 } 5786 5787 out_unlock: 5788 mutex_unlock(&wq_pool_mutex); 5789 } 5790 #endif /* CONFIG_FREEZER */ 5791 5792 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5793 { 5794 LIST_HEAD(ctxs); 5795 int ret = 0; 5796 struct workqueue_struct *wq; 5797 struct apply_wqattrs_ctx *ctx, *n; 5798 5799 lockdep_assert_held(&wq_pool_mutex); 5800 5801 list_for_each_entry(wq, &workqueues, list) { 5802 if (!(wq->flags & WQ_UNBOUND)) 5803 continue; 5804 /* creating multiple pwqs breaks ordering guarantee */ 5805 if (wq->flags & __WQ_ORDERED) 5806 continue; 5807 5808 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5809 if (IS_ERR(ctx)) { 5810 ret = PTR_ERR(ctx); 5811 break; 5812 } 5813 5814 list_add_tail(&ctx->list, &ctxs); 5815 } 5816 5817 list_for_each_entry_safe(ctx, n, &ctxs, list) { 5818 if (!ret) 5819 apply_wqattrs_commit(ctx); 5820 apply_wqattrs_cleanup(ctx); 5821 } 5822 5823 if (!ret) { 5824 mutex_lock(&wq_pool_attach_mutex); 5825 cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 5826 mutex_unlock(&wq_pool_attach_mutex); 5827 } 5828 return ret; 5829 } 5830 5831 /** 5832 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5833 * @cpumask: the cpumask to set 5834 * 5835 * The low-level workqueues cpumask is a global cpumask that limits 5836 * the affinity of all unbound workqueues. This function check the @cpumask 5837 * and apply it to all unbound workqueues and updates all pwqs of them. 5838 * 5839 * Return: 0 - Success 5840 * -EINVAL - Invalid @cpumask 5841 * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5842 */ 5843 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5844 { 5845 int ret = -EINVAL; 5846 5847 /* 5848 * Not excluding isolated cpus on purpose. 5849 * If the user wishes to include them, we allow that. 5850 */ 5851 cpumask_and(cpumask, cpumask, cpu_possible_mask); 5852 if (!cpumask_empty(cpumask)) { 5853 apply_wqattrs_lock(); 5854 if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5855 ret = 0; 5856 goto out_unlock; 5857 } 5858 5859 ret = workqueue_apply_unbound_cpumask(cpumask); 5860 5861 out_unlock: 5862 apply_wqattrs_unlock(); 5863 } 5864 5865 return ret; 5866 } 5867 5868 static int parse_affn_scope(const char *val) 5869 { 5870 int i; 5871 5872 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { 5873 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) 5874 return i; 5875 } 5876 return -EINVAL; 5877 } 5878 5879 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 5880 { 5881 struct workqueue_struct *wq; 5882 int affn, cpu; 5883 5884 affn = parse_affn_scope(val); 5885 if (affn < 0) 5886 return affn; 5887 if (affn == WQ_AFFN_DFL) 5888 return -EINVAL; 5889 5890 cpus_read_lock(); 5891 mutex_lock(&wq_pool_mutex); 5892 5893 wq_affn_dfl = affn; 5894 5895 list_for_each_entry(wq, &workqueues, list) { 5896 for_each_online_cpu(cpu) { 5897 wq_update_pod(wq, cpu, cpu, true); 5898 } 5899 } 5900 5901 mutex_unlock(&wq_pool_mutex); 5902 cpus_read_unlock(); 5903 5904 return 0; 5905 } 5906 5907 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 5908 { 5909 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 5910 } 5911 5912 static const struct kernel_param_ops wq_affn_dfl_ops = { 5913 .set = wq_affn_dfl_set, 5914 .get = wq_affn_dfl_get, 5915 }; 5916 5917 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 5918 5919 #ifdef CONFIG_SYSFS 5920 /* 5921 * Workqueues with WQ_SYSFS flag set is visible to userland via 5922 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 5923 * following attributes. 5924 * 5925 * per_cpu RO bool : whether the workqueue is per-cpu or unbound 5926 * max_active RW int : maximum number of in-flight work items 5927 * 5928 * Unbound workqueues have the following extra attributes. 