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_work(struct pool_workqueue *pwq, 1459 struct work_struct *work) 1460 { 1461 unsigned long *wdb = work_data_bits(work); 1462 1463 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); 1464 trace_workqueue_activate_work(work); 1465 if (list_empty(&pwq->pool->worklist)) 1466 pwq->pool->watchdog_ts = jiffies; 1467 move_linked_works(work, &pwq->pool->worklist, NULL); 1468 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); 1469 } 1470 1471 /** 1472 * pwq_activate_work - Activate a work item if inactive 1473 * @pwq: pool_workqueue @work belongs to 1474 * @work: work item to activate 1475 * 1476 * Returns %true if activated. %false if already active. 1477 */ 1478 static bool pwq_activate_work(struct pool_workqueue *pwq, 1479 struct work_struct *work) 1480 { 1481 struct worker_pool *pool = pwq->pool; 1482 1483 lockdep_assert_held(&pool->lock); 1484 1485 if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE)) 1486 return false; 1487 1488 pwq->nr_active++; 1489 __pwq_activate_work(pwq, work); 1490 return true; 1491 } 1492 1493 /** 1494 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq 1495 * @pwq: pool_workqueue of interest 1496 * 1497 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is 1498 * successfully obtained. %false otherwise. 1499 */ 1500 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq) 1501 { 1502 struct workqueue_struct *wq = pwq->wq; 1503 struct worker_pool *pool = pwq->pool; 1504 bool obtained; 1505 1506 lockdep_assert_held(&pool->lock); 1507 1508 obtained = pwq->nr_active < READ_ONCE(wq->max_active); 1509 1510 if (obtained) 1511 pwq->nr_active++; 1512 return obtained; 1513 } 1514 1515 /** 1516 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq 1517 * @pwq: pool_workqueue of interest 1518 * 1519 * Activate the first inactive work item of @pwq if available and allowed by 1520 * max_active limit. 1521 * 1522 * Returns %true if an inactive work item has been activated. %false if no 1523 * inactive work item is found or max_active limit is reached. 1524 */ 1525 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq) 1526 { 1527 struct work_struct *work = 1528 list_first_entry_or_null(&pwq->inactive_works, 1529 struct work_struct, entry); 1530 1531 if (work && pwq_tryinc_nr_active(pwq)) { 1532 __pwq_activate_work(pwq, work); 1533 return true; 1534 } else { 1535 return false; 1536 } 1537 } 1538 1539 /** 1540 * pwq_dec_nr_active - Retire an active count 1541 * @pwq: pool_workqueue of interest 1542 * 1543 * Decrement @pwq's nr_active and try to activate the first inactive work item. 1544 */ 1545 static void pwq_dec_nr_active(struct pool_workqueue *pwq) 1546 { 1547 struct worker_pool *pool = pwq->pool; 1548 1549 lockdep_assert_held(&pool->lock); 1550 1551 pwq->nr_active--; 1552 pwq_activate_first_inactive(pwq); 1553 } 1554 1555 /** 1556 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1557 * @pwq: pwq of interest 1558 * @work_data: work_data of work which left the queue 1559 * 1560 * A work either has completed or is removed from pending queue, 1561 * decrement nr_in_flight of its pwq and handle workqueue flushing. 1562 * 1563 * CONTEXT: 1564 * raw_spin_lock_irq(pool->lock). 1565 */ 1566 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1567 { 1568 int color = get_work_color(work_data); 1569 1570 if (!(work_data & WORK_STRUCT_INACTIVE)) 1571 pwq_dec_nr_active(pwq); 1572 1573 pwq->nr_in_flight[color]--; 1574 1575 /* is flush in progress and are we at the flushing tip? */ 1576 if (likely(pwq->flush_color != color)) 1577 goto out_put; 1578 1579 /* are there still in-flight works? */ 1580 if (pwq->nr_in_flight[color]) 1581 goto out_put; 1582 1583 /* this pwq is done, clear flush_color */ 1584 pwq->flush_color = -1; 1585 1586 /* 1587 * If this was the last pwq, wake up the first flusher. It 1588 * will handle the rest. 1589 */ 1590 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1591 complete(&pwq->wq->first_flusher->done); 1592 out_put: 1593 put_pwq(pwq); 1594 } 1595 1596 /** 1597 * try_to_grab_pending - steal work item from worklist and disable irq 1598 * @work: work item to steal 1599 * @is_dwork: @work is a delayed_work 1600 * @flags: place to store irq state 1601 * 1602 * Try to grab PENDING bit of @work. This function can handle @work in any 1603 * stable state - idle, on timer or on worklist. 1604 * 1605 * Return: 1606 * 1607 * ======== ================================================================ 1608 * 1 if @work was pending and we successfully stole PENDING 1609 * 0 if @work was idle and we claimed PENDING 1610 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1611 * -ENOENT if someone else is canceling @work, this state may persist 1612 * for arbitrarily long 1613 * ======== ================================================================ 1614 * 1615 * Note: 1616 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1617 * interrupted while holding PENDING and @work off queue, irq must be 1618 * disabled on entry. This, combined with delayed_work->timer being 1619 * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1620 * 1621 * On successful return, >= 0, irq is disabled and the caller is 1622 * responsible for releasing it using local_irq_restore(*@flags). 1623 * 1624 * This function is safe to call from any context including IRQ handler. 1625 */ 1626 static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1627 unsigned long *flags) 1628 { 1629 struct worker_pool *pool; 1630 struct pool_workqueue *pwq; 1631 1632 local_irq_save(*flags); 1633 1634 /* try to steal the timer if it exists */ 1635 if (is_dwork) { 1636 struct delayed_work *dwork = to_delayed_work(work); 1637 1638 /* 1639 * dwork->timer is irqsafe. If del_timer() fails, it's 1640 * guaranteed that the timer is not queued anywhere and not 1641 * running on the local CPU. 1642 */ 1643 if (likely(del_timer(&dwork->timer))) 1644 return 1; 1645 } 1646 1647 /* try to claim PENDING the normal way */ 1648 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1649 return 0; 1650 1651 rcu_read_lock(); 1652 /* 1653 * The queueing is in progress, or it is already queued. Try to 1654 * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1655 */ 1656 pool = get_work_pool(work); 1657 if (!pool) 1658 goto fail; 1659 1660 raw_spin_lock(&pool->lock); 1661 /* 1662 * work->data is guaranteed to point to pwq only while the work 1663 * item is queued on pwq->wq, and both updating work->data to point 1664 * to pwq on queueing and to pool on dequeueing are done under 1665 * pwq->pool->lock. This in turn guarantees that, if work->data 1666 * points to pwq which is associated with a locked pool, the work 1667 * item is currently queued on that pool. 1668 */ 1669 pwq = get_work_pwq(work); 1670 if (pwq && pwq->pool == pool) { 1671 debug_work_deactivate(work); 1672 1673 /* 1674 * A cancelable inactive work item must be in the 1675 * pwq->inactive_works since a queued barrier can't be 1676 * canceled (see the comments in insert_wq_barrier()). 1677 * 1678 * An inactive work item cannot be grabbed directly because 1679 * it might have linked barrier work items which, if left 1680 * on the inactive_works list, will confuse pwq->nr_active 1681 * management later on and cause stall. Make sure the work 1682 * item is activated before grabbing. 1683 */ 1684 pwq_activate_work(pwq, work); 1685 1686 list_del_init(&work->entry); 1687 pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 1688 1689 /* work->data points to pwq iff queued, point to pool */ 1690 set_work_pool_and_keep_pending(work, pool->id); 1691 1692 raw_spin_unlock(&pool->lock); 1693 rcu_read_unlock(); 1694 return 1; 1695 } 1696 raw_spin_unlock(&pool->lock); 1697 fail: 1698 rcu_read_unlock(); 1699 local_irq_restore(*flags); 1700 if (work_is_canceling(work)) 1701 return -ENOENT; 1702 cpu_relax(); 1703 return -EAGAIN; 1704 } 1705 1706 /** 1707 * insert_work - insert a work into a pool 1708 * @pwq: pwq @work belongs to 1709 * @work: work to insert 1710 * @head: insertion point 1711 * @extra_flags: extra WORK_STRUCT_* flags to set 1712 * 1713 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1714 * work_struct flags. 1715 * 1716 * CONTEXT: 1717 * raw_spin_lock_irq(pool->lock). 1718 */ 1719 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1720 struct list_head *head, unsigned int extra_flags) 1721 { 1722 debug_work_activate(work); 1723 1724 /* record the work call stack in order to print it in KASAN reports */ 1725 kasan_record_aux_stack_noalloc(work); 1726 1727 /* we own @work, set data and link */ 1728 set_work_pwq(work, pwq, extra_flags); 1729 list_add_tail(&work->entry, head); 1730 get_pwq(pwq); 1731 } 1732 1733 /* 1734 * Test whether @work is being queued from another work executing on the 1735 * same workqueue. 1736 */ 1737 static bool is_chained_work(struct workqueue_struct *wq) 1738 { 1739 struct worker *worker; 1740 1741 worker = current_wq_worker(); 1742 /* 1743 * Return %true iff I'm a worker executing a work item on @wq. If 1744 * I'm @worker, it's safe to dereference it without locking. 1745 */ 1746 return worker && worker->current_pwq->wq == wq; 1747 } 1748 1749 /* 1750 * When queueing an unbound work item to a wq, prefer local CPU if allowed 1751 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1752 * avoid perturbing sensitive tasks. 1753 */ 1754 static int wq_select_unbound_cpu(int cpu) 1755 { 1756 int new_cpu; 1757 1758 if (likely(!wq_debug_force_rr_cpu)) { 1759 if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1760 return cpu; 1761 } else { 1762 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1763 } 1764 1765 new_cpu = __this_cpu_read(wq_rr_cpu_last); 1766 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1767 if (unlikely(new_cpu >= nr_cpu_ids)) { 1768 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1769 if (unlikely(new_cpu >= nr_cpu_ids)) 1770 return cpu; 1771 } 1772 __this_cpu_write(wq_rr_cpu_last, new_cpu); 1773 1774 return new_cpu; 1775 } 1776 1777 static void __queue_work(int cpu, struct workqueue_struct *wq, 1778 struct work_struct *work) 1779 { 1780 struct pool_workqueue *pwq; 1781 struct worker_pool *last_pool, *pool; 1782 unsigned int work_flags; 1783 unsigned int req_cpu = cpu; 1784 1785 /* 1786 * While a work item is PENDING && off queue, a task trying to 1787 * steal the PENDING will busy-loop waiting for it to either get 1788 * queued or lose PENDING. Grabbing PENDING and queueing should 1789 * happen with IRQ disabled. 1790 */ 1791 lockdep_assert_irqs_disabled(); 1792 1793 1794 /* 1795 * For a draining wq, only works from the same workqueue are 1796 * allowed. The __WQ_DESTROYING helps to spot the issue that 1797 * queues a new work item to a wq after destroy_workqueue(wq). 1798 */ 1799 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 1800 WARN_ON_ONCE(!is_chained_work(wq)))) 1801 return; 1802 rcu_read_lock(); 1803 retry: 1804 /* pwq which will be used unless @work is executing elsewhere */ 1805 if (req_cpu == WORK_CPU_UNBOUND) { 1806 if (wq->flags & WQ_UNBOUND) 1807 cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1808 else 1809 cpu = raw_smp_processor_id(); 1810 } 1811 1812 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 1813 pool = pwq->pool; 1814 1815 /* 1816 * If @work was previously on a different pool, it might still be 1817 * running there, in which case the work needs to be queued on that 1818 * pool to guarantee non-reentrancy. 1819 */ 1820 last_pool = get_work_pool(work); 1821 if (last_pool && last_pool != pool) { 1822 struct worker *worker; 1823 1824 raw_spin_lock(&last_pool->lock); 1825 1826 worker = find_worker_executing_work(last_pool, work); 1827 1828 if (worker && worker->current_pwq->wq == wq) { 1829 pwq = worker->current_pwq; 1830 pool = pwq->pool; 1831 WARN_ON_ONCE(pool != last_pool); 1832 } else { 1833 /* meh... not running there, queue here */ 1834 raw_spin_unlock(&last_pool->lock); 1835 raw_spin_lock(&pool->lock); 1836 } 1837 } else { 1838 raw_spin_lock(&pool->lock); 1839 } 1840 1841 /* 1842 * pwq is determined and locked. For unbound pools, we could have raced 1843 * with pwq release and it could already be dead. If its refcnt is zero, 1844 * repeat pwq selection. Note that unbound pwqs never die without 1845 * another pwq replacing it in cpu_pwq or while work items are executing 1846 * on it, so the retrying is guaranteed to make forward-progress. 1847 */ 1848 if (unlikely(!pwq->refcnt)) { 1849 if (wq->flags & WQ_UNBOUND) { 1850 raw_spin_unlock(&pool->lock); 1851 cpu_relax(); 1852 goto retry; 1853 } 1854 /* oops */ 1855 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 1856 wq->name, cpu); 1857 } 1858 1859 /* pwq determined, queue */ 1860 trace_workqueue_queue_work(req_cpu, pwq, work); 1861 1862 if (WARN_ON(!list_empty(&work->entry))) 1863 goto out; 1864 1865 pwq->nr_in_flight[pwq->work_color]++; 1866 work_flags = work_color_to_flags(pwq->work_color); 1867 1868 /* 1869 * Limit the number of concurrently active work items to max_active. 1870 * @work must also queue behind existing inactive work items to maintain 1871 * ordering when max_active changes. See wq_adjust_max_active(). 1872 */ 1873 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq)) { 1874 if (list_empty(&pool->worklist)) 1875 pool->watchdog_ts = jiffies; 1876 1877 trace_workqueue_activate_work(work); 1878 insert_work(pwq, work, &pool->worklist, work_flags); 1879 kick_pool(pool); 1880 } else { 1881 work_flags |= WORK_STRUCT_INACTIVE; 1882 insert_work(pwq, work, &pwq->inactive_works, work_flags); 1883 } 1884 1885 out: 1886 raw_spin_unlock(&pool->lock); 1887 rcu_read_unlock(); 1888 } 1889 1890 /** 1891 * queue_work_on - queue work on specific cpu 1892 * @cpu: CPU number to execute work on 1893 * @wq: workqueue to use 1894 * @work: work to queue 1895 * 1896 * We queue the work to a specific CPU, the caller must ensure it 1897 * can't go away. Callers that fail to ensure that the specified 1898 * CPU cannot go away will execute on a randomly chosen CPU. 1899 * But note well that callers specifying a CPU that never has been 1900 * online will get a splat. 1901 * 1902 * Return: %false if @work was already on a queue, %true otherwise. 1903 */ 1904 bool queue_work_on(int cpu, struct workqueue_struct *wq, 1905 struct work_struct *work) 1906 { 1907 bool ret = false; 1908 unsigned long flags; 1909 1910 local_irq_save(flags); 1911 1912 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1913 __queue_work(cpu, wq, work); 1914 ret = true; 1915 } 1916 1917 local_irq_restore(flags); 1918 return ret; 1919 } 1920 EXPORT_SYMBOL(queue_work_on); 1921 1922 /** 1923 * select_numa_node_cpu - Select a CPU based on NUMA node 1924 * @node: NUMA node ID that we want to select a CPU from 1925 * 1926 * This function will attempt to find a "random" cpu available on a given 1927 * node. If there are no CPUs available on the given node it will return 1928 * WORK_CPU_UNBOUND indicating that we should just schedule to any 1929 * available CPU if we need to schedule this work. 1930 */ 1931 static int select_numa_node_cpu(int node) 1932 { 1933 int cpu; 1934 1935 /* Delay binding to CPU if node is not valid or online */ 1936 if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 1937 return WORK_CPU_UNBOUND; 1938 1939 /* Use local node/cpu if we are already there */ 1940 cpu = raw_smp_processor_id(); 1941 if (node == cpu_to_node(cpu)) 1942 return cpu; 1943 1944 /* Use "random" otherwise know as "first" online CPU of node */ 1945 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 1946 1947 /* If CPU is valid return that, otherwise just defer */ 1948 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 1949 } 1950 1951 /** 1952 * queue_work_node - queue work on a "random" cpu for a given NUMA node 1953 * @node: NUMA node that we are targeting the work for 1954 * @wq: workqueue to use 1955 * @work: work to queue 1956 * 1957 * We queue the work to a "random" CPU within a given NUMA node. The basic 1958 * idea here is to provide a way to somehow associate work with a given 1959 * NUMA node. 1960 * 1961 * This function will only make a best effort attempt at getting this onto 1962 * the right NUMA node. If no node is requested or the requested node is 1963 * offline then we just fall back to standard queue_work behavior. 1964 * 1965 * Currently the "random" CPU ends up being the first available CPU in the 1966 * intersection of cpu_online_mask and the cpumask of the node, unless we 1967 * are running on the node. In that case we just use the current CPU. 1968 * 1969 * Return: %false if @work was already on a queue, %true otherwise. 1970 */ 1971 bool queue_work_node(int node, struct workqueue_struct *wq, 1972 struct work_struct *work) 1973 { 1974 unsigned long flags; 1975 bool ret = false; 1976 1977 /* 1978 * This current implementation is specific to unbound workqueues. 1979 * Specifically we only return the first available CPU for a given 1980 * node instead of cycling through individual CPUs within the node. 1981 * 1982 * If this is used with a per-cpu workqueue then the logic in 1983 * workqueue_select_cpu_near would need to be updated to allow for 1984 * some round robin type logic. 1985 */ 1986 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 1987 1988 local_irq_save(flags); 1989 1990 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1991 int cpu = select_numa_node_cpu(node); 1992 1993 __queue_work(cpu, wq, work); 1994 ret = true; 1995 } 1996 1997 local_irq_restore(flags); 1998 return ret; 1999 } 2000 EXPORT_SYMBOL_GPL(queue_work_node); 2001 2002 void delayed_work_timer_fn(struct timer_list *t) 2003 { 2004 struct delayed_work *dwork = from_timer(dwork, t, timer); 2005 2006 /* should have been called from irqsafe timer with irq already off */ 2007 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 2008 } 2009 EXPORT_SYMBOL(delayed_work_timer_fn); 2010 2011 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 2012 struct delayed_work *dwork, unsigned long delay) 2013 { 2014 struct timer_list *timer = &dwork->timer; 2015 struct work_struct *work = &dwork->work; 2016 2017 WARN_ON_ONCE(!wq); 2018 WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 2019 WARN_ON_ONCE(timer_pending(timer)); 2020 WARN_ON_ONCE(!list_empty(&work->entry)); 2021 2022 /* 2023 * If @delay is 0, queue @dwork->work immediately. This is for 2024 * both optimization and correctness. The earliest @timer can 2025 * expire is on the closest next tick and delayed_work users depend 2026 * on that there's no such delay when @delay is 0. 2027 */ 2028 if (!delay) { 2029 __queue_work(cpu, wq, &dwork->work); 2030 return; 2031 } 2032 2033 dwork->wq = wq; 2034 dwork->cpu = cpu; 2035 timer->expires = jiffies + delay; 2036 2037 if (unlikely(cpu != WORK_CPU_UNBOUND)) 2038 add_timer_on(timer, cpu); 2039 else 2040 add_timer(timer); 2041 } 2042 2043 /** 2044 * queue_delayed_work_on - queue work on specific CPU after delay 2045 * @cpu: CPU number to execute work on 2046 * @wq: workqueue to use 2047 * @dwork: work to queue 2048 * @delay: number of jiffies to wait before queueing 2049 * 2050 * Return: %false if @work was already on a queue, %true otherwise. If 2051 * @delay is zero and @dwork is idle, it will be scheduled for immediate 2052 * execution. 2053 */ 2054 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 2055 struct delayed_work *dwork, unsigned long delay) 2056 { 2057 struct work_struct *work = &dwork->work; 2058 bool ret = false; 2059 unsigned long flags; 2060 2061 /* read the comment in __queue_work() */ 2062 local_irq_save(flags); 2063 2064 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 2065 __queue_delayed_work(cpu, wq, dwork, delay); 2066 ret = true; 2067 } 2068 2069 local_irq_restore(flags); 2070 return ret; 2071 } 2072 EXPORT_SYMBOL(queue_delayed_work_on); 2073 2074 /** 2075 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 2076 * @cpu: CPU number to execute work on 2077 * @wq: workqueue to use 2078 * @dwork: work to queue 2079 * @delay: number of jiffies to wait before queueing 2080 * 2081 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 2082 * modify @dwork's timer so that it expires after @delay. If @delay is 2083 * zero, @work is guaranteed to be scheduled immediately regardless of its 2084 * current state. 2085 * 2086 * Return: %false if @dwork was idle and queued, %true if @dwork was 2087 * pending and its timer was modified. 2088 * 2089 * This function is safe to call from any context including IRQ handler. 2090 * See try_to_grab_pending() for details. 