5929 * 5930 * nice RW int : nice value of the workers 5931 * cpumask RW mask : bitmask of allowed CPUs for the workers 5932 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 5933 * affinity_strict RW bool : worker CPU affinity is strict 5934 */ 5935 struct wq_device { 5936 struct workqueue_struct *wq; 5937 struct device dev; 5938 }; 5939 5940 static struct workqueue_struct *dev_to_wq(struct device *dev) 5941 { 5942 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 5943 5944 return wq_dev->wq; 5945 } 5946 5947 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 5948 char *buf) 5949 { 5950 struct workqueue_struct *wq = dev_to_wq(dev); 5951 5952 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 5953 } 5954 static DEVICE_ATTR_RO(per_cpu); 5955 5956 static ssize_t max_active_show(struct device *dev, 5957 struct device_attribute *attr, char *buf) 5958 { 5959 struct workqueue_struct *wq = dev_to_wq(dev); 5960 5961 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 5962 } 5963 5964 static ssize_t max_active_store(struct device *dev, 5965 struct device_attribute *attr, const char *buf, 5966 size_t count) 5967 { 5968 struct workqueue_struct *wq = dev_to_wq(dev); 5969 int val; 5970 5971 if (sscanf(buf, "%d", &val) != 1 || val <= 0) 5972 return -EINVAL; 5973 5974 workqueue_set_max_active(wq, val); 5975 return count; 5976 } 5977 static DEVICE_ATTR_RW(max_active); 5978 5979 static struct attribute *wq_sysfs_attrs[] = { 5980 &dev_attr_per_cpu.attr, 5981 &dev_attr_max_active.attr, 5982 NULL, 5983 }; 5984 ATTRIBUTE_GROUPS(wq_sysfs); 5985 5986 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 5987 char *buf) 5988 { 5989 struct workqueue_struct *wq = dev_to_wq(dev); 5990 int written; 5991 5992 mutex_lock(&wq->mutex); 5993 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 5994 mutex_unlock(&wq->mutex); 5995 5996 return written; 5997 } 5998 5999 /* prepare workqueue_attrs for sysfs store operations */ 6000 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 6001 { 6002 struct workqueue_attrs *attrs; 6003 6004 lockdep_assert_held(&wq_pool_mutex); 6005 6006 attrs = alloc_workqueue_attrs(); 6007 if (!attrs) 6008 return NULL; 6009 6010 copy_workqueue_attrs(attrs, wq->unbound_attrs); 6011 return attrs; 6012 } 6013 6014 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 6015 const char *buf, size_t count) 6016 { 6017 struct workqueue_struct *wq = dev_to_wq(dev); 6018 struct workqueue_attrs *attrs; 6019 int ret = -ENOMEM; 6020 6021 apply_wqattrs_lock(); 6022 6023 attrs = wq_sysfs_prep_attrs(wq); 6024 if (!attrs) 6025 goto out_unlock; 6026 6027 if (sscanf(buf, "%d", &attrs->nice) == 1 && 6028 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 6029 ret = apply_workqueue_attrs_locked(wq, attrs); 6030 else 6031 ret = -EINVAL; 6032 6033 out_unlock: 6034 apply_wqattrs_unlock(); 6035 free_workqueue_attrs(attrs); 6036 return ret ?: count; 6037 } 6038 6039 static ssize_t wq_cpumask_show(struct device *dev, 6040 struct device_attribute *attr, char *buf) 6041 { 6042 struct workqueue_struct *wq = dev_to_wq(dev); 6043 int written; 6044 6045 mutex_lock(&wq->mutex); 6046 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6047 cpumask_pr_args(wq->unbound_attrs->cpumask)); 6048 mutex_unlock(&wq->mutex); 6049 return written; 6050 } 6051 6052 static ssize_t wq_cpumask_store(struct device *dev, 6053 struct device_attribute *attr, 6054 const char *buf, size_t count) 6055 { 6056 struct workqueue_struct *wq = dev_to_wq(dev); 6057 struct workqueue_attrs *attrs; 6058 int ret = -ENOMEM; 6059 6060 apply_wqattrs_lock(); 6061 6062 attrs = wq_sysfs_prep_attrs(wq); 6063 if (!attrs) 6064 goto out_unlock; 6065 6066 ret = cpumask_parse(buf, attrs->cpumask); 6067 if (!ret) 6068 ret = apply_workqueue_attrs_locked(wq, attrs); 6069 6070 out_unlock: 6071 apply_wqattrs_unlock(); 6072 free_workqueue_attrs(attrs); 6073 return ret ?: count; 6074 } 6075 6076 static ssize_t wq_affn_scope_show(struct device *dev, 6077 struct device_attribute *attr, char *buf) 6078 { 6079 struct workqueue_struct *wq = dev_to_wq(dev); 6080 int written; 6081 6082 mutex_lock(&wq->mutex); 6083 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) 6084 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", 6085 wq_affn_names[WQ_AFFN_DFL], 6086 wq_affn_names[wq_affn_dfl]); 6087 else 6088 written = scnprintf(buf, PAGE_SIZE, "%s\n", 6089 wq_affn_names[wq->unbound_attrs->affn_scope]); 6090 mutex_unlock(&wq->mutex); 6091 6092 return written; 6093 } 6094 6095 static ssize_t wq_affn_scope_store(struct device *dev, 6096 struct device_attribute *attr, 6097 const char *buf, size_t count) 6098 { 6099 struct workqueue_struct *wq = dev_to_wq(dev); 6100 struct workqueue_attrs *attrs; 6101 int affn, ret = -ENOMEM; 6102 6103 affn = parse_affn_scope(buf); 6104 if (affn < 0) 6105 return affn; 6106 6107 apply_wqattrs_lock(); 6108 attrs = wq_sysfs_prep_attrs(wq); 6109 if (attrs) { 6110 attrs->affn_scope = affn; 6111 ret = apply_workqueue_attrs_locked(wq, attrs); 6112 } 6113 apply_wqattrs_unlock(); 6114 free_workqueue_attrs(attrs); 6115 return ret ?: count; 6116 } 6117 6118 static ssize_t wq_affinity_strict_show(struct device *dev, 6119 struct device_attribute *attr, char *buf) 6120 { 6121 struct workqueue_struct *wq = dev_to_wq(dev); 6122 6123 return scnprintf(buf, PAGE_SIZE, "%d\n", 6124 wq->unbound_attrs->affn_strict); 6125 } 6126 6127 static ssize_t wq_affinity_strict_store(struct device *dev, 6128 struct device_attribute *attr, 6129 const char *buf, size_t count) 6130 { 6131 struct workqueue_struct *wq = dev_to_wq(dev); 6132 struct workqueue_attrs *attrs; 6133 int v, ret = -ENOMEM; 6134 6135 if (sscanf(buf, "%d", &v) != 1) 6136 return -EINVAL; 6137 6138 apply_wqattrs_lock(); 6139 attrs = wq_sysfs_prep_attrs(wq); 6140 if (attrs) { 6141 attrs->affn_strict = (bool)v; 6142 ret = apply_workqueue_attrs_locked(wq, attrs); 6143 } 6144 apply_wqattrs_unlock(); 6145 free_workqueue_attrs(attrs); 6146 return ret ?: count; 6147 } 6148 6149 static struct device_attribute wq_sysfs_unbound_attrs[] = { 6150 __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 6151 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 6152 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 6153 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), 6154 __ATTR_NULL, 6155 }; 6156 6157 static struct bus_type wq_subsys = { 6158 .name = "workqueue", 6159 .dev_groups = wq_sysfs_groups, 6160 }; 6161 6162 static ssize_t wq_unbound_cpumask_show(struct device *dev, 6163 struct device_attribute *attr, char *buf) 6164 { 6165 int written; 6166 6167 mutex_lock(&wq_pool_mutex); 6168 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6169 cpumask_pr_args(wq_unbound_cpumask)); 6170 mutex_unlock(&wq_pool_mutex); 6171 6172 return written; 6173 } 6174 6175 static ssize_t wq_unbound_cpumask_store(struct device *dev, 6176 struct device_attribute *attr, const char *buf, size_t count) 6177 { 6178 cpumask_var_t cpumask; 6179 int ret; 6180 6181 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6182 return -ENOMEM; 6183 6184 ret = cpumask_parse(buf, cpumask); 6185 if (!ret) 6186 ret = workqueue_set_unbound_cpumask(cpumask); 6187 6188 free_cpumask_var(cpumask); 6189 return ret ? ret : count; 6190 } 6191 6192 static struct device_attribute wq_sysfs_cpumask_attr = 6193 __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6194 wq_unbound_cpumask_store); 6195 6196 static int __init wq_sysfs_init(void) 6197 { 6198 struct device *dev_root; 6199 int err; 6200 6201 err = subsys_virtual_register(&wq_subsys, NULL); 6202 if (err) 6203 return err; 6204 6205 dev_root = bus_get_dev_root(&wq_subsys); 6206 if (dev_root) { 6207 err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6208 put_device(dev_root); 6209 } 6210 return err; 6211 } 6212 core_initcall(wq_sysfs_init); 6213 6214 static void wq_device_release(struct device *dev) 6215 { 6216 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 6217 6218 kfree(wq_dev); 6219 } 6220 6221 /** 6222 * workqueue_sysfs_register - make a workqueue visible in sysfs 6223 * @wq: the workqueue to register 6224 * 6225 * Expose @wq in sysfs under /sys/bus/workqueue/devices. 