2091 */ 2092 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 2093 struct delayed_work *dwork, unsigned long delay) 2094 { 2095 unsigned long flags; 2096 int ret; 2097 2098 do { 2099 ret = try_to_grab_pending(&dwork->work, true, &flags); 2100 } while (unlikely(ret == -EAGAIN)); 2101 2102 if (likely(ret >= 0)) { 2103 __queue_delayed_work(cpu, wq, dwork, delay); 2104 local_irq_restore(flags); 2105 } 2106 2107 /* -ENOENT from try_to_grab_pending() becomes %true */ 2108 return ret; 2109 } 2110 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 2111 2112 static void rcu_work_rcufn(struct rcu_head *rcu) 2113 { 2114 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 2115 2116 /* read the comment in __queue_work() */ 2117 local_irq_disable(); 2118 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 2119 local_irq_enable(); 2120 } 2121 2122 /** 2123 * queue_rcu_work - queue work after a RCU grace period 2124 * @wq: workqueue to use 2125 * @rwork: work to queue 2126 * 2127 * Return: %false if @rwork was already pending, %true otherwise. Note 2128 * that a full RCU grace period is guaranteed only after a %true return. 2129 * While @rwork is guaranteed to be executed after a %false return, the 2130 * execution may happen before a full RCU grace period has passed. 2131 */ 2132 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 2133 { 2134 struct work_struct *work = &rwork->work; 2135 2136 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 2137 rwork->wq = wq; 2138 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 2139 return true; 2140 } 2141 2142 return false; 2143 } 2144 EXPORT_SYMBOL(queue_rcu_work); 2145 2146 static struct worker *alloc_worker(int node) 2147 { 2148 struct worker *worker; 2149 2150 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2151 if (worker) { 2152 INIT_LIST_HEAD(&worker->entry); 2153 INIT_LIST_HEAD(&worker->scheduled); 2154 INIT_LIST_HEAD(&worker->node); 2155 /* on creation a worker is in !idle && prep state */ 2156 worker->flags = WORKER_PREP; 2157 } 2158 return worker; 2159 } 2160 2161 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 2162 { 2163 if (pool->cpu < 0 && pool->attrs->affn_strict) 2164 return pool->attrs->__pod_cpumask; 2165 else 2166 return pool->attrs->cpumask; 2167 } 2168 2169 /** 2170 * worker_attach_to_pool() - attach a worker to a pool 2171 * @worker: worker to be attached 2172 * @pool: the target pool 2173 * 2174 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 2175 * cpu-binding of @worker are kept coordinated with the pool across 2176 * cpu-[un]hotplugs. 2177 */ 2178 static void worker_attach_to_pool(struct worker *worker, 2179 struct worker_pool *pool) 2180 { 2181 mutex_lock(&wq_pool_attach_mutex); 2182 2183 /* 2184 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 2185 * stable across this function. See the comments above the flag 2186 * definition for details. 2187 */ 2188 if (pool->flags & POOL_DISASSOCIATED) 2189 worker->flags |= WORKER_UNBOUND; 2190 else 2191 kthread_set_per_cpu(worker->task, pool->cpu); 2192 2193 if (worker->rescue_wq) 2194 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2195 2196 list_add_tail(&worker->node, &pool->workers); 2197 worker->pool = pool; 2198 2199 mutex_unlock(&wq_pool_attach_mutex); 2200 } 2201 2202 /** 2203 * worker_detach_from_pool() - detach a worker from its pool 2204 * @worker: worker which is attached to its pool 2205 * 2206 * Undo the attaching which had been done in worker_attach_to_pool(). The 2207 * caller worker shouldn't access to the pool after detached except it has 2208 * other reference to the pool. 2209 */ 2210 static void worker_detach_from_pool(struct worker *worker) 2211 { 2212 struct worker_pool *pool = worker->pool; 2213 struct completion *detach_completion = NULL; 2214 2215 mutex_lock(&wq_pool_attach_mutex); 2216 2217 kthread_set_per_cpu(worker->task, -1); 2218 list_del(&worker->node); 2219 worker->pool = NULL; 2220 2221 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 2222 detach_completion = pool->detach_completion; 2223 mutex_unlock(&wq_pool_attach_mutex); 2224 2225 /* clear leftover flags without pool->lock after it is detached */ 2226 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2227 2228 if (detach_completion) 2229 complete(detach_completion); 2230 } 2231 2232 /** 2233 * create_worker - create a new workqueue worker 2234 * @pool: pool the new worker will belong to 2235 * 2236 * Create and start a new worker which is attached to @pool. 2237 * 2238 * CONTEXT: 2239 * Might sleep. Does GFP_KERNEL allocations. 2240 * 2241 * Return: 2242 * Pointer to the newly created worker. 2243 */ 2244 static struct worker *create_worker(struct worker_pool *pool) 2245 { 2246 struct worker *worker; 2247 int id; 2248 char id_buf[23]; 2249 2250 /* ID is needed to determine kthread name */ 2251 id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 2252 if (id < 0) { 2253 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 2254 ERR_PTR(id)); 2255 return NULL; 2256 } 2257 2258 worker = alloc_worker(pool->node); 2259 if (!worker) { 2260 pr_err_once("workqueue: Failed to allocate a worker\n"); 2261 goto fail; 2262 } 2263 2264 worker->id = id; 2265 2266 if (pool->cpu >= 0) 2267 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2268 pool->attrs->nice < 0 ? "H" : ""); 2269 else 2270 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2271 2272 worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2273 "kworker/%s", id_buf); 2274 if (IS_ERR(worker->task)) { 2275 if (PTR_ERR(worker->task) == -EINTR) { 2276 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 2277 id_buf); 2278 } else { 2279 pr_err_once("workqueue: Failed to create a worker thread: %pe", 2280 worker->task); 2281 } 2282 goto fail; 2283 } 2284 2285 set_user_nice(worker->task, pool->attrs->nice); 2286 kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 2287 2288 /* successful, attach the worker to the pool */ 2289 worker_attach_to_pool(worker, pool); 2290 2291 /* start the newly created worker */ 2292 raw_spin_lock_irq(&pool->lock); 2293 2294 worker->pool->nr_workers++; 2295 worker_enter_idle(worker); 2296 kick_pool(pool); 2297 2298 /* 2299 * @worker is waiting on a completion in kthread() and will trigger hung 2300 * check if not woken up soon. As kick_pool() might not have waken it 2301 * up, wake it up explicitly once more. 2302 */ 2303 wake_up_process(worker->task); 2304 2305 raw_spin_unlock_irq(&pool->lock); 2306 2307 return worker; 2308 2309 fail: 2310 ida_free(&pool->worker_ida, id); 2311 kfree(worker); 2312 return NULL; 2313 } 2314 2315 static void unbind_worker(struct worker *worker) 2316 { 2317 lockdep_assert_held(&wq_pool_attach_mutex); 2318 2319 kthread_set_per_cpu(worker->task, -1); 2320 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2321 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2322 else 2323 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2324 } 2325 2326 static void wake_dying_workers(struct list_head *cull_list) 2327 { 2328 struct worker *worker, *tmp; 2329 2330 list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2331 list_del_init(&worker->entry); 2332 unbind_worker(worker); 2333 /* 2334 * If the worker was somehow already running, then it had to be 2335 * in pool->idle_list when set_worker_dying() happened or we 2336 * wouldn't have gotten here. 2337 * 2338 * Thus, the worker must either have observed the WORKER_DIE 2339 * flag, or have set its state to TASK_IDLE. Either way, the 2340 * below will be observed by the worker and is safe to do 2341 * outside of pool->lock. 2342 */ 2343 wake_up_process(worker->task); 2344 } 2345 } 2346 2347 /** 2348 * set_worker_dying - Tag a worker for destruction 2349 * @worker: worker to be destroyed 2350 * @list: transfer worker away from its pool->idle_list and into list 2351 * 2352 * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2353 * should be idle. 2354 * 2355 * CONTEXT: 2356 * raw_spin_lock_irq(pool->lock). 2357 */ 2358 static void set_worker_dying(struct worker *worker, struct list_head *list) 2359 { 2360 struct worker_pool *pool = worker->pool; 2361 2362 lockdep_assert_held(&pool->lock); 2363 lockdep_assert_held(&wq_pool_attach_mutex); 2364 2365 /* sanity check frenzy */ 2366 if (WARN_ON(worker->current_work) || 2367 WARN_ON(!list_empty(&worker->scheduled)) || 2368 WARN_ON(!(worker->flags & WORKER_IDLE))) 2369 return; 2370 2371 pool->nr_workers--; 2372 pool->nr_idle--; 2373 2374 worker->flags |= WORKER_DIE; 2375 2376 list_move(&worker->entry, list); 2377 list_move(&worker->node, &pool->dying_workers); 2378 } 2379 2380 /** 2381 * idle_worker_timeout - check if some idle workers can now be deleted. 2382 * @t: The pool's idle_timer that just expired 2383 * 2384 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 2385 * worker_leave_idle(), as a worker flicking between idle and active while its 2386 * pool is at the too_many_workers() tipping point would cause too much timer 2387 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 2388 * it expire and re-evaluate things from there. 2389 */ 2390 static void idle_worker_timeout(struct timer_list *t) 2391 { 2392 struct worker_pool *pool = from_timer(pool, t, idle_timer); 2393 bool do_cull = false; 2394 2395 if (work_pending(&pool->idle_cull_work)) 2396 return; 2397 2398 raw_spin_lock_irq(&pool->lock); 2399 2400 if (too_many_workers(pool)) { 2401 struct worker *worker; 2402 unsigned long expires; 2403 2404 /* idle_list is kept in LIFO order, check the last one */ 2405 worker = list_entry(pool->idle_list.prev, struct worker, entry); 2406 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2407 do_cull = !time_before(jiffies, expires); 2408 2409 if (!do_cull) 2410 mod_timer(&pool->idle_timer, expires); 2411 } 2412 raw_spin_unlock_irq(&pool->lock); 2413 2414 if (do_cull) 2415 queue_work(system_unbound_wq, &pool->idle_cull_work); 2416 } 2417 2418 /** 2419 * idle_cull_fn - cull workers that have been idle for too long. 2420 * @work: the pool's work for handling these idle workers 2421 * 2422 * This goes through a pool's idle workers and gets rid of those that have been 2423 * idle for at least IDLE_WORKER_TIMEOUT seconds. 2424 * 2425 * We don't want to disturb isolated CPUs because of a pcpu kworker being 2426 * culled, so this also resets worker affinity. This requires a sleepable 2427 * context, hence the split between timer callback and work item. 2428 */ 2429 static void idle_cull_fn(struct work_struct *work) 2430 { 2431 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2432 LIST_HEAD(cull_list); 2433 2434 /* 2435 * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2436 * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2437 * path. This is required as a previously-preempted worker could run after 2438 * set_worker_dying() has happened but before wake_dying_workers() did. 2439 */ 2440 mutex_lock(&wq_pool_attach_mutex); 2441 raw_spin_lock_irq(&pool->lock); 2442 2443 while (too_many_workers(pool)) { 2444 struct worker *worker; 2445 unsigned long expires; 2446 2447 worker = list_entry(pool->idle_list.prev, struct worker, entry); 2448 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2449 2450 if (time_before(jiffies, expires)) { 2451 mod_timer(&pool->idle_timer, expires); 2452 break; 2453 } 2454 2455 set_worker_dying(worker, &cull_list); 2456 } 2457 2458 raw_spin_unlock_irq(&pool->lock); 2459 wake_dying_workers(&cull_list); 2460 mutex_unlock(&wq_pool_attach_mutex); 2461 } 2462 2463 static void send_mayday(struct work_struct *work) 2464 { 2465 struct pool_workqueue *pwq = get_work_pwq(work); 2466 struct workqueue_struct *wq = pwq->wq; 2467 2468 lockdep_assert_held(&wq_mayday_lock); 2469 2470 if (!wq->rescuer) 2471 return; 2472 2473 /* mayday mayday mayday */ 2474 if (list_empty(&pwq->mayday_node)) { 2475 /* 2476 * If @pwq is for an unbound wq, its base ref may be put at 2477 * any time due to an attribute change. Pin @pwq until the 2478 * rescuer is done with it. 2479 */ 2480 get_pwq(pwq); 2481 list_add_tail(&pwq->mayday_node, &wq->maydays); 2482 wake_up_process(wq->rescuer->task); 2483 pwq->stats[PWQ_STAT_MAYDAY]++; 2484 } 2485 } 2486 2487 static void pool_mayday_timeout(struct timer_list *t) 2488 { 2489 struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2490 struct work_struct *work; 2491 2492 raw_spin_lock_irq(&pool->lock); 2493 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2494 2495 if (need_to_create_worker(pool)) { 2496 /* 2497 * We've been trying to create a new worker but 2498 * haven't been successful. We might be hitting an 2499 * allocation deadlock. Send distress signals to 2500 * rescuers. 2501 */ 2502 list_for_each_entry(work, &pool->worklist, entry) 2503 send_mayday(work); 2504 } 2505 2506 raw_spin_unlock(&wq_mayday_lock); 2507 raw_spin_unlock_irq(&pool->lock); 2508 2509 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2510 } 2511 2512 /** 2513 * maybe_create_worker - create a new worker if necessary 2514 * @pool: pool to create a new worker for 2515 * 2516 * Create a new worker for @pool if necessary. @pool is guaranteed to 2517 * have at least one idle worker on return from this function. If 2518 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 2519 * sent to all rescuers with works scheduled on @pool to resolve 2520 * possible allocation deadlock. 2521 * 2522 * On return, need_to_create_worker() is guaranteed to be %false and 2523 * may_start_working() %true. 2524 * 2525 * LOCKING: 2526 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2527 * multiple times. Does GFP_KERNEL allocations. Called only from 2528 * manager. 2529 */ 2530 static void maybe_create_worker(struct worker_pool *pool) 2531 __releases(&pool->lock) 2532 __acquires(&pool->lock) 2533 { 2534 restart: 2535 raw_spin_unlock_irq(&pool->lock); 2536 2537 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 2538 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2539 2540 while (true) { 2541 if (create_worker(pool) || !need_to_create_worker(pool)) 2542 break; 2543 2544 schedule_timeout_interruptible(CREATE_COOLDOWN); 2545 2546 if (!need_to_create_worker(pool)) 2547 break; 2548 } 2549 2550 del_timer_sync(&pool->mayday_timer); 2551 raw_spin_lock_irq(&pool->lock); 2552 /* 2553 * This is necessary even after a new worker was just successfully 2554 * created as @pool->lock was dropped and the new worker might have 2555 * already become busy. 2556 */ 2557 if (need_to_create_worker(pool)) 2558 goto restart; 2559 } 2560 2561 /** 2562 * manage_workers - manage worker pool 2563 * @worker: self 2564 * 2565 * Assume the manager role and manage the worker pool @worker belongs 2566 * to. At any given time, there can be only zero or one manager per 2567 * pool. The exclusion is handled automatically by this function. 2568 * 2569 * The caller can safely start processing works on false return. On 2570 * true return, it's guaranteed that need_to_create_worker() is false 2571 * and may_start_working() is true. 2572 * 2573 * CONTEXT: 2574 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2575 * multiple times. Does GFP_KERNEL allocations. 2576 * 2577 * Return: 2578 * %false if the pool doesn't need management and the caller can safely 2579 * start processing works, %true if management function was performed and 2580 * the conditions that the caller verified before calling the function may 2581 * no longer be true. 2582 */ 2583 static bool manage_workers(struct worker *worker) 2584 { 2585 struct worker_pool *pool = worker->pool; 2586 2587 if (pool->flags & POOL_MANAGER_ACTIVE) 2588 return false; 2589 2590 pool->flags |= POOL_MANAGER_ACTIVE; 2591 pool->manager = worker; 2592 2593 maybe_create_worker(pool); 2594 2595 pool->manager = NULL; 2596 pool->flags &= ~POOL_MANAGER_ACTIVE; 2597 rcuwait_wake_up(&manager_wait); 2598 return true; 2599 } 2600 2601 /** 2602 * process_one_work - process single work 2603 * @worker: self 2604 * @work: work to process 2605 * 2606 * Process @work. This function contains all the logics necessary to 2607 * process a single work including synchronization against and 2608 * interaction with other workers on the same cpu, queueing and 2609 * flushing. As long as context requirement is met, any worker can 2610 * call this function to process a work. 2611 * 2612 * CONTEXT: 2613 * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2614 */ 2615 static void process_one_work(struct worker *worker, struct work_struct *work) 2616 __releases(&pool->lock) 2617 __acquires(&pool->lock) 2618 { 2619 struct pool_workqueue *pwq = get_work_pwq(work); 2620 struct worker_pool *pool = worker->pool; 2621 unsigned long work_data; 2622 #ifdef CONFIG_LOCKDEP 2623 /* 2624 * It is permissible to free the struct work_struct from 2625 * inside the function that is called from it, this we need to 2626 * take into account for lockdep too. To avoid bogus "held 2627 * lock freed" warnings as well as problems when looking into 2628 * work->lockdep_map, make a copy and use that here. 2629 */ 2630 struct lockdep_map lockdep_map; 2631 2632 lockdep_copy_map(&lockdep_map, &work->lockdep_map); 2633 #endif 2634 /* ensure we're on the correct CPU */ 2635 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2636 raw_smp_processor_id() != pool->cpu); 2637 2638 /* claim and dequeue */ 2639 debug_work_deactivate(work); 2640 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2641 worker->current_work = work; 2642 worker->current_func = work->func; 2643 worker->current_pwq = pwq; 2644 worker->current_at = worker->task->se.sum_exec_runtime; 2645 work_data = *work_data_bits(work); 2646 worker->current_color = get_work_color(work_data); 2647 2648 /* 2649 * Record wq name for cmdline and debug reporting, may get 2650 * overridden through set_worker_desc(). 2651 */ 2652 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 2653 2654 list_del_init(&work->entry); 2655 2656 /* 2657 * CPU intensive works don't participate in concurrency management. 2658 * They're the scheduler's responsibility. This takes @worker out 2659 * of concurrency management and the next code block will chain 2660 * execution of the pending work items. 2661 */ 2662 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2663 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2664 2665 /* 2666 * Kick @pool if necessary. It's always noop for per-cpu worker pools 2667 * since nr_running would always be >= 1 at this point. This is used to 2668 * chain execution of the pending work items for WORKER_NOT_RUNNING 2669 * workers such as the UNBOUND and CPU_INTENSIVE ones. 2670 */ 2671 kick_pool(pool); 2672 2673 /* 2674 * Record the last pool and clear PENDING which should be the last 2675 * update to @work. Also, do this inside @pool->lock so that 2676 * PENDING and queued state changes happen together while IRQ is 2677 * disabled. 2678 */ 2679 set_work_pool_and_clear_pending(work, pool->id); 2680 2681 pwq->stats[PWQ_STAT_STARTED]++; 2682 raw_spin_unlock_irq(&pool->lock); 2683 2684 lock_map_acquire(&pwq->wq->lockdep_map); 2685 lock_map_acquire(&lockdep_map); 2686 /* 2687 * Strictly speaking we should mark the invariant state without holding 2688 * any locks, that is, before these two lock_map_acquire()'s. 2689 * 2690 * However, that would result in: 2691 * 2692 * A(W1) 2693 * WFC(C) 2694 * A(W1) 2695 * C(C) 2696 * 2697 * Which would create W1->C->W1 dependencies, even though there is no 2698 * actual deadlock possible. There are two solutions, using a 2699 * read-recursive acquire on the work(queue) 'locks', but this will then 2700 * hit the lockdep limitation on recursive locks, or simply discard 2701 * these locks. 2702 * 2703 * AFAICT there is no possible deadlock scenario between the 2704 * flush_work() and complete() primitives (except for single-threaded 2705 * workqueues), so hiding them isn't a problem. 2706 */ 2707 lockdep_invariant_state(true); 2708 trace_workqueue_execute_start(work); 2709 worker->current_func(work); 2710 /* 2711 * While we must be careful to not use "work" after this, the trace 2712 * point will only record its address. 2713 */ 2714 trace_workqueue_execute_end(work, worker->current_func); 2715 pwq->stats[PWQ_STAT_COMPLETED]++; 2716 lock_map_release(&lockdep_map); 2717 lock_map_release(&pwq->wq->lockdep_map); 2718 2719 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2720 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2721 " last function: %ps\n", 2722 current->comm, preempt_count(), task_pid_nr(current), 2723 worker->current_func); 2724 debug_show_held_locks(current); 2725 dump_stack(); 2726 } 2727 2728 /* 2729 * The following prevents a kworker from hogging CPU on !PREEMPTION 2730 * kernels, where a requeueing work item waiting for something to 2731 * happen could deadlock with stop_machine as such work item could 2732 * indefinitely requeue itself while all other CPUs are trapped in 2733 * stop_machine. At the same time, report a quiescent RCU state so 2734 * the same condition doesn't freeze RCU. 2735 */ 2736 cond_resched(); 2737 2738 raw_spin_lock_irq(&pool->lock); 2739 2740 /* 2741 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2742 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2743 * wq_cpu_intensive_thresh_us. Clear it. 2744 */ 2745 worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2746 2747 /* tag the worker for identification in schedule() */ 2748 worker->last_func = worker->current_func; 2749 2750 /* we're done with it, release */ 2751 hash_del(&worker->hentry); 2752 worker->current_work = NULL; 2753 worker->current_func = NULL; 2754 worker->current_pwq = NULL; 2755 worker->current_color = INT_MAX; 2756 pwq_dec_nr_in_flight(pwq, work_data); 2757 } 2758 2759 /** 2760 * process_scheduled_works - process scheduled works 2761 * @worker: self 2762 * 2763 * Process all scheduled works. Please note that the scheduled list 2764 * may change while processing a work, so this function repeatedly 2765 * fetches a work from the top and executes it. 2766 * 2767 * CONTEXT: 2768 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2769 * multiple times. 2770 */ 2771 static void process_scheduled_works(struct worker *worker) 2772 { 2773 struct work_struct *work; 2774 bool first = true; 2775 2776 while ((work = list_first_entry_or_null(&worker->scheduled, 2777 struct work_struct, entry))) { 2778 if (first) { 2779 worker->pool->watchdog_ts = jiffies; 2780 first = false; 2781 } 2782 process_one_work(worker, work); 2783 } 2784 } 2785 2786 static void set_pf_worker(bool val) 2787 { 2788 mutex_lock(&wq_pool_attach_mutex); 2789 if (val) 2790 current->flags |= PF_WQ_WORKER; 2791 else 2792 current->flags &= ~PF_WQ_WORKER; 2793 mutex_unlock(&wq_pool_attach_mutex); 2794 } 2795 2796 /** 2797 * worker_thread - the worker thread function 2798 * @__worker: self 2799 * 2800 * The worker thread function. All workers belong to a worker_pool - 2801 * either a per-cpu one or dynamic unbound one. These workers process all 2802 * work items regardless of their specific target workqueue. The only 2803 * exception is work items which belong to workqueues with a rescuer which 2804 * will be explained in rescuer_thread(). 2805 * 2806 * Return: 0 2807 */ 2808 static int worker_thread(void *__worker) 2809 { 2810 struct worker *worker = __worker; 2811 struct worker_pool *pool = worker->pool; 2812 2813 /* tell the scheduler that this is a workqueue worker */ 2814 set_pf_worker(true); 2815 woke_up: 2816 raw_spin_lock_irq(&pool->lock); 2817 2818 /* am I supposed to die? */ 2819 if (unlikely(worker->flags & WORKER_DIE)) { 2820 raw_spin_unlock_irq(&pool->lock); 2821 set_pf_worker(false); 2822 2823 set_task_comm(worker->task, "kworker/dying"); 2824 ida_free(&pool->worker_ida, worker->id); 2825 worker_detach_from_pool(worker); 2826 WARN_ON_ONCE(!list_empty(&worker->entry)); 2827 kfree(worker); 2828 return 0; 2829 } 2830 2831 worker_leave_idle(worker); 2832 recheck: 2833 /* no more worker necessary? */ 2834 if (!need_more_worker(pool)) 2835 goto sleep; 2836 2837 /* do we need to manage? */ 2838 if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2839 goto recheck; 2840 2841 /* 2842 * ->scheduled list can only be filled while a worker is 2843 * preparing to process a work or actually processing it. 2844 * Make sure nobody diddled with it while I was sleeping. 2845 */ 2846 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2847 2848 /* 2849 * Finish PREP stage. We're guaranteed to have at least one idle 2850 * worker or that someone else has already assumed the manager 2851 * role. This is where @worker starts participating in concurrency 2852 * management if applicable and concurrency management is restored 2853 * after being rebound. See rebind_workers() for details. 2854 */ 2855 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2856 2857 do { 2858 struct work_struct *work = 2859 list_first_entry(&pool->worklist, 2860 struct work_struct, entry); 2861 2862 if (assign_work(work, worker, NULL)) 2863 process_scheduled_works(worker); 2864 } while (keep_working(pool)); 2865 2866 worker_set_flags(worker, WORKER_PREP); 2867 sleep: 2868 /* 2869 * pool->lock is held and there's no work to process and no need to 2870 * manage, sleep. Workers are woken up only while holding 2871 * pool->lock or from local cpu, so setting the current state 2872 * before releasing pool->lock is enough to prevent losing any 2873 * event. 2874 */ 2875 worker_enter_idle(worker); 2876 __set_current_state(TASK_IDLE); 2877 raw_spin_unlock_irq(&pool->lock); 2878 schedule(); 2879 goto woke_up; 2880 } 2881 2882 /** 2883 * rescuer_thread - the rescuer thread function 2884 * @__rescuer: self 2885 * 2886 * Workqueue rescuer thread function. There's one rescuer for each 2887 * workqueue which has WQ_MEM_RECLAIM set. 2888 * 2889 * Regular work processing on a pool may block trying to create a new 2890 * worker which uses GFP_KERNEL allocation which has slight chance of 2891 * developing into deadlock if some works currently on the same queue 2892 * need to be processed to satisfy the GFP_KERNEL allocation. This is 2893 * the problem rescuer solves. 2894 * 2895 * When such condition is possible, the pool summons rescuers of all 2896 * workqueues which have works queued on the pool and let them process 2897 * those works so that forward progress can be guaranteed. 2898 * 2899 * This should happen rarely. 2900 * 2901 * Return: 0 2902 */ 2903 static int rescuer_thread(void *__rescuer) 2904 { 2905 struct worker *rescuer = __rescuer; 2906 struct workqueue_struct *wq = rescuer->rescue_wq; 2907 bool should_stop; 2908 2909 set_user_nice(current, RESCUER_NICE_LEVEL); 2910 2911 /* 2912 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2913 * doesn't participate in concurrency management. 2914 */ 2915 set_pf_worker(true); 2916 repeat: 2917 set_current_state(TASK_IDLE); 2918 2919 /* 2920 * By the time the rescuer is requested to stop, the workqueue 2921 * shouldn't have any work pending, but @wq->maydays may still have 2922 * pwq(s) queued. This can happen by non-rescuer workers consuming 2923 * all the work items before the rescuer got to them. Go through 2924 * @wq->maydays processing before acting on should_stop so that the 2925 * list is always empty on exit. 2926 */ 2927 should_stop = kthread_should_stop(); 2928 2929 /* see whether any pwq is asking for help */ 2930 raw_spin_lock_irq(&wq_mayday_lock); 2931 2932 while (!list_empty(&wq->maydays)) { 2933 struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2934 struct pool_workqueue, mayday_node); 2935 struct worker_pool *pool = pwq->pool; 2936 struct work_struct *work, *n; 2937 2938 __set_current_state(TASK_RUNNING); 2939 list_del_init(&pwq->mayday_node); 2940 2941 raw_spin_unlock_irq(&wq_mayday_lock); 2942 2943 worker_attach_to_pool(rescuer, pool); 2944 2945 raw_spin_lock_irq(&pool->lock); 2946 2947 /* 2948 * Slurp in all works issued via this workqueue and 2949 * process'em. 2950 */ 2951 WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2952 list_for_each_entry_safe(work, n, &pool->worklist, entry) { 2953 if (get_work_pwq(work) == pwq && 2954 assign_work(work, rescuer, &n)) 2955 pwq->stats[PWQ_STAT_RESCUED]++; 2956 } 2957 2958 if (!list_empty(&rescuer->scheduled)) { 2959 process_scheduled_works(rescuer); 2960 2961 /* 2962 * The above execution of rescued work items could 2963 * have created more to rescue through 2964 * pwq_activate_first_inactive() or chained 2965 * queueing. Let's put @pwq back on mayday list so 2966 * that such back-to-back work items, which may be 2967 * being used to relieve memory pressure, don't 2968 * incur MAYDAY_INTERVAL delay inbetween. 2969 */ 2970 if (pwq->nr_active && need_to_create_worker(pool)) { 2971 raw_spin_lock(&wq_mayday_lock); 2972 /* 2973 * Queue iff we aren't racing destruction 2974 * and somebody else hasn't queued it already. 2975 */ 2976 if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2977 get_pwq(pwq); 2978 list_add_tail(&pwq->mayday_node, &wq->maydays); 2979 } 2980 raw_spin_unlock(&wq_mayday_lock); 2981 } 2982 } 2983 2984 /* 2985 * Put the reference grabbed by send_mayday(). @pool won't 2986 * go away while we're still attached to it. 2987 */ 2988 put_pwq(pwq); 2989 2990 /* 2991 * Leave this pool. Notify regular workers; otherwise, we end up 2992 * with 0 concurrency and stalling the execution. 2993 */ 2994 kick_pool(pool); 2995 2996 raw_spin_unlock_irq(&pool->lock); 2997 2998 worker_detach_from_pool(rescuer); 2999 3000 raw_spin_lock_irq(&wq_mayday_lock); 3001 } 3002 3003 raw_spin_unlock_irq(&wq_mayday_lock); 3004 3005 if (should_stop) { 3006 __set_current_state(TASK_RUNNING); 3007 set_pf_worker(false); 3008 return 0; 3009 } 3010 3011 /* rescuers should never participate in concurrency management */ 3012 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 3013 schedule(); 3014 goto repeat; 3015 } 3016 3017 /** 3018 * check_flush_dependency - check for flush dependency sanity 3019 * @target_wq: workqueue being flushed 3020 * @target_work: work item being flushed (NULL for workqueue flushes) 3021 * 3022 * %current is trying to flush the whole @target_wq or @target_work on it. 3023 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 3024 * reclaiming memory or running on a workqueue which doesn't have 3025 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 3026 * a deadlock. 3027 */ 3028 static void check_flush_dependency(struct workqueue_struct *target_wq, 3029 struct work_struct *target_work) 3030 { 3031 work_func_t target_func = target_work ? target_work->func : NULL; 3032 struct worker *worker; 3033 3034 if (target_wq->flags & WQ_MEM_RECLAIM) 3035 return; 3036 3037 worker = current_wq_worker(); 3038 3039 WARN_ONCE(current->flags & PF_MEMALLOC, 3040 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 3041 current->pid, current->comm, target_wq->name, target_func); 3042 WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 3043 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 3044 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 3045 worker->current_pwq->wq->name, worker->current_func, 3046 target_wq->name, target_func); 3047 } 3048 3049 struct wq_barrier { 3050 struct work_struct work; 3051 struct completion done; 3052 struct task_struct *task; /* purely informational */ 3053 }; 3054 3055 static void wq_barrier_func(struct work_struct *work) 3056 { 3057 struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 3058 complete(&barr->done); 3059 } 3060 3061 /** 3062 * insert_wq_barrier - insert a barrier work 3063 * @pwq: pwq to insert barrier into 3064 * @barr: wq_barrier to insert 3065 * @target: target work to attach @barr to 3066 * @worker: worker currently executing @target, NULL if @target is not executing 3067 * 3068 * @barr is linked to @target such that @barr is completed only after 3069 * @target finishes execution. Please note that the ordering 3070 * guarantee is observed only with respect to @target and on the local 3071 * cpu. 3072 * 3073 * Currently, a queued barrier can't be canceled. This is because 3074 * try_to_grab_pending() can't determine whether the work to be 3075 * grabbed is at the head of the queue and thus can't clear LINKED 3076 * flag of the previous work while there must be a valid next work 3077 * after a work with LINKED flag set. 3078 * 3079 * Note that when @worker is non-NULL, @target may be modified 3080 * underneath us, so we can't reliably determine pwq from @target. 3081 * 3082 * CONTEXT: 3083 * raw_spin_lock_irq(pool->lock). 3084 */ 3085 static void insert_wq_barrier(struct pool_workqueue *pwq, 3086 struct wq_barrier *barr, 3087 struct work_struct *target, struct worker *worker) 3088 { 3089 unsigned int work_flags = 0; 3090 unsigned int work_color; 3091 struct list_head *head; 3092 3093 /* 3094 * debugobject calls are safe here even with pool->lock locked 3095 * as we know for sure that this will not trigger any of the 3096 * checks and call back into the fixup functions where we 3097 * might deadlock. 3098 */ 3099 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 3100 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 3101 3102 init_completion_map(&barr->done, &target->lockdep_map); 3103 3104 barr->task = current; 3105 3106 /* The barrier work item does not participate in pwq->nr_active. */ 3107 work_flags |= WORK_STRUCT_INACTIVE; 3108 3109 /* 3110 * If @target is currently being executed, schedule the 3111 * barrier to the worker; otherwise, put it after @target. 3112 */ 3113 if (worker) { 3114 head = worker->scheduled.next; 3115 work_color = worker->current_color; 3116 } else { 3117 unsigned long *bits = work_data_bits(target); 3118 3119 head = target->entry.next; 3120 /* there can already be other linked works, inherit and set */ 3121 work_flags |= *bits & WORK_STRUCT_LINKED; 3122 work_color = get_work_color(*bits); 3123 __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3124 } 3125 3126 pwq->nr_in_flight[work_color]++; 3127 work_flags |= work_color_to_flags(work_color); 3128 3129 insert_work(pwq, &barr->work, head, work_flags); 3130 } 3131 3132 /** 3133 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 3134 * @wq: workqueue being flushed 3135 * @flush_color: new flush color, < 0 for no-op 3136 * @work_color: new work color, < 0 for no-op 3137 * 3138 * Prepare pwqs for workqueue flushing. 3139 * 3140 * If @flush_color is non-negative, flush_color on all pwqs should be 3141 * -1. If no pwq has in-flight commands at the specified color, all 3142 * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3143 * has in flight commands, its pwq->flush_color is set to 3144 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 3145 * wakeup logic is armed and %true is returned. 3146 * 3147 * The caller should have initialized @wq->first_flusher prior to 3148 * calling this function with non-negative @flush_color. If 3149 * @flush_color is negative, no flush color update is done and %false 3150 * is returned. 3151 * 3152 * If @work_color is non-negative, all pwqs should have the same 3153 * work_color which is previous to @work_color and all will be 3154 * advanced to @work_color. 3155 * 3156 * CONTEXT: 3157 * mutex_lock(wq->mutex). 3158 * 3159 * Return: 3160 * %true if @flush_color >= 0 and there's something to flush. %false 3161 * otherwise. 3162 */ 3163 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 3164 int flush_color, int work_color) 3165 { 3166 bool wait = false; 3167 struct pool_workqueue *pwq; 3168 3169 if (flush_color >= 0) { 3170 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3171 atomic_set(&wq->nr_pwqs_to_flush, 1); 3172 } 3173 3174 for_each_pwq(pwq, wq) { 3175 struct worker_pool *pool = pwq->pool; 3176 3177 raw_spin_lock_irq(&pool->lock); 3178 3179 if (flush_color >= 0) { 3180 WARN_ON_ONCE(pwq->flush_color != -1); 3181 3182 if (pwq->nr_in_flight[flush_color]) { 3183 pwq->flush_color = flush_color; 3184 atomic_inc(&wq->nr_pwqs_to_flush); 3185 wait = true; 3186 } 3187 } 3188 3189 if (work_color >= 0) { 3190 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3191 pwq->work_color = work_color; 3192 } 3193 3194 raw_spin_unlock_irq(&pool->lock); 3195 } 3196 3197 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 3198 complete(&wq->first_flusher->done); 3199 3200 return wait; 3201 } 3202 3203 /** 3204 * __flush_workqueue - ensure that any scheduled work has run to completion. 3205 * @wq: workqueue to flush 3206 * 3207 * This function sleeps until all work items which were queued on entry 3208 * have finished execution, but it is not livelocked by new incoming ones. 3209 */ 3210 void __flush_workqueue(struct workqueue_struct *wq) 3211 { 3212 struct wq_flusher this_flusher = { 3213 .list = LIST_HEAD_INIT(this_flusher.list), 3214 .flush_color = -1, 3215 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 3216 }; 3217 int next_color; 3218 3219 if (WARN_ON(!wq_online)) 3220 return; 3221 3222 lock_map_acquire(&wq->lockdep_map); 3223 lock_map_release(&wq->lockdep_map); 3224 3225 mutex_lock(&wq->mutex); 3226 3227 /* 3228 * Start-to-wait phase 3229 */ 3230 next_color = work_next_color(wq->work_color); 3231 3232 if (next_color != wq->flush_color) { 3233 /* 3234 * Color space is not full. The current work_color 3235 * becomes our flush_color and work_color is advanced 3236 * by one. 3237 */ 3238 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 3239 this_flusher.flush_color = wq->work_color; 3240 wq->work_color = next_color; 3241 3242 if (!wq->first_flusher) { 3243 /* no flush in progress, become the first flusher */ 3244 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 3245 3246 wq->first_flusher = &this_flusher; 3247 3248 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 3249 wq->work_color)) { 3250 /* nothing to flush, done */ 3251 wq->flush_color = next_color; 3252 wq->first_flusher = NULL; 3253 goto out_unlock; 3254 } 3255 } else { 3256 /* wait in queue */ 3257 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 3258 list_add_tail(&this_flusher.list, &wq->flusher_queue); 3259 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 3260 } 3261 } else { 3262 /* 3263 * Oops, color space is full, wait on overflow queue. 3264 * The next flush completion will assign us 3265 * flush_color and transfer to flusher_queue. 3266 */ 3267 list_add_tail(&this_flusher.list, &wq->flusher_overflow); 3268 } 3269 3270 check_flush_dependency(wq, NULL); 3271 3272 mutex_unlock(&wq->mutex); 3273 3274 wait_for_completion(&this_flusher.done); 3275 3276 /* 3277 * Wake-up-and-cascade phase 3278 * 3279 * First flushers are responsible for cascading flushes and 3280 * handling overflow. Non-first flushers can simply return. 3281 */ 3282 if (READ_ONCE(wq->first_flusher) != &this_flusher) 3283 return; 3284 3285 mutex_lock(&wq->mutex); 3286 3287 /* we might have raced, check again with mutex held */ 3288 if (wq->first_flusher != &this_flusher) 3289 goto out_unlock; 3290 3291 WRITE_ONCE(wq->first_flusher, NULL); 3292 3293 WARN_ON_ONCE(!list_empty(&this_flusher.list)); 3294 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 3295 3296 while (true) { 3297 struct wq_flusher *next, *tmp; 3298 3299 /* complete all the flushers sharing the current flush color */ 3300 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 3301 if (next->flush_color != wq->flush_color) 3302 break; 3303 list_del_init(&next->list); 3304 complete(&next->done); 3305 } 3306 3307 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 3308 wq->flush_color != work_next_color(wq->work_color)); 3309 3310 /* this flush_color is finished, advance by one */ 3311 wq->flush_color = work_next_color(wq->flush_color); 3312 3313 /* one color has been freed, handle overflow queue */ 3314 if (!list_empty(&wq->flusher_overflow)) { 3315 /* 3316 * Assign the same color to all overflowed 3317 * flushers, advance work_color and append to 3318 * flusher_queue. This is the start-to-wait 3319 * phase for these overflowed flushers. 3320 */ 3321 list_for_each_entry(tmp, &wq->flusher_overflow, list) 3322 tmp->flush_color = wq->work_color; 3323 3324 wq->work_color = work_next_color(wq->work_color); 3325 3326 list_splice_tail_init(&wq->flusher_overflow, 3327 &wq->flusher_queue); 3328 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 3329 } 3330 3331 if (list_empty(&wq->flusher_queue)) { 3332 WARN_ON_ONCE(wq->flush_color != wq->work_color); 3333 break; 3334 } 3335 3336 /* 3337 * Need to flush more colors. Make the next flusher 3338 * the new first flusher and arm pwqs. 3339 */ 3340 WARN_ON_ONCE(wq->flush_color == wq->work_color); 3341 WARN_ON_ONCE(wq->flush_color != next->flush_color); 3342 3343 list_del_init(&next->list); 3344 wq->first_flusher = next; 3345 3346 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 3347 break; 3348 3349 /* 3350 * Meh... this color is already done, clear first 3351 * flusher and repeat cascading. 3352 */ 3353 wq->first_flusher = NULL; 3354 } 3355 3356 out_unlock: 3357 mutex_unlock(&wq->mutex); 3358 } 3359 EXPORT_SYMBOL(__flush_workqueue); 3360 3361 /** 3362 * drain_workqueue - drain a workqueue 3363 * @wq: workqueue to drain 3364 * 3365 * Wait until the workqueue becomes empty. While draining is in progress, 3366 * only chain queueing is allowed. IOW, only currently pending or running 3367 * work items on @wq can queue further work items on it. @wq is flushed 3368 * repeatedly until it becomes empty. The number of flushing is determined 3369 * by the depth of chaining and should be relatively short. Whine if it 3370 * takes too long. 3371 */ 3372 void drain_workqueue(struct workqueue_struct *wq) 3373 { 3374 unsigned int flush_cnt = 0; 3375 struct pool_workqueue *pwq; 3376 3377 /* 3378 * __queue_work() needs to test whether there are drainers, is much 3379 * hotter than drain_workqueue() and already looks at @wq->flags. 