6226 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 6227 * which is the preferred method. 6228 * 6229 * Workqueue user should use this function directly iff it wants to apply 6230 * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 6231 * apply_workqueue_attrs() may race against userland updating the 6232 * attributes. 6233 * 6234 * Return: 0 on success, -errno on failure. 6235 */ 6236 int workqueue_sysfs_register(struct workqueue_struct *wq) 6237 { 6238 struct wq_device *wq_dev; 6239 int ret; 6240 6241 /* 6242 * Adjusting max_active or creating new pwqs by applying 6243 * attributes breaks ordering guarantee. Disallow exposing ordered 6244 * workqueues. 6245 */ 6246 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 6247 return -EINVAL; 6248 6249 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 6250 if (!wq_dev) 6251 return -ENOMEM; 6252 6253 wq_dev->wq = wq; 6254 wq_dev->dev.bus = &wq_subsys; 6255 wq_dev->dev.release = wq_device_release; 6256 dev_set_name(&wq_dev->dev, "%s", wq->name); 6257 6258 /* 6259 * unbound_attrs are created separately. Suppress uevent until 6260 * everything is ready. 6261 */ 6262 dev_set_uevent_suppress(&wq_dev->dev, true); 6263 6264 ret = device_register(&wq_dev->dev); 6265 if (ret) { 6266 put_device(&wq_dev->dev); 6267 wq->wq_dev = NULL; 6268 return ret; 6269 } 6270 6271 if (wq->flags & WQ_UNBOUND) { 6272 struct device_attribute *attr; 6273 6274 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 6275 ret = device_create_file(&wq_dev->dev, attr); 6276 if (ret) { 6277 device_unregister(&wq_dev->dev); 6278 wq->wq_dev = NULL; 6279 return ret; 6280 } 6281 } 6282 } 6283 6284 dev_set_uevent_suppress(&wq_dev->dev, false); 6285 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 6286 return 0; 6287 } 6288 6289 /** 6290 * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 6291 * @wq: the workqueue to unregister 6292 * 6293 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 6294 */ 6295 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 6296 { 6297 struct wq_device *wq_dev = wq->wq_dev; 6298 6299 if (!wq->wq_dev) 6300 return; 6301 6302 wq->wq_dev = NULL; 6303 device_unregister(&wq_dev->dev); 6304 } 6305 #else /* CONFIG_SYSFS */ 6306 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 6307 #endif /* CONFIG_SYSFS */ 6308 6309 /* 6310 * Workqueue watchdog. 6311 * 6312 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 6313 * flush dependency, a concurrency managed work item which stays RUNNING 6314 * indefinitely. Workqueue stalls can be very difficult to debug as the 6315 * usual warning mechanisms don't trigger and internal workqueue state is 6316 * largely opaque. 6317 * 6318 * Workqueue watchdog monitors all worker pools periodically and dumps 6319 * state if some pools failed to make forward progress for a while where 6320 * forward progress is defined as the first item on ->worklist changing. 6321 * 6322 * This mechanism is controlled through the kernel parameter 6323 * "workqueue.watchdog_thresh" which can be updated at runtime through the 6324 * corresponding sysfs parameter file. 6325 */ 6326 #ifdef CONFIG_WQ_WATCHDOG 6327 6328 static unsigned long wq_watchdog_thresh = 30; 6329 static struct timer_list wq_watchdog_timer; 6330 6331 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 6332 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 6333 6334 /* 6335 * Show workers that might prevent the processing of pending work items. 