3380 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 3381 */ 3382 mutex_lock(&wq->mutex); 3383 if (!wq->nr_drainers++) 3384 wq->flags |= __WQ_DRAINING; 3385 mutex_unlock(&wq->mutex); 3386 reflush: 3387 __flush_workqueue(wq); 3388 3389 mutex_lock(&wq->mutex); 3390 3391 for_each_pwq(pwq, wq) { 3392 bool drained; 3393 3394 raw_spin_lock_irq(&pwq->pool->lock); 3395 drained = pwq_is_empty(pwq); 3396 raw_spin_unlock_irq(&pwq->pool->lock); 3397 3398 if (drained) 3399 continue; 3400 3401 if (++flush_cnt == 10 || 3402 (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3403 pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3404 wq->name, __func__, flush_cnt); 3405 3406 mutex_unlock(&wq->mutex); 3407 goto reflush; 3408 } 3409 3410 if (!--wq->nr_drainers) 3411 wq->flags &= ~__WQ_DRAINING; 3412 mutex_unlock(&wq->mutex); 3413 } 3414 EXPORT_SYMBOL_GPL(drain_workqueue); 3415 3416 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3417 bool from_cancel) 3418 { 3419 struct worker *worker = NULL; 3420 struct worker_pool *pool; 3421 struct pool_workqueue *pwq; 3422 3423 might_sleep(); 3424 3425 rcu_read_lock(); 3426 pool = get_work_pool(work); 3427 if (!pool) { 3428 rcu_read_unlock(); 3429 return false; 3430 } 3431 3432 raw_spin_lock_irq(&pool->lock); 3433 /* see the comment in try_to_grab_pending() with the same code */ 3434 pwq = get_work_pwq(work); 3435 if (pwq) { 3436 if (unlikely(pwq->pool != pool)) 3437 goto already_gone; 3438 } else { 3439 worker = find_worker_executing_work(pool, work); 3440 if (!worker) 3441 goto already_gone; 3442 pwq = worker->current_pwq; 3443 } 3444 3445 check_flush_dependency(pwq->wq, work); 3446 3447 insert_wq_barrier(pwq, barr, work, worker); 3448 raw_spin_unlock_irq(&pool->lock); 3449 3450 /* 3451 * Force a lock recursion deadlock when using flush_work() inside a 3452 * single-threaded or rescuer equipped workqueue. 3453 * 3454 * For single threaded workqueues the deadlock happens when the work 3455 * is after the work issuing the flush_work(). For rescuer equipped 3456 * workqueues the deadlock happens when the rescuer stalls, blocking 3457 * forward progress. 3458 */ 3459 if (!from_cancel && 3460 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3461 lock_map_acquire(&pwq->wq->lockdep_map); 3462 lock_map_release(&pwq->wq->lockdep_map); 3463 } 3464 rcu_read_unlock(); 3465 return true; 3466 already_gone: 3467 raw_spin_unlock_irq(&pool->lock); 3468 rcu_read_unlock(); 3469 return false; 3470 } 3471 3472 static bool __flush_work(struct work_struct *work, bool from_cancel) 3473 { 3474 struct wq_barrier barr; 3475 3476 if (WARN_ON(!wq_online)) 3477 return false; 3478 3479 if (WARN_ON(!work->func)) 3480 return false; 3481 3482 lock_map_acquire(&work->lockdep_map); 3483 lock_map_release(&work->lockdep_map); 3484 3485 if (start_flush_work(work, &barr, from_cancel)) { 3486 wait_for_completion(&barr.done); 3487 destroy_work_on_stack(&barr.work); 3488 return true; 3489 } else { 3490 return false; 3491 } 3492 } 3493 3494 /** 3495 * flush_work - wait for a work to finish executing the last queueing instance 3496 * @work: the work to flush 3497 * 3498 * Wait until @work has finished execution. @work is guaranteed to be idle 3499 * on return if it hasn't been requeued since flush started. 3500 * 3501 * Return: 3502 * %true if flush_work() waited for the work to finish execution, 3503 * %false if it was already idle. 3504 */ 3505 bool flush_work(struct work_struct *work) 3506 { 3507 return __flush_work(work, false); 3508 } 3509 EXPORT_SYMBOL_GPL(flush_work); 3510 3511 struct cwt_wait { 3512 wait_queue_entry_t wait; 3513 struct work_struct *work; 3514 }; 3515 3516 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 3517 { 3518 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 3519 3520 if (cwait->work != key) 3521 return 0; 3522 return autoremove_wake_function(wait, mode, sync, key); 3523 } 3524 3525 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3526 { 3527 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3528 unsigned long flags; 3529 int ret; 3530 3531 do { 3532 ret = try_to_grab_pending(work, is_dwork, &flags); 3533 /* 3534 * If someone else is already canceling, wait for it to 3535 * finish. flush_work() doesn't work for PREEMPT_NONE 3536 * because we may get scheduled between @work's completion 3537 * and the other canceling task resuming and clearing 3538 * CANCELING - flush_work() will return false immediately 3539 * as @work is no longer busy, try_to_grab_pending() will 3540 * return -ENOENT as @work is still being canceled and the 3541 * other canceling task won't be able to clear CANCELING as 3542 * we're hogging the CPU. 3543 * 3544 * Let's wait for completion using a waitqueue. As this 3545 * may lead to the thundering herd problem, use a custom 3546 * wake function which matches @work along with exclusive 3547 * wait and wakeup. 3548 */ 3549 if (unlikely(ret == -ENOENT)) { 3550 struct cwt_wait cwait; 3551 3552 init_wait(&cwait.wait); 3553 cwait.wait.func = cwt_wakefn; 3554 cwait.work = work; 3555 3556 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 3557 TASK_UNINTERRUPTIBLE); 3558 if (work_is_canceling(work)) 3559 schedule(); 3560 finish_wait(&cancel_waitq, &cwait.wait); 3561 } 3562 } while (unlikely(ret < 0)); 3563 3564 /* tell other tasks trying to grab @work to back off */ 3565 mark_work_canceling(work); 3566 local_irq_restore(flags); 3567 3568 /* 3569 * This allows canceling during early boot. We know that @work 3570 * isn't executing. 3571 */ 3572 if (wq_online) 3573 __flush_work(work, true); 3574 3575 clear_work_data(work); 3576 3577 /* 3578 * Paired with prepare_to_wait() above so that either 3579 * waitqueue_active() is visible here or !work_is_canceling() is 3580 * visible there. 3581 */ 3582 smp_mb(); 3583 if (waitqueue_active(&cancel_waitq)) 3584 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 3585 3586 return ret; 3587 } 3588 3589 /** 3590 * cancel_work_sync - cancel a work and wait for it to finish 3591 * @work: the work to cancel 3592 * 3593 * Cancel @work and wait for its execution to finish. This function 3594 * can be used even if the work re-queues itself or migrates to 3595 * another workqueue. On return from this function, @work is 3596 * guaranteed to be not pending or executing on any CPU. 3597 * 3598 * cancel_work_sync(&delayed_work->work) must not be used for 3599 * delayed_work's. Use cancel_delayed_work_sync() instead. 3600 * 3601 * The caller must ensure that the workqueue on which @work was last 3602 * queued can't be destroyed before this function returns. 3603 * 3604 * Return: 3605 * %true if @work was pending, %false otherwise. 3606 */ 3607 bool cancel_work_sync(struct work_struct *work) 3608 { 3609 return __cancel_work_timer(work, false); 3610 } 3611 EXPORT_SYMBOL_GPL(cancel_work_sync); 3612 3613 /** 3614 * flush_delayed_work - wait for a dwork to finish executing the last queueing 3615 * @dwork: the delayed work to flush 3616 * 3617 * Delayed timer is cancelled and the pending work is queued for 3618 * immediate execution. Like flush_work(), this function only 3619 * considers the last queueing instance of @dwork. 3620 * 3621 * Return: 3622 * %true if flush_work() waited for the work to finish execution, 3623 * %false if it was already idle. 3624 */ 3625 bool flush_delayed_work(struct delayed_work *dwork) 3626 { 3627 local_irq_disable(); 3628 if (del_timer_sync(&dwork->timer)) 3629 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 3630 local_irq_enable(); 3631 return flush_work(&dwork->work); 3632 } 3633 EXPORT_SYMBOL(flush_delayed_work); 3634 3635 /** 3636 * flush_rcu_work - wait for a rwork to finish executing the last queueing 3637 * @rwork: the rcu work to flush 3638 * 3639 * Return: 3640 * %true if flush_rcu_work() waited for the work to finish execution, 3641 * %false if it was already idle. 3642 */ 3643 bool flush_rcu_work(struct rcu_work *rwork) 3644 { 3645 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 3646 rcu_barrier(); 3647 flush_work(&rwork->work); 3648 return true; 3649 } else { 3650 return flush_work(&rwork->work); 3651 } 3652 } 3653 EXPORT_SYMBOL(flush_rcu_work); 3654 3655 static bool __cancel_work(struct work_struct *work, bool is_dwork) 3656 { 3657 unsigned long flags; 3658 int ret; 3659 3660 do { 3661 ret = try_to_grab_pending(work, is_dwork, &flags); 3662 } while (unlikely(ret == -EAGAIN)); 3663 3664 if (unlikely(ret < 0)) 3665 return false; 3666 3667 set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3668 local_irq_restore(flags); 3669 return ret; 3670 } 3671 3672 /* 3673 * See cancel_delayed_work() 3674 */ 3675 bool cancel_work(struct work_struct *work) 3676 { 3677 return __cancel_work(work, false); 3678 } 3679 EXPORT_SYMBOL(cancel_work); 3680 3681 /** 3682 * cancel_delayed_work - cancel a delayed work 3683 * @dwork: delayed_work to cancel 3684 * 3685 * Kill off a pending delayed_work. 3686 * 3687 * Return: %true if @dwork was pending and canceled; %false if it wasn't 3688 * pending. 3689 * 3690 * Note: 3691 * The work callback function may still be running on return, unless 3692 * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3693 * use cancel_delayed_work_sync() to wait on it. 3694 * 3695 * This function is safe to call from any context including IRQ handler. 3696 */ 3697 bool cancel_delayed_work(struct delayed_work *dwork) 3698 { 3699 return __cancel_work(&dwork->work, true); 3700 } 3701 EXPORT_SYMBOL(cancel_delayed_work); 3702 3703 /** 3704 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3705 * @dwork: the delayed work cancel 3706 * 3707 * This is cancel_work_sync() for delayed works. 3708 * 3709 * Return: 3710 * %true if @dwork was pending, %false otherwise. 3711 */ 3712 bool cancel_delayed_work_sync(struct delayed_work *dwork) 3713 { 3714 return __cancel_work_timer(&dwork->work, true); 3715 } 3716 EXPORT_SYMBOL(cancel_delayed_work_sync); 3717 3718 /** 3719 * schedule_on_each_cpu - execute a function synchronously on each online CPU 3720 * @func: the function to call 3721 * 3722 * schedule_on_each_cpu() executes @func on each online CPU using the 3723 * system workqueue and blocks until all CPUs have completed. 3724 * schedule_on_each_cpu() is very slow. 3725 * 3726 * Return: 3727 * 0 on success, -errno on failure. 3728 */ 3729 int schedule_on_each_cpu(work_func_t func) 3730 { 3731 int cpu; 3732 struct work_struct __percpu *works; 3733 3734 works = alloc_percpu(struct work_struct); 3735 if (!works) 3736 return -ENOMEM; 3737 3738 cpus_read_lock(); 3739 3740 for_each_online_cpu(cpu) { 3741 struct work_struct *work = per_cpu_ptr(works, cpu); 3742 3743 INIT_WORK(work, func); 3744 schedule_work_on(cpu, work); 3745 } 3746 3747 for_each_online_cpu(cpu) 3748 flush_work(per_cpu_ptr(works, cpu)); 3749 3750 cpus_read_unlock(); 3751 free_percpu(works); 3752 return 0; 3753 } 3754 3755 /** 3756 * execute_in_process_context - reliably execute the routine with user context 3757 * @fn: the function to execute 3758 * @ew: guaranteed storage for the execute work structure (must 3759 * be available when the work executes) 3760 * 3761 * Executes the function immediately if process context is available, 3762 * otherwise schedules the function for delayed execution. 3763 * 3764 * Return: 0 - function was executed 3765 * 1 - function was scheduled for execution 3766 */ 3767 int execute_in_process_context(work_func_t fn, struct execute_work *ew) 3768 { 3769 if (!in_interrupt()) { 3770 fn(&ew->work); 3771 return 0; 3772 } 3773 3774 INIT_WORK(&ew->work, fn); 3775 schedule_work(&ew->work); 3776 3777 return 1; 3778 } 3779 EXPORT_SYMBOL_GPL(execute_in_process_context); 3780 3781 /** 3782 * free_workqueue_attrs - free a workqueue_attrs 3783 * @attrs: workqueue_attrs to free 3784 * 3785 * Undo alloc_workqueue_attrs(). 3786 */ 3787 void free_workqueue_attrs(struct workqueue_attrs *attrs) 3788 { 3789 if (attrs) { 3790 free_cpumask_var(attrs->cpumask); 3791 free_cpumask_var(attrs->__pod_cpumask); 3792 kfree(attrs); 3793 } 3794 } 3795 3796 /** 3797 * alloc_workqueue_attrs - allocate a workqueue_attrs 3798 * 3799 * Allocate a new workqueue_attrs, initialize with default settings and 3800 * return it. 3801 * 3802 * Return: The allocated new workqueue_attr on success. %NULL on failure. 3803 */ 3804 struct workqueue_attrs *alloc_workqueue_attrs(void) 3805 { 3806 struct workqueue_attrs *attrs; 3807 3808 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 3809 if (!attrs) 3810 goto fail; 3811 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 3812 goto fail; 3813 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 3814 goto fail; 3815 3816 cpumask_copy(attrs->cpumask, cpu_possible_mask); 3817 attrs->affn_scope = WQ_AFFN_DFL; 3818 return attrs; 3819 fail: 3820 free_workqueue_attrs(attrs); 3821 return NULL; 3822 } 3823 3824 static void copy_workqueue_attrs(struct workqueue_attrs *to, 3825 const struct workqueue_attrs *from) 3826 { 3827 to->nice = from->nice; 3828 cpumask_copy(to->cpumask, from->cpumask); 3829 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 3830 to->affn_strict = from->affn_strict; 3831 3832 /* 3833 * Unlike hash and equality test, copying shouldn't ignore wq-only 3834 * fields as copying is used for both pool and wq attrs. Instead, 3835 * get_unbound_pool() explicitly clears the fields. 3836 */ 3837 to->affn_scope = from->affn_scope; 3838 to->ordered = from->ordered; 3839 } 3840 3841 /* 3842 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 3843 * comments in 'struct workqueue_attrs' definition. 3844 */ 3845 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 3846 { 3847 attrs->affn_scope = WQ_AFFN_NR_TYPES; 3848 attrs->ordered = false; 3849 } 3850 3851 /* hash value of the content of @attr */ 3852 static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 3853 { 3854 u32 hash = 0; 3855 3856 hash = jhash_1word(attrs->nice, hash); 3857 hash = jhash(cpumask_bits(attrs->cpumask), 3858 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 3859 hash = jhash(cpumask_bits(attrs->__pod_cpumask), 3860 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 3861 hash = jhash_1word(attrs->affn_strict, hash); 3862 return hash; 3863 } 3864 3865 /* content equality test */ 3866 static bool wqattrs_equal(const struct workqueue_attrs *a, 3867 const struct workqueue_attrs *b) 3868 { 3869 if (a->nice != b->nice) 3870 return false; 3871 if (!cpumask_equal(a->cpumask, b->cpumask)) 3872 return false; 3873 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 3874 return false; 3875 if (a->affn_strict != b->affn_strict) 3876 return false; 3877 return true; 3878 } 3879 3880 /* Update @attrs with actually available CPUs */ 3881 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 3882 const cpumask_t *unbound_cpumask) 3883 { 3884 /* 3885 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 3886 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 3887 * @unbound_cpumask. 3888 */ 3889 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 3890 if (unlikely(cpumask_empty(attrs->cpumask))) 3891 cpumask_copy(attrs->cpumask, unbound_cpumask); 3892 } 3893 3894 /* find wq_pod_type to use for @attrs */ 3895 static const struct wq_pod_type * 3896 wqattrs_pod_type(const struct workqueue_attrs *attrs) 3897 { 3898 enum wq_affn_scope scope; 3899 struct wq_pod_type *pt; 3900 3901 /* to synchronize access to wq_affn_dfl */ 3902 lockdep_assert_held(&wq_pool_mutex); 3903 3904 if (attrs->affn_scope == WQ_AFFN_DFL) 3905 scope = wq_affn_dfl; 3906 else 3907 scope = attrs->affn_scope; 3908 3909 pt = &wq_pod_types[scope]; 3910 3911 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 3912 likely(pt->nr_pods)) 3913 return pt; 3914 3915 /* 3916 * Before workqueue_init_topology(), only SYSTEM is available which is 3917 * initialized in workqueue_init_early(). 3918 */ 3919 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 3920 BUG_ON(!pt->nr_pods); 3921 return pt; 3922 } 3923 3924 /** 3925 * init_worker_pool - initialize a newly zalloc'd worker_pool 3926 * @pool: worker_pool to initialize 3927 * 3928 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3929 * 3930 * Return: 0 on success, -errno on failure. Even on failure, all fields 3931 * inside @pool proper are initialized and put_unbound_pool() can be called 3932 * on @pool safely to release it. 3933 */ 3934 static int init_worker_pool(struct worker_pool *pool) 3935 { 3936 raw_spin_lock_init(&pool->lock); 3937 pool->id = -1; 3938 pool->cpu = -1; 3939 pool->node = NUMA_NO_NODE; 3940 pool->flags |= POOL_DISASSOCIATED; 3941 pool->watchdog_ts = jiffies; 3942 INIT_LIST_HEAD(&pool->worklist); 3943 INIT_LIST_HEAD(&pool->idle_list); 3944 hash_init(pool->busy_hash); 3945 3946 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 3947 INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 3948 3949 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 3950 3951 INIT_LIST_HEAD(&pool->workers); 3952 INIT_LIST_HEAD(&pool->dying_workers); 3953 3954 ida_init(&pool->worker_ida); 3955 INIT_HLIST_NODE(&pool->hash_node); 3956 pool->refcnt = 1; 3957 3958 /* shouldn't fail above this point */ 3959 pool->attrs = alloc_workqueue_attrs(); 3960 if (!pool->attrs) 3961 return -ENOMEM; 3962 3963 wqattrs_clear_for_pool(pool->attrs); 3964 3965 return 0; 3966 } 3967 3968 #ifdef CONFIG_LOCKDEP 3969 static void wq_init_lockdep(struct workqueue_struct *wq) 3970 { 3971 char *lock_name; 3972 3973 lockdep_register_key(&wq->key); 3974 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3975 if (!lock_name) 3976 lock_name = wq->name; 3977 3978 wq->lock_name = lock_name; 3979 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3980 } 3981 3982 static void wq_unregister_lockdep(struct workqueue_struct *wq) 3983 { 3984 lockdep_unregister_key(&wq->key); 3985 } 3986 3987 static void wq_free_lockdep(struct workqueue_struct *wq) 3988 { 3989 if (wq->lock_name != wq->name) 3990 kfree(wq->lock_name); 3991 } 3992 #else 3993 static void wq_init_lockdep(struct workqueue_struct *wq) 3994 { 3995 } 3996 3997 static void wq_unregister_lockdep(struct workqueue_struct *wq) 3998 { 3999 } 4000 4001 static void wq_free_lockdep(struct workqueue_struct *wq) 4002 { 4003 } 4004 #endif 4005 4006 static void rcu_free_wq(struct rcu_head *rcu) 4007 { 4008 struct workqueue_struct *wq = 4009 container_of(rcu, struct workqueue_struct, rcu); 4010 4011 wq_free_lockdep(wq); 4012 free_percpu(wq->cpu_pwq); 4013 free_workqueue_attrs(wq->unbound_attrs); 4014 kfree(wq); 4015 } 4016 4017 static void rcu_free_pool(struct rcu_head *rcu) 4018 { 4019 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 4020 4021 ida_destroy(&pool->worker_ida); 4022 free_workqueue_attrs(pool->attrs); 4023 kfree(pool); 4024 } 4025 4026 /** 4027 * put_unbound_pool - put a worker_pool 4028 * @pool: worker_pool to put 4029 * 4030 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 4031 * safe manner. get_unbound_pool() calls this function on its failure path 4032 * and this function should be able to release pools which went through, 4033 * successfully or not, init_worker_pool(). 4034 * 4035 * Should be called with wq_pool_mutex held. 4036 */ 4037 static void put_unbound_pool(struct worker_pool *pool) 4038 { 4039 DECLARE_COMPLETION_ONSTACK(detach_completion); 4040 struct worker *worker; 4041 LIST_HEAD(cull_list); 4042 4043 lockdep_assert_held(&wq_pool_mutex); 4044 4045 if (--pool->refcnt) 4046 return; 4047 4048 /* sanity checks */ 4049 if (WARN_ON(!(pool->cpu < 0)) || 4050 WARN_ON(!list_empty(&pool->worklist))) 4051 return; 4052 4053 /* release id and unhash */ 4054 if (pool->id >= 0) 4055 idr_remove(&worker_pool_idr, pool->id); 4056 hash_del(&pool->hash_node); 4057 4058 /* 4059 * Become the manager and destroy all workers. This prevents 4060 * @pool's workers from blocking on attach_mutex. We're the last 4061 * manager and @pool gets freed with the flag set. 4062 * 4063 * Having a concurrent manager is quite unlikely to happen as we can 4064 * only get here with 4065 * pwq->refcnt == pool->refcnt == 0 4066 * which implies no work queued to the pool, which implies no worker can 4067 * become the manager. However a worker could have taken the role of 4068 * manager before the refcnts dropped to 0, since maybe_create_worker() 4069 * drops pool->lock 4070 */ 4071 while (true) { 4072 rcuwait_wait_event(&manager_wait, 4073 !