6336 * The only candidates are CPU-bound workers in the running state. 6337 * Pending work items should be handled by another idle worker 6338 * in all other situations. 6339 */ 6340 static void show_cpu_pool_hog(struct worker_pool *pool) 6341 { 6342 struct worker *worker; 6343 unsigned long flags; 6344 int bkt; 6345 6346 raw_spin_lock_irqsave(&pool->lock, flags); 6347 6348 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6349 if (task_is_running(worker->task)) { 6350 /* 6351 * Defer printing to avoid deadlocks in console 6352 * drivers that queue work while holding locks 6353 * also taken in their write paths. 6354 */ 6355 printk_deferred_enter(); 6356 6357 pr_info("pool %d:\n", pool->id); 6358 sched_show_task(worker->task); 6359 6360 printk_deferred_exit(); 6361 } 6362 } 6363 6364 raw_spin_unlock_irqrestore(&pool->lock, flags); 6365 } 6366 6367 static void show_cpu_pools_hogs(void) 6368 { 6369 struct worker_pool *pool; 6370 int pi; 6371 6372 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6373 6374 rcu_read_lock(); 6375 6376 for_each_pool(pool, pi) { 6377 if (pool->cpu_stall) 6378 show_cpu_pool_hog(pool); 6379 6380 } 6381 6382 rcu_read_unlock(); 6383 } 6384 6385 static void wq_watchdog_reset_touched(void) 6386 { 6387 int cpu; 6388 6389 wq_watchdog_touched = jiffies; 6390 for_each_possible_cpu(cpu) 6391 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 6392 } 6393 6394 static void wq_watchdog_timer_fn(struct timer_list *unused) 6395 { 6396 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 6397 bool lockup_detected = false; 6398 bool cpu_pool_stall = false; 6399 unsigned long now = jiffies; 6400 struct worker_pool *pool; 6401 int pi; 6402 6403 if (!thresh) 6404 return; 6405 6406 rcu_read_lock(); 6407 6408 for_each_pool(pool, pi) { 6409 unsigned long pool_ts, touched, ts; 6410 6411 pool->cpu_stall = false; 6412 if (list_empty(&pool->worklist)) 6413 continue; 6414 6415 /* 6416 * If a virtual machine is stopped by the host it can look to 6417 * the watchdog like a stall. 6418 */ 6419 kvm_check_and_clear_guest_paused(); 6420 6421 /* get the latest of pool and touched timestamps */ 6422 if (pool->cpu >= 0) 6423 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 6424 else 6425 touched = READ_ONCE(wq_watchdog_touched); 6426 pool_ts = READ_ONCE(pool->watchdog_ts); 6427 6428 if (time_after(pool_ts, touched)) 6429 ts = pool_ts; 6430 else 6431 ts = touched; 6432 6433 /* did we stall? */ 6434 if (time_after(now, ts + thresh)) { 6435 lockup_detected = true; 6436 if (pool->cpu >= 0) { 6437 pool->cpu_stall = true; 6438 cpu_pool_stall = true; 6439 } 6440 pr_emerg("BUG: workqueue lockup - pool"); 6441 pr_cont_pool_info(pool); 6442 pr_cont(" stuck for %us!\n", 6443 jiffies_to_msecs(now - pool_ts) / 1000); 6444 } 6445 6446 6447 } 6448 6449 rcu_read_unlock(); 6450 6451 if (lockup_detected) 6452 show_all_workqueues(); 6453 6454 if (cpu_pool_stall) 6455 show_cpu_pools_hogs(); 6456 6457 wq_watchdog_reset_touched(); 6458 mod_timer(&wq_watchdog_timer, jiffies + thresh); 6459 } 6460 6461 notrace void wq_watchdog_touch(int cpu) 6462 { 6463 if (cpu >= 0) 6464 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 6465 6466 wq_watchdog_touched = jiffies; 6467 } 6468 6469 static void wq_watchdog_set_thresh(unsigned long thresh) 6470 { 6471 wq_watchdog_thresh = 0; 6472 del_timer_sync(&wq_watchdog_timer); 6473 6474 if (thresh) { 6475 wq_watchdog_thresh = thresh; 6476 wq_watchdog_reset_touched(); 6477 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 6478 } 6479 } 6480 6481 static