(pool->flags & POOL_MANAGER_ACTIVE), 4074 TASK_UNINTERRUPTIBLE); 4075 4076 mutex_lock(&wq_pool_attach_mutex); 4077 raw_spin_lock_irq(&pool->lock); 4078 if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 4079 pool->flags |= POOL_MANAGER_ACTIVE; 4080 break; 4081 } 4082 raw_spin_unlock_irq(&pool->lock); 4083 mutex_unlock(&wq_pool_attach_mutex); 4084 } 4085 4086 while ((worker = first_idle_worker(pool))) 4087 set_worker_dying(worker, &cull_list); 4088 WARN_ON(pool->nr_workers || pool->nr_idle); 4089 raw_spin_unlock_irq(&pool->lock); 4090 4091 wake_dying_workers(&cull_list); 4092 4093 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 4094 pool->detach_completion = &detach_completion; 4095 mutex_unlock(&wq_pool_attach_mutex); 4096 4097 if (pool->detach_completion) 4098 wait_for_completion(pool->detach_completion); 4099 4100 /* shut down the timers */ 4101 del_timer_sync(&pool->idle_timer); 4102 cancel_work_sync(&pool->idle_cull_work); 4103 del_timer_sync(&pool->mayday_timer); 4104 4105 /* RCU protected to allow dereferences from get_work_pool() */ 4106 call_rcu(&pool->rcu, rcu_free_pool); 4107 } 4108 4109 /** 4110 * get_unbound_pool - get a worker_pool with the specified attributes 4111 * @attrs: the attributes of the worker_pool to get 4112 * 4113 * Obtain a worker_pool which has the same attributes as @attrs, bump the 4114 * reference count and return it. If there already is a matching 4115 * worker_pool, it will be used; otherwise, this function attempts to 4116 * create a new one. 4117 * 4118 * Should be called with wq_pool_mutex held. 4119 * 4120 * Return: On success, a worker_pool with the same attributes as @attrs. 4121 * On failure, %NULL. 4122 */ 4123 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 4124 { 4125 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 4126 u32 hash = wqattrs_hash(attrs); 4127 struct worker_pool *pool; 4128 int pod, node = NUMA_NO_NODE; 4129 4130 lockdep_assert_held(&wq_pool_mutex); 4131 4132 /* do we already have a matching pool? */ 4133 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 4134 if (wqattrs_equal(pool->attrs, attrs)) { 4135 pool->refcnt++; 4136 return pool; 4137 } 4138 } 4139 4140 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 4141 for (pod = 0; pod < pt->nr_pods; pod++) { 4142 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 4143 node = pt->pod_node[pod]; 4144 break; 4145 } 4146 } 4147 4148 /* nope, create a new one */ 4149 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 4150 if (!pool || init_worker_pool(pool) < 0) 4151 goto fail; 4152 4153 pool->node = node; 4154 copy_workqueue_attrs(pool->attrs, attrs); 4155 wqattrs_clear_for_pool(pool->attrs); 4156 4157 if (worker_pool_assign_id(pool) < 0) 4158 goto fail; 4159 4160 /* create and start the initial worker */ 4161 if (wq_online && !create_worker(pool)) 4162 goto fail; 4163 4164 /* install */ 4165 hash_add(unbound_pool_hash, &pool->hash_node, hash); 4166 4167 return pool; 4168 fail: 4169 if (pool) 4170 put_unbound_pool(pool); 4171 return NULL; 4172 } 4173 4174 static void rcu_free_pwq(struct rcu_head *rcu) 4175 { 4176 kmem_cache_free(pwq_cache, 4177 container_of(rcu, struct pool_workqueue, rcu)); 4178 } 4179 4180 /* 4181 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 4182 * refcnt and needs to be destroyed. 4183 */ 4184 static void pwq_release_workfn(struct kthread_work *work) 4185 { 4186 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 4187 release_work); 4188 struct workqueue_struct *wq = pwq->wq; 4189 struct worker_pool *pool = pwq->pool; 4190 bool is_last = false; 4191 4192 /* 4193 * When @pwq is not linked, it doesn't hold any reference to the 4194 * @wq, and @wq is invalid to access. 4195 */ 4196 if (!list_empty(&pwq->pwqs_node)) { 4197 mutex_lock(&wq->mutex); 4198 list_del_rcu(&pwq->pwqs_node); 4199 is_last = list_empty(&wq->pwqs); 4200 mutex_unlock(&wq->mutex); 4201 } 4202 4203 if (wq->flags & WQ_UNBOUND) { 4204 mutex_lock(&wq_pool_mutex); 4205 put_unbound_pool(pool); 4206 mutex_unlock(&wq_pool_mutex); 4207 } 4208 4209 call_rcu(&pwq->rcu, rcu_free_pwq); 4210 4211 /* 4212 * If we're the last pwq going away, @wq is already dead and no one 4213 * is gonna access it anymore. Schedule RCU free. 4214 */ 4215 if (is_last) { 4216 wq_unregister_lockdep(wq); 4217 call_rcu(&wq->rcu, rcu_free_wq); 4218 } 4219 } 4220 4221 /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4222 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4223 struct worker_pool *pool) 4224 { 4225 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4226 4227 memset(pwq, 0, sizeof(*pwq)); 4228 4229 pwq->pool = pool; 4230 pwq->wq = wq; 4231 pwq->flush_color = -1; 4232 pwq->refcnt = 1; 4233 INIT_LIST_HEAD(&pwq->inactive_works); 4234 INIT_LIST_HEAD(&pwq->pwqs_node); 4235 INIT_LIST_HEAD(&pwq->mayday_node); 4236 kthread_init_work(&pwq->release_work, pwq_release_workfn); 4237 } 4238 4239 /* sync @pwq with the current state of its associated wq and link it */ 4240 static void link_pwq(struct pool_workqueue *pwq) 4241 { 4242 struct workqueue_struct *wq = pwq->wq; 4243 4244 lockdep_assert_held(&wq->mutex); 4245 4246 /* may be called multiple times, ignore if already linked */ 4247 if (!list_empty(&pwq->pwqs_node)) 4248 return; 4249 4250 /* set the matching work_color */ 4251 pwq->work_color = wq->work_color; 4252 4253 /* link in @pwq */ 4254 list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4255 } 4256 4257 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4258 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4259 const struct workqueue_attrs *attrs) 4260 { 4261 struct worker_pool *pool; 4262 struct pool_workqueue *pwq; 4263 4264 lockdep_assert_held(&wq_pool_mutex); 4265 4266 pool = get_unbound_pool(attrs); 4267 if (!pool) 4268 return NULL; 4269 4270 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4271 if (!pwq) { 4272 put_unbound_pool(pool); 4273 return NULL; 4274 } 4275 4276 init_pwq(pwq, wq, pool); 4277 return pwq; 4278 } 4279 4280 /** 4281 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 4282 * @attrs: the wq_attrs of the default pwq of the target workqueue 4283 * @cpu: the target CPU 4284 * @cpu_going_down: if >= 0, the CPU to consider as offline 4285 * 4286 * Calculate the cpumask a workqueue with @attrs should use on @pod. If 4287 * @cpu_going_down is >= 0, that cpu is considered offline during calculation. 4288 * The result is stored in @attrs->__pod_cpumask. 4289 * 4290 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 4291 * and @pod has online CPUs requested by @attrs, the returned cpumask is the 4292 * intersection of the possible CPUs of @pod and @attrs->cpumask. 4293 * 4294 * The caller is responsible for ensuring that the cpumask of @pod stays stable. 4295 */ 4296 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, 4297 int cpu_going_down) 4298 { 4299 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 4300 int pod = pt->cpu_pod[cpu]; 4301 4302 /* does @pod have any online CPUs @attrs wants? */ 4303 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 4304 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); 4305 if (cpu_going_down >= 0) 4306 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); 4307 4308 if (cpumask_empty(attrs->__pod_cpumask)) { 4309 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 4310 return; 4311 } 4312 4313 /* yeap, return possible CPUs in @pod that @attrs wants */ 4314 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); 4315 4316 if (cpumask_empty(attrs->__pod_cpumask)) 4317 pr_warn_once("WARNING: workqueue cpumask: online intersect > " 4318 "possible intersect\n"); 4319 } 4320 4321 /* install @pwq into @wq's cpu_pwq and return the old pwq */ 4322 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 4323 int cpu, struct pool_workqueue *pwq) 4324 { 4325 struct pool_workqueue *old_pwq; 4326 4327 lockdep_assert_held(&wq_pool_mutex); 4328 lockdep_assert_held(&wq->mutex); 4329 4330 /* link_pwq() can handle duplicate calls */ 4331 link_pwq(pwq); 4332 4333 old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4334 rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq); 4335 return old_pwq; 4336 } 4337 4338 /* context to store the prepared attrs & pwqs before applying */ 4339 struct apply_wqattrs_ctx { 4340 struct workqueue_struct *wq; /* target workqueue */ 4341 struct workqueue_attrs *attrs; /* attrs to apply */ 4342 struct list_head list; /* queued for batching commit */ 4343 struct pool_workqueue *dfl_pwq; 4344 struct pool_workqueue *pwq_tbl[]; 4345 }; 4346 4347 /* free the resources after success or abort */ 4348 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 4349 { 4350 if (ctx) { 4351 int cpu; 4352 4353 for_each_possible_cpu(cpu) 4354 put_pwq_unlocked(ctx->pwq_tbl[cpu]); 4355 put_pwq_unlocked(ctx->dfl_pwq); 4356 4357 free_workqueue_attrs(ctx->attrs); 4358 4359 kfree(ctx); 4360 } 4361 } 4362 4363 /* allocate the attrs and pwqs for later installation */ 4364 static struct apply_wqattrs_ctx * 4365 apply_wqattrs_prepare(struct workqueue_struct *wq, 4366 const struct workqueue_attrs *attrs, 4367 const cpumask_var_t unbound_cpumask) 4368 { 4369 struct apply_wqattrs_ctx *ctx; 4370 struct workqueue_attrs *new_attrs; 4371 int cpu; 4372 4373 lockdep_assert_held(&wq_pool_mutex); 4374 4375 if (WARN_ON(attrs->affn_scope < 0 || 4376 attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 4377 return ERR_PTR(-EINVAL); 4378 4379 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 4380 4381 new_attrs = alloc_workqueue_attrs(); 4382 if (!ctx || !new_attrs) 4383 goto out_free; 4384 4385 /* 4386 * If something goes wrong during CPU up/down, we'll fall back to 4387 * the default pwq covering whole @attrs->cpumask. Always create 4388 * it even if we don't use it immediately. 4389 */ 4390 copy_workqueue_attrs(new_attrs, attrs); 4391 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 4392 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 4393 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 4394 if (!ctx->dfl_pwq) 4395 goto out_free; 4396 4397 for_each_possible_cpu(cpu) { 4398 if (new_attrs->ordered) { 4399 ctx->dfl_pwq->refcnt++; 4400 ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 4401 } else { 4402 wq_calc_pod_cpumask(new_attrs, cpu, -1); 4403 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 4404 if (!ctx->pwq_tbl[cpu]) 4405 goto out_free; 4406 } 4407 } 4408 4409 /* save the user configured attrs and sanitize it. */ 4410 copy_workqueue_attrs(new_attrs, attrs); 4411 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 4412 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 4413 ctx->attrs = new_attrs; 4414 4415 ctx->wq = wq; 4416 return ctx; 4417 4418 out_free: 4419 free_workqueue_attrs(new_attrs); 4420 apply_wqattrs_cleanup(ctx); 4421 return ERR_PTR(-ENOMEM); 4422 } 4423 4424 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 4425 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 4426 { 4427 int cpu; 4428 4429 /* all pwqs have been created successfully, let's install'em */ 4430 mutex_lock(&ctx->wq->mutex); 4431 4432 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 4433 4434 /* save the previous pwq and install the new one */ 4435 for_each_possible_cpu(cpu) 4436 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 4437 ctx->pwq_tbl[cpu]); 4438 4439 /* @dfl_pwq might not have been used, ensure it's linked */ 4440 link_pwq(ctx->dfl_pwq); 4441 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 4442 4443 mutex_unlock(&ctx->wq->mutex); 4444 } 4445 4446 static void apply_wqattrs_lock(void) 4447 { 4448 /* CPUs should stay stable across pwq creations and installations */ 4449 cpus_read_lock(); 4450 mutex_lock(&wq_pool_mutex); 4451 } 4452 4453 static void apply_wqattrs_unlock(void) 4454 { 4455 mutex_unlock(&wq_pool_mutex); 4456 cpus_read_unlock(); 4457 } 4458 4459 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4460 const struct workqueue_attrs *attrs) 4461 { 4462 struct apply_wqattrs_ctx *ctx; 4463 4464 /* only unbound workqueues can change attributes */ 4465 if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4466 return -EINVAL; 4467 4468 /* creating multiple pwqs breaks ordering guarantee */ 4469 if (!list_empty(&wq->pwqs)) { 4470 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4471 return -EINVAL; 4472 4473 wq->flags &= ~__WQ_ORDERED; 4474 } 4475 4476 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 4477 if (IS_ERR(ctx)) 4478 return PTR_ERR(ctx); 4479 4480 /* the ctx has been prepared successfully, let's commit it */ 4481 apply_wqattrs_commit(ctx); 4482 apply_wqattrs_cleanup(ctx); 4483 4484 return 0; 4485 } 4486 4487 /** 4488 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 4489 * @wq: the target workqueue 4490 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 4491 * 4492 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 4493 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 4494 * work items are affine to the pod it was issued on. Older pwqs are released as 4495 * in-flight work items finish. Note that a work item which repeatedly requeues 4496 * itself back-to-back will stay on its current pwq. 4497 * 4498 * Performs GFP_KERNEL allocations. 4499 * 4500 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4501 * 4502 * Return: 0 on success and -errno on failure. 4503 */ 4504 int apply_workqueue_attrs(struct workqueue_struct *wq, 4505 const struct workqueue_attrs *attrs) 4506 { 4507 int ret; 4508 4509 lockdep_assert_cpus_held(); 4510 4511 mutex_lock(&wq_pool_mutex); 4512 ret = apply_workqueue_attrs_locked(wq, attrs); 4513 mutex_unlock(&wq_pool_mutex); 4514 4515 return ret; 4516 } 4517 4518 /** 4519 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug 4520 * @wq: the target workqueue 4521 * @cpu: the CPU to update pool association for 4522 * @hotplug_cpu: the CPU coming up or going down 4523 * @online: whether @cpu is coming up or going down 4524 * 4525 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 4526 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of 4527 * @wq accordingly. 4528 * 4529 * 4530 * If pod affinity can't be adjusted due to memory allocation failure, it falls 4531 * back to @wq->dfl_pwq which may not be optimal but is always correct. 4532 * 4533 * Note that when the last allowed CPU of a pod goes offline for a workqueue 4534 * with a cpumask spanning multiple pods, the workers which were already 4535 * executing the work items for the workqueue will lose their CPU affinity and 4536 * may execute on any CPU. This is similar to how per-cpu workqueues behave on 4537 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 4538 * responsibility to flush the work item from CPU_DOWN_PREPARE. 4539 */ 4540 static void wq_update_pod(struct workqueue_struct *wq, int cpu, 4541 int hotplug_cpu, bool online) 4542 { 4543 int off_cpu = online ? -1 : hotplug_cpu; 4544 struct pool_workqueue *old_pwq = NULL, *pwq; 4545 struct workqueue_attrs *target_attrs; 4546 4547 lockdep_assert_held(&wq_pool_mutex); 4548 4549 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 4550 return; 4551 4552 /* 4553 * We don't wanna alloc/free wq_attrs for each wq for each CPU. 4554 * Let's use a preallocated one. The following buf is protected by 4555 * CPU hotplug exclusion. 4556 */ 4557 target_attrs = wq_update_pod_attrs_buf; 4558 4559 copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 4560 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 4561 4562 /* nothing to do if the target cpumask matches the current pwq */ 4563 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); 4564 pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu), 4565 lockdep_is_held(&wq_pool_mutex)); 4566 if (wqattrs_equal(target_attrs, pwq->pool->attrs)) 4567 return; 4568 4569 /* create a new pwq */ 4570 pwq = alloc_unbound_pwq(wq, target_attrs); 4571 if (!pwq) { 4572 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 4573 wq->name); 4574 goto use_dfl_pwq; 4575 } 4576 4577 /* Install the new pwq. */ 4578 mutex_lock(&wq->mutex); 4579 old_pwq = install_unbound_pwq(wq, cpu, pwq); 4580 goto out_unlock; 4581 4582 use_dfl_pwq: 4583 mutex_lock(&wq->mutex); 4584 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 4585 get_pwq(wq->dfl_pwq); 4586 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 4587 old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq); 4588 out_unlock: 4589 mutex_unlock(&wq->mutex); 4590 put_pwq_unlocked(old_pwq); 4591 } 4592 4593 static int alloc_and_link_pwqs(struct workqueue_struct *wq) 4594 { 4595 bool highpri = wq->flags & WQ_HIGHPRI; 4596 int cpu, ret; 4597 4598 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 4599 if (!wq->cpu_pwq) 4600 goto enomem; 4601 4602 if (!(wq->flags & WQ_UNBOUND)) { 4603 for_each_possible_cpu(cpu) { 4604 struct pool_workqueue **pwq_p = 4605 per_cpu_ptr(wq->cpu_pwq, cpu); 4606 struct worker_pool *pool = 4607 &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); 4608 4609 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 4610 pool->node); 4611 if (!*pwq_p) 4612 goto enomem; 4613 4614 init_pwq(*pwq_p, wq, pool); 4615 4616 mutex_lock(&wq->mutex); 4617 link_pwq(*pwq_p); 4618 mutex_unlock(&wq->mutex); 4619 } 4620 return 0; 4621 } 4622 4623 cpus_read_lock(); 4624 if (wq->flags & __WQ_ORDERED) { 4625 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 4626 /* there should only be single pwq for ordering guarantee */ 4627 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 4628 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 4629 "ordering guarantee broken for workqueue %s\n", wq->name); 4630 } else { 4631 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 4632 } 4633 cpus_read_unlock(); 4634 4635 /* for unbound pwq, flush the pwq_release_worker ensures that the 4636 * pwq_release_workfn() completes before calling kfree(wq). 4637 */ 4638 if (ret) 4639 kthread_flush_worker(pwq_release_worker); 4640 4641 return ret; 4642 4643 enomem: 4644 if (wq->cpu_pwq) { 4645 for_each_possible_cpu(cpu) { 4646 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 4647 4648 if (pwq) 4649 kmem_cache_free(pwq_cache, pwq); 4650 } 4651 free_percpu(wq->cpu_pwq); 4652 wq->cpu_pwq = NULL; 4653 } 4654 return -ENOMEM; 4655 } 4656 4657 static int wq_clamp_max_active(int max_active, unsigned int flags, 4658 const char *name) 4659 { 4660 if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 4661 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4662 max_active, name, 1, WQ_MAX_ACTIVE); 4663 4664 return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 4665 } 4666 4667 /* 4668 * Workqueues which may be used during memory reclaim should have a rescuer 4669 * to guarantee forward progress. 4670 */ 4671 static int init_rescuer(struct workqueue_struct *wq) 4672 { 4673 struct worker *rescuer; 4674 int ret; 4675 4676 if (!(wq->flags & WQ_MEM_RECLAIM)) 4677 return 0; 4678 4679 rescuer = alloc_worker(NUMA_NO_NODE); 4680 if (!rescuer) { 4681 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 4682 wq->name); 4683 return -ENOMEM; 4684 } 4685 4686 rescuer->rescue_wq = wq; 4687 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name); 4688 if (IS_ERR(rescuer->task)) { 4689 ret = PTR_ERR(rescuer->task); 4690 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 4691 wq->name, ERR_PTR(ret)); 4692 kfree(rescuer); 4693 return ret; 4694 } 4695 4696 wq->rescuer = rescuer; 4697 kthread_bind_mask(rescuer->task, cpu_possible_mask); 4698 wake_up_process(rescuer->task); 4699 4700 return 0; 4701 } 4702 4703 /** 4704 * wq_adjust_max_active - update a wq's max_active to the current setting 4705 * @wq: target workqueue 4706 * 4707 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and 4708 * activate inactive work items accordingly. If @wq is freezing, clear 4709 * @wq->max_active to zero. 4710 */ 4711 static void wq_adjust_max_active(struct workqueue_struct *wq) 4712 { 4713 struct pool_workqueue *pwq; 4714 4715 lockdep_assert_held(&wq->mutex); 4716 4717 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { 4718 WRITE_ONCE(wq->max_active, 0); 4719 return; 4720 } 4721 4722 if (wq->max_active == wq->saved_max_active) 4723 return; 4724 4725 /* 4726 * Update @wq->max_active and then kick inactive work items if more 4727 * active work items are allowed. This doesn't break work item ordering 4728 * because new work items are always queued behind existing inactive 4729 * work items if there are any. 4730 */ 4731 WRITE_ONCE(wq->max_active, wq->saved_max_active); 4732 4733 for_each_pwq(pwq, wq) { 4734 unsigned long flags; 4735 4736 /* this function can be called during early boot w/ irq disabled */ 4737 raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4738 4739 while (pwq_activate_first_inactive(pwq)) 4740 ; 4741 4742 kick_pool(pwq->pool); 4743 4744 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 4745 } 4746 } 4747 4748 __printf(1, 4) 4749 struct workqueue_struct *alloc_workqueue(const char *fmt, 4750 unsigned int flags, 4751 int max_active, ...) 