int wq_watchdog_param_set_thresh(const char *val, 6482 const struct kernel_param *kp) 6483 { 6484 unsigned long thresh; 6485 int ret; 6486 6487 ret = kstrtoul(val, 0, &thresh); 6488 if (ret) 6489 return ret; 6490 6491 if (system_wq) 6492 wq_watchdog_set_thresh(thresh); 6493 else 6494 wq_watchdog_thresh = thresh; 6495 6496 return 0; 6497 } 6498 6499 static const struct kernel_param_ops wq_watchdog_thresh_ops = { 6500 .set = wq_watchdog_param_set_thresh, 6501 .get = param_get_ulong, 6502 }; 6503 6504 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 6505 0644); 6506 6507 static void wq_watchdog_init(void) 6508 { 6509 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 6510 wq_watchdog_set_thresh(wq_watchdog_thresh); 6511 } 6512 6513 #else /* CONFIG_WQ_WATCHDOG */ 6514 6515 static inline void wq_watchdog_init(void) { } 6516 6517 #endif /* CONFIG_WQ_WATCHDOG */ 6518 6519 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) 6520 { 6521 if (!cpumask_intersects(wq_unbound_cpumask, mask)) { 6522 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", 6523 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); 6524 return; 6525 } 6526 6527 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); 6528 } 6529 6530 /** 6531 * workqueue_init_early - early init for workqueue subsystem 6532 * 6533 * This is the first step of three-staged workqueue subsystem initialization and 6534 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 6535 * up. It sets up all the data structures and system workqueues and allows early 6536 * boot code to create workqueues and queue/cancel work items. Actual work item 6537 * execution starts only after kthreads can be created and scheduled right 6538 * before early initcalls. 6539 */ 6540 void __init workqueue_init_early(void) 6541 { 6542 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 6543 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 6544 int i, cpu; 6545 6546 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6547 6548 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 6549 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 6550 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); 6551 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); 6552 if (!cpumask_empty(&wq_cmdline_cpumask)) 6553 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); 6554 6555 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6556 6557 wq_update_pod_attrs_buf = alloc_workqueue_attrs(); 6558 BUG_ON(!wq_update_pod_attrs_buf); 6559 6560 /* initialize WQ_AFFN_SYSTEM pods */ 6561 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 6562 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); 6563 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 6564 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 6565 6566 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 6567 6568 pt->nr_pods = 1; 6569 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 6570 pt->pod_node[0] = NUMA_NO_NODE; 6571 pt->cpu_pod[0] = 0; 6572 6573 /* initialize CPU pools */ 6574 for_each_possible_cpu(cpu) { 6575 struct worker_pool *pool; 6576 6577 i = 0; 6578 for_each_cpu_worker_pool(pool, cpu) { 6579 BUG_ON(init_worker_pool(pool)); 6580 pool->cpu = cpu; 6581 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 6582 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 6583 pool->attrs->nice = std_nice[i++]; 6584 pool->attrs->affn_strict = true; 6585 pool->node = cpu_to_node(cpu); 6586 6587 /* alloc pool ID */ 6588 mutex_lock(&wq_pool_mutex); 6589 BUG_ON(worker_pool_assign_id(pool)); 6590 mutex_unlock(&wq_pool_mutex); 6591 } 6592 } 6593 6594 /* create default unbound and ordered wq attrs */ 6595 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 6596 struct workqueue_attrs *attrs; 6597 6598 BUG_ON(!