4752 { 4753 va_list args; 4754 struct workqueue_struct *wq; 4755 int len; 4756 4757 /* 4758 * Unbound && max_active == 1 used to imply ordered, which is no longer 4759 * the case on many machines due to per-pod pools. While 4760 * alloc_ordered_workqueue() is the right way to create an ordered 4761 * workqueue, keep the previous behavior to avoid subtle breakages. 4762 */ 4763 if ((flags & WQ_UNBOUND) && max_active == 1) 4764 flags |= __WQ_ORDERED; 4765 4766 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4767 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4768 flags |= WQ_UNBOUND; 4769 4770 /* allocate wq and format name */ 4771 wq = kzalloc(sizeof(*wq), GFP_KERNEL); 4772 if (!wq) 4773 return NULL; 4774 4775 if (flags & WQ_UNBOUND) { 4776 wq->unbound_attrs = alloc_workqueue_attrs(); 4777 if (!wq->unbound_attrs) 4778 goto err_free_wq; 4779 } 4780 4781 va_start(args, max_active); 4782 len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4783 va_end(args); 4784 4785 if (len >= WQ_NAME_LEN) 4786 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", wq->name); 4787 4788 max_active = max_active ?: WQ_DFL_ACTIVE; 4789 max_active = wq_clamp_max_active(max_active, flags, wq->name); 4790 4791 /* init wq */ 4792 wq->flags = flags; 4793 wq->max_active = max_active; 4794 wq->saved_max_active = max_active; 4795 mutex_init(&wq->mutex); 4796 atomic_set(&wq->nr_pwqs_to_flush, 0); 4797 INIT_LIST_HEAD(&wq->pwqs); 4798 INIT_LIST_HEAD(&wq->flusher_queue); 4799 INIT_LIST_HEAD(&wq->flusher_overflow); 4800 INIT_LIST_HEAD(&wq->maydays); 4801 4802 wq_init_lockdep(wq); 4803 INIT_LIST_HEAD(&wq->list); 4804 4805 if (alloc_and_link_pwqs(wq) < 0) 4806 goto err_unreg_lockdep; 4807 4808 if (wq_online && init_rescuer(wq) < 0) 4809 goto err_destroy; 4810 4811 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4812 goto err_destroy; 4813 4814 /* 4815 * wq_pool_mutex protects global freeze state and workqueues list. 4816 * Grab it, adjust max_active and add the new @wq to workqueues 4817 * list. 4818 */ 4819 mutex_lock(&wq_pool_mutex); 4820 4821 mutex_lock(&wq->mutex); 4822 wq_adjust_max_active(wq); 4823 mutex_unlock(&wq->mutex); 4824 4825 list_add_tail_rcu(&wq->list, &workqueues); 4826 4827 mutex_unlock(&wq_pool_mutex); 4828 4829 return wq; 4830 4831 err_unreg_lockdep: 4832 wq_unregister_lockdep(wq); 4833 wq_free_lockdep(wq); 4834 err_free_wq: 4835 free_workqueue_attrs(wq->unbound_attrs); 4836 kfree(wq); 4837 return NULL; 4838 err_destroy: 4839 destroy_workqueue(wq); 4840 return NULL; 4841 } 4842 EXPORT_SYMBOL_GPL(alloc_workqueue); 4843 4844 static bool pwq_busy(struct pool_workqueue *pwq) 4845 { 4846 int i; 4847 4848 for (i = 0; i < WORK_NR_COLORS; i++) 4849 if (pwq->nr_in_flight[i]) 4850 return true; 4851 4852 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4853 return true; 4854 if (!pwq_is_empty(pwq)) 4855 return true; 4856 4857 return false; 4858 } 4859 4860 /** 4861 * destroy_workqueue - safely terminate a workqueue 4862 * @wq: target workqueue 4863 * 4864 * Safely destroy a workqueue. All work currently pending will be done first. 4865 */ 4866 void destroy_workqueue(struct workqueue_struct *wq) 4867 { 4868 struct pool_workqueue *pwq; 4869 int cpu; 4870 4871 /* 4872 * Remove it from sysfs first so that sanity check failure doesn't 4873 * lead to sysfs name conflicts. 4874 */ 4875 workqueue_sysfs_unregister(wq); 4876 4877 /* mark the workqueue destruction is in progress */ 4878 mutex_lock(&wq->mutex); 4879 wq->flags |= __WQ_DESTROYING; 4880 mutex_unlock(&wq->mutex); 4881 4882 /* drain it before proceeding with destruction */ 4883 drain_workqueue(wq); 4884 4885 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4886 if (wq->rescuer) { 4887 struct worker *rescuer = wq->rescuer; 4888 4889 /* this prevents new queueing */ 4890 raw_spin_lock_irq(&wq_mayday_lock); 4891 wq->rescuer = NULL; 4892 raw_spin_unlock_irq(&wq_mayday_lock); 4893 4894 /* rescuer will empty maydays list before exiting */ 4895 kthread_stop(rescuer->task); 4896 kfree(rescuer); 4897 } 4898 4899 /* 4900 * Sanity checks - grab all the locks so that we wait for all 4901 * in-flight operations which may do put_pwq(). 4902 */ 4903 mutex_lock(&wq_pool_mutex); 4904 mutex_lock(&wq->mutex); 4905 for_each_pwq(pwq, wq) { 4906 raw_spin_lock_irq(&pwq->pool->lock); 4907 if (WARN_ON(pwq_busy(pwq))) { 4908 pr_warn("%s: %s has the following busy pwq\n", 4909 __func__, wq->name); 4910 show_pwq(pwq); 4911 raw_spin_unlock_irq(&pwq->pool->lock); 4912 mutex_unlock(&wq->mutex); 4913 mutex_unlock(&wq_pool_mutex); 4914 show_one_workqueue(wq); 4915 return; 4916 } 4917 raw_spin_unlock_irq(&pwq->pool->lock); 4918 } 4919 mutex_unlock(&wq->mutex); 4920 4921 /* 4922 * wq list is used to freeze wq, remove from list after 4923 * flushing is complete in case freeze races us. 4924 */ 4925 list_del_rcu(&wq->list); 4926 mutex_unlock(&wq_pool_mutex); 4927 4928 /* 4929 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 4930 * to put the base refs. @wq will be auto-destroyed from the last 4931 * pwq_put. RCU read lock prevents @wq from going away from under us. 4932 */ 4933 rcu_read_lock(); 4934 4935 for_each_possible_cpu(cpu) { 4936 pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4937 RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL); 4938 put_pwq_unlocked(pwq); 4939 } 4940 4941 put_pwq_unlocked(wq->dfl_pwq); 4942 wq->dfl_pwq = NULL; 4943 4944 rcu_read_unlock(); 4945 } 4946 EXPORT_SYMBOL_GPL(destroy_workqueue); 4947 4948 /** 4949 * workqueue_set_max_active - adjust max_active of a workqueue 4950 * @wq: target workqueue 4951 * @max_active: new max_active value. 4952 * 4953 * Set max_active of @wq to @max_active. 4954 * 4955 * CONTEXT: 4956 * Don't call from IRQ context. 4957 */ 4958 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4959 { 4960 /* disallow meddling with max_active for ordered workqueues */ 4961 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4962 return; 4963 4964 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4965 4966 mutex_lock(&wq->mutex); 4967 4968 wq->flags &= ~__WQ_ORDERED; 4969 wq->saved_max_active = max_active; 4970 wq_adjust_max_active(wq); 4971 4972 mutex_unlock(&wq->mutex); 4973 } 4974 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4975 4976 /** 4977 * current_work - retrieve %current task's work struct 4978 * 4979 * Determine if %current task is a workqueue worker and what it's working on. 4980 * Useful to find out the context that the %current task is running in. 4981 * 4982 * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 4983 */ 4984 struct work_struct *current_work(void) 4985 { 4986 struct worker *worker = current_wq_worker(); 4987 4988 return worker ? worker->current_work : NULL; 4989 } 4990 EXPORT_SYMBOL(current_work); 4991 4992 /** 4993 * current_is_workqueue_rescuer - is %current workqueue rescuer? 4994 * 4995 * Determine whether %current is a workqueue rescuer. Can be used from 4996 * work functions to determine whether it's being run off the rescuer task. 4997 * 4998 * Return: %true if %current is a workqueue rescuer. %false otherwise. 4999 */ 5000 bool current_is_workqueue_rescuer(void) 5001 { 5002 struct worker *worker = current_wq_worker(); 5003 5004 return worker && worker->rescue_wq; 5005 } 5006 5007 /** 5008 * workqueue_congested - test whether a workqueue is congested 5009 * @cpu: CPU in question 5010 * @wq: target workqueue 5011 * 5012 * Test whether @wq's cpu workqueue for @cpu is congested. There is 5013 * no synchronization around this function and the test result is 5014 * unreliable and only useful as advisory hints or for debugging. 5015 * 5016 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 5017 * 5018 * With the exception of ordered workqueues, all workqueues have per-cpu 5019 * pool_workqueues, each with its own congested state. A workqueue being 5020 * congested on one CPU doesn't mean that the workqueue is contested on any 5021 * other CPUs. 5022 * 5023 * Return: 5024 * %true if congested, %false otherwise. 5025 */ 5026 bool workqueue_congested(int cpu, struct workqueue_struct *wq) 5027 { 5028 struct pool_workqueue *pwq; 5029 bool ret; 5030 5031 rcu_read_lock(); 5032 preempt_disable(); 5033 5034 if (cpu == WORK_CPU_UNBOUND) 5035 cpu = smp_processor_id(); 5036 5037 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 5038 ret = !list_empty(&pwq->inactive_works); 5039 5040 preempt_enable(); 5041 rcu_read_unlock(); 5042 5043 return ret; 5044 } 5045 EXPORT_SYMBOL_GPL(workqueue_congested); 5046 5047 /** 5048 * work_busy - test whether a work is currently pending or running 5049 * @work: the work to be tested 5050 * 5051 * Test whether @work is currently pending or running. There is no 5052 * synchronization around this function and the test result is 5053 * unreliable and only useful as advisory hints or for debugging. 5054 * 5055 * Return: 5056 * OR'd bitmask of WORK_BUSY_* bits. 5057 */ 5058 unsigned int work_busy(struct work_struct *work) 5059 { 5060 struct worker_pool *pool; 5061 unsigned long flags; 5062 unsigned int ret = 0; 5063 5064 if (work_pending(work)) 5065 ret |= WORK_BUSY_PENDING; 5066 5067 rcu_read_lock(); 5068 pool = get_work_pool(work); 5069 if (pool) { 5070 raw_spin_lock_irqsave(&pool->lock, flags); 5071 if (find_worker_executing_work(pool, work)) 5072 ret |= WORK_BUSY_RUNNING; 5073 raw_spin_unlock_irqrestore(&pool->lock, flags); 5074 } 5075 rcu_read_unlock(); 5076 5077 return ret; 5078 } 5079 EXPORT_SYMBOL_GPL(work_busy); 5080 5081 /** 5082 * set_worker_desc - set description for the current work item 5083 * @fmt: printf-style format string 5084 * @...: arguments for the format string 5085 * 5086 * This function can be called by a running work function to describe what 5087 * the work item is about. If the worker task gets dumped, this 5088 * information will be printed out together to help debugging. The 5089 * description can be at most WORKER_DESC_LEN including the trailing '\0'. 5090 */ 5091 void set_worker_desc(const char *fmt, ...) 5092 { 5093 struct worker *worker = current_wq_worker(); 5094 va_list args; 5095 5096 if (worker) { 5097 va_start(args, fmt); 5098 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 5099 va_end(args); 5100 } 5101 } 5102 EXPORT_SYMBOL_GPL(set_worker_desc); 5103 5104 /** 5105 * print_worker_info - print out worker information and description 5106 * @log_lvl: the log level to use when printing 5107 * @task: target task 5108 * 5109 * If @task is a worker and currently executing a work item, print out the 5110 * name of the workqueue being serviced and worker description set with 5111 * set_worker_desc() by the currently executing work item. 5112 * 5113 * This function can be safely called on any task as long as the 5114 * task_struct itself is accessible. While safe, this function isn't 5115 * synchronized and may print out mixups or garbages of limited length. 5116 */ 5117 void print_worker_info(const char *log_lvl, struct task_struct *task) 5118 { 5119 work_func_t *fn = NULL; 5120 char name[WQ_NAME_LEN] = { }; 5121 char desc[WORKER_DESC_LEN] = { }; 5122 struct pool_workqueue *pwq = NULL; 5123 struct workqueue_struct *wq = NULL; 5124 struct worker *worker; 5125 5126 if (!(task->flags & PF_WQ_WORKER)) 5127 return; 5128 5129 /* 5130 * This function is called without any synchronization and @task 5131 * could be in any state. Be careful with dereferences. 5132 */ 5133 worker = kthread_probe_data(task); 5134 5135 /* 5136 * Carefully copy the associated workqueue's workfn, name and desc. 5137 * Keep the original last '\0' in case the original is garbage. 5138 */ 5139 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5140 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5141 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5142 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5143 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 5144 5145 if (fn || name[0] || desc[0]) { 5146 printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 5147 if (strcmp(name, desc)) 5148 pr_cont(" (%s)", desc); 5149 pr_cont("\n"); 5150 } 5151 } 5152 5153 static void pr_cont_pool_info(struct worker_pool *pool) 5154 { 5155 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 5156 if (pool->node != NUMA_NO_NODE) 5157 pr_cont(" node=%d", pool->node); 5158 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 5159 } 5160 5161 struct pr_cont_work_struct { 5162 bool comma; 5163 work_func_t func; 5164 long ctr; 5165 }; 5166 5167 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5168 { 5169 if (!pcwsp->ctr) 5170 goto out_record; 5171 if (func == pcwsp->func) { 5172 pcwsp->ctr++; 5173 return; 5174 } 5175 if (pcwsp->ctr == 1) 5176 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5177 else 5178 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5179 pcwsp->ctr = 0; 5180 out_record: 5181 if ((long)func == -1L) 5182 return; 5183 pcwsp->comma = comma; 5184 pcwsp->func = func; 5185 pcwsp->ctr = 1; 5186 } 5187 5188 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 5189 { 5190 if (work->func == wq_barrier_func) { 5191 struct wq_barrier *barr; 5192 5193 barr = container_of(work, struct wq_barrier, work); 5194 5195 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5196 pr_cont("%s BAR(%d)", comma ? "," : "", 5197 task_pid_nr(barr->task)); 5198 } else { 5199 if (!comma) 5200 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5201 pr_cont_work_flush(comma, work->func, pcwsp); 5202 } 5203 } 5204 5205 static void show_pwq(struct pool_workqueue *pwq) 5206 { 5207 struct pr_cont_work_struct pcws = { .ctr = 0, }; 5208 struct worker_pool *pool = pwq->pool; 5209 struct work_struct *work; 5210 struct worker *worker; 5211 bool has_in_flight = false, has_pending = false; 5212 int bkt; 5213 5214 pr_info(" pwq %d:", pool->id); 5215 pr_cont_pool_info(pool); 5216 5217 pr_cont(" active=%d refcnt=%d%s\n", 5218 pwq->nr_active, pwq->refcnt, 5219 !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 5220 5221 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 5222 if (worker->current_pwq == pwq) { 5223 has_in_flight = true; 5224 break; 5225 } 5226 } 5227 if (has_in_flight) { 5228 bool comma = false; 5229 5230 pr_info(" in-flight:"); 5231 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 5232 if (worker->current_pwq != pwq) 5233 continue; 5234 5235 pr_cont("%s %d%s:%ps", comma ? "," : "", 5236 task_pid_nr(worker->task), 5237 worker->rescue_wq ? "(RESCUER)" : "", 5238 worker->current_func); 5239 list_for_each_entry(work, &worker->scheduled, entry) 5240 pr_cont_work(false, work, &pcws); 5241 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 5242 comma = true; 5243 } 5244 pr_cont("\n"); 5245 } 5246 5247 list_for_each_entry(work, &pool->worklist, entry) { 5248 if (get_work_pwq(work) == pwq) { 5249 has_pending = true; 5250 break; 5251 } 5252 } 5253 if (has_pending) { 5254 bool comma = false; 5255 5256 pr_info(" pending:"); 5257 list_for_each_entry(work, &pool->worklist, entry) { 5258 if (get_work_pwq(work) != pwq) 5259 continue; 5260 5261 pr_cont_work(comma, work, &pcws); 5262 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 5263 } 5264 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 5265 pr_cont("\n"); 5266 } 5267 5268 if (!list_empty(&pwq->inactive_works)) { 5269 bool comma = false; 5270 5271 pr_info(" inactive:"); 5272 list_for_each_entry(work, &pwq->inactive_works, entry) { 5273 pr_cont_work(comma, work, &pcws); 5274 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 5275 } 5276 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 5277 pr_cont("\n"); 5278 } 5279 } 5280 5281 /** 5282 * show_one_workqueue - dump state of specified workqueue 5283 * @wq: workqueue whose state will be printed 5284 */ 5285 void show_one_workqueue(struct workqueue_struct *wq) 5286 { 5287 struct pool_workqueue *pwq; 5288 bool idle = true; 5289 unsigned long flags; 5290 5291 for_each_pwq(pwq, wq) { 5292 if (!pwq_is_empty(pwq)) { 5293 idle = false; 5294 break; 5295 } 5296 } 5297 if (idle) /* Nothing to print for idle workqueue */ 5298 return; 5299 5300 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 5301 5302 for_each_pwq(pwq, wq) { 5303 raw_spin_lock_irqsave(&pwq->pool->lock, flags); 5304 if (!pwq_is_empty(pwq)) { 5305 /* 5306 * Defer printing to avoid deadlocks in console 5307 * drivers that queue work while holding locks 5308 * also taken in their write paths. 5309 */ 5310 printk_deferred_enter(); 5311 show_pwq(pwq); 5312 printk_deferred_exit(); 5313 } 5314 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 5315 /* 5316 * We could be printing a lot from atomic context, e.g. 5317 * sysrq-t -> show_all_workqueues(). Avoid triggering 5318 * hard lockup. 5319 */ 5320 touch_nmi_watchdog(); 5321 } 5322 5323 } 5324 5325 /** 5326 * show_one_worker_pool - dump state of specified worker pool 5327 * @pool: worker pool whose state will be printed 5328 */ 5329 static void show_one_worker_pool(struct worker_pool *pool) 5330 { 5331 struct worker *worker; 5332 bool first = true; 5333 unsigned long flags; 5334 unsigned long hung = 0; 5335 5336 raw_spin_lock_irqsave(&pool->lock, flags); 5337 if (pool->nr_workers == pool->nr_idle) 5338 goto next_pool; 5339 5340 /* How long the first pending work is waiting for a worker. */ 5341 if (!list_empty(&pool->worklist)) 5342 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5343 5344 /* 5345 * Defer printing to avoid deadlocks in console drivers that 5346 * queue work while holding locks also taken in their write 5347 * paths. 5348 */ 5349 printk_deferred_enter(); 5350 pr_info("pool %d:", pool->id); 5351 pr_cont_pool_info(pool); 5352 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 5353 if (pool->manager) 5354 pr_cont(" manager: %d", 5355 task_pid_nr(pool->manager->task)); 5356 list_for_each_entry(worker, &pool->idle_list, entry) { 5357 pr_cont(" %s%d", first ? "idle: " : "", 5358 task_pid_nr(worker->task)); 5359 first = false; 5360 } 5361 pr_cont("\n"); 5362 printk_deferred_exit(); 5363 next_pool: 5364 raw_spin_unlock_irqrestore(&pool->lock, flags); 5365 /* 5366 * We could be printing a lot from atomic context, e.g. 5367 * sysrq-t -> show_all_workqueues(). Avoid triggering 5368 * hard lockup. 5369 */ 5370 touch_nmi_watchdog(); 5371 5372 } 5373 5374 /** 5375 * show_all_workqueues - dump workqueue state 5376 * 5377 * Called from a sysrq handler and prints out all busy workqueues and pools. 5378 */ 5379 void show_all_workqueues(void) 5380 { 5381 struct workqueue_struct *wq; 5382 struct worker_pool *pool; 5383 int pi; 5384 5385 rcu_read_lock(); 5386 5387 pr_info("Showing busy workqueues and worker pools:\n"); 5388 5389 list_for_each_entry_rcu(wq, &workqueues, list) 5390 show_one_workqueue(wq); 5391 5392 for_each_pool(pool, pi) 5393 show_one_worker_pool(pool); 5394 5395 rcu_read_unlock(); 5396 } 5397 5398 /** 5399 * show_freezable_workqueues - dump freezable workqueue state 5400 * 5401 * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5402 * still busy. 5403 */ 5404 void show_freezable_workqueues(void) 5405 { 5406 struct workqueue_struct *wq; 5407 5408 rcu_read_lock(); 5409 5410 pr_info("Showing freezable workqueues that are still busy:\n"); 5411 5412 list_for_each_entry_rcu(wq, &workqueues, list) { 5413 if (!(wq->flags & WQ_FREEZABLE)) 5414 continue; 5415 show_one_workqueue(wq); 5416 } 5417 5418 rcu_read_unlock(); 5419 } 5420 5421 /* used to show worker information through /proc/PID/{comm,stat,status} */ 5422 void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 5423 { 5424 int off; 5425 5426 /* always show the actual comm */ 5427 off = strscpy(buf, task->comm, size); 5428 if (off < 0) 5429 return; 5430 5431 /* stabilize PF_WQ_WORKER and worker pool association */ 5432 mutex_lock(&wq_pool_attach_mutex); 5433 5434 if (task->flags & PF_WQ_WORKER) { 5435 struct worker *worker = kthread_data(task); 5436 struct worker_pool *pool = worker->pool; 5437 5438 if (pool) { 5439 raw_spin_lock_irq(&pool->lock); 5440 /* 5441 * ->desc tracks information (wq name or 5442 * set_worker_desc()) for the latest execution. If 5443 * current, prepend '+', otherwise '-'. 5444 */ 5445 if (worker->desc[0] != '\0') { 5446 if (worker->current_work) 5447 scnprintf(buf + off, size - off, "+%s", 5448 worker->desc); 5449 else 5450 scnprintf(buf + off, size - off, "-%s", 5451 worker->desc); 5452 } 5453 raw_spin_unlock_irq(&pool->lock); 5454 } 5455 } 5456 5457 mutex_unlock(&wq_pool_attach_mutex); 5458 } 5459 5460 #ifdef CONFIG_SMP 5461 5462 /* 5463 * CPU hotplug. 5464 * 5465 * There are two challenges in supporting CPU hotplug. Firstly, there 5466 * are a lot of assumptions on strong associations among work, pwq and 5467 * pool which make migrating pending and scheduled works very 5468 * difficult to implement without impacting hot paths. Secondly, 5469 * worker pools serve mix of short, long and very long running works making 5470 * blocked draining impractical. 5471 * 5472 * This is solved by allowing the pools to be disassociated from the CPU 5473 * running as an unbound one and allowing it to be reattached later if the 5474 * cpu comes back online. 5475 */ 5476 5477 static void unbind_workers(int cpu) 5478 { 5479 struct worker_pool *pool; 5480 struct worker *worker; 5481 5482 for_each_cpu_worker_pool(pool, cpu) { 5483 mutex_lock(&wq_pool_attach_mutex); 5484 raw_spin_lock_irq(&pool->lock); 5485 5486 /* 5487 * We've blocked all attach/detach operations. Make all workers 5488 * unbound and set DISASSOCIATED. Before this, all workers 5489 * must be on the cpu. After this, they may become diasporas. 5490 * And the preemption disabled section in their sched callbacks 5491 * are guaranteed to see WORKER_UNBOUND since the code here 5492 * is on the same cpu. 5493 */ 5494 for_each_pool_worker(worker, pool) 5495 worker->flags |= WORKER_UNBOUND; 5496 5497 pool->flags |= POOL_DISASSOCIATED; 5498 5499 /* 5500 * The handling of nr_running in sched callbacks are disabled 5501 * now. Zap nr_running. After this, nr_running stays zero and 5502 * need_more_worker() and keep_working() are always true as 5503 * long as the worklist is not empty. This pool now behaves as 5504 * an unbound (in terms of concurrency management) pool which 5505 * are served by workers tied to the pool. 5506 */ 5507 pool->nr_running = 0; 5508 5509 /* 5510 * With concurrency management just turned off, a busy 5511 * worker blocking could lead to lengthy stalls. Kick off 5512 * unbound chain execution of currently pending work items. 5513 */ 5514 kick_pool(pool); 5515 5516 raw_spin_unlock_irq(&pool->lock); 5517 5518 for_each_pool_worker(worker, pool) 5519 unbind_worker(worker); 5520 5521 mutex_unlock(&wq_pool_attach_mutex); 5522 } 5523 } 5524 5525 /** 5526 * rebind_workers - rebind all workers of a pool to the associated CPU 5527 * @pool: pool of interest 5528 * 5529 * @pool->cpu is coming online. Rebind all workers to the CPU. 5530 */ 5531 static void rebind_workers(struct worker_pool *pool) 5532 { 5533 struct worker *worker; 5534 5535 lockdep_assert_held(&wq_pool_attach_mutex); 5536 5537 /* 5538 * Restore CPU affinity of all workers. As all idle workers should 5539 * be on the run-queue of the associated CPU before any local 5540 * wake-ups for concurrency management happen, restore CPU affinity 5541 * of all workers first and then clear UNBOUND. As we're called 5542 * from CPU_ONLINE, the following shouldn't fail. 5543 */ 5544 for_each_pool_worker(worker, pool) { 5545 kthread_set_per_cpu(worker->task, pool->cpu); 5546 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5547 pool_allowed_cpus(pool)) < 0); 5548 } 5549 5550 raw_spin_lock_irq(&pool->lock); 5551 5552 pool->flags &= ~POOL_DISASSOCIATED; 5553 5554 for_each_pool_worker(worker, pool) { 5555 unsigned int worker_flags = worker->flags; 5556 5557 /* 5558 * We want to clear UNBOUND but can't directly call 5559 * worker_clr_flags() or adjust nr_running. Atomically 5560 * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5561 * @worker will clear REBOUND using worker_clr_flags() when 5562 * it initiates the next execution cycle thus restoring 5563 * concurrency management. Note that when or whether 5564 * @worker clears REBOUND doesn't affect correctness. 5565 * 5566 * WRITE_ONCE() is necessary because @worker->flags may be 5567 * tested without holding any lock in 5568 * wq_worker_running(). Without it, NOT_RUNNING test may 5569 * fail incorrectly leading to premature concurrency 5570 * management operations. 5571 */ 5572 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5573 worker_flags |= WORKER_REBOUND; 5574 worker_flags &= ~WORKER_UNBOUND; 5575 WRITE_ONCE(worker->flags, worker_flags); 5576 } 5577 5578 raw_spin_unlock_irq(&pool->lock); 5579 } 5580 5581 /** 5582 * restore_unbound_workers_cpumask - restore cpumask of unbound workers 5583 * @pool: unbound pool of interest 5584 * @cpu: the CPU which is coming up 5585 * 5586 * An unbound pool may end up with a cpumask which doesn't have any online 5587 * CPUs. When a worker of such pool get scheduled, the scheduler resets 5588 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 5589 * online CPU before, cpus_allowed of all its workers should be restored. 5590 */ 5591 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 5592 { 5593 static cpumask_t cpumask; 5594 struct worker *worker; 5595 5596 lockdep_assert_held(&wq_pool_attach_mutex); 5597 5598 /* is @cpu allowed for @pool? */ 5599 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 5600 return; 5601 5602 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 5603 5604 /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5605 for_each_pool_worker(worker, pool) 5606 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 5607 } 5608 5609 int workqueue_prepare_cpu(unsigned int cpu) 5610 { 5611 struct worker_pool *pool; 5612 5613 for_each_cpu_worker_pool(pool, cpu) { 5614 if (pool->nr_workers) 5615 continue; 5616 if (!create_worker(pool)) 5617 return -ENOMEM; 5618 } 5619 return 0; 5620 } 5621 5622 int workqueue_online_cpu(unsigned int cpu) 5623 { 5624 struct worker_pool *pool; 5625 struct workqueue_struct *wq; 5626 int pi; 5627 5628 mutex_lock(&wq_pool_mutex); 5629 5630 for_each_pool(pool, pi) { 5631 mutex_lock(&wq_pool_attach_mutex); 5632 5633 if (pool->cpu == cpu) 5634 rebind_workers(pool); 5635 else if (pool->cpu < 0) 5636 restore_unbound_workers_cpumask(pool, cpu); 5637 5638 mutex_unlock(&wq_pool_attach_mutex); 5639 } 5640 5641 /* update pod affinity of unbound workqueues */ 5642 list_for_each_entry(wq, &workqueues, list) { 5643 struct workqueue_attrs *attrs = wq->unbound_attrs; 5644 5645 if (attrs) { 5646 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5647 int tcpu; 5648 5649 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5650 wq_update_pod(wq, tcpu, cpu, true); 5651 } 5652 } 5653 5654 mutex_unlock(&wq_pool_mutex); 5655 return 0; 5656 } 5657 5658 int workqueue_offline_cpu(unsigned int cpu) 5659 { 5660 struct workqueue_struct *wq; 5661 5662 /* unbinding per-cpu workers should happen on the local CPU */ 5663 if (WARN_ON(cpu != smp_processor_id())) 5664 return -1; 5665 5666 unbind_workers(cpu); 5667 5668 /* update pod affinity of unbound workqueues */ 5669 mutex_lock(&wq_pool_mutex); 5670 list_for_each_entry(wq, &workqueues, list) { 5671 struct workqueue_attrs *attrs = wq->unbound_attrs; 5672 5673 if (attrs) { 5674 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5675 int tcpu; 5676 5677 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5678 wq_update_pod(wq, tcpu, cpu, false); 5679 } 5680 } 5681 mutex_unlock(&wq_pool_mutex); 5682 5683 return 0; 5684 } 5685 5686 struct work_for_cpu { 5687 struct work_struct work; 5688 long (*fn)(void *); 5689 void *arg; 5690 long ret; 5691 }; 5692 5693 static void work_for_cpu_fn(struct work_struct *work) 5694 { 5695 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5696 5697 wfc->ret = wfc->fn(wfc->arg); 5698 } 5699 5700 /** 5701 * work_on_cpu_key - run a function in thread context on a particular cpu 5702 * @cpu: the cpu to run on 5703 * @fn: the function to run 5704 * @arg: the function arg 5705 * @key: The lock class key for lock debugging purposes 5706 * 5707 * It is up to the caller to ensure that the cpu doesn't go offline. 5708 * The caller must not hold any locks which would prevent @fn from completing. 5709 * 5710 * Return: The value @fn returns. 5711 */ 5712 long work_on_cpu_key(int cpu, long (*fn)(void *), 5713 void *arg, struct lock_class_key *key) 5714 { 5715 struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 5716 5717 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); 5718 schedule_work_on(cpu, &wfc.work); 5719 flush_work(&wfc.work); 5720 destroy_work_on_stack(&wfc.work); 5721 return wfc.ret; 5722 } 5723 EXPORT_SYMBOL_GPL(work_on_cpu_key); 5724 5725 /** 5726 * work_on_cpu_safe_key - run a function in thread context on a particular cpu 5727 * @cpu: the cpu to run on 5728 * @fn: the function to run 5729 * @arg: the function argument 5730 * @key: The lock class key for lock debugging purposes 5731 * 5732 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 5733 * any locks which would prevent @fn from completing. 5734 * 5735 * Return: The value @fn returns. 5736 */ 5737 long work_on_cpu_safe_key(int cpu, long (*fn)(void *), 5738 void *arg, struct lock_class_key *key) 5739 { 5740 long ret = -ENODEV; 5741 5742 cpus_read_lock(); 5743 if (cpu_online(cpu)) 5744 ret = work_on_cpu_key(cpu, fn, arg, key); 5745 cpus_read_unlock(); 5746 return ret; 5747 } 5748 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); 5749 #endif /* CONFIG_SMP */ 5750 5751 #ifdef CONFIG_FREEZER 5752 5753 /** 5754 * freeze_workqueues_begin - begin freezing workqueues 5755 * 5756 * Start freezing workqueues. After this function returns, all freezable 5757 * workqueues will queue new works to their inactive_works list instead of 5758 * pool->worklist. 5759 * 5760 * CONTEXT: 5761 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5762 */ 5763 void freeze_workqueues_begin(void) 5764 { 5765 struct workqueue_struct *wq; 5766 5767 mutex_lock(&wq_pool_mutex); 5768 5769 WARN_ON_ONCE(workqueue_freezing); 5770 workqueue_freezing = true; 5771 5772 list_for_each_entry(wq, &workqueues, list) { 5773 mutex_lock(&wq->mutex); 5774 wq_adjust_max_active(wq); 5775 mutex_unlock(&wq->mutex); 5776 } 5777 5778 mutex_unlock(&wq_pool_mutex); 5779 } 5780 5781 /** 5782 * freeze_workqueues_busy - are freezable workqueues still busy? 5783 * 5784 * Check whether freezing is complete. This function must be called 5785 * between freeze_workqueues_begin() and thaw_workqueues(). 5786 * 5787 * CONTEXT: 5788 * Grabs and releases wq_pool_mutex. 5789 * 5790 * Return: 5791 * %true if some freezable workqueues are still busy. %false if freezing 5792 * is complete. 5793 */ 5794 bool freeze_workqueues_busy(void) 5795 { 5796 bool busy = false; 5797 struct workqueue_struct *wq; 5798 struct pool_workqueue *pwq; 5799 5800 mutex_lock(&wq_pool_mutex); 5801 5802 WARN_ON_ONCE(!workqueue_freezing); 5803 5804 list_for_each_entry(wq, &workqueues, list) { 5805 if (!(wq->flags & WQ_FREEZABLE)) 5806 continue; 5807 /* 5808 * nr_active is monotonically decreasing. It's safe 5809 * to peek without lock. 5810 */ 5811 rcu_read_lock(); 5812 for_each_pwq(pwq, wq) { 5813 WARN_ON_ONCE(pwq->nr_active < 0); 5814 if (pwq->nr_active) { 5815 busy = true; 5816 rcu_read_unlock(); 5817 goto out_unlock; 5818 } 5819 } 5820 rcu_read_unlock(); 5821 } 5822 out_unlock: 5823 mutex_unlock(&wq_pool_mutex); 5824 return busy; 5825 } 5826 5827 /** 5828 * thaw_workqueues - thaw workqueues 5829 * 5830 * Thaw workqueues. Normal queueing is restored and all collected 5831 * frozen works are transferred to their respective pool worklists. 5832 * 5833 * CONTEXT: 5834 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5835 */ 5836 void thaw_workqueues(void) 5837 { 5838 struct workqueue_struct *wq; 5839 5840 mutex_lock(&wq_pool_mutex); 5841 5842 if (!workqueue_freezing) 5843 goto out_unlock; 5844 5845 workqueue_freezing = false; 5846 5847 /* restore max_active and repopulate worklist */ 5848 list_for_each_entry(wq, &workqueues, list) { 5849 mutex_lock(&wq->mutex); 5850 wq_adjust_max_active(wq); 5851 mutex_unlock(&wq->mutex); 5852 } 5853 5854 out_unlock: 5855 mutex_unlock(&wq_pool_mutex); 5856 } 5857 #endif /* CONFIG_FREEZER */ 5858 5859 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5860 { 5861 LIST_HEAD(ctxs); 5862 int ret = 0; 5863 struct workqueue_struct *wq; 5864 struct apply_wqattrs_ctx *ctx, *n; 5865 5866 lockdep_assert_held(&wq_pool_mutex); 5867 5868 list_for_each_entry(wq, &workqueues, list) { 5869 if (!(wq->flags & WQ_UNBOUND)) 5870 continue; 5871 /* creating multiple pwqs breaks ordering guarantee */ 5872 if (wq->flags & __WQ_ORDERED) 5873 continue; 5874 5875 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5876 if (IS_ERR(ctx)) { 5877 ret = PTR_ERR(ctx); 5878 break; 5879 } 5880 5881 list_add_tail(&ctx->list, &ctxs); 5882 } 5883 5884 list_for_each_entry_safe(ctx, n, &ctxs, list) { 5885 if (!ret) 5886 apply_wqattrs_commit(ctx); 5887 apply_wqattrs_cleanup(ctx); 5888 } 5889 5890 if (!ret) { 5891 mutex_lock(&wq_pool_attach_mutex); 5892 cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 5893 mutex_unlock(&wq_pool_attach_mutex); 5894 } 5895 return ret; 5896 } 5897 5898 /** 5899 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5900 * @cpumask: the cpumask to set 5901 * 5902 * The low-level workqueues cpumask is a global cpumask that limits 5903 * the affinity of all unbound workqueues. This function check the @cpumask 5904 * and apply it to all unbound workqueues and updates all pwqs of them. 5905 * 5906 * Return: 0 - Success 5907 * -EINVAL - Invalid @cpumask 5908 * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5909 */ 5910 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5911 { 5912 int ret = -EINVAL; 5913 5914 /* 5915 * Not excluding isolated cpus on purpose. 5916 * If the user wishes to include them, we allow that. 5917 */ 5918 cpumask_and(cpumask, cpumask, cpu_possible_mask); 5919 if (!cpumask_empty(cpumask)) { 5920 apply_wqattrs_lock(); 5921 if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5922 ret = 0; 5923 goto out_unlock; 5924 } 5925 5926 ret = workqueue_apply_unbound_cpumask(cpumask); 5927 5928 out_unlock: 5929 apply_wqattrs_unlock(); 5930 } 5931 5932 return ret; 5933 } 5934 5935 static int parse_affn_scope(const char *val) 5936 { 5937 int i; 5938 5939 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { 5940 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) 5941 return i; 5942 } 5943 return -EINVAL; 5944 } 5945 5946 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 5947 { 5948 struct workqueue_struct *wq; 5949 int affn, cpu; 5950 5951 affn = parse_affn_scope(val); 5952 if (affn < 0) 5953 return affn; 5954 if (affn == WQ_AFFN_DFL) 5955 return -EINVAL; 5956 5957 cpus_read_lock(); 5958 mutex_lock(&wq_pool_mutex); 5959 5960 wq_affn_dfl = affn; 5961 5962 list_for_each_entry(wq, &workqueues, list) { 5963 for_each_online_cpu(cpu) { 5964 wq_update_pod(wq, cpu, cpu, true); 5965 } 5966 } 5967 5968 mutex_unlock(&wq_pool_mutex); 5969 cpus_read_unlock(); 5970 5971 return 0; 5972 } 5973 5974 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 5975 { 5976 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 5977 } 5978 5979 static const struct kernel_param_ops wq_affn_dfl_ops = { 5980 .set = wq_affn_dfl_set, 5981 .get = wq_affn_dfl_get, 5982 }; 5983 5984 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 5985 5986 #ifdef CONFIG_SYSFS 5987 /* 5988 * Workqueues with WQ_SYSFS flag set is visible to userland via 5989 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 5990 * following attributes. 5991 * 5992 * per_cpu RO bool : whether the workqueue is per-cpu or unbound 5993 * max_active RW int : maximum number of in-flight work items 5994 * 5995 * Unbound workqueues have the following extra attributes. 5996 * 5997 * nice RW int : nice value of the workers 5998 * cpumask RW mask : bitmask of allowed CPUs for the workers 5999 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 6000 * affinity_strict RW bool : worker CPU affinity is strict 6001 */ 6002 struct wq_device { 6003 struct workqueue_struct *wq; 6004 struct device dev; 6005 }; 6006 6007 static struct workqueue_struct *dev_to_wq(struct device *dev) 6008 { 6009 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 6010 6011 return wq_dev->wq; 6012 } 6013 6014 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 6015 char *buf) 6016 { 6017 struct workqueue_struct *wq = dev_to_wq(dev); 6018 6019 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 6020 } 6021 static DEVICE_ATTR_RO(per_cpu); 6022 6023 static ssize_t max_active_show(struct device *dev, 6024 struct device_attribute *attr, char *buf) 6025 { 6026 struct workqueue_struct *wq = dev_to_wq(dev); 6027 6028 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 6029 } 6030 6031 static ssize_t max_active_store(struct device *dev, 6032 struct device_attribute *attr, const char *buf, 6033 size_t count) 6034 { 6035 struct workqueue_struct *wq = dev_to_wq(dev); 6036 int val; 6037 6038 if (sscanf(buf, "%d", &val) != 1 || val <= 0) 6039 return -EINVAL; 6040 6041 workqueue_set_max_active(wq, val); 6042 return count; 6043 } 6044 static DEVICE_ATTR_RW(max_active); 6045 6046 static struct attribute *wq_sysfs_attrs[] = { 6047 &dev_attr_per_cpu.attr, 6048 &dev_attr_max_active.attr, 6049 NULL, 6050 }; 6051 ATTRIBUTE_GROUPS(wq_sysfs); 6052 6053 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 6054 char *buf) 6055 { 6056 struct workqueue_struct *wq = dev_to_wq(dev); 6057 int written; 6058 6059 mutex_lock(&wq->mutex); 6060 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 6061 mutex_unlock(&wq->mutex); 6062 6063 return written; 6064 } 6065 6066 /* prepare workqueue_attrs for sysfs store operations */ 6067 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 6068 { 6069 struct workqueue_attrs *attrs; 6070 6071 lockdep_assert_held(&wq_pool_mutex); 6072 6073 attrs = alloc_workqueue_attrs(); 6074 if (!attrs) 6075 return NULL; 6076 6077 copy_workqueue_attrs(attrs, wq->unbound_attrs); 6078 return attrs; 6079 } 6080 6081 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 6082 const char *buf, size_t count) 6083 { 6084 struct workqueue_struct *wq = dev_to_wq(dev); 6085 struct workqueue_attrs *attrs; 6086 int ret = -ENOMEM; 6087 6088 apply_wqattrs_lock(); 6089 6090 attrs = wq_sysfs_prep_attrs(wq); 6091 if (!attrs) 6092 goto out_unlock; 6093 6094 if (sscanf(buf, "%d", &attrs->nice) == 1 && 6095 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 6096 ret = apply_workqueue_attrs_locked(wq, attrs); 6097 else 6098 ret = -EINVAL; 6099 6100 out_unlock: 6101 apply_wqattrs_unlock(); 6102 free_workqueue_attrs(attrs); 6103 return ret ?: count; 6104 } 6105 6106 static ssize_t wq_cpumask_show(struct device *dev, 6107 struct device_attribute *attr, char *buf) 6108 { 6109 struct workqueue_struct *wq = dev_to_wq(dev); 6110 int written; 6111 6112 mutex_lock(&wq->mutex); 6113 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6114 cpumask_pr_args(wq->unbound_attrs->cpumask)); 6115 mutex_unlock(&wq->mutex); 6116 return written; 6117 } 6118 6119 static ssize_t wq_cpumask_store(struct device *dev, 6120 struct device_attribute *attr, 6121 const char *buf, size_t count) 6122 { 6123 struct workqueue_struct *wq = dev_to_wq(dev); 6124 struct workqueue_attrs *attrs; 6125 int ret = -ENOMEM; 6126 6127 apply_wqattrs_lock(); 6128 6129 attrs = wq_sysfs_prep_attrs(wq); 6130 if (!attrs) 6131 goto out_unlock; 6132 6133 ret = cpumask_parse(buf, attrs->cpumask); 6134 if (!ret) 6135 ret = apply_workqueue_attrs_locked(wq, attrs); 6136 6137 out_unlock: 6138 apply_wqattrs_unlock(); 6139 free_workqueue_attrs(attrs); 6140 return ret ?: count; 6141 } 6142 6143 static ssize_t wq_affn_scope_show(struct device *dev, 6144 struct device_attribute *attr, char *buf) 6145 { 6146 struct workqueue_struct *wq = dev_to_wq(dev); 6147 int written; 6148 6149 mutex_lock(&wq->mutex); 6150 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) 6151 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", 6152 wq_affn_names[WQ_AFFN_DFL], 6153 wq_affn_names[wq_affn_dfl]); 6154 else 6155 written = scnprintf(buf, PAGE_SIZE, "%s\n", 6156 wq_affn_names[wq->unbound_attrs->affn_scope]); 6157 mutex_unlock(&wq->mutex); 6158 6159 return written; 6160 } 6161 6162 static ssize_t wq_affn_scope_store(struct device *dev, 6163 struct device_attribute *attr, 6164 const char *buf, size_t count) 6165 { 6166 struct workqueue_struct *wq = dev_to_wq(dev); 6167 struct workqueue_attrs *attrs; 6168 int affn, ret = -ENOMEM; 6169 6170 affn = parse_affn_scope(buf); 6171 if (affn < 0) 6172 return affn; 6173 6174 apply_wqattrs_lock(); 6175 attrs = wq_sysfs_prep_attrs(wq); 6176 if (attrs) { 6177 attrs->affn_scope = affn; 6178 ret = apply_workqueue_attrs_locked(wq, attrs); 6179 } 6180 apply_wqattrs_unlock(); 6181 free_workqueue_attrs(attrs); 6182 return ret ?: count; 6183 } 6184 6185 static ssize_t wq_affinity_strict_show(struct device *dev, 6186 struct device_attribute *attr, char *buf) 6187 { 6188 struct workqueue_struct *wq = dev_to_wq(dev); 6189 6190 return scnprintf(buf, PAGE_SIZE, "%d\n", 6191 wq->unbound_attrs->affn_strict); 6192 } 6193 6194 static ssize_t wq_affinity_strict_store(struct device *dev, 6195 struct device_attribute *attr, 6196 const char *buf, size_t count) 6197 { 6198 struct workqueue_struct *wq = dev_to_wq(dev); 6199 struct workqueue_attrs *attrs; 6200 int v, ret = -ENOMEM; 6201 6202 if (sscanf(buf, "%d", &v) != 1) 6203 return -EINVAL; 6204 6205 apply_wqattrs_lock(); 6206 attrs = wq_sysfs_prep_attrs(wq); 6207 if (attrs) { 6208 attrs->affn_strict = (bool)v; 6209 ret = apply_workqueue_attrs_locked(wq, attrs); 6210 } 6211 apply_wqattrs_unlock(); 6212 free_workqueue_attrs(attrs); 6213 return ret ?: count; 6214 } 6215 6216 static struct device_attribute wq_sysfs_unbound_attrs[] = { 6217 __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 6218 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 6219 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 6220 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), 6221 __ATTR_NULL, 6222 }; 6223 6224 static struct bus_type wq_subsys = { 6225 .name = "workqueue", 6226 .dev_groups = wq_sysfs_groups, 6227 }; 6228 6229 static ssize_t wq_unbound_cpumask_show(struct device *dev, 6230 struct device_attribute *attr, char *buf) 6231 { 6232 int written; 6233 6234 mutex_lock(&wq_pool_mutex); 6235 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6236 cpumask_pr_args(wq_unbound_cpumask)); 6237 mutex_unlock(&wq_pool_mutex); 6238 6239 return written; 6240 } 6241 6242 static ssize_t wq_unbound_cpumask_store(struct device *dev, 6243 struct device_attribute *attr, const char *buf, size_t count) 6244 { 6245 cpumask_var_t cpumask; 6246 int ret; 6247 6248 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6249 return -ENOMEM; 6250 6251 ret = cpumask_parse(buf, cpumask); 6252 if (!ret) 6253 ret = workqueue_set_unbound_cpumask(cpumask); 6254 6255 free_cpumask_var(cpumask); 6256 return ret ? ret : count; 6257 } 6258 6259 static struct device_attribute wq_sysfs_cpumask_attr = 6260 __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6261 wq_unbound_cpumask_store); 6262 6263 static int __init wq_sysfs_init(void) 6264 { 6265 struct device *dev_root; 6266 int err; 6267 6268 err = subsys_virtual_register(&wq_subsys, NULL); 6269 if (err) 6270 return err; 6271 6272 dev_root = bus_get_dev_root(&wq_subsys); 6273 if (dev_root) { 6274 err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6275 put_device(dev_root); 6276 } 6277 return err; 6278 } 6279 core_initcall(wq_sysfs_init); 6280 6281 static void wq_device_release(struct device *dev) 6282 { 6283 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 6284 6285 kfree(wq_dev); 6286 } 6287 6288 /** 6289 * workqueue_sysfs_register - make a workqueue visible in sysfs 6290 * @wq: the workqueue to register 6291 * 6292 * Expose @wq in sysfs under /sys/bus/workqueue/devices. 