(attrs = alloc_workqueue_attrs())); 6599 attrs->nice = std_nice[i]; 6600 unbound_std_wq_attrs[i] = attrs; 6601 6602 /* 6603 * An ordered wq should have only one pwq as ordering is 6604 * guaranteed by max_active which is enforced by pwqs. 6605 */ 6606 BUG_ON(!(attrs = alloc_workqueue_attrs())); 6607 attrs->nice = std_nice[i]; 6608 attrs->ordered = true; 6609 ordered_wq_attrs[i] = attrs; 6610 } 6611 6612 system_wq = alloc_workqueue("events", 0, 0); 6613 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6614 system_long_wq = alloc_workqueue("events_long", 0, 0); 6615 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6616 WQ_MAX_ACTIVE); 6617 system_freezable_wq = alloc_workqueue("events_freezable", 6618 WQ_FREEZABLE, 0); 6619 system_power_efficient_wq = alloc_workqueue("events_power_efficient", 6620 WQ_POWER_EFFICIENT, 0); 6621 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 6622 WQ_FREEZABLE | WQ_POWER_EFFICIENT, 6623 0); 6624 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 6625 !system_unbound_wq || !system_freezable_wq || 6626 !system_power_efficient_wq || 6627 !system_freezable_power_efficient_wq); 6628 } 6629 6630 static void __init wq_cpu_intensive_thresh_init(void) 6631 { 6632 unsigned long thresh; 6633 unsigned long bogo; 6634 6635 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); 6636 BUG_ON(IS_ERR(pwq_release_worker)); 6637 6638 /* if the user set it to a specific value, keep it */ 6639 if (wq_cpu_intensive_thresh_us != ULONG_MAX) 6640 return; 6641 6642 /* 6643 * The default of 10ms is derived from the fact that most modern (as of 6644 * 2023) processors can do a lot in 10ms and that it's just below what 6645 * most consider human-perceivable. However, the kernel also runs on a 6646 * lot slower CPUs including microcontrollers where the threshold is way 6647 * too low. 6648 * 6649 * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 6650 * This is by no means accurate but it doesn't have to be. The mechanism 6651 * is still useful even when the threshold is fully scaled up. Also, as 6652 * the reports would usually be applicable to everyone, some machines 6653 * operating on longer thresholds won't significantly diminish their 6654 * usefulness. 6655 */ 6656 thresh = 10 * USEC_PER_MSEC; 6657 6658 /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 6659 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 6660 if (bogo < 4000) 6661 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 6662 6663 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 6664 loops_per_jiffy, bogo, thresh); 6665 6666 wq_cpu_intensive_thresh_us = thresh; 6667 } 6668 6669 /** 6670 * workqueue_init - bring workqueue subsystem fully online 6671 * 6672 * This is the second step of three-staged workqueue subsystem initialization 6673 * and invoked as soon as kthreads can be created and scheduled. Workqueues have 6674 * been created and work items queued on them, but there are no kworkers 6675 * executing the work items yet. Populate the worker pools with the initial 6676 * workers and enable future kworker creations. 6677 */ 6678 void __init workqueue_init(void) 6679 { 6680 struct workqueue_struct *wq; 6681 struct worker_pool *pool; 6682 int cpu, bkt; 6683 6684 wq_cpu_intensive_thresh_init(); 6685 6686 mutex_lock(&wq_pool_mutex); 6687 6688 /* 6689 * Per-cpu pools created earlier could be missing node hint. Fix them 6690 * up. Also, create a rescuer for workqueues that requested it. 6691 */ 6692 for_each_possible_cpu(cpu) { 6693 for_each_cpu_worker_pool(pool, cpu) { 6694 pool->node = cpu_to_node(cpu); 6695 } 6696 } 6697 6698 list_for_each_entry(wq, &workqueues, list) { 6699 WARN(init_rescuer(wq), 6700 "workqueue: failed to create early rescuer for %s", 6701 wq->name); 