6293 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 6294 * which is the preferred method. 6295 * 6296 * Workqueue user should use this function directly iff it wants to apply 6297 * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 6298 * apply_workqueue_attrs() may race against userland updating the 6299 * attributes. 6300 * 6301 * Return: 0 on success, -errno on failure. 6302 */ 6303 int workqueue_sysfs_register(struct workqueue_struct *wq) 6304 { 6305 struct wq_device *wq_dev; 6306 int ret; 6307 6308 /* 6309 * Adjusting max_active or creating new pwqs by applying 6310 * attributes breaks ordering guarantee. Disallow exposing ordered 6311 * workqueues. 6312 */ 6313 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 6314 return -EINVAL; 6315 6316 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 6317 if (!wq_dev) 6318 return -ENOMEM; 6319 6320 wq_dev->wq = wq; 6321 wq_dev->dev.bus = &wq_subsys; 6322 wq_dev->dev.release = wq_device_release; 6323 dev_set_name(&wq_dev->dev, "%s", wq->name); 6324 6325 /* 6326 * unbound_attrs are created separately. Suppress uevent until 6327 * everything is ready. 6328 */ 6329 dev_set_uevent_suppress(&wq_dev->dev, true); 6330 6331 ret = device_register(&wq_dev->dev); 6332 if (ret) { 6333 put_device(&wq_dev->dev); 6334 wq->wq_dev = NULL; 6335 return ret; 6336 } 6337 6338 if (wq->flags & WQ_UNBOUND) { 6339 struct device_attribute *attr; 6340 6341 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 6342 ret = device_create_file(&wq_dev->dev, attr); 6343 if (ret) { 6344 device_unregister(&wq_dev->dev); 6345 wq->wq_dev = NULL; 6346 return ret; 6347 } 6348 } 6349 } 6350 6351 dev_set_uevent_suppress(&wq_dev->dev, false); 6352 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 6353 return 0; 6354 } 6355 6356 /** 6357 * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 6358 * @wq: the workqueue to unregister 6359 * 6360 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 6361 */ 6362 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 6363 { 6364 struct wq_device *wq_dev = wq->wq_dev; 6365 6366 if (!wq->wq_dev) 6367 return; 6368 6369 wq->wq_dev = NULL; 6370 device_unregister(&wq_dev->dev); 6371 } 6372 #else /* CONFIG_SYSFS */ 6373 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 6374 #endif /* CONFIG_SYSFS */ 6375 6376 /* 6377 * Workqueue watchdog. 6378 * 6379 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 6380 * flush dependency, a concurrency managed work item which stays RUNNING 6381 * indefinitely. Workqueue stalls can be very difficult to debug as the 6382 * usual warning mechanisms don't trigger and internal workqueue state is 6383 * largely opaque. 6384 * 6385 * Workqueue watchdog monitors all worker pools periodically and dumps 6386 * state if some pools failed to make forward progress for a while where 6387 * forward progress is defined as the first item on ->worklist changing. 6388 * 6389 * This mechanism is controlled through the kernel parameter 6390 * "workqueue.watchdog_thresh" which can be updated at runtime through the 6391 * corresponding sysfs parameter file. 6392 */ 6393 #ifdef CONFIG_WQ_WATCHDOG 6394 6395 static unsigned long wq_watchdog_thresh = 30; 6396 static struct timer_list wq_watchdog_timer; 6397 6398 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 6399 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 6400 6401 /* 6402 * Show workers that might prevent the processing of pending work items. 6403 * The only candidates are CPU-bound workers in the running state. 6404 * Pending work items should be handled by another idle worker 6405 * in all other situations. 6406 */ 6407 static void show_cpu_pool_hog(struct worker_pool *pool) 6408 { 6409 struct worker *worker; 6410 unsigned long flags; 6411 int bkt; 6412 6413 raw_spin_lock_irqsave(&pool->lock, flags); 6414 6415 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6416 if (task_is_running(worker->task)) { 6417 /* 6418 * Defer printing to avoid deadlocks in console 6419 * drivers that queue work while holding locks 6420 * also taken in their write paths. 6421 */ 6422 printk_deferred_enter(); 6423 6424 pr_info("pool %d:\n", pool->id); 6425 sched_show_task(worker->task); 6426 6427 printk_deferred_exit(); 6428 } 6429 } 6430 6431 raw_spin_unlock_irqrestore(&pool->lock, flags); 6432 } 6433 6434 static void show_cpu_pools_hogs(void) 6435 { 6436 struct worker_pool *pool; 6437 int pi; 6438 6439 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6440 6441 rcu_read_lock(); 6442 6443 for_each_pool(pool, pi) { 6444 if (pool->cpu_stall) 6445 show_cpu_pool_hog(pool); 6446 6447 } 6448 6449 rcu_read_unlock(); 6450 } 6451 6452 static void wq_watchdog_reset_touched(void) 6453 { 6454 int cpu; 6455 6456 wq_watchdog_touched = jiffies; 6457 for_each_possible_cpu(cpu) 6458 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 6459 } 6460 6461 static void wq_watchdog_timer_fn(struct timer_list *unused) 6462 { 6463 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 6464 bool lockup_detected = false; 6465 bool cpu_pool_stall = false; 6466 unsigned long now = jiffies; 6467 struct worker_pool *pool; 6468 int pi; 6469 6470 if (!thresh) 6471 return; 6472 6473 rcu_read_lock(); 6474 6475 for_each_pool(pool, pi) { 6476 unsigned long pool_ts, touched, ts; 6477 6478 pool->cpu_stall = false; 6479 if (list_empty(&pool->worklist)) 6480 continue; 6481 6482 /* 6483 * If a virtual machine is stopped by the host it can look to 6484 * the watchdog like a stall. 6485 */ 6486 kvm_check_and_clear_guest_paused(); 6487 6488 /* get the latest of pool and touched timestamps */ 6489 if (pool->cpu >= 0) 6490 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 6491 else 6492 touched = READ_ONCE(wq_watchdog_touched); 6493 pool_ts = READ_ONCE(pool->watchdog_ts); 6494 6495 if (time_after(pool_ts, touched)) 6496 ts = pool_ts; 6497 else 6498 ts = touched; 6499 6500 /* did we stall? */ 6501 if (time_after(now, ts + thresh)) { 6502 lockup_detected = true; 6503 if (pool->cpu >= 0) { 6504 pool->cpu_stall = true; 6505 cpu_pool_stall = true; 6506 } 6507 pr_emerg("BUG: workqueue lockup - pool"); 6508 pr_cont_pool_info(pool); 6509 pr_cont(" stuck for %us!\n", 6510 jiffies_to_msecs(now - pool_ts) / 1000); 6511 } 6512 6513 6514 } 6515 6516 rcu_read_unlock(); 6517 6518 if (lockup_detected) 6519 show_all_workqueues(); 6520 6521 if (cpu_pool_stall) 6522 show_cpu_pools_hogs(); 6523 6524 wq_watchdog_reset_touched(); 6525 mod_timer(&wq_watchdog_timer, jiffies + thresh); 6526 } 6527 6528 notrace void wq_watchdog_touch(int cpu) 6529 { 6530 if (cpu >= 0) 6531 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 6532 6533 wq_watchdog_touched = jiffies; 6534 } 6535 6536 static void wq_watchdog_set_thresh(unsigned long thresh) 6537 { 6538 wq_watchdog_thresh = 0; 6539 del_timer_sync(&wq_watchdog_timer); 6540 6541 if (thresh) { 6542 wq_watchdog_thresh = thresh; 6543 wq_watchdog_reset_touched(); 6544 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 6545 } 6546 } 6547 6548 static int wq_watchdog_param_set_thresh(const char *val, 6549 const struct kernel_param *kp) 6550 { 6551 unsigned long thresh; 6552 int ret; 6553 6554 ret = kstrtoul(val, 0, &thresh); 6555 if (ret) 6556 return ret; 6557 6558 if (system_wq) 6559 wq_watchdog_set_thresh(thresh); 6560 else 6561 wq_watchdog_thresh = thresh; 6562 6563 return 0; 6564 } 6565 6566 static const struct kernel_param_ops wq_watchdog_thresh_ops = { 6567 .set = wq_watchdog_param_set_thresh, 6568 .get = param_get_ulong, 6569 }; 6570 6571 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 6572 0644); 6573 6574 static void wq_watchdog_init(void) 6575 { 6576 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 6577 wq_watchdog_set_thresh(wq_watchdog_thresh); 6578 } 6579 6580 #else /* CONFIG_WQ_WATCHDOG */ 6581 6582 static inline void wq_watchdog_init(void) { } 6583 6584 #endif /* CONFIG_WQ_WATCHDOG */ 6585 6586 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) 6587 { 6588 if (!cpumask_intersects(wq_unbound_cpumask, mask)) { 6589 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", 6590 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); 6591 return; 6592 } 6593 6594 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); 6595 } 6596 6597 /** 6598 * workqueue_init_early - early init for workqueue subsystem 6599 * 6600 * This is the first step of three-staged workqueue subsystem initialization and 6601 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 6602 * up. It sets up all the data structures and system workqueues and allows early 6603 * boot code to create workqueues and queue/cancel work items. Actual work item 6604 * execution starts only after kthreads can be created and scheduled right 6605 * before early initcalls. 6606 */ 6607 void __init workqueue_init_early(void) 6608 { 6609 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 6610 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 6611 int i, cpu; 6612 6613 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6614 6615 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 6616 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 6617 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); 6618 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); 6619 if (!cpumask_empty(&wq_cmdline_cpumask)) 6620 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); 6621 6622 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6623 6624 wq_update_pod_attrs_buf = alloc_workqueue_attrs(); 6625 BUG_ON(!wq_update_pod_attrs_buf); 6626 6627 /* initialize WQ_AFFN_SYSTEM pods */ 6628 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 6629 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); 6630 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 6631 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 6632 6633 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 6634 6635 pt->nr_pods = 1; 6636 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 6637 pt->pod_node[0] = NUMA_NO_NODE; 6638 pt->cpu_pod[0] = 0; 6639 6640 /* initialize CPU pools */ 6641 for_each_possible_cpu(cpu) { 6642 struct worker_pool *pool; 6643 6644 i = 0; 6645 for_each_cpu_worker_pool(pool, cpu) { 6646 BUG_ON(init_worker_pool(pool)); 6647 pool->cpu = cpu; 6648 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 6649 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 6650 pool->attrs->nice = std_nice[i++]; 6651 pool->attrs->affn_strict = true; 6652 pool->node = cpu_to_node(cpu); 6653 6654 /* alloc pool ID */ 6655 mutex_lock(&wq_pool_mutex); 6656 BUG_ON(worker_pool_assign_id(pool)); 6657 mutex_unlock(&wq_pool_mutex); 6658 } 6659 } 6660 6661 /* create default unbound and ordered wq attrs */ 6662 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 6663 struct workqueue_attrs *attrs; 6664 6665 BUG_ON(!(attrs = alloc_workqueue_attrs())); 6666 attrs->nice = std_nice[i]; 6667 unbound_std_wq_attrs[i] = attrs; 6668 6669 /* 6670 * An ordered wq should have only one pwq as ordering is 6671 * guaranteed by max_active which is enforced by pwqs. 6672 */ 6673 BUG_ON(!(attrs = alloc_workqueue_attrs())); 6674 attrs->nice = std_nice[i]; 6675 attrs->ordered = true; 6676 ordered_wq_attrs[i] = attrs; 6677 } 6678 6679 system_wq = alloc_workqueue("events", 0, 0); 6680 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6681 system_long_wq = alloc_workqueue("events_long", 0, 0); 6682 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6683 WQ_MAX_ACTIVE); 6684 system_freezable_wq = alloc_workqueue("events_freezable", 6685 WQ_FREEZABLE, 0); 6686 system_power_efficient_wq = alloc_workqueue("events_power_efficient", 6687 WQ_POWER_EFFICIENT, 0); 6688 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 6689 WQ_FREEZABLE | WQ_POWER_EFFICIENT, 6690 0); 6691 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 6692 !system_unbound_wq || !system_freezable_wq || 6693 !system_power_efficient_wq || 6694 !system_freezable_power_efficient_wq); 6695 } 6696 6697 static void __init wq_cpu_intensive_thresh_init(void) 6698 { 6699 unsigned long thresh; 6700 unsigned long bogo; 6701 6702 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); 6703 BUG_ON(IS_ERR(pwq_release_worker)); 6704 6705 /* if the user set it to a specific value, keep it */ 6706 if (wq_cpu_intensive_thresh_us != ULONG_MAX) 6707 return; 6708 6709 /* 6710 * The default of 10ms is derived from the fact that most modern (as of 6711 * 2023) processors can do a lot in 10ms and that it's just below what 6712 * most consider human-perceivable. However, the kernel also runs on a 6713 * lot slower CPUs including microcontrollers where the threshold is way 6714 * too low. 6715 * 6716 * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 6717 * This is by no means accurate but it doesn't have to be. The mechanism 6718 * is still useful even when the threshold is fully scaled up. Also, as 6719 * the reports would usually be applicable to everyone, some machines 6720 * operating on longer thresholds won't significantly diminish their 6721 * usefulness. 6722 */ 6723 thresh = 10 * USEC_PER_MSEC; 6724 6725 /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 6726 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 6727 if (bogo < 4000) 6728 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 6729 6730 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 6731 loops_per_jiffy, bogo, thresh); 6732 6733 wq_cpu_intensive_thresh_us = thresh; 6734 } 6735 6736 /** 6737 * workqueue_init - bring workqueue subsystem fully online 6738 * 6739 * This is the second step of three-staged workqueue subsystem initialization 6740 * and invoked as soon as kthreads can be created and scheduled. Workqueues have 6741 * been created and work items queued on them, but there are no kworkers 6742 * executing the work items yet. Populate the worker pools with the initial 6743 * workers and enable future kworker creations. 6744 */ 6745 void __init workqueue_init(void) 6746 { 6747 struct workqueue_struct *wq; 6748 struct worker_pool *pool; 6749 int cpu, bkt; 6750 6751 wq_cpu_intensive_thresh_init(); 6752 6753 mutex_lock(&wq_pool_mutex); 6754 6755 /* 6756 * Per-cpu pools created earlier could be missing node hint. Fix them 6757 * up. Also, create a rescuer for workqueues that requested it. 6758 */ 6759 for_each_possible_cpu(cpu) { 6760 for_each_cpu_worker_pool(pool, cpu) { 6761 pool->node = cpu_to_node(cpu); 6762 } 6763 } 6764 6765 list_for_each_entry(wq, &workqueues, list) { 6766 WARN(init_rescuer(wq), 6767 "workqueue: failed to create early rescuer for %s", 6768 wq->name); 6769 } 6770 6771 mutex_unlock(&wq_pool_mutex); 6772 6773 /* create the initial workers */ 6774 for_each_online_cpu(cpu) { 6775 for_each_cpu_worker_pool(pool, cpu) { 6776 pool->flags &= ~POOL_DISASSOCIATED; 6777 BUG_ON(!create_worker(pool)); 6778 } 6779 } 6780 6781 hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 6782 BUG_ON(!create_worker(pool)); 6783 6784 wq_online = true; 6785 wq_watchdog_init(); 6786 } 6787 6788 /* 6789 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 6790 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 6791 * and consecutive pod ID. The rest of @pt is initialized accordingly. 6792 */ 6793 static void __init init_pod_type(struct wq_pod_type *pt, 6794 bool (*cpus_share_pod)(int, int)) 6795 { 6796 int cur, pre, cpu, pod; 6797 6798 pt->nr_pods = 0; 6799 6800 /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 6801 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 6802 BUG_ON(!pt->cpu_pod); 6803 6804 for_each_possible_cpu(cur) { 6805 for_each_possible_cpu(pre) { 6806 if (pre >= cur) { 6807 pt->cpu_pod[cur] = pt->nr_pods++; 6808 break; 6809 } 6810 if (cpus_share_pod(cur, pre)) { 6811 pt->cpu_pod[cur] = pt->cpu_pod[pre]; 6812 break; 6813 } 6814 } 6815 } 6816 6817 /* init the rest to match @pt->cpu_pod[] */ 6818 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 6819 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); 6820 BUG_ON(!pt->pod_cpus || !pt->pod_node); 6821 6822 for (pod = 0; pod < pt->nr_pods; pod++) 6823 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 6824 6825 for_each_possible_cpu(cpu) { 6826 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 6827 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 6828 } 6829 } 6830 6831 static bool __init cpus_dont_share(int cpu0, int cpu1) 6832 { 6833 return false; 6834 } 6835 6836 static bool __init cpus_share_smt(int cpu0, int cpu1) 6837 { 6838 #ifdef CONFIG_SCHED_SMT 6839 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 6840 #else 6841 return false; 6842 #endif 6843 } 6844 6845 static bool __init cpus_share_numa(int cpu0, int cpu1) 6846 { 6847 return cpu_to_node(cpu0) == cpu_to_node(cpu1); 6848 } 6849 6850 /** 6851 * workqueue_init_topology - initialize CPU pods for unbound workqueues 6852 * 6853 * This is the third step of there-staged workqueue subsystem initialization and 6854 * invoked after SMP and topology information are fully initialized. It 6855 * initializes the unbound CPU pods accordingly. 6856 */ 6857 void __init workqueue_init_topology(void) 6858 { 6859 struct workqueue_struct *wq; 6860 int cpu; 6861 6862 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 6863 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 6864 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 6865 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 6866 6867 mutex_lock(&wq_pool_mutex); 6868 6869 /* 6870 * Workqueues allocated earlier would have all CPUs sharing the default 6871 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU 6872 * combinations to apply per-pod sharing. 6873 */ 6874 list_for_each_entry(wq, &workqueues, list) { 6875 for_each_online_cpu(cpu) { 6876 wq_update_pod(wq, cpu, cpu, true); 6877 } 6878 } 6879 6880 mutex_unlock(&wq_pool_mutex); 6881 } 6882 6883 void __warn_flushing_systemwide_wq(void) 6884 { 6885 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 6886 dump_stack(); 6887 } 6888 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6889 6890 static int __init workqueue_unbound_cpus_setup(char *str) 6891 { 6892 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 6893 cpumask_clear(&wq_cmdline_cpumask); 6894 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 6895 } 6896 6897 return 1; 6898 } 6899 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 6900