6702 } 6703 6704 mutex_unlock(&wq_pool_mutex); 6705 6706 /* create the initial workers */ 6707 for_each_online_cpu(cpu) { 6708 for_each_cpu_worker_pool(pool, cpu) { 6709 pool->flags &= ~POOL_DISASSOCIATED; 6710 BUG_ON(!create_worker(pool)); 6711 } 6712 } 6713 6714 hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 6715 BUG_ON(!create_worker(pool)); 6716 6717 wq_online = true; 6718 wq_watchdog_init(); 6719 } 6720 6721 /* 6722 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 6723 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 6724 * and consecutive pod ID. The rest of @pt is initialized accordingly. 6725 */ 6726 static void __init init_pod_type(struct wq_pod_type *pt, 6727 bool (*cpus_share_pod)(int, int)) 6728 { 6729 int cur, pre, cpu, pod; 6730 6731 pt->nr_pods = 0; 6732 6733 /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 6734 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 6735 BUG_ON(!pt->cpu_pod); 6736 6737 for_each_possible_cpu(cur) { 6738 for_each_possible_cpu(pre) { 6739 if (pre >= cur) { 6740 pt->cpu_pod[cur] = pt->nr_pods++; 6741 break; 6742 } 6743 if (cpus_share_pod(cur, pre)) { 6744 pt->cpu_pod[cur] = pt->cpu_pod[pre]; 6745 break; 6746 } 6747 } 6748 } 6749 6750 /* init the rest to match @pt->cpu_pod[] */ 6751 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 6752 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); 6753 BUG_ON(!pt->pod_cpus || !pt->pod_node); 6754 6755 for (pod = 0; pod < pt->nr_pods; pod++) 6756 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 6757 6758 for_each_possible_cpu(cpu) { 6759 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 6760 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 6761 } 6762 } 6763 6764 static bool __init cpus_dont_share(int cpu0, int cpu1) 6765 { 6766 return false; 6767 } 6768 6769 static bool __init cpus_share_smt(int cpu0, int cpu1) 6770 { 6771 #ifdef CONFIG_SCHED_SMT 6772 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 6773 #else 6774 return false; 6775 #endif 6776 } 6777 6778 static bool __init cpus_share_numa(int cpu0, int cpu1) 6779 { 6780 return cpu_to_node(cpu0) == cpu_to_node(cpu1); 6781 } 6782 6783 /** 6784 * workqueue_init_topology - initialize CPU pods for unbound workqueues 6785 * 6786 * This is the third step of there-staged workqueue subsystem initialization and 6787 * invoked after SMP and topology information are fully initialized. It 6788 * initializes the unbound CPU pods accordingly. 6789 */ 6790 void __init workqueue_init_topology(void) 6791 { 6792 struct workqueue_struct *wq; 6793 int cpu; 6794 6795 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 6796 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 6797 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 6798 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 6799 6800 mutex_lock(&wq_pool_mutex); 6801 6802 /* 6803 * Workqueues allocated earlier would have all CPUs sharing the default 6804 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU 6805 * combinations to apply per-pod sharing. 6806 */ 6807 list_for_each_entry(wq, &workqueues, list) { 6808 for_each_online_cpu(cpu) { 6809 wq_update_pod(wq, cpu, cpu, true); 6810 } 6811 } 6812 6813 mutex_unlock(&wq_pool_mutex); 6814 } 6815 6816 void __warn_flushing_systemwide_wq(void) 6817 { 6818 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 6819 dump_stack(); 6820 } 6821 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6822 6823 static int __init workqueue_unbound_cpus_setup(char *str) 6824 { 6825 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 6826 cpumask_clear(&wq_cmdline_cpumask); 6827 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 6828 } 6829 6830 return 1; 6831 } 6832 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 6833